Math 654

Introduction to Mathematical Fluid

Dynamics

Professor Charlie Doering

Transcription by Ian Tobasco

University of Michigan, Winter 2011

Math 654, Lecture 1 1/5/11 1

Lecture 1: Vectors, Tensors, and Operators

1 Vectors: Notation and Operations
Given a vector x R3 , we can write it with respect to the canonical basis {i, j, k}
as x = xi + yj + z k.
In
this manner, we can define vector fields as
v(x, y, z) = u(x, y, z)i + v(x, y, z)j + w(x, y, z)k.

Note that sometimes the canonical basis is written as { 2 , e

e1 , e 3 }, and similarly
x 1 + x2 e
= x1 e 2 + x3 e
3 ,
v(x1 , x2 , x3 ) = v1 (x1 , x2 , x3 )
e1 + v2 (x1 , x2 , x3 )
e2 + v3 (x1 , x2 , x3 )
e3 .
In this way we easily generalize to Rd , with a vector field being
d
X
v(x1 , . . . , xd ) = vi (x1 , . . . , xd )
ei
i=1

or just
v(x) = vi (x)
ei
using Einstein notation. 1
The function vi is commonly referred to as the ith component of the vector
field.
We have the following operations on pairs of vectors.
Definition 1. The dot product (or inner product) of v, w Rd is defined as
v w = vi wi .

We can arrive at this with the following formalism. First, define the dot product on the canonical basis as
(
1 i=j
i e
e j = ij = .
0 i 6= j
i and w = wj e
Then write v = vi e j , and define
v w = vi wj ( j ).
ei e
Carrying out the implied summation yields the earlier definition.
Definition 2. The outer product of v, w Rd is a linear self-mapping of Rd defined via
i e
vw = vi wj e j ,
j is the linear self-mapping of Rd having as matrix representation in the canonical bases
i e
where e
( j )mn = im jn .
ei e

So, the outer product of v, w Rd produces the linear map with the canonical matrix representation
(vw)mn = vm wn ,
a so-called dyadic tensor. This brings us to the next topic.
1A handy notation where repeated indices imply summation.
Math 654, Lecture 1 1/5/11 2

2 Tensors
Definition 3. A 2-tensor on Rd is a bilinear form on Rd . Specifically, T : Rd Rd R is a 2-tensor if it
satisfies

1. Component-wise additivity:

T (v + v0 , w) = T (v, w) + T (v0 , w)
T (v, w + w0 ) = T (v, w) + T (v, w0 ),

2. Component-wise homogeneity:

T (v, w) = T (v, w)
T (v, w) = T (v, w),

given , R.
Proposition 1. The set of 2-tensors on Rd is isomorphic to the set of linear self-maps of Rd .

In other words, 2-tensors are matrices; we pursue this idea throughout the rest of this section. First, just
as the set of linear self-maps of Rd forms a linear space, the set of 2-tensors on Rd forms a linear space.
Moreover, one can easily show that { j }1i,jd is a basis. Given a 2-tensor T , we use this basis to write
ei e

i e
T = Tij e j ,

so T has the canonical matrix representation

(T )ij = Tij .

We have a way to write the operation of T on v Rd using tensor notation:

T v = i e
(Tij e j ) (vk e
k )
i e
= Tij vk e j e
k
= Tij vk jk
i .
= (Tij vj ) e

The key thing to notice is that e i e

j e
k = jk , which follows from the definition of e
i e
j . Composition of
2-tensors is just composition of linear maps; in tensor notation,

T T0 = 0
Tij Tkl ( j ) (
ei e l )
ek e
0
= Tij Tkl i e
jk e l
= Tij Tjl0 e
i e
l .

Again, the key step here is ( j ) (

ei e l ) = jk e
ek e l . Notice that the canonical matrix representation of T T 0
i e
is the matrix product of the representations of T and T 0 , a sensible proposition.
Square matrices can be symmetric or anti-symmetric, and so can 2-tensors:
Math 654, Lecture 1 1/5/11 3

Definition 4. T , a 2-tensor on Rd , is said to be symmetric if Tij = Tji , and is said to be anti-symmetric if

Tij = Tji .

Generally, a 2-tensor can be written as a sum of its symmetric and anti-symmetric parts. Indeed,
Tij = Sij + Aij
where
1
Sij = (Tij + Tji ) ,
2
1
Aij = (Tij Tji ) .
2
Just like matrices, we have a trace operation.
Definition 5. Given a 2-tensor T on Rd , its trace is
X
Tr (T ) = Tii = Tij ij .

And we may contract tensors as well.

Definition 6. Let T 0 be another 2-tensors on Rd , then the tensor contraction of T and T 0 is
T : T 0 = Tij Tij0 .

Note that just as with the inner product, we can also define tensor contraction formally by writing
T : T0 0
= Tij Tkl ( k ) (
ei e l )
ej e
0
= Tij Tkl ik jl
and summing up over all indices.
Remark. Weve seen quite a few definitions regarding 2-tensors, motivated by linear algebra. We can gener-
alize the idea of a tensor by considering mappings T : Rd Rd R which are multilinear (i.e., linear
| {z }
k products
in each component). These so-called k-tensors do not have analogs in linear algebra, but are sometimes
visualized as multi-index arrays.
Example. (Cross-product) We all know that the cross-product of two vectors v, w R3 can be written as
a formal determinant
1 e
e 2 e3
v w = det v1 v2 v3 .
w1 w2 w3
By defining a tensor known as the Levi-Civita symbol,


1 {i, j, k} a cyclic perm. of {1, 2, 3}
ijk = 1 {i, j, k} an anti-cyclic perm. of {1, 2, 3} ,


0 otherwise
we can write the cross-product in a compact way:
(v w)i = ijk vj wk .
Math 654, Lecture 1 1/5/11 4

3 Derivative Operators
Given coordinates x1 , . . . , xd in Rd we know how to take derivatives. It will be convenient to make the
following notation. Let f : Rd R be differentiable, and write
f
= xi f = i f = f,i
xi
to mean the partial derivative of f w.r.t. the ith coordinate. In such a way, we have Leibnizs rule:

(f g),i = f,i g + f g,i .

Definition 7. The gradient operator is the differential operator

i i () ,
grad () = e

and is usually written as .

Remark. In Cartesian coordinates,
i i = i e
=e i ,
but watch out in other coordinate systems! For example, this does not hold in polar coordinates. For the
rest of this lecture, well work in Cartesian coordinates, unless otherwise specified.

For differentiable f : Rd R, we have

i ,
f = (i f ) e
and for differentiable v(x),

v = (
ei i ) (
ej vj )
vj
= i e
e j
xi
i e
= vj,i e j .

Composing the gradient operator is allowed, e.g.,

2f
i e
f = e j i e
=e j f,ij .
xi xj
And we can even take the gradient of a tensor, e.g.,

(T )ijk = Tjk,i

for some differentiable 2-tensor T .

Definition 8. The divergence operator is the differential operator on vector fields defined by

div v = Tr (v) = vi,i ,

and is usually written as v. Similarly, we have the curl operator which is defined by

curl v = ijk j vk ,

and is usually written as v.

Math 654, Lecture 1 1/5/11 5

With these definitions, we can begin to develop some useful identities.

Proposition 2. Given a vector field v,

v = ( v) 4v.

Remark. This is used to define the Laplacian in general coordinates for vector fields a non-trivial job in
arbitrary coordinate systems because derivatives of unit vectors must be considered.

Proof. On the left, we have

ijk j (klm l vm ) = ijk klm vm,jl .
Its easy to show that
ijk klm = il jm im jl ,
so the left hand side becomes

(il jm im jl ) vm,jl = il vj,jl vi,jj

= vj,ji vi,jj
= i (vi,i ) 4vi

which is exactly the ith component of the right hand side.

Definition 9. The vorticity of a vector field is exactly its curl:

= v.

In components,
i = ijk vk,j .

Now consider a 2-tensor with components

ij = ijk k
= ijk klm vm,l
= (il jm im jl ) vm,l
= vj,i vi,j
= 2 (anti-symmetric part of vj,i ) .

This tensor has canonical matrix representation

0 3 2
= 3 0 1 .
2 1 0

Proposition 3. Given a vector field v,

(v v) = v v + ( v) .
Math 654, Lecture 1 1/5/11 6

Proof. The ith component on the left hand side is

ijk j (vl l vk ) = ijk j (vk,l vl )

= ijk vk,lj vl + ijk vk,l vl,j .

The first term in the sum becomes

ijk vk,lj vl = vl l (ijk vk,l )

= (v )i .

Before we deal with the second term, we note that the following hold:

ijk vl,k vl,j = 0,

vk,l vl,j vl,k vl,j = lkm m vl,j .

So, the second term becomes

ijk vk,l vl,j = ijk (vk,l vl,j vl,k vl,j )

= ijk lkm m vl,j
= (im jl il im ) m vl,j
= i vj,j j vi,j
= ( v + ( v) )i .

Hence, the result.

Math 654, Lecture 2 1/10/11 1

Lecture 2: Convective Derivatives and Conservation Equations

What do we mean by a fluid? A state of matter that cannot support shear stresses, as opposed to a solid.
Well primarily consider ideal fluids (non-viscous, etc.).
Our basic setup is in R3 . Consider a subset R3 which contains fluid matter.

u

V
x

Figure 0.1: A subset of R3 containing fluid.

We can consider several ideas:

Density, (x, t)
Mass (infinitesimal), m = (x, t)V

Velocity field u(x, t) = iu + jv + kw

=e
i ui = (ux , uy , uz )
Pressure field, p(x, t).
and
Definition 1. Consider the force into a fluid element across the face with outward pointing normal n
area a to be
npa,
which implicitly defines the scalar quantity p(x, t). This quantity is called the pressure.

n
a

npa

Figure 0.2: Force on an infinitesimal area a.

The dynamics of these variables (all presumed to be smooth functions) are related by
Math 654, Lecture 2 1/10/11 2

Conservation of mass
Newtons 2nd Law
Thermodynamics.

1 The Convective Derivative

This is a fundamental kinematic principle. Consider f (x, t). There are two ways of computing the rate of
change of this quantity with respect to time. First, we could fix position and calculate the time derivative:
 
df f f (x, t + t) f (x, t)
= = lim .
dt fixed position t t0 t

Or, we may consider calculating the rate of change of f while moving along with fluid elements:
 
df f (x + u(x, t)t) f (x, t)
= lim ,
dt moving t0 t

to first order (which suffices in the limit). Moving with the flow means

X(t) = u (X(t), t) ,

where X(t) is the path we take through the fluid. In this manner, we find
  
df d 
= f (X(t), t)
dt moving dt X(t)=X
f i f (x, t)
= (x, t) + X
t xi
f
= + u.f.
t
Df
Well denote the convective derivative by Dt .

Proposition. The convective derivative is a differential operator, i.e., it satisfies the product rule, the chain
rule, etc.

2 Conservation Equations
2.1 Conservation of Mass

Consider a fluid element, having mass m = V . As the fluid flows, the location, volume, and shape of this
element may change. Conservation of mass means
Dm
0 =
Dt
D DV
= V +
Dt
 Dt

DV
= + u. V + .
t Dt
Math 654, Lecture 2 1/10/11 3

Suppose V = xyz, then  

DV Dx
= yz + . . . .
Dt Dt

z
y x x
x x 2
x x+ 2

Figure 2.1: Infinitesimal volume element.

The rate of change of this line segment depends on the difference in fluid velocity at its endpoints. So,
Dx
= u1 (x + x/2, y, z, t) u1 (x x/2, y, z, t)
Dt
x u1 x u1
= u1 (x, y, z, t) + (x, y, z, t) + u1 (x, y, z, t) + (x, y, z, t) + . . .
2 x 2 x
u1
= x
x
and similarly for y, z. Hence,
DV
= (.u) V,
Dt
and so conservation of mass becomes
 

0= + u. V + (.u) V.
t
We conclude that

0= + u. + (.u) ,
t
and after applying a vector identity,

0= + .(u).
t
This is called the continuity equation.
Remark. The mass flux of a flow field is given by u.

n
a

mass
(u) .
na = time flowing through a

Figure 2.2: Mass flux through an infinitesimal area a.

Remark. The continuity equation is linear in u, so it would seem thus far that things are straightforward.
Math 654, Lecture 2 1/10/11 4

2.2 Conservation of Momentum (Newtons 2nd Law)

d
Recall that F = dt p. Applied to a fluid element,
   
D D D
(mu) = m u + m u
Dt Dt Dt
 
D
= m u
Dt
due to conservation of mass. Continuing,
  
D
(mu) = m + u. iu + jv + kw

Dt t
= m (t u + u.u)
= F

by Newtons 2nd law.

Definition 2. An ideal fluid is one in which there are no shear stresses.

Note that this definition permits normal forces, which we account for in the pressure field.

top = k
n

y
x

bot = k
n

Figure 2.3: Normal forces on a fluid element.

The net force in the z-direction is

3
kF = ntop ptop xy n bot pbot xy
=
kxy(p top + pbot ).

Taylor expanding, we find

ptop + pbot = p(x, y, z + z/2, t) + p(x, y, z z/2, t)

z p z p
= p(x, y, z, t) (x, y, z, t) + p(x, y, z, t) (x, y, z, t)
2 z 2 z
p
= z.
z
Math 654, Lecture 2 1/10/11 5

Hence,  
3 = k

kF p V.
z
Similarly, we have
 
iF1
= i p V,
x
 
jF2
= j p V.
y

So,
m (t u + u.u) = F = V p.
Dividing through by m, we find
1
t u + u.u = p.

The is the momentum equation for ideal fluid motion.

2.3 Thermodynamic Closure

We have

t + .(u) = 0,
1
t u + u.u = p.

But since there are d + 2 variables (, u1 , . . . , ud , p), we need a relation between p and . For example,
we could take p = (kB T ) for an isothermal gas. There are closures for polytropic gases: p = Kn/(n+1) ,
for different values of n. Once we select a thermodynamic closure, we obtain a system of three dynamical
equations for fluid motion. These are known as the Euler equations for an ideal fluid.

2.4 The Incompressible Limit

There are simplifications to be made. One is to consider so-called incompressible fluids (e.g., water, or air
for low Mach number.) Suppose = 0 + r(x, t) with 0 constant, and further suppose p = p(). Our plan
is to linearize the momentum equation for ideal fluids (i.e., take kuk to be small, so neglect terms which
are order kukr (??), and take |r| to be small, i.e., |r| /0  1.) Linearizing, we find

t r + 0 .u = 0,
1
t u + p() = 0.

The second equation becomes
1
0 = t u + p0 ()r
0 + r
1
= t u + p0 (0 )r,
0
Math 654, Lecture 2 1/10/11 6

so we arrive at the system

t r + 0 .u = 0,
1
t u + p0 (0 )r = 0.
0

Taking the time derivative of the first equation and operating with 0 .() on the second yields

2
r + 0 .t u = 0,
t2
0 .t u + p0 (0 )4r = 0.

Subtracting, we arrive at
t2 r p0 (0 )4r = 0,
which is a wave equation for acoustic waves. If we call c2 = p0 (0 ), then the wave speed is
s
p
c= .

For a very compressible fluid, /p is large. For a nearly incompressible fluid, /p is small. In the
limit /p 0, we reach incompressibility. Note that the sign of /p is important (otherwise wed have
complex wave speeds, a rather unfortunate situation).
Definition 3. An incompressible fluid is one in which /p = 0 or equivalently in which c = .

For an incompressible, ideal fluid, we have the incompressible Euler equations:

= constant > 0,
.u = 0,
1
t u + u.u + p = 0.

Now, there are d + 1 variables and d + 1 equations.
Math 654, Lecture 3 1/12/11 1

Lecture 3: Boundary Conditions and Helmholtz-Hodge

Last time, we discussed ideal fluids. We applied conservation of mass and momentum, and discussed ther-
modynamic closure. For low Mach number, we arrived at the incompressible Euler equations:

1 1
t u + u.u + p = f (x, t),

= constant,
.u = 0.

Here, f is an externally applied body force (force/volume). Sometimes, we set = 1 for simplicity, which is
just a rescaling of mass units. We have d equations and d unknowns. Note that p is determined implicitly;
well see how this works later.

1 Boundary Conditions

1.1 Velocity

Suppose the fluid lies in a domain , with rigid boundary which is at rest.

(at rest)
fluid

Figure 1.1: Rigid boundary.

For inviscid solutions, no-slip does not apply. However, suppose n is an outward pointing normal on .
Then, the boundary condition is simply that no fluid flows across the boundary, i.e.,

. u| = 0.
n

Or suppose the fluid lies in a infinite domain . Then, we often take the far-field to be at rest, i.e.,

kuk 0 as kxk .

Or we could specify u at infinity.

A mathematically convenient scenario is to consider periodic boundary conditions.

Math 654, Lecture 3 1/12/11 2

Figure 1.2: Periodic boundary conditions.

This corresponds to considering periodic solutions in an infinite domain,

u(x + iL, t) = u(x, t),

u(x + jL, t) = u(x, t).

Topologically, this means we consider flows on a two- or three-torus.

On the other hand, we can consider physically motivated boundary conditions, such as pipes or channels.
On the top or bottom of the channel, wed have rigid boundary conditions; if the channel is long enough,
then we can consider solutions which are periodic along the length of the channel.

y
j.u = 0

fluid
x
j.u = 0

Figure 1.3: Fluid in a periodic channel.

1.2 Pressure

How do we evolve the pressure of the fluid? There is no explicit evolution equation in our setup. But take
the divergence of the momentum equation to get

1 1
t (.u) + . (u.u) + .p = .f .

Enforcing the divergence-free condition, we arrive at

4p = . (u.u) .f .

This is a Poisson equation for the pressure. Now we need boundary conditions for p.
Math 654, Lecture 3 1/12/11 3

For a periodic domain, we also have

p(x + iL, t) = p(x, t),

p(x + jL, t) = p(x, t).

This is convenient, as the Poisson equation is solvable by separation of variables when the boundary conditions
are periodic. On the other hand, if velocity is specified at , then we also must specify p at . What about
rigid boundaries?
Suppose we have a domain with rigid boundary which is at rest. Dot the momentum equation with
the outward pointing normal on to get

1 1
t ( n+ n
n.u) + (u.u) . .p = n .f .

Evaluating at the boundary, we see that

. p| = (u.u) .
n .f .
n+n

This is an inhomogeneous Neumann boundary condition. This is also good for solving the elliptic problem
posed above for the pressure.

= k.
There are simplifications. Suppose f = 0 and suppose is flat, e.g., n

fluid

= k
n

Figure 1.4: Flat boundary.

Then,
(u.u) .
n = u.u3 = 0.

So for no body forces and flat boundary,

. p| = 0,
n

a homogeneous Neumann boundary condition.

If the boundary is curved, this is no longer true. Consider, though, that at the boundary, u.
n = 0. Hence,

0 = u. (u.
n)
= (u.u) .
n + u. (
n) .u.

So even if the boundary is curved, so long as f is zero, we arrive at

. p| = (u.
n n) .u.
Math 654, Lecture 3 1/12/11 4

n

n

Figure 1.5: Boundary with curvature.

This last condition encodes the curvature of the boundary.

1.3 Initial Value Problem

In summary, we have
1 1
t u + u.u + p = f (x, t),

= constant,
.u = 0,
along with the implicit equation for the pressure,
4p = . (u.u) .f .
As the boundary conditions we have
u.
n = 0 (or periodic b.c.),
. p|
n = (u.u) .
n+n .f .
So the initial value problem is the following:
Given

domain, b.c.s, f
initial conditions, u(x, 0)

Ask: is there a solution to the evolution equations for t > 0? If so, does the smoothness of the solution
depend in a regular way on the smoothness of the initial conditions?

2 Another View of the Euler Equations

If there was an applied F = f V , we would like to say that m a = F, i.e.,
d 1
u f = 0.
dt
Taking the divergence, we get
1
(.u) = . (u.u) + .f .
t
This is a problem! Consider that the right hand side does not have to be zero, so the divergence-free condition
on u can fail in our naive approach. To fix this, we want to take only the component of the acceleration that
preserves .u = 0.
Math 654, Lecture 3 1/12/11 5

Theorem. (Helmholtz-Hodge Decomposition) Suppose is a simply-connected domain. A vector field v on

can be decomposed uniquely into the sum of a divergence-free part that is parallel to , and of a remainder
that is a gradient.

given v(x)

Figure 2.1: Problem domain.

Consider a simply-connected domain with boundary . If were given v(x), the theorem says that

v = u + q

. u| = 0.
for u(x), q(x) and for .u = 0 and n

Proof. Define q(x) as the unique (up to constant) solution of the Poisson equation

4q = .v,
. q|
n = . v| .
n

Then, define u = v q. Check that

.u = .v 4q = 0
and that
. u| = n
n . v| n
. q| = 0.

Remark. This operation defines a linear map v 7 u. Call this map P, so that P{v} = u. Then P = P2 , i.e.
P is a projection operator on the space of vector fields. (Proven by application of the theorem twice.) In
this grain, we may write
v = P{v} + (v P{v}) .
Note that the first and second terms here are orthogonal, given the right inner product.

