124

6.3.3 An introduction to weakly nonlinear theory

Weakly nonlinear theory is the study of the dynamics of a system that is only
weakly nonlinear, that is, a system where the amplitude of the perturbations
is just large enough for the nonlinear terms to become relevant. This happens
when the control parameter (in our case, Ra) is close to the critical value Rac , in
which case this also means that there are usually only very few unstable modes.
The idea is to create a reduced set of equations that describes the nonlinear
interaction between these few modes.
There are a number of different ways of constructing weakly nonlinear equa-
tions – the procedure is not unique, and neither is the result. However, only
one way of studying the problem will actually yield equations whose behavior
correctly models observations. This is the main reason why weakly nonlinear
theory is so difficult to approach at first, and somewhat frustrating to learn.
However, with experience and practice, things get a lot easier. The key is to
already have an idea, ahead of time, of the types of dynamics we need to end up
with (based on experiments, or numerical simulations), and use this information
to find our way more directly.
Before we start, however, a small mathematical detour (that will make our
lives infinitely nicer later) is necessary.

Mathematical detour: Fredholm’s alternative and the solvability con-

dition
See textbook “Applied Partial Differential Equations” by Haberman, for more.

Consider the following ODE problem:

Lu(x) = F (x) (6.1)

on the interval [a, b], subject to homogeneous boundary conditions (that is,
u(a) = u(b) = 0). The operator L is assumed non-singular, and so is the forcing
F (x).
Fredholm’s alternative states that if L is a self-adjoint operator for the inner
product h·i (see definition below) and if there exists a non-trivial (i.e. u(x) is
not identically 0) solution to the homogeneous equation

Luh (x) = 0 (6.2)

then the original problem Lu(x) = F (x) has a solution only if

hF, uh i = 0 (6.3)

This is called a solvability condition. Note that there is actually more to Fred-
holm’s alternative, but for the purpose of what we are about to do, this is all
we need to know. See the textbook listed above for more detail.
Let’s now recast this jargon in slightly less obscure terms, and actually prove
that this statement is correct. First, note that a self-adjoint operator L is one
 125

for which there exists a weight function r(x), and a corresponding inner product
Z b
hf, gi = f (x)g(x)r(x)dx (6.4)
a

such that
hf, Lgi = hLf, gi (6.5)
Note that the inner product is symmetric, so we always have hf, gi = hg, f i for
all f , g.
If L is self-adjoint with that inner product, then it’s fairly easy to show that
different solutions vn to the eigenvalue problem
Lvn (x) = λn vn (x) (6.6)
are orthogonal if they have different eigenvalues λn . Indeed,
hvm , Lvn i = λn hvm , vn i (6.7)
but also, since L is self-adjoint, then
hvm , Lvn i = hLvm , vn i = λm hvm , vn i (6.8)
This can only be true if hvm , vn i = 0 since we assumed that λn 6= λm . Similarly,
it can be shown that the eigenvalues λn have to be real.
It is then also possible to show that while any two eigenfunctions vn and
vm with the same eigenvalue λ are not necessarily orthogonal, it is possible to
find linear combinations of the latter that are orthogonal. This process is called
Gram-Schmidt orthogonalization. The eigenfunctions (whose amplitudes are
always arbitrary) can also be normalized so that hvn , vn i = 1.
Finally, with this set of orthonormal eigenfunctions, it is also possible (but
not very easy) to show that any function f (x) defined on the interval [a, b] and
satisfying the homogeneous boundary conditions f (a) = f (b) = 0 can be written
as X
f (x) = fn vn (x) (6.9)
n
This is often called a generalized Fourier expansion.
What does this buy us? Well, we can now use this information in our original
problem, and let its solution u(x) be
X
u(x) = an vn (x) (6.10)
n

