Lecture notes for Mathematical Physics

Joseph A. Minahan1
Department of Physics and Astronomy
Box 516, SE-751 20 Uppsala, Sweden

1
E-mail: [email protected]

1
1 Introduction
This is a course in Mathematical methods in physics. I should stress at the very beginning
that I am a physicist and not a mathematician, so many of the proofs and exercises offered,
will not be as rigourous as would be given by a proper mathematician. In stead, our goal
will be to develop a set of tools that will be useful for a would be physicist. Much of what
I hope to cover in this course is commonly used by a practicing theoretical physicist.
This course is still a work in progress, but I hope to cover the following topics:
• Group Theory and Lie Algebras
• Path Integrals
• Topology
• Differential Geometry
• Yang-Mills
Each one of these topics is a course in itself, so much of the presentation here will be
somewhat sketchy.

2 Group Theory
A physics problem can be simplified if there is symmetry in the problem. For example, you
probably remember that in ordinary one dimensional quantum mechanics, if the potential
is invariant under the symmetry x → −x, that is,

V (−x) = V (x), (2.0.1)

then we immediately know that the eigenfunctions of the Hamiltonian are either even or
odd under this transformation. This transformation is known as parity and is an element
in one of the simplest examples of a group. We say that the wavefunctions transform
under a representation of the group.
A less trivial example that you should know from your quantum mechanics courses
is the example of a central potential, where the potential only depends on a radius and
not on the angles. This potential is invariant under rotations and the wave functions
can be classified by different eigenvalues of the angular momentum operator. These three
dimensional rotations form a group, the rotational group, and the wave functions are in
representations of this group, the different representations classified by `, where

~ 2 ψ` (r, θ, φ) = h̄2 `(` + 1)ψ` (r, θ, φ).

L (2.0.2)

2
 These two different groups have some significant differences. In the first case, the
group is finite – there is only one nontrivial transformation (x → −x). In the second case,
the group is continuous. The group of rotations is described by 3 continuous angles, the
Euler angles.
There is another significant difference between these groups. In the first case, under a
series of transformations, it does not matter in which order you perform the transforma-
tions. The end result is the same. However, in the second case, the order does matter. A
rotation in the x − y plane followed by a rotation in the x − z plane leads to a different
result than a rotation in the x − z plane followed by a rotation in the x − y plane. The
first case is an example of an Abelian group, and the second case is the example of a
Non-Abelian group.
Now that we have given these examples, let us try and give a more concrete definition
of a group.

2.1 Groups (the definitions)

A group is a set G with a collection of objects, gi in G. The group has associated with
it an operation which we write as “·” and usually call “multiplication”. The elements of
the group have to satisfy the following properties
• If g1 , g2 ∈ G, then g1 · g2 ∈ G.
• There is a unique identity, 1 ∈ G, such that gi · 1 = gi , ∀gi ∈ G.
• For each gi ∈ G, there is a unique inverse, gi−1 ∈ G, such that gi · gi−1 = 1.
• Associativity: (g1 · g2 ) · g3 = g1 · (g2 · g3 ).
That’s it! Any collection of objects with some operation that satisfies these 4 properties
is a group. Note that g1 · g2 = g2 · g1 is not a requirement for this to be a group. However,
if this is true for all elements of the group, then we say that the group is Abelian. Another
term that we will use is the order of a group, which is the number of elements.

2.2 Examples
Here are some examples.
1) The integers under addition. In this case “·”= “+”. n1 + n2 = n3 . If n1 and n2
are integers then clearly n3 is an integer, so the first property is satisfied. The identity
element is 0, and it is obviously unique. The unique inverse of n1 is −n1 . Finally, addition
is associative. Therefore this is a group. Moreover, addition is commutative, that is

3
n1 + n2 = n2 + n1 , so the group is Abelian. Notice that the integers under multiplication
is not a group, since in general the inverse is not an integer.
2) Parity: x → −x. This group has two elements, 1 and Π, where Π2 = 1,2 hence Π
is its own inverse. This group is clearly Abelian, and has order 2.
3) Permutation groups on N elements. The elements of this group are generated by,
gij , the operations that exchange element i with j. The generators of a group are a subset
of G out of which all group elements can be constructed. As an example, consider the
permutation group on 3 elements. The generators of this group are g12 , g23 and g13 .
Actually, we don’t even need g13 , since we can generate it with the other two elements:

g13 = g12 g23 g12 . (2.2.1)

To see this, act with g12 g23 g12 on 3 elements (a, b, c).

g12 g23 g12 (a, b, c) = g12 g23 (b, a, c) = g12 (b, c, a) = (c, b, a) = g13 (a, b, c). (2.2.2)

We also see that this group is nonabelian, since g12 g23 6= g23 g12 . Since there are N ! ways
to order N elements, the order of the permutation group is N !.
4) Rotations in the plane. There are many ways to represent this. In matrix notation,
we can express the transformation corresponding to a rotation by angle θ as
!
cos θ − sin θ
(2.2.3)
sin θ cos θ

The transformation of a two dimensional vector is

! ! ! !
x cos θ − sin θ x x cos θ − y sin θ
→ = (2.2.4)
y sin θ cos θ y y cos θ + x sin θ

The product of two such rotations is

! ! !
cos θ1 − sin θ1 cos θ2 − sin θ2 cos(θ1 + θ2 ) − sin(θ1 + θ2 )
= (2.2.5)
sin θ1 cos θ1 sin θ2 cos θ2 sin(θ1 + θ2 ) cos(θ1 + θ2 )

Clearly, this is an abelian group. There is also another way to represent this group. Note
that if we let z = x + iy, then under the transformation in (2.2.4), z transforms as

z → eiθ z (2.2.6)
2
Note, that where the notation is clear, we will often drop the “·”

4
In other words, we can represent the group elements as eiθ . Since these can be represented
by 1 dimensional complex numbers, we call this group U (1). The “U ” stands for unitary,
since for every element g of the group, we have that

g † = g −1 (2.2.7)

U (1) is an example of what is called a compact group. Roughly speaking, this is because
every group element is described by a θ over a compact space, 0 ≤ θ < 2π. This will be
made more precise later on.
5) SL(2, R). Another interesting continuous group that we will come across is SL(2, R),
which stands for “Special Linear 2 dimensional, Real”. The group elements are made up
of 2 by 2 matrices with real entries whose determinant is 1. In other words, we have the
elements !
a b
, ad − bc = 1. (2.2.8)
c d
We can show that this is a group, because if we multiply two such matrices, say A and
B, we are left with a matrix whose entries are real. Furthermore, since det(AB) =
det A det B, we see that the determinant of the product is also 1, hence the product ! is
1 0
an element of the group. There is clearly a unique identity element I = , and
0 1
since the determinant of the matrices is 1, the determinant of their inverses are also 1
with integer entries. Therefore, SL(2, R) is a group. However, the group is nonabelian,
since usually, 2 by 2 matrices don’t commute with one another.
It turns out that SL(2, R) is an example of a group that is noncompact. To see this,
note that for any given a, b, and c, the expression for d is given by d = (1 + bc)/a. Hence
as long as a is nonzero, there is a solution for d. If a = 0, then we must have bc = −1.
In any event, a, b and c can range over all real values. Now this in itself does not make
the group noncompact. It turns out that we will have to define something called a group
measure, such that integrating over this measure gives an infinite result, in the sense that
integrating dx over all x gives an infinite result.
6) SU (2). The group U (N ) has elements which are unitary N × N complex matrices.
If we restrict ourselves to those matrices whose determinant is 1, then these matrices form
a group called SU (N ), where the “SU” stands for special unitary. Let us concentrate on
the case of SU (2). We note that a unitary matrix U can be written as U = exp(iH),
where H is hermitian (H † = H). In 2 dimensions, any Hermitian matrix can be expressed
as
H = a0 I + a1 σ1 + a2 σ2 + a3 σ3 , (2.2.9)

5
 !
1 0
where I = , σi are the Pauli matrices
0 1
! ! !
0 1 0 −i 1 0
σ1 = , σ2 = , σ1 = , , (2.2.10)
1 0 i 0 0 −1

and a0 and ai are real numbers. The determinant is given by det U = exp(iTrH) (prove
this!), therefore, we see that TrH = 0 and so a0 = 0. Hence, we see that our unitary
matrix is described by three numbers a1 , a2 and a3 .
We can now say more about SU (2). If we use the fact that σi σj + σj σi = 2Iδij , then
it must be true that any element U can be written as

U = b0 I + i(b1 σ1 + b2 σ2 + b3 σ3 ) (2.2.11)

where all coefficients are real and where

b20 + b21 + b22 + b23 = 1 (2.2.12)

This last condition insures that the determinant is 1. But the condition in (2.2.12) tells
us something else. Namely that the elements of SU (2) map to the 3 sphere! In other
words, for every element of SU (2), there is a corresponding point on a 3 dimensional
sphere. This will be of importance later on in this course. This 3-sphere is called the
group manifold of SU (2).

2.3 Subgroups and Cosets

A subgroup H of a group G is a subset of the elements of G, such that the elements of the
subset form a group. In other words, we have that hi ∈ H so that the hi satisfy all of the
properties for a group mentioned above. Every group has at least one subgroup, namely
the trivial group made up of one element, the identity element 1. Note that the identity
element for G is the identity element for H.
Given a subgroup we can define a coset. A coset of the group G by the subgroup H
is written as G/H (or sometimes as H\G) and is defined as an identification. In other
words, two elements gi and gj are considered identical elements in the coset if

gi = gj · h for some h ∈ H. (2.3.1)

Actually, this is what is known as a right coset. A left coset has the identification

gi = h · gj for some h ∈ H. (2.3.2)

6
If the group is abelian, then there is no difference between right and left. However, if the
group is nonabelian, then there could be a difference in the cosets, although there will be
a one to one map of the elements in the left coset to the elements of the right coset. n.b.
The coset is not always a group!
One other thing to note is that the order of G divided by the order of the subgroup
is the order of the coset, in other words, the number of elements in the coset. Of course,
this must mean that the order of any subgroup H divides the order of the original group
G (prove this!)

2.4 Examples of Subgroups and Cosets

1)The integers. A subgroup of the integers is all integers that are a multiple of m, where
m is some positive integer greater than 1. Clearly the “product” of any two such elements
gives an integer which is a multiple of m. The coset of such a subgroup is the integers
mod m. Note that these subgroups have an infinite number of elements but the coset
only has a finite number. Note further that the cosets also form a group, the group of
modular addition.
2) Parity. This group only has the identity element as a subgroup.
3) Permutations. Examples of subgroups of the permutation group on N elements are
the permutation groups on M elements where M < N . So for example, the permutation
group on 3 elements has as a subgroup, the permutation group on two elements. This
subgroup has the elements 1 and g12 . A left coset for this subgroup has 3 elements up to
identification: 1, g13 and g23 . Note that 1 ≡ g12 , g13 ≡ g12 g23 and g23 ≡ g12 g13 . The coset
2
is not a group, since for example g13 = 1, but (g12 g23 )2 = g12 g13 , so the identification is
not preserved under group multiplication.
4) U (1). The subgroups of this group are given by ZN , whose elements are given by the
solutions of the equation z N = 1. The solutions are given by e2πin/N . Clearly these form
a group under multiplication. The cosets of these groups are made up of the elements eiφ ,
where we identify elements if φ1 = φ2 + 2πn/N .
5) SL(2, R). One interesting subgroup is the group SL(2, Z), where the entries of the
matrix !
a b
, ad − bc = 1. (2.4.1)
c d
are all integers. Clearly any product gives a matrix whose entries are integers. The
identity also has integer entries. Less trivial to check is that the inverse of any element
only has integer entries. However, this is guaranteed by the fact that the determinant of

7
the matrix is 1. Therefore, the inverse of (2.4.1) is
!
d −b
(2.4.2)
−c a

and so the inverse also has integer entries.

Another subgroup of SL(2, R) is U (1). This is clear since our original formulation of
U (1) was given in terms of real 2 × 2 matrices with determinant 1. We will defer the
discussion on the coset.
6) SU (2). SU (2) also has a U (1) subgroup, actually, it has many U (1) subgroups.
For example, we can choose the subgroup to be those U = exp iφσ3 . The coset is very
interesting. Let a general SU (2) element be given by (2.2.11). Then if we multiply it by
an element of the subgroup, we get

U = b0 cos φ−b3 sin φ+i(b1 cos φ+b2 sin φ)σ1 +i(b2 cos φ−b1 sin φ)σ2 +i(b3 cos φ+b0 sin φ)σ3 .
(2.4.3)
Let us define
w = b0 − ib3 z = b1 + ib2 . (2.4.4)
Therefore under the transformation of U in (2.4.3), we see that

w → e−iφ w z → e−iφ z. (2.4.5)

The constraints on the b’s results in the constraint

|z|2 + |w|2 = 1 (2.4.6)

Identification under the subgroup means that equal rotations of z and w in the complex
plane correspond to the same coset element. Now let us define ze = ρz and we = ρw,
where ρ is any positive real. Then it is clear from z and w, that we have the same point
in the coset if we identify {ze, w}
e ≡ {ρze, ρw},
e since both expressions can come from the
same z and w. If we also include the identification under the phase rotation, then we
see that we can describe the coset using the identification {ze, w}
e ≡ {λze, λw},
e where λ is
any complex number other than zero. This space is known as CP (1), for 1 dimensional
complex projective plane. The dimension of this space is one complex dimension, which
is the same as two real dimensions. This is because ze and we each have one complex
dimension, but the identification removes one dimension. Even though ze and we can be
any complex number (except, both can’t be zero), the space is compact. We will show
this when we discuss topology.

8
2.5 Representations of Groups
A representation, is a mapping of group elements that preserves the group multiplica-
tion law. In fact, up to now, we have been describing examples of groups through their
representations. The elements of the group are a somewhat abstract notion. The repre-
sentations give a concrete description of the groups.
That is not too say that all representations are equivalent. For example, when we
write group elements as matrices, this is a representation of the group. Let us call an
element of this representation M (gi ), where M (gi ) is the matrix corresponding to the
element gi in G. Therefore, we have that

M (gi )M (gj ) = M (gi · gj ). (2.5.1)

Then an equivalent representation is where

f (g ) = AM (g )A−1 ,
M (gi ) → M (2.5.2)
i i

where A is a matrix that is the same for all group elements. If an A exists such that all
M (gi ) can be transformed to the form
!
M
f (g )
1 i 0
M
f (g
i) = , (2.5.3)
0 M2 (gi )
f

where the matrices are all in block diagonal form, then we say that the representation is
reducible. Notice that all of the blocks form representations of the group. So our original
representation is just a combination of smaller representations of the group.
As an example, consider the case of the exchange between two elements. We can then
write the elements as
! !
1 0 0 1
M (1) = M (g12 ) = . (2.5.4)
0 1 1 0

This is one representation of the group. However, if we perform the similarity transfor-
mation M (gi ) → AM (gi )A−1 , where
!
1 1 1
A= √ , (2.5.5)
2 −1 1

then the new matrices are

! !
1 0 1 0
M
f (1) = M
f (g ) =
12 . (2.5.6)
0 1 0 −1

9
Hence the representation is comprised of two representations. One of these is the trivial
represenation where M
f (g ) = 1. The other representation has M
1 i
f (1) = 1 and M
2
f (g ) =
2 12
−1. Obviously, these representations cannot be reduced further. Representations that
cannot be reduced any further are said to be irreducible.
Next consider the group U (1). The matrices in (2.2.3), can be diagonalized to the
form !
eiθ 0
(2.5.7)
0 e−iθ
Hence the original two dimensional representation is reducible to two one dimensional
representations. However, these are not the only representations of U (1). We can easily
see that einθ where n is any integer is a representation of U (1). In fact, these are all of
the irreducible representations.
In the case of SU (2), the representation in (2.2.11) is irreducible. There is no way
to block diagonalize all elements simultaneously, since the different Pauli matrices don’t
commute with each other. This is the smallest nontrivial representation of SU (2). The
one dimensional representation is trivial.

2.6 Lie Groups and Lie Algebras

A Lie group is a continuous group that is specified by a set of parameters. For a compact
Lie Group, the parameter space is compact. The groups U (1) and SU (N ) are examples of
Lie groups. A Lie group has associated with it a set of generators, Ta , that can generate
the entire group. If a group element can be smoothly continued to the identity, then such
a group element can be written as
!
X
U = exp i θa Ta = exp(iθa Ta ), (2.6.1)
a

where the θa are the parameters.

Let us now assume that the parameters are very small, so that U is very close to the
identity. I will write the parameters as θa = a hence
1
U ≈ 1 + ia Ta − a b Ta Tb + O(3 ). (2.6.2)
2
Suppose I consider the product U V U −1 V −1 , where V is also close to the identity, with
parameters 0a . Then the resulting product is also a group element. Performing the
infinitesimal multiplication, we find that

U V U −1 V −1 ≈ 1 − a 0b Ta Tb + a 0b Tb Ta + ... (2.6.3)

10
Since this is a group element, comparing with (2.6.2), we see that the lowest order cor-
rection to the identity in (2.6.3) is a linear combination of the generators. Hence the
generators must satisfy an algebra of the form

[Ta , Tb ] = ifabc Tc . (2.6.4)

This algebra is called a Lie algebra and the coefficients fabc are called the structure con-
stants. The number of independent generators is called the dimension of the algebra. We
will denote this number by D.
If we write the generators in matrix form, then we say that these are in a representation
of the algebra. The structure constants obviously satisfy the relation fabc = −fbac . But
they also have other properties of note. Notice that by the Jacobi identity

[[Ta , Tb ], Tc ] + [[Tc , Ta ], Tb ] + [[Tb , Tc ], Ta ] = 0 (2.6.5)

Therefore, using the algebra in (2.6.4) in (2.6.5) we find

−fabd fdce Te − fcad fdbe Te − fbcd fdae Te = 0. (2.6.6)

Since the Te are assumed to be independent, this reduces to just an equation for the
structure constants. Using their antisymmetry properties, we find

facd fbde − fbcd fade = −fabd fdce . (2.6.7)

In other words, ifacd is a representation for the generator Ta , where the c and d label in
the structure constant refers to the row and column of the matrix. This representation
is called the adjoint representation and its dimension is equal to the dimension of the
algebra, D.
Associated with the Lie algebra (2.6.4) is a subalgebra known as the Cartan subalgebra.
The Cartan subalgebra is made up of a subset of generators, which we call Hi , that satisfy
the algebra
[Hi , Hj ] = 0. (2.6.8)
The Hi can all be chosen to be Hermitian. The number of generators in the Cartan
subalgebra is known as the rank, n of the Lie algebra. Note that there are many ways to
choose a Cartan subalgebra among the elements of the Lie algebra.
Let us now turn to the specific example of SU (2). From eq. (2.2.11) we see that one
representation of the group has generators which are the Pauli matrices, so the dimension
of the algebra is 3. Let us call the generators Ta = 21 σa . Then the algebra is given by

[Ta , Tb ] = iabc Tc , 123 = 1. (2.6.9)

11
Hence the algebra of SU (2) is isomorphic to the algebra for angular momentum in quan-
tum mechanics. The Cartan subalgebra has one generator, namely T3 (although we could
have chosen any other generator), hence the rank of SU (2) is 1. We should also expect
the irreducible representations of SU (2) to be those representations that we found for the
angular momentum, that is, the representations should be related to the allowed spins for
particles consistent with the algebra.

