Lecture notes in fluid mechanics

Laurent Schoeffel, CEA Saclay

These lecture notes have been prepared as a first course in fluid mechanics up to the presentation of the millennium
problem listed by the Clay Mathematical Institute. Only a good knowledge of classical Newtonian mechanics is assumed.
We start the course with an elementary derivation of the equations of ideal fluid mechanics and end up with a discussion
of the millennium problem related to real fluids. With this document, our primary goal is to debunk this beautiful problem as
much as possible, without assuming any previous knowledge neither in fluid mechanics of real fluids nor in the mathematical
formalism of Sobolev’s inequalities. All these items are introduced progressively through the document with a linear increase in
the difficulty. Some rigorous proofs of important partial results concerning the millennium problem are presented. At the
end, a very modern aspect of fluid mechanics is covered concerning the subtle issue of its application to high energetic
hadronic collisions.

1
§1. Introduction
§2. Continuum hypothesis
§3. Mathematical functions that define the fluid state
§4. Limits of the continuum hypothesis
§5. Closed set of equations for ideal fluids
§6. Boundary conditions for ideal fluids
§7. Introduction to nonlinear differential equations
§8. Euler’s equations for incompressible ideal fluids
§9. Potential flows for ideal fluids
§10. Real fluids and Navier-Stokes equations
§11. Boundary conditions for real fluids
§12. Reynolds number and related properties
§13. The millennium problem of the Clay Institute
§14. Bounds and partial proofs
§15. Fluid mechanics in relativistic Heavy-Ions collisions

2
§1. Introduction
Fluid mechanics concerns the study of the motion of fluids (in general liquids and gases) and the forces
acting on them. Like any mathematical model of the real world, fluid mechanics makes some basic
assumptions about the materials being studied. These assumptions are turned into equations that must
be satisfied if the assumptions are to be held true. Modern fluid mechanics, in a well-posed
mathematical form, was first formulated in 1755 by Euler for ideal fluids.

Interestingly, it can be shown that the laws of fluid mechanics cover more materials than standard
liquid and gases. Indeed, the idea of exploiting the laws of ideal fluid mechanics to describe the
expansion of the strongly interacting nuclear matter that is formed in high energetic hadronic
collisions was proposed in 1953 by Landau. This theory has been developed extensively in the last
60 years and is still an active field of research. This gives a very simple 3-steps picture of a non-trivial
phenomenon observed in hot nuclear matter after the collision of high energetic heavy ions, composed of a
large collection of charged particles.

(i) After the collision a nuclear medium, a zone of high

density of charges, is formed with high pressure in
the middle (center of the collision).

(ii) According to the laws of fluid mechanics, as we shall

prove them, this implies that an acceleration field is
generated from high pressures to low
pressures.

3
 (iii) This implies that particles will flow in a certain
transverse direction, as indicated on the figure. This is
known as the transverse flow property,
well established experimentally.

We come back on these ideas and their developments in the last section of this document. It requires a
relativistic formulation of fluid mechanics. Up to this section, we always assume that the dynamics is
non-relativistic.

§2. Continuumhypothesis

Fluid mechanics is supposed to describe motion of fluids and related phenomena at

macroscopic scales, which assumes that a fluid can be regarded as a continuous medium. This means that
any small volume element in the fluid is always supposed so large that it still contains a very great
number of molecules. Accordingly, when we consider infinitely small elements of volume, we mean
very small compared with the volume of the body under consideration, but large compared with
the distances between the molecules. The expressions fluid particle and point in a fluid are to be
understood in this sense. That is, properties such as density, pressure, temperature, and velocity are
taken to be well-defined at infinitely small points.

These properties are then assumed to vary continuously and smoothly from one point to another.
Consequently, the fact that the fluid is made up of discrete molecules is ignored. If, for example, we deal
with the displacement of some fluid particle, we do mean not the displacement of an individual
molecule, but that of a volume element containing many molecules, though still regarded as a point in
space. That’s why fluid mechanics is a branch of continuum mechanics, which models matter from a
macroscopic viewpoint without using the information that it is made out of molecules (microscopic
viewpoint).

4
§3. Mathematical functions that define the fluid state
Following the continuous assumption, the mathematical description of the state of a moving fluid can be
characterized by functions of the coordinates x, y, z and of the time t. These functions of (x, y, z, t) are
related to the quantities defined for the fluid at a given point (x, y,
z) in space and at a given time t, which refers to fixed points in space and not to fixed particles of the fluid.
For example, we can consider the mean local velocity v(x, y, z, t) of fluid particles or fluid points, called the
fluid velocity, and two thermodynamic quantities that characterize the fluid state, the pressure p(x, y, z, t)
and the mass density (x, y, z, t), the mass per unit volume of fluid. Following the discussion of §2, two
remarks can be done at this stage:
i. The fluid is assumed to be a continuum. This implies that all space-time derivatives of all
dependent variables exist in some reasonable sense. In other words, local properties
such as density pressure and velocity are defined as averages over elements large
compared with the microscopic structure of the fluid but small enough in comparison with the
scale of the macroscopic. This allows the use of differential calculus to describe such a
system.
ii. All the thermodynamic quantities are determined by the values of any two of them, together with
the equation of state. Therefore, if we are given five quantities, namely the three components of
the velocity v, the pressure p and the mass density , the state of the moving fluid is completely
determined. We recall that only if the fluid is close to thermodynamic equilibrium, its
thermodynamic properties, such as pressure, density, temperature are well-defined. This
requires (as a very former hypothesis) that local relaxation times towards thermal equilibrium are
much shorter than any macroscopic dynamical time scale. In particular, microscopic
collision time scale (between elementary constituents of the fluid) needs to be much shorter
than any macroscopic evolution time scales. This hypothesis is almost a tautology for standard
fluids build up by molecules at reasonable density, but becomes not trivial in the case of some hot
dense matter state created in high energetic hadronic collisions.

5
In the following, we prove that these five unknown quantities describe completely the case of what we
define as ideal fluids, in which we take no account of processes of energy dissipation. Energy
dissipation may occur in a moving fluid as a consequence of internal friction (or viscosity) within the
fluid and heat exchange between different parts of it. Neglecting this phenomenon, we can find a set
of five equations that are sufficient obtain a closed system: 5 equations for 5 unknown quantities.
Interestingly, we can gain some intuition about the behavior of the ideal flow by expressing in more
details its pressure field. An ideal fluid, in particular, is characterized by the assumption that each particle
pushes its neighbors equally in every direction. This is why a single scalar quantity, the pressure, is
sufficient to describe the force per unit area that a particle exerts on all its neighbors at a given time.
Also, we know that a fluid particle is not accelerated if its neighbors push back with equal force, which
means that the acceleration of the fluid particle results from the pressure differences. In short, the
pressure force can be seen as a global interaction of all fluid particles.

When the energy dissipation inside the fluid is not neglected, we need to consider also the internal
energy density e(x, y, z, t) and heat flux density q(x, y, z, t) as four additional unknown functions to
be determined by a proper set of closed equations: nine equations are needed in such cases. This
concerns what we define as real fluids. We discuss the case of real fluids in more details later in the
document. However, a few intuitive arguments can be made with no mathematical formalism. When the
energy dissipation is not neglected, this means that we take into account frictional forces inside the
fluid. Their main effect is that they enhance the local coherence of the flow. They counteract at each
point the deviation of the velocity field from its local average. This means that if a fluid particle moves
faster than the average of its neighbors, then friction slows it down.

When fluid quantities are defined at given fixed points (x, y, z) in space and at a given time t, we speak of
the Eulerian description of the fluid. When fluid quantities are defined as associated to a (moving)
particle of fluid, followed along its trajectory, we speak of the

6
Lagrangian point of view. An important notion can be derived from this last view point: the flow map.

A fluid particle moving through the fluid volume is labeled by the (vector) variable X (defined at t=0). At
some initial time, we define a subset of fluid particles (of the entire fluid volume) Ω0. The fluid particles of
this subset will move through the fluid within time. We introduce a function 𝚽(𝚽, t) that describes the
change of the particle position from the initial time up to t > 0. The function 𝚽(. , t) is itself a vector,
with 3 coordinates corresponding to the 3 coordinates of space (Φx, Φy, Φz). This means that we can
denote the position of any particle of fluid at time t by 𝚽(𝚽, t) which starts at the position X at t=0. Then,
we can label the new subset of fluid particles at time t, originally localized in Ω0, as:

Ωt = {𝚽(𝚽, t) ∶ 𝚽 belongs to Ω0 }.

Stated otherwise, the function t → 𝚽(𝚽, t), represents the trajectory of particles: this is what we also call
the flow map. In particular, the particle velocity is given by:

𝚽𝚽
𝚽(x, y, z, t) = (𝚽, t) with 𝚽(𝚽, t = 0) = 𝚽.
𝚽t
We can think of Ωt as the volume moving with the fluid.

7
§4. Limits of the continuum hypothesis
According to the continuous assumption (§2), the physical quantities (like velocity, pressure and density)
are supposed to vary smoothly on macroscopic scales. However, this may not be the case everywhere
in the flow. For example, if a shock front of the density appears at some values of the coordinates at a
given time, the flow would vary very rapidly at that point, over a length of order the collision mean free
path of the molecules. In this case, the continuum approximation would be only piecewise valid and
we would need to perform a matching at the shock front. Also, if we are interesting by scale invariant
properties of fluid in some particular cases, we need to keep in mind that there is a scale at which the
equations of fluid mechanics break up, which is the molecular scale characterized by the mean free path of
molecules between collisions. For example, for flows where spatial scales are not larger than the mean
distance between the fluid molecules, as for example the case of highly rarefied gazes, the continuum
assumption does not apply.

§5. Closed set of equations for ideal fluids

The derivation of equations underlying the dynamics of ideal fluids is based on three
conservation principles:
i. Conservation of matter. Matter is neither created nor destroyed provided there is no source or
sink of matter;
ii. Newton’s second law or balance of momentum. For a fluid particle, the rate of change of
momentum equals the force applied to it;
iii. Conservation of energy.

In turn, these principles generate the five equations we need to describe the motion of an ideal fluid: (i)
Continuity equation, which governs how the density of the fluid evolves locally and thus indicates
compressibility properties of the fluid; (ii) Euler’s equations of motion for a fluid element which indicates
how this fluid element moves from regions of high pressure to those of low pressure; (iii) Equation of
state which indicates the mechanism of energy exchange inside the fluid.

8
We derive first the expression for the conservation of mass (i). Consider a fluid of mass density ,
fluid particle velocity v and some volume , fixed in space (i.e. fixed in some Newtonian reference
frame). The mass of the fluid in this volume is∫ ρdV, where the integration is taken on the volume
. If the fluid moves, then there is a flow of mass across each element of surface d𝚽 on the boundary of
the volume , where the magnitude of the vector d𝚽 is equal to the area of the surface element and its
direction is along the normal to the surface. Provided that there are no sources or sinks of fluid, the
elementary mass of fluid flowing in unit time through an element d𝚽 of the surface bounding  is ρ𝚽. d𝚽.
By convention d𝚽 is taken along the outward normal, which means that ρ𝚽. d𝚽 is positive when the
fluid is flowing out of the volume and negative for a flow into the volume .

The total mass of fluid flowing out of the volume in unit time is thus: ∮ ρ𝚽. d𝚽, where the integration is
taken over the whole of the closed surface bounding . Therefore, we can write:

𝚽
∫ ρdV = − ∮ ρ𝚽. d𝚽,
𝚽t

∫ is the net decrease of the mass of fluid in  per unit time. Using the Green’s
where − 𝚽 ρdV
𝚽t

formula to express ∮ ρ𝚽. d𝚽 as a volume integral over : ∫ div(ρ𝚽)dV. We obtain:

𝚽ρ
∫[ + div(ρ𝚽)]dV = 0.
𝚽t

Since this equation must hold for any volume, the integrand must vanish. This gives:

𝚽ρ
+ div(ρ𝚽) = 0.
𝚽t

(1)

This is the continuity equation, the first fundamental equation of fluid mechanics. The vector
𝚽 = ρ𝚽 is called the mass flux density. Its direction is along the motion of the fluid and its magnitude
equals the mass of the fluid flowing per unit time though a unit area perpendicular to the velocity of the
fluid.

9
By expanding div(ρ𝚽) as (𝚽. 𝚽𝚽𝚽𝚽)(ρ) + ρdiv(𝚽), we can also write this equation as:

𝚽ρ 𝚽
+ (𝚽. =[ + (𝚽. (ρ) = −ρdiv( 𝚽).
𝚽t 𝚽t
𝚽𝚽𝚽𝚽)(ρ) 𝚽𝚽𝚽𝚽)]

(1′)

We identify the operator [𝚽 + (𝚽. 𝚽𝚽𝚽𝚽)] that we define below as the material derivative, or
𝚽t

the derivative following the flow.

Before developing some consequences of the continuity equation, we establish the Newton’s second law
(ii) for some volume  of the fluid, assumed to be ideal, characterized by its mass density, pressure and
fluid particle velocity. In the absence of external force, by definition of the pressure, the total force acting
on the ideal fluid in volume  is equal to: − ∮ pd𝚽.

This last formula represents the integral of the pressure taken over the surface bounding the volume, with
similar conventions as previously defined for the surface element d𝚽. This surface integral can be
transformed to a volume integral over : − ∮ pd𝚽 = − ∫ 𝚽𝚽𝚽𝚽(p)dV.

Thus, the fluid surrounding any elementary volume dV exerts a force −dV 𝚽𝚽𝚽𝚽(p) on that element.
Moreover, from Newton’s second law applied to this elementary volume dV, the
mass times the acceleration equals: ρdV d𝚽.
dt

Then, the Newton’s second law of motion for the fluid per unit volume reads:

d𝚽
ρ = −𝚽𝚽𝚽𝚽(p).
dt

(2)

Here, we need to be careful with the mathematical expression d𝚽/dt. It does not represent (only) the
rate of change of the fluid velocity at a fixed point in space, which would be mathematically
written as 𝚽𝚽/ ∂t.

10
Rather, d𝚽/dt is the rate of change of the velocity of a fluid particle as it moves in space (see
§2), called the material derivative, namely:

d𝚽 [𝚽(x + dx, y + dy, z + dz, t + dt) − 𝚽(x, y, z, t)]

=
.
dt dt

This can be developed as:

d𝚽 dx ∂𝚽 dy ∂𝚽 dz ∂𝚽 ∂𝚽 ∂𝚽
= + + + = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) +
.
dt dt 𝚽x dt 𝚽y dt 𝚽z 𝚽 ∂t
t

There are two terms in the expression of d𝚽 = d𝚽 . dt: the difference between the velocities of
dt

the fluid particle at the same instant in time at two points distant of (dx, dy, dz), which is the distance
moved by the fluid particle during dt, and the change during dt in the velocity at a fixed point in space (x,
y, z). Combining the vector equation (2) and the expression of the material derivative, we get:
1
∂𝚽 )( ) ( )
(
+ 𝚽. 𝚽𝚽𝚽𝚽 𝚽 = − 𝚽𝚽𝚽𝚽 p .
∂t ρ

(3)

This vector equation (3) represents a set of three equations (in three dimensions of space) that
describe the motion of an ideal fluid, first obtained by Euler in 1755. That’s why it is called the Euler’s
equations, the second fundamental set of equations of fluid mechanics (ii). It is trivial to expand the vector
equation (3) on the 3 Cartesian coordinates of space (x, y, z) as a set of 3 equations (in a compact
form):

∂ ∂ ∂ vx ∂p/ ∂x
( + vx + vy +vz ∂ ) [vy] = − 1 [∂p/ ∂y].
∂t ∂x ∂y ∂z vz ρ ∂p/ ∂z

If external forces have to be considered these equations become:

∂𝚽 1 1
+ (𝚽. 𝚽 = − 𝚽𝚽𝚽𝚽(p) + 𝚽ext.
) ρ ρ
𝚽𝚽𝚽𝚽)(
∂t
11
Here, 𝚽𝚽𝚽𝚽 can be for example the gravity force ρ𝚽.

Before continuing the derivation of fundamental equations of fluid mechanics, we give some hints on how
Euler’s and continuity equations can be derived using the notion of flow map (§4). As we have seen, the
flow map is a function 𝚽(𝚽, t) that describes the change of the fluid particle position X at initial time (t=0)
up to t > 0. Then, we have established the following relations (§4):
𝚽(x, y, z, t) = 𝚽𝚽 (𝚽, t) with 𝚽(𝚽, 0) = 𝚽.
𝚽t

This means that the acceleration of a fluid particle can be written as:
d d ∂𝚽 ∂vx ∂ϕx
𝚽(𝚽, t) = 𝚽(𝚽(𝚽, t), t) = +{ + ⋯ }.
dt dt 𝚽t 𝚽 𝚽t
x

We rewrite the last term as:

∂
∂vx ∂ϕx ∂ϕx ∂ ∂ϕy ∂ ∂ϕz v } = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
{ + ⋯} = { v + v +
𝚽x 𝚽 𝚽t 𝚽x x 𝚽 𝚽y y 𝚽 𝚽z z
t t t

We obtain:
d d ∂𝚽
𝚽(𝚽, t) = 𝚽(𝚽(𝚽, t), t) = + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
dt dt ∂t

We find again the derivative of the velocity following the flow that leads to Euler’s vector equations.
Similarly, we can use the notion of flow map to write the conservation of mass and then the continuity
equation. We consider fluid particles X initially localized in a subset Ω0 of the entire fluid volume. At time t >
0, they are contained in the subset (volume moving with the fluid):

Ωt = {𝚽(𝚽, t) ∶ 𝚽 belongs to Ω0 }.

Note that Ωt is not fixed in space (i.e. not fixed in some Newtonian reference frame), but moving with
the flow according to the flow map function.

12
The conservation of mass can then be written as:

∫ ρ(𝚽, t)dV = ∫ ρ(𝚽, 0)dV.

Ωt Ω0

Since the right-hand side is independent of the time, we can write:

d d
∫ ρ(𝚽, t)dV = ∫ ρ(𝚽(𝚽, t), t)dV = 0.
dt Ωt dt Ωt
In this expression, the time derivative represents the material derivative, following the movement
of fluid particles. The situation is not that easy a priori as the domain of integration depends on
the time in the above formula. It can be shown that the following relation holds:
d 𝚽ρ
∫ ρ(𝚽(𝚽, t), t)dV = ∫ [ + div(ρ𝚽)]dV.
dt Ωt Ωt 𝚽t

Then, this leads to the continuity equation.

We come back to the Euler’s equations (3). An important vector identity is the following:
1
𝚽𝚽𝚽𝚽(𝚽𝚽) = 𝚽 × 𝚽𝚽𝚽𝚽(𝚽) + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
2

Then, equations (3) can be rewritten as:

1
∂𝚽 1 ( 𝚽) ( ) ( )
+ 𝚽𝚽𝚽𝚽 − 𝚽 × 𝚽𝚽𝚽𝚽 =− 𝚽𝚽𝚽𝚽 p
∂t 2 ρ
𝚽
𝚽

This expression of equations (3) has many interests that we shall see later. For example, in the case of
constant mass density, taking the curl of this equation makes the gradients vanishing and we obtain
a differential equation involving only the velocity field. Another interesting physical case appears
when 𝚽𝚽𝚽𝚽(𝚽) = 0.

13
Then the velocity field can be written as the gradient of a scalar function and the above expression
leads to an interesting simple equation.

In this section we have ignored all processes related to energy dissipation, which may occur in a moving
fluid as a consequence of internal friction (or viscosity) in the fluid as well as heat exchange between
different parts of the fluid. Thus we have treated only the case of ideal fluids, for which thermal
conductivity and viscosity can be neglected. With the continuity equation, the Euler’s equations make a
set of 4 equations, for five quantities that characterize the ideal fluid (§2).

This means that we are missing one equation, which is coming with the last conservation principle,
namely the conservation of energy (iii). The absence of heat exchange between the different parts of the
fluid implies that the motion is adiabatic: thus the motion of an ideal fluid is by definition considered as
adiabatic. In other words, the entropy of any fluid particle remains constant as that particle moves in
space inside the fluid. We label the entropy per unit mass as s. We can easily write the condition for
an adiabatic motion as:

ds/dt = 0.

This represents the rate of change of the entropy (per unit mass) of a fluid particle as it moves in the fluid.
This can be reformulated as:

𝚽s
+ (𝚽. 𝚽𝚽𝚽𝚽)(s) = 0.
𝚽t

(4)

This expression is the general condition for adiabatic motion of an ideal fluid. This condition usually
takes a much simpler form. Indeed, as it usually happens, the entropy is constant throughout some
volume element of the fluid at some initial time, then it retains the same constant value everywhere in
the fluid volume, at all times for any subsequent motion of the fluid. In this case, equation (4) becomes
simply: s=constant.

14
Such a motion is said to be isentropic, which is what we assume in general for an ideal fluid, unless
stated otherwise. This condition, together with an equation of state for the fluid provides a relation
between the pressure and the mass density, and then the fifth equation: p=p(,s).

