Download as pdf or txt
Download as pdf or txt
You are on page 1of 73

Advanced Methods of

Applied Mathematics
Spring 2023

Contents

1 Asymptotic Methods 3
1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Asymptotic evaluation of integrals 7


2.1 Integration by parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Laplace’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Watson’s lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Method of steepest descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Method of stationary phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Ordinary differential equations 25


3.1 Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Regular Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Singular Perturbation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 WKB theory (geometric optics, ray theory) 45

5 Partial differential equations 51


5.1 First-Order Nonlinear Hyperbolic PDEs . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1.1 Breaking: shocks and shock fitting . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1.2 Conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1.3 Traffic Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.4 Erosion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Higher Order Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . 70
5.2.1 Derivation of the Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.2 Derivation of the Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . 72

1
2 CONTENTS
Chapter 1

Asymptotic Methods

1.1 Notation
Big O: f (z) = O(g(z)) as z → z0 if f (z)/g(z) is bounded as z → z0 , or |f (z)| ⩽ A|g(z)| as z → z0 ,
where A is a constant.
Small o: f (z) = o(g(z)) as z → z0 if f (z)/g(z) → 0 as z → z0 . Alternatively: f (z) ≪ g(z) as
z → z0 .
Asymptotic: f (z) ∼ g(z) as z → z0 if f (z)/g(z) → 1 as z → z0 .
sin 2z 2 = o(z), sin 2z 2 ∼ 2z 2 ,
 
sin (2z) = O(z), as z → 0.

Asymptotic series: The sequence {φn (z)} is an asymptotic sequence if for all n we have φn+1 (z) ≪
φn (z) (or φn+1 = o(φn )) as z → z0 .
X∞
an φn (z) is an asymptotic expansion of f (z) as z → z0 to M terms if
n=1

N
X
f (z) − an φn (z) ≪ φN (z) as z → z0
n=1

for each N ⩽ M . This is usually written


M
X
f (z) ∼ an φn (z) as z → z0 .
n=1

Examples: {(z − 1)n } as z → 1, {z −n } as z → ∞.


The main example is z = ε, z0 = 0 and ϕn (ε) = εn . The expansion is uniform if

X
f (x, ε) ∼ an (x)εn as ε → 0,
n=0

uniformly on x ∈ [a, b] for example.


If the expansion breaks down somewhere, such as if some of the coefficients go to zero or
infinity, then the expansion is non-uniform.
An example of a non-uniform expansion:

x X εn
∼ (−1)n n as ε→0
x+ε x
n=0

is not valid for small x and breaks down when x = O(ε). (When x = O(ε), terms are not of
increasingly smaller order.)

3
4 CHAPTER 1. ASYMPTOTIC METHODS

Perturbation methods: Find approximate solutions to problems for small values of some pa-
rameter ε when the solution for ε = 0 is known. Perturbation problems are either regular or
singular.
Regular problem: No difficulties in determining asymptotic expansions and nothing unusual
happens, much like a Taylor series.
Singular problem: Not regular. Fundamental change in the character of the solution from ε = 0
to ε > 0. For example non-uniform expansion, no solution for ε = 0.

Simple example:
x2 − x + ε = 0, |ε| ≪ 1 (1.1)
For ε small (taking ε = 0), we just solve: x2 − x = 0, which factorises to give x(x − 1) = 0, such
that the solutions are x = 0, 1.
Proceeding with an asymptotic expansion of x:

X
x= an εn = a0 + a1 ε + a2 ε2 + . . .
n=0

Expand up to and including quadratic term


2
x2 = a0 + a1 ε + a2 ε2 + . . .
2
= a20 + 2a0 (a1 ε + a2 ε2 + . . . ) + a1 ε + a2 ε2 + . . .
= a20 + 2a0 a1 ε + (2a0 a2 + a21 )ε2 + . . .
Substituting these expansions into (1.1):
a20 + 2a0 a1 ε + (2a0 a2 + a21 )ε2 − a0 − a1 ε − a2 ε2 + ε + . . . = 0.
Matching coefficients of each order of εn :
ε0 : a20 − a0 = 0 =⇒ a0 = 0, 1
1
ε1 : 2a0 a1 − a1 + 1 = 0 =⇒ a1 = if a0 ̸= 21
1 − 2a0
a21 1
ε2 : 2a0 a2 + a21 − a2 = 0 =⇒ a2 = =
1 − 2a0 (1 − 2a0 )3
For a0 = 0:
x ∼ ε + ε2 + . . . (1.2)
For a0 = 1:
x ∼ 1 − ε − ε2 + . . . (1.3)
Now let us solve (1.1) by applying the quadratic formula:

2 1± 1 − 4ε
x − x + ε = 0 =⇒ x = . (1.4)
2
Binomial expansion (1 + x)n = 1 + nx + n(n−1) 2
2! x +
n(n−1)(n−2) 3
3! x + ···.
Applying a binomial expansion of the square root using that (1 − a)1/2 = 1 − 21 a − 18 a2 + . . . ,
with a = 4ε we see from (1.4) that
1 ± (1 − 2ε − 2ε2 + . . . )
x± = ,
2
which gives the two solutions
x + = 1 − ε − ε2 + . . . , x− = ε + ε2 + . . . (1.5)
in agreement with the regular perturbation expansions (1.2) and (1.3).
1.1. NOTATION 5

Example:

εx2 + x − 1 = 0, |ε| ≪ 1 (1.6)


As before, for ε small (taking ε = 0) we just solve: x − 1 = 0, which gives that x = 1. Now we find
that we have only one solution for a quadratic equation! This is a singular perturbation problem.
The solution obtained by applying the quadratic formula is

−1 ± 1 + 4ε
x= .

Applying a binomial expansion of the square root using that (1 + a)1/2 = 1 + 12 a − 18 a2 + . . . , with
a = 4ε we see
−1 ± (1 + 2ε − 2ε2 + . . . )
x± = (1.7)

to give the two solutions

1
x+ = 1 − ε + . . . , x− = − − 1 + ε + . . . (1.8)
ε
The first solution is a regular perturbation, as in the previous example, while the second solution
x− = −ε−1 − 1 + ε + . . . is singular as ε → 0. As ε → 0, one of the roots is sent to infinity.
Note that the regular root can be found as in the previous example, so we will only consider
the singular root.
Let us consider the singular solution. We note that its leading order is O(1/ε), and so we
introduce a new variable by letting x = X/εα . Then (1.6) becomes

=ε0
z}|{
ε1−2α X 2 + ε−α X − 1 = 0. (1.9)

We apply the principle of dominant balance to determine the value of α. Balancing the exponents
of ε for the first two terms:
1 − 2α = −α =⇒ α = 1,
which gives x = X/ε and will recover the singular O(1/ε) solution.
Alternatively we can consider the balance of the second and third terms, given by α = 0 in this
case, which will recover the regular solution.

Taking an asymptotic expansion of the form



X
X= an εn = a0 + a1 ε + a2 ε2 + . . . ,
n=0

gives
X 2 = a20 + 2a0 a1 ε + (2a0 a2 + a21 )ε2 + . . .
To find the singular solution we substitute these expressions into (1.9) with α = 1, i.e. into

ε−1 X 2 + ε−1 X − 1 = 0,

to give
ε−1 a20 + 2a0 a1 + (2a0 a2 + a21 )ε + ε−1 a0 + a1 + a2 ε − 1 = 0.
Solving this by collecting terms at each order of ε, we have at O(ε−1 ):

a20 + a0 = 0,
=⇒ a0 (a0 + 1) = 0.
6 CHAPTER 1. ASYMPTOTIC METHODS

There are two roots: a0 = 0, which corresponds to the regular solution (ε−1 term vanishes), and
a0 = −1 which corresponds to the singular solution. Thus take the solution a0 = −1.
At O(ε0 ):

2a0 a1 + a1 − 1 = 0,
1
=⇒ a1 = = −1.
2a0 + 1

Finally at O(ε1 ):

2a0 a2 + a2 + a21 = 0,
−a21
=⇒ a2 = = 1.
2a0 + 1

Therefore we find that the singular solution is given by

X ∼ −1 − ε + ε2 + . . . ,

or in the original variables,


1
x ∼ − − 1 + ε + ...
ε
This agrees with the (exact) solution from the quadratic formula.
Chapter 2

Asymptotic evaluation of integrals

2.1 Integration by parts


Exponential integral

The exponential integral is a special function defined as



e−t
Z
E1 (x) = dt,
x t

which is divergent as x → 0.
We can determine the value of this integral for large values of x, x ≫ 1, using integration by
parts:

∞ ∞ −t
e−t

e
Z
E1 (x) = − − dt
t x x t2
 −t ∞ Z ∞ −t
e−x e 2e
= + 2
+ dt
x t x x t3
 −t ∞ Z ∞
e−x e−x e 2 · 3e−t
= − 2 − 2 3 − dt
x x t x x t4
∞ Z ∞
e−x e−x 2e−x 2 · 3e−t 2 · 3 · 4e−t

= − 2 + 3 + + dt
x x x t4 x x t5
∞ Z ∞
e−x e−x 2e−x 2 · 3e−x 2 · 3 · 4e−t 2 · 3 · 4 · 5e−t

= − 2 + 3 − − − dt
x x x x4 t5 x x t6
2·3 2·3·4
 
−x 1 1 2
=e − + − 4 + + ...
x x2 x3 x x5

The pattern is clear, so that higher order terms in the asymptotic series can be deduced.
This is an asymptotic series, but it is divergent due to the (n − 1)! factor, so it is useful for a
small number of terms with x large. If we keep adding many terms then we get worse results. For
a finite number of terms, the accuracy increases as x increases.

Gamma function

The gamma function is a special function defined as


Z ∞
Γ(x) = e−t tx−1 dt.
0

The gamma function is the generalisation of the factorial x! to non-integer x.

7
8 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

The basic property of the gamma function from which it can be calculated is Γ(x + 1) = xΓ(x).
This is shown by integration by parts.
Z ∞
∞
Γ(x) = −e−t tx−1 0 + (x − 1) e−t tx−2 dt

0
Z ∞
= (x − 1) e−t t(x−1)−1 dt = (x − 1)Γ(x − 1).
0

The gamma function can then be calculated for x a positive integer using this recurrence relation.
Z ∞
∞
e−t t0 dt = −e−t 0 = 1.

Γ(1) =
0

Then for positive integer n,

Γ(n) = (n − 1)Γ(n − 1)
= (n − 1)(n − 2)Γ(n − 2)
= (n − 1)(n − 2)(n − 3)Γ(n − 3)
= ...
= (n − 1)(n − 2)(n − 3)(n − 4) . . . 3.2.1
= (n − 1)!

as Γ(1) = 1.
The gamma function generalises factorials for non-integer n. It can be evaluated for x a positive
integer and x a half positive integer. For other values of x, mathematical tables or numerical
methods are needed.

Z ∞
Γ( 12 ) = e−t t−1/2 dt
0
Z ∞
2
= e−y y −1 2y dy, t = y2
0
Z ∞
2
= 2 e−y dy
0

π
= 2×
√ 2
= π.

From this the recurrence relation can be used to evaluate Γ(n + 12 ), as follows

1 1 π
Γ( 32 ) = Γ( ) =
2 2 2√ √
3 3 3 π 3 π
Γ( 52 ) = Γ( ) = =
2 2 2 2√ 4√
5 5 53 π 15 π
Γ( 72 ) = Γ( ) = =
2 2 2 4 8
..
.

Incomplete gamma function


The lower incomplete gamma function is given by
Z x
γ(a, x) = e−t ta−1 dt.
0
2.2. LAPLACE’S METHOD 9

We determine the large x behaviour of this function using integration by parts. First, introduce
the upper incomplete gamma function

Z ∞
Γ(a, x) = Γ(a) − γ(a, x) = e−t ta−1 dt,
x

R∞
such that γ(a, x) = Γ(a) − x e−t ta−1 dt. Then, integrating by parts:

Z ∞ ∞
Z ∞
e−t ta−1 dt = −e−t ta−1 x + (a − 1) e−t ta−2 dt

x x
∞
Z ∞
−x a−1
e−t ta−2 xe−t ta−3 dt

=e x − (a − 1) + (a − 1)(a − 2)
x
∞
= e−x xa−1 + (a − 1)e−x xa−2 − (a − 1)(a − 2) e−t ta−3 x

Z ∞
+ (a − 1)(a − 2)(a − 3) e−t ta−4 dt
x
− − 1)(a − 2) (a − 1)(a − 2)(a − 3)
 
−x a−1 (a 1) (a
=e x 1+ + + + ... ,
x x2 x3

which is again a divergent asymptotic expansion as x → ∞.

2.2 Laplace’s Method

Let
Z ∞
ψ(x) = e−x cosh θ dθ.
−∞

We want to find an asymptotic expansion of this integral for large x, x ≫ 1. As x is large, we note
that the integrand rapidly decays as θ increases from 0. The only non-negligible contribution to
the integral then comes from near θ = 0.
2
Taylor series: near θ = 0, cosh θ = 1 + θ2! + . . . . Good for small values of θ. Large θ? Does not
matter as integrand is negligible.
Then
Z ∞
2 /2)
ψ(x) ∼ e−x(1+θ dθ
−∞

r r
2 2
Z
−x −xθ2 /2
=e e dθ, let θ = t, dθ = dt
−∞ x x
r Z ∞ r
−x 2 −t2 −x 2π
=e e dt = e ,
x −∞ x

R∞ −t2 √
as −∞ e dt = π.
In general
∞ √
π
Z
−at2
e dt = √ ,
−∞ a

which is important to remember for later work.


How about higher order terms in the asymptotic expansion? To mimic the local Taylor series
10 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

at θ = 0, we set cosh θ = 1 + t2 . Then

sinh θ dθ = 2t dt
2t dt
dθ = p
cosh2 θ − 1
2t dt
dθ = p
(1 + t2 )2 − 1
2t dt
dθ = √
1 + 2t2 + t4 − 1
2t dt
dθ = √
t 2 + t2

2 dt
dθ = q .
1 + 12 t2

Then
2 2
√ −x
Z ∞
e−xt √ −x Z ∞
e−xt
ψ(x) = 2e 1/2 dt = 2 2e 1/2 dt
−∞ 1 + 12 t2 0 1 + 21 t2
by symmetry.
We can expand the denominator as a binomial series so that

√ ∞  
1 3 5 6
Z
2
ψ(x) = 2 2e−x e−xt 1 − t2 + t4 − t + ... dt.
0 4 32 128
All these integrals can be expressed in terms of gamma functions.
Z ∞ Z ∞
2n −xt2 1 1
n− 2 −y
t e dt = 1 y e dy, (y = xt2 )
0 2xn+ 2 0
1 1
= x−n− 2 Γ(n + 12 ).
2


n=0: x−1/2 Γ( 21 ) = πx−1/2

−3/2 −3/2 1 π −3/2
n=1: x Γ( 32 ) =x Γ( 1 )
x =
2 2 2
3 3 √ −5/2
n=2: x−5/2 Γ( 52 ) = x−5/2 Γ( 32 ) = πx
2 4
5 15 √ −7/2
n=3: x−7/2 Γ( 27 ) = x−7/2 Γ( 52 ) = πx
2 8

So, r  
2π −x 1 9 75
ψ(x) ∼ e 1− + − + . . . as x → ∞.
x 8x 128x2 1024x3
This process of using an exact transformation, cosh θ = 1 + t2 here, to generate higher order
terms in the asymptotic series is called Watson’s Lemma. The transformation used is deduced from
the local Taylor series of the exponent, as for this example.

Gamma function
Z ∞
Γ(x) = e−t tx−1 dt
0

Let us consider using Laplace’s method to explore the large x behaviour. Laplace’s method: The
2.3. WATSON’S LEMMA 11

integral is not dominated by the behaviour near t = 0.


Z ∞
Γ(x + 1) = e−t+x log t dt
0

Let f (t) = x log t − t, f ′ (t) = x/t − 1 = 0 at t = x, and f ′′ (t) = −x/t2 , so f has a maximum at
t = x. The integral is then dominated around t = x for x ≫ 1.
So let t = xs. Then Z ∞
Γ(x + 1) = xx+1 e−x(s−log s) ds.
0

Now let’s consider when x is large, x ≫ 1. Letting

1 1
f (s) = −s + log s, f ′ (s) = −1 + =0 at s = 1, f ′′ (s) = − ,
s s2

we see that there is a maximum at s = 1. Applying a Taylor series to the exponent about s = 1,

(1 − s)2
s − log s = s − log(1 − (1 − s)) = 1 + + ...
2

Note that log(1 + x) = x − 12 x2 + 31 x3 + . . .


