D

R
A
F
T
Chapter 1
INTRODUCTION
1.1 Historical perspective
Classical dynamics was developed and advanced without the use of geometry by Lagrange
(1736-1813) and Hamilton (1805-1865). A geometric view of classical dynamics is based on
the work of Riemann
1
(1826-1866), who conceived geometry in n-dimensions
2
. His work
was subsequently used by Darboux (1842-1917) who treated a dynamical system as a point
in n-dimensional space. This was followed by the development of tensor calculus by Ricci
(1853-1925) and Levi-Civita (1873-1941).
Tensor calculus was not received with great enthusiasm until the advent of the general
relativity theory of Einstein (1879-1955), which made heavy use of geometric methods.
The geometric treatment of dynamics has led to great new insights in the works of
Kolmogorov, Moser, Arnold and others. Other branches of physics (e.g., electromagnetism,
thermodynamics, continuum mechanics) have made use of geometric methods with success.
1.2 Euclidean and Riemannian geometry
Euclid (c.325-c.265) founded geometry by stating and admitting ve postulates in his famous
work, the Elements.
I. It is possible to draw a straight line from A to B.
II. It is possible to continue a nite straight line in a straight line.
1
student of Gauss
2
Time as a 4
th
dimension was explored by D Alembert (1717-1783) and Lagrange.
1
D
R
A
F
T
2 Introduction
III. It is possible to describe a circle with any center and radius.
IV. All right angles are equal.
V. If a straight line falling on two other straight lines makes the interior angles on the
same side less than two right angles, then the two line intersect, if extended indenitely
along that side.
insert gure
An alternative (more well-known) version of the fth postulate is due to Proclus: given
a point and a straight line that does not contain the point, there is only one straight line
through the point that never intersects the original line.
insert gure
Many centuries of frustration and little progress ensued until Gauss started thinking
about the consequences of doing away with the fth postulate (he never published his
thoughts). Those that did (albeit with limited success) were Lobachevski (1793-1856) and
Bolyai (1775-1856). Both of their works came in the 1820s but did not gain much attention
until 1867.
insert gure
In Euclidean geometry, the notions of
distance
angle
parallelism of lines
straightness of lines
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Euclidean and Riemannian geometry 3
parallel translation along a curve
make excellent sense, say on a at plane. Not all of these notions make sense on a curved
surface, e.g.,
insert gure
Why is parallel transport important? Recall the denition of derivative of a vector
function x(t) at t = t
0
:
v

(t
0
) = lim
t0
v(t
0
+ t) v(t
0
)
t
(1.1)
In that denition, it is implicit that we can compare the vectors v(t
0
) and v(t
0
+ t)
associated with two distinct points t
0
, t
0
+ t.
insert gure
This requires that we parallel transport one to the otherwhich is trivial for Euclidean
spaces, but less trivial for Riemannian spaces.
Where do we encounter non-at spaces in mechanics?
Some typical examples are:
(a) Planar pendulum The conguration space is a circle, S
1
.
(b) Planar double pendulum The conguration space is a two-torus, T
2
.
(c) Take a time-independent, n-dimensional Hamiltonian system associated with the func-
tion H(q
i
, p
i
), i = 1, 2, . . . , n, where
p
i
=
H
q
i
, q
i
=
H
p
i
(1.2)
(q
i
: generalized coordinates in R
n
, p
i
: generalized momenta in R
n
) Clearly