This allows us to know what happens to the density when pressure changes and
consequently to close the system of equations describing the mechanics of ideal fluids: 5 equations
for 5 variables (2 thermodynamic variables and the 3 coordinates of the velocity).

We rewrite the Euler’s equations (3) in case of steady flow, for which the velocity is constant in time at
any point occupied by the fluid:

∂𝚽
= 0.
∂t

This means that the velocity field is a function only of the coordinates. Taking ρ = constant for simplicity,
we obtain:

1
𝚽𝚽𝚽𝚽(𝚽𝚽) − 𝚽 × 𝚽𝚽𝚽𝚽(𝚽) = −𝚽𝚽𝚽𝚽(p/ρ)
2

(3′)

Then, we define streamlines: the tangent to a streamline at any point gives the direction of the velocity at
that point. Streamlines are thus defined by the set of equations:

dx dy dz
= = .
vx vy vz

One interest of steady flow is that the streamlines do not vary with time and thus coincide with the
paths of the fluid particles. Obviously, this coincidence between streamlines and trajectory of fluid
particles does not hold in non-steady flow: indeed, the tangents to the streamlines give the
directions of the velocities of fluid particles as a function of the coordinates in space at a given
instant, whereas the tangents to the path (trajectory) of a given fluid particle provide the direction of
the velocities as a function of time.

15
It follows that the quantity 𝚽𝚽/2 + P/ρ is constant along a streamline, which coincides with the particle
trajectory for steady flow. In general this constant is different with different streamlines: this is what
is called the Bernoulli’s equation.

Ideal fluids present an interesting property: mass and momentum conservation principles are
uncoupled from energy conservation. Indeed, if we consider the entropy to be constant throughout the
fluid, it is not required to consider explicitly the energy conservation to describe the motion of the
fluid. We can show that the relation corresponding to energy conservation is a consequence of the
continuity and Euler’s equations under the condition of isentropic flow. We consider some volume
element  of the fluid, fixed in space, and we find how the energy of the fluid contained in this element
varies with time. In the absence of external force, the energy density, per unit volume of the fluid,
can be written as:

1
U = ρ𝚽2 + ρϵ.
2

The first term is the kinetic energy density and the second the internal energy density, noting
𝚽 the internal energy per unit mass.

The change in time of the energy contained in the volume element  is then given by the partial

derivative with respect to time ∂∫UdV, where the integration is taken over . Then,
∂t

following a similar reasoning as for the continuity equation, we can write a general expression
of energy conservation for the fluid in that volume element, namely:

∂ ∫ UdV
=− ∮ 𝚽. d𝚽 = − ∫ div(𝚽)dV,
∂t

where 𝚽 represents the energy flux density. We let as an exercise (below) to show that 𝚽 is not equal to
U𝚽, but:

1 P 1
𝚽 = ρ𝚽 ( 𝚽 + ϵ + ) = ρ𝚽 ( 𝚽 + h).
22 ρ 22

16
With the notation h = 𝚽 + P/ρ, which corresponds to the enthalpy of the fluid per unit mass. The proof
uses only the continuity and Euler’s equations. Since the equation of energy conservation must
hold for any volume element, the integrand must vanish in:

∂U
∫ [ ∂t + div(𝚽)]dV = 0.

We end up with the local expression of energy conservation:

1
∂ρ( 𝚽2+𝚽)
2
+ div (ρ𝚽(1 𝚽2 + h)) = 0.
∂t 2

The zero on the right hand side of this relation comes from the condition for adiabatic motion: ds/dt = 0,
which is a necessary condition for an ideal fluid (see above).

Stated differently, if ds/dt would be non-vanishing, this right hand side term would be necessarily
proportional to: ds/dt. In fact, in the presence of heat flow within the fluid, which
means that the fluid is not supposed to be ideal, the rate of heat density change reads: ρT ds,
dt

which leads to the general equation for non-vanishing ds/dt:

1
∂ρ( 𝚽2+𝚽) 1 ds
2
+ div (ρ𝚽( 𝚽2 + h)) = ρT .
∂t 2 dt

Exercise: For an ideal fluid. Prove that the energy flux density can be written as:
ρ𝚽(1 𝚽 2 + h) with h = 𝚽 + P/ρ. Then, derive the local expression of the conservation of
2

energy (per unit mass):

1
∂ρ( 𝚽2+ϵ) 1
2
+ div (ρ𝚽( 𝚽2 + h)) = 0.
∂t 2
1
∂ρ( 𝚽2+ϵ)
2
Solution:The idea is to compute the partial derivative using equations of fluid
∂t

mechanics that we have established together with a thermodynamic relation involving the
internal energy (per unit mass).

17
We write:
1 2)
∂ρ(2 𝚽 1 2 ∂ρ
∂
=𝚽
ρ𝚽
+ 𝚽 .
∂t ∂t 2 ∂t

In this identity, ∂ρ can be replaced by −div(ρ𝚽) using the continuity equation and ∂𝚽 is given
∂t ∂t

by the Euler’s equations. We obtain:

1 2 1 2
∂ρ(2 𝚽 ) ( ) ( )
[( ) ]
= ρ𝚽. 𝚽. 𝚽𝚽𝚽𝚽 − 𝚽. 𝚽𝚽𝚽𝚽 p − 𝚽 div ρ𝚽 .
∂t 2
𝚽
With the vector identity: 𝚽. [(𝚽. 𝚽𝚽𝚽𝚽)𝚽] = 1 𝚽. 𝚽𝚽𝚽𝚽(𝚽𝚽) and
2

𝚽𝚽𝚽𝚽(p) = ρ𝚽𝚽𝚽𝚽(h) − ρT𝚽𝚽𝚽𝚽(s) (since dh=Tds+1/ dp), we obtain:

1 1 2
∂(2 ρ𝚽2) 1𝚽 ( ) ( )
= −ρ𝚽. [𝚽𝚽𝚽𝚽( 𝚽 + h)] + ρT𝚽. 𝚽𝚽𝚽𝚽 s − 𝚽 div ρ𝚽 .
∂t 2 2

Also, using the thermodynamic relation: d𝚽=Tds+(p/²)d, we can write: d(ρ𝚽)=hd+Tds. Using the
adiabatic condition of motion, this leads to:
∂(ρϵ)
= −h. div(ρ𝚽) − ρT𝚽. 𝚽𝚽𝚽𝚽(s).
∂t
Combining the results, we find the expected result:
1
∂ρ(2 𝚽 2 + ϵ) 12
= −div (ρ𝚽 ( 𝚽2 + h)).
∂t
In integral form, it reads:
∂ 1 2 1 2

∫ ρ( 𝚽 + ϵ)dV = − ∮ ρ𝚽 ( 𝚽 + h) d𝚽 .
∂t 2 2
The left hand side is the rate of change of energy of the fluid in some given volume. The right hand side is
therefore the amount of energy flowing out of this volume per unit time. Hence,
ρ𝚽 (1 𝚽2 + h) is the energy flux density vector. Its magnitude is the amount of energy passing
18
2

19
per unit time through a unit area perpendicular to the direction of velocity. This means that
anyunit mass offluid carries with it duringits motionthe amount ofenergy 1𝚽2 + h
2

(andnot1𝚽2 +ϵ).
2

The fact that enthalpy appears and not internal energy simply comes from the relation:
1 1 2
− ∮ ρ𝚽 ( 𝚽 2+ h) d𝚽 = − ∮ ρ𝚽 ( 𝚽 + ϵ) d𝚽 − ∮ p𝚽d𝚽.
2 2
The first term is the total energy transported through the surface in unit time by the fluid and
thesecondtermisthework donebypressureforcesonthefluidwithinthesurface.

20
Summary of the three conservation principles for ideal fluids is provided in the table below:

(i) Conservation of matter Continuity equation:

𝚽ρ 𝚽
+ (𝚽. 𝚽𝚽𝚽𝚽)(ρ) = [ + (𝚽. 𝚽𝚽𝚽𝚽)] (ρ)
𝚽t 𝚽t
= −ρdiv(𝚽).
(ii) Balance of Momentum (Newton’s Euler’s (vector) equation:
second law) ∂𝚽 1 1
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = − 𝚽𝚽𝚽𝚽(p) + 𝚽 .
∂t ρ ρ ext

Useful identity:
1
𝚽𝚽𝚽𝚽(𝚽𝚽) = 𝚽 × 𝚽𝚽𝚽𝚽(𝚽) + (𝚽.
𝚽𝚽𝚽𝚽)(𝚽).
2
(iii) Conservation of energy or absence of heat Local form of the conservation of energy
exchange between the different parts of the fluid reads:
1
which implies that the motion is adiabatic ∂ρ( 𝚽2 + 𝚽) 1
2 + div (ρ𝚽( 𝚽2 + h)) = 0.
∂t 2

For ideal fluids, this is equivalent to the

condition:
ds 𝚽s
= + (𝚽. 𝚽𝚽𝚽𝚽)(s) = 0 or s = constant. dt
𝚽t
Together with an equation of state this
provides a relation of the form: p=p(,s).

21
Exercise: Momentum flux. We label the spatial component x, y, z of vectors by one index i.
Prove that there exists a quantity Πik that depends on two indices that verifies the relation:
∂
∂t
∫ρvidV = −∑k=x,y,z∮ ΠikdΣk foreachi=x, y, z.

Solution: It can easily be shown that: Πik = pδik + ρvivk, where δik = 1 if i = k and 0 otherwise.
The integral relation above can be interpreted as usual. The left hand side is the rate of change of the
component i of the momentum contained in the volume considered. The right hand side is therefore the
amount of momentum flowing out through the bonding surface per unit time.
Thus, Πik corresponds to the component i of the momentum flowing in unit time through unit area
perpendicular to the axis labeled by k. The energy flux is given by a vector (depending on one index), the
energy itself being a scalar. Here the momentum flux is given by a quantity
dependingon two indices, atensorof rank2, themomentum itself beingavector.

We consider now the kinetic energy contained in a volume Ωt defined by the flow map (Ωt is

moving with the fluid), namely EK = ∫ 1 ρ𝚽²dV.

Ωt 2
Note that Ωt is not fixed in space (i.e. not

fixed in some Newtonian reference frame), but moving with the flow according to the flow map
function. We are interested by the variation along the flow of the kinetic energy contained in this
volume EK. To get this information, we need to compute the (material) derivative dEK /dt, knowing
that EK is defined as an integral over the moving volume Ωt . As discussed previously, this is not obvious
to commute the derivative and the integral as the integral domain depends on time. In fact, it can be
shown that:

dEK d 1 d𝚽
= [∫ ρ𝚽²dV] = ∫ 𝚽ρ. dV .
dt dt Ωt 2 Ωt dt

22
 i. As a first case, we assume that all the energy is kinetic. This means that we consider the internal
energy as a constant which does not matter in the expression related to energy balance. The
principle of conservation of energy states that the rate of change in time of the kinetic energy in a
portion of fluid (following the flow) equals the rate at which the pressure forces work. For
simplicity, we neglect other external forces that may apply. Mathematically, this gives:
dEK d 1 d𝚽
= [∫ ρ𝚽²dV] = ∫ 𝚽ρ. dV = − ∮ p𝚽. d𝚽.
dt dt Ωt 2 Ωt dt St

The last integral is taken on a closed surface bounding the volume Ωt. Also, the

quantity − ∮ St p𝚽. d𝚽 equals − ∫Ωt div(p𝚽)dV = − ∫Ωt 𝚽. 𝚽𝚽𝚽𝚽(p)dV − ∫Ωt pdiv(𝚽)dV.

We can replace −𝚽𝚽𝚽𝚽(p) by ρ. d𝚽 in this identity using the Euler’s equations, which
𝚽𝚽

leads to:
dEK d𝚽 d𝚽
= ∫ 𝚽ρ. dV = ∫ 𝚽ρ. dV − ∫ pdiv(𝚽)dV.
dt Ωt dt Ωt dt Ωt

This equality can only be realized if div(𝚽) = 0, or (using the continuity equation) d/dt=0.
This corresponds to the condition of incompressible fluid (see §8).

ii. As a second case, we consider that the internal energy is not constant. Following the previous
discussion, we can write in this more general case:
d(EK + ρϵ)
= − ∮ p𝚽. d𝚽.
dt St

After developing the above expression asinthefirst case, we can easily show that this
leads to the relation ρ dϵ = p dρ or equivalently ρ² dϵ = p. This corresponds to a re-
dt ρ dt dρ

writing of the variation of enthalpy as: dh = dp/ρ, which is also the condition for an isentropic
motion.

23
§6. Boundary conditions for ideal fluids
The equations of motion have to be supplemented by the boundary conditions that must be satisfied at
the surfaces bounding the fluid. For an ideal fluid, the boundary condition is simply that the fluid
cannot penetrate a solid surface. This means that the component of the fluid velocity normal to the
bounding surface must vanish if that surface is at rest: 𝚽. 𝚽 = 0.

In the general case of a moving surface, 𝚽. 𝚽 must be equal to the corresponding component of the
velocity of the surface. At a boundary between two immiscible fluids, the condition is that the pressure
and the velocity component normal to the surface of separation must be the same for the two fluids,
and each of these velocity components must be equal to the corresponding component of the
velocity of the surface.

Given boundary conditions, the continuity and Euler’s equations, together with the relation for adiabatic
motion, established in §5 form a closed set of equations necessary to determine the 5 unknown quantities,
once initial conditions are assumed. Solving this problem means that, if we consider a moving fluid
contained in a volume  at any instant t (t the corresponding moving volume), at each point x of
t, we can find a well-defined solution for the 5 quantities to be determined. Such that also the
equations of fluid mechanics must contain non-diverging terms at all points of t. In particular, the total
kinetic energy density
integrated over the moving volume t, 1 ρ𝚽2∫dV, must remain finite, not diverging to infinity
2

at any time. Stated differently, once we define initial conditions, a fluid mechanics problem is solved
(globally) if we can show the existence and unicity of smooth solutions of equations (§5) for the velocity,
pressure and mass density, for all points in the fluid volume and at all times, with dedicated boundary
conditions. This is highly non trivial as we shall see next.

§7. Introduction to nonlinear differential equations

A central issue in the study of nonlinear differential equations, as the Euler’s equations, is that
solutions may exist locally in time (that is, for short periods of time) but not globally in time. This is
caused by a phenomenon called blow-up, illustrated in this section.

24
We first discuss this phenomenon with three simple ordinary differential equations for a real function
u(t):
du
(a) = u, (b) du = u², (c) du = −u + u².
dt dt dt

Equation (a) is linear and its solution is:

u(t) = Aet.
This solution is defined globally in time and grows exponentially as time becomes infinite. Generally,
global existence and exponential growth are typical features of linear differential equations. Equation
(b) is nonlinear. Its solution is:
u(t) = 1/(τ − t),
where τ is a parameter. The solution of (b) is then diverging when t is approaching τ. This example (b)
shows that nonlinearities which grow super-linearly in u(t) can lead to blow-up and a loss of global
existence. Equation (c) is also nonlinear and its solution is:
u(t) = 1/(1 − Aet).

If A ≤ 0, which corresponds to 0 < u(0) < 1, then the solution exists globally in time. If 0 < A < 1,
which corresponds u(0) > 1, then the solution blows up at t = log(1).
A

Thus, there is a global existence of solutions with small initial data and local existence (in time) of
solutions with large initial data. This type of behavior also occurs in many partial differential equations
of more general functions of space and time: for small initial data, linear damping terms can dominate
the nonlinear terms, and one obtains global solutions whereas for large initial data, the nonlinear blow-
up dominates and only local solutions may exist.

Consider now a less simple example of nonlinear partial differential equation. A real function
∂( ) u2
of x and t, u(x, t), is a solution of the equation: ∂u + 2
= 0, with the initial condition:
∂t ∂x

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

25
 (x)
We can show that this equation cannot have a global smooth solution if∂u0 < 0 at any point.
∂x

The proof is simple, based on the previous discussion. Suppose that u(x, t) is a smooth
∂( ) u2
solution. We take the x derivative of the partial differential equation: ∂u + 2
= 0.
∂t ∂x

We obtain:
∂ux
+ uuxx + u2 = 0.
∂t x

The subscript x represents a derivative with respect to x: u = ∂u. We

x can rewrite this
∂x

expression as:
[ ∂ + u ∂ ](ux) + u2 = 0.
∂t ∂x x

It is interesting to note as: d = [ ∂ + u ∂ ] the derivative operator appearing in the last

dt ∂t ∂x

formula (already seen in §5), which is the derivative along the characteristic curves associated
with the function u. Then, we obtain an equation close to equation (b) above:
dux
= −u2.
dt x

Therefore, if ux < 0 at initial time, the solution of this equation follows exactly what we have computed for
equation (b) and it blows up at some positive time. A global smooth solution cannot exist.

∂( ) u2
The interest of this (partial) differential equation: ∂u + 2
= ∂u + u∂(u) = 0 lies in the fact
∂t ∂x ∂t ∂x

that it is equivalent to a one dimensional Euler’s equation, u being the velocity field, with p=0 and in the
absence of external forces. Already with this simple form, we remark that a (unique) global smooth
solution may not exist in general. We discuss further these issues next in the particular (but so
important) case of incompressible fluids.

26
§8. Euler’s equations for incompressible ideal fluids
For many (ideal) flows of liquids (and even gases), mass density () can be supposed constant
throughout the volume of the fluid and along its motion. This is equivalent to neglecting
compression and expansion of these fluids. We speak of incompressible fluids:
ρ = constant.
Equations of fluid mechanics are much simplified for an incompressible fluid. The continuity equation
becomes:
div(𝚽) = 0.

Euler’s equations in the presence of a gravitational field become:

∂𝚽 ∂𝚽 1 p
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = ( 𝚽) ( )
+ 𝚽𝚽𝚽𝚽 𝚽 − 𝚽 × 𝚽𝚽𝚽𝚽 𝚽 = −𝚽𝚽𝚽𝚽 ( ) + 𝚽.
∂t ∂t 2 ρ

Obviously, we can take the curlof the above formula, which leads to an expression involving only the
velocity field:
∂𝚽𝚽𝚽𝚽(𝚽)
= 𝚽𝚽𝚽𝚽[𝚽 × 𝚽𝚽𝚽𝚽(𝚽)].
∂t

Interestingly, as the mass density is not an unknown function any longer for incompressible fluids, the
closed set of equations for (ideal) fluids can be reduced to equations related to velocity only.

The vector 𝚽 = 𝚽𝚽𝚽𝚽(𝚽) is called the vorticity. The equation for the vorticity can be re-written after a
proper development of:
𝚽𝚽𝚽𝚽[𝚽 × 𝚽] = 𝚽. (div(𝚽)) − 𝚽(div(𝚽)) + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) − (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).

Using the condition div(𝚽) = 0 and the identity: div(𝚽) = 𝚽, we obtain:

∂𝚽/ ∂t + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
(5)

27
Together with the definition: 𝚽 = 𝚽𝚽𝚽𝚽(𝚽), these equations completely determine the velocity in
terms of the vorticity. The vector equation (5) is of fundamental importance. To understand it, we need
first to give a hint of what vorticity is physically.

We intend to show that vorticity encodes the magnitude and direction of the axis about which a fluid parcel
rotates locally. For simplicity, we consider a 2-dimensional case in the (xOy) plane. We observe the
deformation along the flow of a rectangular fluid parcel ABCD parameterized at time t by A(x,y),
B(x+dx,y), C(x+dx,y+dy) and D(x,y+dy). Its surface is
=dx.dy and after the time interval dt, the points ABCD at time t have evolved to A’B’C’D’ at time t+dt. It
can easily be shown that: d/dt=(vx/x+vy/y). Hence, the relative variation of the surface of
the fluid parcel is given by the divergence of the velocity.

We label the angles generated by the flow as:

d=(AB,A’B’) and d=(AD,A’D’). The global
rotation of the fluid parcel is given by the
rotation of the diagonal of the rectangle, which we
define asdt. By construction, it is equal to:
½( d+ d) (see figure).

Again, this is easy to show that d+ d=(vy/x-vx/y)dt, or equivalently:

=½( vy/x-vx/y).
Then the quantity , which is characteristic of the rotation, is half the vorticity (in 2- dimensional
space). This result can be generalized to 3-dimensional space with the vector result: =½. Vorticity is
thus directly related to the magnitude and direction of the axis about which a fluid parcel rotates.

We now come back on the vector equation (5) characterizing the evolution in space and time of the
vorticity. It can be written in another very important form using the flow map function. We can show that it
is equivalent to:

28
 𝚽(𝚽(𝚽, t), t) = 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t)). 𝚽𝚽 (𝚽) with 𝚽𝚽 (𝚽) = 𝚽(𝚽(𝚽, 0), 0) = 𝚽(𝚽, 0).
(5’)

This expression is a bit unusual. Indeed, 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t)) can be expended under the 3
components of the gradient with respect to X (in 3-dimensional space), but each component is itself a
vector due to the presence of 𝚽. That’s why the quantity defined by
𝚽(𝚽(𝚽, t), t) = 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t)). 𝚽𝚽 (𝚽) is a vector, where the scalar product is taken between the 3
components of the gradient and the 3 components of the vector 𝚽𝚽 (𝚽).
In order to prove the equivalence between expressions (5) and (5’), we differentiate the relation (5’)
with respect to time (material derivative). We obtain:
d𝚽 ∂𝚽
= + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽[𝚽(𝚽, t), t])𝚽 (𝚽).
𝚽𝚽𝚽𝚽
𝚽 𝚽
dt ∂t

On the right hand side, we have the derivative of a composition of functions which can be easily
computed (exercise below). Knowing that 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t))𝚽𝚽 (𝚽), is equal to
𝚽(𝚽(𝚽, t), t), we end up with:
∂𝚽
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
∂t
This is the vector equation (5), which completes the proof.

