9-Richtmyer - Principles of Advanced Mathematical Physics II

Texts and
Monographs
in Physics
w. BeiglbOck
M. Goldhaber
E. H. Lieb
W. Thirring
Series Editors
Robert D. Richtmyer
Principles of Advanced
Mathematical Physics
Volume II
With 60 Figures
ill Springer-Verlag
New York Heidelberg Berlin
Robert D. Richtmyer
Department of Mathematics
University of Colorado
Boulder, Colorado 80309
USA
Editors:
Wolf Beiglbock Maurice Goldhaber
Institut fUr Angewandte Mathematik Department of Physics
Universitat Heidelberg Brookhaven National Laboratory
1m Neuenheimer Feld 5 Associated Universities, Inc.
D-6900 Heidelberg 1 Upton, NY 11973
Federal Republic of Germany USA
Elliott H. Lieb Walter Thirring

Department of Physics Institut fUr Theoretische Physik
Joseph Henry Laboratories der Universitat Wien
Princeton University Boltzmanngasse 5
P.O. Box 708 A-I090 Wi en
Princeton, NJ 08540 Austria
USA
Library of Congress Cataloging in Publication Data
Richtmyer, Robert D
Principles of advanced mathematical physics.
(Texts and monographs in physics)

Bibliography: v. 1, p.
Includes indexes.
1. Mathematical physics. 1. Title.
QC20.R56 530.1'5 78-16494
(v. 2) AACRI
1981 by Springer-Verlag Inc.

Softcover reprint of the hardcover 1st edition 1981
All rights reserved. No part of this book may be translated or reproduced

in any form without written permission from Springer-Verlag, 175 Fifth
Avenue, New York, New York 10010, U.S.A.
9 8 76 54 3 2 1
ISBN 978-3-642-51078-6 ISBN 978-3-642-51076-2 (eBook)

DOl 10.1007/978-3-642-51076-2
Contents
Preface to Volume II XI
18 Elementary Group Theory 1

18.l The group axioms; examples
18.2 Elementary consequences of the axioms; further definitions 3
18.3 Isomorphism 5
18.4 Permutation groups 6
18.5 Homomorphisms; normal subgroups 8
18.6 eosets 10
18.7 Factor groups 10
18.8 The Law of Homomorphism 1J
18.9 The structure of cyclic groups II
18.10 Translations, inner automorphisms 12
18.l1 The subgroups of /1'4 13
18.l2 Generators and relations; free groups IS
18.13 Multiply periodic functions and crystals 16
18.l4 The space and point groups 17
18.15 Direct and semidirect products of groups; symmorphic space
groups 20
19 Continuous Groups 25
19.1 Orthogonal and rotation groups 25
19.2 The rotation group SO(3); Euler's theorem 27
19.3 Unitary groups 28
v
vi Contents
19.4 The Lorentz groups 29

19.5 Group manifolds 34
19.6 Intrinsic coordinates in the manifold of the rotation group 35
19.7 The homomorphism of SU(2) onto SO(3) 37
19.8 The homomorphism of SL(2, q onto the proper Lorentz
group ~p 38
19.9 Simplicity of the rotation and Lorentz groups 38
20 Group Representations I: Rotations and Spherical Harmonics 40

20.1 Finite-dimensional representations of a group 41
20.2 Vector and tensor transformation laws 41
20.3 Other group representations in physics 44
20.4 Infinite-dimensional representations 45
20.5 A simple case: SO(2) 46
20.6 Representations of matrix groups on Xoo 47
20.7 Homogeneous spaces 48
20.8 Regular representations 49
20.9 Representations of the rotation group SO(3) 50
20.10 Tesseral harmonics; Legendre functions 53
20.11 Associated Legendre functions 55
20.12 Matrices of the irreducible representations of SO(3); the
Euler angles 57
20.13 The addition theorem for tesseral harmonics 59
20.14 Completeness of the tesseral harmonics 60
21 Group Representations II: General; Rigid Motions;

Bessel Functions 62
21.1 Equivalence; unitary representations 62
21.2 The reduction of representations 63
21.3 Schur's Lemma and its corollaries 65
21.4 Compact and noncompact groups 66
21.5 Invariant integration; Haar measure 67
21.6 Complete system ofrepresentations of a compact group 71
21.7 Homogeneous spaces as configuration spaces in physics 72
21.8 M 2 and related groups 73
21.9 Representations of M 2 73
21.10 Some irreducible representations 74
21.11 Bessel functions 75
21.12 Matrices of the representations 76
21.13 Characters 77
22 Group Representations and Quantum Mechanics 80

22.1 Representations in quantum mechanics 80
22.2 Rotations of the axes 81
22.3 Ray representations 82
22.4 A finite-dimensional case 83
Contents vii
22.5 Local representations 83

22.6 Origin of the two-valued representations 84
22.7 Representations of SU(2) and SL(2, IC) 85
22.8 Irreducible representations of SU(2) 87
22.9 The characters of SU(2) 89
22.10 Functions of z and z 89
22.11 The finite-dimensional representations of SL(2, IC) 90
22.12 The irreducible invariant subspaces of xro for SL(2, IC) 92
22.13 Spinors 93
23 Elementary Theory of Manifolds 96

23.1 Examples of manifolds; method of identification 96
23.2 Coordinate systems or charts; compatibility; smoothness 98
23.3 Induced topology 101
23.4 Definition of manifold; Hausdorff separation axiom 101
23.5 Curves and functions in a manifold 103
23.6 Connectedness; components of a manifold 104
23.7 Global topology; homotopic curves; fundamental group 105
23.8 Mechanical linkages: Cartesian products 111
24 Covering Manifolds 114

24.1 Definition and examples 114
24.2 Principles of lifting 117
24.3 Universal covering manifold 119
24.4 Comments on the construction of mathematical models 121
24.5 Construction of the universal covering 123
24.6 Manifolds covered by a given manifold 125
25 Lie Groups 129

25.1 Definitions and statement of objectives 130
25.2 The expansions of m( " . ) and I( " .) 132
25.3 The Lie algebra of a Lie group 133
25.4 Abstract Lie algebras 135
25.5 The Lie algebras of linear groups 135
25.6 The exponential mapping; logarithmic coordinates 136
25.7 An auxiliary lemma on inner automorphisms; the mappings Ad p 139
25.8 Auxiliary lemmas on formal derivatives 141
25.9 An auxiliary lemma on the differentiation of exponentials 143
25.10 The Campbell-Baker-Hausdorf (CBH) formula 144
25.11 Translation of charts; compatibility; G as an analytic manifold 146
25.12 Lie algebra homomorphisms 149
25.13 Lie group homomorphisms 151
25.14 Law of homomorphism for Lie groups 155
25.15 Direct and semidirect sums of Lie algebras 160
25.16 Classification of the simple complex Lie algebras 162
25.17 Models of the simple complex Lie algebras 167
25.18 Note on Lie groups and Lie algebras in physics 170
Appendix to Chapter 25-Two nonlinear Lie groups 171
VlIl Con ten ts
26 Metric and Geodesics on a Manifold 174

26.1 Scalar and vector fields on a manifold 175
26.2 Tensor fields 180
26.3 Metric in Euclidean space 182
26.4 Riemannian and pseudo-Riemannian manifolds 183
26.5 Raising and lowering of indices 185
26.6 Geodesics in a Riemannian manifold 186
26.7 Geodesics in a pseudo-Riamannian manifold 9Ji 190
26.8 Geodesics; the initial-value problem; the Lipschitz condition 190
26.9 The integral equation; Picard iterations 192
26.10 Geodesics; the two-point problem 193
26.11 Continuation of geodesics 194
26.12 Affinely connected manifolds 195
26.13 Riemannian and pseudo-Riemannian covering manifolds 197
27 Riemannian, Pseudo-Riemannian, and Affinely

Connected Manifolds 198
27.1 Topology and metric 199
27.2 Geodesic or Riemannian coordinates 199
27.3 Normal coordinates in Riemannian and pseudo-Riemannian
manifolds 202
27.4 Geometric concepts; principle of equivalence 203
27.5 Covariant differentiation 206
27.6 Absolute differentiation along a curve 208
27.7 Parallel transport 209
27.8 Orientability 210
27.9 The Riemann tensor, general; Laplacian and d'Alembertian 211
27.10 The Riemann tensor in a Riemannian or pseudo-Riemannian
manifold 214
27.11 The Riemann tensor and the intrinsic curvature of a manifold 216
27.12 Flatness and the vanishing of the Riemann tensor 218
27.13 Eisenhart's analysis of the Stackel systems 221
28 The Extension of Einstein Manifolds 223

28.1 Special relativity 223
28.2 The Einstein gravitational field equations 224
28.3 The Schwarzschild charts 227
28.4 The Finkelstein extensions of the Schwarzschild charts 231
28.5 The Kruskal extension 233
28.6 Maximal extensions; geodesic completeness 235
28.7 Other extensions of the Schwarzschild manifolds 235
28.8 The Kerr manifolds 237
28.9 The Cauchy problem 240
28.10 Concluding remarks 243
29 Bifurcations in Hydrodynamic Stability Problems 244

29.1 The classical problems of hydrodynamic stability 244
29.2 Examples of bifurcations in hydrodynamics 245
29.3 The Navier-Stokes equations 247
Contents IX
29.4 Hilbert space formulation 248

29.5 The initial-value problem; the semiflow in,5 248
29.6 The normal modes 249
29.7 Reduction to a finite-dimensional dynamical system 250
29.8 Bifurcation to a new steady state 254
29.9 Bifurcation to a periodic orbit 255
29.10 Bifurcation from a periodic orbit to an invariant torus 257
29.11 Subharmonic bifurcation 261
Appendix to Chapter 29-Computational details for the invariant torus 261
30 Invariant Manifolds in the Taylor Problem 263

30.1 Survey of the Taylor problem to 1968 263
30.2 Calculation of invariant manifolds 265
30.3 Cylindrical coordinates 268
30.4 The Hilbert space 270
30.5 Separation of variables in cylindrical coordinates 27l
30.6 Results to date for the Taylor problem 272
Appendix to Chapter 30-The matrices in Eagles' formulation 274
31 The Early Onset ofTurbulence 276

31.1 The Landau~Hopfmodel 276
31.2 The Hopf example 278
31.3 The Ruelle~ Takens model 279
31.4 The w-limit set of a motion 280
31.5 Attractors 282
31.6 The power spectrum for motions in [Rn 283
31.7 Almost periodic and aperiodic motions 284
31.8 Lyapounov stability 285
31.9 The Lorenz system; the bifurcations 286
31.10 The Lorenz attractor; general description 288
31.11 The Lorenz attractor; aperiodic motions 290
31.12 Statistics of the mapping! and 9 293
31.13 The Lorenz attractor; detailed structure I 294
31.14 The symbols [i,j] of Williams 297
31.15 Prehistories 299
31.16 The Lorenz attractor; detailed structure II 300
31.17 Existence of I-cells in F 301
31.18 Bifurcation to a strange attractor 302
31.19 The Feigenbaum model 303
Appendix to Chapter 3 I (Parts A~H)-Generic properties of systems: 304
31.A Spaces of systems 304
31.B Absence of Lebesgue measure in a Hilbert space 304
31.C Generic properties of systems 305
31.D Strongly generic; physical interpretation 305
31.E Peixoto's theorem 306
31.F Other examples of generic and nongeneric properties 306
31.G Lack of correspondence between genericity and Lebesgue
measure 308
31.H Probability and physics 308
References 313
Index 317
Preface to Volume II
The first eleven chapters in this volume, 18 through 28, contain material
that was developed in the third year of the three-year mathematical physics
sequence at the University of Colorado. The central concepts are groups,
manifolds, and differential geometry. I wish to thank Professors Wesley
Brittin and Russel Dubisch for extensive discussions of this material, and
I wish to thank Professor Wolf Beiglbock for advice and suggestions on the
overall plan and on the material on group representations.
The material in the last three chapters, related broadly to recent work in
differentiable dynamical systems, has been discussed in special courses on
hydrodynamic stability and seminars on mathematical physics. That
material is somewhat less well organized than the older subjects, but has
been included because it contains various concepts of great potential value
in physical science.
Boulder, August 1981 Robert D. Richtmyer

CHAPTER 18
Elementary Group Theory
The group axioms; Abelian group; cyclic group; subgroup; order;

isomorphism; homomorphism; automorphism; permutation; symmetric
group; cycle; transposition; parity; alternating group; kernel of a
homomorphism; normal subgroup; simple group; conjugate elements; cosets;
Lagrange's theorem; factor group; law of homomorphism; translations; inner
automorphisms; Cayley's theorem; conjugate subgroups; simplicity of
.9115; composition series; Jordan-Holder theorem; generators and relations;
free group; free abelian group; the word problem; space and point groups;
direct and semidirect product; symmorphic space groups.
Prerequisite: Elementary algebra.
This chapter contains a survey of elementary group theory. For applications

in later chapters, the high point is the law of homomorphism and the concepts
associated with it.
18.1 The Group Axioms; Examples
A group G is any set or collection of elements {a, b, c, ... , x, y, z, ... }, finite

or infinite, together with a law of composition, denoted by 0, such that:
I. If a and b are any two elements of G, then a b is an element of G.

0
11. If a, b, and c are any three elements of G, then (a 0 b) 0 c = a 0 (b 0 c)

(associative law).
Ill. If a and b are any two elements of G, then there exist unique elements
x and y in G such that a x = band y a = b.
0 0
If the elements are numbers, matrices, quaternions, etc., the composition

a 0 b may be either the sum or the product of a and b; in the examples below,
the word "under" is used to identify the law of composition. In the case of
mappings, transformations, rotations, permutations, etc., the law is under-
stood as the usual law of composition; if a and b are transformations, then
a b is the transformation that results from performing b first, then a.
0
2 Elementary Group Theory
Note. In some books, the axiom iii above is replaced by the fully equivalent
axiom that G contains a unique identity element e and that each element
a of G has a unique inverse a-I-see next section.
As a first example, let G be the set of all rotations in the plane: let R",
denote the transformation in which a point x, y is moved to, or mapped
onto, the point x', y', where
x' = x cos cP - Y sin cp,
(18.1-1)
y' = x sin cp + Y cos cpo
U the transformations R"" and R"'2 are performed in succession, the result
is a rotation through the angle CPI + CP2, i.e., it is the transformation R"" +"'2'
It is easily verified that the set {R",: 0 s cP < 2n} of all such rotations satisfies
the group axioms.
A rotation in 3 dimensions may be described by first choosing a direction
through the origin and then performing a rotation through some angle about
that direction as a fixed axis. It follows from Euler's theorem, proved in
Section 19.2, below, that the resultant of two such transformations, per-
formed in succession, is another such, i.e., is a rotation through some angle
about some axis. [This seems evident (because everyone knows that it is true)
until one tries to prove it.] In consequence, the set of all rotations in 3
dimensions is a group. The group of all rotations in n dimensions is denoted
by SO(n), for reasons that will appear.
As a third example, consider the set of all rotations in 3 dimensions
under which a cube, centered at the origin, is invariant (i.e., is mapped into a
cube that coincides with the original cube). One can rotate the cube through
90, 180, or 270 about an axis through the midpoints of opposite faces,
through 1800 about an axis through the midpoints of opposite edges, or
through 1200 or 240 about an axis through opposite vertices. It is easily
verified that these transformations (including the identity transformation)
form a group of 24 elements. More generally, the set of all transformations
of a specified kind (e.g., rotations, general linear transformations, rigid
motions, conformal mappings) under which a given figure is invariant is a
group, because the figure is clearly invariant under composition and inverses
of such mappings. The rigid motions under which a crystal lattice is invariant
constitute the space group of the crystal-see Section 18.13.
The set of all permutations of n objects is a group; such groups are dis-
cussed in Section 18.4.
Certain sets of real or complex numbers or quaternions are groups with
respect to addition or multiplication, e.g., the set of all integers (positive,
negative, and zero) under addition, the set of all positive real numbers under
multiplication, the integers 0, 1, ... , m - 1 under addition modulo m, or
the set of all nonzero (real) quaternions under multiplication.
When addition is the rule of composition, a 0 b is denoted by a + b, the
inverse of a by - a, and the identity by O. Often the little circle is omitted
and the composition of two elements a and b is written simply as a product abo
Elementary Consequences of the Axioms; Further Definitions 3
A finite group can be fully described by its multiplication table. For example,
Klein's 4-group V4 is defined by
e a b c
e e a b c
a a e c b
b b c e a
c c b a e
which means that a b = c, etc. Each group element appears just once in
0
each row and once in each column; furthermore, all rows are different and
all columns are different. Any square arrangement of letters having this
property is called a latin square (Euler). Any latin square defines an abstract
group, provided that the multiplicative structure thus determined has an
identity and satisfies the associative law.
Abstract group theory deals with the relations indicated in the multiplica-
tion table and completely ignores the inherent nature of the elements, a, b,
etc. In contrast with calculus, real and complex analysis, differential equations,
and other subjects in analysis (group theory belongs to algebra), numerical
quantities hardly ever appear, except integers for the purposes of enumera-
tion and counting.
The theory of groups plays a role in quantum mechanics, in the theory of
spectra, in the analysis of classical dynamical systems, in the theory of auto-
morphic functions, in the theory of algebraic equations, and so on.
18.2 Elementary Consequences of the Axioms;

Further Definitions
The following laws are consequences of axioms i, ii, and iii of the preceding
section:
Law of cancellation: If a, b, c are any elements of a group G, then
a b = a c implies
0 0 b=c
and
boa = c a implies
0 b = c.
Identity: In G there is a unique element e such that a 0 e = eo a = a for
all a in G.
Inverses: If a is any element of G, there exists in G a unique element a- 1
such that a a- 1 = a-loa = e; furthermore, (a b)-l = b- 1 a- 1 .
0 0 0
Extended associative law: (a (b (c d) e = a b cod e, etc. Un-

0 0 0 0 0 0 0
necessary parentheses will be omitted from now on. Also

(aobo ... oxoy)-l = y-l ox- 1 0 ob- 1 oa- 1 .
If a is in G, and m is any integer, then am is defined as follows:

aO = e,
a 1 = a,
a 2 = a a, 0
a- m = (a-1r
Clearly, these powers all commute, and an am = an+ m. Generally, two
0
elements a and b of G are said to commute if a b = boa. If all pairs of

0
elements of G commute, G is said to be a commutative or Abelian group. If

all the elements an (n = 0, 1, 2, ... ) are distinct, then the element a is of
infinite order; otherwise, as it is easily seen, there is a smallest positive integer
I, the order of a, such that al = e; then am = e if and only if I is a divisor of m,
and every power of a is equal to one of the elements {e, a, a 2 , , d- 1 }.
A subgroup of G is a subset G' of the elements of G which is itself a group
under the same law of composition that appears in G. The rotations about
0
the z-axis constitute a subgroup of the group of rotations in 3-space. The

distinct powers of an element a constitute a subgroup called the subgroup
generated by the element a; such a subgroup is a cyclic group of finite or
infinite order. The order of a group is the number of elements in it (finite or
infinite). If G' is a subgroup of G, we write G' < G. In any case, G < G and
{e} < G. If G' - G, G' is a proper subgroup; if G' = {e}, G' is the trivial
subgroup.
MISCELLANEOUS QUESTIONS AND EXERCISES
1. What is the inverse of the element R", in SO(2)? What is the identity element?
2. Show that SO(2) is commutative, while SO(3) is not.
3. Show that the group of rotations that leave a cube invariant is of order 24, as
claimed in Section 18.1.
4. Describe the group of rotations under which a right circular cylinder is in-
variant; same for a regular icosahedron.
5. Derive the three laws at the beginning of this section from the group axioms.
6. Determine which of the following are groups:
(a) The set of all nonzero complex numbers, under multiplication.
(b) The set of all nonzero n x n matrices under multiplication.
(c) The set of all positive rational numbers, under multiplication.
(d) The set of all positive irrational numbers, under multiplication.
(e) The set of all positive algebraic numbers, under multiplication.
(f) The set of all n x n matrices under addition.
(g) The set of all n x n matrices of the form eA , under multiplication.
(h) The integers 1,2, ... , P - 1, under multiplication modulo p, p a prime.
(i) The integers 1,2, ... ,m - 1, under multiplication modulo m, m composite.
(j) The set of all vectors in E 3 , under vector addition.
Isomorphism 5
(k) The set of all nonzero vectors in 3, under vector multiplication.

(I) The set of all complex numbers z such that Iz I = 1, under multiplication.
(m) The set of all n x n unitary matrices, under addition.
(n) The set of all n x n unitary matrices, under multiplication.
(0) The set of all Mobius transformations
az + b
z----tz'=-- (ad - be =1= 0)
ez + d
in the complex plane.
18.3 Isomorphism
If there is a one-to-one mapping cp of a group G onto a group G' such that

cp(a b) 0 = cp(a) cp(b)
0 (18.3-1)
for all a and bin G, then cp is an isomorphism, and the groups are isomorphic;
in symbols, G ~ G'. [In (18.3-1), the first little circle denotes the law of
composition in G, the second that in G'.] One says that products are mapped
onto products. In this case, G and G' may be regarded as merely two different
realizations of the same abstract group.
For example, if G is the set of numbers {I, i, - 1, - i} under multiplication,
and G' is the set of matrices {I, A, B, C} under matrix multiplication, where
1= G ~), A = (
-10 ~), B= (-1 0 -~),
C = (~ -1)o '
then the mapping
cp: 1 -> I, i -> A, -1 -> B, -i -> C
is an isomorphism of G onto G'; the law (18.3-1) is easily verified for each of
the 16 possible pairs (a, b) of elements of G. For example, (-i) = (-1)0);
hence cp( - i) ought to be = cp( - 1)cp(i), i.e., C ought to = BA, which in fact
it is. It should be noted that the mapping
-1 -> B, -i -> A
is another isomorphism of G onto G'.

If G is the group of all complex numbers z such that Iz I = 1, under multi-
plication, then the mapping
elo." (cos e -sin e\

(j): -> .
SIll e cos eJ
is an isomorphism of G onto the 2-dimensional rotation group SO(2).
An isomorphism of a group onto itself is an automorphism. An example

in the group {I, A, B, C} of matrices described above is the mapping
I ~ I, A~C, B~B, C~A.
An automorphism of SO(2) onto itself is given by
(
COS 8
sin 8
- sin
cos 8
8) (-sincos 88
~ sin8).
cos 8
Any mapping cp (not necessarily one-to-one or onto) of a group G into a
group G' such that (18.3-1) is satisfied is a homomorphism. If G is the group
GL(n, C) of all n x n nonsingular complex matrices under multiplication,
then the mapping A ~ det A is a homomorphism of G onto the group of all
nonzero complex numbers under multiplication. As a second example, let G
be the group M 2 of all rigid motions in a plane, i.e., the group of all trans-
formations of the form
'E {x ~ x' = x cos 8 - y sin 8 + a,
(18.3-2)
9,a,b y ~ y' = x sin 8 + Y cos 8 + b,
where 0 :$; 8 < 2n and where a and b are arbitrary real numbers. Then the
mappmg
Teab~ ( .
,,
COS
sm 8
8 -cos
sin 8)
8
(18.3-3)
is a homomorphism of G onto SO(2) as can be seen by performing two

transformations of the form (18.3-2) in succession. Since SO(2) is a sub-
group of G (with a = b = 0), the mapping (18.3-3) may be regarded as a
homomorphism of G into itself.
18.4 Permutation Groups
A permutation is a one-to-one mapping of a set C (usually finite) of objects

or symbols onto itself. For example, if C consists of the first seven digits,
C = {I, 2, ... , 7}, then a particular permutation is the mapping n: j ~ n(j),
where the function n(j) is given by
n(l) = 7, n(2) = 3, n(3) = 1, n(4) =4
n(5) = 6, n(6) = 5, n(7) = 2.
This permutation is written in condensed notation as
1 2 3 4 5 6 7)
(18.4-1)
n = (7 3 146 5 2 '
where it is understood that each symbol in the upper row is mapped onto
the symbol below it. A cycle is a permutation which can be obtained by
arranging the symbols in a circle and mapping each symbol onto the one
following it (say clockwise) around the circle, for example, as in Figure 18.1;
Permutation Groups 7
Figure IS.1 A cyclic permutation.
this cycle is written in a still further condensed notation as (a bed e),

which is of course the same as (b c d e a), etc. Any permutation of the set C
can be expressed in terms of cyclic permutations of various subsets of C;
for example, the permutation n given by (18.4-1) can be written as
n = (1723)(4)(56). (18.4-2)
The length of a cycle is the number of symbols in it. Cycles of length 1 [e.g.,
(4)J are usually omitted, since they represent the identity mapping, in which
nothing is permuted. A cycle of length 2 is a transposition; it simply inter-
changes two of the symbols and leaves the rest unaJtered.
Any permutation can be expressed as the resultant of successive trans-
positions. For example, if
n l = (17), n2 = (72), n3 = (23), n 4 = (56),
then the permutation (18.4-2) can be written as

n = nl 0 n2 0 n3 0 n4 = (17)(72)(23)(56), (18.4-3)
where it is understood that the transpositions are to be performed in the
order reading from right to left. The decomposition of a given permutation
into transpositions is not unique, but it will now be proved that for a given
permutation the number of transpositions is either always even or always odd.
Namely, let f(- ..) be the function of n real or complex variables defined by
f(Xl, ... , xn) = Il (Xk - x); (18.4-4)
l~j<k~n
let n: j --+ n(j) be a permutation of the integers 1, 2, ... , n, and call

frlxb ... , xn) = Il (X,,(k) - X"w)' (18.4-5)
1 ~j<k~n
If Xk - Xj is anyone of the factors in (18.4-4), then either Xk - Xj or Xj - Xk

appears precisely once as one of the factors in (18.4-5), so that either f" == f
or f" == -f; in these cases n is called an even or an odd permutation, respec-
tively. This property is called the parity of the purmutation n. If two permuta-
tions have the same parity (i.e., if they are either both even or both odd), then
their product or resultant is even; if they have opposite parities, their product
is odd. Clearly, the transposition (12) is odd, for then all the factors of
(18.4-5) have the same sign as in (18.4-4), except (X2 - Xl)' Next, it is seen
that the transposition (11) = (21)(12)(2l) is necessarily odd, and finally

that the general transposition Uk) = (lj)(1k)(lj) is always odd. Hence, in
any decomposition of an even (or odd) permutation into transpositions, the
number of factors is always even (or always odd).
There are just as many odd as even permutations of a given set C, because
if 1[1 is any fixed odd permutation, then the mapping 1[ 11[ -> 1[ of even permu-
tations onto odd ones is a one-to-one mapping.
It is clear that the set of all permutations of n symbols (including, of
course, the identity permutation in which each symbol is mapped onto
itself) is a group of order n! under the usual law of composition of mappings
(from right to left); this group is called the symmetric group on n symbols,
and is denoted by Y n' The subgroup of all n!j2 even permutations is called
the alternating group on n symbols, and is denoted by d n
18.5 Homomorphisms; Normal Subgroups

It is recalled that a homomorphism is a mapping cp: G -> G' of a group G
into a group G' such that cp(ab) = cp(a)cp(b) for all a and b in G. (If cp is also
one-to-one and onto, then it is an isomorphism.) Certain subsets (it will be
seen that they are subgroups) of G and G' are defined as follows:
Go = {x: cp(x) = e' (identity of G')}
= kernel of the mapping = ker( cp)
G'1 = {cp(x): xEG}
= image of Gunder cp = cp( G).
G~ = q;(G)
e'
Figure 18.2 A group homomorphism.
The main interest is in the case in which the image G'1 = cp(G) is a simpler
group than G (in this case cp cannot be one-to-one) but yet cp(G) is not merely
the trivial group {e'}. One may then regard the image as having the main
features of G but without some of the fine detail. Insofar as cp( G) approximates
G, it does so accurately, in that the image of a product of two elements in G is
always the product of their images.
Theorem. Go and G'1 are subgroups of G and G' respectively; furthermore,

yxy-1 E Go for any x in Go and any y in G.
Homomorphisms; Normal Subgroups 9
The proof of this theorem is given in some detail, to serve as a sort of

model. Proofs of the other theorems in this Chapter are mostly left to the
reader.
(1) PROOF (that Go < G). It must be proved that the group axioms are satisfied
in Go. First, suppose that x and yare in Go, i.e., that ep(x) = e' and ep(y) = e'. It must
be shown that xy is in Go, but ep(xy) = ep(x)ep(y) (by the homomorphism property of
the mapping ep) = e' e' = e' (by one of the properties of the identity in any group).
Therefore, ep(xy) = e', but Go was defined as the set of all elements in G that are
mapped onto e'; hence xy is in Go. Second, the associative axiom is automatically
satisfied in Go because it is satisfied in all of G [if x, y, and z are in Go, then (xy)z =
x(yz), because x, y, and z are in G, which is a group]. The third axiom will be verified
in the alternative form given in Section 18.1: Identity: Go contains the identity e,
for if a is any element of G, then ep(a)ep(e) = ep(ae) = ep(a); therefore epee) = e'.
Inverse: Suppose that a is in Go (hence, ep(a) = e'). Then, e' = epee) = ep(aa- l ) =
ep(a)ep(a- l ) = e'ep(a- l ) = ep(a- l ); therefore a-I is in Go.
(2) PROOF (that G~ < G'). It will be similarly shown that the group axioms
are satisfied in G'l' First, if x' and y' are in G'l, then x' = ep(x) and y' = ep(y), for some
x and yin G, but x'y' = ep(x)ep(y) = ep(xy), and G'l is defined as the set of all elements
of G' which are = ep(z) for some z in G; therefore x' y' is in G'l' Second, multiplication
is associative in G'l, because it is so in all G'. Third, it has been shown above that
epee) = e'; hence e' is in G~. Lastly, if x' is in G'l, then x' = ep(x) for some x in G. Now,
X'ep(X-l) = ep(x)ep(x- 1) = ep(xx- l ) = epee) = e', which shows that ep(x- l ) is X'-l;
hence x' - I is in G'l'
(3) PROOF (that yxy-l is in Go if x is in Go, while y is arbitrary in G).
ep(yxy-l) = ep(y)ep(X)ep(y-l) = ep(y)e'ep(y-l) = ep(y)ep(y-l) = ep(y)(ep(yWI = e';
therefore, yxy-l is in Go.
Note. All the conclusions hold, even if G' is not a group, but merely a set
in which a product x'y' is defined. In particular, it then follows that the subset
G'l is necessarily a group, even though G' may not be. All the arguments are
unchanged except the second step in part (2). To prove associativity in G'l'
let x', y', and z' be elements in G'l' hence elements of the form <p(x), <p(y), and
<p(z). Then (x'y')z' = (<p(x)<p(ycp(z) = cp(xy)cp(z) = cp'xy)z) = cp(x(yz be-
cause multiplication is associative in G) = <p(x)<p(yz) = <p(x)(<p(y)<p(z =
x'(y' z').
Any subgroup Go of a group G such that yxy-l E Go for any x EGo,

any y E G, is called a normal subgroup (or normal divisor) or invariant or
self-conjugate subgroup of G; in symbols, Go <l G. Clearly, a homomorphism
of the kind we are interested in can exist only if G contains a nontrivial
Cot {e}) proper Ct G itself) normal subgroup Go. If G contains no such, it is
called simple, because it cannot be mapped homomorphic ally onto any
simpler nontrivial group. It will be shown in Section 18.8 that, conversely,
if G contains a normal subgroup, one can construct a homomorphism of
which Go is the kernel, and that all (in the sense of isomorphism) homomor-
phic images of G are obtained by this construction.
For any element x, the elements of the form yxy-1 (y E G) are the conju-
gates of x. A subgroup is normal if and only if it contains all the conjugates
of all its members. In an Abelian group, yxy-1 is always equal to x, hence
every subgroup is normal.
18.6 Cosets
Let Go be a subgroup of G, and y any fixed element of G. The subsets
are called left and right cosets, respectively, of Go in G. The element y is

called a representative of S~) (or of S~). Any member of a coset may be
taken as its representative. Go itself is both a left and a right coset; its
representative may be taken as the identity element e of the group. The
association x -+ y x, for fixed y, all x in Go, is a one-to-one mapping of
0
Go onto S~) (and similarly for right cosets); hence, each coset contains the
same number (finite or infinite) of elements as Go. Furthermore, it is easily
proved that any two left co sets (or any two right cosets) are either identical
or disjoint, so that the number of elements in G (if finite) is equal to the
number in Go times the number of (say left) co sets (including Go itself).
Lagrange's theorem for finite groups follows: the order of a subgroup divides
the order of the group. It follows, for example, that if the number of elements
in G is prime, then G has no subgroups other than {e} and G itself. The co sets
S~) and S~) are also denoted by yG o and GoY. Note that the co sets, except Go
itself, are not subgroups.
18.7 Factor Groups
Theorem 1. A subgroup Go in G is normal if and only if every left coset

(of Go in G) is a right coset, and conversely; the coset containing y is then
denoted by Sy. (The proof is elementary and is left as an exercise.)
Definition. If S1 and S2 are any two subsets (not necessarily subgroups or

cosets) of a group, their product is defined as the subset
(18.7-1)
note that generally S1S2 =1= S2Sl'

If S1 and S2 are co sets (left or right) their product is not generally a coset
(it is usually larger than any coset), unless Go <l G.
Theorem 2. If Go is a normal subgroup ofG (Go <l G), then Sy,SY2 = SylY2
for all Y1 and Yz in G. Conversely, if Go is a subgroup such that the product
of any two (say left) cosets is always a (left) coset, then Go <l G (in which
The Structure of Cyclic Groups 11
case the distinction between right and left cosets disappears). Under these
circumstances, the collection of cosets
{Sy: y E G}, o = multiplication defined by (18.7-1),
is a group, called the factor (or quotient) group of G with respect to Go
and is denoted by GIGo. Furthermore, the mapping ({In: G -+ GIGo defined
by ({In(Y) = Sy (each element ofG is mapped onto the coset in which it resides)
is a homomorphism called the natural homomorphism of G onto GIGo.
According to this theorem, whose proof is also left to the reader, one
can always construct a homomorphic image and a homomorphism corre-
sponding to any normal subgroup Go. The theorem ofthe next section shows
that these are essentially the only homomorphisms of the given group G.
18.8 The Law of Homomorphism
Theorem. Let tf;: G -+ G' be any homomorphism. Let Go be its kernel,

and let tf;(G) be the image ofG in G' under the mapping tf;. Let Sy denote the
coset of Go that contains y. Then tf;(G) and the factor group GIGo are iso-
morphic, and, specifically, the mapping 'II: GIGo -+ tf;(G) given by
(18.8-1)
is (1) well-defined (i.e., independent of the particular choice of the repre-
sentative y of Sy) and is (2) an isomorphism. Schematically,
tf;(G) +-- G -+ GIGo
'" 'P"
~ '(i50)
The reader is urged to complete the proof in detail, but let us point out
what has to be proved. To prove that 'II is well defined, one must show that
if SYI = SY2' then tf;(Yl) = tf;(Y2). To show that 'II is one-to-one, we must show
that if tf;(Yl) = tf;(Yl), then SYI = SY2. To show that 'II has the homomorphism
property, one must show that
'P(Sy,Sy,) = 'P(Sy,)'P(Sy,).
Lastly, it is obvious that 'II is onto, for any element of tf;(G) is tf;(y) for some
yin G.
18.9 The Structure of Cyclic Groups
As noted in Section 18.6, a group of prime order p has no nontrivial proper

subgroups, hence is necessarily cyclic (hence Abelian), and can be generated
by any of its elements (except the identity) because the order of any element
(that is, the order of the subgroup generated by that element) must be either
1 or p, by Lagrange's theorem. (The only element of order 1 in a group is the
identity e, for to say that a1 = e, when I = 1, is equivalent to saying that
a = e). If G = {e, a, a 2 , , an - 1} is a cyclic group of order n, and m is a
divisor of n, then the elements e, am, a 2m , ... constitute a cyclic subgroup of
order n/m, and these are the only subgroups of G. In an infinite cyclic group,
any element that is =F e generates an infinite cyclic subgroup; these are the
only nontrivial subgroups, and they are all distinct, but isomorphic.
18.10 Translations, Inner Automorphisms
If a is an element of a group G, we denote by T. the mapping of G onto itself

given by x ~ ax (x E G); this is a left translation in G; a mapping of the form
x ~ xa, for fixed a, is a right translation.
EXERCISE
1. Show that the set .'T of all left translations in G is itself a group under the usual
law for the composition of mappings, and that this group .'T is isomorphic to G.
N.B.I. Any homomorphism of a group G (abstract or otherwise) onto a
group of mappings is called a representation of G. The isomorphism G ~ ff
is called the regular representation of G. If a representation is an isomorphism
(not merely a homomorphism), it is called faithful. The regular representa-
tion is faithful.
N.B.2. If G is a finite group, the mapping T. is a permutation of the elements
of G; therefore, ff is some subgroup of Y n' where n is the order of G, i.e.,
any finite group is isomorphic to some group of permutations. (This is Cayley's
theorem.)
If a is an element of a group G, we denote by Aa the mapping of G onto

itself given by x ~ axa - 1.
EXERCISE
2(a). Show that the set ~ of all the mappings Aa in G is itself a group, under the
usual law of composition of mappings.
(b). IfG = d 3 , find~.
(c). If G = d 4 , find f
N.B.3. f is in any case some subgroup of Y 6/2 = Y 3 for exercise 2(b)
and some subgroup of Y 12 for exercise 2(c).
N.B.4. "Find f" means to idyntify f as being isomorphic to some known
group.
N.B.5. Each of the mappings Aa is an automorphism called an inner auto-
morphism of G: (1) It is one-to-one and onto, because the equation axa - 1 = Y
can always be solved for a unique x(x = a- 1 ya); (2) it has the homomorphism
The Subgroups of g' 4 13
property that products are mapped into products, because axwa- I =

~ is called the inner automorphism group of G.
axa - 1 awa - 1.
Inner automorphisms of Y'n have a special property. Let n E Y'n, and let
n be written as a product of independent cycles, as in (18.4-2) [two cycles are
independent if they contain no common symbol; (173) and (24) are indepen-
dent, but (173) and (34) are not], the longer cycles being written first and the
cycles of length 1 being included. Then the lengths of the cycles constitute a
partition of n, that is, a set of positive integers whose sum is n. The special
property referred to is that the image of n under an inner automorphism of
Y'n (n -+ (In(J-1, where (J is a given element of Y'n) always corresponds to the
same partition of n as n itself, because if (a, b, ... , f) is any cycle, then
(J(a, b, ... , f)(J-1 is the cycle (J(a), (J(b), ... , (JU)).
If a and x are in a group G, then the element axa- I is called conjugate
to x. The inner automorphism Aa: x -+ axa- 1 (a fixed) maps each group
element x onto one of its conjugates. If Go is a subgroup of G, then the set
G~ = {axa- 1 : xEG o}
is also a subgroup, and is often denoted by aG o a- 1 ; it is said to be conjugate to
Go. If aGo a-I = Go,for all a in G, then Go<l G. Hence, normal subgroups are
sometimes called "self-conjugate" subgroups. The conjugate elements of
Y'n are those that have the same structure when written as products of
independent cycles, e.g., (1732)(56)(4) and (4531)(76)(2).
18.11 The Subgroups of !/'4
The symmetric group 51'4 has 22 subgroups in addition to the trivial sub-
group {e} and G itself. Arranged in classes of conjugate subgroups, they are:
1. {e, (12)}, etc-six subgroups of this type.

11. fe, (123), (132)}, etc.-four subgroups.
111. fe, (1234), (13)(24), (1432)}, etc.-three subgroups.
IV. fe, (12)(34)}, etc.-three subgroups.
v. {e, (12)(34), (13)(24), (l4)(23)} = V4 -one subgroup.
VI. {e, (123), (124), (134), (234), (321), (421), (431), (432), (12)(34), (13)(24),
(l4)(23))} = d 4 -one subgroup.
Vl1. {e, (12), (13), (23), (123), (321)}-four subgroups.
Any subgroup in a given line of the table can be obtained from any other
subgroup in the same line by an inner automorphism of the whole group 51'4.
For instance, under the mapping n -+ (l2)n(12), the elements of the group
fe, (134), (431)} go into the elements ofthe group fe, (234), (432)}. Therefore,
the only normal (i.e., invariant, i.e., self-conjugate) subgroups are V4 and d 4,
each of which occupies a line of the table by itself. However, each of the
subgroups in line iv is a normal subgroup of V4 (which shows, incidentally,
that G 1 <l G z <l G 3 does not imply G 1 <l G3 ). A complete so-called com-
position series (see below) of!/'4 is the series
(18.11-1)
For n 2 5, d n is a simple group (it has no nontrivial proper normal sub-
groups), so the composition series is merely
{e}<l dn<l!/'n-
EXERCISE
1. Show that d s is a simple group. Outline of the proof: Assume that Go <:J d s,
but Go #- {e}; then, it must be proved that Go = d s . Show that Go must contain an
element TC of one of the types
(a) (a be),
(b) (a b)(c d),
(c) (a bed e).
Then it contains all elements aTCa-1, where a E d s; show that it contains all elements of
the type TC. Show that if Go contains all elements of one of the above types, then it contains
elements (hence all elements) of both the other types, e.g., if Go contains type (a), then it
contains (1 2 3)(2 3 4) = (21)(34). Why does this proof fail for d 4? The simplicity of
d s is a key step in Galois proof that the quintic equation cannot be solved by radicals.
A composition series of a group G is a sequence of subgroups {G;} such

that
(18.11-2)
where G i =f. G i + 1, and where no refinement is possible, i.e., if G i <l H <l G i + 10
then either H = Gi or H = Gi + 1 A finite group always has a composition
series. If G is an infinite group, it may happen (see Exercise 3, below) that
every series of the above type has a refinement. The famous Jordan-Holder
theorem says that if G has a composition series (18.11-2), and if
{e} = Ho<l Hi <l ... <l HI = G
is any other composition series for G, then (1) I = k, and (2) although the
subgroups HI, ... , H k - 1 may be different from G1 , . . . , Gk - 1 , at least the
factor groups
HdHo, H z/H Io , H k /H k - 1 (18.11-3)
are the same (except for order) as
GdG o , GZ/G 1, , Gk /G k - 1; (18.11-4)
that is, the list (18.11-3) can be arranged in such an order that each group in
it is isomorphic to the corresponding group in the list (18.11-4).
EXERCISES
2. How does the Jordan-Holder theorem apply to the composition series (18.11-1)
for [//4?
Generators and Relations; Free Groups 15
3. Show that the infinite cyclic group has no composition series.

4. Let Y 00 denote the set of all finite permutations of the positive integers {I, 2, ... }.
(Each permutation in 9" 00 maps all but a finite number of these integers onto themselves.)
Show that Y 00 has the composition series {e} <J d 00 <J Y 00'
18.12 Generators and Relations; Free Groups
Let G be a group; a subset S = {a, b, ... } of the elements of G is said to

generate G if every element of G can be written as a (finite) product of elements,
each of which is either an element in S or the inverse of an element in S.
(Equivalently, any g in G is a product of powers of elements in S-the
powers of an element were defined in Section 18.2.) To describe a group
completely, one must not only say what its generators are, but also give
various relations among them. For example, a cyclic group of order n can
be described by one generator a and one relation an = e.
A group without any relations is called a free group and is constructed
as follows: Let S be a (finite or infinite) set of letters, S = {a, b, ... }, called
generators. A word is a finite string of symbols X 1 X2 .. X k , where each Xi
is either one of the letters in S or a symbol of the form a-l, where a is in S.
Two words are equal if one of them can be obtained from the other by insert-
ing and deleting symbol pairs of the form aa- 1 and a- 1 a (a in S). It is
obviously desirable to delete as many such pairs as possible; this may result
in the empty word - the word with no letters in it; that word is denoted bye,
and it is understood that "e" is not one of the letters in S. (For example,
abc- 1cb- 1a- 1 = e.) The product of two words is obtained by simply
stringing them together: Ifw = X1X2 .. Xk and u = YlY2 ... Yz, then
It is an elementary exercise to verify that the set of all words, using a given
set S of generators, is a group G under this definition of product. It is called
the free group generated by S. The identity element is the empty word e,
and the inverse of X1X2 .. Xk is the word YkYk-l .. Yl, where Yi = a-lor ex
according as Xi is ex or a- 1
Relations can be established among the elements of the group (it is then no
longer free) by means of equations Wl = e, W2 = e, etc., where Wi> W2, etc.
are certain words. The relations establish a structure in the group.
If among the relations we have aba- 1 b- 1 = e (which is equivalent to
ab = ba) for every pair a, b in S, then the group is Abelian. If these are the
only relations, then G is called a free Abelian group. Free groups and free
Abelian groups appear in Section 23.7, in the study of the kinds of multiple
connectedness that a manifold can possess. The structure of a free group or
of a free Abelian group is determined solely by the number of generators.
Any finite group G is equivalent (i.e., isomorphic) to a group defined by
generators and relations: S can be taken as the set of all group elements in G,
and the relations taken so as to give all the information provided by the
multiplication table of the group; whenever the table says that ab = c,

then one of the relations is abc - I = e.
If a group is defined by finitely many generators and finitely many relations,
it is called finitely presented; it may still be a group of infinite order, and
such groups have provided many intricate and difficult problems for current
research. The so-called word problem, though by now not exactly new,
illustrates the flavor of the subject. In addition to the defining relations
WI = e, ... , Wn = e, there are always other equations of the form W = e;
for instance, W can be WI W2, or it can be Wo WI Wo I, where Wo is an arbitrary
word. The word problem, which was first posed in 1912, is this: Given a
finitely presented group, find a procedure (i.e., an algorithm) such that if
an arbitrary word W in the generators of the group is given, the procedure will
allow one to decide, in a finite number of steps, whether the equation W = e
is true or false. W. Magnus showed in 1932 how to produce such an algorithm
for any group with a single defining relation. In 1955, Novikov gave a (very
long) proof that the word problem is in general unsolvable, and by now
groups are known, defined by a quite small number of generators and rela-
tions, for which it can be proved that no algorithm of the desired kind exists.
18.13 Multiply Periodic Functions and Crystals
A crystal, ideally, is composed of an enormous number of identical unit

structures, called cells, arranged in a triply periodic array or lattice in
space. If f(x) is the density of mass or charge or a similar quantity within the
structure, then f(x) is a triply periodic function. Generally, a function
f(x!> ... , xn) = f(x) of n real variables is called n-tuply periodic if there are n
linearly independent vectors v(l), ... , yen), i.e., vectors for which
VI (1) ... VI (n)

det ( : : =f. 0, (18.13-1)
vn(l) vn(n)
such that f(x) satisfies the equations
f(x + v(1 = f(x),
f(x + v(2 = f(x), (18.13-2)
f(x + yen~ = f(x)

for all x. The vectors v(i) are called periods of f(x); if w is any vector of the
form
(18.13-3)
where the mi are integers (positive, negative, or zero), then f(x) = f(x + w)
for all x; hence w is also a period of f(x). It is furthermore assumed that every
The Space and Point Groups 17
period of f(x) is of the form (18.13-3) with integer coefficients; then, the
vectors v(l), ... , v(n) are said to constitute a fundamental set of periods.
Certain functions may not have a fundamental set, even though they are
periodic in the strict sense, for example, constant functions and functions that
are periodic in some coordinates and independent of the others. Such func-
tions will be called degenerate, and are excluded on the physical ground that
each atom occupies a certain volume, and that functions like the potential
and the charge density vary from the center of an atom to the outside, so that
some variation is unavoidable in any given direction in space. A multiply
periodic function is called nondegenerate if it has a fundamental set of periods.
The set of points x in IRn determined by
x = mlv(l) + ... + mnv(n), (18.13-4)
where the mi are integers, is called the lattice of f(x).
If new vectors v'(l), ... , v'(n) are given by
v'(j) = mjl v(l) + ... + mjn v(n), (18.13-5)
where the mjk are integers such that
mIn)
:. = -+1 , (18.13-6)
... mnn
then v'(1), ... , v'(n) also constitute a fundamental set of periods, because
when (18.13-5) is solved by the use of determinants, the v(j) are seen to be
linear combinations of the v'(j) with integer coefficients. Both fundamental
sets generate the same lattice.
18.14 The Space and Point Groups
If, for any fixed w, Tw denotes the translation

(for all x) (18.14-1)
of [Rn onto itself, then a multiply periodic function f(x) is invariant under all
transformations of the Abelian group
g- = {Tw: w is a period off(x)}, (18.14-2)
which is called the translation group of f(x). The function lex) may, of
course, also be invariant under other transformations, such as certain
rotations about certain axes, reflections in certain planes, and the like. The
group Gs of all (homogeneous and inhomogeneous) linear transformations
of [Rn under which f(x) is invariant is called the space group of f(x) or, if
f(x) represents a crystal, the space group of the crystal.
An element of Gs is a transformation of the form
x ...... x' = Mx + 1;, (18.14-3)
where M is a nonsingular matrix and ~ is a vector; the transformation is

denoted by (~, M). The law of composition in the group Gs is found by
applying two such transformations in succession:
xl! = MIx' + ~l = M 1 (M 2 x + ~2) + ~l.
It is seen that
(18.14-4)
A transformation (18.14-3) under whichf(x) is invariant is called a symmetry
operation of f(x).
The transformation (18.14-3) generally combines translation, rotation,
dilation, and shear. However, it is easy to see that if f(x) is continuous and
non degenerate (hence, in particular if it represents a real crystal), dilation
and shear are ruled out. To show the general nature of the argument, we
discuss shear in two dimensions: The square lattice of points with integer
coordinates in the x, y plane is invariant under the translation group ff
consisting of translations x ~ x + k, y ~ y + I (k, I integers) and also under
various transformations involving shear, such as the transformation
(n integer). (18.14-5)
If f(x, y) is the function that is equal to 1 at the lattice points (x, y integers)
and equal to 0 elsewhere, thenfis also invariant under (18.14-5). However, a
continuous nondegenerate doubly periodic function f(x, y) cannot be
invariant under (18.14-5). If it were, then the identity
f(x + ny - k, y) == f(x, y)
would hold, for all nand k; for y irrational, the numbers ny - k are dense
on IR, so by continuity f(x, y) would have to be independent of x for all
irrational y, hence (by continuity again) for all y; hence f(x, y) would be
degenerate. Therefore shear must be excluded. By arguments of this sort, it is
concluded that the matrix M in (18.14-3) must be an orthogonal matrix.
The set of all orthogonal matrices M such that (~, M) is in Gs for some ~
is also a group; it is called the point group of f(x) and is denoted by G p
Clearly the mapping
(18.4-6)
is a homomorphism, whose kernel is ff; hence, by the homomorphism law
for groups, ff is a normal subgroup of G" and Gp is isomorphic to the factor
group Gs/ff.
EXERCISE
1. Using (18.14-4), find the formula for (/;, M)-l. Then verify directly that !Y
is a normal subgroup of Gs by showing that, if (/;, I) is any pure translation (J being the
unit matrix), then any group element of the form (1), M)(/;, 1)(1), M)-l is also a pure
translation.
The Space and Point Groups 19
In crystallography, much information about a crystal structure can be

given by specifying Gp and:Y (the latter by describing the lattice it generates).
However, that does not generally give as much information as is given by
the space group Gs In particular, Gs mayor may not contain Gp as a sub-
group, because Gs may contain (~, M) for some ~ #- 0 but not (0, M).
Note. :Y, as an abstract group, is isomorphic to the free Abelian group on n

generators, hence gives no information. However, the lattice generated by :Y,
when a fundamental set of periods is given, does give information about
f(x). The space group under which the lattice is transformed into itself
contains the space group of f(x) as a subgroup.
For n = 3, a detailed description of the possible symmetry operations and
of the space and point groups, is contained in the International Tables for
X-ray Crystallography, Henry and Lonsdale (1965). The symmetry opera-
tions are: pure translation, pure rotation, reflection in a plane, reflection
together with a rotation about an axis perpendicular to the plane of reflection,
reflection together with a translation parallel to the plane, and a rotation
with a translation parallel to the axis of rotation. Whenever a rotation occurs,
the possible angles of rotation are 2n/n, where n = 1, 2, 3, 4, or 6, as
shown for the case of pure rotations by Exercises 2 and 3 below. There are
32 point groups, 14 types of lattice, and 230 space groups.
EXERCISES
2. (The purpose of this exercise is to show that the only possible pure rotational
symmetries of a 2-dimensional crystal are ones with an n-fold axis, where n = 1, 2, 3, 4,
or 6.) Consider a non degenerate doubly periodic functionJ(x, y), and write it asJ(z),
a (nonanalytic) real function of the complex variable z = x + iy. Let IX and fJ be a
fundamental pair of period; then Re(lX/fJ) =1= 0, and J(z + nIX + mfJ) == J(z), when nand
m are integers. By suitable scaling and suitable orientation of the x and y axes, take
fJ = 1 for simplicity. Assume thatJ(z) is also invariant under a rotation z -> ei8 z. From
the equations
J(z) = J(ze i8 ),
J(z + 1) = J(ze iB + ei8 ),
J(z + IX) = J(ze i8 + lXe iB ),
conclude that ei8 and lXe i8 are also periods of J(z). Show from that that IX satisfies the
equation
ra 2 + (5 - p)1X - q = 0,
_ra 2 + (5 - p)1X - q = 0,
where p, q, r, and 5 are integers such that

p5 - rq = 1,
and hence that ei8 is given by
i8 I jr-=4
e =--'-2--' 1= p + s.
For (J real, I must be in [ - 2,2]. Conclude that the possible values of (J are 0, 11/3,
11/2, 211/3, 11.
3. Extend the conclusion of Exercise 2 to the 3-dimensional case, as follows:
Assume that the function J(x, y, z) = J(x) is triply periodic and that {D, v, w} is a
fundamental set of periods. Suppose, furthermore, thatJ(x) is invariant under a rotation
through an angle (J about some axis in space. By choosing the origin to lie on the axis,
the rotation can be written as x --> Rx, where R is a 3 x 3 orthogonal matrix of deter-
minant = 1. Show that the vectors
D' = RD - D,
v' = Rv - v,
w' = Rw - w,
are all periods ofJ(x), are all perpendicular to the axis of rotation, and are not collinear.
It follows thatJ(x) is doubly periodic in any plane perpendicular to the axis of rotation,
hence Exercise 2 applies.
In some books, the restrictions to 2-, 3-, 4-, and 6-fold rotation axes is
derived from a somewhat mysterious "principle of rational indices," which
is said to be of empirical origin. We have seen, however, that the restriction
follows directly from the existence of a triply periodic structure; hence, the
"principle ofrational indices" is unnecessary.
18.15 Direct and Semidirect Products of Groups;

Symmorphic Space Groups
If Go is a normal subgroup of G, one cannot in general think of G as a product
of Go and GIGo; in fact, G does not in general even contain a subgroup
isomorphic to GIGo. Below, two exceptions to that general rule are discussed.
In the simplest case, Hand K are subgroups of G such that every g in G
can be written uniquely as hk, with h in Hand k in K, and such that every
h in H commutes with every k in K. Then, G is said to be the direct product
of Hand K. In symbols, G = H x K or K x H. The identity e is the only
element in H n K (for if any other a were in H n K, it would have two
representations in the form hk, namely ae and ea). Furthermore, Hand K are
normal subgroups of G, for if a = hkl is any element of G and k2 is any
element of K, then ak 2a- 1 = hklk2kllh-l but this is = klk2k 1 because t.
h commutes with klok2' and k1 1 ; hence, ak 2a- 1 is in K, and K<J G;
similarly, H <J G. Furthermore, the factor group GIH is isomorphic to K,
and GIK is isomorphic to H, because any element of GIH is a coset aH =
{ah: hE H} and can be written uniquely as a coset kH (where a = kh,
k E K, and hE H); the mapping k -+ kH is clearly an isomorphism of K onto
GIH, because klHk2H = klk2H.
Another point of view is to assume that H 0 and Ko are arbitrary given
groups and to construct a group G, called their direct product, whose elements
are pairs (h, k), where hE Ho and k E K o, with the law of composition
(hI' k l ) 0 (h2' k 2) ~ (hIh2' kIk2)' (18.15-1)
Direct and Semidirect Products of Groups; Symmorphic Space Groups 21
The identity element of G is the pair (e, e'), where e and e' are the identity
elements of H 0 and of K o , respectively; furthermore, (h, k) - 1 = (h - 1, k - 1).
Let Hand K be the subsets of G defined as
H = {(h, e'): hE Ho} and K = {(e, k): k E Ko}.
It is easy to verify that G, H, and K are groups, that Hand K are normal
subgroups of G, that H ~ Ho and K ~ K o, and lastly that G = H x K.
EXERCISE
1. Verify the statements made in the last sentence.
Because of the isomorphisms H ~ H and K ~ K o , G is also called the

direct product of Ho and Ko; in fact, it is usual to identify Ho and Ko with
the subgroups Hand K, and to drop the subscript 0 altogether.
EXERCISE
2. Assume that the identity e is the only element common to two subgroups H
and K of a group G. Show that h in H commutes with every k in K if and only if H <l G
and K <l G.
In the next simplest case, the semi direct product, it is still assumed that
every 9 in G can be uniquely expressed as hk, with hE Hand k E K, but it is
assumed only that H <J G, while K is not necessarily normal. Then G is a
so-called semidirect product of Hand K. Any coset of H in G (i.e., any element
of the factor group G/H) has a unique representation kH = Hk, with k in
K; furthermore, Hk1Hk2 = Hk1k2' and hence the factor group G/H is iso-
morphic with K. If 91 and 92 in G are expressed uniquely as 91 = h1k1 and
92 = h2k2' then the unique expression of 9192 = 93 is h3k3' where
h3 = h1k1h2kI1, k3 = k 1k 2
(Note that k1h2 kI1 is in H, because H is normal, but is not necessarily = h2
unless K is also normal.)
The group G of rigid motions in the plane (or in n-space) provides an
example of a semidirect product. A rigid motion is a transformation
X -+ x' = Mx + ~, (18.15-2)
where M is a 2 x 2 (or n x n) real orthogonal matrix with determinant = 1,
and ~ is an arbitrary vector. G is generated by the group f7 of translations
x-+x'=x+~
and the group f31l of rotations about the origin

x -+ x' = Mx.
(The law of composition in f7 is vector addition ~1 + ~2' and that in f7t is
matrix multiplication M 1 M 2. Here, f7 stands for the group of all translations,
not merely those of a lattice.) The combined transformation (18.15-2) is

denoted by (~, M), as in the preceding section, where it was shown that the
law of composition in G is given by
(18.15-3)
This would make G into the direct product !Y x fll if the right member of
the above equation were replaced by (~1 + ~2' M 1M 2)' The significance of
the term M 1 ~ 2 for the two groups is as follows: First, for fixed M, the mapping
~ --+ M~, for all ~
of!Y onto itself is an automorphism of !Y, because it is one-to-one and onto

(M has an inverse), and it maps ~ + T) onto M~ + MT); this automorphism
is called reM). Second, as M varies in all of fll, these automorphisms form a
group .91, a subgroup of the group of all automorphisms of !Y. Third, the
mapping
M --+ r (M), for all M in fll,
is a homomorphism of fll onto .91, because if the automorphism r(M):
~ --+ ~' = M~ is first performed on an element ~ of!Y, followed by r(N): ~' --+
~" = N~', the result is ~ --+ NM~; that is, r(N)r(M) = r(NM).
Definition. Let Hand K be any two groups (the law of composition will
be written multiplicatively in both), and let there be given a homomorphism
k --+ r(k) of K into the automorphism group of H [for fixed k, r(k) maps h
onto r(k)h, for all h in H]. Then, the set of all pairs (h, k), with h in Hand k
in K, and with the law of composition of such pairs given by
(18.15-4)
is a group, called the semidirect product of Hand K (or of H by K) and is
denoted by
The main properties of the semidirect product are summarized in the

following exercises. The reader is urged to do these exercises, because they
reveal that the definition given above, which may have seemed rather
arbitrary, contains just the right ingredients to give the semi-direct product
its various desirable properties.
EXERCISES
3. Show that the identity elements of G = H x, K is (e, e'), where e and e' are the
identity elements of Hand K.
4. Show that the element (r(k-1)h- 1, k- 1) is the inverse (i.e., both right and left
inverse) of (h, k).
5. Show that the associative law holds in G. Warning: r(k)[h 1h2J is not the same as
[r(k)hlJh2' because r(k) is a mapping, not a group element. Exercises 3, 4, and 5 show
that G is a group, as claimed in the definition of semidirect product.
Direct and Semidirect Products of Groups; Symmorphic Space Groups 23
6. Now identify Hand K with the subgroups {(h, e'): all h in H} and {(e, k): all
kin K}, respectively, and show that H is a normal subgroup of G.
7. Construct the factor group G/H, and show that it is isomorphic to K.
8. Conversely, suppose that a group G contains subgroups Hand K, of which H
is normal, such that H n K = {e}, and such that the factor group G/H is isomorphic to
K; show that G is the semidirect product H x, K, where, for any k in K, T(k) is the map-
ping h -> khk- 1 , for all h in H.
9. Show that the semidirect product is equal to the direct product H x K if and
only if T(k) == J [that is, the homomorphism k -> T(k) of K into the automorphism group
of H maps all of K onto the identity element (identity mapping of H onto itself)].
10. Show that K, as well as H, is a normal subgroup of G if and only if T(k) == J,
i.e., if and only if the product is direct.
Exercise 8 shows that the automorphisms -r(k) that appear in the semi-
direct product of two given groups Hand K become inner automorphisms
of the group H x rK that is being defined. This fact is somewhat obscured,
in the case of the rigid motion group G, by the use of the additive notation
for the translation group !!T. If, instead, the translation x --+ x' = x + ~ is
denoted by T; and the multiplicative notation is used, so that {~, M} is
simply the combined operation ~M, then TM~ = M~M-l, so that the
automorphism -reM): ~ --+ M~ in!!T takes the form
-reM): ~ --+ MT~M-\
hence is an inner automorphism in G.
According to the preceding section, the translation group !!T of a crystal
structure, given by (18.14-2), is a normal subgroup of the space group G.,
and the point group Gp is isomorphic to the factor group Gs/!!T [it is recalled
that the point group is the group of all rotations and reflections x --+ Mx
such that (~, M) is in G., for some~]. Gs mayor may not contain a subgroup,
say G~, isomorphic to Gp ; if it does, then!!T and G~ can have only the element
e in common, because all the other elements of !!T are of infinite order (if
T E !!T, then T m =1= I for all m =1= 0), while all elements of Gp are of finite
order. Therefore, according to Exercises 5 and 6, above, Gs contains such a
subgroup if and only if it is a semidirect product!!T x r Gp' In this case, the
space group is called symmorphic by the crystallographers.
When an x-ray crystallographer starts to analyze a set of x-ray reflection
data to determine a crystal structure, he often knows the point group in
advance, from measurement of the angles between crystal faces and cleavage
planes, and from other macroscopic properties of the crystals. However,
he cannot assume that the space group contains a copy of the point group,
i.e., that the space group is symmorphic.
A simple 2-dimensional example of a nonsymmorphic space group is
that of the function f(x, y), which is equal to 1 in the shaded triangles in
Figure 18.3, and 0 otherwise. In complex notation, the space group is
generated by the translations z --+ z + a, Z --+ Z + i/3 and the so-called
glide-reflection z --+ Z + !a; hence the point groups contains the reflection
z --+ z, while the space group does not.
Figure 18.3
The semidirect product H x t K is an instance of a so-called extension

of the group H by the group K. For a fuller discussion of group extensions,
see Chapter 12 of Kurosh, Theory of Groups (1956) or 50 of Redei, Algebra
(1959) (in German).
CHAPTER 19
Continuous Groups
General linear, special linear, orthogonal, and unitary groups; rotation and
Lorentz groups; Euler's theorem, the four components of the full Lorentz
group, the Thomas presession; group manifolds; intrinsic coordinates;
double connectivity of the rotation group; homomorphism of SU(2) onto
50(3) and of SL(2) onto ~; simplicity of the rotation and Lorentz groups.
Prerequisites: Chapter 18 and a little algebra.
Groups of matrices (or of the corresponding linear transformations) in

which the group elements depend continuously on certain parameters, like
the Euler angles in the case of the rotation group, are called continuous (the
formal definition will be given in Chapter 27). Here, properties of some
familiar continuous groups are described.
The group of all nonsingular (generally complex) n x n matrices is called
the general linear group and is denoted by GL(n, C). The subgroup of real
matrices is denoted by GL(n, IR). The subgroups of GL(n, C) and GL(n, IR)
consisting of matrices with determinant = 1 are called special (or unimodular)
linear groups and are denoted by SL(n, C) and SL(n, IR). The notation for
other subgroups of GL(n, C) will be given in the course of the discussion.
19.1 Orthogonal and Rotation Groups
In matrix-vector notation, the plane rotation (18.1-1) is

x -+ X' = Rx, (19.1-1)
where
x= (xy), X' = (Xy:), R = R

cP
= (cos qJ
sin qJ
-sin qJ).
cos qJ
(19.1-2)
Under the transformation (19.1-1), the length of any vector and the
angle between any two vectors are preserved, so that ifx' = Rx and Wi = Rw,
then x' . Wi = X w for any two vectors x and w.
25
26 Continuous Groups
We now look for transformations in n dimensions having this same pro-

perty; i.e., we call
Rll
and R =( :
Rn1
and we look for matrices R such that (Rx)' (Ry) = X' Y for all vectors
x and y. If ~(j) denotes the vector whose jth component is = 1 and whose
other components are = 0, then, in particular, R must be such that
Since the vector
is the jth column of R, it follows that the columns of R are pairwise orthogonal
unit vectors. Any such matrix is called orthogonal. Conversely, if R has that
property, then Rx' Ry = X' Y for all x, y. If RT denotes the transpose of R,
then the rows of RT are the columns of R, so that the law of matrix multi-
plication gives
(1 ... 0)

RTR =: . : =1 ...
I, (19.1-3)
... 1
which is another characterization of an orthogonal matrix. Since RT = R- 1
is the matrix ofthe inverse transformation, which also preserves dot products,
it follows that RT is also an orthogonal matrix; hence, the columns of R T,
that is, the rows of R, are another set of n pairwise orthogonal unit vectors.
Since det RT = det R, equation (19.1-3) shows that det R = 1. We now
define
O(n) = {R: R = n x n real orthogonal matrix}
( 0= matrix multiplication) as the orthogonal goup in n dimensions. Then,
the subgroup
SO(n) = {R E O(n): det R = I}
is the special (or unimodular) orthogonal group in n dimensions. We can

think of these groups either as groups of matrices or as groups of the cor-
responding transformations x -+ Rx of n-space.
The Rotation Group SO(3); Euler's Theorem 27
19.2 The Rotation Group SO(3);

Euler's Theorem
We now show that if R is a 3 x 3 orthogonal (real) matrix with det R = 1,

then the transformation x --+ Rx can be obtained by first choosing a fixed
direction in space through the origin and then rotating the coordinate
system through a suitable angle about this direction as an axis. This is
Euler's theorem. Let Ai and Vi (i = 1,2, 3) be the eigenvalues and eigen-
vectors of R (they may be complex, even though R is real); they satisfy the
equations
(i = 1, 2, 3). (19.2-1)
Since R is also a unitary matrix, we have IIRvil1 = Ilvill, where, for any
(generally complex) vector v, Ilvll denotes(lv x l2 + IVyl2 + IV z I2 )1 /2;therefore,
lAd = 1 (i = 1,2,3). (19.2-2)
The Ai are the roots of the real cubic equation
det(AI - R) = 0, (19.2-3)
and the product of the roots is
(19.2-4)
At least one of the roots is real; if the other two (say A2 and A3) are complex,
then A3 = X2 and, by (19.2-2), A2 A3 = 1; hence A1 = 1. If all three roots are
real, they can be 1, 1, 1 or 1, -1, -1. In any case there is always one root,
say A10 equal to + 1; hence
which shows that the straight line through the origin in the direction of V1
(V1 can be taken as real) is invariant under the transformation x --+ Rx;
evidently this line is the axis of rotation.
[Reminder. If M is any normal matrix, i.e., a matrix that commutes with

its Hermitian conjugate, i.e., if MM* = M*M (in particular, Hermitian
matrices and unitary matrices are normal), then the eigenvectors of M can
be taken to be a complete orthonormal (with respect to the Hermitian inner
product) system of vectors in the n-dimensional space. If the eigenvalues are
all distinct, the eigenvectors are automatically orthogonal; if A is an r-fold
eigenvalue, the corresponding eigenspace is r-dimensional, and one can
choose an orthonormal system in it; if that is done for each eigenspace, a
complete orthonormal system results.]
Let Al = 1, A2 = ei6 , A3 = e- i6, and suppose that Vb V2, V3 form an ortho-

normal set (they are the eigenvectors). Call
1 .
U2 = .J2 (VI + V2), (19.2-5)
i
U3 = .J2 (VI - v 2);
these also form an orthonormal set (they can all be taken as real, for V2
and V3 can be taken as complex conjugates), and
RU 1 = u 1,
RU2 = cos OU2 + sin OU3, (19.2-6)
RU 3 = -sin OU2 + cos OU3'
It is seen that the transformation of R is a rotation in planes perpendicular
to Ul'
The practical calculation of the angle and axis of rotation, when the
matrix R is given, proceeds as follows: Since the sum of the eigenvalues
of a matrix is equal to its trace, the angle 0 is given by
or
cos 0 = !(R l l + R22 + R33 - 1). (19.2-7)
Next, the axis of rotation is in the direction of the eigenvector V (called VI
above) that corresponds to the eigenvalue A = 1; hence Rv = v. But, since
R is an orthogonal matrix, RTR = J; hence V = RTv. Therefore, (R - RT)V
= 0, so that the components VI, V2' and V3 of v are in the ratio
(19.2-8)
EXERCISE
Make a similar analysis of the group SO(n) for arbitrary n.
19.3 Unitary Groups
Orthogonal groups can be generalized to the complex case in two ways.

First, a complex orthogonal matrix is any complex matrix M which satisfies
MTM = J, just as for the real case. The resulting groups do nofseem to be
of much importance.
Second, a matrix U such that U*U = J (then UU* = J) is called a unitary
matrix. Under a unitary transformation X --+ Ux of the n-dimensional
The Lorentz Groups 29
complex space en, the Hermitian scalar product

n
of any two vectors x and y is an invariant. Conversely, if this form is in-

variant, for all x and y, then U is unitary. The unitary group U(n) is the group
of all n x n unitary matrices (or unitary transformations in en).
Since det U* is the complex conjugate of det U, the equation U*U = I
shows that Idet U I = 1, i.e., det U is a number on the unit circle in the
complex plane. The subgroup of U(n) consisting of unitary matrices such
that det U = 1 is called the special (or unimodular) unitary group, and is
denoted by SU(n).
19.4 The Lorenz Groups
According to the theory of special relativity, if x, y, z, t and x', y', Zl, t' are
Cartesian coordinates in two inertial frames of reverence whose axes are
parallel, but are such that the second frame is moving relative to the first
with speed V in the + x direction, and if the origins are coincident at time
t = t ' = 0, then
y' = y, Z' = z,
Other inertial frames can be obtained by relative velocities in other directions,

by rotations in space, by displacements of the origin of space-time, by spatial
reflections, and by time reversal. When reflections and time reversal are
included, we have the full Lorentz group, otherwise the proper Lorentz
group. Displacements of the origin will not be discussed here, so that the
transformations considered are homogeneous, i.e., the equations contain no
constant terms.
If the notation (customary in relativity theory)
(sometimes, ct is called XO), is introduced, and qJ is defined by the equation
Vic
sinh qJ = (19.4-2)
then (19.4-1) is written as

4
X 'll = L p~xv, (19.4-3)
v= 1
where the coefficients p~, when written in matrix format, are

cosh cP o 0 sinh CP)
1 0 0
[p~J = peep) =
(
~ o 1 0 .
(19.4-4)
sinh cP o 0 cosh cP
In the remainder of this section, the summation convention will be used,
according to which any term containing a repeated index, say v as in (19.4-3),
is understood to be summed for v = 1, 2, 3, 4, so that, with this convention,
(19.4-3) is written simply as x'" = p~xv. In relativity theory, Greek indices
usually go from 1 to 4, and Latin indices from 1 to 3.
The set {P( cp)} of matrices (or transformations) of this kind, obtained by
letting cP take on all real values, is a group that will be denoted by 2 x-it
is a subgroup of the Lorentz group. Note that
(19.4-5)
from this equation, the law of composition of (collinear) velocities can be
obtained, that is, the formula for the velocity with which a third frame moves
relative to the first in terms of the relative velocities of the second with
respect to the first and of the third with respect to the second; the derivation
is left as an exercise.
If the second frame is obtained by merely rotating the first one in space,
then the transformation is given by a matrix of the form
R
= ((R') ~)
0 ' (19.4-6)
000 1
where R' denotes a 3 x 3 proper rotation matrix-an element of SO(3).
The set of all such matrices (or transformations) is the rotation subgroup of
the Lorentz group, and will be denoted by r!lt.
The group generated by the elements of 2 x together with those of fJ,
i.e., the group consisting of all finite products Ql Q2 ... Qj' where each Qi
is either of the form (19.4-4) or of the form (19.4-6), is called the proper (or
restricted) Lorentz group and is denoted by 2 p' This group is connected, in
the following sense:
Lemma. IfQo is any element of 2 p , there is a one-parameter family Q(A)

of elements of 2 p such that the matrix elements all depend continuously on A,
for 0 ::; A ::; Ao, and such that Q(O) = I, Q(Ao) = Qo.
PROOF. First, any element P(cp) of 2?x can be connected to the identity by letting cp
vary continuously from zero to its final value; second, any element R of ,'!/l can be
connected to the identity by letting the angle of rotation vary continuously from zero
to its final value; therefore, if Qo = Q 1 Qz ' .. Qj' where each Qi is in 2? x or fYl, then the
interval [0,10 ] can be taken as [O,j], and Q(l) can be chosen so that Q(O) = I,
Q(l) = Q1' Q(2) = Q1 Q2' and so on, and finally QU) = Q1 Q2 ... Qj' with continuous
variations in between.
Inspection shows that the fundamental (quadratic) form (Xl)2 + (X 2)2

+ (X 3)2 - (X4)2 is an invariant under all transformations of 2 x and of f!Il,
hence also under all transformation of 2p. This form is written as gl'vxl'X V
,
where
o o
o
o
1 o
o -1
~)= G,
(19.4-7)
and where the summation convention applies to both J.l and v. If the 4-vector
xl' in gl'vxl'X is replaced first by xl' + yl', then by xl' - yl', and the results
V
subtracted, it is seen that the in variance of the quadratic form gl'vxl'X is V
equivalent to the invariance of the symmetric bilinear form gl'vxl'yV =

xlyl + x 2y2 + X3y 3 _ x4y4.
The full (homogeneous) Lorentz group is now defined as the group of all
homogeneous linear transformations of Xl, ... , X4 under which gl'vxl'X is V
invariant for all 4-vectors xl'; it is denoted by 2 f' If the transformation

xl' --+ x'l' = qex is any element of 2 f' the invariance of the bilinear form
V
shows that gl'vq~q~ = g"A or, in matrix notation, that
(19.4-8)
in other words, the columns of Q are pseudo-orthogonal pseudo-unit

vectors, in the sense that
for v = 1, 2, 3,
(19.4-9)
for v = 4,
and
for v =J A. (19.4-10)
From this it follows that the inverse of Q is its transpose with signs changed
according to the pattern
Since Q-l is also a Lorentz transformation, it follows that the rows of Q

are also pseudo-orthogonal pseudo-unit vectors.
Equation (19.4-8) shows that the determinant of Q is 1. If {Xl, ... , X4}

is taken as {O, 0, 0, 1}, the invariance of the fundamental form shows that
3 3
L (X,j)2 - (X'4)2 = L (qi)2 - (q!)2 = -1;
j= 1 j= 1
therefore, q! is either ~ 1 or ~ - 1.
Theorem. The proper Lorentz group fi'p, defined above as the group
generated by fi' x and ~, consists of all those transformations Q of fi' f for
which det Q = + 1 and q! ~ + 1.
PROOF. It is shown first that the connectedness of the group .Ii' p implies that
det Q = + 1 and that q! ~ 1 for any Q in .Ii' p: Let Q be connected to the identity J,
as in the lemma; since Q(O) = J, we have det Q(O) = 1 and q(O)! = 1; det Q(A.) and
q(A.)! are continuous, and hence cannot jump to negative values as A. varies. (That
det Q is equal to 1 can also be seen directly from the decomposition Q = QIQ2'" Qj,
where each Qi is either in .Ii'x or in 9l). Conversely, let Q be any transformation in .Ii'f
such that det Q = 1 and q! ~ 1. It will be shown that Q can be expressed as RIPR 2.
where R 1 and R2 are in 9l and P is in .Ii'x; hence, Q is in .Ii' p' (This shows, furthermore,
that three factors always suffice in the decomposition QI Q2 ... Qj.) First, let R3 and
R4 be rotations that take the 3-vectors (q1, qi, ql) and (qt, qt q~), respectively, into
the direction of the positive Xl axis. Then,
q~l)
R3 QR 4 = ( (X') ~ = Q', (19.4-11)
q~4 0 0 q~4
where X' is some 3 x 3 matrix and where q~l and q'14 are ~ 0, and q~4 = q! ~ 1. Since
(q~I)2 _ (q~4)2 and (q'14)2 _ (q~4)2 both = -1, a parameter <p can be chosen so that
q~4 = cosh <po
Therefore, ifQ' is multiplied by P( - <p), the last row and the last column of the product
are the same as those of P( - <p )P(<p) = J; hence
P(-.)Q' ~( (X") : ) = Q".
001
This matrix, like Q and Q', is in .Ii'p, and hence leaves the fundamental quadratic form
invariant. Therefore, X" leaves (Xl)2 + (X 2)2 + (X3)2 invariant, hence is in 0(3),
but det X" = + 1, hence X" is in SO(3); i.e., Q" is an element of 9l, say Rs. That is,
P(-<P)R3QR4 =R s ,
and it follows that Q is of the required form
Comment. Elements of the form L = R - lpR, where R is in f7l and P is in 2 x

are called pure Lorentz transformations or boosts: In this case the primed
and unprimed axes are still parallel, just as for P( qJ), but the relative velocity
can be in any direction. L is a symmetric matrix, but if Ll and L2 are two
pure Lorentz transformations, it is readily seen that L 1 L2 is not symmetric
unless the relative velocities are parallel; hence, the pure Lorentz transforma-
tions do not form a subgroup of 2 p' Furthermore, if L1L2 is not symmetric, a
third pure Lorentz transformation L3 can be found such that L1L2L3 is a
pure rotation (an element of f7l), not = I. This leads to the pseudoparadoxical
result that a body can be rotated by subjecting it to three successive stages
of purely linear acceleration, in three different directions, in such a way that
the body is brought back to rest by the combined action of the three stages.
This phenomenon, expressed a little differently, was discovered by L. H.
Thomas in connection with relativistic corrections to atomic energy levels, in
the early days of quantum theory. [An electron, orbiting a nucleus, undergoes
a (continuous) sequence of Lorentz translations, somewhat as described,
and an effect similar to that of electron spin results.]
The group 2t generated by the elements of 2 p together with the trans-

formation of time reversal, given by the matrix
o o
1
o
o
1 ~),
o o -1
is a subgroup of 2 J because T clearly preserves the fundamental form.

Similarly, the group 2 s generated by the elements of 2 p together with spatial
inversion, given by the matrix
-1 o o
S=( 0
-1 o
o o -1
o o o
is a subgroup of 2 J' Lastly, 2 J itself is generated by the elements of 2 p
together with T and S.
EXERCISES
Prove or disprove the following statements:

1. Every proper Lorentz transformation with a symmetric matrix is a pure
Lorentz transformation.
2. If PI and P z are two pure Lorentz transformations, then PIP zp 1 1P 2 1 is a
pure rotation.
3. The rotation subgroup is a normal subgroup of !l' p'
The Lorentz groups can be generalized in an obvious way to groups of

transformations that preserve the value of a fundamental form
(X 1)2 + ... + (xr)2 _ (xr+l)2 _ ... _ (xr+l)2;
the possibility of spatial reflections appears when r is odd, and of time reversal
when I is odd.
19.5 Group Manifolds
If each element
of the orthogonal group 0(3) is represented by a point in a 9-dimensional

(real) space V 9 , with coordinates R l l , R 12 , ... , R3 3 (taken, let us say, in
the order in which they appear in the matrix), then the group is represented
by the set of points in V 9 that satisfy the six algebraic equations
RL + RL + RL = 1, j = 1,2, 3, (19.5-1)
RljRlk + R2jR2k + R3jR3k = 0, (j k) = (1 2), (1 3), (2 3),
(19.5-2)
which assert that the columns of R are pairwise orthogonal unit vectors, as
required by the discussion in Section 19.2. The resulting 3-dimensional
algebraic surface in V 9 is called the manifold of the group 0(3). The exact
size, shape, and curvature of this surface are of no interest, but its overall
topological properties are closely connected with the structure of the group.
With each continuous group of linear transformations is similarly as-
sociated an algebraic surface in some v rn However, care must be used in
drawing conclusions from the number of equations. For SO(3), in addition
to the six equations (19.5-1,2) above, there is a seventh algebraic equation
det R = 1. (19.5-3)
As will be seen, the surface determined by the first six equations has two
separate parts or components, like the two branches of a hyperbola, and
equation (19.5-3) merely eliminates one of the components, but does not
reduce the dimensionality of the surface. For the full linear group GL(n, ~),
there is only an algebraic inequality det M =f. 0; for GL(n, IC) there is the
algebraic inequality (Re det M)2 + (1m det M)2 =f. 0.
EXERCISE
Find an algebraic equation F(x, y, z) = 0 which determines the surface of the torus
in V 3 just as the equation x 2 + y2 + Z2 - 1 = 0 determines the surface of a sphere.
Find a group of which the torus is the group manifold.
Intrinsic Coordinates in the Manifold of the Rotation Group 35
19.6 Intrinsic Coordinates in the Manifold of the

Rotation Group
Although any manifold can be regarded, in principle, as an n-dimensional

surface embedded in a space vm of higher dimension, as in the foregoing
examples, the embedding is not always easy to find or describe, and the
manifold is more appropriately described by intrinsic coordinates, like the
polar angles e, cp on the surface of a sphere, or more generally by two or
more overlapping systems of intrinsic coordinate systems.
Each transformation of the group SO(3) can be obtained by first choosing

a fixed axis and then performing a rotation about that axis (Section 19.2).
If R is the matrix of a rotation through the angle e ;;:: (seen as clockwise
when looking along the positive direction of k) about an axis having the
direction of the unit vector k, then the numbers ek x, ek y , ek z (= ex, ey , ez )
may be taken as intrinsic coordinates in SO(3), and we write R = R(9),
where 9 is the vector ke. To make the coordinate system unique, 9 must be
restricted to the spherical ball K = {9: 11911 :::; n} in the coordinate space [R3,
in which ex, ey , and ez are taken as Cartesian, and we must observe that
opposite ends of a diameter in K correspond to the same element of SO(3)-
otherwise, each point of K corresponds to a unique element of SO(3), and
conversely.
The matrix R = R(9) is given explicitly in terms of the intrinsic coordinates
by the equation
R(9) = exp ( ~z (19.6-1)

-e y
To see this, note first that since R is a nons in gular normal matrix, its logarithm
A is well defined (though multivalued), so that
therefore, A is antisymmetric and hence must be of the form
A= (-~
-b
~ o~). -c
To see that a, b, and c have been correctly interpreted in (19.6-1), note that,
with this interpretation, (l) the eigenvalues of A are 0, ie, where e =
J e; + e; + e;, so that the eigenvalues of exp A are 1 and e i8, (2) the first
eigenvector of A (hence also of exp A) is proportional to 9, (3) since A is a
normal matrix, its eigenvectors can be taken as an orthonormal set, and (4)
it now follows, just as in Section 19.2, that R represents a rotation through
an angle e about an axis in the direction of 9. It is left as an exercise to verify
that (19.6-1) defines R(O) rather than R( -0), if the coordinate system is
right-handed .
.By means of these intrinsic coordinates ex, ey, ez , the properties of the
surface !I' in V 9 determined by the algebraic equations (19.5-1,2,3) can be
found. Each point of the ball K corresponds to a unique point of the surface
!I', except that opposite ends of a diameter in K correspond to the same point
of !I', and the nine coordinates of a point in !I', are continuous functions of
the intrinsic coordinates in K, according to equation (19.6-1). Therefore, !I'
is a connected surface, because any point of the ball K can be connected to
any other point of K by a curve (in fact by a straight line segment) lying in K.
However, !I' is not simply connected.
A connected surface is called simply connected if, given any two points
A and B and any two curves C 1 and C 2 going from A to B in the surface, C1
can be deformed into C 2 by a continuous deformation without leaving the
surface. The plane and the sphere are simply connected, while the torus, the
surface of a cylinder, and the annulus Ri < x 2 + y2 < R~ are not. Among
solids, the ball, the cube, and the spherical shell Ri < x 2 + y2 + Z2 < R~
are simply connected, while the solid torus and the pretzel are not.
To show that the manifold !I' of 80(3) is not simply connected, let A be
the point of !I' corresponding to the center of the ball K, and let B be the point
of !I' that corresponds to the two ends of a diameter of K. Then the two radii
which make up this diameter correspond, in !I', to two curves or paths going
from A to B, and it is evident that neither can be deformed into the other.
However, any other path from A to B can be continuously deformed, in !I',
into one of these two.
The Homomorphism of SU(2) onto SO(3) 37
19.7 The Homomorphism of SU(2) onto SO(3)

Let x, y, z be Cartesian coordinates in a Euclidean space E3, and call A
the matrix
A = C: iy x =:y) , (19.7-1)
which is evidently Hermitian and of trace zero. (The trace of a matrix is the
sum of its eigenvalues and is equal to the sum of its diagonal elements.) Let
U be any 2 x 2 unitary matrix of determinant = 1 [i.e., U E SU(2)], and call
A' = UAU*. (19.7-2)
Since the eigenvalues of A' are the same as those of A, the trace of A' is
also zero; A' is also Hermitian (A'* = A'), so it can be written as
A' = (z,
x' + iy'
x' - iY')
-z' ,
(19.7-3)
where x', y' and z' are real. Furthermore, det A' = det A, because of (19.7-2),
so
(19.7-4)
The relation between x, y, z and x', y', z' is obviously linear for given U;
hence, if we define a 3 x 3 real matrix R = R( U) by
then R is an orthogonal matrix.
A General Comment on Orthogonal, Unitary, and Lorentz Transformations

Suppose that a homogeneous linear transformation x ---> x' in real n-space
is such that the quadratic form X' x = xi + ... + x~ is invariant"; i.e.,
x . x = x' . x', for all x. Then the bilinear form x . y is also invariant, for any
x, any y, because
x y = t(x + y) . (x + y) - tcx - y) . (x - y).
That is, if all lengths are preserved, then all angles are preserved, too.
Similarly, if (-, .) denotes the Hermitian symmetric (i.e., sesquilinear) inner
product in complex n-space, then (x, y) is invariant, for all x, all y, if and only
if (x, x) is invariant, for all x. Therefore, orthogonal and unitary groups can
be characterized by invariance of either the quadratic, or the corresponding
bilinear form.
EXERCISE
Formulate the analogous result for the Lorentz group.