H =
n

i=1
_
H
q
i
q
i
+
H
p
i
p
i
_
= 0 (1.3)
therefore H = constant.
(2n1-dimensional surface embedded in a 2n-dimensional space) in phase space (q
i
, p
i
).
Version: April 5, 2010, 14:18 ME281
D
R
A
F
T
4 1.3. CLASSICAL VIEW OF TENSORS
insert gure
insert gure
(d) Shell structures
Need to determine deformation, strain, stress at every point P.
(e) Curved space-time in general relativity. 4-dimensional spacetime, studied for relativity
by Minkowski (1864-1909) special relativity: spacetime at general relativity: spacetime
curved (due to gravitation)
1.3 Classical view of tensors
Start with the classical view: a tensor is a mathematical object, whose representation with
respect to one coordinate system is related to its representation with respect to any other
coordinate system by means of a denite transformation law.
The independence from coordinate system choice is very important in formulating in-
trinsically meaningful mathematical statements of physical laws. To x our thoughts, take a
point P in n-dimensional Euclidean point space E
n
and parametrize it by a set of coordinates
{(x
1
, x
2
, . . . , x
n
) , x
n
R} which live in the associated Euclidean vector space E
n
.
insert gure
Also, take any other set of coordinates {(y
1
, y
2
, . . . , y
n
) , y
i
R, i = 1, 2, . . . , n} such that
there exists a C
1
-dieomorphism (i.e. a one-to-one C
1
function with C
1
inverse) y : E
n
E
n
y
i
= y
i
(x
1
, x
2
, . . . , x
n
) (1.4)
where y
i
is the i-th component of the function y. There should be open subsets of E
n
containing (x
1
, x
2
, . . . , x
n
) and (y
1
, y
2
, . . . , y
n
). Also, recall that, by the inverse function
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Classical view of tensors 5
theorem, the existence of a C
1
inverse of y is guaranteed if
det
_
y
i
x
j
_
= 0 (1.5)
in a neighborhood of P. In such a case, x
i
= x
i
(y
1
, y
2
, . . . , y
n
), where x
i
is the i-th component
of x = y
1
. Now, take a scalar function F : E
n
R and notice that its value depends only
on the particular point P in E
n
and not on the coordinates used to identify this point.
Therefore, there is a function f : E
n
R such that
F(P) = f
_
x
1
, x
2
, . . . , x
n
_
= (f x)
_
y
1
, y
2
, . . . , y
n
_
= g
_
y
1
, y
2
, . . . , y
n
_
(1.6)
where f x is the composition of f and x, dened by
(f x)
_
y
1
, y
2
, . . . , y
n
_
= f( x
_
y
1
, y
2
, . . . , y
n
_
) = f
_
x
1
, x
2
, . . . , x
n
_
(1.7)
Examples:
(a) Fix Q and dene F
1
(P) = |PQ| = the Euclidean distance between P and Q
(b) (P): mass density of a particle that at a xed time occupies point P.
We can identify F by the totality of components f, g, etc. relative to dierent coordi-
nate systems (x
1
, x
2
, . . . , x
n
), (y
1
, y
2
, . . . , y
n
), etc.
f
_
x
1
, x
2
, . . . , x
n
_
= g
_
y
1
, y
2
, . . . , y
n
_
= (1.8)
related by the transformation laws
g = f x = (1.9)
Classically, F is a 0
th
-order tensor in E
n
at a point. Likewise, start with
F(P) = f
_
x
1
, x
2
, . . . , x
n
_
= g
_
y
1
, y
2
, . . . , y
n
_
(1.10)
Version: April 5, 2010, 14:18 ME281
D
R
A
F
T
6 Introduction
where f is assumed to be in C
1
and take the partial derivatives
_
f
x
1
,
f
x
2
, . . . ,
f
x
n
_
. Using the
chain rule and summation convention,
g
y
1
=
n

j=1
f
x
j
x
j
y
1
=
f
x
j
x
j
y
1
g
y
2
=
n

j=1
f
x
j
x
j
y
2
=
f
x
j
x
j
y
2
.
.
.
g
y
n
=
n

j=1
f
x
j
x
j
y
n
=
f
x
j
x
j
y
n
(1.11)
or, in more compact notation
g
y
i
=
x
j
y
i
f
x
j
(1.12)
The above furnishes a transformation law between the component
_
f
x
1
,
f
x
2
, ,
f
x
n
_
and
_
g
y
1
,
g
y
2
, ,
g
y
n
_
of vectors (or 1
st
-order covariant tensors) in E
n
. Again, this tensor
can be identied by the totality of components related by transformation laws (1.12)
Since the choice of function F is arbitrary, the law (1.12) can be more descriptively
expressed as