2.7 Roots
Having chosen a cartan subalgebra, we can then classify the other generators. To do
this, let us first recall something we have learned from quantum mechanics. In quantum
mechanics we learned about “bra” and “ket” states. For our purposes, we will define a
set of D independent ket states as
|Ta i. (2.7.1)
In other words, for every independent Ta in the Lie algebra we have a corresponding
independent ket state. These states satisfy the necessary linearity properties of quantum
mechanics, namely that
|αTa + βTb i = α|Ta i + β|Tb i. (2.7.2)
For the bra states we have
hTa | = (|Ta i)† . (2.7.3)
We also need an inner product that satisfies the requirement that

hTa |Tb i∗ = hTb |Ta i. (2.7.4)

It is easy to see that

hTa |Tb i = Tr(Ta† Tb ) (2.7.5)
satisfies the requirement in (2.7.4), since

Tr(Ta† Tb )∗ = Tr(TaT Tb∗ ) = Tr((Tb† Ta )T ) = Tr(Tb† Ta ). (2.7.6)

Now define the operator ΠTc , whose action on the states is

ΠTc |Ta i = |[Tc , Ta ]i = ifcab |Tb i. (2.7.7)

The adjoint of this operator satisfies (ΠTc )† = ΠTc† (show this). ΠTc is obviously a linear
operator since

ΠTc |Ta + Tb i = |[Tc , (Ta + Tb )]i = |[Tc , Ta ]i + |[Tc , Tb ]i = ΠTc |Ta i + ΠTc |Tb i. (2.7.8)

12
It is then straightforward to show using the Jacobi identity and the linearity of the oper-
ators that
[ΠTa , ΠTb ] = Π[Ta ,Tb ] = ifabc ΠTc , (2.7.9)
and so
[ΠHi , ΠHj ] = 0 (2.7.10)
for Hi and Hj in the Cartan subalgebra. Since these operators commute, the states can
be simultaneously eigenstates for all such operators coming from the Cartan subalge-
bra. Furthermore, if the Hi are Hermitian matrices, then the operator ΠHi is Hermitian.
Therefore, its eigenvalues are real.
Thus, let us suppose that we have a Lie algebra with rank n, and so n independent
generators in the Cartan subalgebra. Let us write these as a vector

(H1 , H2 ..., Hn ). (2.7.11)

Let us also suppose that the Hi are orthonormal, so that

Tr(Hi Hj ) = kδij . (2.7.12)

Using our arguments from the preceding paragraphs, a basis can be chosen for the gener-
ators outside the Cartan subalgebra, with basis vectors G~a where ~a is an n dimensional
vector
~a = (a1 , a2 ...an ) (2.7.13)
such that
[Hi , G~a ] = ai G~a . (2.7.14)
The different ~a are called the roots and the G~a are called the root generators. Since the
Hi are Hermitian, the components of the roots are real. Furthermore, if we take the
Hermitian conjugate on (2.7.14), we find that

G~a† = G−~a (2.7.15)

We can also establish some other properties. For one thing, it is clear from (2.7.14)
that commutator of two roots satisfies

[G~a , G~b ] ∼ G~a+~b ~a + ~b 6= 0. (2.7.16)

If ~a + ~b is not one of the roots, then the commutator must be zero. It is also clear from
the hermiticity of ΠHi , that
Tr(G~a G~b ) = kδ~a+~b , (2.7.17)

13
where the δ~0 = 1, and is zero otherwise. The G~a have been rescaled so that they have
the same factor of k as in (2.7.12). It is also clear that all Hi commute with [G~a , G−~a ],
therefore
X
[G~a , G−~a ] = λi Hi . (2.7.18)
Now suppose we act with ΠG~a on |G~b i, then we have

ΠG~a |G~b i ∼ |G~a+~b i, (2.7.19)

Hence ΠG~a is a raising operator and ΠG−~a is the corresponding lowering operator. If we
act with ΠG~a on the state |G−~a i, then we have
X
ΠG~a |G−~a i = λi |Hi i. (2.7.20)
i

But we also have that

hHj |ΠG~a |G−~a i = Tr(Hj [G~a , G−~a ])

= Tr([Hj , G~a ]G−~a ) = aj Tr(G~a G−~a ) (2.7.21)

Hence, using the orthogonality of the Hi , we find that

X
[G~a , G−~a ] = ai Hi . (2.7.22)

One can also check that G~a is traceless, since

Tr[Hi , G~a ] = 0 = ai Tr(G~a ). (2.7.23)

Finally, we can show that the root generator G~a is unique for a given ~a. To see this,
let us suppose that there exists another root generator G~a0 that is orthogonal to G~a . In
other words
hG~a0 |G~a i = Tr[(G~a0 )† G~a ] = 0, (2.7.24)
where (G~a0 )† = G0−~a . Now consider the inner product

hHi |Π†G0 |G~a i = h[G~a0 , Hi ]|G~a i = −ai hG~a0 |G~a i = 0. (2.7.25)

~
a

But going the other way, we have that this is

hHi |[G0−~a , G~a ]i. (2.7.26)

14
But we also know that
[G0−~a , G~a ] =
X
σi Hi , (2.7.27)
i

for some constants σi . Hence consistency with (2.7.25) requires that

[G0−~a , G~a ] = 0. (2.7.28)

Now consider the inner product

hG~a |ΠG~0a Π†G0 |G~a i = 0. (2.7.29)

~
a

That this inner product is zero follows from (2.7.28). But this inner product can also be
written as

hG~a |ΠG~a0 Π†G0 |G~a i = hG~a |[ΠG~0a , Π†G0 ]|G~a i + hG~a |Π†G0 ΠG~0a |G~a i
~
a ~
a ~
a

= hG~a |ai ΠHi |G~a i + hG~a |Π†G0 ΠG~0a |G~a i

~
a

= |~a|2 + hG~a |Π†G0 ΠG~0a |G~a i > 0. (2.7.30)

~
a

Hence we have a contradiction.

2.8 Classifying Groups (Part 1)

In this section we will begin Cartan’s proof of the classification of simple compact Lie
algebras.
We say that a Lie algebra and its corresponding Lie group are reducible, if we can break
up the generators into two groups, say Ta and Ta0 , such that [Ta , Tb0 ] = 0 for all primed and
unprimed generators. We call the Lie algebra simple if it cannot be reduced this way. One
upshot of this is that the generators in the Cartan subalgebra are traceless. For example,
the groups U (N ), and hence their Lie algebras are reducible, since U (N ) = U (1)×SU (N )
and the generator of the U (1) group commutes with all generators of the SU (N ) group.
It turns out that the SU (N ) groups are simple groups.
Let’s now classify the groups. Since the number of generators of the Lie group is finite,
we can always choose a root generator G~b , so that the length of the root vector, |~b|, is
greater than or equal to the length of any other root vector. Let us now act with the
operator ΠG~a on the state for the root ~b

ΠG~a |G~b i. (2.8.1)

15
In order for this to be nonzero |~a + ~b| ≤ |~b| since we have assumed that no root is longer
than ~b. Now by the triangle inequality, at least one of |~a + ~b| or | − ~a + ~b| is greater than
|~b|. We will assume that this is true for the latter case. Now consider the positive definite
inner product
hG~b |ΠG−~a ΠG~a |G~b i. (2.8.2)
Using (2.7.9) and (2.7.22), we have

hG~b |ΠG−~a ΠG~a |G~b i = hG~b |Π−~a·H~ |G~b i = −k ~a · ~b. (2.8.3)

So up to a phase, we have that

q
[G~a , G~b ] = −~a · ~b G~a+~b . (2.8.4)

So among other things, (2.8.4) tells us that orthogonal roots commute with each other,
assuming that at least one of the roots is a longest root.
Now consider the inner product
n 
m−1

n n
−~a · ~b −
Y
hG~b |(ΠG−~a ) (ΠG~a ) |G~b i = kn! ~a · ~a , (2.8.5)
m=1 2

where we obtained the righthand side by using the relation

n−1  n−1
 n  
[ΠG~a , ΠG−~a ] = n ΠG−~a − ~a · ~a + Π~a·H~ . (2.8.6)
2
The relation in (2.8.6) can be proved by induction.
Now we see that there can be a potential problem. The inner product in (2.8.5) is
positive definite, but there are terms in the product on the rhs of (2.8.5) which are negative
if m − 1 is large enough. In order to avoid this problem, we must have
n−1
−~a · ~b − ~a · ~a = 0 (2.8.7)
2
for some positive integer n. Even though we started with the state |G~b i, we could have
started with the state |G~a i and obtained a similar result, since there still cannot be a
state |G~a−~b i since we assumed that no root was longer than |~b|. Hence proceding as in
(2.8.5), we can also derive a relation that

n0 − 1~ ~
−~a · ~b − b · b = 0. (2.8.8)
2

16
 Let us now go through the possible solutions for (2.8.7) and (2.8.8). Since we assumed
that |~b| ≥ |~a|, we see that n ≥ n0 .
1) The first possiblity is that n = n0 = 1. Therefore,

~a · ~b = 0, (2.8.9)

so the roots are orthogonal.

For the other cases, neither n or n0 is equal to 1. Thus, we have that
q
n0 − 1~ ~ (n − 1)(n0 − 1)
~a · ~a = b·b ~a · ~b = − |~a||~b| = cos θ|~a||~b|. (2.8.10)
n−1 2
Therefore, it is necessary that
q
(n − 1)(n0 − 1)
≤1 (2.8.11)
2
2) The second possibility is that n = n0 = 2. In this case
1 1
~a · ~b = − ~a · ~a = − ~b · ~b. (2.8.12)
2 2
Hence these roots have equal length and are at an angle of 120 degrees from each other.
3) The third possibility is that n = n0 = 3. Now we have

~a · ~b = −~a · ~a = −~b · ~b. (2.8.13)

Hence this case has ~a = −~b.

4) The fourth possibility is n = 3, n0 = 2. In this case
√
1 2
~a · ~a = ~b · ~b cos θ = − . (2.8.14)
2 2
5) The fifth possibility is n = 4, n0 = 2, hence this has
√
1 3
~a · ~a = ~b · ~b cos θ = − . (2.8.15)
3 2
6) The sixth and last possibility is n = 5, n0 = 2. In this case
1
~a · ~a = ~b · ~b cos θ = −1. (2.8.16)
4
However, this last case will not be a legitimate choice. This solution requires that ~b = −2~a
and so ~b + ~a = −~a. Consider then

ΠG−~a ΠG~a |G~b i = Π−~a·H~ |G~b i = −~a · ~b|G~b i = 2|~a|2 |G−2~a i. (2.8.17)

17
But we also have, using the fact that root vectors are unique

ΠG−~a ΠG~a |G~b i = ΠG−~a |[G~a , G−2~a ]i ∼ ΠG−~a |G−~a i = |[G−~a , G−~a ]i = 0. (2.8.18)

Therefore, this is a contradiction. This also tells us that different roots cannot be parallel.
In deriving the above, we said that no vector was longer than ~b, but we could have
derived the same result so long as either ~b − ~a or ~b + ~a is not a root. But what if both
are roots, how should we proceed? Well let us suppose that ~b · ~a ≤ 0 (if this were not
true, then we could replace ~a with −~a). Then it must be true that |~b − m~a| > |~a| and
|~b − m~a| > |~b| where m ≥ 1. So for some m we will find that ~b − m~a − ~a is not a root. In
which case, we can proceed as before. Hence we have that
n−1
~a · (~b − m~a) + ~a · ~a = 0
2
2m + 1 − n
⇒ ~a · ~b = ~a · ~a, (2.8.19)
2
where n is some integer. But we have already learned that n can only be 1, 2, 3 or 4. For
n = 1, 2 we have that ~a · ~b > 0 which violates are previous assumption. For n = 3, we can
have ~a · ~b = 0 if m = 1, but for other m it violates the assumption. For n = 4, we have
that ~a · ~b = − 21 ~a · ~a if m = 1. All other values of m violate the assumption. In this last
case, we also have that
2−1 ~
~a · (~b − ~a) + (b − ~a) · (~b − ~a) = 0, (2.8.20)
2
hence we find that ~a · ~a = ~b · ~b. Therefore, these vectors have equal length and are at 120
degrees angle from each other. In other words, we do not get any new possibilities for the
relations between root vectors.
Combining all that we know, and allowing for ~a → −~a, we have that any two root
vectors must satisfy one of the following (assuming that |~b| ≥ |~a|):
1) ~a · ~b = 0
2) ~a · ~b = ± 12~b · ~b ~a · ~a = ~b · ~b
3) ~a · ~b = ± 12~b · ~b ~a · ~a = 21~b · ~b
4) ~a · ~b = ± 12~b · ~b ~a · ~a = 31~b · ~b

2.9 Positive roots and simple roots

Let us write the root vector in component form, where

~a = (a1 , a2 ...an ). (2.9.1)

18
We say a root is a positive root if the first nonzero component in the row vector in (2.9.1)
is positive. A simple root is a positive root that cannot be written as the sum of two other
positive roots. Clearly, any positive root can be written as a linear combination of simple
roots with nonnegative coefficients:
~b =
X
ni α
~i (2.9.2)

where α~ i refers to one of the simple roots. To prove this, suppose that ~b is simple, then
clearly it is equal to a linear combination of simple roots. If ~b is not simple, then we can
write it as the sum of two positive roots. These positive roots are either simple or equal
to the sums of positive roots. We keep breaking things down until we are left with simple
roots.
We also can easily show that if α ~ 1 and α ~ 2 are simple, then α ~1 −α~ 2 is not a root. To see
this, note that either α ~1 − α ~2 − α
~ 2 or α ~ 1 is positive. In the first case, we would then have
α α1 − α
~ 1 = (~ ~ 2) + α
~ 2 , so α
~ 1 is the sum of two positive roots and is therefore not simple.
In the second case we have α α2 − α
~ 2 = (~ ~ 1) + α ~ 1 so α
~ 2 is not simple. Since α ~1 − α
~ 2 is not
a root, then we immediately see based on the discussion in the last section that

~1 · α
α ~ 2 ≤ 0. (2.9.3)

We can now show that the number of simple roots is equal to the rank n of the group.
To see this, let us first show that the simple roots are linearly independent. If they were
not, then it would be possible to write the equation
X X
ci α
~i = dj α
~ j, (2.9.4)

where the cofficients ci and dj are nonnegative and the simple roots on the lhs of the
equation are different from those on the rhs. But then this would imply that if I took the
scalar product of both sides of the equation with the rhs, then the rhs would be positive
~i · α
definite, but the lhs would be less than or equal to zero, since α ~ j ≤ 0, if i 6= j. Hence
we have a contradiction. Then since the roots live in an n-dimensional space and the
simple roots generate all positive roots and are linearly independent, there must then be
n of them. Note that if the roots spanned a space that was less than n dimensional, it
would mean that there is a combination of Cartan generators
n
X
ci Hi (2.9.5)
i=1

that commutes with all the root generators. This means that the generator in (2.9.5)
commutes with all generators of the Lie Algebra, and so this is not a simple Lie algebra.

19
 In any event, we now see that the properties of the groups are determined by the
properties of the simple roots, since these generate all the roots, and so it is only necessary
to classify the allowed sets of simple roots.

2.10 Classifying groups continued

It is useful to use a picture description to describe the simple roots and their relationship
to one another. This is shown in figure 1. In the figure, a simple root is represented by a

!="/2

!=2"/3

!=3"/4

!=5"/6

Figure 1: Diagrammatic relations between simple roots

circle. If two simple roots are orthogonal to each other, then the circles are not attached
by a line segment. If the simple roots are at an angle of 120 degrees (2π/3) then we draw
one line segment between the circles. If the simple roots are at 135 degrees (3π/4), then
we draw two lines between the circles with the direction of the arrow pointing toward the
longer root. Finally, if the roots are at an angle of 150 degrees (5π/6), then we draw three
lines connecting the circles, with the direction of the arrow pointing toward the longer
root.
By including all of the simple roots, we can make a chain of these circles with the
circles attached to each other by the line segments. Since the group we are considering
is simple, we cannot have any disconnected chains. Otherwise this would correspond to
having two sets of simple roots spanning orthogonal spaces. In this case we could break
up all of the roots and the generators of the Cartan subalgebra into two parts, with every
generator in one part commuting with every generator in the other part.
We now show that we cannot have the chains of simple roots shown in figure 2.

20
 Figure 2: Chains with at least one triple link.

If we can find some linear combination of the simple roots such that the square is zero,
then it means that the simple roots are not linearly independent. In some cases, we can
find linear combinations such that the square is negative. For the chains in figure 2, we
have the following results:
(~
α1 + 2~ ~ 3 )2 = 0
α2 + α (~
α1 + 2~ α3 )2 = 0
α2 + 3~ (~
α1 + 2~ α3 )2 = −(~
α2 + 3~ α2 )2
α3 )2
(~
(~
α1 + 2~ ~ 3 )2 = −(~
α2 + α α 1 )2 (~
α1 + α ~ 3 )2 = −
~2 + α (~
α1 + 2~ ~ 3 )2 = −(~
α2 + α α2 )2
2
(~
α1 + 2~ ~ 3 )2 = −2(~
α2 + α α 1 )2 (~
α1 + 2~ ~ 3 )2
α2 + α α 2 )2
= −2(~ (~
α1 + α ~ 3 )2 = −(~
~2 + α α1 )2
(2.10.1)
Clearly, we cannot have any chains where these above chains are subchains, since the
relations in (2.10.1) do not rely on the possible connections to other simple roots. This
gives a very strong constraint on the possible lie algebras. It means that the only allowed
chain with a triple line is

that is, the chain with only two simple roots. The corresponding Lie Group is one of
the exceptional Lie groups and is called G2 . This chain is the first example of a Dynkin
diagram, a chain that corresponds to a Lie group. We will say more about the group G2
later.
We can also rule out the sets of chains with no triple lines but at least one double
line shown in figure 3. The numbers inside the circles in figure 3 indicate the number of
simple roots in a linear combination whose length squared is zero. For example, for the
first chain, the inner product with these coefficients is
(2~
α1 + 2~ ~ 3 )2 = 4~
α2 + α α12 + 8~
α1 · α α22 + 4~
~ 2 + 4~ α2 · α ~ 32 = (4 − 8 + 8 − 8 + 4)~
~3 + α α12 . (2.10.2)

21
 2 2 1 1 1 1 1 2 1

1 1
2 2 1 2
1 1

1 2 3 2 1 2 4 3 2 1

Figure 3: Chains with no triples but at least one double.

Furthermore, if we replace the middle circle in the chain with coefficient m with the
chain

m m m .... m m

where the coefficients of every simple root in the chain are also m, then we find that the
linear combination of simple roots still has zero length.
Hence, we learn that the only sets of chains, that is Dynkin diagrams, that do not
have linearly dependent simple roots have at most one double line. Following our previous
arguments, we learn that the allowed chains with one double line are highly restricted and
have the form in figure 4.
The first two sets of Dynkin diagrams are called Bn and Cn , each with n simple roots.
The corresponding groups are the groups SO(2n + 1) and Sp(n). The first group is the
special orthogonal group, that is the group of rotations in 2n + 1 directions. The other
group is the symplectic group in n complex dimensions. The last Dynkin diagram is for
another exceptional group, known as F4 .
To complete the classification, we need to study chains with single lines only. Such sets
of chains, and the allowed Dynkin diagrams, are called simply laced. It is straightforward
to check that the chains in figure 5 are not allowed, since they will lead to linearly
dependent simple roots. In the first chain in figure 5, we can replace one simple root with

22
 .... Bn

.... Cn

F4

Figure 4: Dynkin Diagrams for root systems with one double line

1 1
1
2
1 1
1 1

2 2

1 2 3 2 1 1 2 3 4 3 2 1

1 2 3 4 5 6 4 2

Figure 5: Simply laced diagrams with linear dependencies for the simple roots.

the chain of simple roots in figure 3, and again find that the there is a linear dependence
in the simple roots. Hence, it is not possible to have any closed loop in a chain. Likewise,
for the second chain in figure 5, we can replace the center simple root with the chain

23
in figure 3 with coefficient 2, and still find that the simple roots are linearly dependent.
The last three diagrams also have linearly dependent simple roots, with the coefficients
relating the simple roots given in the circles. Hence, we have shown that the only simply
laced Dynkin diagrams are of the form in figure 6. The Dynkin diagrams An is for the

.... A
n

.... D
n

E
6

E
7

E
8

Figure 6: The simply laced Dynkin diagrams

Lie groups SU (n + 1). The diagrams Dn are for the groups SO(2n). The last three are
for the exceptional groups E6 , E7 , E8 .
So that is it! The only simple compact Lie groups are SU (n), SO(n), Sp(n) and the
exceptional groups G2 , F4 , E6 , E7 and E8 .

