NOTES ON TOPOLOGICAL QUANTUM FIELD THEORY

DYLAN G.L. ALLEGRETTI

Abstract. These notes are meant to provide an introduction to topological quantum eld
theory for mathematicians. Topics include applications of Chern-Simons theory to the study
of knots, Atiyahs denition of a topological quantum eld theory, and the invariants of
Reshetikhin and Turaev. We also include a self-contained introduction to Yang-Mills theory
for physical motivation.
1. Introduction
1.1. Feynman integrals. Quantum eld theory is the formalism developed by physicists for
describing quantum mechanical systems that look classically like elds. It provides a unied
framework for describing many dierent phenomena. For example, light is understood to be
the result of excitations in a medium called the electromagnetic eld, and there is a quantum
eld theory called quantum electrodynamics that describes the dynamics of this eld and its
interaction with matter. Other quantum eld theories are used to describe the strong and
weak nuclear forces and certain condensed matter systems.
Mathematically, a quantum eld theory consists of very simple data. To specify a quan-
tum eld theory, we rst of all need a smooth manifold M called spacetime (usually four-
dimensional with a Lorentzian metric g). We must also specify a vector space T(M) whose
elements are called the elds on M (these could be smooth functions or vector elds on M
for example). Finally, we must specify a function S : T(M) R called the action and
a measure D on T(M). To perform computations in quantum eld theory, we have to
compute certain integrals called Feynman integrals. In particular, we have to compute the
integral
Z =
_
F(M)
e
iS()
D
which is known as the partition function. Once we know the value of this integral, we can
use it to compute the expected values of various physical quantities. By physical quantity,
we mean any attribute of a physical system to which we can assign a number by making
the system interact with a measuring device. Such a quantity is represented by a function
f : T(M) R whose value f() on a eld represents the value of the quantity if the eld
conguration happens to be . If f is the function representing some physical quantity, then
the average measured value of this quantity in the vacuum is
f =
1
Z
_
F(M)
f()e
iS()
D.
Last update: January 8, 2013
1
Performing computations in quantum eld theory is thus a simple matter in principle as it all
boils down to computing integrals. The problem that makes computations extremely dicult
in practice is that the measure D usually cannot be dened rigorously. We are therefore
forced to resort to nonrigorous and heuristic methods to extract physical predictions from
quantum eld theory.
Despite this shortcoming, it turns out that we can evaluate Feynman integrals in certain
special cases. Interestingly, there are quantum eld theories in which the values of these
integrals are topological invariants of M, and in the case of Chern-Simons theory, knots em-
bedded in M. Theories that compute topological invariants in this way are called topological
quantum eld theories (TQFTs). The invariants turn out to be easily computable once we
have made sense of the integration, and for this reason, quantum eld theory has recently
become an extremely important topic in pure mathematics.
1.2. Prerequisites. The purpose of these notes is to give a physically motivated introduc-
tion to TQFTs and their applications. They are primarily aimed at mathematicians who wish
to understand recent developments in mathematical physics. The mathematical prerequisites
therefore consist of standard graduate-level mathematics, including some understanding of
basic dierential geometry and category theory. We assume, for example, that the reader is
familiar with the notion of a smooth manifold, tangent space, and vector bundle. We do not
assume any familiarity with more advanced topics in dierential geometry such as the the-
ory of principal bundles or cobordisms. As far as category theory is concerned, these notes
assume the reader is comfortable with the basic ideas of categories, functors, and natural
transformations, but the prerequisites stop short of any advanced topics like monoidal or
higher categories.
Since these notes are intended primarily for mathematicians, they do not presuppose any
knowledge of physics and they do not discuss the physical ideas related to TQFTs in much
detail. Instead, the physics is used to motivate the mathematical denition of a TQFT. The
purpose of these notes is not to teach physics but to explain how the physicists idea of a
Feynman integral can serve as a source of inspiration for mathematicians.
1.3. Organization. We begin with a discussion of the physical aspects of quantum eld
theory in Section 2. In this section, we explain how physical elds are represented mathe-
matically and explain the notion of gauge symmetry. We then introduce principal bundles
and principal bundle connections and use these mathematical tools to discuss classical Yang-
Mills theory. To illustrate how Yang-Mills theory arises in physics, we briey discuss its role
in quantum electrodynamics. This discussion of Yang-Mills theory is necessary to motivate
our later discussion of Chern-Simons theory and Dijkgraaf-Witten theory. The treatment in
this section is based on the excellent notes [1].
In Section 3, we discuss Chern-Simons theory, a topological theory that gives rise to many
topological invariants of knots. We begin by reviewing the basic invariants of knots and links,
namely the writhe, Kauman bracket, and Jones polynomial. Then we dene the Chern-
Simons action and introduce the notions of holonomy and Wilson loops. Finally, we sketch
the remarkable connection between Chern-Simons theory and link invariants introduced
previously. The treatment in this section borrows heavily from [2].
2
In Section 4, we give a physically motivated introduction to Atiyahs axiomatization of a
TQFT, and we explain how a theory satisfying the axioms can be used to compute topological
invariants. In the process, we have to introduce the theory of cobordisms and symmetric
monoidal categories. Once we have dened TQFTs, we discuss Dijkgraaf-Witten theory as
an example. The material in this section comes from [3], [4], and [5].
Finally, in Section 5, we discuss the theory of Reshetikhin-Turaev invariants, a mathe-
matically rigorous counterpart to Chern-Simons theory. To dene these invariants of links
and 3-manifolds, we rst introduce the notions of ribbon categories and modular tensor
categories. We then explain how such categories give rise to invariants coming from three-
dimensional TQFTs. The treatment in this section is based on [6].
1.4. Convention. In these notes, we will follow the Einstein summation convention and
implicitly sum over all indices which appear once as a superscript and once as a subscript.
For example, when we write an expression y = c

, we really mean the sum y =

.
2. Yang-Mills Theory
2.1. The mathematical description of elds. Before we begin our discussion of topo-
logical quantum eld theories, we will spend some time looking at how quantum eld theory
is used in physics. The main example that will motivate our discussion in this section is
the theory of an n-component vector eld. In other words, we will model physical elds as
smooth functions : M R
n
where for simplicity we take M = R
4
. We usually write a
eld in terms of its components as = (
1
, . . . ,
n
). To specify an action for the theory,
we rst form the expression
/ =
n

i=1
_
1
2

1
2
m
2

2
i
_
where m > 0 is a constant. Here we have dened

= g

where (g

) is the inverse of
the matrix (g

). An expression of this sort is called a Lagrangian density. Once we have a

Lagrangian density, the action is obtained by integrating over all of spacetime:
S() =
_
R
4
d
4
x/.
Observe that in this theory, the elds are represented by sections of a certain ber bundle,
namely MR
n
M. In general, physical elds are represented by smooth sections of some
bundle over the spacetime manifold. This point of view will be useful later on when we begin
our study of gauge elds.
2.2. Gauge invariance. An important feature of our n-component vector eld theory is
its invariance under the action of certain transformations. In this theory, the physical elds
are functions M R
n
. Now there is an action of the group O(n) on the n-dimensional
vector space R
n
given by matrix multiplication, so there is an action of O(n) on functions
3
: M R
n
given by (g)(p) = g(p). Let g = (a
ij
) be an orthogonal matrix. Then

(g)
i
=

j=1
a
ij

j
=
n

j=1
a
ij

j
= g(

)
i
and since orthogonal transformations preserve the inner product on R
n
, it follows that

(g)
i

(g)
i
=

i
. We also have

i
(g)
2
i
=

2
i
, so the action of O(n)
preserves the full expression for the Lagrangian density. Since the Lagrangian density deter-
mines the partition function, it follows that all the physics is invariant under the action of
O(n). In general, whenever we have a group G acting on the space of elds and preserving the
Lagrangian density, we say that the theory is invariant under global gauge transformations.
In this case, the group G is called the gauge group of the theory.
The idea of Yang-Mills theory is to require the physics to be invariant under not only
global gauge transformations but arbitrary local gauge transformations as well. In other
words, we wish to consider a theory with gauge group G, and for any function g : M G,
we wish to write down a Lagrangian density that is invariant under the transformation
(x) g(x)(x). Unfortunately, the Lagrangian density that we have been using is not
invariant under these local gauge transformations because in general

g(x)(x) ,= g(x)