Furthermore, det R = 1, by continuity, since the only possible values of

det Rare 1, since det R depends continuously on U, and since SU(2) is a
connected group. [If U is the 2 x 2 unit matrix, then R(U) is the 3 x 3 unit
matrix and has determinant = 1.J R(U l)R(U 2) is equal to R(U 1U 2), because
R(U i )R(U 2) is the result of transforming A first into U 2AU!, then into
U 1 U 2AU!U!. Therefore the mapping
<1>: U ~ R(U)
is a homomorphism. We state without proof that it is onto.
The kernel is easily seen to be {I, - I}, so with each R E SO(3) there are
associated two elements U and - U of SU(2). It will be proved in Section
19.9 that SO(3) is simple, so no further nontrivial homomorphism of SO(3)
is possible.
19.8 The Homomorphism of SL(2, C) onto the

Proper Lorentz Group !l' p
The homomorphism found in the preceding section is very easily extended.
Let x, y, z, t be coordinates in space-time (in units such that c = 1). Then
the matrix
A= (t + ~ x - iY)
x + lY t - Z
is Hermitian, and det A = t 2 - x 2 - y2 - Z2. If P is an arbitrary 2 x 2

(complex) matrix with det P = 1, then the matrix A' = PAP* is also
Hermitian, and hence can be written as
A'= ( t , + Z' x' - iY')

x' + iy' t' - z' .
Now det A' = det A; hence
That is, the matrix P of SL(2, C) induces a Lorentz transformation T. Clearly

the mapping P ~ T has the homomorphism property that if Pi ~ Ti and
P 2 ~ T2, then PiP 2 ~ Ti T2 . It is in fact a 2-to-1 homomorphism of SL(2, C)
onto !l'p.
19.9 Simplicity of the Rotation and Lorentz Groups
Theorem. The rotation group SO(3) is simple

PROOF. Suppose that Go is a nontrivial normal subgroup; Go <l SO(3). It must be
proved that Go = SO(3). Let Ro EGo, Ro =I- I. It was proved in Section 6.4 that, for
any R in SO(3), the transformation x --> Rx can be described as a rotation through
Simplicity of the Rotation and Lorentz Groups 39
some angle 8 about a fixed direction. Let 0 be a vector in that direction and having
length 11011 = 8, and call R = R(O), as in Section 19.6. Then Ro = R(Oo) for some
0 0 i' O. Now let 0 1 be any other vector having the same length as 0 0 , 110 1 11 = 1100 11,
and let RI be a rotation that carries 0 1 into 0 0 . Then Rjl R(Oo)RI is in the subgroup
Go, because Go is normal; but RjIR(Oo)RI = R(OI)' and hence the subgroup Go
contains every R(O) for 11011 = 1100 11. Next, Go also contains every element R(O)R(Oo),
for 11011 = 1100 11; this is R(O') for some 0' = 0'(0,0 0). Clearly 110'11 is a continuous
function of the components of O. [Recall the explicit formula for R(O) in terms of
8x , 8y , 8z given in Section 19.6.J For 0 = -00 and +0 0, 118'11 = 0 and 2118 0 11, re-
spectively. Therefore, for any 8 in [0, 2110011J, Go contains at least one element R(O)
such that 11011 = 8. By the first argument, again, Go therefore contains every R(O) such
that 0 :::; 11011 :::; 2110 0 11; but for any such R(O), Go also contains R(O)R(O) = R(20),
R(30), etc.; hence Go contains every R(O) for 0 :::; 11011 :::; n, that is, Go = SO(3).
The proper Lorentz group is also simple, although the proof is more
complicated.
CHAPTER 20
Group Representations I:
Rotations and Spherical Harmonics
Representation; faithful representation; dimension; reducibility; irreducible

representation; tensor transformation laws; representations of 50(2) and
50(3); infinitesimal operators; raising and lowering operators; effective and
transitive action of a group; homogeneous space; regular representation;
tesseral harmonics; Legendre polynomials and associated Legendre
functions; recurrence relations and the differential equation for PI;
Rodrigues' formula for PI; orthonormality and completeness of the tesseral
harmonics; addition theorem.
Prerequisites: Chapters 18 and 19; elementary theory of matrices; famil iarity

with the special functions of mathematical physics.
In this chapter and the next two, three apparently remote subjects are shown
to have important interconnections: group representations, the classical
special functions, and quantum mechanics.
Group representations are intimately connected with various special
functions of mathematical physics. In a sense, the primary role of those
functions is to exhibit symmetry relations. For instance, the Legendre
functions appear (via spherical harmonics) in problems with sl?herical
symmetry in such diverse fields as electrostatics, acoustics, heat flow,
neutron transport, and the quantum mechanics of hydrogen-like atoms.
Bessel functions (of integer or half-odd-integer order) appear mostly in
problems of wave motion, but a closer examination shows that they are
associated more with certain symmetries than with the mechanism of the
waves. In fact, wave motion in a nonconstant (even spherically symmetric)
potential generally involves other functions, for example Laguerre functions
in hydrogen-like atoms, while Bessel functions appear when the system is
invariant under the full rigid-motion groups, not merely the rotation sub-
groups-see next chapter. It will be seen in Section 20.5 that the trigonometric
functions appear, in the form eim"" in the representations ofthe two-dimension-
al rotation group, and hence are associated with symmetry about an axis.
40
Vector and Tensor Transformation Laws 41
20.1 Finite-Dimensional Representations of a Group
Let G be any group, and X an n-dimensional (often complex) vector space.

A homomorphism p: g -+ p(g) of G onto a group of linear transformations
in X is called an n-dimensional representation of G (on X). The linear ,trans-
formations are normally described by matrices, and the matrix that describes
the transformation p(g) is also denoted by p(g). Hence, we may also :r;egard
a representation of G as a homomorphism of G onto a group of matrices;
then the identity in G, the inverse of an element in G, and the law of compo-
sitions in G are represented by the unit matrix, the matrix inverse, and the
matrix product. If the homomorphism is an isomorphism, then the repre-
sentation p is called faithful. If G itself is a group of linear transformations in a
vector space Y, then the identity mapping g -+ g of G onto itself is a faithful
representation. Even in this case, however, it is worthwhile, for many
reasons, to consider other representations also.
Note. The usage here differs slightly from that of Section 18.10, where a
representation of G was a homomorphism of G onto any group of (not
necessarily linear) transformations.
If X contains a proper subspace Xl that is invariant under all the ttans-

formations p(g), g E G, then the representation is reducible, and the relStric-
tion of the transformations p(g) to Xl gives a representation of smaller
dimension, called a subrepresentation. If there is no such proper subspace
X 1> P is called irreducible.
If G is a continuous group (a Lie group), it is also required that the matrix
elements of p(g) depend continuously on g. The only continuous groups to
be considered here are matrix groups, like the unitary, orthogonal, and
Lorentz groups. In this case the requirement is that the matrix elemeQts of
p(g) be continuous functions of the matrix elements of g, as is immediately
seen to be the case in all the examples below.
20.2 Vector and Tensor Transformation Laws
We shall now describe some naturally occurring group representatio*s in

physics. According to Chapter 19, the transformations in 3-dimensional
space that result from rotations of the Cartesian axes about the origin
constitute a group called SO(3). Various representations of this group are
given by the transformation laws for the components of tensors. If the
Cartesian coordinates are transformed by x -+ x', where
3
x; = L gikXk, (20.2-1)
k= 1
42 Group Representations I: Rotations and Spherical Harmonics
and where (gik) is a rotation matrix [an element of SO(3)], then the com-
ponents Ii j of a second rank tensor transform according to the law
T;j = L gikgjll1d'
(k,l)
(20.2-2)
If the nine quantities Ii j are called Xl, ' , , , X 9 and are regarded as the co-
ordinates of a point X in ~9, then each transformation x ---+ x' induces a
transformation X ---+ X', and those transformations give a representation of
SO(3) on ~9,
Transformation ofthe components of a third rank tensor give a representa-
tion on ~Z 7, and so on,
More special representations can appear when certain symmetry relations
exist among the tensor components, Suppose that the Ii j are the com-
ponents of the strain-rate tensor at some point in a fluid:
avo
Ii j = -a 1,
Xj
where v = v(x) is the velocity vector field. If the flow is irrotational, then
Ii j = Tji (i,j = 1,2,3). (20.2-3)
It is readily verified that these equations are invariant under rotations;
that is, if Ii j = Tji for all i andj, then T;j = Tji (see Exercise 2 below). Hence,
in this case, only six of the Ii j are independent, say the quantities Yi defined as
Y1=T1Z , YZ = TZ3 , Y3 =T31 ,
Y4 =T1 1> Y5 = TZ2 , Y6 =T33
Then the rotation x ---+ x' induces a linear transformation Y ---+ Y', and a
6-dimensional representation of SO(3) results.
If, furthermore, the fluid is incompressible, there is an additional relation
(20.2-4)
which is also invariant under rotations (see Exercise 3 below). In this case
Y6 can be dropped (i.e., it can always be computed as - Y4 - Y5 ), and the
transformations of Y1 , .. , Y5 give a 5-dimensional representation of SO(3).
The relations (20.2-3) and (20.2-4) determine subspaces of ~9 that are
invariant under the transformations (20.2-2); the invariance of those sub-
spaces permits the 9-dimensional representation to be reduced to the 6-
dimensional and 5-dimensional ones. [An 8-dimensional representation can
be obtained by using (20.2-4) alone.] It will appear below that the 5-dimension-
al subspace cannot be further decomposed into smaller invariant subspaces;
hence the 5-dimensional representation is irreducible. It will appear that for
each odd integer m there is an irreducible m-dimensional representation of
SO(3); when m > 1, that representation is faithful.
The original transformation law (20.2-1) for the components of a vector
x provides of course a 3-dimensional representation. (If the rotations are
regarded as constituting an abstract group, then that representation is no
Vector and Tensor Transformation Laws 43
more trivial than the others.) Furthermore, for the sake of completeness
(in a sense to be made precise later), we include the 1-dimensional representa-
tion given by the transformation law for scalars, which says simply that they
are not transformed at all. Each scalar is a real number x, and each group
element (rotation) is mapped onto the identity transformation x -+ x in IR.
[It should not be concluded that a 1-dimensional representation of a group
always consists only of the identity transformation. A 1-dimensional repre-
sentation of GL(n, IR) or GL(n, IC) is given by mapping the group element
M-an n x n matrix-onto the transformation x -+ (det M)x in IR or C.]
Representations of the Lorentz groups (see Section 19.4) are given similarly
by the transformation laws of scalars, vectors (i.e., 4-vectors), and tensors,
under Lorentz transformations in space-time; they are of dimension
1,4,16, ....
A 6-dimensional representation of the (restricted) Lorentz group !l' p
is given by the transformation law for the components of the electric and
magnetic fields E and H in free space. (The interpretation of this transforma-
tion law in terms of tensors in explained in Exercise 5 below.) Under the
particular Lorentz transformation (19.4-1), the electric and magnetic field
components transform according to the law
E~ = Ex, E~ = Y( Ey - ~ H} E~ = Y( Ez+ ~ Hy),

(20.2-5)
H~ = Hx, H~ = Y(Hy + ~E} H~ = Y( Hz - ~Ey),
where
Y= (1 _~~rl/2.
(See any good book on electromagnetic theory.) Under a rotation, the
components of E and those of H transform independently according to
the usual vector law (20.2-1). According to Section 19.4 any element of!l'p
can be written in the form R 1 P R 2, where R 1 and R2 are rotations in space and
P is a transformation of the form (19.4-1), hence the general transformation
for E and H can be obtained by combining (20.2-1) and (20.2-5).
One of the aims of the theory of group representations is to find all possible
transformation laws for physical quantities, that is, all possible repre-
sentations of physical symmetry groups. As seen in the above examples, there
are two main procedures; the building up of representations from simpler
ones by means of tensors, and the breaking down of representations into sub-
representations. Still another procedure, based on the action of a group on
spaces of functions, is described below, starting in Section 20.5.
EXERCISES
1. Show that the representation of rotations (20.2-1) by the transformations
(20.2-2) is a homomorphism, as required. That is, show that if peg) is the transforma-
tion induced in [R9 by the rotation g, then p(g)p(g') = p(gg').
2. Show that the symmetry or anti symmetry of a second rank tensor is preserved
under rotations, when the law (20.2-2) is used; that is, show that if 1';j = ~; (or 1';j =
-~) for all i,j, then" T;j = Tj; (or T;j = - Tj;) for all i,j.
3. Show that the trace of a second-rank tensor is preserved under rotation; that is,
show that
T'J J + T~ z + T~ 3 = T) J + Tz z + T3 3 .
4. Consider a general linear transformation of the form (20.2-1) in rr;\;", where (g;)
is an n x n nonsingular matrix, and show that the transformation is orthogonal
(gT 9 = ggT = 1) if and only if the trace of every second rank tensor is invariant under
(20.2-2).
5. Show that tbe transformation law (20.2-5) for the electromagnetic field under
the Lorentz transformation (19.4-1) can be obtained by transforming the anti symmetric
second rank tensor T" v defined by
according to the law
when the coordinates are transformed according to

4
x'tl = L q~xa.
u=l
20.3 Other Group Representations in Physics
Pauli's 1927 theory of electron spin and Dirac's 1928 relativistic wave equation
led to transformation laws under rotations and Lorentz transformations of a
different kind from the familiar transformation laws for vectors and tensors.
That led in turn to the theory of new objects called spinors, which take their
place in relativistic quantum mechanics alongside scalars, vectors, and
tensors, and are discussed in Chapter 22. The transformation laws for spinors
give so-called two-valued representations of SO(3) and Y p' which, however,
are true representations of the covering groups SU(2) and SL(2, C) discussed
in Sections 19.7 and 19.8. Why the latter groups should appear at all in
physical problems seemed a paradox and is still left quite unclear in most
books on quantum mechanics. The resolution of the paradox, due mainly
to Hermann Weyl, is described in Chapter 22. It involves the so-called ray
representations, which are not true representations (as defined in this chapter),
but are nevertheless appropriate for quantum mechanical phenomena. In
Wey11928 it is shown that the ray representations of a group are precisely
determined by the true representations of its covering groups; hence,
the representations ofSU(2) and SL(2, C) playa role. The theory shows further
Infinite-Dimensional Representations 45
that since the manifolds of SU(2) and SL(2, IC) are simply connected, these
groups are the so-called universal covering groups of SO(3) and !l' p' re-
spectively, (see Chapters 24 and 27) and that, in consequence, there are no
multivalued representations of SO(3) and !l' p of multiplicity > 2. It is thus
concluded, on the basis of group theory, that vectors, tensors, and spinors
provide the only possible transformation laws of quantum mechanical
phenomena.
In classical physics a distinction is made between polar vectors (like
momentum and electric field) and axial vectors (like angular momentum
and magnetic field). The former change sign under an inversion x ~ - x,
while the latter do not. Hence there are two (in fact only two) ways of ex-
tending the transformation law for vectors to the full orthogonal group
0(3).
The symmetry or anti symmetry of a many-particle wave function under
interchange (or more generally permutation) of identical particles gives a
simple representation of the relevant permutation group. More complicated
representations of the permutation groups appear in the theory of para-
statistics.
20.4 Infinite-Dimensional Representations

Infinite-dimensional representations, for example of the Lorentz and
Poincare groups, are encountered in applications like quantum field theory.
Normally, such a representation is a homomorphism of the group G onto
a group of bounded linear transformations peg) in a Banach or Hilbert
space X. If G is a continuous matrix group, peg) is required to be a strongly
continuous function of g; that is, it is required that for every u in X,
IIp(g)u - p(go)ull ~ 0,
as the matrix elements of g converge to those of go.
While all irreducible representations of compact groups like SO(3) are
finite-dimensional, there are infinite-dimensional representations of the
Lorentz, Poincare, and rigid motion groups such that no finite-dimensional
subspace of the Banach or Hilbert space X is invariant under all the trans-
formations peg), gE G, which therefore cannot be broken down into finite-
dimensional subrepresentations. In the next chapter we describe infinite-
dimensional representations of the 2-dimensional rigid motion group M 2,
which is fairly typical, and which involves Bessel functions. We refer to
Boerner 1955, Gel'fand, Minlos, and Shapiro 1963, Gel'fand, Graev, and
Vilenkin 1966, Vilenkin 1968, Warner 1972, and Barut and Raczka (1977)
for further material on the subject (which is a fairly vast one, as the sizes of
those books indicates) and in particular for the representations of the Lorentz
and Poincare groups.
In Chapter 27 we give an example of a continuous group (Lie group) that
has no finite-dimensional faithful representation, and hence cannot be
realized as a matrix group at all.
20.5 A Simple Case: SO(2)
The general method of finding representations can be illustrated by the

(rather trivial) case of the group SO(2) of rotations about the origin in the
plane. Under this group any circle centered at the origin is invariant. Let
XOO be the infinite-dimensional space of infinitely differentiable functions
f(ep) defined on the unit circle. (For the present qualitative discussion, we
shall not introduce a norm or topology in this space.) Then, with the rotation
ga through an angle ~ [an element of SO(2)] we associate the linear trans-
formation
Pa: f(ep) ~ f(ep - ~) (20.5-1)
in the space XOO. The correspondence ga ~ Pa is a representation of SO(2)
on X 00" (In this example, the transformation f(ep) ~ J(ep + ~) would do
equally well-see Note in the next section.)
An important technique for finding subrepresentations is the use of the so-
called infinitesimal operators of a representation, such as the operator T
obtained in the present instance by differentiating the operator Pa with
respect to ~ at ~ = 0, namely
(Tf)(ep) = -f'(ep), (20.5-2)
so that T may be regarded as the limit of (1/~)(Pa - Po) as ~ ~ o. Evidently,
any subspace of XOO that is invariant under all the transformations Pa is
invariant under T. We look for invariant subspaces of smallest possible
dimension. Hence we look for a functionf(ep) such that the minimal subspace
that containsf(ep) and is invariant under T contains no other functions, or as
few such as possible. For a one-dimensional subspace Tf should be a
multiple off, say Af; that leads to the eigenvalue problem
-f'(ep) = AJ(ep). (20.5-3)
In the present example, since the functions in XOO are required to be single-
valued on the unit circle, the eigenvalues and eigenfunctions are
A = im (m = 0, 1, 2, ...),
(20.5-4)
f(ep) = fm(CP) = e- im",.
For each ~, the transformation Pa given by (20.5-1) also causes each eigen-
functionfm of T to be merely multiplied by a constant, namely
(Pafm)(ep) = eim'im(ep)
That is, for each m the 1- (complex) dimensional subspace
Xm = {Ae-imep: A E C} (20.5-5)
is invariant not only under T but also under all the transformations Pa.
By Fourier's theorem each function in XOO can be expanded in these functions,
hence the subspaces X rn span XOO.
Representations of Matrix Groups on XOO 47
It will now be shown that the only finite-dimensional representations

of SO(2) on XCO are those that can be built up from the ones obtained above,
for it will be shown that any finite-dimensional invariant subspace X' of
XCO is the direct sum of finitely many of the subspaces (20.5-5). Namely, let
J(ep) be any function in X' and write it as a Fourier series
J(ep) = L cmeiml{J. (20.5-6)
(m)
It will be proved that the subspace X' not only contains this sum but contains
each term cmeiml{J individually; hence, it contains the subspace X -m for each
m such that Cm oF O. Namely, since X' is invariant under all the transformations
(20.5-1) it contains all translatesJ(ep - oc) of the given functionJ(ep); hence
it contains any function of the form
K
L hd(ep -
k= 1
OCk),
where the hk and OC k are constants. By choosing this sum as a Riemann sum
that approximates an integral and then going to the limit, it is seen that X'
contains any function of the form
J h(oc)J(ep - oc)doc,
hence, in particular, the function
Therefore, X' contains X _m if Cm oF 0, as was to be proved. (Since X' was

assumed finite-dimensional, it is now seen that Cm is oF 0 only for finitely
many m).
For non-Abelian groups, the minimal invariant subs paces are generally
or more than one dimension.
20.6 Representations of Matrix Groups on Xoo
The ideas in the preceding section can be generalized. If G is a group of

n x n matrices, then its elements g determine linear transformations in an
n-dimensional space vn ( = IW or en). Let Xco denote the space of all infinitely
differentiable functions J(x) on vn or on any surface in vn that is invariant
under G. With each element g of G we associate a linear transformation
peg): J(x) -'> f'(x) = J(g-lX) (20.6-1)
of XCO into itself (the prime does not denote differentiation). If the transforma-
tion peg) = P(gl) is followed by another transformation P(g2), namely
P(g2): f'(x) -'> f"(x) = f'(gi1x),
the result is
P(g2)P(gl): f(x) --+ f"(x)
= f'(gzlx)
= z
f(g 1 1(g I X
= f((g2gl)-IX),
so that
(20.6-2)
Therefore, the association 9 --+ peg) is a representation of G on XOO.
Note. The appearance of 9 - I (rather than 9 itself) in (20.6-1) is necessary to

obtain the correct order of the factors in (20.6-2). It corresponds to the minus
sign in (20.5-1); for an Abelian group the order is of course irrelevant, hence
f(ep + ct) would have been equally satisfactory in (20.5-1).
Replacingf(x) by f(g-IX) is equivalent to performing the mapping x --+ gx

in vn and carrying the values of f along during the mapping, just as the
substitution f(t) --+ f(t - a) carries the values of f to the right on IR, for
a > 0, by a distance a.
If G is a continuous group and the group elements depend on parameters
ct, {3, ... , that is 9 = ga.p, ... ' where go, 0, is the identity elements of G, then
00.
the infinitesimal operators corresponding to the operator T in the preceding

section are given by differentiating p(ga, p, ooJ with respect to each parameter
ct, {3, ... in turn and then setting ct = {3 = ... = O. The infinitesimal operators
of SO(3) are given in Section 20.9.
20.7 Homogeneous Spaces
If the space vn contains a smooth surface S of some lower dimension that is

invariant under all transformations of G, like the unit circle in the case of
SO(2), then XOO can be replaced by a space of functions defined on S rather
than on all v n In the example above, S could be any circle centered at the
origin. If G is SO(3), S can be any sphere centered at the origin of V 3 In
each of these cases, the action of the group on S has the following properties:
a. The group G is effective for S; that is, only the identity e of G gives the
identity mapping in S; in other words, if 9 =f. e, then there is at least
one x in S such that gx =f. x.
b. The group is transitive for S; that is, if x and yare any two points of S,
there is a gin G that carries x into y, i.e., y = gx.
If S is invariant under G, and a and b hold, then S is called a homogeneous
space for G.
Regular Representations 49
Suppo.se, fo.r example, that G co.nsists o.f all transfo.rmatio.ns in real 3-space
o.fthe fo.rm
x ~ x' = ( ae bd fe) x,
001
where ad - be * O. Any plane X3 = co.nstant is invariant under G and
satisfies co.nditio.n b, but the plane X3 = 0 fails to. satisfy a, because every
po.int o.fthe plane X3 = 0 is mapped onto. itself under the mappings
x ~(0
1 0
1 f x.
e)
001
*
Hence, o.nly the planes X3 = co.nstant 0 are ho.mo.geneo.us spaces fo.r this
group.
A general procedure fo.r finding irreducible representatio.ns o.f a co.ntinuo.us
gro.up G o.f linear transfo.rmatio.ns in vn co.nsists o.f finding a ho.mo.geneo.us
space S fo.r G in V n , then letting p be the representatio.n f(x) ~ f(g-IX)
o.f G o.n a space XOO o.f functio.ns o.n S, usually L 2(S), then finding a co.mplete
set o.f infinitesimal o.perato.rs, and then finding minimal invariant subspaces
o.f XOO with the help o.f the infinitesimal o.perato.rs. The special functio.ns
asso.ciated with the symmetries described by G are the elements o.f the in-
variant subspaces o.f XOO.
The pro.cedure is described in mo.re detail in Sec. 20.9 belo.w fo.r the case
o.f the ro.tatio.n group SO(3). In that case, the invariant subspaces that will
be fo.und are all finite-dimensio.nal. It will be seen in the next chapter that the
same is true fo.r any co.mpact group. In that chapter, the questio.n whether all
irreducible representatio.ns are fo.und in this way is also. answered, fo.r co.m-
pact groups.
The theo.ry fo.r no.nco.mpact gro.ups, where infinite-dimensio.nal irreducible
representatio.ns can o.ccur, is co.nsiderably beyo.nd the sco.pe o.f this bo.o.k,
and we shall be co.ntent with a discussio.n o.f so.me o.f the main features o.f
three examples, in the next chapter: the rigid mo.tio.n group M 2, where
Bessel functio.ns appear, and the Lo.rentz gro.up and SL(2, IC), where spino.rs
appear.
20.8 Regular Representations

Fo.r Lie groups, the group manifo.ld itself can always serve as a ho.mo.geneo.us
space. Co.nsider G = SO(3), fo.r example. It is recalled that the manifo.ld o.f
SO(3) is a certain 3-dimensio.nal algebraic surface S in a space o.f 9 real
dimensio.ns; each po.int o.(S represents an element g o.f G. Fo.r fixed h in G,
the left translation g ~ hg maps S o.nto. itself. Let XOO be the space o.f all Coo
functio.ns f(g) o.n S, o.r the Hilbert space L 2(S). Fo.r any h in G, the mapping
p(h): f(g) ~ f(h-Ig)
is a linear transformation in Xoo, and the association h -+ p(h) is a representa-

tion of G. Similarly, the association with h of the mappingf(g) -+ f(gh) (note
that h itself appears here, rather than its inverse) is also a representation of G.
Representations of this type, the so-called left and right regular representa-
tions, are discussed in the next chapter.
EXERCISE
Show that the set of left translations is effective and transitive for the group mani-
fold.
20.9 Representations of the Rotation Group 80(3)
The methods outlined in Section 20.6 and 20.7 are here applied to the
rotation group. Our approach contrasts with the one often taken, in that we
assume nothing in advance about the spherical harmonic functions, Yi(O, <p),
but derive those functions and their properties from group theory.
Let gO) = gw x , Wy, w. be the matrix of a rotation through an angle liroll about
an axis in the direction of roo That is, goo = R(ro), where R() is defined by
(19.6-1). Let XOO be the space of all infinitely differentiable functions f(x)
in [R3. For each g = gO), an operator peg) on XOO is defined, according to
(20.6-1), by the equation
(p(g)f)(x) = f(g-IX). (20.9-1)
These operators peg) constitute a representation of SO(3).
The infinitesimal operators of this representation are obtained as follows:
Since the functions in XOO are differentiable, the operator
1
~ [p(gw, 0, 0) - p(go, 0, 0)]
on XOO has a limit, called L I, as w -+ O. [Recall that p(go, 0, 0) is the identity

operator f -+ f on Xoo.] Any finite-dimensional subspace X I of XOO that is
invariant under all peg) is also invariant under (i.e., is transformed into itself
by) L l . If g = gw,o,o andf(x) = f(x, y, z), then
f(g-IX) = f(x, y cos w + Z sin w, -y sin w + Z cos w),
and it follows that
= -d
Ll def d p(gw,oo) I a - y-;-.
= Z-;- a (20.9-2)
W w=O uy uZ
Similarly,
L2 def
=
d p(go,w,o) I
-d =
a-
x ;:)- a
Z -;-,
w w=O uZ uX
(20.9-3)
d p(go,o,w) I
dcl
L 3 =-d a a
=Y:l-x-;-.
w w=O uX uy
Representations of the Rotation Group SO(3) 51
Except for a factor ih, these are the quantum-mechanical operators for the
components of angular momentum-see Schiff 1955, Chapter IV. They obey
the commutation rules
(i j k = 1 2 3, 2 3 1, 3 1 2). (20.9-4)
Note. The infinitesimal operators of any representation of SO(3) obey these

rules, because the infinitesimal group elements themselves obey them:
Namely, if
T = agwl.w2~1 (i = 1,2,3),
I awi w=o
then, according to (19.9-1),
o o -1
o o o
1 o o
hence [7;, 1j] = 1'", where ij k = 123, 231, or 312. Since products are
mapped onto products in the representation g -+ peg), the rules (20.9-4)
follow for any representation.
Now let S be the unit sphere X Z + yZ + zZ = 1 (a homogeneous space

for the rotation group), and let Xoo(S) be the space of all Coo functions on S.
Starting with the operators L i , the invariant subspaces can be found by
the so-called raising and lowering operators introduced by Dirac (see his
Quantum Mechanics, 1958, Chapter 6) for the quantization of the harmonic
oscillator and of angular momentum. (See also W. Miller 1973.) In polar
coordinates
. a e a
Ll = sm <p ae + cot cos <p a<p ,
a e. a
L z = - cos <p ae + cot sm <p a<p' (20.9-5)
a
L3 = - a<p
The operators L , defined as L 1 iLz, are given by
(20.9-6)
and it is noted that
(20.9-7)
Suppose that f(8, cp) is a function in XOO(S), not identically zero. We wish
to find the minimal invariant subspace Xl containing f(8, cp) and to choose
f(8, cp) so as to make that subspace as small as possible, in a sense. Iff(8, cp) =
L gm(8)e im q>, and if, for some m*, gm*(8) is not =0, then, by the argument of
Section 20.5, the subspace contains all multiples of the single term gm*(8)e im*q>.
The operators L + and L -, when applied to a function of the formf(8)e im q>,
give functions of the form f1(8)e i(m + llq> andfi8)e i(m-1 l q> (for this reason, L +
and L - are called raising and lowering operators), and these functions must
be in Xl, since Xl is invariant; hence, Xl contains functions of the form
(20.9-8)
for m = m* + 1, m* + 2, etc., and for m = m* - 1, m* - 2, ... , etc.
It will appear that the functions gm(8) can be so chosen as to make Xl
finite-dimensional; to achieve this, L +l/1m must be zero for some m, say,
m = l, and L -1/1 m must be zero for some m ~ l, say m = 1'; it will be seen
below that l' = -l. From the first ofthese conditions, g;(8) - 1cot 8g,(8) = 0,
according to the formula (20.9-6) for L +. The solution of this differential
equation is g,(8) = const. (sin 8)', and since the functions in XOO have no
singularities on the unit sphere, it follows that 1 ~ 0; hence,
(20.9-9)
where C is a constant to be determined later. Starting with this function, a
sequence of functions 1/1,-1, 1/1,-2, ... , is obtained by repeated use of the
operator L -, which transforms a function of the form g(8)e im q> into one of the
form h( 8)e i (m - 1 lq>. All these functions are in X 1. We shall now show first that
no new functions are obtained from these by the raising operator, i.e., that
L +l/1m-1 is the same function as I/1m, except for normalization, and second that
L -1/1 -I = 0, so that the sequence terminates at m = -1. We use induction
on decreasing m, starting with m = l: Assume that, for some m, L + 1/1m ex 1/1 m + 1,
i.e., that L - L +l/1m ex I/1m, and note that this last is in any case true for m = l,
because L + 1/1, = O. According to (20.9-7),
(20.9-10)
and it follows that L + L -l/1m is also exl/1m; hence L +l/1m-1 ex I/1m, and the
induction follows.
We now determine the functions I/1m more explicitly. Let the proportionali-
ties referred to be written as
(20.9-11)
Since each I/1m contains an arbitrary factor, these equations determine only
the product am 13m , by the equation L - L +l/1m = -amf3ml/1m. Hence, we can
take 13m = am, for all m. It then follows from (20.9-10) that
for all m < l,

Tesseral Harmonics; Legendre Functions 53
and this equation holds also for m = I, if IXI is set = 0.1t follows by an induction
on decreasing m that
IX~ = (I + m + 1)(1 - m) (m = 1,1- 1, ... , -I);
hence we can choose
IXm = J(I + m + 1)(1 - m) (m = I, I - 1, ...), (20.9-12)

where the positive square root is understood. In particular, IX -1-1 = 0;
hence L -l/l -I = 0, as stated.
The conclusion is that equations (20.9-9, 11, 12) determine all the functions
l/lm (-I S m S l) up to the constant C. These functions span a (21 + 1)-
dimensional subspace X 21 + 1 of XCO(S) that is invariant under L1> L 2 , L3 and
also, as will be seen at the end of Section 20.14, under the transformations
p(g) for all g in SO(3); there is one such subspace for each I = 0, 1,2, .... The
transformations peg), when restricted to the subspace X 21 + 1, are called pl(g);
they constitute a finite-dimensional irreducible representation of G, and it
will be shown in Section 21.13 that these are the only irreducible representa-
tions, up to equivalence.
20.10 Tesseral Harmonics; Legendre Functions
It will be shown in this section that the properties of the spherical harmonics
follow from the representation theory of the rotation group, and that the
tesseral harmonics form a basis for the representation of SO(3).
If the functions l/lm((}, cp) of the preceding section are taken as a basis in
X 21 + 1, then the transformations peg), when restricted to x 2l+ \ are given by
(21 + 1) x (21 + 1) matrices. Before these matrices can be computed, the
functions l/lm must be discussed further; they will be denoted henceforth
by Yi(}, cp), to acknowledge the dependence on l. They are called tesseral
(surface) harmonics. A surface harmonic is a functionf(}, cp) such that rPf(}, cp)
satisfies Laplace's equation in x, y, Z, for some integer p, and it will be seen
that rIYi(}, cp) satisfies Laplace's equation. A tessera (which comes through
Latin from a Greek word meaning "four-cornered") is a curvilinear rec-
tangle such as the ones into which the sphere is divided by the zeros or nodal
lines ofRe Yi orIm Yi, which occur on certain circles oflattitude () = const.
and certain meridians cp = const.
An inner product is defined in the space XCO(S) of functions on the unit
sphere S, as follows:
(f1, f2) = f" ff1(}' cp)fzC(}, cp)sin () d(} dcp. (20.10-1)
The completion of XCO(S) with respect to the norm II f I = (f,j)1/2 is the

Hilbert space L 2(S). The operators p(g), g E SO(3), are unitary in L 2(S),
because they are defined in all L 2(S) and are invertible, and [since the
integral (20.10-1) is invariant under rotations] because
(20.10-2)
for all il and i2' It will be shown that the functions Y?, are orthogonal with
respect to the inner product (20.10-1). If the constant C in (20.9-9) is suitably
chosen (it can depend on I), they are also normalized. It will be proved that
they form a complete orthonormal set of functions on the sphere.
The ({J integration alone shows immediately that Yl:' and YG' are or-
thogonal , if m 1 -r-
-J.. m
2, because ym,I, ym1
1, contains a factor ei(m1-m,)q> . It is
evident from (20.9-2, 3) that the operators Li are antisymmetric, i.e., (Li i, g) =
-(f, Lig), and from this it follows that L - L + is symmetric. Furthermore,
from (20.9-11),
L- L + Y?, = _(a?,)2y?" (20.10-3)
where, with a slight improvement of notation,
(a?,)2 = (l + m + 1)(1 - m), (20.10-4)
according to (20.9-12). Therefore, the equation
(L - L + yr:, Y G) = (Yr:, L - L + Y G)
is equivalent to
(a!';)2(y!';, Y G) = (aG)2(y!';, Y G);
hence, since a!'; i= aG for 11 i= 12, it is seen that the Y?, are orthogonal.
We now show how to choose the constant C in (20.9-9) so as to normalize
the functions Y?,. The adjoint of the operator L + is - L - ; hence
(L + Y?" Y?,+ 1) = (Y?" _ L - Y?,+ 1).
It follows from (20.9-11), since Pm = am = a?" that
( - ia?, Y?, + 1, Y?, + 1) = (Y?" ia?, Y?'),
from which it is seen that I Y?, 112 is independent of m, for given 1. From the
equation (20.9-9) for l/II = Yl,
I YW = 2n Iq2 f sin 21 + 10 dO
_ 2 24 .. (21)
- 4n IC I 1. 3 ... (21 + 1)
2 (21l!)2
(20.10-5)
= 4n Iq (21 + 1)! .
Therefore, if the constant C is chosen as
(-1Y (21+1)!
C = CI = 2lz! 4n (20.10-6)
then all the functions Y?, are normalized.

Associated Legendre Functions 55
With Yl given by (20.9-9) and (20.10-6), and the other Yi given in terms of
Yl by the recurrence relation (20.9-11), which says that L - Yi+ 1 = -io:iYi,
we define new functions Pi(w), called the associated Legendre junctions,
for -1 S w S 1 by the equation
Y ml (e, m) = (_l)m 21 + 1 (l - m)! pm( e) imq> (-1 S m S 1).

't' 4;-(l+m)! 1 cos e
(20.10-7)
Notes. (1) The factors ( -lY in (20.10-6) and ( -lr

in (20.10-7) are arbitrary,
but conventional. (2) Historically, Pi(w) was first defined by equation
(20.11-6) below, and Yi(e, <p) was then defined by (20.10-7). (3) The symbol
Pie w) is used by different authors to denote slightly different functions. The
usage here agrees with that of Talman 1968 and others, for all m (-1 s m s 1),
and agrees with the original definition of Ferrers (see Whittaker and Watson
1927), for m ;;:::: O.
20.11 Associated Legendre Functions

The properties of the functions Pi(w) are derived in this section from con-
siderations of group representations.
Starting from the recurrence relations for the Yi and the formula (20.9-6)
for the operators L , a short calculation gives recurrence relations for the
Pi, which are
~ Pi'(w) + J1n: w2 Pi(w) = Pi+ 1(w) (-1 S m < 1), (20.11-1)
~Pi'(w) - mw Pi(w) = -(1 + m)(I- m + 1)Pi- 1(w)

J1 - w2
(-1 < m s 1), (20.11-2)
where the prime indicates differentiation with respect to w. If the first of these
equations is differentiated once more, and Pi + 1 and Pi + l ' are eliminated,
using the second equation, the result is the associated Legendre differential
equation
(1 - w 2 )Pi" - 2wPi' + [12 + 1- 1 :2W 2 JPi = 0, (20.11-3)
for Pi(w). The first recurrence relation (20.11-1) can be written as
(1 - w 2 )-(m+l)/2Pi+ 1 (w) = d~ [(1 - w 2)-m/2Pi(w)],
from which it is deduced, by an evident induction, that
(1 - w 2)-m/2Pi(w) = (d~)l+m[(l - W 2Y/2P 1- 1(W)]. (20.11-4)

It will be shown below that
P-I(w) = (-lY (1 _ W 2 )1/2 (20.11-5)

I 211! '
from which the so-called Rodrigues formula follows:
Pi(w) = 2;l! (1 - w2)m/2(d~r+m(W2 - 1Y (m = -I, -I + 1, ... , I).
(20.11-6)
[For the present purpose, it would have been more reasonable to fix y l- 1
initially, rather than YL and then determine the other Yi by means of the
raising operator L +, rather than the lowering operator L -. However, the
general relation between Yi and y 1 m , which is now needed, has independent
interest.] Complex conjugation interchanges L + and L -; hence, the con-
jugates of equations (20.9-11) are
L -1/1 m = irxm 1/1 m + 1 , L + 1/1 m + 1 = irxm 1/1 m'
from which it is seen that the quantities ( -lrl/1m satisfy the same equations
as the quantities 1/1 -m; hence
y 1m = k( -lrYi,
where k is a constant, which will soon be seen to be = 1. Since C in (20.10-6)
is real, equations (20.9-9) and (20.10-7) show that pl(w) is real; then (20.11-1)
shows that all the Pi(w) are real; hence, by (20.10-7), Y? is real. The above
equation, with m = 0 then shows that k = 1. Therefore,
(20.11-7)
Since Yl = C(eiq> sin ey, with C given by (20.10-6), the above equation gives
y l- 1 explicitly, from which equation (20.11-5) for P I- I follows from (20.lQ-7).
(Some authors define y l- m to be the complex conjugate of Yr, after defining
the latter for m 2: O. The procedure followed here has some advantages; for
example, the matrices P~'m of the irreducible representations of the rotation
group, given below, are symmetric.)
Clearly Pi(w) is a polynomial, for even m. p?(w), usually denoted by Plw),
is the Legendre polynomial of degree I.
EXERCISES
n
1. Show that SA (1 - wZ)' dw = [21/(21 + 1)] (1 - WZ)'-l dw, and use this
result to show, by an obvious induction, that the integral in (20.10-5) has been correctly
evaluated.
2. Express the operators L in terms of the variables wand cp, where w = cos e,
and derive the recurrence relationships (20.10-8,9) from (20.9-11).
3. For the special case m = 0, verify that the Rodrigues formula (20.11-6) gives a
solution of Legendre's differential equation, which is (20.11-3) with m = O. [The
solution is PiCw).] You are welcome to do the same for m #- 0; it is just more work.
Matrices of the Irreducible Representations of SO(3); the Euler Angles 57
4. Since Pi' and P1- m satisfy the same equation (20.11-3) (this equation is un-
altered by replacing m by - m), which can have at most one solution regular at w = 1,
they must be proportional. Find the proportionality constant. Further warning about
notation: some authors define p l- m to be = Pi'.
5. Show that, as an alternative to (20.11-6), the equation
pm(w) = (_l)m _ _ _. _ 1 (1 _ w2)-m/2 _

(/+m)' (d)l-m(w 2 _ 1)1 (20.11-8)
1 (t-m)!21l! dw
holds, for -1 :S m :S t.
20.12 Matrices of the Irreducible Representations

of 80(3); the Euler Angles
For given I, the functions yr
(m = I, I - 1, ... , -1) are taken as basis vectors
in the space X 21 + 1 of the (21 + i)-dimensional representation of SO(3)
found in the preceding sections. For any functionf = fee, cp), p(g)f is the
function obtained by carrying the values off(e, cp) around the sphere by the
rotation g. Hence, the matrix pl(g) ofthe transformation peg), when restricted
to X 21 + 1, has components P~'m given by
I
(p(g)Yr)(e, cp) = L p~'m(g)Yne, cp) (m = I, I - 1, ... , -I).
m'=-l
(20.12-1)
It is convenient to express the rotation by its Euler angles a, /3, y, and to
write pea, /3, y) instead of peg). Then, g is the result of the following rotations
in succession:
1. A rotation through the angle y about the z axis,
2. A rotation through the angle /3 about the x axis,
3. A rotation through the angle a about the z axis.
(See Exercises 3, 4, and 5 in Section 21.5.) The matrix pi is decomposed
accordingly as
The first and third factors are diagonal matrices; the transformation pea, 0, 0)
merely replaces cp in a function by cp - a and hence multiplies yr
by e - ilZm;
that is,
P~'m(a, 0, 0) = e- ilZm 'c5 m'm'
Therefore, P~'m(rx, /3, y) can be written in the form
P~'m(rx, /3, y) = e-ilZm'p~'m(cos /3)e- iym . (20.12-2)
The functions P~'m(w) are closely related to the Jacobi polynomials; their
properties are discussed at length in Gel'fand, Minlos, and Shapiro 1963
and in Vilenkin 1968, to which the reader is referred for details. (The definition
of P~'m given below agrees with that in Gel'fand et al. and gives the complex
conjugate of the function defined by Vilenkin.) The P~'m are defined by the
equation
P~'m(w) = C(1 + w)-<m+m')/2(l - wym-m')/2(d~ y-m'

x [(w - l)z-m(w + 1y+m], (20.12-3)
where
C = im'-m2- Z( + m')f
(I
)1/2 (20.12-4)
(I - m)!(l + m)!(l- m')!
(It would perhaps be more logical to incorporate the factor im'-m explicitly
in P~'m' rather than in P~'m' which would then be a real function, for
-1 :-:; w :-:; 1, but it is not customary to do so.)
For m = 0 (and for m' = 0), these functions are proportional to the
associated Legendre functions. Comparison of (20.12-3) with (20.11-8)
shows that
P~'O(w) = (_i)m' (1- m')! pm'(w) (20.12-5)

(I + m')! Z
[in particular, P~o(w) = Pz(w)]; hence,
z
Pm'O(rx, /3, 0) = J 4n
21 + 1 Y m'(
z /3, rx - 2n) . (20.12-6)
[This is the same as p~'o(!X, /3, y), because P~'o is independent ofy.]
EXERCISES
1. Verify (20.12-1) for 9 = rotation about the z axis.

2. Verify (20.12-1) for 9 = rotation by n about the x axis.
3. For I = 1, show that the P~'m are given in matrix format by
l+w
--
2
-iJl ~w 2 w-l
2
(P~'m) = .p
-I --
2
w -iJ] ~ w
2
,
w-l
2
-I .p --
2
l+w
--
2
where the rows are numbered by m' = - 1, 0, 1 and the columns by m = -1,0, 1. Note
that this matrix is unitary.
4. Identify the left member of (20.12-1) with Y;"W, q/), multiply the equation
through by r, call z' = r cos e', x' iy' = r sin e'e;cp', and similarly for x, y, z. Again,
take I = 1. Then, using the result of Exercise 3, show that, for the case 9 = rotation
The Addition Theorem for the Tesseral Harmonics 59
through the angle fl about the x axis (i.e., rx = y = 0), the transformation (20.12-1)
reduces to
X' = x,
y' = y cos fl + z sin fl,
z' = - y sin fl + z cos fl.
5. Show that P~'m = P~m"
6. Show that the appearance of rx - n12, rather than rx itself, in (20.12-6) could be
avoided if the second step in the definition of the Euler angles were taken to be a rotation
through the angle fl about the y axis rather than the x axis.
20.13 The Addition Theorem for the Tesseral

Harmonics
In equation (20.12-1), which tells how the functions YT, for given 1, transform
among themselves under a rotation g, we set m = 0, and we take 9 to be the
rotation with Euler angles a, {3, 0:
I
(p(a, {3, O)YP)(), cp) = L p~'o(a, {3, O)YT'(), cp). (20.13-1)

m'=-l
Now the operator pea, {3, 0) carries the values of a function around the sphere
by the rotation g, which takes the polar axis into the direction with coordinates
{3, a - n12. Therefore, the left member above is equal to
YP() 12 , 0),
where ()12 is the angle between the direction {3, a - nl2 and the direction (), cP
[recall that Yp(), cp) does not depend on cp]. We rename the first of these
directions () 1, cp 1 and the second () 2, cp 2, and we use the trigonometrical
identity
cos ()12 = cos ()l cos ()2 + sin ()l sin ()2 COS(CPI - C(2)' (20.13-2)
We also substitute (20.12-6) into (20.13-1); the result is
YP()12' -) = J~ 21 1 mt-l YT() I, CPI) YT()2, C(2), (20.13-3)
which is the required addition theorem. It can also be written as

PI (cos ()12) = Plcos ()1)P1(cos ()2)
+2 (l - m)! PT(cos ()1)PT(cos ( 2)cos m(CPI - C(2)'

m=dl + m)!
(20.13-4)
It is noted in passing that the special case 1 = 1 gives the trigonometric
formula (20.13-2) back again, because PI(w) = w, PHw) = J1 - w 2
20.14 Completeness of the Tesseral Harmonics

It will be shown that the tesseral harmonics form a complete set for the
expansion of functions defined on the sphere. Since (20.11-3) is a differential
equation of Sturm-Liouville type in the interval (-1, 1) (with singular
endpoints of the limit-point type at each end), the completeness can be
established by the methods of Section 10.6 in Volume I. Another approach is
followed here, based on potential theory.
Rodrigues' formula (20.11-6) shows that the function P?,(w) contains only
even powers of w for I + m even, and only odd powers for I + m odd, together
J
with a factor 1 - w2 , for m odd. It follows from (20.1-7), by repeated use
of the equations
r cos = z, e
r sin eeiq> = x + iy,
r2 = x2 + y2 + Z2,
that r'Y?,(e, cp) is a homogeneous polynomial of degree I in x, y, z; i.e., each
term is of the form: constant times xiyizk, where i + j + k = l.
It will now be shown that these polynomials satisfy Laplace's equation.
The Laplacian, when written in polar coordinates, can be expressed in terms
of the operators L i , by use of (20.9-5, 6), as follows:
1 0 2 0 1
-o.SIll e-oe0 + - -1 - -0-
2
2
V =- - r - +
r2 or or r2 sin e oe r2 sin e Ocp2
2
=
1 [0 0+ L -L + + L3 -
2
r2 or r or
2
1L3 .
] (20.14-1)
Now Y?, is an eigenfunction of L - L + and of L 3, with eigenvalues - (a7')2 and

- im, respectively. Therefore,
1
V 2r' Y?,(e, cp) =2 [1(1 + 1) - (I + m + 1)(1 - m) - m 2 - m]r'y?,(e, cp) = 0,
r
where (20.9-12) has been used.
The general homogeneous polynomial p(x, y, z) of degree I has
!(I + 1)(1 + 2) terms xiyizk, because ifi = 0 there are 1+ 1 possible values of
j, if i = 1 there are I possible values, etc.; in each case k is then fixed as I - i - j;
hence the number of terms is 1 + 2 + ... + (I + 1) = !(I + 1)(1 + 2). Now
suppose that p(x, y, z) satisfies Laplace's equation. Since V 2 p is a homogeneous
polynomial of degree I - 2, its vanishing imposes!(I - 1)1 conditions on the
coefficients of p, and it is easy to see that these conditions are independent.
Hence, the harmonic polynomials of degree I form a space of
(I + 1)(1 + 2) _ (I - 1)1 = 21 1
22+
dimensions. This space is precisely spanned by the polynomials
r'Y?, (m = I, I - 1, ... , -I),
Completeness of the Tesseral Harmonics 61
for there are 21 + 1 of them, and they are obviously independent because of
the orthogonality of the Y7'. It is concluded that any harmonic polynomial can
be expressed in terms of the functions r1y7'(0, ({J).
According to the theory of the Dirichlet problem in potential theory, if
f(O, ((J) is any continuous function on the unit sphere, there is a function
tjJ(x, y, z) that satisfies V2tjJ = 0, for x 2 + y2 + Z2 < 1, and is continuous,
for x 2 + y2 + Z2 ::; 1, and takes the values f(O, ({J) on the sphere x 2 +
y2 + Z2 = 1. (In fact, tjJ can be expressed in terms offby the Poisson integral
formula.) tjJ is analytic in the ball and can be expressed as a power series in
x, y, and z. The terms of this expansion, of a given degree I constitute a
harmonic polynomial of degree I, and hence can be expressed in terms of the
functions r1y7'(0, ({J);"that is,
00 1
tjJ(x, y, z) = I
1=0 m=-l
L A7'r1y7'(0, ({J).
It can be shown that, for a continuous boundary functionf, the solution tjJ
of the Dirichlet problem converges to f, uniformly in angle, as r -+ 1. Hence,
it converges in L 2 ; therefore,
f(O, ({J) = L A7'Y7'(O, ((J), (20.14-2)

(l,m)
in the sense of mean convergence. Since the continuous functions are dense
in L 2(S), it follows that the tesseral harmonics form a complete orthonormal
set of functions on the sphere. It follows also that any distribution f(O, ((J)
in L 2(S) can be expanded as (20.14-2), where the coefficients are given by
A7' = f"f Ir(O, ({J)f(O, ({J)sin dO d({J,
and where the series converges tof(O, ((J) in the mean.

A gap in the discussion of the representations of SO(3) in Section 20,9
can now be filled in. It is recalled that XOO was defined as the space of all COO
functions on the sphere, and X 21 + 1 was the subspace spanned by
Y7' (m = I, I - 1, ... , -I).

It was proved that X 21 + 1 is invariant under (is transformed into itself by) the
infinitesimal operators L 1 , L 2 , L3 of the representation p, and it can now be

proved that X2l+ 1 is invariant also under the transformations peg), g E SO(3):
Namely, peg) Y7' is a function of and ({J obtained by rotating the sphere by the
rotation g and carrying the values of Y7'(O, ({J) along; hence, rlp(g)Y7' is also
a homogeneous harmonic polynomial of degree I in x, y, z; hence, it can be
expressed linearly in the polynomials r1y;"'(0, ((J), m' = I, 1- 1, ... , -I.
Therefore, X 21 + 1 is invariant.
The completeness of the set of tesseral harmonics shows that the space
L 2(S2), where S2 is the unit sphere in [R3, is the direct sum (with respect to the
L2 norm) of the spaces X 21 + 1 (l = 0, 1,2, ... ).
CHAPTER 21
Group Representations II:

General; Rigid Motions; Bessel Functions
Unitary representation; reduction; decomposition; direct sum; complete

reduction; Schur's Lemma; compact and noncom pact groups; invariant
integration; Haar measure; right and left translations; invariant integration in
SU(2); area of the n-sphere; regular representations; invariant integration in
50(3); completeness theorems of Peter and Weyl and Vilenkin; wave
functions of the symmetric top; rigid motion groups; Bessel functions;
recurrence relations, differential equations, and generating equations;
expansion of plane wave; characters; completeness of representations of
50(3).
Prerequisites: Chapters 18-20.
The irreducible representations of SO(2) found in Section 20.5 were one-

dimensional; those of SO(3) found in Section 20.9 were multidimensional,
with dimension 21 + 1, 1 = 0, 1, 2, ... ; those of the rigid motion groups, to
be discussed below, are infinite-dimensional. These differences reflect
various group properties, as will be shown.
21.1 Equivalence; Unitary Representations
The first step in classifying representations is to decide when two representa-

tions are equivalent, so that only one of them need be described. Two
representations p(g) and p'(g) on a (finite-dimensional) space X are equivalent
if the first can be transformed into the second by a suitable change of the
coordinates in X, i.e., if there is a matrix A such that p(g) = A-1p'(g)A,
for all g. More generally, if p and p' are representations on different spaces
X and X' of the same dimension, and if there is a matrix A such that the
association x -+ x' = Ax is a one-to-one mapping of X onto X', and if p(g)
and A -1 p'(g)A are the same matrix, for every g, then p and p' are equivalent
representations.
62
The Reduction of Representations 63
In the infinite-dimensional case, as noted in Section 2004, peg) is a bounded

linear transformation in a Banach or Hilbert space for each g. If p and pi are
representations in spaces X and X' of the same kind, say both separable
Hilbert spaces, and if there is a bounded linear transformation A from X
onto X' with an inverse A-I, and if peg) and A -1 p'(g)A are the same operator
for every g, then p and pi are equivalent representations.
This equivalence relation is reflexive, symmetric, and transitive, hence
divides the set of all representations of a given group into equivalence classes.
It is natural to try to choose from each equivalence class a representation
with convenient properties. In most cases it is possible to choose a unitary
representation, i.e., one in which each peg) is a unitary matrix or unitary
operator, for all gin G; this is advantageous because of the theorem in the
next section.
It should be noted that the equivalence of two representations p and pi
is not merely a matter of the isomorphism of the two matrix groups {peg)}
and {p'(g)}: The matrices have to be of {he same order in the two repre-
sentations, and the representations have to be so related that peg) =
A -1 p'(g)A, for all g, for some fixed A. The representations pl(l = 0, 1,2, ... )
of SO(3) found in Section 20.3 are all inequivalent (they have different
dimensions), but the resulting matrix groups {pl(g)} are all isomorphic, for
I =F 0: In fact, they are all isomorphic to SO(3). The following example shows
that two representations may have the same dimension but be inequivalent:
Let G be the torus group T2 , consisting of 2 x 2 diagonal unitary matrices,
i.e., matrices of the form
lal = 1, 1131 = 1.
The two representations on IC given by

peg): z ~ az, p'(g): z ~ 13z
are inequivalent; no transformation of the z plane can take peg) into p'(g)
for all g. For g = (6 _~), peg) is the identity mapping in the plane, and
p'(g) is not.
21.2 The Reduction of Representations
The next step in the classification is to decompose a given representation,

as far as possible, into irreducible components. If p is a representation on a
finite-dimensional space X, and if a subspace Xl is invariant under all p(g)-
i.e., if x E Xl implies p(g)x E Xl> for all g-then p is called reducible, as in the
preceding chapter (otherwise, p is irreducible). If another invariant subspace
X 2 can be found such that X = X 1 EEl X 2 (by which is meant that X 1 and X 2
have no vector in common except x = 0 and that any x in X can be written as
Xl + x 2 , where Xl EX 1 and X 2 E X 2), then p is called fully reducible or
decomposable. In this case, if e1 , ... , en is a basis in X such that e1 , ... , em is
64 Group Representations II: General; Rigid Motions; Bessel Functions
a basis in Xl while em + l , ... , en is a basis in X 2 , then, with respect to this

basis, all the matrices peg) have the form
peg) =
B (0)
(n - m)
(21.2-1)
(0) x
(n - m)
i.e., all matrix elements that connect the two subspaces are zero. (If P is
reducible but not decomposable, then it is possible to choose the basis so
as to make all matrix elements zero in the lower left rectangular block
shown, but not also in the upper right block.) If, for each g, PI(g) and pig)
denote the m x m and (n - m) x (n - m) matrices shown, respectively,
then each of the mappings g ~ PI(g) and g ~ pig) is a representation of G,
and the representation P is their direct sum; in symbols, P = PI P2' +
If X is an infinite-dimensional Banach or Hilbert space, X I is understood
as a closed linear manifold in X; there is no loss of generality here, because
the operators peg) are all bounded, hence the closure of an invariant linear
manifold is invariant. Again, if X = X I E8 X 2, and if X I and X 2 are invariant
under all peg), then we write
P = PI + P2'
where PI and P2 are the restrictions of the representation p to X I and X 2,
respectively. (One of PI' P2 may be finite-dimensional.)
It may be that X I or X 2 contains further invariant subspaces such that
PI and P2 (or both) can be further decomposed, and so on. Then, with respect
to a suitable basis in X, the matrices peg) contain a number of square blocks
straddling the main diagonal, and all elements outside those blocks are
zero. Each square block gives a representation of G, and P is the direct sum
of those representations. When this process has been carried as far as possible,
it may turn out that the resulting representations into which P has been
decomposed are all irreducible. In this case, P is called completely reducible.
Then one can find the structure of all representations of G by finding all
minimal invariant subspaces of a sufficiently large initial space X, as was
done in the preceding chapter for SO(2) and SO(3).
Theorem. Any finite-dimensional unitary representation of a group G

is completely reducible, i.e., either is irreducible or can be written as a
direct sum of irreducible representations.
PROOF. The result follows from the fact that if p is unitary and Xl is invariant, then,
as is easily seen, X t is also invariant.
Schur's Lemma and Its Corollaries 65
21.3 Schur's Lemma and Its Corollaries

The next theorem depends on the famous lemma proved by 1. Schur in
1905, an apparently trivial but quite powerful bit of elementary linear algebra.
Lemma (Schur). If PI and P2 are irreducible representations of G on

V k and Vi, respectively, and if A is an 1 x k matrix such that
Apl(g) = P2(g)A, for all gin G, (21.3-1)
then either k = 1, and A has an inverse A -1 (hence PI and P2 are equivalent),
or A is the zero matrix. In the first case, A is uniquely determined, up to a
scalar multiplier, by (21.3-1).
PROOF. Let S be the null space of A, consisting of all vectors x in Vk such that
Ax = 0. S is invariant under all P1(g), for if Ax = 0, then Apl(g)X = 0, by (21.3-1).
Since P1 is irreducible, either S is all of V\ in which case A = 0, or S consists of the
zero vector alone. In the latter case, no nonzero vector is orthogonal to all the rows
of A; hence the number 1 of rows is ;::0: k, and the mapping x ~ y = Ax is invertible
because any k x k submatrix in A is nonsingular and can be used to solve the equation
y = Ax for x. Let S' be the image of Vk in Vi under the mapping x --> Ax; then S'
is k-dimensional and is invariant under all pz(g), because if y is Ax for some x, then
pz(g)y is Ax', for x' = P1(g)X. The representation pz is irreducible; hence S' = Vi,
k = 1, and A -1 is the matrix of the inverse mapping. We now prove the uniqueness
of A: Suppose that B is another nonzero matrix such that Bp1(g) = pz(g)B, for all g;
then it must be proved that A = const. B. Clearly,
pz(g)AW 1 = AW 1 pzCg), for all g.
It follows from this equation that if v is an eigenvector of AB- 1 (every matrix has at
least one eigenvector), then pz(g)v is also an eigenvector corresponding to the same
eigenvalue, say A. That is, the one-dimensional subspace containing v is invariant
under pz(g) for all g. Since pz is irreducible, the complete eigenspace corresponding to
A must be all of Vk; that is, AB- 1v = AV, for all v; hence B- 1 = AA- 1 or A = AB,
as required.
Comment. No use has been made of the fact that G is a group. The same
conclusions hold if the sets {Pl(g)} and {P2(g)} are any two irreducible
sets of square matrices. A set {MJ of k x k matrices is irreducible if there is
no nontrivial proper subspace of V k that is invariant under all mappings
x --+ MiX.
The following was established during the proof of Schur's lemma:
Corollary. Any matrix that commutes with all the matrices of an irreducible
representation (or an irreducible set of matrices) is a multiple of the identity.
A further corollary is the following:
Theorem. Every finite-dimensional irreducible representation p of an

Abelian (commutative) group is one-dimensional.
PROOF. Each p(h) commutes with all peg); hence each p(h) is a multiple of the identity
matrix; hence, p would be reducible if it were not one-dimensional.
This explains why only one-dimensional representations of SO(2) were

found in Section 20.5.
21.4 Compact and Noncompact Groups

It was pointed out in Section 19.5 that an n x n matrix (generally complex)
can be represented by a point in a space V of 2n 2 (real) dimensions, and that
the set of all matrices constituting any group like GL(n), SL(n), U(n),O(n), etc.,
is an algebraic surface Y' in V. For the groups mentioned (they are continuous
groups or Lie groups), Y' is always a closed point set, but it may not be
bounded. For O(n), U(n), SO(n), SU(n), the surface Y' is bounded [for
example, equation (19.5-1) shows that, for the group 0(3), no point of Y'
can have a coordinate Rjk larger than 1 in absolute value], whereas, for
GL(n), SL(n), M n , and the Lorentz groups, Y' is unbounded (extends to
infinity in V). In the first case, the group is called compact (Y' is a compact
set in V), and in the second case noncompact. The theory of compact groups
is much simpler than that of the noncompact ones; that shows an interesting
interplay between different parts of mathematics: The compactness of a
group manifold is a matter of geometry, while the simplifications referred
to are mainly algebraic.
The following theorems, stated without proof, show that for compact
groups it suffices to consider unitary representations.
Theorem 1. Let p be a finite-dimensional representation of a compact

group G. Then p is equivalent to a unitary representation; that is, there is a
fixed nonsingular matrix A such that the matrix Ap(g)A -1 is unitary for all
gin G.
This can be restated as follows: If a new inner product (', ')1 is defined
in the representation space X by the equation (~, tt)l ~ (A~, Att), where
(-, -) is the ordinary inner product, given by (x, y) = x*y, then
(p(g )~, peg )tt) 1 = (~, tt) 10 for all g in G, all ~, tt in X; (21.4-1)
in other words, the matrices peg) are unitary with respect to the new inner
product. The same holds generally:
Theorem 2. Let p be a representation of a compact group G on a Hilbert

space f); i.e., each p(g) is an invertible bounded linear operator in f), and
P(glg2) = P(gl)P(g2)' Then there is a new inner product (', ')1 in f) such
that
(p(g)u, P(g)V)l = (u, V)l' for all gin G, all u, v in f); (21.14-2)
Theorem 3. All irreducible unitary representations of a compact group G

are finite-dimensional.
Invariant Integration; Haar Measure 67
By contrast, the Lorentz groups have infinite-dimensional irreducible

unitary representations.
21.5 Invariant Integration; Haar Measure
Let G be a compact group of n x n matrices; the manifold 51' of G is a

compact (closed, bounded) surface (of some dimension m ::;; 2n 2 ) in a space
V of 2n 2 (real) dimensions (see Section 19.5). It can be proved that there is a
positive continuous function w(g) defined on 51'~ here, the symbol 9 is
used both for a group element and for the corresponding point of 51' -with
the remarkable property that if f is any continuous function on 51' and h
is any fixed group element, then
J/(h9)W(9)dd(9) = L/(g)W(9)dd(9)
(21.5-1)
for all h in G, all continuous f on 51',
where dd(g) is the m-dimensional volume element, or "element of area" on
g.
If G is noncompact, so that the surface 51' extends to infinity in V, then
the same is true~i.e., there is a weight function w(g) on 51' such that (21.5-1)
holds~provided that f is a function such that the integrals converge.
The mapping 9 ~ hg of G onto itself, for fixed h, is called a left translation
in G. The equation above shows that the integral of a continuous function on
51' with respect to the weight function w is invariant under all left translations
in the group. The integral in (21.5-1) is called a left-invariant integral. A
proof that such a function w exists can be found in Wigner's book Gruppen-
theorie (1931) under the heading "Hurwitzsches Integral." See also Nachbin
1965.
There is a similar invariant integral for right translations. It can be proved
that, in particular, if the group G is compact, then the above integral (with
the same weight function w) is also invariant under right translations and
under inversions; that is
f f(g)w(g)dd(g) = ff(gh)w(g)dd(g) = ff(g-l)W(g)dd(g). (21.5-2)
The proof can be found in Weyl's book, The Theory of Groups and Quantum
Mechanics (1932), Chapter III, Section 12.
EXERCISES
1. Show that any matrix 9 in SU(2) can be written as

where a and b are complex numbers subject only to the restriction 1a 12 + 1b 12 = 1.

Hence, if a = Xl + iX2 and b = X3 + ix 4 , the manifold S of SU(2) is homeomorphic
to the 3-sphere S3, i.e., the unit sphere xi + xi + x~ + xi = 1 in [R4.
2. Show that a left translation g --> hg in SU(2), where h is a fixed group element,
induces a rotation of S3 about its center, and hence that if dd(g) is the element of
3-dimensional area on S3, the weight function w(g) in (20.7-3) can be taken to be constant,
and the integral S f(g)dd(g) is invariant under left translations (also right translations)
in the group.
3. Let a and b in Exercise 1 be written as
b = I.(. 2. exp (.1IJ( --2-

sm {3) y) ,
where
0::; IJ( < 2n,
o ::; {3 ::; n,
-2n::; y < 2n;
then g is written as g(lJ(, {3, y), and the variables IJ(, {3, yare called the Euler angles of g.
[Under the homomorphism of SU(2) onto SO(3) given in Section 19.7, they become
the Euler angles of the rotation R(g), with y restricted to the range 0 ::; y < 2n-note
that replacing y by y + 2n replaces g by -g and leaves R(g) unaltered.] Show that if the
Euler angles are taken as intrinsic c~ordinates in the group SU(2), then the element of
area on S3 is given by
sin {3
dd(g) = -8- dlJ( d{3 dy.
4. Show that
g( IJ(, {3, y) = g( IJ(, 0, O)g(O, {3, 0) g(O, 0, y). (21.5-3)
5. Show that, under the homomorphism of SU(2) onto SO(3) given in Section
19.7, if R(g(lJ(, {3, y)) is called R(IJ(, {3, y), then R(IJ(, 0, 0) is =R(O, 0, IJ() and is a rotation
through IJ( about the z axis, while R(O, {3, 0) is a rotation through {3 about the x axis. Give
the geometrical interpretation of the result of Exercise 4 as the law of composition of an
arbitrary rotation in terms of successive rotations about the z, x, and z axes, respectively.
6. Derive the formula for the area An of the n-sphere (the unit sphere in En+l)
from the evident equation
using the gamma function, and verify directly that the formula of Exercise 3 is correctly
normalized. Conclude, on the basis of the 2-to-l homomorphism of SU(2) onto SO(3)
that the (3-dimensional) area of the surface that has been identified with the manifold
of SO(3)-see Section 19.5-is equal to n 2 Show that the volume of an n-dimensional
ball of radius R is
2nn/2
v" = ___ Rn.
n['(n/2)
Invariant Integration; Haar Measure 69
The system {Y;"} oftesseral harmonics is not the only orthogonal function
system that comes from the representations of SO(3). The tesseral harmonics
are orthogonal on the unit 2-sphere, which was taken as the homogeneous
space for the representation. However, it was pointed out in Section 20.8
that the group manifold can also be taken as the homogeneous space;
then, a larger class of orthogonal functions appears-they are functions of
the Euler angles a, {3, ')I, which can be taken as intrinsic coordinates in SO(3).
The theorem below deals with such function systems in general.
It is customary to denote the expression w(g)dd(g) that appears in the
left-invariant integration over the group manifold simply by dg, and to write
equation (21.5-1) as
Lf(hg)dg = Lf(g)dg . (21.5-4)
Let Co(G) denote the space of continuous functions of compact support

on the group manifold, or, as one says, on G (if G itself is compact, then this
includes all continuous functions on G). Then fG f(g)dg, f ECo(G), is a
continuous linear functional on Co(G) and is a measure (see Chapter 13).
Hence, one sometimes speaks of a left-invariant measure on G-or H aar
measure, because it was discussed in a paper by A. Haar in 1933.
Let L 2( G) denote the Hilbert space of quadratically integrable distributions
defined on the group manifold, with the inner product
(21.5-5)
where, as above, dg is the left-invariant measure (G not necessarily compact).

The left-regular representation of G is the association with each h in G of the
mapping
p(h):f(g) --+ f(h- 1g) (21.5-6)
of L2(G) onto itself. (The right-invariant integral leads similarly to a right-
regular representation.) Since the inner product is based on the left-invariant
integral, it is seen that
(p(h)f1' p(h)f2) = Lf1(h 19)f2(h-1g)dg
= Lf1(g)fig)d g
= (f1' f2), for all f1' f2 in L 2 and all h in G.

That is, the representation p is unitary.
It can be proved that, if G is a compact group, then all irreducible repre-
sentations can be obtained by the decomposition of the (left- or right-)
regular representation p. That is, if P1 is any irreducible representation, then
there is a subspace Xl of L2(G) such that the restriction of p to Xl is equiva-
lent to Pl.
Often, even if G is noncompact, one can find irreducible representations

by operators as in (21.5-6) on some space of functions defined on the group,
although it may be necessary to go outside the Hilbert space L2(G) to find
the functions. We shall not attempt to discuss the general case but merely
give an example: in Section 21.10, the functions that appear in connection
with irreducible representations of the noncompact group M 2 of rigid
motions in the plane are not square integrable over G.
We consider now the calculation of w(g). Suppose that 81> ... ,8n are
intrinsic coordinates in the manifold of a group G and that we wish to find
the weight function w(9) such that w(9)d8 1 ... d8 n is the element of invariant
measure dg.
We rewrite equation (21.5-4) as
J = Lf(hg')dg' = Lf(g)dg .
We consider the case where the function f(g) is zero except for elements g in
a small neighborhood % of the identity element of the group, and f(g) = 1
for those elements. They occupy a small volume V in the coordinate space
near 9 = 0; hence from the right member of the above equation we have
J ~ Vw(O). The left member comes from group elements hg' in the neighbor-
hood % of the identity, hence from g' in a neighborhood of h-l, having
volume V', so that J ~ V'w(h- 1 ); hence we must determine V' to determine
w(h - 1). If we call g' = h - 1 k, then k varies in %. We denote by 9(g) the
coordinates of any group element g. Hence, if 9 and 9' denote the coordinates
of k and g' = h- 1 k, we have
9 = 9(k), 9' = 9(h- 1 k) = 9'(9).
As 9 ranges through the volume V, 9' ranges through V'; hence V' is given in
terms of the Jacobian as
V' ~ aC8'b ... , 8~) I V,

0(8 1 , , 8n ) 0
where the subscript 0 indicates that the Jacobian is to be taken at 9 = O.

We conclude that
w(9(h - 1 = c [0(8'1, ... , 8~) I ] -1, (21.5-7)

0(8 1 , ... , 8n ) 0
where C is = w(O). Since h was arbitrary, h - 1 is also arbitrary; hence this

equation determines w(9) for all 9. If the group is compact, C can be chosen
J
so as to normalize G dg to 1.
EXERCISE
7. Let G be the rotation group SO(3), let I3 x , l3 y , I3z be the intrinsic coordinates
introduced in Section 19.6, and let Obe the vector with components I3 x , l3 y , I3z Specifically,
let 0 represent the group element k in the above discussion, where 11011 ~ 1, and let 0'
Complete System of Representations of a Compact Group 71
represent h-1k. To first order in small quantities [see (19.6-1)],
We can take h- 1 as a rotation through an angle rx about the x axis, since w(h- 1 ) is
independent of the direction of the rotation-axis, i.e., we can take
Show that the coordinates e~, e~, e~ of the element h-1k are, to first order,
e~ = rx + ex' e~ = rx[e 1 + cos rx - ~J,

y e~ = rx[e z 1 ; .cos rx + ~2YJ,
2 sin rx 2 Sillrx
hence that the Jacobian is
a(e~, e~, e~) I rx 2

(21.5-8)
a(e x , e ez )
y, O~O = 2(1 - cos rx)'
hence that the normalized weight function is given by
1 - cos rx
w(9') = -4---'n2'--rx~2- , rx = 119'11. (21.5-9)
Hint: The angle of rotation and the direction of the axis are given by (19.2-7 and 8) for
a given rotation matrix.
21.6 Complete System of Representations of a

Compact Group
We come now to one of the most important theorems on compact groups.
Theorem. Let pk (k = 1,2, ...) be a complete set of inequivalent irreducible

unitary representations of a compact group G; let dk be the dimension of pk,
and let P~n<g) be the matrix elements of the transformation l(g). Then the
functions
k = 1,2, ... ,
{
1 ::;; m, n ::;; db
form a complete orthonormal system on G with respect to the inner product
based on invariant integration on G.
Remarks. (1) Recall that right- and left-invariant integration are the same
on a compact group. (2) It has been assumed that SG dg = 1. (3) The matrix
elements are assumed to refer to an orthonormal set of vectors, so that the
matrices (P~n(g of the unitary transformations leg) are unitary matrices.
For the proof of the theorem, see Vilenkin 1968. The orthonormality of the
functions is a straightforward matter; their completeness is somewhat
deeper and was proved by F. Peter and H. Weyl in 1927.
For G = SO(3), the theorem says that the functions
I = 0, 1,2, ...
{
-1:::;; m', m:::;; 1
form a complete orthonormal system on the manifold of SO(3) with respect

to the inner product
see (20.12-2) and Exercise 3 in the preceding section.

The expansion offunctions in L 2(G) in (generalized) Fourier series in the
functions described in the theorem is called harmonic analysis on the group.
Harmonic analysis on noncompact groups involves generalized Fourier
integrals; see, for example, Gel'fand, Graev, and Vilenkin 1966 for harmonic
analysis on SL(2, IC) and Warner 1972 for harmonic analysis on semi-
simple Lie groups.
A similar theorem holds, under certain circumstances, for the functions
obtained on other homogeneous spaces, with respect to a suitably chosen
inner product. See Vilenkin 1968, Section 4.5 of Chapter 1. Anexample that
has already been established above is that the functions Yi(e, cp) form a
complete orthonormal set on the unit 2-sphere.
21.7 Homogeneous Spaces as Configuration

Spaces in Physics
Considerations of physics often indicate which choice of homogeneous space

will yield the functions of interest in a given problem, for a given symmetry
group. For the quantum mechanical motion in a central force field, the
coordinates are r, e, cp; hence there are only two angle variables, and the 2-
sphere is the appropriate homogeneous space. This choice leads to the
functions Yi(e, cp), which ,can be taken as giving the angular dependence of
the wave function. For the motion of a rigid body about its center of mass,
on the other hand, there are three angular variables, the Euler angles 0(, /3, y,
and the configuration space may be regarded as identical with the manifold
ofthe group SO(3). It was shown in 1929 by A. Sommerfield that the quantum
mechanical problem of the symmetric top (a rigid body with two equal
moments of inertia), which is of interest for the theory of molecular spectra,
can be solved by use of the Jacobi polynomials. Indeed, the functions (20.12-2)
can be taken as the wave functions of the symmetric top.
Representations of M2 73
21.8 M 2 and Related Groups
A rigid motion in real n-space V" consists of a rotation x ~ Rx [R E SO(n)]

followed by a translation x ~ x + ~, where ~ is a constant vector. (If the same
rotation is preceded by a translation x ~ x + ~1' the result is the same, for
suitably chosen ~1' namely ~1 = R-l~.) Note: The words "rotation" and
"motion" are possibly slightly misleading, because nothing depends on the
time t. A rotation is merely a fixed change of orientation, and a translation is
merely a fixed displacement.
The relations among some of the groups of interest in physics are
SO(2) - - + SO(3) --+!l' p
~ ~ ~
M2 --+ M3 - - + f!J>p,
where an arrow indicates the subgroup relationship and where f!J> p is the
proper Poincare group (consisting of proper Lorentz transformations com-
bined with displacements in space and time). Groups involving spatial
inversion and time reversal are also of interest but would complicate the
diagram. M 2 will be considered in some detail in this chapter.
An element of M 2 is a mapping g = g~,~, 0 of the x, y plane onto itself given
by
g: (x)y ~ (XI)
y'
= (X c~s
sm X
f) -
f)
y sin f) + ~).
+ Y cos f) + 1]
(21.8-1 )
Note. In goup theory generally, and, in particular, in the definition of

"representation," a "linear transformation" is understood to be homo-
geneous; hence, (21.8-1) will be called simply a "mapping." It is easily verified
that a 3-dimensional faithful representation of M 2 is given by the association
COS f) -sin f)
(
g +-+ Si~ f) cos f) (21.8-2)
o
EXERCISE
Verify that under this association the composition 929, of two mappings of the
form (21.8-1) is associated with the product of the corresponding matrices of the form
(21.8-2).
21.9 Representations of M2
To find other representations, let XOO denote the space of all infinitely dif-
ferentiable functions f(x, y) defined for all x and y. Evidently, the plane is a
homogeneous space for M 2 A representation of M2 on XOO is obtained by
transforming each f(x, y) into

(p(g)f)(x, y) = f([x - + [y - '1]sin 8,
~]cos 8
- [x - ~]sin 8 + [y - '1]cos 8), (21.9-1)
according to the rule (20.6-1). Three operators Lb L 2 , and L3 ("infinitesimal
operators") are defined by
(Ld) (x, y) = :~ (p(g)f) (x, Y)I~='1=9=O'

(Ld)(x, y) = :'1 (p(g)f) (x, Y)I~='1=o=o' (21.9-2)
(Ld)(x, y) = :8 (p(g)f) (x, Y)I~='1=9=O

From (21.9-1) it is seen that
a a a
L3=y--x-.
a (21.9-3)
L1 = - ax' L2 = oy' ax oy
If rand q> are polar coordinates in the x, y plane, and if L + and L "- are
defined as L1 + iL2 and L1 - iL2' respectively, it is found that
oq>,
(a
L -_ e icp -+-- i
or - r oq> ,
a) (21.9-4)
L + L - = L - L + = V2 .
21.10 Some Irreducible Representations

The representation (21.9-1) is highly reducible, and we look for an invariant
subspace Xl of XOO that is as small as possible in some sense. Xl must be
transformed into itself by each of L 1 , L 2 , L 3 By Fourier series expansion with
respect to q>, any f E XOO can be expanded in the functions
(21.10-1)
where gm is some Coo function on (0, 00). The reason for the factor i- m will
be apparent soon. Each IjJ mis transformed into a multiple of itself by L 3 ,
and we wish to start with one of the IjJm and find out what is the smallest
collection of additional functions that must be included to get an invariant
subspace. From (21.9-4) it is seen that L +ljJm is of the form ei(m+ l)'Ph(r)
and hence that L - L +ljJm is of the form eim'Pg(r). In the interest of obtaining
the smallest possible invariant subspace Xl' it is now assumed that g(r) is
= gm(r) times a constant, say - rt~, and we investigate whether the gm can
be so chosen that this assumption is valid. Hence, we wish to choose the
gm(r) so that
(21.10-2)
Bessel Functions 75
In this way, a sequence {I/Im}'~) 00 of functions is generated, assuming that

none of the a's vanish; however, none can vanish (unless all do), because L +
and L - commute, from which it follows that a! = a!-l; hence all the a's can
be taken as equal. Then, from (21.9-4) it is seen that
(21.10-3)
for each m. This requires, of course, that gm(r) be proportional to the Bessel
function J m( ar); Bessel functions are discussed in the next section.
Conclusion. Xl is the subspace of XOO consisting ofthe (nonsingular) solutions

of "121/1 + a21/1 = 0: it will be called X a. Each nonzero value of a leads in this
way to an irreducible representation of M 2. Under a transformation x -+ /lX,
where /l is real, "12 -+ /l- 2"12; hence we can assume, without loss of generality,
that [a [ = 1. Furthermore, IX and -a determine the same subspace; hence
the relevant values of a may be taken as e iP , 0 :.:; f3 < 1[, For each such a,
the representation of M 2 on Xa is given by the transformations (21.9-1).
21.11 Bessel Functions
We normalize the system of functions {I/Im}~oo by setting 1/10(0,0) = 1.

Then, for each m, the Bessel function of order m can be defined as Jm(z)
= gm(z/a), so that (21.10-1) becomes
I/Im(X, y) = i-meim"'Jm(ar).
The equations (21.10-2) then take the form
(:z - ~)Jm(Z) = -Jm+1(z),

(21.11-1)
which are the recurrence relations for the Bessel functions. Elimination of
J m + 1 gives
(21.11-2)
which is Bessel's differential equation.

The Bessel functions Jm(z) are completely determined by these equations
and the starting conditions J 0(0) = 1 and J m(O) = 0 for m =1= O. From now on,
we assume familiarity with these functions, and, in particular, with the
integral representation
Jm(z) = ~
2n
In eizsint+imt dt.
-1[
(21.11-3)
21.12 Matrices of the Representations
For some purposes, it is more convenient to have the transformations

represented by (infinite) matrices; the matrix elements can be obtained as
follows: Let rt. be a vector whose (generally complex) components are
rt. cos X and rt. sin X, where rt. is as above and X is a real angle. Then, a solution
of the reduced wave equation V 2 u + rt. 2 u is
(21.12-1)
(as before, x = r cos cp, y = r sin q, and the general solution is
(21.12-2)
where! is an arbitrary function in some class of admissible functions, whose

exact properties are not important here. [In the case where rt. is real, if
fer, cp) is interpreted as a wave function, then rt. cos X and a sin X are the
momentum variables, and lex) is closely related to the momentum repre-
sentation of fer, cp).] If the plane wave exp{ia' x} is subjected to the trans-
formation (21.9-1), it becomes exp{ia' (x - ~)o}, where the subscript f)
indicates that the vectors x and ~ have been rotated (counterclockwise)
through the angle f). If the displacement components are written as ~ =
'cos w, '1 = 'sin W, then it is seen from (21.9-1) that
(p(g)f)(r, cp) = f/arcos(<p-X)I'(X)dX,
where
(21.12-3)
(the prime does not denote differentiation); a change of variable X - f) ...... X
has been made in the integral, without altering the limits, because the inte-
grand has period 2n. Hence, the group element g of M 2, which consists of a
clockwise rotation through f) followed by a translation by;, induces the
transformation (21.12-3) in the space of functions! of period 2n. lex) and
I'(X) are now expanded in the Fourier series L
crneirnx and I c~eirnx, res-
pectively; it is found that
00
c~ = L (n)Pmncn,
-00
where
Pmn = p~~(g) = eimOe i(m-n)(w-rr/2) 21 f" eiassint+i(m-n)t dt

n -rr
(21.12-4)
where (21.11-3) has been used. It is seen that the Bessel functions appear
not only in the characterization of the invariant subspaces of Xoo, but also
Characters 77
in the dependence of the matrix elements Pmn on the parameters ~, 1], 8 or

(, W, 8 in the group M 2 .
Irreducible representations of the group M 3 of rigid motions in Euclidean
space E3 can be obtained in a similar way, taking E3 as the homogeneous

space. Let Xa denote the space of all functions u(x, y, z) that satisfy the
3-dimensional reduced wave equation V 2 u + IX 2 U = in all of E 3 , for fixed IX.
Then the representation of M 3 on X a given by the association with each g
in M 3 of the transformation
p(g):f(x) --+ f(g-l X ) (21.12-5)
of Xa onto itself is irreducible. From the infinitesimal operators of this
representation, a basis in Xa is found, consisting of the functions
Yi (8 ,cpr-
) 1/2 J 1+ 1 / 2 (IXr,
) {I = 0, 1, 2,. . . (2 1.12-6)
m = -I, -I + 1, ... , l.
Therefore, the so-called spherical Bessel functions
Hz) = JY; J 1+ l/z{Z) (21.12-7)
have their origin in the irreducible representations of M 3.

