An Introduction To Geometrical Physics (R. Aldrovandi, J. G. Pereira)

An Introduction to
GEOMETRICAL PHYSICS
R. Aldrovandi & J.G. Pereira
Instituto de Fsica Teorica
State University of Sao Paulo UNESP
Sao Paulo Brazil
To our parents
Nice, Dina, Jose and Tito
ii
PREAMBLE: SPACE AND GEOMETRY
What stuff tis made of, whereof it is born,

I am to learn.
Merchant of Venice
The simplest geometrical setting used consciously or not by physicists in their everyday work is the 3-dimensional euclidean space E3 . It consists of the set R3 of ordered triples of real numbers such as p = (p1 , p2 , p3 ), q
= (q 1 , q 2 , q 3 ), etc, and is endowed with a very special characteristic, a metric
defined by the distance function
" 3
#1/2
X
d(p, q) =
(pi q i )2
.
i=1
It is the space of ordinary human experience and the starting point of our
geometric intuition. Studied for two-and-a-half millenia, it has been the
object of celebrated controversies, the most famous concerning the minimum
number of properties necessary to define it completely.
From Aristotle to Newton, through Galileo and Descartes, the very word
space has been reserved to E3 . Only in the 19-th century has it become clear
that other, different spaces could be thought of, and mathematicians have
since greatly amused themselves by inventing all kinds of them. For physicists, the age-long debate shifted to another question: how can we recognize,
amongst such innumerable possible spaces, that real space chosen by Nature
as the stage-set of its processes? For example, suppose the space of our everyday experience consists of the same set R3 of triples above, but with a
different distance function, such as
d(p, q) =
3
X
|pi q i |.
i=1
This would define a different metric space, in principle as good as that

given above. Were it only a matter of principle, it would be as good as
iii
iv
any other space given by any distance function with R3 as set point. It so
happens, however, that Nature has chosen the former and not the latter space
for us to live in. To know which one is the real space is not a simple question
of principle something else is needed. What else? The answer may seem
rather trivial in the case of our home space, though less so in other spaces
singled out by Nature in the many different situations which are objects of
physical study. It was given by Riemann in his famous Inaugural Address1 :
... those properties which distinguish Space from other conceivable triply extended quantities can only be deduced from experience.
Thus, from experience! It is experiment which tells us in which space we
actually live in. When we measure distances we find them to be independent
of the direction of the straight lines joining the points. And this isotropy
property rules out the second proposed distance function, while admitting
the metric of the euclidean space.
In reality, Riemanns statement implies an epistemological limitation: it
will never be possible to ascertain exactly which space is the real one. Other
isotropic distance functions are, in principle, admissible and more experiments are necessary to decide between them. In Riemanns time already
other geometries were known (those found by Lobachevsky and Boliyai) that
could be as similar to the euclidean geometry as we might wish in the restricted regions experience is confined to. In honesty, all we can say is that
E3 , as a model for our ambient space, is strongly favored by present day
experimental evidence in scales ranging from (say) human dimensions down
to about 1015 cm. Our knowledge on smaller scales is limited by our capacity to probe them. For larger scales, according to General Relativity, the
validity of this model depends on the presence and strength of gravitational
fields: E3 is good only as long as gravitational fields are very weak.
These data are like all data not logically necessary,
but only of empirical certainty . . . one can therefore investigate
their likelihood, which is certainly very great within the bounds of
observation, and afterwards decide upon the legitimacy of extending them beyond the bounds of observation, both in the direction of
the immeasurably large and in the direction of the immeasurably
small.
1
A translation of Riemanns Address can be found in Spivak 1970, vol. II. Cliffords
translation (Nature, 8 (1873), 14-17, 36-37), as well as the original transcribed by David
R. Wilkins, can be found in the site http://www.emis.de/classics/Riemann/.
v
The only remark we could add to these words, pronounced in 1854, is
that the bounds of observation have greatly receded with respect to the
values of Riemann times.
. . . geometry presupposes the concept of space, as well as
assuming the basic principles for constructions in space .
In our ambient space, we use in reality a lot more of structure than
the simple metric model: we take for granted a vector space structure, or
an affine structure; we transport vectors in such a way that they remain
parallel to themselves, thereby assuming a connection. Which one is the
minimum structure, the irreducible set of assumptions really necessary to
the introduction of each concept? Physics should endeavour to establish on
empirical data not only the basic space to be chosen but also the structures
to be added to it. At present, we know for example that an electron moving
in E3 under the influence of a magnetic field feels an extra connection (the
electromagnetic potential), to which neutral particles may be insensitive.
Experimental science keeps a very special relationship with Mathematics. Experience counts and measures. But Science requires that the results
be inserted in some logically ordered picture. Mathematics is expected to
provide the notion of number, so as to make countings and measurements
meaningful. But Mathematics is also expected to provide notions of a more
qualitative character, to allow for the modeling of Nature. Thus, concerning
numbers, there seems to be no result comforting the widespread prejudice
by which we measure real numbers. We work with integers, or with rational
numbers, which is fundamentally the same. No direct measurement will sort
out a Dedekind cut. We must suppose, however, that real numbers exist:
even from the strict experimental point of view, it does not matter whether
objects like or e are simple names or are endowed with some kind of an
sich reality: we cannot afford to do science without them. This is to say that
even pure experience needs more than its direct results, presupposes a wider
background for the insertion of such results. Real numbers are a minimum
background. Experience, and logical necessity, will say whether they are
sufficient.
From the most ancient extant treatise going under the name of Physics2 :
When the objects of investigation, in any subject, have first
principles, foundational conditions, or basic constituents, it is
through acquaintance with these that knowledge, scientific knowledge, is attained. For we cannot say that we know an object before
2
Aristotle, Physics I.1.
vi
we are acquainted with its conditions or principles, and have carried our analysis as far as its most elementary constituents.
The natural way of attaining such a knowledge is to start
from the things which are more knowable and obvious to us and
proceed towards those which are clearer and more knowable by
themselves . . .
Euclidean spaces have been the starting spaces from which the basic geometrical and analytical concepts have been isolated by successive, tentative,
progressive abstractions. It has been a long and hard process to remove the
unessential from each notion. Most of all, as will be repeatedly emphasized,
it was a hard thing to put the idea of metric in its due position.
Structure is thus to be added step by step, under the control of experiment. Only once experiment has established the basic ground will internal
coherence, or logical necessity, impose its own conditions.
Contents
I
MANIFOLDS
1 GENERAL TOPOLOGY
1.0 INTRODUCTORY COMMENTS
1.1 TOPOLOGICAL SPACES . . .
1.2 KINDS OF TEXTURE . . . . .
1.3 FUNCTIONS . . . . . . . . . . .
1.4 QUOTIENTS AND GROUPS . .
1.4.1 Quotient spaces . . . . .
1.4.2 Topological groups . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 HOMOLOGY
2.1 GRAPHS . . . . . . . . . . . . . . . . . . .
2.1.1 Graphs, first way . . . . . . . . . .
2.1.2 Graphs, second way . . . . . . . . .
2.2 THE FIRST TOPOLOGICAL INVARIANTS
2.2.1 Simplexes, complexes & all that . .
2.2.2 Topological numbers . . . . . . . .
3 HOMOTOPY
3.0 GENERAL HOMOTOPY . . . . .
3.1 PATH HOMOTOPY . . . . . . . .
3.1.1 Homotopy of curves . . . .
3.1.2 The Fundamental group .
3.1.3 Some Calculations . . . .
3.2 COVERING SPACES . . . . . .
3.2.1 Multiply-connected Spaces
3.2.2 Covering Spaces . . . . . .
3.3 HIGHER HOMOTOPY . . . . .
vii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
5
15
27
36
36
41
.
.
.
.
.
.
49
50
50
52
57
57
64
.
.
.
.
.
.
.
.
.
73
73
78
78
85
92
98
98
105
115
viii
CONTENTS
4 MANIFOLDS & CHARTS

4.1 MANIFOLDS . . . . . . . . . . . . .
4.1.1 Topological manifolds . . . . .
4.1.2 Dimensions, integer and other
4.2 CHARTS AND COORDINATES . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
121
. 121
. 121
. 123
. 125
5 DIFFERENTIABLE MANIFOLDS
133
5.1 DEFINITION AND OVERLOOK . . . . . . . . . . . . . . . . . 133
5.2 SMOOTH FUNCTIONS . . . . . . . . . . . . . . . . . . . . . . 135
5.3 DIFFERENTIABLE SUBMANIFOLDS . . . . . . . . . . . . . . 137
II
DIFFERENTIABLE STRUCTURE
6 TANGENT STRUCTURE
6.1 INTRODUCTION . . . . . . . . . . . . .
6.2 TANGENT SPACES . . . . . . . . . . . .
6.3 TENSORS ON MANIFOLDS . . . . . . .
6.4 FIELDS & TRANSFORMATIONS . . . .
6.4.1 Fields . . . . . . . . . . . . . . .
6.4.2 Transformations . . . . . . . . . .
6.5 FRAMES . . . . . . . . . . . . . . . . . .
6.6 METRIC & RIEMANNIAN MANIFOLDS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
141
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 DIFFERENTIAL FORMS
7.1 INTRODUCTION . . . . . . . . . . . . . . . . . . .
7.2 EXTERIOR DERIVATIVE . . . . . . . . . . . . . .
7.3 VECTOR-VALUED FORMS . . . . . . . . . . . . .
7.4 DUALITY AND CODERIVATION . . . . . . . . . .
7.5 INTEGRATION AND HOMOLOGY . . . . . . . . .
7.5.1 Integration . . . . . . . . . . . . . . . . . . .
7.5.2 Cohomology of differential forms . . . . . . .
7.6 ALGEBRAS, ENDOMORPHISMS AND DERIVATIVES
8 SYMMETRIES
8.1 LIE GROUPS . . . . . . . . . . . . . . .
8.2 TRANSFORMATIONS ON MANIFOLDS .
8.3 LIE ALGEBRA OF A LIE GROUP . . .
8.4 THE ADJOINT REPRESENTATION .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
143
. 143
. 145
. 154
. 161
. 161
. 167
. 175
. 180
.
.
.
.
.
.
.
.
189
. 189
. 197
. 210
. 217
. 225
. 225
. 232
. 239
.
.
.
.
247
. 247
. 252
. 259
. 265
CONTENTS
ix
9 FIBER BUNDLES
9.1 INTRODUCTION . . . . . . . . . .
9.2 VECTOR BUNDLES . . . . . . . . .
9.3 THE BUNDLE OF LINEAR FRAMES
9.4 LINEAR CONNECTIONS . . . . . .
9.5 PRINCIPAL BUNDLES . . . . . . .
9.6 GENERAL CONNECTIONS . . . .
9.7 BUNDLE CLASSIFICATION . . . .
III
273
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
FINAL TOUCH
.
.
.
.
.
.
.
273
275
277
284
297
303
316
321
10 NONCOMMUTATIVE GEOMETRY
323
10.1 QUANTUM GROUPS A PEDESTRIAN OUTLINE . . . . . . 323
10.2 QUANTUM GEOMETRY . . . . . . . . . . . . . . . . . . . . 326
IV
MATHEMATICAL TOPICS
331
1 THE BASIC ALGEBRAIC STRUCTURES

1.1 Groups and lesser structures . . . . . . . . .
1.2 Rings and fields . . . . . . . . . . . . . . .
1.3 Modules and vector spaces . . . . . . . . . .
1.4 Algebras . . . . . . . . . . . . . . . . . . .
1.5 Coalgebras . . . . . . . . . . . . . . . . . .
333
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
334
338
341
344
348
2 DISCRETE GROUPS. BRAIDS AND KNOTS

351
2.1 A Discrete groups . . . . . . . . . . . . . . . . . . . . . . . . . 351
2.2 B Braids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
2.3 C Knots and links . . . . . . . . . . . . . . . . . . . . . . . . . 363
3 SETS AND MEASURES
371
3.1 MEASURE SPACES . . . . . . . . . . . . . . . . . . . . . . . . 371
3.2 ERGODISM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
4 TOPOLOGICAL LINEAR SPACES
4.1 Inner product space . . . . . . .
4.2 Norm . . . . . . . . . . . . . . . .
4.3 Normed vector spaces . . . . . . .
4.4 Hilbert space . . . . . . . . . . .
4.5 Banach space . . . . . . . . . . .
4.6 Topological vector spaces . . . . .
379
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
379
380
380
380
382
382
CONTENTS
4.7
Function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 383
5 BANACH ALGEBRAS
5.1 Quantization . . . . . . .
5.2 Banach algebras . . . . . .
5.3 *-algebras and C*-algebras
5.4 From Geometry to Algebra
5.5 Von Neumann algebras . .
5.6 The Jones polynomials . .
385
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
385
387
389
390
393
397
6 REPRESENTATIONS
403
6.1 A Linear representations . . . . . . . . . . . . . . . . . . . . . 404
6.2 B Regular representation . . . . . . . . . . . . . . . . . . . . . 408
6.3 C Fourier expansions . . . . . . . . . . . . . . . . . . . . . . . 409
7 VARIATIONS & FUNCTIONALS
7.1 A Curves . . . . . . . . . . . . . . . . . .
7.1.1 Variation of a curve . . . . . . . . .
7.1.2 Variation fields . . . . . . . . . . .
7.1.3 Path functionals . . . . . . . . . . .
7.1.4 Functional differentials . . . . . . .
7.1.5 Second-variation . . . . . . . . . .
7.2 B General functionals . . . . . . . . . . . .
7.2.1 Functionals . . . . . . . . . . . . .
7.2.2 Linear functionals . . . . . . . . .
7.2.3 Operators . . . . . . . . . . . . . .
7.2.4 Derivatives Frechet and Gateaux
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 FUNCTIONAL FORMS
8.1 A Exterior variational calculus . . . . . . . . . .
8.1.1 Lagrangian density . . . . . . . . . . .
8.1.2 Variations and differentials . . . . . . .
8.1.3 The action functional . . . . . . . . .
8.1.4 Variational derivative . . . . . . . . . .
8.1.5 Euler Forms . . . . . . . . . . . . . . .
8.1.6 Higher order Forms . . . . . . . . . .
8.1.7 Relation to operators . . . . . . . . .
8.2 B Existence of a lagrangian . . . . . . . . . . .
8.2.1 Inverse problem of variational calculus
8.2.2 Helmholtz-Vainberg theorem . . . . . .
8.2.3 Equations with no lagrangian . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
415
. 415
. 415
. 416
. 417
. 418
. 420
. 421
. 421
. 422
. 423
. 423
425
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
426
426
427
428
428
429
429
429
430
430
430
431
CONTENTS
8.3
xi
. . . . .
The homotopy formula .
Examples . . . . . . . .
Symmetries of equations
C Building lagrangians
8.3.1
8.3.2
8.3.3
.
.
.
.
.
.
.
.
9 SINGULAR POINTS
9.1 Index of a curve . . . . . . . . . . .
9.2 Index of a singular point . . . . . .
9.3 Relation to topology . . . . . . . .
9.4 Basic two-dimensional singularities
9.5 Critical points . . . . . . . . . . . .
9.6 Morse lemma . . . . . . . . . . . .
9.7 Morse indices and topology . . . .
9.8 Catastrophes . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
432
432
434
436
439
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10 EUCLIDEAN SPACES AND SUBSPACES

10.1 A Structure equations . . . . . . . . . . .
10.1.1 Moving frames . . . . . . . . . . .
10.1.2 The Cartan lemma . . . . . . . . .
10.1.3 Adapted frames . . . . . . . . . . .
10.1.4 Second quadratic form . . . . . . .
10.1.5 First quadratic form . . . . . . . .
10.2 B Riemannian structure . . . . . . . . . .
10.2.1 Curvature . . . . . . . . . . . . . .
10.2.2 Connection . . . . . . . . . . . . .
10.2.3 Gauss, Ricci and Codazzi equations
10.2.4 Riemann tensor . . . . . . . . . . .
10.3 C Geometry of surfaces . . . . . . . . . . .
10.3.1 Gauss Theorem . . . . . . . . . . .
10.4 D Relation to topology . . . . . . . . . . .
10.4.1 The Gauss-Bonnet theorem . . . .
10.4.2 The Chern theorem . . . . . . . . .
11 NON-EUCLIDEAN GEOMETRIES
11.1 The old controversy . . . . . . . .
11.2 The curvature of a metric space .
11.3 The spherical case . . . . . . . . .
11.4 The Boliyai-Lobachevsky case . .
11.5 On the geodesic curves . . . . . .
11.6 The Poincare space . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
439
442
443
443
444
445
446
447
449
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
450
450
450
450
451
451
452
452
452
453
453
455
455
457
457
458
459
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
459
460
461
464
466
467
xii
CONTENTS
12 GEODESICS
12.1 Selfparallel curves . . . . . . . . . . .
12.1.1 In General Relativity . . . . . .
12.1.2 The absolute derivative . . . . .
12.1.3 Selfparallelism . . . . . . . . .
12.1.4 Complete spaces . . . . . . . .
12.1.5 Fermi transport . . . . . . . . .
12.1.6 In Optics . . . . . . . . . . . .
12.2 Congruences . . . . . . . . . . . . . .
12.2.1 Jacobi equation . . . . . . . . .
12.2.2 Vorticity, shear and expansion .
12.2.3 LandauRaychaudhury equation
471
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
472
472
473
474
475
475
476
476
476
480
483
PHYSICAL TOPICS
485
1 HAMILTONIAN MECHANICS
1.1 Introduction . . . . . . . . . . .
1.2 Symplectic structure . . . . . .
1.3 Time evolution . . . . . . . . .
1.4 Canonical transformations . . .
1.5 Phase spaces as bundles . . . .
1.6 The algebraic structure . . . . .
1.7 Relations between Lie algebras .
1.8 Liouville integrability . . . . . .
487
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
MORE MECHANICS
2.1
HamiltonJacobi . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Hamiltonian structure . . . . . . . . . . . . . . .
2.1.2 Hamilton-Jacobi equation . . . . . . . . . . . . .
2.2 The Lagrange derivative . . . . . . . . . . . . . . . . . .
2.2.1 The Lagrange derivative as a covariant derivative
2.3 The rigid body . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Frames . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 The configuration space . . . . . . . . . . . . . .
2.3.3 The phase space . . . . . . . . . . . . . . . . . . .
2.3.4 Dynamics . . . . . . . . . . . . . . . . . . . . . .
2.3.5 The space and the body derivatives . . . . .
2.3.6 The reduced phase space . . . . . . . . . . . . . .
2.3.7 Moving frames . . . . . . . . . . . . . . . . . . .
2.3.8 The rotation group . . . . . . . . . . . . . . . . .
487
488
490
491
494
496
498
501
503
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
503
503
505
507
507
510
510
511
511
512
513
513
514
515
CONTENTS
xiii
2.3.9 Left and rightinvariant fields . . . . . . . . . . . . . 515

2.3.10 The Poinsot construction . . . . . . . . . . . . . . . . . 518
3 STATISTICS AND ELASTICITY
3.1 A Statistical Mechanics . . . . . . . . . . . .
3.1.1 Introduction . . . . . . . . . . . . . .
3.1.2 General overview . . . . . . . . . . .
3.2 B Lattice models . . . . . . . . . . . . . . .
3.2.1 The Ising model . . . . . . . . . . . .
3.2.2 Spontaneous breakdown of symmetry
3.2.3 The Potts model . . . . . . . . . . .
3.2.4 Cayley trees and Bethe lattices . . .
3.2.5 The four-color problem . . . . . . . .
3.3 C Elasticity . . . . . . . . . . . . . . . . . .
3.3.1 Regularity and defects . . . . . . . .
3.3.2 Classical elasticity . . . . . . . . . .
3.3.3 Nematic systems . . . . . . . . . . .
3.3.4 The Franck index . . . . . . . . . . .
4 PROPAGATION OF DISCONTINUITIES
4.1 Characteristics . . . . . . . . . . . . . .
4.2 Partial differential equations . . . . . . .
4.3 Maxwells equations in a medium . . . .
4.4 The eikonal equation . . . . . . . . . . .
5 GEOMETRICAL OPTICS
5.0 Introduction . . . . . . .
5.1 The light-ray equation .
5.2 Hamiltons point of view
5.3 Relation to geodesics . .
5.4 The Fermat principle . .
5.5 Maxwells fish-eye . . . .
5.6 Fresnels ellipsoid . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 CLASSICAL RELATIVISTIC FIELDS

6.1 A The fundamental fields . . . . . . .
6.2 B Spacetime transformations . . . . .
6.3 C Internal transformations . . . . . .
6.4 D Lagrangian formalism . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
521
. 521
. 521
. 522
. 526
. 526
. 529
. 531
. 535
. 536
. 537
. 537
. 542
. 547
. 550
.
.
.
.
553
. 553
. 554
. 558
. 561
.
.
.
.
.
.
.
565
. 565
. 566
. 567
. 568
. 570
. 571
. 572
.
.
.
.
575
. 575
. 576
. 579
. 579
xiv
CONTENTS
7 GAUGE FIELDS
7.1 A The gauge tenets . . . . . . . . . . .
7.1.1 Electromagnetism . . . . . . . .
7.1.2 Nonabelian theories . . . . . . .
7.1.3 The gauge prescription . . . . .
7.1.4 Hamiltonian approach . . . . .
7.1.5 Exterior differential formulation
7.2 B Functional differential approach . . .
7.2.1 Functional Forms . . . . . . . .
7.2.2 The space of gauge potentials .
7.2.3 Gauge conditions . . . . . . . .
7.2.4 Gauge anomalies . . . . . . . .
7.2.5 BRST symmetry . . . . . . . .
7.3 C Chiral fields . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
589
. 590
. 590
. 591
. 593
. 594
. 595
. 596
. 596
. 598
. 601
. 602
. 603
. 603
8 GENERAL RELATIVITY
605
8.1 Einsteins equation . . . . . . . . . . . . . . . . . . . . . . . . 605
8.2 The equivalence principle . . . . . . . . . . . . . . . . . . . . . 608
8.3 Spinors and torsion . . . . . . . . . . . . . . . . . . . . . . . . 612
9 DE
9.1
9.2
9.3
9.4
9.5
SITTER SPACES
General characteristics . . . . .
Curvature . . . . . . . . . . . .
Geodesics and Jacobi equations
Some qualitative aspects . . . .
Wigner-Inon
u contraction . . .
.
.
.
.
.
.
.
.
.
.
10 SYMMETRIES ON PHASE SPACE

10.1 Symmetries and anomalies . . . . .
10.2 The Souriau momentum . . . . . .
10.3 The Kirillov form . . . . . . . . . .
10.4 Integrability revisited . . . . . . . .
10.5 Classical Yang-Baxter equation . .
VI
Glossary and Bibliography
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
615
. 615
. 619
. 620
. 621
. 621
625
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
625
628
629
630
631
635
Part I
MANIFOLDS
Chapter 1
GENERAL TOPOLOGY
Or, the purely qualitative properties of spaces.
1.0
INTRODUCTORY COMMENTS
1.0.1 Let us again consider our ambient 3-dimensional euclidean space

E3 . In order to introduce ideas like proximity between points, boundedness
of subsets, convergence of point sequences and the dominating notion
continuity of mappings between E3 and other point sets, elementary real
analysis starts by defining open r-balls around a point p:1

Br (p) = q E3 such that d(q, p) < r .
The same is done for n-dimensional euclidean spaces En , with open r-balls
of dimension n. The question worth raising here is whether or not the real
analysis so obtained depends on the chosen distance function. Or, putting
it in more precise words: of all the usual results of analysis, how much is
dependent on the metric and how much is not? As said in the Preamble,
Physics should use experience to decide which one (if any) is the convenient
metric in each concrete situation, and this would involve the whole body
of properties consequent to this choice. On the other hand, some spaces of
physical relevance, such as the space of thermodynamical variables, are not
explicitly endowed with any metric. Are we always using properties coming
from some implicit underlying notion of distance ?
Defining balls requires the notion of distance function , which is a function d taking
pairs (p, q) of points of a set into the real positive line R+ and obeying certain conditions.
A complete definition is found in the Glossary. Recall that entries in the Glossary are
indicated by an upper dagger .
1
CHAPTER 1. GENERAL TOPOLOGY
1.0.2 There is more: physicists are used to metrics which in reality do

not lead to good distance functions. Think of Minkowski space, which is R4
with the Lorentz metric :
1/2

.
(p, q) = (p0 q 0 )2 (p1 q 1 )2 (p2 q 2 )2 (p3 q 3 )2
It is not possible to define open balls with this pseudo-metric, which allows
vanishing distances between distinct points on the light cone, and even
purely imaginary distances. If continuity, for example, depends upon the
previous introduction of balls, then when would a function be continuous on
Minkowski space?
1.0.3 Actually, most of the properties of space are quite independent of
any notion of distance. In particular, the above mentioned ideas of proximity,
convergence, boundedness and continuity can be given precise meanings in
spaces on which the definition of a metric is difficult, or even forbidden.
Metric spaces are in reality very particular cases of more abstract objects,
the topological spaces, on which only the minimal structure necessary to
introduce those ideas is present. That minimal structure is a topology, and
answers for the general qualitative properties of space.
1.0.4 Consider the usual 2-dimensional surfaces immersed in E3 . To begin with, there is something shared by all spheres, of whatever size. And
also something which is common to all toruses, large or small; and so on.
Something makes a sphere deeply different from a torus and both different
from a plane, and that independently of any measure, scale or proportion. A
hyperboloid sheet is quite distinct from the sphere and the torus, and also
from the plane E2 , but less so for the latter: we feel that it can be somehow
unfolded without violence into a plane. A sphere can be stretched so as to become an ellipsoid but cannot be made into a plane without losing something
of its spherical character. Topology is that primitive structure which will
be the same for spheres and ellipsoids; which will be another one for planes
and hyperboloid sheets; and still another, quite different, for toruses. It will
be that set of qualities of a space which is preserved under suave stretching,
bending, twisting. The study of this primitive structure makes use of very
simple concepts: points, sets of points, mappings between sets of points. But
the structure itself may be very involved and may leave an important (eventually dominant) imprint on the physical objects present in the space under
consideration.
1.0.5 The word topology is like algebra used in two different
senses. One more general, naming the mathematical discipline concerned
1.1. TOPOLOGICAL SPACES
with spacial qualitative relationships, and another, more particular, naming

that structure allowing for such relationships to be well defined. We shall
be using it almost exclusively with the latter, more technical, meaning. Let
us proceed to make the basic ideas a little more definite. In order to avoid
leaving too many unstated assumptions behind, we shall feel justified in
adopting a rather formal approach,2 starting modestly with point sets.
1.1
TOPOLOGICAL SPACES
1.1.1 Experimental measurements being inevitably of limited accuracy,

the constants of Nature (such as Plancks constant ~, the light velocity c,
the electron charge e, etc.) appearing in the fundamental equations are not
known with exactitude. The process of building up Physics presupposes this
kind of stability: it assumes that, if some value for a physical quantity is
admissible, there must be always a range of values around it which is also
acceptable. A wavefunction, for example, will depend on Plancks constant.
Small variations of this constant, within experimental errors, would give other
wavefunctions, by necessity equally acceptable as possible. It follows that,
in the modeling of nature, each value of a mathematical quantity must be
surrounded by other admissible values. Such neighbouring values must also,
by the same reason, be contained in a set of acceptable values. We come thus
to the conclusion that values of quantities of physical interest belong to sets
enjoying the following property: every acceptable point has a neighbourhood
of points equally acceptable, each one belonging to another neighbourhood
of acceptable points, etc, etc. Sets endowed with this property, that around
each one of its points there exists another set of the same kind, are called
open sets. This is actually the old notion of open set, abstracted from
euclidean balls: a subset U of an ambient set S is open if around each
one of its points there is another set of points of S entirely contained in U .
All physically admissible values are, therefore, necessarily members of open
sets. Physics needs open sets. Furthermore, we talk frequently about good
behaviour of functions, or that they tend to some value, thereby loosely
conveying ideas of continuity and limit. Through a succession of abstractions,
the mathematicians have formalized the idea of open set while inserting it in
a larger, more comprehensive context. Open sets appear then as members
of certain families of sets, the topologies, and the focus is concentrated on
the properties of the families, not on those of its members. This enlarged
2
A commendable text for beginners, proceeding constructively from unstructured sets

up to metric spaces, is Christie 1976. Another readable account is the classic Sierpi
nski
1956.
context provides a general and abstract concept of open sets and gives a clear
meaning to the above rather elusive word neighbourhood, while providing
the general background against which the fundamental notions of continuity
and convergence acquire well defined contours.
1.1.2 A space will be, to begin with, a set endowed with some decomposition allowing us to talk about its parts. Although the elements belonging
to a space may be vectors, matrices, functions, other sets, etc, they will be
called, to simplify the language, points. Thus, a space will be a set S of
points plus a structure leading to some kind of organization, such that we
may speak of its relative parts and introduce spatial relationships. This
structure is introduced as a well-performed division of S, as a convenient family of subsets. There are various ways of dividing a set, each one designed to
accomplish a definite objective.
We shall be interested in getting appropriate notions of neighbourhood,
distinguishability of points, continuity and, later, differentiability. How is a
fitting decomposition obtained? A first possibility might be to consider S
with all its subsets. This conception, though acceptable in principle, is too
particular: it leads to a quite disconnected space, every two points belonging
to too many unshared neighbourhoods. It turns out (see section 1.3) that
any function would be continuous on such a pulverized space and in consequence the notion of continuity would be void. The family of subsets is
too large, the decomposition would be too fine-grained. In the extreme
opposite, if we consider only the improper subsets, that is, the whole point
set S and the empty set , there would be no real decomposition and again
no useful definition of continuity (subsets distinct from and S are called
proper subsets). Between the two extreme choices of taking a family with
all the subsets or a family with no subsets at all, a compromise has been
found: good families are defined as those respecting a few well chosen, suitable conditions. Each one of such well-bred families of subsets is called a
topology.
Given a point set S, a topology is a family of subsets of S (which are
called, by definition, its open sets) respecting the 3 following conditions:
(a) the whole set S and the empty set belong to the family;
(b) given a finite number
T of members of the family, say U1 , U2 , U3 , . . . , Un ,
their intersection ni=1 Ui is also a member;
(c) given any number (finite or infinite) of open sets, their union belongs to
the family.
Thus, a topology on S is a collection of subsets of S to which belong the

union of any subcollection and the intersection of any finite subcollection, as
well as and the set S proper. The paradigmatic open balls of En satisfy, of
course, the above conditions. Both the families suggested above, the family
including all subsets and the family including no proper subsets, respect
the above conditions and are consequently accepted in the club: they are
topologies indeed (called respectively the discrete topology and the indiscrete
topology of S), but very peculiar ones. We shall have more to say about them
later (see below, s 1.1.18 and 1.3.5). Now:
a topological space is a point set S
on which a topology is defined.
Given a point set S, there are in general many different families of subsets
with the above properties, i.e., many different possible topologies. Each such
family will make of S a different topological space. Rigour would require that
a name or symbol be attributed to the family (say, T ) and the topological
space be given name and surname, being denoted by the pair (S, T ).
Some well known topological spaces have historical names. When we say
euclidean space, the set Rn with the usual topology of open balls is meant.
The members of a topology are called open sets precisely by analogy with
the euclidean case, but notice that they are determined by the specification
of the family: an open set of (S, T ) is not necessarily an open set of (S, T 0 )
when T 6= T 0 . Think of the point set of En , which is Rn , but with the discrete
topology including all subsets: the set {p} containing only the point p of Rn
is an open set of the topological space (Rn , discrete topology), but not of the
euclidean space En = (Rn , topology of n-dimensional balls).
1.1.3 Finite Space: a very simple topological space is given by the set
of four letters S = {a, b, c, d} with the family of subsets
T = {{a}, {a, b}, {a, b, d}, S, }.
The choice is not arbitrary: the family of subsets
{{a}, {a, b}, {b, c, d}, S, },
for example, does not define a topology, because the intersection
{a, b} {b, c, d} = {b}
is not an open set.
1.1.4 Given a point p S, any set U containing an open set belonging

to T which includes p is a neighbourhood of p. Notice that U itself is not
necessarily an open set of T : it simply includes3 some open set(s) of T . Of
course any point will have at least one neighbourhood, S itself.
1.1.5 Metric spaces are the archetypal topological spaces. The notion of
topological space has evolved conceptually from metric spaces by abstraction:
properties unnecessary to the definition of continuity were progressively forsaken. Topologies generated from a notion of distance (metric topologies) are
the most usual in Physics. As an experimental science, Physics plays with
countings and measurements, the latter in general involving some (at least
implicit) notion of distance. Amongst metric spaces, a fundamental role will
be played by the first example we have met, the euclidean space.
1.1.6 The euclidean space En The point set is the set Rn of n-uples
p = (p1 , p2 , . . . , pn ), q = (q 1 , q 2 , . . . , q n ), etc, of real numbers; the distance
function is given by
" n
#1/2
X
d(p, q) =
(pi q i )2
.
i=1
The topology is formed by the set of the open balls. It is a standard practice
to designate a topological space by its point set when there is no doubt as
to which topology is meant. That is why the euclidean space is frequently
denoted simply by Rn . We shall, however, insist on the notational difference: En will be Rn plus the ball topology. En is the basic, starting space,
as even differential manifolds will be presently defined so as to generalize it.
We shall see later that the introduction of coordinates on a general space S
requires that S resemble some En around each one of its points. It is important to notice, however, that many of the most remarkable properties of the
euclidean space come from its being, besides a topological space, something
else. Indeed, one must be careful to distinguish properties of purely topological nature from those coming from additional structures usually attributed
to En , the main one being that of a vector space.
1.1.7 In metric spaces, any point p has a countable set of open neighbourhoods {Ni } such that for any set U containing p there exists at least one Nj
included in U . Thus, any set U containing p is a neighbourhood. This is not
a general property of topological spaces. Those for which this happens are
said to be first-countable spaces (Figure 1.1).
3
Some authors (Kolmogorov & Fomin 1977, for example) do define a neighbourhood
of p as an open set of T to which p belongs. In our language, a neighbourhood which is
also an open set of T will be an open neighbourhood.
Figure 1.1: In first-countable spaces, every point p has a countable set of open neighbourhoods {Nk }, of which at least one is included in a given U 3 p. We say that all
points have a local countable basis. All metric spaces are of this kind.
1.1.8 Topology basis In order to specify a topological space, one has to

fix the point set and tell which amongst all its subsets are to be taken as
open sets. Instead of giving each member of the family T (which is frequently
infinite to a very high degree), it is in general much simpler to give a subfamily
from which the whole family can be retraced. A basis for a topology T is a
collection B of its open sets such that any member of T can be obtained as
the union of elements of B. A general criterium for B = {U } to be a basis
is stated in the following theorem:
B = {U } is a basis for T iff, for any open set V T and all p V , there
exists some U B such that p U V .
The open balls of En constitute a prototype basis, but one might think of open
cubes, open tetrahedra, etc. It is useful, to get some insight, to think about
open disks, open triangles and open rectangles on the euclidean plane E2 . No
two distinct topologies may have a common basis, but a fixed topology may
have many different basis. On E2 , for instance, we could take the open disks,
or the open squares or yet rectangles, or still the open ellipses. We would
say intuitively that all these different basis lead to the same topology and
we would be strictly correct. As a topology is most frequently introduced
via a basis, it is useful to have a criterium to check whether or not two basis
correspond to the same topology. This is provided by another theorem:
10
B and B 0 are basis defining the same topology iff, for every U B and
every p U , there exists some U0 B 0 such that p B0 U and
vice-versa.
Again, it is instructive to give some thought to disks and rectangles in E2 . A
basis for the real euclidean line E1 is provided by all the open intervals of the
type (r 1/n, r + 1/n), where r runs over the set of rational numbers and n
over the set of the integer numbers. This is an example of countable basis.
When a topology has at least one countable basis, it is said to be secondcountable. Second countable topologies are always first-countable ( 7) but
the inverse is not true. We have said above that all metric spaces are firstcountable. There are, however, metric spaces which are not second countable
(Figure 1.2).
Figure 1.2: A partial hierarchy: not all metric spaces are second-countable, but all of
them are first-countable.
We see here a first trial to classify topological spaces. Topology frequently

resorts to this kind of practice, trying to place the space in some hierarchy.
In the study of the anatomy of a topological space, some variations are sometimes helpful.
An example is a small change in the concept of a basis, leading to the idea of a network.
A network is a collection N of subsets such that any member of T can be obtained as the
union of elements of N . Similar to a basis, but accepting as members also sets which are
not open sets of T .
11
1.1.9 Induced topology The topologies of the usual surfaces immersed

in E3 are obtained by intersecting them with the open 3-dimensional balls.
This procedure can be transferred to the general case: let (S, T ) be a topological space and X a subset of S. A topology can be defined on X by taking
as open sets the intersections of X with the open sets belonging to T . This is
called the induced (or relative) topology, denoted X T . A new topological
space (X, X TP
) is born in this way. An n-sphere S n is the set of points of
n+1
i 2
E
satisfying n+1
i=1 (p ) = 1, with the topology induced by the open balls
of En+1 (Figure 1.3). The set of real numbers can be made into the euclidean
topological space E1 (popular names: the line and rather oldish the
continuum), with the open intervals as 1-dimensional open balls. Both the
set Q of rational numbers and its complement, the set J = E1 \Q of irrational
numbers, constitute topological spaces with the topologies induced by the
euclidean topology of the line.
Figure 1.3: The sphere S 2 with some of its open sets, which are defined as the intersections of S 2 with the open balls of the euclidean 3-dimensional space.
1.1.10 The upper-half space En+ . The point set is

Rn+ = p = (p1 , p2 , . . . , pn ) Rn such that pn 0 .
(1.1)
The topology is that induced by the ball-topology of En . This space, which

will be essential to the definition of manifolds-with-boundary in 4.1.1, is
not second-countable. A particular basis is given by sets of two kinds: (i)
12
all the open balls entirely included in Rn+ ; (ii) for each ball tangent to the
hyperplane pn = 0, the union of that ball with (the set containing only) the
tangency point.
1.1.11 Notice that, for the 2-dimensional case (the upper-half plane, Figure
1.4) for example, sets of type ,
including intersections with the horizontal line,
2
are not open in E but are open in E2+ . One speaks of the above topology as
the swimmers topology: suppose a fluid flows upwardly from the horizontal
borderline into the space with a monotonously decreasing velocity which is unit at
the bottom. A swimmer with a constant unit velocity may start swimming in any
direction at any point of the fluid. In a unit interval of time the set of all possible
swimmers will span a basis.
Figure 1.4: The upper-half plane E2+ , whose open sets are the intersections of the point
set R2+ with the open disks of E2 .
1.1.12 A cautionary remark: the definitions given above (and below) may
sometimes appear rather queer and irksome, as if invented by some skewminded daemon decided to hide simple things under tangled clothes. They
have evolved, however, by a series of careful abstractions, starting from the
properties of metric spaces and painstakingly checked to see whether they
lead to useful, meaningful concepts. Fundamental definitions are, in this
sense, the products of Experimental Mathematics. If a simpler, more direct
definition seems possible, the reader may be sure that it has been invalidated
by some counter-example (see as an example the definition of a continuous
function in section 1.3.4).
13
1.1.13 Consider two topologies T1 and T2 defined on the same point set
S. We say that T1 is weaker than T2 if every member of T1 belongs also to
T2 . The topology T1 is also said to be coarser than T2 , and T2 is finer than
T1 (or T2 is a refinement of T1 , or still T2 is stronger than T1 ). The topology
T for the finite space of 1.1.3 is clearly weaker than the discrete topology
for the same point set.
1.1.14 We have said that the topology for Minkowski space time cannot
be obtained from the Lorentz metric, which is unable to define balls. The
specification of a topology is of fundamental importance because (as will
be seen later) it is presupposed every time we talk of a continuous (say,
wave) function. We could think of using an E4 topology, but this would
be wrong because (besides other reasons) no separation would then exist
between spacelike and timelike vectors. The fact is that we do not know the
real topology of spacetime. We would like to retain euclidean properties both
in the space sector and on the time axis. Zeeman4 has proposed an appealing
topology: it is defined as the finest topology defined on R4 which induces an
E3 topology on the space sector and an E1 topology on the time axis. It
is not first-countable and, consequently, cannot come from any metric. In
their everyday practice, physicists adopt an ambiguous behaviour and use
the balls of E4 whenever they talk of continuity and/or convergence.
1.1.15 Given the subset C of S, its complement is the set C 0 = {p S
such that p
/ C}. The subset C is a closed set in the topological space
(S, T ) if C 0 is an open set of T . Thus, the complement of an open set is (by
definition) closed. It follows that and S are closed (and open!) sets in all
topological spaces.
1.1.16 Closedness is a relative concept: a subset C of a topological subspace Y of S can be closed in the induced topology even if open in S; for
instance, Y itself will be closed (and open) in the induced topology, even if
Y is an open set of S.
Retain that closed, just as open, depends on the chosen topology. A set
which is open in a topology may be closed in another.
1.1.17 A connected space is a topological space in which no proper
subset is simultaneously open and closed.
In this case S cannot be decomposed into the union of two disjoint open sets.
One should not confuse this concept with path-connectedness, to be defined
4
Zeeman 1964, 1967; later references may be traced from Fullwood 1992.
14
later ( 1.3.15) and which, intuitively, means that one can walk continuously
between any two points of the space on a path entirely contained in it. Pathconnectedness implies connectedness, but not vice-versa. Clearly the line E1
is connected, but the line-minus-zero space E1 {0} (another notation:
E1 \{0}) is not. The finite space of 1.1.3 is connected.
1.1.18 The discrete topology : set S and all its subsets are taken as
open sets. The set of all subsets of a set S is called its power set, denoted
P (S), so that we are taking the topological space (S, P (S)). This does yield
a topological space. For each point p, {p} is open. All open sets are also
closed and so we have extreme unconnectedness. Lest the reader think this
example to be physically irrelevant, we remark that the topology induced on
the light cone by Zeemans topology for spacetime ( 1.1.14) is precisely of
this type. Time is usually supposed to be a parameter running in E1 and
a trajectory on some space S is a mapping attributing a point of S to each
instant in E1 . It will be seen later (section 1.3) that no function from
E1 to a discrete space may be continuous. A denizen of the light cone, like
the photon, would not travel continuously through spacetime but bound
from point to point. The discrete topology is, of course, the finest possible
topology on any space. Curiously enough, it can be obtained from a metric,
the so-called discrete metric: d(p, q) = 1 if p 6= q, and d(p, q) = 0 if p = q.
The indiscrete (or trivial ) topology is T = {, S}. It is the weakest possible
topology on any space andbeing not first-countablethe simplest example
of topology which cannot be given by a metric. By the way, this is an
illustration of the complete independence of topology from metrics: a nonmetric topology may have finer topologies which are metric, and a metric
topology can have finer non-metric topologies. And a non-metric topology
may have weaker topologies which are metric, and a metric topology can
have weaker non-metric topologies.
1.1.19 Topological product Given two topological spaces A and B, their
topological product (or cartesian product) AB is the set of pairs (p, q) with
p A and q B, and a topology for which a basis is given by all the pairs of
type U V , U being a member of a basis in A and V a member of a basis in
B. Thus, the cartesian product of two topological spaces is their cartesian set
product ( Math.1.11) endowed with a product topology. The usual torus
imbedded in E3 , denoted T2 , is the cartesian product of two 1-dimensional
spheres (or circles) S 1 . The n-torus Tn is the product of S 1 by itself n times.
1.1.20 We have clearly separated topology from metric and found examples of non-metric topologies, but it remains true that a metric does define
1.2. KINDS OF TEXTURE
15
a topology. A reverse point of view comes from asking the following question: are all the conditions imposed in the definition of a distance function
necessary to lead to a topology? The answer is no. Much less is needed. A
prametric suffices. On a set S, a prametric is a mapping
: S S R+ such that (p, p) = 0 for all p S.
1.1.21 The consideration of spatial relationships requires a particular way
of dividing a space into parts. We have chosen, amongst all the subsets of S,
particular families satisfying well chosen conditions to define topologies. A
family of subsets of S is a topology if it includes S itself, the empty set , all
unions of subsets and all intersections of a finite number of them. A topology
is that simplest, minimal structure allowing for precise non-trivial notions of
convergence and continuity. Other kinds of families of subsets are necessary
for other purposes. For instance, the detailed study of convergence in a nonmetric topological space S requires cuttings of S not including the empty
set, called filters. And, in order to introduce measures and define integration
on S, still another kind of decomposition is essential: a -algebra. In other
to make topology and integration compatible, a particular -algebra must be
defined on S, the Borel -algebra. A sketchy presentation of these questions
is given in Mathematical Topic 3.
1.2
KINDS OF TEXTURE
We have seen that, once a topology is provided, the set point acquires a kind
of elementary texture, which can be very tight (as in the indiscrete topology),
very loose (as in the discrete topology), or intermediate. We shall see now
that there are actually optimum decompositions of spaces. The best behaved
spaces have not too many open sets: they are compact. Nor too few: they
are of Hausdorff type.
There are many ways of probing the topological makeup of a space. We
shall later examine two external approaches: one of them (homology) tries
to decompose the space into its building bricks by relating it to the decomposition of euclidean space into triangles, tetrahedra and the like. The other
(homotopy) examines loops (of 1 or more dimensions) in the space and their
continuous deformations. Both methods use relationships with other spaces
and have the advantage of providing numbers (topological numbers) to
characterize the space topology.
For the time being, we shall study two internal ways of probing a space
(S, T ). One considers subsets of S, the other subsets of T. The first considers
16
samples of isolated points, or sequences, and gives a complete characterization of the topology. The second consists of testing the texture by rarefying
the family of subsets and trying to cover the space with a smaller number of
them. It reveals important qualitative traits. We shall start by introducing
some concepts which will presently come in handy.
1.2.1 Consider a space (S, T ). Given an arbitrary set U S, not necessarily belonging to T , in general there will be some closed sets C containing
U . The intersection C of all closed sets containing U is the closure of
U , denoted U . An equivalent, more intuitive definition is U = {p such that
every neighbourhood of p has nonvanishing intersection with U }. The bestknown example is that of an open interval (a, b) in E1 , whose closure is the
closed interval [a, b].
The closure of a closed set V is V itself, and V being the closure of itself implies
that V is closed.
Given an arbitrary set W S, not necessarily belonging to T , its interior ,
denoted int W or W 0 , is the largest open subset of W . Given all the open sets
O contained in W , then
W 0 = O .
W 0 is the set of points of S for which W is an open neighbourhood. The boundary
b(U ) of a set U is the complement of its interior in its closure,
U0 = U
\U 0 .
b(U ) = U
\b(U ). If U is an open set of T , then U 0 = U and
It is also true that U 0 = U
\U . If U is a closed set, then U
= U and b(U ) = U \U 0 . These definitions
b(U ) = U
correspond to the intuitive view of the boundary as the skin of the set. From
this point of view, a closed set includes its own skin. The sphere S 2 , imbedded in
E3 , is its own interior and closure and consequently has no boundary. A set has
empty boundary when it is both open and closed. This allows a rephrasing of the
definition of connectedness: a space S is connected if, except for and S itself, it
has no subset whose boundary is empty.
Let again S be a topological space and U a subset. A point p U is an isolated point
of U if it has a neighbourhood that contains no other point of U . A point p of S is a
limit point of U if each neighbourhood of p contains at least one point of U distinct of
p. The set of all the limit points of U is called the derived set of U , written D(U ). A
= U D(U ): we may obtain the closure of a set by adding to it all
theorem says that U
its limiting points. U is closed iff it already contains them all, U D(U ). When every
neighbourhood of p contains infinite points of U , p is an accumulation point of U (when
such infinite points are not countable, p is a condensation point ). Though we shall not be
using all these notions in what follows, they appear frequently in the literature and give a
17
taste of the wealth and complexity of the theory coming from the three simple axioms of
1.1.2.
1.2.2 Let U and V be two subsets of a topological space S. The subset

U is said to be dense in V if U V . The same U will be everywhere dense
if U = S. A famous example is the set Q of rational numbers, which is
dense in the real line E1 of real numbers. This can be generalized: the set
of n-uples (p1 , p2 , . . . , pn ) of rational numbers is dense in En . This is a
fortunate property indeed. We (and digital computers alike) work ultimately
only with rational (actually, integer) numbers (a terminated decimal is, of
course, always a rational number). The property says that we can do it even
to work with real numbers, as rational numbers lie arbitrarily close to them.
A set U is a nowhere dense subset when the interior of its closure is empty:
U 0 = . An equivalent definition is that the complement to its closure is
everywhere dense in S. The boundary of any open set in S is nowhere dense.
The space E1 , seen as subset, is nowhere dense in E2 .
1.2.3 The above denseness of a countable subset in the line extends to a
whole class of spaces. S is said to be a separable space if it has a countable
everywhere dense subset. This separability (a name kept for historical
reasons) by denseness is not to be confused with the other concepts going
under the same name (first-separability, second-separability, etc see below,
1.2.14 on), which constitute another hierarchy of topological spaces. The
present concept is specially important for dimension theory (section 4.1.2)
and for the study of infinite-dimensional spaces. Intuitively, it means that
S has a countable set P of points such that each open set contains at least
one point of P . In metric spaces, this separability is equivalent to secondcountability.
1.2.4 The Cantor set A remarkable example of closed set is the Cantor
ternary set.5 Take the closed interval I = [0, 1] in E1 with the induced
topology and delete its middle third, the open interval (1/3, 2/3), obtaining
the closed interval E1 = [0, 1/3] [2/3, 1]. Next delete from E1 the two
middle thirds (1/9, 2/9) and (7/9, 8/9). The remaining closed space E2 is
composed of four closed intervals. Then delete the next four middle thirds to
get another closed set E3 . And so on to get sets En for any n. Call I = E0 .
The Cantor set is the intersection
E=
En .
n=0
5
See Kolmogorov & Fomin 1970 and/or Christie 1976.
18
E is closed because it is the complement of a union of open sets. Its interior

is empty, so that it is nowhere dense. This emptiness is coherent with the
following: at the j-th stage of the building process, we delete 2j1 intervals,
each of length (1/3j ), so that the sum of the deleted intervals is 1. On the
other hand, it is possible to show that a one-to-one correspondence exists
between E and I, so that this almost empty set has the power of the
continuum. The dimension of E is discussed in 4.1.5.
1.2.5 Sequences are countable subsets {pn } of a topological space S. A
sequence {pn } is said to converge to a point p S (we write pn p when
n ) if any open set U containing p contains also all the points pn for n
large enough.
Clearly, if W and T are topologies on S, and W is weaker than T , every sequence which is convergent in T is convergent in W ; but a sequence
may converge in W without converging in T . Convergence in the stronger
topology forces convergence in the weaker. Whence, by the way, come these
designations.
We may define the q-th tail tq of the sequence {pn } as the set of all its points pn for
n q, and say that the sequence converge to p if any open set U containing p traps some
of its tails.
It can be shown that, on first-countable spaces, each point of the derivative set D(U )
is the limit of some sequence in U , for arbitrary U .
Recall that we can define real numbers as the limit points of sequences of
rational numbers. This is possible because the subset of rational numbers Q is
everywhere dense in the set R of the real numbers with the euclidean topology
(which turns R into E1 ). The set Q has derivative D(Q) = R and interior Q0 = .
Its closure is the same as that of its complement, the set J = R\Q of irrational
numbers: it is R itself. As said in 1.1.9, both Q and J are topological subspaces
of R.
On a general topological space, it may happen that a sequence converges

to more than one point. Convergence is of special importance in metric
spaces, which are always first-countable. For this reason, metric topologies
are frequently defined in terms of sequences. On metric spaces, it is usual
to introduce Cauchy sequences (or fundamental sequences ) as those {pn }
for which, given any tolerance > 0, an integer k exists such that, for
n, m > k, d(pn , pm ) < . Every convergent sequence is a Cauchy sequence,
but not vice-versa. If every Cauchy sequence is convergent, the metric space
is said to be a complete space. If we add to a space the limits of all its Cauchy
sequences, we obtain its completion . Euclidean spaces are complete. The
space J of irrational numbers with the euclidean metric induced from E1 is
incomplete. On general topological spaces the notion of proximity of two
19
points, clearly defined on metric spaces, becomes rather loose. All we can
say is that the points of a convergent sequence get progressively closer to its
limit, when this point is unique.
1.2.6 Roughly speaking, linear spaces, or vector spaces, are spaces allowing for addition and rescaling of their members. We leave the definitions
and the more algebraic aspects to Math.1, the details to Math.4, and concentrate in some of their topological possibilities. What imports here is that
a linear space over the set of complex numbers C may have a norm, which
is a distance function and defines consequently a certain topology called the
norm topology. Once endowed with a norm, a vector space V is a metric
topological space. For instance, a norm may come from an inner product, a
mapping from the cartesian set product V V into C,
V V C,
(v, u) < v, u >
(1.2)
with suitable properties. In this case the number

||v|| =
< v, v >
will be the norm of v induced by the inner product. This is a special norm,
as norms may be defined independently of inner products. Actually, one
must impose certain compatibility conditions between the topological and
the linear structures (see Math.4).
1.2.7 Hilbert space6 Everybody knows Hilbert spaces from (at least)
Quantum Mechanics courses. They are introduced there as spaces of wavefunctions, on which it is defined a scalar product and a consequent norm.
There are basic wavefunctions, in terms of which any other may be expanded.
This means that the set of functions belonging to the basis is dense in the
whole space. The scalar product is an inner product and defines a topology. In Physics textbooks two kinds of such spaces appear, according to
whether the wavefunctions represent bound states, with a discrete spectrum,
or scattering states. In the first case the basis is formed by a discrete set
of functions, normalized to the Kronecker delta. In the second, the basis is
formed by a continuum set of functions, normalized to the Dirac delta. The
latter are sometimes called Dirac spaces.
Formally, a Hilbert space is an inner product space which is complete
under the inner product norm topology. Again we leave the details to Math.4,
6
Halmos 1957.
20
and only retain here some special characteristics. It was originally introduced
as an infinite space H endowed with a infinite but discrete basis {vi }iN ,
formed by a countably infinite orthogonal family of vectors. This family is
dense in H and makes of H a separable space.
member of the space
P Each
i
can be written in terms of the basis: X = i=1 X vi . The space L2 of all
absolutely square integrable functions on the interval (a, b) R,
Z b
2
L = {f on [a, b] with
|f (x)|2 dx < },
a
is a separable Hilbert space. Historical evolution imposed the consideration

of non-separable Hilbert spaces. These would come out if, in the definition
given above, instead of {vi }iN we had {v }R : the family is not indexed
by a natural number, but by a number belonging to the continuum. This
definition would accommodate Dirac spaces. The energy eigenvalues, for
the discrete or the continuum spectra, are precisely the indexes labeling the
family elements, the wavefunctions or kets. Thus, bound states belong to
separable Hilbert spaces while scattering states require non-separable Hilbert
spaces. There are
new problems in this continuum-label case:
Pnevertheless
the summations i=1 used in the expansions become integrals. As said in

1.1.21, additional structures are necessary in order to define integration (a
-algebra and a measure, see Math.3.
It is possible to show that En is the cartesian topological product of E1
taken n times, and so that En+m = En Em . The separable Hilbert space
is isomorphic to E , that is, the product of E1 an infinite (but countable)
number of times. The separable Hilbert space is consequently the natural
generalization of euclidean spaces to infinite dimension. This intuitive result
is actually fairly non-trivial and has been demonstrated not long ago.
1.2.8 Infinite dimensional spaces, specially those endowed with a linear
structure, are a privileged arena for topological subtlety. Hilbert spaces are
particular cases of normed vector spaces, particularly of Banach spaces, on
which a little more is said in Math.4. An internal product like that above
does define a norm, but there are norms which are not induced by an internal
product. A Banach space is a normed vector space which is complete under
the norm topology.
1.2.9 Compact spaces The idea of finite extension is given a precise
formulation by the concept of compactness. The simplest example of a space
confined within limits is the closed interval I = [0, 1] included in E1 , but
its finiteness may seem at first sight a relative notion: it is limited within
E1 , by which it is contained. The same happens with some closed surfaces
21
in our ambient space E3 , such as the sphere, the ellipsoid and the torus:
they are contained in finite portions of E3 , while the plane, the hyperboloid
and the paraboloid are not. It is possible, however, to give an intrinsic
characterization of finite extension, dependent only on the internal properties
of the space itself and not on any knowledge of larger spaces containing it.
We may guess from the above examples that spaces whose extensions are
limited have a lesser number of open sets than those which are not. In
fact, in order to get an intrinsic definition of finite extension, it is necessary
to restrict the number of open sets in a certain way, imposing a limit to the
divisibility of space. And, to arrive at that restriction, the preliminary notion
of covering is necessary.
1.2.10 Suppose a topological space S and a collection C = {U } of open
sets such that S is their union, S = U . The collection C is called an
open covering of S. The interval I has a well known property, which is the
Heine-Borel lemma: with the topology induced by E1 , every covering of I has
a finite subcovering. An analogous property holds in any euclidean space: a
subset is bounded and closed iff any covering has a finite subcovering. The
general definition of compactness is thereby inspired.
1.2.11 Compactness A topological space S is a compact space if each
covering of S contains a finite subcollection of open sets which is also a
covering.
Cases in point are the historical forerunners, the closed balls in euclidean
spaces, the spheres S n and, as expected, all the bounded surfaces in E3 .
Spaces with a finite number of points (as that in 1.1.3) are automatically
compact. In Physics, compactness is usually introduced through coordinates
with ranges in suitably closed or half-closed intervals. It is, nevertheless, a
purely topological concept, quite independent of the very existence of coordinates. As we shall see presently, not every kind of space accepts coordinates.
And most of those which do accept require, in order to be completely described, the use of many distinct coordinate systems. It would not be possible
to characterize the finiteness of a general space by this method.
On a compact space, every sequence contains a convergent subsequence,
a property which is equivalent to the given definition and is sometimes used
instead: in terms of sequences,
a space is compact if, from any sequence of its points,
one may extract a convergent subsequence.
1.2.12 Compact spaces are mathematically simpler to handle than noncompact spaces. Many of the topological characteristics physicists became
22
recently interested in (such as the existence of integer topological numbers)

only hold for them. In Physics, we frequently start working with a compact
space with a boundary (think of quantization in a box), solve the problem
and then push the bounds to infinity. This is quite inequivalent to starting
with a non-compact space (recall that going from Fourier series to Fourier
integrals requires some extra smoothing assumptions). Or, alternatively,
by choosing periodic boundary conditions we somehow manage to make the
boundary to vanish. We shall come to this later. More recently, it has become
fashionable to compactify non-compact spaces. For example: field theory
supposes that all information is contained in the fields, which represent the
degrees of freedom. When we suppose that all observable fields (and their
derivatives) go to zero at infinity of (say) an euclidean space, we identify all
points at infinity into one only point. In this way, by imposing a suitable
behaviour at infinity, a field defined on the euclidean space E4 becomes a
field on the sphere S 4 . This procedure of compactification is important in
the study of instantons7 and is a generalization of the well known method by
which one passes from the complex plane to the Riemann sphere. However,
it is not always possible.
1.2.13 A topological space is locally compact if each one of its points
has a neighbourhood with compact closure. Every compact space is locally
compact, but not the other way round: En is not compact but is locally
compact, as any open ball has a compact closure. The compactification
above alluded to is possible only for a locally compact space and corresponds
to adjoining a single point to it (see 1.3.20).8
A subset U of the topological space S is relatively compact if its closure is
compact. Thus, a space is locally compact if every point has a relatively compact
neighbourhood. Locally compact spaces are of particular interest in the theory of
integration, when nothing changes by adding a set of zero measure. On topological
groups (section 1.4.2), local compactness plus separability are sufficient conditions
for the existence of a left- and a right-invariant Haar measure (see Math.6.9),
which makes integration on the group possible. Such measures, which are unique
up to real positive factors, are essential to the theory of group representations and
general Fourier analysis. Unlike finite-dimensional euclidean spaces, Hilbert spaces
are not locally compact. They are infinite-dimensional, and there are fundamental
differences between finite-dimensional and infinite-dimensional spaces. One of the
main distinctive properties comes out precisely here:
Riesz theorem: a normed vector space is locally compact
if and only if its dimension is finite.
7
8
Coleman 1977; Atiyah et al. 1978; Atiyah 1979.

For details, see Simmons 1963.
23
1.2.14 Separability Compactness imposes, as announced, a limitation

on the number of open sets: a space which is too fine-grained will find a way
to violate its requirements. As we consider finer and finer topologies, it becomes easier and easier to have a covering without a finite subcovering. Thus,
compactness somehow limits the number of open sets. On the other hand,
we must have a minimum number of open sets, as we are always supposed to
be able to distinguish between points in spaces of physical interest: between
neighbouring states in a phase space, between close events in spacetime, etc.
Such values belong to open sets ( 1.1.2). Can we distinguish points by using
only the notions above introduced? It seems that the more we add open
sets to a given space, the easier it will be to separate (or distinguish) its
points. We may say things like p is distinct from q because p belongs to
the neighbourhood U while q does not. Points without even this property
are practically indistinguishable: p = Tweedledee, q = Tweedledum. But we
might be able to say still better, p is quite distinct from q because p belongs
to the neighbourhood U , q belongs to the neighbourhood V , and U and V
are disjoint. To make these ideas precise and operational is an intricate
mathematical problem coming under the general name of separability . We
shall not discuss the question in any detail, confining ourselves to a strict
minimum. The important fact is that separability is not an automatic property of all spaces and the possibility of distinguishing between close points
depends on the chosen topology. There are in reality several different kinds
of possible separability and which one (if any) is present in a space of physical
significance is once again a matter to be decided by experiment. Technically,
the two phrases quoted above correspond respectively to first-separability
and second-separability. A space is said to be first-separable when, given
any two points, each one will have some neighbourhood not containing the
other and vice-versa. The finite space of 1.1.3 is not first-separable. Notice
that in first-separable spaces the involved neighbourhoods are not necessarily
disjoint. If we require the existence of disjoint neighbourhoods for every two
points, we have second-separability, a property more commonly named after
Hausdorff.
1.2.15 Hausdorff character A topological space S is said to be a Hausdorff space if every two distinct points p, q S have disjoint neighbourhoods.
There are consequently U 3 p and V 3 q such that U V = . This
property is so important that spaces of this kind are simply called separated
by many people (the term separable being then reserved to the separability
by denseness of 1.2.3). We have already met a counter-example in the trivial
topology ( 1.1.18). Another non-Hausdorff space is given by two copies of
24
E1 , X and Z (Figure 1.5), of which we identify all (and only!) the points
which are strictly negative: pX pZ iff p < 0. The points pX = 0 and pZ = 0
are distinct, pX lying in the region of X not identified with Z and pZ lying
in Z. But they have no disjoint neighbourhoods.
Figure 1.5: An example of non-Hausdorff space.
The space has a Y aspect in this case, but not all non-Hausdorff spaces
exhibit such a bifurcation. All Hausdorff spaces are of course necessarily firstseparable, but they go much further, allowing to discern points in an way ideal
for physicists, after the discussion of 1.1.2. Actually, each point is a closed
set. The Hausdorff property is also highly praised by analysts: it ensures
the uniqueness of a converging sequence limit point. And it is fortunate that
most spaces appearing in Physics are Hausdorff, as only on such spaces are the
solutions of differential equations (with fixed initial conditions) assured to be
unique the Hausdorff character is included in the hypotheses necessary to
prove the unicity of solution theorem. On non-Hausdorff spaces, solutions are
only locally unique.9 It would seem that physicists should not worry about
possible violations of so desirable condition, but non-Hausdorff spaces turn
up in some regions of spacetime for certain solutions of Einsteins equations,10
giving rise to causal anomalies.11 Although the Hausdorff character is also
necessary to the distinction of events in spacetime,12 Penrose has speculated
on its possible violation.13
An open set can be the union of disjoint point sets. Take the interval
I = [0, 1]. Choose a basis containing I, and all the sets obtained by omitting from I at most a countable number of points. A perfect though rather
9
Arnold 1973.
Hajicek 1971.
11
Hawking & Ellis 1973.
12
Geroch & Horowitz in Hawking & Israel 1979. An interesting article on the topology
of the Universe.
13
Penrose in Hawking & Israel 1979, mainly in those pages (591-596) dedicated to
psychological time.
10
25
pathological topological space results. It is clearly second-countable. Given

two points p and q, there is always a neighbourhood of p not containing q
and vice-versa. It is, consequently, also first-separable. The trouble is that
two such neigbourhoods are not always disjoint: the space is not a Hausdorff space. Topological spaces may have very distinct properties concerning
countability and separability and are accordingly classified. We shall avoid
such a analysis of the systematic zoology of topological spaces and only talk
loosely about some of these properties, sending the more interested reader to
the specialized bibliography.14
A Hausdorff space which is a compact (adjective) space is called a compact
(noun).
A closed subspace of a compact space is compact. But a compact subspace
is necessarily closed only if the space is a Hausdorff space.
1.2.16 A stronger condition is the following (Figure 1.6): S is normal if it
is first-countable and every two of its closed disjoint sets have disjoint open
neighbourhoods including them. Every normal space is Hausdorff but not
vice-versa.15 Every metric space is normal and, so, Hausdorff, but there are
normal spaces whose topology is not metrizable. The upper-half plane E2+
of Fig.(1.4) is not normal and consequently non-metric. Putting together
countability and separability may lead to many interesting results. Let us
here only state Urysohns theorem: a topological space endowed with a countable basis (that is, second-countable) is metric iff it is normal. We are not
going to use much of these last considerations in the following. Our aim has
been only to give a slight idea of the strict conditions a topology must satisfy
in order to be generated by a metric. In order to prove that a topology T is
non-metric, it suffices to show, for instance, that it is not normal.
1.2.17 Bad E1 , or Sorgenfrey line: the real line R1 with its proper
(that is, non-vanishing) closed intervals does not constitute a topological
space because the second defining property of a topology goes wrong. However, the half-open intervals of type [p, q) on the real line do constitute a
basis for a topology. The resulting space is unconnected (the complement of
an interval of type [) is of type )[ , which can be obtained as a union of
an infinite number of half-open intervals) and not second-countable (because
in order to cover ), for example, one needs a number of [)s which is
an infinity with the power of the continuum). It is, however, first-countable:
14
For instance, the book of Kolmogorov & Fomin, 1977, chap.II. A general resume with
many (counter) examples is Steen & Seebach 1970.
15
For an example of Hausdorff but not normal space, see Kolmogorov & Fomin 1970,
p. 86.
26
Figure 1.6: First-separable, second-separable and normal spaces: (left) first separable
every two points have exclusive neighbourhoods; (center) Hausdorff every two points
have disjoint neighbourhoods; (right) normal disjoint closed sets are included in disjoint
open sets.
given a point p, amongst the intervals of type [p, r) with r rational, there
will be some one included in any U containing p. The Sorgenfrey topology
is finer than the usual euclidean line topology, though it remains separable
(by denseness). This favorite pet of topologists is non-metrizable.
1.2.18 Consider a covering {U } of S. It is a locally finite covering if each
point p has a neighbourhood U 3 p such that U intersects only a finite number of the U . A covering {Vi } is a refinement of {U } if, for every Vi , there
is a U such that U Vi . A space is paracompact if it is Hausdorff and all
its coverings have local finite refinements. Notice: finite subcoverings lead to
compactness and finite refinements (plus Hausdorff) to paracompactness. A
connected Hausdorff space is paracompact if it is second-countable. Figure
1.7 is a scheme of the separability hierarchy. Every metric space is paracompact. Every paracompact space is normal. Paracompactness is a condition
of importance for integration, as it is sufficient for attributing a partition of
unity (see Math.3.5) to any locally finite covering. It is also necessary to
the existence of linear connections on the space.16 Paracompact spaces are
consequently essential to General Relativity, in which these connections are
16
See Hawking & Ellis 1973.
1.3. FUNCTIONS
27
represented by the Christoffel symbols. The Lorentz metric on a Hausdorff

space implies its paracompactness.17
Figure 1.7: A second hierarchy: all metric spaces are normal.
1.3
FUNCTIONS
Continuity of functions is the central notion of general topology. Functions,

furthermore, allow us to compare the properties of different spaces.
1.3.1 The word function is used here unless otherwise explicitly
stated in the modern sense of monodromous function, with a unique value
in the target space for each point of its domain. Perhaps the simplest examples are the permutations of the elements of a finite set ( Math.2.4).
1.3.2 Mathematics deals basically with sets and functions. The sequences
previously introduced as countable subsets {pn } of a topological space S are
better defined as functions p : N S , n pn , from the set of natural
numbers N into S. Only to quote a famous fundamental case, two sets have
the same power if there exists some bijective function between them. In
this sense, set theory uses functions in counting. We have said in 1.1.18
that the power set P (S) of a set S is the set of all its subsets. For S finite
with n points, P (S) will have 2n (> n) elements and for this reason P (S) is
sometimes indicated by 2S . P (S) is larger than S, as there are surjective
17
Geroch 1968.
28
functions, but no injective functions, from P (S) to S. This notion of relative

size of a set was shown by Cantor to keep holding for infinite sets. P (S)
is always larger than S) and led to his infinite hierarchy of infinite numbers.
1.3.3 Let f : A B be a function between two topological spaces. The
inverse image of a subset X of B by f is
f <1> (X) = { a A such that f (a) X }.
1.3.4 The function f is continuous if the inverse images of all the open
sets of the target space are open sets of the domain space.
This is the notion of continuity on general topological spaces. Here we
have a good opportunity to illustrate the cautionary remark made in 1.1.12.
At first sight, the above definition is of that skew-minded type alluded to. We
could try to define a continuous function directly, as a function mapping
open sets into open sets, but the following example shows that using the
inverse as above is essential. Consider the function f : E1 E1 given by
(
x
for x 0
f (x) =
x + 1 for x > 0.
which is shown in Figure 1.8. In reality, the target space is not E1 . The
Figure 1.8: A standard discontinuous function.

function is actually
f : E1 (, 0] (1, ),
1.3. FUNCTIONS
29
the latter with the induced topology. It has an obvious discontinuity at

x = 0. It is indeed discontinuous by the definition given above, but it is easy
to check that it would be continuous if the direct definition was used. With
the induced topology, (, 0] is open in the image space. The function f
will take open sets of E1 into open sets of the union, but its inverse f <1> will
take (, 0] into the same interval in E1 , where it is not open. By the way,
this also illustrates the wisdom of carefully and clearly stating the definition
and value domains of functions, as mathematicians pedantly do!
1.3.5 One should specify the topology whenever one speaks of a continuous
function. Any function defined on a space with the discrete topology is automatically continuous. On a space with the indiscrete topology, no function
of interest can be continuous. That is what we meant when we said that no
useful definition of the continuity concept was possible in these two cases.
1.3.6 In Zeemans topology for spacetime, every function on the light cone
is continuous, as any function will be continuous if the domain space is discrete. With these repeated references to Zeemans work we want to stress
the following point: people talk frequently of continuous wavefunctions and
fields on spacetime but nobody really knows the meaning of that, as the
real topology of spacetime is unknown. Curiously enough, much more study
has been dedicated to the topological properties of functional spaces (such
as the spaces of quantum fields) than to those of the most important space
of all Physics. By the way, some other topologies have been proposed for
spacetime,18 whose properties would influence those of the fields. Actually,
these tentatives proceed in the inverse way: they look for a topology in which
better known quantities (such as classical paths on which integrations are to
be performed in Feynmans formalism, or Greens functions) are continuous.
1.3.7 If a function is continuous on a space with topology T , it will be
continuous in any refinement of T .
Zeemans topology is a refinement of that of E4 , but a function which
is continuous in his spacetime could appear as discontinuous upon euclideanization. In this way euclideanizations, procedures by which quantities of physical interest (such as Greens functions and S matrix elements),
defined on spacetime, are transformed by some kind of analytical continuation into analogous quantities on E4 , may provide constraints on spacetime
possible topologies.
18
Hawking, King & McCarthy 1976; Gobel 1976 a,b; Malament 1977.
30
1.3.8 An isometry is a distance-preserving mapping between metric spaces:

f : X Y such that dY [f (p), f (q)] = dX (p, q).
If a function g is such that target and domain spaces coincide, then g is
an automorphism . Particular cases of continuous automorphisms are the
inner isometries of metric spaces, when X Y . In this case, the old name
motions is physically far more telling.
1.3.9 Consider En , taking into account its usual structure of vector space
(Math.1). This means, of course, that we know how to add two points of En
to get a third one, as well as to multiply them by external real scalars. We
denote points in En by p = (p1 , p2 , . . . , pn ), a = (a1 , a2 , . . . , an ), etc. Then,
translations t : p p + a, homothecies h : p p with 6= 0, and
the invertible linear automorphisms are continuous. Consider in particular
real functions f : E1 E1 . Then f (x) = x + 1 is surjective and injective
(consequently, bijective). On the other hand, f (x) = x2 is neither, but becomes bijective if we restrict both sets to the set R+ of positive real numbers.
The function exp : [0, 2) S 1 , f () = exp(i), is bijective. With the
topologies induced by those of E1 and E2 respectively, it is also continuous.
A beautiful result: every continuous mapping of a topological space into
itself has a fixed point.
1.3.10 The image of a connected space by a continuous function is connected. The image of a compact space by a continuous function is compact.
Therefore, continuous functions preserve topological properties. But only
in one way: the inverse images of connected and/or compact domains are
not necessarily connected and/or compact. In order to establish a complete
equivalence between topological spaces we need functions preserving topological properties both ways.
A bijective function f : A B will be a homeomorphism between the
topological spaces A and B if it is continuous and has a continuous inverse.
Thus, it takes open sets into open sets and its inverse does the same.
Two spaces are homeomorphic when there exists a homeomorphism between them. Notice that if f : A B and g : B C are continuous,
then the composition (f g) : A C is continuous. If f and g are homeomorphisms, so is their composition. By the very definition, the inverse of
a homeomorphism is a homeomorphism. A homeomorphism is an equivalence relation: it establishes a complete topological equivalence between two
topological spaces, as it preserves all the purely topological properties. We
1.3. FUNCTIONS
31
could in reality define a topology as an equivalence class under homeomorphisms. And the ultimate (perhaps too pretentious?) goal of Topology
as a discipline is the classification of all topological spaces, of course up to
such equivalences. Intuitively, A is homeomorphic to B means that A can
be deformed, without being torn or cut, to look just like B. Under a homeomorphism, images and pre-images of open sets are open, and images and
pre-images of closed sets are closed. A sphere S 2 can be stretched to become
an ellipsoid or an egg-shaped surface or even a tetrahedron. Such surfaces
are indistinguishable from a purely topological point of view. They are the
same topological space. The concept of homeomorphism gives in this way
a precise meaning to those rather loose notions of suave deformations we
have been talking about in 1.0.4. Of course, there is no homeomorphism
between either the sphere, the ellipsoid, or the tetrahedron and (say) a torus
T 2 , which has quite different topological properties.
From the point of view of sequences: a homeomorphism is a mapping
h : A B such that, if {pn } is a sequence in A converging to a point
p, then the sequence {h(pn )} in B converges to h(p); and vice-versa: if
the sequence {qn } in B converges to a point q, then the sequence given by
{h1 (qn )} in A converges to {h1 (q)}.
A condensation is a mapping f : A B which is one-to-one [x 6= y
f (x) 6= f (y)] and onto [f (A) = B]; a homeomorphism is a condensation
whose inverse is also a condensation.
The study of homeomorphisms between spaces can lead to some surprises.
It was found that a complete metric space [ 1.2.5] can be homeomorphic to
an incomplete space. Consequently, the completeness of metric spaces is
not really a topological characteristic. Another result: the space of rational
numbers with the usual topology is homeomorphic to any metric countable
space without isolated points.
The main objective of Zeemans cited papers on spacetime was to obtain a
topology whose automorphic homeomorphisms preserve the causal structure
and constitute a group including the Poincare and/or the conformal group.
1.3.11 Take again the euclidean vector space En . Any isometry will be
a homeomorphism, in particular any translation. Also homothecies with
reason 6= 0 are homeomorphisms. From these two properties it follows
that any two open balls are homeomorphic to each other, and any open ball
is homeomorphic to the whole space. As a hint of the fundamental role which
euclidean spaces will come to play, suppose a space S has some open set U
which is by itself homeomorphic to an open set (a ball) in some En : there is
a homeomorphic mapping f : U ball, f (p U ) = x = (x1 , x2 , . . . , xn ).
Such a homeomorphism is a local homeomorphism. Because the image space
32
is En , f is called a coordinate function and the values xk are coordinates of

p. Coordinates will be formally introduced in section 4.2.
We are now in condition to explain better a point raised in 1.2.7.
The separable Hilbert space is actually homeomorphic to E . Once this
is granted, the same topology is given by another metric,
P
|vk uk |
d(v, u) = k 2k [1+|v
k uk |]
The metric space so obtained is called a Frechet space19 and it is important
because a theorem says that any separable metric space is homeomorphic to
some subspace of a Frechet space.
1.3.12 A counter-example: take S1 as the unit circle on the plane,
S 1 = {(x, y) R2 such that x2 + y 2 = 1},
with the topology induced by the usual topology of open balls (here, open
disks) of E2 . The open sets will be the open arcs on the circle (Figure 9).
Take then the set consisting of the points in the semi-open interval [0, 2)
of the real line, with the topology induced by the E1 topology. The open
sets will be of two kinds: all the open intervals of E1 strictly included in
[0, 2), and those intervals of type [0, ), with 2, which would be
semi-open in E1 . The function given by f () = exp(i), or x = cos and
y = sin , is bijective and continuous ( 1.3.9). But f takes open sets of
the type [ ) into semi-open arcs, which are not open in the topology defined
on S 1 . Consequently, its inverse is not continuous. The function f is not a
homeomorphism. In reality, none of such exist, as S 1 is not homeomorphic
to the interval [0, 2) with the induced topology. It is possible, however, to
define on [0, 2) another topology which makes it homeomorphic to S 1 (see
1.4.3).
1.3.13 Of fundamental importance is the following kind of function:
a curve on a topological space S is a function taking the compact
interval I = [0,1] into S. If : I S is a continuous
function, then is a continuous curve, or path.
Notice the semantic shift from the usual meaning of curve, which would
rather refer to the set of values of the mapping and not the mapping itself. A
point on the curve will in general be represented by (t), where the parameter
19
The name Frechet space is also used with another, more fundamental and quite
distinct meaning. See, for instance, Sierpi
nski 1956.
1.3. FUNCTIONS
33
Figure 1.9: The function f : [0, 2) S 1 included in E2 , given by f () = (cos , sin ),

is bijective and continuous; but it takes some open sets into sets which are not open;
consequently its inverse f <1> is not continuous and f is not a homeomorphism.
t I, so that : t (t). Notice also that we reserve the word path to a

continuous curve.
When (0) = (1), is a closed curve, or a loop, which can be alternatively defined as a function from the circle S 1 into S.
1.3.14 Function spaces Take two spaces X and Y and consider the set
of functions between them, {f : X Y }. This set is usually indicated by
Y X . Suppose X compact and Y metric with a distance function d. We may
define the distance between two functions f1 , f2 as
dist (f1 , f2 ) = least upper bound {d[f1 (p), f2 (p)]} ,
for all p X. It can be shown that this distance turns the set of functions into
a metric (hence topological) space, whose topology depends on the topologies
of X and Y but not in reality on the distance function d (that is: any other
distance function leading to the same topology for Y will give the same
topology for Y X ).
The rather strict conditions above can be softened in the following elaborate
way: take any two topologies on X and Y ; call C the set of compact subsets of
X and O the set of open subsets of Y ; for each c C and o O, call (c, o)
34
that subset of Y X whose members are the functions f such that o f (c). Call
(C, O) the collection of such (c, o). A topology on Y X is then generated from
a basis consisting of all the finite intersections of the sets (c, o) (C, O). This
is the compact-open topology, which coincides with the previous one when X is
compact and Y is metric. Other topologies may be defined on function spaces
and their choice is a matter of convenience. A point worth emphasizing is that,
besides other requirements, a topology is presupposed in functional differentiation
and integration. Function spaces are more complicated because their dimension
is (highly) infinite and many of the usual properties of finite spaces do not carry
over to them.20 As stated in 1.2.13, there is one remarkable difference: when the
topology is given by a norm, an infinite-dimensional space is never locally compact.
1.3.15 Connection by paths

A topological space S is path-connected
(or arcwise-connected ) if, for every two points
p, q in S there exists a path with
(0) = p and (1) = q.
We have said that path-connectedness implies connectedness but there are
connected spaces which are not path-connected. They are, however, very
peculiar spaces and in most applications the word connected is used for
path-connectedness. Some topological properties of a space can be grasped
through the study of its possible paths. This is the subject matter of homotopy theory, of which some introductory notions will be given in chapter 3.
Hilbert spaces are path-connected.
1.3.16 Suppose the space S is not path-connected. The path-component
of a point p is that subset of S whose points can be linked to p by continuous
curves. Of course path is not a gratuitous name. It comes from its most
conspicuous example, the path of a particle in its configuration space. The
evolution of a physical system is, most of times, represented by a curve in the
space of its possible states (see Phys.1). Suppose such space of states is not
path-connected: time evolution being continuous, the system will never leave
the path-component of the initial state. Topological conservation laws are
then at work, the conserved quantities being real functions on the state space
which are constant on each path-component. This leads to a relationship21
20
Differences between finite- and infinite-dimensional spaces are summarized in DeWittMorette, Masheshwari & Nelson 1979, appendix A.
21
Ezawa 1978; idem, 1979.
1.3. FUNCTIONS
35
between topological conservation laws and superselection rules 22 and is the

underlying idea of the notion of kinks.23
The relation pRq = there exists a path on S joining p to q is an
equivalence relation. The path-component of p may be seen as the class of
p.
1.3.17 There are two kinds of conserved quantities in Physics: those coming from Noethers theorem conserved along paths which are solutions
of the equations of motion and the topological invariants which come
from the global, topological properties of the space of states. Noether invariants (see Phys.6) are functions which are constant on the solutions, while the
topological invariants are constant on each path-component of the space of
states. By a convenient choice of topology, also the Noether invariants can
be made into topological invariants.24
1.3.18 Spaces of paths are particular cases of function spaces. Spaces of
paths between two fixed end-points, q : [t1 , t2 ] E3 are used in Feynmans
formulation of non-relativistic quantum mechanics. Integrations on these
spaces presuppose additional structures. Notice that such spaces can be
made into vector spaces if q(t1 ) = q(t2 ) = 0. This is a simple example of
infinite-dimensional topological vector space.25
1.3.19 Measure and probability A measure is a special kind of set
function, that is, a function attributing values to sets (see Math.3). We
actually consider a precise family of subsets, a -algebra A, including the
empty set and the finite unions of its own members. A (positive) measure
is a function attributing to each subset a probability, that is, a positive real
value. A good example is the Lebesgue measure on E1 : the -algebra is that
generated by the open intervals (a, b) with b a and the measure function
is
m[(a, b)] = b a.
A set with a sole point has zero measure. The Cantor set E of 1.2.4 has
m(E) = 0. Measure spaces are easily extended to Cartesian product spaces,
so that the Lebesgue measure goes easily over higher dimensional euclidean
spaces.
22
23
24
25
Streater & Wightman 1964.

Skyrme 1962; Finkelstein 1966 and references therein.
Dittrich 1979.
DeWitt-Morette, Masheshwari & Nelson 1979.
36
1.3.20 Compactification Let us give an example of a more technical use

of homeomorphisms and the equivalences they engender. We have talked
about locally compact spaces and their possible compactification. Given a
locally compact space X, we define formally the Alexandrovs compactified
of X as the pair (X 0 , f 0 ) with the conditions:
(i) X 0 is a compact space;
(ii) f 0 is a homeomorphism between X and the complement in X 0 of a
point p
(iii) if (X 00 , f 00 ) is another pair satisfying (i) and (ii), there exists a unique
homeomorphism h : X 0 X 00 such that f 00 = h f 0 .
We say then that p is the point at infinity of X 0 , and that the compact X 0
is obtained from the locally compact space X by adjunction of an infinity
point. The Alexandrovs compactified of the plane E2 is the pair (sphere
S 2 , stereographic projection from (say) p = the north pole), see Math.11.3.
With enlarged stereographic projections (Phys.9.1), the sphere S n comes out
from the compactification of En .
Physically, such a process of compactification is realized when the following two steps are made: (i) suppose all the Physics of the system is
contained in some functions; for example, in field theory (Phys.6) the fields
are the degrees of freedom, the coordinates of spacetime being reduced to
mere parameters; (ii) the functions or fields are supposed to vanish at all
points of infinity, which makes them all equivalent. Any bound-state problem of nonrelativistic Quantum Mechanics in E3 , in which the wavefunction
is zero outside a limited region, has actually S 3 as configuration space. In
the relativistic case, it is frequent that we first euclideanize the Minkowski
space, turning it into E4 , and then suppose all involved fields to have the
same value at infinity, thereby compactifying E4 to S 4 .
1.4
1.4.1
QUOTIENTS AND GROUPS
Quotient spaces
1.4.1 Consider a spherically symmetric physical system in E3 (say, a central potential problem). We usually (for example, when solving the potential
problem by separation of variables) perform some manipulations to reduce
the space to sub-spaces and arrive finally to an equation in the sole variable
r. All the points on a sphere of fixed radius r will be equivalent, so that
each sphere will correspond to an equivalence class. We actually merge all
the equivalent points into one of them, labelled by the value of r, which is
taken as the representative of the class. On the other hand, points on spheres
1.4. QUOTIENTS AND GROUPS
37
of different radii are nonequivalent, will correspond to distinct equivalence

classes. The radial equation is thus an equation on the space of equivalence
classes.
A physical problem in which a large number of points are equivalent
reduces to a problem on the space of equivalence classes (or spaces whose
points are sub-sets of equivalent points). Such spaces may have complicated topologies. These are called quotient topologies. Mostly, they come up
when a symmetry is present, a set of transformations which do not change
the system in any perceptible way.
The merging of equivalent points into one representative is realized by a
mapping, called a projection. In the above spherical example, the whole set
of equivalence classes will correspond to the real positive line E1+ , on which
r takes its values. The projection will be
: E3 E1+ ,
: p = (r, , ) r .
An open interval in E1+ , say, J = (r , r + ), will be taken back by
into the region between the two spheres of radii (r and (r + . This
region is an open set in E3 so that J is an open set in the quotient space
if is supposed continuous. As distinctions between equivalent points are
irrelevant, the physical configuration space reduces to E1+ . The symmetry
transformations constitute the rotation group in E3 , the special orthogonal
group SO(3). In such cases, when the equivalence is given by symmetry
under a transformation group G, the quotient space is denoted S/G. Here,
<1>
E1+ = E3 /SO(3).
1.4.2 Inspired by the example above, we now formalize the general case
of spaces of equivalence classes. Suppose a topological space S is given on
which an equivalence relation R is defined: two points p and q are equivalent,
p q, if linked by the given relation. We can think of S as the configuration
space of some physical system, of which R is a symmetry: two points are
equivalent when attainable from each other by a symmetry transformation.
All the points obtainable from a given point p by such a transformation
constitute the equivalence class of p, indicated in general by [p] when no
simpler label is at hand. This class may be labelled by any of its points
instead of p, of course, and points of distinct classes cannot be related by
transformations. The set of equivalence classes, which we call {[p]} = S/R,
can be made into a topological space, whose open sets are defined as follows:
let : S S/R be the projection : p [p] (class to which p belongs).
Then a set U contained in S/R is defined to be open iff <1> (U ) is an
38
open set in S. Notice that is automatically continuous. The space S/R is

called the quotient space of S by R and the topology is the quotient topology.
The simplest example is the plane with points supposed equivalent when
placed in the same vertical line. Each vertical line is an equivalence class
and the quotient space is the horizontal axis. In another case, if the plane
is the configuration space of some physical system which is symmetric under
rotations around the origin, all the points lying on the same circle constitute
an equivalence class. The quotient will be the space whose members are the
circles.
1.4.3 Take the real line E1 and the equivalence p q iff p q = n Z (Z
is the set of all integers, an additive group). The space E1 /R, or E1 /Z, has as
set point the interval [0, 1), of which each point represents a class. The open
sets of the quotient topology are now of two types: (i) those of E1 included in
the interval; (ii) the neighbourhoods of the point 0, now the unions u v of
intervals as in the Figure 1.10. This topology is able to close the interval
[0, 1) on itself. The same function f which in 1.3.12 failed to take the
neighbourhoods of 0 into open sets of S 1 will do it now (see Figure 1.10).
Consequently, the same interval [0, 2) (we could have used [0, 1) instead),
becomes, once endowed with the quotient topology, homeomorphic to the
circle. It acquires then all the topological properties of S 1 , for instance that
of being compact.
We shall see later that coordinates are necessarily related to homeomorphisms. The use, for the circle, of an angular real coordinate E1 , which
repeats itself every time it arrives at equivalent points of E1 , is just a manifestation of the quotient relation between E1 and S 1 . Actually, the angle
does belong to the interval [0, 2), but with the quotient topology.
We shall see later that is not a good coordinate, because coordinates
must belong to euclidean spaces.
1.4.4 The M
obius band Even rap-singers are by now acquainted with
the usual definition of this amazing one-sided (or one-edged) object: take a
sheet as in Figure 1.11 and identify a = b0 , b = a0 . This corresponds to taking
the product of the intervals (a, a0 ) (a, b) and twisting it once to attain the
said identification. It is possible to use E1 instead of the limited interval
(a, a0 ). To simplify, use E1 (1, 1) and the equivalence (see Figure 1.12)
given by:
(p1 , p2 ) (q 1 , q 2 ) iff
(i) p1 q 1 = n Z, and
(ii) p2 = ()n q 2 .
Experiment with a paper sheet is commendable: begin with the usual definition and then take sheets which are twice, thrice, etc, as long. A simple
39
Figure 1.10: Kinds of open sets in the quotient topology: (left) first-kind: <1> (u) =
nZ ( + n, + n); (center) second-kind: <1> (u v) = nZ ( + n, + n 1); (right)
with the quotient topology, f (x) = exp[i2x] becomes a homeomorphism.
illustration of the quotient definition comes up. A simple cylinder comes

out if we use the condition p2 = q 2 instead of condition (ii) above. The
cylinder topology so introduced coincides with that of the topological (or
cartesian) product E1 S 1 . The Mobius band, on the other hand, is not
a topological product! Also experiments with a waist belt are instructive
to check the one-edgedness and the fact that, turning twice instead of once
before identifying the extremities, a certain object is obtained which can
be deformed into a cylinder. We can examine free quantum fields in the
original sheet (quantization in a plane box). The use of periodic boundary conditions for the vertical coordinates corresponds to quantization on
the cylinder. Quantization on the Mobius band is equivalent to antiperiodic
boundary conditions and leads to quite distinct (and interesting) results.26
For example, the vacuum (the lowest energy state, see Phys.3.2.2) in the
Mobius band is quite different from the vacuum in the cylinder.
1.4.5 The torus Take E2 and the equivalence
(p1 , p2 ) (q 1 , q 2 ) iff
p1 q 1 = n Z and p2 q 2 = m Z.
A product of two intervals, homeomorphic to the torus T 2 = S 1 S 1 , is
obtained. As T 2 is obtained by dividing the plane by twice the group
of integers Z, we write T 2 = E2 /Z2 . The twisted version is obtained by
modifying the second condition to
p2 ()n1 q 2 = m Z.
26
Avis & Isham, in Levy & Deser 1979.
40
Figure 1.11: The cylinder and the Mobius strip.
The resulting quotient space is the Klein bottle. Experiments with a paper
leaf will still be instructive, but frustating the real construction of the
bottle will show itself impossible (because we live in E3 ; it would be possible
in E4 !). Higher dimensional toruses T n = S 1 S 1 . . . S 1 (n times) are obtained
equally as quotients, T n = En /Zn .
1.4.6 The above examples and the previous chapter considerations on
compactification illustrate the basic fact: for spaces appearing in physical
problems, the topological characteristics are as a rule fixed by boundary
conditions and/or symmetry properties.
1.4.7 As previously said, it is customary to write G instead of R in the
quotient, S/G, when the relation is given by a group G as with the toruses,
the spherical case E1+ = E3 /SO(3), and S 1 = E1 /Z in 1.4.3. Consider
now the configuration space of a system of n particles. From the classical
point of view, it is simply E3n , the product of the configuration spaces of
the n particles. If we now take into account particle indistinguishability,
the particle positions become equivalent. The configuration space will be
the same, but with all the particle positions identified, or glued together.
In an ideal gas, for example, the points occupied by the n particles are to
be identified. The space is insensitive to the exchange of particles. These
exchanges constitute the symmetric group Sn (the group of all permutations
41
Figure 1.12: Scheme of E1 (1, +1) showing the equivalent points to form a Mobius
band.
of n objects, see Mathematical Topic 4) and the final configuration space is

E3n /Sn , a rather messy space.27
Many topological spaces of great relevance possess much more structure
than a simple topology. The additional structure is usually of algebraic
nature, as for example, that of a group.
1.4.2
Topological groups
The euclidean space En is a topological space with the standard euclidean

ball topology. But, being also a vector space, it is furthermore an abelian
group, with the addition operation. A natural question arises: is there any
compatibility relation between the algebraic group structure and the topological structure? The answer is positive: the mappings
(X, Y ) X + Y
X X
are continuous functions En En En and En En respectively. In
rough words, the group operations are continuous maps in the underlying
topology. The notion of topological group comes from this compatibility
between algebraic and topological structures.
Let us rephrase the definition of group (see Mathematical Topic 1). A
group is a set G (whose members are its elements) with an operation
m : G G G, given by m : (a, b) m(a, b), and defined in such a way
that:
27
See for instance Laidlaw & DeWitt-Morette 1971.
42
(i) m(a, b) G for all pairs a, b;

(ii) there exists a neutral element e such that m(a, e) = m(e, a) = a;
(iii) every a G has an inverse element a1 which is such that
m(a, a1 ) = m(a1 , a) = e;
(iv) m(a, m(b, c)) = m(m(a, b), c) for all a, b, c G.
We can think of a mapping inv which take each element into its inverse,
inv: G G, inv : a a1 . Suppose now a topological space G endowed
with such a group structure.
1.4.8 A topological group is a group G whose elements are members of
a topological space in whose topology both the mappings m : G G
G given by (a, b) m(a, b) and inv: G G given by a a1 are
continuous (with the product topology for G G).28
1.4.9 The theory of topological groups has three main chapters, concerning their algebraic structure, their topological structure and their representations. But even the algebraic aspects becomes frequently clearer when something else, such as the differential structure, is added. Representation theory
Math.6) involves functions and measures defined on topological groups, which
are the subject of modern harmonic analysis.
1.4.10 A topological group is compact if it is compact as a manifold. This
means that from any sequence {gn } of its elements one can extract a finite
convergent sub-sequence. Any abstract group is a topological group with
respect to the discrete topology .
1.4.11 The additive group of real numbers (or 1-dimensional translations)
is an abelian non-compact group whose underlying group-space is the infinite
real line. It is commonly denoted by R, but (R, +), indicating also which
operation is supposed, would be a better notation.
1.4.12 The sets R\{0} and C\{0} of nonvanishing real and complex numbers are topological groups with the product operation. The set S 1 C,
S 1 = {z C such that |z| = 1} is a topological group (S 1 , ) with the operation of multiplication and the induced topology. S 1 is actually a subgroup
of C\{0}.
1.4.13 This example illustrates a more general result:
if G is any topological group and H a subgroup of G, then H is a
topological group with the induced topology of the subspace.
28
A classical text on topological groups is Pontryagin 1939.
43
The circle S 1 is thus a topological group and the set of positive real numbers
R+ is a topological subgroup of R\{0}. As suggested by the example of En ,
normed vector spaces in general (reviewed in Math.4), and consequently their
vector subspaces, are topological groups with the addition operation
m(a, b) = a + b.
1.4.14 The set Q of quaternions is a vector space on R with basis {1, i, j, k}.
It constitutes topological groups with respect to the addition and multiplication operations, and is isomorphic to R4 as a vector space. An element of
Q can be written as
q = a 1 + bi + cj + dk ,
with a, b, c, d R. The basis member 1 acts as the identity element. The
multiplications of the other members are given in Table 1.1.
i
j
k
i
-1
-k
+j
j
+k
-1
-i
k
-j
+i
-1
Table 1.1: Multiplication table for quaternions.
1.4.15 Let S 3 Q, where

S 3 = {q Q such that |q| =
a2 + b2 + c2 + d2 = 1}.
It turns out that S 3 is a topological group.

1.4.16 Linear groups Let GL(m, K) be the set of all non-singular matrices m m, with entries belonging to a field K = R, C or Q. In short,
GL(m, K) = {(m m) matrices g on K such that det g 6= 0}.
Given a vector space over the field K, the group of all its invertible linear
transformations is isomorphic to GL(m, K). The subsets
2
GL(m, R) Rm
2
GL(m, C) R(2m)
2
GL(m, Q) R(4m)
are open sets and also topological groups with the operation of matrix multiplication and the topologies induced by the inclusions in the respective
44
euclidean spaces. These linear groups are neither abelian nor compact. Generalizing a bit: let V be a vector space; the sets of automorphisms and
endomorphisms of V are respectively
Aut V = {f: V V , such that f is linear and invertible}
End V = {f: V V , such that f is linear} .
Then, Aut V End V is a topological group with the composition of
linear mappings as operation. If we represent the linear mappings by their
matrices, this composition is nothing more than the matrix product, and we
have precisely the case GL(m, K).
The groups O(n) GL(m, R) of orthogonal n n matrices are other
examples. The special orthogonal groups SO(n) of orthogonal matrices with
det = +1 are ubiquitous in Physics and we shall come to them later.
1.4.17 The set of all matrices of the form ( L0 1t ), where L GL(n, R) and
t Rn , with the matrix product operation, is a topological group called the
affine group of dimension n. It is denoted A(n, R).
1.4.18 Linear projective groups Take the set Mn of all n n matrices
with entries in field K. It constitutes more than a vector space it is
an algebra. Each matrix A GL(n, K) defines an automorphism of the
form AM A1 , for all M Mn . This automorphism will be the same if A is
replaced by kA = kIn A, for any k K, and where In is the unit nn matrix.
The subgroup formed by the matrices of the type kIn is indeed ineffective on
Mn . The group of actual automorphisms is the quotient GL(n, K)/{kIn }.
Consider the subalgebra of Mn formed by the n projectors Pk = matrix
whose only nonvanishing element is the diagonal k-th entry, which is 1:
(Pk )rs = rs ks .
The Pk s are projectors into 1-dimensional subspaces. They satisfy the relations Pi Pj = ij Pi . Each one of the above automorphisms transforms this set
of projectors into another set of projectors with the same characteristics. For
this reason the automorphisms are called projective transformations, and the
group of automorphisms GL(n, K)/{kIn } is called the projective group, denoted P L(n, K). The n-dimensional space of projectors, which is taken into
an isomorphic space by the transformations, is the projective space, indicated
by KP n . There are, however, other approaches to this type of space.
1.4.19 Projective spaces Every time we have a problem involving only
the directions (not the senses) of vectors and in which their lengths are
irrelevant, we are involved with a projective space. Given a n-dimensional
vector space V over the field K, its corresponding projective space KP n is
45
the space formed by all the 1-dimensional subspaces of V . Each point of

KP n is the set formed by a vector v and all the vectors proportional to v.
We may be in a finite-dimensional vector space, or in an infinite-dimensional
space like a Hilbert space. Quantum Mechanics describes pure states as rays
in a Hilbert space, and rays are precisely phase-irrelevant. Thus, pure states
are represented by members of a projective Hilbert space.
Take the sphere S n and the equivalence given by: p q if either p
is identical to q, or p and q are antipodes. For S 1 drawn on the plane,
(x, y) (x, y); on S 2 imbedded in E3 , (x, y, z) (x, y, z); etc.
The quotient space S n /R is the n-dimensional real projective space RP n .
Because pairs of antipodes are in one-to-one correspondence with straight
lines through the origin, these lines can be thought of as the points of RP n .
The space RP 1 is called the real projective line , and RP 2 , the real
projective plane (it is the space of the values of orientation fields, see
Phys.3.1). It can be shown that RP n is a connected Hausdorff compact
space of empty boundary. The antipode relation can be related to a group,
the cyclic group ( Math.2.3) of second order Z2 , so that RP n = S n /Z2 .
There are many beautiful (and not always intuitive) results concerning these
spaces. For instance: RP 0 is homeomorphic to a point; RP 1 is homeomorphic
to S 1 ; the complement of RP n1 in RP n is En ; RP 3 is homeomorphic to
the group SO(3) of rotations in E3 ; etc. Complex projective spaces CP n
are defined in an analogous way and also for them curious properties have
been found: CP 1 is homeomorphic to S 2 ; the space CP n is homeomorphic
to S 2n+1 /S 1 (recall that S 1 is indeed a group); the complement of CP n1
in CP n is E2n ; etc. They are ubiquitous in Mathematics and have also
emerged in many chapters of physical theory: model building in field theory,
classification of instantons,29 twistor formalism,30 etc. Another, equivalent
definition is possible which makes no use of the vector structure of the host
spaces En or Cn . It is also a quotient definition. Consider, to fix the ideas,
the topological space Cn+1 of ordered (n+1)-uples of complex numbers, Cn+1
= {z = (z1 , z2 , . . . , zn , zn+1 )}, with the ball topology given by the distance
function
q
0
0
|2 .
d(z, z ) = |z1 z10 |2 + |z2 z20 |2 + . . . + |zn zn0 |2 + |zn+1 zn+1
The product of z by a complex number c is cz = (cz1 , cz2 , . . . , czn , czn+1 ).
Define an equivalence relation R as follows: z and z 0 are equivalent if some
non-zero c exists such that z 0 = cz. Then, CP n = Cn+1 /R, that is, the
29
30
1977.
Atiyah, Drinfeld, Hitchin & Manin 1978; Atiyah 1979.

Penrose in Isham, Penrose & Sciama 1975; Penrose & MacCallum 1972; Penrose
46
quotient space formed by the equivalence classes of the relation. Any function
f on Cn+1 such that f (z) = f (cz) (function which is homogeneous of degree
zero) is actually a function on CP n .
1.4.20 Real Grassmann spaces Real projective spaces are generalized
in the following way: given the euclidean space EN with its vector space structure, its d-th Grassmann space GN d (E) [another usual notation: Gd (EN )] is
the set of all d-dimensional vector subspaces of EN . Remark that, as vectors
subspaces, they all include the origin (the zero vector) of EN . All Grassmannians are compact spaces (see 8.1.14). Projective spaces are particular
cases:
RP 1 = G1 (E2 ), RP 2 = G1 (E3 ), . . . , RP n = G1 (En+1 ).
Projectors are in a one-to-one correspondence with the subspaces of a vector
space. Those previously used were projectors of rank one. In the present
case the euclidean structure is to be preserved, so that the projectors must
be skew-symmetric endomorphisms, or matrices p satisfying p = pT and
p2 = p. If they project into a d-dimensional subspace, they must furthermore
be matrices of rank d. Consequently, GN d may be seen also as the space of
such projectors:
GN d = {p End(EN ) such that p2 = p = pT , rank p = d}.
Notice that GN d = GN,N d and the dimension is dim GN d = d(N d).
The set of orthogonal frames in the d-dimensional subspaces form another
space, the Stiefel space (see 8.1.15), which is instrumental in the general
classification of fiber bundles (section 9.7).
1.4.21 Complex Grassmann spaces Projective spaces are generalized
to the complex case in an immediate way. One starts from a euclideanized
(as above) complex vector space CN = (z1 , z2 , . . . , zn , zn+1 )}. Now, projectors
must be hermitian endomorphisms and the space is
N
GC
N d = Gd (C ) =
GN d (C) = {p End(CN ) such that p2 = p = p , rank p = d }.
GN d (C) is a compact space whose points are complex d-dimensional planes

in CN .
1.4.22 We have tried in this chapter to introduce some notions of what is
usually called general topology. The reader will have noticed the purely qualitative character of the discussed properties. The two forthcoming chapters
47
are devoted to some notions coming under the name of algebraic topology,
which lead to the computation of some numbers of topological significance.
Roughly speaking, one looks for defects (such as holes and forbidden subspaces) in topological spaces, while remaining inside them. One way to find
such faults comes from the observation that defects are frequently related to
closed subspaces which do not bound other subspaces, as it happens with
some of the closed lines on the torus. Such subspaces give origin to discrete
groups and are studied in homology theory (chapter 2). Another way is by
trying to lasso them, drawing closed curves (and their higher dimensional
analogues) to see whether or not they can be continuously reduced to points
(or to other closed curves) in the space. Such loops are also conveniently
classified by discrete groups. This is the subject of homotopy theory (chapter 3). Both kinds of groups lead to some integer numbers, invariant under
homeomorphisms, which are examples of topological numbers. We shall later
introduce more structure on topological spaces and sometimes additional
structure can be used to signal topological aspects. Within differentiable
structure, vector and tensor fields arise which are able to reveal topological
defects (singular points, see Math.9) through their behaviour around them,
in a way analogous to the velocity field of a fluid inside an irregular container.
Later, in section 7.5, a little will be said about differential topology.
48
Chapter 2
HOMOLOGY
Dissection of a space into cellular building bricks provides information about
its topology.
The study of the detailed topological characteristics of a given space can

be a very difficult task. Suppose, however, that we are able to decompose the
space into subdomains, each one homeomorphic to a simpler space, whose
properties are easier to be worked out. Suppose further that this analysis is
done in such a way that rules emerge allowing some properties of the whole
space to be obtained from those of these components. Homology theory
is concerned precisely with such a program: the dissection of topological
spaces into certain basic cells called chains, which in a way behave as
their building bricks. The circle is homeomorphic to the triangle and the
sphere is equivalent to the tetrahedron. The triangle and the tetrahedron
can be build up from their vertices, edges and faces. It will be much easier
to study the circle and the sphere through such rectified versions, which
furthermore can be decomposed into simpler parts. Once in possession of the
basic cells we can, by combining them according to specific rules, get back
the whole space, and that with a vengeance: information is gained in the
process. Chains come out to be elements of certain vector spaces (and so,
of some abelian groups). Underlying algebraic structures come forth in this
way, which turn out to be topological invariants: homeomorphic spaces have
such structures in common.
As will be seen later on, chains are closely related to integration domains in the case
of differentiable manifolds, and their algebraic properties will find themselves reflected in
analogous properties of differential forms. Of course, only a scant introduction will find
its place here. The subject is a whole exceedingly beautiful chapter of high Mathematics,
and an excellent primer may be found in chapters 19 to 22 of Fraleigh 1974. An extensive
mathematical introduction, easily understandable if read from the beginning, is Hilton &
Wylie l967.
49
50
2.1
CHAPTER 2. HOMOLOGY
GRAPHS
Spaces usually appearing in Physics have much more structure than a mere
topology. Exceptions are the graphs (or diagrams) used in perturbation techniques of Field Theory, cluster expansions in Statistical Mechanics, circuit
analysis, etc, whose interest comes in part from their mere topological properties. Graph Theory is a branch of Mathematics by itself, with important
applications in traffic problems, management planning, electronics, epidemic
propagation, computer design and programming, and everywhere else. Only
a very sketchy outline of the subject will be given, although hopefully pedantic enough to prepare for ensuing developments. Graphs provide a gate into
homological language. The first part below introduces them through a mere
formalization of intuitive notions. The second rephrases (and extends) the
first: its approach allows the introduction of the ideas of chain and boundary,
and sets the stage for the basic notions of simplex and complex.
2.1.1
Graphs, first way
2.1.1 We shall call (closed) edge any subspace in E3 which is homeomorphic to the interval I = [0, 1] with the induced topology. Indicating the
edge itself by e, and by h : I e the homeomorphism, the points h(0) and
h(1) will be the vertices of the edge. A graph G E3 is a topological space
defined in the following way:
(i) the point set consists of the points of a union of edges satisfying the
condition that the intersection of any two edges is either or one common
vertex;
(ii) the open sets are subsets X G whose intersection with each edge
is open in that edge.
2.1.2 Because edges are homeomorphic to I with the induced topology
(and not, for example, with the quotient topology of 1.4.3), no isolated
one-vertex bubbles like are admitted. Notice that knots (Mathematical
Topic 2.3) are defined as subspaces in E3 homeomorphic to the circle S 1 . So,
absence of bubbles means that no knots are parts of graphs. Graphs can be
defined in a less restrictive way with no harm to the results involving only
themselves, but the above restrictions are essential to the generalization to
be done later on (to simplexes and the like). From (ii), a subset is closed if its
intersection with each edge is closed in that edge. The open edges, obtained
by stripping the closed edges of their vertices, are open sets in G. The set of
vertices is discrete and closed in G. A graph is compact iff it is finite, that is,
if it has a finite number of edges. It is connected iff any two of its vertices can
2.1. GRAPHS
51
be linked by a sequence of non-disjoint edges belonging to the graph. Take

an open edge e and call e0 its complement, e0 = G e (G with extraction of
e). If, for every edge e, e0 is unconnected, then G is a tree graph. Otherwise,
it contains at least one loop, which is a finite sequence of edges
e1 , e2 , . . . , en
each one like ei = ui vi , with no repeated edges and only one repeated
vertex, vn = u1 , all the remaining ones satisfying vi = ui+1 . These are, of
course, complicated but precise definitions of usual objects. Notice graphs
are drawn in E3 with no intersections. It can be rigorously proved that any
graph can be realized in E3 without crossing.
In the Feynman diagram quantization technique of field theory, the number of loops is just the order in Plancks constant ~. The semiclassical approximation turns up when only tree diagrams (zero order in ~, zero loops)
are considered.
A tree edge is called a branch and an edge taking part in a loop is called
a chord .
2.1.3 Euler number Let us make a few comments on planar graphs,
those that can be drawn in E2 (although always considered as above, built
in E3 . The simplest one is . Call V the number of vertices, E the
number of edges, L the number of loops, and define
(G) = V E + L.
For , (G) = 1. In order to build more complex graphs one has to
add edges one by one, conforming to the defining conditions. To obtain a
connected graph, each added edge will have at least one vertex in common
with the previous array. A trivial checking is enough to see that (G) remains
invariant in this building process. For non-connected graphs, (G) will have
a contribution as above for each connected component. Writing N for the
number of connected components, V E + L N is an invariant. For
connected compact graphs, (G) is called the Euler number . Being an integer
number, it will not change under continuous deformations. In other words,
(G) is invariant under homeomorphisms, it is a topological invariant. It will
be seen later that (G) can be defined on general topological spaces. The
number of loops, with the relation L = 1V +E, was first used in Physics by
Kirchhoff in his well known analysis of DC circuits, and called cyclomatic
number by Maxwell. The result is also valid for Feynman diagrams and,
with some goodwill, for an island in the sea: take a diagram as that pictured
in Figure 2.1, keep the external edges fixed on the plane and pull the inner
52
CHAPTER 2. HOMOLOGY
vertices up the number of peaks (V) minus the number of passes (E) plus
the number of valleys (L) equals one! To compare with the non-planar case,
consider the tetrahedron of Figure 2.2. A simple counting shows that (G)
= V E + L = 2. Its plane, flattened version beside has (G) = 1 because
the lower face is no more counted. Notice that only elementary loops,
those not containing sub-loops, are counted (see 2.2.3 below).
Figure 2.1: Scheme of an island, with peaks (vertices), passes (edges) and valleys (loops).
Figure 2.2: The tetrahedron and one of its planar projections.
2.1.2
Graphs, second way
2.1.4 Consider now E3 as a vector space, and choose any three of its
vectors v0 , v1 , v2 , imposing however that (v1 - v0 ) and (v2 - v0 ) be linearly
independent:
k 1 (v1 v0 ) + k 2 (v2 v0 ) = 0 implies k 1 = k 2 = 0.
Defining a1 = k 1 , a2 = k 2 , a0 = (k 1 + k 2 ), this is equivalent to saying that
the two conditions
a0 v 0 + a1 v 1 + a2 v 2 = 0
a0 + a1 + a2 = 0
2.1. GRAPHS
53
imply a0 = a1 = a2 = 0. Such conditions ensure that no two vectors are

colinear, that (v0 , v1 , v2 ) constitute a triad. Let us define a vector dependent
on the triad (v0 , v1 , v2 ) by the two conditions
b=
2
X
i=0
b vi ;
2
X
bi = 1.
(2.1)
i=0
The points determined by the barycentric coordinates bi describe a plane in

E3 (Figure 2.3).
Figure 2.3: The barycentric coordinates.
2.1.5 Suppose we consider only two of the vectors, say v0 and v1 . They
are of course linearly independent. We can in this case define, just as above,
their dependent vectors by
e = b0 v0 + b1 v1 , b0 + b1 = 1.
Now, the coordinates b0 and b1 determine points on a straight line. In the
same way, we can take only one of the vectors, say v0 , and its dependent
vector as v = v0 itself: this will determine a point.
Add now an extra condition on the dependent vectors: that each coordinate be strictly positive, bi > 0. Then, the vector b above will span the
54
CHAPTER 2. HOMOLOGY
interior of a triangle; the vector e will span an open edge; and v0 again will
span an isolated point, or a vertex. Notice that the coordinates related to
the vector e give actually a homeomorphism between the interval (0, 1) and a
line segment in E3 , justifying the name open edge. If instead we allow bi 0,
a segment homeomorphic to the closed interval [0, 1] results, a closed edge.
With these edges and vertices, graphs may now be defined as previously. An
edge can be indicated by the pair (vi , vk ) of its vertices, and a graph G by a
set of vertices plus a set of pairs. An oriented graph is obtained when all the
pairs are taken to be ordered pairs which is a formal way of putting arrows
on the edges. A path from vertex v1 to vertex vn+1 is a sequence of edges
e1 e2 e3 . . . en with ei = (vi , vi+1 ) or ei = (vi+1 , vi ). We have said that a graph is
connected when, given two vertices, there exists at least one path connecting
them. It is multiply-connected when there are at least two independent, nonintersecting paths connecting any two vertices. It is simply-connected when
it is connected but not multiply-connected. In Physics, multiply-connected
graphs are frequently called irreducible graphs.
2.1.6 A path is not supposed to accord itself to the senses of the arrows.
A path is called simple if all its edges are distinct (one does not go twice
through the same edge) and all its vertices are distinct except possibly v1 =
vn+1 . In this last case, it is a loop. An Euler path on G is a path with all edges
distinct and going through all the vertices in G. The number nk of edges
starting or ending at a vertex vk is its degree (coordination number would
be more to the physicists taste; chemists would probablyP
prefer valence).
Clearly the sum of all degrees on a graph is even, as V1 ni = 2E. The
number of odd vertices (that is, those with odd degrees) is consequently even.
2.1.7 The Bridges of K
onigsberg Graph theory started up when Euler
faced this problem. There were two islands in the river traversing Kants
town, connected between each other and to the banks by bridges as in the
scheme of Figure 2.4. People wanted to know whether it was possible to do
a tour traversing all the bridges only once and finishing back at the starting
point. Euler found that it was impossible: he reasoned that people should
have a departure for each arrival at every point, so that all degrees should
be even which was not the case.
2.1.8 Graph Theory has scored a beautiful victory for Mathematics with
the recent developments on the celebrated (see Phys.3.2.5) four-color problem. The old conjecture, recently demonstrated, was that four colors were
sufficient for the coloring of a map. This is a problem of graph theory. As said
above, graphs are also of large use in many branches of Physics. Through
2.1. GRAPHS
55
Figure 2.4: Scheme of downtown Konigsberg, and the corresponding graph.
Feynmans diagram technique, they have a fundamental role as guidelines

for perturbation calculations in field theory and in the many body problem.
In Statistical Mechanics, besides playing an analogous role in cluster expansions, graphs are basic personages in lattice models. In the Potts model, for
example, where the underlying lattice can be any graph, they become entangled (sorry!) with knots (Phys.3.2.3). They also appear in the generalized
use of the Cayley tree and the related Bethe lattice in approximations to
more realistic lattice models (Phys.3.2.4).
2.1.9 To the path e1 e2 e3 . . . en we can associate a formal sum
1 e1 + 2 e2 + . . . + n en ,
with i = +1 if ei = (vi , vi+1 ), and i = 1 if ei = (vi+1 , vi ). The sum is thus
obtained by following along the path and taking the (+) sign when going in
the sense of the arrows and the (-) sign when in the opposite sense. The sum
is called the chain of the path and i is the incidence number of ei in the
chain.
2.1.10 A further formal step, rather gratuitous at first sight, is to generalize the i s to coefficients which are any integer numbers: a 1-chain on the
graph G is a formal sum
P
i mi ei = m1 e1 + m2 e2 + . . . + mn en ,
with mj Z.
56
CHAPTER 2. HOMOLOGY
2.1.11 We can define the sum of two 1-chains by

P
P 0
P
0
i mi ei +
i mj ej =
k (mk + mk )ek .
Calling 0 the 1-chain with zero coefficients (the zero 1-chain), the set of
1-chains of G constitutes an abelian group, the first order chain group on G,
usually denoted C1 (G). In a similar way, a 0-chain on G is a formal sum
r1 v 1 + r 2 v 2 + . . . + r p v p ,
with rj Z. Like the 1-chains, the 0-chains on G form an abelian group, the
zeroth chain group on G, denoted C0 (G). Of course, C0 (G) and C1 (G) are
groups because Z is itself a group: it was just to obtain groups that we have
taken the formal step i mj above. Groups of chains will be seen to be
of fundamental importance later on, because some of them will show up as
topological invariants.
2.1.12 Take the oriented edge ej = (vj , uj ). It is a 1-chain by itself. We
define the (oriented) boundary of ej as the 0-chain
ej = uj vj . In the same
way, the boundary of a general 1-chain is defined as
P
P
i mi ei = i mi ei
which is a 0-chain.
2.1.13 The mapping
: C1 (G) C0 (G)
preserves the group operation and is called the boundary homomorphism. A
1-cycle on G is a loop, a closed 1-chain. It has no boundary and is formally
defined as an element c C1 (G) for which c = 0 (the zero 0-chain). The
set of 1-cycles on G form a subgroup, denoted Z1 (G).
2.1.14 Consider the examples of Figure 2.5: Take first the graph at the
left in the figure: clearly,
(e1 + e2 ) = v3 v1 ,
(me1 + ne2 ) = nv3 mv1 + (m n)v2 .
In the second graph,
(e1 + e2 + e3 ) = 2v3 2v1 .
2.2. THE FIRST TOPOLOGICAL INVARIANTS
57
Figure 2.5: Examples of low-dimensional chains.

Now, in the graph at the right, (e1 +e2 +e3 ) is clearly a cycle (which illustrates
the extreme importance of orientation). On this graph,
C1 (G) = {me1 + ne2 + re3 ; m, n, r Z};
C0 (G) = {pv1 + qv2 + sv3 ; p, q, s Z};
Z1 (G) = {m(e1 + e2 + e3 ) C1 (G)}.
A very interesting survey of graph theory, including old classical papers (by Euler, Kirchhoff, ... ) is Biggs, Lloyd & Wilson 1977. An introduction to Homology, most commendable
for its detailed treatment starting with graphs, is Gibling 1977. An introduction to graphs
with applications ranging from puzzles to the four color problem is Ore 1963. Finally, a
more advanced text is Graver & Watkins 1977.
2.2
2.2.1
THE FIRST TOPOLOGICAL INVARIANTS
Simplexes, complexes & all that
2.2.1 Let us now proceed to generalize the previous considerations to objects which are more than graphs. Consider En with its structure of vector
space. A set of vertices is said to be linearly independent if, for the set of
vectors (v0 , v1 , v2 , . . . , vp ) fixing them, the two conditions
a0 v 0 + a1 v 1 + . . . + ap v p = 0
a0 + a1 + . . . + ap = 0
imply a0 = a1 = . . . = ap = 0. This means that the vectors (vi v0 )
are linearly independent. We define a vector b dependent on the vectors
v0 , v1 , . . . , vp by
P
P
b = pi=0 bi vi ; pi=0 bi = 1.
58
CHAPTER 2. HOMOLOGY
The points determined by the barycentric coordinates bi describe a p-dimensional subspace of En , more precisely, an euclidean subspace of En .
2.2.2 A (closed) simplex of dimension p (or a p-simplex ) with vertices
v0 , v1 , v2 , . . . , vp is the set of points determined by the barycentric coordinates satisfying the conditions bi 0 for i = 0, 1, 2, . . . , p. Special cases are
points (0-simplexes), closed intervals (1-simplexes), triangles (2-simplexes)
and tetrahedra (3-simplexes). A p-simplex is indicated by sp and is said to
be generated by its vertices. The points with all nonvanishing barycentric
coordinates are interior to the simplex, their set constituting the open simplex. The boundary sp of sp is the set of points with at least one vanishing
coordinate bi = 0. Given sp generated by vo , v1 , v2 , . . . , vp , any subset of
vertices will generate another simplex: if sq is such a subsimplex, we use the
notation sq h sp . It is convenient to take the empty set as a subsimplex of
any simplex. Dimension theory gives the empty set the dimension (-1), so
that the empty simplex is designated by s1 . The edge in the left branch of
Figure 2.6 is a 1-simplex in E3 . Its boundary is formed by the vertices v0
and v1 , which are also subsimplexes:
v 0 h s1 ; v 1 h s1 .
Figure 2.6: The most trivial simplexes in E3 .

Also s1 is taken to be a subsimplex of itself, s1 h s1 . The empty set being
by convention a subsimplex of any simplex,
s1 h s1
s1 h vo
s1 h v1 .
The (full) triangle is a 2-simplex in E3 . Its boundary is the set of points

belonging to the three edges. And so on.
59
2.2.3 The highest possible dimension of a simplex in En is (n 1). If we

are to see loops, even planar graphs are to be considered in E3 . That is
why we have not counted loops encircling other loops in our discussion of
graphs. A (full) tetrahedron is a 3-simplex in En for n 4. Notice that a
simplex is always a convex set in En . It contains, together with any two of its
points, all the points of the straight segment of line between them. Simplexes
will be used as building bricks to construct more complex objects (fittingly
enough called complexes) and this convexity property is very convenient
for the purpose. Notice that, due to condition (ii) of 2.1.1, the vertices
loose their individuality when considered as points of the boundary of a full
triangle: the requirement that the intersection with every edge must be open
excludes half-open intervals.
2.2.4 Up to now, we have been generalizing to higher dimensions the notions of edge and vertex we have seen in the case of graphs. Let us proceed
to the extension of the very idea of graph. A simplicial complex (or cellular
complex) is a set K of simplexes in En satisfying the conditions:
i) if sp K and sq h sp , then sq K;
ii) if sr K and st K, then their intersection sr st is either empty or a
subsimplex common to both.
2.2.5 The dimension of K is the maximal dimension of its simplexes. A
graph is a 1-dimensional simplicial complex. Notice that K is the set of
the building blocks, not a set of points. The set of points of En belonging
to K with the induced topology is the polyhedron of K, indicated by P(K).
Conversely, the complex K is a triangulation or dissection of its polyhedron.
Due to the way in which it was constructed, a polyhedron inherits much
of the topology of En ; in particular, its topology is metrizable.
2.2.6 We now come to the main idea: suppose that a given topological
space S is homeomorphic to some P(K). Then, S will have the same topological properties of K. We shall see below that many topological properties of
P(K) are relatively simple to establish. These results for polyhedra can then
be transferred to more general spaces via homeomorphism. Suppose h is the
homeomorphism, h(K) = S. The points of S constitute then a curvilinear
simplex. For each simplex sp , the image h(sp ) S will keep the properties
of sp . The image h(sp ) will be called a p-simplex on S. Again, the set of
curvilinear simplexes on S. will be a triangulation (or curvilinear complex)
of S.. A simple example is the triangulation determined on Earths surface
by the meridians and the equator.
60
CHAPTER 2. HOMOLOGY
2.2.7 The boundary of a triangle is homeomorphic to the circle S 1 . This

can be seen by first deforming the triangle into an equilateral one, inscribing
a circle and then projecting radially. This projection is a homeomorphism.
An analogous procedure shows that the sphere S n is homeomorphic to the
boundary of an (n + 1)-tetrahedron, or (n + 1)simplex.
2.2.8 An important point is the following: above we have considered a
homeomorphism h. However, it is possible in many cases to use a less stringent function to relate the topological properties of some topological space
to a simplicial complex. Suppose, for instance, that S is a differentiable
manifold (to be defined later: it is endowed with enough additional structure to make differential calculus possible). Then, it is enough that h be a
differentiable function, which in many aspects is less stringent than a homeomorphism: its inverse has not necessarily a good behaviour and h itself may
be badly behaved (singular) in some regions. By taking simplexes from K
to S via a differentiable function, a homology can be introduced on S, the
singular homology. In reality, many homologies can be introduced in this
way, by choosing different conditions on h. It can be shown that, at least for
compact differentiable manifolds, all these homologies are equivalent, that
is, give the same topological invariants. A differentiable manifold turns out
to be homeomorphic to a polyhedron, even if h is originally introduced as a
differentiable function from some polyhedron of En into it (Cairns theorem).
2.2.9 We have said that the region inscribed in a loop is not a simplex
in E2 . That is why graphs are to be considered as 1-simplexes in E3 . The
dimension of the surrounding space is here of fundamental importance. Let us
repeat what has been said in 2.1.3 on the tetrahedron and its flattening.
Take the graph of Figure 2.7 and pull the inner vertex up. In order to obtain
the boundary of a tetrahedron (a simplex in E4 , that is, a new 2-complex,
an extra bottom face has to be added and the Euler number becomes 2. The
boundary of a tetrahedron being homeomorphic to the sphere S 2 , it follows
that also (S 2 ) = 2 . The same will be true for any surface homeomorphic
to the sphere. If the surface of the Earth (or Mars, or Venus, or the Moon)
could be obtained by continuous deformations from the sphere, the number of
peaks minus passes plus valleys would be two. Of course, this would neglect
steep cliffs and any other singularities.
2.2.10 We are now in condition to introduce a topological invariant which
generalizes the Euler number. Call Ni the number of i-simplexes in a given
n-dimensional complex K. The Euler-Poincare characteristic of K is defined
61
Figure 2.7: To get a tetrahedron, a face must be added.

by
(K) =
n
X
()i Ni .
i=0
Being an integer number, it cannot be changed by continuous deformations

of the space: it is an example of topological number , an integer number which
is characteristic of a space and necessarily the same for spaces homeomorphic
to it.
2.2.11 Notice that non-homeomorphic spaces can have some topological
numbers in common, their equality being only a necessary condition for topological equivalence. The circle S 1 and all the toruses T n = S 1 S 1 . . . S 1
(topological product of n circles) have = 0 but are not homeomorphic.
They have other topological characteristics which are different, as for example the fundamental group, a homotopic character to be examined in Chapter
3), Homology, like homotopy, provides only a partial characterization of the
involved topology.
2.2.12 In order to obtain more invariants and a deeper understanding
of their emergence we need beforehand to generalize the oriented graphs of
2.1.5. A simplex sp is generated by its vertices and can be denoted by their
specification:
sp = (v0 , v1 , v2 , . . . , vp ).
Suppose now that (v0 , v1 , v2 , ... , vp ) is an ordered (p+1)-uple of vertices.
Each chosen order is an orientation of the simplex. For instance, the edge
s1 can be given the orientations v0 v1 and v1 v0 . It is only natural to consider
these orientations as opposite, and write
v0 v1 = v0 v1 = v1 v0 = (v1 v0 ) .
2.2.13 It is also convenient to think of a 2-simplex as a (full) triangle
oriented via a fixed sense of rotation, say counter-clockwise (Figure 2.8).
Here a problem arises: the edges are also oriented simplexes and we would
62
CHAPTER 2. HOMOLOGY
like to have for the boundary an orientation coherent with that of the 2simplex. The figure suggests that the faces coherently oriented with respect
to the triangle are v0 v1 , v1 v2 and v2 v0 , so that the oriented boundary is
(v0 v1 v2 ) = v0 v1 + v1 v2 + v2 v0 .
(2.2)
Figure 2.8: Triangle orientation: counter-clockwise convention.
2.2.14 Notice that the opposite orientation would be coherent with the
opposite orientation of the edges. All the possible orientations are easily
found to be equivalent to one of these two. As a general rule, for a p-simplex
(v0 v1 v2 . . . vp ) = (vi0 vi1 vi2 . . . vip ) , the sign being + or - according to
whether the permutation

0 1 2 ... p
i0 i1 i2 ... ip
is even or odd. An equivalent device is to think of (v0 v1 v2 . . . vp ) as an antisymmetric product of the vertices, or to consider that vertices anticommute.
This is consistent with the boundary of an edge,
(v0 v1 ) = v1 v0 .
As a mnemonic rule,
(v0 v1 v2 . . . vp ) = v1 v2 . . . vp v0 v2 . . . vp + v0 v1 v3 . . . vp . . .
The successive terms are obtained by skipping each time one vertex in the
established order and alternating the sign. For a 0-simplex, the boundary is
defined to be 0 (which is the zero chain). Each term in the sum above, with
the corresponding sign, is a (oriented) face of sp .
2.2.15 Let us go back to eq.[2.2], giving the boundary of the simplex
(v0 v1 v2 ). What is the boundary of this boundary? An immediate calculation
shows that it is 0. The same calculation, performed on the above general
definition of boundary, gives the same result because, in (v0 v1 v2 . . . vp ),
63
each (p-2)-simplex appears twice and with opposite signs. This is a result of
fundamental significance:
sp 0.
(2.3)
The boundary of a boundary of a complex is always the zero complex.

2.2.16 A complex K is oriented if every one of its simplexes is oriented.
Of course, there are many possible sets of orientations and that one which is
chosen should be specified. Let us consider the set of all oriented simplexes
belonging to an oriented complex K,
N1
Np
1 2
1 2
0
s10 , s20 , . . . , sN
0 , s1 , s1 , . . . , s1 , , . . . , sp , sp , . . . , sp
where Ni = number of i-simplexes in K.

2.2.17 A p-chain cp of K is a formal sum
cp =
m1 s1p
m2 s2p
+ ... +
p
m Np sN
p
Np
X
mj sjp ,
j=1
with mj Z. The number p is the dimension of Cp . Just as seen in 2.1.11,

the p-chains of K form a group, denoted Cp (K). The boundary of the chain
cp is
cp =
Np
X
mj sjp .
j=1
Each term in the right hand side is a face of cp . Equation [2.3] implies that,
also for chains,
cp 0.
This is one of the many avatars of one of the most important results of all
Mathematics, called by historical reasons Poincare lemma. We shall meet it
again, under other guises. In the present context, it is not far from intuitive.
Think of the sphere S2 , the boundary of a 3-dimensional ball, or of the torus
T2 which bounds a full-torus, or of any other usual boundary in E3 : they
have themselves no boundary.
64
2.2.2
CHAPTER 2. HOMOLOGY
Topological numbers
There are topological invariants of various types: some are general qualities of
the space, like connectedness and compactness; other are algebraic structures
related to it, as the homotopic groups to be seen in chapter 3; still other are
numbers, like the Euler number and dimension. Whatever they may be,
their common point is that they cannot be changed by homeomorphisms.
Homology provides invariants of two kinds: groups and numbers.
2.2.18 The p-chains satisfying
cp = 0
are called closed p-chains, or p-cycles. The zero p-chain is a trivial p-cycle.
The p-cycles constitute a subgroup of Cp (K), denoted Zp (K).
2.2.19 We have been using integer numbers for the chain coefficients:
mj Z. In fact, any abelian group can be used instead of Z. Besides
the integer homology we have been considering, other cases are of great importance, in particular the real homology with coefficients in R. As we can
also multiply chains by numbers in Z (or R), chains constitute actually a
vector space. To every vector space V corresponds its dual space V formed
by all the linear mappings taking V into Z (or R). V is isomorphic to V
(as long as V has finite dimension) and we can introduce on V constructs
analogous to chains, called cochains. A whole structure dual to homology,
cohomology, can then be defined in a purely algebraic way. This would take
us a bit too far. We shall see later that chains, in the case of differentiable
manifolds, are fundamentally integration domains. They are dual to differential forms and to every property of chains correspond an analogous property
of differential forms. Taking the boundary of a chain will correspond to taking the differential of a form, and Poincare lemma will correspond to the
vanishing of the differential of a differential. Differential forms will have the
role of cochains and we shall leave the study of cohomology to that stage,
restricting the treatment here to a minimum.
p is the sum of all (p+1)-chains
2.2.20 Given a chain cp , its coboundary c
of which cp is an oriented face. Although this is more difficult to see, the
Poincare lemma holds also for :

p 0.
c
Cp (K) Cp+1 (K) is a linear operator. Chains
The coboundary operator :
satisfying
p=0
c
are p-cocycles and also constitute a subgroup of Cp (K).
65
2.2.21 An operator of enormous importance is the laplacian , defined

by
p = ( + )
2 cp .
cp := ( + )c
As takes a p-chain into a (p 1)-chain and takes a p-chain into a (p + 1)chain, takes a p-chain into another p-chain. A p-chain satisfying
cp := 0
preserves the group structure.
is a harmonic p-chain. Just as and ,
2.2.22 Harmonic p-chains are, for finite K, simultaneously p-cycles and pcocycles. They constitute still another subgroup of Cp (K), denoted Bp (K),
the most important of all such groups because it is a topological invariant.
The rank of this group is a topological number, called the p-th Betti number ,
denoted bp .
A
C
D
Figure 2.9: A tetrahedron (left) and a triangulation of the sphere S 2 (right).
2.2.23 Consider the complex formed by the surface of a tetrahedron (Figure 2.9). The vertices A, B, C and D can be used to define the edges AB,
AC, BD, etc. Let us begin by listing and commenting on some results:
(i) A = 0 ; (AB) = B A.
Notice: B is a face of AB, while A is not. Only the oriented vertex (A)
is a face.
= BA + CA + DA (notice the correct signs!);
(ii) A
(iii) 2 (AB) = 0 ; 2 (AB) = 0;
(iv) (AB) = 4AB.
66
CHAPTER 2. HOMOLOGY
The triangle ABD is bounded by T = (ABD) = AB + BD + DA. But

T is also the boundary of another chain: T = (ABC + ACD + BDC).
Thus, a chain can be simultaneously the boundary of two different chains.
All 1-chains are generated by the edges. With coefficients a, b, c, etc in Z (or
R ), a general 1-chain is written
U = aAB + bAC + cAD + dBC + eBD + f CD.
From the result above, U = 4U . As a consequence, U = 0 only if U = 0
and there exists no non-trivial harmonic 1-chain on the tetrahedron. The
dimension of the space of harmonic 1-chains vanishes: b1 (tetrahedron)= 0.
Are there harmonic 0-chains? In order to know it, we start by calculating
=3A B C D, and similarly for the other vertices. We look
A = A
then for a general 0-chain which is harmonic:
(aA + bB + cC + dD) = 0.
This means that
(3a b c d)A + (3b a c d)B + (3c a b d)C
+ (3d a b c)D = 0.
The four coefficients must vanish simultaneously. Cramers rule applied to
these four equations will tell that a simple infinity of solutions exists, which
can be taken to be a = b = c = d. Thus, the chain a(A + B + C + D)
is harmonic for any a in Z (or R ). The dimension of the space (or group)
of harmonic 0-chains is one, b0 (tetrahedron)= 1. Proceeding to examine
2-chains, we start by finding that
(ABD) = 3ABD + BDC + ABC + ACD,
and similarly for the other triangles. Looking for a general harmonic 2-chain
in the form
aBCD + bACD + cABD + dABC,
we find in the same way as for 0-chains that there is a simple infinity of
solutions, so that b2 (tetrahedron)= 1. The tetrahedron is a triangulation
of the sphere S 2 (and of the ellipsoid and other homeomorphic surfaces), as
sketched in Figure 2.9 (right). With some abuse of language, we say that
S 2 is one of the tetrahedrons polyhedra. As a consequence, the same Betti
numbers are valid for the sphere:
b0 (S 2 ) = 1 ; b1 (S 2 ) = 0 ; b2 (S 2 ) = 1.
67
2.2.24 We could think of using finer triangulations, complexes with a

larger number of vertices, edges and triangles. It is a theorem that the Betti
numbers are independent of the particular triangulation used. Notice, however, that not everything is a triangulation: the conditions defining a cellular
complex must be respected. Take for instance the circle S 1 : a triangulation
Figure 2.10: Only the lower two simplexes are real triangulations of the circle.
is a juxtaposition of 1-simplexes joined in such a way that the resulting complex is homeomorphic to it. Now, an edge must have two distinct vertices,
so that simplex (1) in Figure 2.10 is not suitable. The two edges in (2) are
supposed to be distinct but they have two identical vertices. The complexes
(3) and (4) are good triangulations of S 1 . Take the case (3). It is easily seen
that:
= CA + BA; B
= AB + CB;
A
C = BC + AC; (AB) = B A = (CB + AC);

(AB) = 2AB CA BC; etc.
Looking for solutions to (aBC + bCA + cAB) = 0, we arrive at a = b = c,
for any a. There is so a single infinity of solutions and b1 (S 1 ) = 1. In the
same way we find b0 (S 1 ) =1.
2.2.25 Triangulations, as seen, reduce the problem of obtaining the Betti
numbers to algebraic calculations. The examples above are of the simplest
kind, chosen only to illustrate what is done in Algebraic Topology.
2.2.26 Let us again be a bit formal. Taking the boundary of chains induces
a group homomorphism from Cp (K) into Cp1 (K), the boundary homomorphism p : Cp (K) Cp1 (K). The kernel of p ,
68
CHAPTER 2. HOMOLOGY
ker p = {cp Cp (K) such that p cp = 0}
is, of course, the set of p-cycles. We have already said that it constitutes a
group, which we shall denote by Zp (K). Consider p-cycles p , p , p Zp (K),
and (p + 1)-chains p+1 , p+1 Cp+1 (K). If
p = p + p+1 and p = p + p+1 ,
then
p = p + (p+1 + p+1 ) .
Consequently, the relation between p-cycles which differ by a boundary is an
equivalence, and divides Zp (K) into equivalence classes. Each class consists
of a p-cycle and all other p-cycles which differ from it by a boundary. The
equivalence classes can be characterized by those p , p such that no p+1
exists for which p p = p+1 . When such a p+1 does exist, p and p
are said to be homologous to each other. The relation between them is a
homology and the corresponding classes, homology classes. Let us be formal
once more: consider the image of p+1 , the operator acting on (p+1)-chains:
Im p+1 = {those cp which are boundaries of some cp+1 } .
The set of these p-boundaries form still another group, denoted Bp (K). From
the Poincare lemma, Bp (K) Zp (K). Every boundary is a cycle although
not vice-versa. Bp (K) is a subgroup of Zp (K) and there is a quotient group
Hp (K) = Zp (K)/Bp (K) .
2.2.27 This quotient group is precisely the set of homology classes referred
to above. Roughly speaking, it counts how many independent p-cycles exist
which are not boundaries. We have been talking of general complexes, which
can in principle be homeomorphic to some topological spaces. If we restrict
ourselves to finite complexes, which can only be homeomorphic to compact
spaces, then it can be proved that the ranks (number of generators) of all the
groups above are finite numbers. More important still, for finite complexes,
the groups H p (K) are isomorphic to the groups
of harmonic p-cycles .
Consequently,
bp (K) = rank Hp (K) .
69
2.2.28 These Hp (K) are the homology groups of K and, of course, of any
space homeomorphic to K. Let us further state a few important results:
(i) the meaning of b0 : b0 (S) is the number of connected components of S;
(ii) the Poincare duality theorem: in an n-dimensional space S,
bnp (S) = bp (S);
(iii) the Euler-Poincare characteristic is related to the Betti numbers by:
(S) =
Pn
j=0 bj (S).
This expression is sometimes used as a definition of (S). Notice that for

the circle, (S 1 ) = 0. For the sake of completeness: also ker p constitutes a
group, the group Z p of p-cocycles. It contains the subgroup Im p1 of those
p-cocycles which are coboundaries of (p 1)-chains. The quotient group
H p = ker p /Im p1
is the p-th cohomology group of S. For finite complexes or compact spaces,
it is isomorphic to the homology group Hp .
2.2.29 Figure 2.11 shows a torus T 2 and a possible triangulation. The
torus is obtained from the complex simply by identifying the vertices and
edges with the same names. After such identification is done, it is easy to
see that:
(i) the simplex (e1 + e2 + e3 ) is a cycle, but not a boundary;
(ii) the simplex (e4 + e5 + e6 ) is a cycle, quite independent of the previous
one, but which is also not a boundary;
(iii) there are many cycles which are boundaries, such as (e4 +e1 e7 e8 ).
Each cycle-not-boundary can give chains which are n-times itself: we can go
along them n times. It can be shown, using algebraic manipulations analogous (though lengthier) to those used above for the circle and the tetrahedron,
that there are two independent families of such 1-cycles which are not boundaries, so that H1 (T 2 ) is isomorphic to Z Z. The Betti number b1 (T 2 ) = 2.
Going on with an intuitive reasoning, we could ask about the 2-cycles. The
torus itself is one such, as are the chains obtained by covering it n times.
Such cycles are not boundaries and there are no other cases at sight. We
would guess (correctly) that b2 (T 2 ) = 1, which would also come from the
meaning of b0 plus Poincare duality. Calculations of course are necessary
to confirm such results, but they are simple (though tedious) adaptations of
what was done for the tetrahedron. The Poincare duality, of course, reduce
70
CHAPTER 2. HOMOLOGY
Figure 2.11: A possible triangulation of T 2 .
the calculations to (about) one half. The Euler characteristic may be found
either by counting in the triangulation
X
()i Ni = 9 27 + 18 = 0,
i
or from
X
()i bi = 1 2 + 1 = 0.
It is a general result for toruses that (T n ) = 0 for any n.