Using this projection operator, we may write the incompressible Euler equations as
 
d 1
P u = P {f } ,
dt
Math 654, Lecture 3 1/12/11 6

or just
1
t u + u.u q = P{f }.

Taking the divergence, we get
. (u.u) 4q = 0.
This is the view taken in many numerical algorithms. Since the projection operator is linear, we often find
1
t u + P {u.u} = P{f }

in the literature.
Math 654, Lecture 4 1/14/11 1

Lecture 4: Projection Formulation, Energy, and Vorticity

Were discussing the incompressible Euler equations:

1 1 ext
t u + u.u + p = f ,

.u = 0,

given u0 (x) and b.c.s. This is a version of Newtons 2nd Law, in the limit where neighboring fluid elements
dont press each other hard enough to compress the fluid.

1 Projection Formulation of the Euler Equations

d
Last time, we saw that the Euler equations are a projection of dt u = 1 f onto divergence-free vector fields.
Explicitly, this is
 
d 1
P u= f
dt
where P is the projection operator provided by Helmholtz-Hodge. This theorem said that if is a simply-
connected domain with boundary and if v(x) is a vector field on , then

v(x) = u(x) + q

. u| = 0. In the proof, we found q by solving the Poisson problem

uniquely, where .u = 0 and n

4q = .v,
. u|
n = . v| .
n

Example. Let = [0, 1]2 and v(x) = j cos(2x) sin(2y).

Figure 1.1: v (x)

Math 654, Lecture 4 1/14/11 2

Then
.v = 2 cos(2x) cos(2y),

but note

v = k(2 sin(2x) sin(2y)

which is not identically zero. What is the divergence-free part? We need

4q = .v = 2 cos(2x) cos(2y),
. q|
n = . v| = 0.
n

y y q = 0

x q = 0 x q = 0

x
y q = 0

Figure 1.2: Boundary conditions.

The solution of this problem is

1
q(x) = cos(2x) cos(2y),
4
so
1 1
q = i sin(2x) cos(2y) + j cos(2x) sin(2y).
2 2

Hence, the divergence-free part is exactly

u = v q
1 1
= i sin(2x) cos(2x) + j cos(2x) sin(2y).
2 2
Math 654, Lecture 4 1/14/11 3

Figure 1.3: u, the divergence-free part of v.

This vector field looks quite different from v.

Back to the equations:

1
P (t u + u.u) = P(f )

1
= t u + P (u.u) = P(f ).

Note that P (u.u) = u.u (p/).
Note 1. Since t u is equal to a divergence-free vector field, if u starts off divergence-free, it will stay
divergence-free for all time.
Consider the operator P. This is a linear projection operator, since P{v1 + v2 } = P{v1 } + P{v2 } and
P{P{v}} = P{v}. So we may write
v = P{v} + (v P{v})
where, in our previous notation,
P{v} = u,
v P{v} = q.
We claim this is an orthogonal projection, with the inner product of two vector fields v1 , v2 being

v1 (x).v2 (x) dx.

Indeed,

u.q dx = .(qu) dx


= qu.
nda

= 0.
This will prove useful.
Math 654, Lecture 4 1/14/11 4

2 Energy Evolution
Consider fluid in 3-space.

u
m = V
x

x y

Figure 2.1: Fluid element.

The kinetic energy of a fluid element is 12 mkuk2 , so the total kinetic energy in some region is

1 1
Kinetic Energy = ku(x, t)k2 dx = ku(, t)k22 ,
2 2
where the L2 -norm is, as usual, s
2
kvk2 = v(x) dx.

How does the energy evolve in time? Suppose u satisfies the incompressible Euler equations with body force:
1 1
t u + u.u + p = f,

.u = 0,
n . u| = 0.
W.l.o.g., we may take f to be divergence-free.1 The kinetic energy evolves in time by

d d 1
KE = ku(x, t)k2 dx
dt dt 2

1
= ku(x, t)k2 dx
2 t

= u.t u dx
 
1 1
= u. u.u p + f dx


= u. (u.u) dx u.p dx + u.f dx.

1 Write f = P{f } , and lump into the pressure term.
Math 654, Lecture 4 1/14/11 5

Immediately, we see that the second integral is zero, for u.p = .(up) and since n . u| = 0. Physically,
this says that the internal pressures do no work. Now consider the first term. The integrand is

u.p = ui (uj xj ui )
= uj ui j ui
 
1
= uj j ui ui
2
 
1
= u. kuk2
2
 
1
= . ukuk2 ,
2
so  
1
u. (u.u) dx = . ukuk2 dx = 0
2
by the divergence-theorem. Hence,
d
KE = u.f dx.
dt

Note 2. The work done on is due only to the divergence-free part of f . Consider f = kg. We can absorb
this into the pressure term by defining the hydrostatic pressure, p = gz + constant.

3 Vorticity
Definition. The vorticity of a vector field u(x, t) is the vector field defined by (x, t) = u(x, t).

u
x

x y

Figure 3.1: Vorticity of a fluid element.

Proposition. Vorticity is twice the angular velocity of a fluid element.

Proof. This is a heuristic derivation. Consider a fluid element with velocity (u(x, y, z), v(x, y, z), w(x, y, z)).
Math 654, Lecture 4 1/14/11 6

u(x, y + y/2, z)

v(x + x/2, y, z) v(x + x/2, y, z)

y

u(x, y y/2, z)

x
x

Figure 3.2: Computing angular velocity.

The z-component of the average angular velocity (of the boundaries) is

 
1 v(x + x/2, y, z) u(x, y + y/2, z) v(x x/2, y, z) u(x, y y/2, z)
z = +
4 x/2 y/2 x/2 y/2
 
1 v u
=
2 x y

to leading order. Hence the claim.

Recall that
 
1
u.u = u + kuk2 ,
2

for

(u.u)i = ui,j uj
= ik jl uk,l uj
= (ik jl il jk + il jk )uk,l uj
= (ijm mkl + il jk )uk,l uj
= ijm (mkl uk,l )uj + uj,i uj
1
= imj m uj + i (uj uj )
2 
1 2


= ( u)i + kuk .
2 i
Math 654, Lecture 4 1/14/11 7

So we have another representation of the Euler equations, as

 
1 1 2 1
t u + u + p + kuk = f,
2
.u = 0.

This decomposes the evolution of the natural kinematics into a spin part and a no-spin part. This has
immediate implications for the solutions, which well see next time.
Math 654, Lecture 5 1/19/11 1

Lecture 5: Two More Views of the Euler Equations

. We saw two views

Recall our setup, fluid in a domain with boundary and outward pointing normal n
of the incompressible Euler equations:

1. View 1:

1 1
t u + u.u + = f,

.u = 0,
. u| =
n 0.

2. View 2:
 
d 1
P u = P {f } .
dt

Now well see a third view.

1 A Third View of the Euler Equations

Recall that
 
1
u.u = u + kuk2
2

where = u. This gives a third view of the dynamics:

 
1 1 2 1
t u + u + p + kuk = f, (1.1)
2
.u = 0,
= u.

The second implicitly determines pressure, and the third closes the system.
Note. In principle, given the vorticity, we could recover the velocity field, for

= u = (u) 4u = 4u.

So we would solve the Poisson equation with boundary conditions:

4u = ,
. u|
n = 0.

This has solution by the Biot-Savaart Law, which uses a Greens function.
Math 654, Lecture 5 1/19/11 2

1.1 Local Evolution of Kinetic Energy

1
Recall that the kinetic energy is 2
kuk2 dx. Dotting u into (1.1), we have
   
1 1 1 1
kuk2 2
+ u. p + kuk = u.f ,
t 2 2

or equivalently,     
1 1 1
kuk2 + . p + kuk2 u = u.f .
t 2 2
This says that the local evolution of kinetic energy is determined by fluxes and work done by an external
force. This has immediate consequences.

1.2 Bernoullis Theorem

Definition 1. A flow is said to be stationary if the time derivative of any flow property is zero.

For stationary flow with no external force,

 
1
u. p + kuk2 = 0,
2
d
and also u. () (), hence
dt  
d 1
p + kuk2 = 0.
dt 2
So weve proved a theorem.

Theorem. (Bernoulli) For stationary ideal flows, the quantity p + 12 kuk2 for any fliud element is constant.

p lower p higher

Figure 1.1: Bernoullis principle.

2 A Fourth View of the Euler Equations

Take the curl of (1.1) to get
1
t + ( u) = f.

Math 654, Lecture 5 1/19/11 3

A quick calculation shows

( ( u))i = ijk j klm l um

= (il jm im jl ) (l,j um + l um,j )
= i,j uj + i uj,j ui j,j ui,j j
= i,j uj ui,j j .

Putting it together, we arrive at the vorticity evolution equation:

1
t + u. .u = f,

= u.

There are some interesting consequences.

For example, suppose there is no external force. Then

d
= .u.
dt
But note that not every component of the velocity gradient tensor contributes. We can write

1 1
(u)ij = uj,i = (ui,j + uj,i ) + (uj,i ui,j ) = Sij + Aij ,
2 2
and the second term is the anti-symmetric tensor

0 3 2
1
A = 3 0 1 .
2
2 1 0

We claim that
. (u) = . (S + A) = .S,
i.e.,
.A 0.
This is a straightforward matrix calculation. So in fact, when there is no external force,

d
= .u = .S,
dt
1
where Sij = 2 (ui,j + uj,i ), the rate-of-strain tensor.
As a consequence, we have the following

Proposition. Suppose we have a smooth solution to the Euler equations, and of a particular fluid element
is zero at t = 0. Then, it will remain zero for all time. In particular, if (x, 0) 0, then 0 forever.

Definition 2. Flows with no vorticity are said to be irrotational.

Math 654, Lecture 5 1/19/11 4

What happen if the vorticity is not zero? In R3 , .u is called the vortex-stretching term, for the following
reason. Suppose of some fluid element is non-zero, and suppose u changes across the fluid element in such
a way that .u 6= 0. Then, the ensuing changes in the fluid elements moment of inertia lead to changes
in , via conservation of angular momentum. Thus, vortex-stretching see the figure below.

u
t ! t + t

Figure 2.1: Vortex-stretching in a fluid element.

Nobody really knows whether vortex-stretching can lead to finite-time blowup, even given regular initial
conditions. Later on, well see why vorticity plays a big role in regularity of solutions.
0. So the dynamics are now
But in R2 , things are more nicely behaved. Indeed, we have z 0 and k.u

1 1
t u + u.u + p = f (x, y, t),

= 0. In components,1
with k.f

1 1
ut + uux + vuy + px = f1 (x, y, t),

1 1
vt + uvx + vvy + py = f2 (x, y, t).

Vorticity satisfies

= i(wy vz ) + j(uz wz ) + k(v

x uy )
x uy ).
= k(v
1 And note that
1 1
t w + uwx + vwy + p z = f3 ,

so t w = 0.
Math 654, Lecture 5 1/19/11 5

Hence,  
3 .u = 3 z u 0.
.u = k

( u) the scalar vorticity, which

So in R2 , there is no vortex-stretching. In this case, we call 3 = k.
satisfies
d
= 0.
dt
Basically, this says that the moment of inertia of fluid elements does not change for flows in R2 . Note that
if there was an external force,
 
d 1 1 f2 f1
= k. ( f ) = .
dt x y

3 Enstrophy
1
Recall that the kinetic energy is 2
kuk2 dx. Similarly, we define another quantity.
Definition 3. The enstrophy is given by

kk2 dx = kk22 .

We saw that in the absence of external forces, energy is conserved. In a way similar to that calculation, one
can show that
d
(enstrophy) = .S. dx.
dt
d
In R2 , dt (enstrophy) = 0.
Exercise. In R2 ,
d
= 0,
dt
t + uwx + vwy = 0.

Show that
d n
((x, y, t)) dxdy = 0.
dt
Math 654, Lecture 6 1/24/11 1

Lecture 6: Introduction to Potential Flow Theory

We have developed various ways of looking at the Euler equations:
1 1
t u + u.u + p = f

.u = 0,
or
 
1 1 1
t + u + p + kuk2 = f
2
= u,
or
1
t + u. = .u + f.

Note that these are all nonlocal equations for the dynamics. This is a consequence of the incompressibility
condition.

1 Irrotational (Potential) Flows

Consider a periodic channel, i.e., consider solutions which are periodic in the y-direction with period Ly .

z
Lz

y
Ly

Lx

x . u| = 0
n

Figure 1.1: Flow in a periodic channel.

One steady solution is

u = jv(x, z),
p = constant.
This is a solution, for
v
.u = 0,
y
p = 0, and

u. = v(x, y) , so
y
u.u = 0.
Math 654, Lecture 6 1/24/11 2

There is a horrible non-uniqueness here we have produced an entire class of solutions!

Are there any special solutions? Consider the vorticity,

= u
v v
= i +k .
z x
A distinguished solution is one with no vorticity. Then, z v = 0 = x v we obtain solid body flow. Flows
of no vorticity are generally distinguished.
Definition 1. Flows with no vorticity are said to be irrotational.

For the rest of this section, well assume that 0, i.e., u 0. Note that in simply connected domains,
u = for a unique (up to constant) function . So we make a definition.
Definition 2. If u = , we call the velocity potential for u.

Why is unique? Consider a simple closed curve in an irrotational flow.

dl

dl
u

Figure 1.2: A simple closed curve C in flow u.

By Stokes theorem,
u.dl = ( u) .da = 0
C A
as the flow has no vorticity. Thus, we may define
x
(x) = u.dl
x0

along any curve connecting x0 and x. This is well-defined, by the line above.
Definition 3. The circulation of a flow about a simple closed curve C is

C = u.dl.
C

So irrotational flows have no circulation.

Math 654, Lecture 6 1/24/11 3

1.1 Pressure in a Potential Flow

What do the Euler equations look like for potential flows? Barring the presence of external forces,
 
1 1
t u + p + kuk2 = 0,
2
.u = 0.

For stationary (t 0) flows,  

1
p + kuk2 = 0,
2
and thus
1
p(x) + ku(x)k2 = constant.
2
This is much stronger than Bernoullis theorem.
For time-dependent irrotational flows,

u =
= t u = (t ) .

Hence,
 
1 1 2
(t ) + p+ kk = 0
2
 
1 1
= t + p + kk2 = 0,
2
and hence
1
t + p + kk2 = C(t),
2
a time-dependent, spatially-invariant constant. This is a diagnostic equation that gives the pressure field:
1
p(x, t) = C(t) t kk2 .
2
We have thus eliminated the need to solve a pde for the pressure. Instead, given a divergence-free velocity
field, we can integrate on curves to get the velocity potential and plug it into the equation on the line above.
Now, the pressure is not a thermodynamic pressure; it is determined by the need to keep C(t) constant in
space.

1.2 The Velocity Potential

We now derive an important relation for . Since u = and .u = 0, we have

4 = 0.

So the velocity potential satisfies Laplaces equation! Wed like to think of this as a bvp for . Recall the
boundary conditions in our setup are n . u| = 0, so

. | = 0.
n
Math 654, Lecture 6 1/24/11 4

n

| = 0
u. n
= 0
| = 0
. n

Figure 1.3: Laplaces equation with Neumann boundary condition.

Our problem is now linear so long as we have which solves Laplaces equation with Neumann boundary
conditions, we have generated a velocity field which satisfies the Euler equations (along with the pressure
field determined above).

Laplaces equation with homogenous Neumann boundary conditions have a unique solution up to a constant.
(See the homework.) But the additive constant does not change the velocity field. In fact,

0= 4 dx = kk2 dx,

after integrating by parts and using the Neumann boundary condition. So the only incompressible, irro-
tational flows existing on simply-connected domains are the flows with no flow at all (c.f., channel flow,
topologically a torus).

A more general bvp comes by using the inhomogenous boundary condition

. u| = un (x, t).
n

Note that

un (x, t) da = 0

must hold for compatibility. So then wed look for solutions of Laplaces equation

4 = 0

subject to
. | = un .
n
Math 654, Lecture 6 1/24/11 5

n

| = un (x, t)
u. n
= 0

Figure 1.4: Laplaces equation with generic boundary condition.

Proposition 1. There exists a unique (up to time-depedent constant) solution to the bvp posed above.

So we can find a unique (x, t) satisfying Laplaces equation with specified flux at the boundary. Given
such , we can then compute u = and p = t 12 kk2 . By allowing non-zero flux across the
boundaries, we allow for nontrivial flows to exist.

2 2D Incompressible, Irrotational, Ideal Fluid Flows

Since we work in R2 , we have z 0, w 0, and thus u = iu(x, y, t) + jv(x, y, t). Incompressibility means

.u = 0,

and irrotationality means

u = 0.
In coordinates,
u v
+ = 0,
x y
v u
= 0.
x y
These are the Cauchy-Riemann equations for the real and imaginary parts of a complex analytic function.
Indeed, let z = x + iy C, and write

F (z) = u(x, y) iv(x, y).

Note the minus sign.

Definition 4. If
F (z + z) F (z)
lim
kzk0 z
exists, then F : C C is called analytic.
Math 654, Lecture 6 1/24/11 6

Proposition 2. (Cauchy-Riemann Equations) If F is analytic, then

u v
= ,
x y
u v
= .
y x

Proof. Enforce the limit in the definition of analyticity in two ways. The first relation follows when z = x,
and the second when z = iy.

Weve demonstrated thus that the real and imaginary parts of any complex analytic function comprise the
components of a potential flow in R2 .
Recall that for analytic F , we may write
dW
F (z) = ,
dz
allowing for multi-valued W (e.g., F (z) = log z).
Definition 5. In the notation above, W is called the complex potential.

Now, write W (z) = (x, y) + i(x, y). Then,

F (z) = u(x, y) iv(x, y)

= +i
x x

= i +
y y
by the Cauchy-Riemann equations. Notice that

u= , v= ,
x y
so is the velocity potential. Also notice that

u= , v= .
y x
Definition 6. In the notation above, is called a stream function.

Note that

u. = u +v
x y
= u(v) + vu
= 0,

and hence
0 = ..
Math 654, Lecture 6 1/24/11 7

Thus, potential flows in R2 have velocity tangent to level sets of , and perpendicular to level sets of . And
if the flow is stationary (t 0), then the level sets of are the particle paths.

=3
=2

=1

=1
=2
=3

Figure 2.1: Level sets of and in potential flow.

A final calculation:

kuk2 = kk2
= u2 + v 2
 2  2

= +
y x
= kk2 .

So the fluid speeds up when the gradient of the stream function is large (when the level sets of bunch
together), and slows down when the gradient is small (when the level sets of spread apart). This can be
understood via conservation of mass.
Math 654, Lecture 7 1/26/11 1

Lecture 7: Potential Flow Theory and Vortex Dynamics

We left off talking about ideal, incompressible, irrotational flow. Taking 1, the Euler equations are

t u + u.u + p = 0,
.u = 0.

Setting = u, we can rewrite as

 
1
t u + u + p + kuk2 = 0.
2
Since the flow is irrotational, = u = 0, so on simply-connected domains,

u = ,
1
t + p + kuk2 = C(t).
2
To find u, we examine solutions of
4 = 0
subject to various boundary conditions.

= 0 | = g(x, t)
. n

Figure 0.1: Setup for potential flow theory.

This amount to harmonic analysis.

1 Potential Flow in R2
and
Recall that in R2 , = k,
u v
+ = 0,
x y
v u
= 0,
x y
Math 654, Lecture 7 1/26/11 2

the Cauchy-Riemann equations for an analytic function

F (z) = u iv.

We saw that we can write the (possibly multivalued) complex potential W so that

dW
F (z) = .
dz

When we write
W (z) = (z) + i(z),

we recover the velocity potential, , and introduce the stream function, .

1.1 Some Elementary Flows

Example 1. Perhaps the simplest analytic function (beside a constant function) is W1 (z) = (U iV ) z. In
this case, W1 (z) = U x + V y + i (U y V x), so u = = iU + jV . The level sets of are given by

U y V x = const.,

which are straight lines along the flow.

y
= c4
= c3
= c2
= c1

= c3

= c2

= c1

Figure 1.1: Elementary flow I.

Example 2. The next simplest would be W2 (z) = 12 z 2 . So W1 (z) = 1
2 x2 y 2 + 2ixy , and hence


= 12 x2 y 2 and = xy. The flow has u = = ix jy.
Math 654, Lecture 7 1/26/11 3

stagnation point

Figure 1.2: Elementary flow II, uniform strain.

Note that the sum of any two analytic functions is also analytic. E.g., we could examine the complex
potential W1 + W2 to get a new flow.


Example 3. W3 (z) = U z + a2 /z for U, a R. Then
 
a2 x iy
W3 (z) = U x + iy +
x + iy x iy
 2
 
a x a2 y
= U x+ 2 + i y .
x + y2 x2 + y 2

So
 
a2
= Ux 1 + 2 ,
r
 
a2
= Uy 1 2 ,
r

with r2 = x2 + y 2 .
Math 654, Lecture 7 1/26/11 4

Figure 1.3: Elementary flow III, flow past a cylinder.