Representations of Mn are discussed in Vilenkin 1968.
In contrast with the irreducible representations of compact groups, which
are finite-dimensional and depend on a discrete parameter [e.g., I =
0, 1,2, ... for SO(3)], the irreducible representations of the rigid motion
groups are infinite-dimensional and depend on a continuous parameter IX.
The irreducible representations of the Lorentz groups are of two kinds:
finite-dimensional ones that depend on a discrete parameter, and infinite-
dimensional ones that depend on a continuous parameter. The former are
obtained from finite-dimensional representations of SL(2, C) (see next
chapter) and appear in the relativistic quantum mechanics of various particles.
Both are needed, in a complete set of irreducible representations; hence, one
must expect that the infinite-dimensional ones may also playa role in physics.
21.13 Characters
The concept of the character X(g) of a representation plays an important
role in representation theory. For compact groups, it is the key to establishing
the completeness of a set of irreducible representations, hence to deciding
whether all representations have been found. If p is a representation on a
finite-dimensional space X n , so that the p(g) are matrices with elements Pjk(g),
then X(g) = tr p(g) = LJ= 1 Pj/g) Hence, X is a scalar-valued function on the
group G. If G is compact, the characters Xl and X2 of two inequivalent irre-
ducible representations are orthogonal with respect to the inner product
(21.5-5) :
(21.13-1)
whereas Xl(g) == X2 (g) if the representations are equivalent. Furthermore,
tlX(gW dg = 1 (21.13-2)
for an irreducible representation, if the Haar measure is so normalized

that the measure (volume) of the whole group is = l.
The character X(g) depends only on the conjugacy class of the group
element g; that is, X(g) = X(hgh - 1) for all g, all h. The set of all the characters
Xi of irreducible representations of a compact group form a complete set of
functions for the expansion of functions f(g) that depend only on the
conjugacy class of g.
In the rotation group SO(3), all rotations through a given angle w belong
to the same conjugacy class, regardless of the direction of the axis of rotation;
hence, the characters are functions of w only. The classical proof of the
completeness of the set of irreducible representations pi (l = 0, 1, 2, ...)
referred to at the end of Section 20.9 consists in showing (see Wigner 1931)
that the corresponding characters Xl(g) form a complete set of functions for
the expansion of functions of w. It then follows, since there is no nonzero
function of w orthogonal to all the Xl(g), that there is no irreducible repre-
sentation inequivalent to all the /.
EXERCISES

1. Show that if two finite-dimensional representations pI and p2 are equivalent
[i.e., if they have the same dimension and if there is a matrix A i=- such that Apl(g) =
p2(g)A for all g], then Xl(g) == X2(g).
2. Show that two rotations gl and g2 through a given angle ()) about two different
axes are conjugate, i.e., that there is a rotation h such that 9 I = hg 2 h- I
3. Show that the characters of the irreducible representations pi of SO(3) are
Xl = sin(/ + t)lX/sin tlX (I = 0, 1, ... ), and show that they satisfy the orthonormality
relation (21.13-1), where dg is =((1 - cos 1X)/(4n 2 1X2))d 3 9, according to Exercise 7 in
Section 21.5. Hint: It suffices to consider rotations about the z axis, for which the
matrices pl(g) are diagonal; see (20.12-2) and the preceding equation.
To show the completeness of the characters i for SO(3), we must show that
if t/I(a) is any continuous function such that (i, t/I) = 0 for alII, then t/I(a) == O.
Since 1 - cos a = 2(sin ta?and d 3 9 ~ 4na 2 da, this is equivalent to
showing that if
2 s: sin(l + t)a sin ta t/I(a)da = 0
for alII, then t/I(a) == O. If we call ta = t and sin ta t/I(a) = x(t), this is equiva-
lent to showing that if
f "/2
o x(t)sin(21 + l)t dt = 0
Characters 79
for all I, then X(t) == o. This is, however, the case, for if x(t) is extended to the
entire interval - n ~ t ~ n by requiring it to be an odd function about
t = 0 and an even one about t = nj2, then the functions sin(21 + l)t
just suffice for the Fourier series for X(t). Since the characters l form a
complete set of functions, the representations pI (l = 0, 1, 2, ...) are all the
irreducible representations of SO(3). This yields the answer to the question
raised in Section 20.2 for the case of rotations of the Cartesian coordinate
axes in 3-space: All possible nonrelativistic transformation laws of physical
quantities are provided by the representations pI of SO(3).
CHAPTER 22
Group Representations and

Quantum Mechanics
Ray space and ray representations in quantum mechanics; extensions of local

representations; effect of double connectedness of 50(3) and simple
con nectedness of 5U (2); double-va Iued representations; spi nors.
Prerequisites: Chapters 18-21 and elementary quantum mechanics.
The purpose ofthis chapter is to elucidate one particular point in the applica-
tion of group theory to quantum mechanics, namely the occurrence of double-
valued or spin representations of the rotation and Lorentz groups.
22.1 Representations in Quantum Mechanics
It has been seen that, in classical physics, various sets of quantities transform,
under rotations of the coordinate axes, so as to give representations of the
rotation group. (The same applies in classical physics to other symmetry
groups, such as the groups of rigid motion, the crystal symmetry groups, the
Lorentz group, and so on.)
In quantum mechanics, on the other hand, some quantities transform, under
rotations of the coordinate axes, like the components of spinors and thus give
representations of SU(2) rather than of the rotation group SO(3). This was
shown by Dirac (in somewhat different language) in his paper (1928) on the
relativistic wave equation, and it was also implicit in Pauli's theory of the
electron spin, published a year earlier. More generally, spinor components
transform under a Lorentz transformation !l' p so as to give representations
of SL(2, q rather than !l' p. This seemed rather surprising at the time, even
though Dirac showed that all observable quantities transform like scalars,
vectors, and tensors, i.e., according to the representations of SO(3) and !l' p. It
was seen in Sections 19.7 and 19.8 that the homomorphisms of SU(2) and
SL(2, q onto SO(3) and !l' P' respectively, are 2-to-l; hence a representation
of the first group can associate two different matrices, M and - M, with each
80
Rotations of the Axes 81
element g of the second group, i.e., with each of the transformations of space-
time. This association is sometimes called a two-valued representation of the
second group. How they arise is discussed in this chapter. It will be seen that
the role of SU(2) and SL(2, C) is to determine the so-called ray representa-
tions of the physically relevant groups SO(3) and !l! p.
Each possible state of a quantum mechanical system corresponds not to a
single vector tjJ in a Hilbert space f>, but to a ray {rxtjJ} consisting of all numerical
multiples of tjJ. If all vectors are normalized (II tjJ II = 1, II rxtjJ II = 1), then rx has
unit modulus (Irxl = 1), but its phase (arg rx) is arbitrary. This arbitrariness
affects the interpretation of representation theory, as will be seen.
22.2 Rotations of the Axes
A state of a system may be regarded as determined in principle by the simul-

taneous measured values {a, b, ... } of a complete set of commuting observables
(self-adjoint operators) {A, B, ... }. Hence, the set {a, b, ... } of numbers
determines a ray {rxtjJ} in a Hilbert space f>. The observables correspond in
principle to an experimental arrangement, or apparatus, for measuring them.
Suppose that the entire apparatus is rotated about some fixed point p into a
new orientation by a rotation g [an element of SO(3)]. It then determines a
new set of similar observables {A', B', ... }. A given state of the system now
corresponds to a new set {a', b', ... } of numbers, which determine a new ray
{(XIV} in f>. Under the action ofthe rotation g, then, each ray {rxtjJ} is mapped
into another ray {rxtjJ'}. These mappings provide a ray representation of the
group, as discussed in the next section.
Suppose that a normalized vector tjJ is somehow chosen in each ray in f>.
Then the rotation g determines a one-to-one mapping among those vectors
in f> that have been thus chosen. We assume, as an axiom of quantum
mechanics, that the tjJ's can be so chosen that the mapping is linear, and hence
can then be defined in all f> by linearity. Since the representing tjJ's were all
normalized, the mapping is a unitary transformation U or U(g). It is not
unique, however, for $iven g, because of the arbitrariness of the phases of the
representing tjJ's. The degree of arbitrariness of U is described by the following
lemma, whose proof is left as an exercise.
Lemma. Let Uland U 2 be two unitary transformations in f> such that,for

every tjJ, the t11'o transformed vectors U 1 tjJ and U 2 tjJ determine the same ray.
That is, there is a complex-valued function P(tjJ) such that U i tjJ = P(tjJ)U2 tjJ,
for all tjJ. Then P(tjJ) = const. = p, where IPI = 1, i.e., U i = pU2
Unitary transformations U and pU, where P is a constant and IPI = 1, are

called equivalent: U ~ pU. We have seen that each rotation g corresponds to
an equivalence class {PU: IPI = 1} of unitary transformations having
different phases arg p.
82 Group Representations and Quantum Mechanics
Now suppose that for each 9 in SO(3) a single unitary transformation U(g)
is somehow chosen from the corresponding equivalence class. If ljJ' = U(g)ljJ
and ljJ" = U(h)ljJ', then the resulting transformation matrix for the mapping
ljJ -> ljJ", i.e., U(h)U(h), is not necessarily = U(hg), but is ~ U(hg). Hence, for
each pair h, 9 of rotations there is a phase factor y(h, g) such that
U(h)U(g) = y(h, g)U(hg), (22.2-1)
where Iy(h, g)1 = 1. Possibilities for the choice of the function y(h, g) are
discussed below.
22.3 Ray Representations

The set S of all rays is called a ray space. It is not a linear space, because if r is
a ray, and c a number, cris not defined, and ifrl and r 2 are two rays, r 1 + r 2 is
not defined. If we drop the requirement of normalization, each ray is a one-
dimensional subspace of ~. From that point of view, the only reasonable
definitions would make cr the same ray as r, even for c =1= 1, and would make
r 1 + r 2 a two-dimensional subspace for ~, and hence not an element of S.
However, each element of S corresponds to a state of the physical system, and
the correspondence is one-to-one. S is a topological, in fact a metric, space in a
very natural way. If r 1 and r 2 are two rays, their distance can be defined as
d(rl' r2) = inf{llljJl -ljJ211: ljJl Er1, ljJ2 Er 2, IIljJ111 = IIljJ211 = 1}.
The physical properties of the two corresponding states (expectations of
observables) are nearly the same, if d(r 1> r 2) is small.
Each rotation in space (each change of orientation of the apparatus)
induces a transformation is S, as described above. It is not a linear trans-
formation, since S is not a linear space, but it is continuous with respect to the
metric in S. The mapping of the elements of SO(3) onto the corresponding
transformations in S is an isomorphism: It is one-to-one, and the product of
any two elements in SO(3) maps onto the compositions of the correspond-
ing transformations in S, etc. Each such transformation in S corresponds to
an equivalence class of unitary transformations in ~. Generally, a homo-
morphism of a group G onto a group of equivalence classes of unity trans-
formations in a vector space V is called a ray represevtation of G on V.
As above, two unitary transformations Uland U 2 in V are equivalent if
U 1 = /3U 2 for some constant /3.
The ray representations of SO(3) on ~ are the physically relevant expres-
sions of spherical symmetry. For calculational purposes, however, it is
desirable to describe the transformations in S by something more tangible,
like matrices. Hence, the question arises of selecting one unitary transforma-
tion U(g) from each equivalence class in some convenient way.
If the phases of the U(g) could be so chosen that the factor y(h, g) in (22.2-1)
were = 1 for all h, all g, then the mapping 9 -> U(g) would be an ordinary
representation of SO(3) on ~. That can't be done in general; what can be done
will now be explained, after first restricting the problem to a finite-dimen-
sionalone.
Local Representations 83
22.4 A Finite-Dimensional Case

Suppose that the physical system has spherical symmetry and that there is a
discrete energy state of energy E with finite multiplicity n. Then the cor-
responding eigenspace of the energy operator is an invariant subspace t)E of
t). Then rays in t)E are transformed under rotations into other rays in t)E;
hence t)E is invariant under each of the operators U(g), and the restriction of
U(g) to t)E can be represented for each g by an n x n unitary matrix, which
will also be denoted simply by U(g). From this point on, the discussion will be
restricted to the finite-dimensional case.
22.5 Local Representations

The physically reasonable assumption is now made that the phases of the
unitary transformations can at least be so chosen that the matrix elements
Uij(g) are continuous functions of g. Then let {)x, {)y, and ()z be the intrinsic
coordinates in SO(3) defined in Section 19.6, and let JV and JV 0 be the sets of
group elements for which 11911 < nand 11911 < n/2, respectively. Although the
group manifold as a whole is doubly connected, the regions JV and JV 0 are
simply connected neighborhoods of the identity. If g and h are in JV 0, gh is in
JV. Now, g, h, and gh are rotation matrices, and the matrix elements of gh are
continuous functions of the matrix elements of g and h; hence, as g and h vary
continuously in JV 0, gh varies continuously in JV.
It will now be shown that the phases of the unitary matrices U(g) described
above can be so chosen in the neighborhood JV that the function y(g, h) in
(22.2-1) is = 1 for all g, h in JV o. When that is done, the mappingg ~ U(g) is
called a local representation ofSO(3). (See Chapter 25.) Namely, since U(g) is
continuous in JV, the multivalued function (det U(gl/" splits into n inde-
pendent continuous branches in JV, because JV is simply connected. Clearly
U(e) is a multiple of the identity matrix I, and we write U(e) = {1"I, where
IPI = 1. Then, (det U(el/n is Ptimes an nth root of unity, and a function lX(g)
can be defined as that branch of (det U(gl/" that is = Pfor g = e. It is now
asserted that if new unitary matrices V(g) are defined in JV by the equations
1
V(g) = lX(g) U(g), (22.5-1)
then
V(g)V(h) = V(gh) for all g, h in JV o . (22.5-2)
To prove this, note that in any case
V(g)V(h) = beg, h)V(gh),
where beg, h) is a continuous function [compare with (22.2-1)]. It is seen from
(22.5-1) that det V(g) = lforallg,henceb(g, h)" = 1,henceb(g, h)issomenth
root of unity for all g, h; but bee, e) = 1, hence beg, h) == 1, by continuity, and
(22.5-2) follows.
22.6 Origin of the Two-Valued Representations

The remaining question is: When can the local representation g ---> V(g) be
extended to a representation of all SO(3)? If H denotes the matrix group
generated by the matrices V(g), g in %, then the mapping g ---> V(g) is a local
homomorphism SO(3) into H. According to Theorem 3 in Section 25.13, a
local homomorphism of a Lie group G into a Lie group H can be extended to a
homomorphism of all of G into H if G is simply connected, but not necessarily
otherwise. G = SO(3) is, of course, not simply connected; however, from the
mapping g ---> V(g) one can construct also a local homomorphism of SU(2)
into H, and that homomorphism can be extended, because SU(2) is simply
connected. Denote by u ---> g(u), where u E SU(2) and g(u) E SO(3), the
homomorphism of SU(2) onto SO(3) that was constructed in Section 19.7.
Then the mapping u ---> W(u) d~ V(g(u)) is a local homomorphism of SU(2)
into H, defined for those values of u for which g(u) E %. Its extension [which
will also be denoted by u ---> W(u)] is a representation of SU(2). Now, the
homomorphism u ---> g(u) is 2-to-l, in fact g( -u) = g(u), hence, the two
equations g = g(u), W = W(u) give either (1) a representation g ---> W of
SO(3)-this in case W(u) = W( -u)-or (2) an association of two unitary
n x n matrices, say VI (g) and Vig), = - VI (g), with each rotation matrix g, in
such a way that each of the four products
V;(g) fj(h) (ij = 11,12,21,22)
is equal to VI (gh) or to V2 (gh). This association is called a two-valued repre-
sentation of SO(3). Clearly every representation of SU(2) thus determines a
two-valued representation, hence a ray representation, of SO(3).
Summary. Since the quantum-mechanical states correspond to rays in the
Hilbert space rather than to vectors, a rotation g of the physical system cor-
responds not to a unique transformation among the vectors of a given in-
variant subspace, with matrix U = U(g), but instead to a set {aU: all a in C
such that Ia I = I} of unitary transformations. It has been shown, however,
that these sets are necessarily so interrelated that, by suitably choosing
matrices from them, one can obtain a representation ofSU(2). This can happen
in either of two ways:
1. It may be possible to choose one matrix U = U(g) from each set so
as to give a representation of SO(3), and hence also a representation of SU(2)
via the homomorphisms
SU(2) ---> SO(3) ---> {U(g)};
(2 x 2) (3 x 3) (n x n)
2. It may be necessary to choose two matrices U and - U from each set
and to associate them with the two elements u and - u ofSU(2), but with only
one element g of SO(3), in such a way that they form an ordinary representa-
tion of SU(2) and a two-valued or spin representation of SO(3). It will be
seen in Chapter 25 that there is no other group related to SO(3) in the way
SU(2) is; hence there are no n-valued representations except two-valued ones.
Representations of SU(2) and SL(2, C) 85
Similarly, for a system that is invariant not merely under the rotation group
SO(3) but also under the entire proper Lorentz group !t' p' the transformation
of the wave functions corresponding to a given transformation g of !t' p is not
unique. Instead, there is a set {aU: Ia I = 1} oftransformations corresponding
to each g, and these sets are so correlated that one can choose transformations
from them so as to give a representation ofSL(2, C) [which is related to!t' pas
SU(2) is to SO(3)]; this may be a representation of!t' p itself, involving scalars,
vectors, or general tensors, or it may be a two-valued representation (spin
representation) of !t'p' In Dirac's theory of the electron, the transformation
laws ofthe four components ofthe electron's wave function give a two-valued
representation of !t'p (see Dirac 1958, p. 258).
It is easy to see that a two-valued irreducible representation cannot be made
into a single-valued one by somehow appropriately choosing one of the two
matrices U and - U that represent each given g of SO(3) (or !t' p); namely, if
U 0 is a matrix that represents a rotation through n, in a two-valued irreducible
representation, it can be shown that U~ = - I, but U~ represents the identity
in SO(3), hence must be = + I in any single-valued representation.
22.7 Representations of SU(2) and SL(2, C)

The discussion so far is rather hypothetical until it can be shown that there
actually exist representations ofSU(2) that give two-valued representations of
SO(3). To be sure, the identity representation ofSU(2) by itself is one such, but
there are many others. In the next few sections the irreducible representations
of SU(2) will be discussed. They are all finite-dimensional, because SU(2) is
compact, while SL(2, C) and !t' p' which are not compact, have also infinite-
dimensional irreducible representations, concerning which the reader is
referred to Valenkin 1968, Naimark 1976, and Sugiura 1975. It turns out that
certain finite-dimensional representations of SL(2, C) remain irreducible
when restricted to SU(2); they give rise to ordinary and spin representations
of the Lorentz and rotation groups.
An element of SL(2, C) is a unimodular transformation of C 2 onto itself
given by
u = (a
y
[3):
i5
(Xl)
X2
--+ (aXI + [3X 2),
yXI + i5x 2
(22.7-1)
where ai5 - yf3 = 1. The matrix of the inverse transformation is
u- l = ( i5 -f3).
-y a
Now the action of the group on C 2 is effective (any transformation u i= e

moves at least one point of C 2 ) and transitive (given any two points x and y,
there is always an element u in the group such that y = ux), that is, C 2 is a
homogeneous space for SL(2, C). Therefore, let XOO be the space of all entire
analytic functions f(x 1 , x 2 ) of two complex variables. Then, according to

(20.6-1) an infinite-dimensional representation of SL(2, IC) is obtained by
associating with u the transformation p(u) in X OCJ given by
(22.7-2)
Certain elements of the subgroup SU(2) are now considered. Let Wi> W 2 ,
W3 be the intrinsic coordinates in SO(3) defined in Section 19.6, let gWl, W2, W3 be
the corresponding rotation matrix [element of SO(3)J, and let U W1 , W2, W3 be
the elements SU(2) that are mapped onto gWl, W2, W3 by the homomorphism of
Section 19.7. In particular, one can take
(
COS w/2 -i sin W/2)
uw,o,o = -i sin w/2 cos w/2 '
Uo = ( .
COSw/2 -sin W/2) (22.7-3)
,W , 0 sm w/2 cos w/2 '
e - iwj2
Uo,o,w -- (
0 ei~j2 ).
because a direct calculation, using the equations of Section 19.7, shows that
the corresponding transformations from x, y, z to x', y', z' are those given by
the matrices
~G
0
g..,OO 0
W
-~).
0
go"o ~ ( -w~ 0
0
en)
o,
0
(22.7-4)
~G
-w
goom 0
0 ~}
in agreement with (19.6-1). Infinitesimal group elements ofSU(2) are obtained
accordingl y :
T, - - u
1 -
o
OW w,O,Olw=O - -
- 2
1( 0-i
T.2 - -u o --
- OW O,w,Olw=O - 2 1
1(0 -1)0' (22.7-5)
T3 =
o uo,o,wlw=o = 21(-i0
ow 0)
i'
Irreducible Representations of SU(2) 87
The corresponding differential operators of the representation pare
(22.7-6)
L3
o p(uo,o,,,,)lro=o ="2i(Xl OX0 -
= ow X2
0)
OX 2
I
The infinitesimal elements and operators satisfy the commutation relations
[T;,1jJ = 7;" (ijk = 123,231,312). (22.7-7)
We note in passing that the matrices (22.7-5) can be regarded also as the
infinitesimal group elements of the larger group SL(2, q, for the following
reason: First, it is easily verified that the matrices (22.7-3) are given in terms
of the matrices T; by the equations
uro,O,O = exp(wTI ),
uo, ro, 0 = exp(w T2 ),

UO,O,ro = exp(wT3)'
By the methods of Chapter 25 (exponential mapping), it is seen that, more

generally,
(22.7-8)
From (22.7-5) it is seen that the right member of this last equation is of the
form exp(iA), where A is a general 2 x 2 Hermitian matrix of trace zero. If,
now, WI' w 2 and W3 are allowed to take complex values, then it is of the form
exp B, where B is a completely general 2 x 2 matrix of trace zero, and then
exp B is a general 2 x 2 matrix of determinant = 1, i.e., a general element of
the group SL(2, q.
22.8 Irreducible Representations of SU(2)
For each value O,!, 1,1,2, ... of an index I, a subspace X 2 l+ I of XOO is defined
as the space of all homogeneous polynomials in X I and X 2 of degree 21. From
(22.7-2) it is seen that each operator p(u) transforms any homogeneous
polynomial into another homogeneous polynomial of the same degree; hence
each subspace X 2 l+ I is invariant under p(u), not only for all u in SU(2), but
also for all u in SL(2, q.
It will be shown that the representation ofSU(2) given by (22.7-2) on each
subspace X 21 + I (it will be called Dl) is irreducible; hence, the representation of
SL(2, IC) on X 21 +1 is a fortiori also irreducible. It will be shown that any

subspace of X 21 + 1 (other than the subspace consisting of the zero vector alone)
which is invariant under SU(2) is all of X 21 + 1. This is done by the now familiar
method of raising and lowering operators: Any such subspace is invariant
under the operators L l' L 2 , L3 of equations (22.7-6), hence also under
L1 iL2
The monomials
(m = -1,-1 + 1, ... , l) (22.8-1)
are a basis in X 21 + 1; fm is an eigenfunction of the operator L3 with eigenvalue
im. Note that m assumes integer or half-odd-integer values according as I is an
integer or half an odd integer. Any function g in X 21 + 1 can be written as
L em fm By an argument like that of Section 20.5 it follows that if an invariant
subspace of X 21 + 1 contains such a function g, then it contains all those
monomials fm' individually, for which em =I- O. Namely, if the subspace con-
tains the function g, then it contains the function L 3 g (because the subspace
is invariant under L 3 ), also the function P(L 3 )g, where P is any polynomial,
but P can be chosen as to annihilate all the terms of L em fm except one (P can
be taken as the Lagrange interpolation polynomial which vanishes for all
eigenvalues im of L3 except one). Hence, the invariant subspace contains at
least one of the functions fm. But L1 + iL2 is a lowering operator, i.e., it
converts fm into a multiple of fm- b except that it converts f-l into zero; and
L1 - iL2 is a raising operator, i.e., it convertsfm into a multiple offm+ 1, except
that it converts j; into zero. Hence, the invariant subspace is all of X 2 l+ 1, as
was to be proved.
It will be shown in the next section that the Dl (l = 0,1, I,!, ... ) are the only
irreducible representations of SU(2).
The two-valued representations of SO(3) come about as follows: If
u E SU(2), then, under the homomorphism of SU(2) onto SO(3) of Section
19.7, u and -u are both mapped onto an element g ofSO(3). If Dl is one of the
representations of SU(2) found above, then the mapping g ~ D1(u) is a
single-valued representation of SO(3) on X 21 +1 if Dl( -u) = Dl(U) and is a
double-valued representation if Dl( -u) =I- Dl(U). It is only necessary to
examine the case of u = 12 ; then Dl(U) = 121 +l' (Here, h denotes the k x k
unit matrix). Under the mapping determined by the matrix -12' Xl and X 2
go into -Xl and -X 2 ; if 21 is even, the monomial fm goes into itself; hence
Dl( -1 2 ) = 121 +1, and the mapping g ~ D1(u) is the ordinary representa-
tion of odd dimension 21 + 1 found in Section 20.9. If 21 is odd, fm goes into
-fm; hence Dl( -12) = -121+ b and the mapping g ~Dl(U) = D1(u) is
a double-valued or spin representation of even dimension.
In a similar fashion, the irreducible representations ofSL(2, IC) given above
[from which those of SU(2) were obtained by restriction] lead to finite-
dimensional ordinary and spin representations of the Lorentz group fi' p'
However, there are still other irreducible finite-dimensional representations
of SL(2, IC), given in Section 22.11 below; they also determine ordinary and
spin representations of fi' p' none of them unitary.
Functions of z and z 89
22.9 The Characters of SU(2)

The conjugacy classes of SU(2) are easily determined. First, matrices U 1 and
U2 in SU(2) have the same eigenvalues if and only ifthere is a unitary matrix U
[element of U(2)] such that u*u1u = U2. Since U can be multiplied by any
complex number of modulus 1, we can assume det U = 1 without loss of
generality, and then U is in SU(2). It follows that U1 and U2 are in the same
conjugacy class in SU(2) if and only if they have the same eigenvalues. There-
fore, each conjugacy class can be represented by a matrix of the form
U- iaI 2
U= (

for some rJ. in [0,2n]. For such u, the operator DI(u) simply multiplies the
basis vector fm (22.8-1) by eima ; hence DI(u) is a diagonal matrix, whose trace is
I( ) _ sin(l + t)rJ.
X rJ. - . 1 ' (22.9-1)
sm zrJ.
just as for the case in which 1is an integer, according to Exercise 3 at the end of
Section 21.13. It was shown in that section that the functions 22.9-1, for
1 = 0, t, 1, ~, . .. form a complete system for the expansion of functions
depending only on the conjugacy class on the manifold of SU(2); hence, the
representations DI exhaust the irreducible representations of SU(2).
22.10 Functions of z and z

The notation introduced in this section is convenient for discussing the
representations of SL(2, IC) and is used in many branches of mathematics.
Let u(x, y) and vex, y) be two real COO functions of two real variables x and y.
We write z = x + iy, fez) = u(x, y) + iv(x, y), so that fez) is a complex-
valued function of the complex variable z, not in general analytic. If fez) is
analytic, then its derivative can be written in various forms, using the Cauchy-
Riemann equations, namely
a a
/,(z) = ax (u + iv) = -i ay (u + iv)
= aiu + iv),
where az is the operator
az ~ (:x - i :y).
= (22.10-1)
An operator az is similarly defined as
az= ~ (:x + i :y). (22.10-2)

and it is noted that Oz fez) == 0, if fez) is analytic, by virtue of the Cauchy-

Riemann equations. On the other hand, iff(z) is a polynomial (or convergent
power series) in Z, then ozf(z) = o. Furthermore, the operators (22.10-1, 2)
are linear differential operators; hence the usual rule for differentiating a
product holds. Therefore, if f is a polynomial (or a convergent power series)
in both z and z[in which case one usually writes fez, z) to indicate that it is not
necessarily analytic in either z or z], then z can be regarded as a constant for
the purpose of computing oz, and z a constant for computing Oz. That is, z
and z may be regarded as independent variables for purposes of differentia-
tion.
22.11 The Finite-Dimensional Representations

ofSL(2, C)
The method of homogeneous polynomials used in Section 22.8 for SU(2) can
also be used for SL(2, iC), but here a new aspect appears. Given a representa-
tion p of a group G, there are many ways in which another representation p'
can be obtained. Among them is
p'(u) = p(u), 'ifu in G, (22.11-1)
[which means that each matrix element Pmn(u) is replaced by its complex
conjugate], for then P'(U 1 U2 ) = p'(u 1 )P'(U 2 ), etc. Another possibility is
(22.11-2)
if the representation is unitary, this is the same as (22.11-1). If G is itself a

group of matrices, then two other possibilities are
p'(u) = p(ii), (22.11-3)
p'(u) = pu T ) - 1). (22.11-4)
If G is a unitary group, e.g., U(n) or SU(n), then (22.11-3) and (22.11-4) are the
same, but otherwise, they are generally different.
We now show that ifG is SU(2), then the representation p' given by (22.11-3)
is equivalent to p; hence, in this case, no new representations are obtained by
these methods, and that is why these methods were not used in Section 22.8.
Namely, call
so that generally
Y
-l(ac b)d Y ( d
-b
-c).
a
(22.11-5)
The Finite-Dimensional Representations of SL(2, q 91
In particular, for any u in SU(2), U = y-1 uy, which can be seen by writing u as
(-b ~). Then, since y is also in SU(2),
p'(u) = p(y- 1uy) = p(y)-l p(u)p(y),
for all u; hence p and p' are equivalent representations.
When the representations (22.11-3) and (22.11-4) are extended from SU(2)
to SL(2, C) by writing
p'(m) = p(m) (22.11-6)
and
(22.11-7)
respectively, for m in SL(2, IC), they are no longer identical, or even equivalent.
The second one is equivalent to p, because (m T )-l = y-1 my, by (22.11-5),
since det m = 1, while (22.11-6) is inequivalent to p, for if the equations
p(m) = V-I p(m)V (22.11-8)
held for all m, then V would have to be = p(y) in order that this equation be
satisfied for mE SU(2), in which case it would not be satisfied for mf/: SU(2),
since y- 1my is not in general =m for such m.
Clearly, then, SL(2, IC) has more representations, in some sense, than
SU(2). To find them, we let XOO denote the set of all complex-valued functions
of two complex variables Xl and X 2 that are Coo in the real sense rather than
entire analytic in contrast with Section 22.10, and we denote these functions
by f(x!> X2, Xl' X2), following the procedure of Section 22.10. In place of
(22.7-2), we write
(p(u)f)(x t , x 2, Xl' X2)

= f(6x 1 - {JX2' -yxi + ax 2, bX1 - PX2' -YX1 + iXx 2), (22.11-9)
where u is the matrix
a6 - {Jy = 1, (22.11-10)
i.e., any matrix in SL(2, C). Now, in addition to the three matrices (22.7-3),
which are in SU(2) and correspond to rotations in space, and which determine
the infinitesimal operators L 1, L2 , and L3 by (22.7-6), we have three additional
matrices,
cosh m/2 sinh mlh)
( (m = CPJ,
sinh m/2 cosh m/2
cosh m/2 -i sinh m/2)
( (22.11-11)
i sinh m/2 cosh m/2
(e~2
which correspond to Lorentz transformations and determine further infin-

itesimal operators K 1, K 2, and K 3' Hence, the infinitesimal operators of
SL(2, IC) are obtained by substituting the matrices (22.7-6) and (22.11-11) into
(22.11-3) and taking the derivative of each with respect to w at w = O. If we
recall that x 1> X 2 , Xl' and x2 are to be regarded as independent for purposes of
differentiation (see preceding section), we find that
(22.11-12)
1 1
K3 = -"2 (Xl aX! - X2 aX) - "2 (Xl aX! - x
2 ax).
The complete commutation relations are
[Li' L j ] = Lk
[Ki' K j ] = Lk
(ijk = 123,231, or 312). (22.11-13)
[Ki' LJ = 0
[Ki' L j ] = -Kk
We introduce also the operators
L = Ll iL2' (22.11-14)
22.12 The Irreducible Invariant Subspaces of

X OO for SL(2, C)
As basis vectors for X oo , or at least for the set of all polynomials in X oo , we
introduce the monomials
1 I-m I+m-l'-m'-l'+m'
l/l = l/llml'm' = eXl X2 Xl X2 , (22.12-1)
where
c2 = (1- m)! (/ + m)! (I' - m')! (I' + m')!,
Spinors 93
where I and I' are any two of the numbers 0, !, 1, !, ... , and where
m = I, I - 1, ... , -I,
m' = 1', I' - 1, ... , -I'.
For given I and 1', the space X(I, 1') spanned by the "'1m I'm' is the space of all
homogeneou~ polynomials of degree 21 in the variables x 1 and X 2 and of degree
21' in Xl and X2' This space has (complex) dimension (21 + 1)(21' + 1). It is
clear from (22.11-9) that each subspace X(l, I') is mapped into itself under
every p(u), and hence is an invariant subspace. From (22.11-12) we see that
L 3", = i(m - m')"',

K 3 ", = (m + m')"';
i.e., m is an eigenvalue of -!iL3 + !K 3, and m' is an eigenvalue of !iL 3 + !K 3.
It follows, by the same kind of argument used in all previous cases, that if an
invariant subspace contains any function (polynomial) in the subspace
X(I, I'), then it contains every monomial "'Iml'm' which appears in that poly-
nomial with a nonzero coefficient. We find furthermore that
L+ + iK- = -2ixl oX2'

L- + iK+ = -2ix2 oXI'
L - - iK+ = 2ixl oX2'
L + - iK- = 2ix2 oXI'
Hence,
L+ + iK- lowers m,
L- + iK+ raises m,
L - - iK+ lowers m',
L+ - iK- raises m',
in the sense that (L + + iK-)"'lml'm' is proportional to "'lm-ll'm" etc. Hence,
for given I and 1', all the "'Iml'm' are linked by the infinitesimal operators. We
conclude that if an invariant subspace contains any function (polynomial) in
X(I, 1'), then it contains the entire subspace X(l, 1'). That is, p, when restricted
to X(I, 1'), is irreducible; it is denoted by p(I,l'); these are the only finite-
dimensional irreducible representations of SL(2, <C).
22.13 Spinors
Spinors are sets of quantities related to SU(2) and SL(2, <C) in the same way
that the tensors (including vectors and scalars) of prequa.ntum physics are
related to the physical groups SO(3) and .ff'p' Their transformation laws
constitute certain (generally reducible) representations of SU(2) and SL(2, C),

and hence certain one- or two-valued representations of the physical groups.
Those spinors whose transformation laws constitute ordinary or one-valued
representations of the physical groups are in fact tensors, slightly disguised
(see Exercise 1, below); in this sense, tensors are special cases ofspinors.
In order to describe a tensor, one associates a set of quantities (say,
TI I, TI 2, ... , T4 4) with each frame of reference, and then the relations between
these sets, for different frames of reference, constitute the transformation laws.
Once an initial frame of reference has been specified, this is equivalent to
associating one such set of quantities with each element of the rotation or
Lorentz group; the corresponding sets, for other frames of reference, are then
given by the transformation laws. In the case of spinors, one associates a set of
quantities for the initial frame of reference, with each element of the group
SU(2) or SL(2, C), and this amounts to associating two sets, differing, however,
only in phase (in fact, only in sign), with each physical frame of reference.
Since phases, hence signs, are physically irrelevant, we shall speak loosely of
the association of a set of quantities with each frame of reference.
A spinor of rank 1, then, is the association of a pair of numbers ~ I and .~2
with each frame of reference, in such a way that under a Lorentz transforma-
tion corresponding to an element
of SL(2, IC), the ~'s transformation according to the Law
~I -+ ~~ = mll~1 + m I 2 ~2' (22.13-1)

~2 -+ ~~ = m21~1 + m22~2'
Higher rank spinors are defined in complete analogy with tensors: A spinor of
rank r is the association with each frame of reference of a set of 2' complex
numbers ~a, ... ar (each index takes on the values 1 and 2), with the transforma-
tion law
2
~~, ... ar = II
(P, ... ,Pr)ma,p, ... marPr~P' ... Pr (22.13-2)
If we were concerned only with rotations of the x, y, z axes, and not with
Lorentz transformations, hence only with the group SU(2), this is all that
would need to be said. It was seen in the preceding section, however, that in
the study of the representations ofSL(2, C), the matrix mplays a parallel role
with m. A dotted spinor ofrank 1 is the association of a pair of complex numbers
~i, ~i with each frame of reference, according to the transformation law
~'i = mll~i + m12~i' (22.13-3)

~i = m21~i + m22 ~i'
Spinors 95
Finally, a mixed spinor having r undotted and s dotted indices is the associa-
tion of 2r + s complex numbers with each frame of reference, with the trans-
formation law
= L(YJ. ... ,Yr,b" ... ,bs) m",lY"" m"'rYrmp,b, m psbs

1
(22.13-4)
EXERCISES
1. Let ~otJi be a mixed spinor of rank 2, and define quantities Vj U = 1, ... ,4) by
VI = ~Ii + ~2i'
V 2 = ~Ii - ~2i,
V3 = ~ti + ~2i,
V4 = i(~li - ~2i)
Show that quantities VI' ... , V4 transform like the components of a vector under rotations
and Lorentz transformations. Show similarly that a mixed spinor of rank 2r having r
dotted and r undotted indices determines a tensor of rank r.
2. A spin or is called symmetric ifit is symmetric in the dotted indices (unchanged
by any permutation of the dotted indices) and also symmetric in the undotted indices;
show that the transformation law of such a spinor gives a representation of SL(2, C),
which is equivalent to the representation p(l, I') defined in the preceding section, where
21 and 21' are the numbers of dotted and undotted indices, respectively.
CHAPTER 23
Elementary Theory of Manifolds
Locally n-dimensional space; sphere; torus; disk; Mobius strip; Klein bottle;
identification of edges; coordinate charts; compatibility of charts; induced
topology; Hausdorff separation axiom; manifold; curves; functions on a
manifold; connectedness; simple connectedness; component; homotopic
curves; homotopy classes of curves; fundamental group; double
connectedness of SO(3); configuration space of a mechanical system;
Cartesian product manifolds.
Prerequisites: Chapters 18 and 19.
The theory of manifolds is basic for the theory of Lie groups and Riemannian
and Einsteinian geometries. The introduction of the manifold concept into
general relativity around 1960, mainly by Martin Kruskal, put a new light
on that subject and clarified the topological properties, both local and global,
of space-time models. Statistical mechanics deals with flows on manifolds.
Other applications of manifolds to physics appear from time to time, because
of their basic geometric nature. Only finite-dimensional manifolds will be
discussed. For more general manifolds, see Lang 1962.
23.1 Examples of Manifolds; Method of

Identification
An n-dimensional manifold is, roughly speaking, a space that is locally
topologically indistinguishable from Euclidean n-space En; that is, each
point of the manifold lies in a region (connected open set) that is homeo-
morphic to a region in En. The formal definition is given in Section 23.4,
below.
Any Euclidean space itself is trivially a manifold. A simple nontrivial
example is the 2-sphere, i.e., the set S2 of points in E3 for which x 2 + y2 +
Z2 = 1, where x, y, and z are Cartesian coordinates. Any sufficiently small
region U on S2 can be mapped onto a plane region in a one-to-one manner
by any of the projections used by cartographers. (If stereographic projection
96
Examples of Manifolds; Method ofIdentification 97
is used, U is all of S2 minus one point.) Hence S2 is a 2-dimensional manifold.

The torus, that is, the surface of a ring, is another 2-dimensional manifold.
Any open subset of a manifold (e.g., the open disk X2 + y2 < 1 in the
plane) is a manifold. The Mobius band (minus its edge) is a manifold.
Boundary points must be omitted, because a boundary point does not lie
in a part of the surface that can be mapped onto an open set in the plane.
The ball and the solid torus (minus their surfaces) are 3-dimensional mani-
folds. For manifolds with boundary, which will not be discussed here, see
Lang 1962.
The surface known as the Klein bottle is a 2-dimensional manifold.
Imagine a surface in the form of a wine bottle with a long neck and a re-
entrant bottom, as in (a) in Figure 23.1. Then imagine that the neck is bent
downward so as to penetrate through the side of the bottle, as at Q in (b), and
then joined, inside the bottle, onto an opening in the reentrant bottom, thus
forming a closed one-sided surface. This surface cannot be represented as
immersed in V 3 without self-intersection, as at Q, but it can be so immersed in
V 4 . To see that, start with the surface immersed in V 3, as at (b); each point of it
then has three coordinates x I' X 2, X 3. It is only necessary to assign a fourth
coordinate X 4 to each point in such a way that X 4 varies smoothly over the
surface, but in a neighborhood of Q takes on one value on the bottle (say 0)
and a different value on the neck (say 1). Then, no two different points of the
surface have the same set of coordinates Xl' X 2 , X 3 , X 4 .
Many 2-dimensional manifolds can be described by the method of
identification or gluing together of edges. The Mobius band is obtained by
starting with a rectangle ABCD, as in Figure 23.2, and then identifying each
point of the edge AB (taking the points in order from A to B) with the cor-
responding point of the edge CD, so that A is identified with C, B with D,
etc. The identified points are regarded as single points of the manifold, just
(a) (b) Figure 23.1 The Klein bottle.

98 Elementary Theory of Manifolds
A B
D C Figure 23.2
as if the rectangle were a narrow strip of paper, which has been bent into a
circle, has had one end twisted through a half turn, and then has had the edges
glued together.
If, in the above example, the edges AD and CB are also identified, in the
same manner, the Klein bottle is the result. (This would require considerable
stretching of the paper, to say nothing of the problem of self-intersection.)
Group manifolds were discussed in Section 19.5. The manifold of SO(3}
was realized as a certain 3-dimensional algebraic surface in a 9-dimensional
space. This surface is homeomorphic, in some neighborhood of each of its
points, to a region in E 3 , but as a whole is not homeomorphic to any region in
E3; it will be seen in Section 23.7 that it has a kind of connectivity that a
region in E3 cannot have.
The manifold of SO(3} can also be regarded as obtainable by a 3-di-
mensional version of the method of identification, used above for the Mobius
band. If ()x, ()Y' ()z are the intrinsic coordinates introduced in Section 19.6,
then each point of the ball 11911 :s; n represents a single element of SO(3),
and conversely, except that any two antipodal points on the surface, 9 and
-9, where 1/91/ = n, represent the same element of SO(3} and must be iden-
tified with each other. The identification cannot be achieved, in analogy with
the Mobius band, by distorting the sphere in 3-dimensional space and gluing
surfaces together, but evidently it can be achieved by suitably distorting the
sphere in a 9-dimensional space.
According to Exercise 1 in Section 20.6, the manifold of SU(2} can be
realized as the 3-sphere, i.e., the unit sphere in E4. This manifold is simply
connected but is also not homeomorphic to any region in E3.
23.2 Coordinate Systems or Charts;

Compatibility; Smoothness
Let 9Jlo be a space-i.e., a set of object called points. (9Jlo may be a group, or j
the like, but is not yet assumed to have any topological structure-see Note
in Section 23.4.) An n-dimensional coordinate chart in 9Jlo is the assignment
Coordinate Systems or Charts; Compatibility; Smoothness 99
of an n-tuple {xl, ... , xn} ~ x of real coordinates to each point P ofa specified
subset U of ffilo in such a way that the assignment P --+ x is a one-to-one
mapping <p of U onto a connected open set N in the coordinate space IRn; one
writes x = <p(P), and one refers to the triple {U, cp, N} as a coordinate chart
in ffil o. The notation is, of course, redundant, since U and cp determine N,
but it is convenient (see Note in Section 23.4). The vector x = cp(P) is
sometimes called the coordinate of P.
For example, if f) and qJ are polar coordinates on the sphere (f) = Xl and
qJ = XZ), then the mapping is from certain points of the sphere onto points
of the open rectangle (0 < f) < n, -n < qJ < n) in the f), qJ plane IRz. To
make the mapping one-to-one, it is necessary to omit the north and south
poles, f) = 0 and () = n, respectively, and the international date line, qJ = n.
To describe the entire sphere, one might use the method of identification,
that is, extend the mapping to the boundary of the rectangle and then decree
that the points () = 0, - n :-:; qJ :-:; n are all one point at the north pole, also
the points () = n, - n :-:; qJ :-:; n at the south pole, and that, for each () in
(0, n), the points with qJ = +n and qJ = -n are the same point. However, in
order to be able to impose smoothness requirements and ensure that the
surface, when glued together, really looks like a sphere, not like a kreplach
or a sopaipilla, a different procedure is needed.
If {U b CPI, N d and {U z , cpz, N z} are two overlapping charts in ffil o ,
they esta1:>lish a relation between the two sets of coordinates for points P in
the intersection U I II Uz , and this relation is one-to-one, because each of the
mappings P --+ CPI(P) and P --+ cPZ{P) is one-to-one. If we write x = CPI(P)
and y = cPz{P), then the resulting relation between x and y and its inverse
involve functions that will be denoted as follows:
i = 1, ... , n, (23.2-1)
i = 1, ... , n. (23.2-2)
These functions are assumed continuous and differentiable a certain number

of times. Also, if U I II Uz is not empty, each of the transformations (23.2-1)
and (23.2-2) is assumed to apply throughout an open region in ~n. That is,
the two charts are called Ck-compatible if
1. the sets {x = CPl(P): P E U I II Uz} and {y = cpz(P): PE U I II Uz}

are open sets in IRn, and
2. the functions (23.2-1) and (23.2-2) are of class Ck
IfU l II Uz is empty, the charts are automatically compatible.
Note. It follows from 2 that if either of the sets referred to in 1 is open in IRn,
the other is, too.
It is intuitively clear than any two reasonable charts in a reasonable space

will always be compatible. The relationships are shown schematically in
P~i_---+-----4
~i
""------.yi
Figure 23.3 Schematic sketch of two charts in a manifold.
Figure 23.3, where the various mappings indicated are the following:
<Pl: P ....... <Pl(P),

<P2: P ....... <piP),
If the polar angles e, cp on the sphere are the coordinates in the first system,
a second system can be chosen so that the coordinates e', cp' are polar angles
with respect to different axes. For example, the north pole N' in the primed
system (()' = 0) might be taken as the point (e = n/2, cp = n/2), in the old
system, and the angle cp' about this new north pole so chosen that the new
international date line is the portion (e = n/2, - n/2 < cp < n/2) of the old
equator. See Figure 23.4. It is clear that these two coordinate systems
together completely cover the sphere.
Figure 23.4
Definition of Manifold; Hausdorff Separation Axiom 101
EXERCISES
1. Find the transformations (23.2-1, 2) for this example, i.e., the relation between
e, !.p and e', !.p'.
2. Describe coordinate systems on the surface of the torus. Show that the torus
can be covered by two charts, but if simply connected charts are required, three are
needed.
23.3 Induced Topology

If {U, <p, N} is a coordinate chart in a space IDlo, if Uo is any subset of U
whose image
<p(U o) = {x = <p(P): P E Uo}
is an open set in the coordinate space ~n, then Uo is called an open set in IDl o.
In particular, U itself is open. In ~n, the intersection of finitely many open
sets is open, and so is the union of an arbitrary collection of open sets.
Hence, the open sets in IDlo determined by a chart have these same properties.
Conditions 1 and 2 for the compatibility of two charts ensure that these
charts determine the same topology in their overlap. In particular, condition 1
ensures that the intersection ofthe open sets U 1 and U2 is open. In the follow-
ing example, two overlapping charts are given that satisfy condition 2
(trivially) but are not compatible because they do not satisfy condition 1:
Let Wl o consist ofthose points (x, y) in ~2 on the x and y axes; define two one-
dimensional charts in IDlo, one on each axis, as follows:
U 1 = {(x, y): x = O}, U2 = {(x, y): y = O},
<Pl x, y)) = y, <P2 x, y)) = x,
Nl =~, N2 =~.
Then, U 1 n U2 consists of a single point (x = 0, y = 0), which is not an open

set in either chart. More generally, two n-dimensional charts might intersect
on a surface of smaller dimensions than n (and even in such a way that con-
dition 2 is satisfied in the surface), and in this case their intersection would
not be an open set in ~n. Condition 1 prevents this sort ofthing by requiring
the intersection to be n-dimensional.
When a space IDlo is covered by a collection of compatible charts, its
topology is completely determined by defining the open sets to be the open
sets determined by the individual coordinate charts, as above, together with
arbitrary unions of such sets. In particular, then, the entire space IDlo is an
open set.
23.4 Definition of Manifold; Hausdorff

Separation Axiom
Note. The present discussion differs from the usual discussion of manifolds
in one respect. In the usual discussion, one starts with a space in which a
topological structure has previously been defined; in fact, it is assumed to be a
Hausdorffspace. (However, see Lang 1962.) One then requires the coordinate
systems to be continuous with respect to that topology. On the other hand,
the existence of coordinate systems restricts the topology considerably,
in fact, in such a way that the space is locally Euclidean (this refers to topo-
logical, not to metric properties). For the purpose of this book it seems better
to let the topological properties be entirely determined by the coordinate
systems. Then, only the familiar topological concepts of Euclidean spaces
are needed, except for one consideration: When a manifold is being con-
structed by piecing together two or more coordinate charts, care must be
used to ensure that the Hausdorff separation axiom is satisfied - this question
will be discussed below.
Note. The discussion starts with a space IDlo, which is simply a collection
(uncountably infinite) of otherwise undefined elements, called points. In
some applications, the space IDlo is given in advance; for example, it may be
a group. In Riemannian geometry or general relativity, on the other hand,
one starts with a set of functions gflV(X I , ... , xn) defined for coordinates
xl, ... , xn lying in a certain domain N of the coordinate space IRn; each
point of N is then assumed to determine a point P of the Riemannian mani-
fold or the physical space being constructed or described. Then, the part of
the manifold or physical space thus described may be extended, by means of
coordinate transformations like (23.2-1, 2), and so on, until one believes, on
the basis of some criterion or other, that the complete manifold has been
defined (see, for example, Kruskal's criterion of geodesic completeness
described in Chapter 28). In this method, nothing is said in advance about
the abstract space IDlo or its subsets U, U/, etc., until the description is complete.
Each chart is specified by describing N, and nothing is said explicitly about
U and (j); hence we prefer to retain" N" in the designation {U, (j), N} of a chart.
A manifold is required to satisfy the following axiom, which expresses an

obvious property of Euclidean spaces:
Hausdorff Separation Axiom. If P and Q are any two distinct points, then
there are neighborhoods U and m
of P and Q, respectively, such that
Un m = o.
When two coordinate systems are pieced together to make a space, it is

possible to violate this axiom, as the following one-dimensional example
shows: The space IDlo consists of three copies of the line IR, and the points in
IDlo are denoted by {x, oc}, where x is a real number, and oc denotes one ofthe
letters a, b, or c. Two charts are defined in IDlo as follows:
UI = {{x,a}:x 2 O} u {{x,b}:x < O},
(j)l ({x,oc}) = x, NI=IR,
U2 = {{x,c}:x 2 o} u {{x,b}:x < O},
(j)2 ({x, oc}) = x, N2 = IR.
Curves and Functions in a Manifold 103
Each chart, by itself, is homeomorphic to IR, but the points P = {O, a}

and Q = {O, c} are not separated, because any two open intervals that con-
tain P and Q, respectively, have points of the form {x, b}, x < 0, in common.
It is clear that this phenomenon can easily be avoided in the practical
construction of manifolds.
Definition. An n-dimensional manifold IDl is a space IDlo together with a (finite or)
countable set of compatible n-dimensional coordinate charts, which together
cover IDlo in such a way that the resulting topology satisfies the Hausdorff
separation axiom. It is understood that compatible coordinate systems can
be added or deleted at will, so long as the space IDlo is kept covered at all
times; the intrinsic properties of IDl are those properties that are unaltered by
such additions or deletions.
IDl is called a Ck-manifold if the transformations (23.2-1,2) between any

two coordinate systems are of class C k , that is, if the functions Xi( ... ) and
yi( ... ) have continuous partial derivatives (pure and mixed) of all orders
up to and including order k. In this case, the addition of new coordinate
systems is restricted to ones that satisfy this requirement. Similarly, IDl is
called a Coo-manifold, if the transformations are of class Coo; it is called a real
analytic manifold if they are analytic. The manifolds of continuous groups
(Lie groups) are real analytic. In general relativity, on the other hand, it is
advisable to admit Ck manifolds with finite k, because the Einstein field
equations are of hyperbolic type, and gravitational waves can in principle
propagate discontinuities of various derivatives of the components gl'v
of the metric tensor.
EXERCISE
Intrinsic coordinates 8x , 8y , 8z were defined in the rotation group SO(3) in Section

19.6. To serve in the above definition, they must be restricted to the open ball 8; + 8; +
8; < 7r. Introduce additional coordinate systems so as to cover the manifold of SO(3).
(A general method of doing this for groups is given in Chapter 27.)
23.5 Curves and Functions in a Manifold
Suppose f(P) is a real- or complex-valued function defined for all points P

in a manifold IDl. If {U, <p, N} is any coordinate chart, then one can define a
function l(x) = l(xl, ... , xn) by the equations l(x) = f(P), x = <p(P) for
all points x in the open set N into which U is mapped by the mapping
P ~ x = <p(P); since this relation is one-to-one, 1{ ... ) is well defined.
If the resulting function 1( ... ) is continuous for every choice of the chart
{U, <p, N} in IDl, then f(P) is called a continuous function on IDl. If IDl is a
Ck-manifold and the 1 are of class C' (r ::; k), then f(P) is called a function
of class C' on IDl. Clearly, there can be no functions of class C' on IDl, if r > k,
other than constants, because allowable coordinate transformations could
destroy the existence of derivatives of order > k. Continuous or C functions

can also be defined in a portion of IDl.
If pet) is a one-parameter set of points in IDl, where t is a real variable, and
if, for every chart {U, <p, N}, the functions fi(t) given by
x(t) = <p(P(t
are continuous functions of t, for all t for which they are defined, then pet)
is called a curve or path in IDl. If the fi(t) are of class C, then pet) is said to be
of class c. As t varies in an interval [t1' t 2], the function pet) describes a
curve ~ going from the initial point P(t1) to the terminal point P(t2). It is
assumed that all curves are either piecewise differentiable or at least recti-
fiable (i.e., that the image in any coordinate chart is such), unless otherwise
specified. To make that possible, it is assumed that all manifolds considered
are at least of class C 1.
Continuous or class C functions of two or more variables pet, s, ... )
are similarly defined.
23.6 Connectedness; Components of a Manifold
If IDl is such that, given any two points PI and P 2 in it, there is a curve, ~
in IDl that goes from PI to P2, then IDl is called pathwise or arc wise connected.
If 9Jl is arcwise connected and is furthermore such that, given any two
curves ~1 and ~2 going from any point PI to any other point P 2, ~I can be
continuously deformed into ~2 in IDl; i.e., if there is a continuous function
from PI to P 2, as t varies, and this curve coincides with ~ 1 for s = and

with ~2 for s = 1, then IDl is called simply connected.

pet, s) such that for each s in an interval, say [0, 1], pet, s) describes a curve
If S is a set of points in a manifold IDl and P is a point of IDl, and if every

neighborhood of P (open set containing P), no matter how small, contains
points of S, then P is called a limit point of S. If S contains all its limit points,
it is called a closed set. The complement of a closed set is obviously an open
set, and conversely.
An alternative definition of connectedness usually given by topologists is
that a topological space is connected if it cannot be decomposed as S 1 U S2,
where S 1 and S2 are non empty disjoint (S 1 n S2 = 0) open sets, or, equiva-
lently, nonempty disjoint closed sets. Any arcwise connected space is con-
nected. (Exercise: prove this.) For a manifold, the converse is also true, so
that then the two concepts are equivalent.
A component of a manifold IDl is a maximal connected set in IDl; that is, if
Po is a given point of IDl, the set
S = {P: P can be joined to Po by a curve}

is a component ofIDl. It was seen in Section 19.5 that the manifold ofthe group
0(3) has two components.
Global Topology; Homotopic Curves; Fundamental Group 105
As a set in 9)1, a component S is both open and closed. First, it is open,

because (1) any point P of S is an interior point of 9)1 (all points of 9Jl are
interior points) and (2) a neighborhood of P that is the image of a ball in [Rn,
in some chart containing P, obviously consists of points in 9)1 that can be
joined to P; hence this entire neighborhood belongs to S. Second, S is closed,
because if P is a limit point of a sequence in S, then P can be joined to any
point lying in a neighborhood of P of the kind just described, and hence
can be joined to points of the sequence. Conversely, if a set S is connected and
is both open and closed, then it is maximal, i.e., is a component of 9)1; this
fact will be used in the theory of Lie groups.
Warning. The concept of an open set in 9)1 has nothing to do with the possible
embedding of 9)1 in a space of higher dimension. For example, if the unit
sphere x 2 + y2 = Z2 = 1 is regarded as a 2-dimensional manifold 9)1, then
a polar cap, i.e., the set of all points north of a given circle of latitude, is an
open set in 9)1 but not an open set in [R3.
Note. Any component of 9)1 is obviously itself a manifold.
23.7 Global Topology; Homotopic Curves;

Fundamental Group
In the remainder of this chapter, only connected manifolds are considered.
Two curves fo and 'i&'1 in a manifold 9)1, both having initial point Po and a
terminal point P 1, are called homotopic if one of them can be deformed into
the other by a continuous deformation in 9)1, i.e., if there is a continuous
function pet, s) (0 S; t, s S; 1) such that, for every fixed s in [0, 1], pet, s)
traces a curve from Po to P 1 as t goes from 0 to 1, and this curve coincides
with 'i&'o (or s = 0 and with 'i&'1 for s = 1. As was said in Section 19.5, a
manifold is called simply connected if any two curves in it have the same
initial and terminal points are homotopic.
It is clear that homotopy is an equivalence relation (it is reflexive, symmetric,
and transitive); hence, for given initial and terminal points Po and P 1,
the class of all curves that are homotopic to a given curve from Po to P 1
form an equivalence class or homotopy class in 9)1. In the manifold
{x, y: x 2 + y2 > 1}
consisting of the region of the x, y plane exterior to the unit disc, the first
drawing in Figure 23.5 shows three paths from P to Q belonging to the same
homotopy class, while the second drawing shows three paths, no two of
which are homotopic. In general, if'i&' denotes any curve in a manifold, ['i&']
denotes the equivalence class of all curves homotopic to 'i&'.
A partial law of composition of homotopy classes of curves is now defined.
First, if~ 1 and ~ 2 are curves going from P to Q and from Q to R, respectively,
then ~ 1 0 'i&' 2 denotes the curve that follows 'i&' 1 from P to Q and then follows
Figure 23.5 Homotopic and

p p nonhomotopic curves.
from Qto R. To put this in formulas, if the functions PI (t) (0

C(! 2 ~ t ~ 1) and
P it)(O ~ t ~ 1) describe these curves, then thefunction
P 3 (t) <!....._ef {P 1 (2t), if 0 ~ t ~ 1, (23.7-1 )

P z (2t - 1), if1 ~ t ~ 1.
describes the curve C(! 1 C(! 2' The law of composition of homotopy classes is
0
now specified by the formula
This applies when the terminal point of the curves of the first class is the same
as the initial point of the curves of the second class; otherwise, [C(! 1] 0 [C(! 2] is
undefined. It is easy to give a formal proof that the result is independent ofthe
particular curves C(! 1 and C(52 chosen from the respective classes. The "product"
[C(! 1] 0 [C(! zJ consists of all curves homotopic to the curve C(! 1 0 C(52, such as the
curve C(!~ in Figure 23.6.
This law of composition is associative but does not make the set of all
homotopy classes into a group, because the composition is not defined for all
pairs of classes, and nothing has been said about inverses. However, a group
can be obtained as follows: A fixed base point Eo is chosen, and consideration
is restricted to curves that start from Eo and return to Eo. (In the definition
of homotopy, it was not excluded that the initial and terminal points might
coincide.) The set of all homotopy classes of such curves is a group, called the
fundamental group of the manifold, and is denoted by 1[1 (m). If C(! 0 is a curve
C'3
Q
p R
Figure 23.6
Figure 23.7
that can be contracted onto the base point Bo by a continuous deformation in

9Jl (such a curve is called nullhomotopic), then "Coo "C t can be deformed into
"C I; hence ["CoJ 0 ["C IJ = ["C IJ; that is, ["CoJ is the identity element of the
group.If"C I is any curve, we denote by "C l i the same curve traced backwards;
i.e., if the function pet) (0 ~ t ~ 1) describes "C I, then the function P(l - t)
describes "C l i . Clearly, the curve "C I 0 "C l i can be contracted onto the basic
point, i.e., is nullhomotopic-see Figure 23.7; therefore
(23.7-2)
that is,
(23.7-3)
On the assumption that the manifold 9Jl is connected, the fundamental
group 11: 1 (9Jl) is independent of the choice of base point, for let BI be any
other point of 9Jl, and let "COl be any fixed curve from Bo to B I . If "C is any
curve beginning and ending at B I, then
"Col"Co"Col
is a curve beginning and ending at Bo-see Figure 23.8. The mapping
(23.7-4)
is an isomorphism of the fundamental group with BI as base point onto the
fundamental group with Bo as base point, for this mapping is obviously
one-to-one and onto, and a prodllct ["CJ 0 ["C'J = ["C 0 "C'J is mapped under
(23.7-4) onto
["COlo "C 0 "C' 0 "Con = ["COlo "CJ 0 ["C' 0 "Con
= ["Co I 0 "CJ 0 (["Co Ir I 0 ["C OIJ) 0 ["C' 0 "Con
= ["COlo "C 0 "Con 0 ["COlo "C' 0 "Col].
...,.-c
Bo--~
Figure 23.8
EXAMPLES
(1) If Wl is simply connected, then its fundamental group nl(Wl) is the trivial
group consisting of an identity element only.
(2) Let Wl be the surface of a cylinder (finite or infinite, but finite in the drawing
of Figure 23.9). Let e, z be cylindrical coordinates, and let values of e, z be represented
on a strip in the plane, as shown. Each point of Wl is multiply represented on the
strip, and in particular the base point Bo of Wl is represented by the points Bo, B' I'
B' 2, etc. A curve in the strip, such as C(j, going from Bo to any other image of Bo, say
B~, is the image of a closed curve in WI, and, conversely, every closed curve beginning
and ending at Bo in illl has such an image; furthermore, C(j can be continuously
deformed, within the strip, keeping its ends fixed, into any other curve going from
Bo to B~, such as C(j'. Consequently, for each of the possible terminal points B",
there is precisely one homotopy class of curves beginning and ending at Bo in Wl;
k is the net number of windings about the cylinder made by a curve of the class.
The composition of two such curves, say with terminal points B~ and B;, is a
curve with terminal point B"+l; hence nIOJJl) is isomorphic to the additive group
of integers, that is, to the infinite cyclic group Coo. The annulus a < x 2 + y2 < b,
the punctured plane x 2 + y2 > 0, and the Mobius strip all have fundamental
groups isomorphic to Coo.
4n A----+--B~ TERMINAL
POINT
91
2n _---If---_ B~
C c'
_z
(J to ) 0 Bo INITIAL
POINT
Bo
-2n B'_I
-4n B'-2
Figure 23.9
(3) Let Wl be the surface of the torus, which is given by the equations
z = a sin ex,
x = (A + a cos ex)cos {3,
y = (A + a cos ex)sin {3,
where x, y, and z are Cartesian coordinates, a and A are constants (A > a > 0),
and a and f3 are two angles or intrinsic coordinates on IDl. See Figure 23.10. If a and f3
are allowed to vary unrestrictedly, then the number pairs (a, f3) and (a + 2nk, f3 + 2nl)
represent the same point ofIDl. LetthebasepointBobegivenbyx = A + a,y = z = 0;
it is then represented by any of the points (a, f3) = (2nk, 2nl) in a lattice in the a, f3
plane. Any curve from (0, 0) to (2nk, 2nl) in the plane represents a closed curve be-
ginning and ending at Bo in IDl and can be continuously deformed into another curve
from (0, 0) to (2nk, 2nl). Therefore, each integer pair (k, I) determines an element
of the fundamental group n 1(!Ill). Clearly, the composition of the elements determined
by (k, l) and (k', I') is the element determined by (k + k', I + 1'); that is, the funda-
mental group of the torus is isomorphic to the direct product Coo x Coo, i.e., to the
free abelian group on two generators.
a-I
A-
1-
:=0
I
1- Figure 23.10 The 2-torus.
(4) Consider the manifold IDl consisting of the plane with two points a and b
removed. In the theory of functions of a complex variable, a contour of integration
is specified by writing an equation such as
J = f(Q+a+.b-)
f(z)dz.
Here, the expression (a +, a +, b - ) indicates that the contour starts from some base
point B (not coinciding with a or with b), encircles the point a twice positively
(counterclockwise), then encircles the point b once negatively, and then returns to B,
as in Figure 23.11. In function theory it is taken as geometrically evident that this
procedure specifies the contour adequately, if fez) is analytic except for branch
points at a and b; that is, any two contours which follow the above prescription
B
Figure 23.11 A contour in the complex plane.
can be deformed continuously into each other without crossing either branch point.
In other words, the expression (a +, a +, b - ) determines a homotopy class of curves
in IDl, i.e., an element of n I (IDl). This point of view will be accepted here. The simplest
nontrivial elemerits of the group n l (IDl) are (a + ) and (b + ) and their inverses (a - )
and (b-); these group elements will be denoted by ex, {J, ex-I, and p-I. The general
group element is of the form
where each 1\ is either ex or {J and each exponent 8; is either + 1 or - 1. In other words,

nl(IDl) is isomorphic to the free group on two generators. Unlike the groups in the
first three examples, this group is noncommutative. The same fundamental group is
obtained for the figure-of-eight region shown shaded in Figure 23.12. If n distinct
points are removed from the plane, then the reSUlting fundamental group is isomorphic
to the free group on n generators.
Figure 23.12
(5) Let IDl be the manifold of the rotation group SO(3). In Section 19.6 intrinsic
coordinates in IDl were introduced as the three components of a vector 0, which lies
in the spherical ball K = {a: 11011 :0; n} in the coordinate space. If opposite ends of
each diameter of K are identified (regarded as the same point), then there is a one-to
one correspondence between the points of IDl and those of K. A nontrivial element
of n l (IDl) is obtained by taking the base point B as the center K, and by considering
a curve C(j that goes from B along a radius to a point A on the surface, then jumps to the
antipodal point A', then returns to B, as shown in Figure 23.13. This curve C(j cannot
be collapsed onto B by a continuous deformation, because (a) any curve that starts
and ends at B and has such a jump has total length (in K) at least 2n, and (b) it is
intuitively clear on grounds of continuity that continuous deformation cannot make
the jump disappear. (This will be established more firmly in the next chapter.) Now
consider a curve C(j that starts at B and returns to B after making a finite number of
such jumps, let us say from Al to A'I' from A2 to A~, etc., where in each case the prime
denotes the antipodal point. By continuous deformation of the curve, consecutive
jumps can be made to disappear two at a time. Consider a portion of C(j that contains
A'
Figure 23.13 A nonnullhomotopic curve in the

manifold of SO(3).
Mechanical Linkages: Cartesian Products III
A'2
Figure 23.14 Homotopic curves from P to Q in

the manifold of SO(3).
two consecutive jumps, as in Figure 23.14, where it consists of the parts PAl' A~A2'
and A;Q. By moving the part A'IA2 to the surface of K and simultaneously drawing
the points A2 and A'l (also Al and A;) together, the part A~A2 can be made to
disappear, and what remains is a curve, like the dashed one, going from P to Q
without a jump. If this procedure is continued, the curve C(j can be collapsed onto the
base point B, if it had initially an even number of jumps, or onto a curve having a
single jump, if it had initially an odd number. Therefore, the group nl(SO(3 is
isomorphic to the cyclic group of order 2, consisting of just two elements. If initially C(j
had infinitely many jumps, then many of these jumps would have to be very close
together, so that the value of 11911 would remain close to n between them, and a
continuous deformation could be made to a curve that has a finite number of jumps.
These results will all be obtained more simply and rigorously in the next chapter by
means of the covering of SO(3) by SU (2).
It should be noted that a knowledge of the fundamental group does not

completely fix the global topology of a manifold. For example, the sphere
x 2 + y2 + Z2 = 1 and the disk x 2 + y2 < 1 are both simply connected
2-dimensional manifolds, but they are topologically different: If a single
point is removed, the sphere remains simply connected, while the disk does
not. For this reason, one might be tempted to introduce also the higher
homotopy groups (see Hocking and Young 1961, Chapter 4) or other
topological characterizations. However, it appears that the first homotopy
group, the fundamental group, is precisely what is needed for many purposes,
such as the theory of the covering of one manifold by another, which is the
subject of the next chapter.
It will be shown in Chapter 27 that the fundamental group of the manifold
of a Lie group is always Abelian (commutative).
23.8 Mechanical Linkages: Cartesian Products
Each configuration of the compound pendulum sketched in Figure 23.15

can be specified by the two angles IX and [3 shown. For any integers k and 1,
the number pair (IX + 2nk, [3 + 2nl) represents the same configuration as
(IX, [3). Hence, according to Example 3 in the preceding section, there is a
PRIMARY
PIVOT
SECONDARY
PIVOT
Figure 23.15 A compound pendulum.
one-to-one correspondence between the configurations of the pendulum

and the points of a torus, of such a kind that as the pendulum moves, the
corresponding point moves continuously on the torus.
If the secondary pivot is perpendicular to the primary one, as in Figure
23.16, then the tip of the secondary pendulum moves on a torus in space.
In either case, each part of the pendulum moves in a circle about its own
pivot, and the resulting motion is the combined effect of these two circular
motions. Correspondingly, the torus is regarded as the so-called Cartesian
product of two circles. Generally, let 9Jl and 9Jl' be any two manifolds, of
dimensions nand n'. Their Cartesian product, 9Jl x 9Jl', is an (n + n')-
dimensional manifold defined as follows: (1) Each point of 9Jl x 9Jl' is an
PRIMARY
PIVOT
SECONDARY
PIVOT
Figure 23.16 A compound pendulum.

Mechanical Linkages: Cartesian Products 113
ordered pair (P, P'), where P and P' are arbitrary points of 9)1 and 9)1/,
respectively. That is, 9)1 x 9)1/, as a set, is the Cartesian product of 9)1 and IDl'
in the set-theoretical sense. (2) If {U, <p, N} and {U/, <p', N/} are any charts in
9)1 and 9)1', respectively, then a chart {U", <p", N"} is defined in Wl x 9)1/,
as follows: U" is the set of all points (P, PI) such that P is in U and P' is in U/,
and <p"P, PI)) is the (n + n')-component vector consisting ofthe components
of <pep) together with those of <p/(P'), i.e.,
~/P PI)) = {<p;(P), if i = 1, 2, ... , n,

<p" ~
<P,-n (PI) , if i = n + 1, ... , n + n'.
It is evident that these definitions make 9)1 x 9)1' into a manifold.
If the pivots of the compound pendulum are replaced by idealized ball-
and-socket joints, then the configuration space is the Cartesian product of
two 2-spheres, and hence is a 4-dimensional manifold. (Rotation of the two
members about their own logitudinal axes has been neglected.)
Lastly, if a small spherical ball moves inside a hollow sphere in such a way
as to maintain contact, then the configuration space is the Cartesian product
of a sphere and the manifold of SO(3), and hence is a 5-dimensional manifold.
Clearly other examples of this kind can be constructed indefinitely.
CHAPTER 24
Covering Manifolds
Local homeomorphism; projection; p-sheeted covering; good

neighborhood; principles of lifting; universal covering manifold;
construction of mathematical models; manifolds covered by a given
manifold.
Prerequisites: Chapter 23 and parts of Chapters 18 and 19.
24.1 Definition and Examples
The 2-to-1 mapping If; ofSU(2) onto SO(3) found in Section 19.7 is more than
merely a group homomorphism. It is also a mapping of the manifold of
StJ(2) onto the manifold of SO(3) of the kind known as a covering. It is
locally a homeomorphism, in the sense that if P is any point on the manifold
of SU(2), and Q is its image in the manifold of SO(3), then P has a neighbor-
hood that is mapped homeomorphic ally by If; onto a neighorhood of Q.
Furthermore, if Q is any point of the second manifold, then there are always
two points P in the first manifold, each of which has such a neighborhood.
If; is a two-sheeted covering of SO(3) by SU(2).
A mapping If;: 9J1 ~ 91 of a manifold 9J1 to a manifold 91 (they will be called
the "upper" and "lower" manifolds, respectively) is called a covering of 91
by 9J1 if it satisfies the following two requirements, of which the first says that
all of 91 is covered, and the second tells just how it is covered: (a) If; is an onto
mapping; i.e., for each point Q in the lower manifold (91) there is at least one
point P in the upper one (9J1) such that If;(P) = Q; (b) each point Q of the
lower manifold is in some neighborhood 'n whose preimage If;-l('n)-this is
the set of all points of the upper manifold that are mapped onto points of
'n-consists of one or more disjoint neighborhoods U 1, U2 , ... ,or components
(one in each" sheet" of 9J1), each of which is homeomorphic with 'n; that is,
for each j, the mapping P ~ If;(P), restricted to Uj , is a one-to-one bicon-
tinuous mapping ofUj onto 'n. A neighborhood 'n in the lower manifold with
these properties is called a good neighborhood. (A mapping is called bi-
continuous if it and its inverse are both continuous.) If Xl, ... ,xn are co-
ordinates of P in the neighborhood Uj in 9J1, and if yl, ... ,yn are coordinates
114
Definition and Examples 115
of the corresponding point IjJ(P) in the neighborhood m in 91, then the Xi

are continuous functions of the i, and conversely, throughout the respective
neighborhoods. Obviously, m and 91 must have the same number of di-
mensions.
If such a mapping IjJ exists, m is called a covering manifold of 91, and IjJ is
called a covering of 91 by m or a projection of m onto 91.
If the manifolds are connected, then the multiplicity of the covering (i.e.,
the number of points of m that are mapped onto a single point on 91) is
constant throughout the manifolds, because this number is a positive integer
or + 00, and it is obviously constant in any neighborhood, and hence is
constant throughout m and 91. If the multiplicity is p, then IjJ is called a
p-sheeted covering. If m and 91 are of class C\ then the Xi as functions of the
Yi are required to be of class Ck ; that is, the mapping IjJ is required to be of
class C k
Note. The following I-dimensional example shows that it would not have
been equivalent to require merely that each point P of the upper manifold
9'Jl have a neighborhood that is mapped homeomorphically onto a neighbor-
hood of the lower manifold 91: Let 91 be the unit circle in a plane, and let m
be an open interval of length greater than 2n wrapped around the unit circle.
Then the points a and b of 91 lying under the ends of m (see Figure 24.1) do
not satisfy the conditions of the definition, although, since m is open, each
of its points has a neighborhood that is mapped homeomorphically into 91.
If m denotes the Riemann surface of any algebraic function F(z), with all
branch points deleted, if 91 denotes the complex plane, with the corresponding
points deleted, and if IjJ is the mapping that maps any point of m onto the
point of 91 directly underneath it (i.e., the point with the same value of z
attached), then IjJ is a covering of 91 by m. If P is any point of 91, then P has a
neighborhood mwhich is simply connected and does not contain any of the
branch points of F(z). If a right cylinder is constructed with mas base, then
this cylinder intersects each sheet of the Riemann surface in a neighborhood
U which looks exactly like m. Hence, mis a good neighborhood.
A given manifold 91 may have many different covering manifolds IDl, and
may have many different coverings by a given m. If 91 is the unit circle Iz I = 1
in the z plane, then 91 can be covered by the real line by the mapping
Figure 24.1
116 Covering Manifolds
---,
I
Z I
e:
0
()
U
Figure 24.2
x ~ z = eix (the line is to be imagined as wrapped around the circle infinitely

many times), or 91 can be covered by the circle Wi: Iwl = 1 in the w plane by
any of the mappings w ~ z = wn , where n is any integer not =0 (the circle Wi
is to be imagined as stretched and wrapped n times around the circle 91).
Consider the mapping 1/1 of a plane onto a cylinder given by setting z
equal to x and e equal to y, where x, yare Cartesian coordinates in the plane,
and z, e are cylindrical coordinates on the cylinder. This mapping is many-
to-one, because the points (x, y), (x, y 2n), (x, y 4n), etc. are all mapped
onto the same point of the cylinder. The pre image of a neighborhood, which,
like U in Figure 24.2 does not encircle the cylinder, consists of an infinite
succession of neighborhoods Uj in the plane, obtained by displacing anyone
of them horizontally through distances 2n, 4n, etc., as in Figure 24.3;
each Uj is homeomorphic to U; hence, U is a good neighborhood. On the
other hand, a neighborhood like 5B in Figure 24.2, which is a band encircling
the cylinder, is not a good neighborhood, because its preimage is an infinite
strip in the plane and is mapped many-to-one onto 5B under 1/1.
EXERCISE
Discuss the various possible coverings of a torus by a plane, by a cylinder, and by

another torus.
() () ( ) etc.
U'_I U'a U'I
Figure 24.3
Principles of Lifting 117
If ID1 and 91 are C k manifolds, then the mapping t/J is required to be of class
C k ; that is, if t/J maps Ponto Q = t/J(P), if IB is a good neighborhood of Q, if
U is the component of t/J - 1(1B) containing P, and if Xl, ... , Xn are coordinates
of P in U, while y1, ... , yn are coordinates of Q in IB, then, as P varies, the
Xi are functions of class Ck of the /, and conversely. (In nearly all cases of
interest, ID1 and 91 are analytic manifolds, and these functions are analytic.)
If the covering t/J is a one-to-one mapping, so that each of ID1 and 91 is a
covering of the other, then t/J is called a (C k ) homeomorphism and the manifolds
are called homeomorphic; they are topologically indistinguishable. In the case
k = 00, a homeomorphism is sometimes called a diffeomorphism.
24.2 Principles of Lifting
If the covering manifold ID1 is connected, then 91 is necessarily also connected.

The converse is, of course, not true; however, only connected manifolds will
be considered, from now on. If ID1 is simply connected (like the line and the
plane the forgoing examples), then, according to the theorem below, ID1
represents the most that can be achieved, in a sense, by way of covering the
given manifold 91 with a connected manifold. Two lemmas are needed.
Lemma 1 (First Principle of Lifting). Suppose that the manifold ID11

covers the manifold ID10 by the projection t/J. Let Bo be an arbitrary base point
ofID1o. Of all the points P in ID1! such that t/J(P) = B o , choose one and call
it the base point B! of ID1!. (See Figure 24.5 in the next section, where,
however, a third manifold has been added, in connection with the theorem
below.) Let ~o be a curve in ID1 o , described by a continuous function poet)
(0 :s;; t :s;; 1) starting from the base point Bo. Then, there is a unique curve
~ 1: P! (t) in ID11 such that PI (0) = BJ, and t/J(P 1 (t = P oCt). One says that
the curve ~ 0 has been lifted up to ID11 from ID1o.
PROOF. As noted under the definition of a covering, a neighborhood U in Wl o is

called a good neighborhood if each component of 1jI-I(U) is homeomorphic to U
under 1jI. If I is any interval contained in [0, 1], the symbol P 0(1) denotes the segment
of the curve '6'0 given by
A partition of [0, 1] into closed intervals [0, t l ], [tl' t 2], ... , [tN-I' 1] is called a
good partition if each segment Po([t j , t j + 1 ]) of the curve '6'0 lies in a good neighbor-
hood. A good partition exists, because each point of '6'0 is in a good neighborhood;
hence each t in [0, 1] is in an open interval I such that the curve segment P 0(1) is
in a good neighborhood. These open intervals cover [0, 1]; by the Heine-Borel
theorem, a finite number of them cover [0, 1]; if these are arranged in order of
increasing t, then t 1 can be chosen in the intersection of the first and second of these
neighborhoods, t2 in the intersection of the second and third, etc. See Figure 24.4.
Thus, a good partition is obtained. Call Uj the good neighborhood in which the curve
segment P o([t j , t j + 1]) lies. For eachj = 0, 1, ... , N - 1, a segment P 1([tj , t j + 1]) of
Figure 24.4
the curve ~I = {PI(t):O:::; t:::; I} in the upper manifold Wli is now defined, in-
ductively, as follows: First let 5!l0 be the component of IjJ-I(UO) that contains the
base point BI of Wl 1, and define PI ([0, t l ]) to be ~-I(P 0[0, t I]), where ~ is the
restriction of IjJ to 5!l0; ~ is a homeomorphism of 5!l0 onto Uo; hence PI ([0, t I]),
thus defined, is a curve segment in Wl i . Now suppose PI([tj , t j + I]) has been defined;
since Po(tj+ I) lies in Uj+ I as well as in Uj , 5!lj+ 1 can be taken as the component of
IjJ-I(Uj + I) that contains the endpoint P I(t j + 1) of the previously defined segment of
'!&'t; then the segment PI([tj + l , t j +2]) is defined as ~-I(PO([tj+I' t j + 2]), where now
~ is the restriction of IjJ to 5!l j+ l' In this way, the curve'!&' j in Wli is constructed. It is
uniquely determined by the curve'!&'o in the lower manifold and the choice of the base
point B I in the upper one; in particular, it is independent of the choice of the good
partition of [0, 1], because any two partitions have a common refinement, and'!&'l
is obviously not altered by refining the partition used (i.e., by adding further points
of subdivision of [0, 1]). Each of the curves <go and '!&'j uniquely determines the other.
Lemma 2 (Second Principle of Lifting). Under the conditions ofLemma 1,

if p o(t, s) is a continuous function of two variables in 9R o , defined in the
square 0 :s; t, s = 1, if p o(t, so) coincides with the base piont B o , for some to
and so, then there is a unique continuous function PI (t, s) in 9R 1 such that
(1) tJ;(P 1 (t, s)) = P oCt, s) and (2) PI (to, so) is the base point B 1 of9R 1
The proof is similar to that of Lemma 1. The Heine-Borel theorem implies

the existence of a finite collection of open sets in the t, s plane, which covers the
square 0 :s; t, s :s; 1, and such that each set determines a good neighborhood.
Let these sets be arranged in a sequence such that each one contains points in
common with at least one of the preceding sets. From here on the argument
is as before.
Corollary. If two curves C(/O and (If?~ in the lower manifold, both runningfrom
Bo to some point A o , are homotopic, i.e., if one of them can be deformed in 9R o
continuously into the other, keeping the endpoints fixed, then the curves that
result from lifting them up to 9R 1 also have a common endpoint A 1 and are
homotopic in 9R 1 .