Exercise: Prove that: 𝚽𝚽𝚽𝚽𝚽 (𝚽[𝚽(𝚽, t), t])𝚽𝚽 (𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).

Solution: we first expand explicitly the first term, using the notation X=(X1, X2, X3) and
𝚽𝚽(𝚽) = (𝚽0,1, 𝚽0,2, 𝚽0,3). We obtain:
∂𝚽[𝚽(𝚽, t), t] ∂𝚽[𝚽(𝚽, t), t] ∂𝚽[𝚽(𝚽, t), t]
𝚽𝚽𝚽𝚽
𝚽
(𝚽[𝚽(𝚽, t), t]). 𝚽𝚽 (𝚽) = 𝚽 0,1 + 𝚽 0,2 + 𝚽 0,3 .
∂X1 ∂X2 ∂X3
∂𝚽[𝚽(𝚽,t),t] ∂𝚽(.) ∂𝚽
Also, in compact notations: = .
∂X1 ∂X1 ∂x

29
The sum above can thus be rearranged as:
∂𝚽(. ) ∂ ∂𝚽(. ) ∂ ∂𝚽(. ) ∂
[ ω 0,1 + ω 0,2 + ω 0,3 ] 𝚽.
∂X1 ∂x ∂X2 ∂y ∂X3 ∂z
This gives the result.

Exercise: Prove that if we are considering a space with only 2 dimensions, the vector equation
(5) reads:
∂𝚽
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = 0.
∂t
Similarly, prove that in 2-dimensional space, the equivalent formula:
𝚽(𝚽(𝚽, t), t) = 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t)). 𝚽𝚽 (𝚽).
Which can be simplified into:
𝚽(𝚽(𝚽, t), t) = 𝚽𝚽 (𝚽).
Comment these last two expressions.
Solution: When we consider a flow in a plane, which means in 2-dimensional space (2D) labeled in
Cartesian coordinates (x,y), the velocity field can be represented as a 2- dimensional vector
(vx, vy). Then, only the z-component of 𝚽 = 𝚽𝚽𝚽𝚽(𝚽) is non-zero by definition. This implies that the
scalar product between 𝚽 = (0,0, ωz) and the gradient in 2D is zero. Therefore, the vector equation
(5) reduces to the relation:
∂𝚽
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = 0. The vorticity is thus a scalar, we write: ω = ω. The
z
last expression
∂t

becomes a one dimensional partial differential equation: ∂ ω + (vx ∂ + vy ∂ ) (ω) = 0.

∂t ∂x ∂y

Applying a similar method as in the text above, we can show that:

ω(Φx(X1, X2, t), Φy(X1, X2, t), t) = ω0(X1, X2).
This proves the second relations. These two equivalent formulae mean that vorticity is conserved
along paths of fluid particles in 2-dimensional flows.

30
Exercise: Conservation of circulation. We consider a closed curve build up by fluid particles at initial time
C0. This closed curve will move with the fluid, and at time t, we can represent this curve using the flow
map notation as: Ct = {𝚽(𝚽, t) ∶ 𝚽 belongs to Ω0 }.
For an ideal incompressible fluid prove that the circulation of the velocity along the closed curve Ct is
conserved along the flow:

∮ 𝚽. d𝚽 = ∮ 𝚽. d𝚽
Ct C0

What can you conclude concerning the vorticity?

Solution: We compute the material derivative (following the flow of particles): d [∮ 𝚽. d𝚽].
dt Ct

We have already discussed this kind of calculus in §5. The difficulty is that the boundary of the integration
depends on time. However, as explained in §5, we can write:
d d𝚽
[∮ 𝚽. d𝚽] = ∮ . d𝚽.
dt Ct Ct dt

Then, using the equations of motion: d𝚽 = −𝚽𝚽𝚽𝚽(p) with the fact the contour of integration
dt ϱ

is closed, we end up with the desired relation:

d
[∮ 𝚽. d𝚽] = 0.
dt Ct
This completes the proof. The circulation of the velocity along the closed curve Ct is conserved
along the flow.
Consequences for the vorticity: Using the circulation (Stokes) theorem, we can write:

∮ Ct 𝚽. d𝚽 = ∫St 𝚽𝚽𝚽𝚽(𝚽). d 𝚽 = ∫St 𝚽. d 𝚽, where St represents any surface moving with the flow bounded

by the closed contour Ct. This means that the ith component of the vorticity vector can be seen as the
limit circulation per unit area in the plane perpendicular to the (xi)- direction. Intuitively, it measures
how much a little leaf carried by the flow would spin about the (xi)-direction.
Also, we can understand physically what happens if we are restricted to 2-dimensional space (2D). In 2D,
incompressibility implies that St is a constant of motion: it derives from the continuity equation which
is equivalent to a volume preserving condition (surface preserving

31
in 2D). Then, the condition that ∫St 𝚽. d 𝚽 is conserved along the flow, in the limit of very small

area implies that the vorticity is conserved along the flow lines.

This can be formulated as: d𝚽 = 0. This corresponds to what we have shown mathematically
dt

in the previous exercise.

Differently, for 3-dimensional space (3D), there is no constraint on St following the continuity equation (the
volume preserving condition for incompressible fluids). Thus, conservation of the flux of vorticity
cannot control the magnitude of the vorticity vector.
To conclude briefly this discussion, we have understood intuitively some differences between the
2D and 3D physics cases. As we shall see later, these differences can explain why
2D equations of fluid mechanics cannot have singularities why 3D equations might.

We consider the right hand side term ((𝚽. 𝚽𝚽𝚽𝚽)(𝚽)) of the vorticity equation:
D𝚽
= (𝚽.
𝚽𝚽𝚽𝚽)(𝚽). Dt
As we have shown in the exercises above, this term is not present for 2-dimensional flows, for which
vorticity is conserved along flow lines. Therefore, we know that this is the term which brings some
complications for 3-dimensional flows. This is interesting to get an intuitive understanding of it.
Rephrased in words, (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) is proportional to the derivative in the direction of ω along a
vortex line: (ω. ∂/ ∂sω)(𝚽). Where δsω is the length of an element of vortex line. We now resolve the
vector v into components vω parallel to the vortex line (of direction ω) and v⊥ perpendicular to ω and
hence to δsω. Projected along the vortex line, we obtain:
∂(𝚽)
1 D𝚽
. δs ω = . δs = ∂(𝚽ω + 𝚽⊥ ) . δs .
ω Dt ∂sω ω ∂sω ω

32
This gives:
∂(𝚽ω) ∂(𝚽⊥ )
1 D𝚽 = . δs + . δs = {𝚽 (𝚽 + δs ) − 𝚽 (𝚽)} + {𝚽 (𝚽 + δs ) − 𝚽 (𝚽)}.
. δs ω ω ω ω ω ⊥ ω ⊥
ω Dt ω ∂sω ∂sω
The first term on the right hand side represents the rate of stretching of the element δsω. The second term
represents the rate of turning of the element δsω. Then, stretching along the length of a vortex line
causes relative amplification of the vorticity field, while turning away from the vortex line causes a
reduction of the vorticity in that direction, but an increase in the new direction.

Summary of important equations for the vorticity (incompressible ideal fluids):

∂𝚽
General equivalent equations for the + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽),
∂t
vorticity, also called the ‘stretching
equivalent to
formulae’.
𝚽(𝚽(𝚽, t), t) = 𝚽𝚽𝚽𝚽𝚽 (𝚽(𝚽, t)). 𝚽𝚽 (𝚽)
with 𝚽𝚽 (𝚽) = 𝚽(𝚽(𝚽, 0), 0) = 𝚽(𝚽, 0).
The equation using the flow map expresses the In 2D, the general vector equation (above)
fact that vortex lines are carried by the flow. In 2- becomes:
dimensional spaces, the vorticity is carried 𝚽(𝚽(𝚽, t), t) = 𝚽𝚽 (𝚽).
along by particle paths, its magnitude Moreover, the vorticity in 2D is a scalar and this
unchanged. In 3-dimensional spaces, vorticity last equation can be written as:
is carried as well, but its magnitude is ω(Φx(X1, X2, t), Φy(X1, X2, t), t)
amplified or diminished by the gradient of the = ω0(X1, X2).
flow map. Or equivalently:
∂ ∂ ∂
ω + (v +v ) (ω) = 0.
∂t x ∂x y ∂y

33
§9. Potential flows for ideal fluids
We start by restating the property of conservation of circulation (see §8). We have shown that for an
ideal incompressible fluid the circulation of the velocity along the closed curve Ct (moving with the flow)
is conserved:

∮ 𝚽. d𝚽 = ∮ 𝚽. d𝚽.
Ct C0

However, in the course of the proof (in §8), we have used a mathematical trick (with no justification)
concerning the inversion of the integral whose boundary depends on time and
the material derivative: d [∮ 𝚽. d𝚽] = ∮ d𝚽 . d𝚽. In this section we redo this important proof
dt Ct Ct dt

completely. It will allow us to state clearly the minimal assumptions needed to derive this property
(theorem).
We are interested by the material derivative of the circulation of velocity on a closed contour:

∮ 𝚽. d𝚽. The closed contour is supposed to be drawn in the fluid at some instant of time and we
assume that this corresponds to a fluid contour, build up by fluid particles, which lie on this contour. See
the figure below for illustration, where only a few fluid particles have been pictured for simplicity.

In the course of time, these fluid particles move about with the flow and thus the contour moves with
them accordingly. We calculate the material derivative of the velocity circulation bounded by this
contour:
d
[∮ 𝚽. d𝚽].
dt

34
First, this is really the material derivative that we need to evaluate, as we are interested in the change of
the circulation round the fluid contour moving with the flow (displayed in the figure above), and not
round a fixed contour in space coordinates.
Then, we write the element of length in the
contour as d𝚽 = δ𝚽, where δ𝚽 is the
difference between the radius vectors of the
points at the ends of the element of the
contour d𝚽. In this proof, we use the symbol
′δ′ for the differentiation with respect to
space coordinates and ′d′ for the
differentiation with respect to time.

Then, the circulation of velocity can be written as ∮ 𝚽. d𝚽 = ∮ 𝚽. δ𝚽. In differentiating this integral
with respect to time, we need to consider that not only the velocity but also the contour itself changes,
alongthe flow. That’s why we need to differentiate both vand δ𝚽. This leads to:
d d𝚽 dδ𝚽
[∮ 𝚽. δ𝚽] = ∮ . δ𝚽 + ∮ 𝚽. .
dt dt dt
The second integral on the right hand side is trivial to compute after writing the integrand in
the following form using simple algebra: 𝚽. dδ𝚽 = 𝚽. δ d𝚽 𝚽2

dt dt
= δ ( 2 ). This gives after integration

on a closed contour:
dδ𝚽
∮ 𝚽. = 0.
dt
This proves the mathematical identity presented previously: d [∮ 𝚽. δ𝚽] = ∮ d𝚽 . δ𝚽. From
dt dt

Euler’s equations for isentropic fluid (not necessarily incompressible), we can write:
d𝚽
= −𝚽𝚽𝚽𝚽(h).
dt

35
Then, this is easy to conclude that d𝚽 . δ𝚽 = 0, since:
∮
dt
d𝚽 d𝚽
∮ ( )
. δ𝚽 = ∮ 𝚽𝚽𝚽𝚽 ( ) . δ𝚽 = ∮ 𝚽𝚽𝚽𝚽(−𝚽𝚽𝚽𝚽 h ). δ𝚽 = 0.
dt dt

We end up with the property of conservation, which completes the proof:

d
[∮ 𝚽. d𝚽] = 0.
dt

∮ 𝚽. d𝚽 = constant (along the flow).

Therefore, for an ideal fluid, the velocity circulation round a closed fluid contour is constant in time: this is
also called the Kelvin’s theorem. We can remark that this property assumes
Euler’s equations for an isentropic flow. In fact, we need to be able to write 𝚽𝚽𝚽𝚽(p) as a
ρ

gradient of some function. This is the case for an isentropic flow for which the relation
s(p,)=constant poses a one to one relation between pressure and mass density.

From the law of conservation of circulation (along the flow), we can derive another essential property
concerning vorticity. We consider an infinitely small contour (build up by fluid particles) moving with
the flow and we assume that the vorticity is zero at some point along this path. We know (Stokes
theorem) that the velocity circulation round this (infinitely small) closed contour is equal to
𝚽𝚽𝚽𝚽(𝚽). d𝚽 = 𝚽. d𝚽, where d𝚽 is the element of area enclosed by this small contour. At the point
where 𝚽 = 𝚽, the velocity circulation round this small contour is thus also zero. In the course of time, this
contour moves with the fluid, always remaining infinitely small. Since the velocity circulation is
conserved along the flow, it remains equal to zero for all points of this path, and it follows that the
vorticity also must be zero at any point of this path. Therefore, we can state that: if at any point of some
trajectory followed by fluid particles the vorticity is zero, the same is true at any point of this trajectory.

36
Note that if the flow is steady (∂𝚽 = 𝚽), streamlines coincide with paths described in the
∂t

course of time by some fluid particles. In this case, we can consider a small contour that encircles
the streamline. In particular, if it encircles the streamline at the point where the vorticity is zero, this
property is conserved along the streamline. This means that for steady flow, the previous statement
holds for streamlines: if at any point on a streamline, the vorticity is zero, the same is true at all
other points on that streamline.

We continue the argument assuming the flow is steady. We consider a steady flow past a material
body with the much reasonable hypothesis that the incident flow is uniform at infinity. This means
that its velocity is constant at infinity, and thus its vorticity is zero at infinity. Following the previous
statement, we conclude that the vorticity is zero along all streamlines and thus in all space. In fact, this
is not exactly correct as the proof that vorticity is zero along a streamline is invalid for a line which lies on
the surface of the solid body, since the presence of the surface makes it impossible to draw a closed
contour encircling such a streamline! Of course, the physical problem of flow past a given body has a
well-defined solution. The key point is that ideal fluids do not really exist: any real fluid has a certain
viscosity, even small. This viscosity may have in practice no effect on the motion of most of the fluid, but,
no matter how small it is, it will become essentially important in a thin layer adjoining the surface of the
body. We come back on the mathematical description of how real fluids behave in the following sections.
For the moment, we keep the point of view of ideal fluids, knowing that there is a boundary layer around
a solid body inside which this (ideal) description does not apply.

A flow for which the vorticity is zero in all space is called a potential flow, or irrotational flow. Rotational
flows correspond to flow where the vorticity is not zero everywhere. Following the discussion above, a
steady flow past some material body, with a uniform incident flow at infinity must be potential.

37
Another consequence of the theorem of conservation of circulation is that, if at some instant, the flow is
potential throughout the volume of the fluid, we can deduce that this will hold at any future instant. This is
also in agreement with the equation for the vorticity, derived from
Euler’s equation: ∂𝚽 = 𝚽𝚽𝚽𝚽[𝚽 × 𝚽], which shows that if 𝚽 = 𝚽 at time t, it holds at time t+dt.
∂t

We now derive some general simple properties of potential flows, which are very useful in practice:
i. First, we have proved the property (theorem) of conservation of circulation, under the
assumption that the flow is isentropic. This means that if the flow is not isentropic, this law does not
hold. Therefore, if a non-isentropic flow is potential at some instant, the vorticity will in general be
non-zero at subsequent instants, and the concept of potential flow is useless. Therefore, all
what we discuss next assumes that the flow is isentropic.

ii. For a potential flow: ∮ 𝚽. d𝚽 = 0. It follows that closed streamlines cannot exist. Only in
rotational flows, closed streamlines can be present.
iii. If the vorticity (vector) is zero 𝚽 = 𝚽, this implies that there exist a scalar potential
such that:
𝚽 = 𝚽𝚽𝚽𝚽(ϕ).
With the Euler’s vector equation, we get:
∂ϕ v2
𝚽𝚽𝚽𝚽 ( + + h) = 𝚽.
∂t 2
Then, the function inside the gradient is not a function of coordinates and only depends
on time:
∂ϕ v2
()
+ +h=ft.
∂t 2
Here, f(t) is an arbitrary function of time. As the velocity is the space coordinate gradient of
the scalar potential: ϕ, we can add to ϕ any function of time without modifying the velocity
field. In particular, we can make the substitution:
ϕ → ϕ + ∫ fdt.

38
 This means that we can take f(t)=0 without loss of generality in the above equation.
iv. For a steady (and potential) flow, we can simplify the equation given in (iii) with ∂ϕ =
∂t

0 and f(t)=constant:
v2
+ h = constant.
2
This is the Bernoulli’s equation. However, there is an important difference with the Bernoulli’s
equation established in the general case (§5), where the ‘constant’ on the right hand side is
constant along any given streamline, but different for different streamlines. In potential flows,
the ‘constant’ (above) is constant throughout the fluid.
v. An important physics case, where potential flow occurs, concerns small oscillations of an
immersed body in a fluid. It can be shown that if the amplitude of oscillations is small
compared with the dimension of the body, the flow past the body is approximately
potential. The proof is left as an exercise below: the idea is to show that throughout the fluid 𝚽𝚽/
∂t = 𝚽 and thus the vorticity in the fluid is constant. In oscillatory motion, the average of the
velocity is zero, and then we establish that this constant is zero.
vi. Potential flows for incompressible fluids. We first recall that we define an
incompressible fluid by: ρ = constant, throughout the volume of the fluid and its motion. This
means that there cannot be noticeable compression or expansion of the fluid. Following the
continuity equation, this implies: div(𝚽) = 0. We finally recall that, for an incompressible fluid, we
have: dϵ = 0 (always with the isentropic hypothesis). This implies that ϵ is constant, and since
constant terms in the energy do not matter,
the energy flux density for an incompressible fluid becomes: ρ𝚽 (1 𝚽 2 + P/ρ).
2

Similarly, the enthalpy h can be replaced by P/ρ in the equation established in (iii). This leads
to:
∂ϕ v2
+ + P/ρ = f (t).
∂t 2
Now, we combine the two equations: div(𝚽) = 0 and 𝚽 = 𝚽𝚽𝚽𝚽(𝚽) = 𝚽 (or 𝚽
= 𝚽𝚽𝚽𝚽(ϕ)).

39
 We get:
∆(ϕ) = 0.
This is the Laplace’s equation for the potential ϕ. In order to solve this equation, it must be
supplemented by boundary conditions (see §6). For example, at fixed solid surfaces, where the
fluid meets solid bodies, the fluid velocity component normal to the surface (vn) must be zero.
For moving surfaces, it must be equal to the normal component of the velocity of the surface
(which can be a function of time). Note that
the following relation holds: v = ∂ϕ.nTherefore, the general expression of boundary
∂n

conditions is that ∂ϕ is a given function of coordinates and time at boundaries. We

∂n

show how to solve the Laplace’s equation with specific boundary conditions in some exercises
below.

Two other (less simple) consequences of the Kelvin’s theorem:

i. Vortex lines move with the fluid: consider a tube of particles, which at some instant forms a
vortex tube, which means a tube of particle with a given value of:

∮ 𝚽. d𝚽 = K.

Then, at that time, the circulation of the velocity round any contour C’ lying in the tube without
embracing the tube is zero, while for any contour embracing the tube (once), the circulation of
velocity is equal to K. These values of the velocity circulation do not change moving with the fluid.
This means that the vortex tube remains a vortex tube

with an invariant: K = ∮ 𝚽. d𝚽. A vortex line is a limiting case of vortex tube and therefore
vortex lines moves with the fluid (under the hypothesis of the Kelvin’s theorem).
ii. The direction of vorticity turns as the vortex line turns, and its magnitude increases
as the vortex line is stretched: the circulation round a thin vortex tube remains the same. As it
stretches the area of the section decreases and thus the vorticity (~circulation/area)
increases in proportion to the stretch.

40
Summary of some important relations for ideal fluids:
=constant. Incompressibility condition.
∂𝚽𝚽𝚽𝚽(𝚽)
= 𝚽𝚽𝚽𝚽[𝚽 × 𝚽𝚽𝚽𝚽(𝚽)]. Vorticity equations (general).
∂t
∂𝚽 Vorticity equations (incompressible fluid).
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽).
∂t
In 2-dimensional space the term on the right hand
side comes to zero and the vector equation
is reduced to:
d𝚽
= 𝚽.
dt
The circulation of the velocity along the Incompressible (ideal) fluid (isentropic flow).
closed curve Ct (moving with the flow) is
conserved (Kelvin’s theorem):

∮ 𝚽. d𝚽 = ∮ 𝚽. d𝚽.
Ct C0

Potential flow: 𝚽 = 𝚽𝚽𝚽𝚽(ϕ). Potential flow.

