Functorial Field Theories and Factorization Algebras

January 20, 2014

Contents
1 Introduction 2
2 Classical eld theories 3
2.1 Classical mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Digression: Riemannian manifolds and connections . . . . . . . . . . 7
1
1 INTRODUCTION 2
1 Introduction
Here is a cartoon map of the area of mathematics this course will be exploring.
?
?
?
?
quantum
eld theories
classical
eld theories
factorization
algebras
vertex operator
algebras
functorial
eld theories
quantization
q
u
a
n
t
i
z
a
t
i
o
n
q
u
a
n
t
i
z
a
t
i
o
n
As the picture shows, the elephant in the middle is quantum eld theory, a theory that so
far has evaded a mathematical rigorous denition/construction, a fact that is indicated pic-
torially by squiggly lines. There are many competing mathematical approaches to quantum
eld theories, among them
Functorial eld theories, pioneered by Segal [Se2], Atiyah [At] and Kontsevich in the
1980s;
Vertex operator algebras, rst introduced by Borcherds in 1986 [Bo];
Factorization algebras, dened by Beilinson and Drinfeld in 2004 [BD], is a more
geometric and much broader notion than vertex operator algebras.
Factorization algebras and functorial eld theories should be thought of as mathematical
approximations to what a physicist might mean by a quantum eld theory; in particular,
2 CLASSICAL FIELD THEORIES 3
such a eld theory should give rise to a factorization algebra and a functorial eld theory as
indicated by the unlabeled squiggly arrows. Moreover, there are mathematically well-dened
objects called classical eld theories which are expected to lead to quantum eld theories by a
process called quantization, a process that is even less understood than the formal denition
of quantum eld theories. One important benchmark for the quality of a mathematical theory
that hopes to be a contender for a mathematical denition of quantum eld theory is the
success it has in constructing objects in that theory that should be considered quantizations
of classical eld theories. This has only been partially done for factorization algebras and
functorial eld theories which is indicated by the gray arrows originating from classical eld
theories.
In this course we start out discussing various classical eld theories. Then we will motivate
the denition of functorial eld theories by the heuristic discussion of a particular classical
eld theory known as the non-linear -model. Next will be the discussion of functorial
eld theories that will be based on the survey paper [ST]. Then we will be talking about
factorization algebras, drawing from [Gw] and [CG]. The nal topic will be a construction
producing a functorial eld theory from a factorization algebra. This is based on ongoing
joint work with Bill Dwyer and Peter Teichner.
2 Classical eld theories
Denition 2.1. A classical eld theory is given by the following data:
a manifold /, typically innite dimensional, which is called the space of elds;
a smooth map o : / R called the action.
This is a rst, very preliminary denition of what a classical eld theory is. As we will
see in the examples we will discuss, the space space of elds / is typically a space of smooth
maps on a manifold M or, more generally, the space of smooth sections of a bundle over M.
Moreover, for /, the action o() is given as an integral over M. This implies that the
condition for to be a critical point of o amounts to a dierential equation for , which is
called the Euler-Lagrange equation for the action functional o.
In classical physics, we often describe physical quantities in terms of functions which
satisfy dierential equations, for example
the path of a classical point particle satises Newtons equation F = ma, where F is the
net force acting on the particle, m is the mass of the particle, and a is its acceleration.
the electromagnetic eld satises Maxwells equation.
the gravitational eld satises Einsteins equation.
2 CLASSICAL FIELD THEORIES 4
To go from these dierential equations to an action function o : / R, we need to do a bit
of reverse engineering: in each of these cases, we need to construct a suitable action function
such that the associated Euler-Lagrange equation characterizing the critical points of o is
the dierential equation we start with.
2.1 Classical mechanics
For simplicity we restrict to a particular simple classical mechanical system, namely a point
particle moving in R
3
, or more generally in a Riemannian manifold X, under the inuence
of a force eld F. Now, physicists dont typically call classical mechanics a eld theory,
but from a mathematical perspective it seems ok to do that, the same way that a scalar can
be thought of as a 1 1-matrix.
Mathematically, the movement of a point particle in the manifold X is described by a
smooth map : [a, b] X with (t) X giving the position of the particle at time t [a, b].
According to Newtons law it satises the dierential equation F = ma, where F is the net
force acting on the particle, m is the mass of the particle, and a is its acceleration. Written
more carefully, Newtons law is the second order dierential equation
F((t)) = m

(t). (2.2)
Here both sides are vectors in the tangent space T
(t)
X. It is clear how to dene the acceler-
ation vector

(t) T
(t)
X if is a path in R
n
. The denition of the acceleration vector

(t)
for a path in a general Riemannian manifold X requires a discussion of connections, which
we will do in the next subsection.
Thats ne, but what is the space of elds / and the action functional o : / R?
The rst question is easy to answer:
/ := C

([a, b], X) := : [a, b] X [ is smooth.