Example 4. W3 (z) = 2i log az for , a R. Recall that log z = i + log r, so the complex logarithm is
multivalued. But thats okay for our analysis.

dW3
F (z) =
dz
1
=
2i z
x iy
= ,
2i x2 + y 2

so the velocity field is

   
y j x .
u = i +
2 r2 2 r2

The complex potential and stream function satisfy

= ,
2
r
= log .
2 a

Note that the speed satisfies

2 1
kuk2 = .
4 2 r2
Math 654, Lecture 7 1/26/11 5

>0

Figure 1.4: Elementary flow IV, elementary vortex.

A quick calculation. For a simple closed curve C,

C = u.dl = da
A
C

by Stokes theorem. If C does not enclose the origin, C = 0. If C encloses the origin, deform it to a circle
to get

C = 2r = 6= 0.
2r

So if we expect Stokes theorem to hold, we could write

u = (x).

At the origin, there is a point vortex of infinite strength. We have found a Greens function for the curl
operator.
Math 654, Lecture 7 1/26/11 6

1.2 Point Vortex Dynamics

Consider a pair of elementary vortices.

2 < 0

1 < 0

Figure 1.5: A pair of elementary vortices.

Now consider N elementary vortices which interact. We demand that each of the vortices moves in the flow
induced by the others, as to satisfy the Euler equations1 . So write
N
X  
k z k (t)
W (z, t) = log ,
2i a
k=1

with
k (t) = xk (t) + iyk (t).
dW
Recalling that F = dz = u iv, we define (CHECK THE SIGNS)

N
X  
d d j z j 
k = log 
dt dz 2i a 
j=1,j6=k 
z=k
N
X j k j
= .
2i |k j |2
j=1,j6=k

Then, the dynamics are given by

X j yk yj
x k = ,
2 (xk xj )2 + (yk yj )2
j6=k
X j xk xj
y k = .
2 (xk xj )2 + (yk yj )2
j6=k

This is a 2N -dimensional system of ordinary differential equations. But in fact, there is additional structure.
This is a Hamiltonian system:
H
k x k = ,
yk
H
k y k = ,
xk
1 To be specific, this description of vortex dynamics satisfies the weak formulation of the Euler equations.
Math 654, Lecture 7 1/26/11 7

with
X X k j  
kxj xk k
H= log .
j
2 a
k6=j

Here, we have xj = (xj , yj ). In the sense of Hamiltonian dynamics, xk is a canonical position and yk is a
canonical momentum. This is used as a computational tool e.g., we can model a shear layer as many, many
point vortices in a line.

Figure 1.6: Vortex sheet modeling an interface.

Math 654, Lecture 8 1/31/11 1

Lecture 8: Vortex Dynamics and DAlemberts Paradox

The equations:
1
t u + u.u = f,

.u = 0.

. u| = 0. With the definition = u, we have

The boundary condition is n
 
1 1 1
t u + u + p + kuk2 = f,
2
1
t + u. = .u + f .

Recall that irrotationality, = 0, means we can write u = . (The potential is unique up to constant in
simply-connected domains.) Incompressiblity implies .u = 0, so we arrive at

4 = 0.

The pressure is determined by

1 1
t + p + kk2 = C(t).
2

In R2 there is even more structure. The vorticity becomes a scalar,

 
v u
= k = k ,
x y

and irrotationality and incompressibility now read

v u
= ,
x y
u v
= .
x y
These are the Cauchy-Riemann equations for F = u iv. Setting z = x + iy, as usual, we can write
F (z) = W 0 (z) with W (z) = + i, allowing for multivalued antiderivatives (e.g., log ()).

1 Point Vortex Dynamics (ctd.)

The following is a review.

Example 1. Point vortex:

z 
W (z) = log .
2i a
Note that there is circulation about the origin, but the flow is everywhere irrotational (except the origin,
where it is not defined).
Math 654, Lecture 8 1/31/11 2

If we set N point vortices in the plane at locations k = xk +iyk , k = 1, . . . , N , we can ensure the flow satisfies
the Euler equations by requiring that each point vortex movies in the velocity induced by the others. Then
we arrive at 
X  
d z j 
x k iy k = =
j
log  .
dz 2i a 
j6=k
z=k
Then the equations of motion are
X j yk yj
x k = ,
2 (xk xj )2 + (yk yj )2
j6=k
X j xk xj
y k = .
2 (xk xj )2 + (yk yj )2
j6=k

This system has a Hamiltonian structure:

H
k x k = ,
yk
H
k y k = ,
xk
with
X j k  
kxj xk k
H= log .
4 a
j6=k

What does the Hamiltonian structure tell us? Look for symmetries. The hamiltonian has both translational
and rotational invariance, so we arrive at three other conserved quantities
N
X
X = j xj ,
j=1
N
X
Y = j yj ,
j=1
N
X

L = j x2j + yj2 .
j=1

The three quantities of interest here are thus H, L, and X2 + Y2 .

2 DAlemberts Paradox
Consider a body B in a potential flow with far-field velocity given by the requirement that
u iU + jV as kxk = |z| .
Thus,
F (z) U iV,
W (z) (U iV ) z
Math 654, Lecture 8 1/31/11 3

as |z| . We ask: what is the magnitude and direction of the force felt by the object?

Figure 2.1: Flow over a body.

Let f = ifx + jfy be the force1 felt by the object. Then we define a complex force, F = fx ify . We have
the following theorem.

Theorem 1. (Blasius) The force on an object in a stationary, incompressible, irrotational flow with complex
velocity F (z) outside the body is

i 2
F= (F (z)) dz.
2 B

Proof. Observe that f is due to pressure differences around B, in that

f = p ds.
n
B

In components,
 
fx = i.f = p i.
n ds,
B  
fy = j.f = p j.
n ds.
B

We can write an infinitesimal length element as dl = idx + jdy and the normal component at the boundary
= i sin + j cos , so that
is n
dl = i ( cos ) ds + j sin ds.

Since i.
n ds = sin ds = dy and j.
n ds = cos ds = dx, we have

dl = idy jdx.
1 As were in R2 , really this is force/length. But recall that the object is infinite in and out of the page.
Math 654, Lecture 8 1/31/11 4

dy
ds

dx

Figure 2.2: Infinitesimal length elements on B.

Hence,

fx = p dy,
B

fy = p dx.
B

Therefore,

F = fx + ify

= p (dy i dx)
B
= i p dz,
B

as i dz = i (dx + i dy). Since

2

p= u + v 2 + C(t),
2
we arrive at
i

F= u2 + v 2 dz.
2 B

To finish, consider that

2
F 2 dz = (u iv) (dx + i dy)
 2

= u v 2 2iuv (dx + i dy)



= u2 v 2 dx + 2uv dy + i u2 v 2 dy 2uv dx .

Recall that flow is tangent to the boundary. The tangent to the boundary is dl = idx + jdy. Since u k dl,

0 = u dl = k (udy vdx) ,

and hence u dy = v dx on B. So we can write

F 2 dz = u2 + v 2 dx i u2 + v 2 dy


= u2 + v 2 dz.
Math 654, Lecture 8 1/31/11 5

Hence,
i 2
F= (F (z)) dz,
2 B
and hence
i 2
F = fx ify = (F (z)) dz.
2 B
This proves the claim.
Theorem 2. (Kutta-Joukowski Theorem) Given the setup above,
 
f = C U k ,

where
C = u.dl
C
with C being any closed curve containing the object.
Remark. Thus DAlemberts paradox is that f u, so potential flow exerts no drag.

Proof. In our setup, F (z) is complex analytic in the domain outside of the object B. W.l.o.g., we can place
the origin inside B. So F (z) has a convergent Laurent expansion
a1 a2
F (z) = a0 + + 2 + ....
z z
Note that as |z| , F (z) a0 , hence a0 = U iV and there are no positive powers in the expansion
above. Cauchys theorem states that
F (z) dz = 2ia1 .
C
We can take C = B to get

1
a1 = F (z) dz
2i B

1
= (u iv) (dx + i dy)
2i B

1 1
= (u dx + v dy) + (u dy v dx)
2i B 2 B

1
= (u dx + v dy) ,
2i B
as u dy = v dx on the boundary. Thus,

1 C
a1 = u.dl = .
2i B 2i
So we can write
C 1 a2
F (z) = U iV + + 2 + ...
2i z z  
2 2 2 C (U iV ) 1 1
= (F (z)) = (U iV ) + +O ,
2i z z2
Math 654, Lecture 8 1/31/11 6

and by the residue theorem,

2
(F (z)) dz = 2C (U iV ) .
B
Blasius theorem tells us that
i
F = fx ify = 2C (U iV ) ,
2
and hence

fx = C V,
fy = C U.

But this is exactly f = C (U k), so were done.

Kutta-Joukowski tells us that there is lift exerted on the object, but there is no drag. This is a non-physical
state of affairs. Note that there is a three-dimensional version of this theorem, but still there is no drag. So
what are we missing? Viscosity! Well start this next time.
Math 654, Lecture 9 1/31/11 1

Lecture 9: Introduction to Viscous Fluids

In real fluids viscosity plays a role. This lecture well try to understand the physics of viscosity. Later well
see how to incorporate viscosity into Navier-Stokes.

1 Viscosity and the Stress Tensor

As a first example, consider two fluids.

Figure 1.1: Two fluids at an interface.

Shear stress comes about in this example from the diffusion of momentum across the fluid boundary. To
incorporate this on a macroscopic level, we consider a fluid element of volume V = xyz.
y

22 (x, y + y/2, z)

21
12 12

(x, y, z)
11 (x + x/2, y, z)
11
22

21

Figure 1.2: Stress on a fluid element.

Math 654, Lecture 9 1/31/11 2

In our setup, ij is the force per unit area in the jth direction across an area element with normal in the ith
direction. Well adopt the following sign convention: ij is the force on the () side due to the fluid on the
(+) side. So the net force on the element in the 1-direction is (to first order)
     
x y x
F1 = 11 x + , y, z yz + 21 x, y + , z xz 11 x , y, z yz
2 2 2
     
y z z
21 x, y , z xz + 31 x, y, z + xy 31 x, y, z xy
2 2 2
 
11 21 31
= + + xyz
x y z
= (.)1 V.
Carrying out this analysis in each direction, we arrive at
F = .V,
or just
ji
Fi = V.
xj
Definition 1. In the above notation, we call the stress tensor of the fluid.
Proposition. The stress tensor is symmetric.


One way to see this is to consider the fluid as having a velocity distribution u (x, t) + u0 . Then ij = u0i u0j
which is obviously symmetric. Another way is to consider the macroscopic dynamics: well see that the
stress tensor must be symmetric to prevent infinite angular accelerations. E.g., in the 3-direction, the torque
on the fluid element is (to first order)
X
N3 = k. r F
faces
   
x x x x
= 12 x + , y, z yz + 12 x , y, z yz
2 2 2 2
   
y y y y
21 x, y + , z xz 21 x, y , z xz
2 2 2 2
= V (12 21 ) .
We claim this must be zero to prevent infinite angular accelerations. To conclude this, we first compute the
moment of inertia. Since is locally constant, we have



I33 = x2 + y 2 dxdydz
1  2 2

= (x) + (y) V,
12
and thus
N3 12 21
angular acceleration = 2 2.
I33 (x) + (y)
So if 12 6 21 , then as x, y 0, the angular acceleration blows up.
=
Math 654, Lecture 9 1/31/11 3

2 Rate of Strain Tensor for Newtonian Fluids

Now well seperate the stresses into reversible and irreversible parts. W.l.o.g., write

ij = ij p + ij

as to isolate the normal stresses (due to p), which are reversible. The deviatoric stress tensor will represent
the irreversible (frictional) stresses. Now we have a definition.

Definition 2. A Newtonian fluid is one in which the frictional stresses are proportional to the rate of
strain tensor.

The rate of strain tensor describes the deviation of the elements motion from rigid body motion. To
account for these motions, consider x which is a small displacement of two points in the fluid, x(t) and
x(t) + x(t).

x(t) + x(t) x(t) + x(t)

t + t t + t

x(t) x(t)
(a) Rigid motion. (b) Non-rigid motion.

Figure 2.1: Rigid vs. non-rigid body motion.

In time t, rigid body motion will preserve the length of x. But non-rigid body motion leads to changes in
the length of x. The rate of strain tensor will account for the property of the fluid motion which gives rise
to the latter motions. So we have a calculation. On the one hand,

d d
kxk2 = 2kxk kxk,
dt dt
and on the other hand,
 
d d
kxk2 = 2x. x
dt dt
= 2x. (u (x + x, t) u (x, t))
ui
= 2xi xj
xj
 
ui uj
= xi + xj .
xj xi
Math 654, Lecture 9 1/31/11 4

Thus,
 
1 d xi 1 xj
kxk = (ui,j + uj,i ) .
kxk dt kxk 2 kxk
So we have a definition.

Definition 3. The rate of strain tensor is the symmetric part of the velocity gradient tensor,

1
rij = (ui,j + uj,i ) .
2

Now recall we have written

= pI + ,

and we saw that a Newtonian fluid is one for which r, with r = (u)symm. . The most general linear,
isotropic relation between two tensors is
ij = Aijkl rkl

with
Aijkl = aij kl + bik jl + cil jk .

So

ij = aij tr (r) + brij + crji

= aij tr (r) + (b + c) rij ,

since r is symmetric, and thus

ij = c1 rij + c2 tr (r) ij .

Recalling the definition of rij , we find that

tr (r) = rkk = uk,k = .u,

and hence that  

c1 ui uj
ij = + + c2 (.u) ij .
2 xj xi
Adding and subtracting, we may write this as
 
ui uj 2
ij = + uk,k ij + uk,k ij
xj xi 3

where , are constants determined by c1 , c2 . Why would we write this as such? One reason is that the first
term is trace-free, so it represents the volume-preserving deformations of fluid elements. The second term
accounts for deformations which change the volume of fluid elements.

Definition 4. In the notation above, is the coefficient of dynamic (shear) velocity, and is the coefficient
of bulk viscosity.
Math 654, Lecture 9 1/31/11 5

So we have arrived at  
1
ij = pij + 2 rij rkk ij + rkk ij .
3
The rate at which work is being done on the fluid element is given by ij rij . So for Newtonian fluids,
 
2 2 2
rate at which work is being done on fluid element = prkk + 2rij rij (rkk ) + (rkk ) .
3

The first term is the rate of p dv work; the second is the rate of viscous energy dissapation,
2 2 2
= 2rij rij (rkk ) + (rkk ) .
3
For thermodynamic reasons, we expect that 0. This is the same as asking that , 0.
Next time, well see how to put stresses into Navier-Stokes, mainly by
d
u = total force per volume.
dt
Math 654, Lecture 10 2/9/11 1

Lecture 10: Navier-Stokes Equation

A quick review of what we had last time.

1 Stress/Rate-of-Strain Tensors (review)

Recall we had the stress tensor, , with
ij = force/area in j-direction acting across area element with normal in i-direction.
More generally,
m
n = force/area in m-direction
-direction.
acting across area element with normal in n

Figure 1.1: Area element.

The force on a fluid element of volume V is

F = .V,

Fi = ji V.
xi
Note that ij = ji . We also developed the rate of strain tensor, r, has
1
rij = (ui,j + uj,i ) .
2

At each point in the flow, r is symmetric, so we can decompose the dynamics w.r.t. an orthonormal basis of
eigenvectors of r. Then a rod oriented along ei instantaneously stretches at rate i .

3
e e2

e1

Figure 1.2: Eigenvectors of r.

Math 654, Lecture 10 2/9/11 2

Motion in R3 has the following components:

Translation u
Rotation =u
Compression/extension .u
Deformation (pure strain) rij 31 rkk ij

Note that 1 + 2 + 3 = tr r = .u, so compression/extension is captured by r. And so is deformation,

which is the trace-free part of r. Have we lost information or put more in somehow? If we take the velocity
vector field and all of its derivatives, we have 3 + 9 = 12 pieces of information. In the four types of motion
listed above, we have 3 + 3 + 1 + 5 = 12 pieces of information. So it seems that indeed all of the kinematic
properties needed to propagate a fluid element forward one step in time is included in the velocity field (and
its derivatives).
Now observe that the rate at which work done on the fluid element is given by

stress rate = : r = ij rij = tr ( r) .

| {z }
forcevelocity

W.l.o.g., we wrote
ij = pij + ij ,
with being the deviatoric stress tensor. And for Newtonian fluids, we saw that r along with isotropy
forces

ij = Aijkl rkl
= (c1 ij kl + c2 ik jl + c3 il jk ) rkl
= c1 ij rkk + (c2 + c3 ) rij
   
1 c2 + c3
= (c2 + c3 ) rij ij rkk + c1 + ij rkk .
| {z } 3 3
2 | {z }

Note that , are physical properties of the fluid in consideration. So for Newtonian fluids,

ij = pij + ij
 
1
= pij + 2 rij ij rkk + ij rkk .
3

What should the signs be on , ? First, note that for any tensor M we define

kM k2 = tr (M M ) = Mij Mij .

Then the power/unit volume is

    
1 1 1
ij rij = prii + 2 rij ij rkk + ij rkk rij ij rll + ij rll
3 3 3
 
1 2
= p (.u) + 2kr tr (r) T k2 + (tr r) .
3
Math 654, Lecture 10 2/9/11 3

So as long as , > 0, then the work done by the deviatoric stresses will be positive. So it is like a friction.
The first term is the p dv work, and the second term is the dissapative work. Lets give the second term a
name:
1 2
= 2kr tr (r) T k2 + (tr r) ,
3
which is the rate of dissapation in the fluid.

2 Viscous Navier-Stokes
Consider that the total force/volume acting on a fluid element is given by

. + f ext .

For Newtonian fluids, we can make the following calculation:

Fi
= ji
V xj
   
1
= pji + 2 rij ji rkk + ij rkk + fiext
xj 3
p 2
= + (ui,i + uj,i ) (.u) + (.u) + fiext
xi xj 3 xi xi
 
p 1
= + 4ui + + (.u) + fiext .
xi 3 xi

Now were ready to present full Navier-Stokes.

Fluid properties were interested in are , p, u, and the internal energy/mass E. The dynamics are thus:

Conservation of mass:
d
+ (.u) = 0
dt

+ . (u) = 0.
t

Conservation of momentum:
d
u = . + f ext .
dt
For Newtonian fluids,
  

+ u. u = p + 4u + + (.u) + f ext .
t 3

This is know as the Navier-Stokes equation.

Conservation of (internal) energy:

d
E = p (.u) + + . (kT ) ,
dt
Math 654, Lecture 10 2/9/11 4

where T is the temperature at a point and k is a material coefficient called the thermal conductivity
coefficient. The third term comes from Fouriers law. We have

E = cT,

with c being a material parameter called the specific heat.

Thermodynamically determined variables:

pressure: p = p (, T ) ,
material coefficients: , , k, c.

Note that the material coefficients generally are all functions of , T .

We have five unknowns (, p, u, E, T ) and five equations. Throughout the rest of the course, well assume
that all the material parameters are constants. This will greatly simplify the calculations, but well still find
a wealth of information in the equations. Wed also like to only consider incompressible flows, i.e., wed
like = constant and .u = 0. But consider a special class of fluids flows driven by bouancy due to a
gravatational field. Certiainly we cant demand incompressibility everywhere for these flows!
For many fluids, the coefficient of expansion
 
1
= >0
T p

is a relatively small number. E.g., p = kB T for an ideal gas and

 
1 p 1 1
= = 102 -103 .
T kB T p T K

So then Taylor expanding (T ) to get

(T ) = 0 (T0 ) (T0 ) (T T0 ) + . . .
(T )
= = 1 (T T0 ) + 1 %.
0

Sometimes those %s will matter, and in particular for bouyancy-driven flows. Thus we take the Oberbeck-
Boussinesq approximation we say the flow is incompressible except for bouancy forces. In particular, we
have the following approximations:

1. We model density as
= 0 + 0 (T T0 ) ,
and so for external forces due to gravity we write
= k
f ext = kg 0 g k
0 g (T T0 ) .

2. We neglect viscous heating, i.e., we neglect .

Math 654, Lecture 10 2/9/11 5

To justify the second assumption, recall the internal energy equation

 
T d
c + u.T = E = p (.u) + + . (kT ) .
t dt

Well try to identify the (temperature-driven) situations which might be characterized as having negligible
compared to the convective term, cu.T . Neglecting should be OK if

U2 T
= kuk2  0 c0 U .
l2 l
Note that as we already assume flows to be incompressible, the term in doesnt enter into this calculation
(as .u = 0 from the start). If the flow is driven by bouancy, the momentum equation tells us that

U
0 0 gT,
t
or just
U
gt.
T
Hence,
2
Ul2 0 U
=
0 c0 U T
l
0 c0 lT
0 gt
= .
0 c0 l

For a typical ideal gas, the first term 108 sK. Typically, g = 10 m/s and = 1/300 K. So,
2
Ul2  s t
1010 .
0 c0 U T
l
m l

Thus for dynamics with reasonable t/l, the second approximation is justified.
Math 654, Lecture 11 2/14/11 1

Lecture 11: Boussinesq Approximation, Boundary Conditions, and Rayleigh-Benard

Convection
We developed the continuum equations of motion, including Navier-Stokes. The dynamics are given by
d
t + . (u) = 0 + .u = 0,
dt
(t u + u.u) = . + f ext .