24
 One outgrowth of this, is that we can see that some groups are the same, or at least
almost the same. For example, we see that SU (4), which has the Dynkin diagram A3
and SO(6), which has the Dynkin diagram D3 , are the same since the diagrams are the
same. Likewise SO(4) is actually a product of SU (2) groups, since D2 is actually two A1
diagrams. In this same way, we can see that Sp(1) and SO(3) are very similar to SU (2)
and that Sp(2) is the same as SO(5).

2.11 Examples
1) SU (2) (A1 ). This group has one simple root, α~ , and hence only two roots, ±~
α. There
is one element in the Cartan subalgebra, Hα~ , and so the commutation relations of H with
G±~α are
[Hα~ , G±~α ] = ±αG±~α [Gα~ , G−~α ] = αHα~ . (2.11.1)
Hence, after an appropriate scaling of H and G, this is the algebra for the angular mo-
mentum operators, with Hα~ = αJz and G±~α = √α2 J± .
It is customary to choose the length of the simple roots in the simply laced diagram to
have length squared 2. However, this choice is arbitrary. Some books choose the length
squared to be 1, which will be the convention that we follow, unless specifically mentioned
otherwise.
2) SU (3) (A2 ). This has 2 simple roots of length squared 1 at 120 degrees from each
other. If we add the two roots together, we have one more root with length squared 1.
These three roots, along with the three negative roots gives 6 roots in total. Combined
with the two elements of the Cartan subalgebra, we find 8 generators. This is as expected,
since 3 by 3 unitary matrices have 9 generators, but if the the matrices are constrained
to have determinant 1, then one of the generators is removed. The roots for SU (3) are
shown in figure 7. Note that if we take the subset of roots consisting of one simple root
and its negative, then this forms the root system of SU (2), hence SU (2) is a subgroup of
SU (3).
3) G2 . This has 2 simple roots, with one root of length squared 1 and the other of length
squared 1/3. The root diagram is shown in figure 8. There are 12 roots, and hence 14
generators in all. G2 is a subgroup of SO(7), the group of rotations in 7 dimensions. The
roots are shown in figure 8.

25
 Figure 7: Roots for SU (3) (A2 ). The simple roots are denoted by the bold lines.

Figure 8: Roots for G2 . The simple roots are denoted by the bold lines. Note that the
long roots of G2 are identical to the SU (3) roots. Hence SU (3) will be a subgroup of G2 .

2.12 The Classical Lie Groups

The groups SU (n), SO(n) and Sp(n) are called the classical Lie groups. The others
are called the exceptional Lie groups. The classical groups are most easily described by
considering what is left invariant under their transformations.
1) SU (n). Consider two complex n dimensional vectors zi and wi , where i is an index
that runs from 1 to n. Consider the quantity wi∗ zi = w† z. Then if we transform z and w
by a unitary transformation, they become Uij zj and Uij wj , where the repeated j index is
summed over. Clearly, w† z is invariant under this transformation.
Now let us see why the An−1 Dynkin diagrams correspond to the SU (n) Lie algebra.
To show this, suppose we consider a basis in an n dimensional space. Let us consider
vectors in this space which have length squared 2 and which have the form

(1, −1, 0...), (0, 1, −1, 0..0), ...(0..0, 1, −1, 0..0), (0..0, 1, −1) (2.12.2)

Clearly there are n − 1 such vectors and that the inner product of these vectors with
themselves is described by the An−1 dynkin diagram. Let us assume that these are the

26
simple roots. Then all of the positive roots have the form

(0, 0..0, 1, 0..0, −1, 0..0) (2.12.3)

Hence there are n(n − 1)/2 positive roots and an equal number of negative roots. Note
that these roots span an n − 1 dimensional space, since the sum of the components is
always 0, and so all of the roots lie in the plane x1 + x2 + ..xn = 0.
Given that these are the root vectors, let us find matrices that have the correct commu-
tation relations. It is not hard to see that the matrix elements for the Cartan subalgebra
should have the form (Hi )kl = δik δil − n1 δkl and that the root vectors should have the
form (Gij )kl = δik δjl . The Cartan subalgebra has been constructed to be traceless. The
trace piece will commute with all the other generators. Since the Cartan subalgebra has
traceless matrices, it has only n − 1 linearly independent generators. The matrices Hi ,
Gij +Gji and iGij −iGji generate all n×n traceless hermitian matrices and these generate
all n × n unitary matrices with determinant 1.
2) SO(m). SO(m) transformations leave invariant xi yi , where xi and yi are chosen to
be real. The generators of a rotation are given by

(Mij )kl = iδik δjl − iδil δjk (2.12.4)

which generates a rotation in the ij plane. These generators form the Lie algebra

[Mij , Mkl ] = i(δil Mjk − δik Mjl − δjl Mik + δjk Mil ). (2.12.5)

If m = 2n, then we can choose as a Cartan basis

Hi = M2i−1,2i . (2.12.6)

Using (2.12.5) and (2.12.6), it is straightforward to show that the root generators are
given by

G±i±j = M2i,2j−1 ± iM2i−1,2j−1 − (±)iM2i,2j − (±)(±)M2i−1,2j (2.12.7)

so in terms of the root vectors, these are

(0, ..0, ±1, ..0, ±1, ..0) (2.12.8)

where the i and j entry are nonzero. It is then easy to show that the simple roots are

(1, −1, 0..0), (0, 1, −1, ..0)..(0, ..0, 1, −1), (0, ..0, 1, 1) (2.12.9)

27
Hence there are n of these roots. It is straightforward to show that the inner product of
the simple roots with themselves is given by the Dn Dynkin diagram.
For m = 2n + 1, we can have the same generators in the Cartan subalgebra and the
root generators in (2.12.7) are also included. In addition, we have the root generators

G±i = M2i,2n+1 ± iM2i−1,2n+1 , (2.12.10)

whose vectors are

(0..0, ±1, 0..0) (2.12.11)
where the i index is nonzero. The simple roots, can then be shown to have the form

(1, −1, 0..0), (0, 1, −1, ..0)..(0, ..0, 1, −1), (0, ..0, 0, 1) (2.12.12)

Note that the last root has a length squared that is 1/2 the other simple roots. It is then
straightforward to show that the inner product of these roots is given by the Bn Dynkin
diagram.
3) Sp(n). The symplectic group leaves invariant the symplectic product

a1i b2i − a2i b1i = (a1 )T b2 − (a2 )T b1 , (2.12.13)

where a1 , a2 , b1 and b2 are assumed to be n dimensional complex vectors. Notice that

the symplectic product is preserved under the transformation aα → exp(iW )aα , bα →
exp(iW )bα where W is any hermitian and antisymmetric matrix and α = 1, 2. But it
is also invariant under aα → exp(i(−1)α V )aα , bα → exp(i(−1)α V )bα where V is any
hermitian symmetric matrix. Finally, we can have transformations that rotate a1 into a2 ,
which have the form a1 → cos V a1 + i sin V a2 and a2 → i sin V a1 + cos V a2 , where V is
hermitian and symmetric (and hence real), or have the form a1 → cos V a1 + sin V a2 and
a2 → − sin V a1 + cos V a2 . Hence, we can express the complete set of generators as a
tensor product
1 ⊗ W + σi ⊗ Vi , (2.12.14)
where the σi are the Pauli matrices and W is antisymmetric hermitian n × n matrix and
the Vi are symmetric hermitian n×n matrices. In matrix form, the generators in (2.12.14)
can be written as !
W + V3 V1 − iV2
. (2.12.15)
V1 + iV2 W − V3
There are n elements in the Cartan subalgebra, which in terms of the 2n × 2n matrices
have the form
(Hi )kl = δik δil − δi+n,k δi+n,l . (2.12.16)

28
There are then 3 distinct types of root vectors. The first have the form

(Gij )kl = δik δjl − δj+n,k δi+n,l , (2.12.17)

the second have the form

(G0i )kl = δi+n,k δil , (2.12.18)
and the third have the form

(G00ij )kl = δi+n,k δjl + δj+n,k δi,l . (2.12.19)

The commutators of these root generators with the Cartan subalgebra is given by

[Hi , Gjk ] = (δij − δik )Gjk

[Hi , G0j ] = −2δij G0j
[Hi , G00jk ] = −(δij + δik )G00jk . (2.12.20)

Including the adjoints of these root vectors, we see that the complete set of roots have
components of the form

(0..0 ± 1, 0..0, ±1, 0..0) or (0..0, ±2, 0..0). (2.12.21)

The n simple roots then are given by

(1, −1, 0..0), ...(0, 0..0, 1, −1), (0..0, 2) (2.12.22)

Clearly, the last root has length squared that is twice the length of the other roots. It
is also clear that the inner products between the simple roots is given by the Cn Dynkin
diagram.

2.13 Representations and weights

Up to now, we have been considering the groups in terms the adjoint representation. This
is reflected in the fact that the transformation on the ket state |Tb i is

ΠTa |Tb i = |[Ta , Tb ]i = fabc |Tc i. (2.13.1)

Hence the number of states in this representation is equal to the number of generators
of the Lie algebra. So for example, for SU (2), we would find three states in the adjoint
representation. This corresponds to the spin 1 angular momentum states.

29
 But we know from our experience with angular momentum that there are an infinite
number of different angular momentum states, labeled by the quantum number j, with j
either integer or half integer. These different values for j correspond to different SU (2)
irreducible representations, with j = 1 being the adjoint representation. In particular,
there is a smaller, but nontrivial representation, with j = 1/2. This nontrivial represen-
tation is known as a fundamental representation. In the case of SU (2), it is 2 dimensional
(spin “up” and spin “down”).
We now show how to find other representations for the other compact Lie groups.
In the case of SU (2), remember that the states in the representation were found by
acting with raising and lowering operators J+ and J− , until eventually the states were
annihilated. This put a constraint on the values of j. It should be clear from the discussion
in sections (2.7)-(2.9), that the positive roots play the role of the raising operators and
the negative roots play the role of the lowering operators.
We can write a state in terms of its weight vector µ~ , as |~µi which satisfies

ΠHi |~µi = µi |~µi (2.13.2)

for all elements in the Cartan subalgebra. For any given representation, there must be
some state |~µmax i such that this state is annihilated by all positive root operators. The
weight µ~ max is called the highest weight of the representation. So just as in SU (2), where
the representation was labeled by j, which is the maximum eigenvalue of Jz , we have that
the representation in this general case is labeled by the value of the highest weight.
Now if the weight is annihilated by all the positive root operators, then it is clearly
annihilated by all the simple root operators. In this case, we can proceed as we did in
section (2.8) and hence find essentially the same equation as in (2.8.5), that is
ni 
m−1
 ni  ni Y 
h~µ| ΠGα~ i ΠG−~αi |~µi = Cn! ~i · µ
α ~− ~i · α
α ~i , (2.13.3)
m=1 2

where C is a normalization constant. Thus we find that for every simple root, the highest
weights must satisfy an equation of the form
qi
~i · µ
α ~= ~i · α
α ~ i, (2.13.4)
2
where the qi are nonnegative integers. This is to insure that a manifestly positive definite
quantity is not actually negative. A particular useful class of representations are the
fundamental representations, where the fundamental representation for root α ~ i has qj =
δi,j .

30
 Let us consider some examples:
1) SU (2): We have only one simple root, so we find for the fundamental representation
that the weight has 1/2 the length of the root.
2) SU (3): Now we have two simple roots, hence there are two fundamental represena-
tions. We can write the highest weights as a linear combination of roots with fractional
coefficients
µ
~ 1 = c1 α
~ 1 + c2 α
~2 µ ~ 1 + c02 α
~ 2 = c01 α ~ 2. (2.13.5)
In the first case we have
1 2 1
(c1 α ~ 2) · α
~ 1 + c2 α ~1 = α~1 · α
~ 1, (c1 α ~ 2) · α
~ 1 + c2 α ~2 = 0 ⇒µ
~1 = α~1 + α~ 2 , (2.13.6)
2 3 3
while in the second case we have
1 2
µ
~2 = α~1 + α~ 2. (2.13.7)
3 3
Let us now find the other weights in these representations. We will write the ket states
in terms of the components of the simple roots that make up the weights. In the first
case, we act on the state |2/3, 1/3i first with Π−~α1 giving C| − 1/3, 1/3i, where C is a
normalization constant. We now act with Π−~α2 on this state, giving | − 1/3, −2/3i. The
three elements |2/3, 1/3i, | − 1/3, 1/3i and | − 1/3, −2/3i make up this representation.
This fundamental representation is written as 3 (or sometimes as 3)
We can derive the weight vectors for the other fundamental representation by inter-
changing the two simple roots. Hence, the states in this representation, which is called
3 are |1/3, 2/3i, |1/3, −1/3i and | − 2/3, −1/3i. Comparing these two representations,
we see that weights of one are negative the weights of the other representation. Recalling
that the adjoint conjugation takes the root vectors to minus themselves, we see that the
fundamental SU (3) representations are conjugate to each other. But this also means that
the fundamental representations are not real representations. Figure 9 shows the two
fundamental representations.

31
Figure 9: Fundamental representations for SU (3). One representation has triangles point-
ing up, while the other has the triangles pointing down. The root vectors are also shown.

2.14 A more general weight formula

Let us assume that we have a weight state |µi that is not necessarily a highest weight.
However, if we act with a root vector G~a enough times, eventually we will annihilate the
state. Hence, there exists a nonnegative integer p such that3

(G~a )p+1 |~µi = 0. (2.14.1)

Likewise, there exists a nonnegative integer q such that

(G−~a )q+1 |~µi = 0. (2.14.2)

Consider then the inner product of the commutator

~ µi = ~a · µ
h~µ|[G~a , G−~a ]|~µi = h~µ|~a · H|~ ~ (2.14.3)

where the state |~µi is assumed to be properly normalized. Now we have that

G~a |~µi = N~a,~µ |~µ + ~ai, (2.14.4)

where |~µ + ~ai is the normalized state, that is

h~µ + ~a|G~a |~µi = N~a,~µ . (2.14.5)

Likewise, we have that

G−~a |~µi = N−~a,~µ |~µ − ~ai. (2.14.6)
Taking the complex conjugate of (2.14.4), we see that

(N~a,~µ )∗ = N−~a,~µ+~a . (2.14.7)

3
From now on, we will write ΠTa simply as Ta , where the fact that these are operators is implied.

32
 Now these last sets of equations hold for a general weight in any representation, so in
particular they hold for |µ + n~ai where n is any integer between −q and +p. Hence, using
(2.14.3) we can write the series of equations

|N~a,~µ+p~a |2 − |N~a,~µ+(p+1)~a |2 = ~a · (~µ + p~a)

|N~a,~µ+(p−1)~a |2 − |N~a,~µ+p~a |2 = ~a · (~µ + (p − 1)~a)
|N~a,~µ+(p−2)~a |2 − |N~a,~µ+(p−1)~a |2 = ~a · (~µ + (p − 2)~a)
..... = ....
|N~a,~µ−(q−1)~a |2 − |N~a,~µ−(q−2))~a |2 = ~a · (~µ − (q − 1)~a)
|N~a,~µ−q~a |2 − |N~a,~µ−(q−1))~a |2 = ~a · (~µ − q~a) (2.14.8)

Because of (2.14.1) and (2.14.2) we have that N~a,~µ+(p+1)~a = N~a,~µ−q~a = 0. Then, if we add
up both sides of the equations in (2.14.8), we see that the left hand side all cancels and
we are left with
p
X p
X
0 = (p + q + 1)~a · µ
~+ p~a · ~a − q~a · ~a
n=1 m=1
!
p(p + 1) q(q + 1)
= (p + q + 1)~a · µ
~ + ~a · ~a −
2 2
p−q
 
= (p + q + 1) ~a · µ~+ ~a · ~a . (2.14.9)
2
Hence, a general weight µ and a general root ~a satisfy
q−p
~a · µ
~= ~a · ~a. (2.14.10)
2
Let us use this to find the weights of one of the fundamental representations of SU (5).
Let us write the simple roots as

~ 1 = (1, −1, 0, 0, 0), α

α ~ 2 = (0, 1, −1, 0, 0), α~ 3 = (0, 0, 1, −1, 0), α ~ 4 = (0, 0, 0, 1, −1).
(2.14.11)
The highest weight state for the first fundamental representation has a weight µ ~ 1 which
satisfies
1
~1 · α
µ ~1 = α ~1 · α~ 1 = 1, µ ~1 · α~ i = 0, i 6= 1. (2.14.12)
2
Thus,
µ
~ 1 = (1, 0, 0, 0, 0), (2.14.13)

33
and the next weight in the representation is

µ ~1 − α
~2 = µ ~ 1 = (0, 1, 0, 0, 0). (2.14.14)

~2 · α
Now, we note that µ ~ 2 = −1. Therefore, for this root and weight, q − p = 1. Hence, it
must be true that
µ3 = µ2 − α
~ 2 = (0, 0, 1, 0, 0) (2.14.15)
is a weight. In order for µ2 − 2~
α2 to be a weight, it would be necessary for µ2 + α
~ 2 to also
be a weight. But notice that

~ 2 = µ1 − (~
µ2 + α α1 − α
~ 2) (2.14.16)

so in order for µ2 + α
~ 2 to be a weight α~1 − α ~ 2 would have to be a root. But the difference
of simple roots is not a root. We can continue these arguments and generate the other
two weights
µ4 = (0, 0, 0, 1, 0) µ5 = (0, 0, 0, 0, 1) (2.14.17)
to fill out the representation.
As another example, let us consider the adjoint representation of SO(9). The weights
in the representation are the zero weight states corresponding to the elements of the
Cartan subalgebra, and the roots. The simple roots are given by

~ 1 = (1, −1, 0, 0),

α ~ 2 = (0, 1, −1, 0),
α ~ 3 = (0, 0, 1, −1),
α α
~ 4 = (0, 0, 0, 1). (2.14.18)

~ 0 = (1, 0, 0, 0) = −~
Consider the root α α1 − α
~2 − α
~3 − α
~ 4 . Its inner product with α
~ 4 is

~0 · α
α ~ 4 = 0. (2.14.19)

So we conclude that if α
~0 + α ~0 − α
~ 4 is a root, then so is α ~ 4 in order that p − q = 0. Both
combinations are indeed roots.

