Costello

Factorization Algebras in Quantum Field
Theory
Volume 1 (8 May 2016)
Kevin Costello and Owen Gwilliam

To Lauren
&
To Sophie
Contents
Chapter 1. Introduction 1
1. The motivating example of quantum mechanics 2
2. A preliminary definition of prefactorization algebras 6
3. Prefactorization algebras in quantum field theory 7
4. Comparisons with other formalizations of quantum field theory 9
5. Overview of this volume 13
6. Acknowledgements 14
Part 1. Prefactorization algebras 17
Chapter 2. From Gaussian measures to factorization algebras 19

1. Gaussian integrals in finite dimensions 20
2. Divergence in infinite dimensions 22
3. The prefactorization structure on observables 26
4. From quantum to classical 28
5. Correlation functions 30
6. Further results on free field theories 32
7. Interacting theories 33
Chapter 3. Prefactorization algebras and basic examples 37

1. Prefactorization algebras 37
2. Associative algebras from prefactorization algebras on R 42
3. Modules as defects 43
4. A construction of the universal enveloping algebra 49
5. Some functional analysis 52
6. The factorization envelope of a sheaf of Lie algebras 61
7. Equivariant prefactorization algebras 65
Part 2. First examples of field theories and their observables 71
Chapter 4. Free field theories 73

1. The divergence complex of a measure 73
2. The prefactorization algebra of a free field theory 76
3. Quantum mechanics and the Weyl algebra 87
4. Pushforward and canonical quantization 92
5. Abelian Chern-Simons theory 94
6. Another take on quantizing classical observables 101
iii
iv CONTENTS
7. Correlation functions 106

8. Translation-invariant prefactorization algebras 108
9. States and vacua for translation invariant theories 114
Chapter 5. Holomorphic field theories and vertex algebras 119

1. Holomorphically translation-invariant prefactorization algebras 122
2. A general method for constructing vertex algebras 129
3. The βγ system and vertex algebras 140
4. Kac-Moody algebras and factorization envelopes 154
Part 3. Factorization algebras 167
Chapter 6. Factorization algebras: definitions and constructions 169

1. Factorization algebras 169
2. Factorization algebras in quantum field theory 175
3. Variant definitions of factorization algebras 176
4. Locally constant factorization algebras 180
5. Factorization algebras from cosheaves 184
6. Factorization algebras from local Lie algebras 187
Chapter 7. Formal aspects of factorization algebras 191

1. Pushing forward factorization algebras 191
2. Extension from a basis 191
3. Pulling back along an open immersion 197
4. Descent along a torsor 198
Chapter 8. Factorization algebras: examples 201

1. Some examples of computations 201
2. Abelian Chern-Simons theory and quantum groups 206
Appendix A. Background 225

1. Simplicial techniques 226
2. Operads and algebras 232
3. Lie algebras and Chevalley-Eilenberg complexes 240
4. Sheaves, cosheaves, and their homotopical generalizations 244
5. Elliptic complexes 250
Appendix B. Functional analysis 255

1. Introduction 255
2. Differentiable vector spaces 256
3. Locally convex topological vector spaces 267
4. Bornological vector spaces 270
5. Convenient vector spaces 273
6. The relevant enrichment of DVS over itself 277
7. Vector spaces arising from differential geometry 285
Appendix C. Homological algebra in differentiable vector spaces 289

CONTENTS v
1. Introduction 289
2. Linear algebra and homological algebra in DVS 290
3. Spectral sequences 292
4. Differentiable pro-cochain complexes 294
5. Homotopy colimits 296
Appendix D. The Atiyah-Bott Lemma 307
Bibliography 309
Index 315
CHAPTER 1
Introduction
This two-volume book will provide the analog, in quantum field theory, of the
deformation quantization approach to quantum mechanics. In this introduction, we
will start by recalling how deformation quantization works in quantum mechanics.
The collection of observables in a quantum mechanical system forms an asso-
ciative algebra. The observables of a classical mechanical system form a Poisson
algebra. In the deformation quantization approach to quantum mechanics, one
starts with a Poisson algebra Acl and attempts to construct an associative algebra
Aq , which is an algebra flat over the ring C[[~]], together with an isomorphism of
associative algebras Aq /~ Acl . In addition, if a, b ∈ Acl , and e
a, e
b are any lifts of
q
a, b to A , then
1 e
lim [e a, b] = {a, b} ∈ Acl .
~→0 ~
Thus, Acl is recovered in the ~ → 0 limit, i.e., the classical limit.

We will describe an analogous approach to studying perturbative quantum field
theory. In order to do this, we need to explain the following.
• The structure present on the collection of observables of a classical
field theory. This structure is the analog, in the world of field theory, of
the commutative algebra that appears in classical mechanics. We call
this structure a commutative factorization algebra.
• The structure present on the collection of observables of a quantum
field theory. This structure is that of a factorization algebra. We view
our definition of factorization algebra as a differential geometric ana-
log of a definition introduced by Beilinson and Drinfeld. However,
the definition we use is very closely related to other definitions in the
literature, in particular to the Segal axioms.
• The additional structure on the commutative factorization algebra as-
sociated to a classical field theory that makes it “want” to quantize.
This structure is the analog, in the world of field theory, of the Poisson
bracket on the commutative algebra of observables.
• The deformation quantization theorem we prove. This states that, pro-
vided certain obstruction groups vanish, the classical factorization al-
gebra associated to a classical field theory admits a quantization. Fur-
ther, the set of quantizations is parametrized, order by order in ~, by
the space of deformations of the Lagrangian describing the classical
theory.
1
2 1. INTRODUCTION
This quantization theorem is proved using the physicists’ techniques of perturba-

tive renormalization, as developed mathematically in Costello (2011b). We claim
that this theorem is a mathematical encoding of the perturbative methods developed
by physicists.
This quantization theorem applies to many examples of physical interest, in-
cluding pure Yang-Mills theory and σ-models. For pure Yang-Mills theory, it is
shown in Costello (2011b) that the relevant obstruction groups vanish and that the
deformation group is one-dimensional; thus there exists a one-parameter family of
quantizations. In Li and Li (2016), the topological B-model with target a complex
manifold X is constructed; the obstruction to quantization is that X be Calabi-Yau.
Li and Li show that the observables and correlations functions recovered by their
quantization agree with well-known formulas. In Grady et al. (n.d.), Grady, Li, and
Li describe a 1-dimensional σ-model with target a smooth symplectic manifold
and show how it recovers Fedosov quantization. Other examples are considered in
Gwilliam and Grady (2014), Costello (2010, 2011a), and Costello and Li (2011).
We will explain how (under certain additional hypotheses) the factorization
algebra associated to a perturbative quantum field theory encodes the correlation
functions of the theory. This fact justifies the assertion that factorization algebras
encode a large part of quantum field theory.
This work is split into two volumes. Volume 1 develops the theory of fac-
torization algebras, and explains how the simplest quantum field theories — free
theories — fit into this language. We also show in this volume how factorization
algebras provide a convenient unifying language for many concepts in topological
and quantum field theory. Volume 2, which is more technical, derives the link be-
tween the concept of perturbative quantum field theory as developed in Costello
(2011b) and the theory of factorization algebras.
1. The motivating example of quantum mechanics

The model problems of classical and quantum mechanics involve a particle
moving in some Euclidean space Rn under the influence of some fixed field. Our
main goal in this section is to describe these model problems in a way that makes
the idea of a factorization algebra (Section 1) emerge naturally, but we also hope
to give mathematicians some feeling for the physical meaning of terms like “field”
and “observable.” We will not worry about making precise definitions, since that’s
what this book aims to do. As a narrative strategy, we describe a kind of cartoon of
a physical experiment, and we ask that physicists accept this cartoon as a friendly
caricature elucidating the features of physics we most want to emphasize.
1.1. A particle in a box. For the general framework we want to present, the
details of the physical system under study are not so important. However, for con-
creteness, we will focus attention on a very simple system: that of a single particle
confined to some region of space. We confine our particle inside some box and
occasionally take measurements of this system. The set of possible trajectories of
the particle around the box constitute all the imaginable behaviors of this particle;
we might describe this space of behaviors mathematically as Maps(I, Box), where
1. THE MOTIVATING EXAMPLE OF QUANTUM MECHANICS 3
I ⊂ R denotes the time interval over which we conduct the experiment. We say the
set of possible behaviors forms a space of fields on the timeline of the particle.
The behavior of our theory is governed by an action functional, which is a
function on Maps(I, Box). The simplest case typically studied is the massless free
field theory, whose value on a trajectory f : I → Box is
Z
S(f) = ( f (t), f¨(t))dt.
t∈I
Here we use (−, −) to denote the usual inner product on Rn , where we view the box
as a subspace of Rn , and f¨ to denote the second derivative of f in the time variable
t.
The aim of this section is to outline the structure one would expect the observ-
ables — that is, the possible measurements one can make of this system — should
satisfy.
1.2. Classical mechanics. Let us start by considering the simpler case where
our particle is treated as a classical system. In that case, the trajectory of the particle
is constrained to be in a solution to the Euler-Lagrange equations of our theory,
which is a differential equation determined by the action functional. For example,
if the action functional governing our theory is that of the massless free theory,
then a map f : I → Box satisfies the Euler-Lagrange equation if it is a straight line.
(Since we are just trying to provide a conceptual narrative here, we will assume
that Box becomes all of Rn so that we do not need to worry about what happens at
the boundary of the box.)
We are interested in the observables for this classical field theory. Since the
trajectory of our particle is constrained to be a solution to the Euler-Lagrange equa-
tion, the only measurements one can make are functions on the space of solutions
to the Euler-Lagrange equation.
If U ⊂ R is an open subset, we will let Fields(U) denote the space of fields on
U, that is, the space of maps f : U → Box. We will let
EL(U) ⊂ Fields(U)
denote the subspace consisting of those maps f : U → Box that are solutions to
the Euler-Lagrange equation. As U varies, EL(U) forms a sheaf of spaces on R.
We will let Obscl (U) denote the commutative algebra of functions on EL(U)
(the precise class of functions we will consider will be discussed later). We will
think of Obscl (U) as the collection of observables for our classical system that only
depend upon the behavior of the particle during the time period U. As U varies, the
algebras Obscl (U) vary and together constitute a cosheaf of commutative algebras
on R.
1.3. Measurements in quantum mechanics. The notion of measurement is

fraught in quantum theory, but we will take a very concrete view. Taking a mea-
surement means that we have physical measurement device (e.g., a camera) that
we allow to interact with our system for a period of time. The measurement is then
how our measurement device has changed due to the interaction. In other words,
4 1. INTRODUCTION
we couple the two physical systems so that they interact, then decouple them and
record how the measurement device has modified from its initial condition. (Of
course, there is a symmetry in this situation: both systems are affected by their
interaction, so a measurement inherently disturbs the system under study.)
The observables for a physical system are all the imaginable measurements we
could take of the system. Instead of considering all possible observables, we might
also consider those observables which occur within a specified time period. This
period can be specified by an open interval U ⊂ R.
Thus, we arrive at the following principle.
Principle 1. For every open subset U ⊂ R, we have a set
Obs(U) of observables one can make during U.
Our second principle is a minimal version of the linearity implied by, e.g., the
superposition principle.
Principle 2. The set Obs(U) is a complex vector space.
We think of Obs(U) as being the collection of ways of coupling a measure-
ment device to our system during the time period U. Thus, there is a natural map
Obs(U) → Obs(V) if U ⊂ V is a shorter time interval. This means that the space
Obs(U) forms a precosheaf.
1.4. Combining observables. Measurements (and so observables) differ qual-
itatively in the classical and quantum settings. If we study a classical particle, the
system is not noticeably disturbed by measurements, and so we can do multiple
measurements at the same time. (To be a little less sloppy, we suppose that by
refining our measuring devices, we can make the impact on the particle as small as
we would like.) Hence, on each interval J we have a commutative multiplication
map Obs(J) ⊗ Obs(J) → Obs(J). We also have maps Obs(I) ⊗ Obs(J) → Obs(K)
for every pair of disjoint intervals I, J contained in an interval K, as well as the
maps that let us combine observables on disjoint intervals.
For a quantum particle, however, a measurement typically disturbs the system
significantly. Taking two measurements simultaneously is incoherent, as the mea-
surement devices are coupled to each other and thus also affect each other, so that
we are no longer measuring just the particle. Quantum observables thus do not
form a cosheaf of commutative algebras on the interval. However, there are no
such problems with combining measurements occurring at different times. Thus,
we find the following.
Principle 3. If U, U 0 are disjoint open subsets of R, and
U, U 0 ⊂ V where V is also open, then there is a map
? : Obs(U) ⊗ Obs(U 0 ) → Obs(V).
If O ∈ Obs(U) and O0 ∈ Obs(U 0 ), then O ? O0 is defined
by coupling our system to measuring device O during the
period U and to device O0 during the period U 0 .
Further, there are maps for an finite collection of dis-
joint time intervals contained in a long time interval, and
these maps are compatible under composition of such maps.
1. THE MOTIVATING EXAMPLE OF QUANTUM MECHANICS 5
(The precise meaning of these terms is detailed in Section

1.)
1.5. Perturbative theory and the correspondence principle. In the bulk of
this two-volume book, we will be considering perturbative quantum theory. For us,
this adjective “perturbative” means that we work over the base ring C[[~]], where
at ~ = 0 we find the classical theory. In perturbative theory, therefore, the space
Obs(U) of observables on an open subset U is a C[[~]]-module, and the product
maps are C[[~]]-linear.
The correspondence principle states that the quantum theory, in the ~ → 0
limit, must reproduce the classical theory. Applied to observables, this leads to the
following principle.
Principle 4. The vector space Obsq (U) of quantum ob-
servables is a flat C[[~]]-module such that modulo ~, it is
equal to the space Obscl (U) of classical observables.
These four principles are at the heart of our approach to quantum field theory.
They say, roughly, that the observables of a quantum field theory form a factor-
ization algebra, which is a quantization of the factorization algebra associated to
a classical field theory. The main theorem presented in this two-volume book is
that one can use the techniques of perturbative renormalization to construct factor-
ization algebras perturbatively quantizing a certain class of classical field theories
(including many classical field theories of physical and mathematical interest). As
we have mentioned, this first volume focuses on the general theory of factorization
algebras and on simple examples of field theories; this result is derived in volume
2.
1.6. Associative algebras in quantum mechanics. The principles we have
described so far indicate that the observables of a quantum mechanical system
should assign, to every open subset U ⊂ R, a vector space Obs(U), together with a
product map
Obs(U) ⊗ Obs(U 0 ) → Obs(V)
if U, U 0 are disjoint open subsets of an open subset V. This is the basic data of a
factorization algebra (see Section 1).
It turns out that in the case of quantum mechanics, the factorization algebra
produced by our quantization procedure has a special property: it is locally con-
stant (see Section 4). This means that the map Obs((a, b)) → Obs(R) is an isomor-
phism for any interval (a, b). Let A be denote the vector space Obs(R); note that A
is canonically isomorphic to Obs((a, b)) for any interval (a, b).
The product map
Obs((a, b)) ⊗ Obs((c, d)) → Obs((a, d))
when a < b < c < d, becomes, via this isomorphism, a product map
m : A ⊗ A → A.
The axioms of a factorization algebra imply that this multiplication turns A into
an associative algebra. As we will see in Section 2, this associative algebra is the
6 1. INTRODUCTION
Weyl algebra, which one expects to find as the algebra of observables for quantum
mechanics of a particle moving in Rn .
This kind of geometric interpretation of algebra should be familiar to topolo-
gists: associative algebras are algebras over the operad of little intervals in R, and
this is precisely what we have described. As we explain in Section 4, this relation-
ship continues and so our quantization theorem produces many new examples of
algebras over the operad En of little n-discs.
An important point to take away from this discussion is that associative al-
gebras appear in quantum mechanics because associative algebras are connected
with the geometry of R. There is no fundamental connection between associative
algebras and any concept of “quantization”: associative algebras only appear when
one considers one-dimensional quantum field theories. As we will see later, when
one considers topological quantum field theories on n-dimensional space times,
one finds a structure reminiscent of an En -algebra instead of an E1 -algebra.
Remark: As a caveat to the strong assertion above (and jumping ahead of our story),
note that for a manifold of the form X → R, one can push forward a factorization
algebra Obs on X × R to a factorization algebra π∗ Obs on R along the projection
map π : X × R → R. In this case, π∗ Obs((a, b)) = Obs(X × (a, b)). Hence, a quan-
tization of a higher dimensional theory will produce, via such pushforwards to R,
deformations of associative algebras, but knowing only the pushforward is typi-
cally insufficient to reconstruct the factorization algebra on the higher dimensional
manifold. ^
2. A preliminary definition of prefactorization algebras

Below (see Section 1) we give a more formal definition, but here we provide
the basic idea. Let M be a topological space (which, in practice, will be a smooth
manifold).
2.0.1 Definition. A prefactorization algebra F on M, taking values in cochain
complexes, is a rule that assigns a cochain complex F (U) to each open set U ⊂ M
along with
((i)) a cochain map F (U) → F (V) for each inclusion U ⊂ V;
((ii)) a cochain map F (U1 ) ⊗ · · · ⊗ F (Un ) → F (V) for every finite collection
of open sets where each Ui ⊂ V and the Ui are disjoint;
((iii)) the maps are compatible in a certain natural way. The simplest case of
this compatibility is that if U ⊂ V ⊂ W is a sequence of open sets, the
map F (U) → F (W) agrees with the composition through F (V)).
Remark: A prefactorization algebra resembles a precosheaf, except that we tensor

the cochain complexes rather than taking their direct sum. ^
The observables of a field theory, whether classical or quantum, form a pref-
actorization algebra on the spacetime manifold M. In fact, they satisfy a kind of
local-to-global principle in the sense that the observables on a large open set are
3. PREFACTORIZATION ALGEBRAS IN QUANTUM FIELD THEORY 7
determined by the observables on small open sets. The notion of a factorization

algebra (Section 1) makes this local-to-global condition precise.
3. Prefactorization algebras in quantum field theory

The (pre)factorization algebras of interest in this book arise from perturbative
quantum field theories. We have already discussed in Section 1 how factorization
algebras appear in quantum mechanics. In this section we will see how this picture
extends in a natural way to quantum field theory.
The manifold M on which the prefactorization algebra is defined is the space-
time manifold of the quantum field theory. If U ⊂ M is an open subset, we will in-
terpret F (U) as the collection of observables (or measurements) that we can make
which only depend on the behavior of the fields on U. Performing a measurement
involves coupling a measuring device to the quantum system in the region U.
One can bear in mind the example of a particle accelerator. In that situation,
one can imagine the space-time M as being of the form M = A × (0, t), where A is
the interior of the accelerator and t is the duration of our experiment.
In this situation, performing a measurement on an open subset U ⊂ M is some-
thing concrete. Let us take U = V × (ε, δ), where V ⊂ A is some small region in
the accelerator and where (ε, δ) is a short time interval. Performing a measurement
on U amounts to coupling a measuring device to our accelerator in the region V,
starting at time ε and ending at time δ. For example, we could imagine that there is
some piece of equipment in the region V of the accelerator, which is switched on
at time ε and switched off at time δ.
3.1. Interpretation of the prefactorization algebra axioms. Suppose that

we have two different measuring devices, O1 and O2 . We would like to set up our
accelerator so that we measure both O1 and O2 .
There are two ways we can do this. Either we insert O1 and O2 into disjoint
regions V1 , V2 of our accelerator. Then we can turn O1 and O2 on at any times we
like, including for overlapping time intervals.
If the regions V1 , V2 overlap, then we can not do this. After all, it doesn’t make
sense to have two different measuring devices at the same point in space at the
same time.
However, we could imagine inserting O1 into region V1 during the time interval
(a, b); and then removing O1 , and inserting O2 into the overlapping region V2 for
the disjoint time interval (c, d).
These simple considerations immediately suggest that the possible measure-
ments we can make of our physical system form a prefactorization algebra. Let
Obs(U) denote the space of measurements we can make on an open subset U ⊂ M.
Then, by combining measurements in the way outlined above, we would expect to
have maps
Obs(U) ⊗ Obs(U 0 ) → Obs(V)
whenever U, U 0 are disjoint open subsets of an open subset V. The associativity
and commutativity properties of a prefactorization algebra are evident.
8 1. INTRODUCTION
3.2. The cochain complex of observables. In the approach to quantum field

theory considered in this book, the factorization algebra of observables will be a
factorization algebra of cochain complexes. That is, Obs assigns a cochain complex
Obs(U) to each open U. One can ask for the physical meaning of the cochain
complex.
We will repeatedly mention observables in a gauge theory, since these kinds of
cohomological aspects are well-known for such theores.
It turns out that the “physical” observables will be H 0 (Obs(U)). If O ∈ Obs0 (U)
is an observable of cohomological degree 0, then the equation dO = 0 can often
be interpreted as saying that O is compatible with the gauge symmetries of the
theory. Thus, only those observables O ∈ Obs0 (U) that are closed are physically
meaningful.
The equivalence relation identifying O ∈ Obs0 (U) with O + dO0 , where O0 ∈
−1
Obs (U), also has a physical interpretation, which will take a little more work
to describe. Often, two observables on U are physically indistinguishable (that
is, they can not be distinguished by any measurement one can perform). In the
example of an accelerator outlined above, two measuring devices are equivalent if
they always produce the same expectation values, no matter how we prepare our
system, or no matter what boundary conditions we impose.
As another example, in the quantum mechanics of a free particle, the observ-
able measuring the momentum of a particle at time t is equivalent to that measuring
the momentum of a particle at another time t0 . This is because, even at the quan-
tum level, momentum is preserved (as the momentum operator commutes with the
Hamiltonian).
From the cohomological point of view, if O, O0 ∈ Obs0 (U) are both in the
kernel of d (and thus “physically meaningful”), then they are equivalent in the
sense described above if they differ by an exact observable.
It is a little more difficult to provide a physical interpretation for the non-zero
cohomology groups H n (Obs(U)). The first cohomology group H 1 (Obs(U)) con-
tains anomalies (or obstructions) to lifting classical observables to the quantum
level. For example, in a gauge theory, one might have a classical observable that re-
spects gauge symmetry. However, it may not lift to a quantum observable respect-
ing gauge symmetry; this happens if there is a non-zero anomaly in H 1 (Obs(U)).
The cohomology groups H n (Obs(U)), when n < 0, are best interpreted as
symmetries, and higher symmetries, of observables. Indeed, we have seen that the
physically meaningful observables are the closed degree 0 elements of Obs(U).
One can construct a simplicial set, whose n-simplices are closed and degree 0 ele-
ments of Obs(U) ⊗ Ω∗ (∆n ). The vertices of this simplicial set are observables, the
edges are equivalences between observables, the faces are equivalences between
equivalences, and so on.
The Dold-Kan correspondence (see 1.2.2) tells us that the nth homotopy group
of this simplicial set is H −n (Obs(U)). This allows us to interpret H −1 (Obs(U))
as being the group of symmetries of the trivial observable 0 ∈ H 0 (Obs(U)), and
H −2 (Obs(U)) as the symmetries of the identity symmetry of 0 ∈ H 0 (Obs(U)), and
so on.
4. COMPARISONS WITH OTHER FORMALIZATIONS OF QUANTUM FIELD THEORY 9
Although the cohomology groups H n (Obs(U)) where n ≥ 1 do not have a clear

physical interpretation as clear as that for H 0 , they are mathematically very natural
objects and it is important not to discount them. For example, let us consider a
gauge theory on a manifold M, and let D be a disc in M. Then it is often the
case that elements of H 1 (Obs(D)) can be integrated over a circle in M to yield
cohomological degree 0 observables (such as Wilson operators).
4. Comparisons with other formalizations of quantum field theory

Now that we have explained carefully what we mean by a prefactorization
algebra, let us say a little about the history of this concept, and how it compares to
other mathematical approaches to quantum field theory. We will make no attempt
to state formal theorems relating our approach to other axiom systems. Instead we
will sketch heuristic relationships between the various axiom systems.
4.1. Factorization algebras in the sense of Beilinson-Drinfeld. For us, one
source of inspiration is the work of Beilinson and Drinfeld on chiral conformal
field theory. These authors gave a geometric reformulation of the theory of vertex
algebras in terms of an algebro-geometric version of the concept of factorization
algebra. For Beilinson and Drinfeld, a factorization algebra on an algebraic curve
X is, in particular, a collection of sheaves Fn on the Cartesian powers X n of X. If
(x1 , . . . , xn ) ∈ X n is an n-tuple of distinct points in X, let F x1 ,...,xn denote the stalk
of Fn at this point in X n . The axioms of Beilinson and Drinfeld imply that there is
a canonical isomorphism
F x1 ,...,xn F x1 ⊗ · · · ⊗ F xn .
In fact, Beilinson-Drinfeld’s axioms tell us that the restriction of the sheaf Fn to
any stratum of X n (in the stratification by number of points) is determined by the
sheaf F1 on X. The fundamental object in their approach is the sheaf F1 . All the
other sheaves Fn are built from copies of F1 by certain gluing data, which we can
think of as structures put on the sheaf F1 .
One should think of the stalk F x of F1 at x as the space of local operators in a
field theory at the point x. Thus, F1 is the sheaf on X whose stalks are the spaces of
local operators. The other structures on F1 reflect the operator product expansions
of local operators.
Let us now sketch, heuristically, how we expect their approach to be related to
ours. Suppose we have a factorization algebra Fe on X in our sense. Then, for every
open V ⊂ X, we have a vector space Fe(V) of observables on V. The space of local
operators associated to a point x ∈ X should be thought of as those observables
which live on every open neighbourhood of x. In other words, we can define
Fex = lim Fe(V)
x∈V
to be the limit over open neighbourhoods of x of the observables on that neigh-
bourhood. This limit is the costalk of the pre-cosheaf Fe.
Thus, the heuristic translation between their axioms and ours is that the sheaf
F1 on X that they construct should have stalks coinciding with the costalks of our
10 1. INTRODUCTION
factorization algebra. We do not know how to turn this idea into a precise theorem
in general. We do, however, have a precise theorem in one special case.
Beilinson-Drinfeld show that a factorization algebra in their sense on the affine
line A1 , which is also translation and rotation equivariant, is the same as a vertex
algebra. We have a similar theorem. We show in Chapter 5 that a factorization
algebra on C that is translation and rotation invariant, and also has a certain “holo-
morphic” property, gives rise to a vertex algebra. Therefore, in this special case,
we can show how a factorization algebra in our sense gives rise to one in the sense
used by Beilinson and Drinfeld.
4.2. Segal’s axioms for quantum field theory. Segal has developed and stud-
ied some very natural axioms for quantum field theory Segal (2010). These axioms
were first studied in the world of topological field theory by Atiyah, Segal, and
Witten, and in conformal field theory by Kontsevich and Segal (2004).
According to Segal’s philosophy, a d-dimensional quantum field theory (in Eu-
clidean signature) is a symmetric functor from the category CobRiem
d of d-dimensional
Riem
Riemannian cobordisms. An object of the category Cobd is a compact d − 1-
manifold together with a germ of a d-dimensional Riemannian structure. A mor-
phism is a d-dimensional Riemannian cobordism. The symmetric monoidal struc-
ture arises from disjoint union. As defined, this category does not have identity
morphisms, but they can be added in formally.
4.2.1 Definition. A Segal field theory is a symmetric monoidal functor from CobRiem
d
to the category of (topological) vector spaces.
We won’t get into details about what kind of topological vector spaces one
should consider, because our aim is just to sketch a heuristic relationship between
Segal’s picture and our picture.
In our approach to studying quantum field theory, the fundamental objects are
not the Hilbert spaces associated to codimension 1 manifolds, but rather the spaces
of observables. Any reasonable axiom system for quantum field theory should
be able to capture the notion of observable. In particular, we should be able to
understand observables in terms of Segal’s axioms.
Segal (in lectures and conversations) has explained how to do this. Suppose
we have a Riemannian manifold M and a point x ∈ M. Consider a ball B(x, r)
of radius r around x, whose boundary is a sphere S (x, r). Segal explains that the
Hilbert space Z(S (x, r)) should be thought of as the space of operators on the ball
B(x, r).
If r < r0 , there is a cobordism S (x, r) → S (x, r0 ) given by the complement of
B(x, r) in the closed ball B(x, r0 ). This gives rise to maps Z(S (x, r)) → Z(S (x, r0 )).
Segal defines the space of local operators at x to be the limit
lim Z(S (x, r))
r→0
of this inverse system.
One can understand from this idea of Segal’s how one should construct some-
thing like a prefactorization algebra on any Riemannian manifold M of dimension
4. COMPARISONS WITH OTHER FORMALIZATIONS OF QUANTUM FIELD THEORY 11
n from a Segal field theory. Given an open subset U ⊂ M whose boundary is a

codimension 1 submanifold ∂U, we define the space F (U) of observables on U to
be Z(∂U). If U, V, W are three such opens in M, such that the closures of U and V
are disjoint in W, then there is a cobordism
W \ (U q V) : ∂U q ∂V → ∂W.
This cobordism induces a map
F (U) ⊗ F (V) = Z(∂U) ⊗ Z(∂V) → Z(∂W) = F (W).
There are similar maps defined when U1 , . . . , Un , W are opens with smooth bound-
ary such that the closures of the Ui are disjoint and contained in W. In this way, we
can construct from a Segal field theory something which is very like a prefactoriza-
tion algebra; the only difference is that we restrict our attention to those opens with
smooth boundary, and the prefactorization algebra structure maps are only defined
for collections of opens whose closures are disjiont.
Remark: Below we discuss how certain universal factorization algebras relate to
topological field theories in the style of Atiyah-Segal-Lurie. Dwyer, Stolz, and Te-
ichner have also proposed an approach to constructing Segal-style non-topological
field theories, such as Riemannian field theories, using factorization algebras. ^
4.3. Topological field theory. One class of field theories for which there ex-
ists an extensive mathematical literature is topological field theories (see, for in-
stance, Lurie (2009b)). One can ask how our axiom system relates to those for
topological field theories.
There is a subclass of factorization algebras that appear in topological field
theories, called locally constant factorization algebras. A factorization algebra F
on a manifold M, valued in cochain complexes, is locally constant if, for any two
discs D1 ⊂ D2 in M, the map F (D1 ) → F (D2 ) is a quasi-isomorphism. A theorem
of Lurie (n.d.b) shows that, given a locally constant factorization algebra F on Rn ,
the complex F (D) has the structure of an En algebra.
This relationship matches with what one expects from the standard approach
to the axiomatics of topological field theory. According to the standard axioms
for topological field theories, a topological field theory (TFT) of dimension n is a
symmetric monoidal functor
Z : Cobn → Vect
from the n-dimensional cobordism category to the category of vector spaces. Here,
Cobn is the category whose objects are closed smooth n − 1-manifolds and the
morphisms are cobordisms between them. (It is standard, following Freed (1994),
to also consider higher-categorical objects associated to manifolds of higher codi-
mension).
For an n − 1-manifold N, we should interpret Z(N) as the Hilbert space of the
TFT on N. Then, following standard physics yoga, we should interpret Z(S n−1 ) as
the space of local operators of the theory. There are natural cobordisms between
disjoint unions of the n − 1-sphere which make Z(S n−1 ) into an En algebra. For
12 1. INTRODUCTION
example, in dimension 2, the pair of pants with k legs provides the k-ary operations
for the E2 algebra structure on Z(S 1 ).
This story fits nicely with our approach. If we have a locally constant factoriza-
tion algebra F on Rn , then F (Dn ) is an En algebra. Further, we interpret F (Dn ) as
being the space of observables supported on an n-disc. Since F is locally constant,
this may as well be the observables supported on a point, because it is independent
of the radius of the n-disc.
4.4. Correlation functions. In some classic approaches to the axiomatics of

quantum field theory — such as the Wightman axioms or the Osterwalder-Schrader
axioms, their Euclidean counterpart — the fundamental objects are correlation
functions. While we make no attempt to verify that a factorization algebra gives
rise to a solution of any of these axiom systems, we do show that the factorization
algebra has enough data to define the correlation functions. Let us briefly explain
how this works for two different classes of examples.
Suppose that we have a factorization algebra F on a manifold M over the ring
R[[~]] of formal power series in ~. Suppose that
H 0 (F (M)) = R[[~]].
This condition holds in some natural examples: for instance, for Chern-Simons
theory on R3 or for a massive scalar field theory on a compact manifold M.
Let U1 , . . . , Un ⊂ M be disjoint open sets. The factorization product gives us a
R[[~]]-multilinear map
h−i : H 0 (F (U1 )) × · · · × H 0 (F (Un )) → H 0 (F (M)) = R[[~]].
In this way, given observables Oi ∈ H 0 (F (Ui )), we can produce a formal power
series in ~
hO1 , . . . , On i ∈ R[[~]].
In the field theories we just mentioned, this map does compute the expectation
value of observables. (See Section 6, where we describe this map in terms of
Green’s functions. We also recover the Gauss linking number there from Abelian
Chern-Simons theory.)
This construction doesn’t give us expectation values in every situation where
we might hope to construct them. For example, if we work with a field theory
on Rn , it is rarely the case that H 0 (F (Rn )) is isomorphic to R[[~]]. We would,
however, expect to be able to define correlation functions in this situation. To
achieve this, we define a variant of this construction that works well on Rn . Given a
factorization algebra on Rn with ground ring R[[~]] as before, we define in Section
9 a vacuum to be an R[[~]]-linear map
h−i : H 0 (F (Rn )) → R[[~]]
that is translation-invariant and satisfies a certain cluster decomposition principle.
After choosing a vacuum, one can define correlation functions in the same way as
above.
5. OVERVIEW OF THIS VOLUME 13
5. Overview of this volume

