Você está na página 1de 177

Imperial College QFFF MSc

Differential Geometry 2015-16


Lecture Notes

Chris Hull
Oct 2015

Contents
1 Introduction
1.1 Heuristic definition of (real) manifold . . . . . . . . . . . . .
1.2 Example: Stereographic projections of a 2-sphere; the manifold S 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Open sets in Rm in the usual topology . . . . . . . . . . . .
1.4 Topological space . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Topological structure of an m-dimensional real manifold M
1.6 Example: circle S 1 . . . . . . . . . . . . . . . . . . . . . . .
1.7 Topological invariants and compactness . . . . . . . . . . . .
1.8 Formal definition of smooth m-diml real manifold . . . . .
1.9 Example: real projective space RP2 . . . . . . . . . . . . . .
1.10 Product manifold . . . . . . . . . . . . . . . . . . . . . . . .
1.11 Example: Torus . . . . . . . . . . . . . . . . . . . . . . . . .
1.12 Complex manifold . . . . . . . . . . . . . . . . . . . . . . . .
1.13 Example: S 2 as a complex manifold . . . . . . . . . . . . . .
1.14 Example: CPn complex projective space . . . . . . . . . . .
1.15 Examples of not being a manifold . . . . . . . . . . . . . . .
1.16 Manifolds with boundary . . . . . . . . . . . . . . . . . . . .
1

5
5

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

8
10
12
14
15
18
21
23
27
28
29
30
31
33
34

2 Differential maps, tangent vectors and tensors


2.1 Differentiable maps . . . . . . . . . . . . . . . .
2.2 Embedding and submanifolds . . . . . . . . . .
2.3 An example: CY 3 . . . . . . . . . . . . . . . . .
2.4 Diffeomorphism . . . . . . . . . . . . . . . . . .
2.5 Functions . . . . . . . . . . . . . . . . . . . . .
2.6 Curves . . . . . . . . . . . . . . . . . . . . . . .
2.7 Tangent vectors . . . . . . . . . . . . . . . . . .
2.8 An aside: Dual vector spaces . . . . . . . . . . .
2.9 Cotangent vectors . . . . . . . . . . . . . . . . .
2.10 Tensors . . . . . . . . . . . . . . . . . . . . . .
2.11 Contraction of tensors . . . . . . . . . . . . . .
2.12 Tangent vectors act on functions . . . . . . . . .
2.13 Directional derivatives . . . . . . . . . . . . . .
2.14 Differentials as Cotangent Vectors . . . . . . .
2.15 Tensors in a coordinates basis . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

3 Induced maps, tensor fields and flows


3.1 Induced maps: push-forward . . . . . . . . . . . . . . . . . .
3.2 Pull-back . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Vector, covector and tensor fields . . . . . . . . . . . . . . .
3.4 Tensor fields and induced maps . . . . . . . . . . . . . . . .
3.5 Induced maps and diffeomorphisms . . . . . . . . . . . . . .
3.6 Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7 Lie Derivative of a vector field . . . . . . . . . . . . . . . . .
3.7.1 Components of the Lie Derivative of a vector field . .
3.8 Lie Bracket of two vector fields . . . . . . . . . . . . . . . .
3.9 Commuting flows . . . . . . . . . . . . . . . . . . . . . . . .
3.10 Lie Derivative of a tensor field . . . . . . . . . . . . . . . . .
3.11 Example: Active coordinate transformations in GR and symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Differential forms
4.1 Cartan wedge product .
4.2 Exterior product . . . .
4.3 Differential Form Fields
4.4 Exterior derivative . . .
4.5 Pullback of forms . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
2

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

36
36
38
39
40
41
42
43
46
48
49
50
51
53
58
61

.
.
.
.
.
.
.
.
.
.
.

62
62
65
66
69
72
73
77
78
81
83
84

. 85

.
.
.
.
.

87
88
91
92
93
94

4.6
4.7
4.8
4.9
4.10
4.11
4.12
4.13
4.14
4.15
4.16

3d vector calculus . . . . . . . . . . . . . . . . . . . . . . . . . 94
Coordinate free definition of the exterior derivative . . . . . . 94
Interior product . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Lie derivative of forms . . . . . . . . . . . . . . . . . . . . . . 96
Closed and Exact forms . . . . . . . . . . . . . . . . . . . . . 97
Physical application: Electromagnetism . . . . . . . . . . . . . 98
Orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Top forms and volume forms . . . . . . . . . . . . . . . . . . . 100
Integrating a top form over a chart . . . . . . . . . . . . . . . 102
Integrating a top form over an orientable manifold . . . . . . . 104
Integration of r-forms over oriented r-dimensional submanifolds105

5 Stokes theorem and cohomology


5.1 Cycles and boundaries . . . . . . . .
5.2 Stokes theorem . . . . . . . . . . . .
5.3 Example. Gausss and Stokes law . .
5.4 de Rham cohomology . . . . . . . . .
5.5 Example: Cohomology of R . . . . .
5.6 Example: Cohomology of S 1 . . . . .
5.7 Cohomology and Topology . . . . . .
5.8 Stokes Theorem and Cohomology . .
5.9 Poincares Lemma . . . . . . . . . . .
5.10 Example: Electromagnetism (again!)
5.11 Poincare duality . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

106
106
107
108
109
111
111
113
113
115
116
117

6 Riemannian Geometry I: The metric


6.1 The metric . . . . . . . . . . . . . . . . .
6.2 Metric inner product . . . . . . . . . . .
6.3 Volume element . . . . . . . . . . . . . .
6.4 Epsilon symbol . . . . . . . . . . . . . .
6.5 Hodge Star . . . . . . . . . . . . . . . .
6.6 Inner product on r-forms . . . . . . . . .
6.7 Adjoint of d: d . . . . . . . . . . . . . .
6.8 3d vector calculus again. . . . . . . . . .
6.9 The Laplacian . . . . . . . . . . . . . . .
6.10 Ex. Electromagnetism again.... . . . . .
6.11 Hodge theory . . . . . . . . . . . . . . .
6.12 Harmonic representatives for cohomology

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

118
118
120
121
123
124
125
126
128
129
130
131
134

.
.
.
.
.
.
.
.
.
.
.

6.13 Maxwells Equations on a Compact Riemannian Manifold . . . 135


7 Riemannian Geometry II: Geometry
137
7.1 Induced metric and volume for a submanifold . . . . . . . . . 138
7.2 Length of a Curve and the Line Element . . . . . . . . . . . . 139
7.3 Integration of n-forms over oriented n-dimensional submanifolds140
7.4 Hypersurfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.5 Electric and Magnetic Flux . . . . . . . . . . . . . . . . . . . 145
7.6 Stokess Theorem in 3D . . . . . . . . . . . . . . . . . . . . . 145
7.7 Gausss Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.8 Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.9 Torsion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.10 Parallel transport and Geodesics . . . . . . . . . . . . . . . . . 152
7.11 Interpretation of Torsion . . . . . . . . . . . . . . . . . . . . . 153
7.12 Covariant derivative . . . . . . . . . . . . . . . . . . . . . . . 155
7.13 Metric connection . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.14 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.15 Holonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.16 Non-coordinate basis . . . . . . . . . . . . . . . . . . . . . . . 164
7.17 Connections and curvature in non-coordinate bases . . . . . . 169
7.18 Cartans structure equations . . . . . . . . . . . . . . . . . . . 171
7.19 Change of Basis and the Local Frame . . . . . . . . . . . . . . 174
7.20 Tangent Bundle . . . . . . . . . . . . . . . . . . . . . . . . . . 176

Introduction
A manifold is a space which locally looks like Rm (or Cm )
Calculus can be extended from Rm to the manifold. i.e. we have a
notion of differentiability. For a complex manifold, complex analysis
can be extended from Cm to the manifold, so we have a notion of
analyticity.

1.1

Heuristic definition of (real) manifold

We essentially require 2 things - an atlas which is smooth


1. Atlas: m-diml manifold M is covered by patches, and there are maps
which take patches of M into patches in Rm

Each patch of M, Ui , pairs with a map i to give a chart.


The maps i are invertible and
i : M Rm
p i (p)

(1)
(2)

A set of charts (Ui , i ) covering M, so that i Ui = M, is called an


atlas.
The inverse i1 maps coordinates (x1 , . . . , xm ) in Ui0 Rm onto Ui
M.
5

We explicitly give the maps i by giving


i : M Rm
p x(p) = (x1 (p), x2 (p), . . . , xm (p))

(3)
(4)

Example of an altas: An altas! M = The World (i.e. a 2-sphere),


Ui =regions of the world, Ui0 =pages in altas, i =some appropriate
coordinates eg. latitude and longitude.

2. Differentiability: transition from one patch to another is smooth.

If 2 patches Ui and Uj overlap, so Ui Uj 6= 0, then there is a map ji


so
ji : Rm (Ui0 ) Rm (Uj0 )
x y = ji (x) j

(5)
i1 (x)

(6)

for points such that p = i1 (x) Ui Uj . Recall that i is invertible.


ji is called the transition function.
Note that in the diagram, ij should be ji .
From the definition we see that ij is invertible since i , j are, and
furthermore 1
ij = ji .
Since this is a map from Rm Rm , our usual notions of differentiability
(smoothness) can be used.
Explicitly using coordinates where x = (x1 , . . . , xm ) and y = (y1 , . . . , ym ),
we can write the transition function y = ji (x) as ya = ya (x1 , . . . , xm )
where a = 1, . . . , m.
Smoothness requirement: Each ya (x1 , . . . , xm ) for a = 1, . . . , m is C
in each argument xa .

1.2

Example: Stereographic projections of a 2-sphere;


the manifold S 2

Think of S 2 as embedded in R3 as x2 + y 2 + z 2 = 1. Note that x, y, z provide


a unique labelling of points but are not coordinates (there are obviously too
many of them!).

Consider projecting a point (x, y, z) on the S 2 into the plane z = 0 from the
North pole. Let the intersection point be (X, Y, 0). Geometrically,
X=

x
1z

Y =

y
1z

(7)

and these define a map N : S 2 R2 .


Now X, Y , which are functions of the point p = (x, y, z) look like good coordinates. BUT there is a problem at the North pole i.e. p = (0, 0, 1) where
obviously something looks suspicious. The pole is mapped to infinity in R2 ,
but which infinity? eg. is it at (0, ) or (, 0), or (0, )? Hence the map
N is not invertible due to the inclusion of the N-pole.
(technically the map fails to be a homeomorphism - see later...)
So take 2 charts: project onto plane z = 0 from N-pole for p UN , where UN
excludes the N-pole, and project from S-pole onto plane z = 0 for p US ,
where US excludes the S pole.

x
y
Y0 =
(8)
1+z
1+z
We should make these charts cover the S 2 and overlap eg. UN = S 2
{N pole}, US = S 2 {Spole}
X0 =

Differentiability?
Y0
Y
=
0
X
X

X2 + Y 2 =

1+z
1
= 02
1z
X + Y 02

(9)

Transition functions explicitly given as


X 0 = X 0 (X, Y ) =

X2

X
+Y2

Y 0 = Y 0 (X, Y ) =

define the map

X2

Y
+Y2

(10)

This is smooth, C , in overlap (note that this is true as both poles are excluded, where things are not smooth).
Remark: We thought of S 2 as a subset of R3 in order to simply label the
points. This is a common method of explicitly constructing manifolds, although it is not necessary to do this. A very important point is that an atlas
provides an intrinsic definition of a manifold eg. an Atlas is as good as a
globe, and describes the same Earth 2-sphere.

1.3

Open sets in Rm in the usual topology

The canonical example of an open set in Rm is an open ball.


The open ball B (p0 ) of radius  centred on a point p0 in Rm are the points p in
Rm whose distance from p0 is less than . If p has coordinates x(p) and p0 has
coordinates x0 , then the distance between p and p0 is the Euclidean distance
|x(p) x0 | (i.e. for a point y = (y1 , . . . , ym ), |y|2 = (y1 )2 + (y1 )2 + . . . + (ym )2 )
Then
B (x0 ) = {p Rm with |x(p) x0 | < }
(11)

10

This is the interior of a sphere sm1 in Rm .


Example: For m = 1 this gives the familiar open interval (x0 , x0 + ).
Definition of an open set in Rm in the usual or metric topology:
Set of points forms an open set in Rm if for any point in the set we can find
an open ball of non-zero radius, centred on the point and also contained in
the set. i.e. there are no boundary points.

Hence: all open sets in Rm can be formed by a union of open balls.


This notion of open set defines topology on Rm - called the metric topology
as it used a distance measure to define the open sets.

11

1.4

Topological space

Topological space: A set of points X, together with a set of subsets J =


{Oi |i I} so that,
1. 0 J and X J

( 0 - empty set )

2. aA Oa J for any A I, finite or infinite.


3. bB Ob J for any finite subset B I.
Then the {Oi } are called open subsets of X.
Examples:
trivial topology on any set of points X: J = {0, X} - not very useful
discrete topology on any set of points X: J = all subsets of points - also
not very useful
metric topology on Rm : J =set of all possible unions of open balls.
metric topology on R: J =set of all possible unions of open intervals. An
open interval (a, b) is the set of all points x R with a < x < b.
Continuous map: Given two topological spaces, a continuous map f : X
Y is a map from points in X to points in Y such that the inverse image of
an open set in Y is an open set in X.
(Note: For a map g : A B the inverse image of a subset C B is the set
{a A|g(a) C}. This is not related to the inverse map g 1 .)
Homeomorphism: invertible continuous map between topological spaces
X, Y such that the inverse map is also continuous. Then open sets in X are
mapped to open sets in Y and vice versa.
If for topo spaces X, Y there exists a homeomorphism we say that X, Y are
homeomorphic - then they have the same topology.
Intuitively: we can continuously deform one into the other.
Closed Set: A subset U of X is closed if the complement of U , X \ U , is
open.
12

Examples:
metric topology on R: A closed interval is closed. A closed interval [a, b] is
the set of all points x R with a x b.
metric topology on Rm : Closed balls are closed. A closed ball of radius 
centred on point x0 in Rm is
C (x0 ) = {p Rm with |x(p) x0 | }

(12)

Hausdorff (topological smoothness) :


for any 2 distinct points x, x0 there exists open subsets Ux , Ux0 containing
x, x0 respectively s.t. Ux Uy = 0.

13

1.5

Topological structure of an m-dimensional real manifold M

1. M is a Hausdorff topological space with open sets J


2. M has an Altas; a family of pairs (the charts) {(Ui , i )} where Ui J
such that:
(a) the charts cover M so i Ui = M
(b) the i are homeomorphisms from Ui into open subsets Ui0 of Rm .
The set of open sets {Ui } is called an open cover. This means it is a set of
open sets that covers M, so that each point of M is in at least one of the
open sets Ui in the cover. It is a subset of the set of all open sets J.
Thus locally our manifold is homeomorphic to (looks like) Rm .

14

1.6

Example: circle S 1

Can think of S 1 as embedded in R2 so x2 + y 2 = 1.

Take angular coordinate on circle, x = cos , y = sin . This gives a chart


(U, ) where U is given by (x, y) for = (0, 2), and the map is,
: S1 R
p = (x, y)

(13)
(14)

BUT this misses the point on S 1 corresponding to = 0. How annoying.


(U, ) does not provide an altas as it doesnt cover all S 1 .

15

Try U given by (x, y) for = (0, 2 + ) with  > 0. This definitely now does
cover the circle. However now is not a homeomorphism!

= x and = 2 + x for x <  are the same point on S 1 so clearly not


invertible.
So even for a circle we actually need 2 charts to provide an atlas. eg.
U1 :
1 :

= (0, 2)
p = (x = cos , y = sin )

(15)
(16)

0 = (, +)
p = (x = cos 0 , y = sin 0 ) 0

(17)
(18)

and
U2 :
2 :

16

This now covers S 1 , and on the intersections = 0 (0 < 0 < ) or = 0 +2


( < 0 < 0) so both transition functions are clearly smooth.

17

1.7

Topological invariants and compactness

Topological invariants are quantites that depends only on the topological


structure of a manifold.
A famous example is the Euler character. If we think of a 2d surface in 3d
space and triangulate it, then the quantity
= number(f aces) number(edges) + number(points)

(19)

is independent of the triangulation and measures the topology. eg. (S 2 ) =


2, (T 2 ) = 0, and for a Riemann surface of genus G, (Riem) = 2 2 genus.
Another famous example is given by the Betti numbers br , r = 0,P
1, . . . , m for
r
an m-dimensional manifold. The Euler number in general is, = m
r=0 (1) br
A basic topological invariant is compactness. A manifold may be compact
(eg. S 1 ) or non-compact (eg. R).
Definition: A topological space is compact if for every open cover there
exists a finite subset (possibly the whole set itself) which is also an open
cover.
Example. S 1 is compact. Some examples of covers:
take cover {U1 , U2 }. then finite subset {U1 , U2 } is indeed a cover.

take cover {U1 , U2 , U3 }. then finite subset {U1 , U2 } or {U1 , U2 , U3 } are covers.

18

Example. R1 is not compact


take cover {U1 , U2 }. then finite subset {U1 , U2 } is indeed a cover.

BUT for the following cover with unit length open sets Ui , there is no finite
subset which covers R1 , so it is not compact.

Example. An open interval (a, b) in R1 is not compact. It is covered by the


open intervals
Un = (a, b L/n),

n = 2, 3, 4, . . . ,

L=ba

but no finite subset of these covers the interval.


(Heine-Borel) Theorem: If a topological space, X, is a subset of Rm s.t.
it is a closed subset (i.e. its complement is open) and is bounded in extent
19

in Rm then X is compact.
A subset of Rm is bounded in extent if the distance between any two points
is less than some bound R, with the same bound for all pairs of points, i.e.
a subset of X Rm is bounded if there is some positive real number R
such that |x x0 | < R for all points x, x0 X. A subset of Rm is closed if
its complement in Rm is open, i.e. the set of points {x Rm : x
/ X} is open.
The closed interval [a, b] is compact, as it is closed and bounded.
Example. S 2 is compact. It can be embedded as a subset in R3 as x2 + y 2 +
z 2 = 1, so all point in S 2 lie in interval x, y, z [1, 1] so extent of S 2 is
bounded.

20

1.8

Formal definition of smooth m-diml real manifold

1. M is a Hausdorff topological space with open sets J


2. M has an Altas: a family of pairs (the charts) {(Ui , i )} where Ui J
such that:
(a) the charts cover M so i Ui = M
(b) the i are homeomorphisms from Ui into open subsets Ui0 of Rm .
3. Given Ui Uj 6= 0, then the map (the transition function) ij = i j1
from Rm Rm is C , smooth.
The conditions 1), 2) imply that M can locally topologically smoothly be
thought of as Rm .
The condition 3) implies that M can locally differentially smoothly be
thought of as Rm .
The transition function is a map ij from j (Ui Uj ) Rm to i (Ui Uj )
Rm :
ij : j (Ui Uj ) i (Ui Uj )
A point p Ui Uj has coordinates x(i) = i (p) in the (Ui , i ) chart and
coordinates x(j) = j (p) in the (Uj , j ) chart. The two coordinate systems
are related by
x(i) = ij (x(j) )
Note that x(i) Rm is an m-vector. Introducing a coordinate index =
1, 2, . . . , m, we have coordinates x(i) = (x1(i) , x2(i) , . . . , xm
(i) ) for points in Ui ,

and similarly we have coordinates x(j) for points in Uj . Then for each
there is a transition function ij , so that we have
x(i) = ij (x(j) )
.

Note: any open set of a manifold is itself a manifold!

21

Remark:
A manifold typically admits (infinitely) many different altases. This simply
expresses the freedom in choosing a covering of open sets, and coordinates
on those sets.
Two altases {(Ui , i )}, {(Ui0 , i0 )} for M are compatible if the union of these
is also an atlas for M. The equivalence class of all compatible atlases for M
constitute the differentiable structure of M.

Remark:
In the definition of a smooth manifold, the transition functions are smooth,
i.e. infinitely differentiable, C . This can relaxed to define a C r manifold
as one with transition functions that are C r , i.e. they can be differentiated
r times and the rth derivative is continuous.

22

1.9

Example: real projective space RP2

set of undirected lines through the origin in R3


Geometrically:

A directed line is labeled by a point on S 2 . Hence RP2 is S 2 with its antipodal points identified

Take point in R3 (x1 , x2 , x3 ) which labels a line provided 6= (0, 0, 0), i.e. take
point in R3 {0}.
However, (x1 , x2 , x3 ) labels the same line as (x1 , x2 , x3 ) for any R.

23

Hence RP2 R3 {0} with the identification of points (x1 , x2 , x3 ) (x1 , x2 , x3 )


for all R {0}.
The labels (x1 , x2 , x3 ) are sometimes called the homogenous coordinates.
Now we want coordinate charts on RP2 .
Take chart:
U1 :
1 :

all p RP2 s.t. x1 6= 0


RP2
R2


x2 x3
,
= (a1 , a2 )
p = (x1 , x2 , x3 )
x1 x1

(20)
(21)
(22)

This correctly encodes the identification (x1 , x2 , x3 ) (x1 , x2 , x3 ).


