Anda di halaman 1dari 166

Table of contents

1 Preliminaries 1
1.1 Continuum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Questions of notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Coordinate and component transformations . . . . . . . . . . . . . . . . . . 3
1.4 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Tensor analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Time derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Conservation laws and stress 12


2.1 Mass density and equation of continuity . . . . . . . . . . . . . . . . . . . . 13
2.2 Forces and conservation of linear momentum . . . . . . . . . . . . . . . . . 15
2.3 Stress tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Center of mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Conservation of angular momentum . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Conservation of energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.8 Eigenvalue problem for symmetric tensors . . . . . . . . . . . . . . . . . . . 23
2.9 Normal and shear stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Deformation 29
3.1 Finite deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Infinitesimal deformation or small strain . . . . . . . . . . . . . . . . . . . . 34
3.3 Geometric interpretation of small strain . . . . . . . . . . . . . . . . . . . . 36
3.4 Rate of deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5 Principal strains and special deformations . . . . . . . . . . . . . . . . . . . 40
3.6 Compatibility conditions for small strain . . . . . . . . . . . . . . . . . . . . 42

i
Table of contents ii

4 Constitutive relations 44
4.1 Young’s experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Frame-indifference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Isotropy and Reiner-Rivlin fluids . . . . . . . . . . . . . . . . . . . . . . . . 50
4.4 Linear isotropic media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5 Viscoelastic media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.6 Kramers-Kronig’s relations . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5 Ideal fluids 63
5.1 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2 Simplifying hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3 Bernoulli’s law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Solar wind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Hydrostatic equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.6 Sound waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.7 Linear waves in magnetized plasmas . . . . . . . . . . . . . . . . . . . . . . 81
5.8 Large amplitude ion-acoustic solitons . . . . . . . . . . . . . . . . . . . . . 88
5.9 Large amplitude circularly polarized waves in plasmas . . . . . . . . . . . . 93
5.10 Newtonian cosmology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.11 Complex potential theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

6 Viscous fluids 120


6.1 Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2 Dimensional analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.3 Turbulence : Reynolds equations . . . . . . . . . . . . . . . . . . . . . . . . 128
6.4 Turbulence : Kolmogorov’s law . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.5 Matched asymptotic expansions . . . . . . . . . . . . . . . . . . . . . . . . 136
6.6 Boundary layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

7 Linear elasticity 145


7.1 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.2 Elastostatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.3 Elastic waves in general linear media . . . . . . . . . . . . . . . . . . . . . . 149
7.4 Elastic waves in linear isotropic media . . . . . . . . . . . . . . . . . . . . . 151
Table of contents iii

8 Epilogue: Kinetic theory 155


8.1 Kinetic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.2 Macroscopic averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.3 Macroscopic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
List of figures

2.1 Sequence of domains enclosing P . . . . . . . . . . . . . . . . . . . . . . . . . 13


2.2 Surface elements around P . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Comparison of initial and final states. . . . . . . . . . . . . . . . . . . . . . . 30


3.2 Decomposition of neighbouring displacements. . . . . . . . . . . . . . . . . . 35
3.3 Deformation of angles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Velocities of neighbouring particles. . . . . . . . . . . . . . . . . . . . . . . . 38

4.1 Young’s experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45


4.2 Contour of integration in the complex ω-plane. . . . . . . . . . . . . . . . . . 60

5.1 Pressure-density relations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66


5.2 Streamlines and the velocity field. . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3 Steady outflow of a tank. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4 Hydrostatic equilibrium of a star. . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5 Propagation of radio signals through the ionosphere. . . . . . . . . . . . . . . 87
5.6 Profile of a solitary wave joining constant states at η = ±∞ and localized in η. 93
5.7 Two observers O and O′ and a distant galaxy P . . . . . . . . . . . . . . . . . 98
5.8 Flow of a source or sink in the origin. . . . . . . . . . . . . . . . . . . . . . . 108
5.9 Flow around a point vortex in the origin. . . . . . . . . . . . . . . . . . . . . 109
5.10 Flow around a circle or cylinder. . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.11 Flow along an edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.12 Flow in a corner. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6.1 Reynolds experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129


6.2 Velocity diagram in turbulent flow. . . . . . . . . . . . . . . . . . . . . . . . 129
6.3 Comparison between turbulent and Hagen-Poiseuille velocity profiles. . . . . 133
6.4 Energy spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.5 Power law in the inertial range. . . . . . . . . . . . . . . . . . . . . . . . . . 136

iv
Table of contents v

6.6 Graphs of the exact solution and the inner and outer approximations . . . . 140
6.7 Velocity profile. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

7.1 P - and S-earthquake waves. . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

8.1 Description in phase space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156


Chapter 1

Preliminaries

In this chapter we

• give a short introduction to continuum mechanics

• recall concepts from coordinate transformations, vector and tensor algebra


and analysis

• finally address time derivatives of material quantities, both in punctual as


in integral form

1
Table of contents 2

1.1 Continuum mechanics


The basic idea behind continuum mechanics is to describe the kinematical and dynamical
properties of certain classes of material as if the material under consideration were filling
part of physical space in a continuous way and had characteristics which depend continuously
upon space and time. Of course, this is a model, an idealization if one thinks of the molecular
or (sub)atomic structure of all matter. On the microscopic scale much more discontinuity
than continuity is to be found. Nevertheless, experience has learned that many macroscopic
properties can be described with an amazing degree of precision by indeed pretending that
certain materials are continua, in the mathematical sense of the word. To quote just one
example at this stage, for the flow of water through a pipe the knowledge of the molecular
structure of water will learn strictly nothing as far as the flow speed is concerned, if only
because of the vast number of molecules to be accounted for.
Treating matter as a continuum can be done for matter in the solid state, where it comple-
ments the notion of rigid bodies and leads to the theory of elasticity, and also for matter
in the liquid or gas states, where one gets hydro- and aerodynamics which are sometimes
grouped together into fluid dynamics. So it seems that matter under certain circumstances
permits us to use the mathematical concept of a continuum, and the whole question then
revolves around knowing the limits of accuracy and applicability of this continuum concept
in practical cases. This is a rather difficult problem, because it stands almost totally di-
vorced from the mathematical modelling which can proceed once basic concepts and laws
are defined and postulated, as will be done in the next chapter.

1.2 Questions of notation


As is the case with most branches of physics, continuum mechanics deals with certain phys-
ical quantities which exist and can be talked of independently of a particular coordinate
system. In mathematical terms this means that such physical quantities are represented by
tensors and that the physical laws are expressible as tensor equations. Of course, these same
physical quantities will be measured and attributed numbers or values with respect to an
appropriate coordinate system. As soon as a coordinate system is chosen, a tensor, although
a mathematical entity on its own, is completely specified through its components. These
components change when one passes from one coordinate system to another according to
specific rules.
This dual way of thinking of a tensor, i.e as a mathematical entity on its own and through
its coordinate representation, will reveal itself also in the notation. We write T for a certain
tensor, or else we will use its components Tij... when referred to a coordinate system. The
indices i, j, . . . run from 1 to 3, in physical space that is, for one could as easily define
tensors in n-dimensional space and the range of the indices would then be from 1 to n.
Some of the mathematical properties recalled in the following paragraphs are indeed valid
for general tensors in n-dimensional space. The coordinate systems used henceforth will
always be orthogonal Cartesian coordinate systems. This restriction avoids the complication
of having to distinguish between covariant and contravariant tensors, but still allows enough
Table of contents 3

interesting properties and applications for a first course in continuum mechanics. The word
“tensor” will then mean “Cartesian tensor”.
A scalar is a tensor of order zero and has one, invariant component in each coordinate
system. A vector is a tensor of order one and will be written in boldface (e.g. a, v ...). Its
three components will be denoted by the same letter subscripted once (e.g. ai , vi ...).
Finally, a tensor of order two will be called a tensor, in short, and is denoted e.g. by T. It
has nine components which are written Tij for i, j = 1, 2, 3. The unit tensor will be written
as 1, with components δij , the familiar Kronecker deltas.
When a certain expression is written in index notation, use will be made of the Einstein
summation convention in order to save the writing of summation signs with the appropriate
summation ranges. The summation convention implies that, whenever an index appears
twice in a product, the index or subscript takes the values 1,2,3 consecutively and the three
terms thus formed are summed. In the few cases where the summation convention cannot
be followed, this will be mentioned explicitly.

1.3 Coordinate and component transformations


Physical space is seen as a three dimensional affine Euclidean space, modelled on IR3 . A
reference frame in space is given by the choice of an origin and of a positive (i.e. right-handed)
orthonormal basis e(i) (i = 1, 2, 3). The position of a point P is given through its position
vector x = OP or, equivalently, through its Cartesian coordinates, which will be denoted
xi for i = 1, 2, 3 (rather than through x, y, z). The triplet (e(1) , e(2) , e(3) ) determines the
orientation of the physical space. Once a base of a coordinate system is known, the relation
between the position vector and the coordinates of a point P is simply

x = xi e(i) .

By a coordinate system we will always mean a Cartesian coordinate system and we will
denote it by Ox1 x2 x3 .
The next step is the transformation of one coordinate system Ox1 x2 x3 , with unit vectors e(i) ,
to another one Ox′1 x′2 x′3 , with unit vectors e′ (j) . Both coordinate systems have for simplicity
the same origin. The unit vectors then transform according to
(i)
e′ = Aij e(j) ,

where Aij are the direction cosines of the new axes referred to the old ones,
 
′ (i) (j) ′ (i) (j)
Aij = e · e = cos e , e .

Since we are dealing with orthogonal coordinate systems, this yields two sets of orthonor-
mality relations:
Aij Aik = δjk , Aij Akj = δik ,
Table of contents 4

i.e. the matrix [Aij ] is an orthogonal matrix. In addition, the transformation is orientation
preserving, i.e.: det[Aij ] = 1. Thus the required coordinate transformations between the old
coordinates xi and the new ones x′j of a point P is

x′i = Aij xj .

The inverse transformation is


xi = Aji x′j ,
as can easily be checked,
Aji x′j = Aji Ajk xk = δik xk = xi .
Defining an arbitrary vector v in space as the difference between the position vectors of its
two endpoints P and Q, allows us to deduce the transformation laws of the components of
v, which are denoted by vi in Ox1 x2 x3 and by vj′ in Ox′1 x′2 x′3 ,

vi′ = Aij vj

We could also turn the argument around and postulate that a vector is defined in such a
way that its components obey this transformation law. It is easy to check and left to the
reader to verify that vectors thus defined are the elements of a vector space.
Next come the transformation laws for the components of a tensor of order two or higher.
For a tensor of order two these are

Tij′ = Aip Ajq Tpq

and for more general tensors



Tijk... = Aip Ajq Akr · · · Tpqr... .

We will say that a tensor is isotropic if it is invariant under any orthogonal coordinate
transformation, i.e. Tijk... = Aip Ajq Akr · · · Tpqr... , for every orthogonal matrix A.
Each component of a tensor of order n has n indices, and there are just n factors Aip in the
transformation equation. If one thinks of a tensor of rank two T as a linear transformation
which maps a vector u into a vector v,

v = T · u,

or in index notation
vi = Tij uj ,
then the transformation laws for a tensor of order two can be deduced from those for a
vector.
Tensors of the same order can again be proved elements of a vector space. More precisely,
if S and T are tensors of order two, say, and λ is an arbitrary real number, then λ S and
S + T are again tensors of order two, and all the traditional properties of a vector space are
Table of contents 5

satisfied. Let us denote the space of second order tensors by V. For each element S of V
there exists an ‘inverse’ in V, denoted −S, and the following distributive properties hold:
λ(S + T) = λ S + λ T, (λ + µ)T = λ T + µ T,
where λ and µ are scalars.
Finally, let’s go back to the transformation rule for the components of a general tensor

Tijk... = Aip Ajq Akr · · · Tpqr... . These hold not only for transformations from one right-handed
basis to another right-handed basis, but for any transformation of the basis. Quantities for
which the transformation rule only holds for an orthogonal transformation with det[Aij ] = 1
are called pseudo-tensors. Stated otherwise, if the transformation equations for vector or
tensor components only hold when det[Aij ] = +1, then we are dealing with pseudo-tensors
rather than with tensors in the proper sense. An often-used pseudo-tensor of order three is
the alternating pseudo-tensor of Levi-Cività εijk with components

 = +1 if i, j, k are an even permutation of 1,2,3,
εijk = −1 if i, j, k are an odd permutation of 1,2,3,

= 0 otherwise, when two or three indices are equal.
Some useful identities involving the components of the pseudo-tensor of Levi-Cività and of
the unit tensor of Kronecker are
εijk εiℓm = δjℓ δkm − δjm δkℓ ,
εijk εijℓ = 2δkℓ ,
εijk εijk = 6,
δii = 3.
For a given tensor T, with components Tij , one can always construct the transposed tensor,
e such that
denoted by Tt or T,
(T t )ij = Tji .
A coordinate independent definition would be through
u · Tt = T · u
for all u. A tensor is called symmetric iff
Tij = Tji or Tt = T.
Any tensor T of order two can be split in a unique way into a symmetric and an antisymmetric
part,
1 1
T = (T + Tt ) + (T − Tt ),
2 2
or in index notation
1 1
Tij = (Tij + Tji ) + (Tij − Tji ).
2 2
It is possible to generalize the concept of the transposed of a tensor to tensors of higher order,
where such a transposed tensor is called an isomer, but extreme care has to be exercised as
to which pair of indices is involved in such a permutation. The pseudo-tensor of Levi-Cività
is totally antisymmetric, because any permutation of two indices changes the sign of the
components.
Table of contents 6

1.4 Tensor products


Starting with scalars λ and µ we have only one kind of product, the usual one, λµ, in the
sense of number theory.
For vectors u and v in Euclidean space, one is supposed to be familiar with the scalar or
inner product
u · v = ui v i ,
which yields a scalar, and (if the dimension of the space is three) with the vector or cross
product u × v
(u × v)i = εijk uj vk ,
which is really a pseudo-vector.
Next, we can define a product on vectors which yield a (second-order) tensor: the so-called
tensor or outer product
(uv)ij = ui vj .

One can indeed verify that the object defined by the right-hand side is a second-order tensor
by checking the transformation property.
The ideas of inner and outer products can be extended to tensors of order two by putting

(S · T)ij = Sik Tkj

for the inner product and


(S T)ijkℓ = Sij Tkℓ
for the outer product. In more general terms even, the inner product of a tensor of order m
with a tensor of order n yields a tensor of order m + n − 2,

(S · T )i···jk···ℓ = Si···jp Tpk···ℓ ,

whereas the outer (or tensor) product gives a tensor of order m + n,

(S T )i···jk···ℓ = Si···j Tk···ℓ .

If m + n − 2 is 2 or more, the inner product or contraction can be repeated to yield the


double inner product or the double contraction of two tensors,

(S : T )i···jk···ℓ = Si···jpq Tqpk···ℓ .

In particular, the double inner product of two tensors of order two is a scalar,

S : T = Sij Tji

One sometimes finds the double inner product defined in another way,

S ··T = Sij Tij ,


Table of contents 7

which usually gives a different result. It is to be noted that both definitions coincide if at
least one of the two tensors is symmetric.
Remark. In mathematical textbooks one will usually adopt the notation ⊗ for tensor
product (e.g. u ⊗ v, S ⊗ T, . . .).
One could also try to extend the idea of the cross product of two vectors, but it seems only
useful to do so for the cross product of a vector with a tensor of order two, in that order, in
view of the introduction of the curl of a tensor in the next paragraph,

(v × T)ij = εipq vp Tqj .

This cross product is a pseudo-tensor, as was already the case with the cross product of two
vectors.
To round off this paragraph, some forms of

(T · v)i = Tij vj and (v · T)i = vj Tji

are given in the special case when the tensor itself is the tensor product of two other vectors,

(uv) · w = u(v · w) = uv · w,
u · (vw) = (u · v)w = u · vw.

This shows that a notation such as uv · w is unambiguous, and there is no need to write any
brackets.
If one introduces the trace of a tensor of order two,

tr T = Tii ,

which is invariant under orthogonal coordinate transformations (i.e. tr T = Tii′ ), one finds
the following relations as special cases:

tr (uv) = ui vi = u · v,
tr (S · T) = (S · T)ii = Sij Tji = S : T.

1.5 Tensor analysis


In this paragraph the notions of gradient, divergence and curl, which are known from vector
analysis, are recalled and extended to general tensors.
The usual gradient of a scalar function φ of x (and possibly t) is a vector with components

∂φ
(grad φ)i = (∇ φ)i = = ∂i φ.
∂xi
Table of contents 8

The partial derivative operator with respect to the coordinate xi will henceforth be denoted
∂i . The symbolic vector nabla, ∇, has components ∂i and could formally be regarded
as a vector obeying the transformation laws, although it retains its derivative operator
characteristics. Hence the (generalized) gradient of a tensor T of order n is a tensor of order
n + 1, constructed as the tensor product of the ‘nabla’ vector with the said tensor,

(grad T )ij... = (∇ T )ij... = ∂i Tj... .

Special cases of this equation are the gradient of a vector,

∂vj
(grad v)ij = (∇ v)ij = = ∂ i vj
∂xi

which is a tensor of order two, and the gradient of a tensor T of order two,

(grad T)ijk = (∇ T)ijk = ∂i Tjk ,

which is a tensor of order three.


On the other hand, the usual divergence of a vector is a scalar,

div v = ∇ · v = ∂i vi .

Hence we define the divergence of a general tensor T of order n as a tensor of order n − 1,


constructed as the inner product of the nabla vector with T :

(div T )ij... = (∇ · T )ij... = ∂p Tpij...

This includes as a special case the divergence of a tensor (of order two),

(div T)i = (∇ · T)i = ∂j Tji

Finally, in vector analysis, the curl of a vector is a pseudo-vector,

(curl v)i = (∇ × v)i = εijk ∂j vk ,

and, generalizing this concept, the curl of a tensor T (of order two) is a pseudo-tensor of the
same order,
(curl T)ij = (∇ × T)ij = εipq ∂p Tqj .
The gradient, divergence and curl operations can be repeated several times, if the resulting
expressions permit, as in the following example of the double gradient of a scalar function,
which is a tensor of order two,

∂ 2φ
(∇∇ φ)ij = = ∂i ∂j φ.
∂xi ∂xj
Table of contents 9

Once these operations are established, it becomes straightforward to write a Taylor expansion
of a tensor around a point with position vector xo , or with coordinates xoi , either in index
notation,
Tij... (x) = Tij... (xo ) + (xℓ − xoℓ )∂ℓ Tij... (xok ) + . . .
or in vector notation,
1
T (x) = T (xo ) + (x − xo ) · ∇ T (xo ) + (x − xo )(x − xo ) : ∇∇ T (xo ) + . . . .
2
Some useful identities involving the nabla operator are:

∇(u · v) = u · ∇v + v · ∇u + u × (∇ × v) + v × (∇ × u),
∇ · (uv) = (∇ · u)v + u · ∇ v,
∇ · (T · v) = v ∇ : T + Tt : ∇ v,
∇ · (λ v) = λ ∇ · v + v · ∇ λ.

1.6 Time derivatives


In continuum mechanics we are interested, among others, in the time evolution of certain
objects. The partial time derivative of a tensor depending explicitly on t will be denoted by
∂T
∂t T = ,
∂t
with
∂Tij...
(∂t T)ij... =
∂t
If the continuum moves in space, such that position vector of a material point becomes a
differentiable function of time, the change of a tensor along a ‘flow line’ is given by the total
or material time derivative (i.e. the change of the tensor when following the material in its
motion, hence the name of material derivative):
dT d
Ṫ = = T(x(t), t)
dt dt
∂T d x
= + · ∇ T = ∂t T + v · ∇ T.
∂t dt
The operator
d
= ∂ t + v · ∇ = ∂ t + vi ∂ i
dt

is called the material time derivative. The material derivative, as calculated above, is for
functions which are defined in each point of the material. Note, in particular, that the
material derivative of the position vector x at a point P is the velocity of the material in
that point (or the velocity of P for short).
Table of contents 10

Further on, the problem will also arise of finding the total time derivative of quantities which
are expressed as a time-dependent volume integral, when the volume under consideration
coincides at all times with a well-defined part of the material in motion. Something analogous
arises when in a one-dimensional integral both the integrand and the integration limits
depend on a parameter,
Z ξ1 (α)
I(α) = f (ξ, α) dξ,
ξ0 (α)

and the derivative of the integral is sought with respect to this parameter,
Z ξ1 (α)
dI ∂f dξ1 dξ0
= dξ + f (ξ, α) − f (ξ, α) .
dα ξ0 (α) ∂α ξ=ξ1 (α) dα ξ=ξ0 (α) dα

The first term on the right-hand side of this equation comes from the dependence of the
integrand on the parameter, the two remaining terms occur because the limits vary with the
same parameter.
Let ψ(x, t) be a continuous function of coordinates and time and φ(t) the integral of ψ over
a certain material volume V , which evolves in time in such a way that the same material
points remain contained in V , Z
φ(t) = ψ(x, t) dV.
V
To compute the material derivative
Z
d
φ̇(t) = ψ(x, t) dV
dt V

we choose some time-independent reference configuration, in such a way that a particle which
was at position a at a given (original and fixed) instant of time is at x at time t. In this way
we can express
x = x(a, t).
For this transformation the volume-elements dV = d3 x and dV0 = d3 a are related through

dV = J dV0 ,

where dV0 refers to the initial state, and J is the Jacobian of the transformation,
 
∂xi
J = det .
∂aj
Hence
Z Z
d d
ψ(x, t) dV = ψ[x(a, t), t]J dV0
dt V dt V0
Z
d
= {ψ[x(a, t), t]J} dV0
V0 dt
Z  
= ψ̇J + ψ J˙ dV0 ,
V0
Table of contents 11

using the definition of the material time derivative of ψ. To work out what J˙ is, we also
need  
d ∂xi ∂ ẋi ∂vi ∂vi ∂xk
= = = .
dt ∂aj ∂aj ∂aj ∂xk ∂aj
The time derivative of J thus gives rise to
     
d ∂x1 d ∂x1 d ∂x1 ∂xk ∂xk ∂xk

dt ∂a1 dt ∂a2 dt ∂a3 ∂a1 ∂a2 ∂a3


∂x2 ∂x2 ∂x2 ∂v1 ∂x2 ∂x2 ∂x2
J˙ = + ... = + ...
∂a1 ∂a2 ∂a3 ∂xk ∂a1 ∂a2 ∂a3

∂x3
∂x3 ∂x3 ∂x3 ∂x3 ∂x3

∂a1 ∂a2 ∂a3 ∂a1 ∂a2 ∂a3
∂v1 ∂v2 ∂v3
= J+ J+ J = J∂i vi = J ∇ · v,
∂x1 ∂x2 ∂x3
where the dots refer to similar determinants with the second or the third row differentiated
instead of the first one. The transition from the second to the third expression for J˙ in the
previous derivation comes about because only the value k = 1 gives a nontrivial result for
the first determinant, and similarly for the other two determinants. Once J˙ is known, we
obtain Z Z   Z  
d
ψ(x, t) dV = ψ̇ + ψ ∇ · v JdV0 = ψ̇ + ψ ∇ · v dV,
dt V V0 V
or
Z
φ̇(t) = (ψ̇ + ψ ∇ · v)dV
V

This expression will be used several times later on. It is instructive, however, to expand the
definition of the material time derivative ψ̇ and get
Z
φ̇ = (∂t ψ + v · ∇ ψ + ψ ∇ · v)dV
ZV Z
= ∂t ψ dV + ∇ · (ψ v)dV.
V V

The second volume integral can be changed into a surface integral by using the Gauss-
Ostrogradski theorem Z Z
φ̇ = ∂t ψ dV + n · (ψ v)dS,
V ∂V

where ∂V is the boundary of V and n is an outer unit normal to ∂V . The expression for φ̇
thus consists of two parts: one related to the time change of ψ inside V , the other one giving
the total flux of ψ v through the boundary ∂V .
Chapter 2

Conservation laws and stress

Conservation laws lead to basic equations:

• Conservation of mass gives the equation of continuity

∂t ρ + ∇ · (ρv) = 0 = ρ̇ + ρ∇ · v

• Conservation of momentum gives the equation of motion, for which we need


the concept of stress to describe internal forces in a continuous medium

ρ v̇ = f + ∇ · T

• Conservation of angular momentum leads to the symmetry of the stress


tensor

For the (symmetric) stress tensor the eigenvalue problem yields the notions
of principal normal and shear stresses

12
Table of contents 13

2.1 Mass density and equation of continuity


Let us start by choosing a reference frame (coordinate system) Ox1 x2 x3 for the physical
space. Next, we can think of a lump of matter, or a portion of a liquid or gas, which in
the continuum concept fills a domain D0 in space having a volume V0 and with boundary
∂D0 . In an interior point P the description of the continuum and its behaviour requires the
definition and knowledge of certain physical properties. How do we proceed?
First, we introduce the notion of mass density. This is done in a way which might also serve
to get other required properties in P . The continuum enclosed in D0 has a total mass M0 .
The notions of mass and volume should be familiar from theoretical mechanics and give no
difficulties at this level of the description. The average mass density in each interior point
of D0 is simply
M0
ρaverage = .
V0
The mass density in P would then be somehow the limit of this expression, if the volume
around P were tending to zero. To see what this implies, consider a sequence of domains Di
enclosed in each other, as shown in Figure 2.1. The matter enclosed in a domain Di with
volume Vi has a mass Mi . The domains Di can be fictitious and need not at all have physical
boundaries, they are just a theoretical construction used to define the concept of mass density.
This is a stratagem we will often resort to in the development of continuum mechanics:
mentally isolating part of the continuum under consideration. Each of the domains Di
should contain P , of course. With the sequence of domains we get a corresponding sequence
of average mass densities. The mass density ρ in P is then defined as the limit of this
sequence,
Mi i→∞
ρi = −→ ρ(P ) (Mi → 0, Vi → 0).
Vi
This limit should exist and be unique, regardless of the sequence of domains Di used to
construct the limit. Nevertheless, if we go back for a moment to the physical side of the

D0 Di

Figure 2.1. Sequence of domains enclosing P .


Table of contents 14

argument, it is clear that even an elementary volume around P should still contain a large
enough number of molecules in order not to violate the whole idea of a continuum. Indeed,
if volumes of molecular or intra-molecular scale are considered, chopping off part of such
a volume will result in a discontinuous change in the mass density whenever a molecule
is discarded. Here comes a natural limitation of the kind of phenomena one can hope to
describe adequately with a continuum theory: phenomena that are on a scale which is large
compared to the atomic interparticle distance and a fortiori to the size of an atom or a
molecule.
This process of defining a mass density can be repeated for each interior point P . The
result is a mass density ρ(x, t) defined everywhere in D0 and by virtue of the ever present
continuum hypothesis, this mass density is a continuous function of x in D0 at all times.
In a similar vein other densities can be defined, such as an energy density or a momentum
density. Instead of the preceding, more intuitive way of defining the mass density, one can
also use measure theory to define a mass density in a unique way, from the moment one
deals with a continuum such that one can associate with every volume V in D0 a number
giving the mass in V . What are needed are properties of a measure µ defined in D0 such
that in particular

µ(U ∪ U ′ ) = µ(U ) + µ(U ′ ) (U ∩ U ′ = ∅; U, U ′ ⊂ D0 ),


µ(U ) ≥ 0,
µ(∅) = 0.

Clearly, both volume and mass are two such measures defined in D0 , mass only in the con-
tinuum limit which precludes masses concentrated in one point. Then the Radon-Nikodym
theorem states the existence of a unique (mass) density ρ such that mass and volume are
related according to Z Z
M (D) = dM = ρ dV.
D D
This enables us at once to get the total mass contained in a domain D ⊂ D0 with volume
V . Henceforth, such integrals will be written as volume integrals over V rather than over D:
Z
M= ρ dV.
V

If the volume V of a continuum is followed in its motion, it is assumed that V will always
consist of the same material particles, since it is precisely their motion which is constituting
the motion of V . Hence the mass in V is invariant when V is followed in its motion and the
material time derivative of M vanishes,

Ṁ = 0.

This expresses the first conservation law, the conservation of mass, in extensive form for a
specific volume V ,
Z Z
Ṁ = (∂t ρ + ∇ · (ρv))dV = (ρ̇ + ρ∇ · v)dV = 0.
V V
Table of contents 15

If the conservation of mass is to hold throughout the continuum for arbitrary volumes, even
imagined ones, then the integrands in the previous integrals have to vanish everywhere,
which yields the point or differential form of mass conservation,

∂t ρ + ∇ · (ρv) = 0 = ρ̇ + ρ∇ · v

This equation is known as the equation of continuity.

2.2 Forces and conservation of linear momentum


From point mechanics one knows Newton’s (second) law in the form: the time derivative
of the total linear momentum of all particles equals the sum of all forces acting upon these
particles. Going then to the notion of rigid bodies with a finite extent, one had Euler’s laws
as the appropriate generalizations of Newton’s law of motion.
Continuum mechanics can be viewed as a limit of point mechanics with an ever increasing
number of particles of ever decreasing mass, hence it would seem natural to postulate a
rephrasing of Euler’s first law for rigid bodies: the material (i.e. total) time derivative of the
total linear momentum of the mass contained in a volume V is equal to the sum of all forces
acting upon V . The translation of this principle into mathematical form requires a suitable
definition of the total linear momentum of the mass in V , which we take as
Z
P= ρv dV,
V

as well as a suitable definition of all forces acting upon V . Instead of Euler’s first law for a
rigid body, we get for a continuum something formally very similar,
Z
d
Ṗ = ρ v dV = F.
dt V
Two things remain to be done, the computation of Ṗ and a proper definition of F. It is
straightforward to find that
Z  Z Z
d
Ṗ = ρ v dV = (ρ̇v + ρ v̇ + ρ v ∇ · v) dV = ρ v̇ dV.
dt V V V

The simplification in the integrand comes from invoking the continuity equation and is a
result which is valid for more general integrands besides v. As we will need this property
several times afterwards, we check that for a tensor T of any order
Z  Z   Z
d
ρT dV = ρ̇T + ρṪ + ρT ∇ · v dV = ρṪ dV,
dt V V V

as soon as the equation of continuity is valid. Hence Newton’s law now stands in the form
Z
ρ v̇ dV = F.
V
Table of contents 16

As far as the forces are concerned, one usually distinguishes two kinds of forces: the body or
volume forces and the surface forces.
Body forces arise as a consequence of action at a distance and are intuitively thought of
as the generalization of external forces in point mechanics. Typical examples are gravity,
electromagnetic forces and the like. Such forces act in each point of the continuum and
we will characterize them through a force density per unit volume f (x, t), or else through a
normalized force density b(x, t) with the dimensions of an acceleration. The relation between
both densities is
f (x, t) = ρ(x, t) b(x, t).
The sum of the body forces acting upon a volume V of the continuum is thus
Z Z
Fbody = f dV = ρ b dV.
V V

In the case of gravity one finds that


Z  Z
Fgrav = M g = ρ dV g= ρ g dV
V V

and hence

f (x, t) = ρ(x, t) g,
b(x, t) = g.

Surface forces, on the other hand, act on the surface of a continuum, and could in some
rare cases belong to the external forces, such as when the atmospheric pressure has to be
considered at the surface of a body. Usually, however, surface forces have to be introduced
to translate the action of one part of the continuum upon another, if part of the continuum
is considered separately but yet in the same state of motion and deformation. In this sense
the surface forces could be thought of as generalizing the idea of internal forces in a set
of particles, and they represent something as the cohesion of the material. These forces
cancel each other in each interior point as the result of the law of action and reaction, and
only survive with a non-null resultant at the boundary of the continuum, whether the real
physical boundary of a lump of matter or a fictitious boundary which is thought of isolating
part of the continuum from the rest.
Think of a point P , interior in the volume V0 , and imagine an internal volume V in such a
way that P lies on the boundary ∂V of V , which we assume to be continuous. Define n as
a unit vector in the external direction, normal to ∂V , and passing through P . As part of V0
this volume V is in a certain state of motion and deformation. If we now want to describe
only V and forget about the remainder of V0 without changing the motion or deformation
of V , the action which the mass in V0 \V exerts on V has to be taken into account explicitly.
Consider now an elementary surface element ∆S (⊂ ∂V ) at P , as shown in Figure 2.2, and
(n)
call ∆F(n) the total force exerted by V0 \V upon V , acting through ∆S, and ∆NO the
Table of contents 17

n
V0 V ∆S
P

Figure 2.2. Surface elements around P .

resulting moment (with respect to O). Let ∆S shrink to zero and we get, according to the
stress postulate of Euler-Cauchy, a unique and well defined limit:

∆F(n) (n) 
lim∆S→0 =t  
∆S (∆S → 0 k n ⊥ ∆S, P ∈ ∆S ⊂ ∂V ).
(n)
∆NO 

lim∆S→0 =0 
∆S
The vector t(n) is called stress vector, associated with the choice of the normal n in P , and it
has the dimension of force per area. Note that this postulate tells us that there is no stress
moment density.
Several remarks are in order here. First of all, the Euler-Cauchy postulate is a postulate (i.e.
can not be proven), even though introduced on fairly obvious physical grounds. Furthermore,
the stress vector t(n) is uniquely defined under this hypothesis regardless of the choice of V
or its boundary, as long as n is the external unit vector in P normal to ∂V . The argument
to have no net stress moment is based intuitively on the shrinking of the dimensions of the
surface around P to zero. We will follow this argument for the sake of simplicity, although
it is by no means required to construct a consistent theory. Suffice it here to mention that
for so-called Cosserat continua the stress moment density is taken into account in a fully
coherent development of the continuum theory. The principle of action and reaction in an
interior point P means that
t(−n) = −t(n) .
To know the stress field fully at this stage is equivalent to knowing a triple infinity of stress
vectors at each interior point P . Obviously, something simpler will be necessary for a more
workable description.
The sum of the surface forces acting upon a continuum with volume V and boundary ∂V is
Z
Fsurface = t(n) dS.
∂V
Table of contents 18

If we return to the law of motion, which was left in the form


Z
ρ v̇ dV = F,
V

we find Z Z Z
ρ v̇ dV = f dV + t(n) dS.
V V ∂V
For the present, no point or differential form of this conservation law is possible. This will
have to wait until after the introduction of the stress tensor.