We then have
X X
Lu(x) = an Lvn (x) = an λn vn (x) = F (x) (6.11)
n n

Taking the dot-product of this equation with vm , we then get

X
an λn hvm , vn i = am λm = hF, vm i (6.12)
n
126

As long as none of the λm are zero, this can be solved with

hF, vm i
am = (6.13)
λm
and yields a unique solution for u(x).
However, if one of the λm is zero and the projection of F on its corresponding
eigenmode vm is not (hF, vm i 6= 0), then there is no solution to this equation,
because (6.12) is contradictory. For a solution to exist, we must have hF, vm i = 0
for the eigenmode vm whose eigenvalue is 0.
In our original problem, we are told that there is a non-zero solution to

Luh = 0 (6.14)

so in other words, uh is an eigenmode of L with eigenvalue 0. We are exactly

in the conditions of the theorem, and therefore know that Lu(x) = F (x) only
has a solution if hF, uh i = 0.
These considerations can easily be generalized to operators acting in multiple
spatial dimensions, and operating on more than 1 function. This will come in
very handy later.

A simple example of application of weakly nonlinear theory

This section is adapted from Chapter 5.2 of the textbook “Introduction to hydro-
dynamic stability” by Drazin.

Let’s consider a toy problem that nicely illustrates how one may study weakly
nonlinear stability. Consider the following equation that contains a diffusion
term, and a nonlinear term:
∂u 1 ∂2u
− sin u = (6.15)
∂t R ∂z 2
with boundary condition u = 0 at z = 0 and z = π. This equation has a simple
null solution, u = 0, which forms the background around which we will linearize.
In what follows, we therefore assume u is small. Using the fact that sin u ' u
for small enough u, the linearized equation is
∂u 1 ∂2u
−u= (6.16)
∂t R ∂z 2
Seeking solutions of the kind

u(z, t) = û(z)eλt (6.17)

we find that û satisfies

d2 û
= R(λ − 1)û (6.18)
dz 2
which, given the required boundary conditions, has solutions ûn ∝ sin(nz) with
λn = 1 − n2 /R. For λn to be positive, R must be greater than n2 . Hence the
 127

most unstable mode will be n = 1, and the critical value of R above which this
mode is excited is Rc = 1. For R above Rc (but smaller than 4), the only mode
excited is the n = 1 mode.
Suppose R = 1 + , with  small (but not too small). The n = 1 mode is
expected to grow at a rate
1
λ=1− ' (6.19)
1+
This implies that the instability will develop on a slow timescale T = t, which
leads us to suppose that the weakly nonlinear solution is of the form u(z, T ).
Furthermore, it is likely of small amplitude – the closer to onset we are, the
weaker we expect the saturated amplitude of the perturbation to be. Without
any more information, we can at least suppose that

u(z, T ) = α u1 (z, T ) + 2α u2 (z, T ) + . . . (6.20)

The correct value of α is one of the outcomes of this problem that we actually
wish to find.
Let’s plug this expansion into the governing equation. We get
∂u1 ∂u2 1
α+1 + 2α+1 + . . . − (α u1 + 2α u2 + . . .) + (α u1 + 2α u2 + . . .)3
∂T  ∂T 6
2 2

1 ∂ u1 ∂ u 2
= α + 2α + ... (6.21)
1+ ∂z 2 ∂z 2

To the lowest order, regardless of the value of α, we clearly have

∂ 2 u1
+ u1 = 0 (6.22)
∂z 2
which has the solution
u1 (z, T ) = A1 (T ) sin(z) (6.23)
Note that the sine solution was chosen here to satisfy the boundary conditions.
Also note, in view of using the results of the previous section, that we have
found a non-zero solution to the equation Lu1 = 0, where L = d2 /dz 2 + 1. It’s
trivial to show that this operator is self-adjoint with weight function r(x) = 1.
We now subtract this zeroth-order system, and divide the governing equation
by α . Since we do not know a priori what α is, we have to be careful in which
order of  to keep. For instance, we do not know if α ≥ 1 or α ≤ 1, or, for that
matter, whether 2α ≥ 1 or 2α ≤ 1, and so forth. However, we do know that
2α > α, and that α + 1 > α. For this reason, as a first pass at the problem, we
keep all orders up to α + 1 and 2α. We are then left with
∂u1 ∂u2 1
 + α+1 − α u2 − 2α u3 + 2α u31
∂T ∂T 6
∂ 2 u1 ∂ 2 u2 2
α+1 ∂ u2
2
2α ∂ u3
= − 2 + α −  +  (6.24)
∂z ∂z 2 ∂z 2 ∂z 2
128