This two-volume work concerns, as the titles suggests, factorization algebras
and quantum field theories. In this introduction so far, we have sketched what a
prefactorization algebra is and why it might help organize and analyze the behavior
of the observables of a quantum field theory. These two volumes develop these
ideas further in a number of ways.
The first volume focuses on factorization algebras: their definition, some for-
mal properties, and some simple constructions of factorization algebras. The quan-
tum field theory in this volume is mostly limited to free field theories. In a moment,
we give a detailed overview of this volume.
In the second volume, we focus on our core project: we develop the Batalin-
Vilkovisky formalism for both classical and quantum field theories and show how
it automatically produces a deformation quantization of factorization algebras of
observables. In particular, Volume 2 will introduce the factorization algebras as-
sociated to interacting field theories. We also provide there a refinement of the
Noether theorems in the setting of factorization algebras, in which, roughly speak-
ing, local symmetries of a field theory lift to a map of factorization algebras,. This
map realizes the symmetries as observables of the field theory. For a more detailed
overview of Volume 2, see its introductory chapter.
5.1. Chapter by chapter. Chapter 2 serves as a second introduction. In this
chapter we explain, using informal language and without any background knowl-
edge required, how the observables of a free scalar field theory naturally form a
prefactorization algebra. The reader who wants to understand the main ideas of this
two-volume work with the minimum amount of technicalities should start there.
Chapter 3 gives a more careful definition of the concept of a prefactorization
algebra, and analyzes some basic examples. In particular, the relationship between
prefactorization algebras on R and associative algebras is developed in detail. We
also introduce a construction that will play an important role in the rest of the
book: the factorization envelope of a sheaf of Lie algebras on a manifold. This
construction is the factorization version of the universal enveloping algebra of a
Lie algebra.
In Chapter 4 we revisit the prefactorization algebras associated to a free field
theory, but with more care and in greater generality than we used in Chapter 2. The
methods developed in this chapter apply to gauge theories, using the BV/BRST
method to treat gauge symmetry. We analyse in some detail the example of Abelian
Chern-Simons theory, and verify that the expectation value of Wilson lines in this
theory recovers the Gauss linking number.
Chapter 5 introduces the concept of a holomorphic prefactorization algebra on
Cn . The prefactorization algebra of observables of a field theory with a holomor-
phic origin — such as a holomorphic twist Costello (2013) of a supersymmetric
gauge theory — will be such a holomorphic factorization algebra. We prove that
a holomorphic prefactorization algebra on C gives rise to a vertex algebra, thus
linking our story with a more traditional point of view on the algebra of operators
of a chiral conformal field theory.
14 1. INTRODUCTION
The final chapters in this book develop the concept of factorization algebra,
by adding a certain local-to-global axiom to the definition of prefactorization alge-
bra. In Chapters 6, we provide the definition, discuss the relation between locally-
constant factorization algebras and En algebras, and explain how to construct sev-
eral large classes of examples. In 7, we develop some formal properties of the
theory of factorization algebras. Finally, in 8, we move beyond the formal and
analyze some interesting explicit examples. For instance, we compute the fac-
torization homology of the Kac-Moody enveloping factorization algebras, and we
explain how Abelian Chern-Simons theory produces a quantum group.
5.2. A comment on functional analysis and algebra. This book uses an un-
usual array of mathematical techniques, including both homological algebra and
functional analysis. The homological algebra appears because our factorization al-
gebras live in the world of cochain complexes (ultimately, because they come from
the BV formalism for field theory). The functional analysis appears because our
factorization algebras are built from vector spaces of an analytic nature, such as the
space of distributions on a manifold. We have included an expository introduction
to the techniques we use from homological algebra, operads, and sheaf theory in
Appendix A.
It is well-known that it is hard to make homological algebra and functional
analysis work well together. One reason is that, traditionally, the vector spaces
that arise in analysis are viewed as topological vector spaces, and the category of
topological vector spaces is not an Abelian category. In Appendix B, we introduce
the concept of differentiable vector spaces. Differentiable vector spaces are more
flexible than topological vector spaces, yet retain enough analytic structure for our
purposes. We show that the category of differentiable vector spaces is an Abelian
category, and indeed satisfies the strongest version of the axioms of an Abelian
category: it is a locally-presentable AB5 category. This means that homological al-
gebra in the category of differentiable vector spaces works very nicely. We develop
this in Appendix C.
A gentle introduction to differentiable vector spaces, containing more than
enough to follow everything in both volumes, is contained in Chapter 3, section 5.
6. Acknowledgements
The project of writing this book has stretched over many more years than we
ever anticipated. In that course of time, we have benefited from conversations with
many people, the chance to present this work at several workshops and confer-
ences, and the feedback of various readers. The book and our understanding of
this material is much better due to the interest and engagement of so many others.
Thank you.
We would like to thank directly the following people, although this list is un-
doubtedly incomplete: David Ayala, David Ben-Zvi, Dan Berwick-Evans, Damien
Calaque, Alberto Cattaneo, Ivan Contreras, Vivek Dhand, Chris Douglas, Chris
Elliott, John Francis, Dennis Gaitsgory, Sachin Gautam, Ezra Getzler, Greg Ginot,
6. ACKNOWLEDGEMENTS 15
Ryan Grady, Rune Haugseng, Theo Johnson-Freyd, David Kazhdan, Si Li, Ja-
cob Lurie, Takuo Matsuoka, Pavel Mnev, David Nadler, Thomas Nikolaus, Fred
Paugam, Dmitri Pavlov, Toly Preygel, Kasia Rejzner, Nick Rozenblyum, Clau-
dia Scheimbauer, Graeme Segal, Thel Seraphim, Yuan Shen, Jim Stasheff, Stephan
Stolz, Dennis Sullivan, Matt Szczesny, Hiro Tanaka, Peter Teichner, David Treumann,
Philsang Yoo, Brian Williams, and Eric Zaslow for helpful conversations. K.C. is
particularly grateful to Graeme Segal for many illuminating conversations about
quantum field theory over the years. O.G. would like to thank the participants in
the Spring 2014 Berkeley seminar and Fall 2014 MPIM seminar for their interest
and extensive feedback; he is also grateful to Rune Haugseng and Dmitri Pavlov
for help at the intersection of functional analysis and category theory. Peter Te-
ichner’s feedback after teaching a Berkeley course in Spring 2016 on this material
improved the book considerably. Finally, we are both grateful to John Francis and
Jacob Lurie for introducing us to factorization algebras, in their topological incar-
nation, in 2008.
Our work has taken place at many institutions — Northwestern University,
the University of California, Berkeley, the Max Planck Institute for Mathematics,
and the Perimeter Institute for Theoretical Physics — which provided a supportive
environment for writing and research. Research at the Perimeter Institute is sup-
ported by the Government of Canada through Industry Canada and by the Province
of Ontario through the Ministry of Economic Development and Innovation, and
K.C.’s research at the Perimeter Institute is supported by the Krembil Foundation.
We have also benefitted from the support of the NSF: K.C. and O.G. were par-
tially supported by NSF grants DMS 0706945 and DMS 1007168, and O.G. was
supported by NSF postdoctoral fellowship DMS-1204826. K.C. was also partially
supported by a Sloan fellowship.
Finally, during the period of this project, we have depended upon the unrelent-
ing support and love of our families. In fact, we have both gotten married and had
children over these years! Thank you Josie, Dara, and Laszlo for being models of
insatiable curiosity, ready sources of exuberant distraction, and endless fonts of joy.
Thank you Lauren and Sophie for joining and building our lives together. Amidst
all the tumult and the moves, you always encourage us, you always share your wit
and wisdom and warmth. And it makes all the difference.
Part 1
Prefactorization algebras
CHAPTER 2
From Gaussian measures to factorization algebras
This chapter serves as a kind of second introduction, demonstrating how the

free scalar field theory on a manifold M produces a factorization algebra on M. In
the first chapter, we have already sketched why the observables of a field theory
ought to form a factorization algebra, without making precise what we meant by
a quantum field theory. Here we will meet the second main theme of the book —
the Batalin-Vilkovisky (BV) formalism for field theory — and see how it naturally
produces a factorization algebra.
Our approach to quantum field theory grows out of the idea of a path integral.
Instead of trying to directly define such an integral, however, the BV formalism
provides a homological approach to integration, similar in spirit to the de Rham
complex. (As we will see in the next section, in the finite dimensional setting,
the BV complex is isomorphic to the de Rham complex.) The philosophy goes
like this. If the desired path integral were well-defined mathematically, we could
compute the expectation values of observables, and the expectation value map E
is linear, so we obtain a linear equivalence relation between observables O ∼ O0
whenever E(O − O0 ) = 0. We can reconstruct, in fact, the expectation value by
describing the inclusion RelE := ker E ,→ Obs and taking the cokernel of this
inclusion. In other words, we identify “integrands with the same integral.” The
BV formalism approaches the problem from the other direction: even though the
desired path integral may not be well-defined, we often know, from physical argu-
ments, when two observables ought to have the same expectation value (e.g., via
Ward identities), so that we can produce a subspace RelBV ,→ Obs. The BV for-
malism produces a subspace RelBV determined by the classical field theory, as we
will see below, but this subspace is typically not of codimension 1. Further input,
like boundary conditions, are often necessary to get a number (i.e., to produce a
codimension 1 subspace of relations). Nonetheless, the relations in RelBV would
hold for any such choice, so any expectation value map coming from physics would
factor through Obs /RelBV .
In fact, the BV formalism produces a cochain complex, encoding relations
between the relations and so on, whose zeroth cohomology group is the space
Obs /RelBV . (Here, we are using Obs to denote the “naive” observables that one
would first write down for the theory, not observables involving the auxiliary “anti-
fields” introduced when applying the BV formalism.)
That description is quite abstract; the rest of this chapter is about making the
idea concrete with examples. In physics, a free field theory is one where the action
functional is a purely quadratic function of the fields. A basic example is the free
19
20 2. FROM GAUSSIAN MEASURES TO FACTORIZATION ALGEBRAS
scalar field theory on a Riemannian manifold (M, g), where the space of fields is
the space C ∞ (M) of smooth functions on M, and the action funtional is
Z
S (φ) = 12 φ(4g + m2 )φ dvolg .
M
Here 4g refers to the Laplacian with the convention that its eigenvalues are non-
negative, and dvolg denotes the Riemannian volume form associated to the metric.
The positive real number m is the mass of the theory. The main quantities of
interest in the free field theory are the correlation functions, defined by the heuristic
expression
Z
hφ(x1 ) · · · φ(xn )i = φ(x1 ) . . . φ(xn ) e−S (φ) dφ,
φ∈C ∞ (M)
where x1 , . . . , xn are points in M. Note that there is an observable

O(x1 , . . . , xn ) : φ 7→ φ(x1 ) · · · φ(xn )
sending a field φ to the product of its values at those points. The support of this
observable is precisely the set of points {x1 , . . . , xn }, so this observable lives in
Obs(U) for any open U containing all those points. The standard computations in
quantum field theory tell us that this observable has the same expectation value
as linear combinations of other correlation functions. For instance, Wick’s lemma
tells us how the two-point correlation function relates to the Green’s function for
4g + m2 .
Our task is to explain how the combination of the BV formalism and prefactor-
ization algebras provides a simple and natural way to make sense of these relations.
We will see that for a free theory on a manifold M, there is a space of observables
associated to any open subset U ⊂ M. We will see that that the operations we
can perform on these spaces of observables give us the structure of a prefactoriza-
tion algebra on M. This example will serve as further motivation for the idea that
observables of a field theory are described by a prefactorization algebra.
1. Gaussian integrals in finite dimensions

As in many approaches to quantum field theory, we will motivate our defini-
tion of the prefactorization algebra of observables by studying finite dimensional
Gaussian integrals. Thus, let A be an n × n symmetric positive-definite real matrix,
and consider Gaussian integrals of the form
Z X
exp − 21 xi Ai j x j f (x) dn x
x∈Rn
where f is a polynomial function on Rn . Note the formal analogy to the correlation
function we wish to compute: here x replaces φ, 12 xi Ai j x j is quadratic in x as S
P
is in φ, and f (x) is polynomial in x as O(φ) is a polynomial in φ.
Most textbooks on quantum field theory would explain, at this point, Wick’s
lemma, which is a combinatorial expression for such integrals. It reduces the in-
tegral above to a sum involving the quadratic moments of the Gaussian measure.
1. GAUSSIAN INTEGRALS IN FINITE DIMENSIONS 21
Then, such a textbook would go on to define similar infinite-dimensional Gauss-

ian integrals using the analogous combinatorial expression. The key point is the
simplicity of the moments of a Gaussian measure, which allows immediate gener-
alization to infinite dimensions.
We will take a different approach, however. Instead of focusing on the com-
binatorial expression for the integral, we will focus on the divergence operator
associated to the Gaussian measure. (This operator provides the inclusion map
Rel ,→ Obs discussed at the beginning of this chapter.)
Let P(Rn ) denote the space of polynomial functions on Rn . Let Vect(Rn ) denote
the space of vector fields on Rn with polynomial coefficients. If dn x denotes the
Lebesgue measure on Rn , let ωA denote the measure
X
ωA = exp − 21 xi Ai j x j dn x.
Then, the divergence operator DivωA associated to this measure is a linear map
DivωA : Vect(Rn ) → P(Rn ),
defined abstractly by saying that if V ∈ Vect(Rn ), then
LV ωA = DivωA V ωA

where LV refers to the Lie derivative. Thus, the divergence of V measures the
infinitesimal change in volume that arises when one applies the infinitesimal dif-
feomorphism V.
In coordinates, the divergence is given by the formula
X ∂ ! X X ∂ fi
(†) DivωA fi =− fi x j Ai j + .
∂xi i, j i
∂x i
(This formula is an exercise in applying Cartan’s magic formula: LV = [d, ιV ].

Note that this divergence operator is therefore a disguised version of the exterior
derivative.)
By the definition of divergence, we see
Z
DivωA V ωA = 0

for all polynomial vector fields V, because

Z Z Z
DivωA V ωA = LV ω A = d(ιV ωA )

and then we apply Stokes’ lemma. By changing basis of Rn to diagonalize A, one

sees that the image of the divergence map is a codimension 1 linear subspace of
the space P(Rn ) of polynomials on Rn . (This statement is true as long as A is
non-degenerate; positive-definiteness is not required).
Let us identify P(Rn )/ Im DivωA with R by taking the basis of the quotient
space to be the image of the polynomial function 1. What we have shown so far
can be summarized in the following lemma.
1.0.1 Lemma. The quotient map

P(Rn ) → P(Rn )/ Im DivωA R
is the map that sends a function f to its expected value
n f ωA
R
h f iA := RR .
ω
Rn A
This lemma plays a crucial motivational role for us. If we want to know ex-
pected values (which are the main interest in the physics setting), it suffices to
describe a divergence operator. One does not need to produce the measure directly.
One nice feature of this approach to finite-dimensional Gaussian integrals is
that it works over any ring in which det A is invertible (this follows from the ex-
plicit algebraic formula we wrote for the divergence of a polynomial vector field).
This way of looking at finite-dimensional Gaussian integrals was further analyzed
in Gwilliam and Johnson-Freyd (n.d.), where it was shown that one can derive the
Feynman rules for finite-dimensional Gaussian integration from such considera-
tions.
Remark: We should acknowledge here that our choice of polynomial functions and
vector fields was important. Polynomial functions are integrable against a Gaussian
measure, and the divergence of polynomial vector fields produce all the relations
between these integrands. If we worked with all smooth functions and smooth
vector fields, the cokernel would be zero. In the BV formalism, just as in ordinary
integration, the choice of functions plays an important role. ^
2. Divergence in infinite dimensions

So far, we have seen that finite-dimensional Gaussian integrals are entirely
encoded in the divergence map from the Gaussian measure. In our approach
to infinite-dimensional Gaussian integrals, the fundamental object we will define
is such a divergence operator. We will recover the usual formulae for infinite-
dimensional Gaussian integrals (in terms of the propagator or Green’s function)
from our divergence operator. Further, we will see that analyzing the cokernel of
the divergence operator will lead naturally to the notion of prefactorization algebra.
For concreteness, we will work with the free scalar field theory on a Riemann-
ian manifold (M, g), which need not be compact. We will define a divergence
operatorR for the putative Gaussian measure on C ∞ (M) associated to the quadratic
form 12 M φ(4g + m2 )φ dvolg . (Here 4g + m2 plays the role that the matrix A did in
the preceding section.)
Before we define the divergence operator, we need to define spaces of polyno-
mial functions and of polynomial vector fields. We will organize these spaces by
their support in M. Namely, for each open subset U ⊂ M, we will define polyno-
mial functions and vector fields on the space C ∞ (U).
The space of all continuous linear functionals on C ∞ (U) is the space Dc (U) of
compactly supported distributions on U. In order to define the divergence operator,
we need to restrict to functionals with more regularity. Hence we will work with
2. DIVERGENCE IN INFINITE DIMENSIONS 23
Cc∞ (U), where every element f ∈ Cc∞ (U) defines a linear functional on C ∞ (U) by
the formula Z
φ 7→ f φ dvolg .
U
ym (People sometimes call them “smeared” because these do not include beloved
functionals like delta functions, only smoothed-out approximations to them.)
As a first approximation to the algebra we wish to use, we define the space of
polynomial functions on Cc∞ (U) to be
e ∞ (U)) = Sym Cc∞ (U),
P(C
i.e., the symmetric algebra on Cc∞ (U). An element of P(C e c∞ (U)) that is homoge-
neous of degree n can be written as a finite sum of monomials f1 · · · fn where the
fi ∈ Cc∞ (U). Such a monomial defines a function on the space C ∞ (U) of fields by
the formula
Z
φ 7→ f1 (x1 )φ(x1 ) . . . fn (xn )φ(xn ) dvolg (x1 ) ∧ · · · ∧ dvolg (xn ).
(x1 ,...,xn )∈U n
Note that because Cc∞ (U) is a topological vector space, it is more natural to use
an appropriate completion of this purely algebraic symmetric power Symn Cc∞ (U).
Because this version of the algebra of polynomial functions is a little less natural
than the completed version, which we will introduce shortly, we use the notation
e The completed version is denoted P.
P.
We define the space of polynomial vector fields in a similar way. Recall that if
V is a finite-dimensional vector space, then the space of polynomial vector fields
on V is isomorphic toP(V) ⊗ V, where P(V) is the space of polynomial functions
on V. An element X = f ⊗ v, with f a polynomial, acts on a polynomial g by the
formula
∂g
X(g) = f .
∂v
In particular, if g is homogeneous of degree n and we pick a representative e g ∈
(V ) , then ∂g/∂v denotes the degree n − 1 polynomial
∗ ⊗n
w 7→ e
g(v ⊗ w ⊗ · · · ⊗ w).
In other words, for polynomials, differentiation is a version of contraction.
In the same way, we would expect to work with
g ∞ (U)) = P(C
Vect(C e ∞ (U)) ⊗ C ∞ (U).
We are interested, in fact, in a different class of vector fields. The space C ∞ (U)
has a foliation, coming from the linear subspace Cc∞ (U) ⊂ C ∞ (U). We are actually
interested in vector fields along this foliation, due to the role of variational calcu-
lus in field theory. This restriction along the foliation is clearest in terms of the
divergence operator we describe below, so we explain it after Definition 2.0.1.
Thus, let
g c (C ∞ (U)) = P(C
Vect e ∞ (U)) ⊗ Cc∞ (U).
Again, it is more natural to use a completion of this space that takes account of the
topology on Cc∞ (U). We will discuss such completions shortly.
g c (C ∞ (U)) can be written as an finite sum of monomials of

Any element of Vect
the form
∂
f1 · · · fn
∂φ
∂
for fi , φ ∈ Cc∞ (U). By ∂φ we mean the constant-coefficient vector field given by
infinitesimal translation in the direction φ in C ∞ (U).
Vector fields act on functions, in the same way as we described above: the
formula is
∂ X Z
f1 · · · fn (g1 · · · gm ) = f1 · · · fn g1 · · · gbi · · · gm gi (x)φ(x) dvolg
∂φ U
where dvolg is the Riemannian volume form on U.
2.0.1 Definition. The divergence operator associated to the quadratic form
Z
S (φ) = φ(4 + m2 )φ dvolg
U
is the linear map
Div g c (C ∞ (U)) → P(C
g : Vect e c∞ (U))
defined by
n
g f1 · · · fn ∂ = − f1 · · · fn (4+m2 )φ+
! X Z
(‡) Div f1 · · · b
fi · · · fn φ(x) fi (x) dvolg .
∂φ i=1 U
Note that this formula is entirely parallel to the formula for divergence of a
Gaussian measure in finite dimensions, given in formula (†). Indeed, the formula
makes sense even when φ is not compactly supported; however, the term
f1 · · · fn (4 + m2 )φ
need not be compactly supported if φ is not compactly supported. To ensure that
the image of the divergence operator is in P(C e c∞ (U)), we only work with vector
fields with compact support, namely Vect g c (C ∞ (U)).
As we mentioned above, it is more natural to use a completion of the spaces
P(C ∞ (U)) and Vectc (C ∞ (U)) of polynomial functions and polynomial vector fields.
e
We now explain a geometric approach to such a completion.
Let dvolgn denote the Riemannian volume form on the product space U n arising
from the natural n-fold product metric induced by the metric g on U. Any element
F ∈ Cc∞ (U n ) then defines a polynomial function on C ∞ (U) by
Z
φ 7→ F(x1 , . . . , xn )φ(x1 ) · · · φ(xn ) dvolgn .
Un
This functional does not change if we permute the arguments of F by an element
of the symmetric group S n , so that this function only depends on the image of F
in the coinvariants of Cc∞ (U n ) by the symmetric group action. This quotient, of
course, is isomorphic to invariants for the symmetric group action.
2. DIVERGENCE IN INFINITE DIMENSIONS 25
Therefore we define
M
P(C ∞ (U)) = Cc∞ (U n )S n ,
n≥0
where the subscript indicates coinvariants. The (purely algebraic) symmetric power
Symn Cc∞ (U) provides a dense subspace of Cc∞ (U n )S n Cc∞ (U n )S n . Thus, P(C
e ∞ (U))
∞
is a dense subspace of P(C (U)).
In a similar way, we define Vectc (C ∞ (U)) by
M
Vectc (C ∞ (U)) = Cc∞ (U n+1 )S n ,
n≥0
where the symmetric group S n acts only on the first n factors. A dense subspace
of Cc∞ (U n+1 )S n is given by Symn Cc∞ (U) ⊗ Cc∞ (U) so that Vect
g c (C ∞ (U)) is a dense
∞
subspace of Vectc (C (U)).
2.0.2 Lemma. The divergence map
Div g c (C ∞ (U)) → P(C
g : Vect e ∞ (U))
extends continuously to a map
Div : Vectc (C ∞ (U)) → P(C ∞ (U)).
Proof. Suppose that
F(x1 , . . . , xn+1 ) ∈ Cc∞ (U n+1 )S n ⊂ Vectc (C ∞ (U)).
The divergence map in equation (‡) extends to a map that sends F to
n Z
X
−4 xn+1 F(x1 , . . . , xn+1 ) + F(x1 , . . . , xi , . . . , xn , xi ) dvolg .
i=1 xi ∈U
Here, 4 xn+1 denotes the Laplacian acting only on the n + 1st copy of U. Note that
the integral produces a function on U n−1 .
With these objects in hand, we are able to define the quantum observables of a
free field theory.
2.0.3 Definition. For an open subset U ⊂ M, let
H 0 (Obsq (U)) = P(C ∞ (U))/ Im Div .
In other words, H 0 (Obsq (U)) be the cokernel of the operator Div. Later we
will see that this linear map Div naturally extends to a cochain complex of quantum
observables, which we will denote Obsq , whose zeroth cohomology is what we just
defined. This extension is why we write H 0 .
Let us explain why we should interpret this space as the quantum observables.
We expect that an observable in a field theory is a function on the space of fields.
An observable on a field theory on an open subset U ⊂ M is a function on the fields
that only depends on the behaviour of the fields inside U. Speaking conceptually,
the expectation value of the observable is the integral of this function against the
“functional measure” on the space of fields.
Our approach is that we will not try to define the functional measure, but in-
stead we define the divergence operator. If we have some functional on C ∞ (U) that
is the divergence of a vector field, then the expectation value of the corresponding
observable is zero. Thus, we would expect that the observable given by a diver-
gence is not a physically interesting quantity, since its value is zero. Thus, we
might as well identify it with zero.
The appropriate vector fields on C ∞ (U) — the ones for which our divergence
operator makes sense — are vector fields along the foliation of C ∞ (U) by com-
pactly supported fields. Thus, the quotient of functions on C ∞ (U) by the subspace
of divergences of such vector fields gives a definition of observables.
3. The prefactorization structure on observables

Suppose that we have a Gaussian measure ωA on Rn . Then every function
on Rn with polynomial growth is integrable, and this space of functions forms a
commutative algebra. We showed that there is a short exact sequence
DivωA EωA
0 → Vect(Rn ) −→ P(Rn ) −→ R → 0,
where EωA denotes the expectation value map for this measure. But the image of
the divergence operator is not an ideal. (Indeed, usually an expectation value map
is not an algebra map!) This fact suggests that, in the BV formalism, the quantum
observables should not form a commutative algebra. One can check quickly that
for our definition above, H 0 (Obsq (U)) is not an algebra.
However, we will find that some shadow of this commutative algebra structure
exists, which allows us to combine observables on disjoint subsets. This residual
structure will give the spaces H 0 (Obsq (U)) of observables, viewed as a functor on
the category of open subsets U ⊂ M, the structure of a prefactorization algebra.
Let us make these statements precise. Note that P(C ∞ (U)) is a commutative
algebra, as it is a space of polynomial functions on C ∞ (U). Further, if U ⊂ V
there is a map of commutative algebras ext : P(C ∞ (U)) → P(C ∞ (V)), extending
a polynomial map F : C ∞ (U) → R to the polynomial map F ◦ res : C ∞ (V) → R
by precomposing with the restriction map res : C ∞ (V) → C ∞ (U). This map ext
is injective. We will sometimes refer to an element of the subspace P(C ∞ (U)) ⊂
P(C ∞ (V)) as an element of P(C ∞ (V)) with support in U.
3.0.1 Lemma. The product map
P(C ∞ (V)) ⊗ P(C ∞ (V)) → P(C ∞ (V))
does not descend to a map
H 0 (Obs(V)) ⊗ H 0 (Obs(V)) → H 0 (Obs(V)).
If U1 , U2 are disjoint open subsets of the open V ⊂ M, then we have a map
P(U1 ) ⊗ P(U2 ) → P(V)

3. THE PREFACTORIZATION STRUCTURE ON OBSERVABLES 27
obtained by combining the inclusion maps P(Ui ) ,→ P(V) with the product map on
P(V). This map does descend to a map
H 0 (Obs(U1 )) ⊗ H 0 (Obs(U2 )) → H 0 (Obs(V)).
In other words, although the product of general observables does not make
sense, the product of observables with disjoint support does.
Proof. Let U1 , U2 be disjoint open subsets of M, both contained in an open V.