Then
Z ∞ 1 2
Γ(x + 1) ∼ x x+1
e−x− 2 x(1−s) ds
0
Z ∞
1 2
x+1 −x
=x e e− 2 x(1−s) ds.
0

We cannot evaluate this integral, set y = s − 1 and the lower limit goes to y = −1. However,
the integrand is dominated for x large near s = 1. We can then write
Z ∞ 1 2
Γ(x + 1) ∼ x x+1 −x
e e− 2 x(1−s) ds
−∞

r

Z
1
x+1 −x − 2 xy 2 x+1 −x
=x e e dy = x e ,
−∞ x

as the integrand is negligible at s = 0, so that anything away from s = 1 makes no contribution to


the integral.
We note that this expansion for large x leads to the standard, well known Stirling’s formula
(1730):
r
n+1 −n 2π
Γ(n + 1) = n! ∼ n e as n → ∞.
n

2.3 Watson’s lemma


The extension of Laplace’s method to include higher order terms in the asymptotic expansion is
Watson’s Lemma, named after G.N. Watson who was at the University of Edinburgh in the early
1900s. Watson’s Lemma is performed as an exact substitution in the exponent of the exponential,
as in the example above. This substitution is deduced from the Taylor series expansion of the
exponent about the dominant point. The use of Watson’s Lemma is as in the above example for
ψ(x).
12 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

2.4 Method of steepest descent


Laplace’s method is used when the integral is dominated by a point on the contour of integration.
If the dominant point is not on the contour of integration, then the extension of Laplace’s Method
is called the method of steepest descent. Note that the dominant point can lie in the complex plane
and so does not have to be real.
The Laplace transform and its inverse are defined as
∞ γ+i∞
1
Z Z
f¯(s) = f (t)e −st
dt, f (t) = f¯(s)est ds.
0 2πi γ−i∞

The contour of integration passes through γ on the real axis, is parallel to the imaginary axis and
lies to the right of all singularities of f¯.

s-plane

To illustrate the method of steepest descent, we consider the integral

1 √ dz
Z
ϕ(x) = ex(z− z) .
2πi C z

The integrand has a simple pole at z = 0 (and a branch point at z = 0 with a branch cut along the
negative real axis). The integral itself is related to a modified Bessel function.

We wish to look at the behaviour of this integral as x → ∞. Let f (z) = z − z. As with
Laplace’s method, the integral is dominated by contributions from the stationary points of the
exponent. The stationary point is given by
1
f ′ (z) = 1 − √ = 0
2 z

at z = z0 = 14 .
Note that analytic functions cannot have minima or maxima, only saddle points, as follows
from the Cauchy-Riemann equations, f (z) = u + iv,

ux = vy and vx = −uy .

Then
uxx + uyy = vxy + uyy = −uyy + uyy = 0.
2.4. METHOD OF STEEPEST DESCENT 13

Then for the Hessian test applied to the real part of f (z), u

H = uxx uyy − u2xy = −u2xx − u2xy < 0.

Hence, any stationary point of u is a saddle point. Similarly for v.


For this reason, the method of steepest descent is sometimes known as the saddle point method.
Thus we note that the stationary point f (z0 ) is a saddle point. The contours of |f | go down in
one direction and go up in the orthogonal direction around a saddle point. We want the contour
going down as this is the path on which the integrand decreases most rapidly away from the saddle
point, the path of steepest descent.
The idea behind the method of steepest descent is to deform the integration contour C so that
f (z) decreases as rapidly as possible along it, the path of steepest descent. We do this by applying
Laplace’s method which involves a Taylor expansion of the exponent f (z) about the stationary
point z0 :
f ′′ (z0 )
f (z) = f (z0 ) + f ′ (z0 )(z − z0 ) + (z − z0 )2 + . . .
2
2
= − 41 + z − 14 + . . .

as f ′ (z0 ) = 0 and f ′′ (z0 ) = 2. To make the exponential exf decrease away from the saddle point,
we set z = 41 + iy, with −∞ < y < ∞, so that

f (y) = − 41 − y 2 + . . .

This is just using i2 = −1. In this manner we get exponential decay away from the saddle point.
Then
Z ∞ −x/4−xy2 +...
1 xf (z) dz 1 e
Z
ϕ(x) = e = 1 i dy
2πi C z 2πi −∞ 4 + iy
2
e−x/4 ∞ e−xy
Z
∼ dy (as the dominant contribution is around y = 0),
2π −∞ 14
2 −x/4 ∞ −xy2
Z
= e e dy
π −∞
r
2 −x/4 π
= e
π x
2 −x/4
=√ e .
πx
This is just a rough guide to how the method of steepest descent is used. Now we discuss the
method in more detail. The key issue is how to find the steepest descent path.
To find the steepest descent path, we need some properties of the gradient vector

1. ∇f is orthogonal to contours of f , curves on which f is constant.

2. ∇f is in the direction of most rapid increase of f at a point.

3. −∇f is in the direction of most rapid decrease of f at a point.

Analytic function f (z) = u + iv. Cauchy-Riemann equations


∂u ∂v ∂u ∂v
= , =− .
∂x ∂y ∂y ∂x
We note that the gradients of the real and imaginary parts are orthogonal as

∇u · ∇v = (ux , uy ) · (vx , vy )
= −ux uy + uy ux = 0.
14 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

The direction of most rapid decrease of u, the direction of steepest descent, is in the direction
of −∇u, which is orthogonal to ∇v. But ∇v is orthogonal to curves v = constant. So the
direction −∇u is in the direction of the contours of v, the imaginary part of f . Hence, the curve
Im (f ) = const. through the saddle point is the steepest descent contour for f . This is the key
point of the method of steepest descent and is how the contour of steepest descent is calculated.

• The contour of steepest descent is the curve on which Im (f ) is constant and takes its value
at the saddle point, the contour of steepest descent is given by Im (f (z)) = Im (f (z0 )).

Taking f (z) = z− z for example, at the saddle point z0 = 41 , f ( 14 ) = − 41 , so that Im (f (z0 )) = 0.
In this case it is simplest to calculate the contour of steepest descent using the polar form z = reiθ
such that

f (r, θ) = reiθ − reiθ/2
√ √
= (r cos θ − r cos 2θ ) + i(r sin θ − r sin 2θ ).

Then the steepest descent contour is Im (f ) = 0 and is found by solving the equation r sin θ =

r sin 2θ , which is equivalent to

2 r sin 2θ cos 2θ = sin 2θ ,

which has the solutions sin 2θ = 0 and 2 r cos 2θ = 1.
The first has the solutions θ = 0, 2π. This is the steepest ascent contour as on the real axis

f = x − x → ∞ as x → ∞.
The second solution is found with some algebra:

2 r cos 2θ = 1
4r cos2 θ
2 =1
r(1 + cos θ) = 1
2 using 2 cos2 α = 1 + cos 2α
1
r+x= 2
p
1
x2 + y 2 = 2 −x
=⇒ x = 1
4 − y2, a parabola.
2.4. METHOD OF STEEPEST DESCENT 15

So C0 : x = 14 − y 2 is the path of steepest descent. We want to deform the original contour


C → C0 . There is a pole at z = 0 but we won’t pass through it while deforming the contour so we
can ignore it. This is based on Cauchy’s Theorem.
In the previous rough use of the method of steepest descent, we have near z = 14 , f = − 14 +

(z − 14 )2 + . . . = − 41 − y 2 + . . .. As for Watson’s Lemma, we then set z − z = − 14 − τ 2 , with τ
real, so that
√ √ 2 √
z − z + 14 + τ 2 = 0 =⇒ z − z + 14 + τ 2 = 0
 

√ 1 ± 1 − 1 − 4τ 2
=⇒ z =
2
√ 1 ± 2iτ
=⇒ z =
2
√ 2
1
=⇒ z = 2 + iτ so z = 12 + iτ
where we chose the positive root (but the negative one would’ve worked as well). Recall that
2 √
f = − 41 + z − 14 + . . . When z = 14 , τ = 0. τ is real as Im (z − z) is constant on the steepest
2
descent path. Then, since we have z = 12 + iτ =⇒ dz = 2 21 + iτ i dτ ,


ex(− 4 −τ )
1 2

1
Z
1

ϕ= 2 · 2 2 + iτ i dτ
2πi −∞ 1
+ iτ
2
∞ 2
2 1 e−xτ
Z
= e− 4 x dτ
π −∞ 1 + 2iτ
We cannot evaluate this integral so we will need to use Watson’s Lemma: first we expand (1+2iτ )−1
as a geometric series. Recall that

−1
X
(1 + x) = (−1)n xn = 1 − x + x2 − x3 + x4 + . . .
n=0

Hence
 
∞ 2 ∞ 0 0
e−xτ
Z Z
2
dτ ∼ e−xτ 1 − |{z}
2iτ 3
7 − 4τ 2 − 8iτ  4
|{z} + 16τ + . . . dτ
7 
−∞ 1 + 2iτ −∞  
odd odd
Z ∞
2
= e−xτ (1 − 4τ 2 + 16τ 4 + . . .) dτ
−∞
Z ∞
2
=2 e−xτ (1 − 4τ 2 + 16τ 4 + . . .) dτ
0
Since for odd powers of τ we get an integral of an odd function over a symmetric interval, those
terms will go to 0 by definition. We are left with the even powers of τ , and hence a series of even
functions over a symmetric intervalr so we can change the lower limit to 0 and multiply the integral
y
by 2. Letting y = xτ 2 =⇒ τ = , dy = 2xτ dτ we have
x
Z ∞
4y 16y 2
 
2 −1x −y dy
ϕ∼ e 4 ·2 e 1− + 2 + ... √
π 0 x x 2 xy
Z ∞  
2 −1x −y −1/2 4 1/2 16 3/2
= √ e 4 e y − y + 2 y + . . . dy
π x 0 x x
 
2 −1x 1
 4 3
 16 5

= √ e 4 · Γ 2 − · Γ 2 + 2 · Γ 2 + ...
π x x x

√ π 16 3 1 √
 
2 −1x 4
= √ e 4 · π− · + 2· · π + ...
π x x 2 x 2 2
 
2 1 2 12
= √ e− 4 x · 1 − + 2 + . . .
πx x x
16 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

This example shows us how the method of steepest descent and Watson’s Lemma can be combined
in order to find higher order terms in the asymptotic expansion of an integral.

Example:
Find the first term in the asymptotic expansion of
Z x(z−√z)
1 e
ψ= dz, x≫1 (2.1)
2πi C z 2 + ω 2

where C is the contour in the diagram shown below. Again we have f (z) = z − z, so our contour
of steepest descent is the same, however now we have simple poles at z 2 + ω 2 = 0 =⇒ z = ±iω.
This gives us two cases:
Case (i): When ω < 1/2 we won’t run into any issues, because when we deform the original
contour C to the steepest descent contour (the parabola) we will not pass through any poles, so
we can proceed as usual.

z-plane

1
2
C

w C0

1
4
−w
1
x= − y2
4

1

2

Case (ii): When ω > 1/2 however we will pass through both poles while deforming the contour to
the steepest descent contour, therefore we will need to account for this contribution to the integral
by adding the residues at the poles.
Let us do the calculations for both cases:
Case (i): This is the same as the previous example, with our saddle point at z = 41 , so
√ 2
f (z) = z − z ∼ − 41 + z − 14 + . . .
= − 14 − y 2 + . . .
1
where z − 4 = iy. Hence,
1 2

1 e− 4 x−xy
Z
ψ∼ 2 · i dy
2πi −∞ 1
+ iy + ω 2
4
− 14 x ∞ 2
e e−xy
Z
∼ dy
2π −∞ ω 2 + 1/16
1 √
16e− 4 x π
= 2
·√
2π(1 + 16ω ) x
1
8e− 4 x
=√
πx(1 + 16ω 2 )
2.4. METHOD OF STEEPEST DESCENT 17

z-plane

1
2 C

C0

1
4

1
x= − y2
4

1
2

−w

In the second line we were able to ignore the terms iy/2 − y 2 in the denominator because we are
only concerned with the first term in the asymptotic expansion for this question— if we wanted to
find the second or third terms we would have to use Watson’s Lemma and expand the denominator
using the binomial series.
Case (ii): Since we cross the poles while deforming C to C0 we need to add the residues at these
poles to the answer we found in Case (i) (this is how we account for their contribution to the
integral). Let √
ex(z− z )
g= 2 .
z + ω2
Then using the residue theorem for both poles:

Resz=iω g = lim (z − iω)g


z→iω

ex(z− z )
= lim (z − iω) ·
z→iω (z + iω)(z − iω)

ex(iω− iω)
=
2iω
√ iπ
 
x iω− ωe 4
e
=
2iω √ √
1 −x ω
√ iωx− ix√ ω
= ·e 2 e 2
2iω

Resz=−iω g = lim (z + iω)g


z→−iω

ex(z− z )
= lim (z + iω) ·
z→−iω (z + iω)(z − iω)

x(−iω− −iω )
e
=
−2iω

x(−iω− ωe−iπ/4 )
e
=
−2iω
√ √
1 −x√ ω −iωx+ ix√ ω
=− e 2 e 2
2iω
18 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

To get our final answer we need to sum the residues and add it to our previous answer. Therefore,
√  √   √ 
−x ω
√ i ωx− x√ ω −i ωx− x√ ω
e 2 e 2 −e 2
Resz=iω g + Resz=−iω g = ·
ω√ 2i
e−x ω/2  p 
= · sin xω − x ω/2
ω

Hence,

1

8e− 4 x e−x ω/2  p 
ψ∼√ + · sin xω − x ω/2 .
πx(1 + 16ω 2 ) ω

Note that the second term is exponentially small compared to the first term, because ω > 12 .

Hankel’s Integral for the Gamma Function:

1 1
Z
= es s−z ds (2.2)
Γ(z) 2πi C

where C is the contour from s = −iε−∞ to s = iε−∞ around the origin, with ε > 0 (see diagram).

There is a branch cut along the negative real axis due to the s−z = e−z log s term, and there is
a branch point at s = 0. Let s = zt. Then

1 1
Z
= es−z log s ds
Γ(z) 2πi C
1
Z
= ezt−z log z−z log t · z dt
2πi C
1 1−z
Z
= z ez(t−log t) dt
2πi C

1
We can now use the method of steepest descent. Let f (t) = t − log t =⇒ f ′ (t) = 1 − and
t
f ′′ (t) = t12 , thus f ′ (t) = 0 at t = 1 and f ′′ (1) = 1. Taylor expanding around our saddle point at
t = 1:

f ∼ 1 + 12 (t − 1)2 + . . .

First we look at the simplest approximation of this integral, and find the first term in the asymptotic
expansion. Let t − 1 = iy to obtain exponential decay away from t = 1. (t = 1 + iy is the steepest
2.5. METHOD OF STATIONARY PHASE 19

descent path locally near t = 1):

1 1 1−z
Z
z 2
∼ z ez e 2 (t−1) dt
Γ(z) 2πi C
1 1−z z
Z
z 2
= z e e− 2 y · i dy
2πi
ZC
1 z 1−z ∞ − z y2
= e z e 2 dy
2π −∞
r
1 z 1−z 2π
= e z
2π z
1 1
= √ z 2 −z ez

which is the same as Stirling’s Formula.

2.5 Method of stationary phase


We use this method for rapidly oscillating integrals. The value of these kinds of integrals are nearly
0 (since the areas will cancel out) except when the phase of the integral is stationary i.e. at the
stationary points. E.g. sin(xf (θ)) with x ≫ 1, or in complex form as exp(ixf (θ)), which is rapidly
oscillating except where f ′ (θ) = 0. See Figure 2.1, which shows the behaviour of such an integrand
with a stationary point at θ = 0. The method of stationary phase is basically a special case of the
method of steepest descent for oscillatory integrals when the stationary point is on the contour of
integration. Historically, it pre-dates the method of steepest descent.

Figure 2.1: Example of a function which oscillates rapidly except for a region with stationary phase.

Example:
∞   3 
t
Z
G(x) = cos x −t dt, x≫1 (2.3)
0 3
This is a type of Bessel function, and it is a rapidly oscillating integral due to the cosine term.

t3
Let f (t) = − t =⇒ f ′ (t) = t2 − 1, f ′′ (t) = 2t
3
f ′ (t) = 0 at t = ±1 =⇒ t = 1 since 0 < t < ∞
20 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

Then, near the point of stationary phase at t = 1,


1 1 1
f (t) = f (1) + f ′′ (1)(t − 1)2 ∼ − 1 + · 2 · 1 · (t − 1)2 + . . .
2 3 2
2
= − + (t − 1)2 + . . .
3
So, truncating the Taylor series at the t2 term we have
Z ∞   
2 2
G(x) ∼ cos x − + (t − 1) dt
0 3
Z ∞   
2 2
∼ cos x − + (t − 1) dt
−∞ 3
as the integral is dominated near t = 1 and so the integrand is negligible for t < 0.
Z ∞ 
eix(− 3 +(t−1) ) dt
2 2
G(x) ∼ Re
 −∞ Z ∞ 
− 32 ix ix(t−1)2
= Re e e dt
−∞

We need to rotate the contour of integration. Let t−1 = yeiπ/4 =⇒ (t−1)2 = iy 2 and dt = eiπ/4 dy.

