Introduction To Vector and Tensor Analysis: Jesper Ferkinghoff-Borg September 6, 2007

Download as pdf or txt
Download as pdf or txt
You are on page 1of 61

Introduction to vector and tensor analysis

Jesper Ferkinghoff-Borg

September 6, 2007
Contents

1 Physical space 3
1.1 Coordinate systems . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Scalars and vectors 5


2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Basic vector algebra . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Scalar product . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Cross product . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Coordinate systems and bases . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Components and notation . . . . . . . . . . . . . . . . . . 9
2.3.2 Triplet algebra . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Orthonormal bases . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 Scalar product . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.2 Cross product . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Ordinary derivatives and integrals of vectors . . . . . . . . . . . . 13
2.5.1 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.2 Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.2 Partial derivatives . . . . . . . . . . . . . . . . . . . . . . 15
2.6.3 Differentials . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.4 Gradient, divergence and curl . . . . . . . . . . . . . . . . 17
2.6.5 Line, surface and volume integrals . . . . . . . . . . . . . 20
2.6.6 Integral theorems . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Curvilinear coordinates . . . . . . . . . . . . . . . . . . . . . . . 26
2.7.1 Cylindrical coordinates . . . . . . . . . . . . . . . . . . . 27
2.7.2 Spherical coordinates . . . . . . . . . . . . . . . . . . . . . 28

3 Tensors 30
3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Outer product . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Basic tensor algebra . . . . . . . . . . . . . . . . . . . . . . . . . 31

1
3.3.1 Transposed tensors . . . . . . . . . . . . . . . . . . . . . . 32
3.3.2 Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.3 Special tensors . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Tensor components in orthonormal bases . . . . . . . . . . . . . . 34
3.4.1 Matrix algebra . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4.2 Two-point components . . . . . . . . . . . . . . . . . . . . 38
3.5 Tensor fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.1 Gradient, divergence and curl . . . . . . . . . . . . . . . . 39
3.5.2 Integral theorems . . . . . . . . . . . . . . . . . . . . . . . 40

4 Tensor calculus 42
4.1 Tensor notation and Einsteins summation rule . . . . . . . . . . 43
4.2 Orthonormal basis transformation . . . . . . . . . . . . . . . . . 44
4.2.1 Cartesian coordinate transformation . . . . . . . . . . . . 45
4.2.2 The orthogonal group . . . . . . . . . . . . . . . . . . . . 45
4.2.3 Algebraic invariance . . . . . . . . . . . . . . . . . . . . . 46
4.2.4 Active and passive transformation . . . . . . . . . . . . . 47
4.2.5 Summary on scalars, vectors and tensors . . . . . . . . . . 48
4.3 Tensors of any rank . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.3 Basic algebra . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.4 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Reflection and pseudotensors . . . . . . . . . . . . . . . . . . . . 54
4.4.1 Levi-Civita symbol . . . . . . . . . . . . . . . . . . . . . . 55
4.4.2 Manipulation of the Levi-Civita symbol . . . . . . . . . . 56
4.5 Tensor fields of any rank . . . . . . . . . . . . . . . . . . . . . . . 57

5 Invariance and symmetries 60

2
Chapter 1

Physical space

1.1 Coordinate systems


In order to consider mechanical -or any other physical- phenomena it is neces-
sary to choose a frame of reference, which is a set of rules for ascribing numbers
to physical objects. The physical space has three dimensions, which means that
it takes exactly three numbers -say x1 , x2 , x3 - to locate a point, P , in space.
These numbers are called coordinates of a point, and the reference frame for the
coordinates is called the coordinate system C. The coordinates of P can con-
veniently be collected into a a triplet, (x)C (P ) = (x1 (P ), x2 (P ), x3 (P ))C , called
the position of the point. Here, we use an underline simply as a short hand no-
tation for a triplet of quantities. The same point will have a different position
(x0 )C 0 (P ) = (x01 (P ), x02 (P ), x03 (P ))C 0 in a different coordinate system C 0 . How-
ever, since difference coordinate systems are just different ways of representing
the same physical space in terms of real numbers, the coordinates of P in C 0
must be a unique function of its coordinates in C. This functional relation can
be expressed as
x0 = x0 (x) (1.1)
which is a shorthand notation for
x01 = x01 (x1 , x2 , x3 )
x02 = x02 (x1 , x2 , x3 ) (1.2)
x03 = x03 (x1 , x2 , x3 )

This functional relation is called a coordinate transformation. The inverse trans-


formation must also exist, because both systems are assumed to cover the same
space.

1.2 Distances
.

3
We will take it for given that one can ascribe a unique value (ie. indepen-
dent of the chosen coordinate system), D(O, P ), for the distance between two
arbitraily chosen points O and P . Also we assume (or we may take it as an
observational fact) that when D becomes sufficiently small, a coordinate system
exists by which D can be calculated according to Pythagoras law:
v
u d
uX
D(O, P ) = t (xi (P ) − xi (O))2 , Cartesian coordinate system, (1.3)
i=1

where d = 3 is the dimension.


We will call such a coordinate system a Cartesian or a Rectangular coordinate
system. Without loss of generality we can choose one of the points as origen,
-say O-, so x(O) = (0, 0, 0). The neighborhood, Ω(O), around O for which
distances between any two points, P1 , P2 ∈ Ω(O) can be calculated according
to Eq. (1.3) for the same fixed coordinate system is coined a (local) Euklidean
space. In an Euklidean space, we also write |P1 P2 | for D(P1 , P2 ), representing
the length of the line segment connecting P1 and P2 . Trivially, O ∈ Ω(O).
Classical physics takes place in a 3-dimensional globally Euclidean space
Ω(O) = R3 . In general relativity space are intrinsically curved and the as-
sumption of an Euklidean space can only be applied locally. The principle of
curved space is easier to envisage for 2d-surfaces. For instance all points on the
surface of earth can be represented by two numbers -say x1 , x2 - representing
the longitude and latitude. However, the law of Pythagoras (with d = 2) can
only be applied for small rectangular triangles1 on the surface, ie. locally. For
larger rectangular triangles the sum of the angles will be larger than 1800 and
Pythagoras’ law will not be correct. All geometric analysis, however, rely on
the assumption that at sufficently small scales the space will appear flat. In
differential geometry one only requires flatness in a differential sence.

1.3 Symmetries
For classical mechanical phenomena it is found that a frame of reference can
always be chosen in which space is homogeneous and isotropic and time is ho-
mogeneous. It implies that nature has no sence of an absolute origin in time
and space and no sence of an absolute direction. Mechanical or geometrical laws
should therefore be invariant to coordinate transformations involving rotations
and translations. Motivated by this fact we are inclined to develop a mathe-
matical formalism for operating with geometrical objects without referring to a
coordinate system or, equivalently, a formalism which will take the same form
(ie. is covariant) for all coordinates. This fundamental principle is called the
principle of covariance. The mathematics of scalar, vector and tensor algebra is
precisely such a formalism.

1 Small would mean that the length of line segments are much smaller than the radius of

earth

4
Chapter 2

Scalars and vectors

2.1 Definitions
A vector is a quantity having both magnitude and a direction in space, such as
displacement, velocity, force and acceleration.
Graphically a vector is represented by an arrow OP from a point O to a
point P , defining the direction and the magnitude of the vector being indicated
by the length of the arrow. Here, O is called the initial point and P is called the
terminal point. Analytically, the vector is represented by either OP~ or OP and
~
the magnitude by |OP | or |OP|. We shall use the bold face notation in these
notes. In this chapter will assume that all points P belong to an Euklidean
space, P ∈ Ω(O), meaning that lengths of line segments can be calculated
according to Pythagoras.
A scalar is a quantity having magnitude but no direction, e.g. mass, length,
time, temperature and any real number.
We indicate scalars by letters of ordinary types. For example, a vector a will
have length a = |a|. Operations with scalars follow the same rules as elementary
algebra; so multiplication, addition and substraction (provided the scalars have
same units) follow the usual algebraic rules.

2.2 Basic vector algebra


The operations defined for real numbers are, with suitable definitions, capable
of extension to an algebra of vectors. The following definitions are fundamental
and define the basic algebraic rules of vectors:

1. Two vectors a and b are equal if they have the same magnitude and
direction regardless of the position of their initial point.

2. A vector having direction opposite of a vector a but having the same


magnitude is denoted −a.

5
3. The sum of resultant of vectors a and b is a vector c formed by placing
the initial point of b on the terminal point of a and then joining the initial
point of a to the terminal point of b. The sum is written c = a + b.
4. The difference between two vectors, a and b, represented by a − b is
defined as the sum a + (−b).
5. The product of a vector a with a scalar m is a vector ma with magnitude
|m| times the magnitude of a, ie |ma| = |m||a|, and with direction the
same as or opposite to that of a, according as m is positive or negative.

We stress that these definitions for vector addition, substraction and scalar
multiplications are defined geometrically, ie. they have no reference to coordi-
nates. It then becomes a pure geometric exercise to prove the following laws:

a+b =b+a Commutative law for addition


a + (b + c) = (a + b) + c Associate law for addition
m(na) = (mn)a Associate law for multiplication
(2.1)
(m + n)a = ma + na Distributive law
m(a + b) = ma + nb Distributive law
am =def am Commutative law for multiplication

Here, the last formula should just be read as a definition of a vector times a
scalar. Note that in all cases only multiplication of a vector by one or more
scalars are defined. One can define different types of bilinear vector products.
The three basic types are called scalar product (or inner product), cross product
and outer product (or tensor product). We shall define each in turn. The
definition of the outer product is postponed to chapter 3.
Following definition will become useful:

A unit vector is a vector having unit magnitude. If a is not a null vector


then a/|a| is a unit vector having the same direction as a.

2.2.1 Scalar product


The scalar product between two vectors, a and b is defined by

a · b = ab cos(θ), 0≤θ≤π (2.2)

where a = |a|, b = |b| and θ is the angle between the two vectors. Note that
a · b is a scalar. Following rules apply:

a·b =b·a Commutative law for scalar product


(a + b) · c =a·c+b·c Distributive law in 1. argument
a · (b + c) =a·b+a·c Distributive law for 2. argment
m(a · b) = (ma) · b = a · (mb) m is a scalar
(2.3)

6
The three last formulas make the scalar product bilinear in the two arguments.
Note also,

1 a · a = |a|2
2 If a · b = 0 and a and b are not null vectors,
then a and b are perpendicular. (2.4)
3 The projection of a vector a on b is equal to a · eb ,
where eb = b/|b| is the unit vector in direction of b.

2.2.2 Cross product


The cross product, a × b between two vectors a and b is a vector defined by

a × b = ab sin(θ)u, 0 ≤ θ ≤ π, (2.5)

where θ is the angle between a and b and u is a unit vector in the direction
perpendicular to the plane of a and b such that a, b and u form a right-handed
system1 .
The following laws are valid:

a×b = −b × a Cross product is not commutative.


(a + b) × c =a×c+b×c Distributive law for 1. argument
a × (b + c) =a×b+a×c Distributive law for 2. argument
m(a × b) = (ma) × b = a × (mb) m is a scalar
(2.6)
The last three formulas make the cross product bilinear.
Note that:
1. The absolute value of the cross product |a × b| has a particular geometric
meaning. It equals the area of the parallelogram spanned by a and b.
2. The absolute value of the triple product a·(b×c) has a particular geometric
meaning. It equals the volume of the parallelepiped spanned by a, b and
c

2.3 Coordinate systems and bases


We emphasize again that all definitions and laws of vector algebra, as introduced
above, are invariant to the choise of coordinate system2 . Once we introduce a
way of ascribing positions to points by the choise of a coordinate system, C, we
obtain a way of representing vectors in terms of triplets.
1 It is seen that the definition of the cross-product explicitly depends on an arbitrary choise

of handedness. A vector whose direction depends on the choise of handedness is called an


axial vector or pseudovector as opposed to ordinary or polar vectors, whose directions are
independent of the choise of handedness.
2 With the one important exception that we have assumed the absolute notion of handedness

in the definition of the cross product.

7
By assumption, we can choose this coordinate system to be rectangular
(R), CR , cf. chapter 1. By selecting four points O, P1 , P2 and P3 having the
coordinates

xCR (O) = (0, 0, 0), xCR (P1 ) = (1, 0, 0), xCR (P2 ) = (0, 1, 0), xCR (P3 ) = (0, 0, 1)
(2.7)
we can construct three vectors

e1 = OP1 , e2 = OP2 , e3 = OP3 . (2.8)

Furthermore, from the choise of an origin O, a position vector, r = OP, can be


associated to any point P by the formula
X
r(P ) = r(xCR (P )) = xi (P )ei CR rectangular. (2.9)
i

The position vector is an improper vector, meaning that it explicitly depends on


the position of its initial point O. It will therefore change upon a transformation
to a coordinate system with a different origin. However, we shall see that all
physical laws only involve positional displacements,

∆r = r(P2 ) − r(P1 ) = OP2 − OP1 = P1 P2 ,

which is a vector invariant to the choise of the coordinate system.


Any point will have one and only one associated position vector by Eq. (2.9).
Thus, the three vectors R = {e1 , e2 , e3 } form a basis for the vector space3 . In
any coordinate system, C, a well-defined procedure exists for constructing the
associated basis, G, of the vector space4 . However, there are two definining
characteristica for a rectangular system
1. The basis is spatially constant, ie. the basis vectors are not themselves a
function of position.
2. The basis is orthonormal, meaning that the basis vectors are mutually
orthogonal, ei · ej = 0 for i 6= j, and unitary |ei | = 1.
The basis G = {g1 , g2 , g3 } associated with a general coordinate system, C, may
satisfy none of the above. There will always be a one to one relation between the
coordinates and the position vector in geometrical problems of classical physics,
but it is only for bases satisfying 1, that this relation will be on the form of Eq.
(2.9) with the replacement ei → gi . In thes cases, G and the choise of origin,
O, fully specifies the coordinate system, C = (O, G).
3 Recall that the definition of a basis is a set of vectors, G, that span the vector space

and are linear independent. A set of vectors will be a basis if and only if the triple product
g1 · (g2 × g3 ) 6= 0.
4 The simple procedure of Eq. (2.7) and (2.8) for obtaining the basis vectors is not generally

applicable. We shall illustrate the general method in section (2.7).