Let us view the spaces Vectc (C ∞ (Ui )) and P(C ∞ (Ui )) as subspaces of Vectc (C ∞ (V))
and P(C ∞ (V)), respectively.
We denote by DivUi , the divergence operator on Ui , namely,
DivUi : Vectc (C ∞ (Ui )) → P(C ∞ (Ui )).
We use DivV to denote the divergence operator on V.
Our situation is then described by the following diagram:
q12
ker q12 ,→ P(U1 ) ⊗ P(U2 ) −→ H 0 (Obs(U1 )) ⊗ H 0 (Obs(U2 ))
↓
q
Im DivV ,→ P(C ∞ (V)) −→ H 0 (Obs(V))
The middle vertical arrow is multiplication map. We want to show there is a vertical
arrow on the right that makes a commuting square. It suffices to show that the
image of ker q12 in H 0 (Obs(V)) is zero.
Note that H 0 (Obs(U1 )) ⊗ H 0 (Obs(U2 )) is the cokernel of the map
P(C ∞ (U1 )) ⊗ Vectc (C ∞ (U2 )) ⊕ Vectc (C ∞ (U1 )) ⊗ P(C ∞ (U2 ))
1⊗DivU2 + DivU1 ⊗1
−−−−−−−−−−−−−−−→ P(C ∞ (U1 )) ⊗ P(C ∞ (U2 )).
Hence, ker q12 = Im 1 ⊗ DivU2 + DivU2 ⊗1 .

We will show that the image of 1 ⊗ DivU2 + DivU1 ⊗1 sits inside Im DivV . This
result ensures that we can produce the desired map.
Thus, it suffices to show that for any F is in P(C ∞ (U1 )) and X is in Vectc (C ∞ (U2 )),
e = F DivU2 (X). (The same argu-
e in Vectc (C ∞ (V)) such that DivV (X)
there exists X
ment applies after switching the roles U1 and U2 .) We will show, in fact, that
Div(FX) = F Div(X),
viewing F and X as living on the open V.
A priori, this assertion should be surprising. On an ordinary finite-dimensional
manifold, the divergence Div with respect to any volume form has the following
property: for any vector field X and any function f , we have
Div( f X) − f Div X = X( f ),
where X( f ) denotes the action of X on f . Note that X( f ) is not necessarily in the
image of Div, in which case f Div X is not in the image of Div, and so we see that
Im Div is not an ideal.
The same equation holds for the divergence operator we have defined in infinite
dimensions. If X ∈ Vectc (C ∞ (V)) and F ∈ P(Cc∞ (V)), then
Div(FX) − F DivX = X(F).
This computation tells us that the image of Div is not an ideal, as there exist X(F)
not in the image of Div.
When F is in P(C ∞ (U1 )) and X is in Vectc (C ∞ (U2 )), however, X(F) = 0, as
their supports are disjoint. Thus, we have precisely the desired relation F Div(X) =
Div(FX).
We now prove that X(F) = 0 when X and F have disjoint support.
We are working with polynomial functions and vector fields, so that all compu-
tations can be done in a purely algebraic fashion; in other words, we will work with
derivations as in algebraic geometry. Let ε satisfy ε2 = 0. For a polynomial vector
field X and a polynomial function F on a vector space V, we define the function
X(F) to assign to the vector v ∈ V, the ε component of F(v + εXv ) − F(v). Here Xv
denotes the tangent vector at v that X produces.
In our situation, we know that for any φ ∈ C ∞ (V), we have the following
properties:
• F(φ) only depends on the restriction of φ to U1 , and
• Xφ is a function with support in U2 and hence vanishes away from U2 .
Thus, F(φ + εX(φ)) = F(φ) as the restriction of φ + εX(φ) to U1 agrees with the
restriction of φ. We see then that
d
X(F)(φ) = F(φ + εXφ ) − F(φ) = 0,
dε
as asserted.
In a similar way, if U1 , . . . , Un are disjoint opens all contained in V, then there
is a map
H 0 (Obsq (U1 )) ⊗ · · · ⊗ H 0 (Obsq (Un )) → H 0 (Obsq (V))
descending from the map
P(U1 ) ⊗ · · · ⊗ P(Un ) → P(V)
given by inclusion followed by multiplication.
Thus, we see that the spaces H 0 (Obsq (U)) for open sets U ⊂ M are naturally
equipped with the structure maps necessary to define a prefactorization algebra.
(See Section 2 for a sketch of the definition of a factorization algebra, and Sec-
tion 1 for more details on the definition). It is straightforward to check, using the
arguments from the proof above, that these structure maps satisfy the necessary
compatibility conditions to define a prefactorization algebra.
4. From quantum to classical

Our general philosophy is that the quantum observables of a field theory are a
factorization algebra given by deforming the classical observables. The classical
observables are defined to be functions on the space of solutions to the equations
4. FROM QUANTUM TO CLASSICAL 29
of motion. We will now examine how our construction can be seen as just such a
deformation.
Let us see first why this holds for a class of measures on finite dimensional
vector spaces. Let S be a polynomial function on Rn . Let dn x denote the Lebesgue
measure on Rn , and consider the measure
ω = e−S /~ dn x,
where ~ is a small parameter. The divergence with respect to ω is given by the
formula
X ∂ ! 1 X ∂S X ∂ fi
Divω fi =− fi + .
∂xi ~ ∂xi ∂xi
As before, let P(Rn ) denote the space of polynomial functions on Rn and let Vect(Rn )
denote the space of polynomial vector fields. The divergence operator Divω is a lin-
ear map map Vect(Rn ) → P(Rn ). Note that the operators Divω and ~ Divω have the
same image so long as ~ , 0. When ~ = 0, the operator ~ Divω becomes the
operator
X ∂ X ∂S
fi 7→ − fi .
∂xi ∂xi
Therefore, the ~ → 0 limit of the image of Divω is the Jacobian ideal
∂S
!
Jac(S ) = ⊂ P(Rn ),
∂xi
which corresponds to the critical locus of S in Rn . Hence, the ~ → 0 limit of the ob-
servables P(Rn )/ Im Divω is the commutative algebra P(Rn )/ Jac(S ) that describes
functions on the critical locus of S .
Let us now check the analogous property for the observables of a free scalar
field theory on a manifold M. We will consider the divergence for the putative
Gaussian measure Z !
exp − 1~ φ(4 + m2 )φ dφ
M
on C ∞ (M). For any open subset U ⊂ M, this divergence operator gives us a map
Div~ : Vectc (C ∞ (U)) → P(C ∞ (U))
with
∂ X Z
f1 · · · fn 7 − ~1 f1 · · · fn (4 + m2 )φ +
→ f1 · · · b
fi f φ.
∂φ i M
Note how ~ appears in this formula; it is just the same modification of the operator
(‡) as Divω is of the divergence operator for the measure e−S dn x.
As in the finite dimensional case, the first term dominates in the ~ → 0 limit.
The ~ → 0 limit of the image of Div~ is the closed subspace of P(C ∞ (U)) spanned
by functionals of the form f1 · · · fn (4 + m2 )φ, where fi and φ are in Cc∞ (U)). This
subspace is the topological ideal in P(C ∞ (U)) generated by linear functionals of the
form (4 + m2 ) f , where f ∈ Cc∞ (U). If S (φ) = φ(4 + m2 )φ is the action functional
R
of our theory, then this subspace is precisely the topological ideal generated by all
the functional derivatives
∂S
( )
: φ ∈ Cc (U) .
∞
∂φ
In other words, it is the Jacobian ideal for S and hence is cut out by the Euler-
Lagrange equations. Let us call this ideal IEL (U).
A more precise statement of what we have just sketched is the following. De-
fine a prefactorization algebra H 0 (Obscl (U)) (the superscript cl stands for classi-
cal) that assigns to U the quotient algebra P(C ∞ (U))/IEL (U). Thus, H 0 (Obscl (U))
should be thought of as the polynomial functions on the space of solutions to the
Euler-Lagrange equations. Note that each constituent space H 0 (Obscl (U)) in this
prefactorization algebra has the structure of a commutative algebra, and the struc-
ture maps are all maps of commutative algebras. In short, H 0 (Obscl ) forms a com-
mutative prefactorization algebra. Heuristically, this terminology means that the
product map defining the factorization structure is defined for all pairs of opens
U1 , U2 ⊂ V, and not just disjoint pairs.
Our work in the section is summarized as follows.
q
4.0.1 Lemma. There is a prefactorization algebra H 0 (Obs~ ) over C[~] such that
when specialized to ~ = 1 is H 0 (Obsq ) and to ~ = 0 is H 0 (Obscl ).
This prefactorization algebra assigns to an open set the cokernel of the map
~ Div~ : Vectc (C ∞ (U))[~] → P(C ∞ (U))[~],
where Div~ is the map defined above.
q
We will see later that H 0 (Obs~ (U)) is free as an R[~]-module, although this is
a special property of free theories and is not always true for an interacting theory.
5. Correlation functions
We have seen that the observables of a free scalar field theory on a manifold M
give rise to a factorization algebra. In this section, we will explain how the structure
of a factorization algebra is enough to define correlation functions of observables.
We will calculate certain correlation functions explicitly and recover the standard
answers.
Suppose now that M is a compact Riemannian manifold, and, as before, let
us consider the observables of the free scalar field theory on M with mass m > 0.
Then we have the following result.
5.0.1 Lemma. If the mass m is positive, then H 0 (Obsq (M)) R.
Compare this result with the statement that for a Gaussian measure on Rn ,
the image of the divergence map is of codimension 1 in the space of polynomial
functions on Rn . The assumption here that the mass is positive is necessary to
ensure that the quadratic form M φ(4 + m2 )φ is non-degenerate.
R
This lemma will follow from our more detailed analysis of free theories in
Chapter 4, but we can sketch the idea here. The main point is that there is a family
5. CORRELATION FUNCTIONS 31
of operators over R[~] connecting H 0 (Obsq (M)) to H 0 (Obscl (M)). It is straight-

forward to see that the algebra H 0 (Obscl (M)) is R. Indeed, observe that the only
solution to the equations of motion in the case that m > 0 is the function φ = 0,
since the assumption that 4 has nonnegative spectrum ensures that 4φ = −m2 φ
has no nontrivial solution. Functions on a point are precisely R. To conclude that
q
H 0 (Obsq (M)) is also R, we need to show that H 0 (Obs~ (M)) is flat over R[~], which
will follow from a spectral sequence computation we will perform later in the book.
There is always a canonical observable 1 ∈ H 0 (Obsq (U)) for any open subset
U ⊂ M. This element is defined to be the image of the function 1 ∈ P(C ∞ (U)). We
identify H 0 (Obsq (M)) with R by taking this observable 1 to be a basis vector.
5.0.2 Definition. Let U1 , . . . , Un ⊂ M be disjoint open subsets. The correlator is
the prefactorization structure map
h−i : H 0 (Obsq (U1 )) ⊗ · · · ⊗ H 0 (Obsq (Un )) → H 0 (Obsq (M)) R
We should compare this definition with what happens in finite dimensions. If

we have a Gaussian measure on Rn , then the space of polynomial functions modulo
divergences is one-dimensional. If we take the image of a function 1 to be a basis
of this space, then we get a map
P(Rn ) → R
from polynomial functions to R. This map is the integral against the Gaussian
measure, normalized so that the integral of the function 1 is 1.
In our infinite dimensional situation, we are doing something very similar. Any
reasonable definition of the correlation function of the observables O1 , . . . , On ,
with Oi in P(C ∞ (Ui )), should only depend on the product function O1 · · · On ∈
P(C ∞ (M)). Thus, the correlation function map should be a linear map P(C ∞ (M)) →
R. Further, it should send divergences to zero. We have seen that there is only one
such map, up to an overall scale.
5.1. Comparison to physics. Next we will check explicitly that this correla-
tion function map really matches up with what physicists expect. Let fi ∈ Cc∞ (Ui )
be compactly-supported smooth functions on the open sets U1 , U2 ⊂ M. Let us
view each fi as a linear function on C ∞ (Ui ), and so as an element of the polyno-
mial functions P(C ∞ (Ui )).
Let G ∈ D(M × M) be the unique distribution on M × M with the property that
(4 x + m2 )G(x, y) = δDiag .
Here δDiag denotes the delta function supported on the diagonal copy of M inside
M × M. In other words, if we apply the operator 4 + m2 to the first factor of G,
we find the delta function on the diagonal. Thus, G is the kernel for the operator
(4 + m2 )−1 . In the physics literature, G is called the propagator; in the mathematics
literature, it is called the Green’s function for the operator 4 + m2 . Note that G is
smooth away from the diagonal.
5.1.1 Lemma. Given fi ∈ Cc∞ (Ui ), there are classes [ fi ] ∈ H 0 (Obsq (Ui )). Then
Z
h[ f1 ] [ f2 ]i = f1 (x)G(x, y) f2 (y).
x,y∈M
Note that this result is exactly the standard result in physics.

Proof. Later we will give a slicker and more general proof of this kind of
statement. Here we will give a simple proof to illustrate how the Green’s function
arises from our homological approach to defining functional integrals.
The operator 4 + m2 is surjective. Thus, there is a preimage
φ = (4 + m2 )−1 f2 ∈ Cc∞ (M),
which is unique because 4 + m2 is also injective. Indeed, we know that
Z
φ(x) = G(x, y) f2 (y).
y∈M
Now consider the vector field
∂
f1 ∈ Vectc (C ∞ (M)).
∂φ
Note that
∂
! Z
Div f1 = f1 (x)φ(x) − f1 (4 + m2 )φ dvol
∂φ x∈M
Z !
= f1 (x)G(x, y) f2 (y) · 1 − f1 f2 .
M
The element f1 f2 ∈ P(C ∞ (M)) is a cocycle representing the factorization product
[ f1 ][ f2 ] in H 0 (Obsq (M)) of the observables [ fi ] ∈ H 0 (Obsq (Ui )). The displayed
equation tells us that
Z !
[ f1 f2 ] = f1 (x)G(x, y) f2 (y) · 1 ∈ H 0 (Obsq (M)).
M
Since the observable 1 is chosen to be the basis element identifying H 0 (Obsq (M))
with R, the result follows.
With a little more work, the same arguments recover the usual Wick’s formula
for correlators of the form h f1 · · · fn i.
Remark: These kinds of formulas are standard knowledge in physics, but not in
mathematics. For a more extensive discussion in a mathematical style, see Chapter
6 of Glimm and Jaffe (1987) or Lecture 3 by Kazhdan in Deligne et al. (1999). ^
6. Further results on free field theories

In this chapter, we showed that if we define the observables of a free field the-
ory as the cokernel of a certain divergence operator, then these spaces of observ-
ables form a prefactorization algebra. We also showed that this prefactorization
algebra contains enough information to allow us to define the correlation functions
7. INTERACTING THEORIES 33
of observables, and that for linear observables we find the same formula that physi-
cists would write.
In Chapter 3 we will show that a certain class of factorization algebras on
the real line are equivalent to associative algebras, together with a derivation. As
Noether’s theorem suggest, the derivation arises from infinitesimal translation on
the real line, so that it encodes the Hamiltonian of the system.
In Chapter 4, we will analyze the factorization algebra of free field theories
in more detail. We will show that if we consider the free field theory on R, the
factorization algebra H 0 (Obsq ) corresponds (under the relationship between fac-
torization algebras on R and associative algebras) to the Weyl algebra. The Weyl
algebra is generated by observables p, q corresponding to position and momentum
satisfying [p, q] = 1. If we consider instead the family over R[~] of factoriza-
q
tion algebras H 0 (Obs~ ) discussed above, then we find the commutation relation
[p, q] = ~ of Heisenberg. This algebra, of course, is what is traditionally called the
algebra of observables of quantum mechanics. In this case, we will further see that
the derivation of this algebra (corresponding to infinitesimal time translation) is an
inner derivation, given by bracketing with the Hamiltonian
H = p2 − m2 q2 ,
which is the standard Hamiltonian for the quantum mechanics of the harmonic
oscillator.
More generally, we recover canonical quantization of the free scalar field the-
ory on higher dimensional manifolds as follows. Consider a free scalar theory
on the product Riemannian manifold N × R, where N is a compact Riemannian
manifold. This example gives rise to a factorization algebra on R that assigns to
an open subset U ⊂ R, the space H 0 (Obsq (N × U)). We will see that this fac-
torization algebra on R has a dense sub-factorization algebra corresponding to an
associative algebra. This associative algebra is the tensor product of Weyl alge-
bras, where each each eigenspace of the operator 4 + m2 on C ∞ (N) produces a
Weyl algebra. In other words, we find quantum mechanics on R with values in the
infinite-dimensional vector space C ∞ (N). Since that space has a natural spectral
decomposition for the operator 4 + m2 , it is natural to interpret as the algebra of
observables, the associative algebra given by tensoring together the Weyl algebra
for each eigenspace. This result is entirely consistent with standard arguments in
physics, being the factorization algebra analog of canonical quantization.
7. Interacting theories
In any approach to quantum field theory, free field theories are easy to con-
struct. The challenge is always to construct interacting theories. The core results
of this two-volume work show how to construct the factorization algebra corre-
sponding to interacting field theories, deforming the factorization algebra for free
field theories discussed above.
Let us explain a little bit about the challenges we need to overcome in order to
deal with interacting theories, and how we overcome these challenges.
Consider an interacting scalar field theory on a Riemannian manifold M. For

instance, we could consider an action functional of the form
Z Z
S (φ) = − 2
1
φ(4 + m )φ +
2
φ4 .
M M
In general, the action functional must be local: it must arise as the integral over M
of some polynomial in φ and its derivatives. This condition is due to our interest in
field theories, but it is also necessary to produce factorization algebras.
We will let I(φ) denote the interaction term in our field theory, which consists
of the cubic and higher terms in S . In the above example, I(φ) =R φ4 . We will
R
always assume that the quadratic term in S is similar in form − 12 M φ(4 + m2 )φ,
i.e., we require an ellipticity condition. (Of course, the examples amenable to our
techniques apply to a very general class of interacting theories, including many
gauge theories.)
If U ⊂ M is an open subset, we can consider, as before, the spaces Vectc (C ∞ (M))
and P(C ∞ (M)) of polynomial functions and vector fields on M. By analogy with
the finite-dimensional situation, one can try to define the divergence for the putative
measure exp S (φ)/~dµ (where dµ refers to the “Lebesgue measure” on C ∞ (M))) by
the formula
∂ ∂S X
! Z
Div~ f1 . . . fn = ~ f1 . . . fn
1
+ f1 . . . fi . . . fn
b fi φ.
∂φ ∂φ M
This formula agrees with the formula we used when S was purely quadratic.
But now a problem arises. We defined P(C ∞ (U)) as the space of polynomial
functions whose Taylor terms are given by integration against a smooth function
on U n . That is, M
P(C ∞ (U)) = Cc∞ (U n )S n .
n
∂S
Unfortunately, if φ ∈ Cc (U)), then ∂φ is not necessarily in this space of
∞ functions.
For instance, if I(φ) = φ4 is the interaction term in the example of the φ4 theory,
R
then
∂I
Z
(ψ) = φψ3 dvol.
∂φ M
∂I
Thus ∂φ provides a cubic function on the space C ∞ (U), but it is not given by inte-
gration against an element in Cc∞ (U 3 ). Instead it is given by integrating against a
distribution on U 3 , namely, the delta-distribution on the diagonal.
We can try to solve this issue by using a larger class of polynomial functions.
Thus, we could let M
P(C ∞ (U)) = Dc (U n )S n
n
where Dc (U n ) is the space of compactly supported distributions on U. Similarly,
we could let M
Vectc (U) = Dc (U n+1 )S n .
n
The spaces P(C ∞ (U)) and Vectc (C ∞ (U)) are dense subspaces of these spaces.
7. INTERACTING THEORIES 35
If φ0 ∈ Dc (U) is a compactly-supported distribution, then for any local func-

tional S , ∂S ∞
∂φ is a well-defined element of P(C (U)). Thus, it looks like we have
resolved our problem.
However, using this larger space of functions gives us a new problem: the
second term in the divergence operator now fails to be well-defined! For example,
if f, φ ∈ Dc (U), then we have
∂ ∂S
! Z
Div~ f = 1~ f + f φ.
∂φ ∂φ M
Now, M f φ no longer make sense, because we are trying to multiply distributions.

R
More explicitly, this term involves pairing the distribution f φ on the diagonal in
M 2 with the delta-function on the diagonal.
If we consider the ~ → 0 limit of this putative operator ~ Div~ , we nonetheless
find a well-defined operator
Vectc (C ∞ (U)) → P(C ∞ (U))

X → XS ,
which sends a vector field X to its action on the local functional S .
The cokernel of this operator is the quotient of P(C ∞ (U)) by the ideal IEL (U)
generated by the Euler-Lagrange equations. We thus let
cl
H 0 (ObsS (U)) = P(C ∞ (U))/IEL .
As U varies, this construction produces a factorization algebra on M, which we call
the factorization algebra of classical observables associated to the action functional
S.
7.0.1 Lemma. If S = − 21 φ(4+m2 )φ, then this definition of classical observables
R
coincides with the one we discussed earlier:

cl
H 0 (ObsS (U)) H 0 (Obscl (U))
where H 0 (Obscl (U)) is defined, as earlier, to be the quotient of P(C ∞ (U)) by the
ideal IEL (U) of the Euler-Lagrange equations.
This result is a version of elliptic regularity, and we prove it later, in Chapter 4.
Now the challenge we face should be clear. If S is the action functional for
the free field theory, then we have a factorization algebra of classical observables.
This factorization algebra deforms in two ways. First, we can deform it into the
factorization algebra of quantum observables for a free theory. Second, we can
deform it into the factorization algebra of classical observables for an interacting
field theory. The difficulty is to perform both of these deformations simultaneously.
To construct the observables of an interacting field theory, we use the renormal-
ization technique of Costello (2011b). In Costello (2011b), the first author gives a
definition of a quantum field theory and a cohomological method for constructing
field theories. A field theory as defined in Costello (2011b) gives us (essentially
from the definition) a family of divergence operators

Div[L] : Vectc (M) → P(C ∞ (M)),
one for every L > 0. These divergence operators, for varying L, can be conjugated
to each other by continuous linear isomorphisms of Vectc (M)) and P(C ∞ (M)).
However, for a proper open subset U, these divergence operators do not map
Vectc (U)) into P(C ∞ (U)). However, roughly speaking, for L very small, the op-
erator Div[L] only increases the support of a vector field in Vectc (U)) by a small
amount outside of U. For small enough L, it is almost support-preserving. This
property turns out to be enough to define the factorization algebra of quantum ob-
servables.
This construction of quantum observables for an interacting field theory is
given in Volume 2, which is the most technically difficult part of this work. Be-
fore tackling it, we will develop more language and explore more examples. In
particular, we will
• develop some formal and structural aspects of the theory of factoriza-
tion algebras;
• analyze in more detail the factorization algebra associated to a free
theory;
• construct and analyze factorization algebras associated to vertex alge-
bras such as the Kac-Moody vertex algebra;
• develop classical field theory using a homological approach arising
from the BV formalism; and
• flesh out the description of the factorization algebra of classical ob-
servables we have sketched here.
The example of this chapter, however, already exhibits the central ideas.
CHAPTER 3
Prefactorization algebras and basic examples
In this chapter we will give a formal definition of the notion of prefactorization

algebra. With the definition in hand, we proceed to examine several examples that
arise naturally in mathematics. In particular, we explain how associative algebras
can be viewed as prefactorization algebras on the real line, and when the converse
holds.
We also explain how to construct a prefactorization algebra from a sheaf of Lie
algebras on a manifold M. This construction is called the factorization envelope,
and it is related to the universal enveloping algebra of a Lie algebra as well as
to Beilinson-Drinfeld’s notion of a chiral envelope. Although the factorization
envelope construction is very simple, it plays an important role in field theory. For
example, the factorization algebra for any free theories is a factorization envelope,
as is the factorization algebra corresponding to the Kac-Moody vertex algebra.
More generally, factorization envelopes play an important role in our formulation
of Noether’s theorem for quantum field theories.
Finally, when the manifold M is equipped with an action of a group G, we
describe what a G-equivariant prefactorization algebra is. We will use this notion
later in studying translation-invariant field theories (see Section 8) and holomor-
phically translation-invariant field theories (see Chapter 5).
1. Prefactorization algebras
In this section we will give a formal definition of the notion of a prefactor-
ization algebra, starting concretely and then generalizing. In the first subsection,
using plain language, we describe a prefactorization algebra taking values in vec-
tor spaces. The reader is free to generalize by replacing “vector space” and “linear
map” with “object of a symmetric monoidal category C” and “morphism in C.”
(Our favorite target category is cochain complexes.) The next subsections give a
concise definition using the language of multicategories (also known as colored op-
erads) and allow an arbitrary multicategory as the target. In the final subsections,
we describe the category (and multicategory) of such prefactorization algebras.
1.1. The definition in explicit terms. Let M be a topological space. A pref-

actorization algebra F on M, taking values in vector spaces, is a rule that assigns
a vector space F (U) to each open set U ⊂ M along with the following maps and
compatibilities.
• There is a linear map mU
V : F (U) → F (V) for each inclusion U ⊂ V.
37
38 3. PREFACTORIZATION ALGEBRAS AND BASIC EXAMPLES
1 ,...,U n
• There is a linear map mU V : F (U1 ) ⊗ · · · ⊗ F (Un ) → F (V) for
every finite collection of open sets where each Ui ⊂ V and the Ui are
pairwise disjoint. The following picture represents the situation.
V
U1
1 ,...,U n
mU
V : F (U1 ) ⊗ · · · ⊗ F (Un ) → F (V)
U2
... Un
• The maps are compatible in the obvious way, so that if Ui,1 t· · ·tUi,ni ⊆
Vi and V1 t · · · t Vk ⊆ W, the following diagram commutes.
Nk Nni Nk
i=1 j=1 F (U j ) i=1 F (Vi )
F (W)
Thus F resembles a precosheaf, except that we tensor the vector spaces rather than
take their direct sum.
For an explicit example of the associativity, consider the following picture.
V1 U1,2 W
F (U1,1 ) ⊗ F (U1,2 ) ⊗ F (U2,1 )
U1,1
V2
U2,1 F (V1 ) ⊗ F (V2 ) F (W)
The case of k = n1 = 2, n2 = 1
These axioms imply that F (∅) is a commutative algebra. We say that F is
a unital prefactorization algebra if F (∅) is a unital commutative algebra. In this
case, F (U) is a pointed vector space by the image of the unit 1 ∈ F (∅) under the
structure map F (∅) → F (U). In practice, for our examples, F (∅) is C, R, C[[~]],
or R[[~]].
Example: The crucial example to bear in mind is an associative algebra. Every
associative algebra A defines a prefactorization algebra A f act on R, as follows.
` To
each open interval (a, b), we set A f act ((a, b)) = N
A. To any open set U = j I j,
where each I j is an open interval, we set F (U) = j A. The structure maps simply
arise from the multiplication map for A. Figure 1 displays the structure of A f act .
Notice the resemblance to the notion of an E1 or A∞ algebra. (One takes an infinite
1. PREFACTORIZATION ALGEBRAS 39
a⊗b⊗c ∈ A⊗A⊗A
ab ⊗ c ∈ A⊗A
abc ∈ A
Figure 1. The prefactorization algebra A f act of an associative al-

gebra A
tensor products of unital algebras, as follows. Given an infinite set I, consider the
Nby inclusion. For0each finite subset J J⊂ I, we
poset of finite subsets of I, ordered
J0
can take the tensor product A J = j∈J A. For J ,→ J , we define a map A → A
by tensoring with the identity 1 ∈ A for every j ∈ J 0 \J. Then AI is the colimit over
this poset.) ^
Example: Another important example for us is the symmetric algebra of a pre-

cosheaf. Let F be a precosheaf of vector spaces on a space X. (For example,
consider F = Cc∞ the compactly supported smooth functions on a manifold.) The
functor F = Sym F : U 7→ Sym(F(U)) defines a precosheaf of commutative al-
gebras, but it also a prefactorization algebra. For instance, let U and V be disjoint
opens. The structure maps F(U) → F(U t V) and F(U) → F(U t V) induce a
canonical map
F(U) ⊕ F(V) → F(U t V),
and so we obtain a natural map Sym(F(U) ⊕ F(V)) → Sym F(U t V). But
F (U) ⊗ F (V) Sym(F(U)) ⊗ Sym(F(V)) Sym(F(U) ⊕ F(V)),
so there is a natural map
F (U) ⊗ F (V) → F (U t V).
In a similar way, one can provide all the structure maps to make F a prefactoriza-
tion algebra. ^
In the remainder of this section, we describe two other ways of phrasing this
idea, but the reader who is content with this definition and eager to see examples
should feel free to jump ahead, referring back as needed.
1.2. Prefactorization algebras as algebras over an operad. We now pro-

vide a succinct and general definition of a prefactorization algebra using the effi-
cient language of multicategories. (See Appendix 2.3 for a quick overview of the
notion of a multicategory, also known as a colored operad. Note that we mean the
symmetric version of such definitions.)
1.2.1 Definition. Let Disj M denote the following multicategory associated to M.
• The objects consist of all connected open subsets of M.
• For every (possibly empty) finite collection of open sets {Uα }α∈A and
open set V, there is a set of maps Disj M ({Uα }α∈A | V). If the Uα are
pairwise disjoint and all are contained in V, then the set of maps is a
single point. Otherwise, the set of maps is empty.
• The composition of maps is defined in the obvious way.
A prefactorization algebra just is an algebra over this colored operad Disj M .