2.2.30 Genus : one of the oldest topological invariants. Denoted g, it is
the largest number of closed non-intersecting continuous curves which can
be drawn on a surface without separating it into distinct domains. It is zero
for S 2 , one for the torus, two for the double torus, etc. One may always
think of any -dimensional connected, compact and orientable manifold as
consisting of a sphere with n handles, for some value of n, including n = 0.
In this case, g = n. In general, the genus counts the number of toruses of
the surface. It is also half the first Betti number, 2g = b1 .
2.2.31 We have considered a homeomorphism h : K S between a
complex and a space (or between polyhedra) to establish a complete identity
of all homology groups. We might ask what happens if h were instead only
a continuous function. The answer is the following. A continuous function
f : P P 0 between two polyhedra induces homomorphisms
71
fk : Hk (P ) Hk (P 0 )
between the corresponding homology groups. Such homomorphisms become
isomorphisms when f is a homeomorphism.
The homology group H1 is the abelianized subgroup of the more informative fundamental homotopy group 1 , to be examined in section 3.1.2.
2.2.32 Once the Betti numbers for a space S are found, they may be put
together as coefficients in the Poincare polynomial
PS (t) =
Pn
j
j=0 bj (S) t
in which t is a dummy variable. In the example of the sphere S 2 , PS 2 (t) =

1 + t2 . For the torus T 2 , PT 2 (t) = 1 + 2t + t2 . And so on. Of course,
(S) = PS (1). There is more than mere bookkeeping here. Polynomials
which do not change under transformations are called invariant polynomials and are largely used, for example, in knot classification. Of course,
distinct polynomials are related to spaces not equivalent under the transformations of interest. Poincare polynomials are invariant under homeomorphic
transformations.
2.2.33 Latticing a space provides a psychological frame, helping to grasp
its general profile. We have seen how it gives a clue to its homological and homotopical properties. In Physics, lattices provide the framework convenient
to treat solid media but, more than that, they give working models for real
systems. In the limit of very small spacing, they lead to a description of continuum media. But lattices are regular patterns, only convenient to modeling
well-ordered systems such as crystals and metallic solids. Introducing defects
into lattices leads to the description of amorphous media. We are thus led
to examine (static) continuum media, elastic or not. Starting from crystals,
the addition of defects allows a first glimpse into the qualitative structure of
glasses. The very word tensor rings of Elasticity Theory, where in effect
it had its origin. Deformed crystals provide the most material examples
of tensor fields such as curvature and torsion (another resounding name).
Physical situations appear usually in 3 dimensions, but 2 dimensional cases
provide convenient modeling and there is nowadays a growing interest even
in 2 dimensional physical cases, both for physical surfaces and biological
membranes (see Phys.3).
2.2.34 Homology is an incredibly vast subject and we have to stop somewhere. It has been intermittently applied in Physics, as in the analysis of
72
CHAPTER 2. HOMOLOGY
Feynman diagrams1 or in the incredibly beautiful version of General Relativity found by Regge.2 Much more appears when differential forms are
at work and we shall see some of it in chapter 7. Let us finish with a few
words on cubic homology. In many applications it is easier to use squares
and cubes instead of the triangles and tetrahedra we have being employing.
Mathematicians are used to calculate homotopic properties (next section is
devoted to applying complexes to obtain the fundamental group) in cubic
complexes. For physicists they are still more important, as cubic homology
is largely used in the lattice approach to gauge theories.3 Now: cubes cannot,
actually, be used as simplexes in the simple way presented above. The resulting topological numbers are different: for instance, a space with a single
point would have all of the bj = 1. It is possible, however, to proceed to
a transformation on the complexes (called normalization), in such a way
that the theory becomes equivalent to the simplex homology.4 Triangulations
are more fundamental because (i) any compact space (think on the square)
can be triangulated, but not every compact space can be squared (think on
the triangle), and (ii) the topological numbers coming up from triangulations
coincide with those obtained from differential forms. In a rather permissive
language, it is the triangles and tetrahedra which have the correct continuum
limits.
Following a proposal by M. Froissart. See Hwa & Teplitz 1966.

Regge 1961; on the subject (called Regge Calculus) see Misner, Thorne & Wheeler
1973, 42.
3
Becher & Joos 1982; the appendix contains a resume of cubic homology.
4
See Hilton & Wylie l967.
2
Chapter 3
HOMOTOPY
3.0
GENERAL HOMOTOPY
We have said that, intuitively, two topological spaces are equivalent if one can
be continuously deformed into the other. Instead of the complete equivalence
given by homeomorphisms in general difficult to uncover we can more
modestly look for some special deformations preserving only a part of the
topological characteristics. We shall in this section examine one-parameter
continuous deformations. Roughly speaking, homotopies are function deformations regulated by a continuous parameter, which may eventually be
translated into space deformations. The topological characterization thus
obtained, though far from complete, is highly significant.
In this section we shall state many results without any proof. This is
not to be argued as evidence for a neurotic dogmatic trend in our psychism.
In reality, some of them are intuitive and the proofs are analogous to those
given later in the particular case of homotopy between paths.
3.0.1 Homotopy between functions Let f and g be two continuous
functions between the topological spaces X and Y , f, g : X Y . Let again
I designate the interval [0, 1] included in E1 . Then
f is homotopic to g (f g) if there exists a continuous function
F : X I Y such that F (p, 0) = f (p) and F (p, 1) = g(p) for every
p X. The function F is a homotopy between f and g.
Two functions are homotopic when they can be continuously deformed into
each other, in such a way that the intermediate deformations constitute a
73
74
CHAPTER 3. HOMOTOPY
family of continuous functions between the same spaces: F (p, t) is a oneparameter family of continuous functions interpolating between f and g. For
fixed p, it gives a curve linking their values.
3.0.2 Two constant functions f, g : X Y with f (p) = a and g(q) = b for
all p, q X are homotopic if Y is path-connected. In this case, a and b can
be linked by a path : I Y , with (0) = a and (1) = b. Consequently,
F (p, t) = (t) for all (p, t) X I is a homotopy between f and g (Figure
3.1).
Figure 3.1: On the left, the cartesian product X I; on the right, the space
Y where the images of f and g lie and, for each value of t, of the mediating
function F (p, t).
3.0.3 We have mentioned in 1.3.14 that the function set Y X can be

made into a topological space, for instance via the compact-open topology.
We can then define paths on Y X . Given two points in this space, like the
above functions f and g, a homotopy F is precisely a continuous path on Y X
connecting them. As a consequence, two homotopic functions lie necessarily
on the same path-component of Y X .
3.0.4 Homotopy is an equivalence relation between continuous functions,
which consequently decomposes the set Y X of continuous functions from X
to Y into disconnected subsets, the equivalence classes. These classes are the
homotopy classes, and the class to which a function f belongs is denoted by
[f ].
The homotopy classes correspond precisely to the path-components of
Y X . The set of all classes of functions between X and Y is denoted by {X, Y }.
Figures 3.9 and 3.10 below illustrate the decomposition in the particular case
of curves on X.
3.0. GENERAL HOMOTOPY
75
3.0.5 Composition preserves homotopy: if f, g : X Y are homotopic

and h, j : Y Z are homotopic, then h f j g h g j f .
Consequently, composition does not individualize the members of a homotopy
class: it is an operation between classes,
[f g] = [f ] [g].
3.0.6 Homotopy between spaces The notion of homotopy may be
used to establish an equivalence between spaces. Given any space Z, let
idZ : Z Z be the identity mapping on Z, idZ (p) = p for every p in Z. A
continuous function f : X Y is a homotopic equivalence between X and
Y if there exists a continuous function g : Y X such that g f idX
and f g idY . The function g is a kind of homotopic inverse to f .
When such a homotopic equivalence exists, X and Y are said to be of the
same homotopy type. Notice that we would have a homeomorphism if above,
instead of g f idX and f g idY , we had g f = idX and f g = idY .
Homotopy is a necessary condition for topological equivalence, though not
a sufficient one. Every homeomorphism is a homotopic equivalence but not
every homotopic equivalence is a homeomorphism. This means that homeomorphic spaces will have identical properties in what concerns homotopy
(which is consequently a purely topological characteristic) but two spaces
with the same homotopical properties are not necessarily homeomorphic
they may have other topological characteristics which are quite different. It
will be seen in examples below that even spaces of different dimension can
be homotopically equivalent.
3.0.7 Contractibility is the first homotopic quality we shall meet. Suppose that the identity mapping idX is homotopic to a constant function.
Putting it more precisely, there must be a continuous function h : X I X
and a constant function f : X X, f (p) = c (a fixed point) for all p X,
such that h(p, 0) = p =idX (p) and h(p, 1) = f (p) = c. When this happens,
the space is homotopically equivalent to a point and said to be contractible
(Figure 3.2).
The identity mapping idX simply leaves X as it is, while f concentrates
X into one of its points. Contractibility means consequently that to leave X
alone is homotopically equivalent to letting it shrink to a point. The interval
[0, 2) with the induced topology (see 1.3.12) is contractible, but the circle
S 1 is not. With the quotient topology of 1.4.3 the same interval [0, 2) is
no more contractible it becomes equivalent to S 1 .
76
CHAPTER 3. HOMOTOPY
Figure 3.2: There is a homotopy between X and a point: X is contractible.
3.0.8 A special, important result: every vector space is contractible. Let

us see a special example. Take X = En and Y = {0} (see Figure 3.3).
Let f be the constant function f : En {0}, f (x) = 0 x En . Let
g : {0} En be the canonical injection of {0} into En , g(0) = 0. Then,
(f g)(0) = f [g(0)] = f (0) = idY (0).
As 0 is the only point in Y , we have shown that f g = idY , and so, that
f g idY . Also
(g f )(x) = g[f (x)] = g(0) = 0.
Now, let h : En I En , h(x, t) = tx En (because of its vector space
structure). Clearly h(x, 0) = 0 = (g f )(x) and h(x, 1) = x = idX (x). So,
h is a homotopy idX g f . The space En is consequently homotopic to a
point, that is, contractible. The same is true of any open ball of En .
3.0.9 Take X = E2 {0} and Y = S 1 (as in Figure 3.4). Let f : X Y ,
f (x) = x/|x|, and let g : S 1 E2 - {0} be the canonical injection g : ei
S 1 ei E2 {0}. Then,
(f g)(ei ) = f (ei ) = ei = idS 1 (ei ).
On the other hand,
(g f )(x) = g(x/|x|) = x/|x|.
Now, h : X I X given by
h(x, t) = (1 t)x/|x| + tx
is such that h(x, 0) = (g f )(x) and h(x, 1) = x = idX (x). The conclusion
is that E2 {0} and S 1 are homotopically equivalent.
3.0. GENERAL HOMOTOPY
77
Figure 3.3: (left) The plane is homotopically equivalent to a point (here, the
origin), that is, it is contractible.
Figure 3.4: (right) The punctured plane E2 /{0} is homotopically equivalent
to the circle.
This example is of significance in the study of the Aharonov-Bohm effect ( 4.2.18). It

has higher dimensional analogues. The following result is relevant to the Dirac monopole:
the space E4 \{points on the line x1 = x2 = x3 = 0} is homotopically equivalent to the
sphere S 2 . It can also be shown that E3 \{points on an infinite line} S 1 . These, and the
examples of 3.0.8 and 3.0.9, are cases of homotopical equivalences that are not complete
topological equivalences, as the spaces involved do not even have the same dimensions.
Notice also the relationship with convexity in the last two examples.
3.0.10 If either X or Y is contractible, every continuous function f :

X Y is homotopic to a constant mapping (use f = idY f or f =
f idX at convenience). As said above, every vector space is contractible.
Contractibility has important consequences in vector analysis. As we shall
see later, some of the frequently used properties of vector analysis are valid
only on contractible spaces for example, the facts that divergenceless fluxes
are rotationals and irrotational fluxes are potential and will not hold if,
for example, points are extracted from a vector space.
78
CHAPTER 3. HOMOTOPY
3.0.11 We have repeatedly said that homotopies preserve part of the topological aspects. Contractibility, when extant, is one of them. We shall in the
following examine some other, of special relevance because they appear as
algebraic structures, the homotopy groups. The homology groups presented
in chapter 2 are also invariant under homotopic transformations.
3.0.12 General references Two excellent and readable books on homotopy are: Hilton 1953 and Hu 1959. A book emphasizing the geometrical
approach, also easily readable, is Croom 1978.
3.1
PATH HOMOTOPY
Paths defined on a space provide essential information on its topology.
3.1.1
Homotopy of curves
3.1.1 Let us recall the definition of a path on a topological space X: it

is a continuous function f : I X. This is not the kind of path which will
be most useful for us. Notice that I is contractible, so that every path is
homotopic to a constant and, in a sense, trivial. We shall rather consider,
instead of such free-ended paths, paths with fixed ends. The value x0 = f (0)
is the initial end-point, x1 = f (1) is the final end-point and f (t) is the path
from x0 to x1 .
Given two paths f and g with the same end-points, they are homotopic
paths (which will be indicated f g) if there exists a continuous mapping
F : I I X such that, for every s, t I,
F (s, 0) = f (s); F (s, 1) = g(s);
F (0, t) = x0 ; F (1, t) = x1 .
3.1.2 The function F is a path homotopy between f and g and represents,
intuitively, a continuous deformation of the curve f into the curve g. For
each fixed t, F (s, t) is a curve from x0 to x1 , intermediate between f and
g, an interpolation between them (Figure 3.5). For each fixed s, F (s, t) is
a curve from f (s) to g(s) (Figure 3.6). These transversal curves are called
variations (see Math.7).
3.1.3 Path homotopy is an equivalence relation:
(i) given f , trivially f , f as the mapping F (p, t) = f (p) is a homotopy;
by the way, it is the identity function on X I ;
3.1. PATH HOMOTOPY
79
Figure 3.5: For each value of t in I, F (s, t) gives a path, intermediate between
f and g.
Figure 3.6: For each fixed value of s, F (s, t) gives a path going from f (s) to
g(s).
(ii) suppose F is a homotopy between f and g; then,

G(p, t) = F (p, 1 t) is a homotopy between g and f ;
(iii) suppose F is a homotopy f g and G a homotopy g h (Figure
3.7); define H : I I X by the equations

F (s, 2t)
for t [0, 21 ]
H(s, t) =
G(s, 2t 1) for t [ 12 , 1]
H is well defined as, when t = 1/2,
[F (s, 2t)]t=1/2 = F (s, 1) = g(s) = G(s, 0) = [G(s, 2t 1)]t=1/2 .
As H is continuous on the two closed subsets I [0, 1/2] and I [1/2, 1] of
80
CHAPTER 3. HOMOTOPY
Figure 3.7: If F is a homotopy f g and G is a homotopy g h, then there

exists also a homotopy H : f h.
I I, it is continuous on I I. It is a homotopy f h.
3.1.4 By definition, every pair of points inside each path-component of a
given space X can be linked by some path. Let {C } be the set of pathcomponents of X, which is denoted by 0 (X):
0 (X) = {C }.
This notation is a matter of convenience: 0 will be included later in the
family of homotopy groups and each one of them is indicated by n for some
integer n. The relation between the path homotopy classes on a space X and
the path components of the space X I of all paths on X is not difficult to
understand (Figure 3.8). A homotopy F between two paths on X is a path
on X I and two homotopic paths f and g can be thought of as two points of
X I linked by the homotopy F .
3.1.5 The intermediate curves representing the successive stages of the
continuous deformation must, of course, lie entirely on the space X. Suppose
X to have a hole as in Figure 3.9. Paths f and g are homotopic to each other
and so are j and h. But f is homotopic neither to j nor to h. The space X I
is so divided in path-components, which are distinct for f and j.
There are actually infinite components for X I when X has a hole: a
curve which turns once around the hole can only be homotopic to curves
turning once around the hole; a curve which turns twice around the hole
will be continuously deformable only into curves turning twice around the
hole; and the same will be true for curves turning n times around the hole,
suggesting a relation between the homotopy classes and the counting of the
number of turns the curves perform around the hole. Another point is that
3.1. PATH HOMOTOPY
81
Figure 3.8: Relation between path homotopy classes on a space X and the
path components of the space X I of all paths on X.
Figure 3.9: Effect of the presence of a hole.
a curve turning in clockwise sense cannot be homotopic to a curve turning

anticlockwise even if they turn the same number of times (like the curves k
and r in Figure 3.10). This suggests an algebraic counting of the turns. We
shall in what follows give special emphasis to some algebraic characteristics
of paths and proceed to classify them into groups. An operation of path
composition is defined which, once restricted to classes of closed paths, brings
forth such a structure.
3.1.6 Towards a group structure Let us start by introducing the announced operation between paths. Take again the space X and f a path
between x0 and x1 . Let g be a path between x1 and x2 . We define the
82
CHAPTER 3. HOMOTOPY
Figure 3.10: Effect of curve orientations.
composition, or product f g of f and g as that path (Figure 3.11) given by

f (2s)
for s [0, 12 ]
h(s) = (f g)(s) =
g(2s 1) for s [ 12 , 1].
The composition h is a well defined curve from x0 to x2 . It can be seen as
Figure 3.11: Path composition.

a path with first half f and second half g. The order in reading these
products is from left to right, opposite to the usual composition of functions.
3.1.7 We shall show below that the operation is well defined on the
path-homotopy classes. It has the following group-like properties:
(i) associativity: if [f ] ([g] [h]) and ([f ] [g]) [h] are defined, then
[f ] ([g] [h]) = ([f ] [g]) [h] .
3.1. PATH HOMOTOPY
83
Thus, path-homotopy classes with the operation of composition constitute a

semigroup.
(ii) there are identity elements: for any x X, let ex : I X be such
that ex (I) = x; thus, ex is a constant path on X, with in particular both
end-points equal to x. Given a path f between x0 and x1 , we have
[f ] [ex1 ] = [f ] ;
[ex0 ] [f ] = [f ].
Notice however that the identity elements are different in different points
(consequently, path-homotopy classes with the operation of composition constitute neither a monoid nor a groupoid see Math.1).
(iii) existence of inverse elements: the inverse of a path f from x0 to
x1 is a path f <1> from x1 to x0 given by f <1> (s) = f (1 s); thus,
[f ] [f <1> ] = [ex0 ];
[f <1> ] [f ] = [ex1 ] .
3.1.8 Before we start showing that operation is well defined on homotopy
classes, as well as that it is an associative product, let us repeat that, despite a certain similitude, the above properties do not actually define a group
structure on the set of classes. The identity elements are distinct in different points, and the product [f ] [g] is not defined for any pair of equivalence
classes: for that, f (1) = g(0) would be required. The set of path-homotopy
classes on X is not a group with the operation . We shall see in the next
section that the classes involving only closed paths do constitute a group,
which will be called the fundamental group.
3.1.9 In order to show that is well defined on the homotopy classes,
we must establish that, if f f 0 and g g 0 , then f g f 0 g 0 . Given
the first two homotopies, we shall exhibit a homotopy between the latter.
Supposing F and G to be path-homotopies respectively between f, f 0 and
g, g 0 , a homotopy between f g and f 0 g 0 is (see the scheme of Figure 3.12)

F (2s, t)
for s [0, 21 ]
H(s, t) =
G(2s 1, t) for s [ 12 , 1].
For s = 1/2, F (1, t) = x1 = G(0, t) for any t. The function H is, thus, well
defined and continuous. On the other hand,

f (s)
for s [0, 21 ]
H(s, 0) =
g(2s 1) for s [ 12 , 1].
84
CHAPTER 3. HOMOTOPY
defines f g, and

H(s, 1) =
f 0 (s)
for s [0, 21 ]
g 0 (2s 1) for s [ 12 , 1].
defines f 0 g 0 . As H(0, t) = x0 and H(1, t) = x2 , the mapping H is indeed a

path-homotopy between f g and f 0 g 0 . To obtain the associativity property,
Figure 3.12: Composition of homotopy classes, which leads to a group structure.

we should show that
f (g h) (f g) h.
Through the mapping f (g h), the image of s goes through the values of f
when s goes from 0 to 1/2, the values of g when s lies between 1/2 and 3/4,
and the values of h when s goes from 3/4 to 1:

f (2s)
for s [0, 12 ]
(f (g h))(s) =
(g h)(2s 1) for s [ 12 , 1].
In more detail,
(f (g h))(s) =
f (2s)
for s [0, 12 ]
g[2(2s 1)] for s [ 12 , 34 ]

h[2(2s 1)] for s [ 34 , 1].
On the other hand,

((f g) h)(s) =
(f g)(2s) for s [0, 12 ]
h(2s 1)
for s [ 12 , 1].
3.1. PATH HOMOTOPY

or
((f g) h)(s) =
85
f (4s)
for s [0, 14 ]
g(4s 1) for s [ 14 , 12 ]
h(2s 1) for s [ 12 , 1].
Through the mapping (f g) h the image of s goes through the values of f

when s goes from 0 to 1/4, the values of g when s lies between 1/4 and 1/2,
and the values of h when s goes from 1/2 to 1. The paths f (g h) and
(f g) h have the same image, traversed nevertheless at different rates.
The desired homotopy is formally given by
4s
f ( t+1
)
for s [0, t+1
]
, t+2
]
g(4s t 1) for s [ t+1
F (s, t) =
4
4
h( 4st2
)
for s [ t+2
, 1].
2t
4
The arguments of f, g and h are all in [0, 1]. It is easily checked that
F (s, 1) is the same as (f (g h))(s), and that F (s, 0) is the same as ((f
g) h)(s).
3.1.2
The Fundamental group
Classes of closed curves constitute a discrete group, which is one of the main
topological characteristics of a space. Here we shall examine the general
properties of this group. Important examples of discrete groups, in particular
the braid and knot groups, are described in Math.2. Next section will be
devoted to a method to obtain the fundamental group for some simple spaces
as word groups.
3.1.10 We have seen in the previous section that the path product
does not define a group structure on the whole set of homotopy classes. We
shall now restrict that set so as to obtain a group the fundamental group
which turns out to be a topological characteristic of the underlying space.
3.1.11 Let X be a topological space and x0 a point in X. A path whose
both end-points are x0 is called a loop with base point x0 . It is a continuous
function f : I X, with f (0) = f (1) = x0 .
3.1.12 The set of homotopy classes of loops with a fixed base point constitutes a group with the operation . More precisely:
86
CHAPTER 3. HOMOTOPY
Figure 3.13: Loops with a fixed base point.
(i) given two loops f and g with the same base point x0 , their composition
f g is well defined and is a loop with base point x0 (Figure 3.13);
(ii) the properties of associativity, existence of an identity element [ex0 ],
and existence of an inverse [f <1> ] for each [f ] hold evidently .
3.1.13 The group formed by the homotopy classes of loops with base point
x0 is called the fundamental group of space X relative to the base point x0
and is denoted 1 (X, x0 ). It is also known as the Poincare group (a name
not used by physicists, to avoid confusion) and first homotopy group of X at
x0 , because of the whole series of groups n (X, x0 ) which will be introduced
in section 3.3.
Figure 3.14: Changing from one base point to another.
3.1.14 A question arising naturally from the above definition is the following: how much does the group depend on the base point? Before we answer
3.1. PATH HOMOTOPY
87
to this question, some preliminaries are needed. Let be a path on X from

x0 to x1 (Figure 3.14). Define the mapping
# : 1 (X, x0 ) 1 (X, x1 )
through the expression
# ([f ]) = [ 1 ] [f ] [].
This is a well defined mapping. If f is a loop with base point x0 , then
1 ] f
is a loop with base point x1 . Consequently, # maps indeed 1 (X, x0 ) into
1 (X, x1 ). Finally, we have the following theorem:
the mapping # is a group isomorphism.
Let us go by parts: a map of a group G with operation () into a
groupG0 with operation () is a group homomorphism if, for any a and b
in G, (a b) = (a) (b). In this case, is also called a representation
(see Math.6) of G in G0 . The map will be a group isomorphism 1 if a
homomorphism : G0 G exists such that the compositions and
are the identity mappings of G0 and G respectively. In order to prove the
theorem, we have first to show that # is a homomorphism, that is, that it
preserves the operation . But
( # ([f ])) ( # ([g])) = ([ 1 ] [f ] []) ([ 1 ] [g] [])
= ([ 1 ] [f ] [g] []) = # ([f ] [g]).
So good for the homomorphism. Now, taking the inverse path 1 , let
us see that ( 1 )# is the inverse to # . Take [f ] in 1 (X, x0 ) and [h] in
1 (X, x1 ). Then,
( 1 )# [h] = [] [h] [ 1 ];
# (( 1 )# [h]) = [ 1 ] ([] [h] [ 1 ]) [],
and finally
# (( 1 )# [h]) = [h].
It is similarly verified that ( 1 )# ( # [f ]) = [f ] .
3.1.15 Wherever X is path-connected, there exist such mappings # relating the fundamental group based on different points. So, a corollary is:
1
Fraleigh 1974.
88
CHAPTER 3. HOMOTOPY
if X is path-connected and both x0 and x1 are in X, then
1 (X, x0 ) is isomorphic to 1 (X, x1 ) .
3.1.16 Suppose now that X is a topological space and C is a path-component of X to which x0 belongs. As all the loops at x0 belong to C and
all the groups are isomorphic, we may write 1 (C) = 1 (X, x0 ). The group
depends only on the path-component and gives no information at all about
the remaining of X. In reality, we should be a bit more careful, as the
isomorphism # above is not natural (or canonical ): it depends on the path
. This means that different paths between x0 and x1 take one same element
of 1 (X, x0 ) into different elements of 1 (X, x1 ), although preserving the
overall abstract group structure. The isomorphism is canonical only when
the group is abelian. We shall be using this terminology without much ado
in the following, though only in the abelian case is writing 1 (C) entirely
justified. These considerations extend to the whole space X if it is pathconnected, when we then talk of the group 1 (X).
Figure 3.15:
3.1.17 A class [f ] contains its representative f and every other loop continuously deformable into f . Every space will contain a class [c(t)] of trivial,
or constant, closed curves, c(t) = x0 for all values of t(Figure 3.15). If this
class is the only homotopy class on X, 1 (X, x0 ) = [c(t)], we say that X is
simply-connected. More precisely,
X is simply-connected when it is path-connected
and 1 (X, x0 ) is trivial for some x0 X.
3.1. PATH HOMOTOPY
89
Figure 3.16:
3.1.18 When X is simply-connected, every two loops on X can be continuously deformed into each other, and all loops to a point. When this is
not the case, X is multiply-connected. This is a first evidence of how 1 (X)
reflects, to a certain degree, the qualitative structure of X. The fact that
closed laces in a multiply-connected space cannot be caused to shrink to a
point while remaining inside the space signals the presence of some defect,
of a hole of some kind. The plane E2 is clearly simply-connected, but this
changes completely if one of its points is excluded, as in E2 {0}: loops not
encircling the zero are trivial but those turning around it are not (Figure
3.16). There is a simple countable infinity of loops not deformable into each
other. We shall see later (on 3.1.33) that 1 (E2 \{0}) = Z, the group of
integer numbers. We shall also see ( 3.1.34) that the delection of two points,
say as in E2 \{ 1, +1}, leads to a non-abelian group 1 . The space E3 with
an infinite line deleted (say, x = 0, y = 0) is also a clear non-trivial case.
Some thought dedicated to the torus T 2 = S 1 S 1 will, however, convince
the reader that such intuitive insights are yet too rough.
3.1.19 The trivial homotopy class [c(t)] acts as the identity element of
1 and is sometimes denoted [ex0 ]. More frequently, faithful to the common
usage of the algebrists, it is denoted by 0 when 1 is abelian, and 1 when
it is not (or when we do not know or care).
3.1.20 All the euclidean spaces En are simply-connected: 1 (En ) = 1. The
n-dimensional spheres S n for n 2 are simply-connected: 1 (S n ) = 1 for
90
CHAPTER 3. HOMOTOPY
n 2. We shall see below that the circle S 1 is multiply-connected and that

1 (S 1 ) = Z.
3.1.21 Let us see now why the fundamental group is a topological invariant. The root of this property lies in the fact that continuous functions between spaces induce homomorphisms between their fundamental groups, even
when they are not homeomorphisms (Figure 3.17). Indeed, let : X Y
be a continuous
Figure 3.17: A continuous mapping between spaces X and Y induces a homomorphism between their fundamental groups.
mapping with (x0 ) = y0 . If f is a loop on X with base point x0 , then the

composition f : I Y is a loop on Y with base point y0 . Consequently
induces a mapping between 1 (X, x0 ) and 1 (Y, y0 ), defined by
: 1 (X, x0 ) 1 (Y, y0 )
([f ]) = [ f ] .
This mapping is the induced homomorphism, relative to the base point
x0 . It can be shown that it is well defined: if f and f 0 are homotopic and
F : I I X is their homotopy, then F is a homotopy between the loops
f and f 0 . It can also be shown that is a homomorphism,
([f ] [g]) = ([f ]) ([g]) .
Notice that depends not only on but also on the base point. This homomorphism has some properties of great importance, known in the mathematical literature under the name functorial properties. We list them:
3.1. PATH HOMOTOPY
91
(i) if : (X, x0 ) (Y, y0 ), and : (Y, y0 ) (Z, z0 ), then ( ) =

. If i : (X, x0 ) (X, x0 ) is the identity mapping, then i is the identity
homomorphism;
(ii) if : (X, x0 ) (Y, y0 ) is a homeomorphism, then is an isomorphism between 1 (X, x0 ) and 1 (Y, y0 ). This is the real characterization of
the fundamental group as a topological invariant;
(iii) if (X, x0 ) and (Y, y0 ) are homotopically equivalent, then 1 (X, x0 ) and
1 (Y, y0 ) are isomorphic. Recall that two spaces are homotopically equivalent (or of the same homotopic type) when two functions j : X Y and h:
Y X exist satisfying j h idX and h j idY . Here, these functions
must further satisfy j(x0 ) = y0 and h(y0 ) = x0 .
Summing up: continuous functions induce homomorphisms between the
fundamental groups; homeomorphisms induce isomorphisms.
3.1.22 A contractible space is clearly simply-connected. The converse,
however, is not true. We shall show later that the sphere S 2 is simplyconnected, but it is clearly not contractible. The vector analysis properties
mentioned in 3.0.10 (divergenceless rotational, irrotational gradient)
require real contractibility. They are consequently not valid on S 2 , which is
simply-connected.
3.1.23 Before we finish this section let us mention an important and useful
result. With the same notations above, let X and Y be topological spaces
and consider the fundamental groups with base points x0 and y0 . Then, the
fundamental group of (X Y, x0 y0 ) is isomorphic to the direct product of
the fundamental groups of (X, x0 ) and (Y, y0 ):
1 (X Y, x0 y0 ) 1 (X, x0 ) 1 (Y, y0 ) .
3.1.24 As mentioned above, 1 (S 1 ) = Z, the additive group of the integer
numbers. This will be shown in section 3.2.2 through the use of covering
spaces. Although it does provide more insight on the meaning of the fundamental group, that method is rather clumsy. Another technique, much
simpler, will be seen in section 3.1.3. As an illustration of the last result
quoted above, the torus T 2 = S 1 S 1 will have 1 (T 2 ) = Z Z.
3.1.25 For the reader who may be wondering about possible relations between homology and homotopy groups, let it be said that H1 is the abelianized subgroup of the fundamental group. It contains consequently less information than 1 .
92
CHAPTER 3. HOMOTOPY
3.1.26 The fundamental group, trivial in simply-connected spaces, may be

very complicated on multiply-connected ones, as in the last examples. Quantum Mechanics on multiply-connected spaces2 makes explicit use of 1 (for a
simple example, the Young double-slit experiment, see 4.2.17). Feynmans
picture, based on trajectories, is of course a very convenient formulation to
look at the question from this point of view. The total propagator from
a point A to a point B becomes a sum of contributions3 of the different
homotopy classes,
P
K(B, A) = ([])K (B, A) .
K is the propagator for trajectories in the class [], and is a phase providing a one-dimensional representation of 1 (configuration space). It would be
interesting to examine the case of many-component wavefunctions, for which
such representation could perhaps become less trivial and even give some not
yet known effects for non-abelian 1 .
3.1.3
Some Calculations
3.1.27 Triangulations provide a simple means of obtaining the fundamental group. The method is based on a theorem4 whose statement requires
some preliminaries.
(i) Take a path-connected complex with vertices v1 , v2 , v3 , . . . , and edges
(v1 v2 ), (v1 v3 ), (v2 v3 ), etc. To each oriented edge we attribute a symbol,
(vi vk ) gik .
1
We then make the set of such symbols into a group, defining gik
= gki and
imposing, for each triangle (vi vj vk ), the rule
gij gjk gki = 1,

the identity element. Roughly speaking, the element gik corresponds to going along the edge (vi vk ), the product gij gjk to going along (vi vj ) and then
along (vj vk ), and so on. The rule would say that going around a triangle,
one simply comes back to the original point. We emphasize that one such
rule must be imposed for each triangle (2-simplex).
(ii) It can be proven that on every path-connected polyhedron there exists
at least one Euler path, a path through all the vertices which is contractible
(that is, it is simple but not a loop).
2
3
4
Schulman 1968.
DeWitt-Morette 1969; Laidlaw & DeWitt-Morette 1971; DeWitt-Morette 1972.
Hu 1959; see also Nash & Sen 1983.
3.1. PATH HOMOTOPY
93
3.1.28 Once this is said, we enunciate the calculating theorem:

Take a vertex v0 in a polyhedron K and a Euler path P starting from v0 ;
in the group defined above, put further gjk = 1 for each edge (vj vk )
belonging to P. Then, the remaining group G is isomorphic to the
fundamental group of K with base point v0 , 1 (K, v0 ).
The proof is involved and we shall not even sketch it here. The theorem
is a golden road to arrive at the fundamental group. The best way to see
how it works is to examine some examples.
Figure 3.18: A tetrahedron, and a connected path.
3.1.29 Take again the (surface) tetrahedron as in Figure 3.18 and, for P ,
the connected path 1 2 3 4. For each triangle there is a condition:
g12 g24 g41 = 1; g41 g13 g34 = 1 ; g23 g34 g42 = 1; g12 g23 g31 = 1.
We now impose also g12 = g23 = g34 = 1, because the corresponding edges
belong to the chosen path. As a consequence, all the group elements reduce
to the identity. The fundamental group is then 1 (tetrahedron) = {1}. The
tetrahedron is simply-connected, and so are the spaces homeomorphic to it:
the sphere S 2 , the ellipsoid, etc.
3.1.30 The disc in E2 , that is, the circle S 1 and the region it circumscribes
(Figure 3.19); it is perhaps the simplest of all cases. The triangulation in the
figure makes it evident that
94
CHAPTER 3. HOMOTOPY
Figure 3.19:
Figure 3.20:
g12 = g23 = g41 = 1,

because of the chosen path. The two conditions then reduce to
g24 = g34 = 1.
Consequently 1 = {1}.
3.1.31 The M
obius band: with the triangulation and path given in Figure 3.20, it becomes quickly obvious that g13 = g24 = g35 = g46 = 1, and that
g62 = g52 = g16 = some independent element g. This means that we have the
discrete infinite group with one generator, which is Z: 1 = Z. Consequently,
1 does not distinguish the Mobius band from the cylinder.
3.1.32 The torus. With the triangulation and path of Figure 3.21, the
18 conditions are easily reduced to the forms
g49 = g89 = g78 = g68 = g36 = g26 = 1 ; g27 = g39 = g29 = g17 = g35 =: g 0 ;
g24 = g64 = g14 = g65 = g75 = g13 = g18 =: g 00 ; g13 g35 g51 = 1 ; g75 g51 g17 = 1.
The last two conditions can be worked out to give
3.1. PATH HOMOTOPY
95
Figure 3.21: A torus triangulation, and a connected path.
g51 = g 0 g 00 = g 00 g 0 ,
so that g 0 and g 00 commute. The group has two independent generators
commuting with each other. Consequently, 1 (T 2 ) = Z Z.
Figure 3.22: A triangulating for the punctured disk.
3.1.33 The disc with one hollow. A triangulating complex for the oncepunctured disk is given in Figure 3.22, as well as a chosen path. When two
sides of a triangle are on the path, the third is necessarily in correspondence
with the identity element. From this, it comes out immediately that g12 =
96
CHAPTER 3. HOMOTOPY
g23 = g45 = g78 = g48 = 1. It follows also that g68 = 1. We remain with two
conditions, g51 g16 = 1 and g16 g67 = 1, from which g51 = g67 = g61 =: g. One
independent generator: 1 = Z. In fact, the hollowed disk is homotopically
equivalent to S 1 . Notice that the dark, absent triangle was not used as a
simplex: it is just the hollow. If it were used, one more condition would be
at our disposal, which would enforce g = 1. Of course, this would be the disk
with no hollow at all, for which 1 = {1}.
3.1.34 The twice-punctured disk. Figure 3.23 shows, in its left part,
a triangulation for the disk with two hollows and a chosen path, the upper
part being a repetition of the previous case. We find again g51 = g67 =
g61 =: g. The same technique, applied to the lower part gives another,
independent generator: g4,13 = g3,13 = g3,11 =: g 0 . The novelty here is that
the group is non-commutative. Of course, we do not know a priori the
relative behaviour of g and g 0 . To see what happens, let us take (right part
of Figure 3.23) the three loops , and starting at point (10), and examine
their representatives in the triangulation:
Figure 3.23: A triangulating for the twice-punctured disk.
[] = g10,12 g12,5 g56 g67 g78 g84 g4,12 g12,10 = g67 = g;

[] = g10,12 g12,4 g4,13 g13,10 = g 0 ;
[] = g10,13 g13,4 g46 g67 g78 g84 g4,13 g13,10 = (g 0 )1 gg 0 .
3.1. PATH HOMOTOPY
97
As and are not homotopic, their classes are different: [] 6= []. But
[] = [ 1 ][][] ,
so that [][] 6= [][], or gg 0 6= g 0 g. The group 1 is non-abelian with two
generators and no specific name. It is an unnamed group given by its presentation (Math 2.2). It might be interesting to examine the Bohm-Aharonov
effect corresponding to this case, with particles described by wavefunctions
with two or more components to (possibly) avoid the loss of information on
the group in 1-dimensional representations.
3.1.35 The projective line RP1 (see 1.4.19): as the circle is homotopically equivalent to the hollowed disk, we adapt the triangulation of
3.1.33 for S 1 (see Figure 3.24, compared with Figure 3.22) with identified
antipodes: 1 = 1, 2 = 2, etc. Notice that this corresponds exactly
to a cone with extracted vertex. The path is 1 2 3 4 5 6 and,
of course, its antipode. An independent generator remains,
g = g14 = g26 = g16 = g36 = g56 = g15 ,
so that 1 (RP1 ) = Z. This is to be expected, as RP1 S 1 .
Figure 3.24:
3.1.36 The projective plane RP2 : the sphere S 2 with identified antipodes may be described as in Figure 3.25 (see 1.4.19). From the equations
g13 g34 = 1; g34 g42 = 1; g24 g41 = 1; g31 g14 = 1 ,
98
CHAPTER 3. HOMOTOPY
Figure 3.25:
we obtain the identifications

g := g13 = g43 = g42 = g41 = g31 = g 1 ,
so that g 2 = 1. The group has one cyclic generator of order 2. Consequently
(see Math.2.3), 1 (RP2 ) = Z2 . This group may be represented by the multiplicative group {1, 1}. This example of a non-trivial finite group is of
special interest. Compare the following three different kinds of loops at (say)
vertex 1. Loop (1341) is trivial: it can be continuously contracted to a
point. It corresponds to the identity element of 1 . Another loop is obtained
by going from 1 to 4 and then to the antipode 1 of 1 (which is identified to it). This is non-trivial, as such a loop cannot be deformed to a point
and corresponds to the element -1 of the group representation. Now, take
this same loop twice: 100 400 100 400 100 . We see in the Figure
3.25 that such a loop can be progressively deformed into a point; this effect
corresponds to the property (1)2 = 1 in the representation. The projective
plane RP2 is thus doubly-connected.
3.2
3.2.1
COVERING SPACES
Multiply-connected Spaces
3.2.1 A remarkable characteristic of multiply-connected spaces is that

functions defined on them are naturally multivalued. We have been using
3.2. COVERING SPACES
99
the word function for single-valued mappings but in this paragraph we

shall be more flexible. To get some insight about this point, let us consider
a simple case, like that illustrated in Figure 3.26. Suppose in some physical
situation we have a box in which a function , obeying some simple differential equation, describes the state of a system. Boundary conditions
say, the values of on L1 and L2 are prescribed by some physical reason.
Under very ordinary conditions, will have a unique value at each point inside the box (the configuration space), found by solving the equation. In
frequent cases, we could even replace the boundaries by other surfaces inside
the box, using the values on them as alternative boundary conditions. Let
us now deform the box so that it becomes an annular region, and then L1 =
L2 .
Figure 3.26:
Unless we had carefully chosen the boundary values at the start, will
become multivalued on L1 = L2 . Of course, this operation is a violence
against the space: it changes radically its topology from a topology analogous
to that of 1.3.12 to another, akin to that of 1.4.3. The initial box is simplyconnected, the final annulus is multiply-connected. But the point we wish to
stress is that, unless the boundary conditions were previously prepared by
hand so as to match, will become multivalued. Had we started with the
annulus as configuration space from the beginning, no boundaries as L1 and
L2 would be present.
3.2.2 If we want to be single-valued, we have to impose it by hand (say,
through periodic conditions). This happens in Quantum Mechanics when
100
CHAPTER 3. HOMOTOPY
the wavefunction is supposed to be single-valued: recall the cases of fine

quantum behaviour exhibited by some macroscopic systems, such as vortex
quantization in superfluids,5 and flux quantization in superconductors.6 Such
systems are good examples of the interplay between physical and topological
characteristics. They are dominated by strong collective effects. So stiff
correlations are at work between all their parts that they may be described
by a single, collective wavefunction. On the other hand, they have multiplyconnected configuration spaces, and the quantizations alluded to come from
the imposition, by physical reasons, of single-valuedness on the wavefunction.
This leads to topological-physical effects.7
3.2.3 Physical situations are always complicated because many suppositions and approximations are involved. We can more easily examine ideal
cases through mathematical models. Consider, to begin with, on the plane
E2 included in E3 (see Figure 3.27, left), the function defined by
Rx
(a) = 0 ; (x) = a A dl ,
Figure 3.27:
where A is some vector. The integration is to be performed along some curve,
such as 1 or 2 . The question of interest is: consider 1 and 2 to be curves
linking points a and b. Are the two integrals
R
R
1 (b) = 1 A dl and 2 (b) = 2 A dl
equal to each other? It is clear that
5
6
7
See, for example, Pathria 1972.

For example, Feynman, Leighton & Sands 1965, vol. III.
See Dowker 1979.
101
1 (b) 2 (b) =
R
1 2
A dl =
A dl,
the last integration being around the closed loop starting at a, going through
1 up to b, then coming back to a through the inverse of 2 , which is given
by (2 )1 = 2 . We see that (b) will be single-valued iff
H
A dl = 0 .
If the region S circumvented by 1 2 = 1 + (2 )1 is simply-connected,
then 1 2 is just the boundary S of S, and Greens theorem implies
R
R
A dl = S rotA d.
1 2
For general x, the single-valuedness condition for (x) is consequently given
by rot rotA = 0. In a contractible domain this means that some (x) exists
such that A = grad. In this case,
Rx
(x) = a grad dl = (x) (a) ,
so that we can choose (a) = 0 and (x) = (x). When this is the case, A
is said to be integrable or of potential type ( is its integral, or potential ), a
nomenclature commonly extended to (x) itself.
Figure 3.28:
3.2.4 Now to some Physics: suppose (x) to be the wavefunction of an

electron moving under the influence of a magnetic field B =rotA, with A
e
(x) is just the phase
the vector potential. Forgetting about other effects, ~c
acquired by (in the JWKB approximation of nonrelativistic Quantum Mechanics) when we go from a to x (Figure 3.27, right):

Z
2e x
(x) = exp i
A dl (a).
hc a
102
CHAPTER 3. HOMOTOPY
The monodromy condition, B = rot A = 0, is the absence of field. Nonvanishing electromagnetic fields are, for this reason, non-integrable phase
factors.8
3.2.5 Consider now a multiply-connected configuration space, as the annulus in Figure 3.28. Because the central region is not part of the space,
1 2 is no more the boundary of a domain. We can still impose A =
rot [v], for some v(x), but then v(x) is no longer single-valued.9 We can
still write B =rotA, but neither is A single-valued! In Quantum Mechanics
there is no reason for the phases to be single-valued: only the states must be
unique for the description to be physically acceptable. A state corresponds
to a ray, that is, to a set of wavefunctions differing from each other only
by phases. Starting at a and going through 1 2 , the phase changes in
proportion to the integral of A around the hole which, allowing multivalued
animals, can be written
I
Z
Z
2e
2e
2e
2e
rotA d =
B d =
A dl =
,
(phase) =
~c
hc S
hc S
hc
being the flux of B through the surface S circumvented by 1 2 (which
is not its boundary now!). In order to have
e
(a) = ei ~c (a)
single-valued, we must have (2e/hc) = 2n, that is, the flux is quantized:
=n
hc
.
e
It may happen that this condition does not hold, as in the Bohm-Aharonov
effect (see 4.2.18).10
3.2.6 All these considerations (quite schematic, of course) have been made
only to emphasize that multi-connectedness can have very important physical
consequences.
3.2.7 We shall in the next section examine more in detail the relation between monodromy and multiple-connectedness. Summing it up, the following
will happen: let X be a multiply-connected space and a function on X.
The function will be multivalued in general. A covering space E will be
8
9
10
Yang 1974.
Budak & Fomin 1973.
Aharonov & Bohm 1959; 1961.
103
an unfolding of X, another space on which becomes single-valued. Different functions will require different covering spaces to become single-valued,
but X has a certain special covering, the universal covering U (X), which
is simply-connected and on which all functions become single-valued. This
space is such that X = U (X)/1 (X). The universal covering U (X) may be
roughly seen as that unfolding of X with one copy of X for each element of
1 (X).
3.2.8 Recall the considerations on the configuration space of a system of
n identical particles ( 1.4.7; see also 3.2.29 below, and Math.2.9), which is
E3n /Sn . In that case, E3n is the universal covering and 1 Sn . Consider,
to fix the ideas, the case n = 2. Call x1 and x2 the positions of the first
and the second particles. The covering space E6 is the set {(x1 , x2 )}. The
physical configuration space X would be the same, but with the points (x1 ,
x2 ) and (x2 , x1 ) identified. Point (x2 , x1 ) is obtained from (x1 , x2 ) by the
action of a permutation P12 , an element of the group
S2 : (x2 , x1 ) = P12 (x1 , x2 ).
A complex function (x1 , x2 ) (say, the wavefunction of the 2-particle system)
will be single-valued on the covering space, but 2-valued on the configuration
space. To make a drawing possible, consider instead of E3 , the two particles
on the plane E2 . The scheme in Figure 3.29 shows how (x1 , x2 ) is singlevalued on E, where (x1 , x2 ) 6= (x2 , x1 ), and double-valued on X, where the
two values correspond to the same point (x1 , x2 ) (x2 , x1 ). There are two
sheets because P12 applied twice is the identity.
Wavefunctions commonly used are taken on the covering space, where they
are single-valued. The function [P12 (x1 , x2 )] is obtained from (x1 , x2 ) by
the action of an operator U (P12 ) representing P12 on the Hilbert space of
wavefunctions:
(x2 , x1 ) = [P12 (x1 , x2 )] = U (P12 )(x1 , x2 ).
This is a general fact: the different values of a multivalued function are
obtained by the action of a representation of a group, a distinct subgroup
of 1 (X) for each function. Above, the group S2 is isomorphic to the cyclic
group Z2 (notice
the analogy of this situation with the covering related to
the function z in next section). The whole fundamental group will give all
the values for any function. There are as many covering spaces of X as there
are subgroups of 1 (X).
3.2.9 Percolation Phase transitions are more frequently signalled by singularities in physical quantities, as the specific heat. However, sometimes
104
CHAPTER 3. HOMOTOPY
Figure 3.29:
ithey show themselves as clear alterations in the topology of configuration

space,11 as in all the phenomena coming under the headname of percolation.12 In its simplest form, it concerns the formation of longer and longer
chains of (say, conducting) guest material in a different (say, isolating)
host medium by the progressive addition of the former. The critical point
(say, passage to conductivity) is attained when a first line of the guest material traverses completely the host, by that means changing its original
simply-connected character. As more and more complete lines are formed,
the fundamental group becomes more and more complicated.
3.2.10 Covering for braid statistics Instead of the above covering,
braid statistics (see Math.2.9) requires, already for the 2-particle configuration space, a covering with infinite leaves (Figure 3.30).
3.2.11 Poincar
e conjecture Homotopy is a basic instrument in the taxonomic program of classifying topological spaces, that is, finding all classes
of homeomorphic spaces in a given dimension. This project has only been
successful for 2-dimensional spaces. The difficulties are enormous, of course,
but progress has been made in the 3-dimensional case. The main technique
used is surgery, through which pieces of a given space are cut and glued
in a controlled way to get other spaces. The long debate concerning the
Poincare conjecture gives a good idea of how intricate these things are. We
11
Broadbent & Hammersley 1957; Essam 1972. Old references, but containing the
qualitative aspects here referred to.
12
An intuitive introduction is given in Efros 1986.
105
Figure 3.30: The infinite unfolding of the 2-particle configuration space for
braid statistics.
have seen that the sphere S 2 is simply-connected. Actually, it is the only

simply-connected closed surface in E3 . The conjecture was that the same
holds in higher dimensions: S n would be the only simply-connected closed
surface in En+1 . It has been (progressively) proved for n 4. Only recently13
there seems to have been found a proof for Poincares original case, n = 3.
3.2.2
Covering Spaces
3.2.12 The Riemann surface of a multivalued analytic function14 is a covering space of its analyticity domain, a space on which the function becomes
single-valued. It is the most usual example of a covering space. The considerations of the previous section hint at the interest of covering spaces to
Quantum Mechanics. As said above, if X is the configuration space and
13
14
For a popular exposition see Rourke & Stewart 1986.

See, for instance, Forsyth 1965.
106
CHAPTER 3. HOMOTOPY
1 (X, x0 ) is non-trivial for some x0 in X, there is no a priori reason for the

wavefunction to be single-valued: this must be imposed by hand, as a physical principle. Some at least of the properties of their phases are measurable.
What follows is a simplified description of the subject with the purpose of
a fascicule des resultats. For details, the reader is sent to the copious
mathematical literature15 and to good introductions by physicists.16
3.2.13 Let us begin with the standard example (Figure 3.31).
Consider
on the complex plane C the function f : C C, f (z) = z. It is not
analytic at z = 0, where its derivatives explode. If we insist on analyticity,
the point z = 0 must be extracted from the domain of definition, which
becomes C {0} S 1 ( 3.0.9). The function f is continuous, as its inverse
takes two open sets (and so, their union) into an open set in C {0}. If
we examine
how a loop circumventing the zero is taken by f (z), we discover
that z simply takes one into another two values which are taken back to a
same value by the inverse. Only by going twice around the loop in C {0}
can we obtain a closed curve in the image space. On the other hand, a loop
not circumventing the zero is taken into two loops in the image space. The
trouble, of course, comes from
Figure 3.31:
the two-valued character of z and the solution is well known: the function
becomes monodromous if, instead of C {0}, we take as definition domain
a space formed by two Riemann sheets with a half-infinite line
(say, E1+
) in
common. This new surface is a covering of C\{0}. The function z is + | z|
15
16
See the references given at 3.0.12

See Dowker 1979 and Morette-DeWitt 1969; 1972.
107
on one sheet and - | z| on the other. They are related by a representation

of the cyclic group Z2 given by (+1, 1) and the multiplication. The group
Z2 is a subgroup of 1 (C {0}) = Z. The analogy with the statistical case
of 3.2.8 comes from the presence of the same group. The function ()would
require a covering formed by n Riemann sheets.
3.2.14 The idea behind the concept of covering space of a given multiplyconnected space X is to find another space E on which the function is singlevalued and a projection p : E X bringing it back to the space X (Figure
3.32, left). Different functions require different covering spaces. A covering on
which all continuous functions become single-valued is a universal covering
space. Such a universal covering always exists and this concept provides
furthermore a working tool to calculate fundamental groups.
3.2.15 Let us be a bit more formal (that is, precise): consider two topological spaces X and E, and let p : E X be a continuous surjective mapping.
Suppose that each point x X has a neighbourhood U whose inverse image
p1 (U ) is the disjoint union of open sets V in E, with the property that each
V is mapped homeomorphically onto U by p (see Figure 3.32, right). Then
the set {Va } is a partition of p1 (U ) into sheets; p is the covering map, or
projection;
and E is a covering space of X. Strictly speaking, the covering is given by
the pair (E, p). As a consequence of the above definition:
(i) for each x X, the subset p1 (x) of E (called fiber over x) has a
discrete topology;
(ii) p is a local homeomorphism;
(iii) X has a quotient topology obtained from E.
3.2.16 If E is simply-connected and p : E X is a covering map, then
E is said to be the universal covering space of X. From this definition,
the fundamental group of a universal covering space is 1 (E) = {1}. Up
to homotopic equivalence, this covering with a simply-connected space is
unique. The covering of 3.2.13 is not, of course, the universal covering of
C\{0}. As C\{0} is homotopically equivalent to S 1 , it is simpler to examine
S 1.
3.2.17 The mapping E1 S 1 given by
p(x) = (cos 2x, sin 2x) = ei2x
is a covering map. Take the point (1, 0) S 1 and its neighbourhood U
formed by those points in the right-half plane. Then,
[
1
1
1
p (U ) =
(n , n + ).
4
4
n=
108
CHAPTER 3. HOMOTOPY
Figure 3.32:
The open intervals Vn = (n1/4, n+1/4) are (see Figure 3.33) homeomorphically mapped onto U by p. As E1 is simply-connected, it is by definition
the universal covering space of S 1 . Other covering spaces are, for instance,
S 1 itself given as
S 1 = {z C such that |z| = 1}
with the mappings pn : S 1 S 1 , pn (z) = z n , n Z+ .
3.2.18 Consider the torus T 2 = S 1 S 1 . It can be shown that the product
of two covering maps is a covering map. Then, the product
p p : E1 E1 S 1 S 1 , (p p)(x, y) = (p(x), p(y)),
with p the mapping of 3.2.17, is a covering map and E2 , which is simplyconnected, is the universal covering of T 2 .
3.2.19 There are several techniques to calculate the fundamental groups
of topological spaces, all of them rather elaborate. One has been given in
109
Figure 3.33:
Figure 3.34:
section 3.1.3. We shall here describe another, which exploits the universal
covering space. As our aim is only to show the ideas, we shall concentrate in
obtaining the fundamental group of the circle, 1 (S 1 ).
3.2.20 Let p : E X be a covering map. If f is a continuous function of
Y to X, the mapping f : Y E such that p f = f is the lift, or covering,
of f . Pictorially, we say that the diagram 3.34 is commutative.
This is a very important definition. We shall be interested in lifts of two
kinds of mappings: paths and homotopies between paths. In the following,
some necessary results will simply be stated and, when possible, illustrated.
They will be useful in our search for the fundamental group of S 1 .
3.2.21 Let (E, p) be a covering of X and f : I X a path. The lift f is
the path-covering of f. If F : I I X is a homotopy, then the homotopy
F : I I E such that p F = F is the covering homotopy of F .
3.2.22 Take again the covering mapping of 3.2.17, p(x) = (cos 2x,
sin 2x) = ei2x . Then,
110
CHAPTER 3. HOMOTOPY
Figure 3.35:
i) the path f : I S 1 given by f (t) = (cos t, sin t), with initial endpoint
(1, 0), has the lift f(t) = t/2, with initial endpoint 0 and final endpoint 1/2
(Figure 3.35, left);
Figure 3.36:
ii) the path f : I S 1 , given by f (t) = (cos t, sin t), has the lift
f(t) = t/2 (Figure 3.35, right);
iii) the path h(t) = (cos 4t, sin 4t) traverses twice the circle S 1 ; it has the
= 2t (Figure 3.36).
lift h(t)
3.2.23 Let us enunciate some theorems concerning the uniqueness of pathand homotopy-coverings.
Theorem 1: Let (E, p) be the universal covering of X, and f : I X
be a path with initial endpoint x0 . If e0 E is such that p(e0 ) = x0 , then
there is a unique covering path of f beginning at e0 (Figure 3.37).
Theorem 2: Let (E, p) be the universal covering of X, and F : II X
be a homotopy with F (0, 0) = e0 . If e0 E is such that p(e0 ) = x0 , then
there is a unique homotopy-covering F : I I E such that F (0, 0) = e0
(Figure 3.38).
111
Figure 3.37:
Theorem 3: Finally, the monodromy theorem, establishing the relation

between covering spaces and the fundamental group. Let (E, p) be the universal covering of X, and let f and g be two paths on X from x0 to x1 .
Suppose f and g are their respective lifts starting at e0 . If f and g are
homotopic, then f and g have also the same final endpoint,
f(1) = g(1) ,
and are themselves homotopic.
3.2.24 We shall now proceed to apply these results to show that 1 (S 1 ) =
Z, the additive integer group. To do it, we shall exhibit a group isomorphism
between Z and 1 (S 1 , s0 ), with the point s0 = (1, 0) S 1 included in C.
Let p : E1 S 1 be defined by p(t) = (cos 2t, sin 2t). If f is a loop on
S 1 with base point s0 , its lift f is a path on E1 beginning at 0. The point
f(1) belongs to the set p1 (s0 ), that is, f(1) = some n Z. The monodromy
theorem tells us that the integer n depends only on the homotopy class of f .
We may then define a mapping
: 1 (S 1 , s0 ) Z by [f ] = f(1) = n Z.
It remains to show that is a group isomorphism. To do that, we should
show that it is onto, one-to-one and preserves the group structure.17
(i) is onto: let n p1 (s0 ). Being E1 path-connected, we can choose
the path f : I E1 from 0 to n. Then f = p f is a loop on S 1 with base
point s0 , f is its lift and by definition
[f ] = f(1) = n.
(ii) is injective: suppose ([f ]) = ([g]) = n. Let us show that [f ] = [g].
Take the respective lifts f and g from 0 to n. They are homotopic to each
17
Fraleigh 1974.
112
CHAPTER 3. HOMOTOPY
Figure 3.38:
other, because E1 is simply-connected. If F is the homotopy between f and

g, the mapping F = p F will be the homotopy between f and g.
(iii) is a homomorphism: let f and g be two loops on S 1 at the point
s0 , f and g their lifts on E1 at the point 0. Define a path on E1 by
=
h(t)
f(2t)
for t [0, 21 ]
n + g(2t 1) for t [ 12 , 1].
begins at 0. It
and suppose that f(1) = n and g(1) = m. By construction, h
is the lift of f g, as the functions sine and cosine have
is easy to see that h
= f g. On the other hand,
periods 2[p(n + t) = p(t)]. Consequently p h
([f g]) = h(1)

= n + m = ([f ]) + ([g]).
In simple words, a loop on S 1 will always be deformable into some loop of
the form
fn (t) = exp[fn (t)],
with fn (t) = nt, and such that [fn ] = fn (1) = n. Notice that, in order to
show that is a homomorphism, we have used another structure present in
E1 , that of additive group. Covering spaces in general do not present such
a structure. Even so, some information on the fundamental group can be
obtained through the following theorem:
113
3.2.25 Theorem: Let p : (E, s0 ) (X, x0 ) be a covering map. If E is

path-connected, there exists always a surjective mapping
: 1 (X, x0 ) p<1> (x0 ).
If E is simply-connected, is bijective.
3.2.26 Intuitively, 1 (X) counts the number of sheets necessary to make
of X a simply-connected space, or to obtain its universal covering. It is usual
to write, for this reason, X = E/1 (X). What about the other covering
spaces, those which are not universal? We shall only state the general answer:
there is a covering space C for each subgroup K of 1 , obtained by
factorizing the universal covering by that subgroup : C = E/K.
Another general fact is the following:
if a space C is locally homeomorphic to another space X,
then C is a covering space of X.
So, 1 and its subgroups characterize all the covering spaces. Notice that
the counting it provides comes between quotation marks because 1 is not
necessarily Z, nor even an abelian group.
3.2.27 The projective plane RP2 ( 1.4.19 and 3.1.36) is obtained from S 2
by identifying each point x S 2 to its antipode x = x. This identification
being an equivalence relation on S 2 , RP 2 is the set of equivalence classes. A
projection
p : S 2 RP 2
exists, given by p(x) = [x] = |x|. The topology is the quotient topology:
U in RP2 is open if p1 (U ) is open in S 2 . It is possible to show that p is
a covering map. As S 2 is simply-connected (1 (S 2 ) = {1}), the theorem of
3.2.25 tells us that there is a bijection between 1 (RP 2 , r0 ) and p1 (r0 ).
But then, 1 (RP 2 , r0 ) is a group of rank 2, because p1 (r0 ) is a set with 2
elements. Any group of rank 2 is isomorphic to the cyclic group Z2 , so that
1 (RP 2 ) Z2 . We write RP2 = S 2 /Z2 . It is a general result that, for any
n 2,
1 (RP n ) Z2 , or RPn = S n /Z2 .
As for the complex projective spaces CPn , they are all simply-connected:
1 (CPn ) = {1}, for any n .
114
CHAPTER 3. HOMOTOPY
An alternative means to obtain 1 has already been used ( 3.1.23 and

3.1.24) for cartesian products. Another example is the cylinder.
3.2.28 The cylinder: as the cylinder is S 1 I,
1 (cylinder) 1 (S 1 ) 1 (I) Z {1} Z .
Notice that, when calculating 1 for cartesian products, one simply drops
homotopically trivial spaces.
3.2.29 Consider again ( 3.2.8) the configuration space of a system of n
identical particles, which is E3n /Sn . Recall that E3n is the universal covering
and 1 Sn is a nice example of non-abelian fundamental group (when
n 3). When, in addition, the particles are impenetrable, the fundamental
groups would be the braid groups Bn of Math.2.6. Such groups reduce to
the symmetric group Sn in E3 , but not on E2 . As a consequence, quantum
(and statistical) mechanics of identical impenetrable particles on E2 will be
governed by braid groups. Instead of the usual permutation statistics given
by the symmetric groups, which leads to the usual bosons and fermions, a
braid statistics will be at work. By the way, knots in E3 are characterized by
the fundamental group of their complement in the host space (Math.2.14).
3.2.30 Suppose a physical system with configuration space E3 , described
in spherical coordinates by the wavefunction (r, , ). We shall consider
rotations around the 0z axis, and write simply (). When we submit
the system to a rotation of angle around 0z, the transformation will be
represented by an operator U () acting on ,
( + ) = U ()().
When = 2, we would expect ( + 2) = (), that is, U (2) = 1. The
operator U () represents an element of SO(3), the rotation group in E3 (a
topological group). Roughly speaking, there is a mapping
U : [0, 2) SO(3) , U : U ()
defining a curve on the SO(3) space. Supposing U (0) = U (2) is the same
as requiring this curve to be a closed loop. Now, it happens that SO(3) is
a doubly-connected topological space, so that U () is not necessarily singlevalued. Actually,
SO(3) RP 3 S 3 /Z2 .
3.3. HIGHER HOMOTOPY
115
By the way, this manifold is the configuration space for the spherical top,
whose quantization18 is consequently rather involved. In fact, that is why
there are two kinds of wavefunctions: those for which U (2) = U (0) = 1, and
those for which U (2) = - 1. The first type describes systems with integer
angular momentum. Wavefunctions of the second type, called spinors, describe systems with half-integer angular momentum. The universal covering
of SO(3) is the group SU (2) of the unitary complex matrices of determinant
+ 1, whose manifold is the 3-sphere S 3 . Consequently, SO(3) = SU (2)/Z2 .
The group SU (2) stands with respect to SO(3) in a way analogous as the
square-root covering of 3.2.13 stands to C {0}: in order to close a loop
in SU (2), we need to turn twice on SO(3), so that only when = 4 is the
identity recovered. As SU (2) = S 3 is simply-connected, it is the universal
covering of SO(3).
3.2.31 The simplest case of Quantum Mechanics on a multiply-connected
space comes out in the well known Young double-slit interference experiment.
We shall postpone its discussion to 4.2.17 and only state the result. On a
multiply-connected space, the wavefunction behaves as the superposition of
its values on all the leaves of the covering space.
3.3
HIGHER HOMOTOPY
Besides lassoing them with loops, which are 1-dimensional objects, we can
try to capture holes in space with higher-dimensional closed objects. New
groups revealing space properties emerge.
3.3.1 The attentive (and suspicious) reader will have frowned upon the
notation 0 (A) used for the set of path-connected components of space A
in 3.1.4. There, and in 3.1.13, when the fundamental group 1 was
also introduced as the first homotopy group, a whole series of groups n
was announced. Indeed, both 0 and 1 are members of a family of groups
involving classes of loops of dimensions 0, 1, 2, . . . We shall now say a few
words on the higher groups n (X, x0 ), for n = 2, 3, 4, etc.
3.3.2 With loops, we try to detect space defects by lassoing them, by
throwing loops around them. We have seen how it works in the case of
a point extracted from the plane, but the method is clearly inefficient to
apprehend, say, a missing point in E3 . In order to grasp it, we should tend
18
Schulman 1968; Morette-DeWitt 1969, 1972; Morette-DeWitt, Masheshvari & Nelson

1979.
116
CHAPTER 3. HOMOTOPY
something like a net, a 2-dimensional loop. Higher dimensional spaces

and defects19 require analogous higher dimensional loops. As happens for
1-loops, classes of n-loops constitute groups, just the n .
3.3.3 The fundamental group was due to Poincare. The groups n for
general n Z have been introduced in the thirties by E. Cech and W.
Hurewicz. The latter gave the most satisfactory definition and worked out
the fundamental properties. His approach was restricted to metric spaces
but was extended to general topological spaces by R. H. Fox in the forties.
We shall here try to introduce the higher groups as natural extensions of the
fundamental group.
3.3.4 The group 1 (X, x0 ) is the set of homotopy classes of (1-dimensional)
loops on X with base point x0 . With this in mind, our initial problem is to
define 2-dimensional loops. A 1-dimensional loop is a continuous mapping
f : I X with f ( I) = x0 , where I is a provisional notation for the
boundary of I, that is, the set of numbers {0, 1}. We can then define a 2dimensional loop as a continuous mapping from I2 = I I into X, given
by f : I I X such that f ( I2 ) = x0 , where I2 is the boundary of I2 .
3.3.5 To extend all that to higher dimensions, we need beforehand an
extension of the closed interval: denoted by In , it is an n-dimensional solid
cube:
In = {x = (x1 , x2 , . . . , xn ) En such that 0 xi 1, i}.
The set In , the boundary of In , is the cube surface, which can be
defined in a compact way by
Q
In = {x In such that ni=1 xi (1 xi ) = 0}.
3.3.6 Let (X, x0 ) be a topological space with a chosen point x0 . Denote
by n (X, x0 ) the set of continuous functions f : In X such that f ( In ) =
x0 . Given two of such functions f and g, they are homotopic to each other
if there exists a continuous mapping F : In I X satisfying
F (x1 , x2 , . . . , xn ; 0) = f (x1 , x2 , . . . , xn )
F (x1 , x2 , . . . , xn ; 1) = g(x1 , x2 , . . . , xn )
where (x1 , x2 , . . . , xn ) In , and F (x1 , x2 , . . . , xn , s) = x0 when (x1 , x2 , . . .
, xn ) In for all s I. The function F is the homotopy between f and g.
In shorthand notation,
19
Applications of homotopy to defects in a medium are examined in Nash & Sen 1983.
117
F (In ; 0) = f ; F (In ; 1) = g;
F ( In , I) = x0 .
That this homotopy is an equivalence relation can be shown in a way analogous to the 1-dimensional case. The set n (X, x0 ) is consequently decomposed into disjoint subsets, the homotopy classes. The class to which f
belongs will be once again indicated by [f ].
3.3.7 Notice that, in the process of closing the curve to obtain a loop, the
interval I itself becomes equivalent to a loop from the homotopic point
of view, we could take S 1 instead of I with identified endpoints. Actually,
each loop could have been defined as a continuous mapping f : S 1 X. In
the same way, we might consider the n-loops as mappings f : S n X. This
alternative definition will be formalized towards the end of this section.
3.3.8 Let us introduce a certain algebraic structure by defining an operation analogous to that of 3.1.6. Given f and g n (X, x0 ),
h(t1 , t2 , . . . , tn ) =

(f g)(t1 , t2 , . . . , tn ) =
f (2t1 , t2 , . . . , tn )
for t1 [0, 21 ]
g(2t1 1, t2 , . . . , tn ) for t1 [ 12 , 1].
Operation induces an operation on the set of homotopy classes of n :

[f ] [g] = [f g].
With the operation , the set of homotopy classes of n (X, x0 ) constitutes a
group, the n-th homotopy group of space X with base point x0 , denoted by
n (X, x0 ).
3.3.9 Many of the results given for the fundamental group remain valid
for n (X, x0 ) for n 2. We shall now list some general results of the theory
of homotopy groups:
(i) if X is path-connected and x0 , x1 X, then
n (X, x0 ) n (X, x1 ) for all n 1 ;
(ii) if X is contractible, then n (X) = 0 n Z+ ;
(iii) if (X, x0 ) and (Y, y0 ) are homotopically equivalent, then for any n
Z+ ,
n (X, x0 ) n (Y, y0 ) ;
118
CHAPTER 3. HOMOTOPY
(iv) take X and Y topological spaces and x0 , y0 the respective chosen

points; then for their topological product,
n (X Y, (x0 , y0 )) n (X, x0 ) n (Y, y0 ) .
3.3.10 From the property (iv), we can see that, for euclidean spaces,
n (Em ) {0}, for any n and m.
3.3.11 The functorial properties of 3.1.21 keep their validity for n 2.
Let : X Y be a continuous mapping with (x0 ) = y0 . If [f ] n (X, x0 )
for some n, then f : In Y is a continuous mapping with base point y0 ,
that is, ( f )( In ) = y0 ; thus, f is an element of the class [ f ]
n (Y, y0 ). Consequently, induces a map
: n (X, x0 ) n (Y, y0 )
given by
([f ]) = [ f ]
for every [f ] n (X, x0 ). This mapping, which can be shown to be well
defined, is the induced homomorphism relative to the base point x0 . It has
the following functorial properties:
i) if : (X, x0 ) (Y, y0 ) and : (Y, y0 ) (Z, z0 ), then ( ) = .
Given the identity mapping i : (X, x0 ) (X, x0 ), then i is the identity
homomorphism;
ii) if : (X, x0 ) (Y, y0 ) is a homeomorphism, then is an isomorphism
between n (X, x0 ) and n (Y, y0 ).
3.3.12 In two important aspects the higher homotopy groups differ from
the fundamental group:
(i) let (E, p) be the universal covering of X, and let e0 E such that
p(e0 ) = x0 X; then, the induced homomorphism
p : n (E, e0 ) n (X, x0 )
is a group isomorphism for n 2; this means that the universal covering,
which for the fundamental group is trivial, keeps nevertheless all the higher
homotopy groups of the space.
(ii) for X any topological space, all the n (X, x0 ) for n 2 are abelian
(which again is not the case for the fundamental group).
119
3.3.13 For any n Z+ , the n-th homotopy group of the sphere S n is

isomorphic to Z, that is, n (S n ) Z. In section 3.2.17 we have considered
a particular case of the family of loops on S 1 given by
fm (t) = ei2mt .
For each n, the corresponding lift is f(t) = mt. The group homomorphism
: 1 Z takes each class [fm ] into m: ([fm ]) = m. The parameter m
is the number of times a member of the class [fm ], say fm (t) itself, covers
S 1 . Or better, the image space of fm covers S 1 m times. In the same
way, n (S n ) contains the classes of functions whose image space covers
S n . The functions of a given class covers S n a certain number of times,
this number being precisely the labeling m given by . As the m-loops on a
space X correspond to mappings f : S m X, we have here maps S n S n
in which the values cover the target S n m times. This number m is known
in the physical literature as winding number and turns up in the study of
monopoles and of the vacuum in gauge theories (see Phys.7).20 When the
target space is a quotient as S m /Z2 , the winding number can assume halfinteger values, as in the case of the Franck index in nematic systems (see
Phys.3.3.3).
3.3.14 Consider the covering space (E1 , p) of S 1 . As
p : n (E1 , e0 ) n (S 1 , x0 )
is an isomorphism for all n 2, then
n (S 1 ) n (E1 ) {0} n 2.
3.3.15 Take the covering (S n , p) of RPn . As p : n (S n ) n (RPn ) is an
isomorphism for n 2,
n (RPn ) n (S n ) Z,
n 2.
3.3.16 Let us now mention two alternative (and, of course, equivalent)

definitions of n (X, x0 ). The first has just been alluded to, and said to be
relevant in some physical applications. We have defined a 1-loop at x0 as a
continuous mapping f : I X such that f ( I) = x0 . On the other hand,
the quotient space obtained by the identification of the end-points of I is
simply S 1 . We can then consider a 1-loop on X as a continuous mapping
f : S 1 X, with f (1, 0) = x0 . In an analogous way, a 2-dimensional loop
will be a continuous mapping f : S 2 X. The definition of a homotopy of
functions S n X, necessary to get n (X, x0 ), is the following:
20
See the classical series of lectures by Coleman 1977; 1979.
120
CHAPTER 3. HOMOTOPY
let (X, x0 ) be a topological space with a chosen point, and call = n n

the set of all continuous mappings f : S n X, n Z+ , satisfying
f (1, 0, 0, . . . , 0) = x0 . Two functions f and g are homotopic if a
continuous F : S n I X exists such that
F (x, 0) = f (x); F (x, 1) = g(x);
F (1, 0, 0, . . . , 0, s) = x0 , s I.
Of course, F is a homotopy between f and g.
3.3.17 Another definition, due to Hurewicz, involves the idea that a 2-dimensional
loop is a loop of loops. In other words: a 2-dimensional loop is a function f : I X
such that, for each t I, the image f (t) is itself a loop on X, and f ( I) = x0 . With
this in mind, we can endow the set (X, x0 ) of loops on X at x0 with a topology (the
compact-open topology of 1.3.14), making it into a topological space. It turns out that
2 (X, x0 ) can be defined as the fundamental group of (X, x0 ).
More generally, we have the following: let X be a topological space and (X, x0 ) be the
set of loops on X with base point x0 , itself considered as a topological space with the
compact-open topology. If n 2, the n-th homotopy group on X at x0 is the (n 1)-th
homotopy group of (X, x0 ) at c, where c is the constant loop at x0 :
n (X, x0 ) n1 ((X, x0 ), c) .
3.3.18 A last word on 0 : 0-loops would be mappings from {0} I into X; such
loops can be deformed into each other when their images lay in the same path-component
of X. It is natural to put their classes, which correspond to the components, in the family
{n (X)}.
3.3.19 Hard-sphere gas A gas of impenetrable particles in E3 has, of course, nontrivial 2 . The classical problem of the hard-sphere gas is a good example of the difficulties
appearing in such spaces. After some manipulation, the question reduces to the problem
of calculating the excluded volume21 left by penetrable particles, which is as yet unsolved.
3.3.20 General references on homotopy For beginners, giving clear introductions

to the fundamental group: Munkres l975, or Hocking & Young 1961. Very useful because
they contain many detailed calculations and results, are: Greenberg 1967, and Godbillon
1971. An introduction specially devoted to physical applications, in particular to the problems of quantization on multiply-connected spaces, the Bohm-Aharonov effect, instantons,
etc, is Dowker l979. A good review on solitons, with plenty of homotopic arguments is
Boya, Cari
nena & Mateos 1978. A pioneering application to Gauge Theory is found in
Loos 1967.
21
See, for instance, Pathria 1972.
Chapter 4
MANIFOLDS & CHARTS
4.1
4.1.1
MANIFOLDS
Topological manifolds
Topological manifolds are spaces on which coordinates make sense.

We have up to now spoken of topological spaces in great (and rough)
generality. Spaces useful in the description of physical systems are most
frequently endowed with much more structure, but not every topological
space accepts a given additional structure. We have seen that metric, for
instance, may be shunned by a topology. So, the very fact that one works
with more complicated spaces means that some selection has been made.
Amongst all the possible topological spaces, we shall from now on talk almost
exclusively of those more receptive to additional structures of euclidean type,
the topological manifolds.
4.1.1 A topological manifold is a topological space S satisfying the
following restrictive conditions:
i) S is locally euclidean : for every point p S, there exists an open set
U to which p belongs, which is homeomorphic to an open set in some
En ; the number n is the dimension of S at the point p. Given a
general topological space, it may have points in which this is not true:
1-dimensional examples are curves on the plane which cross themselves
(crunodes), or are tangent to themselves (cusps) at certain points. At
these singular points the neighbourhood above required fails to exist.
Points in which they do exist are called general points. Topological
manifolds are thus entirely constituted by general points. Very important exceptions are the upper-half spaces En+ ( 1.1.10). Actually, so
important are they that we shall soften the condition to
i) around every p S there exists an open U which is either homeomorphic
to an open set in some En or to an open set in some En+ . Dimension is
121
122
CHAPTER 4. MANIFOLDS & CHARTS

still the number n. Points whose neighbourhoods are homeomorphic to
open sets of En+ and not to open sets of En constitute the boundary S
of S. Manifolds including points of this kind are called manifolds-withboundary. Those without such points are manifolds-without-boundary,
or manifolds-with-null-boundary.
ii) the space S has the same dimension n at all points. The number n is then
the dimension of S, n = dim S. The union of a surface and a line in
E3 is not a manifold. This condition can be shown to be a consequence
of the following one, frequently used in its stead:
ii) S is connected. When necessary, a non-connected S can be decomposed
into its connected components. For space-time, for instance, connectedness is supposed to hold because we would have no knowledge of
any disconnected component not our own. Nowadays, with the belief
in the existence of confined quarks and shielded gluons based on an
ever increasing experimental evidence, one should perhaps qualify this
statement.
iii) S has a countable basis (i.e., it is second-countable).
This is a pathology-exorcizing requirement the Sorgenfrey line of
1.2.17, for example, violates it.
iv) S is a Hausdorff space.
Again to avoid pathological behaviours of the types we have talked
about in 1.2.15.
4.1.2 Not all the above conditions are really essential: some authors call
topological manifold any locally-euclidean or locally-half-euclidean connected topological space. In this case, the four conditions above define
a countable Hausdorff topological manifold. We have already said that
Einsteins equations in General Relativity have solutions exhibiting nonHausdorff behaviour in some regions of spacetime, and that some topologies
proposed for Minkowski space are not second-countable. The fundamental
property for all that follows is the local-euclidean character, which will allow
the definition of coordinates and will have the role of a complementarity
principle: in the local limit, the differentiable manifolds whose study is our
main objective will be fairly euclidean. That is why we shall suppose from
now on the knowledge of the usual results of Analysis on En , which will be
progressively adapted to manifolds in what follows.
4.1. MANIFOLDS
123
It suffices that one point of S have no euclidean open neighbourhood to

forbid S of being made into a manifold. And all non-euclidean opens sets
are lost in a manifold.
4.1.2
Dimensions, integer and other
4.1.3 The reader will have noticed our circumspection concerning the concept of dimension. This seems intuitively a very topological idea, because
it is so fundamental. We have indeed used it as a kind of primitive concept.
Just above, we have taken it for granted in euclidean spaces and defined dimensions of more general spaces in consequence. But only locally-euclidean
spaces have been contemplated. The trouble is that a well established theory for dimension only exists for metric second-countable spaces, of which
euclidean spaces are a particular case. The necessity of a theory to provide a well defined meaning to the concept became evident at the end of
last century, when intuition clearly showed itself a bad guide. Peano found
a continuous surjective mapping from the interval I = [0, 1] into its square
I I, so denying that dimension could be the least number of continuous
real parameters required to describe the space. Cantor exhibited a one-toone correspondence between E1 and E2 , so showing that the plane is not
richer in points than the line, dismissing the idea of dimension as a measure
of the point content of space and even casting some doubt on its topological nature. Mathematicians have since then tried to obtain a consistent and
general definition.
4.1.4 A simplified rendering1 of the topological dimension of a space X is
given by the following series of statements:
i) the empty set , and only , has dimension equal to 1: dim = 1;
ii) dim X n if there is a basis of X whose members have boundaries of
dimension n 1;
iii) dim X = n if dim X n is true, and dim X n 1 is false;
iv) dim X = if dim X n is false for each n.
This definition has at least two good properties: it does give n for the euclidean En and it is monotonous (X Y implies dim X dim Y ). But a
close examination shows that it is rigorous only for separable metric (equivalently, metric second-countable) spaces. If we try to apply it to more general
1
A classic on the subject is Hurewicz & Wallman 1941; it contains a historical introduction and a very commendable study of alternative definitions in the appendix.
124
spaces, we get into trouble: for instance, one should expect that the dimension of a countable space be zero, but this does not happen with the above
definition. Furthermore, there are many distinct definitions which coincide
with that above for separable metric spaces but give different results for less
structured spaces, and none of them is satisfactory in general.2
4.1.5 In another direction, explicitly metric, we may try to define dimension by another procedure: given the space as a subset of some En , we count
the number N () of n-cubes of side necessary to cover it. We then make
smaller and smaller, and calculate the Kolmogorov capacity (or capacity
dimension)
ln N ()
0 ln(1/)
dc = lim
Suppose a piece of line: divide it in k pieces and take = 1/k. Then, the
number of pieces is N () = k, and dc = 1. A region of the plane E2 may
be covered by k 2 = N () squares of side = 1/k, so that dc =2. The
capacity dimension gives the expected results for simple, usual spaces. It is
the simplest case of a whole series of dimension concepts based on ideas of
measure which, unlike the topological dimension, are not necessarily integer.
Consider a most enthralling example: take the Cantor set of 1.2.4. After
the j-th step of its construction, 2j intervals remain, each of length ( 31j ).
Thus, N () = 2j intervals of length = 31j are needed to cover it, and
dc =
ln 2
.6309...!
ln 3
Spaces with fractional dimension seem to peep out everywhere in Nature and
have been christened fractals by their champion, B. B. Mandelbrot.3 Notice
that the fractal character depends on the chosen concept of dimension: the
topological dimension of the Cantor set is zero. Fractals are of great (and
ever growing) importance in dynamical systems.4
4.1.6 Spaces with non-integer dimensions have been introduced in the
thirties (von Neumann algebras, see Math.5.5).
2
A sound study of the subject is found in Alexandrov 1977.

Much material on dimensions, as well as beautiful illustrations on fractals and a
whole account of the subject is found in Mandelbrot 1977.
4
For a discussion of different concepts of dimension which are operational in dynamical
systems, see Farmer, Ott & Yorke 1983.
3
4.2. CHARTS AND COORDINATES
4.2
125
CHARTS AND COORDINATES
4.2.1 Let us go back to item (i) in the definition of a topological manifold: every point p of the manifold has an euclidean open neighbourhood U ,
homeomorphic to an open set in some En , and so to En itself. The homeomorphism
: U open set in En
will give local coordinates around p. The neighbourhood U is called a coordinate neighbourhood of p. The pair (U, ) is a chart, or local system of
coordinates (LSC) around p. To be more specific: consider the manifold En
itself; an open neighbourhood V of a point q En is homeomorphic to another open set of En . Each homeomorphism of this kind will define a system
of coordinate functions, as u in Figure 4.1. For E2 , for instance, we can use
Figure 4.1: Coordinates, and coordinate functions around a point p.

the system of cartesian coordinates (u1 (q) = x, u2 (q) = y); or else the system
of polar coordinates (u1 (q) = r (0, ); u2 (q) = (0, 2)); and so on.
4.2.2 Take a homeomorphism x : S En , given by
x(p) = (x1 , x2 , . . . , xn ) = (u1 (p), u2 (p), . . . , un (p)).
The functions xi = ui : U E1 will be the local coordinates around p. We
shall use frequently the simplified notation (U, x) for the chart. What people
usually call coordinate systems (e.g. cartesian, polar, elliptic) are actually
126
systems of coordinate functions, corresponding to the above u. This is a

relevant distinction because distinct systems of coordinate functions require
a different number of charts to plot a given space S. For E2 itself, one cartesian system is enough to chart the whole space: U = E2 , u = the identity
mapping. However, this is not true for the polar system: the coordinate
is not related to a homeomorphism on the whole plane, as its inverse is not
continuous it takes points near to 0 and 2 back to neighbouring points
in E2 . There is always a half-line which is not charted (usually taken as R+ );
at least two charts are required by this system. One sometimes forgets this
point, paying the price of afterwards finding some singularity. Some of the
most popular singularities in Physics are not real, but of a purely coordinate origin. A good example of a singular line which is not real, but only
a manifestation of coordinate inadequacy, is the string escorting the Dirac
magnetic monopole in its elementary formulation, which disappears when
correct charts are introduced.5 Another case is the Schwarzschild radius, not
a singularity when convenient coordinate systems are used.6 The word convenient here may be a bit misleading: it means good for formal purposes.
The fact that singularities are of coordinate origin does not mean that they
will not be physically observed, as measuring apparatuses can presuppose
some coordinate function system.
4.2.3 The coordinate homeomorphism could be defined in the inverse sense, from an
open set of some En to some neighbourhood of the point p S. It is then called a
parameterization .
4.2.4 Of course, a given point p S can in principle have many different

coordinate neighbourhoods and charts. Remember the many ways used to
plot the Earth in cartography.
On purpose, cartography was the birth-place of charts, whose use was pioneered by
Hipparchos of Nicaea in the second century B.C.
4.2.5 As the coordinate homeomorphism x of the chart (U, x) takes into

a ball of En , which is contractible, U itself must be contractible. This gives
a simple criterium to have an idea on the minimum number of necessary coordinate neighbourhoods. We must at least be able to cover the space with
charts with contractible neighbourhoods. Let us insist on this point: an LSC
is ultimately (U, x = u ), and x has two pieces. The homeomorphism
takes U into some open V of En ; the coordinate function u chooses coordinates for En itself, taking it into some subspace [for instance, spherical
5
6
Wu & Yang 1975.

Misner, Thorne & Wheeler l973, 31.2.
127
coordinates (r, , ) on E3 involve u: E3 E1+ (0, ) (0, 2)]. Recall that V

is homeomorphic to En . If a space N could be entirely covered by a single
chart, would be a homeomorphism between N and En . If N is not homeomorphic to En , it will necessarily require more than one chart. The minimum
number of charts is the minimum number of open sets homeomorphic to En
covering N , but the real number depends also of the function u. The sphere
S 2 , for example, needs at least two charts (given by the stereographic projections from each pole into an E2 tangent in the opposite pole, see Math.11.3)
but the imposition of cartesian coordinates raises this number to eight.7
Thus, summing up: the multiplicity of the necessary charts depends (i) on the minimum number of euclidean open sets really needed to cover the space; (ii) on the system
of coordinate functions.
4.2.6 The fact that a manifold may require more than one chart has a
remarkable consequence. Transformations are frequently treated in two supposedly equivalent ways, the so called active (in which points are moved)
and passive (in which their coordinates are changed) points of view. This
can be done in euclidean spaces, and its generality in Physics comes from the
euclidean supremacy among usual spaces. On general manifolds, only the
active point of view remains satisfactory.
4.2.7 Given any two charts (U, x) and (V, y) with U V 6= , to a given
point p U V will correspond coordinates x = x(p) and y = y(p) (see
Figure 4.2). These coordinates will be related by a homeomorphism between
open sets of En ,
y x<1> : En En ,
which is a coordinate transformation and can be written as
y i = y i (x1 , x2 , . . . , xn ).
(4.1)
xj = xj (y 1 , y 2 , . . . , y n ).
(4.2)
Its inverse is x y <1> , or
4.2.8 Given two charts (U , ) and (U , ) around a point, the coordinate transformation between them is commonly indicated by a transition
function
g :(U , ) (U , )
1
and its inverse g
.
7
See Flanders l963.
128
Figure 4.2: Two distinct charts around p.
4.2.9 Consider now the euclidean coordinate spaces as linear spaces, that
is, considered with their vector structure. Coordinate transformations are
relationships between points in linear spaces. If both x y <1> and y x<1>
are C (that is, differentiable to any order) as functions in En , the two local
systems of coordinates (LSC) are said to be differentially related .
4.2.10 An atlas on the manifold S is a collection of charts {(U , )} such
that
S
U = S.
4.2.11 The following theorem can be proven:
any compact manifold can be covered by a finite atlas,
that is, an atlas with a finite number of charts.
4.2.12 If all the charts are related by linear transformations in their intersections, it will be a linear atlas.
4.2.13 If all the charts are differentially related in their intersections, it
will be a differentiable atlas. This requirement of infinite differentiability can
be reduced to k-differentiability. In this case, the atlas will be a C k -atlas.
Differentiating [4.1] and [4.2] and using the chain rule,
ki =
y i xj
.
xj y k
(4.3)
129
4.2.14 This means that both jacobians are 6= 0. If some atlas exists on S
whose jacobians are all positive, S is orientable . Roughly speaking, it has
two faces. Most commonly found manifolds are orientable. The Mobius strip
and the Klein bottle are examples of non-orientable manifolds.
4.2.15 Suppose a linear atlas is given on S, as well as an extra chart not belonging
to it. Take the intersections of the coordinate-neighbourhood of this chart with all the
coordinate-neighbourhoods of the atlas. If in these intersections all the coordinate transformations from the atlas LSCs to the extra chart are linear, the chart is admissible to
the atlas. If we add to a linear atlas all its admissible charts, we get a (linear) complete
atlas, or (linear) maximal atlas. A topological manifold with a complete linear atlas is
called a piecewise-linear manifold (usually, a PL manifold).
4.2.16 A topological manifold endowed with a certain differentiable atlas
is a differentiable manifold (see Chapter 5 , section 5.1 ). These are the most
important manifolds for Physics and will deserve a lot of attention in the
forthcoming chapters.
4.2.17 Electron diffraction experiment The simplest case of Quantum Mechanics on a multiply connected space appears in the well known
Young double-slit interference experiment, or in the 1927 Davisson & Germer electron diffraction experiment. We suppose the wave function to be
represented by plane waves (corresponding to free particles of momentum
p = mv) incident from the left (say, from an electron source S situated far
enough to the left). A more complete scheme is shown in Figure 4.5), but let
us begin by considering only the central part B of the future doubly-slitted
obstacle (Figure 4.3). We are supposing the scene to be E3 , so that B extends
to infinity in both directions perpendicular to the drawing. Of course, the
simple exclusion represented by B makes the space multiply-connected, actually homeomorphic to E3 \E1 . A manifold being locally euclidean, around
each point there is an open set homeomorphic to an euclidean space. When
the space is not euclidean, it must be somehow divided into intersecting regions, each one euclidean and endowed with a system of coordinates. Here,
it is already impossible to use a unique chart, at least two as in Figure 4.4
being necessary. We now add the parts A and C of the barrier (Figure 4.5).
Diffraction sets up at the slits 1 and 2, which distort the wave. The slits
act as new sources, from which waves arrive at a point P on the screen after
propagating along straight paths 1 and 2 perpendicular to the wavefronts.
If the path lengths are respectively l1 and l2 , the wavefunction along k will
k
. The two waves will have at P a phase difference
get the phase 2l
2|l1 l2 |
2mv|l1 l2 |
=
.
130
Figure 4.3: Barrier B extends to infinity in both directions perpendicular to

the drawing.
Let us recall the usual treatment of the problem8 and learn something from
Quantum Mechanics. At the two slit-sources, we have the same wavefunction
0 . The waves superpose and interfere all over the region at the right, in
particular at P . The waves arrive at P as
1
2l2
1 = 0 exp i 2l
and
exp
i .
2
0
Their superposition leads then to the relative probability density

2
1
2
| exp i 2l
+ exp i 2l
]| = 2 + 2 cos[ 2|l1l2 | ].
This experimentally well verified result will teach us something important.

The space is multiply-connected and the wavefunction is actually 1 on the
first chart (on the first covering leaf) and 2 on the second chart (on the
second covering leaf). We nevertheless obtain a single-valued at P by taking the superposition. Thus, we simply sum the contributions of the distinct
leaves! The morality is clear: wavefunctions on a multiply-connected manifold should be multiply-valued; Quantum Mechanics, totally supported by
experiment, tells us that we can use as wavefunction the unique summation
of the leaves contributions.
8
Furry 1963; Wootters & Zurek 1979.
131
Figure 4.4: At least two charts are necessary to cover E3 \E1 .
Figure 4.5: Scheme or the double-slit diffractio nexperiment.
4.2.18 Aharonov-Bohm effect We may go ahead9 with the previous

example. Drop parts A and C of the intermediate barrier, and replace part
B by an impenetrable infinite solenoid orthogonal to the figure plane and
carrying a magnetic field B (see Figure 4.6). The left side is replaced by
a unique electron point source S. A single wavefunction 0 is prepared at
point S but, the region of the solenoid being forbidden to the electrons, the
domain remains multiply-connected. There is no magnetic field outside that
region, so that the exterior (tri-) vector potential is a pure gauge, A = f.
Waves going north and south from the forbidden region will belong to
distinct leaves or charts. In the eikonal approximation, the wavefunction at
P obtained by going along 1 will be
9
See Furry 1963, Wu & Yang 1975 and Dowker 1979.
132
Figure 4.6: A scheme for the Aharonov-Bohm effect.
1 = 0 e
i2
h
R
1
pdq
= 0 e
i2
h
e
1 [mv+ c A]dl
= 0 e
i2
h
R
1
mvdl+ i2
h
e
c
e
c
Adl
Adl
The wavefunction at P obtained by going along 2 will be

2 = 0 e
i2
h
R
2
pdq
= 0 e
i2
h
e
2 [mv+ c A]dl
= 0 e
i2
h
R
2
mvdl+ i2
h
Taking a simplified view, analogous to the double slit case, of the kinematic
part, the total phase difference at P will be given by
h
i
H
2|l1 l2 |
i
+ 2e
Adl
hc
2 = 1 e
The closed integral, which is actually a line integral along the curve 2 1 , is
the magnetic flux. The effect, once the kinematical contribution is accounted
for, shows that the vector potential, though not directly observable, has an
observable circulation. The wavefunctions 1 and 2 are values of on
distinct leaves of the covering space, and should be related by a representation
of the fundamental group, which is here 1 = Z. A representation of the
group Z, acting on any (one-dimensional) wavefunction, will be given by any
phase factor like exp[i2], for each real value of , :n exp[i2 n]. The
value of is obtained from above,
H
e
2|
= |l1 l
+ hc
A dl.
Once this is fixed, one may compute the contribution of other leaves, such as
those corresponding to paths going twice around the forbidden region. For
each turn, the first factor will receive the extra contribution of the length of
a circle around B, and the contribution of many-turns paths are negligible
in usual conditions.
Chapter 5
DIFFERENTIABLE
MANIFOLDS
5.1
DEFINITION AND OVERLOOK
5.1.1 Suppose a differentiable atlas is given on a topological manifold S,

as well as an extra chart not belonging to it. Take the intersections of the
coordinate-neighbourhood of this chart with all the coordinate-neighbourhoods of the atlas. If in these intersections all the coordinate transformations
from the atlas LSCs to the extra chart are C , the chart is admissible to
the atlas. If we add to a differentiable atlas all its admissible charts, we get a
complete atlas, or maximal atlas, or C -structure . The important point is
that, given a differentiable atlas, its extension obtained in this way is unique.
A topological manifold with a complete differentiable atlas
is a differentiable manifold.
One might think that on a given topological manifold only one complete
atlas can be defined in other words, that it can become only one differentiable manifold. This is wrong: a fixed topological manifold can in principle
accept many distinct C -structures, each complete atlas with charts not admissible by the other atlases. This had been established for the first time in
1957, when Milnor showed that the sphere S 7 accepts 28 distinct complete
atlases. The intuitive idea of identifying a differentiable manifold with its
topological manifold, not to say with its point-set (when we say a differentiable function on the sphere, the space-time, etc), is actually dangerous
(although correct for most of the usual cases, as the spheres S n with n 6)
and, ultimately, false. That is why the mathematicians, who are scrupulous
and careful people, denote a differentiable manifold by a pair (S, D), where
D specifies which C -structure they are using. Punctiliousness which pays
133
134
CHAPTER 5. DIFFERENTIABLE MANIFOLDS
well: it has been found recently, to general surprise, that E4 has infinite
distinct differentiable structures! Another point illustrating the pitfalls of
intuition: not every topological manifold admits of a differentiable structure.
In 1960, Kervaire had already found a 10-dimensional topological manifold
which accepts no complete differentiable atlas at all. In the eighties, a whole
family of non-smoothable compact simply-connected 4-dimensional manifolds was found. And, on the other hand, it has been found that every
non-compact manifold accepts at least one smooth structure.1
5.1.2 The above definitions concerning C -atlases and manifolds can be
extended to C k -atlases and C k -manifolds in an obvious way. It is also possible to relax the Hausdorff conditions, in which case, as we have already
said, the unicity of the solutions of differential equations holds only locally.
Broadly speaking, one could define differentiable manifolds without imposing
the Hausdorff condition, second-countability and the existence of a maximal
atlas. These properties are nevertheless necessary to obtain some very powerful results, in particular the Whitney theorems concerning the imbedding
of a manifold in other manifolds of higher dimension. We shall speak on these
theorems later on, after the notion of imbedding has been made precise.
5.1.3 A very important theorem by Whitney says that a complete C k -atlas
contains a C k+1 -sub-atlas for k 1. Thus, a C 1 -structure contains a C structure. But there is much more: it really contains an analytic sub-atlas.
The meaning here is the following: in the definition of a differentiable atlas,
replace the C -condition by the requirement that the coordinate transformations be analytic. This will define an analytic atlas, and a manifold with an
analytic complete atlas is an analytic manifold . The most important examples of such are the Lie groups. Of course not all C functions are analytic,
as the formal series formed with their derivatives as coefficients may diverge.
5.1.4 General references An excellent introduction is Boothby l975. A
short reference, full of illuminating comments, is the 5-th chapter of Arnold
l973. Nomizu 1956 is a very good introduction to the very special geometrical
properties of Lie groups. The existence of many distinct differentiable structures on E4 was found in l983. It is an intricate subject, the proof requiring
the whole volume of Freed & Uhlenbeck l984. Before proceeding to the main
onslaught, Donaldson & Kronheimer 1991 summarize the main results in a
nearly readable way. According to recent rumors (as of November 1994), a
suggestion of Witten has led people to obtain all the main results in a much
simpler way.
1
Quinn 1982.
5.2. SMOOTH FUNCTIONS
5.2
135
SMOOTH FUNCTIONS
5.2.1 In order to avoid constant repetitions when talking about spaces and
their dimensions, we shall as a rule use capital letters for manifolds and the
corresponding small letters for their dimensions: dim N = n, dim M = m,
etc.
5.2.2 A function f : N M is differentiable (or C k , or still smooth) if,
for any two charts (U, x) of N and (V, y) of M , the function
y f x<1> : x(U ) y(V )
is differentiable ( C k ) as a function between euclidean spaces.
Figure 5.1: Coordinate view of functions between manifolds.
5.2.3 Recall that all the analytic notions in euclidean spaces are presupposed. This function y f x<1> , taking an open set of En into an open set
of Em , is the expression of f in local coordinates. We usually write simply
y = f (x), a very concise way of packing together a lot of things. We should
keep in mind the complete meaning of this expression (see Figure 5.1): the
point of N whose coordinates are x = (x1 , x2 , . . . , xn ) in chart (U, x) is taken
by f into the point of M whose coordinates are y = (y 1 , y 2 , . . . , y m ) in chart
(V, y).
5.2.4 The composition of differentiable functions between euclidean spaces
is differentiable. From this, it is not difficult to see that the same is true for
functions between differentiable manifolds, because
136

z (g f ) x<1> = z g y <1> y f x<1> .
If now a coordinate transformation is made, say (U, x) (W, w) as in

Figure 5.1, the new expression of f in local coordinates is y f w<1> . Thus,
the function will remain differentiable, as this expression is the composition
y f x<1> x w<1>
of two differentiable functions: the local definition of differentiability given
above is extended in this way to the whole manifold by the complete atlas.
All this is easily extended to the composition of functions involving other
manifolds (as g f in Figure 5.2).
Figure 5.2: A function composition.
5.2.5 Each coordinate xi = ui is a differentiable function

xi : U N open set in E1 .
5.2.6 A most important example of differentiable function is a differentiable curve on a manifold: it is simply a smooth function from an open set
of E1 into the manifold. A closed differentiable curve is a smooth function
from the circle S 1 into the manifold.
5.3. DIFFERENTIABLE SUBMANIFOLDS
137
5.2.7 We have seen that two spaces are equivalent from a purely topological point of view when related by a homeomorphism, a topology-preserving
transformation. A similar role is played, for spaces with a differentiable
structure, by a diffeomorphism:
A diffeomorphism is a differentiable homeomorphism
whose inverse is also smooth.
5.2.8 Two smooth manifolds are diffeomorphic when some diffeomorphism
exists between them. In this case, besides being topologically the same, they
have equivalent differentiable structures. The famous result by Milnor cited
in the previous section can be put in the following terms: on the sphere S 7 one
can define 28 distinct smooth structures, building in this way 28 differentiable
manifolds. They are all distinct from each other because no diffeomorphism
exists between them. The same holds for the infinite differentiable manifolds
which can be defined on E4 .
5.2.9 The equivalence relation defined by diffeomorphisms was the starting point of an ambitious program: to find all the equivalence classes of
smooth manifolds. For instance, it is possible to show that the only classes
of 1-dimensional manifolds are two, represented by E1 and S 1 . The complete
classification has also been obtained for two-dimensional manifolds, but not
for 3-dimensional ones, although many partial results have been found. The
program as a whole was shown not to be realizable by Markov, who found
4-dimensional manifolds whose class could not be told by finite calculations.
5.3
DIFFERENTIABLE SUBMANIFOLDS
5.3.1 Let N be a differentiable manifold and M a subset of N . Then

M will be a (regular) submanifold of N if, for every point p M , there
exists a chart (U, x) of the N atlas, such that p U , x(p) = 0 En and
x(U M ) = x(U ) Em (as in Figure 5.3). In this case M is a differentiable
manifold by itself.
5.3.2 This decomposition in coordinate space is a formalization of the
intuitive idea of submanifold we get when considering smooth surfaces in
E3 . We usually take on these surfaces the same coordinates used in E3 ,
adequately restricted. To be more precise, we implicitly use the inclusion i:
Surface E3 and suppose it to preserve the smooth structure. Let us make
this procedure more general.
138
5.3.3 A differentiable function f : M N is an imbedding when

(i) f (M ) N is a submanifold of N ;
(ii) f : M f (M ) is a diffeomorphism.
Figure 5.3: M as a submanifold of N .
The above f (M ) is a differentiable imbedded submanifold of N . It corresponds precisely to our intuitive idea of submanifold, as it preserves globally
all the differentiable structure.
5.3.4 A weaker kind of inclusion is the following. A smooth function
f : M N is an immersion if, given any point p M , it has a neighbourhood U , with p U M , such that f restricted to U is an imbedding. An
immersion is thus a local imbedding and every imbedding is automatically an
immersion. The set f (M ), when f is an immersion, is an immersed submanifold . Immersions are consequently much less stringent than imbeddings. We
shall later ( 6.4.33 below) give the notion of ntegral submanifold.
5.3.5 These things can be put down in another (equivalent) way. Let us go back to
the local expression of the function f : M N (supposing n m). It is a mapping
between the euclidean spaces Em and En , of the type y f x<1> , to which corresponds
a matrix (y i /xj ). The rank of this matrix is the maximum order of non-vanishing
determinants, or the number of linearly independent rows. It is also (by definition) the
rank of y f x<1> and (once more by definition) the rank of f . Then, f is an immersion
iff its rank is m at each point of M . It is an imbedding if it is an immersion and else an
homeomorphism into f (M ). It can be shown that these definitions are quite equivalent to
those given above.
5.3. DIFFERENTIABLE SUBMANIFOLDS
139
5.3.6 The mapping f : E1 E2 given by

f (x) = (cos 2x, sin 2x)
is an immersion with f (E1 ) = S 1 E2 . It is clearly not one-to-one and so it
is not an imbedding. The circle f (E1 ) is an immersed submanifold but not
an imbedded submanifold.
5.3.7 The mapping f : E1 E3 given by the expression
f (x) = (cos 2x, sin 2x, x)
is an imbedding. The image space f (E1 ), a helix (Figure 5.4), is an imbedded
submanifold of E3 . It is an inclusion of E1 in E3 .
Figure 5.4: A helix is an imbedded submanifold of E3 .
5.3.8 We are used to think vaguely of manifolds as spaces imbedded in

some En . The question naturally arises of the validity of this purely intuitive
way of thinking, so convenient for practical purposes. It was shown by Whitney that an n-dimensional differentiable manifold can always be immersed in
E2n and imbedded in E2n+1 . The conditions of second-countability, completeness of the atlas and Hausdorff character are necessary to the demonstration.
These results are used in connecting the modern treatment with the so-called
classical approach to geometry (see Mathematical Topic 10). Notice that
eventually a particular manifold N may be imbeddable in some euclidean
manifold of lesser dimension. There is, however, no general result up to now
140
fixing the minimum dimension of the imbedding euclidean space of a differentiable manifold. It is a theorem that 2-dimensional orientable surfaces
are imbeddable in E3 : spheres, hyperboloids, toruses are perfect imbedded
submanifolds of our ambient space. On the other hand, it can be shown that
non-orientable surfaces without boundary are not, which accounts for our inability to visualize a Klein bottle. Non-orientable surfaces are, nevertheless,
imbeddable in E4 .
Part II
DIFFERENTIABLE
STRUCTURE
141
Chapter 6
TANGENT STRUCTURE
6.1
INTRODUCTION
In this chapter vector fields will be defined on differentiable manifolds, as

well as tensor fields and general reference frames (or basis). All these objects constitute the tangent structure of the manifold. Metrics, for instance,
are particular tensors. Vector fields provide the background to the modern
approach to systems of differential equations, which we shall ignore. They
also mediate continuous transformations on manifolds, which we shall examine briefly. Differential forms would actually belong here, but they are so
important and useful to Physics that a special chapter will be devoted to
them.
Think of an electron falling along the lines of force of an electric field created by some point charge. The lines of force are curves on the configuration
space M (just E3 in the usual case, but it will help the mind to consider some
curved space) and the electron velocity is, at each point, a vector tangent to
one of them. Being tangent to a curve in M , it is tangent to M itself. A
vector field is just that, a general object which reduces, at each point of the
manifold, to a tangent vector. The first thing one must become aware of is
that tangent vectors do not belong to M . At a fixed point p of M , they
may be added and rescaled, two characteristic properties of the members of
a linear (or vector) space, which in general M is not. They actually belong
to a linear space, precisely the tangent space to M at p. The differentiable
structure gives a precise meaning to such ideas.
6.1.1 Topological manifolds are spaces which, although generalizing their
properties, still preserve most of the mild qualities of euclidean spaces. The
differentiable structure has a very promising further consequence: it opens
the possibility of (in a sense) approximating the space M by a certain Em in
143
144
CHAPTER 6. TANGENT STRUCTURE
the neighbourhood of any of its points. Intuitively: the simplest euclidean

space is the real line E1 . A 1-dimensional smooth manifold will be, for
instance, a smooth curve on the plane E2 . In a neighbourhood of any point
(say, point A in Figure 6.1(a)), the curve can be approximated by a copy
of E1 . Recall the high-school definition of a vector on the plane: given two
points A and B, they determine the vector VAB = B A. If the tangent to
the curve at A is the line
a(t) = A + t(B A)
with parameter t, a(t) is the best linear approximation to in a neighbourhood of A = a(0). The vector
VAB =
da(t)
dt
can then be defined at the point A, since this derivative is a constant, and
will be a vector tangent to the curve at A. This high-school view of a
vector is purely euclidean: it suggests that point B lies in the same space
as the curve. Actually, there is no mathematical reason to view the tangent
space as attached to the point on the manifold (as in Figure 6.1(b)). This
is done for purely intuitive, picturesque reasons.
Figure 6.1: (a) A vector as difference between points: VAB = B A; (b)

Tangent space at p M seen as touching M at p, a mere pictorial resource.
6.2. TANGENT SPACES
145
6.1.2 Another characterization of this vector is, however, more convenient

for the generalization we intend to use in the following. Let f be a smooth
function on the curve , with real values: f [(t)] E1 . Then, (t) = a(t) in
a neighbourhood of A and
df da
df
df
=
= VAB
.
dt
da dt
da
Thus, through the vector, to every function will correspond a real number.
The vector appears in this case as a linear operator acting on functions.
6.1.3 A vector V = (v 1 , v 2 , . . . , v n ) in En can be viewed as a linear operator: take a point p En and let f be a function which is differentiable in
a neighbourhood of p. Then, to this f the vector V will make to correspond
the real number
f

2 f
n f
V (f ) = v 1 r
+
.
.
.
+
v
,
1 p + v
2
r p
rn p
the directional derivative of f along V at p. Notice that this action of V on
functions has two important properties:
(i) it is linear: V (f + g) = V (f ) + V (g) ;
(ii) it is a derivative, as it respects the Leibniz rule
V (f g) = f V (g) + g V (f ).
This notion of vector a directional derivative suits the best the
generalization to general differential manifolds. The set of real functions on
a manifold M constitutes with the usual operations of addition, pointwise
product and multiplication by real numbers an algebra, which we shall
indicate by R(M ). Vectors will act on functions, that is, they will extract
numbers from elements of R(M ). This algebra will, as a consequence, play
an important role in what follows.
6.2
TANGENT SPACES
6.2.1 A differentiable curve through a point p N is a differentiable curve

a : (1, 1) N such that a(0) = p. It will be denoted by a(t), with
t (1, 1). When t varies in this interval, a 1-dimensional continuum of
points is obtained on N . In a chart (U, ) around p these points will have
coordinates
ai (t) = ui [a(t)].
146
6.2.2 Let f be any differentiable real function on U 3 p, f : U E1 , as

in Figure 6.2. The vector Vp tangent to the curve a(t) at point p is given by
i
d
da
f
Vp (f ) =
[(f a)(t)]t=0 =
.
(6.1)
dt
dt t=0 ai
Figure 6.2: A curve maps E1 into N , and a real function proceeds in the
converse way. The definition of vector uses the notion of derivative on E1 .
Notice that Vp is quite independent of f , which is arbitrary. It is an

operator acting on the algebra R(N ) of real functions on N ,
Vp : R(N ) E1 .
An alternative way of introducing a tangent vector is the following: suppose two curves through p, a(t) and b(t) with a(0) = b(0) = p. They are
equivalent (intuitively: tangent to each other) at p if limt0 [a(t)b(t)]/t = 0.
This is indeed an equivalence relation (tangency) and is chart-independent.
The vector Vp is then defined as this equivalence class.
Now, the vector Vp , tangent at p to a curve on N , is a tangent vector to
N at p. In the particular chart used in eq. [6.1], (dai /dt) are the components
of Vp . Notice that, although the components are chart-dependent, the vector
itself is quite independent.
6.2.3 From its very definition, Vp satisfies:
(i) Vp (f + g) = Vp (f ) + Vp (g), , E1 ; f, g R(N );
(ii) Vp (f g) = f Vp (g) + g Vp (f ) (the Leibniz rule) .
The formal definition of a tangent vector on the manifold N at the point
p is then a mapping Vp : R(N ) E1 satisfying conditions (i) and (ii).
6.2. TANGENT SPACES
147
Multiplication of a vector by a real number gives another vector. The sum

of two vectors gives a third one. So, the vectors tangent to N at a point p
constitute a linear space, the tangent space Tp N of the manifold N at the
point p.
6.2.4 Given some LSC around p, with x(p) = (x1 , x2 , . . . , xn ), the operators {/xi } satisfy the above conditions (i) and (ii). They further span the
whole space Tp N and are linearly independent. Consequently, any vector can
be written in the form
Vp = Vpi i .
(6.2)
x
Notice that the coordinates are particular functions belonging to the algebra R(N ). The basis {/xi } is the natural basis, or coordinate basis
associated to the given LSC, alternatively defined through the conditions
xj
= ij .
i
x
(6.3)
The above Vpi are, of course, the components of Vp in this basis. If N =

E3 , eq. [6.2] reduces to the expression of the usual directional derivative
following the vector Vp = (Vp1 , Vp2 , Vp3 ), which is given by Vp . Notice
further that Tp N and En are finite vector spaces of the same dimension and are
consequently isomorphic. In particular, the tangent space to En at some point
will be itself an En . Differently from local euclidean character, which says
that around each point there is an open set homeomorphic to the topological
space En , Tp N is isomorphic to the vector space En .
6.2.5 In reality, euclidean spaces are diffeomorphic to their own tangent
spaces, and that explains part of their relative structural simplicity in
equations written on such spaces, one can treat indices related to the space
itself and to the tangent spaces (and to the cotangent spaces defined below)
on the same footing. This cannot be done on general manifolds, by reasons which will become clear later on. Still a remark: applying [6.2] to the
coordinates xi , one finds Vpi = Vp (xi ), so that
Vp = Vp (xi )
.
xi
(6.4)
6.2.6 The tangent vectors are commonly called simply vectors, or still
contravariant vectors. As it happens to any vector space, the linear mappings
p : Tp N E1 constitute another vector space, denoted here Tp N , the
dual space of Tp N . It is the cotangent space to N at p. Its members are
covectors, or covariant vectors, or still 1-forms. Given an arbitrary basis
148
{ei } of Tp N , there exists a unique basis {j } of Tp N , called its dual basis,

with the property j (ei ) = ij . Then, for any Vp Tp N ,
Vp = i (Vp )ei .
(6.5)
p = p (ei ) i .
(6.6)
For any p Tp N ,
6.2.7 The dual space is again an n-dimensional vector space. Nevertheless,
its isomorphism to the tangent space is not canonical (i.e, basis independent),
and no internal product is defined on Tp N . An internal product will be
present if a canonical isomorphism exists between a vector space and its
dual, which is the case when a metric is present.
6.2.8 Let f : M N be a C function between the differentiable manifolds M and N . Such a function induces a mapping
f : Tp M Tf (p) N
between tangent spaces (see Figure 6.3). If g is an arbitrary real function on
N , g R(N ), this mapping is defined by
[f (Xp )](g) = Xp (g f )
(6.7)
for every Xp Tp M and all g R(N ). This mapping is a homomorphism of vector spaces. It is called the differential of f , by extension of
the euclidean case: when M = Em and N = En , f is precisely the jacobian matrix. In effect, we can in this case use the identity mappings as
coordinate-homeomorphisms and write p = (x1 , x2 , . . . , xm ) and f (p) = y =
(y 1 , y 2 , . . . , y n ). Take the natural basis,
Xp (g f ) = Xpi
[g(y)]
xi
= Xpi
g y j
y j xi
Thus, the vector f [ (Xp )] is obtained from Xp by the (left-)product with the
jacobian matrix:
y j i
[f (Xp )]j =
X.
(6.8)
xi p
The differential f is also frequently written df . Let us take in the above
definition N = E1 , so that f is a real function on M (see Figure 6.4). As g
in [6.7] is arbitrary, we can take for it the identity mapping. Then,
f (Xp ) = df (Xp ) = Xp (f ).
(6.9)
6.2. TANGENT SPACES
149
Figure 6.3: A function f between two manifolds induces a function f between

their tangent spaces.
6.2.9 In a natural basis,

f
.
(6.10)
xi
Take in particular for f the coordinate function xj : M E1 . Then,
df (Xp ) = Xpi
dxj (Xp ) = Xp (xj ).
(6.11)
For vectors belonging to the basis,
) = ij .
(6.12)
xi
Consequently, the mappings {dxj } form a basis for the cotangent space,
dual to the natural basis {/xi }. Using [6.4], [6.10] and [6.11], one finds
dxj (
df (Xp ) =
f
xi
dxi (Xp ) .
As this is true for any vector Xp , we can write down the operator equation
f
dxi ,
(6.13)
xi
which is the usual expression for the differential of a real function. This is
the reason why the members of the cotangent space are also called 1-forms:
the above equation is the expression of a differential form in the natural
basis {dxj }, so that df Tp M . One should however keep in mind that, on a
general differentiable manifold, the dxj are not simple differentials, but linear
operators: as said by [6.11], they take a member of Tp M into a number.
df =
150
Figure 6.4: Case of a real function on M .
6.2.10 As Tp E1 E1 , it follows that df : Tp M E1 . If f is a real function

on the real line, that is, also M = E1 , then all points, vectors and covectors
reduce to real numbers.
6.2.11 Taking in [6.6] ei = /xi and i = dxi , we find the expression of
the covector in a natural basis,
p = p (
)dxi .
i
x
(6.14)
As we have already stressed, if f is a function between general differentiable

manifolds, f = df will take vectors into vectors: it is a vector-valued form.
We shall come back to such forms later on.
6.2.12 Contact with usual notation Let c : (a,
db) M be a differentiable curve on M . Given the point t0 (a, b), then dt t0 is a tangent vector
to (a, b) at t0 . It constitutes
by itself a basis for the tangent space Tt0 (a, b).
d
Consequently, c dt t0 is a vector tangent to M at c(t0 ) (see Figure 6.5).
Given an arbitrary function f : M E1 ,

c dtd t0 (f ) = dtd t0 (f c).
Let us take a chart (U, x) around c(t0 ), as in Figure 6.6, in which the points
c(t) will have coordinates given by x c(t) = (c1 (t), c2 (t), . . . , cm (t)). Then,
f c(t) = [f x<1> ] [x c(t)] = (f x<1> )(c1 (t), c2 (t), . . . , cm (t)).
6.2. TANGENT SPACES
Figure 6.5: c
151

d
dt t0
is a vector tangent to M at c(t0 ).
Notice that (f x<1> ) is just the local expression of f , with the identity
coordinate mapping in E1 . In the usual notation of differential
calculus, it is
simply written f . In that notation, the tangent vector c dtd t0 is fixed by
[ dtd f c(t)]t0 =
f dcj
|
cj dt t0
f
= cj (t0 ) c
j.
Figure 6.6: Using a chart to show the connection with usual notation.
The tangent vector is in this case the velocity vector of the curve at the
point c(t0 ). Notice the practical way of taking the derivatives: one goes first,
by inserting identity maps like x x<1> , to a local coordinate representation
of the function, and then derive in the usual way.
152
6.2.13 Up to isomorphisms, f takes Em into En . Suppose p is a critical

point of f . Then, at p, f = 0, as the jacobian vanishes (see Mathematical
Topic 9.5). Consider the simple case in which
M = S 1 = {(x, y) such that x2 + y 2 = 1} .
Let f be the projection
f (x, y) = y = (1 x2 )1/2 ,
as in Figure 6.7. The jacobian reduces to
dy
dx
= x(1 x2 )1/2 ,
whose critical points are (0, 1) and (0, 1). All the remaining points are
Figure 6.7: Two regular points corresponding to one regular value by a horizontal projection of the circle.
regular points of f . The points f (x, y) which are images of regular points are
the regular values of f . To the regular value P in Figure 6.7 correspond two
regular points: p and p0 . At p, the jacobian is positive; at p0 , it is negative. If
we go along S 1 starting (say) from (1, 0) and follow the concomitant motion
of the image by f , we see that the different signs of the jacobian indicate the
distinct senses in which P is attained in each point of the inverse image. The
sign of the jacobian at a point p is called the degree of f at p: degp f = +1;
6.2. TANGENT SPACES
153
degp0 f = 1. This is a general definition: the degree of a function at a

regular point is the jacobian sign at the point. Now, the degree of f at a
regular value is the sum of the degrees of all its counterimages:
deg P f = sum of degrees at f <1> (P ).
In the example above, we have that degP f = 0.
6.2.14 An important theorem says the following: given a C function
f : M N , then if M and N are connected and compact,
(i) f has regular values;
(ii) the number of counterimages of each regular value is finite ;
(iii) the degree of f at a regular value is independent of that regular value
.
The degree so defined depends solely on f . It is the Brouwer degree of f ,
denoted deg f .
6.2.15 Consider the case of two circles S 1 , represented for convenience on
the complex plane, and the function f : S 1 S 1 , f (z) = z n . All points are
regular in this case. All counterimages contain n points, each one of them
with the same degree: +1if n > 0, and 1if n < 0. As a consequence, deg
f = n. If we go along the domain S 1 and follow the corresponding motion
in the image space, we verify that the degree counts the number of times the
closed curve defined by f winds around the image space when the domain is
covered once. The Brouwer degree is also known as the winding number . It
gives the number n labelling the homotopy classes in 1 (S 1 ). There is some
variation in this nomenclature: the winding number is sometimes defined as
this purely topological number. As given above, it requires the differentiable
structure and, on the other hand, is more general as the function f is not
necessarily a loop. Including the differentiability requirement, however, does
add something, because of the following important result (a special case of a
theorem by Hopf):
two differentiable maps are homotopic if and only if
they have the same Brouwer degree.
Thus, the winding number characterizes the homotopy class of the mapping.
In the higher dimensional case, we have mappings S n S n , related to the
higher-order homotopy groups described in section 3.3. It is used to classify
magnetic monopoles1 and the vacua in gauge theories (Physical Topic 7.2.1).
1
Arafune, Freund & Goebel 1975.
154
6.3
TENSORS ON MANIFOLDS
6.3.1 Tensors at a point p on a differentiable manifold are defined as tensors on the tangent space to the manifold at p. Let us begin by recalling
the invariant (that is, basis-independent) definition of a tensor in linear algebra. Take a number r of vector spaces V1 , V2 , . . . , Vr and an extra vector
space W . Take the cartesian product V1 V2 . . . Vr . Then, the set
Lr (V1 , V2 , . . . , Vr ; W ) of all the multilinear mappings of the cartesian product of r-th order into W is itself another vector space.2 The special case
L1 (V ; W ), with W the field over which V is defined, is the dual space V of
V ; for a real vector space, V = L1 (V ; R).
6.3.2 Consider now the cartesian product of a vector space V by itself
s times, and suppose for simplicity V to be a real space. A real covariant
tensor of order s on V is a member of the space L(V V . . . V ; R), that
is, a multilinear mapping
Ts0 : V
{z. . . V} R.
| V
s times
By multilinear we mean, of course, that the mapping is linear on each copy
of V . On the other hand, given two such tensors, say T and S, the linear
(vector) structure of their own space is manifested by
(a T + b S)(v1 , v2 , . . . , vs ) = a T (v1 , v2 , . . . , vs ) + b S(v1 , v2 , . . . , vs ).
6.3.3 The tensor product T S is defined by
T S(v1 , v2 , . . . , vs , vs+1 , vs+2 , . . . vs+q )
= T (v1 , v2 , . . . , vs ) S(vs+1 , vs+2 , . . . vs+q ),
(6.15)
if T and S are respectively of orders s and q. The product T S is, thus,

a tensor of order (s + q). The product is clearly noncommutative in
general. We have already said that the space of all tensors on a vector space
is itself another linear space. With the operation , this overall tensor space
constitutes a noncommutative algebra (Mathematical Topic 1).
6.3.4 Let {i } be a basis for V , dual to some basis {ei } for V . A basis
for the space of covariant s-tensors can then be built as
{i1 i2 i3 . . . is } , 1 i1 , i2 , . . . , is
2
dim V .
Vector space, here, of course, in the formal sense of linear space.
6.3. TENSORS ON MANIFOLDS
155
In this basis, an s-tensor T is written

T = Ti1 i2 i3 is i1 i2 i3 . . . is
(6.16)
where summation over repeated indices is implied. The components Ti1 i2 i3 is

of the covariant s-tensor T are sometimes still presented as the tensor, a
practice mathematicians have abandoned a long while ago. Tensors as we
have defined above are invariant objects, while the components are basisdependent. On general manifolds, as we shall see later, the differentiable
structure allows the introduction of tensor fields, tensors at different points
being related by differentiable properties. Basis are extended in the same
way. On general manifolds, distinct basis of covector fields are necessary
to cover distinct subregions of the manifold. In this case, the components
must be changed from base to base when we travel throughout the manifold,
while the tensor itself, defined in the invariant way, remains the same. Consequently, it is much more convenient to use the tensor as long as possible,
using local components only when they may help in understanding some particular feature, or when comparison is desired with results known of old. In
euclidean spaces, where one coordinate system is sufficient, it is always possible to work with the natural basis globally, with the same components on
the whole space. The same is true only for a few very special kinds of space
(the toruses, for instance), a necessary condition for it being the vanishing
of the Euler characteristic (technically, they must be parallelizable, a notion
to be examined below (see 6.4.13).
6.3.5 In an analogous way, we define a contravariant tensor of order r: it
is a multilinear mapping
T0r : V
. . . V } R.
| V {z
r times
The space dual to the dual of a (finite dimensional) vector space is the space
itself3 : (V ) = V . Given a basis {ei } for V , a basis for the contravariant
r-tensors is
{ei1 ei2 ei3 . . . eir }, 1 i1 , i2 , . . . ir
dim V .
A contravariant r-tensor T can then be written

Unlike the isomorphism between V and V , the isomorphism between (V ) and V
is canonical (or natural), that is, basis independent. (V ) and V can consequently be
identified. This isomorphism (V ) = V , however, only exists for finite-dimensional linear
spaces.
3
156

T = T i1 i2 i3 ir ei1 ei2 ei3 . . . eir ,
the T i1 i2 i3 ir being its components in the given basis. Of course, the same
considerations made for covariant tensors and their components hold here.
6.3.6 A mixed tensor , covariant of order s and contravariant of order r, is
a multilinear mapping
Tsr : V
. . . V } R.
{z. . . V} |V V {z
| V
s times
r times
Given a basis {ei } for V and its dual basis {i } for V , a general mixed
tensor will be written
...ir
e ei2 ei3 . . . eir j1 j2 j3 . . . js . (6.17)
T = Tji11ji22ji33 ...j
s i1
It is easily verified that, just like in the cases for vectors and covectors, the
components are the results of applying the tensor on the basis elements:
...ir
= T (i1 , i2 , i3 , . . . , ir , ej1 , ej2 , ej3 , . . . , ejs ).
Tji11ji22ji33 ...j
s
(6.18)
Important particular cases are:

(i) the T00 , which are numbers;
(ii) the T01 , which are vectors;
(iii) the T10 , which are covectors.
We see that a tensor Tsr belongs to V1 V2 . . . Vr V1 V2 . . . Vs . Each copy
Vi of V acts as the dual of V , as the space of its linear real mappings, and vice-versa. A
r1
,
tensor contraction is a mapping of the space of tensors Tsr into the space of tensors Ts1
in which Vi is supposed to have acted on some Vk , the result belonging to
V1 V2 . . . Vi1 Vi+1 . . . Vr V1 V2 . . . Vk1

Vk+1
. . . Vs .
The components in the above basis will be (notice the contracted index j)
n n n ...n
jn
...n
3
i1
i+1
r
Tm11 m22 m
.
3 ...mk1 j mk+1 ...ms
6.3.7 Important special cases of covariant tensors are the symmetric ones,
those satisfying
T (v1 , v2 , . . . , vk , . . . , vj , . . .) = T (v1 , v2 , . . . , vj , . . . , vk , . . .)
157
for every j, k. An analogous definition leads to symmetric contravariant

tensors.
The space of all the symmetric covariant tensors on a linear space V can
be made into a commutative algebra in the following way. Call Sk (V ) the
linear space of covariant symmetric tensors of order k. The total space of
symmetric covariant tensors on V will be the direct sum S(V ) =
k=0 Sk (V ).
Now, given T Sk (V ) and W Sj (V ), define the product T W by
T W (v1 , v2 , . . . , vk+j ) =
X
1
T (vP (1) , vP (2) , . . . , vP (k) ) W (vP (k+1) , vP (k+2) , . . . , vP (k+j) ),
(k+j)!
{P }
the summation taking place over all the permutations P of the indices. Notice
that T W Sk+j (V ). With this symmetrizing operation, the linear space
S(V ) becomes a commutative algebra. The same can be made, of course, for
contravariant tensors.
6.3.8 An algebra like those above defined (see 6.3.3 and 6.3.7), which
is a sum of vector spaces, V ==
k=0 Vk , with the binary operation taking
Vi Vj Vi+j , is a graded algebra.
6.3.9 Let {i } be a basis for the dual space V . A mapping p : V E1
defined by first introducing
P V
j1
j2
j3
jk
P = dim
ji =1 Pj1 j2 j3 jk . . . ,
with Pj1 j2 j3 ...jk symmetric in the indices, and then putting
P
j1 j2 j3
jk
p(v) = P (v, v, . . . , v) = dimV
ji =1 Pj1 j2 j3 ...jk v v v . . . v
gives a polynomial in the components of the vector v. The definition is
actually basis-independent, and p is called a polynomial function of degree
k. The space of such functions constitutes a linear space Pk (V ). The sum of
these spaces, P (V ) =
k=0 Pk (V ), is an algebra which is isomorphic to the
algebra S(V) of 6.3.7.
6.3.10 Of special interest are the antisymmetric tensors, which satisfy
T (v1 , v2 , . . . , vk , . . . , vj , . . .) = T (v1 , v2 , . . . , vj , . . . , vk , . . .)
for every pair j, k of indices. Let us examine the case of the antisymmetric
covariant tensors. At a fixed order, they constitute a vector space by themselves. The tensor product of two antisymmetric tensors of order p and q is
158
a (p + q)-tensor which is no more antisymmetric, so that the antisymmetric

tensors do not constitute an algebra with the tensor product. We can however introduce another product which redresses this situation. Before that,
we need the notion of alternation Alt(T ) of a covariant tensor T of order s,
which is a tensor of the same order defined by
X
(sign P )T (vp1 , vp2 , . . . , vps ),
(6.19)
Alt(T )(v1 , v2 , . . . , vs ) = s!1
(P )
where the summation takes place on all the permutations P of the numbers (1, 2, . . . , s). Symbol sign P represents the parity of P . The tensor
Alt(T ) is antisymmetric by construction. If n is the number of elementary
transpositions (Mathematical Topic 2) necessary to take (1, 2, . . . , s) into
(p1 , p2 , . . . , ps ), the parity sign P = ()n .
6.3.11 Given two antisymmetric tensors, of order p and of order q,
their exterior product is the (p + q)-antisymmetric tensor given by
=
(p+q)!
p!q!
Alt( ).
This operation does make the set of antisymmetric tensors into an associative
graded algebra, the exterior algebra, or Grassmann algebra. Notice that only
tensors of the same order can be added, so that this algebra includes in reality
all the vector spaces of antisymmetric tensors. We shall here only list some
properties of real tensors which follow from the definition above:
(i)
(ii)
(iii)
(iv)
(v)
( + ) = +
( + ) = +
a ( ) = (a ) = (a ), for any a R;
( ) = )
(6.20)
= ()
In the last property, concerning commutation, and are the orders respectively of and .
If {i } is a basis for the covectors, the space of s-order antisymmetric
tensors has a basis
{i1 i2 i3 . . . is }, 1 i1 , i2 , . . . , is dim V .
An antisymmetric s-tensor can then be written
=
1
s!
i1 i2 i3 ...is i1 i2 i3 . . . is ,
(6.21)
159
the i1 i2 i3 ...is s being the components of in this basis. The space of antisymmetric s-tensors reduces automatically to zero for s > dim V .
Notice further that the dimension of the vector space formed by the antisymmetric covariant s-tensors is dims V . The dimension of the whole Grassmann algebra is 2dim V .
6.3.12 The exterior product is preserved by mappings between manifolds.
Let f : M N be such a mapping and consider the antisymmetric s-tensor
f (p) on the vector space Tf (p) N . The function f determines then a tensor
on Tp M through
(f )p (v1 , v2 , . . . , vs ) = f (p) (f v1 , f v2 , . . . , f vs ).
(6.22)
Thus, the mapping f induces a mapping f between the tensor spaces,
Figure 6.8: A function f induces a push-foward f and a pull-back f .

working however in the inverse sense (see the scheme of Figure 6.8): f is
suitably called a pull-back and f is sometimes called, by extension, pushfoward . To make [6.22] correct and well-defined, f must be C 1 . The pull-back
has the following properties:
(i)
(ii)
(iii)
f
is linear;
f ( ) = f f ;
(f g) = g f .
The pull-back, consequently, preserves the exterior algebra.
(6.23)
(6.24)
(6.25)
160
6.3.13 Antisymmetric covariant tensors on differential manifolds are called

differential forms. In a natural basis {dxj },
=
1
s!
j1 j2 j3 ...js dxj1 dxj2 dxj3 . . . dxjs .
The well defined behaviour when mapped between different manifolds renders
the differential forms the most interesting of all tensors. But of course we
shall come to them later on.
6.3.14 Let us now go back to differentiable manifolds. A tensor at a point
p M is a tensor defined on the tangent space Tp M . One
can
choose a jchart
around p and use for Tp M and Tp M the natural bases xi and {dx }. A
general tensor can be written
Tsr =
...ir
Tji11ji22ji33 ...j
s
i2 i3 . . . ir dxj1 dxj2 dxj3 . . . dxjs .

i
1
x
x
x
x
(6.26)
0
In another chart, the natural basis will be {/xi } and {dxj }, the same
tensor being written
0
0
0
j10
dxj2 . . . dxj3 dxjs
0
0
0
0 . . . dx
i
i
i
i
x r
x 1
x 2
x 3
0
0
0
i2
i3
i1
0
0
0
0
x
x
xir
xj1 xj2
xjs
i1 i2 i3 ...ir x
= Tj 0 j 0 j 0 ...js0
. . . j1 j2 . . . js
0
0
0
1 2 3
xi0r
x x
x
xi1
xi2
xi3
i2 i3 . . . ir . . . dxj1 dxj2 dxj3 . . . xjs , (6.27)

xi1
x
x
x
i0 i0 i0 ...i0
Tsr = Tj 01j 02j30 ...jrs0

1 2 3
which gives the transformation of the components under changes of coordinates in the charts intersection. Changes of basis unrelated to coordinate
changes will be examined later on. We find frequently tensors defined by
eq.[6.27]: they are those entities whose components transform in that way.
6.3.15 It should be understood that a tensor is always a tensor with respect to a given group. In [6.27], the group of coordinate transformations
is involved. General basis transformations (section 6.5 below) constitute
another group, and the general tensors above defined are related to that
group. Usual tensors in E3 are actually tensors with respect to the group of
rotations, SO(3). Some confusion may arise because rotations may be represented by coordinate transformations in E3 . But not every transformation
is representable through coordinates, and it is better to keep this in mind.
6.4. FIELDS & TRANSFORMATIONS
6.4
6.4.1
161
FIELDS & TRANSFORMATIONS
Fields
6.4.1 Let us begin with an intuitive view of vector fields. In the preceding
sections, vectors and tensors have been defined at a fixed point p of a differentiable manifold M . Although we have been lazily negligent
about this
aspect, the natural bases we have used are actually { xi p }. Suppose now
that we extend these vectors throughout the whole charts coordinate neighbourhood, and that the components are differentiable functions f i : M R,
f i (p) = Xpi . New vectors are then obtained, tangent to M at other points
of the coordinate neighbourhood. Through changes of charts, vectors can
eventually be got all over the manifold. Now, consider a fixed vector at p,
tangent to some smooth curve: it can be continued in the above way along
the curve. This set of vectors, continuously and differentiably related along
a differentiable curve, is a vector field. At p, Xp : R(M ) R. At different
points, X will map R(M ) into different points of R, that is, a vector field is
a mapping X : R(M ) R(M ). In this way, generalizing that of a vector,
one gets the formal definition of a vector field:
a vector field X on a smooth manifold M is a linear mapping
X : R(M ) R(M ) obeying the Leibniz rule:
X(f g) = f X(g) + g X(f ), f, g R(M ).
6.4.2 The tangent bundle A vector field is so a differentiable choice of
a member of Tp M for each p of M . It can also be seen as a mapping from M
into the set of all the vectors on M , the union T M = pM Tp M , with the
proviso that p is taken into Tp M :
X : M T M,
X : p Xp Tp M.
(6.28)
Given a function f R(M ), then (Xf )(p) = Xp (f ). In order to ensure

the correctness of this second definition, one should establish a differentiable
structure on the 2m-dimensional space TM. Let be a function
: T M M, (Xp ) = p,
to be called projection from now on.
6.4.3 As for covering spaces ( 3.2.15), open sets on T M are defined as
those sets which can be obtained as unions of sets of the type 1 (U ), with
162
U an open set of M (so that is automatically continuous). Given a chart

(V, x) on M , such that V 3 p, we define a chart for T M as (V , x) with
V
x : V
= 1 (V )
x(V ) Em ,
x : Xp (x1 (p), x2 (p), . . . , xm (p), Xp1 , Xp2 , . . . , Xpm ),
(6.29)
(6.30)
(6.31)
where Xpi are the components in

Xp =
Xpi
xi

.
p
, y) on M , with W
V 6= , the mapping y x<1> ,
Given another chart (W
defined by
y x<1> (x1 (p), x2 (p), . . . , xm (p), Xp1 , Xp2 , . . . , Xpm )
= (y 1 x<1> (x1 , x2 , . . . , xm ), y 2 x<1> (x1 , x2 , . . . , xm ), . . .
. . . y m x<1> (x1 , x2 , . . . , xm ), Yp1 , Yp2 , . . . , Ypm ),
where (see 6.2.8) YpI = (jacobian of y x<1> )ij Xpj is differentiable. The
two charts are, in this way, differentiably related. A complete atlas can in
this way be defined on T M , making it into a differentiable manifold. This
differentiable manifold T M is the tangent bundle, the simplest example of
a differentiable fiber bundle, or bundle space. The tangent space to a point
p, Tp M , is called, in the bundle language, the fiber on p. The field X itself,
defined by eq.[6.28], is a section of the bundle. Notice that the bundle space
in reality depends on the projection for the definition of its topology.
6.4.4 The commutator Take the field X, given in some coordinate neighbourhood as X = X i x i . As X(f ) R(M ), one could consider the action of
another field Y = Y j x j on X(f ):
2
i f
j i f
(X
)
+
Y
X
.
xj
xi
xj xi
This expression tells us that the operator Y X, defined by
Y Xf = Y j
(Y X)f = Y (Xf ),
does not belong to the tangent space, due to the presence of the last term.
This annoying term is symmetric in XY , and would disappear under antisymmetrization. Indeed, as easily verified, the commutator of two fields

j
j
i Y
i X
[X, Y ] := (XY Y X) = X
Y
(6.32)
i
i
x
x
xj
does belong to the tangent space and is another vector field.
163
6.4.5 ... and its algebra The operation of commutation defines on the
space T M a structure of linear algebra. It is also easy to check that
[X, X] = 0,
[[X, Y ], Z] + [[Z, X], Y ] + [[Y, Z], X] = 0,
the latter being the Jacobi identity. An algebra satisfying these two conditions is a Lie algebra. Thus, the vector fields on a manifold constitute, with
the commutation, a Lie algebra.
6.4.6 Notice that a diffeomorphism f preserves the commutator:
f [X, Y ] = [f X, f Y ].
Furthermore, given two diffeomorphisms f and g, (f g) X = f g X.
6.4.7 The cotangent bundle Analogous definitions lead to general tensor bundles. In particular, consider the union
T M = pM Tp M.
A covariant vector field, cofield or 1-form is a mapping
: M T M
such that (p) = p T M, p M .
6.4.8 This corresponds, in just the same way as has been seen for the
vectors, to a differentiable choice of a covector on each p M . In general,
the action of a form on a vector field X is denoted
(X) = < , X >,
so that
: T M R(M ).
(6.33)
6.4.9 In the dual natural basis (in other words, locally),

= j dxj .
(6.34)
Fields and cofields can be written respectively

X = dxi (X)
= (
= < dxi , X >

,
i
x
xi
) dxi = < , i > dxi .

i
x
x
(6.35)
(6.36)
164
6.4.10 The cofield bundle above defined is the cotangent bundle, or the
bundle of forms.
We shall see later (chapter 7) that not every 1-form is the differential of
a function. Those who are differentials of functions are the exact forms.
6.4.11 We have obtained vector and covector fields. An analogous procedure leads to tensor fields. We first consider the tensor algebra over a point
p M , consider their union for all p, topologize and smoothen the resultant
set, then define a general tensor field as a section of this tensor bundle.
6.4.12 At each p M , we have two m-dimensional vector spaces, Tp M
and Tp M , of course isomorphic. Nevertheless, their isomorphism is not natural (or canonical ). It depends on the chosen basis. Different basis fix
isomorphisms taking the same vector into different covectors. Only the presence of an internal product on Tp M (due for instance to the presence of a
metric, a case which will be seen later) can turn the isomorphism into a
natural one. By now, it is important to keep in mind the total distinction
between vectors and covectors.
6.4.13 Think of En as a vector space: its tangent vector bundle is a Cartesian product. A tangent vector to En at a point p can be completely specified
by a pair (p, V ), where also V is a vector in En . This comes from the isomorphism between each Tp En and En itself. Forcing a bit upon the trivial,
we say that En is parallelizable, the same vector V being defined at all the
different points of the manifold En . Given a general manifold M , it is said
to be parallelizable if its tangent bundle is trivial, that is, a mere Cartesian
product T M = M Em . In this case, a vector field V can be globally (that
is, everywhere on M ) given by (p, V ). Recalling the definition of a vector
field as a section on the tangent bundle, this means that there exists a global
section. Actually, the existence of a global section implies the triviality of the
bundle. This holds for any bundle: if some global section exists, the bundle
is a Cartesian product.
All toruses are parallelizable. Of all the spheres S n , only S 1 , S 3 and
7
S are parallelizable. The sphere S 2 is not a result sometimes called the
hedgehog theorem: you cannot comb a hairy hedgehog so that all its prickles
stay flat. There will be always at least one point like the crown of the head.
The simplest way to find out whether M is parallelizable or not is based on
the simple idea that follows: consider a vector V 6= 0. Then, the vector field
(, V ) will not vanish at any point of M . Suppose that we are able to show
that no vector field on M is everywhere nonvanishing. This would imply that
T M is not trivial. A necessary condition for parallelizability is the vanishing
165
of the Euler-Poincare characteristic of M . All Lie groups are parallelizable

differentiable manifolds.
6.4.14 Dynamical systems Dynamical systems are described in Classical Physics by vector fields in the phase space (q, q).
Consider the free fall
of a particle of unit mass under the action of gravity: call x the height and
y the velocity, y = x.
From
y = g (constant),
one gets the velocity in phase space (vx , vy ) = (y, g). A scheme of this
vector field is depicted in Figure 6.9.
Figure 6.9: Vector field scheme for y = g.