∂ϕ v2
+ +h=ft.()
∂t 2
Potential and steady flow: Potential flow (steady).
v2
+ h = constant.
2
This ‘constant’ is constant throughout the fluid.

Potential and incompressible flows: Potential flow (incompressible).

∂ϕ v2
+ + P/ρ = f t(. )
∂t 2
∆(ϕ) = 0.

41
§10. Real fluids and Navier-Stokes equations
We now study the effect of energy dissipation, occurring during the motion of a fluid, on that motion itself.
This process is the result of the thermodynamic irreversibility of the motion. This irreversibility is always
present to some extent, and is due to internal friction (viscosity) and thermal conduction. In the following,
we always assume that the fluid is incompressible (=constant).
In order to obtain the equations of motion of a viscous incompressible fluid, we have to include
some additional terms in the equations of motion of an ideal fluid.
i. First the equation of continuity, as it is clear from its derivation, is equally valid for any fluid,
whether viscous or not.
ii. Then, Euler’s equations requires modifications, that are described this section.

We have seen in §5 that Euler’s equations can be written in the form:

∂
∫ ρv dV = − ∑ ∮ Π dΣ .
ik k
∂t i
k=x,y,z

Where Πik is the momentum flux density tensor (of rank 2): it depends on 2 indices i and k, running for x,
y, z in 3-dimensional space. For ideal fluids, we have Πik = pδik + ρvivk (with δik = 1 if i = k and 0
otherwise), which represents a completely reversible transfer of momentum, originating from the
mechanical transport of the different particles of the fluid from place to place with pressure forces
acting in the fluid. As already mentioned, the hypothesis behind ideal fluids is that each particle
pushes its neighbors equally in every direction. This is why a single scalar quantity, the pressure (p), is
sufficient to describe the force per unit area that a particle exerts on all its neighbors at a given time.
Then, the acceleration of the fluid particle results from the pressure differences.

For real fluid, for which energy dissipation in the fluid is not neglected any longer, the viscosity or
internal friction is due to another, irreversible, transfer of momentum from points where the velocity
is large to those where it is small. This means that if a fluid particle

42
moves faster than the average of its neighbors, then friction slows it down. The equation of motion for
real fluids can then be obtained by modifying the momentum flux tensor accordingly Πik = pδik +
ρvivk + Π′ik, where Π′ik designs the part of the flux density due to viscosity. A general form of Π′ik can be
established using the fact that processes related to internal friction occur in a fluid only when different
particles move with different velocities, such that there is a relative motion between various parts of
the fluid. Hence, Π′ik must depend on the space derivatives of the velocity. As a first approximation,
Π′ik can thus be
written as a linear combination of terms of the form ∂vi .
∂xk

This way of thinking can be continued to obtain the general vector equation of motion of incompressible
viscous fluid, for which the viscosity is determined by only one coefficient:
1 η
d𝚽 ∂𝚽 )( ) ( )
= (
+ 𝚽. 𝚽𝚽𝚽𝚽 𝚽 = − + ∆𝚽.
dt ∂t ρ ρ
𝚽𝚽𝚽𝚽 p
(6)
This is the Navier-Stokes (vector) equation, first established by Navier in 1822 and then by Stokes in
1845. Other derivations have been proposed in between by Cauchy in 1828 and Poisson in 1829, but
the history has only kept the names of Navier and Stokes. Here, η > 0 is
called the viscosity coefficient (precisely the dynamic viscosity), while ν = η is called the
ρ

kinematic viscosity. In Cartesian coordinates, we can write equations (6) as a set of 3 equations
for the 3 components in x, y and z:
∂p
∂ ∂ ∂ vx 1 ∂p ∂x vx
( +v +v +v ∂ ) [vy] = − + η
∆ [vy].
∂t x ∂x y ∂y z ∂z
vz ρ ∂y ρ vz
∂p
[∂z]

Some general comments are in order:

i. In general the viscosity coefficient is a function of pressure and temperature of the fluid. As
pressure and temperature may not be constant throughout the fluid, the

43
 viscosity coefficient also may not be constant throughout the fluid. Obviously, the viscosity
coefficient is much larger for glycerin than for water. Typical values at 20°C
are η = 0.01 cm²/s for water, 0.15 cm²/s for air and above 6 cm²/s for glycerin.
ρ

Remark that the unit for η is in cm²/s as be seen immediately from equations (6). Also,
ρ

we can mention that the dynamic viscosity (η) of a gas at a given temperature is
independent of the pressure, while the kinematic viscosity (η) is inversely
ρ

proportional to the pressure. Also, it can be shown that the kinematic viscosity of a gas is
about: Λυ, where Λ is the mean free path and υ the thermal speed of molecules (of the gas), of the
same order of magnitude as sound speed. It presents the correct unit (in length/time²). With this
expression, we understand that viscosity is a vestige of the continuum limit based on the
molecular nature of the fluid while Euler’s equations, corresponding to zero viscosity, set this
molecular length scale to zero (see
§2).
ii. The Navier-Stokes equations (6) represent a vector equation (with as many equation as
dimensions of space), which generalizes the Euler’s vector equation due to the
presence of a diffusive term in η ∆𝚽. The convective term in (𝚽. 𝚽𝚽𝚽𝚽)(𝚽), which
ρ

appears in the material derivative, is not affected by the internal friction phenomenon.
iii. The Navier-Stokes equations (6) need to be complemented by the continuity relation for
incompressible (real) fluids, div(v)=0, as well as initial and boundary values to compute the
velocity field for later times t > 0 (next sections). Of course, in the
presence of external forces, we can add a term in ρ1 𝚽ext to equations (6).

iv. A simple derivation of the viscosity term can be proposed, without the use of the tensor
formalism. The idea is to make the parallel between two properties: (1) the physics property
that frictional forces counteract at each point the deviation of the velocity field from its local
average and (2) the mathematical fact that the deviation of a function at a point from its average
value on small surrounding spheres is measured by the negative of the Laplace’s operator (∆).
This implies trivially that frictional (vector) forces must be proportional to ∆𝚽. Adding such
a term to Euler’s vector

44
 equation, we obtain the Navier-Stokes vector equation (6) for incompressible fluids. We let the
justification of the equivalence between (1) and (2) above as an exercise. The key point here
is to see the intricacy between the physical intuition and mathematics: once we have
the physics idea (1), then equations (6) is a direct consequence of the mathematical fact
(2).

Using equations (6), we can verify that the presence of viscosity results in the dissipation of energy,
which is finally transformed into heat: this must be a direct consequence of the Navier-Stokes
equations (6). The calculation of energy dissipation is quite simple for an incompressible fluid filling a
volume of space Ω. Indeed, the total kinetic energy for the fluid contained in Ω reads:
1
E = ρ ∫ 𝚽²dV (with ρ = constant) .
K
Ω 2
Then, it can be shown easily that EK is decreasing in time under a flow following the Navier- Stokes
equations (6). Precisely, we can compute for a 3-dimensional space:
dEK 3 3
= −η ∑ ∫ |𝚽𝚽𝚽𝚽(u )|2dV = −η ∑ ∫ |𝚽u |2dV < 0.
i i
dt Ω i=1 Ω
i=1

When the viscosity is zero (Euler’s equations), we find that the kinetic energy is conserved
for an incompressible fluid (as already shown), while for viscous fluid dEK < 0: friction
dt

transforms kinetic energy into heat.

Exercise: For a 3-dimensional space, prove that:

3 3
dE K
= −ρη ∑ ∫ |𝚽𝚽𝚽𝚽(u )|i dV
2
= −ρη ∑ ∫ |𝚽u | dV <i 0.
2
dt Ω Ω
i=1 i=1

Where EK is the kinematic energy of the incompressible fluid contained in the volume Ω.
At boundaries of the volume containing the fluid (labeled as S in the following), we assume that the fluid
is at rest with vn=0.
Solution: We compute the time derivative of the kinematic energy:

45
 dEK d 1 ∂𝚽
= ρ [∫ 𝚽²dV] = ∫ 𝚽. ρ [ dV = ∫] 𝚽. [−ρ(𝚽. 𝚽𝚽𝚽𝚽)(𝚽) − 𝚽𝚽𝚽𝚽(p) + η∆𝚽]dV.
dt dt Ω 2 Ω
∂t Ω

There are 3 terms to be calculated. The first 2 terms are zero. For the first one, we find using the well-
known vector identity for (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) and integrating by parts:
1 1 1
∫ 𝚽. [(𝚽. 𝚽𝚽𝚽𝚽)(𝚽)]dV = ∫ 𝚽. 𝚽𝚽𝚽𝚽(𝚽𝚽)dV = − ∫ div(𝚽) 𝚽𝚽dV + ∫ 𝚽𝚽 𝚽. d𝚽 = 0.
Ω Ω 2 2 Ω 2 S

Similarly, we have: ∫Ω 𝚽.[𝚽𝚽𝚽𝚽(p)]dV = −∫Ω div(𝚽) pdV +∫S p 𝚽.d𝚽 = 0. Finally, we are
left with the last term proportional to the viscosity:
dEK
= η∫ 𝚽.[∆𝚽]dV.
dt Ω

The integrand: 𝚽. [∆𝚽] can be expanded as: v1. [∆v1] + v2. [∆v2] + v3. [∆v3]. Each of the 3

integrals (like ∫Ωv1.[∆v1]dV) can be transformed using integration by parts. We obtain:

3
dE K
= −η∑∫ |𝚽𝚽𝚽𝚽(v )| idV2 < 0.
dt Ω i=1

Consequently, the kinetic energy is decreasing in time, which reflects the losses due to
friction in a viscous flow. It is also possible to write the last formula as:
dEK
= −η∫ 𝚽𝚽𝚽𝚽(v²).d𝚽.
dt S

Finally, we write the vector equation for the vorticity by taking the curl of the Navier-Stokes equations. We
obtain:
∂𝚽𝚽𝚽𝚽(𝚽)
= 𝚽𝚽𝚽𝚽[𝚽 × 𝚽𝚽𝚽𝚽(𝚽)] + ν∆(𝚽𝚽𝚽𝚽(𝚽)).
∂t
Equivalently, in a form generalizing the stretching formula to (incompressible) viscous fluids:
d𝚽 ∂𝚽
= + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) + ν∆(𝚽) with 𝚽 = 𝚽𝚽𝚽𝚽(𝚽).
dt ∂t

46
In this expression, we recognize transport, stretch as well as diffusion of vorticity. Note that even if the
vorticity is null at initial time, it does not imply that it will last for later time since vorticity can be created by
boundary conditions. Also in 2-dimensional space, we obtain:
d𝚽
= ν∆(𝚽). dt
Here, only one component of 𝚽 is non zero: the vorticity is thus a scalar quantity. Vorticity is transported
through convection and diffuses. Of course, in the absence of viscosity (ν = 0) we find again that
vorticity is a conserved quantity in 2-dimensional space.

§11. Boundary conditions for real fluids

We must also write down the boundary conditions on the equations of motion of a viscous fluid. There
are always forces of molecular attraction between viscous fluid and the surface of a solid body, and these
forces have the result that the layer of fluid immediately adjacent to the surface is brought completely to
rest, and adheres to the surface. Accordingly, the boundary conditions require that the fluid velocity
should vanish at fixed solid surfaces: 𝚽 =
0. It can be emphasized that both the normal and tangential velocity must vanish, whereas for an ideal
fluid, the boundary conditions require only that the normal component vanish. Obviously, in the general
case of a moving surface, the velocity (vector) v must be equal to the velocity of the surface.

We consider one example to illustrate how these conditions operate together with the equations of
motion for a real fluid. Consider the following viscous incompressible flow between two stationary
plates located at y=0 and y=1, with the notations indicated on the figure. This is obviously a 2D
configuration, where only x and y component (in Cartesian coordinates) should be considered.

We are looking for a stationary solution of Navier-Stokes equations of the form v=(u(x, y), 0). Indeed,
the flow is directed only along the x-axis. With: p=p(x) and p(0)=p1, p(L)=p2. The continuity equation
for incompressible fluid gives ∂xu = 0, where we use the standard

47
notation ∂x = ∂/ ∂x. Hence u(x,y)=u(y) and v=(u(y), 0). The Navier-Stokes equation for the only
component to be considered reads:
1 η ∂ 2 ∂ 2
∂tu + u ∂xu = − ∂xp + [∂xxu + ∂yyu] where ∂xx = ( ) and ∂yy = ( ) .
ρ ρ
∂x ∂y
Where we have: ∂tu = 0, ∂xu = 0 and ∂xxu = 0. We obtain:

1 η
− ∂xp + ∂yyu = 0 or equivalently ∂xp = η ∂yyu.
ρ ρ

With the boundary conditions: u(x, 0) = 0 and u(x, 1) = 0. Since both sides of the equation
∂xp = η ∂yyu, depends on different variables, we conclude that ∂xp = η ∂yyu = constant, which gives
trivially the result. The velocity profile is a parabola.

We now add a level of complexity to the previous example. There is one plate at y=0. Assume that the half
space y > 0 is filled with a fluid, bounded by a plate (xz-plane) at
y = 0. The fluid is not moving for t ≤ 0. The plate at the fluid boundary starts to move at initial time (t = 0) with
constant velocity U in the x-direction. We may assume that the fluid starts to flow due to friction. In this
case we look for a solution of the Navier Stokes equations of the form 𝚽 = (u(x, y, t), 0, 0), where u(x, y, t) =
u(y, t) in order to satisfy the continuity relation. The Navier-Stokes equations read:
1 η
∂tu = − ∂xp + ∂yyu, with ∂yp = ∂zp = 0.
ρ ρ
This implies that ∂xp = constant, that we may choose to be zero. Then, we are left a the pure diffusion
equation ∂tu = ν ∂yyu (with ν = η) together with the boundary conditions u(0,t)=U
ρ

and the initial condition u(y,0)=0. We can pose this problem for the dimensionless variable: u̅= u/U. We
obtain: ∂t u̅= ν ∂yy u̅, u̅(𝚽, 0) = 0 and ̅u(0, t) = 1. Since u̅is dimensionless, it can only be a function of
dimensionless variable(s), combining the variables y, ν and t. There is
only one such variable that can be built y .
√νt

Therefore, we can look for u as a function of the variable ξ = y .

2√νt

48
Hence, we can write: u
̅ = F(ξ), with: ∂ξξF + 2ξF = 0, F(0) = 1 and F(∞) = 0. The solution is well known: the
complementary error function. It reads:
F(ξ) = erfc(ξ) = 2 ∫∞ exp(−s2) ds.
ξ
√π

Finally, we can transform easily this solution to obtain:

y
u(y, t) = U. erfc ( ).
2√νt
In particular, we observe that at fixed distance to the plate, the velocity of the fluid will converge to
the velocity of the plate at the limit of infinite time.

From these examples, we can remark that the solution of a problem for a viscous fluid for which we
take the limit ν → 0 is not equivalent to the similar problem solved in the ideal case. Indeed, the boundary
conditions are different in both cases. Then, even if the equations of motion would be similar in the limit
of zero viscosity, the general solution of a problem with given boundaries would not be equivalent.

Finally, we come back to the formulation of the Navier-Stokes equations in the form:
∂
∫ ρv dV = − ∑ ∮ Π dΣ .
ik k
∂t i
k=x,y,z

Here, the index i represents the Cartesian coordinates x, y and z. The correspondence with
equations (6) leads to: Πik = pδik − η (∂vi +∂vk) + ρvivk. Then, the expression above is
∂xk ∂xi

exactly equivalent to Navier-Stokes equations (6). The interest of the last formula is that it is easy to write
down from it an expression of the force acting on a solid surface bounding the fluid. The force acting on
an element of the surface is just the momentum flux through this element. The momentum flux through
surface d is ∑k=x,y,z ΠikdΣk. We write the element of surface along the k-axis as dΣk = dΣ. nk, where n is
the unit vector along the normal (along the outward normal to the fluid).

49
Then, asv=0at the solid surface, we conclude that the force (along the i-axis) acting on a unit surface of
the body is equal to:
∂vi ∂vk ∂vi ∂vk
∑ [pδ − η ( + )] n = pn − η ∑ ( + )n.
ik i
∂xk ∂xi k ∂xk ∂xi k
k=x,y,z k=x,y,z

The first term is the ordinary pressure of the fluid acting on the surface of the body, while the second is the
force of friction, due to viscosity.

Note also that if we have a surface of separation between 2 immiscible fluids, the conditions at the
surface are that the velocities of the fluids must be equal and the forces which they exert on each other
must be equal and opposite. This is a generalization to the case of real fluids of the condition of the
continuity of pressure for the ideal case.

§12. Reynolds number and related properties

We define a macroscopic (dimensionless) number corresponding to the ratio of the strength of the non-
linear effects to the strength of the linear viscous effects. In order to define this quantity, we need to
introduce a characteristic scale U for the velocity and a characteristic length scale L of the flow. In
addition, we write the characteristic time scale as T=L/U. Then, we can pose the dimensionless
parameter we are looking for, the Reynolds number, Re=UL/. It is obvious to check that it
corresponds to the ratio mentioned above. We remark that the (kinematic) viscosity  has dimension
L²/T, using these characteristic scales. This corresponds to the diffusivity of the velocity: namely, in
time T, velocity diffuses over a
distance of order √νT. For example, for water at ambient temperature and pressure, ν~1mm/s²,
so direct viscous effects diffuse velocity at a distance of order 1mm in one second. Also, if we consider the
flow of water with a speed of 1 m/s in a tank of 1 m, we obtain: Re=106.

50
We now introduce the dimensionless variables: (x’, y’, z’)=(x/L, y/L, z/L), t’=t/T, v’=v/U. The Navier-
Stokes equations can then be re-written using the prime (dimensionless) variables in the form:
∂𝚽′ 1
+ (𝚽′. 𝚽𝚽𝚽𝚽′)(𝚽′) = −𝚽𝚽𝚽𝚽′(p′) + ∆′𝚽′.
∂t′ Re

Here, we can check that p’=p/(U²). With this expression, we can see that solutions present a scaling
invariance 𝚽 → λ𝚽(λ𝚽, λ2t), up to the limit of the continuum hypothesis. This leads to the law of
similarity: flows which can be obtained from one another by simply changing the unit of measurement of
coordinates and velocities are said similar. Thus flows of the same type and same Reynolds numbers
are similar.

As the Reynolds number is a standard referenced quantity for various flows, this last expression
is useful in order to make some comparisons between the relevant terms of the equation:
i. For example, we may think that we can neglect the viscous term the Navier-Stokes equations
in comparison with the convective term when the Reynolds number is sufficiently large (at
small viscosity). However, the Navier-Stokes equations correspond to a singular
perturbation of the Euler equations, since the viscosity multiplies the term that contains the
highest-order spatial derivatives. As a result, this is not always possible to operate this
simplification. The high Reynolds number limit of the Navier-Stokes equations is a very difficult
problem, where turbulent effects may dominate the dynamic of the flow.
ii. On the other hand, the limit of small Reynolds numbers is particularly simple. For steady
flow of incompressible viscous fluid, the Navier-Stokes equations read:
1 η
(𝚽. 𝚽𝚽𝚽𝚽)(𝚽 = − 𝚽𝚽𝚽𝚽(p) + ∆𝚽.
) ρ ρ

Using the same notations for characteristic scales as defined above, the convective term is of
the order U²/L,. The diffusive term is of the order of magnitude η U/L². The
ρ

ratio of the two is the Reynolds number (by definition). Hence, the convective term

51
 may be neglected if the Reynolds number is small, and the equations become linear:
−𝚽𝚽𝚽𝚽(p) + η∆𝚽 = 0. Together with the continuity equation div(𝚽) = 0, it completely
determines the motion of the fluid. It is useful to note that:
−𝚽𝚽𝚽𝚽(p) + η∆𝚽 = 0 implies ∆𝚽𝚽𝚽𝚽(𝚽) = 0.

Using these formulae, it is possible to determine the force exerted on a ‘fixed’ sphere (of radius R)
by a moving fluid, having a given velocity U constant at infinity, far away from the sphere body.
This force is also called the drag force (F). As already shown in previous sections, the solution
of this problem with the Euler’s formalism gives a resulting null force. Using the Navier-Stokes
formalism, in the small Reynolds number approximation, we expect that the answer is non-
zero due to the presence of the viscosity coefficient. Indeed, it can be shown that F =
6πηR||𝚽||, with a force parallel to the velocity direction U. This formula is called the Stokes
formula. After a trivial change of reference frame, it gives the drag force on a sphere moving
slowly in a fluid at rest at infinity. We notice that the drag is proportional to the first power of
the velocity and viscosity coefficient, as well as linear in the dimension of the body. These are
general properties, whatever the exact shape of the body considered.