8
In the following we shall retain the notation E = {e1 , e2 , e3 } for the par-
ticular case of an orthonormal basis5 . If ei is non-constant we explicitly write
ei (x). On the other hand, if E is constant the coordinate system is rectangular
and we use the letter R for the basis.

2.3.1 Components and notation


From the definition of any basis, G, spatially dependent or not, an abitrary
vector a will have a unique triplet of quantities a = (a1 , a2 , a3 ) ∈ R3 with
respect to G such that X
a= a i gi . (2.10)
i

These quantities are called the components of a with respect to G. Mathemat-


ically speaking, we obtain a one-to-one map φG : V → R3 between the vector
space, V , and the triplet of scalars R3 , with the choise of a basis. It is useful
to have a short-hand notation for this map (φ) and its inverse (φ−1 ), ie. for the
resolvent of a vector into its components and for the vector obtained by expand-
ing a set of components along the basis vectors. We define the the following two
bracket functions for the two operations respectively:
[a]G =def. φG (a) = a ∈ R3
−1 P
(a)G =def. φG (a) = i ai gi = a ∈ V

The notation for the set of components of a vector a, a = [a]G , also mean that
we shall refer to the i0 ’th component as ai = [a]G,i . Furthermore, when it is
obvious from the context (or irrelevant) which basis is implied, we will omit the
basis subscript and use the notation a = [a] and ai = [a]i . For the collection of
components, ai , into a triplet, ordinary brackets are commonly used, so

(ai ) =def. (a1 , a2 , a3 ) = a.

In order not to confuse this notation with the vector obtained from a set of
components X
(a)G =def. a i gi ,
i

whe shall always retain the basis subscript in the latter expression.
Trivially, we have

g1 = (1, 0, 0)G , g2 = (0, 1, 0)G , g3 = (0, 0, 1)G

or in the more dense, index-based notation6

[gi ]j = δij , wrt. basis G = {g1 , g2 , g3 }


5 This basis need not be spatially independent. In the case of a constant orthonormal basis,

ie. a Cartesian basis, the notation {i, j, k} is also often used. However, we shall retain the
other notation for algebraic convenience.
6 Expressing vector identities in terms of their components is also referred to as tensor

notation.

9
where δij is the Kronecker delta

1 if i = j
δij = (2.11)
0 if i 6= j

Furthermore, if G is constant in space the relation between coordinates and the


position vector is on the form of Eq. (2.9)
X
r(x) = (x)G = xi gi ⇔ [r]G,i = xi G const. (2.12)
i

With the risk of being pedantic let us stress the difference between a vector and
its components. For two different bases G 6= G 0 , a vector a will be represented
by the components a ∈ R3 and a0 ∈ R3 respectively, so that

a = (a)G = (a0 )G 0 but a 6= a0 .

Though representing the same vector, the two sets of components a and a0 will
differ because they refer to different bases. The distinction between a vector
and its components becomes irrelevant in the case where one basis is involved
only.

2.3.2 Triplet algebra


By choosing a basis G for the vector space, the basic algebraic vector laws of
addition, substraction and scalar multiplication, Eq. (2.1), can be expressed in
terms of normal triplet-algebra for the components. For instance, since
X X X
a+b= a i gi + b i gi = (ai + bi )gi
i i i

we have

a + b = (a1 , a2 , a3 )G + (b1 , b2 , b3 )G = (a1 + b1 , a2 + b2 , a3 + b3 )G ,

or equivalently (for all i)


[a + b]i = [a]i + [b]i (2.13)
The same naturally holds for vector substraction. Also, since
X X
ma = m( a i gi ) = mai gi ,
i i

for a scalar quantity m, we have

ma = m(a1 , a2 , a3 )G = (ma1 , ma2 , ma3 )G ,

or
[ma]i = m[a]i . (2.14)

10
The bilinear scalar and cross product are not as easy to operate with in
terms of the components in an arbitrary basis G as in the case of vector addition,
substraction and scalar multiplication. For instance
a · b = (a1 , a2 , a3 )G· (b1 , b2 ,b3 )G
P P
= ( i a i gi ) · j b j gj
P (2.15)
= ij ai bj gi · gj
P
= ij ai bj gij , gij = gi · gj
The quantities gij , are called the metric coefficients. If G is spatially dependent
then gij = gij (x) will be functions of the coordinates. In fact,the functional form
of the metric coefficients fully specifies the type of geometry involved (Euklidean
or not) and they are the starting point for extending vector calculus to arbitrary
geometries. Here it suffices to say that a Cartesian coordinate system uniquely
implies that the metric coefficients are constant and satisfy gij = δij .

2.4 Orthonormal bases


An orthonormal basis, E, satisfies by definition
ei · ej = δij . (2.16)
This relation is particular handy for obtaining the components of a vector a.
Indeed, we have
P
a = i ai ei ⇒ ai = a · ei , E ortonormal. (2.17)
So we obtain the i’th component of a vector a by taking the scalar product
between a and the i’th basis-vector.

2.4.1 Scalar product


In an ortonormal basis the scalar product, Eq. (2.15), is a simple expression in
terms of the components
X X X
a·b=( ai ei ) · ( bj e j ) = a i bi . (2.18)
i j i

Furthermore, if the orthonormal basis is spatially independent, the displacement


vector, PQ, between two points P and Q with coordinates p = (p1 , p2 , p3 ) and
q = (q1 , q2 , q3 ) respectively, will be
X
PQ = r(Q) − r(P ) = (qi − pi )ei
i

The distance between P and Q is given by the length of |PQ| and


X sX
2 2
|PQ| = (qi − pi ) ⇒ |PQ| = (qi − pi )2 . (2.19)
i i

11
In other words, we recover the law of Pythagoras, cf. Eq. (1.3), which is
reassuring since a constant orthonormal basis is equivalent to the choise of a
rectangular coordinate system.

2.4.2 Cross product


Let E be an ortonormal right-handed basis, so e3 = e1 × e2 . Using the laws of
cross-product, Eq. (2.6), one can show
 
e1 e2 e3
a × b = det  a1 a2 a3  = (a2 b3 − a3 b2 , a3 b1 − a1 b3 , a1 b2 − a2 b1 )E (2.20)
b1 b2 b3

where det() denotes the determinant of the array considered as a matrix. An


alternative expression defined only in terms of the components, is
X
[a × b]i = ijk aj bk (2.21)
jk

The symbol ijk with the three indices is the Levi-Civita symbol. It consists of
3 × 3 × 3 = 27 real numbers given by

123 = 231 = 312 = +1


321 = 132 = 213 = −1 (2.22)
ijk = 0 otherwise

In words, ijk is antisymmetric in all indices. From this it follows that the
ortonormal basis vectors satisfy
X
ei × e j = ijk ek , (2.23)
k

i.e. e3 = e1 × e2 , e1 = e2 × e3 , and so on.


Applying eq. (2.3) and (2.6) one can show that the triple product a · (b × c)
in an ortonormal right-handed basis is given by
 
X a1 a2 a3
a · (b × c) = ijk ai bj ck = det  b1 b2 b3  (2.24)
ijk c1 c2 c3

The problem of handedness


Let us emphasize that Eq. (2.20,2.21,2.23,2.24) would have been the same had
we defined the cross-product by a left-hand rule and expressed the components
in a left-handed coordinate system. Indeed, the definition of the Levi-Civita
symbol, Eq. (2.22), does not itself depend on the handedness of the coordinate
system. Since it is in fact not possible to give an absolute prescribtion of how to
construct a right-handed (or left-handed) coordinate system it is more common

12
in modern vector analysis to define the cross product through the determinant,
Eq. (2.20) og equivalently through the Levi-Civita symbol, Eq. (2.21). In
replacing the geometric definition, Eq. (2.5) which pressumes an absolute notion
of handedness, with an algebraic one the direction of the cross-product will
simply be determined by what-ever convention of handedness that has been
adopted with the ortonormal coordinate system in the first place.
Physically, the arbitrariness in the choise of handedness implies that the
orientation of the cross product of two ordinary vectors is not a physical ob-
jective quantity, in contrast to vectors representing displacements, velocities,
forces etc. One therefore distinguishes between proper or polar vectors whose
direction are independent on the choise of handedness and axial or pseudovectors
whose directions depend on the choise of handedness. The distinction between
the two types of vectors becomes important when one considers transformations
between left- and right-handed coordinate systems.

Dimensionality
For all vector operations the only explicit reference to dimension of the physical
space appears in the cross product. This can be seen from the definition of
the -symbol, which explicitly has three indices running from 1 to 3. All other
vector operations are valid in arbitrary dimensions.
The generalization of the cross product to any dimension d is obtained by
constructing a Levi-Civita symbol having d indices each running from 1 to d
and being totally antisymmetric.
For instance for d = 2 :
12 =1
21 = −1 (2.25)
11 = 22 = 0

Here, we get the “cross-product”, â, of a vector a = (a1 , a2 )E by


X
[â]j = ji ai
i

or in doublet notation

â = (â1 , â2 )E = (a2 , −a1 )E ,

which is the familiar expression in d = 2 for obtaining a vector orthogonal to a


given vector. Generally, in d > 1 dimensions the cross product will take d − 1
vectors as arguments.

2.5 Ordinary derivatives and integrals of vectors


Let c(t) be a vector depending on a single scalar variable t. In this section we
will consider how to define derivatives and integrals of c with respect to t.

13
2.5.1 Derivatives
The definition for ordinary derivatives of functions can directly be extended to
vector functions of scalars:
dc(t) c(u + ∆u) − c(u)
= lim . (2.26)
du ∆u→0 ∆u
Note that the definition involves substracting two vectors, ∆c = c(u+∆u)−c(u)
which is a vector. Therefore, dr(t)
du will also be a vector provided the limit, Eq.
(2.26), exists. Eq. (2.26) is defined independently of coordinate systems. In an
orthonormal basis independent of u:
d X dai
a= ei , (2.27)
du i
du

whence the derivative of a vector is reduced to ordinary derivatives of the scalar


functions ai .
If u represents the time u = t and c(t) is the position vector of a particle r(t)
with respect to some coordinate system then dr(t)dt will represent the velocity v(t).
The velocity will be a tangentiel vector to the space curve mapped by r(t). Even
though the position vector itself is not a proper vector due to its dependence on
the choise of origin of the coordinate system, cf. section 2.3, the velocity is a
proper vector since it is defined through the positional displacement, ∆r, which
is a proper vector. Consequently, the acceleration a(t) =def. dv dt is also a proper
vector.
Ordinary differentiation of a vector co-exists nicely with the basic vector
operations, e.g.
d da
(a + b) = du + db
du (2.28)
du
d da
(a · b) = du · b + a · db
du (2.29)
du
d da
(a × b) = du × b + a × dbdu (2.30)
du
d dφ
(φa) = du a + φ da
du (2.31)
du
where φ(u) is a normal scalar function. The above formulas can of course be
expressed in terms of the vector components, ai and bi , using Eq. (2.27).

2.5.2 Integrals
The definition of ordinary integral of a vector depending on a single scalar
variable, c(u), also follows that of ordinary calculus. With respect to a basis E
independent of u the indefinite integral of c(u) is defined by
Z XZ
c(u)du = ci (u)duei ,
i

14
R
where ci (u)du is the indefinite integral of an ordinary scalar function. If s(u)
d
is a vector satisfying c(u) = du s(u) then
d
Z Z
c(u)du = (s(u)) du = s(u) + k
du
where k is an arbitrary constant vector independent of u. The definite integral
between two limits u = u0 and u = u1 can in such case be written
Z u1 Z u1
d u
c(u)du = (s(u))du = [s(u) + k]u10 = s(u1 ) − s(u0 ).
u0 u0 du
This integral can also be defined as a limit of a sum in a manner analogous to
that of elementary integral calculus.

2.6 Fields
2.6.1 Definition
In continuous systems the basic physical variables are distributed over space.
A function of space is known as a field. Let an arbitrary coordinate system be
given.
Scalar field. If to each position x = (x1 , x2 , x3 ) of a region in space the
corresponds a number or scalar φ(x1 , x2 , x3 ), then φ is called a scalar function of
position or scalar field. Physical examples of scalar fields are the mass or charge
density distribution of an object, the temperature or the pressure distribution
at a given time in a fluid.
Vector field. If to each position x = (x1 , x2 , x3 ) of a region in space there
corresponds a vector a(x1 , x2 , x3 ) then a is called a vector function of position
or a vector field. Physical examples of vector fields are the gravitational field
around the earth, the velocity field of a moving fluid, electro-magnetic field of
charged particle systems.
In the following we shall assume the choise of a rectangular
P coordinate system
C = (O, R). Introducing the position vector r(x) = i xi ei the expressions φ(r)
and a(r) is taken to have the same meaning as φ(x) and a(x), respectively.

2.6.2 Partial derivatives


Let a be a vector field a(r) = a(x1 , x2 , x3 ). We can define the partial derivative
∂a
∂xi in the same fashion as in Eq. (2.26). Here we will use the short hand
notation:

∂i = ∂x i
2
2
∂ij = ∂x∂i ∂xj
In complete analogy to the usual definition of partial derivatives of a scalar
function, the partial derivative of a vector field with respect to xi (i = 1, 2, 3) is
a(r + ∆xi ei ) − a(r)
(∂i a)(r) = lim (2.32)
∆xi →0 ∆xi

15
Equation (2.32) expresses how the vector field changes in the spatial direction
of ei . For the same arguments as listed in previous section, ∂i a will for each i
also be a vector field. Note that in general, ∂j a 6= ∂i a when j 6= i.
Rules for partial diffentiation of vectors are similar to those used in elemen-
tary calculus for scalar functions Thus if a and b are functions of x then
∂i (a · b) = a · (∂i b) + (∂i a) · b
∂i (a × b) = (∂i a) × b + a × (∂i b)
2
∂ji (a · b) = ∂j (∂i (a · b)) = ∂j ((∂i a) · b + a · (∂i b))
2 2
= a · (∂ji b) + (∂j a) · (∂i b) + (∂i a) · (∂j b) + (∂ji a) · b

We can verify these expressions by resolving the vector fields into R = {e 1 , e2 , e3 }.