Note we must exclude all points with x1 = 0 or else 1 not a homeomorphism.
[If one tried to extend U1 to contain points with x1 = 0, then 1 would formally take such points to coordinates with (a1 , a2 ) both infinite. Then there
are no points in R2 with finite coordinates (a1 , a2 ) that correspond to points
with x1 = 0. This means there is no region U 0 R2 for which 1 can extend
to a homeomorphism between a region containing points with x1 = 0 and U 0 .]
So require further charts:
U2 :
2 :

all p RP2 s.t. x2 6= 0


RP2
R2


x1 x3
,
= (b1 , b2 )
p = (x1 , x2 , x3 )
x2 x2

(23)
(24)

all p RP2 s.t. x3 6= 0


RP2
R2


x1 x2
p = (x1 , x2 , x3 )
,
= (c1 , c2 )
x3 x3

(26)
(27)

(25)

and
U3 :
3 :

24

(28)

Now the charts {(U1 , 1 ), (U2 , 2 ), (U3 , 3 )} form an altas covering RP2 .
Check the transition functions are smooth:
In overlap p U1 U2 we have,

(a1 , a2 ) =

21 : R2 R2



x1 x3
x2 x3
,
(b1 (a1 , a2 ), b2 (a1 , a2 )) =
,
x1 x1
x2 x2

(29)
(30)

Hence b1 = a11 and b2 = aa21 . This is indeed C smooth, since a1 6= 0 in U2


and hence in the overlap.

There are lots of overlaps!


Generalisation to RPn .
Now RPn set of undirected lines in Rn+1 .
Label a line using homogeneous coordinates (x1 , x2 , . . . xn+1 ) in Rn+1 {0}
and identifying (x1 , . . . , xn+1 ) (x1 , . . . , xn+1 ) for all R {0}.
Construct (n + 1) charts:
Um :
m :

all p RPn s.t. xm 6= 0


RPn
Rn

(31)
(32)
25


p = (x1 , . . . , xn+1 )


xm1 xm+1
xn+1
x1 x2
,
,...,
,
...,
(33)
xm xm
xm
xm
xm

Transition functions from Ui0 Uj0 are again smooth. [Ex for reader!]

26

1.10

Product manifold

If X is an m-diml manifold with atlas {(Ui , i )} and Y is an n-diml manifold


with atlas {(Vj , j )}, then X Y is the product manifold, an (m + n) diml
manifold with altas {(Wij , ij )} defined as
points: a point in X Y is labelled as (p, q) with p X and q Y
patches: (p, q) Wij iff p Ui and q Vj
maps:
ij : X Y Rm+n
(p, q) (i (p), j (q))
i.e. vector (x1 , . . . , xm , y1 , . . . , yn ) Rm+n .
Obviously satisfies manifold conditions.

27

(34)
(35)

1.11

Example: Torus

Torus T n = S 1 . . . S 1 , product of n circles


eg. T 2 is homeomorphic to the surface of a donut/bagel

28

1.12

Complex manifold

An m-diml complex manifold, M, is defined by:


1. (Hausdorff) topological space with open sets J
2. Atlas of charts {(Ui , i )} where Ui J such that:
(a) the charts cover M so i Ui = M
(b) the i are homeomorphisms from Ui into open subsets Ui0 of Cm .
3. Given Ui Uj 6= 0, then the map (the transition function) ij = i j1
from Cm Cm are analytic (or holomorphic) complex functions.
Explicitly in coordinates: a point in Ui0 = (z1 , . . . , zm ), and a point in Uj0 =
(u1 , . . . , um ). Then
ij : Cm Cm
(z1 , . . . , zm ) (u1 (z), . . . un (z))

(36)
(37)

so that all the ui (z1 , . . . , zn ) are analytic (holomorphic) functions in their


arguments zi .
Recall: An analytic (holomorphic) function f (z1 , . . . , zn ) is a function that
is complex differentiable i.e. any derivatives wrt some zi depend only on the
position z1 , . . . , zn that the derivative is evaluated at, and these derivatives
are finite. A result of this is that the function only involves zi and not zi
eg. f (z1 , z2 ) = 1/z1 + 1/z2 is analytic away from z1 = 0 and z2 = 0, while
f (z1 , z2 ) = |z1 |2 + |z2 |2 is not. Recall that an analytic function obeys the
Cauchy-Riemann relations; write zi = xi + iyi with real xi , yi . Then an
analytic function f (z1 , . . . zn ) = fr (z) + ifi (z) obeys,
fi
fr
=
xi
yi

and

fi
fr
=
,
xi
yi

(38)

Note that we may always think of an m diml complex manifold as a 2m diml


real manifold by defining an altas using twice the number of real coordinates
with zi = xi + iyi .

29

1.13

Example: S 2 as a complex manifold

Recall: real manifold S 2 is defined by atlas of 2 charts {(UN,S , N,S )}. Coor0
dinates in R2 are (XN,S , YN,S ) with UN,S
being defined by XN,S , YN,S being
finite. Then the transition functions were;
XN,S =

XS,N
,
2
+ YS,N

YN,S =

2
XS,N

YS,N
2
+ YS,N

2
XS,N

(39)

Now take complex coordinates ZN,S = XN,S iYN,S which define two complex
charts - again ZN,S is finite (i.e. the N and S pole are excluded in the N and
S chart respectively). We then see the transition functions are simply,
ZN =

1
ZS

(40)

which is analytic in the overlap of the 2 charts. (Note the overlap region
excludes ZS = 0).

30

1.14

Example: CPn complex projective space

CPn the set of undirected complex lines through origin in Cn+1 .


[A complex line is z = a + b, for a, b, C, so it is 2 real dimensional!]
The definition is exactly as for RPn except x0 s z 0 s.
Take homogeneous coordinates (z1 , . . . , zn+1 ) in Cn+1 {0} where we identify (z1 , . . . , zn+1 ) (z1 , . . . , zn+1 ) for any C {0}.
Require (n + 1) charts:
Um :
m :

all p CPn s.t. zm 6= 0


CPn
Cn


zm1 zm+1
zn+1
z1 z2
, ,...,
,
...,
p = (z1 , . . . , zn+1 )
zm zm
zm
zm
zm


(m)
= Z1 , . . . Zn(m)

(41)
(42)
(43)
(44)

Then e.g.


(1)
Z1 , . . . Zn(1)


=

z2
zr
zn+1
,..., ,...,
z1
z1
z1



 z z
zr
zn+1
1
3
(2)
(2)
Z1 , . . . Zn =
, ..., ,...,
z2 z2
z2
z2
Transition functions eg. for U1 , U2 overlap:




(1)
(2)
(1)
(1)
(2)
(1)
21 : Z1 , . . . , Zn Z1 (Z ), . . . , Zn (Z )

(45)

with,
(2)

Z1 =

,
(1)

Z1

(1)

(2)

Zi

Zi

(1)

for i 6= 1

(46)

Z1

Hence the transition functions are indeed analytic on the overlaps (where
(1)
Z1 6= 0).
Note: Consider CP1 . This consists of lines through the origin in C2 , specified
by homogeneous coordinates (z1 , z2 ). We have 2 patches. The patch U1 has
31

z1 6= 0, and we define a single complex coordinate Z (1) = z2 /z1 , so Z (1) 6= 0


in U1 . The patch U2 has z2 6= 0, and we define a single complex coordinate
Z (2) = z1 /z2 , so Z (2) 6= 0 in U2 . Then we have a manifold with one complex
dimension (corresponding to 2 real dimensions) with an atlas of two charts.
In the overlap U1 U2 , Z (1) 6= 0 and Z (2) 6= 0, and the transition function is
Z (2) = Z1(1) . But this is exactly the same as our S 2 example earlier, replacing
Z (1) = 1/ZN and Z (2) = 1/ZS . Hence the complex manifold S 2 is simply
CP1 . This is special for CP1 . For n > 1, CPn is not s sphere.

32

1.15

Examples of not being a manifold

Any space with a boundary eg. [0, ) or a [0, ) S 1 etc....

Some simple topological spaces

Orbifolds: ex. R3 /Z2 defined by taking R3 and identifying points by


the Z2 action (x, y, z) (x, y, z). This is not a manifold. The
origin is fixed under this action and leads to an orbifold singularity.
[A space with manifold structure everwhere except at distinct points
where there is a singularity which is locally of the type Rm / for some
discrete group is called an orbifold.]

33

1.16

Manifolds with boundary

We can encompass manifolds with boundary by modifying our definition of a


manifold so that the maps i are homeomorphisms into H m = {(x1 , . . . , xm )
Rm |xm 0}.

The boundary of H m is the space Rm1 with xm = 0, which we denote H m .


The open sets of H m are given by the intersection of open sets of Rm with H m ,
so that if U is an open set of Rm , then V = U H m is an open set of H m . An
open ball lying entirely in H m is then open in H m . A half-ball with centre p
lying in H m is the intersection B (p) H m . This is the hemisphere given by
the part of B (p) with xm 0, with boundary B (p)H m that is an m1 dimensional closed ball in H m . For example if p is the point with co-ordinates
(0, 0, . . . 0), then the half ball is the region {x Rm : |x| <  and xm 0}
with boundary the region {x Rm : |x| <  and xm = 0}. For m = 3,
the boundary of half of the n-dimensional ball is a disc, the 2-dimensional
interior of a circle. General open sets are formed by taking arbitrary unions
of open balls in Rm , and then taking the intersection with H m .
The set of points in M that are mapped to the boundary of H m form the

34

boundary of M, denoted M. The boundary points of a manifold form a


manifold themselves, or a disconnected sum of manifolds if the boundary has
more than one component.
Consider an open set Ui of M, which contains part of the boundary, so that
Ui M 6= 0. The map i takes Ui to an open set Ui0 of H m . The coordinates (x1 , . . . , xm1 , xm ) of a point p M are x(p) = i (p), and points in
the boundary p M Ui ) will have coordinates (x1 , . . . , xm1 , 0).
The transition functions ij are required to be smooth maps from H m to
H m.

35

Differential maps, tangent vectors and tensors

Manifolds allow calculus to be extended from Rm onto them - hence the differential in differential geometry. We shall now explore how this happens.

2.1

Differentiable maps

Define a map f ;
f:

M N
p q = f (p)

(47)

where M is an m diml manifold with altas {(Ui , i )} and N is an n diml


with altas {(Va , a )}.
Let p Ui and q Va with q = f (p).

Now the map,


a f i1 :
36

Rm R m

(48)

and so we can use usual calculus. If we have coordinates {x } on Ui0 and


{y } on Va0 then explicitly,
y (x ) = a f i1 (x )

(49)

for each = 1, . . . , n.
The map f is smooth at p if the (n-vector of) functions a f i1 are C
(smooth) in their arguments x . The map is said to be C r if the functions
a f i1 are C r . Sometimes (as in e.g. Nakaharas book) smooth maps
are referred to as differentiable.
For complex manifolds, the map f is analytic at p if the functions a f i1
are analytic in their arguments x .
Consider a point p in an overlap of charts p Ui Uj . Take coordinates
{x } in Ui0 and {y } in Uj0 . For p Ui Uj , x = i (p) and y = j (p), with
x = ij (y) where ij = i j1 .

Now a f i1 (x) = a f i1 (ij (y)) = a f j1 (y).


If a f i1 (x) is smooth, then since the transition functions are smooth,
then we are guaranteed that a f i1 (y) is also smooth. i.e. the manifold
structure ensures that if our map is smooth in one coordinate system, it will
be in any other.

37

2.2

Embedding and submanifolds

Consider a smooth map f : M N where dimM dimN .


The map f defines an embedding of the manifold M into N if the map f is
one-to-one.
Then the image f (M) is a submanifold of N that is diffeomorphic to M.

Here is an example where the map does not define an embedding as it is not
one-to-one.

[ Technically this is called an immersion, and the map f which we meet


later is one-to-one. ]
Whitney Embedding Theorem Any real smooth m-dimensional manifold
can be smoothly embedded in R2m .
Note that some manifolds can be embedded in Rn for some n < 2m, but 2m
is the smallest dimension that will work for all manifolds.

38

2.3

An example: CY 3

Low energy string theory reduces to a theory of gravity and matter in 10


dimensions.
Simple supersymmetric vacuum solutions giving realistic phenomenology
have a spacetime manifold R4 CY3 (i.e. a product manifold).
Here note that R4 refers to Minkowski spacetime, R1,3 [we will not mention
metrics for a while, and as a manifold Minkowski spacetime is simply R4 ]
CY3 is a Calabi-Yau manifold, a compact 3 complex dimensional manifold.
In fact there are many of them, the only restriction being a topological one
(vanishing first chern class which implies you can always find a metric with
vanishing Ricci tensor).
The fact that CY3 is compact physically allows low energy observers to only
see 4 spacetime dimensions.
All these CY3 manifolds can be thought of as submanifolds of CPn for some
(n > 3).
For example:
The Quintic - studied a lot - is embedded in CP4 as z15 + . . . z55 = 0, where
zi are homogeneous coordinates for CP4 .
Ex. Have a think about what that looks like!

39

2.4

Diffeomorphism

If a smooth map f : M N is invertible and the inverse map f 1 is also


smooth then f is a diffeomorphism.
For such a diffeomorphism f we say M and N are diffeomorphic, meaning
they are the same manifold i.e. M N as manifolds. [Recall for topo spaces
X Y if there exists a homeomorphism between them.]
We denote the set of diffeomorphisms of a manifold M as diff(M), so any
f diff(M) takes,
f :M M
p p0 = f (p)

(50)

We may think of these diffeomorphisms as smooth relabelings of points.


Obviously a diffeomorphism also defines a homeomorphism at the level of the
topological spaces M and N . But two manifolds with the same topology are
not necessarily diffeomorphic! Only seem to be highly non-trivial examples
- eg R4 due to Donaldson. The 7-sphere with the usual topology allows 28
different differentiable structures, while the 11-sphere allows 992 different
differentiable structures. This is rather unusual in most cases, there is only
one differentiable structure.

40

2.5

Functions

The simplest smooth maps. A function f is a smooth map such that,


f:

M R
p f (p)

(51)

Using a chart (Ui , i ) with coordinates x in Ui0 we have a coordinate representation of a function; for a point p Ui ,
f i1 : Rm R
x = i (p) f i1 (x)

(52)

Denote set of functions on M as F(M).


This inherits the usual ring structure from R. Namely we may add/multiply
two functions to get a third simply by adding/multiplying their values at all
points;
(f + g)(p) = f (p) + g(p)
(f g)(p) = f (p)g(p)
A constant function is one where f (p) = c for all p M, for some c R.
41

2.6

Curves

Essentially the opposite of a function! An open curve C is a smooth map


such that;
C: R M
(a, b) p = C()

(53)

In a chart (Ui , i ) with coordinates x in Ui0 we may write,


i C :

R Rm
x = i C()

(54)

An open curve has ends. One can also consider closed curves as maps
S 1 M.

42

2.7

Tangent vectors

Consider an m-dim manifold M and a point p M and two curves C1 , C2


passing through p, so that,
p = C1 (0) = C2 (0)

(55)

We can use the fact that our manifold locally looks like Rm to say whether
or not the two curves have the same direction and speed at p. Taking a
chart covering the neighbourhood of p, (U, ), with coordinates x then we
say that the curves are tangent at p iff;




d
d

x (C1 ())
=
x (C2 ())
(56)
d
d
=0
=0

d
i.e. the curves are tangent to each other at p. We note that d
x (C1,2 ()) =0
Rm , and hence the direction and speed are governed by a point in Rm .

We define the equivalence at p, denoted C1 C2 , iff C1 , C2 are tangent at p.


Note that if two curves are tangent in one coordinate chart, they are tangent
in any other coordinates i.e. our definition of this equivalence is coordinate
independent.
Tangent vector: Geometrically we then define a tangent vector, Vp , at the
point p to be the class of all curves passing through p which are tangent to
43

each other. Given a curve C through p, we denote this equivalence as [C].


We say that C is a representative of the class.
Tangent space: We then define the tangent space at p as the set of all tangent
vectors as p. We denote the tangent space Tp (M).
This tangent space has a linear/vector structure, again inherited from the
fact that the manifold is locally Rm . Consider a point p in a chart (U, ),
where w.l.o.g. we choose the homeomorphism, , so that (p) = 0 Rm .
Then given a, b R, and two tangent vectors Vp,1 , Vp,2 Tp (M) with representatives C1,2 () so that Vp,1 = [C1 ], Vp,2 = [C2 ], we define,
Vp,3 = a Vp,1 + b Vp,2 Tp (M)

(57)

where Vp,3 = [C3 ()] and,


C3 () 1 [a ( C1 ()) + b ( C2 ())]

(58)

with the righthand side being defined from the usual linearity of Rm . Whilst
this definition uses a particular chart, one obtains the same curve C3 whichever
chart one uses, and hence the definition is independent of the local choice of
coordinates.
More explicity we could write this relation as






d
d
d

x (C3 ())
=a
x (C1 ())
+b
x (C2 ())
.
d
d
d
=0
=0
=0

(59)

Hence we see that the tangent space Tp (M) is actually isomorphic to the
vector space Rm . Thus we can write any element as
Vp = v ep,()

(60)

where ep,() Tp (M) are a set of m basis tangent vectors (each with representative curves C() ), and v Rm are the components of the tangent vector
in this basis.
As for any vector space, the components in one basis are related to those in
another by an element of the general linear group. If we have two bases
{ea } and {e0i }, then these are related as e0i = Ai a ea , where the matrix
44

Ai a GL(m, R) (i.e. is an invertible m m matrix - since both are bases


this must be invertible). Then since Vp = V 0i e0i = V a ea then V 0i Ai a = V a
relates the components.

45

2.8

An aside: Dual vector spaces

Take a vector space V , with basis {ei } so v = v i ei V .


Consider a linear function, f , on V ;
f:

v f (v) =

v i f (ei )

(61)

since it is linear. Hence f is determined by knowing the set {f (ei )}.


The space of linear functions is also a vector space due to the linear structure;
f3 (v) = (a1 f1 + a2 f2 )(v) = a1 f1 (v) + a2 f2 (v)

(62)

for a1 , a2 R. Hence given linear maps defined by {f1 (ei )} and {f2 (ei )} we
can construct a third, {a1 f1 (ei ) + a2 f2 (ei )}.
This is called the dual vector space to V , denoted V . Note that since to
define linear map we require {f (ei )}, we must have dim(V ) = dim(V ).
Take basis {ei } for V so f V is f = fi ei (note the position of indices)
Make the choice that basis {ei } is dual to the basis {ei }: i.e. ei (ej ) = i j .
One can always choose this basis.
Note then that under a basis transformation, e0i = Ai a ea for A GL, then
we require a corresponding transform of the dual basis, e0i = (A1 )ai ea .

46

We have an inner product:


< , >:

VV R
(f, v) < f, v > f (v)

(63)

where,
f (v) = fi ei (v j ej ) = fi v j ei (ej ) = fi v i

(64)

and the inner product is bilinear in its arguments.


From the bilinearity of this inner product we see that we may also think
of v V as being a linear function acting on f V , giving the value
v(f ) < f, v >. Hence we see that (V ) = V (at least for finite diml vector
spaces).

47

2.9

Cotangent vectors

Tp M is a vector space. Tp M is the dual vector space - the co-tangent space


Elements of Tp M are cotangent vectors - linear functions on vectors (also
known as one-forms, as we will see later!)
Take p Tp M. Then we can think of p as the map,
p :

Tp M R
vp p (vp )

(65)

The bilinear inner product < , > is the map;


< , >:

Tp M Tp M R
(p , vp ) < p , vp > p (vp )

(66)

Take a basis {e } for Tp M. and a dual basis {e } in Tp M. Then p = e


and vp = v e , and p (vp ) = v .
Now (Tp M) = Tp M, and hence a vector acts as a linear function on the
cotangent space. So given vp Tp M we have,
vp :

Tp M R
p vp (p ) < p , vp >

48

(67)

2.10

Tensors

Tensors are the natural generalization of these maps. A (q, r) tensor is a


linear function on q cotangent vectors and r tangent vectors:
Ap :

Tp M . . . Tp M . . . R
((1) , . . . , (q) , v(1) , . . . , v(r) ) Ap ((1) , . . . , (q) , v(1) , . . . , v(r) ) (68)

Let us take a basis {e } and dual {e } at p. Then (i) = (i) e |p and

v(i) = v(i)
e . Since the tensor map is a linear map, we have,
1
r
Ap ((1) , . . . , (q) , v(1) , . . . , v(r) ) = (1)1 . . . (q)q v(1)
. . . v(r)
Ap (e1 , . . . , eq , e1 , . . . , er )
(69)
Hence the map Ap is simply defined by the collection {Ap (e1 , . . . , eq , e1 , . . . , er )}.
These are the components of the tensor. Denote them,
...