2.3 Stress tensor


In an interior point P the body force density f or ρb is a given quantity. On the other hand,
with each choice of the normal n in P there corresponds a stress vector t(n) . It is generally
admitted that there exists a linear relation between both, expressed in terms of a tensor T
of order two, such that

t(n) = n · T

This tensor T is called the stress tensor and the knowledge of the stress tensor determines
the stress field in P completely. We are now in a position to further transcribe the equations
expressing the conservation of linear momentum.
The surface integral appearing in the momentum equation can be transformed into a volume
integral with the help of the Gauss-Ostrogradski theorem,
Z Z Z
(n)
t dS = n · T dS = ∇ · T dV
∂V ∂V V

or componentwise, Z Z Z
(n)
tj dS = ni Tij dS = ∂i Tij dV.
∂V ∂V V
The momentum equation itself thus becomes
Z Z Z
ρ v̇ dV = f dV + ∇ · T dV,
V V V

and it is now possible to go to the differential or pointwise form of the equation expressing
the conservation of momentum:

ρ v̇ = f + ∇ · T

or equivalently in index notation,

ρv̇j = fj + ∂i Tij .

These are known under the name of equations of motion.


Table of contents 19

2.4 Center of mass


We define, with respect to the given reference frame and again in analogy to what is known
from point mechanics and rigid body mechanics, the center of mass of a continuum contained
in V as R Z
V
ρ x dV 1
xC = R = ρ x dV ,
V
ρ dV M V
with M the total mass of V . When the material time derivative of this expression is taken,
we use the fact that M remains constant and get
Z Z
1 d 1
ẋC = ρ x dV = ρ v dV,
M dt V M V
with the help of the equation of continuity. Hence
P = M ẋC ,
just as for a rigid body. Taking once again the material derivative, we find with the help of
Z Z Z
ρ v̇ dV = f dV + t(n) dS
V V ∂V

that
Z Z
1 d 1
ẍC = ρ v dV = ρ v̇ dV
M dt V M V
Z Z
1 1 1
= f dV + t(n) dS = F.
M V M ∂V M
This states that the acceleration of the center of mass times the total mass is equal to the
sum of the forces acting upon V . The details of the distribution of matter in V seem to
be forgotten, and the actual shape of the volume only plays a role when computing the
sum of the forces. This allows a coarser description of deformable media in terms of the
total mass, center of mass and total forces. Precisely this property has ensured the success
of the classical point mechanics and the mechanics of rigid bodies, where the deformation is
negligible compared to other characteristics, or when all the mass of a body can be thought
of as being concentrated in one point, its center of mass.

2.5 Conservation of angular momentum


In point or rigid body mechanics the conservation of angular momentum is formulated as
follows: the time derivative of the total angular momentum equals the resulting momentum
of all forces acting upon the system, all momenta being computed with respect to a given
fixed point. A similar principle will be invoked in continuum mechanics. The total angular
momentum (with respect to chosen the origin) of a continuum in a volume V is
Z
LO = x × ρ v dV.
V
Table of contents 20

The total momentum results only from the momenta of the forces acting upon V and is
Z Z
NO = x × f dV + x × t(n) dS.
V ∂V
The conservation of angular momentum for continua is hence
L̇O = NO ,
formally reminiscent of classical mechanics, or in more explicit form,
Z Z Z
ρ x × v̇ dV = x × f dV + x × t(n) dS,
V V ∂V
since Z  Z Z
d d
x × ρ v dV = ρ (x × v) dV = ρ x × v̇ dV.
dt V V dt V
One has to change the surface integral into a volume integral according to the theorem of
Gauss-Ostrogradski,
Z Z
(n)

x × t i dS = (x × (n · T))i dS
∂V ∂V
Z Z
= εijk xj nℓ Tℓk dS = ∂ℓ (εijk xj Tℓk ) dV
∂V V
Z Z
= εijk (δℓj Tℓk + xj ∂ℓ Tℓk ) dV = (ε ··T + x × (∇ · T))i dV.
V V
The computation was done in index notation because in vector notation it is well-nigh
impossible and would be at any rate confusing. There remains in the angular momentum
equation that
Z Z Z
x × ρ v̇ dV = x × f dV + (x × ∇ · T + ε ··T) dV
V V V
or with the help of the equation of motion,
Z
ε ··T dV = 0.
V
If stress momenta would have been included in the description, this result would no longer
be valid. Again one finds the differential or point form of the equation expressing the
conservation of angular momentum,
ε ··T = 0 or εijk Tjk = 0.
There follows that
εipq εijk Tjk = δpj δqk Tjk − δpk δqj Tjk = Tpq − Tqp = 0,
and conversely, so that
ε ··T = 0 ⇐⇒ Tt = T.
What remains of the conservation of angular momentum is that the stress tensor has to be
a symmetric tensor. Such a result would not exist for Cosserat continua where a stress
momentum distribution is present.
Table of contents 21

2.6 Conservation of energy


Many things could be said about energy, but most of them straightaway involve thermody-
namical considerations. As we restrict ourselves in the next chapters to problems which can
be studied from an almost purely mechanical point of view (and there are enough interesting
problems falling in this category) we will give the balance of energy in this paragraph in
very simple terms. We recall that the forces acting upon the part of the continuum under
consideration were the body forces f and the stress forces t(n) . The total mechanical power
then is Z Z
P= f · v dV + t(n) · v dS,
V ∂V
which can be computed to be
Z Z
P = f · v dV + n · T · v dS
V ∂V
Z Z
= f · v dV + ∇ · (T · v) dV
V V
Z Z
 
= f · v dV + (∇ · T) · v + Tt : ∇ v dV
ZV V
Z
= (f + ∇ · T) · v dV + T : ∇ v dV.
V V

We define the rate of deformation tensor D, which will be treated in more detail in the next
chapter, as
1 
D= ∇ v + (∇ v)t ,
2
i.e. it is the symmetric part of ∇ v. In view of the symmetry of T it is easy to prove that
1  1 
T : ∇v = T : ∇ v + (∇ v)t + T : ∇ v − (∇ v)t = T : D.
2 2
This will be used, together with the equation of motion, to rewrite
Z Z
P = (f + ∇ · T) · v dV + T : ∇ v dV
V V

as Z Z Z Z
1 dv 2 d 1 2
P= ρ dV + T : D dV = ρv dV + T : D dV.
V 2 dt V dt V 2 V
With the definition of the kinetic energy as
Z
1 2
T = ρv dV,
V 2

we find that Z
P = Ṫ + T : D dV,
V
Table of contents 22

stating that the total mechanical power equals the change in kinetic energy per unit time plus
something which will be related to the behaviour of the medium as a deformable continuum
(and is sometimes called the stress power).
The kinetic energy is not the only form of energy for a continuum. There is the internal
energy, E, characterized by a density ε of internal energy:
Z
E= ρε dV.
V

The remainder in the energy balance, the non-mechanical power, will be lumped together
and called the heat Q, in such a way that the total mechanical power plus the heat is given
by the time variation of the kinetic plus the internal energy:

P + Q = Ṫ + Ė

This is the first law of thermodynamics. Comparing both expressions for P, one finds that
Z Z 
d
Q+ T : D dV = ρε dV .
V dt V

To get a point form of this balance of energy requires some additional hypotheses about how
the heat changes. We will split the heat contained in a volume V in two parts, one connected
with the internal heat sources or sinks, with density ν (possibly from a radiation field), and
one connected with the heat flux through the boundary, with heat flux vector q, such that
Z Z
Q= ρν dV − n · q dS .
V ∂V

As a result, we finally obtain that

ρε̇ = ρν − ∇ · q + T : D

The time rate of change in the internal energy is given by the heat source density, the
divergence of the heat flux and the stress power T : D. For adiabatic processes, where
there are no heat sources/sinks nor heat fluxes, the stress power changes the internal energy
density in a unique way,
ρε̇ = T : D.
This will be as far as we need to push the reasoning about energy for the purpose of this
and subsequent chapters.
Table of contents 23

2.7 Basic equations


We are now in a position to collect the basic equations obtained sofar: the equation of
continuity, the equation of motion and the symmetry of the stress tensor :

ρ̇ + ρ ∇ · v = ∂t ρ + ∇ · (ρv) = 0
ρv̇ = f + ∇ · T = ρb + ∇ · T
Tt = T

To this one could add the energy balance, but we leave it out of the discussion, as it simply
determines the internal energy once T and v (and hence ∇v and D) are known. If in the
preceding set of equations one counts in scalar components, the equation of continuity is one
equation in four unknowns, the mass density and the three components of the velocity. The
equation of motion corresponds to three scalar equations in nine additional unknowns, the
components of the stress tensor, because the volume force density f has to be considered a
given quantity, as the example of gravity shows. Finally, the symmetry of the stress tensor
adds three further equations.
To sum up, we have at our disposal seven scalar equations containing thirteen unknown scalar
quantities. The theoretical description is thus not closed at all. This should not be surprising,
because we only expressed very general principles which are supposed to hold for all kinds
of continua, be they elastic material or fluids, all of which can be encountered in various
guises. The closure of the theoretical description will only be obtained by the introduction of
constitutive equations, where something more specific is said about given classes of continua.
Constitutive equations will have to wait, however, until chapter 4, because first deformation
has to be treated and this will occupy chapter 3. The remainder of the present chapter will
be devoted to some properties of the stress tensor as a symmetric tensor.

2.8 Eigenvalue problem for symmetric tensors


We recall the interpretation of a tensor of order two as a linear operator mapping a vector
u into another one. One can then ask the question if and when u and T · u can be parallel,
i.e. if and when there exists a scalar λ such that
T · u = λu.
If this is possible, λ is called an eigenvalue and u is the corresponding eigenvector of T. The
terminology has been carried over from what is used in matrix algebra. In index notation
we have that
Tij uj = λ ui ,
which form in three-dimensional space a set of three homogeneous equations in the three
components of u. If an eigenvector u is to be found at all, the determinant of the coefficients
must vanish,
det(Tij − λ δij ) = det(T − λ 1) = 0.
Table of contents 24

This is the characteristic equation of T which yields in general three, possibly coinciding,
values for λ. There are three eigenvalues thus, and with each one corresponds an eigenvector.
We know that a symmetric tensor T has three real eigenvalues. If they are different from
each other, the corresponding eigenvectors form an orthogonal system. If the eigenvalues
are not all different, then it is still possible to construct three eigenvectors which form an
orthogonal system. In this orthogonal system the representation of T is diagonal, with the
three eigenvalues as diagonal elements,
 
λ1 0 0
T =  0 λ2 0  .
0 0 λ3

Expansion of the characteristic equation leads to

λ3 − IT λ2 + IIT λ − IIIT = 0,

where the three (principal) invariants of T are given by

IT = tr T = Tii ,
1 1
IIT = (Tii Tjj − Tij Tji ) = tr (T tr T − T2 ),
2 2
IIIT = det T.

From
T · u = λu
there follows after repeated multiplication with T that

Tm · u = λm u.

Since each eigenvalue satisfies the characteristic equation and in view of the diagonal repre-
sentation, T itself satisfies its own characteristic equation,

T3 − IT T2 + IIT T − IIIT 1 = 0

This is known as the Hamilton-Cayley equation and it can be used to express higher powers
of T as combinations of T2 , T and 1.

2.9 Normal and shear stresses


The symmetry of the stress tensor has several interesting consequences. First comes the
concept of normal stress, associated with a direction n at a point P , which is defined as
the component of the stress vector t(n) along n itself. This component can be written in
equivalent ways as
(n)
σN = n · t(n) = nn : T = T : nn = n · T · n.
Table of contents 25

The normal stress is the same for both the inner and the outer normal to a given plane,
(n) (−n)
σN = σN ,
(n)
and the stress is simply called the traction or pressure according to whether σN is positive
or negative. The shear stress is the component of the stress vector t(n) perpendicular to the
normal, in the plane itself,
(n) (n)
tS = t(n) − σN n.
Note that
m · t(n) = m · T · n = n · t(m)
i.e.: the projection of the stress vector t(n) associated with a direction n upon a direction m
equals the projection of the stress vector t(m) associated with m upon n. For two given or-
thogonal directions m and n with their associated stress vectors one speaks of the reciprocity
of the shear stresses, given by
(n) (m)
m · tS = n · tS = mn : T = nm : T (m ⊥ n).

All these properties simply follow from the symmetry of T.


In general one has for a given direction n at a point both normal and shear stresses. A
relevant question in this context is whether there exist planes in the continuum for which
there is a normal but no shear stress. The corresponding stress vector t(n) should be parallel
to the normal n for such planes, i.e.
t(n) = σ n,
which amounts to
T · n = n · T = σ n.
Thus σ is required to be a solution of

det(T − σ 1) = det(Tij − σδij ) = 0.

We are then dealing, of course, with the eigenvalue problem for the stress tensor, for which
can use the results for symmetric tensors recalled in the previous section. The (real) eigen-
values of T are called the principal stresses, and it is possible to find a coordinate system in
which the stress tensor has a diagonal matrix representation. In such a coordinate system
(based on a system of principal axes for stress) the coordinate planes at least satisfy the
requirement that there be no shear stress along these planes. According to whether some
eigenvalues are equal or not, or vanish or not, different states of stress exist. They are
summarized in Table 2.1.
The coordinate planes in the system of principal axes carry no shear stresses. In the converse
problem one could ask for which directions the shear stresses are maximum, or more general,
extremum. For simplicity, the computations will be done in the system of principal axes.
We start from the magnitude squared of the shear stress vector,
 
σS2 = t(n) − σN n · t(n) − σN n = t(n) · t(n) − σN 2
.
Table of contents 26

state of stress T in principal axes t(n) in principal axes invariants


   
σ1 0 0 n1 σ 1
1. general stress  0 σ2 0   n2 σ 2  IIIT 6= 0 or det T 6= 0
0 0 σ3 n3 σ 3
with possibly
σ1 = σ2 but σ2 6= σ3
   
σ1 0 0 n1 σ 1
2. plane stress  0 σ2 0   n2 σ 2  IIIT = 0
0 0 0 0
t(n) k plane (1,2) IIT 6= 0
   
σ1 0 0 n1 σ 1
3. linear stress  0 0 0  0  IIT = 0 = IIIT
0 00 0
t k e(1)
(n)
IT = tr T 6= 0
     
σ 0 0 n1 σ n1
4. spherical stress  0 σ 0   n2 σ  =⇒ σ  n2  IIT = 31 (IT )2
0 0σ n3 σ n3
(n) 1
σ1 = σ2 = σ3 = σ t = σ n (∀n) IIIT = (I )3
27 T
T = σ1
Table 2.1. Different stress states

With respect to principal axes we have

σN = n2i σi , t(n)
α = nα σ α (i summed, α not),

and hence
σS2 = n2i σi2 − (n2i σi )2 .
The extrema of this function have to be sought depending on the direction of n, with the
constraint that n be a unit vector,

ni ni = n2 = 1.

With the use of a Lagrangian multiplier λ we have to find the extrema of the function

φ = σS2 − λ(ni ni − 1)

as a function of the components of n. The condition that φ be extremum is


∂φ
= 0, (α = 1, 2, 3)
∂nα
Table of contents 27

which gives the set of equations


∂φ
= 2nα σα2 − 2(n2j σj )2nα σα − 2λnα
∂nα

= 2nα σα2 − 2σα (n21 σ1 + n22 σ2 + n23 σ3 ) − λ = 0 (α not summed).

The discussion is first given for the case when all eigenvalues σα of the stress tensor are differ-
ent. Then it can be checked that at least one component of n has to vanish in order to find a
solution, while on the other hand not all three components of n may vanish simultaneously.
A first class of solutions is with two components of n zero, the third component then being
1, which yields

λ = −σ12 if n = e(1) ,
σS2 = 0.

This type of solution yields nothing but the coordinate planes in the system of principal
axes, where the shear stresses vanish. The other class of solutions is found by putting one
component of n, say n3 , zero and solving

σ12 − 2σ1 n21 σ1 + n22 σ2 − λ = 0,

σ22 − 2σ2 n21 σ1 + n22 σ2 − λ = 0,
n21 + n22 = 1,

for n1 and n2 . This gives


1
n21 = n22 = , λ = −σ1 σ2 ,
2
 2
2 σ1 − σ2
σS = .
2
The other cases are analogous, and the principal shear stresses occur along the bisector
planes of the system of principal axes,

2 2 1 σ2 − σ1 
n1 = n2 = and n3 = 0 ⇒ σS = 

2 2 


2 2 1 σ 3 − σ 1
n1 = n3 = and n2 = 0 ⇒ σS = (σ1 < σ2 < σ3 )
2 2 


1 σ3 − σ2  

n22 = n23 = and n1 = 0 ⇒ σS =
2 2
These are the maximal shear stresses, and the discussion of the solutions is herewith ex-
hausted, if we suppose that all σi are different from each other.
Suitable but straightforward corrections can be made to the above reasoning if some of the
eigenvalues coincide. We will leave this to the reader, except to say that in the case of three
equal eigenvalues
σ1 = σ2 = σ3 = σ
Table of contents 28

one has a state of isotropic or spherical stress,

T = σ1,

where every direction corresponds to an eigenvector and there are no shear stresses at all,
only (equal) normal stresses.
Chapter 3

Deformation

• Finite and infinitesimal deformation are obtained by comparing initial and


final states, after subtracting translation and rigid body rotation. In its
simplest form the deformation and rotation tensors are
1 1
E = (∇u + u∇) and W = (∇u − u∇)
2 2

• Rates of strain similarly address the evolutionary aspects of deformation,


with rate of strain and vorticity tensors
1 1
D = (∇v + v∇) and Ω = (∇v − v∇)
2 2

• Eigenvalues for the strain and rate of strain tensors give principal strains
and special deformations

29
Table of contents 30

3.1 Finite deformation


Recall that the motion of a rigid body can be decomposed, for the purpose of the theoretical
description, into a translation and a rotation. For deformable bodies, which cannot be
regarded as strictly rigid, the motion will have a third aspect, that of deformation.
Two ways of looking at the motion are possible. First, without bothering about the evolution
through intermediate states, one can simply compare the initial and the final states. This
comparison will yield a certain translation, rotation and deformation. On the other hand,
one might be interested in the motion itself, and then one gets at each time a translation
velocity, an angular velocity and a rate of deformation.
In the present section we start with the comparison of the initial and final states and leave
the velocities for a subsequent section. In Figure 3.1 both states to be compared are drawn
with respect to the same reference system. Fix a reference frame with a Cartesian coordi-
nate system. The coordinates in the initial state, a, are called the material or Lagrangian
coordinates, whereas those in the final state, x, are called the space or Eulerian coordinates.
The particle formerly at P0 is displaced to P , and has a displacement vector u given by
u = x − a.
In view of the continuum hypotheses, we will suppose that the correspondence between
P0 and P is one-to-one, which simply means that the continuum under consideration is

V0
R0
Q0
db V
da R
P0
Q
u dy
dx
P
a
3
x

O
2
1
Figure 3.1. Comparison of initial and final states.
Table of contents 31

nowhere torn apart nor joined onto another material. This requirement stems from the usual
deterministic view that the initial conditions uniquely determine the subsequent evolution of
a given particle. The one-to-one correspondence between P0 and P , which is now built into
the theory, demands quite some caution and adaptation before the theory can be applied to
problems such as fractures of elastic materials. As a further requirement, the displacement
vector is a continuous function of the position in either the initial or the final state.
Using the material coordinates or, equivalently, tagging all particles by their initial positions,
we find that
x=x e(a, t),
and thus for the displacement vector

e(a, t) − a.
e (a, t) = x
u=u

Conversely, using the space coordinates, we can express the original positions of the particles
now at x through
a=b a(x, t).
Similarly, the displacement vector becomes

b (x, t) = x − b
u=u a(x, t).

Both descriptions refer to the same comparison between an initial and a final state, hence the
transformations x = xe(a, t) and a = ba(x, t) should be each others inverse. This in particular
implies that the Jacobian determinant of the transformation be different from zero, and even,
for material continua, that it be positive,
   
∂ −1 ∂
J = det x
ej > 0, J = det aj > 0.
b
∂ai ∂xi

This is based on the reasoning that if the initial and final states are smoothly connected, then
the Jacobian tends to 1 when the final state approaches the initial state. The uniqueness
requirements prevent the Jacobian from going through zero, so that it has to keep a positive
sign. The transformation between initial and final state is orientation preserving.
In order to separate the translation and rotation aspects of the motion from the deformation
aspects, we set out with three neighbouring points, P0 , Q0 and R0 in the initial state and
consider the corresponding points in the final state, P , Q and R. How these particles got
there is irrelevant for the time being. From P0 to Q0 we have the infinitesimal vector da,
corresponding in the final state to dx from P to Q. With the help of the transformations
relating a and x one gets
 
∂ ∂e
xi
dx = da · e
x dxi = daj ,
∂a ∂aj
 
∂ ∂b
ai
da = dx · b
a dai = dxj .
∂x ∂xj
Table of contents 32

Repeating the argument for the pair of points (P0 , R0 ) and (P, R), denoting the separation
vector resp. db and dy, we find analogously
 
∂ ∂exi
dy = db · e
x dyi = dbj ,
∂a ∂aj
 
∂ ∂bai
db = dy · b
a dbi = dyj .
∂x ∂xj
For the simplicity of subsequent calculations we abbreviate the position gradients as
∂ ∂
F= e,
x F−1 = G = b
a,
∂a ∂x
and rewrite the respective differentials as

dx = da · F, dy = db · F
da = dx · G, db = dy · G.

The scalar product dx · dy contains information about the distances between P and Q and
P and R, as well as about the angle QP [ R. Subtracting similar information from the initial
state, contained in da · db, we calculate successively

dx · dy − da · db = (da · F) · (db · F) − da · db
= da · F · Ft · db − da · db
= da · (F · Ft − 1) · db = 2da · L · db.

The symmetric tensor L is called the finite or nonlinear Lagrangian strain tensor of Green
and StVenant. That we do indeed find a tensor can be proved in different ways, either
directly on the definition
1
L = (F · Ft − 1)
2

or via the fact that in


dx · dy − da · db = 2da · L · db
L yields a scalar for an arbitrary choice of the (infinitesimal) vectors da and db.
In a similar vein we get, via the Eulerian description,

dx · dy − da · db = dx · dy − (dx · G) · (dy · G)
= dx · (1 − G · Gt ) · dy = 2dx · E · dy,

and the tensor E is the finite or nonlinear Eulerian strain tensor of Almansi and Cauchy.
The definition
1
E = (1 − G · Gt )
2
Table of contents 33

shows that it is also a symmetric tensor.


To see indeed that L and E merit the name of strain tensors, or put differently, that these
tensors measure pure strain or deformation, we let for a while Q0 and R0 , or Q and R, refer
to the same particle and find that

dx · dx − da · da = dx2 − da2 = 2L : da da = 2E : dx dx.

The left-hand side of this expression refers only to the distance between two particles in the
initial state and to the distance between the same two particles in the final state. Note that
for rigid bodies we could pick any two (not even neighbouring) particles and, by definition
of a rigid body, the distance always remains invariant. This implies that for rigid bodies L
and E have to vanish (i.e. F and G are represented by orthogonal matrices). Conversely, if L
and E vanish in the neighbourhood of a particle P , we find that the distance between P and
a neighbouring particle Q remains invariant during the motion, hence the neighbourhood
of P behaves as if it were part of rigid body. This shows that L and E indeed measure the
deformation or the strain around P0 or P .
Instead of expressing L and E in components of the coordinates, we may just as well use the
displacements. Combining
u=u e(a, t) − a
e (a, t) = x
with the original definition of L gives us in index notation
   
1 ∂e
ui ∂eui
Ljk = δij + δik + − δjk
2 ∂aj ∂ak
 
1 ∂e
ui ∂eui ∂e ui ∂eui
= δij δik + δij + δik + − δjk
2 ∂ak ∂aj ∂aj ∂ak
 
1 ∂e uj ∂e uk ∂e ui ∂eui
= + + .
2 ∂ak ∂aj ∂aj ∂ak

By similarly substituting
b (x, t) = x − b
u=u a(x, t)
into the expression for E we obtain
   
1 ∂b
ui ∂b
ui
Ejk = δjk − δij − δik −
2 ∂xj ∂xk
 
1 ∂b
ui ∂b
ui ui ∂b
∂b ui
= δjk − δij δik + δij + δik −
2 ∂xk ∂xj ∂xj ∂xk
 
1 ∂b uj ∂buk ∂bui ∂bui
= + − .
2 ∂xk ∂xj ∂xj ∂xk

These expressions will be used in the next section to determine the small strain tensors.
Table of contents 34

3.2 Infinitesimal deformation or small strain


If the components of the gradient of the displacement are sufficiently small, in such a way that
the products of those components can be neglected compared to the components themselves,

∂e
u i ∂e
u i ∂e
u j ∂b
u i ∂b
u i ∂b
u j
≪ or ≪
∂aj ∂ak ∂ak ∂xj ∂xk ∂xk ,

one says that the continuum is subjected to a small or infinitesimal strain or deformation.
Applying the above approximation in the last obtained expressions for L and E reduce these
to
 
1 ∂e uj ∂e uk
Ljk = + ,
2 ∂ak ∂aj
 
1 ∂b uj ∂b
uk
Ejk = + .
2 ∂xk ∂xj

Written in tensor notation this is


1 
L= ∇a u e )t ,
e + (∇a u
2
1 
E= ∇x u b )t ,
b + (∇x u
2
and it is obvious that the distinction between a Lagrangian and an Eulerian description can
usually be omitted, the coordinates then being designated by x in both cases. Keeping this
in mind, henceforth we will use E to denote also the linear or small strain tensor,

1  1
E= ∇u + (∇u)t = (∇u + u∇)
2 2

The notation u∇ will be used instead of (∇u)t , whenever no confusion is possible, and is
inspired by the fact that for ordinary vectors (vu)t is just uv.
Now E is the symmetric part of the gradient of the displacement, and the antisymmetric
part of ∇u will be called the rotation tensor W, with

1  1
W= ∇u − (∇u)t = (∇u − u∇)
2 2

This decomposition of ∇u is unique:

∇u = E + W.

We yet have to show that, for small strain, W merits the name of “rotation tensor”. To do
Table of contents 35

V0
Q
u′
Q0

dx

P0 u

Figure 3.2. Decomposition of neighbouring displacements.

so, the displacement u′ = u(Q0 ) of a particle Q0 in the neighbourhood of P0 , as indicated


in Figure 3.2, is expressed through a Taylor expansion around the displacement of P0 ,

u′ = u(xQ0 ) = u(xP0 + dx) = u(xP0 ) + dx · ∇u(xP0 )


= u + dx · ∇u = u + dx · E + dx · W.

The first term on the right-hand side of the last equation is the translation part of the
motion, the second term refers to the pure deformation and the third part is nothing but the
rotation. To see this, we introduce the concept of the dual vector of an antisymmetric tensor
of order two such as W. The notion of a dual vector is restricted to a three-dimensional
space, because only there a vector and an antisymmetric tensor of order two both have three
components, so that a one-to-one correspondence between the information contained in these
is possible. For an antisymmetric tensor with components Wkℓ , define in a unique way the
vector w with components
1
wj = εjkℓ Wkℓ .
2
Conversely, it follows that
1 1 1
εpqj wj = εpqj εjkℓ Wkℓ = δpk δqℓ Wkℓ − δpℓ δqk Wkℓ
2 2 2
1 1
= Wpq − Wqp = Wpq ,
2 2
Table of contents 36

or summarized together,
1
wj = εjkℓ Wkℓ , Wkℓ = εkℓj wj .
2
If W vanishes, so does w and vice versa. Furthermore, one can check that

(v · W)i = vk Wki = vk εkij wj = εijk wj vk = (w × v)i

and hence the expression for u′ can be rewritten as

u′ = u + w × dx + E · dx.

This shows indeed that W is giving the rotation part of the motion, namely a rotation over
an angle kwk. The explicit expression of w is
1 1
wj = εjkℓ Wkℓ = εjkℓ (∂k uℓ − ∂ℓ uk )
2 4
1 1 1 1
= εjkℓ ∂k uℓ + εjℓk ∂ℓ uk = εjkℓ ∂k uℓ = (∇ × u)j
4 4 2 2
or
1 1
W = (∇u − u∇) ⇐⇒ w = ∇×u
2 2

3.3 Geometric interpretation of small strain


For the interpretation of the diagonal elements of E, we return to the Lagrangian view point,

dx · dx − da · da = 2E : dada

and take da parallel to the first axis,

dx2 − da2 = 2E11 da2 ,

so that
dx2 − da2
E11 = .
2da2
This yields several expressions which are equivalent within the small strain approximation,

(dx + da)(dx − da) 2da(dx − da) dx − da


E11 = 2
≃ 2
= .
2da 2da da
This shows that the diagonal elements of the small strain tensor give the relative stretching
along the coordinate axes. This is a quantity which can be measured and is often used in
experimental setups.
Table of contents 37

∂u1
2 tan β =
∂x2
β

∂u2
tan α =
∂x1
α

1
O

Figure 3.3. Deformation of angles.

Once the meaning of the diagonal elements is intuitively clear, we proceed to the off-diagonal
elements. We use
dx · dy − da · db = 2E : dadb
with da parallel to the first axis and db parallel to the second axis and find
π
dx dy cos θ − da db cos = 2E12 da db.
2
Up to second order terms, we have that
dx dy
2E12 = cos θ ≃ cos θ.
da db
As shown in Figure 3.3, we can express the cosine of θ as follows
π 
cos θ = cos − α − β = sin(α + β) ≃ α + β
2
∂u2 ∂u1
≃ tan α + tan β = + = 2E12 ,
∂x1 ∂x2
where it has been taken into account that α and β are assumed to be very small. Hence we
have that the off-diagonal elements of the small strain tensor give half of the change in angle
between two line elements which were perpendicular to each other before the deformation.
This change in angle is called shear. The reason that there is a factor two is that notions like
relative stretching or shear evolved experimentally long before it was recognized that these
were parts of what is now known to behave as a (small) strain tensor.

3.4 Rate of deformation


Up to now the discussion of the deformation was confined to a comparison between an initial
and a final state, without paying any attention to the intervening states or to the motion
Table of contents 38

v′
Q

dx

P v

Figure 3.4. Velocities of neighbouring particles.

involved in getting from one state to the other. To get some idea of the velocities occurring,
we Taylor decompose the velocity of a particle Q in the neighbourhood of a particle P , in
much the same way as we did with the displacements before and as shown in Figure 3.4, so
that

v′ = v(xQ ) = v(xP ) + dx · ∇v(xP ) = v + dx · ∇v


1 1
= v + dx · (∇v + v∇) + dx · (∇v − v∇) = v + dx · D + dx · Ω
2 2
or also
v′ = v + ω × dx + dx · D.
The new quantities introduced in the above are respectively the rate of deformation tensor
D,
1 1
D = (∇v + v∇) or Dij = (∂i vj + ∂j vi ),
2 2

the vorticity tensor Ω,

1 1
Ω = (∇v − v∇) or Ωij = (∂i vj − ∂j vi ),
2 2

and its dual vector, the vorticity (vector) ω

1 1
ω = ∇×v or ωi = εijk ∂j vk
2 2

The rate of deformation tensor D is a symmetric tensor, the vorticity tensor Ω is anti-
symmetric, and they are respectively the symmetric and antisymmetric part of the velocity
Table of contents 39

gradient tensor ∇v. The decomposition of v′ shows that the motion contains three different
elements, at least conceptually, since all aspects occur simultaneously in real motion. These
three different elements are a translation velocity v, a instantaneous rotation with angular
velocity ω and a pure rate of deformation given through D.
Another way of arriving at D is by taking the time derivative of dx·dx−da·da (= 2E : dxdx),
which yields
d d d
(dx2 − da2 ) = (dx · dx) = 2dx · (dx).
dt dt dt
We need to compute separately
d d d
(dx) = (da · ∇a xe) = da · (∇a xe)
dt dt dt
d
= da · ∇a x = da · ∇a v = dv = dx · ∇x v,
dt
and finally get
d
(dx2 − da2 ) = 2dx · dv = 2dx dx : ∇v
dt
= 2dx dx : D + 2dx dx : Ω = 2dx · D · dx.