A classical system is completely specified by its velocity field in phase
space, which fixes its time evolution (Physical Topic 1). Initial conditions
simply choose one of the lines in the flow diagram. Well, we should perhaps qualify such optimistic statements. In general, this perfect knowledge
does not imply complete predictability. Small indeterminations in the initial
conditions may be so amplified during the system evolution that after some
time they cover the whole configuration space (see Mathematical Topic 3.2).
This happens even with a simple system like the double oscillator with noncommensurate frequencies. The above example is precisely the field vector
characterization of the system of differential equations
x = y ; y = g .
The modern approach to systems of differential equations is based on the
idea of vector field.4
4
A detailed treatment of the subject, with plenty of examples, is given in the little
masterpiece Arnold 1973.
166
6.4.15 Dynamical systems: maps Dynamical systems are also described, mainly in model building, by iterating maps like
xn+1 = f (xn ) ,
where x is a vector describing the state of some system. To help visualization,
we may consider n as a discrete time. The state at the n-th stage is given
by a function of the (n 1)-th stage, and so by its n-th iterate applied on
the initial seed state x0 . The set of points {xn } by which the system evolves
is the orbit of the system. An important concept in both the flow and the
map pictures is the following: suppose there is a compact set A to which the
sequence xn (or, in the flow case, the state when t becomes larger and larger)
converges for a given subset of the set of initial conditions. It may consist of
one, many or infinite points and is called an attractor . It may also happen
that A is a fractal, in which case it is a strange (or chaotic) attractor.5 This
is the case of the simple mapping
f : I I, I = [0, 1], f (x) = 4x(1 x),
popularly known as the logistic map, which for certain values of I
tends to a strange attractor akin to a Cantor set. Strange attractors are
fundamental in the recent developments in the study of chaotic behaviour in
non-linear dynamics.6
6.4.16 Let us go back to the beginning of this chapter, where a vector at
p M was defined as the tangent to a curve a(t) onM , with a(0) = p. It is
interesting to associate a vector to each point of the curve by liberating the
variation of the parameter t in eq.[6.1]:
Xa(t) (f ) =
d
(f a)(t).
dt
(6.37)
Then, Xa(t) is the tangent field to a(t), and a(t) is the integral curve of
X through p. In general, this is possible only locally, in a neighbourhood
of p. When X is tangent to a curve globally, the above definition being
extendable to the whole M , X is said to be a complete field . Let us for the
sake of simplicity take a neighbourhood U of p and suppose a(t) U , with
coordinates (a1 (t), a2 (t), . . . , am (t)). Then, from [6.37],
Xa(t) =
5
6
dai
.
dt ai
(6.38)
See Farmer, Ott & Yorke 1983, where a good discussion of dimensions is also given.
For a short review, see Grebogi, Ott & Yorke 1987.
167
Thus,
dai
(6.39)
dt
i
. In this sense, the field whose integral curve is a(t)
is the component Xa(t)

da
is given by dt . In particular, Xp = da
. Conversely, if a field is given by
dt t=0
k 1
2
m
its components X (x (t), x (t), . . . , x (t)) in some natural basis, its integral
curve x(t) is obtained by solving the system of differential equations X k =
dxk
. The existence and unicity of solutions for such systems is in general
dt
valid only locally.
Xa(t) (ai ) =
6.4.2
Transformations
Let us now address ourselves to what happens to differentiable manifolds

under infinitesimal transformations, to which vector fields in a way preside.
More precisely, we examine the behaviour of general tensors under continuous
transformations. The basic tool is the Lie derivative, which measures the
variation of a tensor when small displacements take place on the manifold.
We start with the study of 1-dimensional displacements along a field (local)
integral curve.
6.4.17 The action of the group R of the real numbers on the manifold M
is defined as a differentiable mapping
:RM M
: (t, p) (t, p)
satisfying
(i) (0, p) = p;
(ii) (t + s, p) = (t, (s, p)) = (s, (t, p)),
for all p M , and all s, t R.
6.4.18 At fixed t, (t, p) is a mapping
t : M M
t : p (t, p),
a collective displacement of all the points of M . At fixed p, it is a mapping
p : R M
p : t (t, p),
which for each p M describes a curve (t) = p (t), the orbit of p generated
by the action of the group R. The mapping is a 1-parameter group on M .
168
6.4.19 The action so defined is a particular example of actions of Lie

groups (of which R is a case) on manifolds. We shall see later (section
8.2) the general case. Notice that, being 1-dimensional, group R is abelian.
Mathematicians use to call, by a mechanical analogy, M the phase space,
the flow , and R M the enlarged phase space. Due to the group character,
it can be shown that only one orbit goes through each point p of M .
6.4.20 Take a classical mechanical system and let its phase space M (see Physical
Topic 1) be specified as usual by the points (q, p) = (q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ). The
time evolution of the system, if the hamiltonian function is H(q, p), is governed by the
hamiltonian flow , which for a conservative system is
(q0 ,p0 ) (t) = etH(q0 ,p0 ) ;
(q0 , p0 ) (qt , pt ) .
Given a domain U M , the Liouville theorem says that the above flow preserves
its volume: vol [(t)U ] = vol [(0)U ]. Suppose now that M itself has a finite volume.
Then, after a large enough time interval, forcibly ((t)U ) U 6= . In words: given any
neighbourhood U of a state point (q, p), it contains at least one point which comes back
to U for t > some tr . For large enough periods of time, a system comes back as near as
one may wish to its initial state. This is Poincares theorème du retour.
6.4.21 Let M = E3 , and x = (

x1 , x2 , x3 ) a fixed point different from zero.
Then,
t (x) = (x1 + x1 t, x2 + x2 t, x3 + x3 t)
defines a C action of R on M . For each t R, t : E3 E3 is a translation
taking x into x + xt. Each vector x determines a translation. The orbits are
the straight lines parallel to x.
6.4.22 To each group corresponds a vector field: the infinitesimal operator (or generator ) of is the field X defined by

Xp f = lim
t0

1
[f (p (t)) f (p)] ,
t
(6.40)
on each p M and arbitrary f R(M ). A field X is thus a derivation along

the differentiable curve (t) = p (t), which is its integral curve.
169
With q = p (t0 ), we have the following:

p (t0 )f = p
d
dt
!
t0
d
f=
dt

(f p )
t0

1
= lim
[f p (t0 + t) f p (t0 )]
t0
t
1
= lim
{f [(q + t)] f (q)}
t0 t
1
{f [q (t)] f (q)} = Xp (t0 ) f.
lim
t0 t
As f is any element of R(M ), we have indeed
p (t0 ) = Xp (t0 ) .
(6.41)
The above definition generalizes to manifolds, though only locally, the well
known case of matrix transformations engendered by an invertible matrix
g(t) = etX , of which the matrix X is the generator,

X = dg
= XetX |t=0 .
dt t=0
A matrix Y will transform according to
Y 0 = g(t)Y g 1 (t) = etX Y etX (1 + tX)Y (1 tX) Y + t[X, Y ]
to first order in t, and we find that the first derivative is
[X, Y ] = limt0
1
t
{g(t) Y g 1 (t) Y } .
6.4.23 Take M = E2 and : R M M given by (t, (x, y)) = (x + t, y),

d
translations along the x-axis. The infinitesimal operator is then X = dx
.
6.4.24 We have seen that, given the action , we can determine the field
X which is its infinitesimal generator. The inverse is not true in general
but holds locally: every field X generates locally a 1-parameter group. The
restriction is related to the fact that to find out the integral curve we have to
integrate differential equations ( 6.4.16), for which the existence and unicity
of solutions is in general only locally granted.
6.4.25 Lie derivative In section 6.2 we have introduced the derivative of
a differentiable function f along the direction of a vector X: df (Xp ) = Xp f .
It was a generalization to a manifold M of the directional derivative of a
function on Em . Things are a bit more complicated when we try to derive
170
more general objects. We face, to begin with, the problem of finding the
variation rate of a vector field Y at p M with respect to Xp . This can be
done by using the fact that X generates locally a 1-parameter group, which
induces an isomorphism t : Tp M Tt (p) M , as well as its inverse t .
It becomes then possible to compare values of vector fields. We shall just
state three different definitions, which can be shown to be equivalent. The
Lie derivative of a vector field Y on M with respect to the vector field X on
M , at a point p M , is given by any of the three expressions:

(LX Y )p = limt0 1t t (Y(t,p) ) Yp

1
Yp t (Y(t,p) )
t0 t

= dtd {t (Y )} t=0 .
= lim
(6.42)
Each expression is more convenient for a different purpose. Notice that the
vector character of Y is preserved by the Lie derivative: LX Y is a vector
field.
Let us examine the definition given in the first equality of eqs.[6.42] (see
Figure 6.10). The action p (t) induces an isomorphism between the tangent
spaces Tp M and Tq M , with q = p (t). By this isomorphism, Yp is taken into
p (t) (Yp ), which is in general different from Yq , the value of Y at q. By
using the inverse isomorphism p (t) we bring Yq back to Tp M . In this last
vector space we compare and take the limit. As it might be expected, the
same definition can also be shown to reduce to
LX Y = [X, Y ].
One should observe that the concept of Lie derivative does not require any
extra structure on the differentiable manifold M . Given the differentiable
structure, Lie derivatives are automatically present.
6.4.26 Let us consider, as shown in Figure 6.11, a little planar model for
the water flow in a river of breadth 2a: take the velocity field X = (a2 y 2 )e1 ,
and e2 = y
. It generates the 1-parameter group
with e1 = x
X (t, p) = (a2 y 2 )te1 + p,
or
X (t, p) = [x + (a2 y 2 )t]e1 + ye2 ,
with p = (x, y). The flow leaves the border points invariant. Consider now a
constant transversal field, Y = e2 . It generates the group
171
Figure 6.10: Scheme for the Lie derivative.
Y (s, p) = se2 + p = xe1 + (s + y)e2 .

A direct calculation shows that
(Y X X Y )(p) = s t (s + 2y)e1
or, to the lowest order, (2sty)e1 . The commutator [X, Y] is precisely 2ye1 ,
with the group
[X,Y ] (r, p) = 2 y r e1 + p = (x + 2 y r)e1 + ye2 .
From another point of view: examine the effect of X : Tp M TX (t,p) M :
X (Yp )(f (x, y)) = Yp (f X ) =
h
= 2yt x
+
[f (x + (a2 y 2 )t, y)]
f = t [X, Y ]f + Y f,
so that
1t {X (t, p)(Yp ) Yp } (f ) = [X, Y ]f ,
which is the expression of the third definition in [6.42]. Thus, the Lie derivative turns up when we try to find out how a field Y experiences a small
transformation generated by another field X.
172
Figure 6.11: A planar model for a river flow.
6.4.27 Consider on the plane (Figure 6.12) the two fields X = e1 (con
stant) and Y = xe2 , again with e1 = x

and e2 = y
. The integral curves
are: X (s) = s along e1 and Y (t) = xt along e2 . The groups generated by
X and Y are:
X (s, p) = p + se1 = (x + s)e1 + ye2 ;
Y (t, p) = p + xte2 = xe1 + (y + xt)e2 .
The Lie derivative measures the non-commutativity of the corresponding
groups. We check easily that
X [s, Y (t, p)] Y [t, X (s, p)] = st e2 .
On the other side, [X,Y ] (r, p) = re2 + p. We have drawn these transformations in Figure 6.13, starting at point p = (2, 1), and using s = 2, t = 1.
The difference is precisely that generated by the above commutator.
6.4.28 Lie derivatives are a vast subject.7 We can here only list some of
their properties:
(i) commutator:
[LX , LY ] = L[X,Y ] ;
(6.43)
[[LX , LY ], LZ ] + [[LZ , LX ], LY ] + [[LY , LZ ], LX ] = 0,
(6.44)
(ii) Jacobi identity:
For more details, see for instance Schutz 1985.
173
Figure 6.12: Scheme for fields X = e1 and Y = xe2 .
(iii) function multiplying field:

LX (f Y ) = (LX f )Y + f LX Y ;
(iv) In a natural basis {i =
},
xi
(6.45)
they satisfy
Lj (Y ) =
Y i
;
xj xi
(6.46)
the Lie derivative appears as a coordinate-independent version of the partial

derivative; (v) take a basis {ei }, in which X = X i ei , Y = Y j ej . Then,
LX Y = X(Y i )ei Y (X i )ei + X i Y j Lej ei .
(6.47)
6.4.29 The Lie derivative of a covector field is defined by

(LX )Y = LX ((Y )) (LX Y ).
(6.48)
Thus,
(LX )Y = X((Y )) ([X, Y ]).
6.4.30 This comes out as a consequence of the general definition for the
Lie derivative of a tensor field of any kind, which is (cf. eq.[6.42])
(LX T )p == lim
t0

1
Tp t (T(t,p) ) .
t
(6.49)
The maps induced by the 1-parameter group are to be taken as push-forward

and/or pull-backs, according to the contravariant and/or covariant character
of the tensor. Applied to a function f , this gives simply X(f ).
174
Figure 6.13: Particular transformations for the previous scheme.
Once the action of the Lie derivative is known on fields and cofields, the
general definition is given as
(LX T )(Y1 , Y2 , . . . Ys , 1 , 2 , . . . r )
= LX [T (Y1 , Y2 , . . . Ys , 1 , 2 , . . . r )]
T (LX Y1 , Y2 , . . . , Ys , 1 , 2 , . . . r )
T (Y1 , LX Y2 , . . . Ys , 1 , 2 , . . . r ) . . .
T (Y1 , Y2 , . . . , LX Ys , 1 , 2 , . . . r )
T (Y1 , Y2 , . . . , Ys , LX 1 , 2 , . . . r )
T (Y1 , Y2 , . . . Ys , 1 , LX 2 , . . . r ) . . .
T (Y1 , Y2 , . . . , Ys , 1 , 2 , . . . , LX r ). (6.50)

6.4.31 Notice that LX preserves the tensor character: it takes an rs
tensor into another tensor of the same type. In terms of the components: in
a natural basis {i } the components of LX T are
ab...r
a
ib...r
b
ai...r
(LX T )ab...r
ef ...s = X(Tef ...s ) (i X )Tef ...s (i X )Tef ...s . . .
ab...i
i
ab...r
i
ab...r
i
ab...r
(i X r )Tef
...s + (e X )Tif ...s ) + (f X )Tei...s ) . . . + (s X )Tei...i ). (6.51)
6.4.32 The Lie derivative LX provides the infinitesimal changes of tensorial objects under the 1-parameter group of transformations of which X is the
generator. For this reason, Lie derivatives are basic instruments in the study
of transformations imposed on differentiable manifolds (automorphisms) and,
a fortiori, in the study of symmetries (see section 8.2).
6.5. FRAMES
175
6.4.33 We have seen in 6.4.16 that, given a vector field X on a manifold

M , there exists locally a curve on M which integrates it. Thus, there is a
1-dimensional manifold which is tangent to X. Unless the field is complete,
the curve can only be an immersed submanifold (remember what has been
said in section 5.3). We may consider many fields at a time, and ask for the
general condition to relate fields to an imbedded submanifold N of M . To
begin with, a submanifold is a manifold, so that, if X and Y are tangent to
N , so must be [X, Y ] ( 6.4.4). Consider then a set of n( m) fields tangent
to M . At each point p, they will generate some subspace of Tp M . If they
are linearly independent, they generate a subspace of dimension n. Suppose
this linear independence holds for all p M . Such an assignment of an
n-dimensional subspace Dp M of Tp M for each p M is called a distribution
(not to be mistaken by singular functions!). If around each p there is an open
set U and fields X1 , X2 , . . . , Xn forming a basis for Dq M for all q U , then
the distribution is said to be a differentiable distribution . The distribution is
said to be involutive if it contains the commutator of every pair of its fields.
Suppose now that N is an imbedded submanifold of M , with i : N M
being the imbedding. The N is an integral manifold of the distribution if
i (Tp N ) = Dp M for all p M . When there is no other integral manifold
containing N , N is a maximal integral manifold. This gives the complete
story of the relationships between the spaces tangent to a manifold and the
spaces tangent to a submanifold. Now there is a related and very strong
result, concerning integrability around each point, the Frobenius theorem:
given an involutive differentiable distribution on a manifold M , then
through every point p M there passes a (unique) maximal integral
manifold N (p), such that any other integral manifold through p is a
submanifold of N (p).
Thus, the main condition for local integrability is the involutive character
of the field set: all things being differentiable, there is a local submanifold
around the point p whenever at p the fields do close a Lie algebra.
6.5
FRAMES
6.5.1 Given a differentiable manifold M and an open set U M , a set

{Xi } of m vector fields is a local basis of fields (or local frame) if, for any
p U , {X(p)i } is a basis for Tp M . This means that each X(p)i is a tangent
vector to M at p and that the X(p)i s are linearly independent. In principle,
any set of m linearly independent fields can be used as a local basis. For
some manifolds there exists a global basis. For most, only local bases exist.
176
6.5.2 In particular, around every p M

there is a chart (U, x) and the
set of fields { xi } : U T U , p { xi p } forms a basis. Field bases of

this kind, directly related to local coordinates, are called holonomous (or
holonomic) bases, or coordinate bases. A condition for a basis {Xi } to be
holonomous is that, for any two of its members, say Xj and Xk ,
[Xj , Xk ](f ) = 0
for all f R(M ). Of course, this happens for { x i } but it should be clear
that this property is exceptional: most bases do not consist of all-commuting
fields, and are called anholonomic, or non-coordinate bases.
6.5.3 Take for example the ordinary spherical coordinates (r, , ) in E3 .
The related holonomous basis is (r , , ). We have seen that in E3 a vector
is precisely the directional derivative; nevertheless, this basis does not give
the usual form of the gradient. The velocity, for example, would be
V = V r r + V + V
with components
Vr =
dr
;
dt
V =
d
;
dt
V =
d
.
dt
Usually, however, the velocity components are taken to be

Vr =
dr
;
dt
V = r d
;
dt
V = r sin d
,
dt
which correspond to the anholonomous basis

Xr =
;
r
X =
1
;
r
X =
.
r sin
(6.52)
These fields do not all commute with each other.

6.5.4 We have seen that the commutator of two fields is another field. We
can expand the commutator of two members of an anholonomic basis in that
same basis,
[Xi , Xj ] = C k ij Xk ,
(6.53)
where the Cijk s are called the structure coefficients of the basis algebra. For
the above spherical basis the non-vanishing coefficients are
C r = C r = - 1r ; C = -
1
r sin
6.5. FRAMES
177
and their permutations in the lower indices (in which the coefficients are
clearly antisymmetric). Notice: the coefficients are not necessarily constant
and depend on the chosen basis. Clearly, a necessary condition for the basis
to be holonomic is that Cijk = 0 for all commutators of the basis members.
This condition, Cijk = 0 for all basis members, may be shown to be also
sufficient for holonomy. The Jacobi identity, required by the Lie algebra,
implies
C n kl C i jn + C n jk C i ln + C n lj C i kn = 0.
(6.54)
6.5.5 Let us re-examine the question of frame transformations. Given two
natural basis on the intersection of two charts, a field X will be written
0
X = X i x i = X i
.
xi0
The action of X on the function X j leads to

0
xj i
X =
X.
xi
j0
(6.55)
This expression gives the way in which field components in natural bases
change when these bases are themselves changed. Here, basis transformations
are intimately related to coordinate transformations. However, other basis
transformations are possible: for example, going from the holonomic basis
(r , , ) to the basis (Xr , X , X ) of eq.[6.52] in the spherical case above
is a basis transformation unrelated to a change of coordinates.
6.5.6 Given an anholonomous basis {Xi }, it will always be possible to
write locally each one of its members in some coordinate basis as
Xi = Xij
.
xj
By using the differentiable atlas, the components can be in principle obtained

all over the manifold. Each change of natural basis will give new components
according to
k0
j x
k0
Xi = Xi
.
(6.56)
xj
Notice that basis {Xi } would be holonomous only if Xij =
xj
,
y i
where {y i } is
xj
some other coordinate system. In that case, {Xi = yi }. General matrices

(Xij ) are not of this form, and an holonomous basis is more of an exception
than a rule. More generally, a basis transformation will be given by
0
Xik = Xij Aki ,
(6.57)
178
where A is some matrix. Notice that each basis is characterized by the

matrix (Xik ) of its components in some previously chosen basis. Just above,
a natural basis was chosen. The tangent spaces, being isomorphic to Em ,
possess each one a canonical basis of the type
v1 = (1, 0, 0, . . . , 0), v2 = (0, 1, 0, 0, . . . , 0), . . . , vm = (0, 0, 0, . . . , 1).
The important point is that we can choose some starting basis from which
all the other basis are determined by the matrices of their components. Such
m m matrices belong to the general linear space of m m real matrices. As
they are forcibly non-singular (otherwise the linear independence would fail
and we would have no basis) and consequently invertible, they constitute the
linear group GL(m, R). Starting from one basis we obtain each other basis
in this way, one basis for each transformation, one basis for each element of
the group. The set of all basis at each point p M is thus isomorphic to
the linear group. But the transformation matrices A of eq.[6.57] also belong
to the group, so that we have a case of a group acting on itself. Due to the
peculiar form of the action shown in [6.57], we say that the transformations
act on the right on the field basis, or that we have a right-action of the group.
The frequent use of natural basis (in general more convenient for calculations)
is responsible for some confusion between coordinate transformations and
basis transformations, which are actually quite distinct.
6.5.7 The case of covector field basis is analogous. Two natural basis are
related by
0
xj
0
dxi .
(6.58)
dxj =
xi
The elements of another basis {i } can be written as i = ji dX j and will
transform according to
xk i
i
j 0 =
.
(6.59)
xj 0 k
Under a general transformation,
ji 0 = Akj0 ki ,
(6.60)
so that the group of transformations acts on the left on the 1-form basis. Dual
basis transform inversely to each other, so that, under the action, the value
< , X > is invariant. That is to say that < , X > is basis-independent.
6.5.8 The bundle of linear frames Let Bp M be the set of all linear basis
for Tp M . As we have said, it is a vector space and a group, just GL(m, R).
In a way similar to that used to build up T M as a manifold, the set
6.5. FRAMES
179
BM = pM Bp M
of all the basis on the manifold M can be viewed as a manifold. To begin

with, we define a projection : BM M , with ({X(p)i } Bp M ) = p.
A topology is defined on BM by taking as open sets the sets 1 (U ), for
U open set of M . Given a chart (U, x) of M , a basis at p U is given
m
), where Xij is the j-th
by (x1 , x2 , . . . , xm , X11 , X12 . . . , X1m , X21 . . . X2m . . . Xm
component of the i-th basis member in the natural basis. This gives the
(m + m2 ) coordinates of a point on BM . It is possible to show that the
mapping U GL(m, R) Em+m2 is a diffeomorphism. Consequently, BM
becomes a smooth manifold, the bundle of linear frames on M. We arrive
thus to another fundamental fiber bundle. Let us list some of its characteristics:
(i) the group GL(m, R) acts on each Bp M on the right (see eq.[6.57]);
Bp M is here the fiber on p; this group of transformations is called the structure group of the bundle;
(ii) BM is locally trivial in the sense that every point p M has a neighbourhood U such that 1 (U ) is diffeomorphic to U GL(m, R).
(iii) concerning dimension: dim BM = dim M + dim GL(m, R) = m +
m2 .
6.5.9 The fiber itself is GL(m, R). A fiber bundle whose fiber coincides
with the structure group is a principal fiber bundle. A more detailed study of
bundles will be presented later on. Let us here only advance another concept.
The tangent bundle has the spaces Tp M as fibers. The action of GL(m, R)
on the basis can be thought of as an action on Tp M itself: it is the group
of linear transformations, taking a vector into some other. A bundle of this
kind, on whose fibers (as vector spaces) the same group acts, is said to be an
associated bundle to the principal bundle. Most common bundles are vector
bundles on which some group acts. The main interest of principal bundles
comes from the fact that properties of associated bundles are deducible from
those of the principal bundle.
Coordinates, which are in general local characterizations of points on a
manifold, are usually related to a local frame. One first chooses a frame at a
certain point, consider the euclidean tangent space supposing it as glued
to the manifold at the point, make its origin as a vector space (that is, the
zero vector) to coincide with the point, then introduce cartesian coordinates,
and finally move to any other coordinate system one may wish. By a change
180
of frame,
the set of coordinates will transform according to x0 = Ax, or xj
j0 i
= Ai x , as any contravariant vector. This leads to
0
dxj = dAji xi + Aji dxi .

Many physical problems involve comparison of rates of change of vector quantities in two different frames (recall for example the case of the body and
the space frames in the rigid body motion, Physics Topic 2, from section
2.3.5 on). Consider a general vector u, with
0
duj = dAji ui + Aj dui .

The rate of change with a parameter t (usually time) will be
duj
dt
dAji
dt
u i + Aj
dui
.
dt
A velocity, for example, as seen from two frames, will have its components
related by
0
dAji i
j0 i
j0
v.
v = Ai v +
dt
Of course, we are here supposing that also the frames are in relative motion.
We shall come back to such moving frames later ( 7.2.17 , 7.3.12 and
9.3.6).
6.6
METRIC & RIEMANNIAN MANIFOLDS
The usual 3-dimensional euclidean space E3 consists of the set R3 of ordered

triples plus the topology defined by the 3-dimensional balls. Such balls are
defined through the use of the euclidean metric, a tensor whose components
are, in the global cartesian coordinates, constant and given by gij = ij . We
may thus say that E3 is R3 plus the euclidean metric. We use precisely this
metric to measure lengths in our everyday life. It happens frequently that another metric is simultaneously at work on the same R3 . Suppose, for example,
that the space is permeated by a medium endowed with a point-dependent
refractive index (that is, a point-dependent electric and/or magnetic permeability) n(p). Light rays (see Physical Topic 5) will in this case feel another
metric, which will be gij0 = n2 (p)ij if n(p) is isotropic. To feel means that
they will bend, acquire a curved aspect if looked at by euclidean eyes (like
ours). Light rays will become geodesics of the new metric, the straightest
possible curve if measurements are made using gij0 instead of gij . As long as
we proceed to measurements using only light rays, distances will be different
from those given by the euclidean metric. Suppose further that the medium
6.6. METRIC & RIEMANNIAN MANIFOLDS
181
is some compressible fluid, with temperature gradients and all which is necessary to render point-dependent the derivative of the pressure with respect
to the
density at fixed entropy. The sound velocity will be given by
fluid
p
and the sound propagation will be governed by geodesics of still
c2s =
S
another metric, gij00 = c12 ij . Nevertheless, in both cases we use also the euS
clidean metric to make measurements, and much of geometrical optics and
acoustics comes from comparing the results in both metrics involved. This is
only to call attention to the fact that there is no such a thing like the metric
of a space. It happens frequently that more than one is important in a given
situation (for an example in elasticity, see Physical Topic 3, section 3.3.2).
Let us approach the subject a little more formally.
6.6.1 In the space of differential forms, a basis dual to the basis {Xi } for
fields in T M is given by those j such that
j (Xi ) = < j , Xi > = j i,
(6.61)
so that = < , Xj > j . Given a field Y = Y i Xi and a form z = zj j ,

< z, Y > = zj Y i .
(6.62)
6.6.2 Bilinear forms are covariant tensors of second order, taking T M

T M into R(M ). Recall that the tensor product of two linear forms w and z
is defined by
(w z)(X, Y ) = w(X) z(Y ).
(6.63)
Given a basis { j } for the space of 1-forms, the products i j , with
i, j = 1, 2, . . . , m, form a basis for the space of covariant 2-tensors, in terms
of which a bilinear form g is written
g = gij i j .
(6.64)
g = gij dxi dxj .
(6.65)
Of course, in a natural basis,
The most fundamental bilinear form appearing in Physics is the Lorentz

metric on R4 , which defines Minkowski space and whose main role is to endow
it with a partial ordering, that is, causality.8
6.6.3 A metric on a smooth manifold is a bilinear form, denoted g(X, Y ),
X Y or < X, Y >, satisfying the following conditions:
(i) of course, it is bilinear :
8
See Zeeman 1964.
182

X (Y + Z) = X Y + X Z
(X + Y ) Z = X Z + Y Z;
(ii) it is symmetric:
X Y = Y X;
(iii) it is non-singular :
if X Y = 0 for every field Y , then X = 0.
6.6.4 In the basis introduced in 6.6.2, we have
g(Xi , Xj ) = Xi Xj = gmn m (Xi ) n (Xj ),
so that
gij = g(Xi , Xj ) = Xi Xj .
(6.66)
The relationship between metrics and general frames (in particular, tetrads)
will be seen in 9.3.6. As gij = gji and we commonly write simply i j for
the symmetric part of the bilinear basis, then
i j = (i j) =
1
2
( i j + j i ) ,
we have
g = gij i j
(6.67)
g = gij dxi dxj .
(6.68)
or, in a natural basis,
6.6.5 This is the usual notation for a metric. Notice also the useful symmetrizing notation (ij) for indices. All indices (ijk . . .) inside the parenthesis
are to be symmetrized. For antisymmetrization the usual notation is [ijk . . .],
meaning that all the indices inside the brackets are to be antisymmetrized.
Knowledge of the diagonal terms is enough: the off-diagonal may be obtained
by polarization, that is, by using the identity
g(X, Y ) = 12 [g(X + Y, X + Y ) g(X, X) g(Y, Y )].
6.6.6 A metric establishes a relation between vector and covector fields:
Y is said to be the contravariant image of a form z if, for every X,
g(X, Y ) = z(X).
183
If, in the dual bases {Xi } and { j }, Y = Y i Xi and z = zj j , then gij Y j = zi .

In this case, we write simply zj = Yj . That is the usual role of the covariant
metric, to lower indices, taking a vector into the corresponding covector. If
the mapping Y z so defined is onto, the metric is non-degenerate. This
is equivalent to saying that the matrix (gij ) is invertible. A contravariant
metric g can then be introduced whose components are the elements of the
matrix inverse to (gij ). If w and z are the covariant images of X and Y ,
defined in a way inverse to the image given above, then
g(w, z) = g(X, Y ).
(6.69)
6.6.7 All this defines on each Tp M and Tp M an internal product

(X, Y ) := (w, z) := g(X, Y ) = g(w, z).
(6.70)
A beautiful case of the field-form duality created by a metric is found in

hamiltonian optics, in which the momentum (eikonal gradient) is related to
the velocity by the refractive index metric (see Physical Topic 5.2). There are
many other in Physics. Let us illustrate by a howlingly simple example not
only the relation of 1-forms to fields, but also that of both to linear partial
differential equations. Consider on the plane the function (x, y are cartesian
coordinates, a and b real constants)
f (x, y) =
x2
a2
y2
.
b2
Each case f (x, y) = C (constant) represents an ellipse. The complete family

of ellipses is represented by the gradient form df ; that family is just the set
of solutions of the differential equation df = 0. But f is also solution of the
set of differential equations X(f ) = 0, where X is the field
X=
a2
x
b2
y
y .
Thus, a differential equation is given either by df = 0 or by a vector field.

In the first case the form is the gradient of the solution, which vanishes at
each value C. In the second case the solution must be tangent to the given
field. The form is orthogonal to the solution curve, that is, it vanishes
when applied to any tangent vector: df (X) = X(f ) = 0. Thus, a curve is
the integral of a field through tangency, and of a cofield through gradiency.
The word orthogonal was given quotation marks because no metric connotation is given to df (X) = 0. Of course, multiplying f by a constant will
change nothing. The same idea is trivially extended to higher dimensions.
In the example, we have started from a solution. We may start at a region
around a point (x, y) and eventually obtain from the form
184

df =
2x
a2
dx +
2y
b2
dy
some local solution f = T df (see 7.2.12 for a systematic method to get it);
this solution can be extended to the whole space, giving the whole ellipse.
This is a special case, as of course not every field or cofield is integrable. In
most cases they are only locally integrable, or nonintegrable at all.
Suppose now that a metric gij is present, which relates fields and cofields.
In the case above gij = diag(1/a2 , 1/b2 ) is of evident interest, as f (v) =
g(v, v), with v the position vector (x, y). To the vector v of components (xj )
will correspond the covector of components (pk = gkj xj ) and the action of
this covector on v will give simply p(v) = pk xk = gij xi xj . As we are also in a
euclidean space, the euclidean metric mij = ij may be used to help intuition.
We may consider p and v as two euclidean vectors of components (pk ) and
(xk ). Comparison of the two metrics is made by using g(v, v) = m(p, v).
Consider the curve p(v) = g(v, v) = m(p, v) = C, which is an ellipse. The
vector v gives a point on the ellipse and the covector p, now assimilated to an
euclidean vector, is orthogonal to the curve at each point, or to its tangent
at the point. This construction, allowing one to relate a 1-form to a field
in the presence of a non-trivial metric, is very much used in Physics. For
rigid bodies, the metric m is the inertia tensor,, the vector v is the angular
velocity and its covector is the angular momentum. The ellipsoid is the
inertia ellipsoid, the whole construction going under the name of Poinsot
(more details can be found in Physical Topic 2, section 2.3.10). In crystal
optics, the Fresnel ellipsoid ij xi xj = C regulates the relationship between
the electric field E and the electric displacement D = E, where the metric is
the electric permeability (or dielectric) tensor. In this case, another ellipsoid
is important, given by the inverse metric 1 : it is the index, or Fletchers
ellipsoid (Physical Topic 5.6). In all the cases, the ellipsoid is defined by
equating some hamiltonian to a constant.
6.6.8 An important property of a space V endowed with an internal product is the following: given any linear function f R(V ), there is a unique
vf V such that, for every u V , f (u) = (u, vf ). So, the forms include all
the real linear functions on Tp M (which is expected, they constituting its dual
space), and the vectors include all the real linear functions on Tp M (equally
not unexpected, the dual of the dual being the space itself). The presence of
a metric establishes a natural (or canonical) isomorphism between a vector
space (here, Tp M ) and its dual.
6.6.9 The above definition has used fixed bases. As in general no base covers the whole manifold, convenient transformations are to be performed in the
intersections of the definition domains of every pair of bases. If some of the
185
above metric-defining conditions are violated at a point p, it can eventually

come from something wrong with the basis: for instance, it may happen that
two of the Xi are degenerate at p. A real singularity in the metric should be
basis-independent. Non-degenerate metrics are called semi-Riemannian. Although physicists usually call them just Riemannian, mathematicians more
frequently reserve this denomination to non-degenerate positive-definite metrics, g : T M T M R+ . As it is not definite positive, the Lorentz metric
does not define balls and is consequently unable to provide for a topology on
Minkowski spacetime.
6.6.10 A Riemannian manifold is a smooth manifold on which a Riemannian metric is defined. A theorem (see Mathematical Topic 3.6) due to
Whitney states that
it is always possible to define at least one
Riemannian metric on an arbitrary differentiable manifold .
6.6.11 A metric is presupposed in any measurement: lengths, angles, volumes, etc. We may begin by introducing the length of a vector field X
through
||X|| = (X, X)1/2 .
(6.71)
The length of a curve : (a, b) M is then defined as
Z b
d
dt.
L =
dt
a
(6.72)
6.6.12 Given two points p, q M , a Riemannian manifold, we consider

all the piecewise differentiable curves with (a) = p and (b) = q. The
distance between p and q is the infimum of the lengths of all such curves
between them:
Z b
d
dt.
d(p, q) = inf
(6.73)
{(t)} a dt
In this way a metric tensor defines a distance function on M .
6.6.13 A metric is indefinite when ||X|| = 0 does not imply X = 0. It is
the case of Lorentz metric for vectors on the light cone.
6.6.14 Motions are transformations of a manifold into itself which preserve a metric given a priori. They are also called isometries in modern
texts, but this term in general includes also transformations between different spaces. When represented by field vectors on the manifold, eq.[6.51] will
give the components of the Lie derivative:
186

(LX g) = X g + ( X )g + ( X )g .
Using the properties

( X )g = X X g and ( X )g = X X g ,
it becomes
(LX g) = X ( g g g ) + X + X .
If we define the Christoffel symbol
= = 21 g [ g + g g ] ,
(6.74)
whose meaning will become clear later ( 9.4.23), then the Lie derivative
acquires the form
(LX g) = X X + X X .
Introducing the covariant derivative
X; = X X ,
(6.75)
(LX g) = X; + X; .
(6.76)
it can be written as
The condition for isometry, LX g = 0, then becomes
X; + X; = 0,
(6.77)
which is the Killing equation .9 A field X satisfying it is a Killing field (the

name Killing vector is more usual). There are powerful results concerning
Killing fields.10 For example, on a manifold M , the maximum number of
Killing fields is m(m + 1)/2 and this number is attained only on spaces with
constant curvature. It is a good exercise to find that the generators of the
motions on Minkowski space are of two types:
J() = ,
which generate translations, and
J() = x x ,
9
10
See Davis & Katzins 1962.

See Eisenhart 1949, chap.VI.
187
generators of Lorentz transformations. Together, these operators generate

the Poincare group. Invariance under translations bespeaks spacetime homogeneity. Invariance under Lorentz transformations means spacetime isotropy.
Such properties are seldom present in other manifolds: they may have analogues only on constant curvature spacetimes.11
The metrics concerned with light rays and sound waves, referred to in
the introduction of this section, are both obtained by multiplying all the
components of the euclidean metric by a given function. A transformation
like gij gij0 = f (p) gij is called a conformal transformation. Because in
the measurements of angles the metric appears in a numerator and in a
denominator, both metrics will give the same angle measurements. We say
that conformal transformations preserve the angles, or the cones.
6.6.15 Geometry, the very word witnesses it, has had a very strong historical relation to metric. Speaking of geometries has been, for a long
time, synonymous to speaking of kinds of metric manifolds. Such was, for
instance, the case of the last centurys discussions on non-euclidean geometries (see Mathematical Topic 11). The first statement of the first book
of Descartes Geometry is that every problem in geometry can easily be reduced to such terms that a knowledge of the lengths of certain straight lines
is enough for its construction. This comes from the impression, cogent to
cartesian systems of coordinates, that we measure something (say, distance
from the origin) when attributing coordinates to a point. Of course, we do
not. Only homeomorphisms are needed in the attribution, and they are not
necessarily isometric.
Nowadays, geometry both the word and the concept behind it
has gained a much enlarged connotation. We hope to have made it clear
that a metric on a differentiable manifold is an additional structure, chosen
and introduced at convenience. As said, many different metrics can in principle be defined on the same manifold (see more about that in Mathematical
Topic 11). Take the usual surfaces in E3 : we always think of a hyperboloid,
for instance, as naturally endowed with the (in the case, indefinite) metric induced by the imbedding in E3 . Nevertheless, it has also at least one
positive-definite metric, as ensured by Whitneys theorem. This character of
metric, independence from more primitive structures on a manifold, is not
very easy to reckon with. It was, according to Einstein, a difficulty responsible for his delay in building General Relativity: why were more seven years
required for the construction of the general theory of relativity? The main
reason lies in the fact that it is not so easy to free oneself from the idea that
11
For applications in gravitation and cosmology, see Weinberg 1972, chap. 13.
188
coordinates must have an immediate metrical meaning.12
12
Misner, Thorne &Wheeler 1973, page 5.
Chapter 7
DIFFERENTIAL FORMS
7.1
INTRODUCTION
7.1.1 Exterior differential forms1 are antisymmetric covariant tensor fields

on smooth manifolds ( 6.3.10). Roughly speaking, they are those objects
occurring under the integral sign. Besides being the central objects of integration on manifolds, these integrands have a lot of interest by themselves.
They have been introduced by Cartan mainly because of the great operational simplicity they provide: they allow a concise shorthand formulation of
the whole subject of vector analysis on smooth manifolds of arbitrary kind
and dimension.
We are used to seeing, in the euclidean 3-dimensional space, line integrals
written as
R
(Adx + Bdy + Cdz),
surface integrals as
RR
(P dxdy + Qdydz + Rdzdx),
and volume integrals as

RRR
T dxdydz.
The differential forms appearing in these expressions exhibit a common

and remarkable characteristic: terms which would imply redundant integration, such as dxdx, are conspicuously absent. Intuition might seem enough
1
A very good introduction to differential forms, addressed to engineers and physicists,

but written by a mathematician, is Flanders l963; a book containing a huge amount of
material, written by a physicist, is Westenholz l978. Slebodzinski l970 is a mathematical
classic, containing an extensive account with applications to differential equations and Lie
groups. Perhaps the most complete of the modern texts is Burke 1985. Other good texts
are Warner 1983 and Lovelock & Rund 1975.
189
190
CHAPTER 7. DIFFERENTIAL FORMS
to eliminate redundancy, but there is a deeper reason for that: integrals are
invariant under basis transformations and the corresponding jacobian determinants are already included in the integration measures, which are henceby
antisymmetric. We could almost say that, as soon as one thinks of integration, only antisymmetric objects are of interest. This is a bit too strong as,
for instance, a metric may be involved in the integration measure. However,
differential calculus at least is basically concerned with antisymmetric objects
with a well defined behaviour under transformations, that is, antisymmetric
tensors.
7.1.2 In the case of 1-forms (frequently called Pfaffian forms), of course,
antisymmetry is of no interest. We have seen, however, that they provide
basis for higher-order forms, obtained by exterior product ( 6.3.11). Recall
that the exterior product of two 1-forms (say, two members of a basis { i })
is an antisymmetric mapping
: T10 (M ) T10 (M ) T20 (M ),
where Tsr (M ) is the space of (r, s)-tensors on M . In the basis formed in this
way, a 2-form F , for instance, will be written
F = 12 Fij i j .
7.1.3 We shall denote k (M ) the space of the antisymmetric covariant
tensors of order k on the space M , henceforth simply called k-forms. Recall
that they are tensor fields, so that in reality the space of the k-forms on the
manifold M is the union
S
k (M ) = pM k (Tp M ).
In a way quite similar to the previously defined bundles, the above space can
be topologized and made into another fiber bundle, the bundle of kforms
on M . A particular k-form is then a section
: M k (M )
: p p k (Tp M ).
It is a universally accepted abuse of language to call k-forms differential
forms of order k.
We say abuse, of course, because not every differential form is the
differential of something else.
7.1.4 The exterior product also called wedge product is the generalization of the vector product in E3 to spaces of any dimension and thus,
through their tangent spaces, to general manifolds. It is a mapping
7.1. INTRODUCTION
191
: p (M ) q (M ) p+q (M ),
which makes the whole space of forms into a graded associative algebra.
Recall that dim p (M ) = mp , and the spaces of order p > m reduce to
zero. Thus, if p is a p-form and q is a q-form, p q = 0 whenever
p + q > m. The space of 0-forms has as elements the real functions on M
whose compositions, by the way, exhibit trivially the pull-back property.
7.1.5 A basis for the maximal-order space m (M ) is a single m-form
1 2 3 . . . m.
In other words, m (M ) is a 1-dimensional space. The nonvanishing elements
of m (M ) are called volume elements, or volume forms. Two volume elements v1 and v2 are said to be equivalent if a number c > 0 exists such that
v1 = cv2 . This equivalence divides the volume forms into two classes, each
one called an orientation. We shall come back to volume forms later.
This definition of orientation can be shown to be equivalent to that given
in 4.2.14.
Some nave considerations in euclidean spaces provide a more pictorial
view of Pfaffian forms. Let us proceed to them.
7.1.6 Perhaps the most elementary and best known 1-form in Physics is
the mechanical work, a Pfaffian form in E3 . In a natural basis, it is written
W = Fk dxk ,
with the components Fk representing the force. The total work realized in
taking a particle from a point a to point b along a line is
R
Wab [] = Fk dxk ,
and in general depends on the chosen line. It will be path-independent only
when the force comes from a potential as a gradient Fk = (gradU )k . In
this case W is W = dU , truly the differential of a function, and
Wab = U (a) U (b).
A much used criterion for this integrability is to see whether W [] = 0
when is any closed curve. However, the work related to displacements in
a non-potential force field is a typical non-differential 1-form: its integral
around a closed curve does not vanish, and its integral between two points
will depend on the path. Thus, the simplest example of a form which is not
a differential is the mechanical work of a non-potential force. We shall later
find another simple example, the heat exchange ( 7.2.10).
192
Of a more geometrical kind, also the form appearing in the integrand in

eq.[6.72] is not a differential, as the arc length depends on the chosen curve.
That is why the distance has been defined in eq.[6.73] as an infimum.
7.1.7 The gradient of a function like the potential U (x, y, z) may be pictured as follows: consider the equipotential surfaces U (x, y, z) = c (constant).
The gradient field is, at each point p E3 , orthogonal to the equipotential
surface going through p, its modulus being proportional to the growth rate
along this orthogonal direction. The differential form dU can be seen as
this field (it is a cofield, but the trivial metric of E3 identifies field and
cofields).
For a central potential, these surfaces will be spheres of radii
p
r = x2 + y 2 + z 2 , which are characterized by the form dr. That is to
say, the spheres are the integral surfaces of the differential equation dr = 0.
Despite the simplicity of the above view, it is better to see a gradient in E3 as
the field of tangent planes to the equipotential surfaces and regard 1-forms
in general as fields of planes.
A first reason for this preference is that we may then imagine fields of
planes that are not locally tangent to any surface: they are non-integrable
forms. They vary too quickly, in a non-differentiable way (as suggested
by the right-up corner in the scheme of Figure 7.1). A second reason is that
this notion is generalizable to higher order forms, which are fields of oriented
continuum trellis of hyperplanes, a rather unintuitive thing. For instance, 1forms on a space of dimension m are fields of (m1)-dimensional hyperplanes.
Integrable forms are those trellis locally tangent to submanifolds. The final
reason for the preference is, of course, that it is a correct view. The lack
of intuition for the higher order case is the reason for which we shall not
insist too much on this line2 and take forms simply as tensor fields, which is
equivalent. Let us only say a few more words on Pfaffian forms.
7.1.8 A 1-form is exact if it is a gradient, like = dU . Being exact is not
the same as being integrable. Exact forms are integrable, but non-exact forms
may also be integrable if they are of the form dU . The same spheres dr
of the previous paragraph will be solutions of dr = 0, where = (x, y, z)
is any well behaved function. The field of planes is the same, the gradients
have the same directions, only their modulus change from point to point (see
Figure 7.2). Of course, this is related to the fact that fields of planes are
simply fields of directions. The general condition for that is given by the
Frobenius theorem ( 6.4.33 and below, 7.3.14). Given a Pfaffian form ,
2
A beautiful treatment, with luscious illustrations, is given in Misner, Thorne &

Wheeler 1973.
7.1. INTRODUCTION
193
Figure 7.1: (a) Integrable field of planes; (b) Non-integrable field of

planes.
the differential equation = 0 is the corresponding Pfaffian equation. It will

have solutions if may be put into the form = df . Otherwise, will be
a field of planes which are not (even locally) tangent to families of surfaces,
as happens with non-potential forces.
Let us consider Pfaffian forms on E2 (with its usual global cartesian coordinates (x, y)), on which fields of straight lines will replace those of planes.
The line field formed by the axis Ox and all its parallels is fixed by dy, or
(x, y)dy for any , as the solutions of dy = 0 are y = constant. The fact
that in dy the modulus change from point to point (see Figure 7.3) does not
change the line field, which is only a direction field. The line field of vertical
lines, x = constant, is (x, y)dx.
The form = adx + bdy, where a and b are constants, will give straight
lines y = (a/b)x + c (Figure 7.4 a), whose tangent vectors are
v = (x,
y)
= x x + y y = x x + (a/b) y .
The form is orthogonal to all such tangent vectors: (v) = 0.
Next in complication would be the form
= A(x, y)dx + B(x, y)dy.
194
Figure 7.2: (a) The field dr; (b) The field dr: the moduli change from
point to point, but the directions remain radial.
The equation = 0 will be always integrable, as the ordinary differential

equation
dy
= A/B
dx
will have as solution the one-parameter family of curves f (x, y) = c. There
will always exist an (x, y, z) such that = df . The form / is exact and
is consequently called an integrating denominator . Every field of straight
lines on the plane will find locally a family of curves to which it is tangent.
It follows that d = (d/) . The particular case of = 2xdx + dy,
depicted in Figure 7.4 (b), has for the Pfaffian equation solutions y = x2 + C,
with tangent vectors x
x + 2xx
y.
All this holds no more in higher dimensions: fields of (hyper-) planes are not necessarily
locally tangent to surfaces. When generalized to manifolds, all such line- and plane fields
are to be considered in the euclidean tangent spaces.
7.1.9 A very useful object is the Kronecker symbol , defined by
k k k k
j11j22j33jpp
+ 1 if the js are an even permutation of the ks

= 1 if the js are an odd permutation of the ks
0 in any other case.
This symbol is a born antisymmetrizer. It may be seen as the determinant
7.1. INTRODUCTION
195
Figure 7.3: (a) The line field dy; (b) The line field dy; (c) The line field
dx.
k k k k
j11j22j33jpp
jk11
jk21
...
jkp1
jk12
jk22
...
jkp2
...
...
...
...
j1p
k
j2p
...
k
jpp
(7.1)
It satisfies the relation

k k k k
q n1 n2 n3 np
2 k3 kq n1 n2 n3 np
= q! km11km
.
j11j22j33jq q jm1 1j2mj32j
m3 mq+p
2 m3 mq+p
(7.2)
When no doubt arises, we may write simply

j1 j2 j3 jp = 1j12j23j3jpp .
(7.3)
j1 j2 j3 jm j1 j2 j3 jm = m!
(7.4)
When p = m = dim M ,
7.1.10 In form [7.3] the Kronecker symbol is of interest in the treatment

of determinants. Given an n n matrix A = (Aij ), some useful formulae
involving its determinant are:
detA = i1 in A1i1 A2i2 . . . Anin ;
(7.5)
i1 in Ai1 j1 Ai2 j2 . . . Ain jn = j1 j2 j3 jn detA;
(7.6)
j1 jn i1 ...in A
i1 j1
i2 j2
...A
in jn
= n!detA.
(7.7)
Notice that we are using here upper indices only for notational convenience.
We shall later meet Kronecker symbols of type [7.3] with indices raised by
the action of a metric.
196
Figure 7.4: (a) The line field w = bdy adx; (b) The line field w = 2xdx +
dy.
7.1.11 Kronecker symbols are instrumental in calculations involving components of forms. Given a p-form
=
1
p!
j1 j2 jp j1 j2 . . . jp ,
(7.8)
one particular component is obtained as

j1 j2 jp =
1
p!
k k k k
j11j22j33jpp k1 k2 k3 kp .
(7.9)
The basis for the space of p-forms can be written as

j j j j
j1 j2 jp = j1 j2 . . . jp = k11 k22 k33 kpp k1 k2 . . . kp .
(7.10)
7.1.12 Given the p-form and the q-form , the components of the wedge
product are, in terms of the components of and ,
( )i1 i2 ip+q =
1
p!
k k k k j j j jp
q 1 2 3
3
i11i22i3
iq+p
k1 k2 k3 kq j1 j2 j3 jp .
(7.11)
7.1.13 A practical comment: in eq.[7.8], one is supposed to sum over

the whole range of all the indices. Many authors prefer to use only the
independent elements of the basis: for example, as 1 2 = - 2 1 , they
are of course not independent. Instead of [7.8], those authors would write
= j1 j2 jp j1 j2 . . . jp ,
(7.12)
without the factor 1/p! but respecting j1 < j2 < j3 < . . . < jp in the
summation. We shall use one or another of the conventions, according to
convenience.
7.2. EXTERIOR DERIVATIVE
197
7.1.14 The main properties of the exterior product have been outlined in
eqs.[6.20]. Let us only restate the rule concerning commutation:
p q = ()pq q p
(7.13)
If p is odd, p p = - p p = 0. In particular, this holds for the elements

j of the basis. For a natural basis,
dxi xj = dxj dxi ,
so that dx dx = 0, dy dy = 0, etc. When no other product is present
and no confusion is possible, we may omit the exterior product sign and
write simply dxi dxj = dxj dxi or, using anticommutators, {dxi , dxj } = 0.
A function f is a 0-form and
(f ) = (f ) = f ( ).
(7.14)
f = f = f.
(7.15)
Of course,
Given any p-form , we define the operation of exterior product by a 1-form
through
() : p (M ) p+1 (M )
() = ,
p < m.
(7.16)
7.1.15 It is easy to check that the vanishing of the wedge product of two
Pfaffian forms is a necessary and sufficient condition for their being linearly
dependent.
7.2
EXTERIOR DERIVATIVE
7.2.1 The 0-form f has the differential

df =
f
f
dxi =
dxi ,
i
i
x
x
(7.17)
is a 1-form. The generalization of differentials to forms of any order is the

exterior differential, an operation d with the following properties:
(i) d : k (M ) k+1 (M ); that is, the exterior differencial of a (k-form) is
a certain (k + 1)-form;
(ii) d( + ) = d + d;
198
(iii) d( ) = (d) + () , being the order of ;

(iv) dd =2 0, for any form .
These properties define one and only one operation.
7.2.2 To grasp something about condition (iv), let us examine the simplest
case, a 1-form in a natural basis {dxk }: = i dxi . Its exterior differencial
is
d = (di ) dxi + i d(dxi ) =
i
xi
dxi .
If is exact, = df or, in components, i = i f , then

h 2
i
f
2f
dxi dxj
d = d2 f = 12 xi x
j xj xi
and the property d2 f 0 is just the symmetry of the mixed second derivatives of a function. Along the same lines, if is not exact, we can consider
i
h 2
2
2 k
i
1
j
k
i
i
d2 = xj x
dx
dx
dx
=
dxj dxk dxi = 0.

k
2!
xj xk
xk xj
Thus, the condition d2 0 comes from the equality of mixed second derivatives of the functions i , and is consequently related to integrability conditions. It is usually called the Poincare lemma. We shall see later its relation
to the homonym of 2.2.17.
7.2.3 It is natural to ask whether the converse holds: is every form satisfying d = 0 of the type = d? A form such that d = 0 is said to be
closed. A form which can be written as a derivative, = d for some ,
is said to be exact. In these terms, the question becomes: is every closed
form exact? The answer, given below as the Poincare inverse lemma, is: yes,
but only locally. It is true in euclidean spaces, and differentiable manifolds
are locally euclidean. More precisely, if is closed in some open set U , then
there is an open set V contained in U where there exists a form (the local
integral of ) such that = d. In words, every closed form is locally exact.
But attention: if is another form of the same order of and satisfying d0,
then also = d( + ). There are, therefore, infinite forms of which
is the differential. The condition for a closed form to be exact on the open
set V is that V be contractible (say, a coordinate neighbourhood). On a
smooth manifold, every point has an euclidean (consequently contractible)
neighbourhood and the property holds at least locally. When the whole
manifold M is contractible, closed forms are exact all over M . When M is
not contractible, a closed form may be non-exact, a property which would
199
be missed from a purely coordinate point of view. Before addressing this

subject, let us examine the use of the rules above in some simple cases.
Notice that, after the considerations of 7.1.8 a form may be integrable without being closed. The general problem of integrability is dealt with by the Frobenius theorem
( 6.4.33), whose version in terms of forms will be seen later ( 7.3.14).
By what we have said in 7.1.6, the elementary length ds is a prototype of form which
is not an exact differential, despite its appearance. Obviously the integral
Rx
ds
a
depends on the trajectory, leading thus to a multi-valued function of x (see Mathematical
Topic 7 for more).
7.2.4 Take again the 2-form

F = 12 Fij i j .
Its differential is the 3-form

dF = 21 dFij i j + Fij (d i ) j Fij i d j
(7.18)
(7.19)
The computation is done by repeated use of properties (iii) and (ii) of 7.2.1.
One sees immediately the great advantage of using the natural basis i = dxi .
In this case, from property (iv), only one term remains:
dF = 21 dFij dxi dxj .
The component is a function, so its differential is just as given by eq.[7.17]:
dF =
1
2
Fij
dxk dxi dxj .
xk
(7.20)
We would like to have this 3-form put into the canonical form [7.8], with the
components fully symmetrized. If we antisymmetrize now in pairs of indices
(k, i) and (k, j), we in reality get 3 equal terms,
dF =
1
3!
[k Fij + j Fki + i Fjk ] dxk dxi dxj

k
i
j
= 3!1 2!1 pqr
kij p Fqr dx dx dx . (7.21)
7.2.5 For a general q-form

=
1
q!
j1 j2 ...jq dxj1 dxj2 . . . dxjq ,
200
the differential will be

d =
1
q!
d(j1 j2 ...jq ) dxj1 dxj2 . . . dxjq

j1 j2 ...jq
dxj0 dxj1 dxj2 . . . dxjq
xj0

1 j0 j1 j2 ...jq j1 j2 ...jq
(7.22)
dxi0 dxi1 dxi2 . . . dxiq ,

q! i0 i1 i2 ...iq
x0
=
1
(q+1)!
1
q!
which gives the components

(d)i0 i1 i2 ...iq =
1
q!
j j j ...j
i00i11i22...iqq
j1 j2 ...jq
.
x0
(7.23)
7.2.6 It is convenient to define the partial exterior derivative of with

respect to the local coordinate xj0 by
=
x0
1
q!
j1 j2 ...jq
dxj1 dxj2 . . . dxjq
xj0
(7.24)
so that
.
(7.25)
x0
The expression for the exterior derivative in an arbitrary basis will be found
below (see eq.[7.74]). We shall see later (eq.[7.161) the real meaning of
eq.[7.24], and give, in consequence, still another closed expression for d.
d = dxj0
7.2.7 The invariant, basis-independent definition of the differential of a

k-form is given in terms of its effect when applied to fields:
(k + 1)d(k) (X0 , X1 , . . . , Xk ) =
k
X
bi , Xi+1 , Xi+2 , . . . Xk )]+

()i Xi [(X0 , X1 , X2 , . . . , Xi1 , X
i=0
k
X
i+j
()

b
b
[Xi , Xj ], X0 , X1 , . . . , Xi , . . . , Xj , . . . , Xk , (7.26)
i<j
bn means that Xn is absent.

where, wherever it appears, the notation X
7.2.8 Let us examine some facts in E3 , where things are specially simple.
There exists a global basis, the cartesian basis consisting of
e1 =
=
;
e
=
=
;
e
=
=
.
2
3
x1
x
x2
y
x3
z
201
Its dual is {dx, dy, dz}. The euclidean metric is, in the related cartesian
coordinates,
g = ij dxi dxj .
Given a vector V = V i x i , its covariant image Z is a form such that, for any
vector U ,
Z(U ) = Zi U i = g(U, V ) = ij V j U i ,
so that Zi = ij Vj . One uses the same names for a vector and for its covariant
image, writing Zi = Vi . So, to the vector V corresponds the form V = Vi dxi .
Its differential is
dV =
Vi k
dx dxi = 12 (k Vi i Vk ) dxk dxi ,
xk
or
dV = 21 (rotV )ki dxk dxi .
Think of electromagnetism in E3 : the vector potential is the 1-form A =
Ai dxi , and the magnetic field is H = dA = rotA.
The derivative of a 0-form f is
df = (gradf )i dxi .
Suppose the form V above to be just this gradient form. Then,
d2 f = rot gradf
and the Poincare lemma is here the well known property rot grad f 0.
When the vector potential is the gradient of a function, A = df , the magnetic
field vanishes:
H = d2 f 0.
Consider now the second order tensor
T = 21 Tij dxi dxj .
In E3 , to this tensor will correspond a unique vector (or 1-form) U , fixed by
Tij = ijk Uk . The differential is
dT = 21 k Tij dxk dxi dxj
= 21 ijk k Uk dxk dxi dxj = (div U ) dx1 dx2 dx3 .
Taking Ui = 21 ijk (rotV )jk , the Poincare lemma assumes still another well
known avatar, namely, div rot V 0. The expression for the laplacian of a
0-form f , div gradf = i i f , is easily obtained.
202
A criterion to see the difference between a true vector and a second order tensor is the
behaviour under parity (xi xi ) transformation. A true vector changes sign, while a
second order tensor does not. The magnetic field is such a tensor, and Maxwells equation
divH = 0 is dH = 0, actually the identity d2 A 0 if H = dA.
7.2.9 Maxwells equations, first pair Consider the electromagnetic

field strength in vacuum (0 = 0 = c = 1). It is a second order antisymmetric tensor in Minkowski space, with the components
0
H3 H2 E1
H3
0
H1 E2
[F ] =
(7.27)
H2 H1
0
E3
E1 E2 E3 0
The fourth row and column in the matrix correspond to the zeroth, or time
components. The field strength can be written as a 2-form
F = 21 F dx dx ,
(7.28)
with , = 1, 2, 3, 0. In detail,
F = H1 dx2 dx3 + H2 dx3 dx1 + H3 dx1 dx2
+ E1 dx1 dx0 + E2 dx2 dx0 + E3 dx3 dx0 ,
or
F = 21 ijk Hi dxj dxk + Ej dxj dx0 .
(7.29)
From [7.21],
dF =
1
3!
{ F + F + F } dx dx dx .
(7.30)
From [7.29],
~ Hdx
~ 1 dx2 dx3 +
dF =
+
H3
x0
E2
x1
E1
x2
H
x0
dx0 dx1 dx2 +
E2
x3
H2
x0
E3
x2
E1
x3
dx0 dx2 dx3
E3
x1
dx0 dx3 x1 .
Thus, the equation

dF = 0
(7.31)
~ H
~ = 0 and 0 H
~ = rot E.
~
(7.32)
is the same as
203
This is the first pair of Maxwells equations. Of course, this could have been
seen already in eq.[7.30], which gives them directly in the usual covariant
expression
F + F + F = 0.
(7.33)
Equation [7.31] says that the electromagnetic form F is closed. In Minkowski pseudo-euclidean space (supposedly contractible; recall that we do
not know much about its real topology, 1.1.14, 1.1.18 and 1.3.6), there
exists then a 1-form
A = A dx
such that
F = dA =
1
2
[ A A ] dx dx ,
(7.34)
or, in components,
F = A A .
The potential form A is not unique: given any 0-form f , we can also write
F = d(A + df ). The potentials A and A0 = A + df,
A0 = A + f,
(7.35)
give both the same field F . This is a gauge transformation. The gauge
invariance of F is thus related to its closedness and to the arbitrariness born
from the Poincare lemma. We could formally define F as dA. In that case, the
first pair of Maxwells are not really equations, but constitute an identity.
This point of view is justified in the general framework of gauge theories.
From the quantum point of view, the fundamental field is the potential A,
and not the field strength F . Although itself not measurable, its integral
along a closed line is measurable (Aharonov-Bohm effect, seen in 4.2.18).
Furthermore, it is the field whose quanta are the photons. Even classically,
there is a hint of its more fundamental character, coming from the lagrangian
formalism: interactions with a current j are given by A j . There is also a
further suggestion of the special character of dF = 0: unlike the second pair
of Maxwells equations (see below, 7.4.17), the first pair does not follow
from variations of the electromagnetic lagrangian L = 41 F F .
7.2.10 Thermodynamics of very simple systems We call very simple systems those whose states are described by points on a two-dimensional
manifold with boundary, usually taken as diffeomorphic to the upper right
quadrant of the plane E2 . Thermodynamical coordinates are conveniently
chosen so as to represent measurable physical variables. We shall use the
entropy S and the volume V . The remaining physical quantities are then
204
functions of these two variables (this is sometimes called the entropy-volume

representation). The internal energy, for example, is U = U (S, V ). With
obvious notation, the first principle of thermodynamics reads
dU (S, V ) = T (S, V )dS P (S, V )dV.
The heat variation T dS is usually denoted in textbooks by Q or some
other notation which already indicates that something is amiss. It is in reality
another simple physical example of a 1-form which is not a differential: it is
not an exact form, there exists no such a function as Q that makes this
form into T dS = dQ. Though the same is true of the work P dV , the first
principle says that the difference dU is an exact form. Taking the derivative,
d2 U = 0 = dT dS dP dV.
But
dT =
T
S V
dS +
T
V
dP =
P
S V
dS +
P
V
S
S
dV,
dV.
Thus,
T
V

S
dV dS =
P
S V
dS dV.
Consequently,
T
V

S
P
S V
which is one of Maxwells reciprocal relations. The other relations are obtained in the same way, however using different independent variables from
the start. All of them are integrability conditions, here embodied in the
Poincare lemma. A mathematically well founded formulation of Thermodynamics was initiated by Caratheodory3 and is nowadays advantageously spelt
in terms of differential forms, but we shall not proceed to it here.4
7.2.11 We have introduced 1-forms, to start with, as differentials of functions (or 0-forms). We have afterwards said that not every 1-form is the
differential of some function, and have found some examples of scu nondifferential forms: mechanical work ( 7.1.6) and thermodynamical heat and
work exchanges ( 7.2.10). This happens also for forms of higher order: not
every p-form is the differential of some (p 1)-form. This is obviously related
to integrability: given an exact form
3
4
Very nice resumes are found in Chandrasekhar 1939 and Born 1964.
See for instance Mrugala 1978 and references therein.
205
= d ,
is its integral. The expression stating the closedness of ,

d = 0,
when written in components, becomes a system of differential equations
whose integrability (i.e., the existence of a unique integral ) is only granted
locally.
7.2.12 The inverse Poincare lemma says that every closed form is locally
exact and gives an expression for the integral of . Locally has a precise
meaning: if d = 0 at the point p M , then there exists a contractible
neighbourhood of p in which exists such that = d. To be more precise,
we have to introduce still another operation on forms: given, in a natural
basis, the p-form
(x) = i1 i2 ...ip (x)dxi1 dxi2 . . . dxip ,
the transgression of is the (p 1)-form given by
T =
Pp
j1
j=1 ()
R1
0
dt tp1 xij i1 i2 ...ip (tx)dxi1 dxi2 . . .
. . . dxij1 dxij+1 . . . dxip .
(7.36)
Notice that, in the x-dependence of , x is replaced by (tx) in the argument.

As t ranges from 0 to 1, the variables are taken from the origin to x. In each
term of the summation, labelled by the subindex j, the j-th differential dxij
is replaced by its integral xij . In reality, the T operation involves a certain
homotopy, and the above expression is frequently referred to as the homotopy
formula. The operation is clearly only meaningful in a starshaped region, as
x is linked to the origin by the straight line tx, but can be generalized to
a contractible region. The limitation of the result to be given below comes
from this strictly local property. Well, the lemma then says that, locally, any
form can be written in the form
= d(T ) + T (d).
(7.37)
The proof of this fundamental formula is rather involved and will not be
given here.5 It can nevertheless be directly verified from eq.[7.36], by using
the identity
d
d
[t(tx)] t [(tx)].
(tx) =
dt
dt
5
A constructive proof for general manifolds is found in Nash & Sen 1983; on En , proofs
are given in every textbook: Goldberg l962, Burke l985, etc.
206
7.2.13 The expression [7.37] tells us that, when d = 0,

= d(T ),
(7.38)
so that is indeed exact and the looked for above is just = T (up to s
such that d = 0). Of course, the formulae above hold globally on euclidean
spaces, which are contractible.
~ = constant. It is closed by
7.2.14 Take a constant magnetic field, B
~ B
~ = 0. It is the rotational of A
~ = TB
~ =B
~ ~r, as
Maxwells equation
comes directly from eq.[7.36].
7.2.15 A simple test: take in E2 the form
p
xdx+ydy
v = dr = d x2 + y 2 =
.
2
2
x +y
Then, as expected,
Tv =
R1
0
2 +ty 2
dt tx2
t x2 +t2 y 2
x2 + y 2 = r.
7.2.16 A fundamental property of the exterior derivative is its preservation

under mappings: if f : M N , then the derivative of the pull-back is the
pull-back of the derivative,
f (dw) = d(f w).
(7.39)
The way to demonstrate it consists in first showing it for 0-forms and

then using induction. Let f be given by its local coordinate representation
y f x1 , or y i = f i (x1 , x2 , . . . , xm ), with i = 1, 2, . . . , n (see the scheme of
Figure 7.5).
A field X on M , locally X = X i x i , is pushed forward to a field f X
on N , such that
(f X)(g) = X(g f ) = X i x i g[f (x)] =
= Xi
r
r
g
1 2
n
i g y
i f
g[y
,
y
,
.
.
.
,
y
]
=
X
=
X
.
i
r
i
i
x
y x
x y r
This holds for any g, so that we may write

r
(f X) = X i f
xi
f r
= X(f r ) f r .
The pull-back of the 1-form on N is that 1-form f on M satisfying
207
Figure 7.5:
(f )(X) = (f X) = ( f )(X).
In a local basis, we have = j dy j and
(f X) = j dy j [X i
In the case case = dg =
g
y r
f r
]
xi y r
= j
f r
xi
Xi .
dy r :
(f dg)(X) = dg(f X) =
g f r
y r xi
X i.
A function g is pulled back to the composition, f g = g f . Then,

g f j
g f j
i
k
i
{d[f g]}(X) = {d[g f ]}(X) = { y
= y
j xi dx } X
j xi X ,
xk
so that
f dg = d[f g].
This is eq.[7.39] for 0-forms. As already said, we now proceed by induction.
Suppose the general result holds for (p 1)-forms. Take a p-form . Its
pull-back is
f = f [ p!1 i1 i2 i3 ...ip dy i1 dy i2 dy i3 . . . y ip ]
=
1
p!
[f i1 i2 i3 ...ip dy i1 dy i2 y i3 . . . y ip1 ] f dy ip .
Therefore
d[f ] =
1
p!
{d[f i1 i2 i3 ...ip dy i1 dy i2 dy i3 . . . y ip1 ] f dy ip +
()p1 [f i1 i2 i3 ...ip dy i1 dy i2 dy i3 . . . dy ip1 ] d[f dy ip ]}.

But
208

d[f dy ip ](X) = d[dy ip (f X)],
which means that

d[f dy ip ] = d[dy ip f ] = 0,
and the second term above vanishes. Now, using induction,
d[f ] =
1
p!
f d[i1 i2 i3 ...ip dy i1 y i2 dy i3 . . . dy ip1 ] f dy ip
d[f ] =
1
p!
f d[i1 i2 i3 ...ip dy i1 y i2 dy i3 . . . dy ip1 ] f dy ip .
Consequently,
d[f ] = f [d] .
7.2.17 General Basis As far as derivations are involved, calculations
are simpler in natural bases, but other bases may be more convenient when
some symmetry is present. Let us go back to general basis and reexamine the
question of derivation. Suppose {e } is a general basis for the vector fields
in an open coordinate neighbourhood U of M , and { } its dual basis. They
can be related to the natural basis of the chart (U, x):
e = e
;
x
(7.40)
= dx .
(7.41)
= e e ,
x
(7.42)
dx = ,
(7.43)
Conversely,
where e e = and e e = . The duality relations show that e =

and e = . From eq.[7.40],
[e , e ] = C e ,
(7.44)
with the structure coefficients given by

C = [e (e ) e (e )] e .
We can now calculate
d = d[ dx ] = 12 [ ] dx dx
(7.45)
209
= 12 [e e (e ) e e (e )] e e
= 12 [e e (e ) e e (e )] .
Using the relations like e e = and the derivatives
e e (e ) = e e (e ),
etc, we finally get
d = 21 C .
(7.46)
This equation is a translation of the commutation relations [7.44] to the

space of forms. It tells us in particular that
d (e , e ) = C
.
(7.47)
and eqs.[7.42], [7.43], the differential of a 0-form in an

From df = x
dx
anholonomic basis is
df = e (f ) .
(7.48)
We shall see later that, on Lie groups, there is always a basis in which
the structure coefficients are constant (the structure constants). In that
case, eq.[7.46] bears the name of Maurer-Cartan equation.
Suppose now the 1-form A = A . By using [7.45], one easily finds
dA = 12 [e (A ) e A ) C ] .
(7.49)
Basis { } will be holonomic (or natural, or coordinate) if some coordinate

system {y } exists in which
y
dx .
(7.50)
= dy =
x
This means that is an exact form, and a necessary condition is d = 0.
Conversely, this condition means that a coordinate system such as {y } above
exists, at least locally. From [7.46] comes the equivalent condition
C = 0.
7.2.18 We can go back to the anholonomic spherical basis of 6.5.3 for
E3 , and find the dual forms to the fields Xr , X and X . They are given
respectively by: r = dr, = rd and = r sin . We may then write
the gradient in this basis,
df = ek (f ) k = Xr (f ) r + X (f ) + X (f )
and check that this is the same as
df = (r f )dr + ( f )d + ( f )d,
which would be the expression in the natural basis related to the coordinates
(r, , ). The invariance of the exterior derivative is clear.
210
7.3
VECTOR-VALUED FORMS
7.3.1 Up to now, we have been considering ordinary q-forms, antisymmetric linear mappings taking q vector fields into the real line. A vectorvalued q-form will take the same fields into a vector space. If V is the vector
space and q (M ) is the space of ordinary q-forms on the manifold M , a qform with values in V is an element of the direct product V q (M ); if
{Va } is a basis for V , is written
= Va a ,
(7.51)
where a are ordinary forms. Thus, a vector-valued form may be seen as a

vector whose components are ordinary forms, or as a column of forms.
7.3.2 Vector-valued forms are of fundamental importance in the theory of
fiber bundles, where they appear as representatives of connections, curvatures, soldering, etc. They turn up everywhere in gravitational and gauge
theories: gauge potentials and field strengths are in reality connections and
curvatures respectively. They have been defined as direct products, and in
consequence the operations on q (M ) ignore eventual operations occurring in
V and vice-versa. The exterior derivative of the above form , for example,
is defined as
d = Va d a .
(7.52)
Notice that the above definitions yield objects independent of the basis
chosen for V . Under a basis change to Va0 = Uaa0 Va , the component forms
0
0
change to a = (U 1 )aa a , so that
= Va a = Va0 a
(7.53)
remains invariant. Usual forms a have already been introduced as basisindependent objects in q (M ), so that the whole object s basis-independent
in both spaces.
7.3.3 Algebra-valued forms Of special interest is the case in which the
vector space V has an additional structure of Lie algebra. The generators
will satisfy commutation relations [Ja , Jb ] = f c ab Jc and may be used as a
basis for the linear space V . It is precisely what happens in the examples
quoted above: gauge fields and potentials are forms with values in the Lie
algebra of the gauge group. There are two possible operations on such forms,
the exterior product and the Lie algebra operation. Due to the possible
anticommutation properties of exterior products, one must be careful when
handling operations with the complete vector-valued forms. A bracket can be
7.3. VECTOR-VALUED FORMS
211
defined which dutifully accounts for everything: given two forms = Ja a

and = Jb b of any orders, the bracket is6
[, ] := [Ja , Jb ] a b .
(7.54)
7.3.4 Because we wish to stand by usual practice, we are employing above

the same symbol [ , ] with two different meanings: at the left-hand side, the
defined bracket; at the right-hand side, the commutator of the Lie algebra.
The general definition of the bracket is consequently the following:
|[A, B]| = A B ()A B B A,
(7.55)
C being the order of the algebra-valued form C. This is a graded commutator . When at least one of the involved forms is of even degree, no signs
will come from the exterior product and the bracket reduces to a simple
commutator. Otherwise, an anticommutator comes out.
7.3.5 For example, a 1-form will be
A = Ja Aa dx .
(7.56)
Start by calculating
A A = Ja Jb Aa Ab dx x = 21 [Ja , Jb ]Aa Ab dx x
= 12 Jc f c ab Aa Ab dx x ,
(7.57)
the fc ab s being the Lie algebra structure constants. If we compare with [7.54],
it comes out that
A A = ( 12 [A, A].
(7.58)
As announced, this is actually a graded commutator, though we use for it
the usual commutator symbol.
7.3.6 A particular example is given by the algebra su(2) of the Lie group
SU (2) of special (that is, with determinant = +1) unitary complex matrices.
The lowest-dimensional representation has generators Ji = 21 i , where for
the i s we may take the Pauli matrices in the forms
1 = [ 01 10 ] ; 2 = [ 0i 0 i ] ; ; 3 = [ 10 01 ] ; .
The 1-form [7.56] is then the matrix
h
A3 dx
A = 21 A1 dx +iA
2 dx
A1 dx iA2 dx
A3 dx
It is also an example of a matrix whose elements are noncommutative.

6
Lichnerowicz 1955.
212
7.3.7 The differential of the 1-form [7.56] is easily obtained:

dA = Ja ( Aa ) dx dx = 21 Ja ( Aa Aa ) dx x .
(7.59)
In a non-holonomic basis, we should have found (using [7.46])

dA = 12 Ja (e Aa e Aa C Aa ) .
(7.60)
7.3.8 In gauge theories, the gauge field (strength) F is a 2form on a

4-dimensional space, given in terms of the (1-form) gauge potential A by
F = dA + 21 [A, A] = dA + A A.
(7.61)
The Ja s are the generators of the gauge group Lie algebra. To obtain the
relations between the components, use [[7.61], [[7.59] and [[7.57] to write
F = 12 Ja ( Aa Aa + f a bc Ab Ac )dx dx .
(7.62)
Then, defining the components of F through

F = 12 Ja F a dx x ,
(7.63)
F a = Aa Aa + f a bc Ab Ac .
(7.64)
we find the expression
7.3.9 Notice en passant that even-order forms behave under commutation

just as normal elements in the algebra: the bracket defined in eq.[7.54] reduces
to the algebra commutator when a 2-form, for example, is involved. For
instance, [A A, A] = 0. Take the differential of [7.61]:
dF = 0 + dA A A dA = [dA, A] = [F A A, A],
so that
dF + [A, F ] = 0.
(7.65)
This relation, an automatic consequence of the definition [7.61] of F , is the

Bianchi identity. Notice that it has taken us just one line to derive it.
7.3.10 We shall see now that the expression
DA () = d() + [A, ()]
can be interpreted as a covariant derivative of a 2-form () according to the
connection A, just as
213
DA () = d() + A ()
is the covariant derivative of a 1-form. This will be a bit more formalized
below, in 7.3.11. As we stand, such names are mere analogies to the Riemannian case. Within this interpretation, the field is the covariant derivative
of the connection proper, and the Bianchi identity establishes the vanishing
of the covariant derivative of the field. By the same analogy, the field F is
the curvature of the connection A. In components, eq.[7.65] reads
0 = dF + [A, F ] =
1
3!
Ja {[ F a ] + f a bc Ab [ F c ] } dx dx dx ,
where the symbol [] indicates that complete antisymmetrization is to be

performed on the enclosed indices. It follows the vanishing of each component,
[ F a ] + f a bc Ab [ F c ] = 0.
If we define the dual tensor Fea = 12 F a , the above expression may
be written as
Fea + f a bc Ab Fec = 0.
(7.66)
7.3.11 Covariant derivatives In order to understand the meaning of all
that, we have to start by qualifying [7.52] and [7.53]. The form has been
supposed to take values on some unique vector space V which is quite independent of the manifold. Transformations in that vector space do not affect
objects on the manifold. Consider now the case in which transformations in
V , defined by matrices g, depend on the point on the manifold. If W = Ja W a
is a form of order W and the Ja s are matrices as in 7.3.6, transformations
in V will lead to
W 0 = gJa g 1 W a = gW g 1 .
This is the usual way transformations act on matrices, and will be seen
later (section 8.4) to be called an adjoint action. The matrices g are now
supposed to be point-dependent, g = g(x). We say that they are gaugefied.
Everything goes as before for W itself, but there is a novelty in dW : the
derivative of the transformed form will now be
dW 0 = d(gW g 1 ) = dg g 1 + gdW g 1 + ()W gW dg 1 .
Only the second term of the r.h.s. has a good behaviour under the transformation, just the same as the original form W . We call covariant derivative
of a tensor W a derivative DW which has the same behaviour as W under
the transformation. Obviously this is not the case of the exterior derivative
dW . Let us examine how much it violates the covariance requirement. The
1-form = g 1 dg will be of special interest. By introducing at convenient
places the expression I = gg 1 for the identity, as well as its consequences
214

dgg 1 + gdg 1 = 0 and dg 1 = g 1 dgg 1 ,
dW 0 may be written as
dW= g dW g1 + g{ - (-)W W }g1 = g{dW + |[,]|}g1 ,
in terms of the graded commutator [7.55]. We look now for a compensating
form, a 1-form A transforming according to
A0 = gAg 1 + gdg 1 = g{A }g 1 ,
and we verify that
dW 0 + |[A0 , W 0 ]| = g{dW + |[A, W ]|}g 1 .
In geometrical language, an A transforming as above is a connection on the
manifold. It follows from the last expression that
DW = dW + |[A, W ]|
is the covariant derivative for any form transforming according to the expression W 0 = gW g 1 . This reduces to the previous expressions in the case
of gauge fields. On the other hand, if W is a column vector of forms, and
the matrices act as usual on column vectors, W 0 = gJa W a = gW , then the
covariant derivative is
DW = dW + AW,
with A the same connection as above. Of this type are the covariant derivatives of the source fields in gauge theories. In general, the gs generate a
group G. Quantities transforming as gW g 1 are said to belong to the adjoint representation of G, and quantities transforming as gW belong to linear
representations. In gauge theories, the potentials A and their curvatures
(field strengths) F belong to the adjoint representation. Source fields usually belong to linear representations. Let us retain that the expression of
the covariant derivative depends both on the order of the form and on the
representation to which it belongs in the transformation group.
It is good to keep in mind that some of the so-called covariant derivatives found in
many texts are actually covariant coderivatives, to be seen later ( 7.4.20).
7.3.12 Moving frames Some remarkable simplifying properties show up

in euclidean spaces En . The metric can be taken simply as gij = ij . There
is a global canonical basis of column vectors K , and also a dual basis
of rows KT = (0, 0, . . . , 1, 0, . . . , 0)T with 1 only in the -th entry (T
means transpose). En itself is diffeomorphic to both Tp En and Tp En for each
215
p En . Given a basis {ei }, each member is a section in the tangent bundle,

but can be seen here as a mapping ei : En En , ei : p ei (p), with ei (p)
the vector ei at the point p. Consequently, dei is a 1-form taking En into En ,
a vector-valued form. We write it
dei = ij ej ,
(7.67)
with ij some usual 1-forms. Differentiating this expression and using it

again, one arrives immediately at
dij = ik kj .
(7.68)
Writing ei in the canonical basis,

ei = ei K ,
(7.69)
the elements of the dual basis { j } will be

j = j K T
(7.70)
with j ei = ij . Differentiating [7.69] and comparing with [7.67], one finds

ij = j dei .
(7.71)
From < ei , ej > = ij , one obtains < dei , ej > + < ei , dej > = 0, with the
consequence
ij = ji .
(7.72)
We can define matrices inverse to those appearing in [7.69] and [7.70] so that
K = ej ej and (K T ) = i i . Basis duality enforces ej = j and i = ei .
Differentiating j ei = ij , one finds that
dj = j dei ei = ij ei .
Consequently,
d j = dj (K T ) = ij ei (K T )
and
d j = ij i .
ij
(7.73)
The forms are the Cartan connection forms of space En . Basis like {ei },
which can be defined everywhere on En , are called moving frames (repères
mobiles). Equations [7.68] and [7.73] are the (Cartan) structure equations
of En . We have said that, according to a theorem by Whitney, every differentiable manifold can be locally imbedded (immersed) in some En , for n
large enough. Cartan has used moving frames to analyze the geometry of
general smooth manifolds immersed in large enough euclidean spaces. More
about this subject can be found in Mathematical Topic 10. For an example
of moving frames in elasticity, see Physical Topic 3.3.2.
216
7.3.13 Let us come back to the expressions of the exterior derivative. They
have been given either in the basis-independent form [7.26] or in the natural
basis [7.25]. When a form is given in a general basis,
=
1
p!
j1 j2 jp j1 j2 . . . jp ,
things get more involved as the derivatives of each basis element jk must be
taken into account also. Suppose the basis { i } and its dual {ej }. Writing
the Cartan 1-form in the basis { i } as ij = j ik k , where the j ik s are
the connection components, eqs.[7.73] and [7.45] tell us that the structure
coefficients are the antisymmetric parts of these components, C j ik = j [ik] ,
and we can choose
j ik = ejm ek (em
i ).
In terms of this connection, a covariant derivative j is defined whose
components just appear in the expression for the exterior derivative in a
general basis:
1
d = ()p (p+1)!
e[jp+1 j1 j2 jp ] + k [jp+1 jr j1 j2 jr1 kjr+1 jp ]
j1 j2 . . . jp jp+1
1
= ()p (p+1)!

[p+1 j1 j2 jp ] j1 j2 . . . jp jp+1 ,
(7.74)
so that finally, eq.[7.25] generalizes to

d = j0 j0 ( j0 )j0 .
(7.75)
This is a bit more general than the well known formulae giving the differential
in general coordinate systems, and reduce to them for natural basis.
7.3.14 Frobenius theorem, alternative version We have said ( 6.4.33)
that a set of linearly independent tangent vector fields X1 , X2 , . . . , Xn on a
manifold M are locally tangent to a submanifold N (of dimension n < m)
around a point p M if they are in involution, [Xj , Xk ] = ci jk Xi . This
means that, if we take such fields as members of a local basis {Xa } on M ,
with a = 1, 2, . . . , n, n + 1, . . . , m, the structure coefficients ca jk vanish whenever a n + 1. The dual version is the following: consider a set of Pfaffian
forms 1 , . . . , n . Linear independence means that their exterior product is
nonvanishing (Cartan lemma, Mathematical Topic 10.1.2). If such forms are
to be cotangent to a submanifold, they must close the dual algebra to the
involution condition, and we must have eq.[7.46] for the ci jk s restricted to
7.4. DUALITY AND CODERIVATION
217
the indices i, j, k = 1, 2, . . . , n, and with the others vanishing. This is to say

that di has only contributions along the k s, that is,
di 1 2 . . . n = 0.
This may be shown to be equivalent to the existence of a system of functions
f 1 , . . . , f n such that j = ajk df k . The set {df k } constitutes a local coordinate
basis cotangent to the submanifold N , which is locally fixed by the system of
equations f k = ck (constants). The characterization of (local) submanifolds
by forms is a most convenient method. Global cases are well known: the
(x, y)-plane in the euclidean E3 with cartesian coordinates may be characterized by dz = 0. The sphere S 2 given by r = (x2 + y 2 + z 2 )1/2 , simply by
dr = 0. Notice that the forms characterizing a surface is orthogonal to it.
The straight line given by the x-axis will be given by dy = 0 and dz = 0.
7.4
DUALITY AND CODERIVATION

In a metric space of dimension m, there is a duality between p-forms and
(m p)-forms, and the exterior derivative has a doppelganger.
7.4.1 In some previous examples, we have been using relationships between forms like Uk = 2!1 kij Tij (in E3 ) and Fe = 2!1 F (in Minkowski
space). These are particular cases of a general relation between p-forms and
(n p)-forms on a manifold N . Recall that the dimension
of the space p of

n
n
p-forms on an n-dimensional manifold is p = np . Of course, the space of
(n p)-forms is a vector space of the same dimension and so both spaces are
isomorphic. The presence of a metric makes of this isomorphism a canonical
one.
7.4.2 Given the Kronecker symbol i1 i2 ...in and a metric g, we can define
mixed-index symbols by raising some of the indices with the help of the
contravariant metric,
j j ...j
p
1 2
j1 k1 j2 k2
ip+1
g
. . . g jp kp k1 k2 ...kp ip+1 ...in .
...in = g
A detailed calculation will show that

j1 j2 ...jp ip+1 ...in i1 ...ip ip+1 ...in =
(np)!
g
j j ...j
i11...i2 p p ,
where g = det(gij ) is the determinant of the covariant metric.
218
7.4.3 The dual of a form We shall give first a definition in terms of

components, which is more appealing and operational. For a basis i1 i2
. . . ip for the space p of p-forms, we shall define a duality operation ,
called the Hodge star-operation, by
: p (N ) np (N )
|g|
jp
j2
j1
[ . . . ] = (np)! j1 j2 ...jp jp+1 ...jn jp +1 . . . jn .
(7.76)
Here |g| is the modulus of g = det(gij ). A p-form will be taken into its
dual, the (n p)-form
h
i
j1
j2
jp
1
= p! j1 j2 jp . . .
|g|
= (np)! p! j1 j2 ...jp jp+1 ...jn j1 j2 ...jp jp+1 . . . jn ,
or
|g|
(np)! p!
j1 j2 ...jp j1 j2 ...jp jp+1 . . . jn .
(7.77)
Notice that the components of are
|g|
( )jp+1 ...jn = p! j1 j2 ...jp jp+1 ...jn j1 j2 ...jp .
(7.78)
The examples referred to at the beginning of this paragraph are precisely of

this form, with the euclidean metric of E3 and the Lorentz metric respectively.
Although we have used a basis in the definition, the operation is in reality
independent of any choice of basis. The invariant definition is given below
eq.[7.83].
7.4.4 Consider the 0-form which is constant and equal to 1. Its dual will
be the n-form
|g|
v = 1 = n! j1 j2 ...jn j1 j2 . . . jn
(7.79)
p
1
2
n
(7.80)
= |g| . . . ,
which is an especial volume form ( 7.1.5) called the canonical volume form
corresponding to the metric g. Given the basis {ej }, dual to { j },
v(e1 , e2 , . . . , en ) =
|g|
n!
j
1 (e1 ) j1 (e2 )
j
2 (e1 ) j2 (e2 )
j1 j2 ...jn
...
j. . .
n (e1 ) j2 (e1 )
. . . j1 (en )
. . . j2 (en )
...
...
jn
. . . (en )
219
or
v(e1 , e2 , . . . , en ) =
|g|
n!
j1 j2 ...jn j1 j2 ...jn =
p
|g|
(7.81)
by [7.1] and [7.4]. This could of course have been obtained directly from
[7.80]. Given an arbitrary set of n fields X1 , X2 , . . . , Xn ,
v(X1 , X2 , . . . , Xn ) =
|g|
n!
X1j1 X2j1
X1j2 X2j2
... ...
X1jn X2jn

j1 j2 ...jn

. . . Xnj1
. . . Xnj2
. . . . . .
. . . Xnjn
or
v(X1 , X2 , . . . , Xn ) =
|g| det (Xij )
(7.82)
7.4.5 Invariant definition Take p fields X1 , X2 , . . . , Xp and call X10 , X20 ,

. . . , Xp0 their respective covariant images. Then, the form , dual to the
p-form , is defined as that unique (n-p)-form satisfying
(X1 , X2 , . . . , Xp ) v = ( ) X10 X20 . . . Xp0
(7.83)
for all sets of p fields {X1 , X2 , . . . , Xp }. This is the invariant, basis independent, definition of operator . It coincides with [7.77], and this tells us that
what we have done there (and might be not evident), supposing that the
operator ignores the components, is correct.
7.4.6 Let us go back to eq.[7.76] and check what comes out when we apply
twice the star operator:
[ j1 j2 . . . jp ]
=
=
|g|
(np)! p!
|g|
(np)! p!
j1 j2 ...jp jp+1 jp+2 ...jn jp+1 jp+2 ...jn i1 i2 ...ip i1 i2 . . . ip

g j1 k1 g j2 k2 . . . g jp kp k1 k2 ...kp jp+1 jp+2 ...jn g jp+1 ip+1 g jp+2 ip+2 . . .
. . . g jn in ip+1 ...in i1 ...ip i1 . . . ip
Now,
k1 k2 ...kp jp+1 jp+2 ...jn g jp+1 ip+1 g jp+2 ip+2 . . . g jn in g j1 k1 g j2 k2 . . . g jp kp
= j1 j2 ...jp ip+1 ip+2 ...in det (g ij ) = j1 j2 ...jp ip+1 ip+2 ...in g 1 ,
220
so that
[ j1 j2 . . . jp ]
=
|g|
1
g (np)! p!
j1 j2 ...jp ip+1 ip+2 ...in ()p(np) i1 i2 ...ip ip+1 ...in i1 . . . ip
j j ...j
|g|
1
()p(np) i11i22...ipp (n p)! i1 . . . ip
g (np)! p!
1
= |g|
()p(np) (n p)!p! i1 . . . ip
g (np)! p!
= |g|
()p(np) i1
g
. . . ip .
Thus, taking twice the dual of a p-form yields back the original form up to
a sign and the factor |g|/g. The components of a metric constitute always a
symmetric matrix, which can always be put in diagonal form in a convenient
basis. The number of positive diagonal terms (P ) minus the number of
negative diagonal terms (N ), s = P N = (n N ) N = n 2N is
an invariant property of the metric (a theorem due to Sylvester), called its
signature.7 Minkowski metric, for instance, has signature s = 2. The factor
|g|/g is simply a sign ()N = ()(ns)/2 . We find thus that, for any p-form,
p = ()p(np)+(ns)/2 p ,
(7.84)
so that the operator inverse to is

1 = ()p(np)+(ns)/2
(7.85)
when applied to a p-form and the metric has signature s. Of course, |g|/g = 1
for a strictly Riemannian metric and for this reason the signature dependence
of 1 is frequently ignored in textbooks.
7.4.7 Let {ej } be a basis in which the metric components are gij . The
metric volume element introduced in 7.4.4 could have been alternatively
defined as the n-form v such that
v(e1 , e2 , . . . , en ) v(e1 , e2 , . . . , en ) = det (gij ) = g.
(7.86)
In reality, this only fixes v up to a sign. The manifold N has been supposed
to be orientable and the choice of the sign in this case corresponds to a choice
of orientation.
7.4.8 An inner product (, ) between two forms of the same order can
then be introduced: it is such that
( ) = (, ) v.
(7.87)
It generalizes the inner product generated by the metric on the space of the
1-forms.
7
We could have defined s = N P instead, without any change for our purposes.
221
7.4.9 In E3 , it comes out immediately that

dx1 = dx2 dx3 ;
dx2 = dx3 dx1 ;
dx3 = dx1 dx2 .
Given two 1-forms i dxi and j dxj ,

= (i i ) dx1 dx2 dx3 .
7.4.10 The possibility of defining the above inner product comes from the
following property (true for forms and of the same order):
( ) = ( ).
(7.88)
7.4.11 We have already used the star operator in 7.2.8. The trivial character of the euclidean metric has hidden it somewhat, but the correspondence
between vectors and second order antisymmetric tensors in E3 is given precisely by the star operator. It is essential to the definition of the laplacian of
a function f, f = divgradf . It is a simple exercise to check that
d df = (f )dx dy dz.
7.4.12 Take Minkowski space, with g00 = 1 and gii = 1, in the cartesian basis {dxi }, and with the convention 0123 = +1:
dx1 =
1
3!
p
|g| 1 dx x dx = 1 230 dx2 dx3 dx0
= dx2 dx3 dx0 = dx0 dx2 dx3 ;
dx2 = dx0 dx3 dx1 ;

dx3 = dx0 dx1 dx2 ;
dx1 dx2 ) = dx0 dx3 , etc;
(dx0 dx1 ) = dx2 dx3 ;
(dx1 dx2 dx3 ) = dx0 ;
(dx0 x1 dx2 ) = dx3 , etc.
The dual to the 1-form A = A dx will be
A =
For the 2-form F =
F =
1
2!
1
3!
|g| A x x x .
F dx dx ,
1 1
[
2! 2!
F ]dx dx =
1
2!
Fe dx dx .
222
7.4.13 In a 4-dimensional space, the dual of a 2-form is another 2-form.

One could ask in which circumstances a 2-form can be self-dual (or antiselfdual), F = F . This would require, from eq.[7.84], that
F = F = [ F ] = F = ()(4s)/2 F = ()s/2 F.
In Minkowski spaces, self-duality of F implies the vanishing of F . In an
euclidean 4-dimensional space non-trivial selfduality is quite possible. In
gauge theories, selfdual euclidean fields are related to instantons.
7.4.14 The coderivative de of a p-form is defined by
e := ()p 1 d = ()n(p1)+(ns)/2 d .
d
(7.89)
Perhaps codifferential would be a more appropriate name, but coderivative is more usual. A quick counting will tell that this additional exterior
differentiation takes a p-form into a (p 1)-form. There is more:
dede = ()p1 ()p 1 d 1 d = 1 d d 0.
(7.90)
7.4.15 A form such that de = 0 is said to be coclosed . A p-form such

that a (p + 1)-form exists satisfying = de is coexact. In components,
dep+1 = p!1 [ j ji1 ...ip ] dxi1 xi2 . . . dxip
in a natural basis. Notice what happens with the components: each one will
consist of a sum of derivatives by all those basis elements whose duals are
not in the form basis at the right. The coderivative is a generalization of
the divergence, and is sometimes also called divergence. In a general basis,
corresponding to eq.[7.74] for the exterior derivative, a lengthy calculation
gives the expression
dep+1 = p!1 [j ji1 ...ip ] i1 i2 . . . ip .
(7.91)
Still another expression will be given in 7.6.13. Only after that an expression
will be found for the coderivative de(p q ) of the wedge product of two
forms.
7.4.16 par:7laplacian Now, a laplacian operator can be defined which acts
on forms of any order:
e 2 = d de + ded.
:= (d + d)
(7.92)
223
On 0-forms, reduces (up to a sign!) to the usual Laplace-Beltrami operator acting on functions. Notice that the laplacian of a p-form is a p-form.
Harmonic analysis can be extended to antisymmetric tensors of any order.
A p-form such that = 0 is said to be harmonic. The harmonic p-forms
constitute a vector space by themselves. From the very definition of , a
form simultaneously closed and coclosed is harmonic.
The laplacian has the commutation properties
d = d;
= ;
e
de = d.
If A is a 1-form in E3 , in which the trivial metric allows identification of

1-forms and vectors, = d de + ded is the usual formula of vector calculus
A = grad div A rot rot A .
7.4.17 Maxwells equations, second pair Using the results listed in
7.4.12, eq.[7.29] gives easily
F = H1 dx1 dx0 H2 dx2 dx0 H3 dx3 dx0
+ E1 dx2 dx3 + E2 dx3 dx1 + E3 dx1 dx2 , (7.93)
or
F = Hi dx0 xi + 21 ijk Ei dxj dxk .
(7.94)
Comparison with eq.[7.29] shows that the Hodge operator takes H E and
E H. It corresponds so to the usual dual transformation in electromagnetism, a symmetry of the theory in the sourceless case. If we calculate the
coderivative of the electromagnetic form F , we find
~ Edx
~ 0 + (0 E
~ rotH)
~ d~x.
deF =
(7.95)
We have seen in 7.2.9 that the first pair of Maxwells equations is summarized in dF = 0. Now we see that, in the absence of sources, the second
pair
~ E
~ = 0 and 0 E
~ = rotH
~
(7.96)
is equivalent to deF = 0. The first pair is metric-independent. The coderivative is strongly metric-dependent, and so is the second pair of Maxwells
equations. Equations [7.31] and [7.96] tell us that, in the absence of sources,
F is a harmonic form. The first pair does not change when charges and
currents are present, but the the second does: the first equation in [7.96]
acquires a term 4 in the right-hand side, and the second a term 4J. If
we define the current 1-form
h
i
j := 4 dx0 J~ d~x ,
(7.97)
224
Maxwells equations become

dF = 0 and deF = j.
(7.98)
The current form is coexact, as the last equation implies

dej = 0,
(7.99)
~ J~ = 0,
0 +
(7.100)
or, in components,
which is the continuity equation: charge conservation is a consequence of
the coexactness of the current form. Notice that this is metric dependent.
In the presence of charges, the electromagnetic form is no more harmonic,
but remains closed. Consequently, in every contractible region of Minkowski
space there exists a 1-form A = A dx such that F = dA. From eq.[7.98] we
see that this potential form obeys the wave equation
dedA = j,
(7.101)
A + A = j .
(7.102)
or, in components,
The Lorenz gauge is the choice deA = A = 0.
7.4.18 Here is a list of relations valid in Minkowski space, for forms of
degrees 0 to 4 (we use the compact notations dx = dx dx , dx =
dx dx dx , dx dx dx dx dx ; the bracket [, , . . .] means a
complete antisymmetrization in the included indices).
form
f
A dx
1
F dx
2!
1
W dx
3!
1
V
dx
4!
f dx1230
1
A dx
3!
1 1
[ F ]dx
2! 2!
W [ dx]
V1230
d
( f )dx
1
A dx
2! [ ]
1
F dx
3! [ ]
1
W dx
4! [ ]
0
de
0
A
F dx
2!1 W dx
1
V1230 dx
3!
7.4.19 With the notation of ( 7.3.8), the field equations for gauge theories
are the Yang-Mills equations
F a + f a bc Ab F c = J a ,
(7.103)
where J a is the source current. This is equivalent to

c
[ Fea ] f a bc Ab[ Fe]
= Jea []
(7.104)
7.5. INTEGRATION AND HOMOLOGY
225
In invariant notation, [7.103] reads

e + 1 [A, F ] = J.
dF
(7.105)
Thus, in the sourceless case, the field equations are just the Bianchi identities
[7.65], [7.66] written for the dual of F . A point of interest of self-dual fields
( 7.4.13) is that for them the sourceless Yang-Mills equations coincide with
the Bianchi identities. Any F of the form F = dA + A A will solve the field
equations. Recall the discussion of 7.3.11 on types of covariant derivatives,
depending on the form degrees. The expression at the left-hand side of the
equations above is the covariant coderivative of the 2-form F according to
the connection A, as will be seen below.
7.4.20 Covariant coderivative To adapt the discussion ( 7.3.11) on
covariant derivatives to the case of coderivatives, we simply notice that the
transformations in the value (vector) space ignore the tensor content of the
form, so that ( W )0 = W 0 = g W g 1 . We can therefore apply the same
reasoning to obtain
d W 0 + |[A0 , W 0 ]| = g{d W + |[A, W ]|}g 1 .
e = ()W 1 d W , we apply ()W 1 to this expression to find the
As dW
covariant coderivative
e + ()W 1 |[A, W ]|.
e = dW
DW
(7.106)
For W belonging to a linear representation,

e + ()W 1 A W.
e = dW
DW
7.5
(7.107)
INTEGRATION AND HOMOLOGY
We present here a quick, simplified view of an involved subject. The excuse for quickness
is precisely simplicity. The aim is to introduce the crucial ideas and those simple notions
which are of established physical relevance.
7.5.1
Integration
7.5.1 Let us go back to our small talk of Section 7.1, where it was said that
exterior differential forms are entities living under the shadow of the integral
sign. We shall not have space for any serious presentation of the subject
only the main ideas and results will find their place here. The fundamental
226
idea is as simple as follows: suppose we know how to integrate a form on

a domain D of an euclidean space Em+n . Suppose else that a differentiable
mapping f is given, which maps D into some subset of the differentiable
manifold M ,
f : D f (D) M .
A form defined on M will be pulled back to a form f on D . The integral
of on f (D) is then defined as
Z
Z
=
f .
(7.108)
f (D)
7.5.2 Notice the importance of the pulling-back behaviour of forms. It

is possible to show (this unforgivable phrase) that, given two diffeomorphisms
between the interiors of D and f (D), then they both lead to the same result:
the definitions would then differ by a change of coordinate systems. Interior, let us recall, means the set minus its boundary. Boundaries have been
defined in 2.2.17 for chains, which directs us to another point of relevance:
the integration domains will be chains.
7.5.3 We have in 2.2.5 introduced polyhedra, which are the sets of points
of simplicial complexes in euclidean spaces. We went further in 2.2.6, defining curvilinear polyhedra and curvilinear simplexes on a general topological
space as subsets of these spaces which are homeomorphic to polyhedra and
simplexes in some euclidean space. The line of thought is the following:
first, define integration on euclidean simplexes; second, choose the mapping
f to be a differentiable homeomorphism, fixing curvilinear simplexes on the
manifold M ; third, by using eq.[7.108], define integration on these simplexes;
finally, extend the definition to general p-chains on M . These last chains are
defined in the same way, as sets homeomorphic to euclidean chains: they are
usually called singular p-chains.
The qualification singular is added because the homeomorphism f is only
one-way differentiable and is not a diffeomorphism (see below, 7.5.16 and
7.5.22).
7.5.4 The integration of a p-form on a simplex or polyhedron P on Ep ,
in which it has the expression
= (x1 , x2 , . . . , xp )dx1 dx2 . . . xp ,
is defined simply as the usual Riemann integral
Z
Z
:=
(x1 , x2 , . . . , xp )dx1 dx2 . . . dxp ,
P
as a limit of a summation.
(7.109)
227
7.5.5 The definition of integration on a domain in a differentiable manifold

M will require:
(i) the choice of an orientation in Ep ;
(ii) the choice of a differentiable homeomorphism f , defining on M a
corresponding curvilinear polyhedron or simplex f (P ).
Then, the integral of a p-form will be
Z
Z
:=
f ,
(7.110)
P Ep
f (P )M
f .
:=
f (Sp )
(7.111)
Sp
7.5.6 A singular p-chain p on M is simply obtained: given on Ep a pchain

cp = m1 Sp1 + m2 Sp2 + . . . + mj Spj ,
then
p = m1 p1 + m2 p2 + . . . + mj pj ,
where pj = f (Spj ) is a singular p-simplex.
The coefficients mj are the multiplicities of the corresponding curvilinear simplexes
pj . Intuitively, they give the number of times one integrates over that simplex.
7.5.7 Integration on the p-chain p is now

Z
:=
p
j
X
i=1
Z
mi
=
pi
j
X
i=1
Z
mi
f .
(7.112)
Spi
The boundary of a singular chain is defined in a natural way: first, the boundary of a curvilinear simplex is the image of the restriction of the mapping
f to the boundary of the corresponding euclidean simplex, pj = f (Spj ).
Thus,
j
X
p =
mi pi .
(7.113)
i=1
As the mapping f is a homeomorphism, it will take the empty set into the
empty set. Thus, pj = f (Spj ) = f () = 0, and
p 0.
(7.114)
228
7.5.8 The above definition of singular chains carries over to them all the
algebraic properties of euclidean chains. It is then possible to define closed
singular chains, or cycles, as well as exact singular chains, and transpose the
same homology groups to chains on differentiable manifolds.
7.5.9 We are now ripe to state one of the most important theorems of all
Mathematics: the integral of the exterior derivative of a form, taken on a
singular chain, equals the integral of the form taken on the boundary of that
chain:
Z
Z
d =
.
(7.115)
It is called by a great mathematician of today (i.e., Arnold) the Newton Leibniz - Gauss - Ostrogradski - Green - Stokes - Poincare theorem, a name
tracing its historical evolution. Its fantastic generality was recognized by the
last member of the illustrious team, and includes as particular cases all the
vector analysis theorems associated to those eminent names. The first two
patriarchs belong because, if f is a real function and is the real oriented
interval (a, b), then we have that = b a, and
Z
Z
df =
Z
f = f (b) f (a).
f=
a
(7.116)
Most authors call the theorem after the last-but-one member of the list.
7.5.10 Although we shall not pretend to demonstrate it, let us give only
one indication: it is enough to suppose it valid for the euclidean case:
Z
Z
Z
d =
f d =
d(f );
=f (S)
if it holds for the euclidean chain S, we can proceed:

Z
Z
Z
=
f =
=
.
S
f (S)
(7.117)
As to the demonstration for euclidean chains, it follows the general lines

of the demonstrations of (say) Gauss theorem in vector analysis, with the
necessary adaptations to general dimension and order.
7.5.11 An immediate consequence of Stokes theorem, eq.[7.115], is that
the integral of any closed p-form on any p-boundary is zero.
229
7.5.12 A few examples in electromagnetism on E3 . We shall alternate the

short-hand notation we have been using with more explicit and usual ones.
(i) take a closed curve in E3 , and let S be some piece of surface bounded
by : = S. The circulation of the vector potential A along will be
Z
Z
Z
Z
Z
Z
A=
A dl =
dA = (rotA) d =
H d =
H,
=S
the flux of H through S.

(ii) take now the flux of H through a closed surface (say, a sphere S 2
enclosing a ball B 3 ), which is
Z
Z
Z
Z
Z
H=
H d =
H d =
dH =
(divH).
S2
S2
B 3
On the other hand,

Z
Z
H=
S2
B3
Z
(rot A) d =
dA =
S2
B3
S2
A = 0.
S 2 =0
As the ball is of arbitrary size, this implies div H = 0.
(iii) Faraday law of induction: the circulation of E along a closed line is

Z
I
Z
Z
Z
Z
E = E dl =
E=
dE = (rotE) d =
0 H d.
=S
As already said, H is a 2-tensor; by using the Hodge operation, it can be confused with
e = rot H.
an (axial) vector. Then dH = div H and dH
7.5.13 Total electric charge in a closed universe In the Friedmann

model for the universe, strongly favored by observational evidence,8 it remains to be decided by measurements whether the space part of the universal space-time is an open (infinitely extended, infinitevolume) or a closed
(finite volume) manifold. The model reduces the possibilities to only these
two. Let us consider the closed case, in which the manifold is a 3-dimensional
(expanding) sphere S 3 . The total electric charge is given by Gauss law,
Z
Z
Z
Q = 4
=
E =
(i Ei )d3 x.
S3
S3
S3
But i Ei = dE, with E = Ei dxi . Consequently,

Z
Z
Q=
dE =
E = 0,
S3
S 3
because S 3 = 0. The argument can be adapted to an open universe, but

then convenient conditions at large distances are required.
8
See for example Weinberg 1972, chap.15.
230
7.5.14 Consider a force field on E3 : suppose that the work necessary for a
displacement between two points a and b is independent of the path (Figure
7.6). This is the same as saying that the integral along any closed path
through a and b is independent of that closed path, and thus vanishes.
For a
H
i
conservative system, the work W = Fi dx is a closed form. From Fi dxi = 0
Figure 7.6: Paths between two points in a force field.

we deduce that the force is of potential origin, F = dU for some U , or
Fi = (U/xi ).
7.5.15 This is a particular case of a very important theorem by De Rham. Let us,
before enunciating it, introduce a bit more of language: the integral of a p-form on a
p-cycle is called the period of on . Clearly, the Stokes theorem implies that all the
periods of an exact form are zero. The theorem (first theorem of) De Rham has proved
says that a closed form whose periods are all vanishing is necessarily exact. Notice that
this holds true for any smooth manifold, however complicated its topology may happen
to be.
7.5.16 Let us make a few comments on the simplified approach adopted

above. Integration is a fairly general subject. It is not necessarily related
to topology, still less to differentiability. It requires a measure space, and
for that the presupposed division of space is not a topology, but a -algebra
(Mathematical Topic 3). We are therefore, to start with, supposing also a
compatibilization of this division with that provided by the topology. In
other words, we are supposing a covering of the underlying set by Borel sets
constituting the smallest -algebra generated by the topology. It is not excluded that two similar but different topologies generate the same -algebra.
Only up to these points do the above statements involve (and eventually
probe) the topology. But there are restrictions. For example, to be sure
of the existence of a Borel measure, the topology must be Haussdorf and
locally compact. The mapping f of 7.5.1 is clearly a bit too strong to be
231
assumed to exist for any domain D. We have later assumed the existence of
a differentiable homeomorphism from euclidean spaces to the differentiable
manifolds to the extent of taking chains from one to the other, etc, etc. It
should however be said, in favour of the above exposition, that it allows a
fair view of the main relationships of integration to the global characteristics of the space. The interested reader is urged to consult more complete
treatments.9
7.5.17 On hypersurfaces10 A hypersurface is an (n 1)-dimensional
space immersed in En . It may be an imbedded manifold, or be only locally differentiable. We shall here suppose it an imbedding for the sake of
simplicity. Suppose in En a hypersurface given by the equation (x) =
(x1 , x2 , . . . , xn ) = 0. Then to will correspond a special (n 1)-form, its
volume form , in the following way. A requirement will be that the surface
be nonsingular, grad = d 6= 0, at least locally. Let v be the En volume
form, v = dx1 dx2 dxn . Then is defined by v = d . Around
a point p one is in general able to define new coordinates {uk } with positive
x
|| and such that one of them, say uj , is (x). Then,
jacobian || u

x
v = du1 du2 duj1 d duj+1 dun .
(7.118)
u
If around a point p it so happens that j 6= 0, we may simply choose
ui6=j = xi and uj = (x), in which case
= ()j1
dx1 dx2 dxj1 dxj1 dxn

.
j
(7.119)
A trivial example is the surface x1 = 0, for which = dx2 dx3 dxn .

These notions lead to the definition of hypersurface-concentrated distributions (generalized Dirac functions), given through test functions f on En
by
Z
Z
Z
((), f ) =
()f = f = f (x) .
En
Also the generalized step-function () = 1 for (x) 0, and () = 0 for

(x) < 0, can be defined by 0 () = (), with the meaning
j = (j ) ().
It is sometimes more convenient to use an inverted notation. If is the
boundary of a domain D, we may want to use the characteristic function of
9
10
Such as Choquet-Bruhat, DeWitt-Morette & Dillard-Bleick 1977.

Gelfand & Shilov 1964.
232
D, a function which is 1 inside D and zero outside it. If D is the set of points
x such that (x) 0, we define () = 1 for (x) 0 and () = 0 for
(x) > 0, so that () is the characteristic function of D. In this case we
have that 0 () = (), and
grad = () grad .
This corresponds to the usual practice of using, on a surface, the normal
directed outward, and leads to general expressions for well known relations
of vector analysis.
Such distributions are of use, for example, in getting Maxwells equations
in integral form (Physical Topic 4.3).
7.5.2
Cohomology of differential forms
We begin by recalling and in the meantime rephrasing some of the comments about
chains and their homologies, made in chapter 2. Then we describe the dual structures
which are found among the forms, cohomologies. Although cohomology is an algebraic
structure with a very wide range of applications, differential forms are a subject of choice
to introduce the main ideas involved.11 Here we shall suppose all the chains already
displayed on a general smooth manifold M .
7.5.18 Homology, revisited A chain is a cycle (or is closed ) if it has

null boundary: = 0. It is a boundary if another chain exists such
that = . Closed p-chains form a vector space (consequently, an additive
group) Zp . Boundaries likewise form a vector space Bp for each order p. Two
closed p-chains and whose difference is a boundary are homologous. In
particular, a chain which is itself a boundary is homologous to zero. Now,
homology is an equivalence relation between closed forms. The quotient of
the group Zp by this relation is the homology group Hp = Zp /Bp . For compact
manifolds all these vector spaces have finite dimensions. The Betti numbers
bp = dim Hp are topological invariants, that is, characteristics of the topology
defined on M .
7.5.19 Cohomology Now for forms: a form is closed (or is a cocycle)
if its exterior derivative is zero, d = 0. It is exact (or a coboundary)
11
An excellent short introduction to this subject can be found in Godbillon 1971.