y
i
=
x
j
y
i

x
j
(1.13)
In addition, starting from
y
i
= y
i
_
x
1
, x
2
, . . . , x
n
_
(1.14)
take the total dierential to obtain
dy
i
=
y
i
x
j
dx
j
(1.15)
The above furnishes another transformation law between the components (dx
1
, dx
2
, . . . , dx
n
)
and (dy
1
, dy
2
, . . . , dy
n
) of a dierent type of vectors (or 1
st
order contravariant tensors) at
a point in E
n
. Again, this special tensor can be identied by the totality of its components
related by (1.15).
(1.13): covariant transformation
(1.15): contravariant transformation
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Classical view of tensors 7
We can dene covariant, contravariant and mixed tensors of order greater than one. For
instance, a covariant tensor of order r is such that its components relative to the coordinate
systems (x
1
, x
2
, . . . , x
n
) and (y
1
, y
2
, . . . , y
n
) are related by
T
i
1
,i
2
,...,ir
_
y
1
, y
2
, . . . , y
n
_
=
x
j
1
y
i
1
x
j
2
y
i
2

x
jr
y
ir
S
j
1
,j
2
,...,jr
_
x
1
, x
2
, . . . , x
n
_
(1.16)
Such tensors are often denoted as
_
0
r
_
type. Likewise, for a contravariant tensor of order s,
or a
_
s
0
_
-type tensor
T
i
1
,i
2
,...,is
_
y
1
, y
2
, . . . , y
n
_
=
y
i
1
x
j
1
y
i
2
x
j
2

y
is
x
js
S
j
1
,j
2
,...,js
_
x
1
, x
2
, . . . , x
n
_
(1.17)
and for a mixed tensor of type
_
s
r
_
which is covariant to order r and contravariant to order s
T
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
_
y
1
, y
2
, . . . , y
n
_
=
y
i
1
x
k
1
y
i
2
x
k
2

y
is
x
ks
x

1
y
j
1
x

2
y
j
2

x
r
y
kr
S
k
1
,k
2
,...,ks

1
,
2
,...,r
_
x
1
, x
2
, . . . , x
n
_
(1.18)
Remarks:
The above equations involve two types of transformation occurring simultaneously:
(a) the covariant/contravariant transformations of the components, and (b) the coor-
dinate system transformation. As noted in Sokolniko, whatever the nature of the
covariant/contravariant transformation, it always depends on the coordinate system
transformation.
Assume that x
i
= y
i
, i = 1, 2, . . . , n, i.e. x is the identity transformation. Then
x
i
= x
i
_
y
1
, y
2
, . . . , y
n
_
= y
i
y
j
= y
j
_
x
1
, x
2
, . . . , x
n
_
= x
j
(1.19)
hence
x
i
y
j
=
_
1 : i = j
0 : i = j
=
i
j
(1.20)
and, also,
y
i
x
j
=
i
j
(1.21)
Hence
T
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
_
y
1
, y
2
, . . . , y
n
_
=
i
1
k
1

i
2
k
2
. . .
is
ks

1
j
1

2
j
2
. . .
s
js
S
k
1
,k
2
,...,ks

1
,
2
,...,r
_
x
1
, x
2
, . . . , x
n
_
= S
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
_
x
1
, x
2
, . . . , x
n
_
= S
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
_
y
1
, y
2
, . . . , y
n
_
(1.22)
Version: April 5, 2010, 14:18 ME281
D
R
A
F
T
8 Introduction
i.e., if the coordinate system transformation is the identity, then so is the tensorial
transformation. The above proof makes use of the substitution property of
i
j
, namely
that