The action functional o : / R emerges in a reformulation of Newtons law called the
Lagrangian formulation of classical mechanics. This requires the assumption that the force
vector eld F can be written in the form F = grad V for some function V : X R that
physicists call the potential. Here grad V is the gradient vector eld of V (see Denition
2.12 for the denition of the gradient vector eld of a function on a general Riemannian
manifold). The Riemannian metric and the potential allow us to dene the function
L: TX R by L(x, v) :=
m
2
[[v[[
2
V (x)
that physicists call the Lagrangian. Here x X, v T
x
X, and [[v[[
2
= v, v
x
is the norm-
squared of the tangent vector v as measured in the Riemannian metric on X (see Denition
2.10 for the denition of a (pseudo) Riemannian metric). Physicists call
m
2
[[v[[
2
the kinetic
2 CLASSICAL FIELD THEORIES 5
energy and V (x) the potential energy. The Lagrangian in turn can be used to dene an
action functional o : / R on the space / of all smooth maps : [a, b] X by setting
o() :=
_
b
a
L((t),

(t))dt. (2.3)
According to the main result of the Lagrangian formulation of classical mechanics, solu-
tions of Newtons dierential equation (2.2) correspond essentially to critical points of the
action functional o. However, we need boundary conditions to obtain the correct statement
(without them, the functional o on / has no critical points): x points x, y X and
let /
x,y
be the subspace of / consisting of smooth maps : [a, b] X with (a) = x,
(b) = y. Then the precise statement is the following.
Theorem 2.4. A path
0
/
x,y
is a critical point of the action functional o : /
x,y
R
if and only if it is a solution of Newtons equation (2.2).
If /
x,y
were a nite dimensional smooth manifold, we would know what it means for
the function o to be smooth, and what it means that /
x,y
is a critical point of o. Due
to the innite dimensional nature of /
x,y
we rst need to dene what this means in order
to be provide a proof of the above theorem rather than just a heuristic argument.
Let us rst dene when a map /
x,y
R is smooth. More generally, if M, X are
nite dimensional manifolds, possibly with boundary, and /

is a subset of the set / :=

C

(M, X) of smooth maps M X, we want to dene what it means for maps S /

and /

N to be smooth, where S, N are nite-dimensional smooth manifolds.

Denition 2.5. Let S, M, X, N be nite dimensional manifolds. A map
: S /

/ := C

(M, X)
is smooth if the corresponding map

: S M X given by

(s, x) = ((s))(x) is smooth.
A map f : /

N is smooth if for all smooth maps : S /

the composition
S

/

f
N
is smooth.
Next we want to dene what it means to say that
0
/

/ = C

(M, X) is a
critical point of a smooth function f : /

R. We note that a point

0
W of a nite
dimensional smooth manifold W is a critical point of a smooth function f : W R if and
only if for all smooth paths : R W with (0) =
0
the point 0 R is a critical point of
the composition R

W
f
R. This motivates the following denition.
2 CLASSICAL FIELD THEORIES 6
Denition 2.6. A point
0
/

/ = C

(M, X) is a critical point of a smooth function

f : /

R if for all smooth maps : R /

with (0) =
0
the point 0 R is a critical
point of the composition R

/

f
R.
Remark 2.7. The above abstract denition of smooth functions on subsets /

of mapping
spaces and their critical points works for much more general objects /

, namely contravari-
ant functors /

from the category of smooth manifolds to the category of sets (we note that
a subset /

of the set C

(M, X) of smooth maps determines such a functor by sending a

manifold S to the set of smooth maps from S to /

). This general nonsense of talking about

functors from some category to the category of sets is surprisingly useful and goes under the
name functor of point formalism.
Proof. Let us rst show that the action functional o : /
x,y
R is in fact smooth. By
Denition 2.5 we need to show that for every smooth map : S /
x,y
the composition
S

/
x,y
S
R
is smooth. The smoothness of is equivalent to the smoothness of the corresponding map

: S [a, b] X (s, t) ((s))(t).