For Newtonian fluids,    

2
ij = pij + ui,j + uj,i ij uk,k + ij uk,k .
3
| {z }
ij
diss

A usual example of external forces is f ext = kg. And we had
d
E = .u + + . (kT ) ,
dt
where E = cT and = ij
diss
rij . To close it, we need an equation of state

p = p (, T ) .

Note that all material parameters , , c, k are generally functions of T, .

1 Boussinesq Approximation (ctd.)

Wed like to study a subset of the equations which capture most of the phenomena we might see elsewhere.
So we make the approximation:

.u = 0,
= constant,
, , c, k = constant.

Also, we assume that the thermal expansion coefficient

 
1
= >0
T p

is approximately constant:
0 0 (T T0 ) .

Note that once we make these assumptions, the external force (supposing gravity only) can be written as
k0 gT [0 (1 + T0 ) z] .
f ext = kg

The second term is the gradient of a hydrostatic pressure. The first is the buoyancy force. Last time, we
argued that if the flow is driven by buouancy and if the l, t-scales of the flow satisfy tl  1010 m/s, then we
can neglect the viscous heating terms in the energy equation.
Math 654, Lecture 11 2/14/11 2

All of the above constitute the so-called Boussinesq approximation. In this setup, the evolution equations
become

.u = 0,
1
t u + u.u + p = 4u + kgT,

k
t T + u.T = 4T.
c

Note that compressional heating can make a difference in gaseous flow! So there is a ghost term here,
coming from the term .u in the original energy equation. But if we define
(
cv liquids
c= ,
cp gases

this we can capture this piece of the dynamics.

What has changed from incompressible Navier-Stokes? The only new term showing up in either the continuity
or momentum equations is the term

4u.

And there is a similar term in the energy equation:

k
4T.
c

So we make two new definitions.

Definition 1. The kinematic viscosity is

h L2 i
= .
T
The thermal diffusion coefficient is
k
= .
c

Note that there are some inconsistancy in the Boussinesq approximation. For example, the approximation
does not strictly satisfy mass or energy conservation. But in these approximations we can (to first order)
capture the dynamics of bouancy-driven flows.
There are other applications that take Boussinesq-like approximations. One such application is oceanography.
In this setting, we are interested in salinity throughout the fluid,

S (x, t) = concentration of salt.

Then one might take the approximation

= 0 + (S S0 ) ,
f ext
= kgS,
Math 654, Lecture 11 2/14/11 3

and write the evolution equation

d
S = S 4S.
dt
Here,
T
= Lewis number  1
S
in practice.

2 Boundary Conditions

As usual, we require
. u| = 0,
n

i.e., we expect that no flow occurs across the walls.

fluid

Figure 2.1: Outward pointing normal at .

But now that we have the Laplacian operator to deal with, we need more boundary conditions. A basic
modelling assumption is that the fluid near a wall remains stuck to the wall. This is called the no-slip
boundary condition:
u| = 0.

Note that in many situations, this is empirically valid. But in some applications, the boundary applies
negligible shear stress (e.g., the ocean/air boundary, or the liquid-core/mantle boundary). In these cases,
one would use the free-slip condition, as follows. Consider a boundary with normal component in the
negative z-direction.

fluid w=0

= k
n

Figure 2.2: Boundary with downward pointing normal.

Write
u = iu + jv + kw

Math 654, Lecture 11 2/14/11 4

and note w = 0 at the wall. Free-slip means

0 = (
n. ) n
1 1


x u 2 (y u + x v) 2 (z u + x w) 0
= 0 0 1 y v 1
(z v + y w) 0
2
z w 1

12 (z v + y w)
= 1 (z u + x w)
2
0
= (z v, z u, 0)

at the rigid surface, in our setup. In general, free-slip conditions are

. u|U
n = 0,
. utang |U
n = 0.

So the velocity follows no-slip or stress-free (free-slip) boundary conditions. Temperature is specified on the
boundary, or heat flux
. T |U
n

is specified through the boundary.

3 Rayleigh-Benard Convection

Fluid sits between two flat plates which are held at different temperatures. Both no-slip and stress-free
boundary conditions are interesting to study (or some combination). Everything is assumed to be periodic
in x and y with periods Lx and Ly .

z
T = Ttop

g , , , w=0

x
T = Tbottom

Figure 3.1: Setup for Rayleigh-Benard convection.

Math 654, Lecture 11 2/14/11 5

For all parameter values, there exists an exact solution:

u 0,
z
T = Tbot + (Ttop Tbot ) ,
 h 
1 z2
p = g (Ttop Tbot ) z .
2 h

This is know as the conduction solution. Note that for some choices of parameter values, this can be proven
to be the unique long-time solution. And then it is globally stable.
Lets count the parameters in the problem:

, , , , g, h, Lx , Ly , Ttop , Tbot

make up 10 parameters. But note that these parameters are not really all indepedent, at least from a
mathematical point of view. We can pass to non-dimensional variables. We can rescale time and space (and
consequently velocity), temperature, and pressure. In so doing, we can non-dimensionalize the dynamics. So
first, define a new time variable by
t
t0 = 2 ,
h
h2
motivated by the physical fact that time for heat to diffuse across layer. Also, define

1
(x0 , y 0 , z 0 ) = (x, y, z, ) ,
h
T Ttop
T0 = ,
Tbot Ttop Tbot Ttop
h
u0 = u, and

 2
h p
p0 = .

Then the problem now looks as follows:

z
T =0
z=1

z=0 x
T =1

Figure 3.2: Non-dimensionalized Rayleigh-Benard convection.

Math 654, Lecture 11 2/14/11 6

The equations of motion are

.u = 0,
(t + u.) u + p Ra T,
= Pr4u + kPr
(t + u.) T = 4T,

in the new variables. The dimensionless parameters are

Lx Ly
, , Pr, Ra
h h
with the following definitions.
Definition 2. The Prandtl number is

Pr = .

The Rayleigh number is
g (Ttop Tbot ) h3
Ra = .

In our new variables, the conductive solution is

u = 0,
T = 1 z,
 
1 2
p = z z Pr Ra.
2
L
Note that for given Lhx , hy , there is a critical Rac so that if Ra < Rac , the conduction solution is absolutely
stable. If Ra > Rac , then the conduction solution is unstable. But the Prandtl number does not play a role
in the stability, only in the ensuing fluid dynamics.
Math 654, Lecture 12 2/16/11 1

Lecture 12: Incompressible Navier-Stokes and Plane Coutte Flow

We added more physics (viscosity) to our fluid model. Then we started to boil down the equations to capture
a large variety of dynamics were interested in. The equations are now:

.u = 0,
(t u + u.u) + p = 4u + f ext .

The dynamics happen in some domain with a boundary . We wrote down boundary conditions last
. u| = 0,
time beyond u

no-slip: u specified on ;

stress-free: no stress at the walls.

Note that the incompressibility condition gives

0= (.u) dx = (u.
n) da,

an implicit condition any boundary conditions must satisfy.

1 Pressure
As with Euler, the pressure is determiend instantaneously everywhere by u. Taking the divergence of the
equations, we arrive at
4p = u. (u.u) + .f ext .
And the boundary conditions look like

. p| = n
n . (everything else)| .

So in principle, if were given at time t = 0 the velocity

u (x, 0) = u0 (x) ,

then (numerically) wed solve for the pressure and update forward in time.

2 Case Study: Plane Couette Flow

Consider fluid between two plates, periodic in the directions tangent to the plates. An exact solution is
y
u = iU ,
h
p = constant.

This is known as plane Couette flow.

Math 654, Lecture 12 2/16/11 2

y
u = iU at y = h
h

x
Lx

Lz
z
u = 0 at y = 0

Figure 2.1: Plane Couette flow.

In fact this is how a Couette viscometer measures viscosity of a fluid: the instrument measures how much
force/torque it takes to maintain steady flow in the given setup. Even though the viscosity does not show up
in the above solution, it is intrinsic to the equations. Now well consider various properties of this solution.

2.1 Vorticity

Recall that = u, so here

= k U .
h
The fluid elements are rolling and slipping over each other. (D2)

2.2 Shear Stress

By definition,

ij = pij + (ui,j uj,i )

 U
= = pI + ij + ji ,
h
so we have
p Uh 0
= Uh p 0 .
0 0 p
So here
U
= vertical flux of horizontal momentum.
h
What is the force felt by the top plate? The shear stress at the top plate is
U
12 |y=h = .
h
Math 654, Lecture 12 2/16/11 3

Consider that the force opposes the velocity of the top plate, which makes sense. So we see that
 
U U2
work/time done on fluid = AU = Lx Lz ,
h h

and also
U
force applied = F = Lx Lz .
h
So given such a flow, the applied force is proportional to the velocity of the top plate. This gives us a way to
measure viscosity. But what geometric/material conditions assure that (steady) plane Couette flow appears?
Under what conditions is plane Couette flow stable? Before we can answer this, we need to discuss energy.

2.3 Energy

Recall that the kinetic energy present in the fluid is

1
K (t) = kuk2 dx.
2

Supposing the plane Couette geometry and u (x, 0) = u0 (x), we have

d
K (t) = u. u dx
dt t

= u. (u.u p + 4u) dx.

Lets consider each term here individually. Since the flow is divergence free,

u.p dx = (u.p) dx


= pu.n da

= 0

due to the periodicity. Now the advection term:

ui (uj j ui ) dx = ui (u.ui ) dx

 
1
= u. ui ui dx
2
 
1
= . u kuk2 dx
2

1
= kuk2 u.n da
2
= 0
Math 654, Lecture 12 2/16/11 4

for the same reasons. What about the dissapative term? Consider that

u.4u dx = ui 4ui dx


= ui . (ui ) dx

= [. (ui ui ) (ui ) . (ui )] dx

= .ui da
ui n ui,j ui,j dx


= U u dxdz kuk2 dx.
top y

So the rate of change of total kinetic energy is

d u
K (t) = U dxdz kuk2 dxdydz.
dt top y

The first term here is

u
U dxdz = U 12 dxdz = U F,
top y top

so the first term is the power being put into the flow. The second term is

kuk2 dxdydz

which is negative definite, and is the power dissapated in the flow (by heat).1
Lets check this power balance for plane Couette flow:
 
dK U U2
=U Lx Lz 2 hLx Lz = 0.
dt h h

Now we can ask if plane Couette flow is stable, and if so, how quickly will arbitrary flows converge to it?

2.4 Stability of Plane Couette Flow

Consider starting in the plane Couette geometry with a general initial condition, i.e., u (x, 0) = u0 (x).
Suppose there exists a unique solution u (x, t). Lets compare this with the plane Couette solution
y
uPC (x) = iU .
h
Define
v (x, t) = u (x, t) uPC (x) .
Note that
v|y=0 = 0 = v|y=h
1 So an experimentalist would worry about removing the extra heat through the plates.
Math 654, Lecture 12 2/16/11 5

as the solutions match the same boundary conditions. If we can conclude that v 0 in some sense, then
we can conclude that uPC is absolutely stable.
Since u = uPC + v, Navier-stokes tells us that






t uPC + v + uPC + v . uPC + v + p = 4 uPC + v .

Carrying this out, we have

   
y U
t v + U x v + v. ji + v.v + p = 4v.
h h
Then dividing by yields
U U
t v + v.v + x v + ivy + p = 4v,
h h
where weve written v = ivx + jvy + kv
z . And we have the boundary conditions from above.

Now we make a definition:

Definition 1. The energy in the perturbation is

1
E (t) = kv (x, t)k2 dxdydz.
2

Note that E (t) 0 at all time. If E(t) 0 as t 0, then v 0 with respect to L2 () and u uPC with
respect to L2 (). So we look for conditions yielding this result.
We have

d
E (t) = v.t v dx
dt
 
U U
= v. v.v x v ivy p + 4v dxdydz.
h h

After distributing (in the order above), the first term and the fourth term are zero exactly by the same
reasons as before. What about the second term? Observe that
 
1
v.x v = x kvk2 = 0
2
by the fundamental theorem of calculus and the periodic boundary conditions. So we have

d U
E (t) = vx vy dxdydz + v. (4v) dxdydz
dt h

U
= vx vy dxdydz kvk2 dxdydz.
h | {z }
vi,j vi,j

The second term is 0. But the first term is indefinite. So perhaps if the viscosity is large enough, then
the second term here dominates the first. Next time well derive conditions on to ensure that the energy
decreases.
Math 654, Lecture 13 2/18/11 1

Lecture 13: Plane Couette Flow (ctd.)

The equations are now
p
t u + u.u + = 4u,

.u = 0.

1 Plane Couette Flow

Lets recall what we did last time. We were considering the following setup:

y
u = iU at y = h
h

x
Lx

Lz
z
u = 0 at y = 0

Figure 1.1: Setup for plane Couette flow.

We exhibited a steady solution,

y
uPC = iU ,
h
pPC = constant,

called plane Couette flow. The force required on the top plate to maintain steady flow is
U A
F = Lx Lz 21 =
h
and so requires a power
U 2 A
F U = .
h
The energy dissipation per unit mass is

U 2 A 1 U2
= = 2.
h Ah h
Math 654, Lecture 13 2/18/11 2

1.1 Stability of Plane Couette

A major question we studied last time concerned the stablity of this solution. So we considered an arbitrary
initial condition u0 (x) and subesquent solution u (x, t). We defined
v (x, t) = u (x, t) uPC .
Setting u = uPC + v into the equations left us with
y U 1
t v + v.v + U x v + i vy + p = 4v,
h h
.v = 0.
Similarly we obtained boundary conditions
v|y=0 = 0 and v|y=h = 0.

To check stability, we considered the evolution of

1 1
E (t) = kv (x, t)k2 dxdydz = kv (, t)k22 .
2 2
We arrived at  
d U
E (t) = kvk2 + vx vy dxdydz.
dt h
So we ask: what explicit conditions will lead to the integrand being positive? These conditions will yield
stability of the plane Couette solution. But first we need a mathematical result.

1.2 Poincares Inequality

Theorem 1. (Poincare) Suppose f (y) is a (smooth enough) function on y [0, h] and f (0) = 0 = f (h).
Then h
2 2 2 h 0 2 2
kf k2 = (f (y)) dy 2 (f (y)) dy = 2 kf 0 k22 .
0 h 0 h

Proof. Write

X  ny 
f (y) = fn sin with
n=1
h
h  ny 
2
fn = f (y) sin dy.
h 0 h
Recall Parsevals theorem states

hX 2
kf k22 = f .
2 n=1 n
Assuming f 0 is nice enough to have its own Fourier representation,

X n  ny 
f 0 (y) = fn cos .
n=1
h h
Math 654, Lecture 13 2/18/11 3

Then calculate
h
2
kf 0 k22 = (f 0 (y)) dy
0

h X n2 2 2
= f
2 n=1 h2 n

h X 2 2
f
2 n=1 h2 n

!
2 hX 2
= f
h2 2 n=1 n
2
= kf k22 ,
h2
which is what we wanted to show.

This can be generalized.

Theorem 2. Suppose f (x) is a function on a domain satisfying f | = 0. Then

kf k22 0 kf k22

where 0 is the smallest eigenvalue of 4 on with homogenous Dirichlet boundary conditions.

In our case we have functions f (x, y, z) which are periodic in x, z but not y. So we write

X
X
X  ny  mx lz
f (x, y, z) = fn,m,l sin ei2 Lx ei2 Lz
n=1 m= l=
h

with fn,m,l being complex in general. Explicitly,

Lx h Lz  ny 
2 2imx 2ilz
fn,m,l = dx dy dz sin e Lx e Lz f (x, y, z) .
hLx Lz 0 0 0 h

By Parseval we have
Lx h Lz X
X
h 2
kf k22 = dx dy dz = Lx Lz |fn,m,l | .
0 0 0 2 n=1 m,l=

Here, Poincares inequality becomes

f 2 2
k k2 2 kf k22 .
y h
This analysis will apply to vx (x, y, z, t) , vy (x, y, z, t) , and vz (x, y, z, t).
Math 654, Lecture 13 2/18/11 4

1.3 Energy of Perturbation

Now we continue tracking the evolution of the perturbation. So write

   
d 1 2 2 U
kvk2 = kvk + vx vy .
dt 2 h
Then by Poincares inequality,
vx 2 vy 2 vz 2
kvk22 = k k +k k +k k
x 2 x 2 x 2
vx 2 vy 2 vz 2
+k k +k k +k k
y 2 y 2 y 2
vx 2 vy 2 vz 2
+k k +k k +k k
z 2 z 2 z 2
vx 2 vy 2
k k +k k
y 2 y 2
2 2
2
kvx k22 + 2 kvy k22 .
h h
Its a fact that
1 2 c 2
ab a + b
2c 2
for any 0 < c < .1 Therefore,
   
  1 2 1 2 1

 vx vy dxdydz  v + v dxdydz = vx2 + vy2 .
  2 x 2 y 2
So,
2

kvk22 2
kvx k22 + kvy k22 ,
h
and we have arrived at  
2 2  
.
kvk22 2  v v
x y dxdydz 
h

Setting this back into our energy estimate, we have

 
d 1 2 U
kvk2 = kvk22 vx vy
dt 2 h
 
U  
kvk22 +  vx vy 
h  
U h2
kvk22 + kvk22
h 2 2
and thus    
d 1 Uh
kvk22 2 kvk22 .
dt 2 2

a 2
1 This follows from noting that 0
c
cb .
Math 654, Lecture 13 2/18/11 5

So we have the following result:  

d 1
kvk22 0
dt 2
if
Uh
> ,
2 2
i.e., if
Uh
< 2 2 20.

Definition. The nondimensional parameter
Uh
Re =

is called the Reynolds number.
Note. The Reynolds number is often reffered to as (inertial forces)/(damping forces).
Weve shown that the energy cannot increase if the Reynolds number is less than 2 2 . Now lets show that
this is sufficient to ensure stability. Rearranging things and reapplying Poincares inequality, we have
   
d 1 U  
kvk22
kvk2 +  vx vy 
2
dt 2 h
2 U1

2 kvk22 + kvx k22 + kvy k22
 2h  h 2
U
2 kvk22 .
h 2h
Multiplying through by two, we arrive at
 
d 2 2 U
kvk22 kvk22 .
dt h2 h
| {z }
r
2
Note that r > 0 when Re < 2 . To show exponential decay of perturbations, we need one more result
Gronwalls inequality.
Theorem 3. (Gronwall) Given f (t) with |f (0)| < and
df
rf,
dt
then
f (t) f (0) ert .

Proof. Write
df
rf
 dt

df
+ rf 0
dt
d rt

ert e f (t) 0
dt
Math 654, Lecture 13 2/18/11 6

and conclude
ert f (t) f (0) ,
which is what we wanted to show.

Applying this here, we get

kv (, t)k22 kv (, 0)k22 ert .
So weve proved a theorem.

Theorem 4. Plane Couette flow is absolutely stable (in L2 sense) for

Uh
Re = < 2 2 .

Remark. Its a fact that plane Couette flow is absolutely stable if Re < 88.

How would we obtain the sharper result? Write

U

!
d 2 kvk2 + h vx vy
kvk22 = kvk22
dt kvk22

2 U


2 kwk + h wx wy 2
min kvk2 , (1.1)
.w=0 kwk22
w|y=0 =0= w|y=h

then proceed by variational calculus to calculate this minimum (which is just the norm of an operator). The
number that results is exactly 88.

1.4 Energy Dissipation

We wrote
energy dissipation rate
=
unit mass
and we had
U2
PC = .
h2
As per convention, consider2
h
(Re) = = function of Re.
U3
This quantity is called the dissipation factor or friction factor. (In general, this would be a function of
the aspect ratio of the box, but were suppressing that here.) For plane Couette flow,

PC h U 2 h 1
PC = = = .
U3 h2 U 3 Re
1
T
2 Here, f = limT T 0 f (t) dt.
Math 654, Lecture 13 2/18/11 7

Re = 88
Re
1 10 100 1000 104 105

0.1
absolutely
stable

0.01

0.001

10!4

10!5 PC (Re)

Figure 1.2: PC vs. Re.

We showed today that if Re < 88, then PC for arbitrary initial conditions. Thus plane Couette flow
is absolutely stable for Re < 88. Next time well investigate Re 88.
Math 654, Lecture 14 2/21/11 1

Lecture 14: Plane Couette Flow (ctd.)

Recall we considered stability of plane Couette flow,

y
uPC = iU ,
h
p = constant.