Therefore,
 Z ∞ 
− 23 ix −xy 2 iπ/4
G(x) ∼ Re e e e dy
−∞
 √ 
− 23 ix+i π4 π
= Re e ·√
x
√  
π 2 π
= √ cos − x +
x 3 4
√  
π 2 π
= √ cos x− .
x 3 4
Therefore, √  
π 2 π
G ∼ √ cos x−
x 3 4
for x ≫ 1.
We could have also done this another way, using trig identities:
Z ∞   
2 2
G∼ cos x − + (t − 1) dt
−∞ 3
Z ∞     
2 2
x cos x(t − 1)2 + sin x sin x(t − 1)2 dt
 
= cos
−∞ 3 3
 Z ∞  Z ∞
2 2
cos xθ2 dθ + sin sin xθ2 dθ
 
= cos x x
3 −∞ 3 −∞
2.5. METHOD OF STATIONARY PHASE 21

with θ = t − 1. To evaluate the integrals in the above expression, we use the Fresnel integrals
(Problem Set 2): √
Z ∞ Z ∞
2 2 π
cos(at ) dt = sin(at ) dt = √ (2.4)
−∞ −∞ 2a
Thus,
√   √  
π 2 π 2
G ∼ √ · cos x + √ · sin x
2x 3 2x 3
√     
π π  2 π  2
= √ · cos · cos x + sin · sin x
x 4 3 4 3
√  
π 2 π
= √ · cos x−
x 3 4

Using the complex form and rotating the contour of integration is easier as you do not have to
know the Fresnel integrals. In fact, the Fresnel integrals are evaluated in the first place by rotating
the contour.

Example:
The Bessel function of order zero is expressed in integral form as
π/2
2
Z
J0 (x) = cos(x cos(θ)) dθ, x≫1 (2.5)
π 0

J0 (x) is a solution of Bessel’s equation

d2 y dy
x2 2
+x + x2 y = 0,
dx dx
which arises in the solution of PDEs.
This is another example of a highly oscillatory integral (due to the cosine term), so we use the
method of stationary phase. Let

f (θ) = cos(θ) =⇒ f ′ (θ) = − sin(θ), f ′′ (θ) = − cos(θ)


f ′ (θ) = 0 =⇒ θ = 0, π, . . . , f ′′ (0) = −1

The only stationary point in the region of integration is θ = 0, so it’s the only one we consider.
Near θ = 0,
f (θ) ∼ 1 − 12 θ2 + . . .
So for x ≫ 1

2 π/2
Z
cos x 1 − 12 θ2 dθ

J0 (x) ∼
π 0
2 ∞ x 
Z 
∼ cos x − θ2 dθ
π 0 2

We cannot extend the lower limit to −∞ since our stationary point is at θ = 0; if we were to extend
the limit then it would add a contribution to our integral.
Continuing with the integral:
Z ∞ 
2 i(x− x2 θ2 )
J0 (x) ∼ · Re e dθ
π
 0Z ∞ 
2 ix − ix θ 2
= · Re e e 2 dθ
π 0
22 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

To rotate the contour: let θ = ye−iπ/4 =⇒ dθ = e−iπ/4 dy and so


 Z ∞ 
2 xy 2
J0 (x) ∼ · Re eix e− 2 e−iπ/4 dy
π 0
2   1 r 2π
ix− iπ
= · Re e 4 · ·
π 2 x

2  π 
= √ cos x −
πx 4

The Telegrapher’s Equation


The Telegrapher’s equation is the PDE
∂2u ∂2u
= − u.
∂t2 ∂x2
This equation describes the transmission of electrical signals in cables. It was introduced to describe
the propagation of telegraph signals and the propagation of signals in undersea cables. Lord Kelvin
at Glasgow University used it to model the trans-Atlantic cable from Ireland to Newfoundland and
to show why the original cables failed to transmit signals and how to develop a successful cable and
transmission system. In doing so he developed the method of stationary phase. There is a BBC
documentary on this.
It can be solved using Fourier transforms, which we will see in more depth later in the course.
It can be found that the solution is


1
Z h i p
u= ei(k(ω)x−ωt) + e−i(k(ω)x+ωt) dω, with k(ω) = ω 2 − 1. (2.6)
4π −∞

Evaluating this integral is difficult, so we look at the large time behaviour as t → ∞. When t is
x
large, x is large as well as the signal travels at a finite speed. So we can look at t → ∞ with α =
t
fixed. For now we will only look at the right-going wave (the calculations are similar for the other
wave):
Z ∞ Z ∞
1 1
R= ei(k(ω)x−ωt) dω = eit(k(ω)α−ω) dω (2.7)
4π −∞ 4π −∞

Let f (ω) = k(ω)α − ω = α ω 2 − 1 − ω. Then
αω 1
f ′ (ω) = √ − 1 = 0 when α2 ω02 = ω02 − 1 =⇒ ω0 = ± √
ω2 − 1 1 − α2
For real stationary points we need α < 1 i.e. x < t. We only detail the calculation for +ω0 here,
but the calculation is similar for −ω0 , and the final solution (for α < 1) is the sum of the results
for both stationary points. When α < 1 we can use the method of stationary phase (when α > 1
we will use steepest descent).
1
For ω0 = √
1 − α2
α αω 2 −α
f ′′ (ω) = √ − 3/2
=
ω2 −1 (ω 2 − 1) (ω 2 − 1)3/2
3/2
′′ −α 1 − α2
f (ω0 ) = =− <0
ω02 − 1
3/2 α2
r
1 1 α2 1 p
f (ω0 ) = α 2
− 1 − √ = √ −√ = − 1 − α2 .
1−α 1 − α2 1 − α2 1 − α2
2.5. METHOD OF STATIONARY PHASE 23

Hence, the Taylor expanding around ω0 gives us

1 |f ′′ (ω0 )|
f (ω) ∼ f (ω0 ) + f ′′ (ω0 )(ω − ω0 )2 + . . . = f (ω0 ) − (ω − ω0 )2 + . . .
2 2
as f ′′ (ω0 ) < 0. We take the absolute value of f ′′ (ω0 ) to make our calculations easier and to make
sure we have the right sign in the exponent, so that (as usual) we can bring in a Gaussian integral
later on by rotating contour. Then,
Z ∞
1 it ′′ 2
R∼ eitf (ω0 )− 2 |f (ω0 )|(ω−ω0 ) dω
4π −∞
1 itf (ω0 ) ∞ − it |f ′′ (ω0 )|(ω−ω0 )2
Z
= e e 2 dω
4π −∞

Let ω − ω0 = e−iπ/4 y =⇒ (ω − ω0 )2 = e−iπ/2 y 2 = −iy 2 . Hence

1 itf (ω0 ) ∞ −ty2 |f ′′ (ω0 )| −iπ/4


Z
R∼ e e 2 e dy
4π −∞

1 itf (ω0 ) −iπ/4 2π
= e e p
4π ′′
t|f (ω0 )|
1 1 iπ
= ·p · eitf (ω0 )− 4
2 ′′
2πt|f (ω0 )|

1
Now recall α = x/t and ω0 = √ , so
1 − α2
3/2


x2
− 1−
r 3/2
x2 t2 − x2 ′′ t2 − t2 − x2
f (ω0 ) = − 1 − =− , f (ω0 ) = x2
= .
t2 t tx2
t2

Thus,

1 x √
−i t2 −x2 − iπ
R∼ ·√ e 4
2 2π (t2 − x2 )3/4

for t > x i.e. behind the wavefront at x = t.


Ahead of the wavefront (when t < x), i.e. when α > 1 we get complex saddle/stationary points,
hence we need to use the method
√ of steepest descent.
As before, let f (ω) = α ω 2 − 1 − ω.
αω −α
Then f ′ (ω) = √ − 1 and f ′′ (ω) =
ω2 − 1 (ω 2 − 1)3/2
±i
f ′ (ω) = 0 when (α2 − 1)ω02 = −1 =⇒ ω0 = √
α2 − 1
As usual with the method of steepest descent, we Taylor expand around both stationary points
and see which one will give us steepest descent/exponential decay:
r
−1 i
q p
2
Near ω = +ω0 : f (+ω0 ) = α ω0 − 1 − ω0 = α −1− √ = i α2 − 1
2
α −1 α2 − 1
3/2
′′ α −α i α2 − 1
f (ω0 ) = − 3/2 =  3/2 =
ω02 − 1 −1 α2
2
α −1
− 1
p i 3/2
=⇒ f (ω) ∼ i α2 − 1 + 2 α2 − 1 (ω − ω0 )2 + . . .

24 CHAPTER 2. ASYMPTOTIC EVALUATION OF INTEGRALS

Using the same calculations we also have


p i 3/2
Near ω = −ω0 : f (ω) ∼ −i α2 − 1 − 2 α2 − 1 (ω − ω0 )2 + . . .

Hence we choose ω = +ω0 for steepest descent as then eitf decays away from the saddle point.
Near ω = −ω0 , eitf grows away from saddle point, path of steepest ascent. (Note that we actually
have to compute the steepest descent contours to verify these observations.) Therefore, shifting
the contour to ω = +ω0 (and letting y = ω − ω0 ),
Z ∞ √
1 3/2 2
e−t α −1− 2α2 (α −1) (ω−ω0 ) dω
2 t 2
R∼
4π −∞
1 −t√α2 −1 ∞ − t 2 (α2 −1)3/2 y2
Z
= e e 2α dω
4π −∞

1 −t√α2 −1 2πα
= e √
4π t (α2 − 1)3/4
1 x √
− x2 −t2
= ·√ · e
2 2π (x2 − t2 )3/4

We have a wavefront at x = t, with an oscillatory decay behind for x < t and exponential decay
ahead for x > t (in between is a transient region). The transition region is given by Airy functions.

|
x
x = t: wavefront
Chapter 3

Ordinary differential equations

So far we have seen regular perturbation problems, which present no difficulties in determining their
asymptotic expansion via Taylor series (e.g. solutions of x2 − x + ε = 0).
In this chapter, we will study further in depth singular perturbation problems, characterized by
a fundamental change in the behaviour of the solution when ε changes from ε = 0 to ε > 0 (e.g.
solving εx2 − x − 1 = 0).

3.1 Dimensional Analysis


Example:
The example ordinary differential equations in this section arise in applications, which will some-
times be mentioned. These applications include mechanical systems and electrical circuits. This
will be illustrated by the following mechanical mass-spring system. This equation is the basis for
civil engineering analysis of structures such as bridges and buildings.
Let us consider the oscillations of a mass connected to a spring. Let m be the mass of an
object, k the spring constant, B the damping coefficient and Y the distance that the object has
been displaced, as in the following figure:

Figure 3.1: Basic example – damped mass spring system.

Using Newton’s Second Law and Hooke’s Law, the dynamics of the damped mass spring system
can be represented by the ODE
d2 Y dY
m 2
+B + kY = 0 (3.1)
dT dT
with initial conditions
I0
Y (0) = 0, Y ′ (0) = .
m

25
26 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

This represents the mass initially at 0 displacement and is set in motion by being hit at t = 0 to
give it the velocity I0 /m.
Let us first assume small damping, i.e. B ≪ 1.
We use dimensional analysis and rewrite (3.1) in terms of dimensionless variables, which will
allow us to see the relative magnitude of the different terms, and in particular to isolate small
perturbative terms. In addition, setting the equation in non-dimensional form will enable us to
eliminate the number of parameters to the minimum needed and put the equation in its simplest
possible form.
In dimensional analysis, the notation [.] denotes the dimensions of a quantity, what is measured
in, for instance metres for displacement and metre/sec for velocity. It says nothing about the actual
size of the quantity. Let us denote by L the unit of length, τ the unit of time and M the unit of
mass.
The dimensions, or units, of the displacement, time, and mass are given as

[Y ] = L, [T ] = τ and [m] = M.

We now determine the dimensions of the remaining parameters [k],[B] and [I0 ]. We do this by
noting that all terms in equation (3.1) must have the same units so that the equation makes sense.
Then
d2 Y
 
L M
m 2
= [kY ] =⇒ M 2 = [k] L =⇒ [k] = 2 .
dT τ τ
Similarly,
d2 Y
   
dY L L M
B = m =⇒ [B] = M 2 =⇒ [B] =
dT dT 2 τ τ τ
and    
dY I0 L [I0 ] LM
= =⇒ = =⇒ [I0 ] = .
dT m τ M τ
Now, let t and y be the non-dimensional time and space, respectively. In other words these are just
numbers with no units.
First we have that
  r
k M 1 1 2
hmi M
= 2 = 2 =⇒ τ = =⇒ τ = .
m τ M τ k K
p
Thus M/K has dimensions of time and so can be used as a time scale to make a non-dimensional
time. Thus, t = qTM is a non-dimensional time.
K
Second, we have that
LM I0
[I0 ] = =⇒ L = √ .
τ mk

Then I0 / mk has the dimension of length and can be used as a length scale. We can then define
Y
the non-dimensional length variable y = (I /√ mk)
.
0
Hence, we can define two non-dimensional variables t and y as follows;

T Y mk
t = pm , y = .
k
I0

We can now set the ode (3.1) in non-dimensional form. We can rewrite the derivatives in terms of
these non-dimensional variables by employing the chain rule:
dY dY dt I0 dy dt I0 1 dy I0 dy
= =√ =√ p = ,
dT dt dT mk dt dT mk m/k dt m dt

d2 Y I0 d2 y 1 I0 k d2 y
 
d dY dt
= = = 3 .
dT 2 m dt2 m/k 2
p
dt dT dT m 2 dt
3.1. DIMENSIONAL ANALYSIS 27

So upon substitution into (3.1), the ODE becomes,



I0 k d2 y BI0 dy I0
√ + + k√ y = 0.
m dt2 m dt mk
Dividing each term by the coefficient of the highest derivative, we have the dimensionless form of
(3.1) given by

d2 y B dy
2
+√ + y = 0. (3.2)
dt mk dt
In the new variables, the initial conditions are transformed,

Y (0) = 0 =⇒ y(0) = 0.
I0 dy(0) I0 dy(0)
Y ′ (0) = = =⇒ = 1.
m dt m dt
The original equation (3.2) has been set in non-dimensional form as

d2 y dy B
2
+ε + y = 0, ε= √ (3.3)
dt dt mk
The four parameters in the original equation have been reduced to one in the non-dimensional
equation. So the equation has been greatly simplified.
We can consider the limit of small damping by taking ε = √B ≪ 1. The appropriate way to
√ mk
say that damping is small is that B ≪ mk. You can only say something is small by comparing
it with another quantity.
We seek a perturbation solution of the form

y ∼ y0 + εy1 + ε2 y2 + . . . (3.4)

To proceed, substitute (3.4) into (3.3) to obtain

y0′′ + εy1′′ + ε2 y2′′ + ... + εy0′ + ε2 y1′ + ... + y0 + εy1 + ε2 y2 + ... = 0. (3.5)

Consider terms of the same order of ε in (3.5) and solve at each order for y0 , y1 etc.

O(1) : y0′′ + y0 = 0. (3.6)


O(ε) : y1′′ + y0′ + y1 = 0. (3.7)

First consider (3.6). The general solution is

y0 = A cos t + B sin t.

Then apply the initial conditions,

y(0) = 0 =⇒ y0 (0) = 0, y1 (0) = 0, y2 (0) = 0, . . .


dy(0) dy0 (0) dy1 (0) dy2 (0)
= 1 =⇒ = 1, = 0, = 0, . . .
dt dt dt dt
y0 (0) = 0 gives A = 0. y0′ (0) = 1 gives B = 1
Therefore we find the leading order solution

y0 (t) = sin t.

Now consider (3.7) and use the solution we found for y0 to find.

y1′′ + y1 = −y0′ = − cos t.


28 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

The homogeneous equation is given by

y1′′ + y1 = 0,

which has the solution


y1 = A1 cos t + B1 sin t.

We observe that the particular solution is resonant (cos t appears in the complementary solution),
so we take an ansatz of the form
y1 = αt cos t + βt sin t.

From this we can find the first and second derivatives

y1′ = −αt sin t + βt cos t + α cos t + β sin t,


y1′′ = −αt cos t − βt sin t − α sin t + β cos t − α sin t + β cos t
= −αt cos t − βt sin t − 2α sin t + 2β cos t.

