Lecture notes for Math 256B, Version 2019

Lenya Ryzhik∗
March 14, 2019

Nothing found here is original except for a few mistakes and misprints here and there.
These notes are simply a record of what I cover in class, to spare the students the necessity
of taking the lecture notes. The readers should consult the original books for a better pre-
sentation and context. We plan to follow the following books: C. Doering and J. Gibbon
“Applied Analysis of the Navier-Stokes Equations”, A. Majda and A. Bertozzi “Vorticity and
Incompressible Flow”, P. Constantin and C. Foias “The Navier-Stokes Equations”, as well as
lecture notes by Vladimir Sverak on the mathematical fluid dynamics that can be found on
his website.

1 The derivation of the Navier-Stokes and Euler equa-

tions
The state of the fluid is characterized by its density ρ(t, x) and fluid velocity u(t, x), and our
first task is to derive the partial differential equations that govern their evolution. They will
come from the conservation of mass, Newton’s second law and, finally, an assumption on the
material properties of the fluid.

The continuity equation

Each fluid particle is following a trajectory governed by the fluid velocity u(t, x):

dX(t, α)
= u(t, X(α, t)), X(0, α) = α. (1.1)
dt
Here, α is the starting position of the particle, and is sometimes called “the label”, and the
inverse map At : X(t, α) → α is called the “back-to-the-labels” map. If the flow u(t, x)
is sufficiently smooth, the forward map α → X(t, α) should preserve the mass. Let us
first assume that the fluid density ρ(t, x) = ρ0 is a constant, and see what we can deduce
from the mass preservation – the fluid is neither created nor destroyed. In the constant
density case, mass preservation is equivalent to the conservation of the volume. That is,
if V0 ⊂ Rd , (d = 2, 3) is an initial volume of a parcel of the fluid, then the set

V (t) = {X(t, α) : α ∈ V0 }
∗
Department of Mathematics, Stanford University, Stanford, CA 94305, USA; [email protected]

1
of where the particles that started in V0 at t = 0 ended up at a later time t > 0, should have
the same volume as V0 . In order to quantify this property, let us define the Jacobian
∂Xi (t, α)
J(t, α) = det( ).
∂αj
The change of variables formula, for the coordinate transformation α → X(t, α), implies that
volume preservation means that J(t, α) ≡ 1. As J(0, α) ≡ 1, this condition is equivalent
to dJ/dt ≡ 0. It follows from (1.1) that the full derivative matrix
∂Xi (t, α)
Hij (t, α) =
∂αj
obeys the evolution equation
n
dHij X ∂ui ∂Xk
= , (1.2)
dt k=1
∂x k ∂αj

that is, in the matrix form,

dH ∂ui
= (∇u)H, (∇u)ik = . (1.3)
dt ∂xk
The matrix Hij is also known as the deformation tensor. For example, if u = ū is a constant
vector, so that
X(t, α) = α + ūt,
then H = Id is the identity matrix. In order to find dJ/dt, with J(t, α) = det H(t, α), we
consider a general n × n matrix Aij (t) and decompose, for each i = 1, . . . , n fixed:
n
X
detA = (−1)i+j Mij Aij .
j=1

Note that the minors Mij 0 , for all 1 ≤ j ≤ n, do not depend on the matrix element Aij , hence
∂
(det A) = (−1)i+j Mij .
∂Aij
We conclude that n
d X dAij
(det A) = (−1)i+j Mij .
dt i,j=1
dt

Recall also that (A−1 )ij = (1/ det A)(−1)i+j Mji , meaning that
n
X
(−1)j+i Mij Akj = (det A)δik .
j=1

We apply this now to the matrix Hij :

n
dJ X dHij
= (−1)i+j Mij ,
dt i,j=1
dt

2
and n
X
Jδik = (−1)j+i Mij Hkj (1.4)
j=1

Here, Mij are the minors of the matrix Hij . Using (1.2) gives
n n
dJ X
i+j ∂ui X ∂ui
= (−1) Mij Hkj = Jδik = J(∇ · u). (1.5)
dt i,j,k=1
∂x k
i,k=1
∂x k

Preservation of the volume means that J ≡ 1. As H(0) = Id and J(0) = 1, this is equivalent
to the incompressibility condition:
∇ · u = 0. (1.6)
Here, we use the notation
n
X ∂uk
∇ · u = divu = .
k=1
∂xk
More generally, if the density is not constant, mass conservation would require that for
any initial volume V0 we would have (recall that ρ(t, x) is the fluid density)
ˆ
d
ρ(t, x)dx = 0, (1.7)
dt V (t)

where
V (t) = {X(t, α) : α ∈ V0 }.
Using the change of variables α → X(t, α) and writing
ˆ ˆ
ρ(t, x)dx = ρ(t, X(t, α))J(t, α)dα, (1.8)
V (t) V0

we see that mass conservation is equivalent to the condition

d
(ρ(t, X(t, α))J(t, α)) = 0. (1.9)
dt
Using (1.1) and (1.5) leads to

∂ρ
J + (u · ∇ρ)J + ρ(∇ · u)J = 0. (1.10)
∂t
Dividing by J we obtain the continuity equation
∂ρ
+ ∇ · (ρu) = 0. (1.11)
∂t
We note briefly some basic properties of (1.11). First, the total mass over the whole space is
conserved: ˆ ˆ
ρ(t, x)dx = ρ(0, x)dx. (1.12)
Rd Rd

3
This follows both from (1.11) after integration over Rd (assuming an appropriate decay at
infinity), and, independently, from our derivation of the continuity equation. If (1.11) is posed
in a bounded domain Ω then, in order to ensure mass preservation, one may assume that the
flow does not penetrate the boundary ∂Ω:

u · ν = 0 on ∂Ω. (1.13)

Here, ν is the outward normal to ∂Ω. Under this condition, we have

ˆ ˆ
ρ(t, x)dx = ρ(0, x)dx. (1.14)
Ω Ω

This may be verified directly from (1.11) but it also follows from our derivation of the con-
tinuity equation since (1.13) implies that Ω is an invariant region for the flow u: if α ∈ Ω
then X(t, α) ∈ Ω for all t > 0.
Furthermore, (1.11) preserves the positivity of the solution: if ρ(0, x) ≥ 0 then ρ(t, x) ≥ 0
for all t > 0 and x – this also follows from common sense: density can not become negative.

Newton’s second law in an inviscid fluid

The continuity equation for the evolution of the density ρ(t, x) should be supplemented by
an evolution equation for the fluid velocity u(t, x). This will come from Newton’s second law
of motion. Consider a fluid volume V . If the fluid is inviscid, so that there is no “internal
friction” in the fluid, the only force acting on this volume is due to the pressure:
ˆ ˆ
F =− pνdS = − ∇pdx, (1.15)
∂V V

where ∂V is the boundary of V , and ν is the outside normal to ∂V . Taking V to be an

infinitesimal volume around a point X(t), which moves with the fluid, Newton’s second law
of motion leads to the balance

ρ(t, X(t))Ẍ(t) = −∇p(t, X(t)). (1.16)

We may compute Ẍ(t) from (1.1):

d ∂uj (t, X(t)) X ∂uj (t, X(t))
Ẍj (t) = (uj (t, X(t)) = + Ẋk (t) (1.17)
dt ∂t k
∂xk
∂uj (t, X(t))
= + u(t, X(t)) · ∇uj (t, X(t)).
∂t
Therefore, we have the following equation of motion:
 ∂u 
ρ + u · ∇u + ∇p = 0. (1.18)
∂t
Equations (1.11) and (1.18) do not form a closed system of equations by themselves –
they involve n + 1 equations for n + 2 unknowns (the density ρ(t, x), the pressure p(t, x) and
the fluid velocity u(t, x)). The missing equation should provide the connection between the

4
density and the pressure, and this comes from the physics of the problem, that go into the
assumptions on the material properties of the fluid. In gas dynamics, it often takes the form
of a constitutive relation p = F (ρ), where F (ρ) is a given function, such as F (ρ) = Cργ with
some constant γ > 0. Then, the full system becomes
ρt + ∇ · (ρu) = 0
1
ut + u · ∇u + ∇p = 0, (1.19)
ρ
p = F (ρ).
The pressure may also depend on the temperature, and then the evolution of the local tem-
perature has to be included as well but we will not discuss this at the moment.

The linearized equations

The simplest solution of (1.19) is the constant density and pressure, zero fluid velocity state:
ρ = ρ0 , p = p0 = F (ρ0 ) and u = 0. (1.20)
Let us consider a small perturbation around this state:
ρ = ρ0 + εη + O(ε2 ),
p = p0 + εF 0 (ρ0 )η + O(ε2 ) (1.21)
u = εv + O(ε2 ),
with ε  1. Inserting these expansions into (1.19) gives, in the (leading) order O(ε):
ηt + ρ0 ∇ · v = 0
F 0 (ρ0 )
vt + ∇η = 0. (1.22)
ρ0
It is common to write this system in terms of v and the pressure perturbation p̃ = F 0 (ρ0 )η.
After dropping the tilde it becomes the linearized acoustic system
κ0 pt + ∇ · v = 0 (1.23)
ρ0 vt + ∇p = 0. (1.24)
Here, κ0 = 1/(F 0 (ρ0 )ρ0 ) is the compressibility constant. equations (1.23)-(1.24) form what
is known as the linearized acoustics system. Differentiating (1.23) in time and using (1.24)
leads to the wave equation for pressure:
1
ptt − ∆p = 0, (1.25)
c20
with the sound speed
1 p
c0 = √ = F 0 (ρ0 ). (1.26)
ρ0 κ0
The linearized acoustics is what governs most of the “real-world” applications at “bearable”
sound levels.

5
Euler’s equations in incompressible fluids
A common approximation in the fluid dynamics is to assume that the fluid is incompressible,
that is, its density is constant: ρ(t, x) = ρ0 , as the fluid can not be compressed. Using this
condition in (1.11), leads to another form of the incompressibility condition:
∇ · u = 0, (1.27)
that we have already seen before in (1.6) as the volume preservation condition for the flow.
That is natural: conservation of density means exactly that the volume of a fluid is preserved.
Equations (1.18) and (1.27) together form Euler’s equations for an incompressible fluid:
∂u 1
+ u · ∇u + ∇p = 0, (1.28)
∂t ρ0
∇ · u = 0. (1.29)
Unlike in the acoustics system, the pressure p(t, x) is not prescribed but is rather determined
by the fluid incompressibility condition. In other words, p(t, x) has to be chosen is such a
way that the solution to (1.28) remains divergence free. In order to find the pressure, we may
take the divergence of (1.28) , leading to the Poisson equation for the pressure in terms of the
velocity field:
n n
X ∂  ∂uj  X ∂uk ∂uj
∆p = −ρ0 ∇ · (ut + u · ∇u) = −ρ0 uk = −ρ0 . (1.30)
i,j=1
∂x j ∂x k i,j=1
∂x j ∂x k

We used the incompressibility condition (1.29) in the second and third equalities above.
Equations (1.28)-(1.30) together may be thought of as a closed system of equations for the
velocity u(t, x) alone since p(t, x) is determined by u(t, x) via (1.30). An extremely important
point is that the Poisson equation (1.30) for the pressure means that p(t, x) is a non-local
function of the velocity. Hence the Euler equations are a non-local system of equations for
the fluid velocity – the pressure field at a given point depends on the velocity distribution in
the whole space.
When the problem is posed in a bounded domain, we need to prescribe the boundary
conditions for the fluid velocity and pressure. If the physical domain Ω is fixed and the fluid
does not penetrate through its boundary, a natural physical condition for the fluid velocity is
that the normal component of the velocity vanishes at the boundary:
ν · u = 0 on ∂Ω, (1.31)
where ν is the outward normal to the boundary. It follows that
∂u
ν· = 0 on ∂Ω, (1.32)
∂t
thus the pressure satisfies the Neumann boundary conditions
∂p
= −ρ0 ν · (u · ∇u) on ∂Ω. (1.33)
∂ν
Often, as a simplification we will consider the Euler equations either in the whole space,
with the decaying boundary conditions at infinity, or with the periodic boundary conditions
on a two- or three-dimensional torus, as the boundaries bring extra (and, admittedly, very
interesting) difficulties into an already difficult problem.

6
The viscous stress and the Navier-Stokes equations
The previous discussion did not take into account the viscosity of a fluid, which comes from the
forces that resist the shearing motions because of the microscopic friction. The forces normal
to a given area element are associated to the pressure (which we did take into account), while
those acting in the plane of the area element are associated to the shear stress. In order to
derive the fluid motion equations, as a generalization of the force on a volume element V
coming from the pressure field:
ˆ ˆ
F =− pνdS = − ∇pdx, (1.34)
∂V V

we may write, for the force that acts on an infinitesimal surface area dS of a volume element V :
n
X
dFj = νk τkj dS, (1.35)
k=1

where ν is the outward normal to dS, and τ is the total stress tensor that includes both the
pressure and the shear stress. We will soon start making assumptions on the stress tensor
but for moment, we simply assume that the surface force has the form (1.35) with some
tensor τkj . Integrating this expression over the boundary ∂V leads to the total force acting
on the volume V : n ˆ n ˆ
X X ∂τkj
Fj = νk τkj dS = dx. (1.36)
k=1 ∂V k=1 V
∂xk
We will use the notation ∇ · τ for the vector with the components
n
X ∂τkj
(∇ · τ )j = , (1.37)
k=1
∂xk

as well as denote n
X
(ν · τ )j = νk τkj . (1.38)
k=1

In addition to the surface forces, there may internal forces that act inside the volume V , that
need to be balanced with the surface forces. Let us assume for the moment that the fluid is
in equilibrium, and let f be the internal forces, τ be the stress tensor, and V be an arbitrary
volume element. Then the balance of forces says that
ˆ ˆ
f dx + (∇ · τ )dx = 0, (1.39)
V V

which means that in an equilibrium we have

f + ∇ · τ = 0. (1.40)

The total angular momentum of the force should also vanish, meaning that (in three dimen-
sions) ˆ ˆ
(f × x)dx + ((ν · τ ) × x)dS = 0, (1.41)
V ∂V

7
for each volume element V . The surface integral above can be re-written as1
ˆ ˆ ˆ
∂  ∂τ
lj

εijk νl τlj xk dS = εijk (τlj xk )dx = εijk xk + τkj dx, for each i = 1, 2, 3.
∂V V ∂xl V ∂xl
(1.42)
Here, εink is the totally anti-symmetric tensor: (v × w)i = εijk vj wk , and εijk = 0 if any pair
of the indices i, j, k coincide, while if all i, j, k are different, then εijk = (−1)p+1 , where p = 1
if (ijk) is an even permutation, and p = 0 if it is odd. Using (1.40) in (1.42), we get
ˆ ˆ  
εijk νl τlj xk dS = εijk − fj xk + τkj dx, for each i = 1, 2, 3. (1.43)
∂V V

Returning to (1.41), and combing it with (1.43), we obtain

ˆ ˆ   ˆ
0= εijk fj xk dx + εijk − fj xk + τkj dx = εijk τkj dx, for each i = 1, 2, 3. (1.44)
V V V

As a consequence,
εijk τjk = 0, for each i = 1, 2, 3, (1.45)
which means that the tensor τij has to be symmetric.
Exercise. Modify the above computation to show that the stress tensor is symmetric
even if the fluid is not in an equilibrium.
We may now go back to the derivation of the Euler equations and proceed as before, the
difference being that the force term in the Newton second law is not −∇p but ∇ · τ . This
will lead to the equation of motion
∂u 1
+ u · ∇u = ∇ · τ. (1.46)
∂t ρ
As for the Euler equations, the evolution equation for the fluid velocity needs to be supple-
mented by the continuity equation
∂ρ
+ ∇ · (ρu) = 0. (1.47)
∂t
Previously, we needed also to prescribe the equation of state – the relation between the
pressure and the density. Now, we need to postulate, or derive from physical considerations,
an expression for the stress tensor. We will decompose it as

τij = −pδij + σij . (1.48)

The first term comes from the pressure – it leads to a force acting on a surface element in the
direction normal to the surface element. The second term comes from the shear stress, and
comes from the friction inside the fluid. It is natural to assume that it depends locally on ∇u
– if the flow is uniform there is no shearing force. In order to understand this dependence,
recall that, given a flow
dX
= u(t, X(t)), X(0) = α, (1.49)
dt
1
From now we will use the convention that the repeated indices are summed unless specified otherwise.

8
the deformation tensor Hij = ∂Xi /∂αj obeys
dHij ∂ui
= Hmj , Hij (0) = δij . (1.50)
dt ∂xm
Therefore, the skew-symmetric part of the matrix ∇u (locally in time and space) leads to a
rigid-body rotation and does not contribute to the shearing force. Hence, it is also natural to
assume that the shear stress σij depends only on the symmetric part of ∇u:
1  ∂ui ∂uj 
Dij = + . (1.51)
2 ∂xj ∂xi
In a Newtonian fluid, the shear stress depends linearly on the deformation tensor Dij :

σ = L(D),

for some linear map L between symmetric matrices. The map L should not depend on the
point x and it should be isotropic: for each rotation matrix Q we should have

L(QDQ∗ ) = QL(D)Q∗ . (1.52)

Exercise. Show that the above conditions imply that the map L has to have the form

[L(D)]ij = 2µDij + λδij Tr(D), (1.53)

with some constants λ and µ. These constants are called the Lamé parameters in the context
of the elasticity theory.
For an incompressible fluid, we have

TrD = ∇ · u = 0, (1.54)

hence the stress tensor has a simpler form

σij = 2µDij . (1.55)

We will make an additional assumption that µ and λ are constants that do not depend on
other physical parameters such as temperature, density or pressure. Then the force term
in (1.46) can be written as
∂τjk ∂ h ∂uj ∂uk i
[∇ · τ ]k = = − pδjk + µ( + ) + λ(∇ · u)δjk (1.56)
∂xj ∂xj ∂xk ∂xj
∂p ∂
=− + µ∆uk + (µ + λ) (∇ · u).
∂xk ∂xk
This leads to the Navier-Stokes equations of compressible fluid dynamics
∂u 1 µ (µ + λ)
+ u · ∇u + ∇p = ∆u + ∇(∇ · u) (1.57)
∂t ρ ρ ρ
∂ρ
+ ∇ · (ρu) = 0, (1.58)
∂t
p = F (ρ). (1.59)

9
As with the Euler equations, the equation of state may also involve the temperature, and
then the evolution equation for the temperature should also be prescribed.
The incompressibility constraint ∇ · u = 0, or, equivalently, the constant density ap-
proximation ρ = ρ0 , simplifies the system (1.57)-(1.59) to the incompressible Navier-Stokes
equations
∂u 1 µ
+ u · ∇u + ∇p = ∆u (1.60)
∂t ρ0 ρ0
∇ · u = 0. (1.61)
Note that Euler’s equations are formally recovered from the Navier-Stokes equations by setting
the viscosity µ = 0, or, equivalently, assuming that the shear stress vanishes.
From now on, unless specified otherwise, we will consider only the incompressible Euler
and Navier-Stokes equations.

Two-dimensional flows
We will sometimes consider the two-dimensional version of the Navier-Stokes equations, which
has exactly the same form as the three-dimensional equations (1.60)-(1.61) but with the fluid
velocity that has only two components: u = (u1 , u2 ), and, in addition, the problem is posed
for x ∈ R2 . These can be interpreted as the solutions of the three-dimensional Navier-Stokes
system of a special form u = (u1 (x1 , x2 ), u2 (x1 , x2 ), 0) with the pressure p = p(x1 , x2 ) – that
is, they are independent of x3 and the third component of the fluid velocity vanishes. It is
straightforward to check that, indeed, they satisfy (1.60)-(1.61) provided that ũ = (u1 , u2 )
satisfies
∂ ũ 1 µ
+ ũ · ∇ũ + ∇p = ∆ũ (1.62)
∂t ρ0 ρ0
∇ · ũ = 0, (1.63)
posed in R2 and not in R3 .

2 The vorticity evolution

The vorticity
An important role in the theory of fluids is played by the fluid vorticity. It is defined in terms
of the fluid velocity u(t, x) as a vector
ω = curlu = ∇ × u, ωi = εijk ∂j uk , (2.1)
in three dimensions, and as a scalar
∂u2 ∂u1
ω= − , (2.2)
∂x1 ∂x2
in two dimensions. The two-dimensional vorticity can be understood as the x3 -component of
the three-dimensional vorticity of the flow (u1 (x1 , x2 ), u2 (x1 , x2 ), 0) – the other two compo-
nents of the vorticity vanish for such flows.
The vorticity vector field in three dimensions is always divergence free:
∇ · ω = εijk ∂i ∂j uk = 0. (2.3)

10
Vorticity conservation in two dimensions
Let us now compute the evolution equation for the vorticity in two and three dimensions.
In the two-dimensional case, we start with the Navier-Stokes equations (we will set the den-
sity ρ0 = 1 for simplicity from now on, unless specified otherwise)

∂u
+ u · ∇u + ∇p = ν∆u, (2.4)
∂t
and compute
∂ω ∂  ∂p ∂u2 ∂u2  ∂  ∂p ∂u1 ∂u1 
= ν∆u2 − − u1 − u2 − ν∆u1 − − u1 − u2
∂t ∂x1 ∂x2 ∂x1 ∂x2 ∂x2 ∂x1 ∂x1 ∂x2
2 2 2
∂u1 ∂u2 ∂ u2 ∂u2 ∂u2 ∂ u2 ∂u1 ∂u1 ∂ u1
= ν∆ω − − u1 2
− − u2 + + u1 (2.5)
∂x1 ∂x1 ∂x1 ∂x1 ∂x2 ∂x1 ∂x2 ∂x2 ∂x1 ∂x1 ∂x2
∂u2 ∂u1 ∂ 2 u1 ∂ ∂u2 ∂u1 ∂ ∂u2 ∂u1
+ + u2 = ν∆ω − u 1 ( − ) − u2 ( − ) = ∆ω − u · ∇ω.
∂x2 ∂x2 ∂x2 2 ∂x1 ∂x1 ∂x2 ∂x2 ∂x1 ∂x2
In the last step, we computed that
∂u1 ∂u2 ∂u2 ∂u2 ∂u1 ∂u1 ∂u2 ∂u1 ∂u1  ∂u1 ∂u2  ∂u2  ∂u1 ∂u2 
− − + + = − + −
∂x1 ∂x1 ∂x1 ∂x2 ∂x2 ∂x1 ∂x2 ∂x2 ∂x1 ∂x2 ∂x1 ∂x2 ∂x2 ∂x1
= −ω∇ · u = 0. (2.6)

The “miracle” is that in two dimensions the term we have calculated in (2.6), and which in
three dimensions will contribute to the vorticity growth, cancels out completely because of
the incompressibility condition. Thus, in two dimensions, the vorticity satisfies an advection-
diffusion equation
∂ω
+ u · ∇ω = ν∆ω. (2.7)
∂t
This is very remarkable, as (2.7) obeys the maximum principle: with appropriate decay
conditions at infinity if (2.7) is posed in the whole space R2 , or in the periodic case, we can
immediately conclude that
kω(t, ·)kL∞ ≤ kω0 kL∞ , (2.8)
where ω0 (x) = ω(0, x) is the initial condition for the vorticity, as long as u(t, x) satisfies some
very basic regularity assumptions. Furthermore, in an inviscid fluid, when ν = 0 the vorticity
is simply advected along the flow lines; solution of
∂ω
+ u · ∇ω = 0 (2.9)
∂t
is simply
ω(t, x) = ω0 (t, A(t, x)), (2.10)
where A(t, x) is the ”back-to-labels” map for (1.1). This will help us later to prove the
regularity of the solutions of the Euler and Navier-Stokes equations in two dimensions, though
it will not imply the regularity immediately.

11
The Biot-Savart law in two dimensions
Note also that the pressure term is nowhere to be seen in the vorticity equation (2.7). Thus,
in order to close the problem, we only need to supplement the evolution equation (2.7) for
vorticity by an expression for the fluid velocity u(t, x) in terms of the vorticity ω(t, x). To
this end, observe, that, as u(t, x) is divergence free, and the problem is posed in all of R2 ,
there exists a function ψ(t, x), called the stream function, so that u(t, x) has the form
u(t, x) = ∇⊥ ψ(t, x) = (−ψx2 (t, x), ψx1 (t, x)). (2.11)
To see this, note that, because of the divergence-free condition for u(t, x), the flow
v(t, x) = (u2 , −u1 ), (2.12)
satisfies
∂v1 ∂v2
= , (2.13)
∂x2 ∂x1
hence there exists a function ψ(t, x) so that v(t, x) = ∇ψ(t, x), which is equivalent to (2.11).
The vorticity can be expressed in terms of the stream function as
∆ψ = ω, (2.14)
or, more explicitly, ˆ
1
ψ(t, x) = log(|x − y|)ω(t, y)dy. (2.15)
2π R2
Differentiating (2.15) formally, we obtain an expression for the fluid velocity in terms of its
vorticity ˆ
u(t, x) = K2 (x − y)ω(t, y)dy, (2.16)
R2
with the vector-valued integral kernel
1 x2 x1 
K2 (x) = − 2, 2 . (2.17)
2π |x| |x|
Thus, the Navier-Stokes equations in two dimensions can be formulated purely in terms of
vorticity as the advection-diffusion equation for the scalar vorticity
∂ω
+ u · ∇ω = ν∆ω, (2.18)
∂t
with the velocity u(t, x) given in terms of ω(t, x) by (2.16). A potential danger is that
the function K2 (x) is singular, homogeneous of degree (−1) in x. Thus, it is not obvious
that (2.17) gives a sufficiently regular velocity field u(t, x) for the coupled problem to have a
smooth solution even if the initial conditin ω0 (x) = ω(0, x) is smooth and rapidly decaying at
infinity. However, the ”1/x” singularity in two dimensions is sufficiently mild: writing (2.16)
in the polar coordinates gives (with x⊥ = (−x2 , x1 ))
ˆ ˆ ∞ ˆ 2π
1 (x − y)⊥ 1
u(t, x) = ω(y)dy = (− sin φ, cos φ)ω(x1 −r cos φ, x2 −r sin φ)dφdr,
2π R2 |x − y|2 2π 0 0
(2.19)

12
There is no longer a singularity in (2.19), and the expression for the velocity “makes sense”.
The system (2.16), (2.17), (2.18) is an example of an active scalar – the vorticity ω(t, x) is
a solution of an advection-diffusion equation with the velocity coupled to the advected scalar
itself.

Vorticity evolution in three dimensions

The situation in three dimensions is very different. In order to compute the evolution equation
for the vorticity vector, first, note that the advection term in the Navier-Stokes equations can
be written as
∂ui ∂ui ∂uj ∂uj
(u · ∇u)i = uj = uj ( − ) + uj , (2.20)
∂xj ∂xj ∂xi ∂xi
and that

(ω × u)i = εijk ωj uk = εijk εjmn (∂m un )uk = (δin δkm − δim δkn )(∂m un )uk
= (∂k ui )uk − (∂i uk )uk . (2.21)

We used above the identity

εjik εjmn = δim δkn − δin δkm (2.22)
and anti-symmetry of εijk . We see that
 |u|2 
u · ∇u = ω × u + ∇ . (2.23)
2
Therefore, the Navier-Stokes equations can be written as
 |u|2 
ut + ω × u + ∇ + p = ν∆u. (2.24)
2
The formula
curl(a × b) = −a · ∇b + b · ∇a + a(∇ · b) − b(∇ · a) (2.25)
helps us to take the curl of (2.24), leading to the vorticity equation:

ωt + u · ∇ω = ν∆ω + V (t, x)ω, (2.26)

with
∂ui
V (t, x)ω = ω · ∇u, Vij = . (2.27)
∂xj
We can decompose the matrix V into its symmetric and anti-symmetric parts:
1 1
V = D + Ω, D = (V + V T ), Ω = (V − V T ), (2.28)
2 2
and observe that, for any h ∈ R3
1 1 1
Ωij hj = [∂j ui − ∂i uj ]hj = ∂m uk [δik δjm − δim δjk ]hj = εlij εlkm (∂m uk )hj
2 2 2
1 1 1 1
= − εlij εlmk (∂m uk )hj = − εlij ωl hj = εilj ωl hj = [ω × h]i , (2.29)
2 2 2 2

13
that is,
1
Ωh = ω × h. (2.30)
2
The matrix Ω has an explicit form
 
0 −ω3 ω2
1
Ω =  ω3 0 −ω1  . (2.31)
2
−ω2 ω1 0

As a consequence, we have Ωω = 0, thus V ω = Dω, and the vorticity equation has the form

ωt + u · ∇ω = ν∆ω + D(t, x)ω, (2.32)

with
1  ∂ui ∂uj 
Dij = + . (2.33)
2 ∂xj ∂xi
The term Dω in the vorticity equation is known as the vortex stretching term, and it is maybe
the main reason why the solutions of the three- dimensional Navier-Stokes equations exhibit
such rich behavior and complexity. As we have done in two dimensions, it is possible to
express the velocity u(t, x) in terms of the vorticity – this relation is known as the Biot-Savart
law, leading to the “pure vorticity” formulation of the Navier-Stokes equations, but we will
postpone this computation until slightly later.
The evolution of the matrix D itself is obtained by differentiating the Navier-Stokes equa-
tions to get an evolution equation for the matrix V :

∂  ∂ui  ∂  ∂ui  ∂um ∂ui ∂ 2p ∂ui

+ um + + = ν∆ , (2.34)
∂t ∂xj ∂xm ∂xj ∂xj ∂xm ∂xi ∂xj ∂xj

which, in the matrix form is

∂V
+ u · ∇V + V 2 + H = ν∆V, (2.35)
∂t
where H is the Hessian of the pressure. Taking the symmetric part gives
∂D
+ u · ∇D + D2 + Ω2 + H = ν∆D. (2.36)
∂t

An analogy to the Burgers’ equation

The vorticity equation (2.32) has a quadratic term in ω in the right side. Such quadratic
nonlinearities may potentially lead to a blow up. This is easily seen on the simple ODE
example
ż = z 2 , z(0) = z0 . (2.37)
Its explicit solution is
z0
z(t) = . (2.38)
1 − z0 t

14
If z0 > 0, the solution becomes infinite at the time
1
tc = . (2.39)
z0
At a slightly more sophisticated level, we can look at the familiar Burgers’ equation on
the line:
ut + uux = 0, u(0, x) = u0 (x). (2.40)
Its solutions develop a finite time singularity if the initial condition u0 (x) is decreasing on
some interval. Such discontinuities are known as shocks. In order to make a connection to
the vorticity equation, note that the function ω = −ux satisfies

ωt + uωx = ω 2 , ω(0, x) = ω0 (x) = −u00 (x). (2.41)

This equation is analogous to the vorticity equation with ν = 0, except the nonlinearity has
a different form: D(ω)ω is replaced by ω 2 . As in the case of the quadratic ODE (2.37), the
function ω(t, x) becomes infinite in a finite time if there are points where ω0 (x) > 0. One
should mention that there are two regularizations of the inviscid Burgers’ equation (2.40):
first, adding a diffusive (dissipative) term gives the viscous Burgers’ equation

ut + uux = νuxx , u(0, x) = u0 (x), (2.42)

which has global in time smooth solutions if u0 (x) is smooth. A natural question which we
may revisit later is why is the uxx term sufficiently regularizing? More precisely, one may
consider equations of the form

ut + uux = Au, u(0, x) = u0 (x), (2.43)

where A is a linear dissipative operator in the sense that

ˆ
(Au, u) = (Au(x))u(x)dx ≤ 0. (2.44)
R

If A commutes with differentiation, the “vorticity” equation will have the form

ωt + uωx = Aω + ω 2 , ω(0, x) = ω0 (x) = −u00 (x). (2.45)

Then, the dissipative effect of Aω will compete with the growth caused by ω 2 in the right
side. The issue of when the dissipation will win is rather delicate – we will revisit it later if
we have time.
There is a different approach to the blow up in the Burgers’ equation that illustrates a
general strategy of trying to control integral functionals of the solution rather than solutions
themselves. Let us consider, for simplicity, the solution of the Burgers’ equation on the line
with a periodic initial condition u0 (x):

u0 (x + 2π) = u0 (x).

Then the solution of

ut + uux = 0, u(0, x) = u0 (x) (2.46)

15
will stay periodic for all t > 0 (as long as it exists):
u(t, x + 2π) = u(t, x). (2.47)
If, in addition, the initial data is odd: u0 (−x) = −u0 (x), then solution remains odd as well:
we have u(t, x) = −u(t, x) for all t > 0. This means that, as long as the solution remains
smooth, the functional ˆ π
u(t, x)
L(t) = dx (2.48)
−π x
is well-defined and finite – the function u(t, x) vanishes at x = 0. Differentiating L(t) in time
gives ˆ π ˆ π ˆ
dL(t) ut (t, x) 1 1 π u2 (t, x)
= dx = − uux dx = − dx. (2.49)
dt −π x −π x 2 −π x2
The Cauchy-Schwartz inequality implies that
ˆ π 2 ˆ π 2
2 u(t, x) u (t, x)
L (t) = dx ≤ 2π dx. (2.50)
−π x −π x2
Hence, the function L(t) satisfies a differential inequality
dL 1
≤ − L2 (t). (2.51)
dt 4π
Integrating this inequality in time gives
1 1 t
− ≤− . (2.52)
L0 L(t) 4π
Hence, we have
4πL0
L(t) ≤ . (2.53)
4π + L0 t
We conclude that if L0 < 0 then L(t) = −∞ at some time t < −4π/L0 , thus solution may not
remain smooth past this time. The condition that L0 < 0 distinguishes between the initial
data that “look like” u0 (x) = sin x and like u0 (x) = − sin x. The latter is decreasing at x = 0,
hence the shock is expected to form there, thus it is reasonable to expect that L(t), which
has x in the denominator in the integrand, will blow-up. On the other hand, the former is
increasing at x = 0, thus the shock would not form there, and L(t) should not capture the
singularity formation. A different functional should be considered to capture the blow-up.
Another very interesting regularization of the inviscid Burgers’ equation is via dispersion:
ut + uux = µuxxx , u(0, x) = u0 (x). (2.54)
This is the Kortweg-de Vries equation which describes a regime of the shallow water waves. Its
mathematics is incredibly rich and is connected by now with nearly every area of mathematics.
If we have time, we will go back to it as well. For now, we just mention that solutions of (2.54)
also remain smooth for all t > 0 provided that u0 (x) is, say, a smooth rapidly decaying
function. However, the mechanism for regularity is not dissipative but rather dispersive – the
high frequencies spread faster, hence an oscillation will ”fly away towards infinity very fast”,
and there u is small, hence the nonlinearity does not play a big role there. On the other hand,
the balance between dispersion and nonlinearity leads to extremely interesting effects.

16
Flows with spatially homogenous vorticity
As an example, we consider flows that have a spatially uniform vorticity ω(t). Let us choose
a symmetric matrix D(t) with TrD(t) = 0, and a vector-valued function ω(t) 6= 0 such that

dω
= D(t)ω(t), ω(0) = ω0 . (2.55)
dt
We also define the anti-symmetric matrix Ω(t) via (2.31), so that
1
Ω(t)h = ω(t) × h, for any h ∈ R3 , Ωij = εimj ωm . (2.56)
2
A direct computation, using the symmetry of D, the assumption TrD = 0, and (2.31), gives

Ω̇ + DΩ + ΩD = 0. (2.57)

The observation is that the flow

1
u(t, x) = ω(t) × x + D(t)x (2.58)
2
gives an exact solution of the three-dimensional Euler and Navier-Stokes equations, with the
vorticity curlu = ω. Indeed, first, as the trace of D(t) vanishes, both components in (2.58)
are divergence-free:

∇ · u = ∂j (εjkl ωk xl ) + ∂j (Djk xk ) = εjkl ωk δjl + Djk δjk = 0. (2.59)

Moreover, the second term in (2.58) is the gradient of the function (1/2)(D(t)x · x), hence its
vorticity vanishes, while identity (2.25) means that
1 1 1 1 3
curlu = curl(ω(t) × x) = − ω · ∇x + ω(∇ · x) = − ω + ω = ω. (2.60)
2 2 2 2 2
Next, we compute
1
ut = ω̇ × x + Ḋx, (2.61)
2
and
1 1
∂j uk = ∂j (εkmn ωm xn ) + ∂j (Dkm xm ) = εkmj ωm + Dkj , (2.62)
2 2
so that
1 1
u · ∇uk = uj ∂j uk = εkmj uj ωm + uj Dkj = ω × u + Du. (2.63)
2 2
Putting these equations together and using (2.56) leads to
1 1 1
ut + u · ∇u = ω̇ × x + Ḋx + ω × u + Du = ω̇ × x + Ḋx (2.64)
2 2 2
1 1  1 
+ ω× ω × x + Dx + D ω × x + Dx
2 2 2
= (Ḋ + Ω̇ + Ω + D + DΩ + ΩD)x = (Ḋ + Ω2 + D2 )x = −∇p(t, x)
2 2

17
We have used (2.57) in the next to last equality above. The pressure is given explicitly by
1  ∂D 
p(t, x) = − + D2 + Ω2 x · x. (2.65)
2 ∂t
We conclude that, given any symmetric trace-less matrix D(t), we may construct a solution
of the Euler equations as above.
Example 1. A jet flow. As the first example of using the above construction, we may
take ω0 = 0, so that ω(t) = 0 and D(t) = diag(−γ1 , −γ2 , γ1 + γ2 ) with γ1 , γ2 > 0. The flow is

u(t, x) = (−γ1 x1 , −γ2 x2 , (γ1 + γ2 )x3 ). (2.66)

The particle trajectories are

X(t, α) = (e−γ1 t α1 , e−γ2 t α2 , e(γ1 +γ2 )t α3 ), (2.67)

and have the form of a jet, going toward the x3 -axis, and up along this line for x3 > 0, and
down this direction for x3 < 0.
Example 2. A strain flow. Consider D = diag(−γ, γ, 0) with γ > 0, and, once again,
vorticity ω = 0, so that
u(t, x) = (−γx1 , γx2 , 0). (2.68)
Then the particle trajectories are

X(t, α) = (e−γt α1 , eγt α2 , α3 ). (2.69)

The particle trajectories stay in a fixed plane orthogonal to the x3 -axis and are stretched in
this plane: nearby two particles starting near the x1 -axis with α2 > 0 and α2 < 0 will separate
exponentially fast in time.

Shear layer solutions

Here, we will generalize the second example above: we will be looking at flows of the form
generalizing (2.68):
u(t, x) = (−γx1 , γx2 , w(t, x1 )), (2.70)
that is, the third flow component depends only on x1 and t. Such flows satisfy the Navier-
Stokes equations with the pressure p(t, x) = γ(x21 +x22 )/2, provided that the vertical component
of the flow w satisfies a linear advection-diffusion equation
∂w ∂w ∂ 2w
− γx1 =ν . (2.71)
∂t ∂x1 ∂x1 2
The vorticity is given by
∂w
ω(t, x) = (0, − , 0), (2.72)
∂x1
and its second component ω̃ = −wx1 satisfies (after dropping the tilde)

∂ω ∂ω ∂ 2ω
− γx1 =ν + γω. (2.73)
∂t ∂x1 ∂x1 2

18
Here, we see clearly the three competing effects in the vorticity evolution: the diffusive (dis-
sipative) term νωx1 x1 , the convective term −γx1 ωx1 and the vorticity growth term γω. It is
instructive to look at the three effects in this very simple setting.
First, let us note that when γ > 0, the vorticity equation (2.73) admits steady solutions:

−γx1 ω̄ 0 = ν ω̄ 00 + γ ω̄. (2.74)

Indeed, setting y = λx1 leads to

−γy ω̄y = λ2 ν ω̄yy + γ ω̄, (2.75)

p
thus, choosing λ = γ/ν, we arrive at

−y ω̄y = ω̄yy + ω̄. (2.76)

This equation has an explicit steady solution

2 /2
ω̄(y) = e−y , (2.77)

hence a steady solution of (2.74) is

2
ω̄(x1 ) = e−γx1 /(2ν) . (2.78)

Such solutions do not existpwhen γ = 0 – they are sustained by the stretch, and are localized
in a layer of the width O( ν/γ) around the plane {x1 = 0}. They may also not exist at zero
viscosity: if γ = 0 then (2.74) has no non-trivial bounded steady solutions – thus, they are a
result of a balance between the stretch and the friction.
Equation (2.73) can be solved explicitly. Fitst, writing

ω(t, x) = eγt z(t, x1 ) (2.79)

gives
∂z ∂z ∂ 2z
− γx1 =ν . (2.80)
∂t ∂x1 ∂x1 2
Next, making a change of variables:

z(t, x) = η(τ (t), eγt x1 ) (2.81)

with the function τ (t) to be determined, leads to

∂η ∂η ∂η ∂ 2η
τ̇ + γeγt x1 − γx1 eγt = νe2γt 2 . (2.82)
∂τ ∂ξ ∂ξ ∂ξ
Taking
τ̇ = νe2γt , (2.83)
or
ν  2γt 
τ (t) = e −1 , (2.84)
2γ

19
leads to the standard heat equation

∂η ∂ 2η
= 2 , τ > 0, ξ ∈ R, (2.85)
∂τ ∂ξ

with the initial condition η(0, ξ) = ω0 (ξ). Therefore, the vorticity is

ˆ 
γt ν 2γt 
ω(t, x1 ) = e G (e − 1), eγt x1 − y ω0 (y)dy, (2.86)
2γ

where G(t, x1 ) is the standard heat kernel:

1 −|x1 |2 /(4t)
G(t, x1 ) = √ e . (2.87)
4πt
Let us look at the long time behavior of vorticity:
−1/2 ˆ n |eγt x − y|2 o
γt 4πν 2γt

1
ω(t, x1 ) = e (e − 1) exp − 4ν 2γt ω0 (y)dy (2.88)
2γ (e − 1)
 γ 1/2 ˆ 2γ
2
→ ω̄(x) = e−γ|x1 | /(2ν) ω0 (y)dy,
2πν
1
provided that the initial vorticity ω0 ∈ Lp (R). Thus, the vorticity is localized as t → +∞
around x1 = 0, in a layer of the width O( ν/γ), and its long time limit is a multiple of the
steady solution (2.78).