Another excellent text, with an involved approach to many physical problems, including
a rather detailed treatment of the decomposition theorems (and much more) is Marsden
1974. Finally, a treatise whose reading requires some dedication, and deserves it, is De
Rham l960. For some pioneering applications of cohomological ideas to Physics, see Misner
& Wheeler l957.
233
if another form exists such that = d. Closed p-forms constitute a

vector space, denoted Z p . The same happens to exact forms, which are in a
vector space B p . Two closed p-forms are said to be cohomologous when their
difference is an exact form. In particular, an exact form is cohomologous to
zero. Cohomology is an equivalence relation. The quotient of the cocycle
group Z p by this relation is another group, De Rhams cohomology group
H p = Z p B p . Again, for compact manifolds, all these vector spaces are of
finite dimension, bp = dim H p . Another fundamental result by De Rham is
the following:
for compact manifolds, the homology group Hp and the cohomology group
Hp are isomorphic.
So, the Betti numbers are also the dimension of H p : bp = bp .
Roughly speaking, the number of independent p-forms which are closed
but not exact is a topological invariant. This establishes a strong relation
between forms and the topology of the manifold. Differential forms play just
the role of the cochains announced in 2.2.19.
7.5.20 The above results, and the two identities, 2 0 and d2 0,
show the deep parallelism between forms (integrands) and chains (integration
domains). For compact unbounded manifolds and closed forms, this parallelism is actually complete and assumes the characteristics of a duality: given
a closed p-form p and a closed p-chain p , we can define a linear mapping
Z
p
p
p
p
p R,
(7.120)
, < , > :=
p
which has all the properties of a scalar product, just as in the case of a vector
space and its dual. This product is in reality an action between homology
and cohomology classes, and not between individual forms and chains.
This is a consequence of the two following properties:
(i) The integral of a closed form over a cycle depends only on the homology class
of the cycle : if = , then
Z
Z
Z
Z
Z
Z
Z
=
= +
= + d = .
(ii) The integral of a closed form over a cycle depends only on the cohomology
class of the form : if = d, then
Z
Z
Z
Z
Z
Z
Z
= ( + d) =
+ d =
+
d =
.
234
7.5.21 On a compact metric manifold, the star operator allows the definition of a global inner product between p-forms through an integration over
the whole manifold M :
Z
(, ) :=

(7.121)
M
This is a symmetric bilinear form. A one-to-one relation between forms and

chains is then obtained by using (7.120) and (7.121): the chain can be
identified with the form if, for every form ,
Z
Z
.
(7.122)
< , >= = ( , ) =
7.5.22 A few comments: first, the compactness requirement is made to ensure the
existence of the integrals. It can be softened to the exigency that at least one of the
forms involved has a compact carrier: that is, it is different from zero only on a compact
domain. Second, the complete duality between forms and chains really requires something
else: chains on general smooth manifolds have been introduced through mappings which
are only locally homeomorphisms, and nothing else has been required concerning the
differentiability of their inverses. That is why they are called singular chains. On the
other hand forms, as we have introduced them, are fairly differentiable objects. The above
relation between forms and chains only exists when forms are enlarged so as to admit
components which are distributions. Then, a general theory can be built with forms and
chains as the same objects this has been done by Schwartz and De Rham, who used for
the new general objects, including both forms and chains, the physically rather misleading
name (proposed by Schwartz) current.
7.5.23 We have seen that the homology groups on a topological manifold

can be calculated in principle by the methods of algebraic topology. For
compact manifolds, those results can be translated into results concerning
the cohomology groups, that is, into properties of forms defined on them.
The simplest compact (bounded) manifolds are the balls B n imbedded in
En . Their Betti numbers are
b0 (B n ) = bn (B n ) = 1 ; bp (B n ) = 0 for p = 1, 2, . . . , n 1.
(7.123)
Here, another caveat: we have mainly talked about a particular kind of homology, the so-called integer homology, for chains with integer multiplicities.
The parallelism between forms and chains is valid for real homology: chains
with real multiplicities are to be used. The vector spaces are then related to
the real numbers, and the line R takes the place of the previously used set of
integer numbers Z. In the above example, the space H 0 (B n ) is isomorphic
to the real line R. This means that 0-forms (that is, functions) on the balls
235
can be closed but not exact, although in this case they will be constants:
their set is isomorphic to R. This trivial result is no more valid for p 6= 0: in
these cases, every closed form is exact. This is simply a pedantic rephrasing
of what has been said before, since B n is contractible.
7.5.24 The next simplest compact manifolds are the n-dimensional spheres S n imbedded in En+1 . For them,
H 0 (S 0 ) n Rn2
Hn0)(Sn )0 when
H (S ) p6=R0, n.
H p (S
(7.124)
On S 4 , for instance, closed 1-, 2-, and 3-forms are exact. All 4-forms (which
are of course necessarily closed) and closed functions are constants. On the
sphere S 2 , every closed 1-form (irrotational covariant vector field) is exact
(that is, a gradient). All 2-forms and closed functions are constant.
7.5.25 Spheres are the simplest examples of compact manifolds. Life is
much simpler on such manifolds, on which the internal product (7.121) has
many important properties. Take for instance a p-form and a (p 1)-form
. Then,
Z
(d, ) =
d
M
Z
Z
Z

p
p1
d
d( ) + ()
=
d( ) () d =
M
M
M
Z
Z
Z

p 1
=
+
() d =
de.
M
Consequently,
e
(d, ) = (, d).
(7.125)
The operators d and de are, thus, adjoint to each other in this internal product.
It is also easily found that 1 is adjoint to and that the laplacian
e
= dde + dd
is self-adjoint:
(, ) = (, ).
(7.126)
There is more: on compact-without-boundary strictly Riemannian manifolds,

the internal product can be shown to be positive-definite. As a consequence,
each term is positive or null in the right hand side of
e ) + (dd,
e ) = (d,
e d)
e + (d, d).
(, ) = (dd,
Hence, in order to be harmonic, as to be both closed and coclosed. This
condition is, on such manifolds, necessary as well as sufficient.
236
7.5.26 Kodaira-Hodge-De Rham decomposition theorem This theorem, in its present-day form, is the grown-up form of a well known result,
of which primitive particular versions were known to Stokes and Helmholtz.
Called a fundamental theorem of vector analysis by Sommerfeld,12 it says
that a differentiable enough vector field V in E3 , with a good enough behaviour at infinity, may be written in the form
V = grad f + rot T + c,
where c is a constant vector. In its modern form, it is perhaps the deepest
result of the above general harmonic analysis. It says that the inner product
divides the space of p-forms into three orthogonal sub-spaces. In a rather
weak version,
on a compact-without-boundary manifold, every form can be decomposed in
a unique way into the sum of one exact, one co-exact and one harmonic
form:
= d + de + h,
(7.127)
with h = 0. The authors the theorem is named after have shown that in
e and =
reality, with the above notation, a form exists such that = d
d, which puts as the sum of a laplacian plus a harmonic form:
= + h.
(7.128)
In consequence, no exact form is harmonic unless it is also coexact and

belongs to the harmonic subspace: no harmonic form is purely exact, and so
on. In particular, no harmonic form can be written down as h = d.
7.5.27 All these properties have been found when people were studying
the solutions of the general Poisson problem
= .
(7.129)
It has solutions only when belongs exclusively to the laplacian sector, or

when is not harmonic.
7.5.28 It is not difficult to verify that the harmonic forms constitute themselves still another vector space. Another fundamental theorem by De Rham
says the following:
12
Sommerfeld 1964a, 20.
237
On a compact-without-boundary manifold, the space of harmonic p-forms is

isomorphic to the cohomology spaceH p (M ).
Thus, if p is the laplacian on p-forms,
bp = dim ker p = dim H p (M ).
This is of course very useful, because it fixes the number of independent
harmonic forms on the manifold. This number is determined by its topology.
7.5.29 No electromagnetism on S 4 In order to fix the ideas and check
the power of the above results, let us examine a specially simple case: abelian
gauge theories on the sphere S 4 . Instantons are usually defined as solutions
of the free Yang-Mills equations on S 4 , for a given gauge theory on Minkowski
spacetime. The abelian case includes electromagnetism, for which the group
is the 1-dimensional U (1). The Bianchi identity and the Yang-Mills equations
are simply
e a = 0,
dF a = 0 and dF
(7.130)
with an index a for each generator of the gauge group. Each F a is a 2form and the above equations require that F a be simultaneously closed and
coclosed. This is equivalent to require F a to be harmonic. How is the space
of harmonic 2-forms on S 4 ? De Rhams fundamental theorem tells us that
it is isomorphic to H 2 (S 4 ). This is a space of zero dimension. We arrive
to the conclusion that the only solution for [7.130] is the trivial one, the
vacuum F a = 0. We might say then that no instantons exist for abelian
theories or, in particular, that no nontrivial electromagnetism exists on such
a space-time as S 4 . Notice that, the result coming ultimately from purely
topological reasons, it keeps holding for any space homeomorphic to S 4 .
7.5.30 Extensions of the decomposition theorem for the physically more
useful cases of compact manifolds with boundary have been obtained. The
question is far more complicated, because the different possible kinds of
boundary conditions have to be analyzed separately. The operators d and de
are no more adjoint to each other. The boundary term now survives in the
steps leading to eq.[7.125]:
Z
e
(d, ) = (, d) +
.
(7.131)
M
The boundary conditions must be stated in an invariant way. For that, let
us introduce some notation. Let i: M M be the inclusion mapping
238
attaching the boundary to the manifold. Let us introduce the following

metric-dependent notions:
normal part of the form : n = i ( );
(7.132)
tangent part of the form : t = i ().
(7.133)
The form will be parallel (or tangent) to M if n = 0. It is perpendicular

to M if t = 0. Adaptation to a field X is got by recalling that X is in
relation to two forms: (i) its covariant image, a 1-form, and (ii) the (m 1)form iX v, obtained from the volume form v through the interior product (see
7.6.6). Then, X is tangent to M iff its covariant image is tangent to
M , or iff iX v is normal to M . X is normal to M iff iX v is tangent to
M . Also a stronger definition of harmonic forms is needed now: a form h
is harmonic iff d h = 0 and deh = 0 hold simultaneously. Then, a version13 of
the decomposition theorem valid for manifolds-with-boundary is
= d t + den + h.
(7.134)
Other versions are
(i) = d + t with t satisfying det = 0;
(7.135)
(ii) = de + n with n satisfying dn = 0.
(7.136)
7.5.31 A last remark: the inner product is used to obtain invariants. The
action for electromagnetism, for example, is the functional of the basic fields
A given by
Z
I[A] = (F, F ) =
F F,
(7.137)
where F is given by eq.[7.34]. For more general gauge theories, there is

one such field for each generator in the Lie algebra of the gauge group. In
order to obtain an invariant, an invariant metric has to be found also on
the algebra. Once such a metric Kab = K(Ja , Jb ) is given (see 8.4.12), the
action functional is taken to be
Z
a
b
(7.138)
I[A] = K(Ja , Jb )(F , F ) = Kab F a F b ,
with the F a s now as given in [7.64]. Due to their intuitive content, a chapter
on differential forms is a place of choice to introduce cohomology. It should
be kept in mind, however, that cohomology is a very general and powerful
algebraic concept with applications far beyond the above case, which is to be
seen as a particular though important example. We shall meet cohomology
again, in different contexts.
13
See Marsden 1974.
7.6. ALGEBRAS, ENDOMORPHISMS AND DERIVATIVES
7.6
239
ALGEBRAS, ENDOMORPHISMS AND DERIVATIVES
While physicists have been striving to learn some geometry, geometers were progressively
converting their subject into algebra. We have been leaning heavily on analytical terminology and way-of-thinking in the last chapters. In this paragraph we shall rephrase some
of the previous results and proceed in a more algebraic tune.
7.6.1 Let us start by stating (or restating) briefly some known facts. We
refer to Mathematical Topic .1 for more detailed descriptions of algebraic
concepts. An algebra is a vector space V with a binary operation V V V ,
submitted to certain general conditions. It may be associative or not, and
commutative or not.
7.6.2 A given algebra is a Lie algebra if its operation is anticommutative
and satisfies the Jacobi identity. When the algebra is associative but not a
Lie algebra, it can be made into one: starting from any binary operation a
Lie bracket can be defined as the commutator
[, ] = ,
and makes of any associative algebra a Lie algebra.
7.6.3 An endomorphism (or linear operator) on a vector space V is a
mapping V V preserving its linear structure. If V is any algebra (not
necessarily associative, it may be merely a vector space), End V = {set of
endomorphisms on V} is an associative algebra. Then the set [End V ] of its
commutators is a Lie algebra. The generic name derivation is given to any
endomorphism D : V V satisfying Leibniz law:
D() = (D) + (D).
The Lie algebra [End V ] contains D(V ), the vector subspace of all the derivations of V , and the Lie bracket makes of D(V ) a Lie subalgebra of [End V ].
This means that the commutator of two derivations is a derivation. Given an
element V , it defines an endomorphism ad(a) = ada , called the adjoint
action of a, by
ada (b) = [a, b].
This is a derivative because
ada (bc) = [a, bc] = b[a, c] + [a, b]c = ada (b)c + bada (c).
The set ad(A) = {ada for all a A} contains all the internal derivations of
a Lie algebra A, and is itself a Lie algebra homomorphic to A.
240
7.6.4 A graded algebra is a direct sum of vector spaces, V = k Vk , with

a binary operation taking Vi Vj Vi+j . If Vk , we say that k is the
degree (or order) of , and write = k. The standard example of graded
algebra is that formed by the differential forms of every order on a manifold
M,
(M ) k k (M ),
with the exterior product
: p (M ) q (M ) p+q (M )
as the binary operation. Let us go back to forms and introduce another
endomorphism.
7.6.5 Exterior product Let M be a metric manifold, X = X i i be a
vector field on M written in some natural basis, and X 0 its covariant image
X 0 = Xi dxi . As said in 7.1.14, the operation of exterior product by X 0 on
any form is defined by
(X 0 ) : k (M ) k+1 (M )
(X 0 ) = X 0 , k < m.
(7.139)
7.6.6 Interior product On the other hand, given a vector field X, we

define the operation of interior product by X, denoted i(X) or iX , acting on
p-forms, as
iX : p (M ) p1 (M )
iX
The image i(X) = iX is that (p 1)-form which, for any set of fields
{X1 , X2 , . . . , Xp1 }, satisfies
(iX )(X1 , X2 , . . . , Xp1 ) = (X, X1 , X2 , . . . , Xp1 ).
(7.140)
If is a 1-form,
iX = i(X) = < , X > = (X).
(7.141)
The interior product of X by a 2-form is that 1-form satisfying iX (Y ) =

(X, Y ) for any field Y . For a form of general degree, it is enough to know
241
that, for a basis element,

i(X)[1 2 3 . . . p ] = [i(X)1 ] 2 3 . . . p +
1 [i(X)2 ] 3 . . . p + 1 2 [i(X)3 ] . . . p + . . .
p
X
()j1 1 2 . . . [i(X)j ] . . . p
=
j=1
p
X
()j1 1 2 . . . [j (X)] . . . p . (7.142)
j=1
7.6.7 If the manifold is a metric manifold, an alternative definition is

i(X) = 1 (X 0 ) = ()n(p1)+(ns)/2 (X 0 ).
(7.143)
This operation is adjoint to the exterior product defined just above:

Z
0
p1
p
0
p1
p
((X ) , ) = (X , ) = X 0 p1 p
Z
Z
p1
p1
0
p
p1
= ()
X = ()
p1 [1 X 0 p ]
Z
p1+(np+1)(p1)+(ns)/2
= ()
p1 [X 0 p ]
Z
n(p1)+(ns)/2
= ()
p1 [(X 0 ) p ]
Z
= p1 [i(X) p ] = (p1 , i(X) p ).
7.6.8 Some properties of interest are (f being a real function and g a
mapping between manifolds):
iX f = 0;
(7.144)
iX iX 0;
(7.145)
if X = f iX = iX (f );
g (ig X ) = iX (g );
iX ( ) = (iX ) + () iX .
(7.146)
(7.147)
(7.148)
7.6.9 We have already introduced a good many endomorphisms on the

graded algebra
m
[
(M ) =
k (M ).
k=0
242
e (), ada , LX , and i(X). An endomorphism E has

They are: d, d,
degree r if E : p (M ) p+r (M ). Thus, LX and have degree zero, d and
() have degrees r = +1, de and i(X) have degrees r = 1. We can be more
specific about derivations. For forms p and q , an endomorphism E in is
a derivation if its degree is even and
E( ) = E() + E().
(7.149)
It is an antiderivation if its degree is odd and

E( ) = E() + () E().
(7.150)
7.6.10 Consequently, LX is a derivation, and d and i(X) are antiderivations. In reality, these three endomorphisms are not independent. Let us
recall the expression for the Lie derivative of a p-form: if
= j1 j2 j3 ...jp xj1 dxj2 . . .dxjp ,
then
(LX ) =
+
+
... +
X(j1 j2 j3 ...jp ) dxj1 dxj2 . . .dxjp

(j1 X k )kj2 j3 ...jp dxj1 dxj2 . . .dxjp
(j2 X k )j1 kj3 ...jp dxj1 dxj2 . . . xjp
(jp X k )j1 j2 j3 ...jp1 k dxj1 dxj2 . . . dxjp .
(7.151)
Take p = 1, = j dxj :
LX = X i (i j ) dxj + (j X i ) i dxj .
On the other hand,
i(X)d = i(X)[(j i ) dxj dxi ]
= < X, j i dxj > dxi (j i ) dxj < X, dxi >
= X j (j i ) dxi X i (j i ) dxj
.
(7.152)
(7.153)
(7.154)
Therefore,
d[iX ] = d[X i i ] = (i X j ) j dxi + X i (j i )dxj .
We see that
LX = iX d + d iX = {iX , d}.
(7.155)
This is actually a general result, valid for s of any order and extremely
useful in calculations. It also illustrates another general property: the anticommutator of two antiderivations is a derivation. There is more in this
243
line. Consider derivatives generically indicated by D, D0 , etc, and antiderivatives A, A0 , etc. Using the definitions, one finds easily that the square of an
antiderivative is a derivative, AA = D. In the same token, one finds for
the anticommutator {A, A0 } = D0 , as well as the following relations for the
commutator: [D, D0 ] = D00 and [D, A] = A.
7.6.11 Consequences of eqs.[7.155] and (7.145) are the commutation properties
L X iX = iX L X
(7.156)
d(LX ) = LX (d).
(7.157)
The Lie derivative commutes both with the interior poduct and the exterior
derivative. Other interesting properties are:
Lf X = f LX + df iX
(7.158)
[LX , iY ] = i[X,Y ]
(7.159)
L[X,Y ] = [LX , LY ] .
(7.160)
7.6.12 By the way, eq.[7.151] gives us the real meaning of

and provides a new version of it:
d = dxj L j
xj0
in eq.[7.25],
(7.161)
In a general basis {ek }, with {} its dual,

d = j Lej = ( j )Lej .
(7.162)
f
This equation, which generalizes the formula df = dxi x
i valid for 0forms, is called the Koszul formula and shows how Lie derivatives generalize
partial derivatives. We may use it to check the coherence between the Lie
derivative and the exterior derivative. As the Lie derivative is a derivation,
d( ) = dxj L j ( ) = dxj (L j ) + (L j )] =
= (dxj L j ) + () (dxj L j ) = (d) + () d).
7.6.13 It is also possible to establish a new expression for the codifferential. From [7.89] and [7.25],
de = ()n(p1)+(ns)/2 d = ()n(p1)+(ns)/2 (dxj ) 1 j .

x
244
Therefore,

1
n(np)+(ns)/2
de = i(j )
= ()
i(j ) j . (7.163)
xj
x
In an euclidean space with cartesian basis,

j = ()p(np) j ,
x
x
and we have
de = i(j ) j = i(j )L j
x
In a general basis, the latter is written as center
de = i(ej )Lej .
(7.164)
(7.165)
This is not unexpected if we look at [7.161], and remember the given adjoint
relation between (dxj ) and i(j ). With [7.155], it leads directly to
de = i(j ) d i(j ).
(7.166)
7.6.14 Using [7.164], the derivation character of the Lie derivative, and
[7.144]-[7.148] we find that
de( ) = de + () de
() (L j ) [i(j )] [i(j )] (L j ). (7.167)
7.6.15 From eq.[7.142] it turns out that, if {Xj } is the basis dual to {k },
i(Xj )[i ] = ji i i(Xj ),
so that
i

i(Xj ) + i(Xj ) i

= (i ) i(Xj ) + i(Xj ) (i )
= {(i ), i(Xj )} = ji (7.168)
for any . Consequently, we find the anticommutator
{(i ), i(Xj )} = ji .
(7.169)
245
Using this and again

we find that, applied to any form of degree
Pn eq.[7.142],
j
p, the operator j=1 ( )i(Xj ) behaves like a number operator:
n
X
(j )i(Xj ) p = p p .
(7.170)
j=1
From eq.[7.140] it follows that

{i(Xi ), i(Xj )} = 0.
(7.171)
It is evident from the very definition of () that

{(i ), (j )} = 0.
(7.172)
The last four equations are reminiscent of those respected by fermion

creators aj (= (j )), annihilators ai (= i(Xi )) and the corresponding fermion
P
P
number operator j aj aj (= nj=1 (j )i(Xj )). Again, in euclidean space
with the cartesian basis, we may use [7.164] and d = (dxj ) x

j to find a
curious relation of the laplacian to the anticommutator. Of course
e
= {d, d},
but the above expressions give also
= {(i ), , i(Xj )}i j = ji i j ,
(7.173)
as they should. These formulas are the starting point of what some people call
supersymmetric quantum mechanics and have been beautifully applied14
to Morse theory and in the study of instantons.
7.6.16 The Lie algebra (M ) of fields on a smooth manifold M acts on
the space C of differentiable functions on M . We have an algebra, and
another space on which it acts as endomorphisms. The second space is a
module and we say that C is a (M )-module. With the action of the Lie
derivatives, which due to eq.[7.160] represent the Lie algebra, also the space
of p-forms is a module.
14
Witten 1982a, b.
246
Chapter 8
SYMMETRIES
8.1
LIE GROUPS
The study of a topological group is much easier when the group operation
is analytic. Even the algebraic structure becomes more accessible. This,
however, requires that the topological space have an analytic structure, more
precisely: that it be an analytic manifold. We have already said ( 5.1.3)
that every C structure has an analytic substructure. That is why most
authors prefer to define a Lie group as a C manifold, ready however to
make good use of the analytic subatlas when necessary. A topological group
has been defined in section 1.4.2 as a topological space G endowed with a
group structure and such that both the mappings m : G G G given by
(g, h) g h, and inv: G G given by g g 1 are continuous.
8.1.1 A Lie group is a C manifold G on which a structure of algebraic
group is defined, in such a way that the mappings
G G G , (g, h) g h
and
G G , g g 1
are all C .
8.1.2 It follows from this definition that the mappings
Lg : G G , h g h
and
Rg : G G , h h g
are diffeomorphisms for every g G. These mappings are called respectively
lefttranslation and righttranslation induced by the element g.
247
248
CHAPTER 8. SYMMETRIES
8.1.3 All the continuous examples previously given as topological groups

are also Lie groups. Some manifolds (for example, the sphere S 3 ) accept
more than one group structure, others (for example, the sphere S 2 ) accept
none.
8.1.4 An observation: there is a one-to-one correspondence
f : GL(n, R) Rn AL(n, R)
given by
(L, t)

L t
,
0 1
so that it is possible to introduce, on the affine group, a structure of differentiable manifold.

This makes of f a diffeomorphism. With this structure, the group operation is C .
8.1.5 Other examples may be obtained as direct products : let G1 and G2 be two Lie
groups. The product G = G1 G2 can be endowed with the C structure of the cartesian
product, which makes of G the direct product group G1 G2 if the following operation is
defined:
GGG
((g1 , g2 ), (g10 , g20 )) (g1 g10 , g2 g20 ) .
It is important to notice that the affine group is not to be considered as the Lie group
GL(n, R) Rn , because the product for AL(n, R) is
GGG
((L, t), (L0 , t0 )) (LL0 , Lt0 + t) .
This is an example of semi-direct product, denoted usually by GL(n, R) Rn .
8.1.6 As S 1 is a Lie group, the product of n copies of S 1 is a Lie group, the toral
group or ntorus T n = S 1 S 1 . . . S 1 .
8.1.7 A subgroup of a topological group is also a topological group. An

analogous property holds for Lie groups, under the conditions given by the
following theorem:
let G be a Lie group, and H an algebraic subgroup which is also
an imbedded submanifold. Then, with its smooth structure of submanifold, H is a Lie group.
8.1.8 It is possible to show that the following subgroups of the linear
groups ( 1.4.16) are Lie groups:
8.1. LIE GROUPS
249
1. The orthogonal group O(n), of dimension n(n 1)/2,

O(n) = {X GL(n, R) such that X X T = I} ,
where X T is the transposed of X. The special orthogonal group,
SO(n) = {X O(n) such that det X = +1} ,
is the group of rotations in En . For the special value n = 1, it is trivial:
SO(1) = I. Consecutive quotients of such groups are spheres:
SO(n)/SO(n 1) = S n1 .
Thus, SO(2) = S 1 . The groups SO(n) are homotopically non-trivial:
for n = 2, 1 [SO(2)] = Z; for the other cases, they are doublyconnected:
1 [SO(n)] = Z2 , for n 3 .
2. The unitary group U (n), of dimension n2 ,
U (n) = {X GL(n, C) such that X X = I},
where X is the adjoint (complexconjugate transposed) of X. The
groups U (N ) are multiply connected, 1 [U (n)] = Z for n 1. . The
special cases SU (N ), those for which furthermore det X = +1, are
of enormous importance in the classification of elementary particles.
With the exception of SU (1) = SO(2), they are simply-connected:
1 [SU (n)] = 1 for n > 1. The group SU (2) is isomorphic to the sphere
S 3 (see 3.2.30) and describes the spin (see 8.1.13 below). It is the
universal covering of the rotation group in our ambient E3 , SU (2)/Z2
= SO(3). Such relationships are a bit more difficult to predict for
higher dimensional cases, as shown by the example SU (4)/Z2 = SO(6).
Consecutive quotients are spheres:
U (n)/U (n 1) = SU (n)/SU (n 1) = S 2n1 .
3. The symplectic group Sp(n),
Sp(n) = {X GL(n, Q) such that X X = I} .
It is the group of linear canonical transformations for a system with n
degrees of freedom. It is simply-connected for each value of n. Again,
spheres come out from consecutive quotients: Sp(n)/Sp(n1) = S 4n1
and in particular Sp(1) = SU (2) = S 3 . This is, by the way, a good
example of two group structures defined on the same manifold.
250
4. The special linear group, of dimension n2 1:

SL(n, R) = X GL(n, R) such that det X = 1.
5. The special complex linear group, of dimension 4n2 2:
SL(n, C) = X GL(n, C) such that det X = 1.
8.1.9 We know that the orthogonal group O(n) preserves the euclidean
scalar product: if x, y En , then
hT x, T yi = hx, yi
for all T O(n). This can be generalized to obtain a group preserving the
non-definite (pseudo-euclidean) scalar product
hx, yi =
p
X
i i
xy
i=1
n
X
xj y j .
j=p+1
This group is the set of matrices

1
O(p, q) = {X GL(n, R) such that Ip,q X T I1
},
p,q = X
with p + q = n and Ip,q the diagonal matrix with the first p elements equal
to (+1), and the q remaining ones equal to ( 1). A case of particular
importance is
SO(p, q) = {X O(p, q) such that det X = 1}.
These are non-compact groups, the noblest example of which is the Lorentz
group SO(3, 1). More about these groups will be said in 9.3.4.
8.1.10 One more definition: let be an algebraic homomorphism between the Lie
groups G1 and G2 . Then, will be a homomorphism of Lie groups iff it is C .
8.1.11 Given the toral group T n , then

: En T n
(t1 , t2 , . . . , tn ) (e2t1 , e2t2 , . . . , e2tn )
is a homomorphism of Lie groups.
8.1. LIE GROUPS
251
8.1.12 A topological group is locally compact if around each point there

is an open set whose closure is compact ( 1.2.13). This is a very important
notion, and that by a technical reason. We are used to applying Fourier
analysis without giving too much thought to its fundamentals. Putting it
in simple words, the fact is that Fourier analysis is always defined on a
group (or on a quotient of groups, or still on the ring of a group). Standard
spaces are the circle S 1 (the 1-torus group T 1 ) for periodical functions, the
real line additive group (R, +) and their products. The technical point is
the following: the summations and integrals involved in Fourier series and
integrals presuppose a measure, and this measure must be group-invariant
(the same over all the group-space). And only on locally compact groups
is the existence of such an invariant measure assured. These measures are
called Haar measures (see 8.2.20). Classical Fourier analysis is restricted
to abelian locally compact groups like the above mentioned ones, but local
compactness allows in principle extension of harmonic analysis to non-abelian
groups, such as the special unitary group SU (2) S 3 (Mathematical Topic
6.3).
8.1.13 As topological spaces, we have said ( 3.2.30 and 8.1.8 above)
that the rotation group SO(3), related to the angular momentum, is doublyconnected, with the special unitary group SU (2) as covering space. This is
very important for Physics, as only SU (2) possesses the half-integer representations necessary to accommodate the fermions. The group SO(3) is a
subgroup of the Lorentz group SO(3, 1), which is also unable to take halfinteger spins into account. Every elementary particle must belong to some
representation of the Lorentz group in order to have a well-defined relativistic behaviour. The same must hold for the relativistic fields which represent
them in Quantum Theory. To accommodate both fermions and bosons, it
is necessary to go to the covering group of SO(3, 1), which is the complex
special linear group SL(2, C) (Physical Topic 6).
8.1.14 Grassmann manifolds The Grassmann spaces Gnd of 1.4.20,
whose points are ddimensional planes in euclidean ndimensional spaces,
can be topologized and endowed with a smooth structure, becoming manifolds. They can be obtained (or, if we wish, defined) as double quotients,
Gnd = Gd (En ) = O(n)/(O(d) O(n d)).
If we consider oriented d-planes, we obtain a space G#
nd , which is a double
covering of Gnd . In the complex case, the manifolds are
n
GC
nd = Gd (C ) = U (n)/(U (d) U (n d)).
As quotients of compact spaces, they are themselves compact.
252
8.1.15 Stiefel manifolds

Denoted Snd or Sd (En ), these are spaces
whose members are ddimensional orthogonal frames in En . They are found
to be (or can be alternatively defined as) Snd = O(n)/O(n d) and their
dimensions are: dim Snd = d. Stiefel manifolds have curious homotopic
properties: their lower homotopy groups vanish. More precisely, r (Snd )
= 0 for (n d 1) r 0. As with Grassmann manifolds, we may
C
consider complex Stiefel manifolds: Snd
= Sd (Cn ) is the space of unitary d
n
C
dimensional frames in C . For them, r (Snd
) = 0 for (2n 2d 1) r 0.
Because of these peculiar homotopic properties, Stiefel manifolds are of basic
importance for the general classification of fiber bundles (section 9.7).
8.1.16 After what we have seen in 6.4.3 and 6.5.8, the very matrix
elements can be used as coordinates on a matrix group.
8.2
TRANSFORMATIONS ON MANIFOLDS
We proceed now to a short analysis of the action of groups on differentiable

manifolds. In particular, continuous transformations on manifolds, in general, constitute continuous groups which are themselves manifolds, topological or differentiable. The literature on the subject is very extensive we
shall only occupy ourselves of some topics, mainly those essentially necessary
to the discussion of bundle spaces. We have seen in section 6.4.2, when the
concept of Lie derivative was introduced, the meaning of the action of R on
a manifold M , R being considered as the additive (Lie) group of the real
numbers. We shall now generalize that idea in a way which applies to both
topological and Lie groups.
8.2.1 Action of a group on a set Let G be a group, and M a set.
The group G is said to act on M when there exists a mapping
:M GM
(p, g) (p, g)
satisfying:
(i) (p, e) = p, p M , where e is the identity element of G;
(ii) ((p, g), h) = (p, gh), p M and g, h G.
The mapping is called the right action of G on M , and is generally
denoted by R: we indicate the right action of g by Rg x = x0 M , or more
simply by xg = x0 . Left action is introduced in an analogous way. Once
such an action is defined on a set M , M is said to be a Gspace and G
a transformation group on M . A subset M 0 of M is invariant if xg M 0
8.2. TRANSFORMATIONS ON MANIFOLDS
253
whenever x M 0 . Every subset M 00 of M is contained in some invariant

subset M 0 , which is said to be the generator of M 00 .
8.2.2 When G is a topological group and M a topological space, the action
R is required to be continuous. When G is a Lie group and M a C manifold,
R is required to be C .
8.2.3 As already said, we shall frequently use the abbreviated notation pg
for R(p, g) = (p, g), so that condition (ii) of 8.2.1 above becomes (pg)h
= p(gh). In the C case, the mapping
Rg : M M
p Rg (p) = pg
is a diffeomorphism and allows one to rewrite (ii) as Rh Rg p = Rgh p.
8.2.4 Effective action An action is said to be effective when the identity
element e is the only element of G preserving all the points of M , i.e., when
Rg p = p, p, implies g = e. This means that at least some change comes
out through the action of each group element.
8.2.5 Transitive action The action is said to be transitive when, given
any two points p and q of M , there exists a g in G such that pg = q. We
can go from a point of M to any other point by some transformation of the
group. The space E3 is transitive under the group T3 of translations. If g is
unique for every pair (p, q), the action is simply transitive.
8.2.6 Given the point p of a manifold M , the set of group members Hp =
{h G such that ph = p} is a subgroup of G, called the isotropy group (or
stability group) of the point p.
8.2.7 Homogeneous spaces If G acts transitively on M , M is homogeneous under G. A homogeneous space has no invariant subspace, except
itself and the empty set. Simple groups are homogeneous by the action defined by the group multiplication. E3 is homogeneous under the translation
group T3 but not under the rotation group SO(3). On the other hand, a
sphere S 2 is homogeneous under SO(3). Rotations in E3 leave fixed a certain point (usually taken as the origin). If p is a point of a homogeneous
manifold M , another point q of M will be given by q = pg for some g, or q =
phg for any h Hp . Thus, q(g 1 hg) = phg = q, so that g 1 hg Hq . Given
any member of Hp , g will give a member of Hq and g 1 will do the same
in the inverse way. As a consequence, the isotropy groups of all points on a
254
homogeneous manifold are isomorphic. The Lie algebra has then a canonical
decomposition of the form G0 = H 0 + T , for some T .
A homogeneous space G/H has always a Riemann metric which is invariant under
the action of G, and is said to be a homogeneous Riemannian space. We might in larger
generality define a homogeneous Riemannian space as a Riemannian space M whose group
of motions acts transitively on M .
The fact that a space may be obtained as a quotient of two Lie groups as G/H has
very deep consequences, particularly in the case of homogeneous symmetric spaces. This
happens when G has an involutive automorphism, : G G, 2 = 1. The Lie algebra
canonical decomposition G0 = H 0 + T , in the presence of such an involution, has the form
[H 0 , H 0 ] H 0 ; [H 0 , T ] T ; [T, T ] H 0 .
The bundle G = (G/H, H) admits an invariant connection, called its canonical connection,
which is determined by the space T . It has vanishing torsion and a very special form for
the curvature. This special canonical connection is a restriction, to the quotient space, of
the MaurerCartan form of the group G.
8.2.8 The action of G is said to be free when no element but the identity
e of G preserves any point of M , that is, Rg p = p for some p M implies g
= e. Thus a free action admits no fixed point.
8.2.9 The set of points of M which is obtained from a given point p by
the action of G is the orbit of p:
orbit(p) = {q = pg, g G} .
Thus, one obtains the orbit of p by acting on p with all the elements of G.
An orbit may be seen as the invariant subset generated by a single point:
orbit(p) = p G .
Every orbit is a transitive Gspace by itself. A transitive space is clearly an
orbit by one of its points.
8.2.10 Everything which was said above about right action can be repeated for left action, with the notation L, and Lg (p) = gp. No confusion
should arise by the use of the same notation for the (left or right) action on
a manifold M and the (left or right) translation on the group G itself. For a
physical example illustrating the difference between right and leftactions,
see Physical Topic 2.3.9.
8.2.11 As R is a Lie group, a oneparameter group ( 6.4.17) on a manifold
M is an example of action, as it satisfies conditions (i) and (ii) of 8.2.1,
and is C .
255
8.2.12 Let G1 and G2 be two Lie groups and : G1 G2 a group homomorphism.

It is possible to show that
L : G1 G2 G 2
(g1 , g2 ) (g1 )g2
is a leftaction.
8.2.13 The best known case is the usual action of matrices on column
vectors. We can rephrase it in a pedantic and precise way and call it the
natural action (on the left) of GL(n, R) on Rn :
L : GL(n, R) Rn Rn
(A, x) Ax.
In an analogous way, we have the action on the right
R : Rn GL(n, R)) Rn
(x, A) xT A.
8.2.14 The left-action of the affine group ( 8.1.4, 8.1.5) on Rn is given
by

x1
x1
2
2

x
x
L
t
L
t

... ,
, ...
0 1
n
n
0 1
x
x
1
1
where L GL(n, R) and t Rn . If L O(n), then we have action of a
subgroup of A(n, R) called the group of rigid motions, or euclidean group on
Rn .
8.2.15 The Poincar
e group P O(4, R) (or inhomogeneous Lorentz group,
see Physical Topic 6) is a subgroup of the affine group for n = 4: A(n, R)
P O(4, R), where

L t
P O(4, R) =
such that L SO(3, 1) and t Minkowski space .
0 1
8.2.16 Linear and Affine basis
Let B(Rn ) be the set of linear basis (see

6.5.6) of R (here taken as synonym of E with the structure of vector space),
n
B(Rn ) = {f := {fi }, i = 1, 2, . . . , n such that the fi are linearly independent} .
256
Let us define the action

B(Rn ) GL(n, R) B(Rn )
X
X
X
(f, L) f L :=
fj Lj1 ,
fj Lj2 , . . . ,
fj Ljn .
Given two basis f and f in B(Rn ), there exists a unique L GL(n, R) such that
f = f L .
Thus, the action is simply transitive. This means that there is a one-to-one correspondence
between GL(n, R) and B(Rn ) given by
L e L ,
where e = {e1 , e2 , . . . , en } is the canonical basis for Rn :
e1 = (1, 0, 0, . . . , 0),
e2 = (0, 1, 0, . . . , 0), etc.
Once this correspondence is established, it is possible to endow the set B(Rn ) with a
topology and a C structure so as to make it diffeomorphic to GL(n, R). B(Rn ) is called
the basis space of Rn .
8.2.17 What was done above can be repeated with a general vector space V instead
of Rn . Let us examine now the set of the affine basis on a vector space V :
A(V ) = {[f1 , f2 , . . . , fn ; x] =: [f, x] such that f B(V )
and x V } .
The action of the affine group on the affine basis is given by

A(V ) A(n, R) A(V )

[f, x],

L t
L t
[f L, f t + x] := [f, x]
0 1
0 1
Pn
where f t := i=1 fi ti . Given any two basis f and f A(V ), there is a unique element of
A(n, R) such that

L t
f = f
.
0 1
Thus, there is a one-to-one correspondence between A(n, R) and A(V ), given by

L t
L t
a
,
0 1
0 1
where a = {e1 , e2 , . . . , en , 0} is the canonical basis for A(V ). With this correspondence,
we can endow the set A(V ) with a topology and a C structure making it diffeomorphic
to A(n, R). As a manifold, A(V ) is the affine basis space on V . If, instead of a, another
basis a0 were used, the same structure would result, up to a diffeomorphism.
257
8.2.18 We have learned (section 1.4.1) how to obtain new manifolds from
a given one, as quotient spaces by an equivalence relation. Suppose the
relation between two points p and q M defined by
p q there exists some g G such that q = Rg p.
It is an equivalence. The set [p] = {q M such that q p}, which is the
orbit(p), is the equivalence class with representative p. The set of all these
classes is denoted by M/G: it is said to be the quotient space of M by G.
The canonical projection is defined by : M M/G, p [p]. A quotient
topology can then be introduced on M/G, as well as a C structure.
An important particular case appears when M is itself a Lie group and
G a subgroup of M . Then the quotient space M/G is an analytic manifold. The action G M/G M/G is transitive and M/G is consequently
a homogeneous manifold under the action of G. Manifolds of this type
(group/subgroup) have many interesting special characteristics, as also the
action is an analytic mapping, as well as the projection M M/G.
8.2.19 Suppose the group G acts simultaneously on two manifolds M and N , with
actions denoted mg and ng (see Figure 8.1). A mapping : M N is equivariant (or
an intertwining map) when mg = ng . The diagram at the right of Figure 8.1 is
commutative.
Figure 8.1:
8.2.20 Invariant measure Consider a set M on which an action of the

group G is defined. Suppose there is a Borel measure on M (Mathematical
Topic 3). The measure is said to be an invariant measure when, for each
measurable subset A of M and for all g G, (gA) = (A). Here gA means
258
the set resulting from whatever action of G on the set A. In particular, the
measure is leftinvariant if (LgA) = (A) and rightinvariant if (Rg A)
= (A). The Lebesgue measure on an euclidean space is invariant under
translations and rotations; the Lebesgue measure on the sphere is invariant
under the action of the rotation group on the sphere. Such Haar measures
are sure to exist only on locally compact groups ( 1.2.13, 8.1.12). On
such groups, they are unique up to positive factors. For the groups T n and
(R, +), the Haar measures are simply the (normalized) Lebesgue measures.
Haar measures provide a characterization of compactness: the Haar measure
on G is finite iff G is compact. This property stays behind the well known
fact that Fourier expansions on compact spaces (groups!) are series while
Fourier expansions on non-compact spaces are integrals (see Mathematical
Topic 6.3).
8.2.21 Invariant integration Given
an invariant Rmeasure on M , the corresponding
R
integral is invariant in the sense that M f (gx)d(x) = M f (x)d(x). An integral may be
left-invariant, right-invariant, or both.
8.2.22 Function spaces, revisited Let us go back to 8.2.7 and consider the space
C(M ) of complex functions on a homogeneous space M . The space C(M ) may be made
into a Gspace by introducing the action (Rg f )(m) = f (mg 1 ). ). Now, C(M ) is a vector
space, which is fair, but it is not necessarily homogeneous. It has, in general, invariant
subsets which are themselves linear spaces. Take for example a simple group G and M
= G, the action being given by the group multiplication, say Rg : x xg. Take C h (G),
the set of all homomorphisms of G into the non-vanishing complex numbers. Then if
h C h (G), we have that h C(G) and the set of all constant multiples of h constitutes
an invariant (one-dimensional) subspace. Such subspaces are independent for distinct hs.
. When G is finite and commutative, each member of C(G) is a unique sum of members,
one from each invariant subspace. When G is infinite, such sums become infinite and
some extra requirements are necessary. In general, one restricts C(G) to a subspace of
measurable (square integrable) functions with some (say, Lebesgue) Haar measure. Take
for instance G as the circle S 1 . The only measurable homomorphisms are hn (x) = einx ,
with n = 0, 1, 2, . . .. Then, any square integrable function f on G may be written in a
unique way as
X
f (x) =
fn einx ,
n
which is the Fourier theorem. To each irreducible representation hn (x) corresponds a

harmonic. As long as G is compact, the sums as above are discrete. Things are more
involved when G is not compact. We have said that only locally compact groups have
Haar measures. In the non-compact but still abelian cases, the number n above becomes
continuous and the sums are converted into integrals. The best known case is G = R =
{realnumberswithaddition}, when
Z
f (x) =
f (s)eisx .
R
Let us say a few words on the compact non-commutative case. Take G as the rotation
group in E3 , and M = S 2 , the spherical surface in E3 . The space of all squareintegrable
8.3. LIE ALGEBRA OF A LIE GROUP
259
functions on M divides itself into invariant subspaces Mj , one for each odd number (2j +1)
and such that dim Mj = (2j+1). They are formed by 3variable homogeneous polynomials
of degree j which are harmonic, that is, satisfy the Laplace equation on M . Each Mj
provides an irreducible representation, the fj Mj being
P the surface harmonics, and
any function f C(S 2 ) is uniquely expanded as f =
n fj , with fj Mj (see also
Mathematical Topic 6.3).
8.3
LIE ALGEBRA OF A LIE GROUP
We have said in section 6.4.5 that the vector fields on a smooth manifold
constitute a Lie algebra. Lie groups are differentiable manifolds of a very
special type. We describe now (very superficially) the general properties of
fields and forms on Lie groups.
8.3.1 Consider the left action of a Lie group G on itself:
Lg : G G
h Lg (h) = gh .
(8.1)
A first thing which is peculiar to the present case is that this action is a
diffeomorphism. It induces of course the differential mapping between the
respective tangent spaces,
Lg = dLg : Th G Tgh G .
(8.2)
An arbitrary field X on G is a differentiable attribution of a vector Xg at

each point g of G. Under the action of Lg , its value Xh at the point h will
0
= Lg (Xh ) at the point gh (Figure 8.2,
be taken into some other field Xgh
left).
Suppose now that the field X is taken into itself by the left action of G:
Lg (Xh ) = Xgh .
(8.3)
In this case, X is said to be a leftinvariant field of G and one writes

Lg X = X .
(8.4)
This means that, for any function f R(G),

(Lg Xh ) = Xh (f Lg ) = Xgh (f ) .
(8.5)
260
Figure 8.2:
8.3.2 Notice that, in particular,

Lg (Xe ) = Xe
(8.6)
when X is left-invariant (Figure 8.2, right). Consequently, left-invariant fields

are completely determined by their value at the group identity e. But not only
the fields: their algebras are also completely determined, as diffeomorphisms
preserve commutators ( 6.4.6). Thus,
Lg [Xe , Ye ] = [Xg , Yg ]
(8.7)
for any leftinvariant fields X, Y . This is to say that the Lie algebra of
left-invariant fields at any point on a Lie group G is determined by the Lie
algebra of such fields at the identity point of G.
8.3.3 This algebra of invariant fields, a subalgebra of the general Lie algebra of all the fields on G, is the Lie algebra of the Lie group G. It is usually
denoted by L(G), or simply G0 . The vector space of G0 is T eG. A basis for
G0 will be given by d (= dim G) linearly independent leftinvariant fields X ,
which will satisfy
[X , X ] = C X .
(8.8)
8.3.4 According to (8.7), this relation must hold (with the same C s) at
any point of G, so that the structure coefficients are now point-independent.
261
They are, for this reason, called the structure constants of G. The Lie algebra
of G is thus a small (as compared with the infinite algebra of all fields) Lie
algebra of d fields fixed by their values at one point of G.
8.3.5 Right-invariant fields can be defined in an analogous way. They
constitute a Lie algebra isomorphic to that of the left-invariant fields.
8.3.6 A p-form w on G is leftinvariant if
Lg w = w .
(8.9)
Let us see how things work for 1-forms: given a form wgh at gh, its pull-back
is defined by
hLg1 w, Xih = hw, Lg1 Xigh .
(8.10)
If w is invariant,
hw, Xih = hw, Lg1 Xigh .
(8.11)
hw, Xih = hw, Xigh .
(8.12)
If also X is invariant,
Therefore, an invariant form, when applied to an invariant field, gives a
constant.
8.3.7 Invariant Pfaffian forms on Lie groups are commonly called Maurer
Cartan forms. They constitute a basis {w } for L (G), dual to that of
invariant fields satisfying eq.(8.8). As a consequence, eq.[7.46] tells us that
they obey
(8.13)
dw = 12 C w w .
This is the MaurerCartan equation, which can be put in a basis-independent
form: define the vector-valued canonical form
w = X w .
(8.14)
When applied on a field Z, the canonical form simply gives it back:

w(Z) = X w (Z) = X Z = Z .
(8.15)
Then, a direct calculation puts the MaurerCartan equation in the form

dw + w w = 0 .
(8.16)
8.3.8 For a matrix group with elements g, it is easy to check that

w = g 1 dg
(8.17)
are matrices in the Lie algebra satisfying eq.(8.16) (not forgetting that dg 1
= - g 1 dgg 1 ).
262
8.3.9 Given any n n matrix A, its exponential is defined by the (convergent) series
P
1 j
eA = I + A + 21 A2 + . . . =
(8.18)
j=0 j! A .
The set of matrices of type exp(tA), with t R, constitutes an abelian group:
etA esA = e(t+s)A ; etA etA = I ; etc.
The mapping a: R GL(n, R), a(t) = exp(tA), is a curve on GL(n, R)
whose tangent at t = 0 is

d tA
e = AetA t=0 = A .
(8.19)
dt
t=0
So, A TI GL(n, R), or A G0 L(n, R). The set of matrices exp(tA) is the
group generated by A. As A is arbitrary, we have shown that any n n
matrix belongs to G0 L(n, R). Thus, G0 L(n, R) is formed by all the n n
matrices, while GL(n, R) is formed by those which are invertible.
8.3.10 A very important result is Ados theorem:
every Lie algebra of a Lie group is a subalgebra of G0 L(n, R), for some
value of n.
For Lie groups, an analogous statement holds, but only locally: every Lie
group is locally isomorphic to a subgroup of some GL(n, R).
Concerning matrix notation and the use of a basis: a general matrix in
GL(n, R) will be written as g = exp(X p ), where the X s constitute a
basis in G0 L(n, R). The components p are the group parameters of g.
The vector-valued form
w = X w = g 1 dg = g 1 (X dp )g = g 1 X g dp
will be a matrix of forms, with entries
wi k = (X )i k w = [g 1 X g]i k dp .
8.3.11 Exponential mapping We have seen in 6.4.17 how the group
R acts on a manifold. Let us apply what was said there to the case in which
the manifold is itself a Lie group:
:RGG
: (t, h) (t, h) .
(8.20)
263
Take the orbits through the identity,

(0, e) = e;
(t, e) = e (t) .
The theory of ordinary differential equations tells us that, in this case, there
is an open U G around e in which the solution of eq.[6.41],
d
e (t) = Xe (t)
dt
(8.21)
is unique, for any X. Then, a(t) = e (t) is the integral curve of X through
the identity and Xe G0 . Now, when the manifold is a Lie group, this is
a global result: the field X is complete, that is, a(t) is defined for every
t R. Still more, the set {aX (t)}, for all t R is a oneparameter subgroup of G generated by X. We can then introduce the exponential mapping,
generalizing the case of GL(n, R), as
exp : G0 G
exp(X) = aX (0) ,
(8.22)
so that the subgroup is given by

aX (t) = exp(tX) .
(8.23)
8.3.12 Normal coordinates This mapping is globally C and, in a

neighbourhood around e, a diffeomorphism. In such a neighbourhood, it
allows the introduction of a special LSC. Take a basis {J } of G0 and X =
X J . The algebra G0 can be identified with Rd by X (X 1 , X 2 , . . . , X d ).
As aX (1) = exp(X J , we can ascribe these coordinates to aX (1) itself. By
eq.[8.23], aX (t) would then have coordinates {tX }:
[a(t)] = tX ; [a(s)] = sX .
But a(s)a(t) = a(s + t), so that
[a(s)a(t)] = tX + sX = [a(t)] + [a(s)] .
Such local coordinates, for which the coordinates of a product are the sum
of the factor coordinates, are called the canonical, or normal coordinates.
8.3.13 The Heisenberg Algebra and Group
The usual Poisson bracket relation of classical mechanics for n degrees of freedom q k ,
{pi , pj } = {q k , q l } = 0 ; {pi , q j } = ij ,
264
and the commutation relations for their quantum correspondent operators,

[
pi , pj ] = [
q k , ql ] = 0 ; [
pi , qj ] = i~ij I ,
are formally the same. They constitute a Lie algebra going under the name of Heisenberg
algebra. The corresponding Lie group is the Heisenberg group Hn . The algebra may be
characterized by parameters (p, q, s) in R2n+1 = Rn Rn R1 . Consider the (n+2)(n+2)
matrix
0 p1 p2 . . . pn s
0 0 0 . . . 0 q 1
0 0 0 . . . 0 q 2
h(p, q, s) = . .
..
..
..
.. .
.. ..
.
.
.
.
0 0 0 . . . 0 q n
0 0 0 ... 0
0
It is immediate that the products of two matrices of this kind are given by
h(p, q, s) h(p0 , q 0 , s0 ) = h(0, 0, pq 0 ) ;
[h(p, q, s)]2 = h(0, 0, pq) and [h(p, q, s)]m = 0 for m > 2 .
The commutator
[h(p, q, s), h(p0 , q 0 , s0 )] = h(0, 0, pq 0 p0 q)
will define the Lie algebra. The group Hn is arrived at by the exponential map
h(p, q, s) H(p, q, s) = exp[h(p, q, s)] ,
which is
1 p1
0 1
0 0
H(p, q, s) = . .
.. ..
0 0
0 0
p2
0
1
..
.
...
...
...
..
.
pn
0
0
..
.
0
0
...
...
1
0
s + 21 pq
q1
q2
.
..
q
1
The group law may be expressed as

H(p, q, s)H(p0 , q 0 , s0 ) = H(p + p0 , q + q 0 , s + s0 + 21 (pq 0 p0 q)) .
The centre, which here coincides with the commutator subgroup (Mathematical Topic
1.5), is given by C = {H(0, 0, s)}. The Lebesgue measure on R2n+1 is a biinvariant Haar
measure on Hn . Notice that other matrix realizations are possible, but the above one has
the advantage that the inverse to H(p, q, s) is simply H(p, q, s)1 = H(p, q, s).
8.3.14 A first hint of what happens in a noncommutative geometry is the following.

Taking the Lie group algebraic structure into account, a field defines both a left derivative,

d
XL f (g) =
f (etX g)
dt
t=0
and a right derivative

d
tX
XR f (g) =
f (ge )
.
dt
t=0
There is no reason for them to coincide when the group is non-abelian.
8.4. THE ADJOINT REPRESENTATION
8.4
265
THE ADJOINT REPRESENTATION
A representation of a group G is a homomorphism of G into some other

group H, and a representation of a Lie algebra G is a homomorphism of
G into some other algebra H. The simplest cases are the linear group representations, those for which H is the group AutV of the linear invertible
transformations of some vector space V, which is the carrier space, or
representation space of the representation. More details can be found in
Math.6. We shall here consider some representations in terms of fields defined on the group manifold itself. They are essential to the understanding
of Lie groups and of their action on other spaces.
The adjoint representation is a representation of a Lie group on the vector
space of its own Lie algebra. Its differential is a representation of this Lie
algebra on itself or, more precisely, on its derived algebra (Math. Topic 1.23).
8.4.1 Isomorphisms of a Lie group G into itself and of a Lie algebra G
into itself are automorphisms. The set Aut(G) of all such automorphisms
is itself a Lie group. For every j Aut(G), the differential dj = j is an
automorphism of G, such that j(exp X) = exp(j X). The diagram
j = dj
G0
> G0
exp
exp
>
is commutative.
8.4.2 When working with matrices, we are used to seeing a matrix h be
transformed by another matrix g as ghg 1 . This is a special case of a certain very special representation which is rooted in the very nature of a Lie
group. The automorphisms j above belong to the group Aut(G) of the
linear transformations of G (seen as a vector space). The differential operation j j = dj takes Aut(G) into Aut(G). It is a homomorphism, since
d(j k) = dj dk. Consequently, it is a representation of Aut(G) on G. An
important subgroup of Aut(G) is formed by the inner automorphisms of G,
which are combinations of left-and right-translations ( 8.1.2) induced by an
element g and its inverse g 1 :
jg = Lg Rg1 = Rg1 Lg
266
jg (h) = ghg 1 .
(8.24)
Each jg is in reality a diffeomorphism, and jg (hk) = jg (h) jg (k). Thus, the

mapping g jg is a group homomorphism. The mapping
djg = jg = Lg R(g1 ) = R(g1 ) Lg
belongs to Aut(G).
8.4.3 Now we arrive at our objective: the mapping
Ad : G AutG0
Ad(g) = Adg = djg
(8.25)
is the adjoint representation of G. Given a field X G, the effect of Ad(g)

on X is described by
Adg X = (R(g1 ) Lg )X.
(8.26)
Being X left-invariant, then

Adg X = R(g1 ) X.
(8.27)
8.4.4 Using [8.23], expression [8.26] may be written as

etAdgX = getX g 1 .
(8.28)
Thus: take the curve exp(tX); transform it by jg ; then, Adg X is the tangent
to the transformed curve at the identity (Figure 8.3). This representation is
of fundamental importance in modern field theory, as both gauge potentials
and fields belong to it (see sections 7.3 and 7.4; see also Physical Topic 7).
8.4.5 The mapping
ad := d(Ad)
ad : G0 (AutG)0
X adX
(8.29)
is the adjoint representation of the Lie algebra G. To each field X in G

corresponds, by this representation, a transformation on the fields belonging
to G, of which X will be the generator. We know that a field generates
transformations on its fellow fields through the Lie derivative, so that
adX Y = LX Y = [X, Y ].
(8.30)
267
Figure 8.3:
In a basis {Xi },
adXi Xj = [Xi , Xj ] = C k ij Xk .
(8.31)
Thus, the adjoint representation is realized by the matrices Ci whose elements

(Ci )k j are the structure constants,
(Ci )k j = [ad(Xi )]k j = C k ij .
(8.32)
8.4.6 Notice that if g has its actions given by the matrices U (g), and X
is also a matrix (case of GL(n, R)), eq.[8.26] gives simply the usual rule for
matrix transformation U XU 1 . From a purely algebraic point of view, the
adjoint representation is defined by [8.32]: it is that representation by matrices whose entries are the structure constants. It is sometimes called regular
representation,1 but we shall use this name for another kind of representation
(see Mathematical Topic 6).
We may consider also the representations given by the action on the
forms, through the pull-back L(g1 ) and R(g1 ) . Such representations on
the covectors are called coadjoint representations. In the matrix notation of
8.3.10, the adjoint representation will be given by a matrix A, such that
X0 = g 1 (X )g = A X
1
See, for instance, Gilmore 1974.
268
For g = eX p , the vector-valued form

w = X w = g 1 (X )gdp
will be A dp , so that w = A dp .
8.4.7 Let us go back to the action of groups on manifolds, section 8.2.
Consider the right-action:
Rg : M M
Rg (p) = pg with Re (p) = p.
Figure 8.4:
Let us change a bit the point of view (see Figure 8.4): take p fixed and
change g G. The action Rg becomes a mapping of G into M , which we
shall denote pe:
pe : G M
pe(g) = Rg (p) = pg.
The set of points pe(g) is, as said in 8.2.9, the orbit of p by G.
8.4.8 Then, the differential mapping
(8.33)
269
pe : Te G Tp M
of Tp M ,
will take a field X of G into some field X
= pe (X).
X
(8.34)
This mapping is an algebra homomorphism of G into the Lie algebra of

fields on some U 3 p. The following results are very important:
is then a fundamental
1. if the action is effective, pe is one-to-one; X
e0 of the
field on M , corresponding to X; taking all the X G, the set G
corresponding fundamental fields is a representation of G.
e0 .
2. if G acts also freely, pe is an isomorphism: G0 G
Summing up, the action of a group G on a manifold M around one of its
points is thus realized in the following way:
on M ,
(i) for each group generator X there will be a deputy field X
the nomination being made through the mapping pe ;
(ii) each fundamental field will be the infinitesimal operator of transformations ( 6.4.22) of a one-parameter group;
(iii) the set of fundamental fields will engender the group transformations
on M .
(iv) The representation in the general case will be non-linear (see Physical
Topic 10).
8.4.9 Let us try to put it all in simple words: given a group of transformations on a manifold under some conditions, its generators are represented,
in a neighbourhood of each point of the manifold, by fields on the manifold.
Each generator appoints a field as its representative. This is what happens,
for instance, when we represent a rotation around the axis 0z in E3 by the
infinitesimal operator xy yx , which is a field on E3 . This operator acts on
functions defined on E3 which carry a representation of the rotation group.
8.4.10 We may ask now: under a group transformation, what happens to
the fundamental fields themselves? On M ,
(Rg X)p = pe Rg X = pe [Adg1 X] = Adg1 X,
where use was made of eq.[8.27]. We have been using the same notation Rg
for the actions on G and M , so that we shall drop the bars when not strictly
necessary:
Rg X = Adg1 X.
(8.35)
If we examine the left-action, we find
Lg X = Adg X.
(8.36)
270
Notice the change g g 1 between the two cases. The process is rather
knotty. When we want to know how a fundamental field X changes in some
point p of M , we begin by going back to the field X in G which X represents;
transform X by the adjoint representation and then bring the result back by
the mapping pe (Figure 8.5).
Figure 8.5:
This process is pictorially represented in the commutative diagram
p
> Tp M
G0
Adg1
Rg
G0
> Tp M
8.4.11 All this discussion is strictly local. It shows where the importance
of the adjoint representation comes from and will be instrumental in the
study of fiber bundles.
8.4.12 The Killing form is a bilinear form on G defined by
ij = (Xi , Xj ) = tr [ad (Xi ) ad(Xj )],
(8.37)
271
or
ij = C k im C m jk .
(8.38)
A theorem by Cartan says that det(ij ) 6= 0 is a necessary and sufficient

condition for G to be a semisimple algebra (and G be a semisimple group,
that is, without abelian invariant subgroups). Examples of non-semisimple
groups are the linear groups, the affine group and the Poincare group. On the
other hand, the orthogonal (or pseudo-orthogonal, like the Lorentz group)
and the unitary groups are semisimple. In the semisimple case, as is nondegenerate, it can be used as an invariant metric on G (invariant under the
action of G itself), the Cartan metric. It is used in gauge theories, being the
K metric of the basic lagrangian given by eq.[7.138]. Of course, it only makes
sense when the gauge group (the structure group) is semisimple. Usual gauge
theories use orthogonal or unitary groups.
Suppose G is a compact and semisimple group. A form on G is said to
be invariant (not to be confused with a left-invariant form!) if it is a zero of
the Lie derivatives with respect to all the generators X . As these constitute
a vector basis, we can recall eq.[7.162] to conclude that an invariant form
is closed. With the Cartan metric, the coderivative and the laplacian are
defined. From eq.[7.165], every invariant form on G is also co-closed, and
consequently harmonic. From these considerations follow very restrictive
results on the topology of such groups.2 For instance, the Betti numbers b1
and b2 are zero and b3 1. For simple groups, b3 = 1.
Goldberg l962.
272
Chapter 9
FIBER BUNDLES
9.1
INTRODUCTION
We have already met a few examples of fiber bundles: the tangent bundle, the
cotangent bundle, the bundle of linear frames. Also fibered spaces of another
kind have been seen: the covering spaces, whose fibers are discrete spaces
acted upon by the fundamental group. We shall in the following consider only
bundles with differentiable fibers differential fiber bundles. Of this type,
we have glimpsed tensorial bundles in general, which include the tangent
and the cotangent bundles as particular cases. Locally, bundles are directproduct manifolds, but globally they are nothing of sort. In reality, their
importance comes out mainly when global effects, or effects in the large,
are at stake. As Physics is largely based on (local) differential equations,
such effects are usually taken into account in the boundary conditions.
We start with an intuitive presentation of the bundle idea, then proceed
to examine the simplest cases, vector bundles. In vector bundles, whose prototypes are the tangent bundles, the fibers are vector spaces. We shall then
proceed to the more involved bundle of frames, on which linear connections
play their game. The frame bundle is the prototype of principal bundles,
whose fibers are groups and which are the natural setting summing up all
differential geometry. The frame bundle provides the natural background for
General Relativity, and abstract principal bundles do the same for Gauge
Theory.
9.1.1 Intuitively, a fiber bundle is a manifold (the base) to every point
of which one glues another manifold (the fiber). For example, the sphere
S 2 and all the planes ( E2 ) tangent to it (the sphere tangent bundle); or the
same sphere S 2 and all the straight half-lines ( E1+ normal to it (the normal
bundle). The classical phase space of a free particle is a combination of the
configuration space (base) and its momentum space (fiber), but it is a simple
cartesian product and, as such, a trivial bundle (Physical Topic 1). Notice
however that the base and the fiber do not by themselves determine the
273
274
CHAPTER 9. FIBER BUNDLES
bundle: it is necessary to specify how they are glued together. This is

the role of the projection mapping. For instance, with the circle S 1 and
the straight line E1 we can construct two different bundles: a trivial one
the cylinder and the Mobius band, which is nontrivial. They cannot be
distinguished by purely local considerations. By the way, the word trivial
is here a technical term: a bundle is trivial when it is globally a cartesian
product.
9.1.2 To illustrate the possible import to Physics, the simplest example
is probably the following1 : suppose we have a scalar field on S 1 , evolving in
time according to the 2-dimensional Klein-Gordon equation
(u
t + m2 ) = 0.
The dAlembertian u
t must be defined on the curved space (it is a LaplaceBeltrami operator, eq.7.92) formed by S 1 and the time in E1 . Now, how
can we account for the two different possible spaces alluded to above? The
equation is local and will have the same form in both cases. The answer lies
in the boundary conditions: in the cylinder, it is forcible to use periodic conditions, but on the Mobius band, which is twisted, one is forced to use antiperiodic conditions! Avis and Isham have performed the second-quantized
calculations and found a striking result: the vacuum energy (lowest energy
level) is different in the two cases! Fields on non-trivial bundles (twisted
fields) behave quite differently from usual fields. We usually start doing
Physics with some differential equations and suppose reasonable boundary
conditions. If the comparison with experiments afterwards show that something is wrong, it may be that only these conditions, and not the equations,
should be changed. Purely local effects (such as the values of fields around a
point at which the values are known) are relatively independent of topological (boundary) conditions. We say relatively because quantization is, as a
rule, a global procedure. It assumes well-defined boundary conditions, which
are incorporated in the very definition of the Hilbert space of wavefunctions.
Energy levels, for example, are clearly global characteristics.
9.1.3 Let us go back to the beginning: to constitute a bundle one appends
a fiber to each point of the base. This is of course a pictorial point of view.
One does not really need to attach to each point of the base its tangent space,
for example, but it is true that it helps conceiving the whole thing. The fiber
is, on each point, a copy of one same space, say, Em for the tangent space,
GL(m, R) for the frame bundle, etc. This abstract space, of which every fiber
is but a copy, is called the typical fiber .
1
Avis & Isham l978, 1979; Isham 1978.
9.2. VECTOR BUNDLES
9.2
275
VECTOR BUNDLES
9.2.1 Given a differentiable manifold M , a vector space F and an open set

U M , the cartesian product U F is a local vectorial bundle, for which
U is the base space. If x U , the product {x} F is the fiber on x. The
mapping : U F U such that (x, f ) = x for every f in F is the bundle
projection. Of course, the fiber on x is also 1 (x). As F is open, U F is
also open in M F .
9.2.2 A local fibered chart is a pair (U, ), where is a bijection
: U F U 0 F 0 En ,
with n large enough. Such a chart provides a local system of coordinates
(LSC) on the local bundle. A further condition is necessary: when we define
local bundles as above for two open sets Ui and Uj in M , the coordinate
transformation ij in the intersection Ui Uj , of the form ij = j 1
i ,
must be a diffeomorphism and obey
(x, f ) = (1 (x), 2 (x)f ),
(9.1)
where: x and f are coordinates of a point in the intersection and of a point

in F ; 1 is a coordinate transformation in the intersection; 2 is a mapping
taking the point represented by x into the set of linear mappings of F into
F 0 ; the result, 2 (x), is an x-dependent mapping taking f into some f 0 . A
set of charts (Ui , i ) satisfying the conditions of a complete atlas is a vector
fibered atlas. As usual with their kin, such an atlas allows one to extend the
local definitions given above to the whole M .
9.2.3 A vector bundle is thus built up with a base space M , a typical fiber
F which is a vector space and an atlas. Suppose further that a Lie group
G acts transitively on F : it will be called the structure group of the bundle.
The bundle itself, sometimes called the complete space, is indicated by
P = (M, F, G, ).
(9.2)
An important existence theorem2 states that:

given a Lie group G acting through a representation on a vector space F ,
and a differentiable manifold M , there exists at least one bundle
(M, F, G, ).
2
Steenrod l970.
276
9.2.4 A section is any C mapping

: U U F
(9.3)
such that, for every p U M , ((p)) = p. Such sections constitute by

themselves an infinite-dimensional linear function space.
9.2.5 We have already met the standard example of vector bundle, the
tangent bundle
T M = (M, Em , GL(m, R), T ).
(9.4)
A vector field is a section X : U M T U such that X(p) = Xp , that
is,
T X(p) = p.
9.2.6 This is a typical procedure: in general, points on the bundle are
specified by sections. In physical applications, F is frequently a Hilbert
space, wave-functions playing the role of sections. Why are bundles and
all that almost never mentioned? Simply because in most usual cases the
underlying bundles are trivial, simple cartesian products. In wave mechanics,
it is the product of the configuration space by a Hilbert space. Bundles
provide the geometrical backstage for gauge theories (see Physical Topic 7)
and it was precisely the flourishing of these theories that called attention to
the importance of non-trivial bundles to Physics, mainly after Trautman3
and Yang4 uncovered their deeply geometrical character.
9.2.7 We have above defined local sections. The reason is fundamental:
it can be shown that only on trivial bundles there are global sections. The
simple existence of a section defined everywhere on the base space (such as
the usual wavefunctions) of a bundle ensures its direct-product character.
Every bundle is locally trivial, only cartesian products are globally trivial.
For the tangent bundle T M , this would mean that a field X can be defined
everywhere on M by a single section. As we have said, M is, in this case,
parallelizable. Lie groups are parallelizable manifolds. The sphere S 2 is not
and, consequently, accepts no Lie group structure. In reality, only a few
spheres can be Lie groups (S 1 , S 3 and S 7 ), for this and other reasons.
9.2.8 In general, many different fibers F constitute bundles like [9.2], with
the same group G. They are called associated bundles. Bundles with a given
base space and a given structure group can be classified. The classification
3
4
Trautman l970.
Yang 1974.
9.3. THE BUNDLE OF LINEAR FRAMES
277
depends, fundamentally, only on the topologies of the base and the group.
It does not depend on the fiber. Consequently, the classification can be
realized by taking into account only the principal bundles, in which the fiber
is replaced by the group itself (see section 9.7).
9.2.9 In gauge theories, the source fields belong usually to (associated)
vector bundles.
9.2.10 Fibration All the fibers are isomorphic to the typical fiber in a
fiber bundle. The notion may be generalized to that of a fibration, in which
the fibers are not necessarily the same on each point of the base space. Let us
briefly describe the idea. Given two spaces E and M , a fibration is a mapping
: E M possessing a property called homotopy lifting, which consists
in the following. Consider another space K, and a map f : K E. This
leads to the composition g = f : K M . The map f is the lift of g.
Consider now the homotopy class of g, a set of homotopic maps gt with g0 = g.
If this homotopy lifts to a homotopy ft of f , with f0 = f , then is said to
have the homotopy property (a fiber bundle is a particular case, the bundle
projection being a locally trivial fibration). Thus, the only requirement now
is that all the fibers 1 (x), for x M , be of the same homotopy type. An
example of recent interest is the loop space: start by taking as total space the
set M0I of curves x(t) with initial endpoint x0 = x(0). The final endpoint of
each curve will be x1 = x(1). Take as fibration the mapping : x(t) x1
taking each path into its final endpoint. The mapping 1 (x) is the set of
all curves c0 (x) from x0 to x. Now, choose some fixed path c0 (x) from x0 to
x. Any other path of 1 (x) will differ from c0 (x) by a loop through x0 ,
so that each new c0 (x) determines an element of the loop space LM. The
mapping 1 (x) is homotopically equivalent to LM. Thus, in the fiber above,
the initial endpoint is LM, and satisfies the fibration requirements.
9.3
THE BUNDLE OF LINEAR FRAMES
There are many reasons to dedicate a few special pages to the bundle of linear
frames BM. On one hand, it epitomizes the notion of principal bundle and
is fundamental for geometry in general. This makes it of basic importance
for General Relativity, for example (Physical Topic 8). On the other hand,
precisely because of this particular link to the geometry of the base space, it
has some particular properties not shared by other principal bundles. The
intent here is to make clear its general aspects, which have provided the
historical prototype for the general bundle scheme, as well as stressing those
aspects well known to gravitation physicists which find no counterpart in the
bundles which underlie gauge theories.
278
9.3.1 Introduction Each point b of the bundle of linear frames BM on

the smooth manifold M is a frame at the point p = (b) M , that is, a
set of linearly independent vectors at p. The structural group GL(m, R) acts
on the right on BM as follows: given a = (aij ) GL(m, R) and a frame
b = (b1 , b2 , b3 , . . . , bm ), then
b0 = ba = (b01 , b02 , b03 , . . . , b0m )
with
b0i = aj i bj = bj aj i .
(9.5)
In the natural basis of a chart (U, x) around (b), bi will be written

bi = bi j x j
and {xi , bk l } provides a chart on BM around b. Take on Em the canonical
basis {Ki }, columns whose j-th element is ij : K1 = (1, 0, 0, . . . , 0), K2 =
(0, 1, 0, 0, . . . , 0), etc. The frame b BM can be seen as a mapping taking
the canonical basis {Kk } into {bk } (look at Figure 9.1, page 281). More
precisely: the frame b given by b = (b1 , b2 , . . . , bm ) is the linear mapping
b : Em T(b) M
b(Kk ) = bk .
(9.6)
Being a linear mapping between two vector spaces of the same dimension, b
is an isomorphism. It appoints the base member bj as the representative
of the canonical vector Kj belonging to the typical fiber Em . Consequently,
two vectors X and Y of T(b) M are images of two vectors r and s on Em :
X = b(r) = b(ri Ki ) = ri b(Ki ) = ri bi
and
Y = b(s) = b(si Ki ) = si b(Ki ) = si bi .
9.3.2 Structure group To see more of the action of the structure group
on BM, notice that GL(m, R) acts on the space Em on the left (actually,
we might build up an associated vector bundle with fiber Em ): if V Em ,
V = V k Kk , and a = (aj i ) GL(m, R),
(aV )j = aj i V i or aV = Kj aj i V i .
The mapping b will give
279
b(aV ) = b(Kj )aj i V i = bj aj i V i

and
(ba)(V ) = (ba)(V i Ki ) = V i (ba)(Ki ) = V i ui = bj aj i V i ,
where use has been made of eq.[9.5]. Consequently,
b(aV ) = (ba)(V ).
(9.7)
It is instructive to interpret through this equation the action of GL(m, R) on

BM: it says that the diagram
Em
a Em
ba&
b
T(b) M
is commutative. The mapping

b : GL(m, R) BM
b(a) := Ra (b) = ba,
(9.8)
when applied to the group identity, gives just the frame b: b(e) = be = b.
The derivative mapping will be
b : Ta GL(m, R) Tba BM
b (Ja ) = Xba .
(9.9)
A group generator J will be taken into a fundamental field (section 8.4):

J = b (Je ) = a certain Xb .
We may choose on the algebra G0 L(m, R a convenient, canonical basis given
by the m m matrices i j whose elements are
(i j )a b = ib aj .
(9.10)
In this basis, an element Je G0 L(m, R will be written J = J i j i j . The

Lie algebra will be defined by the matrix commutator, with the commutator
table
[i l , j k ] = ik j l jl i k .
(9.11)
280
The mapping b can be shown to be an algebra homomorphism between

G0 L(m, R and the Lie algebra of fields on BM at b. This allows us to introduce
a basis {Ei j } for the fundamental fields through
Ei j = b (i j ) = (i j ) .
(9.12)
Thus, J = Jj i Ei j . The homomorphism leads to

[Ei l , Ej k ] = ik Ej l jl Ei k .
(9.13)
Now, b is a constant mapping G (b), so that b = 0. Fields

X such that (X) = 0 are called vertical . The fundamental field is are
clearly of this kind,
(Ei j ) = 0.
(9.14)
They also obey eq.[8.35],
Ra (J ) = b [(Ada1 J)e ].
(9.15)
There is a one-to-one correspondence between the structure group and the

space of frames ( 6.5.6 and 8.2.16). The fiber coincides with the group, so
that the bundle is a principal bundle.
9.3.3 Soldering is a very special characteristic of the bundle of linear
frames, not found in other bundles (see Figure 9.1). It is due to the existence
of a peculiar vector-valued 1-form on BM, defined by
: Tb BM Em
:= b1 .
(9.16)
This composition of two linear mappings will be an Em -valued form. Called

canonical form, or solder form, can be written, in the canonical basis {Ki }
of Em as
= Ki i .
(9.17)
Each form k is called a soldering form. It is possible to show that, under
the right action of the structure group,
Rg = g 1 .
(9.18)
The name is not gratuitous. The presence of the solder form signals a coupling between the tangent spaces, to the bundle and to the base manifold,
which is much stronger for BM than for other principal bundles. A consequence is that a connection on BM will have, besides the curvature it shares
with all connections, another related form, torsion. Soldering is absent in
bundles with internal spaces, which consequently exhibit no torsion.
281
Figure 9.1:
9.3.4 Orthogonal groups Let us say a few more words on the relation
between the orthogonal groups and bilinear forms, introduced in 8.1.9. A
group of continuous transformations preserving a symmetric bilinear form
is an orthogonal group or, if the form is not positive-definite, a pseudoorthogonal group. If an element of the group is written , this defining
property takes the matrix form
T = ,
(9.19)
where T indicates the transpose. As a consequence, any member A of the

algebra, for which = eA for some , will satisfy
AT = A 1
(9.20)
and will have trA = 0. Let us go back to the real linear group GL(m, R, and
the basis [9.10]. Given the bilinear form , both basis and entry indices can
be lowered and raised, as in (a b )i j = ai bj . One finds for instance
[ab , cd ] = bc ad da cb .
(9.21)
In this basis a member K = K ab ab of the algebra will have as components its own matrix elements: (K)ij = K ij . The use of double-indexed basis
for the algebra generators is the origin of the double-indexed notation (peculiar to the linear and orthogonal algebras) for the algebra-valued forms, as
282
for example the connection = Ja b a b dx . A special basis for the (pseudo)

orthogonal group SO() corresponding to is given by the generators
Jab = ab ba = Jba .
(9.22)
All this leads to the usual commutation relations for the generators of
orthogonal and pseudo-orthogonal groups,
[J ab , J cd ] = bc J ad + ad J bc bd J ac ac J bd .
(9.23)
The usual group of rotations in 3-dimensional euclidean space is the special orthogonal group, indicated SO(3). Being special means connected
to the identity, that is, represented by 3 3 matrices of determinant = +1.
Orthogonal and pseudo-orthogonal groups are usually indicated by SO()
= SO(p, q), with (p, q) fixed by the signs in the diagonalized form of .
The group of rotations in n-dimensional euclidean space will be SO(n), the
Lorentz group will be SO(3, 1), etc.
9.3.5 Reduction We may in many cases replace GL(m, R) by some subgroup in such a way as to obtain a sub-bundle. The procedure is called
bundle reduction. For example, GL(m, R) can be reduced to the orthogonal subgroup O(m) (or to its pseudo-orthogonal groups). The bundle BM
of the linear frames reduces to the sub-bundle OM = (M, Op M, O(m), ),
where Op M is the set of orthogonal frames on Tp M . Now, Tp M is isomorphic
to Em , the typical fiber of the associated tangent bundle and on which there
exists an internal product which is just invariant under the action of O(m).
Let us consider a consequence of the reduction to SO().
9.3.6 Tetrads The most interesting point of reduction is that, in the
process, each basis of BM defines on M a Riemannian metric. Suppose on
Em the invariant internal product is given by the euclidean (or a pseudoeuclidean) metric , with
(r, s) = r s .
(9.24)
Given X = b(r) = ri bi and Y = b(s) = si bi as introduced in 9.3.1, a

Riemannian (or pseudo-Riemannian) metric on M can be defined by
g(X, Y ) = (b1 X, b1 Y ) = (r, s).
(9.25)
It is possible to show that g is indeed Riemannian (or pseudo-Riemannian,

if O(m) is replaced by some pseudo-orthogonal group). The procedure can
be viewed the other way round: given a Riemannian g on M , one takes the
283
subset in BM formed by the b = (b1 , b2 , . . . , bm ) which are orthogonal according to g. The resulting bundle, OM , is the bundle of orthogonal frames on
M . In the case of interest for General Relativity, it is the pseudo-orthogonal
Lorentz group SO(3, 1) which is at work in the tangent space, and we must
take for the Lorentz metric of Minkowski E3,1 space. Of course, SO(3, 1)
is also a subgroup of GL(4, R). Given a natural basis { }, a general basis
{h } of BM has elements
h = h ,
(9.26)
and its dual basis is {h },
h = h dx
(9.27)
h (h ) = h h = .
(9.28)
with
To reduce to the bundle OM , we impose orthogonality by some g:
g(h , h ) = g h h =
(9.29)
(here plays the role of a pseudo-euclidean Kronecker delta). We can

calculate, for X = X and Y = Y ,
g(X, Y ) = g dx (X)dx (Y ) = g h h (X)h h (Y )
= h (X)h (Y ) = h h X Y .
Thus, X and Y being arbitrary,
g = h h .
(9.30)
From this expression and [9.29], we recognize in these (pseudo-)orthogonal

frames the tetrad fields h (or four-legs, or still vierbeine). Each base {h }
determines a metric by eq.[9.30]: it translates the Minkowski metric into
another, Riemannian metric. It is important to notice that the tetrads belong
to the differentiable structure of the manifold. They are there as soon as some
internal product is supposed on the typical tangent space. Unlike connections
to be introduced in next section they represent no new, additional
structure. Their presence will be at the origin of torsion. Only the metric
turns up in the Laplace-Beltrami operator appearing in the Klein-Gordon
equation of 9.1.2, which governs the behavior of boson fields (Physical
Topic 8.2). Tetrads only come up explicitly in the Dirac equation. This fact
makes of fermions privileged objects to probe into tetrad fields. In particular,
they exhibit a direct coupling to torsion (Physical Topic 8.3).
284
9.4
LINEAR CONNECTIONS
Connections materialize in the notion of parallel transport. Consider a (piecewise differentiable) curve on a manifold. A vector X undergoes parallel
transport if it is displaced along in such a way that its angle with the curve
(i.e. with the tangent to ) remains constant (Figure 9.2). This intuitive view
supposes a metric (see 9.4.23), but actually a connection suffices. A connection determines a covariant derivative of the type described in 7.3.11, and
the vector is parallel-transported when its covariant derivative vanishes. The
covariant derivative can be extended to any tensor, and a tensor is paralleltransported when its covariant derivative vanishes. The notion of parallel
transport for general principal bundles will be introduced (in 9.6.21) in an
analogous way.
9.4.1 Ask a friend to collaborate in the following experiment.5 He must
stand before you, with his arm straight against his side but with his thumb
pointing at you. He must keep rigidly the relative position of the thumb with
respect to the arm: no rotation of thumb around arm axis allowed. He will
then (i) lift the arm sideways up to the horizontal position; (ii) rotate the
arm horizontally so that it (the arm) points at you at the end; you will be
seeing his fist, with the thumb towards your right; (iii) finally he will drop
the arm back to his side. The thumb will still be pointing to your right. The
net result will be a 90 rotation of the thumb in the horizontal plane. Notice
that his hand will have been moving along three great arcs of a sphere S 2 of
radius L (the arms length). Its is just as if you looked at the behaviour of a
vector on earth: (a) initially at the south pole S and pointing at you (see
Figure 9.3); (b) transported
Figure 9.2:
This simple example is adapted from Levi 1993.
9.4. LINEAR CONNECTIONS
285
Figure 9.3:
along the rim from S to H, all the time pointing at you; (c) taken along
the equator to the meridian just facing you (the vector becomes progressively visible and, at F , it will be entirely toward your right; (d) transported
southwards back to S. The net rotation is a measure of earths curvature.
For a small parallel displacement dxk , the variation of a vector X will be
given by
X i = i jk X j dxk ,
where i jk represents precisely the connection. Along a curve, one must
integrate. It so happens that the curvature is the rotational of , and that,
in the case above, the curve bounds one-eigth of the whole earths surface,
4L2 (see details in Mathematical Topic 10.3.1). Stokes theorem transforms
the line integral into the surface integral of the curvature over the octant.
The curvature is constant, and equal to 1/L2 , so that what remains is just
an octants area divided by L2 , which gives the rotation angle, /2.
The same procedure, if followed on the plane, which is a flat (zero curvature) space, would take the vector back to its initial position quite unmodified.
9.4.2 There is no such a thing as curvature of space. This is perhaps
still more evident when general connections are introduced (section 9.6).
Curvature is a property of a connection, and a great many connections may
be defined on the same space. Different particles feel different connections
and different curvatures. There might be a point for taking the Levi-Civita
connection as part of the very definition of spacetime, as is frequently done. It
seems far wiser, however, to take space simply as a manifold, and connection
(with its curvature) as an additional structure (see Physical Topic 8.2). We
shall here be concerned with the general notion of linear connection and
286
its formal developments, such as the relations involving curvature, frames

and torsion. Though nowadays presented in a quite intrinsic way, all this
has evolved from the study of subspaces of euclidean spaces. The historical
approach, embodied in the so-called classical differential geometry, is very
inspiring and a short account of it is given in Mathematical Topic 10.
9.4.3 A linear connection is a G0 -valued form leading fundamental fields
back to their corresponding generators:
(J ) = J.
In particular,
(Ei j ) = i j .
(9.31)
As is a form with values on G0 , it may be written in basis {i j } as

= i j i j , with i j usual 1-forms. Then,
k l (Ei j ) = ik lj .
(9.32)
A field X on BM such that (X) = 0 is said to be horizontal .

9.4.4 Notice that the definition of vertical fields as in eq.[9.14] is inherent to the smooth structure and quite independent of additional structure.
But horizontal fields are only defined once a connection is given. Distinct
connections will define distinct fields as horizontal.
9.4.5 5. A connection can be proven to satisfy, besides the defining relation given by (J ) = J, the covariance condition (Ra )(X) = Ada1 (X).
Conversely, any 1-form on the complete space satisfying both these conditions is a connection. Because of the form of the covariance condition, we
say that the connection belongs to the adjoint representation.
9.4.6 The presence of a connection has an important consequence on the
solder form . Once a connection is given, becomes an isomorphism of
vector spaces (though not of algebras see 9.4.8 below). There will be a
unique set {Ei } of horizontal vectors such that
(Ei ) = Ki or j (Ei ) = ij .
(9.33)
Notice that the fields Ei are exactly dual to the solder forms. Also
(Ei ) = b (Ei ) = bi .
Thus, on the base manifold and in a given basis {bi }, the soldering forms are
represented by that base of forms which is dual to {bi }.
287
Given any vector V = V j Kj in Em , the horizontal vector V j Ej on BM is

the basic or standard vector field associated to V .
For each frame b, the vectors Ei and Ei j can be shown to be linearly independent, so that the set {Ei , Ei j } constitute a basis for Tb BM . Although
connection-dependent, this basis is independent of the charts. The supertangent fiber bundle T BM obtained in this way has a structure of direct
product, and the complete space of the bundle of frames is consequently a
parallelizable manifold. Any vector at b may be decomposed into a vertical
and a horizontal part,
X = V X + HX = X i j Ei j + X i Ei .
(9.34)
9.4.7 Actually, the connection might be defined as a form vanishing on the

horizontal space Hb spanned by the Ei . Using the horizontal projection
X HX, the covariant differential of a p-form is that (p + 1)-form D
which, acting on (p + 1) vectors X1 , X2 , . . . , Xp+1 , gives
D(X1 , X2 , . . . , Xp+1 ) = d(HX1 , HX2 , . . . , HXp+1 ).
(9.35)
It is consequently the horizontalized version of the exterior derivative.

The covariant derivative is thus defined on the bundle complete manifold.
In order to bring it down to the base manifold, use must be made of a
section-induced pull-back.
9.4.8 We can further show that
[Ei j , Ek ] = kj Ei .
(9.36)
Nevertheless, unlike b , is not an algebra homomorphism. The algebra

basis {Ei , Eij } can be completed by putting
[Ei , Ej ] = Fm n ij En m + T k ij Ek .
(9.37)
The notation is not without a purpose. The detailed calculations give for
the coefficients Fm n ij and T k ij the values of the curvature and the torsion,
whose meaning we shall examine in the following. Let us retain here that the
curvature appears as the vertical part of the commutator of the basic fields,
and the torsion as its part along themselves.
The projection is a linear mapping from the start, which becomes
an isomorphism of vector spaces (between the horizontal subspace Hb and
T(b) M ) when a connection is added. The decomposition of Tb BM into horizontal and vertical is, therefore, complete in what concerns the vector space,
but not in what concerns the algebra.
288
9.4.9 Torsion Given a connection , its torsion form T is the covariant

differential of the canonical form ,
T = D.
(9.38)
It is consequently given by
T (X, Y ) = d(HX, HY ) = Ki di (X j Ej , X k Ek ).
We find that
T = 21 Kk T k ij i j .
(9.39)
In detail, the invariant expression of T is

T = d + +
(9.40)
If the solder form is written as Ki i = K h dx in a natural basis, the

components of the torsion tensor are
T = h h + h h .
(9.41)
If the tetrad field is trivial, that is, if it is a simple change of coordinates,

h =
y
,
x
which furthermore can be chosen to be locally the same for the manifold and
for the tangent space, h = , then
T = .
(9.42)
9.4.10 Curvature The curvature form R of the connection is its own

covariant differential, which is found to be
F = D = d + .
(9.43)
Being G0 -valued, its components are given by

F = 12 R .
(9.44)
R = + .
(9.45)
In a natural basis,
289
9.4.11 From the above expressions, by taking derivatives and reshuffling

the terms, we can find the first Bianchi identity
DF = dF + [, F ] = 0,
(9.46)
as well as the second Bianchi identity

dT + [, T ] + [, F ] = 0.
(9.47)
Putting T = 0 implies that F must obey the extra condition

[, F ] = 0.
(9.48)
9.4.12 A vector field X is parallel-transported along a curve (s) if it is

the projection of a horizontal field all along . Its covariant derivative must
then vanish. This means that, when displaced of d along , it satisfies
dX k
d j
+ k ij X i
= 0,
ds
ds
(9.49)
dX k = k ij X i d j .
(9.50)
or
A geodesic curve is a self-parallel curve, that is, a curve along which its
i
is parallel-transported. It obeys conseown tangent vector (velocity) d
ds
quently the geodesic equation
d2 k
ds
+ k ij d
ds
d j
ds
= 0.
9.4.13 More about geodesics is said in Mathematical Topic 12. Here, only
a few general aspects of them will interest us. Geodesics provide an easy
view of the meaning of curvature. Consider two bits of geodesics, d and
ds starting at O as in Figure 9.4. Displace d along d to obtain d0 , and d
along d to obtain d 0 , thereby constituting an infinitesimal parallelogram.
Take then a field X. If we parallel-transport X along d, it will change
according to [9.50]; if we propagate the resulting field along d 0 , it will again
be changed analogously. We shall find a vector X 0 at point c. Now start
again from O, going however first along d and then along d0 . We find X 00
at point c. The difference will then be fixed by the curvature:
X k = Rk ipq X i dp d q .
(9.51)
290
Figure 9.4:
9.4.14 Geodesics allows also an intuitive view of the effect of torsion, and
this in a rather abrupt way: torsion disrupts the above infinitesimal geodesic
parallelograms. In effect, if we transport as above the very geodesic bits in
the presence of torsion (as in Figure 9.5), we find that a gap between the
extremities shows up, and such that
k = [k ij k ji ] d i dj = T k ij d i dj .
(9.52)
If we look at the geodesic equation, it is clear that only the symmetric

part of the connection contributes. Torsion does not affect the form of individual geodesics, though it forbids geodesic parallelograms. A beautiful
manifestation of this effect is found in Elasticity Theory, where it is measured by the Burgers vector (Physical Topic 3.3.2). Spinor fields do couple
to torsion and are, probably, the best candidates for its eventual detection
in real space (Physical Topic 8.3).
9.4.15 Given two fixed fields X and Y , the curvature R can be seen as a
family of mappings R(X, Y ) taking a field into another field according to
R(X, Y )Z = [X Y Y X [X,Y ] ]Z
= X (Y Z) Y (X Z) [X,Y ] Z, (9.53)
where is the covariant derivative and X its projection along X (see
eq.[9.60] below). In the same token, the torsion can be seen as a field-valued
291
Figure 9.5:
2-tensor T such that

T (X, Y ) = X Y Y X [X, Y ].
(9.54)
The Bianchi identities (9.46,9.47) become, in this language,

(X R)(Y, Z) + R(T (X, Y ), Z) + (cyclic permutations) = 0,
(9.55)
and
(Z T )(X, Y ) R(X, Y )Z + T (T (X, Y ), Z) + (cyclic permutations) = 0.
(9.56)
Notice that, even when T = 0, neither is trivial. The covariant derivative
has the properties
f X Y = f X Y,
(9.57)
X+Z Y = X Y + Z Y.
(9.58)
9.4.16 The relation to the usual component form is given by

R(ea , eb )ec = Rf cab ef .
(9.59)
Let us profit to give some explicit expressions. Take a vector field basis
{ea }, with [ea , eb ] = f c ab ec and find first that
X Y = X a [ea Y c + Y b c ba ] ec ,
(9.60)
of which a particular case is

ea eb = c ba ec .
(9.61)
292
Then, we calculate
X Y = X a ea (Y b eb ) = X a [ea Y c + Y b c ba ]ec .
(9.62)
We find next:
[X, Y ] = [X a ea , Y b eb ] = [X a ea (Y c ) Y a ea (X c ) + X a Y b f c ab ]ec ;
(9.63)
[X,Y ] Z = [X(Y d ) Y (X d ) + X a Y b f d ab ]ed Z;
(9.64)
R(ea , eb )Z = ea (eb Z) eb (ea Z) f c ab ec Z.
(9.65)
and
Finally, using eq.[9.59], we obtain
Rf cab = ea f cb eb f ca + d cb f da d ca f db f g ab f cg .
(9.66)
In a holonomic basis, R( , ) = R . Some useful expressions are:

R(X, Y )Z = [X Y Y X [X,Y ] ]Z = X a Y b R(ea , eb )Z
= X a Y b Z d R(ea , eb )ed = X a Y b Z c Rf cab ef ; (9.67)
T (ea , eb ) = ea eb eb ea f g ab eg = (g ba g ab f g ab )eg ;
T ( , ) = T ;
T (X, Y ) = X a Y b T (ea , eb ) = X a Y b {c ba c ab f c ab }ec .
(9.68)
(9.69)
(9.70)
9.4.17 The horizontal lift of a vector field X on the base manifold M is

that (unique! see below) horizontal field X # on the bundle space which is
such that (Xb# ) = X(b) for all b on BM. We only state a few of the most
important results concerning this notion:
(i) for a fixed connection, the lift is unique;
(ii) the lift of the sum of two vectors is the sum of the corresponding lifted
vectors;
(iii) the lift of a commutator of two fields is the horizontal part of the
commutator of the corresponding lifted fields.
9.4.18 A horizontal curve on BM is a curve whose tangent vectors are all
horizontal. The horizontal lift of a smooth curve : [0, 1] M , t t on
the base manifold M is a horizontal curve t# on BM such that (t# ) = t .
Given (0), there is a unique lifted curve starting at a chosen point b0 = 0# .
This point is arbitrary in the sense that any other point on the same fiber
will be projected on 0 .
293
9.4.19 Parallel transport (or parallel displacement) along a curve: take

a point b0 on BM as above, such that (b0 ) = 0 . The unique horizontal lift
going through b0 will have an end point b1 , on another fiber 1 (1 ), or, if
we prefer, such that (b1 ) = 1 = (1). Now, if we vary the point b0 on the
initial fiber 1 (0 ), we shall obtain other points on the final fiber 1 (1 ).
This defines a mapping # between the two fibers. As any horizontal curve
is mapped into another horizontal curve by the group action Rg ,
# Rg = Rg # ,
then the mapping # : 1 (0 ) 1 (1 ) is an isomorphism. This isomorphism, by which each point of the initial fiber is taken into a point of the final
fiber, is the parallel displacement along the curve . In these considerations,
it is only necessary that the curve be piecewise C 1 .
9.4.20 Formal characterization Recall the mapping [9.6]: each frame
b is seen as a map from Em to T(b) M , which appoints the base member
bj as the representative of the j-th canonical vector Kj of Em . And it will
take a general vector X = X j Kj of the typical fiber Em into b(X) = X j bj on
M . Also vice-versa: given any vector V = V j bj tangent to M at p = (b),
its inverse will provide a vector b1 (V ) = V j Kj on Em . We may call b1 (V )
the paradigm of the tangent vector V . Now each frame b is a point on
BM. Consider b0 1 (0 ) as above. It will put any Vp of T(b) M into
correspondence with an euclidean vector V0 = b1
0 (Vp ). The same will be
true of any point b(t) = bt along the horizontal lift of the curve t . The
parallel-transported vector Vt at each point (t) of the curve is defined as
bt (V0 ) = bt [b1
0 (Vp )].
Thus, at each point, one takes the corresponding nominee of the same euclidean vector one started with. We say then that V is kept parallel to
itself.
9.4.21 Associated bundles (a more formal approach) The considerations on horizontal and vertical spaces, parallel displacements, etc, may be
transferred to any associated bundle AM on M , with typical fiber F . Let
us first give a formal definition of such an associated bundle. We start by
defining a right-action of the structure group G on BM F as follows: given
(b, v) BM F , we define
Rg : (b, v) (bg, g 1 v),
294
for each g G. Then AM may be defined as the quotient of BM F by this

action. There is then a natural mapping of BM F into AM , given by the
quotient projection. Given a point u on AM , the vertical space is defined
as the space tangent to F at u. Things are more involved for the horizontal
space. Consider the above natural mapping of BM F into AM and choose
a point (b, v) BM F which is mapped into u. Fix v F and this will
become a mapping of BM into AM . Then the horizontal subspace in AM
is defined as the image of the horizontal subspace of BM by this new mapping.
These considerations allow one to define the covariant derivative of a
section along a curve on an associated bundle AM , that is, of any tensor
field, given a connection on BM. Lifts on AM are defined in the same way
as those of BM. A section of AM on a curve t will be given by a mapping
such that AM (t ) = t all along the curve (Figure 9.6). For each
1
fixed t, call t#t+ the parallel displacement of the fiber AM
(t+ ) from t+
to t . Then the covariant derivative, which measures how much the section
deviates from horizontality in an infinitesimal displacement, is
1 #t+
{t
[(t+ )] (t )}.
0
Dt = lim
(9.71)
Figure 9.6:
The covariant derivative in the direction of a field X at a point p is the
covariant derivative along a curve which is tangent to X at p. From the
above definition of covariant derivative, a section (in general, a tensor) is
said to be parallel-displaced along a curve iff the corresponding covariant
derivative vanishes. When its covariant derivative is zero along any field,
it will be parallel-transported along any curve. In this case, we say that
the connection preserves the tensor. The definition is actually an usual
295
derivative, only taking into account the whole, invariant version of a tensor.
This means that, besides derivating the components, it derivates also the
basis members involved. Take a tensor like T = T e w . The covariant
derivative will be
DT = dT e w + T de w + T e dw .
Using eqs.[7.67] and [7.73] for the adapted frames of Mathematical Topic
10.1.3, it becomes
DT = [e (T ) + T T ] w e w .
The covariant derivative along a curve will be the contraction of this
derivative with the vector tangent to the curve at each point (that is, if u is
the tangent field, its index will be contracted with the derivative index). In
the above example, it will be
Du T = u [e (T ) + T T ] e w .
9.4.22 In this way, the above notion of covariant derivative applies to
general tensors, sections of associated bundles of the frame bundle, and gives
the usual expressions in terms of components (say) in a natural basis, duly
projected along the curve, that is, contracted with its tangent vector. Say,
for a covariant vector, we find the expressions used in 6.6.14,
D X = X; = X X .
(9.72)
This semicolon notation for the covariant derivative, usual among physicists, does not include the antisymmetrization. In invariant language,
DX = 21 D[ X] dx dx .
For a contravariant field,
D X = X; = X + X .
(9.73)
9.4.23 The Levi-Civita connection The covariant derivative of a metric

tensor will have components
D g = g; = g g g
(9.74)
This will vanish when the connection preserves the metric. The compo
nents of the torsion tensor in a natural basis are T
= . In
principle, there exists an infinity of connections preserving a given metric,
296
but only one of them has vanishing torsion. In this case, the connection is
symmetric in the lower indices and we can solve the above expression to find
= = 12 g [ g + g g ],
(9.75)
just the Christoffel symbol of eq.[6.74]. Summing up: given a metric, there
exists a unique torsionless connection which preserves it, whose components
in a natural basis are the usual Christoffel symbols and is called the LeviCivita connection of the metric. Usual Riemannian curvature is the curvature of this connection, which is the connection currently used in General
Relativity (see Phys.Topic 8). The hypothesis of gravitation universality
gives priority to this connection, as it says that all particles respond to its
presence in the same way.
It is with respect to this connection that parallel transport acquires the
simple, intuitive meaning of the heuristic introduction: a vector is paralleldisplaced along a curve if its modulus and its angle with the tangent to the
curve remain constant. Of course, measuring modulus and angle presupposes
the metric. And it is the curvature of this connection which is meant when
one speaks of the curvature of a (metric) space. The discovery of curved
spaces, or non-euclidean geometries (Math.Topic 11), has been historically
the germ of modern geometry.
9.4.24 Consider a manifold M and a point p M . We define the symmetry sp at p as a diffeomorphism of a neighbourhood U of p into itself which
sends exp(X) into exp(X) for all X Tp M . This means in particular that
normal coordinates change signs. When such a symmetry exists, the space
is said to be locally symmetric.
Suppose then that a linear connection is defined on M . We denote M
with this fixed connection by (M, ). A differentiable mapping f of M into
itself will be an affine transformation if the induced mapping f : T M
T M maps horizontal curves into horizontal curves. This means that f maps
each parallel vector field along each curve into a parallel vector field along
the curve f (). The affine transformations on M constitute a Lie group. If
the symmetry sp above is an affine transformation, (M, ) is an affine locally
symmetric manifold. This only happens when T = 0 and R = 0.
On the other hand, (M, ) is said to be an affine symmetric manifold if,
for each p M , the symmetry sp can be extended into a global affine transformation (compare with section 8.2.7). On every affine symmetric manifold
M the group of affine transformations acts transitively. Thus, M may be
seen as a homogeneous space, M = G/H. The connection on G will be the
torsion-free connection above referred to.
9.5. PRINCIPAL BUNDLES
9.5
297
PRINCIPAL BUNDLES
In a principal bundle the fiber is a group G. Other bundles with G as

the structure group are associated to the principal. General properties are
better established in principal bundles and later transposed to the associated.
The paradigm is the linear frame bundle.
9.5.1 We have already met the standard example of principal fiber bundle,
the bundle of linear frames
BM = (M, BP M, GL(m, R), B ),
(9.76)
in which the fiber Bp M is isomorphic to the structure group. Let (M, F, G, )

be a vector bundle, with G acting on F on the left. We can obtain a new
bundle by replacing F by G and considering the left action of G on itself. Such
will be a principal bundle, indicated by (M, G, ), and the bundle (M, F, G, )
is said to be associated to it. We have already seen that the tangent bundle
T M is associated to BM.
Recall the formal definition of an associated bundle for BM in 9.4.21.
It is a particular example of the general definition, in which, as there, we
start by defining a right-action of the structure group G on P F as follows:
given (b, v) P F , we define
Rg : (b, v) (bg, g 1 v),
for each g G. Then the associated bundle AM , with the fiber F on which a
representation of G is at work, is defined as the quotient AM = (P F )/Rg .
There is then a natural mapping : P F AM . Parametrizing b = (p, g),
is the quotient projection
(b, v) = ((p, g), v) = (p, v).
Or, if we prefer, (b, v) = (class of b on (P, v)) = (orbit of b by the action of
G, v).
9.5.2 Conversely, consider a principal bundle (M, G, ). Take a vector
space F which carries a faithful (i.e, isomorphic) representation of G:
: G Aut(F )
: (g).
(9.77)
The space F may be, for example, a space of column vectors on which G is
represented by n n matrices (g):
gf := (g)f.
(9.78)
298
A bundle (M, F, G, ) is got in this way, which is associated to (M, G, ). Notice that different representations lead to different associated bundles. There
are therefore infinite such bundles for each group.
9.5.3 Locally, a point in (M, F, G, ) will be represented by coordinates
(p, f ) = (x1 , x2 , . . . , xm , f 1 , . . . , f n ). A LSC transformation will lead to some
(p0 , f 0 ). Both f and f 0 belong to F . If the action of G on F is (simply)
transitive, there will be a (unique) matrix (g) such that f 0 = (g)f . Thus,
the group action accounts for the LSC transformations in the fiber.
9.5.4 A point on the principal bundle will be found by a section :
(p, f ) = (p),
(p, f ) = (p) = p.
(9.79)
9.5.5 Let us now proceed to the formal definition. It requires a lot of things
in order to ensure that the bundle as a whole is a differentiable manifold.
A C principal fiber bundle is a triplet
P = (M, G, )
such that P (the complete space) and M (the base space) are C differentiable manifolds, and G (the structure group) is a Lie group satisfying the
following conditions:
(1) G acts freely and effectively on P on the right, Rg : P G P ;
this means that no point of P is fixed under the action, and no subgroup of
G is the stability group of some point of P ;
(2) M is the quotient space of P under the equivalence defined by G,
M = P/G; the projection : P M is C ; for each p of M , G is simply
transitive on the fiber 1 (p); so, the fiber is homogeneous, and we say that
the group preserves the fiber;
(3) P is locally trivial: for every p of M , there exists a neighbourhood
U 3 p and a C mapping
FU : 1 (U ) G
such that FU commutes with Rg for every g in G; the combined mapping
fU : 1 (U ) U G
fU (b) = ((b), FU (b))
(9.80)
is a diffeomorphism, called a trivialization. Notice that there is a FU for each

U (see Figure 9.7).
299
With the first condition, we may generalize here the mapping b of eq.[9.6].
When the action is free and effective, there is a Lie algebra isomorphism
between the group Lie algebra and the tangent to the fiber ( 8.4.8). Take
an associated bundle AM , with a typical fiber F which is usually a vector
space and whose copies in the bundle we shall call realized fibers. Choose
a starting basis on F . A point b of the principal bundle may be seen as a
mapping from F into the realized fiber on (b), with image zb = Lb z0 , and
with z0 indicating the zero-section, which will be defined in 9.5.8. The
group identity will deputize ze = Le z0 = z0 as the set of representatives of
the starting basis members. Each member of F will be thus translated into a
point of the realized fiber, and each point z of the realized fiber will come
from a member b1 (z) of F , its paradigm. The typical fiber F is in this
way installed on (b).
Figure 9.7:
9.5.6 We have seen in sections 8.2 and 8.4 that, when a group G acts on
a manifold M , each point p of M defines a mapping
pe : G M, pe(g) = pg = Rg (p).
If the action is free and we restrict ourselves to orbits of p, this mapping
is a diffeomorphism. Well, given a point b in a fiber, every other point in
the fiber is on its orbit since G acts in a simply transitive way. Thus, eb is a
diffeomorphism between the fiber and G,
eb : G 1 ((b)) P
eb(g) = Rg (b) = bg.
(9.81)
300
9.5.7 To say that M = P/G is to say that (bg) = (b). To say that FU
commutes with Rg means that FU (bg) = FU (b)g, or
FU Rg (b) = Rg FU (b).
As a consequence (see Figure 9.8),
Rg (Xb ) = (Xb )
FU Rg = Rg FU .
(9.82)
(9.83)
Figure 9.8:
9.5.8 An important theorem says that

a bundle is trivial if and only if there exists a global C section,
that is, a C mapping : M P with idM . In the general case,
sections are only locally defined. Each trivialization defines a special local
section: for p M and b P with (b) = p, such a section is given by
U : U 1 (U )
U (p) = b[FU (b)]1
so that
b = U (p)FU (b).
(9.84)
Thus, if FU (b) = g, then U (p) = bg 1 . But then

FU (U (p)) = FU (bg 1 ) = FU (b)[FU (b)]1 = e.
This section takes p into a point U (p) such that fU (U (p)) = (p, e). It
is called the zero section of the trivialization fU .
301
9.5.9 FU is a mapping of P into G; when restricted to a fiber, it is a

diffeomorphism. Within this restriction, coordinate changes are given by the
transition functions
gU V : U V G
gU V (p) = FU (b)[FV (b)]1 .
(9.85)
They are C mappings satisfying

gU V (p) gV W (p) = gU W (p).
(9.86)
Notice that something like [FU (b)]1 will always take b into the point
(p, e), in the respective trivialization, but the point corresponding to the
identity e may be different in each trivialization. And, as each chart (U, x)
around p will lead to a different trivialization, the point on the fiber corresponding to e will be different for each chart. It is usual to write FU (x),
where x Em is the coordinate of p = (b) in the chart (U, x).
The bundle commonly used in gauge theories has an atlas (Ui , x(i) ) with
all the Ui identical. Changes of LSC reduce then to changes in the coordinate
mappings, and the transition functions gU V represent exactly the local gauge
transformations,6 which correspond here to changes of zero sections. From
b = U (p)FU (b) and V (p) = bFV1 (b)
it follows that
V (p) = U (p)FU (b)FV1 (b) = U (p)gU V (p),
(9.87)
which shows precisely how sections change under LSC transformations. Equation [9.85] says else that, given in the fiber,
U = gU V V .
(9.88)
If is a column vector in fiber space, this is written in matrix form:

i0 = gij
j .
(9.89)
Gauge transformations are usually introduced in this way: source fields

belonging to some Hilbert space are defined on Minkowski space. They
transform according to [9.89]
0i (x) = Sij (x)j (x).
6
Wu & Yang 1975.
(9.90)
302
This means that LSC transformations are at work only in the internal
spaces, the fibers. The fields carry a representation of the gauge group
in this way. In the fiber, a change of LSC can be looked at in two ways:
either as a coordinate change as above or as a point transformation with
fixed coordinates. The non-triviality appears only in the point dependence
of the transition function.
9.5.10 Giving the coordinate neighbourhoods and transition functions
completely characterizes the bundle. It is possible to show that:
(i) if either G or M is contractible, the bundle is trivial;
(ii) if a principal bundle is trivial, every bundle associated to it is trivial.
9.5.11 Sub-bundles Fiber bundles are differentiable manifolds. Can we
introduce the notion of immersed submanifolds, while preserving the bundle
characteristics? In other words, are there sub-bundles? The answer is yes,
but under rather strict conditions. We have seen a particular case in 9.3.5.
As there is a lot of structure to preserve, including group actions, we must
begin by defining homomorphisms between bundles. Given two bundles P
and P, with structure groups G and G, a bundle homomorphism between
them includes a mapping f : P P 0 and a group homomorphism h : G
G0 , with
f (bg) = f (b)h(g).
(9.91)
If f is an immersion (an injection) and h is a monomorphism (an injective
homomorphism), then we have an immersion of P in P . In this case, P is a
sub-bundle of P . If furthermore P and P have the same base space, which
remains untouched by the homomorphism, then we have a group reduction,
and P is a reduced bundle of P .
9.5.12 Induced bundles We may require less than that. Suppose now
a bundle P with base space B. If both another manifold B 0 and a continuous mapping f : B 0 B are given, then by a simple use of function
compositions and pull-backs one may define a projection, as well as charts
and transition functions defining a bundle P 0 over B 0 . It is usual to say that
the map f between base spaces induces a mapping f : P 0 P between
complete spaces and call P 0 = f P the induced bundle, or the pull-back
bundle. Suppose there is another base-space-to-be B 00 and another analogous
map f 0 : B 00 B leading to a bundle P 00 over B 00 in just the same way. If
B 0 = B 00 and the maps are homotopic, then P 0 and P 00 are equivalent. Such
maps are used to show the above quoted results on the triviality of bundles involving contractible bases and/or fibers. They are also used to obtain
general bundles as induced bundles of Stiefel manifolds, which allows their
classification ( 9.7.2)
9.6. GENERAL CONNECTIONS
303
9.5.13 It might seem that the above use of a general abstract group is farfetched and that Physics is concerned only with transformation groups acting
on physical spaces, such as spacetime and phase spaces. But the above
scheme is just what appears in gauge theories (Physical Topic 7). Gauge
groups (usually of the type SU (N )) are actually abstract, acting on some
internal spaces of wavefunctions defined on Minkowski base space. The
first statement in 9.5.10 would say that, if Minkowski space is contractible,
the bundles involved in gauge theories are trivial if no additional constraints
are imposed via boundary conditions.
9.6
GENERAL CONNECTIONS
A connection is a structure defined on a principal bundle. We have called

linear connections those connections on the bundle of linear frames. Let
us now quote the main results on connections in general.
9.6.1 Consider the tangent structure of the complete bundle space P . At
a given point b, Tb P has a well defined decomposition into a vertical space
Vb , tangent to the fiber at b and its linear complement, which we shall (quite
prematurely) call the horizontal space Hb . The vertical space is defined as
Vb = {X Tb P such that X = 0}.
(9.92)
In words: projects a vector on P into a vector on the base space M . A

vertical vector lies along the fiber and projects into the zero of T(b) M . As
to Hb , the mere fact that it is the linear complement to Vb fixes it at b, but
is not sufficient to determine it in other points in a neighbourhood of b.
9.6.2 The mapping eb : G P given in eq.[9.81] induces the differential
deb = eb : Tg G Tbg P
eb : Xg X bg ,
taking fields on G into fields on P . The composition eb : G M is a
constant mapping, taking all the points of G into the same point (b) of M :
( eb)(g) = (bg) = (b).
Thus, d(( eb) = eb = 0. Consequently, given any X on G, eb (X) is
vertical. Recall what was said in 8.4.8: applied to the present case, G is
304
acting on P and eb (X) is a fundamental field. Given the generators {Ja } in

G0 = Te G, eb will take them into the fundamental fields
Ja = eb (Ja ) Tbe P = Tb P.
(9.93)
Each Ja will be a vertical field and the algebra generated by {Ja }, which
is tangent to the fiber, will be isomorphic to G0 and will represent it on P .
Given a vertical field X, there is a unique X in G0 such that X = eb (X).
The field X is said to be engendered by X, which it represents on P .
The set {Ja } may be used as a basis for Vb .
9.6.3 3.A fundamental field X, under a group transformation, will change
according to eq.[8.35]:
Rg (X) = Adg1 (X) = eb [Adg1 (X)].
(9.94)
9.6.4 Strong relations between the tangent spaces to P , M and G arise

from the mappings between them. A trivialization around b will give
fU (b) = (, FU )(b) = ((b), FU (b)) = (p, g).
It is convenient to parametrize b by writing simply b = (p, g). The mapping
eb will take a Xe in G0 into its fundamental representative
eb (Xe ) = X b = X (p,g) .
(9.95)
If b0 = (p, e), then eb0 (Xe ) = X (p,e) . But

X (p,g) = Rg X (p,e) = Rgeb (Xe ).
This is also
X (p,g) = eb (Xg ) = eb Rg (Xe ).
Recall that FU commutes with Rg .
9.6.5 A general vector Xb can be decomposed into vertical and horizontal
components,
Xb = Vb X + Hb X.
(9.96)
Trouble comes out when we try to separate the components in this equation. The local trivialization shows, as discussed above, that a part of Vb X
comes from G as a fundamental field. We have obtained a purely vertical
contribution to Xb , that coming from G0 . Let us see how a contribution may
305
come from M , the base space. Take a field Xp Tp M . It can be lifted to

Tb M by a section, such as the zero section of eq.[9.84]:
b(p,e) .
U Xp = X
(9.97)
b(p,g) , it is enough to displace the point by the group action,

To obtain X
bb = X
b(p,g) = Rg U (Xp ).
X
(9.98)
A field Xb on the complete space will thus have two contributions: this one
from Tp M , and another, purely vertical, coming from G as in eq.[9.95]:
bb + X b .
Xb = Rg U (Yp ) + eb (Xe ) = X
(9.99)
bb is
Now comes the trouble: X b is purely vertical by construction, but X
not necessarily purely horizontal! It may have a vertical part (Figure 9.9).
The total vertical component will be, consequently, X b plus some contribubb .
tion from X
Figure 9.9:
9.6.6 Consequently, the decomposition [9.96] is, in the absence of any

further structure, undefined. Putting it into another language: P is locally
a direct product U G and we are looking for subspaces of Tb P which are
tangent to the fiber ( G) and to an open set in the base space ( U ). The
306
fields tangent to a submanifold must constitute a Lie algebra by themselves,

because a submanifold is a manifold. The horizontal fields do not form
a Lie algebra: the commutator of two of them has vertical components.
When acted upon (through the Lie derivative) by another horizontal field, a
horizontal field Hb X comes up with a vertical component. In a neighbouring
point b0 , Hb0 X will not necessarily be that same field Hb X at b0 .
9.6.7 Summing up: in a bundle space it is impossible to extricate vertical
and horizontal subspaces without adding further structure. The additional
structure needed is a connection. Strictly speaking, a connection is a differentiable distribution ( 6.4.33), a field H of (horizontal) subspaces Hb of
the tangent spaces Tb P satisfying the conditions:
(i) at each b, Hb is the linear complement of Vb , so that every vector may
be decomposed as in eq.[9.96];
(ii) Hb is invariant under the right action of G, that is, Rg Hb = Hbg .
This field of horizontal subspaces is the kernel of a certain 1-form which
completely characterizes it. It is indeed quite equivalent to this 1-form, and
we shall prefer to define the connection as this very form.
9.6.8 A connection is a 1-form on P with values on G0 which spells out
the isomorphism between G0 and Vb :
: Vb G0
(Xb ) = Z G0 such that eb Z = Xb ,
or Z = Xb .
(9.100)
It is a vertical form: when applied to any horizontal field, it vanishes:

(Hb X) = 0 X.
(9.101)
(J a ) = Ja .
(9.102)
In particular, from [9.93],

9.6.9 Being a form with values in G0 , it can be written
= Ja a
(9.103)
a (J b ) = ba .
(9.104)
It is, in a restricted sense, the inverse of eb . It is equivariant ( 8.2.19)
because it transforms under the action of G in the following way: for any
fundamental field X,
(Rg )(X) = (Rg X) = (Adg1 X) =
eb (Adg1 X) = Adg1 X = Adg1 X(X).
307
Therefore,
Rg = Rg = Adg1 ,
(9.105)
(Rg X) = (Rg )(X) = (g 1 Ja g)a (X).
(9.106)
or
A scheme of the various mappings involved is given in Figure 9.10.
Figure 9.10:
9.6.10 A connection form defines, through [9.101], horizontal spaces at

every point b of P . Notice that, in principle, there is an infinity of possible
connections on P , and distinct connections will determine different horizontal
spaces at each b. Gauge potentials are connections on principal fiber bundles
(a statement to be slightly corrected in 9.6.14). From a purely geometrical
point of view they are quite arbitrary. Only under fixed additional dynamical
conditions (in the case, they must be solutions of the dynamical Yang-Mills
equations with some boundary conditions) do they become well determined.
In order to transfer the decomposition into vertical and horizontal spaces
to an associated bundle, we again recall what happened in the frame bundle
case ( 9.4.21). Given a point u on an associated bundle AM with fiber F ,
the vertical space is simply defined as the space tangent to F at u. For the
horizontal space, we take the natural mapping of P F into AM ( 9.5.1)
and chose a point (b, v) P F such that (b, v) = u. Fix v: this will become
a mapping v of P into AM , v (b) = u0 . Then the horizontal subspace in
AM is defined as the image of the horizontal subspace of P by v .
308
9.6.11 Well, all this may be very beautiful, but is a form on P , it inhabits (the cotangent bundle of) the bundle space. We need a formulation
in terms of forms on the base space M . In order to bring down to M ,
we will be forced to resort to the use of sections the only means to pull
back. Let us consider two zero sections related to two charts on M , with
open sets U and V . They will be related by eq.[9.87],
V (p) = U (p)gU V (p),
(9.107)
where the transition function gU V , given by [9.85], mediates the transformation of section U to section V . It is a mapping
gU V : U V G.
Its differential will take a field X on M into a field on G,
gU V : Tp M TgU V G
gU V : X gU V (X).
(9.108)
9.6.12 Recall the behaviour of the Maurer-Cartan form w on G: if applied

to a field on G, it gives the same field at the identity point, that is, the same
field on G0 (section 8.3.6). It will be necessary to pull w back to M , which
will be done by gU V . We shall define the G-valued form on M as
wU V (X) = (gU V w)(X) = w gU V (X).
(9.109)
To the field gU V (X) on G will correspond a fundamental field on P by

the mapping of eq.[8.34]. As here b = U (p), we can write
gU V (X) =
eU [gU V (X)].
(9.110)
As to the connection , it will be pulled back to M by each section:

U = U ;
V = V
(9.111)
Notice that U and V are forms (only locally defined) on M with values
in G.
9.6.13 Now, let us take the differential of [9.107] by using the Leibniz
formula, and apply the result to a field X on M :
V (X) = U (X)gU V (p) + U (p)gU V (X),
(9.112)
309
whose detailed meaning is

bU [gU V (X)].
V (X) = RgU V [U (X)] +
(9.113)
When we apply to this field on P , the first term on the right-hand side
will be
{RgU V [U (X)]} = (RgU V )(U X) = (Adg1 )U X
UV
by [9.105], and thus also equal to

Adg1 (U )(X) = Adg1 U (X).
UV
UV
The second term, on the other hand, will be {gU V (X)} by eq.[9.110].
This will be a certain field of G, in reality the field gU V (X) on G brought to
G0 , of which gU V (X) = U gU V (X) is the fundamental representative field
on P . A field on G is brought to G0 by the Maurer-Cartan form w, so that
{gU V (X)} = w[gU V (X)] = (gU V w)(X) = wU V (X).
Consequently, (V (X)) = (V )(X) will be
V (X) = Adg1 U (X) + wU V (X),
(9.114)
V = Adg1 U + wU V .
(9.115)
UV
or
UV
This gives pulled back to M , in terms of a change of section defined

by gU V . In general, one prefers to drop the (U, V ) indices and write this
equation in terms of the group-valued mapping g = gU V . Using eqs.[9.106]
and [8.17], it becomes
0 = g 1 g + g 1 dg.
(9.116)
In this notation, we repeat, and 0 are G0 -valued 1-forms on M . In a
natural basis, = dx = Ja a dx ,
0 = g 1 g + g 1 g
(9.117)
Ja 0a = g 1 Ja ga + g 1 g.
(9.118)
and
310
9.6.14 In these expressions the reader will have recognized the behaviour
of a gauge potential under the action of a gauge transformation, or the change
in the Christoffeln due to a change of basis. Here, the small correction we
promised in 9.6.10: gauge potentials are pulled-back connections on principal fiber bundles with the gauge group as structure group and space-time
as base space.7 They are defined on the base space and so they are sectiondependent. A section is what is commonly called a gauge. Changes of
sections are gauge transformations. The geometrical interpretation of the
underlying structure of gauge theories, pioneered by Trautman and a bit resented at first, is nowadays accepted by everybody.8 It helps clarifying many
important points. Let us only call attention to one of such. The vacuum
term g 1 dg in [9.116] is a base-space representative of the Maurer-Cartan
form, eq.[9.109]. This form is a most important geometrical characteristic of
the group, in reality connected to many of its topological properties. The
vacuum of gauge theories is thereby strongly related to the basic properties
of the gauge group.
9.6.15 A 1-form satisfying condition [9.101] is said to be a vertical form.
On the other hand, a form on P which vanishes when applied to any vertical
field,
(Vp X) = 0, X,
(9.119)
is a horizontal 1-form. Clearly, the canonical form [9.103] on the frame
bundle is horizontal. Vertical (and horizontal) forms of higher degrees are
those which vanish when at least one horizontal (respectively, vertical) vector
appears as their arguments.
9.6.16 Given a connection , horizontal spaces are defined at each point
of P . Given a p-form on P with values in some vector space V , its absolute
derivative (or covariant derivative) according to the connection is the (p +
1)-form
D = Hd = d H,
(9.120)
where H is the projection to the horizontal space:
D(X1 , X2 , . . . , Xp+1 ) = d(HX1 , HX2 , . . . , HXp+1 ).
(9.121)
The covariant derivative is clearly a horizontal form. An important property of D is that it preserves the representation: if belongs to a representation, so does D. For example, if
Rg = Adg1 ,
7
8
Trautman l970.
Daniel & Viallet l980; Popov l975; Cho l975.
311
then also
Rg D = Adg1 D.
(9.122)
This property justifies the name covariant derivative. A connection also

defines horizontal spaces in associated bundles, and a consequent covariant
derivative. As fibers are carrier spaces to representations of G, D will in that
case take each element into another of the same representation.
9.6.17 Going back to principal bundles, D is the horizontally projected
exterior derivative on P . If we take V = G0 and = itself, the resulting
2-form
F = D
(9.123)
is the curvature form of . From eq.[9.122],
Rg F (X1 , X2 ) = F (Rg X1 , Rg X2 ) = Adg1 F (X1 , X2 ).
(9.124)
Being a form with values in G0 , it can be written

F = 12 Ja F a ,
(9.125)
with { } a base of horizontal 1-forms. Equation [9.124] is then

Rg F = F 0 = 12 g 1 Ja gF a ,
(9.126)
F 0 = g 1 F g.
(9.127)
or
9.6.18 A closed global expression of F in terms of is got from eq.[7.26],

which for the present case is
2d(X, Y ) = X[(Y )] Y [(X)] ([X, Y ]).
(9.128)
A careful case study for the vertical and horizontal components of X and Y
leads to9
F (X, Y ) = d(X, Y ) + 12 [(X), (Y )],
(9.129)
which is the compact version of Cartans structure equations. This is to
be compared with the results of 7.3.12: there, we had a zero-curvature
connection (also called a flat connection) on En . A better analogy is found
in submanifolds (Mathematical Topic 10).
9
Kobayashi & Nomizu l963, vol. I.
312
9.6.19 We can write more simply

F = d + .
(9.130)
An immediate consequence is the Bianchi identity

DF = 0,
(9.131)
which follows from DF = dF H and H = 0.

9.6.20 The curvature form can be pushed back to the base manifold by
a local section , as was done for the connection form. It is customary to
simplify the notation by writing ( F ) = F . In a natural basis, eq.[9.130]
becomes
F = 21 Ja F a dx dx ,
(9.132)
F a = a a + C a bc b c .
(9.133)
where
Here, we recognize the expression of the field strength in terms of gauge

potentials. For linear connections, it is convenient to use the double-index
notation for the generators Ja b of the linear group GL(m, R) (or of one of its
subgroups). In that case, the above a becomes a b , F a becomes Fa b ,
and so on. Using the structure constants for GL(m, R), eq.[9.133] acquires its
usal form for linear connections. With the Lorentz group generators, it becomes the usual expression of the curvature tensor in terms of the Christoffel
symbols. Notice here a very important point: curvature is a characteristic of
a connection. What we have is always the curvature of a connection.
9.6.21 The simplest way to introduce the notion of parallel transport for
general principal bundles is through the covariant derivative. An object is
parallel-transported when its derivative is zero along the curve. Parallelism
on associated bundles may be introduced along the lines of 9.6.20. A
member of a realized fiber is taken parallelly along a curve when its paradigm
in the typical, abstract fiber remains the same (Figure 9.11).
In such a case, G acts through some representation (section 8.4). A
member of the fiber belongs to its carrier space. In the case of a gauge
theory, a source field will belong to some Hilbert space fiber on which G
performs gauge transformations. Think it as a column vector and let {Ta }
be a basis for the corresponding matrix generators. The connection will be
313
Figure 9.11:
a matrix = Ta a or, in a local natural basis, = Ta a dx . The covariant

derivative will be
D = d + = ( + a Ta )dx .
(9.134)
The covariant derivative D allows one to introduce a notion of parallelism in

neighbouring points: (x) is said to be parallel to (x + dx) when
(x + dx) (x) = dx ,
(9.135)
that is, when D0 at x. What is the meaning of it? For linear connections
and tangent vectors, this generalizes the usual notion of parallelism in euclidean spaces. In gauge theories, in which the associated fibers contain the
source fields, it represents a sort of internal state preservation. Suppose
a gauge model for the group SU (2), as the original Yang-Mills case. The
nucleon wavefields will have two components, being isotopic-spin Pauli
spinors. Suppose that at point x the field is a pure up spinor, representing a proton. If D = 0 at x, then (x + dx) given by eq.[9.135] will also be
a pure proton at (x + dx). All this is summed up again in the following: all
over a parallel transport of an object, the corresponding element in typical
fiber stays the same (Figure 9.12).
314
Figure 9.12:
9.6.22 Holonomy groups Consider again a principal fiber bundle P =

(M, G, ). Fix a point p on M and consider the set of all closed curves
(supposed piecewisesmooth) with endpoints p. The parallel displacement
according to a connection along each loop will define an isomorphism of the
fiber 1 (p) into itself. The set of all such isomorphisms constitute a group
Hol(p), the holonomy group of at the point p. The subgroup engendered
by those loops which are homotopic to zero is the restricted holonomy group
of at point p, denoted Hol0 (p). In both cases, the group elements take
a member of the fiber into another, so that both groups may be thought of
as subgroups of the structure group G. It is possible to show (when M is
connected and paracompact) that:
(i) Hol(p) is a Lie subgroup of G;
(ii) Hol0 (p) is a Lie subgroup of Hol(p);
(iii) Hol0 (p) is a normal subgroup of Hol(p), and Hol(p)/Hol0 (p) is countable.
Take an arbitrary point b belonging to the fiber 1 (p), and consider the
parallel transport along some loop on M : it will lead b to some b0 = ba, so
that to each loop will correspond an element a of G. In this way one finds
a subgroup of G isomorphic to Hol(p). We call this group Hol(b). Had we
chosen another point c on the fiber, another isomorphic group Hol(c) would
be found, related to Hol(b) by the adjoint action. Thus, on any point of the
fiber we would find a representation of Hol(p).
315
Consider an arbitrary b 1 (p). Call P (b) the set of points of P which

can be attained from b by a horizontal curve. Then it follows that (i) P (b) is
a reduced bundle with structure group Hol(b), and (ii) the connection is
reducible to a connection on this sub-bundle. P (b) is called the holonomy
bundle at b.
9.6.23 The Ambrose-Singer holonomy theorem Take the holonomy
bundle P (b) and fix a point h P (b). Consider the curvature form F of the
connection at h, Fh . The theorem states that the Lie algebra of Hol(b) is
a subalgebra of G spanned by all the elements of the form Fh (X, Y ), where
X and Y are arbitrary horizontal vectors at h.
9.6.24 Berrys phase Fields or wavefunctions respond to the presence
of connections by modifying the way they feel the derivatives. We have
seen that they can answer to linear connections and to gauge potentials. If a
wavefunction represents a given system, the effect can be seen by displacing
along a line. As a member of an associated bundle, will remain in the
same state if parallel transported. In the abelian case, integration of the
condition D = 0 along the line adds a phase to (as in 4.2.18). In the
non-abelian case, moves in the internal space in a well-defined way (as
in 9.6.21). There is, however, still another kind of connection to which a
system may react: connections defined on parameter spaces.
The effect was brought to light in the adiabatic approximation.10 Suppose
we have a system dependent on two kinds of variables: the usual coordinates
and some parameters which, in normal conditions, stay fixed. We may think
of an electron in interaction with a proton. The variables describing the
position of the proton, very heavy in comparison with the electron, can be
taken as fixed in a first approximation to the electron wavefunction . Consider now the proton in motion its variables are slow variables and can
be seen as parameters in the description of the electron. Suppose that it
moves (follows some path on the electron parameter space) somehow and
comes back to the initial position. This motion is a closed curve in the parameter space. Normally nothing happens the electron just comes back to
the original wavefunction . But suppose something else: that the parameter space is curved (say, the proton is constrained to move on a sphere
S 2 ). A non-flat connection is defined on the parameter space. The loop described by the proton will now capture the curvature flux in the surface it
circumscribes (again as in 4.2.18, or in the experiment of 9.4.1). The
wavefunction will acquire a phase which, due to the curvature, is no more
vanishing. This is Berrys phase.11 The connection in parameter space has
10
11
Berry 1984.
Simon 1983.
316
the role of an effective vector potential. This is of course a very crude example, which is far from doing justice to a beautiful and vast subject.12 Even
in a fixed problem, what is a parameter and what is a variable frequently
depends on the conditions. And it was found later13 that this geometrical
phase (sometimes also called holonomy phase) can appear even in the absence of parameters, in the time evolution of some systems. There is now a
large amount of experimental confirmation of its existence in many different
physical situations.
9.7
BUNDLE CLASSIFICATION
Let us finish the chapter with a few words on the classification of fiber bundles. Steenrods theorem ( 9.2.3) is a qualitative result, which says that
there is always at least one bundle with a base space M and a group G. We
have seen ( 9.1.1) that with the line and the circle at least two bundles are
possible: the trivial cylinder and the twisted Mobius band. Two questions14
come immediately to the mind, and can in principle be answered:
(1) In how many ways can M and G be assembled to constitute a complete
space P ? The answer comes out from the universal bundle approach.
(2) Given P , is there a criterion to measure how far it stands from the
trivial bundle? The answer is given by the theory of characteristic classes.
Notice to begin with that each associated bundle is trivial when the principal bundle P is trivial. Consequently, an eventual classification of vector
bundles is induced by that of the corresponding principal and we can concentrate on the latter.
9.7.1 Back to homogeneous spaces Let us recall something of what
was said about homogeneous spaces in 8.2.7. If M is homogeneous under
the action of a group G, then we can go from any point of M to any other
point of M by some transformation belonging to G. A homogeneous space
has no invariant subspace but itself. Simple Lie groups are homogeneous
by the actions (right and left) defined by the group multiplication. Other
homogeneous spaces can be obtained as quotient spaces. The group action
establishes an equivalence relation, p q, if there exists some g G such
that q = Rg p. The set
[p] = { q M such that q p } = orbitG (p)
12
13
14
A very good review is Zwanziger, Koenig & Pines 1990.

Aharonov & Anandan 1987 and 1988.
Nash & Sen l983.
9.7. BUNDLE CLASSIFICATION
317
is the equivalence class with representative p. The set of all these classes is the
quotient space of M by the group G, denoted by M/G and with dimension
given by: dim M/G = dim M - dim G. The canonical projection is
: M M/G,
p [p].
All the spheres are homogeneous spaces: S n = SO(n + 1)/SO(n). Consider in particular the hypersphere S 4 = SO(5)/SO(4). It so happens that
SO(5) is isomorphic to the bundle of orthogonal frames on S 4 . We have thus
a curious fact: the group SO(5) can be seen as the principal bundle of the
SO(4)-orthogonal frames on S 4 . This property can be transferred to the de
Sitter spacetimes (Physical Topic 9): they are homogeneous spaces with the
Lorentz group as stability subgroup, respectively
DS(4, 1) = SO(4, 1)/SO(3, 1) and DS(3, 2) = SO(3, 2)/SO(3, 1).
And the de Sitter groups are the respective bundles of (pseudo-) orthogonal
frames. Can we generalize these results? Are principal bundles always related
to homogeneous spaces in this way? The answer is that it is almost so. We
shall be more interested in special orthogonal groups SO(n). In this case, we
recall that the Stiefel manifolds ( 8.1.15) are homogeneous spaces,
Snk = SO(n)/SO(k).
They do provide a general classification of principal fiber bundles.
9.7.2 Universal bundles Given a smooth manifold M and a Lie group
G, it is in principle possible to build up a certain number of principal fiber
bundles with M as base space and G as structure group. One of them is
the direct product, P = M G. But recall that the general definition of
a principal bundle includes the requirement that M be the quotient space
of P under the equivalence defined by G, that is, M = P/G. Let us take
a bit seriously the following easy joke. Say, to start with, that principal
bundles are basically product objects obeying the relation M = P/G, which
are trivial when in this relation we can multiply both sides by G in the nave
arithmetic way to obtain P = M G. Thus, nontrivial bundles are objects
P = M G, where is a twisted product generalizing the cartesian
product as the Mobius band generalizes the cylinder. how many of such
twisted products are possible? Answering this question would correspond
to somehow classifying the principal fiber bundles. It is actually possible to
obtain a homotopic classification of the latter, and this is done through a
construct which is itself a very special fiber bundle. This special fiber bundle
is the universal bundle for G.
318

The principal bundle P = (M, G, ) is universal for the group G
if the complete space P is contractible, that is, if all
the homotopy groups k (P ) = 0.
When k (P ) = 0 only for k < m, P is called m-universal. We shall

only consider the cases in which G is either an orthogonal or an unitary
group. Given G, it would be enough to find some contractible bundle with
G as structure group. We shall find it, the base space being a Grassmann
manifold and the complete space being a Stiefel manifold.
Take the orthogonal group O(m) and its subgroup O(d), with d m.
Let us consider, in the notation of 9.5.1, which is
complete space = (base space, structure group, projection),
the bundle
O(m)/O(m d) = (O(m)/O(d)xO(m d), O(d), quotient projection).
If we recall what has been previously said on Grassmann spaces ( 1.4.20,
1.4.21 and 8.1.14) and on Stiefel manifolds ( 8.1.15), we see that this is
Stiefel = (Grassmann, O(d), quotient projection)
or
Smd = (Gmd , O(d), projection).
The situation, of which a local scheme is given in Figure 9.13, is just that
of a bundle of base space Gd , fiber O(d) and complete space Sd . We may as
Figure 9.13:
9.7. BUNDLE CLASSIFICATION
319
well obtain Smd = (G#

md , SO(d), projection) for the covering of Gmd . In the
C
complex case, we can construct bundles Smd
= (GC
md , U (d), projection).
As said in the mentioned sections, r (Smd ) = 0 for (m d) > r 0,
C
) = 0 for 2(m d) > r 0. As a consequence, we may take
and r (Smd
C
will have all the
m to infinity while retaining d fixed: then S d and Sd
homotopy groups trivial and will be contractible. We have thus universal
bundles whose base spaces are Grassmann manifolds of d-planes of infinitedimensional euclidean (real or complex) spaces. We shall call these base
spaces generically B.
Now, we state the Classification theorem:
Consider the principal bundles of base M and group G; the set of bundles
they form is the set of homotopy classes of the mappings M B.
It may seem rather far-fetched to employ infinite-dimensional objects,
but in practice we may use the base for the m-universal bundle. In this way
the problem of classifying bundles reduces to that of classifying mappings
between base manifolds. The number of distinct twisted products is thus
the number of homotopy classes of the mappings M B.
In reality there is another, very strong result:
Given P as above, then for a large enough value of m there exists a map
f : M Gmd such that P is the pull-back bundle of Smd :
P = f Smd .
Thus, every bundle is an induced bundle (see 9.5.12) of some Stiefel manifold. The Stiefel spaces appear in this way as bundle-classifying spaces.
Formally at least, it would be enough to consider real Stiefel spaces and
orthogonal (or pseudo-orthogonal) groups, because any compact group is a
subgroup of some O(q) for a high enough q. U (n), for instance, is a subgroup
of O(2n). For a general Lie group G, the Stiefel manifold as base space is
replaced by O(m)/GxO(m d). There will be one bundle on M for each
homotopy class of the maps M O(m)/GxO(m d). The question of how
many twisted products there exist can thus be answered. Unfortunately, all
this is true only in principle, because the real calculation of all these classes
is usually a loathsome, in practice non realizable, task.
9.7.3 Characteristic classes These are members of the real cohomology
classes {H n (M )} of differential forms on the base space M . They answer to
the question concerning the degree of twistness of a given twisted product,
320
or how far it stands from the direct product. To obtain them, one first
finds them for the universal bundle, thereby estimating how far it is from
triviality; then one pull the forms back to the induced bundles; as the pullback preserves the cohomological properties, the same twisting measure
holds for the bundle of interest. Thus, the first task is to determine the
cohomology groups H n [O(m)/GxO(m d)]. For the most usual groups,
such classes are named after some illustrious mathematicians. For G = O(n)
and SO(n), they are the Pontryagin classes, though the double covering of
special groups produce an extra series of classes, the Euler classes. For U (n),
they are the Chern classes. These are the main ones. Other may be arrived
at by considering non-real cohomologies.
Part III
FINAL TOUCH
321
Chapter 10
NONCOMMUTATIVE GEOMETRY
Only a glimpse into promising new developments.
Everything said up to now concerns the geometry involved in classical

physics, commutative geometry. We shall here briefly broach the subject
of noncommutative geometry,1 and then proceed to some related aspects in
Quantum Mechanics.
10.1
QUANTUM GROUPS A PEDESTRIAN OUTLINE
10.1 Think of a group of usual matrices. Matrices in general do not commute. Their entries consist of real or complex numbers ai j . We are used to
multiplying matrices (that is, to perform the group operation) and, in doing
so, the entries themselves get multiplied according to the well-known rules.
Complex numbers commute with each other and we do not trouble with their
order when multiplying them: ai j am n = am n ai j .
10.2 Matrix groups are Lie groups, that is, smooth manifolds whose points
are the group elements. To each group element, a point on the group manifold, corresponds a matrix. Thus, each matrix will have its coordinates.
What are the coordinates of a matrix? Just the entries am n if they are real
and, if they are complex, their real and imaginary parts ( 6.4.3 and 6.5.8).
Thus, although themselves noncommuting, matrices are represented by sets
of commuting real numbers. Their non-commutativity is embodied in the
rules to obtain the entries of the product matrix from those of the matrices
being multiplied.
Suppose now that, in some access of fantasy, we take the very entries
i
a j as noncommutative. This happens to be not pure folly, provided some
rules are fixed to ensure a minimum of respectability to the new structure
so obtained. With convenient restrictions, such as associativity, the new
structure is just a Hopf algebra (Math.1.25).
1
Connes 1990.
323
324
CHAPTER 10. NONCOMMUTATIVE GEOMETRY
10.3 Very roughly speaking, quantum groups (the physicists unhappy

name for Hopf algebras)2 are certain sets of matrices whose elements ai j are
themselves non-commutative. They are not groups at all (hence the unhappiness of the name), but structures generalizing them. The non-commutativity
is parametrized in the form
Rrs mn am i an j = Rpq ij ar p as q ,
where the Rrs mn s are complex coefficients and the respectability conditions
are encoded in constraints on them. In particular, a direct calculation shows
that the imposition of associativity, starting from the above general commutativity assumptions, leads to the Yang-Baxter equation (Math.2.11; see also
Phys.10.5) in the form
Rjk ab Rib cr Rca mn = Rij ca Rck mb Rab nr.
Tecnically, these R-matrices, satisfying the Yang-Baxter equation, belong to
a particular kind of Hopf algebras, the so-called quasi-triangular algebras
(see Math.1.27).
10.4 Well, the Hopf structure would probably remain an important tool
of interest in limited sectors of higher algebra if it did not happen that the
Yang-Baxter equation turned up in a surprisingly large number of seemingly
disparate topics of physical concern: lattice models in Statistical Mechanics (about the relation with braids and knots, see Phys.3.2.3), integrability
of some differential equations,3 the inverse scattering method,4 the general
problem of quantization (see section 10.2), etc.
10.5 A first question comes immediately to the mind: if now the entries
themselves are no more commutative, what about the coordinate roles they
enjoyed? The coordinates themselves become noncommutative and that is
where noncommutative geometry comes to the scene. While the new matrices
of noncommutative entries constitute Hopf algebras, the entries themselves
constitute other spaces of mathematical predilection, von Neumann algebras
(Math.5.5).
2
Woronowicz uses the term pseudo-group. Concerning the name quantum groups,
the physicists are not alone in their guilt: people looking for material on Hopf algebras
in the older mathematical literature will have to look for annular groups ... The best
name is probably bialgebras.
3
McGuire 1964.
4
See different contributions in Yang & Ge 1989.
10.1. QUANTUM GROUPS A PEDESTRIAN OUTLINE
325
10.6 A first idea would be to take the above noncommuting matrix elements also as matrices. Matrices of matrices then turn up. This is where
the large use of direct-product matrices ( Math.2.10) comes out. In diagonal block matrices, the blocks work independently from each other and
will correspond to abelian bialgebras. Drop the diagonal character and you
will have general block matrices. By the way, finite von Neumann algebras
are precisely algebras of block matrices (well, actually almost everything a
physicist bothers about is a von Neumann algebra. The novelty is that up
to recent times most of these were commutative).
10.7 Groups are usually introduced in Physics as sets of transformations
on some carrier space. Would quantum groups also preside over transformations? If so, the above matrices with non-commuting entries would be
expected to act upon column vectors, these with non-commuting components. This is actually so,5 and the study of the carrier spaces (called quantum spaces, or Manin spaces) seems simpler than that of the Hopf algebras
themselves.6 The general case is, however, very involved also.7
10.8 There are other gates into the realm of bialgebras. Hopf introduced
them originally in homology theory, but other simpler cases have been found
since then. One may, for instance, proceed in a way similar to that used to
introduce the classical groups as transformation groups ( 8.1.9). These are
sets of transformations preserving given sesquilinear forms.8 This is probably
the most apealling approach for physicists. In knot theory, they appear in
presentations of groups and algebras. In physics, the original approach was
related to Lie algebras deformations.
Classical Sine-Gordon equation is related to the Lie algebra sl(2, R). Once
quantized, instead of this algebra, this integrable equation exhibitted another
structure, which was recognized by Kulish and Reshetikin as a deformation
of sl(2, R), and called it its quantum version. Drinfeld9 has then given
the new structure a general descripton, through the consideration of phase
spaces defined on Lie groups (see Phys.10).
10.9 A most interesting approach is that pioneered by Woronowicz,10
which we could call the Fourier gate. It relates to harmonic analysis on
groups. It goes from the Pontryagin duality for abelian groups, through
5
6
7
8
9
10
Manin 1989.
See for instance Fairlie, Fletcher & Zachos 1989, and references therein.
See Ocneanus postface to Enoch & Schwartz 1992.
Dubois-Violette & Launer 1990.
Drinfeld 1983.
Woronowicz 1987.
326
the Tanaka-Krein duality for non-abelian groups, to still more general theories (see Math.6.14 and those following it). The whole subject is very
involved in its formal aspects, and still a research subject for physicists and
mathematicians.11 It involves deep points of the theory of Banach algebras
(Math.5) and Hopf-Banach algebras. We retain here only the point that coordinates can become non-commutative and proceed to a heuristic discussion
of a formalism well known to physicists.
10.2
QUANTUM GEOMETRY
10.10 People may think that it is evident that Quantum Mechanics is

concerned with a noncommutative geometry. Actually, that the geometry is
noncommutative is not so evident. Despite the foresight of Dirac who, in his
basic paper,12 calls commutators quantum derivations, the well known noncommutativity in Quantum Mechanics is of algebraic, not geometric, character. The difference rests, of course, in the absence of specifically geometric
structures in the algebra of operators, such as differentiable structure, differential forms, connections, metrics and the like in a word, in the absence of
a differential geometry. On the other hand, some noncommutativity comes
up even in Classical Mechanics: the Poisson bracket is the central example.
But this is precisely the point the Poisson bracket is a highly strange
object.
10.11 Consider once again (see Phys.1) the classical phase space E2n of
some mechanical system with generalized coordinates q = (q 1 , q 2 , . . . , q n )
and momenta p = (p1 , p2 , . . . , pn ). Dynamical quantities F (q, p), G(q, p), etc
on E2n constitute an associative algebra with the usual pointwise product
like F G as operation. Given any associative algebra, one may get
a Lie algebra with the commutator as operation. Of course, due to the
commutativity, the classical Lie algebra of dynamical functions coming from
the pointwise product is trivial. It is the peculiar noncommutative Lie algebra
defined by the Poisson bracket which is physically significant. This is a
rather strange situation from the mathematical point of view, as the natural
algebraic brackets are those coming as commutators. The Poisson bracket
stands apart because it does not come from any evident associative algebra
of functions. We know, however, that a powerful geometric background, the
hamiltonian (or symplectic) structure, lies behind the Poisson bracket, giving
to its algebra a meaningful and deep content. On the other hand, in Quantum
11
12
An idea of the present state of affairs can be got in Gerstenhaber & Stasheff 1992.
Dirac 1926.
10.2. QUANTUM GEOMETRY
327
Mechanics, the product in the algebra of dynamical functions (the operators)

is noncommutative and the consequent commutator is significant but there
seems to exist no structure of the symplectic type. Now, it is a general belief
that the real mechanics of Nature is Quantum Mechanics, and that classical
structures must come out as survivals of those quantal characteristics which
are not completely erased in the semiclassical limit. It is consequently
amazing that precisely those quantal structures somehow leading to the
basic hamiltonian formalism of Classical Mechanics, mainly the symplectic
structure be poorly known. The answer to this problem, however, is
known. And it just requires the introduction of more geometric structure in
the quantum realm. The geometry coming forth is, however, of a new kind:
it is noncommutative.
10.12 In rough words, the usual lore of noncommutative geometry13 runs
as follows.14 Functions on a manifold M constitute an associative algebra
C(M ) with the pointwise product (Math.5.4). This algebra is full of content,
because it encodes the manifold topology and differentiable structure. It contains all the information about M . The differentiable structure of smooth
manifolds, for instance, has its counterpart in terms of the derivatives acting
on C(M ), the vector fields. On usual manifolds (point manifolds, as the phase
space above), this algebra is commutative. The procedure consists then in
going into that algebra and work out everything in it, but forgetting about
commutativity (while retaining associativity). In the phase space above, this
would mean that F G is transformed into some non-abelian product F G,
with a new operation.15 The resulting geometry of the underlying manifold M will thereby become noncommutative. Recall that a manifold is
essentially a space on which coordinates (an ordered set of real, commutative
point functions) can be defined. When we go to noncommutative manifolds,
the coordinates, like the other functions, become noncommutative. Differentials come up in a very simple way through the (then nontrivial) commutator
[F, G] = F G G F.
Associativity of the product F G implies the Jacobi identity for the commutator, i.e., the character of Lie algebra. With fixed F , the commutator is
13
Dubois-Violette 1991.
Coquereaux 1989.
15
In this chapter, the symbol is taken for the so-called star-product of quantum
Wigner functions. It has nothing to do with its previous use for the composition of
mapppings. It is the Fourier transform of the convolution, for which the symbol * is
used of old.
14
328
a derivative with respect to F : the Jacobi identity

[F, [G, H]] = [[F, G], H] + [G, [F, H]]
is just the Leibniz rule for the new product defined by [ , ].
10.13 In order to approach this question in Quantum Mechanics, the convenient formalism is the Weyl-Wigner picture. In that picture, quantum
operators are obtained from classical dynamical functions via the Weyl prescription, and the quantum formalism is expressed in terms of Wigner functions, which are c-number functions.
Let us only recall in general lines how the Weyl prescription16 works
for a single degree of freedom, the coordinate-momentum case (see also
Math.2.1). The Wigner functions AW (q, p) are written as Fourier transforms F [A] of certain Wigner densities A(a, b),17
ZZ
2
ei2(aq+ibp)/h A(a, b).
(10.1)
AW (q, p) = F [A] =
h
Then the Weyl operator A(q, p), a function of operators q and p corresponding to the Wigner function AW , is
ZZ
2
A(q, p) =
ei2(aq+ibp)/h A(a, b).
(10.2)
h
We may denote by Fb this operator Fourier transform, so that
A = Fb[F 1 [AW ]]
(10.3)
AW = F [Fb1 [A]].
(10.4)
and
The Wigner functions are, despite their c-number appearence, totally quantum objects. They are representatives of quantum quantities (they will include powers of ~, for example) and only become classical in the limit ~ 0,
when they give the corresponding classical quantities. They embody and
materialize the correspondence principle. The densities
A = F 1 [AW ] = Fb1 [A]
16
17
See for instance Galetti & Toledo Piza 1988.

