Stability Theory∗(Math45031/65031)

Dr. Joel Daou

University of Manchester

First Semester 2021/22

(Previously Math45132/65132, second semester)

Introduction
This is an interesting and useful course due to its applicability to most scientific prob-
lems described by time-dependent differential equations as well as to the mathematical
approach and tools used. The mastery of this approach and these tools, mainly an-
alytical tools, is our main focus as it will empower you to tackle a large variety of
stability problems. A stability problem may be typically formulated as follows. Given
a solution (often a stationary solution) to a mathematical time-dependent problem
(often a PDE representing a physical model), is this solution robust in the presence
of small perturbations? This means, roughly, that if a ‘small’ perturbation is added
to this solution, will this result in a new solution which remains ‘close’ to the original
one as time increases to infinity? If the answer is no for any such perturbation, the
solution is said to be unstable; otherwise, it is said to be stable. Of course, we shall
make this statement more precise mathematically.
From a physical point of view, whether a solution is stable or not is crucial: an
unstable solution is unlikely to be observed in Nature; for example case (b) in fig-
ure 0.1. This observation is encapsulated by the following often cited quotation from
the opening of chapter 3 of the book on Fluid Mechanics by Landau and Lifshitz:
‘Yet not every solution of the equations of motion, even if it is exact, can actually
occur in Nature. Those which do must not obey the equations of fluid dynamics, but
also be stable’. The stability of a given solution is especially relevant in non-linear
∗ All feedback to [email protected]

1
 (d) linearly stable,
(a) stable (b) unstable (c) neutrally stable non-linearly unstable

Figure 0.1: Equilibrium position of a smooth ball.

systems where multiple solutions may exist for the same value of the physical param-
eters. Again only stable solutions are expected to be observed, although unstable
ones are not unimportant as they often play a crucial part in determining the basin
of attraction of stable solutions.
In this course, we shall often use simple mathematical models to illustrate ideas
and techniques. Many applications will be considered however. These are mostly
drawn from fluid mechanics, given that hydrodynamic stability theory is well devel-
oped. However we shall use models from other disciplines involving transport phe-
nomena, reaction-diffusion equations, combustion theory and mathematical biology.
The course will be in three parts. Part I is concerned with basic stability problems
in the context of PDEs and ODEs, deliberately without worrying about physical
applications in order to minimise distraction. Part II will be concerned with linear
stability theory (i.e. stability to small perturbations), applied to applications, drawn
essentially, but not exclusively, from fluid mechanics. Part III is an introduction to
weakly non-linear stability theory.
The course is largely self-contained, with reminders of the needed background
material peppered throughout the text. The background assumed of students is some
basic knowledge of PDEs (method of separation of variables, Fourier series, Fourier
and Laplace integrals,...), ODEs (standard first and second year university material),
and a basic knowledge of Fluid Mechanics (material derivative, flux, conservation
equations including the Navier-Stokes equation,...).

2
1 Basic ideas and simple examples
In most cases, we are dealing with mathematical models represented by a (non-linear)
evolution equation of the form1
ut = N(u) , (1.1)
where N is an operator which is typically non-linear and differential and u a depen-
dent variable which is vector quantity typically depending on time t, space x and a
parameter µ; u = u(x, t; µ). If there is no dependence on x, we have an ODE prob-
lem which is subject to an initial condition (IC), i.e., an Initial Value Problem (IVP).
Otherwise, we have an Initial and Boundary Value Problem (IBVP) with additional
boundary conditions prescribed on the boundary of the x-domain; if the x-domain
is unbounded we typically require the magnitude of u to be finite. Assume that ū
is a stationary solution of the problem, satisfying boundary conditions (if any). Let
u = ū + ũ be a solution obtained by adding to ū a perturbation ũ. Using (1.1), and
noting that ut = ũt since ūt = 0, we see that ũ is governed by
ũt = N(ū + ũ) (non-linear perturbation equation). (1.2)
This is to be solved subject to an IC, say ũ = ũ0 at t = 0, and BCs which follow from
the BCs associated with (1.1) (if any are given). This initial (and boundary) value
problem for ũ is called the non-linear stability problem.
The method of linear stability is based on the assumption that the perturbation
is small, at least initially. As long as this assumption remains valid, the non-linear
stability problem can be linearised (i.e., approximated by a linear problem). To this
end, let ũ = ǫv where ǫ is a small positive number representing the amplitude of the
(initial) perturbation. Then
N(ū + ũ) = N(ū) +ǫLv + o(ǫ) , (1.3)
| {z }
=0

where L is a linear operator representing the differential of N(u) evaluated at u = ū

and o(ǫ) represent terms of order smaller than ǫ. Then, neglecting the o(ǫ) terms,
equations (1.2) and (1.3) imply that
vt = Lv (linear perturbation equation) . (1.4)
This is to be solved subject to an IC, say v = v0 at t = 0. If boundary conditions
are present in the original problem, then these also need be linearised; this results in
1
Partial derivatives will be indicated by subscripts or two other standard notations; e.g. ut =
∂u
∂t = ∂t u.

3
homogeneous BCs for v. Therefore, we have again an initial (and boundary) value
problem for v, with homogeneous BCs (if any), which is called the linear stability
problem.
If any of the solution to (1.4) grows in amplitude without bound, then the sta-
tionary solution ū is said to be unstable. Otherwise, it is said to be stable. If all
solutions decay in amplitude to zero, ū is said to be asymptotically stable.
Note that the parameter ǫ, introduced somewhat artificially, has dropped out from
the linear stability problem. However, its introduction is recommended, as a useful
book-keeping device to derive approximations (linear or non-linear) to the operator
N, especially when perturbation methods are used. Non-linear approximations are
typically derived by retaining some of the o(ǫ) terms in (1.3), such as terms of O(ǫ2 )
and O(ǫ3 ), but the perturbation is still assumed to be small, i.e. ǫ is still assumed to
be small. The resulting problem for v, is called a weakly non-linear stability problem.
Examples of such problems will be considered later in the course.
Returning to the linear stability problem, we note that the problem is autonomous
if, as we shall assume, the original problem is (no explicit dependence on t in N and
in the BCs). In this case, we expect the existence of separable solutions to the linear
problem which are exponential in time, called normal modes, given by

v = Vest (Normal Modes) , (1.5)

where s and V are, using (1.4), any eigenvalue-eigenvector pair of the eigenvalue
problem
LV = sV (eigenvalue problem). (1.6)
Typically, (1.6) admits an infinite ‘complete’ set of solutions (s, V). This means
that any perturbation can be expressed as a linear combination of the normal-modes
(1.5). The linear combination is typically in the form of a Fourier series in a bounded
x-domain where we have a countable set of eigenvalues (a discrete spectrum), or a
Fourier integral in an unbounded domain (where we have a continuous spectrum); see
examples 2 and 3 below and also PDE books.
Now let
s = σ − iω {σ = sr ≡ Re(s) , −ω = si = Im(s)}
Equation (1.5) implies that |v| = |v|t=0 |e(σ−iω)t |. Hence

|v| = |v|t=0 eσt ,

which shows that the growth/decay of the amplitude of any normal mode is deter-
mined by σ = Re(s), which is called the growth rate of the normal mode; ω is called

4
 Im(s) Im(s)
(γ < 0) (γ > 0)

Eigenvalues Eigenvalues

Re(s) Re(s)

its angular frequency. Modes with σ < 0 will decay in amplitude and those with σ > 0
will grow. Therefore, we expect an arbitrary perturbation to decay to zero if all modes
decay, and to grow to infinity (in our linear approximation) if any mode grows. More
formally and conveniently, we introduce the following parameter γ, which we could
also denote by σmax , representing the maximum growth rate, defined by

γ = max{Re(s), all s}. (1.7)

Then, the basic solution ū is stable if γ < 0, unstable if γ > 0, and neutrally stable
if γ = 0.
As the parameter µ is varied, the basic solution may pass from being stable (γ < 0)
to being unstable (γ > 0) (see figure below) as the parameter µ crosses a critical value
µc for which γ = 0. The change in the stability of the solution represents a bifurcation
at µ = µc . This indicates that as µ crosses µc , the real part of an eigenvalue changes
sign, and hence is zero at µ = µc . At µ = µc , this eigenvalue with zero real part, say
s = sc , can be either real or complex. If it is real, then sc = 0, and the corresponding
normal mode (1.5), v = V, is time-independent or stationary; we talk about a ‘zero-
eigenvalue bifurcation’. If it is complex, its complex conjugate is typically (but not
always) also an eigenvalue; therefore we have a pair of complex conjugate eigenvalues
sc = ±iωc . The corresponding normal mode (1.5), v = e±iωc t V, is oscillatory. We
talk about an ‘oscillatory bifurcation’ or ‘Hopf bifurcation’. See examples below for
illustration.

5
Example 1 (a simple ODE)
Determine the stability of the stationary solutions of the following scalar ODE with
parameter µ:
ut = µ(u − u2 ) .

Solution
The stationary solutions u = ū are obtained by setting ūt = 0; hence ū = 0 and
ū = 1. The linear stability problem for each stationary solution is obtained by adding
to ū a perturbation of amplitude ǫ, so that u = ū + ǫv, substituting into the ODE,
and linearising (i.e. Taylor-expanding for ǫ → 0 and retaining terms of O(ǫ) only):
⇒ ǫvt = µ(ū + ǫv − (ū + ǫv)2 ) ⇒ ǫvt = µ (ū − ū2 ) +ǫµ(1 − 2ū)v − ǫ2 µv 2 . Neglecting
| {z }
0
the O(ǫ2 ) term, we have

vt = µ(1 − 2ū)v (linear stability problem) ,

with solution v = v0 eµ(1−2ū)t where v0 is the initial perturbation. For the solution
ū = 0, v = v0 eµt , and therefore this solution is unstable for µ > 0, neutrally stable
for µ = 0 and stable for µ < 0. For the solution ū = 1, v = v0 e−µt , and therefore this
solution is unstable for µ < 0, neutrally stable for µ = 0 and stable for µ > 0.

Example 2 (a PDE on an unbounded domain)

Determine the stability of the constant solutions of the following PDE with parameter
µ:
ut = uxx + µ(u − u2 ) (−∞ < x < ∞) .

Solution
The constant solutions u = ū are obtained by setting ūt = 0 and ūxx = 0; hence ū = 0
and ū = 1. The linear stability problem for each stationary solution is obtained by
adding to ū a perturbation of amplitude ǫ, so that u = ū + ǫv, substituting into the
PDE, and linearising, as in previous example:

vt = vxx + µ(1 − 2ū)v (linear stability problem) , (1.8)

to be solved with the initial condition v = f (x) at t = 0 and the requirement that v
remains finite as x → ±∞. Since this partial differential equation does not depend

6
on t and x explicitly, we expect the existence of solutions exponential in t and x of
the form
v = eikx est (Normal modes) , (1.9)
where s is complex and k real; the reason why k must be real is because otherwise |v|
will tend to infinity2 either as x → −∞ or x → ∞. Substituting into (1.8), we find
that s must be related to k by the following dispersion relation:

s = s(k) = µ(1 − 2ū) − k 2 (1.10)

Thus the normal modes can be indexed by the continuous variable k, the wave
number. Any mode such that s(k) > 0 is unstable. Note that s is real in this case,
hence σ = s and the maximum of σ(k) is attained for k = 0, hence γ = µ(1 − 2ū).
Since stability of the solution ū corresponds to γ > 0, we conclude that the stationary
solution ū = 0 is stable for µ < 0 and unstable for µ > 0 as in the space independent
case of example 1. We shall see in the following example that this conclusion becomes
invalid if the domain is bounded. As for the stationary solution ū = 1, it is stable for
µ > 0 and unstable for µ < 0, again as in example 1.

Remark
The normal modes are given by (1.9) with s satisfying the dispersion relation (1.10).
Since the problem is linear, any linear combination of normal modes is a solution. In
fact, as we shall show more generally in section 2.1, any perturbation v with initial
profile f (x) can be written here as a sum of normal modes in the form of the Fourier
integral Z ∞
v(x, t) = fˆ(k)es(k)t+ikx dk (1.11)
−∞

The coefficient fˆ(k) can be computed by putting t = 0 in this equation and using the
initial condition v(x, 0) = f (x), hence
Z ∞
f (x) = fˆ(k)eikx dk . (1.12)
−∞

This is shows that fˆ(k) is the Fourier-transform of f (x), and therefore

Z ∞
1
fˆ(k) = f (x)e−ikx dx . (1.13)
2π −∞
2
Indeed, if k = kr + iki , then |eikx | = |e−ki x eikr x | = e−ki x ; therefore |v| → ∞ as x → ∞ if ki < 0
or as x → −∞ if ki > 0.