A11,...,qr Ap (e1 , . . . , eq , e1 , . . . , er )

(70)

so that for any set of covectors (i) and vectors v(i) we can write the results of
the action of A on them simply by the linear combination of the components,
...

1
r
Ap ((1) , . . . , (q) , v(1) , . . . , v(r) ) = (1)1 . . . (q)q v(1)
. . . v(r)
A11,...,qr

(71)

Hence the space of (q, r) tensors is a vector space with dimension mq+r .
q
(M). For convenience
The set of (q, r) tensors at a point p is denoted as Jr,p
we will drop the M here.

We may define a tensor product (outer product), denoted , with action


+q2
: Jrq11,p Jrq22,p Jrq11+r
in the following way;
2 ,p
+q2
Let A Jrq11,p and B Jrq22,p . Then C = A B Jrq11+r
is defined by
2 ,p

C((1) , . . . , (q1 +q2 ) , v(1) , . . . , v(r1 +r2 ) ) A((1) , . . . , (q1 ) , v(1) , . . . , v(r1 ) )
B((q1 +1) , . . . , (q1 +q2 ) , v(r1 +1) , . . . , v(r1 +r2 ) )(72)
for any (i) Tp M and v(i) Tp M. Note that this is compatible with the
linear structure of the tensor spaces.

49

+q2
1
0
Then Jrq11+r
= Jrq11,p Jrq22,p . Since J0,p
Tp M and J1,p
Tp M we may
2 ,p
write in general;
q
Jr,p
= Tp M . . . Tp M Tp M . . . Tp M

(73)

Notation varies; typically mathematicians write a tensor as T (, , . . . , ) emphasizing it is a map. Physicists would write it as
...

Tp = T11,...rq e1 . . . eq e1 . . . er

(74)

giving the tensor components explicitly.

2.11

Contraction of tensors

q+1
q
via the contraction of
Given A Jr+1,p
we may form a new tensor B Jr,p
a specified covector slot and a specified vector slot. Given a basis {ei } and
dual basis {ei } we define the contraction map by

B((1) . . . (q) , v(1) , . . . , v(r) ) A((1) . . . , e , . . . (q) , v(1) , . . . , e , . . . , v(r) )


(75)
Notationally in physics we would write,
......

B = A11,......qr e1 . . . eq e1 . . . er
In maths one would write B = A(, . . . , e , . . . , , e , . . . ).

50

(76)

2.12

Tangent vectors act on functions

Recall that a tangent vector, Vp , at the point p is the class of all curves
passing through p which are tangent to each other. For an m-dimensional
manifold M and a point p M with two curves C, C 0 passing through p, so
that,
p = C(0 ) = C 0 (1 )
(77)
for some 0 , 1 , the curves are tangent at p iff;




d 0
d

x (C())
=
x (C ())
d
d
=0
=1

(78)

This allows us to define the directional derivative of a function. Take a curve


C through p and a function f F(M). Then f C : R R.

The rate of change of f along the curve defines the directional derivative at
at p = C(0 )


d
(f C())
d
=0
Using a chart (U, ) we can break up f C = (f 1 ) ( C) where
f 1 : Rm R

C : R Rm

Then C defines a curve in Rm which we write as


x () x (C()) = C()
51

(79)

while f 1 defines a function on Rm which we write as


f (x) f ( 1 (x)) = f 1 (x)
Then we may write the rate of change at p = C(0 ) as





d
d

f C()
=
x (C())
f
(x)

d
d
x
p=C(0 )
=0
p

(80)

using the chain rule.


An important property of the directional derivative is that any other curve C 0
such that C 0 C at p gives the same rate of change. Thus the definition only
depends on the equivalence class of curves that are tangent to one another,
i.e. depends only on the tangent vector. To see this, consider a point p M
with two curves C(), C 0 () passing through p, so that,
p = C(0 ) = C 0 (1 )
for some 0 , 1 , with the curves are tangent at p, so that




d 0
d

x (C())
x (C ())
=
d
d
=0
=1
Then the directional derivative for the two curves is the same:






d
d


=
f C()
x (C())
f
(x)

d
d
p=C(0 )
=0 x
p




d 0


=
x (C ())
f (x)

d
x
p
=1

d
=
f C 0 ()
d
p=C 0 (1 )

(81)

(82)

(83)

Thus the directional derivative at p for two curves C 0 C that are tangent
at p are the same, so that the directional derivative depends only on the
equivalence class of curves at p, [C]. That is, for any tangent vector Vp Tp M
at p, we can define the directional derivative by


d
Vp [f ] =
(f C())
d
=0
for any curve C() in the class of curves corresponding to Vp .
52

2.13

Directional derivatives

For a tangent vector Xp Tp M at p, we have defined the directional derivative Xp [f ] of a function f . Then for each tangent vector Xp Tp M we have
a directional derivative which is a map from the functions to the reals,
p : F(M) R
X
f Xp [f ]

(84)
(85)

which gives the rate of change of the function at p along any representative
curve. This obeys the properties
For a constant function c, Xp [c] = 0.
Xp [f + g] = Xp [f ] + Xp [g]
Leibnitz rule: Xp [f g] = Xp [f ]g + f Xp [g]
This has the natural linear structure of Tp M, so that for a, b R and Xp , Yp
Tp M we have
(aXp + bYp )[f ] = a Xp [f ] + b Yp [f ]
(86)
for any function f F(M). Then writing Zp = aXp + bYp , we have
p + bYp
Zp = aX
p is isomorphic to the vector space of
so that the vector space of maps X
tangent vectors Xp , Tp M. Then each tangent vector Xp corresponds to a
p , and vice versa.
unique directional derivative X
The coordinate basis
A tangent vector Vp Tp M at p is an equivalence class of curves [C] through
p that defines a directional derivative at p of any function f , Vp [f ]. These
constructions are independent of any choice of coordinates. However, choosing a chart gives useful expressions. First, a tangent vector Vp Tp M defines
a vector in Rm for any curve in the corresponding class of curves [C] with
p = C(0 ) by


d
x (C())
(87)
v =
d
=0
53

As we have seen, v Rm is independent of the choice of curve in [C]. Then


we have






d


f C()
=
x (C())
f
(x)

d
d
x
p=C(0 )
=
p
0

= v
f (x)
(88)
x
p

Next, we define a set of vectors e Tp M ( = 1, . . . , m) by the corresponding maps e : f e [f ] where



1

(f

(x))
= 1, . . . , m
(89)
e [f ]

x
(p)
and we can write
e =

(90)

Then


d
= v e [f ]
f C()
d
p=C(0 )

(91)

so that
Vp [f ] = v e [f ]
The set of vectors e Tp M ( = 1, . . . , m) form a basis for Tp M and the
vector V has components v in this basis:
V p = v e
The
e =

(92)

form a basis for directional derivatives in this coordinate chart. This basis is
called the coordinate basis since it is based on a particular chart.
The basis of tangent vectors e are often written as
!

e =
(93)
x p
54

but strictly speaking this is an abuse of of notation: the tangent vectors e


are equivalence classes of curves and it is the corresponding maps e that are
derivatives.

For a chart (U, ) with p U , consider the line in Rm through (p) in the
x direction for some fixed given by
x() () = (p) + ( 0 )
so that

dx()

d
and x(0 ) = (p). We will take to be in an interval (a, b) so that (x() ())
U 0 = (U ) for all (a, b). For each , this gives a curve in U M through
p with C() () given by
C() () = 1 (x() ())
with p = C() (0 ). Then the class of curves [C() ] in M defines a tangent
vector at p in Tp M, and the directional derivative of a function f with
respect to this tangent vector is






d
d


f C() ()
x (C() ())
f
(x)
=

d
d
p=C(0 )
=0 x
p



d
x() ()
f (x)
=

d
=0 x
p

=
f (x)
x
p

=
f (x)
(94)
x
p
= e [f ]

(95)

Thus we see that the basis tangent vector e is the class of curves [C() ]
through p, and defines the directional derivative
e =

55

(96)

Changing from one coordinate chart to another


The action Vp [f ] of a tangent vector Vp on a function f is defined without
reference to any coordinate system. Choosing coordinates x gives components v of Vp in that coordinates basis. Choosing different coordinates y
defines a different coordinates basis, and the components v 0 will be different in the new coordinate system. The fact that they must define the same
Vp [f ] allows us to calculate the relation between the components v in one
coordinate system to the components v 0 in another.
If p is contained in an overlap of two charts, p Ui Uj with coordinates
{x } in Ui0 and {y } in Uj0 , then the components of a directional derivative in
the two corresponding coordinate bases transform to ensure it is invariant.
In Ui we have

(97)
e =
x
while in Uj we have
e0 =

(98)

Consider Vp acting on any function f F(M). Then we may write this


action as

f (x)
(99)
Vp [f ] = v e [f ] = v
x x(p)
or equivalently as
Vp [f ] =

v 0 e0 [f ]

=v

Since the partial derivative chain rule gives





0 f (x(y))
0 x (y)
v
=
v
y
y
y(p)


f (y)
y y(p)

y(p)


f (x)
x x(p)

(100)

(101)

we see that on the overlap we see that for the two expressions for Vp [f ] to
agree for all functions f , we must have the relation between components

0 x (y)
v =v
(102)
y
y(p)

56

While it is elegant to geometrically consider tangent vectors as equivalence


classes of curves, it turns out to be very powerful to consider them from the
isomorphic point of view, as directional derivatives. In particular the differential in differential geometry is related to the important role that tangent
vectors have acting on functions. Whilst this is obvious when taking the
view of directional derivatives, it is far from obvious from the viewpoint of
equivalence classes of curves that it would be a useful observation. This then
allows one to extend ideas of calculus to manifolds.
Hence from now on, we will take tangent vectors to be directional derivatives and we will not distinguish between tangent vectors Xp and directional
p.
derivatives X
Thus we may write
a general tangent vector in a coordinate basis as


Vp = V x p Tp (M).
Vp acts on functions as Vp [f ] = V

f (x)
x

x(p)

R for f F(M).

In an overlap of two charts, with coordinates {x } and {y }, we have,


!
!



0
Vp = V
=
V
(103)
x p
y p
with (the familar transformation of vector components from GR),

0 x (y)
(104)
V =V

y y(p)
Given a curve C() with parameter , we denote the tangent vector to
d
the curve at the point p = C() as d
. Explicitly in a chart we have,

d
dx ()
=
(105)
d
d x p=C()

57

2.14

Differentials as Cotangent Vectors

Given a function f : M R, there exists a natural cotangent vector at a


point p, i.e. a linear map
: Tp M R
which is given by its action on a vector V Tp M by
: V (V ) = V [f ]
In a coordinate basis V = V e Tp (M) where e =
V [f ] = V

(106)



x p

and


f (x)
x x(p)

for f F(M). This can be written as (V ) = V where = e is the


cotangent vector at p whose components in the coordinate basis are

f (x)
=
x x(p)
These are also the components of the differential df . Given a function f :
M R, the differential df at a point p is

f (x)
(dx )|p
(107)
df |p =
x p
so that df = dx .
We have defined a cotangent vector at p as an element of the dual of the
tangent space Tp M. Geometrically, we could instead define a cotangent
vector at p as the equivalence class of functions that have the same gradient
{ fx(x)
(This is easily shown to be coordinate independent). This
} at p.
is analogous to the definition of a tangent vector as an equivalence class of
curves. Any f in the equivalence class has the same gradient { fx(x)
} at p and
so defines the same map : V (V ) = V [f ], so that there is a one-to-one
correspondence between linear maps from the tangent space Tp M to R and
equivalence classes of functions, so that the two definitions agree.

58

The differential df is an example of a one-form at p, i.e. an expression of


the form w = w (dx |p ). There is a one-to-one correspondence between
cotangent vectors w = w e and one-forms w = w dx , so that there is an
isomorphism between the cotangent space Tp M and the space of one-forms
at p. The isomorphism between cotangent vectors and one-forms is similar
to that between tangent vectors and directional derivatives. Just as we can
regard vectors as directional derivatives, we can regard cotangent vectors as
one-forms. Then it is natural to identify the dual coordinate basis e with
the space of differentials at p
e (dx |p )

(108)

The coordinate basis { x |p } for Tp M and the dual coordinate basis {dx |p }
for Tp M are dual, i.e. they satisfy
< dx ,

>=
x

A general co-vector w Tp M can then be expressed as wp = w (dx |p ).


Then the cotangent vector discussed above corresponding to a function f
is precisely the differential = df , so that the differential df is the one-form
defining the map
df : Tp M R
which is given by its action on a vector V Tp M by
df : V df (V ) = V [f ]

(109)

d
Given a curve C() through p and a function f , then the tangent vector d
|p
df ()
d
and the covector df |p have inner product: < df |p , d |p >= d . Hence the
notation df !

To see this explicitly, use



d
dx ()
=
d
d x p=C()

(110)


f (x)
df |p =
(dx )|p

x p

(111)

and

59

together with

>=
x


dx () f (x)
df ()
d
=
< df |p , |p >=
d
d
x p
d p
< dx ,

to find

Differentials and change of coordinates 



Consider a coordinate basis, so that e = x p . Recall under a change of
coordinates y = y(x), then
!

x (y)

0
=
e
(112)
e e =
y p
y y(p)
Recall that for a dual basis to this coordinate basis we require
< e , e >=< e0 , e0 >=
From the chain rule property of partial derivatives we find the dual coordinate
basis transform as

y (x)

0
e
(113)
e e =
x x(p)
Differentials also transform this way,
dy =

y
dx
x

(114)

so that it is consistent to identify the dual coordinate basis with the space
of differentials at p
e (dx |p )
(115)
The coordinate basis { x |p } and the dual coordinate basis {dx |p } satisfy
< dx ,

>=

60

2.15

Tensors in a coordinates basis

We may now write a general (q, r) tensor at p as


Tp =

...
T11,...rq

1
r

.
.
.

dx

.
.
.

dx

x1
xq
p

(116)

Under a change of coordinates x y = y(x) one can easily see the com ...
ponents of the tensor T11,...rq in the x coordinate basis are related to the
01 ...q
in the y coordinate basis as is familar from GR;
components T1 ,...
r
0 ...

q
1
=
T1 ,...
r

y 1
y q x1
xr 1 ...q
.
.
.
.
.
.
T
x1
xq y 1
y r 1 ,...r

61

(117)

3
3.1

Induced maps, tensor fields and flows


Induced maps: push-forward

Suppose we have a map f : M N . What happens to the tangent space


at a point Tp M? A map is induced by f on Tp M at point p called the
push-forward, denoted f , such that;
f :

Tp M Tf (p) N
V f V

(118)

To define this push forward map f let us first consider a function on N ,


g F(N ). Then we can pull-back the function g onto a function on M by
g:
gf :

N R
MR

(119)
(120)

The pull-back of the function g by f is sometimes written as f (g), so


f (g) g f . Recall that a vector V Tp M can then act on the function g f to give its rate of change V [g f ]. Likewise the vector f V Tf (p) N
can act on g as f V [g] to give its rate of change.
We define the push-forward map f , by equating these rates of change
f V [g] V [g f ]

(121)

for any function g F(M). This has the important property that it is a
linear map:
f (aV + bW ) = af (V ) + bf (W )
(122)
62

for V, W Tp M and a, b R. The map f : Tp M Tf (p) N between tangent spaces can be thought of as a linear approximation to the map f .
Using a chart (U, ) on M containing the point p with coordinates x , and
a chart (V, ) on N containing the point f (p) with coordinates y we may
write V = v x |p and f V = w y |f (p) . Then
f V [g 1 (y)] V [g f 1 (x)]

(123)

y(x) = f 1 (x)

(124)

Next we define
so we can write
g f 1 = (g 1 ) ( f 1 ) = (g 1 )(y(x))
so that







1
g (y)
= v
g f (x)
(125)
w
y
x
y=f (p)
x=(p)




1
= v
(g )(y(x))
x
x=(p)






1
=
v
y (x)
g (y)
x
y
y=f (p)

using the chain rule. Hence



y
w =
v
x p

(126)

are the components of the push forward f V .


The push-forward f generalizes to tensor products, so that
f :

Tp M Tp M . . . Tf (p) N Tf (p) N . . .
V W . . . f V f W . . .

(127)

This then defines a push forward map of tensors of type (q, 0), so
f :

q
q
J0,p
(M) J0,f
(p) (N )

63

(128)

Using coordinates as above, we write


A= a

1 ,...,q

and
f A = b

1 ,...,q


. . . q
x1
x p

(129)


. . . q
y 1
y f (p)

(130)

so that the push forward is given by


b

1 ,...,q


y 1
y q 1 ,...,q
. . . q a
=
x1
x p

(131)

Given two maps f : M N and g : N O we can compose to form


g f : M O. Associated with this is the push forward (g f ) which,
for some vector V Tp M is given by
(g f ) V = g f V

64

(132)

3.2

Pull-back

We may define the pull-back f map. This pulls back the cotangent space
from N to M as
f: M N
f : Tp M Tf(p) N
f w w

(133)

The pull-back is defined as follows. Take a vector v Tp M, so that f v


Tf(p) N . Take a covector w Tf(p) N . Then for any v, w we require,
< f w, v >=< w, f v >

(134)

which defines the pull back map f .


This is a linear map:
f (aw + bv) = af (w) + bf (v)

(135)

for w, v Tf(p) M and a, b R. The map f : Tf(p) N Tp M between


tangent spaces can again be thought of as a linear approximation to the map
f.
Take coordinates x on a chart containing p in M, and y in a chart containing f (p) in N , so that y = y(x). Then we may express v = v x Tp M
and w = w dy Tf(p) N . Since

f v =

we see that for the above to hold we must have,




y

f w = w dx
x

(136)

(137)

As for the push-forward map, the pull-back map f naturally extends to ten0
sor products of covectors and hence to Jr,p
tensors.

65

3.3

Vector, covector and tensor fields

We will define a tensor field as a map that smoothly assigns a tensor to each
point in M. That is, we have a map
A:

q
M Jr,p
p Ap

(138)

We will need to define what it means for a tensor field to depend smoothly on
the point p. We will sometimes write this map as A(p) to emphasise that the
tensor field depends on the point p. With a coordinate chart, this becomes
a tensor A(x) depending on the coordinates x = (p).
We denote the space of (q, r) tensor fields as Jrq .
Example: Vector fields are in J01 , covector fields are in J10 .
Consider first vector fields. A vector field assigns to each point p M a
vector Vp Tp M. Then for any function g : M R we have the directional
derivative Vp [g] which gives a real number for each point p M. This defines
a function on M by
V [g] :

M R
p Vp [g]

(139)

Then the action of a vector on functions generalizes to vector fields on functions as


J01 F F
V , g V [g]

(140)

We can now define a vector field V as being smooth if V [g] is a smooth


function on M for all smooth functions g : M R.
For a vector field in a chart we may write V = v (x) x . Then the condition
that V be a smooth vector field is that V [g](x) = v (x) x g(x) is a smooth
function of x for all smooth functions g(x). This requires that each component v (x) is itself a smooth function.

66

A cotangent vector field assigns to each point p M a covector p Tp M.


Then for any vector field V on M, p (Vp ) gives a real number for each point
p M, so that (V ) is a real function
(V ) :

M R
p p (Vp )

(141)

Then a cotangent vector field is said to be smooth if (V ) : M R is


a smooth function for all smooth vector fields V . Introducing coordinates,
= (x)dx and V = v (x) x and (V ) = (x)v (x). Then (V ) will
be a smooth function of x for all smooth functions v (x) if and only if each
component (x) is itself a smooth function.
A (q, r) tensor at a point p was defined as a linear function on q cotangent
vectors and r tangent vectors:
Ap :

Tp M . . . Tp M . . . R
(142)
((1) , . . . , (q) , v(1) , . . . , v(r) ) Ap ((1) , . . . , (q) , v(1) , . . . , v(r) )

For a (q, r) tensor field, we require that this map from a point p to R be
smooth for all smooth covector fields (1) , . . . , (q) and all smooth vector
fields v(1) , . . . , v(r) . This requires that in some chart with coordinates {x },
each of the components of the tensor in the coordinate basis are smooth functions on M.
The smoothness requirement for a tensor field can then be expressed by
requiring that when we express the tensor at points in some chart with coordinates {x } the components of the tensor in the coordinate basis are smooth
functions on M.
The properties of a manifold then imply that if the components are smooth
in one coordinate system, then the components will be smooth in any other
overlapping coordinates.
Example: The metric tensor field which we will encounter later is g J20 ,
and we can explicitly write locally in coordinates as g = g (x)dx dx .

67

Note that tensor fields inherit the linearity of the tangent space. Given
two tensor fields A, B Jrq (M), and two functions a, b F(M), we define
addition of tensor fields and their multiplication by functions to give a tensor
field C = aA + bB Jrq (M), defined so that C(p) = a(p)A(p) + b(p)B(p) at
each point p. This will be smooth if a, b F(M) are smooth functions and
A, B are smooth tensor fields. In some chart the components of C are given
by a linear combination of the components of A, B:
...

...

...

