Part A1: Differential Equations I

∗
Lecture notes by Janet Dyson

MT 2021
Lecturer: Melanie Rupflin
Version from October 19, 2022

Contents

1 ODEs and Picard’s Theorem 5

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Picard’s method of successive approximation . . . . . . . . . . . . . . . . 7
1.3 Picard’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Extension of solutions and global existence. . . . . . . . . . . . . . . . . . 14
1.5 Gronwall’s inequality and continuous dependence on the initial data. . . . 15
1.6 Picard’s Theorem via the CMT . . . . . . . . . . . . . . . . . . . . . . . . 16
1.7 Picard’s Theorem for systems and higher order ODEs via the CMT . . . . 21
1.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Plane autonomous systems of ODEs 25

2.1 Critical points and closed trajectories . . . . . . . . . . . . . . . . . . . . 26
2.1.1 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Stability and linearisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Classification of critical points . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.2 Further example: the damped pendulum . . . . . . . . . . . . . . . 37
2.3.3 An important example: The Lotka–Volterra predator-prey equations 38
∗
based on notes by Peter Grindrod, Colin Please, Paul Tod, Lionel Mason and others, with modifi-
cations by the lecturer

1
 2.3.4 Another example from population dynamics. . . . . . . . . . . . . 39
2.3.5 Another important example: limit cycles . . . . . . . . . . . . . . . 40
2.4 The Bendixson–Dulac Theorem . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.1 Corollary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3 First order semi-linear PDEs: the method of characteristics 46

3.1 The problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 The big idea: characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 Examples of characteristics . . . . . . . . . . . . . . . . . . . . . . 48
3.3 The Cauchy problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5 Domain of definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 Cauchy data and the Cauchy Problem: . . . . . . . . . . . . . . . . . . . 56
3.7 Discontinuities in the first derivatives . . . . . . . . . . . . . . . . . . . . . 56
3.8 General Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4 Second order semi-linear PDEs 59

4.1 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1.1 The idea: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1.2 The Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2 Characteristics: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Type and data: well posed problems . . . . . . . . . . . . . . . . . . . . . 67
4.4 The Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.1 Poisson’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.2 The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5 Where does this course lead? 79

5.1 Section 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 Section 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3 Section 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4 Section 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

2
Introduction

The solution of problems in most parts of applied mathematics and many areas of pure
can more often than not be reduced to the problem of solving some differential equa-
tions. Indeed, many parts of pure maths were originally motivated by issues arising from
differential equations, including large parts of algebra and much of analysis, and Differ-
ential equations are a central topic in research in both pure and applied mathematics to
this day. From Prelims, and even from school, you know how to solve some differential
equations. Indeed most of the study of differential equations in the first year consisted
of finding explicit solutions of particular ODEs or PDEs. However, for many differential
equations which arise in practice one is unable to give explicit solutions and, for the
most part, this course will consider what information one can discover about solutions
without actually finding the solution. Does a solution exist? Is it unique? Does it
depend continuously on the initial data? How does it behave asymptotically? What is
appropriate data?
So, first we will develop techniques for proving Picard’s theorem for the existence and
uniqueness of solutions of ODEs; then we will look at how phase plane analysis enables
us to estimate the long term behaviour of solutions of plane autonomous systems of
ODEs. We will then turn to PDEs and show how the method of characteristics reduces
the solution of a first order semi-linear PDE to solving a system of non-linear ODEs.
Finally we will look at second order semi-linear PDEs: We classify them and investigate
how the different types of problem require different types of boundary data if the prob-
lem is to be well posed. We then look at how the maximum principle enables us to prove
uniqueness and continuous dependence on the initial data for two very special problems:
Poisson’s equation and the inhomogeneous heat equation – each with suitable data.

Throughout, we shall use the following convenient abbreviations: we shall write

DEs: for differential equations.
ODEs: for ordinary DEs, i.e. differential equations with only ordinary derivatives.
PDEs: for partial DEs, i.e. differential equations with partial derivatives.
The course contains four topics, with a section devoted to each. The chapters are:

1. ODEs and Picard’s Theorem (for existence/uniqueness of solutions/continuous de-

pendence on initial data).

2. Plane autonomous systems of ODEs

3. First order semi-linear PDEs: the method of characteristics.

4. Second-order semi-linear PDEs: classification; well posedness; the Maximum Prin-

ciple and its consequences

3
Books
The main text is P J Collins Differential and Integral Equations, O.U.P. (2006), which
can be used for the whole course (Chapters 1-7, 14, 15).
Other good books which cover parts of the course include

W E Boyce and R C DiPrima, Elementary Differential Equations and Boundary Value

Problems, 7th edition, Wiley (2000).

E Kreyszig, Advanced Engineering Mathematics, 8th Edition, Wiley (1999).

G F Carrier and C E Pearson, Partial Differential Equations – Theory and Technique,

Academic (1988).

J Ockendon, S Howison, A Lacey and A Movchan, Applied Partial Differential Equations,

Oxford (1999) [a more advanced text].

4
PART I Ordinary Differential Equations

1 ODEs and Picard’s Theorem

1.1 Introduction

An ODE is an equation for y(x) of the form

G(x, y, y 0 , y 00 , ..., y (n) ) = 0.

We refer to y as the dependent variable and x as the independent variable. Usually this
can be solved for the highest derivative of y and written in the form
dn y
:= y (n) (x) = F (x, y, y 0 , ..., y (n−1) ).
dxn
Then the order of the ODE is n, the order of the highest derivative which appears.
Given an ODE, certain obvious questions arise. We could ask:

• Does it have solutions? Can we find them (explicitly or implicitly)? If not, can we
at least say something about their qualitative behaviour?

• Given data e.g. the values y(a), y 0 (a), ... of y(x) and its first n − 1 derivatives
at some initial value a of x, does it have a solution? is it unique? does it depend
continuously on the given data?

We shall consider these questions in Part I.

For simplicity, we begin with a first-order ODE with data:

y 0 (x) = f (x, y(x)) with y(a) = b. (1.1)

This is an initial value problem or IVP, since we are given y at an initial, or starting,
value of x. You know how to solve a variety of equations like this. You might expect that
a solution exists, so there is some function that satisfies the equations (even if you cannot
find a formula for it) and perhaps you expect, for a given initial values, the solution is
unique (there is only one function that satisfies the ODE and the initial data) as you
may not have encountered the following difficulties.

Warning examples:

Consider this IVP:

y 0 = 3y 2/3 ; y(0) = 0. (1.2)

5
So separate the variables (a prelims technique):
Z Z
dy
= dx,
3y 2/3

to get y = (x + A)3 , so if y(0) = 0

(i) There is a solution y = x3 ;

(ii) But evidently there is another solution: by direct checking y = 0 will do;

(iii) In fact we can find that there are infinitely many solutions. Pick a, b with a ≤ 0 ≤ b
and define

y = (x − a)3 x<a
= 0 a≤x<b
3
= (x − b) b≤x

The solution does exist but is not unique (in fact far from it, since we’ve found infinitely
many solutions).

Furthermore, even if a solution of (1.1) exists, it may not exist for all x.
For example consider the IVP

y0 = y2; y(0) = 1. (1.3)

1
Using separation of variables we can see this has solution y = ; so y → ∞ as
1−x
x → 1, and the solution only exists on x < 1. We will see later that this solution is in
fact unique.
So, if we want to have a unique solution to the problem (1.1) we must impose conditions
on f , and we cannot necessarily expect to have solutions for all x. This will be the first
existence theorem which you’ve encountered. The proof is quite technical, certainly the
most technical thing in the course. In particular, you need to remember from Prelims
the Weierstrass M-test for convergence of a series of functions.
To be precise then, we shall seek a solution of problem (1.1) in a rectangle R about the
initial point (x, y) = (a, b), so suppose R = {(x, y) : |x − a| ≤ h, |y − b| ≤ k} as in figure
1.1:

6
 y

2h

b 2k
R

a x

Figure 1.1: The rectangle R

Our first assumption is that f : R → R is continuous in R. Note that if y is a

solution of (1.1), say on an interval [a−h, a+h], then integrating (1.1) from a to variable
x ∈ [a − h, a + h] yields
Z x
x
y(x) − y(a) = [y(t)]a = f (t, y(t))dt
a

for any x so rearranging Z x

y(x) = b + f (t, y(t))dt. (1.4)
a
We note that since f : R → R and y : [a − h, a + h] → R are continuous, also the function
x 7→ f (x, y(x)) is a continuous function on [a − h, a + h] so integrable.
Conversely, if y(t) is continuous on [a − h, a + h] and satisfies (1.4), then y(a) = b
and by the Fundamental theorem of Calculus we can differentiate (1.4) to get that y
is a solution of (1.1). Thus (1.1) and (1.4) are equivalent. We have transformed the
differential equation to an integral equation - the unknown y is given in terms of an
integral rather than a differential. There is a general theory of these, but we only need
to deal with the particular case of (1.4). The standard approach is to seek a solution by
iteration or successive approximation.

1.2 Picard’s method of successive approximation

We start with an initial guess and then improve it. The guesses, or successive approxi-
mations or iterates, are labeled yn (x) starting with y0 (x). Take

y0 (x)R = b
x (1.5)
yn+1 (x) = b + a f (t, yn (t))dt

That is, we start with the simplest guess, that y equals its initial value, and at each stage
substitute the current guess into the right-hand-side of (1.4) to get the next guess. We

7
need to know if this process converges, and if it does whether it converges to a solution
of the problem (1.4). Consider the differences between successive approximations:

e0 (x) = b
(1.6)
en+1 (x) = yn+1 (x) − yn (x)
and note that
n
X
yn (x) = ek (x). (1.7)
0
Pn
We want yn to converge, so we want the series 0 ek (x) to converge. So we must
estimate the differences en (x). This will need a condition on f , but here is the key idea.
Notice that
Z x
en+1 (x) = yn+1 (x) − yn (x) = [f (t, yn (t)) − f (t, yn−1 (t))]dt
a

and recall that the modulus of the integral of a function is less than or equal to the
integral of the modulus (because the function can be negative). Therefore
Z x
| en+1 (x)| ≤ |f (t, yn (t)) − f (t, yn−1 (t))| dt . (1.8)
a

(The modulus outside the integral on the right hand side is required to cover the case
x ≤ a.) We want to bound the integrand on the right-hand side in terms of the error
en (t) = |yn (t) − yn−1 (t)| of the previous step which motivates the following definition:
Definition 1.1. A function f (x, y) on a rectangle R satisfies a Lipschitz condition (with
constant L) if ∃ real positive L such that
|f (x, u) − f (x, v)| ≤ L|u − v| for all (x, u) ∈ R, (x, v) ∈ R. (1.9)

This is a new condition on a function, stronger than being continuous in the second
variable but weaker than being differentiable. It turns out to be the right condition to
make Picard’s theorem, which is the existence theorem we want, work, as it allows us to
bound the integrand in (1.8) by
|f (t, yn (t)) − f (t, yn−1 (t))| ≤ L|en (t)|.

Important note: One way to ensure that f satisfies a Lipschitz condition on R is the
following: Suppose that, on R, f is differentiable with respect to y , with |fy (x, y)| ≤ K.
Then for any (x, u) ∈ R, (x, v) ∈ R the mean value theorem (applied to the function
[k − h, k + h] 3 y 7→ f (x, y)) gives
|f (x, u) − f (x, v)| = |fy (x, w)(u − v)| ≤ K|u − v| (1.10)
where w is some intermediate value. So, f clearly satisfies the Lipschitz condition on
such intervals.
On the other hand f (y) = |y| is Lipschitz continuous, but is not differentiable at y = 0.

8
1.3 Picard’s Theorem

Theorem 1.1. (Picard’s existence theorem):

Let f : R → R be a function defined on the rectangle R := {(x, y) : |x−a| ≤ h, |y−b| ≤ k}
which satisfies
P(i): (a) f is continuous in R, with bound M (so |f (x, y)| ≤ M ) and (b) M h ≤ k.
P(ii): f satisfies a Lipschitz condition in R.
Then the IVP
y 0 (x) = f (x, y(x)) with y(a) = b.
has a unique solution y : [a − h, a + h] → [b − k, b + k].

We will prove the existence of a solution by showing that the iterates yn defined in (1.5)
converge as n → ∞ to a solution y of the IVP and will do this by showing that the series
in (1.7) converges. We break the proof into a series of steps:
Claim 1: Each yn is well defined, continuous and |yn (x) − b| ≤ k for x ∈ [a − h, a + h].
Proof of Claim 1: This is clearly true for n = 0, so suppose claim is true for some
n ≥ 0. Then for t ∈ [a − h, a + h] we have that (t, yn (t)) ∈ R so as f is defined and
continuous on R and as yn is continuous we have that t 7→ f (t, yn (t)) is a continuous
function on the interval [a − h, a + h]. Thus yn+1 is well defined and continuous by
properties of integration and by P(i)
Z x
|yn+1 (x) − b| ≤ |f (t, yn (t))|dt
a
Z x
≤ M dt = M |x − a| ≤ M h ≤ k
a

for every x ∈ [a − h, a + h]. Thus the claim is true by induction.

Figure 1.2: successive iterates graphed in R.

We next prove:

9
Claim 2: For |x − a| ≤ h and n ∈ N

Ln−1 M
|en (x)| ≤ |x − a|n (1.11)
n!
where L is such that the Lipschitz condition (1.9) holds.
We remark that this claim in particular implies that

Ln−1 M n
|en (x)| ≤ h for all |x − a| ≤ h. (1.12)
n!
P
which will be the estimate that we will use to show that en converges uniformly using
M-test.
Proof of Claim 2: We recall that the Lipschitz condition P(ii) combined with the fact
that the graph of yn is in the rectangle implies that for all |t − a| ≤ h

|f (t, yn (t)) − f (t, yn−1 (t))| ≤ L|yn (t) − yn−1 (t)| = L|en (t)|. (1.13)

From (1.8) and (1.13) thus

Z x
| en+1 (x)| ≤ |f (t, yn (t)) − f (t, yn−1 (t))|dt
a
Z x (1.14)
≤L |en (t)|dt .
a

Now we prove (1.11) by induction:

Z x
e1 (x) = y1 (x) − b = f (t, b)dt. (1.15)
a

By P(i), f is bounded by M so that

Z x
|e1 (x)| ≤ |f (t, b)|dt ≤ M |x − a|, (1.16)
a

so (1.11) is true for n = 1. Now suppose that (1.11) is true for n, then
Z x
| en+1 (x)| ≤ L |en (t)|dt
a
Z x n−1
L M Ln M
≤ L |t − a|n dt = |x − a|n+1 ,
a n! (n + 1)!

so that (1.11) is true by induction.