Let’s begin by being very naive, and assume α = 1. Then we have, to the
lowest order,
∂u1 ∂ 2 u1 ∂ 2 u2
− u2 = − 2 + (6.25)
∂T ∂z ∂z 2
which implies that

∂ 2 u2
 
∂u1 ∂A1
+ u2 = − u 1 = − A 1 sin z (6.26)
∂z 2 ∂T ∂T

This equation for u2 is no longer homogeneous, but still satisfies the same ho-
mogeneous boundary conditions as u1 did. By Fredholm’s alternative, because
there is a non-zero solution u1 to the homogeneous problem, we know that (6.26)
only has solutions if the solvability condition
 
∂A1
h − A1 sin z, u1 (z)i = 0 (6.27)
∂T

is satisfied. Given that u1 (z) ∝ sin(z), and that hsin z, sin zi 6= 0, this is equiv-
alent to requiring
∂A1
− A1 = 0 (6.28)
∂T
which means that A1 grows exponentially without saturating! This effectively
recovers linear theory – it is not incorrect, but it is also not very helpful. It
merely tells us that as long as the amplitude of the perturbation is smaller or of
the other of , then it continues to grow exponentially. We were clearly a little
bit too naive in our choice for α.
How can we account for nonlinear saturation? The key is to go back to
equation (6.24), and more carefully inspect the various terms that arise. Since
we want the nonlinear terms to come in at the same order as the ∂u1 /∂T term,
we have to have 2α =  – it’s that simple. This means that α = 1/2. Going
back to equation (6.24) and keeping only the lowest-order terms in , we then
get
∂ 2 u2
1/2 = −1/2 u2 (6.29)
∂z 2
which implies, as before, that u2 (z, T ) = A2 (T ) sin(z). So far, that is not very
useful. However, to the next order (that is, to order ), we get

∂u1 1 ∂ 2 u1 ∂ 2 u3
− u3 + u31 = − 2 + (6.30)
∂T 6 ∂z ∂z 2
which implies

∂ 2 u3 ∂A1 1
+ u3 = sin z + A31 sin3 z − A1 sin z
∂z 2 ∂T 6 
∂A1 1 3 1
= + A1 − A1 sin z − A31 sin(3z) (6.31)
∂T 8 24
 129

using the identity sin3 (z) = (3/4) sin(z) − (1/4) sin(3z). By Fredholm’s alterna-
tive, we know that this equation only has solutions if the dot product of sin z
with its RHS is 0. While the sin(3z) term is indeed orthogonal to sin(z), the
sin(z) term is not. We therefore need to have, as a solvability condition, that