1.2.2 Definition. Let C be a multicategory. A prefactorization algebra on M taking
values in C is a functor (of multicategories) from Disj M to C.
Since symmetric monoidal categories are special kinds of multicategories, this

definition makes sense for C any symmetric monoidal category.
Remark: If C is a symmetric monoidal category under coproduct, then a precosheaf
on M with values in C defines a prefactorization algebra valued in C. Hence,
our definition broadens the idea of “inclusion of open sets leads to inclusion of
sections” by allowing more general monoidal structures to “combine” the sections
on disjoint open sets. ^
Note that if F is any prefactorization algebra, then F (∅) is a commutative
algebra object of C.
1.2.3 Definition. We say a prefactorization algebra F is unital if the commutative
algebra F (∅) is unital.
Remark: There is an important variation on this definition where one weakens the
requirement that the composition of structure maps holds “on the nose” and in-
stead requires homotopy coherence. For example, given disjoint opens U1 and U2
1 ,U 2
contained in V, which is then contained in W, we do not require that mU W =
V U1 ,U2
mW ◦ mV but that there is a “homotopy” between these maps. This kind of
situation arises naturally whenever the target category is best viewed as an ∞-
category, such as the category of cochain complexes. We will not develop here the
formalism necessary to treat homotopy-coherent prefactorization algebras because
our examples and constructions always satisfy the strictest version of composition.
The reader interested in seeing this variant developed should see the treatment in
Lurie (n.d.b). (We remark that one typically has “strictification” results that en-
sure that a homotopy-coherent algebra over a colored operad can be replaced by a
weakly-equivalent strict algebra over a colored operad, so that working with strict
algebras is sufficient for many purposes.) ^
1.3. Prefactorization algebras in the style of precosheaves. Any multicate-
gory C has an associated symmetric monoidal category SC, which is defined to be
the universal symmetric monoidal category equipped with a functor of multicate-
gories C → SC. We call it the symmetric monoidal envelope of C. Concretely, an
object of SC is a formal tensor product a1 ⊗ · · · ⊗ an of objects of C. Morphisms
in SC are characterized by the property that for any object b in C, the set of maps
1. PREFACTORIZATION ALGEBRAS 41
SC(a1 ⊗ · · · ⊗ an , b) in the symmetric monoidal category is exactly the set of maps

C({a1 , . . . , an } | b) in the multicategory C.
We can give an alternative definition of prefactorization algebra by working
with the symmetric monoidal category S Disj M rather than the multicategory Disj M .
1.3.1 Definition. Let S Disj M denote the following symmetric monoidal category.
• An object of S Disj M is a formal finite sequence [V1 , . . . , Vm ] of opens
Vi in M.
• A morphism F : [V1 , . . . , Vm ] → [W1 , . . . , Wn ] consists of a surjective
function φ : {1, . . . , m} → {1, . . . , n} and a morphism f j ∈ Disj M ({Vk }k∈φ−1 ( j) |W j )
for each 1 ≤ j ≤ n.
• The symmetric monoidal structure on S Disj M is given by concatenta-
tion.
The alternative definition of prefactorization algebra is as follows. It resembles

the notion of a precosheaf (i.e., a functor out of some category of opens) with the
extra condition that it is symmetric monoidal.
1.3.2 Definition. A prefactorization algebra with values in a symmetric monoidal
category C ⊗ is a symmetric monoidal functor S Disj M → C .
Remark: Although “algebra” appears in its name, a prefactorization algebra only

allows one to “multiply” elements that live on disjoint open sets. The category
of prefactorization algebras (taking values in some fixed target category) has a
symmetric monoidal product, so we can study commutative algebra objects in that
category. As an example, we will consider the observables for a classical field
theory. ^
1.4. Morphisms and the category structure. We now explain how prefac-
torization algebras form a category.
1.4.1 Definition. A morphism of prefactorization algebras φ : F → G consists of
a map φU : F(U) → G(U) for each open U ⊂ M, compatible with the structure
maps. That is, for any open V and any finite collection U1 , . . . , Uk of pairwise
disjoint open sets, each contained in V, the following diagram commutes:
φU1 ⊗···⊗φUk
F(U1 ) ⊗ · · · ⊗ F(Uk ) −−−−−−−−−→ G(U1 ) ⊗ · · · ⊗ G(Uk )
↓ ↓ .
φV
F(V) −−−−−−−→ G(V)
Likewise, all the obvious associativity relations are respected.
Remark: When our prefactorization algebras take values in cochain complexes, we

require the φU to be cochain maps, i.e., they each have degree 0 and commute with
the differentials. ^
1.4.2 Definition. On a space X, we denote the category of prefactorization alge-
bras on X taking values in the multicategory C by PreFA(X, C).
Remark: When the target multicategory C is a model category or dg category or

some other kind of higher category, the category of prefactorization algebras natu-
rally forms a higher category as well. Given the nature of our constructions and ex-
amples in the next few chapters, such aspects do not play a prominent role. When
we define factorization algebras in Chapter 6, however, we will discuss such is-
sues. ^
1.5. The multicategory structure. Let SC denote the enveloping symmetric
monoidal category of the multicategory C (see Section 2.3). Let ⊗ denote the
symmetric monoidal product in SC. Any prefactorization algebra valued in C gives
rise to one valued in SC.
There is a natural tensor product on PreFA(X, SC), as follows. Let F, G be
prefactorization algebras. We define F ⊗ G by
(F ⊗ G)(U) = F(U) ⊗ G(U),
and we simply define the structure maps as the tensor product of the structure maps.
For instance, if U ⊂ V, then the structure map is
V ⊗ m(G)V : (F ⊗ G)(U) = F(U) ⊗ G(U) → F(V) ⊗ G(V) = (F ⊗ G)(V).

m(F)U U
1.5.1 Definition. Let PreFAmc (X, C) denote the multicategory arising from the sym-
metric monoidal product on PreFA(X, SC). That is, if F1 , . . . Fn , G are prefactor-
ization algebras valued in C, we define the set of multi-morphisms to be
PreFAmc (F1 , · · · , Fn | G) = PreFA(F1 ⊗ · · · ⊗ Fn , G),
the set of maps of SC-valued prefactorization algebras from F1 ⊗ · · · ⊗ Fn to G.
2. Associative algebras from prefactorization algebras on R

We explained above how an associative algebra provides a prefactorization
algebra on the real line. There are, however, prefactorization algebras on R that do
not come from associative algebras. Here we will characterize those that do arise
from associative algebras.
2.0.1 Definition. Let F be a prefactorization algebra on R taking values in the
category of vector spaces (without any grading). We say F is locally constant if
the map F (U) → F (V) is an isomorphism for every inclusion of intervals U ⊂ V.
2.0.2 Lemma. Let F be a locally constant, unital prefactorization algebra on R

taking values in vector spaces. Let A = F (R). Then A has a natural structure of
an associative algebra.
Remark: Recall that F being unital means that the commutative algebra F (∅) is
equipped with a unit. We will find that A is an associative algebra over F (∅). ^
Proof. For any interval (a, b) ⊂ R, the map
F ((a, b)) → F (R) = A

3. MODULES AS DEFECTS 43
is an isomorphism. Thus, we have a canonical isomorphism

A = F ((a, b))
for all intervals (a, b).
Notice that if (a, b) ⊂ (c, d) then the diagram
A
/ F ((a, b))
Id i(a,b)
(c,d)

A
/ F ((c, d))
commutes.
The product map m : A ⊗ A → A is defined as follows. Let a < b < c < d.
Then, the prefactorization structure on F gives a map
F ((a, b)) ⊗ F ((c, d)) → F ((a, d)),
and so, after identifying F ((a, b)), F ((c, d)) and F ((a, d)) with A, we get a map
A ⊗ A → A.
This is the multiplication in our algebra.
It remains to check the following.
((i)) This multiplication doesn’t depend on the intervals (a, b) t (c, d) ⊂
(a, d) we chose, as long as (a, b) < (c, d).
((ii)) This multiplication is associative and unital.
This is an easy (and instructive) exercise.
3. Modules as defects
We want to explain another simple but illuminating class of examples, and then
we apply this perspective in the context of quantum mechanics. We will work with
prefactorization algebras taking values in vector spaces with the tensor product as
symmetric monoidal structure.
3.1. Modules as living at points. We described already how to associate a

prefactorization algebra FA on R to an associative algebra A. It is easy to modify
this construction to describe bimodules, as follows. Let A and B be associative
algebras and M an A − B-bimodule, i.e., M is a left A-module and a right B-module
and these structures are compatible in that (am)b = a(mb) for all a ∈ A, b ∈ B, and
m ∈ M. We now construct a prefactorization algebra F M on R that encodes this
bimodule structure.
Pick a point p ∈ R. On the half-line {x ∈ R | x < p}, F M is given by FA :
to an interval I = (t0 , t1 ) with t1 < p, F M (I) = A, and the structure map for
inclusion of finitely many disjoint intervals into a bigger interval is determined by
multiplication in A. Likewise, on the half-line {x ∈ R | p < x}, F M is given by
FB . Intervals containing p are determined, though, by M. When we consider an
a⊗m⊗b ∈ A⊗M⊗B
am ⊗ b ∈ M⊗B
amb ∈ M
Figure 2. The prefactorization algebra F M for the A − B-bimodule M
interval I = (t0 , t1 ) such that t0 < p < t1 , we set F M (I) = M. The structure maps
are also determined by the bimodule structure. For example, given
s0 < s1 < t0 < p < t1 < u0 < u1 ,
we have
F M ((s0 , s1 ) t (t0 , t1 ) t (u0 , u1 )) = F M ((s0 , s1 )) ⊗ F M ((t0 , t1 )) ⊗ F M (u0 , u1 ))
= A⊗M⊗B
and the inclusion of these three intervals into (s0 , u1 ) is the map
F M ((s0 , s1 )) ⊗ F M ((t0 , t1 )) ⊗ F M )(u0 , u1 )) → F M ((s0 , u1 ))
.
a⊗m⊗b 7→ amb
The definition of a bimodule ensures that we have a prefactorization algebra. See
Figure 2.
There is a structure map that we have not discussed yet, though. The inclusion
of the empty set into an interval I containing p means that we need to pick an
element mI of M for each interval. The simplest case is to fix one element m ∈
M and simply use it for every interval. If we are assigning the unit of A as the
distinguished element for F M on every interval to the left of p and the unit of B for
every interval to the right of p, then the distinguished elements, then the structure
maps we have given clearly respect these distinguished elements.
These distinguished elements, however, can change with the intervals, so long
as they are preserved by the structure maps. For an interesting example, see the
discussion of quantum mechanics below.
Let us examine one more interesting case. Suppose we have algebras A, B, and
C, and an A − B-bimodule M and a B − C-bimodule N. There is a prefactorization
algebra on R describing the natural algebra for this situation.
Fix points p < q. Let F M,N be the prefactorization algebra on R such that
• on {x ∈ R | x < p}, it agrees with FA ,
• on {x ∈ R | p < x < q}, it agrees with FB ,
• on {x ∈ R | q < x}, it agrees with FC ,
and
• to intervals (t0 , t1 ) with t0 < p < t1 < q, it assigns M, and
• to intervals (t0 , t1 ) with p < t0 < q < t1 , it assigns N.
In other words, on {x ∈ R | x < q}, this prefactorization algebra F M,N behaves like
F M , and like FN on {x ∈ R | p < x}.
We still need to describe what it does on an interval I of the form (T 0 , T 1 ) with
T 0 < p < q < T 1 . There is a natural choice, dictated by the requirement that we
produce a prefactorization algebra.
We know that F M,N (I) must receive maps from M and N, by considering
smaller intervals that only overlap either p or q. It also receives maps from A,
B, and C from intervals not hitting these marked points, but these factor through
intervals containing one of the marked points. Finally, we must have a structure
map
µ : M ⊗ N → F M,N (I)
for each pair of disjoint intervals hitting both marked points.
Note, in particular, what the associativity condition requires in the situation
where we have three disjoint intervals given by
s0 < p < s1 < t0 < t1 < u0 < q < u1 ,
contained in I. We can factor the inclusion of these three intervals through the pair
of intervals (s0 , t1 ) t (u0 , u1 ) or the pair of intervals (s0 , s1 ) t (t0 , u1 ) Thus, our
structure map
M ⊗ B ⊗ N → F M,N (I)
must satisfy that µ((mb)⊗n) = µ(m⊗(bn)) for every m ∈ M, b ∈ B, and n ∈ N. This
condition means that F M,N (I) receives a canonical map from the tensor product
M ⊗B N.
Hence, the most natural choice is to set F M,N (I) = M ⊗B N. One can make
other choices for how to extend to these longer intervals, but such a prefactorization
algebra will receive a map from this one. The local-to-global principle satisfied by
a factorization algebra is motivated by this kind of reasoning.
Remark: We have shown how thinking about prefactorization algebras on a real
line “decorated” with points (i.e., with a kind of stratification) reflects familiar
algebraic objects like bimodules. By moving into higher dimensions and allow-
ing more interesting submanifolds and stratifications, one generalizes this familiar
algebra into new, largely-unexplored directions. See Ayala et al. (n.d.a) for an ex-
tensive development of these ideas in the setting of locally constant factorization
algebras. ^
3.2. Standard quantum mechanics as a prefactorization algebra. We will
now explain how to express the standard formalism of quantum mechanics in
the language of prefactorization algebras, using the kind of construction just de-
scribed. As our goal is to emphasize the formal structure, we will work with a
finite-dimensional complex Hilbert space and avoid discussions of functional anal-
ysis.
Remark: In a sense, this section is a digression from the central theme of the book.
Throughout this book we take the path integral formalism as fundamental, and
hence we do not focus on the Hamiltonian, or operator, approach to quantum
physics. Hopefully, juxtaposed with our work in Section 3, this example clarifies
how to connect our methods with others. ^
Let V denote a finite-dimensional complex Hilbert space V. That is, there is a
nondegenerate symmetric sesquilinear form (−, −) : V × V → C so that
(λv, λ0 v0 ) = λλ0 (v, v0 )
where λ, λ0 are complex numbers and v, v0 are vectors in V. Let A = End(V)
T
denote the algebra of endomorphisms, which has a ∗-structure via M ∗ = M , the
conjugate-transpose. The space V is a representation of A, and the ∗-structure is
characterized by the property that
(M ∗ v, v0 ) = (v, Mv0 ).
It should be clear that one could work more generally with a Hilbert space equipped
with the action of a ∗-algebra of operators, aka observables.
Now that we have fixed the kinematics of the situation, let’s turn to the dynam-
ics. Let (Ut )t∈R be a one-parameter group of unitary operators on V. Since we are
in the finite-dimensional setting, there is no problem identifying
Ut = eitH
for some Hermitian operator H that we call the Hamiltonian. We view V as a state
space for our system, A as where the observables live, and H as determining the
time evolution of our system.
We now rephrase this structure to make it easier to articulate via the factoriza-
tion picture. Let V denote V with the conjugate complex structure. We will denote
elements of V by “kets” | v0 i and elements of V by “bras” hv | , and we provide a
bilinear pairing between them by
v | v0 = (v, v0 ).
We equip V with the right A-module structure by
hv | M = hM ∗ v |
We will write hv | M | v0 i as can think of M acting on v0 from the left or on v from
the right and it will produce the same number.
Our goal is to describe a scattering-type experiment.
• At time t = 0, we prepare our system in the initial state hvin | .
• We modify the governing Hamiltonian over some finite time interval
(i.e., apply an operator, or equivalently, an observable).
• At time t = T , we measure whether our system is in the final state
| vout i.
If we run this experiment many times, with the same initial and final states and the
same operator, we should find a statistical pattern in our data. If an operator O acts
during a time interval (t, t0 ), then we are trying to compute the number
0
hvin | eitH Oei(T −t )H | vout i.
Our formalism does not include the idealized situation of an operator O acting at
a single moment t0 in time, but in the appropriate limit of shorter and shorter time
intervals around t0 , we would compute
hvin | eit0 H Oei(T −t0 )H | vout i,
which agrees with the usual prescription.
Note that we use a bra hvin | for the “incoming” state and a ket | vout i for
the“outgoing” state so that the left-to-right ordering will agree with the left-to-
right ordering of the real line viewed as a time-line. The prefactorization algebra F
on the interval [0, T ] describing this situation has the following structure. Interior
open intervals describe moments when operators can act on our system. An interval
that contains 0 (but not the other end) should describe possible “incoming” states
of the system; dually, an interval containing the other endpoint should describe
“outgoing” states. Let us now spell things out explicitly.
To open subintervals, our prefactorization algebra F assigns the following vec-
tor spaces:
• [0, t) 7→ V
• (s, t) 7→ A
• (t, T ] 7→ V
• [0, T ] 7→ C.
In light of our discussion about modules in the preceding section, note that the
natural choice for the value F([0, T ]) is the vector space V ⊗A V. However, V ⊗End(V)
V is isomorphic to the ground field C due to the compatibility of the left and right
actions with the inner product. We must now describe the structure maps coming
from inclusion of intervals; we will describe enough so that the mechanism is clear.
The case that determines the rest is that the inclusion
[0, t0 ) t (t1 , t2 ) t (t3 , T ] ⊂ [0, T ]
corresponds to the structure map
V ⊗A⊗V → C
.
hv0 | ⊗ O ⊗ | v1 i 7→ hv0 | ei(t1 −t0 )H Oei(t3 −t2 )H | v1 i
In other words, the system evolves according the Hamiltonian during the closed
intervals in the complement of opens during which we specify the incoming and
outgoing states and the operator.
Note that if we set O = ei(t2 −t1 )H , then we obtain
hv0 | ei(t1 −t3 )H | v1 i
and so recover the expected value of being in state hv0 | at time t0 and going to
state | v1 i at time t3 . Setting t3 = t0 , we recover the inner product on V.
For another example of structure maps, the inclusion (t0 , t1 ) ⊂ (t00 , t10 ) goes to
0 0
O 7→ ei(t0 −t0 )H Oei(t1 −t1 )H . More generally, for k disjoint open intervals inside a big
open interval
(t0 , t1 ) t · · · t (t2k−2 , t2k−1 ) ⊂ (t00 , t10 ),
we again time-order the operators and then multiply with evolution operators in-
serted for the closed intervals between them: the structure map for F is
0 0
O1 ⊗ · · · ⊗ Ok 7→ ei(t0 −t0 )H O1 ei(t2 −t1 )H · · · Ok ei(t1 −t2k−1 )H .
Note that these structure maps reduce to those for an arbitrary associative algebra
if we set H = 0 and hence have the identity map as the evolution operator. (In
general, a one-parameter semigroup of algebra automorphisms can be used like
this to “twist” the prefactorization algebra associated to an associative algebra.)
As further examples, we have:
0
• the inclusion [0, t) ⊂ [0, t0 ) has structure map hv0 | 7→ hv0 | ei(t −t)H .
0
• the inclusion (t, T ] ⊂ (t0 , T ] has structure map | v1 i 7→ ei(t−t )H | v1 i.
• the inclusion [0, t0 ) t (t1 , t2 ) ⊂ [0, t ) has structure map hv0 | ⊗ O 7→
0
0
hv0 | ei(t1 −t0 )H Oei(t −t2 )H .
All these choices ensure that we are free to choose what happens during open inter-
vals but that the system evolves according to H during their closed complements.
In this way, the prefactorization algebra encodes the basic abstract structure of
quantum mechanics. The open interior (0, T ) encodes the algebra of observables,
and the boundaries encode the state spaces.
The attentive reader might notice that we have not discussed, e.g., the structure
map associated to (t0 , t1 ) ⊂ [0, t0 ). Here we need to use the distinguished elements
in F((t0 , t1 )) = A and F([0, t0 )) = V arising from the inclusion of the empty set
into these opens. If we fix hvin | as the “idealized” initial state at time 0, then we
0
set hvin | eit H to be the distinguished element in F([0, t0 )) = V. The distinguished
element of F((t0 , t1 )) is the evolution operator ei(t1 −t0 )H . Thus, the structure map for
(t0 , t1 ) ⊂ [0, t0 ) is naturally
0
O 7→ hvin | eit0 H Oei(t −t1 )H
We now describe the dual situation for intervals containing the other endpoint.
Here we specify an “idealized” final state | vout i so that the distinguished element
of F((t, T ]) is ei(T −t)H | vout i. We do not need these initial or final states to recover
the quantum mechanical formalism from the prefactorization algebra, so it is inter-
esting that the prefactorization perspective pushes towards fixing these boundary
states in the form of the distinguished elements.
Remark: Although we have explained here how to start with the standard ingredi-
ents of quantum mechanics and encode them as a prefactorization algebra, one can
also turn the situation around and motivate (or interpret), via the factorization per-
spective, aspects of the quantum mechanical formalism. For instance, time-reversal
amounts to reflection across a point in R. Requiring a locally constant prefactoriza-
tion algebra on R to be equivariant under time-reversal corresponds to equipping
the corresponding associative algebra A with an involutive algebra antiautomor-
phism. In other words, the prefactorization algebra corresponds to a ∗-algebra A.
Likewise, suppose we want a prefactorization algebra on a closed interval [0, T ]
such that it corresponds to A on the open interior. The right end point corresponds
4. A CONSTRUCTION OF THE UNIVERSAL ENVELOPING ALGEBRA 49
to a left A-module V, viewed as the “outgoing” states. It is natural to want the “in-
coming” states — given by some right A-module V — to be “of the same type” as
V (e.g., abstractly isomorphic) and to want the global sections to be C. This forces
us to have a map
V ⊗ A ⊗ V → C.
This map induces a pairing between V and V, which provides a pre-Hilbert struc-
ture. ^
Remark: Our construction above captures much of the standard formalism of quan-
tum mechanics, but there are a few loose ends to address.
First, in standard quantum mechanics, a state is not a vector in V but a line.
Above, however, we fixed vectors hvin | and | vout i, so there seems to be a dis-
crepancy. The observation that rescues us is a natural one, from the mathematical
viewpoint. Consider scaling the initial and final states by elements of C× . This de-
fines a new factorization algebra, but it is isomorphic to what we described above,
and the expected value “hv0 |Ov1 i” of an operator depends linearly in the rescaling
of the input and output vectors. More precisely, there is a natural equivalence rela-
tion we can place on the factorization algebras described above that corresponds to
the usual notion of state in quantum mechanics. In other words, we could make a
groupoid of factorization algebras where the underlying vector spaces and structure
maps are all the same, but the distinguished elements are allowed to change.
Another issue that might bother the reader is that our formalism only matches
nicely with experiments that resemble scattering experiments. It does not seem
well-suited to descriptions of systems like bound states (e.g., an atom sitting qui-
etly, minding its own business). For such systems, we might consider running
over the whole space of states, which is described as a groupoid in the previous
paragraph. Alternatively, we might drop the endpoints and simply work with the
factorization algebra on the open interval, which focuses on the algebra of opera-
tors. ^
4. A construction of the universal enveloping algebra

Let g be a Lie algebra. In this section we describe a procedure that produces
the universal enveloping algebra Ug as a prefactorization algebra on R. This con-
struction is useful both because it is a model for a more general result about En
algebras (see Section 4) and because it appears in several of our examples (such
as the Kac-Moody factorization algebra). For a mathematician, it may be useful to
see techniques similar to those we use in Section 2 shorn of any connection with
physics, so that the underlying process is clearer.
Let gR denotes the cosheaf on R that assigns (Ω∗c (U) ⊗ g, ddR ) to each open U,
with ddR the exterior derivative. This is a cosheaf of cochain complexes, but it is
only a precosheaf of dg Lie algebras. Note that the cosheaf axiom involves the use
of coproducts, and the coproduct in the category of dg Lie algebras is not given by
direct sum of the underlying cochain complexes. In fact, gR is a prefactorization
algebra taking value in dg Lie algebras equipped with direct sum ⊕ as the symmet-
ric monoidal product (here direct sum means the sum of the underlying cochain
complexes, which inherits a natural Lie bracket).
Let C∗ h denote the Chevalley-Eilenberg complex for Lie algebra homology,
written as a cochain complex. In other words, C∗ h is the graded vector space
Sym(h[1]) with a differential determined by the bracket of h. See Appendix 3 for
the definition and further discussion of this construction.
Our main result shows how to construct the universal enveloping algebra Ug
using C∗ (gR ).
4.0.1 Proposition. Let H denote the cohomology prefactorization algebra of C∗ (gR ).
That is, we take the cohomology of every open and every structure map, so
H(U) = H ∗ (C∗ (gR (U)))
for any open U. Then H is locally constant, and the corresponding associative
algebra is isomorphic to Ug, the universal enveloping algebra of g.
Remark: Recall Lemma 2.0.2, which says that every locally constant prefactoriza-
tion algebra on R corresponds to an associative algebra. This proposition above
provides a homological mechanism for recovering the universal enveloping alge-
bra of a Lie algebra, but the reader has probably noticed that we could apply the
same construction with R replaced by any smooth manifold M. In Chapter 6, we
will explain how to understand what this general procedure means. ^
Proof. Local constancy of H is immediate from the fact that, if I ⊂ J is the
inclusion of one interval into another, the map of dg Lie algebras
Ω∗c (I) ⊗ g → Ω∗c (J) ⊗ g
is a quasi-isomorphism. We let Ag be the associative algebra constructed from H
by Lemma 2.0.2.
The underlying vector space of Ag is the space H(I) for any interval I. To be
concrete, we will use the interval R, so that we identify
Ag = H(R) = H ∗ (C∗ (Ω∗c (R) ⊗ g)).
We now identify that vector space.
The dg Lie algebra Ω∗c (R) ⊗ g is concentrated in degrees 0 and 1 and maps
quasi-isomorphically to its cohomology Hc∗ (R) ⊗ g = g[−1], which is concentrated
in degree 1 by the Poincaré lemma. This cohomology is an Abelian Lie algebra
because the cup product on Hc∗ (I) is zero. It follows that C∗ (Ω∗c (R) ⊗ g) is quasi-
isomorphic to the Chevalley-Eilenberg chains of the Abelian Lie algebra g[−1],
which is simply Sym g. Thus, as a vector space, Ag is isomorphic to the symmetric
algebra Sym g.
There is a map
Φ : g → Ag
that sends an element X ∈ g to ε ⊗ X where ε ∈ Hc1 (I) is a basis element for the
compactly supported cohomology of the interval I; we require the integral of ε to
be 1. We will show that Φ is a map of Lie algebras, where Ag is given the Lie
4. A CONSTRUCTION OF THE UNIVERSAL ENVELOPING ALGEBRA 51
bracket coming from the associative structure. This result immediately implies the
theorem, as we then have an map of associative algebra Φ : Ug → Ag that is an
isomorphism of vector spaces.
Let us check explicitly that Φ is a map of Lie algebras. Let δ > 0 be small
number, and let f0 ∈ Cc∞ (−δ, δ) be a compactly supported smooth function with
f0 dx = 1. Let ft (x) = f0 (x−t). Note that ft is supported on the interval (t−δ, t+δ).
R
If X ∈ g, a cochain representative for Φ(X) ∈ Ag is provided by

f0 dx ⊗ X ∈ Ω1c ((−δ, δ)) ⊗ g.
Indeed, because every ft dx is cohomologous to f0 dx in Ω1c (R), the element ft dx ⊗
X is a cochain representative of Φ(X) for any t.
Given elements α, β ∈ Ag , the product α · β is defined as follows.
((i)) We choose intervals I, J with I < J.
((ii)) We regard α as an element of H(I) and β as an element of H(J) using
the inverses to the isomorphisms H(I) → H(R) and H(J) → H(R)
coming from the inclusions of I and J into R.
((iii)) The product α · β is defined by taking the image of α ⊗ β under the
factorization structure map
H(I) ⊗ H(J) → H(R) = Ag .
Let us see how this works with our representatives.
The cohomology class [ ft dx ⊗ X] ∈ H(t − δ, t + δ) becomes Φ(X) under the
natural map from H(t − δ, t + δ) to H(R). If we take δ to be sufficiently small, the
intervals (−δ, δ) and (1 − δ, 1 + δ) are disjoint. It follows that the product Φ(X)Φ(Y)
is represented by the cocycle
( f0 dx ⊗ X)( f1 dx ⊗ Y) ∈ Sym2 (Ω1c (R) ⊗ g) ⊂ C∗ (Ω∗c (R) ⊗ g).
Similarly, the commutator [Φ(X), Φ(Y)] is represented by the expression
( f0 dx ⊗ X)( f1 dx ⊗ Y) − ( f0 dx ⊗ Y)( f−1 dx ⊗ X).
It suffices to show that this cocycle in C∗ (Ω∗c (R) ⊗ g) is cohomologous to Φ([X, Y]).
Note that the 1-form f1 dx − f−1 dx has integral 0. It follows that there exists a
compactly supported function h ∈ Cc∞ (R) with
ddR h = f−1 dx − f1 dx.
We can assume that h takes value 1 in the interval (−δ, δ).
We now calculate the differential of the element
( f0 dx ⊗ X)(h ⊗ Y) ∈ C∗ (Ω∗c (R) ⊗ g).
We have
d (( f0 dx ⊗ X)(h ⊗ Y)) = ( f0 dx ⊗ X)(dh ⊗ Y) + f0 h dx ⊗ [X, Y]
= ( f0 dx ⊗ X)(( f−1 − f1 )dx ⊗ Y) + f0 h dx ⊗ [X, Y].
Since h takes value 1 on the interval (−δ, δ), we know f0 h = f0 . This equation tells
us that a representative for [Φ(X), Φ(Y)] is cohomologous to Φ([X, Y]).
5. Some functional analysis

Nearly all of the examples of factorization algebras that we will consider in
this book will assign to an open subset, a cochain complex built from vector spaces
of analytical provenance: for example, the vector space of smooth sections of a
vector bundle or the vector space of distributions on a manifold. Such vector spaces
are best viewed as being equipped with an extra structure, such as a topology,
reflecting their analytical origin. In this section we will briefly sketch a flexible
multicategory of vector spaces equipped with an extra “analytic” structure. More
details are contained in Appendix B.
5.1. Differentiable vector spaces. The most common way to encode the an-
alytic structure on a vector space such as the space of smooth functions on a man-
ifold is to endow it with a topology. Homological algebra with topological vector
spaces is not easy, however. (For instance, topological vector spaces do not form
an abelian category.) To get around this issue, we will work with differentiable
vector spaces. Let us first define the slightly weaker notion of a C ∞ -module.
5.1.1 Definition. Let Mfld be the site of smooth manifolds, i.e., the category of
smooth manifolds and smooth maps between them, where a cover is a surjective
local diffeomorphism.
Let C ∞ denote the sheaf of rings on Mfld that assigns to any manifold M the
commutative algebra C ∞ (M). A C ∞ -module is a module sheaf over C ∞ on Mfld.
In other words, to each manifold M, a C ∞ -module F assigns a module F (M)

over the algebra C ∞ (M), and for any map of manifolds f : M → N, the pullback
map f ∗ : F (N) → F (M) is a map of C ∞ (N)-modules. For example, if V is any
topological vector space, then there is a natural notion of smooth map from any
manifold M to V (see, e.g., Kriegl and Michor (1997)). The space C ∞ (M, V) of
such smooth maps is a module over the algebra C ∞ (M). Since smoothness is a
local condition on M, sending M to C ∞ (M, V) gives a sheaf of C ∞ -modules on the
site Mfld.
As an example of this construction, let us consider the case when V is the space
of smooth functions on a manifold N, equipped with its usual Fréchet topology.
One can show that for each manifold M, the space C ∞ (M, C ∞ (N)) is naturally
isomorphic to the space C ∞ (M × N) of smooth functions on M × N.
As we will see shortly, we lose very little information when we view a topo-
logical vector space as a C ∞ -module.
Sheaves of C ∞ -modules on Mfld that arise from topological vector spaces are
endowed with an extra structure: we can always differentiate smooth maps from a
manifold M to a topological vector space V. Differentiation can be viewed as an
action of the vector fields on M on the vector space C ∞ (M, V). Dually, it can be
viewed as coming from a connection
∇ : C ∞ (M, V) → Ω1 (M, V),