We now use these two claims to prove the existence of a solution to the integral equation
(1.4) by showing

10
Claim 3: The iterates yn (x) = nj=0 ej (x) converge uniformly to a continuous function
P
y∞ and y∞ is a solution of the integral equation (1.4).
Proof of Claim 3: The uniform convergence immediately follows from the Weierstrass
Ln−1 M hn
M -test and (1.12), since ∞
P
n=1 Mn for Mn = converges and Mn is a upper
n! Pn
bound on |en (x)| that is independent of x. Thus, yn = 0 ek converges uniformly to a
limit y∞ on [a − h, a + h] and this limit is continuous as it is the uniform limit of the
continuous functions yn .
To see that y∞ is a solution of (1.1), we would like to take the limit in (1.5) and exchange
the limit and the integral to get that
Z x Z x
(∗)
y∞ (x) = lim yn+1 (x) = b + lim f (t, yn (t))dt = b + lim f (t, yn (t))dt
n→∞ n→∞ a a n→∞
Z x (1.17)
=b+ f (t, y∞ (t))dt.
a

The reason that we are allowed to switch limit and integral in (*) is that the integrands
f (t, yn (t)) converge uniformly to f (t, y∞ (t)) since the uniform convergence of the yn and
the Lipschitz condition allow us to estimate
sup |f (t, yn (t)) − f (t, y∞ (t))| ≤ sup L|yn (t) − y∞ (t)| → 0.
t∈[a−h,a+h] t∈[a−h,a+h]

We have thus proven the existence of a solution of the integral equation and as remarked
previously differentiating the integral equation implies that
0
y∞ (x) = f (x, y∞ (x))
and since also y∞ (a) = b, thus y∞ is a solution of the IVP.
This completes the proof of existence and it remains to show
Claim 4: The solution of (1.1) is unique among all functions y : [a − h, a + h] →
[b − k, b + k].
Proof of Claim 4: Let y1 (x) and y2 (x) be two solutions of (1.1) (for the same f , a
and b!) and set e(x) := y2 (x) − y1 (x). We aim to show that e(x) = 0 for all x.
As the IVP is equivalent to the integral equation (1.4) we can subtract the two integral
equations satisfied by y1,2 to see that
Z x
e(x) = y2 (x) − y1 (x) = (f (t, y2 (t)) − f (t, y1 (t)))dt
a

so using the triangle inequality for integrals and the Lipschitz condition we get
Z x Z x
| e(x)| ≤ |f (t, y2 (t)) − f (t, y1 (t))|dt ≤ L |y2 (t) − y1 (t)|dt
a a
Z x (1.18)
≤L |e(t)|dt
a

11
Now e(x) is continuous on [a − h, a + h] therefore, it is bounded say |e(x)| ≤ B so
Z x
| e(x)| ≤ L Bdt = LB|x − a|
a

and inducting on n, using (1.18), for each n

|x − a|n
| e(x)| ≤ BLn
n!
So that for each n
Ln hn
| e(x)| ≤ B → 0 as n → ∞ and e(x) = 0.
n!
Thus the difference is zero, so the solutions are the same which establishes uniqueness
(uniqueness proofs almost always go like this: assume there are two and make their
difference vanish).
This completes the proof of Picard’s Theorem.
Note that the proof of uniqueness of the solution only holds among those solutions whose
graph lies in R. This is however enough since we have:
Remark: Let f be so that P(i) hold. Then the graph of any solution y(x) of (1.1) for
|x − a| ≤ h must lie in R.
Indeed, suppose not. Then, by the continuity of y, there will exist a ‘first’ x0 , with
|x0 − a| < h, where (x0 , y(x0 )) is on the boundary of R. That is such that |x0 − a| < h,
|y(x0 ) − b| = k but |y(x) − b| < k if |x − a| < |x0 − a|, see figure 1.3. But then
Z x0
|y(x0 ) − b| ≤ |f (s, y(s))|ds ≤ M |x0 − a| < M h = k
a

a contradiction. Equivalently we can argue that since |f | is bounded by M we know

that the slope of the graph of any solution of y 0 (x) = F (x, y(x)) is at most M so y(x)
cannot increase from y(a) = b to a value larger than b + M h on an interval of length h.

Figure 1.3: x0 is the first x where the graph meets the boundary of the rectangle

12
Remark: An alternative (and more standard) way of proving uniqueness of solutions
of differential equations is via Gronvall’s inequality and this proof will be carried out in
section 1.5.
Since the warning example doesn’t have a unique solution, something goes wrong for
it. As an exercise, show that the warning example fails the Lipschitz condition (in any
neighbourhood of the initial point).
The following example also fails the Lipschitz condition in any neighbourhood of the
initial point y = 0. However, the Lipschitz condition does hold on any rectangle which
does not contain any point (x, 0).
Example: Consider the IVP

y 0 = x2 y 1/5 , y(0) = b.

So we consider the function f : R2 → R defined by f (x, y) = x2 y 1/5 which is clearly

continuous. (Note that when we write y 1/5 here we mean to take the real root: so that
if y is negative we will take −|y|1/5 .)
Case b = 0: f (x, y) does not satisfy a Lipschitz condition on any rectangle of the form
R0 = {(x, y) : |x| ≤ h, |y| ≤ k}, where h > 0 and k > 0.
Suppose it does, then there exists a finite constant L such that for all |x| ≤ h and
|y|, |ỹ| ≤ k
|x2 ||y 1/5 − ỹ 1/5 | ≤ L|y − ỹ|
so in particular (choosing ỹ = 0 and x = h)

|h2 ||y −4/5 | ≤ L for every y ∈ [−h, h] \ {0}.

But this is a contradiction as |h2 ||y −4/5 | is unbounded as y → 0 so the function does not
satisfy a Lipschitz condition on R0 . So Picard’s theorem does not apply if we take b = 0.
(We saw that f satisfies a Lipschitz condition on any rectangle where its derivative with
respect to y exists and is bounded. The problem here is that the derivative of f is
unbounded as y → 0 – indeed the derivative does not exist at y = 0.)
Case b > 0: However, the assumptions of Picard’s theorem will be satisfied if we take as
initial condition y(0) = b > 0, provided we take a rectangle, Rb , given by Rb = {|x| ≤ h,
|y − b| ≤ k} when 0 < k < b, so that y cannot be zero in this rectangle .
x2 y −4/5
On any such rectangle fy (x, y) = 5 is bounded, so, by (1.10), f satisfies a Lipschitz
condition, and P(ii) is satisfied.
For P(i): f (x, y) is continuous on Rb and

max |x2 y 1/5 | ≤ h2 (b + k)1/5 =: M,

Rb

so Picard’s theorem applies in a rectangle where h > 0 satisfies

h2 (b + k)1/5 h ≤ k.

13
That is
k
h3 ≤ . (1.19)
(b + k)1/5

We can of course solve this problem directly using separation of variables if we wish
giving
 5/4
y = 4x3 /15 + b4/5 ,
so actually the solution exists for all x. Note the solution above is valid for b = 0 BUT
the trivial solution is also valid. So while we still have existence, uniqueness does not
hold.

1.4 Extension of solutions and global existence.

The result in Section 1.5 is a local result in that it guarantees existence and uniqueness
of a solution on the interval [a − h, a + h], where h satisfies M h ≤ k (though this h need
not be the best possible).
As we have seen there are examples of initial value problems where solutions do not
exist for all x ∈ (−∞, ∞). So we would like to find conditions which guarantee that the
solution does exist for all x ∈ (−∞, ∞) (or if f is only defined for x in an certain interval
then on the whole such interval). One such condition is the global Lipschitz condition,
where we can find a constant L such that the Lipschitz condition holds for all y. First we
will see that if the Lipschitz condition is global in y, but L still depends on the interval
[a − h, a + h], then there is existence on all of [a − h, a + h]
Suppose we require that f (x, y) is defined and continuous for all (x, y) ∈ [a−h, a+h]×R
and instead of (P(ii)) we have
(P(iii)): f (x, y) satisfies the Lipschitz condition for all real y and all x ∈ [a − h, a + h].
Then the last condition in claim 1 is not required and hence we don’t need to ask
that M h ≤ k. If we investigate the proof of the Picard existence theorem, we see
that M also appears in (1.16). However for claim 2 to hold it is sufficient to take
M = supx∈[a−h,a+h] |f (x, b)|, which exists as x 7→ f (x, b) is a continuous function on the
closed bounded interval [a − h, a + h] (whereas (x, y) 7→ f (x, y) might be unbounded on
the unbounded set [a − h, a + h] × R) . The rest of the proof applies without change and
we hence obtain that the solution exists and is unique ∀x ∈ [a − h, a + h].
If (P(iii)) holds for each h > 0 (for some L that is allowed to depend on h), then we can
carry out this argument for every h > 0 and, by letting h → ∞ deduce that the solution
exists on all of R. In this case we say that we have a global solution.
Example: If f (x, y) = p(x)y + q(x), where p and q are continuous on |x − a| ≤ h, then
f satisfies (P(iii)).
Remark We do not need the interval in (P(iii)) to be a balanced interval (ie of the form
[a − h, a + h]) because we can deal with x ≤ a and x ≥ a separately. Thus if a ∈ [c, d] and

14
we require that f is continuous on [c, d] × R and that it satisfies the Lipschitz condition
on this set, then a solution of (1.1) exists and is unique for all x ∈ [c, d].

1.5 Gronwall’s inequality and continuous dependence on the initial

data.

We will now prove Gronwall’s inequality which will be used to provide another proof of
uniqueness of solutions, but will also be used to show that solutions depend continuously
on the initial data.
Theorem 1.2. (Gronwall’s inequality) : Suppose A ≥ 0 and b ≥ 0 are constants
and v is a non-negative continuous function satisfying
Z x
v(x) ≤ b + A v(s)ds (1.20)
a

then
v(x) ≤ beA|x−a| .
(The modulus is needed to take care of the case x ≤ a.)

Proof: We use an integrating factor.

Rx
For x ≥ a let V (x) = a v(s)ds, so that V 0 (x) = v(x). As x ≥ a and v ≥ 0 also V (x) ≥ 0
and we have
V 0 (x) ≤ b + AV (x).
Multiply through by the integrating factor e−Ax so

(V 0 (x) − AV (x))e−Ax ≤ be−Ax that is

d
(V (x)e−Ax ) ≤ be−Ax , so, integrating and noting that V (a) = 0
dx Z x
−Ax b
V (x)e ≤ be−As ds = (e−Aa − e−Ax ), so
a A
b A(x−a)
V (x) ≤ (e − 1).
A
Finally, using (1.20)
Z x
b
v(x) ≤ b + A v(s)ds = b + AV (x) ≤ b + A (eA(x−a) − 1) = beA(x−a) ,
a A
as required. Similarly if x ≤ a.
Remark: Gronwall’s inequality says that v is bounded above by the solution of the
integral equation one obtains when there is equality in (1.20). For, if we differentiate
Z x
v(x) = b + A v(s)ds,
a

15
we get
v 0 (x) = Av(x), v(a) = b,
which has solution v(x) = beA(x−a) .
Application: Alternative proof of the uniqueness part of Picard’s theorem
and continuous dependence on initital data
Suppose that y and z are solutions of the ordinary differential equation y 0 (x) = f (x, y(x))
with y(a) = b and z(a) = c, where f satisfies conditions (P(i)) and (P(ii)). We want
to bound the difference between these solutions in terms of the difference |b − c| and
show in particular that if b = c then we must have y(x) = z(x) for all x. This gives an
alternative (and indeed simpler) proof of the uniqueness part of Picard’s theorem.
Setting v(x) = |y(x) − z(x)| (note Gronvall requires v to be non-negative) and using that
Z x
y(x) − z(x) = b − c + (f (s, y(s)) − f (s, z(s))ds
a

we get from the Lipschitz condition that

Z x
v(x) ≤ |b − c| + L v(s)ds .
a

We can thus apply Gronwall’s inequality to get that

|y(x) − z(x)| = v(x) ≤ |b − c|eL|x−a| ≤ |b − c|eLh . (1.21)

In particular, if y and z solve the initial value problem with the same initial value
y(a) = z(a) = b, then the functions y and z are equal, which proves the uniqueness part
of Picard’s theorem.
More generally, we have obtained a bound on |y(x) − z(x)| for all x in the interval where
both solutions are defined in terms of |b − c|.
We say a solution is continuously dependent on the initial data on an interval I if we can
make supx∈I |y(x) − z(x)| as small as we like by taking |b − c| small enough. In other
words the error in the solution will be small provided the error in the initial data is small
enough. To be precise, in this case, solutions are continuously dependent on the initial
data for x ∈ [a − h, a + h] if for all  > 0 there exists δ > 0 such that if y and z are as
above,
|b − c| < δ ⇒ |y(x) − z(x)| ≤ , ∀x ∈ [a − h, a + h].
This is clearly true from (1.21), because given  > 0, we have |y(x) − z(x)| ≤  whenever
|b − c| < e−Lh , so we can take δ = e−Lh .

1.6 Picard’s Theorem via the CMT

We can prove Picard’s theorem in a more efficient way, which is really equivalent to
our previous method, by using the contraction mapping theorem (CMT). This is a very

16
useful method of proving existence and uniqueness of solutions of nonlinear differential
equations and many, many other things besides. The results we need will be discussed in
the course on Metric Spaces and Complex Analysis. We will assume the results proved
there.
Define Ch,k = C([a − h, a + h]; [b − k, b + k]), the space of continuous functions y :
[a − h, a + h] → [b − k, b + k]. As is shown in the Metric Spaces course, for y, z ∈ Ch,k if
we define
d(y, z) := ||y − z||sup := sup |y(x) − z(x)|
x∈[a−h,a+h]

then (Ch,k , d) is a complete metric space (we call || · ||sup the “sup norm”).
Also we say that a map T : Ch,k → Ch,k is a contraction if there exists K < 1 such that

||T (y) − T (z)||sup ≤ K||y − z||sup ,

and then we have the CMT, which says:

Theorem 1.3. (Contraction Mapping Theorem) (Banach) Let X be a complete
metric space and let T : X → X be a contraction. Then there is a unique fixed point
y ∈ X, i.e. a unique y such that T y = y.

To prove Picard’s Theorem via the CMT we will first apply this theorem for X = Cη,k =
C([a − η, a + η]; [b − k, b + k]) for a small enough 0 < η ≤ h that we chose below, which
will give that there exists a unique solution for |x − a| ≤ η. In a second step we will then
discuss how this solution can be extended to all of [a − h, a + h] if M h ≤ k by repeating
the argument with a new choice of the space X.
We again consider the IVP (1.1)
Theorem 1.4. (Picard’s existence theorem.) Let f : R → R be a function defined
on the rectangle R := {(x, y) : |x − a| ≤ h, |y − b| ≤ k} which satisfies conditions P(i)(a)
and P(ii) and let η > 0 be so that Lη < 1 and M η ≤ k.
Then the initial value problem (1.1) has a unique solution for x ∈ [a − η, a + η].

Proof.
The strategy is to express (1.1) as a fixed point problem and use the CMT.
As before, we can write the initial value problem as an integral equation
Z x
y(x) = b + f (s, y(s))ds (1.22)
a

Provided f (s, y(s)) is continuous in s, y is a solution of the differential equation if and

only if y is a solution of the integral equation.
If we define Z x
(T y)(x) = b + f (s, y(s))ds
a

17
then we can write (1.22) as a fixed point problem

y = T y.

We will work in the complete metric space Cη,k = C([a − η, a + η]; [b − k, b + k]), where
we will choose η ≤ h so that T : Cη,k → Cη,k and so that T is a contraction. We begin
by proving Claim 1: If η > 0 is so that M η ≤ k then T : Cη,k → Cη,k

Proof. First we note that from the properties of integration, (T y)(x) ∈ C([a−η, a+η]; R).
All that we require is thus to show that ||T y − b||sup ≤ k if ||y − b||sup ≤ k.
But
Z x
||T y − b||sup = sup f (s, y(s))ds (1.23)
x∈[a−η,a+η] a
Z x
≤ sup |f (s, y(s))|ds (1.24)
x∈[a−η,a+η] a
≤ M η ≤ k, (1.25)

provided M η ≤ k.

Claim 2: If Lη < 1 then T is a contraction (with K = Lη):

Proof. Given y, z ∈ Cη,k we can bound

Z x
||T y − T z||sup = sup f (s, y(s)) − f (s, z(s))ds
x∈[a−η,a+η] a
Z x
≤ sup |f (s, y(s)) − f (s, z(s))|ds
x∈[a−η,a+η] a
Z x
≤ sup L|y(s) − z(s)|ds ≤ Lη||y − z||sup ≤ K||y − z||sup
x∈[a−η,a+η] a

where K := ηL < 1 provided η < 1/L.