∂A1 A3
= A1 − 1 (6.32)
∂T 8
This equation, by contrast with the one we obtained for α = 1, contains non-
linear terms that act to saturate the linear instability. In fact, we find that this
ODE has the general solution
√
2 2eT
A1 (T ) = ± √ (6.33)
K + e2T
where K is an integration constant that depends on the initial conditions. Since
u(z) = 1/2 A1 (T ) sin(z), with  = R − 1, then an approximation solution to
(6.15) for R close to one is given by
p
2 2(R − 1)eT
u(z, T ) = ± √ sin(z) (6.34)
K + e2T
As expected, pfor small T , u grows exponentially with growth rate 1, but as
T → ∞, u → 2 2(R − 1). Note that the limit can be obtained without actually
solving the equation for A1 , simply by requiring that the system be in a steady
3
state. We find that,
√ at steady state, A1 = 8A1 , which either means that A1 = 0
or that A1 = ±2 2.
Hence three steady states exist. The two non-zero steady states are stable
for R > 1, and have an amplitude that scales like the square root of the distance
to the bifurcation point, while the null steady state is stable for R < 1. This is
called a Pitchfork bifurcation.
A very important result of this analysis is that, by contrast with the example
given in Section 6.1.3, we found that the nonlinear terms involved in the satura-
tion of the linear instability were cubic, instead of quadratic. This, in fact, could
have been predicted simply by considerations of symmetry!. Indeed, inspection
of (6.15) shows that it is symmetric in z: if u(z) is a solution of the equation, then
u(−z) is one too. Since u1 (z, T ) = A1 (T ) sin z, then u1 (−z, T ) = −A1 (T ) sin z
must also be a solution. This then implies that in the amplitude equation for
A1 , if A1 is a solution, then −A1 must also be a solution. In other words, the
amplitude equation must be invariant if A1 → −A1 . This would not have been
satisfied had there been a quadratic term – in fact, had there been any even
powers of A1 . Only odd powers of A1 are allowed, simply because of the under-
lying symmetry z → −z! These considerations are very powerful, and enable
us to guess what the form of the amplitude equation of complicated systems
may be without actually performing any asymptotic expansion. The only thing
missing, in that case, is the value of the coefficients of each term (e.g. in this
case, knowing that there must be a cubic term does not tell us its sign, nor the
130

fact that the coefficient in front is 8). But knowing the power of the respective
terms in the amplitude equation can already be enough to tell us what kind of
bifurcation (pitchfork, saddle-node, transcritical, etc...) we might expect.

6.3.4 Weakly nonlinear theory for steady-state Rayleigh-

Benard convection
Armed with a grand total of one previous attempt at deriving weakly nonlinear
equations, we now attack the much harder problem of Rayleigh-Bénard convec-
tion.
Recall that the behavior of Rayleigh-Bénard convection for Ra just above
onset is a rapid transition to another steady state that has 2D convective rolls. In
what follows, we will therefore look for steady-state weakly nonlinear solutions
of the problem. Despite being a rather different problem to the one we studied
above, the method used is fairly similar: define a small parameter that is the
distance to onset, expand all quantities in powers of that small parameter, and
find solutions order-by-order, using the solvability condition.
Let’s begin by re-writing the 2D governing equations in terms of the stream-
function φ and the vorticity ω = ωey . First, recall that

ω = ∇ × u = ∇ × ∇ × (φey ) = −∇2 φey (6.35)

We can therefore define a scalar vorticity ω to be the component of vorticity in

the y-direction. It is related to the streamfunction as

ω = −∇2 φ (6.36)

By taking the curl of the momentum equation, and extracting the y com-
ponent of the resulting vorticity equation (which is the only component left in
2D), we get
∂ω ∂T
+ u · ∇ω = RaPr + Pr∇2 ω (6.37)
∂t ∂x
Using the stream function, the nonlinear term can be re-written as

∂φ ∂ω ∂φ ∂ω
u · ∇ω = − = J(ω, φ) (6.38)
∂z ∂x ∂x ∂z
where J is the Jacobian operator. Similarly, the temperature equation can be
rewritten as
∂T ∂φ
+ J(T, φ) + = ∇2 T (6.39)
∂t ∂x
We now consider an expansion of these equation around marginal stability
(that is, around Rac ). To do so, we let