Taking a perturbation, we considered the energy density of the difference,

= kvk22 .
||

We introduced the Reynolds number,

Uh
Re = ,

and wrote
U2 1
PC = 2
and PC = .
h Re
In our analysis we found plane Couette flow to be absolutely stable for Re < 88. All this is depicted below.

Re = 88
Re
1 10 100 1000 104 105

0.1
absolutely
stable

0.01

0.001

10!4

10!5 PC (Re)

Figure 0.1: PC vs. Re.

Math 654, Lecture 14 2/21/11 2

1 Energy in the Large

Consider a general solution u (x, t) starting from u0 (x). Then consider the energy


 (t) = ku (x, y, z, t)k2 dxdydz.
hLx Lz

Let
y
u = iU + v (x, y, z, t)
h

be a perturbation to plane Couette flow, where

v|y=0 = 0 = v|y=h .

Then

U
u = ji + v
h
 2  2  2  2  2  2
vx vy vz U vx vy vz
= kuk2 = + + + + + + + ...
x x x h y y z
U2 U vx
= kvk2 + 2 + 2 .
h h y

Integrating yields

Lx h Lz
U2 U vx
kvk22 = kvk22 + hLx Lz 2 + 2 dx dy dz
h h 0 0 0 y
2
U
= kvk22 + hLx Lz 2 .
h

So with a perturbation,

U2 U2
 (t) = 2
+ kvk22 2 ,
h hLx Lz h

and for long-time averages,

U2 U2 h 1
 2
= 2
3 = .
h h U Re

Thus for high Re, is at least as large as PC = Re1 .

Math 654, Lecture 14 2/21/11 3

Re = 88
Re
1 10 100 1000 104 105

0.1 PC
absolutely
stable

0.01

0.001

10!4

10!5 PC (Re)

Figure 1.1: PC vs. Re.

2 Stability Analysis

Suppose u0 (x) is an exact steady solution of Navier-Stokes (or any dynamical system in general).

Definition 1. We say that u0 (x) is absolutely stable (globally stable) if starting from any initial condition,

u (x, t) u0 (x)

as t .

Definition 2. We say u0 (x) is nonlinearly stable if there exists > 0 so that

ku (x, 0) u0 (x)k < = u (x, t) u0 (x)

as t .

Definition 3. We say u0 (x) is linearly stable if infinitesimal perturbations decay.

Remark. This last notion of stability is quite weak. So we use it in general to test unstability.

Definition 4. We say u0 (x) is unstable if it is not linearly stable.

Math 654, Lecture 14 2/21/11 4

2.1 Plane Parallel Flows

We want to analyze stability of a more general class of solutions to Navier-Stokes. So we make a definition.
Definition 5. A plane-parallel flow is one in which

u0 (x) = iU (y) .

These flows solve

t u + u.u + p = 4u + if (y) .
Note that at time t = 0, we can write
0 = iU 00 (y) + if (y) ,
so the initial condition arises via
1
U 00 (y) = f (y) .

As usual, write
u (x, t) = u0 (x) + v (x, t)
with u0 = iU (y) and v small. So the dynamics become

t v + v.v + iU 0 (y) vy + U (y) x v + p = 4v + iU 00 (y) + if (y)

= 4v.

After linearizing (for small v) we have

t v + iU 0 (y) vy + U (y) x v + p = 4v,

a linear homogenous equation for the perturbation.

Now we look for solutions of the form v (x, t) = et v (x).1 Plugging this in yields

v (x) = 4v + iU 0 (y) vy + U (y) x v + p ,

.v (x) = 0,
v | = 0,
p (x, t) = p (x) et .

We seek a solution to this system if any < {} < 0, well conclude linear instability.
W.l.o.g., say v (x) has exp (ikx x + ikz z) x, z dependance,2 with
2 2
kx = n and kz = m < n, m < .
Lx Lz
So we write
v (x) = iu (y, ) eikx x+ikz z + jv (y, ) eikx x+ikz z + kw
(y, ) eikx x+ikz z .
1 Time-translation invariance allows us to take the Laplace transform in time. (??)
2 Due to spatial invariance (periodicity) in x, z.
Math 654, Lecture 14 2/21/11 5

The divergence-free condition becomes

0 = ikx u (y) + v 0 (y) + ikz w (y) ,

and the pressure satisfies

 
0 0 2
0 = ikx U (y) v + U (y) ikx v + kx2 2
+ 2 kz p.
y

The pressure is thus determined by

 
d2 2 2 0
+ kx + ky p = 2ikx U (y) v.
dy 2

The components of the perturbation satisfy the linearized equations:

 
d2
0 = u + kx2
2
+ 2 kz u U 0 (y) v U (y) ikx u ikx p (2.1)
dy
 
2 d2 2 dp
0 = v + kx + 2 kz v U (y) ikx v (2.2)
dy dy
 
2 d2 2
0 = w + kx + 2 kz w U (y) ikx w ikz p. (2.3)
dy

Writing the pressure equation similarly, we have

 
d2
0 = 2ikx U 0 (y) v + k 2
x k 2
z p. (2.4)
dy 2

Currently we have a three-dimensional system with a divergence-free constraint. Next we aim to turn this
into a two-dimensional system with a divergence-free constraint. This will allow us to use a stream-function
approach.

2.2 Squires Transformation

Define k 2 = kx2 + kz2 , u

=kx u+kz w
k , and p = kkpx . Here we assume kx 6= 0 (the dynamics are simplified already
if kx = 0). Then enforcing kx (2.1)+k
k
z (2.3)
yields
 
d2 2 kx
0 =
u+ k U 0 (y) v ikx U (y) u
u ikx p.
dy 2 k

Multiplying through by k/kx gives

    
k d2 k
0= u
+ k2 U 0 (y) v ikU (y) u
u ik p.
kx dy 2 kx
k
So now we have equations for kx u
, v, p.
Math 654, Lecture 14 2/21/11 6

We want to renormalize our equations. By rearranging the previous equation and by multiplying (2.2) and
(2.4) by k/kx , we arrive at the system
    2 
k k d 2
0 = u
(y) + k u (y) U 0 (y) v (y) ikU (y) u
(y) ik p (y)
kx kx dy 2
    2 
k k d d
p
0 = v (y) + 2
k 2 v (y) ikU (y) v (y)
kx kx dy dy
 2 
d
0 = 2ikU 0 (y) v + k 2 p (y) .
dy 2

= k/kx and = k/kx . We have now recast the dynamics in terms of u

We can relabel as (y) , v (y) , p (y).
We have now a two-dimensional problem with a pressure induced by the divergence-free condition. Note
that we can write the divergence-free condition in the new variables:
dv
0 = ikx u + + ikz w
dy
dv
= 0 = ik
u+ .
dy
As we expect, this looks like a two-dimensional condition.
Now suppose there exists c so that < c = < ( ) < 0. Then as = k/kx > , we immediately
conclude < () < 0. So then there will exist c so that < c yields instability. And the three-dimensional
instability will be worse than the two-dimensional instability!
Next time well eliminate the pressure and take a stream-function approach.
Math 654, Lecture 15 2/23/11 1

Lecture 15: Plane Couette Stability Reduction to 1-D and Non-Normal Growth
Recall what we did last time we considered stability of plane parallel flows, which satisfy
.u = 0,
t u + u.u + p = 4u + f ,
and which arise via
u0 = iU (y) ,
p = constant.

y
u = iUtop
h

0 x
Lx
0 u = iUbot.

Figure 0.1: Plane parallel flow.

Note that the initial condition is determined by the forcing term, via
U 00 (y) + f (y) = 0
subjected to
U (0) = Ubot and U (h) = Utop .
To determine stability (rather, instability) of the plane parallel flows, we look for linear instability. We set
u (x, t) = u0 + v (x, t)
and after linearizing the dynamics for v we get
t v + v.u0 + u0 .v + p = 4v.
We looked for solutions of the form
 
v = iu (y) + jv (y) + kw
(y) et eikx x eikz z

with
2n 2m
kx = and kz = < n, m < .
Lx Lz
This gave us also
p = p (y) et eikx x eikz z .
Math 654, Lecture 15 2/23/11 2

1 Linear Stability of Plane Parallel Flows

We continue with the equations
 
0 d2
u + U (y) v + ikx U (y) u + ikx p = 2
kx2
+ 2 kz u (1.1)
d
 
d 2 d2 2
v + ikx U (y) v + p = kx + 2 kz v (1.2)
dy dy
 
d2
w + ikx U (y) w + ikz p = kx2 + 2 kz 2 w (1.3)
dy
dv
ikx u + + ikz w = 0
dy
subject to
u (0) = 0 = u (h)
v (0) = 0 = v (h)
w (0) = 0 = w (h) .

First, consider kx = 0 = kz . Then the equations become

d2
u + U 0 (y) v = u
dy 2
d d2
v + p = 2v
dy dy
d2
w = 2w
dy
dv
ikx u + + ikz w = 0,
dy
and hence the solution is
ny
u (y) = sin n = 1, 2, 3, . . . .
h
So here,
n2 2
= >0
h2
and there is no instability.
Next consider the case kx = 0 but kz 6= 0. Then the equations become
 2 
0 d 2
u + U (y) v = kz u
dy 2
 2 
dp d 2
v + = kz v
dy dy 2
 2 
d 2
w + ikz p = kz w
dy 2
dv
+ ikz w = 0.
dy
Math 654, Lecture 15 2/23/11 3

If w = 0, then v = constant = 0, so
d2

2
u = kz2 u
dy
and hence
l2 2
= + kz2 > 0 l = 1, 2, . . . .
h2
If w 6= 0, take p = constant = 0 w.l.o.g., then

d2

2
w = kz2 w
dy

which also gives > 0.

Next, we consider the most general case.

2 Reduction of Dynamics to 1-D

Consider kx 6= 0 (kz arbitrary)? W.l.o.g., take kx > 0. Then we use Squires tranformation. So we define

kx u + kz w p
u
= with k = kx2 + kz2
k
k
p = p,
kx
 
kx x(1.1)+kz x(1.3)
then we take k then multiply the result by kkx to get
    2 
k 0 k d 2
u
+ U (y) v + ikU (y) u
+ ik p = k u (2.1)
kx kx dy 2
 
and we multiply (1.3) by kkx as well to get
    
k d
p k d2 2
v + ikU (y) v + = k v. (2.2)
kx dy kx dy 2

The incompressibility condition is now

dv
0 = ik
u+ .
dy
So weve reduced to a two-dimensional system. Lets now introduce a stream-function and reduce to a
one-dimensional system.
Define a stream-function by
ik (y) = v (y) ,
which implies
d
u
= = 0 (y)
dy
Math 654, Lecture 15 2/23/11 4

= k and = k . Taking
by incompressibility. And write now d
ik(2.2) yields
kx kx dy (2.1)
 


d2

00 + k 2 + ikU 00 + ikU (y) 00 + k 2 =
k2 00 + k 2
dy 2
subject to the boundary conditions

(0) = 0 = (h) and

0
(0) = 0 = 0 (h) .

This is known as the Orr-Sommerfeld equation.

For plane Couette flows, we can simplify this equation even further. For these flows, we have
y
U (y) = U
h
so
U
U 0 (y) = and U 00 (y) = 0,
h
and so the equation becomes
 
(y) + ikU d2 2
f yf (y) = k f (y) with
h dy 2
f (y) = 00 (y) + (y) .
 
This is a complex version of the Airy equation. In 1973, Ramonov proved that < > 0 for all .
So
there is no linear instability for plane Couette flows! But how does this translate back to the non-linear
case?

3 Non-Normal Growth of Solutions to ODE

Example 1. Consider the system

X (t) = X + Y + . . .
Y (t) = Y + . . .

with > 0. There is a unique fixed point at (X, Y ) = (0, 0). Look for a solution et and investigate time
dependence. Doing so, we arrive at the eigenvalue/eigenvector problem
    
X 1 1 X
= .
Y 0 Y

There is a solution if  
1 + 1
0 = det = ( 1) ( ) ,
0 +
so

= 1 > 0 and
= > 0.
Math 654, Lecture 15 2/23/11 5

The eigenvectors are

 
1
=1: and
0
 
1
=: .
1

Supposing 0 < < 1, we have the following flow:

Figure 3.1: Phase portrait with eigendirections.

The general solution is not hard to write down:

     
X (t) 1 1
= Aet + Bet .
Y (t) 0 1
 
1 2 2
So if we define an energy by E (t) = 2 X (t) + Y (t) , we see it evolves as

d
E = X X + Y Y
dt
= X (X + Y ) + Y (Y )
= X 2 + XY Y 2
"
#
2 X 2 XY + Y 2
= E.
X2 + Y 2

Then we may write " #

dE 2 X
X Y + Y 2
2 min E (t) .
dt Y
X, X 2 + Y 2
Math 654, Lecture 15 2/23/11 6

If this max is positive, then Grenwalls inequality tells us solutions decay immediately. (??) Define to be
this maximum, then we can write
  

1 21
X
= min
X Y ,
2 +Y 2 =1
X 12
Y
| {z }
M

so is the smallest eigenvalue of M . The eigenvalues are

1 p 
= 1 + 2 2 + 2 ,
2

and in particular when = 41 , = 0. As 0, < 0, so becomes negative. So for small enough , the
energy in the perturbation can grow before it decays.

X(t), Y (t)

Figure 3.2: Non-normal growth.

If this were a non-linear problem, and if the energy grows too quickly, then the non-linear solution could
become unstable (even though the linear problem is stable). This is known as non-normal growth.

Well take the same approach for plane Couette flow. Now,

U y
t v + i vy + U x v + p = 4v,
h h
.v = 0.

Lets look for a solution with x 0. So write

v = ivx (y, z, t) + jvy (y, z, t) + kv

z (y, z, t) ,

and enforce
vy vz
0= + .
y z
Math 654, Lecture 15 2/23/11 7

In this case,
 
U 2 2
t vx + vy = + 2 vx
h y 2 z
 2 
p 2
t vy + = + 2 vy
y y 2 z
 2 
p 2
t vz + = + 2 vz .
z y 2 z

Now lets look for a solution of the form

 y 
vx = Vx (t) sin (kz) sin
h
 y
vy = Vy (t) sin (kz) sin
h 
 y
vz = Vz (t) cos (kz) cos .
h
Note. Were cheating a little here! The z-equation certainly does not satisfy the correct boundary conditions.
This is for illustrative purposes; if we were to require the correct boundary conditions, wed have sinhs and
coshs instead. So a similar analysis applies.

Continuing on, we write Vz (t) = hk Vy (t) and then we have the system
 2 
U 2
Vx + Vy = + k Vx
h h2
 2 
2
Vy = + k Vy .
h2

A particular solution having Vx (0) = 0 and Vy (0) 6= 0 is

Vy (t) = Vy (0) es
Vx (t) = Vy (0) ses ,

where
 
2 2
s = + k t,
h2
Re
= , and
+ h2 k 2
2

Uh
Re = .

What is the energy here? We have
1
1 2 2
2
E (t) = Vx2 + Vy2 = s + 1 es (Vy (0)) .
2 2
Just as in the example, we have the following diagram:
Math 654, Lecture 15 2/23/11 8

E(t)

E(0) >1

t
t

Figure 3.3: Non-normal growth in plane Couette flow.

The critical value t comes from

r
1
s = 1 + 1 2 .

And the maximum energy is
 r 
1 1 1 12
q
Emax = E (0) 2 1 + 1 2 e 2E (0) e2 2

asymptotically as . Observe that Re, so Emax Re2 . In practice, its quite easy to obtain very
high Reynolds numbers. And small perturbations in those flows can lead to instabilities and turbulence,
even though the linear problem is stable.

Example 2. As a first step towards pde, and towards understanding the result we just saw, consider the
following system:

X 1 = X1 + aX2
X 2 = X2 + aX3
X 3 = X3 + aX4
X 4 = X4 + . . .
..
.

As a matrix equation, this is

1 a 0 0 0
X1 X1
0 1 a 0 0
d X2
0 0 1 a 0

X2

X3 = X3 .
dt 0 0 0 1 a
.. ..
. .. .. .. .. .. .
. . . . .
Math 654, Lecture 15 2/23/11 9

All eigenvalues are 1 (specifically, any truncation has all eigenvalues 1). A general solution is

X m
(at)
Xn (t) = et Xn+m (0) .
m=0
m!

Suppose we take Xn (0) = rm with r < 1, then

Xn (t) = rn e(ra1)t

if a > 1r . This is incredibly degenerate non-normal growth!

Math 654, Lecture 16 3/14/11 1

Lecture 16: Plane Couette Flow (ctd.)

Recall we had been studying plane Couette flow,

Uy
uPC = i ,
h
pPC = constant,

a solution to incompressible Navier-Stokes in the following setup:

y
u = iU at y = h
h

x
Lx

Lz
z
u = 0 at y = 0

Figure 0.1: Setup for plane Couette flow.

We defined the energy dissipation rate as a space-time average, i.e.,


 = kuk2 .

In the plane Couette case,

U2
PC = and
h2
PC h 1
PC = = = .
U3 Uh Re

We found that plane Couette flow is the unique long-time solution in this setup for sufficiently small Re; we
also showed PC in general. All this is indicated on the following Re diagram.
Math 654, Lecture 16 3/14/11 2

Re = 88
Re
1 10 100 1000 104 105

0.1 PC
absolutely
stable

0.01

0.001

10!4

10!5 PC (Re)

Figure 0.2: vs. Re.

Well continue to investigate this diagram today.

1 An Upper Bound on Dissipation

Again, we have the Navier-Stokes equations with the plane Couette geometry/boundary conditions,

t u + u.u + p = 4u
.u = 0.

Well proceed as did E. Hopf in 1941: decompose the flow as

u (x, t) = i (y) + v (x, t)

where is a background flow which satisfies the inhomogenous boundary conditions,

(0) = 0 and (h) = U.

Math 654, Lecture 16 3/14/11 3

(y)

PC

y
h

Figure 1.1: E. Hopfs background flow.

So then
v (x, t)|y=0 = 0 = v (x, t)|y=h .
Having done this, we find

  .v = 0

t v + v.v + v. i (y) + i (y) .v + p = 4v + i00 (y) .
| {z } | {z }
(y)x v
i0 (y)vy

Consider the kinetic energy in v. So dot the second relation above with v and integrate over the space,
which yields
 
d 1 2
kvk2 + v. (v) .v dx + 0 (y) vx vy dx
dt 2
| box
{z } box
=0

+ (y) v. (x v) dx + v.p dx = kvk22 + vx 00 (y) dx.
| box {z } | box {z } box
=0 =0

We saw why the terms which dissapear evaluate to zero in previous lectures. By integrating by parts on the
last term we arrive at
 
d 1 vx
kvk22 + 0 (y) vx vy dx = kvk22 0 (y) dx. (1.1)
dt 2 y

Now observe that

u = ji0 (y) + v,
so then
h
2 vx
kuk22 = Lx Lz 0
( (y)) dy + 2 0 (y) dx + kvk22 .
0 y
Math 654, Lecture 16 3/14/11 4

Multiplying this relation by /2, we get

h
1 2 vx
kuk22 = Lx Lz (0 (y)) dy + kvk22 + 0 (y) dx. (1.2)
2 2 0 2 y

Adding (1.1) and (1.2) yields

  h
d 1 1 2 1
kvk22 + kuk22 = Lx Lz (0 (y)) dy kvk22 0 (y) vx vy dx.
dt 2 2 2 0 2

Then we multiply by two and write

h
d 2  
kvk22 + kuk22 = Lx Lz (0 (y)) dy kvk22 + 20 (y) vx vy dx.
dt 0

We have yet to choose . Our goal is the following: Choose (y) so that c > 0 with

 
kvk22 + 20 (y) vx vy dx ckvk22

for all divergence free v which also satisfy the homogeneous boundary conditions.

(y)

boundary
layers

y
h

Figure 1.2: A candidate .

If we can choose such an , then we could write

h
d d 2
kvk22 kvk22 + kuk22 Lx Lz (0 (y)) dy ckvk22 .
dt dt 0
| {z }
A

Call f (t) = kv (, t)k22 , then we have

df
A c f (t) .
dt
Math 654, Lecture 16 3/14/11 5

By the general Gronwalls inequality (proved in the homework) we conclude

(1 ect )
f (t) f (0) ect + A .
c

f (t)

f (0)

A/c

f (0)

Figure 1.3: Energy bound on f (t) = kv (, t)k22 .

So if such an exists, then well have found that the energy in the perturbation is uniformly bounded in long
T
time. And then we can do the following. Take the time average ( T1 0 [] dt) of the differential inequality
h h
d 2 2
kvk22 + kuk22 Lx Lz 0
( (y)) dy ckvk22 Lx Lz (0 (y)) dy
dt 0 0

to get
T h
1 1 1 2
kv (, T )k22 kv (, 0)k22 + ku (, t)k22 dt Lx Lz (0 (y)) dy.
T T T 0 0

Sending T and dividing the result by Lx Lz h we conclude

h
1 2
 (0 (y)) dy,
h 0

thereby bounding the energy dissipation rate. This all hinges on finding an acceptable , which well do
now.