C11...rq (x) = a(x)A11...rq (x) + b(x)B11...rq (x)

68

(143)

3.4

Tensor fields and induced maps

We have seen how a map between manifolds induces push-forwards and pullbacks of certain tensors at a point p. In this section, we discuss extending
these to tensor fields.
A map f : M N induces the push-forward map
f :

Tp M Tf (p) N
V f V

(144)

on the tangent space at a point p. For a point p the vector f Yf (p) is defined
by




f Y [g]
= Y [g f ]
(145)
f (p)

for any function g on N . We can attempt to extend this to define the pushforward of a vector field Y (p) by requiring this to hold for all p M.
There are two problems with this definition. First, it is not well-defined if f
is not one-to-one. To see this, suppose that there are two points p, p0 M
that map to the same point q N , so that f (p) = f (p0 ) = q. Then in
general f will map Y (p) and Y (p0 ) to two different vectors in Tq N , one with
f Y [g]|q defined by Y [g f ]|p and one by Y [g f ]|p0 . Thus for this to be a
good definition, the map f must be one-to-one, i.e. an embedding.
Embeddings For an embedding f , the image f (M) is a submanifold of N . A
vector field on M can be pushed forward to a vector field on f (M), but this
only defines a vector field on the submanifold f (M), not a vector field on
the whole N .
For a vector field Y , Y [g f ] is a function on M (mapping a point p M to
the directional derivative wrt Y of g f at p, which is a real number) while
f Y [g] is a function on a subspace of N . Then for a one-to-one map f , the
definition (145) of the push-forward of the vector field Y can be stated as
the requirement that for any g,
(f Y [g])(f (p)) = Y [g f ](p)
69

(146)

This can then be written as


(f Y [g]) f = Y [g f ]

(147)

where now both sides are viewed as functions on M. If Y is a smooth vector


field, then Y [g f ] is a smooth function on M for all smooth f, g, so that
(f Y [g]) f is smooth. This requires that (f Y [g]) be smooth on f (M), so
that f Y is a smooth vector field on f (M). This uses the fact that M and
f (M) are diffeomorphic, so that a smooth inverse for f is defined on f (M),
although the definition of the inverse cannot be extended to the whole of N .
The pull-back, however, can be extended to covector fields without such
problems. The pull-back of f
f: M N
f : Tp M Tf(p) N
f w w

(148)

was defined by requiring






< f w, v > =< w, f v >

(149)
f (p)

for all v Tp M and w Tf(p) N . This can be extended to define the pullback of covector fields w(q) on N , by requiring this to be true for all vector
fields v(p) on M and at all points p M. Note that this gives a well-defined
covector field on M, even if f is not one-to-one. The covector field on M is
smooth if the covector field w(q) on N is smooth.
If f is a diffeomorphism, then we may push-forward a vector field defined
over the whole of M to obtain a vector field on the whole of N , or pull-back
a covector field defined over the whole of N to obtain a covector field defined
over the whole of M. Recall that f (M) is diffeomorphic to M for embeddings, so that for embeddings one can relate vector and covector fields on M
to ones on f (M).
Consider a coordinate basis, with coordinates x for an open set U in M
containing p and coordinates y for an open set V in N containing f (p), as set
up in section 3.1. The map f : M N defines a coordinate map x y(x)
70

where y(x) is defined by equation (124). A covector w = w dy Tf(p) N


pulls back to


y

f w = w dx Tp M
(150)
x
For a covector field w = w (y)dy on N , this defines a covector field
f w = (f w) (x)dx
on M with components
(f w) (x) = w (y(x))

y
x

These components are smooth functions of x as the components w (y) are


smooth functions of y (as we are assuming w is a smooth covector field)
and y is a smooth function of x, as f is a smooth map. Then f w has components that are smooth functions of x, so it is a smooth covector field on M.
Consider now the push-forward. The push-forward of a vector V = v x |p
at p is a vector f V = w y |f (p) at f (p). with

y
w =
v
x p

(151)

We can now try to use this to construct a vector field w (y) y on N from
a vector field v (x) x on M. This requires regarding both sides of (151)
as functions of y. To regard the components v (x) as functions of y would
require inverting the y(x) to obtain a function x(y). This is not possible for
general f . If f is not 1-1, then there is no well-defined inverse, as more than
one x can give the same y. If f is 1-1, then for a point q f (M), there is
a point p in M with f (p) = q. Then with f 1-1, a point in V f (U ) with
coordinates y will be the image of a unique point in U with coordinates x, so
that there is a smooth function x(y) defined on V f (U ). Then on V f (U ),

(x) becomes a function of


v (x) defines a function v(x(y)) and similarly y
x

y, so that (151) defines w (y) as a smooth function of y in V f (U ). Of


course, this cannot be extended to points in N outside f (M).

71

3.5

Induced maps and diffeomorphisms

In general;
0
tensor), nor pull back a
We cannot push forward a covector (or Jr,p
q
vector (or J0,p
tensor). This is because the map f is not invertible in
general.

We can push/pull vectors/covectors at points. We cannot push vector


fields unless the map is 1-1. If it is 1-1, it only defines vector fields on
the image space.
However, for diffeomorphism maps we may do much more, because the inverse
map exists and is also smooth. If the map f is a diffeomorphism, we can use
(f 1 ) to pull back a vector or (f 1 ) to push forward a covector. In this
q
1
case any tensor Jr,p
can be pushed or pulled. For example A J1,p
is
pushed forward as



1
y x

dy
A = A dx (f A) = f (f ) A = A
x
x y
y
(152)
Furthermore, we can push-forward not just a tensor at a point, but an entire
tensor field - simply by pushing forward the tensor at each point p M.
Conversely we may pullback an entire tensor field. If M and N are diffeomorphic, any tensor field on M can be mapped to a tensor field on N , and
vice versa.
For an embedding, i.e. a 1-1 map f : M N , the image f (M) is diffeomorphic to M, so tensor fields on M can be mapped to ones on f (M). If a
tensor field on M is smooth, then so is the tensor field on f (M) defined by
the induced map.

72

3.6

Flows

Consider a vector field V . This generates integral curves whose tangents at


any point p are given by Vp .

Consider a curve,
C:

R M
p()

(153)

d
The tangent vector at p() is d
|p . Hence for the curve to be generated by a
d
vector field V we require Vp = d |p .

Take a chart containing p with coordinates x . Then we may write the curve
as x () and our vector field in the chart as
V = V (x)

and the tangent to the curve through p() as



d
dx ()
=
d
d x p()

(154)

(155)

Then if C is an integral curve of V we have,


dx ()
= V (x())
d

(156)

We can use this to find the integral curve explicitly. Assume that at = 0
the curve passes through the point p0 with coordinates x(0) = x (p0 ) so that
73

x (0) = x(0) . Then given we know V (x) the above equation provides a ODE
to integrate to find x () and hence C.
Recall that a smooth ODE has a unique solution for some finite interval
about = 0. If the integral curve through p extends an integral curve for
the vector field V that is defined for all R, we say it is complete. A vector
field V is complete if, at every point p, the integral curve for V through p
is complete. In fact on a compact manifold it is proven that one can extend
, so that the integral curve is complete there is a theorem saying
that every vector field on a compact manifold is complete.
Consider this solution which we denote as V (, p0 ). In coordinates we write,
x () = V (, x(0) )

(157)

so that V (0, x(0) ) = x(0) .


Then taking integral curves starting at all points in M we may think of V
for a complete vector field V as the map
V :

RM M
(, p)

p0 = V (, p)

(158)

This map gives the flow defined by the vector field V . We require the map
V : R M M to be smooth.
Remarks: From the flow equation (156) we see
For a real c R, the vector field cV gives the flow cV () = V (c)
i.e. scaling the vector field by a constant doesnt change the integral
curves, only the parameterization of those curves.
The flow obeys V ( + s, p) = V (, V (s, p)). (For a proof, see Nakahara.)

74

For a fixed , the flow V gives a diffeomorphism V () :


ducing an active coordinate transform).

M M (in-

This diffeomorphism then has the Abelian group structure of an identity


V (0) = IdM

(159)

(V ())1 = V ()

(160)

V () V (s) = V ( + s)

(161)

an inverse
and multiplication
This is the group of (active) coordinate transformations generated by V .

75

Consider the active transformation for an infinitesimal = , so that p0 =


V (, p), so that the inverse transformation gives p = V (, p0 ). In coordinates,

x0 = V (, x ) = x + V (x) + O 2
(162)
where we have Taylor expanded the solution to (156). This result is familiar
from GR.
The inverse transformation gives

x = V (, x0 ) = x0 V (x0 ) + O 2

(163)

Note that this implies for any function f ,


0

f (x ) = f (x) + O () = f (x) + V (x)

76




2
f
(x)
+
O

x

(164)

3.7

Lie Derivative of a vector field

Given a vector field V , we can differentiate a function f by taking its rate of


change at a point along an integral curve of V passing through that point.
The tangent vector to the curve is Vp at the point p, and the derivative of f
at p is given by



df
f (p0 ) f (p)
(165)
Vp [f ] =
lim
d p 0

where the integral curve through p is given by V ()p, and hence p0 = V ()p.

This derivative of a function wrt the vector field V gives a new function
which we denote V [f ] so
V [f ] :

M R
p Vp [f ]

(166)

Likewise we may differentiate a vector field Y wrt the vector field V to obtain
another vector field using the Lie derivative, denoted LV . This is similarly
defined as


V () Y |p0 Y |p
LV [Y ]|p lim
Tp M
(167)
0

where again p0 = V ()p.

77

The crucial addition to the derivative above is pulling-back the vector using
the push-forward V () . Since Y |p0 Tp0 M and Y |p Tp M live in different tangent spaces , at p0 and p, they cannot be subtracted to give a vector.
However, by pulling Yp0 back as V () Y |p0 Tp M then we can use the
linearity of the vector space Tp M to take the difference V () Y |p0 and Y |p
giving another vector in Tp M.

3.7.1

Components of the Lie Derivative of a vector field

Let us use coordinates so V = V (x) x and Y = Y (x) x and


x (p0 ) = x (p) + V (p) + O(2 )

(168)

Now we may evaluate the Lie derivative explicitly. Take the vector field Y
evaluated at p0 = V ()p, so
!


(169)
Yp0 = Y (x(p0 ))

x p0

78

and we can now push-forward by V ()1 = V () to obtain the vector we


want at p,
!

x
(p)
Tp M.
V () Yp0 = Y (x(p0 )) 0
(170)

x (p ) x p
Now using the expression for x (p0 ) above, Y (x(p0 )) can simply be expanded
in the infinitesimal  as
!

+ O(2 )
(171)
Y (x(p )) = Y (x(p)) + V (x(p))
Y (x)

x
x=x(p)
This can be thought of as using (164) for the components Y (x0 ).
From (163) we have
x (p) = x (p0 ) V (x(p0 )) + O(2 )

(172)

This can also be seen from


x (p) = x (p0 ) V (x(p)) + O(2 )

(173)

using

V (x(p)) = V

x (p ) V (x(p)) + O( ) = V (x (p0 )) + O(2 ) (174)

We can now use (172) to compute


x (p)
= 

0
x (p )

!


+ O(2 )
V (x)

x
x=x(p0 )
(175)

But using (164), we can move from x(p0 ) to x(p) up to a term of order :
!
!






V (x)
=
V (x)
+ O()
(176)

x
x
x=x(p0 )
x=x(p)
so that
x (p)
= 
x (p0 )

!


V (x)
+ O(2 )
x
x=x(p)
79

(177)

Then we find





2
=
Y
+
Y (x) Y (x) V (x) + O( )
x
x
x x=x(p)







=
Y (x) +  V (x) Y (x) Y (x) V (x) + O(2 )
(178)
.
x
x
x x=x(p)


V () Yp0

(x)

V (x)

where we see everything in the component of the vector is computed at p. We


may now subtract Yp from this, divide by , and take the vanishing epsilon
limit to give the Lie derivative in components at the point p,





LV Y |p = V (x) Y (x) Y (x) V (x)


.
(179)
x
x
x x=x(p)

80

3.8

Lie Bracket of two vector fields

The Lie Bracket [, ] is a map taking two vector fields into another vector
field,
[, ] : J01 J01 J01
X , Y [X, Y ]

(180)

defined by requiring that for all functions g F(M), the vector field [X, Y ]
satisfies,
[X, Y ][g] X[Y [g]] Y [X[g]].
(181)
In components, so X = X (x) x , Y = Y (x) x one can easily see that,



.
(182)
[X, Y ] = X (x) Y (x) Y (x) X (x)
x
x
x
Then we see that the Lie derivative is simply expressed in terms of the bracket
as
LV Y = [V, Y ].
(183)
The Lie bracket is bilinear. It is also skew-symmetric as
[X, Y ] = [Y, X]

(184)

and satisfies the Jacobi identity,


[[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0.

81

(185)

Something we will require later is how the Lie bracket transforms under a
diffeomorphism. Consider a diffeomorphism f Diff(G). Then we may
show that for two vector fields X, Y and their push forwards f X, f Y the
following holds:
f ([X, Y ]|p ) = [f X, f Y ] |f (p)
Proof: Recall, for a point p the vector f X is defined by





f X[g]
= X[g f ]
f (p)

(186)

(187)

for any function g on N . As




f X[g]

f (p)



= (f X[g] f )

(188)
p

we have that, for a map f : M N and any function g F(N ),


(f X[g]) f = X[g f ]

(189)

where both sides are viewed as functions on M. Then


X [Y [g f ]] = X [(f Y )[g] f ] = {(f X) [(f Y )[g]]} f

(190)

and hence
f ([X, Y ])[g] f = [X, Y ][g f ] = X [Y [g f ]] Y [X[g f ]]
= {(f X) [(f Y )[g]]} f {(f Y ) [(f X)[g]]} f
= {[f X, f Y ][g]} f
(191)
This then implies
f ([X, Y ]) = [f X, f Y ]

82

(192)

3.9

Commuting flows

For 2 vector fields X, Y , the Lie bracket [X, Y ] (or Lie derivative) measures
the closure of the flows generated by X and Y . If [X, Y ] 6= 0 then starting
at a point p and moving along one flow by then the other by parameter 
gives a different point to taking the other order of flows, i.e.
X () Y ()p 6= Y () X ()p

(193)

2 vector fields are said to be commuting if [X, Y ] = 0, and then the order
by which one flow doesnt matter.
Theorem: The two diffeomorphisms X () Y () and Y () X () are
equal iff [X, Y ] = 0.
See example sheet 3.
An example of commuting vector fields are those given by a coordinate basis
{ x }.

83

3.10

Lie Derivative of a tensor field

Notation; let us define the Lie derivative of a function as the usual derivative
so
LV f V [f ]
(194)
We have just defined the Lie derivative on a vector field. The Lie derivative
may be extended to allow derivation of any (q, r) tensor field A Jrq by a
vector field V as


(+) A|p0 A|p
q
(195)
Jr,p
LV [A]|p lim
0

q
Recall that since () is a diffeomorphism we may pull-back any Jr,p
0 tensor
q
to Jr,p . Now the Lie derivative wrt a vector field V defines a map,

LV :

Jrq Jrq
A LV A

(196)

Exercise: using coordinates x , show that the Lie derivative of a covector


field w = w (x)dx is given explicitly in components as




LV A = V
w + w V
dx
(197)
x
x
Show that the same result for LV A can be obtained by demanding that
LV < A, Y >=< LV A, Y > + < A, LV Y >

(198)

for any vector field Y and covector field A.


Check that the Lie derivative as defined obeys the Leibnitz rules,
LV (f A) = V [f ]A + f LV A
LV (A1 A2 ) = (LV A1 ) A2 + A1 (LV A2 )

(199)

for a function f and tensor fields A, A1 , A2 . This can be used to derive the
form of the Lie derivative of any (q, r) tensor in a coordinate basis, using the
results for vectors and covectors.

84

3.11

Example: Active coordinate transformations in


GR and symmetry

Given a vector field V , the flow V () takes a point p to a point p0 = V (, p).
If p and p0 are in the same coordinate patch Ui , then the coordinates x = (p)
will transform to the coordinates of the new point p0 , x0 = (p0 ). For infinitesimal , the diffeomorphism V () gives the infinitesimal active coordinate
transformation
x x0 = x + V + O(2 )
(200)
Such transformations play an important part in GR, particularly in perturbation theory. These transformations are diffeomorphisms, which induce
corresponding coordinate transformations.
Such active coordinate transformations are to be distinguished from passive
coordinate transformations. In an active transformation, the point p is transformed to a different point p0 , with corresponding changes of coordinates. For
a passive transformation, the point remains unchanged, but a different coordinate system is chosen. Consider a coordinate patch U , and two different
charts : U Rm and 0 : U Rm . One leads to coordinates x = (p)
and the other to coordinates x0 = 0 (p). The two coordinates are related
by the smooth function x0 (x) = 0 1 (x). Transforming from coordinates
x to coordinates x0 is a passive coordinate transformation the point p remains unchanged, but one changes from coordinates x for U to coordinates x0 .
Given a tensor field T defined on M, the Lie derivative tells us how the components of the tensor transform under a diffeomorphism. If T has components
r
...r
0
T11...
in the x coordinates and T 0 11...
...p in the x coordinates (related by
p
an active coordinate transformation as above) then the difference in these
components


...r
0 1 ...r
1 ...r
T11...
=
T

T
(201)
1 ...p
1 ...p
p
defines another tensor T , where one finds;
T =  LV T + O(2 )

(202)

Hence the Lie derivative can be used to compute the change in components
of a tensor at a point, under an infinitesimal active coordinate transform.

85

If LV T = 0 then the diffeomorphism V is said to be a symmetry transformation of the tensor T . We see from above that the active coordinate
transform generated by V leaves the components of the tensor unchanged.
For the metric tensor g, the diffeomorphisms that are symmetries of g are
isometries and the vector fields satisfying LV g = 0, so that they generate
isometries, are Killing vectors.

86

Differential forms

These objects generalize the notion of differentials found in calculus.


A form of order 1, or a one-form, is a cotangent vector. A covector at p
0
= Tp M is, in a coordinate basis, V = V dx .
V J1,p
0
A differential form of order r is a totally antisymmetric tensor in Jr,p
. Take
1
r vectors V1 , . . . , Vr J0,p
. Then a form, T , satisfies,

T (V1 , . . . , Vi+1 , Vi , . . . , Vr ) = T (V1 , . . . , Vi , Vi+1 , . . . Vr )

(203)

for any i. i.e. it is antisymmetric under any neighbouring vector pair interchange.
For a permutation of r objects, P : (1, 2, . . . , r) (P1 , P2 , . . . Pr ), then the
sign of the permutation, sign(P ), is +1 for an even permutation (i.e. requires
an even number of neighbour pair interchanges) or 1 for an odd permutation.
Following from the antisymmetry of T above, for any permutation, P , of the
vector arguments of T we see,
T (V1 , V2 , . . . Vr ) = sign(P ) T (VP1 , VP2 , . . . , VPr )

87

(204)

4.1

Cartan wedge product

Define the Cartan Wedge product ,, by


X
sign(P ) dxP1 dxP2 . . . dxPr
dx1 dx2 . . . dxr

(205)

where the sum is over all possible permutations P : 1, 2, . . . , r P1 , P2 , . . . , Pr .


For example, this gives,
dx1 dx2 = dx1 dx2 dx2 dx1

(206)

From the definition of we see this is antisymmetric under interchange of


neighbouring indices. Hence we see that,
dx1 dx2 . . . dxr = sign(P )dxP1 dxP2 . . . dxPr

(207)

and therefore this product forms a coordinate basis for the antisymmetric
tensor r-forms. Thus we may write an r-form w in a coordinate basis,
w=

1
w ... dx1 dx2 . . . dxr
r! 1 2 r

(208)

where the components of the antisymmetric tensor w are given by


w1 2 ...r = w(

, 2 , . . . , r )

1
x x
x

(209)

and are antisymmetric; w1 2 ...r = sign(P )wP1 P2 ...Pr . Note the conventional 1/r! above is designed to make the equation above for the components
simple.
Example:
Take a general (0, 2) tensor, say A = A dx dx . We may decompose this
into its symmetric and antisymmetric parts as
A = A() dx dx + A[] dx dx
where A() = 21 (A + A ) and A[] =
defined by the antisymmetric part,

1
2

(A A ). A 2-form, A0 , is

1
A0 = A[] dx dx = A[] dx dx
2
88

(210)

(211)

In general a (0, r) tensor T = T1 ,...,r dx1 . . . dxr defines an r-form T 0


by
T 0 = T[1 ,...,r ] dx1 . . . dxr =

1
T[ ,..., ] dx1 . . . dxr
r! 1 r

where our basis projects onto only the antisymmetric part of T1 ,...,r .

89

(212)

Comments
Let M be an m-dimensional manifold.
0
Denote r-forms at p as rp (M). These inherit the linearity of Jr,p
and
hence again form a vector space.

1p (M) = Tp M. Hence covectors are one-forms.