In terms of the map

the composition o is given by
o((s)) =
_
b
a
L(

(s, t),

t

(s, t))dt (2.8)

which in fact is a smooth function of s S.
Next we want to discuss when a point
0
/
x,y
is a critical point of the action functional
o. So in view of Denition 2.6 let : R /
x,y
be a smooth path with (0) =
0
. We note
that the corresponding map

: R [a, b] X is smooth, and has the properties

(0, t) =
0
(t)

(s, a) = x

(s, b) = y.
To determine the condition on
0
guaranteeing that 0 is a critical point of the composition
(2.8), we dierentiate at s = 0 and obtain

s|s=0
(o((s)) =

s|s=0
_
b
a
L(

(s, t),

t
(s, t))dt
=

s|s=0
_
b
a
_
m
2

t
(s, t),

t
(s, t) V (

(s, t))
_
dt
2 CLASSICAL FIELD THEORIES 7
Moving the s-derivative inside the integral, and using the product (resp. chain rule) for the
rst (resp. second) term, this is equal to
_
b
a
_
m
_

t
(0, t),

2

ts
(0, t)
_
dV (

s
(0, t)
_
dt
We note that

t
(0, t) =

0
(t) T

0
(t)
X is the tangent vector of the path (t). The map
t

s
(0, t) is a vector eld along the path
0
, which geometrically can be interpreted as the
tangent vector of the path : R /
x,y
at the point (0) =
0
/
x,y
. This is often called
the innitesimal variation of the path
0
, and the notation is common for it. Using this
notation, and rewriting the second summand in terms of the gradient of V we obtain
_
b
a
_
m

0
(t),

t
()(t) grad V (
0
(t)), ()(t)
_
dt
The key step in this calculation is the next one, namely integration by parts in the rst
summand. We note that there are no boundary terms since ()(t) =

s
(0, t) vanishes for
t = a, b due to

(s, a) = x and

(s, b) = y for all s R. Using F = grad V in the second
summand, and combining both summands and putting all these steps together, we obtain

s|s=0
(o((s)) =
_
b
a
_
m

0
(t) +F(
0
(t)), ()(t)
_
dt. (2.9)
This shows that if
0
/
x,y
satises Newtons equation (2.2), then
0
is a critical point of
o.
To prove the converse statement let us assume that
0
is a critical point of o. It is not
hard to show that every section ([a, b],

0
TX) with ()(a) = ()(b) = 0 is obtained
as

s
(0, t) for some smooth map

: R [a, b] X with

(0, t) =
0
(t) and (s, a) = x,
(s, b) = y. It follows that for the corresponding path : R /
x,y
with (0) =
0
the
formula (2.9) holds. Hence our assumption that
0
is a critical point of o implies that the
right hand side of (2.9) vanishes for every ([a, b];

0
TX) with (a) = (b) = 0.
We claim that this implies that := m

+ F(
0
) ([a, b],

TX) vanishes. To see

this, assume (t
0
) ,= 0 for some t
0
(a, b). Then choose to be the product of with a
bump function that is 1 in a neighborhood of t
0
and which vanishes at a and b. Then the
integrand above is non-negative and positive for t = t
0
, leading to the desired contradiction
to our assumption that the integral is trivial.
2.1.1 Digression: Riemannian manifolds and connections
The goal of this section is twofold:
2 CLASSICAL FIELD THEORIES 8
extend the statement of Theorem 2.4 to paths : [a, b] X in a general (pseudo)
Riemannian manifold by making sense of
the gradient vector eld grad V of a smooth function V : X R;
the acceleration vector

(t) T
(t)
X.
explain the basic notion of connection on a vector bundle that will be used throughout
these lectures.
Denition 2.10. A pseudo Riemannian metric on a smooth manifold X is a family
,
x
: T
x
X T
x
X R
of symmetric bilinear forms on the tangent spaces T
x
X of X. These are required to be
non-degenerate in the sense that the map
T
x
X T

x
X := Hom(T
x
X, R) v (w v, w
x
) (2.11)
is an isomorphism from the tangent space T
x
X to the cotangent space T

x
X. Moreover, the
family

,
x
is required to depend smoothly on x. This informal statement means that the
section of the vector bundle (TX TX)