The Biot-Savart law

We now return to the vorticity equation in three dimensions

ωt + u · ∇ω = ν∆ω + ω · ∇u. (2.89)

Our goal is to derive an expression for the velocity u in terms of the vorticity ω, so as
to formulate the Euler and Navier-Stokes equations purely in terms of vorticity. In two
dimensions, this was done using the stream function, solution of

∆ψ = ω, (2.90)

with u given by
u = ∇⊥ ψ = (−ψx2 , ψx1 ), (2.91)
or, equivalently, ˆ
u(t, x) = K2 (x − y)ω(y)dy, (2.92)
R2
with the vector-valued integral kernel
1 x2 x1 
K2 (x) = − 2, 2 . (2.93)
2π |x| |x|

20
 In three dimensions, given a divergence-free vector field ω(x) we need to find a divergence-
free vector field u(t, x) so that

∇ × u = ω, ∇ · u = 0. (2.94)

Attempting the same strategy as in two dimensions, we define the stream vector ψ via

∆ψ = ω, (2.95)

and
u(x) = −∇ × ψ(x). (2.96)
Note that, as ∇ · ω = 0 by assumption, we have

∆(∇ · ψ) = 0. (2.97)

Hence, if we assume that ∇ · ψ is bounded, then ∇ · ψ = 0, and ψ is also divergence-free. The

flow u defined by (2.96) is divergence-free: ∇ · u = 0, and

[∇ × u]i = εijk ∂j uk = −εijk ∂j εkmn ∂m ψn = −εkij εkmn ∂j ∂m ψn = −(δim δjn − δin δjm )∂j ∂m ψn
= −∂i ∂j ψj + ∆ψi , (2.98)

that is, ω is the vorticity of u:

∇ × u = −∇(∇ · ψ) + ∆ψ = ω. (2.99)

We have an explicit expression for the stream-vector ψ(x) as the solution of the Poisson
equation (2.95): ˆ
1 1
ψ(x) = − ω(y)dy. (2.100)
4π R3 |x − y|
The velocity is then given by
ˆ ˆ
1  1  1 xj − yj
ui (x) = εijk ∂j ωk (y)dy = − εijk ωk (y)dy, (2.101)
4π R3 |x − y| 4π R3 |x − y|3
so that
ˆ
1
u(x) = K(x − y) × ω(y)dy, (2.102)
4π R3

with
1 x
K(x) = − . (2.103)
4π |x|3
As in the two-dimensional case, the integral operator defining u(x) in terms of the vortic-
ity ω(x) is not “really singular” – the singularity of the 1/|x|2 type is cancelled in three
dimensions by the Jacobian if we pass to the spherical coordinates. However, unlike in two
dimensions, the vorticity equation in three dimensions

ωt + u · ∇ω = ν∆ω + ω · ∇u, (2.104)

21
involves not only u(x) but also the gradient ∇u. Formally differentiating (2.102) leads to
(this identity is not quite correct because of the singularity of the integrals involved)
ˆ
∇u(x)” = ” ∇K(x − y) × ω(y)dy. (2.105)
R3

The integral kernel ∇K(x) in (2.105) has the singularity of the type x/|x|4 , which can not be
simply cancelled by the Jacobian in three dimensions if we pass to the spherical coordinates.
Integral operators with a singularity of this type are known as singular integral operators,
and we will deal with them in some detail later, leaving for now the vorticity equation on a
formal level.

3 The conserved quantities

We will now discuss the physical quantities conserved by the Euler and Navier-Stokes equa-
tions. They are important both from the physical and mathematical points of view – a system
that possesses sufficiently regular integrals of motion will not have irregular solutions if the
initial condition is smooth. As we will see, the integrals of motion for the fluid equations are
often insufficient to deduce the existence and regularity of solutions.

Kelvin’s theorem
Consider a smooth, oriented, closed curve C0 , and let C(t) be its image under a flow u(t, x):

C(t) = {X(t, α) : α ∈ C0 }, (3.1)

with
dX
= u(t, X), X(0, α) = α. (3.2)
dt
The circulation around C(t) is
˛
ΓC(t) = u(t, x) · d`, (3.3)
C(t)

where d` is the length element along Γ(t). Let us parametrize the initial and evolved curves
as
C0 = {γ(s), 0 ≤ s ≤ 1}, C(t) = {X(t, γ(s)), 0 ≤ s ≤ 1}, (3.4)
then the length element along the evolved curve has the components (prime denotes the
derivative with respect to the parametrization parameter s)
∂Xj 0
d`j = γ ds, (3.5)
∂γk k
or C 0 (t, s) = H(t, X(t, γ(s))γ 0 (s), with the matrix

∂Xi (t, α)
Hij (t, X(t, α)) = , (3.6)
∂αj

22
which, as we recall, satisfies (1.3)
dH
= (∇u)H. (3.7)
dt
Now, we may compute
˛ ˆ ˆ 1
d d 1 0
u(t, x) · d` = u(t, X(t, γ0 (s)) · (Hγ )ds = [(u̇ · Hγ 0 ) + (u · Ḣγ 0 )]ds
dt C(t) dt 0 0
ˆ 1
= [(ut + u · ∇u) · Hγ 0 ) + (u · (∇uH)γ 0 )]ds (3.8)
˛0 ˛
= (ut + u · ∇u) · d` + (∇u)t u · d`.
C(t) C(t)

If u satisfies the Euler equations, we have for the first term in the last line above:
˛ ˛
(ut + u · ∇u) · d` = − ∇p · d` = 0. (3.9)
C(t)

The second term can be written as

˛ ˛ ˛  |u|2 
t ∂uk
(∇u) )u · d` = uk d`j = ∇ · d` = 0. (3.10)
C(t) C(t) ∂xj C(t) 2

We see that
˛
d
u(t, x) · d` = 0. (3.11)
dt C(t)

This is Kelvin’s theorem for the Euler equations: the circulation of the flow along a curve
that evolves with the flow is preserved in time.

Conservation of the total velocity and vorticity

If u is a divergence-free velocity field, and q is a scalar function, and both of them decay
sufficiently fast at infinity, we have
ˆ ˆ
(u · ∇φ)dx = − (∇ · u)φdx = 0. (3.12)
Rn

Therefore, integrating either the Euler or the Navier-Stokes equations with solutions that
decay rapidly at infinity, we conclude that
ˆ
d
udx = 0, (3.13)
dt Rn
both in two and three dimensions. The same identity implies that in two dimensions the total
vorticity is preserved: integrating (2.18), we obtain
ˆ ˆ ˆ
d
ωdx = −ν ∆ωdx − (u · ∇ω)dx = 0. (3.14)
dt R2 R2 R2

23
However, in that case we know more: any regular solution of (2.18) can be decomposed as

ω = ω + (t, x) − ω − (t, x),

where ω ± are the solutions of (2.18) with the initial conditions ω0± (x), respectively. It follows
that ˆ ˆ ˆ ˆ
+ −
|ω|dx ≤ ω (t, x)dx + ω (t, x)dx = |ω0 |dx, (3.15)
R2 R2 R2 R2

that is, not only the integral of the vorticity is preserved but its L1 -norm does not grow in
two dimensions.
In addition, for the solutions of the Euler equations in two dimensions, vorticity satisfies
the advection equation
ωt + u · ∇ω = 0. (3.16)
Therefore, not only the integral of the vorticity but all Lp -norms of ω are preserved, with
any 1 ≤ p ≤ ∞: ˆ ˆ
p
|ω(t, x)| dx = |ω0 (x)|p dx. (3.17)
R2 R2

In three dimensions, the vorticity vector satisfies (2.104). Integrating this equation leads
to ˆ ˆ
d
ωi dx = (ω · ∇ui )dx = 0, (3.18)
dt R3 R3

since ω(t, x) is also a divergence-free field. Thus, the total integral of the vorticity is preserved
also in three dimensions. However, conservation of the Lp -norms does not follow, and vorticity
may grow.

Evolution of energy, dissipation and enstrophy

The kinetic energy of the fluid is
ˆ
1
E(t) = |u(t, x)|2 dx. (3.19)
2 Rn

Differentiating in time, assuming that solutions are smooth, gives

ˆ ˆ
dE ∂uj
= (u · ut )dx = (−uj uk − u · ∇p + νuj ∆uj )dx
dt Rn Rn ∂xk
ˆ  |u|2  ˆ ˆ
2
=− (u · ∇ +p −ν |∇u| dx = −ν |∇u|2 dx. (3.20)
Rn 2 R n Rn

Therefore, the energy of the solutions of the Euler equations (ν = 0) is preserved in time:

E(t) = E(0), (3.21)

while the energy of the solutions of the Navier-Stokes equations is dissipating:

dE
= −νD(t), (3.22)
dt

24
where D(t) is the enstrophy ˆ
D(t) = |∇u|2 dx. (3.23)
Rn
The enstrophy can be expressed purely in terms of vorticity using the identity

|ω|2 = εijk εimn (∂j uk )(∂m un ) = (δjm δkn − δjn δkm )(∂j uk )(∂m un ) = |∇u|2 − (∂j uk )(∂k uj ). (3.24)

Note that ˆ ˆ
(∂j uk )(∂k uj )dx = − uk (∂k ∂j uj )dx = 0. (3.25)
Rn Rn
We used the incompressibility condition on u in the last step. This implies that the enstrophy
for a divergence-free flow is ˆ
D(t) = |ω|2 dx. (3.26)
Rn
Therefore, large vorticity leads to increased energy dissipation – this, however, does not
automatically lead to regularity.
An important comment is that the above computations assume that the solution u(t, x)
of the Navier-Stokes equations is sufficiently smooth. The possibility of energy dissipation as
the solutions potentially develop a singularity is an extremely important open question.

Conservation of helicity
The helicity of a flow is ˆ
H= (u · ω)dx. (3.27)
R3
This definition is non-trivial only in three dimensions, as in two dimensions we have, for any
incompressible flow,
ˆ ˆ ˆ
 ∂u
2 ∂u1   ∂u
1 1 ∂(u21 ) 
u1 ωdx = u1 − dx = − u2 + dx
∂x1 ∂x2 ∂x1 2 ∂x2
R2
ˆ
R2 R2
1 ∂
= (u2 − u21 )dx = 0, (3.28)
2 R2 ∂x2 2
with a similar computaiton for u2 . In three dimensions, however, helicity is a non-trivial
quantity, and, for the solutions of the Euler equations, we may compute
ˆ
dH
= (ut · ω + u · ωt )dx. (3.29)
dt R3

We have
ut · ω + (u · ∇u) · ω + ω · ∇p = 0, (3.30)
and
u · ωt + (u · ∇ω) · u = u · (ω · ∇u). (3.31)
The last term in (3.30) integrates to zero since ∇ · ω = 0:
ˆ
(ω · ∇p)dx = 0. (3.32)
R3

25
The other terms lead to
ˆ
dH
=− (uk (∂k uj )ωj + uk uj ∂j ωk − uj ωk ∂k uj )dx = (3.33)
dt
ˆ R3

1
=− (−uk uj ∂k ωj + uk uj ∂j ωk + |u|2 ∂k ωk )dx = 0. (3.34)
R3 2
Here, we have integrated by parts in the first term in the right side and used incompressibility
of u to show that the first two terms in the right side cancel each other, while the last term
vanishes after integration by parts because ∇ · ω = 0. Thus, helicity is preserved for the
solutions of the Euler equations. In particular, the velocity field and the vorticity can not be
”too aligned” in any growth or blow-up scenario for the Euler equations.

4 The Constantin-Lax-Majda toy model

The formulation of the model
In order to appreciate the difficulties of the problem of the regularity for the solutions of the
Euler and the Navier-Stokes equations, and in particular, focus on the effect vortex stretching
term, we consider here a toy model studied by Constantin, Lax and Majda in 1985. The
vortex stretching term in the three-dimensional vorticity equation for the Euler equation

ωt + u · ∇ω = ω · ∇u, (4.1)

has the form (2.105) – once again, it should not be taken too literally because of the singularity
in the integral,
ˆ
∇u(x)” = ” ∇K(x − y) × ω(y)dy, (4.2)
R3

with
1 x
K(x) = − . (4.3)
4π |x|3
The Constantin-Lax-Majda model aims to imitate three important properties of the right side
in the vorticity equation (4.1): first, it is quadratic in ω, second, its integral vanishes:
ˆ
ω · ∇u dx = 0. (4.4)
R3

The third feature is that the kernel ∇K(x) has the singularity of the type x/|x|4 , which
is of the kind x/|x|n+1 in n dimensions that is ”barely non-integrable”. Integral operators
with such kernels are known as Calderon-Zygmund operators. Constantin, Lax and Majda
considered a one-dimensional model, with an analogous singularity in one dimension

∂ω(t, x)
= H[ω]ω, x ∈ R, (4.5)
∂t

26
with the initial condition ω(0, x) = ω0 (x). Here, H(ω) is the Hilbert transform, a singular
integral operator in one dimension:
ˆ
1 ω(y)
H[ω](x) = P.V. dy. (4.6)
π R x−y

The principal value above is understood as

ˆ ˆ ˆ
1 ω(x − y) 1 ω(x − y) 1 1 ω(x − y) − ω(x)
H[ω](x) = lim dy = dy + dy. (4.7)
π ε↓0 |y|>ε y π |y|>1 y π −1 y

The singularity 1/x in the kernel of the one-dimensional Hilbert transform is analogous to
the singularity x/|x|4 in three dimensions that appears in the kernel ∇K in (4.2): both are
odd, and their size is 1/|x|n .

The toyest model of all

Before proceeding with the analysis of the Constantin-Lax-Majda model, let us pause and
see what would happen if we would consider the simplest model that would preserve only the
quadratic nature of the nonlinearity in the vorticity equation:
dω(t, x)
= ω 2 (t, x), ω(0, x) = ω0 (x), x ∈ R. (4.8)
dt
Its explicit solution is
ω0 (x)
ω(t, x) = . (4.9)
1 − tω0 (x)
If there exist x ∈ R so that ω0 (x) > 0, this solution makes sense until the denominator
vanishes, that is, until the time
h 1 i
Tc = inf : ω0 (x) > 0 . (4.10)
ω0 (x)

Let us assume that the function ω0 (x) attains its maximum at x = xm , so that Tc = 1/ω0 (xm ).
The function ω(t, x) at the time t = Tc has an asymptotic expansion near the point x = xm :

ω0 (x) ω0 (xm )
ω(Tc , x) = ≈ . (4.11)
1 − Tc ω0 (x) −(Tc /2)ω000 (xm )(x − xm )2

Thus, the function ω(t, x) blows up at the point xm and the blow-up profile is O(x − xm )−2 .
As a consequence, all Lp -norms of ω(t, x) blow up as well:
ˆ
|ω(t, x)|p dx → +∞ as t ↑ Tc , (4.12)
R

for all p ≥ 1. Moreover, if we define the “velocity” as

ˆ x
v(t, x) = ω(t, y)dy, (4.13)
−∞

27
then v(t, x) also blows-up at the time Tc and its blow-up profile is O(x − xm )−1 . Therefore,
the Lp -norm of the velocity blows up as well:
ˆ
|v(t, x)|p dx → +∞ as t ↑ Tc , (4.14)
R

for all p ≥ 1. In particular, the kinetic energy blows up:

ˆ
|v(t, x)|2 dx → +∞ as t ↑ Tc . (4.15)
R

This is in contrast to the energy conservation in the true Euler equations. Thus, the toy
model (4.8) can not be even “toyishly” correct.

The Hilbert transform

In order to understand the Constantin-Lax-Majda model, let us first recall some basic prop-
erties of the Hilbert transform and its alternative definition in terms of complex analysis.
Given a Schwartz class function f (x) ∈ S(R) define a function
ˆ
u(x, y) = e−2πy|ξ| fˆ(ξ)e2πixξ dξ, y ≥ 0, x ∈ R.
R

Here, the Fourier transform is defined as

ˆ ˆ
ˆ
f (ξ) = f (x)e −2πiξx
dx, f (x) = fˆ(ξ)e2πiξx dξ. (4.16)
R R

The function u(x, y) is harmonic in the upper half plane:

∆x,y u = 0 in R2+ = R × (0, +∞),

and satisfies the boundary condition on the line y = 0:

u(x, 0) = f (x), x ∈ R.

We can write u(x, y) as a convolution

ˆ
u(x, y) = Py ? f = Py (x − x0 )f (x0 )dx0 ,

with
P̂y (ξ) = e−2πy|ξ| ,
and
ˆ ∞
1 1 y
Py (x) = e−2πy|ξ| e2πiξx dξ = + = .
−∞ 2π(y − ix) 2π(y + ix) π(x + y 2 )
2

Next, set z = x + iy and write

ˆ ˆ ∞ ˆ 0
−2πy|ξ| ˆ
u(z) = e f (ξ)e 2πixξ
dξ = fˆ(ξ)e2πizξ dξ + fˆ(ξ)e2πiz̄ξ dξ.
R 0 −∞

28
Consider the function v(z) given by
ˆ ∞ ˆ 0
iv(z) = ˆ
f (ξ)e2πizξ
dξ − fˆ(ξ)e2πiz̄ξ dξ.
0 −∞

Note that, as f (x) is real-valued, we have fˆ(ξ) = fˆ(−ξ), thus v(z) is real-valued:
ˆ ∞ ˆ 0
iv̄(z) = − ˆ
f (ξ)e −2πiz̄ξ
dξ + fˆ(ξ)e−2πizξ dξ
0 −∞
ˆ ∞ ˆ 0 ˆ ∞ ˆ 0
=− ˆ
f (−ξ)e −2πiz̄ξ
dξ + ˆ
f (−ξ)e −2πizξ
dξ = ˆ
f (ξ)e 2πizξ
dξ − fˆ(ξ)e2πiz̄ξ dξ
0 −∞ 0 −∞
= iv(z).

Moreover, as the function ˆ ∞

u(z) + iv(z) = fˆ(ξ)e2πizξ dξ
0
is analytic in the upper half-plane {Imz > 0}, the function v is the harmonic conjugate of u.
It can be written as
ˆ
v(z) = (−isgn(ξ))e−2πy|ξ| fˆ(ξ)e2πixξ dξ = Qy ? f,
R

with
Q̂y (ξ) = −isgn(ξ)e−2πy|ξ| , (4.17)
and ˆ ∞
1 x
Qy (x) = −i sgn(ξ)e−2πy|ξ| e2πiξx dξ = .
−∞ π x + y2
2

The Poisson kernel and its conjugate are related by

i 1
Py (x) + iQy (x) = = ,
π(x + iy) iπz
which is analytic in {Imz ≥ 0}.
In order to consider the limit of Qy as y → 0, we relate it to the principal value of 1/x
defined as in (4.7): it is an element of the space S 0 (R) of the Schwartz distributions, defined
by
ˆ ˆ ˆ
1 φ(x) φ(x) − φ(0) φ(x)
P.V. (φ) = lim dx = dx + dx, φ ∈ S(R), (4.18)
x ε→0 |x|>ε x |x|<1 x |x|>1 x

which is well-defined for φ ∈ S(R). The conjugate Poisson kernel Qy and the principal value
of 1/x are related as follows.
1 x
Proposition 4.1 Let Qy = , then for any function φ ∈ S(R)
π x + y2
2

ˆ
1 1
P.V. (φ) = lim Qy (x)φ(x)dx.
π x y→0 R

29
Proof. Let
1
ψy (x) = χy<|x| (x)
x
so that ˆ
1
P.V. (φ) = lim ψy (x)φ(x)dx.
x y→0 R

Note, however, that

ˆ ˆ ˆ
xφ(x) φ(x)
(πQy (x) − ψy (x))φ(x)dx = 2 2
dx − dx
R x +y |x|>y x
ˆ ˆ  
xφ(x) x 1
= 2 2
dx + 2 2
− φ(x)dx (4.19)
|x|<y x + y |x|>y x + y x
ˆ ˆ ˆ ˆ
xφ(xy) y 2 φ(x) xφ(xy) φ(xy)
= 2
dx − 2 2
dx = 2
dx − 2
dx.
|x|<1 x + 1 |x|>y x(x + y ) |x|<1 x + 1 |x|>1 x(x + 1)

The dominated convergence theorem implies that both integrals on the utmost right side
above tend to zero as y → 0. 2
It is important to note that the computation in (4.19) worked only because the kernel 1/x
is odd – this produces the cancellation that saves the day. This would not happen, for instance,
for a kernel behaving as 1/|x| near x = 0.
Thus, the Hilbert transform defined as
ˆ
1 f (x − y)
Hf (x) = lim dy. (4.20)
π ε→0 |y|>ε y

can be also written as

Hf (x) = lim Qy ? f (x). (4.21)
y→0

In other words, we take the function f (x), extend it as a harmonic function u(x, y) to the
upper half-plane, and find the conjugate harmonic function v(x, y). Then, Hf (x) = v(x, 0),
the restriction of v(x, y) to the real axis. It follows from (4.17) that

d(ξ) = lim Q̂t (ξ)fˆ(ξ) = −isgn(ξ)fˆ(ξ).

Hf (4.22)
t↓0

Therefore, the Hilbert transform may be extended to an isometry L2 (R) → L2 (R), with

kHf kL2 = kf kL2 , H(Hf ) = −f , (4.23)

and ˆ ˆ
(Hf )(x)g(x)dx = − f (x)(Hg)(x)dx. (4.24)

Back to the Constantin-Lax-Majda model

Let us now return to the CLM model

ωt = H[ω]ω, ω(0, x) = ω0 (x). (4.25)

30
The term H[ω]ω in the right side of (4.25) is similar to the vorticity stretching term Dω in
the true three-dimensional vorticity equation in the three aspects we have discussed above,
below (4.3). It is quadratic in ω, it follows from (4.24) that the operator H is skew-symmetric:
ˆ
H[ω](x)ω(x)dx = 0, (4.26)
R

so the right side of (4.25) integrates to zero, as in (4.4), and the kernel 1/x has the correct
singularity – it is odd and of the size 1/|x|n (where n is the dimension). It follows from (4.26)
that the integral of the solution of the toy model (4.25) is preserved:
ˆ
d
ω(t, x)dx = 0. (4.27)
dt R
Given a function φ, let us now use the “complex analysis” definition of ψ = H[φ], and
set u(x, y) and v(x, y) so that the function f = u+iv is analytic in {y > 0}, with the boundary
values u(x, 0) = φ(x), v(x, 0) = ψ(x). As we may write

−if 2 = 2uv + i(v 2 − u2 ), (4.28)

it follows that the harmonic conjugate of uv is (v 2 − u2 )/2. Restricting this identity to the
real line gives
1 1
H(φH[φ]) = (H[φ])2 − φ2 . (4.29)
2 2
Applying the Hilbert transform to the toy vorticity equation gives then
d 1 ω2
H[ω] = (H[ω])2 − . (4.30)
dt 2 2
Therefore, the function
w(t, x) = H[ω](t, x) + iω(t, x) (4.31)
satisfies the simple quadratic ODE
dw 1 1 1
= (H[ω])2 − ω 2 + iH[ω]ω = w2 . (4.32)
dt 2 2 2
Hence, the function w(t, x) is given explicitly by
w0 (x)
w(t, x) = . (4.33)
1 − 21 tw0 (x)
Taking the imaginary part of (4.33) gives an explicit formula for the solution of the toy
vorticity equation:
w0 (x) 2(H[ω0 ](x) + iω0 (x))
ω(t, x) = Im 1 = Im
1 − 2 tw0 (x) 2 − t(H[ω0 ](x) + iω0 (x))
2(H[ω0 ](x) + iω0 (x))(2 − tH[ω0 ](x) + itω0 (x))
= Im
(2 − tH[ω0 ](x))2 + t2 (ω0 (x))2
tω0 (x)H[ω0 ](x) + ω0 (x)(2 − tH[ω0 ](x)) 4ω0 (x)
=2 = . (4.34)
(2 − tH[ω0 ](x))2 + t2 (ω0 (x))2 (2 − tH[ω0 ](x))2 + t2 (ω0 (x))2

31
The explicit formula
4ω0 (x)
ω(t, x) = . (4.35)
(2 − tH[ω0 ](x))2 + t2 (ω0 (x))2
gives an explicit criterion for the solution of the vorticity to exist for all times t > 0. Namely,
the solution ω(t, x) exists and remains smooth provided that there does not exist a point x ∈ R
so that both ω0 (x) = 0 and H[ω0 ](x) > 0. The explicit breakdown time for a smooth solution
is then n 2 o
Tc = inf : ω0 (x) = 0, H[ω0 ](x) > 0 . (4.36)
H[ω0 ](x)
As an example, consider ω0 (x) = cos x, so that H[ω0 ](x) = sin x, and
4 cos x 4 cos x
ω(t, x) = = . (4.37)
(2 − t sin x)2 + t2 cos2 x 4 + t2 − 4t sin x
The breakdown time Tc = 2, at the point x = π/2, and the corresponding “toy velocity” is
ˆ x
1 t2
v(t, x) = ω(t, y)dy = log(1 + − t sin x). (4.38)
0 t 4
Therefore, ˆ π
|ω(t, x)|p dx → +∞ (4.39)
−π
as t ↑ Tc , for any 1 ≤ p < ∞. On the other hand, the Lp -norms of the velocity stay finite:
ˆ π
|v(t, x)|p dx → Mp < +∞, (4.40)
−π

for all 1 ≤ p < +∞, as t →↑ Tc . In particular, the kinetic energy does not blow-up at the
time Tc : ˆ π
|v(t, x)|2 dx → M2 < +∞, (4.41)
−π

This is in contrast to what happens in the “most toyest” model (4.8), where, the kinetic
energy blows up at the blow-up time. Thus, while the Constantin-Lax-Majda model does
not necessarily capture the physics of the Euler equations, it provides a “reasonable” one-
dimensional playground.

5 The weak solutions of the Navier-Stokes equations

We will now start looking at the existence and regularity of the solutions of the Navier-Stokes
equations. In order to focus on the less technical points, we will consider the periodic solutions
of the Navier-Stokes equations:

ut + u · ∇u − ν∆u + ∇p = f (t, x),

∇ · u = 0, (5.1)
u(0, x) = u0 .

32
Here, f is the forcing term, and u0 (x) is the initial condition. We assume both to be 1-periodic
in all directions: f (t, x + ej ) = f (t, x), u0 (x + ej ) = u0 (x), with j = 1, 2 in R2 and j = 1, 2, 3
in R3 . We will look for periodic solutions of (5.1) in Rn , n = 2, 3.
Note that if f = 0 then the integral of u is conserved if f = 0:
ˆ
hui = u(t, x)dx = 0. (5.2)
Tn

Here, Tn = [0, 1]n is the unit period. When f 6= 0, (5.2) holds, provided that hf i = 0 for
all t ≥ 0. Generally, we have a separate equation for hui:

dhui
= hf i, (5.3)
dt
hence ū(t) = hu(t, ·)i is explicit:
ˆ t
ū(t) = ū(0) + hf (s, ·)ids.
0

Then, we set ˆ t
X(t) = ū(s)ds,
0
and observe that
v(t, x) = u(t, x + X(t)) − ū(t),
satisfies

vt + v · ∇v − ν∆v + ∇p = g(t, x), ,

∇ · v = 0, (5.4)
v(0, x) = v0 ,

with
g(t, x) = f (t, x + X(t)) − hf (t, ·)i,
so that both v0 (x) and g(t, x) are still 1-periodic in x, and hg(t, ·)i = 0 for all t ≥ 0. Thus,
we may assume without loss of generality that hf i = 0, and (5.2) holds.
The two and three dimensional cases are very different. In two dimensions, we will be
able to show existence of regular solutions for all t > 0, provided that the forcing f (t, x) and
the initial condition u0 (x) are sufficiently regular. On the other hand, in three dimensions,
we will only be able to show that there exists a time Tc that depends on the force f and the
initial condition u0 so that the solution of the Navier-Stokes equations remains regular until
the time Tc . However, if the initial data and the forcing are sufficiently small (in a sense to
be made precise later), then solutions of the Navier-Stokes equations remain regular for all
times t > 0. This will be shown using the dominance of diffusion over the nonlinearity for
small data.

33
The weak solutions
The distinction between two and three dimensions is less dramatic if we talk about weak
solutions. As is usual in the theory of weak solutions of partial differential equations, the
definition of a weak solution of the Navier-Stokes equations (5.1) comes from multiplying it
by a smooth test function and integrating by parts. First, we note that any test vector field ψ
can be decomposed as a sum of a gradient field and a divergence-free field:

ψ(x) = φ(x) + ∇η(x), (5.5)

with ∇ · φ(x) = 0. This is known as the Hodge decomposition. In the periodic case it is quite
explicit: write ψ(x) in terms of the Fourier transform
X
ψ(x) = ψk e2πik·x , (5.6)
k∈Zn

and set
X (ψk · k) 2πik·x X (ψk · k)
η(x) = 2
e , ∇η(x) = 2
ke2πik·x . (5.7)
k∈Zn ,k6=0
2πi|k| k∈Zn ,k6=0
|k|

Then, the Fourier coefficients of the difference

X (ψk · k)  2πik·x
φ(x) = ψ(x) − ∇η(x) = (ψk − k e (5.8)
k∈Zn ,k6=0
|k|2

are
X (ψk · k)
φ(x) = φk e2πik·x , φk = ψk − k. (5.9)
k∈Zn
|k|2
They satisfy
φk · k = 0, (5.10)
thus the vector field φ(x) is divergence-free:

∇ · φ(x) = 0. (5.11)

Let now u(t, x) be a smooth solution of the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u + g, (5.12)
∇ · u = 0. (5.13)

We will also decompose the forcing term

g = f + ∇ζ with ∇ · f = 0. (5.14)

The first observation is that if we multiply (5.12) by ∇η(x) and integrate, then we simply
get the Poisson equation for the pressure. Indeed, if w is a smooth periodic vector field,
and ∇ · w = 0, then
ˆ ˆ
w(x) · ∇η(x)dx = − η(x)(∇ · w)(x)dx = 0. (5.15)
Tn Tn

34
It follows that ˆ ˆ
(ut · ∇η)dx = (∆u · ∇η)dx = 0. (5.16)
Tn Tn
For the pressure we have: ˆ ˆ
(∇p · ∇η)dx = − p∆ηdx, (5.17)
Tn Tn
while for the nonlinear term we get, after an integration by parts, using the divergence-free
condition on u:
ˆ ˆ ˆ
((u · ∇u) · ∇η)dx = uj (∂j uk )∂k ηdx = − uj uk (∂j ∂k η)dx. (5.18)
Tn Tn Tn

We deduce that, for any test function η(x), we have

ˆ ˆ ˆ
(p∆η + uj uk (∂j ∂k η)dx = g · ∇η = ∇ζ · ∇η. (5.19)
Tn Tn Tn

This is the weak form of the Poisson equation

−∆p = (∂j uk )(∂k uj ) − ∆ζ. (5.20)

On the other hand, when we multiply (5.12) by a divergence-free smooth vector field w(x),
the pressure term disappears: ˆ
(w · ∇p)dx = 0, (5.21)
Tn
and the nonlinear term may be written as
ˆ ˆ ˆ
((u · ∇u) · w)dx = uj (∂j uk )wk dx = − uj uk ∂j wk dx. (5.22)
Tn Tn Tn

Thus, if w is a C ∞ (Tn ) periodic divergence-free field, integration by parts gives

ˆ ˆ ˆ
[ut · w − uj uk ∂j wk ]dx = ν (u · ∆w)dx + (f · w)dx. (5.23)
Tn Tn Tn

For now, we say that u(t, x) is a weak solution of the Navier-Stokes equations if (5.23) holds
for all periodic smooth divergence-free vector fields w(x). A little later, we will make this
notion more precise, setting up the proper spaces in which the weak solutions live, and relaxing
the C ∞ assumption on the test function.

The Galerkin approximation

In order to construct the weak solutions, we will consider the Galerkin approximation of
the Navier-Stokes equations. In the periodic case, this is equivalent to the projection of the
equations on the divergence-free Fourier modes with |k| ≤ m, where m > 0 is fixed. That is,
given a vector-field X
ψ(x) = ak e2πik·x , (5.24)
k∈Zn

35
we set
X  (ak · k)  2πik·x
ψ (m) (x) = Pm ψ(x) = ak − k e , (5.25)
|k|2
|k|≤m

so that, in particular,
∇ · ψ (m) = 0. (5.26)
Note that if ψ is a divergence-free vector field then ψ (m) is simply the projection on the Fourier
modes with |k| ≤ m.
The Galerkin approximation of the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u + f, (5.27)

with u(0, x) = u0 (x), and a divergence-free force f : ∇ · f = 0, is the system

∂u(m) (m)
+ Pm (u(m) · ∇u(m) ) = ν∆u(m) + f (m) , u(m) (0) = u0 . (5.28)
∂t
This is a finite-dimensional constant coefficients system of quadratic ODE’s for the Fourier
coefficients um of the function u(x) with |k| ≤ m. If the function f is time-independent,
this system is autonomous. The goal is obtain bounds on the solution u(m) of the Galerkin
system that would allow us to pass to the limit m → +∞, leading to a weak solution of the
Navier-Stokes equations.

A bound on the energy for the Galerkin solutions

We fix an arbitrary time T > 0 throughout the analysis of the Galerkin system. As (5.28)
is a system of constant coefficient non-linear ODEs for the coefficients uk , |k| ≤ m, it has a
(m)
solution for a sufficiently small time t > 0 (which a priori may depend on the initial data u0 ,
as well as on m). Unlike partial differential equations, such ODEs may lose solutions only via
the blow-up of the energy X
ku(m) k22 = |uk |2 , (5.29)
|k|≤m

and that, as we will now show, can not happen in a finite time for any finite m. Indeed, we
have ˆ ˆ
(m) (m) (m)
(Pm (u · ∇u ) · u )dx = ((u(m) · ∇u(m) ) · u(m) )dx = 0. (5.30)
Tn Tn
We used the definition of the projection Pm in the first identity, and the incompressibility
of u(m) in the second. Therefore, multiplying (5.28) by u(m) and integrating, we obtain
ˆ ˆ ˆ
1d (m) 2 (m) 2
|u | dx = −ν |∇u | dx + (f (m) · u(m) )dx. (5.31)
2 dt Tn Tn T n

We will now use the Poincaré inequality

ˆ ˆ
2 2
4π |φ| dx ≤ |∇φ|2 , (5.32)
Tn Tn

36
that holds for all mean-zero periodic functions φ. With its help, identity (5.31) implies
that E(t) = ku(m) k22 satisfies
1 dE p 1
≤ −4π 2 νE(t) + kf (m) k2 E(t) ≤ −4π 2 νE(t) + 2π 2 νE(t) + 2 kf k22
2 dt 8π ν
1
≤ −2π 2 νE(t) + 2 kf k22 . (5.33)
8π ν
Therefore, we have the inequality
d 2
 1 2
E(t)e4π νt ≤ 2 kf k22 e4π νt . (5.34)
dt 4π ν
Integrating in time leads to an estimate
ˆ t
−4π 2 νt 1 2 ν(t−s)
E(t) ≤ E(0)e + 2 e−4π kf (s)k22 ds. (5.35)
4π ν 0

The estimate (5.35) relies on the finiteness of the L2 -norm of the forcing f . Another way
to estimate the right side in (5.31), relying only on the finiteness of a weaker norm of f , is to
use the inequality
ˆ X X
!1/2
X |fk |2
!1/2
(f · g)dx = fk gk ≤ 4π 2 k 2 |gk |2 = k∇gk2 kf kH −1 ,
Tn
k∈Zn k∈Zn k∈Zn
4π 2 k 2
(5.36)
−1
with the H -norm defined as in the above inequality. Using this inequality in (5.31) gives
1 dE ν 1
≤ −νk∇u(m) k22 + k∇u(m) k2 kf kH −1 ≤ −νk∇u(m) k22 + k∇u(m) k22 + kf k2H −1
2 dt 2 2ν
ν 1
= − k∇u(m) k22 + kf k2H −1 . (5.37)
2 2ν
Now, we use the Poincaré inequality to obtain:
dE C2
≤ −C1 νE + kf k2H −1 , (5.38)
dt ν
with universal constants C1 and C2 . Integrating this differential inequality in time leads to
another estimate for E(t), which involves only kf kH −1 and not kf k2 :
ˆ
−C1 νt C20 t −C1 ν(t−s)
E(t) ≤ E(0)e + e kf (s)k2H −1 ds. (5.39)
ν 0

An enstrophy bound
The same argument provides a bound on the enstrophy D(t) = k∇u(t)k22 . Indeed, integrating
inequality (5.37) in time leads to
ˆ ˆ ˆ T
1 (m) 2 ν T (m) 2 1 (m) 2 1
ku (T )k2 + |∇u (s, x)| dxds ≤ ku0 k2 + kf (m) (s)k2H −1 ds. (5.40)
2 2 0 Tn 2 2ν 0

37
The function spaces and an executive summary
Now, we need to introduce certain spaces. We denote by H the space of all mean-zero vector-
valued functions u in the space [L2 (Tn )]n , with zero divergence (in the sense of distributions):
H = {u ∈ L2 (Tn ) : ∇ · u = 0, hui = 0}, (5.41)
with the inner product ˆ
(f, g) = (f · g)dx. (5.42)
Tn
In other words, a vector field u ∈ H if its Fourier coefficients in the expansion
X
u(x) = uk e2πik·x (5.43)
k∈Zn

satisfy u0 = 0, k · uk = 0 for all k ∈ Zn∗ = Zn \ {0}, and

X
kuk2H = |uk |2 < +∞. (5.44)
k∈Zn
∗

We also denote by V the space of divergence-free functions in the Sobolev space H 1 (Tn ):
V = {u ∈ H 1 (Tn ) : ∇ · u = 0, hui = 0}, (5.45)
with the inner product ˆ
∂u ∂g
hf, gi = ·( )dx, (5.46)
Ω ∂xi ∂xi
for two vector-valued functions f and g. That is, u ∈ V if its Fourier coefficients satisfy u0 = 0,
as well as k · uk = 0 for all k, and
X
kuk2V = |k|2 |uk |2 < +∞. (5.47)
k∈Zn
∗

The dual space to V consists of all distributions with the Fourier coefficients that satisfy
X |uk |2
kuk2V 0 = 2
< +∞, u0 = 0 and k · uk = 0. (5.48)
k∈Z n
|k|
∗

We will occasionally use the Sobolev spaces H s , s ∈ R, of divergence-free functions: we say

that u ∈ H s (Tn ) if its Fourier coefficients uk satisfy
 1/2
X
u0 = 0, k · uk = 0 and kukH s =  |k|2s |uk |2  < +∞. (5.49)
k∈Zn
∗

We have, with this notation V = H 1 and V 0 = H −1 . The spaces L2 (0, T ; H) and L2 (0, T ; V )
have the respective norms
ˆ T ˆ T
2 2 2
kukL2 (0,T ;H) = ku(t)kH dt, kukL2 (0,T ;V ) = ku(t)k2V dt. (5.50)
0 0

Summarizing our analysis of the Galerkin system so far, and rephrasing the results in terms
of the spaces H, V and V 0 , we have proved the following.

