p154022_cumullect

Contents
1 CHAPTER I — TENSORS AND EXTERIOR CALCULUS ON MANIFOLDS 1

1.1 Vector Spaces and Linear Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Vector spaces in a nutshell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 The space dual to a vector space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Where Do Vectors Live? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Manifolds and coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Curves, directional derivatives and vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 The tangent space of a manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.4 Directional derivatives as vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.5 Differential of a function and basis dual to a coordinate basis . . . . . . . . . . . . . . . . 6
1.2.6 Transformations on bases, cobases, and components . . . . . . . . . . . . . . . . . . . . 6
1.3 At Last, Tensors! . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 The tensor product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Transposition, symmetric and skew-symmetric tensors . . . . . . . . . . . . . . . . . . . 9
1.3.3 Transformations on tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.4 The Levi-Civita symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Two More Ways to Construct Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.1 Contracted tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.2 Inner product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.3 The metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Exterior Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5.1 The exterior product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5.2 Oriented manifolds, pseudo-vectors, pseudo-forms and the volume form . . . . . . . . . . 16
1.5.3 The Levi-Civita pseudotensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.4 The Hodge dual of a p-form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6 Exterior Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6.1 Exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6.2 Laplace-de Rham operator, harmonic forms, and the Hodge decomposition . . . . . . . . 21
1.6.3 Exterior derivative and codifferential operator of a 2-form in Minkowski spacetime . . . . 22
1.7 Integrals of Differential (Pseudo)Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.7.1 Integrals of (pseudo) p-forms over a p-dim submanifold . . . . . . . . . . . . . . . . . . 23
1.7.2 Stokes-Cartan Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.8 Maxwell Differential Forms in 3 + 1 Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Appendices 26
A Tangent Spaces as Vector Spaces 26
B Transformation of Vector Components Between Coordinate Systems 27
C Three-dim Inhomogenous Maxwell Equations in the p-form Formalism 28
2 CHAPTER II — A BRIEF INTRODUCTION TO GROUP THEORY 29

2.1 Introducing the Notion of Group (BF 10.1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.1 Some basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.2 Cayley tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Special Subsets of a Group (BF10.3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.1 Special Ternary Compositions: Conjugacy Classes . . . . . . . . . . . . . . . . . . . . . 31
2.2.2 Subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Lecture Notes on Mathematical Methods 2022
2.2.3 Cosets (BF 10.3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.2.4 Lagrange’s Theorem and quotient groups . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.5 Direct Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 The Mother of All Finite Groups: the Group of Permutations . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Definitions, cycles, products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Some subgroups of Sn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.3 Cayley table of S3 as an example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.4 Cayley’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.5 Conjugates and Classes of Sn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.6 Graphical representation of classes: Young frames . . . . . . . . . . . . . . . . . . . . . 37
2.3.7 Cosets of Sn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Representations of Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.1 Action of a group from the left and from the right . . . . . . . . . . . . . . . . . . . . . . 38
2.4.2 Matrix representations of a group (BF10.4) . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.3 Non-unicity of group representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4.4 The regular representation of finite groups . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.5 Unitary representations (BF10.6) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.6 Invariant Spaces and Kronecker sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4.7 Reducible and irreducible representations (BF10.5) . . . . . . . . . . . . . . . . . . . . . 41
2.4.8 Exploring representations with Young diagrams . . . . . . . . . . . . . . . . . . . . . . . 42
2.5 Schur’s Lemmas and Symmetry in the Language of Group Theory (BF10.6) . . . . . . . . . . . . 44
2.5.1 What is a symmetry in the language of group theory? . . . . . . . . . . . . . . . . . . . . 44
2.5.2 Schur’s Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5.3 An orthogonality relation for the matrix elements of irreducible representations (BF10.6) . 45
2.5.4 Characters of a representation (BF10.7); first orthogonality relation for characters . . . . . 46
2.5.5 Multiplicity of irreducible representations and a sum rule for their dimension . . . . . . . 47
2.5.6 Another orthogonality relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.5.7 Character tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Appendices 52
D The Right and Left Actions of a Group on a Vector, with Sn as Example 52

D.0.1 Right action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
D.0.2 Left action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
E Proof of the Second orthogonality Relation for Characters 54
F Direct Product of Representations 54
G A Second Example of Symmetry-Breaking Lifting a Degeneracy 55
3 CHAPTER III — LIE GROUPS 57

3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 Some Matrix Lie Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.1 Bilinear or quadratic constraints: the metric (or distance)-preserving groups . . . . . . . . 58
3.2.2 Multilinear constraints: the special linear groups . . . . . . . . . . . . . . . . . . . . . . 59
3.2.3 Groups of transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2.4 Differential-operator realisation of groups of transformations: infinitesimal generators . . 60
3.2.5 Infinitesimal generators of matrix Lie groups . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Lie Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3.1 Linearisation of a Lie group product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2
3.3.2 Definition of a Lie algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.3.3 Structure constants of a Lie algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3.4 A direct way of finding Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3.5 Hard-nosed questions about the exponential map — the fine print . . . . . . . . . . . . . 67
3.4 Representations of Lie Groups and Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4.1 Representations of Lie Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4.2 Representations of Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.4.3 The regular (adjoint) representation and the classification of Lie algebras . . . . . . . . . 68
3.4.4 The Cartan-Killing form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.4.5 Cartan subalgebra of a semismple algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.5 Weights and Roots of a Representation of a Compact Semisimple Algebra . . . . . . . . . . . . . 72
3.5.1 Properties of eigengenerators in the Cartan-Weyl basis . . . . . . . . . . . . . . . . . . . 73
3.6 Irreducible representations of semisimple algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6.1 Casimir invariant operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6.2 Irreducible representations of so(3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6.3 Irreducible representations of su(2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6.4 Irreducible epresentations of SU(2) and SO(3) . . . . . . . . . . . . . . . . . . . . . . . 76
3.6.5 su(2) substructure of a semisimple algebra and constraints on its root system . . . . . . . 77
3.7 More on finding irreducible representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.7.1 Tensor product representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.7.2 Irreducible tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.7.3 The Wigner-Eckart theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.7.4 Decomposing product representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Appendices 81
H Commutators of Angular Momentum with Vector Operators 81
I Alternative Derivation of the Master Formula 82
4 CHAPTER IV — Solution of Differential Equations with Green Functions 83

4.1 One-dimensional Linear Differential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.1.1 Existence of the inverse of a linear differential operator . . . . . . . . . . . . . . . . . . . 84
4.1.2 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.1.3 First-order linear ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1.4 Second-order linear ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1.5 Second-order IVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.6 Second-order BVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 Solving One-dimensional Second-order Equations with Green Functions (BF 7.3) . . . . . . . . . 86
4.2.1 Solutions in terms of Green Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2.2 1-dim Green Functions without boundary conditions . . . . . . . . . . . . . . . . . . . . 87
4.3 Green functions for the IVP and the BVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.1 Green’s second 1-dim identity and general solution of a BVP in terms of Green functions . 91
4.4 Differential Equations with Partial Derivatives (PDE) . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5 Separation of Variables in Elliptic Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5.1 An Important and Useful 3-dim Differential Operator . . . . . . . . . . . . . . . . . . . . 92
4.5.2 Eigenvalues of L2 and Lz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5.3 Eigenfunctions of L2 and Lz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5.4 General Solution of a Spherically-Symmetric, 2nd-order, Homogeneous, Linear Equation 94
4.6 Second 3-dim Green Identity, or Green’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.7 3-dim Boundary Value (Elliptic) Problems with Green Functions . . . . . . . . . . . . . . . . . . 97
3
4.7.1 Dirichlet and Neumann Boundary Conditions for an Elliptic Problem . . . . . . . . . . . 97

4.7.2 Green function for the 3-d Elliptic Helmholtz operator without boundary conditions . . . . 98
4.7.3 Dirichlet Green function for the Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.7.4 An important expansion for Green’s Functions in Spherical Coordinates . . . . . . . . . . 101
4.7.5 An Elliptic Problem with a Twist: the Time-independent Schrödinger Equation . . . . . . 103
4.8 A Hyperbolic Problem: the d’Alembertian Operator . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.9 Initial Value Problem with Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.9.1 Second-order Cauchy problem using transverse/longitudinal projections . . . . . . . . . . 105
4.9.2 First-order Cauchy problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Appendices 107
J Solving an Inhomogeneous Equation in Terms of Homogeneous Solutions 107
K Solution of a Homogeneous IVP with Homogeneous B.C. 108
L Modified Green Functions for the One-dim Boundary-value Problem 109
M Counting Electromagnetic Degrees of Freedom in the Lorenz Gauge 110
4
1 CHAPTER I — TENSORS AND EXTERIOR CALCULUS ON MANIFOLDS

This module starts by emphasising a view of vectors as objects that can be discussed without explicit reference
to a basis (or cobasis), as is implicit in the elementary notation u, or ~u. An alternative description in terms of
components in a so-called cobasis, or dual basis will be introduced and its meaning explored. Our powerful
geometric approach will then allow a conceptually simple generalisation of vectors to tensors. While everyone
agrees that vectors are indispensable to the mathematical description of many physical quantities, the equally great
importance of tensors is not always fully appreciated. For example, it is difficult to understand electromagnetism if
one insists on regarding the electric and magnetic fields as just two vector fields connected by Maxwell equations,
instead of the six non-zero components of the rank-2, antisymmetric Faraday tensor, or 2-form, F† . The need to
describe how vectors and p-forms change in time and space will lead to the exterior derivative, of which gradient,
divergence and curl are but special cases. We will also see that differential p-forms in fact are the only objects
that can be meaningfully integrated. The concise language of p-forms can illuminate many other areas of physics,
such as mechanics, thermodynamics, general relativity and quantum field theory.
1.1 Vector Spaces and Linear Mappings

1.1.1 Vector spaces in a nutshell
Definition 1.1. A vector space V over a field F is a (possibly infinite) set of objects on which an
operation called addition and another called s-multiplication (multiplication by a scalar) are defined,
and that is closed under these operations. Therefore, any two elements u and v of V satisfy:
(a + b)(u + v) = (a u + a v + b u + b v) ∈ V
∀ a, b ∈ F; in what follows, F = R. Both the addition and s-multiplication operations are commuta-
tive and associative.. We will call elements of a vector space vectors.
Example 1.1. Rn , the set of all ordered n-tuples of real numbers, with addition defined as adding
entries with the same place in the n-tuple, and s-multiplication by λ defined as multiplying each entry
by λ, is perhaps the best-known and most important vector space.
Definition 1.2. If any v ∈ V can be written as a linear combination‡ :

n<∞
X
v = v α eα ≡ v α eα summation over repeated indices implied! (1.1)
α
of a set {eα ∈ V}, then that set is said to span, or to be a set of generators of, the vector space V.
If, furthermore, this set is linearly independent, in the sense that v = 0 =⇒ v α = 0, then it is a
basis of V. The number n of vectors in a basis defines the dimension of V, and we often write V n .
The (real, and unique!) coefficients v α are called the contravariant components of the vector v in
this basis. This one-to-one correspondence between V n and Rn can be represented by a n × 1 matrix:
 1 Warning ! v and its components are different beasts and
v
v2  should never be confused. Byron and Fuller (BF) do not
 
v 7−→  .  make this distinction clear enough. Also, the index on eα
 .. 
identifies the vector, not a component of the vector.
vn
Example 1.2. The standard, or natural, basis Rn is the set {eα } (α = 1, 2, . . . , n), where each
n-tuple labelled by a value of α has 1 in the αth position and 0 in all other positions.
†
These notes generally follow the conventions set by the ISO (International Standards Organisation) for mathematical typography, with
one important exception: as in Byron and Fuller, vectors and tensors are in bold upright (u) instead of bold italic font (u). Sans-serif fonts
denote matrices, eg. M.
‡
Infinite linear combinations (series) require extra topological structure on V so as to allow the notion of convergence.
1
1.1.2 The space dual to a vector space

Let V and W be two vector spaces that share the same field of scalars. We shall be interested in the set of all linear
mappings, L(V, W) := {T : V → W}, such that, ∀ Ti ∈ L(V, W):
(a Ti + Tj )(V) = a Ti (V) + Tj (V) ∈W a Ti + Tj ∈ L(V, W)
where the operations on the right are defined on W. One can define linear mappings on L (V, W), ie., one can
compose linear mappings into a linear map.
An important subset of the set of linear mappings is L(V, R) that contains all linear, real-valued functions on
a vector space. We say that it forms a space V ∗ dual to V. Since L(V m , W n ) has dimension m × n, V ∗ and V
have the same† dimension. The elements of V ∗ are called covectors, or linear functionals (in linear algebra), or
1-forms. One example would be definite integrals on the vector space of polynomials.
There comes following important definition:
Definition 1.3. If {eα } is a basis of a space V n , its unique dual basis (cobasis) in V ∗ , {ω α }, satisfies:
ω α (eβ ) = δαβ (α, β) = 1, . . . , n (1.2)
where δαβ is the Kronecker delta. The left-hand side is just classic multiplication of row-vectors with
column-vectors. Then a covector σ ∈ V ∗ is written σ = σα ω α , where the σα are the covariant
components of σ in this dual basis.
From this we derive the action of an element ω α of the cobasis of V ∗ on a vector v ∈ V:
ω α (v) = ω α (v β eβ ) = v β ω α (eβ ) = v β δαβ = v α

We conclude that the cobasis element ω α projects out, or picks out, the corresponding component of v. This
will probably come as some surprise to many, who are used to think of v α as the projection of v on eα .
What happens if we act on some eα with a 1-form (covector) σ = σβ ω β ? Well,
σ(eα ) = σβ ω β (eα ) = σβ δβα = σα
Do keep in mind that indices on a bold-character object will always label the object itself, not its components,
which will never be bold. Also, in σ = σβ ω β as well as in v = v ν eν , the left-hand side is explicitly basis-
independent; this notation we shall call index-free, or geometric. The right-hand side, in so-called index notation,
makes explicit reference to a basis even though, taken as a whole, it is still basis-independent. Both notations have
advantages and disadvantages to be discussed later. Fluency in both is highly recommended.
Recall the one-to-one correspondence between a vector v and the n-tuple of its components in a basis {eα },
(v 1 , . . . , v n ) ∈ Rn . An analog correspondence exists between a 1-form, σ, and its components σα :
T
v 7−→ v 1 v 2 . . . v n σ 7−→ σ1 σ2 . . . σn
Therefore, we can also think of σ as as a procedure to obtain the number σ(v) = σα v α out of the vector v via
standard multiplication of a row vector with components σα by a column vector with components vβ :
T T
v1 v2 . . . vn 7−→ σ1 σ2 . . . σn v 1 v 2 . . . v n = σα v α (1.3)
σ
Since L(V, R), or V ∗ , is a vector space, it has its own dual space, L(V ∗ , R), or V ∗∗ . We realise that nothing
prevents us (with finite-dimensional spaces) from considering the elements v ∈ V as themselves linear mappings
on V ∗ , and identifying V ∗∗ with V! Then eα (ω β ) = δα β , and v(σ) = v α σα , exactly as in eq. (1.3) above.
These considerations suggest that just like a vector, we can view a 1-form (covector), as a kind of machine†
but with a vector as input and a number as output. The following tables summarise these results:
†
This assumes that V’s dimension is finite!
†
So far as I know, this metaphor was first proposed by Misner, Thorne and Wheeler (MTW) in their monumental textbook, Gravitation.
2
1-form Input vector Output Vector Input 1-form Output

Cobasis ω α Basis eβ ω α (eβ ) = δαβ Basis eα Cobasis ω β eβ (ω α ) = δαβ
Cobasis ω α v ω α (v) = v α Basis eα σ eα (σ) = σα
σ Basis eα σ(eα ) = σα v Cobasis ω α v(ω α ) = v α
σ v σ(v) = σα v α v σ v(σ) = v α σα
Note that σ(v) = v(σ) = σα v α is basis-independent, but only if σ is referred to the cobasis of the basis in
which v is written. At this stage, there is no unique connection between σ and a vector in V. So, tempting as it is
to identify σα v α with the scalar product of two vectors, let us resist that urge.
For a given n-dimensional vector v, there exists a unique set of parallel (n-1)-dimensional hyperplanes that
can provide a pictorial representation of 1-forms. This is easiest when n = 2. Then a = σ1 v 1 + σ2 v 2 determines a
perpendicular to v with equation σ2 = a/v 2 − σ1 v 1 /v 2 . The lines generated by different a all have slope −v 1 /v 2 .
1.2 Where Do Vectors Live?

The obvious answer has to be: in a vector space! Therefore, we should learn how to identify (or construct) such
vector spaces. This is anything but trivial: not all spaces can be equipped with a vector-space structure, and we
shall see that this happens with as familiar a space as spacetime.
1.2.1 Manifolds and coordinates
Definition 1.4. A differentiable manifold, or just manifold, M is a set of elements (“points”), all of
which have an open ball (or neighbourhood) around them in M , such that M can be entirely covered
by a union of possibly overlapping open (without boundary) subsets Ui , each mapped in a one-to-one
way to an open subset of Rn by a non-unique, differentiable coordinate map: x : Ui −→ Rn .
Each (Ui , x) forms a coordinate chart (local coordinate system), and an atlas is any collection of
charts that covers the whole M . Also, on any overlap Ui ∩ Uj ⊂ M , only charts (Ui , x) and (Uj , y)
for which the coordinate transformation y ◦ x−1 : Rn −→ Ui ∩ Uj −→ n
y R between them is (once)
x−1
differentiable are allowed.
The minimum number n of parameters—each a map xi : U −→ R (1 ≤ k ≤ n)—that uniquely
specify every point in U is its dimension.
Example 1.3. • Rn can be promoted to a manifold; it can be covered with just one coordinate chart, Cartesian
(standard) coordinates. Other charts are possible, eg. polar coordinates. to cover the manifold.
• A conical surface, even a semi-infinite one, can never be a manifold because of its tip.
• A vector space V can be made into a manifold that can be covered with one chart (V, Φ), where Φ maps
elements of V to their components in Rn in that basis. Conversely, however, a manifold is not in general a
vector space! On Earth’s surface, there is no meaning to adding the position of Toronto to that of London.
• EvenPthough Rn can be endowed with a manifold structure, a unit ball in Rn , defined

P 2in Cartesian coordinates
2
by xi ≤ 1, is not a manifold because it has an edge; but the open unit ball, xi < 1, is a manifold. So
n+1
X
is the unit sphere, S n , defined by x2i = 1 and embedded in Rn+1 .
The unit circle in the plane R2 , S 1 , and the 2-dim sphere in R3 . S 2 , are archetypal examples of (closed)
curves in R2 and (closed) surfaces in R3 .
S 1 being
√ a 1-dim manifold, we wish to build an atlas for it. One way of doing this is with two open patches,
y = ± 1 − x2 with x = ±1 excluded (why?), and the +/− sign corresponding to the submanifold in the
3
upper/lower half-plane. Then each point of any of the two submanifolds is in one-to-one correspondence
with some x ∈ R, with x < 1. To cover all of S 1 , we repeat with two submanifolds in correspondence with
x > 0 and x < 0, and an atlas with four charts has been constructed.
p
S 1 also has a local coordinate, θ, related to x by: θ = tan−1 (y/x) = tan−1 1/x2 − 1 . To avoid any
point being mapped to more than one value, θ must map to [0, 2π) in R .
An atlas can also be constructed for S 2 out of patches similar to those for S 1 for each of the Cartesian
±x > 0, ±y > 0 , and ±z > 0. Each point in each patch unambiguously maps to R2 .
On S 2 we could also use spherical coordinates θ and φ that map to he region of R2 : θ ∈ (0, π), φ ∈ [0, 2π),
with the poles removed since φ is undetermined there. More patches are needed to cover S 2 .
Notice that we have looked at S 1 and S 2 as being embedded in a higher-dimensional manifold, R2 and R3 .
Whitney’s embedding theorems guarantee that any smooth M n is a submanifold of Rm>2n , with stronger
results in restricted cases. Embedding curves and surfaces in, eg., R3 is great for visualisation, but we are
more interested in their intrinsic properties which should be independent of the embedding manifold.
Less technically, it is usually enough to view a manifold as a set that can be parametrised in a smooth way.
1.2.2 Curves, directional derivatives and vectors

The naı̈ve notion of a vector as a straight arrow from one point to another in Rn cannot be extended to arbitrary
manifolds M , on which straightness will in general have no well-defined meaning (think of straight arrows on a
sphere). Manifolds are not vector spaces; so where do vectors at a point in M actually live? And is it possible to
think of a vector as a local object that involves only that point, independent of any coordinate chart?
Definition 1.5. A curve Γ : R −→ M on a manifold M is a mapping, at least C 1 (no kinks!), of
each value of a real parameter λ to a unique point P in M : Γ(λ) = P. Then λ is a coordinate on Γ.
Definition 1.6. Now introduce the vector space, C ∞ (M ) := { f : M −→ R }, of all smooth,
real-valued functions f that map a point in M , such that, ∀ (f, g) ∈ C ∞ (M ) and a ∈ R:
[f + g](P) = f (P) + g(P) [a f ](P) = a f (P) (1.4)
where addition and multiplication within square brackets are operations on C ∞ (M ), while those on
the right-hand side are on R. The composition f ◦ Γ : R−→ M −→ R is equivalent to f (λ), with λ ∈ R.
Γ f
Definition 1.7. Let Γ be a curve parametrised by a coordinate λ. The velocity at a point P with
coordinate λ0 on this curve, is the linear map v(Γ,P ) : C ∞ (M ) −→ R, defined as:

v(Γ,P ) (f ) := dλ (f ◦ Γ) λ0 = dλ f λ0 dλ := d/dλ (1.5)
Such a curve is only one of an infinite number containing P each with their own velocity at P. For
instance: wΘ (f ) = dµ f µ0 , and Θ(µ0 ) = P. We say that velocities are tangent to the manifold at P.
We can also view the curve Γ as embedded in a region U of a manifold M parametrised by coordinate functions
denoted collectively by x : U −→ Rn , with x ◦ Γ describing what the curve “looks like” in M . Then an alternate
expression for the velocity is, from definition 1.7:
n h
X i
−1

v(Γ,P) (f ) = dλ (f ◦ x ) ◦ (x ◦ Γ) λ0 = dλ (xν ◦ Γ) λ0
∂ν (f ◦ x −1
)
xν (P)
ν
X
=⇒ v(Γ,P) = d λ xν 0
(∂ν )P (1.6)
ν
where the index ν in the multidimensional chain rule runs over the number of local coordinates that specify each
point in M . The xν (λ) parametrise the curve Γ in M . Since f ◦ x−1 maps Rn to R, its derivatives behave like the
standard ∂xν f (xµ ), that is: ∂xν f P := ∂ν (f ◦ x−1 ) P .
We will interpret this important result a little later, after we have constructed the space where v(Γ,P) lives.
4
1.2.3 The tangent space of a manifold
Definition 1.8. A tangent space TP to a manifold M n at a point P ∈ M n , is a set that can be

equipped with a vector-space structure, consisting of all the velocity vectors vP tangent to M n at P,
In fact, all vectors defined on M n at P live in TP , not in M n . TP = Rn . Strictly speaking, P should
be the origin of TP so that the space contains the zero vector.
Similarly, one defines a cotangent space TP∗ containing all the covectors on M n at P.
This definition rests on a very bold assertion, namely, that the velocities as defined above (definition 1.7) are in
fact vectors. For those who are interested, a proof that TP can be made into a vector space is in Appendix A.
1.2.4 Directional derivatives as vectors

To interpret eq. (1.6), we introduce a different perspective on vectors. We notice that v(Γ,P) (f ) = dλ xν 0 ∂ν (f ) P
looks like the directional derivative of f in the direction of v, which in basic calculus is written ∂v f := v · ∇f .
Then the velocity vector has components v ν = dλ xν . This motivates us to identify any tangent vector t at P with
the directional derivative at P in the direction of t. Thus:
Definition 1.9. Given an arbitrary differentiable function f on a manifold M n , parametrised in a local

coordinate system by f (x1 , . . . , xn ), then the action of a vector t ∈ TP on f at a point P is defined
as:
t(f ) := ∂t f = tν ∂ν f (1.7)
P P
Since the ∂ν in eq. (1.6) span the tangent space of M at P, it would be natural to think of them as basis vectors.
But are they linearly independent? Take f = xν , the coordinates on M ; then aµ ∂xµ xν P = aµ ∂µ (xν ◦ x−1 ) P =
aµ δµν = aν , where aν ∈ R. If aµ ∂xµ xν P = 0, aν = 0, which shows that the ∂ν do form a basis of the tangent
space. Thus:
Definition 1.10. The tangent space TP to a manifold M n admits a basis {∂ν P

} (ν = 1, . . . , n) called
the coordinate basis for the n local coordinates xν that parametrise M n .
To find the coordinate-basis tangent vectors, we freeze all the variables that parametrise the manifold, except
one that is varied to generate a coordinate curve whose tangent vector at a point P is the partial derivative with
respect to the parameter λ on which the coordinates xµ depend. Then the components of ∂µ at P are the derivatives
of x ∈ Rn with respect to the parameters of the manifold at P. An example should make this procedure clearer:
Example 1.4. On S 2 (embedded in R3 ), a point P is mapped into the spherical coordinates (θ, ϕ),
with θ 6= (0, π); P can also be described by the R3 coordinates (sin θ cos ϕ, sin θ sin ϕ, cos θ).
Freezing say, θ, generates a great circle on the sphere. Then these coordinates describe a circle of
radius sin θ at “colatitude” θ, and ∂ϕ is a coordinate-basis vector visualised in R3 by the vector with
components:
∂ϕ (sin θ cos ϕ, sin θ sin ϕ, cos θ) = (− sin θ sin ϕ, sin θ cos ϕ, 0)
At each point on S 2 parametrised by (θ, ϕ), this is a vector tangent to the circle at colatitude θ.
Similarly, there is a spherical-coordinate-basis vector, ∂θ , tangent to a meridian going through that
same P, with components (cos θ cos ϕ, cos θ sin ϕ, − sin θ). ∂θ and ∂ϕ together form a basis for
vectors in the plane tangent to S 2 at P. These vectors do not live in S 2 ! Instead, any vector at a point
on S 2 lives in the R2 plane tangent to the point. Each point on S 2 has its own tangent plane.
Also, notice that ∂ϕ is nor normalised to 1, except at θ = π/2. In general, coordinate bases and
cobases are not normalised. But ∂ϕ̂ := sin1 θ ∂ϕ has components (− sin ϕ, cos ϕ, 0), which are the
components of the unit basis vector ϕ̂ in the standard basis; it is an element of a non-coordinate basis.
5
1.2.5 Differential of a function and basis dual to a coordinate basis
Definition 1.11. Let xµ ∈ Rn (µ = 1, . . . , n) be the coordinate functions of arbitrary coordinates at

a point P ∈ M n , and f ∈ C ∞ (M ) a real-valued differentiable function on M n . Let also t ∈ TP be a
vector tangent to M n at P. Then the differential of f at P, df , is defined as the 1-form in TP∗ which,
when t is inserted in its input slot, yields the basis-independent action of t on f at P from eq. (1.7):
[df ](t) := t(f ) = ∂t f (1.8)
To find the components of df in the coordinate cobasis dual to the coordinate basis {∂µ } at a point P ∈ TP , recall
that the action of a 1-form (covector) on a coordinate-basis vector ∂ν outputs the corresponding component of the
1-form in that (as yet undetermined) cobasis: [df ]ν = df (∂ν ), which, from eq. (1.8), is the ordinary derivative of
f in the direction of the basis vector ∂ν , so ∂ν f . Now, taking f = xµ , the same argument immediately leads to:
[dxµ ] (∂ν ) = ∂ν xµ = δµν (1.9)
which we recognise as the defining equation (1.2) for a cobasis, with eµ = ∂µ and ω µ = dxµ . Then, choosing
{∂µ } as basis for TP , we conclude that {dxµ } is the basis, dual to {∂µ }, of the cotangent space, TP∗ , dual to TP .
When written in a coordinate cobasis, σ = σα dxα is called a differential form.
If we think of f as a 0-form, the differential of f is the gradient 1-form df :
df = ∂µ f dxµ (1.10)
We recognise the well-known expression for the differential of a function in calculus, where it is taken to be a
scalar, a number. But df , interpreted as the infinitesimal change of f does not know in which direction this change
should be evaluated. Only when a vector is inserted in its input slot, as in eq. (1.8), can it output a number, the
change of f in the direction of the vector.
As for the usual calculus interpretation of dxµ as the difference between the components of two coordinate
vectors at infinitesimally close points, this is not valid on an arbitrary manifold, since dxµ , like all 1-forms at
a point, lives in the cotangent space, not the manifold, Only in Rn can one ignore with impunity this crucial
distinction between a base manifold and its tangent and cotangent spaces at a point.
1.2.6 Transformations on bases, cobases, and components

Let (U1 , x) and (U2 , y) be two overlapping charts (see definition 1.4) on a manifold, with x and y their coordinates.
At a point P ∈ U1 ∩ U2 , the transformation on the components of a vector v is shown in Appendix B to be:
vyν = ∂xµ y ν xP
vxµ (1.11)
What is remarkable about this transformation is that it is linear and homogeneous, even though the transforma-
tions beween (U1 , x) and (U2 , y) can be non-linear. Thus, in coordinate bases, the coefficients, ∂xµ y ν , in the
transformation law are the entries of the Jacobian matrix of the transformation evaluated at P. Conversely, if
vxν = vyµ ∂yµ xν , one shows, using the chain rule on partial derivatives, that v is unchanged by the transformation.
In general bases, transformations on components must be assumed homogeneous and linear, and take the form:
′ ′ ′
v α = v µ Lα µ = Lα µ v µ (1.12)
where the prime refers to the y coordinates in (1.11). This is the more traditional definition of a vector still in use
′
in physics. These two ways of writing v α are equivalent, but the second one is a matrix product. Bases transform
as:
′ ′
eµ = Lαµ eα′ = eα′ Lαµ (1.13)
The second expression in the equation, however, is not matrix multiplication, because the subscript of the basis
vector is a label for the vector, not for a component of this vector.
6
L being non-singular, and therefore invertible, the action of the inverse transformation L−1 is represented by:
′
v = L−1 v′ ⇐⇒ v µ = (L−1 )µν ′ v ν (1.14)
Do not confuse matrix and index notation! Whereas matrix notation is readily translated into index notation, the
reverse generally requires some rearrangement. This is because index notation does not care about ordering—one
of its virtues—but matrix notation most certainly does.
′ ′
Let {eµ } and {eµ′ } be two bases in V n , connected by eµ = eν ′ Lν µ , where the Lν µ are the coefficients of
′
the matrix L representing a linear transformation L. Let {ω α } and {ω α } be their two respective cobases in V ∗ .
′
Then, writing ω α = M αβ ′ ω β where the M α β ′ are the matrix coefficients of the corresponding transformation
′
M between the cobases, it can be shown (EXERCISE) that M is the inverse of L, ie. M αν ′ Lν β = δαβ in index
′ ′
notation and M = L−1 in matrix notation. This means that: ω α = Lα β ω β .
In the same way as for vector components, we can then obtain (EXERCISE) the transformation law of the
′
components σα of a 1-form σ. Since σ must be cobasis-independent, σα ω α = σβ ′ ω β yields:
σα′ = σµ (L−1 )µ α′ (1.15)
while the inverse matrix, M−1 = L, takes the components in the opposite direction.
The following table summarises all the possible transformations, both in general and in coordinate bases:
Care should be exercised when com-
′ ′
paring this table to the expressions
eα′ = eβ (L−1 )β α′ = eβ ∂α′ xβ eα = eβ ′ Lβ α = eβ ′ ∂α xβ given in §2.9 and in Box 8.4 of MTW
α′ α′ β α′ β α −1 α β ′ α β ′
v = L β v = ∂β x v v = (L ) β ′ v = ∂β x v ′ which refer to Lorentz transforma-
′ ′ ′ ′ ′
tions. In their potentially confus-
ω α = Lα β ω β = ∂β xα v β ω α = (L−1 )α β ′ ω β = ∂β ′ xα ω β ing but standard notation, the matrix
−1
σα′ = σβ (L ) α′β β ′
σα = σβ ′ L α = σβ ′ ∂α x β ′
with elements Lαβ ′ is actually the
′ inverse of the matrix with elements
σα v α = σβ ′ v β ′
Lβ α ; we prefer making this explicit
by writing (L−1 )αβ ′ .
Another word of caution: transformations in coordinate bases may well produce components in non-normalised
bases, even if one starts from a basis that happens to be normalised. This does not occur in the case of rotations
and Lorentz boosts, but it will when we transform from Cartesian to curvilinear coordinates.
Also, in a coordinate basis, we cannot call ∂µ f the components of the gradient vector, ∇f . They do not trans-
form as vector components, as can be seen by calculating ∂µ′ f in terms of ∂ν f using the chain rule (EXERCISE).
1.3 At Last, Tensors!

Our previous discussions make it straightforward to extend the concept of linear mappings to that of multilinear
mappings, ie. mappings which are linear in each of their arguments, with the other arguments held fixed.
With V and its dual space V ∗ , equipped respectively with coordinate basis {∂νi } and cobasis {dxµi } (1 ≤ i ≤
n), we construct the space of multilinear mappings, {T : V ∗ × . . . × V ∗ × V × . . . × V −→ R }:
Definition 1.12. Contravariant tensors T ∈ T r of type (r, 0) are multilinear functions of r 1-forms:
T(σ1 , . . . , σr ) = σµ1 . . . σµr T(dxµ1 , . . . , dxµr ) = T µ1 ...µr σµ1 . . . σµr ∈ R (1.16)
Covariant tensors S ∈ Ts of type (0, s)s are real multilinear functions of s vectors:
S(u1 , . . . , us ) = uν1 . . . uνs S(∂ν1 , . . . , ∂νs ) = Sν1 ...νs uν1 . . . uνs (1.17)
Mixed tansors Q ∈ Tsr of type (r, s) are real multilinear functions of r covectors and s vectors:
Q(σ1 , . . . , σr , u1 , . . . , us ) = σµ1 . . . σµr uν1 . . . uνs Q(dxµ1 , . . . , dxµr , ∂ν1 , . . . , ∂νs )
= Qµ1 ...µr ν1 ...νs σµ1 . . . σµr uν1 . . . uνs (1.18)
7
T µ1 ...µr , Sν1 ...νs , and Qµ1 ...µr ν1 ...νs are the contravariant components of T, the covariant compo-
nents of S, and the mixed components of Q, respectively, in the chosen basis and cobasis.
Following the metaphor of tensors as machines, to output a number from a (r, s) tensor, one must
supply r 1-forms and s vectors as inputs, one for each slot.
1.3.1 The tensor product

There is an important kind of multilinear mapping we can construct, this time out of known building blocks.
Definition 1.13. The Kronecker (tensor) product space of V1∗ and V2∗ is a set of bilinear mappings
L(V1 , V2 , R), denoted by V1∗ × V2∗ , with as product elements the covariant tensor σ ⊗ τ :
[σ ⊗ τ ](u, v) = σ(u) τ (v) (1.19)
for all u ∈ V1 , v ∈ V2 , σ ∈ V1∗ , and τ ∈ V2∗ .
The product space L(V1∗ , V2∗ , R) = V1 × V2 , has as elements the contravariant tensor of rank 2:
[u ⊗ v](σ, τ ) = u(σ) v(τ ) (1.20)
There are tensor product spaces V1 ⊗ V2∗ with elements u ⊗ σ(τ , v) = u(τ )σ(v), and V1∗ × V2 , with
elements σ⊗v(u, τ ) = σ(u)v(τ ). When a tensor is a tensor product, we say that it is decomposable.
It is important to note that the tensor product is not commutative!
Example 1.5. Let P be the vector space whose elements are polynomials of some degree n. Such
a space can be constructed provided we define addition
R 1 and multiplication of polynomials. Then we
can construct a map T : P × P → R defined by 0 p(x) q(x) dx, where (p, q) ∈ P . This bilinear
map—call it the inner product— is a (0, 2) tensor with two vectors as inputs and a number as output.
Now take V1 = V2 = V. If {∂µ } is a coordinate basis for V, then {∂µ ⊗ ∂ν } is a coordinate basis for V ⊗ V.
Similarly, if {dxα } is a coordinate basis for V ∗ , then {dxα ⊗ dxβ } is a coordinate basis for V ∗ ⊗ V ∗ .
We assert that any contravariant (2, 0) tensor lives in V × V, and any covariant (0, 2) tensor lives in V ∗ × V ∗ :
A = Aµν ∂µ ⊗ ∂ν , B = Bαβ dxα ⊗ dxβ (1.21)
Therefore, the action of A on pairs of one-forms, and of B on pairs of vectors, is given by:
A(σ, τ ) = Aµν ∂µ ⊗ ∂ν (σ, τ ) = Aµν ∂µ (σ) ∂ν (τ ) = Aµν σµ τν
(1.22)
B(u, v) = Bαβ dxα ⊗ dxβ (u, v) = Bαβ dxα (u) dxβ (v) = Bαβ uα v β
As we have said before, both A and B can be viewed as operators, or devices, requiring two 1-forms or two
vectors, respectively, as ordered input, to output a product of numbers. But we can also input a single vector
(1-form) and obtain a 1-form (vector) as output, so long as we specify into which of the two input slots it should
be inserted. For instance, we could write B(u, ), or B( , u):

B(u, ) = Bαβ dxα (u) dxβ = Bαβ uα dxβ = σβ dxβ = σ

B( , u) = Bαβ dxα dxβ (u) = Bαβ uβ dxα = τα dxα = τ
Unless the Bαβ happen to be symmetric in their indices, the two resulting 1-forms σ and τ are not the same!
More generally, ∂µ1 ⊗ · · · ⊗ ∂µr ⊗ dxν1 ⊗ · · · ⊗ dxνs forms a basis for Tsr . Then any tensor can be written:
T = T µ1 ...µr ν1 ...νs ∂µ1 ⊗ · · · ⊗ ∂µr ⊗ dxν1 ⊗ · · · ⊗ dxνs (1.23)
To obtain a number, all input slots must be filled; but, as we saw for rank-2 tensors, we can also input one less
vector and get a 1-form as output, or one less 1-form to get a vector. More generally, inserting m input vectors in
a (r, s) tensor outputs a (r, s−m) tensor; inserting q input 1-forms outputs a (r−q, s) tensor.
It is important to remember that interchanging vectors or 1-forms in the input may result in different output.
8
1.3.2 Transposition, symmetric and skew-symmetric tensors

Interchanging any two contravariant or any two covariant slots of a tensor produces a transpose of this tensor.
Strictly speaking, interchanging a covariant and a contravariant slot of a tensor does not make sense.
Definition 1.14. A tensor that remains unchanged under transposition of two of its input slots of the
same type is said to be symmetric in these slots. Its components are unchanged under permutation of
indices corresponding to those slots. If it switches sign under transposition, we say that it is antisym-
metric in these slots, and the components corresponding to these slots also switch sign. Inserting the
same 1-form (in the contravariant slots) or vector (in the covariant slots outputs zero.
Symmetry and antisymmetry are basis-independent properties.
Example 1.6. Take an antisymmetric (0, 2) tensor: F = F[µν] dxµ ⊗ dxν , where square brackets
mean that indices are antisymmetric. Then F(u, u) = Fµν uµ uν = Fνµ uν uµ = −Fµν uµ uν = 0.
Among important tensors are those whose components are completely symmetric in all their covariant (or
contravariant) indices, and those which are completely antisymmetric (skew-symmetric, alternating) in all their
covariant (or contravariant) indices.
A completely symmetric tensor of rank r in n dimensions has n+r−1 r = (n + r − 1)!/(n − 1)!r! independent
components. A skew-symmetric tensor has nr = n!/(n − r)!r! independent non-zero components.
In three dimensions, many physically relevant tensors are symmetric, eg. examples 1.6, 1.7 and 1.8 (moment
of inertia, electrical polarisation, multipole moment) in B&F, as well as the Maxwell stress tensor. Antisymmetric
3-d rank-2 tensors are not usual, although I will argue toward the end of the chapter that in three dimensions a
magnetic field is more naturally described by an antisymmetric (0, 2) tensor than by a vector.
In four dimensions, we also have symmetric tensors, such as the important energy-momentum tensor which
carries all the information about the energy and momentum density at a point, plus the flux of these quantities at
that point. And there is the famous antisymmetric (0, 2) Faraday electromagnetic field tensor F.
It can be useful to symmetrise or skew-symmetrise a general (r, 0) or (0, s) tensor. To symmetrise the compo-
nents of a (0, s) tensor T, construct:
1 X
T(µ1 ...µs ) = T (1.24)
s! π π(µ1 ...µs )
with round brackets around symmetric indices, and where the sum runs over all permutations π of µ1 . . . µs .
Contravariant components are symmetrised in the same way.
To antisymmetrise the components of a (0, s) or (r, 0) tensor T, construct:
1 ν1 ...νs
T[µ1 ...µs ] = δ Tν ...ν (1.25)
s! µ1 ...µs 1 s
j ... j
with square brackets around antisymmetric indices, and the general permutation symbol, δi11 ... iss , defined as:


 +1 j1 . . . js an even permutation of i1 . . . is



−1 j . . . j an odd permutation of i . . . i
j ... j 1 s 1 s
δi11 ... iss := (1.26)

 0 j1 . . . js not a permutation of i1 . . . is



 0 j = j or i = i for some k, l
k l k l
where even/odd means an even/odd number of transpositions (switches) of two indices. The permutation symbol
is seen to be antisymmetric in its upper and lower indices.
s! is the number of terms in all these summations, ie. the number of permutations of the indices of the tensor.
The normalisation factor 1/s! ensures consistency in the event that the Tµ1 ...µs should already be symmetric or
skew-symmetric. A simple example is that of a (2, 0) tensor:
1 µν 1
T µν = (T + T νµ ) + (T µν − T νµ ) ≡ T (µν) + T [µν]
2 2
9
A (0, 2) tensor can also be similarly decomposed.