Now upon substituting into (3.7) we obtain

−αt cos t − βt sin t − 2α sin t + 2β cos t + αt cos t + βt sin t = − cos t


−2α sin t + 2β cos t = − cos t.

Note that all terms linear in t have cancelled as we would expect. Matching coefficients, we find
sin t: α = 0 and
cos t: β = − 12 .
Hence
y1 = A1 cos t + B1 sin t − 12 t sin t.

Applying the initial conditions to find the values of the constants A1 and B1 . We find

y1 (0) = 0 =⇒ A1 = 0.

Also,

y1′ = B1 cos t − 12 t cos t − 21 sin t,


y1′ (0) = 0 =⇒ B1 = 0,

and so the solution is given by


y1 = − 12 t sin t.

Hence the perturbation solution is

y ∼ y0 + εy1 + ...
= sin t − 12 εt sin t + ... (3.8)

This solution breaks down as a perturbation expansion when εt = O(1), i.e. for t = O(ε−1 ); it is a
non-uniform expansion. This expansion can be fixed up using the method of multiple scales, which
is not part of this course.
Now let us consider the exact solution to (3.3). It is a second order homogeneous ODE so we
take the ansatz y(t) = Aeλt to obtain the auxiliary equation

λ2 + ελ + 1 = 0.
3.1. DIMENSIONAL ANALYSIS 29

Solving for λ gives:



−ε ± ε2 − 4
λ=
2
ε ip
=− ± 4 − ε2
2 2q
ε
= − ± i 1 − 41 ε2
2
Therefore we have the following general solution for y:
q  q 
−εt/2 1 2 −εt/2 1 2
y = Ae cos t 1 − 4 ε + Be sin t 1 − 4 ε .

Note that y(0) = 0 implies that A = 0. Applying the other initial condition to the first derivative
q  q 
dy ε −εt/2
q
1 2 1 2 −εt/2 1 2
= − Be sin t 1 − 4 ε + B 1 − 4 ε e cos t 1 − 4 ε ,
dt 2
shows that if y ′ (0) = 1, then q
B 1 − 14 ε2 = 1,
so that
1
B=q .
1 − 14 ε2
We can expand the exponential and the denominator in Taylor series to obtain an asymptotic
approximation to the solution:

e−εt/2
q 
1 2
y=q sin t 1 − 4 ε
1 − 14 ε2
∼ (1 − 12 εt)(1 + 18 ε2 ) sin (1 − 81 ε2 )t


= (1 − 12 εt + 18 ε2 ) sin (1 − 81 ε2 )t


= (1 − 12 εt + 18 ε2 )[sin(t) cos( 81 ε2 t) − cos(t) sin 81 ε2 t ]




= (1 − 12 εt + 18 ε2 )[sin t − 81 ε2 t cos(t) + . . .]
= sin t − 21 εt sin t + ...

which we see recovers the solution (3.8) from before. The Taylor series expansions are only valid if
εt ≪ 1, as for the asymptotic solution. √
Note that we have defined the small parameter ε = B/ mk, which is small damping. What
happens when the system has a small mass instead of a small damping coefficient? Clearly ε will
not be small, and so the scaling dependency on m is not appropriate. Let us find different length
and time scales that do not depend on the mass. From before,
  M
k τ2 1
= M = ,
B τ
τ
B
giving τ = k as a new time scale. Also,
  M
B τ 1
= ML
= ,
I0 τ
L

so that L = IB0 is a new length scale.


Now, we define the dimensionless variables
Y T
y= I0
and t = B
,
B k
30 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

which are not dependent on the mass m. The first and second derivatives are now given by:
dY dY dt I0 1 dy kI0 dy
= = = 2
dT dt dT B Bk dt B dt
d2 Y kI0 d2 y 1 k 2 I0 d2 y
= = .
dT 2 B 2 dt2 Bk B 3 dt2
Now the ODE (3.1) reads
k 2 I0 d2 y kI0 dy I0
m 3 2
+B 2 + k y = 0.
B dt B dt B
Dividing through by the coefficient of the first derivative yields
mk d2 y dy
+ + y = 0,
B 2 dt2 dt
and we define the new dimensionless parameter
mk
ε= ≪1
B2
which is small since we consider the limit with a small mass. Small mass means that m ≪ B 2 /k,
so that we are comparing mass with another quantity with the same dimensions. Now the ODE in
non-dimensional form is
εy ′′ + y ′ + y = 0. (3.9)
Rewriting the initial conditions in terms of the dimensionless variables, we find
Y (0) = 0 =⇒ y(0) = 0
and
dY I0 kI0 dy I0
(0) = =⇒ 2
(0) = ,
dT m B dt m
dy B2 1
(0) = = .
dt km ε
We seek a perurbation expansion solution of the equation as
1
y = y0 + y1 + εy2 + . . .
ε
Note that the perturbation series starts at O(1/ε) as the initial velocity dy/ dt is O(1/ε). We
substitute into (3.9) to obtain
1 1
y0′′ + εy1′′ + y0′ + y1′ + . . . + y0 + y1 + . . . = 0. (3.10)
ε ε
Now consider terms of the same order of ε in (3.10).
At O(1):
y0′ + y0 = 0 =⇒ y0 = Ae−t .
Applying the initial condition
y(0) = 0 =⇒ y0 (0) = 0, A = 0.
However, if we apply the other initial condition for the first derivative we see
1
y ′ (0) = =⇒ y0′ (0) = 1 =⇒ A = −1,
ε
which is inconsistent. The problem is that we have a first order ODE governing the dynamics
for the leading order solution y0 , yet we have two initial conditions which are impossible to satisfy
simultaneously. This is due to the system having a singular perturbation where the small parameter
ε is the coefficient of the highest derivative in (3.9). Setting ε = 0 reduces the order of the equation
by 1, so that we have a singular perturbation. We now consider regularly and singularly perturbed
ODEs in more detail.
3.2. REGULAR PERTURBATIONS 31

3.2 Regular Perturbations


We consider a boundary value problem governed by the nonlinear ODE

y ′′ + εy 2 = 0, (3.11)

with boundary conditions

y(0) = y(1) = 1.

Taking an asymptotic expansion for the solution of the form

y = y0 + εy1 + ε2 y2 + · · · , (3.12)

and substituting into (3.11) gives


 
0= y0′′ + εy1′′ + ε2 y2′′ + ··· + ε y02 2 2 2
+ 2y0 (εy1 + ε y2 + · · · ) + (εy1 + ε y2 + · · · )

= y0′′ + εy1′′ + ε2 y2′′ + εy02 + 2ε2 y1 y0 + · · ·

As usual we match terms at the same orders of ε and solving order-by-order. At leading order
O(1) we find:

y0′′ = 0,
y0 = A0 x + B0 .

Applying the boundary conditions:

y(0) = 1 =⇒ y0 (0) + εy1 (0) + · · · = 1,


=⇒ y0 (0) = 1, y1 (0) = y2 (0) = · · · = 0,
y(1) = 1 =⇒ y0 (1) = 1, y1 (1) = y2 (1) = · · · = 0.

So,

y0 (0) = 1 =⇒ B0 = 1,
y0 (1) = 1 =⇒ 1 + A0 = 1, A0 = 0.

Hence at leading order the solution is given by constant solution y0 = 1.


At the next order, O(ε) :

y1′′ + y02 = 0
=⇒ y1′′ = −y02 = −1

Integrating gives

y1′ = −x + A1
=⇒ y1 = − 21 x2 + A1 x + B1 .

Applying the boundary conditions

y1 (0) = 0 =⇒ B1 = 0,
y1 (1) = 0 =⇒ 0 = − 12 + A1 , A1 = 12 .

And so we find the O(ε) solution,

y1 (x) = − 12 x2 + 12 x = 12 x(1 − x).


32 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

At O(ε2 ) :

y2′′ + 2y1 y0 = 0
=⇒ y2′′ = −2y0 y1 = −x + x2

Integrating gives

y2′ = − 21 x2 + 13 x3 + A2
y2 = − 61 x3 + 1 4
12 x + A2 x + B 2 .

Applying the boundary conditions we have

y2 (0) = 0 =⇒ B2 = 0
y2 (1) = 0 =⇒ 0 = − 16 + 1
12 + A2 , A2 = 1
12 ,

and so the solution is given by

y2 (x) = − 16 x3 + 1 4
12 x + 1
12 x
1 2 3
= 12 x(1 − 2x + x )
= 1
12 x(1 − x)(1 + x − x2 ).

Substituting the solutions from each order in ε together into (3.12), we find the solution to (3.11)
as
1 2
y(x)∼1 + 21 εx(1 − x) + 12 ε x(1 − x)(1 + x − x2 ) + · · · .
This expansion is uniform on 0 < x < 1.

Example: Forced logistic equation


The logistic equation is a common basic model for population growth, and its extensions include
periodic forcing to represent periodically varying resources, for instance due to seasons. An example
of a forced logistic equation is

u′ = au(1 + ε cos t − u), 0 < ε ≪ 1. (3.13)


| {z }
carrying capacity

For example u represents a population size, a defines a growth rate, and the carrying capacity is
the maximum population size, which is time varying here.
We approach the solution of this equation by considering an asymptotic expansion of the form
u = u0 + εu1 + . . ., and substitute into (3.13) to obtain (ignoring terms O(ε2 ) and above):

u′0 + εu′1 = a(u0 + εu1 )(1 − u0 − εu1 + εcos t)


= au0 (1 − u0 ) + aεu1 (1 − u0 ) − aεu0 u1 + aεu0 cos t.

Proceeding to solve order-by-order, we find at O(1):

u′0 = au0 (1 − u0 ),

which is the standard logistic equation for population growth. The steady states of this equation
(u′0 = 0) are given by u0 = 0 and u0 = 1. Since u0 = 0 represents an absence of any population,
we discard this solution as trivial. Let us choose u0 = 1, so that the population is at steady state
without seasonal variation. We want to find out how seasons affect a steady population.
At O(ε):

u′1 = au1 (1 − u0 ) − au0 u1 + au0 cos t


= −au1 + acos t,
u′1 + au1 = acos t.
3.3. SINGULAR PERTURBATION 33
R
We use the integrating factor I = e a dt = eat to solve the above equation as follows:

d at
(e u1 ) = aeat cos t + C
dt Z
⇐⇒ eat u1 = a eat cos t dt + C
Z 
at+it
= a Re e dt + C
 at+it 
e
= a Re +C
a+i
 at+it
(a − i)

e
= a Re +C
a2 + 1
 at
e (cos t + i sin t)(a − i)

= a Re +C
a2 + 1
a
= 2 eat (a cos t + sin t) + C.
a +1
Therefore,
a
u1 (t) = Ce−at + (a cos t + sin t).
a2 +1
Note that as t → ∞ the solution approaches the periodic solution
a
u1 (t) = (a cos t + sin t).
a2 +1
The Ce−at term is a transient. The population then varies periodically with seasons, but with a
phase lag as the forcing is cos t and the response is a cos t + sin t.

3.3 Singular Perturbation


Let us now return to singular perturbation problems for odes.
Consider the ODE
εy ′′ + (1 + ε)y ′ + y = 0, 0 < ε ≪ 1, (3.14)
with the following boundary conditions:

y(0) = 0 and y(1) = 1.

We note that this is a singularly perturbed ODE since the small parameter ε multiplies the highest
derivative. Setting ε = 0 reduces the order of the equation by 1, which cannot satisfy two boundary
conditions. Let us first study the exact solution before considering an asymptotic approximation
to obtain insight into a suitable perturbation method.
Taking an ansatz of the form y = Aeλx , we obtain the auxiliary equation

ελ2 + (1 + ε)λ + 1 = 0.

Solving for λ: p
−1 − ε ± (1 + ε)2 − 4ε
λ=
p2ε
−1 − ε ± (1 − ε)2
=

−1 − ε ± (1 − ε)
=

1
= −1, − .
ε
34 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Hence, our solution is of the form

y = Ae−x + Be−x/ε ,

where A, B are constants.


By using the boundary conditions we can determine A, B. We see that

y(0) = 0 =⇒ A + B = 0 =⇒ B = −A,

and
y(1) = 1 =⇒ 1 = Ae−1 + Be−1/ε
= A(e−1 − e−1/ε )
 −1  −1
=⇒ A = e−1 − e−1/ε and B = − e−1 − e−1/ε .
Thus, we have that our exact solution is

e−x − e−x/ε e−x − e−x/ε


y= ∼ = e1−x − e1−x/ε
e−1 − e−1/ε e−1

as e−1/ε is exponentially small.


This solution is illustrated in Figure 3.2. Away from x = 0, e1−x/ε ∼ 0, so that

y ∼ e1−x .

This satisfies the the boundary condition at x = 1, but does not satisfy the boundary condition at
x = 0. This is because we have neglected e1−x/ε . This solution away from x = 0 is called the outer
solution.
When x = O(ε), e1−x/ε cannot be neglected as x/ε = O(1) and the solution is

y = e1−x − e1−x/ε ∼ e − e1−x/ε .

This solution valid for x = O(ε) is called the boundary layer solution or inner solution and the
region x = O(ε) is called the boundary layer. The boundary layer solution brings the outer solution
down to satisfy the boundary condition at x = 0. The boundary layer allows the second boundary
condition to be satisfied. In the boundary layer the solution is rapidly varying as y ′ = O(1/ε).
The singular perturbation solution then consists of two parts: (i) the outer solution away from
the boundary layer and (ii) an inner solution in the boundary layer. Most of the solution is the
outer solution, with the boundary layer solution only valid in a narrow layer near x = 0.
Let us now see how to develop the singular perturbation solution without knowledge of the exact
solution. Let us suppose now that we do not know the exact solution. But, suppose that we know
the boundary layer is around x = 0. (We will find out how to do find out where it is later.)
Outer Solution
Let us now construct an asymptotic solution. We know that away from the boundary layer y ′
and y ′′ are O(1) (the solution is not changing rapidly). We have that the outer solution is of the
form
y = y0 + εy1 + ε2 y2 + . . . ,
and we substitute this into (3.14) to obtain

εy0′′ + ε2 y1′′ + · · · + (1 + ε)(y0′ + εy1′ + . . . ) + y0 + εy1 + . . . = 0,


εy0′′ + · · · + y0′ + εy1′ + εy0′ + y0 + εy1 + . . . = 0.

Solving order-by-order in powers of ε, we see that at O(1) we have

y0′ + y0 = 0
=⇒ y0′ = −y0
=⇒ y0 = A0 e−x .
3.3. SINGULAR PERTURBATION 35

Figure 3.2: The boundary layer (inner region) here is a small region of thickness O(ε) near x = 0
where the solution varies rapidly, beyond which there is a slowly varying solution in the outer
region.

We have that the boundary layer is around x = 0. Hence, the outer solution satisfies the boundary
condition at x = 1, i.e.

y(1) = 1 =⇒ y0 (1) = 1, y1 (1) = y2 (1) = · · · = 0.

Using the boundary condition y0 (1) = 1, we can solve for the unknown constant A0 to obtain

1 = A0 e−1 =⇒ A0 = e =⇒ y0 = e1−x .

For O(ε) we have:


y0′′ + y1′ + y0′ + y1 = 0.
Using the solution for y0 we find

y1′ + y1 = −y0′′ − y0′ = −e1−x + e1−x = 0,


=⇒ y1′ = −y1 .

The solution is of the form


y1 = A1 e−x ,
with A1 a constant to be determined using the boundary condition y1 (1) = 0, which implies A1 = 0.
In fact, by induction, we have yn = 0, n ⩾ 1.
Note that the exact solution of the ODE (3.14) is given by
x
e−x − e− ε e−x
y= ∼ = e1−x ,
e−1 − e− ε
1
e−1

since e−x/ε , e−1/ε are transcendentally small as ε → 0. This is exactly our outer solution.
Inner Solution
There is a boundary layer at x = 0 where we need to capture rapid variations of the solution
as it adjusts to meet the outer solution y0 , while simultaneously satisfying the boundary condition
at x = 0. The boundary layer width is O(ε), so we stretch the boundary layer so that its width is
O(1) in a new coordinate.
Let X = x/ε, where X is a boundary layer variable. Let the solution in the boundary layer be
given by Y (X) = y(x). Then we have, using the chain rule,

dy dY dX 1 dY d2 y 1 d dY dX 1 d2 Y
= = , = = .
dx dX dx ε dX dx2 ε dX dX dx ε2 dX 2
36 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

So the ODE (3.14) becomes, in the boundary layer variables,

1 d2 Y 1 dY
2
+ (1 + ε) + Y = 0,
ε dX ε dX
which simplifies to
d2 Y dY
2
+ (1 + ε) + εY = 0.
dX dX
We take an asymptotic expansion for the inner solution of the form

Y (X) = Y0 (X) + εY1 (X) + · · · ,

and substitute it into the differential equation above to obtain

Y0′′ + εY1′′ + (1 + ε) Y0′ + εY1′ + εY0 + · · · = 0,




which expands as
Y0′′ + εY1′′ + Y0′ + εY1′ + εY0′ + εY0 + · · · = 0.
Solving this order-by-order in powers of ε, we see that at O(1) we have

Y0′′ + Y0′ = 0.