(x).
We will therefore have to modify our Lagrangian density, replacing the partial derivatives

by the more sophisticated notion of covariant derivative, which transforms in a nontrivial

way when we apply a local gauge transformation.
2.3. Principal bundles. In order to write down a Lagrangian density that is invariant under
local gauge transformations, we need to develop a theory of bundles with group actions.
Denition 2.3.1. Let F be a smooth manifold. We say that a bundle : P M is locally
trivial with standard ber F if there exists an open covering U

of M and for each a

homeomorphism

called a local trivialization such that

1
(U

H
H
H
H
H
H
H
H
H

F
proj
.w
w
w
w
w
w
w
w
w
U

commutes where proj is the projection onto the rst factor. Let G be a Lie group and let
: P M be a bundle with a continuous right action P G P such that G preserves
the bers of and acts freely and transitively on them. Such a bundle is called a principal
G-bundle if it is locally trivial and the local trivializations are G-equivariant.
4
2.4. Associated bundles. Next we are going to study a method of constructing bundles
from a principal bundle. Let : P M be a principal G-bundle and F a smooth manifold
that admits a left G-action. Then there is a right action (P F) G P F of G on the
product P F given by
(u, f) g = (u g, g
1
f).
Dene P
G
F = (P F)/G with the quotient topology. Since the composition P F
proj

P

M is constant on G-orbits, there is an induced map : P
G
F M. This induced
map has a special name.
Denition 2.4.1. The map : P
G
F M is called the bundle associated to P with
ber F. We use the notation [p, f] for the class of (p, f) P F in the total space of this
bundle.
Proposition 2.4.2. The associated bundle : P
G
F M is a locally trivial bundle with
the same trivializing open cover as : P M and bers dieomorphic to F.
Proof. Let U M be a trivializing open set for the principal bundle : P M, and let
:
1
(U) U G be the associated local trivialization. For any point [p, f]
1
(U),
we have (p) = (x, g) for some x U and g G, and we can dene :
1
(U) U F by
([p, f]) = (x, g f).
It is straightforward to check that this map is well dened. On the other hand, there is a
mapping

: U F
1
(U) given by

(x, f) = [
1
(x, 1), f].
and one can show that is a local trivialization with inverse

. This local trivialization

takes the ber over x to x F and therefore gives rise to a dieomorphism between bers
of and the space F. This completes the proof.
A special case of this construction is when F is a vector space on which G has a smooth
representation : G GL(F). In this case the bundle associated to P is called a vector
bundle. The following examples and notations will appear over and over in our discussion.
Examples 2.4.3.
(1) If G is a group of n n matrices and V is an n-dimensional vector space, then we
can take to be the fundamental representation G GL(V ).
(2) Let g be the Lie algebra of G and take to be the adjoint representation Ad : G
GL(g). Then the associated bundle P
G
g M is called the adjoint bundle of P
and denoted Ad(P).
2.5. Transition functions. If U

and U

are two sets in our open cover having nonempty

intersection U

= U

, then there are two trivializations

:
1
(U

) U

G.
These can be written

(p) = (m, g

(p)) and

(p) = (m, g

(p)) where p
1
(m), and we
dene g

(p) to be the element G for which g

(p) = g

(p)g

(p).
Lemma 2.5.1. There exist functions g

: U

G such that g

(p) = g

((p)).
5
Proof. For any g G, we have
g

(pg) = g

(pg)g

(pg)
1
= g

(p)gg
1
g

(p)
1
= g

(p).
Since G acts transitively on the bers of P, it follows that g

is constant on each ber. The

result follows immediately.
Denition 2.5.2. The functions g

: U

G in the lemma are called transition functions.

2.6. g-valued forms. Normally, when we talk about a dierential p-form, we mean an
operation that takes p vector elds as input and produces a number. Equivalently, one can
dene a p-form to be a smooth section of
p
T

M. Sometimes it is useful to generalize this

notion and consider an operation that takes p vector elds and produces an element of some
vector space X.
Denition 2.6.1. If X is a vector space and M is a manifold, then an X-valued dierential
p-form is a smooth section of X
p
T

M. The space of all p-forms with values in X is

denoted
p
(M, X).
Many operations and notations from the theory of ordinary real-valued dierential forms
can also be applied to vector-valued dierential forms. For example, if and are vector-
valued dierential forms, then we dene their wedge product by the usual formula with real
multiplication replaced by tensor product:
( )(v
1
, . . . , v
p+q
) =
1
p!q!

S
p+q
sgn()(v
(1)
, . . . , v
(p)
) (v
(p+1)
, . . . , v
(p+q)
).
Thus the wedge product of p-form with values in X and a q-form with values in Y is a
(p + q)-form with values in the tensor product X Y . In the special case where and
take values in a Lie algebra g, we can compose the wedge product operation with the Lie
bracket [, ] : g g g to get a new element of g called the bracket:
[, ] = [ ].
We can also extend the notion of exterior derivative to X-valued dierential forms. Once we
choose a basis e
i
for X, such a form can be written =
i
e
i
where
i
are real-valued
dierential forms, and we dene the exterior derivative of by the formula d = d
i
e
i
.
Example 2.6.2. For any Lie group G, the Maurer-Cartan form
1
(G, g) is dened by

g
= (L
g
1)

: T
g
G T
1
G = g
where L
g
1 : G G, x g
1
x. One can show that in the special case where G is a matrix
group, the Maurer-Cartan form is given by
g
= g
1
dg. Other important properties of the
Maurer-Cartan form include the transformation law
R
1
g
= Ad(g
1
)
where R : G G is the function dened by R
g
(x) = xg and the equation
d +
1
2
[, ] = 0
6
which is known as the Maurer-Cartan equation.
2.7. Connections. The concept of a connection is central to dierential geometry because
it allows us to dierentiate sections of bundles and transport vectors from one tangent space
to another. To state the denition, we let : P M be a principal G-bundle on M with
trivializing open cover U

.
Denition 2.7.1. Let R
g
: P P denote right multiplication by g G, and let
p
: G P
be the mapping g pg for p P. For any X g let X be the vector eld on P given by
(X)
p
= (
p
)

X. A connection 1-form on P is an element of

1
(P, g) such that
R

g
= Ad(g
1
)
((X)) = X.
It is a fact, whose proof goes beyond the scope of these notes, that a connection 1-form
exists on any principal bundle. Suppose is a connection 1-form on P. For each index ,
there is a canonical local section s

: U

1
(U

) dened so that for each point m U

we have

(s

(m)) = (m, 1). Dene A

1
(U

, g) by
A

= s

.
This 1-form is known in physics as a vector potential or gauge eld. We will use the following
result to see how the vector potential changes when we pass from one trivializing open set
to another.
Proposition 2.7.2. Let be a connection 1-form, and let

be its restriction to
1
(U

).
Then

can be written

= Ad(g
1

+g

where is the Maurer-Cartan form dened above.

Proof. We begin by proving that the two forms agree on the image of s

. Suppose we are
given m U

and p = s

(m). Then we have a direct sum decomposition

T
p
P = im(s

V
p
,
so we can write any tangent vector v T
p
P uniquely in the form v = (s

(v) + v for some

v V
p
. Since g

is constant, we have (g

= 0 and therefore (g

v = (g

v. It
follows that
_
Ad(g

(p)
1
)

+g

_
(v) = (

)(v) + (g

)(v)
= ((s

v) + ((g

v)
= ((s

v) + ((g

v)
= ((s

v) + ( v)
=

(v),
7
so the two forms are agree on the image of s

. Next we will show that these two forms

transform in the same way under the action of G. Indeed,
R

g
(Ad(g

(pg)
1
)

+g

) = Ad((g

(p)g)
1
)R

+g

= Ad(g
1
g

(p)
1
)

+ g

(Ad(g
1
))
= Ad(g
1
)
_
Ad(g

(p)
1
)

+g

_
= Ad(g
1
)

.
It follows that equality holds everywhere on
1
(U

).
Corollary 2.7.3. If U

and U

overlap, then on the intersection U

, we have
A

= Ad(g

)A

+g

.
Proof. Since

and

are restrictions of a globally dened 1-form, we know that

on
1
(U

). Therefore on the intersection U

, we have
s

= s

= s

(Ad(g

(s

)
1
)

+g

)
= Ad(g

)A

+ g

In the last step, we used the fact that g

= g

.
2.8. Gauge transformations. Like all geometric objects, principal bundles have a natural
notion of automorphism. Since we ultimately would like to describe physical systems using
the principal bundle formalism we have developed, we should make sure that all the physics
is invariant under these automorphisms.
Denition 2.8.1. Let : P M be a principal G-bundle. A gauge transformation of P is
G-equivariant dieomorphism : P P such that the following diagram commutes.
P

A
A
A
A
A
A
A
A

.}
}
}
}
}
}
}
}
M
If is a gauge transformation of P, then maps bers to themselves and therefore
restricts to a gauge transformation of the trivial bundle
1
(U