38
Proposition 5.1 Assume that f ∈ L∞ (0, T ; H). Then, the Galerkin system (5.28) has a
unique solution u(m) ∈ L2 (0, T ; V ) ∩ L∞ (0, T ; H). More precisely, there exist two universal
constants C1 > 0 and C2 > 0 so that
ˆ t
2 −4π 2 νt 1 2
(m) 2
ku (t)kH ≤ ku0 kH e + 2 e−4π ν(t−s) kf (s)k2H ds, (5.51)
4π ν 0
ˆ
(m) 2 2 −C1 νt C2 t −C1 ν(t−s)
ku (t)kH ≤ ku0 kH e + e kf (s)k2V 0 ds (5.52)
ν 0
ˆ T ˆ T
(m) 2 2 1
ν ku (s)kV ds ≤ ku0 k2 + kf (s)k2V 0 ds. (5.53)
0 2ν 0

The Galerkin approximation: bounds on the time derivative

The next step is obtain bounds on the time derivative of u(m) . They will be needed in the
passage to the limit m → +∞, to ensure that the limit is weakly continuous in time. Let us
write the Galerkin approximation of the Navier-Stokes equaitons as
∂u(m) (m)
= ν∆u(m) − Pm (u(m) · ∇u(m) ) + f (m) , u(m) (0) = u0 . (5.54)
∂t
(m)
We will aim to obtain the following bounds on ut . The estimates are slightly different in
two and three dimensions.
Proposition 5.2 Assume that f ∈ L2 (0, T ; V 0 ). There exists a constant C which depends
on the norm ku0 kH of the initial conditionu0 , the L2 (0, T ; V 0 )-norm of the forcing f , and the
viscosity ν but not on m so that the solution of the Galerkin system (5.28) in dimension n = 3
satisfies the estimate ˆ T
∂u(m) 4/3
(t) ≤ C. (5.55)
0 ∂t V0

and in dimension n = 2 it satisfies

ˆ T
∂u(m) 2
(t) ≤ C. (5.56)
0 ∂t V0

For the proof, we will estimate individually each of the terms in the right side of (5.54). As
we assume that f ∈ L2 (0, T ; V 0 ), the forcing term in is not a problem either in dimension two
or three. The Laplacian term in (5.54) is also bounded in L2 (0, T ; V 0 ), as follows from (5.53):
the Fourier coefficients of ∆u are |k|2 uk , hence
X |k|4
k∆uk2V 0 = |u |2 = kuk2V ,
2 k
(5.57)
k∈Zn
|k|

thus
ˆ T ˆ T ˆ T
(m) 1 (m) 1
k∆u (s)k2V 0 ds = (m)
ku (s)k2V ds ≤ ku0 k22 + 2 kf (m) (s)k2V 0 ds. (5.58)
0 0 ν 2ν 0

The nonlinear term will require the most effort. We will establish the following bounds.

39
Lemma 5.3 There exists a constant C that so that in two dimensions we have, for any
function u ∈ V :

k(u · ∇u)kV 0 ≤ CkukH kukV , n = 2, (5.59)

and in three dimensions we have

1/2 3/2
k(u · ∇u)kV 0 ≤ CkukH kukV , n = 3. (5.60)

Together with the uniform energy bound (5.52) and the enstrophy bound (5.53), this implies
the conclusion of Proposition 5.2. Indeed, in dimension n = 2, (5.59) gives
ˆ T ˆ T ˆ T
2 2 2
kPm (u · ∇u)(s)kV 0 ds ≤ k(u · ∇u)(s)kV 0 ds ≤ ( sup ku(t)kH ) ku(s)k2V ds ≤ C,
0 0 0≤t≤T 0

and in dimension n = 3, (5.60) leads to

ˆ T ˆ T ˆ T
4/3 4/3 2/3
kPm (u · ∇u)(s)kV 0 ds ≤ k(u · ∇u)(s)kV 0 ds ≤ ( sup ku(t)kH ) ku(s)k2V ds ≤ C.
0 0 0≤t≤T 0

Thus the proof of Proposition 5.2 is reduced to proving Lemma 5.3.

The proof of Lemma 5.3: bounds on the nonlinear term

Note that
k(u · ∇u)kV 0 = k(−∆)−1/2 (u · ∇u)kH . (5.61)
The operator (−∆)−1/2 is defined via its action on the Fourier coefficients of a mean-zero
function u(x): X uk
(−∆)−1/2 u(x) = e2πik·x . (5.62)
k∈Zn
|k|
This operator commutes with the projection Pm , as, in particular, it preserves the incom-
pressibility of u. Hence, Lemma 5.3 can be restated as follows.
Lemma 5.4 Let u ∈ V , then in three dimensions we have the estimate
1/2 3/2
k(−∆)−1/2 (u · ∇u)kH ≤ CkukH kukV , (5.63)

while in two dimensions we have

k(−∆)−1/2 (u · ∇u)kH ≤ CkukH kukV , (5.64)

Proof. In this proof, we will use interchangeably the notation kukH 1 and kukV , since the
divergence-free property plays almost no role in the proof. Take an arbitrary u ∈ H and w ∈ H
and write, for the inner product in H:

((−∆)−1/2 (u · ∇u), w) = ((u · ∇u), (−∆)−1/2 w). (5.65)

In three dimensions, we will show

40
Lemma 5.5 In dimension n = 3, for any u, v, w ∈ V we have

|((u · ∇v), w)| ≤ CkukH 1/2 kvkH 1 kwkH 1 . (5.66)

Applying this estimate in (5.65) gives

|((−∆)−1/2 (u · ∇u), w)| = |((u · ∇u), (−∆)−1/2 w)| ≤ CkukH 1/2 kukH 1 k(−∆)−1/2 w)kH 1 . (5.67)

As
k(−∆)−1/2 w)kH 1 = kwkH , (5.68)
and
!1/2 !1/2
X X X
kuk2H 1/2 = |k||uk |2 ≤ |k|2 |uk |2 |uk |2 = kukH kukV , (5.69)
k∈Zn k∈Zn k∈Zn

we deduce from (5.65) that in three dimensions we have

1/2 3/2
|((−∆)−1/2 (u · ∇u), w)| ≤ CkukH kukV kwkH . (5.70)

As this estimate holds for all w ∈ H, (5.63) follows.

In two dimensions, we will show

Lemma 5.6 In dimension n = 2, we have

|((u · ∇v), u)| ≤ Ckuk2 kukH 1 kvkH 1 . (5.71)

To see that this implies (5.64), we write, using incompressibility of u:

((−∆)−1/2 (u · ∇u), w) = ((u · ∇u), (−∆)−1/2 w) = −((u · ∇(−∆)−1/2 w), u). (5.72)

Applying estimate (5.71) in (5.72) gives

|((−∆)−1/2 (u · ∇u), w)| = |((u · ∇(−∆)−1/2 w), u)| (5.73)

≤ Ckuk2 k(−∆)−1/2 wkH 1 kukH 1 = Ckuk2 kukH 1 kwkH .

As this holds for any w ∈ H, we conclude that (5.64) holds in two dimensions.
Thus, we only need to verify (5.66) in three dimensions and (5.71) in two dimensions to
finish the proof of Lemma 5.4.
Proof of Lemma 5.5. In three dimensions, we use Hölder’s inequality to get
ˆ
|((u · ∇v), w)| ≤ |uj (∂j vk )wk |dx ≤ kukL3 (T3 ) k∇vkL2 (T3 ) kwkL6 (T3 )
T3
= kukL3 (T3 ) kvkH 1 (T3 ) kwkL6 (T3 ) . (5.74)

The Sobolev inequality says that, for m < n/2,

kf kLq (Tn ) ≤ Ckf kH m (Tn ) , (5.75)

41
as long as
1 1 m
≥ − . (5.76)
q 2 n
Therefore, we have in three dimensions

kukL3 (T3 ) ≤ CkukH 1/2 , (5.77)

and
kwkL6 (T3 ) ≤ CkwkH 1 (T3 ) . (5.78)
It follows then from (5.74) that

|((u · ∇v), w)| ≤ kukL3 (T3 ) kvkH 1 (T3 ) kwkL6 (T3 ) ≤ CkukH 1/2 (T3 ) kvkH 1 (T3 ) kwkH 1 (T3 ) , (5.79)

which is (5.66).
Proof of Lemma 5.6. In two dimensions, we proceed similarly: Hölder’s inequality
implies

|((u · ∇v), w)| ≤ kukL4 (T2 ) kwkL4 (T2 ) kvkH 1 (T2 ) . (5.80)

The Sobolev inequality (5.75) in two dimensions implies that

kf kL4 (T2 ) ≤ Ckf kH 1/2 (T2 ) . (5.81)

Using this in (5.80) leads to

|((u · ∇v), w)| ≤ kukL4 (T2 ) kwkL4 (T2 ) kvkH 1 (T2 ) ≤ CkukH 1/2 (T2 ) kwkH 1/2 (T2 ) kvkH 1 (T2 ) . (5.82)

As
kuk2H 1/2 ≤ kuk2 kukH 1 , (5.83)
we obtain

|((u · ∇v), w)| ≤ C(kukL2 kukH 1 kwkL2 kwkH 1 )1/2 kvkH 1 (T2 ) , (5.84)

hence
|((u · ∇v), u)| ≤ CkukL2 kukH 1 kvkH 1 (T2 ) , (5.85)
which is (5.71). This finishes the proof of Lemma 5.4. 2

A compactness theorem
We have deduced above uniform in m a priori bounds on the solution u(m) of the Galerkin
system
∂u(m) (m)
+ Pm (u(m) · ∇u(m) ) = ν∆u(m) + f (m) , u(m) (0) = u0 . (5.86)
∂t
The next step is to use these uniform bounds to show that the sequence u(m) has a (strongly)
convergent subsequence in L2 (0, T ; H). As we will see, the limit of this subsequence will be a
weak solution of the Navier-Stokes equations. We will use the following result.

42
Proposition 5.7 Let um be a sequence of functions satisfying

kum (t)kH ≤ C, (5.87)

for all 0 ≤ t ≤ T , ˆ T
kum (s)k2V ds ≤ C, for all m = 1, 2, . . . (5.88)
0
and ˆ T
∂u(m) p
(t) ≤ C, for all m = 1, 2, . . ., (5.89)
0 ∂t V0

with some C > 0 and p > 1. Then there exists a subsequence umj of um which converges
strongly in L2 (0, T ; H) to a function u ∈ L2 (0, T ; V ).

Proof. The uniform bound (5.88) implies that there exists a subsequence umj which con-
verges weakly in L2 (0, T ; V ) to a function u ∈ L2 (0, T ; V ), which also obeys the bound (5.88).
In addition, using the diagonal argument, we may ensure that the sequence of time deriva-
(m)
tives ut converges weakly to the derivative ut in Lp (0, T ; V 0 ). Thus, the estimate (5.89) also
holds for the function u. The difference

wj = umj − u

converges weakly to zero in L2 (0, T ; V ), and (5.87)-(5.89) hold for wj as well. Our goal is to
prove that the convergence of wj to zero is strong in L2 (0, T ; H). Note that for any f ∈ V

kf kH ≤ (kf kV kf kV 0 )1/2 , (5.90)

hence, for any δ > 0 we have

1
kf k2H ≤ δkf k2V + kf k2V 0 . (5.91)
δ
The uniform bound (5.88) for the functions wj and (5.91) imply
ˆ T ˆ T
1
kwj k2H dt ≤ Cδ + kwj k2V 0 dt. (5.92)
0 δ 0

Our goal is to estimate the second term in (5.92), and show that it goes to zero as j → +∞,
with δ > 0 fixed. Note that
kwj (t)kV 0 ≤ kwj (t)kH ≤ C. (5.93)
Thus, the Lebesgue dominated convergence theorem shows that it suffices to show that

kwj kV 0 → 0 pointwise in t ∈ [0, T ]. (5.94)

To this end, given a time ε > 0 and ε ≤ t ≤ T , let us write

ˆ t
∂wj (τ, x)
wj (t, x) = wj (s, x) + dτ, (5.95)
s ∂τ

43
and average this identity over s ∈ [t − ε, t]:
ˆ ˆ ˆ t
1 t 1 t ∂wj (τ, x)
wj (t, x) = wj (s, x)ds + ds dτ
ε t−ε ε t−ε s ∂τ
ˆ ˆ
1 t 1 t ∂wj (τ, x)
= wj (s, x)ds + (τ − t + ε) dτ. (5.96)
ε t−ε ε t−ε ∂τ

In order to bound the first term, note that for any 0 ≤ a ≤ b ≤ T the integral
ˆ b
Ij (x) = wj (t, x)dt (5.97)
a

converges weakly to zero in V . Indeed, for any v ∈ V 0 , the function χ[a,b] (t)v(x) is an element
of L2 (0, T ; V 0 ), and wj → 0 weakly in L2 (0, T ; V ), thus we have
ˆ ˆ Tˆ
Ij (x)v(x)dx = wj (t, x)χ[a,b] (t)v(x)dxdt → 0 as j → ∞. (5.98)
Tn 0 Tn

As V is compactly embedded into H, weak convergence in V implies strong convergence in H:

the sequence Ij converges strongly to zero in H. Thus, it also converges strongly to zero in V 0 .
In particular, given any ε > 0 and δ > 0 , for all j sufficiently large we have
ˆ t
1
wj (s, x)ds < δ for j ≥ J(ε, δ, t), (5.99)
ε t−ε V0

giving a pointwise in time estimate for the first term in (5.96). For the second term in (5.96),
we may use the Minkowski inequality, followed by Hölder’s inequality, with 1/q + 1/p = 1:
ˆ t ˆ
1 ∂wj (τ, x) 1 t ∂wj (τ, x)
(τ − t + ε) dτ ≤ (τ − t + ε) dτ (5.100)
ε t−ε ∂τ V0 ε t−ε ∂τ V0
ˆ t 1/q ˆ t p 1/p
1 q ∂wj (τ, x)
≤ (τ − t + ε) dτ dτ
ε t−ε t−ε ∂τ V0
ˆ T p 1/p
1/q ∂wj (τ, x)
≤ Cε dτ ≤ Cε1/q ,
0 ∂τ V 0

for all j ≥ 1. It follows from the above analysis that, given any ε > 0 and δ > 0, we may
find J(ε, δ, t) so that

kwj (t)kV 0 ≤ δ + Cε1/q , for all j ≥ J(ε, δ, t). (5.101)

In other words, we have shown that

kwj (t)kV 0 → 0 as j → ∞, pointwise in t ∈ [0, T ]. (5.102)

As we have explained above, we may use the Lebesgue dominated convergence theorem to
conclude from (5.92) that the sequence wj converges strongly to zero in L2 (0, T ; H). This
finishes the proof of Proposition 5.7. 2

44
The weak solutions as limits of the Galerkin solutions
We will now construct the weak solutions of the Navier-Stokes equations as a limit of the
solutions u(m) of the Galerkin system as m → ∞. In particular, the definition of the weak
solution we will adopt is motivated by the estimates on u(m) we have obtained above. We say
that u ∈ Cw (0, T ; H) if the function ψ(t) = (u(t), h) is continuous for all h ∈ H.
Definition 5.8 A function u is a weak solution of the (periodic) Navier-Stokes equations if
∂u
u ∈ L2 (0, T ; V ) ∩ L∞ (0, T ; H) ∩ Cw (0, T ; H) and ∈ L1loc (0, T ; V 0 ), (5.103)
∂t
and, for any v ∈ V , we have
ˆ ˆ tˆ ˆ tˆ
u(t, x) · v(x)dx + ν ∇u · ∇vdxds + ((u · ∇u) · v)dxds
Tn Tn Tn
ˆ 0
ˆ tˆ 0

= u0 (x) · v(x)dx + f · vdxds, for all v ∈ V and 0 ≤ t ≤ T . (5.104)

Tn 0

Let us check that each term in (5.104) makes sense if u satisfies (5.103), and v ∈ V . The first
term is finite since u ∈ L∞ (0, T ; H). The second is finite since u ∈ L2 (0, T ; V ). The last term
in the left side is finite in three dimensions because of the estimate (5.66):
1/2 3/2
|((u · ∇u), v)| ≤ CkukH 1/2 kukH 1 kvkH 1 ≤ CkukH kukV kvkV , (5.105)
as kukH is uniformly bounded in t, and u ∈ L2 (0, T ; V ). In two dimensions, this term is
bounded because of the estimate (5.71):
|((u · ∇u), v)| = |((u · ∇v), u)| ≤ CkukH kukV kvkV , (5.106)
again, because kukH is uniformly bounded in t, and u ∈ L2 (0, T ; V ).
Finally, the right side in (5.104) is finite provided that f ∈ L2 (0, T ; V 0 ) and u0 ∈ H.
The following theorem, due to Leray, is one of the most classical results in the mathematical
theory of the Navier-Stokes equations (we state here its simpler version for the periodic case).
Theorem 5.9 Given u0 ∈ H and f ∈ L2 (0, T ; V 0 ), there exists a weak solution of the Navier-
Stokes equations
ut + u · ∇u + ∇p = ν∆u + f, t > 0, x ∈ Tn , (5.107)
∇ · u = 0,
u(0, x) = u0 (x).
In addition, this weak solution satisfies the energy inequality
ˆ ˆ tˆ ˆ ˆ tˆ
1 2 2 1 2
|u(t, x)| dx + ν |∇u(s, x)| dxds ≤ |u0 (x)| dx + f (s, x) · u(s, x)dxds.
2 Tn 0 Tn 2 Tn 0 Tn
(5.108)
Moreover, we have
∂u
∈ L4/3 (0, T ; V 0 ) in dimension n = 3, (5.109)
∂t
and
∂u
∈ L2 (0, T ; V 0 ) in dimension n = 2. (5.110)
∂t

45
Proof. Let u(m) be the solutions of the Galerkin system (5.28):

∂u(m) (m)
+ Pm (u(m) · ∇u(m) ) = ν∆u(m) + f (m) , u(m) (0) = u0 . (5.111)
∂t
The estimates we have obtained in the previous section imply that, after extracting a subse-
quence, u(m) converge strongly in L2 (0, T ; H) and weakly in L2 (0, T ; V ) to some u. Moreover,
the functions u(m) satisfy a uniform continuity in time bound in V 0 :
ˆ t (m)
(m) (m) ∂u
u (t) − u (s) = dτ, (5.112)
s ∂τ
thus
ˆ t ˆ t 1/p
(m) (m) ∂u(m) ∂u(m) p
ku (t) − u (s)kV 0 ≤ dτ ≤ (t − s)1/q dτ
s ∂τ V 0 s ∂τ V0
ˆ T 1/p
1/q ∂u(m) p
≤ (t − s) dτ ≤ C(t − s)1/q , (5.113)
0 ∂τ V0

with p = q = 2 in dimension n = 2, and p = 4/3, q = 4 in dimension n = 3. Thus, u obeys

the same estimate, and u ∈ C(0, T ; V 0 ). We also know that

∂u(m) ∂u
→ ,
∂t ∂t
weakly in L4/3 (0, T ; V 0 ) in three dimensions, and weakly in L2 (0, T ; V 0 ) in two dimensions.
Given any v ∈ V we multiply the Galerkin system (5.111) by v and integrate:
ˆ ˆ tˆ
(m)
u (t, x)v(x)dx + (u(m) · ∇u(m) ) · (Pm v)dxds
Tn Tn
ˆ tˆ 0
ˆ ˆ tˆ
(m) (m)
= −ν ∇u · ∇vdxds + u0 (x)v(x)dx + f vdxds. (5.114)
0 Tn Tn 0 Tn

We pass now to the limit in this identity, looking at each term individually. The first term in
the right side is easy:
ˆ tˆ ˆ tˆ
(m)
∇u · ∇vdxds → ∇u · ∇vdxds, (5.115)
0 Tn 0 Tn

because u(m) converges weakly to u in L2 (0, T ; V ). Next, we look at the nonlinear term: set
ˆ tˆ ˆ tˆ
(m) (m)
Am = (u · ∇u ) · (Pm v)dxds − (u · ∇u) · vdxds. (5.116)
0 Tn 0 Tn

Let us recall (5.66):

|((u · ∇v), w)| ≤ CkukH 1/2 kvkH 1 kwkH 1 . (5.117)

46
This inequality holds both in two and three dimensions and implies that
ˆ tˆ ˆ t 
2
(u · ∇u) · (Pm v − v)dxds ≤ ku(s)kV ds kPm v − vkV ≤ CkPm v − vkV → 0,
0 Tn 0
(5.118)
as m → ∞. Hence, Am has the same limit as m → ∞ as
ˆ tˆ
0
Am = (u(m) · ∇u(m) − u · ∇u) · (Pm v)dxds = B1 + B2 , (5.119)
t0 Tn

where B1,2 correspond to the decomposition

u(m) · ∇u(m) − u · ∇u = u(m) · ∇u(m) − u(m) · ∇u + u(m) · ∇u − u · ∇u

= u(m) · (∇u(m) − ∇u) + (u(m) − u) · ∇u. (5.120)

To estimate B1 , we write
ˆ tˆ ˆ tˆ
(m) (m)
B1 = (u · (∇u − ∇u)) · (Pm v)dxds = − (u(m) · ∇Pm v) · (u(m) − u)dxds.
0 Tn 0 Tn
(5.121)
The same proof as for (5.66) shows that

|(u · ∇v), w)| ≤ kukV kvkV kwkH 1/2 . (5.122)

Using this in (5.121) gives

ˆ t
|B1 | ≤ ku(m) (s)kV kvkV ku(m) (s) − u(s)kH 1/2 ds (5.123)
t0
ˆ t 1/2 ˆ t 1/2
(m)
≤ kvkV ku (s)k2V ds ku (m)
(s) − u(s)k2H 1/2 ds
0 0
ˆ t 1/4 ˆ t 1/4
(m)
≤ CkvkV ku (s) − u(s)k2V ds ku (m)
(s) − u(s)k2H ds
0 0
≤ Cku(m) − ukL2 (0,T ;H) → 0, as m → ∞,

as u(m) converges to u strongly in L2 (0, T ; H). As for B2 , we write

ˆ tˆ ˆ t
(m)
|B2 | = ((u − u) · ∇u) · (Pm v)dxds ≤ ku(m) (s) − u(s)kH 1/2 ku(s)kV kvkV ds
0 Tn 0
(m)
≤ kvkV kukL2 (0,T ;V ) ku (s) − u(s)kL2 (0,T ;H 1/2 ) → 0, (5.124)

for the same reason as in (5.123).

In order to pass to the limit in the two terms in (5.114) that do not involve the time
(m)
integration, we first note that u0 converges strongly in H to u0 . Furthermore, as u(m)
converges weakly to u in L2 (0, T ; V ), we may extract a subsequence so that u(m) (t) converges
weakly in V to u(t) (pointwise in t), except for t ∈ E, where E is an exceptional set of times
in [0, T ] of measure zero. Weak convergence in V implies that u(m) (t) converges strongly

47
to u(t) in H for t ∈
6 E. Hence, taking t 6∈ E and passing to the limit m → ∞ in (5.114) we
arrive at
ˆ ˆ ˆ tˆ
u(t, x)v(x)dx = u0 (x)v(x)dx − (u · ∇u) · vdxds
Tn Tn Tn
ˆ tˆ ˆ tˆ 0

−ν ∇u · ∇vdxds + f vdxds. (5.125)

0 Tn 0 Tn

Given the a priori bounds on u, the right side of (5.125) is a continuous function of t, defined
for all t ∈ [0, T ], not just t ∈ E. In addition, we know that u(t) is continuous in Cw (0, T ; V 0 ),
and coincides with the aforementioned right side of (5.125) for t 6∈ E. This continuity implies
that u(t) coincides with the right side of (5.125) for all 0 ≤ t ≤ T , which means that it
satisfies (5.125) for all t ∈ [0, T ], giving us a weak solution of the Navier-Stokes equations.
The fact that u ∈ Cw (0, T ; H), and not just u ∈ C(0, T ; V 0 ) follows from (5.125), the
density of V in H and the uniform in t bound on ku(t)kH .
To obtain the energy inequality, we start with the identity
ˆ t ˆ tˆ
1 (m) 2 (m) 2 1 (m) 2
ku (t)kH + ν ku (s)kV ds = ku0 kH + f · u(m) dxds. (5.126)
2 0 2 0 Tn

The right side converges, as m → ∞, to

ˆ tˆ
1
ku0 k2H + ν f · udxds. (5.127)
2 t0 Tn

In the left side, we may use the Fatou lemma to conclude that, as u(m) (t) converges weakly
in H to u(t) for all t ∈ [0, T ], we have
ˆ t ˆ tˆ
1 2 2 1 2
ku(t)kH + ν ku(s)kV ds ≤ ku0 kH + f · udxds. (5.128)
2 0 2 0 Tn

This completes the proof. 2

Uniqueness of the weak solutions in two dimensions

One of the main issues with weak solutions in general in nonlinear partial differential equations
is the issue of uniqueness – it is often much easier to show that they exist than to prove their
uniqueness. Uniqueness of a weak solution hints that it is a “correct” solution, while non-
uniqueness means that an extra condition is needed to pick the physically meaningful solution.
This happens, for instance, in the theory of conservation laws where the notion of an entropy
solution guarantees uniqueness among all weak solutions. The problem of the uniqueness of
the weak solutions for the Navier-Stokes equations in three dimensions is still open. In two
dimensions, we know that the weak solutions of
ut + u · ∇u + ∇p = ν∆u, t > 0, x ∈ T2 , (5.129)
∇ · u = 0,
u(0, x) = u0 (x).
are unique.

48
Theorem 5.10 Let f ∈ L2 (0, T ; V 0 ) and u0 ∈ H. If u1 and u2 are two weak solutions
of (5.129) which both lie in L2 (0, T ; V ) ∩ L∞ (0, T ; H) ∩ Cw (0, T ; H), then u1 = u2 .

Proof. First, we recall, see Theorem 5.9, that if u is a weak solution of the Navier-Stokes
equations (5.129) in L2 (0, T ; V ) ∩ L∞ (0, T ; H) in two dimensions, then ut ∈ L2 (0, T ; V 0 ). Let
us denote w = u1 − u2 . This function satisfies

wt + u1 · ∇w + w · ∇u2 + ∇p0 = ν∆w, t > 0, x ∈ T2 , (5.130)

∇ · w = 0,
w(0, x) = 0,

with p0 = p1 − p2 , and we know that wt ∈ L2 (0, T ; V 0 ).

Multiplying (5.130) by w and integrating over the torus gives
ˆ ˆ ˆ
2
wt · w + ν |∇w| dx + wk (∂j u2,m )wm dx = 0. (5.131)
T2 T2 T2

As wt ∈ V 0 for a.e. t, and w ∈ V for a.e. t ∈ [0, T ], identity (5.131) holds for a.e. t ∈ [0, T ].
Recall that in two dimensions we have

|(w · ∇u2 , w)| ≤ CkwkH ku2 kV kwkV . (5.132)

As w ∈ L∞ (0, T ; H) and u2 , w ∈ L2 (0, T, H), we conclude from (5.131) and (5.132) that
ˆ T
|(wt (t), w(t))|dt < +∞.
0

Now, (5.131) implies that

d C
kwk2H ≤ CkwkH ku2 kV kwkV − νkwk2V ≤ ku2 k2V kwk2H . (5.133)
dt ν
As ˆ T
ku2 k2V dt < +∞,
0
Gronwall’s inequality implies that
nˆ t o
kw(t)k2H ≤ kw(0)k2H exp ku2 (s)k2V ds = 0, (5.134)
0

since w(0) = 0. This finishes the proof. 2

Note that this proof would fail in three dimensions. The reason is that in three dimensions
the nonlinear term satisfies
1/2 3/2
|(w · ∇u2 , w)| ≤ CkwkH ku2 kV kwkV , (5.135)

rather than
|(w · ∇u2 ), w| ≤ CkwkH kwkV ku2 kV , (5.136)

49
which holds in two dimensions. Thus, instead of (5.133), we would get, using Young’s in-
equality
d 1/2 3/2 C
kwk2H ≤ CkwkH ku2 kV kwkV − νkwk2V ≤ 3 ku2 k4V kwk2H . (5.137)
dt ν
As we do not have a uniform bound on
ˆ T
ku(s)k4V ds,
0

we would not be able to finish the proof using the Gronwall inequality. We will need extra
assumptions for uniqueness, which is what we will discuss next.

6 Strong solutions in two and three dimensions

Uniqueness of strong solutions in three dimensions
We say that u is a strong solution of the Navier-Stokes equations (in either two or three
dimensions) if u is a weak solution, and, in addition, u ∈ Cw (0, T ; V ), and the following
bounds hold: ˆ
sup |∇u(t, x)|2 dx < +∞, (6.1)
t∈[0,T ] Tn

and ˆ ˆ
T
|∆u(t, x)|2 dxdt < +∞. (6.2)
0 Tn
The motivation for this definition comes from two properties that we will prove: first, unlike
for the weak solutions, one can show that strong solutions are unique in three dimensions
(existence of strong solutions in three dimensions is an important open problem). Second, as
we will show, the conditions in the definition of the strong solutions are sufficient to show
that they are actually infinitely differentiable if the initial condition u0 and the forcing f are.
First, we prove their uniqueness in three dimensions.
Theorem 6.1 Let u1,2 be two solutions of the Navier-Stokes equations on T3 with the initial
condition u0 ∈ H and f ∈ L2 (0, T ; H). If both u1,2 satisfy (6.1) and (6.2), and they lie
in Cw (0, T ; V ) then u1 = u2 .
Proof. We argue as in the proof of uniqueness of the weak solutions in two dimensions.
Let w = u1 − u2 , so that
∂w
( , w) + νkwk2V + (w · ∇u2 , w) = 0, (6.3)
∂t
as in (5.131). We now use the estimate
1/2 1/2
|((w · ∇u, w)| ≤ CkwkL2 kwkH 1 kukH 1 k∆uk2 . (6.4)

It is obtained as follows: recall that in three dimensions we have

kwkL3 (T3 ) ≤ CkwkH 1/2 , (6.5)

50
thus
ˆ
|((w · ∇u, w)| ≤ |w||∇u||w|dx ≤ kwkL3 k∇ukL3 kwkL3 ≤ Ckwk2H 1/2 k∇ukH 1/2 (6.6)
T3
1/2 1/2
≤ CkwkL2 kwkH 1 kukH 1 k∆ukL2 ,
which is (6.4). Using the bound (6.4) in (6.3) leads to
1d C
(kwk2L2 ) + νkwk2H 1 ≤ kwk2L2 kukH 1 k∆uk2 + νkwk2H 1 . (6.7)
2 dt ν
It follows that
1d C
(kwk2L2 ) ≤ kukH 1 k∆uk2 kwk2L2 . (6.8)
2 dt ν
Now, Grownwall’s inequality implies that w(t) = 0 provided that w(0) = 0, and
ˆ t
kukH 1 k∆uk2 ds < +∞, (6.9)
0

which is a consequence of (6.1)-(6.2). 2

Construction of the strong solutions in two dimensions

We now use the Galerkin system in two dimensions to show existence of global in time strong
solutions of the Navier-Stokes equations in two dimensions. Once again, we restrict ourselves
to the simpler case of the two-dimensional torus T2 . As in the proof of the existence of weak
solutions, we will use the Galerkin system
∂u(m) (m)
+ Pm (u(m) · ∇u(m) ) = ν∆u(m) + f (m) , u(m) (0) = u0 , (6.10)
∂t
and then pass to the limit m → +∞. However, we will be able to obtain better a priori
bounds on the Galerkin system in two dimensions to conclude that in the limit we actually
obtain strong solutions of the Navier-Stokes equations. Since we have already shown the
uniqueness of the weak solutions in the two-dimensional case, this will also show that weak
solutions are actually strong in two dimensions.

Galerkin solutions are often not large

The first step is to show that solutions of the Galerkin system are “often not large” – this
will be made precise soon. The second step will be to show that if solutions are often not too
large, then they can never be large.
Taking the inner product with u(m) we obtain the familiar identity
1 d (m) 2
ku kH + νk∇u(m) k2H = (f, u(m) ). (6.11)
2 dt
We may use the Poincaré inequality
ˆ ˆ
2
X
2
X
2 2 1
|u(x)| dx = |uk | ≤ |k| |uk | = 2 |∇u|2 dx, (6.12)
T2
k∈Zn k∈Zn
4π T n

51
to conclude from (6.11) that
1 d (m) 2 1 4π 2 ν (m) 2 1 ν
ku kH + νk∇u(m) k2H ≤ 2
kf k 2
H + ku )kH ≤ 2 kf k2H + k∇u(m) )k2H .
2 dt 2 · 4π ν 2 8π ν 2
(6.13)
We deduce the bounds we have seen before: there exist two explicit constants C1,2 > 0, so
that ˆ t ˆ
(m) 2 2 C1 t
ν k∇u kV ds ≤ ku0 kH + kf k2H ds, (6.14)
0 ν 0
and ˆ
(m) 2 2 −C2 νt C1 t −C2 ν(t−s)
ku (t)kH ≤ ku0 kH e + e kf k2H ds. (6.15)
ν 0
In particular, if f ∈ L∞ (0, T ; H), then
C1
ku(m) (t)k2H ≤ ku0 k2H e−C2 νt + 2
kf k2∞ , (6.16)
ν
with
kf k∞ = sup kf (t)kH . (6.17)
t>0

Our next goal is to get uniform in time bounds on ku(m) (t)kV – this is not something we
have done in the construction of the weak solutions, because such bound holds only in two
dimensions, and not in three, while the weak solutions can be constructed both in two and
three dimensions. The first step in that direction is to show that this norm can not be large
for too long a time.
Proposition 6.2 Let u(m) (t) be the solution for the Galerkin system with f ∈ L∞ (0, +∞; H)
and u0 ∈ H, in either two or three dimensions. Then in every time interval of length τ > 0
there exists a time t0 so that
(m) 2 2 2 C1 1 
ku (t0 )kV ≤ ku0 kH + ( + τ )kf k∞ . (6.18)
τν ν ν
Proof. Inequality (6.15) implies that
ˆ t
C1 t
ν k∇u(m) k2V ds ≤ ku0 k2H + kf k2∞ , (6.19)
0 ν
and (6.15) that
C1
ku(m) (t)k2H ≤ ku0 k2H + 2
kf k2∞ . (6.20)
ν
Let us also integrate (6.13) between the times t and t + τ , leading to
ˆ t+τ
C1 C1 1
ν ku(m) (s)k2V ds ≤ ku(m) (t)k2H + kf k∞ τ ≤ ku0 k2H + kf k∞ ( + τ ). (6.21)
t ν ν ν
The right side above does not depend on the time t. Therefore, on any time interval [t, t + τ ]
we may estimate the Lebesgue measure of the set of times when ku(s)kV is large:
1  C1 1 
{s : s ∈ [t, t + τ ] s.t. ku(m) (s)kV ≥ ρ} ≤ 2 ku0 k2H + kf k∞ ( + τ ) . (6.22)
νρ ν ν

52
In particular, taking
  1/2
2 C1 1
ρ0 = ku0 k2H + kf k∞ ( + τ ) ,
τν ν ν
we arrive at the conclusion of Proposition 6.2. 2

Galerkin solutions are never large

Next, we will get rid of the “sometimes not large” restriction in Proposition 6.2, showing
that in two dimensions Galerkin solutions are never large in V . We will prove the following
estimate for the solutions of the Galerkin system

∂u(m) (m)
+ Pm (u(m) · ∇u(m) ) = ν∆u(m) + f (m) , u(m) (0) = u0 . (6.23)
∂t
Proposition 6.3 Let u(m) be the solution of the Galerkin system (6.23) with the initial
data u0 ∈ H and f ∈ L∞ (0, T ; H). There exists a constant α that depends on ν, ku0 kH
and kf k∞ but not on m so that u(m) satisfies the bounds

ku(m) (t)kV ≤ α for all t ≥ 1, (6.24)

and
α
ku(m) (t)kV ≤ for all 0 < t < 1. (6.25)
t
In addition, if u0 ∈ V then there exists a constant α1 which depends on ν, ku0 kH and kf k∞
but not on m so that
ku(m) (t)kV ≤ α1 for all 0 < t < 1. (6.26)
Proof. The idea is to use Proposition 6.2 – we know that for any time t > 1 there is a
time t0 ∈ [t − 1, t] so that the norm ku(m) (t0 )kV ≤ α, with the constant α which depends only
on ν, ku0 kH and kf k∞ . The additional ingredient in this proof will be a control of the growth
of ku(m) kV on the time intervals of length 1.
We multiply (6.23) by ∆u and integrate. The first term gives
ˆ ˆ
(m) (m) (m) 1d
ut · ∆u dx = − ∇ut · ∇u(m) dx = − k∇u(m) (t)k2H , (6.27)
T2 T2 2 dt
so that the overall balance is
1d
k∇u(m) (t)k2H + νk∆u(m) k2H − ((u(m) · ∇u(m) ), ∆u(m) ) = −(f, ∆u(m) ). (6.28)
2 dt
For the nonlinear term, we will use the inequality
1/2 3/2
|((u · ∇u), ∆u)| ≤ kukH kukV k∆ukH , (6.29)

which holds in two dimensions. The proof is similar to that of (5.71): we write
ˆ
|((u · ∇v), w)| ≤ |(uj ∂j vk )wk |dx ≤ ku · ∇vkL2 kwkL2 ≤ kukL4 k∇vkL4 kwkL2 . (6.30)
Tn

53
The Sobolev inequality
1 1 m
kf kLq (Tn ) ≤ Ckf kH m (Tn ) , ≥ − (6.31)
q 2 n
implies that in two dimensions we have

kf kL4 (T2 ) ≤ Ckf kH 1/2 (T2 ) . (6.32)

Using this in (6.30) leads to

1/2 1/2 1/2 1/2
|((u · ∇u), ∆u)| ≤ kukH 1/2 kukH 3/2 k∆ukL2 ≤ kukH kukV kukV k∆ukH k∆ukH
1/2 3/2
= kukH kukV k∆ukH , (6.33)

which is (6.29). It follows that the nonlinear term can be estimated, using the inequality
ν 4/3 C 4
ab ≤ a + 3b
4 ν
as
ν C
|((u · ∇u), ∆u)| ≤ k∆uk2H + 3 kuk2H kuk4V . (6.34)
4 ν
Returning to (6.28), we obtain
1d
k∇u(m) (t)k2H + νk∆u(m) k2H ≤ |((u(m) · ∇u(m) ), ∆u(m) )| + kf k∞ k∆u(m) kH (6.35)
2 dt
ν C ν C
≤ k∆u(m) k2H + 3 ku(m) k2H ku(m) k4V + k∆u(m) k2H + kf k2∞ .
4 ν 4 ν
We conclude that
1 d (m) ν C C
ku (t)k2V + k∆u(m) k2H ≤ 3 ku(m) k2H ku(m) k4V + kf k2∞ . (6.36)
2 dt 2 ν ν
Let us set ˆ t
2C
G(t0 ; t) = 3 ku(s)k2H ku(s)k2V ds, (6.37)
ν t0

then (6.36) implies, for any t ≥ t0 :

d  (m) 2  C
ku kV exp{−G(t0 ; t)} ≤ kf k2∞ exp{−G(t0 ; t)}. (6.38)
dt ν
Integrating between t0 and t gives
ˆ t
(m) C
ku (t)k2V
≤ ku (m)
(t0 )k2V
exp{G(t0 ; t)} + kf k2∞ exp{G(t0 ; t)} exp{−G(t0 ; s)}ds
ν t0
ˆ t
C
≤ ku(m) (t0 )k2V exp{G(t0 ; t)} + kf k2∞ exp{G(s; t)}ds
ν t0
C
≤ ku(m) (t0 )k2V exp{G(t0 ; t)} + kf k2∞ (t − t0 ) exp{G(t0 ; t)}. (6.39)
ν

54
Now we will use the “sometimes small” result in Proposition 6.2. Given τ > 0 and t > τ we
may find t0 ∈ [t − τ, t] such that
1
ku(t0 )kV ≤ α(1 + ), (6.40)
τ
with the constant α > 0 that only depends on ν, ku0 kH and kf k∞ but not on m or ku0 kV .
We may also use (6.21) to estimate G(t0 ; t):

G(t0 ; t) ≤ α(1 + τ ). (6.41)

Using this in (6.39) shows that for all t > τ we have

C
ku(m) (t)k2V ≤ ku(m) (t0 )k2V exp{G(t0 ; t)} + kf k2∞ (t − t0 ) exp{G(t0 ; t)} (6.42)
ν
1
≤ α(1 + )eα(1+τ ) + ατ eα(1+τ ) .
τ
This bound is uniform in t > τ . Hence, if we fix τ = 1, we get a uniform in m estimate
for ku(m) (t)kV for all t > 1, giving the bound (6.24).
In order to deal with times t < 1, we will use (6.42) on the time intervals t ∈ [1/2k+1 , 1/2k ]
with τ = 1/2k+1 . The point is that for such times t and τ are comparable: τ ≤ t ≤ 2τ .
Therefore, for t < 1 we have an estimate

tku(m) (t)k2V ≤ α, (6.43)

with the constant α that only depends on ν, ku0 kH and kf k∞ but not on m or ku0 kV , which
is (6.25).
Finally, if we allow the dependence on the norm ku0 kV , then for times t < 1 we may
simply use the first line in (6.42) with t0 = 0, together with the estimate

G(t0 = 0, t = 1) ≤ 2α, (6.44)

which follows from (6.41). This gives (6.26) and finishes the proof of Proposition 6.3. 2

The strong solutions in two dimensions

The above bounds on the solutions u(m) of the Galerkin system (6.23) allow us to pass to
the limit m → ∞ to construct solutions of the Navier-Stokes equations on a two-dimensional
torus

ut + u · ∇u + ∇p = ν∆u + f, t > 0, x ∈ T2 , (6.45)

∇ · u = 0,
u(0, x) = u0 (x).