i
j
S
j
=
i
1
S
1
+
i
2
S
2
+ . . .
i
n
S
n
= S
i
(1.23)
Let
y
i
= y
i
_
x
1
, x
2
, . . . , x
n
_
z
j
= z
j
_
y
1
, y
2
, . . . , y
n
_
(1.24)
be component forms of two coordinate transformation maps y and z
insert gure
such that
z
i
= z
i
_
y
1
, y
2
, . . . , y
n
_
= z
i
_
y
1
_
x
1
, x
2
, . . . , x
n
_
, y
2
_
x
1
, x
2
, . . . , x
n
_
, . . . , y
n
_
x
1
, x
2
, . . . , x
n
__
= z
i
_
x
1
, x
2
, . . . , x
n
_
(1.25)
Said dierently
z
i
= ( z
i
y) (x
1
, x
2
, . . . , x
n
)
= z
i
(x
1
, x
2
, . . . , x
n
)
_
z
i
= z
i
y (1.26)
It is easy, but tedious, to show that, if
S
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
_
x
1
, x
2
, . . . , x
n
_ (
s
r
)
T
k
1
,k
2
,...,ks

1
,
2
,...,r
_
y
1
, y
2
, . . . , y
n
_ (
s
r
)
U

1
,
2
,...,s
m
1
,m
2
,...,mr
_
z
1
, z
2
, . . . , z
n
_
(1.27)
In eect this means that the composition of two coordinate system transformations induces
the associated composite tensorial transformation. This property is referred to as an iso-
morphism between the coordinate transformation and the tensorial transformation.
If all components S
i
1
,i
2
,...,is
j
1
,j
2
,...,jr
(x
1
, x
2
, . . . , x
n
) = 0 then so are all components relative to any
other coordinate system.
dy
j
=
y
j
x
i
dx
i
=
y
j
x
i
F
i
A
dX
A
=
y
j
x
i
F
i
A

X
A
Y
B
dy
B
(1.28)
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Classical view of tensors 9
Example: The deformation gradient F in continuum mechanics, dx = FdX (in compo-
nents dx
i
= F
i
a
dX
A
). It maps a
_
1
0
_
-tensor to another
_
1
0
_
-tensor.
Tensor algebra can be naturally dened for
_
s
r
_
tensors. In particular, if
T
i
1
i
2
...is
j
1
j
2
...jr
_
y
1
, y
2
, . . . , y
n
_
=
y
i
1
x
k
1
y
i
2
x
k
2
. . .
y
is
x
ks
x

1
y
j
1
x

2
y
j
2
. . .
x
r
y
jr
S
k
1
k
2
...ks

2
...r
_
x
1
, x
2
, . . . , x
n
_
(1.29)
and
Q
i
1
i
2
...is
j
1
j
2
...jr
_
y
1
, y
2
, . . . , y
n
_
=
y
i
1
x
k
1
y
i
2
x
k
2
. . .
y
is
x
ks
x

1
y
j
1
x

2
y
j
2
. . .
x
r
y
jr
P
k
1
k
2
...ks

2
...r
_
x
1
, x
2
, . . . , x
n
_
(1.30)
then
T
i
1
i
2
...is
j
1
j
2
...jr
_
y
1
, y
2
, . . . , y
n
_
= Q
i
1
i
2
...is
j
1
j
2
...jr
_
y
1
, y
2
, . . . , y
n
_
(1.31)
satises the tensorial transformation rule. The zero tensor of type
_
s
r
_
is dened as a
_
s
r
_
-type
tensor whose components are zero relative to any coordinate system.
Likewise, scalar multiplication preserves the tensorial transformation rule. It is also easy,
but tedious, to establish that the outer product of an
_
s
r
_
and a
_
q
p
_
tensor yields a
_
s+q
r+p
_
-type
tensor. Instead of pursuing the full derivation, take a simple example of a
_
1
0
_
tensor
B
i
_
y
1
, y
2
, . . . , y
n
_
=
y
i
x
j
A
j
_
x
1
, x
2
, . . . , x
n
_
(1.32)
outer-multiplied by a
_
0
1
_
tensor
D
k
_
y
1
, y
2
, . . . , y
n
_
=
x

y
k
C

_
x
1
, x
2
, . . . , x
n
_
(1.33)
Dropping the explicit mention of dependence on coordinate system, take
B
i
D
k
=
y
i
x
j
x