). Let

:
1
(U

) U

G
be the associated local trivialization. Then we can write

((p)) = ((p), g

((p))) and
use this to dene

(p) = g

((p))g

(p)
1
.
The fact that g

and are G-equivariant implies

(pg) =

(p) for all g G. Since G

acts transitively on bers of P, it follows that

is constant on bers and hence there exists

: U

G such that

(p) = ((p)). Let m be a point of U

and p a point in the ber

over m. Then we have

(m) = g

((p))g

(p)
1
= g

((p))g

((p))
1
g

((p))g

(p)
1
g

(p)g

(p)
1
= g

(m)

(m)g

(m)
1
8
since ((p)) = m. This prove that

(m) = Ad(g

(m))

(m) and therefore the maps

dene a section of the adjoint bundle P

Ad
G. These maps correspond to the local gauge
transformations that we considered earlier while discussing the physical motivation.
Proposition 2.8.2. Suppose we are given a connection 1-form , and let

= (
1
)

be the corresponding gauge transformed 1-form. Let A

and A

be the gauge elds on U

corresponding to and

. Then
A

= Ad(

)(A

).
Proof. We can write

p
= Ad(g

(p)
1
)

+g

and

p
= Ad(g

(p)
1
)

+g

.
Letting q =
1
(p) and applying (
1
)

to the rst of these equations, we nd

p
= Ad(g

(q)
1
)(
1
)

+ (
1
)

= Ad(g

(q)
1
)(
1
)

+ (g

1
)

= Ad(g

(q)
1
)

+ (g

1
)

= Ad(g

(p)
1

(p))

+ (g

1
)

.
The fact that (g

1
)(p) = g

(q) =

(p)
1
g

(p) implies
(g

1
)

= g

Ad(g

(p)
1

(p))

and therefore

p
= Ad(g

(p)
1

(m))

(A

) + g

.
Combining this with our other formula for

p
yields the desired result.
2.9. Covariant dierentiation. Fix a trivializing open set U M, and let A be a vector
potential on this open set. If : P
G
V M is the bundle dened in Example 2.4.3,
then there is a local trivialization
1
(U)

= U V , so we can think of any local section
s : U
1
(U) of as being essentially the same thing as a smooth function s : U V .
If we choose a basis e
i
for V , then we can write s : U V in terms of coordinates as
s = s
i
e
i
with s
i
C

(M). Since the Lie algebra of GL(V ) is gl(V ), we can view the 1-form
A as a matrix whose entries A
j
i
are ordinary real-valued 1-forms. One then has the following
derivative operation.
Denition 2.9.1. The covariant derivative of s is the section
v
s of V dened by

v
s = ds
i
(v)e
i
+s
i
A
j
i
(v)e
j
.
We think of this new section
v
s as the derivative of s in the direction v. In the
special case when v =

, we write
v
=

. Notice that gauge transformations act on

covariant derivatives, taking
v
to a new operation

v
in which the vector potential A is
replaced by the gauge transformed vector potential A

. The following proposition says that

covariant derivatives behave in a particularly nice way under gauge transformations.
9
Proposition 2.9.2. Suppose G is a matrix group and g

: U

G are functions dening a

gauge transformation. Under this gauge transformation,

s = g

s.
Proof. Let g be the g

dened on the open set U. Since G is a matrix group, we can write

g in terms of its matrix coecients, g = (a
i
j
). Then

(gs)
i
=

(a
i
j
s
j
)
= (

a
i
j
)s
j
+ a
i
j
(

s
j
)
= ((

g)s)
i
+ (g

s)
i
,
and the gauge transformed vector potential can be written
A

= gA

g
1
(

g)g
1
where we denote A

= A(

) and A

= A

).Therefore

gs =

(gs)
i
e
i
+ (gs)
i
(A

)
j
i
e
j
= ((

g)s)
i
e
i
+ (g

s)
i
e
i
+ (gs)
i
(gA

g
1
)
j
i
e
j
(

gg
1
)
j
i
(gs)
i
e
j
= (g

s)
i
e
i
+ (gs)
i
(gA

g
1
)
j
i
e
j
= g

s
as desired.
2.10. An invariant Lagrangian density. We have now developed the mathematics nec-
essary to modify our n-component theory so that the Lagrangian density is invariant under
local gauge transformations. Let be the vector bundle with ber R
n
associated to a prin-
cipal O(n)-bundle : P M. Fix a trivializing open set U M and let e
i
be the basis
of local sections of over U given by
e
i
(p) = (0, . . . , 0, 1, 0, . . . , 0)
with a 1 in the ith position and 0s elsewhere. A physical eld is represented mathemat-
ically as a section of this vector bundle, so locally we can write =

i
e
i
for smooth
functions
i
on U. Then we can rewrite our Lagrangian density using covariant instead of
partial derivatives, and we obtain a new expression with the desired invariance property:
/ =
n

i=1
_
1
2
(

)
i
(

)
i
+
1
2
m
2

2
i
_
.
Notice that in order to write down a Lagrangian density which behaves nicely under local
gauge transformations, we had to choose a g-valued 1-form A. In quantum eld theory, this
object is interpreted as a new physical eld, and if we wish to describe this eld quantum
mechanically, we will need to specify an action.
10
2.11. Curvature and eld strength. The rst step in writing down an action for the
gauge eld A is to construct an auxiliary eld called the gauge eld strength.
Denition 2.11.1. Let be a connection 1-form. Then the curvature of is the 2-form
= d +
1
2
[, ]. If s

: U

P is one of the canonical sections dened above, then the

gauge eld strength is the pullback F

= s

.
Explicitly, the eld strength F

is given by the formula

F

= dA

+
1
2
[A

, A

].
Since the eld strength F

is dened locally, it is natural to ask how it changes when we

pass from one trivializing open set to another. The answer is provided by the following
proposition, which implies that the F

patch together to form a globally dened 2-form with

values in the adjoint bundle Ad(P).
Proposition 2.11.2. Let U

and U

be two trivializing open subsets of M. Then on their

intersection we have
F

= Ad(g

)F

Proof. We have seen that on the intersection U

the gauge eld A

is given by
A

= Ad(g

)A

+g

.
Bracketing this 1-form with itself and using graded commutativity of the bracket operation,
we see that
[A

, A

] = [Ad(g

)A

+ g

, Ad(g

)A

+g

]
= [Ad(g

)A

, Ad(g

)A

] + [Ad(g

)A

, g

]
+ [g

, Ad(g

)A

] + [g

, g

]
= [Ad(g

)A

, Ad(g

)A

] + 2[Ad(g

)A

, g

] + [g

, g

].
Now this last expression contains
[Ad(g

)A

, g

] = Ad(g

)[A

, Ad(g
1

)g

]
and we have
Ad(g
1

)g

= g

Ad(g
1

)
= g

= (R
g

.
The composition R
g
1

equals the constant function 1, so its dierential is zero, and

therefore the pullback of the Maurer-Cartan form by this composition vanishes. It follows
that
[A

, A

] = [Ad(g

)A

, Ad(g

)A

] + [g

, g

]
= Ad(g

)[A

, A

] + g

[, ]
= Ad(g

)[A

, A

] 2g

d.
On the other hand, we have dA

= Ad(g

)dA

+g

d, so the desired result is a consequence

of the formula F

= dA

+
1
2
[A

, A

].
11
2.12. The Yang-Mills action. In order to write down an action for the eld A, we need to
assume that our Lie algebra g admits an Ad-invariant inner product , : g g R. Of
course an inner product of this sort may not exist on an arbitrary Lie algebra, but from now
on we will only consider gauge groups whose Lie algebras do admit such an inner product.
If we assume our manifold M comes equipped with a metric tensor g, then there is also
a natural inner product operation on dierential forms. If =

dx

and =

dx

are 1-
forms on M, then this inner product is the function , = g

on M. More generally,
for 1-forms e
1
, . . . , e
p
and f
1
, . . . , f
p
, we dene
e
1
e
p
, f
1
f
p
= det(e
i
, f
j
)
and extend linearly to get an inner product on the space of p-forms.
Now, if S and T are g-valued dierential forms written in terms of sections S and
T of g and ordinary p-forms and , then we can combine these notions of inner product
and write
S , T = S, T, .
In this way, we obtain an inner product operation on g-valued dierential forms. We will
be particularly interested in the squared norm [F

[
2
= F

, F

where F

is the gauge eld

strength introduced above. By Ad-invariance of the inner product on g, we know that this
norm agrees with [F