55
Theorem 6.4 Assume that T > 0, u0 ∈ H and f ∈ L∞ (0, T ; H). Then there exists a
constant C > 0 which depends only on ν, ku0 kH and kf k∞ , and a solution of the Navier-
Stokes equation (6.45) which satisfies the bounds

ku(t)kH ≤ C, (6.46)
C
ku(t)kV ≤ C for t ≥ 1, and ku(t)k ≤ for 0 < t < 1, (6.47)
t
ˆ T
ku(t)k2V dt ≤ C. (6.48)
0

In addition, for any s > 0 there exists Cs so that

ˆ T
k∆u(t)k2H dt ≤ Cs T. (6.49)
s

Moreover, if u0 ∈ V then there exists a constant C > 0 which depends only on ν, ku0 kV and
kf k∞ so that

ku(t)kV ≤ C for all t ≥ 0, (6.50)

and
ˆ T
k∆u(t)k2H dt ≤ CT. (6.51)
0

These bounds are inherited from the solutions of the Galerkin system, we leave the details
of this passage to the reader, as they are very close to what was done in the corresponding
passage in the construction of the weak solutions. We only mention that the L2 (0, T ; H)
estimate for ∆u follows from (6.36). Note that we do not yet claim that if u0 is an infinitely
differentiable function, then the solution u(t, x) is also smooth but only that u is a strong
solution in the sense that the aforementioned bounds on u(t, x) hold. We will improve them
soon, assuming that u0 is smooth.

Strong solutions in three dimensions: small data

While existence of global in time strong solutions in three dimensions is not known, strong
solutions do exist if the initial condition and the forcing are small.

Theorem 6.5 Let u0 ∈ V and f ∈ L2 (0, T ; H). There exists a constant C > 0 which depend
only on ν, so that if ˆ T
ku0 kH + kf (t)k2H dt ≤ C, (6.52)
0
then the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u + f, t > 0, x ∈ T3 , (6.53)

∇ · u = 0,
u(0, x) = u0 (x),

56
have a strong solution on the time interval [0, T ] that satisfies
ˆ T
2 1
ku(t)kV + k∆u(t)k2H dt ≤ , (6.54)
0 C
for all 0 ≤ t ≤ T .
Note that if f = 0 then solutions exist for all t > 0 if the initial condition is small: ku0 kH ≤ C.
The proof of this theorem, once again, relies on the estimates for the Galerkin solutions
(m) (m)
ut + Pm (u(m) · ∇u(m) ) = ν∆u(m) , u(m) (0, x) = u0 (x), t > 0, x ∈ T3 . (6.55)

Taking the inner product with ∆u(m) , as we did in the two-dimensional case, we obtain, as
in(6.28):
1 d (m)
ku (t)k2V + νk∆u(m) k2H − (u(m) · ∇u(m) , ∆u(m) ) = −(f, ∆u(m) ). (6.56)
2 dt
In three dimensions, we may not use the two-dimensional estimate (6.29) for the nonlinear
term. Instead, we will bound it as
3/2 3/2 C ν
|(u · ∇u, ∆u)| ≤ CkukV k∆ukH ≤ 3
kuk6V + k∆uk2H . (6.57)
ν 4
This comes from the estimate

|(u · ∇u, ∆u)| ≤ CkukL6 k∇ukL3 k∆ukL2 . (6.58)

The Sobolev inequality implies that in three dimensions we have

kukL3 ≤ CkukH 1/2 , kvkL6 ≤ CkvkH 1 . (6.59)

Using this in (6.58) gives

3/2 3/2
|(u · ∇u, ∆u)| ≤ CkukL6 k∇ukL3 k∆ukL2 ≤ CkukH 1 k∇ukH 1/2 k∆ukL2 ≤ CkukH 1 k∆ukL2 ,
(6.60)
which is (6.57). We will estimate the forcing term in (6.56) as
4 ν
|(f, ∆u)| ≤ kf k2H + k∆uk2H . (6.61)
ν 4
Altogether, with the above estimates, (6.56) implies
1 d (m)
ku (t)k2V + νk∆u(m) k2H ≤ (u(m) · ∇u(m) , ∆u(m) ) − (f, ∆u(m) ) (6.62)
2 dt
C ν C ν
≤ 3 ku(m) k6V + k∆u(m) k2H + kf k2H + k∆u(m) k2H .
ν 4 ν 4
This gives
1 d (m) C ν C C ν C
ku (t)k2V ≤ 3 ku(m) k6V − k∆u(m) k2H + kf k2H ≤ 3 ku(m) k6V − ku(m) k2V + kf k2H .
2 dt ν 2 ν ν 2 ν
(6.63)

57
Therefore, the function y(t) = ku(m) (t)k2V satisfies a differential inequality

dy C C
≤ 3 y 3 − νy + kf k2H . (6.64)
dt ν ν
Hence, as long as
ν2
y(s) ≤ √ , for all 0 < s < t, (6.65)
C
we have
dy C
≤ kf k2H , (6.66)
dt ν
and ˆ t
C
y(t) ≤ y(0) + kf (s)k2H ds. (6.67)
ν 0
It follows that if ˆ ∞
C ν2
ku0 k2V + kf (s)k2H ds ≤ √ , (6.68)
ν 0 C
with a universal constant C > 0, then

ν2
ku(m) (t)k2V ≤ √ , (6.69)
C

for all t > 0. This is the part of the bound (6.54) on ku(m) kV . In order to get the bound
on ∆u(m) in L2 (0, T ; H), we go back to (6.62):

1 d (m) ν C C C
ku (t)k2V + k∆u(m) k2H ≤ 3 ku(m) k6V + kf k2H ≤ Cνku(m) k2V + kf k2H , (6.70)
2 dt 2 ν ν ν
leading to
ˆ ˆ T ˆ
ν T (m) 2 (m) 2 (m) 2 C T
k∆u (t)kH dt ≤ ku0 kV + Cν ku (t)kV dt + kf (t)k2H dt. (6.71)
2 0 0 ν 0

As we also have ˆ ˆ
T T
(m) C
ν ku (t)k2V dt ≤ ku0 k2H + kf (t)k2H dt, (6.72)
0 ν 0

we deduce that under the assumptions (6.52) we have

ˆ T
k∆u(m) (t)k2H dt ≤ C. (6.73)
0

Passing to the limit m → ∞ we construct a solution of the Navier-Stokes equations u(t) that
satisfies the same estimates (6.54). Uniqueness of the strong solution finishes the proof.

58
Strong solutions in three dimensions: short times
Next, we show that strong solutions of the Navier-Stokes exist for a sufficiently short time
even if the data is not small.
Theorem 6.6 Let u0 ∈ V and f ∈ L2 (0, T ; H). There exists a constant C0 > 0 which
depends on ν and ku0 kV , so that if
ˆ T0
T0 + kf (t)k2H dt ≤ C0 , (6.74)
0

then the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u + f, t > 0, x ∈ T3 , (6.75)

∇ · u = 0,
u(0, x) = u0 (x),

have a strong solution on the time interval [0, T0 ] that satisfies

ku(t)k2V ≤ C0−1 , (6.76)

for all 0 ≤ t ≤ T0 .
For the proof, we recall (6.70):
1 d (m) ν C C
ku (t)k2V + k∆u(m) k2H ≤ 3 ku(m) k6V + kf k2H , (6.77)
2 dt 2 ν ν
which, in particular, implies that the function y(t) = ku(m) (t)k2V satisfies a differential in-
equality

ẏ(t) ≤ Cy(t)3 + Ckf k2H , (6.78)

with the constant C that depends on ν. Dividing by (1 + y)3 we get

ẏ Cy 3 + Ckf k2H
3
≤ 3
≤ C + Ckf k2H , (6.79)
(1 + y) (1 + y)
Integrating in time leads to
ˆ t
1 1
2
− ≤ Ct + C kf (s)k2H ds. (6.80)
(1 + y0 ) (1 + y(t))2 0

Therefore, as long as the time t is such that (6.80) holds, or, rather, as long as T0 satisfies
ˆ T0
1 1
CT0 + C kf (s)k2H ds ≤ 2 2
≤ , (6.81)
0 2(1 + ku0 kV ) 2(1 + y0 )2
we have, for all 0 ≤ t ≤ T0 :
1 1 1
≥ ≥ . (6.82)
(1 + y(t))2 2(1 + y0 ) 2 2(1 + ku0 k2V )2

59
Therefore, as long as the time t is sufficiently small, so that (6.80) holds, we have

ku(m) (t)k2V ≤ 2(1 + ku0 k2V ). (6.83)

As usual, this uniform bound on the Galerkin approximations u(m) (t) implies that, passing
to the limit m → +∞, we construct a strong solution of the Navier-Stokes equations for
times 0 ≤ t ≤ T0 .
In general, for an arbitrary m > 1 2

Strong solutions are smooth if the data are smooth

We now show that if the initial condition u0 and the forcing f are smooth, then the strong
solution of the Navier-Stokes equations (if it exists) is also infinitely differentiable. We consider
only the three-dimensional case but the analysis applies essentially verbatim to the two-
dimensional case as well.

Theorem 6.7 Let u(t, x) be the strong solution of the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u + f, 0 < t ≤ T, x ∈ T3 , (6.84)

∇ · u = 0,
u(0, x) = u0 (x),

in the sense that there exists C > 0 so that

ˆ T
sup ku(t)kV ≤ C, k∆u(s)k2H ds ≤ C. (6.85)
0≤t≤T 0

Assume that u0 ∈ C ∞ (T3 ) and f ∈ C ∞ (0, T ; T3 ), then u ∈ C ∞ (0, T ; T3 ).

The strategy of the proof will be to estimate k∆m u(t)kH for all m ∈ N, and show that, as
long u satisfies the assumptions of Theorem 6.7, these norms remain finite for 0 ≤ t ≤ T , and
all m ∈ N. As m ∈ N will be arbitrary, the Sobolev embedding theorem will imply that u is
infinitely differentiable in x, while the Navier-Stokes equations themselves will imply that u is
infinitely differentiable in time (using the projection on the divergence free fields, the reader
should convince himself that the pressure term is not a problem).
Multiplying (6.84) by (−∆)m u and integrating over T3 gives

(ut , (−∆)m u) − (u · ∇u, (−∆)m u) = −ν(−∆u, (−∆)m u) + (f, (−∆)m u). (6.86)

Integrating by parts leads to

1d
k(−∆)m/2 uk2H − ((−∆)m/2 (u · ∇u), (−∆)m/2 u) + νk(−∆)(m+1)/2 uk2H
2 dt
≤ k(−∆)m/2 f kH k(−∆)m/2 ukH . (6.87)

The key inequality we will need for the nonlinear term is given by the following lemma.

60
Lemma 6.8 For every m > 3/2 there exists a constant C > 0 so that for any vector-valued
functions u, v such that u0 = v0 = 0, and ∇ · u = ∇ · v = 0, and uk = vk = 0 for all k > M ,
with some M > 0, we have

k(−∆)m/2 P (u · ∇v)kH ≤ Ck(−∆)m/2 ukH k(−∆)(m+1)/2 vkH . (6.88)

Here, P is the projection on divergence-free fields.

Postponing the proof of this lemma, we apply it in (6.87):
1d
k(−∆)m/2 uk2H + νk(−∆)(m+1)/2 uk2H ≤ k(−∆)m/2 f kH k(−∆)m/2 ukH (6.89)
2 dt
+Ck(−∆)m/2 uk2H k(−∆)(m+1)/2 ukH .

Next, we use Young’s inequality in the right side together with the Poincare inequality in the
form
k(−∆)m/2 ukH ≤ Ck(−∆)(m+1)/2 ukH . (6.90)
This leads to
1d C ν
k(−∆)m/2 uk2H + νk(−∆)(m+1)/2 uk2H ≤ k(−∆)m/2 f k2H + k(−∆)(m+1)/2 uk2H
2 dt ν 4
C ν
+ k(−∆)m/2 uk4H + k(−∆)(m+1)/2 uk2H (6.91)
ν 4
C C ν
≤ k(−∆)m/2 f k2H + k(−∆)m/2 uk4H + k(−∆)(m+1)/2 uk2H .
ν ν 2
Therefore, we have
1d ν C C
k(−∆)m/2 uk2H + k(−∆)(m+1)/2 uk2H ≤ k(−∆)m/2 f k2H + k(−∆)m/2 uk4H . (6.92)
2 dt 2 ν ν
Looking at this as the differential inequality for y(t) = k(−∆)m/2 uk2H , we deduce that
C C C
ẏ ≤ k(−∆)m/2 f k2H + k(−∆)m/2 uk2H y(t) ≤ Cf + k(−∆)m/2 uk2H y(t), (6.93)
ν ν ν
with a finite constant Cf as f ∈ C ∞ (0, T ; T3 ). Grownwall’s inequality implies now that y(t)
obeys an upper bound
hC ˆ t i ˆ t hC ˆ t i
m/2 2
y(t) ≤ y(0) exp k(−∆) u(s)kH ds + Cf exp k(−∆)m/2 u(τ )k2H dτ ds.
ν 0 0 ν s
(6.94)
In other words, if we know that
ˆ T
k(−∆)m/2 u(s)k2H ds < +∞, (6.95)
0

then
sup k(−∆)m/2 u(s)k2H ds < +∞. (6.96)
0≤t≤T

61
This, in turn, implies that ˆ T
k(−∆)m/2 u(s)k4H ds < C, (6.97)
0

which can be inserted into (6.92) to conclude that

ˆ T
k(−∆)(m+1)/2 u(s)k2H ds < +∞, (6.98)
0

allowing us to build an induction argument and continue forever, meaning that

sup k(−∆)m/2 u(s)k2H ds < +∞, for any m ∈ N. (6.99)

0≤t≤T

This will, in turn, imply that u ∈ C ∞ by the Sobolev embedding theorem. However, this
argument uses the bound (6.88) which applies only for m > 3/2, and the “free” estimate for
the weak solution is
ˆ T ˆ T
2
k∇u(s)kH ds = k(−∆)1/2 u(s)k2H ds < +∞, (6.100)
0 0

which corresponds to m = 1, and for which we may not use this argument. Hence, to start
the induction we need the assumption that
ˆ T
k∆u(s)k2H ds < +∞, (6.101)
0

which corresponds to taking m = 2 > 3/2, allowing us to proceed.

The proof of Lemma 6.8

Recall that
k(−∆)m/2 P (u · ∇v)kH = sup ((−∆)m/2 (u · ∇v), w). (6.102)
w∈H,kwkH =1

Let us write X  X 
u · ∇v(x) = (2πi) (l · uj )vl e2πik·x , (6.103)
k∈Z3 j+l=k

so that
X  X 
((−∆)m/2 (u · ∇v), w) = (2πi)(4π 2 |k|2 )m/2 (l · uj )vl · w−k (6.104)
k∈Z3 j+l=k
X
= (2πi)(4π 2 |k|2 )m/2 (l · uj )(vl · wk ).
j+l+k=0

Next, we will use the inequality

|j + l|m ≤ (|j| + |l|)m ≤ Cm (|j|m + |l|m ), (6.105)

62
which implies
X X
|((−∆)m/2 (u · ∇v), w)| ≤ C |k|m |l||uj ||vl ||wk | ≤ C (|j|m + |l|m )|l||uj ||vl ||wk |
j+l+k=0 j+l+k=0
X X
≤C |l|m+1 |uj ||vl ||wk | + C |j|m |l||uj ||vl ||wk | = A + B. (6.106)
j+l+k=0 j+l+k=0

For the first term, we may estimate

X X X
A=C |l|m+1 |uj ||vl ||wk | = |uj | |l|m+1 |vl ||w−j−l | (6.107)
j+l+k=0 j∈Z3 l∈Z3
X X 1/2  X 1/2 X
2m+2 2 2
≤ |uj | |l| |vl | |wl | = k(−∆)(m+1)/2 vkH kwkH |uj |.
j∈Z3 l∈Z3 l∈Z3 j∈Z3

For the last sum above we may use the estimate

X X 1/2  X 1 1/2 X 1/2
|uj | ≤ |j|2m |uj |2 ≤ C |j|2m
|uj |2
= Ck(−∆)m/2 ukH .
|j|2m
j∈Z3 j∈Z3 j∈Z3 3 j∈Z
(6.108)
We used in the last step the assumption that m > 3/2 (in a dimension n we would have
needed to assume that m > n/2). For the second term in (6.106) we write
X X X
B=C |j|m |l||uj ||vl ||wk | = |l||vl | |j|m |uj ||w−l−j | (6.109)
j+l+k=0 l∈Z3 j∈Z3
X
≤ Ck(−∆)m/2 ukH kwkH |l||vl |,
l∈Z3

and
X X 1/2  X 1 1/2
|l||vl | ≤ |l|2+2m |vl |2 2m
≤ Ck(−∆)(m+1)/2 vkH , (6.110)
|l|
l∈Z3 l∈Z3 3
l∈Z

as m > 3/2. This shows that for any w ∈ H we have

|((−∆)m/2 (u · ∇u), w)| ≤ Ck(−∆)m/2 ukH k(−∆)(m+1)/2 vkH kwkH , (6.111)

and thus finishes the proof of Lemma 6.8. 2

Local in time existence in higher Sobolev spaces

The arguments of the previous section imply also that the Navier-Stokes equations are locally
well-posed in the higher Sobolev spaces H m (T3 ). We state it for simplicity for the case f = 0.

Theorem 6.9 Let u0 ∈ H m , with m ≥ 2, and f = 0. There exists a constant Cm > 0 which
depends on ν, m ≥ 1 and ku0 kH m , so that if

Tm ≤ Cm , (6.112)

63
then the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u, t > 0, x ∈ T3 , (6.113)

∇ · u = 0,
u(0, x) = u0 (x),

have a strong solution on the time interval [0, Tm ] that satisfies

ku(t)k2H m ≤ C0−1 , (6.114)

for all 0 ≤ t ≤ Tm .

The proof is familiar: we start with (6.115) with f = 0:

1d ν C
k(−∆)m/2 uk2H + k(−∆)(m+1)/2 uk2H ≤ k(−∆)m/2 uk4H . (6.115)
2 dt 2 ν
Looking at this as the differential inequality for y(t) = k(−∆)m/2 uk2H , we deduce that

C 2
ẏ ≤ y (t). (6.116)
ν
As a consequence, y(t) remains finite for a time that depends only on y(0). 2

Infinite time blow-up implies a finite time blow-up

The problem of blow-up of solutions of a nonlinear partial differential equation usually consists
in two separate problems: (1) can solutions blow-up in a finite time, and (2) can they blow-up
in an infinite time, in the sense that the norm of the solutions tends to infinity as t → +∞?
The second notion is usually much weaker. For example, solutions of the heat equation with
a linear growth term
ut = ∆u + u, t > 0, x ∈ Rn , (6.117)
have the long time behavior

et ku0 kL1 −|x|2 /(4t)

u(t, x) ∼ e , (6.118)
(4πt)n/2

and thus “blow-up in an infinite time” – all its Lp -norms, p ≥ 1 tend to infinity as t → +∞.
However, one does not normally think of these solutions as really “blowing-up” – they just
grow in time.
The situation is different for the Navier-Stokes equations: an infinite time blow-up implies
a finite-time blow-up. More precisely, let us assume that there exists a strong solution u(t, x)
of the Navier-Stokes equations

ut + u · ∇u + ∇p = ν∆u, 0 < t ≤ T, x ∈ T3 , (6.119)

∇ · u = 0,
u(0, x) = u0 (x),

64
such that u0 ∈ H, and
lim ku(t)kV = +∞. (6.120)
t→+∞

Assuming that such u exists, and given any T > 0, we will now construct an initial con-
dition v0 ∈ V so that the solution of (6.119) with v(0, x) = v0 (x), blows up before the
time T > 0. That is, there will be a time T1 ∈ (0, T ] such that

lim kv(t)kV = +∞. (6.121)

t→T

The idea is to combine the blow-up assumption that there exists a sequence of times tj → +∞
such that
ku(tj )kV ≥ 2j , (6.122)
with the main result of Proposition 6.2: solutions of the Navier-Stokes are often not large.
Given a sequence tj as in (6.122), we may use the aforementioned Proposition to find a
time sj ∈ [tj − T, tj ] so that
 1
ku(sj )kV ≤ C 1 + = C 0. (6.123)
T
The constant C depends only on ku0 kH , and ν > 0. Thus, if we take u(sj ) as the initial
condition for the Navier-Stokes equations, then the corresponding solution of the Cauchy
problem will have reached the V -norm that is larger than 2j by the time T . As ku(sj )kV is
uniformly bounded in j, we may choose a subsequence jk → +∞ so that vk0 (x) = u(sjk , x)
converges weakly in V and strongly in H to a function v0 ∈ V . Consider now the Cauchy
problem with the initial condition v0 :

vt + v · ∇v + ∇p = ν∆v, 0 < t ≤ T, x ∈ T3 , (6.124)

∇ · v = 0,
v(0, x) = v0 (x).

This problem has a strong solution on some time interval [0, T0 ], which depends only on kv0 kV
and ν. We will now show that (6.124) may not have a strong solution on the time inter-
val [0, T ]. To this end, assume that such solution exists on [0, T ], denote

r = sup kv(t)kV , (6.125)

0≤t≤T

and consider the functions vk (t) = u(t + sjk ), which are solutions of

∂vk
+ vk · ∇vk + ∇pk = ν∆vk , 0 < t ≤ T, x ∈ T3 , (6.126)
∂t
∇ · vk = 0,
vk (0, x) = vk0 (x).

Writing wj = vj − v, and expanding

vj · ∇vj − v · ∇v = (v + wj ) · ∇(v + wj ) − v · ∇v = wj · ∇v + v · ∇wj + wj · ∇wj , (6.127)

65
we see that wj satisfies (as in the proof of the uniqueness of the solutions of the Navier-Stokes
equations):
∂wj
+ wj · ∇v + v · ∇wj + wj · ∇wj + ∇p0 = ν∆wj , 0 < t ≤ T, x ∈ T3 , (6.128)
∂t
∇ · wj = 0,
wj (0, x) = vj0 (x) − v0 (x),
with p0 = pj − p. Multiplying by wj and integrating leads to
1d
kwj k2H + νkwj k2V = −(wj · ∇v, wj ). (6.129)
2 dt
We estimate the right side as
|(wj · ∇v, wj )| ≤ kwj kL3 k∇vkL2 kwj kL6 ≤ Ckwj kH 1/2 kvkV kwj kH 1 (6.130)
1/2 1/2 1/2 3/2 ν C
≤ Ckwj kH kwj kV kvkV kwj kV = CkvkV kwj kH kwj kV ≤ kwj k2V + 3 kvk4V kwj k2H .
2 ν
We used Young’s inequality in the last step, with p = 4/3, q = 4. Using this in (6.129) gives
1d ν C
kwj k2H + kwj k2V ≤ 3 kvk4V kwj k2H . (6.131)
2 dt 2 ν
As v is a strong solution, there exists C > 0, which depends on ν and r in (6.125), so that
kwj (t)kH ≤ kwj (0)kH eCt , (6.132)
meaning that wj (t) → 0 strongly in H, for all 0 ≤ t ≤ T . Furthermore, as
ˆ ˆ T
ν T 2 2
kwj (t)kV dt ≤ kwj (0)kH + C kwj (t)k2H dt, (6.133)
2 0 0

and since kwj (t)kH → 0, pointwise in t, while kwj (t)kH ≤ C, we conclude that
ˆ T
kwj (t)k2V dt → 0 as j → ∞. (6.134)
0

In particular, possibly after extracting another subsequence, we know that kwj (t)kV → 0 for
a.e. t ∈ [0, T ]. Take any t ∈ [0, T ] such that kwj kV ≤ 1, then
kvj (t)kV ≤ kwj (t)kV + kv(t)kV ≤ 1 + r. (6.135)
The local in time existence theorem implies that there exists a time T1 , which depends only
on ν, so that
kvj (s)kV ≤ 10(1 + r), (6.136)
for all s ∈ [t, t + T1 ]. The density of times t so that (6.135) holds, means that (6.136) holds
for all 0 ≤ t ≤ T . This, however, contradicts the assumption that
kv(sj − tj )kV = ku(sj )kV ≥ 2j .
Thus, v(s, x) can not be a strong solution on the time interval [0, T ].

66
The Beale-Kato-Majda regularity criterion
We now describe a sufficient condition for the solution to remain smooth. This time, we will
work in the whole space R3 but the existence and regularity results we have proved for the
three-dimensional torus apply essentially verbatim to the whole space as well. As we have
seen in Theorem 6.9, if the H m -norms of a smooth solution u(t, x) remain finite on a time
interval [0, T ], then the solution may be extended past the time T . In other words, a time T
is the maximal time of existence of a smooth solution u(t, x) if and only if

lim ku(t)kH m = +∞. (6.137)

t↑T

The Beale-Kato-Majda criterion reformulates this condition in terms of the vorticity.

Theorem 6.10 Let u0 ∈ Cc∞ (R3 ), so that there exists a classical solution v to the Navier-
Stokes equations with f = 0. If for any T > 0 we have
ˆ T
kω(t)kL∞ dt < +∞, (6.138)
0

then the smooth solution u exists globally in time. If the maximal existence time of the smooth
solution is T < +∞, then necessarily we have
ˆ T
lim kω(t)kL∞ dt = +∞. (6.139)
t↑T 0

The starting point in the proof is the estimate for the evolution of the H m -norms. We take m
to be an even integer fro convenience. Let us recall the identity (6.87) with f = 0:
1d
k(−∆)m/2 uk2H + νk(−∆)(m+1)/2 uk2H = ((−∆)m/2 (u · ∇u), (−∆)m/2 u). (6.140)
2 dt
Note that the term in the right side that has the highest order derivative, of the order (m+1),
vanishes
((u · ∇(−∆)m/2 u), (−∆)m/2 u) = 0,
hence the right side in (6.140) can be estimated by
3 X
X m
Cm kDm uk2 kDk uj kLp kD(m+1−k) ui kLq , (6.141)
i,j=1 k=1

with 1/p + 1/q = 1/2, and with the notation D = (−∆)1/2 . We recall a Gagliardo-Nirenberg
inequality for Rd :
kDj f kLp ≤ CkDm f ka2 kf k1−a
L∞ , (6.142)
with 0 ≤ j < m, and
1 j 1 m
= + a( − ),
p d 2 d
and a = j/m. We will use it for f = Du and 1 ≤ k < m:

kDk−1 DukLp ≤ ckDm−1 DukaL2 kDuk∞

1−a
, (6.143)

67
that is, the terms in (6.141) with 1 ≤ k < m can be estimated as

kDk ukLp ≤ ckDm ukaL2 kDuk1−a

∞ , (6.144)

with (you want to take a as small as possible)

k−1
a= ,
m−1
so that
1 k−1 k−1 1 m−1 k−1 a
= + ( − )= = .
p d m−1 2 d 2(m − 1) 2
The paired terms kDm+1−k ukq can be estimated similarly:

kDm+1−k ukLq = kDm−k DukLq ≤ ckDm ukbL2 kDuk1−b

∞ , (6.145)

with
m−k
b= ,
m−1
and
1 m−k m−k 1 m−1 m−k b
= + ( − )= = .
q d m−1 2 d 2(m − 1) 2
Luckily, we have a + b = 1, and
1 1 a+b 1
+ = = ,
p q 2 2
so that these p and q can be taken in (6.141). It follows that

kDk uj kLp kD(m+1−k) ui kLq ≤ CkDm ukL2 kDukL∞ .

When k = m or k = 1, we simply use p = 1/2 and q = ∞, getting the estimate

kDm ukL2 kDukL∞

for those terms. Altogether we conclude that

1d
kDm uk2H ≤ CkDm uk2H k∇ukL∞ . (6.146)
2 dt
Summing over m, we conclude that for any s ∈ N we have
d
kukH s ≤ Cs k∇ukL∞ kukH s . (6.147)
dt
Therefore, if u0 ∈ Cc∞ (R3 ), then for any of the H s -norms to become infinite by a time T it is
necessary that ˆ T
k∇u(t)kL∞ dt = +∞, (6.148)
0
and, in general, we have
n ˆ t o
kukH s ≤ ku0 kH s exp Cs k∇u(τ )kL∞ dτ . (6.149)
0

68
In a similar vein, multiplying the vorticity equation

ωt + u · ∇ω = ν∆ω + ω · ∇u (6.150)

by ω and integrating, we see that

d
kω(t)kL2 ≤ k∇ukL∞ kωkL2 , (6.151)
dt
so that nˆ t o
kω(t)kL2 ≤ kω0 kL2 exp k∇u(s)kL∞ ds . (6.152)
0

The conclusion of Theorem 6.10 would follow from (6.148) if we would know that

“k∇ukL∞ ≤ CkωkL∞ ”. (6.153)

One may expect this to be true based on its validity for L2 -norms: recall (3.26)
ˆ ˆ
2
|∇u| dx = |ω|2 dx, (6.154)
R3 R3

because

|ω|2 = εijk εimn (∂j uk )(∂m un ) = (δjm δkn − δjn δkm )(∂j uk )(∂m un ) = |∇u|2 − (∂j uk )(∂k uj ),
(6.155)
and ˆ ˆ
(∂j uk )(∂k uj )dx = − uk (∂k ∂j uj )dx = 0. (6.156)
Rn Rn

Identity (6.153), however, is not quite true for the L∞ -norms – the relation between the
gradient of the velocity and the vorticity is in terms of a singular integral operator which
maps every Lp → Lp for 1 < p < +∞ but does not map L∞ to L∞ . However, it is “almost
true” as shown by the following lemma.

Lemma 6.11 Let u(x) be a smooth divergence free velocity field in L2 ∩L∞ , and let ω = ∇×u.
There exists a constant C > 0 so that

k∇ukL∞ ≤ C(1 + log+ kukH 3 + log+ kωkL2 )(1 + kωkL∞ ). (6.157)

Here, for z > 0, we set log+ z = log z if log z > 0, and log+ z = 0 otherwise. The L2 -norm
of ω(t) that appears in (6.157) can be estimated from (6.152) as
ˆ t
+ +
log kω(t)kL2 ≤ log kω0 kL2 + k∇u(s)kL∞ ds. (6.158)
0

Similarly, the H 3 -norm of u(t) can be bounded as in (6.149):

ˆ t
+ +
log ku(t)kH 3 ≤ log ku0 kH 3 + C k∇u(s)kL∞ ds. (6.159)
0

69
Assuming the result of Lemma 6.11, we deduce that k∇uk∞ satisfies the inequality
 ˆ t 
k∇u(t)kL∞ ≤ C0 1 + k∇u(s)kL∞ ds (1 + kω(t)kL∞ , (6.160)
0

with a constant C0 that depends on the initial data u0 . Setting

ˆ t
G(t) = k∇u(s)kL∞ ds, β(t) = 1 + kω(t)kL∞ ,
0

we have from (6.160):

dG
≤ C0 (1 + G(t))β(t),
dt
so that ˆ t ˆ t
d n o n o
G(t) exp − C0 β(s)ds) ≤ C0 β(t) exp − C0 β(s)ds .
dt 0 0
Integrating in time gives
n ˆ t o n ˆ t o
G(t) exp − C0 β(s)ds) ≤ 1 − exp − C0 β(s)ds , (6.161)
0 0

so that n ˆ t o
G(t) ≤ exp C0 β(s)ds) .
0
In other words, we have
ˆ t n ˆ t o
k∇u(s)kL∞ ds ≤ exp C0 t + C0 kω(s)kL∞ ds) . (6.162)
0 0

As a consequence, as long as ˆ t
kω(s)kL∞ ds < +∞, (6.163)
0

all H m -norms of the velocity remain finite, hence u(t) ∈ C ∞ (R3 ). Therefore, the proof of
Theorem 6.10 boils down to Lemma 6.11.

The proof of the estimate on k∇ukL∞

We now prove Lemma 6.11 using the ideas from the theory of singular integral operators.
The velocity field is related to vorticity by the Biot-Savart law:
ˆ ˆ
u(x) = − K(x − y)ω(y)dy = K(y)ω(x + y)dy, (6.164)
R3 R3

with
1
K(x)h = x × h, (6.165)
4π|x|3

70
for any h ∈ R3 . As the singularity in ∇K(x) is of the order 1/|x|3 which is not integrable in
three dimensions, we have to be careful about computing the gradient of u. Let us write
ˆ
u(x + z) − u(x) = K(y)[ω(x + z + y) − ω(x + y)]dy. (6.166)
R3

As K ∈ L1loc (R3 ), if, say, ω ∈ C0∞ (R3 ), then, passing to the limit z → 0, we get
ˆ
∂uk (x)
= Kkm (y)∂j ωm (x + y)dy. (6.167)
∂xj R3

Because of the singularity in K we can not immediately integrate by parts. Let us write this
integral as
ˆ
∂uk (x)
= lim Kkm (y)∂j ωm (x + y)dy =
∂xj ε→0 |y|≥ε
ˆ ˆ
yj
= − lim Kkm (y)ωm (x + y) dy − lim [∂j Kkm (y)]ωm (x + y)dy = Akj + Bkj .
ε→0 |y|=ε |y| ε→0 |y|≥ε

The first integral can be re-written as

ˆ ˆ
yj 1 1 yj
Akj = − lim Kkm (y)ωm (x + y) dy = − lim 3
[y × ω(x + y)]k dy
ε→0 |y|=ε |y| ε→0 4π |y|=ε |y| |y|
ˆ ˆ
1 1 zj 2 1
= − lim [εz × ω(x + εz)] k ε dz = − [z × ω(x)]k zj dz
ε→0 4π |z|=1 ε3 |z|3 |z| 4π |z|=1
ˆ
1 kmn 1
= − kmn zm ωn (x)zj dz = ωn (x)δmj = − kjn ωn (x). (6.168)
4π |z|=1 3 3

Thus, we have
1
|Akj | ≤ kωkL∞ ,
3
and the main focus is on the second term. We have
krm
Kkm (y) = yr ,
4π|y|3

so that
3krm kjm
∂j Kkm (y) = − 5
yj yr + .
4π|y| 4π|y|3
We conclude that for any h ∈ R3 we have
ˆ h 3
krm kjm i
(Bh)k = − lim − y j y r + ωm (x + y)hj dy
ε→0 |y|≥ε 4π|y|5 4π|y|3
ˆ  3(y · h)[y × ω(x + y)]
k 1 
= lim + [ω(x + y) × h]k dy.
ε→0 |y|≥ε 4π|y|5 4π|y|3

71
We shall split B further as follows: take a smooth cut-off function ρ(r) so that ρ(r) = 0
for r > 2R, and ρ(r) = 1 for r < R, with R to be chosen later, and write
ˆ  3(y · h)[y × ω(x + y)]
k 1 
(Bh)k = lim + [ω(x + y) × h]k ρ(|y|)dy
ε→0 |y|≥ε 4π|y|5 4π|y|3
ˆ  3(y · h)[y × ω(x + y)]
k 1 
+ lim + [ω(x − y) × h]k (1 − ρ(|y|)dy = Ck + Dk .
ε→0 |y|≥ε 4π|y|5 4π|y|3

The Cauchy-Schwartz inequality implies that

ˆ ∞ 1 1/2 C
2
|Dk | ≤ C|h|kωkL2 6
r dr ≤ 3/2 kωkL2 |h|. (6.169)
R r R
The key estimate is for Ck : we will show that for any δ > 0 and any Hölder regularity
exponent γ ∈ (0, 1) we have
n R o
|Ck | ≤ C δ γ kωkC γ + kωkL∞ max(1, log |h|. (6.170)
δ
Here, kωkC γ is the Hölder norm. The Sobolev inequality in dimension n
n
kf kC γ (Rn ) ≤ Ckf kH s+γ (Rn ) , s >
2
implies that in three dimensions we have, for all 0 < γ < 1/2:

kωkC γ ≤ CkωkH 2 ,

so that
n R o n R o
|Ck | ≤ C δ γ kωkH 2 +kωkL∞ max(1, log ≤ C δ γ kukH 3 +kωkL∞ max(1, log . (6.171)
δ δ
Altogether, we have
 C n R o
k∇ukL∞ ≤ C kωkL∞ + 3/2 kωkL2 + δ γ kukH 3 + kωkL∞ max(1, log . (6.172)
R δ
2/3
Thus, we set R = kωkL2 . As far δ is concerned, if kukH 3 ≤ 1, we can take δ = 1, while
if kukH 3 ≥ 1, we can take δ = kuk−γ
H 3 . In both cases, we have

k∇ukL∞ ≤ C(1 + log+ kukH 3 + log+ kωkL2 )(1 + kωkL∞ ), (6.173)

which is the claim of Lemma 6.11. It remains, therefore, only to prove the estimate (6.170).