So we want to see that

kvk2 + 20 (y) vx vy dx ckvk22

for some c > 0. Try

Uy

2 0y
U
(y) = y h .
2
U (yh)
2 h y h
Math 654, Lecture 16 3/14/11 6

(y)

y
h h

Figure 1.4: One-parameter family of s.

Note that
U
2 0y<
0 (y) = 0 <y <h ,

U
2 h <y h
so h
2 U2 U2 U2
(0 (y)) dy = + = .
0 4 2 4 2 2
Wed like to get as small as possible to produce as sharp as possible a bound on . First we need a lemma.

Lemma. (Howard) Suppose f : R R satisfies f (0) = 0 = f (L) and

L
2
kf 0 k22 = (f 0 (x)) dx < .
0

Then,
2 1 0 2
|f (x)| x (L x) kf k2 .
L

Proof. By the fundamental theorem of calculus and the Cauchy-Schwarz inequality, we may write
 x 
 
|f (x)| =   1 f () d 
0
0
 x  21  x  12
2
12 d (f 0 ()) d ,
0 0
| {z }

x

and hence
2 x
|f (x)| 2
(f 0 ()) d.
x 0
Math 654, Lecture 16 3/14/11 7

On the other hand,

 
 L 
 0 
|f (x)| =  f () d 
 x 
L !
1

0 2
(L x) 2
f () d ,
x

and hence
2 L
|f (x)| 2
(f 0 ()) d.
Lx x
Adding these up we get  
2 1 1
|f (x)| + kf 0 k22 ,
x Lx
or equivalently  
2 L
|f (x)| kf 0 k22 .
x (L x)
Hence the result.

Now well use the result. First write (as in the proof)
y
vx
vx (x, y, z) = (x, , z) d
0 y
 2 ! 21
vx
y (x, , z) d
0 y

for 0 y . A similar expression holds for vy , and hence

 2 ! 12  2 ! 21

vx vy
|vx vy (x, y, z)| y d d .
0 y 0 y

Now recall that  

1 1 2
ab a + b2 ,
2
so "  2  2 #

y vx 1 vy
|vx vy (x, y, z)| d + d .
2 0 y 0 y
Well finish this next time.
Math 654, Lecture 17 3/16/11 1

Lecture 17: Plane Couette Flow and RANS

Recall we were in the midst of a calculation which would give an upper bound on the dissipation rate in the
flow.

1 An Upper Bound on Dissipation (ctd.)

We decomposed the flow into the background flow (y) and the rest:
u (x, t) = i (y) + v (x, t)
with
.v = 0
and subject to boundary conditions
v|y=0 = 0 = v|y=h .
At this point, we know that if we can choose (y) so that



kvk2 + 20 (y) vx vy dx ckvk22 (1.1)
box

for some c > 0 and for all divergence-free vector fields v satisfying the homogeneous boundary conditions,
then we can conclude that the energy in v (and the total energy) is bounded for all time, and that


2
 1 h 2
 = kuk (0 (y)) dy.
h 0
We proposed trying of the form

Uy

2 0y
U
(y) = y h
2
U (yh)
2 h y h
so the background flow looks as follows.

(y)

y
h h

Figure 1.1: The proposed .

Math 654, Lecture 17 3/16/11 2

In this case, if the above requirement holds for some (0, h/2], then
U 2h
 .
2
Well pick up the analysis here today.
So write
 
 h vx 
 
|vx (x, y, z)| =  1 (x, , z) d 
 0 y 
 y  21 y  2 ! 21
2 vx
1 d (x, , z) d
0 0 y
! 21
y 2
vx
= y (x, , z) d
0 y

and similarly for vy . Then

2 ! 21 y 
y  2 ! 21
vx vy
|vx (x, y, z) vy (x, y, z)| y d d
0 y 0 y
"  2 y 2 #
y 1 y vx vy
d + d .
2 0 y 0 y

Now
    " h #
 h  U
 0 
2 (y) vx vy dydzdx 2 |vx vy | dydzdx + |vx vy | dydzdx
 0  2 0 h
"  2  2
U 2 1 vx vy
dydzdx + dydzdx
4 0 y 0 y
h  2 h  2 !#
1 vx vy
+ dydzdx + dydzdx
h y h y
 v 2 
U 1
vx 2 y
+
4 y 2 y 2
 v 2 
U
vx 2 y
+ 2 .
4 y 2 y 2
We showed in homework that if

.v = 0 and
v|y=0 = 0 = v|y=h

then v 2 v 2
x y
+ 2 kvk22 .
y 2 y 2
Math 654, Lecture 17 3/16/11 3

Now we take = 2 to acheive
 
 h  U
 
 0
2 (y) vx vy dydzdx kvk22
 0  4 2

and thus


U
kvk2 + 20 (y) vx vy dx kvk22 kvk22
4 2
box
 
U
= kvk22
4 2
  2
U
1 kvk22
4 2 h2
by Poincares inequality. So pick any satisfying

4 2
.
U

We have shown that for any satisfying the above requirement,

h
1 2
 (0 (y)) dy
h 0
 2
U 2
=
2 h
U 2
= .
2h
Once we take
4 2
= ,
U
we have acheived
U 2 U 1 U3
 =
2h4 2 8 2 h
and hence
h 1
= 0.088.
U3 8 2
Note. The above analysis required (0, h/2], so the resulting bound only holds if

4 2 h
,
U 2
i.e., if
Re 8 2 11.
Thus the upper bound is consistent with Re1 .
Math 654, Lecture 17 3/16/11 4

We have proved a general theorem.

Theorem. In the plane Couette flow setup,

1
.
Re

If also Re 8 2, then
1
.
8 2

Remark. Suppose we had chosen = h/2 instead, i.e., (y) = uPC y . Then so long as h/2 4 2/U , i.e.,

Re 8 2, the above calculations would yield Re1 . Motivated by this, one could compare the problem
posed at the start of this lecture (equation (1.1)) with the variational problem posed in Lecture 13 (equation
(1.1) in said lecture).

This was a quick calculation with a one parameter family of s which were easy to work with. One could pose
a variational problem to optimize the shape of and solve for the associated optimal bound computationally.
With this approach it was recently shown that

0.011.

The Re diagram below summarizes everything weve discussed regarding plane Couette flow.

Re = 88
Re
1 10 100 1000 104 105

1
0.1
absolutely 8 2
stable

0.01 0.011

0.001 Re1 0.011

10!4

10!5 PC (Re)

Figure 1.2: vs. Re.

Math 654, Lecture 17 3/16/11 5

This is as far as we can go with analysis alone. For large enough Re, experimental data fits to a curve

h 1
2.
U3 (log Re)

Next well explore a classical theory of turbulence which is capable of predicting such a law.

2 Turbulence: Reynolds Averaged Navier-Stokes (RANS)

At very large Reynolds number, turbulent flows are observed. Tracer particle seem to move about randomly.
Even so,

t u + u.u + p = 4u,
.u = 0.

The idea here is to understand

T
1
u (x, y, z) = lim u (x, y, z, t) dt,
T T 0

a long time average. So we aim to derive differential equations for average quantities.
What do we know so far? In the plane Couette setup,

u (x, 0, z) = 0 = u (x, h, z) .

And time-averaging the divergence-free condition yields

.u = 0.

We have left the momentum equation; time-averaging yields

t u + u.u + p = 4u.

Lets look at each term.

First, we have

1 T u
t u = lim (x, y, z, t) dt
T T 0 t
1
= lim (u (x, y, z, T ) u (x, y, z, t))
T T
= 0

so long as u is bounded pointwise. This is the first modeling assumption. So we arrive at the RANS
equations:

j (ui uj ) + i p = 4ui ,
i ui = 0.
Math 654, Lecture 17 3/16/11 6

Note that weve used the fact that

(u.u)i = uj j ui = j (uj ui )

via the divergence-free condition.

Now write
u (x, y, z, t) = u (x, y, z) + v (x, y, z, t)
with v the fluctuations about the mean flow. Then

v = 0,

which says that the mean fluctuation about the mean is zero, and

.v = 0

and
v|y=0 = 0 = v|y=h .
If we plug this into the full Navier-Stokes and average, we see

(u + v) . (u + v) + p = 4 (u + v)

and thus
u.u + u.v
| {z } + v.u
| {z } + v.v + p = 4u + 4v .
| {z }
=0 =0 =0

So we arrive at

u.u + p = 4u . (vv),
| {z }
j (vj vi )
.u = 0.

This is steady Navier-Stokes with an additional term, vv. So we make a definition.

Definition. The Reynolds stress tensor is

Reynolds = vv.

In components,
ij = vi vj .

With this we can write

u.u + p = 4u + . Reynolds .
Recall that
4u = . viscous .
So now it looks like we have an additional forcing term, due to the Reynolds stress. Recall that viscous
stresses arise due to random motion of molecules across planes. The Reynolds stress arises here due to
chaotic motion of fluid elements across planes. Next time well examine a model for the Reynolds stress.
Math 654, Lecture 18 3/18/11 1

Lecture 18: RANS, Turbulent Fluctuations, and Some Modeling Assumptions

We have been talking about turbulence in the context of plane Couette flow. The following -Re diagram
depicts three important flow regimes: laminar flow, transition to turbulence, and fully turbulent flow.

Re

Re0

laminar turbulent
transition

Figure 0.1: vs. Re.

The transition regime is (purposefully) badly defined in the figure. In reality, it depends on many factors
and changes shape/size based on these. To understand the turbulent regime, we started discussing RANS.
So we considered

1 T
u (x) = lim u (x, t) dt
T T 0

= iu + jv + kw.

For plane Couette flow,

u
dxdz = F ,
top y
and then we have
FU
= .
Lx Lz h
So the averaging procedure here is quite natural.
Math 654, Lecture 18 3/18/11 2

As we saw last time, averaging the Navier-Stokes equations yields

u..u + p = 4u.

Then one sets

u (x, t) = u (x) + v (x, t)
with

v = 0 and
v = 0.

Setting this into the averaged Navier-Stokes equations and averaging again yields the RANS equations:

u.u + p = 4u + Rey ,

with
Rey = vv
being the Reynolds stress. So
Rey
ij = vi vj .

Note that

kuk2 = ku + vk2
= kuk2 + 2u : v + kvk2 ,

and averaging yields

kuk2 = kuk2 + kvk2 .
Hence
 = mean + fluct .
Now well investigate the fluctuations in more depth.

1 Fluctuations
Take a change of variables u = u + v in the Navier-Stokes equations, which yields

t v + v.v + v.u + u.v + u.u + p = 4u + 4v.

And also we have

u.u + p = 4u + . Rey .
Subtracting the second from the first yields

t v + v.v + v.u + u.v + (p p) = 4v . Rey .

Now we want to look at energy. So dot v into this relation and integrate. This yields
 
d 1 2 2


kvk2 + v. (u) .v = kvk2 . Rey .v,
dt 2
Math 654, Lecture 18 3/18/11 3

and time-averging gives

 
d 1
kvk22 + v. (u) .v = kvk22 (. Rey ) .v.
dt 2
| {z } | {z }
=0 =0

Therefore, D E


fluct = kvk2 = h (u) : vvi = Rij ij
Rey
,

where  
1 uj ui
Rij = + .
2 xi xj

This is as far as we can go without making modeling assumptions. But wed like to have a way to solve
RANS forward in time. We need physical intuition to link Rey and u.

2 Three Modeling Assumptions

The first big assumption:
Rey R = mean rate of strain.
This is as in molecular dynamics, but as the scales are much larger here this is indeed a modeling assumption.

Definition. Given the assumption above, we call the constant of proportionality between Rey and R the
eddy viscosity, and write
Rey = e R.

Now we ask: what can e depend on? We know that

length2
[e ] = ,
time
and the fluctuations are driven by gradients in the mean velocity, which have

1
[u] .
time
It seems, at least in the plane Couette setup, that our length scale should be

length distance to the wall.

This line of reasoning is due to Prandtl and von Karman. As did they, we now make a second modeling
assumption:
2
e kuk (distance to wall) .

And finally we have a third modeling assumption: We assume the turbulent statistics reflect the maximal
symmetry of the problem.
Math 654, Lecture 18 3/18/11 4

In the plane Couette setup, this last assumption implies that all mean quantities can only depend on y. So
we can write1

u = iu + jv + |{z}

kw
|{z}
=0 =0

= iu (y) ,

and similarly

Rey = vv (y) and

p = p (y) .

Furthermore, by imposing the third assumption on RANS we get

2 Rey
0 = u (y) + (y) , (2.1)
y 2 y 21
p Rey
= (y) . (2.2)
y y 22

Now lets investigate the second assumption for plane Couette flow. Our intuition is that the turbulent flow
should look as follows:

u(y)

y
h

Figure 2.1: Expected turbulent Couette flow profile.

So lets assume u0 (y) 0. Well only consider the bottom half of the layer (by symmetry). Then the distance
to the wall (mixing length) is just y. Thus,

e = 2 u0 (y) y 2

where is the von Karman constant. Note that .4 as per experiments.

1 In particular, v = 0 follows from the divergence-free condition and boundary conditions on v.
Math 654, Lecture 18 3/18/11 5

Lets look at (2.1) and (2.2) more carefully. First, (2.2) says

d d Rey
p = (y)
dy dy 22
d 2
= (v2 ) ,
dy
and thus
2
p (y) = (v2 ) + const.

p(y)

h
y

Figure 2.2: Average pressure profile.

Second, (2.1) says

 
d2 d du
0 = 2 u (y) + e (y)
dy dy dy
"  2 #
2
d d du
= 2u + 2 y 2
dy dy dy
d h i
2
= u0 + 2 y 2 (u0 ) .
dy
So integrating from the boundary up yields

2 wall shear stress

u0 (y) + 2 y 2 (u0 (y)) = u0 (0) = = u2

once we define u by the relation
F
u2 = .
Lx Lz
Hence
2
2 y 2 (u0 (y)) + u0 (y) = u2
s 2
0 u2
= u (y) = 2 + .
22 y 2 2 y 2
Math 654, Lecture 18 3/18/11 6

Recall we expect u0 (y) 0, so s 2

u2
u0 (y) = 2 + + .
22 y 2 2 y 2
Now as u (0) = 0 and as (by symmetry) u (h/2) = U/2, we can determine u . After some work, one finds
s q

 2 2u y 2
2u y 2u y 1 1 +
u (y) = log + 1+ + 2u y

.
u

And demanding U/2 = u (h/2) yields

s q
2
 2
U u h u h 1 1 + u h
= log + 1+ +
u h

,
2u

an implicit equation for u in terms of known constants. Equivalently, one can write
"   # q

2
  r    
1 U u
u 2 1 1 + uU Re
= log Re + 1 + Re +

2 u U U uU Re

with
Uh
Re =

as usual. Note that if Re 0, then
U
= u2 ,
h
so then
U
= .
h
And if Re , then
2 U2
u2
4 (log Re)2
and then
h 2 1
.
U3 4 (log Re)2
2
This matches the experimental data with 0.4. In particular, experiments show (log Re) for large
enough Re.
Math 654, Lecture 19 3/21/11 1

Lecture 19: Kolmogorovs Theory and the Energy Cascade

Today well discuss a general theory of turbulence, attributed to Komogorov. Others associated with this are
Taylor and Richardson. The theory will describe homogenous, isotropic turbulence; the theory is a scaling
theory.

1 Kolmogorovs (Scaling) Theory of (Homogenous, Isotropic) Turbulence

Consider the Navier-Stokes equations without boundaries (i.e., in a periodic box):

t u + u.u + p = 4u + f (x, t)
.u = 0.

Figure 1.1: Turbulent flow in a box.

We can decompose a solution u into its Fourier modes with wave-numbers k as

X
u (x, t) = eik.x u
(k, t) .
k

The Fourier coeffcients satisfy

1
(k, t) =
u eik.x u (x, t) dx
L3 box

with
2
k= n, n = in1 + jn2 + kn
3, ni Z.
L
Math 654, Lecture 19 3/21/11 2

ky

kx

Figure 1.2: Fourier space.

In this decomposition, small scales in real space correspond to high kkk.

The kinetic energy satisfies

1
Kinetic Energy = ku (x, t)k2 dx
2

box
! !
1 X X
ik.x ik0 .x 0
= e (k, t) .
u e (k , t)
u dx
2 box k k0

1X
ei(k+k ).x dx
0
= u u (k0 , t)
(k, t) .
2 0
k,k | box {z }
L3 k+k0 ,0

L X 3
= (k, t) .
u u (k, t) .
2
k

(k, t) = u
It can be shown that u (k, t) , so

L3 X
Kinetic Energy = u (k, t)k2 .
k
2
k

u (k, t)k2 is sometimes called the spectrum of the flow u (x, t).
Note that k
Now we make a key assumption to this theory: We assume isotropy, i.e., we assume k u (k, )k2 depends only
3 3 3
on kkk. Noting k = 2/L and also that (k) = (2) /L = kx ky kz , we can write
 3
1 2
L3 X 2 3 L
k
u (k, )k dx = k
u (k, )k (k)
2 box 2 2
k
 3 X
M L 3
= ku (k, )k2 (k) .
2 2
k
Math 654, Lecture 19 3/21/11 3

The sum now looks like a Riemann integral, so (??)

 3
1 M L
u (k, )k2 dx
k = u (k, )k2 dk
k
2 box 2 2
 3
1 L
= M u (k, )k2 4k 2 dk.
k
2/L 2 2
| {z }
E(k)

Definition. The quantity

 3
1 L
E (k) = u (k, )k2 4k 2
k
2 2

is called the energy spectrum of the flow u (k, t).

Remark. In fact, E (k) is the energy spectrum per unit mass; one could think of E (k) dk as the energy per
unit mass contained in a shell of thickness dk about kkk.

ky

dk

!k!
kx

Figure 1.3: E (k) dk is energy/mass in shells.

Now we have a famous (dimensional) argument. First, consider the flow of energy in k-space as one stirs the
flow at medium-to-large scales. We know that

d  
kuk22 = kuk22 + u.f .
dt 2 | {z }
viscous dissipation | {z }
applied force

Secondary flow is generated from the large scales at sizes smaller than the original scales, resulting in the
well-known energy cascade of three-dimensional turbulence.
Math 654, Lecture 19 3/21/11 4

E(k) cascade

k
energy injected viscous regime
at large scales (heat)

Figure 1.4: The forward energy cascade.

If viscosity were zero, the cascade would go on forever. But note that viscosity has a large effect only at
large wavenumbers. To see why, consider that
X
u = ikeik.x u
(k, t) ,
k

and so

 = kuk2 dx

box

= k 2 E (k) dk.
box
 
So for small k, the dissipation is very small: k 2 = [dissipation], which is the rate of dissipation. (??) If
Navier-Stokes were linear, energy would stay in place; the non-linear term gives rise to the cascade.
Now well determine exactly how E depends on k. The power put into the flow by stirring is given by M L3 .
Since [] = L2 /T 3 , [k] = 1/L, and [E] = L3 /T 2 , a dimensional argument leads us to the well-known relation
2 5
E (k) = Ck  3 k 3 .

Here Ck is known as the Kolmogorov constant. In deriving this weve assumed:

Scales exist;

The energy cascade is driven by the non-linear terms;

A steady-state exists.
Math 654, Lecture 19 3/21/11 5

E(k)
energy in

L3
E(k) can only
depend on k

inertial range
k
2 2
L cascade independent
of viscosity
energy out

Figure 1.5: E (k) in the inertial range.

The quantity in the diagram above is the so-called dissipation lengthscale, and will be determined below.

For now, consider the quantity


U 2 = kuk2 .

We have
2

U2 = 2 E (k) dk
2
L
2
2 5
= 2 Ck  3 k 3 dk
2
L
"  23 #
3 2 L   32
= 2Ck  3 .
2 2 2

So if we neglect the contribution by the (small) term,

2
3Ck  3 2
U2 3 L3
(2) 2

and hence
U3
A
L
3/2
where A = 2/ (3Ck ) . But now recall that = L/U 3 is approximately constant in the turbulent regime!
So dissipation does not scale with the Reynolds number; it is instead universal.
Math 654, Lecture 19 3/21/11 6

Lets try to find . We have

2
2 5
 = k 2 Ck  3 k 3 dk
2
L
2

2 1
=  3 Ck k 2 dk.
2
L

Note that the main contribution here is from high k. Continuing on, we find

  43   43
3 2 2 2
 =  3 Ck


4 L
| {z }
0
3 4 2 4
= (2) 3 Ck  3 3 ,
4
and thus
  34   14
3 3
= 2 Ck .
4 
Now we have a closed picture.

E(k)
cascade

U3

L
E(k) = Ck 2/3 k 5/3

k
2 inertial range 2
L ! "1/4
3
= O(1)

Figure 1.6: Kolmogorovs theory of turbulence.

Math 654, Lecture 20 3/23/11 1

Lecture 20: Turbulent Energy Dissipation

The equations are

t u + u.u + p = 4u + f (x, t) ,
.u = 0.