Finally, we discuss an interesting physics effect for real fluids: boundary layers that we have briefly
mentioned in §8. We consider a simple example first. Assume that the half space
y > 0 is filled with a fluid, bounded by a plate (xz-plane) at y = 0. We consider that the plate is fixed and the
fluid moves at constant velocity U along the x-axis at y = ∞. We have solved a similar problem in the
previous section. Here, only boundary conditions are changed compared to our previous
discussion. They read: u(y = ∞, t) = U and u(0, t) = 0.
The solution follows: ξ
y 2
( ) () ( 2)

u y,t = U. erf ( ) where erf ξ = ∫ exp −s ds

2√νt √π 0

52
The region in which the velocity field departs significantly from the constant flow U is called boundary
layer: according to the expression of u(y, t), this layer is proportional to √νt. Hence, for fixed time, the
boundary layer decreases as 1/√Re.

We summarize the differences between incompressible Euler’s and Navier-Stokes equations:

Navier-Stokes: Euler:
=constant, taken to be 1 (in proper units) =constant, taken to be 1 (in proper units)

∂𝚽 1 ∂𝚽
+ (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = −𝚽𝚽𝚽𝚽(p) + ∆𝚽 + (𝚽. 𝚽𝚽𝚽𝚽)(𝚽) = −𝚽𝚽𝚽𝚽(p)
∂t Re ∂t

div(𝚽)=0 div(𝚽)=0

v=0 at fixed boundary v.n=0 at fixed boundary

We assume that both types of flows coincide at initial time and that both flows are irrotational at initial time.
Under Euler’s formalism, the flow stays irrotational at all times. Under Navier- Stokes equations, the
effect observed in the example above can be generalized: the flow is drastically modified (compared
to the ideal case) near the boundary in a region with thickness proportional to 1/√Re. In addition,
this (local) effect may be a source of vorticity.

53
§13. The millennium problem of the Clay Institute

With all the knowledge of the previous sections, we can posethe millennium problem related to Navier-
Stokes equations.
Given the Navier-Stokes equations for incompressible fluids in 3-dimensional space:
1 η
∂𝚽 )( ) ( ) ( )
(
+ 𝚽. 𝚽𝚽𝚽𝚽 𝚽 = − + ∆𝚽 and div 𝚽 = 0
∂t ρ ρ
𝚽𝚽𝚽𝚽 p
(NSI)
𝚽𝚽𝚽 𝚽𝚽𝚽𝚽𝚽𝚽 − 𝚽𝚽𝚽𝚽𝚽𝚽 −
𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽𝚽 with 𝚽 = 𝚽(𝚽, t ≥ 0)
and p = p(𝚽, t ≥ 0).

Given smooth (divergence free and infinitely differentiable) initial conditions:

𝚽𝚽(𝚽) = 𝚽(𝚽, t = 0).

Given boundary conditions at infinity: the statement of the problem indicates that for physically
reasonable solutions, such that the velocity field does not grow large as ‖𝚽‖ → ∞, the space of initial
conditions is restricted to functions for which the norm of space
derivatives of any order (written schematically as: ‖∂α𝚽 ‖) is bounded. Precisely, for any
𝚽 𝚽

order  and for any constant K, there exists a constant C(α, K) such that:
C(α, K)
‖∂α𝚽 (𝚽)‖ ≤ .
𝚽𝚽
(1 + ‖𝚽‖)K
1
2
Here the norm of a function f: Ω → IR has to be understood as ‖f‖ = (∫ |f(x)|2Ωdx) .

Then, prove that:

Option 1: there exists 𝚽(𝚽, t ≥ 0) and p(𝚽, t ≥ 0) solutions of (NSI) which are infinitely
differentiable in space and time coordinates (for t ≥ 0), with bounded kinetic energy over the all
space. Precisely, there exists a constant such that: ∫‖𝚽(𝚽, t)‖²d3𝚽 < constant for t ≥ 0.

54
 Option 2: there exists a divergence free and infinitely differentiable vector field
𝚽𝚽(𝚽, t = 0) satisfying the boundary conditions and a smooth external force vector field, such
that there exists no solution of (NSI) infinitely differentiable in space and time coordinates (for
t ≥ 0) and with bounded energy over the all space. In this option, the force vector field f(x,t) is
simply added to the vector equation (NSI). It is assumed to be bounded at infinity. Precisely,
for any orders α and β and for any constant K, there exists a constant C(α, β, K) such that:
C(α, β, K)
‖∂α ∂β𝚽(𝚽, t)‖ ≤ .
𝚽t
(1 + ‖𝚽‖ + t)K

Options 1 or 2 are equally good to solve the problem stated by the Clay mathematical institute (CMI) and get
the 1 million dollars prize.
Alternatively, instead of taking boundary conditions with strong decay at infinity, we can consider
periodic boundary conditions in space coordinates for 𝚽𝚽, v and f, and one additional
condition for the external force: ‖∂α ∂β𝚽(𝚽, t)‖ ≤ C(α,β,K). Then, options 1 and 2 can be posed
𝚽t
(1+t)K

in a similar way as done above.

At the end, with the 2 types of boundary conditions, we have 4 possible statements of the problem. The
CMI asks for the proof of one of them to get the prize.

In short, this problem raises the question of whether arbitrary smooth (also called regular) solutions of
the incompressible Navier-Stokes equations in 3-dimensional space can be continued globally from
smooth initial data or not. Either, one needs to prove that initially smooth solutions with strong decay
conditions at infinity (or with periodic boundary conditions) remain smooth for all times, or one
needs to find at least one solution which blows up in finite time.

If this problem is solved positively, it would imply:

55
 i. The existence of solutions: the system described bythe equations must have a wayto evolve in
the future.
ii. The uniqueness: there must not be arbitrary choices for this evolution.
iii. And the continuous dependence on the initial state: any future state of the system is determined,
to arbitrary finite precision, by the initial conditions to a sufficient finite precision.

Obviously, a complete answer to this problem is still open. However, what is known is that this
problem can be solved locally! Starting out from divergence free and infinitely differentiable
(smooth) initial conditions 𝚽𝚽(𝚽), solutions for 𝚽(𝚽, t ≥ 0) and p(𝚽, t ≥ 0) are unique, depend
continuously on the initial conditions, and remain smooth for at least, possibly short, interval of
time: [0, T∗ ], where T* depends on 𝚽𝚽(𝚽). It is not known if the solutions exist for t ≥ T∗ . This last
statement holds for Navier-Stokes equations as well as Euler’s equations. We prove rigorously this
result in the next section.
Then, either a given solution (from a given initial condition) can be continued for all times or it exists only
up to T* and the norm of the solution diverges when t → T ∗ : the solution is blowing up at T*. Then,
schematically, proving the problem globally reduces to finding a bound on the norm of the solution
for all times.

Some interesting partial results are known:

1. In 2-dimensional space, the problem has been solved and there exists smooth and globally
defined solutions. Very roughly, in 2-dimensional space, the dissipation of energy due to
friction is sufficiently strong to prevent blow-up of finite energy solutions. Note that for the
Euler’s formalism, the same conclusion holds, for a simple reason that we discuss in detail later:
in 2-dimensional space, vorticity is conserved as a scalar quantity along flow lines. The key point
is that for Euler’s formalism, it can be shown that any singularity of the velocity field or its
derivatives of any order is necessarily a singularity in the vorticity. Then, the conservation of
vorticity in 2- dimensional space is sufficient to prevent blow-up of finite energy solutions.

56
 2. In 3-dimensional space, if 𝚽(𝚽, t ≥ 0) is sufficiently small, the problem is also solved positively
for Navier-Stokes equations and there exists smooth and globally defined solutions. Much
effort has been spent on characterizing smallness, in terms of smallness of initial
conditions, of the viscosity being large, or of the solution being in some sense close to some
known special solution. Physically, the effect of the diffusion
term η ∆𝚽 is so strong that any perturbation coming from the convective term
ρ

(𝚽. 𝚽𝚽𝚽𝚽)(𝚽) is damped away before it could lead to singularities. Intuitively, if fluid
equations for water are in danger of developing singularities, we replace water by honey,
sufficiently viscous, and no singularity can develop. Obviously, these kinds of results are not
available for Euler’s equation, where the viscosity is absent.
3. Finally, another class of partial results known to hold for Navier-Stokes equations, but not for
Euler’s formalism, concerns the existence of the so-called weak solutions (J. Leray 1933). In
fact, solutions of Navier-Stokes equations can be continued past the time of their first possible
singularity as weak or generalized solutions. We do not want to describe this mathematical
issue here. This means that equations are satisfied by weak solutions in an average sense but
not point by point. Weak solutions exist globally in time. However, they are not known to be
unique!

From these first comments, we understand that the difficulty of the problem comes from the relative
balance between the quadratic term and the diffusive term in equations while the evolution in time is
realized. Similar observations have been done in §7 for much simpler
dynamical systems. Indeed, for the one dimensional equation: du = −u + u2, we have shown
dt

rigorously that there is a global existence of solutions with small initial data and local existence (in
time) of solutions with large initial data. Also, for the one dimensional equation:
∂u ∂( ) u2
+ 2
= ∂u + u ∂(u) = 0, we have noticed that a (unique) global smooth solution may not
∂t ∂x ∂t ∂x

exist in general.

57
Clearly also, when the quadratic term can be neglected in Navier-Stokes equations (NSI), we obtain a
much simpler system following:
∂𝚽
= −𝚽𝚽𝚽𝚽(p) + ν∆𝚽 [with div(𝚽) = 0].
∂t
Where we have posed =1. Taking the divergence of the last expression, we obtain a single equation for
the pressure field ∆p = 0, which decouples the pressure field dependence in space from the velocity
field.
This system of equations (for 𝚽 = 𝚽(𝚽, t ≥ 0) and p = p(𝚽, t ≥ 0)) can be solved according to initial and
boundary conditions as given, which justifies the conclusion of point (2) mentioned above.

For the general form of the Navier-Stokes equations, it is also possible to take the divergence of the
equations. Then, the Laplace’s operator of the pressure is more complex. Indeed we need to consider
the divergence of the quadratic (convective) term. With simple algebra, we obtain:
3
∂v ∂v
−∆p = ∑ i j.
∂xj ∂xi
i,j=1

We have used the notation: (xi)i=1,2,3 = (x, y, z). This means that the pressure field is a given function of the
velocities at the same instant time. Any change in the velocity field at a position x affects the pressure field
immediately everywhere else. This is the subtle reason why no initial and boundary conditions have
been mentioned for the pressure field in the statement of the problem at the beginning of this section.
Otherwise the problem would be ill-posed. This effect is a direct consequence of the incompressible
condition, which assumes that the sound speed is infinite compared to the magnitude of the flow speed.
This implies that sound waves can carry any perturbations of the pressure field (p) instantaneously
across the entire volume of the fluid.

58
Finally, we conclude this presentation of the millennium problem by a remark concerning the interplay
between Euler’s and Navier-Stokes equations. The statement of the millennium problem is related
only to Navier-Stokes equations in 3-dimensional space. However, as we have seen, the case of small
Reynolds numbers is not difficult, when the quadratic term can be neglected, or at least do not dominate
the evolution in time of the velocity field, while the case of high Reynolds numbers is really hard. Indeed, in
this case, turbulent effects can appear that could lead to divergent solutions: that’s what we need to
understand. That’s the reason why the Euler’s formalism can be thought of as a good ‘laboratory’ to
approach the Navier- Stokes one in the latter case. Indeed, we can expect that, under some
conditions, a result which is valid for any finite, but very large, Reynolds number is also compatible with
results concerning infinite Reynolds number (Euler’s formalism). This is not a general statement but it
justifies that we can gain ideas from the infinite Reynolds number scenario to enrich the discussion of
the millennium problem.

§14. Bounds and partial proofs

As it is clear from the statement of the millennium problem (§14), we can gain a deeper view of how to
approach it by proper definitions of bounds on the norms of functions or their derivatives.
Assuming that we have such definitions, it may be possible to apply this formalism to (NSI)
equations in order to find upper limit to the magnitude of velocity field and its derivatives, and to relate
such results to energy dissipation. This is the purpose of this section to give a precise content to the
formalism and answer as much as possible the last open points.

We recall some basic definitions and notations for continuous real valued functions defined on a set Ω
(that can be the set of real numbers IR or an interval [a, b] of IR), f: Ω → IR.
All the mathematics we introduce in this section is useful for the millennium problem. We present them
in the simplest way, avoiding the technical jargon as much as possible. In order to simplify the
presentation, we assume first that f depends on only one variable x (a real

59
number contained in Ω). When f is measurable on Ω, we can define a norm of the function f: Ω → IR as:
1
p
‖f‖Lp(Ω) = (∫ |f(x)|pdx) with p ≥ 1.
Ω

This defines in parallel the set of functions for which ‖f‖Lp(Ω) is finite:
Lp(Ω) = {f: Ω → IR, ‖f‖Lp(Ω) < ∞}.

Note that ‖f‖Lp(Ω) is really a norm in the mathematical sense, with Lp(Ω) being a Hilbert set. It is called a
Lebesgue’s set. We do not need to enter into the details of these notions here. In words, it means that
Lp(Ω) is an abstract vector set (of functions) that allows length and angle to be measured. In this sense, the
norm defined above represents a length in Lp(Ω). In what follows, what we need is to consider mainly
the case: p=2, functions which are square- integrable. Then, we get for the norm of f: Ω → IR
(squared):

‖f‖L²(Ω)
2
= ∫ |f(x)|2dx.
Ω

The related set of square-integrable functions is:

L2(Ω) = {f: Ω → IR, ‖f‖L²(Ω) < ∞}.
We see that these definitions of L2(Ω) and ‖f‖2 represent
L²(Ω) a kind of generalization of what is well known

for the algebra of vectors. These definitions can be extended trivially to functions that depend on 2
or 3 or even more real variables, defined on spaces Ω of 2, 3 or
more dimensions.

In the context of the millennium problem, we are searching for velocity fields (solutions of partial
differential equations if they exist) that must belong to the set of functions L2(Ω), where Ω represents
the 3-dimensional space of coordinates IR3. Indeed, the kinetic energy is required to be bounded over
the all space, which means that ∫‖𝚽(𝚽, t)‖²d3𝚽 is finite for t ≥ 0, where
∫‖𝚽(𝚽, t)‖²d3𝚽 = ∫[v2(𝚽, t) + v2(𝚽, t) + v2(𝚽, t)] d3𝚽.
x y z

60
 3
We write the integral over space of the velocity vector squared (∫‖𝚽(𝚽, t)‖𝚽d 𝚽) as ‖𝚽‖2 according
L2to
(Ω)

the definition above, extended to the 3-dimensional case. In addition to the fact
that ‖𝚽‖22( ) must remain finite, as the equations (NSI) contain also derivatives of the
LΩ

velocity field over space coordinates, these derivatives for regular solutions, if they exist, must also
remain of finite magnitude ( length or norm in L2(Ω)) up to a certain order that we discuss later. We write
generically ∂mvI the derivative of order m for the component i of the velocity field. Precisely, let (m1, m,
m3) be a set of 3 integers, such that
m = m1 + m2 + m3.
Then:
∂mvi = ∂mv = ∂m1 ∂m2 ∂m3 v .
𝚽 i x y z i

Where we use the standard notation: ∂m1 ∂ m1 ∂

.
x = (∂x) . In the following, we use also: ∂i = ∂x
i

 For m=0, this is simply the component i of the velocity field ∂0vi = vi and:
‖∂0𝚽‖2 = ∫[v2 + v2 + v2] d3𝚽.
L²(Ω) x y z

 For m=1:
3 2
‖∂1 𝚽‖2 = ∫ ∑ |∂ v | d3 𝚽.
L²(Ω) ij
i,j=1

This represents the length (norm) squared of the first derivative (1-gradient) of the velocity field
in L2(Ω). Clearly, this length (or magnitude) must remain finite for all times for a regular
solution of (NSI) following the millennium problem.
 In general the norm in L2(Ω) of the derivative of order m (m-gradient) of the velocity field reads:
3 2
‖∂m𝚽‖2 = ∫∑ ∑ |∂m1 ∂m2 ∂m3v | d3𝚽 (with m + m + m = m)
L²(Ω) x y z j 1 2 3
j=1 m1,m2,m3

61
Therefore, following these definitions, it is useful to define special sets of functions (here the velocity
fields 𝚽(𝚽, t): Ω × [0, ∞[→ IR3) in terms of the norms in L2(Ω) of their derivatives, namely a category
of what Sobolev’s sets:
Hs(Ω) = {𝚽: Ω × [0, ∞[→ IR3 such that 𝚽 ∈ L2(Ω) and ∀ m ≤ s, ∂m𝚽 ∈ L2(Ω)}.

Here,  is the 3-dimensional space IR3 or a subset of it. This definition corresponds to Sobolev’s
sets of rank s (integer), defined on L2(Ω), with the definition of the derivatives as written above. A norm
(squared) in Hs(Ω) can then be expressed as:
2
‖𝚽‖H2Ωs( ) = ∑‖∂k 𝚽‖ L2(Ω)
k≤s

We write explicitly some simple examples for single real valued functions (with =IR):
o H0(Ω) = L2(Ω)
o H1(Ω) = {f: Ω → IR such that f ∈ L2(Ω) and ∂f ∈ L2(Ω)}
o H2(Ω) = {f: Ω → IR such that f ∈ L2(Ω), ∂f ∈ L2(Ω) and ∂²f ∈ L2(Ω)}
o Hm(Ω) = {f: Ω → IR such that f ∈ L2(Ω), ∂f ∈ L2(Ω), … , ∂mf ∈ L2(Ω)} (m integer)
2
o ‖f‖2 m = ∑m ‖∂kf‖ = ‖∂0f‖2 + ‖∂1f‖2 + ⋯+ ‖∂mf‖2
H (Ω) k=0 L2(Ω) L2(Ω) L2(Ω) L2(Ω)

Note that Hs(Ω) can also be defined for real values of s. In this case, we need to generalize the formula
above. For simplicity, we provide the new definition for single real valued function:

‖f‖2 s( ) = ∫(1 + |ξ|2)s|̂f(ξ)| ²dξ.

H Ω

Here, f̂ is the Fourier transform of f: Ω → IR. Extension of the definition to dimension 3 is immediate.

Exercise: We consider a single real valued function, f: Ω = IR → IR. We note f̂ its Fourier transform.
Prove that there exist 2 real numbers c1 and c2 such that:
2
𝚽1 ∫(1 + |ξ| 2 )s ̂
|f(ξ)| ²dξ ≤ ‖f‖ 2
Hs( Ω
) ≤ 𝚽2∫(1 + |ξ| 2 )s
|f(ξ)| dξ with s integer.

62
 2
Remark: based this property, it is then possible to prove that ∫(1 +|ξ|2)s|̂f(ξ)| dξ is equivalent to
the standard norm (squared) in Hs(Ω) even when s is a real number.
Solution: a few hints. With ∂̂α f(ξ) = iα ξα ̂f(ξ), we can show that:
‖f‖ 2Hs ( Ω) = ∫{∑ 𝚽≤𝚽|ξ
α |2 ̂
}|f(ξ)|² dξ.
Then, we need to find upper and lower bounds of the term ∑𝚽≤𝚽|ξα|2. They are given by terms
proportional to (1 + |ξ|2)s.

A final important definition, clearly needed to approach the millennium problem, is the set of bounded and
measurable velocity fields on a set Ω = IR3 at a given time t. We define:
‖𝚽‖L∞(Ω) = {smallest C ≥ 0, ‖𝚽(𝚽, t)‖ ≤ C for almost every 𝚽 in Ω} = ‖𝚽‖L∞(Ω)(t).

Then, the set of velocity fields bounded in Ω (for a given instant t), L∞(Ω), is defined as the set of
velocity fields such that: ‖𝚽‖L∞(Ω)(t) < ∞. Obviously, the same definition holds also for any derivative of
the velocity field. For the gradient of the velocity fields, it reads:

‖∂𝚽‖L∞(Ω) = {smallest C ≥ 0, ‖∂𝚽(𝚽, t)‖ ≤ C for almost every 𝚽 in Ω} = max‖∂𝚽(𝚽, t)‖.

𝚽∈ IR3

In practice, Sobolev’s sets of functions are useful because they inform us that a function and its
derivatives up to a certain order belongs to L2(Ω). Then, it would be interesting to conclude from
this property that the functions belonging to a particular Sobolev’s set are (continuously) differentiable
up to a certain order, to be related to the rank of the Sobolev’s set. This kind of statement (if possible)
corresponds to what is called embedding. In general, a set of functions X is said to be embedded in the
set of functions Y (written as

X ⊂ Y) if all elements (functions) in X are also in Y. We start with a trivial example for single real valued
(measurable) functions: L∞(Ω) ⊂ L1(Ω).

Then, L∞(Ω) is embedded in L1(Ω), where L1(Ω) is the Lebesgue’s set of functions such that

(∫Ω |f(x)|dx) is finite. This means also that L1(Ω) contains L∞(Ω). An important consequence that derives
from this definition is that if X ⊂ Y, then there exists a constant C > 0 such that:

63
‖f‖Y ≤ C ‖f‖X. The last inequality can also be taken as a definition of X ⊂ Y. In the following, it will be
clear how to use this notion in very efficient ways.