7
Example 3 (a PDE on a bounded domain)
Examine the stability of the constant solutions of the problem

ut = uxx + µ(u − u2 ) (0 < x < 1) with u = 0 at x = 0, 1 (1.14)

Base state
Clearly, u = ū = 0 is the only constant solution satisfying the boundary conditions.

Small perturbation
Add a perturbation to the basic state, writing

u = ū + ǫv(x, t) (1.15)

Linearise the equations

Now substitute the perturbed form (1.15) into the governing equation (1.14):


(ū + ǫv)t = (ū + ǫv)xx + µ ū + ǫv − (ū + ǫv)2


= ūxx + µ ū − ū2 + ǫ [vxx + µ(1 − 2ū)v] + O(ǫ2 ) · · ·

• At O(ǫ0 ) we have ūt = ūxx + µ (ū − ū2 ). This is just equation (1.14) which is of
course obeyed by any solution including the basic state.
• At O(ǫ), which is the interesting part, we have

vt = vxx + µ(1 − 2ū)v

For any given base state ū(x), this linearised equation governs the dynamics of the
perturbations v(x, t). Below we consider the basic state ū = 0, for which

vt = vxx + µv (1.16)

• Terms O(ǫ2 ) and higher are neglected.

8
Solve the linearised equations using normal modes
We now express v(x, t) as a sum of normal modes, each of which has the general form

v(x, t) = V (x)est .

Substituting this into the linearised equation (1.16) we get

sV = Vxx + µV. (1.17)

Given the boundary conditions u = 0 at x = 0, 1, the function V (x) must clearly

obey boundary conditions V (0) = V (1) = 0. Therefore we have solutions of the form

V (x) = C sin(nπx).

Substituting this into Eqn. (1.17), we get

s = µ − n2 π 2 .

So in this case, we have a discrete spectrum with the eigenvalues s(n) indexed by the
integer n ≥ 1. Note that the eigenvalues are real, and that the nth mode is
• decaying s < 0 for µ < n2 π 2 ,
• neutral s = 0 for µ = n2 π 2 ,
• growing s > 0 for µ > n2 π 2 .

So the basic state is linearly stable (all eigenvalues negative) for µ < π 2 . As µ is
increased, the basic state first becomes unstable at µ = π 2 . In the unstable regime
µ > π 2 , the most unstable mode is the one with the most positive eigenvalue: i.e.
mode n = 1, which has s = µ − π 2 . The fastest growing mode is then given by

v(x, t) = C sin(πx) e(µ−π ) t .

2

Note that in this example, an arbitrary perturbation v(x, t) with an initial profile
f (x) can be written as a sum of normal modes in the form of the sine-Fourier series
∞
X Z 1
Cn sin(nπx) e(µ−n π ) t
2 2
v(x, t) = with Cn = 2 f (x) sin(nπx)dx .
n=1 0

9
Example 4 (a simple oscillatory bifurcation)
In the previous examples the eigenvalues of the linear stability problem were found to
be real. Transition from stable to unstable solutions as µ is varied corresponds to zero-
eigenvalue bifurcations. This example provides a simple illustration of the occurrence
of an oscillatory bifurcation, using the familiar damped oscillator equation

vtt + 2µvt + v = 0 .
p
st
Here the normal modes are given
p by v = Ce , hence s = −µ ± µ2 − 1. If |µ| < 1,
then s = −µ ± iω with ω = 1 − µ2 , and the modes are stable for µ > 0 and
unstable for µ < 0. At µ = µc = 0, s = sc = ±iωc with ωc = 1; this is an oscillatory
bifurcation.

10
2 Linear PDEs with constant coefficients and sta-
bility
Often, the linearisation around a basic solution leads to a linear PDE problem with
constant coefficients, as illustrated in example 2 of previous section. We are then
led to examine the solution of the linear PDE. For simplicity, we shall assume the
domain to be infinite, and illustrate the procedure on a scalar PDE in one dimension.
Generalisation to systems of equations and several dimensions can then be easily
made, as shown in the last two examples of this section.

2.1 Solution using normal modes

Consider the PDE
vt = a0 v + a1 vx + a2 vxx + a3 vxxx + a4 vxxxx + a5 vxxxxx , (−∞ < x < ∞) (2.1)
where a0 , a1 . . . are real constants and require that v is bounded as x → ±∞.
Since the problem is translationally invariant in x and t, we look for exponential
solutions in the form
v = A(k)es(k)t+ikx (2.2)
where k is real (since we require v to be bounded as x → ±∞) , and s(k) and A(k)
are in general complex. Substitution of (2.2) into (2.1) leads to the dispersion relation



s(k) = a0 − a2 k 2 + a4 k 4 + i a1 k − a3 k 3 + a5 k 5 (2.3)
Therefore, writing s(k) = σ(k) − iω(k), the growth rate of a normal mode σ(k) and
its (angular) frequency ω(k) are given by
σ(k) = a0 − a2 k 2 + a4 k 4 and ω = −a1 k + a3 k 3 − a5 k 5 . (2.4)
It is seen that the growth rate σ(k) depends only on the coefficients of the x-derivatives
of even order in (2.1); ω(k) depends only on the coefficients of the x-derivatives of odd
order, and therefore s(k) is real if no derivative of odd order is present in PDE (2.1).
Since the problem is linear, we expect any perturbation to be a sum of normal
modes, i.e. Z ∞
v(x, t) = A(k)es(k)t+ikx dk . (2.5)
−∞
Before proceeding, it is instructive to derive this result by solving the PDE problem
by the Fourier transform method. This is done in the next subsection.

1
2.2 Solution using Fourier transform (optional)
We want to solve the PDE (2.1) with the initial condition v(x, 0) = f (x), using the
Fourier Transform (FT). We begin by reminding that the FT fˆ(k) of a function f (x)
and its inverse are given by (1.12) and (1.13) (used in example 2 of previous section),
which we rewrite here for convenience:
Z ∞ Z ∞
ˆ ikx ˆ 1
f (x) = f (k)e dk , f (k) = f (x)e−ikx dx . (2.6)
−∞ 2π −∞
This definition extends of course to functions having additional independent variables.
For example, the FT of v(x, t) with respect to x is a function v̂(k, t) such that
Z ∞ Z ∞
ikx 1
v(x, t) = v̂(k, t)e dk , v̂(k, t) = v(x, t)e−ikx dx . (2.7)
−∞ 2π −∞

We also remind the following basic properties of the FT which we need here. First
we note, by formally differentiating the first equation in (2.6) with respect to x, that
if fˆ is the FT of f , denoted by f ↔ fˆ, then the FT of fx is ik fˆ, the FT of fxx is
−k 2 fˆ, and so on; this is expressed concisely by the following shorthand notation

f ↔ fˆ , fx ↔ ik fˆ , fxx ↔ −k 2 fˆ , fxn ↔ (ik)n fˆ .

Since v(x, t) ↔ v̂(k, t), we therefore have
vt ↔ v̂t , v ↔ v̂ , vx ↔ ikv̂ , vxx ↔ −k 2 v̂ , vxn ↔ (ik)n v̂ . (2.8)
We now take the FT of (2.1) with respect to x using (2.8), to obtain
v̂t = (a0 + ika1 − k 2 a2 − ik 3 a3 + k 4 a4 + ik 5 a5 )v̂ ,
that is
v̂t = s(k)v̂ , (2.9)
on account of (2.3). This is an ODE for v̂ in which k appears as a parameter, and
which is to be solved with the I.C. v̂(k, 0) = fˆ(k). We have therefore a system of
uncoupled ODEs of infinite dimension, i.e. an ODE for each value of k which can be
solved independently; the solution of each ODE is of course
v̂(k, t) = fˆ(k)es(k)t . (2.10)
Using (2.10) in (2.7) yields the solution of our initial value PDE problem:
Z ∞
v(x, t) = fˆ(k)es(k)t+ikx dk . (2.11)
−∞

2
This justifies the method of normal modes which assumes, from the start, that v(x, t)
is of the form (2.5), a superposition of independent modes of the form A(k)es(k)t+ikx .
In summary: in the method of normal modes, an arbitrary disturbance introduced
as an IC is expanded as a Fourier sum (integral or series) over k. The problem being
linear, each Fourier component is examined on its own, and evolves independently of
other components.

2.3 Stability
As in section 1, let γ = max{σ(k), ∀k}, where σ(k) is given by (2.4). Then, on using
(2.5) or (2.11), we have:

1. If γ < 0 all solutions decay in magnitude to zero as t → ∞ (asymptotic sta-

bility). The zero solution is (asymptotically) stable. We can even say that the
equation is stable (in this linear case corresponding to (2.1)).

2. If γ = 0, the zero solution (or equation) is neutrally stable.

3. If γ > 0, the zero solution (or equation) is unstable.

In the unstable case γ > 0, if γ = ∞ then the IVP is not well-posed as the solu-
tion does not depend continuously on the initial condition (data)3 . Effectively, this
implies that the solution will blow up in finite time, as illustrated in the following
two examples. When this is the case, the model leading to the equation needs to be
revised.

3
Reminder: an IVP or IBVP is well-posed if (a) a solution exists, (b) is unique and (c) depends
continuously on the initial (and boundary) data.

3
Example 1
Show that if a4 > 0 in (2.1), then the IVP is not well-posed.

Solution
Note that (2.4) implies that σ(k) ∼ a4 k 4 as |k| → ∞. It follows, that γ = ∞, and
the problem is not well-posed. To show this, first note that v ≡ 0 is solution to the
IVP with zero initial condition f (x) = 0. We now choose an IC which can be made
arbitrarily close to zero, and show that the corresponding solution is arbitrarily large.
That’s simple: just choose the normal-mode

v = k −1 es(k)t+ikx ,

which clearly solves the IVP with IC f (x) = eikx /k. Note that |IC| = 1/|k| → 0 as
4
|k| → ∞, but |v| = |k|−1 es(k)t ∼ |k|−1 ea4 k t → ∞ in this limit. Therefore, the solution
to the problem with initial data f (x) → 0 does not tend to (in fact diverges from)
the solution to the problem with f (x) = 0, that is from the zero-solution.

Example 2
The backward heat equation vt = −vxx corresponds to choosing a2 = −1 with all
other coefficient zero. Show the corresponding IVP is ill-posed.

Solution
This is a notorious example of an ill-posed IVP. The proof is similar to that of the
previous example with
2
v = k −1 ek t+ikx
which satisfies the IVP with initial condition f (x) = eikx /k. Note in this example,
and in the previous one, that the non well-posedness is associated with the fact that
|v| → ∞ as |k| → ∞ with t fixed. This is very distinct from the concept of instability
which is concerned with the limit t → ∞ with k fixed.

Example 3 (Kuramoto-Sivashinsky equation)

The Kuramoto-Sivashinsky equation is given by
1
vt + vx2 = −vxx − vxxxx .
2
4
This equation may be viewed as a new model to replace (or fix) the backward
heat equation of the previous example, even without the presence of the non-linear
quadratic term. Here the fourth-order derivative is stabilising, leading to γ = 1/4, i.e.
to a well-posed problem for the linearised PDE, as discussed in more detail in example
sheet 1.

2.4 Further definitions and remarks

In the remainder of this section, important basic definitions and remarks are given.
The linear PDE is said to be conservative if s(k) is purely imaginary, i.e. s(k) =
−iω(k) for all values of k. In this case, each normal mode conserves its amplitude
as time varies, i.e. each normal modes satisfies |v(x, t)| = |v(x, 0)|. The quantity
c(k) ≡ ω(k)/k is called the phase speed. If the phase speed c(k) is not constant, the
equation is called dispersive, otherwise it is non-dispersive. In the dispersive case,
different modes (corresponding to different values of k) travel at different speeds, and
therefore an arbitrary initial profile, decomposed of several modes, will loose its shape
as it evolves in time.
The PDE is said to be dissipative if all modes have negative σ(k) (and hence decay
in amplitude to zero as t → ∞) except for a finite number of modes which are neutral
(with σ = 0).

Example 4
The heat equation vt = Dvxx (with diffusion coefficient D > 0) is dissipative since
s(k) = −Dk 2 is negative for all k, except for k = 0 (neutral).

Example 5 (The telegrapher’s equation)

The telegrapher’s equation is given by

vtt − c2 vxx + 2µvt = 0 (µ > 0) .

p
Since s(k) = −µ ± µ2 − c2 k 2 , all modes are decaying (Re(s(k) < 0), except for
k = 0 (neutral). It follows that the equation is dissipative.

5
Example 6 (Extension to a system of equations)
Let v(x, t) be a vector with n components. Consider the PDE system

vt + Cvx = Dvxx + Ev , (2.12)

where C, D and E are constant n × n matrices. Determine the dispersion relation
s(k) to be satisfied by the normal-mode solutions
v = A(k)es(k)t+ikx . (2.13)

Solution
v given by (2.13) is a solution of (2.12) ⇔ sA + ikCA = −k 2 DA + EA ⇔
 
sI + ikC + k 2 D − E A = 0 ,
where I is the n×n identity matrix. This is a homogeneous linear system of equations
with A = 0 as an obvious trivial solution. For a non-trivial solution, we require
 
det sI + ikC + k 2 D − E = 0 (dispersion relation).

