Chapter 1

Mathematical modelling

This book concerns the application of mathematics to problems in the physical sci-
ences, and particularly to problems which arise in the study of the environment. Much
of the environment consists of fluid — the atmosphere, the ocean — and even those
parts which are solid may deform in a fluid-like way — ice sheets, glaciers, the Earth’s
mantle; as a consequence, one way into the study of the environment is through the
study of fluid dynamics, although we shall not follow that approach here. Rather,
we shall approach the study of environmental problems as applied mathematicians,
where the emphasis is on building a suitable mathematical model and solving it, and
in this introductory chapter, we set out the stall of techniques and attitudes on which
the subsequent chapters are based.
There are two particular points of view which we can bring to bear on the math-
ematical models which describe the phenomena which concern us: these are the dy-
namical systems approach, or equivalently the bifurcation theory approach; and the
perturbation theory approach. Each has its place in different contexts, and sometimes
they overlap.
The bifurcation theory approach is most usually (but not always) brought to
bear on problems which have some kind of (perhaps complicated) time-dependent
behaviour. The idea is that we seek to understand the observations through the
understanding of a number of simpler problems, which arise successively through
bifurcations in the mathematical model, as some critical parameter is changed. A
classic example of this approach is in the study of the origin of chaos in the Lorenz
equations, or the onset of complicated forms of thermal convection in fluids.
In its simplest form (e. g., in weakly nonlinear stability theory) the perturbative
approach is similar in method to the bifurcational one; however, the ethos is rather
different. Rather than try and approach the desired solution behaviour through a se-
quence of simpler behaviours, we try and break down the solution by making approxi-
mations, which (with luck) are in fact realistic. In real problems, such approximations
are readily available, and part of the art of the applied mathematician is having the
facility of being able to judge how to make the right approximations. In this book,
we follow the perturbative approach. It has the disadvantage of being harder, but it
is able to get closer to a description of how realistic systems may actually behave.

1
1.1 Conservation laws and constitutive laws
The basic building blocks of continuous mathematical models are conservation laws.
The continuum assumption adopts the view that the physical medium of concern may
be considered continuous, whether it be a porous medium (for example, sand on a
beach) or a fluid flow. The continuum hypothesis works whenever the length or time
scales of interest are (much) larger than the corresponding microscale. For example,
the formation of dunes in a desert (length scale hundreds of metres) can be modelled
as a continuous process, since the microscale (sand grain size) is much smaller than
the macroscale (dune length). Equally, the modelling of large animal populations or
of snow avalanches treats the corresponding media as continuous.
Conservation laws arise as mathematical equations which represent the idea that
certain quantities are conserved — for example, mass, momentum (via Newton’s law)
and energy. More generally, a conservation law refers to an equation which relates
the increase or decrease of a quantity to terms representing supply or destruction.
In a continuous medium, the typical form of a conservation law is as follows:
∂φ
+ ∇.f = S. (1.1)
∂t
In this equation, φ is the quantity being ‘conserved’ (expressed as amount per unit
volume of medium, i. e., as a density; f is the ‘flux’, representing transport of φ within
the medium, and S represents source (S > 0) or sink (S < 0) terms. Derivation of
the point form (1.1) follows from the integral statement
! ! !
d
φ dV = − f .n dS + S dV, (1.2)
dt V ∂V V

after application of the divergence theorem (which requires f to be continuously dif-

ferentiable), and by then equating integrands, on the basis that they are continuous
and V is arbitrary. Derivation of (1.1) thus requires φ and f to be continuously
differentiable, and S to be continuous.
Two basic types of transport are advection (the medium moves at velocity u, so
there is an advective flux φu) and diffusion, or other gradient-driven transport (such
as chemotaxis). One can thus write

f = φu + J, (1.3)

where J might represent diffusive transport, for example.

Invariably, conservation laws contain more terms than equations. Here, for ex-
ample, we have one scalar equation for φ, but other quantities J and S are present
as well, and equations for these must be provided. Typically, these take the form of
constitutive laws, and are usually based on experimental measurement. For example,
diffusive transport is represented by the assumption

J = −D∇φ, (1.4)

2
where D is a diffusion coefficient. In the heat equation, this is known as Fourier’s
law, and the heat equation itself takes the familiar form
∂
(ρcp T ) + ∇.(ρcp T u) = ∇.(k∇T ) + Q, (1.5)
∂t
where Q represents any internal heat source or sink.

1.2 Non-dimensionalisation
Putting a mathematical model into non-dimensional form is fundamental. It al-
lows us to identify the relative size of terms through the presence of dimensionless
parameters. Although technically trivial, there is a certain art to the process of
non-dimensionalisation, and the associated concept of scaling. We illustrate some of
the precepts by consideration of the heat equation, (1.5). We write it in the form
(assuming density ρ and specific heat cp are constant)

∂T
+ u.∇T = κ∇2 T + H, (1.6)
∂t
where H = Q/ρcp . We have taken ∇.u = 0, which follows from the conservation of
mass equation
∂ρ
+ ∇.(ρu) = 0, (1.7)
∂t
together with the supposition of incompressibility in the form ρ = constant.
Suppose we are to solve (1.6) in a domain D of linear magnitude l, on the boundary
of which we prescribe
T = TB on ∂D, (1.8)
where TB is constant. We also have an initial condition

T = T0 (x) in D, t = 0, (1.9)

and we suppose u is given, of order U .

We can make the variables dimensionless in the following way:

x = lx∗ , u = U u∗ , t = tc t∗ , T = TB + (∆T )T ∗ . (1.10)

We do this in order that both dependent and independent dimensionless variables be

of numerical order one, written O(1). If we can do this, then we might suppose a
priori that derivatives such as ∇∗ T ∗ (∇ = l−1 ∇∗ ) will also be of numerical O(1), and
the size of various terms will be reflected in certain dimensionless parameters which
occur.
In writing (1.10), it is clear that l is a suitable length scale, as it is the size of D.
For example, if D was a sphere we might take l as its radius or diameter. We also
suppose that the origin is in D; if not, we could write x = x0 + lx∗ , where x0 ∈ D:
evidently x∗ = O(1) in D.

3
 A similar motivation underlies the choice of an ‘origin shift’ for T . In the absence
of a heat source, the temperature will tend to the uniform state T ≡ TB as t → ∞.
If H '= 0, the final state will be raised above TB (if H > 0) by an amount dependent
on H. We take ∆T to represent this amount, but we do not know what it is in
advance — we will choose it by scaling. The subtraction of TB from T before non-
dimensionalisation is because the model for T contains only derivatives of T , so that
it is really the variation of T about TB which we wish to scale.
In a similar way, the time scale tc is not prescribed in advance, and we will choose
it also by scaling, in due course.
With the substitutions in (1.10), the heat equation (1.6) can be written in the
form " 2 # " # " #
l ∂T ∗ Ul ∗ ∗ ∗ ∗2 ∗ Hl2
+ u .∇ T = ∇ T + . (1.11)
κtc ∂t∗ κ κ∆T
This equation is dimensionless, and the bracketed parameters are dimensionless. They
are somewhat arbitrary, since tc and ∆T have not yet been chosen: we now do so by
scaling.
The solution of the equation can depend only on the dimensionless parameters. It
is thus convenient to choose tc and ∆T so that two of these are set to some convenient
value. There is no unique way to do this.
The temperature scale ∆T appears only in the source term. Since it is this which
determines the temperature rise, it is natural to choose
Hl2
∆T = . (1.12)
κ
It is also customary to choose the time scale so that the two terms of the advective
derivative on the left of (1.11) are the same size, and this gives the convective time
scale
l
tc = . (1.13)
U
It is finally also customary (if sometimes confusing) to remove the asterisks (or what-
ever equivalent symbol is used). If this is done, the dimensionless equation takes the
form $ %
∂T
Pe + u.∇T = ∇2 T + 1, (1.14)
∂t
where the Péclet number is
Ul
Pe = , (1.15)
κ
and the solution of the model depends only on this parameter (as well as the initial
condition). The boundary condition is
T = 0 on ∂D, (1.16)
and the initial condition is
T = θ(x) at t = 0, (1.17)
where
T0 (lx) − TB
θ(x) = . (1.18)
∆T

4
1.2.1 Scaling
A well-scaled problem generally refers to a model in which the dimensionless param-
eters are O(1) or less. Evidently, this can be ensured simply by dividing through
by the largest parameter in any equation. More importantly, if parameters are nu-
merically small, then (as we discuss below) approximate solutions can be obtained
by neglecting them. The problem is well-scaled if the resulting approximation makes
sense. For example, (1.14) is well-scaled for any value of P e. However, the problem
εTt = ε∇2 T + 1, with ε ( 1, is not well scaled. One makes a problem well-scaled in
this situation by rescaling the variables, and we will see examples in our subsequent
discussion.

1.2.2 Approximations
Let us consider (1.14) with (1.16) and (1.17), and suppose that θ ≤ O(1). If P e ( 1,
we obtain an approximation by putting P e = 0: ∇2 T + 1 ≈ 0. Evidently, we cannot
satisfy the initial condition, and this suggests that we rescale t: put t = P e τ , so that
(approximately)
∂T
= ∇2 T + 1; (1.19)
∂τ
now we can satisfy the initial condition (at τ = 0) too. Often one abbreviates the
rescaling by simply saying, ‘rescale t ∼ P e, so that Tt ≈ ∇2 T + 1’.

boundary
layer

sub-characteristics

Figure 1.1: Sub-characteristics and boundary layer for the equation (1.14) when P e ,
1. The sub-characteristics are the flow lines dx/dt = u, and the boundary layer (of
thickness O(1/P e)) is on the part of the boundary where the flow lines terminate.

On the other hand, if P e , 1, then Tt + u.∇T ≈ 0, and we can satisfy the

initial condition; but we cannot satisfy the boundary condition on the whole of the
boundary ∂D, since the approximating equation is hyperbolic (its characteristics are
called ‘sub-characteristics’). To remedy this, one has to rescale x near the part of the

5
boundary where the boundary condition can not be satisfied, and this is where the
sub-characteristics terminate. This gives a spatially thin region, called (evidently) a
boundary layer, of thickness 1/P e (see figure 1.1).
Another case to consider is if θ , 1, say θ ∼ Λ , 1. We discuss only the
case P e , 1 (see also question 1.1). Since T ∼ Λ initially, we need to rescale
1
T , say T = ΛT̃ . Then P e [T̃t + u.∇T̃ ] = ∇2 T̃ + , and with T̃ = O(1), we have
Λ
T̃t + u.∇T̃ ≈ 0 for P e , 1. The initial function is simply advected along the flow
lines (sub-characteristics), and the boundary condition T̃ = 0 is advected across D. In
a time of O(1), the initial condition is ‘washed out’ of the domain. Following this, we
1
revert to T , thus Tt + u.∇T = (∇2 T + 1). Evidently T will remain ≈ 0 in most of
" # Pe
1 χ 1 2
D, and in fact T ∼ O . Putting T = , χ satisifies χt + u.∇χ = ∇ χ + 1,
Pe Pe Pe
and there is a boundary layer near the boundary as shown in figure 1.1. If n is the
1
coordinate normal to ∂D in this layer, then n ∼ in the boundary layer. The final
Pe
1 < O(1).
steady state has T ∼ , and this applies also for θ ∼
Pe
These ideas of perturbation methods are very powerful, but a full exposition is
beyond the scope of this book. Nevertheless, they will relentlessly inform our dis-
cussion. While it is possible to use formal perturbation expansions, it is sufficient in
many cases to give more heuristic forms of argument, and this will typically be the
style we choose.

1.3 Qualitative methods for differential equations

The language of the description of continuous processes is the language of differen-
tial equations, and these will form the instrument of our discussion. The simplest
differential equation is the ordinary differential equation, and the simplest ordinary
differential equation (or ODE) is the first order autonomous equation

ẋ = f (x), (1.20)

dx
where the notation ẋ ≡ indicates the first derivative, and the use of an overdot is
dt
normally associated with the use of time t as the independent variable, i.e., ẋ = dx/dt.
The solution of (1.20) with initial condition x(t0 ) = x0 can be written as the
quadrature ! x
dξ
t = t0 + , (1.21)
x0 f (ξ)

and, depending on the function f , this may be inverted to find x explicitly. So, for
example, the solution of ẋ = 1 − x2 is x = tanh(t + c) (if |x(t0 )| < 1).
Going on with this latter example, we see that x → 1 as t → ∞ (and x → −1
as t → −∞), and in practice, this may be all we want to know. If a population
is subject to constant immigration and removal by mutual pair destruction, so that

6
 2

1
.
x = f(x)
0

-1
-2 -1 0 1 2
x

Figure 1.2: The evolution of the solutions of ẋ = f (x) (here f = 1 − x2 ) depends only
on the sign of x.

ẋ = 1 − x2 , then after a transient (a period of time dependence), the population will

equilibrate stably to x = 1. But to ascertain this, all we need to know is the shape
of the curve f (x) = 1 − x2 . Simply by finding the zeros of 1 − x2 and the slope of the
graph there, we can immediately infer that for all initial values x(0) > −1, x → 1 as
t → ∞, while if x(0) < −1, then x → −∞ as t → −∞: see figure 1.2. And this can
be done for any function f (x) in the equation ẋ = f (x).
This simple example carries an important message. Approximate or qualitative
methods may be just as useful, or more useful, than the ability to obtain exact results.
An extension of this insight suggests that it may often be the case that approximate
analytic insights can provide more information than precise, computational results.