34
2.15 Representations for Subgroups
From the simple roots of a group we can determine what the subgroups are. For example,
the Dynkin diagram of SU (5) looks like

If we were to remove one of the simple roots, then we would be left with

These are the Dynkin diagrams for SU (2) and SU (3), so they both are subgroups of
SU (5). More to the point, they are both subgroups simultaneously, in other words there
is an SU (3) × SU (2) subgroup of SU (5) . Moreover, there is one linear combination of the
SU (5) Cartan subalgebra that commutes with both the SU (2) roots and the SU (3) roots.
Hence there is an additional U (1) subgroup. Hence, SU (5) has an SU (3) × SU (2) × U (1)
subgroup.
A subgroup is called semisimple if it has no U (1) factors, hence SU (3) × SU (2) is
semisimple, but SU (3) × SU (2) × U (1) is not semisimple. A semisimple subgroup is
called maximal if the rank of the subgroup is the same as the original group. Hence
SU (3) × SU (2) is not a maximal subgroup, since its rank is 3, but the rank of SU (5) is 4.
To find maximal subgroups, we need to consider an object known as the extended
Dynkin diagram. Recall that in our classification of the groups, we found many chains
that could not correspond to a Lie algebra since there was a linear relation between the
simple roots. However, we can use this to our advantage to find maximal subgroups. An
extended Dynkin diagram is found by taking the original Dynkin diagram with n simple
roots and adding one circle to the diagram such that the new diagram is connected and
has a linear relation between the n + 1 simple roots. We have basically determined what
these diagrams are already when we ruled out the chains in the previous sections. The
extended Dynkin diagrams for the various groups are shown in figure 12. The tildes over
the diagram labels indicate that these are the extended diagrams. If we now remove a
simple root from the extended diagram, we are left with a diagram that has the same
rank as the original Lie group. Furthermore, this is a subgroup of the original group since
it is made up of roots from the original Lie algebra. What is more, there is no longer a

35
 1

1 1 1 1 .... 1
~
A
n
1

2 2 2 2 .... 2
~
B n
1

1 2 2 2 .... 2 1
~
C n

1 1

2 2 2 .... 2
~
D
n
1 1

1 2 3
~
G 1 2 3 4 2
~
F
2 4
1

2
~
E
1 2 3 2 1 6

2
~
E
1 2 3 4 3 2 1 7

3
~
E
1 2 3 4 5 6 4 2 8

Figure 10: The extended Dynkin diagrams for the compact simple groups. The numbers
inside the circles indicate the coefficients for the linear relation.

36
linear relation between the simple roots, since we have removed one of the simple roots.
Hence this is a Dynkin diagram for a semi-simple Lie algebra.
Let us consider a few examples. First for SU (n), it is clear that when we perform
this surgery on the extended diagram we end up with the original SU (n) diagram, hence
SU (n) has no semi-simple maximal subgroups.
Next consider G2 . If we remove the short root from the extended diagram, we end up
with the SU (3) Dynkin diagram. Hence, SU (3) is a maximal subgroup of G2 .
Next consider E6 . If we consider the extended diagram, then if we remove the middle
simple root, we are left with three SU (3) diagrams. Hence E6 has an SU (3) × SU (3) ×
SU (3) maximal subgroup. But it also has another maximal subgroup. If we remove one
of the simple roots in the middle of one of the legs, then we are left with an SU (6)×SU (2)
diagram. If we remove one of the outer roots, then we get back the E6 diagram again.
Next consider E7 . If we remove the simple root second from the left we get SU (2) ×
SO(10). (We can also get SU (3) × SU (6), SU (3) × SU (3) × SU (2) and SU (8).)
The group E8 also has several maximal subgroups. Among others, it has an SU (2)×E7
subgroup, an SU (3) × E6 subgroup as well as an SO(16) subgroup.
It is also of interest to determine how the representations transform under the sub-
group. One thing that we should definitely expect is that a representation which is
irreducible under the original group will be reducible under the subgroup, since the group
is now smaller. With the smaller group, there are less transformations available and so
it becomes likely that not every element in the representation can be transformed into
every other element in the representation.
How the representations break up into the irreducible representations of the subgroup
is of intense interest in studying grand unified theories. There is a precise way of determing
what these irreducible representations are, but often we can figure out what they are by
simply trying to fit the representations of the subgroup into the representations of the
larger group.
As an example, consider the fundamental representation of G2 , which we write as
7. We would like to determine how this representation reduces under the SU (3) sub-
group. The most obvious reason why it must reduce is that SU (3) has no 7 dimensional
irreducible representation! Since the 7 is a real representation, then under the SU (3) de-
composition, if a complex representation appears, then its conjugate representation must
also appear. The first few irreducible representations of SU (3) are the singlet 1 (that is
the representation that is invariant under all SU (3) transformations), the 3, the 3 and

37
the adjoint 8. Given this, we see that under the subgroup, the 7 decomposes to

7 = 1 + 3 + 3. (2.15.1)

2.16 Real, pseudoreal and complex representations

Given the Lie Algebra in (2.6.4), it is simple to show that

−Ta∗ (2.16.1)

satisfies the same algebra, since the structure constants are real. Therefore, if the Ta are
a representation of the algebra, then so are the −Ta∗ . Clearly, the dimensions of the two
representations are the same, since the complex conjugate of an n × n matrix is also an
n × n matrix. But are they the same representation?
We can investigate this question as follows. Since the generators of the Cartan subal-
gebra Hi are hermitian, it follows that

(−Hi )∗ = −Hi . (2.16.2)

So one of these generators acting on a state of a given weight gives

(−Hi )∗ |~µi = −µi |~µi. (2.16.3)

Hence, if |~µi and | − µ~ i are both in the representation, then we do not get a new repre-
sentation by taking the transformation in (2.16.1). Representations which contain a state
|~µi but not | − µ
~ i are called complex. The representation containing | − µ ~ i is called the
conjugate of the representation containing |~µi. Representations containing both states
are called real representations.
Let us suppose that we have a real representation containing a state |~µi. Since | − µ ~i
is also in the representation, there must exist an invertible linear transformation where

R|~µi = | − µ
~ i. (2.16.4)

for all weight states in the representation. Thus we find

Hi R|~µi = −µi R|~µi ⇒ R−1 Hi R|~µi = −µi |~µi, (2.16.5)

and so
R−1 Hi R = −Hi∗ . (2.16.6)

38
Since the weights are in a representation of the Lie algebra, it must be true for all gener-
ators that
R−1 Ta R = −Ta∗ = −TaT (2.16.7)
where the last equality arises from the hermiticity of the generators.
But we also have that

Ta = −(R−1 Ta R)T = −RT TaT (R−1 )T (2.16.8)

and so
R−1 RT TaT (R−1 RT )−1 = TaT . (2.16.9)
This last equation means that
[R−1 RT , TaT ] = 0 (2.16.10)
for all generators. The only matrix that can commute with all generators in an irreducible
representation is one that is proportional to the identity. To see this, suppose that R−1 RT
is not proportional to the identity. Then its eigenvalues cannot all be equal. But since
it commutes with all Ta , it means that we can block diagonalize the representation such
that those states on the upper block have one eigenvalue of R−1 RT while those on the
lower block have another eigenvalue. But since R−1 RT commutes with Ta , the Ta cannot
mix the states in the different blocks. Hence the representation of the Lie algebra can be
block diagonalized, and hence it is reducible. Hence we conclude that

R−1 RT = λI (2.16.11)

where I is the identity matrix for the representation. Therefore we find that

RT = λR. (2.16.12)

Since (RT )T = R, we see that λ2 = 1. Thus, the possible eigenvalues for λ are

λ = ±1. (2.16.13)

Representations where λ = +1 are called positive real or sometimes just real, while repre-
sentations where λ = −1 are called pseudoreal.
A positive real representation is one where the hermitian generators are completely
imaginary. Therefore, a group element g = exp iθa Ta is made up of only real matrices. If
it is possible to express the generators as entirely imaginary, then there exists a unitary
transformation
Tea = U Ta U −1 , (2.16.14)

39
where Tea = −Tea∗ = −Te T . Taking the transpose of (2.16.14) we have

TeaT = (U −1 )T TaT U T = −U Ta U −1 . (2.16.15)

This then gives

Ta = −U −1 (U −1 )T TaT U T U = (U T U )−1 TaT U T U, (2.16.16)
which means that
R = U T U. (2.16.17)
But then,
RT = (U T U )T = U T U = R. (2.16.18)
Hence, the only representations that can have all real group elements are the positive
reals.

2.17 Group invariants

It is often useful to find a set of invariants under the group action, that is the action of the
group on its representations. For example, we recall from the study of angular momentum
in quantum mechanics that there was an invariant that commuted with all components
of the angular momentum operator, namely J~2 . We would like to find such invariants for
other groups as well.
To do this, recall that we defined the inner product in the adjoint representation to
be the trace of two generators
hTa |Tb i = Tr(Ta† Tb ) (2.17.1)
We can always choose an orthornormal basis such that the generators are hermitian and
orthonormal, in other words
Tr(Ta Tb ) = kδab , (2.17.2)
where k is the same constant as in (2.7.12). For example, we can choose this basis by
combining the root generators into the two separate terms
1 i
T~a,1 = √ [G~a + G−~a ] and T~a,2 = √ [G~a − G−~a ]. (2.17.3)
2 2
Both of these are hermitian and by using eqs. (2.7.21) and (2.7.22) one can easily show
that they satisfy
Tr(T~a,i T~a,j ) = δij k, (2.17.4)

40
 The advantage of this orthornormal basis is that the structure constants have a nice
form. Consider the trace over orthornormal generators

Tr(Ta [Tb , Tc ]) = Tr(Ta fbcd Td ) = kfbca = −kfcba . (2.17.5)

But using the cyclic properties of the trace, we can also show that

Tr(Ta [Tb , Tc ]) = Tr([Ta , Tb ]Tc ) = kfabc = −kfbac = Tr([Tc , Ta ]Tb ) = kfcab = −kfacb .
(2.17.6)
In other words, in this basis the structure constants are completely antisymmetric!
Now consider the sum over all generators in this basis
X
I= Ta Ta . (2.17.7)
a

This sum is an invariant. To show this, consider the commutator of any generator with
the sum (with repeated indices implying the sum)

[Tb , I] = [Tb , Ta ]Ta +Ta [Tb , Ta ] = fbad Td Ta +fbad Ta Td = −fbda Td Ta +fbad Ta Td = 0, (2.17.8)

where we explicitly used the property that fabc = −facb . Hence this sum is an invariant.
In the literature this is known as the quadratic Casimir
Just as in the case of SU (2), the quadratic Casimir will have different values for
different representations. A useful relation can be found between the quadratic casimir
and the trace of the square of any generator. These traces are repesentation dependent,
so let us explicitly put a subscript on the trace

Tr(Ta Tb ) = kR δab (2.17.9)

where R refers to the representation. If we trace the casimir, we find

Tr(I) = dR IR (2.17.10)

where dR is the dimension of the representation and IR is the eigenvalue of the casimir
for this representation. But we also have that
X
Tr(Ta Ta ) = kR = DkR , (2.17.11)
a

where D is the dimension of the algebra. Hence we see that

D
IR = kR . (2.17.12)
dR

41
One consequence of (2.17.12) is that IR = kR if R is the adjoint representation.
Up to now we have been carrying along this factor of kR that comes from the normal-
ization of the trace without giving a specific value to it. There is actually some freedom
in assigning it a value. Notice in (2.6.4) that we could multiply all Ta by a factor of λ
and still satisfy the Lie algebra, so long as we also rescale the structure constants by the
same factor. Rescaling Ta would then rescale kR by a factor of λ2 . However, once we fix
kR for one representation of the algebra, then we fix kR for all representations since the
structure constants are the same for all representations.
The standard choice for normalizing kR is to use
1
Tr(Ta Tb ) = δab (2.17.13)
2
for the fundamental representation of SU (2). Hence, we find that
D 31 3 1 1
 
If = kf = = = +1 (2.17.14)
df 22 4 2 2
which is the standard result we know from the quantization of angular momentum. With
this definition for SU (2), there is a natural way to define kf for the smallest fundamental
representation of SU (n). This representation is n dimensional. SU (n) has an SU (2)
subgroup, which means that we can find 3 generators of the SU (n) Lie algebra that form
an SU (2) Lie algebra amongst themselves. We can choose these generators to be
     
0 1 0 ... 0 0 −i 0 ... 0 ... 0 1 0 0
1 0 0 ... 0 i 0 0 ... 0 ... 0  0 −1 0
     
1 1 1
    
    
T1 =  0 0 0 ... 0
 T2 =  0 0 0 ...
 T3 =  0 ... 0  0 0 0
2
 
2 
2 
 .. ... ..

  .. ... ..

  .. 
 .. ...
0 0 0 0... 0 0 0 ...
0 ... 0 0 0 0
(2.17.15)
Hence we see that we have Pauli matrices in the upper left two by two block and zeroes
everywhere else. The group elements coming from these generators then look like
 
U11 U12 0 0 ... 0

 U21 U12 0 0 ... 0 

3
0 0 1 0 ... 0
X  
 
U = exp(i θa Ta ) =   (2.17.16)
a=1

 0 0 0 1 ... 0 

.. .. .. ..
 
 
0 0 0 0 ... 1
where the two by two matrix in the upper left block is a two dimensional unitary matrix
with determinant one and thus is a fundamental representation of SU (2). The 1’s along

42
the diagonal are also representations of SU (2), but they are trivial representations. Hence,
we see that the n dimensional irreducible representation of SU (n) is reducible under the
SU (2) subgroup to one fundamental representation and n − 2 trivial representations.
These trivial representations are sometimes called singlets.
Anyway, now we see that for the three generators of the subgroup, the trace is
1
Tr(Ta Tb ) = δab . (2.17.17)
2
But (2.17.9) holds for all generators of the algebra, so we must get (2.17.17) for all
generators of the SU (n) Lie algebra, and so kf = 1/2.
The quadratic casimir is the only invariant for SU (2) (this should come as no surprise
since we already know that J~2 is the only invariant of a representation). However, the
other groups have other invariants.

2.18 The Center of the Group

The center of a group is made up of all group elements that commute with every other
element in the group. The center is clearly a group. For SU (n), the center is Zn , as
should be clear, since the elements

diag(e2πim/n , e2πim/n , e2πim/n ...., e2πim/n ) (2.18.1)

are the identity multiplied by a constant and the determinant is 1. Of special interest is
the coset of the group by its center. The coset itself is still a group. The identification
that takes place is between group elements that are not arbitrarily close to each other.
This means for group elements constrained to be close to the identity, the group looks
the same. What we say is that the local structure of the group is the same, but its global
structure has changed.
Since the local structure is the same, the Lie algebra is the same. However, modding
out a group by its center could remove some representations. If a representation is not
invariant under the center, then it can’t be a representation of the coset. One representa-
tion that clearly survives is the adjoint representation, since the action by a group element
is via a commutator, so an element of the adjoint representation must be invariant under
the center since the center commutes with everything.
As an example, let us consider SU (2). This has a Z2 center. The integer spin represen-
tations are invariant under the center, but the half integer representations are not; they
change sign when acted upon by the nontrivial element. Hence the group SU (2)/Z2 only

43
has the integer spin representations, which obviously includes the adjoint. Now consider
SO(3), which has the same Dynkin diagram as SU (2), so locally these groups are the
same. But SO(3) has no center. (Note that diag(−1, −1, −1) is not in SO(3) since it
has determinant −1. This is in the group O(3).) In fact SO(3) is the same as SU (2)/Z2 ,
which obviously has no center since it was taken out. SO(3) is the group corresponding to
the orbital angular momentum, which of course only has integer representations. We also
see that the group manifold for SO(3) is S3 /Z2 . This will have interesting consequences.

44
3 Path Integrals, Wilson Lines and Gauge Fields
We have spent the first few lectures learning about Lie Groups. Lie Groups are of partic-
ular importance for the study of gauge theories. However, before we can put this to good
use, we need to study something called a path integral.

3.1 Path integrals for nonrelativistic particles

In quantum mechanics we often want to compute the probability of finding particles with,
say, particular positions and momentums. One particular problem is this: given that a
particle is at position ~x0 at time t0 , what is the probability that the particle is at position
~x1 at time t1 ? To find the probability, we need to solve the time dependent Schroedinger
equation
∂
ih̄ Ψ(~x, t) = HΨ(~x, t) (3.1.1)
∂t
where H is the Hamiltonian
p~2
H= + V (~x). (3.1.2)
2m
At time t0 , the wave function is 0 unless ~x = ~x0 . (We say that the wave function has
δ function support). Then, as time elapses, the wave function evolves. At time t = t1 , we
then compute the probability amplitude for the wavefunction at ~x = ~x1 . Formally, the
solution to the Schroedinger equation is
i
Ψ(~x, t) = e− h̄ H(t−t0 ) Ψ(~x, t0 ). (3.1.3)
When we first learn quantum mechanics, we almost always assume that Ψ is an eigenfunc-
tion of the Hamiltonian. However, a wave function with δ function support is certainly
not an eigenfunction of the Hamiltonian, since the position operator does not commute
with the momentum operator.
Thus to treat this problem, we do the following. Let us use bra-ket notation. Consider
the ket state |~x0 , t0 i, where we have explicitly inserted a time variable to indicate that
this is the state at time t = t0 . The probability amplitude is then found by computing
the inner product
h~x1 , t1 |~x0 , t0 i. (3.1.4)
This is a formal way of saying that in order to compute the inner product, we have to let
the state evolve over time. But this we can find from (3.1.3), and the inner product in
(3.1.4) can be explicitly written as
i
h~x1 , t1 |~x0 , t0 i = h~x1 |e− h̄ H(t1 −t0 ) |~x0 i. (3.1.5)

45
 To evaluate the expression in (3.1.5), we split up the time interval between t0 and t1
into a large number of infinitesimally small time intervals ∆t , so that
i i
e− h̄ H(t1 −t0 ) = e− h̄ H∆t ,
Y
(3.1.6)

where the product is over all time intervals between t0 and t1 . Since, p~ does not commute
with ~x, we see that
i i 2 i
e− h̄ H∆t 6= e− 2mh̄ p~ ∆t e− h̄ V (~x)∆t , (3.1.7)
however, if ∆t is very small, then it is approximately true, in that
i i 2 ∆t i
e− h̄ H∆t = e− 2mh̄ p~ e− h̄ V (~x)∆t + O((∆t)2 ). (3.1.8)

Hence, in the limit that ∆t → 0, we have that

lim i 2 i
h~x1 | e− 2mh̄ p~ ∆t e− h̄ V (~x)∆t |~x0 i.
Y
h~x1 , t1 |~x0 , t0 i = (3.1.9)
∆t → 0
In order to evaluate the expression in (3.1.9), we need to insert a complete set of
states between each term in the product, the states being either position or momentum
eigenstates and normalized so that
Z
d3 x|~xih~x| = 1 h~x1 |~x0 i = δ 3 (~x1 − ~x0 )
Z
d3 p
|~pih~p| = 1 h~p1 |~p0 i = (2πh̄)3 δ 3 (~p1 − p~0 )
(2πh̄)3
h~x|~pi = ei~x·~p/h̄ (3.1.10)

Inserting the position states first, we have that (3.1.9) is

 
Z
i
− 2mh̄ p|2 ∆t
|~ − h̄i V
d3 x(t)  (~
x)∆t
Y Y
h~x1 , t1 |~x0 , t0 i = h~x(t + ∆t)|e e |~x(t)i , (3.1.11)
t0 <t<t1 t0 ≤t<t1

where the first product is over each time between t0 + ∆t and t1 − ∆t and the second
product is over each time between t0 and t1 − ∆t. We also have that ~x(t0 ) = ~x0 and
~x(t1 ) = ~x1 . Note that for each time interval, we have a position variable that we integrate
over. You should think of the time variable in these integrals as a label for the different
x variables.
We now need to insert a complete set of momentum states at each time t. In particular,
we have that

46
 i 2 i
h~x(t + ∆t)|e− 2mh̄ |~p| ∆t e− h̄ V (~x)∆t |~x(t)i
Z
d3 p i 2 i
= 3
h~x(t + ∆t)|~pie− 2mh̄ |~p| ∆t h~p|~x(t)ie− h̄ V (~x(t))∆t
(2πh̄)
Z
d3 p i∆~x(t)·~p/h̄ − i |~p|2 ∆t − i V (~x(t))∆t
= e e 2mh̄ e h̄ (3.1.12)
(2πh̄)3
where ∆x(t) = x(t + ∆t) − x(t). We can now do the gaussian integral in the last line in
(3.1.12), which after completing the square gives
  !2  
3/2
m i m ∆~x

i
− 2mh̄ p|2 ∆t
|~ − h̄i V (~
x)∆t
h~x(t+∆t)|e e |~x(t)i = exp   − V (~x(t)) ∆t .
2πih̄∆t h̄ 2 ∆t
(3.1.13)
Strictly speaking, the integral in the last line of (3.1.12) is not a Gaussian, since the
coefficient in front of p~2 is imaginary. However, we can regulate this by assuming that the
coefficient has a small negative real part and then let this part go to zero after doing the
integral. The last term in (3.1.13) can be written as
3/2 3/2
m i m ˙2 m i
       
= exp ~x − V (~x(t)) ∆t = exp L(t)∆t , (3.1.14)
2πih̄∆t h̄ 2 2πih̄∆t h̄
where L(t) is the lagrangian of the particle evaluated at time t. Hence the complete path
integral can be written as
3N/2 Z
m i
  
d3 x(t) exp
Y
h~x1 , t1 |~x0 , t0 i = S , (3.1.15)
2πih̄∆t t0 <t<t1 h̄

where S is the action and is given by

Z t1
S= L(t)dt. (3.1.16)
t0

N counts the number of time intervals in the path. In the limit that N goes to infinity,
we see that the constant in front of the expression diverges. It is standard practice to
drop the constant, which is essentially a normalization constant. It can always be brought
back later after normalization.
The expression in (3.1.15) was first derived by Feynman and it gives a very intuitive
way of looking at quantum mechanics. What the expression is telling us is that to compute
the probability amplitude, we need to sum over all possible paths that the particle can
i
take in getting from x0 to x1 , weighted by e h̄ S .