We cannot have an r-form with r > m due to their antisymmetry.
Hence rp (M) vanishes for r > m.
The dimension of the vector space T p is m. The dimension of the
vector space formed by the (0, r) tensors at p given by Tp . . . Tp is
mr i.e. a (0, r) tensor in general has mr components. The dimension
of the vector space rp (M) is given by the number of components of an
r-form,
m!
dim rp (M) =
(213)
(m r)! r!
the no. ways of choosing r from m.
So we have;
dim 1 = m, basis dx
dim 2 = 12 m(m 1), basis dx dx , components are antisymmetric matrices
..
.
dim m = 1, basis dx1 dx2 . . . dxm
We note that dim r = dim mr for r > 0.
We can include the case r = 0 above by simply defining 0p = R. This will
be convenient later.

90

4.2

Exterior product

We may use our wedge product on the basis covectors to construct the exterior product;
:

rp1 (M) rp2 (M) rp1 +r2 (M)


,

where for = r11 ! 1 ...r1 dx1 . . . dxr1 and =


we define the product as

(214)

dx1 . . . dxr2
r2 ! 1 ...r2

1
... ... dx1 . . . dxr1 dx1 . . . dxr2
r1 !r2 ! 1 r1 1 r2

(215)

Note that this is similar to the tensor product, but it automatically antisymmetrises the resulting tensor to give the r1 + r2 -form.
If r1 + r2 > dim M then the wedge product vanishes.
The wedge product is obviously bilinear. Take q , r and s .
Some other properties of the product are
Graded commutativity = (1)qr , eg. dx dy = dy dx
This implies = 0 if q is odd. e.g. dx dx = 0.
Associativity ( ) = ( ).

91

4.3

Differential Form Fields

An r-form field smoothly assigns an r-form to every point. It is a smooth


tensor field in Jr0 whose components are completely antisymmetric. Denote
the set of r-form fields as r (M).
Recall earlier we defined 0p = R. Compatible with this we may define a
0-form field to be a function, i.e. 0 (M) = F(M).

92

4.4

Exterior derivative

Perhaps the most important feature of forms is that there is a very natural
notion of differentiation of a form field - the exterior derivative denoted d:
d:

where for =

r (M) r+1 (M)


d

(216)

dx1
r! 1 ...r

. . . dxr we define,


1

d =
... dx dx1 . . . dxr
r! x 1 r

(217)

(Note that the resulting tensor is indeed antisymmetric.)


This is an interesting derivative. Previously when we differentiated a function
we required a vector i.e. a directional derivative. Here we have a derivative
operator that has no direction built in.
d enjoys a graded Liebnitz rule (which you should check!). For q and
r ,
d( ) = (d) + (1)q (d)
(218)
A very important property of the exterior derivative is that it is nilpotent,
meaning,
d2 = 0
(219)
which follows simply from symmetry,


2
1
2
... (dx dx ) dx1 . . . dxr = 0
d=
r! x x 1 r

(220)

as the partial derivative is symmetric under .


Note that since r vanishes for r > m, then for m , d = 0.
Note we have included the case r = 0; d : 0 1 . Then f F(M) = 0

gives df = fx(x)
dx .

93

4.5

Pullback of forms

Suppose f : M N . Since forms are (0, r) tensors they can be pulled back
using f in the usual way. Additionally one can show
f ( ) = (f ) (f )
d (f ) = f (d)

4.6

(221)

3d vector calculus

See the example sheet


A beautiful example of the power of forms is that they generalize 3d vector
calculus. Recall for a function f (x) we may construct a vector field using
grad, f . For a vector field v(x), we may take the curl to give another
vector field v, or take the divergence to get a function v.
Take M = R3 ;
d of a function f 0 , df gives a one-form with components that are
those of the gradient of f .
A one-form has 3 component, and hence can be associated to a 3-vector
field. d of a one-form gives a 2 form, again with 3 components, where
those components give the curl of the vector field.
A two-form is associated to a vector field. d of it gives a 3-form with
one component which is given by the divergence of the vector field.

4.7

Coordinate free definition of the exterior derivative

For r (M) we may define the exterior derivative without reference to


coordinates as
(d)(X1 , . . . , Xr+1 )

r+1
X

(1)i+1 Xi [(X1 , . . . , Xi1 , Xi+1 , . . . , Xr+1 )]

i=1

(1)i+j ([Xi , Xj ], X1 , . . . , Xi1 , Xi+1 , . . . , Xj1 , Xj+1 , . . . , Xr+1 )(222)

i<j

94

for any vector fields Xi .


For example, take 1 , and vector fields X, Y . Then
d(X, Y ) = X [(Y )] Y [(X)] ([X, Y ])

(223)

Exercise: show that the components this gives for d in a coordinate basis
(with 1 ) are the same as those given earlier.

95

4.8

Interior product

The exterior product and derivative increase the degree of a form. A vector
field X provides a natural map reducing the degree of a form. This is denoted
iX ,
iX :

r r1
iX

(224)

and is defined by
(iX ) (V1 , . . . Vr1 ) (X, V1 , . . . , Vr1 )
Using coordinates we have, =
iX =

dx1
r! 1 ,...,r

(225)

. . . dxr , and then

1
X 1 ...r1 dx1 . . . dxr1
(r 1)!

(226)

It follows from the definition that iX is nilpotent, so that i2X = 0.

4.9

Lie derivative of forms

The Lie derivative of a form can conveniently be written using the interior
and exterior products as
LX = (d iX + iX d)

(227)

Exercise: show that the components this gives for LX in a coordinate basis
(with 1 ) are the same as those given earlier.

96

4.10

Closed and Exact forms

An r-form r is closed if d = 0.
An r-form r is exact if = d where r1 is some r 1
form.
Since d2 = 0, we immediately see that any exact form is also closed.
We are aware of the importance of exact differentials in physics - for example
in thermodynamics - and given a 1-form field, = dx 1 , an important

question is whether it is exact, i.e. can it be written as = df = fx(x)


dx
for a function f ?
We now see that a necessary condition is that d = 0. However, we also see
that this is not sufficient, since might be closed, but not exact.
A natural question then arises which is the starting point for cohomology
which we will discuss later; which forms are closed but not exact?

97

4.11

Physical application: Electromagnetism

= 0. Using speRecall the Gauss law B = 0 and Faraday law E + B


cial relativity, electromagnetism is formulated in terms of an antisymmetric
field strength tensor F . Then these two laws are neatly packaged into the
Bianchi identity,
 F = 0
(228)
Since the field strength tensor F = F dx dx is antisymmetric, i.e.
F (X, Y ) = F (Y, X) for any vectors X, Y , then F is a 2-form,
1
F = F dx dx = F dx dx
2

(229)

Thinking of F as a 2-form we find that the Bianchi identity is precisely the


statement that,
dF = 0
(230)
i.e. F is a closed form!
So using differential forms we have reduced 2 of the 4 Maxwell equations to
a simple statement that the field strength is a closed 2 form.
Note that we often formulate the relativistic theory in terms of a 4-vector
potential A so that F = A A . In fact this is nothing other than
the statement F = dA where A = A dx . If this were true, F would in fact
be an exact 2-form. It is a very important point that Maxwells equations
only imply F is closed. As we will see later locally (i.e. in a contractible
coordinate patch) any closed form is exact. However this is not true globally
(i.e. over the whole manifold). Hence in a chart we may write F = dA but
not necessarily globally. We will return to this example a number of times ...

98

4.12

Orientation

Consider a chart with coordinate basis vectors { x 1 , x 2 , . . . , xm }, where we


are now interested in the order of the list of these basis vectors. Locally this
order defines an orientation (analogous to left or right handed bases in R3 ).
Taking a permutation P : {1, 2, . . . , m} {i1 , i2 , . . . , im } then the ordered
set of basis vectors, { xi1 , xi2 , . . . , xim } defines the same/opposite orientation if P is even/odd.
Consider an overlap between two charts with coordinate bases; { x 1 , . . . , xm }
and { y 1 , . . . , ym }. The Jacobian matrix induced by changing coordinates
locally from x y, is given by y (x)/x . The determinant of this matrix
tells us how the infinitesimal volume element transforms. However, the sign
of the determinant tells us how the orientations of the bases transform. A
positive sign indicates x and y are oriented similarly. A negative sign means
they are oppositely oriented.
A manifold is orientable if there exists an Atlas with charts (Ui , i ) and
coordinates x(i) (p) = i (p) such that on all overlaps Ui Uj 6= 0, then


Jij = det x(i) /x(j) > 0 p Ui Uj .

(231)

Then this Altas defines an orientation consistently over the whole manifold.
The positive definite condition on all the Jacobians over the overlaps ensures
that the orientation defined by one chart is consistent with that defined by
its neighbours. For a non-orientable manifold, such an Altas cannot be found.

99

4.13

Top forms and volume forms

Take an m-form, V on an m-dimensional manifold M. Since it has the highest rank possible it is called a top form.
A top form has only one independent component, and in some coordinates
we may write it as
1
V = v(x)dx1 dx2 . . . dxm = v(x)1 2 ...m dx1 . . . dxm (232)
r!
where the epsilon symbol 1 ...m = sign(P ) for permutation P : (1, 2, . . . , m)
(1 , 2 , . . . , m ). Consider how the component of V transforms as we change
coordinates from a chart Ui to chart Uj :
!
!
x1(i) 1
xm
(i)
m
V = v(i) (x(i) )
...
1 dx(j)
m dx(j)
x(j)
x(j)
!
x1(i)
xm
(i)
= v(i) (x(i) ) 1 ...m 1 . . . m dx1(j) . . . dxm
(j)
x(j)
x(j)



= v(i) (x(i) ) det x(i) /x(j) dx1(j) . . . dxm


(j)
= v(j) (x(j) ) dx1(j) . . . dxm
(233)
(j)


We see that the Jacobian Jij = det x(i) /x(j) of the coordinate transformation x(i) x(j) naturally emerges.
Take an orientable manifold M, with an Altas such that Jij > 0. A volume
form, V , is a top form on M, such that for any point p M, when we write
V in the coordinate basis of a chart containing p, say p Ui , then
V = v(x(i) )dx1(i) . . . dxm
(i)

with v(x(i) ) > 0

(234)

Remarks:
1. From the above discussion we note that one may only find volume forms
on manifolds that are orientable.
2. A top form is generally not a volume form.
3. On an orientable manifold (infinitely) many volume forms can be found.
4. A volume element vanishes nowhere.
100

A volume form, V , defines an orientation on M. If we are interested in the


orientation in some open set U , we simply find a chart (U, ) such that the
coordinates x = (p) are such that the component of V is positive definite
over U . This then smoothly defines an orientation in the tangent space at
each point in that chart. We may then continue to extend this to other charts.
Eventually one will build up an Altas with Jij > 0, oriented in the sense of V .
Note that if V is a volume form, then V is also a volume form but with
the opposite orientation. For an Altas with Jij > 0 oriented in the sense of
V , then one can find an Altas oriented with V 0 = V simply by taking the
same charts, but performing an odd permutation of the coordinate labels i.e.
0P1
0Pm
01
0m
{x1(i) , . . . , xm
(i) } {x(i) , . . . , x(i) } = {x(i) , . . . , x(i) } such that signP = 1.
Example: dx1 dx2 . . . dxm defines the usual orientation on Rm .

101

4.14

Integrating a top form over a chart

Consider an m-dimensional orientable manifold with a top form V m ,


and choose an oriented Atlas so that Jij = det x(i) /x(j) > 0. Then we
define the integral V over a chart Ui (with coordinates x(i) ) to be,
Z
Z
V
dm x(i) v(i) (x(i) )
(235)
Ui0

Ui

where V = v(i) (x(i) )dx1 . . . dxm . The RHS of this is simply the usual
multidimensional integration on Rm from Calculus.
We have used a particular chart to make this definition, and we should check
that the definition is independent of this choice of chart. Consider the integration over an overlap region between two charts R = Ui Uj . Then we
must show that
Z
Z
dm x(i) v(i) (x(i) )
V =
R
Zi (R)
dm x(j) v(j) (x(j) )
(236)
=
j (R)

Now as discussed above, an m-form transforms as


V

= v(i) (x(i) )dx1(i) . . . dxm


(i)


= v(i) (x(i) ) det x(i) /x(j) dx1(j) . . . dxm
(j)
= v(j) (x(j) ) dx1(j) . . . dxm
(j)

Hence we see that


Z
V

dm x(j) v(j) (x(j) )

(237)

j (R)

Z
=
j (R)



dm x(j) v(i) (x(i) ) det x(i) /x(j)

(238)

and then using the standard transformation of the integration measure in


Calculus, so that,
Z
Z




m
m

d x(j) =
d x(i) det x(j) /x(i)
(239)
j (R)

i (R)

102

and then
Z
V








dm x(i) det x(j) /x(i) v(i) (x(i) ) det x(i) /x(j)
(R)
Z
Z i
m
d x(i) (x(i) ))v(i) (x(i) ) =
dm x(i) v(i) (x(i) )
=

i (R)

x(i) /x(j)

i (R)



1



det x(j) /x(i)


and det x(j) /x(i) =

since det
=


det x(j) /x(i) since we have chosen our Altas so Jij > 0.

Let us now consider the top form V to be a volume form, rather than just
an m-form so that its component is positive definite, and hence the volume
form is nowhere vanishing.
This then defines a positive definite integration
R
measure so that Ui f V > 0 for any positive definite function f F(Ui ) so
that f (p) > 0 for all p Ui .

103

4.15

Integrating a top form over an orientable manifold

We may integrate a top form over an orientable manifold by using a Partition of unity on M. Let us take an oriented altas {(Ui , i )} on M so that
any point p is found in a finite number of the Ui s.
[This is always possible for a compact manifold, and technically the sufficient
condition is paracompactness]
Take smooth functions i (p) on M such that;
0 i (p) 1
i (p) = 0 if p
/ Ui .
P

i i (p) = 1 for all p M


For example, any smooth function f F(M) can be decomposed as
X
X
f (p) =
i (p)f (p) =
fi (p)
(240)
i

where fi (p) = i (p)f (p) are new smooth functions that are only non-zero in
the open set Ui .
Using this partition of unity we can integrate a top form V on M, by taking,
Z
XZ
i V
(241)
V
M

Ui

where we already know how to perform the integral of the top form i V over
a chart Ui .

104

4.16

Integration of r-forms over oriented r-dimensional


submanifolds

Suppose we have an m-dimensional manifold M, and a one-to-one map f


defining an embedding of the r-dimensional manifold N into M, where r
m. Thus,
f:

N M

(242)

and the image f (N ) is a submanifold of M (which is diffeomorphic to M).


Now suppose that N is orientable. Then a volume form on N may be pushed
forward to give a volume form on the image submanifold f (M). Hence, given
an orientation on N , the map f induces an orientation on the submanifold
f (M).
Given such an n-dimensional oriented submanifold of M, we may naturally
integrate an r-form r (M) over it using the fact that we may always
pull back a an r-form;
Z
Z
(f )
(243)

f (N )

Comment: Note that if the manifold N has a boundary N , then that


boundary is embedded in M as the (r 1)-dimensional submaniflold f (N ).
Then one can integrate an (r 1)-form over this boundary.

105

Stokes theorem and cohomology

5.1

Cycles and boundaries

We will refer to a closed oriented submanifold V of dimension r as an rsubmanifold.


[ Recall: a closed submanifold is one whose complement is open. ]
The boundary of V is denoted V , and provided the embedding of the submanifold is smooth, then the boundary is of dimension (r 1). Note that if
V is oriented, then V is also oriented. Hence V is an (r 1)-submanifold
or disconnected sum of these.
An r-submanifold is
an r-cycle if it is closed and has no boundary ie, V = 0
an r-boundary if it is closed and is the boundary of some other closed
(r + 1)-submanifold i.e. V = W with W a closed (r + 1)-submanifold.
A very important fact from topology is that a boundary has no boundary!
This is obviously true for elementary submanifolds with boundary, such as
an r-simplex in Rm , and more complicated submanifolds can be build out of
many of these building blocks.
Hence if V is an r-boundary then V = 0. We can think of as being an
operator that is nilpotent. This is the basis for homology theory in topology.

106

5.2

Stokes theorem

For r1 (M) and r-submanifold V , then Stokes theorem states that,


Z
Z
d =

(244)
V

where we emphasize that V is an r-submanifold (i.e. a closed oriented


submanifold).
This is quite a remarkably elegant result.
The righthand side vanishes if V = 0.
If V is a disconnected sum of (r 1)-submanifolds then the righthand side
must be summed over each component (r 1)-submanifold.
The proof is elementary but long winded - see Nakahara for details. One
essentially must prove this is true for the standard simplex and then extend
the result to general simplices, and hence general submanifolds. A 0-simplex
is a point, a 1-simplex is a line segment, a 2-simplex is a triangle with its
interior, a 3-simplex is a solid tetrahedron (a tetrahedron with its interior)
and an r-simplex is an r-dimensional generalisation of the tetrahedron. The
boundary of an r-simplex is made of r 1 simplices, each of which has a
boundary made of r 2 simplices, and so on. In simplicial homology, a toological space is decomposed into simplices glued together along their faces.
An important property of Stokes theorem is integration by parts. Let
m = dim(M). Then take r1 , mr . Then
Z
Z
Z
Z
r1
=
d ( ) =
d + (1)
d
(245)
M

107

5.3

Example. Gausss and Stokes law

Recall that grad, div and curl of 3-d vector calculus arose naturally considering forms in R3 . In 3-d vector calculus we also have the Gauss law,
Z
Z
3
vd x =
v dS
(246)
V

where S is the boundary of V and dS is the directed surface area element,


and the Stokes law,
Z
Z
( v) dS =
v dl
(247)
S

for a surface S with bounding curve C, where again dS is the directed surface
area element and dS is the directed line element.
Take M = R3 . It is a simple exercise to show that Gausss law arises when
we consider Stokes theorem for a 2-form , and a 3-simplex given by the
geometric volume V , so that V = S. Likewise Stokes law arises when we
consider Stokes theorem for a 1-form , and a 2-simplex which is the geometric surface S, and so S = C. These are the two instances of Stokes law
for R3 .

108

5.4

de Rham cohomology

Recall for r , is closed if d = 0 and exact if = da. Then we define


the space of closed r-forms forms, the vector space, Z r (M)
the space of exact r-forms forms, the vector space, B r (M)
As we saw earlier the exterior derivative is nilpotent - hence d2 = 0 for
r . Thus, B r Z r . i.e. if B r then Z r .
Two closed r-forms z1 , z2 Z r (M) are equivalent - cohomologous - if
z1 z2 B r (M)

(248)

i.e. the closed forms differ by an exact form. We write the equivalence
z1 z2 .
We denote the set of equivalent closed r-forms - equivalence classes - by [z]
where z is a representative closed r-form in the class. Hence
[z] = {z 0 Z r (M) where z 0 z}

(249)

is the equivalence class.


The linear structure of Z r , B r is simply inherited from r . However the set of
equivalence classes of closed r-forms, {[z]}, also forms a vector space H r (M)
- the cohomology vector space. It inherits its linear structure from Zr (M)
and the d operator.
Take z1,2 , w1,2 Z r (M) so that z1 w1 and z2 w2 . Hence we can write
w1,2 = z1,2 + dc1,2 with c1,2 r1 (M). Then for , R we have,
w1 + w2 = z1 + z2 + dc1 + dc2
= z1 + z2 + dc

(250)

where c = c1 + c2 r1 (M) and so we see w1 + w2 z1 + z2 .


Thus we find the linear structure,
[z1 ] + [z2 ] = [z1 + z2 ] , R
109

(251)

where we have shown above that this relation is independent of the choice of
representative. i.e. for w1,2 z1,2 then
[w1 ] + [w2 ] = [z1 ] + [z2 ] = [z1 + z2 ] = [w1 + w2 ]

(252)

While the space of closed and exact forms Z r and B r are infinite dimensional
vector spaces, the cohomology vector space may be finite dimensional. i.e.
there are only a finite number of inequivalent closed r-forms on a manifold.
We may provide a basis for the vector space H r (M) by giving a set of inequivalent closed r-forms that give representatives for classes spanning H r .
Let us call these inequivalent closed r-forms { i }, i = 1, . . . , dim(H r ). Then
the basis for H r (M) is {[ i ]}.
For an m-dimensional manifold, Z r (M), B r (M) and H r (M) are only nontrivial for 0 r m.
Remark: For any (connected) M, H 0 (M) = R.
Proof: For r = 0 we have 0 (M) = F(M). For f F(M) we have, in some

chart, df = fx(x)
and so if df = 0 then fx(x)
dx
= 0 in all charts over M and
hence f is constant.
Thus the space of closed 0-forms, Z 0 = R, i.e. the coefficient of the constant
function. Since for 0-forms B 0 (M) = 0, then H 0 = Z 0 = R, with the
representative for a basis being any constant function.

110

5.5

Example: Cohomology of R

Consider H 1 (R). Any 1-form is closed as for any 1 (R) then d = 0.