We decomposed the flow as X

u (x, t) = eik.x u
(k, t)
k

with
1
(k, t) =
u eik.x u (x, t) dx.
L3
The setup here is periodic, boundaryless flow; a cube with side length L:

Figure 0.1: Turbulent flow in a box.

Denoting a time-space average by hi, we have two experimental laws of fully developed turbulence. The
first law is that
 2
ku (, ) u ( + l, )k2 = C (l) 3
where < l < L and viscosity does not matter in this regime. Dimensional analysis leads to this law as
well. The Fourier version of this, assuming homogeneity and isotropy as before, is
2 5
E (k)  3 k 3


for 1  k  L1 . The second law says that  = kuk2 stays finite and non-zero as 0. So for
turbulent flow, kuk2 as 0. This is akin to energy cascading to the small scales.

1 Bounds on Turbulent Energy Dissipation in a Box

Well see how to go from the control parameters , L, f to the energy law. Suppose
X
f (x, t) = eik.x f (k, t) .
2 2
L kkk7 L
Math 654, Lecture 20 3/23/11 2

If we demand the force to be divergence-free, then

X
0 = .f = ieik.x k.f (k, t) ,

and so
k f (k, t)
for all k. Also suppose time dependence of the coefficients of f ei1 t , ei2 t where 1 /2 6 Q. And suppose
max < . Then by dotting the momentum equation with u and integrating over the box, we get
 
d 1
ku (, t)k22 = ku (, t)k22 + f (x, t) .u (x, t) dx.
dt 2

Now if we assume (w.l.o.g.) that u (x, 0) satisfies ui (x, 0) dxdydz = 0, the Poincare lemma reads

4 2
kuk22 kuk22 .
L2
To see this, write
X
kuk22 = L3 u (k)k2
k
k

by Parsevals theorem, and thus

X
kuk22 = L2 kkk2 k
u (k)k2
k6=0
X 4 2
L2 u (k)k2
k
L2
k6=0

4 2
= kuk22 .
L2
And by Cauchy-Schwarz,  
 
 f .u kf k2 kuk2 .
 

Setting these into the differential equation from above yields

 
d 1 4 2
kuk22 2 kuk22 + kf k2 kuk2 .
dt 2 L

Since  
d 1 d
kuk22 = kuk2 kuk2 ,
dt 2 dt
we have
d 4 2
kuk2 2 kuk2 + f (, t)2 .
dt L
Math 654, Lecture 20 3/23/11 3

By Gronwalls inequality,
t
4 2 4 2
ku (, t)k2 ku (, 0)k2 e L2
t
+ kf (, t0 )k2 e L2
(tt0 ) dt0
0
2

4
t
4
2 1e L2
ku (, 0)k2 e L2
t
+ sup {kf (, t00 )k2 } 4 2

.
t00 L2

!u (, t)!2

! $
!f (, t )!2
sup " 42 #
t L2

Figure 1.1: kuk2 stays bounded for all time.

Note that transient time L2 /, and in long enough time ku (, t)k2 supt00 {kf (, t00 )k2 }, i.e., lim supt kuk2
is bounded. This is a neccesary condition for a statistical steady state to exist given any amount of viscosity.
Now we can write
    
1 T d 12 ku (, t)k22 1 T ku (, t)k22 f (x, t) .u (x, t) dxdydz
= +
T 0 dt L3 T 0 L3 L3

and thus  
1 1 1

ku (, T )k22 ku (, 0)k22 = kuk2 T + hf .uiT .
T 2 2
Assuming the terms on the right have finite limits as T we arrive at the power balance

 = hf .ui .

Now define

1
U = kuk2 2 .
Lets look for a lower bound on :


 = kuk2
4 2

2 kuk2
L
4 2
= U 2
L2
Math 654, Lecture 20 3/23/11 4

by Poincares lemma. So
L
=
U3
4 2 L
U 2 3
L2 U

= 4 2
LU
2 1
= 4 .
Re

Now we look for an upper bound on :

 = hf .ui

1
1
kf k2 2 kuk2 2

1
= kf k2 2 U.

Dotting the momentum equation with f and integrating over the box yields

f .t u + f . (u.u) = f .4u + kf k22 .

After integration by parts,

d
(f .u) t f .u u. (f ) .u = 4f .u + kf k22 .
dt
1
T
Dividing by L3 to get a spatial average, then applying () dt yields
T 0


ht f .ui hu. (f ) .ui h4f .ui = kf k2 .

This implies


1
1
kf k2 kt f k2 2 U + sup {kf (x, t)k} U 2 + k4f k2 2 U.
| {z } x,t | {z }
2
1 | {z } C2 2
1
max hkf k i 2 1 L2
hkf k i 2
1
C1 L hkf k2 i 2

Note that C1 depends on the spectrum of f but not the amplitude, and similarly for C2 . Hence,

 21 1 C2
kf k2 max U + C1 U 2 + 2 U,
L L
and hence

1 1 3 C2
 U kf k2 2 max U 2 + C1 U + 2 U 2 .
L L
This all goes to show
L L 1
=
max + C1 + C2 .
U3 U Re
There are multiple ways to send Re ,1 but in any case is bounded above independent of viscosity.
1 Hold U fixed and send 0, or hold fixed and send U , or some combination.
Math 654, Lecture 20 3/23/11 5

The bounds we found on for turbulent flow in a box are depicted in the -Re diagram below.

Re
1 10 100 1000 104 105

100

10 max LU 1 + C1 + C2 Re1

bound
depends
1 on how
Re

0.1
4 2 Re1

0.01

0.001

Figure 1.2: vs. Re.

Math 654, Lecture 21 3/28/11 1

Lecture 21: 2D Turbulence and Introduction to Existence/Uniquenss

Last time we talked about how energy cascades to smaller scales and dissipates in turbulent flow. This has
a particular form in 2D, which well discuss now. Then well switch gears and discuss some of the more
mathematical questions in Navier-Stokes.

1 Turbulence in R2
Consider turbulent flow in a 2-dimensional box.

Figure 1.1: Turbulent flow in a 2D box.

The equations are

t u + u.u + p = 4u + f ,
.u = 0,

where now u = iu + jv. Again,  

d 1
kuk22 = kuk22 + f .u.
dt 2
Recall that  
v u
=u=k = k
x y
in R2 , and  
f2 f1
t + u. = 4 + .
x y
We can derive an evolution equation for the enstrophy via the equation above. This turns out to be
   
d 1 f2 f1
kk22 = kk22 + .
dt 2 x y

The key quantities are

1 2
Energy: 2 kuk2

Enstrophy: kk22
Math 654, Lecture 21 3/28/11 2


Energy dissipation rate:  = kuk2


Enstrophy dissipation rate: = kk2 .

Lets go to spectral space. Decompose u as

X
u (x, t) = eik.k u
(k, t) ,
k

where
k.
u (k, t) = 0;
this is the spectral version of .u = 0. We are in R2 , so vorticity appears as
X
=u=
k eik.k ik u
(k, t).
| {z }
k

k (k,t)

Thus
X
kk22 = L2 (k, t)k2
k
k
X
2
= L kkk2 k
u (k, t)k2
k
= kuk22 ,

and thus


 = 2 and = k4uk2 .

Recall that if = 0, then energy is conserved. And in 2D, enstrophy is conserved as well. Now if we take
k = 2/L in the Fourier decomposition, then

1X L2 1 X 2
k
u (k, t)k = u (k, t)k2 (k)
k
2 4 2 2
k k

E (k) dk

where weve assumed homogenous isotropic turbulence. This shows that the energy spectrum in 2D is

L2
E (k) = u (k)k2 2k.
k
8 2
The important quantities are

Kinetic Energy: E (k) dk

Enstrophy: k 2 E (k) dk

Energy dissipation rate: k 2 E (k) dk
Math 654, Lecture 21 3/28/11 3

Enstrophy dissipation rate: k 4 E (k) dk.

Observe that in 2D, we cant simultaneously conserve kinetic energy and enstrophy if energy only cascades
to larger wave numbers (smaller scales). So instead, 2D flows have both an inverse cascade of energy to
larger scales and a forward cascade to smaller scales.

E(k)

k
2 2
L l

Figure 1.2: The forward/inverse cascade.

Can we see the form of E (k)? Notice that [] = L2 /T 3 , [] = 1/T 3 , [k] = 1/L, and [E (k)] = L3 /T 2 .
Dimensional reasoning gives us two choices: 2/3 k 5/3 and 2/3 k 3 . It has been argued that nature follows
the smaller of these two choices, for a given k.

E(k)

k
3 k 3 3 k 3
2 5 2

Figure 1.3: The 2D energy spectrum.

In particular if we say
E (k) 2/3 k 3
Math 654, Lecture 21 3/28/11 4

for large k, then

2/
= k 4 E (k) dk
2/
k 4 2/3 k 3 dk

2/3
2

for large k. Solving for tells us

  16
3
=

is the smallest length scale at which energy-driven structures are observable before viscosity dominates. But
the most noticeable feature of 2D turbulence is the cascade of energy to large scales. This is featured on the
cover of the text for this course!

2 Introduction to Existence and Uniqueness

Lets begin with a simple example. Consider the following ode for X (t):

dX
= X 2 , X (0) = x0 R.
dt
Questions we might ask:

1. Does a solution exist for t > 0?

2. Can we find smooth solutions?

3. Will these solutions exist for all t > 0 or just some t > 0?

So there are two types of solutions: global solutions which exist for all time, and local solutions which
exist for some time t > 0. In general, the answer to the questions above depend on the initial data. From
algebraic considerations we know a solution is
x0
X (t) = .
1 x0 t

If x0 < 0, then X(t) exists for all t > 0, and X (t) 0 as t . If x0 > 0, then X (t) as t 1/x0 .
What if we try to continue X (t) beyond the singularity at 1/x0 ? Suppose we allow smoothness to fail at
t = 1/x0 . For any 0 a we can produce such a solution:

1x
x0
0 t < x10
0t
X (t) = |a|
t > x10 .

1
1+|a| t x

0

This is horribly non-unique!

Math 654, Lecture 21 3/28/11 5

X(t)

x0
x1
0
t

||

Figure 2.1: Finite time blow-up.

Now well discuss existence and uniqueness of solutions to systems of ode. Consider z (t) = z 1 (t) , . . . , z N (t)
CN with q
2 2 2
|z| = |z 1 | + + |z N | .
The setup is
dz
= F (z)
dt
with F : CN CN , so
dz 1

= F 1 z1, z2, . . . , zN
dt
..
.
dz N

= F N z1, z2, . . . , zN .
dt
The initial data is
z (0) = z0 CN .
Definition. F : CN CN is said to be Lipschitz continuous if there exists K R so that for all z, z CN ,
|F (z) F (
z )| K |z z| .
Here K is called the Lipschitz constant.
Remark. An alternate characterization of K is
|F (z) F (z )|
K = sup .
z CN
z, |z z|

Example. F (x) = x2 is not Lipschitz, for

 2 
|F (x) F (
x)| = x x
2 
= |x + x
| |x x
| ,
x 2

which is unbounded on R. But F (x) = 1+|x| with 0 <   1 is Lipschitz. Here, K O 1 .
Math 654, Lecture 21 3/28/11 6

Theorem. If F is Lipschitz continuous, then there is a unique global solution to the initial value problem
dz
= F (z) , z (0) = z0 CN .
dt

Proof. The idea is to write the ode in integral form, then iterate towards a solution. So we explicitly construct
solutions to t
z (t) = z0 + F (z (t0 )) dt0 .
0
Start from n = 0 and define Picard iterates
t
zn+1 (t) = z0 + F (zn (t0 )) dt0 .
0

So
t
z1 (t) = z0 + F (z0 ) dt0 = z0 + F (z0 ) t
0
t
z2 (t) = z0 + F (z1 (t0 )) dt0
0
..
.

and so forth. We need to demonstrate convergence. Let K be the Lipschitz constant for F , and consider
t [0, T ] for T < K 1 . Each zn is continuous. Well show that the sequence of continuous functions {zn (t)}
converges in the space


C [0, T ] ; CN = f : [0, T ] CN s.t.f is continuous

with the norm

kf k = sup |f (t)| .
t[0,T ]

Well continue this next time.

Math 654, Lecture 22 3/30/11 1

Lecture 22: Existence and Uniqueness (ctd.)

Recall the following definition:

Definition 1. F : CN CN is Lipschitz continuous if there exists K R so that

|F (z) F (z 0 )| K |z z 0 | .

We were in the middle of proving a theorem giving sufficient conditions for existence and uniqueness in ode.

1 Existence and Uniqueness for ODE

Theorem 1. If F is Lipshitz, then the initial value problem

z = F (z) , z (0) = z0 CN

has a unique global (smooth) solution.

Proof. We proceed by Picard iteration. Define a sequence of continuous functions {zn+1 } via
t
zn+1 (t) = z0 + F (zn (t0 )) dt0
0


for t [0, T ] where T < K 1 . Now well show the zn (t) converge in C [0, T ] ; CN in the norm

kzk = sup |z (t)| .

t[0,T ]

Consider that
 t 
 
kzn+1 zn k = sup   0 0
[F (zn (t )) F (zn1 (t ))] dt  0
t[0,T ] 0
t
sup |F (zn (t0 )) F (zn1 (t0 ))| dt0
t[0,T ] 0
t
sup K |zn (t0 ) zn1 (t0 )| dt0
t[0,T ] 0
( )
sup K sup |zn (t0 ) zn1 (t0 )| t
t[0,T ] t0 [0,T ]

= KT kzn zn1 k.

Therefore,

kzk zk1 k (KT ) kzk1 zk2 k

2
(KT ) kzk2 zk3 k
k1
(KT ) kz1 z0 k,
Math 654, Lecture 22 3/30/11 2

and as T < K 1 , we have

kzk zk1 k 0
as k . Now write

zn (t) = z0 + (z1 z0 ) + (z2 z1 ) + + (zn zn1 )

n
X
= |zn (t)| |z0 | + |zk zk1 |
k=1
n
X k1
= kzn k kz0 k + kz1 z0 k (KT ) .
k=1

Since
n
X
X
k1 k1 1
(KT ) (KT ) = ,
1 KT
k=1 k=1

we can conclude that {zn (t)} is a uniformly convergent sequence of continuous functions.1 Thus,

z (t) = lim zn (t)

n

is a continuous function on [0, T ].

To see z (t) satisfies the differential equation, we must show
t t
F (zn (t0 )) dt0 F (z (t0 )) dt0
0 0

as t and in the given norm. So write

 t t 
  t
sup  F (zn (t0 )) dt0 F (z (t0 )) dt0  sup |F (z (t0 )) F (zn (t0 ))| dt0
t[0,T ] 0 0 t[0,T ] 0
KT kz zn k
0

as n . To finish the proof, we must extend the solution to t > 0. Just apply the same process with
new initial conditions at t = T and go forward in steps. But if we had uniqueness, then we could restart the
process at T and smoothness (differentiability) would be guaranteed.
So to prove uniqueness, suppose z (t) and z (t) both satisfy the ivp:

dz d
z
= F (z) and with z (0) = z0 = z (0) .
dt dt
Then define y (t) = z (t) z (t). We have

dy
= F (z (t)) F (
z (t)) = F (z (t)) F (z (t) y (t)) ,
dt
1 Proof of this is by completeness of C and the Weierstrass M-test.
Math 654, Lecture 22 3/30/11 3

so
d 2

|y (t)| = y (F (z) F (z y)) + y F (z) F (z y)
dt
2 |y| |F (z) F (z y)|
2
2K |y|

by the Lipshitz condition. By Gronwall,

2 2
|y (t)| |y (0)| e2Kt = 0

which proves uniqueness.

How can we generalize the idea of Lipshitz continuous to problems like Navier-Stokes?
Definition 2. We say F : CN CN is locally Lipshitz if for each bounded set A CN CN there exists a
bounded K (z, z 0 ) KA < so that

|F (z) F (z 0 )| K (z, z 0 ) |z z 0 |

on A.
Theorem 2. If F is locally Lipshitz, then the initial value problem

z = F (z) , z (0) = z0 CN

has a unique local (smooth) solution.

Proof. Picard iteration, but locally. Choose c R so that c > |z0 | and let

K0 = sup K (z, z) .
|z|,|
z |<2c

1
Let T = (K 0 + |F (z0 )| /c) , then construct a solution on [0, T ] via Picard.
Now show that all iterates stay in a ball of radius 2c. So we proceed via induction. For the base case, observe
that
|z0 | < c < 2c,
and since
z1 (t) = z0 + F (z0 ) t,
we also have

|z1 (t)| |z0 | + |F (z0 )| T

|F (z0 )|
= |z0 | +
K 0 + |F (zc 0 )|
c
c+ 0
1 + |FK(z0 )|
< 2c.
Math 654, Lecture 22 3/30/11 4

For the inductive step, suppose kz1 k, . . . , kzn k < 2c, and argue that kzn+1 k < 2c. Write
 t 
 
kzn+1 zn k = sup  [F (zn (t0 )) F (zn1 (t0 ))] dt0 
t[0,T ] 0
0
K T kzn zn1 k

n
(K 0 T ) kz1 z0 k
n
(K 0 T ) |F (z0 )| T,
then write
n
X
kzn+1 k = kz0 + (zm+1 zm )k
m=0
Xn
kz0 k + kzm+1 zm k
m=0
Xn
m
kz0 k + (K 0 T ) |F (z0 )| T
m=0
|F (z0 )| T
kz0 k +
1 K 0T
2c
by choice of T .
Now proceed as before: iterate forward in T , then prove uniquess.

But the solution is not global in general. The following example shows why.
Example 1. Consider the differential equation
dx
= x2
dt
which has local solution
x0
x (t) = .
1 x0 t
This has finite-time blow-up. But x2 is locally Lipshitz, so the theorem guarantees the uniqueness of this
solution locally.

However, there are global solutions to locally Lipshitz differential equations.

Example 2. Consider
dx
= x3
dt
with > 0. Say x (0) = x0 , then a solution is
x0
x (t) = p .
1 + 2x20 t
This is certainly defined for all t 0, but x3 is only locally Lipshitz.
Math 654, Lecture 22 3/30/11 5

How would we demonstrate the existence of the global solution in the previous example, even though we
only have local Lipshitz? The idea is as follows: If we know a priori that

|x (t)| C (x0 ) < ,

then there is no problem constructing solutions for t 0. So if we have a priori bounds on solutions, then
local Lipshitz is effectively global Lipshitz, and we have continuous and unique solutions for all time.
But note that local Lipshitz is necessary. The canonical example is as follows:

Example 3. Consider the differential equation

dx p
= |x|.
dt
This is not locally Lipshitz. And for x (0) = 0, we can exhibit more than one solution. One solution is
x (t) = 0, and another is x (t) = 14 t2 . In fact there are an infinite number of solutions:
(
0 0 t t0
x (t) = 1 0 2
.
4 (t t ) t t0

So we need at least local Lipshitz for any sort of uniqueness.

Math 654, Lecture 23 4/4/11 1

Lecture 23: Navier-Stokes in Fourier Space and Galerkin Truncations

Last time we gave a proof of the classic existence and uniqueness theorem for the ode z = F (z) with Lipshitz
continuous F . To do so, we constructed Picard iterates
t
zn+1 (t) = z0 + F (zn (s)) ds
0

and demonstrated the existence of a limiting function z which solved the ode in question.
We know the following facts concerning the ode z = F (z) with z CN :

1. If F is Lipshitz continuous then the ode has a unique global solution.

2. If F is locally Lipshitz continuous then there exists a unique local solution.

3. If F is locally Lipshitz and we have an a priori bound on the solution then there exists a global solution.

4. Without the local Lipshitz condition, there may not be a unique solution.

Today well construct a sequence of approximations to Navier-Stokes, each of which falls under the third
case.

1 The Fourier Representation of Navier-Stokes

The equations are

t u + u.u + p = 4u + f (x)
.u = 0

with initial condition u (x, 0) = u0 (x). The domain is a cube with periodic boundary conditions.

Figure 1.1: The domain of interest.

W.l.o.g., we can demand

.f = .u0 = 0.
Math 654, Lecture 23 4/4/11 2

The Fourier decomposition of u is X

u (x, t) = eik.x u
(k, t)
k

with
2  
k= in1 + jn2 + kn
3 , ni Z.
L

k2

k1

Figure 1.2: Wavenumber space.

As usual,
1
(k, t) =
u eik.x u (x, t) dx.
L3
We can write the initial condition as X
u0 (x) = 0 (k) eik.x ,
u
k

and reinterpret the divergence-free condition on u0 as

X
0 = .u0 = u0 (k) eik.x k.
ik. u0 (k) = 0.
k

And because u (x, t) R3 , we must have

(k, t) = u
u (k, t) .