Some important embedding properties are stated below:

(1) s < t implies that: Ht(Ω) ⊂ Hs(Ω)

(2) Hs+k(Ω) ⊂ Ck(Ω) for s > D (where k ≥ 1 and D is the dimension of space Ω),
2

C k(Ω)
= set of functions k − times differentiable in Ω, with
k
‖𝚽‖Ck(Ω) = ∑ max‖∂p𝚽(𝚽)‖.
𝚽∈ Ω
p=1

Summary of the mathematical definitions and properties for 3-dimensional set :

L2(Ω) = {𝚽: Ω × [0, ∞[→ IR3, ‖𝚽‖L²(Ω) < ∞} with ‖𝚽‖2 L²(Ω) = ∫ ‖𝚽(𝚽, t)‖2d3𝚽.
Ω

s(Ω)
𝚽: Ω × [0, ∞[→ IR3 such that 𝚽 ∈ L2(Ω) and ∀ m = m1 + m2 + m3 ≤ s,
H ={ m m m }
∂ m𝚽 = ∂ x 1 ∂ y 2 ∂ z 3 𝚽 ∈ L2(Ω)

2
with ‖𝚽‖2Hs (Ω) = ∑‖∂k 𝚽‖ L2(Ω)
.
k≤s

‖𝚽‖L∞(Ω)(t) = the upper value of ‖𝚽(𝚽, t)‖ in Ω.

Embedding:
X ⊂ Y if all elements (functions) in X are also in Y. If X ⊂ Y,
then there exists a constant C > 0 such that: ‖f‖Y ≤ C ‖f‖X.

3 D
Hs+k(Ω) ⊂ Ck(Ω) for s > = (where k ≥ 1 and D is the dimension of space Ω).
2 2

64
It is time to use the mathematical notions exposed above in order to progress in the
understanding of the millennium problem. However, this would be too hard to start directly with Navier-
Stokes equations. That’s the reason why we apply first the above definitions and properties to the case of
(incompressible) Euler’s equations. We discuss (mathematically) 2 issues related to regularity of
solutions of the Euler’s equations: (i) the local existence and uniqueness of such solutions and (ii)
the role of vorticity. And then, we discuss some mathematical issues for the case of the Navier-Stokes
equations (in the presence of viscosity). We pose: =IR3 (unless explicitly stated otherwise) and
=constant=1.

Also, we write L2 for L2(Ω) and similarly for other sets of functions used hereafter.

A last comment is needed before starting the discussion. We often need to estimate differential

inequalities like dX ≤ b(t)X(t), where X and b are positive functions. The last
dt
t
differential inequality can be solved, it gives: X(t ≥ 0) ≤ X(0)exp[∫ b(s)ds].
0
In particular if b(t) is a
constant (b) strictly positive: X(t ≥ 0) ≤ X(0)exp[bt]. In addition, if we know that X(t) is bounded for all
times, the only possible solution is then X(t ≥ 0) = 0.

(i) We discuss the proof of the local existence and uniqueness of solutions of the
(incompressible) Euler equations in 3-dimensional space, under the hypothesis that the
initial condition is smooth and regular with strong decay at infinity (in space). This is not a
simple proof. We intend to show the important steps. The interest is that it clearly
illustrates how to make useful the above definitions and some properties that come with
them.
1. First, from Euler’s equations, it is possible to show that there exists a constant C > 0
such that:
2
d 2
‖∂ 𝚽‖ ≤ C ‖∂𝚽‖
k ∞ . ‖∂k𝚽‖ dt 2 (for k ≥ 1).
L L
L2
d
We let the proof of this relation as this exercise (below). For k=0, ‖∂0𝚽‖22 is
dt L

the time derivative of the kinetic energy which is constant in time (§9 and 10) and thus:

65
 d
‖∂0𝚽‖2 = 0.
dt L2

2. For m>1+3/2=1+D/2, where D is the dimension of space, we have already mentioned

that Hm(Ω) is embedded in C1(Ω), the set of continuous and differentiable
functions on Ω. This implies that there exists a constant A > 0 such that: ‖𝚽‖C1 ≤ A
‖𝚽‖Hm. Knowing that we also have in general (from a
trivial embedding) ‖∂𝚽‖L∞ ≤ B ‖𝚽‖C1 (with the constant B > 0), we can write:
‖∂𝚽‖L∞ ≤ B‖𝚽‖C1 ≤ A. B ‖𝚽‖Hm.
2
3. Then, using the definition of the norm in H m, ‖𝚽‖2 m = ∑ ‖∂k𝚽‖ , as well as
H k≤m L2

items (1) and (2), it can easily be seen that there exists a constant C′ > 0 such that:
1d
‖𝚽‖2 m ≤ C ′ ‖𝚽‖3 with m > 1 + 3/2 = 1 + D/2.
H Hm
2 dt
Equivalently, we can write:
d
‖𝚽‖ Hm ≤ C′ ‖𝚽‖2 Hm .
dt
4. This implies that 𝚽(𝚽, t) (solution of the Euler’s equations) is bounded in H m × [0, T]. Its
upper bound is given by (see §7 applied to differential inequalities):
‖𝚽0‖Hm 1
for T ≤ T∗ = , m > 1 + 3/2 = 1 + D/2.
1 − C′T‖𝚽0‖Hm C′‖𝚽0‖Hm
This proves the local existence.
5. The uniqueness can be shown by a simple reasoning. Assuming that there exist 2
solutions of the Euler’s equations 𝚽1 and 𝚽2, it is then possible to show that there
exists a constant K > 0 such that:
d 2 2

‖𝚽1 − 𝚽2‖L2 ≤ K [‖𝚽1‖Hm + ‖𝚽2‖Hm ]‖𝚽1 − 𝚽2‖L2 .

dt
As we know that K [‖𝚽1‖Hm + ‖𝚽2‖Hm ] > 0, this implies that 𝚽1 = 𝚽2.

(ii) Next, we discuss the role of vorticity on the regularity of solutions of the Euler’s equations.
We start from the inequality (i.1):

66
 d 2
‖∂k𝚽‖ ≤ C ‖∂𝚽‖ L∞ ‖∂k𝚽‖ 2
2
(for k ≥ 1).
dt L2 L

Or equivalently:
d
‖𝚽‖Hm ≤ K ‖∂𝚽‖L∞ ‖𝚽‖Hm (K > 0).
dt
Here m ≥ 1, as we do not need the embedding relation between ‖∂𝚽‖L∞ and
‖𝚽‖Hm which requires m > 1 + 3/2 [see (i.2)].
Note that ‖∂𝚽‖L∞ = ‖∂𝚽‖L∞ (t) and we have no way to guarantee that ‖∂𝚽‖L∞ (t) <
∞ at all times. The only statement that we have made (i.4) is that there is a finite time up to
which the velocity field is regular (and not blowing up). Indeed, we obtain from the
above the inequality:
t
‖𝚽‖Hm ≤ ‖𝚽0‖Hm exp [K ∫ ‖∂𝚽‖L∞ dt].
0

Thus, as long as ‖∂𝚽‖L∞ < ∞ in [0, T*], the velocity field 𝚽(𝚽, t), solution of Euler’s equations,
exists up to T*. Starting from these relations, it is possible to prove another important
theorem concerning solutions of Euler’s equations (from Beale, Kato and Majda in 1984):

We consider the velocity field 𝚽(𝚽, t), solution of Euler’s equations in 3-

dimensional space, such that 𝚽(𝚽, t) belongs to H m (m>1+3/2) for all t ∈ [0,
T ∗ ]. We write 𝚽 = 𝚽 × 𝚽, the vorticity of this velocity field. Then, T∗ is the
maximum time for the velocity field to be in the above function
T∗ ∗
class if and only if ∫ ‖𝚽‖L∞ dt = ∞. In other words, ∫0 T
‖𝚽‖L∞ dt = ∞
0

is a necessary and sufficient condition for blow-up at time T* for 3-

dimensional Euler’s equations. We let the proof of this theorem as an exercise
(below).
An immediate consequence of this theorem is that there is no finite time blow-up in
2-dimensional Euler’s equations assuming that the initial vorticity field is
bounded. This can be shown in various ways: we let this

67
 as an exercise (below). This implies the existence of global regular
solutions in this context. According to §8 or §9, in 2D, the vorticity is
conserved as a scalar.

Exercise: Consider the velocity field 𝚽(𝚽, t) solution of the (incompressible) Euler equations in 3-
dimensional space, under the hypothesis that the initial condition is smooth and regular
with strong decay at infinity (in space).
2 2
Prove the inequality (i.1): d ‖∂k𝚽‖ ≤ C ‖∂𝚽‖ L∞
. ‖∂k𝚽‖ (for k ≥ 1).
dt L2 L2

A similar proof can easily be done in the case of periodic boundary conditions.
Solution: We start with k=1. We re-write ‖∂1𝚽‖22 as:
L
3 3
2 2
‖∂1 𝚽‖ L2 = ∫ ∑|∂ v
2
|i j dV = ∑ ∫|∂ v | dV.
ij
i,j=1 i,j=1

The integral is taken over the volume of the 3-dimensional space (), knowing that the velocity field
is strongly decaying at infinity (in space). Then, we take the time derivatives of this quantity:
3

1d ∂
‖∂1𝚽‖2L2= ∑ ∫ ∂ v ∂i j v idV. j
2 dt i,j=1
∂t

The term in ∂ vj can be obtained from Euler’s equations in the absence of external field (in
∂t

which we take =constant=1):

∂ 3
v + ∑ v ∂ v + ∂ p = 0.
j α α j j
∂t α=1

In the following, we do not write the sum symbol for repeated indices, this is implicit. This
means that ∑3 vα ∂α is written (simply) as vα ∂α.
α=1

Euler’s equations become: ∂ vj + vα ∂αvj + ∂jp = 0. Next, we introduce ∂ vj in the expression

∂t ∂t
1d 2
of the quantity ‖∂1𝚽‖ 2.
2 dt L

68
We obtain:
1d 1 2 3
‖∂ 𝚽‖ = ∑ − ∫ ∂ v ∂ (v ∂ v + ∂ p) dV
L2 iji α α j j
2 dt
i,j=1

= − ∫ ∂ivj ∂i(vα ∂αvj + ∂jp) dV. (i and j are also repeated indices)

With:

∫ ∂ivj ∂i(vα ∂αvj + ∂jp) dV = ∫[∂ivj ∂ivα ∂αvj + ∂ivjvα ∂i ∂αvj + ∂ivj ∂i ∂jp] dV.

By integration by parts, with the condition of incompressibility ∂αvα = 0 and the fact that the velocity field (at
thus its derivatives) is zero at infinity (boundary surface), we can show very easily that:

∫[∂ivjvα ∂i ∂αvj + ∂ivj ∂i ∂jp] dV = 0.

Explicitly for one term:

∫[∂ivjvα ∂i ∂αvj] dV = − ∫[∂α[∂ivjvα] ∂ivj] dV = − ∫[∂α[∂ivj] vα ∂ivj] dV

1 1
=− ∫[∂α[(∂ivj)²] vα] = ∫[[(∂ivj)²] ∂αvα] = 0.
2 2
Then, we are left with the relation following Euler’s equations:
1d 1 2
‖∂ 𝚽‖ = − ∫[∂ v ∂ v ∂ v ] dV.
L2 i j i α α j
2 dt
The term ∫[∂ivj ∂ivα ∂αvj] dV looks complicated. Also, it does not have a definite sign. However, we
can always write:

− ∫[∂ivj ∂ivα ∂αvj] dV ≤ |∫[∂ivj ∂ivα ∂αvj] dV|.

It follows that:
1d
‖∂1𝚽‖2 ≤ |∫[∂ v ∂ v ∂ v ] dV|.
L2 ijiααj
2 dt

In order to conclude the proof for k=1, it remains to show that ‖∂𝚽‖L∞. ‖∂1𝚽‖22 is a properL upper bound
for the term on the right hand side. This term is built up by 3 gradients of the

69
velocity: we can bound 2 of them by their Euclidian lengths and one by ‖∂𝚽‖L∞ (by definition of ‖. ‖L∞).
Then, the term ‖∂𝚽‖L∞ comes out of the integral. Altogether, we obtain:
3
|∫[∂ v ∂ v ∂ v ] dV| ≤ A. ‖∂𝚽‖ ∞ 2 = A. ‖∂𝚽‖
. ∑ ∫|∂ v | dV ∞ . ‖∂1𝚽‖2 .
ijiααj L ij L L2
i,j=1

Here, A is a strictly positive constant. This completes the proof for k=1.

We now consider the general case: k ≥ 1. The proof follows exactly the same development as the case
k=1. Indeed, we have:
2 2
‖∂k𝚽‖ = ∫ ∑3 ∑ |∂kv | dV with ∂k = ∂k1 ∂k2 ∂k3 (and k = k + k + k ).
L2 𝚽=1 k1,k2,k3 j x y z 1 2 3

It is easier to keep the generic notation ∂ k in the formula. Then, we can easily show:
1 d k 2
‖∂ 𝚽‖ ≤ |∫[∂kvj ∂kvα ∂αvj] dV|.
2 dt L2

Again in a similar way as before, we can bound the 2 k-gradients by their Euclidian lengths and the 1-
gradient by ‖∂𝚽‖L∞ . We obtain finally the inequality:

2 2
d
dt
‖∂k 𝚽‖ 2 ≤ C ‖∂𝚽‖L∞ . ‖∂ k 𝚽‖ 2 (for k ≥ 1), which completes the proof (k ≥ 1).
L L

Exercise: Justify the result from Beale, Kato, and Majda in 1984: we consider the velocity field
𝚽(𝚽, t), solution of Euler’s equations in 3-dimensional space, such that 𝚽(𝚽, t) belongs to H m
(m>1+3/2) for all t ∈ [0, T∗ ].
We write the vorticity 𝚽 = 𝚽 × 𝚽(., t). Then, T ∗ is the maximum time for the velocity field to
T∗
be in the above function class if and only if ∫ ‖𝚽‖ L∞ dt = ∞.
0

Hint: we assume that the following relation holds:

t
‖∂𝚽‖L∞ ≤ ‖∂𝚽0‖L∞ exp [C ∫ ‖∂𝚽‖L∞ dt] (C a positive constant).
0

70
 T∗
Solution: The first part of the proof is trivial: if ∫ 0
‖𝚽‖L∞ dt = ∞,
this implies that ‖𝚽‖ ∞ (t)
L
→∞, and
∗
obviously the velocity function cannot belong to
t→T
m ∗
H (m>1+3/2) for all t ∈ [0, T ]. For the second part of the proof, we need to justify that as
T∗
longas‖𝚽‖L∞ < ∞in [0, T*](or∫0 ‖𝚽‖L∞ dt < ∞), then 𝚽(𝚽,t)solution ofEuler’sequations exists up to
T*.

t
From the inequality (i.1): ‖𝚽‖Hm ≤ ‖𝚽0 ‖Hm exp [C ∫0 ‖∂𝚽‖L∞ dt]. We conclude immediately that as long

as ‖∂𝚽‖L∞ < ∞ in [0, T*], the velocity field 𝚽(𝚽, t), solution of Euler’s equations, exists up to T*.

To complete the second part of the proof, we need to show the following relation:
∗ ∗
T T
∫ 0 ‖𝚽‖L∞ dt < ∞ => ∫0 ‖∂𝚽‖L∞ dt < ∞. This is a direct consequence of the relation given
in the statement of the exercise. This completes the proof.

Exercise: Prove that in 2-dimensional space Euler’s equations admit global regular solutions
𝚽(𝚽, t) under initial and boundary conditions (with strong decay at infinity).
Solution: This exercise is a direct consequence from the 2 previous ones. There are 3 simple methods
to achieve the proof. (1) In 2D, we have the great simplification:
∫[∂ivj ∂ivα ∂αvj] dV = 0. It is equivalent to the absence of the vortex stretching term in the
vorticity equation presented in previous sections. This implies that ‖∂1𝚽‖22 2
L (IR )
is conserved

(and thus bounded once the initial condition is), as well as ‖𝚽‖22 2
L (IR )
. This makes the proof.

(2) Equivalently, we can use the previous result on the behavior of ‖𝚽‖L∞ (t).

Once the vorticity is finite at initial time, there is no possibility for the existence of a blow-up
time: ‖𝚽‖L∞(t) → ∗ t→T
∞. This completes the proof. (3) Finally, it is immediate to see that the
Lp(IR²) norm of the vorticity ‖𝚽‖2p L (IR )2 is preserved for p ≥ 2. This is also a guarantee that

71
‖𝚽‖L∞ (t) remains finite for all times. This ensures the existence of global regular solution of Euler’s
equations.

In summary, we have seen how to apply the mathematical notions introduced at the beginning of
this section in order to search for bounds on the velocity field or its derivatives in the case of
(incompressible) Euler’s equations. This has allowed us to prove some fundamental theorems
related to Euler’s equations. We come back to the case of the Navier- Stokes equations (i.e. in the
presence of viscosity), in the context of the millennium problem. The elements of reasoning follow what we
have developed concerning the Euler’s equations.
Namely, we need to check whether ‖𝚽(t)‖22 and ‖∂1𝚽(t)‖22 are bounded for all times or not.
L L

In the case of Euler’s equations in 3-dimensional space, we have seen that ‖∂1𝚽(t)‖22 is only
L locally

bounded up to a critical time T*. However, we could expect that, even if solutions of Euler’s equations
could blow up, solutions of Navier-Stokes would not due to the viscosity term. The idea is that this
term could be sufficient to control the non-linearity of the equations. That’s the reason why the
proofs done for the Euler’s case do not directly imply similar results for the Navier-Stokes equations.
Therefore we need dedicated proofs for the Navier-Stokes case and thus the millennium problem.

We start with the norm of the velocity field (squared): ‖𝚽(t)‖22 . We have
L already discussed the time
evolution of ‖𝚽(t)‖22 in §10, where
L we have shown that:

1d 3 2
‖𝚽‖2 = −η ∫ ∑|∂ v | dV.
ij
2 dt L2
i,j=1

We recall briefly the key points of the proof. First, we have the trivial identity:
1 d ∂
‖𝚽‖22 = ∫ vj vj dV (summation implicit on repeated indices).
2
dt L ∂t

72
The Navier-Stokes equation for the component j of the velocity reads in the absence of external
force:
∂
v + v ∂ v + ∂ p = η∂ ∂v.
α α j j ααj
∂t j

We could take into account a regular external force. This would not modify the line of the
∂ 1d
proof and the conclusions would be similar. Including vj into ‖𝚽‖22 leads to:
∂t 2 dt L

1d
‖𝚽‖2 = ∫ v [−v ∂ v − ∂ p + η∂ ∂ v ] dV.
L2 j α α j j ααj
2 dt

After integrations by parts with vanishing velocity at boundaries and the incompressibility condition, it is
trivial to show that:

∫[−vjvα ∂αvj − vj ∂jp] dV = 0.

Again, after integration by parts, the last term becomes:

3
2
∫[ηv ∂j ∂αv α] dV
j = −η ∫ ∑ [∂ v ] dV.
αj
α,j=1

This gives the result. This term represents the rate of dissipation of the energy. Therefore the
T 2
cumulative energy dissipation up to time T is equal to: η ∫ ∫ ∑3 |∂ v | dV dt.
t=0 i,j=1 ij

We obtain:

3
1 2 T T
‖𝚽‖ 2 (T) = E(0) − η ∫ ∫ ∑|∂i vj | dV
2
dt = E(0) − η ∫ ‖∂1 𝚽‖22 dt.
L L
2 t=0 t=0
i,j=1

1
In conclusion, ‖𝚽‖22(t) is bounded by the initial energy E(0) for all times as well as the
2 L
T 2
cumulative energy dissipation (as E(0) − η ∫ ∫ ∑3 |∂ v | dV dt ≥ 0). Note that for
t=0 i,j=1 ij

73
Euler’s equations, the energy is a constant of motion, thus there is no implied bound on the
2
space time integral of ∑3 |∂ v | .
i,j=1 ij

The situation looks more favorable for Navier-Stokes equations at this step.
d
We now discuss the other point: the time evolution of the quantity ‖∂1𝚽‖22 . The beginning
dt L

of the computation is easy:

1d ∂
‖∂1𝚽‖2 = ∫ ∂ v ∂ v dV = ∫ ∂ v ∂ [−v ∂ v −∂ p + η∂ ∂ v ] dV.
2 iji j iji α α j j ααj
2 dt L ∂t

Following similar arguments as above, we can simplify this expression into:

1d
‖∂1𝚽‖2 = − ∫ ∂ v ∂ v ∂ v dV − η ∫ ∂ ∂ v ∂ ∂ v dV
L2 ijiααj ii j ααj
2 dt

= − ∫ ∂ivj ∂ivα ∂αvj dV − η‖𝚽²𝚽‖22 .L

With the notation:

‖𝚽²𝚽‖2L2 = ∫ ∂i ∂ivj ∂α ∂αvj dV = ∫‖𝚽²𝚽‖2 dV (can also be written as ∫‖∆𝚽‖2 dV).