A nearly L∞ → L∞ estimate for singular integral operators

We now prove estimate (6.170) for Ck , which we write as
ˆ  3(y · h)[y × ω(x + y)]
k 1 
Ck = lim + [ω(x + y) × h]k ρ(|y|)dy
ε→0 |y|≥ε 4π|y|5 4π|y|3

72
 ˆ
1   dy
= lim 3(ŷ · h)[ŷ × ω(x + y)]k + [ω(x + y) × h]k ρ(|y|) 3
4π ε→0 |y|≥ε |y|
ˆ
1   dy
= lim 3ŷm hm εkjr ŷj ωr (x + y) + εkrm ωr (x + y)hm ρ(|y|) 3
4π ε→0 |y|≥ε |y|
ˆ
hm   dy
= lim 3ŷm εkjr ŷj + εkrm ωr (x + y)ρ(|y|) 3
4π ε→0 |y|≥ε |y|
ˆ
hm
= lim Pmkr (y)ωr (x + y)ρ(|y|)dy. (6.174)
4π ε→0 |y|≥ε

We have denoted here

1
Pmkr = (3ŷm εkjr ŷj + εkrm ). (6.175)
|y|3
The kernel Q(y) = Pmkr (y) (we fix for the moment the indices m, k and r) is homogenous of
degree (−3):
1
Q(λy) = 3 Q(y), for all λ > 0 and y ∈ R3 , y 6= 0. (6.176)
λ
Thus, Q(y) is “barely not in L1 ”: if it were slightly less singular it would have been in L1 .
In addition, the average of Q(y) over the unit sphere (and thus over any sphere centered
at y = 0) vanishes:
ˆ ˆ
Q(y)dy = (3ŷm εkjr ŷj + εkrm )dy = 4π[εkjr δmj + εkrm ] = 4π[εkmr = εkrm ] = 0.
|y|=1 |y|=1
(6.177)
Consider now the term (again, with an index r fixed)
ˆ
Qω(x) = lim Q(y)ωr (x + y)ρ(|y|)dy. (6.178)
ε→0 |y|≥ε

We split the integration in the definition of Qω as follows:

ˆ ˆ
Qω(x) = lim Q(y)ωr (x + y)ρ(|y|)dy + Q(y)ωr (x + y)ρ(|y|)dy = A + B. (6.179)
ε→0 ε≤|y|≤δ |y|≥δ

The second term above is (recall that ρ(|y|) = 0 for |y| > 2R):
ˆ
B= Q(y)ωr (x + y)ρ(|y|)dy, (6.180)
δ≤|y|≤2R

which can be estimated as

ˆ 2R
rn−1 2R
|B| ≤ CkωkL∞ n
dr ≤ CkωkL∞ log . (6.181)
δ r δ
The first term in (6.179) is estimated using the Hölder continuity of ω: the mean-zero prop-
erty (6.177) means that we can write
ˆ ˆ
A = lim Q(y)ωr (x + y)ρ(|y|)dy = lim Q(y)[ωr (x + y) − ωr (x)]ρ(|y|)dy. (6.182)
ε→0 ε≤|y|≤δ ε→0 ε≤|y|≤δ

73
The Hölder continuity of ω implies that the integrand in the last expression above has an
upper bound
C C
|Q(y)[ωr (x − y) − ωr (x)]ρ(|y|)| ≤ n
|y|γ kωkC γ = n−γ kωkC γ , (6.183)
|y| |y|

which is integrable in y at y = 0 for γ > 0. Therefore, we have

ˆ
A= Q(y)[ωr (x − y) − ω(x)]ρ(|y|)dy, (6.184)
0≤|y|≤δ

and ˆ δ
rn−1
|A| ≤ Ckωk Cγ n−γ
dy ≤ CkωkC γ δ γ . (6.185)
0 r
Putting the bounds for A and B together gives (6.170).

7 Vortex lines and geometric conditions for blow-up

The vorticity growth equation
Here, we investigate how vorticity alignment in the regions of high vorticity can prevent blow-
up in the Navier-Stokes and Euler equations. First, we obtain an equation for the magnitude
of vorticity |ω| that shows that it is plausible that the vorticity alignment in the regions of
high vorticity may prevent the growth of vorticity. Recall that the vorticity of the solutions
of the Navier-Stokes equations satisfies the evolution equation

ωt + u · ∇ω − ν∆ω = ω · ∇u (7.1)

Multiplying by 2ω, we obtain

∂t (|ω|2 ) + u · ∇(|ω|2 ) − ν∆|ω|2 + 2ν|∇ω|2 = 2(ω · ∇u) · ω. (7.2)

The right side can be written as

2(ω · ∇u) · ω = 2ωj (∂j uk )ωk = 2(Sω · ω) = 2α(x)|ω|2 ,

with
ω(x)
α(x) = (S(x)ξ(x) · ξ(x)), ξ(x) = , (7.3)
|ω(x)|
and
1
S(x) = (∇u + (∇u)t ). (7.4)
2
When ν = 0 we get a particularly simple form of the vortex stretching balance for the Euler
equations:
∂t |ω| + u · ∇|ω| = α(t, x)|ω|. (7.5)

74
Thus, the vorticity growth may only appear from α(x) large. Our next task is to express α(x)
in terms of the vorticity alignment. We start with the Biot-Savart law
ˆ
1 y
u(x) = × ω(x + y)dy. (7.6)
4π R3 |y|3
Let us recall that
ˆ
∂uk (x)
= lim Kkm (y)∂j ωm (x + y)dy (7.7)
∂xj ε→0 |y|≥ε
ˆ ˆ
yj
= − lim Kkm (y)ωm (x + y) dy − lim [∂j Kkm (y)]ωm (x + y)dy = Akj + Bkj .
ε→0 |y|=ε |y| ε→0 |y|≥ε

The term Akj can be simplified as

ˆ ˆ
yj 1 1 yj
Akj = − lim Kkm (y)ωm (x + y) dy = − lim 3
[y × ω(x + y)]k dy
ε→0 |y|=ε |y| ε→0 4π |y|=ε |y| |y|
ˆ ˆ
1 1 zj 1
= − lim 3 3
[εz × ω(x + εz)]k ε2 dz = − [z × ω(x)]k zj dz
ε→0 4π |z|=1 ε |z| |z| 4π |z|=1
ˆ
1 kmn 1
= − kmn zm ωn (x)zj dz = − ωn (x)δmj = − kjn ωn (x), (7.8)
4π |z|=1 3 3

and B can be written as

ˆ h 3
krm kjm i
Bkj = lim yj yr − ωm (x + y)dy.
ε→0 |y|≥ε 4π|y|5 4π|y|3

Multiplying (7.7) by ijk and summing over j, k, leads now to an integral equation for the
vorticity:
1
ωi (x) = ijk ∂j uk = ijk Akj + ijk Bkj = − ijk kjn ωn (7.9)
ˆ 3
h 3  i
krm kmj
+ lim ijk yj yr + ωm (x + y)dy.
ε→0 |y|≥ε 4π|y|5 4π|y|3

The first term above can be re-written as

−ijk kjn ωn = ijk njk ωn = 2ωi .

In the second term, we use the identities

ijk krm yj yr ωm = kij krm yj yr ωm = [δir δjm − δim δjr ]yj yr ωm = yi (y · ω) − |y|2 ωi ,

and
ijk kmj ωm = kij kmj ωm = 2ωi
Using these transformations in (7.9), gives
ˆ
1 h 3
2 2ωi (x + y) i
ωi (x) = lim [y i (y · ω(x + y)) − |y| ω i (x + y)] + dy.
3 ε→0 |y|≥ε 4π|y|5 4π|y|3

75
so that
ˆ
3 dy
ω(x) = lim σ(ŷ)ω(x + y) (7.10)
4π ε→0 |y|≥ε |y|3

with the matrix σ(ŷ), ŷ = y/|y|, defined as

σ(ŷ) = 3(ŷ ⊗ ŷ) − I. (7.11)

Similarly, we may compute the symmetric part of ∇u:

1
S(x) = (∇u + (∇u)t ).
2
We have
1 1
Skj = (Akj + Ajk ) + (Bkj + Bjk ).
2 2
It is easy to see that the matrix Akj is anti-symmetric, thus

Akj + Ajk = 0.

For the symmetric part of the matrix B we compute

ˆ h 3
krm 3jrm kjm jkm i
Bkj + Bjk = lim yj yr + yk yr − − ωm (x + y)dy
ε→0 |y|≥ε 4π|y|5 4π|y|5 4π|y|3 4π|y|3
ˆ
3 h i dy
= lim krm ŷj ŷr + jrm ŷk ŷr ωm (x + y) 3 .
4π ε→0 |y|≥ε |y|

We conclude that ˆ
3 dy
S(x) = P.V. M (ŷ, ω(x + y)) , (7.12)
4π |y|3
with the matrix-valued function
1
M (ŷ, ω) = [(ŷ × ω) ⊗ ŷ + ŷ ⊗ (ŷ × ω)]. (7.13)
2
Going back to (7.3), we get the following expression for the vorticity stretching coefficient α(x):
ˆ
3 dy
α(x) = (S(x)ξ(x) · ξ(x)) = P.V. (M (ŷ, ω(x + y))ξ(x) · ξ(x)) 3 . (7.14)
4π |y|
The integrand can be re-written as
1
M (ŷ, ω(x + y))ξ(x) · ξ(x)) = [(ŷ × ω(x + y)) ⊗ ŷ + ŷ ⊗ (ŷ × ω(x + y))]ξ(x) · ξ(x)
2
= (ŷ × ω(x + y) · ξ(x))(ŷ · ξ(x)) = D(ŷ, ξ(x + y), ξ(x))|ω(x + y)|,

thus
ˆ
3 dy
α(x) = (S(x)ξ(x) · ξ(x)) = P.V. D(ŷ, ξ(x + y), ξ(x))|ω(x + y)| . (7.15)
4π |y|3

76
Here, we have defined, for three unit vectors e1 , e2 and e3 :

D(e1 , e2 , e3 ) = (e1 · e3 )Det(e1 , e2 , e3 ).

Geometrically, it follows that the regions where ξ(x + y) is aligned with ξ(x) contribute less
to α(x). This applies also to the antiparallel vortex pairing, which is a physically observed
phenomenon. That is, we expect that if the vorticity direction field is aligned or anti-aligned
in the regions of high vorticity, the blow-up might be prevented by the vorticity alignment,
though this requires a careful analysis which we will undertake next.

A priori bounds on the strain matrix

Let us first obtain some bounds on the strain matrix in terms of ω that we will need later.
We have, from (7.12)-(7.13):
ˆ h ˆ
3 i dy 3
Skj (x) = P.V. krm ŷj ŷr + jrm ŷk ŷr ωm (x + y) 3 = P.V. Rkjm (y)ωm (x + y)dy,
8π |y| 8π
(7.16)
with the kernel
1
Rkjm (y) = 3 [krm ŷj ŷr + jrm ŷk ŷr ].
|y|
This kernel is of the singular integral type we have seen before in the Beale-Kato-Majda
criterion: it is homogeneous of degree (−n) (the dimension n = 3), in the sense that

Rkjm (λy) = λ−3 Rkmj (y), (7.17)

and its integral over any sphere centered at y = 0 vanishes:

ˆ
1 1
Rkjm (y)dy = [krm δjr + jrm δkr ] = [kjm + jkm ] = 0. (7.18)
|y|=1 3 3

Let us show that (7.17) and (7.18) imply that the Fourier transform R̂kjm (ξ) is uniformly
bounded:
|R̂kjm (ξ)| ≤ C. (7.19)
Indeed, let us write ˆ
1
Rkjm (y) = 3 Φ(ŷ), Φ(y)dy = 0.
|y| |y|=1

As Rkjm (y) is homogeneous of degree (−n) (in dimension n = 3), its Fourier transform is
homogeneous of degree zero. Then we have:
ˆ 1/δ ˆ ˆ 1ˆ h
1 2πir(ξ·ŷ) 2
i drdŷ
R̂kjm (ξ) = lim 3
e Φ(ŷ)r drdŷ = lim cos(2πr(ξ · ŷ)) − 1 Φ(ŷ)
ε,δ→0 ε S2 r ε→0 ε S2 r
ˆ 1/δ ˆ ˆ 1/δ ˆ
drdŷ drdŷ
+ lim cos(2πr(ξ · ŷ))Φ(ŷ) + i lim sin(2πr(ξ · ŷ))Φ(ŷ)
δ→0 1 S2 r ε,δ→0 ε S2 r
= A1 + A2 + A3 . (7.20)

77
We used the mean-zero property of Φ(ŷ) in the second equality above. For A3 , we may write
ˆ ˆ 1/δ
drdŷ
A3 (ξ) = i lim Φ(ŷ) sin(2πr(ξ · ŷ))
ε,δ→0 S2 ε r
ˆ ˆ
 2π|ξ·ŷ|/δ sin rdr 
= i lim Φ(ŷ)sgn(ξ · ŷ) dŷ.
ε,δ→0 S2 2π|ξ·ŷ|ε r

Recall that there exists a constant C0 > 0 so that for any a, b > 0 we have
ˆ b
sin rdr
≤ C0 ,
a r

hence |A3 (ξ)| ≤ C. For A1 + A2 , we have

ˆ hˆ 1/ε
dr i
A1 (ξ) + A2 (ξ) = lim Φ(ŷ) [cos(2πr(ξ · ŷ)) − 1] dŷ
ε→0 S2 ε r
ˆ h ˆ 2π|ξ·ŷ|/ε ˆ hˆ 1 ˆ ∞
dr i dr cos rdr i
= lim Φ(ŷ) [cos(r) − 1] dŷ = Φ(ŷ) (cos r − 1) + dŷ
ε→0 S2 2π|ξ·ŷ|ε r S2 0 r 1 r
ˆ h ˆ 2π|ξ·ŷ|/ε dr i ˆ hˆ 1 ˆ ∞
dr cos rdr i
− lim Φ(ŷ) dŷ = Φ(ŷ) (cos r − 1) + dŷ
ε→0 S2 r r r
ˆ 1 S 2 0 1

− lim Φ(ŷ) log(2π|ξ · ŷ|/ε)dŷ

ε→0 S2
ˆ hˆ 1 ˆ ∞ ˆ
dr cos rdr i
= Φ(ŷ) (cos r − 1) + dŷ − Φ(ŷ) log(|ξˆ · ŷ|)dŷ. (7.21)
S2 0 r 1 r S2

We used the mean-zero property of Φ(ŷ) in the last step. In particular, it allowed us to
replace ξ by ξˆ under the logarithm sign. Now, the first integral in the last line in (7.21)
does not depend on ξ and is, therefore, uniformly bounded. The second is also bounded,
by an application of the Cauchy-Schwartz inequality on S2 . We conclude that the uniform
bound (7.19) holds. It follows immediately that the strain matrix satisfies an L2 -bound

kSkL2 ≤ CkωkL2 , (7.22)

a bound we have already seen before.

The regularized system

We will follow the paper by P. Constatin and C Fefferman for the analysis of the vorticity
alignment for the Navier-Stokes equations. A similar issue for the Euler equations has been
studied in their joint paper with A. Majda. We will start with a regularized Navier-Stokes
system, obtained by smoothing the advecting velocity:

ut + (φδ ∗ u) · ∇u + ∇p = ν∆u, t > 0, x ∈ Rn (7.23)

∇ · u = 0,
u(0, x) = u0 (x).

78
The convolution is performed in space only:
ˆ
uδ (t, x) = φδ ∗ u(t, x) = φδ (x − y)u(t, y)dy,

and the kernel φδ has the form

1 x
φ
φδ (x) = ,
δ3 δ
with a smooth compactly supported function φ(x) ≥ 0 with kφkL1 = 1. Note that uδ is also
divergence-free: ∇ · uδ = 0. Let us explain why the regularized system (7.23) has a strong
solution, which is smooth if u0 ∈ Cc∞ (R3 ). Of course, the easy bounds on u(t, x) will blow-up
as δ → 0. We argue as in the estimate for the evolution of the H m -norms in the proof of
the Beale-Kato-Majda criterion. First, multiplying (7.23) by u and integrating by parts we
deduce that ˆ ˆ tˆ ˆ
2 2
|u(t, x)| dx + ν |∇u(s, x)| dxds = |u0 (x)|2 dx, (7.24)
R3 0 R3
hence
ku(t)kL2 ≤ ku0 kL2 . (7.25)
It follows from the definition of uδ that

kuδ (t)kC k ≤ Ck (δ), (7.26)

with the constants Ck (δ) that may blow-up as δ → 0. Next, multiplying (7.23) by (−∆)m u
and integrating by parts we obtain
1d
k(−∆)m/2 uk2H + νk(−∆)(m+1)/2 uk2H = ((−∆)m/2 (uδ · ∇u), (−∆)m/2 u). (7.27)
2 dt
As before, the leading order term in the right side vanishes:

((uδ · ∇(−∆)m/2 u), (−∆)m/2 u) = 0,

because ∇ · uδ = 0. Hence, using (7.26), the right side in (7.27) can be estimated by
3 X
X m
m
Cm kD uk2 kDk uδ,j kL∞ kD(m+1−k) ui kL2 ≤ C(δ)kuk2H m . (7.28)
i,j=1 k=1

Summing over m, we conclude that for any s ∈ N we have

d
kukH s ≤ Cs (δ)kukH s . (7.29)
dt
Therefore, if u0 ∈ Cc∞ (R3 ), then u(t) remains in all H m (R3 ) for all t > 0. Of course, the
Sobolev norms of u(t) may blow-up as δ → 0.

79
Vorticity alignment prevents blow-up
We will now show that if the direction of the vorticity of the solutions of the regularized
system (7.23) is sufficiently aligned then solutions of the Navier-Stokes system itself remain
regular. Let us introduce some notation: given a vector e we denote by Pe⊥ the projection
orthogonal to e,
Pe⊥ v = v − (v · e)e.
We will denote by u(t, x) the solution of the regularized system (7.23), let ω(t, x) = ∇×u(t, x)
be its vorticity and ξ(t, x) = ω(t, x)/|ω(t, x)|, while v(t, x) will be the solution of the true
Navier-Stokes equations

vt + v · ∇v + ∇p = ν∆v, t > 0, x ∈ Rn (7.30)

∇ · v = 0,
v(0, x) = u0 (x).

Theorem 7.1 Assume that there exists δ0 , Ω > 0 and ρ > 0 so that for all δ ∈ (0, δ0 ) the
solution u(t, x) of the regularized system (7.23) satisfies

⊥ |y|
Pξ(t,x) (ξ(t, x + y)) ≤ , (7.31)
ρ

for all x, y ∈ R3 and 0 ≤ t ≤ T , such that |ω(t, x)| > Ω and |ω(t, x+y) > Ω. Then the Navier-
Stokes equations (7.30) have a strong, and hence C ∞ -solution on the time interval 0 ≤ t ≤ T .

The strategy will be to get a priori bounds on u(t, x) that do not depend on δ and then pass
to the limit δ → 0. The passage ot the limit is very similar to what we have seen before, so
we focus on the a priori bounds that follow from assumption (7.30).

The a priori bounds for the regularized system

We first get a priori bounds for the regularized system that require no assumptions on the
direction of the vorticity and, in particular, are independent of (7.31). Let us set ω0 = ∇ × u0
and ˆ ˆ
25
Q= |ω0 (x)|dx + |u0 (x)|2 dx.
R3 ν R3

We have then the following bounds, uniform in δ > 0.

Lemma 7.2 The following two bounds hold:

ˆ ˆ tˆ
|ω(t, x)|dx + ν |ω(s, x)|∇ξ(s, x)|2 dxds ≤ Q, (7.32)
R3 0 {x:|ω(s,x)|>0}

for all 0 ≤ t ≤ T , and for any Ω > 0 we have

ˆ Tˆ
Q
|∇ξ(s, x)|2 dxds ≤ . (7.33)
0 {x:|ω(s,x)|>Ω} νΩ

80
Proof. Let us derive the equation for ω(t, x): this derivation follows that for the true Navier-
Stokes equations but the vorticity equation in the presence of the regularization is not identical
to that of the Navier-Stokes equations. The advection term in the regularized Navier-Stokes
equations can be written as

uδ · ∇u = u · ∇u + (uδ − u) · ∇u = u · ∇u − vδ · ∇u, (7.34)

with
vδ = u − uδ .
Recall that

(ω × u)i = εijk ωj uk = εijk εjmn (∂m un )uk = (δin δkm − δim δkn )(∂m un )uk
= (∂k ui )uk − (∂i uk )uk . (7.35)

We used above the identity

εjik εjmn = δim δkn − δin δkm (7.36)
and anti-symmetry of εijk . Thus, as we have previously seen, the advection term can be
written as  |u|2 
u · ∇u = ω × u + ∇ . (7.37)
2
Recall also the formula

∇ × (a × b) = −a · ∇b + b · ∇a + a(∇ · b) − b(∇ · a), (7.38)

which now gives

∇ × (u · ∇u) = ∇ × (ω × u) = −ω · ∇u + u · ∇ω. (7.39)
We also had an observation that
∂ui
ω · ∇u = V (t, x)ω, Vij = . (7.40)
∂xj

The matrix V can be split into its symmetric and anti-symmetric parts:
1 1
V = S + P, S = (V + V T ), P = (V − V T ), (7.41)
2 2
The anti-symmetric part has the form
1 1 1
Pij hj = [∂j ui − ∂i uj ]hj = ∂m uk [δik δjm − δim δjk ]hj = εlij εlkm (∂m uk )hj
2 2 2
1 1 1 1
= − εlij εlmk (∂m uk )hj = − εlij ωl hj = εilj ωl hj = [ω × h]i , (7.42)
2 2 2 2
for any h ∈ R3 . In other words, P satisfies
1
P h = ω × h, (7.43)
2

81
and thus has an explicit form
 
0 −ω3 ω2
1
P = ω3 0 −ω1  . (7.44)
2
−ω2 ω1 0
As a consequence, we have P ω = 0, thus V ω = Sω, so that
∇ × (u · ∇u) = u · ∇ω − Sω. (7.45)
This is, of course, identical to what we have obtained for the true Navier-Stokes equations.
For the term in (7.34), which involves vδ and comes from the regularization, we write
[∇ × (vδ · ∇u)]i = εijk ∂j [vδ,m ∂m uk ] = vδ,m ∂m [εijk ∂j uk ] + εijk (∂j vδ,m )(∂m uk ) (7.46)
= vδ · ∇ωi + εijk (∂j vδ,m )(∂m uk )
Thus, we have
∇ × (uδ · ∇u) = u · ∇ω − Sω − vδ · ∇ω + (∇u) (∇vδ ) = uδ · ∇ω − Sω + (∇u) (∇vδ ). (7.47)
Here, we have introduced the following notation: given two matrices a and b, the vector a b
has the entries
(a b)i = εijk akm bmj . (7.48)
Thus, the vorticity satisfies the evolution equation
ωt + uδ · ∇ω − ν∆ω = Sω − (∇u) (∇vδ ). (7.49)
Once again, we stress that the second term in the right side comes from the regularization.
Note that the vector ξ(t, x) = ω(t, x)/|ω(t, x)| satisfies |ξ|2 = 1, which implies
ξ · ξt = 0, ξj ∂k ξj = 0, for all 1 ≤ k ≤ 3, (7.50)
leading to
(∂k ξj )(∂k ξj ) + ξj ∆ξj = 0, (7.51)
Multiplying (7.49) by ξ(t, x), and using (7.50)-(7.51), we get in the left side
ξ · (ωt + uδ · ∇ω − ν∆ω) = ξ · (|ω|ξt + ξ|ω|t + |ω|(uδ · ∇)ξ + ξ(uδ · ∇|ω|))
−ν(ξ · ξ)∆|ω| − ν(ξ · ∆ξ)|ω| − 2νξk ∂j ξk ∂j |ω| = |ω|t + uδ · ∇|ω| − ν∆|ω| − ν|ω|(ξ · ∆ξ)
= |ω|t + uδ · ∇|ω| − ν∆|ω| + ν|ω||∇ξ|2 .
We deduce an evolution equation for |ω(t, x)| in the region where ω(t, x) 6= 0:
∂|ω|
+ uδ · ∇|ω| − ν∆|ω| + ν|ω||∇ξ|2 = ξ · (Sω − (∇u) (∇vδ )). (7.52)
∂t
Let now f (z) be a C 2 -function of a scalar variable z which vanishes in a neighborhood of z = 0.
Multiplying (7.52) by f 0 (|ω|) and integrating gives
ˆ ˆ ˆ
d 00
f (|ω|)dx + ν 2
f (|ω|)|∇|ω|| dx + ν |ω|f 0 (|ω|)|∇ξ|2 dx (7.53)
dt R3
ˆ R3 R3

= [ξ · (Sω − (∇u) (∇vδ ))]f 0 (|ω|)dx.

R3

82
 Choose a function ψ(y) ≥ 0 such that ψ(y) vanishes for |y| ≤ r0 and y > Ω0 , and such
that ˆ Ω0
ψ(y)dy = 1, (7.54)
0
and set ˆ z
f (z) = (z − y)ψ(y)dy, (7.55)
0
so that ˆ z
0
f (z) = ψ(y)dy, f 00 (z) = ψ(z) ≥ 0. (7.56)
0
In particular, we have 0 ≤ f 0 (z) ≤ 1, f 0 (z) = 0 in a neighborhood of z = 0, and

zf 0 (z) = z, for z > Ω0 . (7.57)

In other words, f (z) is an approximation to z. Then, integrating (7.53) in time gives

ˆ ˆ tˆ ˆ
2
f (|ω(t, x)|)dx + ν |ω(s, x)||∇ξ(s, x)| dx ≤ f (|ω0 (x)|)dx
R3 0 {x:ω(s,x)|>Ω0 } R3
ˆ tˆ
+ [ξ · (Sω − (∇u) (∇vδ ))]f 0 (|ω|)dxds (7.58)
R3
ˆ0
ˆ tˆ 
1 2 1 2 1 2 1 2

≤ |ω0 (x)|dx + |S(s, x)| + |ω(s, x)| + |∇u| + |∇vδ | dxds.
R3 0 R3 2 2 2 2
As ∇ · u = 0, we have
ˆ ˆ ˆ
2 2
|∇u| dx = |ω| dx = 2 TrS 2 dx.
R3 R3 R3

The energy identity (7.24) means that

ˆ ˆ tˆ
f (|ω(t, x)|)dx + ν |ω(s, x)||∇ξ(s, x)|2 dx ≤ Q, (7.59)
R3 0 {x:ω(s,x)|>Ω0 }

with
ˆ ˆ
25
Q= |ω0 (x)|dx + |u0 (x)|2 dx. (7.60)
R3 ν R3

In particular, for any Ω > 0 we obtain

ˆ tˆ
Q
|∇ξ(s, x)|2 dx ≤ . (7.61)
0 {x:ω(s,x)|>Ω} νΩ

We may also let Ω0 → 0 in (7.59), so that f (z) → z, and obtain the estimate in Lemma 7.2
ˆ ˆ tˆ
|ω(t, x)|dx + ν |ω(s, x)||∇ξ(s, x)|2 dx ≤ Q. (7.62)
R3 0 {x:ω(s,x)|>0}

This finishes the proof of this Lemma.

83
Enstrophy bounds when the vorticity direction is regular
Lemma 7.2 does not use assumption (7.31) on the vorticity direction. Now, we will use this
assumption to obtain enstrophy bounds on the solution of the regularized system. We will
show that the solution of the regularized system obeys the following a priori bounds. Here,
we use assumption (7.31): there exists δ0 , Ω > 0 and ρ > 0 so that for all δ ∈ (0, δ0 ) the
solution u(t, x) of the regularized system (7.23) satisfies

⊥ |y|
Pξ(t,x) (ξ(t, x + y)) ≤ , (7.63)
ρ
for all x, y ∈ R3 and 0 ≤ t ≤ T , such that |ω(t, x)| > Ω and |ω(t, x + y) > Ω.
Lemma 7.3 There exists a constant C which depends on the initial data u0 , and Ω, ν, T ,
and the constant ρ in (7.63), so that
ˆ
sup |ω(t, x)|2 dx ≤ C, (7.64)
0≤t≤T R3

and ˆ ˆ
T
|∇ω(t, x)|2 dx ≤ C, (7.65)
0 R3
for all δ ∈ (0, δ0 ).
With these a priori bounds in hand, one can find a subsequence δk ↓ 0, such that the so-
lutions u(t, x) of the regularized Navier-Stokes system converge to a solution v(t, x) of the
true Navier-Stokes equations which obeys the same bounds (7.64) and (7.65). These bounds
imply that v is a strong solution and is therefore smooth if u0 is smooth. Thus, our focus is
on proving Lemma 7.3.
Multiplying the vorticity equation

ωt + uδ · ∇ω − ν∆ω = Sω − (∇u) (∇vδ ) (7.66)

by ω and integrating gives

ˆ ˆ ˆ ˆ
1d 2 2
|ω| dx + ν |∇ω| dx = (Sω · ω)dx − ω · ((∇u) (∇vδ ))dx. (7.67)
2 dt
We will split the vorticity into the ”small” and ”large” components: take a cut-off func-
tion χ(z) such that χ(z) = 1 for 0 ≤ z ≤ 1, χ(z) = 0 for z ≥ 2, and 0 ≤ χ(z) ≤ 1 for
all z ≥ 0. We set
ω(t, x) = ω (1) (t, x) + ω (2) (t, x), (7.68)
with
 |ω(t, x)|    |ω(t, x)| 
ω (1) (t, x) = χ ω(t, x), ω (2) (t, x) = 1 − χ ω(t, x). (7.69)
Ω Ω
Recall that the strain matrix can be written in terms of the vorticity as
ˆ
3 dy y
S(x) = P.V. M (ŷ, ω(x + y)) 3 , ŷ = , (7.70)
4π |y| |y|

84
with the matrix-valued function
1
M (ŷ, ω) = [(ŷ × ω) ⊗ ŷ + ŷ ⊗ (ŷ × ω)]. (7.71)
2
The decomposition (7.68) and (7.70) induce then the corresponding decomposition

S(t, x) = S (1) (t, x) + S (2) (t, x). (7.72)

We can then write

2
X
(Sω · ω) = (S (i) ω (j) · ω (k) ) = X + Y + Z, (7.73)
i,j,k=1

where X comes from the triplets where at least one of ω is ”small”:

2
X X
X= (S (i) ω (j) · ω (k) ),
i=1 (j,k)6=(2,2)

the term Y has S ”small”, and both ω ”large”:

Y = (S (1) ω (2) · ω (2) ),

and, finally, Z has S and both ω ”large”:

Z = (S (2) ω (2) · ω (2) ).

We also set
W = −ω · ((∇u) (∇vδ )).
With this notation, (7.67) has the form
ˆ ˆ ˆ
1d 2 2
|ω| dx + ν |∇ω| dx = (X + Y + Z + W )dx. (7.74)
2 dt

We will estimate the size of each term in the right side of (7.74) separately.
In order to estimate X, we recall that for any incompressible flow v we have
ˆ ˆ
|∇v| dx = |ζ|2 dx, ζ = ∇ × v.
2

As a consequence, the strain matrix

1
Sv = (∇v + (∇v)t )
2
satisfies
3 ˆ  3 ˆ  ˆ ˆ
X ∂vi ∂vj 2 X ∂vi 2
kSv k2L2 = + dx ≤ 4 dx = 4 |∇v| dx = 4 |ζ|2 dx. (7.75)
2

i,j=1
∂xj ∂xi i,j=1
∂xj

85
Then, the term X can be estimated as follows: either ω (j) or ω (k) is ”small” and can be
bounded pointwise by Ω. This allows us to use the Cauchy-Schwartz inequality and (7.75):
ˆ
X(t, x)dx ≤ CΩkSkL2 kωkL2 ≤ CΩkωk2L2 . (7.76)

We have used the bound (7.22)

kSkL2 ≤ CkωkL2 . (7.77)
in the second inequality above.
Next, we note that Y is bounded from above by

|Y (t, x)| ≤ |S (1) (t, x)||ω(t.x)|2 , (7.78)

so that ˆ ˆ 1/2
(1)
|Y (t, x)|dx ≤ kS kL2 |ω(t.x)|4 dx . (7.79)

The Gagliardo-Nirenberg inequality in Rn :

1 1 a
kukLp ≤ Ck∇ukaL2 kuk1−a
L2 , = − ,
p 2 n

implies that in R3 we have

ˆ 1/2 ˆ 3/4  ˆ 1/4
4 2
|ω(x)| dx ≤C |∇ω(x)| dx |ω(x)|2 dx . (7.80)

Using this in (7.78) gives

ˆ
3/2 1/2 3/2 1/2
|Y (t, x)|dx ≤ kS (1) kL2 k∇ωkL2 kωkL2 ≤ Ckω (1) kL2 k∇ωkL2 kωkL2
ν C
≤ k∇ωk2L2 + 3 kω (1) k4L2 kωk2L2 . (7.81)
8 ν
We have used Young’s inequality in the last step, as well as the bound (7.77) for kS (1) kL2 .
The second term in the right side can be bounded with the help of the estimate (7.32) in
Lemma 7.2 as ˆ
(1) 2
kω kL2 ≤ 2Ω |ω(t, x)|dx ≤ 2ΩQ. (7.82)

Thus, the term Y can be estimated as

ˆ
ν C
|Y (t, x)|dx ≤ k∇ωk2L2 + 3 (ΩQ)2 kωk2L2 . (7.83)
8 ν
Before looking at Z, which is the most difficult term, we bound W :

W = −ω · ((∇u) (∇vδ )).

86
This term is only there because of the regularization and should disappear as δ → 0. Note
that
ˆ ˆ
kvδ kL2 = ku − uδ kL2 = ku − φδ ∗ ukL2 = |1 − φ̂δ (ξ)| |û(ξ)| dξ = |1 − φ̂(δξ)|2 |û(ξ)|2 dξ
2 2 2 2 2

ˆ
2
≤ Cδ |ξ|2 |û(ξ)|2 dξ = Cδ 2 k∇uk2L2 = Cδ 2 kωk2L2 . (7.84)

The integral of W is
ˆ ˆ ˆ
∂uk ∂vδ,m
W (t, x)dx = − ωi εijk (∇u)km (∇vδ )mj dx = − εijk ωi dx (7.85)
∂xm ∂xj
ˆ ˆ
∂ωi ∂uk ∂ 2 uk
= εijk vδ,m dx + εijk vδ,m ωi dx.
∂xj ∂xm ∂xj ∂xm

The last integral above can be written as

ˆ ˆ ˆ
∂ 2 uk ∂  ∂uk  ∂ωi
εijk vδ,m ωi dx = ωi vδ,m εijk dx = ωi vδ,m dx = 0, (7.86)
∂xj ∂xm ∂xm ∂xj ∂xm

since vδ is divergence-free. Therefore, we have a bound for W :

ˆ ˆ ˆ
ν 2 C
W (t, x)dx ≤ |∇ω(t, x)| dx + |vδ (t, x)|2 |∇u(t, x)|2 dx
16 ν
ˆ
ν C
≤ |∇ω(t, x)|2 dx + kvδ k2L4 k∇uk2L4 . (7.87)
16 ν
The Gagliardo-Nirenberg inequality implies that
3/2 1/2
kvδ k2L4 ≤ Ck∇vδ kL2 kvδ kL2 . (7.88)

For the gradient term above we can simply bound

k∇vδ k2L2 ≤ Ck∇uk2L2 + Ck∇uδ k2L2 ≤ Ck∇uk2L2 ≤ Ckωk2L2 , (7.89)

and we may use the estimate (7.84) for |vδ kL2 . Therefore, we have

kvδ k2L4 ≤ Cδ 1/2 kωk2L2 . (7.90)

We may also use the same Gagliardo-Nirenberg inequality for k∇ukL4 , leading to
3/2 1/2
k∇uk2L4 ≤ Ck∇ωkL2 kωkL2 . (7.91)

Altogether, this gives

1 Cδ 1/2 3/2 1/2 Cδ 1/2 5/2 3/2

kvδ k2L4 k∇uk2L4 ≤ kωk2L2 k∇ωkL2 kωkL2 = kωkL2 k∇ωkL2
ν ν ν
2
ν Cδ
≤ k∇ωk2L2 + 7 kωk10 L2 , (7.92)
16 ν

87
thus
ˆ ˆ
ν Cδ 2
W (t, x)dx ≤ |∇ω(t, x)|2 dx + kωk10
L2 . (7.93)
8 ν7

Finally, we estimate the most dangerous term Z(t, x),

Z = (S (2) ω (2) · ω (2) ),

and this will be the only estimate that will involve the assumption that the direction ξ(t, x)
of the vorticity is Lipschitz:
⊥ |y|
Pξ(t,x) (ξ(t, x + y)) ≤ , (7.94)
ρ
We write

Z(t, x) = (S (2) ω (2) · ω (2) ) = |ω (2) (t, x)|2 (S (2) (t, x)ξ (2) (t, x) · ξ (2) (t, x)) = |ω(t, x)|2 α(2) (t, x),
(7.95)
with ˆ
(2) 3 dy
α (t, x) = P.V. D(ŷ, ξ(x + y), ξ(x))|ω (2) (x + y)| 3 , (7.96)
4π |y|
where
D(e1 , e2 , e3 ) = (e1 · e3 )Det(e1 , e2 , e3 ).
Assumption (7.94) means that

|y|
|D(ŷ, ξ(x + y), ξ(x))| ≤ , (7.97)
ρ
so that
ˆ ˆ
3 dy 3 dy
|Z(t, x)| ≤ |ω (2) (t, x)|2 (2)
|ω (t, x + y)| 2 ≤ |ω(t, x)|2 |ω(t, x + y)| . (7.98)
4πρ |y| 4πρ |y|2

Therefore, we have
ˆ ˆ 1/2
C
|Z(t, x)|dx ≤ kωk2L4 |I(t, x)| dx 2
, (7.99)
ρ
with ˆ
dy
I(t, x) = |ω(t, x + y)| .
|y|2
In order to compute the L2 -norm of I, we proceed as in the proof of Nash inequality. Let us
compute the Fourier transform of the function ψ(y) = 1/|y|2 :
ˆ ˆ ∞ ˆ π/2 ˆ 2π
e2πiξ·y dy
ψ̂(ξ) = 2
= dr dθ cos θ dφe2πi|ξ|r sin θ
|y| 0 −π/2 0
ˆ ∞ ˆ 1 ˆ ∞
2 sin rdr
= 2π dr due2πi|ξ|ru = .
0 −1 |ξ| 0 r

88
Hence, the L2 -norm of I(t, x) can be bounded as (for any R > 0)
ˆ ˆ
2 ˆ 2 |ω(ξ)|2
kI(t)kL2 = |I(t, ξ)| dξ ≤ C dξ
|ξ|2
ˆ ˆ
|ω(ξ)|2 dξ |ω(ξ)|2 dξ
≤C + C = AR + BR .
|ξ|≤R |ξ|2 |ξ|≥R |ξ|2

Since
|ω̂(ξ)| ≤ kωkL1 ,
the first term can be bounded as,
ˆ R
|AR | ≤ C kωk2L1 dξ ≤ CRkωk2L1 .
0

The second term can be simply bounded by

ˆ
C C
|BR | ≤ 2 |ω(ξ)|2 dξ = 2 kωk2L2 .
R |ξ|≥R R

It follows that for any R > 0 we have

C
kIk2L2 ≤ CRkωk2L1 + kωk2L2 .
R2
Choosing
 kωk2 1/3
L2
R= ,
kωk2L1
we deduce that
4/3 2/3
kIk2L2 ≤ CkωkL1 kωkL2 . (7.100)
Returning to (7.99), we see that
ˆ
C 2/3 1/3
|Z(t, x)|dx ≤ kωk2L4 kωkL1 kωkL2 . (7.101)
ρ

The L4 -norm of ω is estimated using the same Gagliardo-Nirenberg inequality:

3/2 1/2
kωk2L4 ≤ Ck∇ωkL2 kωkL2 , (7.102)

so that
ˆ
3/2 5/6 2/3 ν C 20/6 8/3
|Z(t, x)|dx ≤ Ck∇ωkL2 kωkL2 kωkL1 ≤ k∇ωk2L2 + 3 4 kωkL2 kωkL1 . (7.103)
15 ν ρ

Recalling also the a priori bound (7.32) in Lemma 7.2:

ˆ
|ω(t, x)|dx ≤ Q, (7.104)
R3

89
we see that Z is bounded as
ˆ
ν CQ8/3 10/3
|Z(t, x)|dx ≤ k∇ωk2L2 + 3 4 kωkL2 . (7.105)
15 ν ρ
Recollecting the starting point of our analysis (7.67)
ˆ ˆ ˆ ˆ
1d 2 2
|ω| dx + ν |∇ω| dx = (Sω · ω)dx − ω · ((∇u) (∇vδ ))dx, (7.106)
2 dt
and summarizing the bounds (7.76), (7.83), (7.93), (7.105) that we have obtained for the
terms X, Y , W and Z, respectively, in the right side of the above identity, we get
ˆ ˆ
1d ν C
|ω| dx + ν |∇ω|2 dx ≤ CΩkωk2L2 + k∇ωk2L2 + 3 (ΩQ)2 kωk2L2
2
2 dt 8 ν
2 8/3
ν Cδ ν CQ 10/3
+ k∇ωk2 + 7 kωk10 L2 + k∇ωk2L2 + 3 4 kωkL2 . (7.107)
8 ν 15 ν ρ
Thus, the enstrophy ˆ
E(t) = |ω(t, x)|2 dx,

satisfies a differential inequality

dE
≤ C1 (1 + E 2/3 )E + C1 δ 2 E 5 , (7.108)
dt
with a constant C1 that depends on ν, ρ, Ω and Q. This is a nonlinear inequality and at the
first glance it may seem useless as the solution of an ODE
ż = C1 (1 + z 2/3 )z + C1 δ 2 z 5 , z(0) = z0 > 0, (7.109)
blows up in a finite time. Here, however, we are only concerned with the solution being finite
until time t = T , and, in addition, we have an extra piece of information: the function
k(t) = C1 (1 + E 2/3 )
has a bounded integral:
ˆ T ˆ T ˆ T 2/3
4/3 1/3
k(t)dt ≤ CT + kω(t)kL2 dt ≤ CT + CT kω(t)k2L2 dt ≤ C(1 + T ) = D.
0 0 0
(7.110)
Crucially, the constant D does not depend on δ. Therefore, the solution of (7.109) with δ = 0
does remain finite until the time T , and it is reasonable to expect that so does the solution
with δ > 0 but small. To formalize this observation, let
nˆ t o
Ē(t) = 2E(0) exp k(s)ds .
0

Then E(0) ≤ Ē(0), and we may define τ as the first time such that E(τ ) = Ē(τ ). Until that
time, the function E(t) satisfies
dE
≤ k(t)E + C1 δ 2 Ē 5 , 0 ≤ t ≤ τ. (7.111)
dt

90
Therefore, as long as E(t) ≤ Ē(t), we have a bound for E(t):
nˆ t o ˆ t nˆ t o
E(t) ≤ E(0) exp k(s)ds + C1 δ 2 5
Ē (s) exp k(s0 )ds0 ds.
0 0 s

Thus, if δ is sufficiently small, we have E(t) ≤ Ē(t) for all 0 ≤ t ≤ T . We conclude that there
exists δ0 > 0 so that for all 0 < δ < δ0 the enstrophy is bounded:
ˆ
sup |ω(t, x)|2 dx < +∞. (7.112)
0≤t≤T

The last step is to observe that (7.107) together with (7.112) implies that
ˆ T ˆ
ν |∇ω|2 dx < +∞. (7.113)
0

This completes the proof of Lemma 7.3, and thus that of Theorem 7.1. 2

8 The Caffarelli-Kohn-Nirenberg theorem

In this section, we will describe the results of Caffarelli, Kohn and Nirenberg on the Hausdorff
dimension of the set where the solution of the three-dimensional Navier-Stokes equations

ut + u · ∇u + ∇p = ∆u + f, (8.1)
∇ · u = 0, (8.2)

can possibly be singular. We consider this problem in a smooth bounded domain Ω ⊂ R3 ,

with the no-slip boundary condition

u(t, x) = 0 on ∂Ω. (8.3)

The force f (t, x) is assumed to satisfy the incompressibility condition ∇·f = 0 – this condition
is not really necessary, as otherwise we would write f = ∇Φ + g, with ∇ · g = 0, and absorb Φ
into the pressure term.