5. SOME FUNCTIONAL ANALYSIS 53
where Ω1 (M, V) is defined to be the tensor product

Ω1 (M, V) = Ω1 (M) ⊗C ∞ (M) C ∞ (M, V).
This tensor product is just the algebraic one, which is well-behaved because Ω1 (M)
is a projective C ∞ (M)-module of finite rank.
This connection is flat, in that the curvature
F(∇) = ∇ ◦ ∇ : C ∞ (M, V) → Ω2 (M, V)
is zero. Flatness ensures that the action of vector fields is actually a Lie algebra
action, so that we get an action of all differential operators on M and not just vector
fields.
This flatness property suggests the following definition.
5.1.2 Definition. Let Ω1 denote the C ∞ -module that assigns Ω1 (M) to a manifold
M. For F a C ∞ -module, the C ∞ -module of k-forms valued in F is the C ∞ -module
Ωk (F) that assigns to a manifold M, the tensor product Ωk (M, F) = Ωk (M) ⊗C ∞ (M)
F(M).
A connection on a C ∞ -module F is a map of sheaves on the site Mfld,
∇ : F → Ω1 (F),
that satisfies the Leibniz rule on every manifold M. A connection is flat if it is flat
on every manifold M.
A differentiable vector space is a C ∞ -module equipped with a flat connection.
A map of differentiable vector spaces f : F → G is a map of C ∞ -modules that inter-
twines with the flat connections. We denote the set of all such maps by DVS(F, G).
Almost all of the differentiable vector spaces we will consider are concrete
in nature. Indeed, most satisfy the formal definition of a being a concrete sheaf,
which we now explain. Let Set denote the category of sets and let Set(S , T ) denote
the collection of functions from the set S to the set T . For any sheaf F on Mfld
taking values in Set, there is a natural map F (M) → Set(M, F (∗)): each element
of the set F (M) has an associated function from the underlying set M to the set
F (∗), the value of the sheaf on a point. We say a sheaf F is concrete if this map
F (M) → Set(M, F (∗)) is injective. Hence, for a concrete sheaf, one can think
of any section on M as just a particular function from M to F (∗); a section is
just a “smooth” function on M with values in the set F (∗). As an example of a
concrete sheaf, consider the sheaf X associated to a smooth manifold X, where
X(M) = C ∞ (M, X). In this case, a section of the sheaf X really is just a function to
X. This sheaf just identifies which set-theoretic functions are smooth.
We often work with the differentiable vector space arising from a topological
vector space V, which, just like the example X, simply records which set-theoretic
maps are smooth. For this reason, we will normally think of a differentiable vector
space F as being an ordinary vector space, given by its value on a point F (∗),
together with extra structure. We will often refer to the sections F (M) (i.e., the
value of the sheaf F on the manifold M) as the space of smooth maps to the value
on a point. If V is a differentiable vector space, we often write C ∞ (M, V) for the
space of smooth maps from a manifold M. Abusively, we will also often call a map
of differentiable vector spaces simply a smooth map.
5.2. Differentiable cochain complexes. Although we typically work with dif-
ferentiable vector spaces coming from topological vector spaces, the category DVS
is much easier to use — for our purposes — than that of topological vector spaces.
The key reason is that they are (essentially) just sheaves on a site, and homological
algebra for such objects is very well-developed, as we explain in Appendix C.
5.2.1 Definition. A differentiable cochain complex is a cochain complex in the
category of differentiable vector spaces.
A cochain map f : V → W of differentiable cochain complexes is a quasi-
isomorphism if the map C ∞ (M, V) → C ∞ (M, W) is a quasi-isomorphism for all
manifolds M. This condition is equivalent to asking that the map be a quasi-
isomorphism at the level of stalks.
We use Ch(DVS) to denote the category of differentiable cochain complexes

and cochain maps. It can be enriched over cochain complexes of vector spaces in
the usual way.
5.3. Differentiable prefactorization algebras. We have defined the notion
of prefactorization algebra with values in any multicategory. In order to discuss
factorization algebras valued in differentiable vector spaces, we need to define a
multicategory structure on differentiable vector spaces. Let us first discuss the
multicategory structure on the C ∞ -modules.
5.3.1 Definition. Let V1 , . . . , Vn , W be differentiable vector spaces. A smooth mul-
tilinear map
Φ : V1 × · · · × Vn → W
is a C -multilinear map Φ of sheaves that satisfies the following Leibniz rule with
∞
respect to the connections on the Vi and W. For every manifold M, and for every
vi ∈ Vi (M), we require that
Xn
∇Φ(v1 , . . . , vn ) = Φ(v1 , . . . , ∇vi , . . . , vn ) ∈ Ω1 (M, W).
i=1
We let DVS(V1 , . . . , Vn | W) denote this space of smooth multilinear maps.
In more down-to-earth terms, such a Φ is a C ∞ (M)-multilinear map Φ(M) :

V1 (M) × · · · × Vn (M) → W(M) for every manifold M, in a way compatible with the
connections and with the maps Vi (M) → Vi (N) associated to a map f : N → M of
manifolds.
The category of differentiable cochain complexes acquires a multicategory
structure from that on differentiable vector spaces, where the multi-maps are smooth
multilinear maps that are compatible with the differentials. (Here “compatible”
means precisely the same thing as it does with ordinary vector spaces and cochain
complexes. It means, in the case of unary maps, that we work with cochain maps
and not arbitrary linear maps of graded vector spaces.)
5.3.2 Definition. A differentiable prefactorization algebra is a prefactorization al-

gebra valued in the multicategory of differentiable cochain complexes.
In words, a differentiable prefactorization algebra F on a space X assigns a

differentiable cochain complex F (U) to every open subset U ⊂ X, and a smooth
multi-linear cochain map
F (U1 ) × · · · × F (Un ) → F (V),
whenever U1 , . . . , Un are disjoint opens contained in V.
Note the very different roles played here by the space X, on which the prefac-
torization algebra lives, and the site Mfld of manifolds, on which the differentiable
cochain complexes live. The topology of the space X organizes the algebraic struc-
ture we are interested in. By contrast, the geometry of manifolds encoded in Mfld
organizes the structure of the vector spaces (or cochain complexes) we work with.
Said succinctly, as a substitute for a topology on these vector spaces, we use a sheaf
structure over Mfld.
5.4. Relationship with topological vector spaces. As we have seen, every

topological vector space gives rise to a differentiable vector space. There is a beau-
tiful theory developed in Kriegl and Michor (1997) concerning the precise rela-
tionship between topological vector spaces and differentiable vector spaces. These
results are discussed in much more detail in Appendix B: we briefly summarize
them now.
Let LCTVS denote the category of locally-convex Hausdorff topological vector
spaces, and continuous linear maps. Let BVS denote the category with the same
objects, but whose morphisms are bounded linear maps. Every continuous linear
map is bounded, but not conversely. These categories have natural enrichments
to multicategories, where the multi-maps are continuous (respectively, bounded)
multilinear maps. The category BVS is equivalent to a full subcategory of LCTVS
whose objects are called bornological vector spaces.
Theorem. The functor di ft : LCTVS → DVS restricts to a functor di fβ : BVS →
DVS, which embeds BVS as a full sub-multicategory of DVS.
In other words: if V, W are topological vector spaces, and if di ft (V), di ft (W)
denote the corresponding differentiable vector spaces, then the maps in DVS from
di ft (V) to di ft (W) are the same as bounded linear maps from V to W. More gen-
erally, if V1 , . . . , Vn and W are topological vector spaces, the bounded multi-linear
maps BVS(V1 , . . . , Vn | W) are the same as smooth multi-linear maps DVS(di ft (V1 ), . . . , di ft (Vn ) |
di ft (W)).
This theorem tells us that we lose very little information if we think of a topo-
logical vector space as being a differentiable vector space. We just end up thinking
about bounded maps rather than continuous maps.
Remark: On occasion, we will exploit this relationship by using a concept from
topological or bornological vector spaces without profferring a differentiable ana-
log. For example, in a few places we talk about a dense inclusion and use it to show
that a map out of the bigger space is determined by the subspace. In such cases,
we are working only with bornological vector spaces and depending upon the fact
that di fβ is a full and faithful functor. ^
So far, however, we have not discussed how completeness of topological vector
spaces appears in this context. We need a notion of completeness for a topological
vector space that only depends on smooth maps to that vector space. The relevant
concept was developed in Kriegl and Michor (1997). We will view the category
BVS as being a full subcategory of DVS.
5.4.1 Definition. A topological vector space V ∈ BVS is c∞ -complete, or conve-
nient, if every smooth map c : R → V has an antiderivative.
We denote the category of convenient vector spaces and bounded linear maps
by CVS.
This completeness condition is a little weaker than the one normally studied
for topological vector spaces. That is, every complete topological vector space is
c∞ -complete.
Proposition. The full subcategory di fc : CVS ⊂ DVS is closed under the forma-
tion of limits, countable coproducts, and sequential colimits of closed embeddings.
We give CVS the multicategory structure inherited from BVS. Since BVS is a
full sub multicategory of DVS, so is CVS.
Theorem. The multicategory structure on CVS is represented by a symmetric
⊗β .
monoidal structure b
This symmetric monoidal structure is called the completed bornological tensor
product. If E, F ∈ CVS, this completed bornological tensor product is written
as Eb⊗β F. The statement that it represents the multicategory structure means that
smooth (equivalently, bounded) bilinear maps f : E1 × E2 → F are the same as
⊗β E2 → F, for objects E1 , E2 , F of CVS.
bounded linear maps f 0 : E1b
When it should cause no confusion, we may use the symbol ⊗ instead of b ⊗β
for this tensor product.
5.5. Examples from differential geometry. Let us now give some examples
of differentiable vector spaces. These examples will include the basic building
blocks for most of the factorization algebras we will consider.
Let E be a vector bundle on a manifold X. We let E (X) denote the vector
space of smooth sections of E on X, and we let Ec (X) denote the vector space of
compactly supported sections of E on X.
Let us give these vector spaces the structure of differentiable vector spaces, as
follows. If M is a manifold, we say a smooth map from M to E is a smooth section
of the bundle π∗X E on M × X. We denote this set of smooth maps C ∞ (M, E (X)).
Sending M to C ∞ (M, E (X)) defines a sheaf of C ∞ -modules on the site of smooth
manifolds with a flat connection, and so a differentiable vector space.
Similarly, we say a smooth map from M to Ec (X) is a smooth section of the

bundle π∗X E on M × X, whose support maps properly to M. Let us denote this set
by C ∞ (M, Ec (X)); this defines, again, a sheaf of C ∞ -modules on the site of smooth
manifolds with a flat connection.
Theorem. With this differentiable structure, the spaces E (X) and Ec (X) are in
the full subcategory CVS of convenient vector spaces. Further, this differentiable
structure is the same as the one that arises from the natural topologies on E (X)
and Ec (X).
The proof (like the proofs of all results in this section) are contained in Appen-
dix B, and based heavily on the book Kriegl and Michor (1997).
Let E be a vector bundle on a manifold X. Throughout this book, we will often
use the notation E (X) to denote the distributional sections on X, defined by
E (X) = E (X) ⊗C ∞ (X) D(X),
where D(X) is the space of distributions on X. Similarly, let E c (X) denote the
compactly supported distributional sections of E on X. There are natural inclusions
Ec (X) ,→ E c (X) ,→ E (X),
Ec (X) ,→ E (X) ,→ E (X),

by viewing smooth functions as distributions.
Let E ! denote the vector bundle E ∨ ⊗ DensX , which possesses a natural vector
bundle pairing ev : E ⊗ E ! → DensX . In other words, a smooth section of E and
a smooth section of E ! can be paired to produce a smooth density on X. If this
smooth density has compact support, it can certainly be integrated to produce a
real number. With a little more work, one can show that E c (X) is the continuous
dual to E ! (X), the space of smooth sections of E ! on X. Likewise, one can show
!
that Ec (X) is the continuous dual to E (X), the distributional sections of E ! .
These topological vector spaces E (X) and E c (X) thus obtain natural differ-
entiable structures via the functor di ft . We have the following description of the
differentiable structure.
Theorem. For any manifold M, a smooth map from M to E (X) is the same as a
smooth linear map
Ec! (X) → C ∞ (M).
Similarly, a smooth map from M to E c (X) is the same as a smooth linear map
E ! (X) → C ∞ (M).
This is a consequence of Lemma 5.2.2. See also Section 2.1 for more discus-
sion.
5.6. Multilinear maps and enriched spaces of maps. The category of dif-
ferentiable vector spaces has a natural tensor product. In other words, it is a sym-
metric monoidal category. The tensor product in DVS is very simple: if V, W are
differentiable vector spaces, then they are C ∞ -modules (by forgetting the flat con-
nections) so V ⊗C ∞ W is another C ∞ -module, but it inherits a natural flat connection
∇V ⊗ IdW + IdV ⊗∇W , so we obtain a new differentiable vector space.
One must be careful with this tensor product, however. From the point of view
of analysis, this tensor product is not very meaningful: it is somewhat similar to an
un-completed tensor product of topological vector spaces. For example, if M and N
are manifolds, then C ∞ (M) and C ∞ (N) naturally have the structure of differentiable
vector spaces. It is not true that
(†) C ∞ (M) ⊗C ∞ C ∞ (N) = C ∞ (M × N).
This issue means that our examples will not assign to a disjoint union of opens the
tensor product of values on the components.
The multicategory structure on DVS that we use coincides with this symmetric
monoidal structure. The multicategory structure is better behaved than the sym-
metric monoidal structure: when restricted to the full subcategory CVS the multi-
category becomes the one associated to the symmetric monoidal structure on CVS,
which has good analytical properties (and in particular satisfies the equality in (†)).
Similarly, there is an internal hom in the category of C ∞ -modules, and this
sheaf hom likewise inherits a natural flat connection. We denote this sheaf hom by
HomDVS (V, W). This sheaf hom is not as well-behaved as one would hope, how-
ever, and does not capture the concept of “smooth families of maps”. In particular,
it is not true that the value of the sheaf HomDVS (V, W) on a point is the vector
space DVS(V, W) of smooth maps from V to W. For any reasonable definition of
the notion of smooth family of maps parametrized by a manifold, a smooth family
of maps parametrized by a point should be simply a map, and the self-enrichment
given by HomDVS (V, W) does not satisfy this.
There is, however, another way to enrich the category DVS over itself that
better captures the notion of smooth family of maps. (For a careful treatment, see
Section 6.) Before we define this enrichment, we need the following definition.
5.6.1 Definition. For V a differentiable vector space and M a manifold, let C∞ (M, V)
denote the differentiable vector space whose value on a manifold N is C ∞ (N ×
M, V). The flat connection map
∇N,C∞ (M,V) : C ∞ (N, C∞ (M, V)) → Ω1 (N, C∞ (M, V))
is the composition of the flat connection
∇N×M : C ∞ (M × N, V) → Ω1 (M × N, V) = Ω1 (M, C∞ (N, V)) ⊕ Ω1 (N, C∞ (M, V))
with the projection onto Ω1 (N, C∞ (M, V)).
Here the direct sum decomposition of one-forms is a consequence of the fact

∗
that T M×N ∗ ⊕ π∗ T ∗ , where π : M × N → M and
splits as the direct sum π∗M T M N N M
πN : M × N → N denote the projection maps.
This definition makes the category of differentiable vector spaces into a cate-
gory cotensored over the category of smooth manifolds. Note that C ∞ (M) defines
a ring object in the category DVS, and for any differentiable vector space V, the
mapping space C∞ (M, V) is a C ∞ (M)-module.
This construction generalizes in a natural way. If E is a vector bundle on a
manifold M and V is a differentiable vector space, then we can define
C∞ (M, E ⊗ V) := E (M) ⊗C ∞ (M) C∞ (M, V)
where on the right hand side we are taking tensor products in the category DVS
over the differentiable ring C ∞ (M). Although, in general, tensor products in DVS
are not analytically well-behaved, in this case there are no problems because E (M)
is a projective C ∞ (M) module of finite rank. We will denote C∞ (M, T ∗ M ⊗ E) by
Ω1 (M, E).
5.6.2 Definition. Let V1 , . . . , Vn , W be differentiable vector spaces. Given a mani-
fold M, a smooth family of multi-linear maps V1 × · · · × Vn → W parametrized by
M is an element of
DVS(V1 , . . . , Vn | C∞ (M, W))
where we regard C∞ (M, W) as being a differentiable vector space using the defini-
tion above.
In Section 6 we show that there is a differentiable vector space

HomDVS (V1 , . . . , Vn | W)
defined by saying that its value on a manifold M is DVS(V1 , . . . , Vn | C∞ (M, W)).
The flat connection comes from the natural map
∇C∞ (M,W) : C∞ (M, W) → Ω1 (M, W),
given by applying the projection in the M-direction to the connection ∇ M×N,W
(compare to the connection on C∞ (M, W)).
This definition makes the category DVS into a multicategory enriched in itself,
in such a way that if we evaluate the differentiable Hom-space on a point, we
recover the original vector space of smooth maps.
In this text, whenever we consider a smooth family of maps between differen-
tiable vector spaces, we are always referring to this self-enrichment.
5.7. Algebras of observables. The prefactorization algebras we will use for
most of the book are built as algebras of functions on the convenient vector spaces
E (X), which for us will mean symmetric algebras on their dual spaces. Recall that
E (X) and E c (X) are both convenient vector spaces, and that, in the full subcategory
CVS ⊂ DVS of convenient vector spaces, the multicategory is represented by a
symmetric monoidal structure b ⊗β . Recall as well that b⊗π denotes the completed
projective tensor product in LCTVS.
Proposition. Let E be a vector bundle on a manifold X, and let F be a vector
bundle on a manifold Y. Let E (X) denote the convenient vector space of smooth
sections of E on X, and let F (Y) denote the convenient vector space of smooth
sections of F on Y. Then
E (X)b
⊗β F (Y) Γ(X × Y, E F) E (X)b
⊗π F (Y),
where E F denotes the external tensor product of vector bundles and Γ denotes
smooth sections.
Remark: An alternative approach to the one we’ve taken is to use the category of
nuclear topological vector spaces, with the completed projective tensor product,
instead of the category of convenient (or differentiable) vector spaces. Using nu-
clear spaces raises a number of technical issues, but one immediate issue is the
⊗πC ∞ (Y) = C ∞ (X × Y), it seems not to
following: although it is true that C ∞ (X)b
be true that the same statement holds if we use compactly supported smooth func-
tions. The problem stems from the fact that the projective tensor product does not
commute with colimits, whereas the bornological tensor product does. ^
We can define symmetric powers of convenient vector spaces using the sym-
metric monoidal structure we have described. If, as before, E is a vector bundle on
X and U is an open subset of X, this proposition allows us to identify
Symn (Ec (U)) = Cc∞ (U n , E n )S n .
The symmetric algebra Sym Ec (U) is defined as usual to be the direct sum of the
symmetric powers. It is an algebra in the symmetric monoidal category of conve-
nient vector spaces.
A related construction is the algebra of functions on a differentiable vector
space. If V is a differentiable vector space, we can define, as we have seen, the
space of linear functionals on V to be the space of maps HomDVS (V, R). Because
the category DVS is self-enriched, this is again a differentiable vector space. In a
similar way, we can define the space of polynomial functions on V homogeneous
of degree n to be the space
Pn (V) = HomDVS (V, . . . , V | C ∞ )S n .
| {z }
n times
In other words, we take smooth multilinar maps from n copies of V to R, and
then take the S n coinvariants. The self-enrichment of DVS gives this the structure
of differentiable vector space. Concretely, a smoooth map from a manifold M to
Pn (V) is
C ∞ (M, Pn (V)) = HomDVS (V, . . . , V | C ∞ (M))S n .
| {z }
n times
One can then define the algebra of functions on V by
Y
O(V) = Pn (V).
n
(In this formula, we take the product rather than the direct sum, so that O(V) should
be thought of as a space of formal power series on V. One can, of course, also
consider the version using the sum). The space O(V) is a commutative algebra
object of the category DVS in a natural way.
This construction is a very general one, of course: one can define the algebra
of functions on any object in any multicategory in the same way.
An important example is the following.
6. THE FACTORIZATION ENVELOPE OF A SHEAF OF LIE ALGEBRAS 61
5.7.1 Lemma. Let E be a vector bundle on a manifold X. Then, Pn (E(X)) is iso-

morphic as a differentiable vector space to the S n covariants of the space of com-
pactly supported distributional sections of E ! on X n .
Proof. We know that the multicategory structure on the full subcategory CVS ⊂
DVS is represented by a symmetric monoidal category, and that in this symmetric
monoidal category,
E (X)b⊗β n = Γ(X n , E n ).
It follows from this that, for any manifold M, we can identify C ∞ (M, Pn (E (X)))
with the S n covariants of the space of smooth linear maps
Γ(X n , E n ) → C ∞ (M).
In fact, since the spaces in this equation are bornological (see Appendix B), such
smooth linear maps are the same as continuous linear maps.
We have seen that this space of smooth linear maps is — with its differen-
tiable structure — the same as the space Dc (X n , (E ! )n ) of compactly supported
distributional sections of the bundle (E ! )n .
Note that O(E (U)) is naturally the same as HomDVS (Sym E (U), R), i.e., it is
the dual of the symmetric algebra of E (U).
6. The factorization envelope of a sheaf of Lie algebras

In this section, we will introduce an important class of examples of factoriza-
tion algebras. We will show how to construct, for every fine sheaf of Lie algebras
L on a manifold M, a factorization algebra that we call the factorization envelope.
This construction is our version of the chiral envelope introduced in Beilinson and
Drinfeld (2004). The construction can also be viewed as a natural generalization
of the universal enveloping algebra of a Lie algebra. Indeed, we have shown in
Section 4 that the factorization envelope of the constant sheaf of Lie algebras g on
R is the universal enveloping algebra of g, viewed as a factorization algebra on R.
The factorization envelope plays an important role in our story.
((i)) The factorization algebra of observables for a free field theory is an
example of a factorization envelope.
((ii)) In Section 4, we will show, following Beilinson and Drinfeld, that the
Kac-Moody vertex algebra arises as a (twisted) factorization envelope.
The most important appearance of factorization envelopes appears in our treatment
of Noether’s theorem at the quantum level, which is covered in Volume 2. We show
there that if a sheaf of Lie algebras L acts on a quantum field theory on a manifold
M, then there is a morphism from a twisted factorization envelope of L to the quan-
tum observables of the field theory. This construction is very useful. For example,
for a chiral conformal field theory, it allows one to construct a map of factorization
algebras from a Virasoro factorization algebra to the quantum observables. This
map induces a map of vertex algebras via our construction in Section 2.
6.1. The key idea. Thus, let M be a manifold. Let L be a fine sheaf of dg
Lie algebras on M. Let Lc denote the cosheaf of compactly supported sections of
L. (We restrict to fine sheaves so that taking compactly supported sections is a
straightforward operation.)
Remark: Note that, although Lc is a cosheaf of cochain complexes, and a pre-
cosheaf of dg Lie algebras, it is not a cosheaf of dg Lie algebras. This is because
colimits of dg Lie algebras are not the same as colimits of cochain complexes. ^
We can view Lc as a prefactorization algebra valued in the category of dg Lie
algebras with symmetric monoidal structure given by direct sum. Indeed, if {Ui } is
a finite collection of disjoint opens in M contained in the open V, there is a natural
map of dg Lie algebras
M
Lc (Ui ) = Lc (ti Ui ) → Lc (V)
i
giving the factorization product.

Taking Chevalley-Eilenberg chains is a symmetric monoidal functor from dg
Lie algebras, equipped with the direct sum monoidal structure, to cochain com-
plexes.
6.1.1 Definition. If L is a sheaf of dg Lie algebras on M, the factorization envelope
UL is the prefactorization algebra obtained by applying the Chevalley-Eilenberg
chain functor to Lc , viewed as a prefactorization algebra valued in dg Lie algebras.
Concretely, UL assigns to an open subset V ⊂ M the cochain complex

UL(V) = C∗ (Lc (V)),
where C∗ is the Chevalley-Eilenberg chain complex. The factorization structure
maps are defined as follows: given a finite collection of disjoint opens {Ui } in V,
we have  
O M 
C∗ Lc (Ui ) C∗  Lc (Ui ) → C∗ Lc (V).
i i
This construction is parallel to example 1.1.
We will see later (see Theorem 5.2.1) that this prefactorization algebra is a
factorization algebra.
6.2. Local Lie algebras. In practice, we will need an elaboration of this con-
struction which involves a small amount of analysis.
6.2.1 Definition. Let M be a manifold. A local dg Lie algebra on M consists of the
following data:
((i)) a graded vector bundle L on M, whose sheaf of smooth sections will
be denoted L
((ii)) a differential operator d : L → L, of cohomological degree 1 and
square 0;
6. THE FACTORIZATION ENVELOPE OF A SHEAF OF LIE ALGEBRAS 63
((iii)) an alternating bi-differential operator

[−, −] : L⊗2 → L
that endows L with the structure of a sheaf of dg Lie algebras.
Remark: This definition will play an important role in our approach to interacting
classical field theories, developed in Volume 2. ^
If U ⊂ M, then L(U) is a topological vector space, because it is the space of
smooth sections of a graded vector bundle on U. We would like to form, as above,
the Chevalley-Eilenberg chain complex C∗ (Lc (U)). The underlying vector space
of C∗ (Lc (U)) is the (graded) symmetric algebra on Lc (U)[1]. We need to take
account of the topological structure on Lc (U) when we take the tensor powers of
Lc (U).
We explained how to do this in the Section 5: we define (Lc (U))⊗n to be the
tensor power defined using the completed projective tensor product on the topo-
logical vector space Lc (U). Concretely, if Ln denotes the vector bundle on M n
obtained as the external tensor product, then
(Lc (U))⊗n = Γc (U n , Ln )
is the space of compactly supported smooth sections of Ln on U n . Symmetric
(or exterior) powers of Lc (U) are defined by taking coinvariants of Lc (U)⊗n with
respect to the action of the symmetric group S n . The completed symmetric algebra
on Lc (U)[−1] that is the underlying graded vector space of C∗ (Lc (U)) is defined
using these completed symmetric powers. The Chevalley-Eilenberg differential is
continuous, and therefore defines a differential on the completed symmetric algebra
of Lc (U)[−1], giving us the cochain complex C∗ (Lc (U)).
Example: Let g be a Lie algebra. Consider the sheaf Ω∗R ⊗ g of dg Lie algebras
on R, which assigns to the open U the dg Lie algebra Ω∗ (U) ⊗ g. Then, as we
saw in detail in Section 4, the factorization envelope of this sheaf of Lie algebras
encodes the universal enveloping algebra Ug of g. Indeed, factorization algebras on
R with an additional “locally constant” property give rise to associative algebras,
and the associative algebra associated to U(Ω∗R ⊗ g) recovers the ordinary universal
enveloping algebra.
In the same way, for any Lie algebra g we can construct a factorization algebra
on Rn as the factorization envelope of Ω∗Rn ⊗ g. The resulting factorization algebra
is locally constant: it has the property that the inclusion map from one disc to
another is a quasi-isomorphism. A theorem of Lurie Lurie (n.d.b) tells us that
locally constant factorization algebras on Rn are the same as En algebras. The En
algebra we have constructed is the En -enveloping algebra of g. (See the discussion
in Section 4 for more about locally constant factorization algebras, En algebras,
and the En enveloping functor.) ^
6.3. Shifted central extensions and the twisted envelope. Many interesting
factorization algebras — such as the Kac-Moody factorization algebra, and the
factorization algebra associated to a free field theory — can be constructed from a
variant of the factorization envelope construction, which we call the twisted factor-
ization envelope.
6.3.1 Definition. Let L be a sheaf of dg Lie algebras on M. A k-shifted central
extension of Lc is a precosheaf of dg Lie algebras L
ec fitting into an exact sequence
0 → C[k] → L
ec → Lc → 0
of precosheaves, where C[k] is the constant presheaf that assigns the one-dimensional
vector space C[k] in degree −k to every open.
If L is a local Lie algebra, we require in addition that locally there is a splitting
ec (U) = C[k] ⊕ Lc (U)
L
such that the differential and bracket maps from Lc (U) → C[k] and Lc (U)⊗2 →
C[k] are continuous.
6.3.2 Definition. In this situation, the twisted factorization envelope is the prefac-
torization algebra Ue L that sends an open set U to C∗ (L ec (U)). (In the case that L
is a local dg Lie algebra, we use the completed tensor product as above.)
The chain complex C∗ (L ec (U)) is a module over chains on the Abelian Lie al-
gebra C[k] for every k. Thus, we will view the twisted factorization envelope as a
prefactorization algebra in modules for C[c] where c has degree −k − 1.
Under the assumption that L is a homotopy sheaf, Theorem 6.0.1 shows that
the twisted factorization envelope UL e is a factorization algebra over the base ring
C[c]. Of particular interest is the case when k = −1, so that the central parameter c
is of degree 0.
Let us now introduce some important examples of this construction.
Example: Let g be a Lie algebra and consider the local Lie algebra Ω∗R ⊗ g on the
real line R, which we will denote gR . Given a skew-symmetric, invariant bilinear
form ω on g, there is a natural shifted extension of gRc where
Z
[α ⊗ X, β ⊗ Y]ω = α ∧ β ⊗ [X, Y] + α ∧ β ω(X, Y) c,
R
where we use c to denote the generator in degree 1 of the central extension. Let
Uω gR denote the twisted factorization envelope for this central extension. By mim-
icking the proof of Lemma 4.0.1, one can see that the cohomology of this twisted
factorization envelope recovers the enveloping algebra Ub g of the central extension
of g given by ω. ^
Example: Let g be a simple Lie algebra, and let h−, −ig denote a symmetric invari-
ant pairing on g. We define the Kac-Moody factorization algebra as follows.
Let Σ be a Riemann surface, and consider the local Lie algebra Ω0,∗ ⊗ g on Σ,
which sends an open subset U to the dg Lie algebra Ω0,∗ (U) ⊗ g. The differential
here is ∂,
¯ so we are describing the Dolbeault analog of the de Rham construction
in Section 4.
7. EQUIVARIANT PREFACTORIZATION ALGEBRAS 65
There is a −1-shifted central extension of Ω0,∗

c ⊗ g defined by the cocycle
Z
ω(α, β) = hα, ∂βig
U
where α, β ∈ Ωc0,∗ (U)⊗ g and ∂ : Ω0,∗ → Ω1,∗ is the holomorphic de Rham

operator. Note that this is a −1-shifted cocycle because ω(α, β) is zero unless
deg(α) + deg(β) = 1.
The twisted factorization envelope Uω (Ω0,∗ ⊗g) is the Kac-Moody factorization
algebra. Locally, it recovers the Kac-Moody vertex algebra, just as the real one-
dimensional case recovers the enveloping algebra. (We show this in Section 4.) ^
Example: In this example we will define a higher-dimensional analog of the Kac-
Moody vertex algebra.
Let X be a complex manifold of dimension n. Let Ω∗ (X) denote the de Rham
complex. Let φ ∈ Ωn−1,n−1 (X) be a closed form.
Then, given any Lie algebra g equipped with an invariant pairing h−, −ig , we
e of L = Ω0,∗
can construct a −1-shifted central extension L c ⊗ g, defined as above
by the cocycle Z
ω : α ⊗ β 7→ hα, ∂βig ∧ φ,
X
It is easy to verify that ω is a cocycle. (The case of the Kac-Moody extension
is when n = 1 and φ is a constant.) The cohomology class of this cocycle is
unchanged if we change φ to φ + ∂ψ, where ψ ∈ Ωn−1,n−2 (X) satisfies ∂ψ = 0.
The twisted factorization envelope ULe = Uω L is closely related to the Kac-
Moody algebra. For instance, if X = Σ × Pn−1 where Σ is a Riemann surface, and
the form φ is a volume form on Pn−1 , then the push-forward of this factorization
algebra to Σ is quasi-isomorphic to the Kac-Moody factorization algebra described
above. (Let p : X → Σ denote the projection map. The push-forward is defined by
(p∗ F )(U) = F (p−1 (U))
for each U ⊂ Σ.)
There is an important special case of this construction. Let X be a complex
surface (i.e., dimC (X) = 2) and the form φ is the curvature of a connection on the
canonical bundle of X (i.e., φ represents c1 (X)). As we will show when we discuss
Noether’s theorem at the quantum level, if we have a field theory with an action of
a local dg Lie algebra L then a twisted factorization envelopes of L will map to
the factorization algebra of observables of the theory. One can show that the local
dg Lie algebra Ω0,∗
X ⊗ g acts on a twisted N = 1 gauge theory with matter, and that
the twisted factorization envelope — with central extension determined by c1 (X)
— maps to observables of this theory (following Johansen (1995)). ^
7. Equivariant prefactorization algebras

