TQFTs

[early incomplete draft]

version 1h

Kevin Walker

May 11, 2006

ii
Contents

0 Introduction 1

1 From Path Integrals to Local Relations 3

1.1 Path Integrals and Projections . . . . . . . . . . . . . . . . . . . . . 3
1.2 Functorial Properties of the Path Integral . . . . . . . . . . . . . . . 5
1.3 Locality of Z(Y ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Z2 Homology as a TQFT 9
2.1 The Basic Construction . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Gluing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Cylinder Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Semisimplicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Topological Fields 17
3.1 Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Topological Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Extended Isotopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4 Basic Constructions and Gluing I 25

4.1 Local Relations and Local Subspaces . . . . . . . . . . . . . . . . . . 25
4.2 The Basic Construction in Codimension 1 . . . . . . . . . . . . . . . 26
4.3 Cylinder Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Gluing Codimension-1 Manifolds . . . . . . . . . . . . . . . . . . . . 30

5 Basic Constructions and Gluing II 37

5.1 2-Category Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Gluing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

6 From Local Relations to Path Integrals 55

6.1 Axioms for Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2 Consequences of the Axioms . . . . . . . . . . . . . . . . . . . . . . . 58
6.3 Computing the Path Integral . . . . . . . . . . . . . . . . . . . . . . 61

iii
iv CONTENTS

7 1+1-dimensional Examples 67
7.1 Generalities on 1+1-dimensional Theories . . . . . . . . . . . . . . . 67
7.2 Finite Group Theories in 1 + 1 dimensions . . . . . . . . . . . . . . . 67

8 2+1-dimensional Examples 69
8.1 Temperley-Lieb Theories . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.2 Spherical Categories in General . . . . . . . . . . . . . . . . . . . . . 69
8.3 A2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8.4 G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8.5 Finite Group Theories in 2+1 dimensions . . . . . . . . . . . . . . . 74
8.6 Jones Planar Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 74

9 3+1-dimensional Examples 75
9.1 Theories From Ribbon Categories . . . . . . . . . . . . . . . . . . . . 75
9.2 Decategorification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.3 More on Chern-Simons Theories . . . . . . . . . . . . . . . . . . . . 84
9.4 Contact Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

10 4+1-dimensional Examples 87
10.1 Theories From Khovanov Homology . . . . . . . . . . . . . . . . . . 87

11 n+1-dimensional Examples 89
11.1 Finite Group Theories . . . . . . . . . . . . . . . . . . . . . . . . . . 89
11.2 Finite Total Homotopy Theories . . . . . . . . . . . . . . . . . . . . 89
11.3 String Category . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

12 Reconstruction Results 91

13 [Other Chapters] 93

A Categories 95
A.1 Definitions and Notation . . . . . . . . . . . . . . . . . . . . . . . . . 95
A.2 [Still to do] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

B Semisimple Categories 97
B.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
B.2 Conditions Defining Semisimplicity . . . . . . . . . . . . . . . . . . . 98
B.3 The Density Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 99
B.4 Semisimple Categories . . . . . . . . . . . . . . . . . . . . . . . . . . 101
B.5 To Do List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Chapter 0

Introduction

Notes on this early, incomplete draft.

• There are some inconsistencies, since some of the chapters were salvaged from
an earlier version.

• I need to go through and fix details re dual fields, reversing orientation of

manifolds, etc. (This stuff was deliberately omitted in the first draft to save
time.) (At present this has been done in some chapters but not others.)

• need to add references and bibliography

• some parts have not been proof read

• my notational preferences evolved as this was written, so there are some no-
tational inconsistencies

Notes for this introduction (to-do list):

• give outline of all the chapters

• acknowledgements (many or few?)

• this is partly expository, partly new stuff

• key idea: Computing the path integral by actually integrating against some
measure on the space of fields is notoriously difficult, so to develop a mathe-
matically rigorous theory we need to take a detour around this obstacle. Most
previous approaches follow Segal and Atiyah [need refs] and take a rather
wide detour: a TQFT is defined (roughly) to be any choice of vector spaces
for n-manifolds and vectors for n+1-manifolds which satisfy algebraic prop-
erties suggested by the path integral. Here we make a tighter detour around
the path integral. We retain the notion of fields on manifolds and require

1
2 CHAPTER 0. INTRODUCTION

that the vector spaces for n-manifolds be constructed out of these fields by
local relations (or dually local projections). These local relations carry the
same information as the path integral of the n+1-dimensional ball B n+1 with
all possible boundary conditions. One advantage of this approach is that the
higher algebra / higher codimension / multi-tier structure [cite Freed (and also
xxxx?)] is automatically present all the way down to dimension zero. Another
advantage is that we can construct most (all?) known TQFTs in a uniform
framework.

• I’ve tried to give good citations, but there are bound to be inadvertant omis-
sions (comments welcome) [this is for the future — there are very few citations
at the moment]

• hard to decide what order to put things; some readers might want to skip
ahead to examples before reading all of the general theory

• I’ve gone part way toward making things general in a category-theoretic sense,
but not all the way. (e.g. I usually assume the target category for the TQFT is
complex vector spaces, but the proofs for the most part work more generally.)

• need to comment on references; should I add historical notes at the end of

chapters?
Chapter 1

From Path Integrals to Local

Relations

In this chapter we give a non-rigorous argument showing that topologically invariant

path integrals are more or less equivalent to the systems of fields and local relations
described in detail (and rigorously) in the Chapter 4. We’ll pretend that the integrals
make sense and can be manipulated in the usual ways, and see that this leads directly
to local relations. In turns out that fields and local relations make a good foundation
for a rigorous treatment of TQFTs, especially the high codimension parts of the
theory. In Chapter 6 we will show that in many cases [be more specific here?] the
full path integral can be recovered combinatorially from the local relations.

1.1 Path Integrals and Projections

Fix a “top dimension”, which we denote n + 1 for historical reasons.
Let C be a system of fields for manifolds of dimension less than or equal to n + 1.
Since this is a non-rigorous argument, it’s not worth the trouble to define precisely
what a system of fields is. Standard examples are C(X) is the set of all maps from
X to some fixed space B, or C(X) is the set of all equivalence classes of G-bundles
with connection over X. In both cases, if X has boundary then we usually fix a
boundary condition c ∈ C(∂X) and let C(X; c) denote the set of all fields which
restrict to c on the boundary. We give formal axioms for fields in Chapter 3.
For each n+1-manifold M and each c ∈ C(∂M ), assume we have a function

T : C(M ; c) → T.

Here T ⊂ C denotes the the circle group. (In more traditional notation, T = eiS
for some S : C(M ; c) → R. S is called the action for the theory.) Assume that T is
local, in the sense that it is multiplicative (i.e. S is additive) with respect to disjoint
unions and it commutes with gluing. In other words, we can cut M into pieces and

3
 4 CHAPTER 1. FROM PATH INTEGRALS TO LOCAL RELATIONS

a
d
d

c
c
M Mgl

1.1.1 Fields on glued manifold

compute T on C(M ) in terms of T on C of the pieces. We further assume that T

is invariant under the action of boundary-fixing homeomorphisms on C(M ; c). In
other words, T is topologically invariant.
Now assume that we have defined some sort of integral and class of integrable
functions on C(M ; c), and that T is in this class. Define the path integral
Z
Z(M ; c) = T (x) ∈ C.
x∈C(M ;c)

Suppose ∂M = Y ∪ −Y ∪ W and let Mgl denote M glued to itself along ±Y .

Let c ∈ C(∂Mgl ) = C(Wgl ) and let c also denote the corresponding field on W . Let
d ∈ C(∂Y ) be the restriction of c to ∂Y ⊂ ∂W . (See Figure (1.1.1).) Assume that
we can also integrate a class of functions on fields on n-manifolds, and that
Z Z Z
Z(Mgl ; c) = T (x) = T (y).
1.1.2 x∈C(Mgl ;c) a∈C(Y ;d) y∈C(M ;a,a,c)

In other words, assume that we can “integrate along fibers”.

Let Y be an n-manifold and d ∈ C(∂Y ). Consider the n+1-manifold Y × I
(pinched at ∂Y ×I if Y has boundary, so that ∂(Y ×I) = Y ∪−Y ). For a, b ∈ C(Y ; d)
define Z
KY (a, b) = Z(Y × I; a, b) = T (x).
x∈C(Y ×I;a,b)

Since (Y × I) ∪Y (Y × I) ∼
= Y × I, we have, by (1.1.2),
Z
KY (a, y)KY (y, b) = KY (a, b).
1.1.3 y∈C(Y ;d)
1.2. FUNCTORIAL PROPERTIES OF THE PATH INTEGRAL 5

Let F(Y ; d) denote an appropriate space of functions C(Y ; d) → C. Define πY :

F(Y ; d) → F(Y ; d) by
Z
πY (f )(b) = f (y)KY (y, b).
y∈C(Y ;d)

Then it follows from (1.1.3) that πY is a projection,

πY2 = πY .

Finally, define Z(Y ; d) ⊂ F(Y ; d) to be the image of this projection,

Z(Y ; d) = πY (F(Y ; d)). 1.1.4

1.2 Functorial Properties of the Path Integral

For an n+1-manifold M , let Z(M ) be the function

Z(M ) : C(∂M ) → C
c 7→ Z(M ; c).

Then clearly
Z(M ) ∈ Z(∂M ). 1.2.1
(Proof: Glue a collar of ∂M to M and apply (1.1.2). The result is still M .) It
is also easy to see that for any closed n-manifolds Y1 and Y2 , there is a natural
identification
Z(Y1 ⊔ Y2 ) = Z(Y1 ) ⊗ Z(Y2 ),
and that for any n+1-manifolds M1 and M2

Z(M1 ⊔ M2 ) = Z(M1 ) ⊗ Z(M2 ) ∈ Z(∂M1 ) ⊗ Z(∂M2 ).

Integration gives a pairing

Z(Y ) ⊗ Z(−Y ) → C
Z
f ⊗ g 7→ f (x)g(x).
x∈C(Y )

If ∂M = Y ⊔ −Y ⊔ W , this pairing induces a trace map

Y tr
Z(∂M ) = Z(Y ) ⊗ Z(−Y ) ⊗ Z(W ) −→ Z(W ) = Z(∂Mgl ),

and
Z(Mgl ) = trY (Z(M )).
6 CHAPTER 1. FROM PATH INTEGRALS TO LOCAL RELATIONS

E
x x E
D
Y′
′
Y a′ D a′
∂Y ∂Y
a

1.3.1 Disk in Y (two cases)

The above constitutes a gluing (without corners) formula for n+1-manifolds.

Similarly, if Y is an n-manifold with boundary and d ∈ C(∂Y ), there is a pairing
Z(Y ; d) ⊗ Z(−Y ; d) → C
Z
f ⊗ g 7→ f (x)g(x).
x∈C(Y ;d)

If ∂M = Y ∪ −Y ∪ W (setup for gluing with corners [include figure?]), then for each
c ∈ C(∂Y ) there is a trace map
trY,d
Z(Y ; d) ⊗ Z(−Y ; d) ⊗ Z(W ; d, d) −→ Z(W ; d, d).
These maps (for each d) combine to determine a map
Y tr
Z(∂M ) −→ Z(∂Mgl ),
(see (6.1.4)) and again
Z(Mgl ) = trY (Z(M )).
The above constitutes a gluing (with corners) formula for n+1-manifolds.

1.3 Locality of Z(Y )

We want to show that Z(Y ; d) is defined by local relations. Let D ⊂ Y be in
n-ball (possibly intersecting ∂Y ), and let Y ′ be the closure of Y r D (see Figure
(1.3.1)). Let E = ∂D ∩ Y ′ . Let a ∪∂E a′ ∈ C(∂Y ), where a ∈ C(∂D ∩ ∂Y ; ∂a) and
a′ ∈ C(∂Y ′ ∩ ∂Y ; ∂a). (If D is in the interior of Y then a is the unique field over the
empty n−1-manifold.) Then
[
C(Y ; a ∪ a′ ) = C(D; a, x) × C(Y ′ ; a′ , x).
x∈C(E;∂a)

Restriction then gives an identification

Y
F(Y ; a ∪ a′ ) = F(D; a, x) ⊗ F(Y ′ ; a′ , x).
x∈C(E;∂a)
1.3. LOCALITY OF Z(Y ) 7

D×I
D×I D×I D×I D×I D×I

1.3.2 Decomposing a collar into local collars

(This is a product rather than a direct sum because non-zero components are allowed
for infinitely many x.) Define PD : F(Y ; a ∪ a′ ) → F(Y ; a ∪ a′ ) by
Y
PD = π(D;x) ⊗ id .
x∈C(E;∂a)

(Here we have suppressed a and a′ from the notation on the left hand side.) By the
same argument as before, PD2 = PD . PD is the local projection corresponding to
gluing a collar D × I to D ⊂ Y .
Now the key point: Y × I can be constructed by gluing Di × I to Y × {0}, where
Di runs through an open cover of Y . (See Figure (1.3.2).) It follows that πY can
be written as composition of the PDi ’s. This implies that
\
im(πY ) = im(PDi ).
Di

(Note that πY and the PDi ’s all commute, because the corresponding topological
gluings commute.)
Requiring that f ∈ im(PD ) is a local condition, in the sense that it only depends
on what’s going on inside a disk in Y . Thus we have obtained a local description of
im(πY ).
Dually, we can define y ∼ z if PD (y) = PD (z), where y and z are linear combina-
tions of fields on D. These are the local relations with which we will be concerned
for the rest of the book.
In practice, it turns out to be easier to define systems of local relations than it
is to define a path integral, so in the Chapter 4 we axiomatize these local relations.
Note that PD encodes the same information as Z(D × I) = Z(B n+1 ), so speci-
fying a local relation is tantamount to specifying the path integral of the n+1-ball.
8 CHAPTER 1. FROM PATH INTEGRALS TO LOCAL RELATIONS
Chapter 2

Z2 Homology as a TQFT

[kill this chapter?? is it needed?]

In this chapter we use the Z2 (= Z/2Z) homology of a surface to introduce the
key concepts of local relation, cylinder category, and gluing theorem. All of this is
redone in greater generality in Chapter 4, so some readers may want to skip directly
to that chapter. This chapter can be thought of as a warm-up for Chapter 4.
The proofs in this chapter are overkill if the goal is merely to investigate Z2
homology, but they have the virtue of generalizing to many other contexts.
More specific (less general) generalizations of this “Z2 -homology theory” can
be found in Section 8.1, where we consider a larger class of local relations on 1-
submanifolds, and Section 8.5, where we replace Z2 with an arbitrary finite group.

2.1 The Basic Construction

Let Y be a closed, oriented 2-manifold without boundary. Consider the set C(Y ) of
all unoriented 1-dimensional submanifolds of C of Y (including the empty subman-
ifold), modulo the equivalence relation generated by:

(a) C ∼ C ′ if C and C ′ are related by an isotopy supported in a disk;

(b) C ∼ C ′ if C ′ is obtained from C by removing a small trivial circle; and

(c) C ∼ C ′ if C and C ′ are related by an elementary homology contained in a

disk.

This relation is shown schematically in Figure (2.1.1). Here and throughout we

adopt the well-known convention that relations illustrated on disks apply to all
curves (or fields or whatever) which agree outside of the disk in the figure.
For our purposes, the most important thing about this equivalence relation is
that it is local, in the sense that it is supported within a disk.

9
10 CHAPTER 2. Z2 HOMOLOGY AS A TQFT

(a) ∼ (b) ∼

(c) ∼

2.1.1 Local relations for Z2 homology

Of course, C(Y )/∼ is naturally isomorphic, as a set, to H1 (Y ; Z2 ). But we will

pretend that we know nothing of homological algebra and instead rely mainly on
the fact that ∼ is a local relation.
e ) be the vector space consisting of all finite formal C-linear combina-
Let A(Y
e )/∼. A(Y ) is a vector space of
tions of elements of C(Y ), and define A(Y ) = A(Y
dimension |H1 (Y ; Z2 )|.

We would like to be able to calculate A(Y ) and construct an explicit basis. To

do that we will need a gluing theorem, and to state the gluing theorem we will
need to extend the above constructions to 2-manifolds with boundary, which we
now proceed to do.
If S is a 1-manifold, define C(S) to be the set of unoriented, properly embedded
0-dimensional submanifolds of S (including the empty submanifold). Let Y be a
compact, oriented 2-manifold with boundary, and let a ∈ C(∂Y ). Define C(Y ; a) to
be the set of all unoriented 1-dimensional submanifolds of C of Y such that ∂C = a.
e ; a), ∼, and A(Y ; a) as above, with all isotopies fixing ∂Y .
Define A(Y

2.2 Gluing
Let Y be a 2-manifold with gluing data: ∂Y is decomposed into three disjoint
pieces, ∂Y = S+ ⊔ S− ⊔ S0 , and there is an identification S+ = −S− (i.e. S+
is identified with S− with reversed orientation). (See Figure (2.2.1).) Let Ygl
denote the 2-manifold obtained by gluing S+ to S− . Note that ∂Ygl = S0 . Given
(a, b, c) ∈ C(S+ )×C(S− )×C(S0 ) = C(∂Y ), we have the vector space A(Y ; a, b, c). We
would like to calculate A(Ygl ; c) in terms of the vector spaces A(Y ; a, b, c) for various
a and b. (c ∈ C(S0 ) = C(∂Ygl ) will be fixed throughout the following discussion.)
2.2. GLUING 11

S+ S−

gl

Ygl
Y
S0 S0

2.2.1 Gluing 2-manifolds

For all x ∈ C(S+ ) = C(S− ) gluing yields a map

glx : C(Y ; x, x, c) → C(Ygl ; c),

which induces a linear map

e ; x, x, c) → A(Y
glx : A(Y e gl ; c).

The latter map is compatible with the local relations, and so induces a linear map

glx : A(Y ; x, x, c) → A(Ygl ; c).

By general position, every C ∈ C(Ygl ; c) is isotopic to some gl(C ′ ) where C ′ ∈

C(Y ; x, x, c) for some x. Therefore we have a surjective linear map
M
gl : A(Y ; x, x, c) → A(Ygl , c).
x∈C(S± )

If we can describe the kernel of this map, we will have proved the desired gluing
theorem.
Let C ∈ C(Y ; a, b, c) for some a, b ∈ C(S± ), and let e ∈ C(S± × I; a, b). (See
Figure (2.2.2).) Let Ce = C ∪S+ e ∈ C(Y ; b, b, c) and eC = C ∪S− e ∈ C(Y ; a, a, c).
Then clearly
glb (Ce) ∼ gla (eC) ∈ C(Ygl ; c)

(via an isotopy which shifts along the gluing locus in Ygl ), and so, moving from
C(Ygl ; c) to A(Ygl ; c),
M
Ce − eC ∈ ker(gl) ⊂ A(Y ; x, x, c).
x∈C(S± )

(Here C also denotes its equivalence class in A(Y ; a, b, c).)

12 CHAPTER 2. Z2 HOMOLOGY AS A TQFT

a b
e
C C

Ce b a eC

glb gla

2.2.2 More than one way to glue a collar

We claim that elements of the form Ce − eC generate all of ker(gl). Consider

the following commutative diagram:
L e gl
e gl ; c)
/ A(Y
x A(Y ; x, x, c)
p q
L  
/ A(Ygl ; c)
x A(Y ; x, x, c)
gl

(The quotient maps are denoted by p and q.) The top gluing mapL is injective and
the other three maps are surjective. Let K denote the subspace of x A(Y e ; x, x, c)
generated by ker(p) and elements of the form Ce − eC. It suffices to show that
e gl ; c) contains all elements of the form A − A′ , where A and A′ differ
gl(K) ⊂ A(Y
by a generating local relation supported in some disk D ⊂ Ygl . Arbitrary isotopies
on Ygl are generated by (a) isotopies supported away from the gluing locus in Ygl
and (b) a shift isotopy supported in a collar of the gluing locus which moves the
gluing locus off of itself. If E and E ′ ∈ C(Ygl ; c) differ by an isotopy of type (a),
then E − E ′ = gl(F ) where F ∈ ker(p). If E and E ′ ∈ C(Ygl ; c) differ by an isotopy
of type (b), then E = gl(eC) and E ′ = gl(Ce) for some C and e as above. It follows
that if E and E ′ ∈ C(Ygl ; c) differ by a general isotopy, then E − E ′ ∈ gl(K). Hence
we can assume that the disk D above is disjoint from the gluing locus, since D can
be isotoped off of the gluing locus. Clearly in this case A − A′ = gl(F ) for some
F ∈ ker(p).
2.3. CYLINDER CATEGORIES 13

We have now proved

L
Theorem. With notation as above, let L ⊂ x A(Y ; x, x, c) be the subspace gen-
erated by all elements of the form Ce − eC, where C ∈ C(Y ; a, b, c) and e ∈
C(S± × I; a, b) for some a, b ∈ C(S± ). Then there is a natural isomorphism
M
A(Ygl ; c) ∼
= A(Y ; x, x, c)/L.
x

2.3 Cylinder Categories

The gluing theorem can be usefully restated in term of actions of cylinder categories.
Let S be a 1-manifold. Let A(S), the cylinder category associated to S, be the
category with objects C(S) and morphisms mor(a, b) = A(S × I; a, b). Composition
of morphisms is given in the obvious way by gluing cylinders.
Note that A(−S) is naturally isomorphic to A(S)op , and that A(S ⊔ S ′ ) is nat-
urally isomorphic to A(S) × A(S ′ ).
If ∂Y = S ⊔ R and c ∈ C(R), then Y and c determine a representation of A(S)
as follows. To each object a of A(S) (i.e. a ∈ C(S)), we associate the vector space
A(Y ; a, c). To each morphism f : a → b of A(S) (i.e. f ∈ C(S ×I; a, b)), we associate
the linear map
Gf : A(Y ; a, c) → A(Y ; b, c)
x 7→ x ∪ f.
In other words, Gf is given by gluing a copy of f to Y along S.
Now let Y be a surface with gluing data as above, ∂Y = S+ ⊔ S− ⊔ S0 . Let A =
A(S+ ) = A(S− )op . Fix c ∈ C(S0 ). For (a, b) ∈ C(S+ ⊔ S− ), define Vab = A(Y ; a, b, c).
As described above, the collection of vector spaces {Vab } affords a representation of
A × Aop = A(S+ ⊔ S− ). For all a there is a linear map
gla : Vaa → A(Ygl ; c),
and for every morphism e : x → y of A there is a commutative diagram
Vaa I
{= II gl
ida ×e { { II a
{{ II
{{ II
{ $
Vab C A(Ygl ; c)
CC :
CC uu
C uuu
e×idb CC! uu
uu glb
Vbb

(See Figure (2.2.2).)