y
k
A
j
C

(1.34)
which conrms that the result of the multiplication is a
_
1
1
_
-tensor. The multiplication rule
generalizes readily to products of
_
s
r
_
and
_
q
p
_
tensors.
The contraction of a
_
s
r
_
to a
_
s1
r1
_
tensor is accomplished by equating one covariant and
one contravariant index and summing with respect to it. Resorting, for simplicity to an
example, let the
_
2
1
_
tensor
T
i
1
i
2
j
=
y
j
x
k
1
y
i
2
x
k
2
x

y
j
S
k
1
k
2

(1.35)
be contracted to a
_
0
1
_
tensor as
T
ji
2
j
=
y
j
x
k
1
y
i
2
x
k
2
x

y
j
S
k
1
k
2

=
y
i
2
x
k
2

k
1
S
k
1
k
2

=
y
i
2
x
k
2
S
k
2

(1.36)
Version: April 5, 2010, 14:18 ME281
D
R
A
F
T
10 Introduction
or
T
i
2
=
y
i
2
x
k
2
S
k
2
(1.37)
Another example: a
_
1
1
_
tensor is contracted as T
i
j
T
i
i
to a scalar (i.e. a
_
0
0
_
tensor).
The inner product between two tensors involves the contraction of each one of them by
equating one (or more) covariant/contravariant indices. By example, if we have
T
i
1
i
2
j
=
y
i
1
x
k
1
y
i
2
x
k
2
x

y
j
S
k
1
k
2

_
2
1
_
-tensor (1.38)
Q
o

2
=
y
o
x
m
x
n
2
y

1
x
n
2
y

2
P
m
n
1
n
2
_
1
2
_
-tensor (1.39)
then
T
i
1
i
2
j
Q
j

2
=
y
i
1
x
k
1
y
i
2
x
k
2
x

y
j
y
j
x
m
. .

m
x
n
2
y

1
x
n
2
y

2
S
k
1
k
2

P
m
n
1
n
2
=
y
i
1
x
k
1
y
i
2
x
k
2
x
n
2
y

1
x
n
2
y

2
S
k
1
k
2
m
P
m
n
1
n
2
(1.40)
A tensor is symmetric in a pair of covariant or contravariant indices if the values of the
components remain the same when the indices are interchanged. For instance,
T
i
1
i
2
j
1
= T
i
2
i
1
j
1
(symmetry in the two contravariant indices)
T
i
1
j
1
j
2
j
3
= T
i
1
j
3
j
2
j
1
(symmetry in the rst and third covariant indices)
Symmetry is obviously preserved under coordinate transformation (take the dierence
which is zeroand transform). Likewise, a tensor is skew-symmetric in a pair of covariant
or contravariant indices, if the value of the components reverses sign when the indices are
interchanged. For example,
T
i
1
i
2
j
1
= T
i
2
i
1
j
1
(skew-symmetry in the two contravariant indices)
T
i
1
j
1
j
2
j
3
= T
i
1
j
3
j
2
j
1
(skew-symmetry in the rst and third covariant indices)
In all of the discussion of tensors up to this point, no explicit use is made of the Euclidean
structure of the spacein fact, the preceding analysis applies only to a point P E
n
for which
one can locally nd a coordinate system that can describe the point by way of a n-tuple of
reals (x
1
, x
2
, . . . , x
n
).
When one considers E
n
, then the notion of distance between two points P and P

, which
are innitesimally close to each other, can be given precise mathematical meaning in connec-
tion with the rectangular Cartesian coordinate system and Pythagoras formula: take any
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Classical view of tensors 11
coordinate system blanketing P and P

and assigning to them coordinates (x

in E
n
can be expressed
as
ds =

dz
i
dz
i
(1.43)
Recalling that
dz
i
=
z
i
x
j
dx
j
(1.44)
the same distance can be also expressed as
ds
2
=
z
i
x
j
z
i
x