SKETCH OF THE PROOF. Let poet, s) in the lemma be such that, for each fixed s

in [0, 1], poet, s) traces a curve from Bo to Ao. as t increases from to 1. and such
that for s = this curve is '!&' 0 while for s = 1 it is '!&'o; then use a continuity argument.
based on a good neighborhood of Ao. to show that the terminal point of the lifted
curve, P 1(1, s) cannot jump from one sheet of Wl j to another, as s varies.
Universal Covering Manifold 119
24.3 Universal Covering Manifold

In Section 24.5, it will be proved that, if Wl o is any manifold, then there is a
simply connected manifold that covers Wl o . It is called the universal covering
manifold of Wl o , because of the theorem below, which says (1) that the
universal covering manifold of a given Wl o is unique (up to homeomorphism),
and (2) that it also covers any other manifold that covers Wl o . The proof of
existence is postponed until Section 24.5 because it is a little more abstruse
than the proof of the present theorem.
Theorem. If Wl j and Wl2 are connected covering manifolds of Wl o , and if

Wl2 is simply connected, then Wl 2 is a covering manifold of mi' If Wl j is also
simply connected, then Wl2 and Wll are homeomorphic, i.e., topologically
indistinguishable.
PROOF. Let l/110 and l/1zo denote projections of Wl z and Wl I, respectively, onto
Wl o . A projection l/1ZI from Wl z to WlI will be so constructed that l/110l/1ZI = l/1zo.
Choose base points B z , B I, and Bo in the three manifolds such that l/1zo(Bz) =
l/11O(B I ) = Bo (see Figure 24-5). To describe l/1ZI' we must specify a point l/1ZI(QZ)
in WI I for each point Q2 in 9')1z; we do it as follows: Let 'Il z be a curve Pz(t)
in 9')1z from B z to Q2' namely, a curve such that Pz(O) = B 2 , Pz(l) = Q2' Then the
projection of'll z onto Wl o is a curve poet) = l/1zo(P z(t)) from Bo to Qo. (Comment: 'Il 0
may be self-intersecting, even if 'Il z is not. For example, in Figure 24.6, 'Il z is a curve
in a plane, and 'Il 0 is the result of wrapping the plane around a cylinder.) The mapping
l/1jol is, of course, generally multi valued, but according to the lifting principle of
Lemma 1, there is a unique curve 'Ill: P = P let) in Wl 1, starting from the basepoint
BI and such that l/110('Il 1) = 'Iloltisnow asserted that the terminal point Ql = PI(l)
of'll I is uniquely determined by the terminal point Qz of'll z in the upper manifold Wl z .
To prove this, let 'Il~ be any other curve in Wl z from B z to Qz. Since Wl z is simply
connected, 'Il z and 'Il'z are homotopic-either can be deformed into the other by a
continuous deformation in Wl z . Therefore, their images 'Il 0 and 'Il~ in Wl o are homo-
topic, and, by the corollary to Lemma 2, the curves 'Il 1 and 'Il'1 that result from lifting
'Il 0 and 'Il~ up to Wll have a common terminal point Q I, which is now defined to be
l/1Z1(QZ)' Since Qz was arbitrary, l/1Z1 is defined in all Wl z , and
EXERCISE
Complete the proof by showing (a) that l/1Z1 is an onto mapping, (b) that any Ql
in WlI has a neighborhood msuch that each component of l/1i./( m) is homeomorphic
to munder l/1z j, and (c) that if9'Jl 1 is also simply connected then Ql uniquely determines
Qz, so that l/1Z1 is one-to-one.
Note. The statement that two manifolds are homeomorphic says nothing
about how they might look if they are embedded in some Euclidean space of
higher dimension. A circle in the plane, as a one-dimensional manifold, is
homeomorphic to a simple knot in space; a simple loop of paper is not
homeomorphic to a Mobius band, but it is homeomorphic to a loop of paper
that has two half twists (one end was twisted through a full turn before being
ffil2
B' 0 .. - - .. ,....
1 ...... Q
... 1
''''--_OQl
Figure 24.5 Covering manifolds-see text.
Figure 24.6
Comments on the Construction of Mathematical Models 121
glued to the other end) or any even number of half twists, and a Mobius band
is homeomorphic to such a loop with any odd number ofhalftwists. This can
be seen by considering the method of identification of edges, discussed in
Section 23.1, for constructing these manifolds.
Because of this theorem, a simply connected covering manifold of 91 is

called the universal covering manifold of 91. In the Section 24.5, below, it is
shown that any manifold 91 has a universal covering manifold. This is
important in the theory of Lie groups and for the cosmological interpretation
of Einstein manifolds. Since the construction of the universal covering
manifold is somewhat abstruse, a few comments on constructions in general
are in order.
24.4 Comments on the Construction of

Mathematical Models
The extensive use of construction by mathematicians sometimes seems un-
natural and artificial at first sight. To a physicist, for whom a real number is,
for example, an instantaneous coordinate x = x(t) of a moving point, it
seems wholly inappropriate to regard this number as an infinite equivalence
class of Cauchy sequences ofrational numbers (especially when it is recalled
that each rational number is an infinite equivalence class of ordered pairs of
integers, etc.) However, once the construction has been completed and the
properties of the resulting system have been found, one can then completely
forget the details of the construction, and one can regard a real number as a
thing just as atomic and indivisible as a "point" was in Euclidean geometry.
The construction of mathematical models is now beginning to playa role in
quantum theory and relativity, and for the same reason, basically, for which
it plays a role in mathematics; when a structure is defined by writing down a
set of axioms, one has to prove that the axioms are not mutually contra-
dictory, and the best way to do that is to exhibit a model of the structure based
on simpler axioms that have already been accepted. One of the earliest
instances in mathematics dealt with complex numbers. To some mathe-
maticians it seemed unnatural and dangerous to assert the existence of a
number i such that i2 = -1, and there was much argument about whether
such a number "could exist." The question was settled (by Gauss in 1831
and independently by Hamilton in 1837) by considering ordered pairs (a, b)
of real numbers and by defining the arithmetic operations on those pairs by
(a, b) + (c, d) = (a + c, b + d),
(a, b)(c, d) = (ac - bd, be + ad),
whereupon the system of all ordered pairs acquired exactly the same properties
as the system of numbers a + bi would possess, if the number i existed. One
then felt free to assert the existence of the complex number system, and it
became irrelevant whether one wrote (a, b) or a + bi.
In the development of the relativistic wave equation for the electron,

Dirac encountered the need for four quantities 0(1' 0(2' 0(3, 0(4 which should
multiply according to the rules
U, k = 1, ... ,4). (24.4-1)
Instead of simply postulating the existence of the 0(j and taking the equations
(24.4-1) as axioms, Dirac exhibited four 4 x 4 matrices which multiply
precisely according to (24.4-1) (when the right member is interpreted as
2b jk times the unit matrix).
The theory of Lie algebras starts with abstract definitions and axioms.
After an exceedingly long and complicated development, containing many
difficult lemmas and theorems, one arrives at a classification of the so-called
simple complex Lie algebras, in which the possible algebras! have been
narrowed down to nine types or classes. Two questions then arise: (1) Were
the original axioms free from contradictions? (2) Is it possible that further
work would narrow down the possibilities still further and eliminate some
of the nine types? Both questions have been answered by constructing a
mathematical model of every algebra in each of the nine classes, without
using any axioms other than those of ordinary arithmetic. Some of the con-
structions are rather complicated and arbitrary, but they serve the purpose.
See Hausner and Schwartz 1968.
As a further example, the question of the possible "existence" of certain
systems of non-Euclidean geometry (i.e., based on axioms different from those
of Euclid) was settled by constructing models, without using any new axioms.
In these models, one constructed certain things that were arbitrarily called
"points," one defined what was meant by the" distance" along a curve between
two points, one determined a "straight line" in terms of the shortest curve
between two points, etc., and in the end, one found that these "objects"
satisfy all the axioms of Euclid except that through a point Q not on a straight
line L there are many different straight lines L', L", etc., parallel to L (or with a
different construction that there are none at all). In this way it was proved that
Euclid's parallel axiom could be modified without producing contradictions.
See Chapters 26-28.
Distributions themselves are constructions. Dirac postulated the existence
of something, called b(x - x o), which was supposed to behave like a function
in most respects and to have certain specified properties. The functional
< b, ~ <p(x o), when suitably interpreted, has precisely the required
properties.
The mathematical model of the universal covering manifold IDl of a given
manifold 91 is constructed in the next section by the method used in general
relativity; see Note in Section 23.4. The space IDl is not known in advance as a
topological space or even as a collection of points. Instead, one has a collec-
tion of charts and specifies how they are to be pieced together to form IDl.
According to the Whitney embedding theorem, which will not be proved here,
an n-dimensional manifold such as IDl (abstract or otherwise) is homeo-
morphic to an n-dimensional surface in some Euclidean space EN of higher
Construction of the Universal Covering 123
dimension; this surface thus provides an alternative mathematical model of

9R, once the first model has been constructed.
24.5 Construction of the Universal Covering
We are given a connected manifold 9l of class Ct, and we wish to construct a

simply connected manifold 9R that covers 9l. We assume that the charts
K, L, ... of9l are simply connected. We choose a base point Bo in 9l, and for
any chart K we denote by 0(, {3, .. the homotopy classes of paths from Bo
to K (the terminal point of those paths can be taken as any point of K,
because K is simply connected). We construct duplicates of the chart K,
called K rz , K p ,"" one for each such homotopy class. See Figure 24.7.Jhen
the charts K rz , K p ,"" L(, L", ... , ... are to be pieced together to form the
manifold 9R, and we now specify how those charts overlap. Given any two of
them, say Krz and L", we specify that there is no overlap unless K and L
overlap in 9l. If K and L overlap, we can assume that the paths of 0( and 11
have a common terminal point in the overlap of K and L; then we specify
that there is no overlap of Krz and L" unless the paths of 0( are homotopic to
those of 11, and in that case we write 0( ' " 11 and we specify that Krz and L"
have the same overlap as K and L. In more detail, let
K = {U, cp, N}, L = {U', cp', N'}. (24.5-1)
Then, for each 0(, Krz is a copy of K distinguished from K and from other
copies only by carrying "0(" along as a distinguishing index. We write
L" = {U~, cp', N'}, (24.5-2)
where Urz and U~ are parts of the manifold 9R under construction, and are
defined as follows: We let each point x in the region N of the coordinate space
Figure 24.7 Construction of the universal covering manifold.

[Rndetermine a point p ofUa with coordinates cpj(p) ~ xi, where the xj are the
components of x. m is made up of the points determined in this way by all
the charts K a , K p, ... , L(, L~, ... , ... ; they are all distinct points of m except
for the identifications to be made as we specify the overlap of the charts.
The overlap of K and L in ill is described, according to (24.5-1), by
(24.5-3)
which gives a one-to-one mapping from part of N onto a part of N' and hence
gives two coordinate systems in the region U n U' of m. If a ~ '1, as above,
we specify that the overlap of Ka and L~ is given by the same equations
(24.5-3), and we identify the point of Ua having given coordinates Xl, ... ,xn
with the point ofU~ having corresponding coordinates X'l, ... , xln determined
by those equations.
Clearly, this procedure determines a manifold m of the same differentia-
bility class C k as m. The projection tjJ of m onto ill is easily defined by project-
ing each Ka onto the corresponding K: Each point of Ua in mis projected onto
the point of U in mhaving the same coordinates Xl, ... ,xn. This projection tjJ
is of class C k because in these coordinates it is just the identity mapping.
Lastly, to show that m is simply connected, we first choose a base point Ao
in m as one of the points that lie over the base point Bo of m, as follows: Bo
lies in some chart in m, say L. Then the paths of the homotopy classes (, '1, ...
can be taken as closed paths beginning and ending at Bo. One of the classes,
say '1, is the class of nullhomotopic paths-paths that can be shrunk con-
tinuously in mto the point Bo. Let Ao be the point of L~ that lies over B o ,
i.e., has the same coordinates Xl, ... , xn in L~ that Bo has in L.
Given any chart Ka in m, we choose one of the paths in the class a in mand
lift it, according to the first principle of lifting in Section 24.2, up to m as a
unique path, which we call a', from the new base point Ao to a point in Ka.
Then, whenever charts Ka and L( overlap, we have a ~ (; hence by the second
principle oflifting, the paths a' and" are homotopic in m, or, more precisely,
they become homotopic if they are so chosen as to have a common terminal
point in the intersection of Ka and L,.
Now let~: peA), 0 ~ A ~ 1, be any path in m; we wish to show that it is
homotopic to any other path that also goes from P(O) to P(1), i.e., that m
is simply connected. Each point Pea) of ~ lies in some chart, i.e., peA) lies in
that chart for A in some interval (a - , a + c). By the Heine-Borel theorem,
Figure 24.8
Manifolds Covered by a Given Manifold 125
Figure 24.9
then, we can find a finite subdivision of [0, 1J as

= ..1.0 < ..1.1 < ... < Ak = 1
such that peA) lies in some chart, say L aj , f6r A in [Aj' Aj + 1J, for each
j = 0, 1, ... , k - 1. We call Cfi j the part ofCfi for Aj :.:::: A :.:::: Aj+ 1; Cfi j lies in L aj
For eachj, we let ctj be a path in mi, as in the preceding paragraph, from the
base point Ao to P(A,;). Since P(Aj) and P(A j + 1) both lie in La}' it follows that
ct] C{;j is homotopic to ct]+ l ' See Figure 24.8. Therefore,
0
ct~ 0 Cfi 0 0 Cfi 1 0 0 Cfik _ 1 ~ ct~.
See Figure 24.9. The right member here depends only on the initial and
terminal points P(O) and P(1) of Cfi; hence any other path from P(o) to P(l)
is homotopic to Cfi, as required.
Summary. (Main Theorem). Every (connected) n-dimensional manifold

91 has a universal covering manifold 9]( (n-dimensional), i.e., a simply con-
nected covering manifold. 9]( covers any manifold that covers 91, and the
only simply connected manifolds that cover 9]( are homeomorphic to 9](.
A model of9]( has been constructed by the procedure given above.
24.6 Manifolds Covered by a Given Manifold
A problem inverse to that of finding the universal covering manifold 9]( of a

given manifold 91 is now considered: Given a manifold 9](, construct the
manifolds 91 that can be covered by 9](. The procedure used is the identification
of sets of points in 9](. A few examples are given first.
Let 9]( be the x, y plane, and let it cover the unit cylinder Z by being wrapped
tightly about Z, the axis of Z lying in the y direction. We know, of course,
that the plane then covers the cylinder, but we show how to establish that fact
a priori. We know that, for given (x, y), all of the points (x + 2nl, y), I = 0,
1, 2, ... , of the plane coincide with a single point of the cylinder. There-
fore, a manifold 91 homeomorphic to the cylinder Z can be constructed by
defining each set {(x + 2rcl, y): 1= 0, 1, ... { ~ 1jJ((x, y to be a "point"
of 91 and by defining charts in 91 in an obvious way. The mapping (x, y) ~
1jJ((x, y is then a projection of Wl onto 91. One says that all the points
(x + 2rcl, y) of each set have been identified (i.e., made identical). Note that the
factor 2rc is irrelevant, because only the topological properties of 91 are in-
volved. Identification of the points (x + n, y) or, more generally, of the points
(x + an, y), where a is any nonzero real number, would have the same effect.
Similarly, if, for each x, y in Wl, all the points of the form (x + I, y + m),
where I and m run independently over 0, 1, 2, ... , are identified, then the
resulting manifold 91 is the torus (more properly, is homeomorphic to the
torus).
Let Wl be the infinite strip -1 < x < 1, - 00 < y < 00. For given x, y in Wl,
let the points (( -lYx, y + I), I = 0, 1, 2, ... , be identified. The resulting
manifold 91 is the Mobius band.
To generalize from these examples, let Wl be any (connected) manifold.
Suppose that (J is a homeomorphism (of class ck, ifWl is a C k manifold) ofWl
onto itself. Denote by (Jl the lth iterate of (J, i.e.,
(JI(p) = (J(o{ .. (J(P) . ..
~
,
I repetitions
and denote by (J-l the lth iterate of the inverse mapping (J-1. For any point
P in Wl, consider the point set
IjJ(P) ~ {(JI(p): 1= 0, 1, ... }. (24.6-1)
[In the first example, (J is the displacement (x, y) ~ (x + 2rc, y) in the plane.]
Suppose further that (J is such that the point set IjJ(P) is discrete in Wl, for
every P; that is, suppose that there is a neighborhood of P that contains none
of the other points (JI(p) with I =f. 0. Then, since (J is a homeomorphism, each
~(P) has a neighborhood that contains no (JI(p) with I =f. k.
Under these assumptions, a manifold 91 whose" points" are the sets IjJ(P)
is now constructed:
91 = {1jJ(P): P E Wl}.
To define charts in 91, let {U, <p, N} be a chart in Wl. Assume that U is small
enough so that for no P in U is (J(P) also in U. (If this is not the case, replace
the chart by a suitable subchart.) A chart {U, q" N} is then defined in 91
as follows:
U = {IjJ(P): P in U}.
For P in U, q,(IjJ(P ~f <pep),
hence,
N=N.
It is left as a quite obvious exercise to show that (a) q, is one-to-one, (b) charts
defined in this way are pairwise compatible and cover 91, (c) the mapping
Manifolds Covered by a Given Manifold 127
P ~ ljJ(P) is onto 91, and (d) any neighborhood U of the kind defined above
is a good neighborhood of each of its points, because the components of
ljJ - l(U) are the sets O"I(U), I = 0, 1, .... The conclusion is that ljJ is a covering
of 91 by m.
It will now be shown that every covering of a manifold 91 by a manifold m
is associated with a group of homeomorphisms in m of the kind described
above.
Let m and 91 be connected n-dimensional manifolds, and assume that m
covers 91 by a projection ljJ (which is assumed not to be one-to-ohe, so that
the covering is not trivial). Let B1 and Bo be basepoints in m and 91, where
B1 lies over B o , i.e., ljJ(B1) = Bo. We wish to show that, corresponding to
each other point B'l of m that lies over B o , there is a homeomorphism 0" of
m that carries B 1 into B'l.
The proof is quite simple if m is simply connected, and that case will be
discussed first. Let C(j 1 be a curve in m from B 1 to B~, and let C(j 0 be its image
in 91; C(j 0 is a closed curve beginning and ending at Bo; C(j 1 and C(j 0 will be
kept fixed during the discussion, and by means of them a homeomorphism
0" ofm onto 91 will be constructed, under which B1 goes into B'l. Let
(0 :::;; t :::;; 1),

be a curve in m going from B1 to any point P 1(l) = AI; the image of Al
under 0" will be constructed. Let Po be the image of the curve PI in 91 with
terminal point Ao = ljJ(A 1). Let P'l be the result of lifting the curve Co 0 Po
up to m. Then P'l(l) is a point, which, like AI' lies over Ao. A mapping in m
is defined by
0": Al ~ A'l = P~(l).
It is well defined, because if the same construction starts with any other curve
Q1 in m from B1 to AI' then Ql is homotopic to PI (because m is simply
connected); hence Coo Qo is homotopic to C(j 0 0 Po, and, hence, by the
corollary to the second principle of lifting, P'l is homotopic to Q'l, and the
same point A~ is obtained as the image of Al under 0". Furthermore, 0" is
one-to-one, because COl exists. It is clear that 0" is continuous, because a
small displacement of A 1 can be achieved by a small change of the curve P b
hence by a small change of the curve P'l, thus producing a small displace-
ment of A'l. Therefore, 0" is a homeomorphism ofm.
If m is not simply connected, we can no longer assert that the curves PI
and Q1 from B1 to Al in m are homotopic. Let the terminal point Q'l(l) of
Q~ be called A~; we wish to show that A'{ = A'b so that the mapping is still
well defined. The curves r; 1 P~ and QI1 Q'l in m, which go from Al to A~
and from Al to A~, respectively, are mapped under ljJ onto the curves
POloC o oPO and Q01oC o oQo,
But these are closed curves in 91 beginning and ending at A o , and they are
both homotopic to Co, and it follows from the corollary to the second
principle oflifting that p I l P'l and QI1 Q'l are homotopic, and hence that
a is well defined. The remainder of the discussion is the same as in the

preceding case.
If B'I is taken in succession as each of the points in 9Jl lying over Bo (in-
cluding B I itself) a group of homeomorphisms of 9Jl is obtained. The action
of this group on 9Jl is such that, for any AI, the images {a"(AI): all a} are
discrete. To see that, let U be a good neighborhood of Ao = ljI(AI). The
points a(AI) all lie over Ao, but since ljI is one-to-one in each component U'
of ljI-I(U), there can be at most one point lying over A o in each such com-
ponent.
The problem of finding all manifolds covered by a given manifold is
thereby reduced to the problem of finding all homeomorphisms a of the kind
described above. This idea has applications to the theory of general relativity;
see Chapter 30.
CHAPTER 25
Lie Groups
Lie Group G; linear Lie group; tangent vector; Lie algebra Aof G; Lie product;
Jacobi identity; abstract Lie algebra; structure constants; local isomorphism
ofSU(2) and 50(3); exponential mapping of A into G; logarithmic (or normal)
coordinates in G; adjoint representations of Lie algebras and simply
connected Lie groups; the Campbell-Baker-Hausdorff formula; translation
of charts; ideals; simple Lie algebra; local and global homomorphisms of
groups; homomorphism theory; center of a group; center of an algebra;
covering group; direct and semidirect sums of Lie algebras; classifications of
the simple Lie algebra.
Prerequisites: Chapters 18, 19, 23, 24; Sections 21.1-21.4.
The subject ofthis chapter is the advanced theory of continuous groups, often
called, inaccurately, the theory of Lie groups. Most of the groups themselves
play a role in physics and mathematics at a more elementary level. Among them
are the rotation and rigid motion groups, the Lorentz and Poincare groups,
and the unitary and symplectic groups. What is new is the study of the groups
and the structure they comprise from a deeper analytic, algebraic, and
topological point of view. The key to the study is the theory of Lie algebras
and of the interaction between the groups and their algebras. That inter-
action has played a role in quantum mechanics from the beginning, in that the
elements of the Lie algebras have appeared as operators derived from the
symmetries of a system. In the last 25 years much of the terminology and
certain specific groups, such as the groups derived from the Lie algebra G2 ,
have appeared in particle physics. So far, the applications to particle physics
has been mostly heuristic, but it seems likely that as the physical theory
becomes fully developed, the details of the mathematical structure will be of
greater importance. Most presentations of the theory are quite recondite
and hence rather difficult for the nonspecialist. I have attempted to present
the subject in the most elementary way possible consistent with describing
the complete structure. For instance, a vector field on the group manifold
is defined as consisting of components subject to a transformation law, just
as elsewhere in physics, rather than as an abstract mapping (derivation) in an
algebra of Coo functions.
129
130 Lie Groups
25.1 Definition and Statement of Objectives

A Lie group is a group G whose elements g, h, . .. can be regarded as the points
of a manifold, in such a way that the group-theoretical properties of the
elements vary continuously in the manifold. That is, if g, h, and gh are in
charts {U, <p, N}, {U', <p', N'}, and {U", <p", N"} (these charts are not neces-
sarily distinct), then the coordinates of gh must be continuous functions of the
coordinates of g and h, i.e., the components of <p"(gh) must depend con-
tinuously on the components of <peg) and <p'(h); also, the coordinates of g-1
must be continuous functions ofthe coordinates of g. If G is a group of matrices,
like SU(2) or SO(3), and if the manifold is defined as in Section 19.5, then the
continuous dependence just described is automatic, because the matrix
elements of AB, and of A -1 (for A nonsingular), depend continuously on the
matrix elements of A and B, and of A, respectively. If G is an abstract group,
the continuous dependence must be postulated.
The formal definition can be given in many ways, because a very few basic
properties imply many other properties. For example, the group is often
postulated to be a COO manifold, but Hilbert conjectured in 1900 that it was
only necessary to require it to be a CO manifold, and then it would auto-
matically be a COO manifold; this conjecture was verified in 1952 by Gleason
and independently by Montgommery and Zippin. They showed that the
manifold of every Lie group is actually a real analytic manifold. Furthermore,
it is only necessary to postulate one coordinate chart on the group, namely
one located in a small neighborhood of the identity element and in which gh
and g-1 are continuous; the rest of the manifold structure then follows from
the group axioms. The definition given in this book postulates only what is
necessary for deriving the remaining properties conveniently by elementary
methods.
The familiar Lie groups, including all the ones (to the best of my knowledge)
that have ever appeared in applications, are linear Lie groups, i.e., are
isomorphic to groups of linear transformations in a finite-dimensional
space, or, equivalently, groups of matrices. This applies even to the ones that
originate as groups of nonlinear transformations. For example, the group of
Mobius transformations
az + fJ (a(5 - yfJ oF 0)
+ (5
z~---
yz
in the complex plane is isomorphic to the restricted Lorentz group, which
is linear. Furthermore, much of the theory is simpler for groups of matrices
than for abstract Lie groups; for example, the exponential of a matrix,
exp M, is familiar and elementary, whereas the corresponding construction
for the Lie algebra of an abstract Lie group requires the machinery de-
veloped in the next five sections. For physical applications, therefore, a
theory of the linear groups alone would seem to be desirable. However, the
abstract formulation appears necessary for a complete theory. Even if one
starts with matrices, the theory leads to groups that are not, at least in any
Definition and Statement of Objectives 131
obvious way, groups of matrices, namely quotient groups G/H and semi direct
products. Every compact Lie group is linear, but the proof depends on quite
advanced developments in the theory. See Chevalley 1946. Two nonlinear
Lie groups are described in the Appendix to this chapter. The abstract
theory is presented below, but the specialization to matrices is mentioned at
various points; see Exercises 1-7 in Section 25.14.
Let G be a group. Suppose that, in the space whose points are the elements
of G, there is defined an n-dimensional coordinate chart {U, <p, N} such that
U contains the identity element 1 of the group. (The symbol "1" is used
because" e" is needed for exponentiation). It is assumed, for convenience,
that <p maps 1 onto the origin of [Rn: <p(l) = O. A subset Uo ofU is called open
(as in Chapter 23) if <p(U o) is an open subset of N in [Rn.
We assume that products and inverses of group elements are continuous
in this chart insofar as their coordinates are defined. It then follows that we
can define a smaller chart, with special properties, as follows: Let 9 and h be in
U. If g and h are close enough to 1, that is, if <peg) and <p(h) are close enough
to the origin in [Rn, then gh, g - \ and h - I are also close to 1. In particular, if
g = h = 1, then gh, g- 1, and h- I are = 1 and their coordinates are defined
and are all zero. Hence, by continuity there is a neighborhood UI of 1 such
that if g and h are in U I , the coordinates of gh, g, and h are defined and are in
the open set N of [Rn. It is convenient to consider an even smaller neighborhood
U o = U l (\ U t \ where Uti ~ {g-l: g E Ud, and to call No = (j)(U o) c N.
Then, if g and h are in U o , gh is in U, while g - I and h - I are in Uo. A vector-
valued function m(xI' x z) is therefore defined, for all Xl and Xz in No, by
m(<p(g), <p(h = <p(gh) E N;
similarly, I(x) is defined by
The group G, together with the n-dimensional chart {U, <p, N}, will be called
an n-dimensional Lie group if the functions m(, .) and 1() are defined in an
open set No, as described above, and are of class C 4 . Later, further charts will
be obtained from {U, <p, N}, by use of the group operations, in such a way as
to make G into a manifold.
All the groups described in Chapter 19 are Lie groups when coordinate
charts are suitably defined in them.
(Some authors require the manifold of a Lie group to be connected; for
reasons given in Section 25.11, that requirement is irrelevant.)
For example, let G be the rotation group SO(3) with the intrinsic coordinate
{}x, {}Y' {}z discussed in Section 19.6. Then U can be taken as the set of all group
elements for which 11011 < n (i.e., all for which 11011 i= n), and Uo as the set
for which 11011 < n/2. Hence N is the interior of the ball K in [R3 described in
section, and No is the open ball of half the radius of K. The same coordinates
can be used for 0(3); in that case the entire second component of the mani-
fold is outside U.
132 Lie Groups
To derive the properties of Lie groups from the above definitions, the Lie
algebra A = A( G) of a Lie group G is constructed; A is an n-dimensional
linear space of elements A., JI, ... , in which a multiplicative operation [A., JlJ,
the so-called Lie product, is defined. The structure of A is completely de-
termined by the properties of G in any arbitrarily small neighborhood of 1 ;
on the other hand, A completely determines many of the properties of G.
Then, the so-called exponential mapping from A into G is constructed; it
generalizes the mapping M -+ eM for matrices. In some neighborhood of the
origin of A, the mapping is one-to-one, and the components of an element A.
serve, via the inverse mapping, as the so-called logarithmic coordinates in G.
Other coordinate charts are later obtained from this one by translations in G
and are related to it analytically. An expliCit formula (the CBH formula-
see Sec'tion 25.10) then gives m(A., JI) in terms of A. and JI and shows that the
dependence of gh on g and h is analytic in these coordinates. The formula
depends only on the structure of the Lie algebra, which shows that, in a
neighborhood of 1 where the logarithmic coordinates are defined, the
structure of the group depends entirely on its infinitesimal group elements.
The study of Lie groups combines analysis, algebra, and topology in
almost equal proportions. The application of powerful techniques of linear
algebra yields a complete classification of Lie algebras, from which follows a
classification of Lie groups. That may seem surprising in view of the origin
of Lie groups often as groups of nonlinear transformations-see Eisenhart
1933.
25.2 The Expansions ofm(, .) and 1()
Since the functions m(x, y) and I(x) are of class C 4 , they can be expanded in
Taylor's series in the components Xi and i of x and y, through 3rd order
terms, with remainder terms of order 4. The group relation a1 = 1a = a for
arbitrary a shows that
m(x, 0) == m(O, x) == x (25.2-1)
[it is recalled that .:p(1) = 0]. Therefore, in the expansion ofm(x, y) about the
origin, the constant term vanishes, the linear part is x + y, and the quadratic
part contains terms like xiY', but not terms like XiXk or yil; that is,
(25.2-2)
where the a's, b's, and c's are coefficients and where the summation con-
vention has been used.
The associativity axiom of group theory imposes on m( " . ) the restriction
m(m(x, y), z) = m(x, m(y, z)). (25.2-3)
If only the linear and quadratic terms of the expansion ofm(, -) are included
then (25.2-3) is satisfied automatically; nevertheless, associativity imposes
certain restrictions on the coefficients ajk of the quadratic part, which can
The Lie Algebra of a Lie Group 133
be seen only when the cubic terms are also included. Substitution of (25.2-2)
into (25.2-3) gives, after considerable cancellation,
i. a j xlymzk
aJk lm + b~Jkl (xjyk + xkyj)Zl i ak xjylzm
= aJk lm + c~Jkl Xj(ykzl + ylZk) .
(25.2-4)
The b's and c's will now be eliminated from this relation. Since j, k, I, and m
are summation indices, they can be renamed in each term in such a manner so
that the factors xk/zm appear throughout; then, since the equation is an
identity in x, y, Z, the net coefficient of xkylzm must vanish; this gives
This expression is now summed over the even permutations of k, I, m, and

from the sum is subtracted the result of summing over the odd permutations;
the right member gives zero and the left gives a sum that can be written
as
= L (a~m - a~)(a{l - a{k)
(even permutations of k, I, m)
(25.2-5)
There are no further restrictions imposed on the coefficients a~k by the

group axioms, for the above restrictions just suffice to determine a Lie
algebra, and it will appear eventually that every Lie algebra is the Lie algebra
of some Lie group. (That is a deep result; see Hausner and Schwartz, Section
III,7.)
From the equation
m(l(x), x) == m(x, I(x == 0,
which expresses the group relation that a-1a = aa- 1 = 1 for all a, it is
found that the expansion ofl(x), through quadratic terms, is
(25.2-6)
25.3 The Lie Algebra of a Lie Group
The Lie algebra A of a Lie group G is based on the so-called infinitesimal

elements of G, that is, on the tangent vectors of smooth curves emanating from
some interval : ;
the identity element 1. Such a curve is given by a function get), defined for
t ::;; 8 (8 > 0), such that g(O) = 1 and such that the cor-
responding curve x(t) = <p(g(t in the parameter space [Rn has a tangent at
each point (including t = 0, which is in fact the only point that matters). If
x(t) = <p(g(t = At + ... , (25.3-1)

then the components IV of A transform as the components of a contravariant
vector at the point x = of the manifold, when the coordinates are changed
(see Section 26.1). Namely, since
d
A = dt <p(g(tlt=o,
134 Lie Groups
it is seen that if new coordinates

(i = 1, ... , n)
are introduced, then
This vector is called the tangent vector at 1 to the curve get).

[According to the above definition, a curve is determined by the para-
metrization as well as by the set of group elements in it; get), g(2t), and
g(2t - 1) are different curves emanating from 1, and they have different
tangent vectors, although those vectors all have the same direction.]
The set of all tangent vectors at 1 in an n-dimensional group G constitutes
an n-dimensional vector space over the real field IR, for if At + ... and
Ilt + ... are the coordinates of two smooth curves, then (aA. + bll)t + .. "
where a and b are real numbers, is the coordinate of another smooth curve.
This vector space will be made into an algebra, called the Lie algebra A = A( G)
of the group G, by defining a multiplication based on the group multiplication
in G.
If get) and h(t) are smooth curves emanating from 1 in G, and if the function
k(t) is defined as the commutator of get) and h(t), that is, if
k(t) = g(t)h(t)g(t) -1 h(t) - 1,
and if the coordinates of get) and h(t) are

cp(g(t = A.t + .. "
cp(h(t = Ilt + ... ,
then a direct calculation shows that the coordinate of k(t) is
cp(k(t = vt 2 + ... , (25.3-2)
where
(25.3-3)
Equation (25.3-2) shows that the function k(t) ~ k(Jt) is a curve in G
emanating from 1 and that v is its tangent vector; v is called the Lie product
of A. and 11 and is denoted by the Lie bracket expression
v = [A., 11]. (25.3-4)
[One can also establish directly from (25.3-3) that the Vi transform as the
components of a vector, when coordinates are changed, after the rather
complicated transformation laws of the coefficients a;k have been worked
out.]
From (25.3-3) it follows that the Lie product is linear in each factor and is
antisymmetric: [11, A] = - [A., 11]; from the identity (25.2-5), which was
The Lie Algebras of Linear Groups 135
deduced from the associativity in G, it follows that the Lie product also
satisfies the Jacobi identity
[A, [~,v]] + [~, [v,A]] + [v, [A,~]] = 0. (25.3-5)
Am example of a Lie algebra is the algebra of vectors in 1R 3 , where the Lie
product is defined to be the vector product [A, ~] = A x ~,in the notation of
Gibbs. The Jacobi identity can be verified either by writing (25.3-5) in com-
ponents or by use of the identity A x (~ x v) = (~ . v)~ - (~. ~)v. Lie
algebras of matrices are discussed in Section 25.5, below.
25.4 Abstract Lie Algebras

A finite-dimensional vector space, over IR (or C) as the field of scalars, in
which a multiplication [A, ~] is defined that is linear in each factor, is anti-
symmetric, and satisfies the Jacobi identity (25.3-5), is called a real (re-
spectively complex) Lie algebra. The Lie algebra derived from a group is real.
A Lie algebra is not only generally noncommutative, but also non-
associative; that is, generally, [A, [~, v]] =1= [[A, ~], v]; it has no unity
element because [A, A] = 0, for any A, by the anti symmetry of [A, ~J.
An n-dimensional Lie algebra can be fully specified by choosing a basis
E 1, .. , En in it (a set ofn linearly independent vectors) and then specifying the
3
n constants of structure qk (they are clearly not all independent), defined by
(25.4-1)
25.5 The Lie Algebras of Linear Groups

If G is a group of m x m matrices L, M, ... , that is, a subgroup of GL(m, IR)
or GL(m, C), then its Lie algebra can be realized as a Lie algebra of m x m
matrices. A curve emanating from 1 has the form A(t) = I + tL + ... ;
matrices L obtainable in this way constitute a vector space A of dimension
:s: 2m 2 To find the Lie product in A, let B(t) = I + tM + ... be another
curve emanating from 1, and define
K(t) = A(t)B(t)A(t)-1 B(t)-1
= (1 + tL + .. -)(1 + tM + ...)(1 - tL + ... )(1 - tM + .. -),
(25.5-1)
as was done in Section 25.3 for the function k(t) in the manifold of an abstract
group. When these expressions are multiplied out, the linear terms cancel.
Quadratic terms can arise from the non constant terms of at most two
factors, the constant term I being taken for the other factors. All quadratic
terms arising from A(t)A(t)-1 and from B(t)B(t)-1 cancel exactly, so the only
remaining quadratic terms come from combining a linear term from A(t) or
A(t)-1 with a linear term from B(t) or B(t)-1. From this it is seen that
K(t) = I + t 2 (LM - ML) + .. .. (25.5-2)
136 Lie Groups
According to the definition of Lie product [see equations (25.3-2) and

(25.3-4)], then,
[L,M] = LM - ML. (25.5-3)
This illustrates the general rule that any associative algebra can be made into
a Lie algebra by defining [A, Ji] as AJi - JiA, where AJi and JiA denote products
in the original associative algebra.
Matrices like Land M above were called infinitesimal group elements in
Chapter 20; they are given by L = (d/dt)A(t) 1/= 0' etc., and it was shown there
that for SO(3), which is 3-dimensional, they can be taken as
o o -1
o
1 -r} o
o
o
o
and that these matrices satisfy the relations
[Ei' Ej] = Ek (ijk = 123,231, or 312).
In Chapter 21 it was shown that for SU(2), which is also 3-dimensional,
the infinitesimal group elements can be taken as
Tlz =
1(01 -1)o '
2. Tl3 =
1
2.
(-i
0
and that these matrices satisfy the same relations as the Ei , namely,
(ijk = 123,231, or 312).
According to (25.4-1), these relations determine the structure of the cor-
responding Lie algebra completely. Therefore, if A3 and A z denote the Lie
algebras of SO(3) and SU(2), respectively, then the linear mapping of A3
onto Az induced by Ei ~ Tli (i = 1, 2, 3) is an isomorphism. Regarded as
abstract Lie algebras, A z and A3 are identical, whereas the corresponding
groups are not. As will be seen later, the isomorphism of the algebras induces
a one-to-one mapping of the groups only in a neighborhood of 1. This
mapping is an isomorphism as far as it goes, but, when extended globally,
it becomes the 2-to-1 homomorphism of SU(2) onto SO(3) discussed in
Section 19.7.
Note that Az and A3 are real Lie algebra. Although the matrices Tlb Tlz, Tl3
are complex, Az consists of the linear combinations of these matrices with
real coefficients.
25.6 The Exponential Mapping; Logarithmic

Coordinates
We seek a curve get) in the group G such that
get + s) = g(t)g(s); (25.6-1)
the points on this curve constitute a one-dimensional Abelian subgroup;
The Exponential Mapping; Logarithmic Coordinates 137
g(O) is the identity 1 of G. The corresponding curve x(t) = <p(g(t passes

through the origin of the coordinate space [Rn and satisfies the equation
(25.6-2)
Taking d/ds and then setting s = in this equation gives

(25.6-3)
where q~( ., .) is defined by
. a.
qj(x, y) = ayi m'(x, y) (25.6-4)
and is of class C\ and where ,V U = 1, ... , n) are the components of the

tangent vector Ato x(t) at t = 0. The system of ordinary differential equations
(25.6-3), together with the initial condition x(O) = 0, has a unique solution
x(t) in some neighborhood of t = 0, for given A. From the form of (25.6-3)
it is clear that the solution depends on A and t only through the combina-
tion tA; hence we write the solution as x(t, A) = X(tA). The corresponding
curve get) in G is denoted by exp(tA) or eO.. , because of(25.6-1), which now takes
the form e(t+S)A = etAe sA ; hence X(tA) = <p(etA ). This extends the definition of
the exponential function abstractly to a mapping from a Lie algebra A to a Lie
group G, but in case the elements of G and A are matrices, the definition
agrees with the usual one. The function X() is of class C 3 .
So far, the equation (25.6-1) has been used only for s in the immediate
neighborhood of 0, but it will now be shown that this functional equation or,
what amounts to the same thing, (25.6-2) is satisfied for all t and all s.
Theorem. The solution x(t) of(25.6-3) satisfies (25.6-2)for all t and s such
that x(t), xes), and x(t + s) are defined.
Note 1. This does not follow in any trivial way from the mere forms of the
equations (25.6-2, 3) because, as will be seen, the associative law of group
multiplication has to be used in the proof.
Note 2. Once the functional equation get + s) = g(t)g(s) has been established
for t, s, and t + s in an interval ( - T, T), the equation itself is then used to
define get) for t in (-2T, 2T), then in (-4T, 4T), etc. In consequence, get) is
uniquely defined for all t and satisfies the functional equation for all t and
s-details to be supplied by the reader.
Note 3. If G is a group of matrices, so that the elements Aof A are also matrices,
then the corresponding equation
(25.6-5)
is usually established as follows: The matrix
J1(s) = (e At )-l e A(t+s>,
138 Lie Groups
as a function of s, satisfies the same differential equation and the same initial
condition as the function eM, namely
d
ds pes) = p(s)A.,
p(O) = I;
since the solution of this initial value problem is unique, it follows that
(el.l)-l~(I+S) = eM,
which is equivalent to (25.6-5). This argument is used as a model for the proof,
given below, for the abstract case.
PROOF OF THE THEOREM. It will be shown that the function
y(s) = m(l(x(t, x(t + s, (25.6-6)

which is the coordinate of the group element g(t)-lg(t + s), satisfies the same
differential equation (25.6-3) and the same initial condition as x(s), which is the
coordinate of the group element g(s); since the solution of the initial-value problem
is unique, it follows that g(s) = g(t) -lg(t + s); hence, g(t + s) = g(t)g(s), as required.
That y(s) satisfies the initial condition y(O) = 0 follows from g(t)-l g(t) = 1. Dif-
ferentiation of (25.6-6) with respect to s and substitution of an expression like (25.6-3)
for x/(t + s) gives
i(s) = q~(I(x(t, x(t + s'lIc(x(t + s), OW. (25.6-7)
It will now be shown that the total coefficient of lk in the right member of this equation
is equal to ql(y(s), 0), so that y(s) satisfies the same equation (25.6-3) as x(s). For that
purpose, the associativity of group multiplication is used in the following form:
mi(y(s), z) = mi(m(l(x(t, x(t + s, z)
= mi(l(x(t, m(x(t + s), z, (25.6-8)
which is equivalent to saying that [g(t)-lg(t + s)]h = g(t)-I[g(t + s)h]. Differentia-
tion with respect to Zk gives
ql(y(s), z) = q~(I(x(t, m(x(t + s), zq{(x(t + s), z);
therefore
ql(y(s), 0) = q~<I(x(t, x(t + sq'(x(t + s),O),
as claimed, so that y(s) and x(s) satisfy the same equation, and the theorem is proved.
For t = 1, we have
X(A.) = <p(el.).
Furthermore, at A. = 0, the Jacobian
det(~~;) = det(q~{A., 0
is different from zero, because q~(O, 0) = t5~. Therefore, the function X(A.)
has an inverse in some neighborhood of the origin, and the components of A.
An Auxiliary Lemma on Inner Automorphisms; the Mappings Ad~ 139
can be taken as new coordinates, called logarithmic or normal coordinates of

the group element 9 = e" in some neighborhood of 1 in G. One also writes
A = log g. The (generally many-to-one) mapping 1..-
e" of A into G is
called the exponential mapping. The components of a given vector A depend,
of course, on the choice of the original coordinate system {U, <p, N}, but
they transform as the components of a vector, so that each element of A
(a vector) is mapped into a unique group element.
25.7 An Auxiliary Lemma on Inner

Automorphisms; the Mappings AdJl
The lemma proved below is needed as part of the analytic machinery for
proving the Campbell-Baker-Hausdorff theorem, but has independent
interest. As in Chapter 21, a homomorphism of a group G (abstract or
not) onto a group of nonsingular linear transformations of a vector space
V is called a representation of G on V. If the homomorphism is one-to-one
(i.e., an isomorphism), the representation is called faithful. If every Lie group
had a faithful representation on a finite-dimensional space, the whole theory
would reduce to the manipulation of matrices. Although that is not the case,
any simply connected group has a particular representation (generally
unfaithful), the so-called adjoint representation, described at the end of this
section.
Note. The reader may wish to skip the proofs in this section and the next
three. However, the definitions and the statements of the lemmas and theorems
will be needed later.
If e" is a fixed group element, the mapping 9 - e"ge-" is an inner auto-
morphism of G (see Section 18.10); it induces a linear mapping of A into
itself, to be discussed below.
First, if Jl is any fixd element of A, the linear transformation of A into itself
defined by A - [Jl, A] is denoted by Ad". Relative to a basis in A, this trans-
formation is represented by an n x n matrix. [This matrix is not to be con-
fused with and is generally of different order than the matrices of which A is
composed, in case G is a linear group. If the elements of A are m x m matrices,
then Ad" can be represented by an m 2 x m 2 matrix.] Ad; denotes the trans-
formation
1..- AdiAd"A) = Ad,.[Jl, A] = [Jl, [Jl, A]];
similarly, Ad"Ad v , Ad~, and exp{Ad,,} denote the transformations
A - [Jl, [v, A]],
A - [Jl, [Jl, ... [Jl, A] ... J],
1..- (1 + Ad" + ~2 !Ad" + ...)1..'

2
where 1 denotes the identity transformation in i\..

140 Lie Groups
Lemma. Let e" be afixed group element;for each A, in A let get) be a smooth
curve, with g(O) = 1, whose tangent vector at 1 is A" and let A,' be the tangent
vector at 1 to the curve e"g(t)e-"; then, the mapping A, ~ A,' is a linear
transformation in A given explicitly by
(25.7-1)
PROOF. For each fixed s in the interval [0, IJ, the group automorphism get) ~
e'''g(t)e-''S induces a mapping A ~ A(s), in the manner described in the lemma, and
it will be proved that A(s) satisfies the same differential equation in s as esAd"A. The
logarithmic coordinate of the group element get, s) = e'''g(t)e-''S is
x(t, s) = log get, s) = tA(s) + .. '. (25.7-2)
To find the s derivative at s = So, we write
(25.7-3)
x(t, So + s) = m(m(sl1, x(t, so)), - sl1), (25.7-4)
where, as in the preceding sections, m(, .) gives the coordinate (here, the logarithmic
coordinate) of the product of two group elements in terms of the coordinates of the
factors. We recall the definition (25.6-4) of q~(x, y) and define similarly
. o.
pj(x, y) =- m'(x, y). (25.7-5)
oX j
Differentiating (25.7-4) with respect to t shows that the tangent vector at t = 0 to
the curve x(t, So + s) is given by
(25.7-6)
From the expansion (25.2-2) of mi(x, y), the corresponding expansions of p~ and ql
are, to first order,
p~ = b~ + a~,i + "',
ql = bl + a{k x' + .. '.
(These expansions and the quantities p~ and ql now all refer to the logarithmic co-
ordinates.) From (25.7-6), then,
that is,
[see (25.3-3, 4)]; in terms of a basis in A, this is a system of n first-order differential

equations with constant coefficients, and the solution is
(25.7-7)
In particular, A' = A(l) = eAd"A, as was to be proved.

Auxiliary Lemmas on Formal Derivatives 141
EXERCISE
Show that, for fixed 11, the linear mapping A. -> eAdpA. is an automorphism of A,
that is, (1) that it is one-to-one and (2) that
eAdp[A., v] = [eAdpA., eAdpv]. (25.7-8)
[Note that Ad p itself is not even a homomorphism of A.]
One says that eAdp is the inner automorphism of A induced by the inner
automorphism g --> e"ge-" of G.
Comment. If e"' and e" 2 are any two group elements, and if e" 3 = ell 'ell2 , then
the automorphism
of G induces the automorphism
of A. Therefore, in a neighborhood of 1 such that logarithmic coordinates are

defined, the association of eAdp with the group element e" is (locally) a homo-
morphism of G into a group of linear transformations in the vector space A.
It will be seen later that if G is simply connected, this association can be
extended to a homomorphism of all G, i.e., to the adjoint representation of G.
The group of linear transformations in A generated by transformations of

the form eAdp is called the inner-automorphism group of A and is denoted by
Int(A). Each of its elements is a finite product eAdp1 eAdp2 ... eAdpj and is
called an inner automorphism of A (it cannot necessarily be written as eAd. for
some (J E A). If G is simply connected, this element of Int(A) is the image,
under the homomorphism referred to above (the adjoint representation)
of the element g = e"'e" 2 ... e" j of G.
25.8 Auxiliary Lemmas on Formal Derivatives

It is customary to denote by dg(t)/dt or get) the tangent vector at t to a curve
get) in the group manifold. This is a purely formal notation, unless G is a
group of matrices, for in general no meaning has been attached to a "dif-
ference quotient" (g(t j ) - g(t 2/(t j - t2)' Nevertheless, the notation can be
manipulated according to many of the rules of differentiation, often with
considerable savings in the lengths of formulas and of their derivations.
If hand get) are respectively a single group element and a curve in G, and
if g(to) is the tangent vector to get) at to, then hg(to) is defined to be the tangent
vector to the curve hg(t) at to. The one-to-one mapping of G onto itself, given
by g --> hg for fixed h, is called a left translation in G. It induces a one-to-one
linear mapping of the space of tangent vectors at a point go = g(to) onto the
space of tangent vectors at hg o . Similarly, g(to)h is defined as the tangent
142 Lie Groups
vector to the curve g(t)h; hence a right translation g ~ gh in G induces a

similar mapping of the space of tangent vectors at go into that at go h. Accord-
ing to these definitions,
~ h (t) = h dg(t) ~ ()h = dg(t) h (25.8-1)
dt g dt ' dt g t dt
From the associativity in G it follows that if h1 and h2 are group elements

and Aisatangent vector to some curve, then (h1h2)A = h1(h2 A) and (h1A)h2 =
h 1(Ah 2); in consequence, all parentheses can be omitted. In such a product,
any number of factors can be group elements, but only one can be a tangent
vector, and the product is then a tangent vector. Generally, A, hI.., Ah, etc.
are vectors at different points of G and cannot be compared, since, if 1..1 and
1..2 are vectors at different points, no meaning has even been given to an
equation 1..1 = 1.. 2 . However, if A is in A, hence is a vector at 1, then A and
hAh - 1 are vectors at the same point of G; in particular, the mapping A ~
e"Ae -" is the inner automorphism eAd. of A discussed in the preceding section.
By working with a particular coordinate system, one can easily establish
the following relations:
:t (g(t)h(t)) = get) d~;t) + d~;t) h(t), (25.8-2)
:t (g(t)-1) = _g(t)-1 d~~) g(t)-l, (25.8-3)
d
dt etA = Ae tA = etAA. (25.8-4)
Higher derivatives would generally take us out of the space of tangent

vectors into other (probably less interesting) spaces. However if g = g(s, t) is
a smooth two-parameter family of elements of G, then the quantities defined
by
_1 og
el = el(s, t) = g os '
(25.8-5)
~=~(st)=g-10g
, ot
are tangent vectors through the point g(s, t)-1g(S, t) = 1 for all sand t, i.e.,
are always in A and can be further differentiated.
Lemma. With el and ~ as in (25.8-5),

Oel o~
at - os = [el, n (25.8-6)
[Note. If G is a linear group, so that o2gjotos is well defined as a matrix (in

addition to g, ogjot, and ogjos), then the lemma follows immediately from
(25.8-5), after further differentiation, and (25.8-3); the terms in o2gjotos
cancel.]
An Auxiliary Lemma on the Differentiation of Exponentials 143
PROOF. To verify (25.8-6) for s = So, t = to, write g(s, t) = g(so, to){j(s, t); then
pcan also be written as
IX and
a-
IX(S, t) = {j- I a~'
P(S, t ) = --I
a-9
9 at'
Since {j(so, to) = 1, the expansions of the coordinates of {j and g-I in powers of
s - So = Sl and t - to = tl start with the linear terms [q(I) = 0 is assumed]:
(25.8-7)
hence,
yi(S, t)dJ[ cpi({j(S, t)-I) = -AiS I - thl - Aisi - Bislt l - Citi
+ a~k(A.isl + Jlitl)(AkSI + l t l ) + .... (25.8-8)
Since IX(S, t) is the tangent vector to the curve obtained from {j(s, t)-I{j(S', t) by
varying s', for given sand t, and then setting s' = s, and similarly for p(s, t), it follows
that
. a.
ex'(s, t) = as' m'(y(s, t), x(s't)) 1.,= ..
. a.
P'(s, t) = at' m'(y(s, t), x(s, t)) I., =.;
a direct calculation from (25.8-7) and (25.8-8) then gives
aex i api) i . k k'

( -at - -a = aik(exJp - ex {3J).=.o,'='O'
S s=so. t=to
which is the desired result (25.8-6), by virtue of the definition (25.3-3, 4) of the Lie
product.
25.9 An Auxiliary Lemma on the Differentiation

of Exponentials
Lemma. Ifl.. = I..(t) is a smooth curve in A, and ifl..' or I..'(t) denotes dl../dt
(which is also a curve in A), then
e-1. :t e1. = f(Ad1.),J:, (25.9-1)
where
1 - e- z 1 1 2
f(z) = z
= 1- - z
2!
+ -3! z -'" . (25.9-2)
Comment 1. On the left side of (25.9-1), de1./dt is a tangent vector at the point
e1. of the group G; multiplication of this vector on the left by e -1. transforms
it into a tangent vector at the point 1 of G, i.e., into an element of A.
144 Lie Groups
Comment 2. On the right side of (25.9-1), since the transformation AdA can be
represented by an n x n matrix, and since the series for fez) converges
absolutely for all z, it follows thatf(Ad A) is a well-defined linear transforma-
tion in A. [In particular, in and A,' commute, i.e., if AdA A,' = [A" A,'] = 0, then
the right member of (25.9-1) is just A,' itself.]
Comment 3. If A is a Lie algebra of matrices, the lemma can in principle be

obtained by multiplication of the power series for e-A(t) with the power
series obtained by term-by-term differentiation of the series for eA(t), taking
noncommutativity of the matrices A, and A,' into account by writing [A" A,'] =
A,A,' - A,'A,.1t is easy to verify in this way that the first two or three terms of the
expansion are as claimed in the lemma.
PROOF OF THE LEMMA. Define

a
CI.(s, t) = e-s).(t) as es).(t) = A(t),
so that ~(o, t) == 0, and use the lemma of Sections 25.8:

d a
- A(t) - - ~(s, t) = [A(t), ~(s, t)].
dt as
For fixed t, this is a differential equation of the form
d
A' = ds ~(s) = A~(s),
where A is the matrix of the transformation Ad).(t); the solution that satisfies the
initial condition ~(o) = 0 is
1 - e- As
~(s) = A'.
A
[Note. A is a singular matrix, because Ad). A = O. The expression (1 - eAS)jA denotes
the matrix obtained by substituting A for z in the entire function (1 - eZS)jz.] There-
fore,
as was to be proved.
25.10 The Campbell-Baker-Hausdorff (CBH)

Formula
If A, and Jl are commutative elements of A (i.e., if [A" Jl] = 0) or commutative
matrices, then eAe" = eA+". In general, we seek a 0" such that eAe" = ea. The
CBH formula gives 0" explicitly in terms of A, and Jl and Lie ,brackets con-
taining A, and Jl, for A, and Jl in some neighborhood ofthe origin of A.
The Campbell-Baker-Hausdorff (CBH) Formula 145
Theorem (CBH). Let1/l(z) = z log zl(z - 1) = 1 + w/1 2 - w 2 /2 3 +

w3 /34 - ... , for Iwl < 1, where z = 1 + w; then, for A and Ji in some
neighborhood of the origin in A,
<r ~ log(el.e") = A+ f 1/I(eAd"etAdp)Ji dt. (25.10-1)
Comment 1. The argument of 1/1( .) in the formula can be represented by an

n x n matrix M = I + W, which can be made arbitrarily close to the unit
matrix by taking A and Ji sufficiently small. [Ado is the matrix 0, eAdo is the
matrix I.] The power series for 1/1(1 + w) converges absolutely for Iwl < 1;
hence, the series for 1/1(1 + W) converges elementwise if each element wij
of W satisfies Iwij I < lin, i.e., if A and Ji are restricted to some neighborhood
N of the origin in A.
Comment 2. The vectors A, Ji, and (J are the coordinates (logarithmic) of the
group elements el., e", and eO; hence, <r = m(A, Ji). The CBH formula is there-
fore a formula for m(, .) in logarithmic coordinates, and the structure of A
completely determines multiplication in the group in some neighborhood of 1.
Comment 3. All the expansions implied in (25.10-1) can be carried out;

the first few terms are
log(el.e") = A + Ji + ![A, Ji] + /2[A, [A, Ji]] + /2[Ji, [Ji, A]] + ....
(25.10-2)
PROOF OF THE THEOREM. The function 10g(eAe''') is denoted by a(t); a differential

equation for a(t) will be found, whose solution gives the CBH formula. First, since
the lemma of the preceding section gives

d
II = e-"(t) - e"(t) = f(Ad )a'(t) (25.10-3)
r- dt a(,),
where fez) = (1 - e-Z)jz. If x(z) is defined as z/(1 - e- Z ) = (f(zWl, then

x(M)f(M) = I for any matrix M, so that (25.10-3) can be solved for a'(t):
a'(t) = X(Ada(,)Jl. (25.10-4)
From the definitions of X() and 1/J(.) it follows that X(z) = 1/J(e Z ); hence
a'(t) = 1/J(eAda(')Jl.
This differential equation can be simplified by use of the lemma of Section 25.7,
which says, in the notation of Section 25.8, that the mapping eAd " is the mapping
v ..... e"ve-"; specifically,
eAd,,(,)V = e"(')ve-a(t) = e"e'''ve-'''e-l.
= eAd"e'Adpv.
146 Lie Groups
Now, the unknown appears only on the left, and the CBH formula follows directly
by integrating from t = 0 to t = 1 and by using the conditions
0"(0) = log eA = A., 0"(1) = log eAe" = 0".
Since Ad," is the transformation v ~ [A, v], the matrix elements of Ad," are
linear functions of the components of A. Therefore, the matrix elements of the
transformations exp{Ad,"} and exp{t Ad,,} are analytic functions of the
components of A and Jl. Since t/J(z) is analytic, for 1z - 11 < 1, it follows that,
for A and Jl restricted to the neighborhood IV of the origin in A referred to in
Comment 1, the components of (J = log e'"e" are analytic functions of the
components of A and Jl; by analytic continuation, they are analytic for all A
and Jl such that the logarithm is defined.
When logarithmic coordinates Ai are used, log e'"e" is simply the multipli-
cation function denoted by m(A, Jl) previously. Although this function was
only assumed to be of class C 4 , it is now seen to be analytic, when logarithmic
coordinates are used. In these coordinates, the inversion function I( . ) is given
by 1(1.) = - A, and hence is also analytic.
EXERCISE
Express the matrix elements of AdA in terms of the components Ai of A., for a given
basis E 1, .. , En in 1\., and the corresponding structure constants C)k.
25.11 Translations of Charts; Compatibility;

G as an Analytic Manifold
In this section, a basic chart is chosen, with logarithmic coordinates in it,

and other charts are obtained from it by translations in the group. To simplify
the work, the basic chart is chosen as a rather small one, so as to have
various convenient properties.
First, let U be a neighborhood of 1 in the group G in which the mapping
e'" ~ A into the Lie algebra is one-to-one, so that logarithmic coordinates
can be used in U. Next, let m be a sufficiently small subneighborhood in U
(an open subset of U), also containing 1, so that if g and h are in m, then gh
is in U, and log gh is given by the CBH formula. Finally, let lID be a sub-
neighborhood in msuch that if g and h are in lID, then gh is in m, while g-l
and h- 1 are in lID (this permits use of the CBH formula for triple products
like glgzlg3' etc.), and let N be the image of lID in A; i.e., N = {log g: gElID}.
From now on, {lID, log, N} will be regarded as the basic chart.
It is recalled that, for any fixed a in G, the one-to-one mapping of G onto
itself given by g ~ ag is called a left translation by a and the mapping g ~ ga is
a right translation. For any fixed a in G, a left-translated chart {alID, alP, N}
is defined as follows: First, the subset aIID of G is defined as
aIID = {ag1:g1EIID};
Translations of Charts; Compatibility; G as an Analytic Manifold 147
then for each g = ag 1 in alID, acp(g) is defined as log g 1. Note that the image
of alID under the mapping g ~ acp(g) is the same open set N in the coordinate
space A as the image of lID under the mapping g ~ log g. Right-translated
charts {lIDa, CPa' N} are similar.
Theorem 1. Any two charts obtained by translation of the basic chart

{lID, log, N} are compatible, in fact analytically compatible.
Comment. If a lies in lID, then the left (also right) translation by a is a homeo-
morphism in the basic chart, as far as it is defined, because the coordinates of
ag are continuous (in fact, analytic) functions of the coordinates of g, by the
CBH formula, and the coordinates of g are continuous functions of the co-
ordinates of ag, because g = a- 1 (ag). Hence, any translated chart is com-
patible with the basic chart, and it will be shown that any two translated
charts are also compatible with each other. G thus becomes a manifold, and
Theorem 2 below then shows that the translations are homeomorphisms
in all G.
PROOF OF THEOREM 1. First consider charts obtained by left translation of the

basic chart by a and b, respectively. If the intersection aW n bW is empty, the charts
are automatically compatible. Otherwise, it must be proved first that the inter-
section is mapped by a<fl (also by b<fl) onto an open set in A (see Section 23.2). That is,
if g is any point in the intersection, so that g = ag l = bh l , where gl and hI are in W,
then it must be shown that g is in some subset that is an open subset of each chart,
according to the topology of that chart. Therefore, consider an element g' = ge
that is close to g, so that e is close to I (just how close it has to be will be seen in a
moment). Then, g' = agle = bhle. Now, W is an open set; hence gl and hI are
interior points. Since the coordinates of g'l = gle and h'l = hIe depend continuously
on the coordinates of e, there is a neighborhood Wo of 1 such that, for e in W o , g~ and
h'l areinW. See Figure 25. 1. ThesetsglW o = {gle: einW o} andhlW o = {hIe: einW o}
are neighborhoods in W of gl and hI respectively, i.e., open sets containing gl and
hI' respectively, because, according to the above comment, the left translations by gl
and by hI are homeomorphisms in W. Therefore, since g = ag l = MI' gW o is an
open set in the topology of either chart and is contained in both charts. Hence, the
Figure 25.1
148 Lie Groups
first condition for compatibility is satisfied. Second, it must be proved that the co-
ordinates lP(g) = log gl and b<j)(g) = log hI are analytic functions of each other,
for all 9 that can be written both as ag l and bh l , with gl and hI in illl Since b-Ia =
hl gil, and since hI and gl are in W, it follows that b-Ia is in m(even though a and h
themselves may not even be in the neighborhood U where logarithmic coordinates
are defined). Hence, log hI' which is =log((b-Ia)gl), depends analytically on
log gl, according to the CBH formula. Similarly, log gl depends analytically on
log hI' Lastly, a similar argument shows the compatibility of two right-translated
charts or of a right-translated and a left-translated one.
Theorem 2. When a group G is covered by translated charts, as described

above, the Hausdorff separation axiom is satisfied.
PROOF. Let a and b be any two points of the group, and suppose that they are not
separated, that is, that any neighborhood of a and any neighborhood of b have points
in common. It is to be proved that then a = b. Let 1111 be a norm in A, e.g., the
Euclidean norm relative to some basis E I , ... , En in A. For any b > 0, let Uo be the
neighborhood of 1 consisting of those elements 9 in W such that Illog gil < b. Then
aUoandbUoareneighborhoodsofaandb;letgbeinaU J n bUo,sothatg = a = b1],
where and 1] are in Ub . Then b - la = 1] - I, but since log 1] -I is given by the CBH
formula in terms ofiog 1] and log f.-I = -log 8, both of which can bemade arbitrarily
small by suitable choice of b, it is seen that log 1]8 - I, which is = log b - I a, is = 0;
hence b -Ia = 1, as was to be proved.
Therefore, G, when covered by the translated charts, as will be assumed

from now on, is real analytic manifold. We assume, as in Chapter 23, that
G can be covered by a countable collection of these charts.
Lemma. A left or right translation is an analytic mapping in all G.

PROOF. The left translation 9 ---> ag maps the chart {bW, b<t>, N}(b is arbitrary) onto
the chart {abW, ab<t>, N}; the corresponding coordinate mapping is the identity, for if
9 = bh, then
b<t>(g) = abcp(ag) = log h.

The analyticity of the same mapping in right-translated charts follows from the
compatibility of charts.
Theorem 3. Product and inverse are analytic in all G.

PROOF (For Product). Let a and b be arbitrary elements of G. Let a<t>, b<t>, and ab<t>
denote coordinates in charts obtained from the basic one by left translation by a, b,
and abo It will be shown that, for 9 and h close to 1, the abCjl coordinate of (ag)(bh)
depends analytically on the a<t> coordinate of ag and the b<t> coordinate of bh. Then,
since all translated charts are analytically compatible, the same conclusion holds
for any three charts in which a, b, and ab are respectively located. Now
ab<t>(agbh) = ab<t>(abb-Igbh) = log(b-Igbh);
hence the last member of this equation must be shown to depend analytically on
log 9 and log h. The element b - 19b is obtained from 9 by a left translation by b - I
Lie Algebra Homomorphisms 149
followed (or preceded) by a right translation by b, and is = 1 for 9 = 1;

hence log(b - 19b) exists for 9 near 1 and depends analytically on log g, by the lemma
above. By the CBH formula, log(b - Igbh) depends analytically on log(b -1gb) and
log h, and the conclusion follows.
The proof of the analyticity of the inverse is left as an exercise.
Theorem 4. The principal component ofG, i.e., the component containing 1,

is generated by any neighborhood Wo of 1. That is, any g in that component
can be written as a finite product g = gig 2 , where each g i is in Wo.
(See Section 23.6 for" component.")
PROOF. Let Go be the subgroup of G generated by Wo n Wo I, where Wo 1 is the
set consisting of the inverses of the elements of Wo. The set Woo = Wo n Wo 1 is an
open set containing 1. Clearly Go is connected, because the factors gl, g2, ... in such
a finite product can each, in turn, be smoothly changed into 1 and then removed from
the product, so that as a result 9 is connected to 1. It will be shown that Go is both
open and closed. Therefore, by Section 23.6, it is an entire component of G. [Then,
since Woo cWo, it is aJortiori clear that Go is generated by Wo] If 9 = glg2 gk
is any element of Go, then the open set gW oo , which contains g, consists of elements
of the form g1g2 ... gkgk+ l' where gk+ 1 is also in Woo; therefore gWoo is in Go, ancl
Wo is open. On the other hand, if g' is any limit point of Go, then the set g'Woo =
g'Wool = {g'h- I : h E Woo} is a neighborhood of g', and hence contains an element of
the form 9 1 ... gk; therefore, 9 1 ... gk = g' h - I for some h in Woo; hence, g' = 9 1 ... gk h
is in Go, and Go is closed, as was to be proved.
Remarks. Left translation is a homeomorphism in all G. (It is one-to-one

by group theory and is continuous, indeed analytic, by the lemma above.)
Therefore, each left coset aGo of Go in G is a component, and conversely.
The same argument with right translations shows that each component is
also a right coset; therefore Go is a normal subgroup. The number of com-
ponents is countable, because we have specified that the manifold can be
covered by a countable collection of charts. The theory puts no other re-
striction on the nature of the factor group, because if Go is a connected Lie
group and K any abstract countable group, then the direct product Go x K
can be regarded as a Lie group G having one component for each element in
K, and then GIGo ;:= K. For this reason, GIGo is uninteresting, and many
authors include a requirement of connectedness in their definition of a Lie
group. We prefer not to exclude groups like 0(3) from consideration.
25.12 Lie Algebra Homomorphisms

The homomorphism theory of Lie algebras and Lie groups, described in this
section and the next two, is closely analogous to the homomorphism theory
for groups, described in Sections 18.5-18.8.
Roughly speaking, a homomorphism is a mapping (generally many-to-one)
of a mathematical structure onto a less complicated structure of the same
150 Lie Groups
kind such that any relation that holds among elements of the first structure
also holds among their images in the second structure. Some of these relations
generally turn out to be trivial ones in the second structure, like 1 0 1 = 1 or
o + 0 = 0, while the remaining relations may be regarded as showing some
of the main features of the first structure, but with less fine detail. Just as for
groups, such a mapping exists if and only if the first structure contains a
particular kind of substructure (e.g., a normal subgroup), which can serve as
the kernel of the mapping; the theory shows how to reconstruct the mapping
in question, when the substructure is given, by first forming the so-called
quotient or factor structure (e.g., factor group) and then constructing the so-
called natural homomorphism of the first structure onto the quotient struc-
ture, which is then shown to be equivalent to the original homomorphism.
The execution of this program is straightforward for Lie algebras. For
Lie group, each idea introduced has an immediate parallel in the correspond-
ing Lie algebras, and this interplay between the groups and their algebras
provides powerful techniques for the investigation of the groups.
If A and X are real Lie algebras, a homomorphism A ----> X is a mapping ljJ
of A into X which preserves all the operations of the Lie algebra; that is, if
A, p are in A and a, b are in IR, then ljJ(aA + bp) = aljJ(A) + bljJ(p), and
ljJ([A, p)] = [ljJ(A), ljJ(p)]. For complex Lie algebras, the definition is the
same, except that C replaces IR.
Comment. It is not really necessary to assume in advance that Xis a Lie algebra;
it only needs to be a structure in which A + p, aI., and [A, JI] are defined.
However, the part of it onto which A is mapped by the homomorphism, i.e.,
the image ljJ(A), is then necessarily a Lie algebra.
IfljJ is one-to-one and onto, then it is an isomorphism; in symbols, A ~ X; if,

further, it is a mapping of A onto itself (X = A), then it is an automorphism.
A linear subspace A of A is a subalgebra if [A, p] E A for all A, pEA: if,
further, [A, p] E A for all A E A and all pEA, then A is an ideal of A. A Lie
algebra A of more than one dimension is simple if it contains no ideals except
{O} and A. (The reason for excluding one-dimensional algebras, in this
definition, will appear in Section 25.16.) It is easy to see that the kernel of a
homomorphism ljJ, namely the'set {A: ljJ(A) = O}, is an ideal in A. Therefore,
an alternative definition is that a Lie algebra of more than one dimension
is simple if it cannot be mapped homomorphically onto any less complicated
algebra except the trivial algebra {O}.
Ideals play almost exactly the same role for Lie algebras as normal sub-
groups play for groups.
If Ao is a subspace of A, then the relation A == p (mod Ao), defined as
meaning that A - p E Ao , is an equivalence relation,. which decomposes A
into disjoint classes called residue classes modulo Ao. If, for any fixed A E A,
the residue class {A + p: p E Ao} is denoted by A, and if we define aA = aA
and A + Ji = A + p, then the set of residue classes constitutes a linear vector
Lie Group Homomorphisms 151
space, called the factor space of A modulo Ao. We shall show that if Ao is an
ideal, the factor space can be interpreted as a Lie algebra.
Theorem 1. Let Ao be a subspace of a Lie algebra A. For each choice of

Al and Ji1 in A, the set {[A, JiJ: A - Al E Ao , Ji - Ji1 E Ao} is contained in a
single residue class (namely the class [Ab Ji1J) if and only if Ao is an ideal in A.
In this case, if the Lie product of two residue classes i1 and iiI is defined by
the equation [i b iilJ = [AI, Jil], then the factor space of A modulo Ao is a
Lie algebra, denoted by AIAo, called the factor algebra of A modulo Ao.
The mapping it -+ A is a homomorphism called the natural homomorphism
of A onto AjAo, and is denoted by IjJn.
Comment. Consideration of the commutative case, where every subspace

of A is an ideal, and where every Lie product [A, JiJ is =0, shows that the set
of elements [A, JiJ referred to in the theorem may be only part of the residue
class [AI' JilJ, which is = Ao
PROOF OF THEOREM 1. (1) Suppose that Ao is an ideal. Generally,
hence, if A - AI and ~ - ~I are in Ao, then both terms on the right are in Ao; hence,
[A, ~J and [AI, ~IJ are in the same residue class, as claimed. (2) Conversely, if, for
arbitrary A and ~, [A, ~ + <rJ is always in the same residue class as [A, ~J, for any <r in
Ao. then [A, <r] E Ao, and it follows that Ao is an ideal. (3) The definition [II' i z] =
[AI' Az] of the product of two residue classes shows that the mapping 1..-->1. is a
homomorphism, and then it follows from the comment after the definition of homo-
morphism that the factor space is a Lie algebra.
Theorem 2. (Homomorphism Law). If Ao is the kernel of a homo-

morphism IjJ of A onto A, then Ao is an ideal in A (as already noted), and
A -
A~A.

PROOF. Let t/Jn denote the natural homomorphism of A onto A/Ao. Whenever
t/Jn(A) = t/Jn(A I ), A and Al are in the same residue class, i.e., A - Al is in Ao; hence
t/J(A - AI) = 0, by definition of Ao as the kernel of t/J, hence t/J(A) = t/J(A I ). Therefore,
t/Jt/J;; I is a one-to-one mapping of the set A/Ao of residue classes onto i\.; the mapping
will be called X. It is left to the reader to complete the proof by verifying that X is
linear and satisfies the equation x([i, ji]) = [x(i), X(ji)], where 1. and ji are any
residue classes in A mod Ao , i.e., any elements of A/ Ao.
25.13 Lie Group Homomorphisms
In the case of Lie groups, a homomorphism must preserve not only all
algebraic relations but also all local topological and analytic relations
arising from the manifold structure.
152 Lie Groups
If G and G are Lie groups, then a mapping \f' of G into G is called a Lie-
group homomorphism if:
1. It is a homomorphism in the sense of group theory: \f'(gh) = \f'(g)\f'(h);
\f'(g-l) = \f'(g)-I.
2. It is a continuous mapping; that is, if <t> and (j) are any coordinate
systems in G and G, respectively, then each component of the vector
(25.13-1)
is a continuous function of the components of x, for all x such that the
above expression is defined.
Remark 1. The mapping is generally many-to-one; in fact, G may have more

dimensions than G.
Remark 2. Typical of non-group-theory relations in G that are preserved in G

is the convergence of a sequence gl' g2' ... to a limit h, which means that in a
chart containing h the coordinates of 9 1, 9 2, ... converge to the coordinates
of h; then, because of the continuity of the functions (25.13-1), the sequence
\f'(gl), lfI(g2),' converges to lfI(h) in G. Similarly the image of a curve in G
under a homomorphism is a curve in G.
Remark 3. The property of \f' of being a continuous mapping is invariant

under the addition or deletion of compatible charts in either or both of the
manifolds G and G, because if x' and y' are coordinates in any other charts,
then the schema
x'-x~y-y'
shows that y' is continuous in x' whenever it is defined by the composition of

the three mappings indicated.
Remark 4. If lfI is one-to-one and onto, and if \f'- 1 is also continuous, then lfI
is a Lie-group isomorphism.
Theorem 1. Let lfI be a Lie-group homomorphism (i.e., a continuous homo-

morphism) ofG onto G, and let A and A be the Lie algebras ofG and G. Then,
when expressed in logarithmic coordinates, the mapping \f' is locally a Lie
algebra homomorphism of A onto A.
Remark. As an obvious corollary, since the mapping \f' is linear in these

coordinates, every Lie-group homomorphism is locally an analytic mapping;
it is also globally analytic, because if 9 = go h for arbitrary go, then the element
\f'(g) = \f'(go)\f'(h) = \f'(go)\f'(go Ig) is analytic in 9 for h in some neighbor-
hood of 1, since products and inverses are analytic in each group.
PROOF OF THEOREM 1. For any A sufficiently near the origin in A, we can define
an element of i\. as
}. = log('P(e'-).
Lie Group Homomorphisms 153
We must show that the mapping A --> l is linear and maps [A, JlJ onto [l, ji]. We
show first that it maps tA onto t'i.. for real t, that is, that ?i. = t'i... For fixed A, the set
lJ'(eIA), t E IR, is a one-parameter subgroup of G which includes the group element
e~; hence for each t there is a real number s = f(t) such that
wheref(t) is continuous, because IJ' is continuous, and wheref(O) =

Since IJ' maps products onto products, we have
andf(1) = 1.
eJ(1 + S)A = IJ'(e(t + S)A) = IJ'( eIAe SA )

= lJ'(e'A)IJ'(eSA ) = eJ(I)AeJ(S)A = e[J(t) + J(S)]A.
Hencef(t + s) = f(t) + f(s), and the only continuous functions with this property
are linear. Since also f(O) = 0, f(1) = 1, we see that f(t) == t, as was to be shown.
We now apply the eBH formula to both sides of the equation lJ'(eSAe'l') = esi:e' ii, and
we see that
sA + tJl + 1SI[A, JlJ + ... = s'i.. + tji + 1st [l, jiJ + ... (25.13-2)
for all sand t. We write s = ss', t = st'. Since F,V = Ri, we can cancel one factor s on
each side of the above equation. Then, as s --> 0, the quadratic and higher terms
vanish; hence
S'A + t'Jl = s''i.. + t'ji,
and this establishes the full linearity of the mapping A --> l. Then the linear terms can
be dropped from both sides of (25.13-2), and by a similar argument we see that
[A, Jl J = [l, JiJ,

i.e., the mapping A --> 'i.. is a Lie-algebra homomorphism, as required.
This theorem does not have a global converse, but only a local one, and a
definition is needed, before the converse can be stated. If U is a neighborhood
of the identity 1 of a Lie group G, then an analytic mapping ' of U into a Lie
group G such that ,(gh) = ,(g)'(h) whenever g, h, and gh are in U is called
a local homomorphism of G into G. If the inverse mapping is also a local
homomorphism, i.e., is unique and analytic in some neighborhood of 1 in G,
then ' is a local isomorphism. If, furthermore, G = G, then ' is a local auto-
morphism of G.
We come now to the converse of Theorem 1.
Theorem 2. If A and A are the Lie algebras of G and G, then any Lie-
algebra homomorphism 1jJ: A --> A induces a local homomorphism 'I' : G --> G
given by the exponential mapping, namely '(e A) = e!/l(A), for eA in a suitable
neighborhood of 1 in G.
PROOF. By the eBH formula,
lJ'(eAel') = lJ'(eU I'+1/2[A,I']+"')
= el/l(J,+I'+ 1/2[A, 1'] + "');
154 Lie Groups
since", is a Lie-algebra homomorphism, the last expression is equal to

eo/l O.) + 0/1(,,) + 1/2[0/1(,,),0/1(,,)]+ ,
and since the CBH formula holds also in G, the above is equal to eo/l(")eo/l("l; that is,
'(e"e") = '(e")'(e"), as was to be proved.
As a corollary, if tjJ: A -+ A is a Lie-algebra isomorphism, then 'P is a

local isomorphism of the Lie group; if, furthermore, G = G and A = A, then
'P is a local automorphism of G.
The next theorem indicates circumstances under which a local homo-
morphism can be extended to a global one. This question arose in Section
21.1, where it was shown that quantum mechanics can lead to local repre-
sentations of groups like SO(3) and !l' p' It is recalled that the groups SO(3)
and SU(2) are locally but not globally isomorphic, and that !l' p and SL(2, C)
are also locally but not globally isomorphic. As a second example, let G be
the two-dimensional torus group, regarded as a group of matrices of the form
e~~ e~P).
Its Lie algebra A is the commutative Lie algebra of matrices of the form
Any linear transformation

IX -+ aIX + b{3, {3 -+ CIX + d{3,
where a, b, c, and d are real and ad - bc # 0, is a Lie algebra automorphism
of A. The corresponding local automorphism of G is
'P: (e~~ e~P) -+ ei(a~+bP) ei(C~+dP);

it is valid only for sufficiently small IX and {3, because, for example, the element
has the property that g2 is the identity of G, and this property is not preserved
under 'P, except for special choices of a and b.
Theorem 3. Let 'P be a local homomorphism of a Lie group G into a

Lie group G, defined in a neighborhood U of1 in G. Then:
(a) If there is an extension 'P of the mapping 'P to all of G, which is a
homomorphism in the sense of elementary group theory, then 'P is continuous,
hence analytic, throughout G-i.e., it is a Lie-group homomorphism.
(b) IfG is connected, then there is at most one Lie-group homomorphism
'P which is an extension of 'P.
(c) IfG is simply connected, there is exactly one such extension 'ft.
Law of Homomorphism for Lie Groups 155
PROOF OF PART (a). For arbitrary fixed go,let 9 = go h, where h is in the neighbor-
hood U, so that the mapping h --> 'P(h) is analytic. Then, .p(g) = <P(go)'P(h), but
products and inverses are analytic throughout both groups (Section 25.11); hence h
is analytic in g, and .p(g) is analytic in g.
PROOF OF PART (b). Any 9 in G can be written as glgZ'" gk> where all the factors
are in U; hence, if <P is any extension of the homomorphism 'P, then .p(g) = .p(g 1) ...
.p(gk) = 'P(g I) ... 'P(gk), which is completely determined by 'P; hence any two such
extensions must agree for every g.
PROOF OF PART (c). As on previous occasions, we let!D be a subneighborhood of
U, containing 1, such that if 9 and h are in !D, g-I and h- 1 are in !D, while gh and hg
are in U. The mapping .p will be constructed. Let hand k be elements of G joined by a
smooth curve CfJ. Let CfJ be partitioned into smaller segments by points (group elements)
go, gl, ... , gz so that go = hand gl = k, while g;-l g ;+ 1 is always in !D. We call
9 = 'P(go I gl )'P(gjlgZ )' .. 'P(gz--\ gl);

?i is an element of the group G, and if 'P were a global homomorphism g would be
= 'P(h- lk). From the above formula, it is seen that g is unaltered by a refinement of the
partition of CfJ; hence, since any two partitions have a common refinement, 9 depends
only on hand k and the curve CfJ, and we show next that it is independent ofCfJ for given
hand k. Namely, let cj: be another curve from h to k, close to CfJ, partitioned by points
90 ( = go = h), 91' ... , 9z ( = gl = k) such that 9; is close to gJor each i, and call
9= 'P(goI91)'P(9jI9z )'" 'P(9z-\gz)'
Then 9 can be obtained from 9 by inserting suitable factors, for instance,
'P(g1 191)'P({jI-l g1 )
(which cancel) between the first two factors of 9 and by then noticing that
'P(9jlgl)'P(gjlgZ)'P(g119z) = 'P(9jI 92 )
In this way it is seen that 9 = g, that is, that [] is unaltered by a continuous deforma-
tion ofthe curve CfJ, keeping the endpoints hand k fixed. Lastly, if G is simply connected
any two curves from h to k are homotopic; hence 9 depends only on hand k, so that we
can write [] = [](h, k). Furthermore,
[](gh, gk) = [](h, k) for any 9 in G,
since the same is true of each factor 'P(g;-l g ;+ 1)' It follows that the mapping .p of G
into G given by .p(g) = [](1, g) is the required extension of the local homomorphism
'P, because
<P(?1tgz) = [](l, 91)[](91' glg2) = [](1, gl)[](1, g2),
by addition of curves in G.
25.14 Law of Homomorphism for Lie Groups

The law of homomorphism for Lie groups is the same as that for abstract
groups, except that it has additional topological aspects.
A subgroup of a Lie group G is called closed if its elements constitute a
closed point set in the manifold of G.
156 Lie Groups
Theorem 1. The kernel ofa Lie group homomorphism 'P ofG onto G, that is,
the set
Go = {g E G: 'P(g) = 1 (identity of G)}
is a closed normal subgroup of G.
It is a normal subgroup by elementary group theory, and it is closed,

because of the continuity of the mapping 'P: If {gi} is a sequence in Go that
converges to h, then 'P(gi) converges to 'P(h), but 'P(gi) = 1 for every i;
hence 'P(h) = 1, hence h is in Go.
The next two theorems show that if Go is a closed subgroup of G (whether
normal or not), then logarithmic coordinates can be so chosen in G that,
when restricted to Go, they are logarithmic coordinates in Go, whereby Go
becomes a Lie group in its own right, and its Lie algebra is a subalgebra of the
Lie algebra of G.
If C(/: get) (0 ::::; t ::::; s) is any smooth curve emanating from the identity
(g(O) = 1) and lying in the subgroup Go, then the tangent vector to C(/ at 1 is
called a tangent vector to Go at 1.
Theorem 2. Let Go be a closed subgroup ofG. Then the set Ao of all tangent
vectors to Go at 1 is a subalgebra of the Lie algebra ofG. If Go is a normal
subgroup, then Ao is an ideal.
PROOF. Suppose A is in Ao. Then there is a smooth curve get) that lies in Go and is
such that
log(g(t = At + .. "
where the dots denote terms of order t2 . For each positive integer m, g(tlmr is in
Go, and, by the CBH formula,
where now the dots denote terms of order t 2 lm. By letting m -> 00, we see, since Go
is a closed set, At is the coordinate of a point in Go; hence, for A in A o , eAis in Go.
lfA, and A2 are both in A o , then, by the CBH formula, the vector
log(eAlteA2t) = teAl + A2) + ttZ[A" AzJ + ...
is the coordinate of a point in Go. By an argument similar to the one above, we see
from the linear terms that A, + A2 is in A o, and more generally so is tAl + sA2 for
real t and s, so that Ao is a subspace. In a like manner we see from the quadratic
terms that [A" A2 J is in Ao; hence Ao is a sub algebra. Lastly, if Go is a normal sub-
group, only one of A, and Az need be in Ao , and we see that Ao is an ideal.
It is now clear how to define coordinates in Go. Let lOb . , En be a basic of

the vector space A such that lOb , 10k (k < n) is a basis of Ao. For any A,
we write 1..= 1.. 1 101 + ... + AnIOn, and then Ai, ... , Ak can be taken as co-
ordinates in Go in a suitable neighborhood of 1. To find such a neighborhood,
let N be a neighborhood of 0 in the Lie algebra A of G in which logarithmic
coordinates can be used. Then the intersection N n Ao is an open set in Ao ,

and hence contains a neighborhood No (connected open set) of 0 in Ao.
The set Uo in the group Go given by
Uo = {g E Go: log g E No}
and the mapping <Po(g) = (A 1 , ... , Ak) ofU o into IRk define a chart {U o, <Po, No}
in Go. It, and the other charts obtained from it by translation in Go, as
described in Section 25.11, are called inherited (from G). In this way we arrive
at the following theorem.
Theorem 3. If Go is a closed subgroup of G, and if A and Ao are as above,

then Go, with its inherited charts, is a Lie group, and Ao is its Lie algebra.
The next two theorems show that, if Go is a closed normal subgroup, then
the factor group GIGo can be endowed with a manifold structure which makes
it into a Lie group; the Lie algebra of this group is isomorphic to AIAo;
the natural homomorphism of G onto GIGo is analytic with respect to this
manifold structure. The last theorem of the section is the homomorphism
law itself.
It is easily seen that the last n - k components Ak + 1, ... ,An oU with respect
to the basis E 1, ... , En described above can be taken as coordinates in GIGo,
for they are constant in each coset (of Go in G), or more precisely in the inter-
section of the coset and a suitable neighborhood mof 1 in G in which loga-
rithmic coordinates can be used. Namely, if eA' = eAe", where Ji is in Ao ,
so that eA' and e A are in the same coset, then, by the eBB formula,
A,' = A, + Ji + 1[A" Ji] + ... ;
all the terms on the right, starting with Ji, are in Ao, because Ao is an ideal;
hence A,' and A, have the same last n - k components. It is also clear that these
last components are different (i.e., at least one of them is different) in different
cosets. In these coordinates, the product and inversion functions m(,)
and 1() in GIGo are continuous (in fact analytic): For any product eAe A' = e A"
in G, all n components of A" depend continuously on all the components of A,
and of A,'; hence in particular the same is true of the last n - k components,
which are the coordinates of the corresponding cosets. The natural homo-
morphism of G onto GIGo, in which a group element e A is mapped onto the
coset in which it lies, is continuous in these coordinates, for the mapping
consists in simply ignoring the first k components of A,.
In this way the following theorem is established.
Theorem 4. Let G, Go, A, A o , E1, ... ,En , A1, ... ,An be as above. Then
Ak+ 1, ... , An can be taken as coordinates <peg) (g denotes a coset) in a subset
IT of GIGo (consisting of cosets that intersect the neighborhood 'l5 in G),
thus defining a chart which makes GIGo an (n - k)-dimensional Lie group,
called the Lie factor group (it is also denoted by GIGo) of G with respect to
Go. The natural homomorphism of G onto GIGo is continuous, and hence is
a Lie group homomorphism (hence is analytic).
158 Lie Groups
The following example shows that the conclusions of the theorem need
not hold if Go isnot a closed subgroup: Let G be the 2-dimensional torus group
consisting of matrices
(e~a e~P) (IX, f3 real).
Let Go be the subgroup of matrices
(~t e~e) (t real),
where eis a fixed real irrational number. The manifold of G is a torus, and that
of Go is a helical curve everywhere dense on the torus. Go is a normal subgroup,
but is not closed. The Lie algebra A is a plane, and, under the mapping
e" --+ A of G onto A, Go maps onto a set of parallel straight lines dense in the
plane. Ao is that line of the set that passes through the origin. With respect
to the basis (1' 2) referred to, where now 1 lies in Ao, the second coordinate
A2 has different values on different lines of the set that constitutes Go, and
since any neighborhood of the origin intersects infinitely many of the lines,
,.1,2 is not constant in Go or in any coset.
The last two theorems are now stated without proof.
Theorem 5. The Lie algebra ofG/G o is isomorphic to AjAo.
Theorem 6 (Law of Homomorphism). If Go is the kernel of a Lie group

homomorphism of G onto G, then
G -
-~G.
Go
In other words, if \)' is the isomorphism (in the sense of elementary group
theory) given in Section 18.8, then \)' and its inverse \)' - 1 are analytic
mappings.
The first seven of the exercises below deal with the somewhat elusive
question, mentioned in Section 25.1, as to which Lie groups have faithful
representations, and hence are linear groups, i.e., can be regarded as groups
of matrices (or of the corresponding linear transformations), as is normally
true of the groups that appear in applications. Every Lie algebra A has at
least one representation, the so-called adjoint representation of A on itself
(Exercise 2). By means of the exponential mapping, this gives a local repre-
sentation of a Lie group G on its algebra A. This mayor may not be extendable
to a representation on all of G, and, if it can, it mayor may not be faithful.
Exercise 8 deals with covering groups. SU(2) is the universal covering
group of SO(3). The basic theorem of Section 24.3 on covering manifolds
shows that there is no group not isomorphic to SU(2) that covers SU(2).
Hence, as stated in Section 21.1, there are no multivalued representations of
SO(3) except the 2-valued ones.
EXERCISES
The center e of a group G is the set of those group elements that commute with
every group element, i.e.,
e= {g E G:gh = hgYh in G}.
Similarly, the center Z of a Lie algebra is the set of those elements that commute with
every element in the algebra, i.e.,
1. Show that the center of a Lie group is a closed normal subgroup, and the center
of a Lie algebra is an ideal.
2. Recall that, for any J,., in A, Ad). is the linear transformation 11 --> [A,I1J of A
into itself. Show that if the Lie product of two such transformations is defined in the
usual way
whereupon the set {AdA: J,., E A} of all such transformations becomes a Lie algebra,
then the mapping J,., --> Ad). is a homomorphism of A onto this new algebra. This homo-
morphism is called the adjoint representation of A (on itself).
3. Show that if A is center1ree, which means that Z = {O}, then the adjoint
representation is faithful (i.e., the above homomorphism is an isomorphism).
4. If G is simply connected, then the local homomorphism e). --> eAd). of G onto a
group of linear transformations in A, discussed in Section 25.7, can be extended to a
homomorphism of all G, called the adjoint representation of G on A. Show that a necessary
condition for this homomorphism to be an isomorphism is that G be center1ree, which
means that e = {I}; then the homomorphism is locally an isomorphism.
5. To illustrate Exercise 4, let G be SU(2) and take the 2 x 2 matrices Tj , Tz , T3
defined by (21.2-4) as a basis of the Lie algebra A of SU(2). Then, the transformations
Ad). are represented by 3 x 3 real matrices. Show that the matrices eAdA constitute the
group SO(3) and that the homomorphism e). --> eAd). is then the familiar 2-to-l homo-
morphism ofSU(2) onto SO(3). What are the centers ofSU(2) and SO(3)?
.6. Show that if e is the center ofa group G, then the factor group Gle is not neces-
sarily center-free by considering the finite group
G = {1, i, j, k},
where i,j, k are the quaternion units, which satisfy the equations
i2 = / = k2 = - 1,
ij = -ji = k, jk = -kj = i, ki = -ik = j.