47
 It is natural to ask which path dominates the path integral. Since the argument of the
exponential is purely imaginary, we see that the path integral is a sum over phases. In
general, when integrating over the x(t), the phase varies and the phases coming from the
different paths tend to cancel each other out. What is needed is a path where varying to
a nearby path gives no phase change. Then the phases add constructively and we are left
with a large contribution to the path integral from the path and its nearby neighbors.
Hence, we look for the path, given by a parameterization x(t), such that x(t0 ) = x0
and x(t1 ) = x1 , and such that the nearby paths have the same phase, or close to the same
phase. This means that if x(t) is shifted to x(t) + δx(t), then the change to the action is
very small. To find this path, note that under the shift, to lowest order in δx, the action
changes to
Z t1 " # Z t1 " ! #
∂L ˙ ∂L d ∂L ∂L
S→S+ dt δ~x + δ~x = S + dt − + δ~x. (3.1.17)
t0 ∂ ~x˙ ∂~x t0 dt ∂ ~x˙ ∂~x
Hence there would be no phase change to lowest order in δx if the term inside the square
brackets is zero. But this is just the classical equation of motion! A generic path has a
phase change of order δx, but the classical path has a phase change of order δx2 .
Next consider what happens as h̄ → 0. Then a small change in the action can lead to
a big change in the phase. In fact, even a very small change in the action essentially wipes
out any contribution to the path integral. In this case, the classical path is essentially
the only contribution to the path integral. For nonzero h̄, while the classical path is the
dominant contributor to the path integral, the nonclassical paths also contribute, since
the phase is finite.

3.2 Path integrals for charged particles

Now suppose that we have a charged particle in an electromagnetic field. Recall that we
can write the electric field as

~ d ~ ~ x, t) = −∇
~ × A(~
~ x, t)
E(x) = −∇φ(~x, t) + A(~x, t) B(~ (3.2.1)
dt
We can write these fields as a 4-tensor

Fµν = ∂µ Aν − ∂ν Aµ , A0 = φ. (3.2.2)

with F0i = Ei and Fij = −ijk Bk . A charged particle in such a field feels the Lorentz force

~ + e~v × B.
F = eE ~ (3.2.3)

48
To get such a term in the equation of motion for the particle, we can add to the action
the term Z t1 h i
Sem = −e dt φ + ~x˙ · A
~ , (3.2.4)
t0

so that varying with respect to ~x(t) gives

Z t1  
Sem → Sem − e dt ∇φ ~˙ + ẋi ∇A
~ −A ~ i − ẋi ∂i A
~ · δ~x(t), (3.2.5)
t0

which is the desired result to get the correct force. By the chain rule, we can rewrite the
action in (3.2.4) in terms of the line integral
Z t1
Sem = −e Aµ dxµ , (3.2.6)
t0

which is explicitly Lorentz covariant. Therefore, in the presence of the electromagnetic

field, our path integral for a charged particle is modified to
i ie Z t1
Z  
3
Aµ dxµ ,
Y
h~x1 , t1 |~x0 , t0 i = d x(t) exp S− (3.2.7)
t0 <t<t1 h̄ h̄ t0

The gauge field Aµ has a redundancy. Notice that if we transform Aµ to

Aµ → Aµ + ∂µ φ (3.2.8)

where φ is any function, then the electromagnetic fields are unchanged. Notice that this
transformation is local – φ can take on different values at different space-time points.
Since the fields are unchanged, it must be true that the physics for the charged particle
does not change under the gauge transformation. If we look at the action in (3.2.6), we
see that under the gauge transformation the change in the action is
Z t1
δSem = −e ∂µ φdxµ = −e(φ(~x1 , t1 ) − φ(~x0 , t0 )). (3.2.9)
t0

Hence the path integral in (3.2.7) transforms to

h~x1 , t1 |~x0 , t0 i → e−ieφ(~x1 ,t1 )/h̄ h~x1 , t1 |~x0 , t0 ie+ieφ(~x0 ,t0 )/h̄ . (3.2.10)

In other words, the path integral only changes by a phase and so the physics does not
change, since the amplitude squared is unchanged.

49
3.3 Wilson lines
Not only is a charged particle acted on by external electromagnetic fields, it is also a
source for those same fields. If we assume that the charge particle is very heavy, then it
is not influenced by the electromagnetic fields and in fact can be restricted to a specific
path. So we could instead consider
ie Z x1
 
W (x1 , x0 ) = exp − Aµ dxµ , (3.3.1)
h̄ x0
over a fixed path from x0 to x1 . In quantum field theory, we will have path integrals
where we integrate over the fields Aµ , but let us not worry about that. The expression in
(3.3.1) is called a Wilson line. As we already saw, the Wilson line is not exactly gauge
invariant, it instead changes to

W → e−ieφ(~x1 ,t1 )/h̄ W e+ieφ(~x0 ,t0 )/h̄ (3.3.2)

An object which transforms like this is said to be gauge covariant. Now we notice some-
thing about how W transforms. Its endpoints are transformed under two distinct U (1)
transformations. That is, a gauge transformation transforms the endpoints of the Wilson
line by U (1) group elements. Since W transforms this way, we say that electromagnetism
is a U (1) gauge theory. The fact that the U (1) transformations are different at the differ-
ent endpoints reflects the fact that the gauge transformations are local. We also notice
that shifting φ by 2πh̄/e leaves the endpoints invariant. Hence the gauge transformations
are such that φ is identified with φ + 2πh̄/e. Because of the explicit h̄ dependence, we see
that this identification is a quantum effect.
We can also find an object that is actually gauge invariant, namely the Wilson loop,
which has the form
ie I
 
W = exp − Aµ dxµ . (3.3.3)
h̄
Now there is no endpoint, so W must be invariant. Another way to think of this is that
the endpoints in (3.3.1) are the same, so the two U (1) factors cancel off with each other.
There is also a discrete way of looking at a Wilson line, which will be useful for us
when we consider generalizations to other gauge groups. Suppose that we consider a very
short and straight Wilson line that traverses from point x to x + ∆x. The Wilson line
may then be approximated by

W (x + ∆x, x) = exp (−ieAµ ∆xµ /h̄) . (3.3.4)

50
W (x, x + ∆x) is called the link variable, or just link, between x and x + ∆x. It is clearly
a unitary matrix. Under gauge transformations, the link transforms as

U † (x + ∆x)W (x + ∆x, x)U (x). (3.3.5)

A Wilson line is then given as a product of connected links. The important point here is
that we can basically forget the gauge field Aµ and just worry about the link W .
Now we can see how to generalize a Wilson line that transforms under a U (1) gauge
transformation to one that transforms under some other Lie group. Let us suppose that
we have link variables that take their values in a representation for some Lie group. Let
us now consider a product of attached links

W (x1 , x0 ) = W (x1 , x1 − ∆xN ) . . . W (x0 + ∆x1 + ∆x2 , x0 + ∆x1 )W (x0 + ∆x1 , x0 ) (3.3.6)

Under gauge transformations the links transform as

W (x + ∆x, x) → U † (x + ∆x)W (x + ∆x, x, )U (x) (3.3.7)

hence the Wilson line transforms as

W (x1 , x0 ) = U † (x1 )W (x1 , x0 )U (x) (3.3.8)

For a closed Wilson loop, we would find that

W (x, x) → U † (x)W (x, x)U (x), (3.3.9)

hence the Wilson loop for a nonabelian gauge group is not actually gauge invariant, but
only gauge covariant. A gauge invariant object would be
 
Tr(W (x, x)) → Tr U † (x)W (x, x)U (x) = Tr(W (x, x)). (3.3.10)

3.4 The magnetic monopole

As an application let us consider a magnetic monopole. This section will serve as a prelude
to our discussion on fiber bundles later in the course.
A monopole will have a B ~ field pointing in the radial direction. Let us suppose that
the monopole charge is g, then the B ~ field is given by

~ = g
B r̂ (3.4.1)
4πr2

51
which insures that the flux through a sphere surrounding the monopole is g (I am choosing
units where µ,  and c are all 1.)
Z Z
g= ~ ·B
dS ~ = sin θdθdφr2 Br , (3.4.2)
S2

where dS ~ is the differential on the surface of S2 .

However, there is a problem. Recall that B ~ was given as −∇
~ × A.
~ That means that
the first integral in (3.4.2) is given by
Z Z
~ ·B
dS ~ =− ~ ·∇
dS ~ × A.
~ (3.4.3)
S2 S2

Using Stoke’s theorem, which says that

Z   Z
~· ∇
dS ~ ×A
~ = ~
d~x · A, (3.4.4)
Σ ∂Σ

where Σ is a 2 dimensional surface and ∂Σ is its boundary, we see that

Z
~ ·B
dS ~ = 0, (3.4.5)
S2

since S2 has no boundary.

What went wrong? Our mistake was assuming that A ~ is globally defined over the entire
S2 . Instead, let us break up the sphere into two hemispheres, with A ~ N on the northern
hemisphere which includes the equator and A ~ S on the southern hemisphere which also
includes the equator. Hence, on the equator the gauge fields A ~ N and A~ S overlap. Naively,
we would say that the two fields have to equal each other on the equator, but this is
too stringent a condition. Instead, we only require that they are equal up to a gauge
transformation. Therefore, let us assume that A ~N = A~ S + ∇φ.
~ Now, Stokes law tells us
that Z I
~ ~ ~
dS · B = − d~x · ∇φ, (3.4.6)
S2

where the line integral is around the equator. Now normally, we should expect the integral
to still be zero, since the rhs in (3.4.6) is a total derivative around a loop. But remember
from the previous section that φ is identified with φ + 2πh̄/e, so φ only needs to come
back to itself up to a multiple of 2πh̄/e when going around the equator. Therefore we
find that the magnetic charge can be nonzero and is given by
2πh̄
g=n , (3.4.7)
e
where n is an integer. Hence we find that the magnetic charge is quantized!

52
3.5 Gauge fields for nonabelian Lie groups
We saw previously that the field strength for a U (1) gauge group is given by

Fµν = ∂µ Aν − ∂ν Aµ . (3.5.1)

In this section we construct the corresponding object for a gauge field which has gauge
transformations in a nonabelian gauge group.
Recall that the gauge transformation for Aµ is

h̄
Aµ → Aµ + ∂µ φ = Aµ − i U † ∂µ U, (3.5.2)
e
where we have written the gauge transformation in terms of the U (1) group element
U = eieφ/h̄ . For the case of a nonabelian gauge group, let us consider the short Wilson
line W (x + ∆x, x). This is some unitary matrix, so we can assume that it can be written
as
e a a µ
 
W (x + ∆x, x) = exp −i Aµ T ∆x (3.5.3)
h̄
where the Ta are the generators of the group. We will write

Aµ = Aaµ T a , (3.5.4)

so now Aµ is a hermitian matrix.

To find the gauge transformation of Aµ , note that
e e
   
exp −i Aµ ∆xµ → U † (x + ∆x) exp −i Aµ ∆xµ U (x) (3.5.5)
h̄ h̄
Hence we have that
h̄  † 
Aµ ∆xµ → U † (x + ∆)Aµ U (x)∆xµ + i U (x + ∆x)U (x) − 1 . (3.5.6)
e
Taking the limit that ∆x → 0, we have
h̄
Aµ → U † (x)Aµ U (x) − i U † (x)∂µ U (x) , (3.5.7)
e
where we used the fact that

U † (x + ∆x) = U † (x) + ∆xµ ∂µ U † (x) (3.5.8)

and ∂U † = −U † ∂U U † .

53
 The field strength should at least be gauge covariant. The naive extension of the U (1)
case, ∂µ Aν − ∂ν Aµ does not satisfy this requirement. However, consider the quantity
e
Fµν = ∂µ Aν − ∂ν Aµ − i [Aµ , Aν ] (3.5.9)
h̄
Then under a gauge transformation, this becomes
e
 
†
U ∂µ Aν − ∂ν Aµ − i [Aµ , Aν ] U
h̄
h̄
−i [U † ∂µ U, U † ∂ν U ] + [U † ∂µ U, U † Aν U ] − [U † ∂ν U, U † Aµ U ]
e
h̄ †
+i [U ∂µ U, U † ∂ν U ] − [U † ∂µ U, U † Aν U ] + [U † ∂ν U, U † Aµ U ]
e 
e

†
= U ∂µ Aν − ∂ν Am u − i [Aµ , Aν ] U (3.5.10)
h̄
Therefore, Fµν is the covariant field strength. We will soon see an easier way of deriving
this.

3.6 Gauge field actions

Maxwell’s equations for the U (1) gauge field with no sources (that is no charges or cur-
rents) can be written as
∂µ F µν = 0. (3.6.1)
Notice that the indices in Fµν have been raised. This is to guarantee that the expression
is Lorentz covariant. (Recall from special relativity that an index is raised using the
metric tensor, which in flat space is ηµν = diag(−1, 1, 1, 1).) The repeated indices have
an implied sum. Actually, (3.6.1) gives only half the equations, namely

~ ·E
∇ ~ = 0
~ −∇
∂t E ~ ×B~ = 0. (3.6.2)

The other two equations are guaranteed by the form of Fµν in (3.5.1). Namely, notice
that the equation
∂λ Fµν + ∂ν Fλµ + ∂µ Fνλ = 0 (3.6.3)
is automatic given (3.5.1). The equation in (3.6.3) is an example of a Bianchi identity.
Using (3.2.1), we can rewrite (3.6.3) as

∇ ~+ ∂B
~ ×E ~ = 0, ~ ·B
∇ ~ = 0. (3.6.4)
∂t
54
The first equation is Faraday’s Law and the second assumes that there are no magnetic
monopoles.
We wish to find a lagrangian that leads to the equations of motion in (3.6.1). Actually,
instead of a lagrangian, we want a lagrangian density, where the lagrangian L is given by
Z
L= d3 xL. (3.6.5)

Let us show that

1
L = − Fµν F µν (3.6.6)
4
does the job. Given L, we see that the action is
1Z 4
S=− d xFµν F µν , (3.6.7)
4
where d4 x is the measure for the integral over the 4 dimensional space-time. Notice that
since Fµν is gauge invariant, the action is also gauge invariant.
To find the equations of motion, we should vary the fields in the action and by doing
so, find a field configuration where the action is minimized. Put differently, we want to
find a configuration of Aµ (xσ ) such that the action is unchanged to lowest order when
the fields are changed by δAµ (xσ ) for all points in space-time. Formally, we wish to take
derivatives with respect to the Aµ (xσ ) fields, just as in section (3.1) we took derivatives
with respect to ~x(t). For derivatives with respect to xi (t), we use the fact that

∂xi (t0 )
j
= δ(t − t0 )δij , (3.6.8)
∂x (t)
in other words, the variations at one time are independent of the variations at a different
time. In the case of the fields, we have that the variations at a point in space-time are
independent of the variations at a different point in space-time. Hence we have,
∂Aµ (xσ )
σ
= δ 4 (xσ − y σ )δ ν µ . (3.6.9)
∂Aν (y )

Consider then the variation of the action in (3.6.7). The variation of L(y σ ) is

∂L(y σ ) 1
 
σ
= 4 − F µν (y σ )∂µ δ 4 (xσ − y σ ), (3.6.10)
∂Aν (x ) 4
and so the variation of the action is
∂S Z
σ
= − d4 yF µν (y σ )∂µ δ 4 (xσ − y σ ) = ∂ν F µν (xσ ) = 0, (3.6.11)
∂Aµ (x )

55
where we did an integration by parts to get the final form. This is the equation of motion
for the gauge field.
We can see that the funny factor of −1/4 in (3.6.6) is chosen so that the end result is
(3.5.1). The lagrangian can also be written as
1~ ~ 1~ ~
L= E · E − B · B, (3.6.12)
2 2
~ serves as the canonical momentum for A.
where E ~ The Hamiltonian is then
1~ ~ 1~ ~
H= E · E + B · B, (3.6.13)
2 2
which is the expected energy density for an electromagnetic field.
Now let us assume that there is also a charged particle around that can influence the
electromagnetic fields. Hence we want to add to the action in (3.6.7) the action in (3.2.6).
However, we have to be a little careful since the action in (3.6.7) has an integration over 4
dimensions but the integral in (3.2.6) is a one dimensional line integral. Anyway, varying
the sum of the actions we find
∂(S + Sem ) µν
Z
= ∂ ν F − e dy ν δ 4 (xσ − y σ ) = 0, (3.6.14)
∂Aµ (xσ )
and so the modified equation of motion is
Z
∂ν F µν = e dy ν δ 4 (xσ − y σ ). (3.6.15)

To understand this equation, let us look at the various components. For example, consider
the ν = 0 component equation

~ · E = eδ 3 (xi (t) − y i (t)).

∂ν F 0ν = ∇ (3.6.16)

y i (t) is the position of the particle at time t. Hence the term on the right hand side of
this equation is the charge density, and so this is Gauss’ law in the presence of a charged
particle. If we look at one of the spatial components, then the equation is

~ × B)
~ i=e dy i 3 i
∂ν F iν = ∂t E i − (∇ δ (x (t) − y i (t)). (3.6.17)
dt
The right hand side is now the current density for the moving charged particle and so
our modified equation of motion is Ampere’s law in the presence of a current. From this
discussion, we see that not only does Sem tell us how an electromagnetic field affects a
charged particle, but it also tells us how a charged particle affects the field.

56
 We now wish to generalize the lagrangian in (3.6.6) to nonabelian gauge fields. We saw
in the previous section that an appropriate choice for the field strength is the expresssion
in (3.5.9). It is natural to choose the lagrangian to be − 14 Fµν F µν . There are two slight
problems with this. First, the Lagrangian should be a number, but this is a matrix.
Second, this expression is not gauge invariant, but only gauge covariant. Given the
transformation properties in (3.5.10), we see that

Fµν F µν → U † Fµν F µν U, (3.6.18)

so this is only covariant. To find an invariant expression, we can just take the trace. This
also gives a lagrangian that is a number and not a matrix. Hence, the Lagrangian is
1
L=− Tr[Fµν F µν ]. (3.6.19)
4k
We have the same factor of − 41 , so that the kinetic terms of the U (1) subgroups, that
is those groups generated by the Cartan subalgebra, have the same normalization as the
U (1) gauge fields of electromagnetism. The factor of 1/k in (3.6.19) cancels the factor of
k from (2.7.12).