Here, the norm under the integral is the standard vector norm. We obtain at this step of the calculations:
1d
‖∂1𝚽‖2 + η‖𝚽²𝚽‖2 = − ∫ ∂ v ∂ v ∂ v dV.
L 2 L2 ij iααj
2 dt

The integral on the right hand side is similar to the one obtained in the developments done for Euler’s
equations. There is a new term (on the left hand side) related to the viscosity.

We first make a remark concerning the 2-dimensional case: in 2D, ∫ ∂ivj ∂ivα ∂αvj dV = 0.

74
 1d
Then we obtain (in 2D): ‖∂1𝚽‖22 2 + η‖𝚽²𝚽‖22 2 = 0. We conclude that
2 dt L (IR ) L (IR )
‖∂1𝚽‖22 2 (t) and also ‖𝚽 × 𝚽‖22 2 (t) are uniformly bounded at all times. This is the proof
L (IR ) L (IR )

of the existence of global regular solutions in 2D. This is a consequence from the fact that the term ∫ ∂ivj ∂ivα
∂αvj dV is directly related to the vortex stretching of the vorticity, which disappears in 2D (see
previous sections).

In dimension 3 (3D), this is more complicated. After some algebra that we admit here, we can write:

13 3 2
η 6

− ∫ ∂ivj ∂ivα ∂αvj dV ≤ C ‖∂ 𝚽‖22 ‖𝚽𝚽𝚽‖22 ≤ ‖𝚽𝚽𝚽‖ 2 + 𝚽η‖∂1𝚽‖L2 (C and 𝚽η > 0).
L L 2 L

Note that the algebra needed in order to derive the last formula is presented in the appendix to this
document. It requires the successive use of Hölder’s, Sobolev’s and scaled Young’s inequalities.

Then:
1d
‖∂1𝚽‖2 η 𝚽 2 ‖ 1 ‖6 .
L2 ≤ − ‖𝚽 𝚽‖L2 + 𝚽η ∂ 𝚽 L2
2 dt 2

This inequality can be transformed using the relation:

‖∂1𝚽‖2L2 = ∫ ∂ivj ∂ivj dV = − ∫ vj ∂i ∂ivj dV ≤ ‖𝚽‖ L2

L2
.
‖𝚽𝚽𝚽‖

Finally:

1d η ‖∂1𝚽‖4L2
‖∂1𝚽‖22 ≤ − + Cη‖∂1𝚽‖62 .
2 dt L
2 ‖𝚽‖2L2 L

In particular, the last relation leads to the inequality:

1d
‖∂1𝚽‖2 ≤ C ‖∂1𝚽‖6 .
L2 η L2
2 dt

75
This implies:

‖∂1𝚽‖22 (t) ≤ ‖∂1𝚽‖2L2 (0)

.
L 1
[1 − 2Cη‖∂1𝚽‖42 (0)t]
L
2

This proves the local existence of solutions of the Navier-Stokes equations (similarly to the previous
result obtained for Euler’s equations) up to the critical time T* such that:

1
T∗ = .
[2Cη‖∂1𝚽‖4L(2 )0 ]

4
Note that this is because the positive term Cη‖∂1𝚽‖62 may be too large compared to − η ‖∂1𝚽‖ L2
2
L 2 ‖𝚽‖L2

that we may observe a finite-time blow-up of ‖∂1𝚽‖22 (t). In L

summary, we have a set of coupled inequalities:

T
1 1d η ‖∂1𝚽‖4L2
‖ ‖2 ( ) ( ) ‖ 1 ‖2 2 6
𝚽 L2 T = E 0 − η ∫ ∂ 𝚽 L2 dt ‖∂1𝚽‖L2 ≤ − 2 + Cη‖∂1𝚽‖L2 .
2 t=0
2 dt 2 ‖𝚽‖L2
≤ E(0)

It is important to notice that if there exists a time T for which:

4
η ‖∂1𝚽‖ 2 6 2

− L (T) + Cη‖∂1𝚽‖L2(T) > 0. Then ‖∂1𝚽‖L2(t) can only decrease for t ≥ T, and the
2 ‖𝚽‖22
L

following inequality holds:

η ‖∂1𝚽‖42
L
− (t ≥ T) + Cη‖∂1𝚽‖62L(t ≥ T) > 0.
2 ‖𝚽‖L 22

This means that if the initial conditions are such that:

4
η ‖∂1𝚽‖ 2 6

− L (t = 0) + Cη‖∂1𝚽‖L2 (t = 0) > 0.
2 ‖𝚽‖22
L

76
Then the solution stays regular for all times. Stated otherwise, for small enough initial conditions
‖𝚽0‖22 ‖∂1𝚽0‖22 (given
L
by theL above inequality), ‖∂1𝚽‖22 (t) is finite for all times,
L

this ensures the existence (and uniqueness) of global regular solutions.

Note that instead of searching for inequalities for ‖∂1𝚽‖22, we can

L work with the norm of the vorticity ‖𝚽 =

𝚽 × 𝚽‖22 . The results Lare absolutely equivalent, but it is interesting to prove this rigorously.
Following a similar developments as previously, we can show:
1d η
‖𝚽‖2 ≤ − ‖∂1𝚽‖2 + C ‖𝚽‖6 .
L 2
L2 η L2
2 dt 2
Using the inequality ‖𝚽‖42 ≤ B ‖𝚽‖22 ‖∂1𝚽‖22 (B > 0) (see appendix to this document), we
L L L

get:
1d η
‖𝚽‖2 ≤ − ‖∂1𝚽‖2 + B. C ‖𝚽‖2 ‖𝚽‖2 ‖∂1𝚽‖2 .
L2 L2 η L2 L2 L2
2 dt 2

Altogether:
1d
‖𝚽‖2 ≤ [B.C ‖𝚽‖2 ‖𝚽‖2 η ‖2 .
‖1
L2 η L2 L2 − ] ∂𝚽 L2
2 dt 2

1d
Then, the following inequality holds ‖𝚽‖22 ≤ 0 provided that:
2 dt L
η
B. C ‖𝚽 ‖2 ‖𝚽 ‖2
η 0 L2 0 L2 ≤ .
2
1d
Since ‖𝚽‖22(t) ≤ ‖𝚽0‖22, the inequality ‖𝚽‖22 ≤ 0 implies that ‖𝚽‖22(t) ≤ ‖𝚽0‖22 for
L L 2 dt L L L
all times. Obviously, as ‖𝚽0‖22 = ‖∂1𝚽0‖22 , the above inequality is exactly equivalent to the
L L
conclusion on the smallness of ‖𝚽0‖22 ‖∂1𝚽0‖22 .
L L

77
Finally, we discuss what would be needed to obtain regular solutions for any arbitrary large initial
conditions. We already know that ‖𝚽‖22 (t) is bounded
L for all times. We check at which conditions, we
could find that ‖∂1𝚽‖22 is also bounded atL all times. For this, we can express the upper bound of the 3
gradients of the velocity in a specific way. As previously, we start with:
1d
‖∂1𝚽‖2 + η‖𝚽²𝚽‖2 = − ∫ ∂ v ∂ v ∂ v dV.
L 2 L2 ij iααj
2 dt

After some algebra (admitted here), we can obtain the following inequality:

−∫∂ v ∂ v ∂ v dV ≤ η ‖𝚽 𝚽𝚽‖2 + 𝚽 ‖∂1𝚽‖2 ‖𝚽‖8 (

> 0).
𝚽
i j i α α j 2 η L2 L4 η
2 L

Again, we recall that the algebra needed in order to derive the last formula is presented in the appendix
to this document. It requires the successive use of Hölder’s, Sobolev’s and scaled Young’s inequalities.

Explicitly, we can write:

2
‖𝚽‖84 = ‖𝚽‖84 (t) = [∫‖𝚽(𝚽, t)‖4 d3 𝚽] .
L L

If we could prove that this term is bounded for all times, ‖𝚽‖84(t) ≤ K,Lthen we would get:
1d 1 2
‖∂ 𝚽‖ + η‖𝚽²𝚽‖2 η 𝚽 2 ‖ 1 ‖2 .

L2 L2 ≤ ‖𝚽 𝚽‖L2 + 𝚽η. K ∂ 𝚽 L2
2 dt 2

Or:
1d
‖∂1𝚽‖2 η 𝚽 2 ‖ 1 ‖2 .

L2 ≤ − ‖𝚽 𝚽‖L2 + 𝚽η. K ∂ 𝚽 L2
2 dt 2

This would imply immediately that ‖∂1𝚽‖22 also Lis bounded for all times! And consequently we would
conclude on the regularity of solutions of the Navier-Stokes equations (under the

78
hypothesis of the millennium problem). In fact, this is what we miss in the partial proofs. We would need
to find a bound on the norm of the velocity field, not only in L2(IR3), but also in L4(IR3). An intuitive
explanation is given below.

Beyond all these mathematical inequalities, we try to find an intuitive reason why this is so hard to prove
the global regularity of solutions of the Navier-Stokes equations in the context of the millennium
problem.

A key point is to recall that the Navier-Stokes equations are invariant under the
transformation:

𝚽λ(𝚽, t) = λ𝚽(λ𝚽, λ²t) and pλ(𝚽, t) = λ²pλ(λ𝚽, λ²t).

This means that if the velocity field 𝚽 is a solution of the equations, then the velocity field 𝚽λ
is another acceptable solution (by construction). Also, this implies λ‖𝚽λ‖22 = ‖𝚽‖22 and
L L
2
therefore ‖𝚽 ‖ = λ‖𝚽‖2 . We can think of this transformation, with λ ≫ 1, as taking the
1/λ L2 L2

fine-scale behavior of the velocity field 𝚽 and matching it with an identical (but rescaled and slowed
down) coarse scale behavior of 𝚽1/λ .
Along the mathematical arguments given previously, our goal was to find upper bounds on
T
the maximum energy max‖𝚽‖22 (t) and the cumulative energy dissipation ∫ ‖∂1𝚽‖22 (t) dt.
t≤T L t=0 L

We assume that such bounds exist: we label them respectively as M and C for the field velocity
𝚽. Obviously, for 𝚽1/λ , these bounds become: λM and λC. The last statement means that each time we
have a solution of Navier-Stokes equations with bounds M and C, then the solution
𝚽1/λ is possible, with worsened bounds λM and λC. Blow-up can occur when the solution of
Navier-Stokes equations shifts its energy into increasingly finer and finer scales, thus evolving
more and more rapidly and eventually reaching a singularity in which the scales (in both space and
time) tend towards zero. In such a scenario, we lose obviously the effectiveness of the bounds
(and consequently the control) on the maximum energy and cumulative energy dissipation. For
example, this is possible that at some time, a solution of the equations shifts its energy from a spatial
scale 1/ to 1/2 in a time of order 1/². And if

79
this behavior repeats over and over again, this is clear that the solution is divergent. This simple
argument gives an intuitive view of the problem.

At this point, we can come back to what would happen if we could get a bound on ‖𝚽‖44(t) L
4
for all times. With the relation ‖𝚽 ‖ = 1 ‖𝚽‖4 , we understand that the bound in L4(IR3)
1/λ L4 λ L4

would become increasingly better for the velocity field 𝚽1/λ when the parameter λ is increased.
Under such conditions, it would be possible to understand intuitively that blow- up could not occur.

§15. Fluid mechanics in relativistic Heavy-Ions collisions

We recall that modern aspects of fluid mechanics cover more materials than liquid and gases in the non-
relativistic approximation. As already stated, the idea of exploiting the laws of ideal fluid mechanics to
describe the expansion of the strongly interacting nuclear matter that is formed in high energetic
hadronic collisions was proposed in 1953 by Landau. Of course, in this context of particles moving at
extremely high velocities, fluid mechanics has to be understood in its relativistic limit, which brings some
subtleties to the calculations.

When colliding 2 fast moving heavy ions, built up with many charged particles (picture 1), the idea is
that a zone of high density of charges will be formed (picture 2). Assuming local equilibrium and
depending on the equation of state of this nuclear medium, this overlap region may reach the
conditions that are described by an average (high) density and (high) temperature. The local thermal
equilibrium means that the microscopic collision time scale

80
is much shorter than any macroscopic evolution time scales. The interest of such collisions is that, having
stored part of the available initial energy in compression and thermal excitation, the collisions produce
unique conditions of the nuclear matter not accessible otherwise. In particular, this hot and dense matter
is thought to be composed of strongly interacting quarks and gluons. The next stage of the reaction is the
relaxation of the energy density. The central system is undergoing expansion, into the direction of
largest gradients in density and temperature (pictures 3 and 4), thus reducing its temperature and
density.

Recent experimental results (from 2005 to 2014) provide conclusive evidences that the created hot
and dense matter produced in these collisions behaves collectively and has properties resembling
that of a nearly ideal relativistic fluid, following the theoretical ideas formulated by Landau in 1953. We
can write equations that describe the early stages of the expansion (pictures 2 and 3). The most
important assumption under these equations is that the system can reach local thermodynamic
equilibrium in a very short time. All thermodynamic quantities (see §3) of a fluid element can
then be defined under this hypothesis, in the rest frame of the fluid element. The frame of reference is
thus the local rest frame. Next, we assume that the energy density (E/V) and pressure admit a first
order development around their initial values:
ϵ(t, 𝚽) = ϵ0(t, 𝚽) + δϵ(t, 𝚽) and p(t, 𝚽) = p0(t, 𝚽) + δp(t, 𝚽).

Neglecting second order terms, it can be shown that the relativistic equations of motion read (admitted
here):
∂𝚽 1 𝚽(p).
=−
∂t ϵ +p
The conservation of energy reads:
∂ϵ
= −(ϵ + p) 𝚽. 𝚽.
∂t
The energy density is decreasing as the velocity field is diverging (𝚽. 𝚽 > 𝚽). Similarly, the charge
density (n) conservation reads:

81
 ∂n
= −(n) 𝚽. 𝚽.
∂t
We have 6 unknowns (p, ϵ, n, 𝚽) for 5 equations (stated above). We are missing the 6th equation,
the equation of state:
p = p(n, ϵ).
This closes the system. In practice, this last equation can be expressed using the speed of sound:
∂p(n, ϵ)
cS = .
𝚽ϵ
The last equation is not complete as we need to precise under which line the partial derivative is
taken. This is a subtle issue. However, following ideas developed in the non- relativistic case, we can
show that these lines correspond to: s/n=constant, where s is the entropy density (S/V). We can re-
write the equations of motion as:
∂𝚽 2 𝚽(p)
= − cs .
∂t 1 + c2s p

In fact, the ideal fluid behavior is only approximate: a (small) degree of viscosity is required for a proper
description of the data (pictures 2 and 3) and the late stage of the expansion needs a dedicated
treatment as it appears to be too dissipative for a macroscopic fluid dynamical approach. It must be
described microscopically. The influence of the dissipative stage on measured observables is large as
the system in expansion spends a large fraction of its history in this dissipative stage. Obviously, this
makes the measurements quite difficult to interpret but this is another story.

82
References (books)

Obviously, our presentation of fluid mechanics cannot be exhaustive, in particular concerning real fluids.

To our view, the best reference ever written on fluid mechanics is: L. D. Landau and E. M. Lifshitz
[1968] Fluid Mechanics, Pergamon.

Another useful presentation of general fluid mechanics can be found in: G. K. Bachelor [1974] An
Introduction to Fluid Dynamics, Cambridge University Press, London. For what concerns the regularity
of solutions of Euler’s or Navier Stokes equations, many developments can be found in: P. Constantin
and C. Foias [1988] Navier-Stokes Equations, University of Chicago Press, Chicago and London; R.
Temam [1984] Navier-Stokes Equations, volume 2, Elsevier Science Publishing Company,
Amsterdam; or A. Bertozzi and A.J. Majda [2002] Vorticity and incompressible flow, Cambridge texts
in applied mathematics.

83
 Appendix: useful mathematical inequalities

Basic inequalities
For real numbers: a ≥ 0, b ≥ 0, integers p and q > 0:
1 1 ap b q
+ = 1 → ab ≤ + (Young’s inequality).
p q p q

Proofs: We write: ab = exp[log(a) + log(b)], a ≥ 0, b ≥ 0. Then, by convexity of the exponential

function [e[αx+(1−α)y] ≤ αex + (1 − α)ey], we get:
1 1 1 1 1
ab = exp [ . log(ap) + . log(bq)] ≤ . exp[log(ap)] + . exp[log(bq)] , with
p q p q q
1
= 1 − and a ≥ 0, b ≥ 0.
p

This completes the proof.

Another simple way to make the proof is to consider the real function for positive real numbers:
p
f(x) = x + 1 − x. We can easily show that f(x) ≥ f(1) = 0. Then, using the relation
p q
1/(p−1)
x = a. b , this completes the proof, under the condition 1 + 1 = 1.
p q

From the Young’s inequality, we can deduce immediately the following relation (a, b ≥ 0):
. bq, C = 1 [δp]− . We
q
For any real number δ > 0, there exists C such that ab ≤ δap + C p
δ δ δ q
q q
ap − b
can write equivalently: for any ε > 0, ab ≤ ε p + [ε] p (a, b ≥ 0).
q

From the Young’s inequality, we can derive the important Hölder’s inequality. For sets of real numbers
(ai) and (bi), and positive integers p and q, we have:
1 1
n n p n q
1 1
+ = 1 → ∑|ai bi | ≤ (∑|ai |p ) (∑|bi ) |q (Hölder’s inequality).
p q
i=1 i=1 i=1

84
 1 1
Proof: We define: A = |ai|/(∑n |ai|p)p and B = |bi|/(∑n |bi|q)q. Then, we apply the Young’s
i=1 i=1

inequality to the product AB and sum over the index i. This gives the result, under the
condition: 1 + 1 = 1.
p q

Obviously, this holds also for measurable real valued functions f and g (Ω → IR):
1 1
p p q q 1 1
‖f. g‖L1 (Ω) = ∫|f. g| dx ≤ (∫|f| dx) (∫|g| dx) = ‖f‖Lp(Ω)‖g‖Lq(Ω) (with + = 1).
p q

Here the proof relies on the definitions of A = |f|/‖f‖Lp (Ω) and B = |g|/‖g‖Lp (Ω) .

The Hölder’s inequality, for p=q=2, corresponds to the well-known Cauchy-Schwarz inequality.
It can be formulated explicitly:
(a2 + a2 + ⋯ + a2 )(b2 + b2 + ⋯ + b2 ) ≥ (a b + ⋯ a b )².
1 2 n 1 2 n 1 1 n n

For this particular case, a simple geometric proof can be done. We note u the vector (a1 , a2 , … ,
an ) and v(b1 , b2 , … , bn ), then:

‖𝚽. 𝚽‖2 = ‖𝚽‖2‖𝚽‖2 cos(𝚽, 𝚽) ≤ ‖𝚽‖2‖𝚽‖2.

This gives the result. Similarly, for real valued functions, we obtain (p=q=2):

2
(∫ f. g dx) ≤ (∫ f 2dx) (∫ g2dx) .

Applications of the Hölder’s inequality

For any measurable real valued functions f and g (Ω → IR), the Hölder’s inequality reads:
1 1
p q
‖f. g‖L1 (Ω) = ∫|f. g| dx ≤ (∫|f| p
dx) (∫|g| dx)q

1 1
= ‖f‖Lp(Ω)‖g‖Lq(Ω) (with p, q ≥ 1: + = 1).
p q

85
The integrals are taken over the set Ω. In the most general case, Ω can be equal to the set IRn, where n is
the dimension of space, provided that the integrals are well-defined. The set Ω can also be a bounded
subset of IRn. For n dimensions of space, the integral ∫|f. g| dx becomes
∫|f. g| dV, where dV = dx1 … dxn is the infinitesimal volume element Ω. Also, we have:
1
‖f‖Lp(Ω) = (∫ |f|p dV) . Note that the inequality holds also for the particular case p =
p

∞, then q = 1. We obtain:
‖f. g‖L1(Ω) ≤ ‖f‖L∞(Ω)‖g‖L1(Ω).

Here, ‖f‖L∞(Ω) = {smallest C ≥ 0, f ≤ C for (almost) every point in Ω}. This means that
‖f‖L∞(Ω) is the upper bound of the function f in Ω. The above inequality is thus obvious.

The Hölder’s inequality can be immediately generalized to different cases, assuming that Ω = IRn or a
bounded subset of it.
1
(i) +1+1 = 1 (with p, q, r ≥ 1), f ∈ Lp(Ω), g ∈ Lq(Ω) and h ∈ Lr(Ω), then
p q r

‖f. g. h‖L1(Ω) ≤ ‖f‖Lp(Ω)‖g‖Lq(Ω)‖h‖Lr(Ω).

1
(ii) +1 =1 (with p, q, r ≥ 1), f ∈ Lp(Ω) and g ∈ Lq(Ω), then f. g ∈ Lr(Ω) and
p q r

‖f. g‖Lr(Ω) ≤ ‖f‖Lp(Ω)‖g‖Lq(Ω).

θ 1−θ 1
(iii) + = (with p ≤ r ≤ q and 0 ≤ θ ≤ 1), f ∈ Lp(Ω) ∩ Lq(Ω), then f ∈ Lr(Ω),
p q r

and
‖f‖Lr(Ω) ≤ ‖f‖θp( L Ω )‖f‖
1−θ Lq(Ω)
.