1.3.1 Oscillations
If we move from first order systems to second order systems of the form

ẋ = f (x, y),
ẏ = g(x, y), (1.22)

more interesting phenomena can occur. This is the subject of phase plane analysis,
and the fundamental distinction between first and second order systems is that peri-
odic oscillations can occur. An illuminating example is illustrated in figure 1.3, and
is typified by (but is not restricted to) the equations

ẋ = y − g(x),
ẏ = h(x) − y, (1.23)

7
 h(x)
y

g(x)

Figure 1.3: Nullclines for (1.23).

where the functions g and h are as shown in the figure: g is unimodal (e. g., like
g = xe−x ) and h is monotonic decreasing (e. g., like h = 1/(x − c)). The graphs of
g(x) and h(x) (and more generally, the curves where ẋ = 0 and ẏ = 0) are called
the nullclines of x and y, and it is simple to see that where they intersect, there is
a steady state solution, and also that in the four regions separated by the nullclines,
the trajectories wind round the fixed point in a clockwise manner.
The next issue is whether the fixed point is unstable. If we denote it as (x∗ , y ∗ ),
write x = x∗ + X, y = y ∗ + Y , and linearise for small X and Y , then
" #
−g # 1
U̇ ≈ U, (1.24)
h# −1
" #
X
where U = , and the derivatives are evaluated at the fixed point. The stability
Y " #
−g # 1
of such a two by two system with community matrix A = is governed
h# −1
by the trace and determinant of A. Solutions of (1.24) proportional to eσt exist if
σ 2 − σ tr A + det A = 0, and this delineates the stability regions in the (tr A, det A)
space as indicated in figure 1.4. In the present case, tr A = −g # − 1, det A = g # − h# ,
so that for the situation shown in figure 1.3, where h# < g # < 0, det A > 0, and the
fixed point is an unstable spiral (or node) if g # < −1. When g # = −1, there is a Hopf
bifurcation, and if the system has bounded trajectories (as is normal for a model of
a physical process) then one expects a stable periodic solution to exist. Figure 1.5
illustrates a possible example.

8
 det A

stable unstable
spiral spiral
stable unstable
node node

tr A
saddle

Figure 1.4: Characterisation of fixed point stability in terms of trace and determinant
of the community matrix A. The curve separating spirals from nodes is given by
det A = 14 (tr A)2 .

.
y=0
y

.
x=0

Figure 1.5: Typical form of limit cycle for a system with nullclines as in figure 1.3.

9
 .
y=0
y
.
x=0

Figure 1.6: Typical form of relaxation oscillation in phase plane for (1.25).

1.3.2 Relaxation oscillations

It is a general precept of the applied mathematician that there are three kinds of
numbers: small, large, and of order one. And the chances of a number being O(1)
are not great. Thus for systems of the form (1.22), it is often the case in practice
that the time scales for each equation are different, so that in suitable dimensionless
units, the second order system (1.23) might take the form

εẋ = y − g(x),
ẏ = h(x) − y, (1.25)

where the parameter ε is small. Now suppose that the nullclines y = g(x) and
y = h(x) for the system (1.25) are as shown in figure 1.6, i. e., g has a cubic shape.
Trajectories rotate clockwise, and linearisation about the fixed point yields a commu-
nity matrix A with tr A = −(g # /ε)−1, det A = (g # −h# )/ε, thus with g # > h# , the fixed
point is a spiral or node, and with ε ( 1, tr A ≈ −g # /ε > 0, so it is unstable. Thus we
expect a limit cycle, and because ε ( 1, this takes the form of a relaxation oscillation
in which the trajectory jumps rapidly backwards and forwards between branches of
the x nullcline. For ε ( 1, x rapidly jumps to its quasi-equilibrium y ≈ g(x), and
then y migrates slowly (ẋ ≈ [h(x) − g(x)]/g # (x)) until g # = 0 and x jumps rapidly to
the other branch of g. Figure 1.7 shows the time series of the resulting oscillation.
The motion is called ‘relaxational’ because the fast variable x ‘relaxes’ rapidly to a
quasi-stationary state after each transient excursion.

10
 3.5

3
x
2.5

1.5

0.5

0
t
0 50 100 150 200 250 300

Figure 1.7: Time series for x corresponding to figure 1.6.

1.3.3 Hysteresis
Lighting a match is an everyday experience, but an understanding of why it occurs is
less obvious. As the match is lit, a reaction starts to occur which is exothermic, i. e.,
it releases heat. The amount of heat released is proportional to the rate of reaction,
and this itself increases with temperature (coal burns when hot, but not at room
temperature). The heat released is given by the Arrhenius expression A exp(−E/RT ),
where E is the activation energy, R is the gas constant, T is the absolute temperature,
and we take A as constant (it actually depends on reactant concentration). A simple
model for the match temperature is then
dT
c = −k(T − T0 ) + A exp(−E/RT ), (1.26)
dt
where c is a suitable specific heat capacity, k is a cooling rate coefficient, and T0
is ambient (e. g., room) temperature. The terms on the right represent the source
term due to the reactive heat release, and a Newtonian cooling term (cooling rate
proportional to temperature excess over the surroundings).
We can solve (1.26) as a quadrature, but it is much simpler to look at the problem
graphically. Bearing in mind that T is absolute temperature, the source and sink
terms typically have the form shown in figure 1.8, and we can see that there are three
equilibria, and the lowest and highest ones are stable. Of course, one could have only
the low equilibrium (for example, if k is large or T0 is low) or the high equilibrium
(if k is small or T0 is high). The low equilibrium corresponds to the quiescent state
— the match in the matchbox; the high one is the match alight. If we vary T0 , then
the equilibrium excess temperature ∆ (= T − T0 ) varies as shown in figure 1.9: the
upper and lower branches are stable.
We can model lighting a match as a local perturbation to ∆; the heat of friction
in striking a match raises the temperature excess from near zero to a value above the
unstable equilibrium on the middle branch, and ∆ then migrates to the stable upper

11
 1

0.8

Ae-RT
S
0.6

0.4 k(T-T0 )

0.2
S
U
0
0 2000 4000 6000 8000 10000

Figure 1.8: Plots of the functions A exp[−E/R(T + Tm )] and k(T − T0 ) using values
Tm = 273 (so T is measured in centigrade), with values A = 1, E = 20, 000, R = 8.3,
k = 10−4 , T0 = 15◦ C.

∆
6000

5000 hot
4000

3000 combustion
x
2000

1000
cool
0
0 200 400 600 800 1000
T T+ T0

Figure 1.9: Equlibrium curve for ∆0 as a function of T0 , parameters as for figure

1.8, but E = 35, 000. An initial condition above the unstable middle branch leads to
combustion.

12
branch, where the reaction (like that of a coal fire) is self-perpetuating. Figure 1.9
also explains why it is difficult to light a wet match, but a match will spontaneously
light if held at some distance above a lighted candle.
Figure 1.9 exhibits a form of hysteresis, meaning non-reversibility. Suppose we
place a (very large, so it will not burn out) match in an oven, and we slowly raise
the ambient temperature from a very low value to a very high value, and then lower
it once again. Because the variation is slow, the excess temperature will follow the
equilibrium curve in figure 1.9. At the value T+ , ∆ suddenly jumps (spontaneous
combustion) to the hot branch, and remains on this if T0 is increased further. Now if
T0 is decreased, ∆ remains on the hot branch until T0 = T− , below which it suddenly
drops to the cool branch again (extinction).1 The path traced out in the (T0 , ∆T )
plane is not reversible (it is not an arc but a closed curve).
The reason the multiple equilibria exist (at least for matches) is that for many
reactions, E/R is very large and also A is very large. This just says that it is possible
that Ae−E/RT is very small near T0 but jumps rapidly at higher T to a large asymptote.
To be more specific, we non-dimensionalise (1.26) by putting
T = T0 + (∆T )θ, t = [t]t∗ , (1.27)
and in fact we choose the cooling time scale [t] = c/k. Then we have, dropping the
asterisk, and after some simplification,
" # $ %
A E E∆T θ
θ̇ = −θ + exp − exp , (1.28)
k∆T RT0 RT02 1 + εθ
where ε = ∆T /T0 . The temperature rise scale ∆T has to be chosen, and there are
two natural choices: to set the exponent coefficient E∆T /RT02 to one, or the pre-
multiplicative constant to one. In one way, the latter seems the better choice: it
seems to balance the source with the sink. But because E/R is large, we might then
find E∆T /RT02 to be large, which would ruin the intention. So we choose (but it
does not really matter)
RT02
∆T = , (1.29)
E
so that $ %
θ
θ̇ = −θ + λ exp , (1.30)
1 + εθ
where " #
EA E RT0
λ= 2
exp − , ε= . (1.31)
kRT0 RT0 E
If typical values are T0 = 300 K, E/R = 10, 000 K, we see that ε ( 1, and also, since
" #
λ0 1 AR
λ = 2 exp − , λ0 = , (1.32)
ε ε kE
1
We can understand why T follows the equilibrium curve as follows. We can write (1.26) in terms
˙ = T0 − g(∆), where g(∆) is a cubic-like curve similar to the
of suitable dimensionless variables as ∆
function T0 (∆) depicted in figure 1.9. if T0 is slowly varying, then T0 = T0 (δt) where δ ( 1, and
putting τ = δt, we have δd∆/dτ = T0 (τ ) − g(∆); thus on the slow time scale τ , ∆ will tend rapidly
to a (quasi-equilibrium) zero of the right hand side.

13
λ is extremely sensitive to ε and thus T0 .
So long as θ = O(1), or at least θ ( 1/ε (i.e. T − T0 ( T0 ), we can neglect the
εθ term, so that
θ̇ ≈ −θ + λeθ . (1.33)
This gives the lower part of the S-shaped curve in figure 1.9, and the equilibria are
given by θe−θ = λ, the roots of which coalesce and disappear if λ > e−1 . This
corresponds to the value of T0 = T+ in figure 1.9, and implies
" #
E E
≈ 1 + ln λ0 + 2 ln . (1.34)
RT+ RT+

There are two roots to this, but only one has E/RT+ , 1. Further, since x , 2 ln x
if x , 1, we have, approximately,
E
T+ ≈ . (1.35)
R[1 + ln λ0 + 2 ln{1 + ln λ0 }]

If E/R , T0 , then the fact that one can light matches at room temperature suggests
that λ0 is large, and specifically ln λ0 ∼ E/RT0 . (Note that this does not imply
λ = O(1).)
Carrying on in this vein, let us suppose that we define a temperature Tc by
$ %
E
λ0 = exp , (1.36)
RTc

and we suppose Tc ∼ T0 . It follows that T+ ≈ Tc , or more precisely,

Tc
T+ ≈ , (1.37)
1 + εc {1 + 2 ln(1 + ε−1
c )}

where εc = RTc /E. The stable cool branch and unstable middle branch are then the
roots of $ " #%
−θ 1 1 T0
θe ≈ λ = 2 exp − 1− , (1.38)
ε ε Tc
and in general λ ( 1 (if T0 < Tc ), so that we find the stable cool branch (when
θ ( 1)
" #2 $ " #%
E E 1 1
θ≈λ≈ exp − , (1.39)
RT0 R Tc T0
and the unstable middle branch (where θ , 1),
" # " #
1 T0 E 1 1
θ≈ 1− + O(| ln ε|) ≈ − . (1.40)
ε Tc R T0 Tc

Evidently θ becomes O(1/ε) on the middle branch, and to allow for this, we put

θ = Θ/ε, (1.41)

14
and (1.30) becomes
$ & " #'%
1 1 Θ T0
Θ̇ = −Θ + exp − 1− . (1.42)
ε ε 1+Θ Tc
Equating the right hand side to zero gives an equilibrium which can be written ap-
proximately as 2
Tc − T0
Θ≈ + O(ε| ln ε|), (1.43)
T0
and Θ tends to infinity as T0 → 0. The hot branch is recovered for even higher values
of Θ, so that Θ , 1, in which case the equilibrium of (1.42) is given by
$ %
1 T0
Θ ≈ exp , (1.44)
ε εTc
and increases again with T0 .
At a fixed value of T0 (and thus λ), the critical value of T for ignition is that on
the unstable middle branch, as this gives the necessary temperature which must be
generated in order for combustion to occur. From (1.43) (ignoring terms in ε), this
can be written dimensionally in the simple approximate form

T ≈ Tc , (1.45)

which is approximately the critical temperature at the nose of the curve in figure
1.9. The fact that T is approximately constant on the unstable branch is due to the
steepness of the exponential curve in figure 1.8, which is in turn due to the large value
of E/R. In terms of the parameters of the problem, the critical (ignition) temperature
is thus
E
Tc ≈ " #. (1.46)
AR
R ln
kE
Hysteresis and multiplicity of solutions is a theme which will recur again and again
in this book.

1.3.4 Resonance
Swinging a pendulum is an everyday experience, and one which students learn about
in a first year mechanics course. If the point of suspension itself oscillates, then one has
a forced pendulum, and an interesting phenomenon occurs. At low forcing frequencies,
the pendulum oscillates in phase with the oscillating point of support. At high forcing
frequencies, it oscillates out of phase with the support. Moreover, this change in phase
appears to occur abruptly, at a particular value of the forcing frequency. At the same
time, there is also a sudden rise in amplitude of the motion, although it is less easy to
see this in a casual experiment. This observation is associated with the phenomenon
of resonance, and can be easily experienced by jumping on a springboard.
2
Note that as T0 → Tc , (1.43) matches with (1.40).

15
 |A|

Ω0 ω

Figure 1.10: Resonant amplitude response.

To illustrate the phenomenon of resonance mathematically, we solve the equation

of a forced oscillator, and an example of such a system is the forced pendulum. To
be specific, we take as a model equation

ü + β u̇ + Ω20 sin u = ε sin ωt. (1.47)

This represents the motion of a damped, non-linear pendulum, with a forcing on the
right hand side which mimics (it is not a precise model) the effect on the pendulum
of an oscillating support. We suppose that the model is dimensionless, and that ε
is small, so that the response amplitude of u will be also. We also suppose that the
damping term β is small.
The simplest approximation of (1.47) neglects β altogether, and linearises sin u,
so that
ü + Ω20 u ≈ ε sin ωt, (1.48)
to which the forced solution is
u = A sin ωt, (1.49)
where the response amplitude A is given by
ε
A= . (1.50)
Ω20 − ω 2

Plotting |A| versus ω gives the familiar resonant response diagram of figure 1.10, in
which the amplitude tends to infinity as ω → Ω0 . (If one actually solves (1.48) at
ω = Ω0 , one obtains a solution whose amplitude grows linearly in time.)
The two effects we have neglected, damping and non-linearity, have two separate
effects on this diagram. If we include only damping, so that

ü + β u̇ + Ω20 u = ε Im eiωt , (1.51)

16
 |A|

Ω0 ω

Figure 1.11: Resonant amplitude response with damping.

then the forced solution is again

u = Im [Aeiωt ], (1.52)

where now
ε
A= , (1.53)
Ω20 + iβω − ω 2
and the presence of the damping term causes a phase shift which caps the response
amplitude, as shown in figure 1.11, since
ε
|A| = ; (1.54)
[(Ω20 − ω 2 )2 + β 2 ω 2 ]1/2

the peak amplitude at resonance is |A| = ε/βω.