Taking an ansatz of Y0 = B0 eλX , we see that the auxiliary equation for λ is given by

λ2 + λ = 0 =⇒ λ = 0, −1.

Therefore, the leading order general solution is given by

Y0 = B0 + C0 e−X .

The boundary condition y(0) = 0 implies Y (0) = 0, =⇒ Y0 (0) = Y1 (0) = · · · = 0. As Y0 (0) = 0,


we have C0 = −B0 , and so
Y0 (X) = B0 − B0 e−X .
At O(ε) we have the equation

Y1′′ + Y1′ + Y0′ + Y0 = 0.

Substituting in the solution for Y0 , we find

Y1′′ + Y1′ = − Y0′ − Y0


= −B0 e−X − B0 + B0 e−X
= −B0 .

The homogeneous solution is found solving

Y1′′ + Y1′ = 0,

to obtain
Y1 (X) = B1 + C1 e−X .
The particular solution is of the form Y1 = αX as the forcing is resonant as it satisfies the homo-
geneous equation. Thus with Y1′ = α and Y1′′ = 0 we find that α = −B0 . So the O(ε) general
solution is given by
Y1 = B1 + C1 e−X − B0 X.
From the boundary condition Y1 (0) = 0, we have

B1 + C1 = 0, =⇒ C1 = −B1 .
3.3. SINGULAR PERTURBATION 37

Therefore, the solution is


Y1 = B1 − B1 e−X − B0 X,
where B0 and B1 are undetermined. How do we find them? Note that the inner equation is a
second order ODE, and so we need two conditions. We cannot use the boundary condition at x = 1
(which is in the outer region, not part of the boundary layer we are working in), so we must use
the outer solution to carry the boundary condition at x = 1 into the inner region. Note that the
outer solution (y0 (x)) is not valid in the inner region (inside the boundary layer).
The way to overcome this is to have the inner solution meet the outer solution in a region where
both solutions are valid. This is called asymptotic matching, or the method of matched asymptotic
expansions.
We have from before that the outer solution is given by

y ∼ e1−x as ε → 0, with x fixed,

and the inner solution by

Y (X) ∼ B0 (1 − e−X ) + ε[B1 (1 − e−X ) − B0 X] + . . . as ε → 0, X fixed.

We let the outer variable approach the inner region and vice versa:

x→0 as ε→0
X → ∞ as ε → 0, (X = x/ε).

Expanding the outer solution as x → 0 gives:

y = e1−x = e1 e−x
= e(1 − x) + O(x2 ).

Expanding the inner solution as X → ∞ gives:

Y ∼ B0 + ε(B1 − B0 X) + . . .

Now, matching the solutions yields

e − ex + . . . = e − εeX + . . . = B0 + ε (B1 − B0 X)

holds when

B0 = e, B1 = 0.

The inner and outer asymptotic solutions are given by

Outer solution: y = e1−x


Inner solution: Y = e − e1−X − εeX + . . . .

The inner solution is the same as the exact solution as ε → 0 (for X fixed) as
x
e−x − e− ε
Exact solution: y= 1
e−1 − e− ε
e−εX − e−X
= 1
e−1 − e− ε
(1 − εX) − e−X + . . .

e−1
−X
= e(1 − e ) − εeX + . . .
38 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

Now, how do we determine the boundary layer thickness? The governing ODE is

εy ′′ + (1 + ε)y ′ + y = 0.

We define an inner variable that stretches the outer variable by a factor 1/δ(ε):
x
X= .
δ(ε)
Then, in the new coordinate, the ODE becomes

ε d2 y 1 + ε dy
+ + y = 0,
δ 2 dX 2 δ dX
or equivalently
d2 y dy
ε 2
+ δ(1 + ε) + δ 2 y = 0.
dX dX
The leading order equations for the inner solution now depend on the scale factor δ(ε). There are
now 4 cases to consider:
Case (i) δ≪ε (Inner–inner limit):

d2 y
≃0
dX 2
This limit is no good as the solution y = AX + B cannot be matched to the outer solution
(solution doesn’t exist as X → ∞).

Case (ii) Match orders of first and second terms: δ(ε) = ε (Inner limit):
In this case the leading order balance is between the first and second derivatives to give

d2 y dy
2
+ = 0.
dX dX
This is a distinguished limit. It is the same as the leading order inner equation we found
above.

Case (iii) ε≪δ≪1 (Intermediate limit):


dy
= 0.
dX
This gives nothing.

Case (iv) δ = 1: Match orders of the first and third terms: (Outer limit):
dy
+ y = 0.
dX
This is again a distinguished limit. This is just the outer equation.
The inner equation is then determined by balancing the orders of the first and second terms
in the equation, case (ii).
In the inner and outer equations, when the two largest terms in the equations balance each other
it is known as dominant balance or the distinguished limit, as δ is of a definite order in ε.
We still need to find out how to find the location of the boundary layer.
Let us consider a more general problem— a 2-point boundary value problem:

εy ′′ + a(x)y ′ + b(x)y = 0 (3.15)


y(0) = A, y(1) = B.
3.3. SINGULAR PERTURBATION 39

For the outer solution we consider an expansion of the form y = y0 + εy1 + . . . and substitute into
(3.15) to get
εy0′′ + ay0′ + aεy1′ + by0 + εby1 + . . . = 0.
We now solve at each order of ε. At O(1) we have
b(x)
a(x)y0′ + b(x)y0 = 0, y0′ + y0 = 0.
a(x)
This ODE can be solved by introducing an integrating factor:
b(x)
Z
I = exp dx.
a(x)
Hence,
d  R b dx 
e a y0 = 0
dx R b
e a dx y = C0
R b(x)
− dx
y = C0 e a(x) .

Note that we have only one constant of integration, so we cannot satisfy both boundary conditions.
Furthermore there is trouble when a(x) = 0, which implies that there are internal boundary layers.
Internal boundary layers are not dealt with in this course.
Instead, we assume that a(x) ̸= 0 in what follows, and we now look for the boundary layers.
First we assume there is a boundary layer at x = 0.
x
Let X = δ(ε) , so x = δ(ε)X, and y(x) = Y (X) is the inner solution. With this change of
coordinates, (3.15) becomes

ε d2 Y a(δX) dY
2 2
+ + b(δX)Y = 0.
δ dX δ dX
The correct distinguished limit depends on the behaviour of a and b near x = 0. Assume that they
behave as
a ∼ aα xα , b ∼ bβ xβ as x → 0, for α, β ⩾ 0.
Then,
ε d2 Y δ α X α dY
+ a α + bβ δ β X β Y = 0.
δ 2 dX 2 δ dX
We now consider the dominant balances.
Balancing the orders of the second and first derivative terms
ε δα 1
= =⇒ δ = ε 1+α ,
δ2 δ
or balancing the second derivative and zero derivative terms
ε 1
= δβ =⇒ δ = ε 2+β ,
δ2
1
The first case, δ = ε 1+α . This balance holds if the zero derivative term is of lower order than
α
the second and first derivative terms, which means δδ < δ β i.e. α < β + 1. The boundary layer is
of thickness O(ε1/(1+α) ).
1
The second case, δ = ε 2+β , occurs if the first derivative term is of lower order than the second
and zero derivative terms. Then δ α−1 < δ β , α − 1 > β, or α > β + 1. The boundary layer is of
thickness O(ε1/(2+β) ).
The simplest case is when a(0) = a0 , b(0) = b0 , α = 0 and β = 0. The boundary layer thickness
is then O(ε) with δ = ε.
40 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

The boundary layer equation is

1 d2 Y 1 dY
2
+ a(εX) + b(εX)Y = 0
ε dX ε dX
d2 Y dY
2
+ a(εX) + εb(εX)Y = 0
dX dX
Taking an expansion of the form Y ∼ Y0 + εY1 + ε2 Y2 + . . . , we have

Y0′′ + εY1′′ + a(εX) Y0′ + εY1′ + εb(εX)Y0 + · · · = 0.




Then by Taylor expanding a(εX) and b(εX) we find,

Y0′′ + εY1′′ + (a(0) + εXa′ (0) + . . . ) Y0′ + εY1′ + ε(b(0) + εXb′ (0) + . . . )Y0 + · · · = 0.


So at O(1) we have
′′
Y0 + a0 Y0′ = 0.
Taking the usual ansatz Y0 = αeλX leads to the auxiliary equation

λ2 + a0 λ = 0,

with the solutions λ = 0 or λ = −a0 . Therefore the leading order solution is

Y0 = α0 + β0 e−a0 X .

If the boundary layer is at x = 0, then we match the inner and outer solutions by taking X → ∞.
This limit will only exist if a0 > 0. So for a(0) = a0 > 0 the boundary layer is at x = 0. We then
use the boundary condition at x = 0 for the inner solution.
From the boundary condition y(0) = A we have Y0 (0) = A, and so A = α0 + β0 . Substituting
α0 = A − β0 gives
Y0 = A + β0 (e−a0 X − 1).
If instead a(0) = a0 < 0 then the boundary layer is at x = 1. The inner solution then satisfies
the boundary condition at x = 1.
We then define the boundary layer variable X = 1−x
δ(ε) so that x = 1 − δ(ε)X. Then

ε d2 Y a(1 − δX) dY
2 2
− + b(1 − δX)Y = 0.
δ dX δ dX
Assume that a(1) = a1 ̸= 0 and b(1) = b1 ̸= 0. Then let δ = ε (so that the boundary layer is of
width O(ε)) as before. In the new coordinates the inner equation becomes

d2 Y dY
2
− a(1 − δX) + εb(1 − δX)Y = 0.
dX dX
We take an expansion of the form Y ∼ Y0 + εY1 + ε2 Y2 + . . . and substitute into the ODE above
to obtain

Y0′′ + εY1′′ − (a1 − εXa′ (1))(Y0′ + εY1′ ) + ε(b1 − εXb′ (1))(Y0 + εY1 ) + . . . = 0.

At O(1) we have
Y0′′ − a1 Y0′ = 0.
We make the usual ansatz Y0 = αeλX to obtain the auxiliary equation

λ2 − a1 λ = 0,

with the solutions λ = 0 or λ = a1 . Therefore we see that

Y0 (X) = α0 + β0 ea1 X .
3.3. SINGULAR PERTURBATION 41

From the boundary condition y(1) = B we have Y0 (0) = B (recall that X = (1 − x)/ε, so x = 1 ↔
X = 0), which gives that B = α0 + β0 . Substituting α0 = B − β0 gives that

Y0 = B + β0 (ea1 X − 1).

We match the inner and outer solutions by taking X → ∞, so we must have a1 < 0 for the inner
solution to remain valid.
But since a ̸= 0 then a(0) < 0 and a(1) < 0. a < 0 implies the boundary layer is at x = 1.
To summarise, we have two cases:
(1) If a > 0, then there is a boundary layer at x = 0, and none at x = 1. The inner solution
satisfies the boundary condition at x = 0.
(2) If a < 0, there is a boundary layer at x = 1, and none at x = 0. The inner solution satisfies
the boundary condition at x = 1.

Example:
Consider the boundary value problem

εy ′′ + (1 + x)y ′ + y = 0, y(0) = y(1) = 1.

Note that we identify the functions a(x) = 1 + x, b(x) = 1. There is a boundary layer at x = 0
since a(0) = 1 > 0, and a ̸= 0 in the domain 0 ≤ x ≤ 1.
Outer solution
We take an expansion for the outer solution of the form y = y0 + εy1 + ε2 y2 + ..., and substitute
it into the governing ODE to obtain (up to O(ε2 )):

ε(y0′′ + εy1′′ ) + (1 + x)(y0′ + εy1′ + ε2 y2′ ) + y0 + εy1 + ε2 y2 + . . . = 0.

We proceed to solve order-by-order in powers of ε.


At O(1):

(1 + x)y0′ + y0 = 0
y0
y0′ = − .
1+x
This is a separable ODE which we can solve straightforwardly as
dy0 dx
Z Z
=− ,
y0 1+x
log y0 = − log |1 + x| + C,
A0
=⇒ y0 = .
1+x
Applying the boundary condition at x = 1 (away from the boundary layer)

y(1) = 1 ⇒ y0 (1) = 1, y1 (1) = y2 (1) = ... = 0

we can determine the unknown constant as y0 (1) = 1 ⇒ A0 = 2.

At O(ε):

y0′′ + (1 + x)y1′ + y1 = 0.

Substituting in the solution for y0 , we obtain


4
(1 + x)y1′ + y1 = −y0′′ = −
(1 + x)3
1 4
y1′ + y1 = −
1+x (1 + x)4
42 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS

We can apply an integrating factor


dx
R
e 1+x = elog (1+x) = 1 + x,

so that
d 4
((1 + x)y1 ) = −
dx (1 + x)3
2
(1 + x)y1 = + A1
(1 + x)2
2 A1
y1 = 3 + 1 + x.
(1 + x)

Applying the boundary condition at x = 1, y1 (1) = 0, we determine the unknown constant

1 A1 1
y1 (1) = 0 = + =⇒ A1 = − .
4 2 2
Hence, the outer solution is
!
1
2 2
y(x) = +ε − 2
+ O(ε2 ).
1+x (1 + x)3 1 + x

Inner Solution
We now turn our attention to finding the inner solution. From the general case outlined above
x
the boundary layer variable is X = (since a(0) ̸= 0 and the boundary layer is at x = 0), and
ε
y(x) = Y (X). The inner equation is

Y ′′ + (1 + εX)Y ′ + εY = 0.

We substitute in the expansion Y = Y0 + εY1 + ε2 Y2 + ... to obtain

Y0′′ + εY1′′ + ε2 Y2′′ + (1 + εX)(Y0′ + εY1′ + ε2 Y2′ ) + ε(Y0 + εY1 ) + . . . = 0.

Again solving at each order of ε, we find at O(1):

Y0′′ + Y0′ = 0 =⇒ Y0 = α0 + β0 e−X .

As we are now seeking the inner solution, we need to apply the boundary condition at x = 0, i.e.
inside the boundary layer. In the new variables, the boundary condition is

y(0) = 1 ⇒ Y0 (0) = 1, Y1 (0) = Y2 (0) = ... = 0.

We can thus eliminate one unknown constant to find

Y0 (0) = 1 ⇒ 1 = α0 + β0 , ⇒ α0 = 1 − β0
Y0 = 1 + β0 (e−X − 1).

At O(ε) we solve:
Y1′′ + Y1′ + XY0′ + Y0 = 0.
Substituting in the solution we found for Y0 ,

Y1′′ + Y1′ = −XY0′ − Y0


Y1′′ + Y1′ = β0 Xe−X − β0 e−X + β0 − 1
3.3. SINGULAR PERTURBATION 43

This is equivalent to
d
(Y ′ + Y1 ) = β0 Xe−X − β0 e−X + β0 − 1.
dX 1
Integrating, we obtain
Z
Y1′ + Y1 = β0 (X − 1)e−X dX + (β0 − 1)X + α1

= −β0 Xe−X + (β0 − 1)X + α1 .

Using the integrating factor R


1· dX
e = eX ,
we have
d X
(e Y1 ) = −β0 X + (β0 − 1)XeX + α1 eX
dX
eX Y1 = − 21 β0 X 2 + (β0 − 1)XeX − (β0 − 1)eX + α1 eX + β1
Y1 = − 21 β0 X 2 e−X + (β0 − 1)X − β0 + 1 + α1 + β1 e−X .

Now, applying the boundary condition Y1 (0) = 0 inside the boundary layer, we can eliminate one
of the unknown constants:

Y1 (0) = 0 =⇒ 0 = −β0 + 1 + α1 + β1 =⇒ β1 = β0 − 1 − α1 .

Thus we can express the solution as

Y1 = − 12 β0 X 2 e−X + (β0 − 1)X − β0 + 1 + α1 + (β0 − 1 − α1 )e−X .

Matching
Outer solution !
1
2 2 2
y= +ε − + ...
1+x (1 + x)3 1 + x
As x → 0 (going towards the inner solution)
1 3
 
y → 2(1 − x) + ε 2 − 2 + ... = 2 + ε 2 − 2X + . . .