Example 7 (Extension to several space dimensions)

Consider the linear constant coefficient convection-diffusion-reaction equation
vt + c · ∇v = α∇2 v + βv (2.14)
Determine the dispersion relation s(k) to be satisfied by the normal-mode solutions
v = A(k)es(k)t+ik·x , (2.15)
where c, x and k are 2D or 3D vectors.

Solution
v given by (2.15) is a solution of (2.14) ⇔ sA + ik · cA = −αk 2 A + βA ⇔


s + ik · c + αk 2 − β A = 0 ,
where k = |k|. This is a homogeneous linear equation with A = 0 as an obvious
trivial solution. For a non-trivial solution, we require
s = β − αk 2 − ik · c (dispersion relation).

6
3 ODEs: stability and bifurcations
3.1 General approach: fundamentals
This is a brief outline/reminder of a large topic (dynamical systems). Consider the
autonomous ODE system with parameter µ:
du
= F(u, µ) , (3.1)
dt
where u(t) is a vector with components (ui )i=1,··· n . Let ū be a stationary solution
whose stability we wish to examine. As usual, adding a perturbation of small ampli-
tude ǫ, i.e. letting u = ū + ǫv, linearising and using F (ū, µ) = 0, we obtain

dv
= Lv , (3.2)
dt
where L is the jacobian matrix of F evaluated at ū, namely L = (∂Fi /∂uj )u=ū .
Since the system is autonomous, we look for solutions exponential in time given
by
v = Vest (Normal Modes) , (3.3)
where s and V are, using (3.2), any eigenvalue-eigenvector pair of the eigenvalue
problem
LV = sV (eigenvalue problem). (3.4)
To decide about the stability of ū, we therefore need to examine the eigenvalues s,
which are the roots of the characteristic polynomial equation

det [L − sI] = 0 ⇔ sn + a1 sn−1 + · · · + an−1 s + an = 0 . (3.5)

Theorem: The stationary solution u = ū is asymptotically stable if all eigenvalues

s have negative real parts, and is unstable if any eigenvalue has a positive real part.

Quick explanation: The theorem is obvious when the characteristic equation

(3.5) has n distinct roots (si )i=1,2,···n (eigenvalues of (3.4) with associated eigenvectors
Vi ), since the general solution of (3.2) is then
n
X
v= ci Vi esi t ,
i=1

1
where ci are arbitrary constants. If some eigenvalues are not distinct, the general
solution contains additional terms of the form tk esi t , where k is some positive integer;
in this case however the conclusions of theorem are still clearly applicable.

Conclusion: The stability problem reduces to examining whether the real parts
of the roots of the characteristic polynomial equation (3.5) are negative or not. To
this end, the following stability criterion is very useful.

Routh-Hurwitz stability criterion: A necessary and sufficient condition for

the roots si of the nth order polynomial Pn ≡ sn + a1 sn−1 + · · · + an−1 s + an to all
have negative real parts (Re(si ) < 0) is that Di > 0 for each of the following n
determinants D1 , D2 , · · · Dn :
 
  a1 a3 a5 
a a   
D1 = a1 , D2 =  1 3  , D3 =  1 a2 a4  , · · ·
1 a2  0 a1 a3 

Rule for determining the entries for each Di is simple: first fill the diagonal with
the i coefficients a1 , a2 , · · · ai ; to determine other entries raise the index of the sub-
script by 1 when going upwards (↑), and lower it by one when going downwards (↓)
and set ak = 0 if k ∈ / {0, 1, · · · n}.

Example 1 (Stability criterion for n = 1, 2, 3, 4)

• n = 1: P1 = s + a1 ⇒ D1 = a1 so the stability criterion is a1 > 0 (obvious).
 
a1 0 
• n = 2: P2 = s + a1 s + a2 ⇒ D1 = a1 , D2 = 
2  = a1 a2 . So the stability
1 a2 
criterion requires a1 > 0 and a2 > 0. Note that if s1 and s2 are the roots of P2 ,
i.e. P2 = (s − s1 )(s − s2 ), then T ≡ s1 + s2 = −a1 and D ≡ s1 s2 = a2 , so the
criterion takes the equivalent form T < 0 and D > 0.

• n = 3: P3 = s3 + a1 s2 + a2 s + a3 ⇒
 
  a1 a3 0 
a1 a3   
D1 = a1 , D2 =   = a1 a2 − a3 , D3 =  1 a2 0  = a3 D2 .
1 a2   0 a1 a3 
So stability criterion requires a1 > 0, a3 > 0 and a1 a2 > a3 .

2
 • n = 4: P4 = s4 + a1 s3 + a2 s2 + a3 s + a4 . The stability criterion requires
a1 > 0, a3 > 0, a4 > 0 and a1 a2 a3 > a21 a4 + a23 .

3.1.1 1D systems (n = 1)
Here ut = F (u) and the stability of the stationary solution (or fixed point) u = ū is
determined by L = Fu (ū): ū is stable if Fu (ū) < 0 and unstable if Fu (ū) > 0.
As a simple consequence of this simple stability criterion in 1D, a steady state can
become unstable, as µ varies, only if the real eigenvalue Fu (ū) crosses zero, a ‘zero-
eigenvalue bifurcation’; in particular no oscillations are possible. In 1D, the graphical
approach is often sufficient and is powerful (giving information about global stability),
as illustrated in the following example.

Example 2 (Ignition-Extinction-Hysterisis)
The temperature u in a “well-stirred reactor” is modelled by
du
= F (u; µ) ≡ −u + µ ω(u) ,
dt
where ω is the reaction rate which is a positive function of u, and µ is a parameter
(called the Damköhler’s number). Determine the steady states and their stability in
the case ω = (1 − u) exp(β(u − 1)), with β = 10 (or any constant > 4).

Solution
The steady states are given by ut = 0 ⇔ 0 = −u + µω(u) ⇔ µ = u/ω(u).
Plotting µ versus u and placing the µ-axis horizontally gives figure below. The S-
shaped curve generated represents the steady states and is a fundamental curve in
combustion theory where the upper turning point E defines extinction and the lower
turning point defines ignition. For given values of µ between µE and µI , multiple
solutions exist, which leads to hysterisis. The intermediate solution (dashed curve)
is unstable, upper branch (hot flame) and lower branch (frozen solution) are stable.
Stationary solutions and their (global) stability are all determined by the simple
graphical approach; the S-shaped curve (corresponding to ut = 0) separates regions
in the (µ, u)-plane where ut > 0, i.e. where F (u; µ) > 0 identified with upwards arrows
(↑) from regions with downwards arrows (↓) where ut < 0. The arrows indicate the
direction a perturbation follows for fixed µ.

3
 Remark: It is interesting to repeat this example by changing the reaction term
ω = (1 − u) exp(β(u − 1)) to ω = (1 − u)uβ ; check in this case that the ignition point
I disappears (or is rejected to µ = ∞) and that u = 0 is a linearly stable solution
for all values of µ. However, for any perturbation of finite amplitude, however small,
the u = 0 solution becomes unstable for µ larger than a critical value µc which is a
function of the (initial) amplitude, say A, with limA→0 µc = ∞. In a more complicated
system such behaviour could be puzzling, especially for experimentalists, as it seems
difficult to conclude whether the base solution is stable or not; e.g. Poiseuille flow in
a pipe where the relevant parameter µ is the Reynolds number.

3.1.2 2D systems (n = 2)
 
∂Fi
Here L = ∂u j
is a 2D constant matrix with eigenvalues s1 and s2 , trace T =
u=ū
s1 + s2 , determinant D = s1 s2 , and characteristic equation s2 − T s + D = 0. The
eigenvalues can be located in the (T, D)-plane using
√
T ± T 2 − 4D
s1,2 = .
2
If their locations is in the upper-left quarter plane T < 0 and D > 0, then, as found
in the previous section, ū is stable. As seen in figure, there are two ways to cross
from the stability region to the instability region as µ is varied. (1) ↓: At the critical
condition D = 0, T < 0, implying that one eigenvalue is zero while the other < 0;
this is a zero eigenvalue
√ bifurcation. (2) →: At the critical condition T = 0, D > 0
and hence s1,2 = ±i D; √ this is a Hopf-bifurcation which involves oscillations with
angular frequency ω = D.

4
 D

(Hopf
(2) bifurcation)
Stability Region
(gray area)
2
T - 4D = 0

(Zero-eigenvalue
(1) bifurcation )

Example 3 (Turing instability)

Consider the 2D reaction-diffusion PDE system
 
D1 0
ut = ∇2 u + f(u) ,
0 D2

with u = (u1, u2 ), f = (f1 , f2 ) and diffusion coefficients D1 and D2 . Derive the

condition insuring the stability of a constant solution u = ū.

Solution
Plug u = ū + ǫv into the PDE system and linearise to obtain for the perturbation v
   
D1 0 2 ∂fi
vt = ∇ v + Jv with J≡ .
0 D2 ∂uj ū

Since there is no explicit dependence on space, look for solutions in the form

v = w(t)eik·x ,

hence the following ODE system for w(t)

!
∂f1 ∂f1
dw −D1 k 2 + ∂u1 ∂u2
= Lw with L= ∂f2 2 ∂f2 ,
dt ∂u1
−D2 k + ∂u2

5
where k = |k| and the partial derivatives are evaluated at u = ū. For stability we
require trace(L) < 0 and det(L) > 0, hence
  
∂f1 ∂f2 2 ∂f1 2 ∂f2 2 ∂f1 ∂f2
+ − (D1 + D2 )k < 0 and − D1 k − D2 k − > 0.
∂u1 ∂u2 ∂u1 ∂u2 ∂u2 ∂u1

3.1.3 3D systems (n = 3)
We shall consider Lorentz equations in example sheet 2, question 6. It is a classical
example, but it is very instructive. It will also provides you with a good opportunity
to apply the Routh-Hurwitz stability criterion for n = 3.

3.2 Bifurcation theory

Consider an n-dimensional ODE system ut = F(u, µ) with parameter µ. When a
stationary solution ū(µ) (satisfying F(ū, µ) = 0, of course, and depending on µ) loses
stability as µ crosses a (critical) value µc we say that a bifurcation occurred at µc .
There are two ways the bifurcation can happen:

(1) At µ = µc , and ū = uc say, an eigenvalue of the linearised problem becomes

zero, say sc = 0 (all other eigenvalues s having Re(s) < 0); this is a zero-eigenvalue
bifurcation. Here the Jacobian matrix L = (∂Fi /∂uj )u=ū is non-invertible at µ = µc
because then det(L) = 0 (which is obvious since sc = 0 is an eigenvalue of L ⇔
det(L − 0I) = 0). An important consequence is that the implicit function theorem4
does notes apply at (u, µ) = (uc , µc ). Therefore, the existence close to (uc , µc ) of
a unique (stationary) solution ū(µ) satisfying F(ū(µ), µ) = 0 is not guaranteed;
in practice this means that no solution or multiple solutions are expected close to
(uc , µc ). The bifurcation for such case is well modelled by standard 1D bifurcations
(saddle-node, transcritical, pitchfork), presented below.

(2) A pair of complex conjugate eigenvalues are purely imaginary at µ = µc :

sc = ±iωc (all others having Re(s) < 0). This is a Hopf bifurcation and requires
n ≥ 2. Here the implicit function theorem applies at (uc , µc )), insuring the existence
of a unique stationary solution around this point.
4
The implicit function theorem may be stated for our case as follows: if F(u, µ) is C 1 and
F(uc , µc ) = 0 then the equation F(u, µ) = 0 has a unique solution (u(µ), µ) in a neighbourhood of
(uc , µc ) if det(L) 6= 0 at (uc , µc ).

6
 In the remainder of this section, we introduce the general theory of bifurcations
in the context of some simple “standard” equations. It worth noting though that the
most complicated equations may be shown to reduce to one of these standard forms
when expanded about a bifurcation point.