7. Show that the center Z of a Lie algebra A is an ideal, and that the factor algebra
A/Z is center free.
8. Let G be a connected Lie group, let Wl' be the universal covering manifold of
the manifold ml of G, and let IjJ be the projection ofWl' onto Wl. Wl' is made into a group
G', called the universal covering group of G, by defining a multiplication in it, as follows:
First, choose one of the points in ml' that lie over the identity 1 in G, and call it l' [hence
1jJ(l') = 1; l' will be the identity in G']. Then, let g' and h' be any two points of Wl',
and let g'(s) and h'(s) be curves in Wl' connecting g' and h', respectively, with 1', so that
160 Lie Groups
g'(O) = h'(O) = 1', while g'(I) = g' and h'(1) = h'. Let g(s) and h(s) be the projections of
g'(s) and h'(s) down into Wl; i.e., g(s) = tjJ(g'(s and h(s) = tjJ(h'(s. Then, k(s), defined
as =g(s)h(s), is a curve in Wl starting at 1. Let k'(s) be the curve in Wl' that results from
lifting k(s) up to Wl' in such a way that k'(O) = 1'. (See Section 24.2). Then, the product
g'h' in Wl' is defined to be = k'(I). Show that this definition is consistent (i.e., independent
of the choice of the curves; recall that Wl' is simply connected) and that it makes G'
into a Lie group. Show that the projection tjJ is a Lie group homomorphism of G' onto G.
Show that if g~ lies over 1 [i.e., iftjJ(g~) = 1; i.e., if g~ is in the kernel of the homomorphism
just referred to], then g~ commutes with every h' in G'. Hint: Choose the defining curves
in Wl' in such a way that h'(s) = l' for 0 :s; s :s; 1while g~(s) = g~ for 1 :s; s :s; 1.
25.15 Direct and Semidirect Sums of Lie Algebras
The concepts introduced in this section are analogous to the direct and
semidirect products of groups, defined in Section 22.9, in connection with the
crystallographic space groups. Suppose that a Lie algebra can be decomposed
as the direct sum (in the vector space sense) of two subspaces A1 and A 2 ;
i.e., any A in A can be uniquely decomposed as A1 + A2, where A1 is in A1 and
A2 is in A 2 . Suppose, further, that [A 1, A2] = 0 for every A1 in A1 and A2 in A 2 .
Then A1 and A2 are both ideals in A, and A is said to be their direct sum.
Now suppose that A is the direct sum (in the vector space sense) of Ao
and M, where Ao is an ideal, while M is merely a sub algebra. Then, for A1
and A2 in Ao and Ji1 and Ji2 in M,
The first three terms on the right are all in Ao (because Ao is an ideal) and can
be rewritten as
[A1, A2] + Ad,,! A2 - Ad"2 A1>
while the fourth term, [Ji1, Ji2], is in M.

The same result is obtained if we start with Lie algebras Ao and M and then
construct A, after first noting some of the properties of the linear transforma-
tions Ad". Each of them transforms the ideal Ao into itself, and the mapping
Ji-> Ad"
is a representation of M on Ao, according to Exercise 2 of the preceding
section, because, for Ji and v in M,
Ad[.... l = Ad"Ad v - AdvAd".

For fixed Ji, the transformation Ad" is a derivation, that is,
Ad,,[A 1, A2] = [Ad" A1> A2] + [A1' Ad" A2].

[In any algebra, a derivation is a linear transformation p such that p(x y) = 0
p(x) 0 y+ x 0 p(y), where the circle denotes the multiplication in the algebra.]
Direct and Semidirect Sums of Lie Algebras 161
Now let Ao and M be any given Lie algebras, and let the mapping
be a representation of M by derivations in Ao. A Lie algebra called a semidirect

sum of Ao and M, and denoted Ao EBp M, is defined as follows: As a vector
space, it is the direct sum of Ao and M, so that its elements are ordered pairs
{A, J.l}, where A is in Ao and J.l is in M, and a Lie product is defined in it as
This product is obviously linear in each factor and antisymmetric.
EXERCISE
1. Show that the product just defined satisfies the Jacobi identity.
If A = Ao EBp M, and if Ao i~ identified with the set of elements of the

form {A,O} and M with the set of elements of the form {O, Il}, then p(J.l) is
just the transformation Ad" in A, because
Ad{o. It) {A, O} = [{O, J.l}, {A, O}] = {p(J.l)A, O}.
EXERCISE
2. Let Go and Hbeclosed subgroups ofa Lie group G, where Go is normal. Assume
that each 9 in G has a unique representation as goh, where go and h are in Go and H,
respectively. Let A, Ao, and M be the Lie algebras of G, Go, and H, and show that
A = Ao ffip M, where, for any 11 in M, P(I1) = Ad~. Hint: For A. in Ao and 11 in M, let
fA., 11} denote 10g(eAe~) and find the Lie product of two such curly bracket expressions
by applying the expansion of the CBH formula to
and to each factor separately.
The semi direct sum is a direct sum Ao EB M if p is the trivial homo-

morphism which maps every J.l in M onto the zero transformation; i.e.,

p(J.l)A = for all A. Ao and M are both ideals in Ao EB M.
A fundamental structure theorem, proved at a much later stage of the
theory, says that any Lie algebra can be written as a repeated semidirect sum
of Lie algebras, each of which is either 1-dimensional or simple; hence, a

main objective of the theory is a classification of the simple algebras. The
theorem holds for both real and complex Lie algebras; its proof, which is
quite deep, is found in Hausner and Schwartz 1968.
162 Lie Groups
25.16 Classification of the Simple Complex Lie

Algebras
The relations of the various objects in the theory are indicated in the following
schema:
simple simple
real complex complexEreal
Lie Lie Lie Lie Lie ~ Lie
group --> algebra --> algebra algebra algebras groups
Each Lie group determines a unique real Lie algebra, which in turn deter-
mines a unique complex Lie algebra, by a process called complexification,
described below. The complex case is simpler than the real case, just as in
elementary matrix theory, because the complex number system C is alge-
braically closed, while ~ is not. (It is recalled that a real matrix generally
has complex eigenvalues and eigenvectors.) There exists a complete classifica-
tion of the simple complex Lie algebras into four main series of algebras and
five so-called exceptional algebras. The next step is to find all the simple
real algebras whose complexification leads to a given complex algebra.
This step is carried through in Hausner and Schwartz 1968, where the reader
can find a complete classification of the simple real algebras. The result is
considerably more elaborate than the classification of the complex algebras,
but it is still two steps removed from a classification of the Lie groups; for
this, one must first find all possible repeated indirect sums of I-dimensional
and simple algebras, as described at the end of the preceding section, and
then find all (say connected) Lie groups that yield a given real Lie algebra.
We shall sketch the development very briefly through the classification of
the simple complex algebras. For the algebraic details and the many lemmas
needed for the proofs, the reader is referred to Hausner and Schwartz 1968.
As indicated in the preceding section, we are mainly interested in the simple
algebras, but, in the analysis of them, certain nonsimple algebras appear,
namely the semi simple, solvable, and nilpotent Lie algebras. To define those,
we note first that if Al and A2 are any ideals in a real or complex Lie algebra
A, then [Ab A2], defined as the subspace spanned by elements of the form
[Al' A2], where Al is in Al and A2 is in A2, namely, the subspace
[Al' A2] = span{[Al' A2]: Al E A1, A2 E A2}
is an ideal contained in both Al and A2. We then define two descending
sequences of ideals in A, namely,
Al = A :::J A2 :::J A3 :::J
and
inductively by
N+l = [A, N],
Classification of the Simple Complex Lie Algebras 163
A is said to be solvable if A(k) = 0 for some k and nilpotent if Ak = 0 for some k.

It is easy to see that a nilpotent algebra is solvable; in fact, Nk) c A\ for
all k, by induction on k.
As in Section 25.12, a Lie algebra A of more than one dimension is simple
if it contains no proper nonzero ideals. It is called semisimple if it contains
no proper nonzero solvable ideals (in which case A itself cannot be solvable,
so the word "proper" can be omitted in this last definition).
It turns out, at a considerably later stage of the development, that an
algebra A is semisimple if and only if A2 = A (hence, we require that dim A> 1,
for if dim A = 1, then A2 = 0), also if and only if it can be written as a direct
sum of ideals
A = At EB ... EB Ako
where each Ak is a simple algebra.
If A is a real Lie algebra, its complex(fication is defined as the complex
Lie algebra Awhose elements are formal sums A + iJ!, where A and J! are in
A, and in which the linear combinations and Lie products are defined in the
obvious way; in particular,
[At + iJ!l> 1..2 + iJ!2J = [At, A2J + i[J!t, A2J + i[Al> J!2J - [J!t, J!2l
A. is semisimple if and only if A is semisimple. If A is simple, then its complexifi-
cation is either simple or is the direct sum of two identical (i.e., isomorphic)
simple complex algebras.
Any real or complex Lie algebra A contains nilpotent subalgebras (they
are, of course, not ideals, if A is semisimple); in particular it contains a so-
called Cart an sub algebra M, defined below, which is nilpotent. To investi-
gate the structure of A, one investigates the structure of M and the
relationship between the elements of M and the other elements of A. That
relationship is described by the operators Ad", J! EM; Ad" transforms an
element of A into another element of A, namely A into [J!, Al The mapping
J! --> Ad" is a representation of M on the vector space A; hence the theory starts
by considering general representations of solvable and nilpotent Lie algebras.
The study of a representation P of an abstract algebra A has the merit
that while A in A is an abstract object, p(A) is a linear transformation in a
vector space, and standard linear algebra can be applied; for example, in
the complex case, the transformation p(A) has at least one eigenvalue and one
eigenvector. Also, the Lie product of p(A) and p(J!) is simply p(A)p(J!) -
p(J!)p(A).
If p is a representation of any Lie algebra A on a vector space V, we call v
in V a weight vector of p if it is a simultaneous eigenvector of all the trans-
formations p(')..), A E A, i.e., if
p(A)v = IX(A)v, VAE A,
where IX( . ) is a numerical-valued function, obviously linear, defined on A,
called the corresponding weight of p. A vector v in V is a generalized weight
vector of p corresponding to the weight IX( . ) if, for some integer k,
(p(A) - IX(A)I)kv = 0, VA E A,
164 Lie Groups
I being the identity transformation in V. The set of all generalized weight

vectors for given (X( . ) is called the corresponding weight space and is de-
noted by Va. Thus, weight, weight vector, and weight space correspond to
eigenvalue, eigenvector, and algebraic eigenspace, in the case of a single linear
transformation. In the latter case, if V is a complex vector space, V is the direct
sum of all the algebraic eigenspaces VI EEl ... EEl l'k corresponding to the eigen-
values Zl' .. , zk-this corresponds to the fact that any matrix can be put
into Jordan normal form. Analogous results hold for weights and weight
vectors if the Lie Algebra in question is solvable or nilpotent.
Theorem 1. If P is a representation of a solvable complex Lie algebra M on

a vector space V, then p has at least one weight vector v and corresponding
weight (X( . ).
Under the further assumption that M is nilpotent, we have the following:
Theorem 2. If p is a representation of a nilpotent complex Lie algebra M

on a vector space V, then the weight spaces of p span V as a direct sum
V = Va, EEl ... EB Va k ,
where Va j is the weight space that corresponds to the weight (Xi, ),j = 1, ... , k.
In each case the proof proceeds by induction on the dimension n of M;

the ideal M2 is ::pM, and if N is a subspace of M of dimension n - 1 that
contains M2, then N is a subalgebra (in fact, an ideal) and is solvable in the
case of Theorem 1 and is nilpotent in the case of Theorem 2; hence the
inductive hypothesis can be applied to N. The induction starts for n = 1,
in which case there is only one linear transformation peA) involved up to a
scalar mUltiplier, and the statements in the theorems reduce to the corres-
ponding known facts of linear algebra. The algebraic work in the proofs is
straightforward, but its quantity is enough to discourage all but the strong
at heart.
Next, two important tools in the analysis of a general Lie algebra A are
introduced: the symmetric bilinear form and the notion of a Cartan subalgebra.
The first is the symmetric form (A, J1) defined for all A and J1 in A by the
equation
(A, J1) = tr(Ad" Ad..),
where "tr" denotes trace. It is real- or complex-valued, according as A is a
real or complex Lie algebra, but is not positive definite, except in a special
case mentioned below. A basic theorem, Cartan's Criterion, says that a real
or complex Lie algebra is semisimple if and only if the symmetric bilinear
form is nonsingular, which means that for no A ::p 0 is (A, J1) = 0 for all J1.
If A is any complex Lie algebra and M is a nilpotent subalgebra, we can
apply Theorem 2 to the adjoint representation of M on A, so that the symbols
p(J1) and V are replaced by Ad" and A. The weights, weight vectors, and
Classification of the Simple Complex Lie Algebras 165
weight spaces of this representation are then called roots, root vectors, and
root spaces of M in A. If et( . ) is a root, the corresponding root space is de-
noted by Aa; it is a subspace of A. From the nilpotence of M it follows that
the zero function, et(A) = 0 for all A, is one of the roots, and the corresponding
root space, called AD, contains M. If the nilpotent sub algebra M can be so
chosen that AD is = M, then M is a Cartan subalgebra of A. A basic theorem
that every complex Lie algebra has a Cartan subalgebra.
It turns out that if A is a complex semisimple Lie algebra, then (a) the
Cartan subalgebra M is commutative, (b) for each et f= 0, the root space
Aa is one-dimensional, (c) if et is a root, - et is also a root, and (d) if A and A'
are nonzero vectors in Aa and A-a, then [A, A'] is a nonzero vector in M,
and (A, A') f= O. We number the nonzero roots etl> et z , ... , et k ; we choose
vectors Ai and A-i in Aa, and A-a" so normalized that (Ai, Il-J = 1, and we
call
i = 1, ... , k.
It can be shown that the vectors Ili span M.
It follows from (a) and (b) in the preceding paragraph that for a semisimple
algebra, only ordinary root vectors (i.e., no generalized ones) appear. For
the root vectors Aa (et f= 0), that follows from the one-dimensionality of Aa;
and every vector v in AD = M is a root vector, because Ad" v = 0 for all Il in M.
The Cart an subalgebra is not unique, but it can be shown that if M' is
any other Cartan sub algebra in A, then M and M' have the same dimension,
and there is an automorphism of A that carries M onto M'; hence, either of
them can be used to investigate the structure of A.
It is found that the configuration of the vectors Ili completely determines
the Lie algebra. The description of this configuration is greatly simplified by
the fortunate fact that if Mr denotes the real vector space consisting of
linear combinations of the Ili with real coefficients, then the natural bilinear
from ( ., . ) is real and positive definite in Mr; hence, Mr is a Euclidean space,
if (-, .) is taken as the scalar product. It can be shown that the real dimension
of Mr is the same as the complex dimension of M, and we call it m. The
(complex) dimension of A is then m + 2k. The length of a vector Il in Mr is
111111 = (Il, ll)l/Z, and the angle between two such vectors is
(Il, v)
cos L fl, v = IfJilnv~
The star in M r consisting of the vectors Ili' thought of as radiating out from
the origin, has a rather high degree of symmetry and can be described in the
following terms: (1) For any given simple algebra A, either all the Ili have the
same length, or there are just two lengths, some of the Ili being short, and
the others long. (2) The angle between any two of the vectors is an integer
multiple of 30 or 45. (3) If the angle is 30 or 150, one vector is long and the
other short; the ratio of the lengths is J3.If the angle is 45 or 135, the ratio
ofthe lengths is fl. If the angle is 60 or 120, the two vectors have the same
length. (4) The entire star is symmetric with respect to reflection in each
166 Lie Groups
hyperplane perpendicular to one of the Jli' Every minimal star that satisfies
these conditions determines a unique simple complex Lie algebra, and dif-
ferent stars determine different algebras. If A is merely semisimple and is a
direct sum A = Ai EB ... E8 Ak of simple algebras, then Mr is spanned by k
mutually orthogonal subspaces, each containing the star of one of the simple
algebras.
We now assume that A is simple. When Mr is one-dimensional, the star
consists of two opposed vectors of equal length, and the algebra A, called
At. has dimension I = 3. When Mr is two-dimensional, there are three pos-
sible stars, shown in Figure 25.2 together with the designations and dimen-
sions I of the corresponding algebras, which are called A 2 , B 2 , and G2
When M, is three-dimensional, there are again three possible stars, corres-
ponding to algebras called A 3 , B 3 , and C 3 . The star of A3 consists of six
pairs of opposed vectors Jli and JI-i' all equal in length, and extending from
the origin to the midpoints of the edges of a cube; the angles that occur are
60, 90, 120, and 180. In the star of the algebra B3 there are six pairs of
long vectors, arranged as for A 3 , extending to the edges of a cube, and three
pairs of mutually orthogonal short vectors making angles of 45 with the
nearest long ones, the length ratio being -Ii;the short vectors extend to the
midpoints of the faces of the cube referred to. The star of C 3 is the same as
that of B 3 , but with the long and short vectors interchanged, so that the star
fits into a rhombic dodecahedron. The dimension number I is equal to 8, 10,
14,15,21, and 21, for the algebras A 2 , B 2 , G2 , A 3 , B 3 , and C 3 , respectively.
It is of course meaningless to talk about the lengths and directions of the
vectors Ai, because they lie in the complex space A, for which ( " .) is not
even a Hermitian inner product. What is meaningful is to find the Lie pro-
ducts [A, JI] for sufficiently many pairs A, JI so as to determine the structure
of A. That is best done by means of the models described in the next section.
To determine the possible stars, when Mr is of more than 3 dimensions,
one makes use of a device due to the Soviet mathematician E. B. Dynkin. A
simple set of vectors in the star is a certain set n of just m of the vectors Jli
(m is always < 2k) such that all the vectors of the star can be obtained by
repeated additions and subtractions, starting with the vectors of n, and
such that only one set of vectors satisfying the conditions (1)-(4), above, i.e.,
Figure 25.2 The 2-dimensional stars of simple complex Lie algebras.

Models of the Simple Complex Lie Algebras 167
only one star, can be obtained in this way. It can be proved that it is always
possible to choose a simple set of vectors. Furthermore, although the set n
is not in general unique, if n' is another simple set, then there is an auto-
morphism of A under which M is invariant and n is carried into n'; hence
it is unimportant which simple set is used. The possible angles between any
two vectors of n are 90,120,135, and 150. A Dynkin diagram is a set ofm
points or small circles on a plane, one for each vector in n. If the angle
between two vectors in n is 120, 135, or 150, then the corresponding
points of the diagram are joined by a single, double, or triple line respec-
tively; if the angle is 90, the points are not directly connected. If the angle is
135 or 150, the point corresponding to the shorter vector is indicated by an
asterisk. Then, a number of things can be proved about the diagrams of
simple complex algebras; for example, a diagram can contain no loops, it is
connected, it can contain at most one double or triple line, it can have at
most one branching, and so on. In consequence of these rules, it is found that
there can be just seven types of Dynkin diagrams, as follows (m is the number
of points and is equal to the dimension of M, lis the dimension of A):
type Am(m ~ 1) 0-0 ... 0-0 1= m(m + 2)

Bm(m ~ 2) 0-0 ... C>==G m(2m + 1)
Cm(m ~ 3) 0-0 ... 0=0 m(2m + 1)
Dm(m ~ 4) o-o .. ~ m(2m - 1)
Em(m = 6, 7, or 8) o-o ...~ 78,133,248
F4 ()--()==0-0 52
G2 ()==(V 14
Figure 25.3 Dynkin diagrams for the simple complex Lie algebras.
Types Am, B m, em, and Dm constitute the regular series, and the remaining
five algebras are called exceptional.
25.17 Models of the Simple Complex Lie Algebras
The above classification results from imposing various conditions, arrived at

by very lengthy algebraic considerations, which a simple complex Lie
algebra must satisfy. To show that there are no further conditions, and hence
that all the above algebras actually exist, models of them are constructed.
The models of the regular series are algebras of matrices, defined below. We
168 Lie Groups
continue to denote the elements of the algebras by the symbols A, p, ... ,even
though other symbols might seem more appropriate for matrices.
1. Am consists of all (m + 1) x (m + 1) complex matrices of trace zero.
See exercises below. For Bm and Dm it is necessary to introduce the
antidiagonal matrix
J = p x p matrix(t)i).
1--- 0
2. Bm consists of all (2m + 1) x (2m + 1) complex matrices)" such that
)..J + J).. T = 0 (p = 2m + 1).
3. Dm consists of all 2m x 2m complex matrices).. such that AJ + J).. T = 0
(p = 2m).
For em it is necessary to introduce the 2m x 2m anti diagonal matrix
o
J' =
-1
---(,/ o
4. em consists of all 2m x 2m complex matrices).. such that ,,-J' + J').. T = O.
The following exercises concern the series Am. The series Bm, em, and Dm
are similar. The models of the exceptional algebras are more complicated and
are given in Hausner and Schwartz.
EXERCISES
1. Let N be the Lie algebra of (m + 1) x (m + 1) complex matrices A" with
[A" J1] = A,J1 - J1A,. Compute the natural bilinear form
(A" J1) = tr(AdAAd,J
[Note that AdA and Ad~ are linear transformations in an (m + 1)2 dimensional space,
namely N.] Show that
(A" J1) = 2m + l)tr(A,J1) - (tr A,)(tr J1.
Show that (A" J1) is singular in N but nonsingular in the subalgebra A = Am of matrices
of trace zero, so that Am is semisimple.
2. Let M denote the commutative subalgebra of the Lie algebra A of Exercise 1
consisting of diagonal matrices (with trace zero). Consider the adjoint representation
of M on A: J1-> Ad~, where
(Ad~A,)rs = (Il" - Ilss))"rs (r, S = 1, ... , n).
Consider the roots and root vectors of this representation. Show that the root space Ao ,
which consists of all matrices A, such that (Il" - Ilss)k Ars = 0 for some k, for all J1 in M,
consists also of the diagonal matrices; hence Ao = M, hence M is a Cartan subalgebra.
Show that the other roots cx() and corresponding root vectors A,a are obtained by choos-
ing fixed i and k and setting
cx(J1) = Ilii - Ilkko
A,a = A,(i, k),
Models of the Simple Complex Lie Algebras 169
where A(i, k) is the matrix
and that the vector Il. is the diagonal matrix given by

1
(Il.);; = 2(m + 1)'
-1
(Il.)kk = 2(m + 1)'
otherwise (Il');'k' = o.
Show that a simple set of roots is the set
i = 1, ... , m,
Show that the angle between Il., and Il., + 1 is 120 and that otherwise the angle between
Il., and "' j is 90, so that the Dynkin diagram of A is as given above for Am' namely,
0-0---. ,..-0 (m small circles).
For the classification and models of the simple real Lie algebras, which are needed
for a classification of Lie groups, the reader is referred to Hausner and Schwartz. It is
recalled that if A is a simple real Lie algebra, its complexification Ais either simple or is
the direct sum of two identical (i.e., isomorphic) simple complex algebras, Hence, to
classify the simple real algebras, one must examine each simple complex algebra and then
find all simple real algebras from which it can be obtained by complexification.
Given a simple complex algebra A, one possible choice of A consists of the elements
of A but regarded as a linear space over the real field ~, rather than C, as the field
of scalars, but that is not the only possibility. Other possibilities are found by con-
sidering the so-called conjugations in A. A conjugation in a complex Lie algebra is
an antilinear mapping C [that is, C(aA + bll) = ilCA + bCIlJ, which preserves Lie
products (that is, C[A, IlJ = [CA, CIlJ), and whose square is the identity mapping
[that is, C(CA) = A]. The set of all A in A such that CA = A, with ~ as the field of
scalars, is a simple real Lie algebra. A complete analysis of the conjugations in the
simple complex Lie algebras, and the enumeration of the resulting simple real algebras,
is given in Hausner and Schwartz. As an example, if A is the simple complex algebra
A 1 of 2 x 2 matrices of trace zero, there are three corresponding simple real algebras,
namely A 1 itself (with ~ as the field of scalars) and
RA 1 = {2 x 2 real matrices of trace zero}
and
QA 1 = {2 x 2 matrices of the form iH, where H is Hermitian and of trace zero},
We mention in passing that some of the corresponding Lie groups are SL(2, 1[:), 2p,
SL(2, ~), SU(2), and SO(3).
Corresponding to each simple cpmplex algebra Am' for m > 1, there are
4 + [em + 1)/2J simple real algebras, where [ J denotes integer part.
Corresponding to the exceptional algebra G z there are three real algebras, called G z
(over ~), HGi3 ), and HG~l).
170 Lie Groups
25.18 Note on Lie Groups and Lie Algebras in

Physics
The role of the rotation group SO(3) as a symmetry group in quantum

mechanics was discussed in Chapter 22. In calculations, the corresponding
Lie algebra usually appears, rather than the group itself. The Lie algebra
of SO(3), which is of course the same as the Lie algebra of SU(2), the uni-
versal covering group of SO(3), may be realized either as the Lie algebra of
3 x 3 real skew symmetric matrices or that of2 x 2 skew Hermitian matrices
of trace zero, called QA 1 in the preceding section. It may also be realized
as a Lie algebra of operators in the Hilbert space t) of states of a physical
system. If t) is taken as the space L 2([R3) of wave functions rjJ(x) of a spin-
zero nonrelativistic particle, the mappings
[g E SO(3)]
form a representation of SO(3) on t), as in Section 20.9. The so-called in-
finitesimal operators of this representation were discussed in that section and
are
I a k a Ukl = 123,231,312).
L j = x 8x k - x axl
The corresponding self-adjoint operators ihL j are the angular momentum
observables. The linear combinations of the L j with real coefficients give a
realization of the Lie algebra A of SO(3).
In elementary particle physics, the relevant symmetry group is often not
known, owing to the lack of a complete theory of elementary particles, but
various Lie algebras A are often found to playa role, on empirical grounds.
Confusion can arise because the word "group" is often used for a Lie
algebra. In particular, one finds references to the "group G2 ." According to
Section 25.16, G2 is a Lie algebra, and in fact a complex Lie algebra. The
corresponding group in the physical theory is presumably a group whose Lie
algebra is one of the three real Lie algebras, mentioned in the preceding
section, whose complexification gives G2 The nearest that one can come to
identifying a unique group in such a case is this: Of the real Lie algebras A
whose complexification is a given algebra A, just one is the Lie algebra of a
compact group (or possibly of several compact groups), and of the Lie
groups G having a given real algebra A, just one is simply connected. Hence,
in particular, there is a unique compact simply connected Lie group associa-
ted with the algebra G2 Since many of the symmetry groups of physics are
neither compact nor simply connected, the identification of the group to be
associated with a given Lie algebra in particle physics must presumably
await further theoretical developments.
The algebra G2 has also been used in the study of atoms having partially
filled/shells; see Racah 1949. In that case the theory is sufficiently complete
that presumably there is a clearly defined corresponding group, which,
according to Racah (see also Behrends, Dreitlein, Fronsdal, and Lee 1962)
is a subgroup of SO(7).
Appendix to Chapter 25-Two Nonlinear Lie Groups 171
Appendix to Chapter 25-Two Nonlinear Lie Groups
In this appendix, we give two examples of Lie groups that are not linear, that
is, have no faithful finite-dimensional representations, hence cannot be
realized as groups of matrices.
For the first example, let G denote the so-called Heisenberg group
If the above matrix is denoted by gx,y,z, direct calculation shows that
g;'~,o = g-x,o,O'
-1
go,y,o = go,-y,O, (25.A-1)
-1 -1
gx,o,ogo,y,ogx,o,ogo,y,o = go,O,xy'
It follows that if p is any representation of G, then p(go. 0. z) is a unimodu-
lar matrix for every z, because det(p(go,O,XY)) is equal to
det p(gx,o,o)det p(go,y,o)det p(gx,O,O)-l det p(go,y,O)-l = 1.
Now let Go denote the normal subgroup
It will be shown that every finite-dimensional representation of the factor

group GIGo is nonfaithful; hence GIGo is not a linear group. Let gx,y,z, where
o S z < 1, denote the element of GIGo (a coset in G) that contains gx,y,z'
That is, gx,y,z is the infinite set
g""" ~ {G ~ z ~} ~ 0, t, 2,}
of 3 x 3 matrices. It is easily seen that, in analogy with (25.A-1),
- - --1 --1 -
gx,o,ogo,y,ogx,o,ogo,y,O = go,o,Z'
where z == xy (mod 1). Hence, as above, if p is any representation of GIGo,
then det p(g) = 1 for every g in the subgroup
G
H = {go , , z: 0 S z < 1} < -.
Go
But H is isomorphic to SO(2), with 2nz playing the role of (); hence H is
compact and Abelian. According to the general theory of representations
given in Sections 21.1-21.4, every representation of a compact group is
equivalent to a unitary representation, and every unitary representation of
172 Lie Groups
an Abelian group is completely reducible as the direct sum of I-dimensional

representations. Hence, if p is any m-dimensional representation of GIGo,
then, relative to a suitable basis in V m , the representation p of the subgroup
H ~ SO(2) has the diagonal form
_ (e21tinlZ . (0))
p(go,o.z) = .. .
(0) e21tinrnz
Each of the I-dimensional representations given by the diagonal elements

of this matrix has determinant = 1 for all g in H; hence all the integers
nl, ... , nm are zero. That is, all elements of the subgroup H are presented
by the m x m unit matrix; hence the representation p of GIGo is not faithful.
The second example is less elementary (at least, the present discussion of
it is), because use is made of a moderately deep theorem on representations
of Lie algebras. It will be shown that if G is the universal covering group of
SL(2, ~) (the group of 2 x 2 real matrices with unit determinant), then G
has no faithful finite-dimensional representation.
A canonical form for any matrix M in SL(2, ~) is now obtained. Let ~ be
the rotation [an element of the subgroup SO(2)J that transforms the first
column of M into a vector having components a, 0, where a > 0. Then RM
is of the form
RM = (~ ~), ac = 1.
We write a = eX, c = e-x.Then,
for some real y. Hence, if the rotation R -1 is Ro, we have
M= (
COS e (2S.A-2)
sine
This is the desired canonical form. It follows that the manifold of SL(2, ~)
is the Cartesian product of a circle and two lines, C 1 x ~2. Since the universal
covering of C 1 is ~, the manifold of Gis ~3.
The Lie algebra A of SL(2, ~) has as basis the matrices
L1 = aa~lo = (~ -~),
L2 = ~~Io = (~ _~),
L3 = ~;Io = (~ ~),
Appendix to Chapter 25-Two Nonlinear Lie Groups 173
where, in each case, the subscript zero indicates that the matrix in question is
evaluated at () = x = y = O. A direct calculation shows that
[Ll> L 2] = 2Ll + 4L 3 ,
[L 2, L 3] = 2L 3, (2S.A-3)
[L 3, L 1 ] = L 2
These equations can be solved for L 1 , L 2, and L 3; i.e., the Lie products on the
left also form a basis for A. It follows from the definition of the Lie product
in terms of commutators in the group that the group is generated by com-
mutators, i.e., by elements of the form ghg - 1 h - 1. By the argument used in
the first example, it then follows that if p is any representation, then det p(g) =
1 for all g in SL(2, IR).
Since G and SL(2, IR) are isomorphic in a neighborhood of the identity
element, they have identical Lie algebras, and A is also (in the sense of iso-
morphism) the Lie algebra of G. Hence, if p is any representation of G, it
follows that det peg) = 1 for all g in G. That is, any representation of G is
unimodular.
It can be shown that the Lie algebra A is simple. Namely, if A has an ideal
J that contains a nonzero vector A == aLl + bL 2 + cL 3 , then J also contains
the three vectors [Lj,A] and the nine vectors [Lk> [Lj,AJ]. A direct calcula-
tion, starting with (2S.A-3), shows that L 1 , L 2, and L3 can be expressed as
suitable linear combinations of those 13 vectors (it is not necessary to go to
higher Lie products); hence J coincides with A, i.e., A is a simple Lie algebra.
Let go. x. y denote the element (2S.A-2) of SL(2, IR), where 0 ::; (}::; 2n.
Then (), x, y with () unrestricted, can be taken as coordinates of elements
go, x, y in G in such a way that in the covering of SL(2, IR) by G, the element
gw,X,y of G lies over the element go,x,y of SL(2, IR) if ()' == () (mod 2n).
Now let p be a representation of G on an m-dimensional vector space
vm = em. It will be sHown that p is nonfaithful. A represent:;ttion of A, which
will also be called simply p, is induced in the usual way:
p(L 1 ) = ;() p(go,X,y)!o=x=y=o, etc.
According to Hausner and Schwartz, p. 143, Theorem 2, a representation

of a (real or complex) simple Lie algebra A is completely reducible; i.e., vm
can be written as a direct sum V k1 EE> V k2 EE> . of invariant subspaces on
each of which p is irreducible ('Lkj = m). The group G is generated by ele-
ments of the form eA (A E A), and p( e'l.) is = eP('I.); hence the representation p
of G is also completely reducible. Specifically, each of the subspaces V kj of
vm is invariant under peg), for all 9 in G, and the restriction pj of p to V kj is
irreducible. Let gl denote the element g21< 0 0; gl and all its powers lie over
the identity element of SL(2, IR), hence co'~mute with all 9 in G; hence, by
Schur's lemma, pigl) = AI, where I is the k j x k j unit matrix. Since every
representation is unimodular, det pig) = 1 for all g; hence Akj = 1. This is
true for eachj; hence there is some power gi of gl (e.g., I = nUl k) such that
p(gi) = P(gl)l = I (=m x m unit matrix). But gi is not the identity of G;
hence p is nonfaithful.
CHAPTER 26
Metric and Geodesics on a Manifold
Scalar, vector, and tensor fields; Lie brackets; covariant and contravariant
vectors; transformation laws; inner and outer multiplication; contraction;
quotient law; derivations; metric tensor; definite and indefinite metric;
Riemannian and pseudo-Riemannian manifolds; raising and lowering of
indices; geodesics; Euler variational equation; natural, affine, or preferred
parameter; Christoffel three-index symbols; spacelike, null, and timelike
geodesics; initial-value and two-point problems of geodesics; Volterra
integral equations; Picard iterations; Whitehead's theorem; continuation of
geodesics; affinely connected manifolds; Riemannian and
pseudo-Riemannian covering manifolds.
Prerequisite: Elementary theory of manifolds (Chapters 23 and 24).
A manifold, as defined in Chapter 23, is a thing characterized just by its local

topology: It is a locally n-dimensional space in which the Hausdorff separa-
tion axiom holds. In this chapter and the next two, a manifold is made into a
geometric structure by introducing further notions such as geodesics (a
geodesic is the analogue of a straight line in Euclidean geometry), lengths,
angles, and so on. The most fundamental notion is geodesic, which is derived,
in the main geometries of interest to physics, either from a metric or from an
affine connection; we start with a metric, because of its similarity with
distance in familiar Euclidean geometry.
Rougnly speaking, geometric properties contrast with the global topo-
logical properties of a given manifold in that the latter are expressed in terms
of integer-valued or discrete quantities like the number of components, the
fundamental group, and higher homotopy groups, whereas geometry involves
continuous real quantities like lengths and angles and the natural parameter
along a geodesic (explained in Sections 26.6, 26.7, and 26.12).
In Riemannian geometry, one starts with a metric differential form ds 2 =
L gjk dx j dx k in a given coordinate system, from which geometric ideas are
derived. A simple familiar example of a nonplane 2-dimensional geometry is
the geometry on the unit 2-sphere, where ds 2 = d()2 + sin 2 ed<p2 in terms of
the polar angles () and <po The geodesic (shortest curve) between two points is
174
Scalar and Vector Fields on a Manifold 175
an arc of a great circle, and so on. Various theorems of spherical geometry

follow; for example, the sum of the angles of a triangle (whose sides are
geodesics) exceeds n by an amount equal to the area of the triangle.
Although a general Riemannian manifold can in principle be regarded as
embedded in some Euclidean space EN of higher dimension, just as the
2-sphere can be regarded as the surface of the unit ball in E 3 , it is easier and
more appropriate to use intrinsic coordinates for the study of the intrinsic
geometric concepts determined by the metric.
If the metric form on a manifold Wl is positive definite (see Section 26.4),
then Wl is called Riemannian; if it is indefinite, Wl is called pseudo-Riemannian;
the latter case appears in general relativity. As will be seen, there are rather
profound differences between the two kinds of manifold. It will be assumed
that Wl is connected and of class C\ where k is large enough to accommodate
any derivatives that appear in the discussion-it suffices to take k = 4 in this
chapter and the next two. In relativity, it cannot be assumed that Wl is of
class Coo, because the metric is determined by the distribution of matter,
which is not necessarily infinitely smooth, and because the Einstein field
equations, since they are hyperbolic, can propagate discontinuities of various
derivatives of the gil v in the form of gravitational waves.
26.1 Scalar and Vector Fields on a Manifold

If xl, ... , xn are Cartesian coordinates in a Euclidean space, a vector field in
the space is described by n functions d(xl, ... , x n ), i = 1, ... ,n. If the
Cartesian coordinates are transformed by rotating the axes, then the com-
ponents v l , .. , vn of the vectors are transformed by the same rotation matrix
by which the coordinates Xl, ... , Xn are transformed.
If curvilinear coordinates are introduced, it becomes necessary to dis-
tinguish two kinds of vectors, covariant and contravariant, which have
different physical or mathematical origins. (Still further kinds, the so-called
vector densities of various order, are sometimes introduced; they are often
convenient, but not necessary.)
First, a function f(P) on a manifold Wl (real-valued, unless otherwise
specified) is called a scalar (field) on Wl. When f(P) is given for all points P in
Wl, then, as in Section 23.5, a function l(x l , ... , xn) = lex) is associated with
any given coordinate chart {U, <p, N} by the equations
lex) = f(P), x = <pep), for all P in U. (26.1-1)
This collection of functions l( ... ), one associated with each chart, can be
regarded as determining the scalar field f(P). If {U', <p', N'} is a second
coordinate chart, the relation between the corresponding functions land l'
in the overlap of the two charts is simply
lex) = l'(x'), (26.1-2)
where
x = <pcP) and x' = <p'(P). (26.1-3)
176 Metric and Geodesics on a Manifold
The interpretation of(26.1-2) is that it becomes an identity in xl, ... , xn if the

Xli in the right member are expressed in terms of the Xi, or an identity in
X,l, ... , XIH if the Xi in the left member are expressed in terms of the Xli; the
values of 1 and l' are the same at a given point P in 9Jl.
While a scalar field is a collection of functions 1<- .. ), one associated with
each chart, a vector field or a tensor field is a collection of sets of functions,
one set associated with each chart; in the overlap of two charts, the two
corresponding sets of functions are related by a transformation law which
generalizes the law (26.1-2) for scalars.
As a first example of a vector field, consider the gradient ofthe scalar f(P),
which is assumed to be of class C 1 . In the chart {U, cp, N}, this consists of the
partial derivatives of the function (26.1-1), namely,
1
Vi ( X , . , X
n) a f~(x, ... ,xn),
= --;----' 1 (26.1-4)
ux'
or, more concisely,
a ~
v;(x) = ox f(x)
i (i = 1, ... , n). (26.1-5)
If the relation established by the equations (26.1-3) in the overlap of two

charts is described by writing
(26.1-6)
(26.1-7)
as in Section 23.2, and if, in the second chart, one writes correspondingly
a ~
v;(x') = OX'i f(x'), (26.1-8)
then the relation between the two sets offunctions {v;} and {v;} in the overlap
of the two coordinate systems is
(26.1-9)
(Here, the summation convention is used, according to which the right

member is understood to be summed for k = 1,2, ... , n; k is called a dummy
index.) As in (26.1-1), the interpretation of (26.1-9) is that it is an identity if the
Xi are expressed in terms of the X,i, or the Xli in terms of the Xi.
Henceforth, the circumflex will be dropped; then, equation (26.1-9) is
abbreviated as
(26.1-10)
which is the transformation law for a covariant vector.

The symbol Xk may stand for a variable or for a function; in (26.1-10) it
stands for a function, and one must look at the denominator of the partial
derivative to learn what the independent variables are; if a prime appears, the
independent variables are X '1 , , x in , while if a double prime appears, the
independent variables are X" 1 , ... , x" n , etc. These conventions are standard in
this subject.
A covariant vector field on 9R is defined as a collection of sets {Vi} of n
functions each, one set associated with each chart in 9R, such that the trans-
formation law (26.1-10) holds for any two such sets in the overlap of the cor-
responding charts.
Comments. (1) A vector field is not necessarily the gradient of a scalar, as was
the case in the foregoing example. (2) The two charts may cover exactly the
same portion U of9R, in which case the transformation law (26.1-10) refers to a
"change of independent variables," in the ordinary sense.
The transformation law is transitive: If (26.1-10) is followed by another

transformation from the coordinates x'j to coordinates x"j, then vj and Vj have
the correct relation, because
(26.1-11)
in the overlap of the three charts.

Next, to provide an example of a so-called contravariant vector field, we
consider a fluid flow in 9R. If a fluid particle is at the point P(t) at time t, then,
in a suitable chart {U, <p, N} its coordinates are Xk(t), where x(t) = <p(P(t)),
and the quantities
k = 1, ... , n,
are called its generalized velocity components. [They are the Cartesian velocity
components of the representing point x(t) in the coordinate space \Rn .] If
X'k(t) and V'k(t) are the corresponding coordinates and velocity components
relative to a second chart {U', <p', N' }, then
c:lAlk
uX.
Vi k(t) = -. vJ(t).
ox J
If, to describe the flow of an entire fluid, not just one particle, the velocity
components of the fluid particle which is at point x at some instant t are called
vi(x), then the transformation law is
OX 'k .
V'\X') = - . (x)vJ(x).
ox J
Like (26.1-1) and (26.1-9), this is an identity in the overlap ofthe charts if both
sides are expressed in terms of the Xi or both in terms of the X'i. Again, the
circumflex will be dropped.
A contravariant vector field on m is defined as a collection of sets {Vi} of n

functions each, one set associated with each chart, such that the transforma-
tion law
(26.1-12)
holds for any two such sets in the overlap of the corresponding charts. Contrast
this with (26.1-10) by noting where the prime appears in the derivative.
This transformation law is also transitive.
Note. In Riemannian and pseudo-Riemannian spaces (including Euclidean

spaces), any vector can be represented in either covariant or contravariant
form; the formulas for the raising and lowering of indices by the metric
tensor gjk are given in Section 26.5, below. In some circumstances, however,
for example in Weyl's form of the unified field theory, distances and vector
magnitudes are only relative, and there is no metric tensor, but only a so-
called affine connection (see Section 26.12). In those circumstances, covariant
and contravariant vectors are essentially different entities. Note finally that
the coordinates xl, ... , xn are not the components of a vector, because they
do not transform according to the law (26.1-12), except under homogeneous
linear transformations.
When compatible coordinate charts are added to or deleted from the

manifold, as described in Chapter 23, it is understood that corresponding sets
of vector components {Vj} or {vi} are also added or deleted in accordance with
the transformation laws (26.1-10,12). Properties ofthe vector fields that are
invariant under such additions and deletions are regarded as intrinsic
properties. In this sense, the concept of a covariant or contravariant vector
field is a coordinate-free concept.
EXERCISES
1. Let uj and vj be smooth contravariant vector fields, and call
. k 0 . k 0 .
wJ = vJ uJ (26.1-13)
U -
ox k
- V -
ox k
Show that the quantities {w j }1 transform according to the law (26.1-12), hence con-
stitute another vector field. We write
w = [u, v],
and we call w the Lie bracket of u and v. Clearly [v, u] =- [u, v]. Show that if u j, vj,
and wj are any given smooth vector fields, then
[[u, v], w] + [[v, w], u] + [[w, u], v] = 0 (Jacobi identity).
It follows that if we start two or more Coo vector fields in a Coo manifold, and form all
possible vector fields by repeated construction of linear combinations (with constant
coefficients) and Lie brackets, the result is a Lie algebra (possibly infinite-dimensional).
2. Let ui and vi be smooth contravariant vector fields whose Lie product is =0,
so that
It is to be shown that if a point Q in the manifold is obtained by starting at a point P

and following a solution curve ofthe field ui , i.e., ofthe system
d. .
_Xl = ul (j = 1, ... , n),
dt
for a time t 1 and then following a solution curve of the field vi for a time t 2 , the same
point Q can be reached by first following a solution curve of the field vi for some time t3
and then following one of the field ui for some time t 4 To this end, let a surface xi(s, t)
in the manifold be defined by the initial-value problems
a. .
- Xl(S, 0) = ul(x(s, 0)),
as
xi(O, 0) given (point P),
and
ata.res, t) = .
&(x(s, t)),
Xi(S, O) given (by the preceding problem).

See Figure 26.1.
s
Figure 26.1
Show that then the equation

a. .
- Xl(S, t) = ul(x(s, t))
as
holds also for t "# 0, by showing that the two sides of this equation satisfy identical
initial value problems with respect to t, namely, the same "ordinary" differential
equations (s fixed)
as Xl(s, t) ) = axk
a (a.
at as ~(s, t) ) ,
avi (x(s, t)) (a
that is,
ata. av j
(U 1(X(S, t))) = axk (X(S, t))(uk(x(S, t)))
and the same initial condition, because

a. .
- Xl(S, 0) = U1(X(S, 0)).
as
The surface xj(s, t) consists of points that can be reached from P by zig-zag paths, each
segment of which is either a piece of solution curve of the field u j or one of the field v j
26.2 Tensor Fields

Contravariant tensors Tj\ Tjkl, etc. of arbitrary rank (the rank is the number
of indices) transform according to the law
. ox'j OX'k
T'Jk ... = ox r ox s T rs ... (26.2-1)
The summation convention applies to all the repeated indices r, s, ... on the
right; a multiple sum results. Covariant tensors ~k> ~kl' etc. transform
according to the law
T. = ox r ox s T. (26.2-2)
Jk... ox'j OX,k'" rs ... '
and mixed tensors according to the law illustrated by the example

. ox'j OX'k ox t ox,m
T'Jkt = --r --s - - - Trst. (26.2-3)
ox ox OX'I ox"
All these transformation laws are transitive. The last tensor is said to have
contravariant rank 3 and covariant rank 1.
From the transformation laws, it is seen that if uj , vj are contravariant
vectors and wj ' Zj are covariant vectors, then the quantities defined by
etc. are tensors of the indicated kind. More generally, if

and
are any two tensors, then the products
( Thj,... k,k2... 1,12... )(s"r2." S,S2 ... t,t2"')

are the components of a tensor
whose contravariant and covariant ranks are the sums of the corresponding
ranks of T and S. The process just described is called outer multiplication of
vectors and tensors.
Tensor Fields 181
If a tensor carries the same symbol for a superscript as for a subscript, then
the summation convention applies to that symbol, and the result is a tensor of
lower rank; e.g., given a tensor Rjklm' a tensor Rkl can be defined as Rkl =
R\ Ij. Similarly, from the tensor Si kim there can beformed the scalars Si ki k and
Si\i. This process is called contraction. Outer multiplication followed by
contraction is called inner multiplication; for example, if vi and W j are vectors,
then vjw i is a scalar.
An easily verified converse to the last result is the quotient law which says
that if a collection of sets of quantities {Vi} is given, one set associated with
each coordinate chart that contains some point Po, and if for every covariant
vector {w} defined at Po the quantity viwi is a scalar (an invariant under
coordinate changes), then the sets {Vi} define a contravariant vector (at Po).
The roles of the co- and contravariant vectors can be interchanged. More
generally, if, for example, sets of n 3 quantities {~/} are given and are such that
the quantities
Si = ~klVkWI
transform according to the law (26.1-10) for a covariant vector, for arbitrary
vectors {Vk} and {WI} defined at Po, then the sets {~kl} transform according to
the law for a tensor of the indicated kind-covariant of rank 2 and contra-
variant of rank 1. One says more briefly that ~kl is such a tensor.
Note on the Terminology Found in Some Books on Differential Geometry.

An alternative definition of vector field, given below, is often used. First, if
v) is a contravariant vector field, then the linear operator L = v) ala x), when
. der .
applied to a differentiable scalar f, gives the scalar Lf = vj(ofloxi ). If the

manifold and the functions f and vi are of class C"o, then L maps this class
of function into itself. Furthermore, if f and g are any two such functions,
then
Lfg = jLg + gLf; (26.2-4)
a linear operator having this property is called a derivation, as in Chapter 25.
Conversely, if L is any derivation in the algebra of Coo functions, then a cor-
responding vector field vi can be constructed such that L = vi(olox j),
provided the operator L has a sufficiently local character. In order that L
be so representable, it is clearly necessary that whenever scalars f and g
agree in a neighborhood U in IDl, then Lf and Lg must agree in U. Even more
is true: If L can be so represented, if P is any point, and if the equation
of = 09 (j = 1, ... , n) (26.2-5)
ax} pOx)
1 1
p
holds in any chart (hence in every chart containing P), then Lf and Lg must
agree at P. (It turns out that if L is any linear operator that satisfies this last
condition, it is necessarily a derivation.) To construct the field vi at any point P
in a given chart, we define functions f(j)(x), j = 1, ... , n, by the requirement
that
in a neighborhood of P, and we call

vj(x) = LfU)(x).
This defines functions vj(x) throughout the chart. Then, if f(x) is a COO scalar
field and P any point in the chart, we call
Clearly f and g satisfy (26.2-5); hence
. of I
Lf Ip = Lglp = vJ ax j p' (26.2-6)
Since the left member here is a scalar, it is seen from the quotient law that the
vj transform as the components of a contravariant vector field under changes
of coordinates, and it is seen also that L is a derivation. Hence there is a one-
to-one correspondence between local derivations and vector fields. Because
of this correspondence, some authors define a vector field as a local derivation
in the algebra of coo functions on a manifold. We shall continue to use the older
definition, which has been traditional in most branches of physics.
26.3 Metric in Euclidean Space

If xl, ... , xn are Cartesian coordinates in a Euclidean space and xl, ... , xn
are curvilinear coordinates, in terms of which the Cartesian coordinates are
expressed as Xi = Xi(xl, ... , x n), i = 1, ... , n, then an important tensor gjk is
obtained by writing
ax ax l l ax n ax n
gjk = gjk(XI, ... , xn) = axj axk + ... + axj ox k . (26.3-1)
It is clear that the same functions g j k(' .. ) are obtained, for the given coordinate
system xl, ... , x n , if the X j are replaced by other Cartesian coordinates, say
xj, obtained from the X j by a rotation of the axes (or by a rigid motion,
generally). Now, let Xi and Xi + Llx i (i = 1, ... , n) be the coordinates of two
points P land P 2, and let Xi and Xi + LlX i denote the Cartesian coordinates
of the same two points, i.e.,
Xi = Xi(X l , ... , x n),
(26.3-2)
Xi + LlX = X i(X 1 + Llx 1, ... , Xn
i + Llxn).
The square of the distance from P 1 to P 2 is given by
(d(P 1, P 2))2 = (LlX1)2 + ... + (Llxn)2; (26.3-3)
if the Llx i are regarded as small quantities, then, by expanding (26.3-2) in a
Taylor's series, it is seen that
(26.3-4)
Riemannian and Pseudo-Riemannian Manifolds 183
where 1~x 1 stands for max 1~xj I. This equation is usually paraphrased by
writing
(26.3-5)
ds is called the line element. From (26.3-1) it follows that if a transformation is
made from xl, ... , xn to new coordinates X,l, ... , x rn , then the functions
gjk(' . -) transform like the components of a second-rank covariant tensor, as
the notation indicates. Furthermore, this tensor is symmetric: gjk = gkj'
From (26.3-4) it follows that if the gj k are regarded as the elements of a matrix,
then this matrix is positive definite. A similar metric tensor will be assumed to
exist in any Riemannian space, but will not generally be expressible in the
form (26.3-1) because Cartesian coordinates do not exist if the space is
non-Euclidean.
26.4 Riemannian and Pseudo-Riemannian Manifolds

A Riemannian manifold is an n-dimensional manifold IDl on which a second-

rank tensor gjk is defined, which, in all IDl, is (1) symmetric, i.e., gjk = gkj' and
(2) positive definite, i.e., gj kvjv k > for any non vanishing vector {v j}. The
eigenvalues of the matrix (gjk) are all positive. Note that if gjk is symmetric,
then
oxr oxS
g'jk = ox,j OX'k grs (26.4-1)
is also symmetric. Note also that the positive definiteness is compatible with
this transformation law, because gjkVjvk is a scalar.
The determinant of the matrix (gjk) is denoted by g or g(x\ ... , xn);
it is not a scalar, since its value at a given point in the manifold depends on the
coordinate system.
The transformation law (26.4-1) can be written in matrix notation as
(26.4-2)
where G is the matrix (gjk) and where J is the Jacobian matrix.
In a pseudo-Riemannian manifold, the matrix G is not required to be positive

definite, but only nonsingular and symmetric. Each eigenvalue of G is then
either> or < 0, and the signature s of G is defined as the number of positive
eigenvalues minus the number of negative ones. According to Sylvester's law
ofinertia of quadratic forms, the signature of the matrix G' = J GJT is the same
as that of G-a proof can be found in Bacher 1922; hence, the signature s is
independent of the choice of coordinate system, at each point. The eigenvalues
of G are continuous functions of its elements gjk> hence of the coordinates
Xl, ... , x n , and no eigenvalue is ever zero; hence no eigenvalue can ever switch
its sign as the Xi vary. Therefore, since IDl is connected, the signature s is
constant throughout IDl. In general relativity, IDl has dimension 4 and signa-
ture 2, so that G has three positive eigenvalues and one negative one through-
outIDl.
Since det G i= 0, G has an inverse; the elements of the inverse are denoted
by gjk = gjk(xl, ... , xn); these functions transform according to the law
g'lk
ax X grs
= ____
Alj a Aik
(26.4-3)
aXr aXs
[this equation can be obtained by taking matrix inverses in (26.4-2) and by
noting that the Jacobian matrices of inverse transformations are inverse
matrices]; hence gjk is a contravariant tensor of rank 2.
If [I' is an n-dimensional hyper surface in a Euclidean space EN, where
N> n, then [I' may be regarded as a Riemannian manifold with the metric
that Y' inherits from EN. If Xl, ... , Xn are intrinsic coordinates in a portion U
of [1', if xl, ... , X N are Cartesian coordinates in EN, and if, in U,
i = 1, ... , N,
then, as in Section 26.2, the distance d (in EN) between two points x j , xj + I1xj
in [I' is given by
N
d2 = L [Xi(xl, ... , xn) - Xi(X I + I1xl, ... , xn + I1xn)]2
i= 1
where
N ax i axi
gjk = L -a -a
i=1 X
j
X
k'
One says that [I' is immersed in EN and that gjk is the inherited metric tensor.
EXAMPLE
If Xl = e and X Z = cp are polar coordinates on the unit sphere, where 0 < fJ < n,
-n < cp < n, and if X, Y, Z are Cartesian coordinates in E 3 , then
Z = cos xl,
and
gil = 1, gIZ=gZI=O,
this result is usually written as ds z = dfJ z + sin z fJ dcpz; it can also be obtained from
the line element given by ds z = dr z + rZ dfJ z + r2 sin 2 fJ dcp2 for spherical polar
coordinates in E3 by setting r = 1 and dr = o.
We state, without proof, that an n-dimensional Riemannian manifold can

always be immersed in EN, with N = !n(n + 1), although a higher number of
dimensions may be needed ifthe manifold is to be immersed as a hypersurface
without self-intersections (see discussion of the Klein bottle in Section 23.1).
For the proof, see L. P. Eisenhart, Riemannian Geometry (1966).
Raising and Lowering of Indices 185
A pseudo-Riemannian manifold can be similarly immersed in a flat

(generalized Minkowski) space of suitable dimension and signature. The
converse to this statement, however, has a limitation: A smooth surface in
such a flat space is a pseudo-Riemannian manifold only if it is nowhere
parallel to the light cone.
EXERCISES
1. Consider the Riemannian manifold .Y determined by the single coordinate

system Xl = ~,Xl = I), where the metric tensor is given in matrix form as
and where ~ and I) are allowed to vary over the entire ~, I) plane. Show that .Y can be
immersed in E3 as a hemisphere.
2. Find similarly an immersion in E3 of the manifold with metric tensor
26.5 Raising and Lowering of Indices
In a Riemannian or pseudo-Riemannian manifold, if vj is any contravariant

vector, then the covariant vector Vj = gjkVk obtained by inner multiplication
with the metric tensor is regarded as simply another representation of vj; one
says that Vj has been obtained from vj by lowering the index, and when the
same symbol v is used for a contravariant and a covariant vector, it is under-
stood that they are related in this way. Similarly, if Wj is any covariant vector,
one says that the contravariant vector wj = gjkWk has been obtained by
raising the index. If an index is first raised and then lowered (or conversely),
then the original vector is obtained again, because (gjk) and (gjk) are inverse
matrices, so that gjkgkiwi = Wj' Any index on any tensor can be raised or
lowered in the same way; e.g.,
Sj k1 = gkms jml = glm skm
j , etc.;
note that the horizontal ordering of the indices must be maintained, unless
the tensor is symmetric. Clearly gjk is simply the result of raising both indices
of gjk; the mixed form of the metric tensor is
J. _ J'1
gk - g glk -
_ {I for j = k}_ J.
0 for j #- k - 6b
.
in this special case, it is customary to write the indices without horizontal
separation (g~ rather than gjk or gk j ), which is permissible because of sym-
metry.
26.6 Geodesics in a Riemannian Manifold
The rest of this chapter, except for the last section, deals with geodesics. It is,
however, not quite yet geometry, because we shall be concerned mainly
with analytic tools and relations. The geometry proper starts in the next
chapter with notions like parallel transport along a curve.
Let ~ be a smooth curve in a Riemannian manifold Wl with initial and
terminal points P 1 and P 2' We assume first that ~ lies in a single coordinate
chart and is described in that chart by the functions xi(w), for a ~ w ~ b,
which are assumed to be of class C 2 . Transformations from one chart to
another will be considered later. The quantity
L = fJgkl.xh~;1 dw = f ds (26.6-1)
is called the length of the curve ~. (If Wl is immersed in a Euclidean space EN

of higher dimension, as described in Section 26.4, then L is precisely the
length of ~ as a curve in EN.) The radicand above is positive, because the
metric tensor is positive definite in a Riemannian space-for pseudo-
Riemannian spaces, see the next section. The radicand is understood to be the
function of w given by
1 n dxk(w) dxl(w)
gkl(X (w), ... , x (w ~ ~
def m( 1 n'l .n) (26.6-2)

='I!X, . ,X,X, .. , X ,
where the dot denotes d/dw and where each argument of <1>(... ) is understood
as the corresponding function of w.
We note in passing that if the curve ~ lies in the intersection oftwo charts,
the same length L is obtained from either coordinate system, because the
expression (26.6-2) is a scalar; for any w, the value of is independent of the
coordinate system; hence, the value of L given by (26.6-1) is an invariant. If~
is piecewise smooth, its total length is understood as the sum of the lengths of
its pieces.
If a smooth curve ~ 0 from P 1 to P 2, namely
~o: Xk = I(w) (given), (26.6-3a)
where a ~ w ~ b, can be found such that the integral L is smaller for ~0 than
for any other curve from P lto P 2, then ~ 0 is clearly the shortest such curve.
To compare ~o with neighboring curves, consider curves ~ ofthe form
(26.6-3b)
where 6 is a small parameter and the i'( . ) are arbitrary class C 2 functions such
that i'(a) = i'(b) = 0 (k = 1, ... , n); then
L = L o +.i6
2 l b
a
<1>-1/2 (8<1>
uX
~ k
uX
8<1
_ Z k + _Zk dw + 0(6 2 )
~'k '
(26.6-4)
Geodesics in a Riemannian Manifold 187
where Lo is the length of '{lo. In <1> and its derivatives, the arguments Xk and Xk
are understood to be given the values on '{lo, namely lew) and few). For L
to be a minimum, the above integral must vanish for all choices of the func-
tions Zk(W), and this leads to differential equations for the functions lew) that
describe the curve '{lo. If the integral vanishes for all Zk, i.e., if the l satisfy
these differential equations, then the curve '{lois called a geodesic (or geodesic
curve) in m; '{lo mayor may not be the shortest curve from P1 to P 2 (see
examples, below), but in any case the value of L is stationary on '{lo.
In particular, if m is a Euclidean n-space, and the variables Xl, .. , Xn are
curvilinear coordinates, so that the metric tensor is given by (26.3-1), then '{lo
is a segment of a straight line, expressed in the curvilinear coordinates.
To simplify the following steps, it is convenient to choose the parameter w
on the curve '{loin such a way that <1> is constant on '{lo (but not necessarily on
the neighboring curves '{l; in fact, w cannot be chosen so that <1> is the same
constant on '{lo and on the neighboring curves, because w = a and w = b at
P 1 and P 2, respectively, on these curves; hence, if w could be so chosen, all the
curves would have the same length). This can be done by introducingla new
parameter A = A( w) on '{lo according to the equation
(26.6-5)
for any w1 between a and b (A is arclength along '{lo); using A rather than w as
the variable of integration, the above equation is
therefore, by differentiation with respect to At>
Jgklfl = fo == 1 on '{lo
The factor <1>-1/2 can now be dropped from the integral (26.6-4). That is,
when the parameter A is chosen as arclength on '{lo, the variational problems
c5 ffo dA = 0 and c5 f<1> dA = 0 (26.6-6)
have the same solutions. Although <1> = 1 on '{lo, the partial derivatives of <1> in
(26.6-4) do not vanish, because they involve differentiations in other direc-
tions than merely along '{lo.
Integrating by parts in the second term in (26.6-4) after deleting <1>-1/2 (the
integrated parts vanish because Zk = 0 at P 1 and P 2) and equating the
integral to zero give
fa
b (0<1>
OX k -
d 0<1
dA OXk Z
k
(A)dA = O.
Owing to the arbitrary nature of the Zk( w), therefore,

0<1> d 0<1>
oxk - dA oik = 0 on C(]O (k = 1, ... , n); (26.6-7)
"on C(] 0" means that the functions Xk(A) that appear in (see 26.6-2) are to be
taken as the functions Y\A) that describe the curve C(] o. Equations (26.6-7) are
the Euler variational equations of the problem
(26.6-8)
From the expression (26.6-2) for <1>, it is seen that the Euler equations are
specifically [k is replaced by m in (26.6-7)]
Ogkl k.1 d (.1 .k) 0

oxm X X - dA gml x + gkrn x =
(the two terms in parentheses are of course equal) or
( Ogkl ogml Ogkm) k1 2 1 - O.

(26.6-9)
oxm - oxk - oxl X X - gmlx - ,
these equations (m = 1, ... , n) must be satisfied by the functions Xk(W) =

lew) in order that C(]O be a geodesic.
The equations can be conveniently written in terms of the Christoffel 3-
index symbols of the first and second kind, which are
(26.6-10)
and
{kl} = grm[kl, m] (summed on m), (26.6-11)
respectively. In terms of them, the equations of a geodesic become [after

multiplying (26.6-9) by grm and summing on m]
r
X
+ {r} .k 1 -
kl xx- 0 (r = 1, ... , n) (26.6-12)
(here, as elsewhere, the summation convention is used). The symbol {kl} is a
function of Xl, ... , xn and is understood to be taken at XI(A), ... , Xn(A). If
ce: Xk = Xk(A) is any curve satisfying (26.6-12), then ce is called a geodesic and A
is called a natural parameter (or affine or preferred parameter) on ceo It is
evident that the equations (26.6-12) are invariant under a transformation
A ~ aA + b, where a and b are constants and a 7'= 0; hence the natural param-
eter is only defined up to such a linear transformation. In a Riemannian
manifold, A can be taken as arclength.
Whether L is actually a minimum on ceo (rather than a maximum or merely
a stationary value) and whether the minimum is unique cannot be decided
Geodesics in a Riemannian Manifold 189
SPHERE
MINIMUM-LENGTH
GEODESIC
FROM A TO B
MAXIMUM-LENGTH
GEODESIC
FROM A TO B
CYLINDER
Figure 26.2 Geodesics.
without further consideration (see examples in Figure 26.2, and note also
that there are infinitely many geodesics on a sphere from a given point to its
antipode, all having the same length). It will be seen below, however, that
given any point A of the manifold, there is a neighborhood 91 A of A such that
if B is in 91 A, then there is only one solution !(] 0 of (26.6-12) from A to B lying
in 91 A , and L is a minimum on this curve !(]o.
Notes. The quantities [kl, m] are not the components of a 3rd rank tensor;
neither are the quantities h';,,}' for they don't satisfy the appropriate trans-
formation laws; for example, in a Euclidean space, these quantities are all
identically zero in Cartesian coordinates, but not in curvilinear coordinates.
The quantities Xk are the components of a contravariant vector, but the
quantities j(k are not. Nevertheless, equation (26.6-12) has a certain invariant
character; namely, if it is satisfied, for a given curve peA), in one coordinate
system, then it is satisfied ip any other, because it was derived from the
invariant equation (26.6-8). "Geodesic" and "natural parameter" are
invariant concepts. The n quantities in the left members of (26.6-12), whether
evaluated for a geodesic or not, are, at any point of a curve!(], the components
of a contravariant vector (obtained by the so-called absolute differentiation of
the vector xk-see Section 27.6), although the individual terms of the left
members of (26.6-12) are not, by themselves, the components of a vector. If a
curve!(] runs through several coordinate charts and satisfies (26.6-12) in each
of them, then !(] is also called a geodesic in the manifold.
26.7 Geodesics in a Pseudo-Riemannian Manifold 9Jl
S:
In this case can be negative; hence L, defined as <1>1/2 dw, is meaningless as
S:
a length. Even if L is redefined as 1 11/2 dw, it is still meaningless as a length
in the ordinary sense, because, given any two points P and Q, a (piecewise
smooth) curve can always be found from P to Q for which L = O. Nevertheless,
a curve ~: Xk = Xk(A) that satisfies (26.6-12) is still called a geodesic, and A is
called a natural parameter. The quantity = gikXixk is constant on ~, and
three cases arise:
If > 0, ~ is called a space like geodesic;
if = 0, ~ is called a null geodesic; (26.7-1)
if < 0, ~ is called a timelike geodesic.
Since is quadratic in the xk, this classification is independent ofthe choice of
the natural paramefer. The parameter A can be so chosen that = 1 in the
first case and = -1 in the third; then, A is called distance and proper time,
respectively, along ~. Geodesics playa role in general relativity.
26.8 Geodesics; the Initial-Value Problem;

the Lipschitz Condition
Let a 1 , .. , a" be the coordinates, in a given chart, of an arbitrary point Po in a

Riemannian or pseudo-Riemannian manifold m. The initial-value problem
l
of a geodesic through Po with tangent vector ~1, .. , ~" at Po, and with
natural parameter A, is the following:
dX i .
dA = 1"
diff. eq. . U= 1, ... , n), (26.8-1)
dpJ i k I
- = -hl}PP
dA
Xi(O) = ai
initial condo { pi(O) = ~i U= 1, ... , n). (26.8-2)
It will be proved in the next section that this initial-value problem always has
a unique solution for A in some interval [ - Ao , Ao].
It is convenient to call
l=xk-ak (k = 1, ... , n),
y"+k = pk (k = 1, ... , n),
and to rewrite the differential equations as
dl
dA= fk( y,
1 ... ,y2") (k = 1, ... , 2n), (26.8-3)
Geodesics; the Initial-Value Problem; the Lipschitz Condition 191
where fk denotes the function on the right side of the kth differential equation
of the set (26.8-1), for k = 1, ... , 2n, the Christoffel symbols now being re-
garded as functions of yl, ... ,y2n. The functions fk are defined for all
yn+ 1, ... ,in, and for all i, ... , yn such that the corresponding point
Xl, .. , xn lies in the given chart. It is assumed that the {it} and their first
partial derivatives are continuous throughout the chart, and it is asserted
that the functions fk are Lipschitz continuous in any compact region of the
yl, ... , in space in which they are defined. That is, suppose that (j is a con-
stant such that the fk are defined for all yl, ... , in in the cube W determined
by Iyi I ~ (j (i = 1, ... , 2n). Then there is a constant L = L((j) such that if
{i} and {ji} are two points in W, then
Ifk(yl, ... , y2n) - fk(yl, ... , y2n) I ~ L max Ii - yjl,

(j= 1, ... , 2n)
k = 1, ... , 2n. (26.8-4)
The proof is left as an exercise. It depends on the explicit form of the P given
by (26.8-1) and on the assumption that the Christoffel symbols are functions
of class C 1 of the variables yl, ... , yn. In this new notation, the initial value
problem is
dl
"df= fk( y,
1
... ,yN)
(k = 1, ... , N = 2n). (26.8-5)
leo) = y~ (given)
It will be proved in the next section that this problem has a unique solution
near A = 0, hence we have the following:
Theorem 1. The initial-value problem (26.8-1, 2)for a geodesic starting at

a point Po with initial tangent vector {~k} has a unique solution x\A) in some
interval - Ao ~ A ~ Ao (Ao > 0).
Corollary 1. If~: peA) is a geodesic in m, with A a natural parameter,

satisfying an initial condition (26.8-2), then peA) is unique on its entire
length.
PROOF. Suppose PI (A) and P 2(,1,) are two such curves, and call Al the point at which
they diverge; that is,
Al = sup{A.: PI(A.') = P 2 (A.') for 0::0; A.' ::0; A}.
Let Po = PI (AI) = P 2(,1,1)' and use Theorem 1 (with the origin displaced to ,1,1) to
show that if P 1(A) and PiA) are defined beyond A = ,1,1' then they coincide in some
interval'(Al - AD, ,1,1 + AD), hence that they do not diverge at AI, after all.
If the initial tangent vector {~k} is replaced by a vector of different length in the
same direction, then the geodesic obtained is the same as before, as a point set in 9Jl,
but it has a different choice of the natural parameter Aon it; that is, it follows from the
structure ofthe equations (26.8-1) that if {xk(A), pk(A)} is a solution, for Ain [0, AoJ, and
if a is >0, then {xk(aA), al(aA)} is a solution for A in [0, ADa]. The initial tangent
vector is changed from gk} to {ae}. If a is taken =.10 , it follows that, given any
direction, there is always a solution, valid for all Ain [0, 1], which starts in the given
direction, provided that the components of the initial tangent vector are <ill less than
some positive constant. It can also be proved that the constant depends continuously
on the direction of ~k, and hence has a positive lower bound or minimum. Therefore,
Theorem 2. Let a 1 , . . . , an be the coordinates, in some chart, of the point

Po, as above. Then there is a positive constant K = K(P 0) such that the
geodesic initial value problem (26.8-1, 2) has a solution in the interval
o ~ A ~ 1 for all vectors {~k} such that WI < K (k = 1, ... , n).
This theorem will be used in the next chapter to prove the existence of the
so-called Riemannian coordinates about a point.
26.9 The Integral Equation; Picard Iterations

This section, which is somewhat peripheral to the study of manifolds, is
devoted to a fundamental fact about ordinary differential equations, namely
the existence of a solution of the initial-value problem (26.8-5) when the
Lipschitz condition (26.8-4) is satisfied. That fact and the method used to
prove it are useful in many parts of physics and mathematics.
We first rewrite the initial value problem (26.8-5) as an integral equation of
the Volterra type (that is, one in which the independent variable A appears as
the upper limit of the integral):
leA) = leo) + f f\y(XdX, k = 1, ... ,N, (26.9-1)
where y denotes the vector with components yl, ... , yN. It is convenient to
introduce the norm
Ilxll = max Ixjl;
(j)
then the Lipschitz condition takes the form

I fk(y) - fk(Y)1 ~ LilY - yll, k = 1, ... ,N. (26.9-2)
The solution of the integral equation (26.9-1) is obtained by the Picard
method of iterations, in which leA, q) denote the qth iteration, leA, 0) is
taken as = leO) (independent of A), and the successive improvements are
given, for q = 1, 2, ... , by
leA, q + 1) = leO) + f fk(y(X, qdX. (26.9-3)
To investigate the rate of convergence, we call

L\l(A, q) = leA, q) - leA, q - 1),
Nk(A, q) = fk(Y(A, q - fk(Y(A, q - 1,
Geodesics; the Two-Point Problem 193
for q = 1, 2, .... By subtracting successive equations of the set (26.9-3), we

have
/'ll(A, q + 1) = L"Nk(A', q)dA'. (26.9-4)
This equation holds for q = 0 (as well as for q = 1, 2, ... ), if we define

leA, -1) = 0 and note that fk(O) = O.
The rate of convergence is given by the following:
Lemma. Suppose that

6
Ily(O)11 < :2 (26.9-5)
[i.e., suppose that yeO) lies not merely in the cube W but in the central part of
W] and suppose that IAI ::;; L/log 2, where L is the Lipschitz constant in
(26.9-2). Then,for q = 1,2, ...
6 1
II/'lY(A, q)11 ::;; -2'
q.
(LIAI)q (26.9-6)
and
IIY(A, q)11 < 6. (26.9-7)
PROOF. The case q = 1 follows from (26.9-4) and (26.9-5). The other cases follow
H
by induction on q, because Aq dA = Aq + l/(q + 1) and
IIY(A, q)11 = I ,to l1Y(A, r) II

15 1 15
:s:; - I - (L IAIY < - e10g 2 = 15. (26.9-8)
2 r! 2
This last permits the use of the Lipschitz conditions for each q.
It is clearly the denominator q! in (26.9-6) that gives the method its great
power. Since the sum in (26.9-8) is majorized by the power series for eLIAI , the
sum converges absolutely and uniformly with respect to A in any finite
interval. Hence the sum can be integrated term by term, and we see that the
function
yeA) = lim yeA, q)
satisfies the integral equation (26.9-1); hence it satisfies the initial value
problem (26.8-5). Theorem 1 of the preceding section is thereby proved.
26.10 Geodesics: the Two-Point Problem

In Euclidean geometry, any two points P and Qcan be connected by a unique
straight line segment PQ. The same is true locally in a Riemannian or pseudo-
Riemannian manifold 9.n, with "straight line" replaced by "geodesic."
Consider the problem of finding functions xj(A) (j = 1, ... , n) such that

x j =_{j/}XkX L (O:-s; A:-S; 1,j = 1, ... , n),
(26.10-1)
xj(O) and xj(l) given (j = 1, ... , n).
The following theorem was proved by 1. H. C. Whitehead in 1932:
Theorem. Any point Po in 9J1 has a neighborhood 91 such that if xj(O) and
x j (1) (j = 1, ... , n) are the coordinates of any two points Qo and Q1 in 91,
then there is a unique geodesic segment joining Qo to Ql and lying in 91.
The requirement that the geodesic lie entirely in 91 is often necessary to

obtain uniqueness-see Figure 26.2, in which a geodesic on a cylinder goes the
long way around, instead offollowing the shortest path from its initial point to
its final point.
A somewhat weaker form of the theorem is easily proved; namely, there
exist neighborhoods 91 1 and 91 2, with Po E 91 1 C 91 2 such that if Qo and Ql
are in 91 1, then there is a unique geodesic from Qo to Q1 lying in 91 2 . White-
head's proof that 91 1 and 91 2 can be taken as the same involves a topological
argument.
Two further results, which follow from the proof of Whitehead's theorem,
are now stated, also without proof.
Corollary 1. If the curve Xj(A) is continuous for a :-s; A :-s; b and satisfies the
geodesic equation for a < A < b, then it satisfies this equation also for A = a
and A = b.
Corollary 2. If, to indicate the dependence on the endpoints of the segment,

the solution in Whitehead's theorem is written as Xj(A, Qo, Q1), then for any
A in [0, 1], xj(A, Qo, Q1) depends continuously on the coordinates xj(O) and
x j (1) of the initial and terminal points Qo and Q1
The existence prooffor the two-point problem is similar to the one for the
initial-value problem, in that the system (26.10-1) is converted into an inte-
gral equation, which is then solved by the Picard iterative scheme. The
procedure is somewhat more complicated, because the integral equation is of
the Fredholm type (the upper limit of the integrals is 1, not A), so that one has
neither the factor 1/q! nor the explicit dependence on A that appeared in
(26.9-6). For details, see Whitehead 1932.
26.11 Continuation of Geodesics
It can be proved that, in a Riemannian manifold, if the curve Xj(A) satisfies the
geodesic equation for a < A < b and lies in a compact region of the manifold,
then the limits of Xj(A) exist for A ~ a and A ~ b; hence, in particular, Corol-
lary 1 above applies. In a pseudo-Riemannian manifold, that does not hold, as
Affinely Connected Manifolds 195
the following example shows: Let 9Jl be the surface of a cylinder, let Xl = z
and x 2 = ebe cylindrical coordinates, and let the metric be given by
With z so restricted, 9Jl is a pseudo-Riemannian manifold. The reader can

easily verify that the curve
z = A, (0 < ,1<1)
is a geodesic. As A ~ 0, this geodesic winds infinitely many times around the

cylinder and approaches the circle z = 0.
However, if a curve x j = xj(A) (j = 1, ... , n) satisfies the geodesic equation
for a < A < b and is continuous for a ::; A ::; b, then Corollary 1 of the pre-
ceding section applies, and the quantities xj(b), xj(b) can be used as initial
data for the existence theorem of Section 26.8 to extend the geodesic to some
values of A > b; it can be similarly extended to some values of A < a.
If a geodesic (as opposed to a geodesic segment) is understood as a solution
of the geodesic equations that has been extended as far as possible (perhaps
through many coordinate charts), then the above result may be paraphrased
by saying that in a Riemannian or pseudo-Riemannian manifold a geodesic
cannot have a beginning or an end in the manifold.
Note. This does not imply that A ~ 00 on the geodesic; also it does not
imply that the geodesic may not have a beginning or end in some space in
which the manifold is immersed.
26.12 Affinely Connected Manifolds
Since the transformation laws for tensors are known, and gik is a tensor, it is
easy to find the transformation law for the three-index symbol CD from
equations (26.6-10, 11). If {M' refers to coordinates x'i, ... , x m , and to {M
the unprimed coordinates, then the transformation law is
. [ ax s ax t a2 xr ] ax'i
{),}' = Gt} ax'j ax'k + ax'j ax'k ax r ' (26.12-1)
Since only the tU appear in the geodesic equation (26.6-12), and not the
g jk directly, a more general kind of geometry, called affine geometry, is obtained
if one does not assume the existence of a metric tensor gjk at all but only the
existence of a set of quantities which transform like the CD and which
appear in place of the U} in the geodesic equation. Then there mayor may
not exist a tensor gjk from which these quantities can be obtained via equa-
tions (26.6-10, 11).
An affine connection of a manifold 9Jl is defined in analogy with a tensor as a
collectionofsetsofn 3 functionseach,r)k = rh(xi, ... , xn),onesetassociated
with each chart on 9](, such that in the overlap of two charts the two sets of
functions are related by the transformation law
. [ OX 'ox t
S
02X r ] OX'i
r'jk = r~t ox'j OX'k + ox'j OX'k OX r ' (26.12-2)
which is the same as (26.12-1). Since the r}k are not assumed to be derived
from a metric tensor as the CD were, it is now necessary to verify directly that
this transformation law is transitive (see Section 26.1), in order to be sure that
the definition of a connection is self-consistent. Verification is left as an
exercise.
In an affinely connected manifold 9]( (i.e., a manifold with an affine con-
nection defined on it), a smooth curve ~: Xi = Xi(A.) (i = 1, ... , n) is called a
geodesic, and A. is called a natural parameter on ~, if the equations
(r = 1, ... , n) (26.12-3)
are satisfied on ~; the dot denotes differentiation with respect to A.. Compare
with (26.6-12).
As in Riemannian geometry, the geodesic equations (26.12-3) are invariant
under coordinate changes in the sense that when a curve ~ lies in the overlap
of two charts, it satisfies the equations in one of the coordinate systems if and
only if it satisfies them in the other. The equfltions are also invariant under a
transformation A. ~ aA. + b (a =f. 0) of the natural parameter on a given
geodesic.
The theorems of Section 26.8 on the initial-value problem of geodesics
continue to hold, provided that the components r}k of the connection are
functions of class C~, which, according to (26.12-2), requires that the rranifold
be of class C 3
Whitehead's theorem also continues to hold (in fact, Whitehead stated
and proved it o/iginally for an affinely connected space): each point P has a
neighborhood 9l such that any two points in 9l can be connected by a unique
geodesic in 9l.
The question whether, given an affine connection rh on a manifold, a
metric tensor gjk can be found which is consistent with the metric, i.e., which
is such that rh = {M, is discussed briefly at the end of Section 27.10.
The geometric structure based on the geodesics in an affinely connected
manifold is called the geometry of paths; it is discussed in some detail in the
next chapter. It is clear from (26.12-3) that for that purpose the connection
may be assumed to be symmetric in the subscripts:
(26.12-4)
More generally, a further geometric structure is sometimes introduced, called
torsion, based on the anti symmetric part !(r) k- rL) of the connection -see
Flanders 1963. Since torsion has no effect on the geodesics, it has to be re-
garded as something outside of and superposed on the geometry of the
manifold as determined by its geodesics.
Riemannian and Pseudo-Riemannian Covering Manifolds 197
26.13 Riemannian and Pseudo-Riemannian