Inner solution

Y = 1 + β0 (e−X − 1) + ε − 12 β0 X 2 e−X + (β0 − 1)X − β0 + 1 + α1 + (β0 − 1 − α1 )e−X + . . .




As X → ∞ (go towards the outer solution)

Y → 1 − β0 + ε [(β0 − 1)X − β0 + 1 + α1 ] + . . .

Matching at O(1): 2 = 1 − β0 , β0 = −1.


Matching at O(ε): 32 = −β0 + 1 + α1 , α1 = 32 + β0 − 1 = − 12 .
Note that with these values the −2X term at O(ε) automatically matches. If this did not
happen, then there would be a mistake in the calculations, which would need to be found. Every
term must match at each order.
Hence, the inner solution is

Y (X) = 2 − e−X + ε 32 − 2X + 12 X 2 − 32 e−X + . . . .


 
44 CHAPTER 3. ORDINARY DIFFERENTIAL EQUATIONS
Chapter 4

WKB theory (geometric optics, ray


theory)

Note: WKB theory was named after Wentzel, Kramers, and Brillouin.

Wave Equation
To motivate WKB theory, let us take the standard wave equation

∂2u
= c2 ∇2 u. (4.1)
∂t2
This equation describes some types of wave motion with c the wave speed.
Look for solutions of a single frequency (a Fourier mode)

u(x, t) = v(x)eiωt ,

Then

∂2u
= −ω 2 eiωt v, and ∇2 u = eiωt ∇2 v.
∂t2
So,
 ω 2
c2 ∇2 v + ω 2 v = 0, =⇒ ∇2 v + v = 0, =⇒ ∇2 v + k 2 v = 0,
c
with c = ω/k, where k is the wavenumber (λ = 2π/k is the corresponding wavelength).
In 1D,
d2 v
+ k2 v = 0
dx2
High frequency limit: k ≫ 1 (geometric optics, ray theory). Geometric optics first arose in the
17th century with Cristiaan Huygens in the Netherlands.

Schrödinger’s Equation
Schrödinger’s Equation from quantum mechanics is

ℏ2 ′′
− ψ + V (x)ψ = Eψ (4.2)
2m

where ℏ2 /2m is small. This can be set in the form

1
ψ ′′ + (E − V (x))ψ = 0, (4.3)
ε

45
46 CHAPTER 4. WKB THEORY (GEOMETRIC OPTICS, RAY THEORY)

where ε = ℏ2 /2m ≪ 1. This equation is of the same form as the geometric optics equation with

k = 1/ ε.
In the simplest case, k constant, the geometric optics equation is

vxx + k 2 v = 0, (4.4)

which has the solution


v = A sin kx + B cos kx.
This differential equation and its solution will arise again and again when considering solutions of
pdes, so it should be remembered. It is pointless rederiving it again and again.
So the basic solution of the geometric optics equation is

A sin kx = ℑAeikx

Note that B cos kx = ℜAeikx .


There is also the case of the opposite sign in the geometric optics equation

vxx − k 2 v = 0

Then,
v = Aekx + Be−kx
Again, the solution of this equation should be remembered as it arises again and again. Note that
the solution of this equation is really the same as that of (4.4).

vxx + (−k 2 )v = 0.

Then √ √
−k2 x −k2 x
v = Aei + Be−i = Aei×ikx + Be−i×ikx = Ae−kx + Bekx .
So the solution for either sign is really just the same thing.
Now let us assume that the geometric optics equation does not have constant coefficients. This
arises when light travels through a non-uniform medium, for instance.

u′′ + k 2 q(x)u = 0, (4.5)

where k ≫ 1 is large. The WKB ansatz is

u ∼ A(x)eikϕ(x) .

(k = 1ε ). Here A is the amplitude and ϕ is the phase. Then,

ux = ikϕ′ Aeikϕ + A′ eikϕ


uxx = −k 2 (ϕ′ )2 Aeikϕ + ikϕ′′ Aeikϕ + 2ikϕ′ A′ eikϕ + A′′ eikϕ .

Substitute:
−k 2 (ϕ′ )2 A + ikϕ′′ A + 2ikϕ′ A′ + A′′ + k 2 qA = 0.
At O(k 2 ) we find the equation
2
ϕ′ = q,
which is known as the eikonal equation. Eikon comes from the Greek word (εiκων) for an im-
age. This terminology comes from geometric optics as this equation determines where the light
propagates.
The eikonal equation is easily solved as
Z xp
2
ϕ′ = q ⇒ ϕ = ± q(t) dt.
x0
47

At O(k) we find
ϕ′′ A + 2ϕ′ A′ = 0.
This is called the transport equation. In applications it determines where energy is transported. It
can also be solved as it is a separable equation
ϕ′′ A′
= −
2ϕ′ A
1
⇒ log ϕ′ = − log A + C
2
p eC A0
ϕ′ = =
A A
A0 −1/4
A = √ ′ = A0 q .
ϕ
The WKB solution of the geometric optics equation is then
R √ R √
−1/4 ik xx q(t) dt −1/4 −ik xx q(t) dt
u± (x) = A+ q (x)e 0 + A− q (x)e 0

Note that we can then easily deduce the WKB solution of

u′′ − k 2 q(x)u = 0, (4.6)

by using the discussion above. We eliminate the i to give the WKB solution
R √ R √
k x q(t) dt −k xx q(t) dt
u± (x) = A+ q −1/4 (x)e x0 + A− q −1/4 (x)e 0

Figure 4.1

So for the geometric optics equation (4.5), the WKB solutions are oscillatory for q > 0 and
exponential for q < 0. This is illustrated graphically in Figure 4.1.

Example – Airy’s equation


Airy’s equation is given by
d2 u
− xu = 0.
dx2
Airy’s equation can be solved in terms of Bessel functions, but its solutions are usually given their
own notation. George Airy was Astronomer Royal in the 19th century. Airy’s equation has optical
applications. It arises in the formation of shadow zones. It also arises in quantum tunnelling.
Let us consider the limit |x| ≫ 1. The WKB solutions can be obtained by identifying q(x) = −x,
such that
√ p √ √
q = |x| for x < 0, q = i x for x > 0.

So the solutions can be written,


R√ R√ 2 3/2 2 3/2
u ∼ A+ x−1/4 e x dx
+ A− x−1/4 e− x dx
= A+ x−1/4 e 3 x + A− x−1/4 e− 3 x
48 CHAPTER 4. WKB THEORY (GEOMETRIC OPTICS, RAY THEORY)

for |x| ≫ 1 and x > 0 and


R√ R√ 2 3/2 2 3/2
u ∼ A+ |x|−1/4 ei |x| dx
+ A− |x|−1/4 e−i |x| dx
= A+ |x|−1/4 e 3 i|x| + A− |x|−1/4 e− 3 i|x|

for |x| ≫ 1 and x < 0.

The WKB expansion can be continued to higher orders using



!
X
−n
u ∼ exp ik k Sn (x) (eiS1 = A).
n=0

This higher order theory is not part of this course.


Turning point problem
The WKB solution assumes that k 2 q(x) ≫ 1. This obviously breaks down if q(x) = 0 at some
point. The case in which q(x) = 0 at x = x0 is called the WKB turning point problem. As q changes
sign at x = x0 , the WKB solution changes from oscillatory to exponential or visa versa. While
the turning point problem may seem the exceptional case, it is vitally important. It is related to
shadow zones in optics and acoustics. This is illustrated by the following example.

Whispering Gallery Mode


St. Paul’s Cathedral in London has a large dome over the transept, as shown in Figure 4.2. At
the base of this dome there is a gallery walkway which visitors can walk around. If you whisper
against the wall at this walkway, a person on the other side of the dome can clearly hear you. For
this reason, this gallery is called the whispering gallery. The sound is trapped against the wall
and propagates around the dome. These whispering gallery modes are very common and are an
example of trapped waves. Whispering gallery modes have been used in optical devices to trap and
route light.

Figure 4.2: St. Paul’s Cathedral, London

Figure 4.3: St. Paul’s Cathedral dome

Whispering gallery modes can be analysed using WKB theory. The propagation of sound waves
is governed by the wave equation (acoustic equation)

∂2u
c2 ∇2 u = ,
∂t2
49

where c is the speed of sound, which is ∼ 340m/s.


Let us take the base of the dome to be a circle of radius R, as illustrated in Figure 4.3.
Let u = veiωt , so that we look at a single (Fourier) mode
ω
⇒ ∇2 v + k 2 v = 0, k=
c
This equation is termed Helmholtz’ equation. In plane polar coordinates the Laplacian is

1 ∂2u
 
2 1 ∂ ∂u
∇ u= r + 2 2.
r ∂r ∂r r ∂θ
So in plane polar coordinates the acoustic equation is

∂ 2 v 1 ∂v 1 ∂2v
+ + + k 2 v = 0.
∂r2 r ∂r r2 ∂θ2
We now seek a separation of variables solution of the form v(r, θ) = w(r)y(θ). Substituting into
Helmholtz’ equation gives
1 1
w′′ y + w′ y + 2 wy ′′ + k 2 wy = 0.
r r
w′′ w′ y ′′
r2 + r + k2 r2 + =0
w w y
w′′ w′ y ′′
r2 + r + k2 r2 = − = λ (4.7)
w w y
as the left hand side is a function of r and the right hand side is a function of θ. Here λ is a
separation constant.
y equation:
y ′′ + λy = 0.
We seek periodic solutions, so take λ = ξ 2 , such that the solution is given by

y = A sin ξθ + B cos ξθ.

Now the solution must be single valued, so that when you go once around the dome, θ goes from
0 to 2π, y must return to the same value. So we want 2π periodic solutions for y, so that ⇒ ξ = n
for integer n. Then, substituting into (4.7) we obtain

n2
 
′′ 1 ′ 2
w + w + k − 2 w = 0.
r r
This is Bessel’s equation of order n with linearly independent solutions w = Jn and w = Yn .
However, Bessel functions as such are not part of this course.
Let us consider the regime of large k such that we can apply the WKB approximation. First,
to put the equation into the WKB form (4.5), let w = r−1/2 g. Then
dw
= r−1/2 g ′ − 12 r−3/2 g
dr
d2 w
= r−1/2 g ′′ − r−3/2 g ′ + 34 r−5/2 g
dr2
Substituting into (4.7) gives

n2 −1/2
 
− 12 ′′ −3/2 ′ 3 −5/2 −3/2 ′ 1 −5/2 2
r g −r g + 4r g +r g − 2r g + k − 2 r g=0
r
" #
1
n2 −
g ′′ + k 2 − 4
g = 0.
r2
50 CHAPTER 4. WKB THEORY (GEOMETRIC OPTICS, RAY THEORY)

This is of WKB form with


n2 − 41
q(x) = k 2 − .
r2
The WKB solution for k large is then given by
r r
1
n2 − 4 1
n2 − 4
A+ A−
Rr Rr
i k2 − dt −i k2 − dt
g(r) =  t2 t2
1/4 e + 1/4 e .
r0 r0

n2 − 14 n2 − 14
k2 − r2
k2 − r2

This WKB solution is valid away from the turning points, which are the zeroes of the equation
1
n2 −
q(x) = k 2 − 4
= 0.
r2
These turning points are given by
q
1
kr = kr0 = n2 − 4 ∼ n,

such that the solutions change form approximately at the radii


q
n2 − 41 n
r0 = ∼ .
k k
For r > r0 the solution is oscillatory, which means that there are sound waves. For r < r0 , the
solution is exponential and so there are no waves, that is no sound. So the sound waves are trapped
against the dome wall, which is the whispering gallery mode. This is illustrated in Figure 4.4.

r
0
n/k R

Figure 4.4: Change of WKB solution at the turning point r ∼ n/k.


Chapter 5

Partial differential equations

We begin this chapter by introducing the so-called wave equation, given by

∂2u 2
2∂ u
= c (5.1)
∂t2 ∂x2
where c is the wave speed. The wave equation is used to describe sound waves, shallow water waves,
electromagnetic waves and so on. However, it does not describe waves on the ocean. Ocean waves
are a different type of wave termed dispersive waves and are not described by the wave equation.

D’Alemebert’s Solution
The basic solution of the wave equation is D’Alembert’s solution. We will now derive d’Alembert’s
solution to the wave equation (5.1), given by

u = f (x − ct) + g(x + ct),

where and f , g are arbitrary functions given by initial conditions.


The lines x − ct = constant and x + ct = constant are called characteristics and x − ct and
x + ct are called characteristic variables.
dx
Note that, for x − ct constant, we have that = c, so f (x − ct) is a wave moving to the right
dt
with speed c.
dx
Similarly, for x + ct constant, we have that = −c, so g(x − ct) is a wave moving to the left
dt
with speed c.
Now, let 
ξ = x − ct
.
η = x + ct
We can rewrite the wave equation in terms of ξ and η by using the chain rule to find the first and
∂ 2 u ∂u ∂2u
second order partial derivatives of u(x, t), i.e. ∂u
∂t , ∂t2 , ∂x and ∂x2 :

∂u ∂u ∂ξ ∂u ∂η ∂u ∂u
= + = −c +c ,
∂t ∂ξ ∂t ∂η ∂t ∂ξ ∂η

∂2u ∂ 2 u ∂ξ ∂ 2 u ∂η ∂ 2 u ∂ξ ∂ 2 u ∂η
= −c − c + c + c
∂t2 ∂ξ 2 ∂t ∂ξ∂η ∂t ∂ξ∂η ∂t ∂η 2 ∂t
∂2u ∂2u ∂2u
= c2 2 − 2c2 + c2 2 ,
∂ξ ∂ξ∂η ∂η

Similarly
∂u ∂u ∂ξ ∂u ∂η ∂u ∂u
= + = + ,
∂x ∂ξ ∂x ∂η ∂x ∂ξ ∂η

51
52 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

and
∂2u ∂ 2 u ∂ξ ∂ 2 u ∂η ∂ 2 u ∂ξ ∂ 2 u ∂η
= + + +
∂x2 ∂ξ 2 ∂x ∂ξ∂η ∂x ∂ξ∂η ∂x ∂η 2 ∂x
∂2u ∂2u ∂2u
= + + .
∂ξ 2 ∂ξ∂η ∂η 2

Substituting these into the wave equation (5.1) gives

∂2u 2
2 ∂ u
2
2∂ u
2
2∂ u
2
2 ∂ u
2
2∂ u
c2 − 2c + c = c + 2c + c .
∂ξ 2 ∂ξ∂η ∂η 2 ∂ξ 2 ∂ξ∂η ∂η 2

Finally, cancelling the terms, we get that

∂2u
=0
∂ξ∂η
 
∂ ∂u
=⇒ =0
∂ξ ∂η
∂u
=⇒ = g̃(η).
∂η

Integrating again with respect to η we have


Z
u = g̃(η) dη + f (ξ) =⇒ u = g(η) + f (ξ),

which gives d’Alembert’s solution after substituting for η, ξ.

5.1 First-Order Nonlinear Hyperbolic PDEs


We shall come back to the wave equation later. We shall first consider simpler wave equations.
The simplest wave equation is first order and is given by

∂u ∂u
+ c0 = 0, (5.2)
∂t ∂x
where c0 is a constant. Let ξ = x − c0 t and u = f (ξ). Next, we find the partial derivatives in terms
of ξ using the chain rule

∂u df ∂ξ df
= = −c0 ,
∂t dξ ∂t dξ
∂u df ∂ξ df
= = .
∂x dξ ∂x dξ

Substituting back into (5.2), we get that

∂u ∂u df df
+ c0 = −c0 + c0 = 0.
∂t ∂x dξ dξ

Thus, the solution of this wave equation is u = f (x−c0 t), where f is arbitrary. For initial conditions
given by u = F (x) at t = 0,
=⇒ F (x) = f (x)
=⇒ u = F (x − c0 t),
which is a wave moving to the right with speed c0 .
The solution to equation (5.2) was just stated, so we need to know how to derive this solution.
The method to do this is the method of characteristics.
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 53

Method of characteristics
In this section we will consider solving wave equations or first-order hyperbolic pdes of the form

∂u ∂u
q(x, t) + p(x, t) = h(x, t, u)
∂t ∂x
∂u ∂u
=⇒ + c(x, t) = r(x, t, u).
∂t ∂x
We look for the derivative along special curves known as characteristics. On characteristics we
have positions parameterised by time such that x = x(t), so that on characteristics the solution is
of the form u = u(x(t), t).

t
x = x(t)

dx
dt = C0

u = f (x), x = ξ x

Figure 5.1: Simple straight line characteristics (constant slope) propagating the initial data u(x, t =
0) = f (x) forward in time.