EXERCISE: Symmetrise and antisymmetrise F(σ, τ , θ). If F = F µνλ eµ ⊗ eν ⊗ eλ , write Fs and Fa . How
many components do Fs and Fa have when F is defined over a 3-dim space? a 4-dim space? Is it possible to
reconstruct F µνλ from F (µνλ) and F [µνλ] ?
1.3.3 Transformations on tensors

Using the transformations in the table of section 1.2.6, it is straightforward to generalise the transformation laws
obeyed by tensor components. First, write T in the original basis and in the new (primed) basis:
′ ′ ′ ′
T = T µ1 ...µr ν1 ...νs ∂µ1 ⊗ · · · ⊗ ∂µr ⊗ dxν1 ⊗ · · · ⊗ dxνs = T α1 ...αr β1′ ...βs′ ∂α′1 ⊗ · · · ⊗ ∂α′r ⊗ dxβ1 ⊗ · · · ⊗ dxβs
We obtain:
′ ′ ′ ′
T α1 ...αr β1′ ...βs′ = T µ1 ...µr ν1 ...νs Lα1µ1 . . . Lαrµr (L−1 )ν1β ′ . . . (L−1 )νsβ ′ (1.27)
1 s
In traditional treatments, this transformation law actually defines a tensor. Scalars ((0, 0) tensors) remain invariant;
and we know how the components of vectors and 1-forms transform. What about, say, those of a (2, 0) tensor?
′ ′ ′ ′ ′
T α β = T µν Lα µ Lβ ν = Lα µ T µν L e β′ ⇐⇒ T′ = L T L e
ν
where Le is the transpose of L. Sometimes, as with 3-dim rotations, Le = L−1 ; sometimes, as with Lorentz boosts,

e = L. Tensors of rank 2 (2, 0), (0, 2), (1, 1) can be represented by n × n-dim matrices T, where n is the
L
dimension of the spaces V and V ∗ on which they are defined.
An immediate consequence of eq. (1.27) is that a tensor that is zero in a basis will remain zero in any other
basis. Thus, any equation made of tensors (or components) that is valid in one basis must hold in any other basis.
In the older view of tensors defined by transformations, an object may have tensor character under certain
transformations, but not others. For instance, 4-dim tensors might owe their tensor character to how they transform
under Lorentz transformations, while 3-dim tensors might be tensors only under rotations.
The transformation rules can always be used to establish whether an object is a tensor. For instance, on a space
of dimension n, the Kronecker delta, with components δνµ , is represented by the n × n identity matrix. It is a
mixed rank-2 tensor. Indeed, from the transformation law, eq. (1.27):
′ ′ ′ ′
δµ ν ′ = Lµ λ (L−1 )ρ ν ′ δλρ = Lµ λ (L−1 )λ ν ′ = Iµ ν ′
which are the components of the identity matrix. Here we learn that there is something more to δµν than just being
a tensor: its components remain the same under changes of basis!
1.3.4 The Levi-Civita symbol

Definition 1.15. In a Cartesian orthonormal basis of a n-dim space, the Levi-Civita symbol, ǫµ1 ...µn ,
j ... j
is defined in terms of the general permutation symbol, δi11 ... inn (eq. (1.26)), as:
ǫµ1 ...µn = δµ1 1......µ
n
n
It is skew-symmetric in its n indices, with ǫ1...n = +1, where the indices are in ascending order. In
pseudo-Riemannian manifolds, it is traditional to use ǫ0...n−1 , the 0 index corresponding to time.
The determinant of a n × n matrix L is a product of its elements antisymmetrised with respect to rows (or
columns):
det L = ǫν1 ...νn Lν11 · · · Lνnn (1.28)
If the Levi-Civita symbol is to be a tensor, the transformation laws on its components demand that:
′ ′
1 = ǫ1...n = ǫν1′ ...νn′ Lν1 1 · · · Lνn n = det L
This is the case when L is a 3-dim rotation or a Lorentz-boost matrix, under which ǫµ1 ...µn , like δµ ν , is invariant.
We shall discover a little later how the general Levi-Civita tensor can be constructed.
Fortunately, often we can avoid using the transformation law (1.27) if we build tensors from other objects
known to be tensors. The following section presents some important examples.
10
1.4 Two More Ways to Construct Tensors

1.4.1 Contracted tensors
r−1
Definition 1.16. The contraction of a mixed-type tensor is a linear map Tsr → Ts−1 , (r ≥ 1, s ≥ 1).
More precisely, going back to eq. (1.16), insert a basis and its cobasis into only two input slots:
T(. . . , dxγ , . . . , ∂γ , . . .) = T ... γ ... ... γ ... . . .⊗∂µi−1 ⊗∂µi+1 ⊗. . . . . .⊗dxνj−1 ⊗dxνj+1 ⊗. . . (1.29)
In terms of components, one just makes a contravariant index µ the same as a covariant index, ν, by multiplying
the component by δνµ , thus forcing a summation over these indices.
For instance, consider T ∈ T11 . The contraction of T = T α β ∂α ⊗ dxβ is a scalar, called its trace:
Tr T = T(dxµ , ∂µ ) = T α β ∂α (dxµ ) dxβ (∂µ ) = T µ µ = T µ ν δνµ
When contracting tensors of type higher than 2, it is important to specify which indices are being contracted. Thus,
the tensor T µν λ ∂µ ⊗ ∂ν ⊗ dxλ has two possible contractions: the vectors T µν µ ∂ν and T µν ν ∂µ .
1.4.2 Inner product

Up to now, there has been no unique link between tensors of type (r, 0), (0, r), or (r−q, q). To set up such a link,
a new object is needed: a (0, 2) tensor g = gµν dxµ ⊗ dxν in a coordinate basis.
Indeed, let us insert only one vector in, say, the second slot of g, with as result a covector that we call ũ:
ũ := g( , u) = gµν dxµ dxν (u) = gµν uν dxµ
The correspondence will be unique if we demand that g be symmetric, because then g(u, ) = g( , u). Inserting
vectors in the two input slots yields the number: g(u, v) = gµν uµ v ν . In effect, g may be thought of as a map
from V to its dual space! Once defined, it establishes a unique correspondence between a vector u ∈ V and a
1-form ũ ∈ V ∗ .
Definition 1.17. The inner product of two vectors u and v, < u, v >, can now be defined as:
< u, v > ≡ u · v := g(u, v) = gµν uµ v ν (1.30)
In a coordinate basis, gµν = < ∂µ , ∂ν >, which is just the naı̈ve scalar product of the two basis vectors
∂µ and ∂ν . In a general basis, gµν = < eµ , eν >.
g(u, u) = gµν uµ uν is called the norm of u. If it is positive (negative) ∀ u, we say that g is positive
(negative) definite. But if g(u, u) = 0 for some non-zero vector (null vector) u, then g is indefinite.
g must be invertible (ie. det g 6= 0) , and we can always write: u = g−1 g(u) = g−1 (ũ). Now g−1 must take
1-forms to vectors, which means it must be a (2, 0) tensor g−1 = (g−1 )µν ∂µ ⊗ ∂ν . Then:
u = uµ (g−1 )αβ ∂α ∂β (dxµ ) = uµ (g−1 )αβ ∂α δµ β = uµ (g−1 )αµ ∂α
As will be justified soon, we identify (g−1 )µν with the contravariant components of g, gµν , and, comparing with
u = uα ∂α , we conclude that uµ = gµν uν , uµ being thought now as the contravariant components of the 1-form.
In that sense u and ũ can both have covariant and conravariant components.
These mappings between V and V ∗ can be applied to any tensor T ∈ T ; in other words, g may be used to
convert any contravariant index of a given tensor into a covariant one (“lowering the index”), while g−1 may be
used to convert any covariant index of a given tensor into a contravariant one (“raising the index”). Thus, we say
that the inner product sets up an isomorphism between a vector space and its dual. Because of this connection, we
also have: ∂µ = gµν dxν . One tensor can have all-contravariant, all-covariant, or mixed components! In particular:
gµ ν = g(dxµ , ∂ν ) ≡ < dxµ , ∂ν > = dxµ (∂ν ) = δµν (1.31)

11
Since, as we have seen, δµν is basis-independent, so is g µ ν , unlike gµν and g µν . But g µλ gλν = g µ ν = δµν , which
justifies our earlier assertion that gµν are the components of g−1 . On a n-dim space, gµ µ = δµµ = n.
If δµν are the components of the identity matrix, I, δµν = gµρ δρν = gµν will not in general be the entries of I.
A final word of caution: we always wrote our matrices as Lµν , with the left index a row index. Why do we
not write the matrix of g’s components the same way? Because Lµν is a transformation between two bases in V n ,
whereas gµν transforms a basis in V n to its dual basis. For instance, in R3 , the basis dual to {e1 , e2 , e3 } cannot be
reached by any combination of rotations and translations. Also, Lµν is not a tensor component, but gµν is.
1.4.3 The metric

On top of telling us how to calculate the length of vectors (and angle between them) in the tangent space, the inner
product plays another extremely important rôle: it allows us to define distances on the manifold Mn :
Definition 1.18. As a metric tensor (metric for short), g tells us how to calculate lengths in a vector
space tangent to a point on a manifold, as well as distances on the manifold itself. The name is often
extended (abusively) to its components gµν . Thus,
∆s2 = gµν ∆xµ ∆xν (1.32)
gives the interval between two points labelled by xµ and xµ + ∆xµ .
In old-style notation, one often writes the metric in terms of an infinitesimal interval, or line element:
ds2 = gµν dxµ dxν (sometimes called the first fundamental form of the manifold) with the dxµ the
components of an infinitesimal displacement. In modern notation, however, one identifies the bilinear
form ds2 with g = gµν dxµ ⊗ dxν which then represents the interval ∆s2 for a ∆x to be specified:
∆s2 =< ∆x, ∆x >, identical to the standard eq. (1.32).
Example 1.7. Consider the position vectors x1 and x2 in R3 , with Cartesian components:
x1 7−→ (x1 , y1 , z1 )T , x2 7−→ (x2 , y2 , z2 )T
If we choose a positive-definite g with matrix representation g = I:   
1 0 0 x1 − x2
g(∆x, ∆x) = gµν ∆xµ ∆xν = x1 − x2 , y1 − y2 , z1 − z2 0 1 0  y1 − y2 
0 0 1 z1 − z2
The result, (∆s)2 = (x1 − x2 )2 + (y1 − y2 )2 + (z1 − z2 )2 = (∆x)2 + (∆y 2 ) + (∆z)2 , is recognised
to be the “Pythagorean” distance squared between two points: |x1 − x2 |2 .
Example 1.8. In R4 , let xi (i = 1, 2) be two position vectors with (cti , xi , yi , zi ) as contravariant and
(−cti , xi , yi , zi ) as covariant components. Then take the indefinite η ≡ g with matrix representation:
 
−1 0 0 0
 0 1 0 0
ηµν = 
 0 0 1

0 = diag (−1, 1, 1, 1)
0 0 0 1
g(∆x, ∆x) = −c2 (t1 − t2 )2 + (x1 − x2 )2 + (y1 − y2 )2 + (z1 − z2 )2

is the spacetime distance between two events in Special Relativity, with c the speed of light. Because
η is indefinite, there exist null (zero norm) vectors such that g(x, x) = 0. And, just as ∂ν xµ = δµ ν ,
we must write ∂ν xµ = ηµν .
ROn
bp R b pthe distance between two points λ = a and λ = b on a curve
general manifolds,
µ
parametrised by λ is given
µ ν
by a g(v, v)dλ = a gµν dλ x dλ x dλ, where v is the velocity vector and x are the coordinates describing
the curve on the manifold. The metric, or line element, is said to define the geometry of a manifold. Two manifolds
of the same dimension can have different geometries, eg. R4 with a positive-definite (∆s2 > 0) metric is not the
metric of 4-dim “flat” spacetime of special Relativity.
Quite often, we will wish to work in bases other than coordinate bases. The formal properties of g that we have
reviewed still hold, but its covariant and contravariant components can be different, even in the same coordinates.
12
Definition 1.19. A basis {eµ } such that g(eµ , eν ) = ±1 when µ = ν and 0 otherwise is said to be
orthonormal. A useful notation to distinguish it from a coordinate basis is {eµ̂ }, extending the usual
definition of orthonormality which admits only +1 and 0, and useful in the case of indefinite metrics.
Let n+ (n− ) denote the number of diagonal elements g(eµ̂ , eµ̂ ) equal to +1 (−1). The signature of
the metric is defined by s = n+ − n− . Since n+ + n− = n, the dimension of the space, we also have
s = n − 2n− , and det g = (−1)n− . n+ and n− are basis-independent, and so is the signature.
The sign of the signature of an indefinite metric is arbitrary and is set by convention, which can be a source of
confusion. Example 1.8 sets s = +2, a nice choice when spatial indices are often raised/lowered. In more general
spacetimes, s = −2 is often used (but not always. . . see Misner, Thorne and Wheeler’s Gravitation ). Thus, beware!
Definition 1.20. A n-dim space endowed with a metric of signature ±n is called Euclidean. If
n− = 1 (or n− = n − 1), the space is pseudo-Euclidean, or Lorentzian (aka Minkowski when
n = 4). Example 1.8 has a Minkowski metric in four-dimensional space.
Thanks to the metric, we recover the vector gradient of a function defined in calculus. You may have noticed
that throughout our discussion of manifolds and tangent spaces, no mention was made of an inner product, because
none was needed—until now. A metric g pairs the 1-form df with a vector, ∇f ; indeed, from eq. (1.30):
< ∇f, v > = g(∇f, v) = gµν (∇f )µ v ν = (gµν ∂ µ f ) v ν = (∂ν f ) v ν = [df ](v) (1.33)
where v is an arbitrary vector, and the components of ∇f in a coordinate basis are given by: (∇f )µ = gµν ∂ν f .
Only in a Euclidean metric with a standard basis are the components of ∇f the same as those of df .
Example 1.9. In Minkowski spacetime with coordinates (ct, x1 , x2 , x3 ) and metric ηµν = diag(−1, 1, 1, 1):
df = ∂t f dt + ∂i f dxi ∇f = − ∂ct f ∂ct + ∂ i f ∂i (∂ i = gij ∂j = ∂i )
There is something interesting about the determinant of the metric which we find by writing the transformation
′
law: gµν = ∂µ′ xα gαβ ∂ν ′ xβ , as a matrix equation, and taking the determinant. Defining g = det gαβ , we obtain:
2
∂x
g′ = g (1.34)
∂x′
where |∂x/∂x′ | is the Jacobian of the transformation matrix from x to x′ coordinates. Then g is not invariant!
Definition 1.21. A quantity that has extra powers of |∂x/∂x′ | as factors in its transformation law in
′
addition to the usual ∂µ′ xα and/or ∂α xµ factors is called a tensor density. Thus, g is a scalar density.
This might seem no more than an exotic property until we consider the n-dim volume element as usually written in
an integral. This, as we know from calculus, transforms as : dn x′ = |∂x′ /∂x|dn x (note the position of the prime
in the Jacobian!), so is not invariant. As a result, the volume integral of a scalar function is not
p invariant, yet there
n
should be no memory of the integration variables left after integrating. But if we transform |g| d x, we obtain:
p ∂x p ∂x′ n p
|g ′ | dn x′ = |g| d x = |g| dn x
∂x′ ∂x
Rp
which is seen to be a scalar! Then integrals written as |g|f (x)dn x are invariant. This concept of tensor density
as a notational device has been widely used in General Relativity, although post-1970 literature largely dispenses
with it when p-forms are involved. Indeed, later in section 1.5.2, we shall introduce a deeper definition of the
volume element.
13
1.5 Exterior Algebra

1.5.1 The exterior product
Definition 1.22. The exterior (wedge) product of two 1-forms is the antisymmetrised tensor product
2-form:
σ ∧ τ := σ ⊗ τ − τ ⊗ σ
In general, the wedge product can be used to construct a simple (or decomposable) skew-symmetric
covariant (0, p) tensor out of p 1-forms:
σ1 ∧ . . . ∧ σp = δµ1 1......pµp σµ1 ⊗ · · · ⊗ σ µp (1.35)
Applied to a n-dimensional cooordinate tensor-product cobasis dxρ1 ⊗ · · · ⊗ dxρp , where 1 ≤ ρ1 < . . . <
ρ ... ρ
ρp ≤ n, this becomes dxρ1 ∧ · · · ∧ dxρp = δµ11 ... µpp dxµ1 ⊗ · · · ⊗ dxµp , so that:
ρ ... ρ ρ ... ρ
[dxρ1 ∧ · · · ∧ dxρp ](∂ν1 , . . . , ∂νp ) = δµ11 ... µpp δµ1 ν1 · · · δµpνp = δν11 ... νpp(1 ≤ ρ1 < . . . < ρp ≤ n)
Vp ∗ (1.36)
(0, p) skew-symmetric tensors live in a space denoted by (V ), whose elements are also called p-forms (p is
traditionally used instead of s); very often, V ∗V= Rn .
Thus, from {dxρ } (1 ≤ ρ ≤ n) a basis of p (V ∗ ) can be constructed which contains n!/(p!(n − p)!) indepen-
dent, non-zero elements. In particular, a n-form on a n-dimensional space is a one-component object, a multiple
of the unique basis element, dx1 ∧ dx2 ∧ . . . ∧ dxn , with indices in increasing order. Skew-symmetry forces the
maximum rank of a non-trivial p-formVin n dimensions V to be n (why?).
p q
The
V p+q exterior product of a basis of and a basis of is a basis, dxρ1 ∧ . . . ∧ dxρp ∧ dxρp+1 ∧ . . . ∧ dxρp+q ,
of , again with indices in increasing order, and p + q ≤ n. V V
Then we construct a (p + q)-form out of the antisymmetrised tensor product of σ ∈ p and τ ∈ q :
µ ...µ ν ...νq
[σ ∧ τ ](uρ1 , . . . , uρp+q ) = δρ11...ρp+q
p 1
σ(uµ1 . . . uµp ) τ (uν1 . . . uνq ) µ1 < µ2 . . . < µp , ν1 < . . . < νq
(1.37)
µ ...µ ν ...νq
(σ ∧ τ )ρ1 ...ρp+q = δρ11...ρp+q
p 1
σµ1 ...µp τν1 ...νq µ1 < µ2 . . . < µp , ν1 < ν2 . . . < νq
The exterior product, in contrast to the vector (“cross”) product of vector analysis which it generalises, is
associative: σ ∧ (τ ∧ θ) = (σ ∧ τ ) ∧ θ.
Another very important property of the exterior product of a p-form and a q form is that:
σ ∧ τ = (−1)pq τ ∧ σ (1.38)
ν ...ν µ ...µp µ ...µ ν ...νq
This follows directly from eq. (1.37) by noting that it takes pq transpositions to get δρ11...ρp+q
q 1
into δρ11...ρp+q
p 1
.
It means that the exterior product commutes except when both forms have odd rank.
Eq. (1.37) is easier to use than it might appear. Here are three examples:
Example 1.10. Some people believe that we live in an 11-dimensional world. Let us work out one
component of the 3-form that is the exterior product of a 2-form, σ, and a 1-form, τ :
µνλ
(σ ∧ τ )11,3,6 = δ11 36 σµν τλ µ < ν
3 6 11 3 11 6 6 11 3
= δ11 3 6 σ36 τ11 + δ11 3 6 σ3 11 τ6 + δ11 3 6 σ6 11 τ3
= σ36 τ11 − σ3 11 τ6 + σ6 11 τ3
Example 1.11. In two dimensions, the exterior product of two 1-forms, σ1 and σ2 , is:
σ1 ∧ σ2 = (σ 1 1 dx1 + σ 1 2 dx2 ) ∧ (σ 2 1 dx1 + σ 2 2 dx2 )
= σ 1 1 σ 2 2 dx1 ∧ dx2 + σ 1 2 σ 2 1 dx2 ∧ dx1 = (σ 1 1 σ 2 2 − σ 1 2 σ 2 1 ) dx1 ∧ dx2
= (det S) dx1 ∧ dx2
where S is the 2 × 2 matrix whose two rows are the components of σ1 and σ2 , respectively.
14
1 2 1 3 2 3
V 2 In three dimensions, {dx ∧ dx , dx ∧ dx , dx ∧ dx } forms a basis of the space
Example 1.12.
of 2-forms, (V). Therefore, the most general (not necessarily simple!) 2-form can be written as:
1 V
τ = τ12 dx1 ∧ dx2 + τ23 dx2 ∧ dx3 + τ31 dx3 ∧ dx1 = τµν dxµ ∧ dxν ∈ 2 (1.39)
2
The summation on the right of the second equality is now unrestricted.
Three-dimensional simple 2-forms σ1 ∧ σ2 , however, have the coordinate form (EXERCISE):
(σ 1 1 σ 2 2 − σ 1 2 σ 2 1 ) dx1 ∧ dx2 + (σ 1 3 σ 2 1 − σ 1 1 σ 2 3 ) dx3 ∧ dx1 + (σ 1 2 σ 2 3 − σ 1 3 σ 2 2 ) dx2 ∧ dx3

(1.40)
In Euclidean R3 with Cartesian coordinates, the components would be those of the vector product of
the two vectors associated with σ1 and σ2 .
V
In four dimensions, a basis for 2 contains 6 elements. EXERCISE: What are the components of the exterior
product of two 1-forms in three and four dimensions? (Hint: the components must look like: σµ1 σν2 − σµ2 σν1 .)
More generally, consider simple p–forms on a n-dimensional space. In terms of a basis {dxν }, we have for
1-forms σµ : σµ = σ µ ν dxν (the superscripts on σ and dx being labels for the 1-forms). Thus, with eq. (1.35):
σ1 ∧ . . . ∧ σ p = δµ1 1......pµp σµ1 ⊗ · · · ⊗ σµp (unrestricted sum over µi )

= ǫµ1 ... µp σ µ1ν1 · · · σ µpνp dxν1 ⊗ . . . ⊗ dxνp (unrestricted sums over µi and νi ) (1.41)
where definition 1.15 has been used, and the summation over each νi (1 ≤ i ≤ p) runs from 1 to n,
If we construct a p × n matrix S whose ith row is the n components of the 1-form σ i , we may notice, referring
back to section 1.3.4, that the expression inside the square backets in the second line is the determinant of the p × p
submatrix extracted from column indices ν1 . . . ν p of S, with ν1 < . . . < ν p. Therefore, in eq. (1.42), each term
in the sum over the νi indices has as coefficient a p × p determinant. Each row of a determinant contains p out of
the n components of the 1-forms σ, and these components, labelled by ν1 < . . . < νp , must be the same as the
ones on dxν1 ∧ . . . ∧ dxνp in that term. Also, the νi indices in the square bracket have been antisymmetrised at the
same time, automatically antisymmetrising the tensor-product basis elements dxν1 ⊗ . . . ⊗ dxνp . Therefore:

σ1 ∧ . . . ∧ σ p = ǫµ1 ... µp σ µ1ν1 · · · σ µpνp dxν1 ∧ . . . ∧ dxνp (ν1 < · · · < νp ) (1.42)
With eq. (1.35), the output (a number!) resulting from inputting u1 , . . . , up into σ1 ∧ . . . ∧ σp is:
σ1 ∧· · ·∧σ p (u1 , . . . , up ) = δµ1 1......µ
p
p
σµ1 ⊗· · ·⊗σ µp (u1 , . . . , up ) = ǫµ1 ... µp σµ1 (u1 ) · · · σµp (up ) = det [σ i (uj )]
(1.43)
i i i µ
ie. the determinant of the p × p matrix S whose entries are: S j = σ (uj ) = σµ uj , with µ running from 1 to n.
Example 1.13. For a 3-dim V ∗ of which the 2-forms dxi ∧ dxj are basis elements, we have:
ui uj
dxi ∧ dxj (u, v) = dxi (u) dxj (v) − dxj (u) dxi (v) = vi vj
In Rn with Cartesian coordinates, we interpret this (up to a sign—see 1.5.2 below!) as the area of the
parallelogram whose defining sides are the projections of u and v on the xi -xj plane.
ν ...ν
Example 1.14. There is another useful definition of the permutation symbol, δµ11 ...µnn , equivalent to
the one given by eq. (1.26), and given by eq. (1.36):
δµν11...ν ν1 νn
...µn = dx ∧ . . . ∧ dx (∂µ1 , . . . , ∂µn )
n
Then eq. (1.43) becomes:

δν1µ1 · · · δν1µn
.. ..
δµν11...ν
...µn
n = . . (1.44)
νn
δ µ1 ··· δνnµn
15
Another application provides an easy test for the linear independence of p 1-forms: if their wedge product is
different from zero, those p 1-forms are linearly independent. If they were not, one of them at least could be written
as a linear combination of the others and antisymmetry would force the wedge product to vanish. Conversely, if
the p 1-forms are linearly independent, σ1 ∧ · · · ∧ σp cannot vanish.
Example 1.15. In general a p-form F is not simple. If it is, then the 2p-form F ∧ F must vanish by
antisymmetry. But to what extent does the converse hold?
V
Take F ∈ 2 . When n = 2, F = F12 dx1 ∧ dx2 is simple by inspection, with no need to invoke
F ∧ F = 0.
V
For n = 3, F ∧ F still trivially vanishes for F ∈ 2 . That vanishing will lead to F being simple, but
the argument is a little more involved. We can always write:
F = F12 dx1 ∧ dx2 + F13 dx1 ∧ dx3 + F23 dx2 ∧ dx3 ≡ σ + τ ∧ dx3
where σ is a 2-form on the 2-dim subspace, spanned by dx1 ∧ dx2 , of the 3-dim
V space, and τ is a
1-form on the same 2-dim subspace. Then σ is simple, ie., σ = α ∧ β (α, β ∈ 1 ) because n = 2.
Thus, F ∧ F = σ ∧ σ + 2σ ∧ τ ∧ dx3 = 0. Also, σ ∧ σ = 0 because σ is simple. Then
α ∧ β ∧ τ ∧ dx3 = 0. But the only possible linear dependence between the four 1-forms is between
α, β and τ since none of them depends on dx3 . Therefore, λ1 α + λ2 β + λ3 τ = 0.
If λ3 = 0, β is a multiple of α, so σ = 0, leaving F = τ ∧ dx3 , a simple form. If λ3 6= 0, τ = aα + bβ,
and: b
F = α ∧ β + (aα + bβ) ∧ dx3 = (α + β) ∧ (β + adx3 )
a
which is simple. Thus, 2-forms on 3-dim space are always simple! EXERCISE: A 2-form in n = 4 is
simple if, and only if, F ∧ F = 0. Also, when F ∧ F 6= 0 in n = 4, F can be written (EXERCISE) as
the sum of two simple 2-forms. These statements hold whether or not a metric has been introduced.
1.5.2 Oriented manifolds, pseudo-vectors, pseudo-forms and the volume form

Definition 1.23. Two bases are said to have the same (opposite) orientation if the determinant of
the matrix of the transformation between them is positive (negative). Therefore, bases fall into two
classes, or orientations. Orienting a manifold then means arbitrarily specifying one orientation to
be positive (right-handed), and the other negative (left-handed). Manifolds on which transport of a
basis around a closed loop reverses orientation are non-orientable (eg. the Möbius strip).
In R3 , for instance, ex ∧ ey ∧ ez , ey ∧ ez ∧ ex and ez ∧ ex ∧ ey can be transformed into one another
by matrices of determinant +1. By convention, they are taken to be right-handed. But ey ∧ ex ∧ ez =
−ex ∧ ey ∧ ez cannot be similarly reached from ex ∧ ey ∧ ez : it is an element of a left-handed basis.
Definition 1.24. An object that behaves in all respects as a vector or a p-form, except that its sign is
reversed under a reversal of orientation of the manifold, is called a pseudovector or a pseudoform.
Example 1.16. Generalising example 1.13 above, the simple n-form dx1 ∧ · · · ∧ dxn , when acting
on the vectors v1 , . . . vn in that order, outputs a mumber of magnitude equal to the volume of the
parallelopiped whose edges are v1 , . . . vn . With p = n in eq. (1.43), this is readily computed as the
determinant of all the vector components. There is also a sign involved, with + corresponding to the
orientation of the vectors being the same as that of the basis. We then say that this volume is oriented.
Because it changes sign under interchange of any two basis vectors, we recognise it as a pseudoform.
Definition 1.25. In general coordinates ui on a n-dim manifold, we define the volume pseudoform:
∂x p
dn u := du1 ∧ · · · ∧ dun = |g| du1 ∧ · · · ∧ dun
∂u
where the xi form an orthonormal basis, usually Cartesian, and we have used eq. (1.34) with |g| = 1
for orthonormal bases.
16
1.5.3 The Levi-Civita pseudotensor

We have already remarked on the fact that the Levi-Civita symbol does not transform as a tensor. Consider,
however, the volume pseudoform
p of definition 1.25. By inspection it is a n-form with the single independent
n
component (d u)1...n = |g|. Its other components are obtained by antisymmetrising with the Levi-Civita
symbol, which we shall now denote by [µ1 . . . µn ] to avoid any confusion later. That is:
p
(dn u)µ1 ...µn = |g| [µ1 . . . µn ]
Thus, the right-hand side is the component of a covariant p pseudotensor, ǫ, of rank n. Henceforth,
p whenever
we write components ǫµ1 ... µn , they are to be understood as |g| [µ1 . . . µn ], so that ǫ1...n = |g|.
We obtain ǫ1 ... n by raising the n indices of ǫ1 ... n with g. In general coordinates:
p p 1 p (−1)n−
ǫ1...n = g1µ1 · · · g nµn ǫµ1 ... µn = g1µ1 · · · gnµn |g| δµ1 1......n µn = det gαβ |g| = |g| = p
(−1)n− |g| |g|
In orthonormal bases, this is simply: ǫ1 ... n = (−1)n− ǫ1 ... n .

Both ǫν1 ... νn and ǫµ1 ... µn being antisymmetric, we can relate the permutation symbol to the Levi-Civita pseu-
ν ...ν
dotensor with: ǫν1 ... νn ǫµ1 ... µn = a δµ11 ...µnn . To determine a, we use: ǫ1 ... n ǫ1 ... n = (−1)n− , and there comes:
δν1µ1 · · · δν1µn
.. ..
ǫν1 ... νn ǫµ1 ... µn = (−1)n− δµν11...ν
...µn = (−1)
n n−
. .
δνnµ1 · · · δνnµn
(1.45)
1
ǫν1 ... νp νp+1 ... νn ǫµ1 ... µp νp+1 ... νn = (−1)n− δµν1 ...µ
... νp
(unrestricted sums)
(n − p)! 1 p
In a Euclidean 3-dim space with an orthonormal metric, n− = 0, and the expanded product has six terms.
When contracted over the last or first indices, we obtain (EXERCISE): ǫijk ǫlnk = δi l δj n − δj l δi n . Other
expressions for the product of Levi-Civita tensors in a 4-dim Minkowski space can be found in MTW, pp. 87-88.
1.5.4 The Hodge dual of a p-form

To a vector v on a n-dim metric-endowed space corresponds a pseudoform σ of rank n-1:
σ = v ν ǫνµ1 ...µn−1 duµ1 ∧ . . . ∧ duµn−1 (µ1 < . . . < µn−1 ) (1.46)
which, like v, has n (independent!) components. In 3-dim R3 this is the pseudo-2-form:

p
σ = |g| v 3 du1 ∧ du2 − v 2 du1 ∧ du3 + v 1 du2 ∧ du3
Vp
Also, there must be a mapping between the 1-form dual to v and the (n-1)-pseudoform. Generalising to :
Definition 1.26. Let V n be endowed with a metric and a coordinate basis {∂µ }. With ǫ the Levi-
Civita pseudo-tensor, the Hodge dual† maps a p-form σ to a (n-p)-form ⋆σ = (⋆σ)|ν1 ...νn−p | duν1 ∧
⋆ ⋆
. . . ∧ duνn−p , where we introduce the compact notation: |µ1 . . . µp | ≡ µ1 < . . . < µp , and with
components:
1
(⋆σ)ν1 ...νn−p =
⋆ σµ ...µ ǫµ1 ...µp ν1 ...νn−p = σ |µ1 ...µp | ǫµ1 ...µp ν1 ...νn−p (1.47)
p! 1 p
The Hodge dual of a p-form is a pseudo-form, and vice-versa. It can be shown that, given a mostly
positive metric g, ⋆⋆σ = (−1)n− (−1)p(n−p) σ. So Hodge duality is idempotent in Euclidean spaces
⋆ ⋆
(n− = 0) of odd dimension, such as R3 . In 4-dim Minkowski space (n− = 1), it is idempotent only
on 1- and 3-forms.
†
Here, the meaning of “dual” has no relation to its other use in “dual” space or basis.
17
One immediate application of eq. (1.47) is that the n-dim volume form is the Hodge dual of the 0-form 1:
⋆ 1
⋆ = ǫ|µ1 ...µn | duµ1 ∧ · · · ∧ duµn
p
= |g| du1 ∧ · · · ∧ dun
A very important consequence of the fact that ⋆⋆σ = ±σ is that a p-form and its Hodge dual contain
⋆ ⋆
exactly the same information! Thus, “dualising” a p-form (or an antisymmetric contravariant tensor) can remove
some (or all!) the redundancy due to anrisymmetry while preserving its information. For instance, in 4-dim
Minkowski space, a 4-form with components σµνλρ is dual to a pseudo-0-form, so one independent number instead
of 44 = 256. Or a 3-form with a ppriori 43 = 64 components can be Hodge-dualised to its dual pseudo-1-form
whose four components are (up to |g|) the only independent components of the 3-form.
Example 1.17. If T is a (2, 0) skew-symmetric tensor:

1
(⋆T )λ = ǫµνλ T µν = ǫµνλ T |µν|
⋆
in R3
2
1
(⋆T )λρ = ǫµνλρ T µν = ǫµνλρ T |µν|
⋆
in R4
2
With an orthonormal metric, it is not hard to work out that in the first line ⋆T1 = T 23 , ⋆T2 = T 31 ,
⋆ ⋆
and ⋆T3 = T 12 , so that the 1-form dual to T contains only the three independent components of T.
⋆
In the 3-dim Euclidean

V space
V of p-forms of example 1.12, {dx1 , dx2 , dx3 } and {dx2 ∧dx3 , dx3 ∧dx1 , dx1 ∧
1 2
dx2 } are bases for and , respectively. The two bases are each other’s Hodge dual. In fact, we can Hodge-
dualise the (co)basis: ⋆(duµ1 ∧ · · · ∧ duµp ) = ǫµ1 ...µp |µp+1 ...µn | duµp+1 ∧ · · · ∧ duµn , (or divide by (n − p)! if the
⋆
summations V are unrestricted), in which case the components are not changed—they are just re-allocated to basis
elements of n−p . There are corresponding expressions for Hodge-dualising coordinate bases or the components
of contravariant tensors, as illustrated by the above example.
Example 1.18. If σ and τ are 3-dim 1-forms, the 2-form: σ ∧ τ = (σ2 τ3 − σ3 τ2 ) dx2 ∧ dx3 +
(σ3 τ1 − σ1 τ3 ) dx3 ∧ dx1 + (σ1 τ2 − σ2 τ1 )dx1 ∧ dx2 has as its Hodge dual on a space with metric
g the pseudo-1-form:
p
⋆(σ ∧ τ ) =
⋆ |g| (σ2 τ3 − σ3 τ2 ) dx1 + (σ3 τ1 − σ1 τ3 ) dx2 + (σ1 τ2 − σ2 τ1 ) dx3
If σ corresponds to the vector u and τ to v via the metric, this says that: ⋆(u ∧ v) = u × v, or,
⋆
with eq. (1.47), (u × v)µ = 21 gµρ ǫνλρ (uν v λ − uλ v ν ) = gµρ ǫρνλ uν v λ . So when calculating a vector
product, one is implicitly taking a Hodge dual, the only way that the result can be a pseudo-vector.
It is easy to recover all the relations of vector analysis in Cartesian R3 . For instance:
u · (v × w) = ǫµνρ uµ v ν wρ
= wρ ǫρµν uµ v ν (cyclic permutation of indices on ǫ)
= w · (u × v).
1.6 Exterior Calculus
Definition 1.27. A (r, s) tensor field T(p) on a n-dim manifold M n is a function of points p ∈
M n whose components T (p) = T(dxν1 , . . . , dxνr , ∂µ1 , . . . , ∂µs ) are real-valued differentiable
functions of coordinates on the manifold.
Examples: the coordinate vector field ∂µ , the gravitational and electric fields, the metric tensor with
components < ∂µ , ∂ν > in a coordinate basis.
18
How do we describe the change of a tensor field at a point? More precisely, how do we differentiate it? We
already know from section 1.2 how to take the directional derivative of a (0, 0) tensor, ie. a function. On a “flat”
(without curvature) manifold, directional derivatives of tensor-field components can be calculated in the same way.
For general (r, s) tensors, however, because of point-dependent bases, defining differentiation requires extra
structure, called a connection, or covariant derivative. Raising this important issue is like opening Pandora’s box
(aka can of worms), because there are a lot of ways to construct such a connection. A few, however, have gained
favour as “natural”. Here we only discuss a particular type of differentiation that offers a neat unification of the
ideas of gradient, divergence and curl in vector calculus, and for which a connection is actually not needed.
1.6.1 Exterior derivative

We introduce the exterior derivative operator† , d, which acts on p-forms σ = σ|µ1 ...µp | dxµ1 ∧. . .∧dxµp , defined
over some manifold M n to give p + 1-forms, also defined on M n . Let σ be a p-form and τ a q-form. The operator
satisfies the following properties:

(a) dσ := ∂µ0 σ|µ1 ...µp | dxµ0 ∧ dxµ1 ∧ . . . ∧ dxµp = dσ|µ1 ...µp | ∧ dxµ1 ∧ . . . ∧ dxµp
Each component of dσ with indices in increasing order is a sum of p + 1 terms:
ν ν ...ν
(dσ)µ1 ...µp+1 = δµ01 ...µ
1 p
p+1 ∂ν0 σ|ν1 ...νp | = ∂µ1 σ|µ2 ...µp+1 | − ∂µ2 σ|µ1 µ3 ...µp+1 | + ∂µ3 σ|µ1 µ2 ...µp+1 | − . . .
(b) d(σ + τ ) = dσ + dτ (p = q).

(c) If σ is a 0-form, ie. just a function, then dσ is the 1-form gradient of that function.
(d) d(σ ∧ τ ) = dσ ∧ τ + (−1)p σ ∧ dτ (aka the antiderivation property of d in the exterior product).
(e) d2 σ = 0 (Poincaré lemma).
We shall not prove the antiderivation property (you can do it as an EXERCISE), but Poincaré’s lemma is so
famous and important that it deserves some proof.
First, for an arbitrary function f (0-form):
d2 f = d(∂ν f dxν ) = ∂µ ∂ν f dxµ ∧ dxν = 0
since ∂µ ∂ν is symmetric in µ and ν. If g is another function, d(df ∧ dg) = d2 f ∧ dg − df ∧ d2 g from the
antiderivation property; this must vanish since d2 f = d2 g = 0. By extension, the exterior derivative of the
exterior product of any number of differential 1-forms also vanishes. Now, from properties (a) and (d) above:

d2 σ = d dσ|µ1 ...µp | ∧ dxµ1 ∧ . . . ∧ dxµp

= d2 σ|µ1 ...µp | ∧ dxµ1 ∧ . . . ∧ dxµp − dσ|µ1 ...µp | ∧ d dxµ1 ∧ . . . ∧ dxµp
The first term vanishes because d is nilpotent on functions; the second vanishes because the exterior derivative of
a wedge product of differential 1-forms vanishes.
Example 1.19. In R3 , with u, v, and w as arbitrary coordinates, the differential of a function f in the
coordinate basis {du, dv, dw} is the 1-form:
df = ∂u f du + ∂v f dv + ∂w f dw (1.48)
This is valid only for a coordinate basis. In a spherical coordinate basis {dr, d θ, d φ}, for instance,
df would keep the above simple form. But if we insist on a basis whose elements are normalised
to unity, such as {dr̂, d θ̂, d φ̂} = {dr, rd θ, r sin θ d φ} — as is almost always the case in vector
analysis applied to physics — consistency demands that we write:
1 1
df = ∂r f dr̂ + ∂θ f dθ̂ + ∂φ f dφ̂ (1.49)
r r sin θ
†
Some authors prefer the notation ∇∧ for the exterior derivative.
19
Example 1.20. The exterior derivative of a 1-form σ is the 2-form:

θ = dσ = ∂µ σν dxµ ∧ dxν
(1.50)
= (∂µ σν − ∂ν σµ ) dxµ ∧ dxν (µ < ν)
The components of θ are: θµν = ∂µ σν − ∂ν σµ . (EXERCISE: what would be the exterior derivative
of a 2-form? What would its components be?)
Example 1.21. What about the exterior derivative of a 1-form σ = σu du + σv dv + σw dw in R3 ?
With the equivalent expression: dσ = dσν ∧ dxν , we obtain:

dσ = ∂v σu dv + ∂w σu dw ∧ du + ∂u σv du + ∂w σv dw ∧ dv + ∂u σw du + ∂v σw dv ∧ dw
= (∂v σw − ∂w σv ) dv ∧ dw + (∂w σu − ∂u σw ) dw ∧ du + (∂u σv − ∂v σu ) du ∧ dv
(1.51)
Taking the Hodge dual gives the pseudo-1-form:
p h i
⋆dσ =
⋆ |g| (∂v σw − ∂w σv ) du + (∂w σu − ∂u σw ) dv + (∂u σv − ∂v σu ) dw (1.52)
p of the 3-dim
By analogy with tensor algebra results, we can recover the contravariant components
curl of a vector, but only in Cartesian coordinates! Only in those coordinates is |g| = 1, with
covariant and contravariant components the same.
As we know all too well, the vector components of the curl of a vector in curvilinear coordinates
can be quite complicated; this is largely due to our insisting on working with objects which are less
natural. Exterior derivatives do not involve raising indices with a metric, and so are more natural.
It is interesting that, in vector calculus with Cartesian coordinates, we could write σ as A · dx, with
A a corresponding vector. Then the right-hand side of eq. (1.51) would correspond to ∇ × A · dS,
where dS is a surface element with Cartesian components dy ∧ dz, dz ∧ dx, and dx ∧ dy. Then we
could write d (A · dx) = ∇ × A · dS.
Example 1.22. Here is an intriguing example: the exterior derivative of a pseudo-2-form τ in R3 with
some metric g. Since this will be a pseudo-3-form, we expect it to be a one-component object. Indeed:
dτ = (∂u τvw du) ∧ dv ∧ dw + (∂v τwu dv) ∧ dw ∧ du + (∂w τuv dw) ∧ du ∧ dv
(1.53)
= (∂u τvw + ∂v τwu + ∂w τuv ) du ∧ dv ∧ dw
Now, in three-dimensions τ can be viewed as the Hodge dual, τ = ⋆σ, of the 1-form σ = σu du +
⋆
σv dv + σw dw. In terms of components, τµν = ǫµνλ σ λ . Inserting and then taking the Hodge dual of
p
the last expression, using ⋆(du ∧ dv ∧ dw) = ǫ123 = (−1)n− / |g| from section 1.5.3, gives:
⋆
1 p
(−1)n− ⋆d ⋆σ = p ∂µ ( |g| σ µ )
⋆ ⋆ (1.54)
|g|
Definition 1.28. Extending to n dimensions, we call the right-hand side the divergence, div B, of
the n-dim vector B with components B µ = σ µ . It holds in any coordinates in a metric-endowed space.
In vector calculus with Cartesian coordinates, τ = B 1 dy ∧ dz + B 2 dz ∧ dx + B 3 dx ∧ dy = B · dS, and eq.

(1.53) could be written as: d(B · dS) = ∇ · B d3 x.
The operator ⋆d ⋆ sends a p-form into a (p-1)-form. In mathematical references, this operator is introduced
⋆ ⋆
(up to a sign!) as the codifferential operator, δ. We quote without proof the relation between them: When
acting on a p-form in a Euclidean manifold, δ σ = (−1)n(p+1)+1 ⋆d ⋆σ, and δ σ = (−1)n(p+1) ⋆d ⋆σ in a
⋆ ⋆ ⋆ ⋆
pseudo-Euclidean manifold. Actually, these expressions happen to hold also in a Riemannian (curved) or pseudo-
Riemannian manifold!
20
Like the exterior derivative, the codifferential operator is nilpotent. Indeed, δ2 = ⋆d ⋆⋆d ⋆ = ±⋆d2 ⋆ = 0.
⋆ ⋆ ⋆ ⋆ ⋆ ⋆
Definition 1.29. We define the divergence of any p-form: div σ := −δσ = (−1)n(p+1)+n− ⋆d ⋆σ. ⋆ ⋆
This ensures consistency between eq. (1.54) and the conversion between ⋆d ⋆ and δ. We extend eq.
⋆ ⋆
(1.54) to the divergence of any p-form σ on a n-dim space:

1 p 1 p
(div σ)µ1 ...µp−1 := p ∂ν |g| σ ν µ1 ...µp−1 = p ∂ν |g| g νρ σρµ1 ...µp−1 (1.55)
|g| |g|
From eq. (1.54) follows the definition of the 3-dim Laplacian of a scalar function f in coordinates ui :
1 p 1 p
∇2 f = p ∂i |g|∂ i f = p ∂i |g| g ij ∂j f (1.56)
|g| |g|
1.6.2 Laplace-de Rham operator, harmonic forms, and the Hodge decomposition
Definition 1.30.V The Laplace-de

Vp Rham operator is defined as ∆ = δ d + d δ = (d + δ)2 . It is a
p
mapping ∆ : −→ . When acting on a scalar function, ∆ = δ d; then we also speak of the
Laplace-Beltrami operator.
It is not hard to show that it reduces to the negative of the Laplacian operator of vector analysis, ie. ∆ = δ d =
−⋆d ⋆d = −∂i ∂ i = −∇2 , when acting on 0-forms on Euclidean R3 with Cartesian coordinates. We shall define
⋆ ⋆
∇2 so that ∇2 = −∆ when acting on any p-form in Euclidean R3 equipped with a standard basis.
Example 1.23. For instance, let it act on a 1-form σ in Euclidean R3 . That is, take ∆σ = ⋆d ⋆dσ −⋆ ⋆
d ⋆d ⋆σ using the conversion formula between δ and ⋆d ⋆. Using eq. (1.52), the first term is the
⋆ ⋆ ⋆ ⋆
curl of a curl, whereas the second is the gradient of a divergence. Thus, we recover the expression
well-known from vector calculus: ∇2 A = ∇(∇ · A) − ∇ × ∇ × A, where A is the vector associated
with the 1-form σ.
When acting on functions (0-forms) in Minkowski space, the Laplace-de Rham operator is related to the
d’Alembertian operator := ∂µ ∂ µ : ∆ = − . This defines the d’Alembertian of any p-form in Minkowski space.
Definition 1.31. A p-form σ is said to be harmonic if ∆σ = 0. This generalises the notion of

functions being called harmonic when they satisfy the Laplace equation.
Definition 1.32. A closed form is one whose exterior derivative vanishes. A p-form that can be written
as the exterior derivative of a (p-1)-form is said to be exact.
Clearly, Poincaré’s lemma states that an exact form is closed. But is a closed form exact, ie. if dσ = 0, does
it follow that σ = dτ , with τ uniquely determined? The answer is no, if only because one can always add the
exterior derivative of an arbitrary (p - 2)-form θ to τ and still satisfy dσ = 0. Also, the converse of Poincaré’s
lemma (not proved) states that only in a submanifold in which all closed curves can be shrunk to a point does
dσ = 0 entail the existence in that submanifold of a non-unique (p - 1)-form whose exterior derivative is σ. In
topology, we say that the submanifold must be simply connected (eg. no doughnuts!).
We quote without proof an important result of Hodge: On finite-volume (compact) manifolds without bound-
aries, such as S n , or on a torus, ∆σ = 0 if, and only if, dσ = 0 and d⋆σ = 0 (or δ σ = 0). Harmonic forms
⋆
are both closed and co-closed! This property also holds on open manifolds (eg. Rn ) if σ has compact support (it
vanishes outside a bounded closed region), or if it goes to zero sufficiently fast at infinity.
21
Definition 1.33. Assuming a compact manifold without boundaries or, failing that, compact support
(sufficiently fast fall-off at infinity), the unique Hodge decomposition writes a p-form σ as a sum of
exact (closed), co-closed, and harmonic p-forms:
σ = dα + δ β + harmonic p-form (1.57)
non-unique. dα, δ β and the harmonic p-form
where α is a (p-1)-form and β is a (p+1)-form, both V
in the decomposition live in orthogonal subspaces of p .
Example 1.24. Let A be a vector field with compact support on Euclidean R3 . Then its Hodge
decomposition says that its associated 1-form can be written as the exterior derivative of a 0-form (ie.
the gradient of a function), plus the divergence of a 2-form, β, plus some harmonic 1-form. Now, since
⋆β is a pseudo-1-form in R3 , δ β = ⋆d ⋆β is a 1-form. Then, from eq. (1.52) this term corresponds
⋆ ⋆ ⋆
to the curl of a pseudovector. Therefore, we obtain in terms of vectors:

A = ∇φ + ∇ × M + H (1.58)
where φ is a scalar field, M a pseudovector field, and H another vector field which satisfies ∇2 H = 0
everywhere. But if H vanishes at infinity in R3 , then it must vanish everywhere, and we have the
Helmholtz decomposition for a vector field with compact support.
The curl of ∇φ vanishes identically, and is often called the longitudinal projection of A; the diver-
gence of ∇ × M vanishes identically, and we can call it the transverse projection of A. Thus, ∇ · A
contains no information whatsoever about the transverse part of A, whereas ∇ × A knows nothing
of its longitudinal part. This provides a very useful and powerful tool for analysing 3-dim first-order
field equations (eg. Maxwell’s equations) which are usually statements about the divergence and the
curl of fields. If ∇ · A = 0 everywhere, we can conclude that A is purely transverse, since then φ in
eq. (1.58) satisfies the Laplace equation everywhere, so must vanish if it has compact support.
1.6.3 Exterior derivative and codifferential operator of a 2-form in Minkowski spacetime

V
Let F ∈ 2 on Minkowski (pseudo-Euclidean) R4 . Demand that F be exact and with compact support. Then
there exists a 1-form A such that F = dA, and Fµν = ∂µ Aν − ∂ν Aµ , in any metric. This means that dF = 0.
It is clear from Poincaré’s lemma that dF = 0 knows nothing about A: we say that it is an identity on A.
In addition, we give the exterior derivative of the Hodge dual of F, the pseudo-3-form d ⋆F, as a “source” J ,
⋆
with compact support and Hodge dual 1-form J = ⋆J . Then we have the inhomogeneous equation:
⋆
d ⋆F = 4π J
⋆ (1.59)
If we take the exterior derivative of the equation, the left-hand side vanishes identically, and the right-hand
side becomes: d J = 0. This is better known as the statement that the 4-divergence of J vanishes: ⋆d ⋆ J = 0. ⋆ ⋆
d J = 0 is actually a (metric-independent) conservation law for the source!

What we have constructed is Maxwell’s theory, with F the Faraday 2-form, A the electromagnetic potential
1-form, and J the 4-current. Our treatment assumes a mostly positive metric, as in MTW or Griffiths’ Introduction
to Electrodynamics. With a mostly negative metric, there is a minus sign on the right-hand side of eq. (1.59).
Because the differential operator d is metric-independent, we have given both the homogeneous and inhomo-
geneous first-order Maxwell equations in terms of exterior derivatives of F and its dual ⋆ F. It is easy to convert
⋆
the inhomogeneous equation to a divergence, simply by taking its Hodge dual:

⋆ d ⋆
⋆ ⋆ F = 4π ⋆J = 4π J
⋆ ⇐⇒ div F = − 4πJ (1.60)
In terms of Cartesian components, this can be shown (EXERCISE) to be equivalent to†
∂ µ Fµν = − 4π Jν ⇐⇒ ∂µ F µν = − 4π J ν
†
Again, with a mostly negative metric, such as in Jackson’s Classical Electrodynamics, there would be no minus sign on the right-hand
side. This is because F has opposite sign between the two conventions so as to obtain the same relations between the electric and magnetic
fields and the vector and scalar potentials.
22
the latter form being more appropriate if we insist on thinking of the source term as a vector. I would argue,
however, that the less conventional form eq. (1.59) is much the more natural. The exterior derivative is metric-
independent, and its index form can be written entirely with covariant indices, the natural ones for p-forms. But
to obtain its equivalent in divergence form, we have to Hodge-dualise the right-hand side, so that the vector J
source depends on the metric (see the paragraph after eq. (1.47)), whereas its 3-form version does not. The price,
of course, is that the 3-form version has 64 explicit components, although still only four independent ones.
It is worth noting that, although dF = 0 and the source equation (1.59) completely determine F, A is deter-
mined only up to an additive term df , where f is an arbitrary differentiable function.
As a 3-form, the homogeneous equation dF = 0 also has a lot of components, and when it comes to solving
the system, we may want to extract only the independent ones. This is the same as d ⋆(⋆F) = 0 whose Hodge ⋆ ⋆
dual is δ⋆F = 0. In other words, the divergence of ⋆ F vanishes, only four equations. Actually, this is a general,
⋆ ⋆
easily shown property (EXERCISE): whenever the exterior derivative of a p-form in some manifold vanishes, so
does the codifferential of its dual, and vice-versa.
Another great advantage of writing Maxwell’s equations as dF = 0 and d ⋆F = 4π J is that, provided the ⋆
source is smoothly varying, they are formally the same in curved spacetime! Only when divergences are written in
index notation are covariant derivatives involving a connection needed. Even in index notation, the first equation
does not involve the connection; it does not even require a metric.
Finally, nothing prevents us from constructing an extended Maxwell-like theory (not describing electromag-
netism) involving F as a 3-form. In the past few decades it has received a good deal of attention in some quarters.
1.7 Integrals of Differential (Pseudo)Forms

R
As we figure out the meaning of σp , where we use the notation σp to show explicitly the rank of a p-form, we
shall discover that pretty much any integral in n-dim calculus is the integral of some (pseudo)p-form.
1.7.1 Integrals of (pseudo) p-forms over a p-dim submanifold

R
As a warm-up, consider the integral of the Hodge dual of a scalar function f , ⋆f , over a n-dim region V in Rn ⋆
(eg., over some volume in R3 ). The Hodge dual of a scalar function f , of course, is a pseudo-n-form whose single
independent component is f . Then:
Z Z p Z Z
⋆f =
⋆ f (u) |g| du1 ∧ · · · ∧ dun = f (x) dx1 ∧ · · · ∧ dxn = f (x) dn x
V V V V
where u are general coordinates and dn x is the volume pseudo-n-form in Cartesian coordinates. Then we define:
Definition 1.34.
Z Z Z
1 n 1 n
f (x) dx ∧ · · · ∧ dx := f (x) dx · · · dx = f (x) dn x (1.61)
V V V
ie. the ordinary multiple integral of a scalar function of n variables in n dimensions.
When a p-dim region R is embedded in a n-dim manifold, it will be described with some coordinates u(x), that
is, n functions ui of the p Cartesian coordinates xj that parametrise Rp . Also, an orientation can be defined for the
region. What is the meaning of the integral of a p-form over such a region? We give two examples in R3 .
Example 1.25. Integral of a 1-form over a curve or “line integral”

We know that a curve C can be parametrised in terms of some real parameter t ∈ [a, b]. Then, if α is
a 1-form field on R3 , eq. (1.61) and the chain rule yield:
Z Z Z b Z b
i i
α = αi du = αi [u(t)](dt u ) dt = α(dt u) dt
C C a a
23
Only if R3 is given a metric and the curve 3

R described in R with Cartesian coordinates is this the usual
integral of a vector A along the curve, A·dx. In general, to integrate a vector along a curve, a metric
Rmust be introduced
R so as to transform the vector components into its associated 1-form’s components:
A · du = gij Aj dt ui dt. But no metric is needed to integrate a 1-form along a curve, and this is
the simpler and more natural operation.
If α is exact, then we immediately have the fundamental theorem of calculus:
Z Z b Z b Z
i
df = ∂ui f dt u dt = df = f (b) − f (a) = f
C a a ∂C
where ∂C is the boundary, ie. the end-points, of the curve.