The other effect is nonlinearity, which is less easy to deal with. In fact, one can
use perturbation methods to assess its effect in a formal manner, but our present
purpose is more rough and ready. Our idea is this: resonance occurs when the forcing
frequency ω equals the frequency of the underlying oscillator. The difference which
occurs for a nonlinear pendulum is that this frequency (call it Ω) now depends on the
amplitude of the oscillation A: Ω = Ω(A).
To be specific, we again put β = 0, and consider simply the unforced pendulum:

ü + Ω20 sin u = 0. (1.55)

A first (energy) integral is

1 2
2
u̇ + Ω20 (1 − cos u) = E, (1.56)

where E is constant (and depends on amplitude, with E(A) increasing with A). The
phase plane is shown in figure 1.12 and is symmetric about both u and u̇ axes. Thus

17
 .
u

Figure 1.12: Phase plane for the simple pendulum.

a quadrature of (1.56) implies the period P is given by

√ !
2 2 A du
P = , (1.57)
Ω0 0 [cos u − cos A]1/2
where we have used the fact that the amplitude A is given by

E = Ω20 (1 − cos A). (1.58)

From (1.57), we find that the frequency Ω = 2π/P is given by

πΩ0
Ω(A) = ! . (1.59)
√ A
du
2
0 [cos u − cos A]1/2
Ω is a monotonically decreasing function of A in (0, π), with Ω(0) = Ω0 and Ω(π) = 0,
and this is represented as the dotted curve in figure 1.13.
Without now actually solving the forced, damped, nonlinear equation, we can
guess intelligently what happens. For small amplitude oscillations, |A| starts to in-
crease as ω approaches Ω0 ; but as |A| increases, the natural frequency Ω decreases,
and as it is the approach of ω to the natural frequency which is the instrument of
resonance, so the amplitude response curve bends round, as shown in figure 1.13,
to try and approach the dotted Ω(A) curve. Finally, the effect of damping can be
expected to be as in the linear case, to put a cap on the two asymptotes to Ω(A).
Thus, we infer the response diagram shown in figure 1.13, and this is in fact correct.
Moreover, (1.50) suggests A > <
< 0 for ω > Ω0 , i. e., the solution is in phase with the
forcing for ω < Ω0 , and out of phase for ω > Ω0 . Extending this to the nonlinear
case, we infer that at low frequencies, the response is in phase, but that it is out of
phase at high frequencies (as observed).
The response also involves hysteresis (if damping is small enough). If ω is increased
gradually, then at a value ω+ < Ω0 , there is a sudden jump to an out of phase

18
 |A| Ω ( |A| )

out of phase

Ω0 ω
ω− ω
in phase +

Figure 1.13: Nonlinearity bends the resonant response curve, producing hysteresis.

oscillation with higher amplitude. Equivalently, as ω is reduced for this high frequency
response there is a sudden jump down in amplitude to an in-phase oscillation at a value
ω− < ω+ . This response diagram explains what one sees in the simple experiment
and illustrates the important effects of nonlinearity.

1.4 Qualitative methods for partial differential equa-

tions
Any introductory course on partial differential equations will provide the classifica-
tion of second order partial differential equations into the three categories: elliptic,
parabolic, hyperbolic; and one also finds the three simple representatives of these:
Laplace’s equation ∇2 u = 0, governing steady state temperature distribution (for
example); the heat equation ut = ∇2 u, which describes diffusion of heat (or solute);
and the wave equation utt = ∇2 u, which describes the oscillations of a string or of a
drum. These equations are of fundamental importance, as they describe diffusion or
wave propagation in many other physical processes, but they are also linear equations;
however the way in which they behave carries across to nonlinear equations, but of
course nonlinear equations have other behaviours as well.

1.4.1 Waves
In the linear wave equation (in one dimension, describing waves on strings) utt =
c2 uxx , the general solution is u = f (x+ct)+g(x−ct), and represents the superposition
of two travelling waves of speed c moving in opposite directions. In more than one
space dimension, the equivalent model is utt = c2 ∇2 u, and the solutions are functions

19
of (k.x ± ωt), where ω is frequency and k is the wave vector; the waves move in the
direction of the vector k, while the wave speed is then c = ω/|k|.
Even simpler to discuss is the first order wave equation

ut + cux = 0, (1.60)

which is trivially solved by the method of characteristics to give

u = f (x − ct), (1.61)

representing a wave of speed c. The idea of finding characteristics generalises to

systems of the form
Aut + Bux = 0, (1.62)
where u ∈ Rn and A and B are constant n × n matrices. We can solve this system
as follows. The eigenvalue problem

λAw = Bw (1.63)

will in general have n solution pairs (w, λ), where each value of λ is one of the roots
of the n-th order polynomial
det (λA − B) = 0. (1.64)
Suppose the n eigenvalues λi , i = 1, . . . , n, are distinct (which is the general case);
then the corresponding wi are independent, and the matrix P formed by the eigen-
vectors as columns (i.e., P = (w1 , ..., wn )) satisfies BP = AP D, where D is the
diagonal matrix diag(λ1 , ..., λn ). P is invertible, and if we write v = P −1 u, then
AP vt + BP vx = 0, whence vt + Dvx = 0, and the general solution is
(
u = Pv = Pij fj (x − λj t)ei , (1.65)
i,j

where ei is the i-th unit vector, and the functions fj are arbitrary; this represents the
superposition of n travelling waves with speeds λi . This procedure works providing
A is invertible, and also (practically) if all the λi are real, in which case we say the
system is hyperbolic.
More generally, we can use the above prescription to solve the nonlinear equation

Aut + Bux = r(x, t, u), (1.66)

where we allow A and B to depend on x and t also. The diagonalisation procedure

works exactly as before, leading to
∂ ∂
A (P v) + B (P v) = r[x, t, P v]; (1.67)
∂t ∂x
now, however, λ, w and therefore also P will depend on x and t. Thus we find

vt + Dvx = P −1 A−1 r − [P −1 Pt + DP −1 Px ]v, (1.68)

20
and the components of v can be solved as a set of coupled ordinary differential
equations along the characteristics dx/dt = λi .
If A and B depend also on u (the quasi-linear case), the procedure is less simple
for systems. The characterisation of the system as hyperbolic based on the reality of
the eigenvalues of (1.63) is still appropriate, but the diagonalisation and reduction to
the equivalent of (1.68) are less clear. In the particular case where P depends only
∂vi
on u (and not on x and t), and if P −1 is a Jacobian matrix (i. e., (P −1 )ij = for
∂uj
some vector v(u)), then the function v is given by the (well-defined) line integral
!
v = P −1 du, (1.69)

and vt = P −1 ut , vx = P −1 ux ; hence we can derive the diagonalised form

vt + Dvx = P −1 A−1 r. (1.70)

This shows how the characteristic equations can be derived, but in general the equa-
tions can not be solved, since the elements of D will depend on all the components
of v. An example of this type occurs in river flow, and will be discussed in chapter 4.
However, the method of characteristics always works in one dimension, so we now
return our attention to this case. Consider as an example the nonlinear evolution
equation
ut + uux = 0, (1.71)
to be solved on the whole real axis. The method of characteristics leads to the
implicitly defined general solution

u = f (x − ut), (1.72)

which is analogous to (1.61), and represents a wave whose speed depends on its am-
plitude. Thus higher values of u propagate more rapidly, and this leads to the wave
steepening depicted in figure 1.14.
In fact, it can be seen that eventually u becomes multi-valued, and this signifies a
break down of the solution. The usual way in which this multi-valuedness is avoided
is to allow the formation of a shock, which consists of a point of discontinuity of
u. The characteristic solution applies in front of and behind the shock, and the
characteristics intersect at the shock, whose propagation forwards is described by an
appropriate jump condition: see figure 1.15.
This seemingly arbitrary escape route is motivated by the fact that evolution
equations such as (1.71) are generally derived from a conservation law, here of the
form !
d B
u dx = −[ 12 u2 ]B
A, (1.73)
dt A
where the square-bracketed term represents the jump in 12 u2 between A and B. The
deduction of the point form (1.71) from (1.73) required the additional assumption
that u was continuously differentiable; however, it is possible to satisfy (1.73) at a

21
 u

Figure 1.14: Nonlinearity causes wave steepening.

Figure 1.15: Intersection of characteristics leads to shock formation.

22
point of discontinuity of u. Suppose u is discontinuous at x = xS (t), and denote the
jump in a quantity q across the shock by [q]+
− = q(xS + , t) − q(xS − , t). Then by letting
B → xS + , A → xS − , we find that (1.73) implies the jump condition

[ 21 u2 ]+
− 1
ẋS = + = 2 (u+ + u− ). (1.74)
[u]−

An example
We illustrate how to solve a problem of this type by considering the initial function
for u
1
u = u0 (x) = at t = 0. (1.75)
1 + x2
The implicitly defined solution is then
1
u= , (1.76)
1 + (x − ut)2
or, in characteristic form,
1
u = u0 (ξ) = , x = ξ + ut. (1.77)
1 + ξ2
This defines a single-valued function so long as ux is finite everywhere. Differentiating
(1.77) leads to
u#0 (ξ)
ux = , (1.78)
1 + tu#0 (ξ)
$ %
1
and this shows that ux → −∞ as t → tc = min − # . Since −u#0 = 2ξ/(1+ξ 2 )2 ,
ξ : u"0 <0 u0 (ξ)
√ 8
we find the relevant value of ξ is 1/ 3, and thus tc = √ and the corresponding
√ 3 3
value of x is xc = 3. Thus (1.76) applies while t < tc , and thereafter the solution
also applies in x < xS (t) and x > xS (t), where

ẋS = 12 [u(xS +) + u(xS −)], (1.79)

with
√ 8
xS = 3 at t = √ . (1.80)
3 3
As indicated in figure 1.16, the characteristics intersect at the shock, and it is geomet-
rically clear from figure 1.14, for example, that u+ and u− are the largest and smallest
roots of the cubic (1.76). An explicit solution for xS is not readily available, but it is
of interest to establish the long term behaviour, and for this we need approximations
to the roots of (1.76) when t , 1.
We write the cubic (1.76) in the form
" #1/2
x 1 1−u
u= ± . (1.81)
t t u

23
 t
xS

tc

0 x

Figure 1.16: Characteristic diagram indicating shock formation.

We know that u ≤ 1, and we expect xS to tend to infinity as t → ∞, so that we

suppose x , 1. In that case u ≈ x/t if u = O(1), and the next corrective term gives
" #1/2
x 1 t−x
u≈ ± . (1.82)
t t x
This evidently gives the upper two roots for x < t (since they coalesce at u = 1 when
x = t). For large x, the other root must have u ( 1, and in fact
1
u≈
, (1.83)
x2
in order that (1.81) not imply (1.82).3 Alternatively, (1.83) follows from consideration
of (1.76) in the form
t2 u3 − 2xtu2 + (x2 + 1)u − 1 = 0, (1.84)
providing x , t1/3 .
To find the location of the ‘noses’ of the solution, we note that the approximation
that u ≈ x/t breaks down (see (1.82)) when x ∼ t1/3 , which is also where (1.83)
becomes invalid. This suggests writing
x x
u = W (X), X = 1/3 , (1.85)
t t
and then W (X) is given approximately, for large t, by
1
W (W − 1)2 = , (1.86)
X3
and for X = O(1) there are three roots providing X > 3/22/3 ; at X = 3/22/3 , the two
lower roots coalesce at W = 13 : this describes the left nose of the curve.
" #
1
3
We need u ∼ O
< in order that the second term in (1.81) be significant (otherwise we regain
x2
(1.82)), and in fact we need the two terms to be approximately equal, so that 0 < u < 1: hence
(1.83).

24
 2
(W-1) W

4/27
1/X 3

W 1/3 W+ 1 W
-

Figure 1.17: Determination of W (X).

1/2
~ x_t +
u~ _ t−x
x( )
u~ 2/3
~ 1/t
u~ 2
~ 1/x
xs x~
~t x
x~
~3 t 1/3/ 2/3
2

Figure 1.18: Large time solution of the characteristic solution.

As X becomes large, the upper two roots approach W = 1, thus u ≈ x/t, while the
lower approaches zero, specifically W ≈ 1/X 3 , and hence u ≈ 1/x2 : see figure 1.17.
Thus these roots match to the approximations in (1.82) and (1.83). As X becomes
small, the remaining root is given by W ≈ 1/X, so that u ≈ 1/t2/3 , and (1.84) shows
that this is the correct approximation as long as |x| ( t1/3 . The situation is shown
in figure 1.18).
In order to determine the shock location xS , we make the ansatz that t1/3 ( xS (
t, i. e., that the shock is far from both noses. In that case
1 xS
u+ ≈ , u− ≈ , (1.87)
x2S t

and at leading order we have

xS
ẋS ≈ , (1.88)
2t

25
whence
xS ≈ at1/2 , (1.89)
confirming our assumption that t1/3 ( xS ( t.
To determine the coefficient a, we may use the equal area rule, which follows from
conservation of mass, and states that the two shaded areas in figure 1.18 cut off by
the shock are equal. We use (1.85) for the left hand area, and (1.82) for the right
hand area. Then
! at1/2 ! t " #1/2
x 2 t−x
[W+ (X) − W− (X)] dx ≈ dx, (1.90)
3t1/3 /22/3 t at1/2 t x

where W+ and W− are the middle and lowest roots of (1.86), as shown in figure 1.17.
We write x = t1/2 ξ in the left integral and x = tη in the right, and hence we deduce
that ! 1 " #1/2
1−η
a≈ 2 dη = π. (1.91)
0 η

1.4.2 Burgers’ equation

Although the presence of a shock for (1.71) is entirely consistent with the derivation
of the equation from an integral conservation law, nature appears generally to avoid
discontinuities and singularities, and it is usually the case that in writing an equation
such as (1.71), we have neglected some term which acts to smooth the shock, so that
the change of u is rapid but not abrupt.
The most common type of neglected term which provides the necessary smoothing
is a diffusion term, which is manifested in the adjusted equation as a second derivative
term. The resulting equation is known as Burgers’ equation:

ut + uux = κuxx . (1.92)

Sometimes, as for example in the smoothing effect of heat conduction or viscosity

on sonic shock waves, such a term genuinely represents a physically diffusive process
(e. g., diffusion of heat or momentum); sometimes it arises for more subtle reasons,
as for example in the smoothing of waves on rivers (see, for example, the derivation
of equation (4.57) in the discussion of the monoclinal flood wave in chapter 4).
∂ 2n u
More generally, even-order derivative terms of the form (−1)n−1 κ 2n are smooth-
∂x
ing. (This can be seen by the fact that solutions of the resulting linearised equation
∂ 2n u
ut = (−1)n−1 κ 2n have damped solutions exp(ikx + σt) in which σ = −κk 2n .) A
∂x
fourth order smoothing term occurs in the smoothing of capillary waves by surface
tension, for example.
How does the presence of a diffusive term modify the structure of the solutions?
If κ is small, we should suppose that it has little effect, so that shocks would start
to form. However, the neglect of the diffusion term becomes invalid when the deriva-
tives of u become large. In fact, the diffusion term is trying to do the opposite of

26
the advective term. The latter is trying to fold the initial profile together like an
accordion, while the former is trying to spread everything apart. We might guess
that a balanced position is possible, in which the nonlinear advective term keeps the
profile steep, but the diffusion prevents it actually folding over (and hence causing a
discontinuity), and this will turn out to be the case.