Here again, the interpretation is that if the neighbourhood of a particle P instantaneously


behaves as a rigid body, then the corresponding rate of deformation tensor vanishes.
In order to get some relation between the strain tensor E and the rate of deformation tensor
D, we calculate that
d d
2dx · D · dx = (dx2 − da2 ) = (2dx · E · dx)
dt dt
= 2dv · 
E · dx + 2dx · Ė · dx + 2dx
 · E · dv
= 2dx · Ė + ∇v · E + E · (∇v)t · dx.

It follows that
D = Ė + ∇v · E + E · (∇v)t

because
dx · A · dx = 0 (∀dx) ⇐⇒ A = 0,
provided A is a symmetric tensor. If the deformation is infinitesimal, the last two terms on
the right-hand side of the previous equation may be neglected and there only remains

D ≃ Ė.
Table of contents 40

3.5 Principal strains and special deformations


Both tensors E and D are symmetric and we will confine the discussion to one of the tensors,
say E. The eigenvalue problem for E,
E · n = εn,
means that one is looking for those directions n which remain parallel to themselves under
pure deformation, that is when neither translation nor rotation occurs. The eigenvalues are
the roots of 
 ε1
3 2
det(E − ε1) = −ε + IE ε − IIE ε + IIIE = 0 ε2

ε3
with invariants
1 
IE = tr E, IIE = tr E tr E − E2 , IIIE = det E.
2
Since E is symmetric, the eigenvalues are real and there exists at least three mutually or-
thogonal eigenvectors. After normalization, these eigenvectors can be used as the base of a
coordinate system in which E is diagonal. There are thus at least three different directions
which remain unchanged when the continuum is subjected to pure deformation.
The first invariant IE has a special meaning and is often called the coefficient of cubic
dilatation in small strain. To see where this name comes from, one considers an elementary
parallelepiped around P0 , with volume
dV0 = da · (db × dc),
where da, db, dc are parallel to the three coordinate axes, respectively. Use
dx = da · F = da · (1 + ∇a u)
= da + da · E + w × da = (1 + ε1 )da + w × da
and analogous relations between dy and db, and between dz and dc, where the computations
are done in the principal coordinate system for E. We thus find that
dV = dx · (dy × dz)
= [(1 + ε1 )da + w × da] · {[(1 + ε2 )db + w × db] × [(1 + ε3 )dc + w × dc]}
≃ (1 + ε1 + ε2 + ε3 )da · (db × dc) + da · [db × (w × dc)]
+ da · [(w × db) × dc] + (w × da) · (db × dc)
= (1 + tr E) dV0 .
The relative change in volume is thus
dV − dV0
= tr E.
dV0
According to whether some of the invariants of E vanish, we can perform a classification
reminiscent of the one given for the stress tensor T. We distinguish the following special
states of strain (and a similar survey could be given for the rate of deformation, of course):
Table of contents 41

• General deformation:

IIIE 6= 0, ε1 , ε2 6= ε3 (possibly ε1 = ε2 ).

The stretching for a particular direction n is defined as

ε(n) = nn : E = ni nj Eji = n2i εi ,

where the last expression is only valid in a principal coordinate system based on the
strain tensor eigenvectors.
• Plane deformation:

IIIE = det E = 0, IIE 6= 0, ε1,2 6= 0, ε3 = 0.

All displacements for pure deformation are parallel to one of the coordinate planes of
the system of principal axes, here the (1,2)-plane.
• Linear deformation:

IIIE = IIE = 0, IE 6= 0, ε1 6= 0, ε2,3 = 0.

All displacements for pure deformation are now parallel to one of the axes of the
principal coordinate system, here the first one.
• Spherical deformation:

ε1 = ε2 = ε3 = ε, E = ε 1.

For every direction there is no shear and the stretching is the same, equal to ε. This
implies that every linear dimension is multiplied by the same number and the trans-
formation from the initial to the final state is a homothetic one, changing the volume
of the continuum but neither the form nor the shape. In every coordinate system the
strain tensor is a multiple of the unit tensor.

For a general state of strain it is possible to distinguish between changes in volume and in
form. Every tensor, not just E or a symmetric one, can be split in a unique way into a
spherical part and a remainder, with the restriction that the remainder be traceless. Such a
traceless part is then called the deviatoric part of the tensor. Thus one finds for E that
1 1
E = ε 1 + E′ , ε = tr E = ∇ · u, tr E′ = 0.
3 3
In the special case of E, the spherical part is a third of the coefficient of cubic dilatation
and indicates a change in volume without a change in form (homothetic transformation).
The deviatoric part E′ is a deformation where the form changes without a change in volume,
since the coefficient of cubic dilatation for E′ vanishes. Every deformation thus consists
conceptually of two parts: a change in volume without change in form, and a change in
form without change in volume. Furthermore, the principal axes for E and E′ coincide, as
the spherical part is diagonal in any representation.
Table of contents 42

3.6 Compatibility conditions for small strain


Up to now, the definitions of E and D have been interpreted as a way of calculating E or
D, given the displacements u or the velocities v. There are six scalar equations giving Eij
(or Dij ), starting from the three components ui (or vi ). The inverse problem consists in
determining the displacement u when the strain tensor E is given. It is clear that without
special precautions this will be impossible, as we will have twice as many equations as
there are unknowns. Consequently, not every tensor is admissible as a strain tensor. Some
relations between the components of E have to exist in order to allow the determination of u.
This question is not only of mathematical relevance but has some experimental implications,
as it is the deformations rather than the displacements which are usually measured in an
experiment. The mathematical compatibility conditions then amount to consistency relations
for the experimental data.
The required compatibility conditions are found upon elimination of u from
1
(∇u + u∇) = E.
2
Taking the curl of both sides gives
1 1
∇ × E = ∇ × (∇u + u∇) = ∇ × u∇ = w∇ = (∇w)t ,
2 2
with the help of the particular expression
1
w = ∇×u
2
for the dual vector w of the rotation tensor W. From
(∇ × E)t = ∇w
we find that
∇ × (∇ × E)t = 0

This is the required compatibility condition of StVenant, rewritten in index notation as


εikℓ εjmn ∂k ∂m Enℓ = 0.
Due to the symmetry of E, this is also symmetric and hence equivalent to six scalar equations,
three of the form
2∂1 ∂2 E12 = ∂1 ∂1 E22 + ∂2 ∂2 E11 ,
and three of the form
∂1 ∂2 E33 = ∂3 (∂1 E23 + ∂2 E13 − ∂3 E12 ).
In order to completely close the discussion, we will show that the compatibility condition
of StVenant is not only necessary but also sufficient to deduce the structure of E, if E is
supposed symmetric. From
∇ × (∇ × E)t = 0
Table of contents 43

there follows at once that


(∇ × E)t = ∇d,
and hence also that
∇ × E = d∇.
Taking the trace of this tensor yields

∇ · d = tr ∇ × E = εijk ∂j Eki = 0,

since the tensor of Levi-Cività is completely antisymmetric and we have assumed that E is
symmetric. As d is divergence-free, it can be written as the curl of yet another vector e and
thus
∇ × E = ∇ × e∇.
From this one finds that
E = e∇ + ∇f .
Because of the symmetry of E this implies that

e∇ + ∇f = ∇e + f ∇,

or also that
∇(f − e) = (f − e)∇.
The index notation,
∂i (fj − ej ) = ∂j (fi − ei ),
shows that the vector expression is equivalent to

∇ × (f − e) = 0.

Hence f − e has to be a gradient of a scalar,

f − e = ∇φ,

and consequently,

  
1 1
E = e∇ + ∇(e + ∇φ) = ∇ e + ∇φ + e + ∇φ ∇
2 2
has the structure required from a strain tensor.
There is also a compatibility condition on W, which is much easier to express through its
dual vector:
1
w = ∇ × u ⇐⇒ ∇ · w = 0.
2
Hence
εijk ∂i Wjk = 0.
The foregoing discussion was carried out for E and W (or w), and a similar discussion can
be given, of course, for D and Ω (or ω).
Chapter 4

Constitutive relations

• Notions of frame-indifference are used to construct constitutive relations


which connect strain or rate of strain with stress

• General isotropic media are known as Reiner-Rivlin fluids, with constitutive


relation

T = C(ρ, D) = φ0 (ρ, ID , IID , IIID )1 + φ1 (ρ, ID , IID , IIID )D


+ φ2 (ρ, ID , IID , IIID )D2

• Simplifications occur for linear and isotropic media, recovering well known
laws as Navier-Stokes (fluids)

T = − p(ρ)1 + λ(tr D)1 + 2µD

and Hooke (elastic solids)

T = λ(tr E)1 + 2µE

• More general constitutive laws are possible, as for viscoelastic media with
memory effects Z +∞
T(t) = C(τ ) : E(t − τ ) dτ
−∞

where causality allows the existence of Kramers-Kronig relations


Z +∞ b′′ (ξ)
b′ (ω) = 1 P
C
C

π −∞ ξ−ω
Z +∞ b′ (ξ)
b′′ (ω) = − 1 P
C
C

π −∞ ξ−ω

44
Table of contents 45

4.1 Young’s experiment


In the previous chapters the notions of stress and strain or deformation were introduced in
general terms and the discussion was given without any attempt to relate the stress in a
given continuum to the strain or deformation it was undergoing. Stress and strain occur
in all continua, but the as yet incomplete description through the basic laws has to be
valid for many different classes of material. The stress response to a certain strain will
vary from one material to another, not just in numerical values but also in the form of the
functional dependency. Hence the name of constitutive relations for the various stress-strain
dependencies we will survey.
The simplest starting point is to suppose that the stress depends on the strain in a linear way,
without other parameters such as the temperature or the mass density entering the relation.
To see where these ideas come from we will briefly, and for the present only superficially,
sketch Young’s experiment, in which a prismatic bar with cross-section S is subjected to
a force F in the principal direction of the bar, as shown in Figure 4.1. For several elastic
media the result was that for moderate forces the elongation ∆ℓ increased linearly with the
original length ℓ of the bar at rest and with the magnitude F of the force applied at one end
of the bar, the other end remaining clamped in position. At the same time ∆ℓ decreased
with the cross-section S of the bar, all in direct proportionality. Accordingly the result is
ℓF
∆ℓ ∼ .
S
If we label the principal direction of the bar the first direction, we can call the relative
elongation in this direction the element E11 of the strain tensor for this problem, keeping in
mind the intuitive geometric interpretation of the diagonal elements of the strain tensor for
small strains. On the other hand, the ratio F/S is the force per unit surface area, with the
dimensions of a pressure (or traction!), nothing else but T11 . Hence with
∆ℓ F
E11 = , T11 =
ℓ S
we have
1
E11 = T11 or T11 = EE11 .
E
The constant of proportionality E is called Young’s modulus of elasticity. In the above
simple relationship it just converts the strain into the stress. The relationship is linear in the

2
ℓ ∆ℓ
F
S
1
O

Figure 4.1. Young’s experiment.


Table of contents 46

sense that doubling the strain results in a doubling of the stress. Furthermore, although this
is not obvious at all yet, in the above experiment we may assume the material is isotropic,
meaning that it displays the same material properties in all directions.
To give a better feeling for this, consider a cube of isotropic material. We first subject it to a
given tension in one direction and measure the corresponding stretching. Then, if we repeat
the experiment in one of the other two directions, we expect to find the same elongations if
the tensions are the same. In other words, for isotropic materials such as metals, the way
in which a bar is cut from an ingot is not supposed to affect its elastic properties. Many
materials, of course, are not isotropic. Crystals are a notable example, because they have
specific symmetry directions, with different properties along different directions. Isotropic
materials have simpler constitutive relations than anisotropic ones and we will confine our
study mostly to isotropic materials.
A further distinction is to whether the strain or the rate of strain causes the stress. In the
example of an extended rubber, the stress is present as long as the extension is applied, even
when constant. If there is no strain, there is no stress, and such a material seems to possess
a natural and stress-free reference state. In generalizing these intuitive ideas, we will call a
material elastic if it has a constitutive relation of the form

T = C(E), C(0) = 0.

On the other hand, if we think of fluids, there will be stress, apart from the residual hydro-
static pressure which is isotropic, only as long as the strain or deformation changes. Thus
we deal with fluids if
T = C(D), C(0) = γ 1.
In the two previous cases, C denotes a (tensor valued) function of its argument, sometimes
called “response function”. These two classes of materials can be generalized in many ways,
one of which is to allow the symbol C to stand for a functional dependence of T with respect
to E or D, rather than being an ordinary function, in that it may involve an integration of
E or D over the medium and over time. This means that the stress at a given point and
instant then depends on the strain at other points or at previous instants. This give rise to
materials with memory. Some of this will be discussed in section 4.5 on visco-elastic media.

4.2 Frame-indifference
For linear and isotropic elastic media, the constitutive relation (to be derived later) is called
Hooke’s law, whereas the analogous relation for fluids is called Navier-Stokes’ law. These
laws have evolved historically in an ad hoc fashion, based on experimental evidence, but can
now be derived in a systematic way under the twin assumptions that the stress be a linear
function of the strain or rate of strain and this for isotropic media. However, how does one
know that, to cite only one parameter, the mass density is not to be included in a constitutive
relation? More generally, are there quantities which could be included in admittedly more
general constitutive relations, whereas some other quantities might not?
Table of contents 47

A powerful instrument in the modern theory of constitutive relations has been the principle
of frame-indifference, saying that only frame-indifferent quantities should appear in consti-
tutive relations. Before we give a definition in mathematical terms of what frame-indifferent
quantities are, we have to go back to the definition of vectors and tensors and their trans-
formation laws under an orthogonal transformation (cf. section 1.3)

x′i = Aij xj , Tijk... = Aip Ajq Akr · · · Tpqr... .

Although it was not explicitly stated, these transformations are to be interpreted in a ge-
ometric sense, where the change from one reference frame to another is really a change of
coordinates involving no time dependence, no motion. If we now admit changes of reference
frame such that the elements Aij of the (orthogonal) transformation matrix can be functions
of t and, in addition, we allow for time-dependent translations of the origin, we will find
that not all quantities encountered sofar will still transform according to the transformation
rule for vectors and tensors given above. Those quantities which still do are called frame-
indifferent or are said to obey the principle of frame-indifference or of material objectivity.
A time-dependent change of reference frame will in general be of the form

x′i = Aij (t)xj + ci (t),

which may be written in matrix notation as

x′ = Ax + c,

where the transformation matrix A = [Aij (t)] is orthogonal. (Strictly speaking one could
also allow for a time translation, t′ = t + a, but we will not consider that.) Note in passing
that we are working in the framework of Galilean relativity, i.e. we do not consider the
modifications imposed by special relativity.
Vectors and tensors will be denoted through their component representation as column and
square matrices,  
x1
x = x2  ,
 T = [Tij ].
x3
In this matrix notation, uv t represents the tensor product uv of two vectors u and v, and
ut v their inproduct u · v. Primes then refer to the same vectors or tensors but represented
in a different reference frame. Similarly, the symbolic nabla vector is
 
∂1
∇ = ∂2  ,

∂3

and we obtain under the above change of reference frame


∂ ∂
∂i′ = ′
= Aij = Aij ∂j ,
∂xi ∂xj
Table of contents 48

so that
∇′ = A ∇.
Surprisingly, in view of its derivative operator character, it is frame-indifferent. A displace-
ment vector u transforms according to

u′ = x′Q − x′P = A(xQ − xP ) = Au.

Displacements hence are frame-indifferent in the sense given above. What about the velocity?
Offhand, we would not suppose the velocity to be frame-indifferent if the change of reference
frame is time-dependent. This is confirmed by the computation
d
v ′ = ẋ′ = (Ax + c) = Av + Ȧx + ċ.
dt
For the stress tensor T there is no way of computing its frame-indifference, hence we have
to impose, on more or less plausible physical grounds, that

T ′ = AT At .

In order to check the strain and rotation tensors, one first has to compute

(∇ut )′ = ∇′ (u′ )t = A∇(Au)t = A(∇ut )At .

Once the gradient of the displacement tensor is proven frame-indifferent, its symmetric part
E and its antisymmetric part W also are frame-indifferent. This conclusion is valid even for
the full finite strain tensor,
1 
E= ∇u + (∇u)t − ∇u · (∇u)t ,
2
which is of the utmost importance if one tries to describe media in which the deformations
are not infinitesimal.
Lastly we compute the gradient of the velocity tensor in transformed coordinate systems,

(∇v t )′ = ∇′ (v ′ )t = A∇(Av + Ȧx + ċ)t


= A(∇v t )At + A(∇xt )Ȧt = A(∇v t )At + AȦt ,

and find that


1
D′ = (∇v t + v∇t )′
2
1 1 1
= A(∇v t + v∇t )At + AȦt + ȦAt = ADAt .
2 2 2
Maybe rather unexpectedly, D is frame-indifferent, because, even though time dependent,
the coordinate transformations still are between orthonormal systems and thus

AAt = 1 =⇒ ȦAt + AȦt = 0.


Table of contents 49

On the other hand, one finds for the vorticity tensor Ω that
1 1 1
Ω′ = (∇v t − v∇t )′ = A(∇v t − v∇t )At + (AȦt − ȦAt )
2 2 2
t t t t
= AΩA + AȦ = AΩA − ȦA ,

which is not frame-indifferent. This should come as no surprise, as it is related to the


instantaneous angular velocity and depends very much upon whether the frame moves and
rotates or not.
The question remaining after having found that T, E, W and D are frame-indifferent is to
see whether their time derivatives are also frame-indifferent or not. Most likely they will
not. The computation will be given for the stress tensor, but is analogous for the others.
We start from
T ′ = AT At
and get
Ṫ ′ = AṪ At + ȦT At + AT Ȧt .
Clearly Ṫ is not frame-indifferent. We will, however, construct a tensor involving Ṫ and
which we can use as a frame-indifferent measure for the rate of change of the stress tensor.
Contracting
Ω′ = AΩAt + AȦt
with At yields

Ȧt = At Ω′ − ΩAt ,
Ȧ = AΩ − Ω′ A.

We use this to rewrite Ṫ ′ as

Ṫ ′ = AṪ At + (AΩ − Ω′ A)T At + AT (At Ω′ − ΩAt )


= A(Ṫ + ΩT − T Ω)At − Ω′ T ′ + T ′ Ω′ ,

and rearrange the terms into

Ṫ ′ + Ω′ T ′ − T ′ Ω′ = A(Ṫ + ΩT − T Ω)At .

This shows the bracketed expression to be frame-indifferent, and it is called the rate of stress
tensor of Jaumann,

T = Ṫ + Ω · T − T · Ω.
It involves Ṫ, as well as combinations of T with the vorticity tensor. When Ω vanishes,
as in the case of no instantaneous rotation, Jaumann’s rate of stress tensor coincides with
Ṫ. Clearly, other choices for the stress rate are possible, by adding or subtracting frame-
indifferent combinations to Jaumann’s tensor. Some of these stress rates are Truesdell’s
stress rate,

T = Ṫ + T∇ · v − T · ∇v − (T · ∇v)t
Table of contents 50

and Green’s stress rate,



T = Ṫ − T · ∇v − (T · ∇v)t .
Similar expressions can be constructed, if need be, for the strain rates involving time deriva-
tives of E and D.
At the end of this section, it is worth stressing that the principle of frame-indifference or
material objectivity can only be applied to those parts of the description which can reason-
ably be thought to be independent of the motion of the reference frame. For constitutive
relations this is valid, to a fair approximation, but it would be the height of folly to apply
such a reasoning to the equations of motion!

4.3 Isotropy and Reiner-Rivlin fluids


As an example of how one would proceed in constructing a constitutive relation for isotropic
media, suppose that for a given medium the stress is thought to depend upon the mass density
ρ, the velocity v and the gradient of the velocity ∇ v. In the above enumeration, ρ is just
taken as a representative to show what happens to other scalar frame-indifferent quantities.
We could have included the temperature as another example. In the end, wherever ρ is
written we may put other scalar quantities as well. The dependence of T upon ∇ v will be
through its symmetric part D and its antisymmetric part Ω. Hence we start from
e v, ∇v) = C(ρ, v, D, Ω).
T = C(ρ,

One observer will write this as


T = C(ρ, v, D, Ω),
while another observer in a moving frame will denote this as

T ′ = C(ρ′ , v ′ , D′ , Ω′ ).

We have here the same function C because the medium is supposed isotropic, which will turn
out to be a severe restriction. If we think again for a moment of Young’s experiment, this
amounts to requiring Young’s elasticity modulus to be independent of a particular choice of
coordinate system. If the medium is anisotropic but has some kind of symmetry, a relation
as written above will only hold for certain observers, not for all of them. For isotropic media
we thus require that
 
T ′ = C ρ, Av + Ȧx + ċ, ADAt , AΩAt + AȦt = AT At = AC(ρ, v, D, Ω)At ,

regardless of the particular choices for A and c. We pick a specific transformation such that
at t = 0:

A(0) = 1, c(0) = 0, ċ(0) = 0, Ȧ(0) = Q 6= 0 (Qt = −Q).


Table of contents 51

An example of this would be


   
cos t sin t 0 1 0 0
A(t) =  − sin t cos t 0  , A(0) =  0 1 0  ,
0 0 1 0 0 1

with    
− sin t cos t 0 0 1 0
Ȧ(t) =  − cos t − sin t 0  , Ȧ(0) =  −1 0 0  .
0 0 0 0 0 0
For t = 0 we find that

C(ρ, v + Qx, D, Ω − Q) = C(ρ, v, D, Ω).

C cannot be a function of v, as we could add whatever we wanted to v through Qx without


changing the value of the function. For the same reason C cannot depend upon Ω.
Generally, all quantities which are not frame-indifferent have to drop out of the constitutive
relations and we are left with
T = C(ρ, D).
We thus should have for isotropic tensor functions that for any A

AC(ρ, D)At = C(ρ, ADAt )

Picking a special form for A,


A = 2nnt − 1,
with n a unit vector (i.e. nt n = 1), the direction of which will be specified later, we first
verify that
AAt = (2nnt − 1)(2nnt − 1) = 1.
The transformation property T ′ = AT At becomes for this choice of A

T ′ = 4n(nt T n)nt − 2n(nt T ) − 2(T n)nt + T.

The transformation represented by A is nothing but a rotation over 180◦ along the direction
given by n. We now choose n to be an eigenvector of D, such that in matrix notation

Dn = δn

and hence

D′ = ADAt = 4n(nt Dn)nt − 2n(nt D) − 2(Dn)nt + D


= 4nδnt − 2nnt δ − 2δnnt + D = D.

Using this in
AC(ρ, D)At = C(ρ, ADAt )
Table of contents 52

gives
   
AC(ρ, D)At = 4n nt C(ρ, D)n nt − 2n nt C(ρ, D) − 2 [C(ρ, D)n] nt + C(ρ, D) = C(ρ, D),
or also    
n nt C(ρ, D) + [C(ρ, D)n] nt = 2n nt C(ρ, D)n nt .
Post-multiplying with n gives
 
C(ρ, D)n = n nt C(ρ, D)n = γn,
showing that n is an eigenvector of C(ρ, D) as well. The above proof indicates that every
eigenvector of D is an eigenvector of C(ρ, D). Hence C(ρ, D) is diagonal whenever D is
diagonal. A first, intermediate result is that if D is isotropic, every direction yields an
eigenvector,
D = δ1
and the resulting stress is uniform,
T = C(ρ, δ 1) = γ(ρ, δ)1.
A uniform rate of deformation gives uniform stress in isotropic media, regardless of whether
the constitutive relation is linear or not, complicated or straightforward.
Returning to the general form of C(ρ, D) for isotropic media, it is now easy to show that as
far as its tensor character is concerned, it has to be of the form

C(ρ, D) = φ0 (ρ, ID , IID , IIID )1 + φ1 (ρ, ID , IID , IIID )D + φ2 (ρ, ID , IID , IIID )D2

If one rewrites this in principal axes for D and C(ρ, D), there comes for the eigenvalues
γ1 = φ0 + φ1 δ1 + φ2 δ12 ,
γ2 = φ0 + φ1 δ2 + φ2 δ22 ,
γ3 = φ0 + φ1 δ3 + φ2 δ32 .
For all γi there exists a solution φ0 , φ1 , φ2 if the determinant of the coefficients does not
vanish,
1 δ1 δ 2
1

1 δ2 δ22 = (δ1 − δ2 )(δ2 − δ3 )(δ3 − δ1 ),

1 δ3 δ32
which is the case when the three eigenvalues of D are all different. When there are only two
different eigenvalues for D, the form of C reduces to
C(ρ, D) = φ0 (ρ, ID , IID , IIID )1 + φ1 (ρ, ID , IID , IIID )D,
because then C(ρ, D) also has only two different eigenvalues. When all eigenvalues are equal
the tensors are isotropic,
C(ρ, D) = φ0 (ρ, ID , IID , IIID )1,
Table of contents 53

a case we discussed already. Hence we can take

C(ρ, D) = φ0 (ρ, ID , IID , IIID )1 + φ1 (ρ, ID , IID , IIID )D + φ2 (ρ, ID , IID , IIID )D2

as the most general form, with the proviso that possibly φ2 or φ1 and φ2 vanish. One can
check that the requirement
AC(ρ, D)At = C(ρ, ADAt )
is fulfilled for isotropic fluids, just because the functions φ0 , φ1 and φ2 only depend on the
components of D through the invariant combinations ID , IID and IIID . As φ0 , φ1 and φ2
are functions of the components of D, the constitutive relations are still highly nonlinear
in general. Fluids with such a constitutive relation are called Reiner-Rivlin fluids. The
discussion in this paragraph of the form C(ρ, D) for isotropic media can be repeated in
exactly the same terms for a constitutive relation involving C(ρ, E). As an example, the
discussion in the next section will be given with E instead of D.

4.4 Linear isotropic media


For linear media, such as seemed to be the case with the metal undergoing moderate tensions
in Young’s experiment, the constitutive law is linear (or, more precisely, affine) and thus of
the form
T = C(ρ) : E + K(ρ),
or in index notation,
Tij = Cijkℓ (ρ)Eℓk + Kij (ρ).
The concept of isotropy implies that after a change of reference frame this constitutive law
becomes
Tij′ = Cijkℓ (ρ)Eℓk

+ Kij (ρ).
There are two ways of finding the form of the tensors C and K. The first one is to linearize the
expression for isotropic but general tensor functions, such as we had obtained in the previous
section. As φ0 is multiplied with the unit tensor, it may contain both a term independent of
E and one linear in E, that is, linear in IE . For the next contribution, φ1 is already multiplied
by E and hence must be independent of E, if the total relation is to be linear. For this reason
φ2 has to vanish, being multiplied with terms quadratic in the components of E. Hence

φ0 = κ(ρ) + λ(ρ)IE = κ(ρ) + λ(ρ)tr E,


φ1 = 2µ(ρ),
φ2 = 0,

and the constitutive relation for linear isotropic media is

T = C(ρ) : E + K(ρ) = κ(ρ)1 + λ(ρ)(tr E)1 + 2µ(ρ)E.

We find that

Cijkℓ (ρ) = λ(ρ)δij δkℓ + 2µ(ρ)δiℓ δjk , Kij = κ(ρ)δij ,


Table of contents 54

if C(ρ) connects a symmetric tensor with another symmetric tensor.


Another way of finding the expression for C(ρ) is to express that

Cijkℓ (ρ) = Aip Ajq Akr Aℓs Cpqrs (ρ)

must hold for all coordinate transformations between orthonormal systems. The components
Cijkℓ (ρ) can only contain a given index in pairs, because a transformation of the form

A = 2nnt − 1

would give zero if an index were to occur in an odd number of locations, due to the rotation
over 180◦ around n, for any n. The only components of C that are not zero are

C1111 (ρ) = C2222 (ρ) = C3333 (ρ) = α(ρ),


C1122 (ρ) = C1133 (ρ) = C2211 (ρ) = C2233 (ρ) = C3311 (ρ) = C3322 (ρ) = λ(ρ),
C1212 (ρ) = C1313 (ρ) = C2121 (ρ) = C2323 (ρ) = C3131 (ρ) = C3232 (ρ) = µ(ρ) + ν(ρ),
C1221 (ρ) = C1331 (ρ) = C2112 (ρ) = C2332 (ρ) = C3113 (ρ) = C3223 (ρ) = µ(ρ) − ν(ρ),

as it is obvious that the particular labels attached to some of the coordinate axes do not
matter for isotropic tensors. A transformation is always possible converting the label at-
tached to a given axis into another one. To find a further relation between α(ρ), λ(ρ), µ(ρ)
and ν(ρ), we rotate over an angle θ around one of the axes, say the third one, so that
 
cos θ sin θ 0
A =  − sin θ cos θ 0  .
0 0 1

Calculating C1111 (ρ) we find

α(ρ) = C1111 (ρ) = A1p A1q A1r A1s Cpqrs (ρ)


= A11 A11 A11 A11 C1111 (ρ) + A11 A11 A12 A12 C1122 (ρ)
+ A11 A12 A11 A12 C1212 (ρ) + A12 A11 A11 A12 C2112 (ρ)
+ A11 A12 A12 A11 C1221 (ρ) + A12 A11 A12 A11 C2121 (ρ)
+ A12 A12 A11 A11 C2211 (ρ) + A12 A12 A12 A12 C2222 (ρ)
= (cos4 θ + sin4 θ)α(ρ) + 2 sin2 θ cos2 θ [λ(ρ) + 2µ(ρ)]
= α(ρ) + 2 sin2 θ cos2 θ [λ(ρ) + 2µ(ρ) − α(ρ)] ,

from which it follows that


α(ρ) = λ(ρ) + 2µ(ρ).
The structure of C(ρ) hence is

Cijkℓ (ρ) = λ(ρ)δij δkℓ + µ(ρ)(δik δjℓ + δiℓ δjk ) + ν(ρ)(δik δjℓ − δiℓ δjk ).

If Cikjℓ (ρ) is to be symmetric in the first two or last two indices, as ultimately required by
the symmetry of both the strain and the stress tensors, ν(ρ) has to vanish and we recover
Table of contents 55

the previous expression. Remark that Cijkℓ is also unchanged when the first two indices are
switched with the last two ones,

Cijkℓ (ρ) = Ckℓij (ρ).

As a byproduct of the computation of isotropic tensors of order two (which are a multiple
of the unit tensor), and of order four, one concludes immediately that there are no isotropic
tensors of odd order, as these must at least have one index which occurs an odd number of
times. (Remember that the so-called tensor of Levi-Cività is only a pseudo-tensor of order
three.)
We defined elastic media by a stress-strain relation which vanished for zero strain, i.e. the
strain is measured from the stress-free reference state,

T = C(E), C(0) = 0.

From
T = κ(ρ)1 + λ(ρ)(tr E)1 + 2µ(ρ)E
we find that for linear and isotropic elastic media

T = λ(ρ)(tr E)1 + 2µ(ρ)E.

For the media usually studied in the theory of elasticity λ(ρ) and µ(ρ) have a negligible
dependence upon ρ and can be treated as effectively constant. These are then called the
constants of Lamé and occur in Hooke’s law

T = λ(tr E)1 + 2µE

On the other hand, linear and isotropic fluids are called Newtonian fluids and have a con-
stitutive relation of the form

T = −p(ρ)1 + λ(ρ)(tr D)1 + 2µ(ρ)D.

We have reverted to the usual notation p for the residual hydrostatic pressure, defined as
the opposite of a stress. In general p, λ and µ would still be functions of ρ and other scalar
quantities included in the description, such as the temperature. For the hydrostatic pressure
p this will indeed be the case and we will have to supplement the description by giving a
law of state relating the pressure to the mass density and to the temperature. On the other
hand, experience has shown that for most fluids under the usual conditions of pressure and
temperature, the viscosity coefficients λ and µ are indeed as good as constant and thus
independent of mass density and temperature. Under those conditions, the constitutive
relation is known as Navier-Stokes’ law,

T = −p(ρ)1 + λ(tr D)1 + 2µD


Table of contents 56

Except for the residual hydrostatic pressure, a Newtonian fluid only experiences stresses as
long as the deformation changes. Such a fluid has no natural state and no memory, as anyone
spilling some liquid will immediately observe!
If the tensors T and E or D are split into their spherical and deviatoric parts, as indicated
in chapter 3 for E, giving
T = σ 1 + T′ , tr T′ = 0, σ = 13 tr T,
E = ε 1 + E′ tr E′ = 0, ε = 13 tr E,
D = δ 1 + D′ , tr D′ = 0, δ = 31 tr D,
Hooke’s law can be used to compute first that
1 1 1
σ = tr T = tr [λ(tr E)1 + 2µE] = (3λ + 2µ)tr E = (3λ + 2µ)ε,
3 3 3
and hence
1
T′ = T − σ 1 = λ(tr E)1 + 2µE − (3λ + 2µ)(tr E)1
 3
1
= 2µ E − (tr E)1 = 2µE′ .
3
Hooke’s law is equivalent to the two relations

σ = (3λ + 2µ)ε, T′ = 2µE′ .