See Baker 1958, and Agarwal & Wolf 1970.
(10.5)
10.2. QUANTUM GEOMETRY
329
include usually Dirac deltas and their derivatives. All this may seem rather
strange, that c-number functions can describe Quantum Mechanics. Actually, this is a lie: the Wigner functions are c-number functions indeed, but
they do not multiply each other by the usual pointwise product. In order to
keep faith to the above correspondence rule with quantum operators, a new
product has to be introduced, as announced in 10.12. This means that
we can describe quantum phenomena through functions, provided we change
the operation in their algebra. The product related to quantization is
called (rather unfortunately) the star-product.18 The simplest way to introduce it is by changing the usual multiplication rule of the Fourier basic
functions
(a,b) (q, p) = ei2(aq+ibp)/h .
(10.6)
We impose
(a,b) (q, p) (c,d) (q, p) = ei(adbc)h/4 (a+c,b+d) (q, p)
(10.7)
instead of the classical

(a,b) (q, p) (c,d) (q, p) = (a+c,b+d) (q, p).
(10.8)
The last expression says that the functions provide a basis for a representation of the group R2 , which is self-dual under Fourier transformations
( Math.6.13). Imposing equation [10.7], with the extra phase, means to
change R2 into the Heisenberg group ( 8.3.13). This is enough to establish
a complete correspondence between the functions and the quantum operators. The commutator of two functions, once the new product is defined, is
no longer trivial:
{AW , BW }M oyal (q, p) = [AW , BW ] (q, p)
Z
h
i 4
2 2
dxdydzdwA(x,y) B (z,w)
sin[ (yz xw)](x+z,y+w) (q, p).
=( )
h
h
4
(10.9)
This quantum bracket, called Moyal bracket after its discoverer, allows in
principle to look at quantum problems in a way analogous to hamiltonian
mechanics. Derivatives are introduced through the bracket: for instance,
{q, F }M oyal is the derivative of F with respect to the function q. A symplectic structure appears naturally,19 whose differences with respect to that
of classical mechanics (shown in Physical Topic 1) correspond exactly to the
18
19
Bayen, Flato, Fronsdal, Lichnerowicz & Sternheimer 1978.

Dubois-Violette, Kerner & Madore 1990.
330
quantum effects. Calculations are of course more involved than those with
the Poisson bracket. Due to the star product, for instance, the derivative of
p with respect to q is no more zero, but just the expected {q, p} = i~. If
we want to attribute coordinates to the quantum phase space, the only way
to keep sense is to use no longer the usual euclidean space with commutative numbers, but to consider coordinate functions with a product. In
this picture, we repeat, quantization preserves the dynamical functions, but
changes the product of their algebra.
10.14 And here comes the crux: in the limit ~ 0, the Moyal bracket
gives just the Poisson bracket. The strangeness of the Poisson bracket is
thus explained: though not of purely algebraic content, it is the limit of an
algebraic and nontrivial bracket coming from a non-commutative geometry.
The central algebraic operation of Classical Mechanics is thus inherited from
Quantum Mechanics, but this only is seen after the latter has been given
its full geometrical content. There is much to be done as yet, mainly when
we consider more general phase spaces, and a huge new field of research on
noncommutative geometrical structures is open.
All this is to say that Physics has still much to receive from Geometry.
It is an inspiring fact that an age-old structure of classical mechanics finds
its explanation in the new geometry.
We shall not cease from exploration
And the end of all our exploring
Will be to arrive where we started
And know the place for the first time.20
20
T. S. Eliot, Little Gidding .
Part IV
MATHEMATICAL TOPICS
331
Math.Topic 1
THE BASIC ALGEBRAIC STRUCTURES
GROUPS AND LESSER STRUCTURES
1
2
3
4
5
Definitions
Transformation groups
Representations
Groupoids, monoids, semigroups
Subgroups
RINGS AND FIELDS

6 Rings
7 Fields
8 Ring of a group
MODULES AND VECTOR SPACES
9 Modules
10 Vector spaces
11 The notion of action
12 Dimension
13 Dual space
14 Inner product
15 Endomorphisms and projectors
16 Tensor product
ALGEBRAS
17
18
19
20
21
22
23
Algebras
Kinds of algebras
Lie algebra
Enveloping algebra
Algebra of a group
Dual algebra
Derivation
COALGEBRAS
24 Coalgebras
25 Bialgebras, or Hopf algebras
26 R-matrices
Current mathematical literature takes for granted the notions given here, as
a minimum minimorum, the threshold of literacy. As a rule, we only retain
below those ideas which are necessary to give continuity to the presentation.
Some concepts are left to the Glossary.
333
334
1.1
MATH.TOPIC 1. THE BASIC ALGEBRAIC STRUCTURES

Groups and lesser structures
Groups are the most important of algebraic structures. Men enjoy their
presence in some special way, through the bias of symmetry, and learning
processes seem to make use of them. Since their unheeded introduction by
an ill-fated young man, its fortune has been unsurpassed in Physics.
1.1 Definitions Asking for the readers tolerance, let us only recall, to
keep the language at hand, that the cartesian set product U V of two sets
U and V is the set of all pairs (u, v) with u U and v V . A group is a set
point G on which is defined a binary operation
: G G G,
taking the cartesian set product of G by itself into G, with the four following
properties:
(a) for all g, g 0 G, the result g g 0 belongs to G;
(b) there exists in G an element e (the identity, or neutral element) such
that, for all g G,
e g = g e = g;
(c) to every g G corresponds an inverse element g 1 which is such that
g 1 g = g g 1 = e;
(d) the operation is associative: for all g, g 0 , g 00 in G,
(g g 0 ) g 00 = g (g 0 g 00 ).
The group (G, ) is commutative (or abelian) when g g 0 = g 0 g holds for all
g, g 0 G. The operation symbol is usually omited for multiplicative groups.
The center of G is the set of the gG which commute with all the elements
of G. The order of a group G, written |G|, is the number of elements of G.
Given a group G, if there is a setQof ai G such that every g G can be
written in the monomial form g = i (ai )ni for some set of exponents {ni },
we say that the set {ai } generates G, and call the ai s generators of G. In
the monomials, as the ai s are not necessarily commutative, there may be
repetitions of each ai in different positions. When the set {ai } is finite, G is
finitely generated . The number of generators of G is, in this case, called the
rank of G.
1.1. GROUPS AND LESSER STRUCTURES
335
1.2 Transformation groups As presented above, we have an abstract

group, which is a concern of Algebra (as a mathematical discipline). A
group formed by mappings from a set S on itself (that is, automorphisms) is
a transformation group. The bijective mappings of S constitute the largest
transformation group on S and the identity mapping constitutes the smallest.
Comment 1.1.1 The main interest of groups to Physics lies precisely on transformations preserving some important characteristic of a system. The word symmetry is
commonly reserved to transformations preserving the hamiltonian.
Comment 1.1.2 The set of all homeomorphisms of a topological space constitutes a

group and so does the set of all automorphic isometries (motions) of a metric space.
1.3 Representations A mapping h : G H of a group (G, ) into
another group (H, ) is a homomorphism if it preserves the group structure,
that is, if it satisfies
h(g g 0 ) = h(g) h(g 0 ).
The set ker h := { g G such that h(g) = idH } is the kernel of h and the
set im h := {h H such that some g G exists for which h = h(g)} is
the image of h.
A homomorphism of an abstract group into a transformation group is a
representation of the group. In common language, abstract groups are frequently identified to some important homomorphic transformation group, as
when we talk of the group SO(3) of real 3 3 special orthogonal matrices
as the group of rotations in the 3-dimensional euclidean space. A homomorphism taking every element of a group into its own identity element is
a trivializer . It leads to the so-called trivial representation. The theory of
group representations is a vast chapter of higher mathematics, of which a
very modest resume is given in Mathematical Topic 6.
1.4 Groupoids, monoids, semigroups1 When one or more of the four
requirements in the definition of a group (1.1) are not met, we have less
stringent structures. When only (b) and (c) hold, G is a groupoid. When
only (d) holds, G is a semigroup (some authors add condition (b) in this
case). When only (b) and (d) are satisfied, G is a monoid. A monoid is a
unital semigroup, a groupoid is a non-associative group, etc. As groups are
1
There are some fluctuations concerning the definitions of monoid and semigroup.
Some authors (such as Hilton & Wylie l967) include the existence of neutral element for
semigroups, others (Fraleigh 1974) only associativity. In consequence, the schemes here
shown, concerned with these structures, are author-dependent. We follow, for simplicity,
the glossary Maurer 1981.
336
more widely known, it may be simpler to get the other structures from them,
by eliminating some conditions, as in the diagram
drop
associativity
GROUPOID
not every element

has an inverse
MONOID
there is no
neutral element
SEMIGROUP
&
GROUP
A general scheme is
Set
G = {g,g 0 ,...}
add internal binary

operation
&
operation
is associative
SEMIGROUP
T HEN
IF
there exists
neutral element
MONOID
GROUPOID
ABELIAN
GROUP
.
GROUP
every element
has an inverse
operation
is commutative
Or we may indulge ourselves with some Veblen-like diagrams (Figure 1.1),

with the structures in the intersections of the regions representing the properties.
Comment 1.1.3 Amongst the lesser structures, the term semigroup is widely
used in physical literature to denote anything which falls short of being a group by failing
to satisfy some condition. For example, the time evolution operators for the diffusion
equation are defined only for positive time intervals and have no inverse. Also without
inverses are the members of the so called renormalization group of Statistical Mechanics
and Field Theory. But be careful. So strong is the force of established language, however
defective, that a physicist referring to the renormalization semigroup, or monoid, runs
the risk of being stoned to death by an outraged audience.
1.1. GROUPS AND LESSER STRUCTURES
337
Figure 1.1: The basic structures and their properties.
1.5 Subgroups A subset H of a group G forming a group by itself with

the same operation is a subgroup. The group G itself, as well as the group
formed solely by its identity element, are improper subgroups. Any other
subgroup is proper. A subgroup H of a group G is a normal subgroup (or
invariant subgroup) if ghg 1 H for all g G and h H. A group is
simple if it has no invariant proper subgroup. A group is semisimple if it
has no abelian invariant subgroup. Every simple group is, of course, also
semisimple.
If the order of G is finite, then the order of any subgroup H must divide
it (Lagrange theorem) and the index of H in G is |G|/|H|. If |G| is a finite
prime number, the Lagrange theorem will say that G has no proper subgroups
and, in particular, will be simple. Given a group G with elements a, b, c, etc,
the set of all elements of the form aba1 b1 constitutes a normal subgroup,
the commutator subgroup indicated [G, G]. The quotient G/[G, G] (the set
of elements of G which are not in [G, G]) is an abelian group, the abelianized
group of G. The classical example is given by the fundamental group 1 (M )
of a manifold M and the first homology group H1 (M ): the latter is the
abelianized group of the former.
Discrete groups are discussed in Mathematical Topic 2.
Comment 1.1.4 The homomorphism : G G/[G, G] taking a group G into
its abelianized subgroup is canonical (that is, independent of the choice of the original
generators) and receives the name of abelianizer .
338
Comment 1.1.5 Solvable and nilpotent groups Groups can be classified according to their degree of non-commutativity. Given two subsets A and B of a group G, we
define their commutator as [A, B] = { all elements in G of the form aba1 b1 , with a A
and b B}. We start with the abelian quotient G/[G, G]. We may then form two sequences of subgroups in the following way. Define G1 = [G, G], called the commutator
group of G; proceed to G2 = [G1 , G1 ] and so on to the n-th term of the series, Gn =
[Gn1 , Gn1 ]. Putting G = G0 , then
Gn1 Gn for all n. If it so happens that there exists an integer k such that Gn =
identity for all n k, the group G is solvable of class k. A second sequence is obtained by
defining G0 = G and Gn = [G, Gn1 ]. If Gn = identity for n k, the group G is nilpotent
of class k. Of course, both sequences are trivial for abelian groups. And the less trivial
they are, the farther G is from an abelian group.
1.2
Rings and fields
These are structures mixing up two internal operations.

1.6 Rings A ring < R, +, > is a set R on which two binary internal
operations, + and , are defined and satisfy the following axioms:
(i) < R, + > is an abelian group;
(ii) is associative;
(iii) both operations respect the distributive laws: for all a, b, c R,
a(b + c) = ab + ac and (a + b)c = ac + bc.
The multiplication symbol is frequently omitted: ab = ab.
Comment 1.2.1 When the operation is commutative, so is the ring. When a
multiplicative identity 1 exists, such that 1a = a1 = a for all a R, then < R, +, >
is a ring with unity (unital, or unit ring). In a unit ring, the multiplicative inverse (not
necessarily existent) to a R is an element a1 such that aa1 = a1 a = 1. If every
nonzero element of R has such a multiplicative inverse, then < R, +, > is a division
ring. The subset R0 of R is a subring if ab R0 when a R0 and b R0 . Let G be an
abelian group. The ring R is Ggraded if, as a group, R is a direct sum of groups R ,
for G, such that R R is contained in R+ . The frequent notation na means
a + a + . . . + a (n times).
Comment 1.2.2 Some examples:

(i) the set Z of integer numbers with the usual operations of addition and multiplication
is a commutative ring, though not a division ring;
(ii) the set Zn of integers modulo n is also a commutative (but not a division) ring,
formed from the cyclic group < Zn , + > with the multiplication modulo n ; it should
be denoted < Zn , +, >;
(iii) the set R(t) of polynomials in the variable t with coefficients in a ring R;
(iv) the set Mn [R] of n n matrices whose entries belong to a ring R is itself a ring
with unity if R is a ring with unity; it is not a division ring;
(v) the set of real or complex functions on a topological space S, with addition and
multiplication given by the pointwise addition and product: (f + g)(x) = f (x) + g(x) and
(f g)(x) = f (x) g(x).
1.2. RINGS AND FIELDS
339
1.7 Fields A field is a commutative division ring: to get at it starting

from an abelian group, we must follow both paths in the diagram
Abelian group with operation

+ and neutral 0
add another
operation x
impose distributive laws:

a x (b + c) = a x b + a x c
(a+b) x c = a x c+b x c
Nonassociative
Ring
every element, up to
Unit Ring
impose associativity
Division
Ring
there is neutral
element for x
0, has an inverse
by x00
by x
RING
FIELD
Commutative
Ring
if x is
commutative
The sets of real numbers (R) and complex numbers (C) constitute fields.
Familiarity with real and complex numbers makes fields the best known of
such structures. It is consequently more pedagogical to arrive at the concept
of ring by lifting progressively their properties:
FIELD
drop product
commutativity
Division ring
drop existence of
product inverse
RING
drop existence of
product neutral
Unit Ring
1.8 Ring of a group We can always linearize a group by passing into a

ring. There are two rings directly related to a given group. A first definition
is as follows. Take
P a ring R and consider the set of all formal summations
of the form a = gG a(g)g, where only a finite number of the coefficients
a(g) R are non-vanishing. Assume then addition and multiplication in the
340
natural way:
a+b=
a(g) g +
gG
b(g) g =
gG
ab =
[a(g) + b(g)] g;
gG
a(g) b(h) g h.
g,hG
The summations constitute a new ring, the group ring R(G) of G over R.
If R is a unital ring, each g G can be identified with that element a of
R(G) whose single coefficient is a(g) = 1. The group is thereby extended
to a ring. If G is non-abelian, so will be the ring. Now, to each a R(G)
corresponds anPR-valued function on G, fa : G R, such that fa (g) =
a(g) and a =
R will
gG fa (g)g.PConversely, to any function f : GP
correspond
a
ring
member
f
(g)g.
Notice
that,
given
a
=
gG
gG a(g)g
P
P
and b = hG b(h)h, the product ab = gH fab (g)g will be given by
X X
ab =
[
fa (h) fb (gh1 )]g.
g inH hH
To the product ab will thus correspond the convolution

X
fa (h) fb (gh1 ).
fab (g) =
hH
We arrive thus at another definition: given a group G, its group ring R(G)
over R is the set of mappings f : G R, with addition defined as
(f1 + f2 )(g) = f1 (g) + f2 (g),
and multiplication in the ring given by the convolution product
X
(f1 f2 )(g) =
f1 (h) f2 (h1 g).
hG
The condition concerning the finite number of coefficients in the first definition is necessary for the summations to be well-defined. Also the convolution
P
requires a good definition of the sum over all members of the group, h .
That is to say that it presupposes a measure on G. We shall briefly discuss
measures in Mathematical Topic 3. If G is a multiplicative group, R(G) is indicated by < R, +, > and is the ring of G over R. The restriction < R, >
contains G. If G is noncommutative, R(G) is a noncommutative ring. R(G)
is sometimes called the convolution ring of G, because another ring would
come out if the multiplication were given by the pointwise product,
(f1 f2 )(g) = f1 (g)f2 (g).
1.3. MODULES AND VECTOR SPACES
341
The analogy with Fourier analysis is not fortuitous (see Mathematical Topic
6). In order to track groups immersed in a ring R, we can use idempotents.
An idempotent (or projector ), as the name indicates, is an element p of the
ring such that p p = p. A group has exactly one idempotent element.
1.3
Modules and vector spaces
These are structures obtained by associating two of the above ones.

1.9 Modules An R-module is given by an abelian group M of elements
, , , . . . and a ring R = {a, b, c, . . . } with an operation of external multiplication of elements of M by elements of R satisfying the four axioms:
(i) a M ;
(ii) a ( + ) = a + a ;
(iii) (a + b) = a + b ;
(iv) (ab) = a (b ).
It is frequently denoted simply by M . The external product has been defined
in the order RM , with the element R at the left, so as to give things like a.
The above module is called a left-module. We can define right modules in
an analogous way. A bilateral module (bimodule) is a left- and right-module.
When we say simply module, we always mean actually a bimodule.
Modules generalize vector spaces, which are modules for which the ring is
a field. As vector spaces belong to the common lore, we might better grasp
modules by starting with a vector space and going backwardly through the
steps of Figure 1.2. But here we are trying to be constructive. We shall
recall their main characteristics in a slightly pedantic way while profiting to
introduce some convenient language.
1.10 Vector spaces A linear space (on the field F , for us the real or
complex numbers) is an abelian group V with the addition operation +,
on which is defined also an external operation of (scalar) multiplication
by the elements of F . For all vectors u, v V and numbers a, b F , the
following 5 conditions should be met:
(i) au V ;
(ii) a(bu) = (ab)u;
(iii) (a + b)u = au + bu ;
(iv) a(u + v) = au + av;
(v) 1u = u.
We say that V is a linear (or vector) space over F . The field F is a vector
space over itself. If the field F is replaced by a ring, we get back a module.
V is a real vector space if F = R. The name linear space is to be preferred
342
Figure 1.2: Modules and vector spaces.
to the more current one of vector space, but as this pious statement seems
useless, we use both of them indifferently.
1.11 The notion of action We repeat that the cartesian set product
U V of two sets U and V is the set of all pairs (u, v) with u U and v V .
A well-defined mapping U V V goes under the general name of action
of U on V . The above axioms for vector spaces define an action of the field
F on V .
Comment 1.3.1 Linear representations of algebras are actions on modules and the
classification of modules is intimately related to that of representations.
1.12 Dimension A family {vk } of vectors is said to be linearlyP

dependent
if there are scalars ak
F
,
not
all
equal
to
zero,
such
that
k ak v k =
P
0. If, on the contrary,
k ak vk = 0 implies that all ak = 0, the vk s are
linearly independent. This notion is extended to infinite families: an infinite
family is linearly independent if every one of its finite subfamilies is linearly
independent. The number of members of a family is arbitrary. The maximal
number of members of a linearly independent family is, however, fixed: it is
the dimension of V , indicated dim V . If {vk } is a family of linear independent
vectors, a vector subspace W , with dim W dim V , will be engendered by
1.3. MODULES AND VECTOR SPACES
343
a subfamily {wk } with dim W members. The set of vectors of V with zero
coefficients ak along the wk s will be the linear complement of W .
1.13 Dual space Vector spaces are essentially duplicated. This is so
because the space V * formed by linear mappings on a vector space is another
vector space, its dual. The image of v V by the linear mapping k V *
is indicated by < k, v >. When V is finite-dimensional, the dual of V is a
twin space, isomorphic to V . In the infinite-dimensional case, V is in general
only isomorphic to a subspace of V **. A mapping of V into its dual is an
involution. The isomorphism V V *, even in the finite-dimensional case, is
in general not canonical. This means that it depends of the basis chosen for
V . The presence of a norm induces a canonical isomorphism between V and
(at least part of) V *, and involutions are frequently defined as a mapping
V V (see below, 1.17). The action of the dual on V may define an inner
product V V F .
1.14 Inner product An inner product is defined as a mapping from the
cartesian set product V V into C, V V C, (v, u) < v, u > with the
following properties:
(a) < v1 + v2 , u > = < v1 , u > + < v2 , u >;
(b) < a v, u > = a < v, u >;
(c) < v, u > = < u, v > (so that < v, v > R );
(d) < v, v > 0;
(e) < v, v > = 0 v = 0.
The action of the dual V * on V defines an inner product if there exists
an isomorphism between V and V *, which is the case when V is of finite
dimension. An inner product defines a topology on V . Vector spaces endowed
with a topology will be examined in Mathematical Topics 4 and 5.
Comment 1.3.2 Infinite dimensional vector spaces differ deeply from finite dimensional ones. For example, closed bounded subsets are not compact neither in the norm
nor in the weak topology.
1.15 Endomorphisms and projectors An endomorphism (or linear
operator) on a vector space V is a mapping V V preserving its linear
structure. Projectors, or idempotents, are here particular endomorphisms p
satisfying p2 = p. Given a vector space E and a subspace P of E, it will be
possible to write any vector v of E as v = vp + vq , with vp P . The set Q of
vectors vq will be another subspace, the supplement of P in E. Projectors p
on P (which are such as p(vp ) = vp , Im p = P ) are in canonical (that is, basis
344
independent) one-to-one correspondence with the subspaces supplementary

to P in E: to the projector p corresponds the subspace Q = ker p.
1.16 Tensor product Let A and B be two vector spaces (on the same
field F ) with respective basis {xi } and {yj }. Consider the P
linear space C
with basis {xi yj }, formed by those sums of formal products m,n amn xm yn
defined by
P
P P
P
P
m,n amn xm yn =
n(
m amn xm )yn =
m xm (
n amn yn ),
when only a finite number of the coefficients amn F is different from zero.
We obtain in this way a novel vector space, the tensor product
P of A and B,
denoted
P AB. The alternative notations xi yj for xi yj and Pm,n amn xm yn
for m,n P
amn xm yn are both usual. Given
P two elements a = m am xm A
and b = n bn yn B, then a b = m,n am bn xm yn . The elements of
A B have the following three properties:
(i) (a + a0 ) b = a b + a0 b;
(ii) a (b + b0 ) = a b + a b0 ;
(iii) r (a + b) = (r a) b = a (rb),
for all a, a0 A; b, b0 B and r F . Conversely, these properties define C
in a basis-independent way.
Comment 1.3.3 Consider the space D of all the mixed bilinear mappings :A B
F , (a, b) (a, b). Then A B is dual to D. In the finite dimensional case, they are
canonically isomorphic if < a b, >= (a, b).
Comment 1.3.4 The product of a space by itself, A A, has special characteristics.
It
P may be divided into a symmetric part, given by sums of formal products of the form
m,n amn xm xn , with amn
P = anm , and an antisymmetric part, formed by those sums of
formal products of type m,n amn xm xn with amn = anm .
1.4
Algebras
The word algebra denotes of course one of the great chapters of Mathematics. But, just as the word topology, it also denotes a very specific
algebraic structure.
1.17 Algebras An algebra is a vector space A over a field F (for us,
R or C), on which is defined a binary operation (called multiplication) m :
A A A, (, ) m(, ) = such that, for all a F and , , A,
the following conditions hold:
(i) (a) = a() = (a);
(ii) ( + ) = + ;
1.4. ALGEBRAS
345
(iii) ( + ) = + .
This defines an action of the vector space on itself. Once so defined, A
is an algebra on F . When F = C, a mapping A A is an
involution, or adjoint operation, if it satisfies the postulates
(i) = ;
(ii) () = ;
(iii) (a + b) = a + b .
In that case is the adjoint of .
1.18 Kinds of algebras The algebra A is associative if further

(iv) () = () , , A.
This property can be rendered by saying that the diagram
AA
m id
&m
A.
%m
AAA
id m &
AA
is commutative. A module can be assimilated to a commutative algebra.
A is a unit algebra (or unital algebra) if it has a unit, that is, an element
e such that e = e = , A. Notice that each A defines a
mapping h :F A by h (a) = a . Of course, = h (1). The existence
of a unit can be stated as the existence of an element e such that he (a) =
a e = a, for all a F . This is the same as the commutativity of the diagram
AA
h id
F A
&
&id h
m
AF .
.
A
Homomorphisms of unital algebras take unit into unit. A unit-subalgebra

of A will contain the unit of A. We can put some of such structures in a
346
scheme:
VECTOR SPACE
add associative
multiplication
ALGEBRA
add inner product
add unit
INNER PRODUCT
SPACE
INVOLUTIVE
UNIT ALGEBRA
add involution
INVOLUTIVE
ALGEBRA
(1.1)
An element of an unit-algebra A is an invertible element if there exists in A
an element 1 such that 1 = 1 = e. The set G(A) of the invertible
elements of A is a group with the multiplication, the group of the algebra
A.
A graded algebra is a direct sum of vector spaces, A = k Ak , with the
binary operation taking Ai Aj Ai+j . If Ak , we say that k is the
degree (or order) of , and write = k.
1.19 Lie algebra An algebra is a Lie algebra if its multiplication (called
the Lie bracket) is anticommutative and satisfies the Jacobi identity
m(, m(, )) + m(, m(, )) + m(, m(, )) = 0.
Starting from any binary operation, the Lie bracket can always be defined
as the commutator [, ] = , and builds, from any associative
algebra A, a Lie algebra AL . If A is any algebra (not necessarily associative,
even merely a vector space), End A = {set of endomorphisms on A} is an
associative algebra. Then, the set [End A] of its commutators is a Lie algebra.
A vector basis 1 , 2 , . . . , n for the underlying vector space will be a basis
for the Lie algebra.
Comment 1.4.1 Lie algebras have a classification analogous to groups. They may be
solvable, nilpotent, simple, semisimple, etc, with definitions analogous to those given for
groups.
Comment 1.4.2 Drop in the algebra A the external multiplication by scalars of its
underlying vector space. What remains is a ring. As long as this external multiplication is
irrelevant, we can talk of a ring. The usual language is rather loose in this respect, though
most people seem to prefer talking about algebras.
1.4. ALGEBRAS
347
Comment 1.4.3 Modules may be obtained by applying a projector (here, an element

p of the algebra such that p p = p) to an algebra.
1.20 Enveloping algebra To every finite Lie algebra A will correspond a

certain unital associative algebra, denoted U and called the universal enveloping algebra of A. Given a basis {1 , 2 , . . . , n } for A, U will be generated
by the elements {11 , 22 , . . . , nn }, with j = 0, 1, 2, . . . n.
Comment 1.4.4 Let us quote a few amongst the main properties of U:
(a) U admits a unique anti-automorphism , called the principal anti-automorphism
of U, which is X = X for every X U;
(b) there exists a one-to-one homomorphism :U U U such that
(X) = X 1 + 1 X for all X U; it is called the diagonal mapping of U;
(c) each derivation of A (see 1.23 below) admits a unique extension to a derivation
of U;
(d) there exists a one-to-one correspondence between the representations of A and the
representations of U: every representation of A can be extended to a unique representation
of U, and the restriction to A of every representation of U defines a representation of A.
1.21 Algebra of a group The algebra of a group G comes out when the
group ring of G is a field, usually R or C. It is frequently called the group
convolution algebra, to distinguish it from the other algebra which would
come if the pointwise product were used.
Comment 1.4.5 This other algebra is the set of real or complex functions on the
group G, with addition and multiplication given by the pointwise addition and product:
(f + g)(x) = f (x) + g(x) and (f g)(x) = f (x) g(x). It is sometimes also called the group
algebra.
1.22 Dual algebra On algebras, the involution is required to submit to
some conditions, by which it transfers the algebraic properties into the dual
space, which thereby becomes a dual algebra.
A norm, for example that coming from an inner product, can define a
topology on a linear space. Addition of a topology, through an inner product
or by other means, turns linear spaces into much more complex objects.
These will be left to the Mathematical Topic 5.
1.23 Derivation The generic name derivation is given to any endomorphism D:A A for which Leibnizs rule holds: D() = (D) + (D).
The Lie algebra [End A] contains D(A), the vector subspace of all the derivations of A, and the Lie bracket makes of D(A) a Lie subalgebra of [End
A], called its derived algebra. This means that the commutator of two
derivations is a derivation. Each member A defines an endomorphism
ad(a) = ada , called the adjoint action of a, by
ada (b) = [a, b] b.
348
Comment 1.4.6 The set of differentiable real or complex functions on a differentiable

manifold M constitutes an algebra. The vector fields (derivations on M ) are derivations
of this algebra, which consequently reflects the smooth structure of the space itself.
Comment 1.4.7 If M is a Lie group, consequently endowed with additional structure,

the algebra of functions will gain extra properties reflecting that fact.
1.5
Coalgebras
1.24 General case Suppose now that spaces A and B are algebras over
a certain field, say C. Then A B is also an algebra (the tensor product of
algebras A and B) with the product defined by (a b)(a0 b0 ) = (a a0 ) (b b0 ).
When A and B are associative unit algebras, so will be their tensor product.
The product in A is in reality a mapping m : A A A, a a0 a a0 .
Its dual mapping : A A A is called the coproduct, or comultiplication
(or still diagonal mapping). It is supposed to be associative. The analogue
to associativity (the coassociativity) of the comultiplication would be the
property
(id ) (x) = ( id) (x)
as homomorphisms of A in A A A:
AA
&id
AAA.
%id
A
&
AA
Once endowed with this additional structure, A is a coalgebra. The coalgebra is commutative if (A) is included in the symmetric part (see Comment
1.3.4 above) of A A. Let us put it in other words: define a permutation
map : A A A A,
(x y = y x;
then the coalgebra is commutative if = .
1.25 Bialgebras, or Hopf algebras An associative unit algebra A is a
Hopf algebra (or bialgebra, or an old name annular group) if it is a
coalgebra satisfying
(i) the product is a homomorphism of unit coalgebras;
(ii) the coproduct is a homomorphism of unit algebras,
(xy) = (x)(y).
1.5. COALGEBRAS
349
The general form is

(x) = I x + x I +
xj yj ,
with xj , yj A. From (ii), I = I I. When (x) = I x + x I, x is

said to be primitive.
Let us present two more mappings:
1. A map : A C defining the counit of the coproduct , which is
given by
( id)(x) = (id )(x) = x,
AA
&id
xA.
%id
xA
&
AA
It is an algebra homomorphism:
(xy) = (x)(y).
2. The antipodal map : A A, an antihomomorphism (xy) =
(y)(x) such that
m(id ) (x) = m( id)(x) = (x)I,
which is described in the diagram
id
A A
AA
&m
xA
(x) I A
&
%m
A A
id
AA
The map (x) is called the antipode (or co-inverse) of x. Given the
permutation map (x y) = y x, then 0 = is another coproduct
on A, whose antipode is 0 = 1 .
Comment 1.5.1 Some people call bialgebras structures as the above up to the existence of the counit, and reserve the name Hopf algebras to those having further the
antipode.
350
Comment 1.5.2 Write x = (1 x, 2 x); then both id)x = (1 x)2 x and

id )x = (2 x)1 x should be x, so that
1 x =
1
(2 x) x
and 2 x =
1
(1 x) x.
Furthermore, x(x) = (x)(1 x)(2 x)I.
1.26 Hopf algebras appear in the study of products of representations of

unital algebras. A representation of the algebra A (see Mathematical Topic
6) will be given on a linear space V by a linear homomorphic mapping of
A into the space of linear operators on V . The necessity of a Hopf algebra
comes out when we try to compose representations, as we do with angular
momenta. We take two representations (1 , V1 ) and (2 , V2 ) and ask for a
representation fixed by 1 and 2 , on the product V1 V2 . In order to keep
up with the requirements of linearity and homomorphism, it is unavoidable
to add an extra mapping, the coproduct . Once this is well established,
the product representation will be = (1 2 ) .
The universal enveloping algebra of any Lie algebra has a natural structure of Hopf algebra with the diagonal mapping (see Comment 1.4.4, item
b) as coproduct. But also algebras of functions on groups may lead to such
a structure. Particular kinds of Hopf algebras, called quasi-triangular, are
known to physicists under the name of quantum groups.
1.27 R-matrices We use the direct-product notation of 2.10 of Mathematical Topic 2. The Hopf algebra is a quasi-triangular algebra if further:
(i) and 0 = are related by conjugation:
(x) = R (x) R1
for some matrix R A A. This means that, for a commutative Hopf
algebra, R = I I;
(ii) (id )(R) = R13 R12 ;
(iii) ( id)(R) = R13 R23 ;
(iv) ( id)(R) = R1 .
Then the Yang-Baxter equation of Mathematical Topic 2, 2.11 follows:
R12 R13 R23 = R23 R13 R12 .
Fraleigh 1974
Kirillov 1974
Warner 1972
Majid 1990
Bratelli & Robinson 1979
Math.Topic 2
DISCRETE GROUPS. BRAIDS AND KNOTS
A DISCRETE GROUPS
1
2
3
4
Words and free groups

Presentations
Cyclic groups
The group of permutations
B BRAIDS
5 Geometrical braids
6 Braid groups
7 Braids in everyday life
8 Braids presented
9 Braid statistics
10 Direct product representations
11 The Yang-Baxter equation
C KNOTS AND LINKS
12
13
14
15
16
2.1
Knots
Links
Knot groups
Links and braids
Invariant polynomials
A Discrete groups
These are the original groups, called into being by Galois. Physicists became
so infatuated with Lie groups that it is necessary to say what we mean by
discrete groups: those which are not continuous, on which the topology is
either undefined or the discrete topology. They can be of finite order (like
351
352
MATH.TOPIC 2. DISCRETE GROUPS. BRAIDS AND KNOTS
the group of symmetries of a crystal, or that of permutations of members

of a finite set) or of infinite order (like the braid groups). In comparison
with continuous groups, their theory is very difficult: additional structures
as topology and differentiability tend to provide extra information and make
things easier. As a consequence, whereas Cartan had been able to classify
the simple Lie groups at his time, the general classification of finite simple
groups has only recently (1980) been terminated.
Comment 2.1.1 As in all highly sophisticated subjects, the landscape here is full of
wonders (results such as: any group of 5 or less elements is abelian; or the order of a
simple group is either even or prime; or still any two groups of order 3 are isomorphic
and there are, up to isomorphisms, only two groups of order 4: the so called Klein 4-group
and the cyclic group Z4 ) and amazements (like the existence and order of the Monster
group).
Practically all the cases of discrete groups we shall meet here are fundamental groups of some topological spaces. These are always found in terms
of some generators and relations between them. Let us say a few words on
this way of treating discrete groups, taking for simplicity a finite rank.
2.1 Words and free groups
Consider a set A of n elements, A = {a1 , a2 , . . . , an }. We shall use the
names letters for the elements aj and alphabet for A itself. An animal with
p times the letter aj will be written apj and will be called a syllable. A finite
string of syllables, with eventual repetitions, is (of course) a word . Notice
that there is no reason to commute letters: changing their orders lead to
different words. The empty word 1 has no syllables.
There are two types of transformations acting on words, called elementary
contractions. They correspond to the usual manipulations of exponents: by
a contraction of first type, a symbol like api aqi becomes ap+q
; by a second type
i
contraction, a symbol like a0j is replaced by the empty word 1, or simply
dropped from the word. With these contractions, each word can be reduced
to its simplest expression, the reduced word. The set F [A] of all the reduced
words of the alphabet A can be made into a group: the product u v of two
words u and v is just the reduced form of the juxtaposition uv. It is possible
to show that this operation is associative and ascribes an inverse to every
reduced word. The resulting group F [A] is the word group generated by the
alphabet A. Each letter ak is a generator .
Words may look at first as too abstract objets. They are actually extremely useful. Besides obvious applications in Linguistics and decoding,1
the word groups are largely used in Mathematics, and have found at least
1
Schreider 1975.
2.1. A DISCRETE GROUPS
353
one surprising application in Geometry: they classify the 2-dimensional manifolds.2 In Physics, they are used without explicit mention in elementary
Quantum Mechanics. Recall the Weyl prescription3 (the correspondence
rule) to obtain the quantum operator Weyl(pm q n ) = W(pm q n ) corresponding to a classical dynamical quantity like pm q n :
m n
W(p q ) =
1
2n
n
X
n
k=0
m nk
q p q
1
2m
m
X
m
k=0
pk qn pmk
(2.1)
where bold-faced letters represent operators. For the first few cases, W(pq)
= 21 (pq + qp), W(pq 2 ) = 13 (pq2 + qpq + q2 p), etc. The quantum operator
corresponding to a polynomial pm qn in the classical degree of freedom q
and its conjugate momentum p is the (normalized) sum of all the words
one can obtain with m times the letter p and n times the letter q.
Now, given a general group G, it will be a free group if it has a set
A = {a1 , a2 , . . . , an } of generators such that G is isomorphic to the word
group F [A]. In this case, the aj are the free generators of G. The number of
letters, which is the rank of G, may eventually be infinite.
The importance of free groups comes from the following theorem:
Every group G is a homomorphic image of some free group F [A].
This means that a mapping f : F [A] G exists, preserving the group operation. In a homomorphism, in general, something is lost: many elements
in F [A] may be taken into a same element of G. F [A] is in general too rich.
Something else must be done in order to obtain an isomorphism. As a rule, a
large F [A] is taken and the freedom of its generators is narrowed by some
relationships between them.
2.2 Presentations
Consider a subset {rj } of F [A]. We build the minimum normal subgroup
R with the rj as generators. The quotient F [A]/R will be a subgroup, corresponding to putting all the rj = 1. An isomorphism of G onto F [A]/R
will be a presentation of G. The set A is the set of generators and each rj
is a relator . Each r R is a consequence of {rj }. Each equation rj = 1 is a
relation. Now, another version of the theorem of the previous section is:
Every group G is isomorphic to some quotient group of a free group.
2
3
Doubrovine, Novikov & Fomenko 1979, vol. III.

Weyl 1932.
354
Comment 2.1.2 In this way, groups are introduced by giving generators and relations
between them. Free groups have for discrete groups a role analogous to that of coordinate
systems for surfaces: these are given, in a larger space, by the coordinates and relations
between them. Of course, such coordinates being non-commutative, things are much
more complicated than with usual coordinates and equations seldom lead to the elimination
of variables. Related to this point, there is a difficulty with presentations: the same group
can have many of them, and it is difficult to know whether or not two presentations refer
to the same group.
2.3 Cyclic groups The simplest discrete groups are the cyclic groups,
which are one-letter groups. A group G is a cyclic group if there is an element
a such that any other element (including the identity) may be obtained as
ak for some k. It is of order n if the identity is an . Good examples are the
n-th roots of 1 in the complex plane. They form a group isomorphic to the
set {0, 1, 2, 3, . . . , n 1} of integers with the operation of addition modulo n.
This is the cyclic group Zn . There is consequently one such group Zn for each
integer n = |Zn |. The simplest case is Z2 , which can be alternatively seen as
a multiplicative group of generator a = 1: Z2 = {1, - 1; the operation is
the usual multiplication}. Every cyclic group is abelian. Every subgroup of
a cyclic group is a cyclic group. Any two cyclic groups of the the same finite
order are isomorphic. Thus, the groups Zn classify all cyclic groups and for
this reason Zn is frequently identified as the cyclic group of order n. Any
infinite cyclic group is isomorphic to the group Z of integers under addition.
Given a group G and an element a G, then the cyclic subgroup of G
generated by a, hai = {an : n Z} is the smallest subgroup of G which
contains a. If hai = G, then a generates G entirely, and G is itself a cyclic
group.
Comment 2.1.3 Consider an element g G. If an integer n exists such that gn = e,
then n is the order of the element g, and g belongs to a cyclic subgroup. When no such
integer exists, g is said to be of infinite order. If every element of G is of finite order, G is
a torsion group. G is torsion-free if only its identity is of finite order. In an abelian group
G, the set T of all elements of finite order is a subgroup of G, the torsion subgroup of G.
2.4 The group of permutations

Let A be an alphabet, A = {a1 , a2 , . . . , an }. A permutation of A is a
one-to-one function of A onto A (a bijection A A). The usual notation
for a fixed permutation in which each aj goes into some apj is

a1 a2 an1 an
(2.2)
ap1 ap2 apn1 apn
The set of all permutations of an n-letter alphabet A constitutes a group
under the operation of composition (product), the n-th symmetric group,
denoted Sn . The order of Sn is (n!).
2.1. A DISCRETE GROUPS
355
Comment 2.1.4 The expression a permutation group is used for any (proper or
improper) subgroup of a symmetric group. This is very important because every finite
group is isomorphic to some permutation group (Cayleys theorem).
The permutation of the type

a1 a2 aj1 aj
a2 a3 aj a1
is a cycle of length j, usually denoted simply (a1 , a2 , . . . , aj ). A product of two
cycles is not necessarily a cycle. A product of disjoint cycles is commutative.
A cycle of length 2 is a transposition. Example: (a1 , a5 ). Any permutation
of Sn is a product of disjoint cycles. Any cycle is a product of transpositions,
(a1 , a2 , . . . , an ) = (a1 , a2 )(a2 , a3 )(a3 , a4 ) . . . (an , a1 ). Thus, any permutation
of Sn is a product of transpositions.
Given a permutation s, the number of transpositions of which s is a
product is either always even or always odd. The permutation s itself is, accordingly, called even or odd. The number of even permutations in Sn equals
the number of odd permutations (and equals n!/2). The even permutations
of Sn constitute a normal subgroup, the alternating group An .
The symmetric group can be introduced through a presentation. Define
as generators the (n 1) elementary transpositions s1 , s2 . . . , sn1 such that
si exchanges only the i-th and the (i + 1)-th entry:

si =
1 2
i
i + 1 n 1 n
1 2 i + 1
i
n 1 n

(2.3)
Each permutation will be a word with the alphabet {sj }. The si s obey
the relations
sj sj+1 sj = sj+1 sj sj+1
si sj = sj sj for |i j| > 1
(si )2 = 1,
(2.4)
(2.5)
(2.6)
which determine completely the symmetric group Sn : any group with generators satisfying these relations is isomorphic to Sn .
Comment 2.1.5 Many groups are only attained through presentations. This is frequently the case with fundamental groups of spaces. A good question is the following:
given two presentations, can we know whether or not they present the same group?
This is a version of the so-called word problem. It was shown by P. S. Novikov that
there can exist no general procedure to answer this question.
356
Suppose that in the permutation s there are n1 1-cycles, n2 2-cycles, etc.

The cycle type of a permutation is given by the numbers (n1 , n2 , . . .). Different permutations can be of the same cycle type, with the same set {nj }.
The importance of the cycle type comes from the following property: permutations of the same cycle type go into each other under the adjoint action of
any element of Sn : they constitute conjugate classes. Repeating: to each set
{nj } corresponds a conjugate class of Sn . We can attribute a variable tr to
each cycle of length r and indicate the cycle structure of a permutation by
the monomial tn1 1 tn2 2 tn3 3 . . . tnr r . Then, to all permutations of a fixed class will
be attributed the same monomial above. Such monomials are invariants under the action of the group Sn . The total number of permutations with such
a fixed cycle configuration is Qn n!nj !j nj . The n-variable generating function
j=1
for these numbers is the so-called cycle indicator polynomial4

Cn (t1 , t2 , t3 , . . . tn ) =
n!
tn1 tn2 tn3 . . . tnr r
nj 1 2 3
n
!j
j
j=1
Qn
{nj }
(2.7)
The summation
takes place over the sets {ni } of non-negative integers for
P
which ni=1 i ni = n. Of course, such a summation of invariant objects is
itself invariant. This is an example of a very important way of characterizing
discrete groups: by invariant polynomials. Though not the case here, it is
sometimes easier to find the polynomials than to explicit the group itself.
This happens for example for the knot groups (see Math.2.16 below). The
above invariant polynomial for the symmetric group appears as partition
functions in Statistical Mechanics of systems of identical particles,5 which
are invariant under particle permutations (Phys.3).
2.2
B Braids
2.5 Geometrical braids

A braid may be seen6 as a family of non-intersecting curves (1 , 2 , . . . n )
on the cartesian product E2 I with
j (0) = (Pj , 0)
for
j (1) = (P(j) , 1) for
4
j = 1, 2, . . . , n,
j = 1, 2, . . . , n,
Comtet 1974.
For a description of the symmetric and braid groups leading to braid statistics, see
Aldrovandi 1992.
6
Doubrovine, Novikov & Fomenko 1979, vol. II.
5
2.2. B BRAIDS
357
where is an index permutation (Figure 2.1). A braid is tame when its

curves are differentiable, i.e., have continuous first-order derivatives (are of
class C 1 ). Otherwise, it is said to be wild.
Figure 2.1:
2.6 Braid groups

Braids constitute groups, which were first studied by Artin. There are
many possible approaches to these groups. We shall first look at them as
the fundamental groups of certain spaces. Consider n distinct particles on
the euclidean plane M = E2 . Their configuration space will be M n = E2n
= {x = (x1 , x2 , . . . , xn )}, the n-th Cartesian product of manifold M . Suppose further that the particles are impenetrable, so that two of them cannot
occupy the same position in E2 . To take this into account, define the set
Dn = {(x1 , x2 , . . . , xn ) such that xi = xj for some i, j} and consider its
complement in M n ,
Fn M = M n \Dn .
(2.8)
Then the pure braid group Pn is the fundamental group of this configuration space: Pn = 1 [Fn M ]. If the particles are identical, indistinguishable,
the configuration space is still reduced: two points bfx and bfx are equivalent if (x1 , x2 , . . . , xn ) and (x01 , x02 , . . . , x0n ) differ only by a permutation, a
transformation belonging to the symmetric group Sn . Let Bn M be the space
obtained by identification of all equivalent points, the quotient by Sn :
Bn M = [Fn M ]/Sn .
(2.9)
Then 1 [Bn M ] is the full braid group Bn , or simply braid group. Artins
braid group is the full braid group for M = E2 , but the above formal definition allows generalization to braid groups on any manifold M . Of course,
Fn M is just the configuration space for a gas of n impenetrable particles,
358
and Bn M is the configuration space for a gas of n impenetrable and identical

particles. Consequently, quantization of a system of n indistinguishable particles must start from such highly complicated, multiply-connected space7
As such a quantization employs just the fundamental group of the configuration space,8 it must be involved with braid groups and, of course, statistical
mechanics will follow suit.
2.7 Braids in everyday life
In reality, braid groups9 concern real, usual braids. They count among the
simplest examples of experimental groups: we can easily build their elements
in practice, multiply them, invert them. They are related to the (still in
progress) study of general weaving patterns, which also includes knots and
links. Figure 2.2 depicts some simple braids of 3 strands: take two copies of
the plane E2 with 3 chosen, distinguished points; link distinguished points
of the two copies in any way with strings; you will have a braid. Figure 2.2(a)
shows the trivial 3-braid, with no interlacing of strands at all.
Figures 2.2(b) and 2.2(d) show the basic, elementary steps of weaving,
two of the simplest nontrivial braids. By historical convention, the strings
are to be considered as going from top to bottom. Notice that in the drawing,
the plane E2 is represented by a line just for facility. In 2(b), the line going
from 2 to 1 goes down behind that from 1 to 2. Just the opposite occurs
in 2.2(c). Braids 2.2(b) and 2.2(c) are different because they are thought to
be drawn between two planes, so that the extra dimension needed to make
strings go behind or before each other is available. Braids are multiplied by
composition: given two braids A and B, A B is obtained by drawing B
below A. Figure 2.3 shows the product of 2.2(b) by itself. Figure Math2Fig4
shows 2.2(b) 2.2(d).
The trivial braid 2.2(a) is the neutral element: it changes nothing when
multiplied by any braid. It is easily verified that 2.2(b) and 2.2(c) are inverse
to each other. The product is clearly non-commutative (compare 2.2(b)
2.2(d) and 2.2(d) 2.2(b)). In reality, any braid of 3 strands may be obtained
by successive multiplications of the elementary braids 2.2(b) and 2.2(d) and
their inverses. Such elementary braids are consequently said to generate the
3rd braid group which is, by the way, usually denoted B3 . The procedure of
7
Leinaas & Myrheim 1977: a very good discussion of the configuration spaces of
identical particles systems. Wavefunctions for bosons and fermions are found without
resource to the summations of wavefunctions of distinguishable particles usually found in
textbooks. Braid groups, although not given that name, are clearly at play.
8
Schulman l968; Laidlaw & DeWitt-Morette 1971; DeWitt-Morette 1972; DeWittMorette, Masheshwari & Nelson 1979.
9
Birman 1975: the standard mathematical reference.
2.2. B BRAIDS
359
Figure 2.2:
Figure 2.3:
building by products from elementary braids may be used indefinitely. The

braid group is consequently of infinite order. Of course, each braid may be
seen as a mapping E2 E2 , and 2.2(a) is the identity map.
Comment 2.2.1 All this can be easily generalized to the n-th braid group Bn , whose
elements are braids with n strands. The reader is encouraged to proceed to real experiments with a few strings to get the feeling of it.
A basic point is the following: consider the Figure 2.3. Each point on
it is, ultimately, sent into itself. It would appear that it corresponds to the
identity, but that is not the case! Identity is 2(a) and Figure 2.3 cannot be
reduced to it by any continuous change of point positions on E2 . It cannot be
unwoven! A short experiment shows that it would be possible to disentangle
it if the space were E3 . As every braid is a composition of the elementary
braids, that would mean that any braid on E3 may be unbraided ... as
360
Figure 2.4:
witnessed by millennia of practice with hair braids. Hair braids on E2 can be

simulated by somehow gluing together their extremities, thereby eliminating
one degree of freedom. Because braids can be unwoven in E3 , the braid group
reduces to the symmetric group and Quantum and Statistical Mechanics in
E3 remain what they are usually. Differences could however appear in the
2-dimensional case. Anyhow, from the point of view of maps on E2 , we
see that the identity exhibits infinite possibilities! Each particle sees the
others as forbidden points, as holes. Repeated multiplication by braid 2(b)
2(b) will lead to paths starting at 1 and turning 2, 3, ... times around
a hole representing 2. Of course, all this strongly suggests a relation to
the fundamental group of E2 with holes. It is indeed through this relation
that mathematicians approach braid groups, as seen below. The modified
identity of Figure 2.3 would be simply twice the transposition of points
1 and 2. More generally, any permutation of points becomes multiform:
the n-th braid group is an enlargement of the group of permutations Sn .
Mathematicians have several definitions for Bn , the above configuration
space definition allowing, as said, generalization to braid groups on any
manifold M .
2.8 Braids presented
The braid group Bn has also (n 1) generators j satisfying the relations
j j+1 j = j+1 j j+1
i j = j j for |i j| > 1
(2.10)
(2.11)
They are the same as the two relations [2.4] and [2.5] for Sn . The absence of
condition [2.6], however, makes of Bn a quite different group. It suffices to
say that, while Sn is of finite order, Bn is infinite.
2.9 Braid statistics
2.2. B BRAIDS
361
The absence of the relation [2.6] has, as we have said, deep consequences.
Unlike the elementary exchanges of the symmetric group, the square of an
elementary braid is not the identity. In many important applications, however, it is found that j2 differs from the identity in a well-defined way. In
the simplest case, j2 can be expressed in terms of the identity and j , which
means that it satisfies a second order equation like (j x)(j y) = 0,
where x and y are numbers. In this case, the j s belong to a subalgebra
of the braid group algebra, called Hecke algebra. This is the origin of the
so-called skein relations, which are helpful in the calculation of the invariant
polynomials of knot theory.
In Quantum Mechanics, a basis for a representation of a braid group will
be given by operators U (j ) acting on wavefunctions according to 0 (x) =
U (j )(x) = ei (x). But now there is no constraint enforcing U (j2 ) = 1, so
that U 2 (j )(x) = U (j2 )(x) = ei2 (x), U (j3 )(x) = ei3 (x), etc. The
representation is now, like the group, infinite. It is from the condition U (j2 )
= 1 that the possibilities of phase values for the usual n-particle wavefunctions are reduced to two: as twice the same permutation leads to the same
state, U (j2 )(x) = (x) so that ei = 1. The two signs correspond to
wave-functions which are symmetric and antisymmetric under exchange of
particles, that is, to bosons and fermions. When statistics is governed by the
braid groups, as is the case for two-dimensional configuration spaces of impenetrable particles, the phase ei remains arbitrary and there is a different
statistics for each value of . Such statistics are called braid statistics.
2.10 Direct product representations Representations of the braid
groups can be obtained with the use of direct products of matrix algebras.
Suppose the direct product of two matrices A and B. By definition, the
matrix elements of their direct product A B are
< ij|A B|mn > = < i|A|m >< j|B|n > .
(2.12)
On the same token, the direct product of 3 matrices is given by

< ijk|A B C|mnr > = < i|A|m >< j|B|n >< k|C|r > .
(2.13)
And so on. The direct product notation compactifies expressions in the

following way. Let T = A B, and E be the identity matrix. Then we write
T12 = A B E,
T13 = A E B,
T23 = E A B, etc.
A useful property of direct products is
(2.14)
(2.15)
(2.16)
362
(A B C)(G H J) = (AG) (BH) (CJ),

and analogously for higher order products. We may also use the notation
b an expression like
T ij mn = < ij|T |mn >. Given a matrix R,
bkj ab R
bbi cr R
bac mn = R
bji ca R
bkc R
bba nr
R
mb
(2.17)
b12 R
b23 R
b12 = R
b23 R
b12 R
b23 ,
R
(2.18)
is equivalent to
which is the braid equation, name usually given to [2.10]. To show it, look
at s1 , s2 as s1 = S12 and s2 = S23 , S being some direct product as above.
Then find < ijk|s1 s2 s1 |mnr > = S ij pq S qk vr S pv mn and < kji|s2 s1 s2 |mnr >
= S ji qs S kq mv S vs nr , so that the braid equation is
S kj ab S bi cr S ac mn = S ji ca S kc mb S ba nr .
We have found above conditions for representations of B3 . Higher order
direct products of projectors will produce representations for higher order
b Aut(V V ) satisfying
braid groups. In the general case, given a matrix R
relations as above and the identity E Aut(V ), a representation of BN on
V N is obtained with generators
bi,i+1 . . . E E = (E)i1 R
bi,i+1 (E)N i . (2.19)
i = E E E . . . R
2.11 The Yang-Baxter equation With the notation above, we may
easily establish a direct connection of the braid relations to the Yang-Baxter
equation, usually written
R12 R13 R23 = R23 R13 R12 ,
(2.20)
Rjk ab Rib cr Rca mn = Rij ca Rck mb Rab nr .
(2.21)
which is the same as
b = P R, R
bij mn =
Define now another product matrix by the permutation R
ji
R mn . The above expressions are then equivalent to
b12 R
b23 R
b12 = R
b23 R
b12 R
b23 ,
R
(2.22)
just the braid equation. The permutation relation is thus a very interesting
tool to obtain representations of the braid groups from Yang-Baxter solutions
and vice-versa. Notice that, due to this equivalence, many people give the
name Yang-Baxter equation to the braid equation. An important point is
that Yang-Baxter equations come out naturally from the representations of
2.3. C KNOTS AND LINKS
363
the Lie algebra of any Lie group. Thus, each such Lie algebra representation
will provide a solution for the braid relations.10
The relation between this matrix formulation and our first informal representation of braids by their plane drawings leads to an instructive matrixdiagrammatic formulation. It is enough to notice the relationship
a /b
c/ d
ab cd
R
and proceed to algebrize diagrams by replacing concatenation by matrix multiplication, paying due attention to the contracted indices. Looking at Figure
2.5, we see that the braid equation [2.10], becomes exactly the Yang-Baxter
equation in its form [2.17].
Figure 2.5: Rab ij Rjc kf Rik de = Rbc ij Rai dk Rkj ef
2.3
C Knots and links
The classification of knots has deserved a lot of attention from physicists like
Tait and Kelvin at the end of the last century, when it was thought that the
disposition of the elements in the periodic table might be related to some
kind of knotting in the ether. Motivated by the belief in the possibility of a
fundamental role to be played by weaving patterns in the background of physical reality, they have been the pioneers in the (rather empirical) elaboration
of tables11 of distinct knots. Nowadays the most practical classification of
knots and links is obtained via invariant polynomials. Braids constitute
10
11
See Jimbos contribution in Yang & Ge 1989.

See for instance Rolfsen 1976.
364
highly intuitive groups of a more immediate physical interest, and there is a

powerful theorem relating braid and knots.
2.12 Knots12
Consider the two knots in Figure 2.6. As anyone can find by experience
with a string, they are non-equivalent. This means that we cannot obtain
Figure 2.6:
one from the other without somehow tying or untying, that is, passing one of
the ends through some loop. The mathematical formalization of this intuitive
notion of different knots is the so called knot problem and leads to an
involved theory. The characterization is completely given by the notion of
knot- (or link-) type, which is as sound as unpractical. Actually, there is no
practical way to establish the distinction of every two given knots. There are
however two methods allowing an imperfect solution of the problem. One
of them attributes to every given knot (or link) a certain group, the other
attributes a polynomial. They are imperfect because two different knots may
have the same polynomials or group. The characterization by the knot groups
is stronger: two knots with the same group have the same polynomials, but
not vice-versa. On the other hand, polynomials are easier to find out.
We must somehow ensure the stability of the knot, and we do it by
eliminating the possibility of untying. We can either extend the ends to
infinity or simply connect them. We shall choose the latter, obtaining the
closed versions drawn more symmetrically as in the Figure 2.7. The example
to the right is equivalent to the circle, which is the trivial knot.
12
For an introductory, intuitive view, see Neuwirth 1979. For a recent qualitative
appraisal, see Birman 1991; an involved treatment, but with a very readable first
chapter is Atiyah 1991.
365
Now, the formal definition: a knot is any 1-dimensional subspace of E3

which is homeomorphic to the circle S 1 . Notice that, as spaces, all knots
are topologically equivalent. How to characterize the difference between the
above knots, and between knots in general? The answer comes from noticing
that tying and untying are performed in E3 , and the equivalence or not is
a consequence of the way in which the circle is plunged in E3 . Two knots
A and B are equivalent when there exists a continuous deformation of E3
into itself which takes A into B. This definition establishes an equivalence
relation, whose classes are called knot-types. The trivial knot, equivalent to
the circle itself, is called the unknot. The trefoil and the four-knot overleaf
are of different and non-trivial types. We shall see below (Math.2.14) how
to define certain groups characterizing knot-types.13
Figure 2.7:
2.13 Links
Links are intertwined knots, consequently defined as spaces homeomorphic to the disjoint union of circles. The left example of Figure 2.8 shows
a false link, whose component knots are actually independent. Such links
are also called unknots. The center and right examples of Figure 2.8 show
two of the simplest collectors favorites. Of course, knots are particular onecomponent links, so that we may use links to denote the general case.
We have above talked loosely of continuous deformation of the host
space taking one knot into another. Let us make the idea more precise. The
first step is the following: two knots A and B are equivalent when there exists
a homeomorphism of E3 into itself which takes A into B. This is fine, but
13
Crowell & Fox 1963; Doubrovine, Novikov & Fomenko 1979, vol. II.
366
Figure 2.8:
there is better. We would like to put arrows along the lines, to give knots
an orientation. The equivalence just introduced would not take into account
different orientations of the knots. A more involved notion will allow A and
B to be equal only if also their orientations coincide once A is deformed
into B. An isotopy (or isotopic deformation) of a topological space M is a
family of homeomorphisms ht of M into itself, parametrized by t [0, 1],
and such that (i) h0 (p) = p for all p M , and (ii) the function ht (p) is
continuous in both p and t. Isotopies provide the finest notion of continuous orientation-preserving deformations of M into itself. They constitute a
special kind of homotopy, in which each member of the one-parameter family
of deformations is invertible. When M is the host space of links, this definition establishes an equivalence relation, whose classes are called link-types.
Link-types provide the complete characterization of links, but it has a serious drawback: given a link, it is a very difficult task to perform isotopies
to verify whether or not it may be taken into another given link. That is
why the experts content themselves with incomplete characterizations, such
as the link group and the invariant polynomials.
2.14 Knot groups
Knots, as defined above, are all homeomorphic to the circle and consequently topologically equivalent as 1-dimensional spaces. We have seen that
knot theory in reality is not concerned with such 1-dimensional spaces themselves, but with how E3 englobes these deformed circles. Given the knot
K, consider the complement E3 \K. The knot group of K is the fundamental
group of this complement, 1 (E3 \K). It is almost evident that the group of
the trivial knot is Z. Simple experiments with a rope will convince the reader
that such groups may be very complicated. The trefoil group is the second
braid group B2 . As already said, knot groups do not completely characterize
367
knot types: two inequivalent knots may have the same group.
2.15 Links and braids
The relation between links and braids is given by Alexanders theorem,
which requires a preliminary notion. Given a braid, we can obtain its closure
simply by identifying corresponding initial and end points. Experiments with
pieces of rope will once again be helpful. For instance, with the sole generator
1 of the two-strand-group B2 we can build the Hopf link and the trefoil: they
are respectively the braids 12 and 13 when their corresponding ends meet.
Alexanders theorem says that
to every link-type corresponds a closed braid,
provided both braid and link are tame .
corresponds to a link K, we
Given the braid whose closure (denoted )
write = K. This means that a link-type may be represented by a word

in the generators of some braid group Bn . Experiments also show that this
correspondence is not one-to-one: many braids may correspond to a given
link-type. Thus, we obtain knots and links (their closures) if we connect
corresponding points of a braid.
2.16 Invariant polynomials14
The relation between links and braids is the main gate to the most practical characterizations of links, the invariant polynomials. Great progress has
been made on this altogether fascinating subject in the last years. The idea
of somehow fixing invariance properties through polynomials in dummy variables is an old one. Already Poincare used polynomials, nowadays named
after him, as a shorthand to describe the cohomological properties of Lie
groups.15 For a group G, such polynomials are
pG (t) = b0 + b1 t + b2 t2 + . . . + bn tn ,
where the bk s are the Betti numbers of G. They are, of course, invariant
under homeomorphisms. Notice that each bk is the dimension of a certain
space, that of the harmonic k-forms on G. Or, if we prefer, of the spaces
of cohomology equivalence classes. We have said in Math.2.4 that the cycle
indicators [2.7] are invariant polynomials of the symmetric group, and that
the coefficients of each monomial is the number of elements of the respective
cycle configuration, or conjugate class.
14
Commendable collections of papers on the subject are Yang & Ge 1989 and Kohno
1990.
15
Goldberg 1962.
368
Invariant polynomials are a characterization of knots, which is weaker

than the knot group: knots with distinct groups may have the same polynomial. But they are much easier to compute. And it may happen that a
new polynomial be able to distinguish knots hitherto undifferentiated. Actually, only recently, thanks to Conway, it became really easy to find some of
them, because of his discovery of skein relations. There are at present a few
different families of polynomials. To the oldest family belong the Alexander
polynomials, found in the thirties. The way to their computation was rather
involved before a skein relation was found for them. Skein relations, which
are different for each family of polynomials, provide an inductive way to obtain the polynomial of a given link from those of simpler ones. Suppose three
links that only differ from each other at one crossing.
/
\
There are three types of crossing: /\ , its inverse /\ , and the identity, or
uncrossing )(.
The polynomial of a knot K is indicated by a bracket < K >. If a
knot K 0 differs from K only in one crossing, then their polynomials differ
by the polynomial of a third knot in which the crossing is abolished. There
are numerical factors in the relation, written in terms of the variable of the
polynomial. Instead of drawing the entire knot inside the bracket, only that
crossing which is different is indicated. For example, the Alexander polynomials of K and K 0 are related by
\
< /\ >A < /\ >A +
t1
(2.23)
< )( >A = 0,
the index A indicating Alexander. This relation says that the j s are
in a Hecke algebra: it is a graphic version of
t1
j1 j + I = 0.
t
(2.24)
The skein relation must be supplemented by a general rule < HL > =

< H >< L > if H and L are unconnected parts of HL, and by a normalization of the bubble (the polynomial of the unknot), which is different for each
family of polynomials. For the Alexander polynomial, < O > = 1. A skein
relation relates polynomials of different links, but is not in general enough for
a full computation. It will be interesting for the knowledge of < K > only if
< K 0 > is better known. Kauffman extended the previous weaving patterns
by introducing the so-called monoid diagrams, including objects like . If
we add then convenient relations like
/
< /\ > = t1/2 < )( > (t1/2 + t1/2 ) <

\
S
T
< /\ > = t1/2 < )( > (t1/2 + t1/2 ) <
S
T
>,
>,
(2.25)
(2.26)
369
we can go down and down to simpler and simpler links, and at the end only
the identity and simple blobs O remain.
The animal represents a projector. Kauffmans decomposition is justified by Jones discovery of representations of the braid group in some special
von Neumann algebras, which are generated by projectors. Jones has thereby
found other polynomials, and also clarified the meaning of the skein relations.
The cycle indicator polynomial appears as the partition function of a system
of identical particles (Phys.3.1.2). Jones polynomials appear as the partition
function of a lattice model (Math.5.6).
Adams 1994
Birman 1991
Crowell & Fox 1963
Fraleigh 1974
Kauffman 1991
Yang & Ge 1989
Neuwirth 1965
370
Math.Topic 3
SETS AND MEASURES
MEASURE SPACES
1
2
3
4
5
6
7
The algebra of subsets

Measurable space
Borel algebra
Measure and probability
Partition of identity
Riemannian metric
Measure and Integration
ERGODISM
8 Types of flow
9 The ergodic problem
3.1
MEASURE SPACES
3.1 The algebra of subsets

A family of subsets of S is a topology if it includes S itself, the empty
set , all unions of subsets and all intersections of a finite number of them.
We shall here describe collections of subsets of another kind, profiting in the
while to introduce some notation and algebraic terminology. Given two subsets A and B of a set S,
A B = A\B = difference of A and B = {p A such that p
/ B}
A B = union of A and B = {p A or p B}
A B = symmetric difference of A and B = (A\B) (B\A) .
3.2 Measurable space
Suppose that a family R of subsets is such that
371
372
MATH.TOPIC 3. SETS AND MEASURES
(i) it contains the difference of every pair of its members;

(ii) it contains the union of every pair of its members.
In this case it will contain also the empty set , and all finite unions
and intersections. More than that, a first algebraic structure emerges. The
operation is a binary internal operation, taking a pair (A, B) of subsets
into another subset, A B. A pair such as (A, B) belongs to twice R,
that is, to the cartesian set product R R of R by itself. The notation is
(A, B) R R. An internal binary operation such as is indicated by
:RRR
(A, B) A B .
With this operation, R constitutes an abelian group. The neutral element is
and each subset is its own inverse. Other binary internal operations are of
course present, such as the difference \ and the intersection
: (A, B) A B = A\(A\B).
The latter is associative, A (B C) = (A B) C. The relationship of
and is distributive:
A (BC) = (A B) (A C);
(A B) C = (A C) (B C) .
The scheme is the same as that of the integer numbers, with for addition
and for multiplication. Such a structure, involving two binary internal
operations obeying the distributive laws, one constituting an abelian group
and the other being associative, is a ring (Math.3). A family R of subsets
as above will be a ring of subsets of S. The power set of any S is a ring of
subsets. Suppose now also that
(iii) S R.
S will work as a unit element for the multiplication : A S = S A =
A. In this case the whole structure is a ring with unity. In the present case,
R is more widely known as the Boolean algebra. Because of the historical
prestige attached to the last name, the ring R is called an algebra of subsets,
and indicated by A. Let us make one more assumption:
(iv) R contains all the countable unions of its members.
A family satisfying (i) (iv) is called a -algebra (sometimes also a field). As seen in chapter 1, a topology is essential to a clear and proper
definition of the notion of continuity. A -algebra is the minimum structure
required for the construction of measure and probability theories. The pair
(S, A) formed by a space S and a particular -algebra is for this reason called
a measurable space.
3.1. MEASURE SPACES
373
3.3 Borel algebra

It is possible, and frequently desirable, to make topology and measure
compatible with each other. This is done as follows. Suppose some family C
of subsets of S is given which does not satisfy (i) (iv). It is then possible to
show that there exists a smallest -algebra A(C) of S including C, and that
it is unique. A(C) is said to be the -algebra generated by C. Consider now
a topology T defined on S. The family T is, as in the case above, such that
there will be a smallest -algebra A(T ) generated by T . This is the Borel
-algebra, and every one of its members is a Borel set. The open intervals of
E1 generate a Borel -algebra. If T = indiscrete topology, little will remain
of it in this procedure.
Comment 3.1.1 Besides topologies and -algebras, there are other dissections of a
set, each one convenient for a certain purpose. For example, filters and ultrafilters, which
are instrumental in the study of continuity and convergence in non-metric topological
spaces.
3.4 Measure and probability

Given a set S and a family A = {Ai } of its subsets, a real set function
is given by f : A R, f (Ai ) = some real number. We shall suppose that
A contains the empty set and the finite unions of its members, and define
a positive set function as a mapping m : A R+ . Suppose further that,
for every finite collection of disjoint sets {Ai A, i = 1, 2, . . . , n}, the two
following conditions hold:
P
S
(a) m ( ni=1 Ai ) = ni=1 m(Ai );
(b) m() = 0.
The function m is then said to be finitely additive. If the conditions
hold even when n is infinite, m is countably additive. A positive measure is
precisely such a countably additive set function on S, with the further proviso
that A be a -algebra on S. The sets Ai A are the measurable subsets of
S and, for each set Ai , m(Ai ) is the measure of Ai . The whole structure
is denoted (S, A, m) and is called a measure space. If S is a countable union
of subsets Ai with each m(Ai ) finite, the measure m is -finite. Given
any set algebra on S, it generates a -algebra, and any positive set function
m is extended into a positive measure. If m is -finite, this extension is
unique (Hahn extension theorem). The measure m is finite if m(S) is finite.
A probability space is a (S, A, m) with m(S) = 1. In this case each set
Ai A is an event and m(Ai ) is the probability of Ai . On locally compact
topological spaces we may choose the closed compact subsets as Borel sets.
A positive measure on a locally compact Hausdorff space is a Borel measure.
A good example is the Lebesgue measure on E1 : the Borel -algebra is that
374
generated by the open intervals (a, b) with b a and the measure function
is m[(a, b)] = b a. The Lebesgue measure extends easily to En .
3.5 Partition of identity
Consider a closed subset U of a differentiable manifold M . Then, there is a
theorem which says that there exists a smooth function fU (the characteristic
function of U ) such that fU (p) = 1 for all p U , and fU (p) = 0 for all p
/ U.
Suppose further that M is paracompact. This means that M is Hausdorff
and each covering has a locally finite sub-covering. Given a smooth atlas,
there will be a locally finite coordinate covering {Uk }. Then, another theorem
says that a family {fk } of smooth functions exists such that
(i) the support of fk Uk ;
(ii) 0 fk (p) 1 for all p M ;
P
(iii) k fk (p) = 1 for all p M .
The family {fk } is a partition of the identity . The existence of a
partition of the identity can be used to extend a general local property to
the whole space, as in the important examples below.
3.6 Riemannian metric
Once assured that a partition of the identity exists, we may show that a
differentiable manifold has always a Riemannian metric (6.6.10). As each
coordinate neighbourhood is euclidean, we may define on each Uk the eu(k)
clidean metric g = . A Riemannian metric on M will then be given
by
X
(k)
g (p) =
fk (p)g
(p).
k
3.7 Measure and Integration

On the same token, as we know how to integrate over each euclidean Uk ,
the integral over M of any m-form is defined as
Z
XZ
=
(k) fk (p),
M
Uk
where (k) is the coordinate form of on Uk .
Kolmogorov & Fomin 1977

Choquet-Bruhat, DeWitt-Morette & Dillard-Bleick 1977
3.2. ERGODISM
3.2
375
ERGODISM
Well known examples of probability spaces are found in Classical Statistical

Mechanics (Phys.3). Each one of the statistical ensembles uses a different
Borel measure F (q, p) and provides a different relationship between microscopic and macroscopic quantities. The Lebesgue measure
dqdp = dq 1 dq 2 . . . dq n dp1 dp2 . . . dpn
R
gives the volume of a domain U in phase space M as U dqdp, that is,
F (q, p) = 1. By the Liouville theorem, this volume is preserved by the microscopic dynamics. Systems in equilibrium are described by time-independent
Borel measures. In this case we usually
R write d = F (q, p)dqdp for the
measure and the measure of U , (U ) = U F (q, p)dqdp, is constant in time.
When this happens for any U , we say that the microscopic hamiltonian flow
is measure-preserving. The expected value of a macroscopic quantity A is
R
R
< A > = M a(q, p)F (q, p)dqdp = M a(q, p)d,
where a(q, p) is the corresponding microscopic quantity. Notice however that
the only thing which is warranted to be preserved is the measure of any
volume element. There is no information on anything else. This subject
evolved into a sophisticated theory involving contributions from every chapter
of Mathematics, the Ergodic Theory.
3.8 Types of flow
A particularly important question is the following: what is the flow of a
volume element U in phase space M ? There are three qualitatively different
possibilities:
(i) Non-ergodic flow: U moves without distortion and returns to its initial
position after some finite interval of time; the total flow of U covers a small
region of M . Consider a point p on M , and think of U as the initial uncertainty on its position; then the position at any time is perfectly determined,
as well as the error, which remains U .
(ii) Ergodic flow: the shape of U is only slightly changed during the flow
but the system never comes back to its initial configuration; the total flow of
U sweeps a large region of M , possibly the whole of it; the points originally
in U become a dense subset of M . If there is an initial error U in the
position of the point p, then, after some time, p can be at any point of M .
Previsibility is lost. This situation of overall sweeping of phase space by an
initially small domain is generically called ergodicity.
(iii) Mixing flow: the shape of U is totally distorted; the distance between two initially neighbouring points diverges exponentially in time, d(t)
376
eat d(0). The coefficient a in the exponent is a much used characterization

of chaoticity, the Lyapunov exponent. Because of the underlying deterministic dynamics, this case is frequently referred to as deterministic chaos.
Mixing implies ergodicity, but the converse is not true. Now, Sina has shown
the billiard theorem: a system of N balls in a box of hard walls is a mixing
system. Even such a simple system as 2 balls enclosed in a box is, thus, very
complicated.
3.9 The ergodic problem
All these considerations stand behind the famous ergodic problem. Suppose an isolated system with fixed energy E. Such a system is described by
the microcanonical ensemble (Phys.3) and the representative point travels
on the hypersurface H(q, p) = E of phase space. There are of course the integrals of motion, which reduce this hypersurface to a smaller subspace. We
consider the average behaviour of the representative point on this reduced
phase space. Any macroscopic observation of the system will last for a time
interval T large in comparison with the microscopic times involved. Thus,
what is really observed is a time-average over the microscopic processes,
something like
Z T
1
dt a[q(t), p(t)].
(3.1)
a
T =
2T T
However, a basic notion of Statistical Mechanics is that the value of a macroscopic quantity is obtained as an ensemble average, that is,
Z
a(q, p) d.
(3.2)
<A>=
M
Boltzmanns ergodic theorem says that this expectancy (average on phase

space) equals the time average for large intervals of time: if you call
a
= lim a
T
(3.3)
<A>=a
.
(3.4)
then
The interval T is supposed to be large not only with respect to the times
involved in the detailed microscopic processes (like scattering times), but also
as compared with those times relevant for the establishment of equilibrium
(relaxation time, free flight between the walls, etc).
The ergodic problem is summarized in the question: is the ergodic theorem valid? Or, which is the same, can we replace one average by the other?
If the answer is positive, we can replace statistics by a dynamical average.
3.2. ERGODISM
377
Roughly speaking, the answer is that the theorem is true provided the measure on phase space has a certain property, so that statistics is actually never
eliminated. To give an idea of the kind of questions involved, we shall briefly
describe the basic results.
A first point refers to the very existence of the limit in [3.3]. A second
point is concerned with the independence of (
a on the particular flow (the
particular hamiltonian).
Concerning the limit, there are two main points of view, depending on
the type of convergence assumed. One is that of Birkhoff, the other that of
von Neumann. Suppose a finite-volume subset S of phase space. Then, we
have two different theorems:
(a) Birkhoff s theorem: if the dynamical function f (q, p) on S is such
that
R
f [q(0), p(0)]d < ,
S
then
1
limT { 2T
RT
T
dtf [q(t), p(t)]}
exists for all points (q, p) of S and is independent of the chosen origin of
time. This limit will be identified with f, but notice that this is a particular
definition, assuming a particular type of convergence.
(b) von Neumanns theorem: consider the Hilbert space of squareintegrable dynamical functions on S. The inner product
R
(f, g) = S f (q, p)g(q, p)dS
defines a norm ||f ||. Then there exists a function f such that
limT ||f f|| = 0.
If f is simultaneously integrable and square-integrable, Birkhoffs and von
Neumanns limits coincide over S, except for functions defined on sets of zero
measures. Of course, everything here holds only up to such sets.
This seems to settle the question of the limit, though it should be noticed
that other topologies on the function spaces could be considered. Equation
[3.4] is valid for both cases above, provided an additional hypothesis concerning the measure d is assumed. In simple words, the measure d should not
divide the phase space into non-communicating sub-domains. Phase space
must not be decomposed into flow-invariant sub-regions.
More precisely: a space is metric-indecomposable (or metrically transitive) if it cannot be separated into two (or more) regions whose measures
378
are invariant under the dynamical flow and different from 0 or 1. The condition for the ergodic theorem to be true is that the phase space be metrically
transitive. This means that there is no tendency for a point to abide in a
sub-domain of phase space, or that no trajectory remains confined to some
sub-region. In particular, there must be no hidden symmetries. It is in general very difficult to know whether or not this is the case for a given physical
system, or even for realistic models.
Balescu 1975
Arnold 1976
Jancel 1969
Mackey 1978
Math.Topic 4
TOPOLOGICAL LINEAR SPACES
1 Inner product space

2 Norm
3 Normed vector spaces
4 Hilbert space
5 Banach space
6 Topological vector spaces
7 Function spaces
Adding topologies to vector spaces leads to more sophisticated algebraic
structures. For infinite dimensional manifolds, the resulting topological linear
spaces play the role analogous to that of euclidean spaces for finite dimensional manifolds, both as purveyors of coordinates and, in the differentiable
cases, as tangent spaces. In general, topology is defined on a vector space
through a norm. Let us begin with a particular case.
4.1
Inner product space
A linear space endowed with an inner product is an inner product space.

Given the inner product V V C, (v, u) < v, u >, the number
||v|| = < v, v >

is the norm of v induced by the inner product. This is a special norm,
as general norms will be defined as in next section, independently of inner
products. Some consequences, valid for this particular case, are:
1) the Cauchy-Schwarz inequality: | < v, u > | ||v|| ||u||;
2) the triangular inequality, or sub-additivity: ||v + u|| ||v|| + ||u||;
3) the parallelogram rule: ||v + u||2 + ||v u||2 = 2 ||v||2 + 2||u||2 .
Let us add some further concepts. Two members u and v of a linear
space (that is, of course, two vectors) are orthogonal, indicated uv, if <
v, u >= 0. For them will hold the Pythagoras theorem:
379
380
MATH.TOPIC 4. TOPOLOGICAL LINEAR SPACES
uv
4.2
||v + u||2 = ||v||2 + ||u||2 .
Norm
A norm on a linear space V over the field C is a mapping V R, v ||v||,

the following conditions holding for all v, u V , and C:
(i) ||v + u|| ||v|| + ||u||;
(ii) ||v|| = || ||v|| ;
(iii) ||v|| 0;
(iv) ||v|| = 0 v = 0.
It will be a seminorm if only the sub-additivity property (i) and condition
(ii) hold.
4.3
Normed vector spaces
Once endowed with a norm, V will be a normed vector space. Internal

product spaces are special cases, as we have seen that an internal product
defines a norm. Norm is however a more general concept, as there are norms
which are not induced by an internal product. The parallelogram law is a
consequence of an internal product and does not necessarily hold for a general
norm. A norm is a distance function, and defines the norm topology (also
called the strong topology, and sometimes uniform topology). Normed spaces
are metric topological spaces.
On normed vector spaces, the linear structure allows the introduction
of one further concept: let V be such a space and a, b two of its points.
Define the straight line between a and b by the curve f : I V, f (t) =
(1 t)a + tb. A subset C of V is convex if, for every pair a, b C, all
the points f (t) also lie on C. The whole V is always convex, and so is also
every vector subspace of V . Convex sets are sometimes called starshaped
sets. Closed differential forms are always exact in a convex domain.
4.4
Hilbert space
A Hilbert space is an inner product space which is complete under the

inner product norm topology. The standard case has for point set the set of
1
2
Helmerg 1969.
Halmos 1957.
4.4. HILBERT SPACE
381
sequences v = (v1 , v2 , v3 , . . .) = {vi }

i=1 = {vi }iN of complex numbers such
that
P
2
i=1 |{vi }i=1 | < .
The inner product is defined as
< v, u >=
i=1
vi ui .
A basis is prescribed by the open balls with the distance function

d(v, u) = | < v, u > |.
Hilbert spaces generalize euclidean spaces to the infinite dimensional case.
They can be shown to be connected. It is possible to establish a one-toone correspondence between the above sequences and bounded functions,
by which such spaces become function spaces. The spaces of wavefunctions
describing negative-energy states in Quantum Mechanics are Hilbert spaces
of this kind. Positive-energy states constitute spaces far more complicated
and are sometimes called Dirac spaces by physicists. Let us now consider
sequences of vectors (in Hilbert space, sequences of the above sequences).
The sequence of vectors {vn = (vn1 , vn2 , vn3 , . . .)} is an orthogonal sequence
if vn vm = 0 for all pairs of distinct members, and is orthonormal if further
it is true that ||vn || = 1 for each member. In these cases we talk of an
orthogonal system for the linear space. A theorem says that an orthogonal
family of non-zero vectors is linearly independent. The Hilbert space H,
defined as above, contains a countably infinite orthogonal family of vectors.
Furthermore, this family is dense in H, so that H is separable. In this
case, consider a vector u. The number < u, vn > is the n-th coordinate,
or the n-th Fourier coefficient 3 of u with respect to the system {vm }. An
example of separable Hilbert space is the following: consider the complexvalued functions on the interval [a, b] R. Then the space L2 of all absolutely
square integrable functions is a separable Hilbert space:
Rb
H = L2 = {f on [a, b] with a |f (x)|2 dx < }.
In greater generality, we may consider also non-separable Hilbert spaces.
These would come out if, in the definition given above, instead of v = {vi }iN ,
we had v = {v }R : the family is not indexed by a natural number, but by a
number in the continuum. This definition would accommodate Dirac spaces.
The energy eigenvalues, for the discrete or the continuum spectra, are precisely the indexes labeling the family elements, wavefunctions or kets. There
3
Dieudonne 1960.
382
are nevertheless
Pnew problems in this continuum-label case: the convergent
summations i=1 used in the very definition of Hilbert space become integrals. In order to define integrals over a set, one needs additional structures:
those of a -algebra of subsets, and that of a measure (see Math.3). Such
Hilbert spaces will depend also on the choice of these structures.
4.5
Banach space
We have seen that Hilbert space is an inner product space which is complete
under the inner product norm topology. More general, a Banach space is a
normed vector space which is complete under the norm topology. Thus, each
one of its Cauchy sequences in the norm topology is convergent.
4.6
Topological vector spaces
A Banach space is a topological vector space when both the addition operation and the scalar multiplication are continuous in the norm topology.
Although these rather abstract concepts hold in finite-dimensional spaces,
they are actually fundamental in the study of infinite-dimensional spaces,
which have quite distinct characteristics.
A general scheme is shown in Figure 4.1. Normed spaces have metric
topologies. If they are also complete, they are Banach spaces. On the other
hand, the norm may come from an inner product, or not. When it does
and furthermore the space is complete, it is a Hilbert space. If the linear
operations (addition and scalar multiplication) are continuous in the norm
topology (inner product or not), we have topological vector spaces.
Comment 4.6.1 The word metrizable, when applied to such spaces, means that its
topology is given by a translation-invariant metric.
Let us recall that the dual space to a given linear space V is that linear
space (usually denoted V ) formed by all the linear mappings from V into
its own field. When V is finite-dimensional, V is related to V by an isomorphism which in general is not canonical, but V is canonically isomorphic
to V . In the infinite-dimensional case, V is in general only isomorphic to a
subspace of V . The image of v V by k V is indicated by < k, v >.
On a topological vector space V , another topology is defined through the
action of the V . It is called the weak topology and may be defined through
convergence: a sequence {vn } converges weakly to v V if, for every k V ,
< k, vn > < k, v > as n . As the names indicate, the norm topology is finer than the weak topology: a sequence may converge weakly and
4.7. FUNCTION SPACES
383
Figure 4.1:
not converge in the norm topology. There are many other possible topologies, in effect. Figure 4.2 shows a scheme of linear spaces and some of their
topologies.
A very useful notion is the following: a subset U of a topological vector
space is a bounded set if it obeys the following condition of archimedean
type: for any neighbourhood V of the origin there exists a number n > 0
such that nV U .
4.7
Function spaces
Consider the space C (M, C) of differentiable complex functions on a manifold M . It is a vector space to start with. Define on this space an internal
operation of multiplication C (M, C) C (M, C) C (M, C). To help
the mind, we may take the simplest multiplication, the pointwise product,
defined by (f g)(x) = f (x)g(x). Then C (M, C) becomes an algebra. Actually, it is as associative, commutative *-algebra (described in Math.5). We
may in principle introduce other kinds of multiplication and obtain other
algebras within the same function space.
As hinted in the discussion on the Hilbert space, there are very important
cases in which the norm involves a measure. They combine in this way the
above ideas with those given in Math.3, and the resulting structure is far more
384
Figure 4.2:
complicated. For this reason we leave Banach and *-algebras to Math.5. Differentiability on infinite-dimensional manifolds, which supposes topological
vector spaces to provide a tangent structure, is discussed in Math.7.
Math.Topic 5
BANACH ALGEBRAS
1 Quantization
2 Banach algebras
3 *-algebras and C*-algebras
4 From Geometry to Algebra
5 von Neumann algebras
6 The Jones polynomials
The wealth of vector spaces endowed with a topology has been seen in
Math.4. We give now a sketchy account of what happens when the vector spaces are, furthermore, algebras. This includes spaces of operators, of
particular interest to quantum physics. That is why we start by recalling
some basic points behind the usual idea of quantization, and use some well
known aspects of quantum theory to announce some notions to be developed
afterwards.
5.1
Quantization
Quantum observables related to a physical system are self-adjoint operators

acting on a complex Hilbert space H. Each state is represented by an operator, the density matrix . In some special cases 2 = and the state,
called pure, can be represented by a single projector | >< |, where the
ket | > is an element of H. This exceptional situation is that supposed in
wave mechanics, in which | > is said to be the state of the system and
everything is described in terms of a wavefunction as, for example,
(x) = < x| >.
All predictions are of statistical character. Expectation values are attributed to an observable A as averages given by < A > = tr(A)/tr. If the
system is in the pure state | >, the value of A is
385
386
MATH.TOPIC 5. BANACH ALGEBRAS
A =
<|A|>
.
<|>
Given A and a function f (z), then f (A) represents (under conditions

given below) another operator. For an isolated system, time evolution is
fixed by the fact that |(t1 ) > is related to the same ket at another time,
|(t2 ) >, by a unitary evolution operator,
|(t2 ) > = ei(t2 t1 )H |(t1 ) >,
H being the hamiltonian of the system.
Thus, ultimately, quantization deals with operators acting on Hilbert
spaces. Such operators constitute by themselves other linear spaces and
submit to some peculiar conditions, imposed by physical and/or coherence
reasons. For example, in scattering problems the final state must be obtained
for times very large as compared to any other time interval characteristic
of the process, so that t2 = for all purposes. For analogous reasons
t1 = . Whether or not this is a well-defined notion depends on the
convergence of the evolution operator
U (t) = exp[i(t2 t1 )H],
and consequently on the topology defined on the space of operators. Some
norm must be introduced to provide a good notion of convergence and boundness of operators. Summarizing, we have a topological linear space of operators. Which leads to a Banach space. And, as operators compose between
themselves by product, an algebra of operators is present, which is a Banach
algebra. The algebra must contain the adjoint of each one of its elements, so
that what really appears is a special type of Banach algebra, called *-algebra.
As expectation values are the only physically accessible results and are given
by matrix elements, it is a weak topology which must be at work. This leads
to a still more specialized kind of algebra, a W -algebra. Let us briefly describe such spaces, pointing whenever possible to the main relationships with
quantum requirements.
Comment 5.1.1 Quantum operators are preferably bounded, in the sense that its
spectrum is somehow limited. Instead of using directly non-bounded operators, like the
momentum in wave mechanics, one considers their exponentials. There must be a norm
and, as suggested by the scattering example, a parameter-dependent operator must be
able to be continued indefinitely in the parameter. The operator linking the initial and
final states must belong to the algebra, wherefrom the completeness requirement.
5.2. BANACH ALGEBRAS
5.2
387
Banach algebras
A Banach space is a complete normed vector space, and a Banach algebra B

(older name: normed ring) is a Banach space with an associative internal
multiplication. One can always consider it to be a unit algebra, with unity
element I (if not, one is always able to make the adjunction of I; this is
not as trivial as it may seem, but is guaranteed by a theorem).
The norm must be consistent with the algebra structure. It must satisfy,
for in the field of the vector space B and all A and B B , the conditions
(i) ||A|| = || ||A|| ;
(ii) ||A + B|| ||A|| + ||B|| ;
(iii) ||AB|| ||A|| ||B|| ;
(iv) ||A A|| = ||A||2 .
Once endowed with the norm, the space becomes a metric space, with the
balls {v such that ||v u|| < } around each u. The algebra is involutive
if, besides the involution postulates (Math.1.17), the norm is preserved by
involution: ||A || = ||A||. The Banach space B is symmetric (or self-adjoint)
if, for each element A B, B contains also A .
Comment 5.2.1 Thus, the space of quantum operators must belong to a symmetric
involutive Banach algebra. The involution is the mapping taking each operator into its
adjoint. If A is an acceptable operator, so is its adjoint, which furthermore has the same
squared values.
In quantum theory, we are always interested in functions of operators.
In order to have them well-defined, we need some preliminary notions. The
spectrum SpA of an element A B is the set of complex numbers for which
the element A I is not invertible. An important theorem says that SpA
is a closed set, a non-empty compact subset of C. Consider the complement
of SpA, that is, the set C\SpA of those complex numbers for which the
element A I is invertible. The function
RA () = [A I]1
is called the resolvent of A. This function is analytic in C\SpA. If B happens
to be a field over C, it contains also z C and Sp z = . Thus, the only
Banach field over C is C itself (this is the Gelfand-Mazur theorem).
As SpA is compact, there exists a real number (A) = supSpA ||, which
is called the spectral radius of A. Actually,
(A) = limn ||An ||1/n .
388
Given a complex polynomial with complex coefficients, P (z) = a0 + a1 z +

a2 z + . . . + an z n , we can form a polynomial belonging to the Banach algebra,
P (A) = a0 + a1 A + a2 A2 + . . . + an An for each A B. As powers of the same
operator commute, the mapping P (z) P (A) is an algebra homomorphism.
The sets of polynomials in a fixed A will constitute a commutative algebra.
Taking now all the elements of B, the set of all polynomials constitute the
polynomial algebra of B, in general far from commutative. This procedure of
obtaining elements of B by extension of complex functions may be taken
further. More precisely, it may be taken up to the following point: consider an
open subset o(f ) of the set SpA and let AN (o(f )) be the algebra of analytic
functions on o(f ). The topology used is that of the uniform convergence on
compact sets. Then, there exists a homomorphism of the algebra AN (o(f ))
into B, given by
R
1
f ()RA ()d.
f (A) = 2
This homomorphism includes the above polynomial case, and the good
subset o(f ), as the notation suggests, depends on the function f . The integration is along , which is any closed curve circumscribing the entire set
SpA.
Comment 5.2.2 We learn thus, by the way, how to get a function of a given quantum
operator.
Consider now the case of commutative Banach algebras, which are the
natural setting for standard harmonic analysis1 (Mat.6.11). To each such
algebra one associates I(B), the set of its maximal proper (that is, 6= B)
ideals. I(B) is compact. An important point is that to each maximal ideal
corresponds a character of B, a homomorphic mapping : B C. Given ,
then 1 (0) is a maximal ideal of B. This interpretation of the characters as
maximal ideals leads to the Gelfand transformation: it relates a function on
b
I(B) to each element of B, given by A()
= (A). The set of values of the
b coincides with SpA.
function A
Let B = R(X) be the algebra of real continuous functions on the compact
X with the pointwise product (f g)(x) = f (x)g(x) as operation. The set of
functions vanishing at a point x is an ideal: f (x) = 0 implies (f g)(x) = 0
for any g. Each closed ideal is formed by those functions which vanish on
some closed subset Y X. There is a correspondence between the maximal
ideals and the points of X: we can identify x X to the maximal ideal I(x)
of functions f such that f (x) = 0 and the space X itself to the quotient
R(X)/I, where I is the union of all I(x).
1
Katznelson 1976.
5.3. *-ALGEBRAS AND C*-ALGEBRAS
5.3
389
*-algebras and C*-algebras
A *-algebra2 is a complete normed algebra with involution. A C*-algebra

is a Banach algebra over the field C of complex numbers, endowed with an
antilinear involution T T such that (T S) = S T and ||T T || = ||T ||2 .
Only in the framework of *-algebras can we talk about self-adjointness:
A is self-adjoint if A = A , and A is normal if it commutes with its adjoint:
A A = AA . Also, A is positive if (i) A = A and (ii) SpA R+ .
In a C*-algebra we can have square-roots. If A is positive, then there
exists a B in the C*-algebra such that A = B B.
Comment 5.3.1 The algebra of quantum observables is, under reasonable conditions,
a C*-algebra. The density matrices, which represent the possible states, are positive
operators. The space of states is contained in the space of positive operators.
The frequent use of C*-algebras to treat operators on Hilbert spaces justifies a more specific definition. In this case, a C*-algebra is denoted by L(H)
and is an involutive Banach algebra of bounded (see Math.7) operators taking a complex Hilbert space H = {} into itself, endowed with the norm ||T||
= Sup||||1 ||T || and the involution T T defined by
< T , > = < , T >, , H.
We can then prove that ||T T || = ||T ||2 .
Comment 5.3.2 We list some of the main topologies defined on L(H) :
(i) The norm topology, defined by the norm ||T || = Sup H,||||1 ||T || ; for finite
dimensions T is a matrix and ||T ||2 is the highest eigenvalue of T T ; for infinite dimensions,
||T ||2 is the spectral radius of T T .
(ii) The strong topology, weaker than the norm topology, is defined in such a way that
the sequence Tn converges to T iff Tn converges to T in H for all T H.
(iii) The weak topology, weakest of the three, is defined by the statement that Tn
converges to T in H iff, for all , H, | < Tn , > | converges to | < T , > |.
(iv) The -weak topology, defined by taking two sequences in H, {i } and {k },
with i ||i ||2 < and k ||k ||2 < . Then |T | := n |(n , T n )| is a seminorm on L(H)
and defines the -weak topology.
Let us recall that, as an algebra, A will have a dual space, formed by all
the linear mappings of A into the real line. We write dual, with quotation
marks, because infinite dimensional vector spaces are deeply different from
the finite dimensional vector spaces and one of the main differences concerns
precisely the dual. For finite dimensional spaces, the dual of the dual is the
2
Dixmier 1982.
390
space itself: (V ) = V . This is no more true in the infinite dimensional

case, the general result being that (V ) V . This is only to prepare for the
caveat dual which is essential in the study of infinite dimensional vector
spaces.
Even if H is of countable dimension, L(H) is not. Consequently, L(H)
is not isomorphic to the dual L (H). There is a (unique) subspace L (H) of
L (H) which is isomorphic to L(H), and this is usually called the predual of
L(H). The predual is a closed subspace of L (H), of which L(H) is the dual,
and it has the -weak topology.
Let us see how this relates to properties of spaces. If X is any compact
Hausdorff space, then the set C(X) of continuous complex functions on X is
a commutative unit algebra over C. It has a natural involution given by
(f )(x) = f (x) , and a norm ||f || = supxX |f (x)| which satisfies ||f f || =
||f ||2 . With this norm, the space C(X) is complete. It is consequently a
C*-algebra. By the way, a general C*-algebra has just this structure, up to
the commutative property. To every compact Hausdorff space corresponds
a commutative C*-algebra. If X has only a single point p, X = {p}, then
C(X) = C and each mapping F : X Y determines a functional
F : C(Y ) C on C(Y ) : (F f )(p) = f (F (p)).
This functional can be shown to be linear and multiplicative. Thus, to
points of Y correspond functionals on C(Y ). The Gelfand-Naimark theorem
states that this correspondence is one-to-one.
This has a deep consequence: each commutative unital C*-algebra A is
the algebra of continuous complex functions on a compact space Y : A =
C(Y ). And Y can be identified with the set of linear multiplicative functionals on the algebra. A linear multiplicative functional on the algebra is a
character. Thus, Y is the set of characters of A. Finally: each continuous
mapping F : X Y between two compact Hausdorff spaces induces a C*
homomorphism F : C(Y ) C(X), with (F f )(x) = f (F (x)), so that these
properties are, at least partially, carried over from compact to compact by
continuous mappings.
5.4
From Geometry to Algebra
We have said in section 5.2 that a compact space X can be seen as a subspace
of the algebra C(X) of continuous functions on X with the pointwise product
as multiplication. Each point of X is an ideal formed by those functions
which vanish at the point. Maximal ideals are identifiable to characters of
the algebra and in section 5.3 we have indeed said that if X is a compact then
5.4. FROM GEOMETRY TO ALGEBRA
391
it can be identified to the characters of C(X). This has been the starting
point of a process by which Geometry has been recast as a chapter of Algebra.
We shall only say a few words on the subject, which provides one of the gates
into non-commutative geometry and may become important to Physics in the
near future.
Consider again the space C (M, C) of differentiable complex functions
on a manifold M . It is clearly a linear space. Define on this space an internal operation of multiplication C (M, C) C (M, C) C (M, C). The
function space becomes an algebra. When there is no unit in this algebra,
we can always add it. What results is a unital algebra. To help the mind, we
can take the simplest, usual commutative pointwise product of complex functions: (f g)(x) = f (x)g(x). Suppose we are able to introduce also some norm.
With the pointwise product, C (M, C) becomes an associative commutative
*-algebra.
We can in principle introduce other kinds of multiplication and obtain
other algebras with the same starting space. Inspired by phase spaces, for
instance, we might think of introducing a Poisson bracket. From a purely
algebraic point of view, a Poisson algebra is a commutative algebra A as
above endowed with a map {, } : A A A such that:
(i) A is Lie algebra with the operation {, };
(ii) the bracket is a derivative in A: {a, bc} = b{a, c} + {a, b}c.
Once endowed with such a function algebra and bracket, a space M is
said to have a Poisson structure. Notice that M is not necessarily a phase
space: a Poisson structure can in principle be introduced on any differentiable manifold. But there is still more. The differentiable structure of M is
encoded3 in the *-algebra C (M, C). Each property of M has a translation
into a property of C (M, C). If we restrict ourselves to the space C 0 (M, C)
of functions which are only continuous, only the topological properties of M
remain encoded, the information on the differentiable structure of M being
forgotten. For instance, the mentioned theorem by Gelfand and Naimark
throws a further bridge between the two structures. It states, roughly, that
any unit abelian *-algebra is isomorphic to some algebra like C 0 (M, C), with
M some compact manifold. The difference between C (M, C) and C 0 (M, C)
is that the former has a great number of derivations, the vector fields on M .
In the algebraic picture, such derivations4 are replaced by the derivations
in the *-algebra, that is, endomorphisms of C (M, C) satisfying Leibniz
rule,5 like the above example with the Poisson bracket. Such derivations
3
4
5
Connes 1980.
392
constitute a Lie algebra if C (M, C) is associative. Vector fields on M have

components which are differentiable up to a given order. Higher order differentiability translates itself into properties of derivations on the algebra.
Summing up, the topological and geometrical properties of the manifold M
are somehow taken into account in the algebra of functions C (M, C).
Comment 5.4.1 Also algebraic properties of M are reflected in C 0 (M, C). If M
has also the structure of a topological group, for example, C 0 (M, C) will get some extra
properties.
Comment 5.4.2 We can define compact quantum spaces (quantum groups) by

dropping the commutativity requirement related to the pointwise product. In this case,
each C*-algebra with unit can be seen as the algebra of continuous functions on a certain
quantum space.
The simplest way to pass into noncommutative geometry6 is first to

go to C (M, C) and there dismiss the commutative character of the *algebra. This means alterating the pointwise product into some other, noncommutative product. The direct relation to the manifold M becomes fuzzy.
The best example comes out in the Weyl-Wigner picture of Quantum Mechanics. As said in section 10.2, you start with usual functions F (q, p) on
the most usual euclidean phase space of Classical Mechanics (Phys.1). Instead of using the pointwise product, you deform it into the star product
. Thus F (q, p) G(q, p) is replaced by F (q, p) G(q, p). But then the
functions themselves become ambiguous they are in general built up from
more elementary objects (monomials, for example) and these are ambiguous. For example, from the classical F (q, p) = qp you have to choose either
F (q, p) = q p or F (q, p) = p q. In building up functions, one usually uses a
lot of function-of-function stuff, and now one is forced to go to the very beginning and redefine each function from the most elementary redefined ones.
The results are the Wigner functions, actually c-number representatives of
quantum dynamical quantities which furthermore belong to an algebra with
a deformed Poisson bracket, the Moyal bracket. Non-commutative geometry comes out clearly. Either you go on using the coordinates q and p as
ordinary euclidean-valued functions but loose any connection with quantum reality; or you take also the coordinate functions q and p as belonging
to the deformed function algebra and then the coordinates are no more
commutative.
Comment 5.4.3 The Moyal bracket endows the algebra of Wigner functions with a
Poisson structure.
6
Connes1986.
5.5. VON NEUMANN ALGEBRAS
5.5
393
Von Neumann algebras
An involutive symmetric subalgebra M of L(H) containing the unit and

closed by the weak topology is a von Neumann algebra7 (or W -algebra). The
distinction between von Neumann algebras and general *-algebras is essential:
von Neumann algebras are closed under the weak topology while *-algebras
are closed under the norm topology. Of course, von Neumann algebras are
particular cases of *-algebras, but they are not, in general, separable by the
norm topology. Figure 5.1 is a scheme summarizing the main definitions. It
should be compared with scheme 1.1 in of Math.1 (page 346) and Figure 4.2
(page 384) of Math.4.
Figure 5.1: A scheme summarizing Banach spaces and algebras.

Let us rephrase all this, while introducing some more ideas.8 A von
Neumann algebra M is a nondegenerate self-adjoint algebra of operators on
7
8
Dixmier 1981.
Takesaki 1978.
394
a Hilbert space H, which is closed under the weak operator topology. This
is the locally convex topology in L(H) induced by the family of semi-norms
x L(H)
| < x| > | for all , H.
The von Neumann bicommutant theorem says that M = M 00 : M is the

commutant of its own commutant.
Comment 5.5.1 More precisely: if S is a subset of L(H), its commutant will be S 0
= { x L(H) such that xs = sx for all s S}. Thus, S 00 = (S 0 )0 . Call alg(S) the algebra
generated by S. Suppose two things:
(i) S is symmetric: x S implies x S;
(ii) 1 S.
Then the theorem says that alg(S) is strongly dense in S 00 (consequently, it is also
weakly dense in S 00 ).
An important fact is that the set of all projectors on a von Neumann

algebra M , with the identity included, generates M . We recall that a subset
U of an algebra M is said to generate M if the set of all the polynomials
obtained with all the members of U is dense in M . As the projectors are
idempotents, polynomials in projectors are simply linear combinations (see
section 5.6 below for a finite dimensional example).
The classification of von Neumann algebras is based on the properties of
its projectors. The center of a von Neumann algebra is abelian. A factor is
a von Neumann algebra with trivial center, M M 0 = C. This means that
the center is formed by the complex multiples of the identity.
One of the greatest qualities of von Neumann algebras is their receptivity
to integration. In effect, it is possible to define on them a measure theory
generalizing Lebesgues, and that despite their noncommutativity. A factor,
as said above, is a von Neumann algebra whose center is C. The space of
factors contained in a von Neumann algebra M is itself Borel-measurable
(Math.3). Call the measure . Each factor can be labelled by an index t
belonging to a borelian set, and denoted by M (t). Then the whole algebra
M is given by the decomposition
R
M = M (t)d(t)
(a theorem by von Neumann). This property justifies the name factor and
reduces the problem of finding all the von Neumann algebras to that of classifying all the possible factors. The last grand steps in this measure-algebraic
program were given recently, mainly by A. Connes. They naturally opened
the gates to noncommutative analysis and to noncommutative geometry. Another recent, astonishing finding (by Jones, see section 5.6 below) is that von
Neumann algebras are intimately related to knot invariants.
5.5. VON NEUMANN ALGEBRAS
395
Projectors are ordered as follows. Let p and q be two projectors in the von
Neumann algebra M . We say that p and q are equivalent, and write p q,
if there exists u M such that p = uu and u u = q. We say that p q
(q dominates p) if there exists u M such that p = uu and u uq = u u.
This means that u u projects into a subspace of qH, that is, that u u q.
If p q and q p, then p q. We say further that pq if pq = 0.
Now, if M is a factor, then it is true that, given two projections p and
q, either p q or q p. The terminology is not without recalling that of
transfinite numbers. The projector q is finite if the two conditions p q
and p q together imply p = q. The projector q is infinite if the conditions
p q, p q and p 6= q can hold simultaneously. The projector p is minimal
if p 6= 0 and q p implies q = 0 (p only dominates 0). The factors are then
classified in types, denoted I, II1 , II , and III:
I: if there exists a minimal projector;
II1 : if there exists no minimal projector and all projectors are finite;
II : if there exists no minimal projector and there are finite and infinite
projectors;
III: if the only finite projection is 0.
On a factor there exists a dimension function d: {projections on M }
[0, ] with the suitable properties:
(i) d(0) = 0;
(ii) d(k pk ) = k d(pk ) if pi pj for i 6= j;
(iii) d(p) = d(q) if p q.
It is possible then to show that, conversely, d(p) = d(q) implies p q.
This Murray-von Neumann dimension d can then be normalized so that
its values have the following ranges:
type I: d(p) { 0, 1, 2, . . . , n, with possibly n = }; if n is finite, type
In ; if not, type I ;
type II1 : d(p) [0, 1];
type II : d(p) [0, ];
type III: d(p) {0, }.
We have been talking about a *-algebra A as a set of operators acting on
some Hilbert space H. For many purposes, it is interesting to make abstraction of the supposed carrier Hilbert space and consider the algebra by itself,
taking into account as far as possible only its own properties, independent
396
of any realization of its members as operators. From this point of view, we

speak of the abstract *-algebra. The realization as operators on a Hilbert
space is then seen as a representation (Math.5) of the algebra, with H as the
carrier space. In this case we speak of a concrete *-algebra.
Comment 5.5.2 The name W*-algebra is frequently reserved to the abstract von
Neumann algebras.
It so happens that the abstract algebra is rich enough to provide even an

intrinsic realization on a certain Hilbert space. This comes out of the GNS
(Gelfand-Naimark-Segal) construction. The GNS construction is a method
to obtain a von Neumann algebra from a *-algebra A. One starts by building
a Hilbert space. The linear forms : A C constitute a vector space. A
form is positive if (x x) 0 for all x A. Given a positive form ,
one defines an inner product by < x, y > = (y x). There may exist zeros
of , elements x 6= 0 but with (x x) = 0. The set of zeros (kernel of )
form an ideal I = ker in A. The Hilbert space H is then the completion
of the quotient of A by this ideal, A/I. Then, if it is a C*-algebra, A will
act on by left multiplication and this action is the GNS representation. The
von Neumann algebra is the completion of the image of this representation.
Even if A is a factor, the GNS von Neumann algebra can have a non-trivial
center, due to the process of completion.
In Quantum Statistical Mechanics, von Neumann algebras are traditionally attained in the following way. One starts by preparing the formalism
for a finite system (finite volume and number of particles; or finite lattice
and lattice parameter), with all operators being finite matrices. Each state
is a density matrix (Phys.3) and the space of states is given by the set of
such positive operators. The expectation value of an observable A in state
is tr(A). In such finite models, no phase transition is ever found. Then
one proceeds to the thermodynamic limit, volumes and/or lattice going to
infinity. And one supposes that all this is well-defined, though the limit procedure is very delicate. Phase transitions are eventually found. Actually,
only a particular type of infinite algebras can be found as the limit of finite
algebras, the so-called hyperfinite algebras. The enormous majority of operator algebras cannot be attained in this way. The direct study of infinite but
non-hyperfinite algebras, which could describe physical systems beyond the
thermodynamic limit, is a major program of Constructive Field Theory.9
9
Of which a remarkable presentation is Haag 1993.
5.6. THE JONES POLYNOMIALS
5.6
397
The Jones polynomials
Let us now examine some particular finite-dimensional von Neumann algebras, of special interest because of their relationship with braids and knots.
A finite-dimensional von Neumann algebra is just a product of matrix algebras, and can be represented in the direct-product notation. In his work10
dedicated to the classification of factors, Jones11 was led to examine12 certain
complex von Neumann algebras13 An+1 , generated by the identity I plus n
projectors p1 , p2 , . . . , pn satisfying
p2i = pi = pi
pi pi1 pi = pi
pi pj = pj pi for |i j|
2.
(5.1)
(5.2)
(5.3)
The complex number , the inverse of which is called the Jones index, is
usually written by Jones as
=
t
,
(1+t)2
where t is another complex number, more convenient for later purposes. For
more involved von Neumann algebras, the Jones index is a kind of dimension
(notice: in general a complex number) of subalgebras in terms of which
the whole algebra can be decomposed in some sense. In lattice models of
Statistical Mechanics, with a spin variable at each vertex, the Jones index is
the dimension of the spin-space (Phys.3). Conditions [5.2] and [5.3] involve
clearly a nearest neighbor prescription, and are reminiscent of the braid
relations. We shall see below that some linear combinations of the projectors
and the identity do provide braid group generators.
Consider the sequence of algebras An . We can add the algebra A0 = C
to the club. If we impose that each algebra An embeds naturally in An+1 , it
turns out that this is possible for arbitrary n only if t takes on some special
values: either
(a) t is real positive, or
(b) t is of the form t = e2i/k , with k = 3, 4, 5, . . . in which case
=
10
11
12
13
1
4 cos2 (/k)
Jones 1983.
Jones 1987.
Jones 1985, and his contribution in Kohno 1990.
Wenzl 1988.
398
For these values of t there exists a trace defined on the union of the An s,
defined as a function into the complex numbers, tr: n An C, entirely
determined by the conditions
tr(ab) = tr(ba);
tr(w pn+1 ) = tr(w) for w An
(5.4)
(5.5)
tr(a a) > 0 if a 6= 0;
tr(I) = 1.
(5.6)
(5.7)
Conditions [5.1] [5.7] determine the algebra An up to isomorphisms.

Such algebras were known to physicists, as An had essentially been used
by Temperley and Lieb14 in their demonstration of the equivalence between
the ice-type and the Potts models,15 the only difference being in the projector
normalization. Define new projectors Ei = dpi , where d is a number. If
= d2 , the conditions become
Ei2 = dEi ;
Ei Ei1 Ei = Ei ;
Ei Ej = Ej Ei for |i j| 2.
(5.8)
(5.9)
(5.10)
A fascinating thing about these algebras is that they lead to a family of

invariant polynomials for knots (Math.2.16), the Jones polynomials. But to
physicists, perhaps the main point is that the partition function of the Potts
model is a Jones polynomial for a certain choice of the above variable t.
This entails a relationship between lattice models and knots.
We shall in what follows make large use of the terminology introduced in
Math.2. A representation of the braid group Bn in this algebra is given as
follows: to each generator corresponds a member of the algebra:
(5.11)
r(i ) = Gi = t [tpi (I pi )].
Actually, these operators Gi are just the invertible elements of the algebra,
so that Bn appears here as the group of the algebra An . The inverse to
the generators are
1 1
G1
(5.12)
i = [t pi (I pi )].
t
Each generator G will satisfy a condition of the type (G a)(G b) = 0.
This means that the squared generators are linear functions of the generators,
14
15
Temperley & Lieb 1971.

See Baxter 1982.
399
(normalized so that
so that we have a Hecke algebra. One has P = Ga
ba
2
P = P ) as the projector on the eigenspace of b, in terms of which G =
(b a)P + aI = a(I P ) + bP . The projectors are
pi = [Gi + t]/[(1 + t) t],
and the condition p2i = pi is equivalent to G2i = t(t 1) Gi + t2 , or
(Gi t t)(Gi + t) = 0,
or still
t1
1
t G1
I = 0.
(5.13)
i t Gi +
t
Let us introduce an inspiring notation. Indicate each braid generator by
/
Gi = /\ , where it is implicit that the first line is the i-th, and all the unspecified lines are identity lines. The identity itself is indicated by )( and G1
i
\
suitably by /\ . Relation [5.13] can then be drawn as
\
t /\ - t1 /\ +
t1
)( = 0 .
(5.14)
This is a skein relation, emerging here as the representative of an algebraic

relation. The representation of the braid group Bn takes place actually in a
Hecke sub-algebra of An , which this relation determines.
Comment 5.6.1 For the Alexander polynomial, the relation is (G t)(G + 1/ t) =

0.
In Kauffmans monoid diagrams, the projectors Ei are represented by

and will give to the relation [5.11] the form
/
/\ =
S
t3/2 +t1/2 T
d
- t1/2
)( .
(5.15)
The first projectors are shown in Figure 5.2 (for the case n = 4).
Figure 5.2:
And a simple blob gives just d as a number multiplier: The bubble is
normalized by
400
= d.
Condition [5.10] is immediate. Conditions [5.9] and [5.8] are shown respectively in Figure 5.3 and Figure 5.4.
Figure 5.3:
Figure 5.4:
The reason for the name monoid diagrams is simple to understand here.
Projectors, with the sole exception of the identity, are not invertible. Once
we add projectors to the braid group generators, what we have is no more
a group, but a monoid (Math.1). The addition of projectors to the braid
generators, or the passing into the group algebra, turns the matrix-diagram
relationship into a very powerful technique.
The Jones polynomials are obtained as follows. Given a knot, obtain it

as the closure b of a braid b. Then the polynomial is

n1
1+t
Vb (t) =
tr [r(b)].
t
401
(5.16)
Given a knot, draw it on the plane, with all the crossings well-defined.
Choose a crossing and decompose it according to [5.14] and [5.15]. Two new
knots come out, which are simpler than the first one. The polynomial of the
starting knot is equal to the sum of the polynomials of these two new knots.
Do it again for each new knot. In this way, the polynomial is related to the
polynomials of progressively simpler knots. At the end, only the identity and
the blob remain.
Jones has shown that his polynomials are isotopic invariants. The Jones
polynomial is able to distinguish links which have the same Alexander polynomial. Perhaps the simplest example is the trefoil knot: there are actually
two such knots, obtained from each other by inverting the three crossings.
This inversion corresponds, in the Jones polynomial, to a transformation
t t1 , which leads to a different Laurent polynomial. This means that the
two trefoils are not isotopic.
Kirillov 1974
Haag 1993
Dixmier 1981
Dixmier 1982
Kaufman 1991
Jones 1991
402
Math.Topic 6
REPRESENTATIONS
0 Introduction
A LINEAR REPRESENTATIONS
1 Generalities
2 Dimension
3 Unitary representations
4 Equivalent representations
5 Characters
6 Irreducible representations
7 Tensor products
B REGULAR REPRESENTATION
8 Invariant spaces
9 Invariant measures
10 Generality
11 Relation to von Neumann algebras
C FOURIER EXPANSIONS
12 The standard cases
13 Pontryagin duality
14 Noncommutative harmonic analysis
15 The Peter-Weyl theorem
16 Tanaka-Krein duality
17 Quantum groups
Introduction
In general, a representation1 of a group G is a homomorphism of G into
some other group H . A representation of a Lie algebra G0 of a Lie group
G will be a homomorphism of G0 into some other Lie algebra H 0 . All this is
1
A very sound though compact text on representations is Kirillov 1974.
403
404
MATH.TOPIC 6. REPRESENTATIONS
rather abstract. Actually, representation theory is a way to look at groups

as sets of transformations: H is a transformation group, formed typically
by transformations on vector spaces. Thus, linear representations are those
for which H is the group Aut V of the linear invertible transformations of a
vector space V , or, roughly speaking, those for which H is a matrix group. A
current, though misleading practice is to use the expression representation
for V itself. For algebras, H 0 is a matrix algebra, or something generalizing
it. We shall here concentrate on group representations. We shall make a
passage through linear representations, finite and infinite,2 and come back
to the group, to examine the case of regular representations, for which the
carrier spaces are spaces of functions on the group itself. We finish with a
short incursion into Fourier analysis3 , a chapter of representation theory4 of
permanent interest to Physics.
6.1
A Linear representations
6.1 Generalities
Consider a vector space V and the set of transformations defined on
it. Of all such transformations, those which are continuous and invertible
constitute the linear group on V , indicated by L(V ). A linear representation
of the group G is a continuous function T defined on G , taking values on
L(V ) and satisfying the homomorphism requirement,
T (gh) = T (g)T (h).
(6.1)
We speak, in general, of the representation T (g). The space V is the

representation space, or the carrier space. The representation is faithful if
T is one-to-one. Call e the identity of G , and E the identity of L(V ).
It follows from the requirement T (g 1 ) = T 1 (g) that T (e) = E. A very
particular case is the trivial representation, which takes all the elements of
G into E. The representation T (g) is said to be exact when the identity of
G is the only element taken by T into E. If {ei } is a basis for V , the matrix
T (g) = (Tij (g)) will have entries given by the transformation
T (g)ei = j Tij (g)ej .
6.2 Dimension
2
3
4
Mackey 1955.
Gelfand, Graev & Vilenkin 1966.
Mackey 1978.
(6.2)
6.1. A LINEAR REPRESENTATIONS
405
Ados theorem (see 8.3.10) may be rephrased as follows: every finitedimensional Lie algebra has a faithful linear representation. Or still: every
finite-dimensional Lie algebra can be obtained as a matrix algebra. If V
is a finite space, such matrices are finite and dT (the dimension of V ) is
the dimension of representation T (g). The homomorphism condition is then
simply
dT
X
Tij (gh) =
Tik (g) Tkj (h).
(6.3)
k=1
When V is infinite, we must worry about the convergence of the involved series. V can be, for instance, a Hilbert space. In the general case of a carrier
space endowed with an inner product (u, v), the adjoint A of a matrix A
satisfies (Au, v) = (u, A v) for all u, v V . It comes out that, if T (g) is a
representation, then T (g 1 ) is another representation, called the representation adjoint to the representation T (g) (this is not to be confounded with
the adjoint representation of a Lie group).
6.3 Unitary representations
A representation T (g) of a group G in the inner product space V is unitary
if, for all u, v in V ,
(T (g)u, T (g)v) = (u, v).
(6.4)
In this case, T (g)T (g) = E and T (g) = T 1 (g) = T (g 1 ), so that a unitary
representation coincides with its adjoint.
Comment 6.1.1 The name comes from the matrix case: in a matrix unitary representation, all the representative matrices are unitary.
6.4 4 Equivalent representations
Let T (g) be a representation on some carrier space V , and let us suppose
there exists a linear invertible mapping f into some other vector space U ,
f : V U . Then, T 0 (g) = f T (g) f 1 is a representation with representation space U . The relation between such representations, obtained
from each other by a linear invertible mapping, is an equivalence. Two such
representations are thus equivalent representations and may be seen as different realizations of one same representation. A well known example is
the rotation group in euclidean 3-dimensional space, SO(3): the vector representation, corresponding to angular momentum 1, is that generated by the
fields
ijk (xj k xk j )
if seen as transformations on functions and fields on E3 ; it is that of the real
orthogonal 33 matrices if seen as acting on column vectors.
406
6.5 Characters
Consider, as a simple and basic example, the case in which the carrier
space V is a one-dimensional space. Matrices reduce to functions, so that
T (g) = (g)I, with I the identity operator and a non-vanishing complex
function on G. As T is a homomorphism, (gh) = (g)(h) for all g, h
G. A function like , taking G homomorphically into C\{0}, is called a
character . If G is a finite group, there exists some integer n such that (g n ) =
n (g) = 1, so that |(g)| = 1 and *(g) = (g)1 .
Every finite abelian group is the direct product of cyclic groups ZN . Such
a cyclic group has a single generator g such that g N = I. Hence, N (g) = 1.
For ZN , consequently, the characters are the N -th roots of the identity, one
for each group element. ZN is, therefore, isomorphic to its own group of
characters.
The characters of the direct product of two groups are the respective
product of characters, and it follows that every finite abelian group is isomorphic to its group of characters. The characters of finite abelian groups
have the following further properties:
(i) gG (g) = 0;
(ii) gG 1 (g)2 (g) = 0, whenever 1 6= 2 .
The characters are thus orthogonal to each other. They are in consequence linearly independent, they span the space of functions on the group,
and form a basis for the vector space of complex functions on G . If the
group is unitary, so is the character representation: *(g) = 1 (g).
6.6 Irreducible representations
Suppose that the carrier space V has a subspace V 0 such that T (g)v 0 V 0
for all v 0 V 0 and g G. The space V 0 is an invariant subspace, and
T (g) is reducible. When no invariant subspace exists, the representation is
irreducible. In the reducible case there are two new representations, a T 0 (g)
on V 0 and another, T 00 (g) on the quotient space V 00 = V /V 0 . There exists
always a basis in which the matrices of a reducible representation acquires
the bloc-triangular form
0

T (g) K(g)
.
0
T 00 (g)
There is in general no basis in which K(g) vanishes. This happens, however,
when V 0 admits a linear complement V 00 which is also an invariant subspace,
so that V is the direct sum V = V 0 V 00 of both. Recall that this means
that any v V can be written as v = v 0 + v 00 , with v 0 V 0 and v 00 V 00 . In
this case
T (g)v = T (g)v 0 + T (g)v 00 = T 0 (g)v 0 + T 00 (g)v 00 .
6.1. A LINEAR REPRESENTATIONS
407
The representations T 0 (g) and T 00 (g) are the restrictions of T (g) to V 0

and V 00 and we write T (g) = T 0 (g) + T 00 (g). Notice that we are incidentally
defining an addition of representations. The best physical example is the
addition of angular momenta. If V 0 or V 00 are themselves direct sums of
other invariant spaces, the procedure may be continued up to a final stage,
when no more invariant subspaces exist.
When T (g) is such that it is possible to arrive at a final decomposition
T (g) = j Tj (g), with each Tj (g) an irreducible representation, T (g) is said
to be completely reducible. An important related result is:
All unitary finite-dimensional representations are
completely reducible.
Comment 6.1.2 This is not always true for infinite-dimensional representations. On
the other hand, all the irreducible representations of a commutative group have dimension
1.
6.7 Tensor products
Let T (g) and S(g) be two representations of G, respectively on spaces
V and U . They define another representation of G , the tensor product (or
Kronecker product) R = T S of T and S, acting on the direct product of
V and U . Operators A, B acting on V , and C, D acting on U will satisfy
(A C)(B D) = (AB) CD) .
With tensor products we go into higher dimensional representations, so that
there is a higher chance of obtaining reducible representations.
Coming back to the characters, we have above defined them for onedimensional representations. For general matrix representations, they are
defined as the trace of T (g) ,
T (g) = tr T (g) =
dT
X
Tkk (g),
(6.5)
k=1
where dT is the dimension of the representation. Due to the trace properties,

the character depends only on the class of equivalent representations. Of
course, when the group is not finite, the summations on the group used above
have to be examined in detail. In particular, for continuum groups they
become integrals and some measure must be previously defined. We shall
come to this point later. Let us only state two properties which hold anyway:
the character of an addition is the sum of characters, and the character of a
tensor product is the product of the characters:
T +S (g) = T (g) + S (g);
T S (g) = T (g)S (g).
(6.6)
(6.7)
408
6.2
B Regular representation
6.8 Invariant spaces

Let M be a homogeneous space under G and consider L the set of functions f : M V , where V is some space. Then L is said to be an invariant
space of functions if f (x) L implies f (gx) L. It may happen that some
subspace of L be invariant by itself. To each such invariant subspace will
correspond a representation, given by
T (g)f (x) = f (g 1 x).
(6.8)
6.9 9 Invariant measures

As introduced above, L is not even a topological space. If there exists an
invariant measure on M , we may instead take for L the space of complex,
square-integrable functions. In this case, given the inner product
Z
(f1 , f2 ) =
f1 (x)f2 (x)d(x),
(6.9)
M
the representation [6.8] is unitary:

Z
(T (g)f1 , T (g)f2 ) =
f1 (g 1 x)f2 (g 1 x)d(x)
M
Z
f1 (x)f2 (x)d(x) = (f1 , f2 ). (6.10)
=
M
Recall that on locally-compact groups, the existence of a left-invariant measure is guaranteed, as is the existence of a right-invariant measure. Such Haar
measures, by the way, do not necessarily coincide (groups for which they coincide are called unimodular ). Given the actions of a group G on itself, the
functions defined on G will carry the left and right representations:
L(h)f (g) = (Lh f )(g) = f (h1 g)
R(h)f (g) = (Rh f )(g) = f (gh).
(6.11)
(6.12)
6.10 Generalities
If G has a left-invariant measure, we may take for function space the set
of square-integrable functions. L(g) is called the left-regular representation.
In an analogous way, R(g) is the right-regular representation. Such representations, given by operators acting on functions defined on the group itself,
are of the utmost importance. First, because it happens that
5
Notice to avoid confusion that, as already mentioned, some authors use the name
regular representation in other senses: see for instance Hamermesh 1962, Gilmore 1974.
6.3. C FOURIER EXPANSIONS
409
any irreducible representation of G is equivalent to

some regular representation.
And second, because their study is the starting point of generalized
Fourier analysis, or Fourier analysis on general groups (or still, non-commutative
harmonic analysis see par:Math.6.14 below).
6.11 Relation to von Neumann algebras6
What we have here are representations in terms of operators on function
spaces. We have been forced to restrict the function spaces through the measure requirements. The inner product introduced above defines a topology,
and we may go a step further. We require that the function space be complete in this topology, so that they become Hilbert spaces. The T (g)
above belong consequently to spaces of operators acting on Hilbert spaces.
They may be added and multiplied by scalars, constituting consequently a
linear space. Such operator spaces are themselves normed spaces, the norm
being here the weak one, i.e., that provided by the internal product. Restricting again the whole scheme, we content ourselves with those subspaces
of normed operators which are complete, so as to obtain a Banach space. A
product of such operators is defined, so that they are Banach algebras. On
the other hand, they are involutive spaces, and complete by the weak topology given by the internal product. Thus, finally, they actually constitute
von Neumann algebras. Summarizing: the regular representations are group
homomorphisms from the group G into von Neumann algebras. Wherefrom
comes a relationship7 between the classification of irreducible group representations and the classification of von Neumann algebras (Math.5.5). Just
as for the latter, the groups are said to be of type I, II, III according to their
representations. Only for type I groups does the simple ordinary decomposition of a representation into a sum (in general, an integral) of irreducible
representations hold. For types II and III, the space of representations is in
general not even (first-)separable.
6.3
C Fourier expansions
We discuss some small bits of Fourier analysis, starting with the main qualitative aspects of the elementary case.
6.12 The standard cases
6
7
Mackey 1968.
Mackey 1978, chap. 8.
410
Consider a periodic square-integrable complex function f on the line E1 .

Being periodic means that f is actually defined on the circle S 1 , which is the
manifold of the group U (1) = SO(2). The Fourier expansion for f is
f (x) =
einx f(n),
(6.13)
dxeinx f (x)
(6.14)
n=
where the discrete coefficients

1
f(n) =
2
are the values of the Fourier transform of f. Notice that:

(i) the original space U (1) is compact and the series is discrete, n Z;
(ii) for each n, einx is a character of U (1);
(iii) the characters form irreducible, unitary representation of U (1);
(iv) the Fourier series is consequently an expansion in terms of nonequivalent irreducible unitary representations of the original group U (1).
Conversely,
(i) also Z is a (non-compact) group with the addition operation, (Z, +),
and f(n) is its Fourier expansion;
(ii) for each x, einx can be seen as a character for Z;
(iii) such characters form irreducible, unitary representation of Z;
(iv) the Fourier integral is consequently an expansion in terms of irreducible unitary representations of Z.
Suppose now we dropped the periodicity condition: we would then have
Z
Z
1
iyx
dxeiyx f (x)
(6.15)
dye f (y) ; f(y) =
f (x) =
2
with both x and y E1 . Analogous statements hold, with the difference that
now both groups are (E1 , +), which is non-compact.
Suppose finally we started with the cyclic group ZN , which is a kind of
lattice circle. Then,
f (m) =
N
1 X i(2/N )mn
e
f (m).
ei(2/N )mn f(n) ; f(n) =
2
m=1
n=1
N
X
(6.16)
Both groups are ZN , compact and discrete.

6.13 Pontryagin duality
According to the above discussion, Z is Fourier-dual to S 1 and viceversa. The line E1 is self-dual and so is ZN . These results keep valid for
411
other cases, such as the euclidean 3-space E3 : the Fourier transformations

establish a duality between the space of functions on the original space (which
may be seen as the translation group T3 in E3 ) and the space of the Fourier
transforms, which are functions on in principle another space. The latter
is the space of (equivalence classes of) unitary irreducible representations
of T3 , and constitutes another group. It is a general result that, when the
original group G is commutative and locally compact, the Fourier-dual is also
a commutative locally compact group. Furthermore, if they are compact
or discrete, the duals are respectively discrete or compact. This duality
appearing in the abelian case is the Pontryagin duality. Local compactness
is required to ensure the existence of a Haar measure, which in the example
above is just the (conveniently normalized) Lebesgue measure.
6.14 Noncommutative harmonic analysis
Harmonic analysis may be extended to other groups, although with increasing difficulty. There is a complete theory for abelian groups. For nonabelian groups, only the compact case is well established and the subject has
been recently christened non-commutative harmonic analysis. We have
said (Math.5.2) that commutative Banach algebras are the natural setting
for standard harmonic analysis, that is, for Fourier analysis on abelian
groups. For non-commutative groups, non-commutative Banach algebras
are the natural setting. As a good measure is necessary, the research has
been concentrated on locally-compact groups.
The reason for the special simplicity of abelian groups is that their unitary irreducible representations have dimension one and the tensor product
of two such representations is another one of the same kind. Each such
representation may be considered simply as a function f C(G),
f : G C, g f (g),
with
f (g1 g2 ) = f (g1 )f (g2 ).
(6.17)
Tensor products are then reduced to simple pointwise products of functions.