Ra = (1 + )Rac (6.40)
 131

which defines . Here we assume that  > 0, since we have proved that there are
no non-zero steady states for Ra < Rac . We then expand φ, ω and T as powers
of α , with α to be determined:
φ(x, z) = α φ1 (x, z) + 2α φ2 (x, z) + . . .
ω(x, z) = α ω1 (x, z) + 2α ω2 (x, z) + . . .
T (x, z) = α T1 (x, z) + 2α T2 (x, z) + . . . (6.41)
Note that it is not obvious a priori that the expansion should be like this: it
could be that a different α may be necessary for each variable, or that the
starting power may be different, etc... However, we will only know whether
we are on the right track or not by plugging these expansions in the governing
equations and finding out whether they give the right behavior.
We can already see that, indeed, the expansion has to be similar for ω and
φ. We get, order by order,
ωi = −∇2 φi (6.42)
For the other equations, at the lowest orders in α , we get
 
2α α ∂T1 2α ∂T2
 J(ω1 , φ1 ) + . . . = Rac Pr(1 + )  + + Pr∇2 (α ω1 + 2α ω2 ) + . . .
∂x ∂x
∂φ1 ∂φ2
2α J(T1 , φ1 ) + α + 2α . . . = ∇2 (α T1 + 2α T2 ) + . . . (6.43)
∂x ∂x
Dividing both equations by α , we get
 
α ∂T1 α ∂T2
 J(ω1 , φ1 ) + . . . = Rac Pr(1 + ) + + Pr∇2 (ω1 + α ω2 ) + . . .
∂x ∂x
∂φ1 ∂φ2
α J(T1 , φ1 ) + + α . . . = ∇2 (T1 + α T2 ) + . . . (6.44)
∂x ∂x
We recover the linear equations at the zeroth order in :
∂T1
Rac Pr − Pr∇4 φ1 = 0
∂x
∂φ1
− + ∇2 T1 = 0 (6.45)
∂x
using ω1 = −∇2 φ1 . Of course, there is a trivial solution to these equations
(T1 = φ1 = 0). But we actually already know another non-trivial solution as
well: at exactly Ra = Rac , Rayleigh-Bénard convection is just becoming linearly
unstable, so there is one non-trivial mode with zero growth rate. We found it
in Section 6.2.3, and it is:
φ1 (x, z) = φ̂ sin(πz) cos(kc x) (6.46)
where kc = π 2 /2, and where we have picked, arbitrarily, the cosine mode in the
x direction. That mode is a solution to (6.45), with ω1 and T1 given by
ω1 (x, z) = φ̂(π 2 + kc2 ) sin(πz) cos(kc x)
kc sin(πz) sin(kc x)
T1 (x, z) = φ̂ (6.47)
π 2 + kc2
132

The only remaining unknown is the mode amplitude φ̂. Unfortunately, we can’t
get it from linear equations – this is also going to be one of the outcomes of the
full weakly nonlinear problem.
At the next order, we would like to see the nonlinearities appear. This
should, as in the toy problem studied earlier, give us an idea of what α should
be. Here, a little bit of forward thinking goes a very long way. Note how the first
order at which the nonlinearities seem to appear in the LHS of the momentum
equation is α J(ω1 , φ1 ), which we would very much like to be of the same order
as the Rac Pr∂T1 /∂x term in the RHS. This would require α = 1. However,
doing that would be a mistake. Indeed,
J(ω1 , φ1 ) = −J(∇2 φ1 , φ1 ) = −(π 2 + kc2 )J(φ1 , φ1 ) = 0 (6.48)
and so the nonlinear terms only appear at the next order, and look like
2α J(ω2 , φ1 ) + 2α J(ω1 , φ2 ) (6.49)
Of course, it could be that saturation is caused by the nonlinear terms in the
thermal energy equation (which is non-zero). After subtracting the lowest-order
terms in equation (6.44) (which we have already equated to one another), and
keeping only terms of order α this time, we get
∂φ2
− + ∇2 T2 = J(T1 , φ1 ) (6.50)
∂x
However, this doesn’t tell us what α should be – and that is a strong hint that
there is a serious problem in following this path (See the Appendix to find out
what would have happened with α = 1).
Going back to the momentum equation, if we actually want the nonlinear
terms to balance Rac Pr∂T1 /∂x, we are then forced to choose α such that
2α = 1, in other words, α = 1/2. Using this in (6.44), and keeping only the
leading order terms, we get to order 1/2 ,
∂T2
−Pr∇4 φ2 + = 0
∂x
∂φ2
− + ∇2 T2 = J(T1 , φ1 ) (6.51)
∂x
where (6.42) was used to eliminate ω2 .
At this point, we are getting close to a problem we have already seen: that
of a forced linear equation, whose homogeneous counterpart has a non-trivial
solution. We would therefore like to use Fredholm’s alternative to get a solv-
ability condition (which would then give us a condition on φ1 and hence on φ̂).
In order to do that, we have to prove that the linear problem is self-adjoint. As
it turns out, this is not too difficult. Consider first that we now have a multidi-
mensional operator L, acting on the 2D vector of 2 functions φ = (φ, T ). The
linear problem can be recast as
∂
−Pr∇4 Rac Pr ∂x
   