Shock structure
We suppose κ ( 1, so that ut + uux ≈ 0, and a shock forms at x = xS (t). Our aim
is to show that (1.92) supports a shock structure, i. e., a region of rapid change for u
near xS from u− to u+ .
To focus on the shock, we need to rescale x near xS , and we do this by writing

x = xS (t) + κX. (1.93)

Burgers’ equation becomes

κut − ẋS uX + uuX = uXX . (1.94)

We expect the characteristic solution (with κ = 0) to be approximately valid far from

xS , and so appropriate conditions (technically, these are matching conditions) are

u → u± as X → ±∞, (1.95)

and we take these values as prescribed from the outer solution (i. e., the solution of
ut + uux = 0 as x → xS ±).
Since κ ( 1, (1.94) suggests that u relaxes rapidly (on a time scale t ∼ κ ( 1)
to a quasi-steady state (quasi-steady, because u+ and u− will vary with t) in which

−ẋS uX + uuX ≈ uXX , (1.96)

whence
K − ẋS u + 12 u2 ≈ uX , (1.97)
and prescription of the boundary conditions implies

K = ẋS u+ − 12 u2+ = ẋS u− − 12 u2− , (1.98)

whence
[ 12 u2 ]+
−
ẋS = , (1.99)
[u]+−

which is precisely the jump condition we obtained in (1.74). The solution for u of
(1.97) is then ) *
u = c − (u− − c) tanh 21 (u− − c)X , (1.100)
where c = ẋS .

27
1.4.3 The Fisher equation
In Burgers’ equation, a wave arises as a balance between nonlinear advection and
diffusion. In Fisher’s equation,

ut = u(1 − u) + uxx , (1.101)

a wave arises as a mechanism for transferring a variable from an unstable steady state
(u = 0) to a stable one (u = 1). Whereas Burgers’ equation balances two transport
terms, Fisher’s equation balances diffusive transport with an algebraic source term.
It originally arose as a model for the dispersal of an advantageous gene within a
population, and has taken a plenary rôle as a pedagogical example in mathematical
biology of how reaction (source terms) and diffusion can combine to produce travelling
waves.
We pose (1.101) with boundary conditions

u → 1, x → −∞,

u → 0, x → +∞. (1.102)
It is found (and can be proved) that any initial condition leads to a solution which
evolves into a travelling wave of the form

u = f (ξ), ξ = x − ct, (1.103)

where
f ## + cf # + f (1 − f ) = 0, (1.104)
and
f (∞) = 0, f (−∞) = 1. (1.105)
In the (f, g) phase plane, where g = −f # , we have

f # = −g,
g # = f (1 − f ) − cg, (1.106)

and a travelling wave corresponds to a trajectory which moves from (1,0) to (0,0).
Linearisation of (1.106) near the fixed point (f ∗ , 0) via f = f ∗ + F leads to
" ## " #" #
F 0 −1 F
= , (1.107)
g 1 − 2f ∗ −c g

with solutions eλξ , where λ2 + cλ + (1 − 2f ∗ ) = 0. We anticipate c > 0; then (1, 0) is

a saddle point, while (0, 0) is a stable node if c ≥ 2 (and a spiral if c < 2). For c ≥ 2,
a connecting trajectory exists as shown in figure 1.19: in practice the minimum wave
speed c = 2 is selected. (Connecting trajectories also exist if c < 2, but because (0,0)
is a spiral, these have oscillating tails as u → 0, which are unstable and also (for
example, if u represents a population) unphysical.)

28
 g
1

0.5

0
f
-0.5

-1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Figure 1.19: Phase portrait of Fisher equation, (1.106), for c = 2. Note how close
the connecting trajectory (thick line) is to the g nullcline. This is why the large c
approximation is accurate for this trajectory.

Explicit solutions for (1.104) are not available, but an excellent approximation is
easily available. We put
ξ = cΞ, (1.108)
so
νf ## + f # + f (1 − f ) = 0, (1.109)
with ν = 1/c2 = 1/4 for c = 2. Taking ν ( 1 and writing f = f0 + νf1 + . . ., we have
f0# + f0 (1 − f0 ) = 0,
f1# + (1 − 2f0 )f1 = −f0## , (1.110)
and thus
e−Ξ
f0 = . (1.111)
1 + e−Ξ
Also, noting that 1 − 2f0 = −f0## /f0# (differentiate (1.110)1 ),
f1 = f0 (1 − f0 ) ln[f0 (1 − f0 )], (1.112)
and so on. Even the first term gives a good approximation, and even for c = 2.

1.4.4 Solitons
The Fisher wave is an example of a solitary travelling wave. Another type of solitary
wave is the soliton, as exemplified by solutions of the Korteweg-de Vries equation
ut + uux + uxxx = 0. (1.113)

29
This has travelling wave solutions u = f (ξ), ξ = x − ct, where

f ### + f f # − cf # = 0, (1.114)

and solitary waves with f → 0 at ±∞ satisfy the first integral

f ## + 12 f 2 − cf = 0, (1.115)

and thus
1 "2
2
f + 16 f 3 − 12 cf 2 = 0, (1.116)
with solution "√ #
3 2 cξ
f= 2
c sech . (1.117)
2
Thus there is a one parameter family of these solitary waves, and they are called
solitons, because they have the remarkable particle-like ability to ‘pass through’ each
other without damage, except for a change of relative phase. Despite the nonlinear-
ity, they obey a kind of superposition principle. Soliton equations (of which there
are many) have many other remarkable properties, beyond the scope of the present
discussion.
Some understanding of the solitary wave arises through an understanding of the
balance between nonlinearity (uux ) and dispersion (uxxx ). The dispersive part of
the equation, ut + uxxx = 0, is so called because waves exp[ik(x − ct)] have wave
speed c = −k 2 which depends on wavenumber k; waves of different wavelengths
(2π/k) move at different speeds and thus disperse. On the other hand, the nonlinear
advection equation ut +uux has a focussing effect, which (from a spectral point of view)
concentrates high wave numbers near shocks (rapid change means large derivatives
means high wavenumber). So the nonlinearity tries to move high wavenumber modes
in from the left, while the dispersion tries to move them to the left: again a balance
is struck, and a travelling wave is the result.

1.4.5 Nonlinear diffusion: similarity solutions

Like travelling wave solutions, similarity solutions are important indicators of solution
behaviour. A particularly illuminating illustration of this behaviour is provided by
the general nonlinear diffusion equation

ut = (um ux )x , (1.118)

which arises in many contexts. We shall illustrate the derivation of this equation for
a fluid droplet below. Typically, (1.118) represents the conservation of the density of
some quantity u with a diffusive flux −um ux . A standard kind of problem to consider
is then the release of a concentrated amount at x = 0 at t = 0. We can idealise this
by supposing that at t = 0 (in suitable units),
! ∞
u = 0 for x '= 0, u(x) dx = 1. (1.119)
−∞

30
This apparently contradictory prescription idealises the concept of a very concentrated
local injection of u. For example, (1.118) with (1.119) could represent the diffusion
of sugar in hot (one-dimensional) tea from an initially emplaced sugar grain. (1.119)
defines the delta function δ(x), an example of a generalised function. One can think
of generalised functions as being (defined by) the equivalence classes of well-behaved
functions un with appropriate limiting behaviour. For example, the delta function is
defined by the class of well-behaved functions un for which
! ∞
un (x)f (x) dx → f (0) (1.120)
−∞

as n → ∞ for all well-behaved f (x). As a shorthand, then,

! ∞
δ(x)f (x) dx = f (0) (1.121)
−∞

for any f , but the ulterior definition is really in (1.120). In practice, however, we
think of a delta function as a ‘function’ of x, zero everywhere except for a (very)
sharp spike at x = 0.
In solving (1.118), we also apply boundary conditions
u → 0 as x → ±∞, (1.122)
and these, together with the equation and initial condition, imply that
! ∞
u dx = 1 (1.123)
−∞

for all time.

A similarity solution is appropriate because there are no intrinsic space or time
scales for the problem. It is in this context that one can expect the solution to look
the same at different times on different scales. In general, as t varies, then the length
scale might vary as ξ(t) and the amplitude of the solution u might vary as U (t). That
is, if we look at u/U as a function of x/ξ, it will look the same for all t. This in turn
implies that the solution takes the form
$ %
x
u = U (t)f , (1.124)
ξ(t)
and this is one of the forms of a similarity solution.
It is often the case that U and ξ are powers of t, and the exponents are to be
chosen so that the problem has such a solution. This is best seen by example. If we
denote η = x/ξ(t), and substitute the form (1.124) into (1.118), (1.122) and (1.123),
we find
U# ξ# Um
f − ηf # = 2 [f m f # ]# , (1.125)
U ξ ξ
where U # = dU/dt, ξ # = dξ/dt, but f # = df /dη. The initial/boundary conditions
become
f (±∞) = 0, (1.126)

31
and the normalisation condition (1.123) is
! ∞
Uξ f dη = 1. (1.127)
−∞

A solution can be found provided the t dependence vanishes from the model, and this
requires U ξ = 1 (the constant can be taken as one without loss of generality), whence
(1.125) becomes
[f m f # ]# + ξ m+1 ξ # (ηf )# = 0, (1.128)
and ξ m+1 ξ # must be constant. It is algebraically convenient to choose ξ m+1 ξ # = 2/m,
thus $ % m+2
1
m
η=x , (1.129)
2(m + 2)t
and a first integral of (1.128) is
2
f mf # + ηf = 0, (1.130)
m
with the constant of integration being zero (because f → 0 as η → ±∞). Thus either
f = 0, or
f = [η02 − η 2 ]1/m , (1.131)
so that the solution has the form of a cap of finite extent, given by+ (1.131) (for
∞
|η| < η0 , and f = 0 for |η| > η0 . The value of η0 is determined from −∞ f dη = 1,
and is
1
η0 = , m .
- m+2 (1.132)
! π/2
m+2
2 cos m θdθ
0

The finite extent of the profile is due to the degeneracy of the equation when m > 0.
(The√limit m → 0 regains
√ the Gaussian solution of the heat equation by first putting
η = mη0 ζ, f = F/ m, and noting that η0 ≈ (πm)−m/2 as m → 0 (this last following
by application of Laplace’s method to (1.132)).) The graph of f (η) is shown in figure
1.20.

1.4.6 The viscous droplet

An example of where the nonlinear diffusion equation can arise is in the dynamics of
a drop of viscous fluid on a level surface. If the fluid occupies the region 0 < z <
h(x, y, t) and is shallow, then lubrication theory gives the approximation

∂2u
∇p = µ ,
∂z 2
pz = −ρg, (1.133)

in which u = (u, v, 0) is the horizontal component of velocity, and ∇ is the horizontal

gradient (∂/∂x, ∂/∂y, 0). With p = 0 at z = h, we have the hydrostatic pressure

32
 f

Figure 1.20: f (η) given by (1.131).

pa = 0

2κ > 0
p
Figure 1.21: The surface shown has positive curvature when the radius of curvature
is measured from below the surface; in this case equilibrium requires p > pa .

p = ρg(h − z), so that ∇p = ρg∇h, and three vertical integrations of (1.133)1 (with
zero shear stress ∂u/∂z = 0 at z = h and no slip u = 0 at z = 0) yield the horizontal
fluid flux ! h
ρg
q= u dz = − h2 ∇h. (1.134)
0 3µ
Conservation of fluid volume for an incompressible fluid is ht + ∇.q = 0, and thus
ρg
ht = ∇.[h3 ∇h], (1.135)
3µ
corresponding to (1.118) (in two space dimensions) with m = 3. A drop of fluid
placed on a table will spread out at a finite rate.
That this does not continue indefinitely is due to surface tension. Rather than
having p = 0 at z = h (where the atmospheric pressure above is taken as zero), the
effect of surface tension is to prescribe
p = 2γκ, (1.136)

33
where γ is the surface tension, and κ is the mean curvature relative to the fluid droplet
(i. e., κ > 0 if the interface is concave4 , as illustrated in figure 1.21). The curvature is
defined as 2κ = ∇.n, where n is the unit normal pointing away from the fluid (i. e.,
upwards). At least this shorthand definition works if we define
(−hx , −hy , 1)
n= ; (1.137)
[1 + |∇h|2 ]1/2
thus $ %
∇h
2κ = −∇. . (1.138)
{1 + |∇h|2 }1/2
It is less obvious that it will work more generally, since there are many ways of defining
the interface in the form φ(x, y, z) = 0 and thus n = ∇φ/|∇φ| (that in (1.137) uses
φ = z − h); but in fact it does not matter, since we may generally take φ = (z − h)P
for some arbitrary smooth function P , so that ∇φ = (−hx , −hy , 1)P on z = h, and
∇φ/|∇φ| is the same expression as in (1.137).
For shallow flows, we replace p = 0 on z = h by p = −γ∇2 h there, and thus

p ≈ ρg(h − z) − γ∇2 h, (1.139)

and (via (1.134)), (1.135) is modified to

$ 3 %
h 2
ht = ∇. ∇{ρgh − γ∇ h} . (1.140)
3µ
The fourth order term is also ‘diffusive’, insofar as it is a smoothing term, as already
mentioned: high wave number (high gradient) modes are rapidly damped. The effect
of surface tension relative to the diffusional gravity term is given by the Bond number
ρgl2
Bo = , (1.141)
γ
where l is the lateral length scale of the drop. This is the (only) dimensionless
parameter which occurs when (1.140) is written dimensionlessly.