Let M be a topological space with an action of a group G by homeomorphisms.
For g ∈ G and U ⊂ M, we use gU to denote the subset {gx | x ∈ U} ⊂ M. We
will formulate what it means to have a G-equivariant prefactorization algebra on
M. When M is a manifold and G is a Lie group acting smoothly on M, we will

formulate the notion of a smoothly G-equivariant prefactorization algebra on M.
7.1. Discrete equivariance. We begin with the case where G is viewed sim-
ply as a group (i.e., we do not require any compatibility between the action and a
possible smooth structure on G). We give a concrete definition.
7.1.1 Definition. Let F be a prefactorization algebra on M. We say F is G-
equivariant if for each g ∈ G and each open subset U ⊂ M we are given iso-
morphisms

σg : F (U) −
→ F (gU)
satisfying the following conditions.
((i)) For all g, h ∈ G and all opens U, σg ◦ σh = σgh : F (U) → F (ghU).
((ii)) Every σg respects the factorization product. That is, for any finite tuple
U1 , . . . , Uk of disjoint opens contained in an open V, the diagram
F (U1 ) ⊗ · · · ⊗ F (Un ) / F (gU1 ) ⊗ · · · ⊗ F (gUn )

F (V) / F (gV)
commutes.
There is a more categorical way to phrase this definition. As G acts contin-

uously, each element g ∈ G provides an endofunctor of the category Opens M of
opens in M. Indeed, g provides an endofunctor of Disj M and hence, via precom-
position, provides an endofunctor on the category of algebras over Disj M : it sends
a prefactorization algebra A to a prefactorization algebra gA, where gA(U) =
A(gU). An equivariant prefactorization algebra is then a collection of isomor-
phisms σg : A → gA of prefactorization algebras such that (hσg ) ◦ σh = σgh .
7.2. A useful reinterpretation. We want to be able to talk about a prefactor-

ization algebra F that is equivariant with respect to the smooth action on a manifold
M of a Lie group G. In particular, we need to formulate what it means to give a
“smooth” action of G on a prefactorization algebra.
As a first step, we will rework the notion of equivariant prefactorization alge-
bra. We will start by constructing a colored operad for each group G. An algebra
over this colored operad will recover the preceding definition of a G-equivariant
prefactorization algebra.
7.2.1 Definition. Let DisjGM denote the colored operad where the set of colors is
the set of opens in M and where the operations DisjGM (U1 , . . . , Un | V) are given by
the set
n o
(g1 , . . . , gn ) ∈ Gn | ∀i, gi Ui ⊂ V and ∀i , j, gi Ui ∩ g j U j = ∅
for any choice of inputs U1 , . . . , Un and output V.
There is a map of colored operads Disj M → DisjGM , sending each open to itself
and sending a nonempty operation U1 , . . . , Un → V to the identity in Gn . Hence,
given an algebra A over DisjGM , there is an “underlying” prefactorization algebra
on M.
Next, observe that for an open set U, the set DisjGM (U | gU) is the coset g ·
Stab(U), where the stabilizer subgroup Stab(U) ⊂ G consists of elements in G
that preserve U. Hence for any algebra A over DisjGM , we see that there is an

isomorphism σg : A(U) → A(gU) for every g ∈ G and that they compose in the
natural way. Hence the “underlying” prefactorization algebra is G-equivariant.
Now observe that every operation in DisjGM factors as a collection of unary
operations σg arising from the G-action followed by a operation from the “under-
lying” prefactorization algebra. (If the input opens Ui already sit in the output open
V, we are done. Otherwise, pick elements gi of G that move the input opens inside
V and keep them pairwise disjoint.) Hence we obtain the following lemma.
7.2.2 Lemma. For G a group, every G-equivariant prefactorization algebra on M
produces a unique algebra over DisjGM , and conversely.
Some notation will make it easier to understand how the G-action intertwines
with the structure of the prefactorization algebra A. If (g1 , . . . , gn ) ∈ DisjGM (U1 , . . . , Un |V),
then we denote the associated operation by
m(g1 ,...,gk ) : A(U1 ) ⊗ · · · ⊗ A(Uk ) → A(V).
It can understood as the composition
O ⊗σgi O
A(Ui ) −−−→ A(gi Ui ) → A(V),
i i
where the second map is the structure map of the prefactorization algebra and the
first map is given by the unary operations arising from the action of G on M.
7.3. Smooth equivariance. From now on, we focus upon the situation where
G is a Lie group acting smoothly on a smooth manifold M. In this situation, an
algebra A over DisjGM has an underlying prefactorization algebra on M but now we
want the operations to vary smoothly as the input opens are moved by elements of
G. To accomplish this, we need to equip with sets of operations with a “smooth
structure” and we need the target category, in which an algebra takes values, to
admit a notion of “smoothly varying family of multilinear maps.”
Throughout this book, our prefactorization algebras take values in the category
DVS of differentiable vector spaces or in the category Ch(DVS) of cochain com-
plexes of differentiable vector spaces. We view these categories as enriched over
themselves. In the case of DVS, the self-enrichement is discussed in Section 5.6.
If V, W are differentiable vector spaces, we denote by HomDVS (V, W) the differen-
tiable vector space of maps from V to W. The key feature of HomDVS (V, W) is that
its value on a point is DVS(V, W) and, more generally, its value on a manifold X is
smooth families over X of maps of differentiable vector spaces from V to W. The
self-enrichment of DVS leads, in an obvious way, to a self-enrichment of Ch(DVS).
7.3.1. The colored operad. Equipping each set of operations DisjGM (U1 , . . . , Un |V)
with some kind of smooth structure is a little subtle. We might hope such a set,
which is a subset of Gn , inherits a manifold structure from Gn , but it is easy to pro-
duce examples of opens Ui and V where the set of operations is far from being a
submanifold of Gn . We take the following approach instead. Given a subset S ⊂ X
of a smooth manifold, there is a set of smooth maps Y → X that factor through S
for any smooth manifold Y. In other words, S provides a sheaf of sets on the site
Mfld of smooth manifolds, sometimes known as a “generalized manifold.” (See
Definition 2.0.1 for the site.)
Let Shv(Mfld) denote the category of sheaves (of sets). Then every collection
of operations
DisjGM (U1 , . . . , Un | V)
provides such a sheaf, so we naturally obtain a colored operad enriched in “gener-
alized manifolds,” i.e., Shv(Mfld).
g GM denote the colored operad where the set of colors is the
7.3.1 Definition. Let Disj
g GM (U1 , . . . , Un | V) are the sheaves
set of opens in M and where the operations Disj
on Mfld determined by the subset
DisjGM (U1 , . . . , Un | V) ⊂ Gn .
G
Let us unpack what this definition means. An algebra F over Disjg M in DVS
associates to each open U ⊂ M, a differentiable vector space F (U). To each finite
collection of opens U1 , . . . , Un , V, we have a map of sheaves
g GM (U1 , . . . , Un | V) → HomDVS (F (U1 ), . . . , F (Un ) | F (V)).
Disj
Thus, for each smooth manifold Y and for each smooth map Y → Gn factoring
through DisjGM (U1 , . . . , Un | V), we obtain a section in
C ∞ (Y, HomDVS (F (U1 ), . . . , F (Un ) | F (V)),
which encodes a Y-family of multilinear morphisms from the F (Ui ) to F (V).
Note that we can evaluate each sheaf Disjg GM (U1 , . . . , Un | V) at the point ∗ ∈
Mfld to obtain the underlying set DisjGM (U1 , . . . , Un | V). Thus, each algebra over
g GM has an underlying G-equivariant prefactorization algebra.
Disj
An algebra over this new colored operad is very close to what we need for our
purposes in this book. What’s missing so far is the ability to differentiate the action
of G on the prefactorization algebra to obtain an action of the Lie algebra g.
Ideally, this action of g would simply exist. Instead, we will put it in by hand,
as data, since that suffices for our purposes. After giving our definition, however,
we will indicate a condition on an algebra over DisjGM that provides the desired
action automatically.
7.3.2. Derivations. First, we need to introduce the notion of a derivation of a
prefactorization algebra on a manifold M. We will construct a differential graded
Lie algebra of derivations of any prefactorization algebra.
Recall that a derivation of an associative algebra A is a linear map D : A → A

that intertwines in an interesting way with the multiplication map:
D(m(a, b)) = m(Da, b) + m(a, Db),
a relation known as the Leibniz rule. We simply write down the analog for an
algebra over the colored operad Disj M taking values in cochain complexes.
7.3.2 Definition. A degree k derivation of a prefactorization algebra F is a collec-
tion of maps DU : F (U) → F (U) of cohomological degree k for each open subset
U ⊂ M, with the property that for any finite collection of pairwise disjoint opens
U1 , . . . , Un , all contained in an open V, and an element αi ∈ F (Ui ) for each open,
the derivation acts by a Leibniz rule on the structure maps:
1 ,...,U n 1 ,...,U n
X
DV mU V (α1 , . . . , α n ) = (−1)k(|α1 |+···+|αi−1 |) mU
V (α1 , . . . , DUi αi , . . . , αn ),
i
where the sign is determined by the usual Koszul rule of signs.
Let Derk (F ) denote the derivations of degree k. It is easy to verify that the
derivations of all degrees Der∗ (F ) forms a differential graded Lie algebra. The
differential is defined by (dD)U = [dU , DU ], where dU is the differential on F (U).
The Lie bracket is defined by [D, D0 ]U = [DU , D0U ].
The concept of derivation allows us to talk about the action of a dg Lie algebra
g on a prefactorization algebra F . Such an action is simply a homomorphism of
differential graded Lie algebras from g to Der∗ (F ).
7.3.3. The main definition. We now provide the main definition. As our pref-
actorization algebra F takes values in the category of differentiable cochain com-
plexes, it makes sense to differentiate an element of F (U) for any open U.
7.3.3 Definition. A smoothly G-equivariant prefactorization algebra on M is an
algebra F over Disj g GM and an action of the Lie algebra g of G on the under-
lying prefactorization algebra of F such that for every X ∈ g, every operation
g GM (U1 , . . . , Un | V), and every 1 ≤ i ≤ n, we have
(g1 , . . . , gn ) ∈ Disj
∂
mg ,...,g (α1 , . . . , αn ) = mg1 ,...,gn (α1 , . . . , X(αi ), . . . , αn ).
∂Xi 1 n
∂
On the left hand side, ∂X i
indicates the action of the left-invariant vector field on
G associated to X in the ith factor of Gk and zero in the remaining factors.
k
Remark: In some cases, an algebra A over Disj g GM should possess a natural action of
g. We want to recover how g acts on each value A(U) from the action of G on A.
Suppose V is an open such that gU ⊂ V for every g in some neighborhood of the
identity in G. Then we can differentiate the structure maps mg : A(U) → A(V) to
obtain a map X : A(U) → A(V) for every X ∈ g. If A(U) = limV⊃U A(V), we ob-
tain via this limit, an action of g on A(U). Hence, if we have A(U) = limV⊃U A(V)
for every open U, we obtain an action of g on the underlying prefactorization alge-

bra of A. (The compatibility of the G-action with the structure ensures that we get
derivations of A.) ^
Here is an example we will revisit.
Example: Let F be a locally constant, smoothly translation invariant factorization
algebra on R, valued in vector spaces. Hence, A = F ((0, 1)) has the structure of an
associative algebra.
Being locally constant means that for any two intervals (0, 1) and (t, t + 1),
there is an isomorphism F ((0, 1)) F ((t, t + 1)) coming from the isomorphism
F ((a, b)) → F (R) associated to inclusion of an interval into R. As F is trans-
lation invariant, there is another isomorphism F ((0, 1)) → F ((t, t + 1)) for any
t ∈ R. Composing these two isomorphism yields an action of the group R on
A = F ((0, 1)). One can check that this is an action on associative algebras, not just
vector spaces.
The fact that F is both smoothly translation-invariant and locally constant
means that the action of R on A is smooth, and thus it differentiates to an infin-
∂
itesimal action of the Lie algebra R on A by derivations. The basis element ∂x of R
becomes a derivation H of A, called the Hamiltonian.
In the case that F is the cohomology of the factorization algebra of observables
of the free scalar field theory on R with mass m, we will see in Section 3 that
the algebra A is the Weyl algebra, generated by p, q, ~ with commutation relation
[p, q] = ~. The Hamiltonian is given by H(a) = 2~ 1
[p2 − m2 q2 , a]. ^
Part 2
First examples of field theories and their

observables
CHAPTER 4
Free field theories
1. The divergence complex of a measure

In this section, we will revisit the ideas and constructions from Chapter 2. Re-
call that in that chapter, we studied the divergence operator associated to a Gaussian
measure on a finite-dimensional vector space and then generalized this construc-
tion to the infinite-dimensional vector spaces that occur in field theory. We used
these ideas to define a vector space H 0 (Obsq ) of quantum observables.
Our goal here is to lift these ideas to the level of cochain complexes. We will
find a cochain complex Obsq such that H 0 (Obsq ) is the vector space we constructed
in Chapter 2. To make the narrative as clear as possible, we will recapitulate our
approach there.
1.1. The divergence complex of a finite-dimensional measure and its clas-

sical limit. We start by considering again Gaussian integrals in finite dimensions.
Let M be a smooth manifold of dimension n, and let ω0 be a smooth measure
on M. Let f be a function on M. (For example, M could be a vector space, ω0
the Lebesgue measure and f a quadratic form.) The divergence operator for the
measure e f /~ ω0 is a map
Div~ : Vect(M) → C ∞ (M)
X 7→ ~−1 (X f ) + Divω0 X.
One way to describe the divergence operator is to contract with the volume form
e f /~ ω0 to identify Vect(M) with Ωn−1 (M) and C ∞ (M) with Ωn (M). Explicitly,
the contraction of a function φ with e f /~ ω0 is the volume form φe f /~ ω0 , and the
contraction of a vector field X with e f /~ ω0 is the n − 1-form ιX e f /~ ω0 . Under
this identification, the divergence operator is simply the de Rham operator from
Ωn−1 (M) to Ωn (M).
The de Rham operator, of course, is part of the de Rham complex. In a similar
way, we can define the divergence complex, as follows. Let
PVi (M) = C ∞ (M, ∧i T M)
denote the space of polyvector fields on M. The divergence complex is the complex
Div~ Div~
· · · → PVi (M) −−−→ PV1 (M) −−−→ PV0 (M)
where the differential
Div~ : PVi (M) → PVi−1 (M)
73
74 4. FREE FIELD THEORIES
is defined so that the diagram
∨e f /~ ω0
PVi (M) / Ωn−i (M)
Div~ ddR

∨e f /~ ω0
PVi−1 (M) / Ωn−i+1 (M)
commutes. Here we use ∨e f /~ ω0 to denote the contraction map; it is the natural ex-
tension to polyvector fields of the contractions we defined for functions and vector
fields.
In summary, after contracting with the volume e f /~ ω0 , the divergence complex
becomes the de Rham complex. It is easy to check that, as maps from PVi (M) to
PVi−1 (M), we have
Div~ = ∨~−1 d f + Divω0
where ∨d f is the operator of contracting with the 1-form d f . In the ~ → 0 limit,
the dominant term is ∨~−1 d f .
More precisely, by multiplying the differential by ~, we see that there is a flat
family of cochain complexes over the algebra C[~] with the properties that
((i)) the family is isomorphic to the divergence complex when ~ , 0, and
((ii)) at ~ = 0, the cochain complex is
∨d f ∨d f
· · · → PV2 (M) −−−→ PV1 (M) −−−→ PV0 (M).
Note that this second cochain complex is a differential graded algebra, which is not
the case for the divergence complex.
The image of the map ∨d f : PV1 (M) → PV0 (M) is the ideal cutting out the
critical locus. Indeed, this whole complex is the Koszul complex for the equations
cutting out the critical locus. This observation leads to the following definition.
1.1.1 Definition. The derived critical locus of f is the locally dg ringed space
whose underlying manifold is M and whose dg commutative algebra of functions
is the complex PV∗ (M) with differential ∨d f .
Remark: We will call a manifold with a nice sheaf of dg commutative algebras a dg

manifold. Since the purpose of this section is motivational, we will not develop a
theory of such dg manifolds. We are using the concrete object here as a way to think
about the derived geometry of this situation. (More details on derived geometry,
from a different point of view, will be discussed in Volume 2.) The crucial point is
that the dg algebra keeps track of the behavior of d f , including higher homological
data. ^
Here is a bit more motivation for our terminology. Let Γ(d f ) ⊂ T ∗ M denote
the graph of d f . The ordinary critical locus of f is the intersection of Γ(d f ) with
the zero-section M ⊂ T ∗ M. The derived critical locus is defined to be the derived
intersection. In derived geometry, functions on derived intersections are defined by
1. THE DIVERGENCE COMPLEX OF A MEASURE 75
derived tensor products:

C ∞ (Crith ( f )) = C ∞ (Γ(d f )) ⊗CL ∞ (T ∗ M) C ∞ (M).
By using a Koszul resolution of C ∞ (M) as a module for C ∞ (T ∗ M), one finds a
quasi-isomorphism of dg commutative algebras between this derived intersection
and the complex PV∗ (M) with differential ∨d f .
In short, we find that the ~ → 0 limit of the divergence complex is the dg
commutative algebra of functions on the derived critical locus of f .
An important special case of this relationship is when the function f is zero. In
that case, the derived critical locus of f has as functions the algebra PV∗ (M) with
zero differential. We view this algebra as the functions on the graded manifold
T ∗ [−1]M. The derived critical locus for a general function f can be viewed as a
deformation of T ∗ [−1]M obtained by introducing a differential ∨d f .
1.2. A different construction. We will define the prefactorization algebra of

observables of a free scalar field theory as a divergence complex, just like we de-
fined H 0 of observables to be given by functions modulo divergences in Chapter
2. It turns out that there is a slick way to write this prefactorization algebra as a
twisted factorization envelope of a certain sheaf of Heisenberg Lie algebras. We
will explain this point in a finite-dimensional toy model, and then we will use the
factorization envelope picture to define the prefactorization algebra of observables
of the field theory in the next section.
Let V be a vector space, and let q : V → R be a quadratic function on V. Let ω0
be the Lebesgue measure on V. We want to understand the divergence complex for
the measure eq/~ ω0 . The construction is quite general: we do not need to assume
that q is non-degenerate.
The derived critical locus of the function q is a linear dg manifold. (The Ja-
cobi ideal is generated by linear equations.) Linear dg manifolds are equivalent to
cochain complexes: any cochain complex B gives rise to the linear dg manifold
whose functions are the symmetric algebra on the dual of B.
The derived critical locus of q is described by the cochain complex W given by
∂/∂q
V −−−→ V ∗ [−1],
∂
where the differential sends v ∈ V to the linear functional ∂q v = q(v, −).
Note that W is equipped with a graded anti-symmetric pairing h−, −i of co-
homological degree −1, defined by pairing V and V ∗ . In other words, W has a
symplectic pairing of cohomological degree −1. We let
HW = C · ~[−1] ⊕ W,
where C · ~ indicates a one-dimensional vector space with basis ~. We give the
cochain complex HW a Lie bracket by saying that
[w, w0 ] = ~ w, w0 .
Thus, HW is a shifted-symplectic version of the Heisenberg Lie algebra of an ordi-
nary symplectic vector space.
Consider the Chevalley-Eilenberg chain complex C∗ HW . This cochain com-

plex is defined to be the symmetric algebra of the underlying graded vector space
HW [1] = W[1] ⊕ C · ~ but equipped with a differential determined by the Lie
bracket. Since the pairing on W identifies W[1] = W ∗ , we can identify
C∗ (HW ) = (Sym W ∗ [~], dCE )

where dCE denotes the differential. Since, as a graded vector space, W = V ⊕

V ∗ [−1], we have a natural identification
Cn (HW ) = PVn (V)[~]
where PV∗ (V) refers to polyvector fields on V with polynomial coefficients and
where, as before, we place PVn (V) in degree −n.
1.2.1 Lemma. The differential on C∗ (HW ) is, under this identification, the opera-
tor
~ Diveq/~ ω0 : PVi (V)[~] → PVi−1 (V)[~],
where ω0 is the Lebesgue measure on V, and q is the quadratic function on V used
to define the differential on the complex W.
Proof. The proof is an explicit calculation, which we leave to the interested
reader. The calculation is facilitated by choosing a basis of V in which we can
explicitly write both the divergence operator and the differential on the Chevalley-
Eilenberg complex C∗ (HW ).
In what follows, we will define the prefactorization algebra of observables of
a free field theory as a Chevalley-Eilenberg chain complex of a certain Heisenberg
Lie algebra, constructed as in this lemma.
2. The prefactorization algebra of a free field theory

In this section, we will construct the prefactorization algebra associated to any
free field theory. We will concentrate, however, on the free scalar field theory on
a Riemannian manifold. We will show that, for one-dimensional manifolds, this
prefactorization algebra recovers the familiar Weyl algebra, the algebra of observ-
ables for quantum mechanics. In general, we will show how to construct correla-
tion functions of observables of a free field theory and check that these agree with
how physicists define correlation functions.
2.1. The classical observables of the free scalar field. We start with a stan-
dard example: the free scalar field. Let M be a Riemannian manifold with metric
g, so M is equipped with a natural density. We will use this natural density both
to integrate functions and also to provide an isomorphism between functions and
densities, and we will use this isomorphism implicitly from hereon.
The scalar field theory has smooth functions as fields, and we use the notation
φ ∈ C ∞ (M) for an arbitrary field. The action functional of the theory is
Z
S (φ) = φ4φ,
M
2. THE PREFACTORIZATION ALGEBRA OF A FREE FIELD THEORY 77
where 4 is the Laplacian on M. (Normally we will reserve the symbol 4 for the
Batalin-Vilkovisky Laplacian, but that’s not necessary in this section.) This func-
tional is not well-defined on a field φ unless φ4φ is integrable, but it is helpful to
bear in mind that for classical field theory, what is crucial is the Euler-Lagrange
equations, which in this case is 4φ = 0 and which is thus well-defined on all
smooth functions. The action can be viewed here as a device for producing these
partial differential equations.
If U ⊂ M is an open subset, then the space of solutions to the equation of
motion on U is the space of harmonic functions on U. In this book, we will always
consider the derived space of solutions of the equation of motion. (For more details
about the derived philosophy, the reader should consult Volume 2.) In this simple
situation, the derived space of solutions to the free field equations, on an open
subset U ⊂ M, is the two-term complex
4

E (U) = C ∞ (U) −→ C ∞ (U)[−1] ,
where the complex is concentrated in cohomological degrees 0 and 1. (The bracket

[−1] denotes “shift up by 1.”)
The observables of this classical field theory are simply the functions on this
derived space of solutions to the equations of motion. As this derived space is a
cochain complex (and hence linear in nature), it is natural to work with the polyno-
mial functions. (One could work with more complicated types of functions, but the
polynomial functions provide a concrete and useful collection of observables.) To
be explicit, the classical observables are the symmetric algebra on the dual space
to the fields.
Let’s make this idea precise, using the technology we introduced in Section 5.
The space E (M) has the structure of differentiable cochain complex (essentially, it
is a sheaf of vector spaces on the site of smooth manifolds). We define the space
of polynomial functions homogeneous of degree n on E (M) to be the space
Pn (E (M)) = HomDVS (E (M) × · · · × E (M), R)S n .
In other words, we consider smooth multilinear maps from n copies of E (M), and
then we take the S n -coinvariants. The algebra of all polynomial functions on E (M)
is the space P(E (M)) = ⊕n Pn (E (M)).
As we discussed in Section 5, we can identify
Pn (E (M)) = Dc (M n , (E ! )n )S n
as the S n -coinvariants of the space of compactly supported distributional sections
of the bundle (E ! )n on M n . In general, if E (M) is sections of a graded bundle E,
then E ! is E ∨ ⊗ Dens. In the case at hand, the bundle E ! is two copies of the trivial
bundle, one in degree −1 and one in degree 0.
For example, the space P1 (E (M)) = E (M)∨ of smooth linear functionals on
E (M) is the space
4
E (M)∨ = Dc (M)−1 −
→ Dc (M)0 ,
where Dc (M) indicates the space of compactly supported distributions on M.
We also want to keep track of where measurements are taking place on the
manifold M, so we will organize the observables by where they are supported. The
classical observables with support in U ⊂ M are then the symmetric algebra of
4