Similarly, the law of Navier-Stokes can be written in the form

σ = −p(ρ) + (3λ + 2µ)δ, T′ = 2µD′ .

Use can now be made of such expressions in order to invert Hooke’s law and express the
strain as a function of the stress,
1 1 λ
E= (T − λ(tr E)1) = T− (tr T)1.
2µ 2µ 2µ(3λ + 2µ)
It is time to come back to Young’s experiment, now that we have the adequate relation
between strain and stress. Young’s experiment is characterized by a state of uniaxial stress,
(1)
with t1 = T11 and all other components of T zero. Hence

tr T = T11

and
1 λ
E= T− T11 1.
2µ 2µ(3λ + 2µ)
Compute now E11 as
1 λ λ+µ 1
E11 = T11 − T11 = T11 = T11 ,
2µ 2µ(3λ + 2µ) µ(3λ + 2µ) E
Table of contents 57

and find an expression for the elasticity modulus of Young as a function of Lamé’s constants,

µ(3λ + 2µ)
E= .
λ+µ
There is, however, another part to the strain-stress relation, which we had not discussed in
our preliminary sketch of the experiment, namely
λ
E22 = E33 = − T11
2µ(3λ + 2µ)
 −1
λ µ(3λ + 2µ) ν
=− T11 = − T11 ,
2(λ + µ) λ+µ E

where ν is Poisson’s coefficient of lateral contraction


λ
ν= .
2(λ + µ)

This shows that in addition to an elongation (or possibly compression), the bar in Young’s
experiment also shows a contraction (or dilatation) in the perpendicular directions. Com-
puting
λ 3λ + 2µ µ(3λ + 2µ) 1 E
1+ν =1+ = = =
2(λ + µ) 2(λ + µ) λ + µ 2µ 2µ
or
1 1+ν
= ,
2µ E
one can express the inverted Hooke’s law in E and ν instead of in Lamé’s constants,
1+ν ν
E= T − (tr T)1,
E E
as used in many engineering textbooks. Further applications of Hooke’s law will be given in
the chapter on linear elasticity.

4.5 Viscoelastic media


The constitutive relations discussed sofar have one common characteristic: they are point
relations, because the stress at x and t depends only on the strain or the strain rate at
the same x and t. There are no memory effects, no dispersive effects. Because there are
different possibilities of including memory effects, we only give the simplest example, where
the relationship between stress and strain is still a linear one, but the stress depends upon
all previous deformations,
Z +∞
Tij (x, t) = Cijkℓ (x, τ )Eℓk (x, t − τ ) dτ
−∞
Table of contents 58

or Z +∞
T(t) = C(τ ) : E(t − τ ) dτ.
−∞

The space dependence still being a localized one, we have suppressed the argument x to
simplify the notation. The stress is now a functional of the strain, involving the elasticity
kernel C. Such media are called viscoelastic. The kernel has to be such that

C(τ ) = 0 if τ < 0,

otherwise the stress would depend upon future deformations, which would be against the
principle of causality. Thus we may as well put

Z ∞
T(t) = C(τ ) : E(t − τ ) dτ ,
0

but in the sequel we will continue to write the integral from −∞ to +∞. Furthermore, in
the case of such a linear response, a bounded strain is supposed to yield a bounded stress,
and this imposes the requirement that the components of the kernel satisfy a condition of
boundedness, Z +∞
|Cijkℓ (t)| dt < +∞.
−∞

The study of relations such as


Z +∞
T(t) = C(τ ) : E(t − τ ) dτ
−∞

usually goes via Fourier transforms, because these get rid of the convolution integrals.
In view of the many applications in other branches of physics of both linear response functions
and Fourier transforms, it is worth pursuing the subject a while here. Several ways exist to
define the Fourier transform. We will use
Z +∞
b
φ(ω) = Fω {φ(t)} = φ(t)eiωt dt,
−∞

together with its inverse transform


Z +∞
b 1 b
φ(t) = Ft−1 {φ(ω)} = φ(ω)e −iωt
dω.
2π −∞

A proper check that this is indeed the inverse requires the theory of distributions, but this
is not needed here. One of the most pleasant properties of the Fourier transform is that
 
∂φ
Fω = −iωFω {φ}.
∂t
Table of contents 59

Furthermore, when applied to a convolution such as the one relating T and E, one has
Z +∞
b
T(ω) = eiωt T(t)dt
−∞
Z +∞ Z +∞
iωt
= e dt dτ C(τ ) : E(t − τ )
−∞ −∞
Z +∞ Z +∞
= dt dτ C(τ ) : E(t − τ )eiωt
−∞ −∞
Z +∞ Z +∞
= du dv C(v) : E(u)eiω(u+v) .
−∞ −∞

The last step is possible because the Jacobian determinant of the transformation
 
t−τ = u τ =v
or
τ =v t = u+v

is equal to one. Hence the Fourier-transformed constitutive relation is


Z +∞ Z +∞
b
T(ω) = iωv
dv C(v)e : b
du E(u)eiωu = C(ω) b
: E(ω).
−∞ −∞

b
The information contained in C(t) is also contained in C(ω), although the principle of causal-
ity has yet to be translated into Fourier language. Because the importance of the resulting
relations transcends the mere case of a linear viscoelastic medium, these relations are dealt
with in the next section.

4.6 Kramers-Kronig’s relations


As said before, the constitutive relation for a linear viscoelastic medium is but an example
among many others of linear responses where causality plays a role. In order to simplify the
subsequent treatment, we deal with only one scalar component Cijkℓ (t) at the time, which
b
we denote C(t) or C(ω) for notational simplicity. Split C(ω)b into its real and imaginary
parts,
b
C(ω) =C b′ (ω) + iC
b′′ (ω).

Then immediately one has the following properties for real ω:

b′ (ω) = C
C b′ (−ω),
b
C(ω) b
= C(−ω) or b′′ (ω) = −C
C b′′ (−ω).

b
The real part of C(ω) is even, the imaginary part odd in ω. Also
Z +∞ iωt
Z +∞
b
C(ω) ≤ |C(t)| e dt = |C(t)|dt < +∞,
−∞ −∞
Table of contents 60

R
γ
r
−R +R
O ω0

Figure 4.2. Contour of integration in the complex ω-plane.

b
which indicates that C(ω) has no poles for real ω. In order to be able to use complex analysis,
b
we analytically continue the function C(ω) into the complex domain for ω and first put

ω = ω ′ + iω ′′ .

Then Z +∞
b
C(ω) =

C(t)eiω t e−ω t dt
′′

−∞

and we still have that


Z ∞ Z ∞
b −ω ′′ t
C(ω) ≤ e |C(t)| dt ≤ |C(t)|dt < +∞,
0 0

b
but only if ω ′′ is nonnegative. This indicates that C(ω) has no poles in the upper half-plane,
with even
b
C(ω) →0 as ω ′′ → +∞.
b
If we choose ω0 real, then the complex function C(ω)/(ω − ω0 ) has no poles in the upper
half-plane, but one on the real axis in ω0 . We thus want to apply Cauchy’s formula,
I b
C(ω)
dω = 0
ω − ω0
for a contour not enclosing a singularity. We pick a special contour, as shown in Figure 4.2
and call γ and Γ the small and large semi-circles, respectively, with radii r and R. Cauchy’s
formula is thus applied as
Z ω0 −r Z Z  b Z +R
C(ω)
+ + + dω = 0.
−R γ ω0 +r Γ ω − ω0
Table of contents 61

b
For the integral around γ we expand ω and C(ω) as

ω = ω0 + reiφ (0 ≤ φ ≤ π)

∂ b
C
b
C(ω) b 0 ) + reiφ
= C(ω + ...
∂ω
ω0

and find
Z b Z 0 b Z π
C(ω) C(ω0 ) + . . . iφ b b 0) + . . .
dω = iφ
ire dφ = −iC(ω0 ) dφ + . . . = −iπ C(ω
γ ω − ω0 π re 0

The dots indicate terms in r and they vanish when the limit r → 0 is taken afterwards. The
integral over the large semi-circle can be shown to vanish if the limit R → ∞ is taken. What
remains of Z ω0 −r Z Z +R Z  b
C(ω)
+ + + dω = 0
−R γ ω0 +r Γ ω − ω0
after both these limiting procedures is
Z ω0 −r Z +∞  b
C(ω) b 0 ) = 0,
lim + dω − iπ C(ω
r→0 −∞ ω0 +r ω − ω0
or Z +∞ b
b 0) = 1 P
C(ω
C(ω)
dω.
iπ −∞ ω − ω0
The expression
Z +∞ Z ω0 −ε Z +∞ 
φ(ω) φ(ω)
P dω = lim + dω,
−∞ ω − ω0 ε→0 −∞ ω0 +ε ω − ω0

if it exists, is called the Cauchy principal value. In the case under consideration, we are
b 0 ). After the limiting procedure, the integration is
assured of its existence as it equals iπ C(ω
along the real axis and we rewrite it as
Z +∞ b
b 1 C(ξ)
C(ω) = P dξ.
iπ −∞ ξ−ω

Splitting this expression into its real and imaginary parts yields

Z +∞ b′′ (ξ)
b′ 1 C
C (ω) = P dξ
π −∞ ξ−ω
Z +∞ b′ (ξ)
b′′ (ω) = − 1 P
C
C

π −∞ ξ−ω
Table of contents 62

These are called the Kramers-Kronig’s relations. The reasoning leading ultimately to them
b
was based on the fact that C(ω), as a complex function of a complex variable, had no poles
nor singularities in the upper half-plane. This in turn is a translation of the boundedness of
C(t), together with the principle of causality, through the fact that C(t) vanishes for negative
times. The Kramers-Kronig’s relations are thus nothing but a mathematical consequence of
the principle of causality.
The relations are often used to compute the real part of C(ω) b when the imaginary part is
known or vice versa. This is a boon to experimentalists, because the real and imaginary
parts of a linear response function often refer to different physical properties, and a measure
of one these properties immediately gives the other one.
In electromagnetic theory e.g. the linear response could be the permittivity tensor linking
the dielectric displacement to the electric field, and then the real part of the permittivity
tensor gives the dispersion in the medium, whereas the imaginary part gives the dissipa-
tion. Dispersion and dissipation are two, at first sight totally different characteristics of
the medium, and the Kramers-Kronig’s relations then deduce the one from the other! If
both characteristics can be measured, the Kramers-Kronig’s relations serve as a check on
the consistency of the data.
Chapter 5

Ideal fluids

Application of Navier-Stokes law to ideal, non-viscous fluids include discussions of

• Bernoulli’s law, applied to solar wind models

• Hydrostatic equilibrium inside stars

• Sound waves

• Waves in magnetized plasmas

• Large amplitude electrostatic and electromagnetic plasma waves

• Newtonian cosmology, leading to Friedman universes

• Complex potential theory

• Surface water waves

63
Table of contents 64

5.1 Basic equations


In this and the next chapter, fluids are defined as those linear isotropic media where the
stress-strain relation is the law of Navier-Stokes. Collecting the set of basic equations to
describe phenomena in these fluids, we need the equation of continuity,

ρ̇ + ρ∇ · v = ∂t ρ + ∇ · (ρ v) = 0

the equation of motion,


ρ v̇ = f + ∇ · T = ρ b + ∇ · T

the symmetry of the stress tensor,


Tt = T,
the constitutive relation of Navier-Stokes,

T = −p1 + λ(tr D)1 + 2µD

together with the definition of the rate of deformation,

1
D = (∇v + v∇)
2

and an equation of state relating the pressure to the density and possibly the temperature,

p = p(ρ)

It is possible to reduce this set of equations, first of all by expressing T as a function of the
velocity gradients,
T = −p1 + λ(∇ · v)1 + µ(∇v + v∇).
This serves to determine T once v and p are known. The next step is to compute
∇·T = ∇ · [−p1 + λ(tr D)1 + 2µD]
= ∇ · [−p1 + λ(∇ · v)1 + µ∇v + µv ∇]
= −∇p + λ∇∇ · v + µ∇2 v + µ∇∇ · v,
= −∇p + (λ + µ)∇∇ · v + µ∇2 v.
The set of equations left is thus
ρ̇ + ρ∇ · v = 0,

ρ v̇ = f − ∇p + (λ + µ)∇∇ · v + µ∇2 v

p = p(ρ).
Table of contents 65

The boxed middle equation in the previous set also carries the name of the Navier-Stokes
equation. By restricting the pressure changes to those involving only the density we steer
away from thermodynamical complications, as explained in chapter 2. The set of equations
is now a closed one, with as many unknowns as equations, and this will be used to discuss
a great variety of problems involving fluids, in this chapter as well as in the next. Fluids in
this sense are not only liquids, but also gases and plasmas, at least for certain applications.
Topics to be treated can be chosen from hydraulics, hydrodynamics, aerodynamics, magne-
tohydrodynamics or electromagnetic fluid theory, astrophysics even. We will try to pick one
or two examples in each of these branches of physics.
However, before embarking upon these applications of continuum mechanics, one has to note
that the set of fluid equations does not contain a reference to the stress any longer. We may
check that the mean stress is
   
1 2 2
σ = tr T = −p + λ + µ tr D = −p + λ + µ ∇ · v.
3 3 3
In absolute value the mean stress equals the pressure in two different cases, first of all when
the viscosity coefficients fulfil the so-called Stokes’ condition
2
λ = − µ,
3
or when the medium is incompressible

tr D = ∇ · v = 0.

As we now check, the latter condition follows from the expression of the incompressibility of
a given volume V of the continuum,
Z  Z
d
V̇ = dV = ∇ · v dV = 0 (∀V ) ⇐⇒ ∇ · v = 0.
dt V V

With the help of the equation of continuity we also find that for incompressible media

ρ̇ = 0.

The mass density is constant when the medium is followed in its motion. Often the mass
density is really constant throughout the medium, but then the condition of incompressibility
has to be coupled to the requirement of homogeneity.

5.2 Simplifying hypotheses


The full set of fluid equations is mathematically complicated, as it consists of a set of coupled
nonlinear partial differential equations. In many cases of physical importance the Navier-
Stokes equations can be simplified to some degree when one or more of the hypotheses listed
below are fulfilled or used.
Table of contents 66

Barotropic pressures and incompressibility

When the pressure in a fluid depends only upon the density, it is called barotropic, i.e. the
equation of state takes the form f (p, ρ) = 0. Applications of this are to be found in gases
when the kinetic equation of state does not include the temperature or when the processes run
isothermally or adiabatically. In isothermal processes in a monatomic gas, pV is constant,
and hence also p/ρ, whereas for adiabatic processes pV γ and thus pρ−γ remains constant,
with γ the ratio of the specific heats at respectively constant pressure and constant volume.
For barotropic pressure variations there exists a pressure function P (p), also called pressure
potential, such that
dP 1
∇P = ∇p = ∇p
dp ρ
and defined through Z p
dp′
P (p) = ′
.
p0 ρ(p )

To find a physical interpretation for the pressure function we compute the energy balance in
the absence of heat and viscosity effects (λ = µ = 0), as discussed in chapter 2,
1 p p
ε̇ = T : D = − 1 : D = − tr D
ρ ρ ρ
 
p p d 1
= − ∇ · v = 2 ρ̇ = −p ,
ρ ρ dt ρ
so that
d p 1 dp dP dp dP
(ε + ) = = = .
dt ρ ρ dt dp dt dt
Hence the pressure function is the specific enthalpy
Z p  p
dp′ p
P = ′
= ε+ ,
p0 ρ(p ) ρ p0

homogeneous barotropic
p

ρ0
ρ

Figure 5.1. Pressure-density relations.


Table of contents 67

up to an arbitrary additive constant.


The other practical possibility for the pressure is that the medium is incompressible and
homogeneous,
ρ(x, t) = ρ0
and then the pressure function is not needed, as
1 1 p
∇p = ∇p = ∇ .
ρ ρ0 ρ0
Both cases can be summarized in one graph, as shown in Figure 5.1. In what follows the
case of a barotropic pressure will tacitly cover also that of a homogeneous medium, where
the pressure function is a special limiting case with a vertical (p, ρ)-characteristic.

Ideal and viscous fluids

A fluid is called ideal when its viscosity coefficients are zero

λ = 0 = µ.

In all other cases we talk about viscuous fluids. The resulting form of the Navier-Stokes
equation for ideal fluids carries the name of Euler’s equation,

1
ρ v̇ = f − ∇p or v̇ = b − ∇p
ρ

Stationary and irrotational flows

The flow of a fluid is stationary (or time-independent, or permanent) if

∂t v = 0, ∂t ρ = 0.

Another type of flow is that of an irrotational flow, without vorticity,


1
ω = ∇ × v = 0.
2
This immediately implies the existence of a potential function for the velocity,

v = ∇φ.

As usual with potentials, φ is not unique, because any arbitrary function of time added to a
given φ will yield the same velocity.
Table of contents 68

Conservative forces

The main simplification concerning the body forces is to suppose that they can be derived
from a potential or, in other words, that the body forces are conservative. Two different ways
of implementing this are possible, either on the force density,
∇ × f = 0, f = −∇Ψ,
or on the force density per mass density,
∇ × b = 0, b = −∇Φ.
Remember that b has the dimension of an acceleration.

Nonlinear equations and linearization techniques

The equations normally are nonlinear. This can be simplified in certain cases by the assump-
tion that the phenomena under study are small perturbations with respect to some equilibrium
state. Dropping squares and products of these small perturbations yields equations that are
linear in the unknown quantities. This technique is called linearization and is often used,
when the full nonlinear equations are too difficult to handle, to obtain a first picture of cer-
tain phenomena. Afterwards then, the nonlinearities have to be brought in. For these there
are few general techniques, compared to the vast arsenal of mathematical methods that are
at our disposal for linear equations.
To get a feeling for the method, first write down the fluid equations for a constant homoge-
neous equilibrium state, in which all equilibrium quantities are labelled with the subscript
zero,
f0 = 0, p0 = p(ρ0 ).
This determines what equilibrium states are possible and shows that the body forces must
vanish in equilibrium. Not all problems admit such a homogeneous equilibrium, as the
example of the stratification of a fluid under the influence of gravity shows. The next step
in the linearization is to split every quantity in its equilibrium and its perturbation value,
ρ = ρ0 + ρ1 ,
v = v0 + v1 ,
p = p0 + p1 ,
f = f0 + f1 (or b = b0 + b1 ).
Finally the linearized form of the set of fluid equations appears as
(∂t + v0 · ∇)ρ1 + ρ0 ∇ · v1 = 0,
1 λ+µ µ
(∂t + v0 · ∇)v1 = b1 − ∇p1 + ∇∇ · v1 + ∇2 v1 ,
ρ0 ρ0 ρ0

dp
p1 = ρ1 .

ρ=ρ0
Table of contents 69

One is thus reduced to a linear problem for the determination of the perturbation quantities
ρ1 , v1 and p1 , given the known, homogeneous equilibrium parameters ρ0 , v0 and p0 . If
necessary, higher-order corrections could be considered by extending the perturbation scheme
to second or third order, in which case the problem is solved by successive steps.

5.3 Bernoulli’s law


For the remainder of this chapter we will deal with ideal fluids and use Euler’s equation
instead of Navier-Stokes’ equation. Hence we start from the basic set of equations

ρ̇ + ρ ∇ · v = 0,
1
v̇ = b − ∇p,
ρ
p = p(ρ).

It will be understood that in the case of an incompressible, homogeneous fluid the pressure
is to be treated as one of the dynamical variables determined together with v from

∇ · v = 0,
1
v̇ = b − ∇p.
ρ0
We now return to the general case of ideal fluids. One of the oldest consequences of Euler’s
equation is Bernoulli’s law, of which several variants are known. We will discuss some of
these in what follows, but first recast v̇ as
1 1
v̇ = ∂t v + v · ∇v = ∂t v + v · ∇v + ∇v 2 − ∇v 2
2 2
1
= ∂t v + v · (∇v − v∇) + ∇v 2
2
1 1
= ∂t v + 2v · Ω + ∇v 2 = ∂t v + 2ω × v + ∇v 2 .
2 2
In the case of irrotational flow this reduces to
 
1 2 1 2
v̇ = ∂t (∇φ) + ∇v = ∇ ∂t φ + v ,
2 2

in a kind of hybrid notation. For barotropic pressures we can use the pressure function,
1
∇p = ∇P,
ρ
and finally we suppose the body forces conservative in the sense that

b = −∇Φ.
Table of contents 70

Under the previous assumptions, Euler’s equation can be written as the gradient of a scalar
function,  
1 2
∇ ∂t φ + v + Φ + P = 0,
2
and upon integration one finds a first form of Bernoulli’s law

1
∂t φ + v 2 + Φ + P = 0
2

The trick to get rid of the integration constant with respect to x, which really means an
arbitrary function of t, is to see that the velocity potential was only determined up to an
arbitrary function of t and to let both arbitrary functions cancel each other.
Bernoulli’s law as given above is valid for the irrotational flow of an ideal fluid with barotropic
pressure variations under the influence of conservative forces. What happens when the
assumption of irrotational flow is dropped, but the other restrictions are kept? Then one
gets  
1 2
∂t v + 2ω × v = −∇ v +Φ+P .
2
A first integral can be obtained by integrating this along one of the streamlines. A streamline
of a flow is defined as a curve such that the velocity vector of the fluid is everywhere tangent
to it, as drawn in Figure 5.2. One first finds that
1 ds
d x · ∂t v = ds ev · ∂t v = ds v · ∂t v = ∂t v 2 = ds ∂t v
v 2v
and furthermore dx · (ω × v) vanishes, as dx is always parallel to v along a streamline.
Hence, after integrating the previous equation, the result is
Z
1
∂t v ds + v 2 + Φ + P = Cs (t),
s 2

which is the more general form of Bernoulli’s theorem, with a constant (with respect to x)
which will depend upon the particular choice of the streamline along which the integration
is carried out.

dx ev v

Figure 5.2. Streamlines and the velocity field.


Table of contents 71

Two special cases can be deduced: a first one for permanent (or stationary) flow, when
everything becomes time-independent,

1 2
v + Φ + P = Cs
2

and a second one for irrotational flow, as given already before. For permanent and irrota-
tional flow, the integration constant Cs becomes independent of the choice of a particular
streamline,
1 2
v + Φ + P = C.
2
If we apply Bernoulli’s theorem to the case of an incompressible and homogeneous fluid
influenced by gravity, we have
p
P = , Φ = gx3
ρ0
and thus Z
1 v2 p
∂t v ds + + + x3 = Cs′ (t).
g s 2g ρ0 g
As an application, chosen among many examples from hydraulics, consider the outflow of a
liquid from a large tank, where at the upper level pressure and height are somehow main-
tained. The outflow is supposed to be permanent and we can apply the corresponding form
of Bernoulli’s law in two different points along the streamline through the outflow orifice.
The situation is shown in Figure 5.3. Writing for simplicity z instead of x3 , one gets
p1 v2 p2 v2
z1 + + 1 = z2 + + 2.
ρ0 g 2g ρ0 g 2g
At the upper level (with label 2) there is no velocity, so that
2
v12 = 2g(z2 − z1 ) + (p2 − p1 )
ρ0

p2
z2

∆z

p1
z1

v1

Figure 5.3. Steady outflow of a tank.


Table of contents 72

or r
2
v1 = 2g∆z + ∆p.
ρ0
The velocity v1 of the outflow is caused partly by the difference in height, partly by the
difference in pressure. If there is no difference in pressure, as is the case when both the
upper level in the tank and the outflow orifice are in the open, at atmospheric pressure, the
velocity is just the free fall velocity from a height ∆z. Another limiting case is when there
is no outflow, because the pressure balances gravity,
p1 = p2 + ρ0 g∆z.
This is the usual hydrostatic law, where the pressure at the lower level is the pressure at the
upper level plus the weight per unit area of a liquid column with density ρ0 and height ∆z.

5.4 Solar wind


As a second example of Bernoulli’s law, chosen this time from solar system physics, we discuss
some aspects of the velocity of the solar wind near the Earth. The solar wind is a stream
of matter which is more or less continuously being emitted from the solar surface. The Sun,
as indeed any other main-sequence star, can be viewed as a sphere of hot, ionized gas (or
plasma) which is in equilibrium due to the opposing influences of gravitation and pressure or
temperature effects. More about this equilibrium is given in the next section. Gravitation
tends to keep matter closely together, pressure pushes it apart again. However, the outer
layer of the Sun, the corona, is, for reasons still not fully understood, much hotter than the
lower layers. Such a thermodynamically unstable situation gives some of the particles in the
corona the energy to escape from the solar gravity, hence the origin of the solar wind. In this
simplified picture the solar wind really is the expanding outer corona of the Sun, blowing
past the Earth.
Again we are assuming we are dealing with a permanent flow. In Bernoulli’s law the gravi-
tational potential of the Sun now enters,
GM⊙
Φ=− ,
r
where G is the universal constant of gravitation, M⊙ the mass of the Sun and r the radial
distance from the center of the Sun. In comparison, the gravitational influence of the Earth
can be neglected. We apply Bernoulli’s law twice along a streamline, once in a point close to
the solar surface but outside the Sun (marked with a subscipt 1), the other time in a point
in the neighbourhood of the Earth (subscript 2),
Z p1 Z p2
1 2 GM⊙ dp 1 2 GM⊙ dp
v1 − + = v2 − + .
2 r1 p0 ρ 2 r2 p0 ρ
Here p0 is some reference pressure, and the velocity of the solar wind in the neighbourhood
of the Earth is thus Z p1  
2 2 dp 1 1
v2 = v1 + 2 − 2GM⊙ − .
p2 ρ r1 r2
Table of contents 73

As r1 is of the order of the radius R⊙ of the Sun, whereas r2 is of the order of the radius of
the Earth’s orbit around the Sun, we can safely approximate the last term in the previous
equation by  
1 1 2GM⊙
2GM⊙ − ≃ ≡ vE2 .
r1 r2 R⊙
Here vE is the escape velocity at the surface of the Sun, defined (as for other astronomical
bodies) as the velocity needed by a particle starting at the surface to reach infinity with zero
velocity,
1 2 GM m
mvE − = 0.
2 R
Incidentally, if the escape velocity were for a particular star equal to the speed of light,
2GM
c2 = ,
R
then one would find the radius of a classic black star with mass M ,
2GM
R= .
c2
The concept of such a black star is due to Laplace, predating general relativity, and has
been revived in black hole theory (but based on general relativity rather than on classical
mechanics).
Coming back to the solar wind velocity,
Z p1
dp
v22 = v12 − vE2 +2 ,
p2 ρ

we try to apply the ideal gas law to the solar wind plasma,

p = CρT,

where T is the temperature and C a constant. At the surface of the Sun T1 is of the
order of the temperature of the corona, some million degrees Kelvin (K), whereas in the
neighbourhood of the Earth T2 still is 2 × 105 K. We can thus replace T by some averaged
value T0 and find Z p1 Z ρ1
dp dρ ρ(r1 )
= CT0 = CT0 ln .
p2 ρ ρ2 ρ ρ(r2 )
We finally suppose that the (quiet) solar wind escapes from the surface of the Sun with
negligible initial velocity, v1 ≃ 0, and thus near the Earth we obtain

ρ(R⊙ )
v22 ≃ 2CT0 ln − vE2 .
ρ(r2 )

This can serve as a first approximation to ρ(R⊙ ) because all other values can be measured
or computed.
Table of contents 74

Parker model of the solar wind

In order to get a better picture of how the solar wind evolves, we look at the outer corona of
a spherically symmetric star as the Sun. For time-stationary flows the equation of continuity
reduces to
∇ · (ρ v) = 0.
Here
x
x = r er and v = v er = v ,
r
since only radial motions are possible. Hence
 x  ρv 
∇ · (ρ v) = ∇ · ρv = ∂i xi
r r
ρv  ρv  ρv d  ρv 
= (∂i xi ) + xi ∂i = 3 + xi (∂i r) ,
r r r dr r
since all coordinate dependence is through r. As
xi
∂i r = ,
r
we find that
ρv d  ρv  ρv dv dρ
∇ · (ρ v) = 3 +r =2 +ρ +v = 0.
r dr r r dr dr
Dividing by ρv yields
1 dρ 1 dv 2
=− − .
ρ dr v dr r
For radial motions outside the star the Euler equation

ρv · ∇v = −∇p + f

reduces to
dv 1 dp GM
v =− − 2 .
dr ρ dr r
Using the ideal gas law gives
dv 2κT dρ GM
v =− − 2 .
dr mρ dr r
Remark that integration of this would give
1 2 2κT GM
v =− ln ρ + + C,
2 m r
nothing but Bernoulli’s law.
Eliminate (1/ρ) dρ/dr between the equations of continuity and of motion to get
 
dv 2κT 1 dv 2 GM
v = + − 2 ,
dr m v dr r r
Table of contents 75

which can be rearranged as

 
2 2κT 1 dv 4κT GM
v − = − 2
m v dr mr r

This is of the structure


dv
A(v) = B(r),
dr
with obvious notations,
v 2 − c2S
A(v) = ,
v
2c2 GM
B(r) = S − 2 .
r r
At the solar surface, however defined, we have
 
2c2 2GM 1 2 1 2
B(R⊙ ) = S − = c2S − vE,⊙ < 0.
R⊙ R⊙ 2R⊙ R⊙ 4
The condition
B(r) = 0
gives a critical distance
GM
rc = .
2c2S
For the Sun this is outside the Sun itself, since

rc,⊙ ≃ 3R⊙ ,

at coronal temperatures of 2 × 106 K. Hence B(r) goes from negative at the solar surface to
positive faraway, and can only change sign once.
Suppose now that at the solar surface

A(v) > 0 =⇒ v 2 > c2S .

This would mean a large outflow velocity, probably larger than the Sun can sustain for the
4.5 × 109 years it has existed already. For reasons of continuity, A(v) can only be zero at the
critical distance rc , otherwise we would get infinite gradients in v. So when we start at the
solar surface with A(v) > 0, we note that
dv
A(v) = B(r) < 0,
dr
and v decreases with increasing r. If at rc

dv
A(v)|rc > 0 =⇒ = 0,
dr rc
Table of contents 76

we still have that A(v) > 0 beyond rc , but now with the other sign for dv/dr,
dv
A(v)
= B(r) > 0.
dr
Now v increases again, forever. The model with
r = R⊙ : v 2 > c2S A(v) > 0
r = rc : v 2 > c2S A(v) > 0
means that v decreases to a minimum at rc but increases again. This gives values of v at
the radius of the Earth’s orbit which are too high, compared to the observations.
Conversely, having
r = R⊙ : v 2 < c2S A(v) < 0
r = rc : v 2 < c2S A(v) < 0
means that v increases to a maximum value (smaller than cS ) at rc and decreases thereafter.
This gives values at 1 AU which are far too small.
The only acceptable models are where A(v) and B(r) both vanish at the same distance rc .
To see what happens then, we Taylor expand everything close to the critical distance,

dv dv
v(r) ≃ v(rc ) + ∆r = cS + ∆r.
dr rc dr rc
Then we have that

c2S dv
A(v) = v − = 2 ∆r,
v dr rc
2c2S GM 2c2S 2GM GM
B(r) = − 2 = − 2 ∆r + 3
∆r = 3 ∆r.
r r rc rc rc
Hence from
dv
A(v) = B(r)
dr rc
we see that 2
dv GM
=
dr rc 2rc3
and dv/dr cannot change sign at rc where both A(v) and B(r) change sign. So the two
possibilities at R⊙ are that
v 2 > c2S =⇒ v decreases forever
v 2 < c2S =⇒ v increases forever
and at
r = rc ⇐⇒ v = cS .
Only the Parker model, where the solar winds starts at subsonic speeds at the solar sur-
face, goes through the sonic point at a critical distance and has supersonic velocity in the
neighbourhood of the Earth can give the right values, as observed here.
Table of contents 77

5.5 Hydrostatic equilibrium


In this paragraph we discuss a static problem, based on Euler’s equation without acceleration
term
1
b − ∇p = 0.
ρ
Two other special cases of Euler’s equation will be dealt with in the next sections. First,
1
when the external forces are small compared to the pressure effects, we have v̇ = − ∇p, and
ρ
this will be the starting point for the treatment of sound waves, in section 5.6, where the flow
is determined essentially by the pressure gradient. On the other hand, the external forces
dominate for some other applications, and Euler’s equation then becomes v̇ = b. Based on
this form we will give an introduction to surface water waves and plasma waves in section
5.7, and to Newtonian cosmology in section 5.11.
A first problem of hydrostatics is the pressure and density stratification with height in the
Earth’s atmosphere due to gravity. This is left as an exercise.
As another example, we now give some features of the hydrostatic equilibrium in stars, when
these stars are viewed as spheres of hot, ionized gas. The star is supposed to be in static
equilibrium and spherically symmetric, as shown in Figure 5.4. Because of the spherical
symmetry, all quantities will only depend on the radial coordinate r from the center of the
star. Hence Euler’s equation becomes in equilibrium
dp
= −bρ.
dr
The minus sign is because the pressure acts outwardly, whereas the gravitation acts inwardly,
or in other words, because the pressure will decrease outwards. However, the force of grav-
itation will also depend on r, due to the mass interior to r. To fix the ideas, we mentally

m′
P M
m
r

O
R

Figure 5.4. Hydrostatic equilibrium of a star.