1.4.7 Advance and retreat: waiting times

The similarity solution (1.131) predicts an infinite slope at the margin (where f = 0)
if m > 1 (and a zero slope if m < 1). If one releases a finite quantity at t = 0,
then one expects the long time solution to be this similarity solution. The question
then arises as to how this similarity solution is approached, in particular if the initial
droplet has finite slope at the margin.
This question can be addressed in a more general way by studying the behaviour
near the margin x = xS (t) of a solution h(x, t) of (1.118),

ht = (hm hx )x . (1.142)
4
Geomorphologists would call this surface convex; see chapter 6.

34
Suppose that h ∼ c(xS − x)ν for x near xS . Then satisfaction of (1.142) requires

ẋS ≈ cm [ν(m + 1) − 1](xS − x)νm−1 . (1.143)

Note that the similarity solution (1.131) has ẋS finite when ν = 1/m, consistent with
(1.143), and more generally we see that the margin will advance at a rate ẋS ≈ cm /m
if h ∼ c(xS − x)1/m .
Suppose now that m > 1, and we emplace a droplet with finite slope, ν = 1. Then
the right hand side of (1.143) is zero at x = xS , and thus ẋS = 0: the front does not
move. What happens in this case is that the drop flattens out: there is transport
of h towards the margin, which steepens the slope at xS until it becomes infinite, at
which point it will move. This pause while the solution fattens itself prior to margin
movement is called a waiting time.
Conversely, if m < 1, then the front moves (forward) if the slope is zero there,
and ν = 1/m. If the slope is finite, ν = 1, then (1.143) would imply infinite speed.
An initial drop of finite margin slope will instantly develop zero front slope as the
margin advances.
(1.143) does not allow the possibility of retreat, because it describes a purely
diffusive process. The possibility of both advance and retreat is afforded by a model
of a viscous droplet with accretion, one example of which is the mathematical model of
an ice sheet.5 Essentially, an ice sheet, such as that covering Antarctica or Greenland,
can be thought of as a (large) viscous droplet which is nourished by an accumulation
rate (of ice formed from snow). A general model for such a nourished droplet is

ht = (hm hx )x + a, (1.144)

where a represents the accumulation rate. Unlike the pure diffusion process, (1.144)
has a steady state
) *1/(m+1)
h = 21 (m + 1)a(x20 − x2 ) , (1.145)
where x0 must be prescribed. (In the case of an ice sheet, we might take x0 to be at
the continental margin.) (1.145) is slightly artificial, as it requires a = 0 for x > x0 ,
and allows a finite flux −hm hx = ax0 where h = 0. More generally, we might allow
accumulation and ablation (snowfall and melting), and thus a = a(x), with a < 0 for
large |x|. In that case the steady state is
$ ! x0 %1/(m+1)
h = (m + 1) s dx , (1.146)
x

where the balance function s is ! x

s= a dx, (1.147)
0
and x0 is defined to be where accumulation balances ablation,
! x0
a dx = 0. (1.148)
0
5
Ice sheets and their marginal movement are discussed further in chapter 10.

35
 This steady state is actually stable, and both advance and retreat can occur.
Suppose the margin is at xS , where a = aS = −|aS | (aS < 0, representing ablation).
If we put h ≈ c(xS − x)ν , then (1.144) implies

νcẋS (xS − x)ν−1 ≈ νcm+1 [ν(m + 1) − 1](xS − x)[ν(m+1)−2] − |aS |, (1.149)

and there are three possible balances of leading order terms.

The first is as before,

ẋS ≈ cm [ν(m + 1) − 1](xS − x)νm−1 , (1.150)

and applies generally if ν < 1. Supposing m > 1, then we have advance, ẋS ≈ cm /m
if ν = 1/m, but if ν > 1/m, this cannot occur, and the margin is stationary if
1/m < ν < 1. If ν = 1, then ν(m + 1) − 2 = m − 1 > 0, so that

ẋS ≈ −|aS |/c, (1.151)

and the margin retreats; if ν > 1, then instantaneous adjustment to finite slope and
retreat occurs.
The ice sheet exhibits the same sort of waiting time behaviour as the viscous
droplet without accretion. For 1/m < ν < 1, the margin is stationary, and if xS < x0
then the margin slope will steepen until ν = 1/m, and advance occurs. On the other
hand, if xS > x0 , then the slope will decrease until ν = 1, and retreat occurs. In the
steady state, a balance is achieved (from (1.146)) when ν = 2/(m + 1).

1.4.8 Blow-up
Further intriguing possibilities arise when the source term is nonlinear. An example
is afforded by the nonlinear (reaction-diffusion) equation

ut = uxx + λeu , (1.152)

which arises in the theory of combustion. Indeed, as we saw earlier, combustion occurs
through the fact that multiple steady states can exist for a model such as (1.30), and
the same is true for (1.152), which can have two steady solutions. In fact, if we solve
u## + λeu = 0 with boundary conditions u = 0 on x = ±1, then the solutions are
, ./ 0-
λ
u = 2 ln A sech Ax , (1.153)
2

where A = exp[u(0)/2], and A satisfies

,/ -
λ
A = cosh A , (1.154)
2

which has two solutions if λ < 0.878, and none if λ > 0.878: the situation is depicted
in figure 1.22. If we replace eu by exp[u/(1 + εu)], ε > 0, we regain the top (hot)
branch also, as in figure 1.9.

36
 10 u(0)

λ
0
0 0.5 1 1.5

Figure 1.22: Maximum value of steady solutions u of (1.152), u(0), as a function of

> 0.878.
the parameter λ. Blow-up occurs if λ ∼

One wonders what the absence of a steady state for (1.152) if λ > λc implies.
The time-dependent problem certainly has a solution, and an idea of its behaviour
can be deduced from the spatially independent problem, ut = λeu , with solution
u = ln[1/{λ(t0 − t)}]: u reaches infinity in a finite time. This phenomenon is known
as thermal runaway, and more generally the creation of a singularity of the solution
in finite time is called blow-up. Numerical solutions of the equation (1.152) including
the diffusion term show that blow-up still occurs, but at an isolated point; figure 1.23
shows the approach to blow-up as t approaches a critical blow-up time tc .
In fact, one can prove generally that no steady solutions exist for λ greater than
some critical value, and also that in that case, blow-up will occur in finite time. To
do this, we use some slightly more sophisticated mathematics.
Suppose we want to solve the more general problem

ut = ∇2 u + λeu in Ω, (1.155)

with u = 0 in the boundary ∂Ω, and u = 0 at t = 0 (these conditions are for

convenience rather than necessity). We will be able to prove results for (1.155) which
are comparable to those for the ordinary differential equation version (cf. (1.33))

ẇ = −µ1 w + λew , (1.156)

because, in some loose sense, the Laplacian operator ∇2 resembles a loss term.
More specifically, we recall some pertinent facts about the (Helmholtz) eigenvalue
problem
∇2 φ + µφ = 0 in Ω, (1.157)
with φ = 0 on ∂Ω. There exists a denumerable sequence of real eigenvalues 0 < µ1 ≤
µ2 . . ., with µn → ∞ as n → ∞, and corresponding (real) eigenfunctions φ1 , φ2 , ...

37
 25
u 3.54737770893526
3.56350943288956
3.56383223992489
20 3.56383806154693

15

10

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1

Figure 1.23: Solution of ut = uxx + eu on [−1, 1], with u = 0 at x = −1, 1 and

t = 0. The solution is shown for four times close to the blow-up time, which in
this computation is tc = 3.56384027594971. The many decimal places indicate the
logarithmic suddenness of the runaway as t → tc , but the value of tc itself will depend
on the numerical approximation used.

which form an orthonormal set (using the L2 norm), thus

!
(φi , φj ) ≡ φi φj dV = δij , (1.158)
Ω

where δij is the Kronecker delta (= 1 if i = j, 0 if i '= j). These eigenvalues satisfy a
variational principle of the form
!
µi = min |∇φ|2 dV, (1.159)
Ω
+
where φ ranges over functions of unit norm, 0φ02 = 1+ { φ2 dV }1/2 += 1, which
2 are
2 2
orthogonal to φj for j < i; (more generally µi = min |∇φ| dV / φ dV if φ is
not normalised on to the unit sphere 0φ02 = 1). In particular
!
µ1 = min |∇φ|2 dV, (1.160)
&φ&2 =1 Ω

and the corresponding φ1 is of one sign, let us say positive. +

We take the inner product of the equation (1.155) with φ1 and divide by φ1 dV ;
defining + !
Ω
uφ1 dV
v(t) = + = u dω, (1.161)
φ dV
Ω 1 Ω

38
 + +
where dω = φ1 dV / Ω
φ1 dV is a measure on Ω (with Ω
dω = 1), and using Green’s
theorem, we find !
v̇ = λ eu dω − µ1 v, (1.162)
Ω

and the equation for v is close to the ordinary differential equation (1.156).
Now we use Jensen’s inequality. This says that if we have an integrable function
g(x) on Ω and a convex function f (s) on R (i. e., one that bends upwards, f ## > 0),
then $! % !
f g(x) dω ≤ f [g(x)] dω (1.163)
Ω Ω
+
for any measure ω on Ω such that Ω dω = 1. We have chosen ω to be so normalised,
and eu is convex: thus
! $! %
exp(u) dω ≥ exp u dω = ev , (1.164)
Ω Ω

so that
v̇ ≥ λev − µ1 v. (1.165)
It is now easy to prove non-existence of steady states and blow-up for λ greater
than some critical value λc . Firstly, u must be positive, and hence also v. (For
suppose u < 0: since u = 0 at t = 0 and on ∂Ω, then u attains its minimum in
Ω at some t > 0, at which point ut ≤ 0, uxx ≥ 0, which is impossible, since then
ut − uxx = λeu ≤ 0.) For any v, ev ≥ ev, thus v̇ ≥ (λe − µ1 )v. In a steady state we
must have v̇ = 0, and also v > 0 (since clearly u = 0 is not a steady solution), and
this pair of conditions is impossible if

λ > µ1 /e. (1.166)

This implies non-existence of a steady solution for λ > λc , where λc ≤ µ1 /e.

In a similar vein, if λ > µ1 /e, then

v̇ > µ1 [ev−1 − v], (1.167)

and v > w, where

ẇ = µ1 (ew−1 − w), w(0) = 0. (1.168)
(This is a standard comparison argument: v = w at t = 0, and v̇ > ẇ there, so v − w
is initially positive. It remains so unless at some future time v − w reaches zero again,
when necessarily v̇ − ẇ ≤ 0 — which is impossible, since v̇ > ẇ whenever v = w.)
But w → ∞ in finite time (ẇ > 0 so that w → ∞ as t increases, and as w → ∞,
e−w ẇ ≈ µe−1 , so e−w reaches zero in finite time); therefore also v reaches infinity in
finite time. Finally !
v= u dω ≤ sup u, (1.169)
Ω Ω
+
since Ω
dω = 1: hence u → ∞ in finite time.

39
 In fact u → ∞ at isolated points, and usually at one isolated point. As blow-up
is approached, one might suppose that the nature of the solution in the vicinity of
the blow-up point would become independent of the initial (or boundary) conditions,
and thus that some form of local similarity solution might be appropriate.
This is indeed the case, although the precise structure is rather complicated.
We examine blow-up in one spatial dimension, x. As a first guess, the logarithmic
nature of blow-up in the spatially independent case, together with the usual square-
root behaviour of the space variable in similarity solutions for the diffusion equation,
suggests that we define
x − x0
τ = − ln(t0 − t), η = , u = − ln[λ(t0 − t)] + g(η, τ ), (1.170)
(t0 − t)1/2

where blow-up occurs at x = x0 at t = t0 ; hence g satisfies

gτ = gηη − 12 ηgη + eg − 1. (1.171)

The natural candidate for a similarity solution is then a steady solution g(η) of (1.171),
satisfying
g ## − 12 ηg # + (eg − 1) = 0, (1.172)
and matching to a far field solution u(x, t0 ) would suggest

g ∼ −2 ln |η| as η → ±∞. (1.173)

Solutions of (1.172) with this asymptotic structure do exist as either η → ∞ or

η → −∞ — but not at both ends simultaneously. (1.172) admits even solutions, and
if we restrict ourselves to these, then we may take

g # (0) = 0, g(0) '= 0. (1.174)

(If g(0) = 0, then g ≡ 0 is the solution.) However, it is found that such solutions
have a different asymptotic behaviour as η → ∞, namely
A ) *
g∼− exp 14 η 2 , (1.175)
|η|

and A = A[g(0)] > 0 for g(0) '= 0 (and A(0) = 0), and these cannot match to the
outer solution. If one alternately prescribes (1.173) as η → +∞, for example, then the
solution is asymmetric, and has the exponential behaviour (1.175) as η → −∞. Thus
the appealingly simple similarity structure implied by steady solutions of (1.171) is
wrong (and actually, the solution of the initial value problem (1.171) satisfying (1.173)
tends to zero as τ → ∞).
However, (1.171) itself develops a local similarity structure as τ → ∞, using a
further similarity variable
η x − x0
z= = . (1.176)
τ 1/2 (t0 − 1/2
t) [− ln(t0 − t)]1/2

40
Rewriting (1.171) in terms of z and τ yields
1) *
gτ + 12 zgz + 1 − eg = gzz + 12 zgz . (1.177)
τ
At leading order in τ −1 this has a solution
) *
g = − ln 1 + 14 cz 2 , (1.178)

where c is indeterminate, and this forms the basis for a formal expansion. It is
algebraically convenient to use (1.178) to define c as a new variable, and also to write

s = ln τ ; (1.179)

Then (1.177) becomes

$ & '%
2 2 2 [c + 12 zcz ]2 1
cz = 3 2c + 4zcz + z czz + z − + c + 2 zcz − cs . (1.180)
τz 1 + 14 cz 2