However every 1-form is also exact. Take coordinate x. Then we can write,

Z x
0
0
(253)
= f (x)dx = d
f (x )dx = dF
0

where F (x) =

Rx
0

f (x0 )dx0 0 (R).

Hence any closed 1-form is equivalent to zero, i.e. for any Z 1 then
[] = 0. Therefore H 1 (R) = 0.

5.6

Example: Cohomology of S 1

Let us now reconsider the above situation for an S 1 . We can embed S 1 in C


as S 1 = {ei [0, 2)}.
As for R, every 1-form is closed. However, now not every 1-form is exact. As
above we can use our coordinate in the chart (0, 2) to write a 1-form
1 (S 1 ) as
Z

0
0
f ( )d = dF ()
= f ()d = d
(254)
0

Now the question is whether F () defined locally over the chart (0, 2)
can be extended to a smooth function globally on the S 1 . If it can, then is
exact. Otherwise it is only closed. Clearly F () can be extended to a smooth
function only if F (0) = F (2). Note that F (0) = 0, hence our condition is
that
Z
Z 2
if
=
f ()d = 0 then B r (S 1 )
(255)
S1

R
Take two Rforms ,
0 Z 1 (S 1 ) such that 6= 0 i.e.
/ B 1 (S 1 ). Then
R
take a = ( 0 )/( ) R. Now 0 a B 1 . Hence [ 0 ] = a[] for any 0 .
Hence H 1 (S 1 ) = R and gives a representative for the basis.

111

For example = d in the chart = (0, 2), which is a volume form on S 1 ,


gives a representative. Choosing an atlas for S 1 , one can see that extends
to a well-defined 1-form on S 1 , but there is no globally defined function f on
the circle with = df .

112

5.7

Cohomology and Topology

A very important result is that the cohomology groups of a manifold depend


only on its topology. In particular, the dimension of the vector space H r (M)
is a topological invariant, known as the rth Betti number br (M).
For an m-dimensional manifold, br = 0 for r > m and the Euler number or
Euler characteristic is defined by
= b0 b1 + b2 b3 + . . . (1)m bm

(256)

de Rhams theorem: If M is compact without boundary then H r (M) is


finite dimensional. Furthermore H r (M) is dual to the rth homology group
Hr (M) .
This means that, as the name suggests, homology is dual to cohomology.
Homology is outside the scope of this course; see Nakahara for an exposition.
Roughly speaking, homology is based on the idea of considering closed cycles
modulo cycles that are boundaries.
De Rhams theorem is a profound statement as it links topology and properties of differential forms. The cohomology groups are often straightforward
to calculate and so can be a useful way of finding information about topology.

5.8

Stokes Theorem and Cohomology

Consider an r-submanifold V and an r-form r . We may integrate


over V .
Firstly suppose that V is a cycle, so V = 0, and that is exact so that
= d. Then
Z
Z
Z
=
d =
=0
(257)
V

Now suppose that B is a boundary so that B = V . We may integrate a


form r1 over B. Suppose is closed, so that d = 0. Then
Z
Z
Z
=
=
d = 0
(258)
B

Hence Stokes theorem implies 2 important things for cohomology;


113

Integrating an exact form over a cycle gives zero.


Integrating a closed form over a boundary gives zero.
In practice these are useful. For example, if we integrate a closed form over
a cycle and we get a non-zero answer, then the form is not exact.
Furthermore, due to the relation of the dimension of the cohomology to the
topological Betti numbers, dim H r = br , there is a useful consequence of de
Rhams theorem. Recall that integrating an exact form over a cycle gives
zero. Then
Theorem: RTake M compact without boundary. Consider a closed r-form
. Then if V = 0 for every possible r-cycle V in M then is exact.
Map on cohomology spaces induced by cycles:
Given an r-cycle V we may define the map
V : H r R
Z
[] V ([])

(259)

We note that while the definition depends on an explicit representative, the


map is independent of the choice of this representative.

114

5.9

Poincar
es Lemma

[We will not prove the Lemma here, although the proof is elementary. Nakahara provides a proof.]
An open set U in M is contractible if it can be smoothly contracted to a
point, p0 M. This means there is a map f
f:

RM M
, p p0 = f (, p)

(260)

such that U = f (0, U ) and p0 = f (1, U ) and f is continuous (C 0 ).


E.g. A disc is contractible but an annulus is not.
Then consider forms over the open set U , which itself is a manifold (diffeomorphic to an open subset of Rdim(M) ). Then Z r (U ),B r (U ) are the space of
closed, exact forms defined over U (not over all M).
Poincar
es Lemma: For contractible U M, any closed form over U is
also exact, i.e. Z r (U ) B r (U ).
We say that any closed form is locally exact - where locally is meant in the
sense above (i.e. in a contractible open set).
We may then think of De Rham cohomology as being the topological obstruction to closed forms being globally exact on M.

115

5.10

Example: Electromagnetism (again!)

Recall we claimed the EM field strength is a 2 form F = 2 (M). Then


locally we may write in some contractible open set F = dA for A 1 (U ).
Hence the usual statement in components F = A A .
Globally F is closed, not exact. Hence we may characterize a solution of EM
by the cohomology class of F . Given a set of closed 2-forms { i } which give
a basis of H 2 , {[ i ]} for i = 1, . . . , b2 (M) then the general field strength can
be globally written in terms of a dynamical vector potential A = A dx as
F = i i + dA,

i R,

A 1 (M)

(261)

Why were we not told of this? Simply because H 2 (Rm ) = 0, by Poincares


lemma - i.e. any open set in Rm is contractible to a point, so any closed form
is also exact.
Suppose we are not on Rm . We may take
R a 2d submanifold M with no boundary - say a 2-sphere - and integrate, M F to measure the flux through this
surface. If this flux does not vanish
R then F is not exact. Note then that
the space cannot be Rm ! The flux M F represents the magnetic charge contained within M, so this would be non-zero if there is a magnetic monopole
contained within M.

116

5.11

Poincar
e duality

Take an m-dimensional compact manifold M. Then let us define an inner


product
< , >:

H r (M) H mr (M) R
Z
[] ,

[] < [], [] >

(262)
M

This inner product is bilinear in the two arguments. It is simple to show


that this definition is independent of the representatives of the cohomology
classes
Z
Z
Z
r
d
d =< [], [] > (1)
+
< [], [ + d] > =
M

= < [], [] >

(263)

and likewise in the first argument. We have used the integration by parts
with M = 0 as M is compact.
Theorem (Poincar
e): The inner product < , > is non-degenerate.
Hence we see that H mr (M) is the dual vector space to H r (M). Note that
this implies the relation of the Betti numbers br = bmr .

117

6
6.1

Riemannian Geometry I: The metric


The metric

Take a manifold M. A Riemannian metric is a (0, 2)-tensor field g J20


that is symmetric and positive, i.e. at any point p M and for any vectors
U, V Tp M,
g(U, V ) = g(V, U )
g(U, U ) 0

(264)

with equality in the latter only if U = 0.


The manifold M together with a metric g form the pair, (M, g), a Riemannian manifold.
A pseudo-Riemannian manifold is a manifold together with a pseudo-Riemannian
metric which is a symmetric (0, 2)-tensor field g, now not positive, but instead with the property that at a point p M
if

g(U, V ) = 0 U Tp M V = 0

(265)

This pseudo-Riemannian metric divides the elements of Tp M into 3 classes:


< 0 timelike
g(U, U ) = 0 null
> 0 spacelike

118

(266)

Taking a chart with coordinates {x } we can express the metric g = g dx


dx . In the Riemannian case positivity implies that the matrix g is invertible - i.e. the metric is non-degenerate. Likewise in the pseudo-Riemannian
case the property in equation (265) ensures invertibility of g . Then Riemannian and pseudo-Riemannian manifolds both have an inverse metric, i.e.
a (2, 0)-tensor field g 1 J02 that is symmetric and whose components g
in any chart form a symmetric matrix that is the inverse of the matrix g
given by the components of the metric.
The metric has components g = g which form a real symmetric matrix, so
the eigenvalues of the matrix g are real, and are non-zero since the metric
is invertible. Then at any point we may compute the signature of the metric;
signature = (# of negative eigenvalues, # of positive eigenvalues) (267)
This is independent of the point at which we evaluate it. For a d-dimensional
M, we find for a Riemannian metric the signature (0, d) and for a pseudoRiemannian metric the signature (1, d 1).
We denote the inverse metric components (g )1 = g so that,
g g =

119

(268)

6.2

Metric inner product

The metric tensor gives the non-degenerate inner product at the point p
g:

Tp M Tp M R
U, V
g(U, V )

(269)

It thus provides the map (lowering an index)


Tp M Tp M
U g(, U )

u (g u ) dx
x

(270)

and also the map (raising an index)


Tp M Tp M
U s.t. g(, U ) = ()

dx (g )
x

(271)

This extends to general (q, r) tensors, as we are familiar with from physics.

120

6.3

Volume element

The metric g on an orientable m-diml manifold gives us a canonical volume


form, g . This is defined as
q
g | det g | dx1 . . . dxm
(272)
It is not a priori obvious that this statement is independent of the coordinates - we note the explicit dependence on the components g . However
the statement is invariant under change to other coordinates, although we
require all the coordinates to have the same orientation (hence the requirement the manifold is orientable). Let us now show this.
Recall from our previous discussion of volume elements that under a change
of coordinates x y ,
 
x
1
m
dx . . . dx = det
dy 1 . . . dy m
(273)
y

where det x /y is the Jacobian. Since our manifold is orientable we
may choose charts with coordinates so that all such Jacobians on any chart
overlaps are positive i.e. they all have the same orientation. We recall that
the metric components transform as


x x

g dy dy
(274)
g = g dx dx = g dy dy =

y y
so that,

| det g | = | det



2
x x
x
g | = det | det g |
y y
y

(275)

Hence we see that,


g

 
q
x
dy 1 . . . dy m
=
| det g | det

y

q

x
= sign det
| det g | dy 1 . . . dy m
y
q
=
| det g | dy 1 . . . dy m
121

(276)

since we have chosen coordinates so that the sign of the Jacobian is +1.
Thus we have seen that provided we have coordinates which preserve orientation the definition of the volume element above is indeed independent of
the coordinate chart.
The volume element is called a pseudo-tensor as the sign of the tensor flips
if you define it using coordinate charts covering the manifold with the opposite orientation. Once defined, it behaves as an mform tensor in all respects.
The component function is always positive by the definition, and since the
determinant of the metric never vanishes, we see g indeed defines a volume
form.

122

6.4

Epsilon symbol

Let us define the epsilon symbol,

1 2 ...m

+1
= 1
0

for (1 , 2 , . . . m ) an even permutation of (1, 2, . . . , m)


for (1 , 2 , . . . m ) an odd permutation of (1, 2, . . . , m)
otherwise
(277)

Then we may raise indices using the metric to give,


1 ...r r+1 ...m = g 1 1 . . . g r r 1 2 ...m

(278)

We note that the epsilon symbol does not transform as thepcomponents of a


tensor. As we have just seen we require the combination |g|1 ...m dx1
. . . dxm which then is a (pseudo-) tensor, where g = det g . In fact the
epsilon symbol is said to transform as a tensor density, namely as the components of a tensor up to multiplication by g.
We note the following useful identities:

a11 . . . amm 1 ...m = det a 1 ...m
1 ...m = (det g)1 1 ...m
1 ...m 1 ...m = m! det g
1 ...r r+1 ...m 1 ...r r+1 ...m = r!(m r)!(det g)1
where T[1 ...r ] =

1
r!

pPerm(1,2,...,r)

sign(p)Tp1 ...pr .

123

[r+1 r+2
1+r r+2

...

m ]
m

6.5

Hodge Star

Recall that dim r = dim mr on an m-dimensional manifold. With a


metric we have a natural map;
?:

r (M)

mr (M)

dx1 . . . dxr ? (dx1 . . . dxr )

p ...
1
|g| 1 r r+1 ...m dxr+1 . . . dxm
(m r)!

Remarks

For a general r-form =

1
dx1

r! 1 ...r

. . . dxr , we have

? =

|g|
... 1 ...rr+1 ...m dxr+1 . . . dxm
r!(m r)! 1 r

(279)

The volume element can be written as ?1 = g


The components of ? are
(? )r+1 ...m

p
|g|
=
1 ...r 1 ...rr+1 ...m
r!

(280)

For r we find
Riem
Lorentzian

(0, m) :
(1, m 1) :

Ex: show these are true.

124

? ? = +(1)r(mr)
? ? = (1)r(mr)

(281)

6.6

Inner product on r-forms

Take , r (M), so that,


=

1
... dx1 . . . dxr
r! 1 r

1
... dx1 . . . dxr
r! 1 r

(282)

and then
(?) = (?) =

1
(1 ...r 1 ...r ) g
r!

(283)

which is natural to integrate over M.


We use this to define the following inner product.
(, ) :

r (M) r (M) R
Z
,

(?)

(, )

(284)

We see that this product is symmetric: (, ) = (, )


For the case of a Riemannian metric, signature (0, m), the product is positive definite so that (, ) 0 with equality only if = 0. This is easily
seen using Riemann normal coordinates.

125

Adjoint of d: d

6.7

We define the adjoint exterior derivative as


d :

r (M) r1 (M)
d (1)m(r+1) ? d ?
d +(1)m(r+1) ? d ?

(0, m)
Riem
(1, m 1) Lorentzian

We find that d is nilpotent so (d )2 = 0. This is since d d = ?d ? ?d? =


? d2 ? = 0 since d2 = 0.
Why all the annoying signs in the definitions? All the signs have been to
ensure the following is nice.
Take a compact orientable manifold with metric (M, g). Take r (M),
r1 (M) relation;
(d, ) = (, d )

(285)

Hence d is the adjoint to d with respect to the inner product (, ).


Let us now show this for the Riemannian (0, m) case - I leave the Lorentzian
(1, m 1) case as an exercise. Using Stokes we have,
Z
Z
Z
Z
r+1
0=
(?) =
d ( (?)) =
d (?) + (1)
d?
M

and so
r

r(m
r)

Z
(? ? d ? )

(d, ) = (1) (1)

(286)

where (d ? ) is an r-form, so r = m r + 1 and we have inserted the identity


(1)r(mr) ? ? acting on this. Then we have,
Z
(mr+1)(r1)+r
(d, ) = (1)
(? ? d ? )
M
Z
mrr2 +rm1
(? ? d ? )
= (1)
M Z
= (1)mr+m (1)r(r1)
(? ? d ? )
M
Z
= (1)r(r1)
(?d ) = (, d )
M

126

in gory detail.
Exercise Show that for a 1-form v with components in a chart given by
v = v dx , d v is the 0-form given in that chart by
q

1

| det g | v
d v=p
| det g |
where v = g v .

127

6.8

3d vector calculus again.

Forms generalize 3d vector calculus. Recall for a function f (x) we may construct a vector field using grad, f . For a vector field v(x), we may take
the curl to give another vector field v, or take the divergence to get a
function v.
Take M = R3 with the usual cartesian coordinates xi and metric gij = ij .
The exterior derivative d of a function f 0 , df gives a one-form with
components that are those of the gradient of f .
df = i f dxi
A one-form v = vi dxi has 3 components vi , and there is a corresponding
3-vector field v i = ij vj . The Hodge star gives the 2-form
1
v = vi i jk dxj dxk
2
The divergence of v is given by
d v = d v = i v i = .v
The Hodge star of a 2-form
1
= ij dxj dxk
2
is the one-form

1
= (ij ij k ) dxk
2
The d of a one-form gives a 2 form, again with 3 components, and taking the
Hodge star gives a 1-form whose components are the curl of the vector field:
dv = (i vj )ij k dxk = ( v)i dxi

128

6.9

The Laplacian

We may now write down the Laplacian, a map operating on r-forms.


4:

r (M) r (M)
4 (dd + d d)

(287)

Note that 4 = dd + d d = (d + d )2 since both d2 = (d )2 = 0.


For example, for a function, a 0-form, f 0 (M), one finds using a coordinate chart that,
p

1

4f = d d f = p
|g|g f (x)
(288)
|g|
The Laplacian is one of the most important differential operators in physics
and mathematics.

129

6.10

Ex. Electromagnetism again....

Recall that the equation  F = 0 can be recast in the statement that


F defines a closed 2-form F , so
dF = 0

(289)

Recall the remaining Maxwell equations F = j for a current j = (, j).


One can check that we may write this as
d F = j

(290)

where j = j dx 1 (M).
Acting with d on this we find d j = 0. One can check that when written in
components this is nothing but current conservation.
The equations dF = 0, d F = j can be written on any 4-dimensional manifold
and give Maxwells equations on any space-time. Note that generalising to
m dimensions is also possible, but requires j to be an m 3 form.

130

6.11

Hodge theory

We now define harmonic , co-closed and co-exact forms; An r-form


r (M) is
closed
exact
co closed
co exact
harmonic

if d =

d =

=
4 =

0
d
0
d
0

r1 (M)
r+1 (M)

As an exact form is closed, a co-exact form is co-closed. Recall we denoted


the space of closed and exact r-forms as Z r (M), B r (M). Likewise we denote
the space of co-closed and co-exact r-forms as Z r (M), B r (M). The space
of harmonic r-forms we denote Harmr (M)
Consider a Riemannian manifold (M, g). Then since the inner product (, )
is positive, so that (, ) 0 with equality on if = 0, then we have,
(, 4) = (, d d) + (, dd ) = (d, d) + (d , d ) 0

(291)

Hence we say the Laplacian is a positive operator.


We see then that is harmonic, so 4 = 0, if and only if d = 0 and d = 0
i.e. a form is harmonic iff it is closed and co-closed.
Remarks
The spaces B r (M), B r (M) and Harmr (M) are othogonal wrt to (, ).
i.e. for any B r (M), B r (M) and Harmr (M) then
(, ) = (, ) = (, ) = 0.
The vector space of r-forms can be decomposed into the vector subspace of harmonic forms plus its orthogonal complement: r (M) =
Harmr (M) Harmr (M). Let us call the projector onto harmonic rforms, P ,
P :

r (M) Harmr (M)


P
131

(292)

Hodge decomposition theorem


On a compact Riemannian manifold the space of r-forms admits a unique
decomposition into exact, co-exact and harmonic forms; i.e.
r = B r (M) B r (M) Harmr (M)

(293)

This implies any r-form r can be written uniquely as


= d + d +

(294)

for some r1 , r+1 and Harmr .


Proof
We will not prove the result but recast the result in terms of another: Consider , j r on a compact Riemannian manifold, such that (, ) =
(j, ) = 0 for all Harmr i.e. both solution and source j are orthogonal to the harmonic forms. Then one may prove that we can uniquely solve
Poissons equation 4 = j to give,
= 41 j

(295)

This result requires hard analysis to prove.


Note that it is simple to make the solution orthogonal to the harmonic
forms by adding the correct complementary function i.e. + solves the
same equation for any Harmr .
A side remark is that one cannot solve the Poisson equation 4 = j if
the source j is not orthogonal to the harmonic forms. This follows from
(4, ) = 0 for any Harmr , which follows from 4 = dd + d d.
Example: The harmonic 0-forms are constant functions. Hence for functions
, one cannot solve 4 = unless (, c) = 0R for any constant function c
and hence, recalling that ?c = c g , we require M = 0.
Given the above result, we may proceed to prove the Hodge decomposition
theorem. Now given a form r we may use the projector P above to
132

compute the harmonic component. Subtracting this harmonic component as


0 = (P ) gives an r-form in the complement to the harmonic forms.
Then
( 0 , ) = 0 Harmr (M)

(296)

From the above result about Poissons equation we see that we may then
write 0 = 4 for a unique . Hence,
= 4 + P = d(d ) + d (d) + P
= d + d +

(297)

for = d , = d and = P Harmr . Hence we have explicitly


constructed the decomposition claimed above.

133

6.12

Harmonic representatives for cohomology

Consider a closed r-form Z r on a compact Riemannian manifold. Hodges


theorem tells us we may write any r-form as = d+d + for harmonic .
However we wish this to be closed. Then d = d2 + dd + d = dd since
being harmonic implies it is closed and co-closed, so d = 0. Consider
(d, ) = (dd , ) = (d , d )

(298)

Since we want d = 0 this must vanish. However the RHS can only vanish if
d = 0. Hence we see Z r = B r Harmr , so we may write a closed form as
= d +

(299)

for Harmr .
Therefore an important consequence of Hodges theorem is that every cohomology class [] H r (M) has a unique harmonic representative on a
compact Riemannian manifold. Take a given representative, , which then
has a unique decomposition, = d + for Harmr . Then is the
harmonic representative, so [] = [].
Note that any other representative of the same class, say 0 , so [] = [ 0 ] will
be written 0 = d0 + for the same , since all representatives differ by an
exact form.
We then see that Harmr ' H r , and the space of harmonic r-forms is a
finite dimensional vector space whose dimension is given by the topological
invariant br . This gives the remarkable fact that the number of linearly
independent solutions to the equation = 0 for r-forms depends only
on the topology of the manifold. There are a number of similar results (e.g.
in index theorems) where the number of solutions to a differential equation
on a manifold is a topological invariant, showing a profound link between
apparently different areas of mathematics.