3.7 The covariant derivative

There is a convenient way to write the field strength that nicely generalizes to the non-
abelian case. Define the covariant derivative Dµ to be
e
Dµ = ∂µ − i Aµ . (3.7.1)
h̄
Next consider the commutator [Dµ , Dν ], which is
e e e
 
[Dµ , Dν ] = −i ∂µ Aν − ∂ν Aµ − i [Aµ , Aν ] = −i Fµν (3.7.2)
h̄ h̄ h̄
Hence, the field strength is proportional to the commutator of two covariant derivatives!
It is also easy to check that the covariant derivative transforms covariantly (hence its
name) under gauge transformations. Under the gauge transformation in (3.5.7), we have
e e
 
Dµ → ∂µ − i U † Aµ U − U † ∂µ U = U † ∂µ − i Aµ U = U † Dµ U , . (3.7.3)
h̄ h̄
From this, the gauge covariance of Fµν automatically follows.

57
 The equations of motion can be written simply using the covariant derivative. Consider
a small transformation of Aµ by δAµ . Therefore, the leading order change to Fµν is
e e
δFµν = ∂µ δAν − ∂ν δAµ − i [Aµ , δAν ] − i [δAµ , Aν ] = [Dµ , δAν ] − [Dν , δAµ ]. (3.7.4)
h̄ h̄
Therefore,
1 Z 4 e e
  
δS = − d x 2Tr ∂µ δAν − ∂ν δAµ − i [Aµ , δAν ] − i [δAµ , Aν ] F µν
4k h̄ h̄
1Z 4 e
 
= − d xTr (∂µ δAν − i [Aµ , δAν ])F µν
k h̄
1Z 4 e
  
µν µν
= d xTr δAν ∂µ F − i [Aµ , F ]
k h̄
1Z 4
= d xTr [δAν [Dµ , F µν ])] . (3.7.5)
k
In going from line 2 to line 3 in (3.7.5) we integrated by parts and used the cyclic properties
of the trace. Hence, the equations of motion are

[Dµ , F µν ] = 0. (3.7.6)

Notice that this is a matrix equation and hence must be true for all components of the
matrix. Notice also that the equation of motion is gauge covariant, and so the solutions
to the equation of motion remain solutions after gauge transformations.

58
4 Topology and Differential Geometry
In this chapter, we introduce the concepts of topology, differential topology and the notion
of a fiber bundle. We will see that fiber bundles are particularly useful in studying the
topics of the last chapter.

4.1 Rudimentary Topology

The study of topology is a fascinating subject and is rather fun. Most of you probably
have some idea of topology, at least in the sense that you know it when you see it. You
are probably aware that a doughnut is topologically distinct from a ball, but not from a
coffee cup with a handle. The doughnut shape can be smoothly deformed into the shape
of a coffee cup, but not into a ball, since there would be no way to smoothly deform
away the hole in the doughnut. When dealing with topology, we often speak of topological
invariants. In this case, one of the topological invariants is the number of holes in the
surface.
In this example it was rather simple to figure out a topological invariant, but to go
further, we need to make a precise definition of what we mean by a topology and its
invariants.
To this end let X be a set and let Y = {Xi } be a collection of subsets of X. Then X
is a topological space and Y is a topology of X if
• ∅ ∈ Y , and X ∈ Y where ∅ is the null set.
• For any finite or infinite collection of sets Xi ∈ Y , Xi ∈ Y .
S

• For any finite collection of sets Xi ∈ Y , Xi ∈ Y .

T

The subsets Xi are called the open sets.

Now here are a few other definitions.
• A neighborhood N of x, is a subset of X, containing an open set Xi in the topology of
X where x ∈ Xi . Notice that N does not necessarily have to be an open set.
•Closed Sets. A closed set is a set whose complement in X is an open set. Since X and
∅ are each open sets and are complements of each other, we see that these are both open
and closed.
•Closure of a Set. Consider a set U and consider all closed sets that contain U . Then the
closure of U , U is the intersection of all those closed sets.
• Interior. The interior of a set U , U I , is the union of all open sets contained by U .
• Boundary. The boundary of a set U , b(U ) is the closure of U with the interior removed.
b(U ) = U − U I

59
• Open cover. An open cover of a set U is a collection of open sets Xi , such that U ⊂ Xi .
S

With this last definition, we can now give a precise definition of compactness. A set
U is compact, if every open cover of the set has a finite subcovering. What this means is
that even if we have an infinite collection of open sets covering the set, we can pick out a
finite number of these same open sets and still cover the set.
A set X is connected if it cannot be written as X = X1 X2 , where X1 X2 = ∅.
S T

4.2 Examples
These definitions seem rather abstract, so let us consider several examples. First the
trivial topology of a space X is where the only open sets are X and ∅. The discrete
topology is where Y contains all subsets of X. This will be the last time we consider such
things.
Usual Topology of R. Consider the real numbers R. Define the open sets to be all
open intervals (a, b), and there unions, where a < b and a and b are not included in the
interval. This then satisfies the properties of a topology over R. Note that if we had
allowed infinite intersections to be open sets, then individual points in R would have
been open sets, at which point we would get the discrete topology for R. We can also
see that this space is not compact, since we can find open coverings that do not have a
finite subcover. For example, we can cover R with the open sets (n, n + 2) for all integer
n. But no finite collection of these open sets covers R. We can also see that the closure
of an open set (a, b) is given by [a, b], the line interval including the endpoints. Note that
the interior of [a, b] is (a, b), while the interior of (a, b) is itself.
The open interval. The open interval (a, b), a < b is also an example of a topological
space. For the usual topology the open sets are all open intervals (ai , bi ) and their unions,
with ai ≥ a, bi ≤ b, ai < bi . These then satisfy the required union and intersection
properties. Like R, this is not a compact space, since there exists open coverings without
a finite subcover. For instance, the union of the open intervals (a + (b − a)/n, b) for all
n positive integer is a cover for (a, b), but we cannot cover (a, b) with a finite number
of these open sets. Note further that finite covers of (a, b) exist. But the criterion for
compactness is that every cover has a finite subcover. Notice that the closed interval [a, b]
is not a topological space for the usual topology, since [a, b] is not an open set (however,
it is compact).
Rn . We can also see that Rn is a topological space in more or less the same way that
R is a topological space. Rn is an example of a metric space, in that the space comes

60
with a distance function
d(x, y) ≥ 0 (4.2.1)
and is 0 only if x = y. It also satisfies a triangle inequality

d(x, z) ≤ d(x, y) + d(y, z). (4.2.2)

Instead of open intervals we consider open balls about all points x, Bx (), y ∈ Bx () if
d(x, y) < . All open balls and their unions forms the topology.
CP (1). Consider the complex projective space, where (z1 , z2 ) ≡ (λz1 , λz2 ) with λ 6= 0
and z1 and z2 not both zero. This is a topological space with the topology given by a
collection of open two dimensional surfaces. However, this space is compact. To see this,
note that all but one point on CP (1) is equivalent to (z, 1) for some z. We can consider
an infinite union of subsets, say the disks where |z| < n, where n is a positive integer.
This has no finite subcovering. On the other hand, this is not a covering of CP (1),
because it misses the point equivalent to (1, 0). We can cover this last point with the
open set containing the points (1, w), where |w| < . But now this covering has a finite
subcover, since we can cover CP (1) with this open set and the previous sets of disks with
|z| < 1/ + 1. This is clearly a finite subcover.

4.3 Homeomorphisms and equivalence classes

Consider two topological spaces X1 and X2 . A homeomorphism is a continuous map

f : X 1 → X2 (4.3.1)

whose inverse map, f −1 is also continuous. If a homeomorphism exists between two

topological spaces, then we say that the two spaces are homeomorphic. Another term to
describe these two spaces is to say that they belong to the same equivalence class.
A topological invariant is some characteristic of a topological space that is invariant
under homeomorphisms. Hence, two spaces with different topological invariants are not
of the same equivalence class.
Some examples of topological invariants are compactness and connectedness. Suppose
that X is compact and that f (X) = Y . If f is a homeomorphism, then f −1 (Y ) = X, and
so the open sets of Y map to the open sets of X. Hence an open cover of Y maps to an
open cover of X. The open cover of X has a finite subcover, which then maps back to a
finite subcover of Y . Connectedness is also easy to show.

61
 One can also show that the dimension of the space is a topological invariant. For
example, let us consider Rn . Start with R and assume that R is homeomorphic to R2 .
That is, there is an invertible map that takes R to R2 . If such a map exists, then it
maps a point in R to a point in R2 and the inverse maps this point back to the original
point. Let us remove a point p from R and remove the point f (p) from R2 . Then if
R is homeomorphic to R2 , then R − {p} is homeomorphic to R2 − {f (p)}. However,
R − {p} is not connected but R2 − {f (p)} is connected. Hence these two spaces cannot
be homeomorphic. Therefore, we have a contradiction, so R is not homeomorphic to
R2 . We can keep continuing this argument for the higher Rn , showing that Rm is not
homeomorphic to Rn if m 6= n.

4.4 Differential Manifolds

Let us try and make the discussion of topological spaces even more concrete. To this end,
let us define a differential manifold M.
Let M be a topological space which has the open cover
[
M⊂ Mα (4.4.1)

The open sets Mα are homeomorphic to open subsets Oα of Rn , through the invertable
map
φα : Mα → Oα . (4.4.2)
In the intersection region where Mα ∩ Mβ 6= ∅, we have the map

φβ ◦ φ−1
α : φα (Mα ∩ Mβ ) → φβ (Mα ∩ Mβ ). (4.4.3)

If this map is infinitely differentiable (also known as C ∞ ), then M is a differentiable

manifold. The dimension of the manifold is given by n.
Examples Let us consider 1 dimensional connected manifolds. The only two examples
are the real line R and the circle S1 . Clearly, R is a manifold. To show that S1 is a
manifold, let us write the circle in terms of its x, y coordinates (cos θ, sin θ). The maps to
the open subsets of R map the (x, y) coordinate to θ. We can cover the circle with the
two open subsets, 0 < θ < 3π/2 and π < θ < 5π/2. Then the maps φ2 ◦ φ−1 1 , map θ → θ
in the first intersection region and θ → θ + 2π in the second intersection region. Clearly
these maps are C ∞ , therefore, S1 is a manifold.

62
4.5 Differential forms on manifolds
Since φα maps Mα into Rn , the function φα (p) where p is a point on M, has n components.
Hence we can write this as

φα (p) = (x1α (p), x2α (p), ...xnα (p)) (4.5.1)

The components xiα are called the local coordinates on M. The open region Mα is called
the coordinate patch for these local coordinates. With these local coordinates, we can
describe derivatives and integrals on manifolds. Since the components are understood to
be local, we will drop the α subscript on them.
Suppose that we have a curve parametrized by a variable τ on our manifold M. Hence
we have a function where p(τ ) corresponds to the point on the curve for parameter τ . The
local coordinates on the curve are given by xi (p(τ )). The tangent to the curve is found
by taking the derivative with respect to τ
d i
x (p(τ )). (4.5.2)
dτ
Suppose that we generalize this to any function f (p), which we reexpress in terms of
the local coordinates. Then the rate of change of this function along the curve is
∂f dxi i ∂f
≡ a (4.5.3)
∂xi dτ ∂xi
Since the function f is arbitrary, we can instead refer to a tangent vector over p,
∂
ai , (4.5.4)
∂xi
where the ai are the components of the vector. If we consider all possible curves through
a point p, then the allowed values for ai span Rn . Hence the tangent vectors over p make
up a linear vector space called the tangent space of p. This is usually written as

Tp (M). (4.5.5)

The operators ∂i ≡ ∂x∂ i are a basis for the tangent space, hence the tangent space over p
is n dimensional, the same dimension as the manifold.
Given this vector space, we next construct what is known as a dual space. To this end
consider a vector space V and a linear map α such that

α:V →R (4.5.6)

63
where R refers to the real numbers. The fact that this is linear means that α(v1 + v2 ) =
α(v1 ) + α(v2 ). But the space of all such linear maps is itself a vector space. This space is
called the dual space and is written as V ∗ .
To find the dual of the tangent space, we note that for any curve parametrized by
τ passing through point p, and any real function f (p), the tangent dτd f is itself a real
number. We also note that the differential df can be written as
∂f i
df = dx (4.5.7)
∂xi
∂f n
Allowing for a general real function, we see that ∂x i spans R at each point p. Since
d(f + g) = df + dg, we see that this too defines a linear vector space and the differentials
dxi form a basis of this space. Hence, this space has the same dimension as M. Moreover,
df i
this defines the dual space to Tp (M), Tp∗ (M) since dτ is a linear map from dx
dτ i
∂ to R at
each point p. We can write these relations in bra-ket notation, with

hdxi |∂j i = δ i j (4.5.8)

so that
dxi ∂f dxi df
hdf | ∂i i = i
= . (4.5.9)
dτ ∂x dτ dτ
The space Tp∗ (M) is called the cotangent space at point p. Elements of this space df are
called 1-forms.
It turns out that we can generalize 1-forms to m-forms, where m ≤ n. To this end,
we define the wedge product between two 1-forms to satisfy

dxi ∧ dxj = −dxj ∧ dxi . (4.5.10)

A 2-form ω is then given by

ω = ωij dxi ∧ dxj (4.5.11)
where ωij = −ωji . We have written this 2-form in terms of the local coordinates on the
coordinate patch, but this can be mapped back to the open region in M. The 2-forms
also form a linear vector space with basis vectors dxi ∧ dxj . This space is n(n − 1)/2
dimensional and is written as Ω2 (M). We can also take several wedge products to make
higher forms . In fact, we can continue doing this all the way up to n-forms. However,
there cannot be m-forms with m > n, since the space of 1-forms is n-dimensional but
the wedge product antisymmetrizes between the forms. All m-forms form a linear vector
space, Ωm (M).

64
 We can also combine these forms together to make a one bigger vector space for all
m-forms,
Ω(M) = Ω0 (M) ⊕ Ω1 (M) ⊕ ...Ωn (M), (4.5.12)
where Ω0 (M) are the 0-forms, which are basically functions.
There is a very useful operator that maps m-forms to m + 1-forms known as the
exterior derivative, d. By definition, we have that d acting on a m-form Λ is

dΛ = ∂i (Λj1 j2 ...jm ) dxi ∧ dxj1 ∧ ... ∧ dxjm , (4.5.13)

which is clearly an m + 1-form. It then follows that d2 = 0, since d2 acting on any m-form
Λ is
d2 Λ = ∂i ∂j (Λk1 k2 ...km ) dxi ∧ dxj ∧ dxk1 ∧ ... ∧ dxkm = 0, (4.5.14)
since ∂i ∂j = ∂j ∂i . Any operator whose square is 0 is said to be nilpotent.
Any form Φ that satisfies dΦ = 0 is said to be closed. Any form that can be written
as dΛ is said to be exact. Obviously, every exact form is closed, but the converse is not
true. However, it is almost true, in that locally every closed form can be written as the
exterior derivative of another form.
Let us now consider some examples. First consider the circle S1 . The 0-forms are then
functions on the circle, that is functions that are periodic under θ → θ + 2π. The 1-forms
are of the form g(θ)dθ where g(θ) is periodic. The exact 1-forms are df (θ). However,
not all closed 1-forms are exact. For example dθ is closed, but not exact, since θ is not a
0-form on the circle (it is not periodic). However, dθ is locally exact. If we consider the
two open regions of the previous section, then θ, or θ + 2π is a 0-form in R1 .
For the next example, let us consider U (1) gauge fields in 4-d space-time. This space-
time is homeomorphic to R4 . Recall that the gauge fields are 4-vectors, and that the
relevant term in the action for a charged particle in an electromagnetic field is Aµ dxµ .
Hence the gauge field can be thought of as a 1-form, A. From (3.5.1), it is also clear that
the field strength satisfies
Fµν dxµ ∧ dxν = F = dA, (4.5.15)
hence F is an exact 2-form. Thus, we see that dF = 0, which is the Bianchi identity in
(3.6.3). We now see that the Bianchi identity is a consequence of d2 = 0. We also see
that gauge tranformations can be written as

A → A + dΦ (4.5.16)

where the gauge parameter Φ is a 0-form. It is then obvious that F is invariant under
the gauge transformations because of the nilpotence of d.

65
4.6 Integrals of forms
As you might have guessed by the dxi terms, m-forms may be integrated over m dimen-
sional surfaces, or over M itself. The definition of an integral of an m-form over an
m-dimensional surface is as follows. Let Σ be a manifold and let {Σα } be an open cov-
ering where the Σα have invertible maps φα into Rm . We also assume that the covering
is locally finite, which means that each point in Σ is covered a finite number of times by
the covering. An integral of an m-form Λ over an open region Σα is then given by
Z Z
Λ= Λ123...m dx1 dx2 ..dxm . (4.6.1)
Σα φα (Σα )

To find the integral over Σ, we consider a partition of unity, where we have the
functions eα (p), with the properties
X
0 ≤ eα (p) ≤ 1, eα (p) = 0 if p ∈
/ Σα eα (p) = 1. (4.6.2)
α

Hence, we can define the full integral to be

Z XZ
Λ= eα (p)Λ123...m dx1 dx2 ..dxm . (4.6.3)
Σ α φα (Σα )

One important result is Stokes Law. Suppose we have an open region Σα which has a
boundary ∂Σα , then the integral of the exact form dΛ on this region satifies
Z Z
dΛ = Λ. (4.6.4)
Σα ∂Σα

4.7 The monopole revisited

Let us recall the magnetic monopole, with a field strength given by (3.4.1). To find the
charge, we integrated Fij over the surface of S2 . As we previously saw, we could not define
Fij in terms of the curl of A~ over the entire sphere, instead we had to break the sphere
up into two regions, with a different gauge field in each region.
We have already seen that F is a 2-form. For the monopole, it is a closed 2-form on
S2 , but it is not an exact 2-form. We know that it is closed because it is locally exact.
On the northern hemisphere we have that F = dA1 and on the southern hemisphere it is
F = dA2 . The monopole charge is given by the integral
Z Z Z Z
F = dA1 + dA2 = − dΦ. (4.7.1)
S2 N S S1

66
4.8 Fiber bundles
We have seen in the previous sections that a manifold has associated with it a tangent
space and cotangent space at each point p in the manifold. One might ask whether or
not one could combine these spaces with the underlying manifold M to make one big
topological space. The answer to this is yes. In fact we can combine other spaces with the
manifold to make a bigger topological space. The whole space is known as a fiber bundle.
A fiber bundle is defined as follows:
• Let E be a topological space, called the total space or the bundle space.
• There is a projection Π : E → X of E onto X, where X is also a topological space,
called the base space or just the base.
• There exists another topological space F called the fiber, along with a group G of
homeomorphisms of F into itself.
• There is a cover {Xα } of X as well as a set of homeomorphisms φα such that
φα : Π−1 (Xα ) → Xα × F. (4.8.1)
The expression Π−1 (Xα ) refers to that part of E that projects down to Xα .
• The inverse function for φα satisfies
Πφ−1
α (x, f ) = x, where x ∈ Xα and f ∈ F. (4.8.2)
In other words, we assume that the maps take us from the same point in x in the base
space into the same point in x that appears in the product space.
In the overlapping regions where Xα ∩ Xβ 6= ∅ we have that
φα ◦ φ−1
β : (Xα ∩ Xβ ) × F → (Xα ∩ Xβ ) × F. (4.8.3)
Since the point x ∈ Xα ∩ Xβ maps to itself, these maps define homeomorphisms of F ,
given by gαβ (x). These homeomorphisms are called transition functions and they are
required to live in the group G, also known as the structure group of the bundle. The
bundle is usually given by the data (E, Π, F, G, X)
A bundle E is said to be trivial if E is homeomorphic to the product space X × F ,
where X is the base. Based on the construction of a fiber bundle, we see that all bundles
are at least locally trivial. We also should note that there is some similarity in the
definition of a fiber bundle and the definition of a manifold. For the manifold we had
local maps to Rn , whereas for the bundle, we had local maps to a product space. In the
case of the manifold, we considered maps from Rn to Rn over the intersection regions.
For the bundle we considered maps from the product space to the product space in the
intersecting regions.