[
2
on the set U

and therefore there is a globally well dened function

[F[
2
on M where F is the eld strength regarded as a section of Ad(P). With this notation,
the Yang-Mills action can be written
S(A) =
_
M
[F[
2
vol,
provided this integral exists.
To rewrite this action in a form familiar to physicists, we need the following important
operation.
Denition 2.12.1. The Hodge star operator is the unique linear map :
p
(M)
np
(M)
such that for all ,
p
(M) we have = , vol. We also dene :
p
(M, g)

np
(M, g) by (T ) = T where T is a g-valued p-form written in terms of a
section T of g and an ordinary p-form .
If and are g-valued 1-forms, it is easy to see that the expression , vol is the same
as the composition of with the inner product , : g g R. Abusing notation
slightly, we write this composition as Tr(F F). Then the Yang-Mills action can be written
S(A) =
_
M
Tr(F F).
2.13. Maxwell theory. To conclude this section, we briey mention how Yang-Mills theory
arises in quantum electrodynamics, the quantum theory describing electrons, positrons, and
photons. According to this theory, electrons and positrons are the particle-like excitations in
a eld called the Dirac eld. Although more complicated than the n-component vector eld
we have been discussing, the Dirac eld has a similar Lagrangian density, which is invariant
under global gauge transformations with gauge group U(1). As in the example above, we
must use covariant derivatives to ensure that the expression is invariant under local gauge
12
transformations. As before, this requires introducing a gauge eld A, and the associated
eld strength F is called the electromagnetic or photon eld. Excitations in this eld are
photons, the quantum mechanical particles that make up light.
3. Chern-Simons Theory
3.1. Basic knot theory. We begin this section with a review of some of the classical results
of knot theory. As we will see, quantum eld theory provides an elegant framework in which
these results can be understood.
Denition 3.1.1. A knot in M is a submanifold of M dieomorphic to a circle. A link is
a submanifold dieomorphic to a disjoint union of circles.
We will typically represent knots and links by drawing two dimensional diagrams. For-
mally, if L is a link, we obtain a link diagram by taking a projection p onto R
2
in such a way
that any point of p(L) has a neighborhood that looks like a single segment, or two segments
crossing at an angle. Below are diagrams of the trefoil, the gure-8 knot, and the Hopf link.
We will often consider knots and links equipped with a vector eld called a framing.
Precisely, if L is a link and p is any point on L, then we view the tangent space T
p
L as a
subspace of T
p
M

= R
3
. Let v be a smooth function L R
3
and denote the value of v at a
point p by v
p
.
Denition 3.1.2. We say that v is a framing for the link L if v
p
, T
p
L for all p L. A link
equipped with a framing is called a framed link.
In knot theory, one is interested in understanding when two knots or links are equivalent.
The appropriate notion of equivalence in this case is the following.
Denition 3.1.3. We say that two links L and L

are ambient isotopic or simply isotopic

if there exists a smooth map F : M [0, 1] M such that F(, 0) is the identity on M,
F(, 1) takes L to L

, and F(, t) is a dieomorphism for each t [0, 1]. If in addition L

and L

have orientation or framing, then we also insist that F(, 1) take the orientation or
framing of L to the orientation or framing of L

.
When doing computations in knot theory, we can transform a link diagram into one
representing an isotopic link by applying certain special operations at crossings.
13
Denition 3.1.4. The operations illustrated in the following three diagrams are called
Reidemester moves I, II, and III.

When dealing with links equipped with framing, we consider the following operation called
Reidemeister move I

Proposition 3.1.5. Two link diagrams of unframed links represent the same isotopy class if
we can get from one to the other by applying Reidemeister moves I, II, and III at crossings.
Two diagrams of framed links represent the same isotopy class if we can get from one to the
other by applying moves I

, II, and III.

3.2. Invariants. Now that we have dened isotopy of links, we can start to look for isotopy
invariants which will enable us to classify links up to isotopy. We begin by studying a simple
invariant of links equipped with an orientation. If we look at a diagram of such a link, then
any crossing looks like

or

and we call these two types of crossings right-handed and left-handed, respectively.
Denition 3.2.1. The writhe w(L) of a link L is dened to be the number of right-handed
crossings minus that number of left-handed crossings. That is,
w(L) = #
_

_
#
_

_
It is easy to see that the writhe is invariant under Reidemeister moves I

, II, and III, and

thus gives an invariant of framed links.
14
Denition 3.2.2. The Kauman bracket of a link L is the function L in the variables A,
B, and d determined by the following skein relations.
_ _
= 1
_ _
= d
_ _
_ _
= A
_ _
+B
_ _
Rather than attempt to explain in words the precise meaning of these relations, we illus-
trate the computation of the Kauman bracket with an example.
Example 3.2.3. The Kauman bracket of the Hopf link is
_ _
= A
_ _
+B
_ _
= A
2
_ _
+AB
_ _
+BA
_ _
+B
2
_ _
= (A
2
+ B
2
)d
2
+ 2ABd.
Theorem 3.2.4. The Kauman bracket is an isotopy invariant of framed links for B = A
1
and d = (A
2
+A
2
).
Proof. By Proposition 3.1.5, it suces to check that the Kauman bracket is invariant under
Reidemeister moves I

, II, and III. Indeed, we have invariance under move I

since
_ _
= A
_ _
+B
_ _
= Ad
_ _
+ B
_ _
= (Ad +B)
_ _
= A
_ _
+B
_ _
=
_ _
.
15
We also have AB = 1 and (A
2
+B(Ad +B)) = 0, and therefore
_ _
= A
_ _
+B
_ _
= A
_
A
_ _
+B
_ __
+ B(Ad +B)
_ _
= AB
_ _
+ (A
2
+B(Ad +B))
_ _
=
_ _
.
Hence the Kauman bracket is invariant under move II. Finally, it is invariant under move III
since
_ _
= A
_ _
+B
_ _
= A
_ _
+B
_ _
=
_ _
.
This completes the proof.
The proof of Theorem 3.2.4 shows in particular that the Kauman bracket is not invariant
under Reidemeister move I. Indeed, our proof of invariance under the modied Reidemeister
move I

shows that
_ _
= (Ad + B)
_ _
= A
3
_ _
,
and a nearly identical computation shows
_ _
= A
3
_ _
.
Although the Kauman bracket is not invariant under Reidemeister move I, these computa-
tions show that if we multiply it by the factor (A
3
)
w(L)
, we will have a Laurent polynomial
which is invariant under this move. This motivates the following denition.
Denition 3.2.5. Let L be an oriented link. Then the Jones polynomial associated to L is
V
L
(A) = (A
3
)
w(L)
L(A).
It is immediate from the above discussion that the Jones polynomial is an invariant of
unframed oriented links.
16
3.3. The Chern-Simons form. In the previous section, we formulated Yang-Mills theory
in terms of the Yang-Mills action and indicated its use in quantum electrodynamics. We
are now going to study a dierent eld theory called Chern-Simons theory. We will see that
this theory is similar to Yang-Mills theory, but instead of using it to describe a physical
system, we will talk about its applications to topology. Specically, we will discuss how
Chern-Simons theory computes topological invariants of knots embedded in M.
To begin, recall that we dened the Yang-Mills action to be the integral over M of the
Lagrangian density Tr(F F). Likewise in Chern-Simons theory, the action is obtained by
integrating a certain 3-form on an orientable 3-manifold M. To dene this 3-form, we will
assume that the gauge group G is simply connected. In this case every principal G-bundle
is trivial, and once we choose a trivialization, we can identify any connection on the bundle
with a g-valued 1-form A on M. Then the action is obtained by integrating the 3-form
Tr(A dA +
2
3
A A A).
which is known as the Chern-Simons form. Let us explain the notation in the above ex-
pression. As in Yang-Mills theory, we are assuming our Lie algebra g admits an invariant
inner product , . The rst term of the Chern-Simons form is obtained by wedging A
with dA to get a g g-valued 3-form, and then composing with the inner product to get an
ordinary real-valued 3-form which we can integrate over M. The second term should really
be thought of as
1
3
Tr([A, A] A)
where we rst construct the g g-valued 3-form [A, A] A and then compose with the inner
product to get a real-valued 3-form.
3.4. Holonomy. Suppose that : [0, T] M is a smooth path from p to q and for each
t [0, T] let u(t) be a vector in the ber of P
G
V over (t). Recall that the covariant
derivative of a section s can be written as partial derivatives plus a vector potential:

s =
n

i=1
(

s
i
)e
i
+ A

s.
We would also like to be able to dierentiate u(t) in the direction is going, namely

(t).
In analogy with the above formula, we dene

(t)
u(t) =
d
dt
u(t) + A(

(t))u(t).
Denition 3.4.1. We will say that u(t) is parallel transported along if

(t)
u(t) = 0 for
all t [0, T].
Intuitively, the statement that u(t) is parallel transported along means we can drag
the vector u(0) along , and u(t) is the resulting vector at (t). Suppose we are given a vector
u in the ber over p. By the familiar existence result for ordinary dierential equations, we
can nd a function u(t) which satises
d
dt
u(t) + A(