Table of contents 78

isolate an interior sphere with radius r and mass m,


Z Z r
m = m(r) = ρ dV = 4πr′2 ρ(r′ ) dr′ .
|x|≤r 0

If we put a test particle with mass m′ on the surface of this sphere in P , the force of
gravitation can be computed as if all the mass of this inner sphere were concentrated in the
center, owing to the spherical symmetry. The outer shell, between the spheres with radii r
and R (the radius of the star), does not exert any force on an interior point P , again because
of the spherical symmetry. Hence the particle feels a total force
Gmm′
m′ b = ,
r2
and one finds that
Gm
b= ,
r2
leading to
dp Gmρ
=− 2 .
dr r
If the density distribution ρ(r) in the interior were known as a function of r, one could
explicitly compute m and p. This density distribution, however, is not a simple matter to
determine and depends critically on the model for the stellar interior. Nevertheless, without
going into such considerations, one can obtain some reasonable estimates of pressure and
temperature in the following way.
Upon differentiation of Z r
m = 4π r′2 ρ(r′ ) dr′
0
with respect to r one gets
dm
= 4πr2 ρ.
dr
So elimination of ρ with the help of the equation of motion yields
dp Gm dm
=−
dr 4πr4 dr
and upon integration over the whole star, from 0 to R (in radius) or from 0 to M (in mass),
we find
Z R Z R
Gm dm G dm
p(0) − p(R) = 4
dr > 4
m dr
0 4πr dr 4πR 0 dr
Z M
G GM 2
= m dm = .
4πR4 0 8πR4
In other words
GM 2
p(0) > p(R) + .
8πR4
Table of contents 79

At the surface, the pressure p(R) is negligible, and a lower bound for the pressure at the
center is
GM 2
p(0) >
8πR4

This estimate has been arrived at without proposing any equation of state for the pressure,
and indeed even without saying anything at all about the density distribution inside the
star. When the numbers for the Sun are substituted, one has
p(0)⊙ > 4.5 × 1013 Nm−2 = 4.5 × 108 atm.
For an estimate of the temperature, we define a mean temperature hT i by
Z Z Z
1
hT i = ρT dV ρ dV = ρT dV,
V V M V
with the mass density as weight factor. We postulate the ideal gas law in the form
p = CρT
and get
Z R Z R
1 2 4π
hT i = 4πr ρT dr = r2 p dr
M 0 M C 0
 R Z R Z R
4π 1 3 4π 3 dp 4π dp
= r p − r dr ≃ − r3 dr,
MC 3 0 3M C 0 dr 3M C 0 dr
because p(R) is approximately zero. Using once again
dp Gm dm
=−
dr 4πr4 dr
there comes
Z R Z R
4π Gm dm G dm
hT i = dr > m dr
3M C 0 4πr dr 3M CR 0 dr
Z M
G GM
= m dm = .
3M CR 0 6CR
Applied again to the Sun,
hT i⊙ > 1.9 × 106 K.
The above information about the minimum pressure and averaged temperature allows us
to conclude that the interior of the Sun (and hence of similar stars), must be composed of
hydrogen plasma, because at such temperatures hydrogen is fully ionized. The differences
between gas and plasma, viewed as a completely ionized and almost electrically neutral
gas, come into play when the electromagnetic properties are studied, much less when the
pressure-temperature-density relations are used. To a first approximation, the use of the
ideal gas law is thus justified.
Table of contents 80

5.6 Sound waves


In this and in the next sections we will give an introductory treatment of wave phenomena
in ideal fluids. On the right hand side of Euler’s equation there normally are two terms, one
for the body forces and the other one giving the pressure effects. When the pressure effects
dominate and the body forces are negligible, this yields sound waves, which are propagating
perturbations in density.
When there are no body forces, the set of basic equations for ideal, barotropic fluids in their
linearized form are

∂t ρ1 + ρ0 ∇ · v1 = 0,

1 1 dp
∂t v1 = − ∇p1 = − ∇ρ1 ,
ρ0 ρ0 dρ ρ0

dp
p1 = ρ1 .
dρ ρ0

We have put v0 equal to zero, as the medium is supposed to be at rest in equilibrium. The
only constants characterizing the equilibrium are the mass density ρ0 and the pressure p0 ,
connected through the equation of state. For fluids p is an increasing function of the mass
density, and hence we can put

dp
= c2S > 0
dρ ρ=ρ0

Now cS has the dimensions of a velocity and will turn out to be the sound speed. The last
equation thus is
p1 = c2S ρ1 ,
whereas the first two equations become

∂t ρ1 + ρ0 ∇ · v1 = 0,
c2
∂t v1 + S ∇ρ1 = 0.
ρ0
Elimination of v1 is possible after taking the time derivative of the first equation and the
divergence of the second, and subtracting both equations. The result is

∂t2 ρ1 = c2S ∇2 ρ1 ,

also denoted as  
1 2
2
∇ − 2 ∂t ρ1 ≡ 22 ρ1 = 0.
cS
Here
1 2
2 2 = ∇2 − ∂
c2S t
Table of contents 81

stands for the d’Alembertian of the problem. Immediately one also has

22 p1 = 0.

The solutions of d’Alembert’s equation are of the form

ρ1 = φ(k · x ± ωt),

with similar expressions for p1 . The function φ is arbitrary. In the phase argument k is
the wave vector (in the direction of wave propagation) and ω the (angular) frequency of the
wave. The phase velocity of the wave is given by
ω
= cS .
k
The sound waves propagate in the direction given by k with a constant velocity which
is independent of ω or k. Such waves are called nondispersive. For nondispersive waves
one could also use the equivalent form ψ(n · x ± cS t), where n is a unit vector in the
direction of propagation. However, for dispersive waves the phase velocity will be frequency
or wavenumber dependent and thus the more general form φ(k · x ± ωt) is to be preferred.
From
c2 c2
∂t v1 = − S ∇ρ1 = − S φ′ k
ρ0 ρ0
we learn that v1 is parallel to k, which is typical for longitudinal waves where the directions
of excitation and propagation coincide. Here φ′ refers to the derivative of φ with respect to
its argument k · x ± ωt.
The sound speed increases as the (p, ρ)-characteristic steepens, and one could formally in-
clude the case of an incompressible medium by saying that an incompressible medium has
an infinite sound speed. This is physically plausible: sound waves are nothing but the prop-
agation of disturbances in density which in incompressible media are felt instantaneously
everywhere.

5.7 Linear waves in magnetized plasmas


A plasma is a fully or almost fully ionized gas, that contains in equilibrium as many positive
(ions, protons) as negative (electrons, negative ions) charge carriers per unit volume and
thus is electrically neutral is. Otherwise large electric fields arise that preclude the existence
of a proper equilibrium.
Plasmas could be viewed as the fourth aggregation state of matter. If one thinks of the
classic aggregation states of water, namely solid (ice), liquid and gas (vapour) for rising
temperatures, then heating the gas further ultimately leads to dissociation of atoms into
nuclei and electrons, to a plasma. As an ionized gas a plasma reacts very strongly to the
presence of electromagnetic fields, which the usual (neutral) gas hardly feels.
The occurrence of plasmas in nature is very frequent. To cite just a couple of examples:
above the neutral atmosphere there are the ionosphere and the Earth’s magnetosphere, both
Table of contents 82

ionized media, or, farther away, stars like our own Sun have to be considered as plasma
spheres rather than spheres of ‘hot gas’. In short, the overwhelming part of the known
matter in the universe is in the plasma state. Grown rapidly after 1950, plasma physics
is the last frontier in the history of classical physics, where relativistic and/or quantum
effects hardly need to be taken into account. The only modern notions needed are that of
electrons and ions, but their nuclear properties are not involved. Astrophysical and energy
applications, like the aim of achieving a workable fusion reactor, have stimulated the great
blossoming of plasma physics during the last half century.
We will start the description of waves in plasmas from the use of the constitutive vacuum
relations,
D = ε0 E, B = µ0 H
and introduce the plasma effects via appropriate definitions for the charge and current den-
sities, σ and J, respectively. We combine the laws of Faraday and Ampère,

∇ × E + ∂t B = 0,
1
c2 ∇ × B = ∂ t E + J,
ε0
and eliminate in the classical way one of the fields, e.g. B. This leads to
1
c2 ∇∇ · E − c2 ∇2 E + ∂t2 E + ∂t J = 0.
ε0
Please note from Gauss’ law
ε0 ∇ · E = σ
that ∇ · E does not always vanish. We will use a complex plane wave representation of the
form
E = Ea ei(k·r−ωt)
and likewise for all relevant other variables. This yields an algebraic wave equation,

c2 k(k · E) + (ω 2 − c2 k 2 )E + J = 0,
ε0
because for harmonic plane wave we substitute

∇ → i k, ∂t → −iω.

Since for the time being we do not know anything about the wave character, and already
have one physical direction, that of wave propagation given by k or by n = k/k, we can
decompose all vector quantities into a parallel and a perpendicular part,

E = E k n + E⊥ ,

where of course

E⊥ · k = kE⊥ · n = 0,
Ek = E · n.
Table of contents 83

With an analogous decomposition for J the algebraic wave equation becomes



c2 k 2 Ek n + (ω 2 − c2 k 2 )(Ek n + E⊥ ) + (Jk n + J⊥ ) = 0.
ε0
Hence we find two separate equations for the parallel and the perpendicular parts,

ω 2 Ek + Jk = 0,
ε0

(ω 2 − c2 k 2 )E⊥ + J⊥ = 0.
ε0
Now is the time to introduce the appropriate form for σ and J in a plasma. We adapt the
usual definitions for moving charges and currents,
X
σ= σs ,
s
X
J= σs v s ,
s

and interpret these as summations over the different species making up the plasma. Here
σs is the charge density per fluid composing the plasma, and ρs its mass density. A plasma
consists minimally, in equilibrium, of equal numbers of electrons (s = e) and protons (s = i),
as in the case of a hydrogen plasma. As all charged particles, the plasma species undergo the
influence of electromagnetic fields via the force density of Lorentz. Hence we now introduce
an Euler equation per species,
σs
v̇s = (E + vs × B) .
ρs
Because the full problem is complicated and nonlinear, the equations will be linearized as
follows,

ρs = ρs0 + ρs1 ,
σs = σs0 + σs1 ,
B = B0 + B1 ,

the remaining quantities being zero in equilibrium. In first instance we suppose that in the
magnetic part of the Lorentz force density, vs × B, we can restrict ourselves to vs × B0 ,
where B0 is an externally applied static magnetic field. In this way the Fourier transformed
Euler equation reduce to
σs0
−iω vs1 = (E1 + vs1 × B0 ).
ρs0
In most books on plasma physics, however, instead of the mass and charge densities a
(species) number density ns is used, with the relations

ρs = m s ns , σ s = q s ns ,
Table of contents 84

where ms and qs are the mass and charge of the species under consideration. For a simple
hydrogen plasma we have that

qe = −e, qi = +e.

Hence we rewrite the Euler equation as


qs
−iω vs = (E + vs × B0 ),
ms
and omit from now on all subscripts 1 on the wave variables. Splitting this in parallel and
perpendicular parts yields
qs
− iω vsk = Ek + (vs × Ωs ) · n,
ms
qs
− iω vs⊥ = E⊥ + (vs × Ωs )⊥ .
ms
We have introduced the gyrofrequency vector Ωs ,
qs B0
Ωs = ,
ms
with the gyrofrequency itself being given by
qs B 0
Ωs = .
ms
Both these expressions include the sign of the charge under consideration.
The external magnetic field now introduces a second physical direction, and in order not to
complicate this introduction to plasma waves too much, we will restrict ourselves to wave
propagation parallel to the external magnetic field B0 . Because of k k B0 or n k Ωs , there
follows that

(vs × Ωs ) · n = 0,
(vs × Ωs )⊥ = vs⊥ × Ωs ,

and the equations for vsk and vs⊥ can be dealt with separately.
From the parallel part there immediately follows that
iqs
vsk = Ek ,
ms ω
and hence also that
i X Ns qs2
Jk = Ek .
ω s ms
Here Ns denotes the equilibrium value ns0 of ns . Thus we find that

ω 2 Ek + Jk = 0
ε0
Table of contents 85

is rewritten as ! !
X Ns q 2 X
s
ω2 − Ek = ω2 − 2
ωps Ek = 0.
s
ε0 m s s

Here the plasma frequencies ωps per species have been defined through

2 Ns qs2
ωps =
ε0 m s
and a global plasma frequency ωp is given through
X
ωp2 = 2
ωps .
s

For a simple electron-proton plasma, generated from the ionization of hydrogen, we note
that
Ne e 2 Ni e 2 N e2 N e2 N e2
ωp2 = ωpe
2 2
+ ωpi = + = + ≃ ,
ε0 m e ε0 m i ε0 m e ε0 m i ε0 m e
because of the charge neutrality in equilibrium

Ne = Ni = N.

The large mass difference me ≪ mi has as a consequence that the global plasma frequency
is principally determined by the electrons. We have thus obtained that

(ω 2 − ωp2 )Ek = 0,

from which only two choices follow.


Either we want Ek 6= 0 and then the allowable wave frequency is fixed as

ω = ±ωp

or ω takes on different values and then Ek vanishes, which leads to transverse waves. With
Ek 6= 0 the electric field contains already a component along the direction of wave propaga-
tion and the waves cannot be purely transverse. We will show that they then are longitudinal.
However, before we can complete the discussion we should first look at the remaining part
of the information, and determine vs⊥ from
qs
− iω vs⊥ + Ωs × vs⊥ = E⊥ .
ms
First a vector multiplication with Ωs gives
qs
− iω Ωs × vs⊥ + Ωs × (Ωs × vs⊥ ) = Ωs × E ⊥
ms
or also
qs
− iω Ωs × vs⊥ − Ω2s vs⊥ = Ω s × E⊥ ,
ms
Table of contents 86

because Ωs · vs⊥ = 0. Elimination of Ωs × vs⊥ between both equations tells us that


qs qs
(ω 2 − Ω2s )vs⊥ = iω E⊥ + Ωs × E ⊥ ,
ms ms
with solution
qs iω E⊥ + Ωs × E⊥
vs⊥ = .
ms ω 2 − Ω2s
In this way we have that
X X iω E⊥ + Ωs × E⊥
2
J⊥ = Ns qs vs⊥ = ε0 ωps ,
s s
ω 2 − Ω2s

and consequently there follows from



(ω 2 − c2 k 2 )E⊥ + J⊥ = 0
ε0
that !
X 2
ωps X ωps2
Ωs
ω 2 − c2 k 2 − ω 2 E⊥ + iω n × E⊥ = 0.
s
ω 2 − Ω2s s
ω − Ω2s
2

This is of the structure


A E⊥ + iC n × E⊥ = 0,
from which vector multiplication with n shows that

A n × E⊥ − iC E⊥ = 0.

Elimination of n × E⊥ between the two equations containing it yield

(A2 − C 2 )E⊥ = 0.

Now the condition that E⊥ 6= 0 gives us the other part of the dispersion law,

A±C =0

or explicitly
X ωps
2
ω(ω ∓ Ωs ) X ωps
2
ω
ω 2 − c2 k 2 − = ω 2 − c2 k 2 − = 0.
s
ω 2 − Ω2s s
ω ± Ω s

When ω obeys this part, then we can choose E⊥ 6= 0, but at the same time ω 2 6= ωp2 , and con-
sequently Ek = 0. The waves are transverse, and the extension of the classic electromagnetic
waves in vacuum.
In the opposite case, when ω 2 = ωp2 , one cannot have that
X ωps
2
ω
ω 2 = c2 k 2 +
s
ω ± Ωs
Table of contents 87

and hence E⊥ = 0. If we want to retain any electric field at all, we need Ek 6= 0, and the
waves are then purely longitudinal.
Special cases of the dispersion law for transverse modes,

X ωps
2
ω
2 2 2
ω =c k +
s
ω ± Ωs

are obtained when there are no plasma species at all. Then

Ns → 0 =⇒ ωps = 0

and the dispersion law reduces to the well known dispersion law for electromagnetic waves
in vacuum,
ω 2 = c2 k 2 .
On the other hand, if there is plasma but no external magnetic field, then all Ωs = 0 and
there remains
ω 2 = c2 k 2 + ωp2

A characteristic of these waves is that the plasma frequency acts as a cut-off frequency.
Whereas in vacuum all frequencies are possible, in a nonmagnetized plasma only those
electromagnetic waves with frequencies above the plasma frequency can propagate. This
is of importance for the propagation of radio waves through the ionosphere, the layer of
ionized gas or plasma above the atmosphere. In the ionosphere the density and hence the
corresponding plasma frequency varies with height. For certain layers, waves with too low a
frequency will be reflected, as they cannot propagate into this layer. This means two things:
if the source of the waves is on Earth, certain waves will be reflected back to the surface of
the Earth, making radio transmission beyond the horizon possible. If on the other hand the

ω ≥ ωp

ωp

ω < ωp
source receiver

Figure 5.5. Propagation of radio signals through the ionosphere.


Table of contents 88

source is outside the ionosphere some radio signals never reach the Earth. This is a handicap
both in radioastronomy and in communicating with spacecraft. This is drawn schematically
in Figure 5.5.
In the following two sections, we will give some further examples of plasma waves of larger
amplitude, where the harmonic Fourier transform and associated linearization procedures
are not used or cannot work.

5.8 Large amplitude ion-acoustic solitons


In this section we consider large amplitude electrostatic waves in a plasma, when the relevant
equations cannot be linearized without losing the typical characteristics of these waves. By
electrostatic one usually means that no magnetic effects come into play, so that Faraday’s
equation
∇ × E + ∂t B = 0
shows that
E = −∇φ.
From the other Maxwell’s equations we henceforth only use Poisson’s equation

∇ · D = σi + σe = e(ni − ne )

in the form
ε∇2 φ = e(ne − ni ).
For ion-acoustic waves the electrons are considered isothermal with

pe = κTe ne (Te = constant)

where κ is Boltzmann’s constant. In the electron equation of motion


1 e κTe e
v̇e = − ∇pe − E=− ∇ne + ∇φ
ρe me ne m e me
one now neglects electron inertia. This means that the wave phenomena to be studied are
sufficiently slow so as to allow the very mobile electrons to adjust almost instantaneously,
maintaining a balance between pressure and electric potential effects,
κTe
∇ne ≃ e ∇φ.
ne
Thus the electron density is  

ne = n0 exp ,
κTe
as a constant equilibrium requires that

ne0 = ni0 = n0 .
Table of contents 89

The ions, on the other hand, will be treated as cold, so that in their equation of motion the
electric potential effects determine the motion,
e
v̇i = (∂t + vi · ∇)vi = − ∇φ.
mi
Collecting now the remaining equations, we need the ion continuity equation, the ion equa-
tion of motion and Poisson’s equation,

∂t ni + ∇ · (ni vi ) = 0,
e
(∂t + vi · ∇)vi = − ∇φ,
mi
 
2 eφ
ε∇ φ = en0 exp − eni .
κTe

In what follows we will drop the subscript i on the ion quantities. We will study the nonlinear
behaviour of ion-acoustic waves and double layers moving in the z-direction and start from

∂t n + ∂z (nv) = 0,
e
∂t v + v∂z v = − ∂z φ,
m 
2 eφ
ε0 ∂z φ = en0 exp − en.
κTe

We now look for wave phenomena and localized structures travelling in the z-direction with
a velocity V , so that all quantities will depend upon

ξ = z − V t.

The above equations may then be integrated with one-sided boundary conditions

φ → 0, → 0, v → 0, n → n0 as ξ → +∞.

The continuity equation is thus rewritten as
dn d
−V + (nv) = 0,
dξ dξ
since
d d
∂t = −V , ∂z = .
dξ dξ
We find that
−V n + nv = −V n0
or
V n0
n= .
V −v
Table of contents 90

The ion velocity v will be determined from the equation of motion


dv dv e dφ
−V +v =−
dξ dξ m dξ
or after integration from
1 e
−V v + v 2 = − φ.
2 m
The only acceptable solution is
r r !
2eφ 2eφ
v=V − V2− =V 1− 1− ,
m mV 2

as in the expression for n (> 0) we see that V > v. One thus finds
,r
2eφ
n = n0 1− .
mV 2

Poisson’s equation can now wholly be rewritten as a differential equation for φ,


  ,r
2
dφ eφ 2eφ
ε0 2 = en0 exp − en0 1− .
dξ κTe mV 2

To integrate this, we multiply with dφ/dξ and find


 2     (r )
ε0 dφ eφ 2eφ
= n0 κTe exp − 1 + n0 mV 2 1− −1 ,
2 dξ κTe mV 2

which is of the form


 2
1 dφ
+ Ψ(φ, V ) = 0
2 dξ

with the Sagdeev pseudopotential Ψ(φ, V ) given by


   ( r )
n0 κTe eφ n0 mV 2 2eφ
Ψ(φ, V ) = 1 − exp + 1− 1− .
ε0 κTe ε0 mV 2

This represents the energy integral for a classical particle of unit mass moving with velocity
dφ/dξ in a potential Ψ(φ, V ), where V appears as a parameter. We see that the mass density
and velocity remain real provided
mV 2
φ≤ .
2e
The rarefactive range of solutions (with φ < 0) is unrestricted, but only a limited compressive
range (with φ > 0) is allowed.
Table of contents 91

For both ion-acoustic solitons and double layers one requires that Ψ(φ, V ) be negative be-
tween φ = 0 (because of the boundary conditions at infinity) and some extreme value φ = φm .
The appearance of localized solitary wave solutions furthermore also requires that
∂Ψ ∂ 2Ψ dφ
Ψ(0, V ) = (0, V ) = 0, 2
(0, V ) < 0, =0 at φ = 0,
∂φ ∂φ dξ
∂Ψ
Ψ(φm , V ) = 0, (φm , V ) > 0 if φm > 0,
∂φ
∂Ψ
Ψ(φm , V ) = 0, (φm , V ) < 0 if φm < 0.
∂φ
To adopt the language of classical mechanics, these requirements ensure that a particle
starting at φ = 0 makes a single transit to the point φm and comes back to its initial
position, so that |φm | represents the (maximum) amplitude of the soliton.
For ion-acoustic double layers one requires in addition that
∂Ψ ∂ 2Ψ
Ψ(φm , V ) = (φm , V ) = 0, (φm , V ) < 0,
∂φ ∂φ2
Ψ(φ, V ) < 0 for 0 < |φ| < |φm |.

This means, in the mechanics analogy, that the particle will transit from 0 to φm only (or
vice versa).
For small-amplitude solitons the expansion of the Sagdeev potential near φ = 0 yields
   
n0 κTe eφ e 2 φ2 e 3 φ3 n0 mV 2 eφ e 2 φ2 e 3 φ3
Ψ(φ, V ) = − − − + + +
ε0 κTe 2κ2 Te2 6κ3e Te3 ε0 mV 2 2m2 V 4 2m3 V 6
   
1 n0 e 2 n0 e 2 2 1 n0 e 3 n0 e 3
= − + φ + − + φ3 .
2 ε0 κTe ε0 mV 2 2 3ε0 κ2 Te2 ε0 m2 V 4
Using the ion plasma frequency ωpi and the ion-acoustic velocity cs , defined through
κTe
c2s = ,
m
we rewrite this expansion up to third order as
2   2  
ωpi 1 1 2
eωpi 1 1
Ψ(φ, V ) = − φ + − φ3 .
2 V 2 c2s 2m V 4 3c4s
Weak ion-acoustic solitons are defined by taking the small-amplitude limit in the sense that

V = cs (1 + µ), φ = O(µ) (µ ≪ 1)

and retaining terms up to third order in µ and/or φ. This yields


2  
ωpi 2 eφ
Ψ(φ, V ) = 2 φ −µ + ,
cs 3mc2s
Table of contents 92

and the energy integral becomes


 2 2  
1 dφ ωpi 2 eφ
= 2 φ µ− .
2 dξ cs 3κTe
Substituting for simplicity dimensionless variables

ψ= ,
κTe
ωpi ωpi ωpi
η= ξ= (z − cs (1 + µ)t) = z − ωpi (1 + µ)t,
cs cs cs
we have that  2  
1 dψ 2 1
=ψ µ− ψ .
2 dη 3
We can deduce from this, after taking the derivative twice (undoing so to say the integrations
performed in deriving the Sagdeev equation) that

dψ dψ 1 d3 ψ
−µ +ψ + = 0.
dη dη 2 dη 3
This can also be obtained from a Korteweg-de Vries (KdV) equation by using a reductive
perturbation analysis ab initio. The KdV equation itself looks like

∂ψ ∂ψ 1 ∂ 3 ψ
+ψ + =0
∂τ ∂ζ 2 ∂ζ 3

so that for
η = ζ − µτ
we recover the ordinary differential equation derived before.
The KdV equation has been derived in many different physical situations and describes e.g.
waves on shallow water, as well as the ion-acoustic plasma waves mentioned here.
For the solution we return to
 2  
1 dψ 2 1
=ψ µ− ψ .
2 dη 3
A soliton solution for this problem is

ψ = 3µsech2 αη = ,
cosh2 αη
where α is yet to be determined. Since
dψ sinh αη
= −6αµ ,
dη cosh3 αη
Table of contents 93

(α = 1, µ = 38 )

Figure 5.6. Profile of a solitary wave joining constant states at η = ±∞ and localized in η.

we find from  
2
2 2 sinhαη 9µ2 µ
18α µ = µ−
cosh6 αη cosh4 αη cosh2 αη
that r
µ
α= .
2
The solution thus is
r
2 µ
ψ = 3µsech η
2

This looks indeed as a localized structure, as sketched in Figure 5.6. Since µ was originally
defined as
V − cs
µ= ,
cs
we find localized waves where the amplitude 3µ is related to the normalized wave speed,
a typical property of nonlinear waves. Moreover, µ has to be positive, making the waves
slightly supersonic with respect to cs .

5.9 Large amplitude circularly polarized waves in plasmas


Here we revert to the description of transverse electromagnetic waves in magnetized plasmas,
but without linearizing the basic equations and hence obtaining large amplitude solutions.
We recall the basic equations per species, consisting of the equation of continuity for the
number densities
∂t ns + ∇ · (ns vs ) = 0
and the equation of motion
qs
∂t vs + vs · ∇vs = (E + vs × B).
ms
Table of contents 94

The electric and magnetic fields E and B couple the plasma species together in Maxwell’s
equations

∇ × E + ∂t B = 0,
1 X
c2 ∇ × B = ∂ t E + ns q s v s .
ε0 s

The other Maxwell’s equations will not be used further on, except
X
ε0 ∇ · E = ns q s
s

for the equilibrium conditions.


We start from
ns = Ns ,
vs = Vs ez + us ,
E = E ,
B = B 0 ez + b ,
where Ns , Vs and B0 are constant equilibrium values and us , E and b refer to the fluctuations
associated with the circularly polarized waves. We emphasize, however, that these fluctua-
tions are not assumed to be small, so ours is not a perturbation analysis. The equilibrium
carries neither charge nor current, hence
X X
Ns qs = 0, Ns qs Vs = 0.
s s

Such a situation is only of interest if there are at least two ion species with different drifts,
for in a classic one-ion-species plasma Vi (subscript i for ions) must equal Ve (subscript e for
electrons) and then results from stationary plasmas can be transposed to drifting reference
frames. With more ion species with different drifts, there is no longer any natural drifting
frame of reference. There is also no equilibrium electric field, as this would only give a
common E0 × B0 drift for all plasma components. Such a common perpendicular drift can
be eliminated by going to a deHoffmann-Teller frame where E0 vanishes.
For circularly polarized waves propagating parallel to the external magnetic field there are no
fluctuations in density. This will entail that

usz = Ez = bz = 0.

To start with, we assume a particular form for the wave magnetic field, corresponding to
circularly polarized modes,
b = aex cos φ ± aey sin φ,
where the phase of the waves is given by

φ = kz − ωt.
Table of contents 95

Here the ± signs refer to the left (with the upper sign) or right (with the lower sign) circular
polarization. For this particular dependence on z and t, we see that
db
= −a ex sin φ ± a ey cos φ = ±ez × b.

Faraday’s law gives
dE db
∇ × E = ez × ∂ z E = k ez × = −∂t b = ω ,
dφ dφ
so that
ω aω aω
E= b × ez = ± ex sin φ − ey cos φ.
k k k
Turning next to the fluid equations, one sees that the equations of continuity are fulfilled
identically, whereas the equations of motion become
qs qs
(∂t + Vs ∂z ) us + Ωs ez × us = (E + Vs ez × b) + us × b.
ms ms
Because us , E and b are all orthogonal to ez , we deduce from the preceding equation that

(us × b) · ez = 0.

If us × b 6= 0, it would be parallel to ez , which clearly cannot be. Hence

us = γs b = γs a(ex cos φ ± ey sin φ),

with γs a scalar, and exact finite-amplitude solutions will be possible because the only re-
maining nonlinear term us × b vanishes. If that were not so, higher harmonics would be
generated which cannot be cancelled against other terms. We introduce for each species the
Doppler-shifted frequency
ωs′ = ω − kVs ,
and find that
db
(∂t + Vs ∂z ) us = γs (∂t b + Vs ∂z b) = γs (−ω + kVs ) = ∓γs ωs′ ez × b.

The equations of motion are rewritten as
qs  ω  qs ωs′
∓γs ωs′ ez × b + Ωs γs ez × b = b × ez + V s ez × b = − ez × b,
ms k ms k
which yields the factor of proportionality
qs ωs′
γs = ± ,
ms k(ωs′ ∓ Ωs )
and thus also the species velocity
qs ωs′
us = ± (aex cos φ ± aey sin φ) .
ms k(ωs′ ∓ Ωs )
Table of contents 96

We are now in a position to finally turn to Ampère’s law. In view of the assumed dependencies
of all quantities on z and t, the expansions and the equilibrium current neutrality condition,
this equation can be rewritten as
1 X
c2 ez × ∂ z b = ∂ t E + Ns qs us .
ε0 s

Collecting the forms for b, E and us eventually yields


!
X ωps 2 ′
ωs
ω 2 − c2 k 2 − ′
(aex cos φ ± aey sin φ) = 0.
s
ωs ∓ Ω s

We see that the dispersion law amounts to

X ωps
2
(ω − kVs )
ω 2 − c2 k 2 − =0
s
ω − kV s ∓ Ωs

nothing but the usual (linear) dispersion law for circularly polarized waves propagating
parallel to the external magnetic field, but now in the presence of parallel equilibrium drifts.
A finite-amplitude wave thus is possible indeed. Adiabatic or isothermal pressures can
easily be included in the foregoing treatment, as constant densities imply constant pressures.
Because the dispersion law was derived without referring to a particular frequency regime,
the low-frequency approximation for Alfvén waves is automatically included, as discussed
below.
A low-frequency regime in magnetized plasmas usually implies that |ω| ≪ |Ωs |. Because
of the interest in plasmas with nonzero equilibrium drifts, a long-wavelength approximation
where |kVs | ≪ |Ωs | is also needed, so that |ω − kVs | = |ωs′ | ≪ |Ωs |. The expansion of

ω − kVs ω′
= ′ s
ω − kVs ∓ Ωs ωs ∓ Ωs

up to second order in ωs′ gives

ωs′ ωs′ ωs′2


= ∓ − .
ωs′ ∓ Ωs Ωs Ω2s
We compute separately that
X ωps
2 ′
ωs X Ns q 2 (ω − kVs )ms X Ns qs (ω − kVs )
s
= · = = 0,
s
Ωs s
ε 0 m s q s B 0 s
ε 0 B 0

because of the absence of global equilibrium charge and current densities,


X X
Ns qs = 0, Ns qs Vs = 0.
s s
Table of contents 97

The second-order term gives


X ωps
2 ′2
ωs X Ns q 2 (ω − kVs )2 m2 c 2 µ0 X
s s
= · 2B2
= 2
Ns ms (ω − kVs )2 .
s
Ω2s s
ε m
0 s q s 0 B 0 s

We define the global Alfvén velocity VA through


B2
VA2 = X0 ,
µ0 Ns m s
s

the true bulk speed of the plasma as a whole by


,
X X
V = Ns m s V s Ns m s
s s

and the mean Doppler-shifted frequency as ωe = ω − kV . Deviations of the mean velocity



are then denoted by Vs = Vs − V and W is a measure for the relative mass-averaged parallel
kinetic energy of all species,
,
X 2
X
W = Ns ms Vs′ Ns m s .
s s

The dispersion law is then up to second order approximated by


c2 2
ω 2 = c2 k 2 − ω + k 2 W ).
(e
VA2
Because VA is usually very small with respect to c, this is without much loss of generality
approximated by
ωe 2 = k 2 (VA2 − W ).
One sees immediately the possibility of unstable Alfvén waves whenever

VA2 < W

in other words, whenever there is enough relative streaming between the different species
of the plasma, measured with respect to the Alfvén velocity. A more extensive form of the
above dispersion law and the associated instability criterion has been used to study Alfvén
wave instabilities in solar wind plasmas.
Recent missions to comets P/Giacobini-Zinner, P/Halley and P/Grigg-Skjellerup have given
new results in solar wind plasmas sufficiently contaminated by ions of cometary origin. Due
to the low gravity of the cometary nucleus, such ions escape as neutral atoms or molecules
from the inner coma. Different processes can then ionize these particles, which are picked-up
at ionization by the solar wind magnetic field, with velocities which are very different from
the usual solar wind velocity. There is an interesting region between 1 to 4 ∼ 7 × 106 km
upstream of the cometary nucleus where the instability criterion would seem to hold, leading
to Alfvén wave turbulence.
Table of contents 98

O x′
a

O′

Figure 5.7. Two observers O and O′ and a distant galaxy P .