3.2.1 The saddlenode (or ‘turning point’ or ‘fold’) bifurcation

Consider the dynamical system defined by
dx
= a − x2 , for x, a real.
dt
Here and below a is a control parameter that can be tuned externally; this has been
denoted by µ above. A steady state solution (dx/dt = 0) is simply
√
x = x̄ = ± a.
Therefore, for a < 0 we have no real solutions, and for a > 0 we have two real solutions.
This leads to the “bifurcation diagram” below, where the dashed line represents
unstable solutions (and the solid line represents stable solutions). As explained more
generally in the previous section, the stability can be determined easily graphically
(arrows in the figure) or by computing the partial derivative
√ Fx (x̄, a); here F (x, a) =
2
a−x √ so Fx (x̄, a) = −2x̄; so for a > 0 the solution x̄ = a is stable and the solution
x̄ = − a unstable.

x
+a1/2

−a1/2

3.2.2 The transcritical bifurcation

Consider the dynamical system
dx
= ax − x2 for x, a real.
dt
7
There are two steady states (dx/dt = 0) given by x = x̄1 = 0 and x = x̄2 = a, which
exist for all values of a. The bifurcation diagram is given below. The bifurcation at
a = 0 corresponds to an exchange of stability between the two solution branches.

x=0
a

x=a

3.2.3 The pitchfork bifurcation

Consider the dynamical system defined by
dx dx
= ax − x3 (supercritical) or = ax + x3 (subcritical) .
dt dt
The steady states, for which dx/dt = 0, are given by
√ √
x = x̄1 = 0 , x = x̄2 = + a , x = x̄3 = − a ,

in the supercritical case, with x̄2 and x̄3 only existing for a > 0. In the subcritical
case, the steady states are given by
√ √
x = x̄1 = 0 , x = x̄2 = + −a , x = x̄3 = − −a ,

with x̄2 and x̄3 only existing for a < 0. The bifurcation diagrams are shown next,
and have the form of a pitchfork. The bifurcation which occurs at a = 0 is termed
“supercritical pitchfork” in the supercritical case (for which the non-linear term −x3
is stabilising) and “subrcritical pitchfork” in the subcritical case (for which the non-
linear term +x3 is destabilising). In the latter situation, the model is often modified
(by adding a stabilising non-linear term) typically to dx/dt = ax + x3 − x5 ; as an
exercise, redo the bifurcation diagram in this case.

8
 supercritical pitchfork bifurcation subcritical pitchfork bifurcation

x x

a a

3.2.4 The Hopf bifurcation

A prototypical 2D model for Hopf bifurcation is given by
dx
= −y + (a − x2 − y 2 )x
dt
dy
= x + (a − x2 − y 2 )y
dt
for real x, y, a. There is a trivial steady state at x = y = 0. To examine its linear
stability, write x = 0 + x̃, y = 0 + ỹ, and linearise, to get
     
d x̃ x̃ a −1
=L with L= .
dt ỹ ỹ 1 a
The eigenvalues s of L are given by s = a ± i; we therefore conclude that
• if a > 0 then Re(s) > 0 and so |x̃|, |ỹ| → ∞ (linear instability);

• if a < 0 then Re(s) < 0 and so |x̃|, |ỹ| → 0 (linear stability).

The fact that s is complex confers a new dynamical feature not encountered in the
previous examples: that of temporal oscillation. For a < 0, for example, the progress

9
 y y

x x

a<0 a>0

of x̃ and ỹ in towards the origin is via a damped oscillation, as sketched in the left
hand plot, rather than a straightforward exponential decay.
As in the other bifurcation examples, the loss of stability at a = 0 gives rise to a new
solution for a > 0. In this case, the new solution is periodic:
√ √
x = a cos(t + t0 ), y = a sin(t + t0 ).

The system orbits round the “limit cycle”, the circle drawn in the right hand sketch
above. The bifurcation diagram is then as follows.

Comparing this to the diagrams of section 3.2.3, you will notice that it looks a bit like
a higher dimensional version of a supercritical pitchfork bifurcation. Indeed, we can
again classify Hopf bifurcations as supercritical or subcritical, according to whether
the nonlinearity is destabilising or stabilising respectively.

10
3.2.5 Bifurcations in the Lorentz equations
The Lorentz equations are given by
dx
= σ(y − x),
dt
dy
= rx − y − xz,
dt
dz
= −bz + xy.
dt
The dependent variables are x, y, z, and σ, r, b are control parameters, which we take
to be real and positive. We will assume σ, b to be fixed, and vary r as the single control
parameter. This leads to interesting bifurcations which are instructive to investigate;
see example sheet 2, question 6.
A qualitative summary is provided here referring to the bifurcation diagram below:
the zero solution is found to be stable for r < 1 and unstable for r > 1. At r = 1
a supercritical bifurcation occurs to a non-zero stationary solution. The bifurcated
solution is found to be stable for r < rcrit where it undergoes a Hopf bifurcation which
can be shown to be subcritical. As usual, dashed lines represent unstable solutions.
x or y

r=1 r=rcrit

Dynamical evolution beyond subcritical bifurcations

A subcritical bifurcation occurs at r = rcrit in the bifurcation diagram sketched above.
What happens for r > rcrit , where stability is lost and there is no “nearby” nonlinear
state to go to? In general, several scenarios are possible:
• Evolution to infinity, typically indicating a breakdown of the model.

• Evolution to a non-local fixed point.

11
 • Evolution to a non-local periodic or quasi-periodic state.

• Evolution to a strange attractor, leading to chaotic dynamics.

In the Lorentz equations, the last of these scenario occurs for r > rcrit .

12
4 Governing equations and conditions at bound-
aries and interfaces
In this section, we provide a summary of the equations and boundary conditions which
will be repeatedly used in this course. These should be familiar, at least partially, to
most students from previous courses. If needed, you could just accept the equations
and the conditions, familiarise yourself with the notation and variables introduced,
and make sure that you are able to use the interfacial conditions. This compact
summary is supported by visualiser notes and solved examples which provide more
details and explanations.

4.1 Governing equations

We shall mostly use the following three equations
∂ρ
+ ∇ · ρv = 0
∂t
Dv
ρ = ∇ · σ + ρg
Dt
µ
= −∇p + µ∇2 v + ∇(∇ · v) + ρg
3
DT
ρcp = λ∇2 T + ST
Dt
These equations are known as the continuity, the Navier-Stokes, and the temper-
ature equations, respectively. They express the conservation of mass, momentum,
and energy. Here, ρ and v are the fluid density and velocity, p the pressure, µ the
dynamic viscosity (assumed constant), σ the stress tensor, g the gravity field, T the
temperature, cp and λ the heat capacity and conductivity (assumed constant). In
the temperature equation, the last term named ST is a “source” term which may be
set to zero in many applications. More generally, it can be adjusted to account for
additional thermal effects if needed. For example,
Dp
ST = + Φ + ωT
Dt
accounts for compressibility effects (first term on rhs), viscous dissipation effect (sec-
ond term, see next paragraph) and heat-generation (say by combustion, with ωT
representing the rate of heat-release by reaction). In addition, an equation of state is
also assumed to apply of the form ρ = ρ(p, T ), or ρ = ρ(T ) (for quasi-isobaric flow),

1
or ρ = Cst (for incompressible flow). Six scalar equations are therefore available for
the six scalar unknowns characterising ρ, v, p, and T .

The stress tensor σ:

By definition σ · n is the stress (i.e. force per unit area) exerted by the fluid towards
which n points on a surface element with unit normal vector n.
A common model for the components of σ in a Cartesian coordinate system is:
   
1 1 ∂vi ∂vj
σij = −pδij + 2µ eij − (∇ · v)δij where eij ≡ +
3 2 ∂xj ∂xi
 
It can be shown that the viscous dissipation may be written as Φ = 2µ eij eij − 13 (∇ · v)2 .

The vorticity and the vorticity equation:

A fundamental variable in fluid mechanics, which we shall use later in the course, is
the flow vorticity ω. This is by definition the curl of the velocity field, ω = ∇×v. On
taking the curl of the Navier-Stokes equation above after dividing it by ρ, we obtain
upon using a few vector identities (see exercise below):

Dω
= ν∇2 ω + Sω
Dt
where ν = µ/ρ is the kinematic viscosity and Sω a vorticity production term given
by
 
1 µ 2 1
Sω = ω · ∇v − (∇ · v) ω + 2 ∇ρ × ∇p − 2 ∇ρ × ∇ v + ∇(∇ · v)
ρ ρ 3

The third term on the rhs side of last equation represents the baroclinic generation
of vorticity which is due to the misalignment of the density and pressure gradients.
The last three terms are in fact zero if the flow is assumed to be incompressible since
then ρ = Cst and ∇·v = 0, leaving Sω = ω·∇v. This term, namely ω·∇v, represents
vorticity generation by a mechanism referred to as vortex stretching; this mechanism
is believed to be essential in turbulent flow dynamics. This term is itself zero for
two-dimensional (2D) flows since then ω · ∇ = 0. Therefore for a 2D incompressible
flow, Sω = 0, and the vorticity equation is significantly simplified.

2
Exercise (vorticity equation)
1. Derive the vorticity equation, i.e. show that Sω is as given above.
Hint: Take the curl of the Navier-Stokes equation and use vector identities A9-
A2-A3-A6-A10 of the appendix provided.

2. Explain why the term ω ·∇v represents vortex stretching, by considering Carte-
sian coordinates with the z-axis chosen to coincide at a given instant with the
direction of ω.

3. Justify that Sω = 0 for a 2D incompressible flow, say given by v = u(x, y, t)i +

v(x, y, t)j.

Solution
See solutions to example sheet 3.

4.2 Boundary conditions at an interface

Let v = (u, v, w) be the fluid velocity and V the velocity of the interface. Assume
that the interface, described by an equation such as F (x, t) = 0 or z = h(x, y, t),
is characterised by a unit normal vector n, and let t be an arbitrary unit vector
perpendicular to n.

V
F(x,t) = 0
or
n
z-h(x,y,t)
=0

4.2.1 No-penetration (or kinematic) condition

This condition applies if the interface is a material surface, i.e. formed by the same
particles or molecules. With the notation of figure above, it may be stated as
DF ∂F Dh
v.n = V.n or ≡ + v · ∇F = 0 or =w
Dt ∂t Dt
3
If the interface separates two phases denoted by subscripts 1 (say water below the
interface, with fluid velocity v1 , density ρ1 , etc.) and 2 (say air above the interface,
with fluid velocity v2 , density ρ2 , etc.), then the kinematic condition is to be applied
on each side of the interface resulting in
−Ft
v1 · n = v2 · n = V · n =
|∇F |

It is important to emphasize that the kinematic condition is not applicable if the

interface is not a material surface. Examples of such penetrable surfaces includes
flames, vaporisation fronts, or mathematical surfaces such as concentration or tem-
perature iso-surfaces. In these cases, the kinematic condition needs to be replaced by
the requirement of continuity of the mass flux across the interface, namely

Jρ (v − V) · nK21 = 0 .

Exercise (Kinematic condition)

Justify the kinematic condition as given above (Solution: See visualiser notes).

4.2.2 No-slip condition

This is an empirically robust assumption and can be stated, with t denoting an
arbitrary tangent vector to interface, as

(v − V) · t = 0 or (v − V) × n = 0

4.2.3 Dynamic condition

Fluid 2

n Fluid 1

This is a balance of forces or action/reaction principle applied to fluids at the

interface:

σ 1 · n − σ 2 · n − γn(∇ · n) + ∇s γ = 0 where ∇s ≡ ∇ − n(n · ∇)

4
Here γ is the surface tension (which is typically a function of temperature and con-
centration) and ∇ · n is twice the mean curvature κm of the interface, ∇ · n = κm .
Note also that the equation for the dynamic condition given above is valid when n is
chosen to point towards5 fluid 1.
By taking the dot product of equation with n, the balance of normal stresses is
obtained. By taking the cross product with n the balance of tangential stresses is
obtained. Hence

J(σ · n) · nK21 = −γ(∇ · n) (normal balance)

and
J(σ · n) × nK21 = ∇s γ × n (tangential balance)

Comment on the mean curvature:

The mean curvature κm of a surface at a given point P with normal n as shown in
the figure, is by definition
κ1 + κ2
κm =
2
where κ1 and κ2 are the curvatures of the curves C1 and C2 , respectively; here C1 and
C2 are normal sections to the surface at P in two orthogonal directions. By definition,
a normal section in the direction of a vector t tangent to the surface at P is the curve
corresponding to the intersection of the normal plane containing P, n and t with the
surface. The curvature κ of a cross section (such as curves C1 and C2 ) is equal to the
reciprocal of its radius of curvature R taken with an appropriate sign; the sign will
be taken to be positive if n points away from the centre of curvature6 and negative
otherwise. It can be shown, that for the calculation of the mean curvature any two
orthogonal normal sections will do. It is common and convenient however to take
κ1 and κ2 as the principal curvatures, that is to correspond to the minimum and
maximum values assumed by κ(t) as t varies.
For example, for a sphere of radius R with n pointing radially outward ∇ · n =
2/R; for a cylinder with n pointing radially outward ∇ · n = 1/R; for a flat surface
∇·n = 0, obviously. These formulas are intuitive geometrically, but can be determined
analytically by using the expression of the divergence in spherical, cylindrical, and
rectangular coordinates (simple exercise).
5
Otherwise, use −σ 1 · n + σ 2 · n − γn(∇ · n) + ∇s γ = 0 when n points towards fluid 2.
6
The radius and centre of curvature of a curve C at point P are those of the circle which best
approximates C at P (the so-called osculating circle).