Weak solutions
Let us recall the notion of a Leray weak solution of the Navier-Stokes equations: u is a weak
solution if, first, it is a solution in the sense of distributions, that is, for any smooth compactly
supported vector-valued function ψ(t, x) we have
ˆ ˆ tˆ ˆ tˆ
∂ψj
[u(t, x) · ψ(t, x) − u0 (x) · ψ(0, x)]dx − (u · ψs )dxds − uk uj dxds
∂xk
Ω
ˆ tˆ ˆ tˆ 0 Ω
ˆ tˆ 0 Ω

− p(∇ · ψ)dxds = (u · ∆ψ)dxds + (f · ψ)df xds. (8.4)

0 Ω 0 Ω 0 Ω

91
The second condition is that u satisfies the energy inequality. Note that if u is a smooth
solution of the Navier-Stokes equations, then for any smooth test function φ we have
ˆ ˆ tˆ ˆ
1 2 2 1
|u(t, x)| φ(t, x)dx + |∇u(s, x)| φ(s, x)dxds = |u0 (x)|2 φ(0, x)dx (8.5)
2 Ω 2 Ω
ˆ tˆ 0 Ω
1
+ |u(s, x)|2 (φs (s, x) + ∆φ(s, x))dxds
2 0 Ω
ˆ tˆ  ˆ tˆ
|u(s, x)|2 
+ + p(s, x) u · ∇φ(s, x)dxds + (f · u)φ(s, x)dxds.
0 Ω 2 0 Ω

Taking, formally, φ ≡ 1, the second condition for u to be a Leray weak solution is that it
satisfies the energy inequality:
ˆ ˆ tˆ ˆ ˆ tˆ
1 2 2 1 2
|u(t, x)| dx + |∇u(s, x)| dxds ≤ |u0 (x)| dx + (f · u)dxds. (8.6)
2 Ω 0 Ω 2 Ω 0 Ω

Suitable weak solutions

Caffarelli, Kohn and Nuremberg consider a slightly stronger class of solutions, which they
call suitable weak solutions, defined on an open (time-space) set D ∈ R × R3 . We will,
obviously, require that u is a weak solution of the Navier-Stokes equations in the sense of
distributions: (8.4) holds for any function φ supported in D. We will assume that f ∈ Lq (D)
with some q > 5/2 – this assumption is not very important, as the main result is interesting
even for f ∈ C ∞ (D). We will also assume that the pressure satisfies

p ∈ L5/4 (D), (8.7)

and that there exist some constants E0 and E1 so that or any fixed time t we have
ˆ
|u(t, x)|2 dx ≤ E0 , (8.8)
Dt

where Dt = D ∩ (R3 × {t}), and

ˆ
|∇u(s, x)|2 dx ≤ E1 . (8.9)
D

In addition, we require that the generalized (or, localized) energy inequality holds: for any
function φ ≥ 0 which is smooth and compactly supported in D, we have
ˆ ˆ
2 1
|∇u(s, x)| φ(s, x)dxds ≤ |u(s, x)|2 (φs (s, x) + ∆φ(s, x))dxds (8.10)
2
D
ˆ  D
ˆ
|u(s, x)|2 
+ + p(s, x) u · ∇φ(s, x)dxds + (f · u)φ(s, x)dxds.
D 2 D

At the moment, it is not clear that a suitable weak solution exists – we will prove it below.

92
The parabolic Hausdorff measure
In order to formulate the main results, we need to define an analog of the Hausdorff mea-
sure H1 but suitable for the parabolic problems. For any set X ⊂ R × R3 , δ > 0 and k ≥ 0
we define ∞
nX [ o
k k
Pδ (X) = inf ri : X ⊂ Qri , ri < δ . (8.11)
i=1 i
Here, Qr is a parabolic cylinder: it has the form
Qr = [t − r2 , t] × Br (x),
where Br (x) is a ball of radius r centered at the point x. Then we set
P k (X) = lim Pδk (X). (8.12)
δ↓0

The standard Hausdorff measure is defined in the same way but with Qr replaced by an
arbitrary closed subset of R × R3 of diameter at most ri , thus we have
H 1 ≤ Ck P k .

The main results

We may now describe the main results of the Caffarelli-Kohn-Nirenberg paper. We say that
a point (t, x) is singular if u is not in L∞
loc in any neighborhood of (t, x). Otherwise, we say
that (t, x) is a singular point. We will denote by S the set of all singular points of u(t, x).
Their first result shows that the singularity set has zero Hausdorff measure H1 .
Theorem 8.1 Assume that either Ω = R3 or Ω ⊂ R3 is a smooth bounded domain, and let
D = (0, T ) × Ω. Suppose that for some q > 5/2 we have
f ∈ L2 (D) ∩ Lqloc (D) ∇ · f = 0
and
u0 ∈ L2 (Ω), ∇ · u0 = 0, u0 · ν|∂Ω = 0.
2/5
If Ω is bounded, we require, in addition, that u0 ∈ W5/4 (Ω). Then the initial boundary value
problem has a suitable weak solution in D whose singular set S satisfies P 1 (S) = 0.
√
Their second result concerns absence of singularities outside of a ball of radius 1/ t.
Theorem 8.2 Consider the Navier-Stokes equations in R3 with f = 0 and assume that the
initial data satisfies ∇ · u0 = 0, and
ˆ
1
G= |u0 (x)|2 |x|dx < +∞. (8.13)
2 R3
Then there exists
√ a weak solution of the initial value problem which is regular in the re-
gion {|x| ≥ K1 / t}, with the constant K1 which depends only on G and E, where
ˆ
E= |u0 (x)|2 |x|dx < +∞.
R3

93
Assumption (8.13) means that u is small at infinity, and this smallness, so to speak, invades
the whole space as t grows. If we assume that u is “small near the origin”, in the sense, that
ˆ
|u0 |2
L= dx = L < +∞, (8.14)
R3 |x|

then we have the following result.

Theorem 8.3 Consider the Navier-Stokes equations in R3 with f = 0 and assume that the
initial data satisfies ∇ · u0 = 0, and (8.14) holds. p
There exists a universal constant L0 so that
if L < L0 , then u is regular in the region {|x| ≤ (L0 − L)t}.

The first key estimate: localizing “small data regularity”

We will denote the cylinders labeled by the top as

Qr (t, x) = {(s, y) : |y − x| < r, t − r2 < s < t},

and those labeled by a point slightly below the top as

7 1
Q∗r (t, x) = {(s, y) : |y − x| < r, t − r2 < s < t + r2 }.
8 8
It is well known that if the initial condition u0 and the force f are small in an appropriate
norm, then the solution of the Navier-Stokes equations remains regular for a short time. The
main issue in proving the partial regularity theorems is to localize this result. The first step
in this direction is an estimate showing that if u, p and f are sufficiently small on the unit
cylinder Q1 = Q1 (0, 0), then u is regular in the smaller cylinder Q1/2 = Q1/2 (0, 0) – this is a
very common theme in the parabolic regularity theory.
Proposition 8.4 There exist absolute constants C1 > 0 and ε1 > 0 and a constant ε2 (q) > 0,
which depends only on q with the following property. Suppose that (u, p) is a suitable weak
solution of the Navier-Stokes system on Q1 with f ∈ Lq , with q > 5/2. Assume also that
ˆ ˆ 0 ˆ 5/4
3
(|u| + |u||p|)dxdt + |p|dx dt ≤ ε1 , (8.15)
Q1 −1 |x|<1

and ˆ
|f |q dxdt ≤ ε2 . (8.16)
Q1

Then we have |u(t, x)| ≤ C1 for Lebesgue-almost every (t, x) ∈ Q1/2 . In particular, u is regular
in Q1/2 .
In order to see how we may scale this result to a parabolic cylinder of length r, let us investigate
the dimension of various terms in the Navier-Stokes equations

ut + u · ∇u + ∇p = ∆u + f. (8.17)

Let us assign dimension L to the spatial variable x. As all individual terms in (8.17) should
have the same dimension, looking at the terms ut and ∆u we conclude that time should have

94
dimension L2 . Comparing the terms ut and u·∇u we see that u should have the dimension L−1 .
Then, f should have the same dimension as ut , which is L−3 . Finally, the dimension of the
pressure term should be L−2 . Summarizing, we have

[x] = L, [t] = L2 , [u] = L−1 , [f ] = L−3 , [p] = L−2 . (8.18)

Let us look at the dimension of each term in the estimate (8.15): the term involving |u|3 has
the dimension
[x]3 [t][u]3 = L2 ,
the term involving |u||p| has the same dimension:

[x]3 [t][u][p] = L2 ,

while the last term in the left side has the dimension

[t][x]15/4 [p]5/4 = L23/4 L−10/4 = L13/4 .

We also should note that the dimension of the Lq -norm of f (to the power q) is

[x]3 [t][f ]q = L5−3q .

Accordingly, for a parabolic cylinder Qr (t, x) we set

ˆ ˆ t ˆ
1 3 1 5/4
M (r) = 2 (|u| + |u||p|)dxdt + 13/4 |p|dx dt, (8.19)
r Qr r t−r2 |y−x|<r

and ˆ
3q−5
Fq (r) = r |f |q dyds. (8.20)
Qr

Therefore, Proposition 8.4 has the following corollary.

Corollary 8.5 Suppose hat (u, p) is a suitable weak solution of the Navier-Stokes system on
a cylinder Qr with f ∈ Lq , with q > 5/2. Assume also that

M (r) ≤ ε1 , (8.21)

and
Fq (r) ≤ ε2 . (8.22)
Then we have |u(t, x)| ≤ C1 /r for Lebesgue-almost every (t, x) ∈ Qr/2 . In particular, u is
regular in Qr/2 .

The second key estimate: the blow-up rate

One can deduce from Corollary 8.5 a heuristic estimate on the possible blow-up rate of the
solution. Assume that (t0 , x0 ) is a singular point. Then, (8.21) has to fail for all Qr (t, x) such
that (t0 , x0 ) ∈ Qr/2 (t, x). Therefore, we must have

M (r) = M (r; t, x) > ε1

95
for a family of parabolic cylinders shrinking to the point (t0 , x0 ). Let us assume that

u(t, x) ∼ r−m ,

near x0 , with
r = (|x − x0 |2 + |t − t0 |)1/2 .
Then we have
1 1 2 3
M (r) ∼ r r = r3−3m .
r2 r3m
hence, a natural guess is m = 1, which translates into
C
|∇u| ≥ , as (t, x) → (t0 , x0 ). (8.23)
r2
The next key estimate verifies that this is qualitatively correct.

Proposition 8.6 There is an absolute constant ε3 > 0 with the following property. If u is a
suitable weak solution of the Navier-Stokes equations near (t, x), and if
ˆ
1
lim sup |∇u|2 dyds ≤ ε3 , (8.24)
r↓0 r Q∗r (t,x)

then (t, x) is a regular point.

Let us explain how Theorem 8.1 would follow. Take any (t, x) in the singular set, then,
by Proposition 8.6 we have
ˆ
1
lim sup |∇u|2 dyds > ε3 . (8.25)
r↓0 r Q∗r (t,x)

Take a neighborhood V of the singular set S and δ > 0. For each (t, x) ∈ S we may choose a
parabolic cylinder Q∗r (t, x) with r < δ and such that
ˆ
1
|∇u|2 dyds > ε3 , (8.26)
r Q∗r (t,x)

and Q∗r (t, x) ⊂ V . We will make use of the following covering lemma.

Lemma 8.7 Let J be a collection of parabolic cylinders Q∗r (t, x) contained in a bounded
set V . Then there exists an at most countable sub-collection J 0 = {Q∗i = Q∗r1 (ti , xi )} of
non-overlapping cylinders such that for any Q∗ ∈ J there exists Q∗i so that

Q∗ ⊂ Q∗5ri (ti , xi ).

The proof is very similar to that of the classic Vitali lemma and we leave it to the reader as
an exercise. Using this lemma, we obtain a disjoint collection of cylinders Q∗ri (ti , xi ) such that
[
S⊂ Q∗5ri (ti , xi ),
i

96
and ˆ ˆ
X 1 2 1
ri ≤ |∇u| dxdt ≤ |∇u|2 dxdt.
i
ε3 Q∗ri ε3 V

We deduce that ˆ
1 1
P (S) ≤ |∇u|2 dxdt. (8.27)
ε3 V

In particular, we deduce that the (three-dimensional) Lebesgue measure of S is zero. Then,

as V is an arbitrary neighborhood of S, and the function |∇u|2 is integrable, we can make the
right side of (8.27) arbitrarily small. It follows that P 1 (S) = 0, proving Theorem 8.1. Thus,
the crux of the matter is the proof of Propositions 8.4 and 8.6.

Serrin’s interior regularity result

Before we proceed with the further discussion of the proofs of the theorems of Caffarelli, Kohn
and Nirenberg, let us explain why we say a solution is regular if it is just bounded, and do
not require further differentiability. The reason is a result of Serrin on the interior regularity
of the weak solutions of the Navier-Stokes equations

ut + u · ∇u + ∇p = ∆u + f, (8.28)
∇ · u = 0.

We will assume for simplicity that f = 0 – the reader should consider the generalization to
the case f 6= 0 as an exercise, or consult Serrin’s original paper. Let us borrow the following
very simple observation from Serrin’s paper: if ψ(x) is a harmonic function, then any function
of the form
u(t, x) = a(t)∇ψ(x)
is a weak solution of the Navier-Stokes equations, as long as the function a(t) is integrable.
Therefore, boundedness of u(t, x) can not, in general, imply any information on the time
derivatives of u. On the other hand, this example does not rule out the hope that relatively
weak assumptions on u would guarantee its spatial regularity.
Here is one version of Serrin’s result, which says that bounded solutions of the force-less
Navier-Stokes equations are essentially as good as the solutions of the heat equation.

Theorem 8.8 Let u be a Leray weak solution of the Navier-Stokes equations in an open
region R = (t1 , t2 ) × Ω of space-time, with f = 0, and such that
ˆ t2 ˆ ˆ
2
|ω(t, x)| dxdt < +∞, sup |u(t, x)|2 dx < +∞, (8.29)
t1 Ω t∈[t1 ,t2 ] Ω

where ω = ∇ × u is the vorticity. Assume, in addition, that u ∈ L∞ (R). Then, u is of the C ∞

class in the space variables on every compact subset of R.
0
The full statement of the Serrin theorem says that if u ∈ Ls,s (R), with
 ˆ t2 0
1/s0
kukLs,s0 = kuksLs (Ω) dt ,
t1

97
with (in three dimensions)
3 2
+ 0 < 1, (8.30)
s s
then u is C ∞ in the spatial variables. If, in addition, we know that ut ∈ L2,p with p ≥ 1, then
the spatial derivatives of u are absolutely continuous in time. We will not need these results
for our purposes, so we will leave them out for now. Let us make one comment, however: if
we take s0 = ∞, then condition (8.30) is satisfied, as long as s > 3. That is, if we would have
known a priori that ˆ
|u(t, x)|3 dx ≤ const,
R3
then we could conclude that u is a smooth solution. Of course, we have this information only
for the L2 -norm of the Leray weak solutions, and not for the L3 -norm.
For the proof of Theorem 8.8, let us recall the vorticity equation in three dimensions:

ωt + u · ∇ω − ∆ω = ω · ∇u. (8.31)

Written in the components, this equation is

∂ωk ∂uk ∂ωk
− ∆ωk = ωj − uj , (8.32)
∂t ∂xj ∂xj
or
∂ωk ∂
− ∆ωk = (ωj uk − uj ωk ). (8.33)
∂t ∂xj
Let Ω̄1 be a compact subset of Ω, and t1 < s1 < s2 < t2 , so that S = (s1 , s2 ) × Ω1 is a proper
subset of R, and define, for s1 ≤ t ≤ s2 :
ˆ tˆ
∂
ω̃k (t, x) = G(t − s, x − y)[ωj (s, y)uk (s, y) − uj (s, y)ωk (s, y)]dyds
∂xj s1 Ω1
ˆ tˆ
∂G(t − s, x − y)
= [ωj (s, y)uk (s, y) − uj (s, y)ωk (s, y)]dyds.
s1 Ω1 ∂xj

Here, G(t, x) is the standard heat kernel. The functions

ˆ tˆ
mkj (t, x) = G(t − s, x − y)[ωj (s, y)uk (s, y) − uj (s, y)ωk (s, y)]dyds
s1 Ω1

satisfy
∂mkj
− ∆mkj = (ωj uk − uj ωk )χ[s1 ,s2 ] (t)χΩ̄1 (x). (8.34)
∂t
Thus, for (t, x) ∈ S, the function ω̃ is the solution of
∂ ω̃k ∂
− ∆ω̃k = (ωj uk − uj ωk ). (8.35)
∂t ∂xj

It follows that the difference

B(t, x) = ω(t, x) − ω̃(t, x)

98
satisfies the standard heat equation

Bt − ∆B = 0,

on the set S.
We will now show that ω ∈ L∞ (S), that is, if u is uniformly bounded on R, then the
vorticity is uniformly bounded on any compact subset of R.

Exercise 8.9 Use the convolution with the heat kernel to show that if φ(t, x) satisfies

∂g
φt − ∆φ = ,
∂xj

in the whole space Rn , then

kφkLr ≤ CkgkLq ,
as long as 1 1
(n + 2) − < 1.
q r
The norms are take in space-time.

As u is a Leray weak solution, we know that ω ∈ L2 (R). As u ∈ L∞ (R), it follows that

the functions
gjk (s, y) = ωj (s, y)uk (s, y) − uj (s, y)ωk (s, y)
are also in L2 (R). The result of the above exercise says that then ω̃ ∈ Lr with
1 1 1 1
= − =
r 2 3 6
But then g ∈ L6 , as well, and, as 1/6 < 1/3, it follows that ω̃ ∈ L∞ (R). We also know
that B ∈ L∞ (S) by the regularity estimates for the heat equation, as B ∈ L2 (R) – it is the
difference of two functions in L2 (R). Moreover, we know that B is Hölder continuous.
Now that we know that ω ∈ L∞ (R), we recall that the velocity and the vorticity are
related by the stream vector ψ, defined as the solution of

−∆ψ = ω, ∇ · ψ = 0,

and
u = −∇ × ψ.
Therefore, if ω ∈ L∞ (R), then ψ is C 1,α in the spatial variable, hence u is Hölder in x, and,
in particular, in L∞ . Then the functions mkj are C 1,α in x, thus ω is Hölder in x. Then, the
functions gkj are Hölder in x, so ωx is Hölder in x, continuing this argument we deduce that
both ω and u are C ∞ .

99
Existence of suitable weak solutions
We now prove the existence of suitable weak solutions, in the sense of Caffarelli, Kohn and
Nirenberg. We will restrict ourselves to the whole space: Ω = R3 . Let us first define the
appropriate function spaces. As usual, we will denote by V the space of smooth divergence-
free vector fields u, by H the closure of V in L2 (R3 ), by V the closure of V in H 1 (R3 ), and
by V 0 the dual space of V . The Sobolev spaces Wql (R3 ) with q ≥ 1 and 0 < l < 1 consists of
functions with l derivatives in Lq , and with the norm
kukWql = kukLq + k(−∆)l/2 ukLq .
We will make the standard assumptions:
Ω = R3 , u0 ∈ H, f ∈ L2 (0, T ; H −1 (R3 )). (8.36)
Theorem 8.10 Assume that Ω = R3 , u0 and f satisfy (8.36). Then there exists a suitable
weak solution
u ∈ L2 (0, T ; V ) ∩ L∞ (0, T ; H),
of the Navier-Stokes equations with the force f and the initial condition u0 , in the sense
that u(t) → u0 weakly in H as t → 0. The pressure satisfies p ∈ L5/3 ((0, T ) × R3 ). In
addition, if φ ∈ C ∞ ([0, T ] × R3 ), φ ≥ 0 and is compactly supported, then
ˆ ˆ tˆ ˆ
1 2 2 1
|u(t, x)| φ(t, x)dx + |∇u(s, x)| φ(s, x)dxds ≤ |u0 (x)|2 φ(0, x)dx
2 R3 3 2 3
ˆ ˆ 0 R R
1 t
+ |u(s, x)|2 (φs (s, x) + ∆φ(s, x))dxds (8.37)
2 0 R3
ˆ tˆ  ˆ tˆ
|u(s, x)|2 
+ + p(s, x) u · ∇φ(s, x)dxds + (f · u)φ(s, x)dxds.
0 R3 2 0 R3

The proof is done via a ”retarded mollification”. The (standard) idea is to take Ψδ (u) to be
a mollifier of u such that Ψδ (u) is divergence-free and depends only on the values of u(s, x)
with s ≤ t − δ. The mollified system
ut + Ψδ (u) · ∇u + ∇p = ∆u + f (8.38)
is then linear on each time interval of the firm (mδ, (m + 1)δ). We will get uniform in δ a
priori bounds on u, and then pass to the limit δ → 0.
Let us recall some basic facts about the linear Stokes equation, whose proof is very similar
to what we have done on the torus previously.
ut + ∇p = ∆u + f, ∇ · u = 0. (8.39)
Lemma 8.11 Suppose that f ∈ L2 (0, T ; V 0 ), u ∈ L2 (0, T ; V ), p is a distribution and (8.39)
holds. Then ut ∈ L2 (0, T ; V 0 ),
ˆ ˆ
d 2
|u| dx = 2 (ut · u)dx,
dt Ω Ω

in the sense of distributions on (0, T ), and u ∈ C([0, T ], H), possibly after a modification on
a set of measure zero.

100
Lemma 8.12 Suppose that f ∈ L2 (0, T ; V 0 ), u0 ∈ H, and w ∈ C ∞ ([0, T ]; Ω) are prescribed,
and ∇ · w = 0. Then there exists a unique function u ∈ L2 (0, T ; V ) ∩ C([0, T ]; H), and a
distribution p so that
ut + w · ∇u + ∇p = ∆u + f, ∇ · u = 0, (8.40)
in the sense of distributions, and u(0) = u0 .

Some pressure bounds and interpolation on the velocity

Note that if u solves (8.40) in the whole space, then the pressure satisfies the Poisson equation
3
X
∆p = − ∂ij2 (wi uj ). (8.41)
i,j=1

The singular integral operator corresponding to the Fourier multiplier

ξi ξj
|ξ|2
is bounded Lp (Rn ) → Lp (Rn ) for all 1 < p < ∞, thus, in particular, we have the bound
ˆ Tˆ ˆ Tˆ
5/3
|p| dxds ≤ C |w|5/3 |u|5/3 dxds (8.42)
0 3 0 3
ˆ T ˆ 1/2  ˆ T ˆ
R R
1/2
10/3
≤C |w| dxds |u|10/3 dxds . (8.43)
0 R3 0 R3

We will now use a Gagliardo-Nirenberg inequality

ˆ ˆ a  ˆ q/2−a
q 2
|u| dx ≤ C |∇u| dx |u|2 dx , (8.44)
R3 R3 R3

with 2 ≤ q ≤ 6 and a = 3(q − 2)/4. Note that when q = 2, a = 0, this is a tautology, and
when q = 6, a = 3, this is the familiar Gagliardo-Nirenberg inequality
ˆ ˆ 3
6
|u| dx ≤ C |∇u|2 dx . (8.45)
R3 R3

Taking q = 10/3, and a = 1 gives

ˆ ˆ  ˆ 2/3
10/3 2
|u| dx ≤ C |∇u| dx |u|2 dx (8.46)
R3 R3 R3

Integrating in time and using the a priori assumptions (8.8) and (8.9) leads to
ˆ Tˆ
2/3
|u|10/3 dxdt ≤ CE1 (u)E0 (u). (8.47)
0 R3

Another useful estimate, obtained, once again, by taking q = 10/3 and a = 1, is

ˆ Tˆ ˆ T ˆ 5/8  ˆ T ˆ 3/8
5/4 2
|w · ∇u| dxdt ≤ |∇u| dxdt |w|10/3 dxdt (8.48)
0 R3 0 R3 0 R3
5/8 3/8 1/4
≤ CE1 (u) E1 (w) E0 (w) , (8.49)

101
which can be restated as

kw · ∇ukL5/4 ≤ CE1 (u)1/2 E1 (w)3/10 E0 (w)1/5 . (8.50)

We will also use the following bound, which follows from (8.45) with q = 5/2 and a = 3/8:
ˆ ˆ 3/8
5/2 7/8
|u| dx ≤ CE0 |∇u|2 dx . (8.51)
R3 R3

As a consequence, we have
ˆ T ˆ 2 ˆ T ˆ 3/4
5/2 7/4
|u| dx dt ≤ CE0 (u) |∇u|2 dx dt (8.52)
0 R3 0 R3
ˆ T ˆ 3/4
7/4 1/4 2 7/4
≤ CE0 (u) T |∇u| dxdt ≤ CT 1/4 E0 E1 (u)3/4 .
0 R3

This can be restated as

7/20
kukL5 (0,T ;L5/2 ) ≤ CT 1/20 E0 E1 (u)3/20 . (8.53)

These bounds allow us to take a solution (in the sense of distributions) u ∈ C([0, T ]; H) ∩
L2 (0, T ; V ) of the Stokes advection equation

ut + w · u − ∆u + ∇p = f, (8.54)

with w ∈ C ∞ , multiply by a test function φ and obtain

ˆ ˆ Tˆ ˆ
2 2
|u| (T, x)φ(T, x)dx + 2 |∇u(t, x)| φ(t, x)dxdt = |u0 (x)|2 φ(0, x)dx (8.55)
R3 0 R3 R3
ˆ Tˆ ˆ Tˆ ˆ Tˆ
2 2
+ |u| (φt + ∆φ)dxdt + (|u| w + 2pu) · ∇φdxdt + 2 (u · f )dxdt.
0 R3 0 R3 0 R3

Exercise 8.13 Justify the integration by parts above by mollifying (in time and space) each
term in the Stokes equation, multiplying by φ, integrating by parts and then removing the
mollification using the a priori bounds obtained above.

The retarded mollifier

We take a C ∞ function ψ(t, x) ≥ 0 such that
ˆ
ψ(t, x)dxdt = 1,

and
suppψ ⊂ {(t, x) : |x|2 < t, 1 < t < 2}.
We also extend u(t, x) by zero to t < 0, and set
ˆ
1 s y
Ψδ (u)(t, x) = 4 ψ( , )ũ(x − y, t − s)dyds. (8.56)
δ R4 δ δ

102
The mollified u is divergence-free:
∇ · Ψδ (u) = 0,
and it inherits the a priori bounds on u:
ˆ
sup |Ψδ (u)|2 (t, x)dx ≤ CE0 (u), (8.57)
0≤t≤t R3

and ˆ ˆ
T
|Ψδ (u)|2 (t, x)dxdt ≤ CE1 (u). (8.58)
0 R3

The approximants
We will use the approximants
∂uN
+ Ψδ (uN ) · ∇uN + ∇pN = ∆uN + f, (8.59)
∂t
∇ · uN = 0,
uN (0, x) = u0 (x),

with δ = T /N . We may apply inductively the existence result for the Stokes equation with a
prescribed advection, on the time intervals of the form (mδ, (m + 1)δ), 0 ≤ m ≤ N − 1. Then
we have
ˆ ˆ tˆ ˆ ˆ tˆ
2 2 2
|uN (t, x)| dx + 2 |∇uN (s, x)| dxds = |u0 (x)| dx + 2 (f · uN )dxds. (8.60)
R3 0 R3 R3 0 R3

In particular, we have
ˆ ˆ tˆ ˆ ˆ t
2 2 2
|uN (t, x)| dx + |∇uN (s, x)| dxds ≤ |u0 (x)| dx + kf k2V 0 ds. (8.61)
R3 0 R3 R3 0

We conclude that uN is uniformly bounded in L∞ (0, T ; V ) ∩ L∞ (0, T ; H), the usual Leray
bound. In addition, we know that pN is bounded in L5/3 ([0, T ] × R3 ). It follows that, after an
extraction of a sub-sequence, we have that pN → p∗ weakly in L5/3 ([0, T ] × R3 ), and uN → u∗ ,
weak-star in L∞ (0, T ; H), and weakly in L2 (0, T ; V ).
∂uN
Exercise 8.14 Show that if uN is bounded in L∞ (0, T ; V )∩L∞ (0, T ; H), and ∂t
is bounded
in L2 (0, T ; H −2 ), then uN has a convergent subsequence in L2 ([0, T ] × R3 ).

Exercise 8.15 Show that if uN → u∗ strongly in Lq and uN is bounded in Lr , 1 ≤ q < r,

then uN → u∗ strongly in Ls for all q, s < r.

We may use this with q = 2 and r = 10/3 to conclude that uN → u∗ strongly in Ls ([0, T ]×R3 )
for all 2 ≤ s < 10/3. Then one may easily check that (u∗ , p∗ ) is the sought suitable weak
solution of the Navier-Stokes equations.

103
The proof of Proposition 8.4
We now turn to the proof of the two main auxiliary results, and begin with Proposition 8.4.
We recall its statement:

Proposition 8.16 There exist two absolute constants C1 > 0 and ε1 > 0 and another con-
stant ε2 (q) > 0, which depends only on q with the following property. Suppose that (u, p) is
a suitable weak solution of the Navier-Stokes system on Q1 (0, 0) with f ∈ Lq , with q > 5/2.
Assume also that
ˆ ˆ 0 ˆ 5/4
3
(|u| + |u||p|)dxdt + |p|dx dt ≤ ε1 , (8.62)
Q1 −1 |x|<1

and ˆ
|f |q dxdt ≤ ε2 . (8.63)
Q1

Then we have |u(t, x)| ≤ C1 for Lebesgue-almost every (t, x) ∈ Q1/2 (0, 0). In particular, u is
regular in Q1/2 .

Outline of the proof

Let us take an arbitrary point (s, x0 ) ∈ Q1/2 (0, 0), where we want to show that |u(s, x0 )| ≤ C1 .
As Q1/2 (s, x0 ) ⊂ Q1 (0, 0), we have an integral estimate
ˆ ˆ s ˆ 5/4
3
(|u| + |u||p|)dxdt + |p|dx dt ≤ ε1 . (8.64)
Q1/2 (s,x0 ) s−1/4 |x−x0 |<1/2

We will consider a sequence of shrinking parabolic cylinders Qk = Qrk (s, x0 ), “centered” at

the point (s, x0 ) with rk = 2−k . Our goal will be to show that for all k ≥ 2 we have
ˆ
2/3
− |u(s, x)|2 dx ≤ C0 ε1 , (8.65)
|x−x0 |<rk
´
where −S f denotes the average of a function f over the set S. Then, if (s, x0 ) is a Lebesgue
point for u, it follows that
2/3
|u(s, x0 )|2 ≤ C0 ε1 , (8.66)
hence (8.66) holds for Lebesgue almost every point in Q1/2 (0, 0), which is exactly the claim
of Proposition 8.16.
In order to prove (8.65) we will show that for all k ≥ 2 we have a more general estimate
ˆ ˆ
2 1 2/3
sup − |u(t, x)| dx + 3 |∇u(t, x)|2 dxdt ≤ C0 ε1 . (8.67)
s−rk2 <t≤s |x−x0 |≤rk rk Qk

Note that (8.65) follows immediately from (8.67). Thus, the conclusion of Proposition 8.4
follows from (8.67).

104
 The induction base. We will prove (8.67) by induction, starting with k = 2. For k = 2,
we may use the localized energy inequality: for every smooth test function φ(t, x) ≥ 0, that
vanishes near |x| = 1 and t = −1, we have, for −1 < s < 0, with B1 = B1 (0, 0):
ˆ ˆ sˆ ˆ sˆ
2 2
|u(s, x)| φ(s, x)dx + 2 |∇u(t, x)| φ(t, x)dxdt ≤ |u(t, x)|2 (φt + ∆φ)dxdt
B1
ˆ sˆ −1 B1
ˆ sˆ −1 B1

+ (|u|2 + 2p)u · ∇φ(t, x)dtdx + 2 (f · u)φ(t, x)dxdt. (8.68)

−1 B1 −1 B1

Taking φ such that 0 ≤ φ ≤ 1, φ ≡ 1 on Q1/2 (0, 0) and φ is supported in Q1 (0, 0), we deduce
that
ˆ ˆ ˆ
2 2
|u(s, x)| dx+ |∇u(t, x)| dxdt ≤ C (|u|2 +|u|3 +|u||p|+|u||f |)dxdt. (8.69)
|x−x0 |≤1/4 Q2 Q1 (0,0)

Now, we may use Young’s inequality on the term |u||f |, together with the Lq -bound on f ,
with q > 5/2, the Hölder inequality, as well as our assumption (8.64), to conclude that the
2/3
left side of (8.69) is smaller than Cε1 , provided that ε1 and ε2 are both sufficiently small.
Thus, (8.67) holds for k = 2.
The induction step. The induction step in the proof of (8.67) will be split into two
sub-steps. First, we will show that if (8.67) holds for all 2 ≤ k ≤ n − 1, and n ≥ 3, then we
have ˆ 3/5 ˆ
1 3 rn 2/3
|u| dxdt + |u||p − p̄n |dxdt ≤ ε1 , (8.70)
|Qn | Qn |Qn | Qn
where ˆ
p̄n (t) = − p(t, x)dx. (8.71)
|x−x0 |<rn

Next, we will show that if (8.70) holds for all 3 ≤ k ≤ n, then (8.67) holds for k = n. That
is, we have the following two lemmas.

Lemma 8.17 Assume that ε1 and ε2 are sufficiently small, and n ≥ 3, and (8.67) holds for
all 2 ≤ k ≤ n − 1, then (8.70) holds.

Lemma 8.18 Assume that (8.70) holds for all 3 ≤ k ≤ n, and ε1 and ε2 are sufficiently
small, then (8.67) holds for k = n.

The proof of these lemmas is the heart of the argument.