5.10 Newtonian cosmology


As a last application of Euler’s equation in ideal fluids we discuss some elements of Newto-
nian cosmology. It is clear that a proper treatment of cosmology at present requires general
relativity. However, Newtonian cosmology leads to the same results for the scale factor as
general relativity, for Friedman models, and thus a fairly simple picture of interesting phe-
nomena can be given, without going through all the mathematical intricacies connected with
general relativity.
The universe on the largest scale is viewed as a continuous fluid. This is then coupled to the
cosmological principle, which says that there are no privileged observers and that the universe
is homogeneous and isotropic on the larger scales, meaning that the universe presents the
same picture at all places and in all directions at a given instant. Both these ideas, a
continuous fluid filling the universe and the cosmological principle, are not peculiarities of
Newtonian cosmology but are used also in general relativity, where they lead to the Friedman
universes.
We start with two observers O and O′ , both recording what happens at P , as indicated in
Figure 5.7. The three different points could e.g. refer to three different galaxies. In view of
the homogeneity of the universe, ρ and p can at most depend upon t, but are independent of
position. If v is the velocity of P with respect to O and v′ with respect to O′ , the principle
of equivalent observers requires that there exist a single function v(x, t) such that
v = v(x, t), v′ = v(x′ , t).
Otherwise, one could class the observers according to the functional dependence included in
v(x, t). Also, O′ has a velocity v(a, t) with respect to O. The orientation of both reference
frames in O and O′ remains the same, there is no rotation, and hence the velocities are
additive in a Galilean sense,
v(x, t) = v(a, t) + v(x′ , t).
However, as
x = a + x′ ,
Table of contents 99

one finds that


v(a + x′ , t) = v(a, t) + v(x′ , t)
for all a, x′ and t. This is only possible when v(x, t) is a linear function of its first argument,
without constant term
v(x, t) = M(t) · x.
Here M(t) is a time-dependent tensor, which can always be decomposed into three parts

M = µ 1 + M′ + M′′ ,

such that
t t
M′ = M′ , M′′ = −M′′ , tr M′ = 0.
This yields
v = µ x + M′ · x + x × m′′ ,
where m′′ is the dual vector of M′′ .
Sofar, only part of the cosmological principle has been used. The material description is given
by the equation of continuity, together with Euler’s equation. In the equation of continuity

ρ̇ + ρ ∇ · v = 0

the density ρ can only depend upon t. This part of the reasoning rests on the identification
of the velocities of the observers (or galaxies) with those of the continuum in which they are
imbedded, much as a floating cork is carried along by the water. We compute

∇ · v = ∇ · (M · x) = M : x ∇ = M : 1 = tr M = 3µ.

It is customary to put
Ṙ(t)
µ(t) = ,
R(t)
where R(t) is called the scale factor. Thus the equation of continuity becomes


ρ̇ + 3ρ = 0,
R
yielding after integration that
ρR3 = ρ0 R03 .
This expresses the conservation of matter with density ρ inside a sphere of radius R. At
time t0
ρ(t0 ) = ρ0 , R(t0 ) = R0 .
Euler’s equation reduces to
v̇ = b,
as there are no pressure gradients, p being independent of x. The normalized force density
b is only due to the gravitational interaction and obeys Poisson’s law,

∇ · b = −4πGρ.
Table of contents 100

Hence for isotropic universes we have


4
b = − πGρ x.
3
When the isotropic part of the cosmological principle is invoked here, we have to be consistent
and note that only the isotropic part of M, namely µ 1, can then be nonzero. Hence

v = µ x,

which expresses Hubble’s law that the velocity of distant galaxies increases with their distance
from us or from any other point. Identifying again v and ẋ, Hubble’s law can be rewritten
as

v = ẋ = x.
R
The solution of this differential equation is

x = R(t) c,

with
x0 = x (t0 ) = R(t0 ) c = R0 c.
Combining this with Euler’s equation and the particular form for b gives
4
v̇ = − πGρ x
3
or in other words,
R̈ 4
ẍ = R̈ c = x = − πGρ x.
R 3
The scale factor R thus obeys the equation

R̈ 4 4 1
= − πGρ = − πGρ0 R03 3
R 3 3 R
with the help of conservation of mass, or more precisely

 
4 1
R̈ + πGρ0 R03 =0
3 R2

This equation is the same as the one which is derived for pressureless Friedman models
in general relativity. Hence the discussion we now give transcends the purely Newtonian
framework.
In principle this cosmological equation determines R(t). Afterwards we have

ρ0 R03
ρ(t) =
R3 (t)
Table of contents 101

and
x(t) = R(t) c,
and the problem is solved. Unfortunately, the determination of R(t) is not straightforward.
A first and perhaps startling conclusion is that when R is constant, the mass density has to
vanish. In other words, a static universe must be empty! There is no nonempty stationary
solution, unless one modifies the laws of gravitation ad hoc, as Einstein at first tried to
do, since he believed the universe had to be stationary. After multiplication by 2Ṙ, the
cosmological equation can be integrated once to get
 
2 8 3 1
Ṙ − πGρ0 R0 = C.
3 R
This is of the form of the conservation of energy for the onedimensional motion of a particle
in a conservative force field, and the discussion can be given in analogous terms. We know
from Hubble’s observations that the universe is currently expanding, with Ṙ > 0. Two cases
are possible: either C is positive or zero, or C is negative.
In the case where C is positive or zero, Ṙ2 is strictly positive and Ṙ cannot vanish sometime.
This means that Ṙ is always positive, as it is so now, and thus the scale factor R is an ever
increasing function of time. This is the model of the ever expanding or open universe.
On the other hand, if C were negative, we could put
 
8 3 1
C≡− πGρ0 R0 ,
3 Rc
which defines Rc for a given C < 0. Substituting this into
 
2 8 3 1
Ṙ = πGρ0 R0 +C
3 R
gives us   
2 8 1 1
Ṙ = πGρ0 R03 − .
3 R Rc
We conclude that
1 1
Ṙ2 ≥ 0 =⇒ ≥ ⇐⇒ R ≤ Rc .
R Rc
It is seen that Rc corresponds to a maximum value for the scale factor. As R̈ is always
negative, R increases to the value Rc , where Ṙ vanishes, and then R starts to decrease
again. It will then be decreasing monotonically until zero, where it all started from. A scale
factor zero corresponds to a singularity, as the density becomes infinite. This singularity is
called the “big bang”, with philosophical implications outside the scope of the present lectures
with their simple approach. The model where R is bounded by Rc is called an oscillating
or closed universe, although it is not clear at all whether such a cycle would repeat itself or
not. As
ρR3 = ρ0 R03 = ρc Rc3 ,
Table of contents 102

there corresponds a critical value ρc of the mass density to Rc , such that

R ≤ Rc ⇐⇒ ρ ≥ ρc .

Thus the mass density, if known accurately, could be an indication of whether the universe
was open or closed, expanding forever or “oscillating”. If there is enough mass, hence a
density above the critical density, the gravitational effects will be able to call the expansion
to a halt. If the density is too low, on the other hand, the expansion will be slowed but
never stopped.
The special case where C is zero corresponds to the Einstein-de Sitter model, with
8 1
Ṙ2 = πGρ0 R03 ,
3 R
and the integration can be carried out explicitly,

R = (6πGρ0 )1/3 R0 t2/3 .

The corresponding mass density


1
ρ=
6πGt2
is getting smaller and smaller in a larger and larger universe. At the end of this section, it
is an interesting but unanswerable question why Newtonian cosmology was not investigated
earlier, before rather than after the advent of general relativity. Maybe because there were
no static solutions and people believed in the permanence of our universe?

5.11 Complex potential theory


As a last application for ideal fluids we introduce complex potential theory, which is a
compact way of describing plane irrotational and stationary flows of an ideal homogeneous
fluid, as we will show. Stationarity means that the flow is time-independent, and homogeneity
reduces the equation of continuity to the incompressibility condition,

∇ · v = 0,

implying the existence of vector potential a for the velocity such that

v = ∇ × a.

Irrotational flow has no vorticity,


1
ω = ∇ × v = 0,
2
so that there also exists a scalar potential φ for the velocity field,

v = ∇φ.
Table of contents 103

Both representations of v, written together,

∇φ = ∇ × a,

give that
∇2 φ = 0, ∇2 a = 0,
if we assume that
∇ · a = 0,
a condition that can be imposed without loss of generality. Hence φ and a are harmonic
functions obeying a Laplace equation.
For plane flows (in the x, y-plane) we will put for simplicity of notation

x = x1 , y = x2 and u = v1 , v = v2 .

Furthermore, there is no z = x3 dependence of any kind. Putting then ψ = a3 , we find from

v =∇×a

that
∂ψ ∂ψ
u= , v=− ,
∂y ∂x
and from
v = ∇φ
that
∂φ ∂φ
u= , v= .
∂x ∂y
The meaning of ψ becomes clear when we look at the equation for the streamlines,
dx dy dx dy
= or = .
u v ∂ψ ∂ψ

∂y ∂x
The result is
∂ψ ∂ψ
dψ = dx + dy = 0,
∂x ∂y
and
ψ(x, y) = C
gives the streamlines. Hence ψ is called the stream function. Now the conditions
∂φ ∂ψ
(u =) = ,
∂x ∂y
∂φ ∂ψ
(v =) =− ,
∂y ∂x
Table of contents 104

are precisely the Cauchy-Riemann conditions for a complex function f of a complex argument
z, defined through

f (z) := {C → C : z = x + iy 7→ φ(x, y) + iψ(x, y)},

to be analytic in z. Moreover, there exists a derivative


df ∆φ + i∆ψ
w= = lim ,
dz ∆x,∆y→0 ∆x + i∆y
and the limiting procedure is independent of the way in which it is computed. Indeed,
df ∆φ + i∆ψ ∂φ ∂ψ
= lim = +i
dz ∆x → 0 ∆x ∂x ∂x
∆y = 0
or
df ∆φ + i∆ψ ∂ψ ∂φ
= lim = −i
dz ∆x = 0 i∆y ∂y ∂y
∆y → 0
are equivalent to a complex velocity

df
w= = u − iv
dz

As φ and ψ are harmonic functions, so are u and v. Putting

w = |w| e−iα

means that √
|w| = u2 + v 2
gives the absolute value of the real velocity in the xy-plane and α the angle between this
velocity and the x-axis.
Another way of recovering the Cauchy-Riemann conditions is by differentiating f to get

df = dφ + idψ = w dz = (u − iv)(dx + idy)


= udx + iudy − ivdx + vdy
∂φ ∂φ ∂ψ ∂ψ
= dx + dy + i dx + i dy,
∂x ∂y ∂x ∂y
from which there follows that
∂φ ∂ψ
u= = ,
∂x ∂y
∂φ ∂ψ
v= =− .
∂y ∂x
Table of contents 105

Now w has to be analytic and uniform in a domain D, hence holomorphic. If D is simply


connected, then f is also analytic and uniform, hence holomorphic, and
Z z
f (z) = f (z0 ) + w dz
z0

is independent of the path followed from z0 to z.


However, when D is multiply connected, then in
Z z
f (z) = f (z0 ) + w dz
z0

the last integral will depend on the path, more precisely on the number of times m the path
encircles one of the patches which do not belong to D. Hence
Z z
w dz = m(Γ + i D)
z0

with m an integer and Γ + iD the residue obtained by circling once around the cut-out,
anticlockwise,
I I
Γ + iD = w dz = (u − iv)(dx + i dy)
C C
I I
= (u dx + v dy) + i (u dy − v dx)
IC I C

= v · dx + i (v × dx) · ez .
C C

The former integral is the circulation of v around C, the latter integral represents the debit
across the surface contained in C. Going around this same C (with interior point a) we find
that Log (z − a) increases by 2πi. Hence
Z z
Γ + iD
Log (z − a) and w dz
2πi z0

have the same nonuniform behaviour along path encircling e.g. a, and thus

Γ + iD
f (z) − Log (z − a)
2πi

has to be holomorphic to yield an acceptable choice for a complex potential with holomorphic
derivative
df
w= .
dz
The procedure can be summarized as follows:
Table of contents 106

• We pick a suitable f (z) and find from

f = φ + iψ

that

φ=C =⇒ equipotentials
ψ=C =⇒ streamlines.

• The complex velocity


df
w= = u − iv
dz
gives

|w| = u2 + v 2 : absolute value of flow velocity
α = − arg w : angle between real velocity and x-axis.

Kelvin’s circle theorem

Most cases of potential flow will be in multiply-connected domains, and are concerned in
particular with flows around a cylinder. Luckily, there is a simple procedure to modify any
complex potential f describing an unbounded fluid so that the resulting potential has the
same characteristics but also describes a flow around a cylinder. This result goes by the
name of Kelvin’s circle theorem.
For any complex potential f (z), we define the complex conjugate function f¯(z) by

f¯(z) = f (z̄).

Note that f¯(z), now consideredP as a function of z, is again holomorphic. For instance, when
f has Laurent series f (z) = +∞ n=−∞ na z n
where the coefficients an are complex constants,
¯
the Laurent series of f is given by
+∞
X
f¯(z) = ān z n .
n=−∞

Kelvin’s circle theorem now says that if f (z) is a complex potential, and R is any real
number, then the function  2
¯ R
fR (z) = f (z) + f ,
z
is a complex potential describing a flow around a circle of radius R centered at the origin.
To prove this, we note first of all that fR (z) is holomorphic whenever f (z) is so. Secondly,
in order for fR (z) to describe a fluid in the exterior of the circle mentioned previously, it has
Table of contents 107

to be the case that the circle of radius R is a streamline for fR (z). This is easily verified: on
the circle of radius R, we have that z̄ = R2 /z, and so
 2
¯ R
f = f¯(z̄) = f (z) for z such that |z|2 = R2 .
z

Note that for all z, we have


   !  
R2 R2 R2
f¯ = f¯ =f .
z z̄ z̄

The stream function ψR of fR (z) is nothing but the imaginary part of fR (z). From the
previous expression, it follows that
 2
R
ψR (z) = ImfR (z) = Imf (z) + Imf¯
z
 2
R
= Imf (z) − Imf

On the circle of radius R, we have that |z|2 = R2 or alternatively that z = R2 /z̄, so that
ψR vanishes there. In other words, the boundary of the circle is a streamline and fR (z)
represents a fluid flowing around the circle.
Kelvin’s procedure is most often used in conjunction with the Riemann mapping theorem,
which states that any simply-connected region can be mapped onto the unit disc by a confor-
mal map. Since conformal mappings take holomorphic functions into holomorphic function,
we can hence study potential flow around any body by mapping it conformally onto the unit
disc and applying the Kelvin circle theorem.
In what follows, some elementary examples of potential flow are discussed. Wherever possi-
ble, we use the Kelvin circle theorem.

1. Uniform flow

Put
f (z) = az + b,
with a and b complex constants. Then
df
w= = a = |a|ei arg a
dz
means that
|w| = |a| and α = − arg a.
This describes a flow with constant velocity in a direction making a constant angle α with
the x-axis.
Table of contents 108

2. Source or sink

Choose now
D D D
f (z) = Log z = log r + i θ,
2π 2π 2π
where
z = r eiθ
is the polar representation. Hence
D
φ= log r = C,

in other words, the equipotentials are concentric circles, whereas
D
ψ= θ

gives the streamlines, straight lines leading into or away from the origin. Also
df D D −iθ
w= = = e = |w|e−iα
dz 2πz 2πr
gives
D
|w| = and α=θ (D > 0),
2πr
or
|D|
|w| = and α=θ−π (D < 0).
2πr
|D| is the debit of the source or sink.

Figure 5.8. Flow of a source or sink in the origin.


Table of contents 109

3. Point vortex

For this we take


Γ Γ Γ
f (z) = Log z = −i log r + θ
2πi 2π 2π
and the equipotentials follow from
Γ
φ= θ

as straight lines through the origin. The streamlines are given by
Γ
ψ=− log r,

and are concentric circles. This configuration is drawn in Figure 5.9. The information about
the flow velocity comes from
df Γ Γ −iθ |Γ| −i(θ± π )
w= = −i = −i e = e 2
dz 2πz 2πr 2πr
as
|Γ| π
|w| = and α=θ± .
2πr 2
The concentric streamlines are traversed counterclockwise around the origin, if Γ > 0, or
clockwise if Γ < 0.

Figure 5.9. Flow around a point vortex in the origin.

4. Flow around a circle or cylinder without vortices

In the first example, we have shown that f (z) = Az, where A is a real constant, represents
a steady flow along the x-axis with velocity A. We now consider the case of uniform flow
Table of contents 110

around a cylinder of radius R placed at the origin. We have that f¯(z) = Az, so that by
Kelvin’s circle theorem,  
R2
fR (z) = A z + .
z
From now on, we denote fR simply by f . In polar coordinates, we have
 
iθ R2 −iθ
f (z) = A re + e .
r
This potential has equipotentials
 
R2
φ=A r+ cos θ
r
and streamlines  
R2
ψ=A r− sin θ.
r
The special streamlines corresponding to ψ = 0 are the circle (or the cross section of the
cylinder) with radius R and from θ = 0 also the x-axis outside ]−R, +R[. The circle inside
R is not accessible to the flow. The flow velocity itself is given in complex form as
   
df R2 R2 −2iθ
w= =A 1− 2 =A 1− 2e .
dz z r
Stagnation points or lines are where w vanishes, here

w=0 ⇐⇒ z = ±R.

Hence one streamline comes from −∞ and ends in z = −R, the other starts at z = +R and
goes to +∞.

Figure 5.10. Flow around a circle or cylinder.


Table of contents 111

5. Flows around corners and edges

To model these we start from the complex potential

f = Az n = Arn eniθ (A real)

to find
φ = Arn cos nθ
for the equipotential curves and
ψ = Arn sin nθ
for the streamlines. The special streamlines corresponding to ψ = 0 are

r=0 : origin
π
sin nθ = 0 : θ=k (k integer).
n
Furthermore the complex velocity is
df
w= = nAz n−1 = nArn−1 ei(n−1)θ
dz
and its magnitude
|w| = nArn−1 .
Stagnation points correspond to w = 0 and have to be discussed for different possibilities.
First we take n < 1, so that
1
z n−1 =
z 1−n
and the stagnation points lie at infinity. This corresponds to flows along an edge, as shown
in Figure 5.11, for which π/n > π. The special streamlines with
π
θ=k
n
are θ = 0 and θ = π/n.

Figure 5.11. Flow along an edge.


Table of contents 112

Figure 5.12. Flow in a corner.

On the other hand, if n > 1, then the flow is in a corner, as shown in Figure 5.12 with
π/n < π. Now
w = 0 =⇒ z = 0 (origin)
indicates that the stagnation point lies in the origin.
To conclude this section, it is clear that different expressions for f (z) can be combined to
describe more intricate flow patterns, e.g.
 
R2 Γ
f (z) = A z + + Log z
z 2πi

corresponds to the flow around cylinder or circle with vortex motion.

Limitations of potential flow

In order to understand the solutions of the Euler equations, we have made various simplifying
assumptions. As a result, we obtained particular solutions with interesting physical and
mathematical properties, which model every-day observations very well. In some cases,
however, we obtain results which are clearly nonsensical, indicating that one or more of
our fundamental assumptions are strictly speaking erroneous. One such inconsistency is the
paradox of d’Alembert, which states that the drag on a circular body in an inviscid flow
vanishes. This is clearly in contradiction with, for instance, the observed behavior of a small
object submerged in a steadily flowing stream.

Blasius Formula. We will prove d’Alembert’s paradox in the context of two-dimensional


potential flow, but similar results can be derived in three dimensions as well. The result
follows from Blasius’ theorem concerning the forces acting on a disc in a potential flow. This
disc can have an arbitary shape — Blasius’ theorem is not limited to circular shapes.1
1
The proof of Blasius’ theorem below holds for arbitrary body shapes. Nevertheless, it is a good exercise
to work out the details of the proof for a circular body of unit radius.
Table of contents 113

Let us consider such a disc, which is submerged in a potential flow with stream function
f (z). We denote the force exerted by the fluid on the body by Fbody = (Fx , Fy ) and we
introduce the complex force Fbody as

Fbody = Fx − iFy .

Blasius’ theorem then asserts that


I  2
iρ0 df
Fbody = dz
2 dz
where ρ0 is the constant density of the fluid, and the circular integral is computed along the
boundary of the rigid body.
To prove this expression, we begin by choosing a parametric representation s →
7 z(s) for
the boundary of the rigid body, where for the sake of convenience s denotes arc length.
The (outward) normal vector to the boundary at a point with parameter s is denoted by
n(s) = (nx (s), ny (s)), and we introduce the complex normal

n(s) = nx (s) − iny (s).

It is easy to verify that n(s) = iz̄ ′ (s).


The force at a point z(s) is now given by

F (s) = −p(s)n(s),

where the minus sign is to be explained by the fact that n(s) is pointing outward of the
boundary, while F (s) by convention points inward. The total force exerted by the fluid on
the body is then given by the circular integral
I
Ftot = − p(s)n(s)ds.

We already have a complex expression for n(s); all that remains is to express p(s) in terms
of the complex potential f (z). This can be done by means of Bernouilli’s law: there are no
external forces and the flow is permanent and irrotational, so that
ρ0 2
p=− v +C
2
2
ρ0 df
= − +C
2 dz
where C is a constant.
We know that the boundary of the disc is a streamline, so that Ψ(z(s)) is constant for all s.
df ′
This allows us to conclude that dz z is a real function:

df df (z(s)) d
(z(s))z ′ (s) = = Φ(z(s)).
dz ds ds
Table of contents 114

We now have all the necessary elements to prove Blasius’ formula. We begin by substituting
the complex expression for the pressure and the normal vector into the integral formula for
the force to obtain
I
Ftot = p(s)n(s)ds
I 2
iρ0 df ′
= dz z̄ ds.
2

The constant C in the pressure function doesn’t matter, since the circular integral of a
constant function is zero. The integral can be further rewritten to give
I  
iρ0 df df ′
Ftot = z ds
2 dz dz
I  2
iρ0 df
= z ′ ds
2 dz

and this finally yields the formula of Blasius:


I  2
iρ0 df
Ftot = dz,
2 dz

where the integral (according to Cauchy’s theorem) can be computed along any contour
encircling the origin.

Paradox of d’Alembert. We now return to our original goal: to show that the drag force
on a circular body in a uniform steady flow is zero. Assume that the body radius is R and
that the fluid flows along the x-axis with velocity U . The complex potential is as before
 
R2
f (z) = U z + ,
z

and Blasius’ formula gives for the total force


I  2
iρ0 U 2 R2
Ftot = 1− 2 dz = 0,
2 z

since the integrand has no simple poles.


The result of d’Alembert is nowadays more of a historical curiosity. As with many paradoxes,
it merely signals that the current description is incomplete, and that a novel element is
needed. That element proved to be the development by Prandtl of boundary-layer theory,
where the body is enveloped by a thin layer in which viscous effects are dominant. Outside
the boundary layer, the fluid then behaves as if inviscid.
Table of contents 115

Surface water waves


In an attempt to understand some recent events, such as the tsunami disaster, it is important
to recall some elements about the propagation of waves or solitary structures on the surface of
water. We will start with the theory of linear water waves, because the nonlinear treatments
are quite involved and dealt with in other lectures.
To give some feeling about the subject of nonlinear water waves, we start in 1834, with
the observation by John Scott Russell of a solitary water wave on the then narrow canal
Edinburgh-Glasgow. His detailed findings (written down in 1844) are worth reading even
today, and provide an insight in a keen and unbiased mind:

I believe I shall best introduce the phaenomenon by describing the circumstances


of my own acquaintance with it. I was observing the motion of a boat which was
rapidly drawn along a narrow channel by a pair of horses, when the boat suddenly
stopped — not so the mass of water which it had put in motion; it accumulated
around the prow of the vessel in a state of violent agitation, then suddenly leaving
it behind, rolled forward with great velocity, assuming the form of a large solitary
elevation, a rounded, smooth and well-defined heap of water, which continued
its course along the channel without change of form or diminution of speed. I
followed it on horseback, and overtook it still rolling on at a rate of some eight
or nine miles an hour, preserving its original figure some thirty feet long and a
foot to a foot and a half in height. Its height gradually diminished, and after a
chase of one or two miles I lost it in the windings of the channel.

For the propagation velocity vf Scott Russell empirically put forward the relation

vf2 = g(h + a),

where h is the depth of the water in the canal or in the laboratory vessel, and a the maximum
amplitude of the waves, measured from the undisturbed surface.
It is worth stressing that this ‘great wave of translation’ was observed to be an elevation,
a hump of water above the level surface, not the familiar succession of ups and downs
that we tend to associate with periodic water waves in a linearized description. Maybe
because of that then standard picture, and Scott Russell’s detailed observation and many
subsequent laboratory investigations notwithstanding, the scientific establishment of the day
would have nothing to do with solitary waves. Nevertheless, great scientists like Boussinesq
(1871) and Lord Rayleigh (1876) confirmed the relation put forward by Scott Russell, and
have found the now famous sech2 profiles for these solitary waves. These typical profiles
were the solution of the differential equation derived in the doctoral thesis (1895) of De
Vries, earned under Korteweg as supervisor. Both their names are now attached to the KdV
equation. As indicated already, the nonlinear derivation is involved and better dealt with in
other lectures.
To continue with the linear theory, water is to a high degree of accuracy incompressible and
homogeneous (ρ ≃ constant), so that there remains from the continuity equation only the
Table of contents 116

incompressibility condition
∇ · v = 0.
The equation of motion is now
∂ 1
v + v · ∇v = g − ∇p.
∂t ρ
In equilibrium the pressure gradient has to balance gravity, so that

∇p0 = ρg.

Taking the z axis of a reference frame along the upward vertical, this becomes
dp0 (z)
= − gρ,
dz
yielding the classical stratification of a fluid or gas under the influence of gravity. After
linearization we see that the perturbations obey

∇ · δv = 0,
∂ 1
δv = − ∇δp.
∂t ρ
Because of the vertical stratification, we can take a plane wave solution in x and t, but not
in z, and put

δv = δv(z) exp[i(kx − ωt)],


δp = δp(z) exp[i(kx − ωt)].

We have assumed that the problem is translationally invariant in the third, y direction, and
get in this way a set of three scalar equations,
d
ikδvx + δvz = 0,
dz
1
ωδvx = kδp,
ρ
1 d
iωδvz = δp.
ρ dz
Elimination of δvx and δvz yields a single equation in the pressure perturbations,
d2
δp = k 2 δp.
dz 2
It is seen, through the elimination of δp, that the components of δv are also solutions of this
same differential equation, and hence of the form

δp = Ae−kz + Bekz ,
δvz = Ce−kz + Dekz .
Table of contents 117

The determination of the boundary conditions goes in two steps.


(1) At the free surface we have in general that

p = patm + ρgξ,

where ξ represents a reference height. For small perturbations in the water or at the surface
this becomes
δp = ρgδξ
or after derivation
∂ ∂
δp = gρ δξ.
∂t ∂t
We assume that for this type of problem the atmospheric pressure is constant. But the
vertical movement of the free surface is at the same time that of the water itself, which leads
to the identification on the surface of

δξ = δvz .
∂t
Inserting our plane solution yields

− iωδp = gρδvz ,

and upon substitution of δp by δvx there comes


iω 2
− δvx = gδvz .
k
Together with the incompressibility condition this finally gives
d gk 2
δvz = 2 δvz ,
dz ω
which is only valid, of course, when evaluated at z = 0, the reference height of the undis-
turbed surface of the water.
(2) At the bottom, located at z = −h, there is no vertical water motion, so

δvz = 0.

Both boundary conditions together give for the solutions that


gk 2
−kC + kD = (C + D),
ω2
Ce+kh + De−kh = 0.

This algebraic and homogeneous set of equations in C and D only has a nontrivial solution
provided
gk gk
2 +1 − 1
ω ω 2 = 0,

ekh e−kh
Table of contents 118

or also when
gk −kh  
2
e − ekh + e−kh + ekh = 0.
ω
The dispersion law is thus
ω 2 = gk tanh kh

Special limiting cases

(1) Deep water


This is expressed by the requirement that

kh ≫ 1,

or, rewritten as
2π h
kh = h = 2π ≫ 1,
λ λ
showing that
h ≫ λ.
The wavelength of the perturbation is small compared to the depth of the channel. For large
values of the argument we can approximate tanh kh by 1 and thus get

ω 2 = gk.

The phase velocity is r r


ω g gλ
vf = = =
k k 2π
and indicates that these waves are dispersive.
(2) Shallow water
Here the opposite limit is taken,
kh ≪ 1,
and the approximation
tanh kh ≃ kh
is used to arrive at the shallow water wave dispersion law

ω 2 = gk 2 h.

The phase velocity now is


ω p
vf = = gh,
k
and the waves are nondispersive. Comparison with the expression given by Scott Russell for
the velocity of the observed solitary wave,
p
vf = g(h + a),
Table of contents 119

clearly demonstrates that the linearization procedure for small perturbations necessitates
that a ≪ h. In other words, we deal with surface waves in shallow water (think of the length
of the soliton as some 30 feet or ≃ 9 m), and the nonlinear considerations give a limitation
of the validity of the linear results.
This is a general conclusion. For linear systems the amplitudes can only be determined up to
a normalization factor, as the set of equations for C and D show. How large the amplitudes
can be, before we exceed the domain of validity of the linearization procedure with its neglect
of squares and products, is something that can only be ascertained by a nonlinear study of
the problem!
Whether tsunamis in the open ocean are periodic waves on shallow water or more a train
of (fairly) weak, well separated solitons is still an open question. However, their velocities
(up to 800 km/h) and separation between the successive solitons or crests (periods of 30
to 45 min) indicate that the shallow water wave picture is essentially correct. This can be
determined from the observational constraints and will be left as an exercise.
Chapter 6

Viscous fluids

For viscous fluids we discuss

• Similarity and dimensional analysis

• Notions of turbulence

• Matched asymptotic expansions to treat boundary layers

120
Table of contents 121

6.1 Similarity
This chapter will be devoted to some phenomena where the viscosity of the fluid plays a
crucial role. The basic equations are the equation of continuity

∂t ρ + ∇ · (ρ v) = 0,

Navier-Stokes’ equation of motion, including viscosity,

ρ v̇ = f − ∇ p + (λ + µ)∇∇ · v + µ∇2 v

and a pressure equation of state


p = p(ρ).
One of the simplest examples which can be treated analytically, is the steady laminar flow of
a viscous liquid through a channel or a pipe. Laminar flow is well-ordered or well-structured
flow, as opposed to turbulent or erratic flow, of which more in subsequent sections. It is for
laminar phenomena that the Navier-Stokes equation is most adequate. Applications of this
can be found in many engineering textbooks and will not be given here.
Let us return to the equation of motion for viscous fluids
µ
ρ(∂t v + v · ∇v) = ρ g − ∇ p + ∇∇ · v + µ∇2 v,
3
where for simplicity and to fix the ideas, the forces are just gravity and Stokes’ condition
has been incorporated. This is to be considered together with the appropriate boundary
conditions, also for the continuity equation and an equation of state for the pressure.
Not all flow problems are as easy as laminar Couette and Poiseuille flow, and in many cases
no valid mathematical approach is possible and one has to rely upon computer simulations
or experimental evidence. In the latter case an often advocated way of dealing with as yet
unknown flows is to construct a scale model before embarking upon the construction of the
real thing. The main question is whether the conclusions drawn from the scale model are
valid for a larger experiment with different parameters, or formulated otherwise, when the
flow in the model is similar to the real flow. As will be shown below, this is a far from trivial
matter.
Two flows will be called similar if there is geometric and dynamic similarity. For dynamic
similarity there must be a constant proportion between the variables, such as density and
velocity, in the model and in reality, in each corresponding point. This seems an infinity of
conditions, but for phenomena describable by a set of differential equations it can be reduced
to the equality of some typical dimensionless numbers. Suppose, to fix the ideas and give
some hypothetical numbers, that the fluid in the model varies in density between 1 and 2
kg/m3 , whereas in reality the density goes from 1000 tot 2000 kg/m3 . We feel that in both
cases the same range is covered, relatively speaking. If we put

ρ = ρ0 ρ∗ ,
Table of contents 122

we can take ρ0 as a characteristic value for the mass density in a given situation, with
the dimensions included, while the dimensionless variable ρ∗ is then typical for the span
covered. In the given theoretical example ρ0 would be 1 kg/m3 in the model and 1000 kg/m3
in reality, and in both cases ρ∗ would vary from 1 to 2. The same will be done with all
variables occurring in the given situation, and we use a typical

• length L, such as the diameter or the length of a pipe, the distance between two walls,
the thickness of a body;
• time T , such as the time constant of a transition, the period of an oscillatory movement;
• velocity V , such as the uniform flow velocity, the sound speed when gases are involved,
the speed of light for electromagnetic waves;
• pressure p0 , taken in some reference point;
• density ρ0 , characteristic for the fluid under consideration.