Example 1.26. Integral of a 2-form over a surface
Let S be some surface described in R3 with three coordinate functions ui (x1 , x2 ). The surface is
parametrised with (x1 , x2 ) ∈ R2 , with two basis vectors ∂xi ≡ ∂i along the xi direction, for which
some orientation has been defined as positive. What meaning can we give to the integral of a 2-form
field β over S? From the chain rule and eq. (1.61) we find:
Z Z Z
β = βjk duj ∧ duk = βjk [u(x1 , x2 )] (∂1 uj ∂2 uk − ∂2 uj ∂1 uk ) dx1 dx2 (j < k)
S S
The integrals in R2 on the right are over a rectangular region of S in parameter space. The two
coordinate vectors (see section 1.2.3), ∂1 u and ∂2 u, are tangent to S at every point, and are usually
linearly independent, so form a basis for the space tangent to the surface at a point, with no metric
required as yet.
The Hodge dual of β , a pseudo-1-form, has an associated pseudo-vector B with, asp
components, the
i ijk 1
contravariant components of the Hodge dual, B = ǫ βjk (j < k), eg., B = β23 / |g|, etc. Then:
p B1 B2 B3
j k j k i j k j k
βjk (∂1 u ∂2 u − ∂2 u ∂1 u ) = ǫijk B (∂1 u ∂2 u − ∂2 u ∂1 u ) = |g| ∂1 u1 ∂1 u2 ∂1 u3
∂2 u1 ∂2 u2 ∂2 u3
From eq. (1.43), we recognise the last member of the equality as the output obtained from inserting the
three vectors whose components are the rows of the determinant into the three input slots of a simple
3-form—more accurately, a pseudo-3-form which, from definition (1.25) can be identified with the
volume pseudo-form d3 u. Then our integral can be written:
Z Z
3
β = d u(B, ∂1 u, ∂2 u) dx1 dx2
S
This makes it obvious that the integral is independent of the orientation of R3 , since switching it flips
the sign of both B and d3 u. At every point on S, we can choose the unit n̂ normal to the surface so
that n̂ and the vectors ∂1 u and ∂2 u tangent to the surface form a right-handed (positive orientation)
system. We also note that only the normal component of B can contribute to the integral (why?).
Then the scalar function d3 u(B, ∂1 u, ∂2 u) is the normal component of B multiplied by the surface of
the parallelogram defined by the coordinate vectors (see example 1.13) . Defining the surface element
dS ≡ |∂1 u × ∂2 u|, there comes:
Z Z Z
β = Bn dS = B · dS (1.62)
S
where the often used last expression is called the flux of the pseudo-vector B through the surface S.
It does not depend on the parametrisation chosen for S which is integrated out. The same result holds
if β is a pseudo-2-form, with B now a vector.
24
1.7.2 Stokes-Cartan Theorem

This famous theorem, which we shall not prove, equates the integral of the exterior derivative of a differentiable
(pseudo)p-form, ω, over a bounded region V in a manifold to the integral of ω over the boundary ∂V of V :
Z Z
dω = ω (1.63)
V ∂V
A technicality is that both V and ∂V must have compatible orientations. But no metric is required. The boundary
need not be connected, and it can be broken up into non-overlapping parts when it cannot be covered by a single
coordinate patch. Then we simply sum the integrals over each part.
Example 1.27. At the end of example 1.25 we had already worked out an application when ω is a
0-form: the fundamental theorem of calculus. When ω is a 1-form and V a 2-dim surface in Euclidean
R3 parametrised
R with Cartesian
H R C, theRsame example gives
coordinates and bounded by a closed curve
immediately: ∂V ω = C A · du. From eq. (1.51) and example 1.26, S dω = S ∇ × A · dS, and
we recover the well-known Kelvin-Stokes formula.
Finally, when ω is a pseudo-2-form
R in Euclidean
H R3 and S a surface enclosing a volume V , we recover
the divergence theorem: V ∇ · B dV = S B · dS, from examples 1.22 and 1.26.
Note that a metric is required for the translation from the Stokes-Cartan theorem to the divergence and
Kelvin-Stokes theorems in vector calculus.
1.8 Maxwell Differential Forms in 3 + 1 Dimensions

With F the Faraday 2-form, define two 3-dim p-forms: an electric field strength 1-form E and a magnetic field
strength 2-form B, by:
F = F|µν| dxµ ∧ dxν := E ∧ dt + B (1.64)
where:
E := F10 dx1 + F20 dx2 + F30 dx3 B := F12 dx1 ∧ dx2 + F31 dx3 ∧ dx1 + F23 dx2 ∧ dx3 (1.65)
Now, formally, d = ~d+dt ∧∂t . where ~d denotes the 3-dim exterior derivative. Then Maxwell’s dF = 0 becomes:

~d + dt ∧ ∂t E ∧ dt + B = ~dE ∧ dt + ~dB + dt ∧ ∂t B = ~dE + ∂t B ∧ dt + ~dB
= 0
The plus sign in the round brackets is the result of applying the commutation formula eq. (1.38) to the 1-form dt
and the 2-form B. In three dimensions, then, the homogeneous Maxwell equation gives rise to:
~dB = 0 ~dE + ∂t B = 0 (1.66)
Eq. (1.66) is metric-independent, and will thus hold in any spacetime in a coordinate basis.
The Hodge duals of eq. (1.66) can be written as:
div ⋆B = 0
⋆ ⋆ ~
⋆ dE + ∂t ⋆B = 0
⋆
If we identify the contravariant components of the pseudo-1-form ⋆B with the usual components of the magnetic-
⋆
field pseudo-vector, and use eq. (1.52), we see that these are equivalent to the homogeneous Maxwell equations in
their vector-calculus form: ∇ · B = 0 and ∇ × E + ∂t B = 0.
We see that it is much more natural to view the 3-dim magnetic field as a 2-form which is the exterior derivative
of a 1-form, than as a pseudo-vector which is the curl of another vector. and the electric field strength with the
1-form E than with the vector E. It is consistent with force and momentum also being more naturally 1-forms
µ
(consider eipµ x !).
The inhomogeneous Maxwell equation requires much more care, and is treated in Appendix C.
25
Appendices
A Tangent Spaces as Vector Spaces
First, let us specify what is meant by addition and s-multiplication on a tangent space TP .
Definition A.1. The addition operation on TP is a map, TP + TP −→ L(C ∞ (M ), R), such that,
∀ f ∈ C ∞ (M ) and any two curves (Γ, Θ) ∈ M intersecting at P ∈ M :
(v(Γ,P) + v(Θ,P) )(f ) := v(Γ,P) (f ) + v(Θ,P) (f )
Again, the addition operation on the left is between mappings, whereas that on the right is on R. As
for s-multiplication, it is a map, R × TP −→ L(C ∞ (M ), R), such that, ∀ a ∈ R:
(a · v(Γ,P) )(f ) := a v(Γ,P) (f )
The question now is: do these operations close? In other words, can we find some curve Θ ∈ M such that:
a · v(Γ,P ) = v(Θ,P) , and perhaps another curve Σ ∈ M such that: v(Γ,P) + v(Θ,P) = v(Σ,P) ?
To construct such a curve for s-multiplication, we first redefine the parameter of the curve Γ as the linear
function, µ : R −→ R, of λ: µ = aλ + λ0 , with λ now the parameter of a curve Θ such that Θ(λ) = Γ(µ)
Therefore, Θ(0) = Γ(λ0 ) = P. As in definition 1.5 we can write: Γ(µ) = Γ ◦ µ(λ). Insert this information into
the expression for the velocity for Θ at P:
v(Θ, P) (f ) = dλ (f ◦ Θ) λ=0
= dλ (f ◦ Γ ◦ µ) λ=0
= dµ (f ◦ Γ) µ(λ=0)=λ0
dλ µ λ=0
= a v(Γ, P) (f )
Therefore, we have found a curve Θ such that the operation a · v(Γ,P ) gives the velocity for that curve at P.
Up to now, in our discussion of tangent spaces, we have not needed any reference to coordinate charts. Unfor-
tunately, when it comes to proving that addition of two velocities in TP gives a velocity in TP , we cannot add the
curve mappings directly since this has no meaning. Instead, as was done in the previous section, assume that both
curves Γ and Θ are in some open subset U ⊂ M parametrised by coordinate functions x. Let Γ and Θ go through
point P at values λ1 and λ2 ot their respective parameter. Then construct a curve Σ parametrised in U by:
(x ◦ Σ)(λ) = (x ◦ Γ)(λ1 + λ) + (x ◦ Θ)(λ2 + λ) − (x ◦ Γ)(λ1 )
Although there might appear to be an obvious cancellation in this expression, it is not allowed because the coor-
dinate functions are not linear and thus do not distribute over the additions in Rn in the arguments on the right.
At λ = 0, however, the cancellation does occur, leaving Σx (0) = Θ(λ2 ) = P, so that our curve Σx runs through
point P at λ = 0.
We also need the derivative of the ν th x coordinate of the curve Σ, evaluated at P:
h
dλ (xν ◦ Σ) 0 = dλ (xν ◦ Γ)(λ1 + λ) + (xν ◦ Θ)(λ2 + λ) − (xν ◦ Γ)(λ1 )
0
ν ν
= dλ1 +λ (x ◦ Γ) λ1
dλ (λ1 + λ) 0
+ dλ2 +λ (x ◦ Θ) λ2
dλ (λ2 + λ) 0
ν ν
= dλ1 +λ (x ◦ Γ) λ1
+ dλ2 +λ (x ◦ Θ) λ2
(A.1)
Now go back to our expression (1.6) for the velocity in coordinates x. The first factor on the right has been
evaluated in eq. (A.1) and, running the chain of equalities in eq. (1.6) backward, there comes:
X h i X h i
vΣ,P) (f ) = ∂ν (f ◦ x−1 ) xν (P) dλ (xν ◦ Γ) λ1 + ∂ν (f ◦ x−1 ) xν (P) dλ (xν ◦ Θ) λ2
ν ν
−1
−1

= dλ (f ◦ x ) ◦ (x ◦ Γ) λ1
+ dλ (f ◦ x ) ◦ (x ◦ Θ) λ2
= v(Γ,P) (f ) + v(Θ,P) (f )
26
Thus, adding the velocities for two curves meeting at some point yields the velocity for some other curve in-
tersecting the others at that same point, and the tangent space of a curve at a point can indeed support a vector
space structure! Do note that the result does not depend on whatever coordinate chart we might have used in the
intermediate steps of the proof.†
B Transformation of Vector Components Between Coordinate Systems

Let (U1 , x) and (U2 , y) be two overlapping charts (see definition 1.4) on a manifold M , with x and y their coordi-
nate functions, respectively. Consider a point P ∈ U1 ∩ U2 .
Let us obtain the relation between ∂xµ x and ∂yν y , the coordinates bases for the two charts. These are
P P
maps, which we let act on some arbitrary differentiable function f . We remember that because f acts on the
manifold, we must write ∂xµ f x = ∂µ (f ◦ x−1 ) x . Insert y −1 ◦ y and use the multidimensional version of the
P P
chain rule (f ◦ g)′ (P) = g ′ (P) f ′ [g(P)] (written in the order opposite the usual one):
h i
∂xµ f = ∂µ (f ◦ y −1 ) ◦ (y ◦ x−1 )
xP xP
−1 ν
= ∂xµ (y ◦ x ) xP
∂yν (f ◦ y −1 ) (y◦x−1 )(xP)
= ∂xµ y ν xP
∂yν f yP
(B.1)
A vector v ∈ TP must remain invariant under change of chart. That is: v = vxµ ∂xµ x = vyλ ∂yλ y . Inserting the
P P
transformation law for the coordinate bases, we immediately find he transformation law for the components of v:
vyν = ∂xµ y ν xP
vxµ (B.2)
†
For a clear and accessible discussion of differentiability, manifolds and tangent spaces, see Frederic Schullers’s first five lectures at the
2015 International Winter School on Gravity and Light, available on YouTube.
27
C Three-dim Inhomogenous Maxwell Equations in the p-form Formalism

Going from the simple 4-dim formalism to three dimensions is more complicated than for the homogeneous equa-
tions, because the Hodge dual in the 4-divergence inevitably involves a metric, and because a 4-dim Hodge dual is
not necessarily like a 3-dim Hodge dual! First, we must derive an expansion of ∗ F in terms of E and B. A safe, if
somewhat inelegant, method is to expand it in terms of the components of F = 12 Fµν dxµ ∧ dxν :
∗ 1 µν
F = F ǫ dxα ∧ dxβ
4 p µναβ

= − |g| F 10 dx2 ∧ dx3 + F 20 dx3 ∧ dx1 + F 30 dx1 ∧ dx2 + F 12 dx3 + F 31 dx2 + F 23 dx1 ) ∧ dt
Now we must write this in terms of the covariant components of F, and this is where the metric must come in,
since F µν = g µα g νβ Fαβ :
F i0 = (g00 gij − gi0 g0j )Fj0 + g ij g0k Fjk , F ij = (gi0 g jl − g il gj0 )Fl0 + gik gjl Fkl
We know that Fj0 and Fjk are the components of the 3-dim p-forms E and B, respectively. If g0i 6= 0, each
contravariant component of F will involve both E and B, which will lead to very complicated results. When
g0i = 0, however, we are left with F i0 = g 00 gij Fj0 , and F ij = gik g jl Fkl , and lowering the spatial components
of F involves only the spatial sector of the metric (ignoring the g00 factor), the same sector that is used p to raise
00
indices on the Levi-Civita tensor. Also, if we take g = −1 (mostly positive) Minkowski metric, the |g| factor
is the same for the three-dimensional metric determinant as for the 4-dim one. Because of all this, we can now
write:
∗ 1 i0 j k 1 ij k
F = − ǫ F dx ∧ dx + ǫijk F dx ∧ dt
2 ijk 2
where the roman indices run from 1 to 3. Now we can relate the two terms to E and B:
1 1 1
ǫ F i0 dxj ∧ dxk = ǫijk g00 g il Fl0 dxj ∧ dxk = g00 ǫijk E i dxj ∧ dxk = g00 ∗ E = − ∗ E
2 ijk 2 2
Also:
1
ǫ F ij dxk = ∗ B
2 ijk
with no assumption needed for the spatial part of the 4-dim metric. Then our expansion is ∗ F = −∗ B ∧ dt + ∗ E
where it is understood that, on the right-hand side only, the 3-dim Hodge dual is taken. It is not difficult to show
(EXERCISE) that: d∗ F = −(~d∗ B − ∂t ∗ E) ∧ dt + ~d∗ E.
We define the Maxwell source pseudo-3-form as the expansion:
J ≡ ρ − j ∧ dt ≡ ρ ǫijk dxi ∧ dxj ∧ dxk − ∗ J ∧ dt (i < j < k)
where ρ is the charge scalar density, ρ the three-dim charge-density pseudo-3-form and J the 3-dim current density
1-form. Inserting these expansions in eq. (1.59) yields the two 3-dim Maxwell field equations:
~d∗ E = 4πρ, ~d∗ B = j + ∂t ∗ E (C.1)
Taking the 3-dim Hodge dual of these equations recovers the vector-calculus form of Gauss’s law for electricity
and the Ampère-Maxwell equation.
28
2 CHAPTER II — A BRIEF INTRODUCTION TO GROUP THEORY

One of the most beautiful and useful concepts in physics and, indeed, mathematics, symmetry identifies patterns
connected with a characteristic behaviour of objects, usually an invariance, under transformations. A problem
where a symmetry exists is amenable to much simplification and might even be solvable. Useful information can
be recovered even if the symmetry is only approximate, or is “broken” in a way that is understood. Equally impor-
tant, a continuous symmetry signals the existence of a conserved quantity. For instance, from space-translation
invariance (aka homogeneity of space) follows linear-momentum conservation, whereas time-translation invari-
ance gives rise to energy conservation; and isotropy of space (invariance under rotations) to angular-momentum
conservation. Conservation of electric charge is embodied in the local gauge invariance of Maxwell’s equations.
In modern mathematics, the language of group theory provides a unified and systematic framework for clas-
sifying and describing symmetries. In part because it is jargon-heavy, group theory is often relegated to the fringes
of most physicists’ training. Yet much insight can be gained from at least a modicum of familiarity with it.
2.1 Introducing the Notion of Group (BF 10.1)

2.1.1 Some basic definitions
Definition 2.1. Let G be a set of distinct objects endowed with an associative, but not necesssarily
commutative, binary composition law, or operation, denoted by ⋆ (or ◦). We say that G is a group if:
• ∀ (a, b) ∈ G, a ⋆ b ∈ G (this is called closure);
• there exists a unique element e ∈ G such that, ∀ a ∈ G, e ⋆ a = a ⋆ e = a;
• ∀ a ∈ G, there exists a unique element a−1 ∈ G such that a−1 ⋆ a = a ◦ a−1 = e.
The composition law is often called group multiplication, a term we shall try to avoid because it almost irresistibly
evokes the much narrower ordinary multiplication. There immediately follows a constraint on any composition
law. Let G = {ai }, with i a positive integer or continuous index. Then, for a fixed element ai , the set {ai ⋆ aj },
with j running over all the elements of G, must itself be G, ie., it must contain all elements of the group once, and
only once. Indeed, suppose that ai ⋆ aj = ai ⋆ ak for some j, k. Since ai must have a unique inverse, this forces
aj = ak . A similar argument can be made for fixed aj in {ai ⋆ aj }.
Definition 2.2. When ⋆ is commutative, ie. a ⋆ b = b ⋆ a, ∀ (a, b) ∈ G, we call G an Abelian group.
Definition 2.3. A group of n elements (n < ∞) is said to be finite and of order n. It is discrete if
it is countable, ie., if each element can be associated with a unique positive integer. All finite groups
are discrete, but infinite discrete groups exist. Non-discrete infinite groups are called continuous.
Example 2.1. Consider the set {e, a, a2 , . . . , an−1 }, where ap := a ⋆ a ⋆ . . ., and n is the smallest
integer such that an = e. The set is closed under ⋆, and (ap )−1 = an−p ∀ p. All ap are distinct, for
supposing ap = aq , we would have ap−q = e, with p − q < n, and n would not be the smallest integer
such that an = e. We conclude that the set is a group called Zn (sometimes Cn ), the cyclic group of
order n. When n is even, only an/2 is its own inverse; when n is odd, each element other than e is
paired with a distinct inverse. Thus, Zn can have at most one self-inverse element.
Let g belong to a group G of order n. There must be an integer m such that gm = e. Then we say that g itself
is of order m. If m < n the group is not cyclic, but {e, g, . . . , g m−1 } is a group Zm . g and its inverse have the
same order, ∀ g ∈ G. If all elements of a group are their own inverse (order 2), the group is Abelian (EXERCISE).
Some other groups: C under addition (e = 0, a−1 = −a); C − {0} under multiplication (e = 1, z −1 = 1/z);
the set, GL(n, C), of all complex n × n matrices with non-zero determinant under matrix multiplication; the n
complex roots of 1 under multiplication. Exercise: spot any discrete and cyclic groups in these examples.
It is important to keep in mind that a given set may be a group under one operation, but not under another.
Thus, Z is a group under addition with e = 0 and a−1 = −a, but it is not a group under multiplication.
29
2.1.2 Cayley tables

Let ai (i = 1, . . . , n) be an element of a finite group. By convention, a1 = e. We can construct a n × n
composition table, or Cayley table, whose (ij)th element is ai ⋆ aj . Then the first row and the first column must
be {e, a2 , . . . , an }. They are sometimes omitted by authors who are not nice to their readers.
To satisfy the above constraint on the group composition law, any column or row of a Cayley table must contain
all elements of the group once, and only once. The ordering of the rows and columns is arbirary.
Constructing Cayley tables for finite groups is easy, if tedious for large groups. Let us do it for n = 2, 3, 4:
e a b c
e a b
e a a b c e
a b e b c e a
a e
b e a c e a b
{e, a}
{e , a, b = a2 } {e , a, b = a2 , c = a3 }
The tables for n = 2 and 3 are the only ones possible. Thus, finite groups of order 2 and 3 are cyclic. The case
n = 4, however, opens up more possibilities: Choosing a ⋆ a = a2 = b, as above, the table is that of the cyclic
group Z4 . But we could take a2 = c; or a2 = e, in which case we can further choose b2 = a or b2 = e, yielding:
e a b c e a b c e a b c
a c e b a e c b a e c b
b e c a b c a e b c e a
c b a e c b e a c b a e
By re-labelling b ←→ c in the first table, and a ←→ b in the second, and re-ordering the rows and columns,
we obtain tables which are identical to the cyclic table, and we conclude that they are really those of Z4 .
The last table is genuinely different. It belongs to a group {e, a, b, a⋆b} called the 4-group—aka Felix Klein’s
Vierergruppe V —in which every element is its own inverse (so of order 2), with the fourth element constructed out
of the other two non-identity elements (otherwise V would be cyclic!). An example is D2 , the symmetry group of
a 2-d rectangle centered on the origin:, with the identity, one rotation by π, and two reflections about the axes as
elements.
The foregoing illustrates very nicely two important features of groups:
• Generators of a group
Definition 2.4. A set of generators of a group G is any subset of G from which all other elements
of G can be obtained by repeated compositions of the generators among themselves. G must
contain all the distinct compositions of its generators, including with themselves.
For instance, we can say that if a generates Zn , ap also generates Zn provided p and n have no common
divisor (EXERCISE). Then any such ap can be taken on its own as the generator of Zn . The 4-group is
obtained from two generators. EXERCISE: construct a Cayley table for the group: {e, a, b, b2 , a ⋆ b, b ⋆ a}.
Another example is a rotation by π/6 as the generator of the finite group of rotations by kπ/6 (0 ≤ k ≤ 11)
about the same axis.
EXERCISE: Is it possible to construct a group of order 6 with all its elements of order 2?
• Isomorphisms
We have just been introduced to the important idea that groups which look different may in some sense be
the same because their Cayley tables are identical or can be made to be identical by relabelling. We now
formalise this idea:
Definition 2.5. If there exists a one-to-one mapping between all the elements of one finite group
{G, ◦} and all the elements of another finite group {H, ⋆} such that under this mapping these
groups have identical Cayley tables, then the mapping is an isomorphism, and G and H are
isomorphic: G ∼ = H.
30
Another definition is more apt for continuous groups, which do not have a Cayley table as such:
Definition 2.6. If there exists a one-to-one mapping f between all the elements of one group
{G, ◦} and all the elements of another group {H, ⋆} such that under this mapping, f (a), f (b) ∈
H and f (a ◦ b) = f (a) ⋆ f (b) ∀ a, b ∈ G, then f is an isomorphism of G onto H, and G ∼ = H.
Other examples of isomorphic groups:
– the group of permutations of two objects (S2 ), the group of rotations by π around the z axis, and the
group {1, −1} (under multiplication);
– the group of complex numbers and the group of vectors in a plane, both under addition;
– the groups {R, +} and {R+ , ×} with the exponential as the isomorphism. Later we will see that
because ex ey = ex+y , ex ∈ {R+ , ×} provides a one-dimensional matrix representation of {R, +}.
Definition 2.7. A homomorphism, like an isomorphism, preserves group composition, but it is

not one-to-one (eg. it could be many-to-one).
2.2 Special Subsets of a Group (BF10.3)

There are a number of useful ways to classify the elements of a group. We look at three of them.
2.2.1 Special Ternary Compositions: Conjugacy Classes
Definition 2.8. Given a ∈ G, any element b ∈ G which can be obtained as b = x ◦ a ◦ x−1 , where
x ∈ G, is called the conjugate of a by x. This conjugation operation, which consists of two binary
compositions, has the following properties:
• Reflexivity: a = e ◦ a ◦ e−1 , or a is self-conjugate.

• Symmetry: let b = x ◦ a ◦ x−1 . Then a = y ◦ b ◦ y −1 , with y = x−1 ∈ G.
• Transitivity: let b = x ◦ a ◦ x−1 and a = y ◦ c ◦ y −1 . Then, since x ◦ y ∈ G, b is conjugate to c :
b = x ◦ a ◦ x−1 = x ◦ y ◦ c ◦ y −1 ◦ x−1 = (x ◦ y) ◦ c ◦ (x ◦ y)−1

Definition 2.9. The subset of elements of a group which are conjugate to one another form a conju-
gacy, or equivalence† , class, often abbreviated to just class. The systematic way of constructing the
class for any element ai of a group is to form the set:
{e ◦ ai ◦ e−1 , a1 ◦ ai ◦ a−1 −1 −1
1 , . . . , ai−1 ◦ ai ◦ ai−1 , ai+1 ◦ ai ◦ ai+1 , . . .}
Then e is always in a class by itself, and each element of an Abelian group is the sole element in its class. eg.,
Zn and the four-group.
Classes are disjoint: they have no common element (EXERCISE: show this). Thus, they partition the group.
Elements in the same class share some properties. In particular, they must all be of the same order (EXER-
CISE). In a particularly important type of group, matrix groups, conjugate matrices are similar to one another; they
could represent the same “thing” in different bases.
EXERCISE: obtain the classes for the group: {e, a, b, b2 , a ⋆ b, b ⋆ a}.
†
Actually, conjugacy is only a particular type of equivalence.
31
2.2.2 Subgroups
Definition 2.10. A subset H of a group G that behaves as a group in its own right, and under the same
composition law as G is said to be a subgroup of G: H ⊆ G. H is proper if it is non-trivial (ie. not
e) and if H ⊂ G (ie. H 6= G). The subgroups of a group may have more elements than e in common.
We have already seen that any element g of order m < n of G generates a cyclic subgroup Zm ⊂ G.
Example 2.2. The four-group V has the proper Z2 subgroups: {e, a}, {e, b}, and {e, c = a ⋆ b},
which are isomorphic. By inspection, the group of order 6 {e, a, b, b2 , a⋆b, b⋆a} contains the proper
subgroup Z3 = {e, b, b2 }.
Notation alert: Henceforth, we drop the cumbersome star (circle) whenever there is no risk of confusion with
usual multiplication. Also, if H and H ′ are two subsets of {G, ⋆}, we can write H H ′ for {h h′ } h ∈ H, h′ ∈ H ′ .
Let us try out our new notation on the following definition:
Definition 2.11. A subgroup N ⊆ G is invariant (or normal) if N = G N G−1 or, more precisely,
if g h g −1 ∈ N ∀ h ∈ N and ∀ g ∈ G. Alternate notation: N ✁ G, G ✄ N .
EXERCISE: Show that ∀ gi ∈ G the set with distinct elements gi−1 gj−1 gi gj forms an invariant subgroup of G.
Definition 2.11 is sometimes written G N = N G, but it does not mean that an invariant subgroup must be
Abelian (though it can be). It means that if hi ∈ N and g ∈ G, there is some element hj ∈ N such that g hi = hj g.
Example 2.3. Because the four-group V is Abelian, its non-trivial subgroups, {e, a}, {e, b}, {e, a b},
are all invariant. Subgroups of any Abelian group are invariant.
Since classes and normal groups are both defined by conjugation, it is hardly surprising that they are related.
Indeed, let H ⊂ G. Then H is invariant if and only if it contains complete classes, ie. if it is a union of classes of
G. Indeed, if H is invariant, all the conjugates (elements in the same class) of any h ∈ H are also in H; this holds
for all classes, which are disjoint; so only complete classes can be in H. Conversely, let a subgroup H ⊂ G be a
union of complete classes; therefore, g h g −1 ∈ H ∀ g ∈ G, which is precisely the definition of a normal subgroup.
Definition 2.12. A simple group has no invariant subgroup other than itself and the identity.
2.2.3 Cosets (BF 10.3)
Definition 2.13. Let H be a subgroup of G, and let g ∈ G. Then g H is a left coset of H for a given
g, and H g is a right coset of H. The set of all left (right) cosets of H is called the left (right) coset
space of H. Every coset g H must contain the same number of elements, equal to the order of H.
If H is invariant, to any ot its left cosets corresponds an identical right coset, and vice-versa, as follows
immediately from Def. 2.11. In particular, the right and left cosets of any Abelian subgroup are identical.
Example 2.4. Let G = R3 under addition, and H be a plane containing the origin. For a given vector
a, a + H ∈ H if a ∈ H; otherwise, a + H is another plane parallel to H, and we would say in this
language that it is a left (or right) coset of H through the origin. And H itself would also be a coset.
The most important property of cosets is that they are either disjoint or else identical. Thus, we can say that
the coset space of a subgroup H ⊂ G provides a partition of G.
Indeed, let g1 h1 = g2 h2 for some (h1 , h2 ) ∈ H and (g1 , g2 ) ∈ G. Therefore, g1 = g2 h2 h−1 1 . Now
consider some other element of the same coset, g1 h3 (h3 ∈ H); then g1 h3 = g2 (h2 h−1 1 h3 ) = g h
2 4 , where
−1
h4 = h2 h1 h3 ∈ H. That is, if two elements of different cosets are the same, then any other element, say g h3 ,
in the first coset, must be equal to some element of the second coset. Since the same argument holds when we
switch g1 and g2 , we conclude that if g1 H and g2 H have one element in common, they have all their elements in
common and are thus identical. The same proof applies to right cosets.
It follows (why?) that e H = H is the only coset of a subgroup H that is a group.
32
2.2.4 Lagrange’s Theorem and quotient groups

If H ⊂ G, every element of G must occur either in H or one (and only one) of its other cosets. This forms the
foundation of the proof of Lagrange’s Theorem: The order n of a finite group is an integer multiple of the order
m of any of its subgroups. Indeed, since every element of the group is either in the subgroup or in one of its other
k distinct cosets, each with m elements, (k + 1)m = n. The ratio n/m is called the index of the subgroup.
Let a ∈ G. Clearly, it generates a cyclic subgroup of G of order m, where m ≤ n is the order of a. Therefore,
the order n of G must be an integer multiple of the order m of any of its elements. If n is prime, m = n or m = 1,
and we have proved that the only non-trivial finite group of order prime is the cyclic group, eg., Z5 , Z7 , and that
such a group has no non-trivial subgroup.
Also, if its order n is odd, a group (not only the cyclic ones!) cannot contain any self-inverse element, for such
an element must generate a Z2 subgroup whose order, 2, is forbidden by Lagrange’s Theorem.
The converse of Lagrange’s Theorem does not hold in general: eg., the group of even permutations of four
objects, A4 , with 12 elements, has no subgroup of order 6. The theorem only gives the possible orders of subgroups.
There are stronger conditions on the order and number of the subgroups of G stipulated by the Sylow Theorems,
which lack of time prevents us from exploring.
Now consider the set whose elements are the subgroup as a whole and all its other cosets, each also as a whole:
Definition 2.14. The set of all left cosets of H ⊂ G, each considered as a whole, is called a factor
space for H. Note that the elements of this space are the cosets themselves, each considered as a
whole, not any individual element within a coset.
Factor spaces of a subgroup H are not necessarily groups; but there is one important exception:
Definition 2.15. To an invariant subgroup N of G is associated a factor group of G, G/N . Its

elements are N and all its cosets as sets (not the elements of N or of the cosets!). Its order is the order
of G divided by the order of N , hence the name quotient group often used for G/N .
To show that the factor space of an invariant subgroup is a group, we note that for any coset g N , (g N ) N =
g N N = g N , and N (g N ) = g N N = g N , where we have used the associativity of the group product
and the invariant nature of N . This establishes N as the identity of the factor group. The composition law follows
from:
(g1 N ) (g2 N ) = g1 g2 N N = (g1 g2 ) N
since g N = N g ∀ g ∈ G. Lastly, (g N ) (g −1 N ) = g g−1 N N = N = e. So g−1 N is the inverse of g N .
Factor groups can be useful when we do not need to distinguish between the elements of subgroups of a group.
Much of the usefulness of normal subgroups comes from providing a quick way to find factor groups.
2.2.5 Direct Products
Definition 2.16. Let H1 and H2 be subgroups of G with H1 ∩ H2 = e, and let h1 h2 = h2 h1 ∀ h1 ∈

H1 , ∀ h2 ∈ H2 . If it is possible to write g = h1 h2 ∀ g ∈ G, then G ≡ H1 × H2 is said to be the
internal direct product of its subgroups H1 and H2 .
Example 2.5. The four-group introduced in section 2.1.2 can be seen as Z2 ×Z2 , or {e, a}×{e, b} =
{e, a, b, a b}. But Z4 6= Z2 × Z2 , even though Z2 is a normal subgroup of Z4 , with Z2 = Z4 /Z2 !
Another well-known way of constructing a (this time, external) direct product of, say, two a priori unrelated
matrix groups with elements A ∈ H1 and B ∈ H2 would be:

A 0 A 0 I 0
=
0 B 0 I 0 B
Or we could construct {1, −1}×{1, −1} = {(1, 1), (1, −1), (−1, 1), (−1, −1)}. the external direct product of
Z2 with itself, in this realisation. This, of course, is the four-group (with normal multiplication as group product).
33
2.3 The Mother of All Finite Groups: the Group of Permutations

2.3.1 Definitions, cycles, products
The most important finite group is the group of permutations of n objects, Sn , aka the symmetric group, which
contains n! elements corresponding to the n! possible rearrangements of the objects. A permutation is by definition
a bijective mapping. Following a standard convention, we notate, with 1 ≤ k ≤ n!:
Definition 2.17.
1 2 3 ... n
πk =
πk (1) πk (2) πk (3) . . . πk (n)
The horizontal ordering of the initial objects is immaterial. Also as a matter of convention, we agree
that it is the objects in the slots which are rearranged, not the slots. Finally, we do not have to use
numbers as labels, but they offer the greatest range.
In a permutation, an object i may be mapped into itself, ie. it stays in the same slot. But more typically object
i is mapped to j, while j is mapped to k; and so on along a chain that ends back at object a after l steps. When
this occurs, we speak of a l-cycle. More precisely:
Definition 2.18. Let πk ∈ Sn , and let l be the smallest integer for which [πk (j)]l = j, for some
1 ≤ j ≤ n. Then the sequence of objects in [πk (j)]l is called a l-cycle (sometimes a r-cycle. . . ).
This suggests a much more compact notation for πk , one in which we bother to write only the l-cycles (l > 1), and
consider a given permutation as the product of simpler permutations.
As an example, we write:

1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
= ≡ (1 5) (2 4 3)
5 4 2 3 1 6 5 2 3 4 1 6 1 4 2 3 5 6 1 2 3 4 5 6
It is easy to see the advantages of the cycle notation introduced at the end of the line! Note that the cycles are
disjoint. Any permutation can be, and usually is, represented by a sequence of disjoint cycles. Warning: do not
confuse the symbols in a l-cycle with the outcome of a permutation in Sn !
Any πk ∈ Sn can always be written as the product† of transpositions, or two-cycles. Indeed, a l-cycle may
always be decomposed as a product of l − 1 transpositions, but these are not disjoint. An element of Sn and its
inverse have the same cycle structure.
Definition 2.19. A permutation is even (odd) if it is equivalent to an even (odd) number of transposi-
tions, or switches; thus, a l-cycle which contains an even number of symbols is equivalent to an odd
permutation, and vice-versa. An even permutation is said to have parity 1, and an odd permutation
parity −1. We expect that parity will put strong constraints on the group product table of Sn .
Single transpositions always have odd parity. The mapping from Sn to the parities {1, −1} is a nice example of a
homomorphism.
Definition 2.20. A cyclic permutation of length l has a single cycle of length l > 1.
In cycle notation, S2 = {e, (1 2)} and S3 = {e, (1 2), (1 3), (2 3), (1 3 2), (1 2 3)} ≡ {π1 , π2, π3 , π4 , π5 , π6 },
are the smallest non-trivial symmetric groups. For S3 , note the three-cycles (1 2 3) = 12 23 31 and (1 3 2) =
1 2 3

3 1 2 . I have deliberately changed the order of the latter from what it is in BF, but if you write out the corre-
sponding permutation in full notation for BF’s (3 2 1), you will see that it is identical to mine. So long as we cycle
through in the same direction (here, to the right), where we start the cycle does not matter! It can be shown that S3
and Z6 are the only groups of order 6, up to isomorphisms.
†
Since there is little scope for confusion in the context of Sn , we replace “group composition” with “group product”.
34
2.3.2 Some subgroups of Sn

One obvious subgroup of Sn is the so-called alternating group, An , of all its even permutations. Odd permutations
do not form a group, because their product is an even permutation. Other important subgroups of Sn are the cyclic
groups of order n, generated by the permutations of all objects which return the initial state to itself after n
products.
Now for subgroups of S3 : Lagrange’s Theorem allows only non-trivial proper subgroups of order 2 or 3. The
alternating subgroup A3 is read off the list of the elements of S3 : {e, (1 3 2), (1 2 3), } which must be cyclic
because all groups of order 3 are isomorphic to Z3 . Note: this is not a general feature as the cyclic subgroups of
higher order generated by odd permutations in Sn>3 contain permutations of both even and odd parity.
Transpositions being self-inverse, the other (isomorphic!) subgroups of S3 are {e, π2 }, {e, π3 }, and {e, π4 }.
2.3.3 Cayley table of S3 as an example

The group-product table of S3 contains 36 entries, “only” 25 of which are non-trivial, from which we have just
found three. But I claim that no more than one other product needs to be worked out with explicit permutations.
Indeed, the entries of the 2 × 2 sub-table for the π5 and π6 rows and columns must be even permutations
(they are the group product of even permutations). The diagonals cannot be e, so as to avoid repetition. Next,
the non-diagonal elements of rows and columns corresponding to π2 , π3 and π4 must be π5 or π6 , the only even
permutations other than e. To fill in this sector only requires calculating one group product, say, π2 π3 :
 
1 2 3
1 2 3 1 2 3   1 2 3
π2 π3 = = 3 2 1 = = (1 3 2) = π5
2 1 3 3 2 1 3 1 2
3 1 2
The other unfilled entries in rows and columns for π5 and π6 must be either π2 , π3 , or π4 . For columns π5 and
π6 , applying π2 to π2 π3 gives π3 = π2 π5 , which determines the rest from the table-building rules. Similarly,
π2 π3 π3 = π2 = π5 π3 , and the rest of the π5 and π6 rows is determined. The comes (in two equivalent forms):
e π2 π3 π4 π5 π6 e π5 π6 π2 π3 π4
π2 e π5 π6 π3 π4 π5 π6 e π4 π2 π3
π3 π6 e π5 π4 π2 π6 e π5 π3 π4 π2
≡
π4 π5 π6 e π2 π3 π2 π3 π4 e π5 π6
π5 π4 π2 π3 π6 e π3 π4 π2 π6 e π5
π6 π3 π4 π2 e π5 π4 π2 π3 π5 π6 e
2.3.4 Cayley’s Theorem

Why is Sn so important? As so often, the Cayley table of a group G of order n gives the key to the answer.
∀ ai ∈ G, the row {ai aj } (1 ≤ j ≤ n) is merely a bijective rearrangement of {ai }, that is:

a1 a2 . . . an a1 a2 ... an
ai 7−→ πai = , ai aj 7−→ πai aj =
ai a1 ai a2 . . . ai an ai aj a1 ai aj a2 . . . ai aj an
But we can also write:

a1 a2 . . . an aj a1 aj a2 ... aj an
πai = =
ai a1 ai a2 . . . ai an ai (aj a1 ) ai (aj a2 ) . . . ai (aj an )

aj a1 . . . aj an a1 . . . an a1 a2 ... an
=⇒ πai πaj = =
ai aj a1 . . . ai aj an aj a1 . . . aj an ai aj a1 ai aj a2 . . . ai aj an
What we have shown is that πai πaj = πai aj ; in other words, by definition 2.6, permutations preserve the group
product of G, and we have Cayley’s Theorem:
35
Every group of order n is isomorphic to a subgroup of Sn whose elements (except for e) shuffle all objects in
the set on which it acts.
We have already seen an example of this: the single instance of the cyclic group of order 3 is a subgroup of S3 .
EXERCISE: How many distinct instances of Z4 ⊂ S4 are there? How many of the four-group?
2.3.5 Conjugates and Classes of Sn

To find the classes of Sn , we must form, for each πi ∈ Sn , all its conjugates πj πi πj−1 . This seemingly daunting
task can actually be performed fairly easily, thanks to the natureof Sn . To keep the following
manipulations
as
1 2 ... n 1 2 ... n
uncluttered as possible, let us write πi = a and πj = b, with a = a1 a2 ... an and b = b1 b2 ... bn . Then:

−1 1 2 ... n 1 2 ... n b1 b2 . . . bn 1 2 ... n b 1 b 2 . . . bn
bab = =
b1 b2 . . . bn a1 a2 . . . an 1 2 ... n b1 b2 . . . b n a 1 a 2 . . . a n

a1 a2 . . . an b1 b2 . . . bn b1 b2 . . . bn
= =
ba1 ba2 . . . ban a1 a2 . . . an ba1 ba2 . . . ban

How did we obtain baa1 baa2 ... an
in the second line from 1 2 ... n
b1 b2 ... bn in the last member of the first line? Well,
1 2 ... ban
a1 must occur in some slot on the top line of the latter; since the order of the slots in the permutation is arbitrary,
we move that slot to first position and rename the upper element a1 . Then we do the same for 2 → a2 , etc. The
bottom elements are then the outcome of permuting ai with permutation b to get bai .
This leads to the important result: All permutations in a class have the same cycle structure, not only for Sn ,
but for all finite groups because of Cayley’s theorem. Classes being disjoint, each class of Sn is associated with a
unique cycle structure of its elements. But in groups other than Sn , although all elements in a class have the same
cycle structure, elements with the same cycle structure may belong to different classes (eg. A4 ⊂ S4 , Z4 ⊂ S4 ).
Also, all elements of a class of Sn must have their inverse in the same class; can you see why?
Take S3 as a simple example. As classes we only have C1 = {e}, C2 = {(1 2), (1 3), (2 3)}, and C3 =
{(1 3 2), (1 2 3)}. Since A3 = {e, (1 2 3), (1 3 2)} is the only non-trivial subgroup of S3 that is the sum of
complete classes, C1 + C3 , A3 is the only normal‡ proper subgroup of S3 .
Now consider S4 . There are two other permutations with the same cycle structure as (1 2)(3 4): (1 3)(2 4) and
(1 4)(2 3). Apart from this and the separate class {e}, the other classes of S4 are easily obtained as (1 2) and its 5
conjugate transpositions, (1 2 3) and its seven conjugates, and (1 2 3 4) and its five conjugates.
It is an instructive EXERCISE to show that A4 has no subgroup of order 6 despite this being allowed by
Lagrange’s Theorem.
Hint: after convincing yourself that such a group would take the form {e, a, b1 , b−1 1 , b2 , b2 }, with a a double transposition
−1
and (b1 , b2 ) 3-cycles, show that because the latter are of order 3, b1 b2 is neither b−1
1 nor b2 , which leaves only one possibility
−1
since b1 and b2 cannot be each other’s inverse. Is that possibility allowed by the parity of these elements?
In the literature, classes of Sn are routinely identified by partitions of n reflecting their cycle structure. Thus,
a given class will be written (iαi . . . j αj ), with (1 ≤ (i, j) ≤ n), where αi is the number of i-cycles in the class.
Start with e, whose cycle structure can be written as a product of n 1-cycles: e = (1) (2) · · · (n). So its class,
which always exists, would be denoted by 1n . A transposition has one 2-cycle and n-2 1-cycles, and Sn must
contain n(n − 1)/2 n−2 ). An arbitrary permutation involves
Pnof them (eg., six for S4 as above); it is denoted by (2 1
αi i-cycles, and i i αi = n. In that sense the cycle structure of a class corresponds to a partition of n.
Once we have noticed this correspondence, it becomes rather easy to find the number and cycle ctructure of
Sn classes. We adopt the usual convention that represents the cycle structure of a class by (λ1 ≥ λ2 ≥ · · · ≥ λn ),
where the λi sum up to n. Thus, the only possible partitions of S3 lead to classes (13 ), (2 1), and (3), ie. a class
with three 1-cycles (the identity), a class with one 2-cycle and one 1-cycle (the transpositions), and a class with
one 3-cycle. As for S4 , the possible partitions of 4 give rise to the five classes (14 ), (2 12 ), (22 ), (3 1), and (4).
It is important not to confuse the cycle notation we first introduced with this standard notation which lists all
the cycles in a class as a whole, including 1-cycles when they occur (whereas the other one ignores them).
‡
Note that this subgroup being Abelian is not sufficient to make it invariant; it must be self-conjugate with respect to all elements in S3 .
36
To find the number of elements in a class of Sn , count the distinct ways of partitioning n numbers into its cycle
structure:
n!
(2.1)
α1 ! . . . αn ! 1α1 . . . nαn
where αi ! is the number of non-distinct ways of ordering αi commuting cycles of a given length, and iαi is the
number of equivalent orderings of the symbols inside each i-cycle occurring αi times. From this expression it
should be easy to recover the number of elements in each class of S4 as given above.
Now we can identify (EXERCISE) the invariant subgroups of S4 without writing down its 24×24 Cayley table!
2.3.6 Graphical representation of classes: Young frames

A useful and visual way of representing the classes of Sn is to take n squares and arrange them in rows and
columns so that each column corresponds to an i-cycle and the number of boxes cannot increase from one column
to the next on the right, and from one row to the one next below. The game then consists in building all possible
arrangements that satisfy this constraint. For instance, with S4 , the possibilities are as follows:
Then we just read off the cycle structure for each: (14 ), (2 12 ), (22 ), (3 1), and (4), respectively. Finding the classes
of such monsters as, say S8 , no longer seems so intimidating. These diagrams are known as Young frames.
2.3.7 Cosets of Sn
Finding the left cosets of the subgroups of S3 is as easy as reading rows in its Cayley table. Take the sub-
group H = {e, π2 }; its left cosets by πk are πk {e, π2 } = {πk , πk π2 } (1 ≤ k ≤ 6). Only three are distinct:
{e, π2 }, {π3 , π6 }, {π4 , π5 }. Following Definition 2.14, this set of cosets is the factor space for H. The same
arguments apply to the subgroups {e, π3 } and {e, π4 }.
Turn now to the remaining non-trivial proper subgroup, A3 = {e, π5 , π6 }, of all even permutations in S3 .
Its left cosets are {πk , πk π5 , πk π6 }. For instance, π2 {e, π5 , π6 } = {π2 , π3 , π4 }, which is identical to the
other cosets π3 {e, π5 , π6 } and π4 {e, π5 , π6 }. Also, e {e, π5 , π6 } = π5 {e, π5 , π6 } = π6 {e, π5 , π6 }, as expected.
So another partition of S3 is provided by {e, π5 , π6 } + π2 {e, π5 , π6 }. Note that these left and right cosets
are identical, another way of saying that {e, π5 , π6 } is invariant, as we had found by simpler means. Then

{e, π5 , π6 }, {π2 , π3 , π4 } is the factor group of S3 . From the Cayley table for S3 , we see that the element
{e, π5 , π6 } is the identity, and that this factor group S3 /A3 is isomorphic to Z2 . It is easy to show that Z2 is a
factor group of Sn ∀ n. Equivalently, An is always a normal subgroup of Sn .
37
2.4 Representations of Groups