Now well rewrite Navier-Stokes in Fourier spaces, i.e., well find a set of odes involving the Fourier coefficients
(k, t). The first step is to multiply Navier-Stokes by eik.x and integrate: we do this term by term. The
u
t u term becomes
 
1 ik.x d 1 ik.x d
e t u (x, t) dx = e u (x, t) dx = u (k, t) .
L3 dt L3 dt
Math 654, Lecture 23 4/4/11 3

The 4u term becomes

1 1 

eik.x 4u (x, t) dx = . eik.x u (x, t) + ik.u (x, t) eik.x dx;
L3 L3
the first integral is zero due to periodic boundary conditions, so

1 ik.x i
e 4u (x, t) dx = k. eik.x u (x, t) dx
L3 L3

i
= k.ik eik.x u (x, t) dx
L3
= kkk2 u (k, t) .
The non-linear term becomes
" X 0
!
X 00
!#
1 1
eik.x u.u dx = e ik.x
e ik .x
u 0
(k , t) . e ik .x 00
(k , t)
u dx
L3 L3
k0 k00
X
1
ei(k +k k).x dx
0 00
0 00 00
= (k , t) .ik u
u (k , t) 3
L
k0 ,k00 | {z }
k0 +k00 ,k
X
= ik00 .
u (k0 , t) u
(k00 , t) .
k0 +k00 =k

Now well deal with the pressure term. Taking the divergence of Navier-Stokes and manipulating the result
yields
4p = . (u.u) ;
taking the Fourier transform then yields
X
kkk2 p (k, t) = ik.i k00 .
u (k0 , t) u
(k00 , t)
k0 +k00 =k

once we recall
d (k) = ik.
.v v (k) .
So
k X
p (k, t) = k00 .
u (k0 , t) u
(k00 , t) ,
kkk2
k0 +k00 =k

and since
c (k) = ik
p p (k)
we find
X
c (k) = i kk
p k00 .
u (k0 , t) u
(k00 , t) .
kkk2
k0 +k00 =k

We can interpret the dot as a matrix-vector product where

k1 k1 k1 k2 k1 k3

kk = k2 k1 . . . .
k3 k1 k3 k3
Math 654, Lecture 23 4/4/11 4

So we have arrived at an infinite set of odes constituting Navier-Stokes in Fourier space:

  X
d kk
2
(k, t) = kkk u
u (k, t) i I k00 . (k00 , t) + f (k, t) .
u (k0 , t) u
dt kkk2 0 00 k +k =k

This relation holds for k 6= 0; well deal with this case in a bit.
kk
Observe that the operator P = I kkk2 is a projection operator onto vectors perpendicular to k. Indeed P
is a projector, as
  
kk kk
P2 = I I
kkk2 kkk2
kk kkkk2 k
= I 2 2
+
kkk kkk4
kk
= I
kkk2
= P.

Notice also that P is symmetric. Finally note that since

 
kk kkk2 k
k. I 2
=k = 0,
kkk kkk2

our set of odes is self-consistent.

For k = 0, we look at the time-evolution of

1
(0, t) =
u u (x, t) dx.
L3
Since
p dx = . (uu) dx = 0,

(0, t) evolves in time via

u
d
(0, t) = f (0, t) .
u
dt
We can solve this forward in time trivially:
t
(0, t) = u
u 0 (0) + f (0, s) ds.
0

2 Galerkin Truncation to Navier-Stokes

From here on, well assume u 0 (0) = 0 and f (0, t) 0 for simplicity, otherwise well complicate things
without gaining any understanding. To build approximate solutions to Navier-Stokes, we could keep only
the Fourier modes of u with kkk < K for positive K. This would yield K 3 modes. Or we could keep the
modes with ki < K, again yielding K 3 modes. Both of these truncations are pictured on the next page.
Math 654, Lecture 23 4/4/11 5

Figure 2.1: Two truncation methods.

The first method of truncation is called Galerkin truncation, and is what well pursue throughout the
remainder of the course.
Consider the Galerkin projection operator PN which projects onto modes into the set {k s.t.kkk < K}:
(
u (k, t) kkk < K
PN (
u (k, t)) = .
0 kkk K

Well build the N th approximate to u by requiring that the u N (k, t) satisfy Navier-Stokes in Fourier space
whenever kkk < K. Otherwise, well require u N (k, t) 0. Now the odes are
  X
d N kk
(k, t) = kkk2 u
u N (k, t) i I k00 . N (k00 , t) + f N (k, t) .
uN (k0 , t) u
dt kkk 2 | {z }
k0 +k00 =k
PN (f (k,t))

Note the sum here is now a finite sum, due to the truncation. The initial conditions are given by
N (k, 0) = PN (
u u0 (k)). This is a set of approximately K 3 odes, and is locally Lipshitz in time. Next
time, well show that X
uN (k, t)k2 < M
k
k

for some M R which is independent of N , so establishing an a priori bound.

Math 654, Lecture 24 4/6/11 1

Lecture 24: Global Existence of Solutions to Galerkin Truncated NSE

We have

t u + u.u + p = 4u + f
.u = 0

in a periodic box with side length L.

Figure 0.1: The domain of interest.

W.l.o.g., we demand
.f = 0 and f (x) dx = 0.

The initial condition is u (x, 0) = u0 (x) and satisfies

.u0 = 0 and u0 (x) dx = 0.

We wrote
1
(k, t) =
u eik.x u (x, t) dx
L3
with
2nx 2ny 2nz
k = i +j +k
L L L
where ni = . . . , 2, 1, 0, 1, 2, . . . . With this we showed NSE is equivalent to
  X
d kk
(k, t) = k 2 u
u (k, t) i I 2 u (k00 , t) + f (k) .
(k0 , t) .k00 u
dt k 0 00 k +k =k

Also we showed .u = 0 is equivalent to

k.
u (k, t) = 0.
This is an infinite set of ordinary differential equations, and as they are all coupled it is truly an infinite set.
Recall we discussed two ways to truncate the equations; the first of these methods was the Galerkin trun-
cation, pictured below.
Math 654, Lecture 24 4/6/11 2

ky

kx

Figure 0.2: Galerkin truncation of Fourier modes.

Note that k = 0 is not in the set. Now let N be the order of a Galerkin truncation, i.e., the number of
wave-numbers considered. Then the Galerkin truncated ode is
" #
d (N ) X

u 2 (N )
(k, t) = k u iP
u (N ) 0 00 (N )
(k , t) .k u (k, t) + f (N ) (x) (0.1)
dt 0 00
k +k =k

where weve set

(N ) (k, t) 0
u
whenever k is outside the collection and
(
f (k, t) k in the set
(N )
f (k) = .
0 otherwise

As initial conditions, (
(N ) u 0 (k) k in the set

u (k, t) = .
0 otherwise
This is a set of N ivps. set is locally Lipshetz, hence there exists a local unique solution. But to get global
existence, we need an a priori bound.

1 A Priori Bound for Galerkin Approximations

Define X
u(N ) (x, t) = eik.x u
(N ) (k, t) .
k in the set
N
Also define a new projection operator P on periodic functions g (x) by

X
PN g (x) = eik.x g (k) .
k in the set
Math 654, Lecture 24 4/6/11 3

This is projection onto the N retained Fourier modes. Note that

2
PN = PN ,

so indeed PN is a projector. Also note

X

h (k)
g (x) PN h (x) dx = L3 g (k) PN
k
X
= L 3 h (k)
g (k) PN

k in the set
X
= L 3
PN g (k) h
(k)
k



= PN g (x) h (x) dx,

so PN is self-adjoint.
Now multiply (0.1) by eik.x :
" #
d (N ) X 0 00
eik.x u (k, t) = k 2 u
(N ) eik.x iP eik .x u
(N ) (k0 , t) .k00 eik .x (N )

u (k, t) + f (N ) (x) eik.x .
dt
k0 +k00 =k

So then write
(N )  

u (x, t) = 4u(N ) (x, t) PN u(N ) .u(N ) + p(N ) + PN f (x)
t
and after rearranging,
(N )  

u (x, t) + PN u(N ) .u(N ) + p(N ) = 4u(N ) (x, t) PN f (x) . (1.1)
t
The initial conditions are

u (x, 0) = PN u0 (x) .
This is almost NSE (in a rough sense, PN I as N ). Note that

X
PN g (x) = eik.x g (k)
k in the set
X
1 0
= eik.x eik.x g (x0 ) dx0
L3
k in the set

= P (x, x0 ) g (x0 ) dx0

with X
(x, x0 ) = 1
P eik.(xx ) .
0

L3
k in the set

looks more and more like a delta function.

As we take more and more wave numbers in the set, P
Math 654, Lecture 24 4/6/11 4

Now dot uN (x, t) into (1.1) and integrate over the box:
 
d 1 (N ) 2 (N )
ku (, t)k2 = ku (, t)k2 + u(N ) (x, t) .f (x) dx.
2
dt 2
By Parseval and Poincare,
X
ku(N ) k22 = k 2 k
u(N ) (k, t)k2
k in the set
4 2 X
u (k, t)k2
k
L2
k in the set
4 2 (N ) 2
= ku k2 .
L2
So by Cauchy-Schwarz,
d (N ) 4 2
ku k2 2 ku(N ) k2 + kf k2 .
dt L
Now by Gronwall:
2
4 t
(N ) 4 2 1e L 2
ku(N ) (, t)k2 ku0 k2 e L2 t + kf k2 4 2
.
L2

Note that
! 21
(N )
X
ku0 k2 L3 u0 (k)k2
k
k in the set
! 12
X
3
L u0 (k)k2
k
k
= ku0 k2 ,
and hence
4 2
4
2 1 e L2 t
ku(N ) (, t)k2 ku0 k2 e L2
t
+ L2 kf k2 .
4 2
Weve obtained a uniform bound for all time and all N .
! !
! (N ) !
!u (, t)!
2

L2 !f !2
4 2

Figure 1.1: A priori bound on ku(N ) (, t)k2 .

Math 654, Lecture 24 4/6/11 5

All this has brought us to the following questions:

Does u(N ) (x, t) possess a limit as N ?

Does the limit satisfy NSE?

Is the limit unique?
Is the limit smooth?

Next time well discuss how difficult these questions really are.
Math 654, Lecture 25 4/10/11 1

Lecture 25: Convergence of Galerkin Approximations and Weak NSE

The setup:

t u + u.u + p = 4u + f
.u = 0

with .f = 0 and initial conditions u (x, 0) = u0 (x). The domain:

Figure 0.1: The domain of interest.

And there is no flow on average, i.e., u0 dx = 0 and f dx = 0. To approximate we projected onto N
Fourier modes via PN to give Galerkin approximations


PN uN = uN (x, t) .

Initial conditions truncate as

uN (x, 0) = PN u0 (x) .
Thus we arrive at the (approximated) conservation laws



t uN + PN uN .uN + pN = 4uN + PN f (x)
.uN = 0.

The first is really an integral-differential equation, as the PN operator acts like an integral against a Dirichlet
kernel.
Next, given X
kuN k22 = L3 uN (k, t)k2 ,
k
kS

where S is the set of retained wave numbers in the truncation, we derived the evolution equation
 
d 1 N 2
ku k2 = kuN k22 + uN .f .
dt 2

This gave an a priori energy bound and thus global existence of (approximate) solutions. Today we look at
convergence of these solutions; in particular, well ask if they converge at all.
Math 654, Lecture 25 4/10/11 2

1 Convergence of Approximate Solutions

Wed like to write uN u (x, t) in some sense. Recall that given a sequence of real numbers in a compact
set theremust exist
 a convergent subsequence. But the limit is not unique. In our situation, we have
3
uN L2 [0, L] . And given the a priori energy bound, we know the uN are all contained in a closed (and
   
3 3
bounded) subset of L2 [0, L] . But L2 [0, L] is an infinite dimensional space! So this is not compactness.

Suppose we integrate the energy evolution equation on [0, T ]:

T T  
1 N 2 N 2 1 N 2 N
ku (, T )k2 + ku (, t)k2 dt = ku0 k2 + u .f dx dt. (1.1)
2 0 2 0

Because PN is a projection operator, and because of Cauchy-Schwarz, we can write

T
1 N 2 1
ku (, T )k2 + kuN (, t)k22 dt ku0 k22 + C T
2 0 2
for some C < uniform in N, T . Thus
T
1
kuN (, t)k22 dt ku0 k22 + C T,
0 2

a uniform bound in N . This is extra information: uN lies in a compact subset of L2 in an appropriate

topology. Recall that in the norm topology (the usual topology), we say n 0 iff kn k2 0.

Definition 1. In the weak topology, we say n 0 iff for all smooth we have n dx 0, and then we
write n * 0.

The fact is that there are weakly convergent subsequences of uN which satisfy the weak form of NSE.
Suppose u (x, t) solves NSE. Then given a smooth, periodic, divergence-free v (x, t), we can write
T T
v.t u dxdt + v. (u.u) dxdt
0 box 0 box
T T
= u.4v dxdt + v.f dxdt,
0 box 0

after dotting NSE by v and integrating. Integrating by parts, we get

T T
v (x, T ) .u (x, T ) dx t v.u dxdt u. (v) .u dxdt
0 0
T T
= v (x, 0) .u0 (x) dx + u.4v dxdt + v.f dxdt. (1.2)
0 0

Also if we have a smooth, periodic q (x), we can write

u.q dx = 0 (1.3)

by the divergence-free condition.

Math 654, Lecture 25 4/10/11 3

Definition 2. We say u (x, t) satisfies the weak form of NSE if, given any smooth, periodic, and divergence-
free v (x, t) and any smooth and periodic q (x), we have (1.2) and (1.3).

Its a fact that weak solutions satisfy the energy equality (and subsequent inequality) (1.1). Wed like to
ask if a weak limit has the same characteristic. First, consider functions which are periodic on [0, L] in the
following example.
Example 1. Consider the sequence {N } given by
 
2N x
N (x) = 2 sin .
L
Then we have
kN k22 = L,
so N B (0, L) for all N . And the N do not converge in the norm topology. Nevertheless,
     
 L  L  L 0 2N x
  
 (x) N (x) dx =  (x) 2 cos dx
 0  2N  0 L 
L ! 1
2
L 0 2 1
( (x)) dx L2
2N 0
0
as N . So N 0 in the weak topology! Let be the weak-limit of N . Weve found
= 0 and kk2 = 0,
but
lim kN k = L 6= 0.
N

Weve come to a general fact. Let u denote the weak limit of the Galerkin approximations, uN . Then
kuk2 lim kuN k2 .
N

So what was the energy equality, (1.1), becomes an inequality in the weak limit. What would happen if this
inequality is strict? This would indicate that some part of the energy dissapeared into the fine scales when
we smear the uN with a test function and integrate. But really we dont know if this inequality is strict or
not. This is unsettling indeed.
Its even worse: the weak form of NSE is non-predictive. Consider the next example, first due to Leray in
1932.
Example 2. Suppose x R3 . Define u by
u (x, t) = a (t) (x)
for some with 4 (x) = 0. We claim u is a weak solution of the (unforced) NSE, so long as a satisfies
T
2
(a (t)) dt < .
0
Math 654, Lecture 25 4/10/11 4

Its clear from the definition that u satisfies (1.3), where now test functions have compact support. What
about (1.2)? Since v is divergence-free (it is a test function), each term in (1.2) dissapears immediately
expect for the third term on the left. But

2
u.v.u dx = (a (t)) ,i vj,i ,j dx

2
= (a (t)) vj ,i ,ij dx
 
2 1 2
= (a (t)) v. kk dx
2
= 0,

for again v is divergence-free. So indeed u is a weak solution. But u lives on an infinite domain, and has
infinite energy! This is entirely unphysical.
Math 654, Lecture 26 4/13/11 1

Lecture 26: Uniqueness and Enstrophy

We had been considering uniqueness of solutions to Navier-Stokes. Well continue that today.

1 Uniqueness of Solutions
satisfying
Suppose we have u, u

t u + u.u + p = 4u + f ,
+u
t u .
u + p = 4
u+f

and the divergence-free condition. Then we define v by

u.
v (x, t) = u

Clearly v is divergence-free, and now

t v + v.v + u.v + v.u +

p = 4v.

Dotting with v and integrating by parts yields

 
d 1
kvk22 + v. (u) .v dx = kvk22 0.
dt 2
Then
   
d 1 v. (u) .v
kvk22 kvk22
dt 2 v22
 
v. (u) .v
max kvk22
u v22
ku (, t)k kvk22

where
ku (, t)k = supku (x, t)k.
x

So by Gronwall,  
t
0 0
kv (, t)k2 kv (, 0)k2 exp ku (, t )k dt .
0
We need more regularity to guarantee a bound.
Here are some sufficient conditions for everything to be nice. . .
t
0
ku (, t0 )k42 dt0 < , or
ku (, t)k22 < .

If we could know either of these, then we could demonstrate both that the solution is C ((0, t)) and that it
is unique. Recall that ku (, t)k22 = kk22 , so we could say the second condition is an enstrophy condition.
Lets look at the second condition some more.
Math 654, Lecture 26 4/13/11 2

2 The Enstrophy Condition

Navier-Stokes again:
t u + u.u + p = 4u + f .
Now if we take the curl of Navier-Stokes we get

t + u. = 4 + .u + f .

Of course we could do all this with Galerkin approximations, then wed get

t N + uN . N = 4 N + N .uN + PN ( f )

and similarly for Navier-Stokes. As always, we could dot into the vorticity equation and integrate. Or we
could dot 4u into Navier-Stokes and integrate. (They are morally the same.) The second results in the
equation
1 d
kuk22 = k4uk22 + u. (u) .4u,
2 dt
after forgetting about the forcing term. Equivalently, we could write (after integration by parts)

1 d
kk22 = kk22 + . (u) ..
2 dt
Well continue with the first of these.
The analysis goes as follows. First,

1 d
kuk22 = k4uk22 + u. (u) .4u
2 dt
k4uk22 + kuk kuk2 k4uk2

by Cauchy-Schwarz. We look to express kuk in terms of kuk2 , k4uk2 .

Proposition 1. If u is mean-zero, we have

1 1
kuk Ckuk22 k4uk22

in R3 .

Proof. We write
 
X 
 
|u (x)| =  e ik.x
(k)
u

k6=0 
X
|
u (k)|
k6=0
X X
= |
u (k)| + |
u (k)|
0<|k|K |k|>K
Math 654, Lecture 26 4/13/11 3

for some K. Then

X
|k| X |k|2
|u (x)| |
u (k)| + 2 |u (k)|
|k| |k|
0<|k|K |k|>K
12 21 12 21
X 1 X 2 2
X 1 X 4 2
2
u (k)| +
|k| | 4
|k| | u (k)|
0<|k|K
|k| k6=0 |k|>K
|k| k6=0
12 12
X   21 X 1   12
1 1 2 1 2
2 kuk2 + 4 k4uk2
|k| L3 |k| L3
0<|k|K |k|>K
21 12
1 X 1 (2)
3
1 X 1 (2)
3
kuk2 + k4uk2
3 2 L3 3 4
(2) 2 |k| (2) |k|>K
|k| L3
L <|k|K

! 21 !
K
1 k 2 dk 1 k 2 dk
3 4 kuk2 + 2 4 k4uk2
(2) 0 k2 (2) K k4
 
1 1 1
K 2 kuk2 + 1 k4uk2 .
2 K2
Note that in our setup dk = 2/L, since we are in a box of side length L. Now we are free to choose
positive K. So pick K = k4uk2 /kuk2 , then
2 1 1
|u (x)| kuk22 k4uk22 .
2


So take C = 2/ 2 to finish the proof.


Since 2/ 2 < 1 we can now write
 
d 1 3 3
kuk22 k4uk22 + k4uk22 k4uk22
dt 2
 
1 3 3
k4uk22 + k4uk22 k4uk22

6
kuk2
C
4
C
= kuk62
3
for some constant C. Setting X (t) = ku (, t)k22 , we have thus
dX 2C
3 X 3.
dt
Going forward, we write  
d 2C 2
X (X (t)) X (t)
dt 3
Math 654, Lecture 26 4/13/11 4

and hence  t 
2C 0 2 0
X (t) X (0) exp X (t ) dt .
3 0

Hence  t 
2C
ku (, t)k22 ku (, 0)k22 exp ku (, t 0
)k42 dt 0
.
3 0

So if we knew ku (, t0 )k42 had only time-integrable singularities, then in fact ku (, t)k22 would be finite for
all time.
We could also write
dX 2C 3
X
dt 3
X(t) t
1 2C 0
= dX dt
X(0) X3 0
3

and hence
 2 2 
1 1 1 1 2C
+ t
2 X (t) X (0)2 3
 2
X (0) 4C 2
= +1 (X (0)) t
X (t) 3
 2
X (0) 4C 2
= 1 (X (0)) t
X (t) 3
 2
X (t) 1
= 2 .
X (0) 1 4C
(X (0)) t
3

Therefore,
X (0)
X (t) q .
2 4C
1 (X (0)) t 3
 
2
Thus X (t) is finite until time t = 3 / 4C (X (0)) . Weve just shown the following:

Proposition 2. If ku (, 0)k22 < , then there exists T > 0 so that ku (, t)k22 < for t < T .

We said earlier that ku (, t)k22 < is enough to guarantee smoothness of solutions, so if enstrophy is
bounded at t = 0 then a smooth and unique solution exists for finite time. See 8 of the text for more
details.