E (U)∨ = Dc (U)[1] −→ Dc (U) ,
where Dc (U) indicates the space of compactly supported distributions on U. The

complex E (U)∨ is thus the graded, smooth linear dual to the two-term complex
E (U) above. Note that, as the graded dual to E , this complex is concentrated in
cohomological degrees 0 and −1. These are precisely the observables that only
depend on the behavior of the field φ on the open set U.
Thus, as a first pass, one would want to define the classical observables as the
symmetric algebra on E (U)∨ . This choice leads, however, to difficulties defining
the quantum observables. When we work with an interacting theory, these diffi-
culties can only be surmounted using the techniques of renormalization. For a free
field theory, though, there is a much simpler solution, as we now explain.
Recall from Section 5 that E ! denotes the vector bundle E ∨ ⊗ Dens M . Our
identification between densities and functions then produces an isomorphism
Ec! (U) Cc∞ (U)[1] → Cc∞ (U) ,

for compactly supported sections of E ! . Note that there is a natural map of cochain
complexes Ec! (U) → E (U)∨ , given by viewing a compactly supported function as
a distribution.
2.1.1 Lemma. The inclusion map Ec! (U) → E (U)∨ is a cochain homotopy equiv-
alence of differentiable cochain complexes.
Proof. This assertion is a special case of a general result proved in Appen-
dix D. Note that by differentiable homotopy equivalence we mean that there is an
“inverse” map E (U)∨ → Ec! (U), and differentiable cochain homotopies between
the two composed maps and the identity maps. “Differentiable” here means that
all maps are in the category DVS of differentiable vector spaces. As these are
convenient cochain complexes, suffices to construct a continuous homotopy equiv-
alence.
This lemma says that, since we are working homotopically, we can replace
a distributional linear observable by a smooth linear observable. In other words,
any distributional observable that is closed in the cochain complex E (U)∨ is chain
homotopy equivalent to a closed smooth observable.We think of the smooth linear
observables as “smeared.” For example, we can replace a delta function δ x by some
bump function supported near the point x.
The observables we will work with is the space of “smeared observables”,
defined by
4
Obscl (U) = Sym(Ec! (U)) = Sym(Cc∞ (U)[1] −
→ Cc∞ (U)),
the symmetric algebra on Ec! (U). As we explained in Section 5, this symmetric

algebra is defined using the natural symmetric monoidal structure on the full sub-
category CVS ⊂ DVS of convenient vector spaces. Concretely, we can identify
Symn (Ec! (U)) = Cc∞ (U n , (E ! )⊗n )S n .
In other words, Symn Ec! (U) is the subspace of Pn (E (U)) defined by taking all
distributions to be smooth functions with compact support.
2.1.2 Lemma. The map Obscl (U) → P(E (U)) is a homotopy equivalence of cochain
complexes of differentiable vector spaces.
Proof. It suffices to show that the map Symn Ec! (U) → Pn (E (U)) is a differen-
tiable homotopy equivalence for each n. But for this, it suffices to observe that the
map
Cc∞ (U, (E ! )⊗n ) → Dc (U, (E ! )⊗n )
is an S n -equivariant differentiable homotopy equivalence.
In parallel to the previous lemma, this lemma says that, since we are working
homotopically, we can replace a distributional polynomial observable (given by
integration against some distribution on U n ) by a smooth polynomial observable
(given by integration against a smooth function on U n ).
2.2. Interpreting this construction. Let us describe the cochain complex
Obscl (U) more explicitly, in order to clarify the relationship with what we dis-
cussed in Chapter 2. The complex Obscl (U) looks like
· · · → ∧2Cc∞ (U) ⊗ Sym Cc∞ (U) → Cc∞ (U) ⊗ Sym Cc∞ (U) → Sym Cc∞ (U).
All tensor products appearing in this expression are completed tensor products in
the category of convenient vector spaces.
We should interpret Sym Cc∞ (U) as being an algebra of polynomial functions
on C ∞ (U), using (as we explained above) the Riemannian volume form on U to
identify Cc∞ (U) with a subspace of the dual of C ∞ (U). There is a similar interpre-
tation of the other terms in this complex using the geometry of the space C ∞ (U) of
fields. Let T cC ∞ (U) refer to the subbundle of the tangent bundle of C ∞ (U) given
by the subspace Cc∞ (U) ⊂ C ∞ (U). An element of a fibre of T cC ∞ (U) is a first-order
variation of a field which is zero outside of a compact set. This subspace T cC ∞ (U)
defines an integrable foliation on C ∞ (U), and this foliation can be defined if we
replace C ∞ (U) by any sheaf of spaces.
Then, we can interpret Cc∞ (U) ⊗ Sym Cc∞ (U) as a space of polynomial sections
of T cC ∞ (U). Similarly, ∧k Cc∞ (U) ⊗ Sym Cc∞ (U) should be interpreted as a space
of polynomial sections of the bundle ∧k T cC ∞ (U).
That is, if PVc (C ∞ (U)) refers to polynomial polyvector fields on C ∞ (U) along
the foliation given by T cC ∞ (U), we have
Obscl (U) = PVc (C ∞ (U)).
So far, this is just an identification of graded vector spaces. We need to explain
how to identify the differential. Roughly speaking, the differential on Obscl (U)
corresponds to the differential on PVc (C ∞ (U)) obtain this complex is given by

contracting with the one-form dS for the function
Z
S (φ) = 12 φ4φ,
the action functional. Before making this idea precise, we recall the finite-dimensional
model of a quadratic function Q(x) = (x, Ax) on a vector space V. The contrac-
tion of the 1-form dQ with a tangent vector v0 ∈ T 0 V gives the linear functional
x 7→ (v0 , Ax). As S is quadratic in φ, we expect that the contraction of dS with a
tangent vector φ0 inC ∞ (U) is the linear functional
Z
φ 7→ 21
φ0 4φ.
But we run into an issue here: the functional S is not well-defined for all fields φ,
because the integral may not converge, and similarly the linear functional above is
not well-defined for arbitrary smooth functions φ0 and φ. However, the expression
∂S
Z
(φ) = 12 φ0 4φ
∂φ0
does make sense for any φ ∈ C ∞ (U) and φ0 ∈ Cc∞ (U). (Note that we now only
consider tangent vectors φ0 with compact support.) In other words, the desired
one-form dS does not make sense as a section of the cotangent bundle of C ∞ (U),
but it is well-defined as a section of the space T c∗C ∞ (U), the dual of the subbun-
dle T cC ∞ (U) ⊂ TC ∞ (U) describing vector fields along the leaves. This leafwise
one-form is closed. Such one-forms are the kinds of things we can contract with
elements of PVc (C ∞ (U)). The differential on PVc (C ∞ (U)) thus matches the differ-
ential on Obscl given by contracting with dS .
2.3. General free field theories. With this example in mind, we introduce a
general definition.
2.3.1 Definition. Let M be a manifold. A free field theory on M is the following
data:
((i)) A graded vector bundle E on M, whose sheaf of sections will be de-
noted E , and whose compactly supported sections will be denoted Ec .
((ii)) A differential operator d : E → E , of cohomological degree 1 and
square zero, making E into an elliptic complex.
((iii)) Let E ! = E ∨ ⊗ Dens M , and let E ! be the sections of E ! . Let d! be the
differential on E ! which is the formal adjoint to the differential on E .
Note that there is a natural pairing between Ec (U) and E ! (U), and this
pairing is compatible with differentials.
We require an isomorphism E → E ! [−1] compatible with differ-
entials, with the property that the induced pairing of cohomological
degree −1 on each Ec (U) is graded anti-symmetric.
The complex E (U) is the derived version of the solutions to the equations of
motion of the theory on an open subset U. More motivation for this definition is
presented in Volume 2.
Note that the equations of motion for a free theory are always linear, so that
the space of solutions is a vector space. Similarly, the derived space of solutions
of the equations of motion of a free field theory is a cochain complex, which is
a linear derived stack. The cochain complex E (U) should be thought of as the
derived space of solutions to the equations of motion on an open subset U. As we
explain in Volume 2, the pairing on Ec (U) arises from the fact that the equations
of motion of a field theory are not arbitrary differential equations, but describe the
critical locus of an action functional.
For example, for the free scalar field theory on a manifold M with mass m, we
have, as above,
4+m2
E = C ∞ (U) −−−−→ C ∞ (U)[−1].
∂ 2
Our convention is that 4 is a non-negative operator, so that on Rn , 4 = −
P
∂xi .
The pairing on Ec (U) is defined by
D E Z
φ0 , φ1 = φ0 φ1
M
for φk
in the graded piece C ∞ (U)[k].
As another example, let us describe Abelian Yang-Mills theory (with gauge
group R) in this language. Let M be a manifold of dimension 4. If A ∈ Ω1 (M) is
a connection on the trivial R-bundle on a manifold M, then the Yang-Mills action
functional applied to A is
Z Z
S Y M (A) = − 2
1
dA ∧ ∗dA = 2
1
A(d ∗ d)A.
M M
The equations of motion are that d ∗ dA = 0. There is also gauge symmetry, given
by X ∈ Ω0 (M), which acts on A by A → A + dX. The complex E describing this
theory is
d d∗d d
E = Ω0 (M)[1] →
− Ω1 (M) −−→ Ω3 (M)[−1] →
− Ω4 (M)[−2].
We explain how to derive this statement in Volume 2. For now, note that H 0 (E ) is
the space of those A ∈ Ω1 (M) which satisfy the Yang-Mills equation d ∗ dA = 0,
modulo gauge symmetry.
For any free field theory with cochain complex of fields E , we define the clas-
sical observables of the theory as
Obscl (U) = Sym(Ec! (U)) = Sym(Ec (U)[1]).
It is clear that classical observables form a prefactorization algebra (recall example
1.1). Indeed, Obscl (U) is a commutative differential graded algebra for every open
U. If U ⊂ V, there is a natural algebra homomorphism
iU cl cl
V : Obs (U) → Obs (V),
which on generators is the extension-by-zero map Cc∞ (U) → Cc∞ (V).
If U1 , . . . , Un ⊂ V are disjoint open subsets, the prefactorization structure map

is the continuous multilinear map
Obscl (U1 ) × · · · × Obscl (Un ) → Obscl (V)
Qn Ui
(α1 , . . . , αn ) 7→ i=1 iV αi .
The product denotes the product in the symmetric algebra on V. On a field φ, this
observable takes the value
Y n
iU
V αi (φ) = α1 (φ|U1 ) · · · αn (φ|Un ),
i
i=1
which is the product of the value that each observable αi takes on φ restricted to
Ui .
2.4. The one-dimensional case, in detail. This space is particularly simple
in dimension 1. Indeed, we recover the usual answer at the level of cohomology.
2.4.1 Lemma. If U = (a, b) ⊂ R is an interval in R, then the algebra of classical
observables for the free field with mass m ≥ 0 has cohomology
H ∗ (Obscl ((a, b))) = R[p, q],
the polynomial algebra in two variables.
Proof. The idea is the following. The equations of motion for free classical
mechanics on the interval (a, b) are that the field φ satisfies (4 + m2 )φ = 0. This
space is two dimensional, spanned by the functions {e±mx }, for m > 0, and by the
functions {1, x}, for m = 0. Classical observables are functions on the space of so-
lutions to the equations of motion. We would this expect that classical observables
are a polynomial algebra in two generators.
We need to be careful, however, because we use the derived version of the
space of solutions to the equations of motion. We will show that the complex
2
!
−1 4+m
Ec ((a, b)) = Cc ((a, b)) −−−−→ Cc ((a, b))
! ∞ ∞ 0
is homotopy equivalent to the complex R2 situated in degree 0. Since the algebra

Obscl ((a, b)) of observables is defined to be the symmetric algebra on this complex,
this will imply the result. Without loss of generality, we can take a = −1 and b = 1.
First, let us introduce some notation. We start in the complex E of fields. If
m = 0, let φq (x) = 1 and φ p (x) = x. If m > 0, let
φq (x) = 21 emx + e−mx ,

φ p (x) = 2m
1
emx − e−mx .

For any value of m, the functions φ p and φq are annihilated by the operator −∂2x +m2 ,
and they form a basis for the kernel of this operator. Further, φ p (0) = 1 and φq (0) =
0, whereas φ0p (0) = 0 and φ0q (0) = 1. Finally, φq = φ0p .
Define a map
π : Ec! ((−1, 1)) → R{p, q}
by Z Z
g 7→ π(g) = q g(x)φq dx + p g(x)φ p dx.
It is a cochain map, because if g = (4 + m2 ) f , where f has compact support, then

π(g) = 0. This map is easily seen to be surjective.
This map π says how to identify a linear observable on the field φ over the time
interval (−1, 1) into a linear combination of the “position” and “momentum.”
We need to construct a contracting homotopy on the kernel of π. That is, if
Ker πk ⊂ Cc∞ ((−1, 1)) refers to the kernel of π in cohomological degree k, we need
to construct an inverse to the differential
4+m2
Cc∞ ((−1, 1)) = Ker π−1 −−−−→ Ker π0 .
This inverse is defined as follows. Let G(x) ∈ C 0 (R) be the Green’s function for
the operator 4 + m2 . Explicitly, we have

 m2 e−m|x| if m > 0

G(x) = 

− 1 |x|

2 if m = 0.
Then (4 + m2 )G is the delta function at 0. The inverse map sends a function
f ∈ Ker π0 ⊂ Cc∞ ((−1, 1))
to Z
G ? f = G(x − y) f (y)dy.
y
The fact that f φq = 0 and f φ p = 0 implies that G ? f has compact support.

R R
The fact that G is the Green’s function implies that this operator is the inverse to
4 + m2 . It is clear that the operator of convolution with G is smooth (and even
continuous), so the result follows.
2.5. The Poisson bracket. We now return to the general case and construct
the P0 algebra structure on classical observables.
Suppose we have any free field theory on a manifold M, with complex of fields
E . Classical observables are the symmetric algebra Sym Ec (U)[1]. Recall that the
complex Ec (U) is equipped with an antisymmetric pairing of cohomological degree
−1. Thus, Ec (U)[1] is equipped with a symmetric pairing of degree 1.
2.5.1 Lemma. There is a unique smooth Poisson bracket on Obscl (U) of cohomo-
logical degree 1, with the property that
{α, β} = hα, βi
for any two linear observables α, β ∈ Ec (U)[1].
Recall that “smooth” means that the Poisson bracket is a smooth bilinear map
{−, −} : Obscl (U) × Obscl (U) → Obscl (U)
as defined in Section 5.
Proof. The argument we will give is very general, and it applies in any reason-
able symmetric monoidal category. Recall that, as stated in Section 5, the category
of convenient vector spaces is a symmetric monoidal category with internal Homs
and a hom-tensor adjunction.
Let A be a commutative algebra object in the category Ch(CVS) of convenient
cochain complexes, and let M be a dg A-module. Then we define Der(A, M) to
be the space of algebra homomorphisms φ : A → A ⊕ M that are the identity on
A modulo the ideal M. (Here A ⊕ M is given the “square-zero” algebra structure,
where the product of any two elements in M is zero and the product of an element
in A with one in M is via the module structure.)
Since the category of convenient cochain complexes has internal Homs, this
cochain complex Der(A, M) is again an A-module in Ch(CVS).
The commutative algebra
Obscl (U) = Sym Ec! (U)
is the initial commutative algebra in the category Ch(CVS) of convenient cochain
complexes equipped with a smooth linear cochain map Ec! (U) → Obscl (U). (We
are simply stating the universal property characterizing Sym.) It follows that for
any dg module M in Ch(CVS) over the algebra Sym Ec! (U),
Der(Sym Ec! (U), M) = HomDVS (Ec! (U), M).
A Poisson bracket on Sym Ec! (U) is, in particular, a biderivation. A biderivation
is something that assigns to an element of Sym Ec! (U) a derivation of the algebra
Sym Ec! (U). Thus, the space of biderivations is the space

Der Sym Ec! (U), Der Sym Ec! (U), Sym Ec! (U) .
What we have said so far thus identifies the space of biderivations with
HomDVS (Ec! (U), HomDVS (Ec! (U), Sym Ec! (U))),
or equivalently with
HomDVS (Ec! (U) ⊗ Ec! (U), Sym Ec! (U))
via the hom-tensor adjunction in the category Ch(CVS).
The Poisson bracket we are constructing corresponds to the biderivation which
is the pairing on Ec! (U) viewed as a map
Ec! (U) ⊗ Ec! (U) → R = Sym0 Ec! (U).
This biderivation is antisymmetric and satisfies the Jacobi identity. Since Poisson
brackets are a subspace of biderivations, we have proved both the existence and
uniqueness clauses.
Note that for U1 , U2 disjoint open subsets of V and for observables αi ∈
Obscl (Ui ), we have
{iU
V α1 , iV α2 } = 0
1 U2
in Obscl (V). That is, observables coming from disjoint open subsets commute with
respect to the Poisson bracket. This property of Obscl (U) is the definition of a P0
prefactorization algebra. (Later we will show in Section 5 that this prefactorization

algebra is actually a prefactorization algebra.)
In the case of classical observables of the free scalar field theory, we can think
of Obscl (U) as the space of polyvector fields on C ∞ (U) along the foliation of
C ∞ (U) given by the subspace Cc∞ (U) ⊂ C ∞ (U). The Poisson bracket we have
just defined is the Schouten bracket on polyvector fields.
2.6. The quantum observables of a free field theory. In Chapter 2, we con-

structed a prefactorization algebra that we called H 0 (Obsq ), the quantum observ-
ables of a free scalar field theory on a manifold. This space is defined as a space of
functions on the space of fields, modulo the image of a certain divergence operator.
The aim of this section is to lift this vector space H 0 (Obsq ) to a cochain complex
Obsq . As we explained in Section 1, this cochain complex will be the analog of the
divergence complex of a measure in finite dimensions.
In Section 1, we explained that for a quadratic function q on a vector space V,
the divergence complex for the measure eq/~ ω0 on V (where ω0 is the Lebesgue
measure) can be realized as the Chevalley-Eilenberg chain complex of a certain
Heisenberg Lie algebra. We follow this procedure in defining Obsq .
We construct the prefactorization algebra Obsq (U) as a twisted factorization
envelope of a sheaf of Lie algebras. Let
Ebc (U) = Ec (U) ⊕ R · ~
where R~ denotes the one-dimensional real vector space situated in degree 1 and
spanned by ~. We give Ebc (U) a Lie bracket by saying that, for α, β ∈ Ec (U),
[α, β] = ~ hα, βi .
Thus, Ebc (U) is a graded version of a Heisenberg Lie algebra, centrally extending
the Abelian dg Lie algebra Ec (U).
Let
Obsq (U) = C∗ (Ebc (U)),
where C∗ denotes the Chevalley-Eilenberg complex for the Lie algebra homology
of Ebc (U), defined using the tensor product b
⊗β on the category of convenient vector
spaces, as discussed in Section 5. Thus,

Obsq (U) = Sym Ebc (U)[1] , d

= Obscl (U)[~], d
where the differential arises from the Lie bracket and differential on Ebc (U). The
symbol [1] indicates a shift of degree down by one. (Recall that we always work
with cochain complexes, so our grading convention of C∗ is the negative of one
common convention.)
Remark: Those readers who are operadically inclined might notice that the Lie al-
gebra chain complex of a Lie algebra g is the E0 version of the universal enveloping
algebra of a Lie algebra. Thus, our construction is an E0 version of the familiar
construction of the Weyl algebra as a universal enveloping algebra of a Heisenberg

algebra. ^
Since this is an example of the general construction we discussed in Section 6,
we see that Obsq (U) has the structure of a prefactorization algebra.
As we discussed above in Section 2.2, we can view classical observables on an
open set U for the free scalar theory as a certain complex of polyvector fields on
the space C ∞ (U):
Obscl (U) = PVc (C ∞ (U)), ∨dS .

By PVc (C ∞ (U)) we mean polyvector fields along the foliation T cC ∞ (U) ⊂ TC ∞ (U)
of compactly-supported variants of a field. Concretely, the cochain complex Obscl (U)
is
∨dS ∨dS ∨dS
· · · −−−→ ∧2Cc∞ (U) ⊗ Sym Cc∞ (U) −−−→ Cc∞ (U) ⊗ Sym Cc∞ (U) −−−→ Sym Cc∞ (U),
⊗β .
where all tensor products appearing in this expression are b
In a similar way, the cochain complex of quantum observables Obsq (U) in-
volves just a modification of the differential. It is
∨dS +~ Div ∨dS +~ Div
. . . −−−−−−−−→ Cc∞ (U) ⊗ Sym Cc∞ (U) −−−−−−−−→ Sym Cc∞ (U).
The operator Div is the extension to all polyvector fields of the operator (see Defi-
nition 2.0.1) defined in Chapter 2 as a map from polynomial vector fields on C ∞ (U)
to polynomial functions. Thus, H 0 (Obsq (U)) is the same vector space (and same
prefactorization algebra) that we defined in Chapter 2.
As a graded vector space, there is an isomorphism
Obsq (U) = Obscl (U)[~],
but it does not respect the differentials. In particular, the differential d on the quan-
tum observables satisfies
((i)) modulo ~, d coincides with the differential on Obscl (U), and
((ii)) the equation
(2.6.1) d(a · b) = (da) · b + (−1)|a| a · b + (−1)|a| ~{a, b}.
Here, · indicates the commutative product on Obscl (U). As we will explain in Vol-
ume 2, these properties imply that Obsq defines a prefactorization algebra valued
in Beilinson-Drinfeld algebras and that Obsq quantizes the prefactorization algebra
Obscl valued in P0 algebras. (See Section 2.2 for the definition of these types of
algebras.) It is a characterizing feature of the Batalin-Vilkovisky formalism that
the Poisson bracket measures the failure of the differential d to be a derivation, and
the language of P0 and BD algebras is an operadic formalization of this concept.
We prove in Section 5 that Obscl is a factorization algebra, which implies Obsq is
a factorization algebra over R[~] .
3. QUANTUM MECHANICS AND THE WEYL ALGEBRA 87
3. Quantum mechanics and the Weyl algebra

We will now show that our construction of the free scalar field on the real line R
recovers the Weyl algebra, which is the associative algebra of observables in quan-
tum mechanics . We already showed in Section 2.4 that the classical observables
recovers functions on the phase space.
First, we must check that this prefactorization algebra is locally constant and
so gives us an associative algebra.
3.0.1 Lemma. The prefactorization algebra Obsq on R constructed from the free
scalar field theory with mass m is locally constant.
Proof. Recall that Obsq (U) is the Chevalley-Eilenberg chains on the Heisen-
4+m2
berg Lie algebra H(U) built as a central extension of Cc∞ (U) −−−−→ Cc∞ (U)[−1].
Let us filter Obsq (U) by saying that
F ≤k Obsq (U) = Sym≤k (H(U)[1]).
The associated graded for this filtration is Obscl (U)[~]. Thus, to show that H ∗ Obsq
is locally constant, it suffices to show that H ∗ Obscl is locally constant, by con-
sidering the spectral sequence associated to this filtration. We have already seen
that H ∗ (Obscl (a, b)) = R[p, q] for any interval (a, b), and that the inclusion maps
(a, b) → (a0 , b0 ) induces isomorphisms. Thus, the cohomology of Obscl is locally
constant, as desired.
It follows that the cohomology of this prefactorization algebra is an associative
algebra, which we call Am . We will show that Am is the Weyl algebra, no matter
the mass m.
In fact, we will show more, showing that the prefactorization algebra does
know the mass m, encoded in a Hamiltonian operator. The prefactorization alge-
bra for the free scalar field theory on R is built as Chevalley-Eilenberg chains of
4+m2 ∂
the Heisenberg algebra based on Cc∞ (U) −−−−→ Cc∞ (U)[−1]. The operator ∂x of
infinitesimal translation acts on Cc (U), commutes with the operator 4 + m , and
∞ 2
preserves the cocycle defining the central extension. Therefore, it acts naturally
on the Chevalley-Eilenberg chains of the Heisenberg algebra. One can check that
∂
this operator is a derivation for the factorization product. That is, the operator ∂x
commutes with inclusions of one open subset into another, and if U, V are disjoint
and α ∈ Obsq (U), β ∈ Obsq (V), we have
∂ ∂ ∂
(α · β) = (α) · β + α · (β) ∈ Obsq (U t V).
∂x ∂x ∂x
In other words, the prefactorization algebra is, in a sense, translation-invariant!
(Derivations, and this interpretation, are discussed in more detail in Section 8.)
∂
It follows immediately that ∂x defines a derivation of the associative algebra
Am coming from the cohomology of Obsq ((0, 1)).
∂
Remark: These observations about ∂x apply to the classical observables, too. In-
deed, one discovers that infinitesimal translation induces a derivation of the Poisson
algebra R[p, q]. As this derivation preserves the Poisson bracket, we know there is
an element H of R[p, q] such that {H, −} is the derivation. Here, H = p2 − m2 q2 .
(On a symplectic vector space, every symplectic vector field is Hamiltonian.) ^
3.0.2 Definition. The Hamiltonian H is the derivation of the associative algebra
∂
Am arising from the derivation − ∂x of the prefactorization algebra Obsq of observ-
ables of the free scalar field theory with mass m.
3.0.3 Proposition. The associative algebra Am coming from the free scalar field
theory with mass m is the Weyl algebra, generated by p, q, ~ with the relation
[p, q] = ~ and all other commutators being zero.
The Hamiltonian H is the derivation
H(a) = 1
2~ [p
2
− m2 q2 , a].
Note that it is an inner, or principal, derivation.
Remark: It might seem a priori that different action functionals, which encode
different theories, should have different algebras of observables. In this case, we
see that all the theories lead to the same associative algebra. The Hamiltonian is
what distinguishes these algebras of observables; that is, different actions differ by
“how to relate the same observable applied at different moments in time.”
There is an important feature of the Weyl algebra that illuminates the role of
the Hamiltonian. The Weyl algebra is rigid: the Hochschild cohomology of the
Weyl algebra A(n) for a symplectic vector space of dimension 2n vanishes, ex-
cept in degree 0, where the cohomology is one-dimensional. Hence, in particular,
HH 2 (A(n) ) = 0, so there are no nontrivial deformations of the Weyl algebra. In
consequence, any action functional whose underlying free theory is the free scalar
field (when the coupling constants are formal parameters) will have the same alge-
bra of observables, as an associative algebra. Since HH 1 (A(n) ) = 0, we know that
every derivation is inner and hence is represented by some element of A(n) . Thus,
the derivation arising from translation must have a Hamiltonian operator. ^
Proof. We will start by writing down elements of Obsq corresponding to the
position and momentum observables. Recall that

Obsq ((a, b)) = Sym Cc∞ ((a, b))−1 ⊕ Cc∞ ((a, b))0 [~]
with a certain differential.

We let φq , φ p ∈ C ∞ (R) be the functions introduced in the proof of Lemma
2.4.1. Explicitly, if m = 0, then φq (x) = 1 and φ p (x) = x, whereas if m , 0 we have
φq (x) = 12 emx + e−mx ,

φ p (x) = 2m
1
emx − e−mx .

Thus, φq , φ p are both in the null space of the operator −∂2x + m2 , with the properties
that ∂ x φ p = φq and that φq is symmetric under x 7→ −x, whereas φ p is antisymmet-
ric.
As in the proof of Lemma 2.4.1, choose a function f ∈ Cc∞ ((− 12 , 12 )) that is

symmetric under x 7→ −x and has the property that
Z ∞
f (x)φq (x)dx = 1.
−∞
The symmetry of f implies that the integral of f against φ p is zero.
Let ft ∈ Cc∞ ((t − 12 , t + 12 )) be ft (x) = f (x − t). We define observables Pt , Qt by
Q t = ft ,
Pt = − ft0
The observables Qt , Pt are in the space Cc∞ (It )[1] ⊕ Cc∞ (I) of linear observables for
the interval It = ((t − 21 , t + 12 ). They are also of cohomological degree zero. If we
think of observables as functionals of a field in C ∞ (It ) ⊕ C ∞ (It )[−1], then these are
linear observables given by integrating ft or − ft0 against the field φ.
Thus, Qt and Pt represent average measurements of positions and momenta of
the field φ in a neighborhood of t.
Because the cohomology classes [P0 ], [Q0 ] generate the commutative alge-
bra H ∗ (Obscl (R)), it is automatic that they still generate the associative algebra
H 0 Obsq (R). We thus need to show that they satisfy the Heisenberg commutation
relation
[[P0 ], [Q0 ]] = ~
for the associative product on H Obsq (R), which is an associative algebra by virtue
0
of the fact that H ∗ Obsq is locally constant. (The other commutators among [P0 ]
and [Q0 ] vanish automatically.)
We also need to calculate the Hamiltonian. We see that
∂
− Qt = − ft0 = Pt = dtd f (x − t) = dtd Qt ,
∂x
∂
− Pt = ft00 .
∂x
Hence, the Hamiltonian acting on [P0 ] gives the t-derivative of [Pt ] at t = 0 and
similarly for Q0 . In other words, the parameter t encodes translation of the real
line, and the infinitesimal action of translation acts correctly on the observables.
In cohomology, the image of −∂2x + m2 is zero, we see that
d
dt [Pt ] = 0 + m2 [ f (x − t)] = m2 [Qt ],
as we expect from usual classical mechanics. In particular, when m = 0, [Pt ] is
independent of t: in other words, momentum is conserved.
Thus, the Hamiltonian H satisfies
H([P0 ]) = m2 [Q0 ]
H([Q0 ]) = [P0 ].
If we assume that the commutation relation [[P0 ], [Q0 ]] = ~ holds, then
h i
H(a) = 2~
1
[P0 ]2 − m2 [Q0 ]2 , a ,
as desired.
Thus, to complete the proof, it only remains to verify the Heisenberg commu-
tation relation.
If a(t) is a function t, let ȧ(t) denote the t-derivative. It follows from the equa-
tions above for the t-derivatives of [Pt ] and [Qt ] that for any function a(t) satisfying
ä(t) = m2 a(t), the observable
a(t)Pt − ȧ(t)Qt
is independent of t at the level of cohomology.
et is the linear observable of cohomological degree −1 given
More precisely, if Q
by the function ft in Cc ((t − 21 , t + 12 )[1], we have
∞
∂
(a(t)Pt − ȧ(t)Qt ) = −d(a(t)Q et )
∂t
where d is the differential on observables. Explicitly, we compute
∂ ∂ ∂ ∂ ∂
(a(t)Pt − ȧ(t)Qt ) = −ȧ(t) ft (x) − a(t) ft (x) − ä(t) ft (x) − ȧ(t) ft (x)
∂t ∂x ∂t ∂x ∂t
∂2
= a(t) 2 ft (x) − ä(t) ft (x)
∂x
∂2
= a(t) 2 ft (x) − m2 a(t) ft (x)
∂x
= −d(a(t)Q et ).
We will define modified observables Pt , Qt that are independent of t at the
cohomological level. We let
Pt = φq (t)Pt − φ0q (t)Qt
Qt = φ p (t)Pt − φ0p (t)Qt .
Since φ p and φq are in the null space of the operator −∂2x + m2 , the observables
Pt , Qt are independent of t at the level of cohomology. In the case m = 0, then
φq (t) = 1 so that Pt = Pt . The statement that Pt is independent of t corresponds,
in this case, to conservation of momentum.
In general, P0 = P0 and Q0 = Q0 . We also have
∂
Pt = −d(φq (t)Q
et ),
∂t
and similarly for Qt .
It follows that if we define a linear degree −1 observable h s,t by
Z t
h s,t = φq (u)Q
eu (x)du,
u=s
then
dh s,t = P s − Pt .
Note that if |t| > 1, the observables Pt and Qt have disjoint support from the
observables P0 and Q0 . Thus, we can use the prefactorization structure map
Obsq ((− 21 , 12 )) ⊗ Obsq ((t − 21 , t + 12 )) → Obsq (R)
to define a product observable

Q0 · Pt ∈ Obsq (R).
We will let ? denote the associative multiplication on H 0 Obsq (R). We defined
this multiplication by
[Q0 ] ? [P0 ] = [Q0 · Pt ] for t > 1
[P0 ] ? [Q0 ] = [Q0 · Pt ] for t < −1.
Thus, it remains to show that, for t > 1,
[Q0 · Pt ] − [Q0 · P−t ] = ~.
We will construct an observable whose differential is the difference between the
left and right hand sides. Consider the observable
S = f (x)h−t,t (y) ∈ Cc∞ (R) ⊗ Cc∞ (R)[1],
where the functions f and h−t,t were defined above. We view h−t,t as being of
cohomological degree −1, and f as being of cohomological degree 1.
Recall that the differential on Obsq (R) has two terms: one coming from the
Laplacian 4 + m2 mapping Cc∞ (R)−1 to Cc∞ (R)0 , and one arising from the bracket
of the Heisenberg Lie algebra. The second term maps

Sym2 Cc∞ (R)−1 ⊕ Cc∞ (R)0 → R ~.
Applying this differential to the observable S , we find that
Z
(dS ) = f (x)(−∂2 + m2 )h−t,t (y) + ~ h−t,t (x) f (x)dx
R
Z
=Q0 · (P−t − Pt ) + ~ f (x)h−t,t (x).
Therefore Z
[Q0 Pt ] − [Q0 P−t ] = ~ f (x)h−t,t (x).
It remains to compute the integral. This integral can be rewritten as
Z t Z ∞
f (x) f (x − u)φq (u)du.
u=−t x=−∞
Note that the answer is automatically independent of t for t sufficiently large, be-
cause f (x) is supported near the origin so that f (x) f (x − u) = 0 for u sufficiently
large. Thus, we can sent t → ∞.
Since f is also symmetric under x → −x, we can replace f (x − u) by f (u − x).
We can perform the u-integral by changing coordinates u → u − x, leaving the
integrand as f (x) f (u)φq (u + x). Note that

φq (u + x) = 12 em(x+u) + e−m(x+u) .
Now, by assumption on f ,
Z Z
f (x)e dx =
mx
f (x)e−mx dx = 1.
π−1 ((−1, 1))
N×R
R
(−1, 1)
Figure 1. The value of π∗ Obsq on (−1, 1) is the value of Obsq on

the preimage π−1 ((−1, 1))
It follows that Z ∞
φq (u + x) f (u)du = φq (x).
u=−∞
We can then perform the remaining x integral φq (x) f (x)dx, which gives 1, as
R
desired.
Thus, we have proven
[[Q0 ], [P0 ]] = ~,
as desired.
4. Pushforward and canonical quantization