The set of inequivalent unitary irreducible representations is consequently
itself a group. This is the property which does not generalize to the noncommutative case.
Consider a representation T on a Hilbert space H . A group representation extends to a ring representation. The general Fourier transform is
f() = hG T (h)df (h),
(6.18)
where f() End H . It takes an element f R(G) of the ring into f G.
412
6.15 The Peter-Weyl theorem8

A compact group G has a countably infinite set of representations, which
we shall label by the index . Each representation T is acting on some
vector space of dimension d = dim T (G), and each group element g will
be represented by the matrix T (g) with elements
[T (g)]ij ; i, j = 1, 2, . . . , d .
of all the unitary irreducible representations of G is a sum
The set G
of all such spaces, and is called the unitary dual of G . The Peter-Weyl
theorem says that, with an invariant normalized measure dg on G , the set
{[T (g)]ij } of all matrix elements of all the representations is a complete
orthogonal system for the square-integrable functions on G. More precisely,
if f : G C and
R
|f |2 dg < ,
then
f (g) =
d p
XX
d [T (g)]ij f ij ,
(6.19)
i,j=1
where
f
Z
ij
= d
dgf (g)[T (g)]ij .
(6.20)
are the Fourier components of f . The matrix elements [T (g)]ij , in terms

of which every square-integrable function can be expanded, are the special
functions we are used to. Every time we have a spherically symmetric problem, for example, we obtain ultimately solutions in terms of Legendre polynomials. These polynomials are exactly the above matrix elements for the
representations of the rotation group.
6.16 Tanaka-Krein duality
The Fourier transformation makes use of the unitary irreducible representations of the group, and that is why harmonic analysis is a chapter of
representation theory. When G is a compact nonabelian group, the dual G

is in fact a category, that of the finite dimensional representations of G . It
is a category of vector spaces (or an algebra of blocs). The representations
of this category (representations of categories are called functors) constitute
a group isomorphic to G. This new duality, between two different kinds of
structures a group and a category is called the Tanaka-Krein duality.
8
Vilenkin 1969.
413
6.17 Quantum groups

Is it possible to enlarge the notion of group to another object, so that
its dual comes to be an object of the same kind? The complete answer has
been found in the case of finite groups: the more general objects required are
Hopf algebras (Math.1), frequently called quantum groups in recent times.
On infinite groups, the operators must stand in Hopf algebras which are
also von Neumann algebras. Amongst such Hopf-von Neumann algebras,
some can be chosen (called Kac algebras) that respect some kind of Fourier
duality. But different notions of duality are possible when we go to the finest
details, and the subject is still a province of mathematical research. Due to
the clear relationship between quantization and Fourier analysis, it would be
interesting to find quantum groups as those generalizations of groups allowing
for a good notion of Fourier duality.
Vilenkin 1969
Mackey 1978
Kirillov 1974
Katznelson 1976
Enoch & Schwartz 1992
414
Math.Topic 7
VARIATIONS & FUNCTIONALS
A CURVES
[1] Variation of a curve
[2] Variation fields
[3] Path functionals
[4] Functional differentials
[5] Second-variation
B GENERAL FUNCTIONALS
[6] Functionals
[7] Linear functionals
[8] Operators
[9] Derivatives Frechet and Gateaux
A geometrical approach to variations is given that paves the way to the introduction of functionals and creates the opportunity for a glimpse at infinitedimensional manifolds.
7.1
7.1.1
A Curves
Variation of a curve
To study the variation of a curve we begin by placing it in the middle of a

homotopy. A curve on a topological space M is a continuous function
: I = [1, +1] M .
415
416
MATH.TOPIC 7. VARIATIONS & FUNCTIONALS
We shall consider only paths with fixed ends, from the initial end-point a
= (1) to the final end-point b = (1). Given paths , , going from a
to b, they are homotopic (we write ) if there exists a continuous
mapping F : I I M such that, for every t, s in I,
F (t, 0) = (t); F (t, 1) = (t); F (t, 1) = (t); F (1, s) = a; F (1, s) = b .
The family of curves F is a homotopy including , and , or, if we wish,
a continuous deformation of the curve into the curves and .
For each fixed s, s (t) = F (t, s) is a curve from a to b, intermediate
between and , with (t) = 0 (t) somewhere in the middle. But also, for
each fixed t, t (s) = F (t, s) is a curve going from the point (t) to the point
(t) while s treads I and meets (t) somewhere in between (see Figure 7.1).
Let us fix this notation: s (t) is a curve with parameter t at fixed s; t (s)
is a curve with parameter s at fixed t. The one-parameter family s of curves
is called a variation of .
Figure 7.1:
7.1.2
Variation fields
We now add a smooth structure: M is supposed to be a differentiable manifold and all curves are differentiable paths. Each s (t) will have tangent
vectors
7.1. A CURVES
417
Figure 7.2:
Vs (t) =
d
(t)
dt s
= s ,
which are its velocities at the points s (t). But also the transversal curves
d
t (s) will have their tangent fields, ds
t (s). Consider now the curve (t):
besides its velocity
V (t) = V0 (t) = (t)
d
(t)
dt 0
it will display along itself another family of vectors, a vector field induced by
the variation, (see Figure 7.2)
X(t) =
d
(s)|s=0
ds t
d
(0).
ds t
X(t) will be a vector field on M , defined on each point of (t). This vector
field X is called an infinitesimal variation of . It is sometimes called a
Jacobi field, though this designation is more usually reserved to the case in
which is a geodesic.
7.1.3
Path functionals
Non-exact Pfaffian forms, whose integrals depend on the curve along which
it is taken, open theR way to path functionals. A trivial example is the length
itself: the integral ds depends naturally on the path. Another example is
418
the work done by a non-potential force, which depends on the path along
which the force is transported from point a to point b: the integral
R
Wab [] = Fk dxk
is actually a function of the path, it will depend on the functional form of
(t). Paths belong to a space of functions, and mappings from function spaces
into R or C are functionals. Thus, work is a functional of the trajectory. But
perhaps the most important example in mechanics is the action functional
related to the motion between two points a and b along the path . Let the
points on be given by coordinates i and the corresponding velocity be
given by i . We can use time as the parameter t, so that the velocity
above is the true speed. Let the lagrangian density at each point be given
by L[ i (t), i (t), t]. The action functional of the motion along will be
Z
(7.1)
A[] = L[ i (t), i (t), t] dt.
To each path going from a to b will correspond an action. Consider the

space Hab = {} of all such paths. The functional A is a mapping A : Hab
E1 . Notice that the underlying space M remains in the background: the
coordinates of its points are parameters. If they appear in the potential
function U in the usual way, through
L = T (x i ) U (xi ),
only its values U ( i ) on the path will actually play a role. Clearly the integral
A does not depend on the points, not even on the values of the sole acting
parameter t. The action A depends only on how depends on t.
7.1.4
Functional differentials
Once we have established that A is a real-valued function on Hab , we may

think of differentiating it. But then, to start with, Hab should be a manifold.
The question is: Hab being of infinite dimension, how to make of it a manifold? As repeatedly said, manifolds are spaces on which coordinates make
sense. The euclidean spaces are essential to finite dimensional manifolds, as
they appear both as coordinate spaces and as tangent spaces at each point
of differentiable manifolds. Here, the place played by the euclidean spaces
would be played by Banach spaces, infinite dimensional spaces endowed with
topology and all that. We keep this point in mind, but actually use an expedient: we take as coordinates the i above, understanding by this that,
through the infinite possible values they may take, we are in reality covering
7.1. A CURVES
419
the space Hab with a chart. The high cardinality involved is hidden in the
apparently innocuous , which contains an arbitrary functional freedom.
This assumption makes life much simpler than a detailed examination of Banach spaces characteristics, and we shall admit it. Let us only remark that,
Banach spaces being vector spaces, we are justified in adding the coordinates
i as we shall presently do. There is, however, a clear problem in multiplying
them by scalars, as then the end-points could change. This means that paths
do not constitute a topological vector space. It is fortunate that we shall only
need addition in the following.
A differential of the action A (which does not always exist) can be introduced as follows. Take the path and one of its neighbours in the variation
family, given by the coordinates i + i . We represent collectively the coordinates of the path points by and + . Suppose then that we may
separate the difference between the values of their actions into two terms,
A[ + ] A[] = F [] + R,
(7.2)
in such a way that F [] is a linear functional of , and R is of second (or

higher) order in . When this happens, A is said to be differentiable, and
F is its differential. By a smart use of Dirac deltas, we can factorize out
of F [], obtaining an expression like
F [] =
(examples can be found in Math.8). The quantity is usually called

is the functional derivative. This derivative
the variation, and the factor F
has many names. In functional calculus its is called Frechet (or strong)
derivative. In mechanics, due to the peculiar form it exhibits, it is called
Lagrange derivative. For the action functional [7.1], we have
A[ + ] A[]
Z
Z
i
i
i
i
= L[ (t) + (t), (t) + (t)]dt [ i (t), i (t)]dt

Z
L i L i
=
+ i dt + O( 2 ),
i
so that
Z
F [] =

L i L i
+ i dt.
i

(7.3)
420
The second term in the bracket is

Z

Z
Z
d L i
d L
L d i
dt =
i dt
dt
i
i
i dt
dt
dt

Z
L i
L i
d L
=
i dt.
i
i
i

dt
t=1
t=0
As the variations are null at the end-points, only the last term remains and
we get

Z
L
d L
i dt.
(7.4)
F [] =
i
i
dt

The integrand is now the Lagrange derivative of L. The Euler-Lagrange
equations
L
d L
=0
(7.5)
i
dt i
are extremum conditions on the action: they fix those curves which extremize A[].
7.1.5
Second-variation
Going further, we may add to the differentiable structure of M a connection

. The covariant derivative along an arbitrary vector field X is generally
indicated by X . Given a curve , whose tangent velocity field will be
V (t) = (t)
d
(t),
dt
the covariant derivative V i Di W of a field or form W along the curve will

or V W . The covariant acceleration, for example, will
be indicated by DW
Dt
be
a=
DV
Dt
= V V.
The field (or form) W will be parallel-transported along if V W = 0. The

infinitesimal variation field X, or Jacobi field, will submit to a second order
ordinary differential equation, the Jacobi equation (see Math.12).
Curves are very special examples of functions. We have discussed functionals on spaces of curves. Let us now address ourselves to more general
function spaces.
7.2. B GENERAL FUNCTIONALS
7.2
421
B General functionals
A heuristic introduction to functionals which generalize the path functionals will be given
here. Attention will be called on the topologies necessarily involved. The approach is voluntarily repetitive. Functionals generalize to operators. Functionals are real- or complexvalued functions on some spaces of functions, while operators take such spaces into other
spaces of the same kind. Only that minimum necessary to the presentation of the very
useful notion of functional derivative will be treated.
7.2.1
Functionals
Let us gather all the courage and be a bit repetitive. A functional is a

mapping from a function space into R or C. Let us insist on the topic by
giving a heuristic introduction1 to the notion of functional as a generalization
of a function of many variables. Consider the function f : RN C, written
as f (x1 , x2 , . . . , xN ). It is actually a composite function, as the argument
may be seen as another function
x: {index set} RN , x:1, 2, 3, . . . , N } (x1 , x2 , . . . , xN ),
which attributes to each integer i the value xi . We have thus a function of
function f [x(k)], starting from a discrete set of indices. Why do we still talk
of f as a function of x? Because the index set is fixed throughout the
process of calculating a value of f . Given this set, the function x will fix
a value (x1 , x2 , . . . , xN ). To each value of x, f will then attribute a real or
complex number. Or, if we prefer, to another function
y: {index set} RN , y:1, 2, 3, . . . , N } (y 1 , y 2 , . . . , y N ),
f will attribute another number. The function f is thus dependent only
on the set {x} of functions from {1, 2, 3, . . . , N } into RN , and not at all
on the set {1, 2, 3, . . . , N }. This set is fixed, so that f is only a function
f : {x} C. Take now, instead of the discrete set {1, 2, 3, . . . , N }, a set of
points in some continuum, say, the interval I = [0, 1]. The set {x} will then
be a set of functions on I, each one some x(t). And f [x(t)] will depend on
which function x(t) is considered, that is to say, f will be a functional on the
space of functions defined on I. This space might, in principle, be anything,
but functionals are usually introduced on topological spaces, such as the set
of square integrable functions on the line, or the set of twice-differentiable
functions on the sphere S 2 .
1
Inspired by Balescu 1975.
422
Functionals appear in this way as functions depending on many (actually, infinite) variables, or on variables whose indices belong to a continuum.
For instance, t may be time, x(t) a trajectory between two fixed points in
E3 , and f will be a functional on trajectories (say, the classical mechanical
action [7.1]). We may of course take the converse point of view and consider
usual few-variable functions as functionals on spaces of functions whose arguments belong to a discrete finite set. Or we may go further in the first
direction, and consider many-dimensional sets of continuous indices. For instance, in relativistic field theory (see Phys.6) such indices are the coordinates
in Minkowski space, the role of the degrees of freedom being then played by
fields (x, y, z, ct). A splendid functional will be the action functional for the
fields .
7.2.2
Linear functionals
The reasoning above makes clear the interest of functionals: they are the
functions defined on infinite dimensional spaces, specially on spaces whose
infinity has the power of the continuum or more. As function spaces can
be always converted into linear spaces, a functional can be defined as any
complex function defined on a linear space. Once we have established that
a functional is a function from a function space {f } to C, we may think of
applying to it the usual procedures of geometry and analysis. But then, to
start with, the function space should be a good (topological) space. Better
still, it should be a manifold. Banach spaces will play here the role played
by the euclidean spaces in the finite case. Some authors2 give another name
to topological vector spaces: they call euclidean spaces any inner-product
linear space, be it finite-dimensional or not. It is necessary to adapt euclidean
properties to the infinite-dimensional case and then consider linear objects.3
We shall below list the main results in words as near as possible to those
describing the finite dimensional case, while stressing the notions, here more
delicate, of continuity and boundedness.
In general a functional is a mapping from a topological linear space into
C. The mapping W is a linear functional when
W [f + g] = W [f ] + W [g] and W [kf ] = kW [f ].
2
Kolmogorov & Fomin 1977.

For a very good short introduction to analysis in infinite-dimensional spaces, see
Marsden 1974.
3
7.2. B GENERAL FUNCTIONALS
7.2.3
423
Operators
Let X and Y be two topological vector spaces. Then, an operator is any

mapping M : X Y . It is a linear operator if M [f + g] = M [f ] + M [g]
and M [kf ] = kM [f ]. It is continuous at a point f0 of X if, for any open
set V Y around M [f0 ], there exists an open set U around f0 such that
M [f ] V whenever f U . It is continuous on the space X if it is continuous
at each one of its points. For normed spaces, this is equivalent to saying that,
for any tolerance > 0, there is a spread > 0 such that ||f1 f2 || < implies
||M f1 M f2 || < . Recall (Math.4) that a subset U of a topological vector
space is a bounded set if, for any neighbourhood V of the space origin, there
exists a number n > 0 such that nV U . The mapping-operator M is a
bounded operator if it takes bounded sets into bounded sets. It so happens
that every continuous operator is bounded.
A linear functional is a particular case of operator, when Y = C. We may
thus pursue our quest by talking about operators in general.
7.2.4
Derivatives Fr
echet and Gateaux
Let X and Y now be two normed spaces and M an operator M : X Y.

The operator M will be a (strongly) differentiable operator at f X if there
exists a linear bounded operator Mf0 such that
M [f + g] = M [f ] + Mf0 [g] + R[f, g],
(7.6)
where the remainder R[f, g] is of second order in g, that is to say,

||R[f,g]||
||g||
when ||g|| 0. When this happens, Mf0 [g] Y is the strong differential
(or Frechet differential ) of M at f ; the linear operator Mf0 is the strong
derivative, (or Frechet derivative), of M at f . The theorem for the derivative
of the function of a function holds for the strong derivative. We may define
another differential: the weak differential (or Gateaux differential ) is

d
M [f + tg] M [f ]
Df M [g] =
M [f + tg]
= lim
,
(7.7)
t0
dt
t
t=0
the convergence being understood in Y s norm topology. In general, Df M [g]
is not linear in g. When Df M [g] happens to be linear in g, then the linear
operator Df M is called the weak derivative, (or Gateaux derivative) at f .
The theorem for the derivative of the function of a function does not hold in
general for the weak derivative.
424
If M is strongly differentiable, then it is weakly differentiable and both

differentials coincide. But the inverse is not true. The existence of the weak
differential is a warrant of the existence of the strong differential only if Df M
is continuous as a functional of f . Anyhow, when the strong differential
exists, one may use for it the same expression
d

M
[f
+
tg]
,
dt
t=0
which is frequently more convenient for practical calculational purposes.
The practical use of all that will be illustrated in Math.8, with applications to Physics, more precisely to lagrangian field theory.
Lanczos 1986
Kobayashi & Nomizu l963
Math.Topic 8
FUNCTIONAL FORMS
0 Introduction
A EXTERIOR VARIATIONAL CALCULUS
1
2
3
4
5
6
7
8
Lagrangian density
Variations and differentials
The action functional
Variational derivative
Euler Forms
Higher order Forms
Relation to operators
Continuum Einstein convention
B EXISTENCE OF A LAGRANGIAN
9 Inverse problem of variational calculus

10 Helmholtz-Vainberg theorem
11 Equations with no lagrangian
a Navier-Stokes equation
b Korteweg-de Vries equation
C BUILDING LAGRANGIANS
12 The homotopy formula

13 Examples
a The Helmholtz-Korteweg lagrangian
b Born-Infeld electrodynamics
c Einsteins equations
d Electrodynamics
e Complex scalar field
f Fermion, second order
14 Symmetries of equations
425
426
MATH.TOPIC 8. FUNCTIONAL FORMS
8.1 Introduction
Exterior differential calculus is a very efficient means to compactify notation and reduce expressions of tensor analysis to their essentials. It has
been for long the privileged language of the geometry of finite dimensional
spaces, though it has become a matter of necessity for physicists only in recent times. On the other hand, functional techniques and the closely related
variational methods have long belonged to the common lore of Theoretical
Physics. What follows is a mathematically nave introduction to exterior
variational calculus,1 which involves differential forms on spaces of infinite
dimension in close analogy with the differential calculus on finite dimensional
manifolds. We insist on the word calculus because the approach will be
purely descriptive, operational and practical. It is of particular relevance in
field theory and specially helpful in clarifying the geometry of field spaces.
Only local aspects will be under consideration, meaning by that properties valid in some open set in the functional space of fields. As it has been
the case with differential forms on finite-dimensional manifolds, functional
forms may come to be of help also in the search of topological functional
characteristics.
Special cases of it have been diffusely applied to the study of some specific problems,
such as the BRST symmetry,2 and anomalies,3 but its scope is far more general.
8.1
8.1.1
A Exterior variational calculus
Lagrangian density
8.2 Let us consider a set of fields = {a } = (1 , 2 , . . . , N ) defined on

a space M which, to fix the ideas, we shall suppose to be the Minkowski
spacetime, even though, as some examples will make clear, all that follows
is easily adaptable to fields defined on other manifolds. The word field
is employed here with its usual meaning in Physics: it may be a scalar, a
vector, a spinor, a tensor, etc. What is important is that it represents an
infinity of functions, vectors, etc, and describe a continuum infinity of degrees
of freedom. Each kind of field defines a bundle with M as base space, and
we shall take local coordinates (x , a ) on the bundle. As in the bundles we
have already met, the basic idea is to use the xs and the s as independent
coordinates:4 the base space coordinates x and the functional coordinates
1
2
3
4
Olver 1986; Aldrovandi & Kraenkel 1988.

Stora 1984; Zumino, Wu & Zee 1984; Faddeev & Shatashvilli 1984.
Bonora & Cotta-Ramusino 1983.
See, for instance, Anderson & Duchamp 1980.
8.1. A EXTERIOR VARIATIONAL CALCULUS
427
a . If L[(x)] is a lagrangian density, its total variation under small changes

of these extended, or bundle, coordinates will be
T L = L + dL = (a L) a + ( L) dx ,
(8.1)
f
L
a
is the
where a L is a shorthand for
a (in analogy to f = x ) , and
a
purely functional variation of . We are thus using a natural basis for these
1-forms on the bundle. Vector fields (in the geometrical sense of the word)
can be introduced, and the set of derivatives {

a } can be used as a natural
local basis for them, dual to {a }. A general Field X (we shall call Field,
with capital F, geometrical fields on the functional space) will be written as
X = X a
a . In the spirit of field theory, all information is contained in the
fields, which are the degrees of freedom. Consequently, L will be supposed
to have no explicit dependence on x , so that
L = (a L) a .
Of course, for the part concerned with variations in spacetime (i.e., in the
arguments of the fields), we have usual forms and
d2 L =
8.1.2
1
2
[ L L] dx dx = 0.
(8.2)
Variations and differentials
8.3 This is precisely one of the results of exterior calculus which we wish
to extend to the -space. This extension is natural for the operator:
2L =
1
2
[a b L b a L] a b = 0.
(8.3)
Here a b is the antisymmetrization of the product a b , just the

exterior product of the differentials of the coordinates a and b . We
proceed to strengthen the parallel with usual smooth forms. In order to
enforce the boundary-has-no-boundary property for the total variation, we
must impose
T2 = ( + d)2 = d + d = 0.
(8.4)
But
T2 L = (d + d)L = a a ( Ldx ) + dx (a La )
= (a L a L)dx a dx ,
so that (8.4) requires
a L = a L.
(8.5)
The anticommutation of and d implies the commutation of the respective

derivatives.
428
Comment 8.1.1 This behaviour is no real novelty, as it appears normally in differential calculus when we separate a manifold into two subspaces.
8.4 The total variation T does not commute with spacetime variations,
as
[ , T ]f = ( x ) f,
(8.6)
but the purely functional variation does.
8.1.3
The action functional
8.5 We shall from now on consider only purely functional variations. Furthermore, instead of densities as the L above, we shall consider only objects
integrated on spacetime, such as the action functional
Z
(8.7)
S[] = d4 xL[(x)].
For simplicity of language, we shall sometimes interchange the terms lagrangian and action. Differentiating the expression above,
Z
Z
4
S[] = d xL[(x)] = d4 x{a L[(x)]}a (x).
(8.8)
8.1.4
Variational derivative
8.6 We shall suppose that the conditions allowing to identify the strong
and the weak derivatives are satisfied (see Math.7). Thus, the differential5
of a functional F [] at a point of the function space along a direction (x)
in that space will be defined by

dF [ + ]
F [ + ] F []
0
=
F [] = lim
.
(8.9)
0

d
=0
It is a linear operator on (x), and F 0 [] = 0 is a linearized version of the
equation F [] = 0. The integrand in (8.8) is the Frechet derivative of L[]
along = /. The presence of the integration, allied to property (8.5),
justifies the usual procedures of nave variational calculus, such as taking
variations (in reality, functional derivatives and not differentials) inside the
common derivatives which, allied to an indiscriminate use of integrations by
parts (that is, assuming convenient boundary conditions), lends to it a great
simplicity.
5
The expression given is actually that of the weak (Gateaux) differential. When the
strong derivative exists, so does the weak and both coincide. We suppose it to be the case,
and use the most practical expression.
8.1. A EXTERIOR VARIATIONAL CALCULUS
8.1.5
429
Euler Forms
8.7 The vanishing of the expression inside the curly bracket {} in (8.8)
gives the field equations, a L[(x)] = 0. Given a set of field equations
Ea [(x)] = 0, we shall call its Euler Form the expression
Z
E[] = d4 xEa [(x)]a (x).
(8.10)
The exterior functional (or variational) differential of such an expression will
be defined as
Z
E[] = d4 xEa [(x)] a (x)
R
= 12 d4 x{b Ea [(x)] a Eb [(x)]}b (x) a (x). (8.11)
The differential of (8.8) is immediately found to be zero: 2 S[] = 0.
8.1.6
Higher order Forms
8.8 In analogy to the usual 1-forms, 2-forms, etc, of exterior calculus, we

shall call 1-Forms, 2-Forms, etc, with capitals, the corresponding functional
differentials such as (8.8) and (8.11). A p-Form will be an object like
Z
1
Z[] = p!
d4 xZa1 a2 ...ap [(x)]a1 (x) a2 (x) . . . ap (x), (8.12)
the exterior product signs indicating a total antisymmetrization quite analogous to that of differential calculus.
8.1.7
Relation to operators
8.9 A thing which is new in Forms is that their components in a natural

coframe {a } as above may be operators, in reality acting on the first a
at the right. Take, for instance, the Euler-Form for a free scalar field,
Z

E[] = d4 x u
tx + m2 a (x)a (x).
(8.13)
Its differential will be
Z

E[] = d4 x{ab u
tx + m2 }a (x) b (x) = 0,
(8.14)
430
because the component {ab [u

tx + m2 ]} is a symmetric operator. The vanishing of the dAlembertian term may be seen, after integration by parts, as
a consequence of
a (x) a (x) = 0.
The use of operatorial components provides an automatic extension to the
larger space containing also the field derivatives, avoiding the explicit use of
jet bundles of rigorous variational calculus.6
8.10 Continuum Einstein convention
The indices {aj } in (8.12) are, of course, summed over, as they are repeated. To simplify notation, we shall from now on extend this Einstein
convention to the spacetime variables x and omit the integration sign, as
well as the (p!) factor. Its implicit presence should however be kept in mind,
as integration by parts will be frequently used. In reality, to make expressions shorter, we shall frequently omit also the arguments. Equation (8.10),
for example, will be written simply
E[] = Ea a .
(8.15)
Finally, we shall borrow freely from the language of differential calculus: a

Form W satisfying W = 0 will be said to be a closed Form, and a Form W
which is a variational differential of another, W = Z, will be an exact Form.
8.2
B Existence of a lagrangian
8.2.1
Inverse problem of variational calculus
In rough terms, the fundamental problem of variational calculus is to find

equations (the field equations) whose solutions lead some functional (the action functional) to attain its extremal values. The inverse problem of variational calculus is concerned with the question of the existence of a lagrangian
for a given set of field equations. It can then be put in a simple way in terms
of the Euler Form E: is there a 0-Form S, as in (8.7), such that E = S?
Or, when is E locally an exact Form?
8.2.2
Helmholtz-Vainberg theorem
8.11 Consider the expression (8.11). The Ea [(x)] are densities just as
L[(x)], and the differentials appearing are Frechet differentials,
Ea = {b Ea []} b = Ea0 [].
6
Anderson & Duchamp 1980.
(8.16)
8.2.
B EXISTENCE OF A LAGRANGIAN
431
As said, Ea0 [] = 0 is a linearized version of the equation Ea [] = 0. The

Helmholtz-Vainberg necessary and sufficient condition7 for the existence of a
local lagrangian8 is that, in a ball around in the functional space,
a Ea0 [] = a Ea0 []
(8.17)
for any two increments , . In our notation, with increments a along a

and b along b , (8.16) tells that this is equivalent to b Ea = a Eb or, from
(8.11),
E = 0.
(8.18)
8.2.3
Equations with no lagrangian
We shall see in next section a variational analogue of the Poincar inverse

lemma of differential calculus: for a Form to be locally exact, it is necessary
and sufficient that it be closed. In this case, Ea = a L for some L. There
are, however, equations of physical interest which are not related to an action
principle in terms of the fundamental physical fields involved.
8.12 Navier-Stokes equation
Let us look at the notorious case of the equation
t vi + vj j vi + i p j j vi = 0,
(8.19)
which, together with the incompressibility condition

i v i = 0
(8.20)
describes the behaviour of an incompressible fluid of density and coefficient

of viscosity . We shall consider the stationary case, in which the first term
in (8.19) vanishes. The physical fields of interest are the velocity components
v j and the pressure p. We learn from Fluid Mechanics that the pressure is
the Lagrange multiplier for the incompressibility condition, so that we write
the Euler Form as
E = [vj j vi + i p j j vi ]v i (j v j )p = 0,
(8.21)
with the relative sign conveniently chosen. After putting E under the form
E = (vj j vi )v i + [ 21 (j vi j v i ) p(j v j )],
7
8
Vainberg 1964.
Aldrovandi & Pereira 1986, 1988.
432
a direct calculation shows that

E = [(vj j vi )v i ] 6= 0.
(8.22)
The offending non-lagrangian term can be immediately identified as (vj j vi ).

The power of exterior variational calculus is well illustrated in these few lines,
which summarizes the large amount of information necessary to arrive at this
result.9
8.13 Korteweg-de Vries equation
This is another example of interest, for which the Euler Form is given by
E = (ut + uux + uxxx ) u,
(8.23)
the indices indicating derivatives with respect to t and x. That no lagrangian

exists can be seen from the simple consideration, for instance, of the first term
in E, given by ut u, which is nonvanishing and cannot be compensated
by any other contribution. This example illustrates an important point: the
existence or not of a lagrangian depends on which field is chosen as the
fundamental physical field. Above, such field was supposed to be u. In terms
of u no lagrangian exists. However, a lagrangian does exist in terms of some
if we put u = x , in which case E becomes the closed Form
E = (tx + x xx + xxxx ) .
(8.24)
However, when the choice of the fundamental physical field is given by some
other reason, as in quantum field theory, it is of no great help that a lagrangian may be found by some smart change of variable.
There is an obvious ambiguity in writing the Euler Form for a set of two
or more field equations, as multiplying each equation by some factor leads
to an equivalent set. Such a freedom may be used to choose an exact Euler
Form and to give the lagrangian a correct sign, for example leading to a
positive hamiltonian.
8.3
C Building lagrangians
8.3.1
The homotopy formula
8.14 Each differential form is given locally by a very convenient expression

in terms of a differential and a transgression, which embodies the Poincare
9
Finlayson 1972.
8.3.
433
inverse lemma ( 7.2.12). We shall adapt that expression to Forms. Let us

begin by defining the operation T on the p-Form Z. If
Z[] = Za1 a2 ...ap [] a1 a2 . . . ap ,
(8.25)
then T Z is defined as the (p 1)-Form given by

P
T Z[] = pj=1 ()j1
Z
dt tp1 Za1 a2 ...ap [t ] aj a1 a2 . . . aj1 aj+1 . . . ap . (8.26)
The fields a appearing in the argument of Za1 a2 ...ap are multiplied by the
variable t before the integration is performed. As t runs from 0 to 1, the
field values are continuously deformed from 0 to a . This is a homotopy
operation10 in -space. A more general homotopy t = t + (1 t) 0
with 0 6= 0 can be used, but without real gain of generality. The important
point is that the -space is supposed to be a starshaped domain around some
zero field (each point may be linked to zero by straight lines). Spaces of
this kind are called affine spaces by some authors. Some important field
spaces, however, are not affine. For example, the space of metrics used in
General Relativity includes no zero, nor does the space of chiral fields with
values on a Lie group. For such cases, the use of (8.26) is far from immediate
(see Phys.7).
8.15 The Poincare inverse lemma says that, on affine functional spaces, Z
can always be written locally as
Z[] = (T Z) + T (Z).
(8.27)
This result may be obtained from (8.26) by direct verification. A consequence

is that a closed Z will be locally exact: Z = (T Z). For a closed Euler Form
E, this gives immediately the Vainberg homotopy formula, which gives the
lagrangian as
L = T E.
(8.28)
As the operator T is the transgression operator, this expression is called
the transgression formula. It provides a systematic procedure to find a
lagrangian for a given equation, when it exists. Equation (8.27) allows furthermore a systematic identification of those pieces of a given E which are
lagrangian-derivable and those which are not. This was done directly in
10
Nash & Sen 1983.
434
(8.22) , but (8.27) may be useful in more complicated cases. No term in

(8.23) is lagrangian-derivable, since there it happens that
T E = E
T E = 0
(8.29a)
(8.29b)
When L does exist, a trivial rule to obtain it from E = L = Ea a comes out

when Ea is a polynomial in the fields and/or their derivatives: replace in E
each factor a by a and divide each monomial of the resulting polynomial
by the respective number of fields (and/or their derivatives).
8.3.2
Examples
8.16 The Helmholtz-Korteweg lagrangian

If the first term in [8.21] is dropped, the remaining terms would come
from
= 12 (j vi j v i ) (pj v j ).
(8.30)
As to eq.[8.24], it comes immediately from the lagrangian
= 12 tx + 13 x xx + 21 xxxx .
(8.31)
8.17 Born-Infeld electrodynamics

A simple example of the use of [8.26] in a non-polynomial theory may be
found in the Born-Infeld electrodynamics.11 With F = A A and
F 2 = F F , its Euler Form is
"
E=
F 2 /(2k)
A .
(8.32)
In this case,

TE = A

R1
0
dt
t F
1t2 F 2 /(2k)
gives, after an integration and a convenient antisymmetrization,

p
= k{ 1 F 2 /(2k) 1}.
11
Born & Infeld 1934.
(8.33)
8.3.
435
8.18 Einsteins equations

It is sometimes possible, by a clever picking-up of terms, to exhibit the
Euler Form directly as an exact Form, thereby showing the existence and
the explicit form of a lagrangian. Take Einsteins equations for the pure
gravitational field. Its Euler Form is

(8.34)
E = g R 12 g (R + ) g ,
with the cosmological constant. We can recognize g = 12 g g g

in the second term and separate
R = g R g R
to write
E = [ g(R + )] g g R .
Of these two terms, the latter is known to be a divergence12 and the first
exhibits the Hilbert-Einstein lagrangian. The factor g is to be expected if

we recall the implicit integration in [8.34]. It plays the role of an integrating
factor, as E would be neither invariant nor closed in its absence.
We next give a few more examples of Euler Forms and lagrangians, taken
from field theory (see Phys.6):
8.19 Electrodynamics
The Euler Form for a 1/2-spin field in interaction with the electromagnetic
field is
E = [i ( ieA ) m] [i( + ieA ) + m]
+ [ F + e ] A .
The corresponding lagrangian is
= 21 {i( ( ieA ) i[( + ieA )] } m 14 F F
= 21 {i i[ ] } m + eA 41 F F .
8.20 Complex scalar field
In this case, the Euler Form is given by

E = u
t ie( A ) 2ieA e2 A A

+ u
t + ie( A ) + 2ieA e2 A A ,
and the lagrangian is
12
Landau & Lifshitz 1975.
436
L = [ + ieA ] [ ieA ] + m + 4 41 F F .
8.21 Second order fermion equation
Applying twice the Dirac operator, we obtain a second order equation
for the fermion, which includes a spin-field coupling introduced by Fermi to
account for the anomalous magnetic moment of the neutron. In this case,
i
e
E = u
t ie( A ) 2ieA e A A F
2
i
h
e
2
+ u
t + ie( A ) + 2ieA e A A F
2

+ F + e + e ( ) A .
h
The Fermi term

EF = 2e [ F ] = e [ F A ] = e [A [ F ]
is an example of non-minimal coupling. The interaction lagrangian is, of
course,
LF = 2e F .
An interesting remark is that, in this case, the current j is given by the
complete Lagrange derivative
j =
and not simply by
theories.
8.3.3
LF
A
LF
A
, as it would be usual for the matter currents in gauge
Symmetries of equations
8.22 Other notions from differential calculus can be implemented in the

calculus of Forms. One such is that of a Lie derivative. Recall that a general
field X will be written X = X a /a . Suppose that X represents a transformation generator on the -space. On Forms, the transformation will be given
by the Lie derivative LX . The Lie derivative LX , acting on Forms, will have
properties analogous to those found in differential calculus. In particular, it
commutes with differentials, so that
LX E = LX = LX .
(8.35)
Consequently, a symmetry of the lagrangian (LX = 0) is a symmetry of the

equation ( LX E = 0), but the equation can have symmetries which are not
8.3.
437
symmetries of the lagrangian. This is a well known fact, but here we find a
necessary condition for that: LX = 0. Still other notions of differential
calculus translate easily to Forms, keeping quite analogous properties. Such
is the case, for example, of the interior product iX W of a field X by a Form
W , which has the usual relation to the Lie derivative,
LX W = iX (W ) + (iX W ).
8.23 We have shown some examples of the power of exterior variational
calculus in treating in a very economic way some involved aspects of field
theories. All examined cases were local, valid in some open set of the field
space. Recent years have witnessed an ever growing interest in the global,
topological properties of such spaces. Anomalies, BRST symmetry and other
peculiarities (see Phys.7) are now firmly believed to be related to the cohomology of the field functional spaces, this belief coming precisely from results
obtained through the use of some special variational differential techniques.
Many global properties of finite dimensional manifolds are fairly understood
and transparently presented in the language of exterior differential forms.
The complete analogy of the infinite dimensional calculus suggests that, besides being of local interest, it is a natural language to examine also global
properties of field spaces.
Aldrovandi & Kraenkel 1988

Marsden 1974
Olver 1986
438
Math.Topic 9
SINGULAR POINTS
1 Index of a curve
2 Index of a singular point
3 Relation to topology
4 Examples
5 Critical points
6 Morse lemma
7 Morse index and topology
8 Catastrophes
9.1
Index of a curve
Given a vector field X on a smooth manifold M , the point p M will

be a singular point of X if Xp = 0. Singular points1 of a vector field give
information on the underlying topology. There may be fields without singular
points on the euclidean plane, but not on the sphere S 2 . The point p will
be a critical point of the function f if it is a singular point of the gradient
of f , X = ( i f )i . As the gradient is actually a 1-form, df = (i f )dxi , this
supposes a metric to introduce X as the contravariant image of the differential
form df . Recall that the presence of singular points on M signals a nontrivial tangent bundle. We shall in what follows (except in section Math.9.8)
suppose that singular points, if existent, are always non-degenerate. Let us
begin with the simplest non-trivial case, which occurs when M = E2 . Take a
field X and a singular point p E2 . Take another point q E2 and suppose
it to move around p, describing a closed curve never touching p. Let us fix
a point q0 on the curve and follow the field Xq as q moves along . It is Xq0
at the start, and is Xq0 again when q arrives back at q0 . As it travels along ,
1
A very complete treatment of the subject is given in Doubrovine, Novikov & Fomenko
1979, Vol. II, 13-15.
439
440
MATH.TOPIC 9. SINGULAR POINTS
X will turn around itself a certain number of times, both in the clockwise
sense (taken by convention as negative) and in the counterclockwise sense
(positive by convention). The algebraic sum of this number of turns is the
index of the curve with respect to the field X. For instance, the index equals
+1 in the case pictured in Figure 9.1. If it so happens that p is not a singular
point of the field X, we can always find, thanks to the continuity of X, a
small enough neighbourhood of p inside which all curves have null index.
In Figure 9.2 we show an example: in the complex plane version of E2 , the
behaviour of the field
X(z) = z 2 = (x + iy)2 = x2 y 2 + i2xy = (x2 y 2 ) x + 2xy y ,
when it traverses the circle |z| = 1 around its singular point z = 0. The index
is + 2. It is easy to see that, had we taken a point outside the curve, the
index would be zero. In general, the circle will have index (+n) with respect
to the field X(z) = z n and (n) with relation to the field X(z) = z n (which
illustrates the role of the orientation). A practical way to find the index is to
draw the vectors directly at the origin and follow the angle it makes with
R+ .
Figure 9.1:
Let us formalize the above picture: suppose a chart (U, c) and a field
X = X 1 e1 + X 2 e2 .
Let U 0 = U { singular points of X}, and p U 0 . Define the mapping
f : U 0 S 1 , from U 0 into the unit circle given by
f (p) =
Xp
|Xp |
9.1. INDEX OF A CURVE
441
Figure 9.2:
In the case X(z) = z 2 , with p = z,

f (z) =
x2 y 2
x2 +y 2
+ i x22xy
= X 1 + iX 2 .
+y 2
Then, in a neighbourhood of f (p) S 1 , let us take a local angular coordinate

, circf (p). Such a coordinate (Figure 9.3), as said in 4.2.5, does not cover
the whole plane including S 1 : its inverse is discontinuous and leaves out
the axis = 0. Two charts are actually necessary, each one covering the
axis left out by the other. In the intersection, a coordinate transformation
= 0 + , with constant , is defined. As their difference is a constant, the
coordinate differentials do coincide outside the origin: d = d0 . Thus,
Figure 9.3:
d = d arctan xx1 =
x2 dx1 x1 dx2
|x|2
442
is a differential form well defined on the whole plane outside the singular
points. The index of a closed curve : S 1 U 0 is then defined as
I
1
ind =
d
(9.1)
2
In the example X(z) = z 2 ,
f = arctan x22xy
y 2
It is easier to define z = rei , X(z) = r2 e2i , so that f = 2 . While goes
from 0 to 2, f goes from 0 to 4. Consequently,
H
1
ind = 2
d(2) = 2.
Notice an important thing: the mapping f (z), defined through the field
X, takes two points of U 0 in one same point of S 1 .
The index is just the number of points from the domain of f taken into
one point of its image. This is a general property, which will allow the
generalization to higher dimensional manifolds. Let us list some results which
can be proven about the indices:
(i) they do not change under continuous deformations of the curve ,
provided never touches any singular point;
(ii) they do not change under continuous deformations of the field, provided X never has a singular point on the curve;
(iii) every disk whose contour has a non vanishing index related to a field
contains some singular point of that field;
(iv) the index of a curve contained in a small enough neighbourhood of
a singular point is independent of the curve, that is, it is the same for any
curve; the index so obtained is the index of the singular point; this index is
chart-independent;
(v) if a curve encloses many singular points, its index is the sum of the
indices of each point. A good example is given by the field V (z) = z 2 (z/2);
it has two singular points in the unit disk. The index of z = 0 is 2, that of z
=1/2 is zero, and the index of the unit curve |z|2 = 1 is 2;
(vi) the index of a singular point is invariant under homeomorphism; this
allows the passage from the plane to any bidimensional manifold, as it is a
purely local property.
9.2
Index of a singular point
In order to generalize all this to singular points of fields on general differentiable manifolds, let us start by recalling what was said in section 5.3: any
9.3. RELATION TO TOPOLOGY
443
differentiable manifold M of dimension m can be imbedded in an euclidean

manifold of high enough dimension. Consider p a singular point of a field
X on M . Around it, there will be a neighbourhood U diffeomorphic to Em ,
f (U ) Em . We may transfer X to Em through the differential mapping f
and consider a sphere S m1 around f (p) with a radius so small that p is the
only singular point inside it. To simplify matters, let us forget the diffeomorphism and write simply p for f (p), Xp for Xf (p) , etc. With this
notation, define then the mapping h : S m1 S m1 by
h(p) =
Xp
f [Xp ]
=: r
.
|f [Xp ]|
|Xp |
(9.2)
The index of the singular point is the Brouwer degree (6.2.15) of this mapping.
9.3
Relation to topology
On a compact manifold, the number of singular points of a fixed field X is

finite. Still a beautiful result:
the sum of all the indices of a chosen vector field on a compact
differentiable manifold M equals the Euler-Poincare characteristic of M .
In this way we see that this sum is ultimately independent of the field
which has been chosen it depends only on the underlying topology of the
manifold. As a consequence, on a manifold with 6= 0, each field must have
singular points! Information on the topology of a manifold can be obtained
by endowing it with a smooth structure and analyzing the behaviour of vector
fields. We had seen other, more direct means to detect defects, holes, etc by
lassoing or englobing them. The present differential method (which points to
differential topology) gives another way and can be pictured out by the image
of throwing a fluid through the manifold and looking for sinks, whirlpools,
sources, etc.
9.4
Basic two-dimensional singularities
In the two dimensional case, drawing the local integral lines is of great help
to get intuition on the corresponding fields. Figure 9.4 shows some simple
singular points, with their names and corresponding indices. Notice that the
index does not change if we invert the field orientation. Figure 9.5 shows
444
two other singularities of great importance: the saddle point and the dipole.
Sources, crosspoints and sinks may be taken as elementary singular points.
For 2-dimensional manifolds, we may cast a bridge towards the homological version of the Euler number by taking a triangulation and putting (i) a
source at each vertex, (ii) a crosspoint replacing each edge, and (iii) a sink
at the center of each loop.
Figure 9.4:
9.5
Critical points
Notice that, from the drawings above, the index is not enough to characterize completely the kind of singular point. Such a characterization requires
further analysis, involving a field linearization: near the singular point, it
is approximated so as to acquire a form x = Ax, x being a set of coordinates
and A a matrix. The eigenvalues of A provide a complete classification.2
On n-dimensional metric spaces, as said, connection may be made with the
critical points of functions, which are singular points of their gradient fields.3
Let us go back to the differentiable function f : M N . We have
defined its rank as the rank of the jacobian matrix of its local expression in
2
Arnold 1973, chap.3.

A huge amount of material on recent developments, mainly concerned with dynamical
systems, is found in Guckenheimer & Holmes 1986.
3
9.6. MORSE LEMMA
445
Figure 9.5:
coordinates. This is clearly a local concept. The points of M in which this

rank is maximal, that is, rank f = min (m, n) are called regular points of f .
The points at which rank f < min (m, n) are critical points (or extrema)
of f . The study of functions in what concerns their extrema is the object of
Morse theory, of whose fascinating results we shall only say a few words.4
9.6
Morse lemma
Consider differentiable real functions f : N E1 . Let p N be a critical

point of f . The critical point p is non-degenerate if the hessian matrix of
the composition of functions x<1> , f and z,
2

(z f x<1> )
(9.3)
xi xj
x(p)
is non-singular for some pair of charts (U, x) and (V, z) on N and E1 respectively. Differentiability conditions ensure that in this case the non-degeneracy
is independent of the choice of the charts. Here comes a first result, the
Morse lemma: in the non-degenerate case, there exists a chart (W, y) around
4
Milnor 1973.
446
the critical point p such that y(p) = 0 and, for y y(W ), the function
f y 1 (y1 , y2 , . . . , yn ) can be written as a quadratic form:
2
2
f y 1 (y1 , y2 , . . . , yn ) = f (p) y12 y22 . . . yk2 + yk+1
+ yk+2
+ . . . + yn2 (9.4)
for some k, with 0 k n. Notice that f y 1 is just the expression of f in

local coordinates, as we are using the trivial chart (E1 , identity mapping) for
E1 . The integer k, the number of negative signs in the quadratic form, is the
Morse index of the critical point. When k = n, the point is a maximum of f ,
as in the quadratic form all the other points in the neighbourhood give lesser
values to f . When k = 0, p is a minimum. Otherwise, it is a saddle-point.
The integer k is independent of the choice of coordinates because it is the
signature of a quadratic form. The relation with the (singular point) indices
is as follows: taking the Morse quadratic form and studying its gradient, we
find that minima have index = + 1, maxima have index = ()n , and saddle
points have alternate signs, 1.
The Lemma has a first important consequence: in the neighbourhood W
in which the quadratic expression is valid, there is no other critical point.
Thus,
each non-degenerate critical point is isolated.
Another important consequence, valid when N is compact and f has only
non-degenerate critical points, is that
the number of critical points is finite.
This comes, roughly speaking, from taking a covering of N by including
charts as the (W, y) above, one for each critical point, and recalling that any
covering has a finite subcovering in a compact space.
Take N = S 2 , the set of points of E3 satisfying x2 + y 2 + z 2 = 1. The
projection on the z axis is a real function, z = (1 x2 y 2 )1/2 . It has a
maximum at z = 1, around which z 1 (x2 + y 2 )/2 (so, index 2), and a
minimum at z = 1, around which z 1 + (x2 + y 2 )/2 (so, index 0).
9.7
Morse indices and topology
Another enthralling result links the critical points of any smooth function to
the topology of the space: if N is compact, and nk denotes the number of
critical points with index k, then
n
X
i=0
()k nk = (N ),
(9.5)
9.8. CATASTROPHES
447
the Euler characteristic of N . This means in particular that the sum is

independent of the function f : every function will lead to the same result.
In order to know the Euler characteristic of a space, it is enough to examine
the critical points of any real function on it. Using the previous example,
one finds immediately (S 2 ) = 2.
The relationship of critical points to topology is still deeper. Again for N
compact, each number nk of critical points with index k satisfies the Morse
inequality
nk bk (N )
(9.6)
with bk (N ) the k-th Betti number of N . All this is only meant to give a flavor
of this amazing theory. There are stronger versions of these inequalities, like
the polynomial expression
Pn
Pn
k
k
i=0 (nk bk )t = (1 + t)
i=0 (qk t
with each n 0, from which the above expression for (N ) comes out
when t = 1. Summing up, all we want here is to call the attention to
the strong connection between the topology of a differentiable manifold and
the behaviour of real functions defined on it. For example, take the torus
imbedded in E3 , as in Figure 9.6, and consider the height function, given by
the projection on the z-axis. It has one maximum, one minimum and two
saddle points. Thus,
(T 2 ) = ()2 1 + ()1 2 + ()0 1 = 0.
9.8
Catastrophes
Morse theory shows a kind of stability concerning isolated singularities. The

index of each critical point is fixed, and immediately exhibited by [9.4] once
the good system of coordinates is found. It was a fantastic discovery that
much of this stability keeps holding when the non-degeneracy condition is
waived. In this case the critical points are no more isolated they constitute
lines of singularity. After what we have said, we could expect lines of minima,
or maxima, or saddle points. Any function is approximated by the unique
Morse quadratic form around each critical point. It was found by Thom5 that
5
Thom 1972. The main stimulus for Thom came from optics and the idea of form
evolution in biology. A more readable text is Thom 1974. Zeeman has applied the theory
to biology, medicine, social sciences, etc. His works are collected in Zeeman 1977, where
much material and references can be found. A general appraisal of the theory, as well as
of the controversy its applications have raised, is found in Woodcock & Davis 1980.
448
Figure 9.6:
functions with lines of singularities (catastrophes) are also described by

elementary expressions, few-variables polynomials, around the singularities.
The general expression is not unique, but it is always one of a few basic forms,
dependent on the dimensions involved. These elementary catastrophes
have been completely classified for dimensions 5. In Optics, where they
appear as caustics, their existence, limited number and standard forms have
been beautifully confirmed.6 We might expect their avatars in many other
fields, as bifurcations in non-linear systems and in phase transitions related to
the vacuum (minimum of some potential) degeneracy. It should be noticed,
however, that the theory is purely qualitative. One can say that, in a given
physical system, the singularities must be there, and have such or such form,
but it does not tell at which scale nor when they will show up. Though
some initial successes have been achieved, the feeling remains that there is
much as yet to be done if we want to make of it a practical tool for physical
applications.
Arnold 1973
Doubrovine, Novikov & Fomenko 1979
Milnor 1973
Berry 1976.
Math.Topic 10
EUCLIDEAN SPACES AND SUBSPACES
0 Introduction
A STRUCTURE EQUATIONS
1 Moving frames
2 The Cartan lemma
3 Adapted frames
4 Second quadratic form
5 First quadratic form
B RIEMANNIAN STRUCTURE
6 Curvature
7 Connection
8 Gauss, Ricci and Codazzi equations
9 Riemann tensor
C GEOMETRY OF SURFACES
10 Gauss Theorem
D RELATION TO TOPOLOGY
11 The Gauss-Bonnet theorem
12 The Chern theorem
0 Introduction
Euclidean spaces are, as we have seen, the fundamental spaces to which
manifolds are locally homeomorphic. In addition, differentiable manifolds
can always be imbedded in some euclidean space of high enough dimension
(section 5.3). For a n-dimensional manifold N , the Whitney theorem only
guarantees that this is always possible in a (2n + 1)-dimensional space, but
sometimes a smaller dimension is enough: for instance, S 2 may be imbedded
449
450
MATH.TOPIC 10. EUCLIDEAN SPACES AND SUBSPACES
in E3 . On the other hand, one more dimension is not always sufficient: the 4dimensional Schwarzschild space, solution of Einsteins equations, cannot be
imbedded in E5 . No general rule giving the minimal dimension for euclidean
imbedding is known. We shall here consider the manifold N immersed in
some En+d , with d large enough. This will allow us to touch on some results
of what is nowadays called classical differential geometry.
10.1
A Structure equations
10.1.1
Moving frames
Recall (section 7.3) that, given a moving frame {ei } on En+d , the structure
equations are
d j = j i i ;
j
d i = i k .
10.1.2
(10.1)
(10.2)
The Cartan lemma
An important general result is the Cartan lemma: let {i , i = 1, 2, . . . , r

m} be a set of rlinearly
1-forms on Em . If another set {i } of r
Pr independent
1-forms is such that i=1 i i = 0, then the i are linearly dependent on
them:
P
i = rk=1 ai k k ,
with aij = ik ak j = aji . With this lemma, it is possible to show that the set
of forms k i satisfying both ij = - ji and equation [10.1] is unique.
10.1.3
Adapted frames
An imbedding i : N En+d is a differentiable mapping whose differential

dip : Tp N En+d is injective around any p N . The inverse function theorem says then that a neighbourhood U of p exists such that i|U (i restricted to
U ) is also injective. It is possible to show that there exists a neighbourhood
V i(U ) En+d small enough for a basis {e1 , e2 , . . . , en , en+1 , . . . , en+d } to
exist with the following property: the first n base members (e1 , e2 , . . . , en )
are tangent to i(U ), and the remaining fields are normal to i(U ). Such a
frame always exists and is called a frame adapted to the imbedding. Notice
that this implies in particular that the commutators of the first n fields are
written exclusively in terms of themselves.
10.1. A STRUCTURE EQUATIONS
451
Comment 10.1.1 We are used, in the simple cases we usually meet in current Physics,
to take a space (preferably euclidean), there fix a frame once for all, and refer everything
to it. We even lose the notion that some frame is always involved. This simple-minded
procedure fails, of course, whenever the space is somehow non-euclidean. The amazingly
simple idea of Cartan was to consider instead a bunch of moving frames, valid actually
in a neighbourhood of each point, and whose set will finally constitute the bundle of frames
(section 9.3).
The imbedding i induces forms i ( j ) and i ( j i ) on U . The pull-back
commutes with the operations of exterior product and exterior differentiation.
The basic point of the method of moving frames comes thereof: the structure
equations valid on V hold also on some open of N . It will be better to use
the indices i, j, k, . . . from 1 to n + d, as above; , , , . . . in the range 1 to n;
and indices a, b, c, . . . from n + 1 to n + d. Separating the structure equations
in an obvious way, they become
d = + a a ;
a
d =
d = + c c ;
(10.6)
c;
(10.7)
d b a = a b + c a b c .
(10.8)
=
=
(10.5)
(10.4)
(10.3)
c;
These equations hold on U N but, if applied only on fields u = u e

on U , they lose some terms because a (u) = 0 for every a. Equation [10.4]
reduces to a = 0 which, by the Cartan lemma, means that
a = ha ; ha = ha .
10.1.4
(10.9)
Second quadratic form
The second order symmetric form with the coefficients ha as components,

a = ha ,
(10.10)
is the second quadratic form of the imbedding along the direction a.
10.1.5
First quadratic form
The first quadratic form is the metric on U induced by the imbedding: given
u and v Tp U , this metric is defined by
< u, v >p := < ip (u), ip (v) > .
(10.11)
The metric and the fields {e } determine the and . We say then that
all these objects belong to the intrinsic geometry of U .
452
10.2
B Riemannian structure
10.2.1
Curvature
Let us compare [10.5] with [10.2]. The latter is valid for the euclidean space.
The 2-forms = c c measure how much N departs from an euclidean
space, they characterize its curvature. In terms of intrinsic animals, that is,
in terms of objects on N itself, they are, from [[10.5]],
= d .
(10.12)
They are the curvature forms on N . With the forms acting on the space
tangent to N , the structure equation [10.3] reduces to
d = .
10.2.2
(10.13)
Connection
It is convenient to include the forms in a matrix R, the in a matrix

and the in a column . The equations above become
R = d ,
d = .
(10.14)
(10.15)
Consider an orthonormal basis transformation given by e0 = A e and

0 = (A1 ) or, equivalently, = A 0 . In matrix language, e0 = Ae
and = A 0 . Taking differentials in the last expression,
d = dA 0 + Ad 0 = dA A1 + A0 0 = (dAA1 + A0 A1 ) .
From the unicity of forms satisfying [10.15],
= dAA1 + A0 A1 ,
or
0 = A1 dA + A1 A = A1 [d + ]A.
(10.16)
This is the very peculiar transformation behaviour of the connection form .

In the same way we find that the curvature form changes according to
0 = A1 A.
(10.17)
Matrix behaves, under base changes, in the usual way matrices do under
linear transformations.
10.2. B RIEMANNIAN STRUCTURE
10.2.3
453
Gauss, Ricci and Codazzi equations
Back to the structure equations, we notice the forms

b a = a b = d b a c a b c .
(10.18)
They are the normal curvature forms. This expression may be combined with
[10.9] and [10.13] to give the Gauss equation
X
(ha ha ha ha )
(10.19)
= 21
a
and the Ricci equation

b a = 12 (ha hb hb ha )
(10.20)
The imbedding divides the geometry into two parts, which are related
by the second quadratic forms and by eq.[10.6], which is called the Codazzi
equation:
d a = a + c a c .
(10.21)
The above equations constitute the basis of classical differential geometry.
The important point is that the geometrical objects on N (fields, forms,
tensors, etc) may be given a treatment independent of the exterior objects.
All the relations involving the indices , , , etc may be written without
making appeal to objects with indices a, b, c, etc. This fact, pointed out
by Gauss, means that the manifold N has its own geometry, its intrinsic
geometry, independently of the particular imbedding. This may be a matter
of course from the point of view we have been following, but was far from
evident in the middle of the nineteenth century, when every manifold was
considered as a submanifold of an euclidean space. The modern approach has
grown exactly from the discovery of such intrinsic character: the properties of
a manifold ought to be described independently of references from without.
10.2.4
Riemann tensor
In components, the curvature 2-forms will be written

= 21 R .
(10.22)
If the connection forms are written in some natural basis as

= dx ,
(10.23)
454
the components in [10.22] are obtained from [10.12]:

R = + .
(10.24)
These components constitute the Riemann curvature tensor. The components of the connection form [10.23] are the Christoffel symbols, which may
be written in terms of derivatives of the components of the metric tensor.
These metric components are, when restricted to the intrinsic sector,
g = e eg .
Of course, they are now point-dependent since the adapted basis vectors
change from point to point.
The Ricci tensor R = R is symmetric on a Riemannian manifold.
A manifold whose Ricci tensor satisfies R = g , with a constant, is
called an Einstein space. There are very interesting theorems concerning the
immersion of Einstein spaces. One of them is the following:
if an Einstein space as above has dimension m and is immersed
in Em+1 , then necessarily 0.
Another curious result is the following:
suppose that, on a connected manifold of dimension m, R = f g ,
with f a function; then, if m 3, f is necessarily a constant.
To get the Christoffel symbols (see Phys.8), we start by differentiating
the function g ,
dg = dx g = de e + e de
= e e + e e = g + g

= g + g dx .
Defining := g , we see that
g = + .
Calculating g + g g , we arrive at
= g = 12 g [ g + g g ] .
(10.25)
10.3. C GEOMETRY OF SURFACES
10.3
C Geometry of surfaces
10.3.1
Gauss Theorem
455
To get some more insight, as well as to make contact with the kernel of
classical geometrical lore, let us examine surfaces imbedded in E3 . Consider
then some surface S, dim S = 2, and an imbedding i : S E3 . Two vectors
u, v Tp S will have an internal product given by [10.11], the metric on
S induced through i by the euclidean metric of E3 . To examine the local
geometry around a point p S, take an open U, S U 3 p, and an open V
such that E3 V i(U ). Choose on V a moving frame (e1 , e2 , e3 ) adapted
to i in such a way that e1 and e2 are tangent to i(U ), and e3 is normal. The
orientation may be such that (e1 , e2 , e3 ) is positive in E3 . Here, to simplify
the notation, we take the imbedding as a simple inclusion, all the geometrical
objects on S being considered as restrictions to S of objects on E3 . Given
any vector v = v 1 e1 + v 2 e2 on S, it follows that 3 (v) = 0. Equation [10.4]
becomes
0 = d 3 = 3 1 1 + 3 2 2 .
It follows from the Cartan lemma that
3 1 = h11 1 + h12 2
3 2 = h21 1 + h22 2 ,
with
h12 = h21 ,
(10.26)
where we have used the simplified notation h3 ij = hij . Notice that

h11 = 3 1 (e1 ) ; h22 = 3 2 (e2 ) ; h12 = 3 1 (e2 ) = h21 = 3 2 (e1 ).
As j i = i j and dei = j i ej ,
de3 (v) = 3 1 e1 3 2 e2 .
This may be put into the matrix form
1

1
v
h11 h12
v
de3 2 =
.
v
h21 h22
v2
(10.27)
Thus, the matrix (h ) represents on basis (e1 , e2 ) the differential of the

mapping e3 : U E3 , p e(p)3 . As |e3 | = 1, this mapping takes values
on a unit sphere of E3 . It is called the Gauss normal mapping. The matrix
(h ) may be diagonalized with two real eigenvalues 1 and 2 . These
456
eigenvalues are the principal curvature radii of S at p. Its determinant is the

total curvature, or Gaussian curvature of S at the point p:
K := det(de3 ) = 1 2 = h11 h22 (h12 )2 .
(10.28)
A quick calculation using [10.19] and [10.28] shows that

2 1 = d 2 1 = K 1 2 .
(10.29)
The form 1 2 has a special meaning: applied to two vectors u and v, it

gives the area of the parallelogram they define: 1 2 (u, v) = u1 v 2 u2 v 1 .
It is the area element, which is in reality independent of the adapted frame
and defined on the whole S. It will be denoted
= 1 2 .
It corresponds, of course, to the volume form on S. Unlike , the connection
form 2 1 depends on the adapted frame. Let us proceed to a change from
the frame (e1 , e2 , e3 ) to another frame (e01 , e02 , e03 ), related to it by
e01 = cos e1 + sin e2 ,
e02 = sin e1 + cos e2 .
(10.30a)
(10.30b)
The dual basis will change accordingly,

0
1 = cos 1 + sin 2 ,
02
= sin 1 + cos 2 .
(10.31a)
(10.31b)
Taking the differentials and using the structure equations, we get

0
d 1 = 2 ( 1 2 + d),
d 2 = 1 ( 2 1 + d).
(10.32)
As the forms satisfying such equations are unique, it follows that the connection form of the new basis is
0
2 1 = 2 1 + d.
0
(10.33)
It follows that d 2 1 = d2 1 and the curvature [10.29] is frame independent. It

depends only on the induced metric. This is the celebrated Gauss theorem
of surface theory, which has lead its discoverer to the idea that the geometry
of a space should be entirely described in terms of its own characteristics.
This was shown to be possible in large generality and, although imbeddings
10.4. D RELATION TO TOPOLOGY
457
were very helpful in finding fundamental properties and making them more
easily understood, all of them can be arrived at in an intrinsic way, the
only difficulty being the necessity of a more involved formalism. In higher
dimensional spaces, the curvature is characterized by all the components
R of the Riemann tensor which, for a n-dimensional space, amount to
n2 (n2 1)/12. When n = 2, only one component is enough to characterize
the curvature, as above. In this case, using equations [10.19], [10.21] and
[10.28], we find that
K = 12 R1212 .
10.4
D Relation to topology
10.4.1
The Gauss-Bonnet theorem
Suppose now that S is a compact surface. A field X on S will have a

finite number of singular points pi , given by Xpi = 0. Consider around each
singular point pi an open set Ui small enough for pi to be the only singular
point inside it. To calculate the index at pi , we should integrate the turning
angle (Math.9) of X along Ui . For that, we need to establish a starting
direction, which we take to be e1 . A useful trick is the following: introduce
on the complement S 0 = S i Ui another adapted frame {e01 , e02 , e03 }, with
X
and e02 , e03 chosen so as to make the frame positively oriented. The
e01 = |X|
angle to be integrated is then just that of equations [10.30], [10.31] and
[10.33]. The index at pi will be
H
1
d
Ii = 2
Ui
or, by [10.33],
2Ii =
21
Ui
21 =
Ui
21
Ui
R
Ui
d 2 1 .
0
Keep in mind that 2R1 is defined on the whole S, while 2 1 is only defined
on S 0 . The integral Ui d 2 1 can be made to vanish by taking Ui smaller
0
and smaller, so that Ui {pi }. The form 2 1 keeps itself out from the pi s.
Recalling that the sum of all the indices is the Euler characteristic, we arrive,
in the limit, at
P R
P R
0
0
2 = i Ui 2 1 = i SUi d 2 1
because, in a compact, the union Ui is also the boundary of S = S - Ui
with reversed orientation. From [10.29],
458
2 =
0
R
SUi
K 1 2 .
However, the form 1 2 is frame independent. In the limit Ui {pi }, the

integral leaves out only a set of zero measure it is identical to the integral
on the whole S. Consequently,
Z
1
= 2
K 1 2 .
(10.34)
S
As is independent of the induced metric and of the chosen field, this relation
depends only on S. It holds clearly also for any manifold diffeomorphic to
S. It is a relation between a differentiable characteristic of the manifold,
the curvature, and the topological substratum. It is a famous result, the
Gauss-Bonnet theorem. For a sphere S 2 of radius r, the Gaussian curvature
is K = 1/r2 , = r2 sin dd, and we obtain = 2 again. Notice that
r2 sin = g, as on S 2 the metric is such that

g dx dx = dl2 = r2 (sin2 d2 + d2 ).
In general, in a coordinate basis, [10.34] is written
Z
1
= 2
K g dx1 dx2 .
(10.35)
10.4.2
The Chern theorem
The theorem above has been generalized to manifolds of dimension 2n. First,
imbeddings were used. Allendoerfer and Weil found its first intrinsic proof
in 1943. Two years later, a simpler (for mathematicians) proof, using fiber
bundles, was given by Chern. For these 2n-manifolds, the Euler characteristic
is
Z
2
= a2n
KT g dx1 dx2 . . . dx2n .
(10.36)
S
where a2n is the area of the 2n-dimensional unit sphere,

n 22n+1 n!
(10.37)
a2n =
(2n)!
and KT is a generalization of the total curvature, given by
1
KT =
12...2n 12...2n R1212 R3434 . . . R2n1 2n 2n1 2n
(2n)!2n g
(10.38)
Lichnerowicz 1955
Spivak 1970
Doubrovine, Novikov & Fomenko 1979
Math.Topic 11
NON-EUCLIDEAN GEOMETRIES
1 The old controversy

2 The curvature of a metric space
3 The spherical case
4 The Boliyai-Lobachevsky case
5 On the geodesic curves
6 The Poincare space
11.1
The old controversy
Euclids postulate of the parallels (his 5-th postulate) stated that, given a
straight line L and a point P not belonging to it, there was only one straight
line going through P that never met L. The eon-long debate on this postulate
was concerned with the question of its independence: is it an independent
postulate, or can it be deduced from the other postulates? Euclid himself
was aware of the problem and presented separately the propositions coming
exclusively from the first four postulates. These propositions came to constitute absolute geometry. Those dependent on the 5-th postulate, as for
example the statement that the sum of the internal angles of a triangle is
equal to , he set apart. The debate was given a happy ending around the
middle of the 19th century through the construction of spaces which kept
in validity the first four postulates, but violated precisely the 5-th. On one
side, the independence was thereby proved. On the other, the very way by
which the solution had been found pointed to the existence of new, hitherto
unsuspected kinds of space. Such new non-euclidean spaces are at present
time called Riemannian spaces and their character is deeply rooted in their
metric properties. Though the word space has since then acquired a much
more general, metric-independent meaning, we shall in this chapter follow
the widespread usage of using it to denote a metric space. The main fact
about non-euclidean spaces is spelled out by saying that they are curved.
On such spaces, the role of straight lines is played by geodesics.
459
460
11.2
MATH.TOPIC 11. NON-EUCLIDEAN GEOMETRIES
The curvature of a metric space
As we have irritatingly repeated, it was found later that curvature is not

necessarily related to a metric. It is actually a property of a connection. It
so happens that a metric does determine a very special connection, the LeviCivita connection (represented by the Christoffel symbols in a convenient
covector basis). It is that unique connection which has simultaneously two
important properties: it parallel-transports the metric and it has vanishing
torsion. By curvature of a space we understand the curvature of the LeviCivita connection of its metric, and the space is said to be curved when
the corresponding Riemann tensor is non-vanishing. A space is flat when
R = 0, from which it follows that also its scalar curvature R = R

is zero. Euclidean spaces are flat because the Levi-Civita connection of an
euclidean metric has vanishing Riemann tensor.
A Riemannian space is said to be of constant curvature when its scalar
curvature R is a constant, and the simplest departures from the euclidean
case are those spaces for which R is still constant but non-vanishing. It was
only natural that the first curved spaces found were of this kind. When the
constant R > 0, the manifold is said to have positive curvature, and when
the constant R < 0 it is said to have negative curvature. A sphere S 2 is an
example of positive curvature, a sheet of a hyperboloid an example of negative
curvature. We shall in what follows briefly sketch these 2-dimensional cases,
though emphasizing the fact that the corresponding metrics can be attributed
to the plane R2 . In this way it becomes clear that distinct metrics can be
defined on the point set R2 , each one leading of course to different measures
of distance. We shall privilege cartesian coordinates and also make some
concessions to current language in this so much discussed subject.
The euclidean space E3 , we recall, is the set R3 , whose points are the
ordered real triples like p = (p1 , p2 , p3 ) and q = (q1 , q2 , q3 ), with the metric
topology given by the distance function
p
(11.1)
d(p, q) = + (p1 q1 )2 + (p2 q2 )2 + (p3 q3 )2 .
In cartesian coordinates (X, Y, Z), a sphere of radius Lcentered at the
origin will be the set of points satisfying X 2 + Y 2 + Z 2 = L2 . It will have
positive curvature. Hyperboloids, or pseudospheres, will have negative
curvature and are of two types. A single-sheeted hyperboloid will have its
points specified by X 2 + Y 2 Z 2 = L2 . A two-sheeted hyperboloid will be
given by the equation X 2 + Y 2 Z 2 = L2 . The cone stands in between
as a very special case of both, given by X 2 + Y 2 Z 2 = 0, and being
asymptotically tangent to them. These surfaces can be imbedded in E3 as
differentiable manifolds and are illustrated in Figure 11.1.
11.3. THE SPHERICAL CASE
461
The sphere S 2 is a Riemannian positive curvature space in which each

straight line (self-parallel curves, or geodesics, great circles in the case of
S 2 ) meets each other sooner or later, so that there are actually no straight
parallels to a previously given straight line. A hyperbolic sheet, on the
contrary, may exhibit many parallels to a given geodesics. What we shall do
Figure 11.1:
will be to consider both S 2 and a hyperbolic space imbedded in E3 , and
project them into a plane, thereby obtaining curved spaces on R2 . The
projection to be used, the stereographic projection, has very nice properties,
in special that of preserving circles. Geodesics are sent into geodesics, and
angles are also preserved. It turns out that, if we want to preserve metric
properties in the hyperbolic case, the imbedding space must be not E3 , but
the pseudo-euclidean space E2,1 instead.
The treatment is, up to the dimension, identical to that of the de Sitter
spaces (Phys.9). We shall, consequently, concentrate here on some basic aspects and refer to that chapter for others, as for example, for the justification
of the above statements on the values of the scalar curvature.
11.3
The spherical case
A point on the sphere

S 2 with radius L will be fixed by the values X, Y, Z
such that Z = L2 X 2 Y 2 . The relation to spherical coordinates are

X = L sin cos ; Y = L sin sin ; Z = L cos .
Consider now its stereographic projection into the plane. We choose the
north-pole N = (0, 0, L) as projection center and project each point of S 2
on the plane tangent to the sphere at the south-pole S = (0, 0, L), as
indicated in Figure 11.2. Cartesian coordinates (x, y) are used on the plane.
462
Consider a point P = (X, Y, Z) and its projection P 0 = (x, y, L). The

points N , P and P 0 are on a straight line, so that the differences between
their coordinates, (N P ) and (N P 0 ), are proportional:
X/x = Y /y = (L Z)/2L =: n.
The transformation is thus
X = nx; Y = ny; Z = L(1 2n).
We find then the solutions n = 0 (corresponding to the isolated point N ),
and
n=
4L2
x2 +y 2 +4L2
LZ
.
2L
If we call
r2 = (x2 + y 2 )/4L2 ,
the proportionality coefficient will be
n=
1
.
1 + r2
(11.2)
The relations between the coordinates are thus

X=
x
y
1 r2
;
Y
=
;
Z
=
L
.
1 + r2
1 + r2
1 + r2
(11.3)
The coordinates on the plane will be

x=
2LX
2LY
; y=
.
LZ
LZ
(11.4)
These (x, y) constitute a local system of coordinates with covering patch

S 2 \{N }. The metric, or the line element will be given by
ds2 = dX 2 + dY 2 + dZ 2 = n2 (dx2 + dy 2 ) =
dx2 + dy 2
.
(1 + r2 )2
(11.5)
Comment 11.3.1 Stereographic projections provide the most economical system of

coordinates for the sphere S 2 : only 2 charts are needed. One is the above one, the other
is obtained by projecting from the south pole S onto the plane tangent to S 2 at the north
pole N . Each projection is a homeomorphism of the plane with the chart S 2 \{projection
center}, which is thereby a locally euclidean set. Cartesian coordinates would need 8
charts to cover the sphere with locally euclidean sets.
11.3. THE SPHERICAL CASE
463
The important point is that this procedure may be seen alternatively as

a means of defining a new, non-euclidean metric
gij = n2 (x, y) ij ,
on the plane, with n(x, y) = 1/(1 + r2 ) and ij the euclidean metric. With
this metric, we agree to define as the distance between two points on the
plane, the length of the shortest arc connecting the corresponding points on
the sphere. This is an example of Riemannian structure on the plane R2 .
The curvature of the corresponding Levi-Civita connection will be constant
and positive.
Given the interpretation of distance, we may expect that the geodesics on
the plane be the projections of the spherical great circles. Indeed, there are
two possible results of projecting circles: straight lines and circles. Notice to
start with that
r2 =
L+Z
,
LZ
so that lines at constant Z (horizontal circles) will be led into points satisfying
r2 = constant. The equator (Z = 0), in particular, is taken into r2 = 1, or
x2 + y 2 = 4L2 . Each great circle meeting the equator at (X, Y ) will meet it
again at (X, Y ), and this will correspond to (x, y) and (x, y). In the
general case, a great circle is the intersection of S 2 with a plane through the
origin, with equation uX + vY + wZ = 0. This plane is orthogonal to the
Figure 11.2:
vector (u, v, w), whose modulus is irrelevant, so that actually the constants
are not independent (if u, v, w are direction cosines, u2 + v 2 + w2 = 1).
464
Figure 11.3:
Examine first the planes going through the axis OZ: w = 0 and then Y =
(u/v)X =: X. The intersection projection will satisfy the equation y =
x, representing straight lines through the plane origin. Now, w 6= 0 when
the circle is in general position, which leads to
Z = X + Y = n(x + y)
with = u/w and = v/w. We then obtain from n = (L Z)/2L the
equation
(x 2L)2 + (y 2L)2 = 4L2 (1 + 2 + 2 ),
representing circles centered at the point (2L, 2L).
Geodesics are great circles. Consequently, all the geodesics starting at a
given point of S 2 will intersect again at its antipode. And we see in this way
how Euclids postulate of the parallels is violated: there are no parallels in
such a space, as any two geodesics will meet at some point.
11.4
The Boliyai-Lobachevsky case
A point on a two-sheeted hyperboloid will be fixed by the values X, Y, Z of

its coordinates satisfying the condition
X 2 + Y 2 Z 2 = L2 or Z = X 2 + Y 2 + L2 .
11.4. THE BOLIYAI-LOBACHEVSKY CASE
465
Again we choose the point (0, 0, L) as projection center (it is now the lowest
point of the upper branch) and project each point of the hyperboloid on
the plane tangent at the point (0, 0, L) (which is now the highest point on
the lower branch), as indicated in Figure 11.3. The same reasoning applied
above to the spherical case will lead again to the relations X/x = Y /y =
(L Z)/2L = n, but the form X 2 + Y 2 Z 2 = L2 leads to another
expression for the function n. Instead of [11.2], we have now
n(x, y) =
1
,
1 r2
(11.6)
so that the relations between the coordinates are

X=
x
y
1 + r2
;
Y
=
;
Z
=
L
.
1 r2
1 r2
1 r2
(11.7)
Now we have a problem. We would like to have the equations defining the
surfaces to represent relations between measured distances, and the interval
to be ds2 = dX 2 +dY 2 dZ 2 . This is obviously impossible with the euclidean
metric. In order to preserve that idea, we must change the ambient space
and consider pseudo-euclidean spaces. The above non-compact surfaces will
be (pseudo-) spheres in such spaces.
The pseudo-euclidean space E2,1 is the set point R3 with, instead of [11.1],
the pseudo-distance function
p
(11.8)
d(p, q) = + (p1 q1 )2 + (p2 q2 )2 (p3 q3 )2 .
Due to the negative sign before the last term, this is not a real distance
function and does not define a topology. Taking the origin q = (0, 0, 0) as
the center, there are spheres of three types, according to the values of their
radius L: precisely spheres of real radius, satisfying X 2 + Y 2 Z 2 = L2 ;
those of vanishing radius, X 2 +Y 2 Z 2 = 0; and those with purely imaginary
radius, satisfying X 2 +Y 2 Z 2 = L2 . Thus, the above surfaces in E3 appear
as the possible spheres in E2,1 , with the extra bonus that now the equations
have a metric sense.
Returning to the specific case of the two-sheeted hyperboloid, we have
that the interval in the plane coordinates turns out to be
ds2 = dX 2 + dY 2 dZ 2 = n2 (dx2 + dy 2 ) =
dx2 +dy 2
.
(1r2 )2
The metric is consequently

gij =
1
ij .
(1 r2 )2
(11.9)
466
Metrics like this one and that defined in [11.5], which are proportional to a
flat metric, are called conformally flat metrics. Though they give measures
of distance different from those of the euclidean metric, they give the same
measure for angles.
The plane R2 with the above metric is called the Lobachevsky plane.
Notice that the metric actually brings the hyperboloid down and up to the
plane. The lower sheet has Z/L < 1 and is mapped into the disc bounded
by the circle r2 =1. The disc r2 < 1 with the metric [11.9] is called the
Poincare space. We shall come to it later. The upper sheet has Z/L > 1
and is mapped into the complement of the disc in the plane. As we go to
infinity in the upper and lower sheets we approach the circle (respectively)
from outside and from inside.
We may analyze the projections of intersections of the hyperboloids with
planes (the now eventually open great circles) in the same way used for
the spherical case. Putting together the two metrics by writing
n = (1 r2 )1 ,
we find on the plane the curves
(x 2L)2 + (y 2L)2 = 4L2 (2 + 2 1).
11.5
(11.10)
On the geodesic curves
Given the metric gij = n2 ij , the components of the Levi-Civita connection

are
k ij = [jk i + ik j ij kr r ] ln n,
(11.11)
which is a general result for conformally flat metrics. The calculations are
given in Phys.9, where also everything concerning the curvature is to be
found. Here, we shall rather comment on the geodesics. The geodesic equation is
i
j
dv k
d ln n 1
d2 xk
k dx dx
+
=
+ 2v k
+ 2 k n2 = 0.
(11.12)
ij
2
ds
ds ds
ds
ds
It is a happy fact that, whenever we may define a momentum by
pi = gij v j ,
(11.13)
the geodesics equation simplify because the first two contributions in the
Christoffeln [11.11] just cancel the term coming from the derivative of the
metric. We remain with
dpk 1
+ 2 [k g ij ]pi pj = 0.
(11.14)
ds
SPACE
11.6. THE POINCARE
467
Notice that this is the same as p(v) = pk v k = 1, which happens always

when the geodesics is parametrized by its length. Here we find a force law
dpk
= 12 k (ln n).
(11.15)
ds
This is just equation [5.13] of Topic Phys.5, which devoted to Optics. There,
n hass the role of the refraction index. The situation is also analogous to
the Poinsot construction of particle Mechanics, described in Phys.2. The
geodesic motion
in that case corresponds to that of a particle for which the
quantity ln n acts as a potential.

It seems simpler here to try to integrate just pk v k = 1, or v k v k = n12 , or
(x)
2 + (y)
2 = n12 = [1 r2 ]2 , or still
r 2 + r2 2 =
[1r2 ]2
.
4L2
We shall prefer, however, to change coordinates before solving the equations.

The Jacobi equation, analogous to the case of Phys.9, is
1
D2 X
+
[X (X V )V ] = 0,
2
2
Ds
L
(11.16)
or
D2 X
1
+
[X g(X, V )V ] = 0.
(11.17)
Ds2
L2
Now, X = [X g(X, V )V ] is the component of X transversal to the curve.
D2
||
As the tangential part X || has Ds
= 0, one arrives at
2X
1
D2 X
+ 2 X = 0.
2
Ds
L
11.6
(11.18)
The Poincar
e space
Consider now the Poincare space. We may consider the plane E2 as the
complex plane. It is known that an open disc on the complex plane can be
taken into the upper-half-plane by a homographic transformation, which is
furthermore a conformal transformation. Actually, there is one transformation for each point of the half-plane. For each arbitrarily chosen point a
on the half-plane, there will be one homographic half-plane disc transformation taking a into the circle center.1 Introducing the complex variables
z = x + iy and w = u + iv, the transformation
wa
z=K
(11.19)
w a
1
Lavrentiev & Chabat 1977.
468
takes the open upper-half-plane onto a disc of radius K whose center is the
transformed of a. We choose for convenience a = iaL. The relations between
the coordinates are:
x = 2L
u2 + v 2 4L2
4Lu
; y = 2L 2
,
2
2
u + (v + 2L)
u + (v + 2L)2
(11.20)
with their inverses

8L2 y
8L3 (1 r2 )
2y
2L(1 r2 )
=
=
;
v
=
.
1 + r2 x/L
(x + 2L)2 + y 2
1 + r2 x/L
(x + 2L)2 + y 2
(11.21)
The last relation shows that:
(i) v = 0 corresponds to r2 = 1;
(ii) v > 0 corresponds to 1 > r2 .
The metric above becomes ds2 = (du2 + dv 2 )/v 2 on the upper-half-plane.
Let us examine the geodesics in this case. With
u=
gij =
1
v2
ij ,
the Christoffeln are (notation: u1 = u; u2 = v)

i jk =
1
[ij k2 + ik j2 jk i2 ].
v
(11.22)
The geodesic equations are:

2
u v = 0,
v
(11.23)
1 2
[v u 2 ] = 0.
v
(11.24)
u
v
Now, we have two cases:

(i) If u = 0, the solution u = C, with C a constant, will work if there is a
solution for v = v1 v 2 . This is solved by v = eAt+B , with A and B constants.
Thus, all the vertical straight lines will be solutions;
(ii) If u 6= 0, we find
u = cv 2 ; v1 v v1 v 2 = c2 v 2 = cu.
Putting
z = ln v; z = v1 v = v1 v v1 v 2 ; z = cu
will lead to z = cu + d. The two remaining equations, v1 v = d cu and
u = cv 2 are put together in
SPACE
11.6. THE POINCARE
469
dv
du
dcu
,
cv
whose solutions are the circles

(u c/2)2 + v 2 = B 2 ,
centered on the horizontal axis and orthogonal to it.
On the disc, these two families of solutions will have the following correspondence:
(1) the vertical straight lines u = C will be taken into circles
(x 2L)2 + (y + 4L2 /C)2 = 16L4 /C 2 .
These circles intersect at right angles the border r2 = 1 when y =
(C/2L)x C (lines a and b in Figure 11.4);
(2) the circles (uc/2)2 +v 2 = B 2 will be transformed into y = (4L/c)x
(line c in Figure 11.4).
Figure 11.4:
Figure 11.5:
The Poincare disc has a very curious zoology of curves. Some are indicated in Figure 11.5. There are equidistant curves (as a and b), circles (c),
470
and circles which are tangent to the infinity circle r2 = 1. The latter goes
under the honest name of limiting circles, but also under those of oricycles, or
still horocycles (curve h in Figure 11.5). They, and their higher-dimensional
analogues (horispheres), are important in the study of representations of the
groups of motions on hyperboloids.2 Comparison with the corresponding
curves in the half-plane is an amusing exercise. It serves at least to illustrate
the vanity of a curve aspect, which depends heavily on the coordinate system.
Whittaker 1958
Rosenfeld & Sergeeva 1986
Lavrentiev & Chabat 1977
Vilenkin 1969.
Math.Topic 12
GEODESICS
A wellknown aphorism says that there is only one experiment in Physics: Rutherfords.
In order to acquire information on a blackbox system (an atom, a molecule, a nucleus),
we examine the resulting paths of a wellknown probe (a particle, a light beam) sent
against it. Comparison is made with the paths of the free case, in which the black
box system is not there. That is to say that Physics dwells frequently with the study of
trajectories. In this chapter we describe how to obtain information on a geometry through
the study of its curves. It turns out that what is best probed by trajectories is always a
connection.
Introduction
The word geodesic is used in more than one sense. In the common lore, it
means the shortestlength curve between two points. In a more sophisticated
sense, it represents a selfparallel curve, whose tangent vectors keep parallel
to themselves along the curve. The first concept assumes a metric to measure
the lengths, whereas the second supposes a welldefined meaning for the idea
of parallelism. These are quite different concepts. The ambiguity comes
from the fact that geodesics have been conceived in order to generalize to
curved spaces the straight lines of euclidean spaces. They are as straight as
possible on noneuclidean spaces. The trouble is that the starting notion
of a straight line is itself ambiguous.
A straight line can be seen either as the shortestlength curve between
two points or as a curve keeping a fixed direction all along. The two ideas
coincide for euclidean spaces, but not for more general spaces. The shortest
length point of view is meaningful only on strictly Riemannian manifolds
(endowed with a positivedefinite metric). The selfparallel idea has a
meaning for pseudoRiemannian spaces, and even in nonmetric spaces. As
it happens, parallelism is an idea related to a connection, which may be or
not related to a metric. As the selfparallel concept is more general and
reduces to the shortestlength point of view in the metric case, it is to be
471
472
MATH.TOPIC 12. GEODESICS
preferred as the general definition of geodesics. We shall here, however, start

introducing the notion through its physical connotations. From the physical
point of view, there are two main approaches to the idea of a geodesic. One
relates to gravitation, the other to optics.
12.1
Selfparallel curves
12.1.1
In General Relativity
In the theory of gravitation, the notion of a geodesic curve comes from a

particular use (and view) of the equivalence principle. Start with Newtons
law for the force per unit mass, written of course in a cartesian coordinate
system {y k } on the euclidean space E3 . It is
F k = Ak =
d2 y k
.
dt2
Time is absolute, so that t has the same value for all events in space. Space
time is thus a direct product E1 E3 . Now pass into another, arbitrary
coordinate system {xk }, in terms of which the distance between two infinitesimally close points in spacetime is written ds2 = gij dxi dxj . Then the
expression for the acceleration becomes
A0k =
i
j
d2 xk
k dx dx
+
,
ij
dt2
dt dt
(12.1)
where the k ij s are coefficients given in terms of the gij s as

k ij = 21 g kr [i grj + j gri r gij ]
(12.2)
There is nothing new in this change of appearance, only a change of coordinates. The equation stating the absence of forces,
i
j
d2 xk
k dx dx
+ ij
=0,
dt2
dt dt
(12.3)
gives simply a straight line written in the new coordinates. The set {gij }
contains the components of the euclidean metric of E3 , which are {ij } in
cartesian coordinates {y k }, written in the new coordinates {xk }. It only
happens that, in euclidean cartesian coordinates, the k ij s vanish, and so
do their derivatives.
Suppose now that this is no more the case, that gij is another, non
euclidean metric. The k ij s are then the Christoffel symbols, representing
12.1. SELFPARALLEL CURVES
473
a connection intimately related to the metric. It happens that, though it is

still possible to make the k ij s to vanish at a point in a suitable (normal)
coordinate system, it is no more possible to make also their derivatives to
vanish. The derivatives appear in the curvature of this connection which,
unlike the euclidean case, does not vanish. But the equation (12.3) keeps
its sense: it gives the path of a particle constrained to move on the new,
curved space, but otherwise unsubmitted to any forces.
12.1.2
The absolute derivative
Take a differentiable manifold M and consider on it a linear connection

of curvature R and torsion T . The covariant derivative along a vector field
X will be indicated by X . We recall the representation of curvature and
torsion as families of mappings ( 9.4.15):

R(X, Y )Z = X Y Y X [X,Y ] Z
= X (Y Z) Y (X Z) [X,Y ] Z (12.4)
and
T (X, Y ) = X Y Y X [X, Y ] .
(12.5)
Given a curve with parameter on M , and its tangent velocity field

V ( ) = (
)=
d
d
( ) ,
(12.6)
the covariant derivative of a field or form W along the curve will be DW

=
D
V W . The derivative
D
= V
(12.7)
D
along the curve is frequently called absolute derivative. The object W will
be paralleltransported along if V W = 0. The acceleration, for example,
will have the invariant expression
A=
DV
= V V .
D
(12.8)
This expression holds in any basis and differs from (12.1) only because
the connection involved is quite general. In the basis {ea = a }, the velocity
field V is dd and the acceleration will be written

d b d c
d2 a
a
A=
+ bc
ea .
d 2
d d

(12.9)
474
Comment 12.1.1 The absolute derivative
DS
D
reduces to the simple derivative
dS
d
if
S is an invariant, or a scalar. For instance,

D
d
(Va V a ) =
(Va V a ) .
D
d
In General (and Special) Relativity, the interval or (squared) proper time is given by d 2
= gab dxa dxb = dxa dxa . If we use the proper time as the curve parameter, Va V a = 1.
Then,
D
(Va V a ) = 2 Va Aa = 0 .
D
The acceleration, when nonvanishing, is orthogonal to the velocity. The property ||V ||2
= 1 is valid everytime the curve is parametrized by the proper time. We have, in the
D
gab = V c gab;c = 0. This supposes that the metric is
simple calculation above, used D
paralleltransported by the connection. This property, gab;c = 0, is sometimes called
metric compatibility of the connection, and also metricity.
Notice that a different convention, with opposite sign for d 2 , is frequently used. This
leads to Va V a = 1. The signs in some definitions given below (of transversal metric, of
Fermi derivative) must be accordingly modified.
12.1.3
Selfparallelism
The curve will be a selfparallel curve when the tangent velocity keeps
parallel to itself along , that is, if
DV
= V V = 0 .
(12.10)
A=
D
In the basis {ea = a },
dV a
d2 a
d b d c
a
+ bc
=
+ a bc V b V c = 0 .
(12.11)
2
d
d d
d
A selfparallel curve is more suitably called a geodesic when the connection
is a metric connection (as said, there are fluctuations in this nomenclature: many people call geodesic any selfparallel curve). The equation above,
spelling the vanishing of acceleration, is the geodesic equation.
Comment 12.1.2 The parametrization t (t) = 0 (t)), 1 (t), 2 (t), 3 (t) is, when
(t) satisfies the geodesic equation, defined up to an affine transformation t s = at + b,
where a 6= 0 and b are real constants. For this reason, such a t is frequently called an
affine parameter.
In euclidean space and arbitrary coordinates, the lefthand side of the

geodesic equation is simply the time (or curve parameter) second derivative
of a . The equation says that no force is exerted on a particle going along
the curve, which is consequently the path followed by a free particle. A
vector X is said to be paralleltransported along a curve if
V V X = 0 .
(12.12)
12.1. SELFPARALLEL CURVES
475
Comment 12.1.3 If paralleltransported around a closed curve, X k will come back

to the initial point modified by an amount X k which is a measure of the curvature flux
through the surface circumscribed by the loop:
R
X k = 21 Rk ijl X i d l d j .
12.1.4
Complete spaces
A linear connection is complete if every geodesic curve of can be extended

to a geodesic (t) with ( < t < +). Every geodesic on M is the
projection on M of the integral curve of some standard horizontal field of
BLF (M ), and viceversa. Thus, a linear connection is complete iff every
standard horizontal field of BLF (M ) is complete.
A Riemannian manifold M is geodesically complete if every geodesic on
M can be extended for arbitrarily large values of its parameter. This nomenclature is extended to the metric proper. There are many results concerning
such complete spaces. We shall quote three of them, only to give an idea of
the interplay between geodesics and topologicaldifferential properties.
Let M be a Riemannian complete manifold with nonpositive curvature. Take a point p M and consider the exponential mapping
expp : Tp M M .
Then expp is a covering map. If M is simplyconnected, expp is a
diffeomorphism. Two important results are:
every homogeneous Riemannian manifold is complete;
every compact Riemannian manifold is complete.
12.1.5
Fermi transport
In the framework of General Relativity, not everything follow a geodesic. A

selfpropelled rocket will not do it, neither will a free but spinning particle
(Papapetrou 1951).
Derivatives different of the above absolute derivative are of interest on
general curves. We only consider the Fermi derivative along (t), which is
DF X
DX
DV
DV
=
+ g(X,
) V g(X, V )
.
D
D
D
D
(12.13)
FX
A vector X such that DD
= 0 is said to be Fermipropagated, or Fermi
transported, along the curve. This derivative has the following properties:
476
1. when is a geodesic,
2.
DF V
D
DF X
D
DX
D
= 0 on any curve;
FX
FY
3. if both DD
= 0 and DD
= 0 hold, then g(X, Y ) is constant along the
curve, meaning in particular that orthogonal vectors remain so along
the curve;
4. take an orthogonal basis {ea }, such that e4 = V at some starting point;

then, if Fermitransported, {ea } will be taken into another orthogonal
basis at each point of the curve, with e4 = V ; the set {e1 , e2 , e3 } constitutes a nonrotating set of axes along (t).
12.1.6
In Optics
In geometrical optics, geodesics turn up as light rays, which obey just the
geodesic equation. There, actually, geodesics justify their primitive geodesical role. Not as the shortest length line between the two endpoints, but as
that line corresponding to the shortest optical length. The refraction index
is essentially a 3dimensional space metric, and the light ray follows that
line which has the shortest length as measured in that metric. This is seen
in some detail in Topic Phys.5. Light rays will follow geodesics also in the
4dimensional spacetime of General Relativity. We only notice here that,
as the interval ds2 = 0 for massless particles like the photon, fourvelocities
cannot be defined as they can [see eq.(12.6)] for massive particles following
a timelike curve. Some other curve parameter must be used.
12.2
Congruences
Our objective in the following will be to give a general, qualitative description of the relative behaviour of neighbouring curves on a manifold. We
shall concentrate on geodesics, and by neighbouring we shall mean those
geodesics which are near to each other in some limited region of the manifold.
A family of neighbouring curves is called a pencil or congruence (or still a
ray bundle, or bunch).
12.2.1
Jacobi equation
If is a geodesic, its infinitesimal variation field X is called a Jacobi field

(Math.7). Due to the particular conditions imposed on to make of it a
geodesic, X will satisfy a second order ordinary differential equation, the
12.2. CONGRUENCES
477
Jacobi (or deviation, or secondvariation, or in the case, geodesic deviation) equation:

V V X + V T (X, V ) + R(X, V )V = 0 .
(12.14)
Comment 12.2.1 The field X = 0 is a trivial solution. Notice that T (V, V ) = 0 and
R(V, V ) = 0. On a geodesic, as A = 0, both A and V are Jacobi fields. On a nongeodesic
curve, to impose that A is a Jacobi field is to say that the second acceleration vanishes:
V V V = V A =
D2 V
=0.
D 2
As
DW
= V W = [V a Da W c ]ec ,
D
eq. (12.14) is the same as
D2 X
D
+
T (X, V ) + R(X, V )V = 0 .
2
D
D
(12.15)
(12.16)
In a basis {ea }, this equation reads

D2 X
D
+Vb
(X a T d ab )ed + X a V b V c Rd cab ed = 0 .
2
D
D
(12.17)
In components,

D2 X
D 2
or

D2 X
D 2
d
d
On a geodesic, as

D2 X
D 2
d
or

D2 X
D 2

+ V b V c Dc (X a T d ab ) + X a Rd cab = 0 ,
d
+V
D
Vb
D

D
a d
a d
c
(X T ab ) + X R cab V = 0 .
D
(12.18)
(12.19)
= 0,

+ V c Dc (X a T d ab V b ) + X a Rd cab V b V c = 0 ,

D
a d
b
a d
b c
+
(X T ab V ) + X R cab V V = 0 .
D
(12.20)
(12.21)
In the study of the general behaviour of curves, one is most frequently

interested in the transversal behaviour, on how things look on a plane
(or space) orthogonal to the curve. More precisely, one considers the space
tangent to the manifold at a point on the curve; one direction will be along
478
the curve, colinear with its velocity vector. The remaining directions are
transversal. The metric can be projected into a metric on that subspace:
hmn = gmn Vm Vn (see eq. (12.33) below). Thus, to each point p on the
curve will correspond a plane Pp orthogonal to the curve at p. The next step
is to examine congruences of curves, together with their variations.
Consider a congruence of curves (s) and a variation of it, (s, t). At fixed
s, these curves will cross (s) and constitute another congruence, parametrid
and U = dtd tangent to the respective
zed by t. Consider the fields V = ds
congruences of curves and normalized to unity. Each one of these fields will
be taken into itself by the others congruence. That is, their Lie derivatives
will vanish: [V, U ] = 0. From this commutativity it follows that V a a U b
= U a a V b . But then it follows also that V a Da U b = U a Da V b , or V a U b ;a =
D
U a V b ;a , or still Ds
Ub = D
V b . This may be written as
Dt
D b
U = U a Da V b = U a V b ;a ,
Ds
or
V U = U V .
(12.22)
Notice the difference with respect to the usual invariant derivative

D
U b = V a U b ;a .
Ds
Equation (12.22) has important consequences. Notice that it is basically a
D
D
matrix equation: Ds
U = V U . Taking the absolute derivative Ds
is the same
b
as multiplying by the matrix V = (V ;a ). It means that, while transported
along (s), the transversal vector field U(s) is taken from U(s+ds) by a
simple matrix transformation. If we consider the orthogonal parts given by
D
U k = V k ;j U j . We
the projector, only U k = hk b U b = g ka hab U b , then Ds
shall come back to this transversal approach later.
Let us obtain the Jacobi equation in the strictly Riemannian case. One
sees that the torsion term appearing in the equation would anyhow vanish
in this case: when [U, V ] = 0 and U V = V U ,
T (U, V ) = U V V U [U, V ] = 0 .
Going back to the definition of curvature, we find
R(U, V )V = U (V V ) V (V U ) [U,V ] V
= U (V V ) V (V U ).
(12.23)
12.2. CONGRUENCES
479
As for a geodesics V V = 0, it follows that

V (V U ) + R(U, V )V = 0 ,
(12.24)
which is the simplest case of the Jacobi equation. Other forms are
D2 U
+ R(U, V )V = 0
D 2
(12.25)
and
D2 U
D2 U
a b
+
U
V
R(e
,
e
)V
=
+ U a V b Rd cab ed = 0 .
a
b
D 2
D 2
In a basis {ea =
(12.26)
},
a
D2 U d
D2 U d
d b d c
a d
+
U
R
=
+ U a Rd cab V b V c = 0 .
cab
D 2
d d
D 2
(12.27)
The equation is sometimes given in terms of the sectional curvature.

Given a plane in the tangent space Tp M with {e1 , e2 } a basis for it, then
the sectional curvature on the plane is K(plane) = g(R(e1 , e2 )e2 , e1 ). Thus,
in the (X, V ) plane, with X V , it is just g(X, R(X, V )V ). We see that, if
we project the equation along X, the curvature term is actually the sectional
curvature.
There are two natural Jacobi fields along a geodesic (s): (s)
and s(s).
It is a theorem that, on a Riemannian manifold, any Jacobi field X can be

uniquely decomposed as
X = a + bs + X ,
(12.28)
where a and b are real numbers, and X is everywhere orthogonal to . If A

is orthogonal to at two points, it will be orthogonal all along and will have
the aspect of Figure 12.1.
Comment 12.2.2 The set of all the Jacobi fields along a curve on an ndimensional
manifold constitutes a real vector space of dimension 2n.
Take a geodesic . Two points p and q on are conjugate to each other

along if there exists some nonzero Jacobi field along which vanish at p
and q. This means that infinitesimally neighbouring geodesics at p intersect
at q. On a Riemannian manifold with nonpositive sectional curvature there
are no conjugate points.
480
Figure 12.1: A field U orthogonal to the curve .
12.2.2
Vorticity, shear and expansion
It is clear that the study of Jacobi fields is the main source of information
on the relative behaviour of neighbouring geodesics. They reflect all the
qualitative behaviour: whether geodesics tend to crowd or to separate, to
cross or to wind around each other. But there are some other tensor an
scalar quantities which can be of great help. In particular, when dealing with
the effect of curvature on families of curves in spacetime, a hydrodynamic
terminology is very convenient, as such curves (if timelike or null) represent
possible flow lines (of a fluid constituted by test particles), or are histories
of massless particles. We can introduce the notions of vorticity and shear,
expansion tensor and volume expansion, all of them duly projected into the
transversal space. The intuitive meaning of such tensor quantities is the
following. Suppose given around the curve a congruence of curves, whose
orthogonal cross section draw a circle on Pp . As we proceed along the curve
from point p to another point q at a distance ds, the congruence will take
the points on the circle at p into points on a line on Pq . Volume expansion
will measure the enlargement of the circle, tensor expansion will measure its
deformations, which may be larger or smaller in each direction (Figure 12.2).
At vanishing tensor and volume expansion, the circle will be taken into an
equalsized circle, as shown in Figure 12.3. The same Figure illustrates the
vorticity of the central curve , which measures how much the neighbouring
geodesics turn around it.
The procedure consists of looking at expansion, tensor and volume, and
12.2. CONGRUENCES
481
Figure 12.2: Transversal view of the behaviour of a general congruence.
see how such tensorial quantities behave along the curve. This means to
D
calculate D
acting on them.
Given a metric g, the general invariant definitions of vorticity and expansion are:
vorticity:
(X, Y ) = g(X, Y V ) g(X V, Y )
expansion: (X,
Y) =
1
2
[g(X, Y V ) + g(X V, Y )] .
They are related to a curve of tangent field V in the following way: vorticity is its covariant curl
(X, Y ) = X a Y b V[a;b]
(12.29)
ab =
(ea , eb ) = Va;b Vb;a .
(12.30)
or, in the basis {ea },
The expansion tensor is

a , eb ) = 1 [Va;b + Vb;a ] .
(e
2
(12.31)
482
Figure 12.3: In the absence of shear and volume expansion, a circle around
the curve is taken into another circle of the same size. Vorticity will, however,
cause a rotation.
a , eb ). If the tangent V to
Comment 12.2.3 Consequently, Va;b = 12 (ea , eb ) + (e
ab = 0.
the curve is a Killing vector,
ab = 2 Va;b and
In spacetime, the fourvelocity has been defined in such a way that g(V, V )
= 1. This allows one to define a transversal metric by
h(X, Y ) = g(X, Y ) g(X, V ) g(Y, V ) ,
(12.32)
which gives no components along V ,

h(X, V ) = g(X, V ) g(X, V ) g(V, V ) = 0 .
In components,
hab = gab gac gbd V c V d .
(12.33)
Then, the transversal vorticity and expansion are given by

(X, Y ) = h(X, Y V ) h(X V, Y ),
(12.34)
(X, Y ) = 21 [h(X, Y V ) + h(X V, Y )] .
(12.35)
12.2. CONGRUENCES
483
At each point of a (timelike) curve, they give the vorticity and the expansion
in space.
When n = 4, a vorticity vector is defined by
a = 21 abcd Vb cd ; ab = ha c hb d V[c;d] .
The volume expansion is = V a ;a . The shear tensor is the traceless part of
the expansion tensor
(12.36)
ab = ab 13 hab
12.2.3
LandauRaychaudhury equation
We quote for completeness this equation, which relates expansion to curvature, vorticity and shear:

DV a
d
a b
2
2
2
1
= Rab V V + 2( ) 3 +
.
(12.37)
ds
Ds ;a
As 2 2 = ab ab 0 and 2 2 = ab ab 0, we see that vorticity induces
expansion, and shear induces contraction. Another important equation is
(with everything transversal)

DV[
d
ab = 2 [ ] +
.
(12.38)
ds
Ds ;]
The use of the above concepts and techniques allowed Penrose and Hawking to show that Einsteins equations lead to a singularity in the primaeval
universe. In General Relativity light rays follow null geodesics, and material objects follow timelike geodesics. In the Standard Cosmological Model
[Weinberg 72], the motion of the always receding galaxies are represented by
an expanding timelike congruence of geodesics in spacetime. If we start
from the presentday remarkably isotropic state of the universe and look
backward in time, the above equations lead, under certain physically reasonable conditions, to a starting point of any geodesics. Cosmic spacetime is
consequently not a complete space.
Synge & Schild 1978
Kobayashi & Nomizu 1963
Synge 1960
Hawking & Ellis 1973
484
Part V
PHYSICAL TOPICS
485
Phys. Topic 1
HAMILTONIAN MECHANICS
1.
2.
3.
4.
5.
6.
7.
8.
Introduction
Symplectic structure
Time evolution
Canonical transformations
Phase spaces as bundles
The algebraic structure
Relations between Lie algebras
Liouville integrability
This summary of hamiltonian mechanics has three main objectives: (i) to provide
a good exercise on differential forms, (ii) to show how forms give a far more precise
formulation of a well-known subject and (iii) to present an introductory resume of this
classical theory in what is nowadays the mathematicians colloquial language.1
1.1
Introduction
1.1 Consider the classical phase space M of some conservative mechanical

system with generalized coordinates q = (q 1 , q 2 , . . . , q n ) and conjugate momenta p = (p1 , p2 , . . . , pn ). States will be represented by points (q, p) of the
2n-dimensional space M and any dynamical quantity will be a real function
F(q, p). The time evolution of the state point (q, p) will take place along
the integral curves of the velocity vector field on M ,
XH =
dq i
dpi
+
.
dt q i
dt pi
(1.1)
By using the Hamilton equations

dq i
H
=
dt
pi
and
dpi
H
=
dt
q i
(1.2)
The emphasis here is neither that usually found in physicists texts, standardized
for instance in Goldstein 1980, nor that highly mathematical of the Bible Abraham &
Marsden 1978. It is rather in the line of Arnold 1976.
487
488
PHYS. TOPIC 1. HAMILTONIAN MECHANICS
this evolution field takes up the form

XH =
H
H
.
i
pi q
q i pi
(1.3)
The hamiltonian flow is the oneparameter group (section 6.4.2) generated

by XH . The hamiltonian function H(q, p) will have as differential the 1form
dH =
H i H
dpi .
dq +
q i
pi
(1.4)
Applied to (1.3), this form gives dH(XH ) = dH/dt = 0, which says that the
value of H is conserved along the integral curve of XH .
The hamiltonian formalism provides deep insights into the details of
any mechanical system, even when the number of degrees of freedom is
continuum-infinite, as they are in Field Theory. We are here confining ourselves to particle mechanics, but those willing to see an example of that kind
can take a look at Phys.7.1.3, where the formalism is applied to gauge fields.
1.2
Symplectic structure
1.2 Equations (1.3) and (1.4) suggest an intimate relationship between

dH and XH . In reality, a special structure is always present on phase spaces
which is responsible for a general relation between vector fields and 1forms
on M . In effect, consider the 2form
= dq i dpi .
(1.5)
It is clearly closed, and it can be shown to be also nondegenerate. The

interior product iXH is just
iXH = dH .
(1.6)
Through the interior product, the form establishes a one-to-one relation

between vectors and covectors on M .
Recall that a metric structure (see section 6.6), defined on a manifold
by a second-order symmetric, nondegenerate tensor, establishes a one-to-one
relation between fields and 1forms on the manifold, which is then called a
metric space. In an analogous way, the form (which is an antisymmetric,
nondegenerate secondorder tensor) gives a one-to-one relation between fields
and 1forms on M :
X iX .
(1.7)
1.2. SYMPLECTIC STRUCTURE
489
This is a mere analogy the structure defined by a closed 2form differs

deeply from a metric structure.
Formally, a symplectic structure, or hamiltonian structure,2 is defined on
a manifold M by any closed (not necessarily exact) nondegenerate 2form
. The manifold M endowed with such a structure is a symplectic manifold.
The 2-cocycle is the symplectic form. In the above case, it is also an exact
form (a coboundary, or a trivial cocycle) as it is, up to a sign, the differential
of the canonical form, or Liouville form3
= pi dq i .
(1.8)
With a clear introduction of the important notions in mind, we have actually

been taking as a model for M the simplest example of phase space, the topologically trivial space E2n . Though a very particular example of symplectic
manifold, the pair (E2n , ) is enough to model any case in which the configuration space is a vector space. It is good to have in mind, however, that many
familiar systems have non-trivial phase spaces: for the mathematical pendulum, for example, it is a cylinder. In such cases, there may be no global
coordinates such as the (q i , pi ) supposed above. On generic, topologically
non-trivial symplectic manifolds, not only global coordinates are absent, but
the basic closed nondegenerate 2form is not exact. Furthermore, in general,
a vector field like X is not well defined on every point of M . In fact, it
is generally supposed that H has at least one minimum, a critical point at
which dH = 0; comparing [1.3] and [1.4], we see that the field XH vanishes
at such a point: it has a singular point. Finally, integral curves of a given
vector field are in general only locally extant and unique. Nevertheless, a
theorem by Darboux ensures that
around any point on a (2n)-dimensional manifold M there exists a
chart of canonical, or symplectic coordinates (q, p) in which
a closed nondegenerate 2form can be written as = dq i dpi .
Consequently the above description, though sound for M = E2n , is in general
valid only locally.
Comment 1.2.1 Spaces of 1-forms defined on Lie groups have natural global symplectic structures (Phys.10).
2
About this terminology, see Phys.2.1.1.

Canonical is a word as much abused in mathematics as in religion, and we shall
profit to do hommage to Liouville in the following.
3
490
1.3
Time evolution
1.3 To a contravariant field X, which is locally written as
+ Xqi i ,
pi
q
(1.9)
iX = Xqi dpi Xpi dq i .
(1.10)
X = Xpi
will correspond the covariant field
Applying XH to any given differentiable function F (q, p) on M , we find that

XH F =
F H
dF
F H
F dq i F dpi
+
=
.
=
i
i
i
q pi
pi q
q dt
pi dt
dt
(1.11)
This expression says that

XH F = {F, H} =
F H
F H
,
i
q pi
pi q i
(1.12)
the Poisson bracket of F and H, and that the equation of motion is

dF
= XH F = {F, H} ,
dt
(1.13)
the Liouville equation. The field XH flows the function along time: this
is precisely the role of a generator of infinitesimal transformations (section
6.4.2).
The operator XH is known in Statistical Mechanics as the Liouville operator, or liouvillian. Functions like F (q, p) are the classical observables, or
dynamical functions. The hamiltonian function presides over the time evolution of the physical system under consideration: we shall say that H(q, p) is
the generating function of the field XH . The time evolution of a dynamical
quantity F (q, p) is given by the solution of equation [1.13],
F (t) = F [q(t), p(t)] = etXH F (0) = F (0) + tXH F (0) +
t2
XH XH F (0) + . . .
2!
t2
{{F (0), H}, H} + . . .
2!
= F [etXH q(0), etXH p(0)] .
= F (0) + t{F (0), H} +
(1.14)
This is a purely formal expression, obtained by carelessly rearranging the

series without checking its convergence. F is an integral of motion if its Lie
1.4. CANONICAL TRANSFORMATIONS
491
derivative LXH F = XH F vanishes, or {F, H} = 0. The Lie derivative of

with respect to XH vanishes,
LXH = 0 ,
(1.15)
because LX = d iX + iX d. This means that the 2form is preserved by

the hamiltonian flow, or by the time evolution. If M is two-dimensional,
is the volume form and its preservation is simply Liouvilles theorem for one
degree of freedom. For (2n)dimensional M , the property
LX ( ) = (LX ) + (LX )
(1.16)
of the Lie derivative establishes with [1.15] the invariance of the whole series
of Poincare invariants . . . , including that one with the number
n of s, which is proportional to the volume form of M :
n = ()n dq 1 dq 2 . . . dq n dp1 dp2 . . . dpn .
(1.17)
The preservation of n by the hamiltonian flow is the general Liouville theorem. Dissipative systems will violate it.
Comment 1.3.1 Another structure is usually introduced on the manifold M : an
euclidean metric (, ) which, applied to two fields X and Y written in the manner of
[1.9], is written (X, Y ) = Xqi Yqi + Xpi Ypi . Amongst all the transformations on the 2n
dimensional space M preserving this structure (that is, amongst all the isometries), those
which are linear constitute a group, the orthogonal group O(2n).
1.4
Canonical transformations
1.4 The properties of the hamiltonian function and its related evolution
field are generalized as follows. A field X is a hamiltonian field if the form
is preserved by the local transformations generated by X. This is to say
that
LX = 0 .
(1.18)
Such transformations leaving invariant are the canonical transformations.
When n = 1, as [1.5] is the phase space area form, canonical transformations
appear as area-preserving diffeomorphisms.4 The 1form corresponding to
such a field will be closed because, being closed,
d(iX ) = LX = 0 .
4
(1.19)
Arnold 1966. Areapreserving diffeomorphisms are more general, as they may act on
non-symplectic manifolds.
492
If iX is also exact, so that

iX = dF
(1.20)
for some F (q, p), then X is said to be a strictly hamiltonian field5 (or globally
hamiltonian field) and F is its generating function. In a more usual language,
F is the generating function of the corresponding canonical transformation.
In reality, most hamiltonian fields do not correspond to a generating function
at all, as iX is not exact in most cases a generating function exists only
locally. On the other hand, as a closed form is always locally exact, around
any point of M there is a neighbourhood where some F (q, p) does satisfy
iX = dF . Notice that any field XF related to some dynamical function F
by (1.20) automatically fulfills
LXF = 0 .
(1.21)
This happens because LXF = d iXF + iXF d = d2 F = 0.

The one-to-one relationship established by allows one to adopt for
fields the same language used for forms. Any dynamical function F (q, p)
will correspond to a strictly hamiltonian field XF by iXF = dF . We may
say that XF is exact. Every field in this adapted language is closed.
Suppose now we have a field X satisfying the hamiltonian condition
LX = 0 at some point of M , but such that iX is not exact. If we force the
existence of a generating function F beyond its local natural validity, it will
be a multivalued function and the corresponding canonical transformation
will not be unique. We could talk in this case of non-integrable canonical
transformations. Of course, when the first real homology group H1 (M, R) is
trivial, every hamiltonian field will be exact.
Let us consider strictly hamiltonian fields. The dynamical function F
generates the strictly hamiltonian field
XF =
F
F
pi q i q i pi
(1.22)
Given two functions F and G, their Poisson bracket will have several different
though equivalent expressions:
F G F G
q i pi pi q i
= (XF , XG )
= XF (G) = XG (F ) = iXF iXG
= dF (XG ) = dG(XF ) .
{F, G} =
We follow here the terminology of Kirillov 1974.
(1.23)
(1.24)
(1.25)
(1.26)
1.4. CANONICAL TRANSFORMATIONS
493
The simplest examples of generating functions are given by F (q, p) = q i ,

corresponding to the field XF = - p i ; and G(q, p) = pi , whose field is XG
= q i . They lead to {q i , pk } = ki . Next in simplicity are the dynamical
functions of the type
fab = aq + bp ,
(1.27)
with a,b real constants. The corresponding fields are Jab = - a p

+ b q
.
The commutator of two such fields is [Jab , Jcd ] = 0, and consequently the
corresponding generating function F[Jab ,Jcd ] = F0 is a constant. The Poisson
brackets are the determinants
{fab , fcd } = (Jab , Jcd ) = ad bc .
(1.28)
Each dynamical function G will generate canonical transformations in a way

analogous to the time evolution [1.13]. The field XG will be the infinitesimal
generator of the corresponding local one-parameter group, the transformations taking place alon its local integral curve. Under a transformation
generated by G, another observable F will change according to
dF
= XG F = {F, G},
dr
(1.29)
r being the parameter along the integral curve of XG . The formal solution
of this equation is alike to [1.14],
F [q(r), p(r)] = erXG F (0) = F (0) + rXG F (0) +
= F (0) + r{F (0), G} +
r2
XG XG F (0) + . . .
2!
r2
{{F (0), G}, G} + . . .
2!
= F [erXG q(0), erXG p(0)] . (1.30)
We should furthermore insist on its local character: it has a meaning only

as long as XG has a unique integral curve. As long as it holds, the fields XG
extend the notion of liouvillian to generators of general canonical transformations.
Let us sum it up: to a contravariant field X like [1.9],
X = Xpi
+ Xqi i ,
pi
q
will make to correspond the covariant field

iX = Xqi dpi Xpi dq i .
(1.31)
494
Suppose another field is given, Y = Ypi p i + Yqi q i . The action of the 2-form
on X and Y will give
(X, Y ) = Xqi Ypi Xpi Yqi .
(1.32)
This is twice the area of the triangle defined on M by X and Y , as it is easy

to see in the example given by eqs.[1.27, 1.28].
Comment 1.4.1 A free particle in E3 will be described by a phase space E6 , with q i
(0)
(0)
and pi = mv i as Darboux coordinates. The symplectic form is simply (0) = dq i dpi .

(0)
A charged particle will have as conjugate momentum components pi = pi ec Ai (q) ) and
e
Fij dq i dq j . Notice that the condition
the consequent symplectic form is = (0) + 2c
dF = 0 is essential for to be closed.
1.5
Phase spaces as bundles
1.5 As said above, phase spaces are very particular cases of symplectic
manifolds.6 Not all symplectic manifolds are phase spaces. Think for instance
of the sphere S 2 . The area form endows S 2 with a symplectic structure
which does not correspond to a phase space. Phase spaces are the cotangent
bundles (6.4.7) of the configuration spaces. As such, they have, for each
point in configuration space, a non-compact subspace, the cotangent space,
dual and isomorphic to the tangent space. Consequently, phase spaces are
always non-compact spaces, which excludes, for instance, all the spheres.
Actually, the cotangent bundle T N of any differentiable manifold N has
a natural symplectic structure. Forms defined on T N are members of the
space T (T N ). Now, it so happens that T (T N ) has a canonical element
which is such that, given any section s : N T N , , the relation s
= s holds. Recall how we have introduced topology and charts (see for
example 6.4.3) on a tensor bundle. Take a chart (U, q) on N with coordinates
q 1 , q 2 , . . . , q n ; let be the natural projection : T N N , : Tq N q
N . Then, on 1 (U ) in T N a chart is given by (q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ).
). These coordinates represent on T N the Pfaffian form which in the natural
basis on U is written
U = pi dq i .
(1.33)
We recognize here the [1.8]. The differentiable structure allows the extension of U to the whole T N , giving a 1form (the Liouville form of the
cotangent bundle) which reduces to U in 1 (U ), to some V in 1 (V ),
etc. Define then the 2form = - d. It will be closed and nondegenerate
6
See for instance Godbillon 1969.
1.5. PHASE SPACES AS BUNDLES
495
and, in the chart (U, q) it will be given by expression [1.5]. Such a 2form
establishes a bijection between Tq N and Tq N at each point q N . To each
X T N corresponds iX T N .
A particularly simple case appears when the infinitesimal transformations
preserve the Liouville form itself, that is, when
LX = 0 .
(1.34)
The transformations are automatically canonical, as

L X = LX d = d L X = 0 .
It follows from LX = iX [d] + d[iX ] = 0 that iX = d[(X)]. A good
example is the angular momentum on the euclidean plane E2 : the rotation
generator
q2
X = q1
2
q
q 1
is related to an integral of motion if X(H) = 0. The integral of motion will
then be (X) = q 1 p2 q 2 p1 . The same holds for the linear momenta p1 and
p2 , generating functions for the translation generators /q 1 and /q 2 . If
we calculate directly the Poisson brackets, we find
{q 1 p2 q 2 p1 , p1 } = p2 ; {q 1 p2 q 2 p1 , p2 } = p1 ;
{q 1 p2 q 2 p1 , q 1 } = q 2 ; {q 1 p2 q 2 p1 , q 2 } = q 1 .
We see that they reproduce the algebra of the plane euclidean group. The
field algebra, nevertheless, is

=0.
X, 1 = 2 ; X, 2 = 1 ; X,
q
q
q
q
pj
In reality, (X, /p1 ) = q 2 and (X, /p2 ) = - q 1 as it should be,
but
(X, 1 ) = (X, 2 ) = 0,
q
q
so that
(X, 1 ) 6= {q 1 p2 q 2 p1 , p1 } .
q
Comment 1.5.1 This is related to lagrangian manifolds. An n-dimensional subspace
of the 2n-dimensional phase space M is a Lagrange manifold if (X, Y ) = 0 for any
two vectors X, Y tangent to it. That is, the restriction of to is zero. Examples
are the configuration space itself, or the momentum space. The angular momentum X
above is a field on configuration space. One must be careful when Lagrange manifolds
are present, since may be degenerate. Of course, canonical transformations preserve
Lagrange manifolds, that is, they take a Lagrange manifold into another one.
496
1.6
The algebraic structure
1.6 We have been using above the holonomic base {/q i , /pj } for the
vector fields on phase space. In principle, any set of 2n linearly independent
fields may be taken as a basis. Such a general basis {ei } will have its dual,
the base { j } with j (ei ) = ij , and its members will have commutators
[ei , ej ] = ck ij ek ,
where the structure coefficients ck ij give a measure of the basis anholonomicity. A general field will be written X = X i ei = i (X)ei ; a general 1-form,
= i i = (ei ) i ; the differential of a function F will be dF = ei (F ) i ; and
so on. The symplectic 2form will be
= 21 ij i j = 12 (ei , ej ) i i .
(1.35)
Consider the (antisymmetric) matrix = (ij ) and its inverse 1 = (ij ),

whose existence is identical to the nondegeneracy condition:
ij jk = ij jk = ki .
(1.36)
As the interior product is that 1form satisfying

iX (Y ) = (X, Y )
for any field Y , its general expression is
iX = X i ij j .
(1.37)
The component of a strictly hamiltonian field can then be extracted:

XFj = ek (F ) kj .
(1.38)
The Poisson bracket is

{F, G} = (XF , XG ) = XFi ij XGj = ek (G) kj ej (F ) .
(1.39)
This gives the Poisson bracket in terms of the inverse to the symplectic
matrix. An interesting case occurs when the Liouville form is preserved by
all the basis elements, which are consequently all strictly hamiltonian: from
Lek = 0 it follows that iek = dfk , with fk = (ek ). As = - d, then

ij = d(ei , ej ) = 12 ej (fi ) ei (fj ) + ck ij fk .
1.6. THE ALGEBRAIC STRUCTURE
497
As also ei (fj ) = {fi , fj }, it follows that {fi , fj } = ck ij fk . The Poisson brackets

of the generating functions mimic the algebra of the corresponding fields.
If we come back to the Darboux holonomic basis related to the coordinates
k
{x } = {q i , pj }, the vector base {/xk } will be ek = {/q k } for k =
1, 2, . . . , n and ek = {/pk } for k = (n + 1), (n + 2), . . . , (2n). The matrices
and 1 will have the forms

0 In
0 In
1
=
and =
,
(1.40)
In 0
In
0
where In is the n-dimensional unit matrix. In terms of the collected coordinates {xk }, Hamiltons equations are then
dxk
= dH(ek ) .
dt
(1.41)
Modern approaches frequently define the hamiltonian formalism via the Poisson bracket, introduced as
{F, G} = hij ei (F )ej (G)
with hij = ij .
This is advantageous (see Phys.2.1.1) in the study of the relationship of the

Poisson algebra to another Lie algebra of specific interest in a particular
problem. Such a relationship may be better seen in some special basis. For
example, the configuration space for a rigid body turning around one of
its points is the group SO(3) and it is convenient to choose a hamiltonian
structure in which the group Lie algebra coincides with the Poisson bracket
algebra. This happens if {Mi , Mj } = ijk Mk , in which case we must choose
hij (x) = ijk Mk . This means that h depends only on the coordinates, not on
the conjugate momenta. In the holonomic basis related to Darboux coordinates, hij (x) is a constant. The case above is a special example of a wide
class of problems in which hij (x) is linearly dependent on the xs.7
Notice that 2 = I2n . In terms of the euclidean scalar product (X, Y )
= Xqi Yqi + Xpi Ypi , we see that
(X, Y ) = Xqi Ypi Xpi Yqi = X T Y = (X, Y ) .
A complex structure may be introduced by using the complex representation
X = (Xq , Xp ) X = Xq + iXp ,
(1.42)
by which the manifold M becomes locally the complex n-dimensional space

Cn . In this representation we see immediately that X = iX, coherently
7
Novikov 1982.
498
with the above remark that 2 = I. The linear transformations preserving

this complex structure constitute the complex linear group GL(n, C).
Matrix may be seen as a twisted metric, defining a skewsymmetric
inner product
hX, Y i := (X, Y ) = ij X i X j = (X, Y ) = hY, Xi,
which is quite equivalent to the symplectic structure. Of course, canonical transformations preserve all that. In particular, we may consider linear
canonical transformations, given by those matrices S preserving the metric : S 1 S = . Such linear transformations on the (2n)dimensional
manifold M are the symplectic transformations and also constitute a group,
the symplectic group Sp(2n).
Let us recapitulate: the linear transformations which preserve the euclidean scalar product form the group O(2n); those preserving the symplectic
structure constitute the symplectic group Sp(2n); and those preserving the
complex structure, the group GL(n, C). Transformations preserving simultaneously these three structures will be in the intersection of the three groups.
As it happens,
O(2n) Sp(2n) = Sp(2n) GL(n, C) = GL(n, C) O(2n) = U (n), (1.43)
the unitary group of the n n unitary complex matrices. Summarizing, the
unitary transformations preserve the hermitian scalar product
H(X, Y ) = (X, Y ) + i(X, Y ).
(1.44)
Comment 1.6.1 A short additional comment on the lagrangian manifolds of Comment 1.5.1. Consider a simple phase space M = E2n . Given an nplane En in M , it is
called a lagrangian plane if, for any two vectors X and Y in En , (X, Y ) = 0. An alternative definition of lagrangian manifold is just as follows: any ndimensional submanifold
of M whose tangent spaces are all lagrangian planes. An interesting point is that the set
of all such planes is itself a manifold, called the lagrangian grassmannian (n) of M . This
manifold has important topological characteristics:
H1 ((n), Z) H 1 ((n), Z) 1 ((n), Z) Z .
Actually, it happens that (n) = U (n)/O(n). The group U (n) acts transitively on (n)
and makes here a rare intrusion in Classical Mechanics.
1.7
Relations between Lie algebras
1.7 The Poisson bracket is antisymmetric and satisfies the Jacobi identity.
It is an operation defined on the space C (M, R) of real differentiable functions on M . Consequently, C (M, R) is an infinite-dimensional Lie algebra
1.7. RELATIONS BETWEEN LIE ALGEBRAS
499
with the operation defined by the Poisson bracket. Actually, F XF is a Lie

algebra homomorphism (that is, a representation, see Math.6) of C (M, R)
into the algebra of strictly hamiltonian fields on M . We shall now say a few
words on the relations between these two Lie algebras.
Unless it is an isomorphism (a faithful representation), a homomorphism such as the above F XF loses information. The connection between commutators of fields and the Poisson brackets of the corresponding
generating functions is not immediate. This may be guessed from the most
trivial one-dimensional case
F = q , G = p , XF = /p , XG = /q .
The commutator [XF , XG ] vanishes while the corresponding Poisson bracket
{q, p} = 1. Notice, however, that the commutator of two hamiltonian fields
is always strictly hamiltonian, as, for any function K,
[XF , XG ]K = XF {K, G} XG {K, F } = {{K, G}, F } + {{F, K}, G}
= {{G, F }, K} = X{F,G} K .
(1.45)
[XF , XG ] = X{F,G}
(1.46)
[XF , XG ]K = d{F, G}(XK ) .
(1.47)
dF[X,Y ] = dFX , FY
(1.48)
Therefore,
and
This means that
(which can be alternatively obtained by using the identity i[X,Y ] = LX iY iX LY ), from which it follows that
F[X,Y ] = {FX , FY } + (X, Y ) ,
(1.49)
(X, Y ) being a constant [as d(X, Y ) = 0] depending antisymmetrically

on the two argument fields. Such a constant is thus typically the effect of
applying a 2form on X and Y . Unless (X, Y ) vanishes, the generating
function corresponding to the commutator is not the Poisson bracket of the
corresponding generating functions. Application of the Jacobi identity to
both sides of [1.49], in a way analogous to the above reasoning concerning
, shows that must be a closed form. The appearance of a cocycle like
is rather typical of relations between distinct representations. It would be
better to use the word action, as things may become very different from the
relationship usually denoted by representation. Because its presence frustrates an anticipation of simple formal algebraic likeness, we might venture
500
to call an anomaly, a word which became popular for analogous failures in

quantization procedures.
The presence of this 2cocycle is related to the cohomology of the field
Lie algebra. Generating functions are defined only up to a constant and
the cohomology classes are connected with this freedom of choice. In effect,
choose new functions
0
FX0 = FX + (X) , FY0 = FY + (Y ) , F[X,Y
] = F[X,Y ] + ([X, Y ]) ,
with (X), (Y ) and (([X, Y ])) constants corresponding to the argument

fields. Then, [1.49] becomes
0
0
0
0
F[X,Y
] = {FX , FY } + (X, Y ) .
where 0 (X, Y ) = (X, Y ) + ([X, Y ]). Now, may be seen as a 1form on

the Lie algebra, which gives the constants when applied to the fields. In this
particular case, from the general expression for the derivative of a 1form,
2d(X, Y ) = X[(Y )] Y [(X)] ([X, Y ]) ,
we see that 0 (X, Y ) = (X, Y ) 2 ([X, Y ]). Consequently, 0 may be put
equal to zero if an can be found such that = 2 d. The 2cocycle
is then exact, that is, a coboundary. Summing up: the general relationship
between generating functions related to global hamiltonian fields and the
generating functions related to their commutators is given by [1.49], with
a cocycle on the Lie algebra of fields. When is also a coboundary, the relationship becomes a direct translation of commutators into Poisson brackets
if convenient constants are added to the generating functions. The cocycle
defines a cohomology class on the field Lie algebra. Field commutators and
Poisson brackets are interchangeable only if this class is trivial.8
Let us finally comment on the closedness of the symplectic form. Why
have we insisted so much that be a cocycle? Using the above relations, we
find that
(X, [Y, Z]) = [Y, Z]FX = {FX , {FY , FZ }} .
(1.50)
Combined with the general expression for the differential of a 2form, which
is
3!(d(X, Y, Z)) = X((Y, Z)) + Z((X, Y )) + Y ((Z, X))
+ (X, [Y, Z]) + (Z, [X, Y ]) + (Y, [Z, X]) ,
that equation gives
3d(X, Y, Z) = FX , FY , FZ FZ , FX , FY FY , FZ , FX = 0 .
8
Arnold 1976, Appendix 5.
(1.51)
1.8. LIOUVILLE INTEGRABILITY
501
We see in this way the meaning of the closedness of : it is equivalent to the

Jacobi identity for the Poisson bracket.
1.8
Liouville integrability
1.8 A hamiltonian system with n degrees of freedom, whose flow is given by

the hamiltonian function H, is integrable if there exists a set of n independent
integrals of motion {Fi (q, p), i = 1, 2, . . . , n} in involution, that is, such that
{H, Fi } = 0 ,
{Fi , Fj } = 0 .
(1.52)
(1.53)
This is the most widely used formulation of integrability, due to Liouville. Of

course, the first equation above declares that the Fi s are integrals of motion.
The second is the involution condition. Notice that H is not independent
of the Fi s. From [1.47] it follows that the corresponding fields commute,
[Xi , Xj ] = 0. Equation [1.12] will say that Xi (Fj ) = 0, which is the same
as dFi (Xj ) = 0. All this means that there will be ndimensional integral
manifolds tangent to the Xj s, which are furthermore level manifolds Fk (q, p)
= constant. As the involution condition is the same as (Xi , Xj ) = 0, such
level manifolds are actually lagrangian manifolds.
Arnold 1976
Goldstein 1980
Abraham & Marsden 1978
Arnold, Kozlov & Neishtadt 1988
502
Phys. Topic 2
MORE MECHANICS
HamiltonJacobi
1 Hamiltonian structure
2 HamiltonJacobi equation
The Lagrange derivative
3 The Lagrange derivative as a covariant object

The rigid body
4
5
6
7
8
9
10
11
12
13
Frames
The configuration space
The phase space
Dynamics
The space and the body derivatives
The reduced phase space
Moving frames
The rotation group
Left- and right-invariant fields
The Poinsot construction
2.1
2.1.1
HamiltonJacobi
Hamiltonian structure
Modern authors prefer to define a hamiltonian structure as follows. Take

a space M and the set R(M ) of real functions defined on M (we do not
fix the degree of differentiability for the moment: it may be C (M ), to fix
the ideas). Suppose that R(M ) is equipped with a Poisson bracket {, }M :
R(M ) R(M ) R(M ), given for any two functions F and G as
{F, G}M = hij (x) i F j G,
503
(2.1)
504
PHYS. TOPIC 2.
MORE MECHANICS
where the functions hij (x) are such that the bracket is antisymmetric and
satisfies the Jacobi identity. The space M is then called a Poisson manifold, and the bracket is said to endow M with a hamiltonian structure. Of
course, hij (x) is just the inverse symplectic matrix ij (x) when it is invertible (compare Eq.(2.1) with Eq.(1.39) of Phys.1). The preference given to
this definition of hamiltonian structure, based on the bracket, comes from
the study of systems for which the matrix [hij (x)] is well defined but not
invertible at every point x.
Consider two Poisson manifolds, M and N . A mapping f : M N is a
Poisson mapping if
{f F, f G}M = {F, G}N
(2.2)
where f is the pullback, f F = F f . Take two symplectic manifolds
(M1 , 1 ) and (M2 , 2 ); call 1 and 2 respectively the projections of M1 M2
into M1 and M2 , 1 : M1 M2 M1 and 2 : M1 M2 M2 . Consider also
a mapping f : M1 M2 , with graph f . Let further if : f M1 M2 be
the inclusion. Then,1
= 1 1 2 2
is a symplectic form on the product M1 M2 . If if = 0, the mapping f is
said to be a symplectic mapping.
Comment 2.1.1 We shall not in the following make any distinction between Poisson
and symplectic manifolds and/or mappings.
The graph f is a lagrangian submanifold. Write locally = - d (it

might be = 1 1 2 2 , but not necessarily). Then, if d = dif , so
that f is symplectic iff if is closed. In that case, locally, if = - dS for
some S: f R. The function S is the generating function for the mapping f . Thus, given f , S is a dependent real function defined on the lagrangian submanifold. If (q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ) are coordinates on M2
and (Q1 , Q2 , . . . , Qn , P1 , P2 , . . . , Pn ) are coordinates on M1 , then there are
many ways to chart the graph f on which the function S is defined. Examples are S(q 1 , q 2 , . . . , q n , Q1 , Q2 , . . . , Qn ) and S(q 1 , q 2 , . . . , q n , P1 , P2 , . . . , Pn ).
Notice that S is only locally defined.
Comment 2.1.2 Lagrangian submanifolds are of great import to the semi-classical
approximation to Quantum Mechanics. One of their global properties, the Maslov index,2
appears as the ground state contribution.3
1
2
3
Abraham & Marsden 1978.

Arnold (Appendix), in Maslov 1972.
Arnold 1976, Appendix 11.
2.1. HAMILTONJACOBI
2.1.2
505
Hamilton-Jacobi equation
Take again,4 as above, (Q1 , Q2 , . . . , Qn , P1 , P2 , . . . , Pn ) as coordinates on M1

and (q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ) as coordinates on M2 :
f (Q1 , Q2 , . . . , Qn , P1 , P2 , . . . , Pn ) = (q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ) .
If we now consider S(q 1 , q 2 , . . . , q n , Q1 , Q2 , . . . , Qn ), the expression
if = dS
(2.3)
S
S
enforces pk = q
k and Pk = - Qk . Suppose we find a symplectic mapping f
such that S is independent of the Qj s. Then, S becomes simply a function
on the configuration space. In this case all the Pj = 0 and the hamiltonian
is a constant, H = E or, in the remaining variables,

S
k
=E .
(2.4)
H q ,
q k
This is the time-independent HamiltonJacobi equation. A curve c(t) in

= dS (c(t)) is an integral curve of the
configuration space such that dc(t)
dt
field XH . The surfaces S = constant are characteristic surfaces and c(t) is a
gradient line of S, orthogonal to them.
The time evolution of a mechanical system of hamiltonian H is given by
the Liouville field
H
H
i
.
XH =
i
pi q
q pi
The hamiltonian flow is precisely the one-parameter group generated by XH .
Given any dynamical function F (q, p, t), it will evolve according to the equation of motion
dF
= {F, H} = XH F ,
dt
the Liouville equation (eq.(1.13) of Phys.1). Consider this equation in some
more detail: in terms of the coordinates {xk } = {q 1 , q 2 , . . . , q n , p1 , p2 , . . . ,
pn }, it is a partial differential equation
2n
X
d
i
F (q, p, t) =
XH
(x) i F (x, t) ,
dt
dx
i=1
(2.5)
to be solved with some given initial condition

F (q, p, 0) = f0 (q, p) .
4
See Babelon & Viallet 1989.
(2.6)
506
PHYS. TOPIC 2.
MORE MECHANICS
If Ft (x) is a flow, the solution will be F (x, t) = fo (Ft (x)). The orbits of
the vector field XH are the characteristics of the above differential equation.
Notice that they will fix the evolution of any function. The curve solving
Hamilton equations (which are ordinary differential equations) in configuration space is the characteristic curve of the solutions of the HamiltonJacobi
equations (which are partial differential equations). More will be said on
characteristics in Phys.4.
A famous application comes out in Quantum Mechanics, where S appears
as the phase of the wavefunction: = exp[iS/~]. Then, with H(q, p) =
p2 /2m + V (q), the Schrodinger equation
~2
+ V = E
2m
leads to the quantum-corrected HamiltonJacobi equation

~
1
(S)2 + V = E + i
2 S ,
2m
2m
(2.7)
where the last term is purely quantal.