(the dual of TX TX) given by

X x (v w v, w) (TX TX)

x
= Hom(T
x
X T
x
X, R)
is smooth. If in addition the form ,
x
is positive denite for all x, then , is a
Riemannian metric.
Denition 2.12. Let V C

(X) be a smooth function on a smooth manifold X, and let

dV
1
(X) = (X, T

X) be the dierential of V . If , is a (pseudo) Riemannian

metric on X, the gradient vector eld grad V (X, TX) is the section corresponding to dV
via the isomorphism
(X, TX)

= (X, T

X)
induced by the vector bundle isomorphism TX

=
T

X given on the bers over x by the

isomorphism (2.11).
Our next goal is to dene the acceleration

(t) T
(t)
X of a smooth path : [a, b] X.
This will be done in two steps:
Construction of

. First we want to discuss

, and we start by asking: What kind of
object is

? We know what we mean by

(t) for t [a, b], namely the tangent vector of
the curve at the point (t) X. So whatever

is, we can evaluate

at any t [a, b]
to obtain the vector

(t) T
(t)
X. Restating this,

is a section of the vector bundle
over [a, b] whose ber over t [a, b] is the vector space T
(t)
X. In other words,

is a
section of the pull-back vector bundle

TX [a, b].
2 CLASSICAL FIELD THEORIES 9
Construction of

. We note that a section of the trivial vector M V M with ber
V over a smooth manifold M is the same thing as a smooth function M V to the
vector space V . Hence a section s (M, E) of a vector bundle E M can be
thought of as a generalization of a vector-valued function on M. So we should try to
interpret

=

t

as some kind of derivative of the section

([a, b],

TX) in the
direction of the standard vector eld

t
on [a, b]. This leads us to introduce covariant
derivatives in Denition 2.15 below, which is exactly the kind of gadget we are looking
for: a covariant derivative on a vector bundle E M allows us to dierentiate any
section s (M, E) in the direction of a vector eld X (M, TM) to obtain a new
section denoted
X
s (M, E) which should be thought of as the derivative of s in
the direction of the vector eld X.
Before giving the formal denition of a connection on a vector bundle E M, let
us rst motivate the denition by discussing derivatives of sections (M, E) of the trivial
vector bundle E = M V M which we identify with C

(M, V ), the smooth V -valued

functions on M. Given a vector eld X (M, TM) and a smooth function f C

(M),
we can form Xf C

(M), the derivative of f in the direction of the vector eld X. More

generally, if s C

(M, V ) = (M, E) is a V -valued function, we can form Xs C

(M),
the derivative of s in the direction of the vector eld X. For pedagogical reasons, we will use
the notation
X
s instead of Xs in this paragraph. We can ask ourselves how the section

X
s (M, E) depends on the section s and the vector eld X. In particular, we can ask
about compatibility of the construction
X
s with respect to multiplication with functions
f C

(M), noting that this multiplication makes the space of sections of any vector bundle
over M a module over the algebra C

(M). Here is the well-known answer:

X
s is C

(M)-linear in the X-variable (2.13)

X
s is a C

(M)-derivation in the s-variable (2.14)

The second property means that
X
s is R-linear in the s-variable and that it satises the
Leibniz rule

X
(fs) = (Xf)s +f
X
s
for all f C

(M).
Denition 2.15. A covariant derivative or connection on a vector bundle E M is a map
: (M, TM) (M, E) (M, E) (X, s)
X
s
with properties 2.13 and 2.14.
Example 2.16. Here are some examples of connections.
2 CLASSICAL FIELD THEORIES 10
1. If E = M V M is the trivial vector bundle, then we have the tautological
connection
taut
given by
taut
X
s := Xs as discussed in the paragraph before Denition
2.15.
2. Any vector bundle E M has a connection.
Homework 2.17. Prove this statement. Hint: Use local trivializations and the tau-
tological connection on trivial bundles to construct connections on the restrictions of
E to open subsets that cover M. Then show that these connections can be glued to
a connection on E using partitions of unity.
3. There are many connections on any vector bundle E, as can seen as follows. Suppose
is a connection on E and
A
1
(M, End(E)) = (M, T