5
 n C2
C1
R1

R2

Example: 2D parallel viscous incompressible flow (to practise)

Consider a 2D viscous incompressible flow v = (u(y), 0) = u(y)i. Determine σ and
write down the equations corresponding to the normal balance and tangential balance
of stresses. Determine the stress exerted by the fluid on a horizontal wall given by
y = ywall .

Example: Interface between fluids at rest and the Young-

Laplace formula (1805)
Consider an interface between two fluids at rest (v = 0). Taking into account surface
tension effects, write down the dynamic condition at the interface. Hence show that
for the particular case of a bubble of air in water, we have
2γ
pin − pout =
R
where R is the bubble radius, and pin and pout are the pressure inside and outside the
bubble, respectively.

6
Worked example: inviscid incompressible flow with an inter-
face (needed later)
Consider an inviscid incompressible flow v = (u, v, w) involving an interface between
fluid 1 (bottom) and fluid 2 (top) given by z = h(x, y, t). Choose the unit normal
n to point towards fluid 2 (upwards) and assume the surface tension to be constant,
γ = cst. Write down all interfacial conditions.

Solution
Kinematic condition:
Since particles at the interface have a z-coordinate satisfying z = h(x, y, t), we have
upon applying the material derivative D/Dt to this equation Dz/Dt = Dh/Dt, hence
w = Dh/Dt, that is
ht + uhx + vhy = w

Dynamic condition:
Since n points towards fluid 2 and γ = cst, we have

σ 2 · n − σ 1 · n − γn(∇ · n)

Also, since the fluid is assumed to be inviscid, µ = 0, we have σij = −pδij or σ = −pI
where I is the identity matrix. Hence, the dynamic condition implies that

p(2) − p(1) + γ∇ · n = 0

i.e. p(z = h+ ) − p(z = h− ) + γ∇ · n = 0. This can be written in the form of the jump
condition +
JpKz=h
z=h− + γ∇ · n = 0

7
5 Kelvin-Helmholtz and Rayleigh-Taylor instabil-
ities
Kelvin-Helmholtz (KH) instability is experienced by an interface across which a dis-
continuity or a steep variation in the tangential fluid velocity exists. Rayleigh-Taylor
instability is experienced by an interface separating a dense fluid lying on top of a
lighter fluid under the effect of gravity.
An important example of KH instability: wind blowing over water. If the wind
speed is sufficiently large, an instability is observed which may lead to wave genera-
tion.

Model
We consider an inviscid, incompressible 2D flow, and account for the presence of
gravity and the surface tension of the interface. In the unperturbed state, the interface
is horizontal, located at z = 0, and separates an upper fluid with density ρ2 and
velocity U2 i from a lower fluid with density ρ1 and velocity U1 i. In the perturbed
state, the interface is given by z = h(x, t). The flow velocity will be denoted by v =
(u, w) and the pressure by p.

U2
ρ2 Disturbed interface z = h(x, t)

z=0
g = −gk
Undisturbed interface
U1 ρ1

The problem is governed on each side of the interface by Euler equations

∇p
∇ · v = 0, vt + (v · ∇)v = − +g for − ∞ < x < ∞ , z ̸= h(x, t),
ρ
subject to the usual kinematic and dynamic conditions at the interface (derived in
an example of last chapter), namely

1
 Dh
w(z = h±) = , JpKh+
z=h− + γ∇ · n = 0 ,
Dt
where n is a unit vector normal to the interface pointing in the direction of increasing
z and γ is the surface tension.

Basic Solution
We want to examine the stability of the basic solution (v, p, h) = (v̄, p̄, h̄) correspond-
ing to a flat interface given by
{
U2 for z > 0
h̄ = 0 , p̄ = p̄(z) and v̄ = ūi with ū = .
U1 for z < 0
From Euler’s equation vt + (v · ∇)v = −∇p/ρ − gk it follows that
{
−ρ2 gz for z > 0
p̄ = ,
−ρ1 gz for z < 0
choosing p̄(0) = 0.

Linear stability problem

Add perturbations of small amplitude ϵ such that
(u, w, p, h) = (ū, 0, p̄, 0) + ϵ(u′ , w′ , p′ , h′ ) ,
and plug these expressions into the governing equations to obtain at O(ϵ)
∇p′
∇ · v′ = 0 , vt′ + ūvx′ = −
ρ
Take the divergence of the second equation and use first ⇒
∇2 p ′ = 0 (5.1)
This is to be solved in the domains z > 0 and z < 0, subject to the linearised
kinematic and dynamic boundary conditions6
( ( ) )2
p′z U2
(z = 0±) = − ∂t + ∂x h′ (5.2)
ρ U1
′ ′
JpK0+
z=0− = (ρ2 − ρ1 )gh + γhxx , (5.3)
with p′ required to vanish as z → ±∞.
6
Full details of the simple algebra involved are provided in the visualiser notes.

2
Normal modes
Since there is no explicit dependence on x and t, look for normal modes in the form

p′ = p̂(z) exp (st + ikx) and h′ = C exp (st + ikx) (5.4)

where s is in general complex, k is real (since the x-domain is infinite), and C is a

constant. Then, ∇2 p′ = 0 implies that
{
Ae−kz (z > 0)
p̂zz − k p̂ = 0 ⇒ p̂ =
2
,
Bekz (z < 0)

assuming k ≥ 0, where A and B are constants. Using this results in equations (5.2)
and (5.3), we obtain for A, B and C the system of equations
  
−k 0 ρ2 (s + ikU2 )2 A
k 0 ρ1 (s + ikU1 )2  B  = 0
1 1 (ρ2 − ρ1 )g − γk 2 C

Dispersion relation
The system of equations above is linear and homogeneous. For non trivial solutions,
set its determinant to zero to obtain the dispersion relation

ρ1 (s + ikU1 )2 + ρ2 (s + ikU2 )2 = (ρ2 − ρ1 )gk − γk 3 .

This can be written as

[ ]
(ρ1 + ρ2 ) s2 + 2ik (ρ1 U1 + ρ2 U2 ) s − α + k 2 (ρ1 U12 + ρ2 U22 ) = 0 ,

where α ≡ (ρ2 − ρ1 )gk − γk 3 . The roots of this quadratic equation are

√
−ik (ρ1 U1 + ρ2 U2 ) ± ∆
s= with ∆ = (ρ1 + ρ2 ) α + k 2 ρ1 ρ2 (U1 − U2 )2 .
ρ1 + ρ2
√
Since Re(s) = ± ∆, a mode with wave number k is unstable if ∆ > 0 (due to +
sign); otherwise it is neutrally stable (since then Re(s) = 0).

3
Discussion
(a) If α ≡ (ρ2 − ρ1 )gk − γk 3 = 0, that is (ρ2 − ρ1 )g = 0 and γ = 0, and U1 ̸= U2
√ ∆ > 0 ∀k. All modes are unstable with growth rate σ(k) = Re(s(k)) =
then
∆/(ρ1 +ρ2 ). Since σ(k) ∝ |U2 −U1 |k, larger values of k have larger growth rates.
Hence smaller wavelengths are more unstable. This is the Kelvin-Helmholtz
instability in its purest form. In fact, here we have σmax ≡ max σ(k) = +∞, so
k
that the linear stability problem is ill posed (see chapter 2).
(b) If α ̸= 0 and ρ2 > ρ1 (top heavy), then ∆(k) > 0 for sufficiently small k and hence
the configuration is unstable. This is the Rayleigh-Taylor instability, which is
in its purest form when U1 = U2 (no shear flow) and γ = 0 (no surface tension).
Note that surface tension plays a stabilising role. More precisely, surface ten-
sion always stabilises modes with sufficiently large k (or short wavelength), since
∆(k) < 0 for sufficiently large k due to the term −γk 3 which is then dominant.
(c) Assume now that ρ2 < ρ1 (bottom heavy configuration, say air over water). In
this case, we can still have an unstable configuration if |U2 −U1 | is above a critical
value. Indeed, the instability of the basic solution is still given by ∆(k) > 0 for
some k. This instability condition can be written in the form
ρ1 ρ2 g
(U2 − U1 )2 > R(k) ≡ γk + (ρ1 − ρ2 ) for some k ,
ρ1 + ρ2 k
or
ρ1 ρ2
(U2 − U1 )2 > minR(k)
ρ1 + ρ2 k

An elementary function study shows that R(k) has a minimum at kc with

√
(ρ1 − ρ2 )g √
kc = and R(kc ) = 2 (ρ1 − ρ2 )gγ
γ
Using these expressions in the last equation, we can determine the critical con-
dition for instability and the critical wavelength λc = 2π/kc at the onset of
instability. These are given by
√
2(ρ1 + ρ2 ) √ γ
(U2 − U1 ) >
2
(ρ1 − ρ2 )gγ and λc = 2π .
ρ1 ρ2 (ρ1 − ρ2 )g
Application: for the case of air (wind) blowing with velocity U over still water,
we have ρ1 = 1000kg/m3 , ρ2 = 1.25kg/m3 , γ = 72. 10−3 N/m, g = 9.8m/s2 ,
U1 = 0, and U2 = U , and the instability criterion requires that
U ≥ 6.5m/s and λc = 1.7cm

4
6 Capillary instability/breakup of a liquid jet
This instability is experienced by cylindrical jets of liquid which are commonly ob-
served to break up into droplets. The mathematical theory describing this instability
is due to a paper by Rayleigh in 1879, although experimental investigation of the in-
stability started about 50 years earlier, notably by Savart and Plateau. The instability
is also known as the Rayleigh-Plateau instability .

n Disturbed interface r = f (θ, z, t)

p∞

r Undisturbed interface
2a
z

Model
We consider a jet in a stagnant atmosphere (or vacuum). We assume that the flow
is incompressible, and we neglect the effect of gravity and viscosity. The jet is rep-
resented by the surface r = f (θ, z, t), where (r, θ, z) are standard cylindrical coor-
dinates relative to an orthonormal basis (er , eθ , ez ). The notation v = (u, v, w) =
uer + veθ + wez is used for the velocity field and n denotes a unit vector normal to
the interface pointing in the direction of increasing r; this is given by
∇(r − f ) er − (fθ /r)eθ − fz ez
n= = 1
|∇(r − f )| (1 + fθ2 /r 2 + fz2 ) 2
The problem is governed by Euler equations
∇p
∇·v = 0, vt + (v · ∇)v = − for − ∞ < z < ∞ & r < f (θ, z, t) ,
ρ
subject to the kinematic and dynamic condition at the interface, namely
Df
u= , p = p∞ + γ∇ · n at r = f
Dt
1
Basic Solution
In the unperturbed state, the jet is a circular cylinder with radius a, and our basic
solution (v, p, f ) = (v̄, p̄, f¯) is given by
γ
v̄ = W ez , f¯ = a and p̄ = p∞ + ,
a
where W ez is the jet velocity assumed constant (clearly the equations and boundary
conditions are satisfied by the basic solution). Choosing the frame of reference to be
moving with the jet velocity, we can set W = 0 henceforth, and hence v̄ = 0.

Linear stability problem

Add perturbations of small amplitude ǫ such that
γ
(v, p, f ) = (0, p∞ + , a) + ǫ(v′ , p′ , f ′) ,
a
and plug these expressions into the governing equations to obtain at O(ǫ)
∇p′
∇ · v′ = 0 , vt′ = −
ρ
Take the divergence of the second equation and use first ⇒
∇2 p ′ = 0 (6.1)
This is to be solved in the domain r < a, subject to the linearised kinematic and
dynamic conditions8
ρftt′ = −p′r (r = a)
γ

p′ (r = a) = − 2 f ′ + fθθ ′
+ a2 fzz
′
a
Note that the linearised kinematic condition, which takes originally the form ft′ =
u′ (r = a), has been rewritten after using the r-component of the linearised Euler
equation ρu′t = −p′r to eliminate u′. Differentiating the linearised dynamic condition
above twice with respect to t and using ρftt′ = −p′r allows f ′ itself to be eliminated,
resulting in the boundary condition
γ
γ

p′tt = 2 p′r + p′rθθ + a2 p′rzz = 2 1 + ∂θθ + a2 ∂zz p′r at r=a (6.2)
a a
The final problem is therefore given by the single Laplace equation (6.1) subject
to the boundary condition (6.2) and the requirement that p′ is bounded as z → ±∞.
8
The full details of the derivation are provided in the video lectures and corresponding see visu-
aliser notes.