Covering Manifolds
Let Wl be a covering manifold of 91, and !/J a projection ofWl onto 91. If f(P) is
any function (say, of class Ck ) in ~n, then the functionl(Q) ~f f(!/J(Q)), defined
in Wl, may be said to be lifted from 91 to Wl, in analogy with the lifting of curves
and surfaces, discussed in Section 24.2.
Now suppose that 91 is a Riemannian manifold. Let epi(p) be the coordin-
ates in a good neighborhood min 91, and let gjk be the components of the
metric tensor in these coordinates. The epi and the gik are all functions in 91,
which can be lifted to Wl as functions ifJi and gjk' Each connected component
U j of!/J-I(m) thus becomes a coordinate chart with a metric tensor in it, and it
is easily established that 9Jl is thereby made into a Riemannian manifold,
called a Riemannian covering manifold of 91. If Wl is the universal covering
manifold of 91, then Wl is called its universal Riemannian covering manifold.
Pseudo-Riemannian covering manifolds are similarly defined.
Consider now the problem of constructing manifolds covered by a given
Riemannian manifold 91. For a general manifold Wl, such an 91 was con-
structed in Section 24.6 by means of a homeomorphism (J ofWl onto itself such
that ifP is any point ofWl, then thesetofpointsP,(J(P),(J((J(P)), ... , (J-I(p), ...
is discrete (has no limit point in Wl); it was constructed by identifying all such
points, for each P, i.e., by regarding the set of these points as a single point of
the manifold 91 being constructed. In order for this process to give a Rieman-
nian manifold 91, it is necessary only to require that (J be an isometric homeo-
morphism, that is, one that preserves the metric. That is, let {U, cp, N} be any
chart in Wl, and let U' be the set of all Q = (J(P), for P in U. Define cp'(Q) as
cp((J- '(Q)). Clearly, {U', cp', N} is a possible chart in Wl (i.e., is compatible with
the charts alread y there), by definition of a homeomorphism. Then, the gj k are
required to be the same functions of the coordinates X,i = ep'i in the second
chart as the gj k are of the coordinates Xi = epi in the first chart. In this case, all
the charts obtained from {U, cp, N} by means of (J and its iterates have
identical metric tensors in them and can be identified as being a single chart
of 91.
These ideas are used in Chapter 28 in the study of the global properties of
Einstein manifolds.
CHAPTER 27
Riemannian, Pseudo-Riemannian,
and Affinely Connected Manifolds
Metric, pseudometric, connection, and topology; geodesic or Riemannian

coordinates; geometry in the sense of Klein; approximate congruence;
congruence of stars; covariant derivative; absolute derivative; parallel
transport; orientability, Lorentz orientability; Laplacian; d' Alembertian;
Riemann tensor; Ricci tensor; Riemann curvature scalar; determination of
the second derivatives of the metric tensor from the Riemann tensor;
conditions for an affine connection to be compatible with a metric; intrinsic
curvature of a manifold; flatness; Stackel, Robertson, and Eisenhart
conditions for separability of the wave equation.
Prerequisites: Chapters 23, 24, 26.
The subject of this chapter is the geometry of a manifold that has a metric, a
pseudo metric, or an affine connection defined on it. The dividing line between
the preceding chapter and this may seem rather fine and arbitrary, since the
last topic in that chapter was geodesics, and the first in this is geodesic
coordinates. However, there is a fundamental difference between the two. The
preceding chapter was mainly analytic, the only geometric notion being that
of the distance between two points, whereas this one is mainly geometric, and
even in the sense of Euclid, except that the concepts are somewhat extended
and the formulation is analytic. The fundamental concepts, such as paral-
lelism, length, curvature, and angle, are really geometric, and must be so
regarded. The use of analytic methods does not detract from the geometric
nature of those concepts any more than did the introduction of numerical
coordinates into Euclidean geometry by Descartes. From that point of view,
one of the main results of the preceding chapter, Whitehead's theorem, serves
the same purpose as Euclid's postulate that through any two distinct points
there can be drawn one and only one straight line, even though, in Whitehead's
theorem, the two points must not be too far apart.
In modern mathematics, where all branches have merged to some extent,
the question what is geometry and what is analysis is somewhat abstruse, but
surely those notions that can be traced directly back to Euclid ought to be
198
Geodesic or Riemannian Coordinates 199
called geometric. In Euclid's time, geometry was regarded as the science that
dealt with the physical space around us. If general relativity is correct, then
Riemannian, pseudo-Riemannian, and affine geometry do also; even if
relativity has to be modified, its primary ideas, which have stood for over
60 years, will surely continue to be basic for physical space-time.
The best understanding of the subject is obtained if one keeps geometric
ideas (like parallel transport of a vector along a curve) in the foreground of
one's thinking, and detailed analytic formulas a little bit in the background.
27.1 Topology and Metric
According to Chapter 23, the topology of a manifold is determined by its

coordinate charts.
The main difference between Riemannian and pseudo-Riemannian (or
affinely connected) manifolds is that in the latter the relation between the
topology arid the metric is less intimate than in the former. In a connected
Riemannian manifold, the distance d(P, Q) between any two points P and Q
is the infimum of the length of a curve between them, the infimum being
taken with respect to all curves from P to Q. The manifold is then a metric
space, which means that a distance function is defined in it and satisfies the
relations:
d(P, Q) :2: for all P, Q;
d(P, Q) = 0 if and only if P = Q;
d(P, Q) ::; d(P, R) + d(R, Q).
A topology in the space is determined by the metric, because the sets
See, P) = {Q: d(P, Q) < e}
(the open balls) are open sets, and the other open sets are unions of open balls.
In a Riemannian manifold, this topology is identical with the topology
previously defined by the coordinate charts.
In a pseudo-Riemannian manifold, with (gjk) indefinite, the distance
d(P, Q), as just defined, is always = 0, and hence is not a metric. Nevertheless,
the manifold has a well-defined topology determined by the coordinate charts,
as explained in Section 23.3, and the same is true of an affinely connected
manifold, where there is not even a concept of distance along a curve.
27.2 Geodesic or Riemannian Coordinates
The key to the investigation of the local geometric properties of a Riemannian

manifold in a neighborhood of a point Po, when the metric tensor gjk is
given in a coordinate system Xl, ... , xn containing Po, is to choose new
coordinates x'i, ... , x ,n , which are as nearly Cartesian as possible near Po. We
200 Riemannian, Pseudo-Riemannian, and Affinely Connected Manifolds
should like the transformed metric tensor gj k to be equal to 0 jk at Po, which

can always be achieved, and to remain close to Ojk nearby. It will be seen that
all the first partial derivatives of the gjk can be made to vanish at Po, and
certain linear combinations of those derivatives can be made to vanish in a
neighborhood of Po. In that sense, the new coordinates will be as nearly
Cartesian as possible. The remaining departure of the gjk from Ojk will then
reflect the non-Euclidean character of the geometry.
The pseudo-Riemannian case is similar, except that there we wish gjk to be
0j k at Po, with the + and - signs determined according to the signature of
the tensor g j k
In an affinely connected manifold, we wish all components f"h of the
connection to vanish at Po and remain as small as possible nearby.
We consider in this section a manifold with an affine connection, with
qk rt
= j , and we specialize the results to the other cases later. Let {U, <p, N}
be a chart containing a point Po, and let Xl, ... , xn denote the coordinates in
that chart. We consider the family of all geodesics through Po, with Il a
natural parameter so chosen on each that Il = 0 at Po. The coordinates of
such a geodesic are given by the solutions xk(ll) of the initial-value problem
d2 x k dx l dxm
dll 2 + rfm dll dll = 0, (27.2-1)
(27.2-2)
dx k
di(O) = y\ (27.2-3)
where yl, ... , yn are real numbers, not all zero, which specify the direction in
which the geodesic starts out from Po. According to Theorem 2 of Section
26.8, these equations have a solution xk(ll; yl, ... , yn), k = 1, ... , n, valid for
o s Il s 1, for all vectors yl, ... , yn in some neighborhood ofthe origin. If this
solution is expanded in powers of Il and the coefficients of the first three terms
are obtained from the above equations, we find
where Iy I denotes max(k) Ill. The subscript 0 indicates that the connection r
is to be evaluated at Po. The above may equally well be regarded as a power
series expansion in the small quantities Ilyl, ... , llyn; hence, without loss of
generality, we may simply set Il = 1, if the l are themselves sufficiently small.
That is,
From this equation it is seen that the Jacobian matrix

ox k (1 ; yl, ... , yn)
oyl
Geodesic or Riemannian Coordinates 201
is equal to 67 at Po, i.e., when all the l are zero, and it follows that the
equation
Xk = xk(l; yl, ... , yn)
can be inverted, in some neighborhood of Po, to give the y's as functions of the
x's. The first few terms of the corresponding expansion are
l = Xk - Xk(O) + !r7m Io(x 1 - xl(O))(xm - xm(o)) + . . .. (27.2-5)
The yl, ... , yn are called geodesic (or Riemannian) coordinates about Po.
The connection components r7m in these equations refer to the original
coordinates. The corresponding components in the geodesic coordinate
system will be denoted by t7m; we can find their values as follows. A geodesic
is given in the geodesic coordinates by l = leA) = Ak, where l, ... , ~n are
constants; hence, d 2l/dA2 = 0, and by comparison with the general equation
(27.2-1) withxk replaced by land rfm by tfm, we see that tfm lmis =0 on the
geodesic, for all A and all l, ... , n. Therefore
t7m = 0 for all k, I, m, at Po, (27.2-6)
and generally
(27.2-7)
throughout the neighborhood of Po in which the geodesic coordinates are
defined.
Note. Except in flat space, geodesic coordinates about Po are not in general
geodesic coordinates about a neighboring point Qo. Stated differently, a
straight line l = Ak + ak in the coordinate space ~n of y1, ... , yn does not
generally represent a geodesic unless the constants a l , ... , an are all zero, i.e.,
unless that straight line passes through the origin of ~n.
The new coordinates about Po are still not unique, and the remaining degree
of non uniqueness is described by the following theorem:
Theorem. An arbitrary transformation of the Xi induces a homogeneous

linear transformation of the associated geodesic coordinates yi about a given
point Po.
PROOF. Let x' 1, ... , x'" be coordinates in any other chart containing Po, and let
y'1, ... , ym be the corresponding geodesic coordinates about Po. Any geodesic peA)
through Po, with P(O) = Po, when expressed in the geodesic coordinates, takes the
forms
peA): yj = A~j and y'j = A~,j,
where the ~j and the ~'j are constants. To lowest order in small quantities (A ~ 0),
the x j agree with the i and the x'j with the y'j. Therefore
hence for any given point peA) on a geodesic through Po, we have
y'j =
ax'j I Y\
-k (27.2-8)
ax Po
which is the required linear transformation.
EXERCISE
Show that the next term in the series (27.2-4) can be written as
(27.2-9)
where
(27.2-10)
the summation being over the cyclic permutations of the subscripts I, m, and r. Still
higher-order terms, with coefficients r~m ... v are discussed in Eisenhart 1926.
In the Riemannian and pseudo-Riemannian cases, where the connection is

obtained from a metric as r~m = {l~}' where {l~} is given by (26.6-10, 11), it is
seen from (27.2-6) that the first partial derivatives of the metric tensor vanish
at the origin of geodesic coordinates, because
Ogk~ = [ki, m] + [mi, k].

ox
27.3 Normal Coordinates in Riemannian and

Pseudo-Riemannian Manifolds
When there is a metric tensor, the geodesic coordinates can be further

specialized by means of a linear transformation of the yi so chosen that, at the
point Po,
+1
+1 (0)
+1 (27.3-1)
-1
(0)
-1
(in the Riemannian case, all the signs are +). That can be done as follows:
First, the yi can be transformed by an orthogonal matrix such that the matrix
(gjk) becomes diagonal at Po. Let the diagonal elements be called d 1 , , dn ,
Geometric Concepts; Principle of Equivalence 203
where the positive ones come first (they are all #0). Then, by a further trans-
formation y'i = Jld;!i (here the summation convention is suspended), the
matrix (gjk) is reduced to the above standard form. Then, the i are unique,
except for an orthogonal transformation (a rotation with possibly an inversion)
in the Riemannian case, and except for a Lorentz transformation in the pseudo-
Riemannian case. The yi are then called normal or normal geodesic or normal
Riemannian coordinates.
If a Riemannian manifold is immersed as an n-dimensional surface S in a
Euclidean space EN of higher dimension, as described in Section 26.4, then the
normal coordinates are Cartesian coordinates in the n-dimensional hyper-
plane tangent to S at Po, the projection from the surface to the hyperplane
being such that geodesics through Pogo into tangent lines, and distances
along the geodesics go into distances along the tangent lines.
(In some of the older literature, any coordinates in which the first partial
derivatives of the gjk vanish at Po are called "geodesic coordinates" about
Po)
27.4 Geometric Concepts; Principle of Equivalence
Geometry deals with points and lines and objects constructed from them. It
deals specifically with spaces of points in which certain sets of points called
"straight lines" or " geodesics" are singled out. In the geometries considered
here, it is also assumed that the space is an n-dimensional continuum, that is,
that it has the topological properties that make it an n-dimensional manifold.
A basic geometric notion is that of congruence. There is generally a group
of transformations in the space, called the congruence group, under which
geometric relations are preserved. Examples are the rigid motion group in
Euclidean space and the Poincare group (nonhomogeneous Lorentz group)
in Minkowski space. Then, two figures (point sets) ,ue said to be congruent if
one can be transformed into the other by a transformation of the group. (In
Felix Klein's "Erlangen Program" of 1872, the point of view was reversed:
When a group of transformations is given in a space, it is taken to determine a
geometry consisting of those relations that are preserved under the transforma-
tions of the group.)
By analogy with Euclidean and Minkowskian geometries, one might
suppose that, in a space with a metric tensor gjb the group ought to consist of
those transformations under which the tensor gjk is mapped into itself. How-
ever, there are generally no such transformations (except the identity), unless
the space is fiat or has constant curvature. Hence, generally, the ordinary
concept of congruence is lost.
An approximate concept of congruence can be defined for very small
figures. We consider first a Riemannian manifold. Let y\ ... , yn be normal
coordinates about a point Po and w1 , . . . , wn normal coordinates about
another point Qo . Let l/J denote the mapping of a neighborhood of Po onto a
neighborhood of Qo obtained by equating the normal coordinates. That is, a
point P with coordinates yl, ... , yn is mapped onto a point Qwith coordinates
wI, ... , wn if i = Wi, i = 1, ... , n. A small figure near Po is said to be ap-
proximately congruent to a small figure near Qo if!/J carries the first figure into
the second. Since a normal coordinate system is unique only modulo ortho-
gonal transformations, the result is an approximate Euclidean geometry.
Geodesics through Po are carried by !/J into geodesics through Qo, and it is
easily seen that the angle between two geode-sics (see formula below) is
preserved under the mapping. (Geodesics not through Po generally do not go
into geodesics at all, unless the space is flat.) If three or more smooth curves
intersect at Po, the star formed by their tangent vectors at Po is mapped into a
similar star at Qo with preservation of all angles. If Po AB is a small triangle
with one vertex at Po, i.e., if Po A, Po B, and AB are short geodesics, and its
image under!/J is a figure Qo CD, then Qo C and QoD are geodesics having the
same length as Po A and Po B; CD is nearly a geodesic, if the triangle is small,
and has nearly the same length as AB. If Po and Qo are the same point and the
land wk are two normal coordinate systems about that point, then the con-
gruences are rotations and reflections that keep Po fixed.
Angles are defined by the formula that gives angles in curvilinear coordinates
in Euclidean space. If two smooth curves, given by Xi = Xi(A) and Xi = xi(/1),
intersect at a point with coordinates Xi(Ao) = xi(/1o), their respective direc-
tions at that point are given by the tangent vectors
and the angle () between those directions is given by
gk~j~k
(27.4-1)
cos () = II~II II~II'
where
(27.4-2)
The absolute value signs in (27.4-2) are unnecessary here but are included for
later use in the pseudo-Riemannian case.
Similar considerations apply to a pseudo-Riemannian manifold. As before,
let yI, ... , yn and wI, ... , wn be normal coordinates about points Po and Qo,
respectively, i.e., geodesic coordinates such that the metric tensor takes the
standard form (27.3-1). These coordinates are unique only modulo Lorentz
transformations; hence the geometry is approximately Minkowskian rather
than Euclidean. Under the mapping !/J given as before by Wi = yi, geodesics
through Po are mapped into geodesics through Qo with preservation of type
(spaceIike, null, or timelike), and the angle between two geodesics, given by
(27.4-1), is preserved. Now, however, cos () can be > 1 or < -1; hence () can
be imaginary. If either ~ or ~ is a null vector, cos () is infinite or undefined,
according as g j k ~j ~k is # 0 or = O.
It should be noted that, even if ~ and ~ are both spacelike or both timeIike,
the value of cos () given by (27.4-1) may lie outside the interval [ -1, + 1].
Geometric Concepts; Principle of Equivalence 205
EXERCISE
Show that a necessary and sufficient condition for Icos eI to be < 1 is that all linear
combinations of ~ and ~ be spacelike, or all timelike. If cos e = 1, there is a null
vector of the form a~ + b~ with a and b not both zero. (In the Riemannian case, that null
vector is the zero vector; hence ~ is a scalar multiple of (.)
For figures in an affinely connected manifold, angles and lengths are not
defined; nevertheless, certain geometric relations are invariant under the
mapping ljJ given by yi = Wi as above. Here, yi and Wi are arbitrary geodesic
coordinates; since there is no notion of normal coordinates, they are defined
only modulo nonsingular linear or affine transformations. For example, if
~,~, and ~ are the tangent vectors at Po to three curves, and if ~ lies in the plane
determined by ~ and ~, i.e., if the three vectors are linearly dep~ndent, then
the same is true after the mapping. In fact, a relation of the form ~ = a~ + b~
is preserved under ljJ, for given a and b. More generally, any set of such vectors
is linearly dependent if and only if the set of their images under ljJ is linearly
dependent.
The further geometric concept of parallel displacement or parallel transport
along a curve, which will be described in Section 27.7, was introduced by
Levi-Civita in 1917. The idea is this: In flat geometries, Euclidean, Minkow-
skian, or affine, where the congruence group, when referred to Cartesian
coordinates, consists of transformations of the form
X --+ Mx + a,
M being an orthogonal, Lorentz, or general nonsingular matrix, the sub-
group of the pure translations
plays a special role. (Note that in the Minkowski case, no relative motion is
involved, only a displacement.) Namely, if one figure can be mapped into a
second by a pure translation, the figures are said to be not only congruent but
also to have the same orientation in space, or to be obtainable from each other
by parallel displacement. The corresponding concept in a curved space is the
parallel transport of a small figure along a curve, with generally different
results, however, depending on what curve is used to connect the initial and
final points.
Our approach to these questions will be in analogy with the "equivalence
principle" of general relativity, according to which certain laws can be
formulated by simply asserting that the corresponding laws of special
relativity hold in an inertial or "freely falling" reference frame. Geodesic
coordinate systems about a point play the role of inertial frames. For example,
the parallel transport of a vector along a curve will be so defined that at the
instant when the vector is passing a point P of the curve, if yl, ... , y" are
geodesic coordinates about P, then the vectors components relative to the yi
system are undergoing no change at that instant, i.e., have zero derivatives
with respect to a parameter on the curve.
27.5 Covariant Differentiation
In this and the following two sections, three closely related concepts are
discussed (covariant differentiation, absolute differentiation, and parallel
transport), anyone of which might be taken as fundamental and the others
derived from it. We consider a general affinely connected manifold.
The partial derivatives of the components of a tensor transform like the
components of a tensor of one-higher rank under linear transformations, but
not under more generai ones. For example, in the Euclidean plane, a vector
field with constant Cartesian components has variable polar-coordinate
components; hence the nonvanishing of the partials with respect to rand eof
the latter is a peculiarity of the r, ecoordinate system. To eliminate effects of
this kind, we define a tensor of higher rank by giving its components at any
point P in a geodesic coordinate system about P as the appropriate partial
derivatives; then, to find its components in other coordinate systems,
we must use the transformation laws.
Let P :xi-= d be any point, and let yl, ... , yn be the corresponding geodesic
coordinates. As in Section 27.2,
Xi = d + yi - !rhyj/ + ... , (27.5-1)
yi = Xi _ ai + !rh(xi _ ai)(xk - ak) + .... (27.5-2)
In these equations, the connection coefficients refer to the Xi coordinate sys-
tem and are to be taken at the point P. If Vi(X 1 , ... , xn) is any covariant vector
field,- we denote by V;(l, ... , yn) its components in the geodesic $ystem. A
second rank covariant tensor Vi; j ' the covariant derivative of Vi' is d~fined at P
by giving its components in the yi system as
vi;ilp = (a: i Vit (i,j=l, ... ,n). (27.5-3)
We wish to transform this equation back to the Xi system. At the point P, we

have Vi = Vi and vi;i = vi;i' because axi/ayi is =b ii at P. However, in the right
member of (27.5-3), Vi has to be known near P as well as at P, for purposes of
differentiation. Near P,
hence
vi;jlp = vi;ilp = (a:i (Vk ~~;) t = (:~~ :;; ~~;t + V a~:;:i jp.
k
At P, ax//ayj = b/ i , etc., and the second derivative is = -rL by (27.5-1).

Therefore,
(27.5-4)
Covariant Differentiation 207
(summed on k); this is the law of covariant differentiation in an arbitrary

coordinate system; viJx\ ... , xn) is a twice covariant tensor field.
This law shows that if Vi; j is defined to be given at P by equation (27.5-3) in a
particular geodesic coordinate system about P, then it is also correctly
given by the same equation in any other geodesic system about P. Namely, if
Xi is also a geodesic system, then the r~j are = 0 at P, and (27.5-4) agrees with
(27.5-3).
An immediate application of covariant differentiation is that if Vi is the
gradient of a scalar f in a Euclidean space, then the Laplacian of f is given in
any curvilinear coordinate system by
because this quantity is a scalar (an invariant) and is obviously equal to the
Laplacian of f in Cartesian coordinates.
EXERCISE
1. (a) Show that if Vi = vi(xl, . .. ,x") is a contravariant vector field, then the
formula
(27.5-5)
defines a second-rank mixed tensor field. (b) Show that if Tij is a second-rank covariant
tensor field, then the formula
(27.5-6)
defines a third-rank covariant field.
The covariant derivative of a general tensor has one additional term (i.e.,
in addition to the partial derivative) as in (27.5-4) for each covariant index
and one additional term as in (27.5-5) for each contravariant index. The
covariant derivative of a scalar field f is simply its gradient: fk = of/oxk.
The operation of covariant differentiation provides the basis for the
invariant formulation of field theories and generally of physical theories
where partial differential equations occur.
EXERCISE
2. Show that the operation satisfies the product rule of differentiation. First, if
Ti j in (27.5-6) is a product ViWj, then
(27.5-7)
More generally, if T::: and S::: are arbitrary tensors, then
(27.5-8)
where the indicated product may be an outer or arbitrary inner product, i.e., contracted
any number of times.
Note. In the older literature, covariant differentiation is often indicated by a

comma in place of the semicolon. Now, the comma often indicates ordinary
partial differentiation, so that Vi;k = Vi,k - rhvl'
EXERCISE
3. Show that in a Riemannian or pseudo-Riemannian manifold the metric tensor
behaves like a constant with respect to covariant differentiation, that is,
gij;k=O, b~;k = O. (27.5-9)
(Recall that in such a manifold r}k = CD.)
A corollary to this exercise is that covariant differentiation commutes with
the raising and lowering of indices: if Vi = gijV j, then Vi;k = gijVj;k'
It should be noted that the result of successive covariant differentiations is
not generally symmetric in the corresponding indices; Vj;k;l is not =Vj;l;b
except in flat space, where they both reduce, in cartesian coordinates, to
02V)OXk OXl. The asymmetry is a manifestation ofthe intrinsic curvature ofthe
space and appears, in Section 27.9 below, in the definition of the Riemann
curvature tensor.
27.6 Absolute Differentiation Along a Curve

Suppose we are given a smooth curve~: P = peA), A any real parameter, in a
Riemannian, pseudo-Riemannian, or affinely connected manifold and we are
given a vector or tensor field, defined at least on ~. We wish to define another
vector or tensor on ~ which represents, at each point of~, the rate of change
of the first vector or tensor with respect to A, in some absolute or invariant
sense.
Specifically suppose that ~ lies in a coordinate chart {U, <p, N}, hence is
given by coordinates Xi = Xi(A) = cpi(P(A)), and suppose that Vi(A) are the
components, in that coordinate system, of a covariant vector defined on ~.
We wish to define another covariant vector W;(A) which represents the rate of
change of V;(A). Clearly we cannot simply write W;{A) = dV;{A)/dA, because
those quantities don't transform properly. Hence we appeal to the principle
of equivalence. Let Po = P(AO) be a point on ~, and let l, ... , yn be the
geodesic coordinates about Po that agree with the Xi to first order. Let V;(A)
be the components of Vi in the geodesic coordinates, so that V;(A) = Vi(A) at
A = AD but not generally for A "# AD. We define
.(A )
W, 0
= dVi(A)
dA
I .
'<='<0
We now transform this equation back to the original coordinates.
d ( OXk) at A == AD
W;(A) = dA Vk(A) oyi
Parrallel Transport 209
At A = ,10' the x's and the y's agree to first order, and the second derivative,
according to (27.5-1), is = -rfj. That is,
.(,1) = dV;{A) _ (A)rk. dxj(A)
W, dA Vk 'J dA
Since the geodesic coordinates no longer appear, this equation is valid at any
point of CfJ, and in any coordinate system. It is customary to denote Wi by
bv/bA, called the absolute derivative of Vi along CfJ; hence, along CfJ
bVi dVi k dx j
bA = dA - rijVk dA (27.6-1)
If Vi is given not only on CfJ but as a vector field defined in a region containing
CfJ, then
bVi dx j
bA = Vi;j dA (27.6-2)
The absolute derivatives along CfJ of other tensors are similarly defined, in
complete analogy with covariant derivatives. For example, if a contravariant
vector Vi is given on CfJ, then <
(27.6-3)
For absolute, as for covariant differentiation, the product rule holds, the
metric tensor behaves like a constant, and the absolute derivative of a scalar f
is its ordinary derivative df IdA (another scalar).
In particular, if Vi is the tangent vector to CfJ, given by
. dx i
v'(A) = dA'
then a comparison of (27.6-3) with the equation (26.12-13) of a geodesic

shows that bdlbA = 0 on CfJ if and only if CfJ is a geodesic and A a natural
parameter. In that sense, a geodesic may be characterized as a curve whose
tangent vector is a constant along its length.
27.7 Parallel Transport
The problem of parallel transport was mentioned in Section 27.4. A vector is

given at a point Po of a curve CfJ, and we wish to define a vector at any other
point of the curve in such a way that it may be regarded as obtained from the
given vector by parallel displacement along CfJ. According to the principle of
equivalence, if P 1 is any other point of the curve and if we set up geodesic
coordinates about P 1, the vector's components relative to that coordinate
system should be unchanging as we move past the point Pl. That is, the
absolute derivative of the vector along CfJ should be zero. Therefore, if ~ 1,, ~n
are the components of the given vector at Po, the transported vector is given
at any point peA) of the curve as the solution Vi(A) of the initial value problem
i5v
i5; = 0 on~, (i = 1, ... , n). (27.7-1)
We note that if the vector V/A) is transformed to another coordinate system

x'i, ... , x m , it satisfies the corresponding initial value problem
i5v'
i5A' = 0 on~, (i = 1, ... , n),
where the ~; are obtained from the ~i by the transformation law for a vector at
Po. That is because i5VJi5A is a vector, so that if all its components vanish in one
coordinate system they also vanish in any other.
Parallel transport of a contravariant vector or a general tensor along a
curve is similarly defined.
Now consider, in particular, a Riemannian or pseudo-Riemannian mani-
fold. If Vi(A) and Wi(A) are any smooth vector-valued functions on ~, then,
since i5g jk/i5A = 0,
d . k
dA (gjk VJW ) = ( .
i5v J k
gjk i5A w
. i5w
+ vl bf .
k)
Therefore, if Vj(A) and Wk(A) are the result of parallel transport of two given
vectors along~, it follows that gjkVjwk is constant on~. That is, in a Rieman-
nian or pseudo-Riemannian manifold, magnitudes of vectors and angles between
vectors are preserved under parallel transport.
27.8 Orient ability
If ~ 1 and ~ 2 are curves, each going from a point P to a point Q, parallel

transport along ~ 1 generally has a different effect from parallel transport
along ~ 2. Stated differently, parallel transport around a closed curve from P
back to P generally transforms each vector at P into another vector at P. In a
Riemannian manifold, that transformation is an orthogonal transformation
at P, because it preserves all magnitudes and angles. If it is a proper orthogonal
transformation (its matrix has determinant = + 1) for all closed curves, the
manifold is said to be orientable. That is the case in particular if the manifold is
simply connected, because if a closed curve is contracted continuously to a
point, the orthogonal transformation in question is changed continuously
into the identity transformation, whose determinant is = + 1.
EXERCISE
A Mobius strip made of flat paper without stretching can be regarded as a 2-di-
mensional Riemannian manifold, which can be covered by two or more coordinate
charts, in each of which gjk is = (jjk throughout. Show that this manifold is not orientable.
The Riemann Tensor, General; Laplacian and d'Alembertian 211
Comment. Orientability is a topological, not a metric concept, and can be

defined for general manifolds, where no metric tensor is assumed. The fore-
going definition has been adopted because it can be generalized to Einsteinian
manifolds, as will now be shown, in a sense that goes beyond topological
orientability.
If 9J1 is a pseudo-Riemannian manifold of dimension n = 4 and signature

s = 2, as in relativity, then the transformation at a point P that results from
parallel transport of vectors around a closed curve beginning and ending at P
is a Lorentz transformation. It was shown in Section 19.4 that the full
Lorentz group has four components; hence, there are four possibilities for the
transformation in question. It may be a proper Lorentz transformation, or it
may involve time reversal or spatial inversion or both. If it is a proper Lorentz
transformation for all closed curves, then 9J1 is called Lorentz-orientable.
In that case it is possible to establish a positive direction of time (distinction
between past and future) and a chirality (distinction between right-handed
and left-handed coordinate systems or helices) throughout 9J1. If the manifold
is simply connected, it is Lorentz-orientable; hence, any Einsteinian manifold
9J1 always has some covering manifold 9J1' (in particular, its universal covering
manifold) that is Lorentz-orientable. Since 9J1 and 9J1' are locally indistinguish-
able, it is never necessary, in general relativity, to consider space-time models
that are not Lorentz-orientable. (That observation was made to the author by
Lipman Bers, in a conversation over a cup of tea at the Courant Institute.)
An affinely connected manifold is orientable ifthe affine transformation of
the geodesic coordinates at a point induced by parallel transport around a
closed curve has a positive determinant for all such curves.
27.9 The Riemann Tensor, General;

Laplacian and d' Alembertian
Let 9J1 be affinely connected. It was noted in Section 27.5 that the result of
differentiating a vector field twice covariantly is not generally symmetric in
the corresponding indices. A direct calculation, using the formulas in that
section, shows that
Vj;k;l - Vj;l;k = RijklVi, (27.9-1)
where Rijkl denotes the quantity
. a. a.. .
R'jkl = axk r j l - + r:..krjl
axl rjk - r:..lrjk' (27.9-2)
Since Vi is arbitrary, the quotient law, applied to (27.9-1), shows that the n4
quantities Rijkl are the components of a tensor of rank 4, which is called the
Riemann tensor or the Riemann curvature tensor.
In Euclidean space, the Riemann tensor vanishes identically in any co-
ordinate system, because it clearly vanishes in Cartesian coordinates, and if
all components of a tensor vanish in one coordinate system, then they all
vanish in any other. It will be seen in Section 27.12 that the vanishing of the
Riemann tensor is also sufficient for a Riemannian manifold to be Euclidean,
for a pseudo-Riemannian one to be Minkowskian, and for an affinely con-
nected one to be flat. In each case, this statement refers just to the metric, but
if the manifold is simply connected, it can be extended to a complete Euclidean,
Minkowskian, or flat space.
In the special case of a Riemannian or pseudo-Riemannian manifold, where
there is a metric and where indices can be raised and lowered, the Riemann
tensor can be expressed in other forms-see next section.
If Rjkl is contracted with respect to its first and fourth indices, we obtain the
Ricci tensor
(27.9-3)
which plays a role in relativity.
The remainder ofthis section is devoted to the Laplacian and d'Alembertian
operators. In n-dimensional Euclidean space, in Cartesian coordinates, the
Laplace operator is given by
V2 = of + o~ + ... + o~,
where
o
Ok = oxk'
In the Minkowski space of special relativity, the d'Alembertian operator is

given by
0 2 = of + o~ + o~ - o~.
Both operators can be written as gjk OJ Ob where (gjk) is the diagonal metric
tensor. In physical theories, these operators are applied to scalar and vector
field. When applied to a scalar field f, their generalization to a Riemannian or
pseudo-Riemannian manifold is simply gj kf j; k, but when they are applied to
a vector field Vi' it is necessary to introduce a tensor denoted by Vi;jb called
the second symmetric extension of Vi (see T. Y. Thomas 1961), defined as the
tensor whose components are given at the origin P of geodesic coordinates
i, ... , y"by
Vj;kllp = o~:di [p' (27.9-4)
where vii, ... , yn) are the components of the vector field relative to the
geodesic coordinates. According to Exercise 4 below, this is not in general
equal to the result of symmetrizing the second covariant derivative Vj; k; I with
respect to the indices k and t. According to the principle of equivalence, we
therefore define the Laplacian or d'Alembertian operator, as applied to the
vector field, as
(27.9-5)
The Riemann Tensor, General; Laplacian and d'Alembertian 213
rather than in terms of the second covariant derivative-but see Exercise 3 in

the next section.
Higher symmetric extensions are similarly defined, for instance,
EXERCISES
1. Show that if qJ is a scalar field, then
qJ;k;1 = qJ;l;k = qJ;kl,
whereas, if Vj is a covariant vector field, then, generally
2. Show that, in an arbitrary coordinate system, the second symmetric extension

is given by
(27.9-6)
where rj'k/ is the coefficient given by (27.2-10).

3. Show that, at the origin of geodesic coordinates,
(27.9-7)
where the superscript indicates that the connection components are those that refer
0
to the geodesic coordinates i, ... , yn.

4. Show that if Vj is a smooth vector field, then
Vj;kl - :Vj;k;l + Vj;l;k)

=i(R m1kj + Rmkl)vm (27.9-8)
5. Show that at the origin of geodesic coordinates
(27.9-9)
The last shows that the Riemann tensor gives all the intrinsic information
about a space in the immediate neighborhood of a point P th~t is given by the
connection components and their first derivatives at P. Namely, by a suitable
choice of coordinates (geodesic coordinates), the r~l can all be made = 0 at P;
hence they give no intrinsic or coordinate-free information, and then their
first derivatives are all determined by the Riemannian tensor as in (27.9-9).
In the next section it will be seen in a similar fashion that when there is a
metric tensor g j k' the Riemann tensor gives all the intrinsic information that is
conveyed by the gjk and their first and second partial derivatives at P.
27.10 The Riemann Tensor in a Riemannian or

Pseudo-Riemannian Manifold
We now suppose that a metric tensor gkl is defined in the manifold. Then, the
connection components r~l' etc., that appear in the definition (27.9-2) of the
Riemann tensor are to be identified with the Christoffel symbols of the second
kind, {M, etc., defined by (26.6-10, 11). A straightforward calculation then
shows that the Riemann tensor R ijk1 (the first index has been lowered) can be
expressed in terms of gkl and its derivatives as follows:
+ gpq(Uk, p] [il, q] - [ik, p] Uf, q]), (27.10-1)
where the square brackets denote the Christoffel symbols of the first kind,
given by (26.6-10).
The number of independent components of this tensor is less than n4
because of the following symmetry relations, which follow from the above
equation:
(27.10-2)
R ijk1 + R ik1j + R i1jk = O. (27.10-3)
We shall now show that, in consequence of these relations, the number of
independent components of the Riemann tensor is
n 2 (n 2 - 1)
(27.10-4)
12
which is = 1,6, and 20, respectively, in 2, 3 and 4 dimensions. First, it follows
from the three relations (27.10-2) that any non vanishing component can be
written (by change of sign, if necessary) as R ijk1 , where i < j and k < f, and
where, if (ij) and (kf) are regarded as two-digit base-n integers, (ij) ::; (kl). The
number of possible values of (ij) or (kl) is then tn(n - 1) = mand the number
of possible pairs (ij), (kl) is
1
zmm+
( 1) n(n - l)(n 2 - n
=--------~------
2) + (27.10-5)
8 .
Next, unless i,j, k, and f are all different, (27.10-3) reduces to a combination of
the preceding identities (27.10-2), and if they are different, it can be assumed,
without loss of generality, that i < j < k < f, for if i',j', k', l'is any permuta-
tion of i,j, k, f, the identity (27.10-3) for i,j, k, f can be obtained from the same
one for i', j', k', l' by use of the preceding identities (27.10-2). Therefore the
number of independent identities oftype (27.10-3) is n(n - 1)(n - 2)(n - 3)/4!,
and subtraction ofthis form (27.10-5) gives (27.10-4). It can be proved that the
R ijk1 satisfy no further algebraic identities independent of (27.10-2,3).
The Riemann Tensor in a Riemannian or Pseudo-Riemannian Manifold 215
As in the preceding section, if Rijkl is contracted once, the Ricci tensor

(27.10-6)
results. In Riemannian and pseudo-Riemannian spaces, this tensor is sym-
metric: Rjk = R kj . A second contraction yields the Riemann curvature scalar
R = gJ'kR jk . ( 27.10-7)
The -covariant derivative of the Riemann tensor satisfies the Bianchi

identity
Rijkl;m + Rij1m;k + Rijmk;l = 0, (27.10-8)
as is easily verified from (27.10-1) by using geodesic coordinates about a
point, so that the covariant derivatives become ordinary derivatives at that
point.
Inner multiplication of the last equation with gikgjl gives
-R;m + Rkm;k + R1m;1 = O.
Here the second and third terms are equal, and the first term is just - oRlox m.
It follows from this equation that the symmetric tensor
Ljk = Rjk - Rgjk , (27.10-9)
which appears in the Einstein field equation, is divergence-free:
L/;k = O. (27.10-10)
We now consider briefly the question, under what conditions an affinely
connected manifold can be made Riemannian by introduction of a metric
compatible with the connection, that is, the question whether, when the
r~lxl, ... , xn) are given, with n;l = rrk, there exist functions gkl(X 1 , , xn),
with gkl = glk> such that r~l = {kl}, i.e., such that
rr
kI
=!2 grm(_ Ogkl + ogml + Ogkm)
oxm ox k ox l '
(27.10-11)
where, as usual, the matrix (grm) is the inverse of the matrix (gkl)' This equation
can be solved to give
(27.10-12)
This equation is just gks;l = 0, by definition of the covariant derivative, and is

known to be satisfied in a Riemannian or pseudo-Riemannian manifold.
Therefore, the necessary and sufficient condition for the existence of a metric is
that this system (27.10-12) of equations have a solution gkl(X\ ... , xn).
EXERCISES
1. Show from the compatibility condition of the system (27.10-12) that the solution
gk I' if it exists, must satisfy the equation
(27.10-13)
which is just the first of the relations (27.10-2). Derive further conditions from this
equation by covariant differentiation.
2. Consider a manifold in which the affine connection is given by
otherwise q k = O.
Show that, in this mal1ifold, (27.10-13) has no symmetric solution at the origin. Show
that the Ricci tensor is not symmetric: R 1z of R z I'
3. Show that in a Riemannian or pseudo-Riemannian manifold, if the Ricci
tensor Rij vanishes, as in empty space-time according to general relativity, the two
natural definitions of the Laplacian or d'Alembertian operator, as applied to a vector
field, agree:
Otherwise, as noted in the preceding section, it is the expression on the left that gives the
correct form at the origin of geodesic coordinates.
4. Show that in a Riemannian or pseudo-Riemannian manifold, at the origin of
geodesic coordinates,
a2
ayl aym gjk = -W~.jklm + Rk1jm )
-W~jklm + Rjmk1 ). (27.10-14)

It follows that the Riemann tensor gives all the information about the geometry in the
immediate neighborhood of a point that is given by the metric tensor and its first and
second partials at the point, because there is a coordinate system (geodesic coordinates)
such that the standard equations gjk = bjk and agjk/aX I = 0 hold at the point; hence
these quantities give no information, and then the second derivatives are given by the
above equation.
27.11 The Riemann Tensor and the Intrinsic

Curvature of a Manifold
Suppose that an n-dimensional manifold m is immersed in (i.e., is regarded as

an n-dimensional surface in) a Euclidean space EN, with its metric inherited
from EN. The geometry ofm then has both intrinsic and extrinsic aspects, the
former being those aspects that are determined by the metric tensor in m
and are independent of the particular way m is immersed in EN. (For example,
a plane sheet of unstretchable material can be rolled up as shown in Figure
27.1, and its inherited metric is then still identical with that of the Euclidean
plane.) A geodesic in m is generally a curve in EN, and its curvature is an
extrinsic property; the center of curvature lies outside m.
On the other hand, the kind of curvature possessed by the surface of a
sphere can be determined from the metric. The curvature of the earth can in
principle be determined by measurements made entirely on its surface. One
constructs a large triangle whose sides are geodesics (shortest paths between
The Riemann Tensor and the Intrinsic Curvature of a Manifold 217
Figure 27.1
vertices); one measures the area A of the triangle and the sum L of its angles.
Then the radius r of the earth is given by the formula
A
2" = L - n. (27.11-1)
r
Clearly, L - n is the angular deviation of a vector from its initial direction

after being carried round the triangle by parallel transport. One may imagine
the vector as an arrow lying on the deck of a ship that steams around the
triangle. When the ship changes bearing at a vertex by turning through a
certain angle to the right, the arrow is then turned through the same angle to
the left with respect to the ship before proceeding.
Note that a surface having this kind of curvature can be immersed in
various ways in E3. A hemisphere can be bent, without stretching, by distort-
ing the equator, as shown in Figure 27.2. According to the differential geo-
metry of surfaces in E 3 , the product of the two principal radii of curvature at a
given point is invariant under such bending without stretching.
A formula that generalizes (27.11-1) will now be derived for a general n-
dimensional affine1y connected manifold W1, not necessarily immersed in a
flat space of higher dimension. Let Xl, ... , xn be coordinates in a chart in W1.
Consider a very small closed curve C(/ in W1 obtained by letting the point Xi
Figure 27.2
trace out a parallelogram in the coordinate space IRn with vertices at

xi(O),
Xi(O) + ai,
Xi = Xi(O) + a i + hi,
Xi(O) + hi,
xi(O),
where the ai and hi are small quantities. A simple calculation shows that if a
vector Vi is carried by parallel transport round Cd} starting with Vi = ~i at the
first vertex, the net deviation ~Vi is given, to lowest order in small quantities,
by
(27.11-2)
This is the required generalization of (27.11-1). From it, the Riemann tensor
can be found in terms of the deviations ~d produced in various vectors ~i by
parallel transport round various small loops.
This formula, as written, is valid only for small loops Cd}, whereas (27.11-1)
was valid for an arbitrarily large triangle. That is because a sphere has con-
stant curvature. If the same procedure were used for a general ellipsoid, it
would have to be restricted to very small triangles.
27.12 Flatness and the Vanishing of the

Riemann Tensor
Given two Riemannian manifolds of the same number of dimensions, the

question arises whether they may not be identical spaces described by different
coordinate systems. This raises global topological questions and also the
following more local one: Given two Riemannian metrics, each determined by
a set of n 2 smooth functions,
(j, k = 1, ... , n)
and
gjk
, ( '1
X , .. , X rn) (j, k = 1, ... , n),
defined in regions Nand N', respectively, of IRn; the question is whether there
is a transformation
(j = 1, ... , n)
of N onto N' or a part of N onto a part of N', such that the gjk transform into
the g}k under the transformation law for second-rank covariant tensors. If so,
the two coordinate systems can be interpreted as determining two charts in a
single manifold. This is a difficult question in general, but can be answered
completely if the functions of one set, say the g}k(- -), are constants, in which
case the space is Euclidean and the Riemann tensor vanishes.
Flatness and the Vanishing of the Riemann Tensor 219
It will be shown that the vanishing ofthe Riemann tensor is sufficient as well
as necessary for a space to be flat. Consequently, it is now assumed that
R ijkZ vanishes everywhere in an affinely connected manifold, and it will be
shown that coordinates yl, ... , yn can be so chosen in any simply connected
region that the corresponding connection coefficients r}k vanish identically.
If there is a metric tensor, the y's can be so chosen that the gjk assume the
standard constant values bj k throughout.
Let Xl, ... , xn be coordinates in a simply connected chart and consider the
following initial-value problem for a covariant vector field vlx\ ... , xn):
h . OVi k (27.12-1)
DE: Vi;j = 0, t at IS, ox j - rijVk = 0,
IC: vlO, ... , 0) given. (27.12-2)

As is known, this problem has a solution for arbitrary initial data (27.12-2)
provided that the compatibility condition
o k 0
OXZ (rijVk) - OXj (ri/Vk) =
k
(27.12-3)
is satisfied. Evaluation of the derivatives in this equation by the product rule

and use of the partial differential equation (27.12-1) once more gives the
equation
(o~Z n o~j nz) + rLrikVm -

j - nzrjkvm = 0.
By renaming the dummy indices, this can be written as

RkjZiVk = 0,
where R k jZi is given by (27.9-2). Therefore, the compatibility condition is
satisfied, because the Riemann tensor vanishes. Let vlx\ ... , xn; p), p =
1, ... ,n, be n solutions of the initial value problem obtained by taking n
linearly independent initial vectors Vi(O, ... , 0; p), p = 1, ... , n; each of these
solutions satisfies Vi;j = 0. Now consider, for each p = 1, ... , n, the further
initial value problem
o y(x,I ... , x)n = vlx I , ... , x n.,p),
-0
Xi
yeO, ... , 0) = 0.
The compatibility contlition of this problem is
a 0
oxj Vi - oxi Vj = 0,

that is
Vi;j - Vj;i =
(the other terms in the covariant derivatives cancel), and hence is satisfied,
because the covariant derivatives are zero. We denote the solution of the new
initial-value problem by yP(Xl, ... , x n), for each p, and we choose new co-
ordinates by setting yP = yP(x 1 , .. ,xn), p = 1, ... ,n, which is possible
because the Jacobian of the y's with respect to the x's at the origin is =J. 0
[because it is the determinant whose columns are the vectors v;(O, ... ,0; p),
which were chosen to be linearly independent]; according to the implicit
function theorem, we can solve for the x's in terms of the y's in some neighbor-
hood No of [Rn, so that the y's become new independent variables in No.
With respect to the new coordinates, the vector fields vj have the components
ox
i
Vj Y 1, ... ,yn.)
o (
,p = Vi (1
X , .. , x ".,p) oi
oyP OX i
- - - - 61!
- OX i oi - J'
Since these vector fields were so constructed as to have vanishing covariant
derivatives, we have
0= Vj;k = 0 - thVi = - t j k6f
-tfk'
Therefore the connection coefficients are all zero in the new coordinates, as
was claimed.
Now suppose the manifold is Riemannian or pseudo-Riemannian, and
hence has a metric tensor. The initial vectors v;{O, ... , 0) in (27.12-2) can be
chosen orthonormal, so that
gjk(X\ ... , x")Vj(x\ ... , xn; P)Vk(X\ ... , xn; q) = 6pq
at x = 0; but gjk and the vi all have vanishing covariant derivatives; hence this
equation holds for all x. The metric tensor (contravariant form) in the yj
coordinate system has components
opq _ jk oyP oyq
9 - 9 ox ox
j k
= gjkvl; p)vl; q);

hence gpq = 6pq throughout, and hence also gpq = 6pq throughout, as
was claimed, because the matrix (gpq) is the inverse of the matrix (gpq).
The chart determined by the new coordinates yl, ... , yn in the region No of
[Rn is in general only a part of the original manifold 9)1, but it can be extended
to a complete flat manifold Wl' by extending No to all of [Rn and requiring the
rh and the gkl to be constant in all[Rn. Then 9)1' is the same as 9)1 if 9)1 was
complete and simply connected.
An example of a nonsimply connected flat manifold 9)1 is the 2-torus with
angles e and <p as the coordinates and with (gjk) equal to the 2 x 2 unit
matrix everywhere. (That is, of course, not the metric that is inherited when the
torus is embedded in the usual way in E3.) In this case, the manifold 9)1' is the
universal covering manifold of 9)1.
Eisenhart's Analysis of the Stackel Systems 221
27.13 Eisenhart's Analysis of the Stackel Systems
The main application of Riemannian (rather, pseudo-Riemannian) geometry

to physics is in general relativity. An example of an application outside
relativity theory, which uses the result of the preceding section, is found in the
paper of Eisenhart 1934 on separable coordinate systems for the wave
equation.
If the function 'I' = 'I'(x\ ... , x n , t) satisfies the wave equation
( \7 2 - -
1
-
C 2 (}t 2
(}2) 'I' + V'I' =
'
(27.13-1)
where \7 2 is the n-dimensional Laplacian and V = V(Xl, ... , xn) is a given

scalar function, then, since V does not depend on t, the first step ofthe separa-
tion process can always be taken: One looks for solutions in the form
(27.13-2)
and then I/J satisfies the reduced wave equation
\721/J + (A + V)I/J = 0, (27.13-2)
where A = W 2 /C 2 . The same equation is also obtained from the Schrodinger
equation. If Xl, ... , X" are general curvilinear coordinates, the Laplacian can
be written, when the metric tensor is known, as gi jl/J; i; j ' and then the red uceq
wave equation takes the form
g
i.[
J
(}21/J
(}Xi (}x j -
k (}I/J]
LJ (}x k + (A + V)I/J
_
- 0. (27.13-4)
Whether this equation for I/J can be solved by the method of separation of
variables depends on the forms of the functions gil -) and V(-).
In the separation procedure (see, for example, Morse and Feshbach 1953),
one starts by looking for special solutions in the form of a product
When this product is substituted into (27.13-4), it turns out that, for certain
choices ofthe coordinate system (hence of the gi) and ofthe function V, one
obtains a system of second-order ordinary differential equations, one for each
of the functions Xi' depending on n arbitrary so-called separatiop constants,
of which A in the above equations is the first. In this way one obtains a large
enough family of special solutions (depending on the separation constants
and constants of integration that appear in the solution of the ordinary
differential equations) to serve as a complete set for the expansion of an
arbitrary function of Xl, ... , X".
In brief outline, the inves~igations of various authors showed that three
conditions on the metric are necessary for separability, i.e., for success of the
procedure just described. First, the coordinates must be orthogonal; that is,
the matrix (gi) must be diagonal. The other two conditions on the gij are the
so-called Stackel and Robertson conditions. It was shown by H. P. Robertson

1927 that those conditions are also sufficient for separability, if the function V
has a suitable form.
At that point, the theory was not yet in a satisfactory state, because a
tensor gi/X1, ... , xn) satisfying the above conditions might not be the metric
tensor in a Euclidean space, or in any other space of interest, for that matter.
Eisenhart 1934 therefore added to the above three conditions the further
condition that the Riemann tensor R ijk1 vanish identically. Any tensor gjk
satisfying all these conditions is then the metric tensor in Euclidean space for
some curvilinear coordinate system. For n = 3, Eisenhart gave a full classifica-
tion of all separable systems in E 3 , and for each of the systems he determined
explicit formulas for the separation procedure.
The separable systems in E3 are the general ellipsoidal coordinates and
special cases obtained therefrom by such limiting processes as making two
semiaxes equal, letting one focus go to infinity, and so on, including prolate
and oblate spheroidal, parabolic, paraboloidal, elliptic cylinder, parabolic
cylinder, circular cylinder, spherical, and Cartesian coordinates. For detailed
descriptions of these systems, see Morse and Feshbach.
CHAPTER 28
The Extension of Einstein Manifolds
Special-relativistic and general-relativistic field equations; stress-energy

tensor; the cosmological constant; Einstein manifold; the Schwarzschild and
Finkelstein charts; Birkhoff's theorem; the meaning of spherical symmetry;
the Kruskal manifold; maximal and geodesically complete manifolds; other
extensions of the Schwarzschild charts; time reversal; the Kerr manifold; the
Cauchy problem of the Einstein field equations.
Prerequisites: Chapters 23,24,26, and 27, and some knowledge of general

relativity.
Of the many mathematical problems connected with general relativity, the

extension problem has been chosen for discussion in this chapter, because it is
concerned with the global geometrical and topological properties of Einstein
manifolds, and those properties seem to me to constitute the most basically
mathematical aspect of the theory. Although no use is made of formulas or
results outside the preceding chapters of this book, the present chapter will
probably be intelligible only to readers with some knowledge of relativity.
In particular, the first two sections do not pretend to be in any sense a deriva-
tion of the principles of relativity, but merely a discussion of them.
28.1 Special Relativity
It is recalled that the electromagnetic theory of J. C. Maxwell not only

perfected and completed classical theoretical physics, but also led to three
major crises: (1) Unlike all previous physical theories, it failed to be invariant
under Galilean or Newtonian transformations from one inertial frame to
another; hence it seemed to imply the existence of an absolute frame of
reference in the universe. (2) When the principles of statistical physics were
applied to it, it led to the so-called ultraviolet catastrophe. (3) When applied to
electrons, regarded as point charges, it led to infinite self-force and self-
energy. The first crisis was resolved by the special theory of relativity, the
second by the quantum hypothesis, both around 1900-1905, and the third
223
224 The Extension of Einstein Manifolds
(not yet perfectly) by the renormalization program of quantum electro-

dynamics in the early 1950's.
To resolve the first difficulty, Lorentz and Fitzgerald introduced the
hypothesis that when an object is in motion relative to the absolute frame of
reference, it becomes contracted by a certain factor in the direction of motion,
and its actions become slowed by a similar factor, owing somehow to the
electromagnetic effects of the extra terms in Maxwell's equations caused by
the motion of its frame of reference. The hypothesis contradicted no known
facts, because very little was then known of the structure of matter, although
it was clear that electromagnetism was involved. The consequence of the
hypothesis was that all purely electromagnetic phenomena were invariant
under the so-called Lorentz transformations, which take the contraction and
time-dilation into account (in fact Maxwell's equations are invariant under
them); hence one could never detect motion through the absolute frame by
electromagnetic effects.
Einstein then introduced the further hypothesis that all physicallaw$ are
invariant under Lorentz transformations (those laws had of course to be
modified, when velocities comparable to e were involved, to achieve this). All
inertial frames thereby acquired equal status; hence, the laws of physics were
simplified in that it was no longer necessary to consider how electromagnetic
effects of motion through the absolute frame might affect mechanics, thermo-
dynamics, atomic structure, and so on. As soon as the laws were known in one
frame, the principle of Lorentz in variance then determined them in every
other. In the next section, similar simplifications of physical laws by general
relativity are pointed out.
28.2 The Einstein Gravitational Field Equations
The classical gravitational field equation is

'lj2cp = 4nGp, (28.2-1)
where cp = cp(x) is the gravitational potential, p = p(x) is the matter density,
and G is the universal constant of gravitation. The equation can be made
Lorentz-invariant, as follows: In special relativity, p is l/e 2 times the 4,4
component of a symmetric second-rank tensor T/1V' the stress-energy tensor.
(As in Section 19.4, xl, x 2 , x 3 are space coordinates, and X4 is ct.) Hence, cp is
taken as the 4,4 component of a tensor cp /1V' and equation (28.2-1) is replaced
by
or
(28.2-2)
The Einstein Gravitational Field Equations 225
where
(h'") ~ ~ r1 1(O~ 11
(h,,) (0;
(2S.2-3)
is the metric tensor of flat Minkowsky space.

[As in Section 19.4, Greek indices go from 1 to 4, and Latin indices from 1
to 3. Physicists often let the Greek indices go from 0 to 3, and they then write
the tensors h Pu and hpu with the first diagonal element = 1 and the others
= -1, so that ds 2 = (dXO)2 - (dXl)2 - (dX 2)2 - (dX 3)2. The conventions
used in this chapter are thought to be a little closer to those ofthe mathematical
literature.]
The law of motion for a point mass in a gravitational field, which says
classically that the acceleration is equal to - V<p, takes the Lorentz-invariant
form
x
Il + hllv 8<ppu
8 pu -
xx -
0,
XV
where the dot stands for differentiation with respect to proper time r along
the trajectory and the x/l( r) are the coordinates of the particle. The equation
resembles that of a geodesic, but of course the second term comes from the
gravitational field and has nothing to do with geometry.
Such a theory may be called the special relativity theory of gravitation. It
was never considered seriously, to any extent, because Einstein discovered the
general theory before observational tests of relativistic effects became feasible.
In the special theory, when there is no gravitational field, so that <Pll v == 0, free
bodies move along straight lines, i.e., geodesics, in the 4-dimensional Minkow-
ski space, whereas, when <P/lV =/=. 0, the trajectories depart from straight lines,
and relative accelerations are observed.
In Einstein's general theory, the trajectories are assumed to be always
geodesics (in the absence of nongravitational forces), and the relative ac-
celerations are attributed to the curvature of space time. That assumption
simplified physics in that it then became unnecessary to consider how a
gravitational field might modify the laws of electromagnetism, quantum
theory, and so on, for those laws were assumed to take their field-free form in
a local inertial or "freely falling" frame of reference, in which the gravitational
field has been "transformed away." A study of the equations of the geodesics
shows that, when space-time is almost flat (almost Minkowskian), a coordinate
system can be found in which the tensor gil v differs only slightly from the hll v
and differs in such a way that the motion is the same as though space time
were exactly flat and there were a gravitational potential given by <P/lV =
(c 2 /2)(g/lv - h/l v). In this limit, the Einstein field equations must reduce to
pu 8 8 _ SnG
9 8x P 8xU gil V - 7 T/l v (2S.2-4)
Einstein therefore chose an equation of the form

8nG
W/l V = -4- 7;, .. (28.2-5)
C
where JiJj, v is a tensor containing the components of the metric tensor and
their first and second derivatives with respect to the x/l and where JiJj,v was
required to have the following properties: (1) It must be symmetric and
divergence-free, i.e.,
gVlTW/lV;U = 0;
this is necessary because the stress-energy tensor 7;, v of the matter has these
properties. (2) When the fields are weak (i.e., space-time is nearly flat), it must
reduce to the left member of (28.2-4) in the coordinates referred to. Einstein
found that the only choice of the JiJj, v that has all these properties is the left
member of the following equation, which is called the Einsteinfield equation:
1 8nG
R/l V - zRg/l v - Ag/l v = -4- 7;,., (28.2-6)
C
here, R/l v is the Ricci tensor defined in Section 27.9, and A is a constant. The
second term in (28.2-6) is necessary to make the left member divergence-free,
according to equations (27.10-9, 10).
In applications to small systems like the solar system or a single galaxy, one
would like to suppose that space-time is asymptotically flat at large distances;
this requires that the so-called cosmological constant A be zero. Einstein
assumed, however, that A is not exactly zero but very small. He believed that
that was necessary in order to obtain closed (i.e., finite) models of the universe
and thus avoid the various forms of Olbers's paradox that appear in an
infinite arid asymptotically uniformly populated universe A. A. Friedman
showed in 1922, however, that closed models could also be obtained for
A = O. Since then, A has usually been taken as zero.
In the remainder of this chapter, A will be taken as zero, and the discussion
will be further restricted to empty regions of space time, where T/l v = O. In
this case, since g/lVg/lV = 4, contraction of (28.2-6) shows that the scalar
curvature R is zero; hence, the gravitational equation reduces to R/l v = O. An
Einstein manifold is therefore defined as a 4-dimensional manifold of signature
2 in which the Ricci tensor R/l v is zero throughout. More general definitions
are clearly possible and are sometimes found, for example in Petro v 1969.
The problem considered in this chapter is how to extend a given Einstein
manifold (usually given by a single chart) so as to obtain a larger Einstein
manifold, in fact, to obtain a maximal extension, in some sense, of the given
manifold.
An extension of that kind played an important role in the development of
relativity. The famous Schwarzschild solution for the field around a spherical
mass, which will be discussed in the next section, appeared to indicate a
singularity of space-time at a certain distance (the so-called "Schwarzschild
radius") from the center. The nature of this "singularity" was much discussed
The Schwarzschild Charts 227
in the early days of relativity. It is now known that it is only a singularity of

the coordinate system used and that the manifold can be extended, by adopting
additional coordinate charts, to a manifold with no singularities, except at the
central point. It is clear that one must understand a" solution" of the Einstein
equations to mean a manifold, not merely a single formula for a line element.
28.3 The Schwarzschild Charts
In 1916, K. Schwarzschild found a stationary spherically symmetric solution

of the Einstein equation Ril v = 0 for empty space, which is asymptotic to the
flat Minkowski metric at large distances. The result is a manifold consisting of
a single chart, which will now be described; it presumably represents space-
time around a central spherical mass.
To say that the solution is stationary means that coordinates can be chosen
in the chart so that the gil v are independent of one coordinate, say X4, which
can be regarded as the time. Clearly, then g44 must be <0, while the metric
form gjk dx j dxk obtained by setting dx 4 = 0 must be positive definite.
To say that the solution is spherically symmetric means that there is a
continuous group G of transformations in the manifold, not involving X4,
which is isomorphic to the rotation group SO(3), and under which the metric
is invariant. At large distances, the transformations of G are assumed to take
the usual form of rotations in Xl, x 2 , x 3 ; otherwise, we can't say much about
them in advance, since we have no theory of nonlinear representations of
groups. For example, we cannot say that there is a point in the manifold
(center of rotation) that is invariant under all transformations of G; there is
in fact no such point in the manifold.
It is assumed that the manifold 9Jl 3 obtained by taking a fixed value of X4 is
asymptotically Euclidean at large distances, in the sense that Wl3 contains a
chart whose coordinate domain N 3 consists of all points in 1R3 such that
(28.3-1)
where a is a constant, and then the metric tensor gjk is asymptotic to (jjk at
large distances in N 3' It is assumed that, at least at large distances, there is a
family of concentric spheres, each invariant under all transformations of the
group G. Consider a geodesic starting at a point P of such a sphere So,
starting inward along a normal to So, and followed inward as far as possible
in the chart. It is invariant under those rotations that keep the point P on So
fixed, because the metric is invariant; hence geodesics are mapped into
geodesics, but the geodesic in question is uniquely determined by its initial
direction at P. If such geodesics from all points P of So are followed in for a
fixed distance, their terminal points determine another invariant surface or
sphere S b and the geodesics can be used to map So onto S 1, so that, in
particular, spherical polar coordinates e, cp on So determine coordinate e, cp
on S b which transform under the transformations ofthe group G just as they
transform on So. Lastly, let r be a variable, increasing smoothly outward,
used to distinguish one of the invariant spheres from another. Then we take
e,
r, <p as coordinates in the chart, and then the metric has spherical symmetry
with respect to those coordinates in the usual sense.
We now have a 4-dimensional chart in 9)(, whose coordinate domain N in
~4 is given by
a< r< 00,
0< e< n,
(28.3-2)
-n < <p < n,
- 00 < X4 < 00,
where a is a constant not yet known.

As usual, this chart fails to cover the north and south poles e = 0, n and the
"international date' line" <p = n on each surface r = const., X4 = const.,
but it can be supplemented by another exactly similar chart oriented in such a
way that the two completely cover all such surfaces. Henceforth, this process
will be automatically assumed, whenever polar angles e and <p are used; a
reference to "a chart" under such circumstances really refers to a manifold
containing these two charts.
It is now shown that when the metric is restricted to a surface r = const.,
X4 = const., it necessarily takes the form
(28.3-3)
where A is a positive constant, which may depend on r. To show this, let

r, X4, and <p all be held constant, and consider the metric form at e = 0, <p = o.
It will be just ds z = A dez, where A = gz z(r, 0, 0, X4); to find ds z for other
values of eand <p, it is only necessary to apply a rotation which transforms e, <p
into 0, O. Since the metric is invariant, ds z must transform into A de 2, but the
line element (28.3-3) is known to be invariant under rotations of the coordinate
system e, <p; hence, (28.3-3) is valid on the whole surface.
Because of the symmetry, i.e., the invariance of the metric under the group
G, it is easily shown that the gil v can depend only on r, and that there can be no
terms in gil v dx ll dx v that connect either of e or <p with either ofr or X4. Hence
the metric is of the form
ds 2 = gll dr2 + A(r)(d(}2 + sin z e d2) + 2g 14 dr dx 4 + g44(dx 4)2.

(28.3-4)
At large distances, g 1 1 ::::: 1, g44 ::::: -1, g 14 ::::: O. Hence, in particular, by

continuity, g44 is #0 down to some "radius" ro (a very small radius, it will
turn out). For r > ro, at least, then, the term in g14 can be removed by a
transformation to a new variable
X 14 = X 4 + f-g14
g44
d r,
The Schwarzschild Charts 229
e, <p, and X4 remaining unchanged. Solving for dx 4 gives dx 4 = dX '4 -

(g14/g44)dr, and if this is put into (28.3-4), the terms in dr dX '4 cancel, at the
expense of two new terms in dr2, which, however, can be incorporated into the
first term g 1 1 dr 2 by changing g 1 l' Then, only diagonal elements appear in
ds 2 ; since the first three terms are positive and the fourth is negative, we can
write
(28.3-5)
where rx, [3, and yare functions of r and where the prime has been omitted from
X4.
A final transformation is now made by replacing r by r' = ey/2 Then, eP is
replaced by eP', and e"(dr)2 is replaced by e"'(dr')2, where rx' and [3' are functions
of r'. Primes are again omitted, and the result is
ds 2 = e" dr2 + r2(d8 2 + sin 2 e d2) - eP(dx 4)2. (28.3-6)
This is the form with which Schwarz schild started.
The determination of the functions rx(r) and [3(r) now proceeds as follows:
One computes the Christoffel 3-index symbols [/lV,o] and Grv} , then the
components of the Riemann tensor RJl v P and, by contraction, those of the
(J
Ricci tensor RJl v. These are all expressions in rx, [3, and their first derivatives, so
that setting RJl v = 0 gives differential equations for rx(r) and [3(r). For the
details of this some somewhat lengthy calculation, the reader is referred to any
book on general relativity (for example, Tolman 1934 or Weber 1961). It is
found that
eP = e-" = 1 - -ro, (28.3-7)

r
where ro is an arbitrary constant, called the Schwarzschild radius, which will

be assumed to be > O. (The corresponding chart with r o < 0 will appear
presently.)
It was pointed out in Section 28.2 that when space-time is nearly flat, the
metric can be interpreted in terms of an equivalent (classical) gravitational
field. For r ~ ro, the above metric is equivalent to a Coulomb field with
potential
GM
r
M being another constant. The relation between the constants M and ro is

2GM
ro = -2-' (28.3-8)
C
The application of the Schwarz schild metric to problems of the solar

system, and the observational tests of relativity based on this metric,form an
important chapter in the theory of relativity. If M in (28.3-8) is taken equal to
the solar mass, 2 .10 33 g, r o is found to be ~3 km. For r ~ ro, space-time is
very nearly flat. For points inside the sun, i.e., for r < R0 ~ 6.5 . 105 km, a
different metric, the Schwarzschild interior metric, must be used, which takes
into account the nonvanishing ofthe stress-energy tensor Til Since ro ~ R 0 ,
V.
space-time is very nearly flat (Mitlkowskian) at all points outside the sun, i.e.,
for r > R0 (also at all points inside, it turns out), and the gravitational field is
very nearly a Coulomb field. In fact, as the reader is doubtless aware, the
observational tests of general relativity require the measurement of exceeding-
ly small effects.
This chapter is concerned with a primarily mathematical problem that
comes out of the theory, namely that of finding the maximal extensions of the
empty space solutions, such as the Schwarz schild exterior solution or the
Kerr solution for the field around a rotating mass, into further regions of
space-time devoid of matter. The astronomical interpretation of these solu-
tions in terms of "black holes in space" or cosmological models, is outside
the scope of the present discussion.
Henceforth, units oflength and time will be used such that ro = c = 1, and
t will be written for X4. Then, the Schwarzschild line element is
ds 2 = (1 - ~) -1 dr2 + r2(d&2 + sin 2 & dcp2) - (1 - ndt2 . (28.3-9)
Three coordinate charts can be constructed, using this metric. For the present,
each of them is to be thought of as a separate Einstein manifold. To define a
coordinate chart, it is only necessary to specify the region N in the coordinate
space ~4 in which r, &, cp, and t vary. Clearly, singularities of the gil v must be
avoided; hence there are three possibilities:
NI : 1< r < 00,
0< & < n,

0< cp < 2n,
-oo<t<oo;
NIl: 0< r < 1
(same as above, for &, cp, t);
Nm: -oo<r<O
(same as above, for &, cp, t).
The resulting charts will be called the (Schwarzschild) Charts I, II, and III,
respectively.
If r is replaced by - r, it is seen that Chart III is simply the solution around
a negative point mass. This is an Einstein manifold, according to the definition
adopted here, and it is even geodesically complete, in the sense of Section 28.6,
below. By itself, it is uninteresting, since negative masses presumably do not
exist. However, it will be seen in Section 28.8, that a metric very similar to that
of Chart III appears in part of the Kerr manifold, which represents the field
around a rotating mass.
The following points concerning the time dependence are noted in passing:
It was shown in 1923 by G. D. Birkhoffthat the metric of Chart I is obtained
The Finkelstein Extensions of the Schwarzschild Charts 231
even if one drops the assumption of stationarity. That is, the metric of Chart I
is the only spherically symmetric one that is asymptotically flat at infinite
distances. This comes about as follows: If the functions r:x and f3 that appear
above are allowed to depend on t as well as on r, then a more complicated
general solution appears. However, this solution can always be transformed
into the stationary one (28.3-9) by a pure transformation of coordinates. It
follows, for example, that the gravitational field around a radially pulsating
star is static. In electromagnetic terminology, there is no monopole radiation
of gravitational waves. There is no dipole radiation, either, because there are
no negative masses; however, quadrupole radiation can occur, and this is
believed to provide one mechanism for loss of energy by pulsars. Finally, it
is noted that the metric of Chart II is not stationary in the sense of the defini-
tion at the beginning of this section; the time like variable is r, not t.
The definition of spherical symmetry adopted above needs one comment.
If P is any point in the manifold IDl3 (x 4 being constant), and if S(P) denotes
the set of all points into which P is carried by the transformations of the group
G, then S(P) is a surface r = const., which is 2-dimensional and is a sphere
(a 2-sphere) by any ordinary criterion. That need not be true in a general
spherically symmetric manifold. Let 9)( be the manifold of the group SO(3),
and let G be the group ofleft translations cp(h): g -+ hg (Vg) in 9)(. Then, the
mapping h -+ cp(h) is an isomorphism of SO(3) onto G. However, if P is any
point of 9)(, then the set S(P) of all points into which P is carried by trans-
formations of G is not 2-dimensional-it is all of 9)(, hence 3-dimensional. In
the Schwarzschild Chart I, the asymptotic flatness at infinite distances
evidently puts an additional restriction on the effect of the group G on the
manifold, so that we have only 2-dimensional invariant sets. The manifold of
SO(3) is compact, so no such restriction can be applied to it.
28.4 The Finkelstein Extensions of the

Schwarzschild Charts
New coordinates were introduced m the Schwarz schild Chart I by D.

Finkelstein (1958), by writing
t' = t + log(r - 1),
(28.4-1)
e,
r, cp unchanged,
whereupon the metric form (28.3-9) becomes
ds 2 = (1 + ~ )dr2 + r2(d()2 + sin 2 e dcp2) - (1 - ~ )dt'2 + ~ dt' dr.

(28.4-2)
Forr> 1,t'goesfrom -ooto +00 when tdoes; hence, the range of variation
ofr, e, cp, t' is the same as that ofr, e, cp, t. The resulting chart covers the same
region of space time as the Schwarz schild Chart I, but in different coordinates.
FINKELSTEIN I
SCHWARZSCHILD I Figure 28.1
However, (28.4-2) contains no singularity at r = l; hence this chart can

by extended, by letting r go down to zero, to give the Finkelstein Chart J,
defined by (28.4-2) together with
N F ,,: 0< r < 00,
0< e < n,
-n < <p < n,
-oo<t'<oo;
it covers a larger portion of space time than the Schwarzschild Chart I, as
indicated schematically in Figure 28.1.
By the further transformation
t" = t' - log(l - r) (28.4-3)
applied in the shaded part of the Finkelstein Chart (the part where 0 < r < 1),
the Schwarzschild Chart II is obtained (in the variables r, e, <p, til); hence the
Finkelstein Chart contains one copy each of the Schwarzschild Charts I and
II. However, it is not simply a matter of eliminating the,barrier at r = 1 in the
metric (28.3-9); to see that, consider the t, rand t', r coordinate spaces (the
e
coordinates and <p being fixed), as in Figure 28.2. When the straight line 'fj'
in the Finkelstein Chart is followed down to r = 1, the corresponding curve 'fj
in the Schwarz schild Chart ('fj and 'fj' represent the same curve in space time)
escapes off to t = + 00, hence cannot be continued further, while the original
curve 'fj' can be continued down to r = O.
r r
c'
o o
_ _ _ _--'-_-L-_ _ _ t'
SCHW ARZSCHILD I FINKELSTEIN I
Figure 28.2
The Kruskal Extension 233
FINKELSTEIN I
FINKELSTEIN II
Figure 28.3
Furthermore, in place of (28.4-1) there is a further transformation

till = t - log(r - 1),
which, when applied to the Schwarz schild metric, gives the same as (28.4-2),
but with + (2/r)dt' dr replaced by -(2/r)dt dr. The resulting chart (with
lll
< r < 00) is called the Finkelstein Chart II. In this case, the curve cor-
responding to ~ of Figure 28.4(2) escapes to t = - 00 instead of t = + 00 ;
hence, this extension of the Schwarz schild Chart I leads to a still different
part of space-time, as indicated schematically in Figure 28.3. Indefinite further
extensions can be obtained by alternate successive use of transformations of
the type t --> t log(r - 1) and t --> t 10g(1 - r) in the intervals (1, 00) and
(0, 1) of r. Throughout the manifolds described so far, the Einstein field equa-
tions are satisfied in the form R/l v = 0, and the signature is 2.
28.5 The Kruskal Extension

In the Schwarz schild Chart I, with metric (28.3-9), M. Kruskal introduced
e,
(1960) new coordinates u, cp, v by the equations
u = Jr-=-i e r/2 cosh~,

e, cp unchanged, (28.5-1)
c- t
v = vr - le r / 2 sinh 2'
whereupon the metric form becomes

ds 2 = feu, v)2(du 2 - dv 2) + r(u, v)2(de 2 + sin 2 e dc(2), (28.5-2)
where f and r are certain functions, of which r is defined as the positive solution
of the equation
(28.5-3)
and then f is given by
4
feu, V)2 = - - e-r(u,v). (28.5-4)
r(u, v)
The Schwarzschild Chart I corresponds to the ranges u > 0, - u < v have their usual ranges); this corresponds to the shaded
quadrant labelled" I" in Figure 28.4. However, the metric form (28.5-2) is
nonsingular and of signature 2, and is a solution of the equation Ril v = 0,
throughout the more extended range
- 00 , as usual, (28.5-5)
-~<v<~,
which corresponds to the entire area between the solid curves in Figure 28.4,
consisting of the regions labeled I, 1', II, and II'. Each of the regions I and I'
is a copy of the Schwarzschild Chart I and each of the regions II and II' is a
copy of the Schwarzchild Chart II. Any two adjacent regions of these four
constitute a copy of one of the Finkelstein Charts, and the origin (u = v = 0)
is a point of space-time not covered by any of the preceding charts.
Kruskal showed that the manifold Sl determined by the metric (28.5-2) and
the region of the u, v plane given by (28.5-5), which will be called the Kruskal
manifold, is a maximal extension, in a sense that will be defined, of the
Schwarzschild Charts.
/I
/i
,n
/i r=O
/i
r= 1
II r=2
l' /
U
/
/
/
/ r=2
II'
/
/ r= 1
/
r=O
Figure 28.4 Diagram of the Kruskal manifold.