We have already mentioned that groups can be associated with symmetries, but we have to make this connection
explicit in the language of group theory. We wish to flesh out the rather abstract ideas and tools we have introduced.
We shall find that linear operators on vector spaces (most often, on a Hilbert space) provide us with this connection.
2.4.1 Action of a group from the left and from the right
Let G be a group of linear transformations Tg on square-integrable functions that live in a scpace called the carrier
space Introduce a set of operators, {Tg }, with each Tg acting on a set of parameters that characterise g ∈ G.
Definition 2.21. We distinguish between an action from the left, [Tg f ](x) := f (g−1 x), and an action
from the right, [Tg f ](x) := f (x g), ∀ f . Note that, here, the operators Tg act on the functions (not on
x !).
Why did we define the left action of g ∈ G as g−1 x, and not g x? Denote by Tgi gj the transformation
associated with gi gj ∈ G. Then, with gi = i and gj = j in subscripts so as to declutter the notation:

Tij f (x) = f (gi gj )−1 x = f (gj−1 gi−1 x) = Tj f (gi−1 x) = Ti Tj f (x)
which means that the T operators do form a group; but what if instead:

Tij f (x) = f (gi gj x) = Ti f (gj x) = Tj Ti f (x)
Something awkward has happened: if we write the left action as g x, the associated transformations do not form a
group! And, as you should verify, neither do they if we write the right action as x g−1 .
So, as a matter of notational consistency, we should always write x g for the right action and g−1 x for the left
action, which is indeed what BF do (without much explanation) for the left action.
2.4.2 Matrix representations of a group (BF10.4)
Definition 2.22. A representation D of a group G is a homomorphic mapping of the group ele-

ments onto a set of finite-dimensional invertible matrices such that D(e) = I, the identity matrix, and
D(gi ) D(gj ) = D(gi gj ), in the sense that matrix multiplication preserves the group composition law.
If the homomorphism is one-to-one, a representation is faithful. The dimension of the representation
is the rank of its matrices or, equivalently, the dimension of the carrier space on which it acts.
Even addition can be represented by matrix multiplication: Dα Dβ = Dα+β , with α and β two values of a group
parameter, eg. the matrix Dv = v1 01 . Do you recognise the transformation that applies it to the vector xt ?
The matrices GL(n, C) of rank n can be thought of as the set of all invertible linear transformations on a
vector space of complex-valued functions V = {f (x)}. We have: x = xi ei , with {ei } a basis for V and xi the
components of x in the basis; the subscript on basis vectors labels a whole vector, not a component of the vector.
Let us focus on the transformations Tg (x). Then the left action of g ∈ G is expressed as:

Tg (x) = g−1 x = xi g−1 ei = xi ej (DgL−1 )j i = ej (DgL−1 )j i xi (2.2)
It is an instructive exercise to show that the proper way of expressing the right action, x g, of the same group
in terms of its (right) representation DR matrices is:

x g = ei xi g = (DgR )i j xj ei (2.3)
in which DR i
g acts on the the x written as a column vector. Because of this, some see the right action as the more
“natural” one. For a given g, a right DL and DR is in general each other’s inverse.
38
2.4.3 Non-unicity of group representations

One might hope to define an algorithm that would churn out the representation of each element of a group. But
there is no such thing as a unique representation! Indeed, suppose we have a set of n-dimensional matrices which
represent a group. It is always possible to obtain another representation, of dimension 1, by mapping these matrices
to the number 1. This is called the identity representation, and it always exists. Also, the homomorphic map
of the same matrices to their determinant preserves the group product since det (AB) = (det A)(det B), which
provides another one-dim representation. Of course, no one will claim that such representations are faithful. . .
Also, we can make a change of basis: e′i = ej S j i , or ei = e′ j (S −1 )j i . Then we have the similarity
transformation: D′ g = S Dg S−1 , and the D′ obey the same product rules as the D matrices.
Definition 2.23. Representations connected by a similarity transformation are said to be equivalent

if the transformation matrix is the same for all group elements. They differ only by a choice of basis.
Example 2.6. Consider the continuous group, called SO(2), of rotations around the z axis. We
parametrise a rotation by g = Rα such that Rα φ = φ − α. This corresponds to a counterclockwise
rotation of the standard basis in R2 by α (passive transformation), so that a vector initially at angle φ
is at angle φ − α in the rotated basis; equivalently, rotate the vector by −α in the initial basis.
We find representations for the left action. One method uses Def. 2.21 (with Tg = Rα ) and eq. (2.2):

Rα fi (φ) = fi Rα−1 φ = fi (φ + α) = Di j (−α) fj (φ)
We look for a set of functions of φ which, under Rα , transform into linear combinations of themselves.
Try f1 = cos φ, f2 = sin φ. Then:

Rα f1 (φ) = cos(φ + α) = (cos α) cos φ − (sin α) sin φ = (cos −α) f1 (φ) + (sin −α) f2 (φ)

Rα f2 (φ) = sin(φ + α) = (sin α) cos φ + (cos α) sin φ = − (sin −α) f1 (φ) + (cos −α) f2 (φ)
Compare this with Di j (−α) fj (φ), and switch the sign of α to obtain the left D(α) matrix:

(1) cos α sin α
D (Rα ) =
− sin α cos α
Well, that’s the 2-dim left defining (fundamental) representation for SO(2), probably the most often
used. But it is not the only one! If instead f1 = eiφ , f2 = e−iφ , the same procedure would yield:
iα
(2) e 0
D (Rα ) =
0 e−iα
so here is another two-dim
representation. But it is equivalent because the transformation S−1 D(1) S,
√
with the matrix S = 1i 1i / 2, diagonalises D(1) into D(2) , ∀ α, ie. for all elements of SO(2).
And there are more: each linearly independent function eiα and e−iα is also a perfectly acceptable
one-dimensional representation of SO(2)! Both D(1) and D(2) can be viewed as a joining of these
one-dimensional representations, which we shall call D(3) and D(4) . Obviously, there is something
special about e±iα . Before we discover what it is, let us look at another instructive example.
Example 2.7. Let us work out a three-dimensional representation of the left action of S3 , π −1 x, on
R3 . Since Sn merely shuffles the components of x it preserves its length, which is the definition of
−1 i j −1 i j
is their inverse. In fact, S3 ⊂ O(3)! Then, from eq.
orthogonal matrices, ie., those whose transpose
(2.2), πk x = x ej D i (πk ) = x Di (πk ) ej so as to view the permutations as a shuffling of the
components of x (written as row vectors—see the Appendix at the end of the chapter!), and we have:
     
1 0 0 0 1 0 0 0 1
D(1) (π1 ) = 0 1 0 , D(1) (π2 ) = 1 0 0 , D(1) (π3 ) = 0 1 0 ,
0 0 1 0 0 1 1 0 0
39
     
1 0 0 0 1 0 0 0 1
D(1) (π4 ) = 0 0 1 , D(1) (π5 ) = 0 0 1 , D(1) (π6 ) = 1 0 0
0 1 0 1 0 0 0 1 0
Such a faithful, n-dim left defining (fundamental) representation can be constructed for any Sn .
Now, I claim that there exists another (two-dimensional!) representation of S3 , which is not faithful:

(2) (2) (2) 1 0
D (π1 ) = D (π5 ) = D (π6 ) =
0 1

(2) (2) (2) 0 1
D (π2 ) = D (π3 ) = D (π4 ) =
1 0
Indeed, the products of these matrices are consistent with the group product of S3 in its Cayley table.
Even less faithful, but no less acceptable, is the one-dim representation of Sn obtained by mapping its
permutations to their parity values. For S3 :
D(3) (π1 ) = D(3) (π5 ) = D(3) (π6 ) = 1
D(3) (π2 ) = D(3) (π3 ) = D(3) (π4 ) = − 1
And, of course, we can always map all the πi to 1 and get another (trivial) representation!
On the other hand, we could join D(1) and D(2) into a D(4) = D(1) ⊕ D(2) (direct sum) representation
whose six matrices are 5-dimensional and block-diagonal, each with the submatrices on the diagonal
taken, one from D(2) (the upper one, say), and the other from D(1) , for a given permutation πi .
2.4.4 The regular representation of finite groups
Definition 2.24. The regular representation of the left action of a finite group G is the set of matrices
DLg , with g ∈ G, derived from a group product, such that:
(
L j j 1 g gi = gj
Dg gi = g gi = gj D i (g) ∀ g ∈ G D i (g) =
0 g gi 6= gj
The regular representation is seen to be closely related to the Cayley table of the group. Its dimension
is equal to NG , the order of the group, and it is faithful. We can also see that D j i (e) = δj i , ie.
DL (e) = I. Also, the other matrices in the representation must have a 1 as their (ji)th element and 0
for all other elements in row j and column i; by inspection, this 1 is never on the diagonal.
Similarly, we define a regular representation, DR
g , for the right action of a group:
(
1 gj g = gi or gi g−1 = gj
DR g gi = gi g
−1
= Di j (g) gj Di j (g) =
0 gj g 6= gi or gi g−1 6= gj
A word of caution: do not confuse the dimension of a representation, ie. of its carrier space (the space of
functions on which group operators act), with the dimension of the coordinate space on which these functions act.
2.4.5 Unitary representations (BF10.6)
A representation Dg is unitary if D†g = D−1 −1 ∗
g ), ∀ g ∈ G. In terms of matrix elements, Dij (g ) = Dji (g). For
example, D(3) and D(4) for SO(2) are unitary. Left and right regular representations are also unitary.
Now, if Dg is not already unitary, we can always find a similarity transformation matrix P S,† the Hermitian
2
square root of the positive semi-definite (ie., with eigenvalues λn > 0) matrix: S = g Dg Dg , such that
D′g = S Dg S−1 is unitary EXERCISE—first, show that D†g′ S2 Dg′ = S2 , then apply S−1 on the left and on

the right . Any representation of a finite group is equivalent to a unitary representation. This is also true for
certain infinite (continuous) groups, such as compact Lie groups.
40
2.4.6 Invariant Spaces and Kronecker sum

To understand what relationship may exist between representations, it is time to bring in a very useful concept:
Definition 2.25. Let {f (i) } be a subspace H(i) of the carrier space H of functions on which the linear
transformations Tg associated with a group G act. If, ∀ f (i) ∈ H(i) and ∀ g ∈ G, Tg f (i) ∈ H(i) ,
the subspace is invariant under G. Also, it can be shown that if a subspace of the carrier space is
invariant under a unitary representation, its complement must also be invariant.
Definition 2.26. Let H(1) and H(2) be subspaces of a Hilbert space H such that H is the sum of the
two subspaces with zero intersection. Then, if any function in H can be written uniquely as the sum
of a function in H(1) and another in H(2) , H is called the Kronecker (or direct) sum of H(1) and
H(2) , written H = H(1) ⊕ H(2) . The dimension of H is the sum of the dimensions of H(1) and H(2) .
2.4.7 Reducible and irreducible representations (BF10.5)

Definition 2.27. If some function space H has a proper subspace invariant under a representation D
of G, in the sense of Def. 2.25, then the representation is said to be reducible.
If no such proper subspace exists. the representation will be called irreducible.
Definition 2.28. If there is another level of invariant subspaces, so that any or all of these block
matrices can themselves be written in diagonal-block form, and so on, until we are left with only
irreducible representations, then Dg is fully reducible, in the sense that:
Dg = a1 D(1) (2) (N )
g ⊕ a2 Dg ⊕ · · · ⊕ aN Dg (2.4)
(i)
where ai is the number of times (its multiplicity) the irreducible representation Dg occurs in the
direct sum, and N is the number of different irreducibl;e representations in the direct sum.
It can be shown that every representation of a finite group is either irreducible or fully reducible.
When the n-dimensional function space H has proper invariant subspaces, it means that there are at least
two subspaces in H, each of which has its own set of linearly independent functions that transform among them-
selves. Indeed, let HA be an invariant subspace of dimension d, and let {e1 , . . . , ed , . . .} be a basis of H with
A A
{e1 , . . . , ed } a basis of H . We write vectors of functions in H in block form B , where A ∈ HA has dimen-
sion d, and B belongs to the complement subspace HB , of dimension n - d. When HB is invariant, as it always is
in cases of interest to physics (see section 2.4.5
just then a block-diagonal representation matrix Dg , with
below),
A A′
block submatrices DA B
g and Dg , maps vectors B to B′ where A′ ∈ HA , B ′ ∈ HB . Also, since:
! ! ! !
DA
g 0 DA
g′ 0 DA A
g Dg′ 0 DA
g g′ 0
= =
0 DB
g 0 DB
g′ 0 DB B
g Dg′ 0 DB
g g′
DA B A B
g and Dg do preserve the group product, as they should. Dg has dimension d, and Dg dimension n - d.
Then, if all the matrices Dg in a representation can be brought into diagonal-block form by the same simi-
larity transformation, the representation is reducible to lower-dimensional representations composed of the block
matrices.
Going back to SO(2), the D(2) representation we have obtained is clearly fully reducible to the irreducible
representations D(3) = eiα and D(4) = e−iα , so we can write it as D(2) = D(3) ⊕ D(4) .
Example 2.8. The 5-dimensional representation, D(4) , we have constructed for S3 in Example 2.7 is
(by construction) reducible since it is in block-diagonal form, so D(4) = D(1) ⊕ D(2) . What about that
1 0 0 1
last two-dimensional representation
0 1 and 1 0 ? The first is already in block-diagonal form and
1 0
the second can be diagonalised to 0 −1 , Therefore, we obtain two 1-dim irreducible representations,
41
one identical to the identity representation D(5) = 1, and the other the “parity” representation D(3) .
Then we can write: D(4) = D(1) ⊕ D(3) ⊕ D(5) . What about the (left) defining representation of S3 ,
D(1) : is it reducible?
The defining representation of SN has dimension N . This always reducible representation reduces
to a 1-d representation and a (N - 1)-dimensional irreducible representation. To see how this comes
about, let (x1 , . . . , xN ) be a set of coordinates in the carrier space of a defining representation. It is
easy to construct a fully symmetric combination of all those coordinates:
x1 + . . . + xN
X =
N
This function spans the 1-dim subspace of RN invariant under any permutation of the coordinates;
the subspace thus qualifies as the carrier space of the 1-dim irreducible representation of SN that in
section 2.4.8 will be labelled by (N ). Since the defining representation is unitary, the complementary
subspace is itself invariant, and is the carrier space of another irreducible representation. Indeed, let
this (N -1)-dim subspace be spanned by N - 1 functions of the mixed-symmetry form:
x1 + . . . + xj−1 − (j − 1) xj (x1 − xj ) + . . . + (xj−1 − xj )

Yj−1 = p = p 2≤j≤N
j(j − 1) j(j − 1)
These N -1 Jacobi coordinates can be shown to be linearly independent, so that there is no proper
invariant subspace, and the representation is irreducible. The functions are symmetrised with respect
to j - 1 coordinates and then antisymmetrised with respect to the j th one. This allows us to identify
the representation with another irreducible representation of SN that we will label (N -1 1), and the
defining representation can be written as (N ) ⊕ (N -1 1).
The defining representation, D(1) , of S3 is reducible to two irreducible representations, D(5) = 1 and
a set of six 2-dim orthogonal matrices, three with determinant +1 (rotations in a plane by angles 0,
±π/3) and three with determinant −1, thus showing that S3 ⊂ O(2)!. As expected, D(4) is fully
reducible. Can you see why these irreducible representations could not all be one-dimensional?
So this reduction algorithm certainly works, but it would be nice not to have to rely on looking for invariant
subspaces and similarity transformations, which can get quite involved.
2.4.8 Exploring representations with Young diagrams

We have already discussed how Young diagrams could be used to find and label classes of SN . But, much more
often, it is representations that they help to label. We will be looking at SN whose classes we have associated with
partitions of N and, noting that the number of irreducible representations is also the number of classes, as will be
shown later, we will construct their Young diagrams with the same partitions λi of N , where the λi sum up to N
and λ1 ≥ . . . ≥ λN . So Young diagrams for irreducible respresentations of SN look exactly like those for classes.
What will be different is the labelling of the Young diagrams: instead of taking the partitions as the number of
boxes in columns from left to right, we take them as their number in rows from top to bottom. For S3 , this gives:
(3) (2 1) 13
The sequence of representation labels is the reverse of that for classes! But if they are not cycles, what are they?
To discover the meaning of these Young diagrams we consider how the corresponding permutations act on
functions in the carrier space of the N !-dimensional regular representation of SN . We start by giving ourselves a
set of functions {ψi } (1 ≤ i ≤ N ), each of one variable, where the choice of the same symbol as for particle wave-
functions in quantum mechanics is intentional (some authors use the Dirac notation for them). Then with products
42
of these we construct functions of N variables xj . For instance, the product ψ(1...N) := ψ1 (x1 ) · · · ψ1 (xN ) spans
a one-dimensional subspace which contains functions which are obviously completely symmetric and invariant
under any of the N ! possible permutations of the variables. Thus, our subspace qualifies as an invariant subspace
for the regular representation, and it makes sense to associate it with the 1-dim irreducible identity representation
which has the same matrix, 1, for all elements of SN . We shall follow the usual convention by associating it with
the single Young diagram with one row of N boxes. Its label will therefore always be (N ).
With the same set {ψi }, we can also construct the completely antisymmetric function:
ψ1 (x1 ) ··· ψ1 (xN )
.. ..
ψ[1...N] = . ··· .
ψN (x1 ) · · · ψN (xN )
This function changes sign under any transposition in its set of variables, and the 1-d subspace it spans is also
invariant, because the function resulting from multiplying ψ[1...N] by ±1 is obviously in the same subspace. We
associate this subspace with the 1-dim irreducible representation which sends each element of SN to its parity, +1
or −1. Again by convention, this in turn corresponds to the single one-column Young diagram with N rows.
Other irreducible representations, and thus Young diagrams, have a mixed symmetry which can be used to find
their dimension. This is even stronger than eq. (2.14) which is only a constraint on the possible dimensions. Here
is one way to do this.
• Take the Young diagram for each irrep, and fill each of its N boxes with numbers from 1 to N in all possible
permutations to generate N ! Young tableaux. Then assign a function with N subscripts, living in the carrier
space of the regular representation of SN , to each tableau. The order of the subscripts follows the order of
numbers in the first row, then the second row, until the last row. These functions represent products of
functions, each of one coordinate, but we no longer treat them explicitly as such. They form a basis for the
carrier space of the regular representation.
• Symmetrise each function with respect to the numbers in each row of the tableau, and antisymmettrise the
result with respect to the numbers in each column. This yields, for each diagram, a new, mixed-symmetry
function, ψ (i) (1 ≤ i ≤ N ), that is a linear combination of the previous N ! basis functions for the carrier
space of the regular representation.
Example 2.9. For the (2 1) irreducible representation of S3 , the Young tableaux and corresponding
mixed-symmetry functions would be:
1 2 1 3
3 Ψ(1) = ψ123 + ψ213 − ψ321 − ψ231 2 Ψ(2) = ψ132 + ψ312 − ψ231 − ψ321
2 1 2 3
3 Ψ(3) = ψ213 + ψ123 − ψ312 − ψ132 1 Ψ(4) = ψ231 + ψ321 − ψ132 − ψ312
3 1 3 2
2 Ψ(5) = ψ312 + ψ132 − ψ213 − ψ123 1 Ψ(6) = ψ321 + ψ231 − ψ123 − ψ213
The question now is, are these mixed functions independent? Since we expect the regular repre-
sentation to be reducible (fully reducible, in fact), there should exist a lower-dimensional invariant
subspace, the carrier space of our irreducible representation of interest, and we should be able to show
that there are only nα < 6 (for S3 ) independent combinations, where nα will be the number of basis
functions for the invariant subspace, and therefore the dimension of the irreducible representation of
S3 carried by that space.
We note immediately that linear combinations that differ by a transposition of numbers in a column of
their tableaux cannot be independent: they are the negative of one another. So we have at most three
linearly independent combinations. But we also see that Ψ(1) − Ψ(2) − Ψ(3) = 0, leaving only two
independent combinations, which we take to be Ψ(1) and Ψ(2) , and which are the basis functions for
the carrier space of a 2-dim irreducible representation.
43
This rather tedious procedure can be made much faster by filling the tableaux in all the possible ways subject
to the following rules: the number 1 fills the uppermost, leftmost box; and the numbers must increase down any
column and to the right along any row. The number of ways this can be done is the dimension of the representation.
For instance, the (2 1) Young diagram of S3 generates the two tableaux with so-called standard numbering:
1 2 1 3
3 2
Ψ(1) Ψ(2)
each corresponding to one basis function in the 2-dimensional invariant subspace carrying the (2 1) irrep of S3 .
There is, however, a much more convenient method for calculating the dimension of the representation associ-
ated with a Young diagram if one does not wish to construct bases for the subspaces:
Definition 2.29. For any box in the Young diagram associated with an irreducible representation,
draw a straight line down to the last box in its column and to the right end of the box’s row. The result
is called a hook and the number of boxes traversed by the hook is the hook length of this box.
Then the dimension of an irreducible representation is the order of SN , N !, divided by the product of the N hook
lengths for the associated diagram.
Definition 2.30. Irreducible representations for which the Young diagrams are the transpose of each
other, ie. for which the length of each row in one is equal to the length of the corresponding column
in the other, are said to be conjugate. Their dimensions are the same.
The Young diagram of a self-conjugate irreducible representation is identical to its transpose.
2.5 Schur’s Lemmas and Symmetry in the Language of Group Theory (BF10.6)
We now present two fundamental results of group theory which provide useful criteria for the irreducibility of
representations as well as insight into symmetries, and which lead to relations that help to classify representations.
2.5.1 What is a symmetry in the language of group theory?
Consider a linear operator L such that, ∀ f ∈ H, [Lx f ](x) = h(x) ∈ H. Under a group G, [Tg Lx Tg−1 ][Tg f ](x) =
[Tg h](x), and [Lx′ f ](x′ ) = h(x′ ), so that L transforms under G as: Lx′ = Tg Lx Tg−1 .
Definition 2.31. When Tg Lx Tg−1 = Lx , ∀ g ∈ G, L is said to be invariant under the action of
group G. Since this condition can also be written as Tg L = L Tg , ∀ g ∈ G, then an operator that
is invariant under a group of transformations must commute with all those transformations. If also
[Tg f ](x) = f (x), f is invariant under G itself as well (eg., f (r) in polar coordinates under rotations).
If L has eigenvalues and eigenfunctions and is invariant under G, then there should exist a set {f i } such that:
L (Tg f i ) = Tg L f i = λ (Tg f i ) (2.5)
Thus, if f i is an eigenfunction of L, so is Tg f i , with the same eigenvalue. Therefore, the distinct Tg f i are all
degenerate with respect to λ. If λ is degenerate, there are N degenerate f i for that λ, which form a basis for a
N -dim subspace of functions, characterised by λ. The Tg f i will then be some linear combination of {f j }. so
that the transformed eigenfunctions Tg f i also form a basis for the same subspace of functions as that spanned by
the eigenfunctions of L: the subspace is invariant under the action of the group, in the sense of Def. 2.25! With
summation up to N over repeated indices implied:
Tg f i = f j (Dg )j i (2.6)
Whenever we find (or observe) a set of degenerate eigenfunctions for some operator, the operator is invariant
under the action of a group, and these functions will be connected with an irreducible representation of the group.
44
2.5.2 Schur’s Lemmas

Consider a matrix M ∈ GL(n, C) with eigenvectors A belonging to eigenvalue λ. Let M and Dg commute
∀ g ∈ G. The same argument as in the previous section shows that Dg A are also eigenvectors of M spanning the
same subspace H as the eigenvectors A.
If Dg is irreducible, H has no proper subspace invariant under G and the Dg A form a basis of H. In other
words, ∀ ψ ∈ H: X X
Mψ = ag M Dg A = λ ag Dg A = λ ψ
g g
so that all transformed vectors in H are eigenvectors of M, with the same eigenvalue λ. This can happen only if
M = λ I, and there comes Schur’s First Lemma:
The only complex matrix M that commutes with all the matrices of a given irreducible representation Dg is a
multiple of the identity matrix.
As a corollary, if a matrix can be found which is not a multiple of I and yet commutes with all matrices in a
representation, that representation must be reducible. This provides one handy test for reducibility.
From this Lemma follows an immediate consequence for Abelian groups, where any matrix Dg in a given
representation commutes with the matrices for all other group elements in this representation. Assuming a (n > 1)-
dim irreducible representation, the Lemma requires that Dg = λ I, ∀ g ∈ G. But the n × n identity matrix, which
is diagonal, cannot be irreducible if it represents all group elements, contradicting our assumption. We conclude
that all irreducible representations of an Abelian group are one-dimensional.
(α) (β)
Schur’s Second Lemma: If a non-zero matrix M exists such that Dg M = MDg ∀ g ∈ G, then D(α) and D(β)
must be equivalent irreducible representations. If D(α) and D(β) are inequivalent, M = 0.
This lemma can be proved (pp. BF615–617) by assuming unitary representations. This makes for no loss of
generality for finite or compact Lie groups, since these eg. O(n) have finite-dimensional representations.
2.5.3 An orthogonality relation for the matrix elements of irreducible representations (BF10.6)
Another important consequence of Schur’s Lemmas is the fact that the matrix elements of all the inequivalent irre-
ducible representations of a finite group, or those for infinite groups that have finite-dimensional representations,
form a set of orthogonal functions of the elements of the group. More specifically, if {Dg } is the set of all matrices
Dg in an irreducible representation D, then, for two such representations labelled by α and β::
NG
X i (β) l NG i
Dg(α) k
Dg−1 = δ j δk l δαβ (2.7)
g
j nα
where NG is the order of the group and nα is the dimension of D(α) . The sum is not matrix multiplication! Each
(α) (β)
term is the product of some ik entry of Dg and lj entry of Dg−1 , with ik and lj the same for each term.
In the usual case of unitary representations, this relation simplifies to:
NG
X i l NG i l
Dg(α) k
Dg(β) ∗ j
= δ j δ k δαβ (2.8)
g
nα
These relations set powerful constraints on the matrix elements of representations.

Eq. (2.7) is so important that it deserves a proof. Fortunately, this proof is not too hard. Construct a matrix:
NG
X (β) −1
M = D(α)
g X Dg ] (2.9)
g
45
where D(α) and D(β) are m-dim and n-dim inequivalent irreducible matrix representations of G, and X is any
arbitrary operator represented by a m × n matrix X. Then, for some g′ ∈ G,
NG
X
(α) (β) (α) (β)
Dg′ M Dg′ ]−1 = Dg′ g X Dg′ g ]−1
g
The sum on the right-hand side is just a different rearrangement of the sum that defines M, so that:
(α) (β)
M = Dg′ M Dg′ ]−1
(α) (β)
Thus, Dg M = M Dg ∀ g ∈ G, and M meets the condition for Schur’s Lemmas. In particular, Schur’s lemma
requires that M = 0 if α 6= β since D(α) and D(β) are inequivalent. Now let us choose X to be a matrix whose
only non-zero element, 1, is its (kl)th entry. We can write this formally as: (Xkl )m n = δm k δn l . Inserting gives:
NG
X NG
X
i (β) n i (β) l
(Mkl )i j = D(α)
g m
(Xkl )m n Dg−1 = D(α)
g k
Dg−1
j j
g g
When α 6= β Mlk = 0, as we have seen. When α = β, Schur’s First Lemma requires that Mlk = λk l I, leading to:
NG
X NG
X
i (α) n i (α) l
(Mkl )i j = D(α)
g m
(Xkl )m n Dg−1 = D(α)
g k
Dg−1 = λl k δi j
j j
g g
Setting i = j and interchanging the D factors to get a matrix product, there comes:
NG
X XNG
(α) l
(α) j (α) l
Dg−1 Dg k
= Dg−1 g = NG δl k = λl k nα
j k
g g
Thus, we find: λl k = NG δl k /nα . Combining

q the results for α = β and α 6= β gives eq, (2.7) or (2.8).
nα (α)
This means that the matrix elements N G
(Dg )i j of a unitary irreducible representation must be orthonormal
functions of the group elements g; they are the components of a NG -dim vector orthogonal to the similarly built
vectors for any other irreducible representation. Therefore, these vectors are linearly independent. Also, they form
a complete set.
EXERCISE: Show that for a finite group the sum over all elements g of the matrix elements (Dg )i j (i and j
fixed) of any irreducible representation other than the identity 1-dim representation (eg., (N ) for SN ) is zero. This
property can provide a useful check.
2.5.4 Characters of a representation (BF10.7); first orthogonality relation for characters

It turns out that a surprising large amount of information about representation matrices is encoded in their trace.
Very often this trace can be found without knowing the full matrix.
Definition 2.32. The character of a representation Dg of a group G is defined as a map from G to C:
χg = Tr Dg
Characters of reducible representations are compound; those of irreducible representations are called
simple. Language alert: Mathematicians speak of the “character” of a representation as the set of
traces of the matrices in the representation.
46
We establish an interesting fact: In a given representation, all matrices associated with elements of the same
class have the same trace. Recall that the class to which g belongs is made of {g′ g g′−1 } ∀ g′ ∈ G. Then the
trace of Dg′ g g′−1 is equal† to the trace of Dg , or χ. Since matrices for equivalent representations have the same
character, any statement about characters is basis-independent!
Now set k = i and l = j in eq. (2.8):
XNG
i j NG i j NG i
Dg(α) i Dg(β) ∗ j = δ j δi δαβ = δ i δαβ
g
nα nα
where repeated indices are summed over. Since δi i = nα , this can be rewritten as:
NG
X
∗ (β)
χ(α)
g χg = NG δαβ (2.10)
g
This provides our first orthogonality relation between the characters of irreducible representations. It can be viewed
as an inner product on the space of functions of the NG -dim “ character vectors”.
Some of the terms in this sum will be identical since they correspond to group elements in the same class. So
we can collect all terms belonging to the same class, which we label with k, and instead sum over the classes:
Nc
X (α)
nk χk χ∗k (β) = NG δαβ (2.11)
k=1
with nk the number of elements in class k and Nc the p number of classespin the group. This looks for all the world
like an orthogonality relation between two vectors, nk /NG χ(α) and nk /NG χ(β) , each of dimension Nc .
For a given irreducible representation, eq. (2.11) becomes:
Nc
X (α) 2
n k χk = NG (2.12)
k=1
This is a necessary and sufficient condition for the representation to be irreducible!
Example 2.10. Take for instance the 3 × 3 representation of S3 found in section 2.4.3. The identity,
with trace 3, is in its own class, the three transpositions are in another class with trace 1, and the two
cyclic permutations) have trace 0. Eq. (2.12) gives: n1 χ21 +n2 χ22 +n3 χ23 = 1 (3)2 + 3 (1)2 + 2 (0)2 =
12. Since this is not equal to 6, the number of elements in S3 , the representation must be reducible.
According to eq. (2.11), the “ character vectors” of the Nr different irreducible representations are orthogonal.
There are Nr such orthogonal vectors, and their number may not exceed the dimensionality of the space, Nc , so
that Nr ≤ Nc . We will need this result a little later.
2.5.5 Multiplicity of irreducible representations and a sum rule for their dimension
Now consider the decomposition of a fully reducible representation into a direct sum of irreducible ones, given in
(α)
eq. (2.4). Taking its trace yields an equation for the compound character χg : χg = aα χg , where the sum runs
over the Nr irreducible representations of the group. The compound character is seen to be a linear combination
of simple characters with positive coefficients equal to the multiplicity of each irreducible representation.
Multiplying this relation by χ⋆ (β) (g) and summing over group elements, we find from eq. (2.10):
XNG NG
X
χg χ⋆g (β) = aα χ(α)
g χg
∗ (β)
= aα NG δαβ = aβ NG
g g
Thus, the multiplicity of each irreducible representation in the decomposition of a reducible representation is:
†
This is because Tr AB = Ai j Bj i = Bj i Ai j = Tr BA.
47
NG Nc
1 X 1 X ∗ (α)
aα = χg χg∗ (α) = n k χk χk (2.13)
NG g NG
k
Also, we can exploit the regular representation to obtain other general results for irreducible representations.
As we have seen in section 2.4.4, the matrix elements of the regular representation can only be 1 or 0. Since
only the identity will map a group element to itself, the only matrix with 1 anywhere on the diagonal is the identity
matrix. Therefore, the characters all vanish except by emailing [email protected] for χ(e) = NG .
Now, with nα the dimension of the αth irreducible representation and g = e, eq. (2.13) gives:
1
aα = χe χe∗ (α) = χe∗ (α) = nα
NG
Only χ(e) can contribute to the sum since χg = 0 in the regular representation when g 6= e.
Therefore, the multiplicity of an irreducible representation in the decomposition of the regular representation is
its dimension, and it is never zero. All the irreducible representations of a group must appear in the decomposition
of its regular representation. by emailing [email protected]
Next, taking the P
trace of the Kronecker decomposition (2.4) for the identity element in the regular representa-
tion yields: NG = α aα nα . Combining those results, there comes an important sum rule:
X
NG = n2α (2.14)
α
√ √
This powerful constraint tells us that nα ≤ NG so that any representation of dimension larger than NG
must be reducible. When NG = 2 or 3, all irreducible representations are one-dimensional. When NG = 4, we can
have only four inequivalent 1-d irreducible representations; nα = 2 is ruled out because there would be no identity
1-d representation. When NG = 5, eq. (2.14) does allow the identity representation plus one 2-d irreducible
representation; but we know that this group, Z5 , is Abelian, and so admits only five inequivalent 1-d irreducible
representations. For NG = 6, six 1-d, or two 1-d plus one 2-d irreducible representations, are allowed.
2.5.6 Another orthogonality relation

Here is another orthogonality relation, whose more complicated proof has been consigned to Appendix E at the
end of this chapter for those who may be interested:
XNr
nk (α) ∗ (α)
χk χk ′ = δk′ k (2.15)
α=1
N G
Thus, the characters in a given class k can be considered as components of Nc vectors forming a basis of a
space whose dimension is Nr , the number of irreducible representations, and which, according to eq. (2.15), are
orthogonal. But in a Nr -dimensional space there cannot be more than Nr orthogonal vectors, so Nc ≤ Nr .
In section 2.5.4, however, we had argued that Nr ≤ Nc . These results together lead to the important statement:
The number of inequivalent irreducible representations of a group is equal to the number of classes: Nr = Nc .
Now it can be shown (see Appendix F) that the direct product of an irreducible representation with a 1-d
representation is itself an irreducible representation, which may be the same (when the 1-d representation is the
identity). This goes for their characters also. When the completely antisymmetric (1N ) 1-d representation exists,
as is the case for SN , the characters of an irreducible representation can always be written, class by class, as the
product of the characters of its conjugate representation and the characters in the (1N ) representation. Therefore,
characters for a given class in a pair of conjugate representations are either identical or differ only by their sign.
Characters of a self-conjugate representation in a class that has negative parity must vanish.
48
2.5.7 Character tables

A character table contains the characters of the classes in a group’s irreducible representations. Each row contains
the characters of all classes in a representation, and each column the characters of a class in all representations.
The first row corresponds to the identity 1-dim irreducible representation, (N ); all its entries must be 1. The
first column corresponds to the identity class; each entry in that column must be the dimension of the representation
(ie. the trace of the identity matrix in each representation) for the row.
If we are dealing with SN , there is another 1-dim irreducible antisymmetric representation (called the sign
representation by mathematicians), (1N ), conjugate to (N ), whose 1 × 1 matrices, and therefore characters, are
the parities ±1 of its classes. We choose to place this representation at the bottom of the table.
What about the other entries? Well, we can assign some algebraic symbol to the unknown characters and then
spot conjugate representations. If there are any, the character in each column of one representation in a conjugate
pair must be the character of its conjugate multiplied by the character (±1) in the antisymmetric 1-d row. If there
are self-conjugate representations, any character sitting in the same column as a −1 in the last row must be zero.
We have shown in Example 2.8 that the defining representation of SN reduces to (N ) and (N -1 1). The
characters of this (N -1 1) representation can be calculated as follows. First, we note that for a class labelled
(. . . 2β 1α ), the characters of the defining representation are equal to the number of objects that the permutations
in the class leave invariant, ie. α. Since these compound characters are the sum of the characters of (N -1 1) and
(N ), we find that the characters of each class labelled by α in the (N -1 1) irreducible representation are just α − 1.
When SN has an N -dim irreducible representation, the permutations in the (N ) class shuffle all N objects,.
The N -dimensional matrices representing them must have diagonal entries 0, resulting in a character that is 0.
Next, we let eq. (2.11) and (2.15) provide constraints on the remaining unknowns:
• The first says that complete rows in the table (each for a different representations) are orthogonal, with the
understanding that each term in the sum is weighted by the number of elements in the class (column).
• The second says that complete columns (each belonging to different classes) are othogonal.
Now, if β refers to the identity representation, then, for any irrep α other than the identity, eq. (2.11) becomes:
Nc
X (α)
n k χk = 0 (2.16)
k=1
When invoking the orthogonality constraints to find characters, it is best to apply the linear ones first. Unfortu-
nately, many of these relations will be automatically satisfied and will not yield new information, because of the
strong constraints on the characters imposed by conjugation and self-conjugation of the irreducible representations.
When all possible information has been extracted from eq. (2.16) and (2.15), and there still remain unknowns, one
can try to spot reasonably simple quadratic relations from eq. (2.11) as well as using the normalisation of rows and
columns.
Two last but important remarks: the characters of any 1-dim representations of any group (eg. those of an
Abelian group) must preserve the group product. Also, although the characters of SN are real, characters of other
groups (eg. Zn ) can be complex.
There exist even more sophisticated methods for determining the characters of a group (eg. by generating them
from the characters of a subgroup, or of a factor group), but lack of time and space prevents us from discussing
them here. In fact, character tables for well-known groups can be found in specialised books and on the web.
Let us use these rules to find the characters of S3 as a 3 × 3 table, with classes corresponding to columns and
irreducible representations to rows. The first and last row can be immediately written down from our knowledge
of the parity of each class (−1 for the transpositions and +1 for the cyclic permutations). Note also that the (2 1)
representation is self-conjugate, so we can put 0 for the character in the (2 1) class, because the parity of that class
(last character in the column) is −1. The (2 1) representation is the (N -1 1) representation discussed above, and
its remaining character is determined by its belonging to a class with α = 0; thus, the character must be −1. The
linear constraint (2.16), as well as the other orthogonality rules, are automatically satisfied. Collecting yields:
49
(13 ) (2 1) (3)
nk 1 3 2
(3) 1 1 1
(2 1) 2 0 −1
(13 ) 1 −1 1
EXERCISE: work out the character table and irreducible representations of Z4 , the cyclic group of order 4. You
may make the task easier by remembering that products of characters belonging to a 1-d irreducible representation,
which are the actual representation matrices, must mimic the group product of the corresponding elements.
Example 2.11. Lifting of a degeneracy by a weak interaction

Consider a physical system in a rotationally-invariant potential that depends only on distance to the
origin. This often occurs in quantum mechanics, and the result is that the eigenstates labelled by the
integers that characterise eigenvalues of L2 and Lz , l and m, with −l ≤ m ≤ l, exhibit a 2l + 1-
fold degeneracy, in the sense that they all have the same energy. This is also manifested by the way
spherical harmonics, which are eigenfunctions of L2 and Lz for a given value of l, as well as of the
Hamiltonian, transform under a rotation by some angle α. Using eq. (2.6), we have:
l
X
m′
Rα Ylm (θ, φ) = Ylm′ (θ, φ) D (l) m
(α)
m′ =−l
where the D(l) matrix is an irreducible representation of the rotation group SO(3) which acts on the
invariant space spanned by the 2l+1 Ylm for that l. SO(3) will be discussed in chapter 3.
We can simplify things by noting that rotations by an angle α about any axis are all equivalent to (in
the same class as) a rotation by that angle around the z-axis. It will be sufficient to calculate the trace
of the matrix representing rotations around that axis. To find this matrix, notice that [Rα Ylm ](θ, φ) =
eimα Ylm (θ, φ) = Ylm (θ, φ + α) because the dependence of the spherical harmonics on φ is eimφ .
Therefore, D(l) (α) = diag(e−ilα , e−i(l+1)α , . . . , eilα ), and its character is not hard to compute:
l 2l
!
X X 1 − ei(2l+1)α sin (l + 1/2)α
(l) iα m −ilα iα n −ilα
χ (α 6= 0) = (e ) = e (e ) = e =
n=0
1 − eiα sin(α/2)
m=−l
(2.17)
where we have recast the sum as a geometric series by redefining the index as m = n − l.
Now let us turn on a weak interaction whose corresponding potential is no longer rotationally-invariant,
but still retains invariance under rotations by a restricted, finite set of angles, which we collectively
denote by β. This would happen, for instance, if we embed our spherically-symmetric atom in a crys-
tal lattice. Suppose this restricted set of rotations actually is a group, or more precisely, a subgroup of
SO(3). Then the matrix D(l) (β) should be a representation of that subgroup, but that representation
may no longer be irreducible. This will certainly happen for any D(l) whose dimension is too large to
satisfy the sum rule (2.14) that applies to the finite subgroup.
m′
The set of Ylm transform as: Rβ Ylm = Ylm′ D (l) m (β), with summation over repeated indices
implied. If the induced representation D of the restricted-symmetry subgroup is reducible, there
exists a matrix S independent of β which transforms all its matrices into block-diagonal matrices
D′ = S D S−1 , something which was impossible when there was no restriction on the angles.
But we do not have to know S to extract useful information. Indeed, because D and D′ have the
same trace, we can calculate the characters of D(l) (β) for all elements of the restricted-angle subset
in SO(3). Then we find the character table of the restricted-symmetry group, which is finite. If
there is a row in the table that exactly matches the SO(3) characters of D(l) (β), then D(l) (β) is
not only an irreducible representation of SO(3), it is also an irrep of its subgroup defined by the
50
angles allowed by the restricted symmetry. The corresponding invariant subspaces are identical, and
the original 2l + 1-fold degeneracy for that value of l is still present after the perturbation has been
turned on. As l increases, however, the dimension 2l + 1 of D(l) (0), which always appears as the first
character corresponding to the identity class of SO(3), will eventually exceed the fixed dimension of
any irreducible representation of the subgroup. Then all the corresponding D(l) (β) will be reducible to
a direct sum of the irreducible representations of the subgroup, given by eq. (2.4), with the multiplicity
of each irrep calculable from eq. (2.13).
For instance, suppose that the perturbation has cubic symmetry. A cube is invariant under1 :
• 6 rotations by ±π/2 around the three axes through its centre that intersect faces through their
centre;
• 3 rotations by π around these same axes;
• 8 rotations by ±2π/3 around the four axes through diagonally opposed corners (vertices).
• 6 rotations by π around the six axes intersecting the centre of two diagonally opposed edges;
With the identity rotation, these add up to 24 elements forming a subgroup of SO(3) isomorphic to
S4 . The correspondence between rotations and permutations is obtained by considering each rotation
as a shuffling of the four pairs of diagonally opposed vertices (or the four principal diagonals through
the centre), each pair labelled 1 to 4. The five classes of S4 are (14 ) (e), (2 12 ) (rotations by π), (22 )
(rotations by π) , (3 1) (rotations by ±2π/3), and (4) (rotations by ±π/2). The S4 character table is:
(14 ) (2 12 ) (22 ) (3 1) (4)
nk 1 6 3 8 6
(4) 1 1 1 1 1
(14 ) 1 −1 1 1 −1
(22 ) 2 0 2 −1 0
(3 1) 3 1 −1 0 −1
(2 12 ) 3 −1 −1 0 1
Here, the irreps of S4 (or of the group of rotational symmetries of the cube) are ordered by increasing
dimension instead of their mixed-symmetry structure. With eq. (2.17), we calculate the characters
of the representations of S4 induced by D(l=1) (β) and D(l=2) (β), with angles β running through the
values corresponding to the five classes of S4 :
(14 ) (2 12 ) (22 ) (3 1) (4) (14 ) (2 12 ) (22 ) (3 1) (4)
(l=1) (l=2)
D 3 1 −1 0 −1 D 5 −1 1 −1 1
The l = 1 irrep of SO(3) restricted to the angles allowed by the cubic-symmetry subgroup has the
same dimension and the same characters as the representation (3 1) of S4 in the above character
table for S4 . The invariant spaces are the same and there is no lifting of the unperturbed 3-fold
degeneracy. The l = 2 irrrep of SO(3), however, has no identical row in the S4 character table,
and must correspond to a reducible representation of S4 . With eq. (2.13), we calculate the following
multiplicity for each irrep of S4 that can appear in the decomposition of D(l=2) (β): a(4) = a(14 ) =
a(2 12 ) = 0, and a(22 ) = a(3 1) = 1. Then we have the S4 decomposition:
D(l=2) (β) = D(22 ) (β) ⊕ D(3 1) (β)
The unperturbed 5-fold degeneracy of the l = 2 states is partially lifted to become two “levels”, one
3-fold and one 2-fold degenerate.
Another example illustrating how symmetry-breaking can remove degeneracy, at least in part, can be
found in Appendix G.
1
See, eg: http://demonstrations.wolfram.com/RotatingCubesAboutAxesOfSymmetry3DRotationIsNonAbelian/.
51
Appendices
D The Right and Left Actions of a Group on a Vector, with Sn as Example
First, we recall some important properties of linear transformations. For simplicity we shall take these transforma-
tions to act on vectors x = xi ei ∈ Rn ,, with thestandard basis {ei }. It is customary to distinguish between active
transformations: x −→ x′ = x′i ei = ei Ai j xj , where the transformed coordinates x′i are those of a new vector;
and passive transformations that leave x invariant but transform the basis to {e′i }: e′i = ej P j i .
Paaive transformations, unlike active ones, require the transformed vector x′ to remain the same as the initial
one, because all we have done is change the basis. To preserve x, we must also transform its components with the
inverse transformation:
j
x′ = e′j x′j )P = ei P i j P −1 k xk = ei δi k xk = ei xi = x
Now, although an active transformation is quite different from a passive one, the result should be mathematically
equivalent to the result of transforming the basis with the inverse transformation. In other words, the components,
(x′A )j , of the new vector produced by the active trnsformation should be the same as the those of the initial vector
j
in the transformed basis, that is, (x′A )j = (x′P )j = P −1 k xk . Comparing with (x′A )j = Aj k xk , we see that
that the passive transformation is indeed inverse to the active one.
D.0.1 Right action

This is often the easiest to work with. According to definition 2.21 directly applied as in eq. (2.3), TgR (x) := x g.
Now x = xi ei , and when it comes to representation matrices, we have the choice between acting on components
xi and acting on basis vectors ei ;

 ⇐⇒ x i g = D R i xj
i g j
x g = ei xi g = ei DgR j xj j
 ⇐⇒ ei g = ej Dg−1 R
i
Notice that the action on the basis vectors involves the matrix for g−1 , as expected for the passive transformation
associated with the active one on the components.
(a) Components
i i
For g = πk ∈ Sn , the expression for xi πk is equivalent to xπk (i) = DgR j
xj , so that DgR j
= δπk (i) j ,
and the ith row of the n-dim DRg matrix is the standard-basis vector eπk (i) . For πk = π5 = (1 3 2) ∈ S3 , for in-
R 1 1 1 1 1 1
stance, we find: Dπ5 1 , Dπ5 2 , DπR5 3 = (0, 0, 1), and for π6 = (1 2 3): DπR6 1 , DπR6 2 , DπR6 3
R
= (0, 1, 0), etc. There comes:

    2      3      1 
0 1 0 x1 x 0 0 1 x1 x 1 0 0 x1 x
i
x π2 =  1 0 0 x 2
= x 1 i
x π3 =  0 1 0  x 2
=  x 2 i
x π4 =  0 0 1 x 2 
= x3 
0 0 1 x3 x3 1 0 0 x3 x1 0 1 0 x3 x2
    3      2 
0 0 1 x1 x 0 1 0 x1 x
i
x π5 =  1 0 0 x 2 
= x1  i
x π6 = 0 0 1 x 2 
= x3 
0 1 0 x 3 x2 1 0 0 x 3 x1
−1
where DR
π
= DR
π6 , and vice-versa, as expected from the Cayley table of S3 .
5
(b) Basis vectors

Now the πk shuffle the basis vectors themselves (not components). The expression for ei πk is not matrix
j
multiplication! Indeed, ei πk = ej DπR−1 gives the entries of the ith column of the n-dim DR
g−1 matrix in
k i
52
1 n
the standard basis: ei πk = DπR−1 , . . . , DπR−1 . More succinctly, because eπk (i) = ei πk , we get:
k i k i
j
DπR−1 = δj π−1 (i) , or the .ith column is the standard-basis ector eπk (i) .
k i k
R 1 2 3
Using basis vectors to find DR
π5 , for instance, we have: e1 π5 = e3 = (0, 0, 1) = Dπ6 1 , DπR6 1 , DπR6 1 ,
1 2 3
and e2 π5 = e1 = (1, 0, 0) = DπR6 2 , DπR6 2 , DπR6 2 , etc. The resulting matrix is the one that was
obtained from the components for the right action of π6 , and its inverse will be the matrix for the right action
of π5 , the same that we found somewhat more directly by acting on components.
D.0.2 Left action