Then all differentiation is reduced to ordinary differention of the scalar func-
tions, ak (x) and bk (x). Notice that for the position vector

∂i r = e i . (2.33)

If the vector field is resolved into a non-cartesian coordinate system, C 0


X
a(x0 ) = aj (x0 )gj (x0 ) (2.34)
j

then one must remember to include the spatial derivatives of the basis vectors
as well
∂a 0 X ∂aj
0 0 0 ∂gj 0

(x ) = (x )g j (x ) + a j (x ) (x ) (2.35)
∂x0i j
∂x0i ∂x0i

2.6.3 Differentials
Since partial derivatives of vector field follow those used in elementary calculus
for scalar functions the same will be true for vector differentials. For example,
for C = (O, R)
X X
if a(x) = ai (x)ei , then da = dai (x)ei (2.36)
i i

The change of the vector components due to a infitesimal change of coordinates


dxj is obtained by the chain rule
X
dai (x) = (∂j ai )(x)dxj (2.37)
j

In vector notation it reads


X X 
da(x) = (∂j ai )(x)dxj ei (2.38)
i j

or simply X
da = ∂i adxi . (2.39)
i

16
P
The position vector r = i xi ei is a special vector field for which Eq. (2.38)
implies X
dr = dxi ei . (2.40)
i

The square, (ds)2 of the infinitesimal distance moved is then given by


X
(ds)2 = dr · dr = dx2i
i

If r(u) is space curve parametrized by u then


 2
ds dr dr
= · ,
du du du

Therefore, the arc length between two points on the curve r(u) given by u = u1
and u = u2 is
Z u2 r
dr dr
s= · du.
u1 du du
In general, we have (irrespective of the choise of basis)

d(a · b) = da · b + a · db
d(a × b) = da × b + a × db

2.6.4 Gradient, divergence and curl


Let a rectangular coordinate system be given C = (O, R).

Nabla operator ∇
The vector differential operator del or nabla written as ∇ is defined by
X
∇(·) = ei ∂i (·) (2.41)
i

where · represents a scalar or –as we shall see later– a vector field. Notice that
the i’th component of the operator is given by

∇i = e i · ∇ = ∂ i .

This vector operator possesses properties analogous to those of ordinary vectors


which is not obvious at first sight, since its operation is defined relative to
the choise of the coordinate system, in contrast to the vector algebra defined
hitherto. We shall postpone the demonstration of its vectorial nature until
section 4.5, where we have become more familiar with tensor calculus.
The gradient of a scalar field φ(r) is defined by inserting a scalar field in the
above definition: X
grad φ = ∇φ = (∂i φ)ei
i

17
The gradient of a scalar field has some interesting geometrical properties. Con-
sider the change of φ in some particular direction. For an infinitesimal vector
displacement, dr, forming its scalar product with ∇φ we have
!  
X X X
(∇φ)(r) · dr = (∂i φ)(r)ei ·  dxj ej  = (∂i φ)(r)dxi = dφ,
i j i

which is the infinitesimal change in φ in going from from position r to r + dr.


In particular, if r depends on some parameter t such that r(t) defines a space
curve then the total derivative of φ with respect to that curve is simply

= ∇φ · dr (2.42)
dt
dr
Setting v = dt we obtain from the definition of the scalar product

= |∇φ||v| cos(θ) (2.43)
dt
where θ is the angle between ∇φ and v. This also shows that the largest increase
of the scalar field is obtained in the direction ∇φ and that, if v is a unit vector
the rate of change in this case equals |∇φ|.
We can extend the above analysis to find the rate of change of a vector field
in a particular direction v = dr dt .

d P dai (r(t))
dt a(r(t)) = ei
Pi Pdt 
= i j j ai )(r))vj ei
(∂
P P
= j vj ∂j ( i ai (r)ei ) (2.44)
P
= j vj ∂j a(r)
= (v · ∇) a(r)
This shows that the operator
X
v·∇= vi ∂i (2.45)
i

gives the rate of change in direction of v of the quantity (vector or scalar) on


which it acts.
A second interesting geometrical property of ∇φ may be found by considering
the surface defined by φ(r) = c, where c is some constant. If r(t) is a space
curve in the surface we clearly have dφ
dt = 0 and consequently for any tangent
vector v in the plane we have ∇φ · v = 0, according to Eq. (2.42). In other
words, ∇φ is a vector normal to the surface φ(r) = c at every point.

Divergence of a vector field


The divergence of a vector field a(r) is defined by
X
(div a)(r) = (∇ · a)(r) = (∂i ai )(r) (2.46)
i

18
The full physical and geometrical meaning of the divergence is discussed in next
section. Clearly (∇ · a)(r) is a scalar field. Now if some vector field a is itself
derived from a scalar field via a = ∇φ then ∇ · a has the form ∇ · ∇φ or, as it
is usually written ∇2 φ where
X  X ∂2 
∇2 = ∂ii2 =
i i
∂x2i

∇2 φ is called the laplacian of φ and is typically encountered in electrostatic


problems or in diffusion equations of physical scalar field such as a tempera-
ture or density distribution. The laplacian can also act on vector through the
components X XX
∇2 a = ∂ii2 a = (∂ii2 aj )ej
i j i

Curl of a vector field


The curl of a vector field a(r) is defined by
  
curl (a) = ∇ × a = ∂2 a3 − ∂3 a2 e1 + ∂3 a1 − ∂1 a3 e2 + ∂1 a2 − ∂2 a1 e3 ,

In analogy to the definition of cross-product between two ordinary vectors we


can express the definition symbolically
 
e1 e2 e3
∇ × a = det  ∂1 ∂2 ∂3  ,
a1 a2 a3

where it is understood that, on expanding the determinant, the partial deriva-


tives act on the components of a. The tensor notation of this expression (for
the i’th component) is even more compact
X
[∇ × a]i = ijk ∂j ak (2.47)
jk

Clearly, ∇ × a is itself a vector field.


For a vector field v(x) describing the local velocity at any point in a fluid,
∇ × v is a measure of the angular velocity of the fluid in the neighbourhood of
that point. If a small paddle wheel were placed at various points in the fluid
then it would tend to rotate in regfions where ∇ × v 6= 0, while it would not
rotate in regions where ∇ × v = 0.
Another insight into the physical interpretation of the curl operator is gained
by considering the vector field v describing the velocity at any point in a regied
body rotation about some axis with angular velocity ω. If r is the position
vector of the point with respect to some origin on the axis of rotation then the
velocity field would be given by v(x) = ω × r(x). The curl of the vector field is
then found to be ∇ × v = 2ω.

19
Combinations of grad,div and curl
There are myriad of identities between various combinations of the three impor-
tant vector operators grad, div and curl. The identities involving cross products
are more easily proven using tensor calculus which is postponed for chapter 4.
Here we simply list the most important identities:

∇(φ + ψ) = ∇φ + ∇ψ
∇ · (a + b) =∇·a+∇·b
∇ × (a + b) = ∇ × a + ∇ × b
∇ · (φa) = φ∇ · a + a · ∇φ
(2.48)
∇ · (a × b) = b · (∇ × a) − a · (∇ × b)
∇ × (∇ × a) = ∇(∇ · a) − ∇2 a
∇ · (∇ × a) =0
∇ × ∇φ =0

Here, φ and ψ are scalar fields and a and c are vector fields. The last identity
has an important meaning. If a is derived from the gradient of some scalar field,
a = ∇φ then the identity shows that a is necessarily irrotational ,∇ × a = 0.
We shall return to this point in the next section.

2.6.5 Line, surface and volume integrals


In the previous section we have discussed continuously varying scalar and vector
fields and discussed the action of various differential operators on the. Often,
the need arises to consider the integration of field quantities along lines, over
surfaces and throughout volumes. In general the integrand may be scalar or
vector, but the evaluation of such intregrals involves their reduction to one or
more scalar integrals, which are then evaluated. This procedure is equivalent to
the way one in practice operates with differential operators on scalar and vector
fields.

Line integrals
In general, one may encounter line integrals of the forms
Z Z Z
φdr, a · dr, a × dr, (2.49)
C C C

where φ is a scalar field, a is a vector field and C is a prescribed curve in space


joining to given points A and B in space. The three integrals themselves are
respectively vector, scalar and vector in nature.
The formal definition of a line integral closely follow that of ordinary integrals
and can be considered as the limit of a sum. We may divide the path joining
the points A and B into N small line elements ∆rp , p = 1, · · · , N . If xp =
(x1,p , x2,p , x3,p ) is any point on the line element ∆rp then the second type of

20
line integral in Eq. (2.49), for example, is defined as
Z XN
a · dr =def lim a(xp ) · ∆rp ,
C N →∞
p=

where all |∆rp | → 0 as N → ∞.


Each of the line integrals in Eq. (2.49) is evaluated over some curve C
that may be either open (A and B beingH distinct points) or closed (A and B
coincide). In the latter case, one writes C to indicate this. The curve C is a
space curve, c.f. section 2.5.1, most often defined in a parametric form. In a
cartesian coordinate system C = (O, R), it becomes
X
C : r(u) = xi (u)ei , u0 ≤ u ≤ u1 , and [r(u0 )] = x(A), [r(u1 )] = x(B)
i
(2.50)
The metod of evaluating a line integral is to reduce it to a set of scalar
integrals. It is usual to work in cartesian coordinates, in which case, dr =
P
i ei dxi . The three integrals in Eq. (2.49) then becomes
R R P P R 
C φ(r)dr = C φ(r) ( i ei dxi) = i C φ(r)dxi ei
(2.51)
R R P P  P R
C
a(r) · dr = C ( i ai (r)ei ) · j ej dxj = i C ai (r)dxi

and
R R P P 
a(r) × dr = ( i a i (r)) × e j dx j
C RC R j  R R 
= CR a 2 (r)dx 3 − CRa3 (r)dx2 e1 + C a3 (r)dx1 − C a1 (r)dx3 e2
+ C a1 (r)dx2 − C a2 (r)dx1 e3
(2.52)
Note, that in the above we have used relations of the form
Z Z 
ai ej dxj = ai dxj ej ,
C C

which is allowable since the cartesian basis is independent of the coordinates.


If E = E(x) then the basis vectors could not be factorised out in the integral.
The final scalar integrals in Eq. (2.51) and Eq. (2.52) can be solved using the
parametrized form for C, Eq. (2.50). For instance
Z u1
dx3
Z
a2 (r)dx3 = a2 (x1 (u), x2 (u), x3 (u)) du.
C u0 du
In general, a line integral between two points will depend on theR specific path
C defining the integral. However, for line integrals of the form C a · dr there
exists a class of vector fields for which the line integral between two points
is independent of the path taken. Such vector fields are called conservative.
A vector field a that has continuous partial derivatives in a simply connected
region R 7 , is conservative if, and only if, any of the following is true
7 A simply connected region R is a region in space for which any closed path can be

continuously shrunk to a point, ie. R has no “holes”

21
RB
1. The integral AHa · dr, where A, B ∈ R, is independent of the path from
A to B. Hence C a · dr = 0 around any closed loop in R.
2. There exists a scalar field φ in R such that a = ∇φ.
3. ∇ × a = 0.
4. a · dr is an exact differential.
We will not demonstrate the equivalence of these statements. If a vector field
is conservative we can write

a · dr = ∇φ · dr = dφ

and Z B Z B Z B
a · dr = ∇φ · dr = dφ = φ(A) − φ(B).
A A A
This situtation is encountered whenever a = f represents the force f derived
from a potential (scalar) field, φ, such as the potential energy in a gravitational
field, the potential energy in an elastic spring, the voltage in a electrical circuits,
etc.

Surface integrals
As with line integrals, integrals over surfaces can involve vector and scalar fields
and, equally, result in either a vector or a scalar. We shall focus on surface
integrals of the form Z
a · dS. (2.53)
S
where a is a vector field and S is a surface in space which may be either open or
closed. Following
R the notation
H of line integrals, for surface integrals over a closed
surface S is replaced by S . The vector differential dS in Eq. (2.53) represents
a vector area element of the surface S. It may also be written dS = ndS where
n is a unit normal to the surface at the position of the element and dS is a
scalar area of the element. The convention for the direction of the normal n
to a surface depends on whether the suface is open or closed. For a closed
surface the direction of n is taken be outwards from the enclosed volume. An
open surface spans some perimeter curve C. The direction of n is then given
by the right-hand sense with respect to the direction in which the perimeter is
traversed, ie. it follows the right-hand screw rule.
The formal definition of a surface integral is very similar to that of a line
integral. One divides the surface into N elements of area ∆Sp , p = 1, · · · , N
each with a unit normal np . If xp is any point in ∆Sp then
Z N
X
a · dS = lim a(xp ) · np ∆Sp ,
S N →∞
p=1

22
where it is required that ∆Sp → 0 for N → ∞.
A standard way of evaluating surface integrals is to use cartesian coordinates
and project the surface onto one of the basis planes. For instance, suppose a
surface S has projection R onto the 12-plane (xy-plane), so that an element of
surface area dS projects onto the area element dA. Then
dA = |e3 · dS| = |e3 · ndS| = |e3 · n|dS.