The simplest case is when c is a constant, such as in

∂u ∂u
+ c0 = 0,
∂t ∂x
with initial condition u = F (x) at t = 0. We use the directional derivative, calculated by using the
chain rule, to find the change of the solution along characteristics:

du du(x(t), t) ∂u dx ∂u ∂u dx ∂u
= = + = + .
dt dt ∂x dt ∂t ∂t dt ∂x
Comparing this to (5.2), we see that these equations are the same if

dx
=⇒ = c0 .
dt
We then have from (5.2) that
du
= 0.
dt
Thus,
du dx
=0 on = c0 .
dt dt
dx
This is called the characteristic form of the equation. Here, = c0 gives the slope of the charac-
dt
teristic curves (see Figure 5.1). Thus, we reduce the PDE to a set of ODEs which we can integrate
to get that
u = A on x = c0 t + B.
To find the constants of integration A, B, recall the initial condition

t = 0, u = F (x).
54 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

Let us take a particular characteristic starting at x = ξ at t = 0, i.e. at


t = 0, x = ξ, u = F (ξ) =⇒ u = F (ξ) = A
x = c0 t + B, t = 0 =⇒ ξ = B.
Thus, the general solution is
u = F (ξ) on x = c0 t + ξ
Eliminating (the parameter) ξ between these equations gives
u = F (x − c0 t) as before.
This represents a wave moving to the right with velocity c0 . Each point on the initial condition
moves to the right with velocity c0 . That is, each point on the initial profile propagates on a
characteristic starting at that point and moving with velocity dx/ dt = c0 , as illustrated in Figure
5.2.

Figure 5.2

In the general case, i.e. when c is not a constant, we have a first order (hyperbolic) PDE of the
form
∂u ∂u
+ c(x, t) = r(x, t, u)
∂t ∂x
dx
where the characteristic = c(x, t) is not a straight line. But the method of characteristics still
dt
dx
entails setting = c(x, t), which gives non-straight lines when c is not constant. Then, again
dt
using the chain rule gives that on the characteristics x = x(t) and
du(x(t), t) ∂u dx ∂u
= + .
dt ∂x dt ∂t
Matching with the original equation gives the characteristic form
du dx
= r(x, t, u) on = c(x, t).
dt dt

Example
Consider the PDE
∂u ∂u
+ e−t = 0,
∂t ∂x
with initial condition u = f (x) at t = 0.
In characteristic form
du dx
= 0 on = e−t
dt dt
=⇒ u = A on x = −e−t + B.
On the characteristic starting at x = ξ at t = 0, we have that u = f (ξ) at t = 0. Thus,
A = f (ξ) on ξ = −1 + B =⇒ B = ξ + 1,
u = f (ξ) on x = −e−t + ξ + 1 =⇒ ξ = x + e−t − 1.
Hence, the solution is, on eliminating ξ,
u = f (x + e−t − 1).
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 55

Figure 5.3: Information flows along characteristic curves.

Example: Signalling Problem


Consider
∂u ∂u
x2 + = −tu, x > 0,
∂t ∂x
with the boundary condition u = g(t) at x = 0. This is called a signalling problem as it is a
boundary value problem with u specified at x = 0 for all t. A signal is being generated at x = 0
which propagates into x > 0.

Figure 5.4: Characteristic curves propagating data from a boundary condition across space.

To use the method of characteristics, we first rewrite the PDE as

∂u 1 ∂u −tu
+ 2 = 2 .
∂t x ∂x x
In this characteristic form we have u = u(x(t), t) on characteristics C (see Figure 5.4).
We see that dx/dt = 1/x2 , so the equation in characteristic form is

du −tu dx 1
= 2 on = 2.
dt x dt x
Integrating the characteristics, we have
Z Z
x2 dx = dt =⇒ 1 3
3x = t + B,

so we have that
du −tu 1 3
= 2 on 3x = t + B.
dt x
56 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

Next, note that


du
dt = −tu =⇒ du = −tu = −( 1 x3 − B)u
dx dx 3

dt
1
Z Z
1 3 1 4

=⇒
u
du = 3 x − B dx =⇒ log u = − 12 x + Bx + A.

Hence, we have that


4 /12+Bx
u = Ae−x on 1 3
3x = t + B.
We choose the characteristic starting at x = 0, t = τ , so that u = g(τ ) at x = 0
1 3
3x = t + B =⇒ 0 = τ + B
=⇒ B = −τ and A = g(τ ).

Hence, we have that


4 /12−τ x
u = g(τ )e−x on 1 3
3x = t − τ =⇒ τ = t − 13 x3 ,

which gives us, on eliminating τ , the solution


4 4  4
u = g t − 31 x3 e−x /12−tx+x /3 = g t − 31 x3 ex /4−tx .


C
t

x
C1 C2

x
(a) With u given on the dotted line as (b) In this case, two separate characteris-
known data, the characteristic C propa- tics C1 and C2 intersect after finite time,
gates the solution from one part of the given leading to a singularity (known as a caustic
distribution to another – this is generally an in optics).
ill-posed problem.

Figure 5.5: Caution – information flows along the characteristics, so it is important to consider
that they are well behaved and maintain uniqueness.

The solution of first order hyperbolic is determined by the characteristics, so to understand


their solution, you must understand how the characteristics behave. Problems can arise, depending
on how the characteristics behave, so that the equation is ill-posed and there is no solution. This
is illustrated in Figure 5.5. Characteristics carry the solution from the initial condition. If a
characteristic intersects the boundary again it will carry the solution with it and this will, in
general, not agree with the solution at the point it intersects again. So the equation is ill-posed
and there is no solution. Two different characteristics carry different values of the solution, so if
they intersect, there will be two different solution values at the same point. Again, the equation is
ill-posed.
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 57

Nonlinear First Order PDEs


The simplest nonlinear first order hyperbolic PDEs have the form

∂u ∂u
+ c(u) = 0,
∂t ∂x
with the initial condition u = f (x) at t = 0.
As before, we look for the characteristics and set the equation in characteristic form. On a
characteristic C, x = x(t), u = u(x(t), t) the chain rule gives

du ∂u ∂u dx
= + .
dt ∂t ∂x dt
Set dx/dt = c(u). Then the characteristic form of the equation is

du dx
= 0 on = c,
dt dt
=⇒ u = A on x = ct + B.

Note that c = c(u), so that c is not constant. However, on a characteristic u is a constant as


du/dt = 0, so that c is constant on a characteristic too.
At t = 0, x = ξ and u = f (ξ)

=⇒ A = f (ξ), B = ξ.

Hence, the solution is given by

u = f (ξ) on x = ct + ξ = c(f (ξ))t + ξ.

In principle, we then would eliminate ξ to obtain a solution of the form u = u(x, t). However,
unless c and f are simple, this cannot be done in general and we are left with an implicit solution
in terms of the parameter ξ.
The solution propagates on characteristics. Therefore, the solution evolves by each point on the
initial condition propagating with velocity dx/dt = c(u). Let us assume that c′ (u) > 0 and refer
to Figure 5.6 which illustrates this case. Let us take two points on the initial condition u1 and u2 ,
with u1 > u2 , and look at the portion on the initial condition which is decreasing. The point u1
then has greater velocity than the point u2 , as c(u1 ) > c(u2 ). Therefore, the point u1 eventually
catches up with the point u2 at the breaking time, as in Figure 5.6(b), so that the solution has an
infinite slope. For times greater than the breaking time t > tb , the point u1 overtakes the point u2
and the solution becomes multi-valued, as in Figure 5.6(c). This is unphysical as in applications u
represents quantities, such as densities of fluids or traffic, so this multi-valued solution is unphysical
and unacceptable. The characteristic solution breaks down at the breaking time tb . After this
breaking time, something else needs to be done which has no counterpart for linear hyperbolic
equations. Figure 5.6(d) shows this breaking behaviour in terms of the characteristics. After the
breaking time the characteristics intersect, which is the same thing as saying that the solution
becomes multi-valued as the solution propagates on the characteristics. For times greater than the
breaking time t > tb the solution is made single valued by the fitting of a shock.
The same behaviour occurs for c′ (u) < 0, but the breaking and multi-valued behaviour occurs
in portions of the initial condition which are increasing. It is the mirror image of the behaviour for
c′ (u) > 0.

Example: Shock Waves


Consider
∂u ∂u
+u =0
∂t ∂x
58 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

(a) t = 0 (b) t = tb
u u
Each point moves with
velocity c(u)

Infinite slope

x
x

(c) t > tb (d) Slope given by dx/ dt = c(u)


t

Figure 5.6: Wave steepening due to the nonlinear wave equation with speed c(u) until an infinite
slope emerges at a breaking time tb . As the wave breaks, it becomes multi-valued at some regions of
x – this corresponds to regions of characteristics intersecting, as demonstrated in the final subfigure.
The intersection of characteristics is also known as a shock.

with the initial condition u = f (x) at t = 0 where f (x) = 1 − x. Note that for this example, c = u.
Then, in characteristic form

du dx
= 0 on = u,
dt dt
=⇒ u = A on x = ut + B.

u is constant on characteristic curves, because the derivative du/ dt = 0.


By the initial condition, at t = 0, x = ξ and u = 1 − ξ

=⇒ A = 1 − ξ, B = ξ.

Hence, we have that

x−t
u=1−ξ on x = (1 − ξ)t + ξ =⇒ ξ = .
1−t
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 59

Therefore, on eliminating ξ the solution is given by


x−t 1−x
u=1− = .
1−t 1−t
Note u = ∞ at t = 1! t = 1 is the breaking time tb . We look at shocks in more detail later.
Expansion Waves
The opposite of a shock is an expansion wave, as demonstrated in Figure 5.7. Let us assume that
c′ (u) > 0 and take two points on the initial condition u1 and u2 with u1 > u2 . On an increasing
portion of the initial condition, c(u1 ) > c(u2 ), so that the initial condition with u1 moves faster
than the point u2 . The initial condition then evolves by pulling apart and expanding out. This is
called an expansion wave. It is the opposite to a shock wave as the solution expands out rather
than contracting down on itself.
u u

velocity c Expansion wave

Stretches out

x x

c = c1 = c(f1 ) t c = c2 = c(f2 )

x
f = f1 f = f2
∆x

Figure 5.7: Expansion wave.

Example
Consider
∂u ∂u
+u =0
∂t ∂x
with the initial condition at t = 0 given by
(
u+ , x > 0
u=
u− , x < 0,

where u+ > u− . This initial condition is shown in Figure 5.8.


In characteristic form the equation is

du dx
= 0 on = u,
dt dt
=⇒ u = A on x = ut + B.

For x < 0 at t = 0,

x = ξ, u = u− =⇒ A = u− , B=ξ
=⇒ u = u− on x = u− t + ξ,
60 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

Figure 5.8: Expansion wave initial condition

So the solution is identically u = u− . For the initial condition in x < 0 there is a limiting
characteristic with ξ = 0 as this is where the solution for x < 0 ends. Hence, x = u− t on this
characteristic.
Similarly, for x > 0,

x = ξ, u = u+ =⇒ A = u+ , B=ξ
=⇒ u = u+ on x = u+ t + ξ,

Again, the solution is identically u = u+ . Again, the limiting value of ξ for the initial condition in
x > 0 is ξ = 0, which gives x = u+ t.
However, this is not the total solution as we are missing a portion. So far we have determined
the solution in x ≤ u− t and x ≥ u+ t due to the limiting characteristics with ξ = 0. So we are
missing the solution in u− t < x < u+ t, as shown in Figures 5.9 and 5.10. This missing portion
of the solution can be found by realising that it is generated by each point in the jump in the
initial condition between u− and u+ , as shown in Figure 5.9. The initial jump between u− and
u+ expands out to form the missing portion of the solution. Each portion of the jump generates
its own characteristic on which the solution propagates, as shown in Figures 5.10 and 5.11. The
characteristics are
x = ut + ξ.

Now the jump is at x = 0 at t = 0, so that ξ = 0. Hence,


x
x = ut, u= .
t

This is the solution in u− t < x < u+ t. This solution is called a centred expansion fan. The limits
of the expansion fan are determined by continuity. It must link to u+ and u− at its two ends. To
match with u = u+ ahead, we require x/t = u+ , and to match with u = u− behind, x/t = u− . The
centred expansion fan then occurs in u− t ≤ x ≤ u+ t.

Figure 5.9: Expansion waves.


5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 61

u
t>0

u2

Expansion flow
u1

u1 t u2 t x

Figure 5.10: Expansion waves.

Expansion flow
c = c1 c = c2

Figure 5.11: Expansion wave characteristics

Therefore, the complete expansion fan solution is given by



u− , x < u− t

u = xt , u− t ⩽ x ⩽ u+ t

u+ , x > u+ t.

We can easily verify that u = x/t is a solution of the equation by direct substitution.
x ∂u −x ∂u 1
u= =⇒ = 2 , = ,
t ∂t t ∂x t
so
∂u ∂u −x x
+u = 2 + 2 =0
∂t ∂x t t
as required.

5.1.1 Breaking: shocks and shock fitting

Figure 5.12: Formation of a shock in a breaking wave.

Consider
∂u ∂u
+ c(u) = 0,
∂t ∂x
62 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

where c′ (u) > 0 and with the initial condition u = f (x) at t = 0.


In characteristic form

du dx
= 0 on = c(u)
dt dt
=⇒ u = A on x = c(u)t + B

since u, and thus c(u), are constant on a characteristic. Using the initial condition, we choose the
characteristic at t = 0, x = ξ and u = f (ξ). Hence

A = f (ξ), B = ξ,

which gives us that


u = f (ξ) on x = c(f (ξ))t + ξ,

which is an implicit solution.


Now, let F (ξ) = c(f (ξ)) and rewrite the previous solution as

u = f (ξ) on x = F (ξ)t + ξ.

We want to show that this solution breaks with its x derivative becoming infinite at the breaking
time tb . We use the chain rule
∂u ∂u ∂ξ ∂ξ
= = f ′ (ξ) .
∂x ∂ξ ∂x ∂x
Using implicit differentiation on the characteristic, we have that

dF ∂ξ ∂ξ ∂ξ 1
x = F (ξ)t + ξ =⇒ 1 = t+ =⇒ = .
dξ ∂x ∂x ∂x dF
1+ t

Therefore,
∂u f ′ (ξ)
= .
∂x dF
1+ t

The problem is when the denominator goes to zero, so that there is a vertical tangent. This occurs
at the breaking time tb , when the characteristic solution ceases to be valid. Breaking occurs when

dF −1
F ′ (ξ) < 0 .

1+ t = 0 ⇐⇒ t = ′
dξ F (ξ)

The breaking time is then given by

−1
F ′ (ξ) < 0 .

tb = min
ξ F ′ (ξ)

This breaking and the formation of a multi-valued solution is shown in Figure 5.12. As explained
above, for t > tb the characteristics intersect and the solution is multi-valued.
To make the solution single valued, we fit in a shock or shock wave, which cuts out the multi-
valued portions of the solution. We want to fit a shock to make the solution single valued for
t ⩾ tb , while conserving quantity described by the equation. The concept of shock waves comes
from compressible gas flow, for u is the gas velocity. A shock wave is the loud bang generated by
supersonic aircraft and explosions. As matter cannot be created or destroyed, a shock wave must
be fitted by ensuring that the underlying quantity is conserved.
This breaking and shock fitting are nonlinear phenomena.
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 63

(a) Multivalued unphysical solution with


crossing of characteristics. (b) Now single valued, no crossing shocks.

Figure 5.14: x1 , x2 fixed. Density ρ for unit length, flux q for unit time.

5.1.2 Conservation laws


Let us derive a general first order pde based on the conservation of a quantity whose density is ρ.
Examples of ρ are the density of gas and the density of traffic on a road.
Let us consider a region containing ρ in x2 ≤ x ≤ x1 , as shown in Figure 5.14. The only way ρ
can change in x2 ≤ x ≤ x1 is that it enters at x = x2 and exits at x = x1 . The flow of ρ across a
boundary is called the flux of ρ and is denoted by q. Conservation of ρ in x2 ≤ x ≤ x1 gives
d x1
Z x1
∂q
Z
ρ dx = q2 − q1 = − dx
dt x2 x2 ∂x

This says that the rate of change of ρ in x2 ≤ x ≤ x1 is given by the difference of the flux q2
entering at x = x2 and the flux q1 leaving at x = x1 . This equation is the integral form of the
conservation of ρ. Then
Z x1 Z x1
∂ρ ∂q
i.e. dx = − dx
x2 ∂t x2 ∂x
Z x1  
∂ρ ∂q
+ dx = 0.
x2 ∂t ∂x

This is true for any x1 , x2 , and so


∂ρ ∂q
+ = 0.
∂t ∂x
This is a conservation law (in differential form). It can then be seen that first order hyperbolic
equations arise as conservation equations.
Let us now consider conservation of ρ if a shock occurs at x = s(t), with x2 < s < x1 , as shown
in Figure 5.15. We allow for discontinuities in ρ and q, that is shocks.
The integral conservation equation is
d x1
Z
ρ dx + q1 − q2 = 0.
dt x2
64 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

Figure 5.15: Schematic of shock location.