14 CHAPTER 2. Z2 HOMOLOGY AS A TQFT

Theorem. A(Ygl ; c) is the universal object with the above two properties. In other
words, if there is a vector space W and linear maps fa : Vaa → W for all x, such
that for all e : x → y the following diagram commutes

= Vaa BB
ida ×e {{{ BB fa
{{{ BB
BB
{{ !
Vab C = W
CC |
CC |||
C
e×idb CC! ||
|| fb
Vbb
then there exists a map g : A(Ygl ; c) → W such that fa = g · gla for all x.
This is just a restatement of the first version of the gluing theorem. It has the
advantage of generalizing to the case where target category is not a vector space.
For example the theorem remains true if we systematically, in the above discus-
sion, replace A(Y, a) with the set of all properly embedded 1-submanifolds in Y ,
modulo isotopy, with boundary a. In this case A × Aop acts on the collection of sets
{Vab }, and A(Ygl ; c) is the universal set having the above properties.
The above universal construction is usually called the coend (dual to end) of the
A × Aop action. (See for example [Mac98, p. 226].) From our point of view, it would
be better to call this the gluing of a representation V of A × Aop , but “coend” is
already well-established so we stick with that terminology.

2.4 Semisimplicity
Back to the Z2 -homology vector spaces A(Y ; a). The categories A(S) (where S is a
closed 1-manifold) have the additional property of being semisimple. (See Appendix
B for details.) This means, among other things, that there is a fixed list L(S) of
irreducible representations of A(S), and that any representation ρ of A(S) can be
written as M
ρ∼= ρα ⊗ α,
α∈L(S)

where the isomorphism is natural, and ρα = hom(α, ρ). It is also the case that
L(A(S ⊔ S ′ )) = L(A(S)) × L(A(S ′ )) (i.e. the irreps of A × A′ , where A and A′ are
semi-simple, are isomorphic to the products of irreps of A and A′ ).
More specifically, let e1 , e2 , e3 , e4 be the morphisms of A(S 1 ) shown in Figure
(2.4.1). The ei ’s are a complete set of minimal idempotents for A(S 1 ).PThat is,
ei ei = ei , ei ej = 0 if i 6= j, any morphism f of A(S 1 ) can be written as 41 gi ei hi
for some {gi , hi }, and no larger set of idempotents has this property.
Let αi = A(S 1 )ei as a left A(S 1 ) representation. Geometrically, αi is S 1 × I with
ei fixed on one end, arbitrary curves-mod-relations on the rest of S 1 × I, and A(S 1 )
acting on the other end (see Figure (2.4.2)).
2.4. SEMISIMPLICITY 15

1 1 1 1
e1 = + e3 = +
2 2 2 2

1 1 1 1
e2 = − e4 = −
2 2 2 2

2.4.1 Idempotents for A(S 1 )

ei

2.4.2 The representation αi

If α is an irrep of A(S(∂Y )), define

A(Y ; α) = A(Y ; ·)α = hom(α, A(Y ; ·)).

More concretely, if ∂Y = S 1 then A(Y ; αi ) is naturally isomorphic to the space of

all curves (mod relations) on Y which coincide with ei in a collar of ∂Y . In general,
we put an idempotent at each closed component of ∂Y .
Reflection in the I direction gives functor A(S) → A(S)op for any closed 1-
manifold S, and the square of this functor is the identity. This allows us to identify
representations of A(S) with representations of A(S)op . For the Z2 homology theory,
α∼ = αop for all irreps α, but this doesn’t hold for general theories. (Also, there are
no Frobenius-Schur complications (see (13.0.1)) for the Z2 homology theory.) If
α ⊗ β is an irrep of A × Aop , then the coend of α ⊗ β is naturally isomorphic to
C if α = β op , and zero if α 6∼
= β op . From this follows a third version of the gluing
theorem:

Theorem. Let Y be a surface with boundary ∂Y = S+ ⊔ S− ⊔ S0 , and let Ygl denote

Y with S+ and S− identified, as above. Let β be an irrep of A(S0 ). Then
M
A(Ygl ; β) ∼
= A(Y ; α, α, β).
α∈L(A(S+ ))

[talk about 3-dim’l part of theory?]

[need to make boundary orientation and left/right action conventions consistent]

16 CHAPTER 2. Z2 HOMOLOGY AS A TQFT
Chapter 3

Topological Fields

The techniques of the previous chapter generalize easily to manifolds of arbitrary

dimension and a wider variety of functors C and local relations. In this chapter
we formalize the properties of C which are needed to make the constructions work.
(General requirements for local relations are given in Chapter 4.) Functors C satis-
fying the requirements will be called topological fields. (The term “field” is borrowed
from physics, where typical examples are the set of maps from a manifold into a
linear space, or the set of connections on a bundle over a manifold.)
In a nutshell, a topological field is a collection of functors from manifolds of
dimensions 0, . . . , n to sets which behave well under cutting and pasting. Our main
examples are the set of maps from a manifold into some target space B, and the
set of embeddings of cell complexes (perhaps with additional structure and specified
local combinatorics) into a manifold. One of the goals of this chapter is to lay the
groundwork for more exotic examples.
(Impatient readers would do well to skip this chapter and just assume that
“topological field” means one of the main examples.)
We start by listing the properties of manifolds that we will use. We then define
topological fields as a collection of functors which preserve these properties in the
appropriate way.
[I haven’t yet carefully checked that the field axioms given below are complete.
I might need to add a couple of more axioms.]

3.1 Manifolds
Our manifolds will always be oriented and compact. They might have additional
structure (e.g. [need examples of additional structure]). They could be PL or
smooth. [for smooth, need to say more about corners; also could be topological?]
Morphisms between manifolds of the same dimension will always be orientation-
preserving homeomorphisms (or diffeomorphisms in the case of smooth manifolds)
unless specified otherwise.

17
18 CHAPTER 3. TOPOLOGICAL FIELDS

Y -Y

gl

Y'
Y
gl

-Y

3.1.1 Gluing manifolds

Let Mi denote the category whose objects are compact, oriented manifolds of
dimension i, and whose morphisms are orientation-preserving homeomorphisms. In-
cluded in each Mi is ∅, the empty manifold of dimension i.
The collection of categories {Mi }i≥0 has the following familiar additional struc-
ture.

Boundary. There is a functor ∂ : Mi → Mi−1 . On objects ∂ is the usual boundary

of manifolds (with the “outward normal last” orientation convention, say [is this the
convention I want to use?]). For f : Y → W a morphism, ∂f is defined to be f
restricted to ∂Y , a morphism from ∂Y to ∂W . If ∂Y = ∅, we say that Y is closed.

Orientation reversal. There is an orientation reversal functor − : Mi → Mi . For

objects, −A is A with the opposite orientation. For morphisms, −f is identical to f
as a map of sets. −2 is the identity functor. − commutes with ∂; i.e., ∂(−A) = −∂A.

Disjoint union. There is a functor ⊔ : Mi × Mi → Mi given by disjoint union.

(Mi ×Mi denotes the abstract product of categories.) Up to canonical isomorphism,
this operation is commutative and associative. The empty i-manifold acts as the
identity element. ⊔ gives Mi the structure of a commutative monoid.

Gluing without corners. Given an i-manifold M and an identification ∂M =

Y ⊔−Y ⊔Y ′ , there is a glued manifold glY (M ) obtained by identifying the two copies
of Y in ∂M . (See Figure (3.1.1).) There is a natural isomorphism ∂(glY (M )) = Y ′ ,
and − glY (M ) can be identified with glY (−M ). If the gluing region ±Y ⊂ ∂M is
clear from context, we write simply gl(M ). If M = M1 ⊔ M2 , Y ⊂ ∂M1 , and
−Y ⊂ ∂M2 , then we sometimes use the notation M1 ∪Y M2 or simply M1 ∪ M2 .
3.2. TOPOLOGICAL FIELDS 19

Y -Y Y'
Y'
gl

Y -Y
Y'
gl

3.1.2 Gluing manifolds with corners

Gluing with corners. Let M be an i-manifold and ∂M = Y ∪ −Y ∪ Y ′ . (That

is, ∂M has an identification with the gluing-without-corners Y ∪ −Y ∪ Y ′ , and
∂Y ′ = ∂(Y ⊔ −Y ).) Then there is a glued manifold glY (M ) obtained by identifying
the two copies of Y in ∂M . (See Figure (3.1.2).) There is a natural isomorphism
∂(glY (M )) = gl∂Y (Y ′ ), and − glY (M ) can be identified with glY (−M ). Note that
gluing without corners is a special case of gluing with corners (corresponding to
∂Y = ∅). If the gluing region ±Y ⊂ ∂M is clear from context, we write simply
gl(M ). If M = M1 ⊔ M2 , Y ⊂ ∂M1 , and −Y ⊂ ∂M2 , then we sometimes use the
notation M1 ∪Y M2 or simply M1 ∪ M2 .
Product with I. There is a functor Mi → Mi+1 given by M 7→ M × I and
f 7→ f × id. (Here I denotes the unit interval [0, 1] ⊂ R.) Unless stated otherwise,
our products will be “pinched” along the boundary, so that ∂(M ×I) = (−M )∪∂M M
(rather than (−M ) ∪ (∂M × I) ∪ M ).
Collar neighborhoods. Let ∂M = Y ∪ Y ′ . Then there is a collaring morphism
(homeomorphism) M → M ∪Y (Y × I), well-defined up to isotopy rel boundary.
[?? add composition of prod bordisms ??]

3.2 Topological Fields

Before defining topological fields, we give some examples. The reader is encouraged
to keep these examples in mind when reading the subsequent abstract axioms for
topological fields.
Mapping spaces. Given a topological space B, define

MFB (X) = {f : X → B},

 20 CHAPTER 3. TOPOLOGICAL FIELDS

the set of all continuous maps from a manifold X to B. We will be particularly

interested in the case where B is the classifying space BΓ of a finite group Γ. When
the space B is either irrelevant or clear from context, we will omit the subscript B
from the notation.
MF is a functor (either covariant or contravariant, since all morphisms of Mi are
invertible), and it preserves monoidal structure (that is, MF(X1 ⊔X2 ) = MF(X1 )×
MF(X2 ), and similarly for morphisms). If f ∈ MF(X), we define ∂f to be f
restricted to ∂X. Maps can be glued together if they agree on the gluing region,
and all maps on the glued manifold are obtained in this way.
3.2.1 Designs. Define
S(X) = {codimension-1 unoriented PL submanifolds of X}.
The submanifolds should be properly embedded and transverse to ∂X.
Submanifolds can be glued together if they agree on the gluing region, but note
that not all elements of S(gl(X)) are obtained from S(X) in this way. A submanifold
of gl(X) is so obtained if and only if it is transverse to the image in gl(X) of the
gluing region. Of course this is, in various senses that can be made precise, almost
all of S(gl(X)), but this “almost all” instead of “all” will complicate our axioms for
topological fields.
[give more precise def: embedded complex, specified local behavior, labels, orienta-
tions]
The above functors can be generalized to fancier “designs” on manifolds. [maybe
find a better term than “designs”.] (See for example [forward refs; e.g. G2].) Our
generic name for such generalizations will be DF. For example, we could define
DF(Y ), for a 2-manifold Y , to be the set of all trivalent graphs G in Y , with
oriented edges and colors assigned to each edge and each component of Y r G.
[need fig] As before, these should be properly embedded graphs, transverse to ∂Y .
Define DF(R), for a 1-manifold R, to be the set of all properly embedded, oriented
0-submanifolds of R (i.e. a finite collection of signed points in the interior of R),
with colors assigned to both the points and the regions between the points. Define
DF(Z), for a 0-manifold Z, to be a coloring of the components of Z. Boundaries of
fields are defined in the obvious ways, restricting orientations and colors.
Let Y = R ∪ −R ∪ R′ . A field (i.e. colored, oriented graph) on Y glues up to one
on glR (Y ) if and only if its restrictions to the two copies of R agree with orientations
reversed. This suggests that we should define, for any 1-manifold S, a bijection
DF(S) ↔ DF(−S), a↔b
a,
where ba has the same colors and points as a, but with reversed orientations on the
points.
We now axiomatize topological fields (i.e. list the properties of the above exam-
ples which are needed to make later constructions work).
3.2. TOPOLOGICAL FIELDS 21

Let S denote the category of sets. A topological field (of top dimension n)
consists of a collection of functors

Ci : Mi → S

(0 ≤ i ≤ n) which satisfy the following conditions. (We will usually omit the
subscript i from the notation.)
Boundary. There are boundary maps

∂ : C(X) → C(∂X).

Also, for c ∈ C(∂X) we define

C(X; c) = ∂ −1 (c) ⊂ C(X);

C(X; c) should be thought of as the set of fields on X with boundary conditions c

on ∂X.
These boundary maps should be compatible with morphisms and the boundary
functor on manifolds: for all i-manifolds X and Y and f : X → Y , the following
diagram commutes.
Ci (f )
Ci (X) / Ci (Y )

∂ ∂
 Ci−1 (∂f ) 
Ci−1 (∂X) / Ci−1 (∂Y )

In other words, ∂ is a natural transformation between the two functors Ci and Ci−1 ◦∂
from Mi to S.
Examples. For α ∈ MF(X), ∂α = α|∂X . For β ∈ S(X) (or β ∈ DF(X)), ∂β is
the (transverse) intersection of β with ∂X (preserving colors and orientations for
DF).
Orientation reversal. If X is closed, there is a bijection

C(X) ↔ C(−X), a↔b

a.

For general X, there are bijections

C(X; c) ↔ C(−X, b
c), a↔b
a.

Examples. For MF, these are the obvious bijections resulting from ignoring the
orientation of X. For DF, the bijections could involve reversing orientations on
parts of the design; see discussion at (3.2.1).
Disjoint union. The fields preserve monoidal structure: there is an identification

C(X1 ⊔ X2 ) = C(X1 ) × C(X2 ),

 22 CHAPTER 3. TOPOLOGICAL FIELDS

and these identifications are compatible with morphisms, orientation reversal, and
boundaries in the obvious ways. Note that this implies that C(∅) consists of a single
element.
3.2.2 Gluing without corners. Given an identification

∂X = Y ⊔ −Y ⊔ W

(see Figure (3.1.1)), the disjoint union axiom provides an identification

C(∂X) = C(Y ) × C(−Y ) × C(W ).

Using the boundary and orientation reversal maps, we get two maps from C(X) to
C(Y ).
∂ / C(∂X) pr / C(Y )

EqY (C(X))  / C(X)
JJ O
JJ pr
JJ −
JJ
J$
C(−Y )
Let EqY (C(X)) denote the equalizer of these two maps. (That is, the set of all fields
in C(X) on which the two maps agree. In the case where Y and −Y lie in separate
components of X, EqY (C(X)) is a fibered product.) Then there is an injection
gl : EqY (C(X)) → C(glY (X)) such that the following diagram commutes.
 / C(X) ∂ / C(∂X)
EqY (C(X)) 
gl prW
 
∂
C(glY (X)) / C(W )

We will often identify EqY (C(X)) is its image gl(EqY (C(X))) ⊂ C(glY (X)). If
X = X1 ⊔ X2 , Y ⊂ ∂X1 and −Y ⊂ ∂X2 , then we will often use the notation α1 ∪ α2
instead of gl((α1 , α2 )). (Here αi ⊂ C(Xi ).)
Furthermore, we require that for any x ∈ C(glY (X)) and any neighborhood N
of the image of Y in glY (X) there exists a homeomorphism f of glY (X), isotopic
to the identity and supported in N , such that f∗ (x) ∈ EqY (C(X)). In other words,
any field on the glued manifold is close to a field obtained by gluing.
[Need to define restriction c|S for codim-0 submanifold S.]

Examples. For MF, EqY (MF(X)) is all of MF(glY (X)). For S, EqY (S(X))
consists of all submanifolds of glY (X) which are transverse to the image of the
gluing region Y in glY (X).
Gluing with corners. Assume an identification

∂X = Y ∪ −Y ∪ W,
3.3. EXTENDED ISOTOPY 23

with Y and −Y disjoint in ∂X. (See Figure (3.1.2).) [correct Y ′ vs W inconsistency?]

Let S = ∂W = ∂(Y ⊔ −Y ), and let EqS (C(Y ⊔ −Y ⊔ W )) ⊂ C(∂X) be as described
in (3.2.2) (gluing without corners on ∂X). Define C ⋔(X) = ∂ −1 (EqS (C(Y ⊔ −Y ⊔
W ))). C ⋔(X) has two maps (one of which involves orientation reversal) to C(Y ); let
EqY (C ⋔(X)) denote the equalizer of these two maps.

∂
C(X) / C(∂X)
O O
gl
?
EqY (C ⋔(X))
 / C ⋔(X) ∂ / Eq (C(Y ⊔ −Y ⊔ W )) pr / C(Y )
S RRR O
RRR pr
RRR
RRR −
R(
C(−Y )

Let Eq∂Y (C(W )) ⊂ C(W ) be as in (3.2.2), for the gluing W → gl∂Y (W ) = ∂(glY (X)).
There is a map ∂ : EqY (C ⋔(X)) → Eq∂Y (C(W )) induced by ∂ : C ⋔(X) → C(∂X).
The gluing with corners axiom requires that there is an injection gl : EqY (C ⋔(X)) →
C(glY (X)) such that the following diagram commutes.

∂
EqY (C ⋔(X)) / Eq (C(W ))
∂Y

gl gl
 
∂
C(glY (X)) / C(gl (W ))
∂Y

Furthermore, we require that for any x ∈ C ⋔(X) and any neighborhood N of the
image of Y in glY (X) there exists a homeomorphism f of glY (X), isotopic (rel
boundary) to the identity and supported in N , such that f∗ (x) ∈ EqY (C ⋔(X)). In
other words, any field on the glued manifold whose boundary is obtained by gluing
is close to a field obtained by gluing.
As before, we will sometimes use the notation α1 ∪α2 instead of gl((α1 , α2 )) when
we are gluing distinct components of X together. The subspace EqY (C ⋔(X)) ⊂
C(glY (X)) is called the gluing subspace relative to this decomposition of glY (X).
Note that the gluing without corners axiom (3.2.2) is a special case of this one.

[add products, collars, homeos of products(?)]

3.3 Extended Isotopy

Plan:

• def: extended isotopy generated by isotopy and gluing of product fields to

collars; gluing e-isotopic fields yields e-isotopic fields
24 CHAPTER 3. TOPOLOGICAL FIELDS

• note that this is a local relation, since both isotopy and collars can be localized
(need further field assumptions here? (for collars))

• work through several examples (ones that will be used later)

• purpose of this section: do some tedious proofs here so as not to interrupt the
flow (w?) in later chapters

to do:

• ? generalize to not-necessarily-linear case?

• need to be clearer about A maybe/maybe-not being linear (examples will help

with this)

• need to define “codim-k” near the start

• ?? need to define “transverse” in field section

Chapter 4

Basic Constructions and

Gluing I

In this chapter we define local relations/subspaces, introduce the basic construc-

tion for codimension 1 (dimension n) manifolds, and prove the codimension 1 glu-
ing theorem. [need to be consistent about hyphenating (or not) codimension 1 or
codimension-1]
Local relations (or dually subspaces) are defined as ones which can be specified
inside an n-ball. Given fields and local relations, the basic construction associates
to each n-manifold Y with field boundary conditions c an object A(Y ; c) (either a
vector space or set in most examples). Given an n−1-manifold S, gluing defines an
associative composition on the various A(S × I; c0 , c1 ), and we get a category A(S).
The category A(∂Y ) acts on the collection {A(Y ; ·)}, and there is a gluing theorem
describing the glued spaces A(Ygl ) in terms of the various A(Y ; c) and the action of
A(∂Y ). We give several versions of this gluing theorem.

4.1 Local Relations and Local Subspaces

Local relations (or dually, local subspaces) carry the same information as the path
integral of B n+1 (see Chapters 1 and 6).
Let C be a field functor and B = B n be the standard n-dimensional ball. A
system of local relations for C is a collection of equivalence relations ∼ in C(B; c)
(for all c ∈ C(∂B)) such that:

• If a, b ∈ C(B; c) are related by an isotopy, then a ∼ b. (In other words, ∼ is 4.1.1

at least as strong as isotopy. This is equivalent to the topological invariance
of the action.)

• If a ∈ C(B; c) and S is a codimension 0 submanifold of ∂B, then ((c|S × 4.1.2

I) ∪ a) ∼ a. In other words, we can glue a product field onto a subset of

25
 26 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

the boundary of a and the result is equivalent to a. (This is in some sense

the closure of the isotopy condition.) [maybe define “extended isotopy” in the
fields chapter]

4.1.3 • Let E ⊂ B be a subball (possibly intersecting ∂B) of B, equipped with

a homeomorphism to the standard ball B. Then ∼ on the various C(E; c′ )
induces relations on the various C(B; c), and we require that this induced
relation be implied by ∼ on C(B; c). (In other words, applying ∼ to a subball
yields no new relations.)

Let A(B; c) = C(B; c)/ ∼.

More generally, we could replace C(B; c) above with A(B; e c), the space of finite
linear combinations of elements of C(B; c). In this case ∼ is equivalent to a quotient
e
map A(B; c) → A(B; c).
e
Dually, let Z(B; c) be the space of all functions from C(B; c) to C. A system of
e
local subspaces for C is a collection of subspaces Z(B; c) ⊂ Z(B; c) (for all c ∈ C(∂B))
such that:

• If a, b ∈ C(B; c) are related by an isotopy, f (a) = f (b) for all f ∈ Z(B; c).

• If a ∈ C(B; c) and S is a codimension 0 submanifold of ∂B, then f ((c|S × I) ∪

a) = f (a) for all f ∈ Z(B; c).

• Let E ⊂ B be a subball (possibly intersecting ∂B) of B, equipped with a

homeomorphism to the standard ball B. Then the various subspaces Z(E; c′ )
induce subspaces Z ′ (B; c), and we require that Z ′ (B; c) ⊂ Z(B; c).

Z(B; c) should be thought of as the image of the local projections in Chapter 1.