φ
Lφ = ∂ = (6.52)
− ∂x ∇2 T
 133

We have already shown, for instance that

Lφ1 = 0 (6.53)

To show that L is self-adjoint, we to find an inner product for which hφ1 , Lφ2 i =
hLφ1 , φ2 i for any two functions-vectors φ1 and φ2 (i.e. not necessarily the ones
associated with the expansion in ). Trial and error shows that it is given by
Z
hφ1 , φ2 i = (φ1 φ2 + Rac PrT1 T2 )dxdz (6.54)

where the integral is taken over the domain D spanning z in [0, 1], and assuming
periodicity in x. Indeed, with successive integrations by parts (making sure that
all of the boundary terms vanish), we have

hφ1 , Lφ2 i
Z     
4 ∂T2 ∂φ2 2
= φ1 −Pr∇ φ2 + Rac Pr + Rac PrT1 − + ∇ T2 dxdz
∂x ∂x
Z  
2 2 ∂φ1 ∂T1
= −Pr∇ φ1 ∇ φ2 − Rac PrT2 + Rac Prφ2 − Rac Pr∇T1 · ∇T2 dxdz
∂x ∂x
Z     
4 ∂T1 ∂φ1 2
= φ2 −Pr∇ φ1 + Rac Pr + T2 −Rac Pr + Rac Pr∇ T1 dxdz
∂x ∂x
= hφ2 , Lφ1 i (6.55)

Given that L is self-adjoint, we now know that, in order for (6.51) to have a
solution, it is necessary that the inner product of φ1 (the non-zero solution of
the homogeneous problem) with its RHS be zero:
   
φ1 0
< , >= 0 (6.56)
T1 J(T1 , φ1 )

This implies Z
Rac Pr T1 (x, z)J(T1 , φ1 )dxdz = 0 (6.57)

Since
kc2 π sin(2πz)
J(T1 , φ1 ) = φ̂2 (6.58)
π 2 + kc2 2
that is, only depends on z, while T1 ∝ sin(kc x), then indeed, (6.57) is true. We
can therefore expect that there is a solution to this equation. A quick inspection
reveals that, because the RHS of (6.51) is independent of x, the solution is also
independent of x. Indeed, let’s try T2 = T2 (z) and φ2 = φ2 (z). We then have
the conditions
d4 φ2
−Pr = 0
dz 4
d2 T2 kc2 π sin(2πz)
= φ̂2 (6.59)
dz 2 π2+ kc2 2
134

The solution to the first, subject to homogeneous boundary conditions, is simply

φ2 (z) = 0 (so ω2 = 0 too). The solution to the second yields

kc2 sin(2πz)
T2 (z) = −φ̂2 (6.60)
π2 + kc2 8π

We can now attack the next order in . Unfortunately, because φ2 = ω2 = 0,

the nonlinear terms in the momentum equation still do not appear. However,
by some stroke of luck, this does not matter, as we shall see. Let’s proceed. We
get,
 