Under the left action of g, we have: TgL (x) = g−1 x. Once again this can act on basis vectors or on components:
 j j
j  ⇐⇒ g−1 ei = ej DgL−1 = ej DgL i
i
g−1 x = xi g−1 ei = x ej DgL−1
i
i i
i  ⇐⇒ g−1 xi = DgL−1 xj = xj DgL j
j
where the last equalities on the right hold when the matrices are orthogonal (eg., rotations). In that case we can
find the left-action matrix for g directly.
(a) Basis vectors

1 2 n
In the case of g = πk ∈ Sn , we have: πk−1 ei = eπ−1 (i) = DπLk i , DπLk i , . . . , DπLk i as the ith row
k
j
of DLπk . In short DπLk i = δπ−1 (i) j . For instance, taking as before πk = (π5 , π6 ) ∈ S3 immediately leads to:
k
   
0 1 0 0 0 1
DLπ5 = 0 0 1 DLπ6 = 1 0 0
1 0 0 0 1 0
These are the DL matrices found in example 2.7. The left and right representations for each single transposition
are identical.
Components
i i
In the case of g = πk ∈ Sn , we have: xπk (i) = xj DπLk j , so that DπLk j = δj πk (i) . Working with our
trusted friends (π5 , π6 ) ∈ S3 , we obtain:
   
0 1 0 0 0 1
DLπ5 = 0 0 1 DLπ6 = 1 0 0
1 0 0 0 1 0
53
E Proof of the Second orthogonality Relation for Characters

p (α) i
Our first orthogonality relation for characters, eq. (2.8). says that the set { nα /nG Dg j
}, with i andj fixed,
of an irreducible representation α, can be viewed as the NG components of a vector orthogonal to any other such
vector corresponding to other matrix elements, whether or not they belong to the same representation as that of the
first vector. There are NG such vectors and they form a complete set with completeness relation expressed as:
XNr Xnα
nα i (α) ∗ i
Dg(α) j Dg′ = δg′ g (E.1)
α i, j
NG j
where Nr is the number of irreducible representations. Again, the left-hand side is not matrix multiplication.
Take the equation for each element g of some class k, and sum over all elements of the class; we can also do
this with g′ over the elements of another class k′ . When k 6= k′ , the right-hand side of the double summation must
vanish because classes are distinct; when k = k′ , the double sum collapses into one which adds up to nk .
Pn k (α) i
g Dg j
in the now quadruple sum on the left-hand side is an element of the matrix M constructed by
Pn k
summing all the matrices Dg in the representation that correspond to elements g of class k: M = g Dg .
If g′ is some arbitrary element of G, we have:
X X
Dg′ M Dg′−1 = Dg′ Dg Dg′−1 = Dg′ g g′−1 = M
g g
where the last equality results from the fact that, since g′ g g′−1 is in class k, the left-hand side of the last
equality is just a rearrangement of the sum defining M. Thus, Dg M = M Dg ∀ g ∈ G, and, from Schur’s First
Lemma, M = λ I, with λ a constant that depends on the class and on the n-dim representation. Then Tr M = nλ.
Because all matrices in a class for a given representation must have the same trace, we have: Tr M = nk χk .
Since that trace is also nλ, we find:
nk
M = χI (E.2)
n
With two of its four sums replaced by matrix elements of M, the completeness relation (E.1) now reads:
Nr X
X nα
nα (α) i (α) ∗ i
Mk j
Mk′ j
= nk δk′ k
α
N G
i, j
(α)
Inserting M = (nk /nα )χ(α) I and carrying out the sums over i and j gives another orthogonality relation:
XNr
nk (α) ∗ (α)
χk χk ′ = δk′ k (E.3)
N
α=1 G
F Direct Product of Representations

Let D i j and D ′α β be the entries of respresentation matrices D, of rank m, and D′ , of rank n, for some goup
element g. We define the direct-product representation constructed out of D and D′ as the (mn)×(mn) matrix
M = D ⊗ D′ with entries M iα jβ := D i j D ′α β , where the superscript iα and subscript jβ, each running from 1 to
(mn) just indicate which matrix elements to multiply.
M can be seen to be a representation because the ordinary matrix product Mg1 Mg2 has entries:
(Mg1 )iα jβ (Mg2 )jβ kγ = (Dg1 )i j (Dg′ 1 )α β (Dg2 )j k (Dg′ 2 )β γ = (Dg1 )i j (Dg2 )j k (Dg′ 1 )α β (Dg2 )′β γ
= (Dg1 g2 )i k (Dg′ 1 g2 )α γ = (Mg1 g2 )iα kγ
Therefore, the matrix product of direct-products for two group elements is the direct product of the matrix product
for the representations of the sane two elements. Also, the character of the direct-product for a group element,
M iα iα = D i i D ′α α , is obviously the product of the characters of the representations in the product.
When D is irreducible, its direct product with a 1-dim (and thus irreducible) representation, results in an
irreducible representation of the same dimension as D, which we call its conjugate representation.
54
G A Second Example of Symmetry-Breaking Lifting a Degeneracy

Take six identical bodies arranged on a circle 60◦ apart, each subject to an identical external linear restoring force
giving rise to small oscillations about their equilibrium position and tangent to the circle. Let X be their displace-
ment vector, with components: x1 , . . . , x6 their displacements away from their respective equilibrium position.
This vector is a solution of Newton’s 2nd Law for the system, Ẍ = −M−1 KX, where: M = diag(m, m, . . . , m),
and K = diag(k, . . . , k), with k the restoring constant associated with the motion. We call M−1 K the dynam-p
ical matrix of the system. Of course, as we all know, all bodies oscillate at the same frequency ω0 = k/m.
The space of solutions is spanned by the six eigenvectors belonging to the same eigenvalue ω0 . This is a six-fold
degeneracy.
Now we couple the bodies by an identical weak interaction to their nearest neighbours ±60◦ away; similarly,
couple them to their second next neighbours, ±120◦ away, by another (even weaker) interaction that is identical
for both these neighbours; finally, a third (weakest) interaction couples each one to its opposite counterpart, 180◦
away. We wish to study the effect of the coupling on the motion of the bodies tangent to the circle.
Because of the symmetry of the interactions and of the system, the dynamical M−1 K matrix must have the
form:  
ω02 −ω12 −ω22 −ω32 −ω22 −ω12
−ω 2 ω02 −ω12 −ω22 −ω32 −ω22 
 1 
 2 
−ω2 −ω12 ω02 −ω12 −ω22 −ω32 
M−1 K = 
−ω 2 −ω 2 −ω 2 ω02 −ω12

−ω22 
 3 2 1 
 2 
−ω2 −ω32 −ω22 −ω12 ω02 −ω12 
−ω12 −ω22 −ω32 −ω22 −ω12 ω02
How can we use the symmetry to find the normal modes of the system? By recognising that the system must be
invariant under 60◦ rotations. This operation is isomorphic to a cyclic permutation: (1 2 3 4 5 6) ∈ Z6 . The regular
representation matrix for this element of Z6 looks like:
 
0 1 0 0 0 0
0 0 1 0 0 0
 
0 0 0 1 0 0
S =  
0 0 0 0 1 0
 
0 0 0 0 0 1
1 0 0 0 0 0
Invariance under S means that M−1 K and S commute. In fact, this last statement can be used to obtain the form
of the M−1 K matrix given above.
The eigenvectors of S now satisfy SA = λA. But since S6 = I, we immediately find that the eigenvalues are
the sixth roots of 1, as expected for the cyclic group. Therefore. λ(m) = eimπ/3 , (0 ≤ m ≤ 5). To each value of m
corresponds an eigenvector A(m) with components Aj(m) = λ(m) j−1
= eim(j−1)π/3 .
These eigenvectors are also the normal modes of the system. Inserting into the eigenvalue equation M−1 K A(m) =
2 A
ω(m) (m) with the coupling parameters ω(5) = ω(1) and ω(4) = ω(2) yields the dispersion relation:
6
X
2 2
ω(m) = ωj−1 eim(j−1)π/3 = ω02 − 2 ω12 cos mπ/3 − 2 ω22 cos 2mπ/3 − (−1)m ω32
j=1
We note that A∗(1) = A(5) , and A∗(2) = A(4) . These modes are complex, which is a problem if they are sup-
posed to correspond to real relative amplitudes. But we also note that ω(1) = ω(5) , and ω(2) = ω(4) ; therefore,
the corresponding eigenvectors span two invariant 2-dim subspaces, which allows us to take appropriate linear
combinations of the eigenvectors to turn them into real modes of the same frequency.
The coupling has lifted the original 6-fold degeneracy of the uncoupled system, but there is still some degen-
eracy left because of the two 2-dim subspaces.
This is as far as we can go without knowing the interaction parameters themselves. But we have succeeded in
nailing down the relative amplitudes of motion of the bodies in each normal mode without that explicit knowledge!
55
3 CHAPTER III — LIE GROUPS

3.1 Definitions
In this chapter we focus on a class of groups with an infinite number of elements. As groups, they of course satisfy
the algebraic properties of a group as set out in definition 2.1. But now we put in an extra requirement: that each
group element gp , or g(P ), be in correspondence with a point P in some manifold, with the index “p” taken as a
set of continuous, real variables. We say that the manifold parametrises the group. More precisely:
Definition 3.1. Let P be any point in a n-dim manifold M n which is obtained from two other points,
P1 and P2 from invertible mappings P = φi (P1 , P2 ). Let g(P1 ) ⋆ g(P2 ) = g(P ) be the group product
of an infinite group G. If the maps φi and their inverse are differentiable, then G is a Lie group.
The important point to remember here is that since they correspond to points in a manifold, elements
of a Lie group can be parametrised in terms of smooth coordinates on this manifold.
A Lie group is real if its manifold is real and complex if its manifold is complex.
The dimension of a Lie group is the dimension of its manifold.
Definition 3.2. A Lie group is said to be path-connected if any pair of points on its manifold is
connected by a continuous path.
A Lie group is compact when the volume of its manifold is finite.
The subset of all elements in a Lie group whose corresponding points in M n are connected by a continuous path to
the identity is a subgroup. Thus, a Lie group that is not path-connected must contain a path-connected subgroup.
Example 3.1. An infinite line with a coordinate patch −∞ < x < ∞ (x ∈ R) is a 1-dim manifold.
In section 2.1.1 we stated that C was a continuous group under addition. So is R itself, and if we write
a group element as g(x) = ex , we can easily deduce the function corresponding to the group product.
Indeed, g(z) = g(x) ⋆ g(y) = g(x + y), and we are not surprised to find that: z = φ(x, y) = x + y.
Example 3.2. Restrict θ = x ∈ R with 0 ≤ θ < 2π, and define group elements g(θ) = eiθ with product:
g(θ1 ) ⋆ g(θ2 ) = g(θ1 + θ2 mod 2π)
The group manifold here is the unit circle, S 1 , whose points are each parametrised by real angle θ,
and φ(θ1 , θ2 ) = θ1 + θ2 . Its elements are complex, but the group is real! It is Abelian, and connected.
Example

3.3. Real invertible 2 × 2 matrices form a group whose elements can be written as g(x) =
x1 x2
x3 x4 . Constraining the matrices to be unimodular (to have determinant 1) lowers the number of
parameters by 1. The group product is:

x1 x2 y1 y2 z1 z2
=
x3 1+xx12 x3 y3 1+yy12 y3 z3 1+zz12 z3
Compute the set of three functions zi = φi (x1 , x2 , x3 , y1 , y2 , y3 ) consistent with this group product:
1 + y2 y3 1 + x2 x3
z1 = x1 y1 + x2 y3 z2 = x1 y2 + x2 z3 = x3 y1 + y3
y1 x1
In this parametrisation, the mappings φi are all differentiable only off the x1 = 0 and y1 = 0 planes.
Whatever the associate manifold is—see later—it cannot be covered with just this coordinate patch.
The inverse mapping corresponding to g−1 (x) can be read off the inverse matrix g−1 .
57
Example 3.4. If we demand instead that invertible complex 2 × 2 matrices be not only unimodular,
but unitary as well, the treatment is simpler. Introduce the parametrisation:

z w a0 + i a3 a2 + i a1
=
−w∗ z ∗ −(a2 − i a1 ) a0 − i a3
with the condition |z|2 + |w|2 = a20 + a21 + a22 + a23 = 1 which guarantees that the matrix is unitary
with determinant equal to 1. The group manifold is thus the unit 3-sphere S 3 embedded in R4 with
coordinates (a0 , a1 , a2 , a3 ) ∈ R; this is a real three-dimensional Lie group.
3.2 Some Matrix Lie Groups

Amazingly enough, it turns out that almost all Lie groups of interest in physics, the so-called classical Lie groups,
are either matrix groups or groups of transformations isomorphic to matrix groups. The only group product we
ever have to consider is matrix multiplication, and inverse elements are just inverse matrices.
Following standard usage, we introduce the diagonal matrix Iqp with p entries +1, and q entries −1, where
p + q = n. In this notation In is the n-dim identity matrix, and also the Cartesian metric in Euclidean n-dim space.
One useful way of classifying Lie groups is to begin with n × n invertible matrices over some field F of
numbers, the general linear group GL(n, F), and identify interesting subgroups by constraining its elements.
Here, we focus on two types of constraints: bilinear and unimodular.
3.2.1 Bilinear or quadratic constraints: the metric (or distance)-preserving groups
Definition 3.3. Unitary transformations T of a complex matrix M ∈ GL(n, C) are defined by:
M 7→ T M T†
where the subgroup of matrices T leaves the Cartesian n-dim metric M = In invariant: TIn T† =
TT† = In . Thus, T−1 = T† , and we call that subgroup U (n) ⊂ GL(n, C): Both U (n) and its
matrices are unitary. Example 3.2 referred to U (1).
Definition 3.4. Orthogonal transformations T of a real matrix M ∈ GL(n, R) are defined by:
M 7→ T M TT
(TT is the transpose of T), such that T leaves In invariant: T In TT = TTT = In , that is, T−1 = TT .
The group of such matrices is called O(n) and is orthogonal.
Be aware that n in O(n) or U (n) refers to the dimension of the matrices, not that the group which is the number
of coordinates on its manifold! O(n) matrices have determinant ±1, whereas the absolute value of the complex
determinant of U (n) matrices is equal to 1. Thus, (can you see why?) O(n) is not path-connected; neither is U (n).
The group manifolds (and thus these groups themselves) are compact because their matrices define closed,
bounded subsets of the manifolds that parametrise GL(n, C) and GL(n, R). O(n) and U (n) preserve the length
(or norm) of n-vectors in Euclidean Rn , and therefore also angles between those vectors (eg., the angles of any
triangle are determined by the lengths of its sides).
We also have the non-compact groups which preserve the indefinite metric Iqp , defined by the transformations:
T Iqp TT = Iqp O(p, q) (3.1)
T Iqp T† = Iqp U (p, q) (3.2)
A famous example is O(3, 1), aka the full Lorentz group, that leaves the mostly positive Minkowski metric on
R4 (or space-time distance) invariant; equivalently, the norm of a 4-vector x is left invariant by 3-dim rotations,
Lorentz transformations (boosts), and space or time reflections. In principle, from the condition: T I13 TT = I13 ,
one could work out detailed constraints on the elements of the O(3, 1) matrices to find that there are six independent
parameters, but this would be needlessly messy. There are far better ways of parametrising the group to extract all
this information, and much more, as we shall see.
58
3.2.2 Multilinear constraints: the special linear groups

The special linear subgroups SL(n, C) ⊂ GL(n, C) and SL(n, R) ⊂ GL(n, R) contain all unimodular matrices.
Example 3.3 actually referred to SL(2, R) with the constraint x1 x4 −x2 x3 = 1, a bilinear constraint. SL(n, R)
is often referred to as the volume-preserving group in Rn . But it does not preserve all lengths—and metrics!
The intersection of special linear and metric-preserving groups can form important subgroups: eg., SO(n) =
O(n) ∩ SL(n, R) and SU (n) = U (n) ∩ SL(n, C). These groups are compact. Example 3.4 was about SU (2).
We stated earlier that O(n) and U (n) are not path-connected, but we know that they must have path-connected
subgroups, ie. groups with elements connected to the identity by a continuous path. These are SO(n) and SU (n).
3.2.3 Groups of transformations

Continuous transformations in physics act on vectors, or on functions of vectors. These transformations belong to
groups which are usually isomorphic to matrix Lie groups.
1. Translations Let f be an analytic function acting on Rn . The left action on f of the operator Ta associated
with Ta x = x + a is:
[Ta f ](x) = f (Ta−1 x) = f (x − a) a ∈ Rn
Except for the identity (a = 0), such transformations leave no x invariant and are called inhomogeneous.
2. Rotations
Parametrise 3-dim rotations in the z = 0 plane of a vector x ∈ R3 by Rα , with Rα φ = φ + α, with
[Rα f ](φ) = f (φ − α), and −π < φ ≤ π. In terms of the left action on the components of x: x′ = Rα x
(ie. x′ obtained by rotating x by +α in the z = 0 plane), the matrix associated with Rα is:
 
cos α − sin α 0
 sin α cos α 0 Then : [Rα f ](x) = f (Rα−1 x) = f (x cos α + y sin α, −x sin α + y cos α, z) .
0 0 1
In terms of example 2.6, this corresponds to rotating the basis by −α.

Arbitrary rotations in 3-dim space: can be written as [R f ](x) = f (R−1 x), where R can be factored
as Rα,β,γ = Rα Rβ Rγ , α, β and γ being the famous Euler angles. We will not write down a matrix
representation. It leaves lengths invariant, is unimodular, and thus is an element of SO(3).
3. We also have scale transformations x′ = ax, with a ∈ R a non-zero positive constant, and x ∈ Rn in
Cartesian coordinates (think of zooming in or out). The restriction to Cartesian coordinates is important: in
spherical coordinates over R3 , only the radial coordinate would scale.
4. Lorentz and Poincaré transformations

Lorentz boosts are given in Jackson’s Classical Electrodynamics, eq. (11.19), for R4 coordinates ct and x:
γ−1
ct′ = γ(ct − β · x) x′ = x + (β · x) β − γβ(ct)
β2
p
where β is the velocity of the primed frame in the unprimed frame, and γ = 1/ 1 − β 2 . Jackson’s eq.
−1 x in
(11.98) expresses this transformation in matrix form. To include 3-dim rotations, just replace x by Rα,β,γ
the second equation. It is not worth writing the resulting 4 × 4 matrix which will be an element of SO(3, 1)
if we exclude time reversal and space reflection; otherwise the relevant group will be O(3, 1) (or O(1, 3) in
a mostly negative metric), the full Lorentz group. The transformation is a homogeneous one, which in the
4-vector formalism is written: a′ = Λa, where a is a any 4-vector (not necessarily position).
59
We can extend the full Lorentz transformations to include space-time translations t:
x′ = Λx + t
Whereas the homogeneous transformations left the norm of a 4-vector invariant, these inhomogeneous trans-
formations leave invariant only the norm of the difference between two 4-vectors.
If we call Λ the full Lorentz transformation matrix, we can construct the matrix for these transformations by
adding to Λ a fifth row and column whose last element is a 1 that does not do anything, that is:
′
x Λ t x
=
1 0 1 1
These matrices form the 10-parameter inhomogeneous Lorentz group, or Poincaré group, ISO(3, 1). Inci-
dentally, setting Λ = I gives a matrix realisation of the 4-dim translation group.
These examples illustrate the isomorphism between physical transformations and matrix Lie groups. We can
then identify, say, a rotation with a SO(3) matrix, and even call SO(3) the rotation group.
3.2.4 Differential-operator realisation of groups of transformations: infinitesimal generators

Now we explore more deeply this isomorphism between groups of transformations of functions and Lie groups.
We shall express the left action of a few transformations as differential operators, a far from gratuitous exercise.
1. Translations
We can first look just at smooth functions f (x), x ∈ R. Then the result of a translation Ta x = x + a, a ∈ R,
on f , with a ≪ x, can be Taylor-expanded about x:

[Ta f ](x) = f (Ta−1 x) = f (x − a) = (1 − a dx + . . .) f (x) = exp(−a dx ) f (x)
In R3 this generalises to:
" ∞
#
X 1
[Ta f ](x) = f (Ta−1 x) = f (x − a) = (−ai ∂i )n f (x) = exp(−ai ∂i ) f (x) (3.3)
n!
n=0
The operators −∂i are called the infinitesimal generators of translations. Quantum mechanics uses instead
the Hermitian momentum operator p = −i~∂ and writes the translation operator as: Ta = e−ia·p/~ .
We note that the Cartesian infinitesimal generators −∂i (or pi ) commute amongst themselves.
2. Rotations
For rotations Rα φ = φ + α in the (z = 0) plane by a small angle α:

[Rα f ](φ) = f (Rα−1 φ) = f (φ − α) = (1 − α dφ + . . .) f (φ) = exp(−α dφ ) f (φ)
As we have seen in the last section, in R3 with Cartesian coordinates, this gives for the left action of a rotation
Rα x = (x cos α − y sin α, x sin α + y cos α, z): f (Rα−1 x) = f (x cos α + y sin α, −x sin α + y cos α, z).
If we Taylor-expand the right-hand side we obtain:

[Rα f ](x) = 1 + α (y ∂x − x ∂y ) + . . . f (x) = exp(α Mz ) f (x) (3.4)
where Mz = y ∂x − x ∂y . Similarly for rotations about the x and y axes, the general rotation operator is:
Rα,β,γ = exp(αMx ) exp(βMy ) exp(γMz ), where:
Mx = z ∂y − y ∂z , My = x ∂z − z ∂x , Mz = y ∂x − x ∂y (3.5)
60
or: Mi = −ǫijk xj ∂ k = 21 ǫijk J jk , where Jjk := x[k ∂j] , with xj and ∂k defining a 2-dim plane of rotation.
The pseudovector operator M is the Hodge dual of the more natural simple 2-form operator J. In quantum
mechanics, it is redefined as L = i~M and interpreted as the (Hermitian) angular-momentum operator.
These infinitesimal generators do not commute. Indeed: [Mi , Mj ] = ǫij k Mk , or [Li , Lj ] = i~ ǫij k Lk .
(Note: we could have written — some do! — defined M as the negative of the above. The cost, however,
would be an extra minus sign in the commutation relations.)
3. Dilations or scale transformations

In a n-dim space with Cartesian coordinates xµ , a scale transformation is: Tκ xµ = (1 + κ)xµ . In the limit
of small κ, Tκ−1 xµ ≈ (1 − κ)xµ . Again we Taylor-expand a function f (Tκ−1 xµ ) in the small parameter κ:

f (Tκ−1 xµ ) = (1 − κ xµ ∂µ + · · · ) f (xµ ) = exp(−κ xµ ∂µ ) f (xµ ) (3.6)
We identify D = −xµ ∂µ = −x · ∂ (compare with M = −x × ∂) as the infinitesimal generator of dilations.
We can now find the infinitesimal generators of an arbitrary group of transformations with m parameters ai
near the identity, such that ai = 0 ∀ i for the identity group element. These transformations map a point in a
manifold M n (not the group manifold!) to another one nearby that can be described by the same coordinate chart.
Let the transformations act (left action!) on a space (aka carrier space) of differentiable functions f on M n :
[Ta f ](x) = f (Ta−1 x)
Focus on Ta f , and take f as a function of the parameters ai . As before, Taylor-expand the right-hand side to first
order around the identity parametrised by a = 0:
h i
[Ta f ](x) = 1 + ai ∂ai (Ta−1 x)j ∂j + . . . f (x)
a=0
where i runs over the number of parameters, ie. the dimension of the group, and j from 1 to the dimension of the
space on which the functions f act.
Definition 3.5. The operators:

Xi = ∂ai (Ta−1 x)j ∂j (3.7)
a=0
are called infinitesimal generators of the group of transformations. In some references, the right-
hand side is multiplied by −i (with appropriate adjustment to the expansion) to ensure hermiticity.
For example, rotations in the z = 0 plane in Cartesian R3 involve one parameter (angle) a1 = α, and only x and
y derivatives can occur since z does not depend on α. Then the second term in the square bracket of eq. (3.4) is
recovered.
3.2.5 Infinitesimal generators of matrix Lie groups

Now we show how to linearise matrix groups and find their infinitesimal generators. This is not hard at all if we
know the matrices. In general, the matrix elements will be analytic functions of some (non-unique!) set of group
parameters ai , and all we have to do is Taylor-expand the matrix to first order in the group parameters around the
identity element In , for which the ai all vanish:
Ma = In + ai Xi Xi = ∂ai Ma (3.8)
a=0
where we understand that differentiating a matrix means differentiating each of its elements. The matrices Xi are
the infinitesimal generators of the group. Again, some prefer the definition Xi = −i ∂ai Ma a=0 .
61
cos θ − sin θ

Example 3.5. Let Mθ ∈ SO(2): sin θ cos θ , for 0 ≤ θ < 2π, that effects rotations in a plane.

1 −θ
Taylor-expand to first order: Mθ ≈ θ 1 = I2 + θ X
Then the infinitesimal generator of SO(2) is:

0 −1
X = ∂θ Mθ θ=0
=
1 0
a matrix fully consistent with the constraints on SO(n) generators as we shall discover in section
3.3.4. We shall write the space it spans as:

0 −θ
so(2) =
θ 0
Another example (EXERCISE) that is quite easy to work out is SL(2, R); it will have three infinitesimal
generators. Similarly, using the parametrisation of example 3.4, we see that an element of SU (2) may be written
as a0 I2 + ai Xi , where Xi = i σi are the generators of SU (2), with σi the Pauli matrices.
When the group matrices are not known we must resort to other methods to be discussed a little later.
An infinitesimal generator is an operator that effects an infinitesimal transformation away from the identity.
We want to reconstruct a finite transformation out of a succession of infinitesimal transformations that use only the
generators, ie., the first-order contribution in the expansion of a transformation or of a matrix:
n
Ma = lim (Ma/n )n = lim 1 + (ai /n)Xi
n→∞ n→∞
Now the following relations on any (well-behaved) linear operator A hold:

n ∞
X
1 A An
exp(A) ≡ lim 1+ A e = (3.9)
n→∞ n n=0
n!
since the right-hand side of the first relation is equal to its derivative when n → ∞. Therefore, Ma = exp(ai Xi ).
This exponential map, then, is the tool that reconstructs finite transformations from infinitesimal ones. But it must
be handled with some care as we shall discover.
Note that the inverse of a group element eA is e−A , and that a generator matrix A = ai Xi need not be invertible.
3.3 Lie Algebras

3.3.1 Linearisation of a Lie group product
To understand the importance of infinitesimal generators, notice that linear combinations of group elements may
not be in the group: for instance, linear combinations of SO(2) matrices are not elements of SO(2). In general,
group products are non-linear in the group parameters, so linear combinations cannot be expected to preserve them.
Linear combinations of infinitesimal generators of rotations, however, are generators of rotations! Indeed,
there is a set {Xi } of infinitesimal generators of a Lie group that forms a basis of a linear vector space. Then an
arbitrary element in the space can always be written as bi Xi , with bi the group parameters.
That vector space arises from linearising the product of a Lie group G around the identity. The result can
considerably simplify the study of the group. First, write (g, g′ ) ∈ G in the neighbourhood of the identity as
g ≈ e + aǫX and g′ ≈ e + bǫY , where ǫ is an arbitrarily small real number and a and b are real, but arbitrary as
well. X and Y are infinitesimal generators of g and g′ , respectively. Expand g g′ ∈ G to first order in ǫ:
g g′ ≈ e + ǫ (aX + bY ) + . . .
Manifestly, aX + bY is a generator for the product g g′ , and the generators indeed form a linear vector space.
62
Now expand the product h = g g′ (g′ g)−1 ∈ G to first non-vanishing order, this time writing g ≈ e + ǫ1 X +
ǫ21 X 2 /2,
and g′ ≈ e + ǫ2 Y + ǫ22 Y 2 /2, with (ǫ1 , ǫ2 ) arbitrarily small:
1 1 1 1
g g′ (g′ g)−1 ≈ e + ǫ1 X + ǫ21 X 2 e + ǫ2 Y + ǫ22 Y 2 e − ǫ1 X + ǫ21 X 2 e − ǫ2 Y + ǫ22 Y 2 + . . .
2 2 2 2
≈ e + ǫ1 ǫ2 (XY − Y X) + . . .
All other contributions of order ǫ2i and ǫ1 ǫ2 cancel out. We define [X, Y ] := XY − Y X, the commutator of the
generators X and Y . As the generator for g g′ (g′ g)−1 , [X, Y ] must be an element of the same vector space as X
and Y . When h = e, g g′ = g′ g, and the commutator of the generators vanishes. Thus, mathematicians often refer
to g g′ (g′ g)−1 as the “commutator” for the group product, but we shall reserve the term for [X, Y ].
It is straightforward to show that the Jacobi identity holds, just by expanding it:

X, [Y, Z] + Y, [Z, X] + Z, [X, Y ] = 0 (3.10)
3.3.2 Definition of a Lie algebra

Now we are ready for an important definition that collects and generalises our findings:
Definition 3.6. An algebra g is a vector space equipped with, on top of the generic addition operation,
a bilinear product g × g −→ g. When the product is the Lie bracket [·, ·], which:
• is linear: [aX + bY, Z] = a [X, Z] + b [Y, Z] ∀ a, b ∈ R or C;
• is antisymmetric: [X, Y ] = − [Y, X];

• satisfies the Jacobi identity: X, [Y, Z] + Y, [Z, X] + Z, [X, Y ] = 0.
we say that g is a Lie algebra. In physics, the Lie bracket is the commutator XY − Y X. Many,
because they
alwaysdeal with the algebra,
not the group, use G to denote g, which can be confusing.
Because X, [Y, Z] − [X, Y ], Z] 6= 0, Lie algebras are non-associative.
It is crucial to keep in mind that the action of a Lie-algebra element X on another one, Y , is not XY , but their
commutator! The closure property of a Lie group in effect translates into the existence of its algebra.
The algebra ±i g is said to be essentially real. Example: the linear and orbital angular-momentum operators
of quantum mechanics related to real infinitesimal generators.
Sometimes, however, it proves very convenient to construct a complex extension of a real or essentially real
algebra, by allowing basis redefinitions that involve complex coefficients. For instance, we might wish to construct
J± = Jx ± iJy . This provides more flexibility in constructing useful bases.
The dimension n of a Lie algebra is the number of parameters of its associated group.
3.3.3 Structure constants of a Lie algebra

The commutators of its n infinitesimal generators Xi which form a basis of a Lie algebra are themselves elements
of the algebra, so they must be written as linear combinations of those basis generators:
[Xi , Xj ] = Cij k Xk (3.11)
The coefficients Cij k are called the structure constants of the Lie algebra, whose structure they are said to
specify. In fact, with some rarely relevant caveats, they pretty much tell us everything about the group itself.
The structure constants inherit the antisymmetry of the commutators: Cji k = −Cij k . When the structure
constants all vanish, ie., when [X, Y ] = 0 ∀ (X, Y ) ∈ g, we say that the algebra is Abelian.
The Jacobi identity on elements of an algebra induces (EXERCISE) a relation between the structure constants:
Cij l Ckl m + Cjk l Cil m + Cki l Cjl m = 0 ⇐⇒ C[ij l Ck]l m = 0 (3.12)

63
Defining a matrix (Di )j k = −Cij k , we find (EXERCISE) that D satisfies the commutation relation (3.11). If
we can take the group’s representations to be unitary, as for compact groups such as SU (n) and SO(n), the
corresponding representations of the algebra are anti-Hermitian and we immediately find (EXERCISE), since they
must satisfy the commutation relations, that the structure constants are real.
The structure constants for the essentially real algebra ±i g are just (exercise) ±i Cij k . Very often, in the case
of essentially real algebras, people will call the Cij k themselves the structure constants instead of ±i Cij k .
Two Lie algebras are said to be isomorphic when they have the same dimension and structure constants, up to
a redefinition (eg. rescaling) of their generators.
3.3.4 A direct way of finding Lie algebras

Suppose we do not have an explicit parametric form for the matrix realisation of a Lie group. All we know are the
constraints on the group elements. This is sufficient to find the Lie algebra and then reconstruct the group matrix.
First, linearise the constraints. At the beginning of section 3.2 we found that for Cartesian metric-preserving
compact groups, M In M† = In ; for non-compact metric-preserving groups (when the metric is indefinite),
M Iqp M† = Iqp , with p + q = n.
Linearising for the compact groups, we get: (In + ǫ A)(In + ǫA)† ≈ In + ǫ (A† + A) = In Therefore the
matrices representing the algebra are antihermitian: A† = −A. Their diagonal matrix elements are pure imaginary
for unitary group algebras u(n); for orthogonal group algebras o(n), A is real skew-symmetric, with n(n − 1)/2
independent parameters. Thus, o(n) algebra is the set of all real skew-symmetric matrices of rank n.
If we choose to use essentially real algebras instead (eg. L as generators of so(3) instead of M in section
3.2.4), then M = In + iǫA, and the A matrices are Hermitian: A† = A.
If the group is an indefinite orthogonal group, which is non-compact, the same process yields: A† Iqp = −Iqp A.
This is a bit messier, but we can simplify it by breaking A into block matrices. If S is a q × q matrix, T a q × p
matrix, U a p × q matrix, and V a p × p matrix, then:
†
S U† −Iq 0 −Iq 0 S T
+ = 0
T† V † 0 Ip 0 Ip U V
Expanding, we arrive (exercise) at three conditions on the block matrices: S† = −S, V† = −V, T† = U.
Both the S and V diagonal blocks are antihermitian. The off-diagonal blocks are each other’s adjoint. Over R, this
means that A has two antisymmetric diagonal block matrices, one q × q and one p × p; the off-diagonal blocks
are the transpose of one another. The number of parameters of the indefinite orthogonal group O(p, q) is then
p(p − 1)/2 + q(q − 1)/2 + pq = n(n − 1)/2, the same as for the compact orthogonal group O(n).
There only remains to notice that the non-zero elements of the infinitesimal generator matrices can only be ±1
(over R) and also ± i (over C) because of the linearisation.
Another important constraint can be imposed on a group matrix M: det M = 1, which defines SL(n, R or C).
Since the determinant of a product of matrices is equal to the product of the determinants of the matrices, and
because—when a matrix A is diagonalisable— there exists a similarity transformation SAS−1 which takes A to
A′ = diag(λ1 , . . . , λi , . . .), we conclude that det A is equal to the product of the eigenvalues of A.
Also, if M = eA , it transforms as:
1 1 ′
S eA S−1 = S IS−1 + S A S−1 + S A S−1 S A S−1 + . . . = I + A′ + (A′ )2 + . . . = eA
2! 2!
′
where eA is a diagonal matrix with eλi as entries. In other words, the eigenvalues of eA are just eλi . Then:
Y X ′
det eA = eλi = exp λi = eTr A
i i
But Tr A′ = Tr(SAS−1 ) = Tr A. We obtain via this elegant (but limited to diagonalisable matrices!) deriva-
tion an important basis-independent relation, valid for any square matrix:
det M = det (eA ) = eTr A (3.13)

64
This extends to det (eA eB · · · ) = eTr (A+B+...) , and since all SL(n, R) matrices can be written as a prodiuct eA eB
(to be shown later), we immediately deduce that all matrices in the algebra sl(n, R) must have vanishing trace,
including those in su(n) and so(n). Thus, it can be said that sl(n, R) is the set of all traceless matrices of rank n.
Since antisymmetric real matrices are traceless, o(n) and so(n) are identical. This is very much related to the
absence of a continuous path from the O(n) identity (which is unimodular) to orthogonal matrices with determinant
−1: O(n) is not path-connected. Spatial inversions cannot be linearised; one cannot invert axes by a “small”
amount! So the infinitesimal generators of O(3) are those of its path-connected SO(3) subgroup of rotations.
We quote an important but difficult to prove expression which says that the familiar rule ea eb = ea+b does not
hold for matrices unless they commute! This is the so-called Baker-Campbell-Hausdorff (BCH) formula:
1 1
eA eB = eC C = A + B + [A, B] + A, [A, B] + [A, B], B + . . . (3.14)
2 12
Example 3.6. To find the matrix realisation of the generators of SO(3), which live in a three-
parameter algebra, consider counterclockwise rotations by a small angle θ around an axis whose di-
rection is specified by the vector n̂. An active transformation rotates a vector x by adding a small
vector that is perpendicular to both the axis and to x, with only vectors along the axis unchanged. By
geometry, we find that, to first-order, the transformed vector is x′ = x + θn̂ × x. Expanding gives:
x′ ≈ x + θ(ny z − nz y) y ′ ≈ y + θ(nz x − nx z) z ′ ≈ z + θ(nx y − ny x)

  
0 −θz θy x
′  θz 0 −θx  y 
⇐⇒ x =x +
−θy θx 0 z
where θ = θn̂. The matrix is an element of the so(3) algebra. How does this compare to the operator
algebra as laid out in eq. (3.5)? By identifying α = θz , etc., we can write the first order in the
expansion of the general rotation operator as:
  
0 −θz
θy ∂x
x y z  θz 0 −θx  ∂y 
−θy θx 0 ∂z
The matrix is indeed the so(3)-algebra matrix. A rotation by a finite angle θ around axis n̂ can be
k
written as: R(θ) = eθ Mk , with generators:
     
0 0 0 0 0 1 0 −1 0
Mx = 0 0 −1 , My =  0 0 0 , Mz = 1 0 0
0 1 0 −1 0 0 0 0 0
The operator and matrix algebras have the same commutator structure, [Mi , Mj ] = ǫij k Mk , estab-
lishing their isomorphism.
Often, SO(3) generators are written as Jij = ǫijk M k , which is arguably more natural. Since
(Mi )jk = −ǫijk , the matrix elements are: (Jij )lm = −ǫijk ǫklm = −(δi l δj m − δi m δj l ). The labels
(ij), i < j for J refer to the plane of rotation. To obtain their commutators, compute (EXERCISE):
i
Jmn , Jpq j = (Jmn )i k (Jpq )k j − (Jpq )i l (Jmn )l j and rearrange the eight resulting terms, yielding† :
[Jmn , J pq ] = δm p Jn q − δm q Jn p − δn p Jm q + δn q Jm p 1 ≤ m < n ≤ N, 1 ≤ p < q ≤ N

(3.15)
Only one term on the right can contribute (EXERCISE), and the commutator vanishes unless one (and
only one) number in the pair (mn) is equal to one (and only one) number in the pair (pq).
†
This relation has a short form: [Jmn , J pq ] = δ[m [p Jn] q] , that is helpful as a mnemonic device. Just start from: δm p Jn q , and generate
the other three terms by atisymmetrising with respect to p and q, then m and n, and finally both pairs together.
65
This result is important because it applies to rotations in dimensions N > 3, for which a plane of
rotation does not uniquely define an axis, as it does for N = 3. But a rotation in a 2-dim plane in
N -dim space is always about a well-defined point where all axes perpendicular to the plane meet.
Two other important and often useful results: scalar operators, ie., those that are invariant under 3-dim
rotations, must commute with the SO(3) generators (eg., the Hamiltonian for a spherically-symmetric
potential). As for a vector operator V, ie., one that transforms as a vector under rotations, it is shown
in Appendix H that it satisfies [Mi , Vj ] = ǫijk V k , or [Li , Vj ] = i ǫijk V k .
Example 3.7. The 6-dimensional so(4) Lie algebra is the set of all antisymmetric 4 × 4 real matrices,
which can be parametrised in the following way:
 
0 −a3 a2 −b1
 a3 0 −a1 −b2 
so(4) = ai Mi + bi Ni = 
−a2 a1 0 −b3 

b1 b2 b3 0
It is now appropriate to use the 4(4 − 1)/2 = 6 Jij generators, introduced in example 3.6, that
generate rotations in the (ij)-plane. With eq. (3.15) it is easy to compute the nine non-trivial so(4)
commutators, by taking Ji4 = Ni and Jij = ǫijk M k (1 ≤ i, j < k ≤ 3). Alternatively, we could use
the isomorphism with differential operators. With R4 coordinates x, y, z, u, there are six of these:
M1 = z ∂y − y ∂z , M2 = x ∂z − z ∂x , M3 = y ∂x − x ∂y
N1 = x ∂u − u ∂x , N2 = y ∂u − u ∂y , N3 = z ∂u − u ∂z
Whether with eq. (3.15) or the operator realisation, we obtain:
[Mi , Mj ] = ǫijk Mk , [Mi , Nj ] = ǫijk Nk , [Ni , Nj ] = ǫijk Mk (3.16)
1 1
The generators can be decoupled by transforming to the basis: Yi = (Mi + Ni ), Zi = (Mi − Ni ),
2 2
from which we immediately obtain the decoupled relations:
[Yi , Yj ] = ǫijk Yk , [Yi , Zj ] = 0, [Zi , Zj ] = ǫijk Zk (3.17)
By inspection, the Yi and Zi are generators of two separate su(2) (or so(3)) algebras, and so(4) =
su(2) ⊕ su(2). In terms of dimensions, 6 = 3 ⊕ 3. At group level, we say that SO(4) is locally
iso-
morphic to the direct product SU (2)×SU (2); it is globally isomorphic to SU (2)×SU (2) /Z2 since
a pair consisting of an element of SU (2) and its negative corresponds to the same SO(4) rotation;
SO(4) is globally isomorphic to SO(3) × SO(3).
i i
Then so(4) = ai Yi + bi Zi , and since [Yi , Zj ] = 0, an element of SO(4) takes the form: ea Yi eb Zi .
Example 3.8. The so(3, 1) algebra of the group SO(3, 1) derived from the metric-preserving con-
straint is:  
0 ζx ζy ζz
ζx 0 −θz θy 
so(3, 1) = 
 ζy θz 0

−θx  = θ µ Mµ + ζ ν Kν (3.18)
ζz −θy θx 0
where the infinitesimal generators can be read off:
     
0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0 0 −1 0
Mx =   My =   Mz =  
0 0 0 −1 0 0 0 0 0 1 0 0
0 0 1 0 0 −1 0 0 0 0 0 0
     
(3.19)
0 1 0 0 0 0 1 0 0 0 0 1
1 0 0 0 0 0 0 0 0 0 0 0
Kx =   Ky =   Kz =  
0 0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 0 0 0
66
One shows (EXERCISE) that the commutators of the infinitesimal generators are:
[Mi , Mj ] = ǫij k Mk [Mi , Kj ] = ǫij k Kk [Ki , Kj ] = − ǫij k Mk (3.20)
Although the number of generators is identical to so(4), there is an important difference between these
relations and the ones derived in example 3.7: the minus sign in the relation for the K, which can also
be obtained by letting N → i K. Then the complex basis in which the commutators decouple is:
L± ± ± k ± ± ∓
i = (Mi ± iKi )/2, yielding (EXERCISE) : [Li , Lj ] = ǫij Lk and [Li , Lj ] = 0.
As in example 3.6, by defining Jij = ǫij k Mk and J0i = Ki , 1 ≤ i ≤ 3, one rewrites the commutator
relations (3.20) as a relation valid for any so(p, q) algebra in N = p + q dimensions.