We seek a solution for (1.180) in the form

1 1
c ∼ c0 (z, s) + c1 (z, s) + 2 c2 (z, s)..., (1.181)
τ τ
and then, since τ d/dτ = d/ds, we have
1 1
cs ∼ ċ0 + (ċ1 − c1 ) + 2 (ċ2 − 2c2 ) + ..., (1.182)
τ τ
where ċi ≡ ∂ci /∂s. Substituting this into (1.180) and equating powers of τ , we find

c0 = C0 (s), (1.183)

where C0 is arbitrary, and

$ & '%
2 2 C02
c1z = 3 2C0 + z − + C0 − Ċ0 . (1.184)
z 1 + 14 C0 z 2

The arbitrary function C0 arises because the order of the approximate equation is
reduced. In order to specify it, and other arbitrary functions of s which arise at each
order, we require that the solutions ci be smooth, and this requires that there be no
term on the right hand side of (1.184) proportional to 1/z as z → 0, in order that
logarithmic singularities not be introduced. Specifically, we require at each stage of
the approximation that
∂ci 2 ) *
= 3 a0i + a1i z + a3i z 3 + . . . ; (1.185)
∂z z
so that z 2 ci is smooth. Applying this to (1.184) requires that

Ċ0 = C0 (1 − C0 ), (1.186)

41
so that C0 → 1 as s → ∞, and then
2C0 ) *
c1 = − 2
+ C1 (s) + C02 ln 1 + 14 C0 z 2 . (1.187)
z
At O(1/τ 2 ), we then have
$ &
2 2 2
c2z = 2c1 + 4zc1z + z c1zz + z −(ċ1 − c1 ) + (c1 + 12 zc1z )
z3
'%
2c0 (c1 + 12 zc1z ) 1 2 2 1 2 2
− + 4 c0 c1 z (1 + 4 c0 z ) , (1.188)
1 + 14 c0 z 2

and applying the regularity condition (1.185), we find, after some algebra,

Ċ1 = 2(1 − C0 )C1 + 52 C03 , (1.189)

so that C1 → C10 + 52 s as s → ∞. Thus finally we obtain the local similarity solution

$ & '%
c(x − x0 )2
u ≈ − ln λ t0 − t + , (1.190)
4[− ln(t0 − t)]

where c ≈ C0 (s), s = ln τ = ln[− ln(t0 − t)].

1.4.9 Reaction-diffusion equations

The development of mathematical biology in the last thirty years has led to one
particular pedagogical example of wave and pattern formation, and that is in the
coupled sets of equations known as reaction-diffusion equations. The general type is
∂ui
= fi (u) + ∇.[Dij ∇uj ], (1.191)
∂t
for n reactants u1 , ..., un , where the summation convention (sum over repeated suffixes,
here j) is implied, but much of what is known about the behaviour of such systems
can be illustrated with the two-species equations

ut = f (u, v) + D1 ∇2 u,
vt = g(u, v) + D2 ∇2 v. (1.192)

The phenomena which we find are closely allied to the behaviour of the underlying
dynamical system

u̇ = f (u, v),
v̇ = g(u, v), (1.193)

and we will discuss three types of behaviour: wave trains, solitary waves, and sta-
tionary patterns.

42
Wave trains
One way in which periodic travelling waves, or wave trains, can arise is when the
underlying kinetics described by (1.193) is oscillatory. Diffusion causes the oscillations
to propagate in space, and a periodic travelling wave results. It suffices to consider
components which diffuse equally rapidly, so that we may consider the suitably scaled
equation
wt = f (w) + ∇2 w, (1.194)
where w ∈ Rn .
Suppose that the reaction kinetics admit an attractive limit cycle for the under-
lying system wt = f (w), and denote this as W0 (t), i.e.

W0# = f (W0 ). (1.195)

Suppose further that we look for solutions which are slowly varying in space. We
define slow time and space scales τ and X as
√
τ = εt, X = εx (1.196)

and seek formal solutions of (1.194) in the form w(X, t, τ ), where

wt + εwτ = f (w) + ε∇2 w, (1.197)

and ∇ = ∇X now. Expanding w as

w ∼ w0 + εw1 + . . . (1.198)

leads to

w0t = f (w0 ),
w1t − Jw1 = −w0τ + ∇2 w0 , (1.199)

and so on; here J = Df (w0 ) is the Jacobian of f at w0 . After an initial transient, we

may take
w0 = W0 (t + ψ), (1.200)
where ψ(τ, X) is the slowly-varying phase, and J = Df (W0 ) is a time-periodic matrix.
Thus we find that w1 satisfies

w1t − Jw1 = −(ψτ − ∇2 ψ)W0# + |∇ψ|2 W0## . (1.201)

Note that s = W0# satisfies the homogeneous equation st − Js = 0. It follows that

the solution of (1.201) is

w1 = −t(ψτ − ∇2 ψ)s + |∇ψ|2 u, (1.202)

where ! t
u = M (t) M −1 (θ)J(θ)s(θ) dθ + M (t), (1.203)
0

43
and M is a fundamental matrix for the homogeneous equation, i. e., M # = JM ,
M (0) = I. Floquet’s theorem implies that

M = P etΛ , (1.204)

where P is a periodic matrix of period T (the same as that of the limit cycle W0 ).
We can take the matrix Λ to be diagonal if the characteristic multipliers are distinct,
and since we assume W0 is attracting, the eigenvalues of Λ will all have negative real
part, except one of zero corresponding to s. With a suitable choice of basis, we then
have
(etΛ )ij → δi1 δj1 as t → ∞, (1.205)
i. e., a matrix with the single non-zero element being unity in the first element. In
this case the first column of P is s, i. e., Pi1 = si .
From (1.203), we have
! t
u = P (t) eηΛ P −1 (t − η)J(t − η)s(t − η) dη + M c. (1.206)
0

The effect of the transient dies away as t → ∞, and if we ignore it, then we can take
Mij = si δj1 , whence M c = c1 s, and thus
$! t %
u=s α(η) dη + c1 , (1.207)
0

where the periodic function α is given by6

α = (P −1 )1m Jmj sj . (1.208)

We define the mean of α to be

! T
1
ᾱ = α(η) dη, (1.209)
T 0

so that ! t
β= (α − ᾱ) dη (1.210)
0
is periodic with period T . Then (1.202) is

w1 = [t{−ψτ + ∇2 ψ + ᾱ|∇ψ|2 } + c1 + β]s, (1.211)

and in order to suppress secular terms (those which grow in t), we require the phase
ψ to satisfy the evolution equation

ψτ = ∇2 ψ + ᾱ|∇ψ|2 . (1.212)

This is an integrated form of Burgers’ equation; in one dimension, u = −ψX /2ᾱ

satisfies uτ + uuX = uXX . Disturbances will form shocks, which are jumps of phase
6
We use the summation convention, which implies summation over repeated suffixes.

44
gradient. More generally, if u = −∇ψ/2ᾱ, then (bearing in mind that ∇ × u = 0)
we find
uτ + (u.∇)u = ∇2 u, (1.213)
which is the Navier-Stokes equation with no pressure term. Phase gradients move
down phase gradients, and form defects where the (sub-)characteristics intersect.7
Solutions of (1.212) which vary with X correspond to travelling wave trains. For
example, in one dimension, waves travel locally at speed dX/dt ≈ −(∂ψ/∂X)−1 . In
general, however, the phase of the oscillation becomes constant at long times if zero
flux boundary conditions ∂ψ/∂n = 0 are prescribed at container boundaries, and
wave trains die away. However, this takes a long time (if ε is small), and while spatial
gradients are present, the solutions have the form of waves. For example, target
patterns are created when an impurity creates a local inhomogeneity in the medium.
Suppose the effect of such an impurity is to decrease the natural oscillation period
by a small amount (of O(ε)) near a point, which we take to be the origin. To be
specific, suppose that the impurity is circular, of radius a; then it is appropriate to
specify
ψ = τ + c at R = a, (1.214)
where R is the polar radius and c is an arbitrary constant (it merely fixes the time
origin), and we expect ψ to tend towards the solution ψ = τ − f (R) as t → ∞, where
f satisfies
1
f ## + f # − ᾱf #2 + 1 = 0, (1.215)
R
together with f (a) = c and an appropriate no flux condition at large R; such a
condition can always be implemented by consideration of a small boundary layer near
the boundary. Alternatively, we can restrict attention to a target pattern centred at
the impurity by suppressing incoming waves (this is known as a radiation condition).
The relevant solution if ᾱ > 0 is
1 √
f (R) = ln K0 ( ᾱR), (1.216)
ᾱ
where K0 is the modified Bessel function of the second kind of order zero. The
other Bessel function I0 is suppressed because
√ of the radiation condition (it produces
incoming waves). At large√ R, ψ ∼ −R/ ᾱ, which represents an outward travelling
wave of speed dR/dt ≈ ᾱ. If, on the other hand, ᾱ < 0, then K0 is replaced by a
combination of the Bessel functions J0 and Y0 , and the solution blows up at finite R,
and travelling wave solutions of this type do not exist. More generally, if ψ = βτ on
R = a, then target patterns exist if ᾱβ > 0.
7
Physicists call (1.212) the KPZ equation (after Kardar et al. (1986)). The substitution u =
exp(ᾱψ) reduces it to the diffusion equation for u; this is the Hopf–Cole transformation (see Whitham
1974).

45
 g=0

v
g>0 g<0 f=0
f>0

f<0

Figure 1.24: Phase diagram for kinetics of (1.217).

Activator-inhibitor system
An example of a system supporting travelling wave solutions is the activator-inhibitor
system
ut = f (u, v) + ∇2 u,
vt = g(u, v) + ∇2 v, (1.217)
where the nullclines of the kinetics are as shown in figure 1.24 (cf. figure 1.6). This
system is called an activator-inhibitor system because ∂f /∂v > 0, thus increased
v activates u, while ∂g/∂u < 0, so increased u inhibits v. When the intersection
is on the decreasing part of f = 0, as shown, then ∂f /∂u > 0, ∂g/∂v < 0, and
−fu /fv > −gu /gv , whence the determinant D of the Jacobian of (u, v)T at the fixed
point is positive. Hence the fixed point is unstable if fu + gv > 0, and a limit cycle
exists in this case if trajectories are bounded. For example, if f = F/ε, ε ( 1, this
is the case, and the limit cycle takes the relaxational form shown in figure 1.6. The
addition of diffusion allows travelling wave trains to exist, as described above.

Solitary waves in excitable media

Suppose now the intersection point of the nullclines f = 0 and g = 0 is as shown in
figure 1.25. The fixed point of the underlying dynamical system is now stable, but
relatively small perturbations to v can cause large excursions in u, as shown. When
diffusion is included, these large excursions can travel as solitary waves. The simplest
way to understand how this comes about is if we allow u to have fast reaction kinetics
and take v as having zero diffusion coefficient.

46
 v

f>0

f<0

g<0

g>0

Figure 1.25: Phase plane for excitable kinetics.

In one dimension, a suitably scaled model is then

εut = f (u, v) + ε2 uxx ,
vt = g(u, v), (1.218)
and we look for a travelling wave solution of the form
u = u(ξ), v = v(ξ), ξ = ct − x, (1.219)
where c (assumed positive) is to be found. Then
εcu# = f + ε2 u## ,
cv # = g, (1.220)
and the idea is to seek a trajectory for which (u, v) → (u∗ , v ∗ ) as ξ → ±∞ (here
(u∗ , v ∗ ) is the fixed point of the system). The form of this trajectory is shown in
figure 1.26. On the slow parts of the wave, f ≈ 0 and cv # ≈ g. On the fast parts, we
put ξ = εΞ; then v ≈ constant, and we denote v+ (= v ∗ ) and v− as the corresponding
values of v; v− is unknown (as is c).
On the fast parts of the wave, we define u# = w (where now u# = du/dΞ), so that
u# = w,
w# = cw − f± (u), (1.221)
where f± (u) = f (u, v± ). The graphs of f+ and f− are similar, and are shown in figure
1.27, where we see that construction of the connecting branches P Q and RS requires
that the fixed points P and Q, or R and S, of (1.221) have a connecting trajectory.
In general, this will not be the case, but we can choose c to connect P to Q (since
v+ is known), and then we choose v− to connect R to S (with this same value of c).
The form of the resulting travelling wave is shown in figure (1.28).

47
 v

v+
P Q

S v- R

Figure 1.26: Phase plane for solitary wave trajectory.

+ (u)
f_

v+
P Q

R u
S
v-

Figure 1.27: Phase plane connection for the fast parts of the travelling wave.

48
 R Q
u

P P

x
ξ

Figure 1.28: Spatial form of the travelling wave.

Pattern formation
We have seen that an activator (v)-inhibitor (u) system
u̇ = f (u, v),
v̇ = g(u, v), (1.222)
admits periodic travelling waves when the uniform state is unstable, and solitary
waves when it is stable (and the activator diffuses slowly). Stationary patterns can
occur when a stable steady state of (1.222) is rendered spatially unstable by different
component diffusivities. Suppose that
ut = f (u, v) + uxx ,
vt = g(u, v) + dvxx , (1.223)
is an activator-inhibitor system with fv > 0, gu < 0; the restriction to one spatial
dimension is inconsequential. The parameter d here represents the ratio of activator
to inhibitor diffusivities. Note that when d → 0, we expect solitary wave propagation,
at least for the phase diagram of figure 1.25, where also fu < 0, gv < 0 at the fixed
point.
With the stationary state denoted as (u∗ , v ∗ ), we assume it is stable in the absence
of diffusion; thus assume
T = fu + gv < 0,
∆ = fu gv − fv gu > 0, (1.224)
both evaluated at (u∗ , v ∗ ). We put
" # " ∗ #
u u
= + weσt+ikx ; (1.225)
v v∗

49
linearisation of (1.223) then yields

(M − k 2 D − σ)w = 0, (1.226)

where " # " #

fu fv 1 0
M= , D= . (1.227)
gu gv 0 d
The eigenvalues σ are the roots of

σ 2 − Td σ + ∆d = 0, (1.228)

where

Td = T − (1 + d)k 2 ,
∆d = ∆ − k 2 (dfu + gv ) + dk 4 . (1.229)

The steady state is stable if and only if Td < 0 and ∆d > 0 (cf. figure 1.4). Now
T < 0 and ∆ > 0 by assumption: hence Td < 0, and thus instability occurs if and only
if ∆d < 0. Since ∆ > 0, we see from (1.227) that this can only occur if dfu + gv > 0.
Thus either fu > 0 or gv > 0, and the system cannot be excitable. Since fu + gv < 0,
we see that a necessary condition for instability is that d '= 1. Because d is the ratio
of two diffusivities, this instability is known as diffusion-driven instability (DDI), or
Turing instability, after the originator of the theory.
To be specific, let us suppose the situation to be that of figure 1.24, i. e., fu > 0,
gv < 0: then we require d > 1 for DDI. The precise criterion for instability is that
min ∆d < 0, and, from (1.229), this is

dfu + gv > 2[∆d]1/2 , (1.230)

and this can be reduced to

$ %2
∆1/2 + {fv |gu |}1/2
d> . (1.231)
fu
The resulting instability is direct and not oscillatory (in time), though it is os-
cillatory in space. We can therefore expect stationary finite amplitude patterns to
emerge as the stable solutions, and this is indeed what often occurs.
The form of these putative steady solutions as d becomes large can be studied by
seeking (spatially) periodic solutions of

uxx + f (u, v) = 0,
vxx + ε2 g(u, v) = 0, (1.232)

where we define ε2 = 1/d ( 1.