(t))u(t) = 0
17
with the initial condition u(0) = u. This means we can parallel translate u along any smooth
path. Let H(, )u = u(T) denote the result of parallel translating u along the path to
the point q. Since the dierential equation dening parallel transport is linear, the map
H(, ) : V V is a linear transformation.
Denition 3.4.2. The map H(, ) : V V is called the holonomy along the path .
Let us investigate the eect of a local gauge transformation on the holonomy. Suppose
u(t) satises the parallel transport equation

(t)
u(t) = 0.
In terms of a vector potential A, this equation says
d
dt
u(t) = A(

(t))u(t)
or in components
d
dt
u(t) =

(t)A

u(t).
Now a local gauge transformation g : M G takes u(t) to a function w(t) = g((t))u(t).
Dierentiating this new function, we obtain
d
dt
w(t) =
_
d
dt
g((t))
_
u(t) + g((t))
d
dt
u(t)
=

(t)(

g)u(t) g

(t)A

u(t)
=

(t)(

g)g
1
w(t)

(t)gA

g
1
w(t)
We are assuming that the gauge group G is a group of matrices, so the gauge transformed
vector potential A

can be written A

= gA

g
1
(

g)g
1
and therefore
d
dt
w(t) =

(t)A

w(t).
This proves that w satises the parallel transport equation

(t)
w(t) = 0 where

is the
connection obtained by applying the gauge transformation g to . By denition then, the
holonomy H(,

) maps g(0)u(0) to g(T)u(T), and we have

H(,

) = g((T))H(, )g((0))
1
.
3.5. Wilson loops. Let be a loop in M based at p. Then the transformation law that we
derived above for holonomy becomes
H(,

) = g(p)H(, )g(p)
1
and therefore the trace of the holonomy is invariant under gauge transformations:
Tr H(,

) = Tr(g(p)H(, )g(p)
1
) = Tr H(, ).
Denition 3.5.1. The trace of the holonomy along is called a Wilson loop and is denoted
W(, D) = Tr(H(, )).
18
3.6. Synthesis. Finally, we are going to relate Chern-Simons theory to the link invariants
that we introduced earlier in this section. Recall that in quantum eld theory any physical
quantity is represented by a function f : T(M) R on the space of elds, and the vacuum
expectation value of such a quantity is given by
f =
1
Z
_
F(M)
f()e
iS()
D.
In a theory with gauge symmetry, two elds are physically equivalent if they are related by a
gauge transformation, so in this setting we require the function f to be gauge invariant. We
have already seen that Wilson loops are invariant under gauge transformations, and indeed,
they are among the simplest observables in gauge theory. Our main claim in this section is
that the expectation values of Wilson loops in Chern-Simons theory are isotopy invariants of
links. More precisely, if L is a framed oriented link in M with components
1
, . . . ,
n
, then
the unnormalized expectation value
_
A
W(
1
, A) . . . W(
n
, A)e
ik
4
S(A)
DA.
is an invariant of L. Here k is an integer called the level, S denotes the Chern-Simons action,
and / denotes the space of all connections on M. As a special case, we can consider this
integral for the empty link, and we nd that the partition function
Z =
_
A
e
ik
4
S(A)
DA
is an invariant of the 3-manifold M. In order to turn these claims into completely precise
results, we would need to rigorously dene the measure DA appearing in the integrals and
make a number of technical modications to the statements. Although we will not attempt
to do this here, we emphasize that these integrals have been made precise in certain special
cases, and they are known to give rise to familiar invariants. For example, let /(L) denote our
alleged knot invariant above, and consider the case where M = S
3
and is the fundamental
representation of U(1). Then we have
/(L) = e
iw(L)/k
where w denotes the writhe of L. Alternatively, if we take to be the fundamental represen-
tation of SU(2), then /(L) is simply the Kauman bracket of L evaluated at A = q
1/4
where
q = e
2i
k+2
. By varying the gauge group of the theory, we obtain a wealth of other interesting
invariants.
4. Atiyahs Definition
4.1. Cobordisms. One of the major advantages of doing topology with TQFTs is the fact
that elds are dened locally. This means that we can compute invariants by breaking our
manifolds into simpler pieces and computing invariants on the pieces. In order to study
TQFTs axiomatically, we should therefore talk about how to break a manifold into simple
pieces, or equivalently, how to build up a closed n-manifold by gluing together simpler n-
manifolds along their (n 1)-dimensional boundaries.
19
The manifolds that we will use to build up more complicated manifolds by gluing are called
cobordisms. Roughly speaking, a cobordism is an n-manifold that connects together two
(n 1)-dimensional manifolds.
Denition 4.1.1. Let
0
and
1
be closed (n1)-manifolds. A cobordism between
0
and

1
is an n-dimensional manifold with boundary M =
0

1
.
Example 4.1.2. The illustrations below show two dierent ways to connect the manifold

0
= S
1

S
1
to the manifold
1
= S
1

S
1

S
1
by a two-dimensional cobordism.

$
$
$
$
As illustrated in the picture on the right, neither the
i
nor the cobordism M is required to
be a connected manifold.
Usually, we think of a cobordism as having a direction so that some of its boundary
components can be viewed as the source and the others can be viewed as the target of
the cobordism. To make this idea precise, consider an oriented manifold M with boundary,
and let be a connected component of the boundary of M. Given a point x , let
[v
1
, . . . , v
n1
] be a positive basis for the tangent space T
x
. Then a vector w T
x
M is
called a positive normal if [v
1
, . . . , v
n1
, w] is a positive basis for T
x
M. If a positive normal
points inward, then is called an in-boundary, while if it points outward, then is called
an out-boundary. (It is a nontrivial theorem in dierential topology that these notions do
not depend on the choice of the point x.)
Denition 4.1.3. Let
0
and
1
be closed oriented (n1)-manifolds. An oriented cobordism
M from
0
to
1
is a compact oriented manifold with a map
0
M taking
0
dieomor-
phically onto the in-boundary of M and a map
1
M taking
1
dieomorphically onto
the out-boundary of M.
For example, either of the manifolds illustrated in the above example can be regarded as
an oriented cobordism once we choose an orientation. When drawing pictures of oriented
cobordisms, we will always think of the cobordism as going from top to bottom so that the
boundary components near the top of the page form the in-boundary, and the ones near the
bottom of the page form the out-boundary.
Denition 4.1.4. We say that two cobordisms
0
M
1
and
0
M

1
are
equivalent if there exists an orientation-preserving dieomorphism : M M

making the
following diagram commute.

B
B
B
B
B
B
B
B

|
|
|
|
|
|
|
|
M

20
4.2. Composition of cobordisms. Since every oriented cobordism has a source and target,
it is natural to think of such cobordisms as the morphisms in a category whose objects are
(n1)-manifolds. Given a cobordism M
0
from
0
to
1
and a cobordism M
1
from
1
to
2
,
we would like to dene their composite to be the cobordism obtained by gluing M
0
to M
1
along the manifold
1
. The illustration below shows an example for n = 2.
g

f
=
gf
While the space M
0

1
M
1
that we obtain by gluing two cobordisms does have a natural
manifold structure, it is not immediately obvious that it can be given smooth structure. We
will now dene the smooth structure in two steps. The rst step is to dene the smooth
structure when the cobordisms being composed are of a special type.
Denition 4.2.1. A cylinder is a manifold of the form [a, b] where is a closed manifold
of dimension n 1.
We will always regard such a cylinder as an oriented cobordism from to itself. Let M
0
and
M
1
be cobordisms that are equivalent (in the sense of Denition 4.1.4) to cylinders
0
[0, 1]
and
1
[1, 2], the equivalences being given by
0
: M
0
[0, 1] and
1
: M
1
[1, 2].
Then there is a homeomorphism

2
: M
0

M
1
[0, 2]
The manifold [0, 2] has smooth structure which agrees with that of
0
[0, 1] and

1
[1, 2], so we can put a smooth structure on M
0

M
1
by pulling back the atlas of
[0, 2] along this homeomorphism.
The next step is to use the idea of the last paragraph to dene the smooth structure on
an arbitrary composition of oriented cobordisms. Here we require a technical notion from
the subject known as Morse theory.
Denition 4.2.2. Let M be a compact manifold and f : M [0, 1] a smooth map to the
closed unit interval. A critical point x of this function f is said to be nondegenerate if the
matrix