If we hence choose η < min{h, k/M, 1/L} then T satisfies the conditions of the CMT
and has a unique fixed point, y(x). As explained before, a (continuous) function y solves
the integral equation T y = y if and only if it is continuously differentiable and a solution
of the initial value problem, so we have established that the initial value problem has a
unique solution on the interval [a − η, a + η].
Note that our proof using CMT produces a more restricted range of x values than did
our proof on one dimension. The range of η depends on L as well as M and k. However,
if M h ≤ k, actually we only need η ≤ h, and we can now extend the range of the solution
to all x ∈ [a − h, a + h], by iteration.

18
Corollary 1.5. If M h ≤ k, then the initial value problem has a unique solution on the
whole interval [a − h, a + h]

As uniqueness follows from Gronvall’s Lemma as explained in Section 1.5 we only have
to prove that the solution exists on the whole interval [a − h, a + h]. There are two
very different ways of proving this, one by iterating the above argument to construct a
solution on larger and larger intervals, and one by arguing by contradiction.
Proof: We look at x ≥ a first. If h < 1/L we are done. (Take η = h.)
Otherwise we choose η1 < 1/L. Then, from Theorem 1.4, there exists a unique solution,
y0 say, on [a, a + η1 ] .
Now choose η2 = min{2η1 , h}, and look for a solution, y1 say, on [a + η1 , a + η2 ], of the
ODE with initial data y1 (a + η1 ) = y0 (a + η1 ).
Now define
y(x) = y0 (x), x ∈ [a, a + η1 ]
y(x) = y1 (x), x ∈ [a + η1 , a + η2 ]

To construct y1 : As in Theorem 1.4, but we now work in the space X1 := C([a + η1 , a +

η2 ]; [b − k, b + k]), and take (for a + η1 ≤ x ≤ a + η2 )
Z x
(T1 y)(x) = y0 (a + η1 ) + f (s, y(s))ds
a+η1
Z a+η1 Z x
= b+ f (s, y0 (s))ds + f (s, y(s))ds. (1.26)
a a+η1

So T1 : X1 → X1 because from (1.26)

||T1 y − b||sup ≤ M η1 + M (x − (a + η1 ) = M (x − a) ≤ M η2

≤ M h ≤ k.
Also T1 is a contraction as the proof of claim 2 only requires that the length of the
interval we work on, which for T1 is η2 − η1 , is less than 1/L. Thus we obtain the
existence of a unique solution on [a, a + η2 ]. Repeating this argument, both in positive
and negative direction, we continue to be able to extend the solution and after finitely
many steps have reached the endpoint a + h of the original interval, since we can carry
out each step except the very last one (where we will be able to choose ηj = h since we’ll
have h − ηj−1 < L1 ) with the same ’stepsize’ ηk − ηk−1 = η1 .
Alternative proof of Corollary 1.5 (via contradiction):
As we know that the solution y(x) exists at least on [a − η, a + η] we can consider the
maximal subset I of [a − h, a + h] on which the solution y(x) of (1.1) exists. We want
to show that I is indeed all of [a − h, a + h]. To see this we first note that I must be a
closed interval, for if y(x) solves (1.1) on some (c, d) ⊂ [a − h, a + h] then we can extend
y to the closed interval simply by setting y(d) := limx%d y(x) and y(c) := limx&c y(x).

19
Note that this is possible as M h ≤ k ensures that |y(x) − b| ≤ k and since we have a
uniform bound on |y 0 (x)| ≤ M and hence uniform continuity of y.
So I = [c, d] and to show the claim we need to exclude the possibility that c > a − h
or d < a + h. So suppose that d < a + h. Then we know that b̃ = y(d) satisfies
|b − b̃| ≤ M (d − a) < k. Choosing k̃ = k − M (d − a) and h̃ ∈ (0, a + h − d) small enough
so that M h̃ ≤ k̃ we can apply Theorem 1.4 to get the existence of a solution ỹ(x) on an
interval [d − η̃, d + η̃] for some η̃ > 0 of our ODE , now with initial value ỹ(d) = y(d).
But then we can use ỹ to extend y to a larger interval which gives a contradiction.
Global Existence: If f is continuous for all x ∈ [a − h, a + h], and all y and satisfies a
global Lipschitz condition (i.e. condition P(iii) on [a − h, a + h] × R), then we instead
work in the spaces Ch = C([a − h, a + h]; R), respectively Cη = C([a − η, a + η]; R). As
before, claim 1 then no longer requires the condition M h ≤ k and we obtain in a first
step that a solution exists on [a − η, a + η] for η < L1 . We can then carry out either or
the two arguments above to see that this solution indeed exists on all of [a − h, a + h].
Comparison of the two methods of the proof of Picard:
(1)The proof using the CMT is shorter and simpler than the direct proof because much
of the work has been done in proving the CMT and once we have chosen a suitable
space, we have only to check that the conditions apply. Also, the CMT automatically
gives uniqueness of solutions, which has to be proved separately in the direct method.
Furthermore CMT can be used in more general situations to prove existence of solutions.
2) By Theorem 1.4.3 of the Prelims Analysis II lecture notes, a sequence of continuous
functions yn converges in the sup norm if and only if it is uniformly convergent. Thus
convergence in the sup norm is equivalent to uniform convergence. Furthermore in the
CMT the fixed point is given by the limit in Ch,k of yn = T yn−1 , with y0 any point in
the space. So if we take y0 (x) = b (for all x) the fixed point is given by the uniform limit
of the successive approximations as in the direct proof.
(3) One feature of the proof using the CMT was that it produced a less delicate result, in
that the range of x for which it applied is more restricted. (Though it was easy to extend
the range using iteration or a contradiction arguement.) This sometimes happens when
we use abstract results rather than direct computations, because the direct computations
can be more delicate. We can see why this happens in this case if we investigate the
direct proof. In the direct proof we are working pointwise and each time we apply the
inductive step using (1.14) we integrate (x − a)n , and thus end up dividing by n!, so we
have a series which converges for all x ∈ [a − h, a + h]. But in the CMT we are working
in C([a − h, a + h]) , so on each integration we take the supremum which, of course, does
not depend on x, and thus we integrate a constant, so the n! is absent.
Remark: There are many other fixed point theorems, and other abstract results, which
can be used to prove existence of solutions of more general equations involving derivatives
(including partial derivatives) and integrals. These powerful theorems generally require
some general theory of Banach and Hilbert spaces (see the B4 courses) and a knowledge

20
of suitable spaces (eg Sobolev spaces) in which to apply them (see the part C courses
on functional analysic methods for PDEs and fixed point methods for non-linear PDEs).
The above proof can in particular be adjusted to prove the existence of solutions of
partial differential equations, such as heat equations with non-linearities.

1.7 Picard’s Theorem for systems and higher order ODEs via the CMT

We now want to look at existence and uniqueness of solutions of systems of ODEs. As

well as being of interest in itself, this will be useful in particular for proving the existence
of solutions of equations with higher order derivatives. We consider a pair of first order
ODEs, for the functions y1 and y2 .
y10 (x) = f1 (x, y1 (x), y2 (x)) (1.27)
y20 (x) = f2 (x, y1 (x), y2 (x)) (1.28)
with initial data y1 (a) = b1 , y2 (a) = b2 . (1.29)

We can introduce vector notation

     
y1 f1 b
y= , f= , b= 1 ;
y2 f2 b2

So we can write equations (1.27)–(1.29) as

y 0 (x) = f (x, y(x)), (1.30)
y(a) = b, (1.31)

Now we want to prove Picard’s Theorem for such systems of ODEs. Our previous proof
using the CMT will extend in a very natural way.
We need a ‘distance’ in R2 . In the Metric Spaces course the various norms l1 , l2 (the
Euclidean distance) and l∞ on Rn were defined. We could use any of these (or any
other norm on Rn ), but we will make the fairly arbitrary choice to use the l1 norm,
||y||1 = |y1 | + |y2 |. In place of the rectangle R we will use the subset S = {(x, y) ∈
R3 : |x − a| ≤ h, y ∈ Bk (b)}, where Bk (b) is the closed disc in R2 centred on b,
radius k with respect to the l1 norm. That is Bk (b) = {y ∈ R2 : ||y − b||1 ≤ k} (ie
Bk (b) = {(y1 , y2 ) ∈ R2 : |y1 − b1 | + |y2 − b2 | ≤ k}.
We will suppose that
(H(i)) f1 (x, y1 , y2 ) and f2 (x, y1 , y2 ) are continuous on S, and are hence bounded (be-
cause f1 and f2 are continuous functions on the closed bounded set S), say |f1 (x, y)| +
|f2 (x, y)| ≤ M on S.
(H(ii)) f1 (x, y1 , y2 ) and f2 (x, y1 , y2 ) are Lipschitz continuous with respect to (y1 , y2 ) on
S. That is, there exist L1 and L2 such that for x ∈ [a − h, a + h] and u, v ∈ Bk (b),
|f1 (x, u1 , u2 ) − f1 (x, v1 , v2 )| ≤ L1 (|u1 − v1 | + |u2 − v2 |) and
|f2 (x, u1 , u2 ) − f2 (x, v1 , v2 )| ≤ L2 (|u1 − v1 | + |u2 − v2 |).

21
It is easy to see that these conditions are equivalent to the following:
(H(i))0 f (x, y) is continuous on S, and bounded by M , say (that is ||f (x, y)||1 ≤ M ).
[M must exist because f is a continuous function on the closed bounded set S.]
(H(ii))0 f (x, y) is Lipschitz continuous with respect to y on S. That is, there exists L
such that for x ∈ [a − h, a + h] and u, v ∈ Bk (b),

||f (x, u) − f (x, v)||1 ≤ L||u − v||1 .

Note that we can take L = L1 + L2 .

We now get the following version of Picard’s existence theorem
Theorem 1.6. (Picard’s existence theorem for systems.) Let f1 , f2 : S → R be
functions for which the conditions (H(i)) and (H(ii)) (or (H(i))0 and (H(ii))0 ) hold
true for the set S = [a − h, a + h] × Bk (b) ⊂ R3 . Then there exists 0 < η ≤ h, such that
the initial value problem (1.31) has a unique solution for x ∈ [a − η, a + η]

Our previous proof using the CMT will extend to this case if we work in the complete
metric space Cη := C([a − η, a + η]; Bk (b)), the space of continuous functions mapping
from [a − η, a + η] to Bk (b) with norm (or distance) on Cη defined by
!
||y||sup = sup ||y(x)||1 := sup (|y1 (x)| + |y2 (x)|).
x∈[a−η,a+η] x∈[a−η,a+η]

As before, we can write the initial value problem as an integral equation

Z x
y(x) = b + f (s, y(s))ds
a

where by the integral we mean that we integrate componentwise. Provided f (s, y(s)) is
continuous in s, y is a solution of the differential equation if and only if y is a solution
of the integral equation.
If we define Z x
(T y)(x) = b + f (s, y(s))ds
a
then we can write this as a fixed point problem

y = T y.

As before we can now work in the complete metric space Cη , to show that, provided
we choose η < min{h, k/M, 1/L}, then T : Cη → Cη and is a contraction (see problem
sheet).
Again we can extend the range of the solution to all x ∈ [a − h, a + h], by iteration.
Corollary 1.7. If M h ≤ k then there is a unique solution for all x ∈ [a − h, a + h]

22
Again if the functions are globally Lipschitz with respect to (y1 , y2 ), then the solution is
global.
This all extends easily to systems of n equations.

Picard for Higher Order ODEs

With Picard extended to first-order systems, it is a small step to extend it to a single,

higher order ODE. For simplicity, we consider just an IVP for linear second-order ODEs
(which will be considered in more detail in DEs2):

y 00 + p(x)y 0 + q(x)y = r(x)

with initial data

y(a) = b y 0 (a) = c,
and p, q, r continuous for |x − a| ≤ h.
To reduce this to a first-order system, introduce z = y 0 and write

y 0 = z := f1 (x, y, z)

z 0 = −pz − qy + r := f2 (x, y, z)
with data y(a) = b, z(a) = c. This is precisely in the form to which the previous
section applies, and it’s easy to check that the global Lipschitz condition is satisfied, so
we get:

Theorem 1.8. ( Picard for second-order linear ODEs)

With the assumptions as above, the solution exists for |x − a| ≤ h, and is unique.

Clearly this method can be extended to the IVP for an n-th order linear ODE. In
particular, this justifies our belief that an n-th order ODE needs n pieces of data to fix
a unique solution.

1.8 Summary

So we have looked at existence and uniqueness of solutions of various initial value prob-
lems. We found that Lipschitz continuity gives existence and uniqueness and that
uniqueness can fail without the Lipschitz continuity. Even with Lipschitz continuity,
the existence is often local, though a global Lipschitz condition on an interval containing
the initial point will give existence and uniqueness of solutions on that interval.
There were two different methods of proof of existence and uniqueness. First we did a
direct proof using successive approximations. Then we saw that using the CMT simplifies
the proof, because the hard work has already been done in the proof of the CMT and

23
proving completeness of C. This proof readily extends to treat systems of ODEs (though
here again we could have used successive approximation). A disadvantage of using the
CMT was that it gave the result only for a restricted range of x, though it was easy to
extend using iteration. This can be a feature of proofs using abstract results, because
sometimes some of the detail is lost in the abstraction.
We also derived Gronwall’s inequality, and used it to show that solutions depend con-
tinuously on the initial data. It also gave another proof of uniqueness.

24
2 Plane autonomous systems of ODEs

The definition: a plane autonomous system of ODEs is a pair of ODEs of the form;
dx
= X(x, y) (2.1)
dt
dy
= Y (x, y)
dt

Here “autonomous” means there is no t-dependence in X or Y , and “plane” means there

are just two equations, so we can draw pictures in the (x, y) - plane, which will then be
called the phase plane.
Given initial values x(0) = a, y(0) = b, expect that there exists a unique solution and this
solution which will define a trajectory or phase path in the phase plane. It is convenient,
though not necessary, to think of t as time, and the trajectory as the curve in the plane
(including orientation) that is traced out by a moving particle. We put an arrow on the
trajectory giving the direction of increasing t. We will denote ẋ = dx
dt etc. We will assume
throughout that X and Y Lipschitz continuous in x and y (on every bounded subset of
R2 ) as this will allow us to apply Picard’s theorem to obtain important properties of
solutions for these plane autonomous system and of the corresponding trajectories.
Important observations

• If (x(t), y(t)) is a solution of (2.1) then for any fixed number t0 ∈ R also

x̃(t) := x(t + t0 ), ỹ(t) := y(t + t0 )

solve (1.5) and they trace out the same trajectories.

• Through every point (x0 , y0 ) there exists a UNIQUE trajectory. In particular,

different trajectories can NEVER intersect, though they might asymptote to the
same point (x∗ , y ∗ ) as t → ∞ or as t → −∞ (and any such point must be a critical
point, see below)

The first point immediately follows when we insert (x̃(t), ỹ(t)) into the equations as this
gives
˙
x̃(t) = ẋ(t + t0 ) = X(x(t + t0 ), y(t + t0 )) = X(x̃(t), ỹ(t))
and
˙ = Y (x̃(t), ỹ(t)).
ỹ(t)
Note that this does not work if the system is not autonomous (i.e. if either X or Y also
depend on t).
The second point is an important consequence of Picard’s theorem (and holds true as
we assume X, Y Lipschitz): First of all Picard guarantees the existence of a solution

25
(x(t), y(t)) with x(0) = x0 and y(0) = y0 and hence there is a trajectory through
the point. If (x̃(t), ỹ(t)) is any other solution that traces out a trajectory through
(x0 , y0 ) then there is a t0 so that (x̃(t0 ), ỹ(t0 )) = (x0 , y0 ). Looking at u(t) := x̃(t − t0 ),
v(t) := y(t − t0 ) we get a new solution of (2.1) which has the same initial values as
the original (x(t), y(t)), namely (u(0), v(0)) = (x0 , y0 ) = (x(0), y(0)). By the uniqueness
part of Picard we thus know that these two solutions (u(t), v(t)) and (x(t), y(t)) must
be the same, so (x̃, ỹ) is nothing else than a time-shift of the original solution so must
trace out the same trajectory.