2
Normal modes
Since there is no explicit dependence on z, θ and t, look for normal modes exponential
in these variables in the form

p′ = p̂(r) exp (st + ikz + inθ) (*)

where s is in general complex, k real (since the z-domain is infinite), and n integer
(so as to satisfy the physical requirement that p′ is unaffected if θ is increased by 2π,
hence exp(2iπn) = 1).
Now equation (6.1), written in terms of cylindrical coordinates as

1 ∂ 1
∇2 p ′ = (rp′r ) + 2 p′θθ + p′zz = 0 ,
r ∂r r
implies for the normal modes(*) that

1 d n2
(r p̂r ) − 2 p̂ − k 2 p̂ = 0 or
r dr r

r 2 p̂rr + r p̂r − (k 2 r 2 + n2 )p̂ = 0 (6.3)

Also the boundary condition (6.2) implies when applied to the normal modes (*) that

γ
p̂r (a)
s2 = 2
1 − n2 − a2 k 2 (6.4)
ρa p̂(a)

Dispersion relation
ODE (6.3) is a modified Bessel equation which admits In (kr) and Kn (kr) as two
independent solutions, known as modified Bessel functions of the first and second
kind, respectively. Since Kn (r) are known to blow up as r → 0 they are eliminated,
and therefore the general solution of (6.3) is given by

p̂ = CIn (kr) ,

in which k and n appear as wavenumbers. Using this formula in equation (6.4)

determines that the dispersion relation is given by

γ 2 2 kIn (ka)

′
s2 = 1 − n2
− a k
ρa2 In (ka)

3
Implications
In terms of α ≡ ak (dimensionless wavenumber), the dispersion relation takes the
form
γ
αIn′ (α)
s2 = 3 1 − n2 − α2
ρa In (α)
The two fractions on the rhs are always positive, and so the sign of the rhs is deter-
mined by the factor 1 − n2 − α2 . Therefore,

(a) if n 6= 0, then s2 < 0 ∀α, and hence s is purely imaginary. So modes with n 6= 0
are neutrally stable; in other words the basic solution is stable for all θ-dependent
(or non-axisymmetic) perturbations.

(b) If n = 0, then
γ 2 αI0 (α)

′
s2 = 1 − α
ρa3 I0 (α)
and so s2 < 0 if |α| > 1 (corresponding to stable modes) and s2 > 0 if |α| < 1
(unstable). Since |α| < 1 ⇔ |ak| < 1 ⇒ λ ≡ 2π/|k| > 2πa, we conclude that the
jet is unstable to axisymmetric perturbations whose wavelength λ is
larger than the circumference 2πa of the jet. This is a long wavelength
instability.
p
A plot of the normalised growth rate S ≡ s/ γ/ρa3 versus |α| shows that S(α)
has a maximum, equal to 0.343, at α = 0.697. Hence the most unstable mode
has a growth rate s∗ and wavelength λ∗ given by
p 2πa
s∗ = 0.343 γ/ρa3 and λ∗ = ≈ 9a
0.697
It follows that thinner jets (corresponding to smaller values of a) disintegrate
(into droplets) faster. The typical size R of the droplets obtained can be roughly
estimated by equating the volume of the liquid in a unit wavelength λ∗ to the
volume of a droplet, namely,

πa2 λ∗ ≈ πa2 (9a) = 4πR3 /3 ⇒ R ≈ 1.9a ,

that is, the typical droplet radius is about twice the jet radius.

4
Physical mechanism of the instability
The mechanism is the coupling between surface tension and the cylindrical geometry
of the problem. More precisely, to explain the instability we can use Laplace (excess)
pressure formula, namely
p = p∞ + γ∇ · n , (**)
where p is the pressure on the liquid side of the interface, p∞ the ambient pressure,
and n the outward unit normal.

n A A′ p∞
B
rA r
rB
2a
z

Consider a sinusoidal perturbation of the interface as in figure. Consider point

A corresponding to a crest on the perturbed interface, and point B corresponding
to a trough. For large wavelengths λ ≫ a, the curvature at A and B is mainly
determined by the normal sections perpendicular to the plane of paper:
1 1
∇ · n|A ≈ ∇ · n|B ≈ ,
rA rB
where rA and rB are the radial distances from the axis of the unperturbed cylinder.
Since rA > rB , ∇ · n|A < ∇ · n|B , and hence using (**) pB > pA , that is the pressure
at B is larger than the pressure at A. This will generate a flow from B to A (and
A′ ), which amplifies the perturbation of the interface9 .

9
See also the little exercise below where the argument above is made more accurate, thus precisely
predicting the range of unstable wavelengths.

5
Exercise (to explain jet instability)
Consider an axi-symmetric interface with equation r = f (z) and unit outward normal
vector n.

1. Determine ∇ · n (twice the mean curvature)

2. In the special case f (z) = a + ǫ cos kz corresponding to a cylindrical interface

with radius a subject to a sinusoidal perturbation with wave number k and
small amplitude ǫ (0 < ǫ ≪ 1), determine ∇ · n at a crest (point A in previous
figure where kz = 0 modulo 2π, e.g. z = 0) and a trough (point B where kz = π
modulo 2π, e.g. kz = π)

3. Applying Laplace formula pin − pout = γ∇ · n at the interface, calculate pB − pA

and hence derive a condition for instability.

6
7 Stability of parallel flows and the Orr-Sommerfeld
equation (OSE)
In this section we examine the stability of a parallel incompressible flow U(y)i. This
reduces to solving the Orr-Sommerfeld equation (OSE). The OSE follows most simply
from the equation of the stream function in 2D.

7.1 Preliminaries: stream function and vorticity

The vorticity is defined by ω = ∇ × v. Starting from the non-dimensional Navier-
Stokes equation
−∇p 1 2
vt + v · ∇v = + ∇ v+g (7.1)
ρ Re
taking the curl of both sides, and using a few vector identities (as done in section 4
and in example sheet 3, Q2), leads to the vorticity equation
1 2
ωt + v · ∇ω = ∇ ω + ω · ∇v (7.2)
Re
For a 2D flow v = (u, v) = v(x, y, t), ω = ωk, with ω = vx − uy , ω · ∇v = 0, and

1 2
ωt + v · ∇ω = ∇ω (7.3)
Re
For a 2D flow a stream function ψ can be defined by

u = ψy , v = −ψx (7.4)

which implies that ∇ · v = 0 automatically. Since v · ∇ψ = 0, it is clear that ψ is

constant on streamlines. Also, we have v = ∇ × (ψk), hence by taking the curl of
v, we have ω = (−∇2 ψ)k and ω = −∇2 ψ. Substituting the last identity in (7.3), it
follows that
∂
∂
∂
1 4
∇2 ψ + u ∇2 ψ + v ∇2 ψ = ∇ψ (7.5)
∂t ∂x ∂y Re
and, on using (7.4),

∂
∂
∂
1 4
∇2 ψ + ψy ∇2 ψ − ψx ∇2 ψ = ∇ ψ (7.6)
∂t ∂x ∂y Re

1
7.2 The Orr-Sommerfeld equation (OSE)
Z
Consider the basic flow U(y)i with stream function ψe = U(y)dy. Add perturbations
of small amplitude ǫ such that ψ = ψe + ǫψ ′ and substitute into (7.6) using the fact
that ψet = 0, ψex = 0, ψey = U, ∇2 ψe = Uy , to get
∂
∂
∂
1 4
∇2 ψ ′ + U ∇2 ψ ′ − ψx ∇2 ψ = ∇ψ (7.7)
∂t ∂x ∂y Re
The OSE is obtained by introducing into (7.7) normal modes of the form

ψ ′ = φ(y)eik(x−ct) (7.8)

This implies that

1
2
(U − c) D 2 − k 2 φ − Uyy φ = D2 − k2 φ (7.9)
ikRe
d
where (D ≡ dy
), or equivalently

1

(U − c) φyy − k 2 φ − Uyy φ = φyyyy − 2k 2 φyy + k 4 φ
ikRe
Equation (7.9) is a fourth-order ODE which requires four boundary conditions. These
are
φ=0 and φy = 0 at y = y1 , y2
where y1 and y2 are typically the locations of the walls bounding the flow domain. The
first condition corresponds to the no-penetration condition (u = 0 ⇒ u′ = 0 ⇒ φ = 0)
and the second to the no-slip condition (v = 0 ⇒ v ′ = 0 ⇒ φy = 0). If y1 or/and y2
is infinite, conditions are still applicable.

Remarks:
1. The normal modes (7.8) have the exponential factor eik(x−ct) , which is the same
as the factor est+ikx (used in all other chapters) provided s = −ikc. Given
our notation s = σ − iω and c = cr + ici , we have σ = kci (growth rate) and
ω = kcr (angular frequency), and normal modes (7.8) take the form

ψ ′ = φ(y)ekcit eik(x−cr t)

2
 These represent amplified/damped travelling waves with phase speed cr often
referred to as Tollmien-Schlichting (or TS) waves, especially when amplified. A
mode is amplified if ci > 0; we are assuming throughout that k > 0 without
loss of generality as the direction of propagation is still free and determined by
the sign of cr .
2. The OSE is the basis for studying the stability of parallel flows, or nearly parallel
flows (such as boundary layers, jets, mixing layers, etc.). Furthermore the 2D
stability analysis undertaken is sufficient to predict the onset of instability as
Re is increased. This is justified by Squire’s theorem (1933): “ To each growing
3D perturbation at a given Re, a 2D growing perturbation can be found for
smaller Re”.
3. An alternative more common way of deriving the OSE is given in Example
sheet 5 (question 1) along with a proof of Squire’s theorem.

7.3 The Rayleigh equation (inviscid stability)

If viscous effects are neglected in the OSE (7.9), we get by letting Re → ∞ the
Rayleigh equation

(U − c) D 2 − k 2 φ − Uyy φ = 0 (7.10)
which is to be solved in a domain (y1 , y2 ) subject to the boundary conditions
φ=0 y = y1 , y2 (7.11)
The boundary conditions follow from the no-penetration condition while the no-slip
condition is not applicable (within the inviscid flow assumption). ODE (7.10) subject
to (7.11) is an eigenvalue problem with eigenvalues c(k) and eigenfunctions φ. Unlike
the OSE, the Rayleigh equation has real coefficients. This implies that if c is an
eigenvalue with eigenfunction φ, then c̄ is an eigenvalue10 with eigenfunction φ̄. To
see this, just take the complex conjugate of (7.10) to obtain


(U − c̄) D 2 − k 2 φ̄ − Uyy φ̄ = 0
It follows that to each eigenvalue with positive ci (growing mode), there corresponds
an eigenvalue with negative ci (decaying mode). Therefore, any stable mode with
wave number k, say, is in fact neutrally stable, i.e. c(k) is real.
10
The bar indicates the complex conjugate.

3
 Two useful necessary conditions for instability of the basic flow U(y) follow from
(7.10) and (7.11):

Rayleigh’s inflection point criterion (1880):

A necessary condition for the inviscid instability of the parallel flow U(y) is that U(y)
has a point of inflection in the flow domain; say I where U ′′ (yI ) = 0.

Fjortoft’s criterion (1950):

A necessary condition for the inviscid instability of the parallel flow U(y) is that U(y)
has a point of inflection I and that (U −UI )U ′′ < 0 in a subinterval of the flow domain
(UI ≡ U(yI )).

Proof of Rayleigh’s and Fjortoft’s criteria:

Assume the existence of an unstable mode, i.e. a mode for which ci > 0, and hence
U − c 6= 0 since U(y) is real. Divide (7.10) by U − c, multiply by φ̄ and integrate
with respect to y from y1 to y2 . On integrating the term φ̄D 2 φ by parts and using
the boundary conditions (7.11), we get
Z y2 Z y2
2 2 2

(U − c̄)U ′′ 2
|φy | + k |φ| dy + |φ| dy = 0 (*)
y1 y1 |U − c|2

The first integral is always real. Taking the imaginary part of the second integral
implies that Z y2
U ′′
ci 2
|φ|2dy = 0 (**)
y1 |U − c|

Since ci 6= 0, the last integral must be zero which implies that the integrand, and
hence U ′′ , must change sign in (y1, y2 ). An inflection point yI where U ′′ (yI ) = 0 must
therefore exist. This completes the proof of Rayleigh’s criterion.
The proof of Fjortoft’s criterion is obtained by taking the real part of (*), which
yields Z y2 Z y2
(U − cr )U ′′ 2

2
|φ| dy = − |φy |2 + k 2 |φ|2 dy
y1 |U − c| y1

and adding to it Z y2
U ′′
(cr − UI ) |φ|2 dy = 0
y1 |U − c|2

4
The fact that the rhs of last equation is zero is an obvious consequence of (**). It
follows that
Z y2 Z y2
(U − UI )U ′′ 2

2
|φ| dy = − |φy |2 + k 2 |φ|2 dy < 0 .
y1 |U − c| y1

We conclude that (U − UI )U ′′ < 0 in a subinterval of the domain (y1 , y2 ), completing

the proof.