Since in the 12-plane dA = dx1 dx2 we have the expression for the surface integral
dx1 dx2
Z Z Z
a(r) · dS = a(r) · ndS = a(r) · n
S R R |e3 · n|
Now, if the surface S is given by the equation x3 = z(x1 , x2 ), where z(x1 , x2 )
gives the third coordinate of the surface for each (x1 , x2 ) then the scalar field

f (x1 , x2 , x3 ) = x3 − z(x1 , x2 ) (2.54)

is identical zero on S. The unit normal at any point of the surface will be given
∇f
by n = |∇f | evaluated at that point, c.f. section 2.6.4. We then obtain

∇f dA dA dA
dS = ndS = = ∇f = ∇f = ∇f dx1 dx2 ,
|∇f | |n · e3 | |∇f · e3 | |∂3 f |
where the last identity follows from the fact that ∂3 f = 1 from Eq. (2.54). The
surface integral then becomes
Z Z
a(r) · dS = a(x1 , x2 , z(x1 , x2 )) · (∇f )(x1 , x2 )dx1 dx2 .
S R

which is an ordinary scalar integral in two variables x1 and x2 .

Volume integrals
Volume integrals are generally simpler than line or surface integrals since the
element of the volume dV is a scalar quantity. Volume integrals are most often
on the form Z Z
φ(r)dV a(r)dV (2.55)
V V
Clearly, the firs form results in a scalar, whereas the second one yields a vector.
Two closely related physical examples, one of each kind, are provided R by the
total mass M , of a fluid contained in a volume V , given byR M = V ρ(r)dV and
the total linear momentum of that same fluid, given by V ρ(r)v(r)dV , where
ρ(r) is the density field and v(r) is the velocity field of the fluid.
The evaluation of the first volume integral in Eq. (2.55) is an ordinary
multiple integral. The evaluation of the second type of volume integral follows
directly since we can resolve the vector field into cartesian coordinates
Z XZ 
a(r)dV = ai (r)dV ei .
V i V

23
Of course we could have written a in terms of the basis vectors of other coordi-
nate system (e.g. spherical coordinates) but since such basis vectors are not in
general constant, they cannot be taken out of the integrand.

2.6.6 Integral theorems


There are two important theorems relating surface and volume integrals of vec-
tor fields, the divergence theorem of Gauss and Stokes theorem.

Gauss theorem
Let a be a (differentiable) vector field, and V be any volume bounded by a
closed surface S. Then Gauss theorem states
Z I

∇ · a (r)dV = a(r) · dS. (2.56)
V S

, The proof goes as follows. Let C = (O, R) be a cartesian coordinate system,


and let B(r0 ) be a small box around r0 with its edges oriented along the direc-
tions of the basis vectors and with volume VB = ∆x1 ∆x2 ∆x3 . Each face can be
labelled as sk where s = ±1 and k = 1, 2, 3 indicates the orientation ofQthe face.
Thus the outward normal of face sk is nsk = sek . If the area Ak = i6=k ∆xi
of each face Fsk is small then we can approximate each surface integral
Z Z
a(r) · dS = a(r) · sek dS
Fsk Fsk

by evaluating the vector field a(r) in the center point of the face r0 + s ∆x
2 ek
k

 
∆xk
Z
a(r) · dS ≈ a r0 + s ek · sek Ak
Fsk 2

The surface integral on the rhs. of Eq. (2.56) for the box B(r0 ) then becomes

a(r) · dS ≈ sk a r0 + s ∆x
H P 
2 ek · sek Ak
k
B(r0 )  
= k a r0 + ∆x ∆xk
P 
2
k
e k − a r 0 − 2 e k · ek Ak
P  (2.57)

= k ak r0 + ∆x ∆xk
 
2 e k − a k r0 − 2 e k Ak
k

P
≈ k (∂kak )(r0 ) ∆xk Ak
= ∇ · a (r0 ) VB

Any volume V can be approximated by a set of small boxes, Bi , centered around


the points r0,i , where i = 1, · · · , N . For the lhs. of Eq. (2.56) we then have
Z X 
∇ · a(r)dV ≈ ∇ · a (r0,i ) VBi . (2.58)
V i

Adding the surface integrals of each of these boxes, the contribution from the
mutual interfaces vanishes (since the outward normal of the two adjacent boxes

24
point in opposite direction). Consequently, the only contributions comes from
the surface S of V .
I XI
a(r) · dS ≈ a(r) · dSi (2.59)
S i B(r0,i )

Since each term on the rhs. of Eq. (2.58) equals a corresponding term on the
rhs. of Eq. (2.59) the Gauss teorem is demonstrated.
Gauss theorem is often used in conjunction with following mathematical
theorem
d ∂
Z Z
φ(r, t)dV = φ(r, t)dV,
dt V V ∂t
where t is time and φ is a time dependent scalar field (The theorem works in
arbitrary spatial dimension). The two theorems are central in deriving partial
differential equations for dynamical systems, in particular the so called conti-
nuity equations, linking a flow field to the time changes of scalar field advected
by the flow. For instance if ρ(r, t) is a density and v(r, t), is the velocity field
of a fluid then the vector j(r) = ρ(r)v(r) gives the density current. It can then
by shown (try it) that under mass conservation then

∂ρ
+ ∇ · j = 0.
∂t
The divergence of a vector field therefore has the physical meaning of giving the
net “outflux” of a scalar advected by the field within an infinitesimal volume.

Stokes theorem
Stokes theorem states that if S is the “curl analogue” of the divergence theorem
and relates the integral of the curl of a vector field over an open surface S to the
line integral of the vector field around the perimeter C bounding the surface.
Z I

∇ × a (r) · dS = a(r) · dr (2.60)
S C

Following the same lines as for the derivation of the divergence theorem the
surface S can be divided into many small areas Si with boundaries Ci and unit
normals ni . For each small area one can show that
I
(∇ × a) · ni Si ≈ a · dr.
Ci

Summing over i one finds that on the rhs. all parts of all interior boundaries
that are not part of C are included twice, being traversed in opposite directions
on each occasion and thus cancelling each other. Only contributions from line
elements that are also part of C survive. If each Si is allows to to tend to zero,
Stokes theorem, Eq. (2.60), is obtained.

25
2.7 Curvilinear coordinates
The vector operators, ∇φ, ∇ · a and ∇ × a , we have discussed so far have
all been defined in terms of cartesian coordinates. In that respect we have
been more restricted than in the algebraic definitions of the analogous ordinary
scalar and cross product, Eq. (2.18) and Eq. (2.20) respectively. Here we only
assumed orthonormality of the basis. The reason is that the nabla operator
involves spatial derivatives which implies that one must account for the possible
non-constancy of E, c.f. Eq. (2.35).
Many systems possess some particular symmetry which makes other coordi-
nate systems more natural, notably cylindrical or spherical coordinates. These
coordinates are just two examples of what are called curvilinear coordinates.
Curvilinear coordinates refer to the general case in which the rectangular coor-
dinates x of any point can be expressed as functions of another set of coordinates
x0 , thus defining a coordinate transformation, x = x(x0 ) or x0 = x0 (x) 8 . Here
we shall discuss the algebraic form of the standard vector operators for the two
particular transformations into cylindrical or spherical coordinates. An impor-
tant feature of these coordinates is that although they lead to non-constant
bases, these bases retain the property of orthonormality.
The starting point for the disgression is to realize that a non-constant basis

E 0 (x0 ) = {e01 (x0 ), e02 (x0 ), e03 (x0 )},

implies that the basis vectors are vector fields, as opposed to the constant basis
R = {e1 , e2 , e3 } associated with cartesian coordinates. Any vector field, a(x 0 )
is then generally written as
X
a(x0 ) = a0j (x0 )e0j (x0 ) (2.61)
j

where a0j are the components of the vector field in the new basis, and x0 = x0 (x)
is the coordinate transformation. One notes, that the functional form of the
components a0j (x0 ) differ from the functional form of the cartesian components,
aj (x), because the same point, P will have two different numerical representa-
tions x0 (P ) and x(P ). For a scalar field one must have the identity

φ0 (x0 ) = φ(x),

where φ0 is the functional form of the scalar field in the primed coordinate
system and φ is the function of the scalar field with respect to the unprimed
system. Again, the two functional forms must be different because the same
point has different numerical representations. However, the value of the two
functions must be the same since x0 and x represent the same point in space.
8 Recall that the notation x0 = x0 (x) is a short hand notation for the three functional
relationships listed in Eq. (1.2).

26
2.7.1 Cylindrical coordinates
Cylindrical coordinates, x0 = (ρ, φ, z), are defined in terms of normal cartesian
coordinates x = (x1 , x2 , x3 ) = (x, y, z) by the coordinate transformation

x = ρ cos(φ)
y = ρ sin(φ), (2.62)
z = z,

The domain of variation is 0 ≤ ρ < ∞, 0 ≤ φ < 2π.


The position vector may be written as

r(x0 ) = ρ cos(φ)e1 + ρ sin(φ)e2 + ze3

Local basis
If we take the partial derivatives with respect to ρ, φ, z, c.f. Eq. (2.33), and
normalize we obtain
e01 (x0 ) = eρ (x0 ) = ∂ρ r = cos(φ)e1 + sin(φ)e2
e02 (x0 ) = eφ (x0 ) = ρ1 ∂φ r = − sin(φ)e1 + cos(φ)e2
e03 = e3

These three unit vectors, like the Cartesian unit vectors ei , form an orthonormal
basis at each point in space. An arbitrary vector field may therefore be resolved
in this basis

a(x0 ) = ar (x0 )er (x0 ) + aφ (x0 )eφ (x0 ) + az (x0 )ez

where the vector components

ar = a · e r , aφ = a · e φ , Vz = a · e z

are the projections of a on the local basis vectors.

Resolution of gradient
The derivatives after cylindrical coordinates are found by differentiation through
the Cartesian coordinates (chain rule)
∂x ∂y
∂ρ = ∂ρ ∂x + ∂ρ ∂y = cos(φ)∂x + sin(φ)∂y
∂x ∂y
∂φ = ∂φ ∂x + ∂φ ∂y = −ρ sin(φ)∂x + ρ cos(φ)∂y

FromP
these relations we can calculate the projections of the gradient operator
∇ = i ei ∂i on the cylindrical basis and we obtain

∇ρ = eρ · ∇ = ∂ρ
∇φ = eφ · ∇ = ρ1 ∂φ
∇z = ez · ∇ = ∂z

27
The resolution of the gradient in the two bases therefore becomes
1
∇ = e ρ ∇ρ + e φ ∇φ + e z ∇z = e ρ ∂ ρ + e φ ∂ φ + e z ∂ z
ρ
Together with the only non-vanishing derivatives of the basis vectors

∂φ eρ = eφ
∂φ eφ = −eρ

we have the necessary tools for calculating in cylindrical coordinates. Note, that
it is the vanishing derivatives of the basis vectors that lead to the simple form
of the vector operators in cartesian coordinates.

Laplacian
The laplacian in cylindrical coordinates takes the form

∇2 = ∇ · ∇ = (eρ ∇ρ + eφ ∇φ + ez ∇z ) · (eρ ∇ρ + eφ ∇φ + ez ∇z )

Using the linearity of the scalar product it can be rewritten to the form
X
e0i ∇0i · e0j ∇0j ,

where e01 = eρ , ∇01 = ∇ρ etc. The interpretation of each of these terms is

e0i ∇0i · e0j ∇0j = e0i · ∇0i (e0j ∇0j ) (2.63)

Applying the chain rule of differentiation and Eq. (2.7.1) one can then show
1 1 2
∇2 = ∂ρρ
2
+ ∂ρ + 2 ∂φφ 2
+ ∂zz
ρ ρ

2.7.2 Spherical coordinates


The treatment of spherical coordinates follow the same line as cylindrical coor-
dinates. The spherical coordinates are x0 = (r, φ, θ) and the coordinate trans-
formation is given by
x = r sin(θ) cos(φ)
y = r sin(θ) sin(φ) (2.64)
z = r cos(θ)
The domain of variation is 0 ≤ r < ∞, 0 ≤ θ ≤ π, 0 ≤ φ < 2π. The position
vector is

r(x0 ) = r sin(θ) cos(φ)e1 + r sin(θ) sin(φ)e2 + r cos(θ)e3

28
Local basis
The normalized tangent vectors along the directions of the spherical coordinates
are
er (x0 ) = ∂r r = sin(θ) cos(φ)e1 + sin(θ) sin(φ)e2 − sin(θ)e3
eθ (x0 ) = 1r ∂θ r = cos(θ) cos(φ)e1 + cos(θ) sin(φ)e2 − sin(θ)e3
1
eφ (x0 ) = r sin(θ) ∂φ r = − sin(φ)e1 + cos(φ)e2

They define an orthonormal basis such that an arbitrary vector field may be
resolved in these directions

a(x0 ) = ar (x0 )er (x0 ) + aθ (x0 )eθ (x0 ) + aφ (x0 )eφ (x0 )

with a0i = e0i · a, where a01 = ar , e01 = er etc.

Resolution of the gradient


The gradient operator may also be resolved on the basis

∇r = e r ∇r + e θ ∇θ + e φ ∇φ
(2.65)
1
= er ∂r + eθ r1 ∂θ + r sin(θ) ∂φ

The non vanishing derivatives of the basis vectors are

∂θ er = eθ ∂φ er = sin(θ)eφ
∂θ eθ = −er ∂φ eθ = cos(θ)eφ (2.66)
∂φ eφ = − sin(θ)er − cos(θ)eθ

This is all we need for expressing vector operators in spherical coordinates.

Laplacian
The laplacian in spherical coordinates becomes

∇2 = (er ∇r + eθ ∇θ + eφ ∇φ ) · (er ∇r + eθ ∇θ + eφ ∇φ ) ,

which, after using the linearity of the scalar product and Eq. (2.66) becomes

2 1 2 cos(θ) 1
∇2 = ∂rr
2
+ ∂r + 2 ∂θθ + 2 ∂θ + 2 2 ∂φφ 2
r r r sin(θ) r sin (θ)

Notice that the two first terms involving radial derivatives can be given alter-
native expressions

2 2 1  1 2
φ) + ( ∂r φ) = 2 ∂r r2 (∂r φ) = ∂rr

(∂rr (rφ) ,
r r r
where φ is a scalar field.