2
f
x
i
x
j
is nonsingular in any coordinate system, and f is called a Morse function if all of its critical
points are nondegenerate and f
1
(0, 1) = M.
A theorem of dierential topology states that Morse functions always exist, so if M
0
and M
1
are the oriented cobordisms we wish to compose, we can take Morse functions f
0
: M
0
[0, 1]
and f
1
: M
1
[1, 2]. By choosing > 0 to be small, we can ensure that f
0
and f
1
are regular
on [1 , 1] and [1, 1 + ]. Then the preimages of these two intervals are dieomorphic to
cylinders, and we know how to get a smooth structure on the composition.
21
4.3. Composition of cobordism classes. Finally, we show that the composition of two
cobordisms does not depend on the actual cobordisms chosen, but only on their equivalence
classes. For this we need the following result from dierential topology.
Proposition 4.3.1. Let M
0
and M
1
be composable cobordisms with common boundary
component , and let M
1
M
0
= M
0

M
1
be their composition. If and are two
smooth structures on M
1
M
0
which both induce the original smooth structure on M
0
and
M
1
, then there is a dieomorphism M
1
M
0
M
1
M
0
taking to .
Suppose we are given composable cobordisms M
0
and M
1
which are equivalent (rel the
boundary) to cobordisms M

0
and M

1
:

B
B
B
B
B
B
B
B

M
0

.|
|
|
|
|
|
|
|

B
B
B
B
B
B
B
B

M
1

.|
|
|
|
|
|
|
|
M

0
M

1
Then we have the composition M
1
M
0
and the composition M

1
M

0
, and the maps
0
and

1
glue to give a homeomorphism : M
1
M
0
M

1
M

0
which restricts to a dieomorphism
on each of the original cobordisms:

H
H
H
H
H
H
H
H
H

M
1
M
0

.v
v
v
v
v
v
v
v
v
M

1
M

0
We can use this homeomorphism to dene a smooth structure on M

1
M

0
. Although this
smooth structure may not coincide with the original, Proposition 4.3.1 implies that the two
structures are dieomorphic rel the boundary. Thus we see that there is a well dened way
to compose equivalence classes of oriented cobordisms.
4.4. The category of cobordisms. The notions introduced so far allow us to dene a
category Cob(n) which is an important ingredient in Atiyahs denition of a TQFT. The
objects of this category are closed oriented (n 1)-dimensional manifolds. Given two such
objects
0
and
1
, a morphism
0

1
is a dieomorphism class of oriented cobordisms
from
0
to
1
. It is simple to check that the composition law dened above is associative.
If M is a cobordism from
0
to
1
and C is the cylinder on
0
, then one can check that
M C = M up to equivalence, and similarly C M = M if C is the cylinder on
1
. This
proves that Cob(n) is in fact a category. This category is called the category of n-dimensional
cobordisms.
4.5. Some terminology. Before we get to the denition of a TQFT, we have to discuss
some ideas from category theory. In the discussion that follows, we will encounter denitions
involving many commutative diagrams, and we would like to use the notion of a natural
transformation to simplify these denitions. Recall that a natural transformation between
a functor F : ( T and a functor G : ( T consists of a morphism
X
: F(X) G(X)
for each object X of (. These morphisms are called the components of . In the following,
if F, G : (
1
(
n
T are functors and
X
1
,...,X
n
: F(X
1
, . . . , X
n
) G(X
1
, . . . , X
n
) are
22
morphisms, then these s are called natural isomorphisms or a natural transformation
if they are the components of a natural isomorphism or natural transformation F G.
4.6. Monoidal categories. The main reason for the present categorical digression is that
we need to equip our categories with an operation so that for any two objects X and Y we
can form the product XY . The example to keep in mind is the disjoint union operation
in the category of n-dimensional cobordisms. We want this operation to be unital and
associative so that our category is in some ways like a monoid. It is unnatural, however, to
require the operation to be strictly unital and associative. Instead, we would like for to be
unital and associative up to isomorphism, and the isomorphisms (XY )Z

= X(Y Z)
and 1X

= X

= X1 should be part of the structure, satisfying certain axioms called
coherence conditions.
Denition 4.6.1. A monoidal category is a category ( together with a functor : (( (,
a distinguished object 1 (, and natural isomorphisms
X,Y,Z
: (XY )Z X(Y Z),

X
: 1X X, and
X
: X1 X. These isomorphisms are required to make the diagram
(WX)(Y Z)
W(X(Y Z))
W((XY )Z) (W(XY ))Z
((WX)Y )Z

W,X,Y Z

O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
1
W

X,Y,Z

W,XY,Z

W,X,Y
1
Z

7
7
7
7
7
7
7
7
7
7
7
7

WX,Y,Z

o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
and the diagram
(X1)Y

X,1,Y

X
1
Y

M
M
M
M
M
M
M
M
M
M
X(1Y )
1
X

Y
.q
q
q
q
q
q
q
q
q
q
XY
commute for all objects X, Y , and Z.
It is easy to see that the above denition gives a categorical analog of the notion of a
monoid. In these notes, what we really need is a categorical analog of the notion of a
commutative monoid. We can get this by requiring our category to include isomorphisms
XY Y X that reverse the order of factors. These isomorphisms are part of the structure
and are required to satisfy additional coherence conditions.
23
Denition 4.6.2. A braided monoidal category is a monoidal category ( together with
natural isomorphisms
X,Y
: XY Y X, such that the diagrams
X(Y Z)

X,Y Z

(Y Z)X

Y,Z,X

(XY )Z

X,Y,Z

X,Y
1
Z

Y (ZX)
(Y X)Z

Y,X,Z

Y (XZ)
1
Y

X,Z

(XY )Z

XY,Z

Z(XY )

1
Z,X,Y

X(Y Z)

1
X,Y,Z

1
X

Y,Z

(ZX)Y
X(ZY )

1
X,Z,Y

(XZ)Y

X,Z
1
Y

commute for all X, Y , and Z. A braided monoidal category is called a symmetric monoidal
category if in addition we have
Y,X

X,Y
= 1
XY
.
Examples 4.6.3. The following symmetric monoidal categories arise frequently in category
theory and the study of TQFTs.
(1) The category Cob(n) or n-dimensional cobordisms with =

.
(2) The category Vect(k) of vector spaces over a eld k with =
k
.
(3) The category Set of sets with =

.
(4) The category Set of sets with = .
Now that we have dened symmetric monoidal categories and given some examples, we
can dene functors that preserve the structure of these categories.
Denition 4.6.4. Let ((, , 1
C
) and (T, , 1
D
) be two monoidal categories. A monoidal
functor between these categories is a functor F : ( T, together with a natural trans-
formation
X,Y
: FXFY F(XY ) and a morphism : 1
D
F1
C
such that the
diagram
(FXFY )FZ

X,Y
1
FZ

FX(FY FZ)
1
FX

Y,Z

F(XY )FZ

XY,Z

FXF(Y Z)

X,Y Z

F((XY )Z)
F

F(X(Y Z))
and the diagrams
FX1
D
1
FX

FX

FXF1
C

X,1
C

FX
F(X1
C
)
F
X

1
D
FY
1
FY

FY

F1
C
FY

1
C
,Y

FY
F(1
C
Y )
F
Y

commute for all X, Y , Z (. A monoidal functor between two symmetric monoidal cate-
gories is called a symmetric monoidal functor if, in addition to satisfying the axioms above,
24
the following diagram is commutative for all X and Y in (.
FXFY

FX,FY

X,Y

FY FX

Y,X

F(XY )
F
X,Y

F(Y X)
4.7. Topological quantum eld theories. We have now introduced the language of sym-
metric monoidal categories, and we are ready to talk about TQFTs. Recall that in order
to specify a quantum eld theory, we must specify a spacetime manifold M, a vector space
T(M) of elds on M, an action S : T(M) R, and a measure D. The goal of topological
eld theory is to construct a topological invariant of a closed manifold M from the action S
and measure D.
In fact, we will associate an invariant not only to closed manifolds, but to any n-dimensional
cobordism as well. Let
1
and
2
be closed (n 1)-dimensional manifolds, and let M be a
cobordism from
1
to
2
. Write Fun(T(
i
)) for the space of functions on T(
i
). Then the
invariant that we associate to M is the integral operator ZM : Fun(T(
1
)) Fun(T(
2
))
with kernel
K
M
(
1
,
2
) =
_
F(M)
|
1
=
1
,|
2
=
2
e
iS()
D.
As a special case, we can take M to be a closed n-manifold, regarded as a cobordism from
the empty (n 1)-manifold to itself. By convention, the vector space T() is the zero space
so that Fun(T())

= R. Thus ZM is a linear map R R, which is the same thing as
an element of R. In this way, our construction assigns a numerical invariant to any closed
oriented n-manifold.
To describe this invariant more explicitly, let M be a closed n-manifold, regarded as a
cobordism from
1
= to
2
= . We write 0

for the unique eld in T(

1
) and 0
+
for the
unique eld in T(
2
). For any f Fun(T(
1
)), we have
ZM(f)(0
+
) =
_