M End(E))
is an endomorphism-valued 1-form. Then +A is again a connection, dened by
(+A)
X
s :=
X
s +A
X
s for X (M, TM), s (M, E).
Here A
X
(M, End(E)) is the bundle endomorphism obtained by evaluating the
1-form A on the vector eld X.
Homework 2.18. Show that +A is in fact a connection. Moreover, show that any
connection

is of the form +A for some A

1
(M, End(E)).
We note that these two statements imply that the space of connections is an ane
vector space for the vector space
1
(M, End(E)).
4. A (pseudo) Riemannian metric on a manifold M (see Denition 2.10) determines a
unique connection on the tangent bundle TM, called the Levi-Civita connection. It
is characterized by the following two properties:
metric property XY, Z =
X
Y, Z+Y,
X
Z for all X, Y, Z (TM); in other
words, we can dierentiate the product Y, Z C

(M) using the product rule.

torsion free property
X
Y
Y
X = [X, Y ] for all X, Y (TM), where [X, Y ]
(TM) is the Lie bracket of the vector elds X, Y .
Homework 2.19. Show that if M is the Euclidean space R
n
with its standard metric,
then the Levi-Civita connection agrees with the tautological connection on TM =
M R
n
.
5. If f : N M is a smooth map and E N is a vector bundle, then a connection
on E induces a pullback connection on the pullback bundle f

E.
REFERENCES 11
Remark 2.20. Given the uniqueness of the Levi-Civita connection on the tangent bundle
TM determined by a Riemannian metric , , it might be tempting to think that a
(pseudo) metric on any vector bundle E M determines uniquely a connection on E.
Here the notion of a (pseudo) metric on E is a slight generalization of that of a Riemannian
metric as described in Denition 2.10 by simply replacing the tangent bundle TM by the
vector bundle E.
We note that of the two properties characterizing the Levi-Civita connection, the metric
property makes sense for Y, Z (E) and X (TM), while the torsion-free property
does not make sense: if is a connection on E, then the term
X
Y requires X to be a
vector eld, and Y to be a section of E, while for the term
Y
X it is the other way around.
So we can insist that a connection on E is metric, but that condition does not determine
the connection uniquely.
Finally we ready to dene the acceleration

of a path in a Riemannian manifold X,
thus making sense of Newtons equation (2.2) in that general case.
Denition 2.21. Let : [a, b] X be a smooth path in a Riemannian manifold X, and
let

([a, b],

TX) be its tangent vector eld. Then the acceleration vector eld


([a, b],

TX) is dened by

:=
t

([a, b],

TX).
Here the connection on

TX is the pull-back of the Levi-Civita connection on TX, and

t
is the coordinate vector eld on [a, b].
If the Riemannian manifold X is the Euclidean space R
n
, then by homework problem
2.19 the Levi-Civita connection on TX is the tautological connection on TX = X R
n
.
This implies in particular that for the pull-back connection on

TX we have

=
t

=
taut

=

t

,
which is the usual denition for acceleration of a curve in Euclidean space.
Homework 2.22. Prove Theorem ?? in the general case where X is a (pseudo) Riemannian
manifold.
References
[At] Atiyah, Michael Topological quantum eld theories. Inst. Hautes tudes Sci. Publ. Math.
No. 68 (1988), 175?186 (1989)
REFERENCES 12
[BD] Beilinson, Alexander; Drinfeld, Vladimir, Chiral algebras. American Mathematical
Society Colloquium Publications, 51. American Mathematical Society, Providence,
RI, 2004. vi+375 pp.
[Bo] R. Borcherds, Vertex algebras, Kac-Moody algebras, and the Monster, Proc. Natl.
Acad. Sci. USA. vol. 83 (1986), 30683071
[CG] K. Costello and O. Gwilliam Factorization algebras in quantum eld theory,
draft monograph, available at http://www.math.northwestern.edu/
~
costello/
factorization.pdf
[Gw] O. Gwilliam, Factorization algebras and free eld theories, Northwestern thesis (2012),
available at http://math.berkeley.edu/
~
gwilliam/thesis.pdf.
[Se2] G. Segal, The denition of conformal eld theory, Topology, geometry and quantum
eld theory, 423 577, London Math. Soc. Lecture Note Ser., 308, Cambridge Univ.
Press, Cambridge, 2004.
[ST] S. Stolz and P. Teichner, Supersymmetric eld theories and generalized cohomology,
Proc. Symp. Pure Math. Vol. 83, available at http://arxiv.org/pdf/1108.0189.
pdf.