This means that we can interpolate between higher and lower Lp(Ω) sets to get
something in between. That’s the reason why we can call this relation an
interpolation inequality for Lp(Ω) sets.

86
 The proof of the last relation follows from the standard Hölder’s inequality after writing:

‖f‖rr = ∫ |f|r dV = ∫ |f|θr |f|(1−θ)rdV ≤ ‖|f|θr‖ p ‖|f|(1−θ)r‖ q .

L (Ω)
Ω Ω Lθr(Ω) L(1−θ)r(Ω)

This gives the result.

When the set Ω is bounded (in IRn), these inequalities become very useful to derive important links
between Lp(Ω) norms (integrability) and regularity of functions. For example, if Ω is an interval [a, b] in
IR (n=1), then we get the following properties.
(i) If q ≥ p, then there exists a constant K which depends on Ω, p, q such that:
‖f‖Lp(Ω) ≤ K ‖f‖Lq(Ω) ( q ≥ p ≥ 1).
This can also be written as:
Lq(Ω) ⊂ Lp(Ω) ( q ≥ p ≥ 1).
Explicitly, Lq(Ω) is embedded in Lp(Ω). A set of functions X is said to be embedded in the set of
functions Y (written as X ⊂ Y) if all elements (functions) in X are also in
Y. In addition, the relation between the norms in both sets X and Y is as written above.
Namely, Lq(Ω) ⊂ Lp(Ω) ( q ≥ p ≥ 1) is equivalent to the inequality
‖f‖Lp(Ω) ≤ K ‖f‖Lq(Ω),
for some constant K(Ω, p, q). Lq(Ω) is embedded in Lp(Ω) can also be stated as: Lp(Ω)
contains Lq(Ω), or Lp(Ω) ⊃ Lq(Ω) ( q ≥ p ≥ 1). We let this proof as an exercise
(below).
This relation means that higher Lp(Ω) norms can control lower ones
(‖f‖Lp(Ω) ≤ K ‖f‖Lq≥p(Ω)). In particular, this implies that functions that belong to higher Lp(Ω)
spaces are more regular. This is only correct for a bounded set Ω.

(ii) This also implies that, for any bounded set Ω:

L1(Ω) ⊃ Lp≥1(Ω) ⊃ Lq≥p(Ω) ⊃ L∞(Ω).

87
 L1(Ω) is the larger set. In particular for a bounded set L1(Ω) ⊃ L∞(Ω), which means that once ∫Ω

|f| dV is finite, then the function f: Ω → IRn is bounded up.

Exercise: Prove that, for a bounded set Ω and q ≥ p ≥ 1, then there exists a constant K which depends on
Ω, p and q such that: ‖f‖Lp(Ω) ≤ K ‖f‖Lq(Ω) (for q ≥ p).
Solution: We have:
1 1 1
p r q
‖f‖Lp(Ω) = (∫ 1. |f|p dV) ≤ (∫ 1r dV) (∫ |f|q dV)
Ω Ω Ω

with 1 − 1 = 1 (Hölder′s inequality).

p q r
1 1
−
p q
This gives the result, with K = (∫ Ω
1 dV) (since Ω is bounded).

First Sobolev’s inequalities

We now intend to present some links between the integrability and the differentiability of these
functions. This is an essential task as, at the end, we aim to study solutions of partial differential
equations (for example in the context of fluid mechanics). The general idea is that it seems to be more
difficult to obtain the differentiability than the integrability. We need to develop this idea quantitatively
with precise definitions. That’s the reason why we will define some sets of functions which involve and
index for differentiability and another index for integrability.

We recall some basic definitions. In order to present first the formulae using simple notations,
we start with single real valued functions h: IR → IR. We assume that the derivatives of h up to
order k are well-defined. We note ∂mh (m ≤ k), these derivatives.

88
Then, we define what is called a Sobolev’s set Wk,p(IR) as the set of functions h: IR → IR such that h ∈
Lp(IR) and ∂m≤kh ∈ Lp(IR). The second condition means that the derivatives of h, up to order k,
belong to Lp(IR). As mentioned above, one index (k) stands for the differentiability and one
index (p) for the integrability.

We note:
Wk,p(IR) = {h: IR → IR such that ∂mh ∈ Lp(IR) for all 0 ≤ m ≤ k}. (with ∂0h = h)
In parallel, we can define a norm on this set as:
‖h‖p = ∑‖∂m h‖p .
Wk,p(IR) Lp(IR)
m≤k

Next, this definition can be extended immediately to functions f: IRn → IR (where n –the dimensions
of space- can be larger than 1). We write the formulae for the particular case n=3, corresponding to 3-
dimensional spaces (3D).

Let (m1, m, m3) be a set of 3 integers, such that m = m1 + m2 + m3. We assume that the derivatives of
the function f are well-defined up to rank k, using the notation:
m3 m1 ∂ m1
∂m f = ∂m f = ∂ m1 m2
∂ ∂ f {with ∂ =( ) }.
𝚽 x y z x
∂x
The Sobolev’s set Wk,p(IR3) of such functions is defined as:
Wk,p(IR3) = {f: IR3 → IR such that ∂mf ∈ Lp(IR3) for all 0 ≤ m = m1 + m2 + m3 ≤ k}.

In parallel, we can define a norm on this set as:

‖f‖p = ∑ ‖∂mf‖p .
Wk,p(IR3) Lp(IR3)
m1+m2+m3≤k

With:

‖∂Mf‖p =∫ ∑ |∂mf|p dV
Lp(IR3)
IR3 m=m +m +m =M
1 2 3

89
Obviously, we observe that the definitions for dimension 3 (or more generally n) follow exactly
those for dimension 1. Note that the case p = ∞, is also well-defined with the definition:
‖f‖Wk,∞(IRn) = max ‖∂mf‖L∞(IRn) .
m1+⋯+mn ≤k

The particular case p=2 is essential for the following discussion. We write:
Hk(IRn) = W k,2(IRn)
Explicitly:
f:IRn → IR such that f ∈ L2(IRn) and ∂mf ∈ L2(IRn) for all 1
Hk(IRn) = { }.
≤ m = m1 + m2 + ⋯+ mn ≤ k

We also use the name Sobolev’s set for Hk(IRn). Moreover, Hk(IRn) is a Hilbert’s space with the inner
product:

〈 u,v〉 Hk(IRn) = ∫ dV [ ∑ ∂mu ∂mv].

IRn m +⋯+m ≤k
1 n

Where u and v are functions that belong to H k(IRn).

Note that Hs(Ω) can also be defined for real values of s. In this case, we need to generalize the formula
above. For simplicity, we provide the new definition for single real valued function:

‖f‖2 s () = ∫(1 + |ξ|2 )s |f̂(ξ)| ²dξ.

H IR

Here, f̂ is the Fourier transform of the function f. Then the definition of Hs(IR) can (also) be written as:
Hs (IR) = {f: IR → IR such that f ∈ L2 (IR) and (1 + |ξ|2 )s/2 ̂f∈ L2 (IR)}.
Extension to dimension 3 (or n) is immediate.

90
We give below 4 important Sobolev’s inequalities that follow the previous definitions and properties.
In general, any inequality which trades differentiability and integrability of functions is called a
Sobolev’s inequality. We understand below that this trade is one way: we cannot generally sacrifice
integrability to gain differentiability (as already mentioned). In the following, we always assume that the
functions are vanishing at the boundary of Ω. If Ω = IRn, this means that functions are assumed to vanish
at infinity. The theorems are not correct otherwise. Also, we systematically mention whether the
inequalities are given for Ω = IRn or for Ω = {bounded subset of IRn}.

2 2
i. From the relation (1 + |ξ|2 )r|f̂(ξ)| ≤ (1 + |ξ|2 )s |f̂(ξ)| (s ≥ r), it follows the first embedding
theorem for Sobolev’s sets: Hs(Ω) ⊂ Hr(Ω) (s ≥ r).
ii. For any functions, u ∈ W1,p(IRn) (1 ≤ p < n), then u ∈ Lnp/(n−p)(IRn) and there exists a
constant C(n, p) such that:
‖u‖Lnp/(n−p)(IRn) ≤ C(n, p) . ‖∂u‖Lp(IRn).

This implies also the embedding relation:

‖u‖Lnp/(n−p)(IRn) ≤ C(n, p) . ‖u‖W1,p(IRn).

Or equivalently:
np
W 1,p(IRn)
⊂ n−p
L (IRn).
iii. More generally, if u ∈ W1,p(IRn) (p < n) and p ≤ q ≤ np/(n − p), then u ∈ Lq(IRn) and there
exists a constant C(n, q, p) such that:
‖u‖Lq(IRn) ≤ C(n, q, p) . ‖∂u‖Lp(IRn).
For q < np , the embedding relation holds only for a bounded subset of IRn (Ω):
n−p

‖u‖Lq(Ω) ≤ C(n, q, p) . ‖u‖W1,p(Ω).

Or equivalently:
W1,p(Ω) ⊂ Lq(Ω).

91
 iv. The last relation can be generalized to Wk,p(IRn) with the updated result: for any functions,
u ∈ Wk,p(IRn) (kp < n), then u ∈ Lnp/(n−kp)(IRn) and

‖u‖Lnp/(n−kp)(IRn) ≤ C(n, k, p) . ‖∂ku‖ p( n).

L IR

This implies also the embedding relation, for any function u ∈ Wk,p(IRn):
‖u‖Lnp/(n−kp)(IRn) ≤ C(n, p) . ‖u‖Wk,p(IRn).

Or equivalently:
np
Wk,p(IRn) ⊂ Ln−kp(IRn).

A simple view of the above Sobolev’s inequalities can be given assuming a certain behavior for the
function u. Let us assume that u is a bump of size H and length L and this for all dimensions of the
space IRn. This means that:
‖u‖Lq (IRn ) ~[Hq Ln ]1/q and ‖∂u‖Lp (IRn ) ~[(H/L)pLn]1/p .

Then, we can verify at which conditions we can get a Sobolev’s inequality (involving
‖u‖Lq(IRn) and ‖∂u‖Lp(IRn)) for all L and H:

‖u‖Lq(IRn) ≤ C . ‖∂u‖ap( L IRn).

Equivalently, we can write:

HLn/q ≤ C. H a L(n−p)a/p.
This implies that: a=1 and 1 = 1 − 1 or q = pn/(n − p).
q p n

This method provides a general intuitive trick to re-derive quickly Sobolev’s inequalities.

92
Exercise: Prove item (ii). For any functions, u ∈ W1,p(IRn) (1 ≤ p < n), then u ∈ Lnp/(n−p)(IRn)
and there exists a constant C(n, p) such that:
‖u‖Lnp/(n−p)(IRn) ≤ C(n, p) . ‖∂u‖Lp(IRn).

Solution: Essentially, the proof uses the Hölder’s inequality several times. This kind of proof is not
difficult per se but requires a good organization. First, we need to show the result in the simpler case:
p=1. Then, it will be possible to generalize (see later) to the general case for all p values. For the case
p=1, we need to show that, for any function u ∈ W1,1(IRn) (1 < n), there exists a constant C(n) such
that:
‖u‖Ln/(n−1)(IRn) ≤ C(n) . ‖∂u‖L1(IRn).

Explicitly, we need to derive the formula:

n
n−1
∫|u|n/(n−1) dV ≤ C(n) (∫|∂u| dV) .

Where the integrals are taken over the all space IRn and thus, dV is a volume element of IRn. The first
step is to make a derivative appear using a simple identity:
xj
u(x1, … , xn) = ∫ ∂ju(… , aj, … )daj.
−∞

The derivative inside the integral is taken on the jth variable of the function, and this is this variable
which is integrated over. Then, taking the absolute values on both sides, the inequality follows:
xj
|u(x1, … , xn)| ≤ ∫ |∂ju(… , aj, … )|daj.
−∞

Repeating this process n-times for all variables of the function u(x1, …, xn), we obtain a product
from j=1 to n. At this stage, we get:
1/(n−1)
n xj
|u(x1, … , xn)|n/(n−1) ≤ [∏ ∫ |∂ju(… , aj, … )|daj] .
j=1 −∞

As the integrand on the right hand side is positive (absolute value), all integrals can be extended up
to +∞. Hence, we can write:

93
 1
n ∞ n−1 n ∞ 1/(n−1)
|u|n/(n−1) ≤ [∏∫ |∂ju|daj] = ∏[∫ |∂ju|daj] .
j=1 −∞ j=1 −∞

The next step of the proof consists in integrating this inequality, over x1, x2,…,xn, in order to derive a bound
for ∫|u|n/(n−1) dV. We start by integrating over x1:
n 1
∞ ∞ ∞ n−1
∫ |u|n/(n−1) . dx1 ≤ ∫ ∏ [∫ |∂j u|daj ] . dx1 .
−∞ −∞ j=1 −∞

But, for what concerns the integral over x1of the right hand side part of the inequality, all terms that do
not depend on x1are thus constant of x1, and can then be taken out of the
1
∞ n−1
integral over x1. This is obviously the case for [∫ |∂1u|da
−∞
1] in which the variable x1is
integrated over. We obtain:
n 1 1 n 1
∞ ∞ n−1 ∞ n−1 ∞ ∞ n−1
∫ ∏[∫ |∂ju|daj] . dx1 = [∫ |∂1u|da1] ∫ ∏[∫ |∂ju|daj] . dx1.
−∞ j=1 −∞ −∞ −∞ j=2 −∞

This is the stage where the Hölder’s inequality can be applied to

∞ 1
∞ [∫ |∂ju|daj]n−1. dx1. We recall that:
∫ −∞∏n j=2 −∞
p1 p1n 1 1
|∫ f . f . . . f dx| ≤ (∫|f |p | |q with + ⋯+ = 1.
1 2 n 1 dx) … (∫ fn dx)
p1 pn
This leads to:
1
1 n−1
∞ n ∞ n ∞
n−1
∫ ∏ [∫ |∂ju|daj] . dx1 ≤ [∏ ∬ |∂ju|daj dx1] .
−∞ j=2 −∞ j=2 −∞

All together at this stage, we have:

1
1 n ∞ n−1
∞ ∞ n−1
∫ |u|n/(n−1) . dx1 ≤ [∫ |∂1 u|da1 ] . [∏ ∬ |∂j u| dx1 daj ] .
−∞ −∞ j=2 −∞

Following the same strategy, as mentioned above, we now integrate the last inequality over x2:

94
 ∞
∫ |u|n/(n−1) . dx1 dx2
−∞
1 1n 1
∞ n−1 ∞ ∞ n−1 ∞ n−1
≤ [∬ |∂2u|dx1da2] .∫ [∫ |∂1u|da1] ∏[∬ |∂ju|dx1daj] dx2.
−∞ −∞ −∞ j=3 −∞

Again, for what concerns the integral over x2of the right hand side part of the inequality, all terms that do
not depend on x2are thus constant of x2, and can then be taken out of the
1
n−1 ∞
integral over x2. This is the case for [∬ |∂2u|dx
−∞
1da2] in which the variable x2 is
integrated over.
∞ 1 ∞ 1
Be careful: this is not [∬ |∂1u|da1dx2]n−1. The term [∫ |∂1u|da1]n−1 is still in the integral
−∞ −∞

over x2 as this is a function which depends on x2. However, following the same line of process as before,
we can apply the Hölder’s inequality to:
∞ ∞ 1 ∞ 1
∫ [∫ |∂1u|da1]n−1 ∏n [∬ |∂ju|dx1daj]n−1 dx2.
−∞ −∞ j=3 −∞

Then obtain:

∞
∫ |u|n/(n−1) . dx1 dx2
−∞
1
1 1 n n−1
∞ n−1 ∞ n−1 ∞
≤ [∬ |∂2u|dx1da2] . [∬ |∂1u|da1dx2] . [∏ ∬ |∂ju| dx1dx2daj] .
−∞ −∞ j=3 −∞

We can follow exactly the same process by integrating over all the other variables. We end up with:
n
n−1
∫|u|n/(n−1) dV ≤ (∫ ∂u| | dV) .

This completes the proof for p=1. For any p values, we need to prove that:

95
 (n−p)/(np) 1/p
[∫|u|pn/(n−p) dV] ≤ C(n, p) (∫|∂u| dV) p
.

The idea is to simply apply the relation derived for p=1, not to the function u, but to |u|γ.

The following inequality holds:

n n
n−1 n−1
γn/(n−1) γ γ−1
∫|u| dV ≤ (∫|∂|u| | dV) = (γ ∫|u| |∂u| dV) .

Using the Hölder’s inequality, we can express:

1 1
γ−1|∂u| (γ−1)q
q 1 1 p
γ ∫|u| dV ≤ γ [∫|u| dV] [∫|∂u| dV] with + = 1.
p
p q
Bychoosing γn =(γ−1)q,wehave:
n−1
n
γn 1−n/q(n−1) n p( n− 1)
[∫|u|n−1 dV] ≤ γn−1 [∫|∂u|p dV] .

Using the relations 1 + 1 = 1 and γn = (γ − 1)q, wecan see immediately that the last
p q n−1

inequality is equivalent to:

(n−p)/(np) 1/p
[∫|u|pn/(n−p) dV] ≤ C(n, p) (∫|∂u| dV) p
.

Note that it is standard to write:

pn
p∗ = .
n −p
Then, we can write:
1/p∗ 1/p
p∗
[∫|u| dV] ≤ C(n, p) (∫|∂u|p dV) .

96
Poincaré’s inequality
For any functions, u ∈ W1,p(Ω) (Ω bounded subset of IRn), there exists a constant C(n, p) such that:
‖u‖Lp(Ω) ≤ C(n, p) . ‖∂u‖Lp(Ω).
This is called the Poincaré’s inequality. The general idea is that this inequality gives a way to estimate the
Lp(Ω) norm of a function in terms of the Lp(Ω) norm of its derivative. Explicitly, this relation reads:
∂u p
∫ |u(x1, … , xn)|p dV ≤ C(n, p) . ∫ ∑ | | dV .
n
Ω Ω i=1 ∂xi

Further Sobolev’s theorems

We have already presented an interpolation inequality for Lp(Ω) sets: for any
function f ∈ Lp(Ω) ∩ Lq(Ω), then f ∈ Lr(Ω) and ‖f‖Lr(Ω) verifies:
θ 1−θ θ 1−θ 1
‖f‖Lr(Ω) ≤ ‖f‖Lp(Ω)‖f‖Lq(Ω) provided that + =
p q r
with 1 ≤ p ≤ r ≤ q < ∞ and 0 ≤ θ ≤ 1.

Similarly, there exists an interpolation inequality for Sobolev’s sets Hp(IRn). For any real values 0 < s1
≤ s2 < ∞, and for any s ∈ [s1, s2]
such that s = θs1 + (1 − θ)s2 with 0 ≤ θ ≤ 1, we have:

If u ∈ Hs1 (IRn) ∩ Hs2 (IRn), then u ∈ Hs(IRn), and

‖u‖Hs(IRn) ≤ ‖u‖θ s1H(IRn)‖u‖1−θ Hs2 (IR n) (s = θs1 + (1 − θ)s2)
This means that we can interpolate between higher and lower Hs(Ω) sets to get something in between.

We now discuss a final useful inequality, usually called the Gagliardo-Nirenberg-Sobolev’s inequality.
This is also an interpolation inequality, complementary of the previous relation

97
between sets Hs(IRn), Hs1 (IRn) and Hs2 (IRn). It follows closely the interpolation inequality for Lp(Ω)
sets, recalled above.

For any function u ∈ Wl,p(IRn) ∩ Wj,q(IRn), then u ∈ Wk,r(IRn) and there exists a universal constant
C>0 (independent of u) such that:
θ 1−θ 1 k
‖∂ku‖ ≤ C. ‖∂lu‖ ‖∂ju‖

Lr(IRn) Lp(IRn) Lq(IRn)

providedthat −
r n
1 l 1 j
= θ( − ) + (1 − θ) ( − ).
p n q n

This is called the Gagliardo-Nirenberg-Sobolev’s inequality or simply the interpolation inequality.

Remark that it holds for 0 ≤ θ ≤ 1, j ≤ k < l and 1 ≤ p ≤ r ≤ q < ∞.
For a bounded subset of IRn (Ω), we have in addition:

θ
‖u‖Wk,r(Ω) ≤ C‖u‖ Wl,p(Ω)
‖u‖ 1−θ
Wj,q(Ω)
.

We can apply this interpolation inequality to a particular (interesting) case:

1 1
‖∂1u‖Lp(IRn) ≤ C‖∂2u‖2p( n) ‖u‖2p( n).
L IR L IR
1 1
Here, we can use the scaled Young’s inequality to C‖∂2u‖2p ( n) ‖u‖2 p( n) .
L IR L IR

ap q q
− b
For any ε > 0, ab ≤ ε + [ε] p (a, b ≥ 0).
p q

Hence, there exists a constant >0 such that:

‖∂u‖Lp(IRn) ≤ δ‖∂2u‖Lp(IRn)+Cδ‖u‖Lp(IRn).

This means that the Lp norm of ∂u is controlled by the Lp norms of ∂2u and u.

98