2.1 Critical points and closed trajectories

A critical point is a point (x0 , y0 ) in the phase plane where X(x0 , y0 ) = Y (x0 , y0 ) = 0.
So a critical point is a particular (very special) trajectory corresponding to solutions
(x(t), y(t)) of (2.1) that are constant in time.
There may be trajectories in the phase plane which are closed i.e. which return to the
same point. Provided they don’t just correspond to constant solutions and so are simply
given by a single point, these correspond to periodic solutions of (2.1) as may be seen
as follows:
Suppose the trajectory is closed so that for some finite value t0 of t, (x(t0 ), y(t0 )) =
(x(0), y(0)), while (x(t), y(t)) 6= (x(0), y(0)) for 0 < t < t0 . Define x̄(t) = x(t + t0 ),
ȳ(t) = y(t + t0 ). Then as before we see that (x̄(t), ȳ(t)) is another solution of (2.1)
with x̄(0) = x(t0 ) = x(0); ȳ(0) = y(t0 ) = y(0). Now by uniqueness of solution (given
Lipschitz again).
x(t + t0 ) = x̄(t) = x(t)
y(t + t0 ) = ȳ(t) = y(t),
but this is now true for all t, so a closed trajectory corresponds to a periodic solution
of (2.1) with period t0 . The converse is trivial.
Note in particular that this means that a trajectory cannot intersect itself, but might
close up to a closed curve (without self-intersections).

2.1.1 An example

Consider the harmonic oscillator equation

ẍ = −ω 2 x.

Turn this into a plane autonomous system by introducing y as follows:


ẋ = y = X(x, y)
(2.2)
so ẏ = −ω 2 x = Y (x, y).

26
(Clearly this trick often works for second-order ODEs arising from Newton’s equations.)
The only critical point is (0, 0), but note that
d 2 2
(ω x + y 2 ) = 2ω 2 xẋ + 2y ẏ = 0
dt
so ω 2 x2 + y 2 = constant. (which, from Prelims dynamics, we know to be proportional
to the total energy). For a given value of the constant this is the equation of an ellipse,
so we can draw all the trajectories in the phase plane as a set of nested (concentric)
ellipses:

Figure 2.1: The phase diagram for the harmonic oscillator; to put the arrows on the trajectories,
notice that ẋ > 0 if y > 0.

The picture in the phase plane is called the phase diagram (or phase portrait) and from
that we see that all trajectories are closed, so all solutions are periodic (as we already
know, from Prelims).

2.2 Stability and linearisation

We want to learn how to sketch the trajectories in the phase plane in general and to do
this we first consider their stability. Intuitively we say a critical point (a, b) is stable if
near (a, b) the trajectories have all their points close to (a, b) for all t greater than some
t0 . We make the formal definition:
Definition A critical point (a, b) is stable if given p
 > 0 there exists δ > 0 and t0 such
that for any solution (x(t), y(t)) of (2.1) for which (x(t0 ) − a)2 + (y(t0 ) − b)2 < δ
p
(x(t) − a)2 + (y(t) − b)2 < , ∀t > t0 .

A critical point is unstable if it is not stable.

(Here we have used the Euclidean distance. We could use other norms such as l1 or l∞ )

27
A common way to analyse the stability of a critical point is to linearise about the point
and assume that the stability is the same as for the linearised equation. There are
rigorous ways of showing when this is true. We will assume it is valid, pointing out the
cases where it is likely fail. Linearising will also enable us to classify the critical points
according to what the trajectories look like near the critical point .
So suppose P = (a, b) is a critical point for (2.1), so

X(a, b) = 0 = Y (a, b). (2.3)

Now x = a, y = b is a solution of (2.1). We linearise by setting

x = a + ζ(t); y = b + η(t)

where ζ and η are thought of as small. From (2.1), and Taylor’s theorem

ẋ = ζ̇ = X(a + ζ, b + η) = X(a, b) + ζXx |p + ηXy |p + h.o.

ẏ = η̇ = Y (a, b) + ζYx |p + ηYy |p + h.o.

where ‘h.o.’ means quadratic and higher order terms in ζ and η. Now use (2.3) and
neglect higher order terms to find
 .    
ζ̇ A B ζ
=


η̇ C D η

    (2.4)
A B Xx |p Xy |p 
= 
C D Yx |p Yy |p

 
ζ
Call this (constant) matrix M and set Z(t) = then (2.4) becomes
η

Ż = M Z. (2.5)
We can solve (2.5) with eigen-vectors and eigen-values as follows: Z 0 eλt is a solution,
with constant vector Z 0 and constant scalar λ if

λZ 0 = M Z 0 ,

i.e. Z 0 is an eigen-vector of M with eigen-value λ. We are considering just 2×2-matrices,

with eigen-values say λ1 and λ2 so the general solution if λ1 6= λ2 is

Z(t) = c1 Z 1 eλ1 t + c2 Z 2 eλ2 t , (2.6)

for constants ci . Recall λ1 , λ2 may be real, in which case the ci and the Z i are real, or
a complex conjugate pair, in which case the ci and the Z i are too.
If λ1 = λ2 = λ ∈ R say, we need to take more care. The Cayley-Hamilton Theorem (see
Algebra I) implies that (M − λI)2 = 0 since the characteristic polynomial is cM (x) =
(x − λ)2 , so either M − λI = 0 or M − λI 6= 0. We have a dichotomy:

28
 (i) if M − λI = 0 then M = λI and the solution is

Z(t) = Ceλt (2.7)

for any constant vector C.

(ii) if M − λI 6= 0 then there exists a constant vector Z 1 with

Z 0 := (M − λI)Z 1 6= 0

but
(M − λI)Z 0 = (M − λI)2 Z 1 = 0.
(So Z 0 is the one linearly independent eigenvector of M .) One now checks that
the solution of (2.5) is
(c1 Z 1 + (c0 + c1 t)Z 0 )eλt . (2.8)

Now we can use (2.6) and (2.8) to classify critical points.

2.3 Classification of critical points

We shall assume that neither eigenvalue of the matrix M is zero, which is the requirement
that the critical point be non-degenerate. A proper discussion of this point would take
us outside the course but roughly speaking if a critical point is degenerate then we need
to keep more terms in the Taylor expansion leading to (2.4), and the problem is much
harder.
Case 1. 0 < λ1 < λ2 (both real of course)
From (2.6), as t → −∞, Z(t) → 0, and Z(t) ∼ c1 Z 1 eλ1 t unless c1 = 0 in which case
Z(t) ∼ c2 Z 2 eλ2 t , while as t → +∞, Z(t) ∼ a large multiple of Z 2 , unless c2 = 0 when
Z(t) ∼ a large multiple of Z 1

Z2

Z1

Figure 2.2: An unstable node.

29
These trajectories converge on the critical point into the past, but go off to infinity in
the future. A critical point with these properties is called an unstable node.

Case 2: λ1 < λ2 < 0 (both real)

This is as above but with t → −t and the roles of Z 1 , Z 2 switched. The trajectories
converge on the critical point into the future and come in from infinity in the past.

Z2

Z1

Figure 2.3: A stable node.

This is a stable node.

Case 3: λ1 = λ2 = λ. If the solution of the linearised equation is given by (2.7) (case

(i)) we have a star , while if the solution is given by (2.8) (case (ii)) there is an inflected
node . In both cases the critical point is stable if λ < 0 and unstable if λ > 0.

30
 Figure 2.4: Unstable star case (i) and unstable inflected node case (ii)

Case 4: λ1 < 0 < λ2 (both real)

If c1 = 0 then Z(t) → ∞ along Z2 as t → ∞

→0 along Z2 as t → −∞.
If c2 = 0 then Z(t) → 0 along Z1 as t → ∞
→∞ along Z1 as t → −∞.

If c1 , c2 6= 0 then Z(t) → ∞ along Z 2 as t → ∞ and along Z 1 as t → −∞.

Most trajectories come in approximately parallel to ±Z 1 and go out becoming asymptotic
to ±Z 2 .

Figure 2.5: A saddle.

This is a saddle (to motivate the name, think of the trajectories as contour lines on a
map; then two opposite directions from the critical point are uphill and the two orthog-
onal directions are downhill).

If the eigen-values are a complex conjugate pair we may write

λ1 = µ − iν, λ2 = µ + iν µ, ν ∈ R,

and the classification continues in terms of µ and ν.

In (2.6) the ci Zi are a conjugate pair so if we put c1 = reiθ , Z 1 = (1, keiφ )T , then

reiθ
 
c1 Z 1 =
rkei(φ+θ)
so that  
µt 2r cos(νt − θ)
Z(t) = e .
2rk cos(νt − (φ + θ))

31
Case 5: µ = 0
Then Z(t) is periodic.

Figure 2.6: An anticlockwise centre (B < 0); X = −x − 3y, Y = x + y

This case is called a centre, and is stable. The sense of the trajectories, clockwise or
anticlockwise, depends on the sign of B; B > 0 is clockwise (take ζ = 0 and η > 0, then
ζ̇ = Bη > 0).
To see that this centre is stable: Take t0 = 0. Consider the path whose maxmum distance
from the critical point, (a, b), is  > 0. Let δ > 0 be the minimum distance of this path
from
p (a, b). Then p
(x(0) − a)2 + (y(0) − b)2 ≤ δ implies (x(t) − a)2 + (y(t) − b)2 ≤ , for all t ≥ 0.

Case 6: µ 6= 0
This is just like case 5, but with the extra factor eµt , which is monotonic in time. We
have another dichotomy:

(i) µ > 0 then |Z(t)| → ∞ as t → ∞ so the trajectory spirals out, into the future.
This is called an unstable spiral.

(ii) µ < 0 this is the previous with time reversed so it spirals in, and is called a stable
spiral.

In case 6, as in case 5, the sense of the spiral is dictated by the sign of B.

[ An alternative method of looking at case 5 and 6:
Case 5: µ = 0
so λ1 = −iν and λ2i = −ν 2 < 0; but, as both the trace and determinant of a matrix are
invariant under P −1 M P transformations, in terms of the matrix M of (2.4), trace M =
A + D = λ1 + λ2 = iν − iν = 0 so det M = AD − BC = −A2 − BC = λ1 λ2 = ν 2 > 0

32
Figure 2.7: A antilockwise unstable spiral; X=-y, Y=x+y. Reverse the arrows for a stable spiral .

Equation (2.4) becomes

    
ζ̇ A B ζ
= . (2.9)
η̇ C −A η

As an exercise, show that now −Cζ 2 + 2Aζη + Bη 2 is constant in time. We know that
B, C have opposite signs with (−BC) > A2 so this is the equation of an ellipse.
This case is called a centre.

Case 6: µ 6= 0
So, in (2.6), we must have Z 1 = Z̄ 2 and c1 = c̄2 and

Z(t) = eµt c1 Z 1 e−iνt + c̄1 Z̄ 1 eiνt ,

 

which is just like case 5, but with the extra factor eµt , which is monotonic in time. So:

(i) µ > 0 then |Z(t)| → ∞ as t → ∞ so the trajectory spirals out, into the future. An
unstable spiral.

(ii) µ < 0 this is the previous with time reversed so it spirals in, a stable spiral.]

33
Important observation:
Both the trace and determinant of a matrix are invariant under P −1 M P transfor-
mations, so in terms of the matrix M of (2.4), trace M = A + D = λ1 + λ2 and
det M = AD − BC = λ1 λ2
Thus: if A + D > 0 then we have one of the cases 1, 4 or 6(i), all of which are unstable
(but if A + D < 0 the critical point can be stable or unstable). Further det M = λ1 λ2 .
So when the eigenvalues are real the sign of det M tells us whether the signs of the
eigenvalues are the same or different. The determinant is always positive in the case of
complex eigenvalues.
Relationship to non-linear problem: One hopes that the linearistion will have the
same type of critical point as the original system. In general if the linearisation has a
node, saddle point or spiral, then so does the original system, but proving this is beyond
the scope of this course. However, a centre in the linearisation does not imply a centre
in the nonlinear system. This is not surprising when one reflects that a centre in the
linear system arises when Re λ = 0 so the perturbation involved when one returns to the
nonlinear system, however small, can change this property.
Analysing the critical points and their local behaviour is important in determining the
general behaviour of trajectories of an ODE system. Connecting the various critical
points together requires care. It helps to remember that trajectories can never intersect
and that while different trajectories can asymptote to the same point (x0 , y0 ) this can
only be the case if (x0 , y0 ) is a critical point. Also that the signs of X(x0 , y0 ) and
Y (x0 , y0 ) give the signs of dx/dt(t) and dy/dt(t) respectively of solutions of (2.1) at the
time where they pass through this point.
We note that a trajectory can only become horizontal in a point (x0 , y0 ) if Y (x0 , x0 ) = 0
as this means that the corresponding solution of (2.1) has velocity dy/dt = 0 in the
moment where it passes through that point.
Similarly, the only points (x0 , y0 ) in the plane where trajectories can become vertical are
points where X(x0 , y0 ) = 0.
To draw a phase diagram it hence helps to draw the ”nullclines”, which are the curves
in the plane on which X(x, y) = 0 respectively Y (x, y) = 0.
Such nullclines obviously cross at critical points. To find the nullclines sketch the curves
X(x, y) = 0 and the curves Y (x, y) = 0. In particular in any region bounded by null-
clines the trajectories must have a single sign for dx/dt and for dy/dt. Hence a simple
examination of the expressions for X and Y in any region will determine if all the arrows
in that region are “up and to the left”, “up and to the right”, “down and to the left” or
“down and to the right”.

34
2.3.1 An example

Find and classify the critical points for the system

ẋ = x − y = X(x, y) (2.10)
ẏ = 1 − xy = Y (x, y)

Solution: for the critical points, from X = 0 deduce x = y, therefore from Y = 0

deduce x2 = 1, and we have either (1, 1) or (−1, −1).
For the classification, calculate
   
Xx Xy 1 −1
M= = ,
Yx Yy −y −x

and evaluate at the critical points:

√
 
1 −1
at (1, 1) : M = : λ2 − 2 = 0 : λ = ± 2
−1 −1

this is a saddle. The corresponding eigenvectors are:

√
 
1√
λ1 = − 2 Z 1 = direction in
1+ 2
√
 
1√
λ2 = 2 Z 2 = direction out
1− 2

 
1 −1
at (−1, 1) : M = : λ2 − 2λ + 2 = 0 : λ = 1 ± i.
1 1

this is an unstable spiral; B < 0, so its described anticlockwise.

35
 Figure 2.8: The phase diagram of (2.10)

Figure 2.9: The phase plane diagram of (2.10) showing the nullclines y = x and xy = 1.

36
2.3.2 Further example: the damped pendulum

Another example from mechanics: a simple plane pendulum with a damping force pro-
portional to the angular velocity. We shall use the analysis of plane autonomous systems
to understand the motion.
Take θ to be the angle with the downward vertical, then Newton’s equation is
mlθ̈ = −mg sin θ − mklθ̇,
where m is the mass of the bob, l is the length of the string, g is the acceleration due
to gravity and k is a (real, positive) constant determining the friction. We cast this as
a plane autonomous system in the usual way: set x = θ and y = ẋ = θ̇ so
ẋ = y
g
ẏ = − sin x − ky
l
4g
For simplicity below, we’ll also assume that k 2 < l , so that the damping isn’t too large.
To sketch the phase diagram, we first find and classify the critical points. The critical
points satisfy y = 0 = sin x, so are located at (x, y) = (N π, 0). Then
 
0 1
M=
− gl cos x −k
The classification depends on whether N is even or odd:
 
0 1
for x = 2nπ M =
− gl −k
which gives a stable spiral (clockwise);
 
0 1
for x = (2n + 1)π M= g
l −k
which gives a saddle.
We now have enough information to sketch the phase diagram (note that ẋ is positive
or negative according as y is).