67
4.9 Examples of fiber bundles
The simplest example of a nontrivial bundle is the Mobius strip. Let the base be S1 and
the fiber be the line segment [−1, 1]. We cover S1 with two overlapping open regions that
intersect in two seperate regions. In the first intersecting region, g12 (x) = 1, in other
words, the map takes the fiber to itself. In the other region, because of the twisting of the
strip as we go around the circle, the map takes every point in the fiber f → −f . Hence
g12 (x) = −1 in this region. Hence the structure group consists of two group elements, the
identity and a nontrivial element g where g 2 = 1. Hence the structure group is Z2 .
The next example we consider is the tangent bundle for M, T (M). The bundle is
given by the union
[
T (M) = Tp (M), (4.9.1)
for all p ∈ M. The base space is M itself and the fiber for any point p is its tangent space
Tp (M). To find the transition functions, we consider the intersection of two coordinate
patches Mα and Mβ , and a point p that lives in this intersecting region. A tangent vector
at p then satisfies
∂ ∂ ∂xj
aiα (p) i = aiβ (p) i = aiβ αi ∂xiβ . (4.9.2)
∂xα ∂xβ ∂xβ
Hence we have that
∂xiα
aiα = ajβ , (4.9.3)
∂xjβ
Hence the transition functions are given by

∂xiα
gαβ (p) = . (4.9.4)
∂xjβ

These are elements of the group GL(n, R), the group of general linear transformations on
an n dimensional real vector space. Hence the structure group is GL(n, R).
We next consider the cotangent bundle, T ∗ (M). This proceeds in a fashion similar to
the tangent bundle. Now the elements of the fibers in a particular coordinate patch have
the form
bαi dxiα . (4.9.5)
Hence the fibers in the intersecting regions have the relation

∂xjβ
bαi = bβj . (4.9.6)
∂xiα

68
Hence the transition functions have the form
∂xjβ
gαβ = , (4.9.7)
∂xiα
and the group is once again GL(n, R). Note the difference in the transition functions
for T ∗ (M) and T (M). The bundles T (M) and T ∗ (M) are examples of vector bundles,
bundles whose fibers are vector spaces.
The next examples we consider are called principle bundles. A principle bundle is
a bundle whose fibers are actually the transition functions. In other words, points in
the fiber are elements of the structure group. It is straightforward to show that the
homeomorphisms of the fiber are actually the group elements. This is because for any
g ∈ G, g : G → G and the map is invertible since every group element has a unique
inverse. Hence the transition functions are elements of the structure group.
The first example of a nontrivial principle bundle is one where we have Z2 over the
circle. The fiber then has two group elements, the identity e and g, where g 2 = e. Breaking
up the circle into two open regions and two intersection regions, we see that in one of the
intersecting regions e(e, g) is mapped to e, g) and in the other intersecting region (e, g) is
mapped to (g, e) = g(e, g), hence g is the transition function in this region.
The second principle bundle we shall consider is known as a frame bundle, where the
fibers are the space of all frames for Tp (M) over a point p. This means the fibers are
all possible bases for the tangent space. Since one basis is rotated into another basis
through a general linear transformations, the frames can be defined as a general linear
transformation of some fixed reference frame. Hence the fibers are essentially elements of
GL(n, R).
We will later see that gauge theories are closely associated with principle bundles of
compact Lie groups.

4.10 More on fiber bundles

A section of a fiber bundle is a continuous map s from the base space to the bundle, which
satisfies
s:X→E Π(s(x)) = x x ∈ X. (4.10.1)
Sometimes this is called a global section. Sections (that is global sections) don’t always
exist.
In fact, for a principle bundle, the existence of a section means that the bundle is
trivial. To see this, note that a section satisfies s(x) ∈ G, where G is the structure group.

69
Hence the fiber over a point x is generated by taking gs(x) for all g ∈ G and the bundle
is generated by taking all g and all points x. Hence an element of the bundle E is given
by gs(x). But given this, there clearly is a continuous map from E to X × G, since s(x)
is continuous and g ranges over G. Namely, the map takes gs(x) → (x, g). The inverse
map back to E is just gs(x), so E is homeomorphic to X × G and so the bundle is trivial.
So for the example of the (E, Π, Z2 , Z2 , S1 ), the only possible sections are s(x) = g
or s(x) = e. But for the nontrivial bundle, it is clear that such an identification is not
possible over the whole circle.
For any bundle E, we can always construct a principle bundle P (E), by replacing the
fibers in E with the transition functions of E, while keeping the transition functions the
same. We now show that if P (E) is trivial, then E itself is trivial.
First we argue that if a bundle E is equivalent to another bundle E 0 with the same
base, fiber and structure group then the transition functions of E 0 are related to those of
E by
0
gαβ (x) = gα−1 (x)gαβ (x)gβ (x), (4.10.2)
where the gα (x) live in the structure group. Suppose that X is covered by Xα and for E
we have the maps φα into the product space and for E 0 we have the maps ψα . This just
constitutes a relabeling of the fiber coordinates. Then

φα ◦ ψα−1 : Xα × F → Xα × F. (4.10.3)

Since the map fixes x, this defines a homeomorphism of F , gα (x), which is in the structure
group. It is then clear that the relation between the structure constants is as in (4.10.2).
We can also argue the converse, that given (4.10.2), then the bundles E and E 0 are
equivalent since we can explicitly construct the homeomorphism between the two bundles.
To show that E is trivial if P (E) is trivial, we note that for any trivial bundle, the
transition functions have the form

gαβ (x) = gα−1 (x)gβ (x), (4.10.4)

since there is a homeomorphism that maps all the transition functions to the identity.
But since the principle bundle is trivial, it means that it has transition functions of this
form, and since E has the same transition functions, it too has functions of this form.
Hence if P (E) is trivial then E is trivial.

70
4.11 Connections and curvatures on bundles
In this section we would like to look at gauge theories from a geometric perspective. We
will assume that the gauge theory lives on some manifold M and that the group associated
with it is G.
Let us consider a principle bundle P over a manifold M with structure group G. Let
us further assume that G is continuous. Locally, the coordinates in the bundle look like
(x, g), where x ∈ M and g ∈ G. Now since M is a manifold and G is continuous, then P
itself is a manifold and as such it has a tangent and a contangent space over each point
in the bundle. If n is the dimension of the manifold and d is the dimension of the group,
then n + d is the dimension of the tangent and cotangent spaces.
We can also construct a tangent bundle for P , T (P ) and a cotangent bundle T ∗ (P )
and consider vectors and 1-forms that live in these bundles. To this end, let us consider
the following 1-form ω which we write as4

ω = ig −1 dg + g −1 Ag (4.11.1)
where g ∈ G and A = Aa µ T a dxµ . In general, g and A are matrices, hence ω is matrix
valued. So for instance, the first term in (4.11.1) has components
h i
g −1 dg = [g −1 ]ik dgkj . (4.11.2)
ij

Hence we see that A is a 1-form on M, while ω is 1-form for P . Note that this is given
for a particular point in P , (x, g). If A = 0, then we see that ω points along the direction
of the fiber, the only differentials are for g and not x.
The differential in terms of a group element may seem a little strange. One could
think about this is in terms of local coordinates, where for a small change in the group
element, we have
g → g exp ia T a , (4.11.3)
where the Ta are the generators of the Lie algebra and the a → 0. The a can be
thought of as the limiting values for the local coordinates θa on the group manifold.
Hence g −1 dg = T a dθa . This is very similar to the 1 − f orms on M, which we often write
in terms of its local coordinates. So for a U (1) group, the group manifold is a circle and
g −1 dg = dθ. For SU (2), we saw that the group manifold was S3 , so the differentials can
be thought of as differentials for the local coordinates on S3 . Having said this, it is more
convenient to leave the form as in (4.11.1).
4
We have dropped the factors of e/h̄. This can be done under a rescaling of the gauge fields.

71
 One problem with the local coordinates is the noncommutativity of g with dg. In
fact, we could have defined the differentials so that g −1 dg = ig −1 T a dθa g. One further
advantage of leaving the differentials as g −1 dg is that they are invariant under a global
transformation, g → hg, where h is a group element. This invariance is called a left
invariance.
Next consider elements of T (P ) which in terms of the group coordinates and local
coordinates on M is given by
!
µ µ ∂ ∂
C Xµ = C iBµ ij + µ . (4.11.4)
∂gij ∂x
We now break up the tangent space into two pieces, a vertical piece V (P ) and a horizontal
piece H(P ). By definition, we define the vertical part to be tangent vectors of the form
∂
γij , (4.11.5)
∂gij
hence these are vectors that point along the fiber and not M. The horizontal piece should
be chosen such that it is orthogonal to V (P ) and so that T (P ) = V (P ) ⊕ H(P ). But
recall that inner products are not taken between tangent vectors, but between a tangent
vector and a cotangent vector. We will define H(P ) then to be those tangent vectors
whose inner product with ω is zero. If A = 0, then clearly H(P ) is M itself, in other
words the horizontal space is just the base of the fiber. If A 6= 0 then the result is more
interesting.
Why bother defining a horizontal space? The reason is that given a point in the
bundle which is locally (x, g), we would like to see how the group element g changes as
we move along a curve in M. In other words, the curve on M can be lifted onto a curve
in the bundle where the curve passes through one point on the fiber. The lifted curve is
determined by H(P ), in that we find the curve by lifting a point in x to a point in the
fiber, and then moving along the curve such that we lie in H(P ). So let a curve be locally
parameterized by x(τ ), where τ is a value that parameterizes the curve. Then the lifted
curve is given by g(x(τ )). If A is zero, then g(x) = g is contant in x. The possible lifts
over a curve in the base space are then parallel since g does not change along the curve.
If A is not zero, then we say that the lifts parallel transport the fibers along the curve.
A is called the connection, since it tells us how to parallel transport a fiber at a point x
to another fiber at x0 . Given ω in (4.11.1), then the elements X ∈ H(P ) satisfy
h i ∂ ∂
hω ij |Xµ i = h[ig −1 ]ik dgkj + g −1 Aa ν T a g dxν | µ + iBµ lm i
ij ∂x ∂glm
h i
= g −1 Ab µ T b g − g −1 Bµ = 0. (4.11.6)
ij

72
Hence we find that h i
Bµ ij = Ab µ T b g . (4.11.7)
ij

This then determines Bµa and hence the horizontal space H(P ).
With the above constraint, a tangent vector in the horizontal piece of the bundle is
given by
∂ ∂
Dµ = µ
+ i [Aa µ T a g]ij . (4.11.8)
∂x ∂gij
As this is written, this does not look particularly enlightening. Let us however, note the
following. Suppose we are considering some matrix valued quantity whose dependence
on g is of the form W f = gW g −1 but which is otherwise independent of g. Under the

transformation g → hg, we see that W f → hW f h−1 . Recall from the previous section that

this is how a gauge covariant object transformed under a gauge transformation h. If we

act with Dµ in (4.11.8) on gW g −1 , we find
  ∂ f
Dµ gW g −1 = W + iAa µ T a gW g −1 − igW g −1 Aa µ T a = [∂µ + iAa µ T a , W
f ], (4.11.9)
∂xµ
where we used that ∂g∂ij gkl = δ i k δ j l . In other words, this is the covariant derivative of the
previous chapter. Hence we see that the covariant derivatives in gauge theories can be
thought of as the covariant derivatives on the principle bundles and that the gauge fields
are the connections. The covariant derivative in (4.11.8) can be used for quantities that
transform other ways under the gauge transformations. For example, in particle physics,
one often considers fields ψ that transform as ψ → hψ. It is straightforward to determine
how the covariant derivative acts on these fields.
Let us now show that the left multiplication of g by h are indeed the gauge trans-
formations. Note that h need not be constant over M. Actually to do this, we need to
say something more about ω. ω is assumed to be invariant under a change of bundle
coordinates that leaves x fixed. In other words, ω is invariant under g → hg. With this
assumption, we see that A must change. In particular, we have that

ig −1 dg + g −1 Ag = ig −1 h−1 d(hg) + g −1 h−1 A0 hg. (4.11.10)

Then this gives that

A0 = −idh h−1 + hAh−1 (4.11.11)
which is our gauge transformation from the previous chapter. Recall that in the last
section we argued that two bundles are equivalent if g → hg for all elements of the fibers.

73
Hence we learn that gauge transformations does not change the equivalency of a bundle.
In other words, equivalent bundles are equal up to a gauge transformation.
Now consider the commutator of two covariant derivatives. This is
!
∂ ∂ ∂ ∂ ∂
[Dµ , Dν ] = i(∂µ Aν − ∂ν Aµ )g − Aa µ Ab ν T a g T b g − T bg T ag . (4.11.12)
∂g ∂g ∂g ∂g ∂g

The quantity in the last parentheses satisfies

∂ b ∂ ∂ a ∂ ∂
Tija gjk Tlm gmn − Tijb gjk Tlm gmn = (T b T a − T a T b )g . (4.11.13)
∂gik ∂gln ∂gik ∂gln ∂g
Therefore, we have
∂ ∂
[Dµ , Dν ] = i(∂µ Aν − ∂ν Aµ − i[Aµ , Aν ])g = iFµν g . (4.11.14)
∂g ∂g
Thus we see that the field strength is associated with the curvature on the principle
bundle.
Note that the 2-form F can be written as

F = dA − iA ∧ A. (4.11.15)

But we also see that we can find a nice relation in terms of ω. Note that

dω − iω ∧ ω = d(ig −1 dg + g −1 Ag) − i(ig −1 dg + g −1 Ag) ∧ (ig −1 dg + g −1 Ag)

= −ig −1 dg ∧ g −1 dg − g −1 dg ∧ g −1 Ag + g −1 dAg − g −1 A ∧ dg
+ig −1 dg ∧ g −1 dg + g −1 dg ∧ g −1 Ag + g −1 Adg − ig −1 A ∧ Ag
= g −1 (dA − iA ∧ A)g. (4.11.16)

In deriving this last equation, we used the fact that d(Ag) = dAg − Adg. Hence, we see
that F = g(dω − iω ∧ ω)g −1 .

4.12 Cohomology
We have argued that principle bundles are equivalent, if the fibers are relabled through the
transformation g → hg. We also showed that this corresponds to a gauge transformation
for the connection 1-form. So it is important to determine whether two principle bundles
are equivalent or not. It turns out that there is a nice way of doing this using something
called charateristic classes. But before doing this, we need to learn a little cohomology.

74
 Let us consider the differential manifold M with dimension n and the vector space of
forms on this n
⊕Ωr (M).
X
Ω(M) = (4.12.1)
r=0

Recall that an r-form Λ is given in terms of local coordinates as

Λi1 ...ir dxi1 ∧ ... ∧ dxir . (4.12.2)

We also recall that the exterior derivative d takes an r-form to an (r + 1) form and that
d2 = 0. We say that d is nilpotent and it turns out that with such nilpotent operators
one can find topological invariants.
In particular, let us consider all r-forms λ that satisfy dλ = 0. Such forms are
called closed. Certainly forms that satisfy λ = dΦ are themselves closed because of the
nilpotence of d. Such forms are called exact. The question is, are all closed forms exact?
The answer is no, but it is this fact that makes them interesting.
Instead of considering all closed forms, we will instead consider all closed forms modded
out by exact forms. In other words, two closed forms are said to be equivalent (or
cohomologous) if
λ1 = λ2 + dΦ → [λ1 ] = [λ2 ] (4.12.3)
for any Φ. This equivalence among closed forms determines a cohomology and the elements
of the cohomology are the closed forms modded out by all exact forms. We write this as

H r (M, R) = Z r (M)/B r (M), (4.12.4)

where Z r refers to all closed r-forms and B r refers to all exact r-forms. The R in
H r (M, R) means that the cohomology takes its values in the reals (it turns out that
there can be cohomologies that take values in the integers, or in Z2 , but we will not
consider them here.) The forms should also be nonsingular, in that integrating over them
should not lead to any divergences.
As an example, let us consider the circle S1 . Since S1 is one dimensional, we see that
all 1 forms are closed. Suppose that the circle is parameterized by θ which runs from
0 → 2π. Then an exact 1-form can be written as df (θ), where f (θ) is periodic in θ. An
example of a closed 1-form that is not exact is dθ. While dθ is single valued around the
circle and hence is an allowed form, θ is not single valued. In any event, any 1-form can
be written as
λ = f (θ)dθ. (4.12.5)

75
Since f (θ) is periodic and nonsingular, it must have the form
X
f (θ) = c0 + [cn cos(nθ) + bn sin(nθ)] (4.12.6)
n=1

where the coefficients can take any real values. But we note that
" #
X cn bn
f (θ)dθ = c0 dθ + d sin(nθ) − d cos(nθ) = c0 dθ + d(g(θ)). (4.12.7)
n=1 n n

Hence any 1-form can be written as c0 dθ plus an exact form. Since c0 is any real number,
we see that H 1 (S1 , R) = R.
Next consider H 0 (S1 , R). In this case no closed forms are exact, since we don’t have −1
forms. The 0-forms are functions, and the closed forms are constants. Since the constants
live in R, we see that H 0 (S1 , R) = R. In fact for any connected space H 0 (M, R) = R.
Let me now state something that I will not prove. This is the Poincaré Lemma, which
states that all closed forms on a contractible space are exact. A space is contractible if
it can be smoothly deformed to a point. The canonical way to contract a space is to
consider maps from M × [0, 1] → M, where (x, t = 0) = x, (x, t = 1) = x0 . If a C ∞
map exits then the flow from t = 0 to t = 1 describes M smoothly shrinking down to the
point x0 . So for example, in 1 dimensions, the closed line segment [−1, 1] is contractible,
since there is the smooth map [−1 + t, 1 − t] that shrinks the segment to the point x = 0.
However, S1 is not contractible. There is no smooth map that takes [0, 2π] to say [0, 0],
because the end points of the interval have to differ by an integer times [2π] and there is
no way that an integer can smoothly jump to another integer.
A manifold M is said to be simply connected if every closed curve is contractible to
a point. For example, the 2-sphere is simply connected since a loop going around the
equator, can be smoothly deformed by pulling it into the northern hemisphere and then
shrinking smoothly to a point at the north pole. Anyway, if the space is simply connected,
then the integral of any closed 1-form over a closed loop is 0.
I
ω = 0. (4.12.8)

To see this, we note that a small deformation of a closed loop does not change the integral.
This is because a small deformation can be written as
I I
ω+ ω (4.12.9)
loop

where loop refers to a small contractible loop. (See figure). Since a closed form is locally

76
 + =

Figure 11: A loop is deformed by adding a small loop to it. Note the orientations of the
loops.

exact, we can write the integral over the small loop as a total derivative. Since the loop is
closed, the integral is zero. Hence, we can keep adding small deformations to the loop and
H
not change the value of ω. But since the loop is contractible, we can smoothly shrink it
to a point by smoothly changing the curve by adding small loops. Since ω is nonsingular,
the integral of the loop when shrunken to a point is zero. (We can also use this argument
to say that for any two noncontractible loops that can be smoothly deformed into one
another, the integrals of ω over the loops are the same.) Since the integral is zero, this
means that the integral is conservative and so ω = dΛ. Hence, for any simply connected
space, H 1 (M, R) = 0.
Let us consider the cohomology for the n dimensional sphere Sn . Since these are simply
connected, we have that H 1 (Sn , R) = 0 if n > 1. We also have that H 0 (Sn , R) = R since
the space is connected. Furthermore, one can show that H p (Sn , R) = H p−1 (Sn−1 , R).
The rough proof of this is as follows. Consider a closed p form ω on Sn . If we consider the
two spaces A1 and A2 , where A1 has the south pole removed and A2 has the north pole
removed, then these spaces are contractible and so ω is exact on each space. Let us say
that ω = dΦ1 on A1 and ω = dΦ2 on A2 . Now consider the intersecting region A1 ∩ A2 .
This space is contractible to an Sn−1 sphere. If ω is not exact, then Φ = Φ1 − Φ2 6= 0
on the intersecting region. But since ω is equal on the two regions, we see that Φ is a
closed form on this region, which contracts to a closed form on Sn−1 . We can also mod
out by the exact forms in this region, since we are free to add exact forms to Φ1 and Φ2
and still satisfy ω = dΦi in the two regions. Hence we see that the cohomology descends
down. Given this, we see that for Sn , we have that H p (Sn , R) = 0 if 0 < p < n and that

77
H n (Sn , R) = R. In general, for an n-dimensional orientable connected compact space,
the surface itself is not contractible, so H n (M, R) = R. These sets of closed forms are
often called the volume forms.
Roughly speaking, the cohomology H p (M, R) counts the number of noncontractible
p-dimensional surfaces in M. So for example, the cohomology of the torus T 2 , which is the
product space of two circles is H 0 (T 2 , R) = R, H 1 (T 2 , R) = R ⊕ R and H 2 (T 2 , R) = R.
The torus has two noncontractible loops, hence the result for H 1 . Since T 2 is compact and
orientable, we have H 2 (T 2 , R) = R. Specifically, we can parameterize the torus by two
angles, θ1 and θ2 . Elements of H 1 (T 2 , R) are c1 dθ1 + c2 dθ2 and the elements of H 2 (T 2 , R)
are cdθ1 ∧ dθ2 .