4.2 The Basic Construction in Codimension 1

The local relations and subspaces of the previous section give rise to vector spaces
for arbitrary n-manifolds. The construction is fairly obvious, but it takes a while
to state it precisely. Briefly, A(Y ; c) is the quotient of A(Ye ; c) by local relations
derived from all balls B ⊂ Y . Dually, Z(Y ; c) in the intersection of the subspaces
e ; c) derived from all B ⊂ Y .
of Z(Y
Let Y n be a closed n-manifold. For each ball B ⊂ Y we have a decomposition
Y = B ∪ Y ′ and [
C(Y ) ⊇ C(B; c) × C(Y ′ ; b
c).
c∈C(∂B)

Hence M
e )⊇
A(Y e
A(B; e ′; b
c) ⊗ A(Y c).
c∈C(∂B)
4.2. THE BASIC CONSTRUCTION IN CODIMENSION 1 27

P
c2

c1
c0 B Y

4.2.1 Subball intersecting boundary

e ) induced by the local relations

Let ∼B be the equivalence relation on C(Y ) or A(Y
on B and the above decompositions. Let ∼Y be the equivalence relation generated
by {∼B }, where B ranges over all subballs of Y . Define A(Y ) = A(Ye )/∼Y (or
A(Y ) = C(Y )/∼Y ).
Dually, we have, for each B ⊂ Y ,
Y
e )⊆
Z(Y e
Z(B; e ′; b
c) ⊗ Z(Y c).
c∈C(∂B)

Let Y
′
ZB ⊆ e ′; b
Z(B; c) ⊗ Z(Y e ).
c) ⊂ Z(Y
c∈C(∂B)

Define \
′
Z(Y ) = ZB ,
B

where B ranges over all subballs of Y . [need to revise notation above; ⊗ is not the
right thing; fix below too]
More generally, let Y be an n-manifold with boundary, c ∈ C(∂Y ), and B ⊂ Y be
a subball, possibly intersecting ∂Y . Let S0 = ∂B ∩ ∂Y , ∂B = S0 ∪ S1 , ∂Y = S0 ∪ S2 ,
and P = ∂S0 = ∂S1 = ∂S2 . (See Figure (4.2.1).) Let c = c0 ∪ c2 with respect to
the decomposition ∂Y = S0 ∪ S2 , and let ∂c0 = ∂c2 = d ∈ C(P ). Let Y = Y ′ ∪ B,
so that ∂Y ′ = (−S1 ) ∪ S2 . Then we have decompositions
[
C(Y ; c) ⊇ C(B; c0 ∪ c1 ) × C(Y ′ ; cb1 ∪ c2 )
c1 ∈C(S1 ;d)

and M
e ; c) ⊇
A(Y e
A(B; e ′ ; cb1 ∪ c2 ).
c0 ∪ c1 ) ⊗ A(Y
c1 ∈C(S1 ;d)

e ; c), and the various ∼B gen-

Then as before we can define ∼B on C(Y ; c) or A(Y
e ; c)/∼Y or A(Y ; c) =
erate an equivalence relation ∼Y . We define A(Y ; c) = A(Y
C(Y ; c)/∼Y .
28 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

Dually, we have, for each B ⊂ Y ,

Y
e ; c) ⊆
Z(Y e
Z(B; e ′ ; cb1 ∪ c2 ).
c0 ∪ c1 ) ⊗ Z(Y
c1 ∈C(S1 ;d)

Let Y
′
ZB ⊆ e ′ ; cb1 ∪ c2 ) ⊂ Z(Y
Z(B; c0 ∪ c1 ) ⊗ Z(Y e ; c).
c1 ∈C(S1 ;d)

As before, define \
′
Z(Y ; c) = ZB ,
B

where B ranges over all subballs of Y .

Clearly both A and Z are functorial; boundary-fixing homeomorphisms of Y act
on A(Y ; c) and Z(Y ; c). (Since all morphisms are invertible, we needn’t distinguish
between covariant and contravariant functors.)
At this point we have given two definitions of A(B n ; c) and Z(B n ; c): input for
local relations or subspaces and also the above constructions for subballs of B n . We
must show that these two definitions agree. More generally we will show that for any
n-manifold Y the above constructions of A(Y ; c) (or Z(Y ; c)) yield the same answer
if we restrict the subballs in the construction to come from some open covering of
Y.
Let Ar (Y ; c) denote the result of the restricted construction. Clearly A(Y ; c) ⊆
r
A (Y ; c) (more subballs means a stronger equivalence relation), so we must show that
A(Y ; c) ⊇ Ar (Y ; c). An arbitrary boundary-fixing isotopy of Y can be decomposed
into isotopies supported in balls of the given open cover of Y . Thus the restricted
equivalence relation is at least as strong as isotopy. Let B ⊂ Y be an arbitrary
subball in the interior of Y . There is an isotopy of Y which carries B into the
interior of one of the covering subballs. But (4.1.3) in the definition of local relation
[need better way of referring here (and below)] now insures that that any relation
supported in B is also part of the restricted relation. If B intersects ∂Y , we convert
B to an interior subball by gluing a collar onto ∂Y . (This relies on (4.1.2) in the
definition.) The collar is added in pieces, each of which is supported in the boundary
of the given covering of Y .

4.3 Cylinder Categories

Our next goal is to describe how A and Z behave under cutting and gluing. This will
entail understanding the effect of shifting a collar of the cutting n−1-submanifold
across the cutting submanifold (see Section 4.4), which motivates the the definitions
in this section.
Let S be a closed n−1-manifold. Define A(S), the cylinder category associated
to S, to be the category consisting of:
4.3. CYLINDER CATEGORIES 29

• Objects: A(S)0 = C(S).

• Morphisms: A(S)1ab = A(S × I; b a, b). In other words, the morphisms from a to

b are fields on the cylinder S × I with boundary conditions given by ba and b,
modulo local relations. (In most of our examples this will be a vector space.)

• Composition: Given by gluing two cylinders together to get another cylinder.

Associativity of composition follows from the associativity of gluing.

More generally, if S has boundary and c ∈ C(∂S) we can define a category A(S; c)
with objects C(S; c) and morphisms A(S × I; b a, b), where a, b ∈ C(S; c). Here, as
usual, S × I denotes a cylinder with the vertical boundary ∂S × I pinched, so that
∂(S×I) is naturally identified with (−S)∪S (the double of S) and b a ∪b ∈ C(∂(S×I)).
If S is the empty n−1-manifold, then A(S) is the trivial category consisting of
a single object and (multiples of) its identity morphism.
There is a natural identification

A(S; c)op = A(−S; b

c).

On objects this isomorphism is given by a 7→ b a, and on morphisms it is induced by

the orientation preserving map r : (x, t) 7→ (x, 1 − t) from S × I to (−S) × I.
Composing the above functor with orientation reversal on (−S) × I gives a 4.3.1
contravariant functor from A(S; c) to itself. On objects this is the identity, a 7→ a.
def d
On morphisms it is y 7→ y ∗ = r(y), where y ∈ A(S × I; b a, b) and r flips the I factor
of S × I as above.
It is easy to see that there is a natural identification

A(S1 ⊔ S2 ) = A(S1 ) × A(S2 ).

Recall that a right representation of a category A is a functor from Aop to V1 ,

the category of vector spaces and linear maps. In other words, for each object x of A
we have a vector space Wx , for each morphism e : x → y of A we have a linear map
ρe : Wx → Wy , and composition of morphisms is preserved (ρef = ρe ρf = ρf ◦ ρe ).
If A has linear morphisms, then ρ above is required to be a linear map.
Let Y be a n-manifold with boundary (possibly empty). Then the collection of
vector spaces A(Y ; ·) affords a representation of A(∂Y ). A morphism of A(∂Y ) is
an equivalence class of fields on ∂Y × I, and gluing this as a collar to Y defines
the action of the representation. Associativity of gluing implies that composition is
preserved. We denote this representation of A(∂Y ) by A(Y ). In other words, A(Y )
consists of the collection of vector spaces A(Y ; ·) together with the above action of
A(∂Y ).
Dually, Z(Y ; ·) affords a left representation of A(∂Y ). The action of e ∈ A(∂Y )1ab
on x ∈ Z(Y ; b) is defined by (e · x)(u) = x(u ∪ e), where u ∈ C(Y ; a). As before, we
 30 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

denote this representation (collection of vector spaces plus left action of A(∂Y )) by
Z(Y ).
For any closed n−1-manifold S define Z(S) = Rep(A(S)), the category whose
objects are left representations of A(S) and whose morphisms are natural transfor-
mations (also called intertwiners in this context). Then we have, for all n-manifolds
Y,
4.3.2 Z(Y ) ∈ Z(∂Y ).
This is the categorified version of (1.2.1) and (6.1.1).
4.3.3 Assume now that A(∂Y ) is semisimple (see Appendix B) and let α be a right
irrep of A(∂Y ). Define
A(Y ; α) = mor(α, A(Y )),
the space of intertwiners from α to A(Y ). Let L be a complete set of irreps for
A(∂Y ). Then there is a canonical isomorphism
M
A(Y ) ∼
= A(Y ; α) ⊗ α.
α∈L

In this context, irreps are sometimes referred to as “labels” or “particles”.

If eα ∈ A(∂Y )1bb is a minimal idempotent for α (see Appendix B [need more
specific reference]), then there is a natural isomorphism

A(Y ; α) ∼
= A(Y ; b)eα .

In other words, A(Y ; α) is naturally isomorphic to the subspace of fields (mod

relations) in A(Y ; b) which have representatives containing eα in a collar of ∂Y .
Dually, if β is a left irrep of A(∂Y ), we define

Z(Y ; β) = mor(β, Z(Y )),

and we have, canonically,

M
Z(Y ) ∼
= Z(Y ; β) ⊗ β
β∈L

(where now L denotes a complete set of left irreps of A(∂Y )). Note: In the old-
fashioned axiomatization of TQFTs (see, e.g., [Walk91]) these are the spaces asso-
ciated to surfaces with labeled boundary.

4.4 Gluing Codimension-1 Manifolds

Let Y be an n-manifold on the verge of being glued without corners. In other words,
we have an identification ∂Y = (−S) ⊔ S ⊔ R (see Figure (4.4.1)), and define Ygl to
be Y with S glued to −S. Fix (for the remainder of this section) c ∈ C(R) = C(∂Ygl ).
4.4. GLUING CODIMENSION-1 MANIFOLDS 31

−S +S

R R
Y Ygl

4.4.1 More gluing

Our goal is to describe A(Ygl ; c) in terms of the various A(Y ; ·, ·, c) and the action
of A(−S ⊔ S) = A(S)op × A(S) on these spaces.
We’ll start with the most abstract formulation of the codimension-1 gluing the-
orem, and then work our way toward more concrete statements.
Theorem. Let Y , S, Ygl , c be as above. 4.4.2
(a) For each object x of A(S) there is a map

b, x, c) → A(Ygl ; c).
glx : A(Y ; x

(b) For each morphism e : x → y of A(S) the following diagram commutes

b, x, c)
A(Y ; x
e×1nnnn
nn7 NNN
NNNglx
n NNN
nn NN'
nnn
A(Y ; yb, x, c) A(Ygl ; c)
PPP 7
PPP
ppppp
PPP
1×e PPP' ppp
ppp gly
A(Y ; yb, y, c)

(c) A(Ygl ; c) is the universal object (vector space, set, or whatever flavor of A
we’re using) with properties (a) and (b). In other words, given a W and maps
gl′x : A(Y ; x
b, x, c) → W (for all x) such that the diagram analogous to the one in (b)
above commutes for all e, there is a unique θ : A(Ygl ; c) → W such that gl′x = θ ◦ glx
for all x.
b, x, c)
A(Y ; x ′
e×1nnnn
nn7 NNN
NNN
glx
NN
nnn glx NNN'
nnn '
A(Y ; yb, x, c) A(Ygl ; c) _θ _ _7/ W
PPP p 7
PPP glyppp
PPP pp
1×e PPP' pp
ppp gl′y
A(Y ; yb, y, c)
In other words, A(Ygl ; c) is the coend (see Appendix A [need more specific reference])
of the action of A(S)op × A(S) on A(Y ; ·, ·, c).
 32 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

Proof. Part (a) is obvious; fields which agree on ±S can be glued to yield a field on
Ygl , and all local relations on Y are also local relations on Ygl .
Part (b) follows from the fact that shifting a collar across the gluing submanifold
changes the field by an isotopy. The top arrows in the diagram correspond to gluing
e to −S, and then gluing −S to S to obtain a field on Ygl . The bottom arrows
correspond to gluing e to S first. The two fields thus obtained differ by an isotopy
of Ygl which shifts a collar across ±S, so they represent the same element of A(Ygl ; c)
and the diagram commutes.
Part (c) requires a little more work. Let g ∈ C(Ygl ; c) and ḡ be the equivalence
class of g in A(Ygl ; c). After an isotopy, we may assume that g is transverse to
the gluing submanifold (image of ±S) in Ygl . Then ḡ = glx (g♯ ) for some g♯ ∈
A(Y ; x b, x, c), where x is g restricted to ±S. Necessarily, we define θ(ḡ) = gl′x (g♯ ).
Thus θ is unique if it exists.
It remains to be shown that the above procedure for defining θ yields well-defined
results. We must show that if g and h are equal in A(Ygl ; c), then gl′x (g♯ ) = gl′y (h♯ )
(where y is h restricted to ±S). If g and h are locally related with respect to a ball
B ⊂ Ygl disjoint from ±S, then g♯ = h♯ , so a fortiori gl′x (g♯ ) = gl′x (h♯ ). If g and h
are related by a collar shift isotopy along ±S, then gl′x (g♯ ) = gl′y (h♯ ) by the assumed
commutativity of the above diagram for W and gl′ . Since collar shift isotopies plus
isotopies supported in balls disjoint from ±S generate arbitrary isotopies on Ygl ,
gl′x (g♯ ) = gl′y (h♯ ) if g and h are related by an arbitrary isotopy. Finally, suppose g
and h are locally related with respect to a ball B not disjoint from ±S. Let ϕ be
an isotopy of Ygl which moves B off of ±S. Note that ϕ(g)♯ is related to ϕ(h)♯ by a
local relation supported in ϕ(B). Then gl′x (g♯ ) = gl′w (ϕ(g)♯ ) = gl′z (ϕ(h)♯ ) = gl′y (h♯ ),
where w and z are the restrictions of ϕ(g) and ϕ(h) to ±S.

There is, of course, and dual version of the gluing theorem for Z:

4.4.3 Theorem. Let Y , S, Ygl , c be as above.

(a) For each object x of A(S) there is a map

rx : Z(Ygl ; c) → Z(Y ; x
b, x, c).

(b) For each morphism e : x → y of A(S) the following diagram commutes

b, x, c)
Z(Y ; x
nn gNNN
e×1nnnn NNNrx
n NNN
nn NN
nw nn
Z(Y ; yb, x, c) Z(Ygl ; c)
gPPP
PPP ppp
PPP pppp
1×eop PPP p r
pw pp y
Z(Y ; yb, y, c)
4.4. GLUING CODIMENSION-1 MANIFOLDS 33

(c) Z(Ygl ; c) is the universal object with properties (a) and (b). In other words,
given a W and maps rx′ : W → Z(Y ; x b, x, c) (for all x) such that the diagram
analogous to the one in (b) above commutes for all e, there is a unique θ : W →
Z(Ygl ; c) such that rx′ = rx ◦ θ for all x.

b, x, c) l
Z(Y ; x
nn gNNN rx′
e×1nnnn NNN
N
nnn rx NNNN
wnnn
Z(Y ; yb, x, c) Z(Ygl ; c) o_θ _ _ W
gPPP p
PPP ry ppp
PPP ppp
1×e PPP p
pw pp r ry′
Z(Y ; yb, y, c)

In other words, Z(Ygl ; c) is the end (See Appendix A [need more specific reference])
of the action of A(S)op × A(S) on Z(Y ; ·, ·, c).
The proof is dual to the proof of (4.4.2). The map rx : Z(Ygl ; c) → Z(Y ; x b, x, c) is
b, x, c)) of C(Ygl ; c).
given by restricting a function in Z(Ygl ; c) to the subset glx (C(Y ; x
In the remainder of this section we state a number of special cases of the
codimension-1 gluing theorem.

Corollary. Let Y , S, Ygl , c be as above. If the target category of A (on n-manifolds) 4.4.4
is the category of vector spaces, then
 
M
A(Ygl ; c) =  A(Y ; xb, x, c) /hev ∼ vei.
x∈C(S)
L
By hev ∼ vei we mean the subspace of x∈C(S) A(Y ; x b, x, c) generated by all ev − ve,
for all morphisms e : x → y of A(S) and all v ∈ A(Y ; y, x, c). Here we write the
action of A(S) as juxtaposition on the right and the action of A(S)op as juxtaposition
on the left.
L
Note that above means finite linear combinations.
Corollary. Let Y , S, Ygl , c be as above. If the target category of Z (on n-manifolds)
is vector spaces, then
Y
Z(Ygl ; c) = {(vx ) ∈ b, x, c) | vx e = evy for all e}.
Z(Y ; x
x∈C(S) 4.4.5

Here e : x → y and vx e, evy ∈ Z(Y ; x

b, y, c).
Q
Note that above means infinite linear combinations.
If ∂Y1 = (−R) ⊔ S, then we can think of Y1 as a bordism from R to S and A(Y1 )
as an A(R)-A(S) bimodule (i.e. it has an action of A(R)op × A(S)).
 34 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

Corollary. Let ∂Y1 = (−R) ⊔ S and ∂Y2 = (−S) ⊔ T . Then

A(Y1 ∪S Y2 ) = A(Y1 ) ⊗S A(Y2 ).

(See Appendix A for the definition of ⊗S .) In other words, we have a functor from
the category of n-dimensional bordisms to the category of bimodules.

(Some authors use the above functorial property as an axiom for the codimension-
1 gluing properties of TQFTs, but note that without auxiliary axioms it fails to
handle the case of gluing an n-manifold to itself, nor does it cover interrelationships
between different bordisms with the same underlying manifold (i.e. moving a part
of the incoming boundary to the outgoing boundary, or vice-versa).)
Let Y , S, R, Ygl , c be as before. Assume now that A(S) and A(R) are semisimple
categories. Let L be a complete set of irreps for A(S) and K be a complete set of
irreps for A(R). Then
M
A(Y ) = A(Y ; α∗ , β, γ) ⊗ α∗ ⊗ β ⊗ γ,
α,β,γ

where the sum is over α, β ∈ L and γ ∈ K. Note that the dual representation α∗ is
a right irrep of A(S)op = A(−S), and so α∗ ⊗ β ⊗ γ runs through a complete set of
irreps for A(∂Y ). (See (4.3.3) above.) Also
M
A(Ygl ) = A(Ygl ; γ) ⊗ γ.
γ∈K

By (B.5.1), if α is an irrep then the coend of α∗ ⊗ α is canonically isomorphic to C,

and if α and β are irreps and α is not isomorphic to β then the coend of α∗ ⊗ β is
0. It follows that

4.4.6 Corollary. Retaining notation from above, there is a natural isomorphism

M
A(Ygl ; γ) ∼
= A(Y ; α∗ , α, γ).
α∈L

Dually, we have that the end of α∗ ⊗ α is canonically isomorphic to C and the

end of α∗ ⊗ β is 0 if α is not isomorphic to β, for irreps α and β. Hence

4.4.7 Corollary. Retaining notation from above, there is a natural isomorphism

M
Z(Ygl ; γ) ∼
= Z(Y ; α∗ , α, γ).
α∈L

The above corollary is the old-fashioned version of the codimension-1 gluing

theorem (see e.g. [Walk91]).
4.4. GLUING CODIMENSION-1 MANIFOLDS 35

−S

+S
P
−P
+P

R Rgl

R Y Ygl
Rgl

4.4.8 Gluing n-manifolds with corners

Next we consider gluing with corners for n-manifolds.

Let Y be an n-manifold with identifications ∂Y = (−S) ∪ S ∪ R, ∂S = P ,
∂R = P ⊔ −P (see Figure (4.4.8)). Let Ygl be Y with −S glued to S. Let Rgl be
R with −P glued to P . Note that ∂Ygl = Rgl .
Choose c ∈ C(Rgl ) and assume that c is the gluing of some c♯ ∈ C(R; zb, z),
z ∈ C(P ). Then A(S; z)op × A(S; z) acts on A(Y ; ·, ·, c♯ ) and we have the following
generalization of (4.4.2).

Theorem. A(Ygl ; c) is the coend of the action of A(S; z)op ×A(S; z) on A(Y ; ·, ·, c♯ ). 4.4.9
(a) For each object x of A(S; z) there is a map

b, x, c♯ ) → A(Ygl ; c).
glx : A(Y ; x

(b) For each morphism e : x → y of A(S; z) the following diagram commutes

b, x, c♯ )
A(Y ; x
nn6
e×1nnnn
NNN
NNNglx
nnn NNN
nnn NN'
A(Y ; yb, x, c♯ ) A(Ygl ; c)
PPP p7
PPP
PPP ppppp
1×e PPP( pp
ppp gly
A(Y ; yb, y, c♯ )

(c) A(Ygl ; c) is the universal object (vector space, set, or whatever flavor of A
we’re using) with properties (a) and (b). In other words, given a W and maps
gl′x : A(Y ; x
b, x, c♯ ) → W (for all x) such that the diagram analogous to the one
in (b) above commutes for all e, there is a unique θ : A(Ygl ; c) → W such that
36 CHAPTER 4. BASIC CONSTRUCTIONS AND GLUING I

gl′x = θ ◦ glx for all x.

b, x, c♯ )
A(Y ; x
n6 NNN gl′x
e×1nnnnn NNN
nnn NN
nnn glx NNN'
'
A(Y ; yb, x, c♯ ) A(Ygl ; c) _θ _ _/7 W
PPP pp 7
PPP gly
PPP ppppp
1×e PPP( p
ppp gl′ y
A(Y ; yb, y, c♯ )

The proof does not differ in any interesting way from the proof of (4.4.2). All of
the other results in this section generalize easily to gluing n-manifolds with corners.
Note, however, that (4.4.9) is a somewhat unsatisfactory result. We want to
know not only A(Ygl ; c) for various c ∈ C(∂Ygl ), but also the action of A(∂Ygl ) on
A(Ygl ) (and also Z(Ygl )). For gluing without corners, ∂Ygl is a closed submanifold of
∂Y and knowledge of the action of A(∂Y ) on A(Y ) translates easily into knowledge
of the action of A(∂Ygl ) on A(Ygl ). For gluing with corners, this is not the case. In
order to state a more powerful result on gluing n-manifolds with corners, we first
need to understand gluing n−1-manifolds without corners. That is, we need to be
able to describe the category A(Rgl ) is terms of the categories A(R; zb, z) and the
action of the 2-category A(P ) on them. This is done in Chapter 5.

still to do in this chapter:

• need to state cat conditions for general version of gluing

• YYYYY: insert examples into later sections (A and Z defs, cylinder cats, ...)