Jµν , Jαβ = ηµα Jνβ + ηνβ Jµα − ηµβ Jνα − ηνα Jµβ 0 ≤ (µ, ν) ≤ N − 1 (3.21)
where Jνµ = −Jµν (subscripts label generators J, not their components!), and ηµν is the Cartesian
Minkowski metric: diag (∓1, ±1, . . . , ±1), depending on the metric sign convention.
One very important realisation of this algebra interprets θi as the three angles rotating around Cartesian
axis 1 ≤ i ≤ 3, and ζi = β̂i tanh−1 β the rapidity parameters for pure Lorentz boosts along the x,
y and z axes, written in terms of the relative velocity β betweeen two inertial frames. Then so(3, 1)
is called the Lorentz algebra for Minkowski spacetime. The relation (3.20) can also be derived
(EXERCISE) in the differential-operator realisation: Jµν = xν ∂µ − xµ ∂ν .
Symmetries under rotations and Lorentz transformations, as well as under translations, are prime example of
global symmettries, in the sense that the transformations have the same form at all points. Local symmetries
involve transformations that can vary arbitrarily from one point to another, so are said to be point-dependent.
3.3.5 Hard-nosed questions about the exponential map — the fine print
Three theorems by Lie, which we have implicitly used, show that for any Lie group an algebra can be found,
characterised by the structure constants. At best only the path-connected part of a group can be recovered from its
algebra. We have relied on the exponential map to do this, but it is not always possible,
at least with just one map.
Here is a counter-example (provided by Cartan). Take: Z = x2x+x 1 x2 −x3
3 −x1
∈ sl(2, R), whose trace vanishes.
Exponentiating gives the basis-independent result (EXERCISE):

 sinh r

 I2 cosh r + Z r2 > 0
X 1  r
eZ = Zn = I2 + Z r2 = 0
n! 

n
I2 cos r + Z sin r

r2 < 0
r
where r 2 = x21 + x22 − x23 = −det Z, which makes the results basis-independent. The structure is reminiscent
of the light-cone structure obtained by endowing the parameter space R3 with an indefinite metric invariant under
SO(2, 1). Inside the light-cone, for any value of x3 , the values of the other two parameters are confined inside
a circle of radius smaller than x3 . The corresponding generators map to compact group elements. Outside the
light-cone, however, r can grow without restriction and maps to non-compact elements of SL(2, R).
So far, so good. But a glance at the above expressions shows that Tr Z
e ≥ −2 always. Yet SL(2, R) has a
large subset of elements with trace smaller than −2: matrices of the type −λ0 −1/λ0 (λ > 1), for instance. These
cannot be reached with the above exponential map.
Cartan argued that all the group elements could nevertheless be reached by writing:

x1 x2 0 −x3
Z = Za + Zb = x2 −x1
+
x3 0
67
and taking the product of the exponentials of Za and Zb , which is not eZ since [Za , Zb ] 6= 0. Then (EXERCISE):

Za Zb z+y x cos x3 − sin x3
e e = x z−y sin x3 cos x3
where z ≡ cosh r ′ ≥ 1, x ≡ xr2′ sinh r ′ , and y ≡ xr1′ sinh r ′ , with r ′2 = x21 + x22 . Each matrix is unimodular, and
the trace of the product is now 2z cos x3 = 2 cosh r ′ cos x3 , which is unrestricted.
In example 3.3 we noted that we needed more tools to tell us what the manifold of SL(2, R) was. Now we
know! The parameters of the non-compact matrix satisfy z 2 − (x2 + y 2 ) = 1 which is the positive-z hyperboloid.
Topologically, it is equivalent to R2 . The parameter values −π ≤ x3 ≤ π map the Zb subalgebra to SO(2) ⊂
SL(2, R), whose manifold is S 1 . We conclude that SL(2, R) is non-compact, and that its manifold is R2 × S 1 .
Every point is path-connected to the origin (x1 = x2 = 0) of R2 and x3 = 0 on S1 , so SL(2, R) is path-connected.
3.4 Representations of Lie Groups and Algebras

3.4.1 Representations of Lie Groups
Definition 3.7. As with finite groups, a representation Tg of a Lie group G (g ∈ G) is a homomor-
phism of G to the group of general linear matrices GL(V) acting on a space V, its carrier space.
For compact Lie groups, V is a finite-dimensional Hilbert space H, ie. a vector space over C equipped
with an inner product. For non-compact groups, it may well happen that H is infinite-dimensional.
Of special interest are irreducible representations. They satisfy Schur’s lemma: A unitary representation Tg is
irreducible if, and only if, the only operator A on H such that: A Tg = Tg A ∀g ∈ G is a multiple of the identity.
The following statements, which we quote without proof, apply to compact Lie groups:
• An irreducible representation of a compact Lie group is equivalent to an unitary representation. All unitary
representations of a compact Lie group are finite-dimensional. Thus, so are all irreducible representations..
• Every representation of a compact Lie group that is not already irreducible is fully reducible, in the sense
that it can be written as the direct sum of irreducible unitary representations.
3.4.2 Representations of Lie algebras

Lie algebras, as we have seen, can be realised as (differential) operators, or also as gl(V), the set of all linear
transformations on a Hilbert space H. We have gl(n, R) or gl(n, C) realised as n × n real or complex matrices. In
fact, a finite-dimensional algebra will always be isomorphic to some matrix algebra.
Definition 3.8. Let g be a Lie algebra. A representation T of g maps elements of the algebra to
elements of the general linear invertible matrix transformations on its carrier space (or module) V.
The mapping is a homomorphism. The dimension of a representation is that of its carrier space.
this product. Thus, if T is a

g has a Lie bracket, the commutator, and its representations must satisfy
representation of g, we must have, ∀ (X, Y ) ∈ g: T[X,Y ] = TX , TY .
3.4.3 The regular (adjoint) representation and the classification of Lie algebras
We have already noted how eq. (3.12) for the structure constants could be written as the commutator of matrices
which we now recognise as providing a new representation of the algebra:
Definition 3.9. The regular (adjoint) representation of a Lie algebra associates with each element
j
Z of the algebra a matrix RZ (or adZ ) such that RZ (Xi ) = [Z, Xi ] = Xj RZ i , where the Xi are
j
the basis generators of the algebra. Some authors use the definition [Z, Xi ] = RZ i Xj .
Clearly, the regular representation of a basis generator is just the structure constants: [Xi , Xj ] =
k
RXi j Xk = Cij k Xk . Its dimension is that of the algebra, the number of generators (or parameters).
68
We confirm that R is a representation (EXERCISE, with the Jacobi identity): [RXi , RXj ]Xk = R[Xi , Xj ] Xk .
Example 3.9. Take the defining, two-dimensional representation of the essentially real version of the
su(2) algebra with basis set Si = σi /2, where σi are the three Hermitian Pauli matrices, and whose
commutators are: [Si , Sj ] = i ǫij k Sk . Then (adSi )k j = i ǫij k , and we have† ;
     
0 0 0 0 0 i 0 −i 0
adS1 = 0 0 −i adS2 =  0 0 0 adS3 =  i 0 0
0 i 0 −i 0 0 0 0 0
Then a generic element Z = ai Si of su(2) has the Hermitian regular representation:

 
0 −i a3 i a2
RZ =  i a3 0 −i a1 
−i a2 i a1 0
Definition 3.10. A subalgebra of an algebra g is just a subspace that closes under commutation. A
subalgebra gsub is invariant if [gsub , g] ⊆ gsub , ie. if, ∀ X ∈ gsub and ∀ Y ∈ g, [X, Y ] ∈ gsub . An
invariant subalgebra is sometimes called an ideal, but we shall not be using this term.
The centre z of an algebra is the largest subalgebra that commutes with all elements of the algebra.
The centre of a commutative (Abelian) algebra is itself. z is always an Abelian invariant subalgebra.
Like the structure constants, the regular representation summarises the structure of the Lie algebra. This
algebra is a vector space spanned by a basis of generators. But we can decide to transform to another basis via a
similarity transformation. The question is: can we transform the regular representation to a basis where it takes a
form that might help classify the algebra?
Definition 3.11. If a sequence of transformations exists that puts the regular representation of a non-
Abelian Lie algebra into block-diagonal form, with the blocks irreducible non-zero subrepresentations,
the representation is said to be fully reducible. In this case, the regular representation can be written
as a direct sum of irreducible representations. Of course, these irreducible representations cannot all
be one-dimensional. In this basis, the block submatrices commute with one another.
Definition 3.12. If an algebra has no non-trivial invariant subalgebra, its regular representation is
irreducible (it leaves no proper subspace of its carrier space invariant), and the algebra is called simple.
Definition 3.13. A Lie algebra that contains no Abelian, invariant subalgebra is said to be semisimple,
ie. it has zero centre (no non-zero element commutes with all other elements). A semisimple algebra is
either simple or the sum of simple algebras (that may occur more than once in the sum). A semisimple
algebra always has at least two complementary invariant subalgebras, and there is a basis in which all
the generators of one commute with all the generators of the other(s), but not amongst themselves.
From these two definitions it follows that all simple algebras are semisimple since they are already in (single)
block form. Non-simple semisimple algebras must contain a proper, non-Abelian, invariant subalgebra.
Abelian Lie algebras (eg. u(1), so(2)) are not semisimple, and therefore not simple. Apart from so(4) (see
example below), the non-Abelian so(n) algebras are all simple, and so are the su(n) and sl(n, R) algebras.
†
The cummutation relations for the adjoint representation are: adSi , adSj = i ǫij k adSk < ahref = ”p154022a ss1.pdf ” >. With

k
our convention, adXi j = Cij k , for the adjoint representation, the structure constants for adjoint and defining represeentations are
k
always identical. With the other convention, adXi j = Cij k , they would differ by a minus sign.
69
Example 3.10. From eq. (3.16), no basis generator of so(4) commutes with all others: the algebra has
no non-zero centre! It is therefore† semisimple. Its structure constants determine the 6-dim regular
representation of a generic element of so(4) in block-diagonal form:
 
0 −a3 a2 0 0 0
 a3 0 −a1 0 0 0
 
−a2 a1 0 0 0 0 
R =  
 0 0 0 0 −b3 b2 
 
 0 0 0 b3 0 −b1 
0 0 0 −b2 b1 0
The blocks cannot be further reduced, so(3) being simple; so(4) is semisimple, but not simple.
3.4.4 The Cartan-Killing form

Again, we recall that a Lie algebra is a vector space. As such, not only does it have a basis which can be chosen at
our convenience, it can also be equipped with a (non-unique!) inner product. One such inner product is:
(Y, Z) = Tr Y Z
Definition 3.14. The Cartan-Killing form (CK-form) is a symmetric, bilinear form whose compo-
nents are the inner product of all pairs of elements of a Lie algebra in their adjoint representation:

(Y, Z) := Tr RY RZ = (RY )kl (RZ )lk (3.22)
The CK-form for basis generators Xi is easily calculated: (Xi , Xj ) = Cil k Cjk l . If the algebra has n
parameters, the CK-form has n(n + 1)/2 components.
An important property of the CK-form is its invariance under the action of any element g in the Lie group associated
with a Lie algebra. Let X and Y be elements of a Lie Algebra. Then:

g X g−1 , g Y g−1 = Tr Rg RX Rg−1 Rg RY Rg−1 = Tr (RX RY ) = (X, Y )
where we have used the property Tr A B = Tr B A. Linearising after writing: g = eǫZ , we obtain (EXERCISE):

[Z, X], Y + X, [Z, Y ] = 0 (3.23)
Definition 3.15. A CK-form is degenerate (or singular) if there exists at least one element Z in the
algebra for which (Z, Y ) = 0 ∀ Y ∈ g, ie., if the matrix (Xi , Xj ) has a row and column entirely pop-
ulated with zeros, which forces its determinant to vanish. Otherwise, the CK-form is non-degenerate.
Equivalently, a CK-form is non-degenerate if there exists a basis in which it is diagonal with all entries
non-zero. Then we say that it induces a Cartan metric g on a Lie algebra, with components gµν =
(Xµ , Xν ), where {Xµ } is that basis. If the algebra is compact, we can transform to an orthonormal
Cartan metric g = kIn ; if the algebra is non-compact, we can transform to an indefinite metric kIqp ,
with p + q = n, the dimension of the algebra. In these two cases, it is habitual to call In and Iqp
themselves the metric, which is then manifestly orthonormal.
Like all metrics, an orthonormal Cartan metric can raise and lower indices. In particular, introduce
fµνλ := Cµν ρ gρλ . Inserting gρλ = (Xρ , Xλ ), one can show (EXERCISE) with eq. (3.23) that fµνλ is antisymmet-
ric.
Now, if an algebra has a non-zero centre z (ie. an Abelian invariant subalgebra that commutes with all the
elements of the algebra), its CK-form is degenerate because the adjoint representation of any element of z vanishes
trivially. Cartan’s criterion asserts that the converse is also true, which leads to a useful alternate definition:
†
For a given i, j Yi and Zj in the decoupled basis of eq. (3.17) form an Abelian subalgebra, but it is not invariant.
70
Definition 3.16. A Lie algebra is semisimple if, and only if, its CK-form is non-degenerate,
Example 3.11. xi ∂j is a basis of the operator realisation of gl(3, R). Then xi ∂i commutes with every
other element of the algebra, and gl(3, R) has a non-zero centre. Therefore, it is not semisimple.
Example 3.12. In example 3.9, we have already obtained the adjoint representation for the generators
of su(2) — and the one for so(3) because the structure constants for the two algebras are now identical.
With S in the adjoint representation, eq. (3.22) then gives:
(Si , Sj ) = Tr (Si Sj ) = 2 δij
The CK-form is then 2I. This confirms that the CK-form for su(2) induces an invertible definite
(Euclidean) orthonormal metric, g = I. Therefore, the group is compact as well as semisimple, and
we can write the structure constants as the skew-symmetric fijk = iǫijk .
The Cartan metric as defined above in terms of the regular representation for each generator can be tedious to
calculate when the matrices are huge. So long as it is non-degenerate, however, we can extract useful information
about it with much less work by instead calculating (R, R), with R = aµ Xµ , the aµ being the parameters:
(R, R) = aµ aν Tr (Xµ Xν ) = aµ aν gµν = aµ aµ (3.24)
and (R, R) contains information about the Cartan metric gµν —more specifically, whether the algebra is compact.
Example 3.13. Go back to the defining representation used for Z ∈ sl(2, R) in section 3.3.5:

x1 x2 + x3 1 0 0 1 0 1
Z = = x1 + x2 + x3
x2 − x3 −x1 0 −1 1 0 −1 0
The corresponding independent non-zero structure constants are: C123 = 2, C312 = −2, and C231 =
−2. From these we build the regular-representation matrix:
 
0 2x3 −2x2
RZ = −2x3 0 2x1 
−2x2 2x1 0
Now, we only need to calculate the diagonal elements of R2 and sum them to get: (R, R) = 8(x21 +
x22 − x23 ). We deduce that the algebra is non-compact. That X1 and X2 are non-compact, while X3 is
compact, was determined earlier in section 3.3.5.
Interestingly enough, using the defining representation directly, we would find (EXERCISE) 2(x21 +
x22 − x23 ). This is because for semisimple algebras the defining and regular representations are both
faithful, and thus contain the same information, opening up the possibility of calculating aµ aµ in eq.
(3.24) directly from the defining representation instead of the more unwieldy regular representation.
3.4.5 Cartan subalgebra of a semismple algebra

Now we would very much like to find whether some elements Hi of a semisimple algebra have a diagonalisable
adjoint-representation matrix, and thus satisfy the eigenvalue equation:
RHi (Y ) = [Hi , Y ] = λY Y (3.25)
for some Y ∈ g, which makes Y an eigengenerator of Hi . In fact, we would like to know the maximal subset of
elements of an algebra that commute between themselves, thus forming an Abelian (non-invariant!) subalgebra.
71
Definition 3.17. A maximal Abelian subalgebra of a semisimple Lie algebra is called a Cartan sub-
algebra h. Its dimension r < n defines the rank of the algebra. It is unique up to isomorphism. The
elements of a Cartan subalgebra are called its Cartan generators. Being Abelian, its irreducible rep-
resentations are one-dimensional, and there exists a basis in which all Cartan generators are diagonal.
Example 3.14. An ordered basis of the complex extension of su(2) (Example 3.9) in its defining
representation is {S− , S0 , S+ }, where S± = √12 (S1 ± iS2 ) and S0 = S3 , with [Si , Sj ] = iǫij k Sk ,
or: [S0 , S± ] = ±S± , and [S+ , S− ] = S0 . Then the adjoint representation for S0 and S± is:
     
−1 0 0 0 0 0 0 1 0
adS0 =  0 0 0 adS+ = 1 0 0 adS− = 0 0 −1
0 0 1 0 −1 0 0 0 0
Because adS0 is diagonal, S0 is a Cartan generator; comparing with eq. (3.25), adS0 has a complete set
{S− , S0 , S+ } of eigengenerators for the corresponding eigenvalues {−1, 0, 1}, which form a basis
of the algebra. But neither S+ nor S− is diagonalisable and they are not Cartan generators, Thus, the
algebra contains only one Cartan generator and is of rank 1.
Another important thing we learn from this is that the structure constants in the complex extension of
an algebra can be quite different from those of the algebra itself, even in its essentially real version.
Indeed, the adjoint representation of S3 found in example 3.9 is not diagonal, and has only zeros on
its diagonal, in contrast with with adS0 , although adS3 does diagonalise to adS0 . Of course, this does
not affect the CK-form which, being a trace, is basis-independent.
It can be shown that the rank of a su(n) algebra is n − 1; also, so(2n) and so(2n + 1) have rank n.
3.5 Weights and Roots of a Representation of a Compact Semisimple Algebra
Definition 3.18. Let |µi be an eigenvector common to all Cartan basis generators Hi , living in the
carrier space of some representation D of the generators. Then Hi |µiD = µi |µiD . The set {µi }
corresponding to each eigenvector can themselves be viewed as the components of a r-dimensional
vector called a weight µ of the representation. The number of these weights is the dimension of D.
To find the n weights (often called a multiplet) of a representation D with matrices of rank n, simply identify
a set of r Cartan generators Hi in D , and diagonalise them if they are not in diagonal form. The ith (1 ≤ i ≤ r)
component of the j th weight (1 ≤ j ≤ n) is the (jj)th entry of the n × n matrix representing Hi . These weights
correspond to a point on a r-dimensional weight diagram, or lattice.
Definition 3.19. In a semisimple algebbra there exists a basis in which the non-Cartan generators Eα
of a semisimple algebra satisfy: [Hi , Eα ] = αi Eα , 1 ≤ i ≤ r The Eα ∈ g are then eigengenerators
(often confusingly called root vectors by mathematicians) of the element Hi of the Cartan subalgebra.
Then the set {αi } of eigenvalues can be viewed as the components of a r-dimensional vector called
the root α. We can also write [H, Eα ] = αEα . In any representation (defining, adjoint), this basis is
called the Cartan-Weyl basis.
Do keep in mind the crucial distinction between the eigengenerators, whose associated eigenvalues are the root
components, and the eigenvectors that live in the carrier space, whose eigenvalues are the components of the
weights. Also, the roots do not depend on the representation D, whereas the weights do. Indeed, one often speaks
of the weights of D as being the representation itself.
We can write an algebra g as the sum of its Cartan subalgebra, with roots zero, and the non-Cartan generators
with non-zero roots. The set of all non-zero roots define the root system of the algebra in a r-dim space.
72
As we are soon to discover, all the information about a semisimple algebra is encoded in its root system. A
Euclidean metric is induced on their space by the metric of the Cartan subalgebra, so that we can represent it as
having r Cartesian axes, each associated with a Cartan generator Hi . The root vectors can then be represented in
a root diagram. The ith component of each root is the projection of the root along the Hi axis. Being of smaller
dimension, this root space is almost always much easier to work with than the algebra itself.
3.5.1 Properties of eigengenerators in the Cartan-Weyl basis

Those eigengenerators of Hi , Eα ∈ g (all of them generators in the complex extension!), which are not Cartan
generators are quite interesting. An important fact, which we shall not prove, is that they are uniquely labelled by
their roots. To each non-zero root corresponds one and only one such generator, which spans a 1-dim subalgebra.
Now let α and β be two non-zero roots. Then, from the Jacobi identity and definition 3.19, there comes:

adHi [Eα , Eβ ] Hi , [Eα , Eβ ] = [Hi , Eα ], Eβ + Eα , [Hi , Eβ ] = (αi + βi ) [Eα , Eβ ]
If α + β is not a root, [Eα , Eβ ] = 0. Otherwise, [Eα , Eβ ] is an eigengenerator with root α + β, so we can write:
[Eα , Eβ ] = Cαβ Eα+β Cαβ ∈ R (3.26)
Using the definition, [Hi , Eα ] = αi Eα , one should now be able to see that in the Cartan-Weyl basis, the
adjoint representation of Hi is a diagonal matrix, with as entries r zeros and the αi component of the n−r roots.
Also, all diagonal entries of the adjoint representation of any other generator Eα must be zero. From this, the
following statements about the CK-form of a semisimple algebra can be derived:
(Hi , Eα ) = 0, (Eα , Eβ ) = 0 unless α + β = 0
Also, according to a theorem from linear algebra about nondegenerate symmetric bilinear forms, there exists a
basis of h in which hij = (Hi , Hj ) = kD δij , where hij are the metric components of the Cartan subalgebra.
These results, although derived in the adjoint representation, apply to the defining representation as well.
To go further, work with Hermitian Cartan generators: Hi† = Hi of the essentially real algebra. Then, if
† † †
[Hi , Eα ] = αi Eα , we immediately find that [Hi , Eα ] = −αi Eα , so that Eα = E−α . Thus, non-Cartan genera-
tors and non-zero roots always come in pairs, {Eα , E−α }. In fact, −α is the only possible multiple of α which
is a root; it always exists, otherwise (Eα , Z) = 0 ∀ Z ∈ g, and the CK-form would be degenerate. Now we
know how to compute the non-Cartan generators in the Cartan-Weyl basis from the pairs Xk and Xl of non-Cartan
generators of the algebra: E±α = A(Xk ± iXl ), with A a normalisation contant.
When β = −α, eq. (3.26) maps [Eα , Eβ ] to a generator with zero root, ie. one that lives in the Cartan
subalgebra. Therefore, [Eα , E−α ] = λi Hi for 1 ≤ i ≤ r. Taking the inner product with Hj , one quickly shows
(EXERCISE), using the cyclicity of the trace, that λi = αi Eα , E−α , so that:

[Eα , E−α ] = Eα , E−α hij αj Hi = Eα , E−α k αi Hi
where we have noted that hij = k δij for a semisimple algebra. Now [Hi, Eα ] = αi Eα determines Eα only up to
a normalisation constant, which can be chosen so as to make Eα , E−α cancel k, leaving the more simple:
[Eα , E−α ] = αi Hi := α · H summation implied (3.27)
Now is a good time to discover what those non-Cartan generators do for a living. We have:

Hi E±α |µi = [Hi , E±α ]|µi + E±α Hi |µi = (µi ± αi ) E±α |µi (3.28)
We see that E±α |µi is an eigengenerator of Hi with eigenvalue µi ± αi . The E±α act as raising/lowering
operators on the carrier space of the Cartan generators, changing weights µ by ±α. Often, the quickest way
to obtain the roots is to work out all the possible differences between neighbouring weights of a low-dimension
representation.
73
3.6 Irreducible representations of semisimple algebras

Each irreducible representation of a Lie algebra can be labelled with the eigenvalues of some function of the basis
generators of the algebra.
3.6.1 Casimir invariant operators

Definition 3.20. A Casimir invariant operator C (H. Casimir’s thesis, 1931) for a representation of
a Lie algebra is an operator that commutes with all the generators of the representation.
When the representation is irreducible, C has to be a multiple of the identity by Schur’s lemma. All elements of
an invariant subspace of the carrier space of the representation will be eigenvectors of C with the same eigenvalue.
When the algebra is semisimple, work by C. Chevalley amd Harish-Chandra (1951) guarantees the existence of a
set of Casimir operators as polynomials in the generators, whose eigenvalues may be used to label the irreducible
representations of the algebra. More precisely, each invariant subspace of the carrier space has a set of basis
vectors, each labelled by an eigenvalue of each Casimir operator. The number of Casimir operators is the rank of
the algebra.
In other words, if f (x) is in an invariant subspace of the carrier space of the algebra, then for each Casimir
operator Ci , Ci f (x) = g(x) is also in that same invariant subspace.
Because a metric can always be defined for a semisimple algebra, I claim that C2 := g µν Xµ Xν is a Casimir
operator, called the quadratic Casimir invariant, and where the Xµ are basis generators of the algebra. Indeed:
µν
g Xµ Xν , Xρ = gµν Xµ [Xν , Xρ ] + [Xµ , Xρ ] Xν = gµν Cµρλ (Xν Xλ + Xλ Xν )
= gµν gαλ fµρα (Xν Xλ + Xλ Xν )
= 0
since g µν gαλ fµρα is antisymmetric in, and the term in round brackets is symmetric in, ν and λ.
Example 3.15. From example 3.12, the metric for so(3) is, up to a constant, gµν = δµν . Then:
C2 = X µ Xµ = Jx2 + Jy2 + Jz2 = J 2
where J is the angular momentum operator of quantum mechanics. Since so(3) is of rank 1, this is the
only Casimir invariant. Then the eigenvalues of J 2 each label an irreducible representation of so(3).
Note that because of its construction, C2 is not in the algebra. In a Cartan-Weyl basis, it takes the form:
X X
C2 = gij Hi Hj + (E−α Eα + Eα E−α ) = gij Hi Hj + 2 E±α E∓α ∓ α · H (3.29)
+roots +roots
3.6.2 Irreducible representations of so(3)

We now show how working in the Cartan-Weyl basis yields without much effort the irreducible representations
of so(3), the 3-parameter algebra of the group of 3-dim rotations. All we will need in this approach are the
commutation relations in the standard basis: [Ji , Jj ] = i ǫij k Jk .
Because of these relations, only one generator can be diagonalised. Then the eigenvalues m of this Cartan
generator label the weights in each irreducible representation. The other two generators are non-Cartan. The single,
one-component root √ α can be normalised to 1. In a Cartan-Weyl basis, {E−1 , H1 , E1 } = {J− , J0 , J+ }, where
J± = (J1 ± i J2 )/ 2. From definition 3.19, [J0 , J± ] = ±J± . Eq. (3.27) then leads directly to: [J+ , J− ] = J0 .
We know enough to find the eigenvalues λ of the Casimir operator J 2 which label all irreducible representa-
tions. By definition of a Casimir operator J0 and J± commute with J 2 , so that: J 2 (J± | λ mi) = J± (J 2 | λ mi) =
λ (J± | λ mi). Also, eq. (3.28) becomes for so(3): J0 (J± | λ mi) = (m ± 1) (J± | λ mi).
74
Then J+ raises, and J− lowers, the weights m by 1, but they cannot transform | mi to an eigenvector of J 2
with a different eigenvalue λ. All the weights in a given invariant subspace are eigenvectors of J 2 with the same λ.
Next, we write relation (3.29) between the C2 Casimir operator and the generators for so(3):
J 2 = 2J± J∓ + J02 ∓ J0 (3.30)
Since an irreducible representation must be finite-dimensional, we expect that for a given λ there exists a
highest weight, mmax ≡ j, and also a lowest weight, mmin ≡ j ′ . Then J+ | ji = 0 and J− | j ′ i = 0. There comes:
J 2 | ji = j 2 | ji + j | ji = j(j + 1) | ji = λ | ji
J 2 | j ′ i = (j ′ )2 | j ′ i − j ′ | j ′ i = j ′ (j ′ − 1) | j ′ i = λ | j ′ i
Comparing yields λ = j(j + 1) = j ′ (j ′ − 1), and thus j ′ = −j. It follows that the weights m go from −j to j
in N integer steps, ie, j = −j + N , so j = N/2.
We conclude that the eigenvalues of the Casimir operator J 2 are j(j + 1), where j is a positive integer or
a half-integer, and that for a given value of j, the weights m can take 2j + 1 values, from −j to j. Therefore,
odd-dimensional irreducible representations correspond to integer j and even-dimensional ones to half-integer j.
With the help of eq. (3.30), we can now exhibit the full action of J− on a weight | jmi of J 2 and J0 . Let
J− | jmi = c− | j, m − 1i. Then, if | jmi is normalised:
hjm|J+ J− |jmi = c∗− c− = |c− |2
But since 2J± J∓ = J 2 − J02 ± J0 , we also have that:

1 1
hjm|J+ J− |jmi = hjm|(J 2 − J02 + J0 ) |jmi = j(j + 1) − m2 + m
2 2
Comparing yields c− up to an unimportant phase factor which we put equal to ±1. We find the coefficient in
J+ | jmi = c+ | j, m + 1i in a strictly analogous way. The results for both ladder operators are, up to an arbitrary
sign:
1 p
J± | jmi = √ j(j + 1) − m(m ± 1) | j, m ± 1i (3.31)
2
Each value of j labels a 2j+1-dim invariant subspace of the carrier space of so(3) of which the 2j + 1 | jmi form
a basis.
The entries of the three representation matrices, D j (J0 ) = hjm′ |J0 | jmi and D j (J± ) = hjm′ |J± |jmi, are:
j j δm′ ,m±1 p
Dm ′ m (J0 ) = m δm′ m Dm ′ m (J± ) = √ (j ∓ m)(j ± m + 1) |m| ≤ j (3.32)
2
This form for the coefficients is often quoted, but the equivalent form in eq. (3.31) is often
√ easier to use since only
√
the second factor in the root changes. The representation matrices for Jx = (J+ +J− )/ 2, Jy = (J+ −J− )/(i 2)
and Jz = J0 are easily recovered if needed. Keeping in mind that the rows and columns are labelled by the values
of m from −j to j, we have for the defining representation of so(3), labelled by j = 1:
     
0 1 0 −1 0 0 0 0 0
D 1 (J+ ) = 0 0 1 D 1 (J0 ) =  0 0 0 D 1 (J− ) = 1 0 0
0 0 0 0 0 1 0 1 0
Any other irreducible representation for integer values of j can be calculated in the same way with eq. (3.32).
Another approach relies on the actual form of the generators. In the defining, irreducible 3-dim representation
of the Cartesian basis, the three generators, which we choose to be Hermitian, are:
     
0 0 0 0 0 i 0 −i 0
J1 = 0 0 −i J2 =  0 0 0 J3 =  i 0 0 (3.33)
0 i 0 −i 0 0 0 0 0
75
Diagonalise, say, J3 with the transformation Ji −→ A−1 Ji A, where A is a unitary matrix so as to preserve
Hermiticity, and construct the non-Cartan generators in the Cartan-Weyl basis as before:
       
0 −i 0 −1 0 0 0 0 0 −i 0 i
1
J+ = 0 0 i J0 =  0 0 0 J− =  i 0 0 A = √  1 √0 1
0 0 0 0 0 1 0 −i 0 2 0 2 0
Although these generators are different from the D 1 matrices obtained from eq. (3.32), they are perfectly acceptable
as an irreduciblerepresentation
since they satisfy
both [J0 , J± ] = ±J± and [J+ , J− ] = J0 . Indeed, any pair of
0a0 0 0 0
the form J+ = 0 0 b and J− = a∗ 0∗ 0 , wih (a, b) ∈ C, satisfies these commutation relations! All these
00 0 0 b 0
irreducible representations in the Cartan-Weyl basis are equivalent.
3.6.3 Irreducible representations of su(2)

Take as the defining representation of the semisimple algebra su(2) the Hermitian generators S = σ/2:

1 0 1 1 0 −i 1 1 0
S1 = S2 = S3 =
2 1 0 2 i 0 2 0 −1
One Cartan and one pair of non-Cartan generators fit, thus one independent 1-dim non-zero root. The diagonal S3
is identified with the sole Cartan generator s0 . In this representation the weights of s0 are 1/2 and −1/2, for the
1 0
doublet of eigenvectors 0 and 1 . The roots are all the possible differences between the weights, ie. ±1.
Fom the definition of roots, [s0 , E±1 ] = ±E±1 . The structure of the algebra then determines the non-Cartan
generators in the Cartan-Weyl basis. [s0 , E±1 ] = ±E±1 gives, up to a normalisation constant A:

0 1 0 0
s+ := E1 = A = A(S1 + i S2 ), s− := E−1 = A = A(S1 − i S2 )
0 0 1 0

With h11 = 1/h11 = (Tr s20 )−1 = 2 in the argument leading to eq. (3.27), we find that: A2 = s+ , s− =
Tr (s+ s− ). The√choice A = 1 recovers the commutator [s+ , s− ] = 2s0 , as in quantum mechanics, whereas the
choice A = 1/ 2 makes Tr (s+ s− ) cancel h11 , yielding [s+ , s− ] = s0 , with the structure constant just the root
component. Then the set {s+ , s0 , s− } forms a Cartan-Weyl basis for the complex extension of su(2).
Now our discussion of the so(3) algebra in section 3.6.2 allows us to take j to be a multiple of 1/2. Inserting
j = 1/2 into eq. (3.32) then yields 2 × 2 generators identical to those of the defining representation of su(2),
whether in the standard or the Cartan-Weyl basis. This confirms the isomorphism of the su(2) and so(3) algebras.
3.6.4 Irreducible epresentations of SU(2) and SO(3)

Finite SU (2) and SO(3) transformations can be reconstructed with the exponential map, corresponding to a rota-
tion parametrised by θ = θn̂: R(θ) = eiθn̂·J for SO(3), and S(θ) = eiθn̂·S = I cos 2θ −2(n̂·S) sin 2θ (EXERCISE)
for SU (2), where the direction of n̂ is the axis of rotation.
But the isomorphism between su(2) and so(3) does not translate into an isomorphism between SU (2) and
SO(3)! A SO(3) rotation by 2π is identical to the identity, but a SU (2) rotation by 2π is equivalent to minus the
identity, because of the factor 1/2 lurking in the s matrices. We say that SU (2) and SO(3) are locally isomorphic.
There is a 2 → 1 homomorphism that maps SU (2) to SO(3): ±S(θ) → R(θ), and because of this SU (2)
can be represented by SO(3) matrices. But the map is not uniquely invertible, and therefore only SU (2) matrices
that correspond to integer j are stricto sensu representations of SO(3). Those with half-integer SU (2) j are called
spinor representations, and we say that integer and half-integer representations of SU (2) together form projective
representations Rg of SO(3), in the sense that Rg1 Rg2 = αg1 ,g2 Rg1 g2 , with α ∈ C.
Wigner matrices Dθj = eiθn̂ ·sj (with sj the triplet of basis generators of the defining representation labelled
j
by j), is the name given to the irreducible representations of SU (2), and the matrix elements are called Wigner
functions. They can be rather complicated, except when n̂ = ẑ and sz = s0 is diagonal, in which case (Dθj )m m =
′
′
eimθ δm m (|m| ≤ j). They are tabulated in many places for small j and are easily calculated by computer.
76
3.6.5 su(2) substructure of a semisimple algebra and constraints on its root system
Because they live in a r-dim space, only r of the n − r roots of a semisimple algebra can be linearly independent.
Definition 3.21. A positive root is one whose first non-zero component is positive; otherwise, it is
negative. The r positive roots which cannot be obtained from a linear combination of other positive
roots are called simple, fundamental, or independent. The other positive roots can be obtained as
linear combinations of the simple roots, with positive coefficients.
√
Each pair e±α := 2E±α /|α| of normalised non-Cartan generators of a semisimple algebra, together with
the combination: hα = 2α · H/|α|2 , forms a su(2) subalgebra. There is a su(2) subalgebra for each pair
of non-zero roots (Chevalley 1955). {hα , e±α } is called the Chevalley basis of the su(2) subalgebra . Indeed,
[eα , e−α ] = hα , but also:
√ √
2 2α 2 2 |α|2
hα, , e±α = · [H, E±α ] = ± E±α = ± 2 e±α
|α|3 |α|3
With hα = 2s0 and e±α = s± , we recover the su(2) structure constants in the Cartan-Weyl basis. Thus, a
semisimple algebra of dimension n and rank r contains (n − r)/2 generally non-distinct su(2) subalgebras, each
associated with a different root and having as Cartan generator a different element of the Cartan subalgebra, plus
two non-Cartan generators corresponding to that root.
Roots are tightly constrained by the su(2) substructure described above. Consider some other root β. Then:
h i √ √
2 2α 2 2α· β α· β
hα , e±β = 2
· H, E±β = ± 2
E±β = ± 2m e±β m :=
|α| |β| |α| |β| |α|2
Since hα /2 is a Cartan generator of su(2), we may have found another su(2) subalgebra if we can make sense of
m. Now let β + kα (k ∈ Z) be a non-zero root. Then, in the same way as above:
hh i α · (β + kα)
α
, eβ+kα = eβ+kα = (m + k) eβ+kα
2 |α|2
Then let p and q be two non-negative integers, with p the largest number for which β + pα is still a root, and q
be the largest number for which β − qα is still a root. So we have a string, or chain eβ−qα , . . . , eβ , . . . , eβ+pα
of su(2) generators that act in the root space, raising or lowering in integer steps. All elements in the set {β +
kα; k = −q, . . . m, . . . p} are roots. We can associate each of these generators with one in a (2j +1)-dim su(2)
representation labelled by j. It takes p unit steps to go from m to the highest root, and q steps to go to the lowest
root in the chain, so that −j + q = m and j − p = m, leading to q − p = 2m, and p + q = 2j. As expected for a
su(2) algebra, m (and j) is an integer or half-integer. We arrive at the master formula† :
α·β
−p≤ 2 = − (p − q) < q (3.34)
|α|2
If we had started instead with eα and added/subtracted integer multiples of β to α, we would have found that
2β · α/|β|2 = − (p′ − q ′ ). Multiplying the two master formulae yields the important expression:
(α · β)2 1
2 2
= cos2 θαβ = (p − q)(p′ − q ′ ) ≤ 1 (3.35)
|α| |β| 4
p
The relative length of the roots is seen to be constrained to |α|/|β| = (p′ − q ′ )/(p − q). Also, if α and β are
simple roots, ±(α − β) cannot be a root; otherwise, one of the two must be positive, and a simple root could be
constructed out of two different positive roots: eg., β = (β − α) + α. Thus, β − kα is not a root for any k 6= 0,
†
A derivation that does not rely on the su(2) substructure can be found in Appendix I; but it involves rather heavier calculations.
77
including k = q, and q = 0 for simple roots. Therefore, from the master formula (3.34), the angle between two
simple roots satisfies cos θαβ ≤ 0, so that π/2 ≤ θαβ ≤ π.
Since (p − q)(p′ − q ′ ) must be an integer, There are only five possible values allowed for cos2 θαβ in eq.
(3.35), and this, for any two roots of any semisimple algebra: 0 ⇒ θαβ = ±90◦ ; 1/4 ⇒ θαβ = (60◦ , 120◦ );
1/2 ⇒ θαβ = (45◦ , 135◦ ); 3/4 ⇒ θαβ = (30◦ , 150◦ ); and 1 ⇒ θαβ = 180◦ (0◦ is forbidden because no two
roots can be a positive multiple of each other).
Thanks to all these constraints, a systematic and exhaustive procedure exists to construct the root space for
all four families of classical semisimple groups, and for the five so-called exceptional groups. The root diagrams
exhibit a high degree of symmetry. All positive roots can be generated by linear combinations of the simple roots.
So-called Weyl reflections about hyperplanes perpendicular to the simple roots through the origin generate the
rest.
With the subscript denoting the rank of the algebra, the four families of semisimple groups are:
• An−1 (n > 1), corresponding to SU (n), SL(n, R), SU (p, q), with p + q = n (not the p and q above!)
• Bn , corresponding to SO(2n + 1) and SO(p, q), with p + q = 2n + 1.
• Cn , corresponding to Sp(n) and Sp(p, q), with p + q = 2n.
• Dn , corresponding to SO(2n) and SO(p, q), with p + q = n.
SU (2), SL(2, R), both A1 , and SO(3) (B1 ) , all have the same one-dim root space with the two roots ±1. Only
five two-dimensional root spaces (four classical and one exceptional) can satisfy all our constraints; but B2 and
C2 are rotated from each other by 45◦ , so are taken to be the same. And there are only four three-dimensional
root spaces. Beyond three dimensions, root spaces can no longer be represented on root diagrams. Instead, one
uses Dynkin diagrams, which are planar and represent only the simple roots and the angle between them. They are
equivalent to a root diagram.
Finally, a few words about weight diagrams. One of the Cartan generators, say H1 , will always be the Cartan
generator of a su(2) (and so(3) - see section 3.6.2) subalgebra. Then weight points are arranged on lines parallel
to the H1 axis, with each line corresponding to an irreducible representation (multiplet) of su(2) labelled with j,
an integer multiple of 1/2, and containing 2j + 1 weights. These weights can be generated by starting from the
highest weight of the representation, defined as the weight µ for which µ + α is not a weight when α is any
positive root. and applying the lowering non-Cartan generator of su(2) to the weights in each su(2) multiplet, ie.,
by repeated addition of the r-dim root, (−1, 0, . . . , 0), to that highest weight. This root, as well as (1, 0, . . . , 0)
(which moves up from the lowest to the highest weight), is always a root of the semisimple algebra. Needless to
say, as one moves parallel to the H1 axis, all other components in the weights remain the same. Subtracting a
simple root from the highest weight yields the highest weight of a neighbouring su(2) multiplet.
The number of weights for these different su(2) multiplets must add up to the dimension of the multiplet of
the semisimple algebra. The su(2) multiplets must fit snugly inside this multiplet. For instance, take the 10-
dim representation (decuplet) of su(3) of rank 2; thus the weights are 2-component vectors. The weights lie on
an inverted-triangle lattice with one horizontal su(2) quadruplet, triplet, doublet and singlet, in the direction of
decreasing H2 eigenvalues.
3.7 More on finding irreducible representations

3.7.1 Tensor product representations
Definition 3.22. Let fj1 m1 and fj2 m2 be two basis functions in the carrier space of irreducible repre-
sentations Dgj1 and Dgj2 , respectively, of g ∈ SU (2) or SO(3), such that:
′ ′
Sg fj1 m1 = fj1 m′1 (Dgj1 )m1m , Sg fj2 m2 = fj2 m′2 (Dgj2 )m2m
1 2
78
Then we form the product representation Dgj1 ⊗ Dgj2 :

′ ′
Sg fj1 m1 fj2 m2 = fj1 m′1 fj2 m′2 (D j1 )m1m1 (D j2 )m2m2 (3.36)
In Dirac notation, the product of the basis functions would read: |j1 m1 , j2 m2 i = |j1 m1 i | j2 m2 i.
Such a product is needed when a system responds to transformations in more than one way, either because of
the coupling of two separate systems (eg. particles) or because two distinct dynamical variables of one system get
coupled. A common transformation on the whole system is to be written as a direct product of transformations on
each of its parts in its own subspace.
Linearise eq. (3.36) using the generic expansion D = I + ai Xi , where X stands for a generator of SU (2)
or SO(3) in that representation. We find that the generators of the composite representation are the sums of the
generators of the distinct terms in the tensor product, so that:
X(1⊗2) (fj1 m1 fj2 m2 ) = (X (1) fj1 m1 ) fj2 m2 + fj1 m1 (X (2) fj2 m2 ) (3.37)
that is: X(1⊗2) = X (1) ⊗ I + I ⊗ X (2) or, more sloppily, X = X(1) + X(2) . When the generators have diagonal
representations, as happens with J0 (SO(3)) or s0 (SU (2)), we find, eg.:
J0 (fj1 m1 fj2 m2 ) = (m1 + m2 ) fj1 m1 fj2 m2
Note that [X (1) , X (2) ] = 0, because they act on distinct subspaces.

As before, we expect the product representation to be reducible, ie. there should exist linear combinations
φjm (or |j mi) of the product basis functions fj1 m1 fj1 m2 which transform among themselves. In other words, we
are looking for invariant subspaces of the Hilbert product space. Those linear combinations take the form of the
invertible transformation: X
φjm = j1 m1 , j2 , m2 |jm fj1 m1 fj2 m2 (3.38)
m1 ,m2

where m = m1 + m2 , and |j1 − j2 | ≤ j ≤ j1 + j2 . The real coefficients j1 m1 , j2 , m2 |jm are known as
Clebsch-Gordan or Wigner coefficients. They are unique up to a phase convention.
One easy way to obtain the φjm in terms of the fj1 m1 fj2 m2 is to start with the highest weight component,
m = j1 + j2 , of the highest j irreducible representation: j = j1 + j2 . Of course, φj1 +j2 ,j1+j2 = fj1 j1 fj2 j2 . Next,
apply J− on the left and on the right, using eq. (3.37), until the lowest weight component of the j irreducible rep-
resentation, φj,−j , is reached. Now obtain the linear combination for the highest weight of the j - 1 representation,
φj−1,j−1 , by demanding that it be orthogonal to φj,j−1 , and repeat with J− . Continue until all values of j allowed
by |j1 − j2 | ≤ j ≤ j1 + j2 have been reached.
3.7.2 Irreducible tensors

Suppose that a set of functions fjm in the carrier space of SU (2) or SO(3) transforms under a group element
′
parametrised by θ = θn as: Rθ fjm = fjm′ (Dθj )m m . Then the set {fjm } form a basis for an irreducible
representation of SU (2) or SO(3) labelled by j.
Definition 3.23. Let {Tjm } be a set of operators on the carrier space that transform as:
′
Rθ Tjm Rθ−1 = Tjm′ (Dθj )m m
(3.39)
Then we say that they are the components of a rank-j irreducible (or spherical) tensor.
If we linearise this equation, we obtain (EXERCISE) a more useful alternative definition of irreducible tensors in
terms of generators J (j) of an irreducible representations of the algebra, preferably in the Cartan-Weyl basis:
′
[J (j) , Tjm ] = Tjm′ (J (j) )m m (no summation on j) (3.40)
79
where j is the label of the irreducible representation. For SU(2) or SO(3):

p
[J0(j) , Tjm ] = m Tjm , [J±(j) , Tjm ] = (j ∓ m)(j ± m + 1) Tj,m±1 (3.41)
As a direct consequence of these commutation relations, the matrix element of Tjm , hj2 m2 |Tjm |j1 m1 i, vanishes
unless m2 = m1 + m and |j1 − j| ≤ j2 ≤ j1 + j. These are a version of the famous vector addition rules.
3.7.3 The Wigner-Eckart theorem

The Wigner-Eckart theorem asserts that if Tj is a spherical tensor under SU (2), then its matrix elements, written
in bra-ket notation, hj2 m2 |Tjm |j1 m1 i, can be factored as:
(j m , jm|j2 m2 )
hj2 m2 | Tjm |j1 m1 i = 1 √1 hj2 || Tj ||j1 i (3.42)
2j2 + 1
Tj is a tensor wh
where hj2 ||Tj ||j1 i is called the reduced matrix element and does not depend on m, m1 or m2 . So the depen-
dence of the matrix element on these numbers is carried entirely by the Clebsch-Gordan coefficient!
The Wigner-Eckart theorem applies to unitary representations of Lie groups, not only SU (2). The Clebsch-
Gordan coefficients and the labelling with eigenvalues of Casimir operators will be appropriate to the Lie group.
As a result, ratios of matrix elements for a given j but different m are just ratios of Clebsch-Gordan coefficients.
Example 3.16. When T transforms as a scalar under some Lie group, the relevant representation
matrix of the group is the identity matrix. For SU (2), j = m = 0, and the vector-addition rules
√
collapse the Wigner-Eckart theorem to: hj2 m2 | T |j1 m1 i = δj1j2 δm1m2 hj2 || T ||j1 i/ 2j2 + 1. Then
matrix elements of scalar operators between weights of different irreducible representations vanish.
The importance of the Wigner-Eckart theorem resides in its separating symmetry-related (“geometrical”) as-
pects of matrix elements from other (“dynamical”) aspects stored in the possibly unknown reduced matrix element.
3.7.4 Decomposing product representations

The problem of decomposing representations of a semisimple group into their irreducible representations can often
be treated in a fairly intuitive way. Consider SO(3) again, and its 3-dim carrier space of functions f (x) and g(y)
(eg. the wave-functions of two particles), each transforming in some known way under 3-dim rotations. We can
form tensor products, f (x) ⊗ g(y) whose transformation properties are derived from those of the functions.
For instance, if our functions were 3-dim vectors, we would have a 9-dim product representation (with nine
weights, or basis vectors for its carrier space) , with components T ij , which under rotations R would tranform as:
T ′ij = Ri k Rj l T kl (3.43)
The 6-dim symmetric part of T ij rotates into a symmetric object, and the 3-dim antisymmetric part into an anti-
symmetric one. Thus, we have easily found invariant subspaces. Moreover, the trace of T ij , T ii , is invariant under
rotations, forming a 1-dim invariant subspace that should be separated out from the symmetric part.
Note that the trace is obtained by contracting T ij with the metric of the carrier space, with components gij ,
which here is just the identity matrix invariant under rotations. Similarly, the antisymmetric part can be obtained
with the Levi-Civita symbol that is also invariant under rotation. Thus, we can write:

ij 1 ij ji
1 ijk lm 1 2 ij k 1 1
T = T + T + ǫ ǫklm T = T + T − g T k + (T ij − T ji ) + gij T kk (3.44)
ij ji
2 2 2 3 2 3
The numerical coefficient of the trace term has been chosen so as to make the symmetric term traceless.
But we can also think of eq. (3.43) as a 3 ⊗ 3 exterior direct product of a rotation with itself, so a 9 × 9 matrix,
with each row labelled by a pair {ij} and each column labelled by a pair {kl}, acting on a 9×1 matrix with entries
T kl labelled by the pairs {kl}. The direct-product matrix is a representation of SO(3). Indeed, under a rotation
80
R1 followed by R2 , T ij −→ (R2 R1 )i m (R2 R1 )j n T mn , where now the 9 × 9 matrix is formed from the matrix
product R2 R1 . Being reducible, the representation can be transformed via an angle-independent similarity matrix
to a block-diagonal matrix with a symmetric traceless 6 × 6 block (which acts only on the symmetric traceless part
of T) , an amtisymmetric 3 × 3 block acting only on the antisymmetric part of T , and a 1 acting only on the trace.
We obtain the following decomposition into irreducible representations: 9 = 5 ⊕ 3 ⊕ 1.
As expected, the total dimensions on the left and right match. The result is also consistent with what we would
find by decomnposing a j1 ⊗ j2 = 1 ⊗ 1 SO(3) product representation with the method of section 3.7.1 to obtain
a direct sum of three irreducible representations labelled by j = 2, j = 1, and j = 0.
Appendices
H Commutators of Angular Momentum with Vector Operators
Take a unit vector û and a vector operator V with components Vu with respect to û. In example 3.6 we found that
under a rotation R(θ) by a small angle θ about an axis n̂, û′ = û + θ n̂ × û. Then:
Vu′ = V · û′ = V · û + θ V · n̂ × û = Vu + θ V · n̂ × û
Also,

Vu′ = R(θ) Vu R† (θ) = e−i θn̂·L Vu ei θn̂·L ≈ (1 − i θn̂ · L) Vu (1 + i θn̂ · L) ≈ Va − i θ n̂ · L, Vu (H.1)

Consistency then demands that: n̂ · L, Vu = i V · n̂ × û = i ǫijk V k ni uj . With n̂ along the the x-axis and
û along the y-axis, there comes:
Li , Vj = i ǫijk V k (H.2)
81
I Alternative Derivation of the Master Formula

This derivation involves finding the (real!) structure constants Cαβ in eq. (3.26), without explicit calculation of
commutators. They satisfy symmetry relations, such as, from (3.26) and its adjoint:
∗
Cβα = − Cαβ C−α,−β = − Cαβ = − Cαβ (I.1)
Also, let α, β, and α + β be non-zero roots; then γ = −(α + β) is also a non-zero root, Using the Jacobi
identity on Eα , Eβ , and Eγ , plus eq. (3.26) and (3.27), leads (EXERCISE) to:
(α Cβγ + β Cγα + γ Cαβ ) · H = 0
The Hi being linearly independent, this can only be satisfied if: α Cβγ + β Cγα + γ Cαβ = α(Cβγ − Cαβ ) +
β(Cγα − Cαβ ) = 0, which yields additional symmetries on the structure constants of a semisimple algebra:
Cβ,−α−β = C−α−β,α = Cαβ (I.2)

Going back to eq. (3.26), we can write: [Eα , Eβ+α ] = Cα,β+α Eβ+2α , . . . , [Eα , Eβ+kα ] = Cα,β+kα Eβ+(k+1)α .
But there must exist a value k = p ≥ 0 such that β + (p + 1)α is not a root, so that Cα,β+pα = 0. Similarly, if
we start from: [E−α , Eβ ] = C−α,β Eβ−α , there must exist a value k = −q ≤ 0 such that β − (q + 1)α is not a
root, and C−α,β−qα = 0.
Next, start from the always useful Jacobi identity and evaluate the commutators using eq. (3.26) and (3.27):

Eα , [Eβ+kα , E−α ] + Eβ+kα , [E−α , Eα ] + E−α , [Eα , Eβ+kα ] = 0
=⇒ [Eα , Eβ+(k−1)α ] Cβ+kα,−α − [Eβ+kα , α · H] + [E−α , Eβ+(k+1)α ] Cα,β+kα = 0
=⇒ Cα,β+(k−1)α Cβ+kα,−α + α · (β + kα) + C−α,β+(k+1)α Cα,β+kα = 0
Applying relations (I.1) and then (I.2) to the first and last term on the left yields the recursion relation:
2 2
Cα,β+(k−1)α = Cα,β+kα + α · (β + kα)
2
We already know that, by definition of p, Cα,β+pα = 0. Then, from our recursion relation, Cα,β+(p−1)α =
2 2 2 2 2
α · β + p|α| , Cα,β+(p−2)α = Cα,β+(p−1)α + α · β + (p − 1)|α| = 2α · β + (p − 2)|α| , etc. Generically:

2 p+k
Cα,β+(k−1)α = (p − k + 1) α · β + |α|2
2
The recursion stops when k = −q, ie. when C−α,β−qα = −Cα,−(β−qα) = −Cβ−(q+1)α,α = 0:

2 p−q
0 = Cα,β−(q+1)α = (p + q + 1) α · β + |α|2
2
or:
α·β
2 = − (p − q) (I.3)
|α|2
82
4 CHAPTER IV — Solution of Differential Equations with Green Functions

Physical quantities are generally represented by functions of up to four (three spatial and one time) variables and
therefore satisfy partial differential equations (PDE). More precisely, let y(x1 , . . . , xn ) be a variable dependent on
the independent variables x1 , . . . , xn , then y may have to satisfy equations of the form:

∂y ∂m y i
f y, , ..., ,x = 0 (4.1)
∂xi ∂xi ∂xj . . .
where 0 ≤ i, j, . . . ≤ m, with the constraint: i + j + . . . = m.