29
Chapter 3

Tensors

In this chapter we will limit ourself to the discussion of rank 2 tensors, unless
stated otherwise. The precise definition of the rank of a tensor will become clear
later.

3.1 Definition
A tensor, T, of rank two is a geometrical object that to any vector a associate
another vector u = T(a) by a linear operation

T(ma) = mT(a) (3.1)


T(a + b) = T(a) + T(b) (3.2)

In other words a rank 2 tensor is a linear vector operator. We will denote tensors
(of rank 2 or higher) with capital bold-face letters.
Any linear transformation of vectors, such as rotations, reflections or pro-
jections are examples of tensors. In fact, we have already encountered tensors
in disguise. The operator T = c× is a tensor. For each vector, a, it associates
the vector T(a) = c × a, obtained by rotating a 900 counter-clockwise around
c and scaling it with the magnitude |c|. Since the cross-product is linear in the
second argument T is linear and thus a tensor. For reasons to become clear we
will also extend and use the dot-notation to indicate the operation of a tensor
on a vector so
T · a =def. T(a) (3.3)
Physical examples of tensors include the inertia tensor, I, or the moment of
inertia, that specifies how a torque, τ (a vector), changes the angular momen-
dl
tum, dt (a vector)
dl
=I·τ
dt
Another example is the (transposed) stress tensor, σ t , that specifies the force f
a continuous medium exerts on a surface element defined by the normal vector

30
n.
f = σt · n
A tensor often associated with the stress tensor is the strain tensor, U, that
specifies how a strained material is distorted ∆u in some direction ∆r

∆u = U · ∆r

We shall be more specific on these physical tensors later.

3.2 Outer product


The outer product a ⊗ b or simply ab between two vectors a and b is a tensor
defined by the following equation, where c is any vector

(ab)(c) = a(b · c) (3.4)

In words, for each c the tensor ab associates a vector in the direction of a and
with a magnitude equal to the projection of c into b. In order to call the object
(ab) a tensor we should verify that it is a linear operator. Using the definition,
Eq. (3.3), allows us to “place the brackets where we want”

(ab) · c = a(b · c),

which will ease the notation. Now, demonstrating that (ab) indeed is a a tensor
amounts to “moving brackets”:

(ab) · (mc) = a(b · (mc)) = ma(b · c)


= m(ab) · c
(3.5)
(ab) · (c + d) = a(b · (c + d)) = a(b · c) + a(b · d)
= (ab) · c + (ab) · d

We note that since the definition, eq. (??) involves vector operations that
bear no reference to coordinates/components, the outer product will itself be
invariant to the choise of coordinate system. The outer product is also known
in the litterature as the tensor, direct, exterior or dyadic product. The tensor
formed by the outer product of two vectors is called a dyad.

3.3 Basic tensor algebra


We can form more general linear vector operators by taking sum of dyads. The
sum of any two tensors,S and T, and the multiplication of a tensor with a scalar
m are naturally defined by

(S + T) · c =S·c+T·c
(3.6)
(mS) · c = m(S · c).

31
Here, c is any vector. It is easy to show that that (S + T) and (mS) are also
tensors, ie. they satisfy the definition of being linear vector operators, Eq.(4.3.2).
Definition Eq. (3.6) and the properties Eq. (3.5) guarantee that dyad products,
sums of tensors and dot products of tensors with vectors satisfy all the usual
algebraic rules for sums and products.
P For instance,
P the outer product between
two vectors on the form c = i mi ai and d = j nj bj , where mi and nj are
scalars, is
! 
X X X
cd = m i ai  nj b j  = m i n j ai b j , (3.7)
i j ij

ie. it is a sum of all dyad combinations ai bj . The sum of two or more dyads
is called a dyadic. As we shall demonstrate in section 3.4 any tensor can be
expressed as a dyadic but not necessarily as a single dyad.
We can also extend the definition of the dot product. For any vector c we
define the dot product of two tensors by the following formula where c is any
vector
(T · S) · c = T · (S · c) (3.8)
In words, application of the operator (T·S) to any vector means first applying S
and then T. Since the association of two linear functions is a linear function (T·
S) is a tensor itself. Sums and products of tensors also obey usual rules of algebra
except that dot multiplication of two tensors, in general, is not commutative

T+S =S+T (3.9)


T · (S + P) = T · S + T · P (3.10)
T · (S · P) = (T · S) · P (3.11)

3.3.1 Transposed tensors


Furthermore, we define a dot product of a dyad with a vector on the left in the
obvious way
c · (ab) = (c · a)b (3.12)
and correspondingly for a dyadic. Since the dot product of a dyadic with a
vector is not in general commutative

T · c 6= c · T.

it makes sence to introduce the transpose Tt of a tensor

Tt · c = c · T (3.13)

Since c · (ab) = (ba) · c we have for a dyad

(ab)t = ba. (3.14)

32
As we shall see in section 3.4 any tensor can be expressed as a sum of dyads.
Using this fact following properties can easily be proved

(T + S)t = T t + St (3.15)
t t t
(T · S) =S ·T (3.16)
t
Tt = T. (3.17)

3.3.2 Contraction
Another useful operation on tensors is that of contraction. The contraction,
ab : cd, of two dyads results in a scalar defined by

ab : cd = (a · c)(b · d) (3.18)

Note the following useful relation for the contraction of two dyads formed by
basis vectors
ei ej : ek el = δik δjl
The contraction of two dyadics is defined as the bilinear operation
!  
X X X
m i ai b i :  nj c j d j  = mi nj (ai · cj )(bi · dj ), (3.19)
i j ij

where mi and nj are scalars.


In some textbook another type of contraction (a “transposed contraction”)
is also defined
ab · ·cd = ab : dc
Similarly for two dyadics
!  
X X X
m i ai b i ·· nj c j d j  = mi nj (ai · dj )(bi · cj )
i j ij

3.3.3 Special tensors


Some special tensors arise as a consequence of the basic tensor algebra.
A symmetric tensor is a tensor that satisfies Tt = T.
A anti-symmetric tensor is a tensor that satisfies Tt = −T
Any tensor T can trivially be decomposed into a symmetric and an antisym-
metric part. To see this, define

S(T) = T + Tt
(3.20)
A(T) = T − Tt

Then St = S is symmetric and At = −A is anti-symmetric and


1 1
T= S+ A
2 2

33
Finally, an important tensor is the identity or unit tensor, 1. It may be
defined as the operator which acting on any vector, yields the vector itself.
Evidently 1 is one of the special cases for which for all tensors A
1·A =A·1

If m is any scalar the product m1 is called a constant tensor and has the property
(m1) · A = A · (m1) = mA
Constant tensors therefore commute will all tensors and no other tensors have
this property.

3.4 Tensor components in orthonormal bases


Just as is the case with vectors we can represent a tensor in terms of its com-
ponents once we choose a basis. Let therefore E be an orthonormal basis and
T be any tensor. Define
Tij = ei · T · ej (3.21)
In words Tij is the i’th component of the image of the j’th basis vector. Tij is
also called the ijth component of T.
Using the linearity of T and scalar products one observes that the tensor
P is
fully specified by these 3 × 3 = 9 quantities. Indeed, for any vector c = j cj ej
one obtains the i’th component of the image u = T · c by
 P 
e i · T · ( j cj e j ) =
P 
ei · (3.22)
j cj T · e j =
P P
j cj (ei · T · ej ) = j Tij cj

Thus, X
T·c = Tij cj ei (3.23)
ij

Since two indices are required for a full representation of T, it is classified as


a tensor of second rank. We will use the notation [T]E,ij for the ij component
of the tensor in basis E or simply [T]ij when it is clear from the context (or
irrelevant) which basis is implied. Note that the kl component of a dyad ei ej is
(ei ej )kl = (ek · ei )(ej · el ) = δik δjl
and that the 3 × 3 dyad combinations ei ej form a basis for the tensor space
X
T= Tij ei ej (3.24)
ij

Thus, the concepts of a dyadic and linear vector operator or tensor are identical
and are equivalent to the concept of linear vector function in the sense that
every linear vector function defines a certain tensor or dyadic, and conversely.

34
Conveniently, the components, Tij , can be arranged in a matrix T = (Tij ),1
 
T11 T12 T13
T = (Tij ) =def.  T21 T22 T23  (3.25)
T31 T32 T33

In analogy with the notation used for vectors we write T = [T]E –or simply
T = [T] when the basis is irrelevant– as a short hand notation for the component
matrix of a tensor with respect to a given basis. Also, for the inverse operation
(ie. the tensor obtained by expanding the components along the basis dyads)
we shall use the notation
X
T = (T )E =def Tij ei ej . (3.26)
ij

The row-column convention in Eq. (3.25) implies that the representation of


the dyad ei ej is a matrix with a 1 in the i’th row and j’th column and zero
everywhere else.
As for vectors one should notice the difference between a tensor and its
matrix representation. The concept of a matrix is purely algebraic; matrices are
arrays of numbers which may be added and multiplied according to particular
rules (see below). The concept of a tensor is geometrical; a tensor may be
represented in any particular coordinate system by a matrix, but the matrix
must be transformed according to a definite rule if the coordinate system is
changed.

3.4.1 Matrix algebra


Our definitions of the tensor-vector dot product and tensor-tensor dot-product
are consistent with the ordinary rules of matrix algebra.

Tensor · vector
For instance, Eq. (3.22) shows that if u = T · c then the relationship between
the components of u, T and c in any ortonormal basis will satisfy
X
ui = [T · c]i = Tij cj , (3.27)
j

In standard matrix notation this is precisely equivalent to


     
u1 T11 T12 T13 u1
 u2  =  T21 T22 T23  ·  u2 
u3 T31 T32 T33 u3
1 In accordance with standard notation in litterature the bracket around the components,

(Tij ), is used to indicate the matrix collection of these. This is in analogy to the notation
used for triplets a = (ai ), cf. section 2.3.

35
or
u = T · a,
where the triplets, u and a are to be considered as a columns. Therefore, we have
the correspondence between the tensor operation and its matrix representation

[T · c] = [T] · [c]

Tensor · tensor
Also, let us consider the implication of definition of the dot product of two
tensors, Eq. (3.8) in terms of the components.
P 
T · (S · c) = T · kj Skj cj ek
P
= kj Skj cj T · ek
P P
= kj Skj cj i Tik ei
P P
= ij k Tik Skj cj ei (3.28)

Comparing this with Eq. (3.23), shows that


X
[T · S]ij = Tik Skj (3.29)
k

which again is identical to the ij’th element of the matrix product T · S. A more
transparent derivation is obtained by noting that for any vector c

(ek el ) · c = ek (el · c) = ek cl

and consequently

(ei ej · ek el ) · c = (ei ej ) · (ek el · c) = (ei ej ) · ek cl = ei (ej · ek )cl


= ei δjk cl = δjk ei (el · c) = δjk (ei el ) · c

Therefore, one obtains the simple result for the dot product of two basis dyads

ei ej · ek el = δjk ei el (3.30)

Now, evaluating T · S in terms of its dyadic representations and collecting


terms leads to P  P
T·S = ij Tij ei ej · ( kl Skl ek el )
P
= ijkl Tij Skl (ei ej ) · (ek el ) (3.31)
P
= ijkl Tij Skl δjk ei el
P P
= ij l Til Slj ei ej
In obtaining the last expression we have exchanged the summation indices j
and l. The final expression confirms Eq. (3.29). In summary,

[T · S] = [T] · [S]

36
Tensor + tensor
One can similarly show that the definition of the sum of two tensors Eq. (3.6),
(S + T) · c = S · c + T · c, (3.6)
implies that tensors are added by adding their component matrices. Indeed,
P P 
(S + T) · c = S ij e i e j + T ij e i e j ·c
P ij ij
P
= ij Sij ei (ej · c) + ij Tij ei (ej · c) (def. Eq. (3.6))
P (3.32)
=  ij (Sij + Tij )ei (ej  · c)
P
= ij (Sij + Tij )ei ej · c

Consequently, X
S+T = (Sij + Tij )ei ej (3.33)
ij

or equivalently,
[S + T]ij = Sij + Tij , (3.34)
which means that the component matrix of S + T is obtained by adding the
component matrices of S and T respectively
[S + T] = [S] + [T]
It is left as an exercise to demonstrate that
[mT]ij = mTij , (3.35)
where m is a scalar.

Transposition
Finally, the definition of the transpose of a tensor, Eq. (3.13), is also consistent
with the algebraic definition of transposition. Specifically, comparing
   
X X X X
c·  Tij ei ej  = Tij (c·ei )ej = Tij ej (ei ·c) =  Tji ei ej  ·c (3.36)
ij ij ij ij

with the definition of the transpose


Tt · c = c · T
the components satisfy  t
T ij = Tji , (3.37)
in agreement with the matrix definition of transposition. In other words, the
matrix representation of the transposed tensor equals the transposed matrix
representation of the tensor itself:
[Tt ] = [T]t

37
Eq. (3.36) also demonstrates
X
[c · T]i = cj Tji (3.38)
j

In keeping with the convention that c represents a column, the matrix form
of this equation is
ct · T ,
where ct and and the image ct · T will be rows.