F(
1
)
f(

)
__
F(M)
|
1
=

,|
2
=0
+
e
iS()
D
_
D

=
__
F(M)
e
iS()
D
_
f(0

)
since the total measure of T(
1
) equals 1. The invariant that we are assigning to M is
therefore just the partition function
_
F(M)
e
iS()
D of the eld theory.
The reason for making our eld theory assign invariants to all n-dimensional cobordisms
and not just closed n-manifolds is that this enables us to compute invariants by breaking
our manifolds into simpler pieces and computing invariants on the pieces. For example, let

1
,
2
, and
3
be closed (n 1)-manifolds, and let M
1
be a cobordism from
1
to
2
, M
2
a cobordism from
2
to
3
. Then
(ZM
2
ZM
1
)(f)(
+
) =
_
F(
2
)
_

F(
1
)
f(

)K(

, )K(,
+
)D

D
25
Let M be the cobordism obtained by gluing M
1
and M
2
along
2
. We know that if is a eld
on M and
1
= [M
1
,
2
= [M
2
, then the action satises S() = S(
1
)+S(
2
). Moreover,
since elds are dened locally, we know that such a eld is completely determined by its
restriction to the manifolds M
i
and the
i
. Therefore
_
F(
2
)
K(

, )K(,
+
)D
=
_
F(
2
)
__

1
F(M
1
)

1
|
1
=

,
1
|
2
=
e
iS(
1
)
D
1
___

2
F(M
2
)

2
|
2
=,
2
|
3
=
+
e
iS(
2
)
D
2
_
D
=
_
F(M)
|
1
=

,|
3
=
+
e
i(S(|M
1
)+S(|M
2
))
D
=
_
F(M)
|
1
=

,|
3
=
+
e
iS()
D
= K(

,
+
).
Plugging this result back into the previous equation gives Z(M
2
M
1
) = ZM
2
ZM
1
. We
can express this fact more abstractly by saying that a topological quantum eld theory is a
functor from Cob(n) to Vect(k) which assigns the vector space Fun(T()) to an object
and assigns the linear map ZM to a cobordism M.
In addition, it is possible to convince oneself that a topological quantum eld theory
should preserve the monoidal structure of these categories. Locality of elds implies that
T(
1

2
) = T(
1
) T(
2
), and it follows that Fun(T(
1

2
))

= Fun(T(
1
))
Fun(T(
2
)). One can argue, using the sort of heuristic arguments involving Feynman inte-
grals that we have been using, that Z(M
1

M
2
) = ZM
1
ZM
2
. We are thus led to make
the following denition.
Denition 4.7.1. An n-dimensional topological quantum eld theory is a symmetric monoidal
functor Cob(n) Vect(k).
We should emphasize at this point that the above computations involving Feynman
integrals are not really well dened because we do not know how to dene the measure D.
The only reason for including them in the discussion is that they provide some physical
motivation for the denition of a TQFT. On the other hand, Denition 4.7.1 is completely
precise, and it provides a rigorous framework for constructing quantum invariants of spaces.
4.8. Dijkgraaf-Witten theory. A simple example of a TQFT in the sense of Atiyah is
the theory introduced by Dijkgraaf and Witten. This example is similar in some ways to a
gauge theory but has a well dened partition function. Precisely, we let G be a nite group
and M any n-dimensional cobordism. We take T(M) to be the set of all isomorphism classes
of principal G-bundles on M. If M has boundary M =
0

1
, then the kernel dening
ZM : Fun(T(
1
)) Fun(T(
2
)) is
K
M
(
0
,
1
) =

F(M)
|
0
=
0
,|
1
=
1
1
[ Aut()[
26
where Aut() is the nite group of automorphisms of . Then
ZM(f) =

0
F(
0
)
K
M
(
0
,
1
)f(
0
).
Since the group G is nite, it follows that the sums in these expressions are nite and there
are no problems of convergence. One can easily check that these data dene a TQFT in the
sense of Denition 4.7.1.
5. Reshetikhin-Turaev Invariants
5.1. Semisimple categories. In this section, we will see how to construct interesting quan-
tum invariants starting from certain categories. These categories, known as modular tensor
categories, come equipped with a lot of of extra structure, and we will spend much of this
section dening the extra structures. We begin by recalling the following denition (for
details see [4]).
Denition 5.1.1. Let k be a eld. A category ( is additive over k if
(1) Each set hom-set in ( is a k-vector space and the compositions are k-bilinear.
(2) There exists an object 0 ( such that Hom
C
(0, V ) = Hom
C
(V, 0) = 0 for every
object V (.
(3) ( has nite direct sums.
An additive category ( is said to be abelian if it has the following additional property:
(4) Every morphism in ( has a kernel and cokernel and is the composition of an
epimorphism followed by a monomorphism. If ker = 0 then = ker(coker ), and
if coker = 0 then = coker(ker ).
In an abelian category (, it makes sense to say to that an object U is simple if any injection
V U is either 0 or an isomorphism. We say that an abelian category is semisimple if every
object is isomorphic to a direct sum of simple ones. If an additive category is also monoidal,
we will require that the monoidal structure be bilinear and that the unit object 1 be
simple and End(1)

= k. In this section, we will assume that all categories are semisimple
with the property End(V )

= k for simple V .
5.2. Braids and braided monoidal categories. Recall that in the previous section, we
dened a braided monoidal category to be a category with the extra structures , 1, , ,
, and satisfying certain coherence conditions. We will now discuss how these categories
are related to three-dimensional topology.
Denition 5.2.1. A braid in n strands is the isotopy class of a union of n nonintersecting
segments of smooth curves in R
3
called strands with endpoints in 1, . . . , n 0 0, 1
such that for each segment the third coordinate is strictly decreasing from 1 to 0.
27
Example 5.2.2. The following diagrams represent equivalent braids in four strands.
If we are given two braids in n strands, we can compose them by putting one on top of the
other. This operation turns the set of all braids in n strands into a group called the braid
group and denoted B
n
.
Let ( be a braided monoidal category, and let V
1
, . . . , V
n
be objects of (. Consider the
expressions obtained by forming the tensor product V
i
1
V
i
n
where (i
1
, . . . , i
n
) is a
permutation of (1, . . . , n) and then inserting parentheses and factors of 1. If X
1
, X
2
are
objects constructed in this way and : X
1
X
2
is any isomorphism obtained by composing
, , , , and their inverses, then we can associate to an element of B
n
as follows: To
each factor of , , and , we assign the identity braid
and to each factor of
V
i
V
i+1
, we assign the braid which interchanges the ith and (i + 1)st
strands
One then has the following result.
Proposition 5.2.3. The morphism depends only on its image in the braid group B
n
.
For a proof of this theorem, see [4].
5.3. Ribbon categories. Just as the notion of braided monoidal category is related to the
topology of braids, there is a notion of ribbon category, which is related to the topology of
braids equipped with a framing.
Denition 5.3.1. Let ( be a monoidal category, and let V be an object of (. A right dual
to V is an object V

of ( together with morphisms e

V
: V

V 1 and i
V
: 1 V V

such that the compositions

V
i
V
1
V

V V

V
1
V
e
V

V
and
V

1
V
i
V

V V

e
V
1
V

28
are identities. Here and throughout this section we skip the canonical associativity and unit
isomorphisms so that, for example, the morphism V V V

V above is really the

composition
V

1
V

1 V
i
V
1
V

(V V

) V

V (V

V ).
A left dual to V is dened to be an object

V of ( together with morphisms e

V
: V

V 1
and i

V
: 1

V V satisfying similar axioms. Finally, a monoidal category ( is said to be
rigid if every object in ( has right and left duals.
Denition 5.3.2. A ribbon category is a rigid braided monoidal category with natural
isomorphisms
V
: V V

satisfying

V W
=
V

W

1
= 1

V
= (

V
)
1
for all objects V and W.
In any rigid braided monoidal category (, there is a natural isomorphism
V
: V

V
given by the composition
V

i1

V V

1
1

V V

1e

V
and thus if ( is a ribbon category, we can dene a natural isomorphism
V
=
V

V
: V V .
This operation is called a balancing isomorphism or twist, and can be shown to satisfy the
following axioms:

V W
=
WV

V W
(
V

W
)

1
= 1

V
= (
V
)

.
For any object V in a ribbon category ( and any endomorphism f of V , we dene the
trace Tr f End
k
(1)

= k to be the composition
1
i
V

V V

f1

V V

V
1

e
V

1.
In particular, for f = 1
V
, this lets us dene the dimension of V by dimV = Tr 1
V
.
5.4. The graphical calculus. Let ( be a ribbon category. A useful way of proving theorems
about about ( is to assign a diagram to each morphism and then argue pictorially using
the diagrams. In this scheme, if f : V W is any morphism in (, then the diagram
corresponding to f looks like
f