• need to be clearer about A maybe/maybe-not being linear (examples will help

with this)

• make clear how empty boundary case works

• need to define “codim-k” near the start

• ?? retain old manifolds → cat → n-vect space thing? put it in some other
chapter?
Chapter 5

Basic Constructions and

Gluing II

This Chapter is analogous to Chapter 4, but one dimension lower and hence one
category level higher. We show that n−1-manifolds with boundary give rise to col-
lections of 1-categories; that these collections afford a representation of a 2-category
associated to the n−2-dimensional boundary of the n−1-manifold; and that these
2-category actions can be used to state and prove gluing theorems for A and Z of
1-manifolds.
[need to finish this intro once chapter is complete]

[WARNING: need to replace a with b

a and X with −X in many places]

[WARNING: in order to be consistent with other chapters, need to switch right and
left actions, and modify most of the figures]

5.1 2-Category Actions

By 2-category we mean a disk-like 2-category, as defined in (A.2.1). In particular,
there are conjugations (corresponding to homeomorphisms of I and D 2 ) defined on
the 1-morphisms and 2-morphisms, and these conjugations satisfy various identities.
Let P be a closed n−2-manifold. We define a 2-category A(P ) as follows:
• 0-morphisms (objects): C(P ).

• 1-morphisms from a to b: C(P × I; a, b).

• 2-morphisms from e to f : A(P × I × I; e, f ), where P × ∂I × I ⊂ ∂(P × I × I)

is pinched, so that ∂(P × I × I) can be identified with P × I × ∂I.

• Composition of 1-morphisms: given by the gluing (P × I) ∪P (P × I) ∼

= P × I.

37
38 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

• Composition of 2-morphisms: given by the gluing (P ×I ×I)∪P ×I (P ×I ×I) ∼

=
P × I × I.

• [need to decide whether other type of composition of 2-mors is part of the data
(as opposed to derived)]

• Associativity 2-morphism for composable 1-morphisms e, f, g: the track of an

isotopy from (ef )g to e(f g) (two distinct but isotopic fields on P × I).

• Conjugation of 1-morphisms: given by an orientation-reversing homeomor-

phisms from I to itself.

• Conjugations of 2-morphsisms: given by various homeomorphisms from I × I

to itself.

[maybe go over above in more detail at (A.2.1)]

A left representation W of a 2-category C consists of:

• For each x ∈ C 0 , a 1-category Wx .

1 , a functor W : W → W . (We usually simplify notation
• For each e ∈ Cxy e y x
and denote We by e.)

• For each e ∈ Cxy 1 and f ∈ C 1 , an invertible natural transformation W c

yz ef
between We Wf (composition of two functors) and Wef (functor associated to
the composition of two 1-morphisms). In other words, composition is only
preserved up to natural transformations, and these natural transformations
are part of the data of the representation. (We usually simplify notation and
denote Wefc by c .)
ef

• For each h ∈ Cef 2 , a natural transformation W : W → W . (We usually

h e f
simplify notation and denote Wh by h.)

• For each e ∈ Cxy1 , u ∈ W 0 and v ∈ W 0 , an adjunction (natural bijection)

x y
between mor(ev, u) and mor(v, e∗ u). If the W· ’s are linear categories, the
these adjunctions are required to be linear. [is my use of “adjunction” here
standard? if not, change or explain why I’m being slightly non-standard]

• There are various relationships between conjugations in C 2 and the adjunc-

tions, between the various cef and associator 2-morphisms, etc. [be specific
here?]

A right representation is defined similarly, reversing the direction of the C 1

1 , a functor W : W → W .
actions: for each e ∈ Cxy e x y
5.1. 2-CATEGORY ACTIONS 39

x y y

e u
x y

eu

y v
g
u
x
eg
e×I

5.1.1 The functor for e

Let S be an n−1-manifold with boundary. For each x ∈ C(∂S) we have the

1-category A(S; x), and this collection of 1-categories affords a left representation
of the 2-category A(∂S) as follows.
Let C denote A(∂S). For e ∈ Cxy 1 = C(∂S × I; x, y) we need a functor (also

denoted e) from A(S, y) to A(S, x). On objects this functor is given by gluing a
copy of e to S along ∂S (i.e. we add a collar to ∂S containing the field e). On
morphisms this functor is given by gluing a copy of e × I to S × I along ∂S × I.
(This involves unpinching parts of the boundaries; see [need ref for this? or is it
clear?].) It’s clear that composition is preserved, so this defines a functor. (See
Figure (5.1.1).)
For e ∈ Cxy1 and f ∈ C 1 we need an invertible natural transformation c
yz ef con-
necting the composition of the actions for e and f with the action of the composition
ef . This is given by the track of an isotopy, supported in a collar neighborhood of
∂S, connecting the corresponding (compositions of) collaring homeomorphisms.
For h ∈ Cef2 = A(∂S × I × I; e, f ) we need a natural transformation between the

functors for e and f . This is given, at u ∈ C(S; y), by gluing a copy of h to u × I (see
Figure (5.1.2)). The proof that this collection of morphisms actually does comprise
a natural transformation, that is, that this diagram
em
eu / ev

hu hv
 
fu / fv
fm

commutes for all u, v ∈ C(S; y) and m ∈ A(S × I; u, v), is illustrated in Figure

(5.1.3). [need to say why this works for general fields; cite specific properties from
40 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

f
h
u
e

hu

h u×I

5.1.2 The natural transformation for h

fv fv
fm hv
fu = ev
hu em
eu eu

5.1.3 Verifying the natural transformation

the field defs; clearly works for our basic examples (pictures and maps); or, better,
do the proof in the fields section/chapter]
For e ∈ Cxy 1 , u ∈ C(S; x) and v ∈ C(S; y). we need a bijection (isomorphism)

between A(S × I; ev, u) and A(S × I; v, e∗ u). This is given by an isotopy of S × I

which shifts a collar of ∂S from S × {0} to S × {1}; see Figure (5.1.4). (Note that
this is one of the few times we consider isotopies of manifolds which do not fix the
boundary.) This completes the proof that {A(S; ·)} affords a left representation of
A(∂S). We will denote the entire representation package by A(S).
A similar argument shows that {Z(S; ·)} affords a right representation of A(∂S).
Denote this representation by plain Z(S). If, for an n−2-manifold P , we define Z(P )

ē u u

↔
v e v

5.1.4 Adjunction isotopy

5.2. GLUING 41

−P x P P

u glx (u)

P ×I

m glx (m)

5.2.2 Gluing along P

to be the 2-category of all right representations of A(P ) [need to say what sort of
2-cat this is and why], then we have, for all n−1-manifolds S,

Z(S) ∈ Z(∂S) 5.1.5

This is a categorified version of (4.3.2) and a doubly categorified version of (1.2.1).

[also forward ref to combinatorial version of (n + 1) gluing?]

5.2 Gluing
Next we use the above representations to prove gluing theorems for A and Z of
n−1-manifolds. Suppose we have the usual gluing (without corners) scenario: an
n−1-manifold S with an identification ∂S = (−P )⊔P ⊔Q. For notational simplicity
we will henceforth ignore Q. We have an action of the 2-category A(−P ⊔ P ) ∼ =
op
A(P ) × A(P ) on {A(S; ·, ·)} (equivalently, commuting right and left actions of
A(P )), and we would like to compute the 1-category A(Sgl ) from this action.
We will start by noting some relationships between A(Sgl ) and A(S). We will
then prove that A(Sgl ) is the universal 1-category with these properties (what one
might (and we will) call a categorified coend or 2-coend construction).
To simplify notation (and also allow for its reuse), denote the 2-category A(P )
by A, denote the 1-category A(S; x, y) by Wxy (where x, y ∈ A0 = C(P )), and denote
the 1-category A(Sgl ) by C.
For each x ∈ A0 we have a functor

glx : Wxx → C. 5.2.1

On objects this functor is given by gluing S along P to obtain Sgl . On morphisms

it’s given by gluing S × I along P × I to obtain Sgl × I. (See Figure (5.2.2).)
 42 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

−P P
x y x y

e
u

P ×I

gly (ue)
sex
glx (eu)

5.2.4 se at x

For each e ∈ A1xy we can construct two functors from Wxy to C, gly ◦(e × 1) and
(1 × e) ◦ glx , and there is an invertible natural transformation se between these two
functors:

Wyy
y< BB
BBgly
e×1 yyy
BB
yy BB
yy
Wxy se =C
FF ||
FF |
FF |
1×e FF ||
" || glx
5.2.3 Wxx

(This is a 2-dimensional diagram with each i-cell labeled by an i-morphism. Here

and below we follow the convention of placing 2-morphisms (e.g. natural transfor-
mations) in the appropriate 2-cell of a diagram with a surrounding box instead of
an accompanying arrow. This is in order to keep the diagrams from getting too
cluttered. The range and domain of the 2-morphism is usually clear from context.)
The morphisms for se are tracks of isotopies (in Sgl × I) which shift a copy of e
across the gluing locus P . (See Figure (5.2.4).) It is easy to verify that this is a
natural transformation.
5.2. GLUING 43

(1 × h) • idglx

sf
=
se

(h × 1) • idgly

5.2.6 A commutativity relation

For all h ∈ A2ef , the following diagram commutes:

1 Wyy sf
I >>
f ×1 >>
>>
>> gl
>> y
h×1 >>
>>
>>
e×1 >>

Wxy se @C
1×e

1×h glx

1×f 
.W 5.2.5
xx

The above diagram should be thought of as a cell decomposition of the 2-sphere

with four 2-cells. (The 2-cell “at infinity” is labeled by sf .) There are two different
ways of composing the natural transformations in the diagram, ((1 × h) • idglx ) ◦ se
and sf ◦ ((h × 1) • idgly ). (Here • denotes “horizontal” composition of natural
transformations (see (A.2.2)) and idglx denotes the identity natural transformation
from the functor glx to itself.) The commutativity of the diagram means that
these give the same natural transformation between the functors gly ◦(e × 1) and
glx ◦(1 × f ). Roughly speaking, the shift morphisms se commute with the (gluings
of) the 2-morphism actions h. The proof of this assertion is indicated in Figure
(5.2.6).
 44 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

f e fe

1 × cf e
sf e

= se

cf e × 1
sf

5.2.8 Another commutativity relation

For all e ∈ A1xy and f ∈ A1yz , the following diagram commutes:

.W
(f e)×1 x < zz==
x ==
cef × 1 xx x ==
xxx f ×1 == sef
==glz
Wyz ==
y< EE
EE1×f sf ==
y yy EE ==
y E ==
y e×1 E"
yy gly
=
WxzE id Wyy /C
@
EE y <

EE1×f yyy


EE yy


E" yy e×1 se



Wxy


FF


glx
FF1×e

1 × cef FF
FF



1×f e "


5.2.7 0 Wxx

Again, the diagram should be thought of as a cell decomposition of the 2-sphere,

this time with six 2-cells. The functors for the two sides of the 2-cell labeled “id”
are both e × f , and so are connected by the identity natural transformation. There
are two ways of composing the natural transformations in the diagram to obtain
a natural transformation between the functors glz ◦(f e × 1) and glx ◦(1 × f e), and
the commutativity of the diagram means that these two natural transformations are
equal. Roughly speaking, this means the the shift morphisms preserve composition
of 1-morphisms; sef can be computed from se and sf . The proof of this assertion is
indicated in Figure (5.2.8).
It turns out that the above properties uniquely characterize A(Sgl ). To make
this statement more precise, we need a definition.
5.2. GLUING 45

Definition. Let A be an arbitrary disk-like 2-category, and let {Wxy } be a collection 5.2.9
of 1-categories affording an Aop × A action. A 1-category C, together with functors
{glx } and invertible natural transformations {se } satisfying (5.2.1), (5.2.3), (5.2.5)
and (5.2.7), is called the 2-coend of the Aop × A action if it is universal in the
following sense. If C ′ , {gl′x } and {s′e } also satisfy (5.2.1) through (5.2.7), then
there exists a functor θ : C → C ′ and, for all x ∈ A0 , a natural transformation
ηx : θ ◦ glx → gl′x , such that

gl′y
Wyy
y< BB s′e
e×1 yyy BB ηy
BB
yy gly BB
yy 
θ
Wxy se =C
/ C′
FF H
FF glx ||
FF ||
1×e FF || ηx
" ||
Wxx
gl′x

commutes for all e ∈ A1xy .

[need to talk about uniqueness of 2-coend]

Theorem. A(Sgl ) is the 2-coend of the A(P )op × A(P ) action on {A(S; x, y)}. 5.2.10

Proof. We will introduce two new categories, G (G for generators and relations) and
P (P for parallelogram). Both will have concrete algebraic descriptions. It will be
easy to show that G is (up to natural isomorphism, of course) the 2-coend of the
theorem, and that P is isomorphic to A(Sgl ). We then show that G and P are
isomorphic. (So G and P provide two alternative, more concrete descriptions of the
2-coend.)
S 0 . The morphisms will be
First we define G. The objects of G are x∈A0 Wxx
defined
S in terms of generators and relations. There are two types of generators,
1 0
x∈A0 Wxx (with the obvious choice of range and domain), and, for all x, y ∈ A ,
1 0
e ∈ Axy and u ∈ Wyx , morphisms

σeu : eu → ue
−1
σeu : ue → eu

The morphism of G corresponding to f ∈ Wxx 1 will be denoted fb, unless there is no

chance of confusion, in which case we denote it as plain f . The relations are

• fbgb = fcg 5.2.11

−1 = id
• σeu σeu −1
eu and σeu σeu = idue 5.2.12
 46 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

5.2.13 • σe· is a natural transformation between the right and left actions of e ∈ A1xy :
c σev = σeu fce
for all f ∈ (Wyx )1uv , ef

5.2.14 • for all h ∈ A2ee′ , σeu (h × 1)u = (1 × h)u σe′ u

• for all e ∈ A1xy and g ∈ A1yz and u ∈ Wzx

0 , σ −1
5.2.15 e,gu σg,ue = α σeg,u α, where α is
the associator for e and g at u

Next we show that G is the 2-coend of the A × Aop action. There are obvious
functors glx : Wxx → G for all x (sending f to fb), and, with σe· playing the role of
se , the above relations for G guarantee that these data satisfy (5.2.1), (5.2.3), (5.2.5)
and (5.2.7). Let C ′ , {gl′x } and {s′e } be as in the definition of 2-coend. We must
define a functor θ : G → C ′ satisfying the conditions of (5.2.9). On the objects of G
define θ so that θ ◦ glx = gl′x for all x. On the generating morphisms coming from
Wxx again define θ so that θ ◦ glx = gl′x for all x. Finally, define θ(σeu ) = s′e,u and
θ(σeu−1 ) = (s′ )−1 .
e u
We must show that the above definition of θ on the generators of G 1 respects
the relations. Since the relations are just translations of the commutative diagrams
defining the 2-coend, this is easy to do. (5.2.11) follows from the fact that gl′x is
a functor. (5.2.12) follows from the fact that s′e and (s′e )−1 are mutually inverse.
(5.2.13) follows from the fact that s′e is a natural transformation. (5.2.14) follows
from (5.2.5). (5.2.15) follows from (5.2.7). Thus θ is a well-defined functor.
Finally, we must define, for all x, an invertible natural transformation ηx : gl′x →
θ ◦ glx . Since these two functors are equal, we can take ηx to be the identity natural
transformation. S
Next we define P. The objects of P are again x∈A0 Wxx 0 . The morphisms from
0
u ∈ Wxx to v ∈ Wyy are0

 ,
M  
 mor(eu, ve) p ◦ (h • idu ) ∼ (idv •h) ◦ p
5.2.16 e∈A1yx

Here mor(eu, ve) = (Wyx )1eu,ve and h ∈ A2ef . The equivalence class of a ∈ mor(eu, ve)
1 . We think of a morphism of P as a parallelogram with
will be denoted gle (a) ∈ Puv
short sides labeled by e and long sides labeled by u and v. See Figure (5.2.17).
(More literally, a morphism of P can be thought of as a field on Sgl × I restricting
to e on P± × I.)
[need to replace most (all?) occurrences of idblah • with simple juxtaposition]
To complete the definition of P we must specify how to compose morphisms. Let
p ∈ mor(eu, ve) and q ∈ mor(f v, wf ), where e ∈ A1xy , f ∈ A1zx , u ∈ Wxx
0 , v ∈ W0 ,
yy
0 . Then define
w ∈ Wzz

glf (q) ◦ gle (p) = glef ((qe) ◦ (f p)).

5.2. GLUING 47

h
e
v v h

h e ∼ f
e f

u u

5.2.17 Morphisms for P

(There should be some associators inserted above, but they have been omitted for
simplicity.) See Figure (5.2.18).
The objects of P are essentially all of the objects of A(Sgl ). (More precisely,
they are all of the objects which are transverse to the gluing surface.) It follow from
(4.4.4) that the morphisms of the two categories are identical. So all that remains
is to show that the two composition rules coincide. This is illustrated in Figure
(5.2.19). The basic idea is that the equivalence relation on morphisms of P allows
us to eliminate ide and idf from the definition of composition in P, yielding the
result of composing in A(Sgl ).
Now we must show that G and P are isomorphic. Predictably, we define functors
α : G → P and β : P → G and show that α ◦ β is the identity functor on P and β ◦ α
is the identity functor on G. Note that G 0 = P 0 , so we can (and do) define both α
and β to be the identity on objects.
First some notation. For e ∈ A1xy , let Vee∗ ∈ A2 be the 2-morphism conjugate to
ide with domain idx and range ee∗ . Let Λee∗ ∈ A2 be the 2-morphism conjugate to
ide with domain ee∗ and range idx . See Figure (5.2.20). [put these defs earlier and
just recall them here??]
Keeping the topological interpretation of the algebra in mind, it’s easy to define
−1 ) = gl (id 1
α. Let α(σeu e eue ) and α(σeu ) = gle∗ (Λe∗ e • idu •Vee∗ ). For h ∈ (Wxx )uv ,
let α(b
h) = glidx (h̃), where h̃ is the morphism from idx •u to v • idx guaranteed by
the definition of 2-category action. [need to make sure we included this in the def]
See Figure (5.2.21). We must verify that the above assignments of generators of
G 1 obey the relations. This straightforward exercise is left to the reader.
Defining β : P → G is a little more complicated. The key topological idea is that
an arbitrary field on Sgl × I is isotopic to a composition of fields (morphisms) which
either: (a) are product fields near P± × I ⊂ Sgl × I, and so are gluings of fields on
S × I; or (b) are tracks of isotopies of Sgl which shift a collar neighborhood across
P± . The proof of this is illustrated in Figure (5.2.22). Accordingly, we define, for
48 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

w v

q f p e
f e
v u

e×I
qe fp
f ×I

w
f
f qe

e fp e

5.2.18 Composition in P

w w
f f
f q
f
= e
e e
e

u u

5.2.19 Proof that composition rules are the same

x y

e
x×I e ē
Λeē Veē
e ē x×I

5.2.20 Vee∗ and Λee∗

5.2. GLUING 49

e u
e u

−1 e
u α(σeu )
e e
α(σeu ) ē e
u
ē
e u
v
x x
h
v u
x×I
α(h)

x×I u

5.2.21 Definition of α : G → P

[No figure yet]

5.2.22 Decomposing a morphism of P

 50 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

a ∈ (Wyx )1eu,ve ,
β(gle (a)) = (id\ b′
v •Λee∗ ) ◦ σe∗ ,ve ◦ a ,
where a′ is the adjoint of a with domain u and range e∗ ve. [here we ignore some
morphisms coming from the action of identity 1-morphisms of A; need to say this
more precisely; or maybe include those morphisms]
Now that α and β have been defined, we must verify that they are mutually
inverse. [relatively easy for P → G → P; harder for G → P → G; proofs illustrated
in figs xxxx, yyyy, zzz]

There is a dual version of the codimension-2 gluing theorem (5.2.10). Retaining

notation from above, we have
5.2.23 Theorem. Z(Sgl ) is the 2-end of the A(P )op × A(P ) action on {Z(S; x, y)}. 
The A(P )op × A(P ) action on {Z(S; x, y)} needs no explanation. The proof of
(5.2.23) is dual to the proof of (5.2.10). All that remains is to define the 2-end of
the A(P )op × A(P ) action. This dual to the definition of 2-coend above, but we
repeat the (dualized) details below.
5.2.24 Definition. Let A be an arbitrary disk-like 2-category, and let {Wxy } be a collection
of 1-categories affording an Aop × A action. A 1-category C, together with functors
rx : C → Wxx
(for all x ∈ A0 ) and invertible natural transformations {se }
Wyy
y `BB
e×1 yyy BBry
y BB
y BB
|yy
Wxy se C
Fb F |
FF
FF |||
1×e FF ||r
}|| x
Wxx
(for all e ∈ A1 ) satisfying (5.2.25) and (5.2.26) below, is called the 2-end of the
Aop × A action if it is universal in the following sense. If C ′ , {rx′ } and {s′e } also
satisfy (5.2.25) and (5.2.7), then there exists a functor θ : C ′ → C and, for all
x ∈ A0 , a natural transformation ηx : rx ◦ θ → rx′ , such that
′
ry
s
Wyy
y B` B s′e
e×1 yyy BB ηy
y BB
y r y BB
|yy
θ
Wxy se Co C′
Fb F |
FF rx |||
FF
1×e FF || ηx
|} |
Wxx k ′
rx
5.2. GLUING 51

commutes for all e ∈ A1xy .