∂T1 ∂T3
0 = Rac Pr + − Pr∇4 φ3
∂x ∂x
∂φ3
J(T2 , φ1 ) + = ∇2 T3 (6.61)
∂x
which can be rewritten as

−Rac Pr ∂T
 
1
Lφ3 = ∂x ≡ N3 (6.62)
J(T2 , φ1 )

As usual, this only has solutions if

hφ1 , N3 i = 0 (6.63)

Given that
∂T2 ∂φ1 k 3 cos(2πz)
J(T2 , φ1 ) = − = −φ̂3 2 c 2 sin(πz) sin(kc x) (6.64)
∂z ∂x π + kc 4

then

hφ1 , N3 i
k 3 cos(2πz)
Z     
∂T1
= −Rac Pr φ1 + T1 φ̂3 2 c 2 sin(πz) sin(kc x) dxdz
∂x π + kc 4
2
k 2 cos(2πz)
Z  
k
= −φ̂2 Rac Pr 2 c 2 sin2 (πz) cos2 (kc x) + φ̂2 2 c 2 sin2 (πz) sin2 (kc x) dxdz
π + kc π + kc 4
2 Z 1 2

2 Rac Pr kc π 2 2 kc cos(2πz) 2
= −φ̂ sin (πz) + φ̂ 2 sin (πz) dz
2 π 2 + kc2 kc 0 π + kc2 4
" #
Rac Pr kc2 φ̂2 kc2
= −φ̂2 2 2
1− (6.65)
4 π + kc 8 π 2 + kc2

For this quantity to be zero, we have to have

π 2 + kc2
φ̂2 = 8 (6.66)
kc2
 135

This finally tells us what the amplitude of the convective roll is! The weakly
nonlinear solution can then be written as:
s
√ 1/2 π 2 + kc2
φ(x, z) = ±2 2 sin(πz) cos(kc x) + O(3/2 )
kc2
√
1/2 2 2 sin(2πz)
T (x, z) = ± p sin(πz) sin(kc x) −  + O(3/2 ) (6.67)
2
π + kc2 π

where we recall that  was defined such that

Ra = (1 + )Rac (6.68)

As before, we see that the weakly nonlinear solution scales like the square
root of the distance to the bifurcation parameter Rac , with two symmetric
solutions. This reveals the onset of convection to be associated with a pitchfork
bifurcation again, as in the previous example. And as in the previous example,
that could have been guessed simply based on the symmetries of the problem!
The steps we went through to get to the weakly nonlinear solution, however,
tell us a lot about the physics behind the nonlinear saturation of convection,
at least close to onset. We see that the nonlinear terms in the momentum
equation never came in. Instead saturation proceeds by the modification of the
background state by the convective rolls. Indeed, the rolls, to order one, induced
an order 2 change in the mean temperature profile. We see that instead of a
linear profile, the nonlinear mean profile at saturation is given by
Z
sin(2πz)
(T̄ (z) + T̃ (x, z))dx = −z −  (6.69)
π
This new profile has a more uniform temperature near the mid-point of the
domain, and sharper gradients near the boundary (see Figure).

Appendix
Suppose we had proceeded anyway, and required that α be equal to 1. The
leading order equations then read
∂T2 ∂T1
−Pr∇4 φ2 + Rac Pr = −Rac Pr
∂x ∂x
∂φ2
− + ∇2 T2 = J(T1 , φ1 ) (6.70)
∂x
The associated solvability condition is
Z    
∂T1
φ1 −Rac Pr + Rac PrT1 J(T1 , φ1 ) dxdz = 0 (6.71)
∂x
It can be shown with very little algebra that the only way of satisfying it is with
φ̂ = 0 (and therefore φ1 , T1 and ω1 are also 0). Arriving at this kind of result
definitely means that our selection of expansion was incorrect.