W
29
The identity morphism can also be written as

V
1
V
.
= V
where the symbol
.
= means that the two diagrams represent the same morphism in (. Also,
if f : U V and g : V W are two morphisms, the diagram corresponding to their
composition can be written

W
gf
.
=
U
V
f

W
g
If V and V

are objects in ( then the diagram corresponding to their tensor product is

obtained by drawing the corresponding arrows parallel to one another. If f : V W and
f

: V

are morphisms, then the diagram corresponding to their tensor product can
be written

V V

WW

ff

.
= f

In our diagrams, arrow corresponding to the dual V

of an object V is obtained by reversing

the arrow corresponding to V . We will also omit the arrows corresponding to 1 ( and
identify V

with V via the isomorphism

V
. Then the diagrams for e
V
: V

V 1 and
i
V
: 1 V V

can be written as
e
V

V
/
/
/
/
/

1
.
=

and
i
V

1
/
/
/
/
/

V
.
=

Finally, the diagrams corresponding to
V V
and
1
V V

can be written as

V V

/
/
/
/

V
/
/
/
/

.
=


30
and

1
V V

V
/
/
/
/

/
/
/
/

V
.
=


5.5. Ribbon tangles. In addition to providing an intuitive graphical notation, the graphical
calculus introduced above actually helps us study the topology of certain braid-like objects.
Let us now dene precisely what these objects are.
Denition 5.5.1. A tangle is the isotopy class of a union of nonintersecting smooth curves
in R
2
[0, 1], which can have endpoints only on the lines R0 0 and R0 1.
If none of the curves have endpoints, then the tangle is simply a link in R
2
[0, 1].
This denition gives rigorous meaning to the pictures used above to represent morphisms
in a ribbon category. It is natural to ask whether, given two morphisms, the corresponding
tangles are distinct. To see that the answer is no, we observe that
V
,= 1
V
in a general
ribbon category and yet the diagram for
V
is

V
.
=

which is clearly isotopic to the picture of 1

V
. We must therefore modify the notion of a
tangle so that distinct morphisms correspond to distinct tangles.
Denition 5.5.2.
(1) A ribbon is a homeomorphic image of a rectangle in R
3
with a distinguished pair of
opposite edges and a distinguished side called the face side.
(2) A coupon is a rectangle in R
3
lying in a plane parallel to R 0 R and having
edges parallel to R 0 0 and 0 0 R.
(3) A generalized ribbon tangle with n strands is the isotopy class of a union of nonin-
tersecting ribbons and coupons in R
2
[0, 1] such that the distinguished edges of the
ribbons lie on the lines R0 0 and R0 1 or on the edges of coupons
parallel to them, and near them the ribbons have their face sides turned upwards.
Consider a generalized ribbon tangle in which each ribbon strand is directed. Label each
ribbon in this generalized ribbon tangle by an object of (, and for any xed coupon, let
V
1
, . . . , V
m
be the labels of the ribbons that end on its top edge. We write
i
= + if the ith
ribbon points towards the coupon and
i
= if it points away from the coupon. Then we
can form the object X = V

1
1
V

m
m
where V
+
= V and V

= V

. (If no objects end

on the top edge of this coupon, then X = 1.) Similarly, we can associate an object Y to the
bottom edge of a coupon, where now we must write = + for ribbons that point away from
the coupon.
31
Denition 5.5.3. A (-colored ribbon tangle is a generalized ribbon tangle in which each
strand is directed, each ribbon is labeled by an object of (, and each coupon is labeled by
a morphism f : X Y where X and Y are the objects associated to its top and bottom
edges by the above construction.
We now use the graphical calculus to associate (-colored ribbon tangles to a morphism
in (. Let us rst x V
1
, . . . , V
n
( and consider the objects obtained by tensoring V
1
, . . . , V
n
in any order, possibly with repetitions, and adding an arbitrary number of left and right
stars and factors of 1. To each expression obtained in this way, we associate a sequence
F(X) of arrows and labels in the following way: To each object

V

, we assign the data

V if the total number of stars is even and V if the total number of stars is odd. For
example, we associate to the object
((V
1
V
2
) V
4
) ((V

1
1)

V
2
) . . .
the sequence
V
1
V
2
V
4
V
1
V
2
. . . .
Let X
1
and X
2
be objects of ( of the above form. Then we can consider all morphisms
: X
1
X
2
which can be obtained as a composition of the elementary morphisms
1
,

1
,
1
,
1
, e, i,
1
as well as a number of other morphisms of (. To such a composition,
we associate a generalized ribbon tangle T = F() in such a way that F(X
1
) is the top of
T and F(X
2
) is the bottom of T. We do this by letting the morphisms , , , and their
inverses correspond to trivial tangles and letting F(e), F(i), F(), and F(
1
) correspond
to the tangles of the graphical calculus, with the blackboard framing. Given two morphisms

1
and
2
in (, we get F(
1

2
) and F(
1

2
) by combining the tangles F(
1
) and F(
2
)
in the obvious way.
Proposition 5.5.4. A morphism : X
1
X
2
as above depends only on the isotopy class
of the tangle F().
This result implies that for every (-colored ribbon tangle T, the morphism F
1
(T) : X
1

X
2
in ( is an isotopy invariant of T. In particular, if T is a (-colored framed link, then we
can view F
1
(T) as a numerical invariant of T.
5.6. Modular tensor categories. The last notion from category theory that we will need
is the notion of a modular tensor category. Let ( be a semisimple ribbon category and let I
be the set of equivalence classes of nonzero simple objects in (. For each i I, let V
i
be a
representative for the equivalence class i. Then we can dene numbers s
ij
k

= End 1 by
the formula
s
ij
=
1
i

1
j
Tr
V

i
V
j
Denition 5.6.1. Let ( and I be as above. We say that ( is a modular tensor category if
( has the following properties:
(1) ( has only nitely many isomorphism classes of simple objects (so [I[ < ).
(2) The matrix s = ( s
ij
)
i,jI
is invertible.
32
In applications to topology, the most interesting and nontrivial examples of modular tensor
categories are categories of representations of quantum groups at roots of unity. We refer
the reader to [6] for more information.
In the following discussion, we will use the notation p

iI

1
i
d
i
where d
i
= dimV
i
and D =

p
+
p

.
5.7. Invariants of 3-manifolds. Finally, let us show how modular tensor categories give
rise to quantum invariants of 3-manifolds. To do this, we need to discuss how 3-manifolds
can be obtained by surgery along links. Therefore let L be a link in S
3
with components

1
, . . . ,
n
and denote by T
i
a small tubuler neighborhood of
i
. Then each T
i
is a solid
torus. Let T
0
be a xed solid torus and choose orientation-preserving homeomorphisms
f
i
: T
i
T
0
. These mappings give rise to an orientation-preserving homeomorphism
f :
n
_
i=1
T
i

n

i=1
T
0
.
Denition 5.7.1. A surgery of S
3
along the link L is a 3-manifold M
L,f
dened by the
formula
M
L,f
=
_
S
3

n
_
i=1
T

i
_

f
_
n

i=1
T
0
_
.
Proposition 5.7.2. Any connected closed three-dimensional manifold can be obtained as a
surgery of S
3
along some link.
In the denition of surgery, M
L,f
depends on both the link L and the attaching map f.
However, if L happens to be a ribbon link, then there is a canonical choice of f and so in
this case we write M
L,f
= M
L
.
Theorem 5.7.3. Let ( be a modular tensor category, and let L be a framed link in R
3
S
3
.
Then the number Z(M
L
) dened by the formula
Z(M
L
) = D
|L|1
F
1
(L)
_
p
+
p

_
(L)/2
,
where [L[ is the number of components of L and (L) is the wreath number, is an invariant
of the 3-manifold M
L
which comes from a topological quantum eld theory.
Acknowledgments
Much of the material in these notes was originally presented in a seminar at Yale University
in the Spring semester of 2012. I thank Giovanni Faonte, Ivan Ip, Hyun Kyu Kim, and Linhui
Shen for participating in the seminar and helping me learn the subject.
References
[1] Figueroa-OFarrill, J. Unpublished notes on gauge theory. 2007.
[2] Baez, J. and Muniain, J.P. Gauge Fields, Knots and Gravity. World Scientic, 1994.
[3] Kock, J. Frobenius Algebras and 2D Topological Quantum Field Theories. Cambridge University Press,
2003.
33
[4] Mac Lane, S. Categories for the Working Mathematician. Springer-Verlag, 1988.
[5] Segal, G. Unpublished notes on topological quantum eld theory. 1999.
[6] Bakalov, B. and Kirillov, A. Lectures on Tensor Categories and Modular Functors. American Mathemat-
ical Society, 2000.
34