3.4.2 Two-point components


As a curiosity, we mention that it is not compulsery to use the same basis for
the left and the right basis-vectors when expressing a tensor in terms of its
components. There are indeed cases where it is advantageous to use different
left and right bases. Such tensor-components are called two-point components.
In these cases one should make it clear in the notation that the first index of the
component refer to a different basis than the second. For a two-point component
matrix T̃ we write X
T = (T̃ )E 0 ,E = T̃ij e0i ej
ij

and similarly, [T]E 0 E,ij for T̃ij . Two-point components do not involve any extra
formalism though. For instance, getting the components in the dyad basis
E 0 E 0 or in the basis EE is only a question of respectively multiplying with the
basis-vectors e0i to the right or the basis-vectors ei to the left on the expression
0
P
kl T̃kl ek el and collect the terms. For instance
P 
0
[T]EE,ij = ei · T̃ e e
kl kl k l · ej
= kl T̃kl (ei · e0k )(el · ej )
P
(3.39)
= kl T̃kl (ei · e0k )δlj
P

= k T̃kj (ei · e0k )


P

showing that that ij’th component in basis EE, Tij , is


X
Tij = T̃kj (ei · e0k )
k

3.5 Tensor fields


As for scalar and vector fields one speaks of a tensor field (here of rank 2) if to
each position x = (x1 , x2 , x3 ) of a region R in space there corresponds a tensor
T(x). The best known examples of tensor fields in physics is the stress and
strain tensor fields.

38
In cartesian coordinates a tensor field is given by
X
T(x) = Tij (x)ei ej (3.40)
ij

Note in particular that for each i = 1, 2, 3


P
ai (x) = T(x) · ei = j Tji (x)ej and
P (3.41)
bi (x) = ei · T(x) = j Tij (x)ej

are vector fields with components [ai ]j = Tji and [bi ]j = Tij , respectively.

3.5.1 Gradient, divergence and curl


Let a cartesian coordinate system be given C = (O, R).

Nabla operator
For any vector field,
P a(r) we can construct a tensor field by the nabla operator.
Inserting a(r) = i ai (r)ei “blindly” into the definition Eq. (2.41) gives
 
X X X
(∇a)(r) = ei ∂i  aj (r)ej  = (∂i aj )(r)ei ej (3.42)
i j ij

Assuming that ∇ is a proper vector operator2 then (∇a)(r) represents a tensor


field with the components

[∇a]ij (r) = (∂i aj )(r) (3.43)

Notice that the components of ∇a are that of an outer product between two
ordinary vectors, [ba]ji = bj ai .
The tensor, ∇a, represents how the vector field a changes for a given dis-
placement ∆r
a(r + ∆r) ≈ a(r) + ∆r · (∇a)(r)
or
da = dr · (∇a)
An alternative expression for the differential da was obtained in Eq. (2.44),

da = dr · (∇a) = (dr · ∇)a

Notice that this identity is in accordance with the usual definition for a dyad
operating on a vector from the left.
2 We recall that the ∇ operator has been defined relative to a chosen coordinate system. It

remains to be proven that operators derived from ∇ such as gradients, curls, divergence etc.
“behave as” proper scalars or vectors. We return to this point in 4.5.

39
Divergence
The divergence of a tensor field is obtained straight forwardly by applying the
definition of ∇ and the dot-product. With respect to a cartesian basis R we
have  P P 
∇ · T (x) = ( i ei ∂i ) · jk Tjk (x)ej ek
P P
= Pi P (3.44)
k ∂i Tik (x)ek
= k ( i ∂i Tik (x)) ek .
P
Consequently, a = ∇ · T is a vector with components ak = i ∂i Tik , jvf. Eq.
(3.41). The result is actually easier seen in pure subscript notation
X
[∇ · T]k = ∂i Tik ,
i

because it follows directly from the matrix algebra, Eq. (3.38). Another way of
viewing the divergence of a tensor field is to take the divergence of each of the
vector fields ai = T · ei X
∇·T= (∇ · ai ) ei
i

Curl
The curl of a tensor field is obtained similarly to the divergence. The result is
most easily obtained by applying the algebraic definition of curl, Eq. (2.47).
Then we obtain X
[∇ × T]ij = imn ∂m Tnj
mn

Thus the curl of a tensor field is another tensor field. As for the divergence
we obtain the same result by considering the curl operator on each of the three
vector field aj = T · ej
X 
∇×T= ∇ × aj ej ,
j

where an outer vector product is involved in each of the terms ∇ × aj ej .

3.5.2 Integral theorems


The two important theorems for vector fields, Gauss theorem and Stokes theo-
rem, can be directly translated to tensor fields. The analogy is most easily seen
in tensor notation. Then for a vector field a(r) the two theorem reads
R P H P
V ( i ∂i ai ) dV  = S ( i ai ni ) dS Gauss
R P H P (3.45)
V ijk ijk ∂j ak ni dS = C ( i ai dxi ) Stokes

40
The corresponding tensor versions are then
R P H P
V
( i ∂i Til ) dV  = S ( i Til ni ) dS Gauss
R P H P (3.46)
V ijk ijk ∂j Tkl ni dS = C ( i Til dxi ) Stokes

Here l refer to any component index l = 1, 2, 3. The reasonP these formulas also
works for tensors is that for a fixed l, al = T · el = i Til ei defines a vector
field, c.f. (3.41), and Gauss’ and Stokes’ theorems works for each of these.
Due to Gauss theorem the physical interpretation of the divergence of a
tensor field is analogous to the divergence of a vector field. For instance, if a
flow field v(r, t) is advecting a vector a(r, t) then the outer product J(r, t) =
v(r, t)a(r, t) is a tensor field where [J]ij (r, t) = vi (r, t)aj (r, t) describes how
much of the j’th component of the vector a is transported in direction ei . The
divergence ∇ · J is then a vector where each component [∇ · J]j corresponds
to the accumulation of aj in an infinitesimal volume due to the flow. In other
words, if a represents a conserved quantity then we have a continuity equation
for the vector field a
∂a
+∇·J=0
∂t

41
Chapter 4

Tensor calculus

In the two preceeding chapters we have developed the basic algebra to for scalar,
vector and rank 2 tensors. We have seen that the notation of vectors and tensors
comes in two flavours. The first notation insists in not refering to coordinates or
components at all. This geometrically defined notation, called direct notation,
is explictly invariant to the choise of coordinate system and is the one adopted
in the first part of chapter two and three (2.1-2.2, 3.1-3.3). The second notation,
based on components and indices, is called tensor notation and is the one that
naturally arises with a given coordinate system. Here, vectorial or tensorial
relations are expressed algebraically.
Seemingly, a discrepancy exists between the two notations in that the lat-
ter appears to depend on the choise of coordinates. In section 4.2 we remove
this descrepancy by learning how to transform the components of vectors and
tensors when the coordinate system is changed. As we shall see these transfor-
mation rules guarantee that one can unambigously express vectorial or tensorial
relations using component/tensor notation because the expressions will preserve
the form upon a coordinate transformation. Indeed, we should already expect
this to be the case since we have not made any specific assumptions about the
coordinate system in the propositions regarding relations between vector or ten-
sor components already presented, except of this system beeing rectangular. It
is possible to generalize the tensor notation to ensure that the formalism takes
the same form in non-rectangular coordinate systems as well. This requirement
is known as general covariance. In the following we will, however, restrict the
discussion to rectangular coordinates.
Since tensor notation often requires knowledge of transformation rules be-
tween coordinate systems one may validly ask why use it at all. First of all,
in quantifying any physical system we will eventually have to give a numerical
representation which for vectors or tensors imply to specify the value of their
components with respect to some basis. Secondly, in many physical problems it
is natural to introduce tensors of rank higher than two1 . To insist on a direct
1 The Levi-Civita symbol being an example of a (pseudo)-tensor of rank 3.

42
notation for these objects become increasingly tedious and unnatural. In tensor
notation no new notation needs to be defined, only those we have already seen
(some in disguise), which is reviewed in section 4.3. It is therefore strongly
recommended to become familiar with the tensor notation once and for all.

4.1 Tensor notation and Einsteins summation


rule
Using tensor notation an index will either be free or bound. A free index occurs
exactly once in a term and has the meaning “for any component”. For instance,
in the formula for the addition of two vectors in terms of its components, Eq.
(2.13),
[a + b]i = [a]i + [b]i ,
index i is free. The same free indices must occur in each term. As in the
above example, an equation with exactly one free index in each term expresses
a vector identity, an expression with two free indices in each term expresses a
tensor identity, etc.
A bound index (or a dummy index) refers to a summation index. For in-
stance the scalar product in an orthonormal basis, Eq. (2.18), reads
X
a·b= a i bi .
i

Here, i is a bound index and can be renamed without changing the meaning of
the expression. The situation where a dummy index appears exactly twice in a
product occurs so often that it is convenient to introduce the convention that
the summation symbol may be left out in this specific case. In particular, it
occurs in any component representation of a “dot” or inner product. Hence we
may write
P
a·b = P i a i bi = a i bi (i a bound index)
[T · a]j = i Tji ai = Tji ai (i a bound index, j a free index)
P (4.1)
[T · S]ij = k Tik Skj = Tik Skj (ij free indices, k bound index)
P
T:S = ij Tij Sij = Tij Sij ij bound indices
Similar, for the expansion of the components along the basis vectors, Eq. (2.10)
X
a= a i gi = a i gi
i

and for the divergence of a vector field


X
∇·a = ∂i ai = ∂ i ai
i

The convention of implicit summation which is very convenient for tensor


calculus is due to Einstein and is referred to as Einsteins summation rule. In
this and the following chapters we shall adopt the implicit summation rule unless
otherwise stated.

43
4.2 Orthonormal basis transformation
Let E and E 0 be two different orthonormal bases. The basis vectors of the primed
basis may be revolved into the unprimed basis

e0 j = aji ei (sum over i), (4.2)

where
aji =def e0j · ei (4.3)
represents the cosine of the angle between e0j and ei . The relation between the
primed and unprimed components of any (proper) vector

v = ci ei = c0j e0 j

can be obtained by dotting with e0 j or ei

vj0 = aji vi (4.4)


vi = aji vj0 (4.5)

Here, vj0 = [v]E 0 ,j and vi = [v]E,i . Note, that the two transformations differ in
whether the sum is over the first or the second index of aij . The matrix version
of Eq. (4.4) reads
v 0 = A · v, A = (aij ), (4.6)
and Eq. (4.5)
v = At · v 0 .
Using the same procedure the primed and unprimed components of a tensor

T = Tij0 e0i e0k = Tjl ej el


2
are related by
0
Tik = e0 i · T · e0 k = aij akl Tjl (4.7)
Tik = ei · T · ek = aij akl Tjl , (4.8)
0
where Tik = [T]E 0 ,ik and Tik = [T]E,ik . Note that the transformation matrix A
is simply the two-point components of the identity tensor

A = [1]E 0 E
2 In keeping with Einsteins summation rule this expression is short hand notation for
X X
0 0 0
T= Tij ei ek = Tjl ej el
ik jl

44
4.2.1 Cartesian coordinate transformation
A particular case of an orthonormal basis transformation is the transformation
between two cartesian coordinate systems, C = (O, R) and C 0 = (O0 , R0 ). Here,
the form of the vector transformation, Eq. (4.4), can be directly translated to
the coordinates themselves. To show this we employ the simple relation between
coordinates and the position vector, Eq. (2.9), valid in any cartesian system.
The displacement vector between any two points is given by ∆r = ∆xi ei . This
vector –or its differential analogue dr = dxi ei – is a proper vector independent
of the coordinate system. Following relation between the coordinate differentials
dxi and dx0j must then be satisfied
dr = dxi ei = dx0j e0 j
Multiplying with e0j we obtain

dx0j = dxi ei · e0j


or equivalently
∂x0j
= e0j · ei = aji , (4.9)
∂xi
where aji is defined as in Eq. (4.3). Since the rhs. is constant (constant basis
vectors) the integration of Eq. (4.9) gives
x0j = aji xi + dj . (4.10)
Here, dj is the j’th coordinate of the origin, O, of the unprimed coordinate
system as seen from the primed coordinate system, d = (x 0 )C 0 (O). In matrix
notation Eq. (4.10) takes the form
x0 = A · x + d, A = (aij ), (4.11)
which is identical to Eq. (4.6) except from the optional displacement.

4.2.2 The orthogonal group


Considering Eq. (4.4) and Eq. (4.5) as a set of two matrix equations we note
that the transpose matrix At operates as the inverse of A
A · At = 1, (4.12)
where 1 is the identity matrix, (1)ij = δij . In index notation
(A · At )ij = aik akj = δij (4.13)
A matrix with the above property is called orthogonal, and the set of of all
orthogonal 3×3 matrices constitutes a continuous group3 called O(3). According
to Eq. (4.12)
det(A · At ) = det(A)2 = 1,
3 Recall, that a mathematical group (G, ∗) is a set G with a binary operator, ∗, that satisfies

following axioms:

45
so
det(A) = ±1
Transformations with the determinant +1 or −1 cannot be continuously con-
nected. Therefore, the set of transformations with determinant +1 itself forms
a group, called SO(3), which represents the set of all rotations. By considering
a pure reflection, R = (rij ), in the origin of a cartesian coordinate system

x0i = −xi = rij xj , rij = −δij

we see that det(R) = −1. Clearly, R · R = 1, so the set Z(2) = {1, R} also
forms a group. Consequently, we may write O(3) = Z(2) ⊗ SO(3). In words,
any orthogonal matrix can be decomposed into a pure rotation and an optional
reflection. Note, that any matrix with det = −1 will change the handedness of
the coordinate system.

4.2.3 Algebraic invariance


In view of the fact that the various vector and tensor operations (section 2.1-
2.3, 3.1-3.3) were defined without reference to a coordinate system, it is clear
that all algebraic rules for computing sums, products, transposes, etc. of vec-
tors and tensors (cf. section 3.4.1) will be unaffected by an orthogonal basis
transformation. Thus for examples

a·b = a0i b0i = aj bj


[a + b]E 0 ,i = a0i + b0i
(4.14)
[T · c]E 0 ,i = Tij0 c0j
[Tt ]E 0 ,ij = Tji0 ,

where, a0i = [a]E 0 ,i , aj = [a]E,j , Tji0 = [T]E 0 ,ji etc.