Let us now break up the density integral to one below the shock and one above the shock
d s(t) d x1
Z Z
ρ dx + ρ dx + q1 − q2 = 0,
dt x2 dt s(t)
Z s(t) Z x1
∂ρ ∂ρ
dx + ṡ(t)ρ|x=s− + dx − ṡ(t)ρ|x=s+ + q1 − q2 = 0,
x2 ∂t s(t) ∂t
on using the fundamental theorem of calculus. Now let x2 → s− and x1 → s+ . Then

ṡ(ρ2 − ρ1 ) = q2 − q1 ,

q2 − q1 [q]
U = ṡ = =
ρ2 − ρ1 [ρ]
is then the shock velocity, where [.] denotes the jump in a quantity across the shock.
The conservation equation
∂ρ ∂q
+ = 0,
∂t ∂x
where ρ and q are differentiable, then holds, but is supplemented by the jump condition U = [ q ]/[ ρ ]
at shocks. This is illustrated in Figure 5.16. The breaking, multi-valued portion of the solution
is cut out in such a manner that, the new, single valued solution with a jump, conserves ρ with a
shock fitted. The characteristics then do not cross, but terminate at the shock. The shock fitted
solution is single valued, but with a jump at the shock.

Figure 5.16: Shock fitting

Example
The details of shock fitting can be done for the specific equation
∂u ∂u
+u = 0, (5.3)
∂t ∂x
with the initial condition at t = 0 given by
(
u2 , x < 0,
u=
u1 , x > 0,
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 65

with u1 < u2 , as shown in Figure 5.17. This solution breaks immediately as there is a jump in the
initial condition at t = 0.

Figure 5.17: Initial distribution which breaks immediately.

In conservation form, we can rewrite (5.3) as


∂ ∂ 1 2

(u) + 2u = 0,
∂t ∂x
which has the density ρ = u and flux q = u2 /2. The shock velocity is thus given by

[ u2 /2 ]
U = ṡ = .
[u]
In characteristic form
du dx
= 0 on =u
dt dt
u = A on x = ut + B

For x < 0, at t = 0, x = ξ and u = u2 . Then B = ξ and A = u2 . The solution is then u = u2 .


For x > 0, at t = 0, x = ξ and u = u1 . Then B = ξ and A = u1 . The solution is then u = u1 .
Behind the shock, u = u2 , and ahead of the shock u = u1 . Therefore,

(u21 − u22 )/2


U = ṡ = = 21 (u1 + u2 ).
u1 − u2
Hence,
s = 12 (u1 + u2 )t + C.
The shock starts at x = 0 at t = 0. Hence s(0) = 0 and C = 0.
So, the shock fitted solution is
(
u1 , x > 21 (u1 + u2 )t
u=
u2 , x < 21 (u1 + u2 )t

The characteristics and shock trajectory for this solution are shown in Figure 5.18.

Example
Let us consider a different initial configuration, which is more complicated, as shown in 5.19 and is

0,
 x < 0,
u(x, 0) = u1 , 0 ≤ x ≤ 1,

0, x > 1.

The initial jump at x = 1 is compressive and so a shock forms there. The jump at x = 0 is
expansive and a centred expansion fan forms there.
66 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

t s(t)

x = u1 t

x = u2 t

Figure 5.18: Fitted shock solution

In characteristic form
du dx
= 0 on =u
dt dt
u = A on x = ut + B.

For x < 0 and x > 1, at t = 0, x = ξ and u = 0. Then A = 0 and so u = 0.


For 0 < x < 1, at t = 0, x = ξ and u = u1 . Then A = u1 and so u = u1 .
Shock
u

u1

→ Shock
Expansion flow →

1 x

Figure 5.19: Initial profile for u(x, 0).

Hence
[q] u2 /2
ṡ = = 1 = 12 u1 .
[ρ] u1
Integrating this simple ODE, we see that the position of the shock is given by
1
1
s= 2 u1 t + C,

The shock starts at x = 1 at t = 0 and so s(0) = 1, that is, C = 1.


Expansion Fan
The expansion fan starts at x = 0 at t = 0. Hence in the expansion fan B = 0 for the
characteristics and u = x/t.
We need to determine the ends of the expansion fan. At the left end continuity gives u = 0, so
that x = ut = 0. At the right end, continuity gives u = u1 , so that x = u1 t.
The total solution is given by



 0, x ⩽ 0,

x/t, 0 < x < u t,
1
u=
u1 , u1 t ⩽ x < 21 u1 t + 1,


x > 21 u1 t + 1.

0,

This solution is valid until the expansion wave hits the shock. The shock moves at velocity s = u1 /2
and the right end of the expansion fan moves at velocity u1 . Hence, the expansion fan catches up
with the shock.
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 67

This occurs when


2
u1 t = 21 u1 t + 1 =⇒ t= .
u1
At times t > 2/u1 we know that the shock velocity is given by

[ u2 /2 ]
ṡ = .
[u]

Now u = 0 ahead of the shock is still valid. However, behind the shock there is now the expansion
fan with u = x/t. So just behind the shock x = s and u = s/t. The jump condition then gives

1 s2
2 t2
ṡ = s ,
t
1s
ṡ = 2 t,
ds dt
Z Z
1
= 2 + Ã,
s t
1
log s =
2 log t + Ã,

=⇒ s = A t.

At t = 2/u1 ,

s = 12 u1 t + 1 = 2,
r
2
∴ 2=A ,
u1

A = 2u1 .

Figure 5.20: Shock interacting with expansion fan.

So for t > 2/u1 the solution is



0,
 x⩽0

u = x/t, 0 < x < 2u1 t
 √
0, x > 2u1 t.

Note that √
u1
ṡ = √ → 0 as t → ∞.
2t
The shock slows down on intersecting with the expansion flow and eventually dies. The character-
istic diagram for this solution is shown in Figure 5.20.
First order pdes have numerous applications of importance in science and engineering. Two
examples will be given: traffic flow and the geophysics of landscape erosion.
68 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

5.1.3 Traffic Flow


Take cars travelling in one lane with no movements between lanes. No entrances and exits. In this
model we have

ρ : Number of cars per unit length,


q : Flux of cars.

Then, the equation for conservation of cars is

∂ρ ∂q
+ = 0.
∂t ∂x
What is the relation between the flux q and the density ρ, i.e. what is q = Q(ρ)? We can deduce
a general form of Q. When there are no cars, ρ = 0 and there is no flux, q = 0. So Q(0) = 0.
When the density of cars on a road reaches a maximum ρf , then there is grid lock and the cars do
not move and there is no flux, Q(ρf ) = 0. In between ρ = 0 and ρ = ρf there must be a maximum
of Q, q = qmax . This general flux is shown in Figure 5.21.
Traffic measurements done in the United States of America give typical values of qmax = 1500
per hour at ρM ≈ 80 per mile and a traffic jam density of ρj ≈ 228 per mile. Since q = ρv, the car
speed is given by v = q/ρ. Then for qmax , v ≈ 20 mph for the maximum flux of cars on a highway.

Figure 5.21: Traffic flux as a function of traffic density.

With q = Q(ρ) the conservation of cars equation becomes

∂ρ ∂ρ
+ Q′ (ρ) = 0.
∂t ∂x
We have Q′ (ρ) > 0 for ρ < ρM , and Q′′ (ρ) < 0 (see Figure 5.21). In characteristic form this
equation is

dρ dx
= 0 on = Q′ (ρ),
dt dt
ρ = A on x = Q′ (ρ)t + B.

Q′′ (ρ) < 0 implies that the wave speed/characteristic slope c = Q′ (ρ) is a decreasing function of
the density ρ, c′ (ρ) < 0.
Let us take an initial lump of cars as shown in Figure 5.22. Since c′ < 0, higher values of ρ
propagate slower than lower values. Hence, the decreasing portion of the initial distribution of
cars forms an expansion fan and the increasing portion develops a shock as shown in Figure 5.23.
A backward propagating shock forms. This solution accords with experience. When there is a
disturbance in traffic, you approach a jam of cars with the information that the car density is
increasing coming towards you, which is the shock propagating towards you. If the shock did not
travel backwards, then you would not know about the disturbance and you would crash. When
you pass the disturbance, the traffic density suddenly decreases and the cars space out, which is
you travelling into the expansion fan.
5.1. FIRST-ORDER NONLINEAR HYPERBOLIC PDES 69

Figure 5.22: Initial traffic density distribution at t = 0.

Expansion wave

Shock

Figure 5.23: The evolution of the traffic density as a shock and an expansion wave forms.

5.1.4 Erosion
Let us now consider erosion, the erosion of a hill or mountain.
Suppose we have a hill with a height h(x, t), as in Figure 5.24.

Figure 5.24: Coordinate for the height of the hill.

Intuition:  
∂h ∂h
= −Q ,
∂t ∂x
i.e. the rate of change of the height due to erosion is some function Q that depends on the slope,
the greater the slope, the more the hill erodes. Let σ = hx . Then

∂σ ∂Q(σ)
=− .
∂t ∂x

So what is Q? If the hill is flat, there is no erosion, Q(0) = 0. The greater the slope, the greater the
erosion. Also the erosion is symmetric in the sign of the slope. Figure 5.25 shows such an erosion
function Q(σ).
By the chain rule, the equation for hill erosion is

∂σ ∂σ
+ Q′ (σ) = 0.
∂t ∂x
In characteristic form
dσ dx
=0 on = Q′ (σ).
dt dt
70 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

Figure 5.25: Height erosion function

(a) t = 0 (b) t > tb

Figure 5.26: Shock and expansion formation for erosion.

In this case we have c = Q′ and c′ = Q′′ (σ) > 0 so that shocks break forward, portions of the initial
hill with σ decreasing form a shock and portions with σ increasing form an expansion fan. This
process is shown in Figure 5.26.
The shock is in σ = hx , i.e. there is a discontinuity in the slope, not the height. So the shock
results in a corner being formed in the hill profile. The expansion fan in σ in regions with σ
increasing results in the slope being spread out and this portion becoming flat. The net result is
the top of formation shown in Figure 5.27. This type of hill shape is familiar from the American
West and American westerns, and is common in the states of Arizona, New Mexico and Utah. It
is called a mesa, which is the Spanish word for a table (la mesa).

Figure 5.27: The profile of the hill height h(x, t) for t > tb .

So simple first order pdes based on the concept of conservation can be used to model physical
phenomena and give results in accordance with observations of the real world.

5.2 Higher Order Partial Differential Equations


The standard second order partial differential equations are the following.

• The Heat (Diffusion) Equation


∂u
= ν∇2 u
∂t
This is a parabolic equation and ν is the diffusivity of the medium. This equation has
applications to heat flow and diffusion, diffusion of water in soil and chemicals in a medium.

• The Wave Equation


∂2u
= c2 ∇2 u
∂t2
5.2. HIGHER ORDER PARTIAL DIFFERENTIAL EQUATIONS 71

This is a hyperbolic equation and c is the wave speed. Applications include sound, electro-
magnetic waves, shallow water waves (not deep water waves). This equation also models
gravity waves, with c the speed of light.
• Laplace’s Equation
∇2 u = 0
This is an elliptic equation and has applications in static problems, gravitation, electrostatics,
perfect fluid flow and water waves. This course will not deal with Laplace’s equation.

5.2.1 Derivation of the Heat Equation


∂u
= ν∇2 u
∂t
We shall now derive the heat equation for the flow of heat through a material, based on the
concept of the conservation of energy.
Let T (x, t) denote temperature. The amount of heat energy per unit volume is ρcT , where ρ is
the density of the material and c is the heat capacity. The density and heat capacity of materials
are measured quantities which can be found in engineering and physical tables. Denote the heat
flux (flow) by q.
Let us consider conservation of heat energy in an arbitrary volume V of the material. The
surface of V is ∂V and the outward unit normal to ∂V is n. The density of heat energy in V is
ρcT .

Figure 5.28: Heat flow in a volume V .

We need an expression for the heat flux q, which is the rate of flow of heat energy across a unit
surface per unit time. We will use Fourier’s Law (1822) which states
q = −k∇T,
where k is the thermal conductivity. Thermal conductivity measures how easily heat can flow
through the material. Fourier’s Law says that heat flows down its gradient from hot to cold.
The total amount of heat energy in a volume V is
Z
ρcT dV.
V

Therefore the rate of change of heat energy in V is


d
Z
ρcT dV.
dt V
For the heat loss over the boundary ∂V , we are only interested in the component of q going
out of V . The component of q parallel to the surface does not result in heat flowing out of V . The
component of q perpendicular to ∂V is n · q. So the total heat flux out of ∂V is
Z
q · n dS.
∂V
72 CHAPTER 5. PARTIAL DIFFERENTIAL EQUATIONS

So by conservation of energy,
d
Z Z Z
ρcT dV = − q · n dS = k∇T · n dS.
dt V ∂V ∂V

Next we use Gauss’ Theorem which states


Z Z
∇ · F dV = F · n dS.
V ∂V

So we have,
d
Z Z
ρcT dV = ∇ · (k∇T ) dV,
dt V V

Z Z
(ρcT ) dV = ∇ · (k∇T ) dV,
V ∂t V
Z  

(ρcT ) − ∇ · (k∇T ) dV = 0.
V ∂t

This must hold for any V and so



(ρcT ) = ∇ · (k∇T ).
∂t
Then since ρ, c, k are constants
∂T
= ν∇2 T,
∂t
k
where ν = ρc is called the thermal diffusivity.

5.2.2 Derivation of the Wave Equation


∂2u
= c2 ∇2 u
∂t2
The wave equation can be derived for sound waves from the equations for the flow of a fluid. It
can also be derived from the equations for waves on instrument strings. But we shall derive the wave
equation from the equations of electromagnetic theory, Maxwell’s equations. James Clerk Maxwell
made the equations of electricity and magnetism consistent and then predicted the existence of
electromagnetic waves, radio waves, microwaves, infrared light, visible light etc. in one of the
greatest scientific achievements.
We start from Maxwell’s Equations:
ρ
∇·E = , (5.4)
ϵ
∇ · H = 0, (5.5)
∂E
∇×H =J +ϵ , (5.6)
∂t
∂H
∇ × E = −µ , (5.7)
∂t
where E(x, t) is the electric field, H(x, t) is the magnetic field, J is the current density, ρ is the
charge density, µ is the magnetic permeability and ϵ is electric permittivity of free space.
Equation (5.4) is Gauss’ Law which gives the electric field generated by electric charges of
density ρ. Equation (5.5) states that you cannot generate isolated north and south poles of a
magnet. Equation (5.6) is Ampere’s Law, which determines the magnetic field generated by flowing
electrical charges. Equation (5.7) is Faraday’s Law which determines the electric field generated
by a varying magnetic field. Faraday’s Law is the equation governing electricity generation by
spinning magnets around electrical wires.
In a vacuum with no charges (ρ = 0) and no currents (J = 0), Maxwell’s equations simplify to
∂E ∂H
∇ · E = 0, ∇ · H = 0, ∇×H =ϵ , ∇ × E = −µ .
∂t ∂t
5.2. HIGHER ORDER PARTIAL DIFFERENTIAL EQUATIONS 73

Differentiating the third of these equations with respect to t gives,

∂H ∂2E
∇× =ϵ 2 .
∂t ∂t
Then using the fourth equation,

∂2E
 
1 ∂H
∇× ∇×E = −∇ × = −ϵ 2 .
µ ∂t ∂t

So we have,
∂2E
µϵ = −∇ × (∇ × E).
∂t2
We have the following identity for the curl of the curl:

∇ × (∇ × F ) = ∇(∇ · F ) − ∇2 F .

Applying this gives


∇ × (∇ × E) = ∇(∇ · E) − ∇2 E = −∇2 E
using the fact ∇ · E = 0. Therefore we have,

∂2E
= c2 ∇2 E
∂t

where c = 1/ µϵ. This is the (vector) wave equation for the electric field E with wave speed c.
In free space ϵ0 = 8.8545 × 10−12 m−3 kg−1 s4 A2 and µ0 = 1.2566 × 10−6 m kg s−2 A−2 . Then
c = 2.9980 × 108 ms−1 , which is the speed of light. This result shows that light is electromag-
netic radiation, which is one of the most important scientific discoveries ever. Maxwell made the
prediction that light is electromagnetic radiation and the existence of electromagnetic radiation in
general in 1862. Heinrich Hertz experimentally verified the existence of electromagnetic radiation
in 1887.

You might also like