Mathematicians tend to call Hamilton-Jacobi equation any equation of
the type H(x, x ) = 0, that is, any equation in which the dependent quantity
f does not appear explicitly.5 The eikonal equation

x1
2

+
x2
2

+ ... +
xn
2
=1
(2.8)
is, of course, an example (see Phys.4 and Phys.5). The function is the
optical length, and the level surfaces of are the wave fronts. The characteristics of H(x, x ) = 0 obey differential equations which are just the
Hamilton equations
H
H
, p =
.
x =
p
x
The projections of the trajectories on the xspace are the rays.
Given a hypersurface in En , define (x) as the distance of the point x
to . Then satisfies the eikonal equation. Actually, any solution of this
equation is, locally and up to an additive constant, the distance of x to some
hypersurface.
5
Arnold 1980.
2.2. THE LAGRANGE DERIVATIVE
2.2
507
The Lagrange derivative
We shall not really examine the lagrangian formalism, which is summarized in Math.8
and Phys.6 and pervades many other chapters. The intention here is only to present the
Lagrange derivative of Classical Mechanics as an example of covariant derivative.
2.2.1
The Lagrange derivative as a covariant derivative
The configuration space M is the space spanned by the values of the degrees
of freedom. Its points are described by a coordinate set x = xk , one xk for
each degree. The velocity space is described accordingly by x = {x k }, where
k
x k = dx
. The lagrangian function L(x, x,
t) is defined on the combined
dt
configurationvelocity space, which is actually the tangent bundle T (M ) of
M . The extremals of L(x, x,
t) are curves (t) satisfying the EulerLagrange
equations, or equations of motion (see Math.7).
Comment 2.2.1 That the combined space is T (M ) is a simplicity assumption. No
system has been exhibited where this hypothesis has been found wanting. There is no
problem in identifying the point set, but the supposition means also that there is a projection, with the charts of the configuration space being related to those of the combined
space, etc.
Comment 2.2.2 When the configuration space is non-trivial, that is, when the degrees of freedom take values in a non-euclidean space, many local systems of coordinates
may be necessary to cover it, but the lagrangian function should be independent of the
number and choice of the charts.
Consider a time-independent change of coordinates in configuration space,
x y k = y k (xj ), which:
(i) is invertible, that is, we can find xj = xj (y k );
(ii) takes time as absolute, so that the velocities simply follow the configk
uration transformation, y k = dydt ;
(iii) leaves invariant the lagrangian function, L(x, x,
t) = L(y, y,
t).
In the lagrangian formalism, the coordinates of the configuration and
velocity spaces are independent in each chart but, once a coordinate transformation is performed, the new velocities may depend on (say) the old
coordinates. A first important thing is that, despite the highly arbitrary
character of the transformation, the velocities (tangent vectors) are linearly
k
k
transformed: y k = dydt = y
x j . This simply says that they are indeed vectors
xj
under coordinate transformations on the configuration space. Notice that
j
x n
xn
=
.
y j
y j
(2.9)
508
PHYS. TOPIC 2.
From
x k =
follows
MORE MECHANICS
xk n
y
y n
x k
2 xk
=
y n .
j
j
n
y
y y
(2.10)
The Euler-Lagrange equations are conditions fixing the extrema of L,

L
so that it is necessary to examine x
j . If L were a simple function of the
coordinates, the derivative would be a vector and all should be well, but the
velocity dependence embroils the things. Notice first that
L
L x n
L xn
=
=
.
y j
x n y j
x n y j
(2.11)
is also a good vector,

Which shows that the conjugate momentum pn = L
x n
though in a converse way: it transforms by the inverse of the matrix transforming the velocity. The velocity is a contravariant vector, or simply vector.
The converse behaviour shows that the momentum is a covariant vector, or
covector. Anyhow, it behaves covariantly under coordinate transformations
on the configuration space. Its time derivative, however, does not:

d L x n
L 2 xk
d L
=
+
y n .
(2.12)
dt y j
dt x n y j
x k y n y j
The first term on the right-hand side would be all right, but the second
L
represents a serious deviation from tensorial behaviour. As to x
j itself, it is
ill-behaved from the start:
L xn
L x k
L xn
L 2 xk
L
=
+
=
+
y n .
y j
xn y j
x k y j
xn y j
x k y j y n
The same offending term of (2.12) turns up, and we conclude that

L
d L
xk L
d L
.
y j
dt y j
y j xk dt x k
(2.13)
(2.14)
The operator acting on L has the formal properties of a derivative: it is

linear and obeys the Leibniz rule. This modified derivative is the Lagrange
derivative. Unlike the usual derivative, it gives a well-behaved, tensorial
object under coordinate changes in configuration space. The terms dtd L
y j
d L
and dt x k are compensating terms, alike to the compensating contributions
of the gauge fields or connections in the covariant derivatives.
2.2. THE LAGRANGE DERIVATIVE
509
Actually, we have not used any characteristic of L as a lagrangian. The

above results say simply that derivatives on configuration space of any function F (x, x)
on the combined configurationvelocity space (which is the tangent bundle of the configuration space) must be supplemented with an extra term in order to have a coordinate-independent meaning. The good,
coordinate-independent derivative of any function is thus the Lagrange derivative
:=
.
(2.15)
k
k
x
x
dt x k
The reasoning remains true for higher-order lagrangians, dependent on the
acceleration, second acceleration, etc. The Lagrange derivative must be accordingly adapted, the compensating terms being then an alternate sum of
higher-order contributions:
d
d2
d3
:=
... + . . .
xk
xk dt x k dt2 xk dt3 x k
(2.16)
Comment 2.2.3 The Lagrange derivative is not invariant under a general transformation in velocity space. Notice that, in the simple L = L(x, x)
case, the non-covariant
d L
d k
compensating term dt
p
,
the
newtonian
expression
for the force. That force
is
just
k
dt
x
has, consequently, no coordinate-independent meaning.
The complete Lagrange derivative is essential to obtain consistent expressions6 for the force Qi coming from velocity dependent potentials V (q, q),
such as the Lorentz force of electrodynamics.

The
generalized
force,
apP
pearing in the expression W = i Qi dq i for the work, is
Qk :=
V
d V
V
= k +
.
k
q
q
dt qk
(2.17)
In this case, with L = T V , the Lagrange equations of motion retain the

form
L
=0.
(2.18)
q k
An example is Webers law7 of attraction which comes from

1
r 2
V =
1+ 2
r
c
and leads to the complete Ampères law.8 The same holds for the conjugate
momentum, if L depends on q.
6
7
8
Whittaker 1944, p. 44 on.

Whittaker 1953, p. 226 on.
Assis 1989.
510
PHYS. TOPIC 2.
MORE MECHANICS
Comment 2.2.4 The degrees of freedom xk are indexed here by the discrete labels
k. When there is a continuum of degrees of freedom, each degree becomes a function
of the labels, which are then indicated (say) by x. Each degree is then a field
(x). Systems with a continuous infinity of degrees of freedom are discussed in Math.8
and Phys.6.
Comment 2.2.5 In a system with several degrees of freedom, the individual Euler
Lagrange equations are not necessarily invariant under coordinate transformations. It is
their set which is invariant. Each equation is said to be covariant, not invariant.
This reminds us of the components of vector fields and forms: components are covariant,
though vector and covector fields are invariant. In effect, the Euler-Lagrange equations
can be assembled in a certain functional differential form, which is invariant and can be
defined also for sets of equations which do not come as extremal conditions on a Lagrange
function (Math.8).
Comment 2.2.6 We have been cheating a little. Recall that t is not necessarily
time: it is actually a curve parameter, and all the above derivatives are concerned
with points on a curve. What really happens is that the Lagrange derivative is the good
derivative on a functional space, the space of functionals on the space of trajectories. A
wellknown example of such a functional is the action functional (see Math.7).
2.3
The rigid body
The study of rigid body motion has many points of interest: (i) it gives an example of
nontrivial configuration space, which is furthermore a Lie group whose Maurer-Cartan
form has a clear physical interpretation; (ii) it gives a simple example of a metricinduced
canonical isomorphism between a vector space and its dual; (iii) it illustrates the use of
moving frames; (iv) it shows the difference between left and rightaction of a group.
2.3.1
Frames
A rigid body is defined as a set of material points in the metric space E3 such
that the distance between any pair of points is fixed. Any three non-colinear
points on the so defined rigid body define a frame, in general nonorthogonal,
in E3 . This frame attached to the body will be indicated by F 0 . It is called
the body frame, in oposition to a fixed frame F given a priori in the vector
space E3 , called the space frame, or laboratory frame. We are taking
advantage of the double character of E3 , which is both a manifold and a
vector space. Given the position of the three points, the position of any
other point of the rigid body will be completely determined. We say that the
configuration is given. Thus, in order to specify the position of any body
point, we need only to give the positions of the three points. These would
2.3. THE RIGID BODY
511
require 9 coordinates, which are however constrained by the 3 conditions

freezing their relative distances. Consequently, the position of any point will
be given by 6 coordinates.
We may take one of the three points as the origin O of the frame F 0 .
From an arbitrary starting configuration, any other may be obtained by
performing two types of transformations: (i) a translation taking O into any
other point of E3 , and (ii) a rotation around O. There is consequently a one
toone correspondence between the set of configurations and the set of these
transformations. This set of transformations is actually a 6dimensional
group, denoted by SO(3) R3 , the direct product of the rotation group
SO(3) by the translation group R3 . Thus, the configuration space of a rigid
body is (the manifold of) SO(3) R3 . This manifold may be identified to
the tangent bundle T SO(3), which is a direct product because SO(3), as any
Lie group, is parallelizable.
Comment 2.3.1 This is not to be mistaken by the euclidean group, the group of
transformations (motions, or isometries) in our ambient E3 , which is the semi-direct product SO(3) R3 , a non-trivial bundle. The difference comes from the fact that, in the
latter, rotations and translations do not always commute.
2.3.2
The configuration space
We shall consider the case in which the body has a fixed point, and take that
point as the origin O for both frames F and F 0 . There are no translations
anymore: the configuration space of a rigid body moving around a fixed
point reduces to SO(3). We are supposing a fixed orientation for the body,
otherwise the configuration space would be O(3). As a manifold, SO(3) is a
half-sphere: SO(3) S 3 /Z2 . When we use angles as coordinates on SO(3)
(say, the Euler angles), the members of the tangent space will be angular
velocities. We may go from such starting coordinates to other generalized
coordinates, of course. It is of practical interest to identify the frame origin
O also to the group identity element. The space tangent to the configuration
space at O will then be identifiable to the Lie algebra so(3) of the group,
whose generators {Li } obey [Li , Lj ] = ijk Lk . As the structure constants are
just given by the Kronecker symbol ijk , the Lie operation coincides with the
usual vector product Li Lj = ijk Lk in E3 .
2.3.3
The phase space
The phase space will be the cotangent bundle T SO(3). The members of
the cotangent space will be the dual to the angular velocities, that is, the
angular momenta. In problems involving rotational symmetry, one frequently
512
PHYS. TOPIC 2.
MORE MECHANICS
starts with phase space coordinates (q i , pk ) and find the angular momenta
M1 = q 2 p3 q 3 p2 , etc, as conserved quantities. They satisfy the Poisson
bracket algebra {Mi , Mj } = ijk Mk . This means that the Mk s provide a
representation of so(3) (the Lie algebra of the group SO(3)) in the Poisson
bracket Lie algebra of all functions on the phase space. In some cases, it
is convenient to use the Mk s themselves as generalized coordinates. The
Poisson bracket
{F, G} = hij (x) i F j G
in this case has hij (x) = ijk M k for i, j = 1, 2, 3; for i, j = 4, 5, 6 it has hij (x)
= 1 when j = i 3; and hij (x) = 0 for all the remaining cases. Furthermore,
the invariant Cartan-Killing metric of SO(3) is constant and euclidean, so
that in this metric we are tempted to write Mi = M i . Nevertheless, another
metric is present, the moment of inertia Iij . A comparison with the case of
optics is helpful here. The Cartan-Killing metric plays the same role played
by the euclidean E3 metric in optics, while the moment of inertia has some
analogy to the refractive metric (see Phys.5).
2.3.4
Dynamics
Dynamics is presided by the hamiltonian

H=
1
2
ij
I ij Mi Mj ,
(2.19)
where I ij = (I 1 )ij are the entries of the inverse to the moment

P of inertia
matrix. The hamiltonian H can be diagonalized as H = 12 i ai Mi2 , and
the angular velocity is defined as k = H/Mk . The Liouville equation
becomes
= {M, H} = M ,
M
which is Eulers equation for the rigid body motion. Actually, all these
quantities are referred to the body frame, and will be indexed with b.
Thus, when a rigid body moves freely around a fixed point O, its angular
momentum Mb and angular velocity b with respect to O are related by
Eulers equation
dMb
+ b Mb = 0 .
(2.20)
dt
Comment 2.3.2 The same equation follows from the lagrangian L =
Mi i H.
Comment 2.3.3 With respect to space, the angular momentum Ms satisfies

0, which expresses the conservation of overall angular momentum.
Ms
dt
2.3. THE RIGID BODY
2.3.5
513
The space and the body derivatives
The space and the body derivatives of the components of a vector quantity G are related by

dG
dt

=
space
dG
dt

+ b G .
(2.21)
body
An example is the relation between the velocities vs = vb + b r. Another is

given by the above relationship between the rates of variation of the angular
momenta.
The linear velocity of a point at position r is given by vb = b r.
The points on the axis instantaneously colinear with the angular velocity ,
given by r = a b for any a, have vanishing velocities. They constitute the
instantaneous axis of rotation.
2.3.6
The reduced phase space
There are, therefore, 4 integrals of motion: the three components of M and

the energy E. The reduced phase space, in which the motion forcibly takes
place, will be a 2dimensional subspace of T SO(3), determined by the constraints M = constant and E = constant. This 2dimensional subspace is the
torus T 2 . That this is so comes from a series of qualitative considerations:9
(i) the subspace admits global motions, i.e., given the initial conditions,
the system will evolve indefinitely along the flow given by a vector field,
which is consequently complete, without singularities; (ii) it is connected,
compact and orientable; (iii) the only 2dimensional connected, compact
and orientable manifolds are the sphere and the multiple toruses with genus
n = 1, 2, . . .. The torus T 2 is the case n = 1. In order to have a complete
vector field, the manifold must have vanishing Euler number. Here, =
b0 b1 + b2 . Connectedness implies b0 = 1, Poincare duality implies b2 = b0 ,
so that we must have b1 = 2. The genus is just b1 /2, so that we are forced to
have n = 1. On the torus, we may choose two angular coordinates, 1 and
2
1
= 1 and d
= 2 . This means
2 , and find the equations of motion as d
dt
dt
that the motion of a rigid body with a fixed point can be reduced to two
periodic motions with independent, possibly incommensurate frequencies. In
the last case, the body never comes back to a given state and we have an
example of deterministic chaotic motion (see Math.3.9).
9
See Arnold 1976.
514
2.3.7
PHYS. TOPIC 2.
MORE MECHANICS
Moving frames
Let us consider again the two frames F and F 0 with the same origin O. Take
a cartesian system of coordinates in each one. A point will have coordinates
x = (x1 , x2 , . . . , xn ) in F and x0 = (x01 , x02 , . . . , x0n ) in F 0 . Let us first simply
consider the motion of an arbitrary particle with respect to both frames.
The coordinates will be related by transformations x0i = Aij xj and xi =
(A1 )ij x0j .
Compare now the velocities in the two frames. To begin with, the point
k
will have velocity vF = x of components x k = xdt in the space frame F and
0k
vF0 0 = x 0 of components x 0k = xdt in the space frame F 0 . Here, of course, the
absolute character of time (t is the same in both frames) is of fundamental
importance. But there is more: we may want to consider the velocity with
respect to F as seen from F 0 , and vice-versa. Let us call vF0 the first and vF 0
the velocity with respect to F 0 as seen from F . Let us list the velocities in
the convenient notation
(seen from)
v(with respect to)
vF = x = velocity with respect to, and seen from, the space frame F ;
vF0 = velocity with respect to F as seen from F 0 ;
vF 0 = velocity with respect to F 0 as seen from F ;
vF0 0 = x 0 = velocity with respect to, and seen from, the rotating frame F 0 .
As velocities are vectors with respect to coordinates transformations, vF0 =
A vF . Also, vF0 0 = A vF 0 and vF 0 = A1 vF0 0 . But x 0 = vF0 0 = Ax + Ax,

1
0
1
so that vF 0 = A vF 0 = vF + A Ax. It will be useful to write this in

components,
dAji i
vFk 0 = vFk + (A1 )kj
x ,
(2.22)
dt
which means that vFk 0 dt = dxk + (A1 )kj dAji xi . For a particle belonging to
the rigid body, consequently fixed in F 0 , vF0 0 = 0 and vF 0 = 0. Thus,
vFk = (A1 )kj A ji xi .
(2.23)
ki = (A1 )kj A ji
(2.24)
Let us call
the angular velocity tensor. Then we have
vFk = ki xi ,
(2.25)
2.3. THE RIGID BODY
515
which is the equation v = r when n = 3. In this case, the usual relationship of antisymmetric tensors to vectors allows one to define the vector
angular velocity j from the tensor angular velocity by ki = kij j , or
j = 12 jki (A1 )kn A ni .
(2.26)
The wellknown consequence is that matrix action on column vectors turns

into vector product. We actually find vF = r. We might invert all
the discussion, taking F 0 as fixed and F as turning. The whole kinematics
is equivalent, with only an obvious change of sign in the angular velocity.
The same treatment holds, of course, for other vectors under the coordinate
transformations.
2.3.8
The rotation group
Each matrix A taking a vector given in F into the same vector in F 0 represents
a rotation. It is a member of the rotation group SO(3), and the form =
A1 dA is the SO(3) canonical (or Maurer-Cartan) form. The angular velocity
tensor is the result of applying this form to the field dtd , tangent to the particle
trajectory,

d
dA
1
.
= A1
A dA
dt
dt
Thus, the angular velocity is the canonical form along the trajectory. But
the role of the canonical form is to take any vector field on the group into
a vector field at the identity. That is, into the Lie algebra of the group. If
i
A = e Ji , then
= e
so that

d
= e
dt
i Ji
i Ji
dk Jk e
j Jj
Jk e
jJ
d
dt
= e
i Ji
Jk e
jJ
dk ,

dk
dk
= (AdA1 )k j Jj
= Jk0
dt
dt
belongs to the Lie algebra. We see in this way how the angular velocities turn
up in the Lie algebra so(3) of SO(3). By the way, the above considerations
show also that = (AdA1 J)k dk = AdA1 (d).
2.3.9
Left and rightinvariant fields
We can transport a tangent vector into the group identity by two other
means: left-translation and right-translation. To each position of the body
516
PHYS. TOPIC 2.
MORE MECHANICS
corresponds an element of the group. Take an initial position of the body

and identify it (arbitrarily) to the identity element. One obtains every other
position by applying group elements. Take some J in the algebra and consider
the one-parameter group of elements g( ) = e J . As we have seen that
angular velocities belong to the Lie algebra, it will be the group of rotations
with angular velocity J. Now,
g =
d J
e = Jg
d
is a tangent vector, and we see that J = gg

1 , that is, the angular velocity
J is obtained by righttranslation. Another angular velocity is obtained by
lefttranslation. The first is identified to the space angular velocity, and
the latter to the body angular velocity. Let us see how it happens.
A point in configuration space is a point of the group manifold. Let us
try to use both the differential and the group structure simultaneously. If
g S0(3), then L = g 1 dg is left-invariant (it is just the Maurer-Cartan
canonical form ) and R = dgg 1 is right invariant. Direct calculations
show that:
R = g L g 1 = Adg (L ) ;
(2.27)
L = g 1 R g = Adg1 (R ) ;
(2.28)
dL + L L = 0 ;
(2.29)
dR R R = 0 .
(2.30)
We may use the holonomic group parameter basis, writing g in terms of

i
the group generators {Ji }, that is g = e Ji , to obtain
R = Ji di and L = g 1 Ji gdi = (Adg1 Ji )di .
The left-invariant form is, as repeatedly said, the MaurerCartan canonical
form, which we write simply L = = Ji i . If we write Adg1 Ji = hi j Jj
for the adjoint representation, the MaurerCartan basis {i } will be related
to the parameter basis {di } by j = hi j di . The parameters i are angles,
i
= di ( dtd ). Notice that dtd =
and the usual angular rate of change is i = d
dt
i
i and that the time variation of the anholonomic form is given by

d
j
= hi j i = hi j Ri .
dt
In matrix notation, g = (gij ), and g 1 = (g ij ) implies R = dgg 1 , the
ijk k
relation with vector notation being ij
= ()

R = R . We check that
d
k
ijk
1
and then that = g ir (g )rj , as it should. The formula
= d dt
2.3. THE RIGID BODY
517
hints that we might use also the

notation (a suggestive convention, though
d
basically incorrect) = dt . Notice finally that, with that convention,
d
= i
= k Jk =
i
dt
d
dt
= .
Right and leftinvariance are related to the presence of two distinct

derivatives on the group. Given a function F (g), a field X can derive it from
the left,

d
L
sX
X F (g) =
F (e g)
= dL F (g)(X) ,
ds
s=0
and from the right,

d
sX
X F (g) =
F (g e )
= dR F (g)(X) .
ds
s=0
R
One sees that dR F (g)(X) = dL F (g)(Adg X). In particular, for the field X =
d
tangent to a curve,
dt
"
#
R
L

L
d
d
d
d
1
1
F (g) = g
g
g F (g) +
F (g) ,
F (g) = dg
dt
dt
dt
dt
or
dL F
dR F
=
R F .
dt
dt
For a vector component,

dR F i
dL F i
dL F i
j
=
ij
F
=
ijk kR F j ,
R
dt
dt
dt
or
dL V
dR V
=
+ R V .
dt
dt
Comparison with (2.21) shows that:

(i) R dtd is the usual space velocity;

(ii) L dtd is the usual body velocity;
R
(iii) ddt is the usual space derivative;
(iv)
dL
dt
is the usual body derivative.
(2.31)
518
2.3.10
PHYS. TOPIC 2.
MORE MECHANICS
The Poinsot construction
The tangent space is constituted by the angular velocities. The cotangent

space is the space of the angular momenta. Thus, the angular momentum
belongs to the coadjoint representation. The well known property by which
the angular momentum is related to the angular velocity through the inertia
operator reveals the metric. Indeed, the inertia operator shows up as a left
invariant metric, relating as usual the tangent and the cotangent spaces. And
just as above, given a covector at a point of the group, the angular momentum
in space is obtained by the rightaction, and the angular momentum in
body by leftaction. The metric appears in the kinetic energy, which is
given by T = 21 (Mb , b ). In the absence of external forces, the rigid body
motion is a geodesic on SO(3) with this metric.
Given the hamiltonian
P
H = 21 ij Iij i j = f () ,
the angular momentum is M = grad f . The inertia ellipsoid is given by f =
constant = E. Using the euclidean metric (, ) and Mi = Iij j , we may define
the inertia ellipsoid as the point set { such that (M, ) = 1}. This means
that we stay at an energy levelsurface H = 12 . To each point on the ellipsoid
will correspond an angular velocity. Draw as the position vector, and get
the tangent at the point. Then M will be the vector perpendicular to the
tangent from the origin taken at the ellipsoid center (Figure 2.1, leftside).
Suppose now that a metric gij is present, which relates fields and cofields.
The case gij = diag (1/a2 , 1/b2 ) is of evident interest, as f (v) = g(v, v), with
v the position vector (x, y). To the vector of components (xj ) will correspond
the covector of components (pk = gkj xj ) and
p(v) = pk v k = gij xi xj .
As we are also in an euclidean space, the euclidean metric mij = ij may
be used to help intuition. We may consider p and r as two euclidean vectors
of components (pk ) and (xk ). Comparison of the two metrics is made by
using g(v, v) = m(p, v). Our eyes are used to the metric m, and we shall
use it to measure angles and define (now, really metric) orthogonality. In the
rightside of Figure 2.1 we show the vector v giving a point on the ellipse
and the covector p, now assimilated to an euclidean vector. The vector p is
orthogonal to the curve at each point, or to its tangent at the point. It has
the direction given by the thin line and we draw it from the origin O at the
ellipse center. The curve equation is p(v) = g(v, v) = m(p, v) = C. As |p|
= m(p, p)1/2 = |df |, we can take
p = v cos
2.3. THE RIGID BODY
519
Figure 2.1: The Poinsot construction.
This construction to relate a form to a field in the presence of a non-trivial

metric, mainly in its 3dimensional version, is very much used in Physics.
For rigid bodies, the metric is the inertia tensor, the vector is the angular
velocity and its covector is the angular momentum. The ellipsoid is the inertia
ellipsoid and the whole construction goes under the name of Poinsot. An
analogous case, the Fresnel ellipsoid, turns up in crystal optics (see Phys.5.6).
Lovelock & Rund l975

Lanczos 1986
Goldstein 1980
Westenholz l978
Arnold 1976
520
PHYS. TOPIC 2.
MORE MECHANICS
Phys. Topic 3
STATISTICS AND ELASTICITY
A STATISTICAL MECHANICS
Introduction
General overview
B LATTICE MODELS
The Ising model
Spontaneous breakdown of symmetry
The Potts model
Cayley tree and Bethe lattice
The four-color problem
C ELASTICITY
Regularity and defects
Classical elasticity
Nematic systems
The Franck index
3.1
3.1.1
A Statistical Mechanics
Introduction
The objective of Statistical Mechanics is to describe the behaviour of macroscopic systems, composed by a large number of elements, assuming the knowledge of the underlying dynamics of the individual constituents. In the effort
to describe real systems, usually very involved objects, it is forced to resort
to simplified models. Some models are actually reference models, supposed
to give a first approximation to a whole class of systems and playing the role
of guiding standards. They are fundamental to test calculation methods and
as starting points for more realistic improvements. For low-density gases, for
instance, the main reference models are the ideal gases, classical and quantal,
and the hard-sphere gas. For solids, lattices with oscillators in the vertices
521
522
PHYS. TOPIC 3. STATISTICS AND ELASTICITY
are standard when the involved atoms or molecules have no structure. The
next step involves attributing some simple internal structure to the atoms.
It is of course very tempting to look at the lattice as a space of which the
cells are building blocks. And then consider the case of negligible spacing
between the atoms as a model for the continuous media.
As implied in section 2.2, the structure of a space is at least in part revealed by its building blocks. However, it is in general difficult to find out
which ones the necessary blocks are. We have there used irregular tetrahedra
to cover E3 . It would have been impossible to do it with regular tetrahedra.
Because historically this problem was at first studied in the 2-dimensional
case, we refer to it as the problem of the tilings (or pavings) of a space.
Thus, we say that we cannot pave E3 with regular tetrahedra, though we
can do it with irregular ones; and we cannot pave the plane E2 with regular
pentagons, though we can pave a sphere with them, and then project into E2 ,
as we do in order to endow the plane with a spherical metric (see Math.11).
This procedure is helpful in modeling some distortions in crystals, caused by
defects which, in the continuum limit, induce a curvature in the medium.
Indeed, once the lattices become very tightly packed, some continuity and
differentiability can be assumed. Vectors become vector fields, and tensors
alike. Whether or not the system is a continuum is a question of the physical scales involved. When we can only see the macroscopic features, we may
look at the limiting procedure as either an approximation (the continuum
approximation) or as a real description of the medium. Elasticity theory
treats the continuum case, but the lattice picture is too suggestive to be
discarded even in the continuum approximation.
We are thus led to examine continuum media, elastic or not. Introducing
defects into regular lattices can account for many properties of amorphous
media. The addition of defects to regular model crystals, for example, provide
good insights into the qualitative structure of glasses.
Modern theory of glasses sets up a bridge between lattice models and
Elasticity Theory. Adding defects changes the basic euclidean character of
regular lattices. It turns out that some at least of the amorphous aspects of
glasses can be seen as purely geometrical and that adding defects amounts
to attributing torsion and/or curvature to the medium.
3.1.2
General overview
Statistical Mechanics starts by supposing that each constituent follows the

known particle mechanics of Phys.1. Basically, this is the mechanicsinvolved
in its name, though the most interesting systems require Quantum Mechanics instead. To fix the ideas, we shall most of time consider the classical
3.1. A STATISTICAL MECHANICS
523
case. But we are none the wiser after assuming microscopic hamiltonian dynamics, which supposes the knowledge of the boundary values. It is clearly
impossible to have detailed information on the boundary conditions for all
the particles in any realistic situation, such as a normal gas with around
1023 particles. It is essential to take averages on these boundary conditions,
and that is where the statistical comes forth. Different assumptions, each
one related to a different physical situation, lead to different ways of taking
the average. This is the subject of the ensemble theory of Statistical Mechanics. The microcanonical ensemble is used when all the energy values
are equally probable; the canonical ensemble describes systems plunged
in a thermal bath, for which only the average energy of the system is conserved; the grand-canonical ensemble describes systems for which only the
average number of particles is preserved; and so on. From a more mathematical point of view, Statistical Mechanics is a privileged province of Measure
Theory. Each ensemble defines a measure on phase space, making of it a
probability space (Math.3).
The volume of a 2n-dimensional phase space M is given by the Lebesgue
measure dqdp dq 1 dq 2 . . . dq n dp1 dp2 . . . dpn . The measure (the volume) of a
domain D will be
R
m(D) = hn D dqdp.
There are, however, two converging aspects leading to the choice of another
measure. First, M is a non-compact space and m(M ) is infinite. Second, in
physical situations there are preferred regions on phase space. An n-particle
gas with total hamiltonian h(q, p), for example, will have a distribution proportional to the Boltzmann factor exp[h(q, p)/kT ]. In general, the adopted
measure on phase space includes a certain non-negative distribution function
F (q, p) 0 which, among other qualities, cuts down contributions from high
qs and ps. Such a measure will give to D the value
R
m(D) = D F (q, p)dqdp.
If we want to have a probability space, we have to normalize F (q, p) so that
m(M ) = 1. The canonical ensemble, for example, adopts the measure
F (q, p) = R
eh(q,p)/kT
.
eh(q,p)/kT dqdp
M
(3.1)
In Classical Statistical Mechanics, macroscopic quantities are described

by piecewise continuous functions of time and of the position in the physical
space. Thus, the energy density H(x, t) of an n-particle gas around the
point x = (x1 , x2 , x3 ) at instant t will be a functional of the microscopical
hamiltonian h(q, p; x, t):
524
H(x, t) =
R
M
F (q, p)h(q, p; x, t)dqdp.
This is quite general: given any microscopic mechanical quantity r(q, p; x, t),
its macroscopic correspondent R(x, t) will be given by the average
Z
R(x, t) =
F (q, p)r(q, p; x, t)dqdp.
(3.2)
M
The normalizing denominator in F is the partition function, from which

thermodynamical quantities can be calculated. All this means that the state
of the system, as far as Statistical Mechanics is concerned, is fixed by the
probability measure dm = m(dqdp) = F (q, p)dqdp, which is interpreted as
the probability for finding the system in a region of volume dqdp around the
point (q, p) of the phase space. The measure translates microscopic into
macroscopic quantities. Actually, F (q, p) is a dynamical quantity F (q, p, t)
satisfying the Liouville equation (Phys.1)
t F = {F, H},
and the time evolution of the whole system is fixed by the behaviour of
F (q, p, t). Systems in equilibrium are described by time-independent solutions, for which F is an integral of motion. Different conditions lead to various solutions for F , each one an ensemble. In the general case, the measure
defines then the evolution of each physical quantity through the weighted
averages
Z
R(x, t) =
F (q, p, t)r(q, p; x)dqdp.
(3.3)
Thus, the state of the system is given by F (q, p, t). In equilibrium, F is

constant in time and so is each R(x, t) the system has fixed values for all
observable quantities. In practice, time-averages instead are observed. That
the expectancy (average on phase space) is equivalent to the time average is
the content of the famous Boltzmanns ergodic theorem (Math.3.2).
Of course, when quantum effects are important, hn dqdp is only a first
approximation. We should actually sum over discrete values, as in the lattice
models of section 3.2 below. Everything lies on the density matrix,
=
eH/kT
.
treH/kT
(3.4)
The expectation of an observable represented by the operator A will be

< A > = tr A.
(3.5)
3.1. A STATISTICAL MECHANICS
525
The state of the system is now given by the density matrix, and the
partition function is the denominator in [3.4]. The canonical ensemble is
particularly convenient, as the fixed number of particles makes it easier to
define the hamiltonian H. Consider for a moment a gas with N particles
of the same species. All average values can be obtained from the partition
function. The semi-classical case is obtained as the limit of Plancks constant going to zero, but it is wise to preserve one quantum characteristic
incorporated in the Gibbs rule: the particles are indistinguishable. For the
state of the system, fixed by the expectation values of the observables, it is
irrelevant which individual particle is in this or that position in phase space.
This means that the partition function, in terms of which the average values
can be obtained, must be invariant under the action of the symmetric group
SN , which presides over the exchange of particles. The canonical partition
function QN (, V ) of a real non-relativistic gas of N particles contained in
a d-dimensional volume V at temperature T = 1/k is an invariant polynomial of SN , just the SN cycle indicator polynomial CN (Math.2). If is the
mean thermal wavelength and bj is the j-th cluster integral which takes into
account the interactions of j particles at a time,
QN (, V ) =
1
N!

CN

V
V
V
b1 d , 2b2 d , 3b3 d , . . .
(3.6)
On 2-dimensional manifolds the exchange of particles is not governed by

the symmetric group, but by the braid group, and the statistics changes
accordingly. Instead of the usual statistics, which leads to particles behaving
either as bosons or as fermions, the so-called braid statistics is at work, giving
to the particles a continuum of possible intermediate behaviours (Math.2.9).
In the quantum case, we have actually to consider a space (an algebra)
of operators, of which the density matrices are the most important. Basically, as long as N is finite, we have a finite von Neumann algebra. Once the
partition function, as well as its logarithm and its derivatives, are obtained,
thermodynamical quantities are arrived at by taking the thermodynamical
limit of large N and V , the volume of the system. The background algebra
will then be a more general, infinite-dimensional von Neumann algebra (see
Math.5).
Balescu 1975
Pathria 1972
526
3.2
3.2.1
B Lattice models
The Ising model
Most of the lattice models suppose a d-dimensional lattice with N vertices

(sites) and some spacing (the lattice parameter) between them. In each
site is placed a molecule, endowed with some internal discrete degree of
freedom, generically called spin. The lattice can be cubic, hexagonal, cubic
centered in the faces, etc. The spin at the sitek is described by a qdimensional vector k (Figure 3.1, right). The interaction takes place along
the edges (the bonds) and is given by a general (Stanley) hamiltonian
of the form
X
X
H=
Jij i j H
k .
(3.7)
i<j
Figure 3.1:
The factor Jij represents the coupling between the molecules situated in the
i-th and the j-th sites. To solve such a model means to obtain the explicit
form of the partition function. A reasonably realistic case is the Heisenberg
model, for which d = q = 3. Though no analytic solution has been obtained,
many results can be arrived at by numerical methods. Some other cases have
been solved, none of them realistic enough. The problem is less difficult in
the nearest-neighbour approximation, which supposes that Jij 6= 0 only when
i and j are immediate neighbours. For 1-dimensional systems with no
external magnetic field (H = 0), exact solutions are known for all values of
the spin dimension q. For higher dimensions, the best known model is the
3.2. B LATTICE MODELS
527
celebrated Lenz-Ising model, for which q = 1 (sk = k = 1) and the same

interaction is assumed for each pair of neighbours:
X
X
H = J
si sj H
sk .
(3.8)
<ij>
The symbol < ij > recalls that the summation takes place only on nearest
neighbours. The partition function is
X
X
X
QN (, H) =
exp{K
si sj + h
sk },
(3.9)
sk =1
<ij>
where K = J and h = H. The 1-dimensional case was solved by Ising in

1925. For H = 0, it is
QN () = [2 cosh K]N = [2 cosh(J)]N .
(3.10)
To illustrate the general method of solution and to introduce the important concept of transfer matrix, let us see how to arrive to this result. The
model consists of a simple line of spins 1/2, disposed in N sites. The left
segment of Figure 3.2 shows it twice, once to exhibit the site numeration and
the other to give an example of possible spin configuration. Identification of
the sites 1 and N + 1 (N +1 = 1 ) corresponds to a periodic boundary
condition. We have in this case an Ising chain (Figure 3.2, right segment),
which becomes a torus in the 2-dimensional case.
Figure 3.2:
The total energy will be

E() = J
N
X
i=1
i i+1 H
N
X
i=1
i ,
(3.11)
528
and the partition function,

QM (H, T ) =
eE() =
eK
i i+1 +h
(3.12)
The symbol represents a configuration of spins, = (1 , 2 , 3 , . . . , N )

and the summation is over the possible values of .
There are many methods to solve the problem. We shall here introduce,
as announced, the symmetric transfer matrix V , a symmetric matrix with
elements
0
0
V0 = eK +(h/2)(+ ) .
(3.13)
Instead of the usual labeling, we use as labels the spin values {+,-}. The
matrix elements will be V+ + = eK+h ; V+ = eK ; V + = eK ; V = eKh ,
so that

K+h
eK
e
(3.14)
V =
eK eKh
One finds then from [3.12] that the partition function can be written as
X
QN (H, T ) =
V1 2 V2 3 V3 4 . . . VN 1 = tr V N .
(3.15)
The partition function is the trace of the N -power of the transfer matrix. In
other cases, more than one transfer matrix can be necessary, as in the Potts
model (see Phys.3.2.3 below), in which each transfer matrix is related to a
generator of the braid group. Here, V alone suffices. It can be diagonalized
with two eingenvalues 1 and 2 . Then, [3.15] will say that
N
QN (H, T ) = N
1 + 2 .
The eingenvalues are solutions of the secular equation

K+h

K
e

e

= 2 + (2eK cosh h) + 2 sinh(2K) = 0.
Kh
eK
e

One finds
= eK cosh h
(3.16)
(3.17)
e2 K cosh2 h 2 sinh(2 K) ,
so that

QN (H, T ) =
N
q
2
2
K
cosh h + e cosh h 2 sinh(2 K)

+ e
cosh h
e2 K
N
cosh h 2 sinh(2 K) . (3.18)
2
529
The first cases are

Q1 (H, T ) = 2eK cosh h , Q2 (H, T ) = 2e2K cosh 2h + 2e2K .
When h = 0,
QN (0, T ) = [eK + eK ]N + [eK eK ]N = [2 cosh K]N + [2 sinh K]N . (3.19)
As cosh x > sinh x, the first term will dominate for large values of N , and
the result [3.10] comes forth.
For the 2-dimensional case, the solution1 for a square lattice and hamiltonian H = 0 has been found by Onsager in 1944:
Z
1
d ln[1 + (1 b2 sin2 )1/2 ]1/2 }]N , (3.20)
QN () = [2 cosh(2J) exp {
2 0
where b = [2 sinh(2J)/ cosh2 (2J)]. The procedure to find it was extremely
difficult. A simpler derivation, due to Vdovichenko,2 has been known since
the sixties. No analytic solution has been found as yet for the d = 3 case.
Comment 3.2.1 If we go to the continuum limit, with the lattice parameter going to
zero, the spins constitute a spin field. We can also go to the classical limit, so that spin
is a variable taking continuum values at each point: it becomes a field whose character
depends on the range of values it is allowed to assume.
3.2.2
Spontaneous breakdown of symmetry
The main interest in these models lies in the study of phase transitions. The
1-dimensional solution [3.10] shows no transition at all, which is an example of
a general result, known as the van Hove theorem: 1-dimensional systems with
short-range interactions between constituents exhibit no phase transition.
The 2-dimensional Onsager solution, however, shows a beautiful transition,
signaled by the behaviour of the specific heat C, whose derivative exhibits
a singularity near the critical (Curie) temperature kTc 2, 269J. The
specific heat itself behaves, near this temperature, as C constant ln(|T
Tc |1 ). A logarithmic singularity is considered to be a weak singularity. The
magnetization M , however, has a more abrupt behaviour, of the form
M
1.2224
N
1
2
See Huang 1987.

Landau & Lifshitz l969.
Tc T
Tc

, with = 1/8.
(3.21)
530
Numerical studies show that in the 3-dimensional case the phase transition
is more accentuated. For a cubic lattice the specific heat near the critical
temperature behaves as
C |T Tc | , with 0, 125.
(3.22)
Thus, the general behaviour depends strongly on the dimension. This

transition may be thought of as a ferromagnetic transition, with an abrupt
change from a state in which the magnets are randomly oriented, at high
temperatures, to a microscopically anisotropic state in lower temperatures,
in which the magnets are aligned. It is also an order-disorder transition.
The magnetization of some real metals is qualitatively described by the Ising
model. The fractional values of the exponents in the behaviours [3.21] and
[3.22] tell us that these critical points are not singular points of the simple
Morse type, which would have a polynomial aspect (Math.9.6). They point
to degenerate points, and are actually obtained via the far more sophisticated
procedures of the renormalization group.
P
With an energy of the form E = J <ij> si sj , there are two configurations with the (same) minimum energy: all the spins up (+1) and all spins
down (-1). Thus, at temperature zero, there are two possible states, and the
minimal entropy is not zero but S0 = k ln 2. When the temperature is high,
the system is in complete microscopic disorder, with its spins pointing along
all directions. Even small domains (large enough if compared to molecular
dimensions) of the medium exhibit this isotropy, or rotational symmetry. As
the temperature goes down, there is a critical value at which the system
chooses one of the two possible orientations and becomes spontaneously
magnetized while proceeding towards the chosen fundamental state. The
original macroscopic rotational symmetry of the system breaks down. Notice that the hamiltonian is, and remains, rotationally symmetric. The word
spontaneous acquired for this reason a more general meaning. We call
nowadays spontaneous breakdown of symmetry every symmetry breaking
which is due to the existence of more than one ground state. The fundamental state is called vacuum in field theory. When it is multiple, we
say that the vacuum is degenerate. There is thus spontaneous breakdown
of symmetry whenever the vacuum is degenerate. A quantity like the above
magnetization, which vanishes above the critical temperature and is different from zero below it, is an order parameter. The presence of an order
parameter is typical of phase transitions of the second kind, more commonly
called critical phenomena.
3.2.3
531
The Potts model
The Potts model3 may be defined on any graph (see section 2.1), that is, any
set of vertices (sites) with only (at most) one edge between each pair. This
set of sites and edges constitutes the basic lattice, which in principle models
some crystalline structure. A variable si , taking on N values, is defined on
each site labelled i. For simplicity of language, we call this variable spin.
Dynamics is introduced by supposing that only adjacent spins interact, and
that with interaction energy eij = J si sj , where is a Kronecker delta. The
total energy will be E = J (ij) si sj , the summation being on all the edges
(i, j). Then, with K = J/kT, the partition function for an M -site lattice will
be
QM = s exp[(ij) si sj ] ,
the summation being over all the possible configurations s = (s1 , s2 , . . . , sM ).

The Ising model, with cyclic boundary conditions, is the particular case with
N = 2 and K replaced by 2K. Despite the great generality of this definition
on generic graphs, we shall only talk of lattices formed with squares. The
main point for what follows is that QM may be obtained as the sum of all
the entries of a certain transfer matrix T analogous to [3.14]. This matrix
T turns out to be factorized into the product of simpler matrices, the local
transfer matrices, which are intimately related to the projectors Ei of the
Temperley-Lieb algebra (Math.5.6).
A surprising outcome is that the partition function for the Potts model
can be obtained as a Jones polynomial for a knot related to the lattice in a
simple way. Given a square lattice as that of Figure 3.1, with the interactions
just defined, consider the N n N n matrices E1 , E2 , . . . , E2n1 , with
(E2i1 )
s,s0
(E2i )s,s0 =
1
N
n
Y
sj s0j ,
j6=i=1
n
Y
N si si+1
sj s0j .
j=1
We follow here the Bible of lattice models: Baxter 1982.
(3.23)
(3.24)
532
Let us give some examples, with the notation |s > = |s1 , s2 , s3 , . . . , sn > :
< s|E2 |s0 > = s1 s2 (s1 s01 s2 s02 . . . sn s0n );
< s|E4 |s0 > = s2 s3 (s1 s01 s2 s02 . . . sn s0n );
...
0
< s|E2n2 |s > = sn1 sn (s1 s01 s2 s02 . . . sn s0n );

< s|E1 |s0 > =
< s|E3 |s0 > =
1
N
s1 s01 N1
s2 s02 . . . sn s0n ;
s3 s03 s4 s04 . . . sn s0n ;
< s|E5 |s > = s1 s01 s2 s02 N1 s4 s04 . . . sn s0n ;

...
< s|E2n1 |s > = s1 s01 s2 s02 . . . sn1 s0n1 N1 .
0
Thus, in the direct product notation (Math.2.10), if E is the identity N N

matrix, the even-indexed matrices are
E2i =
N si si+1 E E E E E . . . E
N si si+1 (E n ). (3.25)
Matrix E2i is, thus, a diagonal matrix, with entries N si si+1 . The oddindexed matrices are
h
i
1
E2i1 = E E E E E . . . N . . . E E
i
h
= E (i1) 1N E (ni) , (3.26)
where
1
N
, which is in the i-th position, is a N N matrix (also a projector)
with all the entries equal to 1N . The notation is purposeful: such Ek s

satisfy just the defining relations of the Temperley-Lieb algebra (Math.5.6)
with M = 2n 1, and Jones index = N . By the way, we see that the Jones
index is in this case just the dimension of the spin space.
We introduce the local transfer matrices
Vj = I +
v
N
E2j , Wj =
v
N
I + E2j1 ,
(3.27)
with I the identity matrix and v = eK 1. We can also introduce the

Kauffman decomposition (see Math.2.16)
/
\ = )( +
v
N
(3.28)
533
which means that the inverse is

\
v
N
/\ =
)( +
.
(3.29)
The bubble normalization is

=
(3.30)
There will be two global transfer matrices, which can be put into the
forms
V = exp{K(E2 + E4 + . . . + E2n2 )} = exp{K
n1
X

sj sj+1 } E n =
j=1
n1
Y
j=1
n1
Y 2j 2j+1
v
)(
+
I + E2j =
N
j=1
2j
v
N
2j+1

=
n1
Y 2j
2j1
2j
(3.31)
j=1
and
W =
n h
Y
vI +
N E2j1 = N
n/2
j=1
n
Y
2j1
v
N
2j
)( +
j=1
=N
n/2
n 2j1
Y
\
/\
(3.32)
j=1
We now look at these transfer matrices in terms of the (2n1) generators

of the braid group B2n . They are
1
V = 2 4 . . . 2n2 and W = 11 31 . . . 2n1
.
In the case of a n m Potts lattice, the partition function is

Qnm = T V W V W . . . V = T T ,
where is a column vector whose all entries are equal to 1. There are mV s
and (m 1)W s in the product. To sandwich the matrix T between T and
is a simple trick: it means that we sum all the entries of T .
Let us now try to translate all this into the diagrammatic language. The
sum over all configurations is already accounted for in the matrix product, as
the index values span all the possible spin values. The question which remains
is: how to put into the matrix-diagrammatic language the summation over
the entries of the overall transfer matrix? The solution comes from the use of
the projectors. In order to see it, let us take for instance the case n = m =2.
In this case the diagrams have 4 strands,
534
V = (I +
v E2 )
N
, W = (vI +
N E1 )(vI +
N E3 ).
The matrix involved will be

and is an element of B4 : V W V = 2 11 31 2 . We have to sum over all the
values of the indices a, b, c, d, e, and f in Figure 3.3.
The desired result is obtained by adding projectors before and after the
diagram as in Figure 3.4, and then taking the trace, that is, closing the
final diagram. This closure is represented by taking identical labels for the
corresponding extreme points which is just closure in the sense of knot
Figure 3.3:
theory. Of course, there will be two extra factors from the bubbles, which
must be extracted. This solution is general: for M = 2n vertices, we add 2n
projectors and then close the result, obtaining n extra bubbles which must
then be extracted. Thus, the partition function is
Q = N n/2 < K > .[4.11]
(3.33)
The general relationship is thus the following: given a lattice, draw its
medium alternate link K, which weaves itself around the vertices going
alternatively up and down the edges. Figure 3.5 shows the case m = n =2,
which corresponds simply to T closed by pairs of cups. To each edge of the
lattice will correspond a crossing, a generator of Bn (or its inverse). Vertices
will correspond to regions circumvented by loops. With the convenient choice
535
Figure 3.4:
Figure 3.5:
of variables given above, the partition function is the Jones polynomial of

the link.
In these lattice models, the lattice itself has been taken as fixed and
regular, dynamics being concentrated in the interactions between the spin
variables in the vertices. In the study of elastic media and glasses, this regularity is weakened and the variables at the vertices acquire different values
and meanings.
3.2.4
Cayley trees and Bethe lattices
Suppose we build up a graph in the following way: take a point p0 as an

original vertex and draw q edges starting from it. To each new extremity
add again (q 1) edges. Thus q is the coordination number, or degree of
536
each vertex in the terminology of section 2.1. The first q vertices constitute
the first shell, the added q(q 1) ones form the second shell. Proceed
iteratively in this way, adding (q 1) edges to each point of the r-th shell to
obtain the (r + 1)-th shell. There are q(q 1)r1 vertices in the r-th shell.
Suppose we stop in the n-th shell. The result is a tree with
V = q[(q 1)n 1]/(q 2).
This graph is called a Cayley tree. It is used in Statistical Mechanics, each
vertex being taken as a particle endowed with spin. The partition function
will be the sum over all possible spin configurations. There is a problem,
though. The number of vertices in the n-th shell is not negligible with respect
to V , so that one of the usual assumptions of the thermodynamical limit
that border effects are negligible is jeopardized. One solution is to take
n , consider averages over large regions not including last-shell vertices,
and take them as representative of the whole system. The tree so obtained
is called the Bethe lattice and the model is the Ising model on the Bethe
lattice. Its interest is twofold: (i) it is exactly solvable and (ii) it is a first
approximation to models with more realistic lattices (square, cubic, etc).
3.2.5
The four-color problem
The intuitive notion of a map on the plane may be given a precise definition
in the following way. A map M is a connected planar graph G and an
embedding (a drawing) of G in the plane E2 . The map divides the plane
into components, the regions or countries. G is the underlying graph of M ,
each edge corresponding to a piece of the boundary between two countries.
Actually, one same graph corresponds to different maps, it can be drawn in
different ways.
However, there is another graph related to a given map M : place a vertex
in each country of M and join vertices in such a way that to each common
border correspond an edge (as we have done in drawing the graph for the
Konigsberg bridges in 2.1.7). This graph is the dual graph of M , denoted
D(M ). When we talk of coloring a map we always suppose that no two
regions with a common border have the same color. The celebrated four-color
conjecture says that 4 colors are sufficient. That 4 colors are the minimum
necessary number is easily seen from some counter-examples. That they are
also sufficient is believed to have been demonstrated in the 70s by Appel
and Haken. The proof, involving very lengthy computer checkings and
some heuristic considerations, originated a warm debate.4 What matters
4
For a thorough account, see Saaty & Kainen 1986.
3.3. C ELASTICITY
537
here is that the question is a problem in graph theory and related to lattice
models. In effect, the problem is equivalent to that of coloring the vertices
of the dual graph with different colors whenever the vertices are joined by
a common edge. Or, if we like, to consider the dual graph as a lattice,
and colors as values of a spin variable, with the proviso that neighbouring
spins be different. Given a graph G, the number P (G, t) of colorings of G
using t or fewer colors may be extended to any value of the variable t. It
is then called a chromatic polynomial . It comes not as a great surprise that
such polynomials are related to partition functions of some lattice models in
Statistical Mechanics.
Pathria 1972
Baxter 1982
3.3
3.3.1
C Elasticity
Regularity and defects
Despite its position as a historical source of geometrical terminology, the

language of Elasticity Theory takes nowadays some liberties with respect to
current geometrical jargon. There are differences concerning basic words, as
happens already with torsion, taken in a more prosaic sense, and also some
shifting in the nomenclature, even inside the Elasticity community. Texts on
elasticity keep much of lattice language in the continuum limit and make use
of rather special names for geometrical notions. Thus, local system of lattice vectors is used for Cartan moving frames and lattice correspondence
functions for moving frame components. Distant parallelism is the eloquent expression for asymptotic flatness. And differentials are frequently
supposed to be integrable.
We try here to present a simple though general formulation, the simplest
we have found seemingly able to accommodate coherently the main concepts.
The formalism is in principle applicable to crystals, elastic bodies and glasses
in the continuum limit. When talking about crystals, we think naturally
of some order or periodicity at the microscopic level. Amorphous media like
glasses, however, can be considered in the same approach, provided some
defects are added to the previous crystalline regularity. We start thus from
the usual supposition about microscopic regularity in crystals and deform
the medium to obtain a description of amorphous solids. The continuum
approximation to an elastic body is taken as the limit of infinitesimal lattice
538
parameter. We shall consequently use the word elasticity in a very broad

sense, so as to include general continuum limits of regular and irregular
crystals, with preference for the latter. Some at least of the amorphous
aspects of glasses can be seen as purely geometrical, and adding defects
amounts to attributing torsion and/or curvature to the medium, which makes
of such systems physical gateways into these geometrical concepts.
Comment 3.3.1 Regularity means symmetry, usually under translations and/or rotations. Take the simple example of a 2-dimensional square lattice (Figure 3.1). Translations
and rotations are discrete: the regular crystal is invariant under discrete translations of
a multiple of the lattice parameter, and rotations of angles which are integer multiples of
/2. These rotations constitute the so-called rotation group of order 4: the only generator
is exp[i/2], so that it is a cyclic group (Math.2.3).
The sources of deformations may be external or internal. The first case

is the main subject of the classical texts on elasticity. Forces are applied
on the system through their surface. The main objective then is to find the
relation between the applied stress and the internal strain, which for small,
reversible deformations, is given by Hookes law. The interaction between
the atoms (or molecules) at the vertices should be represented by realistic
potential wells, but a simple view is given by their first approximation. The
first approximation to any reasonable potential well is a harmonic oscillator,
so that a rough qualitative model is obtained by replacing the bonds by
springs.5
Internal deformations are the principal concern of the theory of glasses
and amorphous media. They arise from defects, and defects are of many
kinds, but there are two main types of internal deformations: dislocations
and disclinations. There are some fluctuations in the very definitions of these
concepts. Some authors define a dislocation simply as any linear defect, and
a disclination as a defect leading to non-integrability of vector fields. For
other, dislocations are failures of microscopic translational invariance and, in
the same token, disclinations are related to failures of microscopic rotational
invariance. In this line of thought, (geometric) torsion is then related to
dislocations, and curvature to disclinations. In our inevitable geometrical
bias, we shall rather adopt this point of view.
Comment 3.3.2 Such notions are not always equivalent. You can, for example, distort a space to become S 3 , which is curved and has rotation invariance.
Comment 3.3.3 Beauty is sometimes related to slightly broken symmetry. And some
masterpieces are what they are because they are slightly uncomfortable to the eye. Some
of Eschers woodprints, such as the celebrated Waterfall, are good illustrations of torsion
5
For an illuminating modern discussion, see Askar 1985.
3.3. C ELASTICITY
539
as engendered by a line of dislocations. Euclidean perspective goes wrong in such a space

and our euclidean eyes are at a loss.
In order to help ideas sinking in, let us see in a simple 2-dimensional

example how such deformations can bring about curvature. We shall talk of
an imaginary 2-dimensional semi-conductor. It is possible to pave the plane
with regular hexagons because the internal angle at each hexagon vertex is
2/3. Regular pentagons would not do it because the angles at the vertices
are not of the form 2/N , with N an integer. If an edge collapses so that one
of the original hexagons becomes a pentagon, an angle defect would come out
(Figures 3.6,3.7). Nevertheless, it is possible to tile a sphere with pentagons:
a possible polyhedron (2.2.5) of S 2 is the pentagon-faced dodecahedron
Figure 3.6:
(Figure 3.8). On the other hand, we can pave a hyperboloid with heptagons,
which signals to positive curvature when we remove an edge, and negative
curvature if we add one. The crystal would be globally transformed if all
the edges were changed, but the presence of a localized defect would only
change the curvature locally. The local curvature will thus be either positive
or negative around a defect of this kind. It is a general rule that, when a
localized defect is inserted in a lattice, it deforms the region around but its
effect dies out progressively with distance. Figure 8 shows a typical case of
(two-dimensional) dislocation through the insertion of a limited extra line.
Experiments with a paper sheet can be of help here.
Well, in real media we have to do with atoms placed on the vertices. It
happens in some amorphous semi-conductors6 that, due to the presence of
6
Harris 1975; Kleman & Sadoc 1979.
540
Figure 3.7:
Figure 3.8:
impurities (sometimes simply hydrogen in the realistic 3-dimensional cases),

some of the edges do collapse, so that two adjoining hexagons become pentagons with only a common vertex. In the rough model with springs, some
of then acquire a large spring constant and the oscillators become very steep.
We might think that this would lead to a situation in which some of the
cells would be irregular, with sides of different lengths. Nevertheless, at the
microscopic level, the edges distances between the atoms are fixed by
the inter-atomic potentials. In principle, they correspond to minimal values
of the energy in these potentials. It happens in some cases that the distances
are kept the same. In the more realistic 3-dimensional case, the suggestion
to drop or to add a wedge comes from the experimental evidence of the presence of rings with one-less or one-more atoms in an otherwise regular lattice
3.3. C ELASTICITY
541
Figure 3.9:
in amorphous semi-conductors.7 Thus, the inter-atomic potentials require

regular polyhedra to tile the system, which is consequently deformed into
some spherical geometry. Some other physical systems require instead that
an extra edge is added so as to form heptagons and leading to a hyperbolic
geometry.
The original flat space is thus curved by the presence of impurities.
Remember that, at 2 dimensions, the sign of the curvature 1/rr at a given
point is very easy to see. Trace two tangent circles perpendicular to each
other at the point. Their radii r and r are the so called curvature radii. If
both have the same sign, the curvature is positive. If they have opposite
signs, negative. If you prefer, in the first case there is an osculating sphere
at the point, and in the second, a hyperboloid. Figure 9 shows the case
of square-to-triangle collapse. At the right, the vector field represented by
the crosses comes back to itself after a trip around a loop circumventing no
defect. At the left, the vector field is taken along a loop around the defect,
and is rotated of an angle at the end of the trip. This vindicates the view
of non-integrable vector fields, which anyhow lies behind the very notion of
curvature. Recall that, taken along an infinitesimal geodesic loop, a vector
k
field V is changed by V k = Rrij
V r dxi xj (9.4.13).
7
Sadoc & Mosseri 1982.
542
Figure 3.10:
3.3.2
Classical elasticity
There are two main properties characterizing the euclidean crystal we

started from. First, we can measure the distance between two neighbouring points using the euclidean metric,
dl = [ab dxa dxb ]1/2 .
(3.34)
Second, we can say whether or not two vectors at distinct points are parallel.
We do it by transporting one of them along the lines defining the lattice into
the others position, in such a way that the angles it forms with the lines are
kept the same. If their direction and sense coincide when they are superposed,
they are parallel. This corresponds to parallel-transporting according to the
trivial euclidean connection (the Levi-Civita connection for the metric ab ),
whose covariant derivatives coincide with usual derivatives in a Cartesian
natural basis, and for which the lattice lines are the geodesics. The lattice
itself play the role of a geodesic grid. We can place at each vertex a set of
lattice vectors {ea }, oriented along the lines and parallel-transported all
over the lattice. Given the set at one vertex, we know it at any other vertex.
The euclidean metric will fix ab = (ea , eb ) for this initial dreibein.
The first clear visible signal of a deformation is that the measure of distance between the points changes. In the general case, the change is different
in different regions. Two neighbouring points initially separated by an eu-
3.3. C ELASTICITY
543
Figure 3.11:
clidean distance [3.34] will, after the deformation, be at a different distance,

though we keep measuring it with an euclidean rule. We represent this by
dl0 = [gij (x)dxi dxj ]1/2 ,
(3.35)
as if the distance were given by some other, point-dependent metric gij . The
{dxi } are the same as the previous {dxa }: we are simply concentrating the
deformation in the metric. The euclidean and the new metric tensors are
related by some point-dependent transformation ha i (x),
gij = ab ha i hb j = ha i hb j (ea , eb ) = (ei , ej ).
(3.36)
Each ei = ha i ea is a member of the new dreibein. The new metric is

given by the relative components of these ei s, measured in the old euclidean
metric. This is to say that the initial covector basis {dxa } is related to another
covector basis { i }, dual to {ei }, by dxa = ha i i . Of course, gij i j = dl2 .
We have, so, just a 3-dimensional example of repere mobile (7.3.12), with
all its proper relationships and a metric defined by it (see also 9.3.6 and
Math.10.1.1). As the deformations are supposed to be contiguous to the
identity, the hai s are always of the form
ha i = ia + ba i ,
(3.37)
544
for some fields ba i , which represent the departure from the trivial dreibein
related to the unstrained state. The new metric has the form gij = ij + 2uij ,
where
uij = 21 (bij + bji + ab ba i bb j )
(3.38)
is the strain tensor .
If the new basis is holonomic, ba i = i ua for some field ua (x), the deformation field . In this case the field of deformations is the variation in the
coordinates, given by x0k = xk + uk (x). This gives
dx0k = dxk + duk (x) = dxk + j uk (x)dxj ,
so that the length element changes by
dl2 dl02 = dxk dxk +j uk (x)dxk dxj +j uk (x)dxj dxk +i uk (x)j uk (x)dxi dxj .
The derivative wij = i uj is the distortion tensor . To first order in u and
its derivatives, dl02 = [jk + 2ujk ]dxj dxk , where
ujk = w(jk) = 21 [j uk + k uj ]
(3.39)
is the strain tensor for this holonomic case. Notice from gij = ij + 2uij =
ab ha i hb j that the Cartan frames (dreibeine) are
ha i = ia + wa i .
(3.40)
The fields bai of [3.37], consequently, generalize the distortion tensor to the
anholonomic case. The holonomic case comes up when they are the derivatives of some deformation field. Some authors define defects as the loci of
singular points, and dislocations as lines of singularity of the deformation
field. This only has a meaning, of course, when this field exists.
Summarizing, a deformation creates a new metric and new dreibeine. It
changes consequently also the connection. We might think at first the new
connection to be the Levi-Civita connection of the new metric, but here
comes a novelty. The connection is, in principle, a metric-independent object and can acquire proper characteristics. For example, it can develop a
non-vanishing torsion. Impurities, besides changing the distances between
the atoms, can also disrupt some of the bonds, in such a way that the original (say) hexagon is no more closed. They may become open rings in the
plane, but they may also acquire a helicoidal aspect in 3 dimensional media.
The euclidean geodesic grid collapses. These deformations are called dislocations and are of different kinds. Figures 3.12, 3.13 show what happens to
a loop in the case of the dislocation of Figures 3.9, 3.10. If we keep using the
original geodesic grid, we find a breach: the grid is destroyed and there are
3.3. C ELASTICITY
545
Figure 3.12:
no more infinitesimal geodesic parallelograms (9.4.14). The new dreibeine

fail to be parallel-transported. The torsion T mesures precisely this failure,
because there is a theorem (the Ricci theorem, see Phys.8) which says that,
given g as above and a torsion T , the connection is unique.
The presence of (geometrical!) torsion in amorphous media is confirmed
by experimental measurements, and is, in physical terminology, related to two
physical quantities: the Nye index and the Burgers vector.8 These quantities
measure the cleavage and are related to torsion as follows. Take a loop in the
undeformed crystal as in Figure 3.13 (it is called a Burgers circuit in elasticity
jargon). Once the crystal is deformed, it fails to close into a loop. The vector
from the starting point to the final point of a curve, which would be a loop
in the undeformed crystal, is called the Burgers vector . The situation is
simpler when a deformation field does exist. Consider in this case a closed
line before the deformation, and a point p on it, which we shall take as
its coincident initial and final endpoints. After the deformation, when the
deformed 0 is no more closed, p goes to two distinct points, p0 and p00 . The
Burgers vector is then defined as
k
B :=
du =
0
Burgers 1940.
uk i
dx =
xi
Z
0
wi k dxi = uk (p00 ) uk (p0 ).
(3.41)
546
Figure 3.13:
It is clearly a measure of the disruption. The distortion tensor appears as a

linear density for the Burgers vector. The Nye index k i is introduced by
Z
k
k i dxi ,
(3.42)
B =2
in general, eventually also in the non-integrable case. It is a line integral of the

form bk i dxi . In the continuum case, the classical notion of torsion is related
precisely to this disruption of infinitesimal geodesic parallelograms. The
failure xk , not necessarily integrable, of such closure is precisely measured
by
xk = T k ij dxi dxj .
(3.43)
Recall that on a 3-dimensional euclidean space, dxi dxj = ij k dxk .
Consequently, xk = T k ij ij s dxs = 2 k s dxs and
rk = 12 rij Tijk .
(3.44)
Thus, the Nye index is precisely the dual to the torsion field. The torsion is
T a ij = i ha j j ha i + a bi hb j a bj hb i ,
so that, by a change of basis, T k ij = ha k T a ij = k ji k ij . Thus,
rk = rij k [ij] ,
(3.45)
3.3. C ELASTICITY
547
where k [ij] is the antisymmetric part of k ij .

What happens to the connection in the deformed case? As we also want
to keep vector moduli and angles with a coherent meaning, we should ask
that the connection preserves the metric,
dgij = (ijr + jir )dxr .
(3.46)
The metric can only fix a piece of the symmetric part of ijr in the two last
indices (see Phys.8.1). To determine ijr completely we need to know the
torsion through experimental measurements of the Nye index.
3.3.3
Nematic systems
In the above discussion of torsion and curvature, only positional degrees of

freedom were supposed for the constituent molecules. They have been treated
as punctual, only the position of their centers of gravity have been considered.
Deformations were supposed to introduce new metrics and connections on
the lattice-manifold. The situation is consequently related only to the linear
frame bundle (section 9.4). We can imagine that adding internal degrees,
like in the spin lattice models of section 3.2 above, gives a situation more
akin to that of gauge fields, with internal spaces as fibers9 , and in which
general principal bundles (sections 9.5 and 9.6) are at work. Molecules have in
general internal degrees, the simplest of which is, for non-spherical molecules,
orientation.
In a solid crystal there is perfect positional order: it represents the case
in which the centers of gravity are perfectly established at the fixed sites.
Classical elasticity theory studies precisely small departures from this regular
case.
Melting takes the crystal into a state of positional disorder, a liquid.
There are systems in which the solid-liquid transition is not so simple, but
takes place in a series of steps in which order is progressively lost. And in
some situations a system can be stable in some intermediate state. It is in
this case a liquid crystal. For instance, the system can lose the ordering in
2 dimensions while retaining a periodic order in the third. Such a phase is
called smectic. There are phase transitions related to change of orientation
in liquids, solids and smectic media.
The quantum Ising model considers in each site a two-valued spin. One
might imagine cases more classical, in which spin takes on values in a
continuous range. In the nematic crystals we consider, instead of spin, an
internal variable describing the orientation of the molecules. This case
9
On such spin glasses and gauge field theories, see Toulouse & Vannimenus 1980.
548
is more involved, as different orientations between neighbouring molecules

imply distinct couplings between them. And also, a direction is less than
a vector, because two opposite vectors correspond to the same direction.
Some organic systems, as well as solid hydrogen, show a high-temperature
phase which is positionally ordered but orientationally disordered (plastic
crystals).
Certain organic liquids have a low-temperature phase which is positionally disordered but orientationally ordered, with the molecules oriented along
a preferential direction. Such systems usually have a parity symmetry: sufficiently large subdomains do not change if all the three axes are reversed.
This is a consequence of the requirement that the molecules orientation be
the only origin of anisotropy. Parity invariant systems (crystals, liquids, or
liquid crystals) with an orientation degree of freedom are generically called
nematic systems. In most cases the molecules are of ellipsoidal shape, so
that only the direction, not the sense of the molecule orientation, is of import (Figure 3.14). The state of a molecule can be characterized by (say) a
versor n along its major axis (called the director ). Thus, the director can be
seen, in the continuum limit, as a field. As states n and - n coincide (Figure
3.15), the director field has values in the half-sphere S 2 /Z2 = P R2 . It is
not a vector field, it is a direction field. The average value of the direction
field has the role of an order parameter, which vanishes above the critical
temperature and becomes significant below it.
Figure 3.14:
3.3. C ELASTICITY
549
Figure 3.15:
Comment 3.3.4 Systems with the same general characteristics, but without parity
invariance, are thermodynamically unstable they decay into stable systems of another
kind, cholesterics.
Consider a finite nematic system. By that we mean that the system covers a compact domain V in E3 with boundary V . The distribution of the
direction field will be fixed up when the values of n are given on V . Two
standard examples10 come to the scene when we consider an infinite cylindrical system with axis (say) along the axis Oz. In cylindrical coordinates
(, z, ), the field n(r) will not depend on z because it is infinite in that
direction, and will not depend on because there is not in the system any
parameter with the dimension of length, in terms of which we could write a
non-dimensional variable like n(r). Thus, we have only to consider a plane
transverse section of the cylinder and the only significant variable is the angle
10
Landau & Lifchitz 1990; the last editions have chapters concerning dislocations and
nematic systems, written respectively in collaboration with A.M. Kossevitch and L.P.
Pitayevski.
550
: n(r) = n(). The two cases correspond to two distinct kinds of boundary
conditions: the values of n at the boundary, n|V , are either orthogonal to
V or parallel to V . Continuity will then fix the field all over the section.
It is clear then, by symmetry, that n(r) is in both cases ill-defined on the Oz
axis, which is supposed perpendicular to the plane at the origin in Figure
3.16. This line of singularity Oz is a disclination line.
Figure 3.16:
3.3.4
The Franck index
3.1 A most beautiful physical example of topological number is the Franck

index.11 It would be a pure case of winding number were not for the fact that
not a true vector field is involved, but a direction field. As a consequence,
the resulting number can take half-integer values. Let us proceed now to
the general winding number definition. We shall traverse a closed path in
the system and look at each point at the value of n(r) in RP 2 . Take r0 as
the starting and final point in the system. As n(r) is a physical, necessarily
single-valued field, we start at a certain point n(r0 ) of RP 2 , go around for
a trip on RP 2 and come necessarily back to the same point n(r0 ). Thus,
a closed curve in V is led into a closed curve in RP 2 . But the curve on
RP 2 can make a certain number of turns before coming back to the original
point. The Franck index is precisely this number, the number of loops of
the curve on RP 2 corresponding to one loop on the system. It is easier to
visualize things on the sphere S 2 , provided we are attentive to the antipodes
(see Figure 3.17). A closed curve on RP 2 has this difference with respect to
a closed curve in S 2 , that when we say the same point we can mean not
only the same point on the sphere, but also its antipode, so that on RP 2 a
11
Franck 1951.
3.3. C ELASTICITY
551
Figure 3.17:
curve is closed also if it connects two antipodes on S 2 . A curve looping to

the same point on the sphere will have integer index. But a curve connecting
two antipodes will have a half-integer value of .
Looking at Figure 3.16, we see the particularly clear relationship between
the topological characteristic, on one hand, and symmetry plus boundary
conditions, on the other.
Timoshenko & Goodier 1970

Love 1944
Landau & Lifchitz 1990
Nabarro 1987
de Gennes & Prost 1993
552
Phys. Topic 4
PROPAGATION OF DISCONTINUITIES
1
2
3
4
4.1
Characteristics
Partial differential equations
Maxwells equations in a medium
The eikonal equation
Characteristics
Given a system of first-order partial differential equations, its solution can always be obtained from the solution of a certain system of ordinary differential
equations. The latter are called the characteristic equations of the original
system. This is also true for some important higher-order equations. Such a
conversion from partial-differential into ordinary equations can be seen as a
mere method for finding solutions. For some physical problems, however, it
is much more than that. The physical trajectory, solution of Hamilton equations (which are ordinary differential equations) in configuration space, is the
characteristic curve of the solutions of the Hamilton-Jacobi equations (which
are partial differential equations). This points to their main interest for us:
the solutions of the characteristic equations (frequently called the characteristics) have frequently a clear physical meaning: particle trajectories, light
rays, etc. The classical example is Hamiltons approach to geometrical optics
(Phys.5.1).
There are two main views on characteristics:
(i) they are lines (or surfaces, or still hypersurfaces, depending on the
dimensionality of the problem) along which disturbances or discontinuities
propagate, in the limit of short wavelength (geometric acoustics and/or geometric optics); thus, they appear as lines of propagation of the quickest
perturbations. The surface (or line, or still hypersurface) bordering the
region attained by the disturbance originated at some point, called the char553
554
PHYS. TOPIC 4. PROPAGATION OF DISCONTINUITIES
acteristic surface, is conditioned by causality, which reigns sovereign in this

point of view.
(ii) they are lines perpendicular to the wavefronts; this approach relates
to the Cauchy problem of partial differential equations.
The first view has been the traditional one, but the second won the front
scene in the forties, with Luneburgs approach to Geometrical Optics. He
emphasized the identity of the eikonal equation and the equation of characteristics of Maxwells equations which, as we shall see, governs the propagation
of discontinuous solutions.
Let us mention the two noblest physical examples. One appears in the
above mentioned relationship between Hamilton equations and HamiltonJacobi equations, described in Phys.2. The hamiltonian flow is generated by
a field XH , which gives the time evolution of a dynamical function F (q, p, t)
according to the Liouville equation (Phys.1), a partial differential equation
2n
d
i
F (q, p, t) =
XH
(x)
F (x, t)
i
dt
x
i=1
(4.1)
with some initial condition

F (q, p, 0) = f0 (q, p).
(4.2)
If Ft (x) is a flow, the solution will be F (x, t) = f0 (Ft (x)). The orbits of the
vector field XH are the characteristics of equation (4.1).
The other example is given by the eikonal equation

2
2

2
+
+ ... +
= 1,
(4.3)
x1
x2
xn
in which is the optical length and the level surfaces of are the wavefronts
(Phys.5).
4.2
Partial differential equations
The classical lore1 on the characteristics of partial differential equations runs

as follows (to make things more visible, we shall talk most of time about
the two-dimensional case, so that we shall meet characteristic curves instead
of characteristic surfaces or hypersurfaces). The most general second order
linear partial differential equation will be of the form
2f
2f
f f
2f
A 2 + 2B
+ C 2 F (x, y, f,
,
) = 0,
x
xy
y
x y
1
See Sommerfeld 1964b.
(4.4)
4.2. PARTIAL DIFFERENTIAL EQUATIONS
555
where the source term F is a general expression, not necessarily linear in

and f
. It is convenient to introduce the notations:
f , f
x
y
p=
f
f
2f
2f
2f
; q=
; r=
;
t
=
;
s
=
,
x
y
x2
xy
y 2
in terms of which the equation is written

Ar + 2Bs + Ct = F.
(4.5)
dp = rdx + sdy
(4.6)
dq = sdx + tdy.
(4.7)
It follows also that

and
Now we ask the following question: given a curve = ( ) in the (xy)
f
) in the normal direction are given,
plane, on which f and its derivative ( n
f
does a solution exist? If f is given along , then ( f
is known. As from ( n
)
f
f
f
and ( ) we can obtain ( x ) and ( y ), f and its first derivatives, p and q,
are known on . In order to find the solution in a neighborhood of , we
should start by (in principle at least) determining the second derivatives r, s
and t on . In order to obtain them we must solve eqs.(4.6) and (4.7). The
condition for that is that the determinant
= Ady 2 2Bdxdy + Cdx2
be different from zero. There are then two directions (dy/dx) at each point
(x, y) for which there are no solutions. The two families of curves on which
= 0 are the characteristic curves. Along them one cannot find r, s and
t from the knowledge of f , p and q. Thus, in this line of attack, one must
require that be nowhere tangent to the characteristics. We shall see later
the opposite case, in which just coincides with a characteristic. Once the
non-tangency condition is fulfilled, there must be a solution in a neighborhood of . The miracle of the story is that, when looking for the higher order
derivatives in terms of the preceding ones, one finds, step by step, always the
same condition, with the same determinant. Thus, if is different from zero,
f can be obtained as a Taylor series.
The characteristic equation
Ady 2 2Bdxdy + Cdx2 = 0,
whose solutions correspond to
B
y
=
x
B 2 AC
C
(4.8)
556
determines, in principle, two families of curves on the plane (xy), which are
the characteristic curves.
There are three quite distinct cases, according to the values of the discriminant B 2 AC, and this leads to the classification of the equations and
of the corresponding differential operators appearing in (4.4):
B 2 AC < 0: the equation is of elliptic type;
B 2 AC > 0: the equation is of hyperbolic type;
B 2 AC = 0: the equation is of parabolic type.
Notice that A, B and C depend at least on x and y, so that the character
may be different in different points of the plane. Thus, the above conditions
are to be thought of in the following way: if the discriminant is negative in
all the points of a region D, then the equation is elliptic on D. And in an
analogous way for the other two cases.
Only for the hyperbolic type, for which there are two real roots 1 and
2 , is the above process actually applied. If the coefficients A, B, C are
functions of x and y only, then these curves are independent of the specific
solution of the differential equation (see the Klein-Gordon example below).
The families of curves are then given by (x, y) = c1 , which is the integral of
y 0 + 1 (x, y) = 0, and (x, y) = c2 , which is the integral of y 0 + 2 (x, y) = 0.
Suppose the differential equation has a fixed solution f = f0 (x, y). In
order to pass into geometric acoustics and/or optics, we
(i) add to it a perturbation f1 . Usually certain conditions are imposed on
such perturbations, conditions related to geometric acoustics or optics:
f1 is small, their first derivatives are small, but their second derivatives are relatively large. This means that f1 varies strongly at small
distances. It obeys the linearized equation,
A
2 f1
2 f1
2 f1
+
2B
+
C
= 0,
x2
xy
y 2
where, in the coefficients A, B and C, the function f is replaced by the

solution f0 . Then, we
(ii) write f1 in the form f1 = aei , with a large function (which is the
eikonal), and a a very slowly varying function (small derivatives), to
find the eikonal equation
2

2
A
+ 2B
+C
= 0.
(4.9)
x
x
y
y
Finally, we
4.2. PARTIAL DIFFERENTIAL EQUATIONS
557
, =
and dx
= d
the
(iii) find the ray propagation by putting k =
x
y
dy
dk
latter being the group velocity. The eikonal equation turns into the
dispersion relation Ak 2 2Bk + C 2 = 0, and then
dx
B Ak
=
,
dy
C Bk
which is an alternative form of
dx
B
=
dy
B 2 AC
.
C
Comment 4.2.1 We have been treating scalar optics: f is a scalar. In the real case
of optics, the procedure above must be followed for each component of the electric and
magnetic fields, as well as the four-potential.
As a very simple though illustrative case, consider the Klein-Gordon equation,
1 2f
2f
+ m2 c2 f = 0.
2
2
2
x
c t
As the equation is already linear, the perturbations will obey the same equation. The characteristics are (dx/dt) = c, that is, the light cone. The
dispersion relation for the eikonal equation will be = + kc; the source
term F of [4.4] is here the term ( m2 c2 f ), but it does not influence the
characteristics. This is general: whenever the coefficients A, B and C dependent only on the independent variables x and t, the characteristics are
independent of the special solution of the starting equation.
Only a few words on geometric acoustics.2 In a medium with sound
velocity c, the differential equations of the two families of characteristic curves
C+ and C are (dx/dt) = v + c and (dx/dt) = v c. The disturbances
propagate with the sound velocity with respect to a local frame moving with
the fluid. The velocities v + c and v c are the velocities with respect to the
fixed reference frame. Along each characteristic, the fluid velocity v remains
constant. Some perturbations are simply transported with the fluid, that is,
propagate along a third characteristic C0 , given by (dx/dt) = v. In general,
a disturbation propagates along the three characteristics passing through a
certain point on the plane (x, t). It can nevertheless be decomposed into
components, each one going along one of them.
Comment 4.2.2 What we gave here is a telegraphic sketch of a large, wonderful
theory. Systems of first order partial differential equations are equivalent to systems of
Pfaffian forms, from which a systematic theory of characteristics can be more directly
formulated.3
2
3
Landau & Lifshitz l989.

Westenholz l978; Choquet-Bruhat, DeWitt-Morette & Dillard-Bleick 1977.
558
4.3
Maxwells equations in a medium
Geometrical optics was traditionally regarded as an asymptotic approximation, for large wave numbers, of the wave solutions of Maxwells equations.
In two series of lectures delivered in the forties, Luneburg changed the tune.
He noticed the identity of the eikonal equation and the equation of characteristics of Maxwells equations. Think of a light signal emitted at t = 0,
which attains at an instant t the points of a surface defined by some function (x, y, z) = ct. The surface is a border, separating the region already
attained by the waves from the region not yet reached by any field. In the
inner side of the surface, the field has some nonvanishing value; at the other
side, the field is zero. The wavefront (x, y, z) represents a discontinuity of
the field, propagating at speed c, which may be point-dependent. Though a
little more involved at the start, this point of view has the great advantage
of treating light propagation no more as an approximation, but as a particular class of exact solutions of Maxwells equations, light rays appearing in
this view as lines along which discontinuities propagate. The equations, of
course, coincide with those obtained in the short wavelength treatment.
Consider an isotropic but non-homogeneous medium which is otherwise
electromagnetically inert (neither macroscopic magnetization nor electric polarization). This means that the electric and the magnetic permeabilities
depend on the positions but not on the directions. In anisotropic media
and become symmetric tensors, but we shall not consider this case here.
Then, with D = E and B = H, the sourceless Maxwell equations are
D
= 0;
t
B
= 0;
c rot E +
t
div D = 0;
div B = 0.
c rot H
(4.10a)
(4.10b)
(4.10c)
(4.10d)
The second and the fourth may be obtained respectively from the first and
the third by the duality symmetry: , E H, H E. The energy
is
W =
1
8
(E D + H B),
and the Poynting vector (energy flux vector) is

S=
Energy conservation is written
c
4
(E H).
4.3. MAXWELLS EQUATIONS IN A MEDIUM

W
t
559
+ div S = 0.
A hypersurface4 in En , we recall, is an (n1)-dimensional space immersed

in En . Consider the wavefront defined by as a closed surface = D
E3 , circumscribing the domain D whose characteristic function is a stepfunction (). Let us calculate some integrals:
Z
(i)
k f =
()f
()k f = k
E3
E3
k ()f
3
ZE
Z
(k )()f =
f (k );
E3
(ii)
Z
()i Vj =
i Vj =
E3
E3
i ()Vj
Z
Z
=
Vj (i )() = Vj (i );
E3
(iii)
(rot V )k = kij
Z
()i Vj = kij
i Vj = kij
3
ZE
i ()Vj
E3 Z
= kij
Vj (i )()f = kij
E3
(i )Vj ,
or
R
D
rot V =
(grad ) V.
The components i are proportional to the direction cosines of the normal

to the surface. The unit normal will have, along the direction k, the
component
k
|grad |
= Pk
i (i )
These results are in general of local validity, the surfaces being supposed to
be piecewise differentiable.
Take the Maxwell equation div D = 0. Its integral form will be obtained
by integrating it on a domain D and using the formula (ii) above:
Z
Z
Z
Z
div D = Di (i ) = D grad = (D grad ) .
D
4
Gelfand & Shilov 1964. More details are given below, in 7.5.17.
560
The same holds for the

R Maxwell equation
R div B = 0, so that these
equations say that both (D grad ) and (B grad ) vanish for arbitrary closed surfaces . The other equations, c rot H D
= 0 and
t
B
c rot E + t = 0, because of the time dependence, are better approached
by considering D as a domain in E4 instead of E3 . One could have done
it for the above Gauss theorems, with the simplifying fact that all timelike
components vanish. The closed surface is now 3-dimensional and the same
expressions above lead to the result that both
Z
Z
(c rot H t D) =
(cgrad H t D)
and
Z
Z
(cgrad E + t B)
(c rot E + t B) =
D
vanish for any closed surface in E4 .

Finally, all this may be used to study the case in which the fields are
discontinuous on a surface. We consider the domain D in E4 to be divided
into two subdomains D1 and D2 (Figure 4.1) by the spacelike surface 0 ,
defined by (x1 , x2 , . . . , x4 ) = 0.
Figure 4.1: Domain D in E4 , divided by 0 into two pieces D1 and D2 .
Integrating the equations on D1 , D2 and D = D1 + D2 , using the above

results separately for each region, and comparing the results, one arrives at
the following conditions for the discontinuities of the fields, indicated by the
4.4. THE EIKONAL EQUATION
561
respective brackets:
[H] c grad + [D] t = 0;
[E] c grad [B] t = 0;
[D] grad = 0;
[B] grad = 0.
(4.11a)
(4.11b)
(4.11c)
(4.11d)
It might seem that these are conditions on the field discontinuities. But
usually the discontinuities are given, or supposed, so that actually these are
conditions on the surface, on the function .
The formulae contain two main possibilities. Either the fields are discontinuous, or the permeabilities are. Using D = E and B = H, the
equations become:
[H] c grad + [E] t = 0;
[E] c grad [H] t = 0;
[E] grad = 0;
[H] grad = 0.
(4.12a)
(4.12b)
(4.12c)
(4.12d)
The case in which the permeabilities are not continuous is the usual one in
optical instruments. In such instruments, the hypersurface is fixed in time,
(x1 , x2 , x3 , x4 ) = (x1 , x2 , x3 ) = 0, so that t = 0 and the conditions
become:
[H] c grad = 0;
[E] c grad = 0;
[E] grad = 0;
[H] grad = 0.
(4.13a)
(4.13b)
(4.13c)
(4.13d)
This says that the tangencial components of E and H, as well as the

normal components of E and H, are continuous on the discontinuity
surface.
4.4
The eikonal equation
Suppose the permeabilities are continuous. In this case, eqs.[4.12] become

[H] c grad + [E] t = 0;
[E] c grad [H] t = 0;
[E] grad = 0;
[H] grad = 0.
562
Vector-multiplying by (cgrad ) the first equation and using the second,

[H] c2 grad grad + [H](t )2 = 0,
from which it follows that
{(grad)2

(t )2 }[E]
c2
= 0.
Using the equations in the inverse order we find instead

{(grad )2

(t )2 }[H]
c2
= 0.
As [E] and [H] are nonvanishing, the eikonal equation is forcible,

(grad )2
With the refraction index n =

(t )2 = 0.
c2
(4.15)
, it becomes
(grad )2
n2
(t )2 = 0.
c2
(4.16)
Because one starts by integrating all over D, the terms without derivatives just compensate and disappear. This will always be the case with
discontinuities: only the derivative terms contribute to the conditions on the
surface. Comparing with the extra terms and looking back to those terms
really giving some contribution to the ultimate result, one sees that only
those constituting a wave equation contribute (which, by the way, answers
for its ubiquity in Physics). What happens to the large-frequency approach
is then clear: the hypotheses made in the asymptotic approach are such that
only the higher derivative terms are left, so that the results coincide.
Of course, once the results are obtained, we may consider the surface to
be, at the beginning, the source of a disturbance, taking the field to be zero
at one side. The disturbance which is propagated is then the field itself.
Notice that the equation
[H] c grad + [E] t = 0
comes from c rot H E
= 0, and that
t
[E] c grad [H] t = 0
comes from c rot E + H
= 0. Recall how we get the wave equation: we
t
take the curl of
c rot H E
=0
t
4.4. THE EIKONAL EQUATION
563
to get
E
= 0,
c2 rot rot H c rot
t
and use the second,

c rot E = H
,
t
to arrive at
rot rot H +
2 H
c2 t2
= H +
2 H
c2 t2
= 0.
We see thus the parallelism of the two procedures, the operation (grad )
playing a role dual to taking the curl.
Luneburg 1966
Abraham & Marsden 1978
Guillemin & Sternberg 1977
564
Phys. Topic 5
GEOMETRICAL OPTICS
1
2
3
4
5
6
7
Introduction
The light ray equation
Hamiltons point of view
Relation to geodesics
The Fermat principle
Maxwells fish-eye
Fresnels ellipsoid
5.0
Introduction
The central equation of Geometrical Optics is the light ray equation. The
usual approach to it is to start by looking for asymptotic solutions of Maxwells
equations, fall upon the eikonal equation and then examine the ray curvature. The result is an equation in euclidean 3-dimensional space, which may
be interpreted as the geodesic equation in a metric defined by the refractive
index. We shall here proceed from the eikonal equation as obtained in Phys.4,
therefore parting with this traditional, asymptotic approach. Consequently,
we look at rays as characteristics, curves along which certain electromagnetic
discontinuities, the wave fronts, propagate.
The characteristic equation for Maxwells equations in an isotropic (but
not necessarily homogeneous) medium of dielectric function (r), magnetic
permeability (r) and refractive index index n = is

2 2 2
2
2
2
2 n
+
+
2
= (grad ) 2
= 0. (5.1)
y
z
c
t
c
t
Looking for a solution in the form (x, y, z, t) = (x, y, z) ct, we fall upon
565
566
PHYS. TOPIC 5. GEOMETRICAL OPTICS
the eikonal equation, or equation for the wave fronts, under the form

5.1
2

+
2

+
2
= (grad )2 = n2 (r).
(5.2)
The light-ray equation
The wave fronts are surfaces given by (x, y, z) = constant, consequently

integrals of the equation d = grad = 0. A light ray is defined as a path
conjugate to the wave front in the following
sense: if r is the position
p
vector of a point on the path and ds = dx2 + dy 2 + dz 2 is the element
of arc length, then u = (dr/ds) is the tangent velocity normalized to unity
(with the path length s as the curve parameter). The light ray is then fixed
by
dr
= d = grad .
(5.3)
n
ds
This means that applying the 1-form d to the tangent vector gives the
refractive index, d(u) = n. We shall later give another characterization
of this conjugacy. We may eliminate by taking the derivative, while
d
= uj j :
noticing that ds

d
dr
d
grad = dr
grad [grad ]
n
= ds
ds
ds
ds
= n1 [grad ] grad [grad ] =
1
2n
grad [(grad )2 ] =
1
2n
grad [n2 ];
thus,
d
ds
dr
n
ds

= grad n.
(5.4)
This is the same as

1
duk
[ k n uk uj j n] = 0,
ds
n
(5.5)
which is the differential equation for the light rays. The Poynting vector, and
thus the energy flux, is oriented along the direction of u. The equations for
the light rays are the characteristic equations for the eikonal equations, which
are themselves the characteristic equations of Maxwells equations. For this
reason they are called the bicharacteristic equations of Maxwells equations.
The curvature of a curve at one of its points is 1/R, with R the radius of
/R, where n
is the unit vector along
the osculating circle. As a vector, it is n
the radius. When u is the unit tangent vector and s is the length parameter,
5.2. HAMILTONS POINT OF VIEW
567
/R = (du/ds). The ray curvature (or curvature vector of a ray) is thus the
n
vector field K = (du/ds). Equation [5.5] is equivalent to
K=
1
(u grad n) u,
n
(5.6)
which is the form most commonly found in textbooks.
5.2
Hamiltons point of view
We are looking at things in E3 , of course. More insight comes up from the

following analogue in particle mechanics. Let us consider the cotangent bundle T E3 and there take coordinates (x, y, z, px , py , pz ), pi being the conjugate
coordinate to xi . This means that the natural symplectic form of T E3 may
be written locally in the form = dxk pk (see Phys.1). When needed,
we may use the euclidean metric to relate vectors and covectors, which in
the cartesian coordinates we are using is the same as identifying them. We
shall see that another metric will be simultaneously present, and playing a
fundamental role.
Consider the hamiltonian
H=
1
2n2
[j p2j ],
write p = grad and rephrase the eikonal equation as H(p) = 21 . We may

think of the particle as endowed with a position-dependent mass n2 . The
corresponding hamiltonian field (bicharacteristic field) is
"
#
1
1 X 2 n
X = 2 k p k k + (
p)
.
(5.7)
n
x
n j j xk pk
This field is symplectically dual to the form
dH = iX = iX (dxk pk ) .
The bicharacteristic curve is an integral curve of X lying on the characteristic manifold H(p) = 12 . Notice that we are here talking about bicharacteristic or characteristic objects (fields, equations and curves) on the phase space,
of which the characteristic objects on the configuration space are projections.
As k pk x k = 2H, the lagrangian corresponding to H above is
L[] =

1 2 2
n x + y 2 + z 2 .
2
(5.8)
568
This gives the kinetic energy which is related to the Riemann metric
p
n dx2 + dy 2 + dz 2 = nds.
(5.9)
We have thus a metric

gij = n2 (x, y, z)ij ,
(5.10)
which we call the refractive metric.

Now an important point: with this new metric, a new relationship between vector and covectors comes up. Let us examine the velocity vector v
= (x,
y,
z).
It is v = r = (dr/d ) , where is the proper time, given by
d 2 = n2 (dx2 + dy 2 + dz 2 ).
Its relation to the unit tangent above is u = nv, and its components satisfy
i v i v i = 1/n2 . Just as (d/ds) = uj j , the derivative with respect to the new
parameter is (d/d ) = v j j . The momentum p above is just its covariant
image by the refractive metric, pi = gij v j = n2 v i . We may even write, as
usual in geometry, the contravariant version as pi = g ij pj = v i .
We arrive in this way at a better characterization of the above mentioned
conjugacy between the wavefront and the trajectory: it is summed up in
d(v) = i pi v i = i gij v i v j = 1.
The gradient defining the family of surfaces = constant, which is the differential form d, applied to the tangent velocity to the path, gives 1. As
shown in the Fresnel construction, which will be given below, this means
that, seen as an euclidean vector, the covector p at each point of the path is
orthogonal (in the euclidean metric) to the plane tangent to the wavefront.
Comment 5.2.1 Let us repeat that there are two different velocities at work here: vj
= (dxj /d ) and uj = (dxj /ds), with d = nds and u = nv. As it happens whenever the
curve is parametrized by the length (s here), the corresponding velocity is unitary: |u|
= 1. Thus, the velocity v along a ray and the one normal to the wavefront, the gradient p
of the surface = constant, are respectively a vector and a covector related by the metric
given by the refraction index: p = nu = n2 v.
5.3
Relation to geodesics
Let us show that light rays are simply the geodesics of the refractive metric
[5.10]. The corresponding Christoffel symbols are

k ij = n1 ik j n + jk i n ij kr r n = (ik j) ij kr r (ln n), (5.11)
5.3. RELATION TO GEODESICS
569
where the notation (i . . . j) indicates index symmetrization. The geodesic

equation in this case is
Dv k
dv k
dv k
=
+ k ij v i v j =
+
D
d
d
1
n
k j

2v v j n i (v i v i ) k n = 0.
(5.12)
Changing variables from and v to s and u, we find just the light ray
equation. In the inverse way, the equation for the light rays is

d
d
d
2 dr
n dr
=
n
= ds
(n2 v)
grad n = ds
ds
ds
d
d
d
d
d 2
n + n2 ds
v = 2vn ds
n + n2 ds
v,
= v ds
or
d
ds
Using
dv k
d
1 dv k
,
n ds
v+
2
n
d
v ds
n
1
n2
grad n = 0.
we find
dv k
d
2
n
v k dd n
1
n3
k n = 0,
the geodesic equation above. By the way, this equation becomes particularly
simple and significant when written in terms of pk : it reads
dpk
= k (ln n).
d
(5.13)
The logarithm of the refractive index acts as (minus) the potential in the
mechanical picture.
Thus, the equation for the ray curvature just states that the light ray is
a geodesic curve in the refractive metric gij = n2 ij . The procedure above is
general. If we write pi = gij vj = i , then the calculation of (dpk /d ) leads
automatically to
dv k
+ k ij v i v j = 0,
(5.14)
d
being the Levi-Civitta connection of the refractive metric. The inverse
procedure works if p is an exact 1-form, that is, a gradient of some , because
in a certain moment we are forced to use i pj = j pi . Anyhow, one always
finds
dpk
= 12 g ij k (pi pj ) = 12 k (g ij ) pi pj .
(5.15)
d
Comment 5.3.1 The condition rot (nu) = rot (n2 v) = 0, known in Optics as the
condition for the existence of the eikonal, is an obvious consequence of the Poincare
lemma, as p = d.
570
Comment 5.3.2 As seen, the relationship between optical media and metrics is deep
indeed. Mathematicians go as far as identifying the expressions optical instrument and
Riemannian manifold.1
Metric [5.10] is the euclidean metric multiplied by a function. This kind

of metric, related to a flat metric (that is, to a metric whose Levi-Civitta
connection has vanishing Riemann tensor) by the simple product of a function, is said to be conformally flat (see Math.11). This means that, though
measurements of lengths differ from those made with the flat metric, the
measurements of angles coincide. The refractive metric, as a consequence,
has a strong analogy with the conformally flat metrics (the de Sitter spaces,
see Phys.9) appearing in General Relativity. There, the corresponding flat
space is Minkowskis. A consequence of this common character is found in
similar behaviour of geodesics, as in the fact that anti-de Sitter universes
have focusing properties quite analogous to that of an optical ideal apparatus, Maxwells fish-eye (see Phys.5.5 below).
We may look for the Euler-Lagrange equation for the lagrangian [5.8].
However, we find that
k
L
2 Dv
=
n
,
(5.16)
xk
D
L
so that x
In this way the
k = 0 is equivalent to the geodesic equation.
equivalence is established between the mechanical and the optical points
of view, at least in what concerns the equations.
5.4
The Fermat principle
We have seen that the differential equations given by the hamiltonian field
[5.7], in the form of Hamilton equations, are equivalent to the geodesic equation for the refractive metric, or still to the Lagrange equations
R written in
hamiltonian form. The geodesics extremize the arc length nds for this
metric. Now,
R
R
nds = d
on a path is the optical length of , or its time of flight. Thus, the light
rays are those paths between two given points which extremize the optical
length. This is Fermats principle.
The surfaces = constant may be seen as surfaces of discontinuity
(Phys.4), or as wavefronts. The higher the value of |grad |, the closer
are these surfaces packed together. The eikonal equation would say that the
1
See, for example, Guillemin & Sternberg 1977.
5.5. MAXWELLS FISH-EYE
571
refractive index is a measure of the density of such surfaces. Thus, the discontinuities propagate more slowly in regions of higher index. If we interpret
grad as a vector, it will be tangent to some curve, it will be a velocity
which is larger in higher-index regions.
5.5
Maxwells fish-eye
Consider the unit sphere S 2 and its stereographic projection into the plane.
A point on S 2 will be fixed by the values X, Y, Z of its coordinates, with X 2
+ Y 2 + Z 2 = 1 (for more details, see Math.11). The relation to spherical
coordinates are X = sin cos ; Y = sin sin ; Z = cos . We choose the
north pole (0, 0, 1) as projection center and project each point of the
sphere on the plane Z = 0. The corresponding plane coordinates (x, y) will
be given by
Y
X
; y=
.
(5.17)
x=
1Z
1Z
Call
X2 + Y 2
2
2
2
r =x +y =
.
(5.18)
(1 Z)2
The line element will then be
ds2 = dX 2 + dY 2 + dZ 2 = 4
dx2 + dy 2
.
(1 + r2 )2
(5.19)
This corresponds to a 2-dimensional medium with refractive index

n=
2
.
1 + r2
(5.20)
It is found that the geodesics are all given by
(x2 R2 1)2 + y 2 = C 2 ,
with some constants R and C. Thus, they are all the circles through the
points (0, 1). All the light rays starting at a given point will intersect
again at another point, corresponding to its antipode on S 2 . This is an
example of perfect focusing.
Comment 5.5.1 A manifold such that all points have this property is called a wiedersehen manifold. The sphere S 2 is a proven example, but it is speculated that others exist,
and also conjectured that only (higher dimensional) spheres may be wiedersehen manifolds.
But things become far more exciting in anisotropic media. Let us say a
few words on crystal optics.
572
5.6
Fresnels ellipsoid
In an electrically anisotropic medium, the electric displacement D is related

to the electric field E by Di = j ij Ej , where ij is the electric permitivity
(or dielectric) tensor. The electric energy density is
We = 21 E D = 12 ij ij Ei Ej .
(5.21)
The Fresnel ellipsoid is given by ij ij xi xj =constant= 2We . The 2tensor = (ij ) is non-degenerate and symmetric, the latter property being a
consequence of the requirement that the work done in building up the field,
dWe = 12 E d D,
be an exact differential form. Consequently, is a metric, E is a field and
D its covariant version according to this metric. Now, the metric can be
diagonalized, in which case the field and cofield are colinear in the 3-space
E3 , or parallel: Di = i Ei . The ellipsoid becomes
i i (xi )2 = 2We .
(5.22)
The metric eigenvalues i are the principal dielectric constants. The construction to get D from E using the ellipsoid is analogous to the Poinsot
construction for obtaining the angular momentum of a rigid body from its
angular velocity (Phys.2.3.10). It is also the same given above to obtain p
from v. Diagonalizing the metric corresponds to taking the three cartesian
axes along the three main axes of the ellipsoid.
Usual crystals are magnetically isotropic, or insensitive, so that the magnetic permeability may be taken as constant: = 0 . The magnetic induction B and the magnetic field H are simply related by B = H = 0 H.
Thus, the (squared) refraction index is given by the tensor (n2 )ij = 0 ij . By
diagonalization as above, Fresnels ellipsoid becomes
i n2i Ei2 = constant = 2We .
The ni s are the principal refraction indices. Along the principal axes of the
ellipsoid, of size (1/ni ), light will travel with the so called principal light
velocities,
c
1
ui =
=
.
(5.23)
ni
i 0
In the above procedure, D is seen as a form, a covector, while E is a vector.2
2
That is why Sommerfeld 1954 (p. 139, footnote) talks of the E components as point
coordinates and of those of D as plane coordinates.
5.6. FRESNELS ELLIPSOID
573
Of course there is an arbitrariness in the above choice: we might instead

have chosen to use the inverse metric 1 , with D as the basic field. In this
case another ellipsoid comes up, given by
i 2
i nxi = 2We ,
which has the advantage that the principal axes are just the principal refraction indices. This is called the index ellipsoid, reciprocal, ellipsoid of
wave normal, Fletchers ellipsoid, or still optical indicatrix. .
Synge 1937
Sommerfeld 1954
Born & Wolf 1975
Gelfand & Shilov 1964
Luneburg 1966
574
Phys. Topic 6
CLASSICAL RELATIVISTIC FIELDS
A THE FUNDAMENTAL FIELDS

1 Introduction
B SPACETIME TRANSFORMATIONS
2 The Poincare group
3 The basic cases
C INTERNAL TRANSFORMATIONS
4 Global and local gauge transformations
D LAGRANGIAN FORMALISM
5 The Euler-Lagrange Equation
6 First Noethers theorem
7 Minimal Coupling Prescription
8 Local phase transformations
9 Second Noethers theorem
10 Using general frames
6.1
A The fundamental fields
6.1 Introduction
Elementary particles must have a well-defined behaviour under changes of
inertial frames in Minkowski spacetime. Such changes constitute the Poincare
group (inhomogeneous Lorentz group). In order to have a well-defined behaviour under the transformations of a group, an object must belong to a
representation, to a multiplet. If the representation is reducible, the object
is composite, in the sense that it can be decomposed into more elementary
objects belonging to irreducible representations. Thus, truly elementary particles must belong to irreducible representations and be classified in multiplets of the Poincare group. Each multiplet will have well-defined mass and
helicity.
575
576
PHYS. TOPIC 6. CLASSICAL RELATIVISTIC FIELDS
In Quantum Field Theory, elementary particles come up as field quanta.

Their quantum numbers, such as mass and helicity, are fixed by the corresponding free fields. This means that, to be related to a particle, a field must
exist in free state, or to be well defined far away from any region of interaction. It is not clear that every field has such asymptotic behaviour. It is
not clear, in particular, that fields such as those corresponding to quarks and
gluons do describe real particles. There is, consequently, a modern tendency
to give priority to fields with respect to particles. We shall talk about fields.
Comment 6.1.1 The word field, as used here, is of course not to be mistaken by
the algebraic structure of Math.1 1.7. Neither is it to be taken as the geometrical fields
which are natural denizens of the tangent structure of any smooth manifold and are, at
each point, vectors of the linear group of basis transformations. Those are linear vector
fields. In Relativistic Field Theory, fields (scalar, vector, spinor, tensor, etc) belong to
representations of the Lorentz group, as specified below (Phys.6.2). They are Lorentz
fields. In particular, the vector fields of Field Theory (like the electromagnetic 4-vector
potential) are mostly represented as 1-forms, covectors of the Lorentz group seen as a
subgroup of the linear group in Minkowski space.
6.2
B Spacetime transformations
6.2 The Poincar

e group
The Poincare group is the group of motions (isometric automorphisms)
of Minkowski spacetime.
Comment 6.2.1 This means that what was said in the Introduction holds for free
systems, in the absence of interaction. Only the total quantum numbers are preserved
in interactions. Individual particles are identified only far from the interaction region,
where their momenta (consequently, masses) and helicities are measured. The status of
confined particles like quarks is not clear.
Acting on spacetime, the Poincare group P is the semi-direct product of

the (homogeneous) Lorentz group L = SO(3, 1) by the translation group T ,
P = L T , its transformations being given in cartesian coordinates as
x0 = x + a .
(6.1)
Let us first consider the Lorentz group. If is the Lorentz metric matrix,
the matrices = ( ) satisfy T = , with T standing for the transpose
matrix. This is the defining condition for the Lorentz group. The matrices for
a general Lorentz transformation will have the form = exp[ 21 J ], where
= are the transformation parameters, and J are the generators
obeying the commutation relations
[J , J ] = J + J J J
(6.2)
6.2. B SPACETIME TRANSFORMATIONS
577
Each representation will have as generators some matrices J . Expression [6.2] holds for anti-adjoint generators. There is no special reason to use
self-adjoint operators, as anyhow the group SO(3, 1), being non-compact, will
have no unitary finite-dimensional representations. Furthermore, SO(3, 1)
cannot accommodate half-integer spin particles. The true Lorentz group of
Nature is SL(2, C), which has also spinor representations and is the covering
group of SO(3, 1). We put, consequently, L = SL(2, C). Thus, the classifying group of elementary particles is the covering group of the Poincare
group. It is more practical to rechristen the Poincare group and write
P = SL(2, C) T . The translation group has generators T , and the remaining commutation relations are the following:
[J , T ] = T T ,
(6.3)
[T , T ] = 0.
(6.4)
Besides mass and helicity, particles (and their fields) carry other quantum
numbers (charges in general: electric charge, flavor, isotopic spin, color, etc),
related to other symmetries, not concerned with spacetime. For simplicity, we
shall call them internal symmetries. If G is their group, the total symmetry
is the direct product P G. A particle (a field) is thus characterized by being
put into multiplets of P and G. A particle with a zero charge is invariant
under the respective transformation and is accommodated in a singlet (zerodimensional) representation of the group.
There are two kinds of internal symmetries. They may be global (as that
related to the conservation of baryon number), independent of the point in
spacetime; or they may be local (those involved in gauge invariance). In the
last case, the above direct product is purely local, the fields being either in
a principal (if a gauge potential) or in an associated fiber bundle (if a source
field).
6.3 The basic cases
Relativistic fields are defined according to their behaviour under Lorentz
transformations, that is, according to the representation they belong to. Fortunately, Nature seems to use only the lowest representations: the scalar, the
vector and the spinor representations.
Scalar fields (belonging to a singlet) are those which remain unchanged:
0 (x0 ) = (x).
(6.5)
They obey the Klein-Gordon equation

t(x) + m2 (x) = 0.
u
(6.6)
578
Vector fields are those which transform according to

i
0 0
(x ) = (x) = exp { M } (x),

2
(6.7)
where each M is a 4 4 matrix with elements

[M ] = i( ).
(6.8)
This matrix basis is chosen so that

= exp { 2i M } = exp[ ],
that is to say, the components coincide with the matrix elements.
Comment 6.2.2 An example of vector field is given by the spacetime cartesian coordinates themselves. Another is the electromagnetic field, a zero mass case involving
furthermore a local gauge invariance. The basic equations are Maxwells equations, examined in Phys.4 and Phys.7:
F + F + F = 0;
(6.9)
F = J
(6.10)
As F = A A , the potential submits to the wave equation

A + A = J .
Spinor fields (bispinor representation) behave as

0 (x0 ) = [exp {
i
}](x),
4
(6.11)
where 21 are 4 4 matrices generating the bispinor representation.

Comment 6.2.3 Under infinitesimal Lorentz transformations with parameters ,
bispinor wavefunctions and their conjugates will change according to
i
i
= ; = .
4
4
(6.12)
Spinor fields in the absence of interactions obey the Dirac equation

i m = 0.
(6.13)
6.3. C INTERNAL TRANSFORMATIONS
6.3
579
C Internal transformations
6.4 Global and local gauge transformations

Fields will further belong to representations U (G) of internal transformation groups G. Under a transformation given by the element g of the
group G, their behaviour is generically represented by
0i (x) = [U (g)]ji j (x) = [exp { a Ta }]ji j (x).
(6.14)
The Ta s are generators in the U -representation.

If the transformation parameters a are independent of spacetime points,
the above expression represents global gauge transformations; if the transformation parameters a are point-dependent, it represents local gauge transformations.
The fields j (x) are source fields. Gauge potentials (interaction mediators) may be written as A = Ja Aa , with Ja the generators in the adjoint
representation. Under a local transformation generated by the group element
g = exp{ a Ja }, they behave according to
A0 = g 1 A g + g 1 g.
(6.15)
The corresponding infinitesimal transformation is

b = f b cd Ac d b .
A
(6.16)
with f b cd the structure constants of the group. Field strengths F = Ja F a

transform according to
0
F
= g 1 F g,
(6.17)
with the corresponding infinitesimal version given by
F b = f b cd F c b .
(6.18)
Gauge fields are the special subject of Phys.7.
6.4
D Lagrangian formalism
6.5 The Euler-Lagrange Equation

A physical system will be characterized as a whole by the symmetryinvariant action functional
Z
S[] = d4 x[(x)],
(6.19)
580
where represents collectively all the involved fields (Math.8). The variation of a field i may be decomposed as
i (x) + x i (x).
i (x) =
(6.20)
The second term in the right-hand side is the variation due to changes
x = x0 x in the coordinate argument. As to the first,
i (x) = 0 (x) i (x)
i
is the change in the functional form of i at a fixed value of the argument.
Notice that
= 0.
[ , ]
(6.21)
The general variation of S[] is given by:
Z
S[] = [d x] [(x)] + d x [(x)] = [d4 x][(x)]

Z
[(x)]
[(x)]
4
i (x) +
i (x) + [(x)]x . (6.22)
+ dx
i (x)
i (x)
with [d4 x] a symbolic notation to signify the variation of the integration
measure, which is ( x )d4 x in Cartesian coordinates. Collecting terms,

Z
[(x)]
[(x)]
4
i (x) +
i (x) + [(x)]x
S[] = d x
.
i (x)
i (x)
(6.23)
where the Lagrange derivative (Math.7) is simply
[(x)]
[(x)]
[(x)]
=

i (x)
i (x)
i (x)
(6.24)
because we are supposing the lagrangian density to depend only on and on

the first derivative . The Euler-Lagrange equations are then [(x)]
= 0.
i (x)
6.6 First Noethers theorem
The first Noethers theorem is concerned with the action invariance under
a global transformation: it imposes the vanishing of the derivative of S with
respect to the corresponding constant (but otherwise arbitrary) parameter
a . The condition for that, from [6.23], is

i (x)
i (x)
[(x)]
[(x)]
x
=
+ [(x)] a .
(6.25)
i (x) a
i (x) a
6.4. D LAGRANGIAN FORMALISM
581
For i satisfying the Euler-Lagrange equation, the current

i (x)
[(x)]
x
Ja =
+ [(x)] a
i (x) a
(6.26)
is conserved. Internal symmetries are concerned with the first term, while
spacetime symmetries are concerned with the last one. Let us then examine
some examples.
(i) Translations are given by:
x0 = x + x = x +
x
a .
a
(6.27)
If we take the x themselves as parameters, then a

= . Fields are Lorentz
i
tensors and spinors, and as that they are unaffected by translations: a
= 0.
Consequently, i = ( i )x . The Noether current related to spacetime

translations in the energy-momentum tensor density
[]
i .
i
As an example, let us consider a fermion field, for which

= 2i [ ] m,
(6.28)
(6.29)
and consequently
=
i
2

[ ] .
(6.30)
(ii) Lorentz transformations are,, in their infinitesimal form, given by

(6.31)
x = 2i M x = 12 [ ] x = x ,
where use was made of [6.8]. Consequently,
x

= ( x x ) ,
and the Noether current will be the total angular momentum current density:
M
=
i
i
[]
x

. (6.32)
i

i
The last term is the spin current density

S =
i

,
i
(6.33)
582
which appears only when the field is not a Lorentz singlet. From the conservation laws M = 0 and = 0, it follows that
S = .
(6.34)
If a vector field has Lagrangean density

=
1
2

m2 ,
(6.35)
its spin current density reads

S = .
(6.36)
= i m
(6.37)
For a fermion field

and the spin current density is
S =
1
4
{ + } .
(6.38)
(iii) Phase transformations related to abelian groups are of the form

0
0 = eiq ; = eiq .
(6.39)
In this case, the Noether current will be the electric current J = q .

For non-abelian transformations like [6.14], the current is
Ja =
i (x)
[(x)]
[(x)]
[(x)]
=
[Ta ]ji j =
Ta . (6.40)
a
i (x)
i (x)
i (x)
Under global transformations, b = 0 in [6.16], and gauge potentials obey

b = f b cd Ac d .
A
(6.41)
The corresponding Noether current will be the self-current

j a = f a bc Ab F c .
(6.42)
The conservation law will be given by

(Ja + ja ) = 0.
(6.43)
583
6.7 Minimal Coupling Prescription

These currents may also be obtained from the free lagrangians for
and G for Ab through the minimal coupling rule 0 , where 0 has
the form of but with usual derivatives replaced by covariant ones:
D = + Aa
,
a
(6.44)
with the total (Lagrange) functional derivative
=

a
a
a
and not simply
.
a
(6.45)
As a consequence,
D = + Aa
+ Aa
.
a
(6.46)
Once this is made, the currents can be written as

Ja
0
G
= a ; ja = a .
A
A
(6.47)
In general, source fields transformations do not depend on the parameter

derivatives and for them the last term of [6.46] does not contribute, but that
term is essential when we take the covariant derivative of the gauge potentials
to obtain the field strength: under local transformations, the gauge potential
transforms according to [6.16] and, consequently,
b
D Ab = Ab Ab + f b ac Aa Ac = F
.
(6.48)
6.8 Local phase transformations

The self-current, as given by [6.42], is covariant only under global transformations, but not under local transformations: in fact, it is just
ja =
G
,
Aa
(6.49)
as we can see by comparing the Yang-Mills equation

E a =
G
= F a + f a bc Ab F c = J a
Aa
(6.50)
with equation [6.42]. This is an example of the well known result of Mechanics
(Phys.2.2), which says that simple functional derivatives are not covariant,
584
whereas Lagrange derivatives are. As the time component of a current is the

charge density, the total charge is its space-integral. The continuity equation
J a = 0 then implies the time-conservation of the charge. In the above
case, as the total current is J a + j a = F a , the continuity equation is
automatically satisfied and the corresponding charge
Z
Z
Z
Z
a0
0
k0
Q=
Ja F
=
F =
k F =
d2 i F i0
(6.51)
V
is time-independent. To ensure the covariance of the charge, we should have

Z

Z
Z
0
2
0
2
1
1
2
Q =
d F =
d g F g = g
d F g = g 1 Qg. (6.52)
V
As g = g(x) = e (x)Ja , in order to extract g from inside the integral, it is

necessary that a (x) be constant at the surface V .
Finally, we should mention that, if we had insisted in using the complete
derivative
G
ja = A
a
instead of the ill-defined ja given by [6.49], the very expression E a in [6.50]

would result (with opposite sign) and the total current would be covariant
but vanishing for solutions of the field equation. We are consequently forced
to keep working with the non-covariant current and consequently restricting
the gauge transformations to become global beyond a certain distance. In
that case, though the currents are not covariant, the charges are.
6.9 Second Noethers theorem
The second Noethers theorem is concerned with local transformations,
with point-dependent parameters. We shall consider here only the case of
a gauge field Aa and a generic source field belonging to a representation
with generators {Ta }. The total lagrangian, supposed invariant, will be
= G + .
(6.53)
The important point is that the pure gauge lagrangian

G = 14 Fa F a
is invariant under gauge transformations. This means that also the source
lagrangian , which is the free lagrangian with the derivatives replaced by
covariant derivatives, is invariant by itself. More than that: if there are many
585
source fields, each one will contribute with a lagrangian chosen so as to be

isolatedly gauge invariant.
Consider again [6.23]. A local invariance means that the action remains
unmoved under transformations in a small region around any point in the
system. Take a point y inside the system and calculate from [6.14] and
[6.16] the following functional derivatives:
a (x)
A
= 4 (x y)f a bc Ac (x) ba 4 (x y)
b
(y)
(6.54)
(x)
= (x) 4 (x y)
a (y)
(6.55)
We see that the deltas concentrate the variations at the interior point y. The
last, derivative term of [6.23] is an integration on the surface of the system,
which for the generic field is
i R
h
i
h
R 4
[(x)]
[(x)]
3
d x i (x) i (x) = S d i (x) i (x) .
The integration variable x will be at the surface of the system, whereas
y represents a point inside the system. Due to the deltas, this term will
vanish. The remaining terms will give, after integration,
S
(y)
G (y) b
G (y)
c
=
(T
)
(y)
+
f
A
(y)
+
. (6.56)
a
ij
j
ac
a (y)
i (y)
Ab (y)
Ab (y)
When there is a local invariance, Sa (y)

= 0. We have said that each piece of
in [6.53] is independently gauge invariant. This means that actually the
variation vanishes for each field, so that
S
(y)
=
(Ta )ij j (y) = 0
a
(y)
i (y)
(6.57)
for each source field , and

G (y)
SG
G (y) b
=
f ac Ac (y) + b
=0
a
b
(y)
A (y)
A (y)
(6.58)
for the gauge field. Relations like [6.57] and [6.58], coming solely from the
invariance requirement and independent of the field equations, are said to be
strong relations. Consider first the latter. The last expression is
[ab + f b ac Ac ]
G (y)
= 0.
Ab (y)
(6.59)
586
We can introduce the notation

Db a = ab + f b ac Ac
(6.60)
for the covariant derivative, and write the strong relation for the gauge field
as
D D Fa = 0.
(6.61)
For the source field, [6.57] gives

(y)
(y)
(y)
(Ta )ij j (y) =

(Ta )ij j (y) = 0.
i (y)
i (y)
i (y)
(6.62)
Take for example the case of a fermion field, whose lagrangian is given by
(see Phys.7)

(6.63)
= 2i [ ] m
The source current is

Ja = i Ta Ta ,
which is also
Ja =
,
Aa
(6.64)
(6.65)
and the strong relation [6.62] takes the form

D Ja = 0.
(6.66)
The strong relations are not real conservation laws, but mere manifestations
of the local invariance. As said in Phys.6.7, only the sum J + j of the
source current and the gauge field self-current has zero divergence. And
this fact only leads to a meaningful (that is, covariant) conserved charge
under the additional proviso that the local transformations become constant
transformations outside the system.
6.10 Using general frames1
We have used above a Cartesian coordinated system. An alternative2 is to
introduce general coordinate systems or, still better, general frames. Notice
to begin with that, even on a curved spacetime of metric g, the tetrads
h = (h ) satisfy, under Lorentz transformations, h0T h0 = hT h; thus, the
1
2
Aldrovandi & Pereira 1991.

Fleming1987
587
metric g = hT h is a Lorentz scalar. The integration measure is

det(h ), and its variation in terms of the tetrads is
g = g x .
g = h =
(6.67)
Usual derivatives must be changed

But then S and
R 4 into covariant ones.
depend on the h s: S[, h] = d xh(i , D i , h ). Thus,

Z
S[, h] =
d x h(i , i , h ) + h h
h

, (6.68)
i + (
i ) + ( )x
+h
i
i
4
where we have used the property h = hh h . This is found by first

recalling that h = det(h ). Calling H the matrix (h ), h = det H =
exp[ tr ln H]. Take then ln h = tr ln H, which implies
h1 h = tr [H 1 H] = h h ,
so that
h = hh h = hh h .
(6.69)
Comment 6.4.1 We take the opportunity to comment upon the notion of covariant
derivative and the behavior of the connection. Let us examine the vector field case: to the
first order, a change in the coordinates leads to
0 (x0 ) = 0 (x + dx) = 0 (x) + dx
= (x)
i
2
(M ) = (x) dx
Thus,
(x) = 0 (x0 ) (x) = [ + ]dx =: D dx = D ,
and we have 0 (x) = (x) when D = 0, that is, when is parallel-transported. To
see how should change, it is enough to compare with
0 (x0 ) = 0 (x + dx) = (x)
i
2
(M ) .
Comment 6.4.2 The variation along a curve of tangent vector (velocity) u will be
(u) = [ + ]dx (u) = u D =
D
Ds
D
When = u itself, we have the acceleration Ds
u . The condition of no variation of the
velocity field along the curve will lead to the geodesic equation
D
Ds u
= u [ u + u ] =
d
ds u
+ u u = 0.
588
6.11 If we were to consider gravitation, h would be taken as an independent field. This is not our interest here (see Phys.8 for that). We shall here
remain on flat Minkowski space, though using tetrads, point-dependent general frames. Lorentz metric, as in general coordinates, becomes coordinate
dependent. The tetrads will here only represent coordinate transformations
on Minkowski space, through which Poincare transformations can be simulated. Thus, h will only relate to the change x and
h = h h x = h x .
The connection , on its side, is a mere transform of a cartesian = 0, so
that, in the basis {h } = {h dx }, it is
= h h .
6.12 For coordinate transformations x0 = x + x , we have h0 = h +
h , with
h h =
(x0 x )
y
x
.
y
Things are simpler if x = y, in which case h = and

h0 = +
x
y
This leads to the expected result for the measure variation:

h = hh h = h
y x
x
=
h
.
x y
x
(6.70)
We have consequently
= ,
h
the energy-momentum tensor previously obtained.
Bjorken & Drell 1964, 1965

Itzykson & Zuber 1980
Bogoliubov & Shirkov 1980
Konopleva & Popov 1981
(6.71)
Phys. Topic 7
GAUGE FIELDS
0 Introduction
A THE GAUGE TENETS
1 Electromagnetism
2 Nonabelian theories
3 The gauge prescription
4 Hamiltonian approach
5 Exterior differential formulation
B FUNCTIONAL DIFFERENTIAL APPROACH
6 Functional Forms
7 The space of gauge potentials
8 Gauge conditions
9 Gauge anomalies
10 BRST symmetry
C CHIRAL FIELDS
11 Some comments on chiral fields
0 Introduction
General Relativity has been for a long time the prototype of a physical theory with a geometric flavor. The badge of this geometrical character is given
by the fact that, under the action of a gravitational field, a test particle
moves freely in a curved space, the curvature being that of the Levi-Civita
connection of the gravitational field, which is a metric. The gravitational
interaction is thereby geometrized. Gauge theories are also of geometrical
character, as the gauge potentials are connections on general principal bundles. In both cases, a significant geometric stage set is supposed as a kind
of kinematical background to which dynamics is added. Gauge theories are,
in a sense, still more geometric than General Relativity, because there is a
duality symmetry between their dynamics and the geometric background.
589
590
PHYS. TOPIC 7. GAUGE FIELDS
Comment 7.0.3 Trautman and Yang have greatly emphasized the geometric approach to gauge theories. This has been roughly presented in the main text. We shall give
here a resume of the physicists approach.
7.1
7.1.1
A The gauge tenets

Electromagnetism
The electromagnetic lagrangian

= 12 i i[ ] m + eA 41 F F
(7.1)
is invariant under the transformations

(x) 0 (x) = ei (x),
(7.2)
where is a constant phase. Such transformations have been called gauge

transformations of the first kind, and the invariance leads to charge conservation by Noethers first theorem. The phase factor ei is an element
of the unitary group U (1) of 11 unitary matrices. Yang and Mills1 noticed that is also invariant when = (x) is point-dependent, provided
that simultaneously the potential changes according to the well known gauge
transformations of the second kind,
1
(7.3)
A (x) A0 (x) = A (x) + (x).
e
The field A must have this behaviour in order to compensate the derivative terms (x) turning up in the point-dependent case. The electromagnetic potential appears in this way as a compensating field, with some
analogy to the extra term needed in the Lagrange derivative in mechanics
(Topic Phys.2.2). The phase factors ei(x) are now point-dependent elements
of the group U(1), which is the gauge group. The phase symmetry then
requires the second Noethers theorem to be related to the charge, besides
some extra assumptions concerning asymptotic behaviour (see Topic Phys.6,
6.7 & 6.9).
The lagrangian can be rearranged as

= 12 i ( ieA ) i[( + ieA )] m 14 F F . (7.4)
It all works as if the effect of the electromagnetic field is solely to change
the derivative acting on the source field:
ieA .
1
Yang & Mills 1954.
(7.5)
7.1. A THE GAUGE TENETS
591
The change of derivative gives a rule to introduce the electromagnetic

interaction: given a source field, we write down its free field equations and
then change the derivatives. This is the minimal coupling prescription,
which will be generalized below. The new derivative was called covariant
derivative by analogy with General Relativity.
7.1.2
Nonabelian theories
But gauge theories2 in their modern sense were really inaugurated3 when
Yang and Mills proceeded to consider the isospin group SU (2), whose nonabelian character made a lot of difference. A fermionic source will now be a
multiplet, a field transforming in a well-defined way under the action of the
group. Utiyama4 generalized their procedure to other Lie groups and we shall
prefer to recall once for all the general case of a group G. The phase factors,
elements of the gauge group G in that representation, are now operators of
a
the form ei (x)Ta , and the source multiplets will transform according to
(x) 0 (x) = U (x)(x) = ei
a (x)T
a
(x).
(7.6)
The indices a, b, c, . . . = 1, 2, 3, . . . dim G, are Lie algebra indices, which are

lowered by the Killing-Cartan metric ab of the gauge group, supposed to be
semi-simple. The Ta s are the generators of the group Lie algebra, written in
the representation to which belongs.
The vector potentials, now one for each group generator, will have a
behaviour more involved than [7.6]. We write A (x) = Ta Aa (x), which
makes of A (x) a matrix in the representation of . In order to keep its
role of compensating field, this matrix gauge potential will have to transform
according to
A (x) A0 (x) = U (x)A (x)U (x)1 + i U (x) U (x)1 .
(7.7)
A certain confusion comes up here. Suppose another field , belonging

to some other representation of the gauge group G, appears as the source
2
The subject is treated in every modern text on Field Theory, from the classical
treatise by Bogoliubov & Shirkov 1980, to the more recent Itzykson & Zuber 1980; there
are many other books, as Faddeev & Slavnov 1978; for an excellent short introduction
covering practically all the main points, see Jackiw 1980.
3
Weyls pioneering version, in which gauge invariance appears as an indifferent scaling in ambient space, is summarized in Weyl 1932 and in Synge & Schild 1978. The original
pioneering literature is of difficult access: it includes Weyl 1919 and Weyl 1929; London
1927; Fock 1926; and O. Klein, contribution in New Theories in Physics, Conference of
the International Union of Physics, Warsaw, 1938. Excerpts can be found as addenda in
Okun 1984.
4
Utiyama 1955.
592
field. If the group generators are Ta0 in the representation of , then A (x)
must be written A (x) = Ta0 Aa (x). An examination of the sourceless case
shows that A (x) must actually be in the adjoint representation of G. By this
representation, the group acts on its own Lie algebra. If the generators of G
are Ja , they will satisfy the general commutation relations [Ja , Jb ] = f c ab Jc ,
with f c ab the structure constants, and will transform by Ja U 1 Ja U . The
vector potential will then be
A (x) = Ja Aa (x).
(7.8)
The particular expression of the covariant derivative changes from one

representation to the other. Acting on a field belonging to the representation Ta , the covariant derivative will be iAa Ta ; on a field in the
representation Ta0 , it will be iAa Ta0 .
There is more. The field strength F , invariant in the electromagnetic
case, must have here at least a well-defined behaviour under the gauge transformations. It is found that the only covariant expression is
F (x) = A A + [A , A ]
(7.9)
a matrix F = Ja F a in the adjoint representation which will, as a consequence of [7.7], transform according to
0
a
F (x) F
(x) = U (x)1 F (x)U (x) = U 1 Ja U F
.
(7.10)
The lagrangian for a fermionic source field will now be

=
1
2
i ( iAa Ta ) i[( + iAa Ta )]
m 14 F a Fa . (7.11)
The gauge field (strength) F is a 2-form,
F = 12 Ja F a dx dx
(7.12)
given in terms of the (1-form) gauge potential

A = Ja Aa dx
(7.13)
F = dA + 12 [A, A] = dA + A A.
(7.14)
as
In components, this is

F = 21 Ja Aa Aa + f a bc Ab Ac dx dx ,
(7.15)
593
from which one gets back relation [7.9],

F a = Aa Aa + f a bc Ab Ac
(7.16)
The Bianchi identity

[ F a ] + f a bc Ab [ F c ] = 0
(7.17)
is an automatic consequence of equation [7.16]. If we define the dual tensor

F a = 12 F a , it may be written as
F a + f a bc Ab F c = 0.
(7.18)
Notice that the dual presupposes a metric, so that the former version is in
principle to be preferred.
The field equations for gauge theories are the Yang-Mills equations
F a + f a bc Ab F c = J a
(7.19)
where the J a s are the source currents. This is equivalent to

[ F a ] + f a bc Ab [ F c ] = Ja
(7.20)
We observe that, in the sourceless case, the field equations are just the
Bianchi identities written for the dual of F . This is the duality symmetry. If
we know the geometrical background, we know the dynamics. For this reason we have said that gauge theories are still more geometric than General
Relativity. While in the latter dynamics is introduced independently, the
Yang-Mills equations are related by duality to the Bianchi identities, which
are purely geometric.
Comment 7.1.1 Notice that, when G is an N -dimensional abelian group, the theory
reduces formally to N electromagnetisms.
7.1.3
The gauge prescription
We arrive in this way at the general gauge prescription. In order to introduce

an interaction invariant under the local symmetry given by a group G, we
change all ordinary derivatives in the free equations into covariant derivatives
acting on each source field ,
= igAa Ta ,
(7.21)
where the Ta s are the group generators in the representation of G to which

belongs. The coupling constant g, which here takes the place of the charge
594
e of electromagnetism, can at this level of the theory be absorbed in A.

We shall, in the formulae which follow, ignore it. This is a particular
case of the general covariant derivative given in Phys.6.7. The field equation for each source field will involve the Lagrange derivative with the usual
derivatives replaced by the covariant ones:
(y)
(y)
(y)
=
D
= 0.
(y)
(y)
(y)
(7.22)
Because the covariant derivative is different when acting on fields in different representations, it is important to pay careful attention to its form on
(y)
each object. It will be [7.21] when acting on , but (y) is a (co-)vector
object alike to A , so that D will have the rotational form given in eq.
[6.48] of Phys.6. Once this is said and retained, it is a common practice to
use always the same symbol, and write in all cases.
7.1.4
Hamiltonian approach
There are many reasons to prefer the hamiltonian approach5 to Yang-Mills

equations6 . Space and time have very distinct roles in this formalism. The
electric and magnetic fields, respectively Ea i = Fa i0 and Ba i = 12 ijk F a jk , also
play different roles. In the hamiltonian formalism, the canonical coordinates
are the Aa k s and, given the sourceless lagrangian
G = 14 F a Fa ,
(7.23)
the conjugate momenta are the electric fields

ai =
G
= E ai = F ai0 .
0 Aai
(7.24)
The hamiltonian can be written in the form

Z

H = d4 x tr 0 A + A0 [A0 , A] E + 21 (E2 + B2 ) ,
(7.25)
and the action can be rewritten as

Z

S = d4 x tr 0 A E + 21 (E2 + B2 ) Aa 0 Ga (x) ,
(7.26)
5
6
Itzykson & Zuber 1980.