All variables are then rewritten as a product of such a characteristic value times a dimen-
sionless variable, which will carry a star
x = L x∗ , t = T t∗ , v = V v∗ , p = p0 p∗ , ρ = ρ0 ρ∗ .
For simplicity a single characteristic length L and velocity V were chosen for the three space
dimensions. This is certainly not a strict requirement and one could very well envisage
situations where e.g. the length in one direction were measured in a different way from
lengths in other directions, with something similar for the velocities. The derivatives thus
become
1 1
∇ = ∇∗ , ∂t = ∂t∗ .
L T
Applying this to the equations of continuity and of motion gives
ρ0 ∗ ∗ ρ0 V ∗
∂ ρ + ∇ · (ρ∗ v∗ ) = 0,
T t L  
V ∗ ∗ V2 ∗ ∗ ∗ (g) p 0 ∇∗ p ∗ µV 1 1 ∗ ∗ ∗ ∗2 ∗
∂ v + v ·∇ v = ge − + ∇ ∇ ·v +∇ v ,
T t L ρ0 L ρ∗ ρ0 L 2 ρ∗ 3
or in dimensionless form
L ∗ ∗
∂t ρ + ∇∗ · (ρ∗ v∗ ) = 0,
TV  
L ∗ ∗ ∗ ∗ ∗ Lg (g) p 0 ∇∗ p ∗ µ 1 1 ∗ ∗ ∗ ∗2 ∗
∂ v +v ·∇ v = 2 e − + ∇ ∇ ·v +∇ v .
TV t V ρ0 V 2 ρ∗ ρ0 LV ρ∗ 3
In these equations several dimensionless combinations of the characteristic values appear.
These characteristic numbers are all named and each of them is significant for some partic-
ular effect or for a specific type of phenomenon. The first is called Strouhal’s number,
L
St = .
TV
Table of contents 123

It only occurs when the flow is not permanent and in this sense it is a measure for the
non-stationarity of the flow. Next we have Froude’s number,

V2 ρ0 V 2
Fr = = ,
gL ρ0 gL
which gives the relation of the kinetic to the potential energy densities, here the potential
energy density in the gravitational field. The following number is
p0
Eu = ,
ρ0 V 2
Euler’s number. This time the kinetic energy density is compared to the pressure. For gases
where the pressure variations are adiabatic, p ∼ ργ , one deduces that
dp p
=γ ,
dρ ρ
or with the characteristic values,
p0
c2S = γ .
ρ0
Hence Euler’s number can be transformed into
c2S 1 1
Eu = 2
= ,
γV γ (Ma)2

where
V
Ma =
cS

is the Mach number. It relates the flow velocity to the ambient sound speed. For Mach
numbers smaller than one, the flow is subsonic, when the Mach number exceeds one the flow
is supersonic. The flow at Mach number one is called transonic. These measures are used in
particular to express the speed of an aircraft. Finally there is the famous Reynolds number,

ρ0 LV ρ0 V 2
Re = = µV
µ L

which measures the ratio of the kinetic energy density to the viscous dissipation.
It has been observed experimentally to be a good measure for the degree of turbulence in
a flow. At low Reynolds numbers the flow is laminar or well structured. A low Reynolds
number implies a small typical length, such as the diameter of a narrow tube, together with
a small typical velocity, such as in quiet flow, and a high viscosity, indicating a large degree
of internal coherence. All these elements together give a steady, regular flow. On the other
hand, when the Reynolds number is high, the typical length is large, coupled with a high flow
Table of contents 124

speed in a fluid where the viscosity is low. We then have turbulent or irregular flow, with
a lot of vortices or eddies. The transition of laminar to turbulent flow occurs at a certain
critical Reynolds number. An introduction to the mathematical description of turbulence is
given in section 6.3.
Having discussed all the characteristic numbers, we rewrite the dimensionless equations of
continuity and of motion as

St ∂t∗ ρ∗ + ∇∗ · (ρ∗ v∗ ) = 0,
 
∗ ∗ ∗ ∗ ∗ 1 (g) Eu ∗ ∗ 1 1 1 ∗ ∗ ∗ ∗2 ∗
St ∂t v + v · ∇ v = e − ∗ ∇ p + ∇ ∇ ·v +∇ v .
Fr ρ Re ρ∗ 3

When this set of equations is applied to two different situations, the equations and hence also
the phenomena described will be the same when the dimensionless combinations are equal.
Then the scale model and the real situation will be similar. But the stumbling block lies in
the fact that not all typical numbers can be made equal in the same comparison between
model and reality!
Suppose that we put an airplane wing model in a wind-tunnel to measure its aerodynamic
shape. The equations for steady flow in the absence of external forces become

∇∗ · (ρ∗ v∗ ) = 0,
 
∗ ∗ ∗ Eu ∗ ∗ 1 1 1 ∗ ∗ ∗ ∗2 ∗
v ·∇ v =− ∗ ∇ p + ∇ ∇ ·v +∇ v .
ρ Re ρ∗ 3

We label with 1 all quantities for the model, with 2 all quantities for the real setup. We set
out with the same Reynolds number,

Re1 = Re2

or    
ρ0 ρ0
L1 V 1 = L2 V 2 .
µ 1 µ 2

If both the model and the real wing are put in an ordinary air flow, ρ0 /µ is the same and
we find that
L2
V1 = V2 .
L1
The velocity in the model is the real velocity divided by the scale factor L1 /L2 . However,
with these values one finds that
V1 V 2 L2 L2
Ma1 = = = Ma2 ,
cS c S L1 L1
if the sound velocity is the same. Even in cases where ρ0 /µ and cS are not the same in
the model and in the real experiment, there is usually not enough range in the kinematic
viscosities µ/ρ0 or sound speeds cS available to compensate for the scaling in length. Hence
when the Reynolds numbers are taken alike, the Mach numbers are nowhere equal! The only
Table of contents 125

way out would be to use for this experiment a gas with a totally different density, which is
generally not possible.
The unavoidable conclusion is that it is almost never possible to get similar flow in a model
experiment and in reality, if one defines similarity by requiring all corresponding dimen-
sionless numbers to be equal. The way past this difficulty is by deciding beforehand what
characteristic number is relevant for a given situation and disregarding other numbers. For
the airplane wing the viscosity of air is totally unimportant compared to the velocity effects,
hence the equality of the Mach numbers, not of the Reynolds numbers, is adhered to.

6.2 Dimensional analysis


In the previous section several characteristic and dimensionless numbers were arrived at,
starting from the governing equations such as the Navier-Stokes equation. Through dimen-
sional analysis one can also get some of these numbers, even in the case when governing
equations are lacking or not generally accepted, as in the case of turbulence, which is dis-
cussed in the next two sections.
The general idea behind dimensional analysis is to list the different parameters such as den-
sity, velocity, viscosity which are thought important for a given problem, and then to try and
form dimensionless numbers, of which there will be fewer, obviously, which can be used as
new parameters in establishing some kind of functional relationship.
In order to fix the ideas, we will start with a simple example from classical mechanics, the
velocity of a body free-falling from a height h in the gravity field of the Earth. We know the
result as p
v = 2gh,
because it was derived from Newton’s equations. Suppose we did not know Newton’s equa-
tions, but had some feeling, probably from experiments, that the velocity v depended only
on h and g. Using the dimensions L for length and T for time, we have the following
dimensionalities
L L
[v] = , [h] = L, [g] = 2 ,
T T

and hence a dimensionless combination of v, h and g is only possible as v/ gh, possibly
raised to some power. The relation

φ(v, g, h) = 0

can thus be rewritten as  


v
ψ √ = 0.
gh
This is now an equation for the dimensionless combination, with solutions
v p
√ =C =⇒ v = C gh.
gh
Table of contents 126

The numerical value of the constant C has to be found from observational evidence. Dimen-
sional analysis is unable to give this.
The more general formulation of the above ideas is embodied in the π-theorem of Bucking-
ham, stating that a functional relation

φ(α1 , . . . , αk , αk+1 , . . . , αn ) = 0

between n different parameters αj can be equivalently replaced by another functional relation

ψ(π1 , . . . , πm ) = 0

between m dimensionless parameters πj , where

m=n−k

and k is the largest number of parameters in the original list which will not combine to a
dimensionless number. For mechanical problems one usually has that

k ≤ r,

if r is the minimum number of independent dimensions required to give the dimensions of


all parameters occurring. In the example of the falling body, one sees that v and h (or v and
g, or h and g) are dimensionally independent, but v, h and g together not. Hence:

n = 3, k = 2, r = 2, m = 1,

as only length and time play a role.


To give an outline of the proof of the π-theorem, we select, through a judicious numbering of
the original parameters, k quantities α1 , . . . , αk , with independent dimensionalities, as basic
quantities,
[α1 ] = δ1 , . . . , [αk ] = δk .
The remaining parameters then all have dimensionalities of the form

[αk+1 ] = δ1p1 . . . δkpk ,


...
[αn ] = δ1q1 . . . δkqk .

If we change the dimensional units of α1 , . . . , αk by the factors β1 , . . . , βk , we get

α1′ = β1 α1 , ..., αk′ = βk αk ,

and consequently for the other parameters



αk+1 = β1p1 · · · βkpk αk+1 , ..., αn′ = β1q1 · · · βkqk αn .

In this new systems of units, the functional relation between the parameters αj′ becomes

φ (β1 α1 , . . . , βk αk , β1p1 . . . βkpk αk+1 , . . . , β1q1 . . . βkqk αn ) = 0.


Table of contents 127

We now take a particular choice for the conversion factors,


1 1
β1 = , ..., βk = .
α1 αk
For the example of the free-falling body this amounts to measuring the length dimension of
the velocity in units of the height.
p The accelerations are measured in units of the gravity
constant g, which imposes then h/g as the unit in which time is measured. Thus√ velocities,
−1/2
with dimensionality length over time, are now expressed in units h(h/g) = gh.
The relation between the parameters is supposed to be independent of the choice of the
scales or units, and with the particular choice for the conversion factors β1 , . . . , βk we get

φ(1, . . . , 1, π1 , . . . , πn−k ) = 0,
| {z }
k

if the new dimensionless parameters are defined through


αk+1 αn
π1 = , ..., πn−k = .
α1p1. . . αkpk α1q1 . . . αkqk
We now have a relation between n − k new parameters, hence something of the form

ψ(π1 , . . . , πm ) = 0.

The set of dimensionless parameters is not unique, as the product of two dimensionless
numbers is again dimensionless. The point is that there are no more than m of those
dimensionless parameters which are independent. The problem has thus been reduced to
dealing with a functional relation between m instead of n parameters. This is of some
consequence when one has experimental data from which to infer the form of the functional
relation. For the falling body one just has to determine the constant in
p
v = C gh,

rather than try to plot v as a function of g and h.


As a further example, we would like to determine the drag force F on a sphere with diameter
d, placed in the steady flow with velocity V of an incompressible, viscous fluid, with material
characteristics ρ and µ. We assume that we have listed all relevant parameters and put

φ(F, d, V, ρ, µ) = 0.

There are 5 parameters with dimensionalities


ML L M M
[F ] = , [d] = L, [V ] = , [ρ] = , [µ] = .
T2 T L3 LT
The independent dimensions are length L, time T and mass M . Furthermore, ρ, V and d
can not be combined to yield a dimensionless number, hence

n = 5, k = 3, r = 3, m = 2.
Table of contents 128

We look for two dimensionless parameters of the form

π 1 = F ρa V b d c , π2 = µρα V β dγ .

Such an initial choice is somewhat arbitrary. We have opted to take π1 linear in F , whereas
π2 will lead to a Reynolds number, always an important quantity in viscous flow problems.
From the requirement that
[π1 ] = 1, [π2 ] = 1,
we find for the exponents

a = −1, b = c = −2, α = β = γ = −1,

and thus the two dimensionless parameters are


F µ 1
π1 = , π2 = = .
ρV 2 d2 ρV d Re
The second one is indeed the inverse Reynolds number. The functional relation between the
two parameters is
ψ(π1 , π2 ) = 0
or, solved for π1 ,
π1 = χ(π2 ).
More explicitly  
2 2 µ
F = ρV d χ = ρV 2 d2 ξ(Re).
ρV d
This might help in rearranging possible experimental data, much more so than the original
relation between five dimensional parameters.

6.3 Turbulence : Reynolds equations


Turbulence often occurs in nature, and one would intuitively describe turbulent flow as some-
thing erratic and unpredictable, quite the opposite of laminar, well-ordered or well-structured
flow. When observing vortices in turbulent flow, it becomes clear that neighbouring material
particles can end up quite far apart, which is not the case in laminar flow. One would feel
that in order to characterize a turbulent flow pattern many more parameters and information
would have to be given than for laminar flow where the velocity gives already quite some
information.
It was said earlier that the transition from laminar to turbulent flow occurs at a critical
Reynolds number, as was shown by Reynolds in one of his famous experiments, in which he
injected a small filament of coloured fluid into a transparent one, as shown in Figure 6.1. At
low Reynolds numbers the coloured fluid flows quietly along with the main stream without
producing vortices. This means that when small disturbances occur, they are immediately
damped out. Quite the contrary happens in turbulent flow, where the coloured fluid ends
Table of contents 129

laminar

turbulent

Figure 6.1. Reynolds experiment.

up in a complicated and seemingly undescribable pattern of vortices. It is also clear that


the mathematical description of turbulence must, per definition as it were, be much more
involved than the description of laminar flow. Experience has shown that the Navier-Stokes
equation provides in this latter case a reasonably accurate picture. Unfortunately, this is
not at all the case for turbulence.
Suppose that we try to plot the velocity, in magnitude, at a certain point as a function of
time, as in Figure 6.2. With some hope we might be able to discern in this erratic graph
a kind of average velocity, around which many fluctuations occur. If we translate this idea
into mathematical terms, we put
v = hvi + v′ ,
where hvi is an average velocity, defined as
Z t+ τ2
1
hvi = v(t′ )dt′ ,
τ t− τ2

and v′ is the fluctuating part of the velocity. The average is still time-dependent, and thus
the time interval τ must be chosen large enough to contain enough fluctuations to give a
meaningful average, and yet τ may not be too large lest the average velocity does not really
depend on time. The notion of such an average requires that within a good approximation
we have that
hhvii ≃ hvi, hv′ i ≃ 0.

hvi

O t

Figure 6.2. Velocity diagram in turbulent flow.


Table of contents 130

Before we even try to apply this averaging procedure to the Navier-Stokes equation, we must
know what happens to space and time derivatives. Space derivatives give no problem, as
Z τ
1 t+ 2
h∇vi = ∇v dt′ = ∇hvi.
τ t− τ2

For time derivatives we have to check that


Z τ
1 t+ 2 ′ 1h τ τ i
h∂t vi = ∂t v dt = v(t + ) − v(t − )
τ t− τ2 τ 2 2
Z t+ τ !
1 2
= ∂t v dt′ = ∂t hvi.
τ t− τ2

The main difficulty lies in the average of quadratic quantities,

hvwi = h(hvi + v′ )(hwi + w′ )i


= hhvihwi + hviw′ + v′ hwi + v′ w′ i ≃ hvihwi + hv′ w′ i.

We will now try to apply this to the Navier-Stokes equation for an incompressible and
homogeneous fluid, really one of the simplest situations one might think of. This implies
that for the time being we will trust the Navier-Stokes equation as being valid also for
turbulence. This is by no means obvious nor generally accepted, and in the light of some
observational evidence probably not so. Nevertheless, when we start from

∇ · v = 0,
1 µ
∂t v + v · ∇v = b − ∇ p + ∇2 v,
ρ0 ρ0
we find that the incompressibility condition is carried over for the average velocity and for
the fluctuations,
∇ · hvi = 0, ∇ · v′ = 0.
For the Navier-Stokes equation we get
1 µ
∂t hvi + hv · ∇vi = hbi − ∇hpi + ∇2 hvi.
ρ0 ρ0
We have to compute

hv · ∇vi = h∇ · vv − v∇ · vi = h∇ · vvi = ∇ · hvvi


= ∇ · hvihvi + ∇ · hv′ v′ i = hvi · ∇hvi + ∇ · hv′ v′ i.

This means that the Navier-Stokes equation becomes


1 µ
∂t hvi + hvi · ∇hvi = hbi − ∇hpi + ∇2 hvi − ∇ · hv′ v′ i
ρ0 ρ0
1
= hbi + ∇ · (−hpi1 + 2µhDi − ρ0 hv′ v′ i) ,
ρ0
Table of contents 131

almost the same as the original equation, but written with average quantities, were it not
that the pressure
−p1 → − (hpi1 + ρ0 hv′ v′ i)

now incorporates a part which is not to be written as a isotropic tensor, except in the case
of irrotational flow, as shown below. The supplementary part in the stress tensor is known
as the turbulent stress tensor of Reynolds, and
1
∂t hvi + hvi · ∇hvi = hbi + ∇ · (−hpi1 + 2µhDi − ρ0 hv′ v′ i)
ρ0
is hence called Reynolds equation.
Suppose now for a moment that the flow is irrotational, meaning there are no vortices. From
1
ω = ∇×v =0
2
we find that
1 1
hωi = ∇ × hvi = 0, ω ′ = ∇ × v′ = 0.
2 2
Furthermore we can compute in this case that

∇ · v′ v′ = v′ ∇ · v′ + v′ · ∇v′ = v′ · (∇v′ − v′ ∇) + v′ · (v′ ∇)


1 1
= 2v′ · Ω′ + ∇ v ′2 = 2ω ′ × v′ + ∇ v ′2
2   2
1 1
= ∇ v ′2 = ∇ · v ′2 1 ,
2 2

so that the pressure remains isotropic


1
p → hpi + ρ0 hv ′2 i.
2
The distinction between turbulent and laminar flow would completely disappear, except for a
change in the effective scalar pressure. This is clearly inadequate, and hence the occurrence
of vortices is an essential characteristic of turbulent flow.
If one wants to continue the description of turbulent flow with Reynolds equations, it is
necessary to introduce hypotheses of a statistical nature about the Reynolds stress, which
correlate v′ with itself. This is a domain still under a lot of discussion, clearly outside the
scope of these lectures.
Another possible approach to turbulence is to Fourier transform the original Navier-Stokes
equations with respect to the space coordinates. First, however, we deduce from the unaver-
aged equation of motion that
1 2
∇ p = −∇ · (v · ∇v) (b = 0),
ρ0
Table of contents 132

which is an equation for the pressure if the velocity is known. The Fourier transform with
respect to the space coordinates is defined as
Z
b
φ(k) = Fk (φ) = φ(x)eik·x dV,
V

where the integration is carried out over the whole physical space. The Fourier transforms of
the incompressibility condition, the equation of motion and the equation for the Laplacian
of the pressure thus yield

k·v
b = 0,
i µk 2
b + Fk (v · ∇v) =
∂t v k pb − b,
v
ρ0 ρ0
k2
pb = −i k · Fk (v · ∇v).
ρ0
Hence we eliminate the pressure to obtain
   
µ 2 1
∂t + k v b= kk − 1 · Fk (v · ∇v).
ρ0 k2

As the Fourier transform of a product, here the scalar product of v with ∇v, is a convolution
integral between the Fourier transforms of each of the factors, the previous equation will be
of the form
  Z
µ 2
∂t + k v b = d3 k′ M(k, k′ ) : v b(k′ )b
v(k − k′ )
ρ0

The discussion of this equation hinges again on what kind of hypotheses one introduces
concerning the kernel M(k, k′ ), so that one is reduced to essentially the same statistical
arguments needed to progress in the Reynolds picture.

6.4 Turbulence : Kolmogorov’s law


The preceding section should not convey the impression that nothing definite is known
about turbulence, just because no satisfactory and consistent mathematical theory has been
achieved sofar.
To begin with, there is a lot of observational information, such as the flattening of the usual
parabolic Hagen-Poiseuille velocity profile when the flow gets turbulent, as shown in Figure
6.3.
Next we can use dimensional analysis to get some information about the turbulent kinetic
energy. If the total kinetic energy T is defined as
Z
1 2
T = ρv dV,
V 2
Table of contents 133

turbulent

H.P.
O v

Figure 6.3. Comparison between turbulent and Hagen-Poiseuille velocity profiles.

then the kinetic energy density for a homogeneous fluid is 21 ρ0 v 2 . The average kinetic energy
density in the turbulent motion is 12 ρ0 hv ′2 i. Per unit mass this becomes
1
W = hv ′2 i,
2
which is then investigated in its spectral representation,
Z Z
W = Wλ dλ = Wk dk.

Here λ represents a wavelength and k is the corresponding wavenumber, k = 2π/λ, and Wλ


or Wk tells us how much energy there is in a mode with wavelength λ or wavenumber k. We
associate λ with the scale of a vortex and know via Wλ or Wk how much vortices of a given
scale contribute to the energy density. The dimensions of this type of problem are
 
2 L2 W L3
[W ] = [V ] = 2 , [Wk ] = = 2.
T k T
We follow the reasoning of Kolmogorov and distinguish, albeit in a very loose and intuitive
way, three different regimes or scales in the turbulence. First of all, the domain of small k
is called the production or hydrodynamic range, because to small k correspond large scales,
hence large vortices or eddies. This is the domain where the vortices are generated, e.g.
when we stir a fluid or put an obstacle in the flow, and where the energy is given to the
vortices. In a sense one could call this a hydrodynamic problem when one views the vortices
as separate entities, to be described together with the remainder of the fluid.
At the other end, the domain of large k is called the dissipation range, because here the small-
scale vortices are vanishing. Returning to the picture of a fluid which is stirred, as soon as
the stirring stops we observe that the vortices get smaller and smaller until all vortices die
out, their energy given up to the main flow.
The domain of intermediate k, corresponding to the domain of medium-sized vortices, is
called the inertial range. The idea of Kolmogorov is to suppose that the energy which is
being fed into the large eddies cascades down via the intermediate vortices to the small-scale
eddies which finally dissipate their energy into the main flow. The graph for the spectral
Table of contents 134

Wk

O k

Figure 6.4. Energy spectrum.

energy density would look like the one given in Figure 6.4.
Kolmogorov hypothesized that in the inertial range the energy density Wk would only depend
on k itself, on the kinematic viscosity ν = µ/ρ0 and on the dissipation Q in the small-scale
eddies. The dimensions of the various parameters are thus

L3 1
[Wk ] = 2
, [k] = ,
T L
   −1
µ M M L2
[ν] = = = ,
ρ0 LT L3 T
 
W L2
[Q] = = 3.
t T

The mass of the particles making up the fluid in turbulent flow does not play a role, because
we started from the turbulent kinetic energy per unit mass and used the kinematic viscosity.
Had we everywhere kept the mass density in, there would have been one supplementary
parameter ρ0 and one supplementary dimension M , leading ultimately to the same two
dimensionless parameters.
The combination of ν and Q never yields a dimensionless number, and there are two dimen-
sions, length and time. Thus

n = 4, k = 2, r = 2, m=2

and the functional relation


φ(Wk , k, ν, Q) = 0
can be rearranged as
ψ(π1 , π2 ) = 0.
The two dimensionless parameters will be taken as

π 1 = Wk ν a Qb , π2 = kν α Qβ ,
Table of contents 135

one linear in Wk , the other linear in k. From


L3+2a+2b L−1+2α+2β
[π1 ] = = 1, [π2 ] = = 1,
T 2+a+3b T α+3β
we deduce that
5 1 3 1
a=− , b=− , α= , β=−
4 4 4 4
and hence  1/4
Wk kν 3/4 ν3
π1 = 5/4 1/4 , π2 = 1/4 = k .
ν Q Q Q
Hence one gets
π1 = χ(π2 )
or
Wk = ν 5/4 Q1/4 χ(kLd ),
where  1/4
ν3
Ld =
Q
is the dissipation length of Kolmogorov.
Another way of arriving at this result is to construct from ν and Q a length scale, as was
done in defining Ld , and a time scale,
 1/2
ν
Td = .
Q
The dimensions of Wk are then
 3/4  −1
L3 ν3 ν3
[Wk ] = d2 = = ν 5/4 Q1/4 ,
Td Q Q
and those of k are  3 −1/4
1 ν
[k] = = .
Ld Q
Kolmogorov then went one step further and assumed that Wk in the inertial range was inde-
pendent of the viscosity, as the viscosity after all was responsible for the ultimate dissipation
of the energy. No matter what the viscosity was, it influenced the dissipation and this in turn
influenced Wk .
Two possible ways are open to compute the form of Wk now. One way is to take
Wk = ν 5/4 Q1/4 χ(kLd ),
and find a form for χ which will make Wk independent of ν. The only possibility is that χ
is some power of its argument times a constant, such that
−5/3
Wk = ν 5/4 Q1/4 Ko kν 3/4 Q−1/4 = Ko Q1/4+5/12 k −5/3
Table of contents 136

or
Wk = Ko k −5/3 Q2/3

where Ko is Kolmogorov’s constant.


Another way of arriving at this is to start the dimensional analysis from

φ(Wk , k, Q) = 0 or ψ(π) = 0,

with
L3 1 L2
[Wk ] = 2 , [k] = , [Q] = 3 ,
T L T
and construct the only dimensionless parameter π which is possible. This then leads again
to the same result, which gives the slope of the graph of Wk in the inertial range, for a given
dissipation Q, as indicated in Figure 6.5.
When the kinematic viscosity ν is no longer deemed important, the concepts of a dissipation
length Ld or a dissipation time Td in terms of ν and Q are not possible. One could construct
new length and time scales from Wk and Q, but these now involve both the energy in
the inertial range and the dissipation themselves, not a very clarifying way to look at the
problem.
Once a more qualitative theory of turbulence emerges and is generally accepted, Kolmogo-
rov’s law will serve as a test, when such a theory is applied to the inertial range as defined
above. However, such is not the case at present.

6.5 Matched asymptotic expansions


If we recall the equations describing the steady laminar flow of a homogeneous fluid, we have
two different sets of equations according to whether the fluid is considered viscous or ideal.

Wk
5
k− 3

i.r.
O k

Figure 6.5. Power law in the inertial range.


Table of contents 137

In the dimensionless notations of section 6.1 (without the asterisks) and in the absence of
body forces we find

∇ · v = 0,
1 2
v · ∇v = −∇ p + ∇ v,
Re
for a viscous fluid and

∇ · v = 0,
v · ∇v = −∇ p,

for an ideal one. In both cases the pressure in equilibrium has been adjusted so as to make
Euler’s number equal to one. The equation of motion for viscous fluids contains a Laplacian,
whereas for ideal fluids only first-order derivatives occur. This changes the mathematical
formulation in the sense that more boundary conditions can be imposed for viscous fluids,
because of the higher order of the partial differential equations.
In order to see how one could deal with problems of this nature, where the vanishing of
a small parameter (such as in the above case with 1/Re small) alters the mathematical
character of the equations, we will expose in the present section the method of matched
asymptotic expansions and defer the discussion of the fluid equations themselves to the next
section, on boundary layer theory. We start from a mathematical model equation,

d2 f df
ε 2
+ =a
dx dx

with ε a small parameter. We let

x ∈ [0, 1], a ∈]0, 1[,

and choose as boundary conditions

f (0) = 0, f (1) = 1.

For such a simple model equation we know the exact solution with the given boundary
conditions,
e−x/ε − 1
fε (x) = (1 − a) −1/ε + ax.
e −1
When ε tends to zero in this solution, one finds

f0 (x) = lim fε (x) = 1 − a + ax.


ε→0, x fixed

The same limiting process in the equations would yield


df0
= a,
dx
Table of contents 138

and hence we see that f0 is the solution of this equation with the boundary condition

f0 (1) = 1.

Unfortunately, as we are dealing with a first-order equation, the other boundary condition
can not be fulfilled and
f0 (0) = 1 − a.
This means that
fε (0) −→
/ f0 (0),
ε→0
so that
f0 (x) = 1 − a + ax
is a fair approximation to
e−x/ε − 1
fε (x) = (1 − a) + ax,
e−1/ε − 1
as long as x is not too small, thus outside a boundary layer in the immediate vicinity of the
origin. Because f0 (x) serves as a first approximation to fε (x) outside this boundary layer,
it is sometimes called the outer expansion f o (x). It is also natural that the outer boundary
condition is satisfied, not the inner one.
The troublesome domain is where x and ε are both small, in such a way as to make the two
terms in the left-hand side of the original model equation comparable,
d2 f df
ε 2
∼ .
dx dx
Then f varies so fast that the second derivative is large compared to the first derivative. We
now use a transformation for this inner or boundary layer which stretches x together with
ε. We formally put
x = εξ
and keep ξ finite, even when x and ε tend to zero. We thus have that

F (ξ) ≡ f (εξ)

and find for the derivatives that


dF df d2 F 2
2d f
=ε , = ε .
dξ dx dξ 2 dx2
The equation for f is transformed into
d2 F dF
2
+ = εa,
dξ dξ
subject to the inner boundary condition that f (0) or F (0) vanishes. When ε now tends to
zero, there only remains
d2 F0 dF0
+ = 0.
dξ 2 dξ
Table of contents 139

However, as this still is a second-order equation, a second boundary condition must be


imposed to determine the solution completely. It would be wholly illogical to use the outer
boundary condition, as the inner expansion, leaving the ratio x/ε finite, cannot possibly be
a good approximation outside the boundary layer. Something better is needed. The solution
for F0 is
F0 (ξ) = Ae−ξ + B,
and the inner boundary condition requires that A and B be opposite and thus

F0 (ξ) = B(1 − e−ξ ).

On the other hand, if we return to the exact solution of the original equation, transform it
into
e−ξ − 1
Fε (ξ) = (1 − a) −1/ε + εaξ
e −1
and let ε tend to zero,

F0 (ξ) = lim Fε (ξ) = (1 − a)(1 − e−ξ ),


ε→0, ξ fixed

we see from the comparison of both forms of F0 (ξ) that

B = 1 − a.

F0 (ξ) is called the inner expansion f i (ξ), to be determined with the inner boundary condition
plus something else.
Sofar, we know the exact solution, fε (x) or Fε (ξ), depending on whether we choose to use x
or ξ as the independent variable. There are two approximations:
Outer expansion

f o (x) = f0 (x) = lim fε (x),


ε→0, x fixed
f o (1) = f0 (1) = 1;

Inner expansion

f i (ξ) = F0 (ξ) = lim Fε (ξ) = lim fε (εξ),


ε→0, ξ fixed ε→0, ξ fixed

f i (0) = F0 (0) = 0.