The proof of Lemma 8.17

We set ˆ ˆ
1 2 1
A(r) = sup |u(t, x)| dx, G(r) = 2 |u|3 dxdt,
s−r2 <t<s r Br (x0 ) r Qr (s,x0 )

and ˆ
1
δ(r) = |∇u(t, x)|2 dxdt.
r Qr (s,x0 )

105
Recalling that the dimension of u is 1/L, and the dimension of t is L2 , while the dimension
of p is 1/L2 , we see that, A(r), G(r), and δ(r) are all dimensionless. The induction hypothesis
is
2/3
A(rk ) + δ(rk ) ≤ Cε1 rk2 , 2 ≤ k ≤ n − 1. (8.72)
In addition, we know that
G(r1 ) + K(r1 ) ≤ Cε1 , (8.73)
which is part of (8.64).
Bound on the first term in (8.70). The two terms in the left side of (8.70) will be
estimated separately. We will extensively use the Gagliardo-Nirenberg inequality in a ball
ˆ ˆ a  ˆ ˆ
q 2 2
q/2−a C  q/2
|u| dx ≤ C |∇u| dx |u| + 2a |u|2 dx , (8.74)
Br Br Br r Br

with 2 ≤ q ≤ 6, and a = 3(q − 2)/4 – this is the only choice of a which makes (8.74)
dimensionally correct. Taking q = 3 and a = 3/4 gives a bound on the L3 -norm that appears
in the left side of (8.70):
ˆ ˆ 3/4  ˆ ˆ
3 2 2
3/4 C  2
3/2
|u| dx ≤ C |∇u| dx |u| + 3/2 |u| dx . (8.75)
Br Br Br r Br

Integrating in time and using Hölder’s inequality leads to

ˆ ˆ s ˆ 3/4  ˆ ˆ s ˆ
3 2 2
3/4 C 3/2
|u| dxdt ≤ C |∇u| dx |u| dx dt + 3/2 |u|2 dx dt
Qr s−r2 r s−r2
ˆ 3/4  ˆ s  ˆ
Br Br
ˆ s ˆ Br

2 2
3 1/4 C 3/2
≤C |∇u| dxdt |u| dx dt + 3/2 |u|2 dx dt
Qr s−r2 Br r s−r2 Br
 3/4
≤ C rδ(r) r1/2 [rA(r)]3/4 + Cr1/2 [rA(r)]3/2 = Cr2 A(r)3/4 [δ(r)3/4 + A(r)3/4 ]. (8.76)

Dividing by |Qr | gives

ˆ ˆ
1 3 C C
|u| dxdt ≤ 5 |u|3 dxdt ≤ 3 A(rn−1 )3/4 [δ(rn−1 )3/4 + A(rn−1 )3/4 ]
|Qrn−1 | Qrn−1 rn−1 Qrn−1 rn−1
C
≤ 3
(A(rn−1 ) + δ(rn−1 ))3/2 ≤ Cε1 , (8.77)
rn−1
which, in turn, means that
ˆ ˆ
1 C0
3
|u| dxdt ≤ |u|3 dxdt ≤ C 00 ε1 . (8.78)
|Qrn | Qrn |Qrn−1 | Qrn−1

Hence, if ε1 is so small that

1
C 00 ε1/3 ≤ ,
2
then
ˆ
1 1 2/3
|u|3 dxdt ≤ ε1 . (8.79)
|Qrn | Qrn 2

106
This is the estimate we need on the first term in the left side of (8.70). Note that (8.78) can
be also restated as

G(rn ) ≤ Cε1 rn3 . (8.80)

Bound on the second term in (8.70). In order to get a bound on the second term in
the left side of (8.70), we need to show that, under the assumption
2/3
A(rk ) + δ(rk ) ≤ Cε1 rk2 , 2 ≤ k ≤ n − 1, (8.81)

we have ˆ
3/5 2/3
rn ε1
|u||p − p̄n |dxdt ≤ , (8.82)
|Qn | Qn 2
provided that ε1 is sufficiently small. The main issue is bounding the pressure. Recall that p
satisfies the Poisson equation (note that this is the first time in the proof of the current lemma
that we use the Navier-Stokes equations)

∂2
−∆p = (ui uj ). (8.83)
∂xi ∂xj

For any cut-off function φ we can write

ˆ ˆ
3 1 3 1
φ(x)p(t, x) = − ∆y (φp)dy = − (p∆φ + 2∇φ · ∇p + φ∆p)dy.
4π R3 |x − y| 4π R3 |x − y|

Using (8.83) and integrating by parts, we may write the above as

φp = p1 + p2 + p3 ,

where
ˆ
3 ∂2 h 1 i
p1 = φui uj dy,
4π R3 ∂yi ∂yj |x − y|
ˆ ˆ
3 xi − yi ∂φ 3 1 ∂ 2φ
p2 = u u
i j dy + ui uj dy,
2π R3 |x − y|3 ∂yj 4π R3 |x − y| ∂yi ∂yj
ˆ ˆ
3 1 3 xi − yi ∂φ
p3 = p∆φdy + p dy.
4π R3 |x − y| 2π R3 |x − y|3 ∂yj

We will take a function φ so that φ(y) ≡ 1 for |y − x0 | ≤ 3/16 and φ(y) = 0 if |y − x0 | ≥ 1/4.
Let us split p1 as
p1 = p11 + p12 ,
with
ˆ
3 ∂2 h 1 i
p11 = φui uj dy,
4π |y−x0 |<2rn ∂yi ∂yj |x − y|
ˆ
3 ∂2 h 1 i
p12 = φui uj dy.
4π |y−x0 |>2rn ∂yi ∂yj |x − y|

107
We can write (dropping the subscript n for the moment)
|p − p̄| ≤ |p11 − p̄11 | + |p12 − p̄12 | + |p3 − p̄3 | + |p4 − p̄4 |.
To estimate p11 , recall that the operators
 1 
Tij (ψ) = ∇2ik ?ψ
|x|
are Calderon-Zygmund operators, hence they are uniformly bounded in Lq , 1 < q < ∞. It
follows that (we denote r = rn and Br = Brn (x0 ))
ˆ 2/3
3
kp11 kL3/2 (Br ) ≤ C |u| dx ,
B2r

and ˆ ˆ
1 1  3/2
2/3
p̄11 ≤ |p|dx ≤ |p| dx ,
|Br | Br |Br |2/3 Br
hence ˆ ˆ
3/2
|p̄11 | dx ≤ |p|3/2 dx.
Br Br
We conclude that
ˆ ˆ 1/3  ˆ 2/3
3 3
|u||p11 − p̄11 |dx ≤ C |u| dx |u| dx . (8.84)
Br Br B2r

The terms |pi − p̄i | for p12 , p2 and p3 are estimated using the following bounds on the
gradients ∇pi for |x − x0 | < r (recall that φ ≡ 1 in the ball B3/16 (x0 ) so that ∇φ = 0 in that
ball):
ˆ ˆ
|u|2 |u|2
|∇p12 (x)| ≤ C 3
dy ≤ C 3
dy,
2r<|y−x0 |<1/4 |y − x| 2r<|y−x0 |<1/4 |y − x0 |
ˆ
|∇p2 (x)| ≤ C |u|2 dy,
B1/4 (x0 )
ˆ
|∇p3 (x)| ≤ C |p|dy.
B1/4 (x0 )

This leads to
ˆ ˆ 1/3
3 2/3 3
|u||p12 − p̄12 | ≤ Cr[ sup |∇p12 (x)|](r ) |u| dx
Br x∈Br Br
ˆ 1/3 ˆ |u|2
3 3
≤ Cr |u| dx dy, (8.85)
Br 2r<|y−x0 |<1/4 |y − x0 |3
and
ˆ 1/3 ˆ
3 2/3
|u||p2 − p̄2 | ≤ Cr[ sup |∇p2 (x)|](r ) |u|3 dx (8.86)
x∈Br
ˆ 1/3 ˆ ˆ 1/3  ˆ
Br Br
2/3
3 3 2 3 3 3
≤ Cr |u| dx |u| dy ≤ Cr |u| dx |u| dy .
Br B1/4 (x0 ) Br B1/4 (x0 )

108
For p3 , we write
ˆ ˆ  ˆ 
|u||p3 − p̄3 | ≤ Cr |u|dy |p| (8.87)
Br Br B1/4 (x0 )
ˆ 1/5  ˆ 1/5  ˆ 
3 3/5 2 3
≤ Cr(r ) |u| dy |u| dy |p|
Br Br B1/4 (x0 )
ˆ 1/5  ˆ 
≤ Cr3 A(r)1/5 |u|3 dy |p| .
Br B1/4 (x0 )

Integrating the above estimates over the time interval s − r2 ≤ t ≤ s, and collecting all the
terms we get ˆ
|u||p − p̄r |dxdt ≤ W1 + W2 + W3 + W4 . (8.88)
Qr

The term
ˆ 1/3  ˆ 2/3
3
W1 = C |u| dxdt |u|3 dxdt = Cr2 G(r)1/3 G(2r)2/3 (8.89)
Qr Q2r

comes from (8.84) and using Hölder’s inequality. Using (8.80), W1 can be bounded as

W1 ≤ Cε1 rn2 rn3 = Cε1 rn5 . (8.90)

The second term arises from (8.85) and also using Hölder’s inequality (note that 13/3 =
3 + 2(2/3)),
ˆ 1/3 ˆ
|u(t, y)|2
13/3 3
W2 = Cr |u| dxdt sup 3
dy. (8.91)
Qr s−r2 <t<s 2r<|y−x0 |<1/4 |y − x0 |

Note that for r = rn = 2−n , the last factor in (8.91) can be estimated with the help of the
induction hypothesis (8.81) as
ˆ n−1 ˆ
|u(t, y)|2 X |u(t, y)|2
dy ≤ dy
2rn <|y−x0 |<1/4 |y − x0 |3 k=3 2−k <|y−x0 |<2−(k−1) |y − x0 |3
n−1 ˆ n−1 n−1 2/3
X X 2/3
X Cε1
≤ 23k
|u(t, y)|2 dy ≤ rk−3 A(rk−1 ) ≤ Cε1 rk−1 ≤ .
k=3 2−k <|y−x0 |<2−(k−1) k=3 k=3
rn

Using this inequality, together with (8.80) in (8.91) gives

2/3
13/3 2 1/3 ε1 2/3
W2 ≤ Crn (rn G(rn )) ≤ Crn4 G(rn )1/3 ε1 ≤ Crn5 ε1 . (8.92)
rn
The third term
ˆ 1/3  ˆ 2/3
3 3
W3 = Cr |u| dxdt |u|3 dxdt (8.93)
Qr Q1/4

109
comes from (8.86) and, of course, using Hölder’s inequality once again, and can be bounded
with the help of (8.80) as

W3 ≤ Crn3 (rn2 G(rn )))1/3 G(1/4)2/3 ≤ Crn14/3 ε1 . (8.94)

Finally, the last term in (8.88) comes from (8.87):

ˆ 1/5  ˆ 0 ˆ 5/4 4/5
3 1/5 3
W4 = Cr A(r) |u| dxdt |p|dx dt . (8.95)
Qr −1/16 B1/4

It can be bounded as (assuming that ε1 ≤ 1):

4/5 2/3 4/5
W4 ≤ Crn3 A(rn )1/5 (rn2 G(rn ))1/5 ε1 ≤ Crn3 (rn2 ε1 )1/5 (rn5 ε1 )1/5 ε1 ≤ Crn22/5 ε1 . (8.96)

Altogether, we conclude that

ˆ
|u||p − p̄rn |dxdt ≤ Crn22/5 ε1 . (8.97)
Qn

We conclude that ˆ
3/5 2/3
rn ε1
|u||p − p̄rn |dxdt ≤ Cε1 ≤ , (8.98)
|Qn | Qn 2
provided that ε1 is small enough. This bounds the second term in (8.70) and finishes the
proof of Lemma 8.17.

Proof of Lemma 8.18

We now assume that
ˆ 3/5 ˆ
1 3 r 2/3
|u| dxdt + k |u||p − p̄n |dxdt ≤ ε1 , (8.99)
|Qk | Qk |Qk | Qk

for all 3 ≤ k ≤ n, and show that then

ˆ ˆ
2 1 2/3
sup − |u(t, x)| dx + 3 |∇u(t, x)|2 dxdt ≤ C0 ε1 . (8.100)
2 <t≤s |x−x |≤r
s−rn 0 n
rn Qn

We will shift the origin so that (s, x0 ) = (0, 0), to simplify the notation. The idea is to use
the generalized energy inequality
ˆ ˆ sˆ ˆ sˆ
2 2
|u(s, x)| φ(s, x)dx + 2 |∇u(t, x)| φ(t, x)dxdt ≤ |u(t, x)|2 (φt + ∆φ)dxdt
B1
ˆ sˆ −1 B1
ˆ sˆ −1 B1

+ (|u|2 + 2p)u · ∇φ(t, x)dtdx + 2 (f · u)φ(t, x)dxdt, (8.101)

−1 B1 −1 B1

with a suitable test function φn . We will set

φn (t, x) = χ(x)ψn (t, x),

110
with the backward heat kernel
1 n |x|2 o
ψn (t, x) = 2 exp − ,
(rn − t)3/2 4(rn2 − t)
and a smooth function χ(x) ≥ 0 so that χ(x) ≡ 1 on Q2 = Q1/4 (0, 0) and χ = 0 outside
of Q1/3 (0, 0). Then we have
∂φn
+ ∆φn = 0, on Q2 ,
∂t
and
∂φn
+ ∆φn ≤ C, everywhere,
∂t
and the following bounds hold:
1 C C
3
≤ φn ≤ 3 , |∇φn | ≤ , on Qn , n ≥ 2 (8.102)
Crn rn rn4
and
1 C C
3
≤ φn ≤ 3 , |∇φn | ≤ , on Qk−1 \ Qk , n ≥ 2. (8.103)
Crk rk rk4
We may now insert this φn into (8.101), and use the lower bound for φn on Qn to get
ˆ ˆ ˆ
1 2 1 2 ∂φn
sup 3
|u(t, x)| dx + 3 |∇u| dxdt ≤ C |u|2 | + ∆φn |dxdt
2 ≤t≤0 rn
−rn |x|<rn rn Qn Q1 ∂t
ˆ ˆ ˆ s
3
+C |u| |∇φn |dtdx + C p(u · ∇φn )dtdx + C |f ||u||φ|dxdt
Q1 Q1 Q1
= C(I1 + I2 + I3 + I4 ). (8.104)

To estimate I1 we simply use Hölder’s inequality:

ˆ ˆ 2/3
2 3 2/3
|I1 | ≤ C |u| dxdt ≤ C |u| dxdt ≤ Cε1 . (8.105)
Q1 Q1

The second term is estimated as

n ˆ n
X 1 3
X 1 2/3 5 2/3
|I2 | ≤ C 4
|u| dxdt ≤ C ε rk ≤ Cε1 .
4 1
(8.106)
r
k=1 k Qk
r
k=1 k

The last term in (8.104) is also easy:

n ˆ n ˆ 1/3  ˆ
X 1 X 1 3 3/2
2/3
|I4 | ≤ C |u||f |dxdt ≤ C |u| |f | (8.107)
r3
k=1 k Qk
r 3 Qk
k=1 k Qk
n n
X 1 2/3 5 1/3 10/3−5/q 1/q 2/9
X 2−5/q 1/q 2/9
≤C (ε
3 1
r k ) kf k q r
L (Q1 ) k ≤ Cε2 ε 1 rk ≤ Cε2 ε1 ,
r
k=1 k k=1

as q > 5/2. Therefore, if ε2 is sufficiently small, we have

2/3
|I4 | ≤ Cε1 . (8.108)

111
 Finally, we deal with I3 . Here, we will use the condition that u is a divergence-free flow.
Let us take smooth functions 0 ≤ χk ≤ 1 such that χk ≡ 1 on Q7rk /8 , and χk ≡ 0 outside
of Qrk , and
C
|∇χk | ≤ .
rk
Then, as χ1 φn = φn , we can write I3 as a telescoping sum:
ˆ n−1 ˆ
X ˆ
I3 = p(u · ∇φn )dtdx = pu · ∇((χk − χk+1 )φn ) + pu · (χn φn ). (8.109)
Q1 k=1 Q1 Q1

Since u is divergence-free, and χk − χ+1 vanishes outside of Qk , we can write for k ≥ 3:

ˆ ˆ ˆ
pu · ∇((χk − χk+1 )φn ) = pu · ∇((χk − χk+1 )φn ) = (p − p̄k )u · ∇((χk − χk+1 )φn ).
Q1 Qk Qk

For k = 1, 2 we simply have

ˆ ˆ
2/3
pu · ∇((χk − χk+1 )φn ) ≤ c |p||u| ≤ Cε1 ,
Q1 Q1

while for the last term in (8.109) we have

ˆ ˆ
pu · (χn φn ) = (p − p̄n )u · ∇(χn φn ).
Q1 Qn

Putting these together, we have

n ˆ n
2/3
X 1 2/3
X 1 2/3 5−3/5 2/3
I3 ≤ Cε1 +C 4
|p − p̄ k ||u| ≤ Cε1 + C ε rk
4 1
≤ Cε1 . (8.110)
r
k=3 k Qk
r
k=3 k

This finishes the proof of Lemma 8.18, and thus that of Proposition 8.4.

9 The weak solutions of the Euler equations

The goal of this section is to give a naive and simplistic glimpse of the recent results on the
Onsager conjecture and the weak solutions of the Euler and Navier-Stokes equations that do
not preserve the energy. This material is mostly based on a recent review by V. Vicol and
T. Buckmaster, but also includes some material from the pioneering papers by C. De Lellis
and L. Székelyhidi.

9.1 The statistical description of turbulence

The starting point of our discussion are the Navier-Stokes equations

uνt + uν · ∇uν + ∇p = ν∆uν + f,

∇ · uν = 0, (9.1)

112
with a small viscosity ν > 0. Our favorite fundamental energy balance says that, as long as
the solution uν (t, x) remains smooth, we have
ˆ ˆ ˆ
1d
|u (t, x)| dx = −ν |∇u (t, x)| dx + (f · uν )dx.
ν 2 ν 2
(9.2)
2 dt
On the other hand, if we consider the Euler equations rather than the Navier-Stokes equations,
with the same forcing

vt + v · ∇v + ∇p = f,
∇ · v = 0, (9.3)

and assume that v(t, x) is also smooth, then the corresponding energy balance is simply
ˆ ˆ
1d 2
|v(t, x)| dx = (f · v)dx. (9.4)
2 dt TL TL

Our interest will be in two issues: first, should we think of the solutions to Euler equations as
the solutions to the Navier-Stokes equations in the limit of a zero viscosity, and, second, how
do the weak solutions to the Euler equations behave when the forcing f is, in some sense,
small. In other words, can a small force f create a large (but oscillatory) solution to the Euler
equations. These issues are quite closely related.
The answer to the first question depends, essentially, on what happens to the energy
dissipation term in the right side of (9.2). Naively, one may expect that this term vanishes
as ν → 0, so that for ν > 0 small it is also small. This, of course, assumes that uν remains
uniformly smooth as ν → 0. As we will see, this is not the case even in much simpler linear
problems. In order to be more specific, we will assume, without any rigorous justification,
that uν satisfies the following hypotheses that reflect the physical observations. First, uν (t, x)
is a space-time stationary random process – its law is the same for all t ∈ R and x ∈ R3 , and
for any a collection of space-time points (t1 , x1 ), . . . , (tN , xN ) and any shifts s ∈ R and y ∈ R3 ,
the joint law of
uν (t1 + s, x1 + y), . . . , uν (tN + s, xN + y)
does not depend on the ”off-sets” s ∈ R and y ∈ R3 but only on the relative times and
positions t1 , . . . , tN and x1 , . . . , xN . Second, we assume that the field uν (t, x) is statistically
isotropic: for any collection of points x1 , . . . , xN , any t ∈ R, and any orthogonal matrix R,
the joint law of
uν (t, Rx), . . . , uν (t, RxN )
is the same as that of uν (t, x), . . . , uν (t, xN ). For the final assumption, let us define the
increments
δuν (t, x, z) = uν (t, x + z) − uν (t, x).
We assume self-similarity of the increments: there is a range of scales `, known as the in-
ertial range, and a constant µ > 0, so that the law of δuν (t, x, λ`ẑ) is the same as that
of λµ δuν (t, x, `ẑ) for all unit vectors ẑ with |ẑ| = 1, and λ > 0 so that both ` and λ` are in
the inertial range.

113
 A basic hypothesis of the theory of turbulence, together with the above space-time homo-
geneity, isotropy and self-similarity properties, is that the average energy dissipation rate

εν = hν|∇uν (t, x)|2 i → ε > 0 as ν → 0, (9.5)

does not vanish in the limit ν → 0. Here, h·i denotes the statistical averaging. This, in a
sense, defines, what it means for uν to be turbulent. This should, naturally, in the limit ν → 0,
lead to the solutions to the Euler equations for which we have an inequality in (9.4) rather
than an equality: ˆ ˆ
1d 2
|v(t, x)| dx < (f · v)dx, (9.6)
2 dt
and which are not smooth. This brings about two fundamental questions: first, how should we
expect the energy dissipation rate to behave for ν small, and, second, for what kind of rough
solutions to the Euler equations should we not expect energy conservation? The former is
addressed by the Kolmogorov theory of turbulence, and the latter by the Onsager conjecture,
though the two are closely related.
Let us define the mean energy per unit volume carried by wave numbers smaller than κ
as h|P≤κ uν |2 i. Here, P≤κ denotes the projection on the wave numbers smaller than κ in the
Fourier space. The energy spectrum of uν is then defined as
d
E(κ) = h|P≤κ uν |2 i. (9.7)
dκ
The main hypothesis of the statistical turbulence theory is that in the inertial range the
energy E(κ) depends only on the limiting average energy density ε in (9.5) and the wave
number κ but not on f or the viscosity ν. The dimensions of these objects are
hd i length2 length3
[E(κ)] = h|P≤κ uν |2 i = length = ,
dκ time2 time2
length2 1 length2
[ε] = = , (9.8)
time time2 time3
1
[κ] = .
length
Therefore, the dimensional analysis implies that the ratio
E(κ)
(9.9)
εa κb
is non-dimensional, and thus should be a constant, if (and only if)

3 = 2a − b, 2 = 3a, (9.10)

so that a = 2/3, b = −5/3. We deduce that in a turbulent flow we should have

E(κ) = CK ε2/3 κ−5/3 , (9.11)

in the inertial range, with some constant CK > 0, that should be determined from the physical
considerations.

114
 The self-similarity exponent µ can also be determined from purely dimensional consider-
ations. Let us define the p-th order absolute structure function as

Sp (`) = h|uν (t, x + `ẑ) − uν (t, x)|p i, |ẑ| = 1, ` > 0.

In the inertial range we should have

Sp (`) = Cdim `pµ , (9.12)

with a dimensional constant Cdim . The physical hypothesis is again that Sp (`) depends only
on ε and `. Note that the corresponding dimensions are

lengthp length2
[Sp (`)] = , [ε] = , [`] = length.
timep time3
We conclude that there exists a non-dimensional constant Dp so that

Sp (`) = Dp (ε`)p/3 .

Comparing to (9.12) we conclude that the self–similarity exponent µ = 1/3.

The inertial range extends from the macroscopic scale of the forcing down to a small
scale `K that should depend only on ε and the viscosity ν. Once again, looking at the
dimensions
length2 length2
[ε] = , [ν] = ,
time3 time
we conclude that the Kolmogorov dissipation length is

cK ν 3/4
λK = , (9.13)
ε1/4
with a constant cK that comes from physical considerations. The constants cK and CK are
not independent – they can be related using the hypothesis that the energy is concentrated
in the inertial scale λK  `  L, together with (9.11) and the relation between ε and E(κ).

9.2 The easy direction of Onsager’s conjecture

Let us now turn to a more mathematical analysis. As usual, we work on the torus T3 . We
say that v(t, x) is a weak solution to the Euler equations

vt + v · ∇v + ∇p = 0, t > 0, x ∈ T3 ,
∇ · v = 0, (9.14)
v(0, x) = v0 (x),

if v ∈ C[0, T ; L2 (R3 )], for any t > 0 the vector field v(t, ·) is divergence-free in the sense of
distributions, and for any divergence-free test function φ(t, x) we have
ˆ ∞ˆ ˆ
v(t, x) · [∂t φ(t, x) + v(t, x) · ∇φ(t, x)]dxdt + v0 (x)φ(0, x)dx = 0. (9.15)
0 T3 T3

115
A smooth solution to Euler’s equations conserves energy:
ˆ ˆ
2
|v(t, x)| dx = |v0 (x)|2 dx. (9.16)
T3 T3

For the weak solutions, Onsager’s conjecture, directly related to the Kolmogorov self-similarity
exponent µ = 1/3, says that (i) a weak solution to the Euler equations that belongs to the
α
Hölder space Ct,x with α > 1/3 conserves energy, and (ii) for any α < 1/3 there exists a weak
α
solution to the Euler equations in the Hölder space Ct,x that does not conserve energy.
The first part of this conjecture is much easier to prove. Let us assume that v(t, x) is C α
in the x-variable, with α > 1/3. Let φ ≥ 0 be a smooth test function in Cc∞ (R3 ) such
that kφkL1 = 1 and set φ` = `−3 φ(x/`), a standard mollifier. Given a function f we will use
the notation
f` = φ` ? f. (9.17)
The mollified vector field v` = v ? φ` satisfies

∂t v` + (v` · ∇v` ) + ([v · ∇v]` − v` · ∇v` ) + ∇p` = 0, t > 0, x ∈ T3 ,

∇ · v` = 0. (9.18)

We can write, using the divergence-free property of v:

[v · ∇v]`,j = φ` ? [vk ∂k vj ] = φ` ? [∂k (vk vj )] = ∂k [φ` ? (vk vj )] = ∂k [(vk vj )` ]. (9.19)

Then, multiplying (9.18) by v` and integrating by parts gives, as v` is also divergence-free:

ˆ ˆ ˆ tˆ
1 2 1 2
|v` (t, x)| dx − |v` (0, x)| dx = − (v`,j [vk ∂k vj ]` − v`,j [v`,k ∂k v`,j )dxds
2 T3 2 T3 T3
ˆ tˆ 0
ˆ tˆ
=− (v`,j ∂k [(vk vj )` ] − v`,j [v`,k ∂k v`,j )dxds = [(vk vj )` − v`,k v`,j ]∂k v`,j dxds. (9.20)
0 T3 0 T3

We have the following lemma.

Lemma 9.1 Let φ ≥ 0 be in Cc∞ (Rd ) and such that kφkL1 = 1, and set φ` (x) = `−d φ(x/`).
Then, for any α ∈ (0, 1) we have

kf ? φ` kC 1 ≤ C`−(1−α) kf kC α , (9.21)

and
k(f g) ? φ` − (f ? φ` )(g ? φ` )kC 0 ≤ C`2α kf kC α kgkC α , (9.22)
with a constant C that depends on φ.

With this lemma in hand, and assuming that v ∈ C α (R3 ), we may estimate the integral in
the right side of (9.20) as
ˆ
[(vk vj )` − v`,k v`,j ]∂k v`,j dx ≤ Ck(vk vj )` − v`,k v`,j kC 0 kv` kC 1
T3
≤ C`2α kvk2C α `−(1−α) kvkC α = C`3α−1 kvk3C α → 0, (9.23)

116
if α > 1/3. Therefore, passing to the limit ` → 0 in (9.20), we obtain
ˆ ˆ
2
|v(t, x)| dx = |v(0, x)|2 dx, (9.24)
T3 T3

thus the energy is conserved.

Let us now prove Lemma 9.1. To prove the first bound in this lemma, we write
ˆ
φ` (x + hek − y) − φ` (x − y)
∂k (f ? φ` )(x) = lim f (y)dy
h→0 h
ˆ
φ` (x + hek − y) − φ` (x − y)
= lim (f (y) − f (x))dy
h→0 h
ˆ
φ((x + hek − y)/`) − φ((x − y)/`) dy
= lim (f (y) − f (x)) n
h→0 h `
ˆ −1
φ(z + h` ek ) − φ(z)
= lim (f (x − `z) − f (x))dz,
h→0 h
so that
ˆ
|φ(z + h`−1 ek ) − φ(z)|
|∂k (f ? φ` )(x)| ≤ lim |f (x − `z) − f (x)|dy
h→0 h
ˆ
|φ(z + hek ) − φ(z)| α
≤ `α−1 kf kC α lim |z| dz ≤ C`α−1 kf kC α . (9.25)
h→0 h
For the second bound, we note that
ˆ
(f g) ? φ` (x) − (f ? φ` )(x)(g ? φ` )(x) = f (y)g(y)φ` (x − y)φ` (x − z)dydz
ˆ
− f (y)φ` (x − y)g(z)φ` (x − z)dydz (9.26)
ˆ
= φ(y)φ(z)[f (x − `y)g(x − `y) − f (x − `y)g(x − `z)]dydz
ˆ
= φ(y)φ(z)[f (x − `z) − f (x − `y)]g(x − `z)dydz
ˆ
= φ(y)φ(z)[f (x − `z) − f (x − `y)][g(x − `z) − g(x)]dydz,

so that

|(f g) ? φ` (x) − (f ? φ` )(x)(g ? φ` )(x)| (9.27)

ˆ
≤ φ(y)φ(z)|f (x − `z) − f (x − `y)||g(x − `z) − g(x)|dydz
ˆ
2α
≤ ` kf kC kgkC
α α φ(y)φ(z)|z − y|α |z|α dydz = C`2α kf kC α kgkC α ,

finishing the proof of Lemma 9.1.

117
9.3 The wild continuous weak solutions of the Euler equations
In this section, we prove existence of a Hölder continuous solution of the Euler equations, with
a sufficiently small Hödler exponent β > 0. The Euler equations written in the divergence
form are
vt + ∇ · (v ⊗ v) + ∇p = 0, t > 0, x ∈ T3 ,
∇ · v = 0, (9.28)
that does not conserve the energy. Here, the torus is normalized as T3 = [0, 1]3 , with the
periodic boundary conditions. We use here and below the notation
[a ⊗ b]ij = ai bj (9.29)
for the standard tensor product of two vectors and
1
[a ⊗tr b]ij = ai bj − (a · b)δij , (9.30)
n
for a traceless tensor product of a pair of vectors a and b in Rn . Only the divergence of the
traceless tensor products will appear below in various equations that also have the pressure
terms, and the trace part can be always added to the gradient of the pressure.
Theorem 9.2 There exists β > 0 and a weak solution v ∈ C([0, T ]; C β (T3 )) to the Euler
equations such that ˆ ˆ
2
|v(1, x)| dx ≥ 2 |v(0, x)|2 dx. (9.31)
T3 T3

The proof proceeds by an induction. We will construct a sequence vq , q = 0, 1, 2, . . . of

solutions to the forced Euler equations
∂vq + ∇ · (vq ⊗ vq ) + ∇pq = ∇ · Rq , t > 0, x ∈ T3 ,
∇ · vq = 0, (9.32)
with a Reynolds stress Rq that goes uniformly to zero as q → +∞, and vq converges uniformly
to a weak solution to the Euler equations satisfying the ”reverse” energy inequality (9.31).
At each induction step, we do not design vq+1 directly but rather use vq to construct the
increment wq+1 = vq+1 − vq , in such a way that vq+1 satisfies
∂t vq+1 + ∇ · (vq+1 ⊗ vq+1 ) + ∇pq+1 = ∇ · Rq+1 , t > 0, x ∈ T3 ,
∇ · vq+1 = 0, (9.33)
with a smaller Reynolds stress Rq+1 . Given wq+1 , the Reynolds stress Rq+1 is determined as
the trace-less symmetric matrix satisfying
∇ · Rq+1 = Eosc + Etr + EN ash , (9.34)
with the ”error” terms in the right side depending on vq and wq+1 :
Eosc = ∇ · (wq+1 ⊗ wq+1 ) − ∇ · Rq + ∇(pq+1 − pq ), (9.35)
Etr = ∂t wq+1 + vq · ∇wq+1 , (9.36)
EN ash = wq+1 · ∇vq . (9.37)

118
These terms are known as the oscillation error, the transport error and the Nash error,
respectively. Given the iterate vq , the goal will be to choose wq+1 so that Rq+1 , the symmetric
trace-less solution to (9.34) with a given right side, is small, and, in addition, the series
X
wq (9.38)
q

converges. In order to make sure that the reverse energy inequality (9.31) holds, we will
choose the first iterate v0 so that v0 (0, x) ≡ 0, and v0 (1, x) does not vanish. This means
that v0 (t, x) satisfies (9.31) trivially. The induction construction will ensure that actually
all vq (t, x) stay sufficiently close to v0 (t, x) for all 0 ≤ t ≤ 1, so that in the limit q → +∞ the
inequality (9.31) will still hold.
The correction wq+1 will consist of two parts:
(p) (c)
wq+1 = wq+1 + wq+1 . (9.39)
(p)
Here, wq+1 is the principal part of the perturbation, chosen so that the low frequency terms in
the trace-less product wq+1 ⊗tr wq+1 essentially cancel those in Rq , so that these contributions
to the oscillation error cancel each other. Roughly speaking, it is of the form
(p)
X
wq+1 ∼ aξ (Rq )Wξ . (9.40)
ξ

Here, Wξ are ”building blocks” oscillating at a high frequency λq+1 , and the coefficients aξ (Rq )
are chosen so that the aforementioned cancellation of the lower frequencies takes place. As
(p)
an additional minor complication, wq+1 will need to be corrected to decrease the transport
(c)
error. The correction wq+1 is chosen to ensure that wq+1 is divergence-free.
In order to see yet another way the threshold 1/3 for the Hödler regularity comes up, let
us assume that the frequencies are chosen so that
λq = λq , (9.41)
with some λ ∈ N. Then, in order for the series in (9.38) to converge to a C β function v, we
should have, at least,
kwq kC 0 ≤ λ−β
q . (9.42)
The Reynolds stress should then satisfy, roughly
−2β
kRq kC 0 ≤ λq+1 , (9.43)
because it is related quadratically to wq+1 . The contribution of the Nash error to the Reynolds
stress Rq+1 is one derivative smoother that EN ash , and oscillates at frequency λq+1 . In the
uniform norm, it should be of the order
Ckwq+1 kC 0 kvq kC 1 Cλ−β
q+1
X −β−1 1−β
N ash
kRq+1 kC 0 ≤ ≤ λm λ−β
m ≤ Cλq+1 λq
λq+1 λq+1 m≤q
= Cλ−β−1
q+2 λ
β+1 1−β 2(β−1)
λq+2 λ −2β 3β−1
≤ Cλq+2 λ . (9.44)
In other words, for the bound (9.43) to be ”iteratable” we need to have β < 1/3, another
indication for why Onsager’s conjecture holds. In reality, we will take the frequencies growing
much faster than in (9.41), and we will also take β to be very small.

119
9.4 The iterative estimate
We now turn to an implementation of the above scheme. We will take the frequencies
q
λ q = a2 , (9.45)
with a ∈ N sufficiently large, to be specified later, so that
λq+1 = λ2q . (9.46)
We also set
δq = λ−2β
q , (9.47)
with β > 0 sufficiently small, also to be specified later. We will assume the following inductive
bounds on vq and Rq :
kvq kC 0 ≤ 1 − δq1/2 , (9.48)
kvq kCt,x
1 ≤ CR δq1/2 λq , (9.49)
kRq kC 0 ≤ cR δq+1 , (9.50)
with a pair of universal constants CR and cR , to be specified below. Let us explain the choices
here. As we have mentioned above, the basic premise is that the increment wq+1 = vq+1 − vq
1/2
is of the size δq+1 in the uniform norm – see (9.52) below, and oscillates at frequency λq+1 .
Then the Reynolds stress Rq should be of the size δq+1 in the uniform norm, simply because it
is quadratic in wq+1 , which gives the induction assumption (9.50). The uniform bound (9.48)
is a convenient induction assumption since
vq+1 = vq + wq+1 ,
so that if (9.48) holds at level q, and we have (9.52) below, then
1/2 1/2
kvq+1 k ≤ 1 − δq1/2 + δq+1 ≤ 1 − δq+1 .
Finally, assumption (9.49) on the C 1 -norm of vq comes about because the frequencies λq grows
sufficiently fast, so that even though kwq+1 kC 0  kwq kC 0 , we still have kwq+1 kC 1  kwq kC 1 ,
so that the main contribution to kvq kC 1 comes from kwq kC 1 , which is of the size
kwq kC 1 ∼ λq+1 kwq kC 0 = λq+1 δ 1/2 .
1/2
Note that δq λq → +∞ as q → +∞ in (9.49), since β > 0 is small – because of the easy part
α
of Onsager’s conjecture, we do not expect vq to converge in a Hölder space Ct,x with α > 1/3,
1
let alone in Ct,x . The induction step is described in the following.
Proposition 9.3 There exists β > 0 sufficiently small and a0 sufficiently large, so that for
any a ≥ a0 there exist vq and Rq , q ≥ 0, that satisfy
∂vq + ∇ · (vq ⊗ vq ) + ∇pq = ∇ · Rq , t > 0, x ∈ T3 ,
∇ · vq = 0, (9.51)
and obey (9.48)-(9.50), and such that
1/2
kvq+1 − vq kC 0 ≤ δq+1 . (9.52)

120
 Let us explain how Proposition 9.3 implies the conclusion of Theorem 9.2. We take the
first iterate to be an oscillatory shear flow
t 1/2
v0 (t, x) = (sin(λ0 x3 ), 0, 0). (9.53)
2
Then we have
1 1/2
kv0 (t, ·)kC 0 ≤≤ 1 − δ0 , (9.54)
2
so that (9.48) is satisfied, for a large enough. We also have
1/2 1/2
kv0 (t, ·)kCt,x
1 ≤ λ0 ≤ λ0 δ0 , (9.55)

as long as δ0−1 ≤ λ0 , which is true as long as β < 1/2, and a is sufficiently large. Hence, (9.49)
also holds for q = 0. To find R0 we note that, as v0 is a shear flow, we have v0 · ∇v0 = 0,
hence  1/2

1 0 0 − cos(λ 0 x 3 )
R0 = 1/2  0 0 0 , (9.56)
2λ0 1/2
− cos(λ0 x3 ) 0 0
so that
1 1/2 ∂v0
∇ · R0 = (sin(λ0 x3 ), 0, 0) = .
2 ∂t
It follows that
1
kR0 kC 0 = 1/2
≤ a−1/2 ≤ cR δ1 = cR a−4β , (9.57)
2λ0
provided that β < 1/8 and a is sufficiently large. Hence, condition (9.50) also holds at q = 0.
A key consequence of (9.57) is that v0 is a solution of the forced Euler equations with a
Reynolds stress that is already very small in the uniform norm, provided that we take a
sufficiently large. In addition, the L2 -norm of v0 vanishes at t = 0 but is not zero at t = 1.
We will now construct a rough weak solution to the unforced Euler equations that will be
close to v0 (t, x) in the uniform norm for all 0 ≤ t ≤ 1, and this will force it to violate the
energy inequality.
We start the iteration as in Proposition 9.3, with the initialization (v0 , R0 ), and obtain a
sequence (vq , Rq ). Let us take β > 0 as in the definition (9.47) of δq . Then, for any α < β
the bounds (9.48), (9.49) and (9.52), together with an interpolation inequality between the
1/2
Hölder norms, and the fact that the sequence δq λq in the right side of (9.49) is monotonically
increasing, imply that
(1−α)/2 α/2 α 1/2 −(β−α)
kvq+1 −vq kCxα ≤ cα kvq+1 −vq k1−α α
Cx0 kvq+1 −vq kCx1 ≤ cα δq+1 δq+1 λq+1 = cα δq+1 λαq+1 = cα λq+1 .
(9.58)
Thus, the limit
v = lim vq
q→+∞

exists in C([0, 1], C α (T3 )) for any α < β. Furthermore, (9.50) implies that

Rq → 0 in C 0 ([0, 1] × T3 ).

121
It follows that v(t, x) is a weak solution to the Euler equations that lies in C([0, 1], C α (T3 ))
for any α < β.
To finish the proof of Theorem 9.2, it remains to show that the reverse energy inequality

kv(1, ·)kL2 ≥ 2kv(0, ·)kL2 (9.59)

holds. The point is that, if a is sufficiently large, then, on one hand, v(t, x) is close in the
uniform norm to v0 (t, x) for all 0 ≤ t ≤ 1, and on the other v0 (0, x) = 0 while v0 (1, x) has a
fixed non-zero L2 -norm that is independent of λ0 . Indeed, we have using (9.52):
∞ ∞ ∞ ∞
q
X X 1/2
X X
kv − v0 kC 0 ≤ kvq+1 − vq kC 0 ≤ δq+1 = λ−β
q = a−β·2
q=0 q=0 q=0 q=0
∞
X 1
≤ a−β(q+1) ≤ , (9.60)
q=0
10000

if a is sufficiently large, so that v and v0 are close. It follows that

1 1
2kv(0, ·)kL2 ≤ 2kv0 (0, ·)kL2 + = ≤ kv0 (1, ·)kL2 − kv0 (1, ·) − v(1, ·)kL2 ≤ kv(1, ·)kL2 ,
100 100
finishing the proof of Theorem 9.2.

9.5 Proof of Proposition 9.3

We now prove Proposition 9.3. We only need to prove the inductive step as we have already
constructed the pair (v0 , R0 ). It will be more convenient to work with the mollified versions
of vq and Rq defined as

v` = (vq ?x φ` ) ?t ϕ` , (9.61)
R` = (Rq ?x φ` ) ?t ϕ` .

We dropped the subscript q above in v` and R` to simplify the notation. Here, φ` (x) and ϕ` (t)
are standard scalar-valued mollifiers of compact support in x and t, respectively. As we are
not aiming to prove an optimal result, we choose the mollification scales in x and t to be the
−1
same, and take ` as an intermediate scale between λ−1
q and λq+1 :

` = λq−3/2 , (9.62)

so that
λ−1 −2 −1
q+1 = λq ≤ ` ≤ λq . (9.63)
Note that, by the induction hypothesis (9.48), we have

kv` kC 0 ≤ kvq kC 0 ≤ 1 − δq1/2 , (9.64)

and for any N ≥ 1 we have, because of the way ` was chosen and the second induction
hypothesis (9.49):

kv` kC N ≤ C`−N +1 kvq kC 1 ≤ C`−N +1 λq δq1/2 ≤ C`−N , (9.65)

122
while
1/2
kvq − v` kC 0 ≤ `kvq kC 1 ≤ C`λq δq1/2 ≤ Cλ−1/2
q δq1/2  δq+1 , (9.66)
as long as β > 0 is sufficiently small.
As in (9.18), we obtain

∂t v` + ∇ · [v` ⊗ v` ] + ∇p` = ∇ · (R` + Rcomm ), (9.67)

∇ · v` = 0,

with
Rcomm = v` ⊗tr v` − [(v ⊗tr v) ?x φ` ] ?t ϕ` . (9.68)
Recall that the traceless tensor product ⊗tr is defined in (9.30). In (9.67), with a slight abuse
of notation, the pressure p` includes both the convolution of pq with the mollifiers and what
should have been the trace part of Rcomm . Note that, as in (9.27), we have, using (9.49)
and (9.30):

kRcomm kCt,x
0 ≤ C`kvkCt,x
1 kvkC 0
t,x
≤ C`δq1/2 λq = Cλ−3/2
q λ−β −β−1/2
q λq = λq  δq+2 , (9.69)

provided that β is sufficiently small and a is sufficiently large.

9.5.1 The Reynolds stress equation

Let us first address the equation for the Reynolds stress:

∇ · R = E(x), x ∈ T3 , (9.70)

with the condition that R(x) is a symmetric trace-free matrix. Here, E is a mean-zero vector-
field on T3 : ˆ
E(x)dx = 0. (9.71)
T3

We claim that a trace-less symmetric solution to (9.70) is given by

1
Rkm (x) = (∂k ∆−1 Em + ∂m ∆−1 Ek ) − (δkm + ∂k ∂m ∆−1 )∆−1 (∇ · E). (9.72)
2
The symmetry and mean-zero properties of R are obvious from (9.72). Its trace vanishes
because
n + 1 −1
TrR = 2∆−1 (∇ · E) − ∆ (∇ · E) = 0
2
in dimension n = 3. To check (9.70) we write
1
(∇ · R)m = ∂k Rkm = ∂k (∂k ∆−1 Em + ∂m ∆−1 Ek ) − ∂k (δkm + ∂k ∂m ∆−1 )∆−1 (∇ · E)
2
1 1
= Em + ∂m ∆−1 (∇ · E) − ∂m ∆−1 (∇ · E) − ∂m ∆−1 (∇ · E) = Em . (9.73)
2 2
The next lemma says that R is similar to (−∆)−1/2 E when E is oscillatory.