Illustration of the application of Rayleigh’s and Fjortoft’s criteria

U ′′ > 0 U ′′ (U − UI ) > 0
′′
U <0
U ′′ < 0 U ′′ (U − UI ) > 0
Stable by Rayleigh Stable by Fjørtoft

U ′′ < 0 U ′′ (U − UI ) < 0
′′
U >0
U ′′ > 0 U ′′ (U − UI ) < 0

Stable by Rayleigh May be unstable

Alternative form of Fjortoft’s criterion:

For a monotone velocity profile U(y) with a single point of inflection at yI , a necessary
condition for instability is that the magnitude of the vorticity ω ≡ −U ′ (y) is maximum
at yI .

Proof:
See Example sheet 5, question 3. Also, recall that ω = ∇ × (U(y)i), hence ω = ωk =
−U ′ (y)k.

5
Example 1 (applying Rayleigh’s and Fjortoft’s criteria)
Refer to the figure below where the profiles of several unidirectional flows U(y) are
plotted. Inflection points where U ′′ is zero are denoted by a thick dot. Which profiles
are potentially unstable according to Rayleigh’s criterion and to Fjortoft’s criterion?

Answer
Flows which could be unstable according to both criteria are those in the last two
figures in row 2, and those in the second, fourth, and fifth figures in row 3.

- Channel flows

- Boundary layers

p <0 p =0 p >0 Twall > Tfluid

x x x
- Unbounded shear layers

6
 Another general result which can be derived similarly to Rayleigh’s and Fjortoft’s
criteria (see Drazin and Reid) is the following:

Howard’s semi-circle theorem:

The eigenvalues c of amplified and neutral modes lie in the complex cr − ci plane
inside or on a semi-circle centred at ( Umax +U
2
min
, 0) with radius Umax −U
2
min
, where Umax
and Umin are the maximum and minimum of U(y) in the flow domain. Thus, all
unstable modes lie in the shaded domain sketched below.

ci

cr

Umin Umax + Umin Umax

2

Howard’s theorem places bounds on the phase speed cr (in the positive x-direction)
and the growth rate kci ; recall that our normal modes are proportional to ekci t eik(x−cr t) .
In particular, Umin ≤ cr ≤ Umax ; hence if e.g. Umin and Umax are ≥ 0, then the propa-
gation is to the right. Also, the growth rate is bounded by k(Umax2−Umin) .

7
Critical layer (and the continuous spectrum (optional))
Consider a mode k, c which is neutral, thus c = cr is real and Umin ≤ c ≤ Umax . It
follows that there exists yc ∈ (y1 , y2 ) such that U(yc ) = c. The plane y = yc is called
a critical layer. At yc , the highest derivative in Rayleigh’s equation (7.10) drops out
and the eigenfunction φ(y) may become singular (typically φ′′ infinite at yc and hence
φ′ discontinuous there). Such singular solutions φ are termed singular neutral modes,
and their eigenvalues c which take all values in [Umin , Umax ] constitute the continuous
spectrum. It should be noted however that since the continuous spectrum is associated
with neutral modes, it may be ignored for deciding whether the basic flow U(y) is
stable or not. It is also worth noting that the singularity of φ at yc does not occur if
U ′′ (yc ) = 0 (unless U ′ (yc ) = 0); indeed a Taylor expansion around yc implies that

U ′′ U ′′′ (yc )
∼ ′ as y → yc
U −c U (yc )

so that yc is no more a singular point of Rayleigh’s equation (7.10). In this case,

which requires U(y) to have a point of inflection at yI , and for which yc = yI with
U ′′ (yI ) = 0, we have a regular neutral mode with eigenvalue c = U(yI ).
We close this section by emphasising that the continuous spectrum is in addition
to the usual discrete spectrum of complex conjugate pairs of eigenvalues associated
with our normal modes. It can be shown11 that the number of pairs (for a given k)
is ≥ 0 and ≤ number of inflection points in the profile of U(y). In particular, if U(y)
has no points of inflection, then the flow is necessarily stable, which is in agreement
with Rayleigh’s inflection point criterion.
A more fundamental question may be raised here, related to the use of the method
of normal modes, which assumes that any initial perturbation may be expressed as
a combination of normal modes, and hence implicitly that the normal modes form a
complete set. This latter assumption is seldom proven rigourously, and it is certainly
incorrect12 in the case of Rayleigh’s equation (see e.g. next example). Still the method
of normal modes is widely used and successful in deciding stability, even in our case.

11
L.N. Howard, The number of unstable modes in hydrodynamic stability problems, J. Mécanique
3 (1964), 433-443.
12
However, when the eigenfunctions of the continuous spectrum are added to the normal modes
then a complete set of eigenfunctions is expected.

8
Example 3 (Couette flow: stability, continuous spectrum)
Determine the discrete spectrum, the continuous spectrum and the stability of Cou-
ette flow given by
U(y) = y , −1 ≤ y ≤ 1

Solution
Rayleigh’s equation (7.10) for this flow is given by

(y − c)(φ′′ − k 2 φ) = 0 , −1 < y < 1 , subject to φ(−1) = φ(1) = 0

Case 1: c 6∈ [−1, 1] (range of U). Then, we have to solve φ′′ − k 2 φ = 0 subject to the
boundary conditions φ(−1) = φ(1) = 0. This implies that φ = Aeky + Be−ky where A
and B satisfy the homogenous system of equations Ae−k + Bek = 0, Aek + Be−k = 0.
For non trivial solutions, the determinant of this system must vanish, which implies
that k = 0, and hence φ ≡ 0 (not acceptable). Therefore, the discrete spectrum is
empty and hence the flow is (inviscidly) stable.
Case 2: c ∈ (−1, 1). Then, for y 6= c, φ′′ − k 2 φ = 0. Solving separately in the two
domains −1 < y < c and c < y < 1, applying the boundary conditions at y = ±1, and
requiring φ to be continuous at y = c implies that φ is given within a multiplicative
constant by
 sinh(1+y)

 sinh(1+c) (−1 < y < c)
φ=

 sinh(1−y)
sinh(1−c)
( c < y < 1)
These are singular neutral modes (with discontinuous derivative at y = c) correspond-
ing to the continuous spectrum of eigenvalues c ∈ (−1, 1). As mentioned earlier, being
neutral, they play no part when deciding for stability of the basic flow.

7.4 Piecewise-linear velocity profiles (simple calculations)

The difficulty in solving the Rayleigh equation (7.10) lies in the y-dependence of the
coefficients. However, for basic profiles made up of linear segments so that U ′′ = 0,
the Rayleigh equation simplifies to

φ′′ − k 2 φ = 0 (7.12)

9
with general solution φ = Aeky + Be−ky in intervals not containing discontinuities of
U and/or U ′ . This suggests a method for estimating the stability of smooth profiles
by approximating them by piecewise-linear profiles, e.g. as in the following figures:

U U

At points of discontinuity of U and/or U ′ , the following jump conditions are to

be satisfied:
Jump condition 1:
J(U − c)φ′ − U ′ φK = 0 (7.13)
Jump condition 2:
s {
φ
=0 (7.14)
U −c
These jump conditions follow from the dynamic condition and the kinematic condition
at the discontinuity surface, say y = y0 + ǫη(x, t), assumed to be a material surface
(see Example sheet 5, question 4).

Example 4: Kelvin–Helmholtz instability

A typical example of a shear instability is revealed in an experiment initially per-
formed by Reynolds (1883) and later by Thorpe (1969), and shown in Fig. 7.1(a). A
horizontal tube is filled with a layer of water over a layer of brine (coloured). The
tube is suddenly tilted by a few degrees, which sets the fluids into motion: the brine
descends while the (lighter) water rises, leading to the basic counterflow of two lay-
ers (i.e. a mixing layer velocity profile, which is characterised by the presence of an
inflexion point). Kelvin–Helmholtz instabilities of this type are found in wide variety
of other situations as illustrated in Fig. 7.1.

10
 (b)

(a)

Figure 7.1: (a) Kelvin-Helmholtz instability at the sheared interface between two
fluids (Thorpe 1971, Van Dyke 1982). (b) Clouds provide flow visualisation for shear
layer instability in the lower atmosphere.

We consider here the simplest model of a shear layer:


U2 if y > 0
where U =
U1 if y < 0
U

In each layer, the Rayleigh equation (7.10) reduces to φ′′ − k 2 φ = 0. Hence, the
general solution satisfying φ = 0 as y → ±∞ is given by

Ae−ky if y > 0
φ=
Be ky if y < 0
Applying the jump condition (7.13) and (7.14), we get
(U2 − c)A + (U1 − c)B = 0
(U1 − c)A − (U2 − c)B = 0
This homogeneous linear system has a non-trivial solution iff its determinant vanishes.
This yields the dispersion relation for the Kelvin–Helmholtz instability on a vortex
sheet
(U1 − c)2 + (U2 − c)2 = 0

11
hence
1 i
c = (U1 + U2 ) ± (U2 − U1 ) (7.15)
2 2
as found and discussed in a previous lecture. In particular, we note that we have
instability to all wavelengths whenever U1 6= U2 , with short wavelengths being most
unstable (since the growth rate kci is proportional to k, that is larger for larger k or
smaller wavelength). This non-physical feature of the growth rate arises because the
interface (between the upper and lower layers) is idealised as being infinitely thin.
In reality, the interface has a finite thickness, and the idealisation is valid only for
perturbations with wavelengths much larger than this thickness.

Example 5: Mixing layer with finite thickness

We now consider a mixing layer of thickness h,

where 
 U2 if y > h/2
U1 +U2 (U2 −U1 )y
U= 2
+ h
if |y| < h/2

U1 if y < −h/2
In each layer, the Rayleigh equation (7.10) reduces to φ′′ − k 2 φ = 0. Hence, the
general solution satisfying φ = 0 as y → ±∞ is given by

 Ae−ky if y > h/2
φ= Be−ky + Ceky if |y| < h/2
 ky
De if y < −h/2

Applying jump condition (7.13) at y = h/2, we get

U2 − U1
−(U2 − c)kAe−kh/2 = (U2 − c)(−kBe−kh/2 + kCekh/2 ) − (Be−kh/2 + Cekh/2 )
h
12
Applying jump condition (7.14) at y = h/2, we get

Ae−kh/2 Be−kh/2 + Cekh/2

=
U2 − c U2 − c
Similarly, applying jump conditions (7.13) and (7.14) at y = −h/2, we get

U2 − U1
(U1 − c)kDe−kh/2 = (U1 − c)(−kBekh/2 + kCe−kh/2 ) − (Bekh/2 + Ce−kh/2 )
h
De−kh/2 Bekh/2 + Ce−kh/2
=
U1 − c U1 − c
The last four equations constitute a homogeneous linear system for A, B, C and D.
Setting its determinant to zero (lengthy) provides the dispersion relation

U1 + U2 U2 − U1  1/2
c= ± (1 − kh)2 − e−2kh (7.16)
2 2kh
In the long-wave limit, i.e. kh ≪ 1, (7.16) reduces to
1 i
c ∼ (U1 + U2 ) ± (U1 − U2 ) (7.17)
2 2
For a given k, we have instability iff f (kh) ≡ (1−kh)2 −e−2kh < 0; an elementary study
of the function f reveals that f < 0 for 0 < kh < 1.278 and that the minimum of f
occurs at kh = 0.797. We conclude that the mixing layer is unstable to perturbations
with large wavelengths λ, such that λ/h = 2π/kh > 2π/1.278 ≈ 5. Perturbations
with smaller wavelengths are stable. Finally we note that in the long wavelength
limit kh ≪ 1, a Taylor expansion shows that the dispersion relation (7.17) reduces to
(7.15). This is in line with physical intuition that for long waves the interface appears
as having zero thickness.

13
7.5 The effect of viscosity

The Orr-Sommerfeld equation (OSE) is difficult to solve, most of all analytically.

Complications arise due to the presence of regions of rapid change at walls (boundary
layers) and near critical layers (of the Rayleigh equation (RE)). The limit Re → ∞
which reduces the fourth order OSE to the second order RE is singular. A numerical
approach is therefore needed and the remainder of this section presents qualitatively
essential results based on numerical calculations. The main points to keep in mind
are:

1. The effect of viscosity can be destabilising; i.e. a flow which is stable for Re = ∞
(no viscosity) can be unstable when Re is finite.

2. Mainly two types of diagrams in the Re-k are obtained depending on the base
flow U(y) having a point of inflection or not.

3. In the case where no point of inflection is present, as e.g. for a 2D Poiseuille flow
U = 1 − y 2, the flow is inviscidly stable (according to Rayleigh’s inflection point
criterion) and the stability regions in the Re-k are as sketched in figure 7.2.
Note that the marginal stability curves (ci = 0) consists of an upper branch
ku (Re), say, and a lower branch kl (Re), with both ku and kl → 0 as Re → ∞.
For any Re > Rec , with Rec ≈ 5772 for the 2D Poiseuille flow, we have linear
instability for a band of wavenumbers kl (Re) < k < ku (Re). Note however that
for any fixed k ∗ the flow is stable in the inviscid limit Re → ∞, consistent with
Rayleigh’s inflection point criterion.