As referred to in the above definition, we have for all h ∈ A2ef ,

Wyy sf
>^ >
f ×1 >>
>>
>> r
>> y
h×1 >>
>>
>>
e×1 >>
Wxy nq se C
U
1×e

1×h rx

1×f
Wxx 5.2.25
Also, for all e ∈ A1xy and f ∈ A1yz ,

(f e)×1
Wzz=^
xx ==
cef × 1 xxx ==
==
x
|xx f ×1 == r sef
Wyz == z
bEE sf ==
yyy EE1×f ==
yye×1 EE ==
yy EE ==
 |y ry
Wxz
P bEE id Wyy o C
EE1×f yyy




EE y


EE yy
y| y e×1 se



Wxy

Fb F


rx
FF1×e


1 × cef FF
FF



1×f e


Wxx 5.2.26

It will be useful to have a more concrete description of the 2-end. Suppose we

have a (not necessarily universal) C ′ , {rx′ } and {s′e } satisfying the conditions above.
def
To each object a of C ′ we can associate the collection of objects {ax = rx′ (a) ∈ Wxx
0 }
0
(indexed by x ∈ A ) as well as the collection of morphisms {se,a : eay → ax e}
(indexed by e ∈ A1xy ). [need to fix: should be s′e,a] It follows from (5.2.25) that for
all h ∈ A2ef the following diagram commutes
se,a
eay / ax e

h•id id •h
 sf,a 
f ay / ax f 5.2.27
 52 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II

It follows from (5.2.26) that for all e ∈ A1xy and f ∈ A1yz the following diagram
commutes
sef,a
(ef )az / ax (ef )

cL
ef cR
ef
 esf,a se,a f 
5.2.28 a(f az ) / eay f / (ax e)f

where cL R
ef [cef ] is the associator associated to the left [right] action of A on W .
def
For each morphism m : a → b of C ′ we have a collection of morphisms {mx =
′
rx (m) : ax → bx } (indexed by x ∈ A0 ). Because se is a natural transformation, the
following diagram commutes for all e ∈ A1

se,a
eay / ax e

emy mx e
 se,b 
5.2.29 eby / bx e

We are thus led to define a 1-category E whose objects are collections {ax ∈
0 } and {s
Wxx e,a : eay → ax e} satisfying (5.2.27) and (5.2.28). Morphisms of E are
collections {mx : ax → bx } satisfying (5.2.29). Composition is given by {mx } ◦
def
{nx } = {mx ◦ nx }. (It is easy to verify that {mx ◦ nx } satisfies (5.2.29).)
There are obvious functors rx : E → Wxx (for all x ∈ A0 ), and the individual
morphisms se,a fit together to give natural transformations se as in (5.2.24). [no-
tational problem here: se,a plays two roles, one from def of E and one from def of
2-end. should fix this.] The above discussion motivating the definition of E can also
be read as a proof that E, {rx } and {se } are universal in the appropriate sense.
Thus we have the desired concrete description of the 2-end.
One of the simplest examples of a 2-end is when A above is a 2-category with
only one object (i.e. a spherical tensor 1-category) and Aop × A acts on itself via left
and right tensor multiplication. The 2-end of this action has a more familiar name:
the Drinfeld center of the tensor 1-category A. Indeed, if we specialize the definition
of E above to this case we obtain the usual definition of the Drinfeld center. (See
for example [Kas95, p. 330]. Note however that we make no assumptions about
strict associativity.) [D. center works for general tensor 1-cat (don’t need spherical).
comment on this?]
5.2.30 So a special case of the codimension-2 gluing theorem (5.2.23) is the following.
Suppose P is a closed n−2-manifold such that the 2-category A(P ) has only one
object, and thus can be thought of as a tensor 1-category. Then Z(P ), the repre-
sentations of A(P ), can also be thought of as a tensor 1-category, and Z(P × S 1 ) is
the Drinfeld center of Z(P ). [need to say more here; need additional assumptions
on A(P )]
5.2. GLUING 53

Note that for any closed P the category Z(P × S 1 ) has the structure of a braided
tensor category. Let X 2 be a twice-punctured disk. Then Z(P ×X) can be thought of
as a functor Z(P × S 1 ) × Z(P × S 1 ) → Z(P × S 1 ), which gives a tensor (monoidal)
structure to Z(P × S 1 ). The geometric braiding of X gives rise to an algebraic
braiding of Z(P × S 1 ). [need to show that in special case of D. center this def of
tensor and braiding agrees with the usual one.] [need to go into more detail on
above]

Still to do:

• remark somewhere that we will (to deobfuscate notation) sometimes fail to

distinguish between morphisms and their conjugates; e.g. add or subtract,
without comment, actions of identity 1-morphisms of A, omit associators, etc.

• annularization

• gluing n-manifolds with corners (revisit)

• discuss higher codim case? or put it in separate chapter?

• refer forward to applications in example chapters

• in earlier (fields) section, prove a few relevant isotopy identities carefully, using
field axioms (and refer to that from this chapter)
54 CHAPTER 5. BASIC CONSTRUCTIONS AND GLUING II
Chapter 6

From Local Relations to Path

Integrals

[this an incomplete draft at the moment; also, it hasn’t been proof-read]

In Chapter 1 we showed (non-rigorously) that topologically invariant path inte-

grals lead to local relations. In this chapter we show (rigorously) how to reconstruct
the path integral from local relations (Theorem (6.3.1)). In order for this reconstruc-
tion to work the local relations need to satisfy some relatively mild semi-simplicity
and finiteness conditions.
The main ideas are that, as argued in Chapter 1, local relations carry essentially
the same information as the path integral of the n+1-ball, and that by gluing (with
corners) n+1-balls together and applying the gluing formula for the path integral,
we can compute the path integral of any n+1-manifold. After these observations all
that remains is to show that these computations yield consistent results.

[mention pairings too?]

6.1 Axioms for Path Integrals

By “path integral” we mean an invariant of n+1-manifolds (denoted as usual by Z)
which relates to the rest of the TQFT as described below. (See Chapter 1 for an
explanation of why one would expect these formal properties to arise from actual
integrals.)

First, the path integral Z(M n+1 ) is a function on fields on its boundary com- 6.1.1
patible with local relations, so

Z(M ) ∈ Z(∂M )

55
 56 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

for all n+1-manifolds M . (If ∂M = ∅, this means that Z(M ) ∈ C, or whatever the
ground ring is.) Equivalently, Z(M ) is a function

Z(M ) : A(∂M ) → C.

6.1.2 Second, for all n-manifolds Y and boundary conditions c ∈ C(∂Y ), we have
nondegenerate pairings

c) → C
Z(Y ; c) ⊗ Z(−Y ; b
x ⊗ y 7→ hx, yi

and

c) → C
A(Y ; c) ⊗ A(−Y ; b
x ⊗ y 7→ hx, yi.

The pairings should be compatible with the actions of A(∂Y ): for all x ∈ A(Y ; c),
y ∈ A(−Y ; bb) and e ∈ A(∂Y )1cb = A(∂Y × I; b
c, b),

6.1.3 hxe, yi = hx, eyi,

and similarly for Z(Y ; c). (Here we are using the identification A(−∂Y ) = A(∂Y )op ,
so A(∂Y ) has a left action on A(−Y ; ·).) More generally, we could replace ∂Y above
with a codimension-0 submanifold of ∂Y .
Recall that associated to reversing the orientation of Y we have an isomorphism
(conjugate linear if the ground ring is C) A(Y ; c) ∼ = A(−Y ; b c). Combining this
isomorphism with the above pairings we get [sesquilinear] inner products

A(Y ; c) ⊗ A(Y ; c) → C
def
x ⊗ y 7→ hx, yi = hx, ybi

and similarly

Z(Y ; c) ⊗ Z(Y ; c) → C
def
x ⊗ y 7→ hx, yi = hx, ybi.

(Note that we have “overloaded” the angle brackets h·, ·i to denote both the pairings
and the inner products. Which meaning is intended can be deduced from what’s
being plugged into the angle brackets.) We require these inner products to be [skew]
symmetric. The inner products induce isomorphisms between A(Y ; c) and Z(Y ; c)
(recall that these spaces are mutually dual), and these isomorphisms should preserve
the pairings.
(In all of the examples we will study, A(Y ; c) and Z(Y ; c) will be finite dimen-
sional, so we won’t worry about completeness.)
6.1. AXIOMS FOR PATH INTEGRALS 57

−Y +Y

W Wgl

W M Mgl
Wgl

6.1.5 Yet another figure illustrating gluing

Finally, there is a gluing relation for n+1-manifolds. Let M be an n+1-manifold 6.1.4

with boundary ∂M = Y ∪ −Y ∪ W and let Mgl be M glued (with corners) along ±Y
(see Figure (6.1.5)). Note that ∂Mgl = Wgl , where Wgl is W glued along ∂(±Y ).
For each c ∈ C(∂Y ) we have maps

cr ctr
Z(∂M ) −→ Z(Y ; c) ⊗ Z(−Y ; b
c) ⊗ Z(W ; b
c, c) −→ Z(W ; b
c, c),

where rc is a restriction map and trc comes from the pairing for Z(Y ; c). Call the
composite map ϕc : Z(∂M ) → Z(W ; b c, c). Recall from (4.4.5) that
Y
Z(Wgl ) = {(vx ) ∈ b, x) | vx e = evy for all e ∈ A(∂(Y ))1 }.
Z(Y ; x
x∈C(∂Y )

Then the gluing relation for n+1-manifolds states that (ϕx (Z(M ))) determines an
element of Z(Wgl ) (i.e. ϕx (Z(M ))e = eϕy (Z(M )) for all e ∈ A(∂(Y ))1xy ) and that
Z(Mgl ) is equal to this element of Z(Wgl ) = Z(∂M ). More succinctly,

Z(Mgl ) = trY (Z(M )), 6.1.6

where trY : Z(∂M ) → Z(∂Mgl ) is induced by the various ϕc (see (6.2.2) below).
Here’s a more concrete version of the gluing relation. Assume that A(Y ; c) is 6.1.7
finite dimensional. Fix c ∈ C(∂Y ) and let x ∈ A(W ; b c, c). We want to evaluate the
function Z(Mgl ) on the glued field glc (x) ∈ A(∂Mgl ) in terms of Z(M ) evaluated on
elements of A(∂M ) = A(Y ∪ −Y ∪ W ). Let {ei } be a basis of A(Y ; c) and let {ei }
def def
be the dual basis of Z(Y ; c). Let gij = hei , ej i and gij = hei , ej i. (Note that the
matrices (gij ) and (gij ) are mutually inverse.) Then straightforward linear algebra
shows that the above gluing relation is equivalent to
X
Z(Mgl )(glc (x)) = Z(M )(ei ∪ ebj ∪ x)gij .
i,j 6.1.8
 58 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

If {ei } is an orthogonal basis then this becomes

X 1
Z(Mgl )(glc (x)) = Z(M )(ei ∪ ebi ∪ x) ,
6.1.9 hei , ei i
i

and if {ei } is an orthonormal basis it becomes

X
Z(Mgl )(glc (x)) = Z(M )(ei ∪ ebi ∪ x).
6.1.10 i

6.2 Consequences of the Axioms

In this section we note some consequences of the above axioms. We will later use
the reasoning of this section to prove that if we have defined a path integral for
n+1-manifolds composed of handles of index less than k, then we can extend the
definition to n+1-manifolds composed of handles of index less than k + 1. Thus,
starting with the path integral of the n+1-ball we construct inductively the path
integral for any n+1-manifold.
The inner product on A(Y ; c) is determined by Z(Y × I). Indeed, applying
(6.1.8) to the gluing (Y × I) ∪Y (Y × I) = Y × I we have, for a basis {ei } of A(Y ; c),
X
Z(Y × I)(ebk ∪ el ) = Z(Y × I ⊔ Y × I)(ebk ∪ ei ∪ ebj ∪ el )gij
i,j
X
= Z(Y × I)(ebk ∪ ei ) · gij · Z(Y × I)(ebj ∪ el ),
i,j

which implies that Z(Y × I)(ebj ∪ ei ) gives the inverse of the matrix gij . In other
words,
hei , ej i = Z(Y × I)(ebi ∪ ej ),
which implies (since the basis {ei } was arbitrary) that

6.2.1 hx, yi = Z(Y × I)(b

x ∪ y)

for all x, y ∈ A(Y ; c). Note that if we use (6.2.1) to define the inner product then it
is automatically compatible with the boundary category actions as in (6.1.3):

hxe, yi = Z(Y × I)(b

xeb ∪ y)
= Z(Y × I)(b
x ∪ ey) (because x
beb ∪ y = x
b ∪ ey)
= hx, eyi.

6.2.2 If (6.2.1) holds for Y then the gluing formulas of (6.1.4) automatically yield
functions on C(∂Mgl ) which lie in Z(∂Mgl ). Let b, c ∈ C(∂Y ), m ∈ A(∂Y × I; bb, c)
6.2. CONSEQUENCES OF THE AXIOMS 59

x x
b b m c c

ek fi
M Y ×I M
el fj
b b m c c
x x

6.2.3 Lots of boundary labels

and x ∈ A(W ; b, b c). Let {ei } be a basis of A(Y ; b) with dual inner product matrix
gij . Let {fi } be a basis of A(Y ; c) with dual inner product matrix hij . (See Figure
(6.2.3).) Then
X
Z(Mgl )(glc (xm)) = Z(M )(fi ∪ fbj ∪ xm)hij
i,j
X
= Z(M )(ek ∪ fbj ∪ x)gkl hel , mfi ihij
i,j,k,l
X
= Z(M )(ek ∪ fbj ∪ x)gkl hel m, fi ihij
i,j,k,l
X
= Z(M )(ek ∪ ebl ∪ mx)gkl
k,l
= Z(Mgl )(glb (mx)).

Let R be an n−1-manifold and c ∈ C(∂R). We have inner products on 6.2.4

A(R; c)1ab = A(R × I; b a, b) for all a, b ∈ A(R; c)0 = C(R; c). Applying (6.1.3) to
R × {0} and R × {1} ⊂ ∂(R × I), and recalling the definition of x∗ from (4.3.1), we
have
hxy, zi = hx, zy ∗ i = hz ∗ x, y ∗ i = hz ∗ , y ∗ x∗ i = hyz ∗ , x∗ i = hy, x∗ zi
for all x, y, z ∈ A(R; c)1 such that the above expression is defined.
Assume now that A(R; c) is semisimple. Let α be an irrep of A(R; c) and let
ex ∈ A(R; c)1xx and ey ∈ A(R; c)1yy be two minimal idempotents for α (see (B.5.2)).
Then there exists u ∈ A(R; c)1xy such that ex = uu∗ and ey = u∗ u. It follows that

hex , ex i = huu∗ , uu∗ i = hu, uu∗ ui = hu∗ u, u∗ ui = hey , ey i.

It is also easy to see that if e, f ∈ A(R; c)1xx are minimal idempotents for two non-
isomorphic irreps, then he, f i = 0.
Next we describe how gluing n-manifolds affects the pairings. Let ∂Y = R ∪ 6.2.5
S ∪ −S, and b ∈ C(R). (For simplicity we are suppressing from the notation labels
for ∂S and ∂R = ∂S ⊔ ∂(−S).) We want to compute the pairing for A(Ygl ; gl(b))
 60 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

M1 W3
Y31 M3
W1 Y12
Y23

M2

W2

6.2.7 Gluing three at a time

in terms of the pairings for A(Y ; b, b

a, a) for various a ∈ C(S). Assume that A(R)
is semisimple, and let {ei } be a complete set of minimal idempotents for A(R).
Let {fij } be an orthogonal basis of A(Y ; b, ei , e∗i ). [need to introduce this notation
somewhere] It follows from [need ref] that {gl(fij )} is a basis of A(Ygl ; gl(b)). To
compute the inner product on A(Ygl ; gl(b)) we use (6.1.9) to compute Z(Ygl × I),
which is obtained by gluing Y × I to itself along a copy of R × I. Note that ei is
part of an orthogonal basis {gk } of A(R × I, ∂ei ) with the property that ei gl ei = 0
unless gl = ei . It now follows from (6.1.9) that
δik hfij , fil i δik δjl hfij , fij i
hgl(fij ), gl(fkl )i = = .
hei , ei i hei , ei i
In other words, {gl(fij )} is an orthogonal basis of A(Ygl , gl(b)), but we need to adjust
by a correction factor hei , ei i−1 .
(We remark that these are the same inner product correction factors that appear
in [Walk91]. One advantage of the present approach over that of [Walk91] is that
here the correction factors arise naturally and explicitly.)
6.2.6 We can now derive a symmetric formula for gluing three n+1-manifolds together.
Let M = M1 ∪ M2 ∪ M3 , Yij = Mi ∩ Mj , and ∂M1 = Y12 ∪ Y31 ∪ W1 (and similarly
for M2 and M3 ). Let R = M1 ∩ M2 ∩ M3 . Note that ∂M = W1 ∪ W2 ∪ W3 . (See
Figure (6.2.7).) Choose fields ci ∈ C(Wi ) which can be glued to yield a field
c = c1 ∪ c2 ∪ c3 ∈ C(∂M ). Let b ∈ C(∂R) be the common restriction of the ci ’s
to ∂R. Assume that A(R; b) is semisimple. Let {ei } be a complete set of minimal
idempotents for A(R; b). Let {fij12 } be an orthonormal basis of A(Y12 ; ei ), and define
31 } and {f 23 } similarly. (We are suppressing from the notation the restriction of
{fik il
c to ∂(Y12 ) r R.) It follows from (6.1.10) that
X
Z(M1 ∪ M2 )(fbik 31
∪ fil23 ) = Z(M1 )(fbik
31
∪ fij12 ) · Z(M2 )(fbij12 ∪ fil23 ).
j
6.3. COMPUTING THE PATH INTEGRAL 61

(We continue to suppress from the notation c and its various restrictions.) It now
follows from (6.2.5) that
X
Z(M1 ∪ M2 ∪ M3 )(c) = Z(M1 ∪ M2 )(fbik
31
∪ fil23 ) · Z(M3 )(fbil23 ∪ fik
31
) · hei , ei i−1
i,k,l
X
= Z(M1 )(fbik
31
∪ fij12 ) · Z(M2 )(fbij12 ∪ fil23 ) · Z(M3 )(fbil23 ∪ fik
31
) · hei , ei i−1 .
i,j,k,l

[need to adjust alignment above] Note that the above expression is symmetric in
1,2,3. In other words, if we applied the gluing formula in a different order (say
by first gluing M1 and M3 , then gluing M2 to M1 ∪ M3 ), the final answer for
Z(M1 ∪ M2 ∪ M3 ) would not change. This self-consistency property of the gluing
formula will be used in the proof of (6.3.1).
While the above expression might look complicated, the idea is simple: Orthog-
onally decompose the three spaces A(Yxy ; ·) according to the minimal idempotents
(irreps) of A(R; b), then do the obvious tensorial contractions, but adjust them by
a correction factor hei , ei i−1 . There is an obvious generalization for any number of
Mi ’s glued together around a “corner” R.
[need to say something about total finiteness in an n-cat]

[need to revisit this section once the following section is complete]

6.3 Computing the Path Integral

Now, after much foreshadowing, we can state and prove

Theorem. Suppose 6.3.1

1. there exists z ∈ Z(S n ) such that the induced inner product A(B n ; c)⊗A(B n ; c) →
C given by a ⊗ b 7→ z(b a ∪ b) is positive definite for all c ∈ C(S n−1 ); and

2. dim A(Y n ; c) < ∞ for all n-manifolds Y and all c ∈ C(∂Y ).

Then there exists a unique path integral Z(M n+1 ) ∈ Z(∂M ) (for all n+1-manifolds
M ) satisfying the axioms of (6.1) and such that Z(B n+1 ) = z.

(Note that we have assumed above standard identifications S n = ∂B n+1 , S n =

B n ∪ B n , and S n−1 = ∂B n .)
[Need to replace assumption that dim A(Y n ; c) < ∞ with some finiteness properties
of the n-cat Z(pt). (But the present version is also useful, so...??) Also, relax
positive definite assumption??]
62 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

Proof. We will define Z(M ) by decomposing M into handles (each of which is home-
omorphic to B n+1 ) and applying the gluing formula (6.1.7). The challenge is to show
that this computation is independent of the choice of handle decomposition. We will
do this by inducting on the index of the handles.
First, some terminology. By a (k, l)-body we mean a k-dimensional manifold
equipped with a handle decomposition with all handles of index ≤ l. We will use the
same letter to denote both a (k, l)-body and the underlying k-dimensional manifold;
which meaning is intended should be clear from context. If X is a (k, l)-body then
X × I will denote the (k + 1, l)-body given by thickening all the handles (increasing
their dimension by one while leaving the index unchanged).
Our inductive hypotheses are

1. For each (n + 1, i)-body M we have defined a Z(M ) ∈ Z(∂M ) which is un-

changed by handle slides and handle cancellations of index ≤ i.

2. For each (n, i)-body Y , A(Y ; c) is finite-dimensional for all c ∈ C(∂Y ) and we
have defined a positive definite inner product on A(Y ; c) which is unchanged
by handle slides and handle cancellations of index ≤ i.

3. For each (n − 1, i)-body R, A(R; b) is semisimple with finitely many irreps for
all b ∈ C(∂R).

The hypotheses in the statement of the theorem imply the inductive hypotheses
for i = 0. Note that a (k, 0)-body is homeomorphic to a disjoint union of copies of
B k , and there are no handle slides or cancellations of dimension ≤ 0 to consider.
A(B n−1 ; c) is semisimple because of the positive definite inner product. It has finitely
many irreps because otherwise A(B n−1 × S 1 ; c × S 1 ) would be infinite dimensional
[refer to relevant gluing theorem].
We’ll verify the inductive hypotheses for i = 1, then i = 2, then the general case.
Let M be an (n + 1, 1)-body. Choose an ordering of the 1-handles, then use this
ordering to define Z(M ) ∈ Z(∂M ) by attaching each 1-handle in turn and applying
(6.1.7). (Here we use the fact that A(B n × S 0 ; c) is finite dimensional for all c.)
Since the attaching targets of the 1-handles are disjoint, this calculation is clearly
independent of the ordering of the 1-handles. Consider a slide of a 1-handle α over
a 1-handle β. If α come before β in the ordering then it is clear that the calculation
of Z(M ) using (6.1.7) is not affected, since if α is already attached then sliding β
is simply an isotopy of the attaching target for β. Therefore Z(M ) is unaffected by
handle slides. Next consider a 0-handle α which is canceled by a 1-handle β. Let
α′ be the 0-handle at the other end of β. Instead of gluing β simultaneously to α
and α′ , we can first glue β to α′ , then glue α to β. Since the gluing targets are
disjoint, the latter order of gluing yields the same computation of Z(M ) as the first.
But each of the gluings in the latter order is equivalent to attaching a collar to a
copy of B n in the boundary of an n+1-maanifold, and thus has no effect on Z(M ).
6.3. COMPUTING THE PATH INTEGRAL 63

Therefore Z(M ) is unaffected by the introduction or elimination of canceling pairs

of 0- and 1-handles.
Next consider an (n, 1)-body Y . A(Y ; c) is finite-dimensional for all c by the
hypotheses of the theorem. [In the future, want to derive finite dimensionality
inductively from properties of Z(pt).] Since Y × I is an (n + 1, 1)-body, Z(Y × I) is
defined and we can use (6.2.1) to define an inner product on A(Y ; c). Handle slides
and cancelations for Y are mirrored by handle slides and cancelations for Y × I,
so the inner product is invariant under these changes of handle decomposition. By
hypothesis the inner product is positive for the handles (copies of B n ) with any
boundary conditions. The argument of (6.2.5) now shows that the inner product on
A(Y ; c) is positive definite.
Next consider an (n − 1, 1)-body R. Since R × I is an (n, 1)-body, the morphism
spaces of A(R; b) have positive definite inner products. We have hab, ci = Z(R ×
I)(abb ∪ c) = Z(R × I)(bb ∪ a∗ c) = hb, a∗ ci (because ab
b ∪ c = bb ∪ a∗ c), so these inner
products are compatible the action of A(R; b) on itself. It follows from (B.5.3) that
A(R; b) is semisimple. A(R; b) has finitely many irreps since otherwise A(R × S 1 ; b ×
S 1 ) would be infinite dimensional.
We have now established the inductive hypotheses for i = 1. Next we establish
them for i = 2, which will present one additional complication.
Let M be an (n + 1, 2)-body. Choose an ordering of the 1- and 2-handles such
that if the attaching map of a 2-handle α goes over a 1-handle β, then α follows β in
the ordering. Use this ordering to define Z(M ) ∈ Z(∂M ) by attaching each handle
in turn and applying (6.1.7). For 2-handles, this requires the use of the recently
defined inner product on A(B n−1 × S 1 ; c). (B n−1 × S 1 has a standard (n, 1)-body
structure with one 0-handle and one 1-handle.) Because of disjointness properties of
the attaching targets and the restrictions on the ordering, the calculation of Z(M )
does not depend on the choice of (restricted) ordering. As before, Z(M ) is not
affected by handle slides, since for an appropriately chosen ordering the handle slide
is merely an isotopy of the attaching target. Invariance under cancelation of 0- and
1-handles is proved as before.
Cancelation of 1- and 2-handles presents a new wrinkle because the attaching
targets of a canceling 1-2 pair intersect. Let α be a 1-handle which is canceled by
a 2-handle β. Assume that α immediately precedes β in the ordering, and let γ
denote the n+1-manifold resulting from all gluings which precede α in the ordering.
Attaching α and then β to γ results in an n+1-manifold which is homeomorphic
to γ. We must show that doing these two gluings in this order has no effect on
Z(γ) ∈ Z(∂γ). Define

def
Yαγ = α∩γ ∼
= Bn × S0
def
Yαβ = α∩β ∼
= B n−1 × B 1
def
Yβγ = β∩γ ∼
= B n−1 × B 1 .
64 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