37
 Figure 2.10: The phase diagram of the damped pendulum

2.3.3 An important example: The Lotka–Volterra predator-prey equations

This is a simplified mathematical model of a predator-prey system. Think of variables x

standing for the population of prey, and y for the population of predators, both functions
of t for time. As time passes, x increases as the prey breed, but decreases as the predators
predate; likewise y increases by predation but decreases if too many predators compete.
We assume that x and y are governed by the following plane autonomous system:

ẋ = αx − γxy (2.11)
ẏ = −βy + δxy,

where α, β, γ, δ are positive real constants. Because of the interpretation as populations,

we only care about x ≥ 0, y ≥ 0 but we shall consider the whole plane for simplicity.
Again, the aim is to use the analysis of plane autonomous systems to lead us to the
phase diagram and an understanding of the dynamics.
For the critical points first, set

X := x(α − γy) = 0

Y := y(−β + δx) = 0.
There are two solutions, (0, 0) and ( βδ , α
γ ). For the matrix:
   
Xx Xy α − γy −γx
M= =
Yx Yy δy −β + δx

so first  
α 0
at (0, 0) : M =
0 −β
which gives a saddle, where, it is easy to see, the out-direction is the x-axis and the
in-direction is the y-axis. Next
!
− βγ
 
β α 0
at , : M= αδ
δ : λ2 + αβ = 0
δ γ γ 0

which gives a centre, described anticlockwise since B < 0.

We have found and classified the critical points. Before sketching the phase diagram, it
is worth noting, from (2.11), that the axes are particular trajectories, and trajectories
can only cross at critical points (as noted before).

38
 Figure 2.11: The phase diagram for the Lotka–Volterra system

Therefore any trajectory which is ever in the first quadrant is confined to the first
quadrant, and no trajectory can enter the first quadrant from outside. Since there is a
centre in the first quadrant, it looks as though all trajectories in the first quadrant may
be periodic. This is true, and can be seen by the following argument: form the ratio
ẏ y(−β + αx) dy
= = .
ẋ x(α − γy) dx
and separate
(α − γy) (−β + δx)
dy − dx = 0;
y x
now integrate
β log x − δx + α log y − γy = C. (2.12)
for a constant C. For different values of C, (2.12) is the equation of the trajectory or
equivalently the trajectories are the level sets or contours of the function on the left in
(2.12). This function is of the form h(x) + k(y) and can easily be seen to have single
maximum (at ( βδ , αγ )) . Also it tends to minus infinity on the axes and at infinity.
Therefore its contours are all closed curves and so all the trajectories are closed and all
the solutions of (2.11) are periodic.
This useful technique can be applied to other examples.

2.3.4 Another example from population dynamics.

This is a simple model for two species in competition. Suppose that, when suitably
scaled, the population on an island of rabbits (x ≥ 0) and sheep (y ≥ 0) satisfies the

39
plane autonomous system:

ẋ = x(3 − x − 2y), ẏ = y(2 − x − y). (2.13)

(The populations are in competition for resources so each has a negative effect on the
other)
If we analyse this system we find that the critical points are (0, 0), (3, 0), (0, 2), (1, 1).
Then at (0, 0):  
3 0
M=
0 2
which has eigenvalues 3 and 2, with eigenvectors are (1,0), and (0,1) and is an unstable
node.
At (3, 0):  
−3 −6
M=
0 −1
which has eigenvalues -3 and -1, with eigenvectors (1,0), and (-3,1) and is a stable node.
At (0, 2):  
−1 0
M=
−2 −2
which has eigenvalues -1 and -2, with eigenvectors (-1,2), and (0,1) and is a stable node.
At (1, 1):  
−1 −2
M=
−1 −1
√ √ √ √
which has eigenvalues −1 − 2 and −1 + 2, with eigenvectors ( 2, 1) and (− 2, 1).
and is a saddle point.
Again, as x and y represent populations we require that any trajectory which starts out
in the first quadrant will remain there. As in the previous example this is indeed the
case as the axes are particular trajectories.
Looking at the phase diagram we can see that, in the long term, depending on the initial
data, either the rabbits or the sheep will survive.
Other values of the coefficients will give different outcomes - see problem sheet 2.

2.3.5 Another important example: limit cycles

Consider the plane autonomous system:

1
ẋ = (1 − (x2 + y 2 ) 2 )x − y
1
ẏ = (1 − (x2 + y 2 ) 2 )y + x.

40
 Figure 2.12: The nullclines for the equations (2.13).

Figure 2.13: Phase diagram for the equations (2.13) for competitive species - no nullclines.

Figure 2.14: The phase diagram for the equations (2.13) for competitive species - with the nullclines
Rabbits or sheep survive, depending on the initial data.
41
Put x2 + y 2 = r2 then
X = x(1 − r) − y
Y = y(1 − r) + x
and one sees that only critical point is (0, 0). One could go through the classification for
this to find that it is an unstable spiral (exercise!).
Alternatively, in this case, we can analyse the full nonlinear system. We shall transform
to polar coordinates. The simplest way to do this is as follows: first

rṙ = xẋ + y ẏ = x[x(1 − r) − y] + y[y(1 − r) + x]

= r2 (1 − r)
or
ṙ = r(1 − r).
Then, with
y = r sin θ,
we find
ẏ = ṙ sin θ + r cos θθ̇ = y(1 − r) + x,
which gives θ̇, so the system becomes

θ̇ = 1
ṙ = r(1 − r).

Unlike the system in its previous form, we can solve this. First

θ = t + const,

and then Z Z Z  
dr 1 1
dt = = + dr
r(1 − r) r 1−r
so
r
log = t + const
|1 − r|
i.e.
r
= Aet .
1−r
Solve for r and change the constant:
1 1
r= −t
= 1
1 + Be 1 + ( r0 − 1)e−t

where r(0) = r0 .
Note that as t → ∞, r → 1, while as t → −∞ either r → 0 if r0 < 1 or r → ∞ at some
finite t if r0 > 1.

42
Now it is clear that the origin is an unstable spiral, and that the trajectories spiral
out of it anticlockwise. We can also see that r = 1 is a closed trajectory and that all
other trajectories (except the fixed point at the origin) tend to it; we call such a closed
trajectory a limit cycle. It is stable because the other trajectories converge on it. (For
an example of an unstable limit cycle we could consider the same system but with t
changed to −t.)

limit cycle

Figure 2.15: Phase diagram with a limit cycle

Another system with a limit cycle arises from the Van der Pol equation:

ẍ + (x2 − 1)ẋ + x = 0

where  is a positive real constant. If  = 0 this is the harmonic oscillator again. If  6= 0

then the usual trick produces a plane autonomous system:

ẋ = y

ẏ = −(x2 − 1)y − x.
The only critical point is (0, 0) and it’s an unstable spiral for  > 0 (exercise!).

Claim: Its beyond us to show this, but this system has a unique limit cycle, which
is stable. There are some good illustrations for this in e.g. Boyce and di Prima (pp
496–500 of the 5th edition).

43
2.4 The Bendixson–Dulac Theorem

It’s important to be able to detect periodic solutions, but it can be tricky. We end this
section with a discussion of a test that can rule them out.

Theorem 2.1. ( Bendixson–Dulac) Consider the system ẋ = X(x, y), ẏ = Y (x, y),
with X, Y ∈ C 1 . If there exists a function ϕ(x, y) ∈ C 1 with
∂ ∂
ρ := (ϕX) + (ϕY ) > 0
∂x ∂y
in a simply connected region R then there can be no nontrivial closed trajectories lying
entirely in R.

Proof. (By nontrivial, I mean I want the trajectory must have an inside i.e. it isn’t just
a fixed point.) So suppose C is a closed trajectory lying entirely in R and let D be the
disc (which also lies entirely in R, as R is simply connected) whose boundary is C. We
apply Green’s theorem in the plane. Consider the integral
Z Z Z Z  
∂ ∂
ρ dxdy = (ϕX) + (ϕY ) dxdy
D D ∂x ∂y
I
= −ϕY dx + ϕX dy
C
I
= −ϕ (−ẏdx + ẋdy) .
C
But on C, dx = ẋdt, dy = ẏdt so this is zero, which contradicts positivity of ρ, so there
can be no such C.

2.4.1 Corollary.

If
∂X ∂Y
+
∂x ∂y
has fixed sign in a simply connected region R, then there are no nontrivial closed tra-
jectories lying entirely in R.

This is just the previous but with ϕ const — in an example, always try this first!

2.4.2 Examples

(i) the damped pendulum (section 2.3.2)

ẋ = y

44
 g
ẏ = − sin x − ky
l
has no periodic solutions.
Here
∂X ∂Y
+ = −k < 0;
∂x ∂y
now use the corollary.

(ii)
ẍ + f (x)ẋ + x = 0
has no periodic solutions in a simply connected region where f has a fixed sign.

By the usual trick we get the system

ẋ = y

ẏ = −yf (x) − x
then
∂X ∂Y
+ = −f (x)
∂x ∂y
and we use the corollary.

(iii) The system

ẋ = y
ẏ = −x − y + x2 + y 2
has no periodic solutions.

The corollary doesn’t help so try the general case:

ρ := (ϕX)x + (ϕY )y = ϕ(−1 + 2y) + Xϕx + Y ϕy .

Now guess: ϕy = 0 then

ρ = ϕ(−1 + 2y) + yϕx = −ϕ + y(ϕx + 2ϕ)

so if we take ϕ = −e−2x the coefficient of y (which can take either sign) is zero
and ρ = 2e−2x > 0 and we are done.

45
PART II Partial Differential Equations.

3 First order semi-linear PDEs: the method of character-

istics

3.1 The problem

In this chapter, we consider first-order PDEs of the following form:

∂z ∂z
P (x, y) + Q(x, y) = R(x, y, z) (3.1)
∂x ∂y

The PDE is said to be semi-linear as it is linear in the highest order partial derivatives,
with the coefficients of the highest order partial derivatives depending only on x and y.
If P and Q depend also on z the PDE is said to be quasi-linear. We will consider only
semi-linear equations.
We will assume throughout this section that, in the region specified, P (x, y) and Q(x, y)
are Lipschitz continuous in x and y and R(x, y, z) is continuous and Lipschitz continuous
in z. This will be enough to ensure that the characteristic equations have a solution
through each point in the region, which is unique except at points where P = 0 and
Q = 0 (see below) .
We want to find a unique solution to (3.1) given suitable data and determine its domain
of definition. This is the region in the (x, y)-plane in which the solution is uniquely
determined by the data. It turns out to depend on both the equation and the data.
The solution of (3.1) will be a function

z = f (x, y)

but can be thought of as the surface defined by this equation, or equivalently defined by
the equation
Σ(x, y, z) := z − f (x, y) = 0. (3.2)
We shall refer to this as the solution surface and call it Σ. The method of solution of
the equation will be to generate Σ.

46
 n
z

P
=0

Figure 3.1: The solution surface

A normal to the solution surface is defined by

 
∂Σ ∂Σ ∂Σ
n = OΣ = , , = (−fx , −fy , 1).
∂x ∂y ∂z
(this is a fact from Prelims for the single-subject mathematicians; Maths and Comp
students should ask their tutors for enlightenment) so consider the vector t = (P, Q, R).
Then
∂f ∂f
t · n = −P −Q + R,
∂x ∂y
which vanishes by (3.1), so t is tangent to the surface Σ.

3.2 The big idea: characteristics

We look for a curve Γ whose tangent is t. If Γ = (x(t), y(t), z(t)) in terms of a parameter
t this means
dx

= P (x, y), (a)  
dt 

dy

= Q(x, y), (b) (3.3)
dt 
dz


= R(x, y, z). (c) 

dt
These are the characteristic equations and the curve Γ is a characteristic curve or just a
characteristic. Given a characteristic (x(t), y(t), z(t)), call the curve (x(t), y(t), 0), which
lies below it in the (x, y)-plane, the characteristic projection or characteristic trace.
The next result shows that characteristics exist, and gives the crucial property of them:
Proposition 3.1. Suppose that P (x, y) and Q(x, y) are Lipschitz continuous in x
and y and R(x, y, z) is continuous and Lipschitz continuous in z. Then
(a) There is a characteristic projection through each point of the plane, and they
can only meet at critical points (ie points where P and Q are both zero).
(b) Given a point p ∈ Σ, the characteristic through p lies entirely on Σ.

47
Proof

(a) This is exactly the same as in Section 2.1, as (3.3)(a),(b) is an autonomous system
dy
and, if P 6= 0, reduces to dx = QP , which has a unique solution for given initial
data. We can then find z uniquely along y = y(x) from

dz R(x, y(x), z)
= .
dx P (x, y(x))

(b) (This is fiddly.) Let (x(t), y(t), z(t)) be the characteristic through p and set

Σ(t) = z(t) − f (x(t), y(t)).

Then to prove that the characteristic lies on the surface Σ we must show that
Σ(t) ≡ 0. To do this we will show that Σ(t) satisfies an IVP which has the unique
solution zero so that Σ(t) = 0 for all t.
From (3.3) and the chain rule

dΣ ∂f
= R(x(t), y(t), z(t)) − P (x(t), y(t))
dt ∂x
∂f
− Q(x(t), y(t)),
∂y
while from (3.1)

∂f ∂f
R(x(t), y(t), f (x(t), y(t)) − P (x(t), y(t)) − Q(x(t), y(t)) = 0.
∂x ∂y
Subtract these
dΣ
= R(x, y, z) − R(x, y, f )
dt
Now put f = z − Σ, then the RHS here is a function F (Σ(t), t) for some F with
F (0, t) = 0. Furthermore, if the point p is (x(0), y(0), z(0)) then Σ(0) = 0.
Thus we have an ODE for Σ(t), with given initial value, and the RHS is continuous
and is Lipschitz continuous in Σ (check!), so the IVP has a unique solution. But
Σ(t) = 0 is a solution, so it is the only one. Therefore, Σ = 0 all along Γ, and Γ
lies on the solution surface. Q.E.D.

Thus the solution surface Σ is generated by a collection of characteristics.

3.2.1 Examples of characteristics

We need to gain proficiency in calculating characteristics.

48
 (a) Calculate the characteristics for the PDE:

∂z ∂z
x +y = z.
∂x ∂y

From (3.3) write down the characteristic equations and solve them:

dx
= P = x; x = Aet
dt
dy
= Q = y; y = Bet
dt
dz
= R = z; z = Cet
dt
with A, B, C constants (trivial to solve).

(b) Calculate the characteristics for the PDE:

∂z ∂z
y + = z.
∂x ∂y

dx t2
= y; x = Bt + + A
dt 2
dy
= 1; y = B + t
dt
dz
= z; z = Cet .
dt
with A, B, C constants. To solve this system, pass over the first, solve the second,
then come back to the first and third. (I am adopting a convention to introduce
the constants A, B, C in the first, second and third of the characteristic equations
respectively.)

In general solving the characteristic equations needs experience and luck; there isn’t a
general algorithm.