4.13 Characteristic classes for principle bundles

We would now like to take what we learned for cohomology on manifolds and apply it to
the study of principle bundles. In doing so, we will want to construct closed forms that
are invariant under gauge transformations.
For a group G there there is a straightforward way of finding the invariants, although
it is not always practical. Consider the determinant

det(tI + ca T a ), (4.13.1)

where I is the identity matrix and t and ca are ordinary coefficients. If the T a are m × m
matrices, then this determinant is given by
m
det(tI + ca T a ) = ti Pm−i (ca ),
X
(4.13.2)
i=0

where Pm−i (ca ) is a polynomial of order m − i in the coefficients ca . Assume that the
ca are objects that transform covariantly, in other words, under a gauge transformation,
ca T a → gca T a g −1 . Then the determinant transforms as

det(tI + ca T a ) → det(tI + gca T a g −1 ) = det(g(tI + ca T a )g −1 ) = det(tI + ca T a ). (4.13.3)

Hence, the determinant is invariant. But since it is invariant for arbitrary t, it must be
true that the individual polynomials in (4.13.2) are invariant.
Now instead of ca , let us put in the curvature two forms F = Fµν a
dxµ ∧ dxν T a = F a T a .
Under gauge transformations we have that F transforms covariantly, so the polynomials in
a
(4.13.2) will be invariant. The lorentz indices of Fµν do not effect the group transformation

78
properties, so we can take them anyway we please and the resulting polynomial will still
be invariant. We therefore choose the polynomials to be
!
F
ci (P ) = Pi (4.13.4)
2π

where the products of the terms in the Pi are assumed to be wedge products. The factor
of 2π is put in for later covenience. If we choose the group G to be SU (n), then the
resulting polynomials ci (P ) are called the Chern classes for the bundle P .
To evaluate these Chern classes, we note that the determinant of a matrix is the
product of its eigenvalues. Suppose for a given F , the eigenvalues are f j with n eigenvalues
for G = SU (n). Hence the determinant satisfies
! ! ! !
F f f f
det tI + = t+ 1 ∧ t+ 2 ∧ ... ∧ t + n (4.13.5)
2π 2π 2π 2π

Hence we have that

1 X
ci (P ) = f j1 ∧ f j2 ∧ ... ∧ f ji . (4.13.6)
(2π)i j1 <j2 ..<ji ≤n

Finally, we use the fact that the trace of a matrix is the sum of its eigenvalues, and that
the trace of a matrix to a power m is the sum of the eigenvalues to the power m. We can
then work through the individual values of i and find the characters. In the end, let me
just state the formula:

(−1)i X Y ij 1

j ij

ci (P ) = (−1) TrF (4.13.7)
(2π)i P ji =i j ij !j ij
j

where all products of F are assumed to be wedge products and where the sum means
over all possible ij such that
X
jij = i. (4.13.8)
j

Let me give the first few Chern classes. The lowest one has

c0 (P ) = 1. (4.13.9)

The next one, known as the first Chern class is

1
c1 (P ) = TrF . (4.13.10)
2π

79
For SU (n), since the generators are traceless c1 (P ) = 0. For U (1), we have that c1 (P ) =
1
2π
F . The second Chern class is given by
1 1 1 1
 
c2 (P ) = (TrF ) ∧ (TrF ) − Tr(F ∧ F ) = 2 ((TrF ) ∧ (TrF ) − Tr(F ∧ F ))
4π 2 2 2 8π
(4.13.11)
For U (1), it is clear that this Chern class is 0. For SU (n), this reduces to
1
c2 (P ) = − Tr(F ∧ F ). (4.13.12)
8π 2
The third Chern class is
1 1 1 1
 
c3 (P ) = 3 (TrF ) ∧ (TrF ) ∧ (TrF ) − Tr(F ∧ F ) ∧ (TrF ) + Tr(F ∧ F ∧ F ) .
8π 6 2 3
(4.13.13)
For U (1), it is simple to check that this is 0, as it should be. For SU (2), it should also be
0. The nontrivial term to check is the TrF ∧ F ∧ F term, but it can be shown that this
is zero based on the properties of the Pauli matrices.
The Chern classes have two important properties. The first one is that they are closed
forms. To see this, we note that the F is proportional to F ∼ D ∧ D, where D is
the covariant derivative form D = d − iA. Note that before we used a commutator of
covariant derivatives to find F , but the commutator is essentially built into the wedge
product, since
1
D ∧ D = Dµ Dν dxµ ∧ dxν = [Dµ , Dν ]dxµ ∧ dxν . (4.13.14)
2
Therefore, we have
D ∧ D = −i(dA − iA ∧ A) = −iF . (4.13.15)
This is the nonabelian version of the bianchi identity. Then the covariant derivative acting
on F is

[D, F ] = dF − i[A, F ] = (D) ∧ i(D ∧ D) − i(D ∧ D) ∧ (D) = 0. (4.13.16)

Now consider the trace of the product of any number of F . Then since the covariant
derivative acting on F is 0, we have

0 = Tr((dF − i[A, F ]) ∧ F ... ∧ F ) + Tr(F ∧ (dF − i[A, F ]) ∧ ... ∧ F ) + ...

+Tr(F ∧ F ∧ ... ∧ (dF − i[A, F ]))
= Tr((dF ) ∧ F ... ∧ F ) + Tr(F ∧ (dF ) ∧ ... ∧ F ) + ...Tr(F ∧ F ∧ ... ∧ (dF ))
= d Tr(F ∧ F ∧ ... ∧ F ), (4.13.17)

80
where we used the cyclic properties of the trace to eliminate the [A, F ] terms. Since the
Chern classes are made up of products of traces, it automatically follows that they are
closed.
Another important feature of the Chern classes is that the closed forms only change
by an exact form when the gauge connection is varied. Let me stress that this is not a
gauge transformation, but an actual change in the connection that can change the field
strengths. To show this, suppose that I change the connection globally by a small amount
. Then the change in F is d − i( ∧ A + A ∧ ). Therefore the change in TrF m to
lowest order in  is

δTrF m = mTr((d − i( ∧ A + A ∧ ) ∧ F m−1 )

= md Tr( ∧ F m−1 ) + mTr( ∧ (d(F m−1 ) − i[A, F m−1 ]))
= md Tr( ∧ F m−1 ) + mTr( ∧ D(F m−1 ))
= md Tr( ∧ F m−1 ). (4.13.18)

Therefore, we see that for an infinitesimal change, the trace changes by an exact form.
To see what happens for a finite change in A, let us replace  by tη, where t will vary
from 0 to 1. The gauge fields as a function of t are then A(t) = A + tη. For a small
change from A(t) to A(t + ∆t) then we have that

δTr(F m (t)) = md Tr(η ∧ F m−1 (t))∆t, (4.13.19)

where F is now t dependent. Hence we have that

d
Tr(F m (t)) = md Tr(η ∧ F m−1 (t)) (4.13.20)
dt
and so  Z 1 
∆Tr(F m ) = d m dtTr(η ∧ F m−1 (t)) . (4.13.21)
0
Since all traces differ by an exact form, then all products of traces also differ by exact
forms. As an example consider the product

TrF m ∧ TrF l → (TrF m + dΦ1 ) ∧ (TrF l + dΦ2 )

= TrF m ∧ TrF l + dΦ1 ∧ TrF l + dΦ2 ∧ TrF m + dΦ1 ∧ dΦ2
= TrF m ∧ TrF l + d(Φ1 ∧ TrF l + Φ2 ∧ TrF m + Φ1 ∧ dΦ2 ),
(4.13.22)

where we used the fact that d2 = 0 and that the traces are closed.

81
4.14 The magnetic monopole (yet again), integer cohomology
and the Hopf fibration
We can use the first Chern class to describe magnetic monopoles. Consider the
1
c1 (P ) = F (4.14.1)
2π
integrated over S2 . Locally, F is exact, so F = dA1 in one hemisphere of the sphere and
F = dA2 on the other hemisphere. The two hemispheres intersect along the S1 at the
equator. Therefore
1 Z 1 Z 1 Z 1 Z
F = A1 − A2 = dΦ = n, (4.14.2)
2π S2 2π S1 2π S1 2π S1
where n is an integer. The fact that we can have nonzero values for n is a consequence of
the nontrivial cohomology of H 2 (S2 , R).
Since the quantities we are computing turn out to be integers, it is often convenient to
consider the integer cohomology as opposed to the real cohomology. They are essentially
the same thing. The only difference is that for the integer cohomology, we assume that
the coefficients are integers. What we would say is that H 2 (S2 , Z) = Z. The advantage of
this is that often times we are integrating things that do have coefficients in the integers.
1
In this case, it is the closed form 2π F . The integer coefficient is n.
If n is nonzero, then we have a nontrivial fiber bundle. Let me make this explicit, with
the example of n = 1. In our case, the base space is S2 and the fiber is the U (1) group.
A group element is given by eiφ where φ ≡ φ + 2π. Hence, the fiber space is essentially
a circle S1 . If the principle bundle P is trivial then it is equivalent to a product space
P ' S2 × S1 .
Let us express the sphere in terms of CP (1), that is we have two complex numbers
and an identification (z1 , z2 ) ≡ (λz1 , λz2 ), where λ is any nonzero complex number. Let
us now include the fiber which is given by φ. Region 1 contains the point (1, 0) and region
2 contains the point (0, 1). Since z1 is never zero in region 1 and z2 is never zero in region
2, let us attempt to identify the fiber to be the phase of z1 , θ1 , in region 1 and the phase
of z2 , θ2 , in region 2. In the overlap region, both phases are well defined, since neither
z1 nor z2 are zero here. In the overlap region, from (4.14.2), we see that θ1 = θ2 + nθ
where θ is the angle around the equator. But the angle around the equator is given by
arg(z1 /z2 ), hence we can make this identification of fibers if n = 1. The whole bundle is
therefore described by the space (z1 , z2 ) ≡ (ρz1 , ρz2 ) where ρ is positive real. Note that
the identification under the phases is gone, since a change in the phase is considered a

82
change along the fiber. By rescaling, we can always reexpress our space as (z1 , z2 ) with
the constraint that
z1 z1∗ + z2 z2∗ = 1. (4.14.3)
But this is S3 ! Hence the principle bundle corresponding to a single monopole is topolog-
ically the same as the three sphere. This is clearly not the same as S2 × S1 , for one thing
the cohomology is different. This fibration is known as the Hopf fibration.

4.15 Instantons and the second chern class

In this final section, we consider an important phenomenon in gauge theories called in-
stantons. Instantons are essentially nontrivial solutions to the equations of motion in
Euclidean space. We will consider such solutions on S4 , the 4 sphere. S4 is almost topo-
logically the same as R4 , but not quite. What S4 is equivalent to is R4 ∪ {∞}, that is
R4 and the point at infinity. However, if we restrict our forms on R4 to be well behaved
at infinity, then this is essentially the same as considering the space to be S4 .
Let us consider the second Class for SU (2), which is given by
1
c2 = − Tr(F ∧ F ). (4.15.1)
8π 2
As we have already seen this is a closed form. If we integrate this over S4 , then the result
is a topological invariant.
Since c2 is closed, it is locally exact. Hence we can break up S4 into two regions that
intersect at the equator which is an S3 . Just as in the monopole case, we have that the
integral of the Chern class is Z Z
c2 = Λ (4.15.2)
S4 S3

where Λ is a closed 3 form. What is this 3-form? Just as in the monopole case, it is
basically a “pure gauge” piece.
To see this, let us look more closely at the TrF ∧ F piece. If we expand this in terms
of the A field as shown in (4.11.15), we find

Tr(F F ) = Tr(dA ∧ dA − 2idA ∧ A ∧ A − A ∧ A ∧ A ∧ A). (4.15.3)

We next note that the last term in the parentheses is zero. To see this, we use the
antisymmetry properties of the wedge product and the cyclic property of the trace. You
can readily verify by expanding out the matrices that

Tr(B p ∧ B q ) = (−1)pq Tr(B q ∧ B p ), (4.15.4)

83
where B p is a matrix valued p-form and B q is a matrix valued q-form. Hence we have
that

Tr([A ∧ A ∧ A] ∧ A) = −Tr(A ∧ [A ∧ A ∧ A]) = −Tr(A ∧ A ∧ A ∧ A) = 0. (4.15.5)

We can then show that the remaining terms are

i
  
Tr(dA ∧ dA − 2idA ∧ A ∧ A) = d Tr F ∧ A + A ∧ A ∧ A . (4.15.6)
3
To verify this last statement, we note that
i
  
d Tr F ∧ A + A ∧ A ∧ A
3
i
 
= Tr dF ∧ A + F ∧ dA + (dA ∧ A ∧ A − A ∧ dA ∧ A + A ∧ A ∧ dA)
3
= Tr (i[A, F ] ∧ A + F ∧ dA + idA ∧ A ∧ A)
= Tr(dA ∧ dA − 2idA ∧ A ∧ A), (4.15.7)

where in going from the second line to the third line, we used the Bianchi identity in
(4.13.15). Hence, in the two hemispheres of S4 , we have that
1 i
 
c2 = dΦ Φ = − 2 Tr F ∧ A + A ∧ A ∧ A . (4.15.8)
8π 3
Next we note that under a gauge transformation, A → U † AU − iU † dU , Φ transforms
as
1 i

Φ → Φ − 2 Tr(F ∧ (−iU † dU )) + 3 Tr(−iU † dU ∧ A ∧ A)
8π 3
i 1

† † † † †
− 3 Tr(U dU ∧ U dU ∧ A) − Tr(U dU ∧ U dU ∧ U dU )
3 3
1 1
 
† † † †
= Φ − 2 −id(Tr(A ∧ (U dU ))) − Tr(U dU ∧ U dU ∧ U dU ) . (4.15.9)
8π 3
Therefore, we find that on the S3 equator, if the two gauge fields from the northern and
southern hemispheres differ by a gauge transformation, then
Z
1 Z 1
c2 = Tr(U † dU ∧ U † dU ∧ U † dU ), (4.15.10)
S4 8π 2 S3 3
where we have thrown away the total derivative term.
To find this integral, let us recall that the group manifold of SU (2) is actually the
3-sphere. In other words, a general element U is given by

U = b0 + b1 σ1 + b2 σ2 + b3 σ3 , b20 + b21 + b22 + b23 = 1. (4.15.11)

84
Since the group manifold is S3 , there exists a one to one and onto map from the physical
S3 to the S3 of the group manifold. For such a map, the trace term in (4.15.10) is 12
times the volume form of S3 . The factor of 12 has a factor of 3! from the combinatorics in
the wedge product and a factor of 2 from the two dimensional trace. For this particular
map, we then have that Z
1 Z
c2 = 2 ω3 = 1 (4.15.12)
S4 2π S3
where ω 3 is the volume 3-form of S3 . A more general class of maps would wind the S3 n
times around the group manifold. In this case
Z
c2 = n n ∈ Z. (4.15.13)
S4

If the second Chern class is nonzero, then the bundle over S4 is nontrivial. In the case
where n = ±1, the bundle is a Hopf fibration which is topologically the same as S7 .
Why should we care about these objects? The reason is that we can use them to find
nontrivial minima for gauge field configurations. In doing physical computations, it is
often convenient to rotate the time variable t to be imaginary, in other words t satisfies
t = iτ , where τ is real. The effect this has on the action that appears in the path integral
is that Z Z
S = dtL → i τ L = iSE . (4.15.14)
SE is called the Euclidean action. This is because the normal minkowski metric is rotated
to a Euclidean metric:
ds2 = −dt2 + d~x2 → dτ 2 + d~x2 . (4.15.15)
Under the rotation, the phase factor in the path integral becomes

eiS → e−SE (4.15.16)

Hence for the Euclidean action, we want to find field configurations that minimize SE .
Let us consider the Euclidean action for a gauge field,
1 Z 4
SE = d xTr(Fµν F µν ), (4.15.17)
4k
where the integral is over R4 . The traces are normalized so that k = 1/2. It is now
convenient to express this using something called the dual form. On R4 , the dual form is
constructed by taking the  tensor and writing
1
Feµν = µνλρ F λρ . (4.15.18)
2
85
Feµν is antisymmetric in its two indices and from it we can write a two tensor, which is
usually written as ∗F ,
1
∗F = Feµν dxµ dxν . (4.15.19)
2
Then it is not too hard to see that the Euclidean action is
Z
SE = TrF ∧ ∗F . (4.15.20)

We also note that

TrF ∧ F = Tr ∗ F ∧ ∗F . (4.15.21)
Hence we have that
1
Tr(F ∧ ∗F ) = Tr([F ± ∗F ] ∧ [∗F ± F ]) ∓ Tr(F ∧ F ). (4.15.22)
2
Now the first term on the rhs of (4.15.22) is positive definite and the second term when
integrated over R4 is a topological invariant. Hence we have a local minimimum of SE in
terms of the field configurations if we set the first term on the rhs to zero. This means
that
F =±∗F (4.15.23)
These configurations are localized in the R4 in the sense that F falls to zero as we approach
infinity. These configurations can be thought of solutions in space-time, but where the
time component is also space-like. Since the solution is also localized in euclidean time,
they are call “instantons”. Actually the solution with the plus sign in (4.15.23) is called
an instanton and the one with a minus sign is called an anti-instanton. We can easily see
what SE is at the minima using (4.15.22) and (4.15.13), namely it is

SE = 8π 2 |n| (4.15.24)

The absolute value arises because the lhs of (4.15.22) is positive definite. Hence the
instantons can only have positive n and the anti-instantons can only have negative n.
Hence we see that for n 6= 0, the instantons are nontrivial configurations since the action
is nonzero.

86