134

6.13

Maxwells Equations on a Compact Riemannian


Manifold

Consider Maxwells equations


d F = j

dF = 0

on a compact Riemannian Manifold, so that Hodges decomposition theorem


applies. As we saw before we may write
F = + dA

(300)

where A 1 (M) and is fixed and gives the cohomology class of the
field strength; [F ] = []. The gauge transformation is then written as
A0 = A + d, for a function 0 (M). We see that F remains invariant
when A transforms, since d2 = 0.
Recall the equation d F = j. Earlier we had commented that this implies local current conservation d j = 0. However, we see now a stronger statement
is implied, namely that in fact j is globally a co-exact form.
We now use our discussion of Hodge theory to choose a canonical gauge and
a canonical representative .
Lorentz gauge
Let us consider the gauge d A = 0 - the Lorentz gauge. Under a gauge
transform we see,
d A0 = d A + d d = d A + (d d + dd ) = d A + 4

(301)

where we have used the fact that d = 0 since is a zero form. Hence we
see that by choosing = 41 (d A), then d A0 = 0.
Note that (on a compact Riemannian manifold) we may always invert the
Laplacian, since the source, d A being co-exact is always orthogonal to the
harmonic forms.
Harmonic representative

135

Recall d F = j. Hence if we choose Harm2 (M) as the representative


of the cohomology class of F , so [F ] = [], then since a harmonic form is
co-closed we simply have,
d dA = j

(302)

Furthermore using Lorentz gauge, d A = 0, we can write this as


4A = j

(303)

and hence the gauge potential can be solved as A = 41 j. Recall (on a


compact Riemannian manifold) we require the source j to be orthogonal to
the harmonic forms for this to be well defined. However, as we said above, j
must be co-exact, and therefore satisfies this restriction.
Much of physics is governed by the Laplacian.

136

Riemannian Geometry II: Geometry

A metric allows us to give a geometry to a manifold. In GR we often see the


formula
ds2 = g dx dx

(304)

What does this mean!? Where is the ?! Where is the geometry?!


At a point p the metric is an inner product on the vector space Tp M = Rm ,
and (in Riemannian signature) it defines the dot product of tangent vectors,
and hence determines their lengths, and the angle between them, just as
usual for vectors in Rm .
However, the metric tensor field allows us a notion of geometry in M itself,
not just in its tangent spaces. The key is through the theory of integration
that we have developed, which can be applied to give the length of a curve.

137

7.1

Induced metric and volume for a submanifold

Given an m-dimensional manifold M and metric g, let us consider an embedded submanifold, diffeomorphic to an n-dimensional manifold N (n m),
defined by a smooth (one-to-one) map,
f:

N M

(305)

The we may induce a metric gN on N by the pull-back gN = f g. This is


called the induced metric.
Take coordinates {x } on M and {y } on N . Then we can write g =
g dx dx , and f is defined by x = x(y). Then the induced metric is
explicitly given as f g = (f g) dy dy , with
(f g) =

x x
g
y y

(306)

We may think of classical geometry as the study of induced metrics on submanifolds of M = Rm equipped with the Euclidean metric g = ij dxi dxj .
All notions of distance and curvature we have and will discuss give the correct classical result in this case.
The induced metric can be used to define a volume form on N
p
N = | det gN | dy 1 dy 2 . . . dy n

(307)

which can be integrated to give the volume of N with respect to the metric
gN = f g. However, N is diffeomorphic to N 0 = f (N ) and the metric gN
on N can be pushed forward to a metric gN 0 = f gN on f (N ) which is the
metric on the embedded submanifold inherited from the metric g on M. The
corresponding volume element N 0 on N 0 = f (N ) is the push-forward of N ,
N 0 = f (N ). As N and N 0 are diffeomorphic,
Z
Z
N 0 =
N
(308)
N0

and defines the volume of the submanifold N 0 = f (N ), and this can be


calculated as an integral over N .

138

7.2

Length of a Curve and the Line Element

Consider a coordinate chart and a curve C given explicitly by x (). This


an example of the set-up of the last subsection, with C = N 0 an embedded
1-dimensional submanifold, with n = 1 and = y 1 the coordinate on (an
open subset of) the parameter space N = R. Then gN is the 1 1 matrix
gN = g

dx dx
d d

(309)

and
p

|gN | d

(310)


Z s
dx dx


|gN | d =
g d d d

(311)

N =
Then the length of the curve is
s=

Z p

Then at any point p C, we have


 2
ds
dx dx
= g
d
d d

(312)

The metric allows us to associate


a Rgeometric distance to a portion of the
R
ds
d. For M = Rm with a Euclidean
curve C 0 C by the integral C 0 ds = C 0 d
metric this agrees with the usual notion of distance, and for a general manifold and metric provides the generalization of it.
For an infinitesimal change in parameter along the curve, , we have
 2
s
x x
= g
(313)


We call s the line element, and identify it with a physical distance measure
along the curve. The GR notation
ds2 = g dx dx

(314)

is an abuse of notation for the line element where we have cancelled the 2 s !

139

7.3

Integration of n-forms over oriented n-dimensional


submanifolds

We have an m-dimensional manifold M, and a one-to-one map f defining an


embedding of an n-dimensional orientable manifold N into M, where n m.
Thus,
f:

N M

(315)

and the image N 0 = f (N ) is a submanifold of M (which is diffeomorphic to


N ).
Given such an n-dimensional oriented submanifold of M, we defined the
integral of an n-form n (M) over it using the pull back
Z
Z
(f )
(316)

f (N )

We now give the coordinate expression for this. Consider a chart (U, ) for
(M with coordinates x (p) = (p) ( = 1, . . . , m) and a chart (V, ) for N
such that f (V ) U with coordinates y (q) = (q) ( = 1, . . . , n). Then in
U
1
= 1 ...n dx1 . . . dxn
n!
and in V
1
x1
xn 1
1 ...n
(f ) =
. . . n dy . . . dy n

1
n!
y
y
xn 1
x1
...
dy . . . dy n
= 1 ...n
y 1
y n
!
1
x1
xn
p
=
1 ...n
...
N
y 1
y n
| det gN |

where the induced volume form is


p
N = | det gN | dy 1 dy 2 . . . dy n

140

(317)

(318)

A point q in N with coordinates y is mapped to a point f (q) in N 0 = f (N )


with coordinates x (y). Consider the n vector fields t , ( = 1, . . . n)



x
t f
=
y
y x
For example, t1 is the tangent to the curve C1 , which is the curve x (y 1 )
given by the functions x (y ) with y 1 varying and y 2 , y 3 , . . . y n fixed, so that
t1 =

d
dy 1

Similarly, t0 for some fixed 0 is the tangent to the curve C0 given by x (y)
and varying the coordinate y 0 while keeping the other n 1 coordinates y
with 6= 0 constant. The tangent vector to this curve at f (q) is
t0

x
d
=
dy 0
y 0 x

The n vectors t defined in this way are tangent vectors to the submanifold
N 0 = f (N ).

141

7.4

Hypersurfaces

A hypersurface in an m-dimensional manifold is an (m 1)-dimensional


submanifold. Consider then the case of the previous subsections with n =
m 1, so that N 0 = f (N ) is an m 1 dimensional submanifold of M (which
is diffeomorphic to N ). The induced volume element N on N defines an
m 1 form N 0 defined for all p N 0 given by the push-forward of N ,
N 0 = f (N ). The Hodge dual of this defines a 1-form n(p) 1 (M)
defined for all p N 0
n = (1)m1 ? N 0
where = 1 for Riemannian signature, and = 1 for Lorentzian signature.
This then implies
N 0 = n
The 1-form n Tf(p) M annihilates all vectors tangent to N 0 , i.e. to all
vectors in the m 1 dimensional subspace f (Tp N ) of Tf (p) M,
hn, f V i = 0
for all V Tp N . and is a smooth nowhere vanishing 1-form field on f (N ).
The pull-back of n is zero,
f (n) = 0

Using the metric to rise the index on n gives a vector N Tf (p) M with
components N = g n . For a pseudo-Riemannian manifold, the vector
N is timelike, spacelike or null, and the corresponding hypersurface is then
said to be a timelike, spacelike or null hypersurface, respectively. This is
orthogonal to all vectors tangent to the hypersurface at f (p),
g(N, f V ) = 0
for all V Tp N and is called a normal vector. The normal vector N (p)
then give a vector field over the hypersurface. For a null hypersurface, the
induced metric gN has vanishing determinant and as a result the volume
form N = 0. Then the volume of a null hypersurface is zero.

142

The normal vector is normalised so that g(N, N ) is 1 for a spacelike hypersurface, 0 for a null one and 1 for a timelike one. (For the Riemannian
case, it is 1.) We now show this. Using (283)
n (?n) = n2 g ,

n2 = g n n

(319)

and
N 0 (?N 0 ) = 2 g

(320)

where 1 ...m1 are the components of N 0 , i.e.


N 0 =

1
1 ...m1 dx1 . . . dxm1
(m 1)!

and
2 =

1
(m 1)!

1 ...m1 1 ...m1

From N 0 = n we obtain
2 = n2
using n = (1)m1 n. For a null hyper surface, n2 = 0. If n2 6= 0, then
as N 0 is the push-forward of
p
(321)
N = | det gN | dy 1 dy 2 . . . dy n
with components
1 ...m1 =

p
| det gN | 1 ...m1

satisfying
1
1 ...m1 1 ...m1 = | det gN |(det gN )1 = 1
(m 1)!
with a plus sign for a spacelike hypersurface and a minus sign for a timelike
one. Then the pushforward gives n2 = 2 = 1.
We will now assume that the hypersurface is not null. Then
g = n N 0
The normal vector is normalised so that g(N, N ) = 1 if it is spacelike or
g(N, N ) = 1 if it is timelike. If N (f (p)) is not null, then it does not lie in
143

the tangent space f (Tp N ) to the hypersurface.


For a 1-form field w = w dx on M with corresponding vector field W =
w /x , there is an (m 1)-form
iW V = ?w
which can be written as
iW (n N ) = n(W )N 0 n iW N 0

(322)

The pull-back of this is


f (?w) = n(W )N

(323)

using f (N 0 ) = N and f (n) = 0. Here n(W ) = n W (x(y) is a function


on N . Then the integral of ?w over the embedded hypersurface is
Z
Z
w =
n(W ) N
(324)
N

f (N )

Physics notation
In section 4, the integral of a top form over a chart was defined. For a
function on a manifold M, g is a top form and for a chart (U, ) with
coordinates x in U 0 = (U ) Rm the integral of the top form over U is
given by an ordinary integral over U 0
Z
Z
p
(325)
dm x | det g|
g
U

U0

The integral over U is sometimes written as


Z
Z
p
dV =
dm x | det g|
U0

(326)

U0

where here dV is not a form but represents an infinitesimal volume element.


R
R
For a hyper surface, the integral f (N ) w gives N w(N ) N . For a chart
(V, ) for N with coordinates y in V 0 = (V ) Rm1 ,
Z
Z
p
(327)
w(N ) N =
dm1 y | det gN | n W
V

V0

and the right hand side is sometimes written as


Z
Z
Z
p
m1

d
y | det gN | n W =
W n dS =
V

144

W dS
0

(328)

7.5

Electric and Magnetic Flux

The Maxwell 2-form is 21 F dx dx . At fixed time x0 =constant, we have


the R3 with coordinates xi , and
Fij = ijk B k ,

(?F )ij = ijk E k

(329)

where E i , B i are the electric and magnetic fields. Then


1
(?F )ij dxi dxj = ?E,
2

1
Fij dxi dxj = ?B,
2

(330)

where
B = Bi dxi ,
Then for a 2-surface N 0 R3
Z
Z
F =
N0

E = Ei dxi
Z

?B =

n(B) N

N0

(331)

In a coordinate chart, the right hand side can be written in the conventional
physics form for the magnetic flux,
Z
Bi dS i
Similarly,
Z

Z
F =

N0

Z
?E =

N0

n(E) N

(332)

gives the conventional form for electric flux


Z
Ei dS i

7.6

Stokess Theorem in 3D

We can now find the original Stokess theorem in 3D as a special case of the
general theorem referred to as Stokess theorem. Consider a 2-dimensional
surface S bounded by a curve C = S which is embedded in flat 3-dimensional
space, R3 . For a 1-form v = vi dxi , the 2-form dv can be integrated over S
and Stokess theorem used to find
Z
Z
dv =
v
S

145

The left-hand-side can be rewritten as


Z
Z
Z
i
dv = (dv)i dS = ( v)i dS i
S

and for the right hand side the notation


Z
vi dli
C

is sometimes used, where dli represents an infinitesimal element of length


tangent to the curve. Then we have the familiar formula
Z
Z
( v).dS =
v.dl
S

7.7

Gausss Theorem

We can now combine the above with Stokess theorem to obtain the generalisation of Gausss Theorem to any number of dimensions. For a 1-form field
w = w dx on M, d w is a top form. Suppose the hypersurface f (N ) is the
boundary of some m-dimensional submanifold P M, so that f (N ) = P.
Then by Stokess theorem
Z
Z
Z
n(W ) N
w =
dw =
P

Using d w = d w this can be rewritten as


Z
Z

( d w) =
n(W ) N
P

The notation
w d w
then allows us to rewrite Gausss theorem in the more familiar form
Z
Z
(.w)dV =
w dS
P

146

7.8

Connections

Recall that given a vector at a point p, say V Tp M, we can take the directional derivative of a function f F(M) at that point as V [f ].
However we currently have no way to take a directional derivative of a vector
field (or indeed any tensor field). (Recall using the Lie derivative we may
differentiate a tensor field using a vector field, V J01 , but at a point, p,
this derivative doesnt just depend on the direction vector V |p . Hence it is
not a directional derivative.)
In order to take the directional derivative of a tensor we first introduce an
affine connection, . This is a map
:

J01 (M) J01 (M) J01 (M)


X, Y
X Y

(333)

with the following properties: for any X, Y, Z J01 (M) and any f F(M)
the map is bilinear so that,
X (Y + Z) = X Y + X Z
(X+Y ) Z = X Z + Y Z
and also
f X Y = f X Y
X (f Y ) = X[f ] Y + f X Y
Of these latter two properties, the first shows the derivative is a directional
one - i.e. at a point p, (f X Y )|p only depends on X|p , and not the value of
X at any other point. The second shows the connection obeys a Liebnitz rule.
Since it is linear, we may specify this map by its action on basis vector fields.
Taking a coordinate basis {e } = { x } in a chart (U, ), we can specify
by its components
e e

(334)

where { } are the connection components, a set of smooth functions defined over the open set U . To specify the connection we must supply these
147

connection components in all the charts in our atlas.


Using these components together with the properties of a connection above,
we may then calculate X Y in a coordinate chart (U, ) as
X Y

= X (Y e )
= X (e [Y ]e + Y e )



e
Y + Y
= X
x

(using directional deriv property)


(using Leibnitz property)
(using connection components)

which is the familiar expression we are used to.


Consider an overlapping chart with coordinates {y } and corresponding coordinate basis {
e } = { y }. Then we can write the connection components
}, as
in the new coordinates, {
e
e e =

Using the relation betweens these bases, e =

(335)
x
e ,
y

we can compute,

 
 
x
x
x

e = x e
e = e
e

y
y
y
y

  
x x
x
=
e e + e
e

y
y
y
 
x
x x

e
+
e

e
=

y y
y

  

x
y
x x

+
e
=

y y
y y
x

(336)

and so see that the two sets of connection components are related as
2

= x x y + x y

y y x
y y x

(337)

We see that the connection components almost transform under coordinates


changes as the components of a (1, 2) tensor, but not quite due to the last
term above.

148

As a comment, suppose we have two connections , 0 which in some co0


ordinate chart have connection components { } and { } respectively.
0
Then consider the difference of their components { = }.
From above we see these do transform as components of a (1, 2) tensor,
=

x x y

y y x

(338)

since the last term in equation (337) cancels. Hence the difference between
two connections does define a (1, 2) tensor.
Alternatively, given a connection and a tensor field H J21 (M) we may
define a new connection 0 . In a chart the connection has components { }
and the tensor has components H , and we define a new connection 0 to
act on the coordinate basis as

0 e = + H e
(339)

149

7.9

Torsion

We decompose the connection into symmetric and antisymmetric parts


1
= S + T
2


1
+
2

S
T

(340)

From the transformation of the connection components (337), we see that


the symmetric part S transforms like the components of a connection
2 x y
x x y
S = S +
y y x
y y x

(341)

while the antisymmetric part T transforms like the components of a tensor


x x y
T = T
y y x
Then the antisymmetric components T
T = T

(342)

define a (1, 2)-tensor,

dx dx
x

(343)

called the Torsion tensor. The symmetric components {S } are given by


S = 21 T where is a connection and 12 T is a tensor, and so
S transform as the components of a connection, S .
Hence we may think of a general connection as being defined from a symmetric connection, S , together with a torsion tensor, T .
There is an elegant basis invariant definition given by the following map,
T :

J01 J01 J01


X, Y
T (X, Y ) X Y Y X [X, Y ]

(344)

In addition to being bilinear, the torsion map has the following important
property (Exercise to check);
T (a X, b Y ) = a b T (X, Y )

150

a, b F(M)

(345)

which implies that it defines a (1, 2)-tensor with components in a basis {e },


T

= < e , T (e , e ) >

(346)

This map gives the torsion tensor since T (e , e ) = e e = 2[] e .


We note the torsion map is antisymmetric,
T (X, Y ) = T (Y, X)

151

(347)

7.10

Parallel transport and Geodesics

Consider a vector field X J01 (M) and a curve C M given by p(). We


may ask how the vector field at a point on the curve varies, i.e. how X|p()
varies. We say the vector X|p() doesnt change along the curve, or more
formally is parallelly transported along the curve if


=0
(348)
d X
d

p()

Explicitly in coordinates {x } so that X = X (x)e and the curve C is given

d
by x() with tangent d
= ( dxd() )e then


d
dx

d X = d (X e ) =
[X ] +
X e
(349)
d
d
d
d
An important point is that whilst we have defined X as a vector field over the
manifold, we see that actually since is a directional derivative, we actually
only require the vector field to be defined over the curve C - since we only
d
[X (x())].
require d
Note that solving the above ODEs set to zero explicitly allows us to parallel
transport a vector along a curve, by specifying the components X at one
end and integrating.

152

Geodesics
A curve C is said to be geodesic if the tangent vector at p, V (p) = d/d|p , is
parallelly transported along the curve
V V |p = 0 p C

(350)

Intuitively we may think of the curve as being straight with respect to the
connection, since its tangent is parallelly transported along the curve.
In coordinates we have,


d2 x dx dx
+

d2
d d


=0

(351)

A geodesic is a curve where x () satisfies this second order differential equation. Note that in general the connection components (x) will be functions of the coordinate x, so that the differential equation will be non-linear
in general. Conversely, given a vector V Tp M at a point p with coordinates
x0 , we can solve the above second order differential equation to generate a
curve x() with


dx ()
d

= e
=V
x (0) = x0 ;
d =0
d =0
to find a geodesic through p with tangent vector V at p. If the connection components vanish, = 0, then the differential equation becomes
d2 x
= 0 with solution given by the straight line x = x0 + V through p
d2
in the direction V .

7.11

Interpretation of Torsion

A remark is that the geodesic curves of a connection only depend on the


symmetric part of . Hence two connections that differ by a torsion will
have the same geodesics. Knowing all the geodesics of a connection allows
us to precisely reconstruct the symmetric components S .

153

Given a geodesic, we may think of torsion as effecting how a vector othogonal


to the tangent parallel transports.
Start at p with two vectors X, Y Tp . Flow a small parameter along
the geodesic of X to the point px . Parallel transport Y along this curve to
obtain Y 0 Tpx . Now flow along the geodesic generated by Y 0 starting at
px to reach the point pyx . Repeat the same operations but interchanging X
and Y to end at the point pxy .
Torsion measures how close the points pxy and pyx are to each other - i.e.
how near the parallelogram is to closing. In coordinates one finds to linear
order in ,

(352)
x (pyx ) x (pyx ) = T Y X p
We remark that in classical geometry torsion vanishes.

154

7.12

Covariant derivative

We now define a directional derivative which acts on a general tensor field,


called the covariant derivative. Using a vector field X J01 we denote the
derivative X . It is the bilinear map
:

J01 Jrq Jrq


X, T X T

(353)

and we encapsulate it being a directional derivative by requiring, as we did


for the connection, that
f X T = f X T
We require our covariant derivative to obey a Leibnitz rule
X (T1 T2 ) = (X T1 ) T2 + T1 (X T2 )

(354)

where T1 Jrq11 and T2 Jrq22 are two tensor fields. Furthermore we require
that under contracting indices we have
(X T ) ( . . . , e , . . . , e , . . . ) = X (T ( . . . , e , . . . , e , . . . )) (355)

Thinking of functions F(M) as (0, 0)-tensor fields, J00 (M), then we may
define the covariant derivative on a (0, 0)-tensor field simply as
X f X[f ]

(356)

X Y f Y X f = X[Y [f ]] Y [X[f ]] = ([X, Y ])[f ]

(357)

This implies that

where [X, Y ] is the Lie bracket.