Any property or relation between vectors and tensors which is expressed in
the same algebraic form in all coordinate systems has a geometrical meaning
independent of the coordinate system and is called an invariant property or
relation.
The invariance of the algebraic rules for vector and tensor operations can
be verified directly by using the orthogonality property of the transformation
matrix, Eq. (4.13). In particular, whenever a “dot” or inner product appears
between two vectors, a vector and a tensor or two tensors we can pull together
two tranformation matrices and exploit Eq. (4.13). For instance, to verify the
1. Closure: ∀a, b ∈ G : a ∗ b ∈ G.
2. Associativity: ∀a, b, c ∈ G : (a ∗ b) ∗ c = a ∗ (b ∗ c)
3. Identity element: There exists an element e ∈ G so that ∀a : a ∗ e = e ∗ a = a.
4. Inverse element: For each a ∈ G the exists an inverse a−1 ∈ G such that a ∗ a−1 =
a−1 ∗ a = e.
Orthogonal matrices forms a group with matrix multiplication as the binary operator and
the identity matrix as the identity element. Notice in particular that the composition or
multiplication of two orthogonal transformations is orthogonal.

46
third invariance above, ie. that the i’th component of the image u = T · c of T
operating on c, is always obtained as

ui = Tij cj , (3.27)

irrespective of the chosen coordinate system, we can tranform the components


of u from one (unprimed) to another (primed) coordinate system and see if these
equal the transformed component matrix of T times the transformed component
triplet of c:

[T]E 0 ,ij [c]E 0 ,j = Tij0 c0j


= (aik ajl Tkl )(ajm cm ) Transf. rules for components
= aik Tkl (ajl ajm )cm
= aik Tkl δml cm Orthogonality (summing over k)
= aik (Tkl cl )
= aik [u]E,k
= [u]E 0 ,i Tranf. rules for components
(4.15)
We shall return to other important invariances in chapter 5

4.2.4 Active and passive transformation


Comparing the transformation of the components of a vector upon a basis trans-
formation, Eq. (4.4) with the matrix representation of a tensor operating on
a vector, Eq. (3.27), it is clear that these two operations are defined in the
same manner algebraically. Since the first equation only expresses the change
of components due to a change of basis (the vector itself remains unchanged) it
is referred to as a passive transformation as opposed to the second case which
reflects an active transformation of the vector. Here, we shall show a simple
algebraic relation between the two type of transformations. Let O represent an
orthogonal tensor, ie. a tensor for which

Ot · O = 1,

where 1 as usual denotes the identity tensor. In any orthonormal basis the
matrix representation of this identity will be that of Eq. (4.12). Further, let a
new basis system E 0 = {e01 , e02 , e03 } be defined by

e0i = O · ei

where E = {e1 , e2 , e3 } is the old basis. We shall use the short-hand notation

E 0 = O(E)

Due to the orthogonality property of O, E 0 will be ortonormal iff E is.

47
The components of a vector v in the primed basis relates to the the unprimed
components as

vi0 = e0i · v
= (O · ei ) · v
= [O · ei ]E,j [v]E,j (Scalar product wrt. E)
= [O]jk [ei ]k vj (Eq. (3.27), all components wrt. E)
(4.16)
= [O]jk δik vj
= [O]ji vj
= [Ot ]ij vj
= (Ot · v)i

This shows that the matrix, A, representing the basis transformation, Eq. (4.6),
satisfies
A = [Ot ]E = O t = O−1 .
Being explicit about the bases involved we may write this identity as

[1]E 0 E = [O−1 ]E , where E 0 = O(E), (4.17)

which is to say that the matrix representing a passive transformation from one
to another basis, E → E 0 , equals the matrix representation wrt. E of the tensor
mapping the new basis vectors onto the old ones, O−1 (E 0 ) = E.

4.2.5 Summary on scalars, vectors and tensors


To summarize the present section we may give an algebraic definition of scalars,
vectors and tensors by refering to the transformation properties of their compo-
nents upon a change of coordinates. For an cartesian coordinate transformation,
Eq. (4.10) following transformation properties of scalars, vectors and tensors
apply.

Scalars
A quantity m is a scalar if it has the same value in every coordinate system.
Consequently it transforms according to the rule

m0 = m

Vectors
A triplet of real numbers, v, is a vector if the components transform according
to
vj0 = aji vi
The matrix notation of the transformation rule is

v0 = A · v

48
Rank 2 Tensors
A rank two tensor is a 3 × 3 matrix of real numbers, T , which transforms as the
outer product of two vectors

Tij0 = aik ajl Tkl

Again it should be noticed that it is the transformation rule which guarantees


that the nine quantities collected in a matrix form a tensor. The matrix notation
of this transformation is
T 0 = A · T · At

4.3 Tensors of any rank


4.3.1 Introduction
Tensors of rank higher than two are obtained by a straightforward iteration of
the definition of the outer product. For instance the outer product between a
dyad a1 a2 and a vector a3 can be defined as the linear operator, a1 a2 a3 , that
maps any vector c into a tensor according to the rule

a1 a2 a3 (c) =def. a1 a2 (a3 · c) (4.18)

In words, for each c the object a1 a2 a3 associates a tensor obtained by multi-


plying the dyad a1 a2 with the scalar a3 · c. Definition Eq. (4.18) preserves the
invariance with respect to the choise of coordinates. Addition of triple products
and scalar multiplication is defined analogous to the algebra for dyads.
In section 3.4 we demonstated that a basis for second order tensors is ob-
tained by forming all 3 × 3 dyad combinations, ei ej , between orthonormal basis
vectors. Similarly, one can demonstrate that a basis for all linear operators,
T3 , which map a vector into a second rank tensor is obtained by forming the
33 combinations of the direct products ei ej ek . Consequently, for any T3 there
exists a unique set of quantities, Tijk , such that

T3 = Tijk ei ej ek NB! Einstein summation convention

Not surprisingly, these quantities, Tijk , are called the components of T3 and
since three indicies are needed, T3 is called a third rank tensor. Upon a co-
ordinate transformation we would need to transform the components of each
individual vector in the triple product. Expressing the original basis vectors in
terms of the new basis
ei = aji e0j
one obtains
0
T3 = Tlmn e0l e0m e0n = Tijk ei ej ek
(4.19)
= Tijk (ali e0 l )(amj e0 m )(ank e0 n ) = ali amj ank Tijk e0l e0m e0n

49
This shows that upon an orthogonal transformation the components of a third
rank tensor transform as
0
Tlmn = ali amj ank Tijk
The inverse transformation follows from expressing the new basis vectors in
terms of the old one, e0 j = aji ei in Eq. (4.19)
0
Tijk = ali amj ank Tlmn .
We can continue this procedure, defining the outer product of four vectors
in terms triple products, cf. Eq. (4.18), and introducing addition and scalar
multiplication of these objects as well.
In keeping with previous notation, the set of all components is written as
T = (Tijk ).
Here, ijk are dummy indices and the bracket around Tijk indicates the set of
all of these
(Tijk ) = { Tijk | i = 1, 2, 3 ; j = 1, 2, 3 ; k = 1, 2, 3 }.
Also, the tensor is obtained from the components as
T = (T )E = ( (Tijk ) )E =def Tijk ei ej ek

4.3.2 Definition
In general, a tensor T of rank r is defined as a set of 3r quantities, Ti1 i2 ···ir , that
upon an orthogonal change of coordinates, Eq. (4.11), transform as the outer
product of r vectors
Tj01 j2 ···jr = aj1 i1 aj2 i2 · · · ajr ir Ti1 i2 ···ir . (4.20)
4
Accordingly, a vector is a rank 1 tensor and a scalar is a rank 0 tensor.
It is common to refer to both T and the set (Ti1 i2 ···ir ) as a tensor, although
the latter quantities are strictly speaking only the components of the tensor
with respect to some particular basis E. However, the distinction between a
tensor and its components becomes unimportant, provided we keep in mind
that Tj01 j2 ···jr in Eq. (4.20) are components of the same tensor only with respect
to a different basis E 0 . It is precisely the transformation rule, Eq. (4.20), that
ensures that the two set of components represent the same tensor.
For completeness, the inverse transformation is obtained by summing on the
first indices of the transformation matrix elements instead of the second ones,
cf. section 4.3.2:
Ti1 i2 ···ir = aj1 i1 aj2 i2 · · · ajr ir Tj01 j2 ···jr .
4 The notation T
i1 i2 ···ir may be confusing at first sight. However, i1 , i2 , · · · ir are simply r
independent indices each taking values in the range {1, 2, 3}. Choosing a second rank tensor
as an example and comparing with previous notation, Tij , it simply means that i1 = i, i2 = j.
For any particular component of the tensor, say T31 , we would have with the old notation
i = 3, j = 1 and in the new notation i1 = 3, i2 = 1.

50
4.3.3 Basic algebra
Tensors can be combined in various ways to form new tensors. Indeed, we have
already seen several examples thereof5 . The upshot of these rules below is that
tensor operations that look “natural” are permissible in the sense that if one
starts with tensors the result will be a tensor ??.

Addition/substraction
Tensors of the same rank may be added together. Thus
Ci1 i2 ···ir = Ai1 i2 ···ir + Bi1 ii 2···ir
is a tensor of rank r if A and B are tensors of rank r. This follows directly from
the linearity of Eq. (4.20),
Cj01 j2 ···jr = A0j1 j2 ···jr + Bj0 1 j2 ···jr
= aj1 i1 aj2 i2 · · · ajr ir Ai1 i2 ···ir + aj1 i1 aj2 i2 · · · ajr ir Bi1 i2 ···ir
= aj1 i1 aj2 i2 · · · ajr ir (Ai1 i2 ···ir + Bi1 i2 ···ir )
= aj1 i1 aj2 i2 · · · ajr ir Ci1i2 ···ir .
Consequently, the 3r quantities (Ci1 i2 ···ir ) transform as the components of r
rank tensor when both (Ai1 i2 ···ir ) and (Bi1 ii 2···ir ) do (We made use of this in
the second line of the demonstation above). It follows directly that substraction
of two equally ranked tensors is also a tensor of the same rank, and that tensorial
addition/substraction is commutative and associative.

Outer product
The product of two tensors is a tensor whose rank is the sum of the ranks
of the given tensors. This product which involves ordinary multiplication of
the components of the tensor is called the outer product. It is the natural
generalization of the outer product of two vectors (two tensors of rank 1) defined
in section 3.2. For example
Ci1 i2 ···ir j1 j2 ···js = Ai1 i2 ···ir Bj1 j2 ···js (4.21)
is a tensor of rank r + s if Ai1 i2 ···ir is a tensor of rank r and Bj1 j2 ···js is a tensor
of rank s. Note, that this rule is consistent with the cartesian components of a
dyad c = ab in the specific case where A = a and B = b are vectors. Also, a
tensor may be multiplied with a scalar m = B (tensor of rank 0) according to
the same rule
Ci1 i2 ···ir = mAi1 i2 ···ir
Note, that not every tensor can be written as a product of two tensors of lower
rank. We have already emphasized this point in case of second rank tensors
which can not in general be expressed as a single dyad. For this reason divison
of tensors is not always possible.
5 The composition, T · S, of two second rank tensors, T and S, forms a new second rank

tensor. The operation of a second rank tensor on a vector (1. rank tensor) gives a new vector.
The addition of two second rank tensors gives a new second rank tensor, etc.

51
Permutation
It is permitted to exchange or permute indices in a tensor and still remain with
a tensor. For instance, if we define a new set of 3r quantities Ci1 i2 ···ir from
a tensor Ai1 i2 ···ir by permuting two arbitrarily chosen index numbers α and
β > α:
Ci1 i2 ···iα−1 iα iα+1 ···iβ−1 iβ iβ+1 ···ir = Ai1 i2 ···iα−1 iβ iα+1 ···iβ−1 iα iβ+1 ···ir ,
this new set will also be a tensor. Its tensorial property follows from the sym-
metry among the a-factors in Eq. (4.20). Tensors obtained from permutting
indices are called isomers. Tensors of rank less than two (ie. scalars and vectors)
have no isomers. A tensor of rank two has precisely one isomer, the transposed
one, obtained by setting α = 1 and β = 2 in the above notation:
(Tt )i1 i2 = Ti2 i1

Contraction
The most important rule in tensor algebra is the contraction rule which states
that if two indices in a tensor of rank r + 2 are put equal and summed over,
then the result is again a tensor with rank r. Because of the permutation rule
discussed above we only have to demonstrate it for the first two indices. The
contraction rule states that if A is a tensor of rank r + 2 then
P3
Bj1 j2 ···jr = Aiij1 j2 ···jr (NB! Aiij1 j2 ···jr = i=1 Aiij1 j2 ···jr )
is also a tensor of rank r. The proof follows
Bj0 1 j2 ···jr = A0iij1 j2 ···jr
= aik ail aj1 m1 aj2 m2 · · · ajr mr Aklm1 m2 ···mr
= δkl aj1 m1 aj2 m2 · · · ajr mr Aklm1 m2 ···mr Akkm1 m2 ···mr
= aj1 m1 aj2 m2 · · · ajr mr Bm1 m2 ···mr
Consequently, Bj1 ···jr does transform as a tensor of rank r as claimed.

Inner product
By the process of forming the outer product of two tensors followed by a con-
traction, we obtain a new tensor called an inner product of the given tensors. If
the rank of the two given tensors are respectively r and s then the rank of the
new tensor will be r + s − 2. Its tensorial nature follows from the fact that an
outer product of two tensors is a tensor and the contraction of two indices in a
tensor gives a new tensor. In the previous chapters we have reserved the “dot”-
symbol for precisely this operation. For instance, a scalar product between two
vectors, a and b is a inner product of two rank 1 tensors giving a 0 rank tensor
(scalar):
[ab]ij = a i bj (Outer product)
a·b = a i bi (Contraction of outer product, setting j = i)

52
A rank two tensor, T, operating on a vector v is a inner product of between a
rank two and a rank one tensor giving a rank 1 tensor:
[Tv]ijk = Tij vk (Outer product)
[T · v]i = Tij vj (Contraction of outer product, setting k = j)
It is left to an exercise to see that the “dot”-operation defined in chapter 2 for
two second rank tensors, T · S, is an inner product.
In general, any of two indices in the outer product between two tensors can
be contracted to define an inner product. For example
Ci1 i2 i4 i5 j1 j3 = Ai1 i2 ki4 i5 Bj1 kj3
is a tensor of rank 6 obtained as an inner product between a tensor Ai1 i2 i3 i4 i5 of
rank 5 and a tensor Bj1 j2 j3 of rank 3 by setting i3 = j2 . Tensors of rank higher
than 4 are an odditity in the realm of physics, fortunately .