Other Extensions of the Schwarzschild Manifolds 235
28.6 Maximal Extensions; Geodesic Completeness

The distinguishing feature of the Kruskal manifold ft is that if a geodesic C(f is
started at an arbitrary point XIl(O) and in a direction given by an arbitrary
initial tangent vector XIl(O), then either it can be continued to infinite values of
the natural parameter A in the manifold or it encounters, for some finite A, a
genuine singularity. A manifold having this property will be called geodesically
complete [some authors use this term a little differently-see Ellis (1972)].
By a genuine singularity is meant a point or set of points where some curva-
ture invariant (that is, some scalar constructed from the functions gil v and
their derivatives of various orders) becomes infinite. Such a singularity
cannot be removed by a change of coordinates, because this scalar, being an
invariant, tends to infinity in any coordinate system, as that point or set of
points is approached.
The proof of the geodesic completeness of the Kruskal manifold is ele-
mentary but tedious. It is shown first that the curves r(u, v) = 0 in the u, v
plane consists of genuinely singular points. It is easy to find curvature scalars
that -+ 00 as r(u, v) -+ O. The Riemann curvature scalar R is not one of these,
because R == 0 in R However, the scalar R(1.py~R(1.PY~, where R(1.py~ is the
Riemann tensor, has the value 6[r(u, V)]-4, hence -+00 as r(u, v) -+ O.
The value ofthefunction r(u, v) in the Kruskal manifold is equal to the value
of the coordinate r in the Schwarz schild manifolds. When r is a polar coordin-
ate, one tends to think of r = 0 as representing a "single point." In the Kruskal
manifold, however, the equation r(u, v) = 0 appears to determine a whole
curve, namely the hyperbola drawn in Figure 28.4. In any case, however, this
"point" or "curve" is not in the manifold ft, and is not in the region of
space-time being described. Whether the singularity is a "point" or is "ex-
tended" is probably a physically meaningless question.
Kruskal then showed that, if a geodesic C(f has arbitrary XIl(O) in ft and
arbitrary XIl(O), then either r(u, v) -+ 0 as A -+ some finite .AO on C(f, or else
A -+ 00 along C(f. Hence ft is geodesic ally complete. In either of the Schwarz-
schild Charts I or II, there are geodesics for which r -+ 1 and t -+ 00 for
finite values ofthe natural parameter A. Since Ais an invariant having physical
significance, while the coordinates r, e, q>, t are purely arbitrary, it seems clear
that it ought to be possible to continue these geodesics in some other charts,
and that is just what the Kruskal extension achieves.
28.7 Other Extensions of the Schwarzschild

Manifolds
Instead of the transformation (28.5-1) introduced by Kruskal, consider the
transformation from r, e, q>, t to ~, e, q>, 17 given by
~ = (r - l)e r ,
e, q> unchanged, (28.7-1)
17 = (r - l)e sinh t.
r
The Schwarzschild metric (28.3-9) then becomes

e-r(~)
ds 2 = r(~)(~2 + 112) [~(d~2 - d11 2) + 211 d~ dl1] + r(~)2[de2 + sin 2 e dcp2],

(28.7-2)
where the function r(~) is given, for any ~ > -1, as the positive solution of the
equation
(28.7-3)
A manifold based on (28.7-2) which will be called Sl', is given by the chart in
which the ranges of the variables are:
N': -1<~<00 }
except ~ = 11 = 0,
-00<11<00
e, cp as usual.
This manifold contains one copy each of the Schwarzschild manifolds I and II;
they are in the regions 0 < ~ < 00 and -1 < ~ < 0, respectively. It will be
seen below that this manifold Sl' is maximal-i.e., it cannot be extended to any
larger one; it might therefore possibly seem superior to the Kruskal manifold,
since the formulas are simpler. However, the singularity at ~ = 11 = 0 is in a
sense a singularity of the coordinate system and not a genuine singularity; it
will be seen that all curvature invariants have finite limits as the point
~ = 11 = 0 is approached.
The manifold Sl' has the further property of time reversal. That is, if tan a
denotes the slope d~/dl1 of a null geodesic in the~, 11 plane (a geodesic on which
ds 2 = 0 while e and cp are constant), and if tan f3 denotes ~/11, it is seen from
(28.7-2) that
tan f3(1 - tan 2 a) = 2 tan a,
which is equivalent to the equation tan f3 = tan 2a; hence, if a point in the
~,11 plane encircles the origin once clockwise, f3 increases by 2n while a
increases by n; i.e., the directions of the null geodesics have been reversed.
Clearly, the type of singularity exhibited by the manifold Sl' at ~ = 11 = 0 is
physically unacceptable; generally, a maximal extension of a given Einstein
manifold is not physically reasonable unless it is geodesically complete.
The nature of the singularity at ~ = 11 = 0 is further explored in the
following exercises:
EXERCISES
1. Consider the one-chart 2-dimensional manifold Wl given by the metric form
(28.7-4)
The Kerr Manifolds 237
and the region

-00 < ( < oo}, except ~ = I] =0
-00<1]<00
of the coordinate space. Show that illl is fiat (i.e., that Ropya == 0), and is of signature zero,
hence is locally Minkowskian. Show that illl exhibits time reversal.
2. Find the geodesics on the manifold 9:11 of the preceding exercise. Show that,
when a geodesic C(J goes to 00 in the ~, I] plane, the natural parameter Aon C(J goes to 00,
whereas, when a point approaches the origin on a geodesic, A tends to a finite value. The
geodesics on which the latter happens are the half-lines ~ = r cos C(, I] = r sin C(, C( a
constant and r > O.
Hints: These exercises become trivial if one transforms to new variables x, t such
that
I] = 2xt, (28.7-5)
which can also be written as ~ + il] = (x + it)2, whereupon ds 2 becomes dx 2 - dt 2. The

transformation ~, I] --> x, t is double-valued, but can be made single-valued in any
simply connected part of the region N;m.
3. The transformation (28.7-5) x, t --> ~, I] makes the punctured x, t plane into a
two-sheeted covering manifold of illl. Discuss the other covering manifolds of illl, and
show that the two-sheeted one is the only one that has a geodesically complete extension
(obtained by putting the origin of the x, t plane back in).
4. By similar use of the transformation from u, v to~, I] given by ~ + il] = {u + iV)2,
show that the Kruskal manifold 5\, with the point u = v = 0 removed, is a two-sheeted
covering of the manifold 5\' defined at the beginning of this section. Show that the most
general covering of 5\' contains n copies of the Schwarzschild I and n copies of Schwarzs-
child II, where n is a positive integer or 00. Show that the two-sheeted covering is the only
one that has a geodesic ally complete extension (obtained by putting the point u = v = 0
back in). Since the projection from 5\ to 5\' preserves the metric, it is evident that all
curvature invariants in 5\' have finite limits as the singularity at ~ = I] = 0 is approached.
28.8 The Kerr Manifolds
An axisymmetric solution of the empty space field equation R" v = 0 was

obtained by Kerr (1963); it can be interpreted-at least, its outer part can-
as the solution around a rotating star. The metric is asymptotically Minkow-
skian at large distance r, as in the Schwarzschild solution. At large but finite
distances, the metric contains a term that represents the Newtonian potential
- GM/r, as in the Schwarzschild solution, where M is the star's mass, and also
a term rx/r 2 times an angular factor, where rx is the star's angular momentum.
[N ote: This is a purely relativistic effect; it arises from the presence of the
momentum density pv of the matter in certain components of T" v on the right
side of (28.2-6); the classical Poisson equation does not give such an effect.
Classically, rotation has an indirect effect on the gravitational field through
the oblateness of a rotating star, but that effect varies as 1/r 3 .] The solution
will now be discussed, without derivation or proof.
One description of the Kerr manifold is in terms of coordinates called

t, x, y, z. (As in similar cases, the reader should be warned that these are
arbitrary symbols; x, y, and z are not necessarily to be interpreted as Cartesian
coordinates in any sense, nor t as time.) In units such that 2M, G, and care
unity (the unit oflength is then the Schwarzschild radius 2MG/c 2 ), the metric
form is
where
z p a
kll dx ll = - dz + 2 2 (x dx + y dy) + 2 2 (x dy - y dx) - dt,
P P +a p +a
(28.8-2)
and where p = p(x, y, z) is a function determined by the equation
(28.8-3)
The axis of rotation is the z axis; a is a constant equal to twice the ratio ofthe
angular momentum to the mass in the units used. For rapidly rotating stars,
a ~ 1, while a can be of the order of 1 or smaller for slowly rotating ones.
Various charts can be based on this solution. There is a singularity on the
circle x 2 + y2 = a 2 , z = 0, where Rafh~RaPy~ _ 00. This corresponds to the
singularity at r = 0 in the Kruskal solution, and in fact the circle contracts to
the origin as a - O.
Equation (28.8-3) has two roots p(x, y, z) of opposite sign at every point
x, y, z except on the disk x 2 + y2 < a 2, z = 0, where p = O. As the point x, y, z
passes through this disk, it is necessary to switch from the one solution to the
other, in order to make the derivatives of the gil v continuous. This is achieved
by introducing charts IDlI' ... , IDl4 as follows. In all of them, ds 2 is given by
(28.8-1), but with the sign of p(x, y, z) specified in each case.
For oni
ar,
{N = {all x, y, z, t} - {x 2 + y2 :::;; a2, z = O},
(28.8-4)
p(x, y, z) > O.
The equation for N means that the closed central disk is excluded from N.
In analogy with function theory, this disk is called a branch cut.
For IDl2 {N = same as for IDlI'

p(x, y, z) < o. (28.8-5)
The next two charts serve to connect IDlI and IDl2 across the disk. For them, N
could be any simply connected region containing the open disk (but not the
circle, of course). It could be some sort ofthin wafer, but for simplicity we take
The Kerr Manifolds 239
it to be all space with the region of the x, y plane exterior to the disk excluded
as a branch cut.
N = {all x, y, z, t} - {x 2 + y2 ~ a2, z = O},
For Wl3 { (28.8-6)
sgn p(x, y, z) = sgn z.
For
on
;I.1l4
{N = same as for Wl3 . (28.8-7)
sgn p(x, y, z) = - sgn z.
The chart Wl3 agrees with Wll for z > 0 and with Wl2 for z < 0, while Wl4 agrees
with Wll for z < 0 and with Wl2 for z > o. A doubly connected manifold Wl o
can now be constructed by the mappings
(7~(~0l
Wl3 Wl 4 , (28.8-8)
(~~o)
Wl2
where, in each case, the mapping is the identity mapping
t ~ t, x~x,
y~y, z~z.
in the half-space (z > 0 or z < 0) referred to in the diagram. In this manifold,

a point can return to its initial position after threading twice through the
circles x 2 + y2 = a2, z = o.
This manifold contains two asymptotically flat regions extending to
infinite distances: Both have Ipl ~ max{1, a}; one has p > 0 and the other
p < O. The first of these exhibits, at large distances, the gravitational field of a
positive mass (=!in the present units), and the second that of a negative mass
(= -!). The two regions are connected through the central ring.
An analysis of the geodesics shows that this manifold Wl o is geodesically
complete, for Ia I > !, but not for Ia I < !. The extension to a larger and geo-
desically complete manifold, for Ia I < ! is carried out in Boyer and Lindquist
1967. The resulting manifold is rather complicated, and the reader is referred
to Boyer and Lindquist for its description; the present discussion will be
restricted to an analogy between the Finkelstein charts and the manifold Wl o
described above.
To investigate the metric at large distances (large values of p), it is convenient
to introduce polar coordinates r, (), cp by writing z = r cos (), etc., with the
usual reminder that this does not necessarily make r a radial distance
coordinate. For large positive p, (28.8-3) shows that p ~ r. Then (28.8-2) gives
k/l dx/l ~ dr - dt, and (28.8-1) then gives, in the approximation p = r,
ds 2 = (1 + ~ )dr2 + r2(d()2 + sin 2 () dc(2) - (1 - ndt2 - ~ dr dt,

which is the metric (28.4-2) of the Finkelstein Chart 1. That is, in the limit
a = 0, (28.8-1) is exactly the Finkelstein Chart I; hence, it is expected that, for
small a, it will be necessary to supplement this chart with another copy of it
and two copies of a chart that reduces to the Finkelstein Chart II for a = 0,
joined together as the Finkelstein Charts are in the Kruskal manifold. These
latter charts are obtained by changing the sign of dt in equation (28. 8-2) for
the quantity kJi dxJi.
There are therefore two versions of the manifold 9Jl o defined by the schema
o
(28.8-8), according to the sign of dt in (28.8-2), say 9Jl and 9Jl;j ; they are both
contained in the geodesically complete manifold constructed by Boyer and
Lindquist.
28.9 The Cauchy Problem

The Cauchy problem of Einstein's equations (for empty space) has one
feature in common with the Cauchy problem of Maxwell's equations. It is
recalled that two of Maxwell's equations, namely
VE = 0, V'H=O (28.9-1)
can be regarded as conditions solely on the initial data; if these conditions are
satisfied at t = 0, it then follows from the other equations,
aE = aH
-
at -cV x H
' at = cV x E (28.9-2)
(these are six equations when written in component form), that the divergence
conditions (28.9-1) are then automatically satisfied for all t ~ 0. Hence, the
full system of eight equations contains enough redundancy so that only the
six equations (28.9-2) in six unknowns need to be regarded as the equations of
evolution.
The Einstein field equations
Rpy = (28.9-3)
have similar properties, but with a somewhat different consequence. First,
since the tensors gpy and Rpy are both symmetric, we may take the gpy with
f3 :::; y as the unknowns, and we only need to consider the equations (28.9-3)
with f3 :::; y. Then there are ten equations in ten unknowns. It will be seen that
these equations impose four conditions or constraints on the initial data and
are such that, if these conditions are satisfied at t = X4 = 0, then they are
automatically satisfied later. Therefore, there are only six independent
equations of evolution for ten unknowns; the solution is hence under-
determined and contains four arbitrary functions. That is just as it ought to be.
The initial-value problem, if properly formulated, ought to determine the
geometry of space-time for X4 > 0, or at least for X4 in some interval (0, T),
but the geometry does not uniquely determine the g Ji v' owing to the possibility
of coordinate changes. Any solution of the initial-value problem can be
The Cauchy Problem 241
modified by subjecting the coordinates in the region X4 > 0 to any arbitrary

transformation that leaves the initial data on the initial surface X4 = 0
unchanged.
Since the differential equations (28.9-3) are of the second order, the initial
data are the values of
g p. V>
og
oxp.4v (11 1
a x, x 2 ,x 3) i.'lor x 4 = 0. (28.9-4)
The choice X4 = 0 of the initial hypersurface Y does not imply that Y is fiat,
because the metric tensor gjk in Y is arbitrary (it is recalled that Latin indices
take values from 1 to 3), but the initial data are assumed to be such that the
3 x 3 matrix (gjk) is positive definite on Y. Also, (gp..) must be nonsingular
and of signature 2 on Y. It follows that [/' is spacelike, and, by use of the
formula for matrix inverse, that g44 < 0 on Y. To simplify the present
discussion, the functions (28.9-4) are assumed to be analytic, so that the power-
series method may be used to solve the Cauchy problem. All partial derivatives
ofthe gp. v that do not involve differentiation with respect to X4 more than once,
are then determined on Y by the functions (28.9-4).
The components Rpy of the Ricci tensor are obtained from equation
(27.10-1) for Ro;PYb by contracting with respect to Il( and <5, that is, by multi-
plying by gO; b and then summing on Il( and <5 from 1 to 4. The result is
(28.9-5)
(28.9-6)
(28.9-7)
where only those terms containing second derivatives with respect to X4 have
been written; the dots stand for terms containing quantities that have been
differentiated at most once with respect to X4. [The summation convention
applies in (28.9-6, 7), but the Latin indices take only the values 1,2, 3.J Since
g44 i= 0, the differential equations R j k = 0 and the initial data determine the
second derivatives of the gjk on [/'. When these second derivatives are sub-
stituted into the other four equations Rp 4 = 0, four conditions on the initial
data are obtained, while the second derivatives of the gP4 with respect to X4
are undetermined.
The equations of evolution can be separated from the auxiliary condition
by a device due to Lichnerowicz (see Adler, Bazin, and Schiffer 1965, where
there is an excellent discussion ofthe Cauchy problem). The mixed form of the
Einstein tensor is
(28.9-8)
where R~ p = ga yRy p is the mixed form of the Ricci tensor and R = gl' vRI' v is
the curvature scalar. It is asserted that the system of differential equations
U,'k= 1,2,3,j~k), (28.9-9)
(/3 = 1, ... ,4) (28.9-10)
is equivalent to the original system (28.9-3).
Note. For any solution gl'v of these equations, R has the value zero, so that
Gpand Rp have the same value (namely zero), but, as expressions containing
the dependent variables g/lv and their derivatives, they are different; hence
(28.9-9, 10) is a set of differential equations different from the set (28.9-3), but
equivalent to it. To prove the assertion, it suffices to note that, according to
(28.9-8), if the Rjk are set equal to zero, then,
G4 k = g 4YR yk = g44R 4k> (28.9-11)
G4 4 = !g44R 44;
hence, since g4 4 1= 0, the system (28.9-10) implies the vanishing of all Rp Y' and
this in turn implies the vanishing of all the G~ p.
If the expressions (28.9-5, 6, 7) for the components of the Ricci tensor are
substituted into (28.9-8), it is seen that the differential equations G4 p = do
not contain any second derivatives with respect to X4. Hence, these equations

play the role of auxiliary conditions; in particular they are conditions that
must be satisfied by the initial data.
An important property of the Einstein tensor (which has already played a
role in Section 28.2) is that it is divergence-free, according to equation
(27.10-10); that is,
(28.9-12)
This equation will be used to show that if the functions g /lV satisfy the differ-
ential equations (28.9-9) for all X4 in some interval [0, TJ and satisfy the
auxiliary condition (28.9-10) for X4 = 0, then they satisfy the auxiliary con-
dition also for all X4 in [0, T]. The formulas for covariant differentiation in
Section 27.5 show that (28.9-12) can be written as
(28.9-13)
where the coefficients A ~ yP6 depend only on the gI' v and their first derivatives.
Since the functions g /l v are such that all R j k are = 0, equations (28.9-8) and
(28.9-11) show that each Gjp (as a function of the gil v and their first derivatives)
can be expressed in terms of the G\ (y = 1, ... , 4). Equation (28.9-13) then
takes the form
a4 G4 p _- BYIij -~-.
-8 8 G4y + CY p G4 y.
X ox}
Concluding Remarks 243
where the coefficients BY/ and C~ depend only on the gil v and their first
derivatives. This is a linear system of differential equations for the G4 p (when
the gil v are given), in which the time derivatives are given explicitly in terms of
the spatial derivatives. According to the Cauchy-Kovalevski theorem, the
solution is unique; hence if all four G4 p vanish on Y (x 3 = 0), then they vanish
also for X4 > 0, as was to be proved.
Lastly, the equations Gjk = 0 (28.9-9) can be regarded as differential
equations for the functions gjk (j, k = 1,2, 3,j ::::; k). However, they also
contain the functions ga4 [in the terms indicated by dots in (28.9-5)]. These
four functions can be specified arbitrarily (but smoothly) in all space-time,
provided only that they match the valuesg a4 and iJg a4 /iJX 4 given on the surface
Y(X4 = 0). Then, since g44 -=f. 0, (28.9-5) shows that the equations G jk = 0
(28.9-9) determine the second time derivatives of all the gjk. By the Cauchy-
Kovalevski theorem, again, these equations have a unique solution, in some
interval 0 ::::; X4 ::::; T, for the given initial data.
28.10 Concluding Remarks
The general problem ofthe extension of Einstein manifolds is far from solved.
Almost nothing is known about the existence or uniqueness of geodesically
complete extensions. The simple manifolds in Section 28.7 have no such
extensions, and the following one has many. Consider a single chart in which
J
the gllv are defined to be Minkowskian
(g,,) ~ ( 1(0)1 1 (O~ (28.9-1)
in a bounded region N of the coordinate space. In one geodesically complete

extension, N' is all of [R4 and the gil v are given by (28.9-1) everywhere. This is
static, fiat, empty space. In another extension, N is again all of [R4, but some
gravitational waves are present that have not yet arrived at the space-time
region represented by N. Almost nothing is known about the Cauchy problem,
except for existence and uniqueness in some time interval 0 ::::; X4 ::::; T.
Even if there is a unique solution for all xl, ... , X4, this is still only one chart,
and it may be an imcomplete manifold.
CHAPTER 29
Bifurcations in Hydrodynamic
Stability Problems
Equation of evolution; Navier-Stokes equation; Poiseuille and Couette

problems; Taylor vortices; wavy vortices; flows and semiflows in a Hilbert
space; normal modes; completeness of the normal mode system; invariant
manifolds; stable and unstable manifolds; fixed points, closed orbits, and
invariant tori; bifurcation; supercritical and subcritical bifurcation;
subharmonic bifurcation; Poincare mappings.
Prerequisites: Chapters 1-8; some knowledge of fluid dynamics.
The work of Lorenz 1963 and of Ruelle and Takens 1971 initiated the
introduction into hydrodynamic stability theory of concepts and principles
from the currently active mathematical field of topological dynamical
systems. It was immediately clear that some of the concepts, for example the
concept of generic properties of systems, have many applications elsewhere
in physics. Also the new concepts and principles put new light on old things,
such as bifurcation phenomena. The idea of strange attractors and their
connection with continuous power spectra gave a new understanding of
chaotic behavior generally.
In this chapter and the next two, the new ideas are presented in the setting
of the study of the onset of turbulence.
29.1 The Classical Problems of Hydrodynamic

Stability
We consider the flow of an ideal incompressible viscous fluid under con-

ditions of mild instability, so that the flow is smoother and simpler than the
chaotic flow that constitutes fully developed turbulence, while the elements
of instability and unpredictability are present in a simple form and can be
studied analytically. In an arrangement in which the Reynolds number is
slowly increased, the regime of interest to us is the early onset of turbulence.
244
Examples of Bifurcations in Hydrodynamics 245
The flow represents a balance between a steady influx of energy from an

external source and the dissipation of energy by viscous friction. In the
classical problems of the Poiseuille type, energy is supplied by an externally
imposed pressure gradient which, for example, causes flow through a long
circular pipe or in the channel between parallel plane walls; in the Couette
problems the energy is supplied by laterally moving walls, such as a sliding
plane wall or a rotating cylinder; in the Benard problems it is supplied by an
external heat source, which causes thermal convection. Other examples are
flow past a circular cylinder (von Karman problem) or in the boundary layer
over a flat plate parallel to the primary flow (Blesi us problem).
A given problem can usually be characterized by a length 10 and a velocity
Do, where 10 may be a diameter or other dimension, and Do may be a mean
velocity of the flow or the velocity of motion of one of the walls. The dimen-
sionless quantity
R = lovo (29.1-1)
v '
where v is the kinematic viscosity coefficient ( = viscosity coefficient divided
by the fluid density), is called the Reynolds number of the problem. It is
usually assumed that R is varied by varying Vo, keeping v and the dimensions
constant. However, the fluid-dynamical equations are invariant under
changes of length and speed scales (keeping all length and speed ratios
constant) and the substitution of one fluid for another, so long as R is kept
fixed.
The Taylor problem of flow between rotating concentric cylinders, some-
times called the circular Couette problem, and the Benard problem of con-
vection in a horizontal layer of fluid heated from below are the examples most
studied so far of the early onset of turbulence.
29.2 Examples of Bifurcations in Hydrodynamics

We consider first the special case of the Taylor problem of flow between
concentric rotating cylinders in which the outer cylinder is at rest. When the
inner cylinder rotates slowly, the motion is laminar; the fluid velocity has only
a () component (where r, e, and z are cylindrical coordinates), and that com-
ponent depends only on r. When a certain critical velocity of rotation is
exceeded, that flow, the so-called Couette flow, becomes unstable and has
superposed on it a perturbation consisting of uniformly spaced ring vortices,
as shown in Figure 29.1. If A denotes some measure of the strength of the
vortices, say the maximum of the r or z component of velocity of the per-
turbation, then the dependence of A on the angular velocity 0 of the inner
cylinder is as shown schematically in Figure 29.2(a). The two possible signs
of A, for given 0, correspond to the two possible directions of vortex rotation,
which are equally likely and are determined by initial conditions (neighbor-
ing vortices rotate opposITely). If 0 exceeds the critical value 0 1 only slightly,
IAI is approximately proportional to JO - 0 1 ,
246 Bifurcations in Hydrodynamic Stability Problems
Figure 29.1 Taylor vortices.
The appearance of the new flow at 0 = 0 1 is called a bifurcation. In the

case just described, it is a bifurcation to another steady flow; at a fixed point
in space the fluid velocity is independent of time; the ring vortices are stable
and persist as long as the inner cylinder is kept rotating.
When a second critical angular velocity O 2 is exceeded, the ring vortices
are unstable, and a second bifurcation takes place to wavy vortices, indicated
schematically in Figure 29.3. The waves rotate about the common axis of the
cylinders at about the mean angular velocity of the fluid in the Couette flow;
hence, at a fixed point in space, the fluid velocity is now periodic in time. If
A1 denotes the amplitude of the waviness, the dependence of A1 on 0 is as
indicated in Figure 29 .2(b).
----------L-------~A --------~--------_+A
(a) (b)
Figure 29.2 Bifurcations in the Taylor problem.

The Navier-Stokes Equations 247
,...::::::=====:::.
I I
<q----F>
<----J>
I
I
~ I
~-:---==~:--~
0::: ______ )
"-
Figure 29.3 Wavy vortices in the Taylor problem.
A perhaps more familiar example of bifurcation is found in the flow past

a circular cylinder. At low speed, the flow is laminar; when a critical velocity
is exceeded, the laminar flow is unstable and is replaced by the so-called von
Karman vortex street, in which vortices form alternately on the two sides of
the cylinder, with their axes parallel to the cylinder, and then move down-
stream with a speed around half the ambient fluid speed. In this case the flow
is periodic after the first bifurcation.
29.3 The Navier-Stokes Equations
In terms of the fluid's velocity field u(x, t) and its pressure field p(x, t) the
Navier-Stokes equations are
AU
at + (u V)u + Vp - vV 2 u = 0, (29.3-1)
V u = 0, (29.3-2)
in a region flA of physical space, together with the boundary condition that
u is given on oflA, (29.3-3)
and a suitable initial condition. The density has been taken = 1 by suitable
choice of units.
Let u(x and p(x) represent a steady solution of these equations, for ex-
ample the Couette flow in the Taylor problem. For study of the effect of
perturbations (either finite or infinitesimal) on that solution, it is convenient
to write the total fields as
u(x) + u(x, t), fJ(x) + p(x, t),
where now u and p represent the departure from the steady solution. The
equations are then
au + (it. V)u + (u . V)it + (u . V)u + Vp - vV 2 u = 0, (29.3-4)

at
V' u = 0, (29.3-5)
u = on afYl. (29.3-6)
Note that the equations have not been linearized. The known function it(x)
now appears in the coefficients of the terms linear in u.
29.4 Hilbert Space Formulation
Corresponding to each hydrodynamic stability problem there is a problem

of evolution in a Banach space b (in practice a Hilbert space), consisting of an
equation of evolution of the form
du u = u(t) b}

E
dt = Lu + B(u, u) { (29.4-1)
t ::?:
with an initial condition
u(o) = Uo gIVen. (29.4-2)
For each t, u(t) is a point in b and represents an instantaneous state of the
system-a velocity field u(x) and a pressure field p(x) in the physical space.
L is a linear operator. The nonlinear terms of the Navier-Stokes equation,
which (29.4-1) represents, are quadratic and are written as B(u,u), where
B(, .) is a bilinear function defined on a suitable domain in b x b. We
shall sometimes write Q(u) for B(u, u).
It is sometimes convenient to consider a more general equation
du
M dt = Lu + B(u, u), (29.4-3)
where M is another linear operator. If M has an inverse, then this equation

can be reduced to the form (29.4-1), but often the evolution takes place not
in all of b but in a subspace bo. Then, M has an inverse in the subspace but
not necessarily in b, and M- 1 may be rather complicated for calculation;
hence, (29.4-3) is often convenient.
A particular Hilbert space suitable for the Taylor problem is described in
the next chapter.
29.5 The Initial-Value Problem; the Semiflow in .s
The do Il1 inant terms of the Navier-Stokes equation (29.3-1) are the first
and last terms on the left, which give the equation the character of a diffusion
equation. With suitable boundary conditions, the corresponding diffusion
The Normal Modes 249
equation has a unique solution, for initial u in a set dense in f), and the solu-
tion depends continuously on the initial u. Because of this continuous de-
pendence, generalized solutions can be defined for arbitrary initial u and
they too depend continuously on the initial u. See Richtmyer and Morton
1967. The solution cannot in general be continued backward for negative t,
although certain solutions can; in particular, normal mode solutions can.
The general behavior of the full nonlinear Navier-Stokes equation is
similar (see Ladyzhenskaya 1969, 1975 and Marsden and McCracken 1976,
Section 9), but the proofs are more difficult and the theory is less complete.
We shall assume that (29.4-1) or (29.4-3) has a unique solution u(t) in
f), for t ;;:: 0, for arbitrary u(O) in f). For given initial u, we call the solution
qJ(u, t), so that
u(t) = qJ(u(O), t). (29.5-1)
For fixed u, qJ(u, t) (t ;;:: 0) is called a motion in f); for fixed t ;;:: 0, the cor-
respondence u ~ qJ(u, t) is a mapping in f), which is assumed continuous;
for t = 0, it is the identity mapping, because qJ(u,O) = u. The function
qJ( ., .) is called a semiflow in f).
Although the motions cannot generally be continued backward in time,
they are unique, insofar as they can be. Stated differently, two distinct
motions never coalesce, at a finite time, so as to be identical thereafter.
29.6 The Normal Modes
We look for solutions of the linearization of (29.4-1), namely of

du
dt = Lu, (29.6-1)
in the form
(29.6-2)
hence we look for eigenfunctions tjJ and eigenvalues A of L, given by
LtjJ = AtjJ, (29.6-3)
where, of course, tjJ =1= O. [We could also start from (29.4-3) and seek solutions
of LtjJ = AMtjJ.]
In the hydro dynamical problems, L is not self-adjoint; hence the usual
spectral theory does not apply. Nevertheless, in most cases, L has a pure
point spectrum, and there are denumerably many eigenvalues, so we write
(j = 1,2, ... ). (29.6-4)
The completeness of the set {tjJ j} of eigenfunctions for the expansion of an
arbitrary u in f) has been discussed for certain hydro dynamical problems by
DiPrima and Habetler 1969, using a theorem of Naimark on operators in a
Hilbert space, and more generally by Sattinger 1970, using a theorem of
Carleman. Completeness is established in the sense that the finite linear

combinations of the t/Jj are dense in 5, provided that we include generalized
eigenfunctions, if there are any, i.e., vectors t/J such that
(L - AI)k+ It/J = 0, (29.6-5)
(If k = 0, t/J is an ordinary eigenfunction.)
Fluid-dynamicists have always assumed that in normal problems there
are only ordinary eigenfunctions. In analogy with the finite-dimensional
case, it seems likely that existence of generalized eigen-functions is a non-
generic property of a fluid system. (See Appendix to Chapter 31.) An n x n
matrix A can be represented by a point in a space of V of n 2 dimensions. A
necessary condition for the matrix A to have a generalized eigenvector is that
the characteristic equation
def,
peA; A) = qet(A - AI) = (29.6-6)
have a multiple root, i.e., that it have a root A in common with the equation
d
dA peA; A) = 0. (29.6-7)
Elimination of A from (29.6-6) and (29.6-7) gives an algebraic equation for A;

hence only matrices A that lie on a surface in V, i.e., on a set of measure zero,
can have generalized eigenvectors.
We shall assume that the operator L has only ordinary eigenfunctions.
Incidentally, it is easy to check for the possible existence of generalized
eigenfunctions in computational problems in a Hilbert space. If one has
solved the equation Lt/J = At/J, then the adjoint equation L *X = 2X has
solutions; if it has a solution such that (X, t/J) f; 0, then there are no generalized
eigenvectors corresponding to the eigenvalue A.
In the hydrodynamic problems, the eigenvalues lie mostly on or near the
negative real axis; specifically they lie in a region bounded by a parabola:
Re A :s; Co - cl(lm ,1,)2 (Cl > 0);
Co generally increases with the Reynolds number R, but only a finite number
of the eigenvalues are in the right half-plane, for any R. Because the Navier-
Stokes equations are real, the eigenvalues either are real or appear in complex
conjugate pairs.
29.7 Reduction to a Finite-Dimensional

Dynamical System
To study the onset of turbulence, it is not necessary to know about all orbits
of the system (29.4-1) in the Hilbert space 5. It suffices to know about a
special family of orbits, which, according to the physical argument given
below, lie in the so-called unstable manifold that emerges from the origin in 5.
Reduction to a Finite-Dimensional Dynamical System 251
The motions in that manifold constitute a finite-dimensional dynamical

system. That idea was used by Davey, DiPrima, and Stuart 1968 and by
Ruelle and Takens 1971.
Stable and unstable manifolds play an important role in the theory of
differentiable dynamical systems; see Smale 1967, Abraham and Robbin
1967, and Kelley 1967. Much of the theory originated in celestial mechanics
and is not directly applicable to our problem because (1) the dynamical
systems of celestial mechanics are finite-dimensional to start with, while
those of hydrodynamics are not, and (2) the former are Hamiltonian, while
the latter are strongly dissipative and their solutions do not generally extend
to negative times, so that the former determine a flow in the Hilbert space
and the latter only a semiflow (but see Sell 1971, where much of the theory is
so formulated as to include semiflows). We shall follow a largely intuitive
approach based on physical considerations.
We assume that the Reynolds number has been maintained for a long time
at a value somewhat above the first critical value, and that for that Reynolds
number a finite number of the eigenvalues of the linearized problem, A.k
(k = 1, ... , K), are in the right half-plane Re A. > 0, while the rest are in the
left half-plane. We assume that at a very early time (very large negative t),
which we shall call the early linear regime, an arbitrary but very small dis-
turbance was present, consisting of a linear combination of all the normal
modes I/lkeAkt (k = 1 to 00). We assume that at a later time (t still negative),
the late linear regime, all the normal modes except the first K of them have
died out, and the disturbance is of the form of a linear combination
(29.7-1)
where the ak are constants. We assume that this disturbance is still suf-
ficiently small that these modes are growing exponentially and independently.
At a still later time, the nonlinear regime (which we think of as including the
"present" instant t = 0), the solution (29.7-1) has continued to grow until,
owing to the nonlinearities, it is no longer of that simple form (although it
still depends on the parameters ai' ... , aK), but may for example begin to
spiral toward a closed orbit or exhibit other complicated nonlinear behavior.
Figure 29.4
If we take a fixed t, say t = 0, and let the parameters a 1 , , aK vary, the

resulting points lie on K -dimensional surface or manifold Wl in the Hilbert
space tangent at the origin to the linear manifold Wl o sparined by the eigen-
vectors ljJ 1, ... , 1/1 K' (See Figure 29.4) Wl is invariant under the semi flow in f,
determined by (29.4-1) in the sense that any orbit that starts in Wl remains in
Wl, because, according to (29.7-1), a shift by to of the origin of time is simply
equivalent to changing the parameter values according to the scheme
ak -+ ak eAktO
The orbits that lie in Wl may be regarded as describing the motions of a
dynamical system with K degrees of freedom, whose properties we wish to
investigate.
Wl is called the unstable manifold in f, that issues or emerges from the
origin. The unstable manifold issuing from any other fixed point in f, can
be similarly defined after first linearizing the equation of evolution about that
point.
For unstable manifolds issuing from closed or quasiperiodic orbits, see
Abraham and Robbin 1967, Appendix C by Al Kelley.
The manifold Wl is locally attracting in the sense that there is a neighbor-
hood of the origin such that any orbit that remains in that neighborhood for
all t > 0 approaches Wl as t -+ 00. We shall not attempt to make that state-
ment more precise except in one case: If Wl results from a supercritical bi-
furcation at R = Ro as discussed in the remaining sections of this chapter,
then, for R > R c , Wl contains a new fixed point (in addition to the origin) or
a new closed orbit or invariant torus, which is near the origin and which is
stable, and in fact attracting, with respect to disturbances within Wl; then,
for R - Rc small enough, it is attracting also for arbitrary small disturbances
(not necessarily in Wl). For example, Davy 1962 showed that the Taylor
vortices are stable in the relevant 2-dimensional unstable manifold, and we
conclude that for small R - Rc they are stable with respect to arbitrary small
disturbances, as observed in experiments.
The local stability of Wl reflects, in the linear approximation, the exponen-
tial decrease with time of all normal modes except those appearing in the
construction of Wl-see (29.7-1).
Coordinates in Wl can be chosen in various ways. More suitable than the
parameters ab"" aK are coordinates Xl"'" XK based on a projection
onto the linear manifold Wl o to which Wl is tangent at the origin of f,. For
any u in f" that projection is given by an operator P defined by
K
Pu = L (Xb U)I/Ik, (29.7-2)
k=l
where the vectors {xd are eigenfunctions of the adjoint problem and form a
biorthogonal system with the {I/Ik}' Hence, for any u in the unstable manifold
Wl, we take the coordinates as
k = I, ... ,K. (29.7-3)

Reduction to a Finite-Dimensional Dynamical System 253
For orbits lying in m, the equations of motion (29.4-1) take the form
k = 1, ... ,K. (29.7-4)
For points near the origin, we have
F k(X 1, ... , x K) = AkXk + higher order terms. (29.7-5)
The calculations of the functions F k is described in the next chapter. It is
based on the idea that if U(X1, ... , x K) is the point ofm (a point in~) corres-
ponding to coordinate values Xl' ... , XK, and if Xk(t) (k = 1, ... , K) is any
solution of (29.7-4), then the quantity
u(t) = U(X1(t), ... , xK(t (29.7-6)
must satisfy the equation (29.4-1) of evolution in~. That requirement suffices
to determine both the dependence of the Xk( . ) on t and of u( . -) on the Xk .
The computational procedure assumes analyticity throughout, so that
U(X1, ... , XK) can be expanded in a power series in the Xk with coefficients
that are elements of ~ and the Fk(Xb ... , XK) as ordinary power series. That
assumption must be regarded as tentative, although the Navier-Stokes flow
in ~ is known to have at least Coo smoothness (see Marsden and McCracken
1976).
For the n-dimensional reversible systems of interest to celestial mechanics,
one defines also the stable manifold that emerges from the origin (similarly
from any other fixed point); it is tangent at the origin to the linear manifold
spanned by the remaining eigenvectors CfJK + 1, . . . , CfJn. It can be characterized
as consisting of motions u(t) such that u(t) ~ 0 as t ~ 00. In fact, the stable
manifold is usually discussed first, and then the unstable manifold is defined
as the stable manifold that would result from replacing t by - t. Although, in
hydrodynamics, most motions cannot be reversed in time, the particular
motions that lie on the unstable manifold m can be, and m can be character-
ized as consisting of these motions such that u(t) ~ 0 as t ~ - 00.
For use in Section 29.10, we mention another version of the unstable
manifold, which refers to mappings, rather than flows. In place of the family
of mappings u ~ CfJ(u, t) in a Hilbert space depending on a continuous
parameter t we consider a family of mappings x ~ <l>m(x) in an n-dimensional
manifold m depending on a discrete parameter m, given by iterating a
mapping <1>:
<l>m(X) = <1>(<1>( ... (x) .. (m iterations).
We suppose that x = 0 is a fixed point of and we linearize near 0:

(x) = Mx + higher order terms,
where M is an n x n matrix. If M has eigenvalues CX b , CXk (k < n) lying
outside the unit circle Icx I = 1 and corresponding to independent eigen-
vectors v1 , . , Vk , while the other eigenvalues all lie inside the unit circle,
then there is an invariant k-dimensional manifold ~n lying in m and tangent
at the origin to the linear manifold 91 0 spanned by v1, ... , Vk; see Abraham
and Robin 1967 or Smale 1967.
Note on Real and Complex Hilbert Spaces. In the

Hilbert space ~ of a fluid dynamical system the functions p, u, v, w (pressure
and fluid velocity components) are allowed to assume complex values, even
though they are real for physical flows, so that physical solutions are restric-
ted to a real subspace ~o of ~. Since the Navier-Stokes equations are real,
the eigenvalues of the linearized problem either are real or occur in complex
conjugate pairs. We can assume that an eigenfunction l/Ik is real if Ak is real
and that l/Ik and l/Ik' are complex conjugates if Ak and Ak, are. Then, in the
representation u = I Ck l/Ik of an element of ~, we can require that Ck be real
if Ak is and that Ck = Ck, if Ak = Xk,; then u lies in ~o. With that understanding,
the manifold m is a real K-dimensional surface tangent at the origin to a
real linear subspace mo of ~o, and the coordinates X k in mgiven by (29.7-3)
also have the property that X k is real if Ak is, and X k = Xk' if Ak = Xk ,.
29.8 Bifurcation to a New Steady State

In the simpler of the two classical Hopf bifurcation theorems, it is assumed
that one simple real eigenvalue Ai (R) crosses into the right half-plane (i.e.,
passes through the origin), as the Reynolds number R increases past the
critical value Rc:
(29.8-1)
Then the unstable manifold is one-dimensional, and the equation of motion
(29.7-4) takes the form
x = F(x; R), (29.8-2)
where we have suppressed the subscript 1 on x and have taken the dependence
on R into account. According to (29.7-5) and (29.8-1), this equation can be
.
R R
+
I I
,.
- ... ,
I
,
- - Rc - - , " - - _ " ' - " - -
.,.'" ,"
,, ,
~
\
-x -x _x
(a) (b) (c)
Figure 29.5
Bifurcation to a Periodic Orbit 255
Figure 29.6
expanded as
x = P(R - Re)x + higher order terms. (29.8-3)
The stationary orbits x = 0 are represented by the points on the locus
F(x; R) = 0 in the x, R plane. The locus consists of the R axis and a curve
passing through the point x = 0, R = RC' as shown in three cases in Figures
29.5a, b, c.
If the next lowest term in (29.8-3) is ax 2 , there is an un symmetric bifurca-
tion; if it is ax 3 , there is a symmetric one, which is supercritical if a < 0 and
subcritical if a > O.
Stability is determined by the sign of x at points near the curves. For
example, in the case of the unsymmetric bifurcation, the motion of points in
the x, R plane is indicated by the arrows in Figure 29.6. In all cases, upturning
branches are stable and down turning ones unstable, while the solution x =
is always unstable for R > R e

In the subcritical bifurcation illustrated in Figure 29.5c, there is no stable
equilibrium in the neighborhood of x = 0 for R > Re. In this case, if R is
increased very slowly past Rc> a typical orbit takes the system from x :::::
to distant points of the configuration space in a relatively short time as soon

as R exceeds Re. This phenomenon is called an explosive transition and
contrasts with the adiabatic sequence of stable states which the orbit follows
in the other cases.
29.9 Bifurcation to a Periodic Orbit
In the second case of the classical Hopf bifurcation theorem, it is assumed

that a single complex conjugate pair of eigenvalues crosses into the right
half-plane, as R increases past RC' while the rest remain in the left half-plane.
We write
,11, ,12 = a iw = a(R) iw(R), (29.9-1)
where
a(RJ = 0, a'(RJ > 0, w(RJ -=1= 0. (29.9-2)
The manifold 9)1 has two dimensions. In place of the complex conjugate
coordinates Xl and X2 in 9)1, we take real coordinates X and y such that
Xl = X + iy, X2 = X - iy.
To lowest order, a motion in ill1 is given, according to (29.7-4, 5), by
dt
~(x + iy) = Al(X + iy) = (a + iw)(x + iy).
Close to the origin, the orbits are approximately the spirals
(x + iy) ~ const. e<Tt(cos wt + i sin wt).
In polar coordinates,
f = ar + 0(r2),
(29.9-3)
(j = w + OCr).
It follows that in some neighborhood of the origin, e is always increasing,

on any orbit, and r always positive; r may increase or decrease; close to the
origin, r increases if a > and decreases if a < 0. We now investigate what
happens a little farther out.
We define the Poincare mapping of the problem as a mapping X ~ <l>(x)
of the x axis in 9)1, by saying that if an orbit has coordinates x, 0, for some t,
then it has coordinates <l>(x), 0, when f) has increased by 2n. Note that x and
<l>(x) can be either both positive or both negative. We write
<l>(x) = x[l + g(x)]. (29.9-4)
The orbit structure depends on the properties of the function g(x). From the
spirals near the origin, we see that
1 + g(O) = e2 "<T/w; (29.9-5)

in particular, g(O) = for R = RC' because then
g(x) in a Taylor series in x and R - RC' we have
(J = 0. Hence, if we expand
g(x) = g(x, R) = ax + b(R - RJ + cx 2 + .... (29.9-6)

The coefficient a must be = 0, for otherwise g(x, Rc) would have opposite
signs for x > and x < near x = 0, and that would imply that an orbit
crosses itself, i.e., its second turn would be farther from the origin than its
first turn on one side and closer to the origin on the other. The coefficient b
is positive, because, under the assumption (29.9-2) about a, orbits near the
origin spiral out for R > Rc and in for R < Rc.
Bifurcation from a Periodic Orbit to an Invariant Torus 257
We consider the locus in an x, R plane of the equation g(x, R) = 0. If

xo, Ro is a point on that locus, then, for R = R o , <D(x o) = xo, that is, there
is a closed orbit that cuts the x axis at x = Xo. The locus contains the point
x = 0, R = Rc; near that point it is a curve that cuts the R axis horizontally;
it turns upward, as in Figure 29.5b, if c < 0, and it turns down, as in Figure
29.5c, if c > 0. The bifurcation is called supercritical in the first of these
cases and subcritical in the second. As in the preceding section, one must
expect an explosive transition, if c > 0.
29.10 Bifurcation from a Periodic Orbit to an

Invariant Torus
The next bifurcation, after one that results in a closed orbit, hence a periodic
motion, can result in an invariant 2-dimensional torus, as shown by the
example of Hopf 1948. Theorems on such a bifurcation have been given by
various authors, including Naimark 1959, Sacker 1964, Ruelle and Takens
1971, and Lanford 1973. The theorems have been based largely on Floquet
theory, but we shall take a more intuitive approach, based on the notion of
a Poincare mapping.
Let R 1 be the critical value of the Reynolds number R for first appearance
of the periodic orbits in a supercritical bifurcation, as discussed in the pre-
ceding sections. We assume that for some R > Rl the unstable manifold
9Jl has dimension K. 9Jl contains the 2-dimensional manifold discussed in
the preceding section, and we suppose the coordinates in 9Jl so chosen that
the first two of them are the coordinates x, y of the preceding section; the re-
mainder are called X3, .. , XK. Then, if R is not too much above Rb the
closed orbit encircles the origin in the x, y subspace and cuts the positive
and negative x axes once each. Let V be the (K - I)-dimensional hyper-
surface in 9Jl given by y = 0; then, V is intersected twice by the closed orbit,
as shown schematically in Figure 29.7, and we denote one of the intersections
v .. x
"\'\'
""
4>(x)
Figure 29.7
by x = ~ = ~(R), where x is a vector with K - 1 components x, X3' . , XK'

For x near ~, let <D(x) denote the second intersection with V of an orbit that
starts at x, as shown. Then
<D: x ~ <D(x)
is a Poincare mapping defined in some neighborhood of ~ in V. Note that
q,(~) = ~. For x near ~, we linearize:
<D(x) - ~ = M(x - ~) + higher order terms, (29.10-1)

where M is a (K - 1) x (K - 1) matrix.
We now assume that M has one pair of eigenvalues r:x [ = r:x(R)] and i'i with
corresponding eigenvectors v, v, such that
Ir:x(R2) I = 1 for some R2 > R i , (29.10-2)
d
dR Ir:x(R)IR=R2 > 0, (29.10-3)
while the other eigenvalues of M all lie inside the unit circle. As R is increased,
the closed orbit loses stability at R > R 2 ; hence, a new bifurcation occurs at
R = R2
For R > R 2 , according to Section 29.7, the mapping <D in V has a 2-
dimensional unstable manifold or surface S in V tangent at ~ to the linear
manifold So spanned by v and v. To simplify visualization, regard V as 2-
dimensional; then S = v. S is invariant under q,; hence we may think of <D
as a mapping in S. We let u and v be real coordinates in S such that the pro-
jection of x - ~ from S onto So is (u + iv)v + (u - iv)v, and we call z = u
+ iv. Then 'the Poincare mapping becomes z ~ z' = r:xz + higher order
terms, or, more explicitly,
<D: z --> z' = r:xZ[1 + I

j+k=i
qjkZjzk], (29.10-4)
where the qjk are coefficients (generally complex) and where the summation
is over all nonnegative integers j and k such that j + k ~ 1.
It will be shown, under certain further assumptions, that for R > R2 there
is a nearly circular invariant closed curve rc in S encircling the point~. Then,
if we let rc be carried along by the flow in m, it leaves the hypersurface V and
traces out an invariant tube in m
which is then closed, forming a torus, when
rc passes through V again near ~.
To facilitate analysis of the Poincare mapping, it is convenient to in-
troduce new coordinates ~ and 1'/ in S in place of u and v, so chosen as to make
the Poincare mapping <D take as simple a form as possible (a normal form),
namely to eliminate some of the nonlinear terms in (29.10-4). If Ir:xl were "# 1,
we could eliminate as many of the nonlinear terms as we wish (see Siegel,
Himmelsmechanik, 1956, Section 21), but in order to let R vary through the
value R 2, where Ir:x(R) I = 1, without a singularity of the transformation, we
must retain the term r:xqi i Z2 z in (29.10-4); otherwise, we can eliminate all
Bifurcation from a Periodic Orbit to an Invariant Torus 259
terms of order 54. (In principle, we could eliminate terms of order 56 by

retaining a term exQ22z3z2, and so on, but the elimination of terms of order
54 is just what is needed for our analysis.) We transform to ( = ~ + il'/ by
the equations
4
Z = (+ L CPlm (l(m, (29.10-5)
l+m=2
4
z' = (' + L CPlm(,IC m (29.10-6)
l+m=2
We shall choose the coefficients ({JIm so that when these equations are sub-
stituted into (29.10-4), the Pointcare mapping takes the normalform
(29.10-7)
where 0((5) contains terms of degree 5 and higher in ( and ( and where /3 is
a new constant. To determine the coefficients CPlm in terms of the given
coefficients Qjk, we substitute the right member of (29.10-5) for z into the
right member of(29.1O-4); then we substitute the right member of(29.10-7) for
(' into the right member of (29.10-6) and take the result as z' for the left
member of (29.10-4). Equation (29.10-4) thereby becomes an identity in ( and
~, and the net coefficient of (P~q can be equated to zero. The resulting equa-
tions, if taken in the right order, can be solved for the coefficients CPlm pro-
vided that
(29.10-8)
as explained in the Appendix to this chapter. (The equation for the case
p = 2, Q = 1, cannot be solved for CP2 1 when Iex I = 1, but can always be
solved for /3, and we simply set CP2 1 = 0.)
For this method of analysis to succeed, we must assume, according to
(29.10-8), that as R is increased and the point ex(R) crosses the unit circle
Iex I = 1 in the complex plane, it crosses at a point which is not a root of
unity of atly order less than 6. What happens if it crosses at such a point is
discussed briefly in the next section.
If we let 11 be a new dimensionless parameter in place of R given by 11 =
IexCR) I - 1, and if rand () are polar coordinates, given by ( = re i6 , the Poincare
mapping takes the form
il: {r' = (1 + Il)r + clr 3 + fer, ()r 5, (29.10-9)
()' = () + C2 + C3 r2 + g(r, ()r4, (29.10-10)
where f and g are smooth functions and where C l , C 2 , C3, f, and g all depend
smoothly on 11 in some neighborhood of 11 = o.
The constant C l plays the role of the Landau constant; if C l < 0, which
we shall assume, we have a supercritical bifurcation.
If the higher order term containing fer, () in (29.10-9) were missing, the
circle 1(60: r = r0, where
would be invariant under the Poincare mapping, for j1 > O. It is centered

on the closed orbit shown in Figure 29.7; hence, as it is carried along by the
flow in Wl, it would generate a torus. To take into account the effect of the
higher order terms, we define a sequence of curves
inductively, starting with ~o, by the equation ~n+l = <D(~n). It can be

shown that for j1 small enough (R - R2 small enough), these curves con-
verge, as n --+ 00, to a limiting curve ~ 00' which is invariant under <D.
The proof of convergence and the derivation of the restrictions on j1 are
quite long and are omitted, except for an outline and a few formulas given in
the Appendix to this chapter. It is shown, for example, that if j1 is small
enough, the annulus
(29.10-11)
is mapped under I> into the smaller annulus
(1 + l7j1)rO < r' < (% - 2~j1)rO; (29.10-12)
that suggests that I> is a contracting mapping. It is also shown that if the
maximum radial displacement of ~ n under I> is denoted by
Ll n = maxlrn+l(8) - rn(8)1,
(8)
then
Lln+ 1 :s; K Lln' (29.1O-l3)
for a constant K < 1, if j1 is small enough. Finally, it is shown that
(29.10-14)
if j1 is small enough, so that the limiting curve ~ 00 is at least Lipschitz con-

tinuous.
Lastly, we consider the tube in the K-dimensional space Wl consisting of
orbits that start from points of the curve ~ 00. That tube joins onto itself,
forming a closed surface, when the orbits return to the surface S, where they
again pass through the curve ~ 00. To see that the surface is homeomorphic
to a torus, not a Klein bottle, we have to be sure that the orientation of~ 00 is
preserved when the tube is closed, but that follows from the Poincare
mapping (29.10-9, 10), which shows that 8' in an increasing function of 8.
It should be noted that we have not shown that any single orbit covers the
resulting torus densely, and hence yields a quasiperiodic function of time.
In fact, it will be seen in Chapter 31 that that is in general unlikely, owing to
Peixoto's theorem.
Appendix to Chapter 29-Computational Details for the Invariant Torus 261
29.11 Subharmonic Bifurcation
If, in the notation of the preceding section, a(R) passes through the unit
circle Ia I = 1 at a root of unity of degree less than 6, as R increases, the
argument given for the existence of an invariant torus breaks down. In that
case, the bifurcation may lead to one or more further periodic orbits.
We consider only the simplest case in which the higher terms of the
mapping (29.10-4) are mostly already missing, and hence do not have to be
eliminated. We assume a = exp{2nip/q} (q ~ 5) and we take the mapping
to be
(f3 real, < 0)
or, in polar coordinates,
r' = (1 + J.1)f + Clf 3 (Cl < 0),
()' = () + 2n ~. (29.11-1)
q
J
In this case there are new orbits, at a distance fo = J.1/( -Cl) (in the surface
S) from the old orbit, which are closed, because each point of the circle f = fo
is transformed into itself after q-fold iteration of the Poincare mapping.
For small positive J.1 the old orbit is unstable, if Cl < 0, and the new ones
stable. As J.1 is increased past 0, the period of the observed orbit is suddenly
increased by a factor q.
In this case there may still be an invariant torus as result ofthe bifurcation,
although our method of finding it breaks down. In the above example, there
is such a torus, because the circle r()) = fo is invariant.
Conversely, even when the invariant torus is established, there may be
points on the curve j 00 that are invariant under a certain number of iterations
of the Poincare mapping, and hence lead to closed orbits on the torus.
Appendix to Chapter 29-Computational Details

for the Invariant Torus
We consider first the problem of determining the coefficients CPlm of (29.10-5,
6) from the given coefficients qjk of (29.10-4) so as to make the Poincare
mapping take the normal form (29.10-7). As noted in the text, we substitute
(29.10-5, 6, 7) into (29.10-4) to give an equation in powers of ( and (, and
we equate the net coefficients of (p(q. For (p, q) = (0, 0), (1, 0), and (0, 1), the
resulting equations are satisfied automatically. The next twelve equations
are taken in the order
(p, q) = (2, 0) (1, 1) (0,2)
(3,0) (2, 1) (1,2) (0, 3)
(4,0) (3, 1) (2,2) (1,3) (0,4)
and are solved for the coefficients C(Jpq in that order; then C(Jpq is the only
unknown in the equation in which it first appears, and it appears in the
following way:
aPfi.qC(Jpq + ... = aC(Jpq + ....
We can solve for C(Jpq except when
Since p +q> 1, this can happen only if Ia I = 1 and then only when
alp+q-11 = 1.
Unless p + q - 1 = 0, this last equation holds only when a is a Ip + q - 11th
root of unity, and in this way tile restrictions (29.10-8) are obtained. The case
p + q - 1 = 0 occurs only for (p, q) = (2,1); that equation cannot be
solved for C(J21 if lal = 1, but it can be solved for fJ; hence we merely set
C(J21 = o.
We also record, without derivation, some bounds on fl, which suffice to
ensure the existence of the invariant limiting curve CC 00 discussed in the text.
First, the annulus (29.10-11) is mapped into the annulus (29.10-12) if
maxlflfl < fl < 133,
1t;~4d and
where max I f I means the maximum of I fer, 8, fl) I in some region r :$ rb
Ifll :$ fl1 in which the normal form (20.10-9, 10) holds. Second, the bound
(29.10-14) on drn/d8 holds if also
maxl.f..ld/13/2 + 5 maxlflc:fl + maxlfoiclh < i

and
21c3ic4h + max!grlc:fl2 + 4 maxlgldfl 3/2 + maxlgo!c:fl < !,
where
4
= .
C4
3J - C1
Lastly, (29.10-13) holds with K = 1 - fl/6, if also

21c3ic4J; + maxlgr lc:fl 2 + 4 maxlgldfl 3/2 + 5 maxi f !C:fl
+ maxl.f..lclfl 3/2 <!.
CHAPTER 30
Invariant Manifolds in the Taylor

Problem
Power series representation of a finite-dimensional manifold in a Hilbert

space; coordinates in the manifold; determination of the expansion
coefficients; dynamical system in the manifold; separation of variables;
Taylor vortices, wavy vortices; helical vortices.
Prerequisites: Chapter 29.
This chapter is devoted to a method for calculating the unstable manifold

that arises from the basic flow in hydrodynamical problems, devised by
Davey 1962, Davey, DiPrima, and Stuart 1968, and Eagles 1971 for the study
of the Taylor problem. Although the results that have been obtained with
it so far are rather limited, it is the only method known at present for such
problems (but see Hassard 1980 for a similar method in finite-dimensional
spaces). Presumably, methods of this sort will be needed to understand the
early onset of turbulence in detail.
30.1 Survey of the Taylor Problem to 1968
Let rl and r2 be the radii of the inner and outer cylinders (assumed infinitely
long) that confine the flow, and let Q 1 and Q 2 be their angular velocities.
The unperturbed flow has angular velocity at radius r (rl ::; r ::; r2) given by
(30.1-1)
where A and B are determined by the no-slip conditions at the walls, namely
Q 1 = A + Bid and Q 2 = A + Bid. The problem is characterized by
various dimensionless parameters, for example rt/r2 (which is fixed in a
given apparatus) and the two Reynolds numbers
(30.1-2)
263
264 Invariant Manifolds in the Taylor Problem
The stability of the basic laminar Couette flow given by (30.1-1) against
small axisymmetric disturbances was studied by G. 1. Taylor 1923, both
theoretically and experimentally. He found stability for values of R1 and It{2
represented by points lying below the curve in Figure 30.1, for the case
rdr2 = 0.880. Rayleigh had shown earlier, by a simple argument, that the
flow is stable for 0 < R1 < R 2 , that is, to the right of the dashed line in the
figure.
Taylor's calculations showed that, as R1 is increased beyond its critical
value (the value on the curve) for fixed R 2 , the basic flow becomes unstable
with respect to a normal mode whose velocity field is of the form
u = Re[f(r)e iIlZ ] (30.1-3)
in cylindrical coordinates, r, e, z. This disturbance has the structure of a
series of circular vortices, uniformly spaced in the z direction, as sketched in
Figure 29.1 in the preceding chapter. Taylor found that the experimentally
observed vortices (which can be made visible by means of particles suspended
in the fluid) were in qualitative agreement with the calculations; in particular,
their axial separation was in agreement with the value of rJ. in (30.1-3) for
which the disturbance first becomes unstable (i.e., becomes unstable for the
smallest value of R 1).
For R1 not too far above its critical value, the vortices (now called Taylor
vortices) are stable; together with the basic flow, on which they are super-
posed, they form a new steady flow (at each point in space, the fluid velocity
vector is independent of time), which persists indefinitely, so long as the
cylinders are kept rotating.
Taylor observed experimentally that when a still higher critical value of
R1 is reached, the vortices become wavy, and the waves rotate around the
axis at approximately the mean angular velocity (0 1 + O2 )/2.
Taylor's analysis, being linear, said nothing about the flow except below
and immediately above the first critical value of R1 and said nothing about
7
/'
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
----------------------~/----------------------~R2
Figure 30.1 Taylor's stability diagram (schematic).
Calculation of Invariant Manifolds 265
the stability or instability of the Taylor vortices, once they have formed.
Evidently, for Rl in the range where the Taylor vortices are observed, the
exponential growth of the normal mode (30.1-3) levels off at a finite ampli-
tude, owing to the effect of nonlinearities.
Subsequent improved theoretical and experimental work has generally
confirmed Taylor's results. However, Taylor considered only axisymmetric
disturbances, and it was shown by Krueger, Gross, and DiPrima 1966 that,
for counter-rotating cylinders with n 2 /n 1 sufficiently negative (beyond the
range studied experimentally by Taylor), the mode that first becomes un-
stable is not axisymmetric, but has an angular dependence of the form eimO ;
m increases through the values 1,2,3, ... as n 2 /n 1 is made more and more
negative. Hence the leftmost part of the curve in Figure 30.1 must be revised
downward, but only by a quite small amount, because the stability of the
mode depends only very slightly on m, in the case of the rather narrow gap
(rdr2 = 0.880) studied by Taylor.
These predictions were confirmed experimentally by Snyder 1970, who
showed also that the nonaxisymmetric modes, when they appear, are helical.
(The linear theory cannot distinguish helical vortices from wavy ring vortices;
the four normal modes containing eiazeime are all equally likely and can be
combined to give real dependence either as ~~~ rl.Z ~~~ me or ~~~(rl.Z me); only
a nonlinear theory can say which of these is preferred.)
In recent years, the nonlinear theory developed by Davey 1962, Davey,
DiPrima, and Stuart 1968, and Eagles 1971 has led to an understanding of the
structure and stability of the finite-amplitude Taylor vortices, the second
bifurcation to the wavy vortices, and the structure and stability of the wavy
vortices, as discussed below.
In a problem like this, involving a sequence of bifurcations, the theory
consists ideally of a sequence of alternately linear and nonlinear investiga-
tions. After each bifurcation the structure and amplitude of the new flow is
found by a nonlinear calculation. Its stability is then investigated by linear-
izing the equations about the new flow and studying the growth of
infinitesimal disturbances, in order to find the next bifurcation, and so on.
30.2 Calculation of Invariant Manifold&

In this section we describe in general terms the method for calculating the
unstable manifold developed by the authors mentioned above on the basis
of earlier work by Stuart and Watson.
We take the equation of evolution in the form (29.4-3), so that the eigen-
values and eigenfunctions of the linearized problem satisfy the equations
(j = 1,2, ... ). (30.2-1)
We assume, as in Section 29.6, that the tf;j form a complete set in f), and that
the adjoint functions, which satisfy the equation
(f = 1,2, ... ), (30.2-2)
form a biorthogonal set with the if;j in the sense that

(Xj' Mif;k) = bjk (30.2-3)
We assume that the unstable manifold Wl is K-dimensional, and we take
coordinates of a point U on Wl by projection onto Wl o, as in Section 29.7:
Xk = (Xb Mu) (k = 1, ... , K). (20.2-4)
Therefore, for U in Wl,
K
U= L xkif;k + u', (30.2-5)
k=l
where u' is orthogonal to Xl' ... , XK:

(Xk, Mu') = 0 (k = 1, ... , K). (30.2-6)
The objective is to find a function U(X1, . , XK), or u(x) for short, with
values in~, which gives the point of9Jl having coordinate values Xl' . , XK,
at least in some neighborhood of the origin. We assume that the function is
analytic, and hence can be expanded in a power series
u(x) = L Xquq, (30.2-7)
(qe.P)
where x and q are K -component vectors, where 2 denotes the lattice

2 = {q: each qi = integer ~ 0; at least one qi > O}, (30.2-8)
where x q is an abbreviation for XilX~2 ... x'Jt, and where each coefficient uq
is an element of ~.
According to (30.2-5, 6), the coefficients uq of the linear terms in (30.2-7)
are the first K eigenfunctions if; 1, . . . , ~f K, while the remaining uq are ortho-
gonal to Xb ... , XK in the sense of (30.2-6). However, the later coefficients
uq are not eigenfunctions, but satisfy certain inhomogeneous equations,
described below; therefore, (30.2-7) is not an eigenfunction expansion.
We assume that the functions in (29.7-4), which describe the dynamical
system in Wl, are also analytic and can be expanded in power series in
Xb ... , XK, and we write
Xj = L ajpx P (j = 1, ... , K), (30.2-9)

(pdf)
where the ajp are numerical coefficients.

We compute it = du/dt from (30.2-7, 9), using the chain rule for differentia-
tion. Then we substitute U and it into the equation of evolution (29.4-3) and
require the result to be an identity in Xl' .. , XK. That requirement suffices
to determine the coefficients uq and ajp.
To compute it from (30.2-7), we must differentiate x q with respect to Xj'
for eachj, and then use (30.2-9). We call ej the K-component vector whosejth
component is equal to 1 while the other components are equal to O. Then,
(30.2-10)
Calculation of Invariant Manifolds 267
hence,
K
U= "L.. "q.a.
L.. j jp
Xp+q-eju
q
(p,qE.2') j~l
K
= I
(SE.2')
XS I I' qjajs+ej_qUq,
j~ 1 (q)
(30.2-11)
where I'
denotes the sum over all q in 2 such that s + ej - q is also in 2;
it is a finite sum, for given s.
It is convenient to introduce a norm for vectors in the lattice 2: Iq I =
If~ 1 qj; Iql is a positive integer. If Iq I = 1, q is one of the vectors ej .
We now substitute U from (30.2-7) and U from (30.2-11) into the equation
of evolution (29.4-3) and equate the net coefficients of X S on the two sides,
for each sin 2. First, ifsis one of the vectors el , the quadratic terms contribute
nothing, and we have
K
I ajelMuej - LU el = 0 (l = 1, ... , K). (30.2-12)
j~ 1
But we have already seen that uej is = I/lj; hence [from (30.2-1)J
AI for j = I,
ajel = { 0 for j =f. l. (30.2-13)
To lowest order for small Xl> ... , XK, (30.2-9) then gives Xj = const. exp(Ajt},
as expected.
Second, if s is not one of the el , that is, if Is I > 1, we find, using (30.2-13),
(f
j ~ 1
sjAjM - L)U s = -
j
f I"
~1 (q)
qjajs+erqMUq + I"'B(uq, U
(q)
S - q ), (30.2-14)
where I"
denotes the sum over all q in 2 such that s + ej - q is also in 2
and q =f. s, and I'" denotes the sum over all q in 2 such that s - q is also in
2; these are finite sums.
We now show that these equations, if taken in a suitable order for the
various s in 2, determine the unknown functions Us and unknown coefficients
ajs inductively. We assume the equations (30.2-14) so ordered that all equa-
tions with a given value of IS I appear earlier than the equations with one
higher value of Is I, and we assume that when any of the equations is en-
countered, all functions and coefficients appearing in previous equations
have been determined. (The ordering of the equations with a given value of
lsi is irrelevant.) We claim that then all the u q appearing on the right of
(30.2-14) are known. In fact, for most of the terms there, Iql is less than
lsi; the only exceptions are terms in which q is of the form s + ej - el ,
for some 1 =f. j (recall that q =f. s); however, the term with that value of q
contains the coefficient qj el' which is = 0 by (30.2-13); hence, all the uq that
appear on the right of the equation may be regarded as known. The un-
knowns in that equation are therefore the function Us and the coefficients
al s (l = 1, ... , K).
To determine the coefficient al s ' for any 1 = 1, ... , K, we take the inner
product with Xl throughout (30.2-14). The left member gives zero, and all
the terms in the first sum on the right give zero, except when q = el andj = I;
hence,
al s = - (Xl' I"'B(u q , US - q )).

(q)
With these coefficients known for 1 = 1, ... , K, equation (30.2-14) can then
be solved for Us' In this way, the invariant K-dimensional manifold 9Jl is
determined in that neighborhood of the origin in f, in which the series
(30.2-7, 9) converge.
This is the method of Davey, of Davey, DiPrima, and Stuart, and of
Eagles. Although those authors don't describe it in those terms, its main
feature is a calculation of the unstable manifold of the origin in f,; then the
equations (30.2-9) represent a finite-dimensional dynamical system in that
manifold and can be studied by any of the standard methods, by analytical
search for fixed points and cycles, or by calculations of orbits by numerical
methods for ordinary differential equations.
An important feature of their method is the assumed orthogonality
of all the uq , except for Iql = 1, to the adjoint functions Xl>"" XK' We
have presented that assumption by equations (30.2-5, 6) as simply a con-
venient way of prescribing the coordinates Xl"'" x K in the unstable
manifold, but it has a deeper significance, and is the fundamental idea that
makes the method successful. Each Us is determined in terms of known
quantities by an inhomogeneous equation (30.2-14), in practice a differential
equation with boundary conditions. For some of those equations, when the
Reynolds number is close to one of the critical values, the operator on the
left is nearly singular, and in fact is singular when the Reynolds number is
equal to that value. According to the alternative theorem of linear algebra,
which applies also to linear equations in a Hilbert space, singular inhomo-
geneous linear equations have no solution unless the right member is
orthogonal to all solutions of the corresponding transposed homogeneous
equation, which are just Xl' ... , XK; if it is, the solutions are nonunique, but
they can be made unique by requiring that they also be orthogonal to
Xl>"" XK'
If this assumption, or something similar, is not made, some of the uq
can be unreasonably large, for Reynolds numbers of interest, presumably
preventing the convergence of the series (30.2-7).
30.3 Cylindrical Coordinates

The Navier-Stokes equations (29.3-4 and 5) are
au
at + (u V)u + (u . V)u + (u . V)u + Vp - vV 2 u = 0, (30.3-1)
V. u = 0, (30.3-2)
Cylindrical Coordinates 269
where u(x) is the basic laminar flow (Couette flow for the Taylor Problem).
We introduce cylindrical coordinates r, e, z and corresponding velocity
components u, v, w, so that
u = ukr + vke + wk z , (30.3-3)

where k" ke, k z are unit vectors in the directions of increasing r, e, and z,
respectively. In the cylindrical coordinates, the operators V2 and u' V
have the forms
(30.3-4)
When these operators are applied to a vector field in the form (30.3-3),
the dependence of the unit vectors kr and ke on emust be taken into account:
o
oe ke = -kr
For the Taylor problem, the basic laminar flow is given, according to
(30.1-1), by u = 0, W = 0, and
v = v(r) = Ar + B/r. (30.3-5)

When the foregoing equations are combined, they give a simultaneous
system of partial differential equations for the quantities u, v, w, and p, as
functions of r, e, z, and t. For the practical application of the method described
in the preceding section, it is convenient to put those equations in the form
introduced by Eagles 1971, by the introduction of six-component vectors
U, V, etc., whose components are p, ov/or, ow/or, u, v, w. The system then
takes the form
o 0
- U - AU - M - U - K(U)U = 0 (30.3-6)
or ot '
where A, M, and K(U) are 6 x 6 operator-valued matrices contammg
%e and %z, but not %r. [To achieve the last, the divergence equation
V . u = 0 has been used to eliminate explicit reference to ou/or, except in
the first term on the left of (30.3-6).J The matrices are written out in the
Appendix to this chapter; they are essentially the same as in the paper of
Eagles, but in a slightly different notation. All matrix elements of M and
K(U) are zero except in the upper right quarter of the matrices. To explain
the notation K(U), we may consider the expression K(U)V; the matrix
elements depend linearly on the components of U and act linearly on the
components of V.
30.4 The Hilbert Space
For the Taylor problem, theory and experiment agree in indicating that,
once the Taylor vortices (or wavy vortices, or helical vortices) have been
established, the entire flow is periodic in the z direction with a period of the
order of twice the separation of the cylinders, at least when the cylinders are
very long, and in the approximation in which end effects are neglected.
Here, we shall simply assume such periodicity, and we shall assume that the
wave number rx is known, so that the period is 2n/rx. We call &It the region
given by
2n
&It: 0::;; e::;; 2n, O::;;z::;;-, (30.4-1)
!Y.
and we take f> as the Hilbert space L2(&It)6 with the inner product
(U, V) = fff D
at
V drdedz. (30.4-2)
With this choice of inner product, the eigenvalue problem adjoint to

a
- U - AU - AMU = 0
or ' (30.4-3)
U4 = U 5 = U 6 = 0 at r = r 1 and
is the problem
a
- - U - A*U - AM*U
- =0
or ' (30.4-4)
U 1 =U 2 =U 3 =0 atr=r 1 and r=r2'
where A* and M* are the transposes of the matrices A and M with %e
and %z replaced by -%e and -a/oz.
As in the preceding section, it is assumed that:
1. For a given value of A, the problem (30.4-3) has a solution if and only
if the problem (30.4-4) has one.
2. Each of the problems has a complete set of eigenfunctions, and for this
purpose it is not necessary to include generalized eigenfunctions of
higher order, e.g., solutions of (a/or - A - AM)2U = 0, etc.
These assumptions have been confirmed as far as possible by numerical
calculations involving the first ~ 40 eigenfunctions. The eigenfunctions
U(j) of the direct problem and those UW) of the adjoint problem can be
chosen to be biorthogonal in the sense that
(30.4-5)
The choice (30.4-2) of inner product is somewhat arbitrary. For example,
it might seem to be natural to have rdr appear in place of dr. Then, the adjoint
Separation of Variables in Cylindrical Coordinates 271
equation (30.4-4) would be somewhat different, and the adjoint functions

utu) would be different, but the biorthogonality relations (30.4-5) would
continue to hold. It is only those relations that are important; they are used
to construct the projections that project the points of the unstable manifold 9Jl
onto the corresponding linear manifold 9Jlo tangent to 9Jl at the origin.
There is no choice of inner product that makes the operators self-adjoint.
30.5 Separation of Variables in Cylindrical Coordinates

The variables can be separated in the linearized problem, and in particular
in the eigenvalue problems (30.4-3, 4), so that the eigenfunctions are of the
form U = V(r)eiPClZ+im6, where p and m are integers; we write the kth eigen-
function as
k = 1,2,3, .... (30.5-1)
One of the advantages of the method of Davey, DiPrima, and Stuart,
described in Section 30.2, is that, although the variables don't separate in
the nonlinear problem, they do separate in each term of tHe expansion
(30.2-7) of the unstable manifold in the coordinates Xl>'" ,'XK' The co-
efficient uq that appears there is an element of the Hilbert: space, hence
represents a vector function of r, e, z in the region f!Jl, and that function
has just the form (30.5-1).
Specifically, those coefficients, which are now called U q [q being a point
in the lattice ! described by (30.2-8)] and the numerical coefficients ajq
that appear in (30.2-9) satisfy relations as follows:
Lemma. Forgeneraljandq,ajq = ounless p(q) = p(e) and m(q) = m(ej);

furthermore, U q is of the form
Uq = V q(r)e ip(q)ClZ+im(q)6.
Here p(q) and m(q) are given by

K
p(q) = L qkPk,
k=l
K
m(q) = L qkmk,
k=l
and ej is that vector in the lattice ! having its jth component = 1 and its
other components = O.
The proof of the lemma is by an easy induction on the norm iqi = ql

+ ... + qK introduced in Section 30.2, and is omitted.
The advantage of the separation of variables is that the linear equations
that have to be solved for the uq (30.2-14) are ordinary differential equations
for functions of r in the interval [r l , r2]. They are sixth-order two-point
boundary-value problems, with three boundary conditions at each end of

the interval. There are generally many such equations per problem (for
example, 800 of them when the number of dimensions K of the unstable
manifold is = 14 and the expansions (30.2-7, 9) are carried through fifth-
degree terms), but the available numerical methods for such equations are
faster and more accurate than methods for partial differential equations.
30.6 Results to Date for the Taylor Problem

We summarize here the general conclusions that come from the calculations
of Davey, Di Prima, and Stuart 1968, those of Eagles 1971, and a few ad-
ditional calculations that I have made recently on the Cray 1 Computer at
NCAR (the National Center for Atmospheric Research, Boulder, Colorado,
sponsored by the National Science Foundation).
The outer cylinder is taken to be at rest (Q 2 = 0). It is then customary to
express the rate of rotation in terms of the Taylor number
2Qiri(r2 -- r )3
T=--=---c:--,--=-- 1
v(rl 1- r2) ,
which is proportional to the square of the Reynolds number Rl (R2 is
zero).
As T is increased, for fixed rdr2' the first eigenvalue of the linearized
problem to cross into the right half-plane is real and corresponds to mk = 0
in the expression (30.5-1) for the eigenfunction. Next is a complex conjugate
pair of eigenvalues corresponding to mk = 2, and so on. Each eigenvalue is of
multiplicity 2 (the degeneracy corresponds to the possibility of shifting the
entire flow pattern in the z direction); hence the number of dimensions K
of the unstable manifold takes on in succession the values 2, 6, 10, 14, ....
Calculations have been made up to K = 14. (At still higher values of T, the
values mk = 0, 1, ... come in again, but corresponding to eigenfunctions with
more complicated radial dependence; no calculations have been made in this
regime.)
The modes of motion (stable and unstable) that have been discovered so
far are all periodic orbits of the dynamical system in the unstable manifold
[see (30.2-9)J, except for the basic laminar (Couette) flow and the Taylor
vortices, which are fixed points of that system. The additional modes are as
follows:
Helical vortices: These are similar to the Taylor vortices, except that if we
follow one of them around in () from 0 to 2n, we find that its end is not con-
nected onto its own beginning, but onto the beginning of the second, or
fourth, or sixth, ... , one above (or below) it. We define a corresponding
integer m = 1,2, 3, .... (It cannot be connected onto the first or third, etc.,
because those vortices rotate in the opposite direction.) The entire pattern
rotates about the axis with an angular velocity not far from the mean velocity
Qd2 of the fluid.
Results to Date for the Taylor Problem 273
(...._ _ _ _ ) ('-1-_ _ _ _) c....__

c. . ____ ) ('-1-_ _ _ _) c. . __
------+-------------~6
Figure 30.2
Nonaxisymmetric simple mode (never stable): The vortex strength varies

sinusoidally around the axis. In effect the vortex cores are connected as
shown schematically in Figure 30.2. The pattern rotates about the axis.
Up-down wavy vortices: These are similar to the Taylor vortices, but the
vortex cores are displaced alternately in the + and - z direction as e
goes from 0 to 2n. We call m (= 1,2, ... ) the number of such displacements
in each direction.
In-out wavy vortices: Similar to the up-down wavy vortices, except that
the cores are displaced alternately in and out in the radial direction. They
are never stable. Both types of wavy vortices rotate about the axis with an
angular velocity ~nd2.
In all cases it is found that as the Taylor number T is increased past a
first critical value Tb the laminar flow becomes unstable and is replaced by
the Taylor vortices, which have a stength roughly proportional to JT - T1 ,
and which are then stable up to a second critical value T2 , where the waviness
J
sets in with an amplitude roughly proportional to T - T2
The helical vortices bifurcate from the basic laminar flow above T1 ,
i.e., after the basic flow has already become unstable; see Figure 30.3. The
Taylor vortices
Figure 30.3 Inaccessible stable helical
vortices.
helical vortices are unstable when they first appear, but then become stable
at slightly higher values of T, as indicated by the solid curves in the figure.
Then, they are stable modes that are inaccessible in the sense that they cannot
be reached from the basic flow by a continuous sequence of stable modes.
The experimental work of Gollub and Swinney 1975 indicates that,
at a Taylor number of the order of 200T 10 a strange attractor should appear,
because they observe a continuous power spectrum. It is not possible to
extend the calculations to such high values of T, because the number of
dimensions of the unstable manifold becomes unmanageably large. It is of
course possible to continue calculating with the manifold of smaller di-
mension, determined by the eigenvalues farthest to the right in the complex
plane. Such a manifold is invariant, but not attracting. Calculations of this
sort suggest that the wavy vortices may be stable to quite high value of T,
thus delaying the appearance of a stange attractor or even of further bi-
furcations.
Appendix to Chapter 30-The Matrices in

Eagles' Formulation
The matrices that appear in equation (30.3-6) are as follows:
A=
0
v
- -Of] -vo z B
2V v
---Of] 0
r r r2
1
-Of]
vr
1
r
0 !v (V' + V)
r
- ~r' Of] 1
--B+-
v
1
r2
0
1 1
!o 0 0 0 --B
v z r v
1 1
0 0 0 - rOo -oz
r
0 1 0 0 0 0
0 0 1 0 0 0
where B is the operator
B = C2+ 2)
v r2 Of] Oz V
- --;:Of];
,,
, -1 0 0
(0) ,, 0 l/v 0
,,
M= ,, 0 0 l/v ,
1__ - _______________________
(0) (0)
Appendix to Chapter 30-The Matrices in Eagles' Formulation 275
u V
- - - 88 - w8 z -V + -8
u
8 u8 z
r r r r
K(U) =
(0) ~v (~+ 8V)
r 8r ~ (~88 + W8 z) 0
l8w
v 8r
0 ~ (~88 + W8 z)
(0) (0)
CHAPTER 31
The Early Onset of Turbulence
Periodic, quasi-periodic, almost periodic, and aperiodic motions; w-limitset;

attractors; power spectrum; Lyapounov stability; strange attractors; the
Lorenz attractor; strongly generic, generic, nongeneric, and strongly
nongeneric properties of systems.
Prerequisites: Chapter 29.
The word "turbulence," as used in practical fluid dynamics, refers to a flow

whose chaotic aspects are so highly developed that statistical methods can
be used for the study of at least many of its characteristics. It occurs at
substantially higher Reynolds numbers than those we shall consider. For
example, fully developed turbulence in air, in which the so-called inertial
subrange is fully developed, requires more violent conditions than can be
achieved in most wind tunnels, and is observed mainly in the free atmosphere.
Between laminar flow and turbulence there is an ill-defined regime called
the transition to or onset of turbulence. We shall be concerned only with the
part of that regime at quite low Reynolds numbers. The word "early" in
the title refers to a physical arrangement in which the Reynolds number is
increased very slowly with time. The flows we consider are simple and smooth,
but, nevertheless, exhibit certain features of unpredictability and chaos at
quite early stages.
31.1 The Landau-Hopf Model
A schematic model of the transition is described in Fluid Mechanics, Landau

and Lifshitz 1959. It is assumed that, at least in some problems, there is a
sequence of supercritical bifurcations forming schematically a tree, as in
Figure 31.1. After the first bifurcation the motion is generally periodic;
after the second it is generally quasi-periodic with two periods, and so on. A
quasi-periodic function with m periods is a function of the form
(31.1-1)
276
The Landau-Hopf Model 277
Figure 31.1 Bifurcation tree.
where g(., ., ... , -) is periodic in each of its arguments with period 2n, and
the frequencies Wi are incommensurable, which means there is no vanishing
linear combination elw l + ... + CmWm with rational coefficients CI' ... ' Cm
If WI> .. , Wm are commensurable, the number of independent frequencies is
less than m. Suppose, for example that m = 2 and WZ/w l = p/q, where p
and q are integers. Then, for
(31.1-2)
we find that WI to = 4nq and W z to = 4np; hence f(t) is periodic (not merely
quasi-periodic) with period given by (31.1-2).
It was shown in the last section of the preceding chapter that if the first
bifurcation leads to a closed orbit, the second can lead to an attracting
invariant torus in the phase space 5. If, furthermore, the motion is such
that its orbit covers the torus densely, then a resulting function of time, such
as one ofthe coordinates in the phase space, is quasi-periodic with two periods.
Specifically, one can define two intrinsic angle-coordinates e and cp on the
torus such that e = WIt + const., cp = Wz t + const., and the orbit is dense
on the torus if and only if WI and Wz are incommensurable. After the next
bifurcation there may be motion on a 3-torus, and so on.
278 The Early Onset of Turbulence
Exactly which branch of the tree in Figure 31.1 is followed depends on the
structure of the infinitesimal perturbation that caused departure from the
basic or laminar flow when the first critical value of the Reynolds number was
reached. More generally, the phases associated with the various frequencies
depend in a random manner on that perturbation, so that (31.1-1) might
better be written as
(31.1-3)
The idea behind the Landau-Hopf model was that as soon as there are
many independent frequencies, the motion is so irregular in appearance that
it must be regarded for practical purposes as chaotic.
There are various ways in which this model can be inappropriate:
1. One of the bifurcations in the tree of Figure 31.1 may be subcritical;
then, as soon as the corresponding critical value of the Reynolds number is
exceeded, there is no nearby stable motion for the system to follow, and there
is a so-called explosive transition to a motion involving more or less remote
parts of the phase space.
2. In some problems, such as flow in a circular pipe, the basic flow is
stable to infinitesimal disturbances at all Reynolds numbers, but is unstable
to finite disturbance of rather small amplitude, and the critical amplitude for
instability decreases toward zero as the Reynolds number increases, so
that stable flow cannot be achieved in practice at high Reynolds number,
owing to the presence of small but finite disturbances.
3. Although an invariant torus generally appears at the second bifurcation,
the orbit need not be dense on it; it may return to its starting point after
winding finitely many times around; then the orbit is closed and the motion is
periodic, as mentioned in Section 29.11. In fact it is now believed, on the basis
of Peixoto's theorem (see Appendix) that closed orbits on the torus are more
likely than dense ones. This may lead to the Feigenbaum model-see
Section 31.19.
4. A possibility discussed by Ruelle and Takens 1971 is that, after a few
bifurcations, there appears an invariant point set in the phase space, which is
not a torus but a so-called strange attractor; then, as explained below, the
motion is not quasi-periodic, but aperiodic.
31.2 The Hopf Example

In 1948 Eberhard Hopf gave an example of a simple dynamical system that
has an infinite sequence of bifurcations, each leading to an attracting torus
of one higher dimension than the preceding one. The functions u(x, t) and
z(x, t) are generally complex-valued functions of real variables x and t and
are periodic in x with period 2n. The equations are
ou -z 0 Z 01
o2 u
+ 't lox-2 '
ot - U
(31.2-1)
oZ 02 Z
ot = Z 0 U +z 0 F + II ox 2 '
The Ruelle~Takens Model 279
The small circle denotes convolution; generally,

1 2IT
fog = (f g)(x) = 2n Jo f(x
0
r + y)g(y)dy;
U 1 is just the average value of u; /1 is a positive constant, and F = F(x) is a
0
given complex-valued even periodic function.

The system may be regarded as a simple analogue of the Navier-Stokes
equations on a compact manifold (the circle), with the nonlinearities written
as convolutions rather than advection terms. F(x) is a forcing term, and /1,
the viscosity, is a parameter that can be varied.
Solutions can be found by expanding all functions in Fourier series with
respect to x, for example u(x, t) = L~ 00 un(t)e inx , with similar notations for
the other functions. Since the Fourier transform of a convolution is the
product of the Fourier transforms, the terms with given n are not coupled
to terms with a different n. In that respect the system fails to model the
Navier-Stokes equations, but its solutions can be found explicitly.
The only restrictions put on the forcing function F(x) are to prevent the
system's being too special, i.e., non generic. Namely, if we write
00
F(x) = L (an + ibn)einX, (31.2-2)

~oo
then it is assumed that infinitely many of the an are positive, that the bn are not
rationally related (any finite set of them is linearly independent over the
rationals), and that no two of the quantities an/n 2 are equal.
The critical values of /1 are the numbers an/n 2 ; they can be arranged in a
sequence /11 > /12 > /13 > ... ~ O. The general solution represents a moving
point in an infinite-dimensional space n with coordinates un(t), zn(t), n =
0, 1, 2, .... Hopf proved that, for /1 > /11' the fixed point at the origin
ofn attracts all other solutions, so that Un ~ 0, Zn ~ 0, as t ~ 00; as /1 decreases
past /11' that solution becomes unstable, and there is a bifurcation to an
attracting periodic orbit that grows out of the origin; when /1 decreases
past /12, that orbit also becomes unstable, and there is a bifurcation to an
attracting torus (2-torus) that grows out of the orbit, and so on. After the
kth bifurcation there is an attracting k-dimensional torus, and the orbits on
the torus are dense on it.
31.3 The Ruelle-Takens Model

In the model of the early onset of turbulence proposed by D. Ruelle and
F. Takens 1971, the first four bifurcations are assumed, as in the Landau-Hopf
model, to be super critical and to lead to invariant tori T\ k = 1,2, 3,4, each
of which is attracting between its appearance and the next bifurcation.
Concerning the existence of these tori, see the discussion of the Feigenbaum
model in Section 31.19. Ruelle and Takens prove that, on T 4 , motion on a
particular kind of strange attractor contained in T4 is rather likely. The
attractor is locally the Cartesian product of a two-dimensional Cantor set and

a two-dimensional surface.
Their theorem can be paraphrased as follows: Consider a Banach space ~,
each point of which represents a vector field on the torus T4, with a norm
containing the magnitudes of the components of the vector field and their
derivatives of order S 3. Two points in ~ differing by a small amount in
norm may be regarded as two physical systems, of which one can be obtained
from the other by a small perturbation of the vector field. Then, given any
constant vector field on T4 (one for which the angle variables on the torus
vary linearly with time), and given any 13 1 > 0, there is a perturbation of that
field of magnitude less than 13 1 which yields a strange attractor of the kind
they describe. Then there is another number 132 > 0 (possibly much smaller
than 13 1 ) such that the strange attract or persists under any further perturba-
tion of magnitude < 13 2 , Hence, the vector fields that yield the strange attractor
cannot be dismissed as unlikely.
Their particular choice of strange attract or is somewhat arbitrary; one
can imagine many variations of it, each having the property stated.
The strange attractor discovered earlier by E. N. Lorenz (see Sections
31.9-31.17 below) arises by a somewhat different mechanism. Apparently,
no one has found a specific vector field on a specific manifold that leads to a
strange attractor precisely according to the Ruelle-Takens model. The
important idea in their paper is that motions on strange attractors are in
some sense likely, or at least not unlikely and are possibly even generic in
certain circumstances. Their theorem does not say that the existence of a
strange attractor is a generic property of vector fields on T4 (see the Appendix
to this chapter). The corresponding subset of the Banach space ~ is open but
not necessarily dense; it merely comes arbitrarily close to every point of ~
that represents a constant vector field on T4. It says simply that motion on a
strange attractor is more likely, once an invariant T4 has been established,
than the quasi-periodic motion on T4.
31.4 The w-Limit Set of a Motion

Following the considerations of Section 29.7, we assume that we are dealing
with a dynamical system
x = F(x) (31.4-1)
on a finite-dimensional manifold m, when~ F() is a smooth vector field
on m. A solution x(t) of this equation is called a motion in m, and the point
set in m given by
y = {x(t): all t} (31.4-2)
is called the orbit or trajectory) of the motion. We assume that the initial-value
problem of (31.4-1) is well posed, and we denote by q>(x o , t) the solution that
starts at xo, for any Xo in m and all t :2: 0; that is, if x(t) is any solution, then
x(t) = q>(x(O), t). (31.4-3)
The w-Limit Set of a Motion 281
For fixed x, cp(x, t) is a motion, while for fixed t 2!: 0, a one-to-one mapping
of 9Jl into itself is given by x ~ <p(x, t). The function <p(x, t) is called a semi-
flow; it obviously has the semigroup property that if t and s are 2!: 0, then
<p(<p(x, t), s) = <p(x, t + s). (31.4-4)
For t = 0, <p is the identity mapping: <p(x,O) == x. We assume that <p is
continuous in x and t.
A point ~ of 9Jl is called an OJ-limit point of a motion x(t) if x(t) comes
arbitrarily close to ~ at times arbitrarily far in the future, that is, if there is a
sequence {t n }1' such that
tn ~
Ix{tn) _ ~I ~
oo}
as n ~ 00. (31.4-5)
The set of all OJ-limit points of a motion is called its OJ-limit set and is denoted
by Qx, where x is the initial point of the motion; it is a closed point set. If y
is any other point of the same orbit, then Q y = Q".
As an example, if a motion tends to a fixed point, as t ~ 00, then that point
is the OJ-limit set of the motion. If a motion in a plane spirals outward toward
a closed curve, then that curve is the OJ-limit set of the motion. If the orbit ofa
motion on a torus is dense on the torus, then the entire torus is the OJ-limit
set of every point of that orbit.
The symbol OJ refers to future time. When a motion exists for all t, a-limit
points and a-limit sets are similarly defined by letting t ~ - 00.
If a motion x(t) lies in a bounded region of 9Jl, for t 2!: 0, then its OJ-limit
set Q is non empty, and x(t) gets closer to Q as times goes on; that is,
distance {x(t), Q} ~ 0, as t ~ 00. (31.4-6)
(See Nemytskii and Stepanov 1960, Chapter V, Section 3.) However, Q
need not be stable or attracting; other nearby motions may move away
from it and never return.
An important property of OJ-limit sets has to do with the reversibility of the
motion. A solution of(31.4-1) cannot generally be continued to all negative t.
For example, the solutions of x = - x 3 on IR cannot be [except the solution
x(t) == 0], and solutions of the Navier-Stokes equations generally cannot
be, because of the parabolic nature of the equations. However, certain
special motions can be.
From the continuity and semigroup property of <p, it follows that if the
initial point x(O) is an OJ-limit point of x(t), then every subsequent point
x(t o) is also an OJ-limit point of x(t). In other words, if a motion starts in its
own OJ-limit set, it stays there. It can be proved that such motions can also
be continued indefinitely backward in time, and hence lie in il,.(O) for all t.
See Sell 1971, Theorem 11.8. This last result may seem paradoxical from the
point of view of physical observation or numerical simulation, where there
is only a finite accuracy. According to (31.4-6), any bounded motion x(t)
lies for practical purposes in its own OJ-limit set after a finite lapse of time.
That doesn't imply, of course, that finite-difference methods can be used to
follow such an orbit indefinitely backward without escaping from QX(O);

it means merely that if a motion has been somehow followed forward for a
sufficiently long time interval, after transients have disappeared, the result
may be regarded as a rather accurate sampling of the kind of motion which
ideally is confined to QX(O) for all t.
31.5 Attractors
An attractor is roughly a set in 9Jl such that any sufficiently nearby motion
gets closer and closer to it as time goes on. Specifically, we shall call a con-
nected closed bounded set S in 9Jl an attractor if
1. S is contained in an open set :Yl o such that for any x in !!lo the motion
<p(x, t) is in :Yl o for all t > 0;
2. If:Yl is any open set containing S (see Figure 31.2), then for any x
in :Yl o there is a time T such that <p(x, t) is in !!l for all t > T.
3. For the given region !!lo , S is the smallest set having the above properties
in the sense that if :Yl(t) is the image of :Yl o under the flow <p, then, as
t -> 00, !!let) shrinks down onto S but no further:
S = n:Yl(t).
t20
For a given attract or S, the largest open set:Yl o having the properties stated
is called the region of attraction of S. We require S to be connected, for if it
consisted of two disconnected pieces S 1 and S 2, then each of these could be
enclosed in a suitable region !!l1 or :Yl2 and be an attractor in its own right.
Smale 1967 imposes the further requirement that there be an orbit dense
in S. An example in which that requirement is not met, while the others are,
is provided by the Taylor problem of flow between rotating cylinders. After
the first bifurcation, there is a closed curve (in fact a circle) in phase space
(Hilbert space) consisting of fixed points, and any nearby motion comes
x
f?to

Figure 31.2
The Power Spectrum for Motions in [p;n 283
asymptotically to rest at one of those fixed points. However, this example is

nontypical in the sense that an arbitrarily small perturbation (for example,
subjecting the entire apparatus to a small velocity in the axial or z direction)
can make points march around the circle, whereupon the circle becomes an
orbit.
The w-limit set of a motion is an attractor if it also attracts all other
nearby motions. More generally, an attractor S is the union of many w-limit
sets, namely the w-limit sets of all motions that start in fJlt. In particular, S
may contain fixed points and closed orbits, which are of course their own
w-limit sets.
Attracting fixed points, attracting closed orbits, and attracting invariant
tori of various dimensions are examples of attractors. Other attractors often
have a more complicated geometric structure, for example involving Cantor
sets. A strange attractor is one on which the motions are unstable in the sense
of Lyapunov and hence are characterized by a continuous power spectrum,
as explained in the rest of this chapter.
31.6 The Power Spectrum for Motions in [Rn

The theory of the power spectrum was discussed in Section 4.6 of Volume I.
It concerns a bounded continuous function f(t), which oscillates more or
less irregularly for all t from - 00 to + 00. Such a function cannot be expres-
sed as a classical Fourier series, because it is not periodic, nor as a classical
Fourier integral, because it is not in L 2 ; but it is a tempered distribution, and
hence has a Fourier transform in the sense of distribution theory. The power
spectrum is a function Sew), which describes the distribution of the power
associated withf(t) over the frequencies ofthe Fourier components, ignoring
their phases. Namely, Sew) is a nondecreasing real function, and S(w 2 ) -
Sew!) is the power contained in the Fourier components with frequencies in
the interval (w!, w 2 ).
We summarize the results here, as applied to a vector-valued function
x(t) in [Rn, in place of f(t).
The autocovariance of x(t) is given by
R(T) = lim _1 fT x(t + T) . x(t)dt, (31.6-1)

T-->oo 2T -T
and then the power spectrum is given by
Sew) = foo
-00
R(T)
eiw! -
2.
1l:lT
1
dT. (31.6-2)
Comments
1. It has been tacitly assumed that all components of the vector x(t)
contribute equally to the energy of the motion. A generalization is to replace
the scalar product by
x(t + T)' Bx(t), (31.6-3)
where B is a positive definite matrix, to be determined by physical considera-

tions so that (31.6-3) represents the energy.
2. If the spectrum is "continuous," i.e., if Sew) is absolutely continuous,
which happens if R(r) tends to zero sufficiently rapidly as r ~ + 00, then the
spectral density is given by an ordinary Fourier transform
S'ew) = -1 fco R(r)e iw! dr. (31.6-4)

2n -co
In numerical calculations, R( . ) is given only at discrete values of rand S()
or S'() only at discrete values of w. Then, S'() is usually obtained from R()
by the fast Fourier transform algorithm, and if the spectrum contains lines,
they show up not as c)-function contributions to S'() but as single values of
S'() that are much larger than the surrounding values.
3. If x(t) comes from experiment or numerical simulation, we do not
take the limit in (31.6-1), but write instead
R(r) ~ b _ a
1 fb x(t + r) x(t)dt,
a (31.6-5)
where (a, b) is a long interval. As noted at the end of Section 31.4, although
x(t) is regarded as an approximation to a motion on its own w-limit set,
hence defined for all t, it is known in practice only for t ;:::: 0; hence in the
above expression, we have 0 < a < b. Still, that expression should be a good
approximation if all transients have decayed sufficiently by the time t = a.
For further details and examples, see Section 4.6 of Volume 1.
31.7 Almost Periodic and Aperiodic Motions
The quasi-periodic motions predicted by the Landau-Hopf model are a

special case of almost periodic motions, and hence have a pure line spectrum.
If the m-tuple periodic function g in (31.1-1) is expanded in an m-tuple
Fourier series,
co
I
then, after suitable renumbering of the terms, the quasi-periodic function
f(t) given by (31.1-1) can be written as
co
f(t) = I cjeiiiJit, (31.7-2)
j=1
where each Wj is a linear combination of WI' .. , Wm with integer coefficients.