About its superiority, as well as for a beautiful general discussion, see Jackiw 1980.
595
where each Ga (x) in tr[Aa 0 Ga (x)] is the expression appearing in the nonabelian Gauss law, which reads
Ga (x, t) = Dk E ak = k E ak + f a bc Ab k E ck = 0.
(7.27)
We are here, as announced, ignoring factors involving the coupling constants. We see in [7.26] that A0 is the Lagrange multiplier enforcing Gauss
law, which appears as a constraint. Another constraint is the magnetic field
expression

(7.28)
Ba i = 12 ijk F a jk = ijk j Aa k + 21 abc Ab j Ac k .
The ensuing dynamical equations are Hamiltons equations: Ampères
law
1 i
E = ( B)i + ijk [Aj , Bk ] + [A0 , Ei ],
c t
and the time variation of the vector potential,
1
A = E A0 + [A0 , A].
c t
(7.29)
(7.30)
The nonvanishing canonical commutation relations are:

{Aa i (x), Eb j } = i ba ij 3 (x y).
(7.31)
The hamiltonian formalism is of special interest to quantization. In the

Schrodinger picture of field theory, the state is given by a functional [A],
a kind of wave function on the coordinates Aa i s. Applied to a general state
functional [A], the Aai s are to be seen as multiplication-by-function operators, while their conjugate momenta are operators,
Ea k [A] = i
[A],
Aa k
(7.32)
in complete analogy with elementary Quantum Mechanics.
7.1.5
Exterior differential formulation
We have given the basic formulae in the main text. Let us only repeat a few
of them for sake of completeness. The field strength [7.14] is the curvature
of the gauge potential A, which is a connection. Taking the differential of F
leads directly to the Bianchi identity
dF + [A, F ] = 0.
(7.33)
596
The Yang-Mills equations [7.19] are, in invariant notation,

+ 1 [A, F ] = J.
dF
(7.34)
Thus, in the sourceless case, we see clearly the invariant version of the duality
symmetry. A self-dual (or antiself-dual) 2-form in a 4-dimensional space,
solution of
F = F,
(7.35)
will respect
F = F = [F ] = F = ()(4s)/2 F = ()s/2 F.
(7.36)
In Minkowski spaces, the signature s = 2 and the self-duality implies

the vanishing of F , but in an euclidean 4-dimensional space there may exist
non-trivial self-dual solutions. Such self-dual euclidean fields are called instantons. Any self-dual F of the form F = dA+AA will solve automatically
the sourceless field equations.
It comes out clearly from the differential approach that the gauge prescription must be improved. There are covariant derivatives, but there are
also covariant coderivatives. The latter are to replace the usual coderivatives
of the free case. This is the case, for example, in the continuity equation.
7.2
B Functional differential approach
Functional forms (Math.8, to which we refer for the calculations) enlarge the
geometrical meaning of gauge fields.
7.2.1
Functional Forms
The Euler Form for a sourceless gauge field is

a
E = F a + f a bc Ab F c Aa = (D F
)Aa
(7.37)
The coefficient, whose vanishing gives the Yang-Mills equations, is the covariant coderivative of the curvature F of the connection A according to that
same connection. Each component Aa is a variable labelled by the double index (a, ), and f a bc are the gauge group structure constants. Let us examine
the condition for the existence of a lagrangian. Taking the differential,

E = F a + f a bc Ab F c + f a bc Ab F c Aa ,
(7.38)
the last term vanishes if we use the complete antisymmetry (or cyclic symmetry) of f a bc : the coefficients become symmetric under the change (a, )
7.2. B FUNCTIONAL DIFFERENTIAL APPROACH
597
(b, ). Integrating by parts the first term, using again the cyclic symmetry and conveniently antisymmetrizing in (, ), we arrive at the necessary
condition for the existence of a lagrangian:
E = 12 dF a dFa = 0.
(7.39)
The cyclic symmetry used above holds for semisimple groups, for which
the Cartan-Killing form is an invariant metric well defined on the group.
Actually, we have been using this metric to raise and lower indices all along.
No lagrangian exists in the nonsemisimple case.7 In the semisimple case, we
obtain
a
= 14 F a Fa .
(7.40)
G = 12 Aa D F
R
The action is just that of eq.[7.137], which is F F . We might consider
the action
Z
CG = F F.
(7.41)
It is not difficult to find that there exists a 3-form K such that F F = dK.
This means that CG is a surface term, that would not lead to local equations
by variation (though a na
R ve variationRwould lead to the Bianchi identities).
In the euclidean case, E4 F F = S 3 K. We now consider the field F
concentrated in a limited region, so that only the vacuum exists far enough,
that is, on a sphere S 3 of large enough radius. The potential will be given
by the last term of equation [7.7] (or [7.43] just below). Examination of the
detailed form of K shows that, with a convenient normalization, CG can be
put in the form
Z

1
d3 xtr ijk g 1 (x) i g(x) g 1 (x) j g(x) g 1 (x) k g(x) .
n = 242
S3
(7.42)
When the gauge group is SU (2), whose manifold is also S , the function
g(x) takes S 3 into S 3 . Once more, we can show8 that the integrand is actually
a volume form on S 3 , so that in the process of integration we are counting
the number of times the values g(x) cover SU (2) = S 3 while the variable
x covers S 3 one time. The normalization above is chosen so that just such
integer number n comes out. This is the winding number (see 3.3.13
and 6.2.15) of the function g. The values of the number n can be used to
classify the vacua, which are of the form g 1 dg. This topological number is a
generalization of the Chern number (section Math.10.4.2), introduced in the
bundle of frames, to general bundles on E4 with structure group SU (2).
3
7
8
Aldrovandi & Pereira 1986, 1988.

see Coleman 1979.
598
7.2.2
The space of gauge potentials
The functional approach gives an important role to the space of the gauge
potentials, on which the state functional [A] is defined. This space is usually
called the A-space and will be denoted by . As a function, the potential A
depends (i) on some fixed starting value a, and (ii) on the group element
g by whose transformation A is obtained from a. Thus, we rewrite [7.7]
as
A(a, g) = g 1 ag + g 1 dg.
(7.43)
This decomposition corresponds to specially convenient coordinates on
A-space. The vacuum term, v = g 1 dg, corresponds to the Maurer-Cartan
form of the group: one checks easily that dv + v v = 0.
In the functional case, A itself becomes a functional of the functions g(x)
and a(x). Notice that, in this case, also g(x) is to be seen not as an element
of the gauge group G, but as a member g of the space of G-valued
Figure 7.1: Local decomposition of into components along and perpendicular to the large group.
functions (g(x) is actually a chiral field, see below). This space, an infinite
group formed by all the gauge transformations on spacetime, is called the
large group, and will be denoted by .
The space is starshaped,9 so that the use of the homotopy formula
to get the Lagrangian G of [7.40] is straightforward, and G will be valid
on the whole as far as no subsidiary gauge condition is imposed. Of
course this is not the real physical space, which requires a choice of gauge
9
Singer 1981.
599
and is far more complicated.10 Given the large group , the physical space
is formed by the gauge-inequivalent points of , the quotient space /
constituted by the gauge orbits, or the space of points a(x). This leads
to a (local!) decomposition of into components along and perpendicular
to the functional large group (Figure 7.1). Variations on may be locally
decomposed into a part along and a part orthogonal to ,
Aa = || Aa + Aa .
(7.44)
The part || parallel to is a gauge transformation. Defined on there are

entities which act as representatives of the geometrical entities defined on the
gauge group G. Such representatives are, however, dependent on the point
in spacetime.
We have here an opportunity to apply functional exterior calculus. As the
exterior differential is that given by the variational differential, we may think
of the form = a Ja = g 1 g as a functional version of the Maurer-Cartan
form v = g 1 dg. We shall find below that a small correction is necessary to
this interpretation. The functional version of the Maurer-Cartan form will
be = g 1 || g, which stands along the large group. We obtain from [7.43]
A(a, g) = d + [A, ] + g 1 ag = D + g 1 ag,
(7.45)
from which we can interpret

|| A = D
(7.46)
as the usual gauge transformation of A:

|| A (a, g) = + [A , ].
(7.47)
A = g 1 ag,
(7.48)
But we see also that

which says that the perpendicular variation is the transformation of the
physical variation. It is the real variation, to the exclusion of any gauge
transformation. Gauge transformations will be given by
|| = a Xa ,
(7.49)
where the Xa s are the generators in the functional representation, to be

found in the following. Another interesting result is
v = d + [v, ].
10
Wu & Zee 1985.
(7.50)
600
It is sometimes more convenient to work with components. Take g =

a
e = e Ja . Then,
= g 1 g = g 1 ()g = g 1 (a Ja )g
= (a )g 1 Ja g = a Ka b Jb = b Jb ,
or
b = a Ka b .
(7.51)
Here, the Ka b s are the coefficients of the adjoint representation.

Take the functional forms { b } as basis and consider the dual basis {Xa }.
On any functional ,
= || + =
a
|| a
A +
A
a
A
Aa
a
=
D a +
A
a
A
Aa

a
a
A . (7.52)
+
= D
Aa
Aa
On the other hand, this must also be
= Xa () a +
Aa
Aa .
We find in this way that the group generators acting on the functionals are
Xa = D
.
Aa
(7.53)
Also the Euler Form [7.37] can be decomposed:

E = tr (E A ) = tr (E || A + E A ) = tr (E D + E A )
= tr [(D E )]+ tr (E g 1 a g) = tr [(D E )]+ tr (gE g 1 a ),
or
E =: tr [(D E )] + tr(e a ).
(7.54)
We have introduced e = gE g 1 , which stands for the expression appearing

in the equation when small a stands for the potential. As D E = 0
identically, e = 0 is the true equation. It is noteworthy that
G = || G + G = E a Aa + || G Ea
= E a Aa a D = E a Aa a Xa G . (7.55)
601
We can also recognize, by integrating by parts, that

0 = || G = (D a )Ea = || Aa Ea ,
(7.56)
which again says that the equation is orthogonal to .

The group parameters a , in terms of which an element of G is written
as g = exp { a Ta } in some representation generated by {Ta }, become fields
a (x). The canonical Maurer-Cartan 1-forms v = g 1 dg on G are represented
by cofields (x) = g 1 (x) || g(x), 1-Forms on , whose expression is enough to
ensure that satisfies a functional version of the Maurer-Cartan equations,
|| =
(7.57)
|| a = 12 f a bc b c .
(7.58)
or
The components a , or the matrix i j = a (Ja )i j = a f i aj , are alternatively
used when convenient, the same holding for A , F , A , etc.
7.2.3
Gauge conditions
Gauge subsidiary conditions correspond to 1-Forms along . Take, for example, the one dimensional abelian case of electromagnetism, for which the
Maxwell Euler Form is
E = ( F )A
(7.59)
As A = || A + A and || A = for some parameter field , an integration by parts shows that the contribution along vanishes. The Lorenz
gauge condition is specified by the 1-Form
H = ( A ) = A A || A .
(7.60)
The complete Euler Form governing electromagnetism in the Lorenz gauge

is consequently
(7.61)
E = ( F ) A 2 || (A A ).
The Form H is exact only along , so that we cannot say that the transgression T H = 2 || (A A ) is a lagrangian in the usual sense. But [7.61] is an
eloquent expression: the equation goes along the physical space, whereas the
gauge condition lies along the large group.
The above considerations can be transposed without much ado to the
nonabelian case. Putting
Aa = D a + Aa
602
in [7.37], the contribution along the group vanishes again. The Lorenz Form
is now
H = ( Aa )a = Aa (D a [A , ]a )
= Aa || Aa = 2 || (Aa Aa ). (7.62)
Supposing a convenient normalization for the Cartan-Killing metric which
we have been using implicitly, the total Euler Form can be written
E = tr {D F A
|| (A A )}.
(7.63)
We have used, for the sector along , the holonomic (or coordinate)
basis {a }, composed of exact Forms. We could likewise have used a nonholonomic basis. With differential forms, the choice of basis is in general
dictated by the symmetry of the problem.
7.2.4
Gauge anomalies
The expressions for the gauge anomaly are components of 1-Forms along
in the anholonomic basis {a }:
U = Ua a .
(7.64)
|| U = 12 [Ta Ub Tb Ua Uc f c ab ]a b .
(7.65)
Using [7.58], we find
The vanishing of the expression inside the brackets is the usual Wess-Zumino
consistency condition, which in this language becomes simply
|| U = 0.
(7.66)
Again, U must be locally an exact Form, but only along , so that TU is not
a lagrangian.
Notice that, unlike the case of the action, the last equation does not
express the invariance of U under gauge transformations. Only when acting
on 0-Forms does || represent gauge transformations. As seen in Math.8, the
situation is again analogous to differential geometry, where transformations
are represented by Lie derivatives. Let us consider vector fields on , say
entities such as = a Ta or X = X a / a . Transformations on Forms will
be given by the Lie derivatives
LX = iX + iX .
7.3. C CHIRAL FIELDS
603
For 0-Forms, only the last term remains, but for U the first will also contribute. The invariance of a Form W under a transformation whose generator
is represented by a Killing Field X is expressed by LX W = 0. In the case
of an Euler Form coming from a lagrangian, E = S. The commutativity between the Lie derivative and the differential operator leads to LX E = LX S,
a well known result: the invariance of S (LX S = 0) implies the invariance of
E (LX E = 0), but not vice-versa (see Math.8.3.3). The invariance of E only
implies the closeness of LX S, and the equations may have symmetries which
are not in the lagrangian.11
7.2.5
BRST symmetry
A final remark concerning gauge fields: we have already used || Aa = D a .

As Aa is a 0-Form, this measures to first order its change under a group
transformation given by g(x) = exp[(x)] 1 (x). By using that
= g 1 ()g, we can write
|| A = D .
(7.67)
A fermionic field will transform according to || 0 = 0 = g 1 g 0 , or

|| =
(7.68)
Let us repeat equation [7.57],

|| = .
(7.69)
The three last equations express the BRST transformations12 provided

the Maurer-Cartan Form is interpreted as the ghost field,13 and Slavnovs
operator is recognized as || . The use of || to obtain topological results14
is a fine suggestion of the convenience of variational Forms to treat global
properties in functional spaces, although it remains, to our knowledge, the
only such application to the present.
7.3
C Chiral fields
We make now a few comments on pure chiral fields, here understood simply
as the group-valued fields g(x) met above.
11
12
13
14
Okubo 1980.
Stora 1984; Baulieu 1984.
Stora 1984; Leinaas & Olaussen 1982.
Ma
nes 1985.
604
(i) The functional space reduces to , and will coincide with the previous
. Neither G nor are starshaped spaces, so that we must work with
tensor fields on the Lie algebra (which, being a vector space, is starshaped)
and their functional counterparts. The variation of the Maurer-Cartan form
= g 1 g is the covariant derivative of its corresponding Form:
||
= g 1 (g)g 1 g + g 1 (gg 1 g)
= + [ , ] = D . (7.70)
(ii) To obtain the Euler Form corresponding to the 2-derivative contribution to the chiral field dynamics, we start from the usual action
S = 21 tr ( ),
(7.71)
from which
E = S = tr( ) = tr { ( + [ , ])}
= tr { } = tr {( )}
= tr { (g 1 g)g 1 g}. (7.72)
(iii) The existence of a lagrangian here is a consequence of the functional
Maurer-Cartan equation. In effect,
E = ( a )a ] = a a + ( a )a
= ( a + f a bc b c ) a + a )a
= ( a )[a + 21 f a bc b c ] = 0.
The presence of in the trace argument in [7.71] would not be evident
from the field equation
(g 1 g) = 0.
(7.73)
The variation was entirely made in terms of and , which belong to
starshaped spaces, and not in terms of g(x). We can consequently follow the
inverse way: put [7.72] in the form E = tr ( ) and get [7.71] back.
(iv) The above considerations are examples of the power of exterior variational calculus, on which more is said in Math.8.
Trautman l970, l979

Yang 1974
Popov l975
Faddeev & Shatashvilli 1984
Phys. Topic 8
GENERAL RELATIVITY
1 Einsteins equation
2 The equivalence principle
3 Spinors and torsion
Little more than a topical formulary, with emphasis on some formal points.
8.1
Einsteins equation
The geometrical stage set is provided by the bundle of linear frames. This
means that the cast of characters will include linear connections
= a b a b dx ,
their curvatures
F = a b Ra b = 12 a b Ra b dx x
= 21 a b [ a b a b + a c c b a c c b ] dx dx ,
and tensors in general. We shall see, however, that the main actors will be
metrics. It is good to keep in mind the different character of the indices
in a b as in Ra b : the first two are algebraic, as they indicate algebra
components, while the remaining indices are those of tensorial components.
In the curvature form F ,
Ra b = 12 Ra b dx dx
is the component of R along a b .
Consider a manifold with a Riemannian metric. Given an arbitrary linear
connection , the covariant derivative of a metric tensor will have components
D g = g; = g g g
605
(8.1)
606
PHYS. TOPIC 8. GENERAL RELATIVITY
This will vanish when the connection preserves the metric, that is, when
the metric is parallel transported by :
g = ()
(8.2)
Recall that we use the notations () and [] for symmetrized and antisymmetrized indices. From this we find that
g + g g = () + [] + []
(8.3)
Comment 8.1.1 Notice that an orthogonal connection, along generators Jab = ab

ba in the algebra, would not contribute to g .
A theorem by Ricci says that, given a metric g , there exists a unique
linear connection which preserves g and has a fixed T for its torsion. In
particular, there is a unique with zero torsion, the Levi-Civita connection
of the metric (9.4.23). The other connections differ from this privileged one
precisely by their torsions.1
Thus, there exist in principle an infinite number of connections preserving
a given metric, but only one of them has vanishing torsion. The components
of the torsion tensor in a natural basis are
T = ,
so that T
= 0 implies that the connection is symmetric in the lower indices.
In order to see it, recall that one changes algebra and tensor indices with the
tetrads and that, for the algebra indices in a connection,
a b = ha hb + ha hb
Replace this in the torsion component
T a = ha ha + a b hb a b hb
(8.4)
to obtain
T a = ha ha
or
T = ha T a = .
The frequently used symmetry = is thus a consequence of a vanishing torsion.
1
Kobayashi & Nomizu l963.
8.1.
EINSTEINS EQUATION
607
Comment 8.1.2 As a comparison of the bundle of frames (section 9.3) with general
bundles involving internal symmetry groups (section 9.5) shows, having a vanishing
torsion is quite distinct from having no torsion at all.
For a symmetric connection, the last two terms in [8.3] vanish and we
find that the connection is given by the Christoffel symbol
} = 21 g [ g + g g ] .
= {
(8.5)
This is the connection at work in gravitation as described by General Relativity. In Gauge Theories, the connection is the basic field. In General
Relativity, it is a tributary field: it is completely fixed by the metric, which
is thus the true fundamental field.
Comment 8.1.3 On the other hand, the two theories have much in common. The
field is in both cases the curvature. The absence of field is given by the vanishing of
the curvature. There is another point. For fields in general, one has a more physical
criterion to know where the field is: the energy-momentum density being a positive
characteristic, the field is present where the energy-momentum density is different from
zero. The gravitational field has no well-defined (that is, covariant) energy momentum and
thus this characterization fails. But also here there is something in common, because the
current density of a gauge field is not well defined (covariant) either. Energy-momentum
is the source of gravitation but there is no way of defining a (covariant) energy-momentum
for the gravitational field proper. Color (say) is the source of a gauge field, but there is
no covariant characterization of the color density of the gauge field proper.
The vanishing of torsion implies some extra symmetries in the curvature
components. Besides the antisymmetry in the second pair of indices in the
Riemann tensor
Ra b = a b a b + a b a b ,
(8.6)
forcible in a 2-form component, there is also an antisymmetry in the exchange

between an algebraic and a tensor indices. Algebraic and spacetime indices
get mixed up.
The Ricci tensor is a symmetric 2-tensor defined as
R = ha hb Ra b
(8.7)
and the scalar curvature (or scalar invariant) is

R = g R .
(8.8)
To introduce dynamics and arrive at Einsteins equation, the simplest

and surest path is his first, compelling though heuristic, derivation.2 Given
a strictly Riemannian manifold, it turns out that the Einstein tensor
G = R 21 Rg
2
Chandrasekhar 1972.
(8.9)
608
is the only symmetric 2-tensor with vanishing covariant derivative. Concerning the source fields, their symmetrized energy-momentum tensor ,
modified by the presence of g , is the only symmetric 2-tensor with vanishing covariant derivative. The source, in Newtonian gravitation, is the mass,
whose concept is broadened into energy by Special Relativity. Energy, which
in field theory is represented by the energy-momentum tensor, is to be the
source of gravitation. It is thus natural to write G = k , where k is
some constant. Comparing with Newtons law in the static weak-field limit,
one determines k = (8G/c4 ) and the field equation, Einsteins equation,
comes as
.
(8.10)
R 12 R g = 8G
c4
In the absence of sources, contraction of R 21 R g = 0 with g leads to
R = 0, and consequently to
R = 0.
(8.11)
This equation can be obtained from the Einstein-Hilbert action
Z
S[g] = d4 x g R,
(8.12)
whose variation is
Z
Z

S[g] = d x g g
R + d4 x g g R
Z
Z

4
1
d x g R 2 g R g + d4 x [total divergence]. (8.13)
Comment 8.1.4 A difference with respect to gauge theories turns up here: unlike
that of gauge fields, action [8.12] is linear in the curvature. Dynamics is in consequence
quite different.
8.2
The equivalence principle
Let us now concern ourselves with the coupling of gravitation to other fields.
Despite well known qualms,3 we shall choose a nave course, whose main
advantage is that of being short.
The manifold metric g relates to the flat Lorentz tangent metric through
the tetrad fields,
g = ab ha hb .
(8.14)
Suppose for a moment that the ha s are trivial four-legs, mere coordinate
choices. We can calculate the corresponding Christoffel and curvature. We
3
See the Preface of Synge 1960.
8.2. THE EQUIVALENCE PRINCIPLE
609
a
find then that Rb
= 0. This is a matter of course, as trivial tetrads will
only lead to other representations, in terms of non-cartesian coordinates,
of the flat Lorentz metric. We pass from the tangent metric to another,
Riemannian metric only through a non-trivial tetrad field. This leads to
a rule, which we shall call equivalence principle. To obtain the effect of
gravitation on sources in general (particles or fields), (i) write all the usual
equations they obey in Minkowski space in general coordinates, represented
by trivial tetrads, and (ii) keep the same formulae, but with the trivial tetrads
replaced by general tetrads, related to the metric by [8.14]. The presence of
tetrads enforces also the passage of simple derivatives to covariant derivatives
with the frames Cartan connection, so that usual derivatives are replaced
by covariant ones. The resulting equation holds in General Relativity. This
is reminiscent of the gauge prescription (Phys.7).
Comment 8.2.1 Of course, the equivalence principle takes its roots in the inverse
reasoning: when gravitation becomes progressively weaker, the equations for the source
fields approach those valid in Special Relativity (Phys.6). There is a particular system of
(normal) coordinates in which = 0.
Comment 8.2.2 Another difference with respect to gauge theories lies in the attribution of particles to the group multiplets. In gauge theories, source particles are placed
in convenient multiplets, the attribution being based on phenomenological grounds. A
particle which is insensitive to a certain gauge field is supposed to be in a singlet representation of the gauge group. The linear group (and its subgroups, Lorentz and Poincare)
acting on spacetime can always act upon fields (x) via the regular representation, which
changes the very arguments (points of spacetime) of the field (Math.6). Thus, every field
feels gravitation through this representation, a property that goes under the name of
universality . Fields endowed with spin will transform according to a direct product of
this representation and that (vector, spinor, etc) representation related to spin.
R
The total action will be [8.12] plus d4 x g S , the action of the remaining source fields, modified as indicated. The resulting Euler-Lagrange
equations will be (i) for the gravitational field, Einsteins equations, to which
the other fields provide the source in the form of their modified energymomentum density tensor, and (ii) for the source fields, simply their modified
free equations. As an example, a free scalar field, liege to the Klein-Gordon
equation,
t (x) + m2 (x) = 0,
u
(8.15)
will obey that same equation formally, but with the dAlembertian replaced
by the four-dimensional Laplace-Beltrami operator
t =
u
1
g

g g .
(8.16)
610
The laplacian is a second order operator, and here all derivatives should
be covariant. It should be
t = D D .
u
(8.17)
And indeed it is. But we should notice that the first covariant derivative,
hitting on the scalar , coincides with the simple derivative. And the second
one, hitting on the vector , will have the Cristoffel coupled with the vector
representation (Phys.6), leading to the above expression.
Non-trivial geometry affects the scalar field only through the presence of
the metric in the dAlembertian operator. We shall see below that spinor
fields probe deeper into the geometry. They are sensitive to the tetrads and,
through their spin, to a part of the connection.
The source energy-momentum current density is
=
S
.
g
(8.18)
It differs from the canonical energy-momentum. As energy-momentum is the

Noether current related to translational invariance (Phys.6), it is meaningless
in the usual overall Riemannian point of view. The source term is actually
not the energy-momentum, but its curvature-modified version, a tensor which
tends to it in the zero curvature limit. We have seen how such a modification
comes out, by taking non-trivial four-legs and covariant differentials. Furthermore, is symmetric. Thus, the source is a symmetrized version of the
modified energy momentum.
Comment 8.2.3 There is one further question concerning the energy-momentum as
a current: it has not the usual current-density dimension and consequently, the constant
k must have a compensating dimension. As a coupling constant, k appears also in the
self-interactions of the gravitational field (included in the term tg n eq.[8.19] below). The
problems of renormalizability with a non-dimensionless coupling constant are well known
and will appear in any theory with energy-momentum as a source. This is the fulcrum of
the short distance (or quantization) problems of General Relativity.
Comment 8.2.4 Although S /g may be used as an energy-momentum for source

fields, it cannot be applied to the gravitational field itself.4 Something analogous happens
for gauge fields, where the self-current is not /A (see Phys.6.7). Similarly to that
case, if we try to obtain the energy-momentum of the gravitational field as S/g , we
find
S [g]
g
4
Fock 1964.
= g R
1
2
g R
8.2. THE EQUIVALENCE PRINCIPLE
611
and the total current density vanishes. Einsteins equations [8.10] may, however, be put
into a form5 analogous to eq.[6.50] of Phys.6, that is,
+ g t = g ,
(8.19)
where is antisymmetric in and , and t represents the energy-momentum of the
gravitational field. The total current equals , so that

(8.20)
g( + t ) = 0.
The quantity t is a pseudo-tensor, not covariant under general frame transformations,
just as the self-current j a nu of the gauge fields (see Phys.6.7) is not gauge covariant. Nevertheless, t is asymptotically a tensor: at large distances, it becomes a tensor under linear
coordinate transformations which mimic isometries of Minkowski space, transformations
of the Poincare group.6 Again, this is in complete analogy with the gauge self-current,
which becomes covariant only at large distances, when the gauge transformations must
become global.
Up to this point, we have seen the standard case, which works when no
fields with spin higher than zero is present.
In the light of the geometric vision acquired by all our previous discussion,
we can indulge in some instructive reflection. As repeatedly stated, there is
no curvature of space. Curvature is a property of a connection, and a great
many connections may be defined on the same space. Take an electron on a
Riemannian spacetime. It responds to the action of the Levi-Civita connection given by the Riemannian metric. Now add an electromagnetic field. The
electron will now answer to the appeal of two connections, the previous one
and the electromagnetic potential. Add further a neutrino: it will feel (probably) the Levi-Civita connection, but not the electromagnetic potential. As
long as it stays far from the electron, there will be no manifestation of the
weak-force connection. Thus, different particles feel different connections,
different curvatures, and will consequently show distinctly curved trajectories to our euclidean eyes. If we included connections in the very definition of
space, the electron and the neutrino would live in different spaces. Now, there
is a point for taking the Levi-Civita connection of spacetime as part of its
definition, as universality of gravitation would imply that all particles would
feel it the same. There are two reasons for not doing this. First, universality
is as yet a pious postulate, based more on simplicity requirements than on
experimental evidence. Second, there is theoretical evidence that spinning
particles feel torsion, that is, that they deviate from the purely metric behaviour.7 As it is, it seems far wiser to take space simply as a manifold, and
connection (with its curvature) as an additional structure.
5
6
7
Landau & Lifshitz 1975.

For an assessment of the requirements of asymptotic flatness, see Faddeev 1982.
Hehl, von der Heyde, Kerlick & Nester 1976.
612
8.3
Spinors and torsion
A linear connection exhibits torsion in the general case.8 We can always

decompose in symmetric and antisymmetric parts,
= 12 () + 12 [] = 21 () 12 T
(8.21)
Comment 8.3.1 There is another, frequently used, decomposition. Indicate the Levi
Civita connection (that is, the Christoffel) of a given metric by . Then, one writes
= + K ,
where K is called the contorsion tensor (the difference between two connections is
always a tensor). The two decompositions do not coincide, as the contorsion can have a
symmetric part. Actually, if also preserves the metric, one can find that
K =
1
2
(T + T T ) ,
so that the torsion is (minus) the antisymmetric part of K.
Comment 8.3.2 Given a linear connection , then each expression of the form
(t) = t + (1 t) ,
for t [0, 1], defines a linear connection. The particular case t = (1/2) has vanishing
torsion. The torsion of is the difference between and (1/2) . As only the symmetric part
appears in the geodesic equation (Math.12), torsion does not contribute. As (t) () =
() , all these connections have the same geodesics.
Now we come to the important point. The equivalence principle is based

on the fact that there is always a local coordinate system (the normal coordinates) in which the symmetric part of the connection vanishes. In that
coordinate system, the equations recover their special-relativistic form. But
this is not true of the whole connection. Torsion being a tensor, it cannot vanish in a particular frame without vanishing in all frames. Thus, the
presence of torsion induces a violation of the equivalence principle.
Comment 8.3.3 This leads to still another difference with gauge theories. There is
no choice of gauge doing such a job for a general connection, so that nothing similar to an
equivalence principle can exist for gauge theories.
8
For a recent discussion on deviations from the geodesic behaviour of particles, see
Yee & Bander 1993.
8.3. SPINORS AND TORSION
613
Spinors (see Phys.6) have interesting properties, which make of them ideal
detectors of torsion. Their treatment requires the explicit use of the four-legs
and their spin couples to torsion. Let us briefly examine what happens to a
Dirac spinor in a background space endowed with curvature and torsion.9
In the presence of a (external) curvature, spinors respond no more to
the usual derivative, but to the Fock-Ivanenko derivative, which takes into
account the spin coupling:

(8.22)
D = iha a 4i ab ab
where the 4 4 matrices Jab = 21 ab generate the bispinor representation.10
The lagrangian is

= 2i ha a + bc a bc ( ) a bc bc a m
(8.23)
and leads to the Dirac equation in the presence of a connection,

ha a 4i bc bc m = 0.
(8.24)
If we recall the expressions for the energy-momentum density (eq. Phys.6.30)

= 2i nu [ ] .
(8.25)
and for the spin density current
S ab = 14 [ ab + ab ] ,
(8.26)
the lagrangian assumes the form

= ha a + 21 ab S ab
(8.27)
The matrices have the property = i , from which we find

that
Sab = 14 { , ab } = Sba .
We see thus that the spin current couples to the connection antisymmetric
part [] , that is, to the torsion. As spinors also require the explicit use of
tetrads, they indeed see more of the geometric details.
Comment 8.3.4 Vector fields will also perceive the connection, as their covariant
derivative will be analogous to [8.22], though with the vector representation matrices
given by eq.[6.8] of Phys.6 instead of ab .
9
Dirac 1958.
Bjorken & Drell 1964, to be consulted also on the little bit of gammalogy used
below.
10
614
Comment 8.3.5 Recall (9.4.14) that torsion does not affect geodesics (though the
symmetric part of contorsion does), but breaks infinitesimal parallelograms.
Of course, it is not in [8.25] which is the source in General Relativity,

but modified by the presence of the gravitational field. Some use of the
gamma matrix properties shows that this modified version encompasses the
spin current. It is
0 = + 12 ab S ab
just what is necessary to have = ha 0a . The spin-torsion coupling is
hidden in the modified energy-momentum.
Detailed calculations in General Relativity can be rather fastidious, although nowadays most of it can be done on algebraic computer resources.
Anyhow, it is essential to have done them by itself at least once and in some
special case. The basic calculations are shown in Phys.9 for the simplest
possible case, that of spaces of constant curvature.
Weinberg 1972
Fock 1964
Misner, Thorne & Wheeler 1973
Synge 1960
Phys. Topic 9
DE SITTER SPACES
1
2
3
4
5
General characteristics
Curvature
Some qualitative aspects
Wigner-In
on
u contraction
For once, though in the specially simple case of constant curvature, we
present some detailed calculations on Riemannian spaces. As de Sitter spaces
and their groups of isometries (the de Sitter groups) are related by WignerInon
u contraction to the Minkowski space and and its group of isometries
(the Poincare group), we profit to give also an example of the contraction
procedure.
9.1
General characteristics
A Riemannian space is said to be of constant curvature when its scalar curvature R = g R is a constant. Four-dimensional constant-curvature spaces
have Riemann tensor components always given locally by
R =
1
12
(g g g g ) .
(9.1)
The space is of positive or negative curvature respectively if R > 0 or

R < 0. The 4-dimensional spaces of constant positive curvature are the
sphere S 4 and the de Sitter spacetime. There exists a unique kind of 4dimensional manifold with constant negative curvature, the anti-de Sitter
spacetime.1 In between, with R = 0, stands the flat Minkowski space. We
shall study 4-dimensional spaces of constant curvature R 6= 0 here because
they are the simplest non-trivial spacetimes.
1
We adopt here the terminology of Hawking & Ellis 1973, though not their sign
conventions. Notice in particular that the sign of R depends on the conventions used.
615
616
PHYS. TOPIC 9. DE SITTER SPACES
Figure 9.1: Stereographic coordinates for the 4-sphere: consider the northpole N = (0, 0, 0, 0, +L) E5 and the euclidean space E4 tangent to the
sphere at the south pole S = (0, 0, 0, 0, L). Given a point S 4 , draw a
straight line from N through . The coordinate mapping will take into the
point x E4 at which the straight line intersects E4 . The coordinates of
are then the cartesian coordinates of x.
Comment 9.1.1 The invariance of Minkowski spacetime M under the transformations of the Poincare group P reflects its uniformity. P is the group of motions of M
(6.6.14), with the maximal possible number of Killing vectors, which is ten for a 4dimensional space. Notice that the Lorentz subgroup provides an isotropy around a given
point of M , and the invariance under translations enforces this isotropy around any other
point. This is the meaning of uniformity: all the points of spacetime are ultimately
equivalent. Amongst curved spacetimes, only those of constant curvature can lodge the
highest number of Killing vectors. Given the metric signature and the value of R, the
maximally-symmetric torsionless space is unique.2 General Relativity does not consider
torsionned spaces and, in its picture, the de Sitter spaces are the only uniform curved
spacetimes.
A de Sitter space [which we call DS(4, 1) for reasons given below] may
be seen as an inclusion in E4,1 of the hypersurface whose points ( 1 , 2 , 3 ,
4 , 5 ) satisfy
( 1 )2 + ( 2 )2 + ( 3 )2 ( 4 )2 + ( 5 )2 = L2 ,
(9.2)
with the induced topology and the metric induced through the inclusion by
the pseudo-euclidean metric of E4,1 . This space is homeomorphic to S 3 E1
2
Weinberg 1972.
9.1. GENERAL CHARACTERISTICS
617
Figure 9.2: The anti-de Sitter space. Points of the upper branch correspond
to points outside the circle. Points of the lower branch, to points inside.
and its group of motions is the pseudo-orthogonal group SO(4, 1).3 An antide Sitter space DS(3, 2) may be seen as an inclusion in E3,2 , the manifold
whose points satisfy
( 1 )2 + ( 2 )2 + ( 3 )2 ( 4 )2 ( 5 )2 = 1,
(9.3)
again with the induced topology and metric. The space is now homeomorphic
to S 1 E3 and its group of motions is the pseudo-orthogonal group SO(3, 2).
Both SO(4, 1) and SO(3, 2) are called de Sitter groups and each contains the
Lorentz group SO(3, 1) as a subgroup.
There are closed timelike geodesics in the anti-de Sitter space DS(3, 2).
Actually, DS(3, 2) has a fascinating property, the wiedersehen faculty
(Phys.5.5): all timelike geodesics passing through a point will converge
unanimously to another point, and then to a third one, und so weiter. Its
3
For a discussion of DS(4, 1), see Schmidt 1993.
618
universal covering is obtained by simply unwrapping the circle. This covering, which no more supports causal bizarreries, is then simply the topological
space E4 . Some authors reserve the name anti-de Sitter space to this covering space.
The metric is, in both cases, simply the Lorentz metric multiplied by a
point function. Metrics differing by the product by a function are conformally
equivalent, meaning that all angle measures are the same in both cases. A
space which is equivalent in this way to a flat space is called conformally flat.
The de Sitter spaces are conformally flat.
Consider in E5 the hypersphere S 4 given in Cartesian coordinates { a }, a =
1, 2, . . . , 5 by
( 1 )2 + ( 2 )2 + ( 3 )2 + ( 4 )2 + ( 5 )2 = L2 .
We can project it stereographically from the point 5 = +L (north
pole) into the hyperplane E4 tangent at the point 5 = L (south pole).
This will provide every point of the hypersphere (except the north pole) with
coordinates (, = 1, . . . , 4)
x =
2
1 5 /L
(9.4)
on E4 . Notice that this is a direct adaptation of a Riemannian metric on an

euclidean space (see what is done in Math.11 in 2 dimensions). Introducing
2 = x x , with the four-dimensional euclidean metric, and calculating
the line element ds2 = da d a in these stereographic coordinates, we find
ds2 = g x dx , where the new metric is
g = n2 (p) ,
(9.5)
n = 21 (1 5 /L)
(9.6)
n(p) being the function

or, in terms of 2 ,
1
.
(9.7)
1 + 2 /4L2
q
5
The scheme of Figure 9.1 shows = L2 ~2 ( 4 )2 .
Consider now, instead of a sphere, a hyperbolic hyperspace in E3,2 , given
by
( 1 )2 + ( 2 )2 + ( 3 )2 ( 4 )2 ( 5 )2 = L2 .
n=
The points 5 = +L and 5 = L are now respectively the lowest point of

the upper sheet and the highest point of the lower sheet of the hyperbolic
space (see Figure 2). The stereographic projection, given again by
= n(p) x
9.2. CURVATURE
619
leads to the metric

g = n2 (p) ,
(9.8)
with n(p) given by (9.6), which now assumes the form

n=
1
,
2
1 4L
2
(9.9)
with 2 = x x , and with the Lorentz metric being = diag(1, 1, 1, 1).

This is the space
q DS(3,2). Figure 9.2 shows the possible values of the coor5
dinate = L2 + ~2 ( 4 )2 . The same projection from the hyperbolic
space fixed by
( 1 )2 + ( 2 )2 + ( 3 )2 ( 4 )2 + ( 5 )2 = L2
would lead to the space DS(4, 1).
9.2
Curvature
We can easily obtain the Riemann curvature for the spherical and the hyperbolic cases. To treat all of them simultaneously, we write
n=
1
,
1 + s 2 /4L2
(9.10)
the sign s = 55 = (+1) referring to the spherical and DS(4,1) cases, the sign
s = 55 = (1) to the DS(3,2) case. The notation will be used for both
the euclidean and the Lorentz metric. Noticing that
n =
s 2
n x ,
2
2L
(9.11)
the Christoffel symbols are found to be (see Math.11)

= s
n
[ x x x ].
2L2
(9.12)
We then calculate
n
[ ]
L2

n2
4 x x ( ) + x x
4L
= s
620
and
=
n2
{x x ( ) + 2 ( ) + x x ( )}.
4
4L
Using again (9.7) and (9.9), the Riemann tensor components
R = +
are found to be
R = s
1
[ g g ].
L2
(9.13)
The Ricci tensor is consequently

R = s
3
g
L2
(9.14)
and the scalar curvature is, as expected if we compare (9.13) and (9.1), the
constant
12
R = s 2.
(9.15)
L
These results are, up to the numerical factors, the same for spacetimes of
any dimension.
9.3
The equation for the geodesics is

d2 x
s
dx dx
+
= 0.
x 2x
ds2
2nL2
ds ds
(9.16)
We can as easily obtain the Jacobi equation, which is

s
D2 X
+ 2 X (X V )V = 0,
2
Ds
L
or
(9.17)
D2 X
s
+ 2 [X g(X, V )V ] = 0.
(9.18)
2
Ds
L
Now, the expression X = [X g(X, V )V ] represents the component of X
D2
||
transverse to the curve. As the tangential part X || satisfies Ds
= 0, we
2X
arrive at
D2 X
s
+ 2 X = 0.
(9.19)
2
Ds
L
9.4. SOME QUALITATIVE ASPECTS
621
Formally, this equation is of harmonic oscillator type, which hints at periodic solutions and to the above mentioned focusing property. The transverse
field simply oscillates around the curve. This phenomenon of perfect focusing is not quite surprising if we recall how similar this space is to a higher
dimensional replica of that favorite chimaera of optics, the perfectly focusing
Maxwells fish-eye (Phys.5.5).
9.4
Some qualitative aspects
The spheres are homogeneous spaces, that is, quotients S n = SO(n+1)/SO(n)

of two Lie groups. In particular, the hypersphere S 4 is S 4 = SO(5)/SO(4).
Actually, SO(5) is isomorphic to the bundle of orthogonal frames on S 4 .
Our de Sitter spaces are homogeneous spaces with the Lorentz group as stability subgroup, respectively DS(4, 1) = SO(4, 1)/SO(3, 1) and DS(3, 2) =
SO(3, 2)/SO(3, 1). The de Sitter groups are the respective bundles of (pseudo)orthogonal frames. This is an amazing fact indeed: the set of all Lorentz
frames on DS(3, 2), for example, is just SO(3, 2). The homogeneous character creates in this way a natural difference between Lorentz transformations
and the remaining ones. We might call the latter de Sitter translations.
By the process of Wigner-Inon
u group contraction, both de Sitter groups reduce to the Poincare group, both de Sitter spacetimes reduce to Minkowski
spacetime, and de Sitter translations reduce to the usual translations in space
and time. Even without contraction, there is a deeper characteristic which
makes de Sitter translations quite distinct of the Lorentz transformations.
The de Sitter groups, algebras and spaces are symmetric (8.2.7). There
is an involution, a mapping : group group, with 2 = 1, leaving the
Lorentz subgroup invariant, but changing the de Sitter translations. According to a theorem by Wang, quotient spaces have a very special connection,
invariant under the group action. As a consequence, there exists a special
canonical connection defined on the de Sitter spaces, with very particular
characteristics.4
9.5
Wigner-In
on
u contraction
We recall that this contraction is a general procedure, by which some groups

are deformed into others by taking the asymptotic values of convenient parameters. The standard example is the deformation of the Poincare group
into the Galilei group when the velocity of light is taken to infinity. We shall
4
Kobayashi & Nomizu l963.
622
here only say a few words on the de Sitter-Poincare contraction. It is easier to

see the procedure in the Lie algebra. It can be summarized as follows. Suppose we have on a (n+1)-dimensional euclidean space the symmetric bilinear
form
(X, Y ) = ab X a Y b ; a, b = 1, 2, . . . , n.
We may diagonalize it, then define convenient Cartesian coordinates, establishing a basis in which all eigenvalues are either +1 or 1. In this case,
the squared length of a vector will be (, ) = ab a b . Orthogonal and
pseudo-orthogonal groups SO(p, n p) are defined by transformations which
preserve such kinds of bilinear forms, with p the number of positive eigenvalues. The above cases of de Sitter spaces fall in this case. The generators
{Jab } of the Lie algebra of SO(p, n p) will then satisfy the commutation
rules
[Jab , Jcd ] = bc Jad + ad Jbc bd Jac ac Jbd .
(9.20)
For the case n = 5 and p = 1 or 2, we define
= L1 J 5 ,
(9.21)
= P 55 (4L2 )1 K
(9.22)
with
x
where P and K = [ 2 2xn2 ]P are respectively the generators of trans()

lations and special conformal transformations. For 55 = +1, and L
are the generators of the de Sitter group SO(4, 1). For 55 = -1, (+) and
L are the generators of the de Sitter group SO(3, 2). In terms of these
generators, the de Sitter algebra becomes
[J , J ] = J + J J J ;
(9.23)
[ , J ] = ;
(9.24)
[ , ] = 55 L J .
(9.25)
lim = P
(9.26)
In the limit L ,
L
and the de Sitter algebra contracts to the usual Poincare algebra of the
generators J and P . Of course, if we look at (9.9), (9.12) and (9.13), we
see that this limit leads exactly to the Minkowsky geometry. This is the
usual Wigner-Inon
u contraction.
Let us now consider another possibility by taking the opposite limit, that
is, L 0. In this case,
(9.27)
lim = 41 55 K
L0
U
CONTRACTION
9.5. WIGNER-INON
623
and the de Sitter algebra contracts to the algebra given by eq.(9.23) and
[K , J ] = K K
(9.28)
[K , K ] = 0.
(9.29)
We see in this way that the Lie group formed by the Lorentz (J ) and the
special conformal generators (K ) has the same Lie algebra as the Poincare
group.
Summing up: by the process of Wigner-Inon
u group contraction with
L , both de Sitter groups reduce to the Poincare group, both de Sitter
spacetimes reduce to Minkowski spacetime, and the de Sitter translations
reduce to the ordinary translations in space and time. In a similar procedure, but considering the limit L 0, the de Sitter translations reduce
to the special conformal transformations, and the resulting group, despite
presenting the same algebra, is deeply different from the Poincare group. We
can conjecture in this way that, if somehow the de Sitter group appears as
the symmetry group of a physical theory, the Poincare group generated by
J and P would be related to the weak field limit (L , R 0) of
this theory, while the Poincare-like group generated by J and K would be
related to the strong field limit (L 0, R ) of the theory.
Eisenhart 1949
G
ursey 1962
Hawking & Ellis 1973
624
Phys. Topic 10
SYMMETRIES ON PHASE SPACE
1
2
3
4
5
Symmetries and anomalies

The Souriau momentum
The Kirillov form
Integrability revisited
Classical Yang-Baxter equation
The study of the action of symmetry groups on phase space is an opportunity
to introduce some topics of contemporary research: non-linear representations, cohomology of Lie algebras, anomalies, etc. It is, however, a theme
of fundamental importance by itself. It leads to a partial but significant
classification of phase spaces and opens a road to the general problem of
quantization and its relationship to representation theory through the orbit
method.
10.1
Symmetries and anomalies
Suppose that a group G acts transitively on the phase space M (so that M is
homogeneous under G), in such a way that its transformations are canonical.
This means that G will act through a representation in a subgroup of the
huge group of canonical transformations. In this case, M is said to be a
symplectic homogeneous manifold. The generators Ja of the Lie algebra G0
of G will have commutation relations
[Ja , Jb ] = f c ab Jc .
(10.1)
Each Ja will be represented on M by a fundamental field, a hamiltonian field

Xa . There will be a representation of G0 by vector fields Xa = (Ja ). We
would expect that the representative fields Xa satisfy the same relations,
[Xa , Xb ] = f c ab Xc .
(10.2)
In this case, the algebra representation is said to be linear. But this is not
what happens in general. Usually, the actions of groups on manifolds are
625
626
PHYS. TOPIC 10.
typically non-linear (non-linear representations are even sometimes defined

as these actions). The very use of the word representation is an abuse (to
which we shall nevertheless stick for the sake of simplicity), as is no true
homomorphism: the word action, less stringent, would be more correct.
A special case occurring with some frequency comes out when the action is
given by
(Ja ) = Xa a ,
(10.3)
where the a s are functions. Recall that a field like Xa acts on functions defined on M , producing other functions. Equation [10.3] says that the action
of each generator of G has an extra contribution which can be accounted for
through multiplication by a function. A dynamical quantity F will change
according to (Ja )F = Xa F a F . As to the function a , it can be interpreted as the result of the action of some form on Xa : a = (Xa ). This
situation corresponds to the minimum departure from the simplest expected
case [10.2]. It comes immediately that
[(Ja ), (Jb )] = [Xa , Xb ] Xa [b ] + Xb [a ].
(10.4)
This kind of action, in which each generator is represented by a field action

plus multiplication by a function, has a nice property: the whole group action
can work in that way, because the representative Lie algebra is closed in this
kind of action: as seen in [10.4], the commutator of two such actions is also a
field plus a multiplicative function. The action is a projective representation
of the algebra G0 when [10.4] can be rewritten as
[(Ja ), (Jb )] = f c ab (Jc ) Kab ,
(10.5)
where the Kab s constitute an antisymmetric set of functions, which can be

taken as the components of a 2-form K: Kab = K(Xa , Xb ). We put it in this
way because [10.5] is just the textbook definition of a projective representation. That it represents the slightest departure from linear representations
may be seen by a simple cohomological reasoning. To begin with, we can
impose the Jacobi identity on the commutator, to ensure the Lie algebra
character. We find easily that the condition is equivalent to dK = 0 in the
subspace generated by the Xa s. General representations satisfying [10.5] will
require that K be a cocycle. We might ask to which one of the cohomology
classes a projective representation given by [10.3] would belong. As
[(Ja ), (Jb )] = f c ab (Jc ) {Xa [b ] + Xb [a ] f c ab c },
K will be
Kab = Xa [b ] + Xb [a ] f c ab c .
(10.6)
10.1. SYMMETRIES AND ANOMALIES
627
Always in the subspace of the Xa s, this expression says that the 2-form K
of components Kab is the (exterior) differential of the 1-form of components
c . The condition reduces to K = d, a coboundary. The cohomology class
of K is trivial for representations like [10.3]. From [10.2], [10.5] and [10.6],
the linear case is seen to require
Xa [b ] + Xb [a ] f c ab c = 0.
(10.7)
This is just d = 0 written in components. Projective representations reduce

to linear representations when also is a cocycle. It is in this sense that they
represent a minimal departure from linearity.
By the way, we may give this discussion a contemporary flavor by calling
anomalies both Kab in [10.5] and a in [10.3], as they represent an expectation failure analogous to that coming out in the quantization processes,1 for
which this terminology has been introduced. The expression [10.7] is quite
analogous to the Wess-Zumino condition and typical of anomaly removal
(Phys.7.2.4).
The above kind of procedure is typical of the applications of cohomology
to representations. In particular, it is an example of the so called cohomology of Lie algebra representations.
Unfortunately, even when the representation (G0 ) is linear, further anomalies insist in showing up. Suppose the representative fields to be all strictly
hamiltonian (Phys.1), so that to each Xa corresponds a function Fa (called
in the present case the hamiltonian related to Ja ) such that
iXa = dFa .
(10.8)
Fa is the generating function of the canonical transformation whose infinitesimal generator is Xa . But this means that there is still another representation at work here, that of the algebra of hamiltonian fields in the
algebra of differentiable functions on M , with the Poisson bracket as algebra
operation. With a simplified notation, it is the homomorphism : (G0 )
C (M, R), : Xa Fa , with the corresponding operations [, ] {}. From
what is seen in (Phys.1.7),
dF[Xa,Xb] (Y ) = i[Xa,Xb] (Y )
for any field Y . The last expression is the same as
([Xa , Xb ], Y ) = f c ab (Xc , Y ),
1
1984.
There is more than a mere analogy here. See for instance Faddeev & Shatashvilli
628
PHYS. TOPIC 10.
so that dF[Xa,Xb] = f c ab dFc . As also

d{Fa , Fb } = dF[Xa,Xb] ,
we have d{Fa , Fb } = f c ab dFc , from which
{Fa , Fb } = f c ab Fc + (Xa , Xb ).
(10.9)
The presence of the constant (Xa , Xb ), which comes out from applying
a 2-form to the two fields Xa and Xb , says that, in principle, also is a
projective representation. This is related to the fact that generating functions
are defined up to constants. We may proceed in a way quite analogous to the
previous case. The Jacobi identity applied to [10.9] will say that the 2-form
is a cocycle. Let us add a constant to each of the above functions, and
consider the modified functions: Fa0 = Fa + a , Fb0 = Fb + b , etc. The relation
becomes
{Fa0 , Fb0 } = f c ab dFc0 + 0 (Xa , Xb ),
with 0 (Xa , Xb ) = (Xa , Xb ) c f c ab . The constants a may be seen as the
result of applying an invariant 1-form to the respective fields, a = (Xa ),
etc. As (d)ab = c f c ab , we see that 0 = + d. If some exists
whose choice leads to 0 = 0, showing as the exact form = d, then
the functions may be displaced by arbitrary constants so that the algebra
reduces to {Fa , Fb } = f c ab Fc . The projective representation reduces to a
linear representation when the cohomology class of is trivial. In this case,
we have {Fa , Fb } = F[Xa,Xb] , and the symplectic manifold M is said to be
strictly homogeneous (or Poisson) under the action of G.
10.2
The Souriau momentum
When the above Fa s exist, we can also consider directly the composite mapping F : G0 C (M, R given by , such that F (Ja ) = Fa (x). We
are supposing that G is a symmetry group of the system. The transformations generated by its representative generators will preserve the hamiltonian
function H. As
{Fa (x), H} = Xa H = LXa H,
the invariance of H under the transformations whose generating function is
Fa , LXa H = 0 gives just the usual {Fa (x), H} = 0. Each Fa is a constant
of motion. This is the hamiltonian version of Noethers theorem (Phys.6.6).
Each symmetry yields a conserved quantity.
Let us place ourselves in the particular case in which the Liouville form
is also preserved: LXa = 0. Then,
10.3. THE KIRILLOV FORM
629
dFa = iXa = iXa d = (LXa diXa ) = d[iXa ]

and, with the generating functions defined up to constants,
Fa (x) = [iXa ](x) = [(Xa )](x).
(10.10)
The composite mapping F , such that F (Ja ) = Fa (x), can be realized as a

cofield on G. Take the Maurer-Cartan basis { a } for G0 and define
P : M G0 ,
P : x P (x) = Px = Fa (x) a .
(10.11)
(10.12)
The hamiltonians are, of course, Fa (x) = Px (Ja ). The mapping P is the

Souriau momentum and is defined up to an arbitrary constant in G0 . Notice
that its existence presupposes the globally hamiltonian character of the Xa s.
Given the action of G on M , it provides the constants of motion related to
its generators. There is more; one can show that:
(i) the mapping P commutes with the group action, so that P (x) is a
G-orbit in G0 ;
(ii) P is a local homeomorphism of M into one of the orbits of G in G0 ;
this will have a beautiful consequence.
10.3
The Kirillov form
Consider an n-dimensional Lie group G acting on itself. This action is a

diffeomorphism and consequently fields and cofields on G will be preserved,
that is, taken into themselves. A set of n left-invariant fields Ja may be
taken as a basis for the Lie algebra G0 . Such a basis will be preserved and
will keep the same commutation relations [Ja , Jb ] = f c ab Jc at any point of
G, so that the f c ab s will be constant. G acts on the Ja s according to the
adjoint representation g 1 Ja g = Ka b Jb . The dual basis to {Ja } is formed by
the Maurer-Cartan 1-forms c such that c (Ja ) = ac and which satisfy
d c = 12 f c ab a b .
The group G acts on the c s according to the coadjoint representation
g 1 b g = Ka b a . Now, each 1-form = a a on G defines a 2-form
the Kirillov form) by the relation (Ja , Jb ) = [Ja , Jb ]) = c f c ab . That is,
= 21 c f c ab a b = c d c .
(10.13)
This form is closed, nondegenerate and G invariant. It defines a symplectic

structure. As is preserved by the group action, the same is defined along
all its orbit,
630
PHYS. TOPIC 10.
Orb () = {Ad*g , all g G},

by G in the coadjoint representation. So, on each such orbit (usually called
coorbit) there is a symplectic structure, which is furthermore strictly homogeneous. The important point is the following: orbits in the coadjoint
representation of Lie groups are, in reality, the only symplectic strictly homogeneous manifolds. Any other strictly homogeneous manifold is locally
homeomorphic to one of these orbits and, consequently, is a covering of it.
The homeomorphism is precisely the Souriau mapping P . In this way, such
orbits classify all symplectic strictly homogeneous manifolds.
10.4
Integrability revisited
Consider2 on a phase space two functions L(q, p) and M (q, p) with values in
some Lie algebra G0 of a Lie group G:
L = Ja La (q, p) ; M = Ja M a (q, p).
(10.14)
They are said to constitute a Lax pair if the evolution equation for L is
d
L = [L, M ] = Ja f a bc Lb M c .
dt
(10.15)
If now we take for the hamiltonian M = g 1 dtd g, the solution of this equation is
L(t) = g 1 (t)L(0)g(t).
(10.16)
Recognizing the action of the adjoint representation, we can write this also
as
L(t) = Adg1 (t) L(0).
(10.17)
The evolution is governed by the adjoint action. Consider now any polynomial I(L) of L which is invariant under the adjoint representation. It will
not change its form under the group action, and therefore
d
I(L) = 0.
dt
(10.18)
Lax pairs provide consequently a very convenient means to find integrable

systems. To obtain integrals of motion, one chooses a representation in which
L and M are well known matrices and check candidate invariants of the type
Ij = tr (Lj ), which are adjoint-invariant, verifying whether or not they are
2
Babelon & Viallet 1989.
10.5. CLASSICAL YANG-BAXTER EQUATION
631
in involution. The secular equation fixing the eigenvalue spectrum of L,

det(L I) = 0, is a polynomial in with coefficients which are themselves
polynomials in the traces of powers of L. The eigenvalues are thus also
adjoint-invariant. The evolution equation dL
= [L, M ] is for this reason
dt
called an isospectral evolution.
10.5
Classical Yang-Baxter equation
The classical Yang-Baxter equation is the Jacobi identity for the Poisson
bracket for phase spaces defined on Lie groups, written in terms of the inverse
to the symplectic matrix. Let us see how it comes out.3 The Poisson bracket
of two functions F and G is related to the symplectic cocycle by
{F, G} = (XF , XG ) = ek (G)kj ej (F ),
(10.19)
where XF and XG are the hamiltonian fields corresponding to the functions,

{ek } is a vector basis and the matrix (ij ) is inverse to the matrix (ij )
formed with the components ij = (ei , ej ) of the symplectic form. The
requirement that be a closed form, or a cocycle, is equivalent to the Jacobi
identity for the Poisson bracket. We look for the Jacobi identity written in
terms of the inverse matrix (ij ).
Any differentiable manifold may in principle become a symplectic manifold, provided there exists defined on it a closed nondegenerate two-form,
leading to a Poisson bracket. In the case of interest, the manifold is a Lie
group endowed with a hamiltonian structure consistent with the group structure.
Consider then as symplectic manifold a Lie group G with Lie algebra G0 .
Choose a basis {Ja } of generators, with [Ja , Jb ] = f c ab Jc . Such generators
correspond to smooth left-invariant (or right-invariant) complete fields on
the manifold G acting on the functions F C (G, R) as derivations. A
curve on G will be given by a one-parameter set of elements g(t) = exp[tX],
where X = X a Ja is the generator corresponding to g = g(1) = exp[X]. Each
generator Ja is represented on the group manifold by a left-invariant field ea .
The set {ea } provides a basis for the vector fields on G, with [ea , eb ] = f c ab ec .
The consistency between the Lie group structure and the symplectic structure
of G is obtained by imposing that the fields ea be hamiltonian fields, that
is, that the infinitesimal transformations they generate are canonical. This
means that they preserve the symplectic cocycle and is expressed by the
vanishing of the Lie derivative
Lea = (d iea + iea d) = 0,
3
Drinfeld 1983.
(10.20)
632
PHYS. TOPIC 10.
where iea is the interior product. With the cocycle condition d = 0, this
implies the closedness of the form iea . It follows that there exists locally
a function Fa such that iea = dFa . The function Fa will be the generating
function of the canonical transformation generated by ea . Consequently,
{Fa , Fb } = ea (Fb ) = ab . The Jacobi identity is then
{{Fa , Fb }, Fc } + {{Fc , Fa }, Fb } + {{Fb , Fc }, Fa }
= {ab , Fc } + {ca , Fb } + {bc , Fa }
= ec (ab ) + eb (ca ) + ea (bc ) = 0. (10.21)
Using ec (ab ) = 12 [ec eb (Fa ) ec ea (Fb )] for each term of [10.21], we find
f d cb ad + f d ac bd + f d ba cd = 0.
Contracting with the product ka jb ic , we arrive finally at
f i ab ka jb + f j ab ia kb + f k ab ja ib = 0.
(10.22)
We shall change to the standard notation in the literature, putting rab = ab

for a symplectic structure defined on a Lie group. Thus,
f i ab rka rjb + f j ab ria rkb + f k ab rja rib = 0.
(10.23)
This is the classical Yang-Baxter equation, though not in its most usual form,
which is given in direct product notation. The contravariant tensors (rkj )
may be seen as a map
G T G T G , g r(g) = rkj ek ej .
This represents on the group manifold a general member of the direct product
G0 G0 which will have the form r = rab Ja Jb . In this notation (Math.2.10),
the algebra is included in higher product spaces by adjoining the identity
algebra. For an element of G0 , we write, for example,
X1 = X 1 = X a (Ja 1); X2 = 1 X = X a (1 Ja ),
or
X1 = X 1 1; X2 = 1 X 1; X3 = 1 1 X,
and so on. Elements of G0 G0 may then be written
r12 = rab Ja Jb 1; r13 = rab Ja 1 Jb ; r23 = rab 1 Ja Jb .
We can make use of the multiple index notation:
10.5. CLASSICAL YANG-BAXTER EQUATION
633
< ij|A B|mn > = < i|A|m >< j|B|n >;

< ijk|A B C|mnr > = < i|A|m >< j|B|n >< k|C|r > ; etc.
If r belongs to G0 G0 , the matrix elements are
< ij|r|mn > = rij mn and < ijr|r E|mns > = sr rij mn .
We can then calculate to find
[r12 , r13 ] = rab rcd [Ja , Jc ] Jb Jd = rdb rec f a de Ja Jb Jc ;
[r12 , r23 ] = rab rcd Ja [Jb , Jc ] Jd = rad rec f b de Ja Jb Jc ;
[r13 , r23 ] = rab rcd Ja Jc [Jb , Jd ] = rad rbe f c de Ja Jb Jc .
The equation takes, therefore, its standard form
[r12 , r13 ] + [r12 , r23 ] + [r13 , r23 ] = 0.
(10.24)
The name classical comes from the fact that, when conveniently parametrized,
this equation is the limit h 0 of the Yang-Baxter equation [2.22] of Math.2.
Arnold 1976
Kirillov 1974
Babelon & Viallet 1989
634
PHYS. TOPIC 10.
Part VI
Glossary and Bibliography
635
GLOSSARY
Abelianizer: a canonical homomorphism : G G/[G,G] taking a group
G into its abelianized subgroup.
Affine space: a subspace of a linear space V whose elements may be
written in the form a = k + v0 , with k in a linear subspace of V and v0 a
fixed point of V.
Baire space: a space which is not a countable union of nowhere dense
subsets. A complete metric space is a Baire space.
Bijective: a mapping which covers all the target space (surjective or
onto) and is one-to-one (see function). Also called a condensation.
Canonical: in general, a basis independent object, or mapping. For an
isomorphism between linear spaces, a basis-independent isomorphism. For
groups, see homomorphism.
C*-algebra: see *-algebra. An involutive Banach algebra satisfying the
further condition ||u*u|| = ||u||2 . Only in such algebras can we talk about
self-adjointness: u is self-adjoint if u = u*.
Center of a group G: the set of elements of G commuting with all the
elements of G. The center of an algebra A is the subalgebra formed by those
elements commuting with all elements of A.
Centralizer of a subset X of a group G: the set of G elements commuting with every member of X; the centralizer of G itself is the center of G;
centralizer of a subset X of an algebra A: the set of elements of A commuting
with every member of X; the centralizer of A itself is the center of algebra A.
Characteristic of a ring: for a in ring R, call (na) the expression a +
a + . . . + a, with n summands. Positive integers ni may exist for which
(ni a) = 0 for all a R. Then n = minimum {ni } is the characteristic of R.
When no such ni s exist, R is of characteristic zero. The rings Z, R and C
are of this kind. Zn is of characteristic n.
Commutant S of a subset S of elements of an algebra A: S = {a A
such that a s = sa for all s S}. The commutant of A itself is its center.
Condensation: a bijective mapping.
Congruency: for a fixed positive integer n, the number r (0 r <n)
637
638
such that h = nq + r for some q is congruent to h modulo n. Notation:
h = r(mod n). The number r such that h + k = nq + r = r(mod n) is the
sum modulo n of h and k. This can be adapted to multiplication: the
multiplication modulo n of two integers p and q is the remainder of their
usual product when divided by n, the number m such that pq = m(mod n).
Conjugate class: if a is an element of a group (G, .), its conjugate
class [a] is the set of all elements of G which can be put under the form x a
x1 for some x G. Two conjugate classes [a] and [b] are either identical or
disjoint. The element a belongs to the center of G if and only if [a] = {a}.
Degree of an element of a graded algebra: see graded algebra.
Deformation of a topological space X: a family {hs } of mappings hs :
X X, with parameter s I [0, 1] such that h0 is the identity mapping
and the function H: I X I X defined by H(s,p) = hs (p) is continuous.
When the mappings hs are homeomorphisms, the deformation is an isotopy,
or isotopic deformation. A deformation into or onto a subspace Y is a
deformation of X such that the image h1 X is contained in (or is equal to) Y.
A subspace Y of a topological space X is a deformation retract of X if there
is a retraction r: X Y which is a deformation. Deformation retractions
preserve homotopy type.
Differential graded algebra: a graded algebra on which is defined a
graded derivative , a derivation D such that D() = (D) + (-) D.
The standard example is the exterior derivative. Of special interest because
especially prone to cohomology.
Distance function: a function d taking any pair (p, q) of points of a
set X into the real line R and satisfying the following four conditions : (i)
d(p, q) 0 for all pairs (p, q); (ii) d(p, q) = 0 if and only if p = q; (iii)
d(p, q) = d(q, p) for all pairs (p, q); (iv) d(p, r) + d(r, q) d(p, q) for any
three points p, q, r. It is thus a mapping d : X X R+ . A space on
which a distance function is defined is a metric space. For vector spaces
this expression is usually reserved to translation-invariant distance functions.
A distance function is sometimes called a metric, but it is better to separate
the two concepts, though in effect a definite-positive metric tensor defines a
distance function.
Divisors of zero: two elements l and r of a ring < R, +, . > are (respectively left and right) divisors of zero if they are nonzero and such that
l.r = 0. In a commutative ring, left (right) divisors of zero are right (left)
divisors of zero. In the ring < Zn , +, . >, all numbers not relatively prime
to n are divisors of zero (and only them). So, 2 and 3 are divisors of zero in
Z6 . In particular, Zn has no divisors of zero when n is a prime number. See
integral domain.
Endomorphism: a homomorphism of a set endowed with some algebraic
639
structure (such as a group, a ring, a linear space, . . . ) into itself.
Epimorphism: a surjective homomorphism, that is, a map whose image
covers entirely the target space and which preserves the algebraic structure.
Equivalence relation: see relation.
Filter: a filter on the set S is a family F of subsets of S such that (i) the
empty set does not belong to F, (ii) the intersection of two members of F
is also a member, and (iii) any subset of S containing a member of F is also
a member. The simplest example is the set of all open neighbourhoods of
a fixed point p S. An ultrafilter is a filter which is identical to any filter
finer to it. Filters and ultrafilters are essential to the study of continuity and
convergence in non-metric topological spaces. The notion of filter is dual to
that of set ideal. See ideal.
Fr
echet space: a metrizable complete topological vector space.
Function: a mapping with a unique value in the target space for each
point of its domain. As a point set function, f: A B will be (i) surjective
(or onto) if f(A) = B; that is, the values of f for all points of A cover the whole
of B; (ii) injective (or one-to-one) if, for all a and a A, the statement
f(a) = f(a) implies a = a; that is, it takes distinct points of A into distinct
points of B; (iii) bijective (also called a condensation) if it is both onto
and one-to-one. If B A, an inclusion is an injective map i : B A with
i(p) = p if p B.
Graded algebra: a sum of vector spaces, V = k Vk , with the binary
operation taking Vi Vj Vi+j . If Vk , we say that k is the degree
(or order) of , and write = k. The standard example is the space of
differential forms of every order on a manifold M.
Graded derivative: see differential graded algebra.
Graph of a function F: M N : the set f(M) of points of N which are the
image by f of some point of M.
Homomorphism: in general, a mapping preserving algebraic structure.
A mapping f: (G, ) (H, # ) between two groups is a homomorphism
when for all a,b M G, f(a b) = f(a) # f(b). When such a mapping exists,
G and H are homomorphic. A homomorphism is canonical when it takes
a given member of G always into the same member of H (see trivializer and
abelianizer). Straightforward generalization for rings. See isomorphism.
Ideal: the sub-ring R of a ring R is a left-ideal if a b M R for all a
M R and b M R; it is a right-ideal if a b M R when a M R and b M
R; and a bilateral ideal if a b M R when either a M R or b M R. When
nothing else is said, the word ideal is used for bilateral ideals. When R is
such a bilateral ideal, there is a natural multiplication defined on the group
R/R. The resulting ring is a quotient ring. The ring Zn , +, . is the
quotient of Z by the ideal formed by all the multiples of n: Zn = Z/nZ. The
640
ring R is a maximal ideal of R if, for any other ideal R, R c R implies
R = R. In the ring of complex functions on a topological space S, those
functions vanishing at a certain point p M S form an ideal. Analogous to,
and sometimes confused with, a normal subgroup. If R is a ring with unity,
and N is an ideal of R containing an element with a multiplicative inverse,
then N = R. On a set S, an ideal is a family I of subsets of S such that (i)
the whole set S does not belong to I, (ii) the union of two members of I is
also a member, and (iii) any subset of S contained in a member of I is also a
member.
Injective mapping: a one-to-one mapping, taking distinct points into
distinct points (see function).
Integral domain: a commutative ring with unity and with no divisors
of zero. Every field is an integral domain. Every finite integral domain is a
field. Zn is a field when n is prime.
Isomorphism: in general, a one-to-one onto (or bijective) mapping preserving algebraic structure. An isomorphism between two groups (G, ) and
(H, # ) is a bijective mapping f: G H such that for all a,b M G, f(a b) =
f(a) # f(b). When such a mapping exists, G and H are isomorphic. The
generalization for rings is straightforward. An isomorphism is a bijective
homomorphism.
Isotopy: see deformation.
Kernel of a mapping f: X Y: the set ker f = {x M X such that f(x)
= 0}. For a homomorphism f: G H, the kernel (ker f) is the set of all the
elements of G which are mapped into the identity element of H.
Metric: a second order nonsingular symmetric tensor. See distance
function.
Metric space: see distance function.
Metrizable space: a space whose topology may be generated by the
balls defined by a distance function. It is always first-countable. For topological vector spaces, this metric must be translation-invariant.
Monomorphism: an injective homomorphism, that is, a one-to-one
mapping preserving algebraic structure.
Multiplication modulo n: see congruency.
One-to-one: same as injective; see function.
Onto: see surjective and function.
Operation: a binary operation o on a set S is a rule assigning to each
ordered pair of elements (a, b) of S another element a o b of S. It establishes
structure on S, and a good notation would be S, o but a structured set is
frequently denoted simply by the symbol of the set point, S. When a o b =
b o a for all a, b M S, the operation is commutative. When a o (b o c) =
(a o b) o c for all a, b, c M S, the operation is associative.
641
Power set P(S) of a given set S: the set of all subsets of S. Sometimes
indicated by the notation Exp S. If S is finite with n elements, P(S) has
2n elements. Also for S infinite, P(S) is larger then S (Cantor theorem).
Prametric: on a set S, a prametric is a mapping : S S R+ such
that (p, p) = 0 for all p S. Once endowed with a prametric, the space S is a
prametric space, and (p, q) is the prametric distance between p and q. It
is possible to have (p, q) = 0 even if p 6= q. If (p, q) 6= 0 implies p 6= q, then
the prametric is separating. If for all pairs of arguments (p, q) = (q, p), the
prametric is symmetric. A symmetric is a separating symmetric prametric.
Metrics are very special cases of symmetrics (cf. distance function). A
prametric is enough to define a topology on S.
Relation on a set S: a subset of S S. Given a relation R between
the points of a set, we write pRq to mean that p is in that relation with
q. The relation R is an equivalence relation when it is reflexive (each
point p is in that relation with itself, pRp, p), symmetric (pRq implies
qRp) and transitive (pRq and qRr together imply pRr). Equality is the
obvious example. Loosely speaking, near is an equivalence but far is
only symmetric. On the plane with coordinates (x, y), the relation has the
same coordinate x = x0 establishes an equivalence between all the points
on the same vertical line. An equivalence relation (for which the notation
instead of R is usual) divides a point set into equivalence classes, subsets
of points having that relation between each other. In the example of the
plane, each vertical line is an equivalence class and can be labelled by the
value of x0 .
Retraction of a topological space X onto a subspace Y : a continuous
mapping r : X Y such that, for any p Y , r(p) = p. When such a
retraction exists, Y is a retract of X.
Rn : the set of ordered n-uples of real numbers. Each n-uple is in general
represented as x = (x1 , x2 , ..., xn ).
Separating: a prametric m is separating if m(p, q) 6= 0 implies p 6= q. A
metric is always separating.
Sum modulo n: see congruency.
Surjective mapping: a map whose image covers entirely the target space
(see function); same as onto.
Topological invariant: a property (quality or number) of a topological
space which is shared by all spaces homeomorphic to it (invariant under
homeomorphisms).
Ultrafilter: see filter.
642
REFERENCES
Abraham R & Marsden J 1978: Foundations of Mechanics (2nd ed.), BenjaminCummings, Reading, Mass.
Adams C C 1994: The Knot Book, W.H. Freeman and Co., New York.
Agarwal G S & Wolf E 1970: Phys. Rev. D2 2161.
Aharonov Y & Anandan J 1987: Phys. Rev. Lett. 58 1593.
Aharonov Y & Anandan J 1988: Phys. Rev. D38 1863.
Aharonov Y & Bohm D 1959: Phys. Rev. 115 485.
Aharonov Y & Bohm D 1961: Phys. Rev. 123 1511.
Aldrovandi R 1992: Fort. Physik 40 631.
Aldrovandi R & Galetti D 1990: J. Math. Phys. 31 2987.
Aldrovandi R & Kraenkel R A 1988: J. Phys. A21 1329.
Aldrovandi R & Pereira J G 1986: Phys. Rev. D33 2788.
Aldrovandi R & Pereira J G 1988: Rep.Math.Phys. 26 237.
Aldrovandi R & Pereira J G 1991: in MacDowell S, Nussenzveig H M
& Salmeron R A (eds.), Frontier Physics: Essays in Honour of J. Tiomno,
World Scientific, Singapore.
Alexandrov P 1977: Introduction `
a la Theorie Homologique de la Dimension et la Topologie Combinatoire, MIR, Moscow.
Anderson I M and Duchamp T 1980: Am. J. Math. 102 781.
Arafune J , Freund P G O & Gobel C J 1975: J. Math. Phys. 16 433.
Arkhangelskii A V & Fedorchuk V V 1990: in Arkhangelskii A V &
Pontryagin L S (eds.), General Topology I, Encyclopaedia of Mathematical
Sciences, vol.17, Springer, Berlin.
Arnold V I 1966: Ann. Inst. Fourier 16 319.
Arnold V I 1973: Ordinary Differential Equations, MIT Press, Cambridge.
Arnold V I 1976: Les Methodes Mathematiques de la Mechanique Classique, MIR, Moscow.
Arnold V I 1980: Chapitres Supplementaires de la Theorie des Equations

Differentielles Ordinaires, MIR, Moscow.
643
644
Arnold V I, Kozlov V V & Neishtadt A I 1988: in Arnold V I (ed.),
Dynamical Systems III, Encyclopaedia of Mathematical Sciences, Springer,
Berlin.
Assis A K T 1989: Found. Phys. Lett. 2 301.
Askar A 1985: Lattice Dynamical Foundations of Continuum Theories,
Atherton R W and Homsy G M 1975: Stud. Appl. Math. 54 31.
Atiyah M F, Drinfeld V G, Hitchin N J & Manin Yu. I 1978: Phys. Lett.
A65 185.
Atiyah M F 1979: Geometry of Yang-Mills Fields, Acad. Naz. Lincei,
Pisa.
Atiyah M F 1991: The Geometry and Physics of Knots (Lezioni Lincee,
Accademia Nazionale dei Lincei, Pisa), Cambridge University Press, Cambridge.
Avis S J & Isham C J 1978: Proc. Roy. Soc. 362 581.
Avis S J & Isham C J 1979: in Levy M & Deser S, Recent Developments
in Gravitation, Proceedings of the 1978 Cargèse School, Gordon and Breach,
New York.
Babelon O & Viallet C -M 1989: Integrable Models, Yang-Baxter Equation, and Quantum Groups, SISSA-ISAS preprint 54 EP.
Baker G A 1958: Phys. Rev. 109 2198.
Balescu R 1975: Equilibrium and Nonequilibrium Statistical Mechanics,
J. Wiley, New York.
Baulieu L 1984: Nucl. Phys. B241 557.
Baxter R J 1982: Exactly Solved Models in Statistical Mechanics,
Academic Press, London.
Bayen F, Flato M, Fronsdal C, Lichnerowicz A & Sternheimer D 1978:
Ann . Phys.
111 61 and 111.
Becher P & Joos H 1982: Z. Phys. C15 343.
Berry M V 1976: Advances in Physics 25 1.
Berry M V 1984: Proc. Roy. Soc. A392 45.
Biggs N L, Lloyd E K & Wilson R J 1977: Graph Theory 1736-1936,
Oxford University Press, Oxford.
Birman J S 1975: Braids, Links, and Mapping Class Groups, Princeton
University Press, Princeton.
Birman J S 1991: Math. Intell. 13 52.
Bjorken J D & Drell S D 1964: Relativistic Quantum Mechanics, McGrawHill,
New York.
645
Bjorken J D & Drell S D 1965: Relativistic Quantum Fields, McGraw-Hill,
New York.
Bogoliubov N N & Shirkov D V 1980: Introduction to the Theory of
Quantized Fields
(3rd ed.), J. Wiley, New York.
Bonora L & Cotta-Ramusino P 1983: Comm. Math. Phys. 87 589.
Boothby W M l975 : An Introduction to Differentiable Manifolds and
Riemannian Geometry, Academic Press, New York.
Born M & Wolf E 1975: Principles of Optics, Pergamon, Oxford.
Born M 1964: Natural Philosophy of Cause and Chance, Dover, New
York.
Born M & Infeld L 1934: Proc. Roy. Soc. A144 425.
Boya L J, Cari
nena J F & Mateos J 1978: Fort. Physik 26 175.
Bratelli O & Robinson D W 1979: Operator Algebras and Quantum Statistical Mechanics I, Springer, New York.
Broadbent S R & Hammersley J M 1957: Proc. Camb. Phil. Soc. 53
629.
Budak B M & Fomin S V 1973: Multiple Integrals, Field Theory and
Series,
MIR, Moscow.
Burgers J M 1940: Proc. Phys. Soc. (London) 52 23.
Burke W L l985: Applied Differential Geometry, Cambridge University
Press, Cambridge.
Chandrasekhar S 1939: An Introduction to the Study of Stellar Structure,
U. Chicago Press, Chicago.
Chandrasekhar S 1972: Am. J. Phys. 40 224.
Chillingworth D 1974: in Global Analysis and its Applications, IAEA,
Vienna, vol.I.
Chinn W G & Steenrod N E 1966: First Concepts of Topology, Random
House,
New York
Cho Y M l975: J. Math. Phys. 16 2029.
Choquet-Bruhat Y, DeWitt-Morette C & Dillard-Bleick M 1977: Analysis, Manifolds and Physics, North-Holland, Amsterdam.
Christie D E 1976: Basic Topology, MacMillan, New York.
Coleman S 1977 : Classical Lumps and their Quantum Descendants, in
A. Zichichi (ed.), New Phenomena in Subnuclear Physics, Proceedings of
the 1975 International School of Subnuclear Physics (Erice, Sicily), Plenum
Press, New York.
Coleman S 1979 : The Uses of Instantons, in A. Zichichi (ed.), The
Whys of Subnuclear Physics , Proceedings of the 1977 International School
646
of Subnuclear Physics (Erice, Sicily), Plenum Press, New York
Connes A 1980: C. R. Acad. Sci. Paris 290A 599.
Connes A 1986: Noncommutative Differential Geometry, Pub. IHES
(Paris) 62 257.
Connes A 1990: Geometrie Non-Commutative, InterEditions, Paris.
Comtet L 1974: Advanced Combinatorics, Reidel, Dordrecht.
Coquereaux R 1989: J. Geom. Phys. 6 425.
Croom F H 1978: Basic Concepts of Algebraic Topology, Springer, Berlin.
Crowell R H & Fox R H 1963: Introduction to Knot Theory, Springer,
Berlin.
Daniel M & Viallet C M l980: Rev. Mod. Phys. 52 175.
Davis W R & Katzins G H 1962: Am. J. Phys. 30 750.
de Gennes P G & Prost J 1993: The Physics of Liquid Crystals (2nd ed.),
Clarendon Press, Oxford.
de Rham G 1960: Varietes Differentiables, Hermann, Paris.
DeWitt-Morette C 1969: Ann. Inst. Henri Poincare XI 153.
DeWitt-Morette C 1972: Comm. Math. Phys. 28 47.
DeWitt-Morette C, Masheshwari A & Nelson B 1979: Phys. Rep. 50
255.
Dirac P A M 1926: Proc. Roy. Soc. A109 642.
Dirac P A M 1958: in W. Frank (ed.), Planck Festschrift, VEB Deutscher
Verlag der Wissenschaften, Berlin, p. 339.
Dirac P A M 1984: The Principles of Quantum Mechanics, Clarendon
Press, Oxford.
Dittrich J 1979: Czech. J. Phys. B29 1342.
Dixmier J 1981: von Neumann Algebras, North-Holland, Amsterdam.
Dixmier J 1982: C*-Algebras, North-Holland, Amsterdam.
Donaldson S K & Kronheimer P B 1991: The Geometry of Four-Manifolds,
Clarendon Press, Oxford.
Doubrovine B, Novikov S & Fomenko A 1979: Geometrie Contemporaine,
MIR, Moscow.
Dowker J S 1979: Selected Topics in Topology and Quantum Field Theory,
lectures delivered at the Center for Relativity, Austin, Texas.
Drinfeld V G 1983: Sov. Math. Dokl. 27 222.
Dubois-Violette M 1991: in Bartocci C, Bruzzo U & Cianci R (eds.),
Differential Geometric Methods in Theoretical Physics, Proceedings of the
1990 Rapallo Conference, Springer Lecture Notes in Physics 375, Springer,
Berlin.
Dubois-Violette M 1988: C. R. Acad. Sci. Paris 307I 403.
Dubois-Violette M, Kerner R & Madore J 1990: J. Math. Phys. 31 316.
Dubois-Violette M & Launer G 1990: Phys. Lett. B245 175.
647
Efros A L 1986: Physics and Geometry of Disorder, MIR, Moscow.
Eisenhart L P 1949: Riemannian Geometry, Princeton University Press,
Princeton.
Enoch M & Schwartz J.-M 1992: Kac Algebras and Duality of Locally
Compact Groups, Springer, Berlin.
Essam J W 1972: in Domb C & Green M S (eds.), Phase Transitions and
Critical Phenomena, vol.2, Academic Press, London.
Ezawa Z F 1978: Phys. Rev. D18 2091.
Ezawa Z F 1979: Phys. Lett. B81 325.
Faddeev L D 1982: Sov. Phys. Usp. 25 130.
Faddeev L D & Shatashvilli S L 1984: Theor. Math. Phys. 60 770.
Faddeev L D & Slavnov A A 1978: Gauge Fields. Introduction to Quantum Theory, Benjamin/Cummings, Reading.
Fairlie D B, Fletcher P & Zachos C Z 1989: Phys. Lett. B218 203.
Farmer J D, Ott E & Yorke J E 1983: Physica D7 153.
Feynman R P, Leighton R B & Sands M 1965: The Feynman Lectures in
Physics, Addison-Wesley, Reading.
Finkelstein D 1966: J. Math. Phys. 7 1218.
Finlayson B A l972: Phys. Fluids 13 963.
Flanders H l963: Differential Forms, Academic Press, New York.
Fleming H 1987: Rev. Bras. Fs. 17 236.
Fock V 1926: Z. Physik 39 226.
Fock V A 1964:The Theory of Space, Time and Gravitation (2nd ed.),
Pergamon, New York.
Forsyth A R 1965: The Theory of Functions, Dover, New York.
Fraleigh J B 1974: A First Course in Abstract Algebra, Addison-Wesley,
Reading.
Franck F C 1951: Phil. Mag. 42 809.
Freed D S & Uhlenbeck K K l984: Instantons and Four-Manifolds, Springer,
Berlin.
Fullwood D T 1992: J. Math. Phys. 33 2232.
Furry W H 1963: in Brittin W E, Downs B W & Downs J (eds.), Lectures in Theoretical Physics, vol. V, 1962 Summer Institute for Theoretical
Physics, (Boulder) Interscience, New York.
Galetti D & Toledo Piza A F 1988: Physica A149 267.
Gelfand I M & Shilov G E 1964: Generalized Functions, vol.1,
Academic Press, New York.
Gelfand I M, Graev I M & Vilenkin N Ya 1966: Generalized Functions,
Vol. 5,
Academic Press, New York.
Geroch R P 1968 : J. Math. Phys. 9 1739.
648
Gerstenhaber M & Stasheff J (eds.) 1992: Deformation Theory and
Quantum Groups with Applications to Mathematical Physics, AMS, Providence.
Gibling P J 1977: Graphs, Surfaces and Homology, Chapman & Hall,
London.
Gilmore R 1974: Lie Groups, Lie Algebras, and Some of Their Applications,
J.Wiley, New York.
Godbillon G 1971: Elements de Topologie Algebrique, Hermann, Paris.
Goldberg S I l962: Curvature and Homology, Dover, New York.
Goldstein H 1980: Classical Mechanics (2nd ed.), Addison-Wesley, Reading.
Gobel R 1976a: J. Math. Phys. 17 845.
Gobel R 1976b: Comm. Math. Phys. 46 289.
Graver J E & Watkins M E 1977: Combinatorics with Emphasis on the
Theory of Graphs, Springer, Berlin.
Greenberg M J 1967: Lectures on Algebraic Topology, Benjamin, Reading.
Grebogi C, Ott E & Yorke J A 1987: Science 238 585.
Guckenheimer J & Holmes P 1986: Nonlinear Oscillations, Dynamical
Systems, and Bifurcations of Vector Fields, Springer, Berlin.
Guillemin V & Sternberg S 1977: Geometric Asymptotics, AMS, Providence.
Gurarie D 1992: Symmetries and Laplacians, North Holland, Amsterdam.
G
ursey F 1962: Introduction to the de Sitter Groups, in G
ursey F (ed.),
Group Theoretical Concepts and Methods in Elementary Particle Physics,
Istanbul Summer School of Theoretical Physics, Gordon and Breach, New
York.
Haag R 1993: Local Quantum Physics, Springer, Berlin.
Hajicek P 1971: Comm. Math. Phys. 21 75.
Halmos P R 1957: Introduction to Hilbert Spaces, Chelsea Pub.Co., New
York.
Hamermesh M 1962: Group Theory and its Applications to Physical Problems,
Addison-Wesley, Reading.
Harris W F 1975: Phil. Mag. B32 37.
Hawking S W & Ellis G F R 1973: The Large Scale Structure of SpaceTime,
Cambridge University Press, Cambridge.
Hawking S W & Israel W (eds.) 1979: General Relativity: An Einstein
Centenary Survey, Cambridge University Press, Cambridge.
Hawking S W, King A R & McCarthy P J 1976: J. Math. Phys. 17 174.
649
Hehl F W, von der Heyde P, Kerlick G D & Nester J M 1976: Rev. Mod.
Phys. 48 393.
Hilton P J & Wylie S l967: Homology Theory, Cambridge University
Press, Cambridge.
Hilton P J 1953: An Introduction to Homotopy Theory, Cambridge University Press, Cambridge.
Hocking J C & Young G S 1961: Topology, Addison-Wesley, Reading.
Hu S T 1959: Homotopy Theory, Academic Press, New York.
Huang K 1987: Statistical Mechanics (2nd ed.), J. Wiley, New York.
Hurewicz W & Wallman H 1941: Dimension Theory, Princeton University Press, Princeton.
Hwa R W & Teplitz V L 1966: Homology and Feynman Integrals, Benjamin, Reading.
Isham C J, Penrose R & Sciama D W (eds.) 1975: Quantum Gravity,
Oxford University Press, Oxford.
Isham C J 1978: Proc. Roy. Soc. 362 383.
Isham C J 1984: in DeWitt B S & Stora R (eds.), Relativity, Groups and
Topology II, Les Houches Summer School, North-Holland, Amsterdam.
Itzykson C & Zuber J.-B. 1980: Quantum Field Theory, McGraw-Hill,
New York.
Jackiw R 1980: Rev. Mod. Phys. 52 661.
Jackson J D 1975: Classical Electrodynamics (2nd ed.) John Wiley, New
York.
Jacobson N 1962: Lie Algebras, Dover, New York.
Jancel R 1969: Foundations of Classical and Quantum Statistical Mechanics,
Pergamon, Oxford.
Jones V F R 1983: Invent. Math. 72 (1983) 1.
Jones V F R 1985: Bull. Amer. Math. Soc. 12 103.
Jones V F R 1987: Ann. Math. 126 335.
Jones V F R 1991: Subfactors and Knots, CBMS, Amer. Math. Soc.,
Providence, Rhode Island.
Katznelson Y 1976: An Introduction to Harmonic Analysis, Dover, New
York.
Kauffman L H 1991: Knots and Physics, World Scientific, Singapore.
Khavin V P 1991: in Khavin V P & Nikolskij N K (eds.), Commutative Harmonic Analysis I, Encyclopaedia of Mathematical Sciences, vol.15,
Springer, Berlin.
Kirillov A 1974: Elements

de la Theorie des Representations, MIR, Moscow.
Kleman M & Sadoc J F 1979: Journal de Physique - Lettres 40 569.
650
Kohno T 1990: New Developments in the Theory of Knots, World Scientific, Singapore.
Kobayashi S & Nomizu K l963: Foundations of Differential Geometry,
Interscience, New York.
Kolmogorov A N & Fomin S V 1970: Introductory Real Analysis, PrenticeHall, Englewood Cliffs.
ements de la Theorie des FoncKolmogorov A N & Fomin S V 1977: El
tions et de
lAnalyse Fonctionelle, MIR, Moscow.
Konopleva N P & Popov V N 1981: Gauge Fields, Harwood, Chur.
Laidlaw M G G & DeWitt-Morette C 1971: Phys. Rev. D3 1375.
Lanczos C 1986:The Variational Principles of Mechanics (4th ed.) Dover,
New York.
Landau L D & Lifshitz E M l969: Statistical Physics (2nd ed.), Pergamon,
Oxford.
Landau L D & Lifshitz E M l975: The Classical Theory of Fields, Pergamon, Oxford.
Landau L D & Lifshitz E M l989: Mechanique des Fluides, MIR, Moscow.
Landau L D & Lifshitz E M l990:Theory of Elasticity, MIR, Moscow.
Lavrentiev M & Chabat B 1977: Methodes de la Theorie des Fonctions
dUne Variable Complexe, MIR, Moscow.
Leinaas J M & Myrheim J 1977: Nuovo Cimento B37 1.
Levi B G 1993: Physics Today, 46 (#3) 17.
Levy M & Deser S (eds.) 1979: Recent Developments in Gravitation,
Cargèse Lectures 1978, Plenum Press, New York.
Lichnerowicz A 1955: Theorie Globale des Connexions et des Groupes
dHolonomie, Dunod, Paris.
London F 1927: Z. Physik 42 375.
Loos H G 1967: J. Math. Phys. 8 2114.
Love A E H 1944: A Treatise on the Mathematical Theory of Elasticity,
Dover, New York.
Lovelock P & Rund J l975: Tensors, Differential Forms and Variational
Principles,
J. Wiley, New York.
Lu Qi-keng 1975: Chin. J. Phys. 23 153.
Luneburg R K 1966: Mathematical Theory of Optics, University of California Press, Berkeley.
Mackey G W 1955: The Theory of Group Representations, Lecture Notes,
Department of Mathematics, University of Chicago.
Mackey G W 1968: Induced Representations of Groups and Quantum
Mechanics, Benjamin, New York.
651
Mackey G W 1978: Unitary Group Representations in Physics, Probability and Number Theory, Benjamin/Cummings, Reading.
Majid S 1990: Int. J. Mod. Phys. A5 1.
Malament D B 1977: J. Math. Phys. 18 1399.
Mandelbrot B B 1977: The Fractal Geometry of Nature, W. H. Freeman,
New York.
Ma
nes J, Stora R & Zumino B 1985: Comm. Math. Phys. 108 157.
Manin Yu I 1989: Comm. Math. Phys. 123 163.
Marsden J 1974: Applications of Global Analysis in Mathematical Physics,
Publish or Perish, Inc., Boston.
Maslov V P 1972: Theorie des Perturbations et Methodes Asymptotiques,
Dunod-Gauthier-Villars, Paris.
Maurer J 1981: Mathemecum, Vieweg & Sohn, Braunschweig, Wiesbaden.
McGuire J B 1964: J. Math. Phys. 5 622.
Michel L 1964: in G
ursey F (ed.), Group Theoretical Concepts and Methods in Elementary Particle Physics, Istanbul Summer School of Theoretical
Physics (1963), Gordon and Breach, New York.
Milnor J 1973: Morse Theory, Princeton University Press, Princeton.
Misner C W & Wheeler J A 1957: Ann. Phys. 2 525.
Misner C W, Thorne K & Wheeler J A 1973: Gravitation, W. H. Freeman,
San Francisco.
Mrugala R 1978: Rep. Math. Phys. 14 419.
Munkres J R l975: Topology: a First Course, Prentice Hall, New York.
Nabarro F R N 1987: Theory of Crystal Dislocations, Dover, New York.
Nash C & Sen S l983: Topology and Geometry for Physicists, Academic
Press, London.
Neuwirth L P 1965:Knot Groups, Princeton University Press, Princeton.
Neuwirth L P 1979: Scient. Am. 240 84.
Nomizu K 1956: Lie Groups and Differential Geometry, Publ. Math.Soc.
Japan.
Novikov S P 1982: Sov. Math. Rev. 3 3.
Okubo S 1980: Phys. Rev. D22 919.
Okun L B 1984: Introduction to Gauge Theories, publications of the
Institute of Theoretical and Experimental Physics ITEP - 43, Moscow.
Olver P J 1986: Applications of Lie Groups to Differential Equations,
Springer, New York.
Ore O 1963: Graphs and their Uses, L. W. Singer, New York.
Papapetrou A 1951: Proc. Roy. Soc. A209 248.
Pathria R K 1972: Statistical Mechanics, Pergamon, Oxford.
Penrose R & MacCallum M A H 1977: Phys. Rep. 6 241.
Penrose R 1972: Rep. Math. Phys. 12 65.
652
Pontryagin L S 1939: Topological Groups, Princeton University Press,
Princeton.
Popov D A l975: Theor. Math. Phys. 24 347.
Porteous I R 1969: Topological Geometry, van Nostrand, London.
Quinn F 1982: J. Diff. Geom. 17 503.
Rasband S N 1990: Chaotic Dynamics of Nonlinear Systems, J. Wiley,
New York.
Regge T 1961: Nuovo Cimento 19 558.
Rolfsen D 1976: Knots and Links, Publish or Perish, Berkeley.
Rosenfeld B A & Sergeeva N D 1986: Stereoghaphic Projection, MIR,
Moscow.
Rourke C & Stewart I 1986: New Scientist, 4th September, pg. 41
Saaty T L & Kainen P C 1986: The Four-Color Problem, Dover, New
York.
Sadoc J F & Mosseri R 1982: Phil. Mag. B45 467.
Schmidt H.-J. 1993: Fort. Physik 41 179.
Schreider Ju. A 1975: Equality, Resemblance and Order, MIR, Moscow.
Schulman L 1968: Phys. Rev. 176 1558.
Schutz B 1985: Geometrical Methods of Mathematical Physics, Cambridge University Press, Cambridge.
Sierpi
nski W 1956: General Topology, University of Toronto Press; Dover
edition, New York, 2000.
Simmons G F 1963: Introduction to Topology and Modern Analysis, McGrawHill- Kogakusha, N.Y.-Tokyo.
Simon B 1983: Phys. Rev. Lett. 51 2167.
Singer I M & Thorpe J A 1967: Lecture Notes on Elementary Topology
and Geometry, Scott Foresman & Co., Glenview.
Singer I M 1981: Physica Scripta 24 817.
Skyrme T H R 1962: Nucl. Phys. 31 556.
Slebodzinski W. 1970: Exterior Forms and their Applications, PWN,
Warszawa.
Sommerfeld A 1954: Optics, Academic Press, New York.
Sommerfeld A 1964a: Mechanics of Deformable Bodies, Academic Press,
New York.
Sommerfeld A 1964b: Partial Differential Equations in Physics, Academic Press,
New York.
Spivak M 1970: Comprehensive Introduction to Differential Geometry,
Publish or Perish, Brandeis.
Steen L A & Seebach, Jr J A 1970: Counterexamples in Topology, Holt,
Rinehart and Winston, New York.
653
Steenrod N l970: The Topology of Fibre Bundles, Princeton University
Press, Princeton.
Stora R 1984: in tHooft G et al. (eds.), Progress in Gauge Field Theory,
Plenum,
New York.
Streater R F & Wightman A S 1964: PCT, Spin and Statistics, and all
That, Benjamin, New York.
Synge J L 1937: Geometrical Optics, Cambridge University Press, Cambridge.
Synge J L 1960: Relativity: The General Theory, J. Wiley, New York.
Synge J L & Schild A 1978: Tensor Calculus, Dover, New York.
Takesaki M 1978: in DellAntonio G et al (eds.), Springer Lecture Notes
in Physics 80, Springer, Berlin.
Temperley H N V & Lieb E H 1971: Proc. Roy. Soc. 322 251.
Thom R 1972: Stabilite Structurelle et Morphegenèse , Benjamin, New
York.
Thom R 1974: Modèles Mathematiques de la Morphegenèse, Union Generale
d Editions, Paris.
Timoshenko S & Goodier J N 1970: Theory of Elasticity (3rd ed.) McGrawHill,
New York.
Toulouse G & Vannimenus J 1980: Phys. Rep. 67 47.
Trautman A l970: Rep. Math. Phys. 1 29.
Trautman A l979: Czech. J. Phys. B29 107.
Utiyama R 1955: Phys. Rev. 101 1597.
Vainberg M M l964: Variational Methods for the Study of Nonlinear Operators,
Holden-Day, San Francisco.
Vilenkin N Ya 1969: Fonctions Speciales et Theorie de la Representation
des Groupes, Dunod, Paris.
Warner F W l983: Foundations of Differential Manifolds and Lie Groups,
Scott-Foreman, Glenview.
Warner G 1972: Harmonic Analysis on Semi-Simple Lie Groups I, Springer,
Berlin.
Weinberg S 1972: Gravitation and Cosmology, J. Wiley, New York.
Wenzl H 1988: Invent. Math. 92 349.
Westenholz C l978: Differential Forms in Mathematical Physics, NorthHolland, Amsterdam.
Weyl H 1919: Annalen der Physik 59 101.
Weyl H 1929: Z. Physik 56 330.
654
Weyl H 1932: Theory of Groups and Quantum Mechanics, E. P. Dutton,
New York.
Wheeler J A & Zurek W H (eds.) 1983: Quantum Theory and Measurement, Princeton University Press, Princeton.
Whittaker E T 1944: Analytical Dynamics, Dover, New York.
Whittaker E T 1953: A History of the Theories of Aether and Electricity,
Thomas Nelson, London.
Woodcock A & Davis M 1980: Catastrophe Theory, Penguin Books, Harmondsworth.
Wootters W & Zurek W H 1979: Phys.Rev. D19 473.
Woronowicz S L 1987: Commun. Math. Phys. 111 613.
Wu T T & Yang C N 1975: Phys. Rev. D12 3843.
Wu Y S & Zee A 1985: Nucl. Phys. B258 157.
Yang C N & Mills R L 1954: Phys. Rev. 96 191.
Yang C N 1974: Phys. Rev. Lett. 33 445.
Yang C N & Ge M L 1989: Braid Group, Knot Theory and Statistical
Mechanics,
Yee K & Bander M 1993: Phys. Rev. D48 2797.
Zeeman E C 1964: J. Math. Phys. 5 490.
Zeeman E C 1967: Topology 6 161.
Zeeman E C 1977: Catastrophe Theory, Addison-Wesley, Reading.
Zumino B, Wu Y S & Zee A 1984: Nucl. Phys. B241 557.
Zwanziger J W, Koenig M & Pines A 1990: Ann. Rev. Phys. Chem. 41
601.
Index
E 3 , 189
RP n , 45
S 1 , 32, 38, 42
S 3 , 43
S 4 , 22, 235, 237
S n , 21, 89, 133, 235, 317
Sn , 354
S n , 11, 45
Rn , 641
0 , 120
1 , 86, 92
n , 116
1 , 113
-algebra, 35
E1 , 20
bad, 25
1
E+ , 37
E2 , 9
E3 , 3, 200, 218, 221
E4 , 22, 29
En , 8, 20, 22, 30, 31, 89, 214
R+ , 42
*-algebra, 391
Cn , 45
E4 , 134
En , 45
abelian, 334
abelianized, 337
abelianizer, 337, 637
accumulation point, 16
acoustics, 556
acoustics., 557
action, 626
defined, 342
of a group, 167
on phase space, 625
of a set, 342
on a basis, 178
action functional, 579
adjoint
action, 239
in algebra, 345
representation, 265
Ados theorem, 405
affine
group, 44
space, 637
transformation, 296
affine: locally symmetric manifold, 296
Aharonov-Bohm effect, 77, 97, 102,
131
Alexander polynomial, 368
Alexandrovs compactification, 35
algebra, 344, 347
Banach, 386
C*, 389
defined, 344
enveloping, 347
exterior, 158
factor, 394
general definition, 239
graded, 157, 240, 346
Grassmann, 158
Hopf, 348
Lie, 163, 239
noncommutative, 154
655
656
INDEX
of a group, 347
Bn , 359, 398
of functions, 145, 392
bad E1 , 25
of sets, 15
Baire space, 637
semisimple, 271
Banach
von Neumann, 393
algebra, 326, 386, 387
W*, 393
space, 20, 382, 386, 387, 409
algebra-valued form, 210
barycentric coordinates, 53
algebraic topology, 46, 67
basic field, 287
Allendoerfer, 458
basis, 175
alphabet, 352
anholonomic, 176, 209
alternating group, 355
canonical, 178
alternation, 158
coordinate, 176
Ambrose-Singer holonomy theorem, 315
covector, 178
analytic
dual, 148
atlas, 134
for a topology, 9
manifold, 134
general, 208
angular
global, 175
momentum
holonomic, 176, 209
current, 581
local, 175
anholonomic basis, 176, 209
non-coordinate, 176
annular group, 348
transformations, 177
anomaly, 602, 625, 627
vector
anti-de Sitter spacetime., 615
coordinate, or natural, 147
antiderivation, 242
Berrys phase, 315
antipode, 349
Bethe lattice, 55, 536
antisymmetric tensors, 157
Betti number, 65, 67, 69, 234, 271,
arcwise-connected, 34
367, 447
Aristotle, iii
bialgebra, 348
Arnold, 165
Bianchi identity, 212, 225, 291, 312,
Artins braid groups, 357
593
associated bundle, 179, 276, 293, 297
first, 289
asymptotic flatness, 537
second, 289
atlas
bijective, 637, 639
analytic, 134
billiard theorem, 376
differentiable, 128
bimodule, 341
fibered, 275
Birkhoffs theorem, 377
in general, 128
Bohm-Aharonov effect, 77, 97, 102,
linear, 128
131
attractor, 166
Boliyai, iv
automorphism, 30, 43, 265
Boltzmann, 524
inner, 265
Borel
INDEX
-algebra, 373
measure, 373
set, 373
Born-Infeld electrodynamics, 434
boundary, 122
homomorphism, 56
conditions, 22, 40, 99
and bundles, 274
homomorphism, 67
of a boundary, 63
of a set, 16
boundary conditions, 39
bounded
operator, 423
set, 423
Braid
statistics, 360
braid
equation, 362
geometrical, 356
group, 114, 357, 360, 525, 533
statistics, 104, 114, 525
tame, 357
branch (in a graph), 51
bridges of Konigsberg, 54
Brouwer degree, 153, 443
BRST transformations, 603
bundle
associated, 179, 276, 293, 297
existence of, 275
fiber, 162
heuristic introduction, 273
induced, 302
normal, 273
of 1-forms, 164
of frames, 179, 284, 605
of k-forms, 190
principal, 179
reduced, 302
reduction, 282
space, 162
657
structure group, 179
tangent, 162
tensor, 163
trivial, 164, 274
vector, 275
Burgers
circuit, 545
vector, 290, 545
C -structure, 133
C*-algebra, 389, 637
calculating theorem, 93
canonical, 637
homomorphism, 639
isomorphism, 164
canonical form, 280, 310
Cantor set, 17, 35, 124, 166
capacity dimension, 124
carrier space, 404
Cartan, 189
connection form, 215
lemma, 450
metric, 271
cartesian product, 14, 35
cartesian set product, 334, 342, 372
cartography, 126
catastrophe, 448
Cauchy sequence, 18
Cauchy-Schwarz inequality, 379
causal
anomalies, 24
structure, 31
Cayley
theorem on finite groups, 355
tree, 55, 536
Cech, 116
center, 637
of a group, 334
centralizer, 637
chain, 226
boundary, 63
658
closed, 64
coboundary, 64
defined, 63
face, 63
harmonic, 65
homologous, 68
of a path, 55
singular, 226, 234
chaos
chaotic behaviour, 166
deterministic, 376
character of a representation, 406, 407
characteristic
class, 319
equation, 565
function, 231, 374
of a ring, 637
characteristic equations, 553
characteristics, 553
charge, 584
chart, 125, 177
Chern
theorem, 458
class, 320
chiral field, 603
cholesterics, 549
chord (in a graph), 51
Christoffel symbol, 27, 186, 296, 454,
568, 607, 619
chromatic polynomial, 537
circle, 32, 42, 75
Classical Mechanics, 330
classification
of differentiable manifolds, 137
of fiber bundles, 319
closed
chain, 64
curve, 32
form, 198
set, 13
closure
INDEX
of a braid, 367
of a set, 16
coadjoint representation, 267, 630
coalgebra, 348
coarser topology, 13
coassociativity, 348
coboundary, 64
operator, 64
cochain, 64
coclosed
form, 222
cocycle, 64
Codazzi equation, 453
coderivative, 222, 243, 596
codifferential, 222
coexact
form, 222
cofield, 163
cohomology, 64
group, 69
of Lie algebra representations, 627
color, 607
commutant, 637
commutator, 162
subgroup, 337
compact, 25
group, 42
manifold, 232, 443
without boundaries, 235
space, 446
surface, 457
compact space, 20
compact-open topology, 33, 74, 120
compactification, 22, 35
compactness, 20
local, 22
compensating form, 214
complement, 13
linear, 343
complete
atlas, 129
INDEX
bundle space, 275, 298, 303
differentiable atlas, 134
field, 166
space, 18, 31
completion, 18
complex, 59, 68
components, 156, 160
composition, 135
of functions, 75, 82
comultiplication, 348
condensation, 31, 637, 639
condensation point, 16
conformal
group, 31
spaces, 618
transformation, 187
conformally flat metric, 466, 570
congruency, 637
conjugate class, 638
connected
graph, 54
space, 13
connectedness, 25, 34, 122
multiple, 89
path, 87
simple, 88
connection, 26, 621
as a distribution, 306
flat, 311
form, 306, 452
Cartan, 215
general, 605
general treatment, 303
Levi-Civitta, 611
linear, 286
continuity, 28, 373
continuity equation, 224
continuous operator, 423
contorsion
tensor, 612
contractible, 75, 91, 101, 198, 206
659
contraction, 156
contravariant
image, 182
tensor, 155
vector, 147
convergence, 15, 18, 373
convex, 59
convolution, 340
Conway, 368
coordinate, 32, 125, 136
function, 149
basis, 147, 176
function, 32
neighbourhood, 161
transformation, 127, 177, 588
coproduct, 348
cotangent
bundle, 164, 567
space, 147
counit, 349
covariant
coderivative, 214, 225
derivative, 186, 212, 213, 225, 284,
287, 583, 586
of a section, 294
tensor, 154
vector, 147
covector, 150
covering, 21
homotopy, 109
map, 107
space, 103, 105, 119
covering space, 105
CPn , 113
critical phenomena, 530
critical point, 445, 446
crosspoint, 444
crunode, 121
crystal optics, 184
Curie, 529
current form, 223
660
curvature, 288, 290, 452, 605, 611
constant, 460
form, 311, 452
of a curve, 566
of a light ray, 567
principal radii, 456
tensor, 454
total, Gaussian, 456
curve
closed, or loop, 32, 136
defined, 32
differentiable, 136
index of a, 440
integral, 166
cusp, 121
cycle, 64, 355
indicator polynomial, 356
structure of a permutation, 356
cyclic group, 354
cylinder, 39, 114
De Rham
current, 234
decomposition theorem, 236
theorem on harmonic forms, 236
theorem on periods, 230
de Sitter spacetime, 317
de Sitter spacetime., 615
defect, 538
deformation, 538, 638
field, 544
retract, 638
degree, 152, 638
Brower, 153
dense subset, 17, 375
derivation, 242, 391
algebraic, 239, 347
derivative
Frechet, 423
functional, 423
INDEX
Gateaux, or weak, 423
Lie, 169
strong, 423
derived
algebra, 347
set, 16
Descartes, iii, 187
determinants, 195
dielectric constant, 572
diffeomorphic manifolds, 137
diffeomorphism, 137, 163
differentiable
atlas, 128, 177
curve, 136, 145
distribution, 175
manifold, 160, 185
operator, 423
structure, 391
differential, 148
Frechet, 423
Gateaux, 423
strong, 423
weak, 423
differential equation, 183
differential form, 64, 160, 181, 189
differential operator
classification, 556
differential topology, 47, 443
differentially related, 128
dimension, 122, 123
of vector space, 342
defined, 342
dimension theory, 17
Dirac, 326
equation, 578
equation, 613
monopole, 77
space, 381
spinor, 613
Dirac space, 19
direct product, 361
INDEX
661
disc
2
in E , 93
once punctured, 95
twice punctured, 96
disclination, 538
discontinuity
propagation of, 558
discrete
group, 351
metric, 14
topology, 7, 14, 29
dislocation, 538
dispersion relation, 557
distance, 185
distance function, iii, 3, 33, 185, 638
distortion tensor, 544
distribution, 175, 231, 306
divergence, 222
division ring, 338
divisors of zero, 638
dual
basis, 148
of a form
definition, 218
invariant definition, 219
space, 154, 343, 382
tensor, 593
transformation, 223
dual
tensor, 213
dual space
defined, 343
duality
of chains and forms, 233
operation, 218
Pontryagin, 411
symmetry, 558, 593
Tanaka-Krein, 412
dynamical system: flows, 165
dynamical system:maps, 166
edge
closed, 50
open, 50
eikonal equation, 554
eikonal equation, 556, 562, 566
Einstein, 187
equation, 435, 608
space, 454
tensor, 607
Einstein-Hilbert action, 608
electrodynamics
Born-Infeld, 434
electromagnetic form, 203
electromagnetism, 202, 223, 229, 238
electron diffraction, 129
elementary particle, 575
ellipsoid
of inertia, or Poinsot, 184
Fletchers, 184, 573
Fresnel, 184
index, 184, 573
of wave normal, 573
reciprocal, 573
elliptic type operator, 556
endomorphism, 43, 239, 343, 638
defined, 343
energy-momentum, 588
ensemble
statistical, 523
enveloping
algebra, 347
epimorphism, 639
equation
Einsteins, 435, 608
Gauss, 453
Killing, 186
Maurer-Cartan, 209
Maxwells, 202
Maxwells second pair, 223
Ricci, 453
equivalence
662
class, 37, 74
classes, 641
relation, 30, 37, 639, 641
equivalence principle, 608
equivalent representations, 405
Ergodic
flow, 375
ergodic
problem, 376
theorem, 376, 524
theory, 375
Euclid, 459
euclidean plane, 9
euclidean space, iii, 3, 7, 21, 147, 214,
449, 460
pseudo-, 465
euclideanization, 29
Euler
characteristic, 447, 457
class, 320
number, 51, 444
Euler-Lagrange equation, 580
Euler-Poincare characteristic, 60, 69,
165, 443
event
in probability theory, 373
exact form, 164, 192, 198
exchange of particles, 41
exterior
derivative
introduced, 197
algebra, 158, 159
coderivative, 217
derivative, 210
in a general basis, 216
invariant definition, 200
product, 158, 159, 190, 197
as operation, 197
variational calculus, 437, 604
Faraday law, 229
INDEX
Fermats principle, 570
fermion field, 582
fermion number, 245
Feynmans picture, 92
fiber, 107, 162, 179
bundle, 162
heuristic introduction, 273
fibered chart, 275
fibration, 277
field, 161, 457, 576
basic, 287
complete, 166
defined, 339
fundamental, 269, 280
hamiltonian, 627
horizontal, 286
Jacobi, 417
relativistic, 577
scalar, 577
spinor, 578
strictly hamiltonian, 627
theory, 576
vector, 578
vertical, 280
fields of planes, 192
filter, 15, 639
finer topology, 13
finite space, 7, 13, 21
finitely generated group, 334
first homotopy group, 86
first quadratic form, 451
first-countable, 8, 10, 13
first-separability, 23, 25
flat
connection, 311
space, 460
Fletchers ellipsoid, 184, 573
flow, 165, 168
ergodic, 375
hamiltonian, 168
mixing, 375
INDEX
non-ergodic, 375
flow diagram, 165
Fock-Ivanenko derivative, 613
focusing
perfect, 571
force, 191
form, 163
algebra-valued, 210
canonical, 310
frame bundle, 280
closed, 198
coclosed, 222
coexact, 222
curvature, 311
differential, 147, 189
exact, 164, 198
harmonic, 223
horizontal, 310
Kirillov, 629
vector-valued, 150
vertical, 310
four-color problem, 54, 536
Fourier
analysis, 22, 409, 410
coefficient, 381
duality, 413
Fox, 116
Frechet
derivative, 419
derivative, 423
space, 32, 639
fractals, 124
frame, 175, 586
adapted, 450
transformations, 177
Franck index, 119, 550
free group, 353
Fresnel ellipsoid, 184, 572
Friedmann model, 229
Frobenius theorem, 175, 192, 199
in terms of forms, 216
663
full braid group, 357
function, 639
between differentiable manifolds,
135
continuous, 28
distance, 185
homotopic, 73
in coordinates, 135
monodromous, 27
polynomial, 157
function space, 33
functional, 418, 421
linear, 422
functorial properties, 90
fundamental field, 269, 280, 304
fundamental group, 83, 86, 112, 116,
352
fundamental sequence, 18
Gd (EN )], 46
Galileo, iii
Galois, 351
Gateaux derivative, or weak derivative, 423
gauge, 203, 210, 212, 301, 303, 310,
312
anomaly, 602
field, 592, 596
group, 591
potential, 307, 592
prescription, 594
theory, 276, 589, 609
transformation, 579, 598
gauge theory, 212, 238
Gauss
theorem, 455
curvature, 456
equation, 453
normal mapping, 455
theorem, 228, 456
Gelfand-Mazur theorem, 387
664
Gelfand-Naimark theorem, 390
Gelfand-Naimark-Segal construction,
396
general basis, 208
General Relativity, iv, 26, 187, 607
generator, 168, 334, 352
genus, 70
geodesic, 587, 612, 620
equation, 569
equation, 289, 612
geometrical phase, 316
Geometry, 187
geometry
intrinsic, 451
of surfaces, 455
GL(m, R), 178
GL(m, C), 43
GL(m, K), 43
GL(m, R), 43, 282
GNS construction, 396
Graded
algebra, 639
graded
algebra, 157, 191, 638
derivative, 638
ring, 338
graded algebra, 240, 346
graded commutator, 211
gradient, 192
Graph
of a function, 639
graph
defined, 50
theory, 54
Grassmann
algebra, 158
space
complex, 46
real, 46
Grassmann space, 318
gravitation, 607
INDEX
Greens theorem, 101, 228
group, 41
1-parameter, 167
abelian, 334
abstract, 335
action, 167
affine, 44
algebra of a, 347
alternating, 355
annular, 348
compact, 42
defined, 334
discrete, 351
finitely generated, 334
free, 353
Heisenberg, 329
nilpotent, 338
of an algebra, 346
of quaternions, 43
orthogonal, 281
presentation, 353
representation, 22, 335, 403
ring, 340
semisimple, 271, 337
simple, 337
solvable, 338
symmetric, 354
topological, 41
torsion-free, 354
transformation, 335
type of, 409
group reduction, 302
groupoid, 335
Haar measure, 22, 408, 411
Hahn extension theorem, 373
Hamilton equation, 554
Hamilton-Jacobi equation, 554
hamiltonian
field, 627
flow, 168
INDEX
formalism, 594
mechanics, 329
harmonic analysis, 409
non-commutative, 411
harmonic chain, 65
harmonic form, 223
on a Lie group, 271
Hausdorff, 25, 122
space, 134, 373
Hausdorff space, 23
Hecke algebra, 361, 399
hedgehog theorem, 164
Heine-Borel lemma, 21
Heisenberg group, 329
Helmholtz-Korteweg lagrangian, 434
hessian, 445
higher homotopy groups, 118
Hilbert space, 19, 32, 103, 380, 385,
409
Hilbert-Einstein lagrangian, 435
Hodge star operator, 218, 219
Hodge theorem
manifolds with boundary, 237
manifolds without boundary, 236
holonomic
basis, 176, 209
holonomy group, 314
homeomorphism, 31, 35, 75, 137
homogeneous
space, 316, 621
homology, 68, 78
class, 68
cubic, 72
group, 68, 69, 71, 228, 233
integer, 64, 234
real, 64, 234
singular, 60
theory, 47
homomorphism, 70, 335, 639
induced, 118
homothecies, 30
665
homotopic
path, 78
functions, 73, 116
homotopy, 73, 80
class, 74, 85, 89, 116, 153
classification of bundles, 319
group, 71, 91
first, 86
n-th, 117
path, 78
theory, 47
type, 75
homotopy formula, 205
Hopf, 153
Hopf algebra, 323, 413
defined, 348
Hopf-Banach algebra, 326
Hopf-von Neumann algebra, 413
horizontal
curve, 292
field, 286
form, 310
lift
of a curve, 292
of a vector field, 292
horocycle, 470
Hurewicz, 116, 120
hyperbolic space, 618
hyperbolic type operator, 556
hyperfinite algebra, 396
hypersurface, 231, 559
ideal, 639
idempotent, 341, 343
image, 335
imbedded submanifold, 138
imbedding, 138, 140, 450
immersed submanifold, 138
immersion, 138
improper subset, 6
incidence number, 55
666
inclusion, 639
indefinite metric, 185
index
Franck, 550
Morse, 446
of a curve, 440
of a singular point, 442, 457
of a subgroup, 337
index ellipsoid, 573
indiscrete topology, 7, 14, 29
indistinguishability, 41
induced bundle, 302, 319
induced topology, 11, 75
inertia
ellipsoid, 184
tensor, 184
infinite-dimensional spaces, 22
infinitesimal operator, 168
infinitesimal variation, 417
injective, 639, 640
inner product, 19, 220, 234, 343
defined, 343
space, 19, 379, 380
instanton, 22, 222, 596
integrability, 199, 630
integrable, 101
forms, 192
integral, 226
integral curve, 166
integral domain, 640
integrating denominator, 194
integration, 374
interference experiment, 129
interior, 16
interior product, 240
interval I = [0,1], 20
invariance
local, 585
invariant
measure, 408
form on a Lie group, 271
INDEX
polynomial, 71, 367, 398
space, 408
subgroup, 337
inverse image, 28
inverse Poincare lemma, 205
invertible element, 346
involution, 631
of algebra, 345, 347
of vector space, 343
involutive
algebra, 387
irrational numbers, 11
Ising model, 527, 531, 536
isolated point, 16
isometry, 30, 31, 185
isomorphism, 640
canonical, 164, 184
natural, 164
isotopy, 638
(or isotopic deformation), 366
Jacobi
equation, 467, 620
field, 417
identity, 163, 177, 346, 631
Jones, 369, 394
index, 397, 532
polynomial, 398
Konigsberg bridges, 54
Kac algebras, 413
Kauffman, 368, 369
kernel, 67, 335, 640
Kervaire, 134
Killing
equation, 186
field, 186
form, 270
vector, 186, 616
kink, 34
Kirillov form, 629
INDEX
667
Klein bottle, 40
algebra, 163, 210, 239, 266, 346,
350
Klein-Gordon equation, 557, 577, 609
bracket, 239, 346
knot, 531
derivative, 185, 242
group, 366
functional, 436
theory, 534
properties, 172
Kolmogorov capacity, 124
group, 134, 165, 276
Koszul formula, 243
Lie
derivative,
169
Kronecker symbol, 194, 217
lift, 109, 277
Lagrange
light cone, 29
derivative, 419, 580, 583, 584
light ray, 566
theorem, 337
limit point, 16
lagrangian, 580
limiting circle, 470
existence of, 596
line, 11
Laplace-Beltrami operator, 223, 274, line field, 193
609
Linear
laplacian, 221, 222, 245
independence, 216
in homology, 65
linear
large group, 599
atlas, 128
lattice
complement, 343
Bethe, 55, 535
connection, 26, 286, 605
models, 526
dependence, 342
parameter, 526
frame, 278
Lax pair, 630
functional, 422
Lebesgue measure, 35, 373
group, 43
left-regular representation, 408
independence, 342
Leibniz
operator, 239, 343, 423
rule, 145
representation, 214, 626
lemma
space, 380
Cartan, 450
linear space
inverse Poincare, 205
defined, 341
Morse, 445
link, 365, 367
Poincare, 198
Liouville
length
equation, 554
of curve, 185
theorem, 168, 375
of vector, 185
liquid crystal, 547
Lenz, 527
Lobachevsky, iv
letter, 352
plane, 466
Levi-Civita connection, 295, 611
local
Lichnerowicz bracket, 211
compactness, 22, 34, 373
Lie
coordinate, 125
668
homeomorphism, 31
invariance, 584
system of coordinates, 125
transformation, 583
locally
euclidean space, 121
locally finite covering, 26
logistic map, 166
loop (curve), 32, 85
higher-dimensional, 116
of loops, 120
loop (in a graph), 51
loop space, 277
loop(curve)
differentiable, 136
Lorentz
group, 576
group, 577, 616
metric, 4, 13, 27, 181, 185, 618
singlet, 582
transformation, 187, 581
Lorenz gauge, 224, 601
LSC, 125, 147
Luneburg, 554, 558
Lyapunov exponent, 376
Mobius band, 38, 94, 274
magnetic
field, 202, 206
monopole, 126, 153
magnetization, 529
manifold
analytic, 134
imbedded, 138
immersed, 138
Poisson, 628
strictly homogeneous, 628
symplectic
homogeneous, 625
manifolds
diffeomorphic, 137
INDEX
manifolds-with-boundary, 11, 122
Manin space, 325
map, 166
Markov, 137
Maurer-Cartan
equation, 209, 601
form, 604
maximal
atlas, 133
ideal, 640
maximum, 446
Maxwell
equations, 558
equations, first pair, 202
equations, second pair, 223
reciprocal relations, 204
measurable
space, 372
subsets, 373
measure, 20, 35, 340, 382, 523
finite, 373
positive, 373
space, 373
measurement, 185
mechanical
work, 191
metric, iii, 4, 148, 164, 605, 640
conformally equivalent, 618
conformally flat, 570
defined, 181
indefinite, 185
Lorentz, 185
refractive, 568
Riemannian, 185
semi-Riemannian, 185
metric space, iii, 8, 10, 25, 240, 638
metric topologies, 8
metric-indecomposable, 377
metrically transitive, 377
metrizable space, 25, 26, 59, 382, 640
Milnor, 133
INDEX
minimum, 446
Minkowski space, 4, 13, 181, 186, 221,
615
forms in, 224
mixed tensor, 156
Mixing flow, 375
module
defined, 341
momentum
Souriau, 628, 629
monodromy, 102
theorem, 111
monoid, 335
monoid diagrams, 368, 399
monomorphism, 640
monopole, 119
Morse
index, 446
inequality, 447
lemma, 445
theory, 445
motion, 30, 185, 576
moving frame, 214, 450
Moyal bracket, 329, 392
multiply-connected
space, 98
graph, 54
space
defined, 89
wavefunctions, 129
wavefunctions on, 102
natural
basis, 147, 149
isomorphism, 164
neighbourhood, 8
nematic system, 548
network, 10
Newton, iii
nilpotent group, 338
Noethers theorem, 35, 628
669
first, 580
second, 584
non-coordinate basis, 176
non-degenerate critical point, 445
non-Hausdorff space, 24
non-integrable phase factors, 102
non-linear representation, 269, 626
non-metric topological space, 15
non-metric topological spaces, 373
non-orientable manifolds, 129
non-potential force, 191
noncommutative algebra, 154
noncommutative geometry, 327, 392
noncommutative geometry,, 323
norm, 19, 34, 380, 386
norm topology, 19, 380
normal
bundle, 273
coordinates, 612
operator, 389
space, 25
subgroup, 337
normed
ring, 387
vector space, 22, 42, 380
nowhere dense subset, 17
O(n), 44
one-parameter
group, 167
one-to-one, 640
Onsager, 529
onto, 640
open r-balls, 3
open ball, 4, 11
open set, 5, 6, 24
operation, internal, 640
operator, 386
coboundary, 64
continuous, 423
defined, 423
670
laplacian, 65
types of, 556
optical indicatrix, 573
optical length, 570
optics, 554, 556, 565
orbit, 167, 268, 630
order
of a group, 334
of an element, 354
oricycle, 470
orientation, 129, 191
of a simplex, 61
orthogonal
group, 281
sequence, 381
system, 381
parabolic type operator, 556
paracompact, 26, 314, 374
parallel
lines, 459
transport, 284, 289, 293, 294
parallelism, 312
parallelizable
manifold, 164, 276, 287
parallelogram rule, 379
parameterization, 126
parity, 202
partial differential equation, 554
partition
function, 524, 525
of a space, 107
of the identity, 374
path, 32, 54, 74, 78
closed, or loop, 54
Euler, 54
simple, 54
path homotopy, 78
path-component, 34, 74, 80
path-connectedness, 13, 34, 74, 87
path-covering, 109
INDEX
Pauli matrices, 211
paving, 522
percolation, 103
permutation, 354, 362
Peter-Weyl theorem, 412
Pfaffian
equation, 193
form, 190, 197
phase space, 168
enlarged, 168
symmetries, 625
phase transformation, 582
phase transition, 529
piecewise-linear manifold, 129
planar graphs, 51
plastic crystals, 548
Poincare, 116, 367
group, 576
duality, 69
group, 31, 187, 616
inverse lemma, 205
lemma, 63, 64, 198, 569
polynomial, 71
space, 466
theorem du retour, 168
theorem on integration, 228
Poincare
conjecture, 104
Poinsot
construction, 572
ellipsoid, 184
point
critical, 445
critical non-degenerate, 445
maximum, 446
minimum, 446
regular, 445
saddle, 446
singular, 439
pointwise product, 340, 391
Poisson
INDEX
bracket, 330
bracket, 326, 392
equation, 236
manifold, 628
structure, 391
polarization, 182
polyhedron, 226
of a complex, 59
polynomial algebra, 388
polynomial function, 157
Pontryagin
class, 320
duality, 411
positive measure, 373
potential, 101
potential form, 203, 224
Potts model, 55, 531
power set, 14, 27, 372, 641
Poynting vector, 558, 566
prametric, 15, 641
predual, 390
presentation, 353
principal curvature radii, 456
principal fiber bundle, 179, 317
probability
space, 373
theory, 373
product topology, 14
projection, 107, 161
projective
group, 44
line, 45, 97
plane, 45, 97, 113
representation, 626
space, 44
transformation, 44
projector, 44, 341, 343
proper subset, 6
pull-back, 159, 206, 451
pull-back bundle, 319
pure braid group, 357
671
pure state, 45
push-foward, 159
Pythagoras theorem, 379
quadratic form
first, 451
second, 451
quantization, 39, 100, 330, 385
quantum group, 324, 350, 392, 413
Quantum Mechanics, 92, 129
quantum mechanics, 326
quantum space, 325
quaternions, 43
quotient
space, 37, 621
topology, 37, 75, 107
rank
of a finitely generated group, 334
of a function, 138
rational numbers, 11
reduction
of bundle, 282, 302
refinement, 13, 29
of a covering, 26
refractive index, 571
regular point, 445
regular values, 152
relation
defined, 641
equivalence, 641
relative
topology, 11
compactness, 22
relator, 353
repère mobile, 215
adjoint, 265
of a group, 266
of an algebra, 266
coadjoint, 267, 630
672
completely reducible, 407
defined, 403
dimension of, 405
equivalent, 405
faithful, 297, 404
generalities, 265, 403
irreducible, 406
linear, 404
non-linear, 269
projective, 626
reducible, 406
regular, 267, 408
right-regular, 408
ring, 411
singlet, 407
space, 404
theory, 404
trivial, 404
unitary, 405
resolvent, 387
restricted holonomy group, 314
retraction, 641
Ricci, 620
equation, 453
tensor, 454, 607
theorem, 606
Riemann, iv, 619
inaugural address, iv
integral, 226
sheet, 106
surface, 105
tensor, 453, 607
Riemannian, 570
manifold, 185
metric, 185, 220, 282, 374
Riesz theorem, 22
right-regular representation, 408
ring
defined, 338
of a group, 339
of subsets, 372
INDEX
RP1 , 97
RP2 , 97, 113
RP3 , 114
RPn , 113
S1 , 32, 75
S2 , 36, 127, 446, 458
S4 , 237
S7 , 137
saddle-point, 446
scalar
curvature, 607
invariant, 607
scalar field, 577
Schwartz, 234
Schwarzschild
radius, 126
space, 450
second quadratic form, 451
second-countable, 11, 25, 122
topology, 10
second-separable, 23
section, 162, 276
self-adjoint, 389, 637
algebra, 387
self-dual forms, 222, 596
semigroup, 335
seminorm, 380
semisimple algebra, 271
semisimple group, 337
separability, 23
first, 23
second, 23
separable space, 17, 23
separated space, 23
sequence, 18, 21, 31
Cauchy, 18
set function, 373
countably additive, 373
finitely additive, 373
positive, 373
INDEX
set product, 372
signature, 220
simple
group, 337
simplex, 58, 226
simply-connected, 88
graph, 54
Sina theorem, 376
singular chain, 226
singular point, 439, 442, 457
sink, 444
skein relations, 361, 368
smectic
phase, 547
smooth function, 446
SO(3), 114
SO(n), 44
soldering, 280
form, 280
solvable group, 338
Sorgenfrey line, 25
source, 444
Souriau momentum, 628, 629
space, 6
compact, 20
concept of, iii, 4, 129
conformally flat, 618
contractible, 75
cotangent, 147
Dirac, 19
Hilbert, 19
hyperbolic, 618
inner product, 19
separable, 23
separated, 23
tangent, 147
vector, 19
spaces of paths, 35
spacetime, 29, 616
spacetime topology, 13
specific heat, 529
673
spectral radius, 387
spectrum, 387
sphere, 235, 276, 458
sphere S 7 , 133
sphere S n , 133
sphere S n , 11
sphere S7 , 137
spherical top, 115
spin current, 581, 582
spinor, 613
field, 578
spontaneous breakdown of symmetry,
530
standard field, 287
star-operation, 218
star-product, 329
starshaped sets, 380
state
of a physical system, 385, 524
stereographic projection, 36, 618
Stiefel manifold, 46, 317, 319
Stokes theorem, 228
straight line, 380
strain tensor, 544
strange attractor, 166
strong
derivative, 419, 423
differential, 423
topology, 380
stronger topology, 13
structure
coefficients, 176
structure equations, 311, 450
En , 215
structure group, 179, 275, 278
su(2) algebra, 211
SU(2) group, 115, 211
sub-bundle, 302
subgroup, 337
topological, 42
submanifold, 137
674
subring, 338
subsets, 15
improper, 6
proper, 6
superconductor, 100
superfluid, 100
superselection rules, 34
supersymmetry, 245
surfaces, 4, 21, 455, 456
surgery, 104
surjective, 639, 641
syllable, 352
symmetric, 641
algebra, 387
symmetric group, 354
symmetric tensor, 156
symmetry, 577
properties, 40
and anomaly, 625
of a lagrangian, 436
on phase space, 625
transformation, 37
system
of coordinate functions, 125
Tn , 40
tail, 18
tame braid, 357
Tanaka-Krein duality, 412
tangent
bundle, 162, 273
field, 166
space, 147, 154
vector, 146
Temperley and Lieb, 398
tensor, 154, 190
antisymmetric, 157
bundle, 163
components, 156
contraction, 156
contravariant, 155
INDEX
covariant, 154
field, 164
mixed, 156
product, 154, 344
of representations, 407
symmetric, 156
tensor product
defined, 344
tetrad, 282
trivial, 588
tetrad field, 283
tetrahedron, 52, 65, 66, 93
theorem
calculating, 93
billiard, 376
bundle classification, 319
du retour, 168
ergodic, 376
hedgehog, 164
thermodynamics, 204
tiling, 522
topological
number, 22
conservation laws, 34
defect, 47
dimension, 123
group, 22, 41
invariant, 35, 64, 90, 641
manifold, 121
number, 15, 47, 61, 64
product, 14, 39
space, 4
defined, 7
vector space, 382
vector space., 35
topology, 4, 31
basis, 9
coarser, 13
compact-open, 33
defined, 6
discrete, 7, 14
INDEX
finer, 13
general, 3
indiscrete, 7, 14
induced, or relative, 11
norm, 19, 380
quotient, 36
strong, 380
stronger, 13
trivial, 14
uniform, 380
weaker, 13
torsion, 280, 288, 290, 606, 612
-free group, 354
group, 354
subgroup, 354
torus, 40, 69, 94, 108
trace, 398
transfer matrix, 528, 531, 533
local, 532
transformation
conformal, 187
Lorentz, 187
transformation group, 335, 404
transgression, 205
transition function, 127
transition functions, 301
translation, 30, 581
tree
Cayley, 55, 535
tree graph, 51
triangular inequality, or sub-additivity,
379
triangulation, 59, 69, 72, 444
trivial
bundle, 164
topology, 14
trivialization, 298, 304
trivializer, 335
twisted field, 274
type
675
of a factor algebra, 395
of a group, 409
typical fiber, 274
ultrafilter, 639
uniform topology, 380
unimodular group, 408
uniqueness of solutions, 24
unit
algebra, 345, 387
ring, 338, 372
universal
bundle, 317
covering space, 107
enveloping algebra, 347, 350
universality of gravitation, 609
universe
Friedmann model, 229
upper-half
plane, 12, 25
space, 11
Urysohns theorem, 25
vacuum, 119, 530
van Hove theorem, 529
variation, 78, 416
vector, 145, 146
bases, or frames, 175
contravariant, 147
covariant, 147
field, 161, 276, 391, 578, 582
space, 8, 19, 76, 77
normed, 379
vector fibered atlas, 275
vector space
defined, 341
vector-valued form, 150, 210
vertex, 50
vertical
field, 280, 286
form, 310
676
space, 294, 303, 307
volume form, 191, 220
canonical, 218
of a hypersurface, 231
von Neumann
algebra, 124, 369, 393, 409, 413
bicommutant theorem, 394
ergodic theorem, 377
W*-algebra, 393
Wang
theorem, 621
wave equation, 224
wave front, 566
wavefunction, 100, 101, 106, 114
weak
derivative, (or Gateaux derivative),
423
differential, 423
topology, 382
weaker topology, 13
wedge product, 190
Weil, 458
Wess-Zumino
condition, 602, 627
Weyl
prescription, 328, 353
Weyl-Wigner picture of Quantum Mechanics, 328, 392
Whitney theorem, 134, 139, 215, 449
on metric, 185
wiedersehen manifold, 571
Wigner
function, 328
Wigner-Inon
u
group contraction, 621
winding number, 119, 153, 550, 597
word, 352
group, 352
work, 230
INDEX
as a differential form, 191
Yang-Baxter equation, 324, 350, 362
classical, 631
Yang-Mills
equation, 224, 237, 583, 593, 596
Young double-slit experiment, 129
Zeeman, 31
topology for spacetime, 13, 14, 29
zero section, 300

An Introduction To Geometrical Physics (R. Aldrovandi, J. G. Pereira)

Uploaded by

Copyright:

Available Formats

An Introduction To Geometrical Physics (R. Aldrovandi, J. G. Pereira)

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Introduction To Geometrical Physics (R. Aldrovandi, J. G. Pereira)

Uploaded by

Copyright:

Available Formats

An Introduction to

PREAMBLE: SPACE AND GEOMETRY

What stuff tis made of, whereof it is born,

This would define a different metric space, in principle as good as that

Aristotle, Physics I.1.

4 MANIFOLDS & CHARTS

1 THE BASIC ALGEBRAIC STRUCTURES

2 DISCRETE GROUPS. BRAIDS AND KNOTS

Function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 383

10 EUCLIDEAN SPACES AND SUBSPACES

2.3.9 Left and rightinvariant fields . . . . . . . . . . . . . 515

6 CLASSICAL RELATIVISTIC FIELDS

10 SYMMETRIES ON PHASE SPACE

Glossary and Bibliography

1.0.1 Let us again consider our ambient 3-dimensional euclidean space

CHAPTER 1. GENERAL TOPOLOGY

1.0.2 There is more: physicists are used to metrics which in reality do

1.1. TOPOLOGICAL SPACES

with spacial qualitative relationships, and another, more particular, naming

1.1.1 Experimental measurements being inevitably of limited accuracy,

A commendable text for beginners, proceeding constructively from unstructured sets

CHAPTER 1. GENERAL TOPOLOGY

1.1. TOPOLOGICAL SPACES

Thus, a topology on S is a collection of subsets of S to which belong the

CHAPTER 1. GENERAL TOPOLOGY

1.1.4 Given a point p S, any set U containing an open set belonging

1.1. TOPOLOGICAL SPACES

1.1.8 Topology basis In order to specify a topological space, one has to

CHAPTER 1. GENERAL TOPOLOGY

We see here a first trial to classify topological spaces. Topology frequently

1.1. TOPOLOGICAL SPACES

1.1.9 Induced topology The topologies of the usual surfaces immersed

1.1.10 The upper-half space En+ . The point set is

The topology is that induced by the ball-topology of En . This space, which

CHAPTER 1. GENERAL TOPOLOGY

set R2+ with the open disks of E2 .

1.1. TOPOLOGICAL SPACES

CHAPTER 1. GENERAL TOPOLOGY

1.2. KINDS OF TEXTURE

CHAPTER 1. GENERAL TOPOLOGY

1.2. KINDS OF TEXTURE

1.2.2 Let U and V be two subsets of a topological space S. The subset

See Kolmogorov & Fomin 1970 and/or Christie 1976.

CHAPTER 1. GENERAL TOPOLOGY

E is closed because it is the complement of a union of open sets. Its interior

On a general topological space, it may happen that a sequence converges

1.2. KINDS OF TEXTURE

with suitable properties. In this case the number

CHAPTER 1. GENERAL TOPOLOGY

is a separable Hilbert space. Historical evolution imposed the consideration

the summations i=1 used in the expansions become integrals. As said in

1.2. KINDS OF TEXTURE

CHAPTER 1. GENERAL TOPOLOGY

recently interested in (such as the existence of integer topological numbers)

Coleman 1977; Atiyah et al. 1978; Atiyah 1979.

1.2. KINDS OF TEXTURE

1.2.14 Separability Compactness imposes, as announced, a limitation

CHAPTER 1. GENERAL TOPOLOGY