Yαβ

Yαγ

Yβγ
∂γ

6.3.2 Canceling 1- and 2-handles

(See Figure (6.3.2).) Note that if we attach β to γ along Yβγ this is equivalent to
attaching a boundary collar to Yβγ ⊂ ∂γ and so has no effect on Z(γ). [need ref
for this] Similarly, attaching α to γ ∪ β along Yαγ ∪ Yαβ is equivalent to attaching a
boundary collar to Yαγ ∪ Yαβ ⊂ ∂(γ ∪ β), and so again does not affect the calculation
of Z(γ). (Yαγ ∪ Yαβ here has an obvious (n, 1)-body structure, and we use this
structure to define the inner product.) So it suffices to show that attaching α then
β yields the same result for Z and attaching first β then α. This follows from (6.2.6).
The proof that Z(M ) is independent of slides and cancelations is now complete.
Let Y be an (n, 2)-body. As above, since Y × I is an (n + 1, 2)-body, we can
use (6.2.1) to define an inner product on A(Y ; c). Handle slides and cancelations
for Y are mirrored by handle slides and cancelations for Y × I, so the inner product
is invariant under these changes of handle decomposition. By hypothesis the inner
product is positive for the handles (copies of B n ) with any boundary conditions.
The argument of (6.2.5) now shows that the inner product on A(Y ; c) is positive
definite.
Let R be an (n−1, 2)-body. Since R×I is an (n, 2)-body, the morphism spaces of
A(R; b) have positive definite inner products. The same argument as before shows
that these inner products are compatible with the action of A(R; b) on itself. It
follows from (B.5.3) that A(R; b) is semisimple. Again, A(R; b) has finitely many
irreps since otherwise A(R × S 1 ; b × S 1 ) would be infinite dimensional.
We have now established the inductive hypotheses for i = 2. The proof for
arbitrary i is very similar. The only part worth commenting on is the cancelation
of i- and i−1-handles.
Let α be an i−1-handle which is canceled by an i-handle β. Assume that α
immediately precedes β in the ordering, and let γ denote the n+1-manifold resulting
from all gluings which precede α in the ordering. Attaching α and then β to γ results
6.3. COMPUTING THE PATH INTEGRAL 65

in an n+1-manifold which is homeomorphic to γ. We must show that doing these

two gluings in this order has no effect on Z(γ) ∈ Z(∂γ). Define
def
Yαγ = α∩γ ∼
= B n−i+2 × S i−2
def
Yαβ = α∩β ∼
= B n−i+1 × B i−1
def
Yβγ = β ∩γ ∼
= B n−i+1 × B i−1 .
Note that if we attach β to γ along Yβγ this is equivalent to attaching a boundary
collar to Yβγ ⊂ ∂γ and so has no effect on Z(γ). [need ref for this] Similarly,
attaching α to γ ∪ β along Yαγ ∪ Yαβ is equivalent to attaching a boundary collar
to Yαγ ∪ Yαβ ⊂ ∂(γ ∪ β), and so again does not affect the calculation of Z(γ).
(Yαγ ∪ Yαβ here has a (n, i − 1)-body structure, with Yαβ playing the role of an
i−1-handle, and we use this structure to define the inner product.) So it suffices to
show that attaching α then β yields the same result for Z and attaching first β then
α. This follows from (6.2.6).
At this point we have defined the path integral Z(M ) for arbitrary handle-body
structures on the n+1-manifold M and shown that it is invariant under handle slides
and cancelations of any index. Thus Z(M ) is independent of the choice of handle
decomposition and depends only on the underlying manifold. Similarly, we have
well-defined positive definite inner products on A(Y ; c) for any n-manifold Y , and
A(R; b) is semisimple for all n−1-manifolds R.
All that remains to be done is to show that the gluing relation (6.1.4) holds.
[Sketch: Start with M , then construct Mgl by adding handles to ±Y , mirroring
a handle decomposition of Y . The effect of these handle additions is the same as
taking inner products in A(Y ; c).]
The proof of (6.3.1) is also a recipe for producing a “state model” for the theory.
If we assemble all the summations for all of the handle attachments into one big sum-
mation, we end up with a sum over labelings of the handles (the labels indexing the
orthogonal bases for various handle attaching targets). If the handle decomposition
is dual to a triangulation, then of course this is also a sum over labelings of the sim-
plexes of the triangulation. Specific examples of this are given in later chapters. For
1+1-dimensional theories, we get the familiar [need name]. For 2+1-dimensional
theories based on a spherical category, we get the Turaev-Viro model. For 3+1-
dimensional theories based on a ribbon category (which, via decategorification, lead
to 2+1 dimensional theories for “extended” manifolds), we get three different mod-
els depending on the type of handle decomposition. For handle decompositions with
a single 0-handle and several 2-handles, we get the the Witten-Reshtikhin-Turaev
surgery formula. For a handle decomposition dual to a triangulation, we get the
Crane-Yetter model. For handle decompositions in the which the 2-skeleton has
tetrahedral singularities, and which allow the the 2-handles to be disk bundles over
higher genus surfaces, we get the Turaev shadow state sums. [need section refs for
all of the above]
66 CHAPTER 6. FROM LOCAL RELATIONS TO PATH INTEGRALS

Still to do:

• need reference for handle stuff

• (is semi-simple case the only one that works? probably.)

• do general state model? only 4d? will need to make assumptions about general
n-cat for general case. maybe just do 4-dim’l case as generally as possible and
say that it’s clear this could be generalized
Chapter 7

1+1-dimensional Examples

7.1 Generalities on 1+1-dimensional Theories

To do:

• Start with 1-category, derive Frobenius algebra etc.

7.2 Finite Group Theories in 1 + 1 dimensions

To do:

• go over finite groups theories in detail, for 1+1-dimensional case

67
68 CHAPTER 7. 1+1-DIMENSIONAL EXAMPLES
Chapter 8

2+1-dimensional Examples

8.1 Temperley-Lieb Theories

To do:

• determine all ideals

• idempotents

• quotients

• 6j?

• do 3-d part here, or wait until next section?

8.2 Spherical Categories in General

[put this section before Temperley-Lieb section?]

Outline (to be filled in later):

• A (disk-like) 2-category with only one object is what’s known as a spherical

tensor category. (Need to give details of equivalence and full definition of
spherical category.)

• Given a spherical category C, with objects C 0 and morphisms C 1 , we can

define fields on surfaces consisting of embedded graphs with oriented edges
labeled by C0 and vertices (“coupons” in R-T terminology) labeled by C 1 .
(Graphs are allowed to have circular edges without vertices.) Frobenius reci-
procity means that we don’t need to distinguish range and domain at the
coupons. Fields on a 1-manifold are oriented points (ends of oriented arcs),
labeled by C 0 . There is a unique (empty) field on a 0-manifold.

69
 70 CHAPTER 8. 2+1-DIMENSIONAL EXAMPLES

• The local relations are generated by

1. isotopy,
2. reversing the orientation of an arc and changing the label to its dual,
3. replacing “identity” coupons with parallel arcs (note that this includes
cups and caps because of Frobenius reciprocity),
4. erasing arcs labeled by the trivial object (and adjusting labels of coupons
if necessary),
5. combining two adjacent coupons into one (using composition of mor-
phisms in C), and
6. replacing a diagram containing a coupon labeled by a linear combination
of morphisms with the corresponding linear combination of fields.

• For this theory, A(pt) is essentially C (thought of as a 2-category, of course).

(Need to give more details on “essentially”.)

• A(S 1 ) is the annularization of C.

• Z(S 1 ) is Drinfeld center of the category of representations of C. (See (5.2.30).)

• Note that any labeled graph in S 2 is equivalent, via the above local relations,
to some multiple of the empty graph. We call this multiple the “standard
evaluation” of a graph in S 2 .

8.2.1 • A(Y 2 ) can be described in terms of labelings of 0- and 1-skeleton of a fixed

cell decomposition of the surface Y . An explicit list of relations corresponding
to 2-cells of the cell decomposition can be given. (Give details.) Note that we
already know that the vector space is independent of the cell decomposition,
so we don’t need a separate proof of independence.

• If C is semisimple, then we can restrict all coupons to be trivalent. We can

also restrict all edge labels to be simple objects (or minimal idempotents or
irreps; need to comment on equivalence between minimal idems, irreps and
simple objects). We get the familiar description in terms of labeled, oriented
trivalent graphs. (Plain “trivalent graphs” for short.)

• (Need to give details on “F” moves (a.k.a. recoupling), etc.)

8.2.2 • (Give refinement of description of A(Y ; c), including case when Y is a disk.
Observe that in this case we have an orthogonal basis.)

Assume now that C is s semisimple spherical category with finitely many irreps.
It follows from (8.2.1) that A(Y ; c) is finite dimensional for all 2-manifolds Y and
c ∈ C(∂Y ). Note that Z(S 2 ) is 1-dimensional, and the standard evaluation of triva-
lent graphs (which evaluates to 1 ∈ C on the empty graph) is a basis. It follows
8.2. SPHERICAL CATEGORIES IN GENERAL 71

from (8.2.2) that any non-zero element z ∈ Z(S 2 ) determines a nondegenerate in-
ner product on A(D 2 ; c) for all c. Assume that these inner products are positive
definite. [comment on this assumption.] We can now apply (6.3.1) to construct
a path integral for 3-manifolds. Following the proof of (6.3.1) we will construct a
state sum description of the path integral in terms of labelings of the cells of a cell
decomposition of a 3-manifold. We will see that this state model turns out to be
the Turaev-Viro state model [TV92, Tur94].
Choose z ∈ Z(S 2 ); z is λ times the standard evaluation of trivalent graphs for
some λ ∈ C. [for positive definiteness, need λ ∈ R+ ] (Recall from (6.3.1) that we
define Z(B 3 ) = z.) Our first task is the compute bases and inner products for
A(Y ; c), where (Y ; c) runs through all attaching targets for handles that we will
need below.
Recall that a field on 1-manifold R consists of a finite number of oriented points
in R each labeled by an irrep. If R is contained in the boundary of a 2-manifold
we will assume that all of the points are oriented inward. Fields will be denoted
as a sequence of irreps, so, for example, A(D 2 ; a, b, c) means a disk with boundary
conditions given by three inward pointing points labeled by a, b and c. [this remark
on notation should go earlier]
Let a and b be irreps (or equivalently minimal idempotents or simple objects) 8.2.3
of C. Then A(D 2 ; ba, b) is 1-dimensional if a and b are equivalent and 0-dimensional
otherwise. A basis for A(D 2 , ba, a) is given by a single oriented arc in D 2 connecting
the boundary points and labeled by a. Call this basis element ebaa . Then

hebaa , ebaa i = z(b

ebaa ∪ ebaa ) = λda ,

where da is the value of a loop labeled by a in the standard evaluation. [need to

introduce da above]
Next consider A(D 2 ; a, b, c) ∼ a, bc) ∼
= hom(b a, c) ∼
= hom(bbb = hom(1, abc) etc. If 8.2.4
2
x, y ∈ A(D , a, b, c) then hx, yi is given by evaluating a “theta” graph with arcs
labeled by a, b, c and vertices labeled by xb and y, then multiplying by λ. By as-
sumption this inner product is nondegenerate, so we can choose an orthogonal basis
{eabci } ⊂ A(D 2 , a, b, c). We have

heabci , eabcj i = z(b

eabci ∪ eabcj ) = δij λθabci ,

where θabci is the value of the standard evaluation on a theta graph labeled by a, b,
c, eabci and ebabci .
Next we consider S 1 × I. We will see below that we will only need to know inner 8.2.5
products on A(S 1 × I; c) when c is the empty boundary condition ∅. For a an irrep
of C, consider the element ea ∈ A(S 1 ×I; ∅) represented by a loop S 1 ×{pt} ⊂ S 1 ×I
labeled by a. It follows from (4.4.6) that {ea }, where a runs through a complete set
of irreps of C, is a basis of A(S 1 × I; ∅). Recalling the definition of ebaa from above,
 72 CHAPTER 8. 2+1-DIMENSIONAL EXAMPLES

[No figure yet]

8.2.6 Cutting a solid torus to get a ball

and applying (6.1.9), we see that

hea , ea i = Z(S 1 × I × I; eba , ea )

= z(ebaa ∪ ebaa ∪ ebaa ∪ ebaa )hebaa , ebaa i−1
= λda (λda )−1
= 1

(see Figure (8.2.6)). A similar argument shows that hea , eb i = 0 if a and b are not
equivalent. So {ea } is an orthonormal basis of A(S 1 × I; ∅).
8.2.7 Finally we consider S 2 . Let e∅ ∈ A(S 2 ) be represented by the empty trivalent
graph; e∅ spans A(S 2 ). Let L be a complete set of irreps of C. Decomposing
S 2 × I = (D 2 × I) ∪ (D 2 × I) and applying (6.1.9) and (8.2.5) we have

he∅ , e∅ i = Z(S 2 × I)(b

e∅ ∪ e∅ )
X
= z(bea ∪ ∅D2 ∪ ∅D2 )z(ea ∪ ∅D2 ∪ ∅D2 )hea , ea i−1
a∈L
X
= (λda )(λda ) · 1
a∈L
2
= λ D,

def P 2
where D = a da .

Armed with the above computations, we can now calculate the path integral
Z(M ) of a 3-manifold M in terms of a generic cell decomposition of M . By generic
we mean dual to a triangulation, so that each 1-cell is incident to three 2-cells, and
the 1- and 2-cells incident to a 0-cell form a tetrahedral pattern. By thickening the
cells we get a handle decomposition of M . For simplicity, we will at first assume
the M is closed. Later we will indicate how to extend the results to more general
handles decompositions and to 3-manifolds with boundary.
We will start by analyzing the effect of adding 3-handles and work our way down
to the 0-handles. Let Mi ⊂ M be the union of all handles of index less than or equal
to i. Let ni be the number of i-handles, and let Hi denote the set of i-handles.
Applying (6.1.9) and (8.2.7) to the decomposition M = M2 ∪ {3-handles}, we see
8.2. SPHERICAL CATEGORIES IN GENERAL 73

that

Y
Z(M ) = Z(M2 )(en∅ 3 ) z(e∅ )he∅ , e∅ i−1
H3
Y
= Z(M2 )(en∅ 3 ) λ−1 D −1 .
H3

Of course, the product is equal to (λD)−n3 , but we prefer to write it as a product

in order to be consistent with the treatment of the other handles below.
Let L2 denote the set of all labelings of the 2-handles by elements of L. Applying
(6.1.9) and (8.2.5) to the decomposition M2 = M1 ∪ {2-handles}, we see that

X Y
Z(M2 )(en∅ 3 ) = Z(M1 )(gα ) λd(α, f ),
α∈L2 f ∈H2

where gα denotes is the graph on ∂M1 consisting of a loop for the attaching target of
each 2-handle labeled according to α, and d(α, f ) = dα(f ) , the standard evaluation
of a loop labeled by α(f ).
Next we consider the decomposition M1 = M0 ∪ {1-handles}. We want to eval-
uate Z(M1 ) on gα . In applying (6.1.9) and (8.2.4) we will place a trivalent vertex
(some eabci ) in each D 2 in ∂(M0 ⊔ {1-handles}) which results from cutting M1 . [need
figure(?)] The resulting field on the boundary of a 1-handle is a theta graph labeled
according to α and the eabci ’s on the two attaching disks. If these two basis elements
are not the same then the theta graph will evaluate to zero, so we may assume that
they are the same. In other words, given α we have fixed arc labels a, b, c for the
theta graph of each 1-handle, and we sum over i’s so that eabci runs through a ba-
sis of A(D 2 , a, b, c). Let L1,α denote the set of all such labelings of the 1-handles
consistent with α. For β ∈ L1,α and e ∈ H1 let Θ(α, β, e) denote the standard eval-
uation of the labeled theta graph on the boundary of e. For each 0-handle v ∈ H0
there is a labeled tetrahedral graph on the boundary of the 0-handle. (Labels of the
edges of the tetrahedron come from α and labels of the vertices of the tetrahedron
from from β.) Let Tet(α, β, v) denote the standard evaluation of this graph. Note
that the contribution of a 1-handle e to the he∗ , e∗ i−1 factor of (6.1.9) is, by (8.2.4),
(λΘ(α, β, e))−2 , since we glue along two pairs of disks for each 1-handle. Putting
this all together, we have

X Y Y
Z(M1 )(gα ) = λΘ(α, β, e)(λΘ(α, β, e))−2 λ Tet(α, β, v)
β∈L1,α e∈H1 v∈H0
X Y Y
= (λΘ(α, β, e))−1 λ Tet(α, β, v)
β∈L1,α e∈H1 v∈H0
74 CHAPTER 8. 2+1-DIMENSIONAL EXAMPLES

Combining all of the above, we have

X X Y Y Y Y
Z(M ) = λ−1 D−1 λd(α, f ) λ−1 Θ(α, β, e)−1 λ Tet(α, β, v)
α∈L2 β∈L1,α H3 f ∈H2 e∈H1 v∈H0
X X Y Y Y Y
= λχ(M ) D −1 d(α, f ) Θ(α, β, e)−1 Tet(α, β, v).
α∈L2 β∈L1,α H3 f ∈H2 e∈H1 v∈H0

Here χ(M ) denotes the Euler characteristic of M , which is always zero for a closed 3-
manifold. We leave it in the formula because we want to indicate how to generalize
to non-closed 3-manifolds, and because we want to emphasize the parallels with
a similar expression for 4-manifolds. [need forward ref] The above expression for
Z(M ) is essentially the Turaev-Viro state sum [TV92, Tur94]. But note that we do
not need to show that it is invariant under Pachner moves on the triangulation —
invariance follows from the easier and more general (6.3.1).
For a general cell decomposition of M (one not necessarily dual to a triangu-
lation), The factors of Θ and above will be replaced with an evaluation of “multi-
barred theta graph” reflecting the number of 2-cells incident to a 1-cell. [need figure]
The factors of Tet will be replaced with an evaluation of the graph which is the link
of the 2-skeleton of the decomposition around the vertex. If M has boundary then
we first fix a graph g on ∂M , which we assume is in a neighborhood of the 1-skeleton
of the cell decomposition restricted to the boundary. Then there is a similar state
sum for Z(M )(g). [need to give more details]

8.3 A2
(follow Kuperberg then extend some; include lots of detail)

8.4 G2
(follow Kuperberg then extend some; include lots of detail)

8.5 Finite Group Theories in 2+1 dimensions

8.6 Jones Planar Algebras

To do:
• Planar algebra is a disk-like 2-cat with two objects and singly generated 1-
mors.

• Comment more generally on relation to subfactor point of view to this one?

Provide translation table? Put this in another section/chapter?
Chapter 9

3+1-dimensional Examples

9.1 Theories From Ribbon Categories

Outline (to be filled in later):

• Note that this section is very similar to Section 8.2, so the reader might want
to compare the two, or treat Section 8.2 as a warm-up for this section.

• A disklike 3-category with only one object (0-morphism) and only one 1-
morphism is a ribbon category. [need to give details] Thus we expect that
a from a ribbon category we can construct a 3+1-dimensional TQFT. We
will see in the next section that (with some additional assumptions on the
ribbon category) this 3+1-dimensional TQFT can be decategorified to yield
the Witten-Reshtikhin-Turaev type 2+1-dimensional TQFT that can be con-
structed directly (but less cleanly) from the same ribbon category.

• Given a ribbon category C, with objects C 0 and morphisms C 1 , we can define

fields on 3-manifolds consisting of embedded ribbon graphs with oriented edges
labeled by C0 and vertices labeled by C 1 . Following Reshetikhin and Turaev,
we will call a labeled vertex a coupon. [need to comment on pinning or cyclic
ordering or ciliation at vertices] (Graphs are allowed to have circular edges
without vertices.) The dualities of the ribbon category mean that we don’t
need to distinguish range and domain at the coupons. Fields on a 2-manifold
are oriented framed points (ends of oriented ribbons), labeled by C 0 . There is
a unique (empty) field on a 1-manifold. There is a unique (empty) field on a
0-manifold.

• The local relations are generated by

1. isotopy,
2. reversing the orientation of a ribbon and changing the label to its dual,

75
 76 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES

3. replacing “identity” coupons with parallel ribbons (note that this includes
cups and caps because of Frobenius reciprocity),
4. erasing ribbons labeled by the trivial object (and adjusting adjacent labels
of coupons as necessary),
5. combining two adjacent coupons into one (using composition of mor-
phisms in C), and
6. replacing a diagram containing a coupon labeled by a linear combination
of morphisms with the corresponding linear combination of fields.

• For this theory, A(pt) is essentially C (thought of as a 3-category, of course).

[Need to give more details on “essentially”.]

• Note that any labeled ribbon graph in (B 3 , ∅) is equivalent, via the above
local relations, to some multiple of the empty graph. We call this multiple the
“standard evaluation” of a graph in (B 3 , ∅).

9.1.1 • A(M 3 ; c) can be described in terms of labelings of 0- and 1-skeleton of a

fixed cell decomposition of the 3-manifold M . An explicit list of relations
corresponding to 2-cells of the cell decomposition can be given. [Give details.]

• [Need to remark that A(M 3 ; c) is the familiar (relative) skein module of M

based on C.]

• If C is semisimple, then we can restrict all coupons to be trivalent. We can

also restrict all edge labels to be simple objects [or minimal idempotents or
irreps; need to comment on equivalence between minimal idems, irreps and
simple objects]. We get the familiar description in terms of labeled, oriented
trivalent graphs. (Plain “trivalent graphs” for short.)