dx
j
dx

= g
j
dx
j
dx

(1.45)
where
g
j
_
x
1
, x
2
, . . . , x
n
_
=
z
i
x
j
z
i
x

(1.46)
are the components of a symmetric covariant tensor of order 2, called the metric tensor.
Why does the set of all g
j
constitute a tensor? Since, in a Euclidean space, the distance
ds is invariant (i.e. unchanged by the choice of coordinate system), take a third system {y
i
},
such that
z
i
= z
i
_
y
1
, y
2
, . . . , y
n
_
= z
_
x
1
, x
2
, . . . , x
n
_
(1.47)
and
y
i
= y
i
_
z
1
, z
2
, . . . , z
n
_
(1.48)
Now
ds
2
=
z
i
x
j
z
i
x

dx
j
dx

=
z
i
y
j
z
i
y

dy
j
dy

_
ds is invariant (1.49)
Version: April 5, 2010, 14:18 ME281
D
R
A
F
T
12 Introduction
Therefore, take
y
i
= y
i
_
z
1
, z
2
, . . . , z
n
_
= y
i
_
z
1
_
x
1
, x
2
, . . . , x
n
_
, z
2
_
x
1
, x
2
, . . . , x
n
_
, . . . , z
n
_
x
1
, x
2
, . . . , x
n
__
= y
i
_
x
1
, x
2
, . . . , x
n
_
(1.50)
and
x
i
= x
i
_
y
1
, y
2
, . . . , y
n
_
(1.51)
Now, one merely needs to establish that
_
z
i
y
j
z
i
y

_
=
x
k
y
j
x
m
y

_
z
i
x
k
z
i
x
m
_
(1.52)
i.e., the tensorial transformation relation for
_
0
2
_
tensors. Indeed,
ds
2
=
z
i
y
j
z
i
y

dy
j
dy

=
z
i
x
k
z
i
x
m
dx
k
dx
m
=
z
i
x
k
z
i
x
m
_
x
k
y
j
dy
j
__
x
m
y

dy

_
(1.53)
hence
_
z
i
y
j
z
i
y

z
i
x
k
z
i
x
m
z
k
y
j
z
m
y

_
dy
j
dy

= 0 (1.54)
which produces the desired result.
Returning to our initial nding,
ds
2
= dz
i
dz
i
= g
j
dx
j
dx

(1.55)
for the Euclidean space E
n
, where
g
j
_
x
1
, x
2
, . . . , x
n
_
=
z
i
x
j
z
i
x

(1.56)
and note that the metric tensor is dened for the coordinate system {x
i
} always in connection
with the existing RCC system in E
n
. The following question arises: how about if the
background E
n
is unavailable? Said dierently: how about if we do not know a priori that
the space is Euclidean? Here, the distance between neighboring points can be dened again
by
ds
2
= g
ij
_
x
1
, x
2
, . . . , x
n
_
dx
i
dx
j
(1.57)
ME281 Version: April 5, 2010, 14:18
D
R
A
F
T
Suggestions for further reading 13
where g
ij
(x
1
, x
2
, . . . , x
n
) is a given symmetric tensor function of (x
1
, x
2
, . . . , x
n
). In this case,
z
k
x
j
z
k
x

= g
ij
_
x
1
, x
2
, . . . , x
n
_
(1.58)
is a system of 1 + 2 + . . . + n =
1
2
n(n + 1) non-linear partial dierential equations for n
unknowns z
k
(x
1
, x
2
, . . . , x
n
). If the system possesses a solution, then the space is Euclidean
(specically, it is E
n
). Otherwise, the space is Riemannian and g
ij
(taken to be symmetric
and positive-denite) is called a Riemannian metric.
Note that a
_
0
2
_
-type tensor is positive-denite if its components form a positive-denite
n n matrix.
1.4 Suggestions for further reading
Section 1.1
[1] F. Cajori. History of Mathematics. Third edition, Chelsea, New York, 1980. [This
book contains an excellent account of the history of pure and applied mathematics up
to the beginning of the twentieth century].
Section 1.2
[1] I.S. Sokolniko. Tensor Analysis: Theory and Applications. John Wiley, New
York, 1951. [A most useful reference to the classical treatment of tensors].
Version: April 5, 2010, 14:18 ME281