3.3 The Cauchy problem

A Cauchy Problem for a PDE is the combination of the PDE together with boundary
data that, in principle, will give a unique solution, at least locally. We will look for
suitable data and determine the domain on which the solution is uniquely determined.
Suppose we are given the solution z of (3.1) along a curve γ0 (the data curve) in the
(x, y)-plane. This produces a curve γ in space (the initial curve):

49
 γ(s) = (x(s), y(s), z(s))
z

γ0 (s) = (x(s), y(s), 0))

Figure 3.2: Geometry of the Cauchy problem.

We introduce a parameter s along γ so it is (x(s), y(s), z(s)), while γ0 is the projection

(x(s), y(s), 0), and we will assume x, y, z are continuously differentiable. Then, to solve
(3.1), we construct the solution surface Σ by taking the characteristics through the points
of γ (because Proposition 3.1(b) tells us that the solution surface is generated by these
characteristics) . Thus the method of solution, the method of characteristics, is

(i) Parametrise γ as (x(s), y(s), z(s)).

(ii) Solve
∂x
=P
∂t
∂y
=Q
∂t
∂z
=R
∂t
for (x(t, s), y(t, s), z(t, s)) with data x(0, s) = x(s); y(0, s) = y(s); z(0, s) = z(s).

then, knowing (x(t, s), y(t, s), z(t, s)), we have found Σ in parametric form. We would
like to find the solution explicitly, that is z in terms of x and y, a question we will explore
below, and there is a restriction on the data for the method to work, also to be found
later.

3.4 Examples

(a) Solve
∂z ∂z
y + = z,
∂x ∂y
with z(x, 0) = x for 1 ≤ x ≤ 2.

50
 z

2
γ

1
y

γ0
1
2
x

Figure 3.3: The data curve for this problem

We introduce a parameter s for the data, say γ(s) = (s, 0, s), for 1 ≤ s ≤ 2, and
then solve the characteristic equations (done in section 3.2.1) with this as data at
t=0
t2
x = Bt + + A; x(0, s) = A = s
2
y = B + t; y(0, s) = B = 0
z = Cet ; z(0, s) = C = s
So, C = s, B = 0, A = s and the parametric form of the solution is

x = s + 12 t2 


y=t (3.4)
z = set


for 1 ≤ s ≤ 2.

(b) (From Ockendon et al) Solve

∂z ∂z
x +y = (x + y)z (3.5)
∂x ∂y

with z = 1 on the segment of the circle (x − 2)2 + y 2 = 2, y ≥ 0.

√ √
So we can take γ(s) = (2 − 2 cos s, 2 sin s, 1), s ∈ [0, π], and solve the charac-
teristic equations:
∂x √
= P = x; x = Aet ; A = (2 − 2 cos s)
∂t
∂y √
= Q = y; y = Bet ; B= 2 sin s
∂t

51
 ∂z √ √
= R = (x + y)z = ((2 − 2 cos s + 2 sin s)et )z.
∂t
We can integrate the final equation to get
√ √ √ √
log |z| = ((2 − 2 cos s + 2 sin s)et ) + C; C = −(2 − 2 cos s + 2 sin s).
So the parametric form of the solution is
√ t

x = (2
√ − 2 cos s)e 
t
y = ( 2 sin s)e√ (3.6)
√
z = exp[((2 − 2 cos s + 2 sin s)(et − 1)]


3.5 Domain of definition

Where is the solution determined uniquely by the data? This is the domain of definition
and is the region in the (x, y)-plane where the solution surface is uniquely determined
and is given explicitly as z = f (x, y).
Bounded Initial Data:
In general the initial curve could be bounded or semi-infinite. The solution surface is
swept out by the characteristics through the initial curve, so the solution will be defined
in the region swept out by the projections of the characteristics through the initial curve.
In particular if the initial curve is bounded, the domain of definition will be bounded by
the projections of the characteristics through the end points of the initial curve.
Blow up:
The method of characteristics reduces the PDE (3.1) to a system of ODEs. As we have
already seen nonlinear ODEs can give rise to solutions which blow up, so the same must
be true of non linear PDEs, even if those that are semi-linear.
Explicit solution: We want z as a function of x and y, so we need to be able to
eliminate t and s in favour of x and y, at least in principle. For this, recall from Prelims
the definition of the Jacobian:
 
∂(x, y) x t yt
J(s, t) = = det . (3.7)
∂(t, s) xs ys
Now if
x = x(t, s), and y = y(t, s)
are continuously differentiable functions of t and s in a neighbourhood of a point, then
a sufficient condition to be able to find unique continuously differentiable functions
t = t(x, y) and s = s(x, y)
in some neighbourhood of the point, is that J be non-zero at the point. We can then
substitute into z = z(t, s) to get
z = z(t(x, y), s(x, y)) = f (x, y),

52
a continuously differentiable function of x and y as required. This comes from a result
known as the Inverse Function Theorem. We will have to take it on trust, but it is the
two dimensional equivalent of the one dimensional result you saw in Analysis in Prelims
- where the function f has a differentiable inverse if f has a non zero derivative.
Thus we require first that J(s, 0) 6= 0 on the initial curve. If this is so, then the problem
has a unique explicit solution, at least close to the initial curve. Then, we can continue
to extend the solution away from the curve so long as J 6= 0 and the solution does not
blow up.
This fails if the data curve touches a characteristic projection. For
   
xt yt P Q
J(s, 0) = det = det . (3.8)
xs ys xs ys

So if, for some s, J(s, 0) = 0 then, if a dash denotes differentiation with respect to s on
γ0 = (x(s), y(s), 0) , we have

P (x, y)y 0 − Q(x, y)x0 = 0, (3.9)

so the data curve is tangent to the projection of the characteristic at that point.
Typically away from the initial curve we will only have J = 0 if the characteristics cross
each other (so that uniqueness fails) and for semi-linear problems this can only happen
at critical points.
Examples:
Example (a) : Let us determine the domain of definition for the example of section
3.4(a):
The solution surface is swept out by the characteristics through γ, so has edges given by
the characteristics through the ends of γ, which are at s = 1 and s = 2.
At s = 1, the characteristic projection is
1 1
x = 1 + t2 ; y=t so x = 1 + y2;
2 2
at s = 2 it’s
1 1
x = 2 + t2 y=t so x = 2 + y 2
2 2

53
 2

x = 1 + 21 y 2
1
x = 2 + 12 y 2

0 x=2
x=1

-1

-2
0.0 0.5 1.0 1.5 2.0 2.5 3.0

Figure 3.4: The domain of definition for this problem

so the solution surface lies above the region

1 1
1 + y2 ≤ x ≤ 2 + y2 (3.10)
2 2

54
and these curves don’t meet .
It is easy to see from (3.4) that the solution can be given explicitly as z(x, y) = (x− 21 y 2 )ey
for 1 + 21 y 2 ≤ x ≤ 2 + 12 y 2 .
Thus the solution does not blow up and can be given explicitly. Therefore (3.10) gives
the domain of definition. Or we could calculate J which is never zero.
Example (b): Now look at the example of section 3.4(b); First consider
√ t
√
(√ 2 sin s)et
 
(2 −
√ 2 cos s)e
J = det t
( 2 sin s)e ( 2 cos s)et
√
= 2e2t (1 − 2 cos s),
This vanishes when s = π4 , that is along the characteristic projection y = x, which
touches the data curve at (1, 1). So we will restrict the data curve to s ∈ [0, π/4).
The data curve starts at s = 0, so the solution surface will have an edge given by the
characteristic through the end of γ at s = 0 and at s = 0 the characteristic projection is
y = 0. But the domain of definition is swept out by the projections of the characteristics
through γ. So the domain of definition is 0 ≤ y < x.
Remark: More generally if the data curve and characteristic projection coincide then,
if the initial curve γ is a characteristic so that z(s) satisfies dz
ds = R, there will be an
infinity of solutions through γ, while otherwise there will be no solution.
For, if γ is a characteristic, then let C be any curve through γ whose projection is nowhere
tangent to a characteristic projection. Then there is a solution surface through C. But
this was any C so there is an infinity of solutions. On the other hand, if the data curve
is a characteristic projection, then there can only be a solution if γ = (x(s), y(s), z(s))
is a characteristic, as if the solution surface is z = f (x, y), then z(s) = f ((x(s), y(s)) so,
using the chain rule,
dz dx dx
= fx + = P fx + Qfy fy = R.
ds ds ds

So, returning to our example above, if one continues along the data curve from s = π4
there will be problems as each characteristic meets the data curve in two points, so the
initial data is likely to be inconsistent. If one changes the data curve to continue to
the left of y = x (for example the curve with the small dashes in the diagram) then it
is likely there would only be problems on the characteristic y = x. But if we continue
the data curve to include a section of the line y = x (curve with larger dashes in the
diagram) the problem is likely to have no solution.

55
 3.0
2.5
2.0
1.5
1.0
0.5
0.0
0 1 2 3 4

Figure 3.5: Different data curves for problem (3.5)

Example (c): Consider the PDE

∂z ∂z
+ = z2,
∂x ∂y

with z(x, 0) = 1.
parametrise the initial curve as (s, 0, 1). Then the solutions of the characteristics are:
1 1
x = t + s, y = t, z = 1−t , so z = 1−y which blows up at y = 1.

3.6 Cauchy data and the Cauchy Problem:

We have seen that a necessary condition on the data for the solution to exist close to
the initial curve is that on γ0

P (x, y)y 0 − Q(x, y)x0 6= 0. (3.11)

If it holds, we call the data Cauchy data.

Thus the Cauchy Problem is the PDE (3.1) with Cauchy data.

3.7 Discontinuities in the first derivatives

The characteristic projections have another property. They are the only curves across
which the solution surface can have discontinuities in the first derivatives. For, suppose
that γ is the curve (x(s), y(s), z(s)) in the solution surface, across which there are discon-
tinuities in the first order partial derivatives but z is continuous. Use the superscript ±
to denote the solution on either side of γ and denote the jumps in the partial derivative

56
by [zx ]+ + − + + −
− = zx − zx and [zy ]− = zy − zy . Then differentiating each solution along γ,
and also using (3.1)

dz + dx dy
= zx+ + zy+ , (3.12)
ds ds ds
+ +
R = P zx + Qzy . (3.13)

and
dz − dx dy
= zx− + zy− , (3.14)
ds ds ds
R = P zx− + Qzy− . (3.15)

dz + dz −
But z is continuous across γ, so ds = ds , and subtracting, we see that

dx dy
0 = [zx ]+
− + [zy ]+
− , (3.16)
ds ds
0 = P [zx ]+ +
− + Q[zy ]− . (3.17)

Thus for there to be a non-zero jump, the solution of these simultaneous equations must
be nonzero, so the determinant of the coefficients must be zero. That is
 
P Q
0 = det dx dy , (3.18)
ds ds

so that the projection of the curve must be a characteristic projection. That it is a

characteristic curve now follows from (3.12) and (3.13). See problem sheet 3 for an
example of this.

Figure 3.6: The curve γ.

57
3.8 General Solution

Another problem we could have considered, is what is the most general solution of (3.1)?
Just as we expect the most general solution of an ODE to have n arbitrary constants, so
we expect the most general solution of a PDE of order n to have n arbitrary functions.
∂z
For example: The first order PDE ∂x (x, y) = 0, has the most general solution z = f (y)
where f is an arbitrary function.

58
4 Second order semi-linear PDEs

4.1 Classification

In this section, we are interested in second-order PDEs of the following form:

a(x, y)uxx + 2b(x, y)uxy + c(x, y)uyy = f (x, y, u, ux , uy ). (4.1)

| {z }
principal part

This PDE is said to be linear if f is linear in u, ux , uy , otherwise it is said to semi-linear.

(If the coefficients a, b, c also depend on u, ux , uy it is said to be quasi-linear. We will
consider only semi-linear equations.) You have seen the following examples in Prelims:

uxx + uyy = 0 Laplace’s equation

uxx − uyy = 0 wave equation if y = ct
uxx − uy = 0 heat equation if y = t/κ.

Equations that are linear (in the dependent variable) have solutions that can be combined
by linear superposition (taking linear combinations). In general PDEs that are nonlinear
(for example where f , above, depends nonlinearly on u or its derivatives) do not have
solutions that are superposable.
We will assume throughout that functions are suitably differentiable.

4.1.1 The idea:

In this section, the key idea is to change coordinates so as to simplify the principal part.
So we make the change of variables

(x, y) → (ϕ(x, y), ψ(x, y));

with non vanishing Jacobian (basically this ensures that the map is locally invertible):

∂(ϕ, ψ)
= ϕx ψy − ϕy ψx 6= 0.
∂(x, y)

We will abuse the notation a little and write (the solution) u as either a function of
(x, y) or (ϕ, ψ) as required.
For the change in the partials, we calculate

ux = uϕ ϕx + uψ ψx ; uy = uϕ ϕy + uψ ψy

59
then
uxx = uϕϕ ϕ2x + 2uϕψ ϕx ψx + uψψ ψx2 + uϕ ϕxx + uψ ψxx


uxy = uϕϕ ϕx ϕy + uϕψ (ϕx ψy + ψx ϕy ) + uψψ ψx ψy + uϕ ϕxy + uψ ψxy (4.2)
uyy = uϕϕ ϕ2y + 2uϕψ ϕy ψy + uψψ ψy2 + uϕ ϕyy + uψ ψyy


so that (4.1) becomes

A(ϕ, ψ)uϕϕ + 2B(ϕ, ψ)uϕψ + C(ϕ, ψ)uψψ = F (ϕ, ψ, u, uϕ uψ ) (4.3)

with
A = aϕ2x + 2bϕx ϕy + cϕ2y


B = aϕx ψx + b(ϕx ψy + ϕy ψx ) + cϕy ψy (4.4)
C = aψx2 + 2bψx ψy + cψy2 .


(Beware, F will include lower order derivatives from (4.2).) In a matrix notation (4.4)
is (check!)
     
A B ϕx ϕy a b ϕx ψx
=
B C ψx ψy b c ϕy ψy
so that, taking determinants,
 2
∂(ϕ, ψ)
(AC − B 2 ) = (ac − b2 )(ϕx ψy − ψx ϕy )2 = (ac − b2 ) . (4.5)
∂(x, y)
(We could obtain (4.5) directly from (4.4) but the matrix notation makes the computa-
tion simpler.) Now (4.5) leads to a classification of second-order linear PDEs:

4.1.2 The Classification

Second-order linear PDEs are classified into three types as follows:

1. ac < b2 hyperbolic: e.g. wave equation;

2. ac > b2 elliptic: e.g. Laplace equation;

3. ac = b2 parabolic: e.g. heat equation.

So, by (4.5) the class of the equation is invariant under transformations with non-
vanishing Jacobian.
We shall look at the classification in terms of the quadratic polynomial

a(x, y)λ2 − 2b(x, y)λ + c(x, y) = 0. (4.6)

Note: We will assume that a 6= 0, in the domain under consideration. If a = 0 but c 6= 0,

we can swap the roles of x and y.

60
Case 1: hyperbolic type
So ac < b2 and the quadratic has distinct real roots λ1 , λ2 , say. So
 2
dy dy
a(x, y) − 2b(x, y) + c(x, y) = 0. (4.7)
dx dx

is equivalent to
dy dy
= λ1 (x, y), = λ2 (x, y). (4.8)
dx dx
Suppose these equations have solutions ϕ(x, y)=constant, ψ(x, y) =constant, respec-
tively. Set as change of variables

ϕ = ϕ(x, y), ψ = ψ(x, y).

Then, on ϕ(x, y)=constant,

dy
ϕx + ϕy =0
dx
so that
λ1 ϕy = −ϕx
and thus
A(ϕ, ψ) = a(x, y)(ϕx )2 + 2b(x, y)ϕx ϕy + c(x, y)(ϕy )2 = 0,
ϕx ψx
and analogously C(ϕ, ψ) = 0. But λ1 6= λ2 , so ϕy 6= ψy , and from (4.5) B 6= 0. Divide
(4.3) by B to obtain the equation in the form

uϕψ = G(ϕ, ψ, u, uϕ , uψ ). (4.9)

This is the normal form (or canonical form) for a hyperbolic equation; the equation
(4.7) is the characteristic equation; ϕ, ψ are characteristic variables; curves on which ϕ
or ψ are constant are characteristic curves. We can often solve (4.9) explicitly.