123
Lemma 9.4 Assume that a(x) ∈ C m,α (T3 ) and Φ ∈ C m,α (T3 ) be smooth R3 -valued functions,
let C be such that
C −1 ≤ |∇Φ(x)| ≤ C for all x ∈ T3 .
Let ω ∈ Z3 , α ∈ (0, 1) and m ≥ 1, and R(x) be the solution to (9.70) with
ˆ
iω·Φ(x)
E(x) = a(x)e − a(y)eiω·Φ(y) dy, (9.74)
T3

given by (9.72). There exists a constant K that depends on C, α and m but not on |ω| such
that  kak 0
C 1 
kRkC α ≤ C + (kak C m,α + kakC 0 k∇ΦkC m,α ) . (9.75)
|ω|1−α |ω|m−α
Proof. To be filled in.
The reason we allow a phase factor Φ(t, x) in Lemma 9.4 is that we will need to modify
the phase to decrease the transport error, as discussed in Section 9.5.3 below. Our strategy
will be to construct wq+1 so that Rq satisfies (9.70) with a right side that is as in Lemma 9.4:
mean-zero and oscillatory, ”essentially” at a single, sufficiently high frequency: in particular,
the terms
kakC 0 kakC m,α
,
|ω|1−α |ω|m−α
in the right side of (9.75) should be small.

9.5.2 The Beltrami flows

The building blocks we will use to construct the principal part of the perturbation wq+1 as
in (9.40): X
(p)
wq+1 ∼ aξ (Rq )Wξ (9.76)
ξ

are the Beltrami waves Wξ . They are defined as follows. Recall that the set Q3 ∩ S2 of
rational points is dense on the unit sphere S2 . To see that, consider the inverse map of the
stereographic projection s(x, y) : R2 → S2
 2y 2x x2 + y 2 − 1 
s(x, y) = , , .
x2 + y 2 + 1 x2 + y 2 + 1 x2 + y 2 + 1

It is clear that s maps Q2 to Q3 ∩ S2 . Since Q2 is dense in R2 and s is a diffeomorphism of R2

onto S2 \ (0, 0, 1), the density of Q3 ∩ S2 in S2 follows.
Next, given ξ ∈ Q3 ∩ S2 , we take Aξ ∈ Q3 ∩ S2 so that

Aξ · ξ = 0, A−ξ = Aξ . (9.77)

The choice of Aξ is not unique: for instance, we can take Aξ = (−ξ2 , ξ1 , 0) for ξ = (ξ1 , ξ2 , ξ3 )
with ξ1 ≥ 0 and extend it to ξ with ξ1 < 0 using the even symmetry in (9.77). We also define
the complex vector
1
Bξ = √ (Aξ + iξ × Aξ ). (9.78)
2

124
By construction, the vector Bξ satisfies

|Bξ | = 1, Bξ · ξ = 0, iξ × Bξ = Bξ , B−ξ = B ξ , (9.79)

with · denoting the standard real inner product, without the complex conjugation, and the
bar denoting the complex conjugation. The third identity above relies on the formula

[ξ × (ξ × Aξ )]k = εkmj ξm εjrs ξr (Aξ )s = [δrk δms − δks δmr ]ξm ξr (Aξ )s
= ξk (ξ · Aξ ) − |ξ|2 (Aξ )k = −(Aξ )k .

It follows that for any λ ∈ Z such that λξ ∈ Z3 , the function

Wξ,λ (x) = Bξ e2πiλξ·x (9.80)

satisfies
[∇ × Wξ,λ ]j = εjkm 2πiλξk Bξ,m e2πiλξ·x = 2πλBξ,j e2πiλξ·x , (9.81)
and is therefore a periodic eigenfunction of the curl operator corresponding to the eigen-
value 2πλ:
∇ × Wξ,λ = 2πλWξ,λ . (9.82)
We can now fix λ ∈ Z and take any finite set Γ ⊂ Q3 ∩ S2 such that −Γ = Γ and λξ ∈ Z3 for
any ξ ∈ Γ. Then for any collection of coefficients aξ ∈ C such that a−ξ = āξ , the vector field
X
W (x) = aξ Bξ e2πiλξ·x (9.83)
ξ∈Γ

is a real-valued divergence free vector field on T3 such that

∇ × W (x) = 2πλW (x). (9.84)

Note that for any vector W we have

(W × [∇ × W ])i = εijk Wj εkmn ∂m Wn = (δim δjn − δin δjm )Wj ∂m Wn = Wj ∂i Wj − Wj ∂j Wi ,

which gives the vector identity

 |W |2 
W × [∇ × W ] = ∇ − W · ∇W. (9.85)
2
It follows from (9.84) and (9.85) that
 |W |2 
W · ∇W = ∇ . (9.86)
2
In other words, any W of the form constructed above is a solution of the Euler equations,
with zero pressure.
Observe also that given any ξ ∈ Q3 ∩ S2 , the vectors ξ, A(ξ) and ξ × A(ξ) form an
orthonormal basis, so that

ξ ⊗ ξ + Aξ ⊗ Aξ + (ξ × Aξ ) ⊗ (ξ × Aξ ) = Id, (9.87)

125
which implies
1
Bξ ⊗ B−ξ = (Aξ + iξ × Aξ ) ⊗ (Aξ − iξ × Aξ ) (9.88)
2
1 i
= (Aξ ⊗ Aξ + (ξ × Aξ ) ⊗ (ξ × Aξ )) + [(ξ × Aξ ) ⊗ Aξ − Aξ ⊗ (ξ × Aξ )]
2 2
1 i
= (Id − ξ ⊗ ξ) + [(ξ × Aξ ) ⊗ Aξ − Aξ ⊗ (ξ × Aξ )].
2 2
It follows that for W of the form (9.83) we have
ˆ X ˆ 0
X
(W ⊗ W )dx = aξ aξ0 e2πiµ(ξ+ξ )·x (Bξ ⊗ Bξ0 )dx = aξ a−ξ (Bξ ⊗ B−ξ )
T3 ξ,ξ 0 ∈Γ T3 ξ∈Γ
1X
= |aξ |2 (Id − ξ ⊗ ξ), (9.89)
2 ξ∈Γ

because
X
|aξ |2 ((ξ × Aξ ) ⊗ Aξ − Aξ ⊗ (ξ × Aξ ) = 0, (9.90)
ξ∈Γ

as the individual terms inside the sum are odd in ξ and the set Γ is symmetric: −Γ = Γ.
We will use the Beltrami flows as building blocks in the decomposition (9.40) for the
principal part of the perturbation wq+1 :
(p)
X
wq+1 ∼ aξ (Rq )Wξ,λq+1 . (9.91)
ξ

The goal will be to cancel out the average of the Rq term in the oscillation error (9.35)

Eosc = ∇ · (wq+1 ⊗ wq+1 ) − ∇ · Rq + ∇(pq+1 − pq ), (9.92)

so that Eosc has the form (9.74) in Lemma 9.4. To this end, we need to know that the family
of the Beltrami flows is rich enough so that the cancellation is achievable for a large class of
given matrices Rq . Keeping in mind expression (9.89), we will now prove the following. We
denote by Br (Id) the closed ball of 3 × 3 symmetric matrices centered at Id, of radius r.

Lemma 9.5 There exist two disjoint finite subsets Λ0 , Λ1 ⊂ Q3 ∩ S2 such that if ξ ∈ Λj
then −ξ ∈ Λj , and r0 > 0, so that for each matrix M ∈ Br0 (Id) and j = 0, 1, we have a
decomposition
1 X (j)
M= (γ (M ))2 (Id − ξ ⊗ ξ). (9.93)
2 ξ∈Λ ξ
j

(j)
Moreover, for each ξ ∈ Λj and j = 0, 1, the coefficients γξ (R) are C ∞ -functions on Br0 (Id).

Proof. To be filled in.

126
9.5.3 The principal part of the perturbation
We would like to take the principal part of the perturbation as a sum of the Beltrami waves.
At the same time, we need to make sure that we have a small transport error in (9.36)

Etr = ∂t wq+1 + vq · ∇wq+1 . (9.94)

To this end, we will replace the phase ξ · x in the definition of the Beltrami wave by a
phase Φ(t, x) that is transported by the vector field vq . We divide the interval 0 ≤ t ≤ 1 into
intervals of length `, and for j = 0, . . . , [`−1 ], we define Φj (t, x) as the T3 -periodic solution to

∂t Φj + v` · ∇Φj = 0, (9.95)
Φj (j`, x) = x.

We have the following standard estimates for Φj : first, differentiating (9.95) in x, and using
Gronwall’s inequality and the inductive assumption (9.49) gives

k∇Φj (t) − IdkC 0 ≤ C`kv` kC 1 ≤ C`λq δq1/2 = Cλ−1/2

q δq1/2  1, for all (j − 1)` ≤ t ≤ (j + 1)`.
(9.96)
Differentiating (9.95) once again gives

k∇Φj (t)kCt,x
1 ≤ Cλq δq1/2 , for all (j − 1)` ≤ t ≤ (j + 1)`, (9.97)

and, more generally,

k∇Φj (t)kC n ≤ C`1−n λq δq1/2  `−n , for all (j − 1)` ≤ t ≤ (j + 1)`. (9.98)

Each Φj (t, x) will play a role only on the time interval [(j − 2)`, (j + 2)`]. For this, we will
make use of time-cutoffs: take a non-negative bump function χ(t) supported in [−1, 1] so
that χ(t) ≡ 1 on [−1/2, 1/2] and such that the shifts

χj (t) = χ(`−1 t − j)

satisfy X
χ2j (t) ≡ 1 for 0 ≤ t ≤ 1. (9.99)
j

Note that each time t at most two of χj (t) are non-zero.

Let us recall the sets Λ0 and Λ1 from Lemma 9.5. For a general j we will set Λj = Λ0
(j)
if j is even and Λj = Λ1 if j is odd. We do the same for the functions γξ appearing in that
(p)
lemma. With this notation, we define for the principal part of the perturbation wq+1 (t, x) as
(p)
XX
wq+1 (t, x) = w(ξ) (t, x), (9.100)
j ξ∈Λj

with each individual wave w(ξ) (t, x) in the form of a modulated Beltrami wave

w(ξ) (t, x) = aq+1,j,ξ (t, x)Wξ,λq+1 (Φj (t, x)) = aq+1,j,ξ (t, x)Bξ exp {2πiλq+1 ξ · Φj (t, x)}. (9.101)

127
Note that
(∂t + vq · ∇)(exp{2πiλq+1 ξ · Φj (t, x)}) = 0, (9.102)
so that

(∂t + vq · ∇)w(ξ) = (∂t + vq · ∇)[aq+1,j,ξ ](t, x)Bξ (exp{2πiλq+1 ξ · Φj (t, x)}), (9.103)

and the potentially dangerous term of the size λq+1 coming from the differentiation of the
exponent vanishes. This is why we use the phases Φj (t, x) rather than simply x. The ampli-
tudes aq+1,j,ξ (t, x) are chosen as
1/4 1/2 (j)
aq+1,j,ξ (t, x) = cR δq+1 χj (t)γξ (M` (t, x)), (9.104)

with the matrix

−1/2 −1
M` (t, x) = Id − cR δq+1 R` (t, x). (9.105)
(j)
As the functions γξ are defined only in the ball Br0 (Id), we need to check that the ma-
trices M` (t, x) are in that ball for all 0 ≤ t ≤ 1 and x ∈ T3 . Recalling the inductive
assumption (9.50), we see that
−1/2 −1 −1/2 −1 1/2
cR δq+1 kR` kC 0 ≤ cR δq+1 cR δq+1 ≤ cR ≤ r0 , (9.106)

with r0 as in Lemma 9.5, provided we take

cR ≤ r02 . (9.107)
(j)
It follows that the matrix M` (t, x) is, indeed, in the domain of definition of the functions γξ
for all j, all t ∈ [0, 1] and x ∈ T3 . As at most two of the functions χj do not vanish for any
given t ∈ [0, 1], and they satisfy 0 ≤ χj (t) ≤ 1, we have a uniform estimate
1/2
(p) 1/4 1/2 δq+1
kwq+1 (t, x)kC 0 ≤ K0 cR δq+1 ≤ , (9.108)
2
provided that we choose cR sufficiently small, depending only on a universal constant K0
(j)
that itself depends only on the uniform norm of the functions γξ (M ) on Br0 (Id) and on
the number of elements in the finite sets Λ0 and Λ1 . The above estimate accounts for the
(p)
contribution of wq+1 to the error bound (9.52).

9.5.4 The incompressibility correction

Let us write the individual terms w(ξ) (t, x) that appear in (9.100) as

w(ξ) (t, x) = aq+1,j,ξ (t, x) exp {2πiλq+1 ξ · φj (t, x)}Bξ exp {2πiλq+1 ξ · x} (9.109)
= aq+1,j,ξ (t, x) exp {2πiλq+1 ξ · φj (t, x)}Wξ,λq+1 (x) = bq+1,j,ξ (t, x)Wξ,λq+1 (x),

with

φj (t, x) = Φj (t, x) − x, bq+1,j,ξ (t, x) = aq+1,j,ξ (t, x) exp {2πiλq+1 ξ · φj (t, x)}. (9.110)

128
Recalling (9.96), we can think of φj (t, x) as small, so the largest contribution to ∇w(ξ) (t, x)
should come from the Beltrami wave Wξ,λq+1 (x). However, the latter is incompressible so one
can think of w(ξ) as incompressible to the leading order. To be more precise, let us use (9.84)
to write
1 h i
bq+1,j,ξ (t, x)Wξ,λq+1 (x) = ∇ × (bq+1,j,ξ (t, x)Wξ,λq+1 (x)) − (∇bq+1,j,ξ (t, x)) × Wξ,λq+1 (x) .
2πλq+1

While the first term above is incompressible, the second is not. Accordingly, to compensate
for the second term, we define

(c) 1
w(ξ) (t, x) = (∇bq+1,j,ξ (t, x)) × Wξ,λq+1 (x)
2πλq+1
1  
= ∇aq+1,j,ξ + 2πiλq+1 aq+1,j,ξ (∇Φj (t, x) − Id)ξ × Bξ exp{2πiλq+1 ξ · Φj (t, x)}
2πλq+1
 ∇a 
q+1,j,ξ
= + iaq+1,j,ξ (∇Φj (t, x) − Id)ξ × Wξ,λq+1 (Φj (t, x)). (9.111)
2πλq+1

The full incompressibility correction is then

(c)
XX (c)
wq+1 (t, x) = w(ξ) (t, x), (9.112)
j ξ∈Λj

and the full perturbation is

(p) (c) 1 XX
wq+1 (t, x) = wq+1 (t, x) + wq+1 (t, x) = ∇ × [bq+1,j,ξ (t, x)Wξ,λq+1 (x)], (9.113)
2πλq+1 j ξ∈Λ
j

so that
∇ · wq+1 = 0, (9.114)
and wq+1 (t, x) is mean-zero. We may also estimate the incompressible correction, starting
with the right side of (9.111), and once again using the fact that χj (t) satisfy 0 ≤ χj (t) ≤ 1,
and only two of χj (t) do not vanish for any t ∈ [0, 1] as
h k∇a
q+1,j,ξ kC 0
i
(c)
kwq+1 kC 0 ≤ K sup sup + kaq+1,j,ξ kC 0 k∇Φj − IdkC 0 , (9.115)
j ξ∈Λj λq+1

with a universal constant K. At the moment, we do not have a good bound on k∇aq+1,j,ξ kC 0
as that would require a bound on k∇R` kC 0 , since R` enters the definition (9.104)-(9.105)
of aq+1,j,ξ . However, a standard mollification estimate, together with (9.104)-(9.105) and the
induction assumption (9.50), show that the first term above can be bounded as
1/2 1/2
k∇aq+1,j,ξ kC 0 1/2 −1 −1 kR` kC 0 δq+1 δq+1
≤ Kδq+1 δq+1 λq+1 ≤K ≤ , (9.116)
λq+1 ` `λq+1 100

because
`λq+1 = λq−3/2 λ2q  1.

129
Here we see that it is important that aq+1,j,ξ oscillate on scales much larger than λ−1
q+1 . The
second term in the right side of (9.115) can be estimated with the help of (9.96) as
1/2
1/2 δq+1
kaq+1,j,ξ k k∇Φj − Idk
C0 C0 ≤ Kδq+1 λ−1/2
q δq1/2 ≤ , (9.117)
100
provided that a is sufficiently large and β is sufficiently small. It follows that
1/2
(c) δq+1
kwq+1 kC 0 ≤ . (9.118)
10
Together with (9.108), this finishes the proof of the error bound (9.52):
3 1/2
kwq+1 kC 0 ≤ δq+1 . (9.119)
4
However, we still need to verify that the induction bounds (9.48)-(9.50) hold for vq+1 and Rq+1 .

9.5.5 The induction estimates on the velocity

We first prove the inductive estimates (9.48)-(9.49) on the velocity vq+1 , as they follow directly
from the construction of the perturbation wq+1 . It is convenient to define vq+1 not as vq +wq+1
but as
vq+1 = v` + wq+1 . (9.120)
The uniform bound in (9.48) for q + 1 follows simply from this estimate at level q and (9.119),
together with (9.66):
1/2
kvq − v` kC 0  δq+1 , (9.121)
which gives
1 1/2 3 1/2 1/2
kvq+1 kC 0 ≤ kvq kC 0 + kvq − v` kC 0 + kwq+1 kC 0 ≤ 1 − δq1/2 + δ + δ ≤ 1 − δq+1 , (9.122)
10 q+1 4 q+1
since we have δq+1 ≤ 4δq if we choose a sufficiently large, for a given fixed small β > 0.
To get the gradient bound (9.49) at the level q + 1 we first recall that for the spatial
derivatives we have (9.96), (9.97) and (9.116):

k∇Φj (t) − IdkC 0 ≤ Cλq−1/2 δq1/2 , (9.123)

k∇Φj (t)kCt,x
1 ≤ Cλq δq1/2 , (9.124)
1/2
δq+1
k∇aq+1,j,ξ kC 0 ≤C . (9.125)
`
Once again, as at most two of χj (t) do not vanish for any t > 0, it follows from (9.101)-(9.105)
that the principal part of the perturbation satisfies
 
(p)
k∇wq+1 kC 0 ≤ K sup sup k∇aq+1,j,ξ kC 0 + kaq+1,j,ξ kC 0 λq+1 k∇Φj kC0 , (9.126)
j j∈Λj

130
with a constant K that depends only on the number of the elements of the sets Λ0 and Λ1 .
The first term above we estimate by (9.125), and the second by (9.123), which gives
1/2
(p) δq+1 1/2 1/2
k∇wq+1 kC 0 ≤C + Cδq+1 λq+1 ≤ Cδq+1 λq+1 . (9.127)
`
(c)
For the spatial derivative of wq+1 , we note that
 ka
(c) q+1,j,ξ kC 2
k∇wq+1 kC ≤ K sup sup
0 + k∇aq+1,j,ξ kC 0 k∇Φj − IdkC 0
j j∈Λj λq+1

(c)
+kaq+1,j,ξ kC 0 k∇Φj kC 1 + λq+1 kwq+1 kC 0 . (9.128)
The first term above, once again, can be bounded using the basic mollification estimate as
1/2 1/2 3/2 1/2
kaq+1,j,ξ kC 2 Kδq+1 h kR` kC 0 kR` k2C 0 i Kδq+1 λq+1 δq+1 λq+1
≤ + 2
≤ ≤ . (9.129)
λq+1 λq+1 δq+1 `2 δq+1 `2 λq+1 100
The second term in the right side of (9.128) is estimated using (9.116) and (9.123) as
1/2 1/2
Kδq+1 1/2 −1/2 1/2 δq+1 λq+1
k∇aq+1,j,ξ kC 0 k∇Φj − IdkC 0 ≤ δq λq = Kδq+1 δq1/2 λq ≤ . (9.130)
` 100
The third and the fourth terms in right side of (9.128) satisfy
(c) 1/2 1/2 1/2
kaq+1,j,ξ kC 0 k∇Φj kC 1 + λq+1 kwq+1 kC 0 ≤ Kδq+1 λq δq1/2 + λq+1 δq+1 ≤ 2λq+1 δq+1 . (9.131)
Putting together the above estimates, we see that
1/2
k∇wq+1 kC 0 ≤ CR λq+1 δq+1 , (9.132)
with a universal constant CR . In particular, we have not used the estimate (9.49) at level q
in deriving (9.131), hence there is no danger that CR may change from step q to step q + 1.
For the time derivative we have (9.95), which, together with (9.96) shows that
k∂t Φj (t)kC 0 ≤ C, (9.133)
while
1/2 1/2
k∂t aq+1,j,ξ kC 0 ≤ Cδq+1 (`−1 + `−1 ) = Cδq+1 `−1 , (9.134)
as in the estimate (9.129) for the gradient of aq+1,ξ,j . These two bounds give
(p) 1/2
k∂t wq+1 kC 0 ≤ Cδq+1 `−1 , (9.135)
(c)
in the same way as (9.125) and (9.126) lead to (9.127). For the time derivative of wq+1 , we
note that
 k∂ ∇a
(c) t q+1,j,ξ kC 0
k∂t wq+1 kC 0 ≤ K sup sup + k∂t aq+1,j,ξ kC 0 k∇Φj − IdkC 0
j j∈Λj λq+1

(c)
+kaq+1,j,ξ kC 0 k∂t ∇Φj kC 0 + λq+1 kwq+1 kC 0 . (9.136)
The first term above is estimated exactly as in (9.129), the second as in (9.130), the third
and the fourth as in (9.131), which gives us
(c) 1/2
k∂t wq+1 kC 0 ≤ Cδq+1 `−1 , (9.137)
finishing the proof of (9.49) at level q + 1.

131
9.5.6 The new Reynolds stress
We finally come to the key estimate in the proof of Proposition 9.3: the proof of the inductive
estimate (9.50) at level q + 1 that shows that the Reynolds stress decreases at each inductive
step and tends to zero in the uniform norm as q → +∞. The analysis is based on Lemma 9.4
that we state again here.

Lemma 9.6 Assume that a(x) ∈ C m,α (T3 ) and Φ ∈ C m,α (T3 ) be smooth R3 -valued functions,
and let C be such that
C −1 ≤ |∇Φ(x)| ≤ C for all x ∈ T3 .
Let ω ∈ Z3 , α ∈ (0, 1) and m ≥ 1, and R(x) be the solution to

∇ · R = E(x), TrR(x) = 0, R(x) is a symmetric matrix, (9.138)

with ˆ
iω·Φ(x)
E(x) = a(x)e − a(y)eiω·Φ(y) dy, (9.139)
T3
given by
1
Rkm (x) = (∂k ∆−1 Em + ∂m ∆−1 Ek ) − (δkm + ∂k ∂m ∆−1 )∆−1 (∇ · E). (9.140)
2
There exists a constant K that depends on C, α and m but not on |ω| such that
 kak 1 
C0
kRkC α ≤ C 1−α
+ (kakC m,α + kakC 0 k∇ΦkC m,α ) . (9.141)
|ω| |ω|m−α

In particular, if E(x) is of the form

E(x) = a(x)W(ξ) (Φj (x)), W(ξ) (x) = Bξ e2πiλq+1 ξ·x , (9.142)

with
kakC n ≤ C`−n , k∇Φj kC n ≤ C`−n , (9.143)
then R(x) satisfies

C  `−m−α  C  `−m−1  C 3/4(m+1)−m+1

 C
kRkC 0 ≤ 1+ m−α ≤ 1+ = (1+λq+1 ≤ 1−α , (9.144)
λ1−α
q+1 λq+1 1−α
λq+1 m−1
λq+1 1−α
λq+1 λq+1

as long as we take m ≥ 8.
To get an equation for Rq+1 we recall that v` satisfies (9.67):

∂t v` + ∇ · [v` ⊗ v` ] + ∇p` = ∇ · (R` + Rcomm ), (9.145)

∇ · v` = 0,

with
Rcomm = v` ⊗tr v` − [(v ⊗tr v) ?x φ` ] ?t ϕ` . (9.146)

132
Hence, vq+1 = v` + wq+1 satisfies

∂t vq+1 + ∇ · (vq+1 ⊗ vq+1 ) = ∂t v` + ∇ · (v` ⊗ v` ) + ∂t wq+1 + ∇ · (wq+1 ⊗ wq+1 )

+∇ · (v` ⊗ wq+1 ) + ∇ · (wq+1 ⊗ v` ) = ∇ · (R` + Rcomm ) − ∇p` + ∂t wq+1 + v` · ∇wq+1
+wq+1 · ∇v` + ∇ · (wq+1 ⊗ wq+1 ), (9.147)

so that the Reynolds stress Rq+1 and pressure pq+1 satisfy, after absorbing p` into pq+1

∇·Rq+1 = ∇pq+1 +∇·(R` +Rcomm )+∂t wq+1 +v` ·∇wq+1 +wq+1 ·∇v` +∇·(wq+1 ⊗wq+1 ). (9.148)

We write
(p) (c)
wq+1 = wq+1 + wq+1 ,
and represent the right side of (9.148) as
(c)
∇ · Rq+1 = Etr + Eosc + EN ash + Ecorr + ∇ · Rcomm + ∇ · Rcorr + ∇pq+1 , (9.149)

with the transport error

(p) (p) (p)
Etr = ∂t wq+1 + v` · ∇wq+1 − h∂t wq+1 i, (9.150)

the oscillation error

(p) (p)
Eosc = ∇ · (wq+1 ⊗ wq+1 + Rq ), (9.151)
and the Nash error
(p)
EN ash = wq+1 · ∇v` , (9.152)
coming from the principal part of the perturbation, and the corrector error
(c) (c) (c)
Ecorr = ∂t wq+1 + v` · ∇wq+1 − h∂t wq+1 i, (9.153)

and the tensor

(c) (c) (c) (p) (c) (c) (p)
Rcorr = wq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 (9.154)

coming from the incompressibility correction to the perturbation. Note that in the definition
(c)
of Rcorr we have replaced the tensor products ⊗ by the trace-free tensor products ⊗tr , with
the difference going into the pressure pq+1 . The notation h·i refers to the spatial average, as
before: ˆ
hf i = f (y)dy.
T3
As wq+1 is a curl, its spatial average vanishes, hence
(p) (c)
h∂t wq+1 i + h∂t wq+1 i = 0, (9.155)

so that the addition of these two terms to (9.150) and (9.153) does not change anything. Then
we can write
(c)
Rq+1 = Rtr + RN ash + Rcomm + Rcorr + Rcorr + Rosc , (9.156)

133
 (c)
with Rcomm and Rcorr defined in (9.146) and (9.154), respectively, and Rtr , RN ash , Rcomm
and Rcorr given by (9.140) with the corresponding E in the right side.
The term Rosc in (9.156) should be a trace-less symmetric solution to

∇ · Rosc = Eosc + ∇posc , (9.157)

with Eosc given by (9.151) and some pressure posc that we will absorb into pq+1 . We can
re-write Eosc as
X X 
(p) (p)
Eosc = ∇ · (wq+1 ⊗ wq+1 + R` ) = ∇ · w(ξ) ⊗ wξ0 + R` . (9.158)
j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0

Note that w(ξ) and w(ξ0 ) have disjoint support in time if ξ ∈ Λj and ξ 0 ∈ Λj 0 with |j − j 0 | > 1.
In addition, if |j − j 0 | = 1, then Λj and Λj 0 are disjoint sets so that ξ + ξ 0 6= 0 – this is why
we took Λ0 and Λ1 as two different sets. Hence, the only terms in the sum in (9.158) that
satisfy ξ + ξ 0 = 0 are those with j = j 0 . Thus, we have
X X  X X 
Eosc = ∇ · w(ξ) ⊗ w(−ξ) + R` + ∇ · w(ξ) ⊗ wξ0 . (9.159)
j ξ∈Λj j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0 ,ξ+ξ 0 6=0

We claim that the divergence of the first sum in (9.159) actually vanishes – and that is the
reason we have chosen the coefficients aq+1,j,ξ in the way we did. Indeed, recall that
1/4 1/2 (j)
aq+1,j,ξ (t, x) = cR δq+1 χj (t)γξ (M` (t, x)), (9.160)
(j)
with the coefficients γξ defined so that

1 X (j)
M` (t, x) = (γ (M` (t, x)))2 (Id − ξ ⊗ ξ), (9.161)
2 ξ∈Λ ξ
j

for each j, where

−1/2 −1
M` (t, x) = Id − cR δq+1 R` (t, x), (9.162)
which implies

1/2 1 X 1/2 (j)

cR δq+1 Id − R` (t, x) = c δq+1 (γξ (Mq (t, x)))2 (Id − ξ ⊗ ξ), (9.163)
2 ξ∈Λ R
j

again for each j. Multiplying (9.163) by χ2j (t) and summing over j, using (9.99), we arrive at

1/2 1XX 2 1/2 (j)

cR δq+1 Id − R` (t, x) = χj (t)cR δq+1 (γξ (Mq (t, x)))2 (Id − ξ ⊗ ξ)
2 j ξ∈Λ
j

1 X X
= |aq+1,j,ξ (t, x)|2 (Id − ξ ⊗ ξ). (9.164)
2 j ξ∈Λ
j

134
On the other hand, as in (9.89), we have, since aq+1,j,−ξ = aq+1,j,ξ , that
X X
w(ξ) ⊗ w(−ξ) = |aq+1,j,ξ |2 Bξ ⊗ B−ξ (9.165)
ξ∈Λj ξ∈Λj
1X
= |aq+1,j,ξ |2 (Aξ + iξ × Aξ ) ⊗ (Aξ − iξ × Aξ )
2 ξ∈Λ
j

1X i
= |aq+1,j,ξ |2 (Aξ ⊗ Aξ + (ξ × Aξ ) ⊗ (ξ × Aξ )) + [(ξ × Aξ ) ⊗ Aξ − Aξ ⊗ (ξ × Aξ )]
2 ξ∈Λ 2
j

1X
= |aq+1,j,ξ |2 (Id − ξ ⊗ ξ).
2 ξ∈Λ
j

Since the set Λj is symmetric: Λj = −Λj , the second term in the third line above vanishes
after summation over ξ ∈ Λj , and for the first term in that line we used (9.87):

ξ ⊗ ξ + Aξ ⊗ Aξ + (ξ × Aξ ) ⊗ (ξ × Aξ ) = Id. (9.166)

We deduce from (9.164) and (9.165) that

X 
∇· w(ξ) ⊗ w(−ξ) + R` = 0, (9.167)
ξ∈Λj

as we have claimed. Recall also that for a scalar-valued function g(x) and a matrix-valued
function F (x) we have

[∇ · (g(x)F (x))]i = ∂j (g(x)Fji (x)) = (∂j g(x))Fji (x) + g(x)(∂j (Fij (x))
= (F t (x)∇g(x))i + g(x)(∇ · F (x))i ,

so that
∇ · (gF ) = F t ∇g + g∇ · F.
Hence, Eosc has the form
X X 
Eosc = ∇ · w(ξ) ⊗ wξ0 (9.168)
j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0 ,ξ+ξ 0 6=0

1X X
= bq+1,j,ξ bq+1,j,ξ0 ∇ · (Wξ,λq+1 ⊗ Wξ0 ,λq+1 + Wξ,λq+1 ⊗ Wξ0 ,λq+1 )
2 j,j 0 ξ∈Λ ,ξ0 ∈Λ ,ξ+ξ0 6=0
j j0
X X
+ (Wξ0 ,λq+1 ⊗ Wξ,λq+1 )∇(bq+1,j,ξ bq+1,j,ξ0 ),
j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0 ,ξ+ξ 0 6=0

with bq+1,j,ξ as in (9.109) and (9.110). In addition, as in the derivation of the Euler equation
 |W |2 
ξ,λ
∇ · (Wξ,λ ⊗ Wξ,λ ) = ∇ ,
2
we also have fill this in

∇ · (Wξ,λ ⊗ Wξ0 ,λ + Wξ,λ ⊗ Wξ0 ,λ ) = ∇(Wξ,λ · Wξ0 ,λ ). (9.169)

135
Therefore, (9.168) becomes
1X X
Eosc = bq+1,j,ξ bq+1,j,ξ0 ∇(Wξ,λq+1 · Wξ0 ,λq+1 )
2 j,j 0 ξ∈Λ ,ξ0 ∈Λ ,ξ+ξ0 6=0
j j0
X X
+ (Wξ0 ,λq+1 ⊗ Wξ,λq+1 )∇(bq+1,j,ξ bq+1,j,ξ0 )
j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0 ,ξ+ξ 0 6=0

1X X h i
= ∇ bq+1,j,ξ bq+1,j,ξ0 (Wξ,λq+1 · Wξ0 ,λq+1 ) (9.170)
2 j,j 0 ξ∈Λ 0 0
j ,ξ ∈Λj 0 ,ξ+ξ 6=0
X X h 1 i
+ Wξ0 ,λq+1 ⊗ Wξ,λq+1 − (Wξ,λq+1 · Wξ0 ,λq+1 )Id ∇(bq+1,j,ξ bq+1,j,ξ0 )
j,j 0 ξ∈Λj ,ξ 0 ∈Λj 0 ,ξ+ξ 0 6=0
2

= ∇posc + Ẽosc (9.171)

The first term in the right side can be incorporated into pressure, so that we can define Rosc
as the solution to
∇ · Rosc = Ẽosc , (9.172)
given by (9.140) with E = Ẽosc . Summarizing, and recalling (9.156), we have the following
expression for Rq+1 :
(c)
Rq+1 = Rtr + RN ash + Rcomm + Rcorr + Rcorr + Rosc , (9.173)
(c)
with Rcomm and Rcorr defined in (9.146) and (9.154), respectively, and the individual contri-
butions Rtr , RN ash , Rcomm , Rcorr and Rosc given by (9.140) with the corresponding E in the
right side.

9.5.7 The inductive estimates on the new Reynolds stress

Now we estimate each individual term in the right side of (9.173).

The transport error

Recall that the transport error is given by (9.150):
(p) (p) (p)
Etr = ∂t wq+1 + v` · ∇wq+1 − h∂t wq+1 i. (9.174)

The last term in the right side does not contribute to (9.140) and only serve to ensure
that hEtr i = 0. In addition, we have

∂t Wξ,λ (Φj ) + vq · ∇Wξ,λ (Φj ) = 0, (9.175)

because Φj is advected by vq : it satisfies (9.95). It follows that

XX
Etr = (∂t aq+1,ξ,j (t, x) + vq · ∇aq+1,ξ,j (t, x))Wξ,λq+1 (Φj (t, x)) (9.176)
j ξ∈Λj

136
As we have seen many times, the standard mollification estimates on the derivatives of R` in
terms of kR` kC 0 , imply the bounds
1/2 1/2
kaq+1,λq+1 ,ξ kC m ≤ Cδq+1 `−m , k∂t aq+1,ξ,j + vq · ∇aq+1,ξ,j kC m ≤ Cδq+1 `−m−1 . (9.177)

Thus, we are in the situation as in (9.142)-(9.144), with C = C 0 δ 1/2 `−1 in (9.144), which gives

1/2 1/2 α−1/4 cR δq+2

kRtr kC 0 ≤ kRtr kC α ≤ Cδq+1 `−1 λα−1
q+1 = Cδq+1 λq+1 ≤ , (9.178)
100
provided that α and β are sufficiently small.

The oscillation error

The estimate for the oscillation error is similar. First, we note that
 
|∇(bq+1,j,ξ bq+1,j,ξ0 )| ≤ |∇(aq+1,j,ξ aq+1,j,ξ0 )| + λq+1 δq+1 |∇Φj − Id| + |∇Φj 0 − Id|
≤ Cδq+1 `−1 + Cλq+1 δq+1 `λq δq1/2 . (9.179)

A very similar argument, using (9.98) yields

k∇(bq+1,j,ξ bq+1,j,ξ0 )kC m ≤ Cδq+1 `−m−1 + Cλq+1 δq+1 `λq δq1/2 `−m . (9.180)

Hence, we can use (9.144) (strictly speaking, we are using its analog for the case when the
right side of (9.138) has the form of a tensor product of two right sides as in (9.142) but the
same argument applies) with

C = C 0 [δq+1 `−1 + λq+1 δq+1 `λq δq1/2 ],

which gives
cR δq+2
kRosc kC 0 ≤ kRtr kC α ≤ C[δq+1 `−1 + λq+1 δq+1 `λq δq1/2 ]λα−1
q+1 ≤ , (9.181)
100
provided that α and β are sufficiently small.

The Nash error

The Nash error comes from (9.152):
(p)
XX
EN ash = wq+1 · ∇v` = aq+1,j,ξ Wξ,λq+1 (Φj ) · ∇v` , (9.182)
j ξ∈Λj

so it is again of the form (9.142) and we can appeal to (9.144). The estimate
1/2
kaq+1,j,ξ ∇v` kC n ≤ Cδq+1 λq δq1/2 `−n (9.183)

then leads to
1/2 cR δq+2
kRN ash kC 0 ≤ kRN ash kC α ≤ Cδq+1 λq δq1/2 λα−1
q+1 ≤ . (9.184)
100

137
The corrector error
The corrector error has two components:
(c)
Rcorr + Rcorr . (9.185)
(c)
Here, Rcorr is given by (9.154):
(c) (c) (c) (p) (c) (c) (p)
Rcorr = wq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 , (9.186)
and Rcorr is given by (9.140) with E = Ecorr , which is defined in (9.153)
(c) (c) (c)
Ecorr = ∂t wq+1 + v` · ∇wq+1 − h∂t wq+1 i. (9.187)
The first term is estimated using the estimates (9.108)
1/2
(p) δq+1
kwq+1 (t, x)kC 0 ≤ , (9.188)
2
and (9.115)-(9.117) which say that
h k∇a
q+1,j,ξ kC 0
i
(c)
kwq+1 kC 0 ≤ K sup sup + kaq+1,j,ξ kC 0 k∇Φj − IdkC 0 (9.189)
j ξ∈Λj λq+1
1/2 1/2
δq+1 1/2 δq+1
≤K + Kδq+1 λ−1/2
q δq1/2 ≤ K 0 . (9.190)
`λq+1 `λq+1
(c)
This allows us to estimate Rcorr simply as
(c) (c) (c) (p) (c) (c) (p)
kRcorr kC 0 = kwq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 + wq+1 ⊗tr wq+1 kC 0
cR δq+2
≤ Cδq+1 `−1 λ−1
q+1 ≤ . (9.191)
100
As for Rcorr , note that, once again, because Wξ,λq+1 (Φj ) solves the transport equation (9.175),
we have
XX  ∇a 
q+1,j,ξ
Ecorr = (∂t + v` · ∇) + iaq+1,j,ξ (∇Φj − Id)ξ Wξ,λq+1 (Φj (t, x)). (9.192)
j ξ∈Λ
λq+1
j

We know that
1 1/2 1/2
k(∂t + v` · ∇)∇aq+1,j,ξ kC n ≤ Kδq+1 λ−1
q+1 `
−2−n
= Kδq+1 λq `−n , (9.193)
λq+1
and
1/2 1/2
k(∂t + v` · ∇)[aq+1,j,ξ (∇Φj − Id)kC n ≤ Kδq+1 [`−n−1 `λq δq1/2 + `1−n λq δq1/2 ] ≤ K 0 δq+1 λq δq1/2 `−n .
(9.194)
Appealing to (9.144) one more time, we obtain
1/2 cR δq+2
kRcorr kC 0 ≤ Cδq+1 λq λα−1
q+1 ≤ , (9.195)
100
if α and β are sufficiently small. This was the last estimate we needed to prove that
cR δq+2
kRcorr kC 0 ≤ , (9.196)
2
and we are done. This completes the proof of Proposition 9.3 and hence that of Theorem 9.2
as well.

138