We begin by seeking solutions with period of O(1). As u varies over distances
of x = O(1), v = v̄ is approximately constant, and thus the equation for u can be
integrated to give the first integral
1 2
u
2 x
+ V (u, v̄) = E, (1.233)

50
where ! u
V (u, v) = f (u, v) du, (1.234)
0
and E is constant.
The forms of the curves f (u, v) = 0 (defining v as a function of u), f (u, v) as a
function of u for various fixed v, and V (u, v) as a function of u are shown in figure
1.29. For constant v, solutions for u will be periodic if they lie in the potential well
of V . Given v̄ and E, these periodic solutions are fully determined, and in particular
their period P is a function of v̄ and E, thus P = P (v̄, E). The choice of v̄ and E
must then be made so that v is periodic. We can choose the origin of x so that u is
maximum there; then in fact u is even, and hence so is g[u(x; v̄, E), v̄]. Integration of
(1.232)2 then yields
! x
2
v = v̄ − ε (x − ξ)g[u(ξ; v̄, E), v̄] dξ, (1.235)
−P/2

where periodicity of v requires that

! P/2
g[u(ξ; v̄, E), v̄] dξ = 0. (1.236)
−P/2

(We also require that

! P/2
ξg[u(ξ; v̄, E), v̄] dξ = 0, (1.237)
−P/2

but this is satisfied automatically since the integrand is odd.)

Given v̄, (1.236) appears to determine E, and thus provide a one-parameter family
of periodic solutions. However, it is unlikely that (1.236) can generally be satisfied
for a given function g. Consideration of figure 1.29 suggests that it is more likely
that, given v̄ and E, satisfaction of (1.236) will depend on the precise location of the
curve g = 0. For a function g(u, v; α) dependent on a single parameter α, such as
g = α − u3 v, this suggests that (1.236) may be satisfied (if at all) for a unique value
of α(v̄, E). Since also P = P (v̄, E), this suggests a one-parameter family of spatially
periodic solutions in which P = P (α).
The other possibility for periodic solutions involves the existence of regions in
which u is constant, separated by boundary layers in which u changes rapidly. In this
case, the longer space scale X = εx comes into play, and the resultant form of the
equations (1.232),

ε2 uXX + f = 0,
vXX + g = 0, (1.238)

is clearly suggestive of a boundary layer structure.

The boundary layers themselves are still described by (1.233), but now we require
that u tends to constants ū+ and ū− as x → ±∞; this requires v̄ to have the particular
value where the local maxima of V (u, v̄) are the same (and these occur at ū− and ū+ ).

51
 20
v
15

v+ 10
v-
v- 5
0
0 1 2 3 4 5 6
u
10
f (u,v) v+
5 v-
0

v-
-5

-10
0 1 2 3 4 5 6

u
4
-
V(u,v)
2

-2

-4

0 1 2 3 4 5 6

Figure 1.29: Definition of the values v± defined by the function f (u, v). The upper
graph shows the curve defined implicitly by f (u, v) = 0 (compare figure 1.24). The
middle graph shows the function f (u, v) as a!function of u for v = v+ , v̄, v− , and
u
the lowest graph is the potential V (u, v) = f (u, v) du for the value of v = v̄
0
corresponding to the middle of these three curves. The choice of v̄ in the figure is
that for which the two maxima of V are equal. The particular function used in the
illustrations is f (u, v) = v−[u3 −8u2 +17u], for which the value of v̄ where the maxima
are equal is v̄ ≈ 7.407; the values of v+ and v− are v+ ≈ 10.879 and v− ≈ 3.935.

52
 20

g=0 f=0
15
u (v)
10 B -
v- A
C
5 u+(v)
D

0
0 1 2 3 4 5 6
u- - u- +

Figure 1.30: The nullclines f (u, v) = 0 and g(u, v) = 0. The f nullcline defines locally
two functions u± (v). During the oscillation, v moves from A to B and back to A,
while g > 0, and similarly from C to D and back to C while g > 0. When v reaches
v̄, a boundary layer in u switches the solution between its two branches.

For the value of E equal to this maximum, there are then boundary layer solutions
in which either u goes from ū− to ū+ as x increases, or from ū+ to ū− .
The periodic solutions are filled out by solving

vXX + g(u, v) = 0, (1.239)

in which u is determined by f (u, v) = 0. There are two branches of the resultant

function u(v), which we denote by u− (v) and u+ (v) (and u± (v̄) = ū± ), as indicated
in figure 1.30; if we define
 ! v

 g [u− (v # ), v # ] dv # for v̄ < v, v+ ,

W (v) = ! v̄v (1.240)

 # # #
 g [u+ (v ), v ] dv for v− < v < v̄,
v̄

then W is a V -shaped function defined in [v− , v+ ], with a minimum at v = v̄. Solutions

for v are determined from
1 2
v + W (v) = F,
2 X
(1.241)
for constant F , and determine a one-parameter family of periodic solutions. Note
that this family occurs for a fixed choice of f and g, and the parameter can be taken
to be the period. This family is then naturally interpreted as the continuation to large
d of the bifurcating family dependent on wave number which arises when (1.231) is
satisfied.

53
1.5 Notes and references
Modelling
By mathematical modelling, I mean the formulation of a problem in mathematical
terms. If the process is continuous, usually the model will take the form of differential
equations, and in this book we further confine ourselves to deterministic models, as
opposed to stochastic models. Stochastic models are of increasing popularity, aiming
as they do to represent the noisiness of a system, but they can also be something of
an excuse to sweep things we don’t understand under the carpet.
The original classic book which set out the applied mathematician’s stall is that
by Lin and Segel (1974). It contains the ethos of applied mathematics, but retained
a somewhat austere choice of applications. Another classic book which dealt much
more with practical (mostly industrial) applications is that by Tayler (1986). My own
book (Fowler 1997) is in a similar spirit.
These books, certainly the latter two, are aimed at graduate level. There are a
number of books which deal more gently, but still genuinely, with modelling. The
classic of this type is perhaps that by Haberman (1998), a reprinted edition of his
1977 text. More recent books in this direction are those by Fowkes and Mahony
(1994), Howison (2005), and Holmes (2009).

Asymptotics and perturbation theory

Like modelling, there are many books on perturbation methods. To my mind, the
pre-eminent ones are those by Kevorkian and Cole (1981), Bender and Orszag (1978)
and Hinch (1991). Van Dyke’s (1975) book is also a classic. Other well known books
are those by Nayfeh (1973) and Holmes (1995).
The flavours of these books are subtly different. Bender and Orszag’s blockbuster,
taught at M. I. T. in a one-semester course (the whole book), has as its central part
the asymptotic study of boundary layers. The book has the novelty of giving many
numerical illustrations of how good (or bad) the approximations are, and when they
appear to break down.
Kevorkian and Cole’s book (an expanded edition of Cole’s original 1968 mono-
graph) focusses more on multiple scale methods, and takes these to levels of sophisti-
cation a good deal beyond more elementary texts, and there are expositions of some
classic problems: the derivation of the Korteweg-de Vries equation describing long
waves in shallow water, and the relaxational van der Pol oscillator, for example.
Van Dyke’s book is slightly more formal in nature, and mostly concerned with
fluid mechanics. It is one of the few places where one can learn the method of
strained coordinates, a method which is particularly useful in dealing with the motion
of margins and fronts. Hinch’s and Nayfeh’s books include a chapter on strained
coordinates also, as well as the other staple contents. Hinch’s is short, to the point,
succinct. Holmes’s book includes a chapter on homogenisation.

54
Combustion, nonlinear diffusion and blow-up
Two early accounts of combustion and exothermic reactions are those by Aris (1975)
and Buckmaster and Ludford (1982). The first of these largely deals with reaction
in (solid) permeable catalysts, while combustion theory of the second tends to deal
with gaseous combustion, where the theory has all the complication of compressible
gas dynamics together with the species reaction kinetics. A more mathematical book
is that by Bebernes and Eberly (1989). Other books on this subject include those
of Williams (1985), Barnard and Bradley (1985) and Glassman (1987), the latter
two more descriptive than Williams’s voluminous work. A similar analytic approach
is that by Liñán and Williams (1993), but this book is more concise than that of
Williams. Combustion really applies to any reaction, but by convention refers specif-
ically to reactions where there is a large change of temperature. If this is such that
the reactants become luminous, we have a flame. If the change of temperature is
rapid, we have a thermal explosion. Since in gases, increase of temperature is associ-
ated with increase of pressure, explosions tend to be associated with shock waves, or
detonation waves, and this is the explosive ‘blast’.
The classical treatment of thermal explosions (in solids) is much as described in
section 1.4.8, and involves the positive feedback associated with exothermic reactive
heating, which causes the runaway. Explosive runaway can also be caused by au-
tocatalytic feedback in the reaction scheme, much as in a nuclear explosion; this is
the ‘chain’ reaction. Systems with autocatalysis are also prone to oscillatory bifurca-
tions and waves, and are dealt with in the book by Gray and Scott (1990). Ignition
of explosions may be caused by impact or friction (as in striking a match). Both
events cause a localised hotspot to occur, that of impact being due to the sudden
compression of small gas bubbles, see Bowden and Yoffe (1985).
Reactions in a diffusive flame (i. e., one where fuel and oxidant are not pre-mixed)
can be analysed using large activation energy asymptotics; the reactions occur in a
narrow front which spreads as a deflagration wave, whose speed is less than the sound
speed, and is rate-limited by the supply of reactant to the front. The detonation wave
is a reactive shock wave, in which the reaction is triggered not by supply of reactant,
but by gas compression and consequent heating within the shock.
The book by Samarskii et al. (1995) provides a wealth of information about non-
linear diffusion equations, and their associated solution properties of compact support
and blow up. The asymptotic description given here of the local similarity structure
for the blow up of solutions of ut = uxx + λeu is based on that of Dold (1985).

Burgers’ equation
Burgers’ equation relates to a model introduced by Burgers (1948) to describe tur-
bulence in fluid flow in a pipe. In its original form, his model is given by the pair of

55
equations
! b
dU νU 1
b = P− − v 2 dy,
dt b b 0
2
∂v ∂v Uv ∂ u
+ 2v = + ν 2. (1.242)
∂t ∂y b ∂y
This is a toy model which aims to mimic the classical procedure of Reynolds aver-
aging, leading to an evolution equation for the mean flow U (t), and another for the
fluctuating velocity field v(y, t). The cross stream variable is y, and the width of the
‘pipe’ is b. Burgers’ equation follows from the assumption that U = 0, and arises in
the original paper as an approximation to describe the transition region near shocks;
Burgers gives the travelling wave front solution for this case. A thprough discussion
of Burgers’ equation is given by Whitham (1974).

Fisher’s equation
The geneticist R. A. Fisher wrote down his famous equation (Fisher 1937) to describe
the propagation of an advantageous gene in a population situated in a one dimensional
continuum — Fisher had in mind a shore line as an example. The genes (or more
properly alleles, i. e., variants of genes), reside in the members of a population, and the
proportion of different alleles of any particular gene is described by Hardy–Weinberg
kinetics. If one allele has a slight evolutionary advantage, then its proportion p will
vary slowly from generation to generation, and its rate of change is given in certain
circumstances by the logistic equation ṗ = kp(1−p). The effect of diffusion allows the
genes to migrate through the migration of the carrier population. See Hoppensteadt
(1975) for a succinct description. Fisher did not bother with all this background,
but simply wrote his equation down directly. As well as this paper, he authored or
co-authored eight other papers in the same volume, as well as being the journal editor!

Solitons
There are many books on solitons. An accessible introduction is the book by Drazin
and Johnson (1989), and a more advanced treatment is that of Newell (1985). The
subject is rich and fascinating, as is also the curious discovery of the ‘first’ soliton,
or ‘great wave of translation’ by John Scott Russell in 1834, as he followed it on
horseback along the Edinburgh to Glasgow canal. The Korteweg–de Vries equation
which appears successfully to describe such waves was introduced by them much later
(Korteweg and de Vries 1895), by which time they are referred to as solitary waves.
Korteweg and de Vries also wrote down the periodic (but unstable) cnoidal wave
solutions.
There are many other equations which are now known to possess soliton solutions,
and their folklore has crept into many subjects. Under the guise of ‘magmons’, for
example, they have appeared in the subject of magma transport, which we discuss in
chapter 9.

56
Reaction–diffusion equations
Any book on mathematical biology (and there are a good number of these) will discuss
reaction–diffusion equations. The gold standard of the type is the book (now in two
volumes) by Murray (2002), which also contains much other subject matter. A more
concise book just on reaction–diffusion equations is that by Grindrod (1991). These
books span the undergraduate/graduate transition. The book by Edelstein-Keshet
(2005) is gentler, and aimed at a lower level.
Kopell and Howard (1973) and Howard and Kopell (1977) studied waves in reaction-
diffusion equations using the ideas of bifurcation theory and multiple scales. Keener
(1980, 1986) studied spiral wave formation in excitable media, using as a template a
singularly perturbed pair of equations, essentially of Fitzhugh–Nagumo type.
Meinhardt (1982) studied pattern formation in reaction-diffusion systems, and
later (Meinhardt 1995) studied the relation between a suite of mathematical models
and actual observed patterns on sea shells. The comparison is striking as well as
pictorially sumptuous.