4. In the case where a point of inflection is present, as e.g. for a boundary layer in
an adverse pressure gradient the stability regions in the Re-k are as sketched in
figure 7.3. Here the upper branch of the marginal stability curve (ci = 0) is such
that ku → kn as Re → ∞, where kn is a nonzero constant which is determined
by the requirement that the critical layer is found at the location yI of the
inflection point (hence c = U(yI ) is known, but kn needs to be computed). The
asymptote corresponds in fact to the regular neutral mode discussed earlier.

14
k
stable

k*
kc unstable
ku

kl
Rec Re

Figure 7.2: Marginal stability curve of Plane-Poiseuille flow.

15
 k stable

kc
unstable
kn

Rec Re

Figure 7.3: Marginal stability curve of boundary layer flow with adverse pressure
gradient.

16
8 Rayleigh-Bénard instability
This instability is experienced by a horizontal layer of fluid heated from below.

8.1 Rayleigh-Benard problem

TC (cold)

g d
y
TH (hot) x
⇒

instability

The basic state corresponds to steady heat conduction, that is zero-flow and a
constant vertical temperature gradient (top figure). If the temperature gradient is
large enough, an instability is observed which typically leads to convection rolls whose
size is of the order of the width d of the layer (bottom figure).
Physical mechanism: a hot particle tends to rise by buoyancy which is op-
posed by viscosity and heat conduction. The instability is controlled by the Rayleigh
number, a non-dimensional parameter defined by

α(TH − TC )gd3
Ra ≡ ,
νκ
where ν is the kinematic viscosity, κ the heat diffusivity, and α the thermal expansion
coefficient. The latter is defined by α = −(∂ρ/∂T )p /ρ; α ≈ 1.9 10−4K −1 for water
and 3.10−3K −1 for air.
It is observed experimentally that convection occurs when Ra ≥ Rac , a critical
value that we need to determine.

1
8.2 The Boussinesq approximation
For problems involving buoyancy, a simplified version of the governing equations (of
section 4) may be used, if the density variations are small, namely if

|∆ρ/ρ| ≪ 1 ⇔ α|∆T | ≪ 1 (by definition of α)

For later reference, we note that under these conditions, we have

ρ = ρ0 [1 − α(T − T0 )] , (8.1)

where T0 is a reference temperature which will be chosen to be equal to the cold

temperature, T0 = TC .
Boussinseq approximation: take density and other transport coefficients con-
stant except in buoyancy term, hence

∇·v =0 (8.2a)
Dv
ρ0 = −∇p + µ∇2 v + ρg (8.2b)
Dt
DT
= κ∇2 T (8.2c)
Dt
For a discussion of the Boussinesq approximation, see the book Hydrodynamic and
Hydromagnetic Stability by Chandrasekhar.

8.3 Vorticity equation within the Boussinesq approximation

Taking the curl of the Navier-Stokes equation (8.2b), we obtain for the vorticity
ω ≡ ∇ × v the equation
∇ρ
ωt + v · ∇ω = ω · ∇v + ν∇2 ω + g, (8.3)
ρ0

where ν = µ/ρ0 . This is essentially the same vorticity equation we derived in chapter 4
and used in chapters and 7, except for the last term which accounts for the gravity
term in the Boussinesq approximation . This term comes from the curl of the last
term in (8.2b), taking into account that ∇ × (ρg) = ∇ρ × g + ρ∇ × g and that
∇ × g = 0, g being constant.

2
 Assume now that the flow is two-dimensional, v = (u, v) = v(x, y, t), then ∇·v = 0
implies that a stream function ψ exists such that

u = ψy , v = −ψx , (8.4)

and ω = ωk with
ω = −∇2 ψ (8.5)
Since ω · ∇ = ωk · (i∂x + j∂y ) = 0, equation (8.3) simplifies, after using equa-
tion (8.1), to
ωt + v · ∇ω = ν∇2 ω − α(∇T × g) · k , (8.6)
hence, since (∇T × g) · k = [(Tx i + Ty j) × (−gj)] · k = −gTx , we have

ωt + v · ∇ω = ν∇2 ω + gαTx (8.7)

Using (8.5), the last equation implies that

D
(∇2 ψ) = ν∇4 ψ − gαTx , (8.8)
Dt
with
D ∂ ∂ ∂ ∂ ∂ ∂
= +u +v = + ψy − ψx
Dt ∂t ∂x ∂y ∂t ∂x ∂y

8.4 The stability problem

8.4.1 Formulation

DT
= κ∇2 T
Dt
D
(∇2 ψ) = ν∇4 ψ − gαTx
Dt
Non-dimensionalisation: referring to the figure given above, we choose as
reference length d and as reference time d2 /κ (noting that κ and ν have as dimension
m2 /s). Now let

1 t T − TC ψ
(x′ , y ′) = (x, y) , t′ = , θ= , ψ′ =
d d2 /κ TH − TC κ

3
In terms of these variables, the non-dimensional problem (after dropping primes, see
example sheet 6, question 1) becomes

Dθ
= ∇2 θ (8.9)
Dt
1 D
(∇2 ψ) = ∇4 ψ − Ra θx , (8.10)
Pr Dt
where
ν α(TH − TC )gd3
Pr = (Prandtl number) and Ra = (Rayleigh number)
κ νκ
Boundary conditions

θ(y = 0) = 1 , θ(y = 1) = 0

and
ψx = 0 , ψy = 0 (case of a rigid wall) ,
since these two conditions are equivalent to the no penetration v = 0 and no slip
u = 0 conditions as seen from (8.4), or

ψx = 0 , ψyy = 0 (case of a free surface) ,

with the second condition expressing a zero-shear stress on the surface in this case
(see example sheet 6, Question 1).

8.4.2 The linear stability problem

Our (non-linear) problem is given by equations (8.9) and (8.10) subject to the bound-
ary conditions just given.
Basic state (stationary solution and no flow):

θ = θ̄ = 1 − y , ψ = ψ̄ = 0

This solution obviously satisfies the equations and boundary conditions

Linearisation: Add perturbations of small amplitude ǫ such that

(θ, ψ) = (1 − y, 0) + ǫ(θ′ , ψ ′ )

4
to get at order ǫ
 
∂ 2
− ∇ θ = −ψx (8.11)
∂t
 
1 ∂
− ∇ ∇2 ψ = −Ra θx ,
2
(8.12)
Pr ∂t
to be solved in the domain −∞ < x < ∞, 0 < y < 1, subject to the boundary
conditions
θ = 0 , ψ = 0 , ψy = 0 (or ψyy = 0) at y = 0, 1 , (8.13)
with ψy = 0 being applicable in the case of a rigid

wall and ψyy = ∂0 in the case of a
∂ 2
free surface. Eliminating θ (by taking ∂t − ∇ (8.12) and using ∂x (8.11)), we get
  
∂ 2 1 ∂
−∇ − ∇ ∇2 ψ = Ra ψxx
2
(8.14)
∂t P r ∂t
This is a sixth order PDE with six boundary conditions given by (8.13). Note that
the boundary condition θ = 0 as y → 0, 1 can be written in terms of ψ. Indeed, θ = 0
as y → 0, 1 implies that RHS (8.12) → 0 and therefore LHS (8.12) → 0 as y → 0, 1,
hence  
1 ∂
− ∇ ∇2 ψ as y → 0, 1
2
Pr ∂t
So the 6 boundary conditions are
 
1 ∂
ψ = 0, ψy = 0 (or ψyy = 0) , − ∇ ∇2 ψ
2
as y → 0, 1 (8.15)
Pr ∂t
Normal modes: Since there is no explicit dependence on x and t, look for
normal modes exponential in these variables in the form

ψ = φ(y) exp (st + ikx) , (*)

where s is in general complex and k is real (since the x-domain is infinite). Now
equation (8.14) implies for the normal modes (*) that

 s 2

D2 − k2 − s D2 − k2 − D − k 2 φ = −k 2 Ra φ (8.16)
Pr
Also the boundary conditions (8.15) imply that
 s 2

φ = 0 , φy = 0 (or φyy = 0) , D2 − k2 − D − k 2 φ as y → 0, 1
Pr
5
The last boundary condition can be simplified to φyyyy = (2k 2 + s/Pr)φyy , by noting
that it can be expanded as
s s
φyyyy − (2k 2 + )φyy + k 2 (k 2 + )φ = 0
Pr Pr
and noting that the last term on the LHS is zero on account of φ = 0. Therefore, we
have
s
φ = 0, φy = 0 (or φyy = 0) , φyyyy = (2k 2 + )φyy as y → 0, 1 (8.17)
Pr
Of course, in the case of free surfaces, the boundary conditions simplify further to

φ = 0, φyy = 0 , φyyyy = 0 as y → 0, 1 (8.18)

In this case, the solution is analytically possible (simple), and is obtained in the next
section.
Remark: Using (8.18) and (8.16) implies that φyyyyyy = 0 as y → 0, 1, and in
fact that all even derivatives are zero as y → 0, 1.

8.4.3 Analytical solution (free surfaces)

We have to solve ODE (8.16) subject to boundary conditions (8.18). For each k, we
have an eigenvalue problem with eigenvalues s and eigenfunctions φ. The eigenfunc-
tions are given by
φ = A sin(nπy) n = 1, 2, 3 · · · ,
which clearly satisfy (8.18). Plugging this expression into (8.16) and noting that
D 2 → −n2 π 2 implies that

 s

−n2 π 2 − k 2 − s −n2 π 2 − k 2 − −n2 π 2 − k 2 φ = −k 2 Ra φ ,
Pr
that is, s 
(s + γ) + γ γ = k 2 Ra with γ ≡ n2 π 2 + k 2 ,
Pr
hence we obtain the dispersion relation:
 
2 2 k 2 Ra
s + γ(1 + Pr) s + γ − Pr = 0 (*)
| {z } γ
a1 | {z }
a2

6
The roots of this quadratic equation are given by
s
γ(1 + Pr) RaPr
s=− ± γ 2 (Pr − 1)2 + 4k 2
2 γ

Since the argument of the square root is positive, the roots s are real.
We now apply the Routh-Hurwitz stability criterion for the second order polyno-
mial (*), which states that the roots of (*) have negative real parts (here they are
simply < 0, since they are real) if and only if

k 2 Ra
a1 > 0 and a2 > 0 ⇔ a2 > 0 (since always a1 > 0) ⇔ γ2 − >0
γ

Therefore, the basic solution is stable iff Ra < γ 3 /k 2 , that is

( )
3 3
(n2 π 2 + k 2 ) (n2 π 2 + k 2 )
Ra < ∀k, n or Ra < Rac ≡ min
k2 k,n k2

Now, the minimum of the function of n and k on the rhs clearly

n occurs foron = 1
3
(as it is an increasing function of n), and hence Rac = min (π + k 2 ) /k 2 . The
2
k
minimum of this function of k is easily evaluated and occurs at kc with

π 27π 4
kc = √ ≈ 2.22 and Rac = ≈ 657.5
2 4
3
The neutral stability curve, Ra = (π 2 + k 2 ) /k 2 , and the region of the stability and
instability are shown in the figure below.
Conclusion: The basic solution becomes unstable when Ra exceeds Rac and
at the onset of instability k = kc . Note that Rac is independent of the Prandtl
number Pr, and this observation is in fact in agreement with experimental findings.
The value of Rac and kc given above are obtained for the case where the boundary
conditions correspond to both upper and lower surfaces being free, where we were able
to fully proceed with the calculations analytically. For other boundary conditions,
some numerics is required, and the main results related to the critical conditions are
summarised in the table below.
We note that the second row, pertaining to the case of two rigid surfaces, is the
case which is most common in experiments. As mentioned at the beginning of this
chapter, for Ra slightly above Rac , convection is observed (typically but not always) to

7
 Ra

unstable
Rac

stable

kc k
Conditions at the onset of instability

Boundary conditions Rac kc λc = 2π/kc

Free-Free 657.5 2.22 2.83
Rigid-Rigid 1700 3.12 2.02
Rigid-free 1101 2.60 2.34

set in in the form of counter-rotating rolls, each roll having size close to d, the spacing
between the horizontal plates; since the spacial period (wavelength) comprises two
rolls, it is equal to 2d. This is in good agreement with the prediction of row 2, where
the non-dimensional wavelength at the onset of instability is λc = 2.02, that is 2.02d
in dimensional terms.
Extensive investigations related to the Rayleigh-Benard instability have been car-
ried out to date, with the classical authoritative reference on the topic being the
massive (but quite readable) book Hydrodynamic and Hydromagnetic Stability by
Chandrasekhar.