If this equation can be split into:

∂y ∂my i
g y, , ..., i j ,x = F (xi )
∂xi ∂x ∂x . . .
it is said to be inhomogeneous. If F (xi ) = 0, g = 0 is said to be a homogeneous equation.

You may be relieved to know that in physics we almost never have to go beyond m = 2. Still, PDEs can
be extremely challenging, and most have to be solved numerically. Very thick books have been written on
techniques for numerically solving PDEs, and we will not even attempt to broach the topic. In some impor-
tant cases, PDEs in n independent variables can be converted into n ordinary differential equations (ODE).via
the technique of separation of variables. To test whether a PDE has completely separable solutions, insert
y(x1 , . . . , xn ) = X1 (x1 )X2 (x2 ) . . . Xn (xn ) into it, and see if it can be written as a sum of terms, each of which
depends on one xi only. If that happens, the PDE can be satisfied only if each term is equal to a constant, called the
separation constant, with all the constants summing to zero. Then we are left with n ODEs, one for each Xi (xi ).
If the solution to each of these ODEs is unique, this solution to the PDE will also be unique.
In the next few sections, we shall discuss linear ODEs of first and second order, returning to PDEs later.
4.1 One-dimensional Linear Differential Operators

A differential operator L of order n is said to be linear over an interval a ≤ t ≤ b if its action L[f ] on all the
functions f (t) in its domain D is linear in the functions and all their derivatives present in the operator. Linearity
means that if f1 and f2 are any two functions f1 , f2 ∈ D, then L[c1 f1 + c2 f2 ] = c1 L[f1 ] + c2 L[f2 ], where c1 and
c2 are constants. Formally (without mention of D), in one dimension L is written as:
n
X
L = pj (t) djt (4.2)
j=0
L is not considered to be specified until D is given. At this stage, this consists of all n-times differentiable
functions, but we will want to restrict it further. First of all, D should be a vector space, H, whose elements are all
the square-integrable functions on [a, b], using the inner product:
Z b
(f, g) := f ∗ (t) g(t) dt (4.3)
a
We might want to know whether L has eigenfunctions, ie., whether there exist some φ ∈ D such that L scales
φ by a constant factor λ. This, of course, is the eigenvalue problem: L[φ] = λ φ. It requires L : H → H, which
is not the case in general for differential operators. Also, if it can be shown that the set {φi } of eigenfunctions is a
basis for H, then L[y] ∈ H, ∀ y ∈ H. Achieving this will restrict the formal L itself, not only its domain.
Can L have 0 as eigenvalue? If so, we say that the associated eigenfunction, fh , is a solution to the homoge-
neous equation L[y] = 0. Or, instead, we can restrict the image of L to be a specific function F (t), the source
or driving term, and look for the c orresponding function(s) in D, if any. This means solving the inhomogeneous
equation L[y] = F , and the existence and uniqueness of a solution, finh , implies the existence of an inverse oper-
ator L−1 such that finh = L−1 [F ]. As we are soon to discover, the existence and uniqueness of L−1 is connected
to the eigenvalue problem with λ = 0.
83
Also, solving a nth -order ODE requires specification of boundary conditions (B.C.) on the solution Thus, the
domain of L also depends on the BC imposed on the functions on which it is allowed to act. Then invertibility of
L is also very much dependent on those BC. We address each each of these questions in turn.
4.1.1 Existence of the inverse of a linear differential operator

In linear algebra it is well known that a matrix operator L has an inverse if, and only if, its determinant is non-zero,
in which case the equation LX = Y has the unique solution X = L−1 Y. When L is a linear differential operator,
we say that it is invertible if it is one-to-one. And it is one-to-one if, and only if, the only function it sends to zero
is the zero function.
To prove this, assume that L[f ] = 0 has for unique solution f = 0, and that L[g] = L[h] = F . Then
L[g] − L[h] = L[g − h] = 0 by linearity, and we have g = h. Therefore, there is only one function mapped to
F (including F = 0!) by L. Conversely, assume that L is one-to-one, and that there exists a function fh 6= 0
such that L[fh ] = 0. But then Cfh , with C an arbitrary constant, is also a solution of this equation, and L is
many-to-one, contradicting our assumption. Thus, L is invertible if, and only if, it has no zero eigenvalue.
The existence of a non-trivial solution to the homogeneous equation depends very much on the boundary
conditions imposed on the most general solution. which completely determine a unique solution if it exists.
4.1.2 Boundary Conditions

The most general solution to an inhomogeneous equation takes the form f = fh + finh , ie., the sum of a homo-
geneous and an inhomogeneous solution, and where L[f ] = L[fh ] + L[finh ] = L[finh ] = F . We require that the
task of satisfying the B.C. fall entirely to fh , so that finh satisfies homogeneous B.C., ie., it contributes nothing to
the B.C. on f . As for the source term F , although it may happen that it vanishes at the boundaries, we do not want
to constrain it other than being piecewise continuous.
In the theory of linear operators of order n, B.C. are often expressed in the form of n linear combinations of
f and all its derivatives of order n-j (1 ≤ j ≤ n), evaluated at the two end-points a and b, with b > a. These
can be written as a matrix equation involving two matrices, A and B, with constant coefficients: Afa + Bfb = C,
where fa and fb are vectors with components {f, f˙, . . . , dtn−1 f } evaluated at a and b, respectively, and C is a
given constant vector. When C = 0, we say that the B.C. are homogeneous. To have fh (and its derivatives) zero
everywhere in the interval [a, b], both fa and fb must vanish, which requires the B.C. to be homogeneous. Then L
is indeed invertible.
But what if the B.C. on fh are inhomogeneous (non-zero)? Then we can consider two linear operators with
the same form L: L[D h ], where D h is the set of functions fh with homogeneous B.C. that satisfy L[fh ] = 0, and
L[D inh ], where D inh is the set of functions fh with inhomogeneous B.C. that satisfy L[fh ] = 0. Here, “inhomoge-
neous” applies to the B.C., not the inhomogeneous solution which always has homogeneous B.C. Only L[D h ] can
be invertible, when D h contains only the zero function.
In practice, B.C. are usually expressed as n arbitrary values assigned to f and/or its derivatives at the bound-
aries, and they come as two main types:
(1) One-Point (Initial) conditions, aka Initial-Value Problem (IVP): In the formal theory, one matrix in Afa +
Bfb = C, say B, is set to zero, so that only one point, the initial “time” a, is involved, and A is diagonal.
Therefore, f and its n-1 derivatives take known values (or can be set arbitrarily) at t = a. Then a theorem
shows that the solution to the one-dim IVP exists and is unique.
(2) Two-point boundary conditions, or Boundary-Value Problem (BVP): this time the n known or specified
values of f and its derivatives can be at both end-points a and b. This is a much more complicated situation,
with neither existence of a solution nor its uniqueness guaranteed.
In the most prevalent case, f (a) and f (b) are known (Dirichlet problem), or its first derivatives at a and b are
known (Neumann problem). Periodic B.C., where f (a) = f (b) and f˙ a = f˙ b , can also occur.
84
4.1.3 First-order linear ODEs

It is not difficult to obtain an explicit general solution to a first-order IVP. First, assume that the ODE has been
converted to its normal form:
L[f ] = dt f + β(t) f = F (t)
Then one shows (EXERCISE) that if β(t) is continuous over [a, b], the general solution of this IVP is:
Z t Rt
i(a) i(t′ ) ′ β(t′ ) dt′
f (t) = f (a) + F (t′ ) dt i(t) = e (4.4)
i(t) a i(t)
Notice that fh = i(a)f (a)/i(t) solves the homogeneous equation: dt f + β(t)f = 0, and that the inhomogeneous
term in eq. (4.4) satisfies homogeneous B.C., as expected. Thus, fh (a) = f (a). When f (a) = 0, f (t) = 0.
With f (a) specified, the solution is unique. Indeed, let g(t), with g(a) = f (a), also satisfy L[g] = F . Then
h = g − f solves the homogeneous ODE with homogeneous B.C. h(a) = 0, which forces h(t) = 0 for t > a.
4.1.4 Second-order linear ODEs

The most general form for a linear, second-order equation over the interval [a, b] is:
L[f ] = α(t) d2t f + β(t) dt f + γ(t) f = F (t) (4.5)
where α 6= 0, β and γ are required to be continuous, while F is piecewise continuous.

Introduce the Wronskian of two differentiable functions, f1 (t) and f2 (t), defined as: W (t) := f1 f˙2 − f2 f˙1 . If
there exists no constant C such that f2 = Cf1 ∀ t, f1 and f2 are said to be linearly independent. The Wronskian
provides a handy test for linear independence: two differentiable functions that do no vanish anywhere in an
interval are linearly dependent over that interval if, and only if, their Wronskian vanishes everywhere (EXERCISE).
The Wronskian of two homogeneous solutions of eq. (4.5) obeys a first-order differential equation whose
solution is Abel’s formula (EXERCISE):
Rt ′ ′ ′
− t0 [β(t )/α(t )]dt
W (t) = W (t0 ) e (4.6)
where t0 is any point in the interval [a, b]. In the important case that β = α̇, eq. (4.6) leads to αW being constant.
Also, if the Wronskian of two homogeneous solutions vanishes anywhere, it vanishes everywhere, because the
exponential cannot vanish in a finite interval.
Given one solution, f1 , of eq. (4.5), an immediate useful application
of the Wronskian generates a second
linearly independent solution. Noticing that W (t)/f12 = dt f2 /f1 and integrating, we find with eq. (4.6) that:
Z t
W (t0 ) − R t′ (β/α)dt′′ ′
f2 (t) = f1 (t) e dt (4.7)
a f12 (t′ )
Discarding any term proportional to f1 leaves a solution that is linearly independent from f1 .
And now comes a surprising fact, courtesy also of the Wronskian: given two independent solutions of the
homogeneous equation, a solution of the inhomogeneous equation: L[f (t)] = F (t), can be generated which
satisfies homogeneous B.C. Appendix J presents a simplified version, leading to eq. (J.1), of this variation of
parameters method discovered by Euler and Lagrange. Shortly, however, we shall explore another method which
yields the same resultsR while providing much deeper insight.
Note also that if β(t)/α(t)dt exists within the interval of interest, it is always possible to eliminate the
first-order derivative term in any linear second-order ODE, with a redefinition of the form f (t) = g(t)eµ(t) (the
substitution f (t) = µ(t)g(t) also works), to arrive (EXERCISE) at the normal Sturm-Liouville form:
β Z t
1 1 β2 F (t) −µ(t) F (t) β ′
g̈(t) + γ − dt − g(t) = e = exp dt (4.8)
α(t) 2α 4 α α(t) α(t) 2α
85
as determined by the requirement that the transformed ODE have no first-order derivativeof g. In the frequent case
α = 1 and β and γ constants, this assumes the much simpler form: g̈(t) + γ − (β/2)2 g(t) = F (t)eβt/2 .
Let f1 (t) and f2 (t) be two independent solutions of L[f ] = 0. Then fh = c1 f1 + c2 f2 , with c1 and c2 deter-
mined from the B.C. on fh , is the general solution of the homogeneous equation (principle of linear superposition).
4.1.5 Second-order IVP

It can be shown (the technical proof is presented in Appendix K) that the only homogeneous solution of a second-
order IVP, L[f ] = F , for which f and dt f both vanish at the initial point t = a, is the trivial solution fh = 0.
Consequently, if there exist two solutions f and g such that f (a) − g(a) = 0 and f˙ a − ġ a = 0, then f = g
everywhere. D h = {0}, and L[D h ] is indeed invertible. We conclude that the general solution to a second-order
IVP always exists and is unique. With the inhomogeneous solution derived in Appendix J, we find:
Z ∞
′ f1 (t′ ) f2 (t) − f2 (t′ ) f1 (t)
f (t) = c1 f1 (t) + c2 f2 (t) + θ(t − t ) F (t′ ) dt′ (4.9)
a α(t′ ) W (t′ )
where θ(t − t′ ) is the step-function which vanishes for t < t′ and equals 1 when t > t′ , and:
f˙2 (a) f (a) − f2 (a) f˙(a) f˙1 (a) f (a) − f1 (a) f˙(a)
c1 = , c2 = −
W (a) W (a)
4.1.6 Second-order BVP

The Dirichlet problem for a 2nd -order differential operator is also addressed in Appendix J, and we will re-visit it
at length in the context of Green functions. The Neumann problem is fiddlier, and we will say much less about it.
For a start, there may be a constraint on the B.C., or on the source. For instance, integrating d2x f = F over [a, b]
b R
immediately yields dx f a = F (x)dx. For homogeneous Neumann B.C., this translates into a constraint on the
source.
4.2 Solving One-dimensional Second-order Equations with Green Functions (BF 7.3)
4.2.1 Solutions in terms of Green Functions
We shall now investigate the conditions that allow the existence and uniqueness of finh formally written as
finh (t) = [L−1 F ](t), where L−1 is an integral operator whose action on F (t) is:
Z
−1
L F (t) = G(t, t′ ) F (t′ ) dt′ (4.10)

Assuming that F is square-integrable over some interval, we want L−1 to return a square-integrable result L−1 F (t),
ie., finh (t); this is the case if the two-point function G(t, t′ ) is itself square-integrable over the interval (see the
end of section BF7.1 for more details). R
Now, acting on the above equation with L gives: [Lf ](t) = F (t) = [Lt G](t, t′ ) F (t′ ) dt′ . This is satisfied
provided G(t, t′ ) obeys:
Lt G (t, t′ ) = δ(t − t′ ) (4.11)
We shall refer to this as the defining equation for a Green function G(t, t′ ) of L. It should be expected that any
indefinite solution of eq. (4.11) must be supplemented with B.C. related to those on f (t).
For the Green function to exist, L must be invertible, which we have seen requires that there be no non-trivial
homogeneous solution with homogeneous B.C.
The link between the existence of the Green function and the criterion for invertibility of L can be made more
tangible if L has a complete set of orthonormal eigenfunctions φj of L, with associated eigenvalues λj , on the
interval. A version of the spectral theorem of operator theory asserts that such a set exists if L is in self-adjoint
86
form, ie., if β = α̇ in eq. (4.5). Even if it isn’t, it can always be put in such a form by multiplying it by a function
w and imposing dt (wα) = wβ, which determines w up to a constant.
Then finh can be expanded over the subset R that satisfies homogeneous B.C. (with unknown coefficients aj ),
and so can F , with known coefficients bj = φ∗j (t′ )F (t′ ) dt′ . Both sets of coefficients are, as usual, projections of
f and F on the eigenfunctions. The eigenvalue equation then yields a relation between them, and, assuming that
integral and summation signs can be interchanged, there comes (EXERCISE) the inhomogeneous solution:
X φj (t) Z Z hX i
∗ ′ ′ ′
finh (t) = [φj (t ) F (t ) dt = φj (t) φ∗j (t′ )/λj F (t′ ) dt′
λj
j j
R
We can write this solution as f (t) = G(t, t′ )F (t′ ) dt′ , so long as the Green function:
X φj (t) φ∗ (t′ )
′ j
G(t, t ) = (4.12)
λj
j
exists, ie., only if there is no non-trivial φj satisfying homogeneous B.C. such that L[φj ] = 0. Note, however, that
even if G(t, t′ ) defined as obeying equation (4.11) does not exist, the solution f (t) might still exist, provided
that the φj associated with λj = 0 satisfy bj = 0. But such a solution would be far from unique, because any
multiple of φj could be added to it (see Appendix L for more details).
4.2.2 1-dim Green Functions without boundary conditions

What restrictions does eq. (4.11) impose on G(t, t′ )? Two, in fact:
(a) G(t, t′ ) is a continuous function of t everywhere, including at t = t′ , otherwise its second derivative at t = t′
would be the derivative of a δ-function, and the differential equation would not be satisfied. Note, however,
that the Green function for a first-order operator is discontinuous, eg., L = −idt has as Green function the
step-function i θ(t − t′ ).
(b) Ġ must have a discontinuity at t = t′ . To see this, integrate eq. (4.11) from t = t′ − ǫ to t = t′ + ǫ. Since the
coefficients in L are continuous, they hardly vary when the interval is arbitrarily small (ǫ → 0). In that limit,
the integrals of G and Ġ both vanish because G is continuous, and only the integral of G̈ contributes:
t=t′ +ǫ 1
lim Ġ(t, t′ ) =
ǫ→0 t=t′ −ǫ α(t′ )
Because of the discontinuity in its derivative at t = t′ , G should be different on either side while satisfying
[LG](t, t′ ) = 0, so that it can be written in terms of f1 and f2 :

a1 (t′ ) f1 (t) + a2 (t′ ) f2 (t) t′ < t
G(t, t′ ) =
b (t′ ) f (t) + b (t′ ) f (t) t′ > t
1 1 2 2
The continuity of G and the discontinuity in Ġ at t = t′ then yield the matrix equation at t′ :

f1 (t′ ) f2 (t′ ) a 1 − b1 0
=
f˙1 (t′ ) f˙2 (t′ ) a 2 − b2 1/α
For the system to have a solution, the determinant of the matrix, ie. the Wronskian, W ≡ f1 f˙2 − f˙1 f2 , cannot
vanish anywhere, or else it would vanish everywhere, and f1 and f2 would not be independent as postulated. Then:

a1 − b1 1 −f2 (t′ )
=
a2 − b2 α(t′ ) W (t′ ) f1 (t′ )
87
Eliminating a1 and a2 with this equation, the Green function for L without B.C. must take the general form:
 ′ ′
b1 (t′ ) f1 (t) + b2 (t′ ) f2 (t) − f1 (t) f2 (t ) − f2 (t) f1 (t )

t′ < t
G(t, t′ ) = [α W ](t′ ) (4.13)

b (t′ ) f (t) + b (t′ ) f (t) ′ >t
1 1 2 2 t
The term with the Wronskian vanishes at t = t′ , ensuring the continuity of G as required. The adjustable parame-
ters b1 and b2 can now be chosen so that G satisfies suitable boundary conditions.
4.3 Green functions for the IVP and the BVP

(1) Initial-value problem
In an IVP, we must impose the following initial conditions on G: G(a, t′ ) = dt G(t, t′ ) t=a = 0. Since a < t′ ,
the relevant expression in the general Green function (4.13) is the one for t < t′ , ie., the solution of the
homogeneous equation which must vanish as seen in section 4.1.5, so b1 = b2 = 0. There comes the unique:
f2 (t)f1 (t′ ) − f1 (t) f2 (t′ )
Givp (t, t′ ) = θ(t − t′ ) (4.14)
[αW ](t′ )
The step-function does not make Givp (t, t′ ) discontinuous because the rest of the expression vanishes at t′ = t.
Eq. (4.9) can now be written as:
Z ∞
f˙2 (a) f (a) − f2 (a) f˙(a) f1 (a) f˙(a) − f˙1 (a) f (a)
f (t) = f1 (t) + f2 (t) + Givp (t, t′ ) F (t′ ) dt′
W (a) W (a) a
(4.15)
with Givp given by eq. (4.14). A physicist is pleased that the B.C. guarantee causality: Givp (t′ > t) = 0.
(2) Two-point boundary-value problem
Although superficially similar, the two-point boundary-value problem (BVP) requires a little more care. Many
treatments enforce homogeneous B.C. on Gbvp through B.C. on f1 and f2 (or their derivatives) at the end-
points, a strong restriction. Here, we follow a slightly different approach that, as in the IVP, initially only
assumes linear independence of f1 and f2 , without B.C. on f1 and f2 .
We focus on the Dirichlet problem, where fh (x) is specified at x = a and x = b, with a < b. But we do
impose homogeneous B.C. on the Dirichlet Green function GD : GD (a, x′ ) = 0 (a < x′ ) immediately leads to:
b2 (x′ ) = − b1 (x′ )f1 (a)/f2 (a), whereas GD (b, x′ ) = 0 (x′ < b) gives:
f2 (a) f2 (b) f1 x′ ) − f1 (b) f2 (x′ ) f1 (a) f1 (b) f2 (x′ ) − f2 (b) f1 (x′ )
b1 (x′ ) = =⇒ b2 (x′ ) =
[α W ](x′ ) f1 (a) f2 (b) − f1 (b) f2 (a) [α W ](x′ ) f1 (a) f2 (b) − f1 (b) f2 (a)
The resulting Dirichlet Green function factorises (EXERCISE) in x and x′ :

′ 1 f2 (b) f1 (x> ) − f1 (b) f2 (x> ) f2 (a) f1 (x< ) − f1 (a) f2 (x< )
GD (x, x ) = (4.16)
[α W ](x′ ) f1 (a) f2 (b) − f1 (b) f2 (a)
where x> := max(x, x′ ) and x< := min(x, x′ ). Linear independence of f1 and f2 guarantees the non-
vanishing of W , but unlike an IVP, GD exists only if f1 (a)f2 (b) − f1 (b)f2 (a) 6= 0.
The most simple case occurs when f1 (a) = f2 (b) = 0; then f1 (b) and f2 (a) drop out, leaving: GD (x, x′ ) =
f1 (x< )f2 (x> )/α(x′ )W (x′ ).
We can now write down the general solution to the Dirichlet problem:
Z b
f2 (b) f (a) − f2 (a) f (b) f1 (a) f (b) − f1 (b) f (a)
f (x) = f1 (x) + f2 (x) + GD (x, x′ ) F (x′ ) dx′
f1 (a) f2 (b) − f1 (b) f2 (a) f1 (a) f2 (b) − f1 (b) f2 (a) a
(4.17)
If the homogeneous B.C. allow the existence of a non-zero homogeneous solution, ie., an eigenfunction of L
with eigenvalue zero satisfying these same conditions, eq. (4.12) forbids a Green function. 88
Example 4.1. A Helmholtz operator

Consider the operator d2t + ω02 with initial conditions on f and f˙ at a single point (IVP). We choose
the linearly independent f1 = sin ω0 t and f2 = cos ω0 t. Also, noting that α = 1 and W = −ω0 , eq.
(4.14) yields the IVP Green function:
′

′ ′ ′ sin ω0 (t − t )
Givp (t, t ) = Givp (t − t ) = θ(t − t )
ω0
Note the dependence of the IVP Green function on the difference t − t′ . Indeed, it can be shown
(EXERCISE) that for the second-order linear differential equation: [Lf ](t) = F (t) with constant
coefficients, Green functions for a one-dim IVP must satisfy G(t, t′ ) = G(t − t′ ), just by using the
general form of the homogeneous solutions: f± (t) = eλ± t . This is a manifestation of the invariance
of the differential operator with constant coefficients under translations of the variable t (eg. time).
By contrast, for the same L, f1 and f2 , (with ω0 = k), but with a Dirichlet problem at a = 0 and b,
we immediately obtain from eq. (4.16):
1
GD (x, x′ ) = sin k(x> − b) sin kx< (4.18)
k sin kb
and, provided kb 6= nπ, the unique inhomogeneous part of the solution to (d2x + k2 )f (x) = F (x) is:
Z x Z b
sin k(x − b) ′ ′ ′ sin kx
finh (x) = sin(kx ) F (x ) dx + sin k(x′ − b) F (x′ ) dx′
k sin kb a k sin kb x
If kb = nπ (n ∈ Z), ie., if b is an integer multiple of the half-period, the condition for the existence
of a Dirichlet Green function, f1 (a)f2 (b) − f1 (b) f2 (a) = − sin kb 6= 0, is violated.
Note that the same result would have followed from the initial choice f1 (0) = f2 (b) = 0, where
f1 = sin kx and f2 = sin k(x − b), with now W = −k sin kb.If k = nπ/b for some integer n 6= 0,
W = 0, so that f1 and f2 are linearly dependent. φ0 (x) = sin nπ(b − x)/b satisfies homogeneous
B.C. (at x = 0 and b) and solves the homogeneous equation. Thus, L = d2x + (nπ/b)2 is not
invertible and the standard Green-function approach fails. As discussed in Appendix L, a modified
Green function could still be constructed if φ0 (x)F (x) integrates to zero over the interval, but the
complete solution would not be unique unless an extra normalisation condition is imposed on the
homogeneous solutions.
Example 4.2. Eq. (4.13) has no explicit dependence on the coefficient of the first–order derivative
in L. This reflects the option we know we have to eliminate it from a second-order equation. For
instance, invoking eq. (4.8) with constant coefficients transforms the homogeneous equation for a
damped harmonic oscillator, (d2t + 2γdt + ω02 )f (t) = 0, into d2t g(t) + (ω02 − γ 2 )g(t) = 0, with
f (t) = g(t)e−γt .Inserting
p a solution
of the form eλtp
, we find the independent homogeneous
p solutions:
−γt ′
f1 (t) = e sin ω0 − γ t , f2 (t) = e−γt cos
2 2 ω02 − γ 2 t . Now W = − ω02 − γ 2 e−2γt , and
a straightforward substitution into eq. (4.14) for an IVP gives:
p 2
′ ′ −γ(t−t )′ sin ω0 − γ 2 (t − t′ )
G(t, t ) = θ(t − t ) e p (4.19)
ω02 − γ 2
89
Example 4.3. While we are talking about the damped harmonic oscillator, let us use it to illustrate
another way to solve differential equations that combines Fourier and Green techniques. The idea is
to write the equation:
f¨(t) + 2γ f˙(t) + ω02 f (t) = F (t)
in the frequency domain, assuming that the driving force dies at t → ±∞ or, alternatively, is turned
on at, say, t = 0, and then off at some later time. In this case the Fourier transform of F (t) exists and,
writing the Fourier representation of a function and of its drivative:
Z Z
1 iωt
iω
dt f (t) = √ dt f (ω) e dω = √ f (ω) eiωt dω
2π 2π
it is easy to see that our differential equation becomes:
Z h i
1
√ f (ω) − ω 2 + 2i γ ω + ω02 − F (ω) eiωt dω = 0
2π
Then, because the Fourier transform of the zero function vanishes everywhere, the differential equa-
tion is turned into the algebraic equation:
F (ω)
f (ω) =
−ω 2 + 2i γ ω + ω02
To go back to the time domain, we just write a solution to the inhomogeneous equation:
Z ∞ Z " Z #
1 1 eiω(t−t′ ) dω
f (t) = √ f (ω) eiωt dω = F (t′ ) dt′
2π −∞ 2π −ω 2 + 2i γ ω + ω02
Z ∞
= G(t, t′ ) F (t′ ) dt′
−∞
where: Z ′
∞
′ ′ 1 eiω(t−t )
G(t, t ) = G(t − t ) = − dω
2π −∞ (ω − ω+ )(ω − ω− )
p
with ω± = ± ω02 − γ2 + iγ.
To calculate G for t > t′ , we use contour integration in the complex ω plane, with the contour C
chosen to be counterclockwise around the upper infinite half-plane. Both poles ω = ω± lie in the
upper half-plane. Breaking up the contour into the real axis plus the semi-circle at infinity, we have:
Z ∞ ′ I ′ Z ′
1 eiω(t−t ) 1 eiω(t−t ) 1 eiω(t−t )
− dω = − dω + dω
2π −∞ (ω − ω+ )(ω − ω− ) 2π C (ω − ω+ )(ω − ω− ) 2π |ω|→∞ (ω − ω+ )(ω − ω− )
With t − t′ > 0, the numerator in the second integral on the right goes to zero as |ω| → ∞, and the
integral vanishes. The contour integral is evaluated with the Residue theorem:
Z ∞ ′ ′ ′
!
′ 1 eiω(t−t ) −1 eiω+ (t−t ) eiω− (t−t )
G(t − t ) = − dω = 2πi −
2π −∞ (ω − ω+ )(ω − ω− ) 2π ω+ − ω− ω+ − ω−
p 2
′ sin ω − γ 2 (t − t′ )
= e−γ(t−t ) p0
ω02 − γ 2
When t−t′ < 0, we must use a contour enclosing the lower infinite half-plane. But the integrand in the
contour integral is analytic in this region, and the integral vanishes by the Cauchy-Goursat theorem.
Thus, G(t, t′ ) = 0 for t < t′ , and we have recovered the result obtained in eq. (4.19) for an IVP. Here,
however, no knowledge of the homogeneous solutions was needed to find the Green function! As for
a BVP, if we can find a particular solution to eq. (4.11), we can enforce, eg., GD = 0 at the end-points
for a Dirichlet problem, by adding a suitable term G̃ that satisfies L[G̃) = 0.
90
4.3.1 Green’s second 1-dim identity and general solution of a BVP in terms of Green functions
Assume that a second-order linear operator Lx = αd2x + βdx + γ has been put in self-adjoint form, that is, β = α′ ,
with α′ = dx α. Then a few manipulations (EXERCISE) lead to Lagrange’s identity:
′
v Lx [u] − u Lx [v] = (α (v u′ − u v ′ )
where u, v ∈ D of L. Integrate over an interval [a, b] to obtain Green’s second identity in one dimension:
Z b
x=b
v Lx [u] − u Lx [v] dx = α (v u′ − u v ′ ) (4.20)
a x=a
Thanks to this identity, the homogeneous part, fh , of the solution to a BVP for L[f ] = F can be expressed in terms
of the same Green function that appears in the inhomogeneous solution, and for any B.C., homogeneous or not.
Indeed, suppose that u = G(x, x′ ) and that v = f (x) is the general solution to the inhomogeneous equation.
Then one easily shows from Green’s identity that for x′ ∈ [a, b]:
Z b
′
x=b
f (x ) = G(x, x′ ) F (x) dx − α G(x, x′ ) ∂x f − f ∂x G(x, x′ ) (4.21)
a x=a
where G(x, x′ ) is a Green function for Lx . We are already familiar with the first (inhomogeneous) term, but the
second one warrants careful examination. Obviously, it must be related to the homogeneous solution. But wait—is
f (x′ ) actually the general solution? Not yet! It is still just an identity. The second term is evaluated at the end-
points of the interval, so it depends on the boundary conditions for f . We cannot freely specify f and f ′ at both
a and b as this would be in general inconsistent. If f is specified at the end-points, then we must first find the
solution for f in order to know what its derivatives are at the end-points.
For a Dirichlet problem, however, we know that GD = 0 at the end-points. After interchanging x and x′ , and
using the symmetry, proved below, of GD in its arguments, there comes the general solution:
Z b h ix′ =b
f (x) = GD (x, x′ ) F (x′ ) dx′ + α f ∂x′ GD ′ GD (x, a) = GD (x, b) = 0 (4.22)
a x =a
Compare this form of the general solution, which explicitly depends only on F (x) and GD , plus f (a) and f (b), to
the solution (4.17) in terms of the linearly independent homogeneous solutions. It is a very instructive EXERCISE
to show their equivalence. We also see that if f happens to obey homogeneous B.C., f (a) = f (b) = 0, there is no
homogeneous part, which we saw in section 4.1.2 was essential to the existence of GD .
One important property of Dirichlet Green functions may be derived by letting v = GD (x′′ , x′ ) and u =
GD (x′′ , x) in Green’s second 1-dim identity (4.20), which holds
for differential operators of the form Lx′′ =
dx′′ (αdx′′ ) + γ. Because GD = 0 at the end-points and Lx′′ G (x , y) = δ(x′′ − y), we immediately find that
′′
GD for such operators is symmetric in its arguments:

GD (x, x′ ) = GD (x′ , x) (4.23)
This symmetry can provide a useful check on calculations.
The Neumann problem, with dx f specified at a and b, is not so straightforward. Just setting ∂GN to zero at
the end-points may be inconsistent, as one can see, eg., for L = d2x , by integrating ∂x2 G = δ(x − x′ ) once to get
b
∂x G a = 1. Instead, since L is assumed to be in self-adjoint form, one can introduce a modified Green function G
with the defining equation Lx G = δ(x − x′ ) − φ0 (x)φ0 (x′ ), where φ0 is a non-zero solution of the homogeneous
equation: [Lf ](x) = 0, with dx φ0 = 0 at the end-points (see Appendix L for the details). In example L.1, the
modified Neumann Green function for L = d2x that satisfies consistent homogeneous small B.C. is found to be:
GN (x, x′ ) = −x2 /2L + ax/L + θ(x − x′ )(x − x′ ) + b2 (x′ ). Insert this into eq. (4.21), and interchange x and x′ to
obtain f (x); then choose b2 (x) = −x2 /2L + ax/L, which makes GN (x, x′ ) symmetric, but does not affect f (x)
(why?). Implementing the symmetry, there comes the (unique up to a constant C) Neumann solution:
Z b h ix′ =b
f (x) = C + GN (x, x′ ) F (x′ ) dx′ − GN dx′ f ′ ∂GN (x, a) = ∂GN (x, b) = 0 (4.24)
a x =a
91
Problems in More than One Dimension (BF 7.4)

In one dimension, Green’s function for a second-order linear differential operator L always exists and is unique
for an IVP. If it exists for a BVP (no zero eigenvalue for L), it is unique. This is closely related to the fact that
boundary conditions are specified at one or two points only. In two or more dimensions, the boundaries contain an
infinite number of points, and Green functions are no longer guaranteed to exist, even for an IVP, But they do exist
in important cases of physical interest.
4.4 Differential Equations with Partial Derivatives (PDE)

Unless you are working on superstrings, it is usually sufficient to study PDEs in no more than four dimensions† .
In accordance with modern usage, we shall use Greek indices in four-dimensional (three spatial and one time)
problems, and roman indices in three spatial dimensions. We also implement the Einstein summation convention
according to which repeated indices in factors are to be summed over; in any such pair, we will try to write one
index as a superscript and one as a subscript so as to spot them more easily. Then the form of a second-order linear
differential operator that we shall use is:
L = αµ (x) ∂µ2 + β ν (x) ∂ν + γ(x) (4.25)
where x is the generalised position and it should be emphasised that Cartesian coordinates are implied. The
coefficients are assumed to be continuous in x.
We follow Hadamard (1923) and classify L according to the coefficients of the second-order derivatives:
Definition 4.1. • If at least one of the αµ vanishes at some point, the operator (and corresponding
homogeneous PDE will be said to be parabolic at that point (eg. heat equation, Schrödinger
equation, in which there is no second-order time-derivative).
• If the coefficients αµ are not zero but one of them has a sign different from all others at some
point, we say that L is hyperbolic at that point (eg., in Minkowski spacetime, the wave equation).
• If all αµ coefficients, all non-zero, have the same sign at some point (as in a Euclidean space), L
is elliptic at that point (eg. Laplace and Helmholtz operators — static 3-dim problems).
4.5 Separation of Variables in Elliptic Problems

Since the Laplacian operator occurs in all elliptic problems, it is worth taking a closer look at it. Our first task is to
separate it into two convenient parts; at the same time this will get us acquainted with a very powerful technique.
4.5.1 An Important and Useful 3-dim Differential Operator

To do this, we introduce the self-adjoint vector operators −i∇ and L = −i x × ∇, or Li = −iǫijk xj ∂ k , where
ǫijk is the completely antisymmetric Levi-Civita symbol, and summation over repeated indices is implied. With
the identity: ǫijk ǫimn = δj m δk n − δj n δk m , the scalar product of L with itself is, in Cartesian coordinates:
L · L = − ǫijk ǫimn xj ∂ k (xm ∂n )

= − xj ∂j + xj ∂ k ∂k − 3∂j − xk ∂ k ∂j = − xj xj ∂ k ∂k + 2xj ∂j − xj ∂j + xj ∂j (xk ∂k )
Extracting the Laplacian and reverting to coordinate-free notation, there comes:
L2 1 L2 1
∇2 = − + ∂r + ∂r (r ∂r ) = − + 2 ∂r (r 2 ∂r ) (4.26)
r2 r r2 r
The distance r to the origin can be expressed in any coordinates we wish, yet this expression obviously wants to
single out the direction along x = r n̂ from the other two. Also, it would be nice if L only involved derivatives
in directions perpendicular to n̂. This is most easily realised in a spherical coordinate system, since its radial
coordinate naturally corresponds to the direction along x; the other two coordinates are angular.
†
Anyway, it is straightforward to generalise our discussion to any number of spatial dimensions plus one time dimension.
92
By transforming the Cartesian components of L to spherical coordinates (r, θ, φ), we obtain (the calculation
is rather tedious, but Maple/Mathematica will readily do it for us):
Lx = − i (y∂z − z∂y ) = − i (− sin φ ∂θ − cot θ cos φ ∂φ )

Ly = − i (z∂x − x∂z ) = − i (cos φ ∂θ − cot θ sin φ ∂φ )
Lz = − i (x∂y − y∂x ) = − i ∂φ
The derivatives with respect to r have cancelled out! We also find that:

2 1 1 2
L = − ∂θ (sin θ ∂θ ) + ∂ (4.27)
sin θ sin2 θ φ
So L2 depends only on the angular coordinates. Eq. (4.26) makes it obvious that the commutator [∇2 , L2 ] = 0.
Now one readily shows (see section 3.2.4) that the following important relations hold:
[Lx , Ly ] = i Lz , [Ly , Lz ] = i Lx , [Lz , Lx ] = i Ly (4.28)
By symmetry, we have immediately that [L2 , L] = 0.
4.5.2 Eigenvalues of L2 and Lz

The operators L and L2 were studied earlier in section 3.6.2. There we found that the eigenvalues of L2 are l(l+1),
where l is a positive integer. We also found that the eigenvalues m ∈ Z of Lz take 2l + 1 values, from −l to l.
Since [L2 , L] = 0, L2 and Lz have a common set of eigenfunctions flm . The action of the raising and lowering
operators L± = Lx ± i Ly on these eigenfunctions is given (up to normalisation) by eq. 3.31:
p
L± flm = l(l + 1) − m(m ± 1) fl,m±1 (4.29)
4.5.3 Eigenfunctions of L2 and Lz

The eigenfunctions of Lz = −i ∂φ are readily obtained by solving the differential equation:
−i ∂φ f (θ, φ) = m f (θ, φ). With a separation ansatz: f (θ, φ) = F (θ)H(φ), the solution for H is:
H(φ) = eimφ (4.30)
Now we require that H (and f ) be single-valued, that is, H(φ + 2π) = H(φ). Thus:
eim(φ+2π) = eimφ =⇒ e2imπ = cos 2mπ + i sin 2mπ = 1
which constrains m to be any integer. Therefore, l := mmax must also be an integer. This is what rules out
the possibility of half-integer eigenvalues allowed for a self-adjoint operator J that just satisfies the canonical
commutation relations: [Ji , Jj ] = iǫijk J k .
The θ dependence of the eigenfunctions must be derived from the eigenvalue equation for L2 . Call f (θ, φ) =
m
Yl (θ, φ) = F (θ)H(φ); these must satisfy:

2 m 1 1
L Yl (θ, φ) = − ∂θ (sin θ ∂θ ) + 2
∂ Y m (θ, φ) = l(l + 1) Ylm (θ, φ)
sin θ sin2 θ φ l
as well as Lz Ylm (θ, φ) = mYlm (θ, φ), ie., Lz H(φ) = mH(φ). Then Ylm (θ, φ) = F (θ)ei mφ , and:

1 m2
− dθ (sin θ dθ ) − F (θ) = l(l + 1) F (θ)
sin θ sin2 θ
93
Instead of solving this equation by brute force, we use a clever technique involving the ladder operators L± :

L± = ± eiφ ∂θ ± i cot θ ∂φ
Now, when m = l, we have:

L+ Yll = eiφ ∂θ + i cot θ ∂φ Yll (θ, φ) = 0
Inserting Yll = F (θ)eilφ , this reduces to the much simpler:
dθ F (θ) − l cot θ F (θ) = 0
whose solution is F (θ) = (sin θ)l . Therefore, Yll = (sin θ)l eilφ . Applying L− the requisite number of times
generates the other Ylm (0 < m < l): Ylm ∝ Ll−m l
− Yl . When normalised, these are the spherical harmonics:
s
(−1)m 2l + 1 (l − m)!
Ylm (θ, φ) = (1 − x2 )m/2 dl+m
x (x2 − 1)l eimφ x = cos θ (4.31)
2l l! 4π (l + m)!
4.5.4 General Solution of a Spherically-Symmetric, 2nd-order, Homogeneous, Linear Equation

Suppose we are presented with the equation ∇2 + γ(x) Ψ(x) = 0. Work in spherical coordinates, and make the
ansatz: Ψ(x) = R(r)F (θ, φ). Using the form for ∇2 derived earlier, eq. (4.26), we have:
L2 Ψ 1
∇2 Ψ + γ(x)Ψ = − 2
+ ∂r Ψ + ∂r (r ∂r Ψ) + γ(x) Ψ
r r
L2 F (θ, φ) F (θ, φ)
= − R(r) 2
+ dr R(r) + dr (r dr R(r)) + γ(x) R(r)F (θ, φ)
r r
Multiplying the second line by r 2 /(R(r)F (θ, φ)), we see that the equation is separable provided γ(x) = γ(r):
R(r)
L2 F (θ, φ) = λ F (θ, φ) dr R(r) + dr (r dr R(r)) + r γ(r) R(r) = λ
r
The first equation is the eigenvalue equation for L2 , whose eigenvalues are λ = l(l + 1) (l ≥ 0 ∈ Z), with the
spherical harmonics Ylm (θ, φ) as eigenfunctions. The radial equation can thus be written:

1 2
l(l + 1)
dr r dr Rl (r) + γ(r) − Rl (r) = 0
r2 r2
When γ(r) = 0, this is the radial part of the Laplace equation which becomes, after the change of variable
r = ex : d2x R+dx R−l(l+1)R = 0. Inserting a solution of the form epx turns the equation into p2 +p−l(l+1) = 0,
that is, p = l or p = −(l + 1), which leads to R = Aelx + Be−(l+1)x = Ar l + Br −(l+1) . Therefore, the general
solution to the Laplace equation in spherical coordinates is:
X∞ X l
Blm
Ψ(r, θ, φ) = Alm r + l+1 Ylm (θ, φ)
l
(4.32)
r
l=0 m=−l
The coefficients Alm and Blm are determined from boundary or matching conditions. In regions either containing
the origin, or extending all the way to infinity, Blm = 0 or Alm = 0, respectively. Clearly, if this solution is to
be regular, and if it holds everywhere, it must vanish. In other words, if the Laplace equation is valid everywhere,
it has no non-vanishing regular solution. For a non-trivial solution, there must be a region of space where there
exists an inhomogeneous term acting as a source.
Note, however, that the general solution holds at any point where there is no source. The effect of sources is
encoded in the coefficients Alm and Blm .
94
When γ(r) = k2 > 0, we get the radial part of the Helmholtz equation in spherical coordinates:

2 l(l + 1)
d2r Rl (r) + dr Rl (r) + k2 − Rl (r) = 0
r r2
Defining dimensionless x = kr readily transforms it into a form of the Bessel equation whose solutions are the
spherical Bessel functions of the first and second (Neumann) kind, usually written as (see also Jackson’s Classical
Electrodynamics, section 9.6):

l  xl x ≪ (1, l)
1 d sin x
jl (x) = (−x)l ∼ (4.33)
x dx x  1 sin(x − lπ/2) x ≫ l
x

 1
l 
− l+1 x ≪ (1, l)
l 1 d cos x x
nl (x) = − (−x) ∼ (4.34)
x dx x 
 1
− cos(x − lπ/2) x ≫ 1
x
The general solution of the Helmholtz equation is a linear combination of the jl and nl .
Here are a few spherical Bessel and Neumann functions as plotted on Maple, with ρ = x:
The nl diverge at the origin and thus are excluded from any solution regular at the origin.
(1,2)
Spherical Bessel functions hl (x) = jl (x) ± i nl (x), aka Hankel functions of the first and second kind, can
(1,2)
come in handy. One can express the general solution of the Helmholtz equation in terms of the hl .
95
4.6 Second 3-dim Green Identity, or Green’s Theorem

Before discussing the all-important subject of boundary conditions, we derive a result that will
prove very useful
in the study of 3-dim elliptic problems. We assume the self-adjoint form: L[f ] = ∂ i α(x)∂i f + γ(x)f .
Write the divergence theorem for ∇ · (αf ∇g) over a connected volume, and expand the divergence to get:
Z h i I I
3
f ∇ · (α ∇g) + α ∇f · ∇g d x = α f ∇g · dS = α f ∇g · n̂ dS (4.35)
V ∂V ∂V
where ∂V is the closed boundary of the volume V of integration, and the unit vector n̂ normal to ∂V , by con-
vention, always points outward from the volume. This is Green’s first identity in three dimensions; when α is a
constant, and introducing the normal derivative ∂n = n̂ · ∇, it reduces to the more familiar form:
Z h i I I
f ∇2 g + ∇f · ∇g d3 x = f ∇g · dS = f ∂n g dS (4.36)
V ∂V ∂V
Interchanging f and g in the first identity (4.35) and subtracting, adding and subtracting γf g in the volume
integral yields the second Green identity in three dimensions—compare with one-dim eq. (4.20):
Z I I
3
f L[g] − g L[f ] d x = α f ∇g − g ∇f · dS = α f ∂n g − g ∂n f dS (4.37)
V ∂V ∂V
With α a constant, this becomes the well-known Green theorem:

Z I
2 2
3
f ∇ g − g∇ f d x = f ∇g − g ∇f · dS (4.38)
V ∂V
Example 4.4. Uniqueness and existence of solutions for the Poisson equation with B.C.
The Poisson (inhomogeneous Laplace) equation is of the form ∇2 Ψ(x) = F (x). We also specify
B.C. for either Ψ or ∂n Ψ on ∂V . With f = g = Ψ3 and α constant, eq. (4.36) becomes:
Z I
2 2 3
[Ψ3 ∇ Ψ3 + (∇Ψ3 ) ] d x = Ψ3 ∂n Ψ3 dS
V ∂V
Suppose there exist two solutions, Ψ1 and Ψ2 , of ∇2 Ψ(x) = F (x) that satisfy the same conditions
on the surface. Define Ψ3 := Ψ2 − Ψ1 . Then ∇2 Ψ3 = 0 inside the R volume. The surface integral
2 3
is zero because either Ψ3 = 0 or ∂n Ψ3 = 0 on the surface; and (∇Ψ3 ) d x = 0 everywhere.
Also, Ψ3 being twice differentiable at all points in the volume, ∇Ψ3 is continuous and therefore zero
everywhere inside, so that Ψ3 is a constant. It follows immediately that if Ψ3 = 0 on ∂V , Ψ1 = Ψ2
everywhere. On the other hand, when ∂Ψ3 /∂n = 0 on ∂V , Ψ3 can be a non-zero constant inside.
We conclude that Ψ1 = Ψ2 inside the volume (up to a possible additive constant), and that the solution,
if it exists, is uniquely determined. The importance of this result cannot be overstated: any function
that satisfies the inhomogeneous Laplace equation and the B.C. is the solution, no matter how it was
found! Moreover, we see that we cannot arbitrarily specify both Ψ and ∂Ψ/∂n on the boundary since
one suffices to determine the solution.
The B.C. determine the solution, but only if it exists. Further conditions must be met for this to happen.
Indeed, integrate ∇2 Ψ(x) = F (x) over (connected!) V ; the divergence theorem yields the condition:
Z Z
F (x) d3 x = ∂n Ψ(x) dS (4.39)
V ∂V
Another condition for the existence of a solution is that the enclosing boundary be “reasonably”
smooth (eg. no spikes . . . ) if we wish to specify ∂n Ψ on ∂V .
Finally, if ∇2 φn = λn φn , and taking f = φ∗n and g = φn in eq. (4.36), one shows (EXERCISE) that
the eigenvalues of the Laplacian are always negative.
96
4.7 3-dim Boundary Value (Elliptic) Problems with Green Functions

Consider Lx in self-adjoint form. Introduce Green functions that satisfy [Lx G](x, x′ ) = δ(x − x′ ) (some authors
multiply the right-hand side by ±4π) in regions with closed boundaries. If we are a little careful, we will find that
for some B.C. this kind of problem can admit unique Green functions. Just as in one dimension, this requires that
there exist no non-trivial solution to Lx f (x) = 0 with homogeneous B.C.
4.7.1 Dirichlet and Neumann Boundary Conditions for an Elliptic Problem

Suppose that Ψ(x) satisfies† Lx Ψ(x) = F (x). Proceeding exactly like in the 1-dim case, take f = Ψ and g = G
in Green’s second identity (eq. (4.37)):
Z I
3
Ψ Lx [G] − G Lx [Ψ] d x = α (Ψ ∂n G − G ∂n Ψ) dS
V ∂V
We obtain: Z I

Ψ(x) δ(x − x′ ) − F (x) G(x, x′ ) d3 x = α (Ψ ∂n G − G ∂n Ψ) dS
V ∂V
With x′ inside the volume, re-arranging then yields:
Z I

Ψ(x′ ) = F (x) G(x, x′ ) d3 x + α Ψ ∂n G − G ∂n Ψ dS (4.40)
V ∂V
where the normal derivatives in the integrand on the right-hand side are to be evaluated on ∂V , the boundary of
the arbitrary volume. This expression for Ψ cannot be considered a solution yet; it is still “just” an identity.
Again, note that Ψ and ∂Ψ/∂n are in general not independent on the boundary. We are not free to specify
them both arbitrarily at any point on ∂V as such values will in general be inconsistent.
As before, specifying Ψ on ∂V gives Dirichlet B.C., whereas specifying ∂n Ψ gives Neumann B.C.
For a Dirichlet problem we demand that GD (x, x′ ) = 0 ∀ x ∈ ∂V . With Green’s 2nd identity, it is then quite
easy to prove (EXERCISE) that GD (x, x′ ) is symmetric in its arguments. After interchanging x and x′ in eq.
(4.40) and implementing the symmetry of GD , the solution for Ψ is:
Z I
Ψ(x) = F (x′ ) GD (x, x′ ) d3 x′ + α Ψ(x′ ) ∂n′ GD (x, x′ ) dS ′ (4.41)
V ∂V
The Dirichlet solution is uniquely determined by the B.C. on Ψ via GD . Note that the total surface ∂V enclosing
the volume may be disjoint, as for instance with the volume between two concentric spheres.
If we have managed to find GD for a particular type of boundary, the source-free solution (F (x′ ) = 0) is just
the surface integral; but if it happens that Ψ = 0 on ∂V , only the volume integral contributes. Many boundary-
value problems in electrostatics, for which the B.C. are reasonably simple, can be solved this way.
Similar considerations apply to Neumann B.C., ie. when ∂Ψ/∂n rather than Ψ is known on the boundary. But
we must be a little careful about the B.C. on ∂n GN : we cannot always put this equal to 0 in eq. (4.40). Indeed, take
for instance L = ∇2 ; then, from the divergence theorem and the defining equation L[GN ] = δ(x − x′ ):
Z I
∇ · ∇GN d3 x = ∂n GN dS = 1
∂V
A consistent B.C. is ∂n GN ∂V
= 1/S, and the (non-unique) solution to the Neumann problem for L = ∇2 reads:
Z I
Ψ(x) = < Ψ >∂V + F (x′ ) GN (x, x′ ) d3 x′ − GN (x, x′ ) ∂n′ Ψ(x′ ) dS ′ (4.42)
V ∂V
unique up to the a priori unknown average of Ψ over the surface. Often (but not always!) the volume is bounded
by two surfaces, one closed and finite and the other at infinity, in which case ∂n GN can be set to zero on the entire
boundary, and < Ψ >∂V vanishes, still leaving an arbitrary additive constant in the solution. Also, GN (x, x′ )
itself is determined only up to an additive function h(x), with Lx h = 0. When L = ∇2 , however, condition (4.39)
prevents h from contributing to the solution, and h can be used to make GN symmetric in x and x′ .
†
Although we call it “inhomogeneous”, nothing in what we will do here prevents F (x) from depending on Ψ(x).
97
4.7.2 Green function for the 3-d Elliptic Helmholtz operator without boundary conditions
We proceed to find a Green function for the operator ∇2 + λ, with λ a constant. The Fourier transform of
(∇2 + λ) Ψ(x) = F (x) is (−k2 + λ)ψ(k) = F (k). We must distinguish between two possibilities:
1. λ = −κ2 ≤ 0, κ ≥ 0
Then, similarly to what happens in one dimension (example 4.3), an “inhomogeneous” solution is:
Z ZZ ′
1 F (k) ik·x 3 1 3 ′ −ik·x′ F (x )
Ψ(x) = − e d k = − d x e eik·x d3 k
(2π)3/2 k2 + κ2 (2π)3 k2 + κ2
R
Compare with the Green-function form of the inhomogeneous solution, V F (x′ )G(x, x′ ) d3 x′ (EXER-
CISE):
Z ik·(x−x′ ) Z ∞ ′
1 e i k eik|x−x |
G(x, x′ ) = − d 3
k = dk
(2π)3 k2 + κ2 (2π)2 |x − x′ | −∞ k2 + κ2
This integral is easily evaluated as part of a contour integral around a semi-circle at infinity in the upper
complex k half-plane. As in the one-dimension example, the contribution at infinity vanishes, and the
′
residue due to the pole at k = iκ is e−κ|x−x | /2. The Residue theorem then yields the (sometimes called
fundamental, or singular) solution:
′
′ 1 e−κ|x−x |
G(x, x ) = − (4.43)
4π |x − x′ |
For λ = 0 (κ = 0), we obtain a Green function for the Laplacian operator.