Summary
In summary, all tensorial algebra of any rank can basically be boiled down to
these few operations: addition, outer product, permutation, contraction and
inner product. Only one more operation is essential to know, namely that of
differentiation.

4.3.4 Differentiation
Let us consider a tensor A of rank r which is a differentiable function of all
of the elements of another tensor B of rank s. The direct notation for this
situation would be
A = A(B) (Rank(A)=r, Rank(B)=s)
In effect, it implies the existence of 3r functions, Ai1 ···ir , each differentiable in
each of its 3s arguments B = (Bj1 j2 ···js ),
Ai1 i2 ···ir = Ai1 i2 ···ir ( (Bj1 j2 ···js ) )
Then the partial derivatives
∂Ai1 i2 ···ir
Ci1 i2 ···ir ,j1 j2 ···js = (4.22)
∂Bj1 j2 ···js
is itself a tensor of rank r + s. As for ordinary derivatives of functions, the
components Ci1 i2 ···ir will in general also be functions of B,
Ci1 i2 ···ir = Ci1 i2 ···ir ( (Bj1 j2 ···js ) )
The direct notation for taking partial derivatives with respect to a set of tensor
components is  
∂A
C(B) = (B)
∂B

53
To demonstrate that C is a tensor, and so to justify this direct notation in the
first place, we will have to look at the transformation properties of its elements.
Indeed, we have
 ∂A0 
Ci01 i2 ···ir ,j1 j2 ···js (B 0 ) = ∂B 0i1 i2 ···ir (B 0 )
j1 j2 ···js 
∂B ∂(ai1 k1 ai2 k2 ···air kr Ak1 k2 ···kr )
= ∂B 0l1 l2 ···ls ∂Bl l ···l (B)
j1 j2 ···js 1 2 s

= ai1 k1 ai2 k2 · · · air kr aj1 l1 aj2 l2 · · · ajs ls Ck1 k2 ···kr ,l1 l2 ···ls (B)
(4.23)
In the second step we have used the chain rule for differentiation. The last step
follows from
Bl1 l2 ···ls = aj1 l1 aj2 l2 · · · ajs ls Bj0 1 j2 ···js .
It is essential that all components of B are independent and can vary freely
without constraints ??.
From this rule follows the quotient rule which states that if the tensor
Ai1 i2 ···ir is a linear function of the unconstrained tensor Bj1 j2 ···js through the
relation
Ai1 i2 ···ir = Ci1 i2 ···ir j1 j2 ···js Bj1 j2 ···js + Di1 i2 ···ir
then C is a tensor of rank r + s and consequently D must also be a tensor of
rank r.

4.4 Reflection and pseudotensors


Applying the general tensor transformation formula, Eq. (4.20), to a reflection
through the origin of a cartesian coordinate system, x0 = −x, we find

Ti01 ···ir = (−1)r Ti1 ···ir (4.24)

There are quantities, notably the Levi-Civita symbol, that do not obey this
rule, but acquire an extra minus sign. Such quantities are called pseudo-tensors
in contradistinction to ordinary or proper tensors. Consequently, a set of 3 r
quantities, Pi1 i2 ···ir , is called a pseudo-tensor if it transforms according to the
rule

Pi01 i2 ···ir = det(A)ai1 j1 ai2 j2 · · · air jr Pj1 j2 ···jr Pseudo-tensor, (4.25)

upon the coordinate transformation, Eq. (4.11) For a pure reflection, aij = −δij ,
so
Pi01 i2 ···ir = −(−1)r Pi1i2 ···ir .
By comparing Eq. (4.25) with Eq. (4.20) one observes that the difference be-
tween a tensor and a pseudotensor only appears if the transformation includes a
reflection, det(A) = −1. A proper tensor can be considered as a real geometrical
object, independent of the coordinate system. A proper vector is an “arrow”
in space for which the components change sign upon a reflection of the axes
of the coordinate system, cf. Eq. (4.24). Direct product of ordinary vectors

54
are ordinary tensors, changing sign once for each index. An even number of
direct products of pseudo-tensors will make an ordinary tensor, whereas an odd
number of direct products of pseudo-tensors leads to another pseudo-tensor.

4.4.1 Levi-Civita symbol


If we apply the transformation rule, Eq. (4.20), to the Levi-Civita symbol, ijk
we obtain
0ijk = ail ajm akn lmn . (4.26)
This expression is antisymmetric in all indices ijk 6 . Consequently, it must be
proportional to ijk . To get the constant of proportionality, we should observe
the following connection between the determinant of a matrix and the Levi-
Civita symbol
det(A) = lmn a1l a2m a3n , A = (aij ). (4.27)
Taking ijk = 123 in Eq. (4.26) we get

0123 = det(A) = det(A)123 ,

since 123 = +1. Thus the constant of proportionality is det(A) and 0ijk =
det(A)ijk . If 0ijk should be identical to ijk we must multiply the transforma-
tion with an extra det(A) to account for the possiblity that the transformation
involves a reflection, where det(A) = −1. The correct transformation law must
therefore be
0ijk = det(A)ail ajm akn lmn ,
whence the Levi-Civita symbol is a third rank pseudo-tensor. The transforma-
tion rule for pseudo-tensors leaves the Levi-Civita symbol invariant under all
orthogonal transformations, 0ijk = ijk . 7
The most important application of the Levi-Civita symbol is in the algebraic
definition of the cross-product, c = a × b between two ordinary vectors

ci = [a × b]i = ijk aj bk .

Since the rhs. is a double inner product between a pseudo-tensor of rank three
and two ordinary vectors, the lhs. will be a pseudo-vector. Indeed, the direc-
tion of c depends on the handedness of the coordinate system, as previously
mentioned. An equivalent manifestation of its pseudo-vectorial nature is that c
does not change its direction upon an active reflection.
6 For instance,

0ikj = ail akm ajn lmn = ail akn ajm lnm = ail ajm akn lnm = −ail ajm akn lmn = −0ijk
The second identity follows from the fact that n and m are bound indices and can therefore
be renamed to one and another.
7 We should not be suprised by the fact that 
ijk are unaffected by coordinate transforma-
tions, since its definition makes no distinction between the 1,2 and 3 directions.

55
4.4.2 Manipulation of the Levi-Civita symbol
Two important relations for the Levi-Civita symbol are very useful to derive
the myriad of known identities between various vector products. The first is a
formula to reduce the product of two Levi-Civita symbols into kronecker deltas
 
δil δim δin
ijk lmn = det  δjl δjm δjn  (4.28)
δkl δkm δkn

which after calculating the determinant becomes

ijk lmn = δil δjm δkn + δim δjn δkl + δin δjl δkm
(4.29)
−δin δjm δkl − δjn δkm δil − δkn δim δjl

Due to this relation an arbitrary tensor expression may be reduced to contain


zero -symbols whereas a pseudotensor expression may be reduced to contain
only one such symbol. Every time two ’s meet in an expression they may be
reduced away ??. This is the secret behind all complicated formules involving
vector products. From Eq. (4.29) we may deduce a few more expressions by
contraction of indices
ijk lmk = δil δjm − δim δjl
ijk ljk = 2δil (4.30)
ijk ijk =6

An other important relation for the -symbol follows from the observation that
it is impossible to construct a non-trivial totally antisymmetric symbol with
more than three indicies. The reason is that the antisymmetry implies that any
component with two equal components must vanish. Hence a non-vanishing
component must have all indices different, and since there are only three possible
values for an index this is impossible. Consequently, all components of a totally
antisymmetric symbol with four indices must vanish. From this follows the rule

ai jkl − aj ikl − ak jil − al jki = 0, (4.31)

because the lhs. is antisymmetric in four indices.

Derivation of Eq. (4.28)


To derive Eq. (4.28) one uses two well-known properties of the determinant of
a matrix. First, the determinant of a product of matrices equals the product of
the determinants of the individual matrices

det(M · N) = det(M ) det(N )

Secondly, the determinant of the transpose, N t , of a matrix, N , equals the


determinant of the matrix

det(N t ) = det(N )

56
Set    
a1 a2 a3 x1 x2 x3
M =  b1 b2 b3  , N =  y1 y2 y3  ,
c1 c2 c3 z1 z2 z3
then the matrix product M · N t becomes
 
a·x a·y a·z
M · Nt =  b · x b · y b·z  (4.32)
c·x c·y c·z

Also, det(M ) = ijk ai bj ck , and det(N ) = lmn xl ym zn according to Eq. (2.24).


Consequently,

det(M · N t ) = det(M ) det(N ) = ijk lmn ai bj ck xl ym zn ,

which together with Eq. (4.32) gives


 
a·x a·y a·z
ijk lmn ai bj ck xl ym zn = det  b · x b·y b·z . (4.33)
c·x c·y c·z

Setting a = ei , b = ej , c = ek , x = el , y = em , z = en we have for the lhs of


Eq. (4.33)
P
 0 0 0 0 0 0a 0b 0c 0x 0y 0z 0 =
P i 0 j 0 k 0 l 0 m0 n 0 i j k l m n i j k l m n (4.34)
ij k lmn
0 0 0 0 0 0  i 0 j 0 k 0  l 0 m0 n 0 δ i 0 i δ j 0 j δ k 0 k δ l 0 l δ m0 m δ n 0 n = ijk lmn ,

where we for pedagogical reasons have reinserted the involved sums, denoting
summation indices with a prime. The above expression equals the lhs of Eq.
(4.28). For the rhs of Eq. (4.33) we also recover the rhs. of Eq. (4.28) by noting
that each scalar product indeed becomes a kronecker delta, ie. a·x = ei ·el = δil
etc. This proves Eq. (4.28).

4.5 Tensor fields of any rank


All that has been said about transformation properties of tensors and pseudo-
tensors can be converted directly to that of tensor fields and pseudo-tensor
fields. One only needs to remember that all the (pseudo) tensor components
are now functions of the coordinates. To be specific, upon the cartesian coor-
dinate transformation, Eq. (4.11) the components of a tensor field T of rank r
transforms as
Ti01 i2 ···ir (x0 ) = ai1 j1 ai2 j2 · · · air jr Tj1 j2 ···jr (x) (4.35)
and the components of a pseudo tensor field P of rank r transforms as

Pi01 i2 ···ir (x0 ) = det(A)ai1 j1 ai2 j2 · · · air jr Pj1 j2 ···jr (x) (4.36)

57
The same transformation rules also hold true for a transformation between any
two orthonormal bases, for instance from a rectangular to a spherical basis.
One must bear in mind, however, than in this case the transformation matrix
elements, aij , will themselves be function of the position, aij = aij (x).
A tensor field is just a particular realization of the more general case con-
sidered in section 4.3.4, with a tensor A being a function of another tensor B,
A = A(B). Here, B = r. Therefore, we can apply Eq. (4.22) to demonstrate
that derivatives of a tensor field is another tensor field of one higher rank. Al-
though r is an improper vector, cf. section 2.3, C = ∂T ∂r will still be a proper
tensor. The reason is that in deriving the tensorial nature of C, Eq. (4.23), we
have used the tensorial nature of B only to show that
∂Bl1 l2 ···ls
= a j 1 l1 a j 2 l2 · · · a j s ls
∂Bj0 1 j2 ···js
However, this also holds true for any improper tensor in cartesian coordinates.
Specifically, for the position vector we have [r]R,l = xl , [r]R0 ,j = x0j and
∂xl
= ajl .
∂x0j
Consequently, if T is a tensor field of rank r, T(r) = (Ti1 i2 ···ir (r))R , then the
operation
∂T
= ∇T
∂r
gives a tensor field of rank r + 1 with the components

[∇T]i1 i2 ···ir ,j (r) = ∂j Ti1 i2 ···ir (r).
Specifically, ∇φ is a vector field when T = φ is a scalar field, c.f. section 2.6.4,
and ∇a is a rank two tensor field when T = a is a vector, cf. section 3.5. 8

Divergence
The fact that spatial derivatives of a tensor always leads to a new tensor of one
higher rank makes it easy to deduce the type of objects resulting from various
vector operations. For instance, if a(x) is a vector field then ∇a is a tensor field
of rank two. By contracting the two indices we therefore obtain a scalar, cf.
section 4.3.2
[∇a]i,i = ∂i ai a scalar quantity.
Thus, the divergence of the vector field is a scalar.
For a rank two tensor, T, ∇T is a rank 3 tensor, and we can perform two
different contractions, aj = ∂i Tij and bi = ∂j Tij , yielding the components of
two different vectors a and b. Only if T is symmetric will a = b. The direct
notation for the two operations are ∇ · T and ∇ · Tt , respectively.
8 Note, that the convention in tensor notation is to place the index of the spatial derivative

at the end of the tensor components, in conflict with the natural convention arising from the
direct notation which for –say a vector a– reads [∇a]ij = ∇i aj . However, in tensor notation
one always explicitly writes the index to be summed over, so no ambiguities arise in practice.

58
Curl
The tensor notation for the curl operator ∇ × a on a vector field a involves the
Levi-Civita symbol
[∇ × a]i = ijk ∂j ak .
If a is a polar vector then the rhs will be a double inner product between a
pseudo-tensor of rank 3 and two polar vectors yielding an axial (or pseudo)
vector field.

59
Chapter 5

Invariance and symmetries

60

You might also like