• (Need to give details on “F” moves (a.k.a. recoupling), etc.)

• Let L be a complete set of irreps of C.

9.1.2 • (Give refinement of description of A(M ; c), including case when M is a disk.
Observe that in this case we have an orthogonal basis.)

Assume now that C is s semisimple ribbon category with finitely many irreps.
It follows from (9.1.1) that A(M ; c) is finite dimensional for all 3-manifolds M
and c ∈ C(∂M ). Note that Z(S 3 ) is 1-dimensional, and the standard evaluation
of trivalent ribbon graphs (which evaluates to 1 ∈ C on the empty graph) is a
basis. It follows from (9.1.2) that any non-zero element z ∈ Z(S 3 ) determines a
nondegenerate inner product on A(B 3 ; c) for all c. Assume that these inner products
are positive definite. [comment on this assumption.] We can now apply (6.3.1) to
construct a path integral for 4-manifolds.
9.1. THEORIES FROM RIBBON CATEGORIES 77

Following the proof of (6.3.1) we will construct a state sum description of the
path integral in terms of labelings of the handles of a handle decomposition of a 4-
manifold. [need to be consistent about handle decomposition vs cell decomposition]
We will see below that this state sum specializes to: (a) the well-known Witten-
Reshtikhin-Turaev framed link surgery formula, for handle decompositions consist-
ing of a single 0-handle and several 2-handles; (b) the Crane-Yetter state sum, for
handle decompositions which are dual to triangulations; and (c) the Turaev shadow
state sum, for handle decompositions which have tetrahedral singularities in their
2-skeleton. [need to be more precise here] [need refs for above]
Choose z ∈ Z(S 3 ); z is λ times the standard evaluation of trivalent graphs for
some λ ∈ C. [for positive definiteness, need λ ∈ R+ ] (Recall from (6.3.1) that we
define Z(B 4 ) = z.) Our first task is the compute bases and inner products for
A(M ; c), where (M ; c) runs through all attaching targets for handles that we will
need below.
Recall that a field on 2-manifold Y consists of a finite number of oriented framed
points in Y each labeled by an irrep. If Y is contained in the boundary of a 3-
manifold we will assume that all of the points are oriented inward. Fields will be
denoted as a sequence of irreps, so, for example, A(B 3 ; a, b, c) means a 3-ball with
boundary conditions given by three inward pointing points labeled by a, b and c.
[remark that the ordering of the irreps (points) in a connected component doesn’t
matter] [this remark on notation should go earlier]
Let a and b be irreps (or equivalently minimal idempotents or simple objects) of 9.1.3
C. Then A(B 3 ; ba, b) is 1-dimensional if a and b are equivalent and 0-dimensional oth-
erwise. A basis for A(B 3 , ba, a) is given by a single oriented ribbon in B 3 connecting
the boundary points and labeled by a. Call this basis element ebaa . Then

hebaa , ebaa i = z(b

ebaa ∪ ebaa ) = λda ,

where da is the value of a 0-framed loop labeled by a in the standard evaluation.

[need to introduce da above]
Next consider A(B 3 ; a, b, c) ∼= hom(ba, bc) ∼ a, c) ∼
= hom(bbb = hom(1, abc) etc. If 9.1.4
3
x, y ∈ A(B , a, b, c) then hx, yi is given by evaluating a “theta” graph with ribbons
labeled by a, b, c and vertices labeled by x b and y, then multiplying by λ. By as-
sumption this inner product is nondegenerate, so we can choose an orthogonal basis
{eabci } ⊂ A(B 3 , a, b, c). We have

heabci , eabcj i = z(b

eabci ∪ eabcj ) = δij λθabci ,

where θabci is the value of the standard evaluation on a theta graph labeled by a, b,
c, eabci and ebabci .
More generally, if c = (a1 , . . . , an ) (n ≥ 3), we can choose a trivalent ribbon 9.1.5
tree with boundary c, and labelings of the tree give an orthogonal basis of A(B 3 ; c)
(see (9.1.2)). The n − 3 internal edges of the tree are labeled by irreps and the n − 2
 78 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES

trivalent vertices of the tree are labeled by some orthogonal bases of the appropriate
trivalent vertex spaces. Call these (multi) labels b and v respectively. Then
Y Y
hebv , ebv i = λ Θ(b, v, i) d(b, j)−1 ,
i∈V j∈E

where V denotes the set of vertices of the tree and E denotes the set of internal
edges. [need to give more detail and improve notation]
9.1.6 Next we consider S 1 × D 2 . We will see below that we will only need to know
inner products on A(S 1 × D 2 ; c) when c is the empty boundary condition ∅. For
a an irrep of C, consider the element ea ∈ A(S 1 × D 2 ; ∅) represented by a loop
S 1 × {pt} ⊂ S 1 × D 2 labeled by a. It follows from (4.4.6) that {ea }, where a runs
through a complete set of irreps of C, is a basis of A(S 1 × D 2 ; ∅). Recalling the
definition of ebaa from above, and applying (6.1.9), we see that

hea , ea i = Z(S 1 × D 2 × I; eba , ea )

= z(ebaa ∪ ebaa ∪ ebaa ∪ ebaa )hebaa , ebaa i−1
= λda (λda )−1
= 1

A similar argument shows that hea , eb i = 0 if a and b are not equivalent. So {ea } is
an orthonormal basis of A(S 1 × D 2 ; ∅).
9.1.7 Next we consider S 2 × I. In applications we will only need to consider empty
boundary conditions. A(S 2 × I; ∅) is 1-dimensional and spanned by e∅ , the empty
field. Let L be a complete set of irreps of C. Decomposing S 2 × I × I = (D 2 × I ×
I) ∪ (D 2 × I × I) and applying (6.1.9) and (9.1.6) we have

he∅ , e∅ i = Z(S 2 × I × I)(be∅ ∪ e∅ )

X
= z(bea ∪ ∅B 3 ∪ ∅B 3 )z(ea ∪ ∅B 3 ∪ ∅B 3 )hea , ea i−1
a∈L
X
= (λda )(λda ) · 1
a∈L
2
= λ D,
def P
where D = a d2a .
9.1.8 Finally we consider S 3 . A(S 3 ) is 1-dimensional and spanned by e′∅ , the empty
field. [need better notation] Decomposing S 3 × I = (B 3 × I) ∪ (B 3 × I) and applying
(6.1.9) and (9.1.7) we have

he′∅ , e′∅ i = Z(S 3 × I)(b

e′∅ ∪ e′∅ )
= z(∅)z(∅)he∅ , e∅ i−1
= λλ(λ2 D)−1
= D −1 .
9.1. THEORIES FROM RIBBON CATEGORIES 79

[No figure yet]

9.1.10 A figure which needs no caption

This completes the inner product calculations we will need for handle attachments.
Before doing general calculations we look at S 2 × I and S 2 × S 1 , and also show
that frequently we can choose λ so that Z(W 4 ) depends only on the bordism class
of W .
Consider A(S 2 ×I; b a, b), where the two ribbon ends lie in different components of
∂(S × I) and a and b are irreps (simple objects). It is easy to see that A(S 2 × I; b
2 a, b)
is 0 if a ≇ b, and that A(S 2 × I; b a, a) is spanned by xa , the obvious unknotted
ribbon connecting the two ribbon ends. So we must determine whether xa = 0 in
(S 2 × I; b
a, a).
Let Hab denote the standard evaluation of the (framed) Hopf link with labels a 9.1.9
and b. Note that H1a = da . Let xab denote xa plus a small linking circle labeled by
b (see Figure (9.1.10)). Then xab = (Hab /da )xa . But in S 2 × I the linking circle is
isotopic to an unlinked circle, so we also have xab = db xa . So (Hab /da − db )xa = 0
for all b ∈ L, which means that xa = 0 unless Hab = da db for all b ∈ L. By (9.1.1)
this is also a sufficient condition for xa 6= 0. Call such an a ∈ L degenerate. [find a
better name for this?] Thus A(S 2 × I; b a, a) is 1-dimensional when a is degenerate
and 0-dimensional otherwise. Note that 1 ∈ L is always degenerate.
Similar arguments show that a basis of A(S 2 × S 1 ) consists of {fa }, where a is
degenerate and fa ∈ A(S 2 × S 1 ) denotes the ribbon graph pt × S 1 ⊂ S 2 × S 1 labeled
by a ∈ L.
Suppose that dim A(S 2 ×S 1 ) = 1 (or equivalently 1 is the only degenerate irrep).
[in what follows, need to make clearer when this assumption is needed.] We will
show that in this case setting λ2 = 1/D yields a path integral Z(W 4 ) which depends
only on the bordism class of W . 4-dimensional bordism is generated by three types
of surgery (and their inverses):
∅ ↔ S4
B4 × S0 ↔ S3 × B1
B 3 × S 1 ↔ S 2 × B 2.
We have
Z(S 4 ) = Z(B 4 )(e′∅ ) · Z(B 4 )(e′∅ ) · he′∅ , e′∅ i−1 = λ2 D,
so the first type of surgery does not affect Z(W ) if λ2 = 1/D. We have
Z(B 4 × S 0 )(∅) = λ2 ,
 80 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES

while
Z(S 3 × B 1 ) = he′∅ , e′∅ i = D −1 ,
so the second type of surgery does not affect Z(W ) if λ2 = 1/D.
Finally, we consider B 3 × S 1 ↔ S 2 × B 2 , with boundary S 2 × S 1 . With a
eye toward future applications, we will temporarily abandon our assumption that
dim A(S 2 × S 1 ) = 1. Let fa ∈ A(S 2 × S 1 ) be as defined above. Note that f1 is
equivalent to the empty field. We have

Z(B 3 × S 1 )(fa ) = dim A(B 3 ; a) = δ1a .

On the other hand, using the gluing formula for S 2 × B 2 = (B 2 × B 2 ) ∪ (B 2 × B 2 ),

we have
X
Z(S 2 × B 2 )(fa ) = λHab · λdb
b∈L
X
= λ2 (da db ) · db
b∈L
2
= da λ D.

(By (9.1.9) if fa 6= 0 then Hab = da db .) So Z(W ) is a bordism invariant if and only

if dim A(S 2 × S 1 ) = 1 and λ2 = 1/D.
[need to say more about the not-bordism-invariant case.]
Consider a 4-manifold W consisting of a single 0-handle and some 2-handles.
Identify S 3 with the boundary of the 0-handle. The attaching curves of the 2-
handles form a framed link L in S 3 . Let L(L) denote the set of all labelings of
L by irreps of L. For m ∈ L(L) let J(L, m) denote the standard evaluation of L
labeled by m. [remark that this is an evaluation of the generalized Jones polynomial
of (L, m)] Let d(m) denote the product dm1 · · · dmk , where m = (m1 , . . . , mk ). It
follows from the gluing formula and (9.1.6) that
X
Z(W )(∅) = (λJ(L, m))(λk d(m)).
m∈L(L)

[remark that this is the well known Witten-Resheitkin-Turaev surgery formula] Note
that we do not need to show that this expression is invariant under handle slides; this
follows from (6.3.1). More generally, let K be a labeled ribbon graph (e.g. a framed
link) in ∂W . By general position, we may assume that K lies in the complement of
L in S 3 . Let J(L ∪ K, m) denote the standard evaluation of L ∪ K, with L labeled
by m. Then X
Z(W )(K) = (λJ(L ∪ K, m))(λk d(m)).
m∈L(L)

9.1.11 Next consider a 4-manifold W equipped with a generic cell decomposition (i.e. a
cell decomposition dual to a triangulation). This means that a neighborhood of each
9.1. THEORIES FROM RIBBON CATEGORIES 81

0-cell looks like an open cone on the boundary of a 4-simplex, and a neighborhood
of a point in a k-cell looks like a neighborhood of a point of a k-cell in this cone.
(In other words, the open cone on the boundary of a 4-simplex is the local model
for the cell decomposition.) We will use the gluing formula to derive a state sum
description of Z(W ) in terms of labelings of the cells. [refer to 2+1 dim’l case? need
to say that this is very similar to 2+1/TV case, and so we’ll use fewer words]
For simplicity, assume that W is closed.
[say something about the fact that there is a closely related handle decomposition
and we will treat the two as equivalent]
Let ni be the number of i-handles, Hi be the set of i-handles, and let Wi denote
the union of all handles of index less than or equal to i.
Applying the gluing formula (6.1.9) and the inner product calculation (9.1.8) to
the decomposition W = W3 ∪ {4-handles}, we have
Y
Z(W ) = Z(W3 )(∅) z(e′∅ )he′∅ , e′∅ i−1
H4
= Z(W3 )(∅)λn4 D n4

Applying the gluing formula and the inner product calculation (9.1.7) to the decom-
position W3 = W2 ∪ {3-handles}, we have
Y
Z(W3 )(∅) = Z(W2 )(∅) z(∅)he∅ , e∅ i−1
H3
−n3
= Z(W3 )(∅)λ D −n3

Let L2 denote the set of labelings of the 2-handles by elements of L. Applying

the gluing formula and the inner product calculation (9.1.6) to the decomposition
W2 = W1 ∪ {2-handles}, we have
X Y
Z(W2 )(∅) = Z(W1 )(Lα ) λd(α, f ),
α∈L2 f ∈H2

where Lα denotes the ribbon link in ∂W1 consisting of an (appropriately framed)

ribbon in the core of the attaching target of each 2-handle, labeled according to α,
and d(α, f ) = dα(f ) , the standard evaluation of an 0-framed unknot labeled by α(f ).
Next we consider the decomposition W1 = W0 ∪ {1-handles}. We want to evalu-
ate Z(W1 ) on Lα . The attaching target of a 1-handle is a pair of 3-balls. The field
Lα restricted to the boundary of one of these 3-balls consists of four labeled framed
points (corresponding to the vertices of a tetrahedral graph in ∂B 3 ). The two ends
of the 1-handle have the same four labels. For each 1-handle e choose an orthogonal
basis of A(B 3 ; c(α, e)), where c(α, e) denotes the four labeled points. Let L1 denote
the set of all labelings of the 1-handles by these basis vectors. We are now ready
to apply (6.1.9) to the gluing of the 1-handles to the 0-handles. The summation
will be over β ∈ L1 . Let E(α, β, e) = hf, f i, where f is the basis vector assigned
 82 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES

by β ∈ L2 to the 1-handle e, and the inner product is computed using the standard
evaluation (λ = 1). For each 1-handle e we have a factor of λE(α, β, e) coming
from the 1-handle and (λE(α, β, e))−2 coming from the gluing correction factor, so
the total contribution of a 1-handle is (λE(α, β, e))−1 . Each 0-handle v contributes
λΦ(α, β, v), where Φ(α, β, v) denotes the standard evaluation of the 1-skeleton of
the boundary of a 4-simplex. The labels of the edges of this graph come from α,
and the labels of the (4-valent) vertices come from β. Combining all of the above
and applying the gluing formula we have
X Y Y
Z(W1 )(Lα ) = λ−1 E(α, β, e)−1 λΦ(α, β, v)
β∈L2 e∈H1 v∈H0

If we choose for each 1-handle e a trivalent ribbon tree with boundary c(α, e) (equiv-
alently, choose a partition of the four labeled points of c(α, e) into two groups of two),
then labelings of the internal edge and two vertices of this tree give an orthogonal
basis of A(B 3 ; c(α, e)) (see (9.1.5)). The above expression becomes
X Y Y
Z(W1 )(Lα ) = λ−1 Θ1 (α, β, e)−1 Θ2 (α, β, e)−1 d(β, e) λΦ(α, β, v).
β∈L2 e∈H1 v∈H0

Here Θ1 and Θ2 are the θ factors for the two trivalent vertices corresponding to e,
with labels coming from α and β. d(β, e) is the loop value for the internal edge of
the tree for e. Φ(α, β, v) can now be interpreted as a “15j” symbol: it is the standard
evaluation of a trivalent ribbon graph with 15 edges and 10 trivalent vertices.
Combining the expressions for all handles we have
X X
Z(W ) =λχ(W ) Dn4 −n3
α∈L2 β∈L2
Y Y Y
d(α, f ) Θ1 (α, β, e)−1 Θ2 (α, β, e)−1 d(β, e) Φ(α, β, v).
f ∈H2 e∈H1 v∈H0

(Here χ(W ) = n4 − n3 + n2 − n1 + n0 is the Euler characteristic of W .) If λ = 1 this

coincides with Crane-Yetter state sum [CY93], after taking into account different
conventions for normalizing trivalent vertices. [more specifically, the Crane-Yetter
15j symbol contributes a factor of d(β, e)2 for each 1-handle e]
Note that we do not need to show that this expression is independent of the
choice of cell decomposition; this follow from (6.3.1).
[remark: the above can be adapted to 4-manifolds with boundary]
9.1.12 Next we consider a different class of cell decompositions of a 4-manifold and show
that it leads to the Turaev shadow state sum [Tur94]. Consider a cell decomposition
of W 4 where (a) the 2-skeleton near a 0-cell looks like a cone on the 1-skeleton of
a 3-simplex, (b) each 1-cell is incident to three 2-cells, and (c) the “2-cells” are
allowed to be arbitrary oriented surfaces with a single boundary component. (Such
9.1. THEORIES FROM RIBBON CATEGORIES 83

[No figure yet]

9.1.13 Cutting a fat graph

a decomposition is called a shadow of W . Turaev allows slightly more generality

(e.g. nonoriented 2-cells), but we will be content with the above.)
Before applying the gluing formula we need to compute the path integral of
a generalized 2-cell. Let Y be an oriented surface with ∂Y ∼ = S 1 . We want to
compute Z(Y × D 2 ; ea , ∅), where ea resides on (∂Y ) × D 2 (see (9.1.6)) and ∅ resides
on Y × ∂D 2 . Y can be realized as thickening of a 1-complex G; see Figure (9.1.13).
Making a cut in the middle of each 1-cell of G and applying the gluing formula we
see that
Y Y
Z(Y × D 2 ; ea , ∅) = λda (λda )−1
v e
χ(Y )
= (λda ) ,

where v runs over 0-cells of G and e runs over 1-cells of G. If the disk bundle over
Y is twisted with respect to the boundary trivialization determined by embedding
of ea , or equivalently the embedding of ea is changed by a twist, this becomes

Z(Y × D 2 ; ea , ∅) = tka (λda )χ(Y ) ,

where ta is the twist factor for a and k is the number of twists (Euler number of the
disk bundle). [need to say this better; also refer back to def of ta ]
Now we proceed as before. Gluing 3- and 4-handles presents nothing new:

Z(W ) = (λD)n4 −n3 Z(W2 )(∅).

Gluing 2-handles to W1 gives

X Y
Z(W2 )(∅) = Z(W1 )(Lα ) t(α, f )kf (λd(α, f ))χ(f ) ,
α∈L2 f ∈H2

where t(α, f ) is the twist factor for the label that α assigns to f , and kf is the euler
number for f . [comment on Turaev “gleams”] Gluing 1-handles to W0 gives
X Y Y
Z(W1 )(Lα ) = λ−1 Θ(α, β, e)−1 λ Tet(α, β, v).
β∈L2 e∈H1 v∈H0

Here L2 is the set of all labelings of the 1-handles by orthogonal basis vectors
associated to the three labels coming from α and the three 2-cells adjacent to the
84 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES

1-handle. Θ(α, β, e) is the standard evaluation of the theta graph corresponding to

e, with edge labels coming from α and vertex labels coming from β. Tet(α, β, v) is
the standard evaluation of the tetrahedron associated to v with edge labels coming
from α and vertex labels coming from β. Combining all this we have
X X
Z(W ) =λχ(W ) Dn4 −n3
α∈L2 β∈L2
Y Y Y
t(α, f )kf d(α, f )χ(f ) Θ(α, β, e)−1 Tet(α, β, v).
f ∈H2 e∈H1 v∈H0

If we set λ2 = 1/D this is essentially the Turaev shadow state sum for the cell
decomposition of W . [need to check normalization] Note that we do not need to
show that this expression is independent of the choice of cell decomposition; this
follow from (6.3.1).
[need to remark that this can be generalized to manifolds with boundary]
[...]

9.2 Decategorification
Plan:
• choose λ to make Z a bordism invariant

• give details for case where S is singular

9.3 More on Chern-Simons Theories

Plan:
• use multiplication str to take square root

• cf K(G, 1) is H-space iff G is cyclic (U(1) theories)

• comment on lack of higher codim stuff [move this elsewhere]

9.4 Contact Structures

Plan:
• Tight contact structure is a local relation, so we can use TQFT machinery to
study the set of tight contact structures on a 3-manifold

• Need to go into details of gluing smooth 3-men w/ corners

• Verify field axioms

9.4. CONTACT STRUCTURES 85

• gluing

• note relation to Honda, Etnyre, etc.

• Eliashberg criterion for 3-ball

• algorithm for handlebody

• A(S 1 × I)
86 CHAPTER 9. 3+1-DIMENSIONAL EXAMPLES
Chapter 10

4+1-dimensional Examples

10.1 Theories From Khovanov Homology

Outline (to be filled in later):

• The main fact about Khovanov homology that we will use is that surface
bordisms in S 3 × I give maps on Khovanov homology. (See [MW06] and [need
ref for Jacobsson].)

• Recall from [MW06] the notion of disoriented 1-manifolds and 2-manifolds...

• We can give Khovanov homology the structure of a 4-category as follows.

– The unique 0-morphism is an undecorated point.

– The unique 1-morphism is an undecorated interval.
– A 2-morphism is a collection of framed, oriented points in the interior
D2 .
– A 3-morphism is a framed, disoriented tangle in B 3 . (The domain and
range are the restriction to the southern and northern hemispheres.)
– Given two 3-morphisms with the same boundary, we can glue them to-
gether and get a disoriented link in S 3 . A 4-morphism is an element of
the Khovanov homology of such a link.
– Composition of i-morphsims (i ≤ 3) is given by gluing.
– Composition of 4-morphisms is given as follows. Let A, B and C be
tangles with common boundary. Let AB, BC and AC denote the links
formed by gluing the tangles. There is an obvious bordism W from
AB⊔BC to AC. (The underlying 4-manifold of W is B 4 with two smaller
balls removed from its interior. The surface in W consists of A × I, B × I
and C × I with some identifications near the common boundary of A,
B and C.) Choose a point p in the 3-ball containing B, disjoint from

87
88 CHAPTER 10. 4+1-DIMENSIONAL EXAMPLES

the tangle. Let α be the corresponding arc in W which joins the two
inner boundary components and is disjoint from the surface. Let W ′ be
W with a neighborhood of α removed. W ′ is a bordism in S 3 × I from
AB ∪ BC (a link in S 3 consisting of distant copies of AB and BC) to
AC. Thus W ′ gives a map from Kh(AB) ⊗ Kh(BC) → Kh(AC). This
is the definition of composition of 4-morphisms. It is not hard to show
that this is independent of the choice of p.
– The various conjugations needed to complete the definition of the 4-
category have obvious geometric definitions.
– [Need to also say something about (s)pin structures. Does this entail a
modification of the definition of 4-category?]