With κ = 0 and F (x) = −4πρ(x) (Gaussian units!), for instance, an inhomogeneous solution is the
generalised Coulomb Law for the electrostatic potential of a localised charge density ρ(x), or one that
vanishes at infinity faster than 1/|x − x′ |.
98
2. λ = κ2 ≥ 0
In order to invert the algebraic equation for ψ(k), we write λ = (q ± iǫ)2 (ǫ ≥ 0). Then we arrive at:
Z ′ ′
(±) ′ 1 eik·(x−x ) 3 1 e±iq|x−x |
Gq (x, x ) = − lim d k = − (4.44)
(2π)3 ǫ→0 k2 − (q ± iǫ)2 4π |x − x′ |
For details of the calculation, see pp. BF415–416.
Do check that these Green functions satisfy (∇2 + λ)G(x, x′ ) = δ(x − x′ ). But note that they are not the
general solution of this equation, since any function that satisfies the homogeneous equation can be added to them!
If the volume integral extends over all space, the surface integral in the Dirichlet solution for the case λ < 0
certainly vanishes at infinity for fairly weak conditions on Ψ(x), because of the exponential factor in Green’s
function. When λ ≥ 0, the surface integral also vanishes provided Ψ(x) → 0 faster than 1/|x − x′ |, (since
dS ∼ |x − x′ |2 ), and we are left with just the inhomogeneous integral:
Z ′
(±) 1 F (x′ ) e±iq|x−x | 3 ′
Ψq (x) = − d x (4.45)
4π V |x − x′ |
If, however, Ψ(x) does not vanish fast enough at infinity, it is more convenient to write it in terms of the
solution of the homogeneous equation (∇2 + q 2 )Ψ(x) = 0, plus the volume integral:
Z ′
(±) iq·x 1 F (x′ ) e±iq|x−x | 3 ′
Ψq (x) = A e − d x (4.46)
4π V |x − x′ |
Note that these expressions for Green’s functions assume no boundary surfaces (except at infinity)!
4.7.3 Dirichlet Green function for the Laplacian

R
When there are no boundary conditions for Ψ on finite surfaces, the volume integral F (x′ )G(x, x′ )d3 x′ can be
taken as the solution to L[Ψ] = F . For instance, in the case of a point-source located at y: F (x′ ) = −4πqδ(y−x′ ),
with q some constant, we see that Ψ(x) = −4πqG(x, y) = q/|x − y| in the case of L = ∇2 .
When there are finite boundaries, however, as in a Dirichlet problem, we know that we have to ensure that
GD (x, x′ ) = 0 when either x or x′ is a point on the surface that encloses the volume in which our solution is
valid. Obviously, with the Green function given in eq. (4.43), which vanishes only on a boundary at infinity, this
is impossible. It is time to exercise our freedom to add to G a function that satisfies the homogeneous equation
L[G] = 0 and contains free parameters that can be set so as to force the combined Green function to vanish on the
boundary. In the case of the Laplacian, we take:

′ 1 1 g
GD (x, x ) = − +
4π |x − x′ | |x − x′′ |
which means that if the second term is to satisfy the Laplace equation ∀ x inside the volume where we are looking
for a solution, x′′ must lie outside the volume containing x and x′ .
Example 4.5. Solution of the Dirichlet problem on a sphere for the Laplacian
Consider a sphere of radius a centered on the origin. We want: GD (an̂, x′ ) = GD (x, an̂′ ) = 0
Symmetry of GD dictates that x′′ and x′ be collinear, which means that, at |x| = r = a, we can write:
!
′ 1 1 g
GD (an̂, x ) = − + ′′ a
4π a n̂ − ra′ n̂′ r r′′ n̂ − n̂′
where rn̂ = x, etc. By inspection, we see that if GD (an̂, x′ ) is to vanish for n̂ in an arbitrary direction,
we must have: 1/a = −g/r ′′ and r ′ /a = a/r ′′ . Then:
g = − a/r ′ , r ′ r ′′ = a2 (4.47)
99
Thus, x′′ does lie outside the sphere if x′ is inside, and vice-versa. Replacing an̂ by rn̂ = x yields:
" #
1 1 1
GD (x, x′ ) = − −
4π |x − x′ | (r ′ /a)x − (a/r ′ )x′
  (4.48)
1  1 1 
= − p − q
4π 2
r 2 + r ′ − 2rr ′ cos γ 2 ′ 2 2 2 ′
r r /a + a − 2rr cos γ
The second form makes is most easy to see that not only GD (x, an̂′ ) = 0, but also GD (an̂, x′ ) = 0,
as desired. In spherical coordinates centered on the sphere, the angle γ between x amd x′ is, from
spherical trigonometry: cos γ = cos θ cos θ ′ + sin θ sin θ ′ cos(φ − φ′ ). The Dirichlet Green function
we have found is valid for any ball since it does not care about which particular B.C. is specified for
Ψ(x) on its spherical boundary.
When Ψ(r ′ = a) = 0, the surface integral in eq. (4.41) vanishes; the volume integral remains the
same since it is independent of the B.C. for Ψ. If Ψ(r ′ = a) 6= 0, we must evaluate ∂n′ GD on the
sphere. In spherical coordinates, this is:
∂GD ∂GD 1 a(a2 − r 2 )

= ± = ±
∂n′ ∂r ′ 4πa2 (r 2 + a2 − 2ar cos γ)3/2
∂V r ′ =a
depending on whether dS′ , the normal to the surface which always points out of the volume, is in the
direction of x′ or in the opposite direction. Then the general solution of the inhomogeneous Laplace
equation with B.C. specified on the surface r = a for Ψ is:
 
Z
1 1 1
Ψ(x) = F (x′ )  q − p  d 3 x′
4π 2 ′ 2 2 2 ′ r 2 + r ′ 2
− 2rr ′ cos γ
r r /a + a − 2rr cos γ
I
1 a2 − r 2
± Ψ(r ′ = a) dS ′ (4.49)
4π a (r 2 + a2 − 2ar cos γ)3/2
where the (+) sign refers to the solution for r < a and the (−) sign applies to r > a. In the latter
case, there is an implicit assumption that the integrand, Ψ∂n′ GD , of the surface integral vanishes at
infinity faster than 1/r ′2 . When F (x) = 0 everywhere inside the volume where the solution is valid,
we are left with the Laplace equation ∇2 Ψ = 0, with solution:
I
′ ′ 1 a(a2 − r 2 )
Ψ(x) = ± Ψ(a, θ , φ ) dS ′ (4.50)
4πa2 (r 2 + a2 − 2ar cos γ)3/2
Clearly also, if Ψ(a, θ ′ , φ′ ) 6= 0 and r > a, F (x) 6= 0 somewhere in the region r < a, and vice-versa.
One can also show that for a ball of radius a and surface Ωn−1 in Rn :

 1
ln |x − x′ | − ln r ′ x − a x′ (n = 2)
 2π a r′
′
GD (x, x ) =
 1 1 1
− − (n > 2)
(n − 2) Ωn−1 |x − x′ |n−2 |(r ′ /a)x − (a/r ′ )x′ |n−2
(4.51)
which leads to a unified expression, valid for n ≥ 2, for the normal derivative of GD on the sphere:
1 a2 − r 2
∂n′ GD = ± an−2 (4.52)
r ′ =a Ωn−1 |x − x′ |n
r ′ =a
100
4.7.4 An important expansion for Green’s Functions in Spherical Coordinates

The angular dependence in the Green functions such as derived above is quite complicated and may well not yield
a solution in closed form when integrated, so it is often sensible to use a expansion appropriate to the coordinate
system selected for the problem. Indeed, let us do this for the Laplacian in spherical coordinates.
In spherical coordinates, Green functions for the Laplacian operator all satisfy:
∇2x G(x, x′ ) = δ(x − x′ )

∞ X
X l
1 ′ ∗
= δ(r − r ) Ylm (θ ′ , φ′ ) Ylm (θ, φ) (4.53)
r2
l=0 m=−l
where the completeness relation for spherical harmonics has been invoked:
∞ X
X l
∗
Ylm (θ ′ , φ′ )Ylm (θ, φ) = δ(x − x′ )δ(φ − φ′ ) (x = cos θ) (4.54)
l=0 m=−l
We shall look for an expansion over separable terms of the form:

∞ X
X l
′
G(x, x ) = gl (r, r ′ ) Ylm
∗
(θ ′ , φ′ ) Ylm (θ, φ)
l=0 m=−l
Inserting into eq. (4.53), we immediately find with eq. (4.26) that gl (r, r ′ ) must satisfy the radial equation:

r 2 ∇2r gl (r, r ′ ) = dr r 2 dr gl (r, r ′ ) − l(l + 1) gl (r, r ′ ) = δ(r − r ′ )
We now find ourselves in the familiar territory of 1-dim Green-function problems and self-adjoint operators. For
instance, we can connect with eq. (4.13) for a 1-dim Dirichlet Green function. We take two concentric spheres of
radius a and b, with b > a.
We have α(r ′ ) = r ′2 and, with f1 = r l and f2 = r −(l+1) , W (r ′ ) = −(2l + 1)/r ′2 . Also, let r< ≡ min(r, r ′ )
and r> ≡ max(r, r ′ ). It takes only a straightforward computation using eq. (4.16) to arrive at (EXERCISE):
∞ X
X l ! !
∗ (θ ′ , φ′ ) Y (θ, φ) l
Ylm lm a2l+1 r> 1
GD (x, x′ ) = l
r< − l+1 − (4.55)
(2l + 1) 1 − (a/b)2l+1 r< b2l+1 l+1
r>
l=0 m=−l
Inspection of the radial factors shows that this expression vanishes at r = a and r = b (and when r ′ = a or
r ′ = b), as it should. We did not have to require this since it is built in the derivation of the 1-dim Dirichlet Green
function. Two important cases:
∞ X
X l !
′ Y ∗ (θ ′ , φ′ )Y
lm lm (θ, φ) l
l
r> 1
GD (x, x ) = r< − (a = 0) (4.56)
(2l + 1) b2l+1 l+1
r>
l=0 m=−l
∞ X
X l !
Y ∗ (θ ′ , φ′ ) Y lm (θ, φ) 1 a2l+1
GD (x, x′ ) = lm
l+1 l+1
l
− r< (b → ∞) (4.57)
(2l + 1) r> r<
l=0 m=−l
The first expression gives the Green function inside a sphere of radius b; the second one, outside a sphere of radius
a and all the way to infinity. When there are no boundary surfaces, we obtain over all space:
∞ X
X l
1 l
r<
G(x, x′ ) = − Y ∗ (θ ′ , φ′ ) Ylm (θ, φ)
l+1 lm
(4.58)
2l + 1 r>
l=0 m=−l
101
This also yields a useful expansion of the ubiquitous distance factor 1/|x − x′ |.
When 0 ≤ r ≤ b (interior case) we can rewrite (EXERCISE) the surface integral in eq. (4.49) as:
∞ X
X l Z
′ ′ ∗ r l
Ψ(b, θ , φ ) Ylm (θ ′ , φ′ ) dΩ′ Ylm (θ, φ)
b
l=0 m=−l
where Ψ(b, θ ′ , φ′ ) is specified on the surface r = b. The normal derivative of the Green function on the surface,
∂G/∂n′ = ∂G/∂r ′ r′ =b , has been evaluated for r< = r and r> = r ′ since r < r ′ = b. Also, the surface element
on a sphere of radius b is dS ′ = b2 dΩ′ . This expression is still rather complicated, but it simplifies considerably
if Ψ(b, θ ′ , φ′ ) exhibits a symmetry (eg. azimuthal). Also, if one can write Ψ(b, θ ′ , φ′ ) as a linear combination of
spherical harmonics, the angular integration becomes trivial due to the orthonormality of the harmonics, and only
a few terms in the sums might contribute.
102
4.7.5 An Elliptic Problem with a Twist: the Time-independent Schrödinger Equation
The time-independent Schrödinger equation (TISE) for a potential V (x) takes the following suggestive form:
2m
(∇2 + λ) ψ(x) = V (x) ψ(x) (4.59)
~2
where λ = 2mE/~2 . Although the right-hand side is not inhomogeneous, our previous results still hold.
For bound states (E < 0) of an attractive potential, λ = −κ2 < 0, and we have the integral equation:
Z −κ|x−x′ |
m e
ψ(x) = − V (x′ ) ψ(x′ ) d3 x′
2π~2 |x − x′ |
A somewhat simpler integral expression may be derived from the convolution [V ∗ ψ](k):
Z
1
[V ∗ ψ](k) := V (k − k′ )ψ(k′ ) d3 k′
(2π)3/2
According to the convolution theorem, the Fourier transform of [V ∗ ψ](k) is just V (x)ψ(x) in eq. (4.59). Then
the Fourier representation of this equation can be written as:
Z Z Z
2 2 ik·x 3 2m 1
− (k + κ )ψ(k) e d k = 2 ′ ′ 3 ′
V (k − k )ψ(k ) d k eik·x d3 k
~ (2π)3/2
and there comes: Z
2m V (k − k′ )ψ(k′ ) 3 ′
ψ(k) = − d k
(2π)3/2 ~2 k2 + κ2
For unbound states (E > 0), κ2 > 0, and we can immediately write the Lippmann-Schwinger equation:
Z ±iq|x−x′ |
(±) A iq·x m e
ψq (x) = e − 2
V (x′ ) ψq(±) (x′ ) d3 x′ (4.60)
(2π) 3/2 2π~ |x − x′ |
p
with q = 2mE/~2 .
The asymptotic√form of the Lippmann-Schwinger equation is of particular interest. When r >> r ′ , we can
expand |x − x′ | = r 2 − 2x · x′ + r ′2 ≈ r − n̂ · x′ , with n̂ = x/r. Inserting into the integral equation yields:
Z
(±) A iq·x m e±iqr ′ (±)
ψq (x) = 3/2
e − 2
e∓iqn̂·x V (x′ ) ψq (x′ ) d3 x′
r→∞ (2π) 2π~ r

A iq·x e±iqr
= e + f± (q)
(2π)3/2 r
This expression represents the spatial dependence of a superposition of a plane wave and a scattered spherical
wave propagating inward or outward from the origin. The function f± (q) is called the scattering amplitude; it
also obeys an integral equation in q (momentum) space, eq. BF7.75, and its square modulus is directly related to
experimental data. See BF p. 414–420 for more details and an application to the Yukawa potential.
4.8 A Hyperbolic Problem: the d’Alembertian Operator

With the Fourier integral representation (note the different normalisation and sign in the exponentials!):
Z ∞
1
Ψ(x, t) = Ψ(x, ω) e−iωt dω
2π −∞
Z ∞ (4.61)
Ψ(x, ω) = Ψ(x, t) eiωt dt
−∞
103
we can transform a typical inhomogeneous wave equation:

1 2
Ψ(x, t) = ∇2 Ψ(x, t) − ∂ Ψ(x, t) = F (x, t)
c2 t
where F (x, t) is a known source, to its Helmholtz form:
(∇2 + k2 )Ψ(x, ω) = F (x, ω) k2 ≡ (ω/c)2 (4.62)
Just as for the Laplacian, there exist Green functions for ∇2 + k2 ; we have found them earlier in eq. (4.44):
1 e±ikR
G(±) (R) = − R ≡ |x − x′ | (4.63)
4π R
Now we are ready to derive the full Green functions for the d’Alembertian operator, which satisfy:
′ ′
x G(x, t; x , t ) = δ(x − x′ ) δ(t − t′ ) (4.64)
With the important representation of the delta-function:

Z ∞
1 ′
δ(t − t′ ) = e−i ω(t−t ) dω (4.65)
2π −∞
the defining equation becomes, in the frequency domain:

′
(∇2x + k2 )G(x, x′ , ω, t′ ) = δ(x − x′ ) eiωt
′ ′
Assume separable solutions of the form G(x, x′ )eiωt ; inserting yields: G± (x, x′ , ω, t′ ) = −ei(±kR+ωt ) /4πR,
after using (4.62). Then, transforming back to the time domain, and replacing k by ω/c, we arrive at the Green
functions:
Z ∞
(±) ′ ′ 1 iω[±R/c+(t′ −t)] 1 δ t′ − [t ∓ R/c]
G (x, t; x , t ) = − 2 e dω = − (4.66)
8π R −∞ 4π R
Thus, in non-elliptic problems, Green functions can contain δ-functions and so may not be actual functions!
Using eq. (4.64), we also recognise that:
Z Z ∞ Z Z ∞
3 ′ (±) ′ ′ ′ ′ ′ 3 ′
x d x G (x, t; x , t ) F (x , t ) dt = d x F (x′ , t′ ) x G(±) (x, t; x′ , t′ )dt′ = F (x, t)
all
space
−∞ −∞
has the generic form Ψ(x, t) = F (x, t), which shows that the general solution of a wave equation with sources
can be written either as the retarded or advanced solutions:
Z Z ∞
Ψ{ ret } (x, t) = Ψ{ in } (x, t) + G(±) (x, t; x′ , t′ ) F (x′ , t′ ) d3 x′ dt′
adv out −∞
Z
1 F (x′ , t ∓ R/c) 3 ′
= Ψ{ in } (x, t) − d x (4.67)
out 4π x − x′ { ret }
adv
where in the integral the position x′ must be evaluated at the retarded time t − R/c, or at the advanced time
t + R/c. This ensures the proper causal behaviour of the solutions, in the sense that, eg., the solution at time t only
depends on the behaviour of the source point x′ at time t − R/c. Ψin and Ψout are possible plane-wave solutions
of the homogeneous wave equation for Ψ. Often they can be dropped.
104
4.9 Initial Value Problem with Constraints

The Initial Value Problem (IVP) consists in finding which data must be specified at a given time for the time
evolution of variables to be uniquely determined by their equations of “motion”.
By initial data, one means the state of the set of variables and their first-order derivatives on a three-dimensional
spacelike hypersurface; usually, this means at some time t0 everywhere in space. The IVP together with the evolu-
tion equations constitute the Cauchy Problem of the theory. If the Cauchu problem can be solved, the dynamical
behaviour of the set of variables can be uniquely predicted from its initial data.
Most often, the equations of “motion” take the form of a set of equations of the form f = F . If they
always told the whole story, the Cauchy problem would be solved by specifying the value of f and its first-order
time derivative at t0 . When there are inherent, built-in constraints on the initial data, however, these constraint
equations must be discovered and solved. Also, we must find which initial data we are allowed to specify freely.
We study in some depth a very important example: Maxwell’s theory. In linear, unpolarised and unmagnetised
media, Maxwell’s equations are:
km
∇ · E = 4πke ρ ∇×B − ∂t E = 4πkm J
ke
(4.68)
∇·B = 0 ∇ × E + ∂t B = 0
where ke and km are constants that depend on the system of units, and ke /km = c2 , with c the speed of light. The
1
source terms ρ and J satisfy a continuity equation: ∂t ρ = 4πk e
∇ · ∂t E = − ∇ · J.
The two homogeneous equations are equivalent to:
E = − ∂t A − ∇Φ B = ∇×A (4.69)
If we perform the gauge transformations Φ → Φ − ∂t f and A → A + ∇f , where f (x, t) is an arbitrary real

differentiable function, neither E nor B change! We say that Maxwell’s theory is gauge-invariant.
The inhomogeneous Maxwell equations (4.68) become second-order equations for Φ and A:
∇2 Φ + ∂t (∇ · A) = −4π ke ρ
(4.70)
1
A − ∇ ∇ · A + 2 ∂t Φ = − 4π km J
c
.
4.9.1 Second-order Cauchy problem using transverse/longitudinal projections

While eq. (4.70) are gauge-invariant, A and Φ themselves are not, at least at first sight. What this means is that
the time-evolution of at least some of the four quantities Φ and A cannot be uniquely determined from their initial
conditions and eq. (4.70) since we can always perform an arbitrary gauge transformation on them at some arbitrary
later time t, as often as we wish. This serious issue must be addressed if Φ and A are to be of any use at all.
One instructive approach is to note that according to the Helmholtz theorem (see section 1.6.2) any differen-
tiable 3-dim vector field that goes to zero at infinity faster than 1/r may be written as the sum of two vectors:
∇u + ∇ × w
A = |{z}
| {z }
AL AT
AL = ∇u, whose curl vanishes identically, is the longitudinal part (or projection) of A; AT = ∇ × w, whose
divergence vanishes identically, is the transverse projection of A. This allows us to decompose Maxwell’s equa-
tions for the fields and the potential into longitudinal and tranverse parts, which are perpendicular to each other.
Project the second equation (4.70). The transverse projection immediately gives:
AT = − 4π km JT (4.71)
105
where we have used the fact a gradient is a longitudinal object. The two transverse components AT satisfy a
proper wave equation and correspond to physically observable quantities, in the sense that being transverse, they
are unaffected by A → A + ∇f , which can change only the longitudinal component AL . Therefore, the time
evolution of the two transverse AT is not arbitrary and they have a well-posed Cauchy problem.
Now, remembering that = ∇2 − (∂t2 )/c2 , take the divergence of the longitudinal projection of (4.70):

1 ∂t ∇2 Φ
∇· AL − ∇ ∇ · AL + 2 ∂t Φ + 4π km JL = (∇ · AL ) − ∇2 (∇ · AL ) − + 4π km ∇ · JL
c c2
1 h i
= − 2 ∂t ∂t (∇ · AL ) + ∇2 Φ + 4π ke ρ
c
where the continuity equation has been invoked in the second line. But the terms in the square bracket on that line
are just the first of equations (4.70). Therefore, the second Maxwell equation for the 3-vector potential contains
no information about ∇ · A that is not in the first equation. But that is really an equation for Φ, with ∇ · Ȧ (more
precisely, ∇ · ȦL ) as a source together with ρ. Therefore, Maxwell’s theory cannot uniquely determine the time
evolution of the divergence of the 3-vector potential. Nor can it uniquely determine the time evolution of Φ, since
Φ is gauge-variant. Systems whose time-evolution involves arbitrary functions are often called singular.
4.9.2 First-order Cauchy problem

Now consider this same Cauchy Problem from the point of view of the fields E and B. Taking the curl of the
first-order curl equations (4.68), we arrive at:
E = 4πke ∇ρ + 4πkm ∂t J
(4.72)
B = − 4π km ∇ × J
These look like wave equations for six quantities. But only those of their solutions which also satisfy the
first-order field equations (4.68), including at initial time t0 , are acceptable.
The two first-order divergence equations contain no time derivatives and are thus constraints on E and B at
t = t0 . The constraint equation on E can be rewritten ∇2 u = ρ, a Poisson-type equation which can be solved for
u at initial time so long as ρ falls off faster than 1/r 2 at infinity). In the case of B, the scalar field u satisfies a
Laplace equation everywhere and is therefore zero. So B has no longitudinal component, only transverse ones. In
both cases, the longitudinal component is either zero or can be solved for at t0 , so cannot be freely specified.
Now look at the two first-order equations (4.68) which contain time derivatives. Suppose we specify E and
∂t E at t = t0 , so as to solve the 2nd -order equations, eq. (4.72). Then the two transverse components of B are
determined by ∇×B = 4πkm J+∂t E/c2 ; ∂t B is determined, also at t = t0 , by the curl equation for E. Therefore,
once we have specified the two transverse components of E and their time derivatives, the first-order equations
take over and determine the others at t = t0 . Alternatively, we could have specified the two transverse components
of B and their time derivatives at t = t0 to constrain all the other field components and time derivatives.
You can also use (EXERCISE) the transverse/longitudinal projections of the first-order equations (4.68) to
show that in source-free space, only the transverse components of E and B obey a classical wave equation.
Thus, the results of the first-order Cauchy-data analysis are fully consistent with the second-order analysis
on A: only two transverse components correspond to independent, physical dynamical degrees of freedom, This
Cauchy analysis does not rely on some particular solution, but is valid for any electromagnetic field and potential.
Since Maxwell’s theory contains no information about ∇ · A, this must be supplied by a so-called gauge
condition. One that is frequently used is the Lorenz condition: ∇ · A = −∂t Φ/c2 ,. Inserting it into Maxwell’s
equation (4.70) for A could lead you to believe that Φ and the three components of A propagate to infinity, whereas
I hope to have convinced you that only the transverse components of A do. In Appendic M, moreover, we show
that AL can be made to disappear without affecting Maxwell’s equations for the fields and the potentials.
106
Appendices
J Solving an Inhomogeneous Equation in Terms of Homogeneous Solutions
Let f1 and f2 be independent solutions to the homogeneous differential equation (4.5). We use them to derive a
particular solution finh (t) to the inhomogeneousequation. The key step is to insert finh (t) = f1 (t)g(t) to obtain
a first-order equation for ġ: g̈ + dt (ln f12 ) + β/α ġ = F/αf1 . Then the general first-order solution (4.4), together
with Abel’s formula (4.6) and W (x)/f12 = dt (f2 /f1 ), yields:
Z t
f2 f1 (t′ ) F (t′ ) ′
ġ(t) = dt B + ′
dt
f1 a [α W ](t )
Z t Z t
f2 f1 (t′ ) F (t′ ) ′ f2 f1 (t′ ) F (t′ ) ′
= dt B + ′
dt − dt ′
dt
f1 a [α W ](t ) f1 a [α W ](t )
Z t
f2 f1 (t′ ) F (t′ ) ′ f2 (t) F (t)
= dt B + ′)
dt −
f1 a [α W ](t [α W ](t)
where B is an arbitrary constant. A final integration leads to:

Z t
f1 (t′ ) f2 (t) − f2 (t′ ) f1 (t)
finh (t) = f1 g = F (t′ ) dt′ + Af1 (t) + Bf2 (t) (J.1)
a [α W ](t′ )
Because we have not implemented homogeneous boundary conditions (B.C.) on this inhomogeneous solution, we
must include the terms Af1 + Bf2 , even though they look like belonging to the homogeneous solution.
We know that finh (t) must satisfy homogeneous (B.C.). We consider the two most important cases.
With one-point B.C. (IVP), finh (t) and its derivative must vanish at t = a. The integral term and its derivative
are automatically zero at t = a. The other contribution also vanishes because the IVP has no non-zero homoge-
neous solution for homogeneous B.C.. In other words, the integral term satisfies the B.C. without any help from
the adjustable constants A and B. Therefore, the inhomogeneous solution to an IVP Is:
Z t Z ∞
f1 (t′ ) f2 (t) − f2 (t′ ) f1 (t) ′ ′ ′ f1 (t′ ) f2 (t) − f2 (t′ ) f1 (t)
finh (t) = ′)
F (t ) dt = θ(t−t ) ′)
F (t′ ) dt′
a [α W ](t a [α W ](t
(J.2)
and it is the general solution when the boundary conditions on the general solution are homogeneous. When they
are not, we must add to finh (t) the homogeneous solution with appropriate non-zero constants A and B. Of course,
α should not vanish for t > a, and neither can the Wronskian, but the latter is guaranteed by our assumption that
f1 and f2 are linearly independent. We conclude that a unique solution to the IVP always exists, provided that
α 6= 0 and that the source term, F (t), is piecewise continuous for t > a.
We should check that our solution (J.2) satisfies the inhomogeneous equation. A surprise awaits us: because
L[f1 ] = L[f2 ] = 0, the integrand does not contiribute to L[finh ] for a ≤ t′ < t; the sole contribution must come
from the point t′ = t. This suggests that the Direc delta-function must somehow be involved, and this is indeed
what happens if we we use the second expression with the step-function, whose derivative is the delta-function.
The other case we wish to address is the Dirichlet problem, ie. the boundary-value problem (BVP) with f
specified at the two end-points. While the integral term in eq. .(J.1) satisfies a homogeneous B.C. at t = a, it does
not at the other end of the interval, at t = b. Enforcing finh (b) = 0 requires adjusting the constants A and B.
finh (a) = 0 immediately leads to B = −Af1 (a)/f2 (a). Then finh (b) = 0 determines A, and we arrive at:
Z t
f1 (t′ ) f2 (t) − f2 (t′ ) f1 (t)
finh (t) = F (t′ ) dt′
a α(t′ ) W (t′ )
Z
f2 (a)f1 (t) − f1 (a)f2 (t) b f1 (t′ ) f2 (b) − f2 (t′ ) f1 (b)
+ F (t′ ) dt′
f2 (a)f1 (b) − f1 (a)f2 (b) a [α W ](t′ )
107
This expression looks more symmetric if we combine the two integrals from a to t. Some tedious algebra yields:
Z
f2 (b)f1 (t) − f1 (b)f2 (t) t f1 (t′ ) f2 (a) − f2 (t′ ) f1 (a)
finh (t) = F (t′ ) dt′
f1 (a)f2 (b) − f2 (a)f1 (b) a [α W ](t′ )
Z
f2 (a)f1 (t) − f1 (a)f2 (t) t f1 (t′ ) f2 (b) − f2 (t′ ) f1 (b)
+ F (t′ ) dt′
f1 (a)f2 (b) − f2 (a)f1 (b) b [α W ](t′ )
which can be written in the compact form:
Z b " #
f2 (b)f1 (t> ) − f1 (b)f2 (t> ) f2 (a) f1 (t< ) − f1 (a) f2 (t< )
finh (t) =
′ ) W (t′ ) f (a)f (b) − f (a)f (b)
F (t′ ) dt′ (J.3)
a
α(t 1 2 2 1
where t> := max(t, t′ ) and t< := min(t, t′ ). The existence of the inhomogeneous solution depends on the
denominator of the integrand not vanishing, as well as piecewise continuity of F (t).
As with the IVP, checking the validity of this solution by calculating L[finh ] reveals the same behaviour: only
the point t′ = t contributes. The presence of the delta-function is easier to see if we write the expression in terms
of step-functions which split the integral over two intervals.
The Green-function formalism introduced in the main body of these notes will shed more light on these results.
K Solution of a Homogeneous IVP with Homogeneous B.C.

We show that the only eigenfunction of the operator L = d2t + β(t)dt + γ(t) with eigenvalue zero, that satisfies
homogeneous one-point boundary conditions, is the zero function. The relevant eigenvalue problem is the homo-
geneous equation: L[f (t)] = 0 over [a, b], with f (a) = f˙ a = 0. The proof is adapted from section 13.3 in
Hassani’s textbook. The strategy relies on the following, easy to show, fact:
If there exists a constant, c > 0, and a differentiable function, h(t), such that ḣ(t) ≤ ch(t), ∀ t ∈ [a, b], then:
h(t) ≤ h(a)ec(t−a) .
Indeed, starting from ḣ ≤ ch and recalling that an exponential is never negative, one has: ḣ e−ct ≤ ch e−ct ,
or: dt (h e−ct ) ≤ 0. Integrating yields: h(t) e−ct − h(a) e−ca ≤ 0, which is the result sought.
Therefore, if we can find a function h of f and f˙ that satisfies ḣ ≤ ch for some positive constant c and that
vanishes at t = a by virtue of the initial conditions on f , then it will be bounded from above by zero; and if
this same function happens to be a linear combination of positive functions of f and f˙, we can conclude that the
function is exactly zero, and so therefore are f and f˙.
We make a prescient choice that does satisfy h(a) = 0: h = f 2 + f˙2 ≥ 0, with appropriate constants (dropped
here to lower clutter) to make units consistent if t is not dimensionless.
Now comes the somewhat fiddly part. First, note that (f ± f˙)2 = f 2 + f˙2 ± 2f f˙, so that |2f f˙| ≤ f 2 + f˙2 .
Now differentiate h, and use the homogeneous equation to eliminate f¨, plus the properties of absolute values:

|ḣ| ≤ (1 + |γ|) 2f f˙ + 2|β|f˙2 ≤ (1 + |γ|) (f 2 + f˙2 ) + 2|β|f˙2 = (1 + |γ|) f 2 + 1 + |γ| + 2|β| f˙2
Since the coefficient of f 2 on the right-hand side is always smaller than the coefficient of f˙2 , we can take the
maximum of the latter over [a, b] as our constant c, and we have proved that our choice h = f 2 + f˙2 ≥ 0 satisfies
ḣ ≤ |ḣ| ≤ c h, which with h(a) = 0 is the condition to have h ≤ h(a)ec(t−a) = 0. The only possibility is then
h = 0, and thus f = f˙ = 0 over the interval.
A shorter argument relies on expressing the general homogeneous solution in a basis {f1 , f2 }: fh = Af1 +
Bf2 . Differentiating and setting f (a) = f˙ a = 0 yields the matrix equation:

f1 (a) f2 (a) A 0
=
f˙1 a f˙2 a B 0
For at least one of A or B to be non-zero, the determinant of the matrix, ie., the Wronskian of f1 and f2 , must vanish
at t = a. But then, these functions being themselves solutions of the homogeneous equation, their Wronskian
vanishes everywhere in [a, b] if it vanishes anywhere because of Abel’s formula (4.6), and the functions are linearly
dependent, a contradiction. we conclude that A and B, and therefore fh , are zero.
108
L Modified Green Functions for the One-dim Boundary-value Problem

In sections 4.1.1 and 4.1.2 we saw that the one-dim BVP has no solution in terms of Green functions unless no
homogeneous solution fh 6= 0 exists that satisfies the homogeneous boundary conditions (B.C.) at x = a and b.
There is an escape clause, however.
For simplicity’s sake, let us assume that there is only one eigenfunction φ0 (x) corresponding to eigenvalue
λ0 = 0 that obeys [Lφ0 ](x) = 0 with homogeneous B.C. If a solution f to L[f ]P = F exists, it can be expanded
over the complete set of orthonormal eigenfunctions of (self-adjoint!) L: f (x) = j6=0 aj φj (x). Then [Lf ](x) =
P j
j6=0 a λj φj (x) = F (x). Orthogonality of φ0 with the other φj immediately gives:
Z b
φ∗0 (x) F (x) dx = 0 (L.1)
a
Thus, a solution exists only if the driving term is itself orthogonal to φ0 over the interval. To discover what form
that solution takes, consider the modified Green function:
X φj (x) φ∗ (x′ )
′ j
G(x, x ) := λj 6= 0 (L.2)
λj
j6=0
While superficially identical to eq. (4.12), this expression specifically omits the now non-zero φ0 (x) φ∗0 (x′ ) term
which simply did not exist in eq. (4.12).
Not surprisingly, and although Pthey satisfy the same homogeneous B.C., G(x, x′ ) and G(x, x′ ) solve different
defining equations: [LG](x, x ) = all j φj (x) φ∗j (x′ ) − φ0 (x) φ∗0 (x′ ), that is:
′

LG (x, x′ ) = δ(x − x′ ) − φ0 (x) φ∗0 (x′ ) (L.3)
because {φj } with φ0 included is a complete set. Then one quickly shows that if condition (L.1) holds, the form:
Z b
f (x) = C φ0 (x) + G(x, x′ ) F (x′ ) dx′ (L.4)
a
is the solution to [Lf [(x) = F (x). But since C is an arbitrary constant, we have lost unicity—in fact the number
of solutions is infinite. Of course, if f must satisfy non-homogeneous B.C., we must also add the homogeneous
solution that satisfies them.
When solving eq. (L.3) to find a modified Green function, we can proceed as in sections 4.2.2 and 4.3. Adding
the particular solution of [LG](x, x′ ) = −φ0 (x) φ∗0 (x′ ) to its homogeneous solution does not change the conditions
on G(x, x′ ) at x = x′ , but that particular solution must be added to (4.13), which will alter the b1 and b2 coefficients
calculated by imposing the relevant homogeneous B.C. on G. Unfortunately, without an explicit φ0 , it becomes
impossible to write a general result for the modified Green function.
Example L.1. The one-dim Laplace equation, d2x f (x) = 0, has for general solution fh (x) = Ax+B,
with A and B constants. The Dirichlet B.C., fh (a) = fh (b) = 0, lead to fh (x) = 0 everywhere,
and the Green function always exists. The homogeneous Neumann B.C., dx fh a = dx fh b = 0,
however, do not determine B, and the homogeneous equation is solved by the non-trivial φ0 (x) = B.
√ once, the same Neumann B.C. on the
Integrating eq. (L.3) modified Green function GN are consistent
only if B = 1/ L, with L := b − a. The equation dx GN = −1/L for x 6= x′ is then solved by:
2
GN (x, x′ ) = −x2 /2L + b1 (x′ )x + b2 (x′ ). Implementing the homogeneous Neumann B.C. leads to:
GN (x, x′ ) = −x2 /2L + ax/L + θ(x − x′ )(x − x′ ) + b2 (x′ ), with b2 (x′ ) arbitrary.
Provided the source in d2x f = F integrates to zero over [a, b], as required by eq. (L.1), a general
solution to this Neumann problem (with homogeneous B.C.!) is given by eq. (L.4), but it is unique
only up to an arbitrary constant.
109
M Counting Electromagnetic Degrees of Freedom in the Lorenz Gauge

The key observation is that one can change both A and Φ to new functions that still obey the Lorenz condition.
Indeed, let f be some scalar function that satisfies the homogeneous wave equation f = ∇2 f − c12 ∂t2 f = 0.
Then add ∇2 f to ∇ · A and c12 ∂t2 f to −∂t Φ/c2 to obtain:
1
∇ · (A + ∇f ) = − ∂t (Φ − ∂t f ) (M.1)
c2
This shows that gauge-transformed potentials still satisfy the Lorenz condition! As noted before, it is important to
keep in mind that since the transformation shifts A by a gradient, which is a longitudinal object, it does not affect
the transverse components of A.
Now, for the first time, we shall have to look at actual solutions of the wave equations for A and Φ. To make
things as simple as possible, take plane-wave solutions A = A0 ei(kx−ωt) , where the x-axis has been aligned along
the direction of propagation, and Φ = Φ0 ei(kx−ωt) . Then:
∇ · A = ∂x Ax = ikA0x ei(kx−ωt) , ∂t Φ = −iωΦ0 ei(kx−ωt)
Inserting into the Lorenz condition with ω/k = c yields, as expected, a relation between the longitudinal compo-
nent Ax and Φ: A0x = Φ0 /c.
Now fold in f = f0 ei(kx−ωt) into eq. (M.1) for the gauge-transformed potentials, to get:
ω
ik (A0x + ik f0 )ei(kx−ωt) = i (Φ0 + i ω f0 ) ei(kx−ωt)
c2
Since f0 is arbitrary, we can choose it to cancel A0x , which at the same time gets rid of Φ0 , leaving us with only
the transverse components of A!
Tte conclusion is the same as that of the analysis of the field equations: only the two transverse components of
A propagate, in the sense that they carry energy to infinity.
110

p154022_cumullect

Uploaded by

Copyright:

Available Formats

p154022_cumullect

Uploaded by

Document Information

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

p154022_cumullect

Uploaded by

Copyright:

Available Formats

Contents

1 CHAPTER I — TENSORS AND EXTERIOR CALCULUS ON MANIFOLDS 1

A Tangent Spaces as Vector Spaces 26

B Transformation of Vector Components Between Coordinate Systems 27

C Three-dim Inhomogenous Maxwell Equations in the p-form Formalism 28

2 CHAPTER II — A BRIEF INTRODUCTION TO GROUP THEORY 29

2.2.3 Cosets (BF 10.3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

D The Right and Left Actions of a Group on a Vector, with Sn as Example 52

E Proof of the Second orthogonality Relation for Characters 54

F Direct Product of Representations 54

G A Second Example of Symmetry-Breaking Lifting a Degeneracy 55

3 CHAPTER III — LIE GROUPS 57

3.3.2 Definition of a Lie algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

H Commutators of Angular Momentum with Vector Operators 81

I Alternative Derivation of the Master Formula 82

4 CHAPTER IV — Solution of Differential Equations with Green Functions 83

4.7.1 Dirichlet and Neumann Boundary Conditions for an Elliptic Problem . . . . . . . . . . . 97

J Solving an Inhomogeneous Equation in Terms of Homogeneous Solutions 107

K Solution of a Homogeneous IVP with Homogeneous B.C. 108

L Modified Green Functions for the One-dim Boundary-value Problem 109

M Counting Electromagnetic Degrees of Freedom in the Lorenz Gauge 110

1 CHAPTER I — TENSORS AND EXTERIOR CALCULUS ON MANIFOLDS

1.1 Vector Spaces and Linear Mappings

Definition 1.2. If any v ∈ V can be written as a linear combination‡ :

1.1.2 The space dual to a vector space

(a Ti + Tj )(V) = a Ti (V) + Tj (V) ∈W a Ti + Tj ∈ L(V, W)

ω α (v) = ω α (v β eβ ) = v β ω α (eβ ) = v β δαβ = v α

1-form Input vector Output Vector Input 1-form Output

1.2 Where Do Vectors Live?

1.2.1 Manifolds and coordinates

• EvenPthough Rn can be endowed with a manifold structure, a unit ball in Rn , defined

1.2.2 Curves, directional derivatives and vectors

1.2.3 The tangent space of a manifold

Definition 1.8. A tangent space TP to a manifold M n at a point P ∈ M n , is a set that can be

1.2.4 Directional derivatives as vectors

Definition 1.9. Given an arbitrary differentiable function f on a manifold M n , parametrised in a local

Definition 1.10. The tangent space TP to a manifold M n admits a basis {∂ν P

∂ϕ (sin θ cos ϕ, sin θ sin ϕ, cos θ) = (− sin θ sin ϕ, sin θ cos ϕ, 0)

1.2.5 Differential of a function and basis dual to a coordinate basis

Definition 1.11. Let xµ ∈ Rn (µ = 1, . . . , n) be the coordinate functions of arbitrary coordinates at

[df ](t) := t(f ) = ∂t f (1.8)

[dxµ ] (∂ν ) = ∂ν xµ = δµν (1.9)

1.2.6 Transformations on bases, cobases, and components

1.3 At Last, Tensors!

1.3.1 The tensor product

1.3.2 Transposition, symmetric and skew-symmetric tensors

A (0, 2) tensor can also be similarly decomposed.

1.3.3 Transformations on tensors

1.3.4 The Levi-Civita symbol

1.4 Two More Ways to Construct Tensors

Tr T = T(dxµ , ∂µ ) = T α β ∂α (dxµ ) dxβ (∂µ ) = T µ µ = T µ ν δνµ

1.4.2 Inner product

ũ := g( , u) = gµν dxµ dxν (u) = gµν uν dxµ

< u, v > ≡ u · v := g(u, v) = gµν uµ v ν (1.30)

gµ ν = g(dxµ , ∂ν ) ≡ < dxµ , ∂ν > = dxµ (∂ν ) = δµν (1.31)

1.4.3 The metric

g(∆x, ∆x) = −c2 (t1 − t2 )2 + (x1 − x2 )2 + (y1 − y2 )2 + (z1 − z2 )2

df = ∂t f dt + ∂i f dxi ∇f = − ∂ct f ∂ct + ∂ i f ∂i (∂ i = gij ∂j = ∂i )

1.5 Exterior Algebra

(σ 1 1 σ 2 2 − σ 1 2 σ 2 1 ) dx1 ∧ dx2 + (σ 1 3 σ 2 1 − σ 1 1 σ 2 3 ) dx3 ∧ dx1 + (σ 1 2 σ 2 3 − σ 1 3 σ 2 2 ) dx2 ∧ dx3