When these different solutions are plotted in the same graph, one gets two complementary
approximations to the exact solutions, as shown in Figure 6.6. Furthermore

lim f o (x) = lim f i (ξ) = 1 − a


x→0 ξ→∞
Table of contents 140

f (x)
1
f o (x)

1−a
f i ( xǫ )

1
O x

Figure 6.6. Graphs of the exact solution and the inner and outer approximations

The inner limit of the outer approximation and the outer limit of the inner approximation
coincide. This property, which we can verify here on the exact solution, is used as a require-
ment when the exact solution is not known, and serves as the missing boundary condition
for the inner expansion. If we call coi the inner limit of the outer approximation,

coi = lim f o (x)


x→0

and cio the outer limit of the inner approximation,

cio = lim f i (ξ),


ξ→∞

then obviously one requires that


coi = cio .
In this sense the inner and the outer expansions are matched, hence the name of the method
of matched asymptotic expansions, of which only the zeroth-order approximation in ε was
exposed here. A composite approximation f c (x), valid both in and outside the boundary
layer, is x x
f c (x) = f i + f o (x) − coi = f i + f o (x) − cio .
ε ε
Indeed, the outer and inner approximations to f c (x) are

f co (x) ≃ cio + f o (x) − coi = f o (x),


x x
f ci (x) ≃ f i + coi − coi = f i .
ε ε
Table of contents 141

To recapitulate what was found from this discussion, as we will need it for those cases where
there is no exact solution to guide us, the method of matched asymptotic expansions runs
as follows. For the outer solution, where ε is small but x finite, f o is the solution of
df o
= a, f o (1) = 1,
dx
using the outer boundary condition to determine the solution completely. This differential
equation is obtained by a simple limiting procedure from the original equation. The solution
is then
f o (x) = 1 − a + ax,
with
f o (0) = 1 − a.
For the inner solution, the substitution

x = εξ

is made and then the limit is taken for small ε afterwards,

d2 f i df i
+ = 0, f i (0) = 0,
dξ 2 dξ
using the inner boundary condition. In view of the matching recipe

lim f i (ξ) = lim f o (x),


ξ→∞ x→0

one gets
lim f i (ξ) = 1 − a,
ξ→∞

determining f i (ξ) completely as in

f i (ξ) = (1 − a)(1 − e−ξ ).

One could use more refined versions of the matching principle to obtain higher-order approx-
imations. Nevertheless, the zeroth-order approximation is still a useful one, in determining
where the boundary layer is situated in a given problem and how its thickness varies with
some small parameters.

6.6 Boundary layer


For the steady, laminar and plane flow of a homogeneous fluid, the fluid can be regarded as
ideal in the middle of the flow, and it is only near the wall of the channel that the viscosity
effects will play a role and give rise to the boundary layer. The velocity profile will then
Table of contents 142

intuitively look as sketched in Figure 6.7. The width δ(ε) of the boundary layer will depend
on the small parameter ε = 1/Re. The boundary conditions are in general

x2 = 0 : v1 = 0, v2 = 0,
x2 = 1 : v1 = 1, v2 = 0, p = 1,

as we are working in nondimensional variables, where distances are measured in units of


the half-width of the channel, which greatly exceeds the width of the boundary layer, and
where the reference velocity and pressure are the values in the region where the flow is well
developed and ideal.
As the picture is two-dimensional in the space coordinates, starting with the outer region
where v only has a parallel component, we write

∇ · v = 0,
v · ∇v = −∇ p,

out in full

∂1 v1 = 0,
v1 ∂1 v1 = −∂1 p = 0,
−∂2 p = 0.

The solution, keeping in mind the outer boundary conditions, is

v1 = v1 (x2 ),
p = 1,

such that
v1 (1) = 1.
In other words, the pressure is constant and v1 can at most vary with the height in the
channel.

x2

δ(ǫ)

O x1

Figure 6.7. Velocity profile.


Table of contents 143

For the inner region, we need to consider

∇ · v = 0,
1 2
v · ∇v = −∇ p + ∇ v = −∇ p + ε∇2 v,
Re
giving explicitly that

∂1 v1 + ∂2 v2 = 0,
(v1 ∂1 + v2 ∂2 )v1 = −∂1 p + ε(∂12 + ∂22 )v1 ,
(v1 ∂1 + v2 ∂2 )v2 = −∂2 p + ε(∂12 + ∂22 )v2 .

The small variables are the coordinate and velocity components across the boundary layer,
as the velocity has no perpendicular component at the wall nor inside the ideal flow region.
The stretching of variables, discussed in the previous section, becomes
1
x2 = ε m η = η,
Rem
1
v2 = εm V = V,
Rem
where the precise value of m is yet to be determined. The reason for an equal stretching
in x2 and v2 lies in the fact that the incompressibility condition does not contain ε. Hence
one would not expect one term or the other to become dominant as a result of a stretching
carried out in another equation. Put differently, any other stretching would eliminate one
of the two terms in the incompressibility condition and lead to a constant solution for the
velocity component left in. With the indicated stretching one has

∂2 = ε−m ∂η

and the set of viscous fluid equations can be rewritten as

∂1 v1 + ∂η V = 0,
(v1 ∂1 + V ∂η )v1 = −∂1 p + ε(∂12 + ε−2m ∂η2 )v1 ,
(v1 ∂1 + V ∂n )V = −ε−2m ∂η p + ε(∂12 + ε−2m ∂η2 )V.

If the viscosity effects are to be comparable to the other terms, which was after all the whole
purpose of introducing a stretching, the only possible choice is
1
1 − 2m = 0 =⇒ m= ,
2
giving
1
x2 = ε1/2 η = √ ,
Re
−1/2

∂2 = ε ∂η = Re∂η ,
1
v2 = ε1/2 V = √ V.
Re
Table of contents 144

Any other choice for the stretching of the second coordinate would either make the viscos-
ity effects vanish also in the boundary layer or effectively decouple these effects from the
remainder of the description of the flow. We thus find

∂1 v1 + ∂η V = 0,
(v1 ∂1 + V ∂η )v1 = −∂1 p + (ε∂12 + ∂η2 )v1 ,
(v1 ∂1 + V ∂η )V = −ε−1 ∂η p + (ε∂12 + ∂η2 )V.

The small ε limit now is

∂1 v1 + ∂η V = 0,
(v1 ∂1 + V ∂η )v1 = −∂1 p + ∂η2 v1 ,
∂η p = 0.

In the boundary layer the pressure can only vary in the parallel direction, whereas outside
it is constant. The condition for the matching of the two approximations yields

lim po (x1 , x2 ) = lim p0 = 1 = lim pi (x1 ),


x2 →0 x2 →0 η→∞

in other words the pressure is constant throughout the entire boundary layer. This simplifies
the description still further, and reverting to more usual notations we see that the equations
for the flow in the boundary layer are

∂ 1 v1 + ∂ 2 v2 = 0
(v1 ∂1 + v2 ∂2 )v1 = ∂22 v1

The analysis given in the previous pages assures us that all important terms which matter
inside the boundary layer, and only those, are retained in the final form of the equations.
This is to be contrasted, at constant pressure, with the outer equations, of which only

∂ 1 v1 = 0

needs to be discussed.
Finally there remains the width of the boundary layer itself, which is seen from the scaling
to go as
1
δ(ε) ∼ ε1/2 = √ .
Re
The more important the viscosity, the larger the boundary layer will be. Boundary layers are
encountered not only in the laminar flow in channels, but also whenever a fluid is flowing past
an obstacle, such as air around the wing of an aircraft. In the latter problem the boundary
layer becomes extremely important, both for the stability of the flight and for the efficiency
with which turbulence can be avoided.
Chapter 7

Linear elasticity

• Elastostatic problems lead to Beltrami-Michell relations

• Elastic waves include seismographic examples

145
Table of contents 146

7.1 Basic equations


For the elastic media considered in this chapter, the stress-strain relation will be Hooke’s
law,
T = λ(tr E)1 + 2µ E

with the restriction that the displacement gradients are sufficiently small so that the small
strain tensor E can be used. Furthermore, we need the equation of motion,

ρ v̇ = f + ∇ · T,

the symmetry of the stress tensor,


T = Tt ,
the definition of the small strain tensor,
1
E = (∇u + u∇),
2
and the definition of the velocity,
v = ẋ = u̇.
The velocity is defined in terms of the displacement away from the time-independent equi-
librium state.
Expressing the stress tensor as a function of the displacement gives

T = λ(∇ · u)1 + µ(∇u + u∇)

and the divergence of this tensor yields

∇ · T = ∇ · [λ(∇ · u)1 + µ(∇u + u∇)]


= (λ + µ)∇∇ · u + µ∇2 u.

Hence the set of basic equations can be reduced to the equation of Navier-Cauchy

ρ ü = f + (λ + µ)∇∇ · u + µ∇2 u

This is an equation determining u when the forces f are given. Although superficially similar
to the Navier-Stokes equation

ρ v̇ = f − ∇p + (λ + µ)∇∇ · v + µ∇2 v,

there are two essential differences. First, in elasticity there is no hydrostatic pressure and
thus no need for a separate equation of state. Second, the material time-derivative is a
second-order one, altering the mathematical structure of the equation as will be obvious
when studying elastic waves in section 7.3.
Table of contents 147

For sufficiently small displacements the acceleration term can be linearized,

ρ ü = ρ(∂t + v · ∇)(∂t + v · ∇)u ≃ ρ0 ∂t2 u,

and from now on we will restrict ourselves to the theory of linear elasticity, where ρ or rather
ρ0 is taken as one of the material constants, together with Lamé’s constants λ and µ. This
is the reason why the equation of continuity was not recalled in the set of basic equations.
Strictly speaking of course, ρ is not constant in the sense of λ and µ, but in the linearized
version of the equation of Navier-Cauchy only the equilibrium value ρ0 enters

ρ0 ∂t2 u = f + (λ + µ)∇∇ · u + µ∇2 u

Note that the concept ‘linear’ enters here in two different ways. When dealing with the
constitutive relations, Hooke’s law is a linear relation between the stress and the strain
tensors in isotropic elastic media. Now, in ‘linear’ elasticity, the nonlinear equation of motion
(obtained with the help of the linear constitutive relation) is itself linearized with respect to
the perturbations away from the equilibrium, stress-free state.

7.2 Elastostatics
For elastostatic problems the Navier-Cauchy equation reduces to

f + (λ + µ)∇∇ · u + µ∇2 u = 0.

The problem consists in the determination of u, given f and some prescribed boundary con-
ditions, here in the form of boundary displacements. Certain problems, however, involve
boundary conditions in terms of the stresses rather than the displacements, and then the
set of basic equations has to be rearranged differently. The inverted form of Hooke’s law is
used instead, and we have to start from

f + ∇ · T = 0,
1+ν ν
E= T − (tr T)1.
E E
The second equation is necessary, because only part of the information about T enters into
the equation of equilibrium. However, here we are not really interested in E. Therefore we
start from the compatibility conditions of St. Venant,

∇ × (∇ × E)t = 0,

written in a different but equivalent form

∇∇ tr E + ∇2 E − ∇∇ · E − (∇∇ · E)t = 0.
Table of contents 148

Using the definition of E, it is easy to verify that indeed

∇∇ tr E + ∇2 E − ∇∇ · E − (∇∇ · E)t
1 1 1 1
= ∇∇∇ · u + ∇2 ∇u + ∇2 u∇ − ∇(∇2 u + ∇∇ · u) − (∇2 u∇ + ∇∇∇ · u) = 0.
2 2 2 2
Substitution of E expressed as function of T in the modified compatibility conditions yields

∇∇ tr T + (1 + ν)∇2 T − ν(∇2 tr T)1 − (1 + ν) ∇∇ · T + (∇∇ · T)t = 0,

or with the help of the equilibrium equation,

∇∇ tr T + (1 + ν)∇2 T − ν(∇2 tr T)1 + (1 + ν)(∇f + f ∇) = 0.

Taking the trace of this equation gives an intermediate result,


1+ν
∇2 tr T = − ∇ · f,
1−ν
which we use to eliminate ∇2 tr T in the preceding equation, to finally obtain

1 ν
∇2 T + ∇∇ tr T + (∇ · f )1 + ∇f + f ∇ = 0
1+ν 1−ν

These equations are called the Beltrami-Michell compatibility conditions for the stress, which
have to be considered together with the equilibrium equations

∇ · T + f = 0.

To solve the Beltrami-Michell conditions for T when f is given is in general an involved


problem.
An important special case is when the deformation of the elastic medium only results from
a surface loading and not from body forces, or when the body forces are constant as in the
case of gravity. Then there follows that

∇2 tr T = 0,

and hence also that


∇2 tr E = ∇2 ∇ · u = 0.
The cubic dilatation coefficient is a harmonic function in this case. Since
1 1 ν
∇ · E = ∇2 u + ∇∇ · u = − ∇ tr T,
2 2 E
one can take the Laplacian of this equation as well as of the Beltrami-Michell conditions for
constant f and find that
∇4 u = 0, ∇4 T = 0.
Table of contents 149

Both the displacement and the stress are thus biharmonic functions.
In particular, for plane stress in the absence of body forces, the equilibrium conditions are
explicitly

∂1 T11 + ∂2 T21 = 0,
∂1 T12 + ∂2 T22 = 0,

and the other components are zero, T3j = Tj3 = 0. From the first equation we learn that

T11 = ∂2 α, T21 = −∂1 α,

whereas the second equation similarly yields

T12 = ∂2 β, T22 = −∂1 β,

with α and β new functions to be determined. However, in view of the symmetry of the
stress tensor,
T12 = T21 = −∂1 α = ∂2 β,
there exists a single function ϕ such that

α = ∂2 ϕ, β = −∂1 ϕ.

This function is called Airy’s stress function and it determines the components of the plane
stress tensor as

T11 = ∂22 ϕ,
T12 = −∂1 ∂2 ϕ,
T22 = ∂12 ϕ.

The requirement that


∇2 tr T = 0
thus becomes
∇2 tr T = ∇2 (T11 + T22 ) = ∇2 (∂12 + ∂22 )ϕ = ∇4 ϕ = 0,
in other words Airy’s stress function is also a biharmonic function. Once this function is
known, keeping in mind the boundary conditions, the stress distribution inside the continuum
can be completely determined.

7.3 Elastic waves in general linear media


In this section a discussion will be given of plane waves in general linear elastic media, which
may be isotropic or anisotropic. The stress-strain relation is supposed to be linear but not
necessarily isotropic
T = C : E.
Table of contents 150

This has to be combined with


ρ0 ∂t2 u = ∇ · T,
1
E = (∇u + u∇),
2
where the equation of motion is written in the absence of body forces. The starting point is
a plane-wave solution for the displacement, say
u = a ϕ(k · x − ωt),
where k is the wave vector and ω the wave frequency. We need to compute some derivatives
of u,
∂t u = −ωa ϕ′ , ∂t2 u = ω 2 a ϕ′′ ,
∇u = ka ϕ′ , u∇ = ak ϕ′ ,
where the dashes refer to derivatives of ϕ with respect to its full argument k · x − ωt. Hence
we find for the strain tensor
1
E = (ka + ak)ϕ′ ,
2
and similarly for the stress tensor
1
T = C : E = (C : ka + C : ak)ϕ′ = C : ka ϕ′ ,
2
in view of the symmetry of the elasticity tensor C with respect to its first two or last two
sets of indices. The divergence of T is thus
∇ · T = k · C : ka ϕ′′ = (k · C · k) · a ϕ′′
and finally the equation of motion yields
ρ0 ω 2 a ϕ′′ = (k · C · k) · a ϕ′′ .
The previous expression can be divided by k 2 , and new notations introduced, n for the unit
vector in the direction of wave propagation and c for the phase velocity of the waves,
k ω
n= , c= .
k k
′′
As ϕ is not identically zero, we deduce that
(n · C · n − ρ0 c2 1) · a = 0.
When written in scalar form, we find a set of three homogeneous algebraic equations in the
three unknown components of the amplitude a of the displacement. If we want a nontrivial
solution, we have to choose c in such a way that

det(n · C · n − ρ0 c2 1) = 0

In general, this dispersion law will yield three solutions for ρ0 c2 , namely the three eigenvalues
of the second-order tensor n · C · n. In more physical terms, this indicates that for a given
direction of wave propagation there are in general three different phase velocities and thus
three different wave modes possible. This is valid for any linear elastic medium, which in
general may be anisotropic.
Table of contents 151

7.4 Elastic waves in linear isotropic media


We will now compute the phase velocities explicitly in the case of a linear isotropic elastic
medium. We can do this in two ways. A first possibility is to use the explicit expression for
C in its component form

Cijkℓ = λ δij δkℓ + µ(δik δjℓ + δiℓ δjk ) (ν = 0)

when the medium is isotropic, and find that

n · C · n = (λ + µ) nn + µ 1.

The dispersion law thus is


 
det (λ + µ) nn + (µ − ρ0 c2 )1 = 0.

Another way is to substitute a plane-wave solution into the Navier-Cauchy equation without
body forces and find
(λ + µ) nn · a + (µ − ρ0 c2 ) a = 0.
Again the dispersion law is the same. As the determinant of a second-order tensor is one of
its invariants, we can choose a suitable reference frame in which the determinant is easily
computed. If we take the third axis of the reference frame along the direction of wave
propagation n, we have explicitly that

µ − ρ0 c 2 0 0

0 µ − ρ c 2
0 =0
0
0 0 λ + 2µ − ρ0 c
2

and hence
(µ − ρ0 c2 )2 (λ + 2µ − ρ0 c2 ) = 0.
If we define two wave velocities cP and cS by

 1/2  1/2
λ + 2µ µ
cP ≡ , cS ≡
ρ0 ρ0

we see that the dispersion law has a double root in cS and a single root in cP . Actually,
each root is to be taken once with the plus and once with the minus sign, but that merely
amounts to wave propagation in the opposite direction without a change in the character of
the wave. To find out the nature of the waves, it is necessary to go back to

(λ + µ) nn · a + (µ − ρ0 c2 ) a = 0.

For the P -wave we have that

c = cP , ρ0 c2 = λ + 2µ,
Table of contents 152

and thus
(λ + µ)(nn · a − a) = 0.
As λ + µ is nonzero, this learns that

n (n · a) = a.

In other words, a and n are parallel, the displacement is in the direction of wave propagation
and the P -wave is a longitudinal wave.
As for the S-wave, with
c = cS , ρ0 c2 = µ,
there remains that
(λ + µ) nn · a = 0
or, equivalently,
nn · a = 0.
Hence a is now perpendicular to n and the S-wave is a transverse wave. Actually, there are
two S waves, for which the wave amplitudes a may be taken orthogonal to each other, in
the plane perpendicular to n. In view of the fact that
 1/2  1/2
µ λ + 2µ
cS ≡ < ≡ cP ,
ρ0 ρ0

of the waves propagating along n the P -wave travels faster than the S-waves. The labels
P and S come from seismology where the waves connected with earthquakes are called the
primary and the secondary waves. From the delay between the arrivals of the P - and S-waves
one can get some estimates of the depth of the epicentre. For the Earth the numerical values
are
cP
cP ≃ 8 − 14km/s, cS ≃ √ ,
3
depending on the rock strata through which the waves travel. From this one can deduce
that for rock layers the Lamé constants are almost equal
3µ λ + 2µ
3c2S ≃ c2P =⇒ ≃ =⇒ λ ≃ µ.
ρ0 ρ0
Another, more far reaching conclusion to be drawn from these two wave types is that the
Earth has a molten or liquid core, because at seismographic observatories located at the
antipodes of the epicentre, only the P -wave is registered, as shown in Figure 7.1. As a
longitudinal and dilatational wave, the P -wave is able to propagate in solids as well as in
fluids, whereas the S-wave as a transverse and shear wave can only propagate in solids.
A fluid cannot support shear waves. From the location of the epicentre and the region in
the antipodes where only the P -wave gets through, one can estimate the dimensions of the
molten core region with surprising accuracy. What cannot be deduced in this way is whether
inside the molten (iron) core there is itself a smaller but solid nucleus.
Table of contents 153

E
S P
S

Figure 7.1. P - and S-earthquake waves.

The branch of seismography which tries to infer information about the interior of the Earth,
from the way in which different wave types propagate but can only be observed at the surface,
is called seismology. Similar concepts have recently been developed for wave propagation
in the Sun, where helioseismology now gives very reliable information about the otherwise
inaccessible stratification of the interior of the Sun, and in stars, where asteroseismology tries
to do the same for stellar interiors. Contrary to the Earth, however, the Sun and stars cannot
be treated as elastic media and one has to go back to gas- or plasma-dynamic descriptions.
In view of the importance of waves in elastic media, one can use
λ+µ
c2P − c2S =
ρ0
to rewrite the Navier-Cauchy equation as

∂t2 u = (c2P − c2S ) ∇∇ · u + c2S ∇2 u

We write the displacement in its Helmholtz representation,

u = ∇ψ + ∇ × a

where ψ is a scalar and a a vector potential. This decomposition is always possible, and one
can even impose the auxiliary gauge condition that

∇ · a = 0.

In this representation one has


∇ · u = ∇2 ψ
Table of contents 154

and hence the Navier-Cauchy equation becomes

∇ ∂t2 ψ + ∇ × ∂t2 a = c2P ∇ ∇2 ψ + c2S ∇ × ∇2 a

or
∇(∂t2 ψ − c2P ∇2 ψ) + ∇ × (∂t2 a − c2S ∇2 a) = 0.
When taking the divergence of this equation, there follows that the scalar potential obeys
an equation of d’Alembert with a right-hand side which is itself a harmonic function,

∂t2 ψ − c2P ∇2 ψ = Ψ, ∇2 Ψ = 0.

For this equation of d’Alembert the characteristic velocity is cP . Similarly, because the
vector potential a could be chosen divergence-free, it obeys another equation of d’Alembert
with again a harmonic function on the right-hand side,

∂t2 a − c2S ∇2 a = A, ∇2 A = 0.

The characteristic velocity in the d’Alembertian is now cS . Both harmonic functions Ψ and
A are furthermore linked by
∇ Ψ + ∇ × A = 0.
For the special choice when these harmonic functions are both zero, one recovers two equa-
tions of d’Alembert, which give plane-wave solutions for both the scalar and the vector
potential and hence for the longitudinal and the transverse part of the displacement in its
Helmholtz decomposition. This brings one back to results derived earlier in a more direct
but less general way.
Chapter 8

Epilogue: Kinetic theory

• Kinetic theory is a more advanced level of description of continuous media

• It is shown how one can derive the basic equations, used in previous chap-
ters, from a general kinetic equation

155
Table of contents 156

8.1 Kinetic equations


Up to now, the description of continua has been given in physical space, with as independent
coordinates the usual space and time coordinates. This implies that the velocity is one of the
macroscopic and dependent quantities, much as the pressure or density, in the sense that it
is given at each physical point and for each moment as an average for the fluid or the elastic
medium. For a fluid with its rather higher coherence, this might be a fair assumption, but
from detailed studies of gases and plasmas it is known that neighbouring material particles
can have quite different velocities. In the usual continuum description these differences in
velocities are just averaged out, but it is conceivable that one would like to get a description
in which these different velocities play a more essential role.
The resulting kinetic theory is in a sense also a continuum description but on a higher level.
Such a description requires the introduction of a 6-dimensional phase space instead of the
usual 3-dimensional physical space, a phase space in which the independent coordinates are
the three (physical) space coordinates xi and the three velocity coordinates wi . These are for
the purpose of kinetic theory all independent coordinates. Using still the 3-vector notation
x for the space coordinates and w for the velocity, we find that the mathematical point
(x, w) in phase space locates a material particle with position in physical space at x and
with velocity w.
The next step in the development of the theory of kinetic equations is the introduction of a
density or distribution function ϕ(x, w, t) in phase space, in such a way that ϕ(x, w, t)d3 x d3 w
gives the mass of material particles which are contained in an elementary range d3 x around
the point x and such that their velocities lie in an elementary range d3 w around w. Schemat-
ically this is indicated in Figure 8.1. Instead of d3 w we will use dW , as we wrote dV instead of
d3 x in the introductory chapters, where the usual continuum description was developed. The
(velocity) space in which w varies is denoted by W . Such a distribution function ϕ(x, w, t)
is only meaningful as a measure in phase space, if upon integration over all possible values

w + d3 w
w

O x x + d3 x x

Figure 8.1. Description in phase space.


Table of contents 157

of the velocities the usual mass density ρ(x, t) in physical space is recovered,
Z
ρ= ϕ dW
W

For the purpose of the present chapter, where we only want to give the link between the
kinetic equations and the usual continuum equations without going into a detailed study of
the kinetic equations themselves, a kinetic equation is an equation of the general form
Dϕ δϕ
= .
Dt δt
On the left-hand side use has been made of the total time derivative in phase space, when
the particles are followed in their motion, whereas on the right-hand side the change in ϕ is
due to the interactions of the particles between themselves. The total time derivative can
be expanded as
Dϕ ∂ϕ ∂ϕ dx ∂ϕ dw
= + · + ·
Dt ∂t ∂x dt ∂w dt
∂ϕ ∂ϕ ∂ϕ
= +w· +b· ,
∂t ∂x ∂w
if the microscopic laws of motion are incorporated
dx dw
w= , b= .
dt dt
Here b is the normalized force density at the microscopic level, and the notation ∂/∂ x
replaces ∇ in this chapter to avoid confusion with ∂/∂ w.
The kinetic equation thus is

∂ϕ ∂ϕ ∂ϕ δϕ
+w· +b· =
∂t ∂x ∂w δt

If the particles do not interact, the right-hand side vanishes and we get the Liouville equation,
which essentially expresses the conservation of particle mass or number in phase space. If the
right-hand side is computed for a gas of hard-sphere particles undergoing binary collisions
only, one gets the Boltzmann equation. There are many more examples of kinetic equations
to be given, according to the hypotheses one introduces about the interactions between the
particles, but a deeper discussion of these would lead us to far.
The study of kinetic equations thus has two aspects: first, the making of suitable assumptions
about the right-hand side involving the interactions, and second, the solution of the equation
itself for the distribution function. This is certainly not the purpose of this chapter, but
essential for our considerations is the fact that the left-hand side of the kinetic equations
has a definite expression, and we will now show how the usual continuum equations can be
deduced from it.
Table of contents 158

8.2 Macroscopic averages


In view of the fact that the distribution gives the mass density upon integration over w, we
define the average velocity v as
Z
ρ v = ρhwi = ϕ w dW.
W

In more general terms, when we have a certain microscopic quantity Q, such as velocity or
energy, we define the average or macroscopic corresponding quantity hQi as
Z
ρhQi = ϕQ dW.
W

This is consistent with both Z


ρ= ϕ dW,
W

if we put Q = 1 and get hQi = 1, and with


Z
ρhwi = ϕw dW,
W

if Q = w and hQi = v. For further use we define the peculiar velocity

w′ = w − v = w − hwi,

with zero average, to express the way in which the individual or microscopic velocities differ
from the averaged or fluid velocity.
In order to get an equation for hQi, we multiply

∂ϕ ∂ϕ ∂ϕ δϕ
+w· +b· =
∂t ∂x ∂w δt
with Q and integrate over all w,
Z  
∂ϕ ∂ϕ ∂ϕ
Q +w· +b· dW = ∆Q,
W ∂t ∂x ∂w

where we define Z
δϕ
∆Q = Q dW
W δt
as the contribution to the changes in Q due to the particle interactions. We will compute
some of the integrals separately. The first term gives
Z Z Z  
∂ϕ ∂ ∂Q ∂ ∂Q
Q dW = (ϕQ)dW − ϕ dW = (ρhQi) − ρ .
W ∂t W ∂t W ∂t ∂t ∂t
Table of contents 159

For the second term we find


Z Z Z
∂ϕ ∂ ∂Q
Qw · dW = · (w Qϕ)dW − ϕw · dW
W ∂x W ∂x W ∂x
 
∂ ∂Q
= · (ρhw Qi) − ρ w · .
∂x ∂x
For the last term we will explicitly use the fact that the distribution function vanishes at
the boundaries in phase space, meaning simply that there are no particles at infinite or very
large distances, nor with infinite or very large velocities,
Z Z Z
∂ϕ ∂ ∂
Qb · dW = · (b Qϕ)dW − ϕ · (b Q)dW
W ∂w W ∂w W ∂w
 

= −ρ · (b Q) .
∂w
Collecting all the results one finds
 
∂ ∂ ∂Q ∂Q ∂
(ρhQi) + · (ρhw Qi) = ρ +w· + · (b Q) + ∆Q
∂t ∂x ∂t ∂x ∂w

8.3 Macroscopic equations


To recover the continuity equations, the equations of motion and the energy balance, we will
successively put Q equal to 1, w′ and 12 w′ 2 . For the case that Q = 1 we find for some of the
expressions that
 
∂Q
hw Qi = v, = 0, ∆Q = 0.
∂w
The last term, ∆Q, vanishes because it is assumed that the interactions between the par-
ticles do not affect their numbers or masses. The idea of the conservation of mass is thus
incorporated into the theory of kinetic equations. Hence
 
∂ ∂ ∂Q
(ρhQi) + · (ρhw Qi) = ρ b · + ∆Q
∂t ∂x ∂w
becomes the equation of continuity,

∂ ∂
ρ+ · (ρ v) = 0
∂t ∂x

For the equation of motion, we put Q = w′ so that hQi = 0, and compute that
1
hw Qi = hww′ i = hw′ w′ i + hvw′ i = hw′ w′ i = − T,
ρ
Table of contents 160

for the terms on the left hand side of the moment equation governing the averages, whereas
for the right hand side we need
   
∂Q ∂ ∂v
= (w − v) = − ,
∂t ∂t ∂t
     
∂Q ∂ ∂ ∂
w· = w· (w − v) = − w · v = −v· v,
∂x ∂x ∂x ∂x
     
∂ ∂ ∂
· (bQ) = · b(w − u) = b · w = hb · 1i = hbi .
∂w ∂w ∂w
Here T is the stress tensor, defined as
Z
′ ′
T = −ρhw w i = − ϕ(w − v)(w − v)dW,
W
and it is worth noting that in this derivation the stress tensor is symmetric by definition.
We have also used the fact that the velocity dependent forces are essentially restricted to
the Lorentz term,
q
b = (w × B),
m
for which

·b=0
∂w
is valid. All this leads to the intermediate result that
 
∂Q ∂Q ∂ ∂v ∂
+w· + · (b Q) = − −v· v + hbi .
∂t ∂x ∂w ∂t ∂x
If in addition the interaction between any two particles conserves their total momentum, as
in the case of elastic collisions between like particles, ∆ w vanishes. Substitution of all this
finally gives the desired equation of motion

∂v ∂ 1 ∂
+v· v = hbi + ·T
∂t ∂x ρ ∂x

Finally, we put Q equal to 21 w′ 2 and compute


1 2 1 2 1 2
hw Qi = hw w′ i = vhw′ i + hw′ w′ i,
2 2 2
for the terms on the left hand side of the moment equation, whereas for the right hand side
we get
     
∂Q 1 ∂ ′2 ′ ∂ ′ ∂
= w = w · w = − hw′ i · v = 0,
∂t 2 ∂t ∂t ∂t
     
∂Q 1 ∂ ′2 ′ ∂ ′ ∂ 1 ∂
w· = w· w = ww: w = − hw′ wi : v= T: v,
∂x 2 ∂x ∂x ∂x ρ ∂x
     
∂ 1 ∂ ′2 ′ ∂
· (bQ) = b· w = wb: w = hw′ b : 1i = hw′ · bi = 0.

∂w 2 ∂w ∂w
Table of contents 161

The last step is evident for velocity independent forces, but also holds for the magnetic part
of the Lorentz force, since
q q q
hw′ · bi = hw′ · (w × B)i = hw′ · (w′ × B)i + hw′ i · (v × B) = 0.
m m m
Again, if the interactions between any two particles conserve their total (kinetic) energy,
1
2
∆w′ 2 vanishes and we get

1∂ 2 1 ∂ 2 2 ∂
(ρhw′ i) + · (ρ vhw′ i + ρhw′ w′ i) = T : v.
2 ∂t 2 ∂x ∂x
Upon expansion of this equation, we have
   
1 ′ 2 ∂ρ ∂ 1 ∂ ∂ 2 1 ∂ 2 ∂
hw i + · (ρ v) + ρ +v· hw′ i + (ρhw′ w′ i) = T : v,
2 ∂t ∂x 2 ∂t ∂x 2 ∂x ∂x

and some terms are found to cancel due to the continuity equation. There only remains
1 d 2 1 ∂ 2
ρ hw′ i = T : D − · (ρhw′ w′ i),
2 dt 2 ∂x
where D is the symmetric part of (∂/∂x)v.
If this energy balance is compared to the previous form

ρε̇ = ρν − ∇ · q + T : D,

the internal energy ε is found as


Z
1 2 1
ε = hw′ i = ϕ(w − v)2 dW
2 2ρ W

and the heat flow q as


Z
1 2 1
q = ρhw′ w′ i = ϕ(w − v)2 (w − v) dW.
2 2 W

The energy balance is thus



ρε̇ = T : D −· q.
∂x
Heat sources are not included here, because then ∆Q would no longer be zero for Q taken
as 21 w′ 2 .
Throughout the above derivations of the continuity equation, the equation of motion and
the energy balance by suitable averaging of the kinetic equation, it is seen that as soon
as the distribution function is known, all the averages (the mean velocity, the stress and
even the internal energy density or the heat flow) can be computed explicitly. But as said
already before, this requires the solution of the corresponding kinetic equation and that is
quite another mathematical problem!

Anda mungkin juga menyukai