Examples:

(a)
uxx − uyy = 0.
We already know how to solve this, but let us apply the method. So

a = 1, b = 0, c = −1, and λ2 − 1 = 0.

We can take
λ1 = 1, λ2 = −1
and solve (4.8)
y 0 (x) = 1 y 0 (x) = −1

61
 to get
ϕ=x−y ψ = x + y.
(There is clearly lots of choice at this stage.) The equation has become

uϕψ = 0,

which we solve at once by

u = f (ϕ) + g(ψ),
a solution known from Prelims. So the characteristic curves of the wave equation
are x + ct=const and x − ct =const.

(b) An example with data: solve

(x + y)
xuxx − (x + y)uxy + yuyy + (ux − uy ) = 0, for y 6= x
(y − x)

with
1
u = (x − 1)2 , uy = 0 on y = 1.
2
Problem is hyperbolic provided x 6= y (check). The quadratic (4.6) is

xλ2 + (x + y)λ + y = 0

= (λ + 1)(xλ + y);
so choose
y
λ1 = −1 λ2 = −
x
and solve
y
y 0 (x) = −1; y 0 (x) = − ,
x
by x + y = const; xy =const, so put

ϕ = x + y; ψ = xy.

Calculate
ux = uϕ + yuψ
uy = uϕ + xuψ
so that
uxx = uϕϕ + 2yuϕψ + y 2 uψψ
uxy = uϕϕ + xuϕψ + yuϕψ + xyuψψ + uψ
uyy = uϕϕ + 2xuϕψ + x2 uψψ .
(It is always better to calculate the derivatives directly, rather than trying to
remember formulae.)

62
Now the PDE becomes

0 = x[uϕϕ + 2yuϕψ + y 2 uψψ ]

−(x + y)[uϕϕ + (x + y)uϕψ xyuψψ + uψ ]

+y[uϕϕ + 2xuϕψ + x2 uψψ ]
+(x + y)uψ
= (4xy − (x + y)2 )uϕψ
so
uϕψ = 0
and the solution is

u = f (ϕ) + g(ψ) = f (x + y) + g(xy).

To impose the data, calculate

uy = f 0 (x + y) + xg 0 (xy)

so on y = 1,
1
u = f (x + 1) + g(x) = (x − 1)2
2
uy = f 0 (x + 1) + xg 0 (x) = 0.
Differentiate the first:
f 0 (x + 1) + g 0 (x) = x − 1
and solve simultaneously with the second:

g 0 (x) = −1,
and integrate to find
g(x) = −x + c.
Substitute back in u(x, 1):
1 1
f (x + 1) = (x − 1)2 + x − c = (x + 1)2 − x − c,
2 2
so
1
f (x) = x2 − x + 1 − c.
2
Finally
1
u = f (x + y) + g(xy) = (x + y)2 − (x + y) + 1 − xy.
2

63
Case 2: elliptic type
Now ac > b2 so (4.6) has a complex conjugate pair of roots, and the integral curves of

y 0 (x) = λ(x, y); y 0 (x) = λ̄(x, y)

are in complex conjugate pairs, ϕ(x, y) =const; ψ(x, y) = ϕ̄(x, y) =const. Then A =
C = 0, B 6= 0 and the equation becomes

uϕϕ̄ = G(ϕ, ϕ̄, u, uϕ , uϕ̄ ).

Introduce new variables, ζ, η, given by ϕ = ζ + iη, ϕ̄ = ζ − iη, to obtain the normal

form for an elliptic equation (check):

uζζ + uηη = H(ζ, η, u, uζ , uη ), (4.10)

which closely resembles Laplace’s equation.

Example: Classify and reduce to normal form the PDE

yuxx + uyy = 0, for y > 0.

ac − b2 = y so the equation is elliptic when y > 0.

The characteristic equation is
y(y 0 )2 + 1 = 0
that is
y 1/2 y 0 = ±i,
so integrating
2y 3/2 ∓ 3ix = const.
So take as variables ζ = 2y 3/2 ; η = 3x. Making the substitution, we find that

3ζ(uζζ + uηη ) + uη = 0,

but ζ 6= 0 so the normal form is

uζ
uζζ + uηη = − .
3ζ

Case 3: parabolic type

Now ac = b2 so (4.6) has a repeated root λ(x, y). Solve y 0 (x) = λ(x, y) for one new
coordinate ϕ, and pick any ψ with

ϕx ψy − ϕy ψx 6= 0 (4.11)

64
as the other, then A = 0 so B 2 = AC = 0. But C =6 0, as ψ =const is not a characteristic
curve by (4.11), so we get the normal form for a parabolic equation:

uψψ = G(u, ϕ, ψ, uϕ , uψ ),

which closely resembles the heat equation.

Example:
Classify and reduce to normal form the equation

x2 uxx + 2xyuxy + y 2 uyy = 0 for x > 0. (4.12)

The relevant quadratic is

x2 λ2 − 2xyλ + y 2 = 0 = (xλ − y)2

y
which has equal roots, so this equation is parabolic; λ = x so solve
dy y y
= , to get, for example, ϕ =
dx x x
and take, for example, ψ = x. Calculate
y
ux = − uϕ + uψ
x2
1
uy = uϕ
x
so that
y2 y y
uxx = 4
uϕϕ − 2 2 uϕψ + uψψ + 2 3 uϕ
x x x
y 1 1
uxy = − 3 uϕϕ + uϕψ − 2 uϕ
x x x
1
uyy = 2 uϕϕ .
x
The equation becomes

y2
 
2y 2y
x2 uϕϕ + uϕψ + u ψψ + uϕ +
x4 x2 x3
   
y 1 1 2 1
+ 2xy − 3 uϕϕ + uϕψ − 2 uϕ + y uϕϕ = x2 uψψ = 0 (4.13)
x x x x2
so the normal form is
uψψ = 0
with general solution u = F (ϕ) + ψG(ϕ). In terms of the original variables this is:
y y
u=F + xG . (4.14)
x x

65
NB Very often, a question like this will be phrased in the form ‘Classify and reduce to
normal form the equation (4.12) and show that the general solution can be written as
(4.14)’. Therefore candidates for ϕ and ψ are proposed by the question itself.

A warning example:
The type can change e.g. classify the equation

uxx + yuyy = 0.

Then
λ2 + y = 0, λ2 = −y,
and this is:

• elliptic in y > 0,

• parabolic at y = 0,

• hyperbolic in y < 0.

4.2 Characteristics:

The characteristics of second order semi-linear PDEs have analogous properties to the
characteristic projections of first order semi-linear PDEs. (Note the difference in termi-
nology.)
Firstly, if there are discontinuities in the second derivatives of a solution across a given
curve, then that curve must be a characteristic curve. To see this, suppose that the curve
Γ, given parametrically by (x(s), y(s)), is a curve across which there are discontinuities
in the second derivatives of the solution. Let u+ xx etc denote values on one side of Γ
−
and uxx denote values on the other side of Γ. Then differentiating ux (x(s), y(s)) and
uy (x(s), y(s)) along Γ, and noting that u, ux , uy are continuous across the curve,

dux dx ± dy ±
= u + u
ds ds xx ds xy
duy dx ± dy
= uyx + u±
ds ds ds yy
and also f (x, y, u, ux , uy ) = a(x, y)uxx + 2b(x, y)uxy + c(x, y)u±
± ±
yy .

Subtracting the ‘minus’ equation from the ‘plus’ equation

dx dy
0= [uxx ]+
−+ [uxy ]+
−
ds ds
dx dy
0= [uyx ]+
−+ [uyy ]+
−
ds ds
0 = a(x, y)[uxx ]− + 2b(x, y)[uxy ]− + c(x, y)[uyy ]+
+ +
−,

66
where [uxx ]+ + −
− = uxx − uxx denotes the jump in uxx across Γ, etc. If there are to be
discontinuities in the second derivatives, then this set of equations in the jumps must
have a nonzero solution, so that the determinant of the coefficients must be zero. Thus
 2  2
dy dx dy dx
a(x, y) − 2b(x, y) + c(x, y) = 0, (4.15)
ds ds ds ds

so Γ is a characteristic.
Furthermore, under suitable smoothness conditions, the Cauchy problem for a second
∂u
order semi-linear PDE, where u and ∂n are given along a curve Γ, will have a unique
local solution provided Γ is nowhere tangent to a characteristic curve. This result is
beyond the scope of this course and will be investigated further in the Part B course,
Applied PDEs. It can be seen that it is necessary that Γ is not a characteristic curve, as
if u exists then it must have unique second order partial derivatives along Γ and exactly
as above this can only be true when Γ is not a characteristic curve.
Remark: Our previous work carried the implicit assumption that a 6= 0. Note that
(4.15) gives a method of calculating the characteristic curves if a = 0. In particular if
a = 0 and c = 0 then the characteristic curves are x =const, y =const.

4.3 Type and data: well posed problems

We want to say something about the notion of well posedness and its connection with
type. Our examples are mostly based on knowledge acquired in Prelims.
A problem, consisting of a PDE with data, is said to be well posed if the solution:

• exists

• is unique

• depends continuously on the data.

Recall that, in Section 1.5, we said that a solution of a DE is continuously dependent

on the data if the error in the solution is small provided the error in the initial data is
small enough.. We then gave a precise definition for ODEs. We now want to extend
this definition to PDEs. Data can be given in different ways, so to be precise we will
consider a problem where u(x, y) is the solution of a certain PDE in a bounded subset of
the plane D, with u given on some curve Γ. Then we will say that the solution depends
continuously on the data if :
∀ > 0 ∃δ > 0 such that if ui , i = 1, 2, are solutions with ui = fi on Γ then

sup |f1 − f2 | < δ ⇒ sup |u1 − u2 | < .

Γ D

67
The definition extends in a fairly obvious way to other types of data. (Note that there
are plenty of other ‘distances ’ we could use in place of taking the sup, but that is what
we will use here.)
Of the three requirements for well posedness it is existence which is the hardest to obtain.
In Prelims solutions were found for a number of linear problems, either by making a
change of variables and then integrating, or by using separation of variables and Fourier
series. Anything more than this is beyond the scope of this course. Uniqueness of
solution was also proved for a number of linear problems (even when you didn’t know
if the solution existed). Proving uniqueness and continuous dependence on the data is
much easier for linear problems as we can then start out by looking at the difference
between two solutions, which will then be the solution of some suitable problem. Later we
will look at the linear equations Poisson’s equation and the heat equation and state and
prove the maximum principle, which will enable us to prove uniqueness and continuous
dependence for suitable boundary data.
But first we need to consider what data might be appropriate. In Prelims you considered
three particular PDEs, each with a different type of data which arose from the particular
physical problem they modelled. These are summarised in the table below. It turns out
that there are mathematical as well as physical reasons why each problem had a different
type of data. So first we will look at some of these problems and consider which may be
well posed.

68
 Some Models from Prelims:

PDE Models Boundary conditions

∂u
Wave Equation: c2 uxx − utt = 0 Waves on a string; u, ∂t given t = 0 (IBVP)
Hyperbolic (Infinite string)
∂u
u gives displacement ∂t given t = 0; plus end point (IVP)
condition - say u = 0 at ends (Finite string)

69
∂u
Laplace’s Equation: uxx + uyy = 0, Potential Theory or u or ∂n given on boundary of D (BVP)
Elliptic (x, y) ∈ D Steady Heat;
u gives potential or temperature

Heat Equation: ut = κuxx Heat; u given at t = 0; (IBVP)

Parabolic plus endpoint conditions (Temperature
u gives temperature - u or ∂u
∂x given at ends in finite bar)
Some examples from Prelims: (We will assume that the data is smooth enough for
the following to hold.)

(a) Hyperbolic equation: The IVP and IBVP (initial-boundary-value problem) for the
wave equation
uxx − uyy = 0 (ct = y).
For the IVP (modelling an infinite string, where u is the displacement), we know
the solution is
1 x+y
Z
1
u = [f (x + y) + f (x − y)] + g(s)ds, (4.16)
2 2 x−y

where the data are u(x, 0) = f (x) and uy (x, 0) = g(x), −∞ < x < ∞. This is
d’Alembert’s solution of the IVP: it exists, and is unique and, intuitively at least,
a small change in ϕ, ψ gives a small change in u (see problem sheet for proof). So
this problem is well-posed.
For the IBVP (modelling a finite string length L, fixed at each end) consider the
data:
u(x, 0) = f (x), uy (x, 0) = g(x) 0 < x < L
u(0, y) = 0 = u(L, y).
So the boundaries are at x = 0, L. This IBVP is solved using separation of variables
and Fourier series to get a solution
X nπx  nπy nπy 
u= sin an cos + bn sin
n
L L L

with an , bn given in terms of f and g. Uniqueness was shown in Prelims, so this

is the unique solution. If we now appeal to intuition for continuous dependence on
the data this problem is well-posed.

(b) Elliptic equation:The BVP for the Laplace/Poisson’s equation (modelling steady
state heat, for example, where u is the temperature)

uxx + uyy = 0.

Do this first with data at the sides of a square, so 0 ≤ x, y ≤ a with

u(0, y) = u(a, y) = u(x, 0) = 0; u(x, a) = f (x).

nπy
Consider separable solutions un = sin nπx
a sinh a , then

X sinh nπy
a nπx
u= an sin
n
sinh(nπ) a

70
 and X nπx
f (x) = an sin
a
which determines the solution as a Fourier series.
Now a different BVP, with data at the circumference of the unit circle:

on r = 1, u = f (θ)

and in polars
1 1
urr + ur + 2 uθθ = 0.
r r
The separable solutions are

Arn + rBn (C cos nθ + D sin nθ)



A + B log r, n = 0
Regularity at r = 0 implies
∞
1 X
u = a0 + rn (an cos nθ + bn sin nθ)
2
1

and the boundary value at r = 1 requires

1 X
a0 + (an cos nθ + bn sin nθ) = f (θ)
2
which again is solved by Fourier methods.
Uniqueness was proved in Prelims so in each of these cases we have the unique
solution. It is plausible, but beyond our scope, to show that there is existence of
solution in general. Later we will prove that there is continuous dependence on
the data. So this problem is well posed

(c) Parabolic equation: The IBVP for the heat equation (modelling heat flow in a bar
length L, with the ends held at zero temperature, u is temperature. )

uxx = uy

on the semi-infinite strip where y = t > 0 and 0 < x < L, and data

u(x, 0) = f (x); u(0, y) = 0 = u(L, y).

The relevant separable solutions are

nπx − n2 π22 t
un = sin e L
L
so that
X nπx − n2 π2 2 t
u= an sin e L
n
L

71
 and the initial value requires
X nπx
f (x) = an sin
L
which is solved by Fourier methods. The solution exists provided the series for u
converges which it will do for positive t. However, note that for negative t the
exponentials grow rapidly with n and there is no reason to expect existence.
Uniqueness for this problem was done in Prelims, so again this is the unique
solution for positive t. Later we will prove continuous dependence on the data.
Thus the problem is well posed forward in time.

(d) What then is not well-posed? We give a few examples:

• BVPs for hyperbolic

e.g. uxx − uyy = 0 on the unit square with data

u(0, y) = u(1, y) = u(x, 0) = 0; u(x, 1) = f (x).