The covariant derivative acting on a (1, 0)-tensor field we take to be the
connection. Then recall we required our connection to have the property:
X (f Y ) = X[f ] Y + f X Y for f J00 and X, Y J01 . We can now think
of this as a giving the Leibnitz rule required by the covariant derivative,
X (f Y ) = (X f ) Y + f (X Y )
155

(358)

where f Y = f Y as f J00 . Using these two definitions and the Liebnitz


property, we can induce the action of the covariant derivative on any tensor
field.
Consider a 1-form and a vector field V , so that
X ( V ) = (X ) V + (X V )

(359)

This has components V . Contracting the indices gives V , the components of the function < , V >= (V ). The condition that the covariant
derivative of a contraction is the contraction of a covariant derivative then
implies that
X h, V i = h(X ), V i + h, (X V )i

(360)

Then for all , V, X we have


h(X ), V i = X [h, V i] h, (X V )i

(361)

and this defines X .


Consider the above with
X = e ,

V = e ,

= e

with the usual coordinate basis vectors


e =
Then we use the fact that <

,
x

, dx
x

h( e ), e i =
=
=
=

e = dx
>= in a coordinate chart to find

e [he , e i] he , ( e )i


[he , e i] e , e
he , e i

(362)

Now e is a one-form and so is a linear combination of basis covectors e


e = H e
156

(363)

for some components H . We then find


H =

(364)

Thus we have


=
+

x
x

dx = dx

(365)

and using these and Liebnitz we can compute the covariant derivative of a
...
general tensor T = T11,...rq x1 . . . xq dx1 . . . dxr where X
...
acts on the component T11,...rq function and each vector and co-vector in the
basis of this (q, r)-form field.
For example
= ( e )
= ( )e + e
= ( )e e

= e

157

7.13

Metric connection

For a given metric g, a connection is a metric connection if at each point


p M it obeys
X g = 0 X Tp (M)

(366)

Another way to say this is that the metric is parallel transported along any
curve.
Given a coordinate basis we may compute

g = g g g e e

(367)

and hence the metric connection must obey


g g g = 0 , ,

(368)

Cyclic permutations of this give,


g g g = 0
g g g = 0

(369)

Adding both equations in (369) and subtracting the one in (368) we obtain
() g [] g [] g = C g

(370)

where C is called the Christoffel connection, defined as


1
C g ( g + g g )
2

(371)

Hence we see that the symmetric part of the connection can simply be written
in terms of the Christoffel connection and the antisymmetric part, given by
the torsion tensor,
S

= C + T ()

(372)

where we note that T = g g T .


In summary the condition that is a metric connection completely determines the symmetric part of the connection in terms of the antisymmetric
158

part, which is unconstrained by the condition. Thus a metric connection is


determined entirely by the metric g (which determines the Christoffel connection) and torsion tensor T .
Levi-Civita connection
In the case of vanishing torsion, T = 0, we see there is a unique metric connection = C , which is symmetric = . This is called the
Levi-Civita connection.
Hence for vanishing torsion the metric g completely determines the metric
connection. In the absence of physics to dynamically determine the torsion,
this is the natural connection arising in GR. In particular this arises from
variation of the Einstein-Hilbert action.
We note that for the Levi-Civita connection one finds that geodesics have
extremal length. One can compute a geodesic curve C by functional variation
of,
Z
Z
q
(373)
I[C] = ds = d x x g [x()]
C

with respect to x () where x =

C
dx ()
.
d

159

One obtains: x + C x x = 0.

7.14

Curvature

We define the curvature map as


R:

J01 J01 J01 J01


X, Y, Z R(X, Y, Z) X Y Z Y X Z [X,Y ] Z

Firstly this map is antisymmetric R(X, Y, Z) = R(Y, X, Z).


Secondly we note it is trilinear and has the following important property
(Exercise to check),
R(a X, b Y, c Z) = a b c R(X, Y, Z)

a, b, c F(M)

(374)

which implies the map defines a (1, 3)-tensor R e e e e , by


R = < e , R(e , e , e ) >

(375)

This (1, 3) tensor is called the Riemann curvature tensor.


Taking the basis {e } to be a coordinate basis, so that [e , e ] = 0, then we
can compute,
R = < e , e e ) >


= < e , e e ) >
< e , e [ ]e + e e [ ]e e ) >

+
=

x
x
=

and from R(X, Y, Z) = R(Y, X, Z) we have the antisymmetry, R =


R .
Geometric interpretation
Consider a point p and coordinates {x } and the two curves C1 , C2 defined by
xp +  , xp + where [0, 1] and we take  , << 1. We close these
curves into a small parallelogram using two more curves, C10 , C20 , defined by
xp + + and xp + + . Call point p0 that with coordinates xp + + .

160

Then take a vector V at x . Parallel transport along C1 then C10 to obtain


a vector V1 Tp0 . Parallel transport along C2 then C20 to obtain a vector
V2 Tp0 . Then the difference in these is given by the curvature,
V1 V2 =  R V

(376)

See the book by Nakahara for a derivation of this result.

7.15

Holonomy

A loop is a closed curve


C : S1 M
[0, 1] p = C();

p(1) = p(0)

(377)

Consider the loops passing through a point p. Given a connection, any vector
X Tp M can be parallel transported round a curve C through p to give a
new vector XC Tp M. This defines a linear transformation
MC : Tp M Tp M
MC : X X C

(378)

This is the holonomy at p of the connection for the loop C. To see this, choose
a contractible patch U and note that each basis vector e will be transported
to a new vector (e )C which can be written as a linear combination of basis
vectors:
(e )C = (MC ) e
(379)
so that X = X e transforms to
XC = X (MC ) e

(380)

(as parallely transporting a function X round a loop leaves it unchanged).


Going back round the loop in the opposite direction should give the inverse
transformation, so the matrix MC is invertible, MC GL(m, R).
Two curves C1 , C2 through p can be combined to give a curve C3 through p,
where you go round C1 first, then round C2 . The matrices clearly combine as
MC 3 = MC 2 MC 1
161

(381)

The set of all linear transformations MC for all curves C through p then form
a group Hp , the holonomy group at p. Clearly this is a subgroup of GL(m, R).
If two points p, q are connected by a curve D, then parallel transporting
any vector X Tp M along D to a vector XD Tq M defines a linear
transformation
L D : Tp M Tq M
(382)
Given a curve Cq through q, there is then a curve Cp through p given by
going from p to q along D, then round Cq , then back along D to p. This can
be written as Cp = D1 Cq D. Then the linear transformations are related by
conjugation
(383)
MCp = L1
D MCq LD
so that Hp = L1
D Hq LD and Hp and Hq are isomorphic. Then if M is (pathwise) connected, i.e. any two points can be joined by a curve in M, then the
holonomy group H is independent of the point p.
For a connected manifold, the holonomy H of a connection is then H
GL(m, R). If the manifold is orientible and the connection preserves the
orientation, then H SL(m, R). For a metric connection, the lengths of
vectors are preserved by the holonomies, so that H O(m) for Riemannian
manifolds or H O(m1, 1) for Lorentzian ones. If the manifold is orientible
and the connection is metric then H SO(m) for Riemannian manifolds or
H SO(m 1, 1) for Lorentzian ones, while for a metric of signature (s, t),
H SO(s, t). For a product space M1 M2 with connection 1 + 2
where i is a connection on Mi with holonomy Hi (i = 1, 2), the holonomy
is H1 H2 .

162

Bergers Classification
In 1955, Berger gave a complete classification of possible holonomy groups
for the Lev-Civita connection for simply connected, Riemannian manifolds
which are irreducible (not locally a product space) and nonsymmetric (not
locally a coset space G/H with Lie groups G, H). Bergers list is as follows:

Holonomy dim
SO(n)
n
U(n)
2n

Type of manifold
Orientable manifold
Khler manifold

SU(n)

2n

CalabiYau manifold

Sp(n)
Sp(1)

4n

Quaternion-Khler
manifold

Sp(n)

4n

Hyperkhler manifold

G2
Spin(7)

7
8

G2 manifold
Spin(7) manifold

Comments

Khler
Ricci-flat,
Khler
Einstein
Ricci-flat,
Khler
Ricci-flat
Ricci-flat

Kahler, hyperkahler and Calabi-Yau manifolds are all complex manifolds.


Quaternion-Kahler manifolds are sometimes called Quaternionic manifolds.

163

7.16

Non-coordinate basis

Given a chart U with coordinates x there is a natural choice of basis for


Tp M and Tp M at every point p in that chart, namely the coordinate basis
{e = x } and its dual {e = dx }. However, it is sometimes useful to
consider a general basis. At each point p U we choose a set of basis vectors
{
ea (p)} for Tp M with a = 1, . . . m and require that and require that the ea (p)
are smooth vector fields on U .
Each basis vector ea can be written as a linear combination of the coordinate
basis vectors e :
ea = ea e ,

ea GL(m, R)

(384)

with components ea (x).


Terminology
The matrices ea are called the zweibeins/dreibeins/vierbeins/vielbeins for a
2/3/4/> 4 dimensional manifold.

Given a metric g, the components of g in the non-coordinate basis will be


gab = g(
ea , eb )

(385)

At a point p, this is diagonalisable: one can change to a new orthonormal


basis e0a
e0a = ab eb ,

ab GL(m, R)

(386)

such that
0
gab
= g(
e0a , e0b )

(387)

is diagonal with eigenvalues 1. This can be done for all points p U , with
a local GL(m, R) transformation ab which is a smooth function of x. We
will always restrict to such orthonormal bases.

164

For a Riemannian metric, we have (in an orthonormal basis)


g(
ea , eb ) = ab

(388)

and for pseudo-Riemannian signature (1, m1) we require (in an orthonormal


basis)
g(
ea , eb ) = ab

(389)

with ab = diag(1, +1, . . .+1) the Minkowski metric. For pseudo-Riemannian


signature (t, s) we require (in an orthonormal basis)
g(
ea , eb ) = ab

(390)

with ab = diag(1, . . . 1, +1, . . . + 1) the metric with t 1s and s +1s.


Hence we see the metric is the constant flat metric in the non-coordinate
basis.
We will deal with the case of signature (t, s) in what follows and use ab for
the metric ab = diag(1, . . . 1, +1, . . . + 1) with t 1s and s +1s, so
that for the Riemannian case with t = 0, we have ab = ab .
We will also require that the non-coordinate basis has the same orientation
as the coordinate basis, ie.
det ea > 0
We shall see later that this is useful.

165

(391)

Index notation
We will use roman labels for non-coordinate basis indices and greek for coordinate basis indices. Our relation above g(
ea , eb ) = ab translates to the
following component statement,
g ea ea = ab

(392)

Let us introduce the following notation to represent the inverse of the m m


matrix ea as
ea (
ea )1

(393)

so that we may conveniently write


ea eb = eb ea = ba

(394)

Using this notation we may invert the relation above to give the metric in
terms of the es
g = ea eb ab

(395)

As an example we can write a vector as


V = V e = V a ea
where the non-coordinate components are given as V a = ea V .

166

(396)

Dual non-coordinate basis


We may define a dual non-coordinate basis {a } so that,
< a , eb >= ba

(397)

(for any signature). Then we see,


a = ea dx

(398)

For example we can now write a covector as,


= dx =
a a

(399)

with the non-cordinate components


a = ea .
We see our index notation has been so that we can think of multiplication
by ea and ea as converting the coordinate basis tensor components to noncoordinate basis ones.

167

Form of the metric


The metric now takes a very simple form in the non-coordinate basis,
g = g dx dx = ab a b

(400)

(in the Riemannian case). However, we must recall that the non-coordinate
basis is not a coordinate basis! Hence,
[e , e ] = 0 ,

but

ec 6= 0
[
ea , eb ] = cabc (p)

(401)

where the coefficients cabc (p) (sometimes called the object of anholonomy)
depend on the matrices ea (x) and their derivatives.
In a non-coordinate basis, so that a = ea dx then we find,
g = 1 2 . . . m

(402)

Let us check this.


1 . . . m = e11 . . . emm dx1 . . . dxm

= det ea dx1 . . . dxm

(403)

2
Now as g = ea eb ab then we have det g = (1)t det ea . Recall that
p
we restricted our vierbeins to have det ea > 0, and hence | det g | =
det ea , showing the result claimed above is indeed true.

168

7.17

Connections and curvature in non-coordinate bases

Using a non-coordinate basis {


ea } and dual {a } we define the connection
components,
a eb ea eb = ab c ec

(404)

We note that
a eb =

ea

(
eb e ) =

ea

e + eb e
x b

and hence the connection components are related as,





c

e + eb
ab = ea e
x b

(405)

(406)

For a metric connection, a g = 0, which implies




0 = a bc b c


d d
c
c b
b d

= bc ad + ad c
= 2a(bc) b c

(407)

where abc = abd dc , and hence a(bc) = 0, so that


abc = acb

(408)

This enables us to write the covariant derivative of a vector field


V = V e = V a ea

(409)

with respect to a vector field


a ea
X = X e = X

(410)

= X (V a ea )
= X[V a ]
ea + V a X ea
b c ec
= (X V a )
ea + V a X
ba

a
a
b

= X ( V +
b V )
ea

(411)

as
X V

169

where

a b = ec bca

(412)

The components of the torsion tensor are given by


Tbc

= < a , T (
eb , ec ) >=< a , b ec c eb [
eb , ec ] >
a
a
a
= bc cb cbc

(413)

where [
eb , ec ] = cbc a (p)
ea .
Likewise the components of the Riemann tensor are given as,
Ra dbc = < a , R(
eb , ec , ed ) >=< a , b c ed c b ed [eb ,ec ] ed >

170

7.18

Cartans structure equations

Since both the torsion and curvature have an explicit antisymmetry, it is


natural to define forms from them. We do so as follows;
Torsion 2-form
We define the vector valued 2-form;
1
Ta Tbc a b c
2

(414)

where we note that the antisymmetry Tbc a = Tcb a makes this natural to
define.
Curvature 2-form We define the matrix valued 2-form;
a d 1 Ra dbc b c
R
2

(415)

where again the antisymmetry Ra dbc = Ra dcb makes this natural to define.
Together with these 2-forms we also define a connection 1-form.
Spin connection 1-form
We define the matrix valued 1-form

a c bc a b

(416)

Using
a b = ec bca this can also be written as

ac =
a b dx

(417)

Cartans structure equations


This enables us to write the torsion and curvature in a particularly elegant
way using differential form technology,
Ta = da +
a b b
a b = d
R
ab +
ac
cb

(418)

and these are known as Cartans structure equations.


[Ex. confirm these give the correct form for the torsion and curvature].

171

We note that for a metric connection, since abc = acb we have

ab =
ba

(419)

ab =
with
ab ac
c b . Then Cartans structure equations imply that R
ba . In a coordinate basis this translates into the result for the curvature
R
for a metric connection
R = R

(420)

which is familiar from GR.


Bianchi identities
Taking the exterior derivative of both of these equations and using the fact
that d is nilpotent, we obtain
a b b
dTa +
a b Tb = R
ab +
cb R
ac
dR
ac R
cb = 0

(421)

These are the Bianchi identities that are familiar from GR, although we note
we have included torsion.
No torsion
In the absence of torsion, we see from the structure equations and Bianchi
identities
da +
a b b = 0
a b b = 0
R
ab +
cb R
ac
dR
ac R
cb = 0

(422)

In a coordinate basis, the 2nd and 3rd equations here become the familiar
expressions from GR,
R [] = 0
[ R ] = 0

172

(423)

Solving Einsteins equations using vierbeins


Taking the Levi-Civita connection, which is a metric connection with no
torsion, then,
ab = ba

and

da +
a b b = 0

(424)

Given a metric, one may then choose a convenient non-coordinate basis, and
hence the set {a }, and the above conditions then determine a b , and hence
the connection a method introduced by Misner. One then solves for the
curvature from the second Cartan structure equation.
An alternative approach is to observe that for vanishing torsion we my use
equation (413) together with the metric condition abc = acb to solve for
the abc components.

173

7.19

Change of Basis and the Local Frame

In a patch U , any 2 sets of basis vector fields ea , e0a will be related at a point
p by
b eb ,
e0a =
a

b GL(m, R)

(425)

b (p) GL(m, R) which will depend smoothly on the point p, i.e.


for some
a
these will be local GL(m, R) transformations. For orthonormal frames, the
ab (p) will be restricted to be in SO(m) for an oriented
local transformations
Riemannian manifold or in SO(m 1, 1) for an oriented Lorentzian mani ab (p) is in SO(t, s). These
fold. For an oriented manifold of signature (t, s),
are then referred to as local frame rotations or local Lorentz transformations.
In an overlap U U 0 , the frames ea in U and e0a in U 0 must be related by a
ab eb transformation of this kind (in GL(m, R), SO(m) or SO(m1, 1))
local
ab eb ,
e0a =

b GL(m, R)

(426)

There will be corresponding transformation of the dual bases; in order that


< a , eb >=< 0a , e0b >= ba

(427)

be preserved,
0a = ab b ,

1 ) b
ab (
a

(428)

A vector field
V = V a ea = V 0a e0a

(429)

will then have frame components transforming as


V 0a = ab V b ,

1 ) b
ab (
a

so that V a ea = V 0a e0a .
The torsion 2-form
Ta = da +
a b b
174

(430)

transforms to
T0a = d0a +
0a b 0b


= a dc +
c b b
c

(431)

so that, using matrix notation,

0 =
1 d1

(432)

The vielbein defines a local frame at a point p, a frame in which the metric
is the flat metric . For a Lorentzian manifold in GR, this can be thought
of as defining a local inertial frame at a point p in the spacetime. The local
frame rotations provide a local SO(t, s) gauge symmetry, with
c b acting as
the gauge field. The metric g is a symmetric matrix and so has m(m + 1)/2
degrees of freedom. The vielbein ea has m2 degrees of freedom, but the gauge
parameters a b have m(m 1)/2 independent components. The local gauge
symmetry can be used to remove m(m 1)/2 of the m2 degrees of freedom
of the vielbein ea , leaving m(m + 1)/2 independent degrees of freedom. In
GR, the m(m+1)/2 (off-shell) degrees of freedom of the gravitational field in
m dimensions can be represented by the components of a symmetric metric
g or by the m2 degrees of freedom of the vielbein ea , identified under the
action of the gauge group of dimension m(m 1)/2.

175

7.20

Tangent Bundle

At every point p in M (an m-diml manifold) we have a tangent space Tp M.

The tangent bundle of M, denoted T M is simply the collection of all these


tangent spaces.
T M pM Tp M
(433)
An element Vp T M gives both a point p M and a tangent vector Vp
at that point. It is also a manifold, with dimension 2m and we will now
construct it explicitly.
Locally the tangent space is a product manifold. Consider a chart Ui on M.
Take coordinates {x(i) }. Then we can represent some vector Vp = v x |p
(i)

using the coordinate basis, where the components v = (v1 , . . . , vm ) = Rm .


An open set Ui is also a manifold, and the tangent space over this open set
is given by the product manifold,
T Ui = Ui Rm

(434)

Hence we can use this fact to provide an altas for T M as follows.


Construct charts on T M as {Vi , i }, where Vi = Ui Rm . These are homeomorphic to Rm Rm = R2m - hence T M is a 2m-diml manifold.
176

A point in Ui Rm is labelled by the pair (p, v ), where p Ui and v =


(v1 , . . . , vm ) labels a point in Rm . Our homeomorphism i is given by
i :

Ui Rm R2m
1
m
(p, v ) (x(i) , v ) = (x1(i) , . . . , xm
(i) , v , . . . , v )

(435)

We now have an altas.


Hence we see that a point in T Ui can be associated with a point p M
and vector Vp Tp M whose components are v in the coordinate basis i.e.
Vp = v x |p
(i)

We must check the transition functions for our altas is smooth. Consider a
point q Vi Vj , which corresponds to a point p Ui Uj . Then any tangent
vector Vp in Tp M can be written as
Vp =

v(i)

x(j)

|p = v(j) |p = v(i)
|p
x(i)
x(j)
x(i) x(j)

(436)

Hence we see the transition function taking coordinates from Vi0 to Vj0 is given
by
ij : (x(i) , v ) (x(j) (x(i) ), M v )
where,
M v

x(j)
x(i)

|p GL(m, R)

(437)

(438)

and so the matrix M is an element of the general linear group (i.e. an m m


matrix with non-zero determinant so it is invertible). Are these transition
functions smooth? Yes - since M is invertible, and x(j) (x(i) ) are smooth, M
is also smooth.
The co-tangent bundle has an analogous construction and is UpM Tp M. It
is denoted T M.

177

Você também pode gostar