Exercises
1.1 Suppose $ %
∂T
Pe + u.∇T = ∇2 T + 1 in D,
∂t
with

T = 0 on ∂D,
T = ΛΘ(x) in D at t = 0,

and Θ = O(1), Λ , 1, P e ( 1. Discuss appropriate scales for the various

phases of the solution.

1.2 The differential equation

ẋ = a − xe−x , x > 0, a > 0,

may have 0, 1 or 2 steady states. Determine how these depend on a, and describe
how solutions behave for a > e−1 and a < e−1 , depending on the value of x(0).

1.3 Each of the equations

z 5 − εz − 1 = 0,
εz 5 − z − 1 = 0,
has five (possibly complex) roots. Find leading order approximations to these
if ε ( 1. Can you refine the approximations?

57
1.4 u and v satisfy the ordinary differential equations
u̇ = k1 − k2 u + k3 u2 v,
v̇ = k4 − k3 u2 v,
where ki > 0. By suitably scaling the equations, show that these can be written
in the dimensionless form
u̇ = a − u + u2 v,
v̇ = b − u2 v,
where a and b should be defined. Show that if u, v are initially positive, they
remain so. Draw the nullclines in the positive quadrant, show that there is a
unique steady state and examine its stability. Are periodic solutions likely to
exist?
1.5 The relaxational form of the van der Pol oscillator is
εẍ + (x2 − 1)ẋ + x = 0, ε ( 1.
A suitable phase plane is spanned by (x, y), where y = εẋ+ 13 x3 −x. Describe the
motion in this phase plane, and find, approximately, the period of the relaxation
oscillation. What happens if ε < 0?
1.6 Find a scaling of the combustion equation
" #
dT E
c = −k(T − T0 ) + A exp − ,
dt RT
so that it can be written in the form
θ̇ = θ0 − g(θ),
where θ0 = RT0 /E and g = θ − αe−1/θ . Give the definition of α. Hence show
that the steady state θ is a multiple-valued function of θ0 if α > 14 e2 .
Find approximations to the smaller and larger positive roots of x2 e−x = ε,
where ε is small and positive. Hence find the approximate range (θ− , θ+ ) of θ0
for which there are three steady solutions.
Suppose that α > 14 e2 , and θ0 varies slowly according to

θ̇0 = ε(θ∗ − θ),

where ε ( 1. Show that there are three possible outcomes, depending on the
value of θ∗ , and describe them.
1.7 A forced pendulum is modelled by the (dimensional) equation
lθ̈ + k θ̇ + g sin θ = α sin λt.
By non-dimensionalising the equation, show how to obtain (1.47), and identify
the parameters ε, β, Ω0 and ω.

58
 1.8 It is asserted after (1.59) that Ω(A) is a decreasing function of A for 0 < A < π,
or equivalently, that the function
! A
1 du
p(A) = √
2 0 [cos u − cos A]1/2
is increasing. Show that this is true by writing p in the form
! 1" #1/2 " #1/2
θ φ dw
p=
0 sin θ sin φ (1 − w2 )1/2
for some functions θ(w, A) and φ(w, A), and using the fact that θ/ sin θ is an
increasing function of θ in (0, π).
" # " #
A−u A+u
[Hint: cos u − cos A = 2 sin sin .]
2 2
1.9 A simple model for the two-phase flow of two fluids along a tube is
αt + (αv)z = 0
−αt + [(1 − α)u]z = 0
ρg [(αv)t + (αv 2 )z ] = −αpz ,
ρl [{(1 − α)u}t + {Dl (1 − α)u2 }z ] = −(1 − α)pz ,
where p is pressure, u and v are the two fluid velocities, α is the volume fraction
of the fluid with speed v, ρg is its density, and ρl is the density of the other
fluid. Show that there are two characteristic speeds dz/dt = λ, satisfying
(λ − u)2 = (Dl − 1)[u2 + 2u(λ − u)] − s2 (λ − v)2 ,
where $ %1/2
ρg (1 − α)
s= .
ρl α
Deduce that the characteristic speeds are real if, when Dl − 1 ( 1, s ( 1,
& '2
> 1+ s(u − v)
Dl ∼ .
u
In particular, show that the roots are complex if Dl = 1 and u '= v. What does
this suggest concerning the well-posedness of the model?
1.10 The function u(x, t) satisfies
ut + uux = α(1 − u2 )
for −∞ < x < ∞, with u = u0 (x) at t = 0, and 0 < u0 < 1 everywhere. Show
that the characteristic solution can be written parametrically in the form
u0 (s) + tanh αt sech αt
u= , exp[α(x − s)] = .
1 + u0 (s) tanh αt 1 − u tanh αt

59
 Sketch the form of the characteristics for an initial function such as u0 (s) =
a/(1 + s2 ). Show that, in terms of s and t, ux is given by

[αsech 2 αt]u#0 (s)

ux = ,
[1 + u0 (s) tanh αt][α + {u#0 (s) + αu0 (s)} tanh αt]
and deduce that a shock will form if u#0 + α(1 + u0 ) becomes negative for some
s. Show that if u0 = a/(1 + s2 ) and a is small, this occurs if
√
< 3a 3
α∼ .
8

1.11 Discuss the formation of shocks and the resulting shock structure for the equa-
tion
ut + uα ux = ε[uβ ux ]x ,
where α, β > 0, and ε ( 1. (Assume u > 0, and u → 0 at ±∞.)
Show that the equation
ut + uux = εuuxx
admits a shock structure when ε ( 1, but that the shock speed is not given by
ẋS = 12 (u+ + u− ) (cf. (1.74)). Why should this be so?

1.12 Use phase plane methods to study the existence of travelling wave solutions to
the equation
ut = up (1 − uq ) + [ur ux ]x ,
when (i) p = 1, q = 2, r = 0; (ii) p = 1, q = 1, r = 1.

1.13 Two examples of integrable partial differential equations which admit soliton
solutions are the nonlinear Schrödinger (NLS) equation

iut = |u|2 u + uxx ,

and the sine-Gordon equation

utt − uxx = sin u.

Show that these equations admit solitary wave solutions (which are in fact
solitons).

1.14 Write down the equation satisfied by a similarity solution of the form u = tβ f (η),
η = x/tα , for the equation

ut = (um ux )x in 0 < x < ∞,

where m >!0, with um ux = −1 at x = 0, u → 0 as x → ∞, u = 0 at t = 0.

∞
Show that f dη = 1, and hence show that in fact f reaches zero at a finite
0
value η0 . Is the requirement that m > 0 necessary?

60
1.15 u satisfies the equation

ut = [D(u)ux ]x in 0 < x < ∞,

with u = 0 at x → ∞ and t = 0. For a general function D (not a power of u),

for what kind of boundary condition at x = 0 does a similarity solution exist?
What if, instead, D = D(ux )? Write down suitable equations and boundary
conditions for the similarity function in each case.

1.16 The depth of a small droplet, h, satisfies the surface-tension controlled equation
γ
ht = − ∇.[h3 ∇∇2 h].
3µ
!
Suppose that a small quantity h dA = M is released at time zero at the origin.
Find a suitable similarity solution in one and two horizontal spatial dimensions.

1.17 A gravity-driven droplet of fluid spreads out on a flat surface. Its viscosity µ is
a function of shear rate, so that a lubrication approximation leads to the model
for its depth h, shear stress τ and velocity u:
∂τ
ρg∇h = ,
∂z
∂u
= A|τ |n−1 τ .
∂z
(A constant viscosity fluid has n = 1.) Show that the horizontal fluid flux is

A(ρg)n
q=− |∇h|n−1 hn+2 ∇h,
n+2
and deduce that
∂h A(ρg)n
= ∇.[hn+2 |∇h|n−1 ∇h].
∂t n+2
Non-dimensionalise the model, assuming initial emplacement of a finite volume
M at the origin, and find similarity solutions in one and two dimensions for the
depth. What happens as n → ∞ or n → 0?

1.18 The depth h of a symmetric (two-dimensional) droplet under the influence of

gravity and surface tension is described by the dimensionless equation
$ & '%
∂h ∂ 3 ∂ ∂2h
= h Bh − 2 ,
∂t ∂x ∂x ∂x

subject to the conditions

! x0
h = 0, hx = ∓ tan θ at x = ±x0 , h dx = A,
−x0

61
where θ is the contact angle. Show that there is a steady state solution h =
h0 u(x), in which ! 1
du √
1/2
= B|x|,
u [(1 − u)(ρ − u)]

the coefficient ρ is defined by

tan2 θ
ρ= ,
Bh20
and the maximum depth h0 is given by
√
A B
= I(ρ),
2h0
where ! 1
u du
I(ρ) = .
0 [(1 − u)(ρ − u)]1/2

By considering (graphically) both sides of the equation for h0 as functions of

ρ, show that there is a unique value of h0 satisfying this equation, and thus a
unique solution for h.
By evaluating the integrals explicitly, show that
7√ 8
2 Bx
u = 1 − (ρ − 1) sinh ,
2

and that ρ is determined by

"√ #
AB ρ+1 ρ+1
= −1 + √ ln √ .
2 tan θ 2 ρ ρ−1

Find explicit approximations for h when AB , tan θ and AB ( tan θ, and

hence show that the margin positions are approximately given by
 √

 A B

 , AB , tan θ,
 tan θ
x0 = " #1/2



 3A
 , AB ( tan θ.
2 tan θ

[Note that if θ is the actual contact angle, then implicitly, the depth scale and
lateral length scale have been taken equal, and the derivation of the equation for
h via lubrication theory is only self consistent if h0 ( 1 or x0 , 1. Since a
length scale can be prescribed from the initial droplet size, we can choose A = 1
without loss of generality. We can then find conditions on B and tan θ which
ensure self consistency.]

62
1.19 Let u satisfy
ut = λup + uxx ,
with u = 1 on x = ±1 and t = 0. Prove that if λ is large enough, u must blow
up in finite time if p > 1. Supposing this happens at time t0 at x = 0, show
that a possible local similarity structure is of the form
f (ξ) x
u= β
, ξ= ,
(t0 − t) (t0 − t)1/2

and prove that β = 1/(p − 1). Show that in this case, f would satisfy

f ## − 12 ξf # + λf p − βf = 0,

and explain why appropriate boundary conditions would be

f ∼ |ξ|−2β as ξ → ±∞,

and show that such solutions might be possible. Are any other limiting be-
haviours possible?

1.20 When an oscillatory reaction-diffusion system has an imperfection of size com-

parable to, or larger than, the wave length, then spiral waves can occur. This
is because the wave trains need not be in phase round the boundary of the
obstacles. For example, consider a slowly varying system (1.194) with solutions
w ≈ W0 (t + ψ), where ψ satisfies the equation

ψτ = ∇2 ψ + ᾱ|∇ψ|2 .

Suppose that the imperfection is of radius a, and that the effect of the surface
is to alter the period, so that we take ψ = βτ + mθ + c on r = a, where m is an
integer (so that w is single valued, if we suppose the period of W0 is normalised
to be 2π); c is an arbitrary constant, which we can choose for convenience.
Put ψ = βτ + mθ − φ(r), and show that φ satisfies the equation
$ %
## 1 # #2 m2
φ + φ − ᾱ φ + 2 + β = 0.
r r
Hence show that
1
φ = − ln w(λr),
ᾱ
where w(z) satisfies Bessel’s equation in the form
$ %
## 1 # ν2
w + w + s − 2 w = 0, (∗)
z z
providing we choose

λ = |ᾱβ|1/2 , ν = i|ᾱm|, s = − sgn (ᾱβ).

63
 The solutions of (∗) when s = 1, i. e., ᾱβ < 0, are the Hankel functions
" #1/2
2 ) 9 :*
Hν(1,2) (z) = Jν (z) ± iYν (z) ∼ exp ±i z − 12 νπ − 14 π
πz

as z → ∞. If ᾱβ > 0, so that s = −1, then the solutions are the modifed Bessel
functions Iν (z) and Kν (z), and we have

1 ; π <1/2
Iν ∼ √ ez , Kν ∼ e−z
2πz 2z

as z → ∞.8
Deduce that solutions of this type exist if ᾱβ > 0, and that in this case the
presumption of outward travelling waves (the radiation condition) requires us
to choose w = Kν (z). Show that as r → ∞ in this case,
, " #1/2 -
β
w ≈ W0 t + βτ + mθ − r + O(ln r) .
ᾱ

This solution represents a spiral wave. Note that the integer m is unconstrained.
Its specification would require a model for the reaction on the surface of the
impurity at r = a. It is plausible to imagine that such angle dependent phases
arise through bifurcation of the surface reaction model as the impurity size
increases.

1.21 The Fitzhugh-Nagumo equations are

εut = u(a − u)(u − 1) − v + ε2 uxx ,

vt = bu − v,

where 0 < a < 1, ε ( 1, and α is positive and large enough that u = v = 0 is

the only steady state. Show that the system is excitable, and show, by means
of a phase plane analysis, that solitary travelling waves of the form u(ξ), v(ξ),
ξ = ct − x, are possible with c > 0 and u, v → 0 as ξ → ±∞.

1.22 u and v satisfy the equations

δut = ε2 uxx + f (u, v),

vt = vxx + g(u, v),

where
f (u, v) = u[F (u) − v], g(u, v) = v[u − G(v)],
and F (u) is a unimodal function (F ## < 0) with F (0) = 0, while G(v) is mono-
tone increasing (G# > 0) and G(0) > 0, and there is a unique point (u0 , v0 ) in
8
See Watson (1944, pp. 199 f.) for these results.

64
the positive quadrant where f (u0 , v0 ) = g(u0 , v0 ) = 0, and F # (u0 ) < 0. (For
example F = u(1 − u), G = 0.5 + v.)
Examine the conditions on δ and ε2 which ensure that diffusive-driven instability
of (u0 , v0 ) occurs.
If the upper and lower branches of F −1 are denoted as u+ (v) > u− (v), explain
why u− is unstable when ε ( 1. By constructing phase portraits for v when
u = 0 and when u = u+ (v), and ‘gluing’ them together at a fixed value v = v ∗ ,
show that spatially periodic solutions exist which are ‘patchy’, in the sense that
u alternates rapidly between u+ (v) and 0.

65