• We can now define a 4+1-dimensional TQFT (for spin or pin 4-manifolds)

based on this 4-category. The fields on a 2-manifold are collections of framed,
oriented points. The fields on a 3-manifold are disoriented framed tangles. The
fields on 4-manifolds are “4-dimensional spaghetti and meatballs pictures”.
(Standard, 2-dimensional spaghetti and meatballs pictures are the meat and
potatoes of the planar algebra literature. This excellent terminology is due
to Vaughan Jones. Here we double the dimensions of both the spaghetti and
the meatballs.) The 2-dimensional spaghetti consists of framed disoriented
surfaces. Each 4-dimensional meatball is a 4-ball which meets the spaghetti
in a framed disoriented link L. It is labeled by an element of Kh(L). The
local relations on these pictures are isotopy plus coalescing a subregion of
spaghetti and/or meatballs into a larger meatball. The latter relation uses the
composition of 4-morphisms defined above.

• In particular, A(B 4 ; L) ∼
= Kh(L) is this theory.
Chapter 11

n+1-dimensional Examples

11.1 Finite Group Theories

Plan:

• do twisted case too

• also mention original D-W approach via triangulations

• ? also Quinn’s finitie total homotopy generalization?

[also twisted case]

11.2 Finite Total Homotopy Theories

(including theories based on homology; or maybe put these in a separate section?)

11.3 String Category

(Include this here? not very n-dimensional. relation to graph invariants)

89
90 CHAPTER 11. N +1-DIMENSIONAL EXAMPLES
Chapter 12

Reconstruction Results

In this Chapter we show how to construct various parts of a full TQFT from various
sorts of combinatorial data (link invariants with certain properties, Moore-Seiberg
data, modular tensor category, ...)
Maybe the point of this chapter should be to make contact with prior TQFT
literature.
Maybe get rid of this chapter and instead put the various reconstruction results
at the ends of the 1+1, 2+1, 3+1 chapters (?)
Plan:

• dehn surgery formula

• moore-seiberg equations

• TV model

• also dim 2 TV model (?)

• hopf algebra with properties blah blah blah

91
92 CHAPTER 12. RECONSTRUCTION RESULTS
Chapter 13

[Other Chapters]

Some other chapters (or subchapters):

• summarize properties (axioms)

• (similarly) a summary chapter before the examples chapters

• local rels from Hopf algebras / quantum groups (and vice-versa?)

• lower bounds (both rank and norm stuff)

• various flavors of extended manifolds (p1, lagrangians, 2-framings)

• ?? tables of small theories

• ?? CS-U(1) theories

• ? general (disklike) n-cats

• (?) mu invariant as a TQFT

• need to put discussion of Frobenius-Schur stuff somewhere (this is a place-

holder) 13.0.1

• ? more category-theoretic chapter

93
94 CHAPTER 13. [OTHER CHAPTERS]
Appendix A

Categories

A.1 Definitions and Notation

A category C consists of the following data:
• Objects C 0 .

• For all a, b ∈ C 0 , a set of morphisms Cab

1 (also denoted mor(a, b)). The collec-
S
tion of all morphisms is denoted C 1 = a,b Cab 1 .

• For all a, b, c ∈ C 0 , a composition function Cab

1 × C 1 → C 1 . This is required
bc ac
to be associative. [should I bother defining associativity?]

• For all a ∈ C 0 an identity morphsism 1a ∈ Caa1 (also denoted id ). For all

a
0 1
a, b ∈ C and f ∈ Cab we require that 1a f = f = f 1b .
Let k be a ring. A k-category is a category where each morphism set Cab 1 is a A.1.1

finitely generated k-module and the composition functions are bilinear in k.

A.2 [Still to do]

(placeholder for definition of disk-like 2-category) A.2.1
(placeholder for discussion of “horizontal” composition of 2-morphisms (or A.2.2
maybe just natural transformations)
[introduce (in appropriate places) convention that juxtaposition composes one way
and ◦ and parens mean the other (more usual) way (arrows vs functions)]

• also 2-cats? or put them in a separate chapter?

• duality/conjugate stuff

• ends and coends

95
96 APPENDIX A. CATEGORIES

• relations to tensor over action, homomorphisms, invariants, coinvariants, ...

• representations = functors (2-cat reps also? if so need frob duality); “we’ll

use ‘rep’ instead of ‘functor’ because...”; also “action”; also “module” (maybe
module should be the main term?)
Appendix B

Semisimple Categories

[need to complete transition from v-space to modules (seach for all ’vector’, ’space’,
’k’)]

[use of “k-category” is inconsistent with the rest of the book (n-category)]

[need to switch from left reps to right reps (in order to be consistent with he rest of
the book)]
In this appendix we prove structure theorems (see (B.4.5) and (B.4.6)) for
semisimple k-categories. We then use this structure theorem to prove various things
about these categories. When k is a field, some authors call k-categories “algebroids”
(in analogy to groups and groupoids), and indeed most facts about semisimple alge-
bras (or more generally semisimple rings) generalize easily to categories. The proofs
given here follow Chapter XVII of [Lang xxxx] closely; in many cases they have
been adapted almost verbatim. In what follows, the reader should keep in mind the
special case of algebras, thought of as categories with only one object.

B.1 Definitions
In this appendix category will always mean k-category (see (A.1.1)), where k is a
ring. An important special case is when k is a field. As in the rest of this book,
we will write the composition of f ∈ Cab1 and g ∈ C 1 as f g, not gf (think sticking
bc
arrows together, rather than composing functions).
For the remainder of this section let C be a k-category.
A left ideal L ⊂ C is a subset of C 1 which is closed under left composition by
morphsisms of C, and such that L ∩ Cab 1 is a submodule of C 1 for all a, b ∈ C 0 . In
ab
1
other words, for all f ∈ C and g ∈ L, f g ∈ L whenever the composition is defined.
Right ideals of C are defined similarly. Two-sided ideals are both left and right
ideals.

97
 98 APPENDIX B. SEMISIMPLE CATEGORIES

A (left) representation of C (or C-representation or C-module) is a functor E

from C to the category of k-modules. For every a ∈ C 0 we have a k-module Ea ,
and for every (f : a → b) ∈ C 1 we have a linear map E(f ) : Eb → Ea . The map
E : Cab1 → hom(E , E ) is required to be k-linear. Composition of morphisms is
b a
preserved. We also say that C acts (via E) on the collection of k-modules {Ea }a∈C 0 .
We sometimes drop E from the notation for morphisms and write the action as
juxtaposition, e.g. for each (f : a → b) ∈ C 1 we have a linear map f : Eb → Ea .
Note that a left ideal of C is a C-representation.
A subrepresentation of E is a collection of submodules Fa ⊂ Ea , for all a ∈ C 0 ,
which is preserved under the action of C (i.e. f Fa ⊂ Fb for all f as above). We also
denote this as F ⊂ E.
Given F ⊂ E we can form the quotient representation E/F with (E/F )a =
Ea /Fa .
If E and E ′ are representations of C, the direct sum E ⊕ E ′ is defined by (E ⊕
E )a = Ea ⊕ Ea′ and the obvious actions of C. If F and F ′ are subrepresentations
′

of E, then the sum F + F ′ ⊂ E is defined by (F + F ′ )a = Fa + Fa′ and the obvious

actions of C. Note that F + F ′ is direct if andL only if Fa ∩ Fa′ = 0 for all a ∈ C 0 .
A free C-representation is one of the form i C∗a 1 , where a ∈ C 0 .
i i
A simple representation is one which contains no nontrivial subrepresentations.
A natural transformation (or intertwiner) h between representations E and E ′
of C consists of linear maps ha : Ea → Ea′ (for all a ∈ C 0 ) such that for all
(f : a → b) ∈ C 1 E(f )hb = ha E ′ (f ).
For a ∈ C 0 , let Ca denote the representation with (Ca)b = Cba 1 and the obvious

action of C. For any representation E and v ∈ Ea , let Cv denote the subrepresenta-

tion of E generated by v, (Cv)b = Cba 1 v. There is an obvious natural transformation

Ca → Cv.

B.2 Conditions Defining Semisimplicity

A representation is called semisimple if it satisfies the conditions in the following
proposition.

B.2.1 Proposition. The following three conditions on a representation E of C are equiv-

alent:

1. E is a sum of simple representations.

2. E is a direct sum of simple representations.

3. Every subrepresentation F ⊂ E is a direct summand: there exists F ′ ⊂ E such

that E = F ⊕ F ′ .
P P
Proof. 1 ⇒ 2. Let E = i∈I F i , with each F i ⊂ E simple. P (Recall that “ ”
here means finite sums.) Choose a maximal J ⊂ I such that j
j∈J F is direct
B.3. THE DENSITY THEOREM 99

L
(i.e. {Faj } are independent subspaces of Ea for all a ∈ C 0 ). Let E ′ = j∈J F j . I
claim that E ′ = E. It suffices to show that for all i ∈ I, F i ⊂ E ′ . Consider the
subrepresentation (F i ∩ E ′ ) ⊂ F i . F i ∩ E ′ = 0 would contradict the maximality of
J. Therefore, by theL simplicity of F i , F i ∩ E ′ = F i and F i ⊂ E ′ .
2 ⇒ 3. Let E = i∈I F i with each F i simple and let G ⊂ E. Choose a maximal
P P
J ⊂ I such that G + ( j∈J F j ) is direct (i.e. Ga ∩ ( j∈J Faj ) = 0 for all a ∈ C 0 ).
P
Let E ′ = G + ( j∈J F j ). Then for all i ∈ I, F i ∩ E ′ = F i (because F i ∩ E ′ = 0
would contradict the maximality of J), so E ′ = E.
3 ⇒ 1. First we show that any subrepresentation G of E contains a simple sub-
representation. Choose a non-zero v ∈ Ga , a ∈ C 0 . Then Cv is a subrepresentation
of G and ker(Ca → Cv) is a left ideal L 6= C. Consider the partially ordered set of
left ideals strictly smaller than Ca and containing L. By Zorn’s Lemma there is a
maximal ideal M in this set. It follows that M v is a maximal subrepresentation of
Cv. By condition 3, E = M v ⊕ M ′ for some M ′ . Then (Cv ∩ M ′ ) ⊂ G is simple by
the maximality of M v. This is the desired simple subrepresentation of G.
Let E ′ be the sum of all the simple subrepresentations of E. If E ′ 6= E then
E = E ′ ⊕ F for some F 6= 0. But by the argument above F contains a simple
subrepresentation, which is a contradiction.

Proposition. Every subrepresentation and every quotient representation of a semi- B.2.2

simple representation E is semisimple.
Proof. Let F ⊂ E. For all G ⊂ F , E = G ⊕ G′ for some G′ and F = G ⊕ (G′ ∩ F ).
Hence F is semisimple.
Write E = F ⊕ F ′ . Then F ′ is semisimple and the canonical natural transforma-
tion E → E/F induces an isomorphism F ′ → E/F . Hence E/F is semisimple.

B.3 The Density Theorem

Let E be a representation of C. Let R = EndC (E), the ring of natural transfor-
mations from E to itself. For each a ∈ C 0 , Ea is an R-module. For each f ∈ Cab
1 ,

f : Eb → Ea is an R-map (i.e. f commutes with the R-actions). We can ask how

much of HomR (Ea , Eb ) is realized by morphisms of C.
Lemma. Let C, E and R be as above and assume E is semisimple. Choose a, b ∈ B.3.1
C 0 and an R-map α : Eb → Ea . Choose v ∈ Ea . Then there exists f ∈ Cab
1 such

that αv = f v.
Proof. Let E = Cv ⊕ E ′ and let π : E → Cv be the projection. Then π ∈ R =
EndC (E), so αv = απb v = πa αv. Hence αv ∈ Cv, which implies that there is an
1 such that αv = f v.
f ∈ Cab

Next we use a diagonal trick to enhance the above result to work for a finite
number of vi ∈ Ea .
 100 APPENDIX B. SEMISIMPLE CATEGORIES

B.3.2 Theorem (Density Theorem). Let C, E and R = EndC (E) be as above and
assume E is semisimple. Choose a, b ∈ C 0 and an R-map α : Eb → Ea . Choose
1 such that αv = f v for all i.
v1 , . . . , vm ∈ Eb . Then there exists f ∈ Cab i i

Proof. Let E (m) be the direct sum of m copies of E. Then EndC (E (m) ) is isomorphic
B.3.3 to m × m matrices with entries in R. (Proof: Clearly Matm (R) ⊂ End(E (m) ). Let
(m) (m)
h ∈ End(E (m) ). Then ha : Ea → Ea decomposes as a matrix (hij a ), where each
hij
a is a linear map from Ea to itself. It is easy to see that for fixed i and j and
varying a the collection {hij (m) : E (m) → E (m)
a } is in EndC (E) = R.) Note that α b a
(m) (m)
is an EndC (E )-map, and (vi ) can be thought of as an element of Ea . It now
follows from the previous lemma that there exists f ∈ Cab 1 such that αv = f v for
i i
all i.

Next consider a family of C-representations E j , j ∈ J. Let R denote the category

of natural transformations amongst the E j . We have R0 = J and Rjl 1 are the

natural transformations from E l to E j . Then each collection of k-modules Ea =

{Eaj | j ∈ J} is a representation of R, and the actions of f ∈ Cab
1 constitute a natural
j
transformation from Eb to Ea . In other words, {Ea } is a two dimensional array of
k-modules, with C acting horizontally (say) and R acting vertically, and the C and
R actions commute. As before, we can ask whether a general natural transformation
from Eb to Ea is realized by a morphism of C.

B.3.4 Theorem (Generalized Density Theorem). Let C, {E j } and R be as above and

assume each E j is semisimple. Choose a, b ∈ C 0 and an R-map (natural transfor-
mation) α : Eb → Ea . Choose a finite set {vij }, where vij ∈ Ebj . Then there exists
1 such that α v j = f v j for all i and j.
f ∈ Cab j i i

e= L j j
Proof. Let E i,j E , where i, j runs through the index set for {vi }. (In general
there are multiple copies of each E j in the direct sum.) E e is a C-representation,
e consists of matrices whose (i, j; i′ , j ′ ) entry is a natural transformation
and EndC (E)
′
from E to E j . (This is similar to (B.3.3).) Note that α determines an EndC (E)-
j e
map from E eb to E
ea , and (v ) can be thought of as an element of E
j ea . It now follows
i
1 j j
from (B.3.1) that there exists f ∈ Cab such that αj vi = f vi for all i and j.

B.3.5 Corollary. Let C be a category and {E i } and collection of C-representations such

that (a) for all non-zero f ∈ C 1 there is some E i where f acts non-trivially, and (b)
for each a ∈ C 0 there are only finitely many i such that Eai 6= 0. Let R = EndC (E ∗ ),
the category of natural transformations amongst {E i }. Then there is a natural
isomorphism C ∼ = EndR (E∗ ).

Proof. Consider the natural map C → EndR (E∗ ). By assumption this map is in-
jective. By (B.3.4) it is surjective: Fix b ∈ C 0 and choose {vij } of (B.3.4) to be a
collection of generators of the finitely many, finitely generated, non-zero Ebj .
B.4. SEMISIMPLE CATEGORIES 101

B.4 Semisimple Categories

1 is a semisimple C-representation for all
A category C is called semisimple if C∗a
0
a ∈ C . In this section we prove structure theorems ((B.4.5) and (B.4.6)) for
semisimple categories.
Proposition. If C is semisimple, then every C-representation is semisimple.
Proof. This follows from (B.2.2) and the observation that every C-representation is
a quotient of a free C-representation.

A left ideal of C is a C-representation, and is called simple if it is simple as

a C-representation. Two ideals L and L′ of C are called isomorphic if they are
isomorphic as C-representations.
We will decompose C as a sum of its simple left ideals and thereby get a structure
theorem for C.
Lemma. Let L be a simple left ideal of C and let E be a simple representation of B.4.1
C. If L is not isomorphic to E, then LE = 0.
Proof. Since CLE = LE, LE is a subrepresentation of E. Since E is simple,
LE = E or LE = 0. If LE = E, choose y ∈ Ea such that Ly 6= 0. Again by the
simplicity of E, Ly = E. The map h : l 7→ ly is a C-map from L onto E. Since L is
simple, ker(h) = 0 and h is an isomorphism, contradicting our hypothesis.

Let {Li }i∈I be a collection of pair-wise nonisomorphic simple left ideals of C

such that every simple left ideal of C is isomorphic to some Li . For i ∈ I, define
Ci to be the sum of all simple left ideals isomorphic to Li . By (B.4.1), Ci Cj = 0 if
i 6= j. Each Ci is a left ideal, and by (B.2.1)
X
C= Ci
i∈I B.4.2
Hence for all i ∈ I
Ci ⊂ Ci C = Ci Ci ⊂ Ci
(the first inclusion because C contains identity morphisms for all objects), so Ci is
a two-sided ideal.
Choose a ∈ C 0 and write X
1a = eaj ,
j∈Ia

wherePIa is a finite subset of I and 0 6= eaj ∈ Cj . (This is possible by (B.4.2); recall

the “ ” means finite sums.) Note that eaj is the identity morphism of Cj at a.
For any x ∈ Cab 1 ,

   
X X X
x = 1a x1b =  eaj  x  ebk  = eai xebi .
j∈Ia k∈Ib i∈Ia ∩Ib
 102 APPENDIX B. SEMISIMPLE CATEGORIES

P
Furthermore, if x = i xi , with xi ∈ Ci , then necessarily xi = eai xebi . It follows
that C is a direct sum M
C= Ci .
i∈I
A category is called simple if it has only one isomorphism class of non-zero simple
left ideal. Clearly Ci above i simple for all i.
We have proved the following structure theorem for semisimple categories.
B.4.3 Theorem. Let C be a semisimple category. Let {Li }i∈I be a complete set of rep-
resentatives for the simple leftLideals of C, and let Ci be the sum of all left ideals
isomorphic to Li . Then C = i∈I Ci and each Ci is a simple category. For each
a, b ∈ C 0 , Ciab
1 is non-zero for only fintely many i.

1 needs to be defined? or is the meaning of this notation clear?]

[Ciab
It follows from (B.4.3) that the hypotheses of (B.3.5) apply to the collection of
C-representations {Li }. In order to better understand the natural transformations
amongst the Li we prove
B.4.4 Lemma (Schur Lemma). Let E, E ′ be simple representations of a category C
and h : E → E ′ be a natural transformation. Then h is an isomorphism or zero.
Assume now that k (the base ring for C) is an algebraically closed field. If h is an
isomorphism and h′ : E → E ′ is another natural transformation, then h′ = λh for
some λ ∈ k.
Proof. Both ker(h) and im(h) are subrepresentations, so by the simplicity of E and
E ′ h is either an isomorphism or zero.
Assume now that k (the base ring for C) is an algebraically closed field. Suppose
h is an isomorphism and h′ is another natural transformation. Choose a ∈ C 0 and
let λ ∈ k be an eigenvalue of h−1 ′
a ha . (Eigenvalues exist because k is algebraically
′
closed.) Then h − λh is a natural transformation with non-zero kernel. By the
simplicity of E, the kernel is all of E and therefore h′ = λh.
It follows that the only natural transformations amongst the Li are automor-
phisms. Let Ri be the division ring of automorphisms of Li . For each a ∈ C 0 , Lia
is a left Ri -module. (B.3.5) now implies
B.4.5 Theorem. Let C be a semisimple category. Let {Li }i∈I be a complete set of rep-
resentatives for the simple left ideals of C. To each object a ∈ C 0 we associate the
collection of Ri -modules {Lia }i∈I . Then C is naturally isomorphic to the category
with objects {Lia } and morphisms consisting of all (graded) Ri -maps from {Lib } to
{Lia }.
If k is an algebraically closed field, then by (B.4.4) Ri = k for all i. A complete
category of graded vector spaces is defined to be one where each object is a graded
finite dimensional vector space and Cab 1 consists of all graded linear maps from a to

b. We have proved
B.5. TO DO LIST 103

Theorem. Let C be a semisimple category over an algebraically closed field k. Let B.4.6
{Li }i∈I be a complete set of representatives for the simple left ideals of C. Then C
is naturally isomorphic to the complete category of graded vector spaces with each
object a ∈ C 0 corresponding to the graded vector space {Lia }i∈I .

Next we discuss representations...

B.5 To Do List
To do:

[placeholder] ends and coends of irreps B.5.1

[placeholder] definition of minimal idempotent for irrep B.5.2
[placeholder] positive definite inner product implies semisimplicity B.5.3

• compatible inner products

• representations

• idempotents

• ends and coends of irreps

• also tensor stuff?? Bratelli diagrams etc.

104 APPENDIX B. SEMISIMPLE CATEGORIES
Bibliography

[NOTE] [Note: Bibliography is not complete. At the moment I’m just testing
different styles for entries.]

[TO DO] [To do: need to add Segal refs, Freed, FQ, Quinn, Barrett, Sawin,
Frohman, JKB, Bullock, more Turaev, BHMV, Przytycki, Kirby-
Melvin, Roberts, ...]

[BFK03] D. Bullock, C. Frohman and J. Kania-Bartoszynska, The Yang-Mills

measure in the Kauffman bracket skein module, Comment. Math. Helv.
78 (2003), 1–17.

[CY93]

[KL94] L. H. Kauffman and S. L. Lins., Temperley-Lieb Recoupling Theory and

Invariants of 3-Manifolds, Princeton University Press, Princeton, NJ,
1994.

[Kas95] C. Kassel, Quantum Groups, Springer-Verlag, New York, 1995.

[Mac98] S. Mac Lane, Categories for the Working Mathematician, second edi-
tion, Springer, New York, 1998.

[MW06] S. Morrison and K. Walker, Fixing the functoriality of Khovanov ho-

mology, in preparation.

[Tur94] V. Turaev, Quantum Invariants of Knots and 3-manifolds, Walter de

Gruyter, Berlin, 1994.

[TV92] V. Turaev and O. Viro, State sum invariants of 3-manifolds and quan-
tum 6-j symbols, Topology 31 (1992), 862–902.

[Walk91] K. Walker, On Witten’s 3-manifold invariants, preprint 1991.

[Wall69] C. T. C. Wall, Non-additivity of the signature, Invent. Math. 7 (1969),

269–274.

105
106 BIBLIOGRAPHY

[Wit89a] E. Witten, Quantum field theory and the Jones polynomial, Comm.
Math. Phys. 121 (1989), 351–399.

[Wit89b] E. Witten, Gauge theories and integrable lattice models, Nucl. Phys. B
322 (1989), 629–697.