If there were a value of r > 0 such that the quantities Wk r (k = 1, ... , m)
were all simultaneously integer multiples of 2n, then f(t + r) would be
-=f(t), and fO would be periodic. Since the wk's are incommensurable,
there is no such r. However, it can be proved that r can be so chosen that the
Lyapunov Stability 285
difference f(t + T) - f(t) is arbitrarily small. Not only that, but given
e; > 0, there is a T = T(e;) > 0 such that every t-interval oflength Tcontains
at least one value T such that
If(t + T) - f(t) I < e; for all t.
Any continuous function having this last property is called almost periodic
in the sense of H. Bohr and can be expanded in a series of the form (31.7-2),
which then converges in a certain L2 sense-see Riesz and Sz. Nagy 1953,
Chapter VI. A quasi-periodic function, as we have defined it, is an almost
periodic function in which only a finite number of the frequencies Wj are
linearly independent over the rationals.
A vector-valued almost periodic function x(t) is similarly defined.
As shown earlier (Section 4.6 of Volume I), the power spectrum of an
almost-periodic function is a pure line spectrum. Hence, in the Landau-Hopf
model of the transition to turbulence, the power spectrum is a pure line
spectrum after any finite number of bifurcations, although the number of
lines in a given frequency interval might increase considerably as the Reynolds
number increases.
The Ruelle-Takens model, as we shall see, predicts that a continuous
spectrum will appear after a rather small number of bifurcations.
The almost periodic character of the motion in the Landau-Hopf model
seems implausible on intuitive grounds. If the motion is in any sense random,
it should not be able to "remember" its past behavior so precisely as to
reproduce that behavior, to any desired accuracy, at times arbitrarily far in
the future, as would be the case if it were almost periodic. This can also be
expressed in terms of the autocovariance R(T), because R(T)jR(O) is the
autocorrelation, that is, the correlation of the functions f(t) and f(t + t). A
theor~m on almost periodic functions says that the convolution of two such
functions, in the sense of (31.6-1), is also almost periodic (see Riesz-Nagy).
Hence R(T) would be almost periodic if f(t) were, and the correlation
coefficient of f(t) and f(t + T) would come arbitrarily close to 1.0 repeatedly
for certain arj>itrarily large values of T.
By con ast, for typical motion on a strange attractor, R(T) decreases
rapidly zero as t - 00, and then the power spectrum is purely continuous
and is iven by(31.6-4). For the case of the Lorentzattractor, see Section 31.11.
31.8 Lyapunov Stability
A motion x(t) [a solution of (31.4-1)] is stable in the sense of Lyapunov if

any other motion x(t) [another solution of (31.4-1)] that is sufficiently close
to it initially stays close at later times, specifically if for any e; > 0 there is a
15 = bee;) > 0 such that
if Ix(O) - x(O) I < 15,
then Ix(t) - x(t) I < e; for all t > O.
If the motion is not stable in this sense, there is a positive e (not necessarily
very small) such that no matter how small c5 is, there is a perturbation of
initial size < c5 that grows to a size ~ e at some later time. One of the motiva-
tions for Lorenz's work, described below, was to show that a simple proto-
type of atmospheric motion is Lyapunov unstable, with obvious implication
for the problem of weather forecasting.
31.9 The Lorenz System; the Bifurcations
The first strange attract or in a problem arising from fluid dynamics was
discovered by E. N. Lorenz in 1963. Lorenz expanded the Benard equations
of thermal convection for a horizontal layer of fluid heated from below in a
triple Fourier series with respect to the space variables, then truncated the
resulting system of ordinary differential equations for the time dependence
of the Fourier coefficients to three equations. If the Fourier coefficients in
those equations are denoted by X(t), yet), and Z(t), the equations are
X = -(JX + (JY,
y= rX - Y - XZ, (31.9-1)
Z = -bZ + XY,
or, more briefly,
x= F(X). (31.9-2)
The constants (J, r, and b are dimensionless; for the physical system
consideied by Lorenz, they have the values
(J = 10, O<r<oo; (31.9-3)
r is proportional to the Rayleigh number and is a measure of the intensity of
the heating.
Lorenz was interested in exhibiting the general kind of instability found
in atmospheric physics and did not intend the above system to be a realistic
model of the atmosphere or of thermal convection. Later Curry 1978 studied
a more realistic model of the Benard convection equations by truncating to
14 rather than three equations. He found a more complicated sequence of
bifurcations, as the Rayleigh number r is increased, but the strange attractor
was still present for certain values of r.
Lorenz showed there is a constant R, depending on (J, r, and b, such that
any solution X(t) of (31.9-1) is eventually trapped in the ball
(31.9-4)
Furthermore, it follows from (31.9-1) that the divergence of the vector field
F(X) has the constant value
V F = -(J + b + 1) = -131, (31.9-5)
The Lorenz System; the Bifurcations 287
so that the volume of a region carried along by the flow (31.9-1) in [f;3
decreases with time as exp( -13.67t). Therefore there is at least one attractor
in the ball (31.9-4), and any such attractor occupies zero volume in [f;3.
The fixed points or stationary solutions of the system (31.9-1) are as follows:
1. For any r, the origin, X = Y = Z = 0 is a fixed point. For 0 < r < 1
it is stable (in fact attracting). For r > 1 it is unstable; the linearized problem
has one positive eigenvalue and two negative ones. There is a one-dimensional
unstable manifold with a horizontal tangent vector at the origin (a vector
parallel to the plane Z = 0) and a two-dimensional stable manifold with a
vertical tangent plane.
2. For any r > 1 there are two more fixed points, called P I and P 2:
P2 : X = Y = J b(r - 1), Z = r - 1,
(31.9-6)
PI: X = Y = - Jb(r - 1), Z=r-1.
Hence, there is a first bifurcation, of the type discussed in Section 29.9, at r = 1.
To determine the stability of the new fixed points, we write X = Xo + Xl>
where Xo is given by (31.9-6), and we linearize with respect to Xl; we find
(31.9-7)
When X o , Yo, and Zo are substituted from (31.9-6), the matrix of this system
becomes
a
-1 -Jb(~ - 1));
Jb(r - 1) -b
It has one negative real eigenvalue and two complex conjugate ones. The
complex eigenvalues are in the left half-plane (hence the new fixed points
are stable) ifr < ro, where
ro = a(a + b + 3)/(a - b - 1) = 24.74. (31.9-8)
Hence there is a second bifurcation at r = ro, and this one is of the kind
discussed in Section 29.10, leading to periodic solutions. However, this
bifurcation is subcritical, as shown by the calculations of Marsden and
McCracken 1976. Hence the periodic solutions 'are present only for r < ro
and are unstable, while for r > ro an explosive transition to something else
must be expected. It turns out, as discussed below, that the transition is not
really "explosive," because of the presence of another attractor (in fact a
strange attractor) in the near vicinity in [f;3; see Section 31.17.
With each of the points PI and P 2 is associated, for r > ro, a one-dimen-
sional stable manifold and a two-dimensional unstable one. In the latter,
solutions spiral outward from the fixed point. Nearby solutions also spiral
outward and, at the same time, are drawn rapidly toward the unstable
manifold, because of the large negative eigenvalue associated with the
stable one.
31.10 The Lorenz Attractor; General Description
To investigate the behavior of the system after the second bifurcation, at

r = ro = 24.74, Lorenz calculated solutions of the system (31.9-1) numeri-
cally, for r = 28. He found that, after transients have decayed, the orbits
appear to lie, as accurately as one can tell from the calculations, on a branched
surface L o , shown schematically in Figure 31.3a, where the directions of
motion are also indicated. It is roughly heart-shaped and lies roughly in the
vertical plane X = Y [which contains the fixed points Pi and P 2 given by
(31.9-6)]. It is symmetric with respect to reflection about the Z axis and has
two holes, one surrounding each of the fixed points Pi and P 2' Below a
branch line BB, which joins the two holes and dips down somewhat below
the horizontal, there is a single sheet. Above it, there are two sheets, one a
little to the left and behind, the other a little to the right and in front. They
are joined along the heavy part of the branch line BE.
Lo is bounded by part of the unstable manifold of the origin, which con-
sists of two orbits Wr and W~ starting out from the origin in opposite direc-
tions. These orbits later continue into the interior of the surface Lo after going
round the holes shown and then crossing the branch line BB at the ends
of the heavy part. The fixed points Pi and P 2 lie inside the holes, and if the
surface Lo were continued into the holes, orbits would spiral outward in
them; it is seen from the figure, however, that an orbit once in Lo can never
subsequently get into the holes.
Every orbit crosses the branch line BB downward; if it crosses at the
center of BB, it comes asymptotically to rest at the origin 0; otherwise, it
encircles one of the holes and goes down across BB again, and so on. We
define a Poincare map s --+ l/Jo(s) of BB as follows: Let s be a parameter
(possibly arclength) along BB, with s = 0 at the center. If an orbit crosses
BB at s = s 1 =I 0, then it next crosses BB at s = l/J o(S 1); l/J 0(0) is undefined.
o
Figure 31.3a The branched surface.
The Lorenz Attractor; General Description 289
I/Io(S)
Figure 31.3b The Poincare mapping.
See Figure 31.3b. Numerical orbit calculations show that this mapping has
the property called locally eventually onto by Williams 1977: If I is any
interval So < S < So + 8, no matter how narrow, then for some n, the n-times
iterated mapping of I is all of BE. In other words, no matter how close to-
gether two orbits on Lo are initially, they eventually become completely
separated as time goes on. This follows, for example, from showing that if the
parameter s is suitably chosen, then l/J'o(s) ~ const. > 1 for all s.
An orbit that goes down across the branch line BB to left of center encircles
p 1 clockwise before returning to BB, and one that goes down to the right of
center encircles P 2 counterclockwise. The number of times an orbit encircles
one of the points P 1 or P 2 before moving to the other depends on how rapidly
it spirals outward from that point and also in a critical way or how far from
the center it first cut BB after coming from the other side. It is the essence of
Lorenz's discovery that the successive numbers of circuits round those points
vary in a pseudorandom way so that the motion is aperiodic.
As Lorenz pointed out, the picture based on the branched surface Lo
cannot be precise, because two orbits that go down across the branch line BB
at the same point of BB would then coincide subsequently, and that would
contradict the unique reversibility of the orbits. Hence, the two sheets of the
surface can't merge into a single sheet, but must remain separated by a
possibly very small distance. If we follow the orbits round again, the two
sheets become four, and so on. We conclude that the attractor must contain
infinitely many sheets, probably shomehow connected together into a single
structure in ~3, which will be called the Lorenz attractor and denoted by L.
However, at least for the parameter values studied by Lorenz (0" = 10,
b = t r = 28), the fine structure just described is really quite fine; the
separation of the sheets is very small, so that, to something like four-decimal
accuracy, the branched surface L o , with the motions on it as sketched,
describes the attract or fully.
31.11 The Lorenz Attractor; Aperiodic Motions
Lorenz observed that the randomness of the motion could be analyzed by

considering the successive maxima Zn, n = 0, 1,2, ... , of Z(t) on the com-
puted orbit. [From the third equation of (31.9-1), we see that these maxima
occur at intersections of the orbit with the hyperboloid given by Z =
(ljb)XY.] Lorenz found that if Zn+ 1 is plotted against Zn, the points lie
quite accurately on a curve Zn+ 1 = f(Zn), which is smooth except for a
central cusp, as shown in Figure 31.4.
As Lorenz pointed out, Zn+ 1 cannot be exactly a single-valued function
of Zn, since its value depends generally also on the values of X(t) and yet)
at the instant when Z(t) = Zn- Hence the Lorenz graph ought to be slightly
fuzzy ::lnd ought to have some transverse structure, although possibly of a
very narrow extent. This structure is related to the fine structure of the
attractor L, and one can begin to see it if the deviations of the points on the
graph from a smooth curve are magnified by about 10 3 (see Richtmyer 1981).
If we ignore the fine structure of the Lorenz graph, the sequence {Zn}O' of
maxima of Z(t) may be regarded as obtained from an initial value Zo by
iteration of the mapping f: Z -+ feZ), so that Zn+ 1 = f(Zn).
In order to study the iterated mapping statistically, we consider the
pro blem of transforming from Z to another variable W by equations
Z = ((W), W= C 1 (Z), (31.11-1)
i:
t ,I
j:
zn+1
so ..,"
..
"./ Zn~
aa ID
e
Figure 31.4 The Lorenz graph.
The Lorenz Attractor; Aperiodic Motions 291
such that the resulting mapping g: W -+ g(W) takes an especially simple

form. In particular, we wish to choose (31.11-1) so that g(W) is the triangular
function
{
2W, if 0 ::; w::; 1,
g(W) = 2 - 2W, if 1::; w::; 1. (31.11-2)
Lorenz considered the mapping g based on this function as a sort of model for
the mapping f that arose in his calculations, in order to predict qualitatively
the statistical properties of the motion. We show how to compute the trans-
formation (31.11-1) under certain assumptions about the function feZ).
First, we assume that a linear change of the coordinate Z has been made so
that the least and greatest possible values of Zn are 0 and 1. Then the function
feZ) maps the interval [0, 1] onto itself. We assume that f(O) = f(l) = O.
[Actually,f(O) is ~0.0035, but we shall ignore this difference along with the
fine structure of the Lorenz graph.] We also assume that feZ) is differentiable,
except at the cusp, and that I1'(Z) I is greater than some constant r:x > 1 for
all Z. Then, it can be shown-see Riissmann and Zehnder 1980 or Richtmyer
1981-there is a unique continuous increasing function ((W) that transforms
the mapping Z -+ feZ) into the mapping W -+ g(W).
To calculate ((W), we denote by CPl(Z) and CP2(Z) the inverses of the rising
and falling parts of feZ), as shown in Figure 31.5. Then, it is seen from
(31.11-2) that ((W) must satisfy the equations
((W) = CPl(((2W for 0 ::; w::; 1,
((W) = CP2(((2 - 2W for 1 ::; W ::; 1.
From these equations, ((W) is calculated in succession for dyadic values of
W in the order W = 0, 1, t, i, i, i, ... , starting with ((0) = 0 and ((1) = 1,
and then for other values by the requirement of continuity. Figure 31.6a
shows the result, and Figure 31.6b shows the result of applying the trans-
formation ((W) to the Lorenz graph.
CPl,2
Z Figure 31.5
.. z;i
(w)
10
III >
0 .1 .z . . .1 .6 .'7 .1 . 1.0
Figure 31.6a
LCR:NZ GRIIPH IN W
1.0
I \
.I
r
\
...,
l
, ... \
I ;.
"'n+1 / ...
.,
/ ,
,,.I
\
\
,I \,
'...
/
.I \
\
\,
I ,
I
/ \
\.
.1 \
""n---;)oo \
.1 .z .1 .6 .'7 .1
1.0
Figure 31.6b
Statistics of the Mappings f and g 293
The function (W) is Holder continuous with Holder exponent log2 IX,
where rx is the greatest lower bound of IF(Z) I, as above, but it is not absolutely
continuous; hence what is true of the mapping g: W ~g(W) for almost all
W is not necessarily true of the mapping f: Z ~ feZ) for almost all Z.
A consequence of the continuity of (W) (which does not require absolute
continuity) is that the mappingfis locally eventually onto, in the terminology
of Williams, referred to in Section 31.10; if I is any interval a < Z < a + 1::,
then for some finite number n of iterations, pnl(J) is the entire interval [0, 1].
Clearly f has that property if and only if g also has it, but for g it is nearly
obvious. Namely, the length of any interval I in W is doubled under g unless
g(I) contains the point W = 1, in which case the length is at least not de-
creased. Hence, under each pair of iterations, the length is at least doubled,
unless, for some I, g(I) and g(g(I)) both contain the point W = 1, but in that
case g(g(I)) contains all of [1, 1], so that g(g(g(I))) is [0, 1].
It follows that iteration of the mapping f is unstable in the sense of
Lyapounov, for no matter how close together two points are initially, they
will eventually be separated by a finite amount (for example at least 1).
Hence, if we can neglect the fine structure of the Lorenz graph, the motion on
the Lorenz attractor is also Lyapounov unstable, and hence has a purely
continuous power spectrum.
31.12 Statistics of the Mappings f and g
The two mappings have slightly different statistical properties; we discuss g

first. Because of the special character of the function g(W), as given by

(31.11-2), we represent W in binary form as W = .aOa 1 a2 ... , where each
ai is or 1. Then g(W) is simply
if ao = 0,
if ao = 1.
where the overbar denotes complementation, i.e., replacement of by 1

and 1 by 0. Now suppose the initial element Wo of the sequence {w,,},';"'=o,

obtained by iterating g, is chosen at random from a uniform distribution in
[0, 1]. Then each binary digit of Wo is equally likely to be or 1, independently
of the choice of the others, and furthermore w" is < 1 or :;::: 1 according to the
choice of the nth digit of Wo. We conclude that each w" is equally likely to be
< 1 or :;:::1, and there is no correlation in this matter between successive
members of the sequence.
To find the corresponding properties of f, a sequence {Zn} of length
200,000 was generated by iterating f numerically (Richtmyer 1981). To
analyze the result, Sn was defined as
- 1, if Zn < 1,
Sn = { .
+ 1, 1
If Zn :;::: 2'
The average of Sn was found to be -0.1416 (about 64 standard deviations)

and significant positive correlations were found of Sn with Sn+1> with Sn+2,
and with Sn + 3' The correlation of Sn with Sn + k' for k 2 4, was not significant
in the sample size available.
The difference between j and g reflects the nonabsolute continuity of the
functions (W) that connects Z and W. Clearly, W is <! if and only if Z =
(W) is <! (because the Lorenz graph, Figure 31.4, is symmetric with high
accuracy), but the behavior for almost all W is not the same as the behavior
for almost all Z. There is a set of measure zero in [0, 1J that is transformed
into a set of positive measure, in fact measure = 1, under" and there are
similar sets for C 1. Nevertheless, the two mappings behave similarly;
hence the qualitative statistical properties of the Lorenz system can be inferred
by study of j, as was done by Lorenz.
31.13 The Lorenz Attractor; Detailed Structure I
According to Section 31.1 0, the attractor L contains infinitely many sheets (in
fact, uncountably many), all lying close to the idealized branched surface L o ,
and somehow connected together to form a many-sheeted structure in [R3.
The structure of L was investigated by R. F. Williams 1977, and we shall
describe his main results in somewhat intuitive geometrical terms, ignoring
certain topological difficulties such as arise from the presence of infinitely
many vertices of the cell complex F in a bounded region of the strip S (see
below).
The attractor is of course completely determined by the differential
equations (31.9-1). At present, however, there is no known method for
determining it precisely, starting from those equations, or even for determin-
ing its general topological properties. Furthermore, the equations are
somewhat specialized and artificial; hence there is more interest in the general
kinds of attractor that can result from equations generally similar to (31.9-1).
The objective of the work of Williams was to find all attractors that are
related to a branched manifold Lo of the kind described above in the way
the attractor L appears to be. He calls them Lorenz attractors generally.
Williams used an abstract topological construction known as the inverse
limit. The topological properties of the resulting attractor and the flow on it
are completely determined by the given semiflow on L o , and in fact by the
so-called kneading sequences of the orbits Wr and W~. Williams showed
that there are uncountably many different such attractors, i.e., topologically
different ones.
The attractors found by Williams can be embedded in [R3, but the question
which of them would result from a given differential equation system of the
type (31.9-1) remains open.
To investigate the structure, following Williams, we consider a piece of the
surface and then continue it by following the orbits in it both forward and
backward in time. If an orbit encircles Pi it moves to a sheet of L behind the
The Lorenz Attractor; Detailed Structure I 295
one it was on before (from the point of view of an observer looking at Figure
31.3), whereas, if it encircles P 2, it moves to a sheet in front. If it encircles first
one, then the other, we don't know whether it is then in front of the starting
point or behind; it might even return to its starting point, and then we
should have a periodic orbit. One of Williams's results is that the periodic
orbits are dense in L.
The unstable manifold WU(O) plays a role. As stated in Section 31.10, it
consists of two orbits that start horizontally in two opposed directions from
0, form the boundary of L o , and then continue into the interior. Conceivably
either or both of them might eventually hit the exact center of BB and then
come to rest asymptotically at o. We shall ignore that possibility as being
unlikely and assume that they continue winding indefinitely round in Lo.
Following Williams, we denote by Wl'(O) the orbit that leaves 0 to the right,
hence first meets BB at its left end and by W~(O) the orbit that leaves 0 to the
left, hence first meets BB at its right end. We define so-called kneading se-
quences for these orbits as ZlZ2Z3 ... , where Zk is = lor =2 according as the
kth circuit of the orbit is around PI or P 2. Below, it will be assumed that these
sequences start as 2111 ... and 1222 ... , respectively, for Wi(O) and W~(O).
It will be seen that they completely determine the topology of the attractor L.
In order to carry out the program, we have to make the following as-
sumptions. [That is, we have to conjecture that these things are consequences
of the differential equations (31.9-1)-numerical calculations give consider-
able support for most of them.]
1. The branched surface Lo and the semiflow on it are as described in
Section 31.10. [We say "semiflow," not "flow," because an orbit x(t) in L o ,
in contrast with those in L, cannot be uniquely determined, for t < 0, from
x(O), owing to the branching.]
2. The Poincare map l/Jo of BB is locally eventually onto.
3. There is a continuous mapping p (projection) of L onto Lo which
carries orbits of L onto orbits of Lo; if x(t) is a motion in L, then p(x(t)) is a
motion in Lo.
4. Every motion in Lo is the image under p of a unique motion in L.
(It would seem to be difficult to get supporting evidence for this assumption
from the numerical work.) Furthermore, the motion in L depends con-
tinuously on the motion in Lo; if orbits xo(t) and xo(t) are close together in
Lo for a long interval ( - T, T), then they are the projections of orbits x(t)
and x'(t) that are close together in space (i.e., in L) for a long interval.
5. The kneading sequences of Wi(O) and W~(O) start as 2111. .. and
1222... , respectively. That is, Wi(O), after initially encircling P 2, then en-
circles Plat least three times (in Lorenz's system it does so 25 times) before
again encircling P 2, and similarly for W~(O).
To follow the orbits properly in three dimensions, we replace the branch

line BB by a strip S transverse to Lo and intersecting Lo on BE. It turns out
to be convenient to pull Sand BB both downward until they pass through the
origin 0, as shown in Figure 31.7. We shall develop a detailed description
of the intersection S (] L ~ F of the strip S and the attract or L. F will be
Figure 31.7 The strip S.
described as a so-called cell complex, consisting of points called vertices and

curves called i-cells, each of which connects one vertex to another. We shall
find that the vertices are the origin 0 and the intersections of WU(O) with S,
and the i-cells are the intersections of the sheets that make up L with S.
F consists of the origin 0 and two parts, one lying on either side of a line
crossing the strip S at O. That line lies in the stable manifold of 0 and serves to
separate the orbits that will next encircle P 1 from those that will next
encircle P 2 .
It will be seen that WU(O) is throughout its length the boundary of infinitely
many sheets of L, which come together at that boundary so as to form what
Williams calls the "spine" of a "Cantor book" -see Figure 31.9. A line
transverse to Lo just inside the boundary intersects those sheets on a Cantor
set, as explained below.
We denote by Ib 12 , the successive intersections of W~(O) with S (they
can lie on either side of 0) and by r b r 2, ... the intersections of W~(O) with S;
they are points in S. When projected onto L o , 11 is at the extreme left of the
branch line BB, and r 1 at the extreme right; the others lie between.
Each sheet of L intersects the strip S on a curve C(j, a i-cell of F, and if we
let C(j be carried round by the flow until its points again meet S in another
curve C(j', it sweeps out a surface element L in IR\ as shown in Figure 31.8,
and we shall construct L be piecing together such surface elements; C(j and
'to' are respectively the initial and final curves of L. We assume tentatively
that there is a curve connecting 11 to r 1 in S and consisting of points of the
attract or L. (This assumption is justified in Section 31.17 below.) That
curve goes through 0, hence consists of two i-cells of F, and we have to
consider the two separately, so let ce be the i-cell that connects 0 to r l ' As C(j
is carried round by the flow to sweep out a surface L, we let its left end,
which starts at 0, be carried to the right from 0 along Wi'(O), so that it, like
all other points of C(j, encircles the point P 2 once counterclockwise; it then
The Symbols [i,)1 of Williams 297
c C'
Figure 31.8 A surface element in the Lorenz attractor.
meets S at the point 11; hence the final curve C(l' of the surface element I:
connects 11 to f2 in S, a little in front of C(l (toward the viewer).
Before we can carry the points of C(l' round again, we must note that since
11 and f2 are on opposite sides of 0 by assumptions 5 above, we must divide
"6" into two I-cells, say C(l'1 and '6'~, one connecting 11 to 0 and the second
connecting 0 to f 2. Then, '6"1 will be carried clockwise round P 1 and C(l~
counterclockwise round P 2, thus creating two new surface elements within
L, say I:'1 and I:~. This process can be continued indefinitely.
Generally, given two points Ii and fj of the intersection of Wr(O) and
W~(O), respectively, with S, then, if they lie on the same side of 0, there may
(or may not) be a I-cell of F connecting Ii to fj (i.e., a curve in S connecting
Ii to fj and consisting of points of L-ifthere is one, there are infinitely many).
If that curve is carried along by the flow, as described above, it sweeps out
a surface element I: of L.
The origin 0 is also denoted by 10 and fo. We denote it by fo if it is then to
be carried along W~(O) to r 1 and by 10 if it is then to be carried along Wr(O)
to 11. The general surface element I: of L is then obtained by carrying a I-cell
connecting Ii to rj in S (where now we permit i = 0 or j = 0) round P 1
or round P 2 to a curve connecting Ii + 1 to f j + 1. (Then, ifli+ 1 and f j + 1 lie on
opposite sides of 0, it is necessary to divide the new curve into two I-cells,
as before.)
31.14 The Symbols [i,J1 of Williams
For given i andj, ifli and fj are connected by a I-cell of F (hence, by infinitely
many, as we shall see), we denote that fact by saying that a symbol [i, j] is
defined or exists. We now state the rules for the existence of these symbols,
following Williams. Our tentative assumption implied that [0, 1] and

[1,0] are symbols (and we concluded, assuming that r2lies to the right of 0,
that [0, 2] is also a symbol). The rules are:
1. [0, 1] and [1,0] are symbols.
2. If [i, j] is a symbol, and
(a) Ii + 1 and r j + 1 lie respectively to the left and right of o. then
[i + 1,0] and [O,j + 1] are symbols;
(b) otherwise, [i + 1,j + 1] is a symbol.
In case 2(a) we say that [i,j] precedes (or is followed by) [i + 1,0] and
[0, j + 1], and in case 2(b) that it precedes [i + 1, j + 1] ("Immediately
precedes" might be a better wording.)
(Note. In contrast with Williams, we are assuming, as said earlier, that

neither Ii or rj ever falls exactly on 0 for any i > 0 or j > 0.)
The rules are so chosen that [i,j] precedes [i',j'] if and only if there is a
surface element L whose initial curve connects Ii to rj and whose final curve
connects Ii' to r j,. [This describes only a part of the final curve in case 2(a).]
The following are immediate consequences of the rules (some of them
require an induction in the proof):
1. If [i, j] is a symbol, then i =1= j.
2. Each symbol has either one or two successors, depending on whether
case (a) or case (b) holds.
3. A symbol [i, j] with i and j both > 0 has exactly one predecessor,
namely [i - 1,j - 1].
4. If [i,j] is a symbol, then either Ii < rj S 0 or 0 S Ii < rj' where <
denotes the ordering obtained by projecting the strip S onto the branch line
BB of L o .
5. Since, by assumption 5 of the preceding section, the orbit WI(O), after
coming to 11' goes at least twice more round the fixed point PI before crossing
over the center into the right half of the strip S, we see that the first few
symbols, starting with [1,0] are
[1,0]
[2,0]
/ \ [0,1]
j
[3,0]
\ /
[0, 1] [1,0]
\
[0,2].
In particular [0, 1] has at least two different predecessors, [1, 0] and [2, 0].
6. Each symbol is an ultimate predecessor of any other, in the sense that
if (J = [i,j] and r = [i',j'] are given, then there is a finite sequence (Jo = (J,
(J l' (J 2, ... , (J n = r such that (J k always precedes (J k+ 1. That is clear if (J 0 = [1,0]
Prehistories 299
(or [0, IJ), since, by the rules, all other symbols follow from [1, OJ or [0, IJ,
and we have just seen that [1, OJ and [0, IJ follow from each other. On the
other hand, if T = [1, OJ (or [0, IJ), and (J is arbitrary, the sequence can be
found by appeal to what is essentially the locally eventually onto property
of the Poincare mapping %(s) of the branch line BB, which says roughly
that the sequence can be so chosen that the width of the interval (Ii' r)
constantly increases. Williams proved that [1, OJ (hence also [0, IJ) can
be reached from arbitrary (J in a finite number of steps if the derivative
o/'o(s) is > fi for all s. In the case studied by Lorenz, the minimum of o/'o(s)
is more like 1.05; however, we can replace the archlength s by a new parameter
on BB, by the method discussed in Section 31.11, so as to convert the graph
of %(s) into two straight lines, and then o/'o(s) is indeed >fi (in fact about
1.9).
7. For each i > 0 (and eachj > 0) there is at least onej (one i) such that
[i, j] is a symbol.
31.15 Prehistories
According to Section 31.4, one of the characterizing features of an atttactor L

is that a point x belongs to L if and only if the orbit x(t) for which x(O) = x
lies in L for all t < 0 as well as all t ~ o. As t -+ - 00, x(t) passes through a
unique sequence of surface elements :!:, :!:', :!:", ... of the kind described in
Section 31.13. If the orbits in :!: are followed backward in time, they lie in
:!:', and so on. Hence, the initial curve of:!: is at least a part of the final curve
of :!:'. If the initial curve of:!: connects Ii to rj and that of :!:' connects Ii' to
rj" then the symbol [i', j'] precedes the symbol [i, jJ. Generally, [i, jJ may
have many predecessors, but as soon as a particular I-cell connecting Ii
to rj is chosen, a unique predecessor is thereby singled out, because the
initial curve of the surface element :!:' is uniquely determined by its final
curve.
It follows that if [i,j] is a symbol, each I-cell connecting Ii to rj is charac-
terized by a unique infinite sequence of symbols ... , (J _ 2, (J -1, ao = [i, jJ
such that (J _ k _ 1 precedes (J _ k' for each k.
It can be shown that if we extend the sequence also indefinitely in the other
direction,
(31.15-1)
we single out not merely a particular I-cell connecting Ii to r j , but a single

point of that I-cell, so that each sequence (31.15-1), where (Jk always precedes
(Jk+ 1> corresponds to a unique orbit on L. That follows from the locally
eventually onto character of the Poincare map in BE. If we follow two orbits
forward in time, then no matter how close together they are at t = 0, they
will eventually separate so as to lie in different surface elements L at some
time t > o.
31.16 The Lorenz Attractor; Detailed Structure II
We show first that if [i,j] is a symbol, there are uncountably many I-cells
in F connecting I; to rj. According to the preceding section, any such I-cell
corresponds to a unique sequence of symbols
(31.16-1)
where each (J - k - l is a predecessor of (J -k' Conversely (see next section), any

such sequence determines a I-cell, and we shall see that there are uncountably
many choices for the sequence, for given (Jo. Once any (J -k has been chosen,
it is possible, according to (6) of Section 31.4, to choose (J -k-l, (J -k-2, ,
so as to reach [0, IJ in a finite number of steps, say (J -k-/ = [0, 1]. Then,
according to (5) of that section, there are at least two choices of (J - k - / - t.
Hence, there is the possibility of a twofold choice infinitely many times in
choosing the sequence. If we represent the nth choice by a binary digit an'
we see that there are at least as many sequences (31.16-1), for given (Jo, as
there are real numbers .aOala2 in the unit interval, namely uncountably
many.
It follows that a line transverse to the branched surface Lo intersects
uncountably many sheets of the attractor L. Let M denote the point set, on
that line, of intersections with L. Since L is a closed set in 1R 3 , M is a closed set
on that line. As noted in Section 31.9, L has zero Lebesgue measure in 1R3;
hence M has zero measure on the line, for otherwise the Cartesian product
of M with a piece of surface of one of the sheets of L would have positive
3-dimensional measure. Hence M is an uncountable closed set of measure
zero. Lastly, M has no isolated points, for the topological arguments of
Williams show that if a I-cell is determined by a sequence (31.17-1), there are
other I-cells arbitrarily close to it obtained from sequences that agree with
(31.16-1) sufficiently far back. Hence, M is a Cantor set.
Figure 31.9 A "Cantor book."

Existence of i-Cells in F 301
According to (7) of Section 31.14, each point Ii and each point fj is the
terminus of uncountably many I-cells in F. If we move those I-cells along
by the flow, we see that the unstable manifold WU(O) is, throughout its entire
length, the spine of a Cantor book, in the terminology of Williams (see Figure
31.9).
A sequence (31.15-1), if periodic, corresponds to a periodic orbit on L.
Given any sequence (31.15-1), we can clearly find a periodic sequence that
agrees with the given one for say - K < k < K, where K is large. Hence,
given any orbit, we can find a periodic orbit arbitrarily close to it; the periodic
orbits are dense in L. (In a physical realization or numerical simulation, of
course, the notation of a strictly periodic orbit is an empty concept, owing to
the finite accuracy and the Lyapuonov instability.)
31.17 Existence of I-Cells in F
In Section 31.13 we found it necessary to assume in advance that there is at

least one I-cell in F, for example one connecting 0 to fl' (We then deduced
the existence of other I-cells by letting the first be carried round by the flow
until it intersects the strip S again, and so on.) That is, we assumed that the
attractor L contains a curve lying in S and connecting 0 to fl' The projection
of F onto the surface Lo is the branch line BB, which is a curve in space.
That much is supported by the numerical evidence. For the detailed structure
of F itself, we must appeal to the assumptions 1-5 that were made in Section
31.13 about the attractor L. Here we shall indicate how those assumptions
imply the existence of a 1-cell connecting 0 to f1 (or, by the same argument,
11 to 0).
We shall consider also the following problem. It was shown that if 0"0
= [i,j] is a symbol, then any 1-cell connecting Ii to fj determines a unique
sequence .. . ,0" -2,0" -1,0"0 such that 0" -k-1 always precedes 0" -k' The problem
is to show that conversely any such sequence determines a 1-cell connecting
Ii to f j . ,
Williams's solution of the first problem starts by letting Xo 0 be any point

of the branch line BB between 0 and the right end fj' constructing an orbit
xo(t) in the branched surface L o , such that xo(O) is =x oo , by specifying its
prehistory, then calling x(t) the orbit in L whose projection onto L o , accord-
ing to assumption 4 in Section 31.13, is xo(t). The construction is such that
the orbit x(t) depends continuously on the position ofx oo on BB; hence the
point x(O) of the orbit depends continuously on Xo 0, hence traces of a curve
in F, as Xoo is varied, and furthermore that curve touches 0 and fl at its
two ends.
There are presumably many ways of choosing the prehistory of xo(t) that
would achieve the desired effect, for there are infinitely many 1-cells con-
necting 0 to fl in F. The one selected by Williams is to make the prehistory
alternate between the two halves of BB; namely, if for some tb the point
X O(t 1 ) of the prehistory is on BB (but not at 0 or either end), the orbit is
continued backward (upward) in the sheet of Lo behind if XO(tl) lies to the

right of 0 and in the sheet in front if it lies to the left. It is easily verified by
reference to Figure 30.3a that that is always possible and never causes the
prehistory to hit 0 or r 1 or 11' To show the continuous dependence on Xo 0,
we observe that if Xo 0 and Xo 0 are close together on the branch line BB, the
resulting orbits xo(t) and xo(t) in Lo are close together for a long interval,
and then the same follows for the orbits x(t) and x'(t) in L by assumption 4
of Section 31.13. To show that the resulting curve in F has the correct end-
points, note first that if Xo 0 is very close to the right end of BB, the prehistory
is very close to the outside edge of that sheet of Lo that lies behind; hence, in
the past, x(t) must have spent a long time near the stagnation point 0, and
hence was close for a long time to the orbit W~(O) that goes from 0 to rl'
Lastly, if Xo 0 is very near the center of the branch line BB, the orbit is already
very near the stagnation point, and hence was close, for a long past interval,
to the constant orbit x(t) == O.
To show similarly that any sequence
(31.17-1)
where (J -k-l is a precedessor of (J - b determines a 1-cell in F is more compli-
cated, and we refer the reader to the paper of Williams, where the point set F
is approximated by so-called retractions. The idea is that if Xo 0 is a point of
BB between the projections onto BB oft; and r j , we let the sequence (31.17-1)
dictate which sheet of Lo the prehistory xo(t) follows each time the branch
line is encountered, as we follow the orbit xo(t) backward (upward) across
BB. For technical reasons, it is expedient to follow the sequence (31.17-1)
in this way for only a finite number, say n, of steps and to decree that prior
to that the prehistory alternated between the two halves of BB as in the above
discussion. The resulting orbit x(t) is then not quite the one we want, but is
close to the one we want and gets closer, as n ~ 00.
The conclusion that each sequence (31.17-1) determines a unique 1-cell
in F has already been used in the preceding section, to show that the sheets
of L are uncountable.
31.18 Bifurcation to a Strange Attractor
We noted in Section 31.9 that the bifurcation in the Lorenz system at r = ro

= 24.74 is subcritical, and we mentioned the possibility of an explosive
transition when the fixed points PI and P 2 lose stability. However, as r is
slowly increased past r 0, the motion on the attract or L takes over without
any explosive transition, except insofar as the sudden appearance of the
motion on L may be regarded as an explosion.
The attract or has been described for r = 28. If we let r decrease down to
r 0, the two holes in the branched surface Lo close up by contracting onto the
fixed points PI and P 2' For still smaller values of r, down to 13.96, L continues
to exist as an invariant set for the motion (though not an attract or), but it is in
The Feigenbaum Model 303
contact with the fixed points PI and P z only at r = roo For r only slightly
above ro, an orbit emerging from PI or P z is immediately an orbit on L.
In this sense the bifurcation at ro results in .an abrupt transition from a
stationary orbit at PI or P z to motion on the attract or L.
31.19 The Feigenbaum Model
While the Lorenz attractor appears in connection with a subcritical Hopf

bifurcation, the Landau-Hopf model and the Ruelle-Takens model both
require a sequence of supercritical bifurcations leading to invariant tori of
successively higher dimension, arbitrarily high in the former model and of
dimension at least 4 in the latter. However, such a sequence is unlikely,
according to Peixoto's theorem. As was pointed out at the end of Section
29.10, the appearance of an invariant 2-torus at a bifurcation from a periodic
orbit does not imply the appearance of orbits dense on that torus. Instead,
the appearance of finitely many periodic orbits and fixed points is generic.
Other orbits tend asymptotically to those periodic orbits and fixed points.
The appearance of an invariant 3-torus at the next bifurcation depends, at
least in the theory of Chenciner and Jooss 1976, on the existence of an orbit
dense on the 2-torus. Hence, the bifurcation to an invariant 3-torus seems
unlikely.
If a periodic orbit on the 2-torus goes round the long way n times before
closing, then the bifurcation is subharmonic with a sudden n-folding of the
period at the bifurcation (see Section 29.11). Recently, M. Feigenbaum has
developed a model based on a sequence of subharmonic bifurcations with
period doublings. (See Feigenbaum 1980 and the references given there.) It
turns out that such doublings occur in many examples of iterated mappings
and simple dynamical systems. Furthermore, as the number n of doublings
increases, the behavior of the system is governed by certain asymptotic laws
that involve universal constants and functions, independent of the system
under study. In addition, the asymptotic laws appear to hold quite accurately
for rather small values of n. In particular, the values f.1n of the dimensionless
parameter f.1 at which the bifurcations (doublings) take place converge to a
valuef.1oo geometrically, with
f.1n+ I - f.1n ~ 0.21416938

f.1n - f.1n-l
for large n. As n ~ 00, at least in the cases studied, the power spectrum of the
motion approaches a continuous spectrum with certain universal features.
At f.1 = f.100' the motion is presumably aperiodic on a strange attractor.
There is evidence (Lorenz 1981) for an example of this behavior in the
Lorenz system at considerably higher values of the dimensionless parameter
r than values studied by Lorenz. Namely, the strange attract or that appears
at r = 24.74 persists up to a value r = r* (~250). For r considerably greater
than r*, there is a periodic orbit, and as r is decreased toward r*, there is a
sequence of doublings at values rn of r that converge to r* from above, with
rn+l - rn::::: 0.214.
rn-rn-l
Appendix to Chapter 31-Generic Properties of

Systems
In this appendix, generic and nongeneric properties of systems are explained
and explored. It might be thought that these notions will in a sense replace
probability in some parts of physics. We shall conclude that that is not so,
but that they may be an important guide to the further development of our
ideas of probability in physics.
31.A Spaces of Systems

Because a physical system cannot be specified exactly, and for other reasons,
it is often desirable to consider not a single system but a large family of
them. If each system is characterized by the values of n parameters 1X1' , IXn'
then each system corresponds to a point in the space [Rn. Conversely, each
point in [Rn or in some region f!ll c [Rn, may correspond to a unique system
of the family.
More generally, the systems of a family may be represented by the points in
a Banach or Hilbert space, or in some more general metric space, or in some
still more general topological space, which we shall call the space of systems.
For example, each system might be a dynamical system in the plane:
x = X(x, y), y=Y(x, y),
where X and Yare the components of a given smooth vector field X(x).
Then each such vector field determines a system and may be represented by
a point in a Banach space ~ with a suitably chosen norm IIXOII. Clearly
~ is infinite-dimensional, because no finite number of parameters can specify
a vector field completely.
We shall suppose generally that the space at least has a metric. Then if
the distance d( Xl, X 2) between two systems is small, we can think of either
one as obtainable from the other by a small perturbation. We shall even
usually assume that the space has a norm, so that d(Xb X 2 ) = IIX l - X211.
31.B Absence of Lebesgue Measure in a

Hilbert Space
If there are finitely many parameters IXb , an and if they are distributed
according to a continuous probability law in [Rn, and if a property holds in all
[Rn except for a set of Lebesgue measure zero, we say that it holds for almost
all systems, or that the probability of its not occurring is zero.
Appendix to Chapter 31-Generic Properties of Systems 305
If we let n --+ 00, so that [Rn is replaced by an infinite dimensional Hilbert

space, then, as shown in Section 13.11 of Volume I, there is no Lebesgue
measure; hence, probability statements of the above kind cannot be made.
(Non-Lebesgue probability measures are discussed in Section 3l.H at the
end of this appendix.)
31.C Generic Properties of Systems

In any space of systems, whether it has Lebesgue measure or not, assuming
it at least has a topology, one can define what is meant by generic and non-
generic properties of the systems. This is often done with the implied sug-
gestion that nongeneric properties can be ignored from some points of view.
A subset of the space is called a Baire set if it is the intersection of count ably
many dense open subsets. The complement of a Baire set is called a meager
set; it is the union of countably many nowhere dense sets.
A property of a system is called generic if it occurs on a Baire set in the
space of systems. A property is called nongeneric if it occurs on a meager set.
Note that "nongeneric" does not mean merely "not generic," because a
set may be neither a Baire set nor a meager set, for example, a half-space
Xl > 0 in [Rn.
The Baire category theorem says that if the space is a complete metric
space a Baire set is dense in it, but a meager set may also be dense; hence the
distinction is not based on denseness. Furthermore, in case the space happens
to be finite-dimensional a Baire set may have Lebesgue measure zero; hence
the distinction is not based on measure, either. See Section 31A.G below.
If a property a is generic, the property "not a" is nongeneric; if a and b
are generic, the property" a and b" is generic. Two contradictory properties
cannot both be generic.
31.D Strongly Generic; Physical Interpretation

We shall not give a physical interpretation of genericity, except in the
following special case: A property is called strongly generic if it occurs on a
dense open set in the space, i.e., not merely on a countable intersection of
such sets. (We note in passing that the intersection of finitely many dense
open sets is again a dense open set.) If a property is strongly generic, if X is
any system, and if Gl is any positive number, then, by a perturbation of X of
magnitude SGl' we can obtain a system Y (II Y - XII S Gl) which has the
property. Furthermore, there is then another positive number G2 <:1) such
that the system retains the property under any arbitrary further perturbation
of magnitude SG2. Nothing is said aqQ1i1t how difficult it may be to find the
first perturbation. (There may alsobt perturbations of X of magnitude
SGI that do not give the system the desired property.) But all sufficiently
carefully prepared systems have the property; hence the property cannot
be dismissed as one "not likely to appear in practice."
If a property is strongly nongeneric, that is, if it occurs on a nowhere

dense set in the space, then, for any e1 ~ 0 there is a perturbation of norm
~e1 that makes it disappear, and then there is an e2 ~ 0 such that no further
perturbation of norm ~ e2 can make it reappear.
31.E Peixoto's Theorem

The theorem of Peixoto 1962 says that for dynamical systems on a compact
2-dimensional manifold, a particular strongly generic property is that every
motion tends asymptotically to one of a finite number of fixed points and
periodic orbits. Hence, in particular, quasi-periodic motions are nongeneric.
We state the theorem for the 2-torus; 9Jl = T2. Let () and q> be angle
variables on T2, so that the dynamical system is
e= F(), q, cp = G(), q,
where the functions F and G are periodic in both () and q> with period 2n.
Let !!3 be the Banach space of all pairs of periodic functions F and G of class
C 1 , with the norm given by
max(8.",) max{IFI, IGI, laoFI, laoGI, la",FI, la",GI},
so that the topology in !!3 is that of uniform C 1 convergence. Each point of
!!3 represents a dynamical system. Then there is a dense open set in !!3 repre-
senting dynamical systems, for each of which
I. There are at most finitely many fixed points and finitely many closed
orbits on T2.
u. Each motion tends asymptotically to a fixed point or closed orbit.
(This is intended to include stationary motions at the fixed points and
periodic motions on the closed orbits.) The theorem actually says a
little more than this; see Peixoto 1962, also Section 31.F below.
This is in sharp contrast with the Kolmogorov-Arnol'd-Moser theorem
(see Moser 1973, Theorem 2.8), which says that for Hamiltonian systems,
under certain assumptions, quasi-periodic motions are stable under almost
all small perturbations. There is no contradiction, here, because the Hamil-
tonian systems are very special (for example, they are conservative) and are
in a sense nongeneric among dynamical systems as a whole, so that the small
perturbations referred to are, in this case, far from arbitrary, but are such
that the system remains Hamiltonian.
31.F Other Examples of Generic and

Nongeneric Properties
Consider the space C[O, 1] of continuous functions defined for 0 ~ x ~ 1
with the usual norm Ilfll = suplf(x)l. For such functions, differentiability
is a nongeneric property, and even differentiability at a single point, and
even Lipschitz continuity at a single point; see Boas 1960, also Exercise 1
below.
The next two examples are part of a more complete version of Peixoto's
theorem than the one given above. As stated, it is generic for a vector field
on a compact 2-manifold to have a finite number of fixed points and periodic
orbits. In a coordinate system with a fixed point at the origin, the vector field
is
F(x) = Ax +"',
where A is a 2 x 2 matrix. The theorem goes on to say that generically each
such fixed point is hyperbolic, which means that no eigenvalue of A is pure
imaginary; hence there are three possibilities: If both eigenvalues have
negative real parts, the fixed point is attracting; if both have positive real
parts, it is repelling; and if one has a positive real part and the other a negative
one, it is a saddle point.
A further conclusion of Peixoto's theorem is that the existence of an orbit
going from one saddle point to another is nongeneric; if there is such an
orbit, for some vector field, the smallest perturbation can make it miss the
second saddle point.
A conjecture that has not been proved, or even fully formulated, was made
in Section 29.6. As was stated there, the existing completeness theorems on
systems of eigenfunctions for hydrodynamic problems require that the
generalized eigenfunctions, if any, be included; see equation (29.6-5). It is
conjectured that in some suitable space of hydrodynamic systems, the exis-
tence of generalized eigenfunctions is nongeneric. An eigenvalue A is said to
have index 1 if there are no corresponding generalized eigenfunctions. It
seems likely to be strongly generic for any eigenvalue Ak to have index 1, but
then only generic for them all to have index 1, since a countable collection
of dense open sets is not necessarily open, but is in any case a Baire set.
EXERCISE
1. For each positive integer n define a subset En of C[O, I J as follows: A function

J is in En ifthere is an Xo in [0, 1 - IlnJ such that
IJ(xo + h~ - J(x o) I ::;; n for 0 < h < ~. (31.F-I)
Show first that En is a closed subset of C[O, IJ by showing that if Jk(X) --+ J(x) uni-
formly, as k --> 00, and if each Jk satisfies (31.F-l), then so does f. Show next that the
complement of En is dense; show in fact that ifJis any function in C[O, IJ, it can be made
differentiable by a small perturbation, using a mollifier (see Volume I), then can be made
to violate (31.F-l) by adding to it, if necessary, a further small perturbation having a
sufficiently large derivative at x = xo. Hence each En is a nowhere dense set, hence
U:~ 1 En is a meager set, and hence it is nongeneric for a function in C[O, 1] to be
Lipschitz continuous on the right at any point.
31.G Lack of Correspondence between Genericity

and Lebesgue
EXERCISES
I.Let X be the unit interval on the real line, as a metric space with the distance
(x, Y) =Ix - YI. Let n be a positive integer. With each rational number p/q, where
o < p < q, assocaite the interval
pIp 1
----<x<-+--,
q qnq2 q q nq2
and let Sn be the union of those intervals. Show that Sn is open and dense in X and has
Lebesgue measure:::; 4/n. Conclude that a Baire set in X can have zero Lebesgue measure,
and a meager set can have measure 1.
2. Let S be the intersection of the sets Sn. According to Exercise 1, S is a Baire set.
Show that it is uncountable by considering numbers x in [0, IJ having a binary repre-
sentation
x = .a j 0 a2 00 ... 0 a 3 00 ... 0 a4 0 0 ... ,
where each al is = 0 or 1. Show that if the number VI of zeros between al and al+ j increases
rapidly enough with I, then x is in all the sets Sn. Show that the set of all such x is un-
countable.
31.H Probability and Physics
Suppose that a topological space X is a space of physical systems. When a

physicist says that a certain property of such a system is very unlikely to
appear, he means presumably that if a large number N of the systems is
somehow selected from X, very few of them have the property in question,
and, as N increases, the fraction of them that have the property tends to zero.
Such a statement, if true, is unavoidably a probability statement. What is
lacking from this description is a statement of how the N systems are selected
from X. Since X is generally infinite-dimensional, and hence has no Lebesgue
measure, there is no possibility of the N systems being chosen at random
from a uniform distribution in X, because no such distribution exists.
The method of selection has to be determined by considerations of physics,
and cannot be decided by considerations of genericity alone.
There are of course many ways of defining probability measures in infinite-
dimensional spaces; the Gaussian measures in a Hilbert space, described in
Volume I, Section 13.11, are examples. Such a measure could be used for the
selection of the N systems from X. The possibilities are presumably very
numerous, and it may be necessary to appeal to many physical principles.
For example, it would seem reasonable to require that the probability
measure be positive on the open sets in X, for otherwise it would seem that
the choice of the space X was itself faulty; parts of it should be discarded. Then
suppose there is a transformation in X that represents some modification
Appendix to Chapter 3 I-Generic Properties of Systems 309
of the systems so as to alter their main properties without causing very

"improbable" properties to become suddenly "probable"; then under such
a transformation in X sets of measure zero should go into sets of measure
zero. In view of the physical interpretation given in 31.D, one requirement
might be that a strongly nongeneric property should appear only on a set of
measure zero.
All this is speculation at present, but it may indicate that genericity cannot
replace probability in physics, but can perhaps be a guide to the appropriate
choice of probability measures.
EXERCISES
1. Show that differentiation is nongeneric also in the Hilbert space fl = L 2 [0, 1],
by filling in the steps of the following outline of a proof: First define a linear manifold
in fl by
(31.H-l)
where the derivative is meant in the distribution sense (see Chapter 5 in Volume I).
For any if; in fl, write
00
if;(x) = I ~ke2nikx. (31.H-2)

-00
Then the manifold '!l can be characterized as
(31.H-3)
Define also the smaller manifold
(31.H-4)
for each M = 1, 2, .... If it can be shown that each '!lM is nowhere dense, i.e., that its
complement C'!lM is dense and open, it will follow that '!l = UKi~ 1 '!lM is a meager set
in fl. Show first that C'!lM is dense by showing that any if; E L 2 that does not already
violate the condition I if; , I :S; M can be made to do so by-adding to it an arbitrarily small
function with a large derivative. Then show that C'!lM is open as follows: Consider any
if; in C'!lM' that is, any if; such that either IWII = CJJ or IWII = M + b for some b > O.
It must be shown that there is a neighborhood of that if; that is contained in C'!lM'
Choose K such that
and show that if Xis any function in L2 such that I xii is < b/4K, then IW + x'il > M + ib,
so that C'!lM is open.
2. Show that the same is true in the real Hilbert space L 2 [0, 1] by using the same
argument as above but assuming throughout that ~-k = ~.
In order to introduce real coordinates and real basis functions, we write
~o = XO,
j:
c.,k -
_ Xk +
fiiX- k (k
= 1, 2, ...), (31.H-5)
IPo(x) == 1, IPk(X) fi cos 2nkx,

=
(31.H-6)
IP-k(X) = fi sin 2nkx.
Then, from (31.H-2) with ~-k = ~,we find
00
tjJ(x) = I Xk({Jk(X), (31.H-7)

-00
00
tjJ'(x) = 2n I kx-k({Jk(X). (31.H-8)

-00
Gaussian measures in a real Hilbert space ~ were described in Section

13.11 of Volume I; the main points are as follows: If9Jl is any finite-dimen-
sional subspace of ~ and S is any Borel set in 9Jl, then the set
Z = S + 9Jl.L,
i.e., the set of all points x + y, where XES and y E 9Jl.L, is called a cylinder set.
To define a so-called Gaussian measure in~, we let B be a positive operator
of the trace class, and we call A = B- 1 . Then, given any cylinder set Z = S
+ 9Jl.L, where dim 9Jl = m, we let {({J j}i be an orthonormal set in 9Jl, and we
define an m x m matrix A(9Jl) by A(9Jl)jk = (({Jj' AIPk), 1 S j, k S m. Then
the probability P(Z) is taken as
P(Z) = Jdet A(9Jl)

(2n)m/2 s exp
f {1 t -2
}
x j A(9Jl)jk Xk dV. (31.H-9)
As stated in Section 13.11, this set function PO can be extended to a unique

probability measure defined on the a-algebra ~ generated by the cylinder
sets.
We now apply these ideas to the real Hilbert space ~ = L2[0, 1J, using
the coordinates Xk and basis functions ({Jk given by (31.H-5, 6). We define the
operator A by the simple formula
00
AtjJ(x) = I akxk({Jk,
-00
where tjJ(x) is given by (31.H-7), and where the ak are POSItIve and
L (l/a k ) < 00, so that B = A -1 is compact. We consider the cylinder sets
Z M.K = ~ Ikx
{'J'I'' E L 2.. 4n 2 -K
1... k 12 <
- M2} ,
so that
(31.H-1O)
where !lM is given by (31.H-4). Then (31.H-9) takes the form
P(ZM,K) = IT Jf.i j-s' fexp { -t ~ akxf }dX_K'" dXK, (31.H-ll)
where S is the ellipsoidal region
(31.H-12)
EXERCISES
3. By letting aU the integration variables except the last in (31.H-ll) go from

- 00 to + 00, while the last one XK is restricted to IXKI ~ M/(2nK), show that
M~
P(ZM K) < ~ .
, y2n 3 K
4. Now take ak = Ikl r , where r> 1 (to make L (l/a k ) converge). Show that if
\
1 < r < 2, then P(ZM,K)---> 0 as K ---> 00, so that from (31.H-I0), PCD M) = 0, and hence
by countable additivity,
This exercise shows that the meager set l) of differentiable functions in

L2[0,1] has Gaussian measure zero if 1 < r < 2. For r > 2, P(!l) = 1.
Roughly speaking, when r is increased, the Gaussian measure is concentrated
more and more toward the origin in ~. The differentiable functions lie near
the origin, according to (31.H-3), so that if the concentration of the proba-
bility measure toward the origin is large, the differentiable functions acquire
positive probability, but for r < 2 they have zero probability.
In this case, therefore, the probability measure can be so chosen that the
nongeneric property (differentiability) is improbable.
References
Abraham, R., and Robbin, J. (1967): Transversal Mappings and Flows. W. A. Benjamin,
New York.
Adler, R., Bazin, M., and Schiffer, M. (1965): Introduction to General Relativity.
McGraw-Hill, New York.
Barut, A. 0., and Ra<;zka, R. (1977): The Theory of Group Representations and Appli-
cations. PWN Polish Scientific Publishers, Warsaw.
Behrends, R. E., Dreitlein, J., Fronsdal, c., and Lee, W. (1962): Simple groups and
strong interaction symmetries. Rev. Mod. Phys., vol. 34, pp. 1-40.
Bacher, M. (1922): Introduction to Higher Algebra. The MacMillan Co., New York.
Boerner, H. (1955): Darstellungen von Gruppen. Springer-Verlag, Berlin, Heidelberg,
New York.
Chevalley, C. (1946): Theory of Lie Groups I. Princeton Univ. Press, Princeton.
Curry, J. H. (1978): A generalized Lorenz system. Comm. Math. Phys., vol. 60, pp.
193-204.
Davey, A. (1962): The growth of Taylor vortices in flow between rotating cylinders. J.
Fluid Mech., vol. 14, pp. 336-368.
Davey, A., DiPrima, R. c., and Stuart, J. T. (1968): On the instability of Taylor vortices.
J. Fluid Mech., vol. 31, pp. 17-52.
Dirac, P. A. M. (1928): The quantum theory of the electron. Proc. Roy. Soc. A, vol.
117, pp. 610-624.
Dirac, P. A. M. (1958): The Principles of Quantum Mechanics. Clarendon Press, Oxford.
Eagles, P. M. (1971): On stability of Taylor vortices by fifth-order amplitude expansion.
J. Fluid Mech., vol. 49, pp. 529-550.
Eisenhart, L. P. (1926): Riemannian Geometry. Princeton Univ. Press, Princeton.
Eisenhart, L. P. (1933): Continuous Groups of Transformations. Princeton Univ. Press,
Princeton.
Eisenhart, L. P. (1934): Separable systems ofStaeckel. Ann. Math., vol. 35, pp. 284 ff.
Ellis, H. G. (1972): Ether flow through a drainhole: a particle model in general relativity.
J. Math. Phys., vol. 14, pp. 104-118.
Feigenbaum, M. (1980): Universal behavior in nonlinear systems. Los Alamos Science,
vol. 1, pp. 4-27.
Finkelstein, D. (1958): Past-future asymmetry of the gravitational field of a point
particle. Phys. Rev., vol. 110, pp. 965-967.
313
314 References
Flanders, H. (1963): Differential Forms. Academic Press, New York.

Gel'fand, I. M., Graev, M. I., and Vilenkin, N. Ya. (1966): Generalized Functions, Vol.
5, Integral Geometry and Representation Theory, (translated from the
Russian) Academic Press, New York.
Gel'fand, I. M., Minlos, R. A., and Shapiro, Z. Ya. (1963): Representations of the Rotation
and Lorentz Groups and Their Applications, (translated from the Russian)
MacMillan (Pergamon), New York.
Gleason, A. (1952): Groups without small subgroups. Ann. of Math. vol. 56, pp.
193-212.
Gollub, J. P., and Swinney, H. L. (1975): Onset of turbulence in a rotating fluid. Phys.
Rev. Letters, vol. 35, pp. 927-930.
Haar, A. (1933): Der Massbegriff in der Theorie der kontinuierlichen Gruppen. Ann.
Math., vol. 34, pp. 147-169.
Hassard, Brian, D. (1980): Computation of invariant manifolds. To appear.
Hausner, M., and Schwartz, J. T. (1968): Lie Groups and Lie Algebras: a Brief Descrip-
tion. Gordon and Breach, New York.
Hocking, J. G., and Young, G. S. (1961): Topology. Addison Wesley, Reading, Mass.
Hopf, E. (1948): A mathematical example displaying the features of turbulence. Comm.
Pure Appl. Math., vol. 1, pp. 303~322.
Kelley, Al (1967): The Stable, Center-Stable, Center, Center-Unstable, and Unstable
Manifolds. Appendix C in Abraham and Robbin 1967.
Kerr, R. P. (1963): Gravitational field of a spinning mass as an example of algebraically
special metrics. Phys. Rev. Letters, vol. 11, pp. 237-238.
Krueger, E. R., Gross, A., and DiPrima, R. C. (1966): On the relative importance of
Taylor-vortex and non-axisymmetric modes in flow between rotating
cylinders. J. Fluid Mech., vol. 24, pp. 521-538.
Kruskal, M. D. (1960): Maximal extension of Schwarzschild metric. Phys. Rev., vol.
119, pp. 1743-1745.
Kurosh, A. G. (1956): Theory of Groups 1,11. Chelsea Pub. Co., New York.
Ladyzhen~kaya, O. (1969): Mathematical Theory of Viscous Incompressible Flow.
Gordon and Breach, New York.
Ladyzhenskaya, O. (1975): Mathematical analysis of Navier-Stokes equations for
incompressible liquids. Annual Rev. of Fluid Mech., vol. 7, pp.
249-272.
Landau, L. D., and Lifshitz, E. M. (1959): Fluid Mechanics. Pergamon Press, London.
Lanford, O. E. (1973): Bifurcation of periodic solutions into invariant tori: the work
of Ruelle and Takens. in Nonlinear Problems in the Physical Sciences and
Biology. Lecture Notes Vol. 322, Springer-Verlag, Berlin, Heidelberg,
New York.
Lang, S. (1962): Introduction to Differentiable Manifolds. Wiley Interscience, New York.
Lorenz, E. N. (1981): Noisy periodicity and reverse bifurcation. To appear.
Lorenz, E. N. (1963): Deterministic nonperiodic flow. Jour. Atmos. Sci., vol. 20, pp.
130-141.
Marsden, J. E., and McCracken, M. (1976): The Hopf Bifurcation and Its Applications.
Springer-Verlag, Berlin, Heidelberg, New York.
References 315
Miller, W. (1973): Symmetry Groups and Their Applications. Academic Press, New York.
Montgomery, D., and Zippin, L. (l952)-see next item.
Montgomery, D., and Zippin, L. (1955): Topological Transformation Groups. Inter-
science Publishers, New York.
Morse, P. M., and Feshbach, H. (1953): Methods of Theoretical Physics I, II. McGraw-
Hill, New York.
Moser, J. (1973): Stable and Random Motions in Dynamical Systems. Princeton Univ.
Press, Princeton.
Nachbin, L. (1965): The Haar Integral. Van Nostrand, Princeton.
Naimark, M. A. (1976): The Theory of Group Representations. Nauka, Moscow.
Naimark, M. A. (1959): On some cases of periodic motions depending on parameters.
Dokl. Akad. Nauk SSSR, vol. 129, pp. 736-739.
Pauli, W. (1927): Zur Quantenmechanik des magnetischen Elektrons. Zeits.! Physik,
vol. 43, pp. 601-623.
Peixoto, M. (1962): Structural stability on two-dimensional manifolds Topology,
vol. 1, pp. 101-120.
Peter, F., and Weyl, H. (1927): Die Vollstandigkeit der primitiven Darstellungen einer
geschlossenen kontinuierlichen Gruppe. Math. Ann., vol. 97, pp. 737-755.
Redei, L. (1959): A(qebra. Akademische Verlagsgesellschaft, Leipzig.
Richtmeyer, R. D., and Morton, K. W. (1967): Difference Methods for Initial- Value
Problems. Wiley-Interscience, New York.
Richtmeyer, R. D. (1981): A study of the Lorenz attractor. Advances in Mathematics,
to appear.
Riesz, F., and Sz. Nagy, B. (1953): Ler;ons d' Analyse Fonctionnelle. Akademiai Kiad6,
Budapest.
Robertson, H. P. (1927): Bemerkung iiber separierbare Systeme in der Wellenmechanik.
Math. Annalen, vol. 98, pp. 749 ff.
Ruelle, D., and Takens, F. (1971): On the nature of turbulence. Comm. Math. Phys.,
vol. 20, pp. 167-192.
Russmann, H., and Zehnder, E. (1980): On a normal form of symmetric maps of[O, IJ.
Comm. Math. Phys., vol. 72, pp. 49-53.
Sacker, R. (1964): (a bifurcation theorem) Thesis (unpublished), New York University.
Schiff, L. (1955): Quantum Mechanics. McGraw-Hill, New York.
Schur, I. (1905): Neue Begriindung der Theorie der Gruppencharaktere. Sitzungsber.
preuss. Akad. Wiss., vol 1905, pp. 406-432.
Sell, G. R. (1971): Topological Dynamics and Ordinary Differential Equations. Van
Nostrand-Reinhold, London.
Siegel, C. L. (1956): Vorlesungen aber Himmelsmechanik. Springer-Verlag, Berlin,
Heidelberg, New York.
Siegel, C. L. and Moser, J. (1971): Lectures on Celestial Mechanics. Springer-Verlag,
Berlin, Heidelberg, New York.
Smale, S. (1967): Differentiable dynamical systems. Bull AMS, vol. 73, pp. 747-817.
Snyder, H. A. (1970): Waveforms in rotating Couette flow. IntI. Jour. Nonlinear Mech.,
vol. 5, pp. 659-685.
316 References
Sommerfeld, A. (1929). Das Molekiil als symmetrische Kreisel, in Atombau und

Spektrallinien, wellenmechanischer Erganzungsband. Friedr. Vieweg u.
Sohn AG, Braunschweig.
Sugiura, M. (1975): Unitary Representations and Harmonic Analysis. Halsted Press,
New York.
Taylor, G. 1. (1923): Stability of a viscous liquid contained between two rotating
cyclinders. Phil. Trans. Roy Soc. A, vol. 223, pp. 289-343.
Thomas, T. Y. (1961): Concepts from Tensor Analysis and Differential Geometry.
Academic Press, New York.
Vilenkin, N. Ya. (1968): Special Functions and the Theory of Group Representations.
AMS Translations, Providence, R.l.
Warner, G. (1972): Harmonic Analysis on Semi-Simple Lie Groups I, II. Springer Verlag.
Weyl, H. (1928): Gruppentheorie und Quantenmechanik. S. Hirzel Verlag, Leipzig.
Weyl, H. (1932). The Theory o/Groups and Quantum Mechanics. E. P. Dutton and Co.,
New York.
Whitehead, J. H. C. (1932): Convex regions in the geometry of paths. Quart. Jour. of
Math., vol. 3, pp. 33-42.
Wigner, E. (1931): Gruppentheorie und ihre Anwendung auf die Quanenmechanik der
Atomspektren. Fr. Vieweg u. Sohn AG, Braunschweig.
Williams, R. F. (1977): The structure of Lorenz attractors. Turbulence Seminar (Univ.
of Calif., Berkeley) 1976-77, pp. 94-112.
Index
Abelian group 4 Boost, relativistic 33

Absolute differentiation along a Branched surface 288
curve 208
Addition theorem for tesseral
harmonics 59 Campbell- Baker- Hausdorff (CBH)
Adjoint representation of a Lie formula 144, 145
algebra 159 Cancellation, law of 3
Adjoint representation of a Lie group 159 Cantor book 301
Affine connection 195 Cantor set 300
Affine parameter on a geodesic 188 Cartan sub algebra 164, 165
Almost periodic motion 284 Cartesian product 111, 112
Aperiodic motion 284 Cauchy problem of Einstein's
Associated Legendre functions 55 equations 240
Associative law 1 Ck -compatible charts 99
extended 3 Center of a group 159
Attractor 282 Center of a Lie algebra 159
Autocovariance 283 Character of a representation 77
Automorphism Chart in a manifold 99
of a group 6 Christoffel symbols 188
of a Lie algebra 14 Closed subgroup 155
Ck-manifold 103
C "'-manifold 103
Bessel function 75 Commutative group 4
Bicontinuous 114 Compact group 66
Bifurcation 245, 246, 254, 286 Completely reducible representation 64
subcritical 255, 257, 287 Completeness of the tesseral
subharmonic 261 harmonics 60
supercritical 255, 257 Complete system of representations 71
Bifurcation to a strange attractor in the Complexification of a real Lie
Lorenz system 302 algebra 163
Bifurcation tree 277 Component of a manifold 104
317
318 Index
Composition of velocities in relativity 30 Factor group 10

Composition series 14 Faithful representation of a group 12, 41
Congruence group 203 Feigenbaum model 278, 303
Conjugate subgroup 13 Finitely presented group 16
Conjugation in a complex Lie Finkelstein charts 232, 233
algebra 169 Flatness 218
Continuation of geodesics 194 Free Abelian group 15
Constants of structure 135 Free group 15
Coordinate chart 99 Full Lorentz group 29, 31
Coset 10 Fully reducible representation 63
Couette flow 264 Functions of z andz 89
Covariant differentiation 206 Fundamental group of a manifold 106
Covariant vector field 177
Covering, covering manifold 115
Curry model of thermal convection 286 General linear group 25
Curve in a manifold 104 Generalized eigenfunctions 250
Cycle 6 Generalized weight vector 163
Cyclic group 4, 11 Generators of a group 15
Cyclic permutation 6 Generic properties of systems 304, 305
Cylindrical coordinates in the Geodesic 186,188, 196
Navier-Stokes equations 268 null 190
space like 190
time like 190
Decomposable representation 63 Geodesic completeness 235
Degenerate periodic function 17 Geodesic coordinates 201
Diffeomorphism 117 Geometry of paths 196
Direct product of groups 20 Good neighborhood 114
Direct sum of Lie algebras 160 Group 1
Direct sum of representations 64 Abelian (commutative) 4
Dotted spinor 94 compact 66
Dummy index 176 congruence 203
Dynkin diagram 167 cyclic 4, 11
factor group 10
finitely presented 16
Early linear regime in hydrodynamic free, free Abelian 15
stability problems 251 fundamental 106
Effective 48 general linear 25
Einstein field equations 226 Heisenberg 171
Einstein manifold 226 inner automorphism 13
Equivalent representations 62 Lorentz 29 - 31
Equivalent unitary transformations 81 orthogonal 25, 26
Euler angles 57 permutation 6
Euler's Theorem 27 Poincare 73
Euler variational equation 188 point 18
Even permutation 7 rigid motion 73
Explosive transition 287 rotation 25, 30, 50
Exponential mapping 136 simple 9
Extended associative law 3 space 17
Extension of a group 24 special linear 25
Index 319
special orthogonal 26 Invariant torus 257,260,261

translation 17 Inverse 3
unimodular linear 25 Irreducible representation 41, 63
unimodular orthogonal 26 Isomorphism 5
unitary 28
universal covering 159
Group manifold 34 Jacobi identity 135, 178
Jordan-Holder Theorem 14
Haar measure 67, 69

Hausdorff separation axiom 102 Kernel of a homomorphism 8
Heisenberg group 171 of a Lie group homomorphism 156
Helical vortices in the Taylor Kerr manifold 237
problem 265, 272, 273 Klein bottle 97
Hilbert space for the Taylor problem 270 Kneading sequence 295
Homeomorphism 117 Kruskal extension of the Schwarzschild
Homogeneous space 48 charts 233
Homomorphism 6, 8 Kruskal manifold 234
Homomorphism Law 11
for Lie groups 155
for Lie algebras 151 Lagrange's Theorem 10
Homomorphism of SL(2, C) onto the Landau-Hopf model of turbulence 276
proper Lorentz group 38 Late linear regime in hydrodynamic
Homomorphism of SU(2) onto SO(3) 37 stability problems 251
Homotopic curves 105 Latin square 3
Homotopy class of curves 105 Law of cancellation 3
Hopf example resembling the onset of Law of composition
turbulence 278 Law of homomorphism-see
Homomorphism Law
Left coset 10
Ideal 160 Left-invariant integral 67
Identification, method of, for Left-invariant measure 69
manifolds 97 Left-regular representation 69
Identiry 3 Left translation 12, 67
[i, j] symbols of Williams 297 Legendre functions 53, 55
Inaccessible stable helical vortices 273 associated 55
Induced topology 101 Legendre polynomial 56
Infinite-dimensional representation 45 Lie algebra 132 -134
Initial-value problem of a geodesic 190 nilpotent 163
Inner automorphism group 13 solvable 163
Inner automorphism of a Lie simple 150
algebra 139, 141 semisimple 163
Intrinsic coordinates in the rotation Lie algebra homomorphism 149
group 35 Lie bracket 134, 178
Intrinsic curvature of a manifold 217 Lie group 131
Intrinsic properties of a manifold 103 linear 130
Invariant integration 67 nonlinear 171
Invariant manifold 265 Lie group homomorphism 151
Invariant subgroup 9 Linear Lie group 130
320 Index
Local homomorphism of a Lie group 153 Null geodesic 190

Local isomorphism of a Lie group 153 Nullhomotopic curve 107
Locally attracting 252
Locally eventually onto 293
Local representation 83 Odd permutation 7
Logarithmic coordinates 136 Omega-limit point 281
Lorentz group 29 Omega-limit set of a motion 280, 281
restricted or proper 29, 30 Order of a group, subgroup, or element 4
full 29, 31 Orientability 210
Lorenz attractor 288, 289 Orthogonal group 25, 26
Lorenz graph 290
Lorenz system 286
Lowering and raising of indices 185 Parallel transport 209
Lowering and raising operators 52 Parity of a permutation 7
Lyapunov stability 285 Path in a manifold 104
Peixoto's Theorem 278,303, 306
Permutation
Manifold 96, 103 cyclic 6
covering 115, 197 even 7
Einstein 226 group 6
intrinsic curvature 217 odd 7
intrinsic properties 103 Picard method of iterations 192
invariant 265 Poincare group 73
of a group 34 Poincare mapping 256
pseudo-Riemannian 183 Point group 18
Riemannian 183 Positive definite 183
stable 263 Power spectrum 283
universal covering 119,121,123,125, Prehistories 299
197 Principle of equivalence 203
unstable 252 Principles of lifting 117
Maximal extension of an Einstein Projection 115
manifold 235 Proper or restricted Lorentz group 29, 30
Metric tensor 182 Proper subgroup 4
Mobius band 97 Pseudo- Riemannian manifold 183
Multiplication table 3 Pure Lorentz transformation 33
Multiplicity of a covering 115
Quasi-periodic function 276

Natural parameter on a geodesic 188,
196
Navier-Stokes equations 247 Raising and lowering of indices 185
Nilpotent Lie algebra 163 Raising and lowering operators 52
Nonaxisymmetric simple mode 273 Rank of a spin or 94
Nonlinear Lie group 171 Ray representation 81, 82
Normal coordinates 203 Ray space 82
Normal divisor 9 Reduced wave equation 221
Normal modes in hydrodynamic stability Reducible representation 41, 63
problems 249 Region of attraction 282
Normal subgroup 9 Regular representation 12, 49
n-tuply periodic function 16 Relations in a group 15
Index 321
Representation of a group 12 Simplicity of the rotation and Lorentz

decomposable 63 groups 38
equivalent 62 Simply connected 36, 104, 105
faithful 12, 41 Solvable Lie algebra 163
fully reducible 63 Space group 17
irreducible 41 Spacelike geodesic 190
left-regular 69 Special linear group 26
ray 81, 82 Special orthogonal group 26
reducible 41, 63 Special-relativity theory of
regular 12, 49 gravitation 225
of matrix groups 47 Spinor 93
right regular 69 dotted 94
spin 84 symmetric 95
two-valued 84 Spin representation 84
unitary 63 Stable manifold 253
Residue class 150 Star in a real Cartan subalgebra 165
Restricted or proper Lorentz group 29, Strange attractor 274, 283
30 Strongly generic 305
Reynolds number 245 Subalgebra 150
Riemannian covering manifold 197 Subcritical bifurcation 255, 257, 287
Riemannian manifold 183 Subgroup 4
Riemann tensor 211, 214 Subharmonic bifurcation 261
Right coset 10 Subrepresentation 41
Right-regular representation 69 Summation convention 30, 176
Right translation 12 Supercritical bifurcation 255, 257
Rigid motion groups 73 Surface harmonic 53
Rodrigues formula 56 Sylvester's law of inertia 183
Rotation group SO(3) 25, 50 Symmetric bilinear form 164
Rotation subgroup of the Lorentz Symmetric spinor 95
group 30 Symmetric top 72
Ruelle - Takens model of the onset of Symmorphic space group 20, 23
turbulence 279
Taylor problem 263

Scalar field on a manifold 175 Taylor vortices 246, 264
Schur's Lemma 65 Tensor field on a manifold 180
Schwarz schild charts 227 Tesseral harmonics 53
Schwarzschild radius 229 addition theorem 59
Schwarzschild line element 230 completeness 60
Self-conjugate subgroup 9 Thomas rotation 33
Semidirect product of groups 20, 22 Timelike geodesic 190
Semidirect sum of Lie algebras 160 Time reversal 236
Semi flow 248, 249 Topology of a manifold 199
Semisimple Lie algebra 163 Transitive 48
Separable coordinate systems 221 Translation 12
Separation of variables in the Taylor Translation group 17
problem 271 Translation of a chart 146
Signature 183 Transposition 7
Simple group 9 Trivial subgroup 4
Simple Lie algebra 150 Turbulence 276
322 Index
Two-point problem of a geodesic 193 Upper and lower manifolds of a

Two-sheeted covering 114 covering 114
Two-valued representation 84
Vector field on a manifold 177

Unimodular linear group 25
Unimodular orthogonal group 26
Unitary group 28 Wavy vortices in the Taylor
Unitary representation 63 problem 247, 273
Universal covering group 159 Weight, weight vector, weight
Universal covering manifold 119, 121, space 163, 164
123, 125 Whitehead's Theorem 194
Universal Riemannian covering Williams' analysis of the Lorenz
manifold 197 attractor 289
Unstable manifold 252 Word problem 16
Texts and Monographs in Physics
Edited by W. Beiglbock, M. Goldhaber, E. Lieb, and W. Thirring
Texts and Monographs in Physics includes books from any field of physics that might be used as
basic texts for advanced training and higher education in physics, especially for lectures and seminars
at the graduate level.
Polarized Electrons Principles of Advanced Mathematical Physics

by J. Kessler By R.D. Richtmyer
1976. ix, 223p. 104 illus. cloth Volume I
1978. xv, 40Op. 45 illus. cloth
Volume II
The Theory of Photons and Electrons 1981. xi, 322 p. 60 illus. cloth
The Relativistic Quantum Field Theory of
Charged Particles with Spin One- Half
Second Expanded Edition Foundations of Theoretical Mechanics
By J. Jauch and F. Rohrlich By R.M. Santilli
1976. xix, 553p. 55 illus. cloth Part I: The Inverse Problem in Newtonian
Mechanics
1978. 288p. cloth
Essential Relativity Part II: Generalizations of the Inverse Problem
Special, General, and Cosmological in Newtonian Mechanics
Revised Second Edition In preparation
By W. Rindler
1980. xvii, 284p. 44 illus. paper
Advanced Quantum Theory and Its
Applications Through Feynman Diagrams
Inverse Problems in Quantum Scattering By M.D. Scadron
Theory 1979. xiv, 386p. 78 illus. cloth
By K. Chadan and P. Sabatier
1977. xxii, 344p. 23 illus. cloth
Operator Aigebras and Quantum Statistical
Mechanics
Quantum Mechanics By O. Bratteli and D. W. Robinson
By A. Biihm Vol I: C and W Algebras. Symmetry Groups.
1979. xvii, 521p. 105 illus. cloth Decomposition of States
1979. 50Op. cloth.
Vol. II: Equilibrium States. Models in Quantum
Relativistic Particle Physics
Statistical Mechanics
By H. Pilkuhn
1981. 517p. approx. cloth
1979. xii, 427p. 89 illus. cloth
The Nuclear Many-Body Problem

The Concepts and Logic of Classical
By P. Ring and P. Schuck
Thermodynamics as a Theory of Heat 1980. 716p. approx. 171 illus. cloth
Engines
Rigourously Constructed upon the Foundation
Laid by S. Carnot and F. Reech Nuclear Reactions with Heavy Ions
By C. Truesdell and S. Bharatha By R. Bass
1977. xxii, 154p. 15 illus. cloth 1980. 41Op. 176 illus. 31 tables. cloth

9-Richtmyer - Principles of Advanced Mathematical Physics II

Uploaded by

Copyright:

Available Formats

9-Richtmyer - Principles of Advanced Mathematical Physics II

Uploaded by

Document Information

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

9-Richtmyer - Principles of Advanced Mathematical Physics II

Uploaded by

Copyright:

Available Formats

Texts and

Elliott H. Lieb Walter Thirring

Library of Congress Cataloging in Publication Data

(Texts and monographs in physics)

1981 by Springer-Verlag Inc.

All rights reserved. No part of this book may be translated or reproduced

ISBN 978-3-642-51078-6 ISBN 978-3-642-51076-2 (eBook)

18 Elementary Group Theory 1

19.4 The Lorentz groups 29

20 Group Representations I: Rotations and Spherical Harmonics 40

21 Group Representations II: General; Rigid Motions;

22 Group Representations and Quantum Mechanics 80

22.5 Local representations 83

23 Elementary Theory of Manifolds 96

24 Covering Manifolds 114

25 Lie Groups 129

26 Metric and Geodesics on a Manifold 174

27 Riemannian, Pseudo-Riemannian, and Affinely

28 The Extension of Einstein Manifolds 223

29 Bifurcations in Hydrodynamic Stability Problems 244

29.4 Hilbert space formulation 248

30 Invariant Manifolds in the Taylor Problem 263

31 The Early Onset ofTurbulence 276

Boulder, August 1981 Robert D. Richtmyer

Elementary Group Theory

The group axioms; Abelian group; cyclic group; subgroup; order;

Prerequisite: Elementary algebra.

This chapter contains a survey of elementary group theory. For applications

18.1 The Group Axioms; Examples

A group G is any set or collection of elements {a, b, c, ... , x, y, z, ... }, finite

I. If a and b are any two elements of G, then a b is an element of G.

11. If a, b, and c are any three elements of G, then (a 0 b) 0 c = a 0 (b 0 c)

If the elements are numbers, matrices, quaternions, etc., the composition

18.2 Elementary Consequences of the Axioms;

Extended associative law: (a (b (c d) e = a b cod e, etc. Un-

necessary parentheses will be omitted from now on. Also

If a is in G, and m is any integer, then am is defined as follows:

elements a and b of G are said to commute if a b = boa. If all pairs of

elements of G commute, G is said to be a commutative or Abelian group. If

the z-axis constitute a subgroup of the group of rotations in 3-space. The

MISCELLANEOUS QUESTIONS AND EXERCISES

(k) The set of all nonzero vectors in 3, under vector multiplication.

If there is a one-to-one mapping cp of a group G onto a group G' such that

is another isomorphism of G onto G'.

elo." (cos e -sin e\

An isomorphism of a group onto itself is an automorphism. An example

An automorphism of SO(2) onto itself is given by

is a homomorphism of G onto SO(2) as can be seen by performing two

18.4 Permutation Groups

A permutation is a one-to-one mapping of a set C (usually finite) of objects

Figure IS.1 A cyclic permutation.

this cycle is written in a still further condensed notation as (a bed e),

then the permutation (18.4-2) can be written as

let n: j --+ n(j) be a permutation of the integers 1, 2, ... , n, and call

If Xk - Xj is anyone of the factors in (18.4-4), then either Xk - Xj or Xj - Xk

that the transposition (11) = (21)(12)(2l) is necessarily odd, and finally

18.5 Homomorphisms; Normal Subgroups

Figure 18.2 A group homomorphism.

Theorem. Go and G'1 are subgroups of G and G' respectively; furthermore,