Recall this data gave a well-posed problem for the Laplace equation, but here
if f = 0, then sin nπx sin nπy will do, for any n, while it can be proved that
there is no solution at all if f 6= 0 (try the Fourier series to see what goes
wrong).
• IVPs for elliptic
e.g. uxx + uyy = 0 in the horizontal strip 0 ≤ y ≤ Y , −∞ < x < ∞, with
data
u(x, 0) = 0, uy (x, 0) = f (x).
We know (from the problem sheet) that this data gives continuous depen-
dence on the data for the wave equation. But not for Laplace’s equation.
For if f (x) = n1 sin nx it can be seen that u(x, y) = n12 sinh ny sin nx. But
sup | n12 sinh ny sin nx| → ∞ as n → ∞, whereas 1/n sin nx → 0. Thus small
changes in the initial data can lead to large changes in the solution. [More
precisely: Suppose that there is continuous dependence on the initial data
about the zero solution. That is: ∀ > 0, ∃δ > 0 such that

sup |f (x) − 0| < δ ⇒ sup |u(x, y) − 0| < .

x∈R x∈R,0≤y≤Y

But taking  = 1, say, there exists N such that for all n > N , sup | n12 sinh ny sin nx| >
1, but for any δ > 0 we can choose n > N such that | n1 sin nx| < δ, giving a
contradiction.

• IBVP for elliptic

e.g. uxx + uyy = 0 on the semi-infinite strip 0 ≤ x ≤ 1, y ≥ 0, with data

u(0, y) = u(1, y) = 0, u(x, 0) = 1, uy (x, 0) = 0.

72
 This data gives a well-posed problem for the wave equation. If we try for
separable solutions here, we have un = sin nπx cosh nπy so
X
u= an sin nπx cosh nπy.
n

Initial conditions need X

1= an sin nπx
whence
an = 0 n even
4
= n odd,
nπ
and then  
1 X 4
u ,y = (−1)n cosh(2n + 1)πy,
2 n
(2n + 1)π
which does not converge for any y > 0 (because the ”cosh” terms grow rapidly
with n) - there is no solution (strictly speaking, we’ve only shown that there
is no solution of the form considered; we need more).
• The BVP for the heat equation is not well-posed, but we won’t show that.

Again, it is beyond our scope to prove it in this course, but these different be-
haviours are universal for the different types of second-order, linear PDEs. In
tabulated form, which problems are well-posed?

IVP IBVP BVP

Hyperbolic yes yes no
Elliptic no no yes
Parabolic yes yes no

where the ‘yes’ for parabolic equations are only valid forward in time.

4.4 The Maximum Principle

4.4.1 Poisson’s equation

The normal form for second-order elliptic PDEs is

uxx + uyy = f (x, y, u, ux , uy ) (4.17)

73
The operator on the left-hand side is referred to as the Laplacian, for which the symbols
∇2 u and ∆u are often used as shorthand. Poisson’s equation is a special case of (4.17),
in which f depends only on x and y. We have already seen that appropriate boundary
data for (4.17) is to give just one boundary condition on u everywhere on a closed curve.
We will consider the Dirichlet problem where u is given on the boundary of D:

uxx + uyy = f (x, y) in D (4.18)

u = g(x, y) on ∂D. (4.19)

It was shown in Prelims, using the divergence theorem, that if a solution exists, then
it is unique. Using the maximum principle we will give another proof of this and also
show that there is continuous dependence on the data. The solution does exist, but
apart from the particular cases considered in Prelims that is beyond the scope of this
course. Existence of solutions of general elliptic problems will be considered in the Part
C courses Functional Analytic Methods for PDEs and Fixed point methods for nonlinear
PDEs.
Theorem 4.1. (The Maximum principle for the Laplacian ) Suppose u satisfies

∆u := uxx + uyy ≥ 0 (x, y) ∈ D, (4.20)

everywhere within a bounded domain D. Then u attains its maximum value on ∂D.

Remark: This Theorem of course applies to the Poisson equation where we ask that
uxx + uyy is given by a prescribed function f , but is equally applicable to get information
for non-linear problems, such as solutions of the equation uxx + uyy = u2 for which we
know that the right hand side has a given sign.
Remark: As u is a continuous function on the set D̄ which is a closed and bounded
subset of R2 and thus compact, we know that u achieves its maximum in some point
p ∈ D̄. The above theorem now tells us that this maximum value will indeed always be
achieved on the boundary, though does not exclude that the maximum is also achieved
at further points which might be in the interior.
In fact however the so called strong maximum principle (which is off syllabus) asserts
that a function u with ∆u ≥ 0 cannot have an interior maximum unless it is constant.
Proof:
If we denote the boundary of D by ∂D , then as D ∪ ∂D is a closed bounded set and
thus compact, u must attain its maximum somewhere in D or on its boundary. The
proof now proceeds in two parts.
Suppose first that uxx + uyy > 0 in D.
If u has an interior maximum at some point (x0 , y0 ) inside D, then the following condi-
tions must be satisfied at (x0 , y0 ):

ux = uy = 0, uxx ≤ 0, uyy ≤ 0.

74
But, as we assumed that uxx + uyy > 0 in all of D it is impossible for both uxx and uyy
to be non positive . Hence u cannot have an interior maximum within D. so it must
attain its maximum value on the boundary ∂D.
Suppose now that we only have uxx + uyy ≥ 0 in D. We perturb u to get a function v
which satisfies vxx + vyy ≥ 0, so we can apply the first part of the proof.
Consider the function

v(x, y) = u(x, y) + (x2 + y 2 ),
4
where  is a positive constant.
Then
vxx + vyy = uxx + uyy +  > 0
in D. So using the result just proved, v attains its maximum value on ∂D.
Now, suppose that the maximum value of u on ∂D is M and the maximum value of
(x2 + y 2 ) on ∂D is R2 , then the maximum value of v on ∂D (and thus throughout D)
is M + (/4)R2 . In other words, the inequality
 
u + (x2 + y 2 ) = v ≤ M + R2
4 4
holds for all (x, y) ∈ D. Letting  → 0, we see that u ≤ M throughout D, i.e. that u
attains its maximum value on ∂D.
It obviously follows (by using the above result with u replaced by −u) that, if ∆u ≤ 0
in D, then u attains its minimum value on ∂D.
In the case ∆u = 0, u therefore attains both its maximum and minimum values on ∂D.
This is an important property of Laplace’s equation.

Corollary 4.2. (a) Consider the Dirichlet problem (4.18), (4.19). Then if the solution
exists, it is unique.
(b) The Dirichlet problem (4.18), (4.19) has continuous dependence on the data.

Proof: (a) Suppose that u1 , u2 are two solutions, so u = u1 − u2 satisfies

uxx + uyy = 0 in D (4.21)

u = 0 on ∂D. (4.22)

By (4.21) the maximum and minimum of u occur on ∂D, so by (4.22) u ≤ 0 and u ≥ 0

in D. Thus u = 0 in D as required.
(b) We have to prove that for all  > 0 there exists δ > 0 such that if ui , i = 1, 2 are
solutions with boundary data gi , then

sup |g1 (x, y) − g2 (x, y)| < δ ⇒ sup |u1 (x, y) − u2 (x, y)| < .
(x,y)∈∂D (x,y)∈D

75
By linearity u = u1 − u2 satisfies

uxx + uyy = 0 in D (4.23)

u = g1 − g2 , on ∂D. (4.24)

We now apply the maximum principle to see that u ≤ max(x,y)∈∂D (g1 −g2 ), and applying
the same result to −u, −u ≤ max(x,y)∈∂D −(g1 − g2 ).
Hence in D

|u1 − u2 | ≤ max |g1 − g2 |, (4.25)

(x,y)∈∂D

so we may take δ = .

4.4.2 The heat equation

In parabolic PDEs it is usually the case that one independent variable represents
time, so we now use x and t as independent variables instead of x and y. The
normal form for second-order parabolic equations is

uxx = F (x, t, u, ut , ux )

and specific examples include the inhomogeneous heat equation, often called the
diffusion equation:
ut = uxx + f (x, t).
and the reaction-diffusion equation

ut = uxx + f (x, t, u),

Well posed boundary data: Typical boundary data for a diffusion equation are
to give an initial condition for u at t = 0 and one boundary condition on each of
two curves C1 and C2 in the (x, t)-plane that do not meet and are nowhere parallel
to the x-axis.
For example: The inhomogeneous heat equation

ut = uxx + f (x, t)

is a simple model for the temperature u(x, t) in a uniform bar of conductive ma-
terial, with heat source f (x, t), where x is position and t is time. Suppose the bar
is of length L, its initial temperature is given via u0 (x), and its ends are kept at
zero temperature. Then the initial and boundary conditions are

u = u0 (x) at t = 0, u = 0 at x = 0; u = 0 at x = L.

76
If, instead of being held at constant temperature, an end is insulated, then the
Dirichlet boundary condition, u = 0, there is replaced by the Neumann boundary
condition, ux = 0. Alternatively , the boundary conditions at x = 0 and x = L
may, in general, be replaced by conditions at moving boundaries, say x = x1 (t)
and x = x2 (t).
Theorem 4.3. (The Maximum principle for the heat equation) Suppose that u(x, t)
satisfies
ut − uxx ≤ 0 (4.26)
in a region Dτ bounded by the lines t = 0, t = τ > 0, and two non-intersecting
smooth curves C1 and C2 that are nowhere parallel to the x-axis. Suppose also that
f ≤ 0 in Dτ . Then u takes its maximum value either on t = 0 or on one of the
curves C1 or C2 .

Proof:
The proof is similar to that for Poisson’s equation.
We first observe that since u is a continuous function on a compact set D̄τ it will
achieve its maximum on D̄τ .
Suppose first that ut − uxx < 0 in Dτ . At an internal maximum inside Dτ , u must
satisfy
ux = ut = 0 uxx ≤ 0, (utt ≤ 0).
On the other hand, if u has a maximum at a point on t = τ , then there it must
satisfy
ux = 0, ut ≥ 0, uxx ≤ 0.
With ut − uxx assumed to be strictly negative, both of these lead to contradictions,
and it follows that u must take its maximum value somewhere on ∂Dτ but not on
t = τ . We are done.
Suppose now that ut − uxx ≤ 0, then define

v(x, t) = u(x, t) + x2 ,
2
where  is a positive constant. Then v satisfies
vt − vxx = ut − uxx −  < 0
in Dτ . So by the earlier step v takes its maximum value on ∂Dτ but not on t = τ .
Now if the maximum value of u over these three portions of ∂Dτ is M , and the
maximum value of |x| on C1 and C2 is L, then
L2 
u≤v≤ + M.
2

77
Now we let  → 0 and conclude that u ≤ M , i.e. u takes its maximum value on
∂Dτ but not on t = τ
If ut − uxx ≥ 0 in Dτ , then a similar argument shows that u attains its minimum
value on ∂Dτ but not on t = τ . Thus, for the homogeneous equation (the heat
equation) u attains both its maximum and its minimum values on ∂Dτ but not on
t = τ.
Remark: Physical interpretation: In a rod with no heat sources the hottest and
the coldest spot will occur either initially or at an end. (Because heat flows from
a hotter area to a colder area.)
Corollary 4.4. Consider the IBVP consisting of (4.26) in Dτ with u given on
∂Dτ \{t = τ }. Then if the solution exists, it is unique and depends continuously
on the initial data.

Proof: As for Poisson’s equation (see problem sheet).

Sketch: Suppose that ui (i = 1, 2) are solutions of (4.26) in Dτ with data
ui = gi on C1 , ui = hi on C2 , ui (x, 0) = ki (x) for x between C1 and C2 .
Let u = u1 − u2 , g = g1 − g2 , h = h1 − h2 , k = k1 − k2 . Then u satisfies (4.26)
with f = 0, and u = g on C1 , u = h on C2 , u(x, 0) = k(x) for x between C1 and
C2 . Thus u and −u take their maximum values on ∂Dτ \{t = τ }. Hence

sup u ≤ max{sup g, sup h, sup k}

Dτ

and
sup −u ≤ max{sup −g, sup −h, sup −k}.
Dτ

So
sup |u| ≤ max{sup |g|, sup |h|, sup |k|},
Dτ

and continuous dependence follows. In particular, if g = h = k = 0 then u = 0

and uniqueness follows.

78
5 Where does this course lead?

The course leads to DEs2, where, among other topics, boundary value problems for
ODEs are discussed. Further discussion of differential equations comes in the Part
B courses ‘Non-linear Systems’ and ‘Applied Partial Differential Equations’. The
use of abstract methods such as the Contraction Mapping Theorem to investigate
the solutions of differential equations is taken further in various C4 courses, which
require some knowledge of Banach and Hilbert spaces.
The techniques taught in DEs1 and DEs2 will be useful in various applied maths
courses such as the Part A short course ‘Modelling in Mathematical Biology’ and
the Part B courses ‘Mathematical Ecology and Biology’, ‘Viscous Flow’ and ‘Waves
and Compressible Flow’.

5.1 Section 1

Fixed point results such as the CMT provide very powerful methods for proving
existence of solution for ODEs and PDEs. For PDEs we have to work in Banach
spaces of functions rather than Rn . For example for parabolic equations such as
the reaction-diffusion equation:

ut = uxx + f (t, u),

with suitable boundary data, our proof of Picard’s theorem can be extended to
prove local existence (in t), provided f is continuous and is Lipschitz in u, where
for each time t the solution u lives within a Banach space of x-dependent functions
(which is infinite dimensional space) rather than the Euclidean n-dimensional space
that we considered previously for ODEs. The technical details of this require
methods from Functional Analysis as covered in the courses B4,1, B4.2 and C4.1
courses, but the basic ideas are just the same - there is just more to do, because
more can go wrong!
An example of a reaction diffusion equations which occurs in applications is Fisher’s
equation:
ut = uxx + u(1 − u),
Another example of a second order parabolic equation is the Black-Scholes equation
in mathematical finance.
Existence theorems play an important role in the theory of PDEs, which is a large
and active field of current research, and you will be able to learn more about this
in the Part C courses on Functional Analytic methods of PDEs and on Fixed Point
Methods for Nonlinear PDEs.

79
5.2 Section 2

Phase plane analysis is a very important tool in mathematical modelling. It will be

used, for example, in the Part A short course ‘Modelling in Mathematical Biology’
and the Part B course, ‘Mathematical Ecology and Biology’.
The theory will be taken further in the Part B course, ‘Nonlinear Systems’.

5.3 Section 3

We have considered only semi-linear first order PDEs. Similar methods can be
extended to quasi-linear and fully non-linear equations. In these cases the char-
acteristic equations are generally more difficult to solve. Such equations allow for
the formation of shocks - see Part B ‘Applied PDEs’.
These first order PDEs model many physical processes, including particularly con-
servation laws, and will appear in many of the modelling courses in Part B and
beyond.
There are other methods for producing explicit solutions of PDEs:
Transform methods are very useful for linear PDEs- see the Part A short course
in HT.
Similarity solutions can be used for both linear or non-linear PDEs - see Part B
course ‘Applied PDEs’ - and involve reducing the PDE to an ODE in a ’similarity
variable’ involving both independent variables in the PDE.

5.4 Section 4

As we have already observed, proving the existence of solution, even of semi-linear

second order PDEs is challenging. The different types of equation demand different
approaches. For example:
Semi-linear hyperbolic equations with the solution u and its normal derivative
prescribed on a given initial curve: One method to prove existence of solutions
proceeds by showing that on initial curves other than characteristic curves we can
find all the derivatives of u and that if all coefficients etc in the equation are very
smooth (analytic) the solution is given by a power series near the initial curve.
This is a version of the Cauchy-Kowalevski theorem.
Elliptic equations are treated in the Part C course ‘Functional analytic methods
for PDEs’.

80
Acknowledgements

I am grateful to colleagues for allowing me to use and adapt their notes from earlier
courses.

81