Consider the free scalar field theory on a manifold of the form N × R, equipped
with the product metric. We assume for simplicity that N is compact. Let Obsq
denote the prefactorization algebra of observables of the free scalar field theory
with mass m on N × R. Let π : N × R → R be the projection map. There is a
pushforward prefactorization algebra π∗ Obsq living on R and defined by
(π∗ Obsq )(U) = Obsq (π−1 (U)).
(This pushforward construction is a version of compactification in physics.) In
this section, we will explain how to relate this prefactorization algebra on R to
an infinite tensor product of the prefactorization algebras associated to quantum
mechanics on R. See Figure 4.
L of the operator 4 + m on
Let {ei }i∈I be an orthonormal basis of eigenvectors 2
C (N), where ei has eigenvalue λi . The direct sum

∞
i∈I Rei is a dense subspace
of C ∞ (N).
For m ∈ R, let Am denote the cohomology of the prefactorization algebra as-
sociated to the free one-dimensional scalar field theory with mass m. Thus, Am
4. PUSHFORWARD AND CANONICAL QUANTIZATION 93
is the Weyl algebra generated by p, q, and ~ with commutator [p, q] = ~. The

dependence on m appears only through the Hamiltonian.
Consider the tensor product of algebras
AN = A √λ1 ⊗R[~] A √λ2 ⊗R[~] · · ·
over the entire spectrum {λi }i∈I . (The infinite tensor product is defined to be the
colimit of the finite tensor products, where the maps in the colimit use the unit in
each algebra. See example 1.1 for more discussion.)
4.0.1 Proposition. There is a dense sub-prefactorization algebra of π∗ Obsq that
is locally constant and whose cohomology prefactorization algebra corresponds to
the associative algebra AN .
Remark: This proposition encodes, in essence, the procedure usually known as
canonical quantization. (In a physics textbook, N is typically a torus S 1 × · · · × S 1
whose radii are eventually “sent to infinity.”) The Weyl algebra A √λi for eigenvalue
λi corresponds to “a quantum mechanical particle evolving in N with mass m and
energy λi .” (An important difference is that a quantum field theory textbook often
works with Lorentzian signature, rather than Euclidean signature, as we do.) The
time-ordered observables of the whole free scalar field π∗ Obsq are essentially de-
scribed by the combination of the algebras for all these energy levels. The different
energy levels do not affect one another, as it is a free theory. When an interaction
term is added to the action functional, these energy levels do interact.
The prefactorization algebra π∗ Obsq has a derivation, the Hamiltonian, coming
from infinitesimal translation in R. The prefactorization algebra A √λi also has a
1
[p2 − λi q2 , −]. The map from i A √λi to
N
Hamiltonian, given by bracketing with 2~
H ∗ (π∗ Obsq ) intertwines these derivations. ^
Proof. The prefactorization algebra π∗ Obsq on R assigns to an open subset
U ⊂ R the Chevalley-Eilenberg chains of a Heisenberg Lie algebra given by a
central extension of
4+m2
Cc∞ (U × N) −−−−→ C ∞ (U × N)[−1].
A dense subcomplex of these linear observables is
M 4R +λi

(†) Cc∞ (U)ei −−−−−→ Cc∞ (U)ei [−1] ,
i∈I
because the topological vector space of smooth functions is a certain completion
of this direct sum of eigenspaces.
Let Fi be
√ the prefactorization algebra on R associated to quantum mechanics
with mass λi . This prefactorization algebra is the envelope of the Heisenberg
central extension of
4R +λi
Cc∞ (U)ei −−−−−→ Cc∞ (U)ei [−1].
Note that Fi is a prefactorization algebra in modules over R[~]. We can define then
the tensor product prefactorization algebra
F = F1 ⊗R[~] F2 ⊗R[~] · · ·
to be the colimit of the finite tensor products of the Fi under the inclusion maps
coming from the unit 1 ∈ Fi (U) for any open subset. Equivalently, we can view
this tensor product as being associated to the Heisenberg central extension of the
complex (†) above.
Because the complex (†) is a dense subspace of the complex whose Heisenberg
extension defines π∗ Obsq , we see that there is a map of prefactorization algebras
with dense image
F → π∗ Obsq .
Passing to cohomology, we have a map
H ∗ F → H ∗ (π∗ Obsq ).
As the prefactorization algebra H ∗ (Fi ) corresponds to the Weyl algebra A √λi , we
see that H ∗ F corresponds to the algebra AN .
5. Abelian Chern-Simons theory

We have discussed the free scalar field in some detail, and we have recovered
several aspects of the “usual” story. (Further aspects of the free scalar field occupy
much of the rest of this chapter.) Another important example of a free field theory
is Abelian Chern-Simons theory, which is a particularly simple example of both
a topological field theory and a gauge theory. The theory examined here is the
perturbative facet of Chern-Simons theory with gauge group U(1).
Let M be an orientable manifold of dimension 3. The space of fields is E =
Ω M [1], equipped with the exterior derivative d as its differential. Integration pro-
∗
vides a skew-symmetric pairing of degree −1:

Z
hα, βi = (−1) |β|
α ∧ β,
M
where α, β live in Ec and each has degree 1 less than its usual degree, due to the
shift. (Inside the integral, we simply view them as ordinary differential forms,
without the shift.) The action functional of Abelian Chern-Simons theory is then
S (α) = hα, dαi,
and the associated equation of motion is that dα = 0. In other words, it picks out
flat connections.
To be more precise, this complex describes the derived space of solutions to
the problem of finding flat connections on the trivial complex line bundle on M,
up to gauge equivalence. If we focus on the fields of degree 0 – namely, 1-forms
– then a solution is precisely a flat connection. Two solutions α and α0 such that
α − α0 = d f , for f ∈ Ω0 [1], are viewed as equivalent solutions. The function s = e f
is a nowhere-vanishing section of the complex line bundle, and we can use it to
relate the two connections, as follows. If g is a function, then
s−1 (d + α0 )(sg) = s−1 (ds)g + sdg + α0 sg

= (dg + α0 g) + (d f )g
= (d + α)g,
5. ABELIAN CHERN-SIMONS THEORY 95
since ds = sd f . In other words, s provides an automorphism, or gauge transforma-

tion, intertwining the two flat connections.
We have explained why Ω1 appears in degree 0 (the connection 1-forms are
the main actors) and why Ω0 appears in degree 1 (functions are the Lie algebra of
gauge automorphisms). We now explain why Ω2 and Ω3 appear.
The action functional S is naturally a function on this linear derived space
d
Ω0 [1] → Ω1 ,
so that its exterior derivative dS is a section of the cotangent bundle of this linear
derived space. As S is quadratic, dS is linear, so the section is linear. From the
physical perspective, we want to study the derived intersection of the zero section
and dS in this cotangent bundle: this amounts to studying the derived critical locus
of S (i.e., the points in the linear derived stack where dS vanishes). This derived
intersection is itself a linear derived stack and is described by the full de Rham
complex.
5.1. Observables. The classical observables for Abelian Chern-Simons the-

ory are easy to describe and to compute. For U an open subset of M, we have
d d d

− Ω1c [1] →
Obscl (U) = Sym∗ Ω0c [2] → − Ω2c →
− Ω3c [−1] .
Thus, we see that

H ∗ Obscl (U) Sym∗ Hc∗ (U)[2] .

In contrast to the free scalar field, where the cohomology of the linear observ-
ables was often infinite-dimensional, the cohomology of the linear observables is
finite-dimensional, for any open U. Thus, the cohomology prefactorization algebra
H ∗ Obscl is relatively easy to understand.
This pleasant situation continues with the quantum observables. Consider the
natural filtration on Obsq (U) by
F k Obsq (U) = Sym≤k Ω∗c (U)[2] [~].

This filtration induces a spectral sequence, and the first page is H ∗ Obscl (U)[~],
which we have already computed. The differential on the first page is zero, but the
differential on the second page arises from the pairing on linear observables (recall
that the pairing provides the Lie bracket of the Heisenberg Lie algebra, and this
bracket provides the differential in the Chevalley-Eilenberg chain complex). Note
that the pairing on linear observables descends to the cohomology of the linear
observables by
h[α], [β]iH ∗ = hα, βi.
Hence, the differential d2 on the second page of the spectral sequence vanishes on
constant and linear terms (i.e., from Sym0 and Sym1 ) and satisfies
d2 ([α][β]) = ~h[α], [β]iH ∗
for a pure quadratic term. The BD algebra axiom (recall equation (2.6.1)) then
determines d2 on terms that are cubic and higher.
Thus, the cohomology of the quantum observables reduces to understanding

the integration pairing between cohomology classes of compactly supported de
Rham forms. In particular, Poincaré duality implies the following.
5.1.1 Lemma. For M a connected, closed 3-manifold,
H ∗ Obsq (M) R[~][1 − dim H 2 (M)].
Proof. Let b2 = dim H 2 (M). Pick a basis α1 , . . . , αb2 for H 2 (M) and a dual
basis β1 , . . . , βb2 for H 1 (M):
hα j , βk iH ∗ = δ jk .
Let ν denote the generator for H 0 (M) and [M] the dual generator for H 3 (M). Thus,
d2 (α j βk ) = ~δ jk and d2 (ν[M]) = ~, and d2 vanishes on all other quadratic monomi-
als. Explicit computation then shows that the pure odd term να1 · · · αb2 is closed
but not exact, and it generates the cohomology.
5.2. Compactifying along a closed surface. The next case to consider is a
3-manifold of the form M = R × Σ, where Σ is a closed, orientable surface. Let
π : M → R denote the obvious projection map. We will show that the pushforward
prefactorization algebra π∗ Obsq is locally constant, so that it corresponds to an
associative algebra. This associative algebra is the Weyl algebra for the graded
symplectic vector space H ∗ (Σ)[1]. (Recall Figure 4.)
Some notation will make it easier to be precise. Let g denote the genus of Σ.
Pick a symplectic basis {α1 , . . . , αg , β1 , . . . , βg } for H 1 (Σ) in the sense that
Z
α j ∧ βk = δ jk .
Σ
Let ν denote the basis for R given by the constant function 1. Let µ denote a
H 0 (Σ)
basis for H 3 (M) such that Σ µ = 1. Thus, the graded vector space H ∗ (Σ)[1] has a
natural symplectic structure given by the integration pairing. The odd elements ν
and µ are dual under this pairing.
5.2.1 Proposition. The prefactorization algebra H ∗ π∗ Obsq is locally constant,
and it corresponds to the Weyl algebra AΣ for H ∗ (Σ)[1]. In terms of our chosen
generators, this Weyl algebra has commutators [ν, µ] = ~, [α j , βk ] = ~δ jk and all
others vanish.
The Hamiltonian in this Weyl algebra is zero.
The fact that the Hamiltonian is zero is what makes Abelian Chern-Simons a
topological field theory . Note as well the appearance of the Weyl algebra.
Proof. This argument is much simpler than the argument for the free scalar
field in one dimension. We will use the Hodge theorem to replace the big complex
Ω∗c (R) ⊗ Ω∗ (Σ) by the smaller complex Ω∗c (R) ⊗ H ∗ (Σ), and then exploit our con-
struction for the universal enveloping algebra (recall Proposition 4.0.1) to obtain
the result.
Let L = π∗ Ω∗M [1]. This is the pushforward of the local Lie algebra of linear
observables (it is an Abelian Lie algebra, as we are working with a free theory).
Then π∗ Obscl C∗ Lc . Let Lb denote the pushforward of the Heisenberg central

extension. Then π∗ Obs C∗ L
q bc , the factorization envelope.
By the Hodge theorem, we can obtain a homotopy equivalence between Ω∗ (Σ)
and its cohomology H ∗ (Σ). Tensoring with Ω∗c on R, we obtain a homotopy equiv-
alence
Lc ' Ω∗c ⊗ H ∗ (Σ)[1] = g
of Abelian dg Lie algebras.Thus, we see that C∗ g is homotopy equivalent to π∗ Obscl .

Similarly, C∗b g is homotopy equivalent to π∗ Obsq , where b g denotes the Heisenberg
central extension (i.e., we use the Poincaré pairing on H ∗ (Σ) tensored with the
integration pairing on R).
The cohomology prefactorization algebra H ∗ π∗ Obsq is isomorphic to the co-
homology prefactorization algebra H ∗ (C∗b g). By Proposition 4.0.1, we know this
corresponds to the associative algebra U H ∗[ (Σ)[1], where the central element is
denoted ~.
Now we compute the Hamiltonian. Recall that the Hamiltonian arose in the
free scalar field via the translation action on R. Below, we will explicitly compute
the Hamiltonian in a manner parallel to Proposition 3.0.3. But there is a conceptual
reason for the vanishing of the Hamiltonian: the theory, because it is simply the
de Rham complex, is homotopy-invariant under diffeomorphism. Not only are
the observables preserved by translation along R; they are preserved under any
diffeomorphism of R.
∂
Let x denote a coordinate on R. Infinitesimal translation ∂x preserves every-
thing in the prefactorization algebra π∗ Obs , just as in Proposition 3.0.3. In this
q
case, however, the solutions to the equations of motion – namely, de Rham co-
homology classes of Σ – are translation-invariant,Rso the Hamiltonian is trivial, as
follows. Fix a function f ∈ Cc∞ ((− 12 , 12 )) such that R f (x)dx = 1. Hence the 1-form
f (x)dx generates Hc1 (R). Let ft (x) = f (x − t). Then we get linear observables in
C∗b g((t − 12 , t + 12 )) by
At, j = ft (x)dx ⊗ α j ,
Bt,k = ft (x)dx ⊗ βk ,
νt = ft (x)dx ⊗ ν,
µt = ft (x)dx ⊗ µ.
These elements are all cohomologically nontrivial, and their cohomology classes
generate the Weyl algebra. Observe that, for instance,
∂
At, j = ft0 (x)dx ⊗ α j = d ft ⊗ α j ,
∂x
so that the infinitesimal translation is cohomologically trivial. This argument ap-
plies to all the generators, so we see that the derivation on the Weyl algebra induced
∂
from ∂x must be trivial.
Remark: There is a closely related free field theory on M = R × Σ. Equip Σ with a

complex structure, and consider the free theory on M whose fields are
Ω∗ (R) ⊗ Ω0,∗ (Σ)[1] ⊕ Ω∗ (R) ⊗ Ω1,∗ (Σ),
concentrated from degrees −1 to 2. (Just to pin down our conventions: Ω0 (R) ⊗

Ω0,0 (Σ) is in degree -1, and Ω0 (R) ⊗ Ω1,0 (Σ) is in degree 0.) This theory is topolog-
ical along R and holomorphic along Σ (see Chapter 5 for more discussion of this
notion). Notice that if we added the differential ∂ for Σ to this complex, we would
recover Abelian Chern-Simons on R × Σ.
q
Let us denote the observables for this theory by Obs . It is straightforward to
∂
q
mimic the proof above to show that the cohomology of π∗ Obs corresponds to the
∂
Weyl algebra on the graded symplectic vector space H ∗,∗ (Σ)[1], the Dolbeault co-
homology of Σ. As this cohomology is isomorphic to the de Rham cohomology of
q
Σ, we see that the cohomology prefactorization algebras H ∗ π∗ Obsq and H ∗ π∗ Obs
∂
are isomorphic. ^
5.3. 3-manifolds with boundary. We now examine the observables on a com-
pact, connected, orientable 3-manifold M with connected boundary ∂M = Σ of
genus g. Let M denote the open interior, so M = M − ∂M. Speaking loosely, we
can view M as a “module” over the “algebra” Σ×R, because we can keep stretching
the end of M to include a copy of Σ × R. We will show that this picture leads to a
precise statement about the observables: the cohomology H ∗ Obsq (M) is a module
for the Weyl algebra AΣ associated to H ∗ (Σ)[1].
Pick a parametrization of a collar neighborhood N of the boundary ∂M such

that we have a diffeomorphism ρ : N − → Σ × [0, 2). Let π : N → [0, 2) denote
composition with the projection map. Let N 0 = π−1 ((0, 1)) denote a neighborhood
of the end of M.
Consider the map p : M → (0, 1] where
π(x), x ∈ π−1 ((0, 1))

(
p(x) = ,
1, else
which collapses everything in M outside of N 0 down to a point. By construction,
p−1 ((a, 1]) is diffeomorphic to M for any 0 < a. Likewise, p−1 ((a, b)) is diffeomor-
phic to Σ × (a, b) for any 0 < a < b < 1. See Figure 5.3.
It is clear that p∗ Obsq on the interval (0, 1) is simply the observables compact-
ified along a surface, as in Proposition 5.2.1. To any open set of the form (a, 1] in
[0, 1], we see that p∗ Obsq ((a, 1]) is quasi-isomorphic to Obsq (M).
5.3.1 Proposition. The cohomology prefactorization algebra H ∗ p∗ Obsq is con-
structible with respect to the stratification (0, 1] = (0, 1) ∪ {1}. In particular,
(1) on the stratum (0, 1), H ∗ p∗ Obsq corresponds to the Weyl algebra AΣ ;
(2) for any open (a, 1], H ∗ p∗ Obsq ((a, 1]) H ∗ (Obsq (M)); and
(3) the structure maps equip H ∗ (Obsq (M)) with the structure of a left AΣ
module, whose annihilator is the ideal generated by the Lagrangian
N Σ × [0, 2)
M
Σ
Figure 2. Everything right of the dotted copy of Σ maps to the

right endpoint
subspace Im i∗ ⊂ H ∗ (Σ)[1] for the map i∗ : H ∗ (M; Σ)[1] → H ∗ (Σ)[1]

arising from i : Σ ,→ M.
Remark: This proposition – more accurately, its proof – indicates how to construct
a topological field theory, in the functorial sense, from the bordism category of
oriented 2-manifolds with oriented cobordisms as morphisms to the category of
graded algebras with bimodules as morphisms. To a surface Σ, we assign the Weyl
algebra AΣ . To a 3-manifold, we associate to the bimodule over Weyl algebras
(extending the proposition above to more general 3-manifolds with boundary). The
final step in constructing the functor – verifying composition – relies on the gluing
properties of prefactorization algebras, which we discuss in Chapter 6, although
one could verify it directly by computation in this case. ^
Proof. We have already explained claims (1) and (2), so we tackle (3). It is
manifest that H ∗ (Obsq (M)) is a left AΣ module, because p∗ Obsq is a prefactoriza-
tion algebra. Thus, it remains to describe the annihilator of this module.
Recall the “half lives, half dies” principle for oriented 3-manifolds: for M a
compact, orientable 3-manifold with boundary i : ∂M ,→ M, the image of the
map i∗ : H 1 (M) → H 1 (∂M) has dimension half that of H 1 (∂M). (For a proof,
see lemma 3.5 of Hatcher (n.d.).) The image L = Im i∗ is, in fact, a Lagrangian
subspace of H 1 (∂M)[1].
Now consider our situation, where M = M − ∂M and ∂M = Σ. We know that
the neighborhood N 0 = p−1 ((0, 1)) of the end of M is diffeomorphic to Σ × (0, 1).
Let j : N 0 ,→ M denote the inclusion. By Poincaré duality for compactly supported
de Rham cohomology, we know Hc2 (N 0 ) H 1 (Σ) and Hc2 (M) H 1 (M). Hence,
the “half lives, half dies” principle shows that for the extension-by-zero map j! :
Hc2 (N 0 ) → Hc2 (M), the kernel Ker j! is a Lagrangian subspace. We also see that
j! : Hc3 (N 0 ) Hc3 (M) and j! : Hc1 (N 0 ) → Hc1 (M) is the zero map.
Thus, the graded subspace Ker j! ⊂ Hc∗ (N 0 ) is Lagrangian.
For the observables, the map j! describes how the linear observables on N 0 be-
have as linear observables on M. The observables in Ker j! thus vanish on M, and
so the ideal they generate in AΣ vanishes on the module H ∗ Obsq (M). We need to
show this ideal is the entire annihilator. Note that the action of AΣ is governed by
the integration pairing in Hc∗ (M), and the extension-by-zero map Hc∗ (N 0 ) → Hc∗ (M)
respects the integration pairing, so there are no nontrivial relations involving non-
linear observables.
5.4. Knots and links. Chern-Simons theory is famous for its relationship with
knot theory, notably the Vassiliev invariants. Here we will explain a classic exam-
ple: how the Gauss linking number appears in Abelian Chern-Simons theory. For
simplicity, we work in R3 .
Let K : S 1 ,→ R3 and K 0 : S 1 ,→ R3 be two disjoint knots. The linking
number `(K, K 0 ) of K and K 0 is the degree of the Gauss map G K,K 0 : S 1 × S 1 → S 2
that sends a pair of points (x, y) ∈ K × K 0 to the unit vector (x − y)/|x − y|. One
way to compute it – indeed, the way Gauss first introduced it – is to pull back the
standard volume form on S 2 and integrate over K × K 0 . Another way is to pick an
oriented disc whose boundary is K 0 and to count, with signs, the number of times
K intersects D (we may wiggle D gently to make it transverse to K).
To relate this notion to Chern-Simons theory, we need to find an observable
associated to a knot. The central object of Chern-Simons theory is a 1-form A,
which we view as describing a connection d + A for the trivial line bundle on R3 .
Thus, the natural observable for the knot K : S 1 → R3 is
Z
OK (A) = A,
K
where we have fixed an orientation on S 1. Note that this operator is linear in A.
Remark: The Wilson loop observable WK is exp(OK ), which is the holonomy of
the connection d + A around K. We do not work with it here because we want to
work with the symmetric algebra on linear observables, rather than the completed
symmetric algebra. In the non-Abelian setting, we do work with the completed
symmetric algebra, so the analogous Wilson loop observable lives in our quantum
observables for non-Abelian Chern-Simons theory. ^
There is a minor issue with this observable, however: it is distributional in na-
ture, as its support sits on the knot K. It is easy to find a smooth (or “smeared”) ob-
servable that captures the same information. (We know such a smooth observable
exists abstractly by Lemma 2.1.1, but here we will give an explicit representative.)
The key is again to use the pushforward.
Consider the 3-manifold T = S 1 × R2 . We view the core C = S 1 × (0, 0) ⊂ T
as the “distinguished” knot in T . Any embedding κ : T ,→ R3 induces a pushfor-
ward of observables so that we can understand the observables supported in tubular
neighborhood κ(T ) of the knot κ(C) by understanding the observables on T .
6. ANOTHER TAKE ON QUANTIZING CLASSICAL OBSERVABLES 101
We now explain how to find a smeared analog of OC , where C is the distin-

guished knot, in Obsq (T ). Fix an orientation of T and an orientation of the em-
bedded circle C ⊂ T . Let µ be a compactly supported volume form on R2 whose
support contains the origin and whose total integral is one. Consider its pullback
µT = π∗ µ along the projection π : T → R2 . Using Cartesian coordinates r, z on R2
and coordinates θ, r, z on T , we have µT = φ(r, z)dr ∧ dz, where φ is a nonnegative
bump function with support at the origin (0, 0). Note that µT is independent of θ.
We thus have a linear function, or observable,
Z
Oµ : A 7→ A ∧ µT
T
on any 1-form A. Using the coordinates from above, we see
Z
Oµ (A) = φ(x, y)OS 1 ×(r,z) (A)dr dz,
R2
so that Oµ is just a smoothed version of OC .
Let’s return to the linking number. Given two knots K and K 0 , we can find
thickened embeddings κ : T → R3 and κ0 : T → R3 whose images are disjoint
and where κ(C) = K and κ0 (C) = K 0 . Let µκ denote the image of µT in Obsq (κ(T ))
under the pushforward by κ, and likewise for µκ0 . This observable µκ is a smeared
version of OK .
5.4.1 Lemma. In H 0 (Obsq (R3 )) R[~], the cohomology class [µκ µκ0 ] of the prod-
uct observable µκ µκ0 is equal to ~`(K, K 0 ).
Proof. We outline here one approach, in the spirit of our preceding work. Be-
low, in example 6.3 we explain the standard proof.
The Poincaré lemma for compactly supported forms tells us that Hc2 (R3 ) = 0.
As µκ0 is closed, we know there is some 1-form ρ such that dρ = µκ0 . The product
observable µκ ρ then has cohomological degree −1, and by construction we have
Z
(d + ~4)µκ ρ = µκ µκ0 + ~ µκ ∧ ρ.
R3
Thus we need to verify that, up to a sign, the integral agrees with the linking num-
ber.
We do this by making a good choice of ρ. There is a smooth embedding of
a closed 2-disc D ,→ R3 whose oriented boundary is K 0 . Extend this embedding
to a thickened embedding i : R3 ,→ R3 , where D sits inside the input copy of
R3 as D × 0 and where this thickened embedding contains the image of κ0 . Then
i∗ µκ0 is cohomologically trivial, by the Poincaré lemma. Pick a 1-form ρ such that
dρ = i∗ µκ0 . Note that ρ has support in i(R3 ). The product µκ ∧ ρ is smeared version,
under Poincaré duality, of the intersection of i(D) with K. Thus its integral counts
the signed intersection.
6. Another take on quantizing classical observables

We have given an abstract definition of the prefactorization algebra of quan-
tum observables of a free field theory as the factorization envelope of a certain
Heisenberg dg Lie algebra. Our goal in this section is to provide another useful
description.
The prefactorization algebra of quantum observables Obsq , viewed as a graded
prefactorization algebra with no differential, coincides with Obscl [~]. In particular,
the structure maps are the same. The only difference between Obsq and Obscl [~] is
in the differential.
In deformation quantization, we deform, however, the product of observables.
We saw in Section 3 that taking cohomology, the structure maps for the quantum
observables do change and correspond to the Weyl algebra. The change in differen-
tial modifies the structure maps at the level of cohomology. However, it is possible
to approach quantization differently and deform the structure maps at the cochain
level, instead of the differential.
Hence, in this section we will give an alternative, but equivalent, description of
Obsq . We will construct an isomorphism of precosheaves Obsq Obscl [~] that is
compatible with differentials. This isomorphism is not compatible, however, with
the factorization product. Thus, this isomorphism induces a deformed factorization
product on Obscl [~] corresponding to the factorization product on Obsq .
In other words, instead of viewing Obsq as being obtained from Obscl by keep-
ing the factorization product fixed but deforming the differential, we will show that
it can be obtained from Obscl by keeping the differential fixed but deforming the
product.
One advantage of this alternative description is that it is easier to construct
correlation functions and vacua in this language. We will now develop these ideas
in the case of the free scalar field.
6.1. Green’s functions. The isomorphism we will construct between the cochain
complexes of quantum and classical observables relies on a Green’s function for
the Laplacian.
6.1.1 Definition. A Green’s function is a distribution G on M× M, preserved under
reflection across the diagonal (i.e., symmetric) and satisfying
(4 ⊗ 1)G = δ4 ,
where δ4 is the δ-distribution on the diagonal M ,→ M × M.
A Green’s function for the Laplacian with mass satisfies the equation
(4 ⊗ 1)G + m2G = δ4 .
∂ 2
(The convention is that 4 has positive eigenvalues, so that on Rn , 4 = −
P
∂xi .)
If M is compact, then there is no Green’s function for the Laplacian. Instead,

there is a unique function G
e satisfying
e = δ4 − 1.
(4 ⊗ 1)G
However, if we introduce a non-zero mass term, then the operator 4+m2 on C ∞ (M)
is an isomorphism, so that there is a unique Green’s function.
If M is non-compact, then there can be a Green’s function for the Laplacian

without mass term. For example, if M is Rn , then a choice of Green’s function is

 4π1d/2 Γ(d/2 − 1) |x − y|2−n if n , 2

G(x, y) = 

− 1 log |x − y|

2π if n = 2.
As the reader will see, the role of the Green’s function is to produce an isomor-
phism, so different choices lead to different isomorphisms. These choices do not
change the quantum observables themselves, merely how one “promotes” a classi-
cal observable to a quantum one.
6.2. The isomorphism of graded vector spaces. Let us now turn to the con-
struction of the isomorphism of graded vector spaces between Obscl (U) and Obsq (U)
in the presence of a Green’s function.
The underlying graded vector space of Obsq (U) is
Sym(Cc∞ (U)[1] ⊕ Cc∞ (U))[~].
In general, for any vector space V, each element P ∈ (V ∨ )⊗2 defines a differential
operator ∂P of order two on Sym V, viewed as a commutative algebra, where ∂P
is uniquely characterized by the conditions that it is zero on Sym≤1 V and that on
Sym2 V it is given by contraction with P. The same holds when we define the
symmetric algebra using the completed tensor product.
In the same way, for every distribution P on U ×U, we can define a continuous,
second-order differential operator on Sym(Cc∞ (U)[1] ⊕ Cc∞ (U)) that is uniquely
characterized by the properties that it vanishes on Sym≤1 , that it vanishes elements
of negative cohomological degree in Sym2 (Cc∞ (U)[1] ⊕ Cc∞ (U)), and that
Z
∂P (φψ) = P(x, y)φ(x)ψ(y)
U×U
for any φ, ψ ∈ Cc∞ (U)
of cohomological degree zero. (On the left hand side, we
view φψ as an element of Sym2 .)
Choose a Green’s function G for the Laplacian on M. Thus, G restricts to a
Green’s function for the Laplacian on any open subset U of M.
Therefore, we can define a second-order differential operator ∂G on the algebra
Sym(Cc∞ (U)[1] ⊕ Cc∞ (U)). We extend this operator by R[~]-linearity to an operator
on the graded vector space Obsq (U).
Now, the differential on Obsq (U) can be written as d = d1 + d2 , where d1 is
a first-order differential operator and d2 is a second-order operator. The operator
4
d1 is the derivation arising from the differential on the complex Cc∞ (U) −
→ Cc∞ (U).
The operator d2 arises from the Lie bracket on the Heisenberg dg Lie algebra; it
is a continuous, ~-linear, second-order differential operator uniquely characterized
by the property that
Z
d2 (φ φ ) = ~
−1 0
φ−1 (x)φ0 (x)
M
for φk in the copy of Cc∞ (U) in degree k, sitting inside of Obsq (U).
The Green’s function G satisfies
((4 + m2 ) ⊗ 1)G = δ4
where δ4 is the Green’s function on the diagonal. It follows that
[~∂G , d1 ] = d2 .
Indeed, both sides of this equation are second-order differential operators, so to

check the equation, it suffices to calculate how the act on an element of Sym2 (Cc∞ (U)[1]⊕
Cc∞ (U)). If φk denotes an element of the copy of Cc∞ (U) in cohomological degree
k, we have
~∂G d1 (φ0 φ−1 ) = ~∂G (φ0 (4 + m2 )φ−1 )

Z
= G(x, y)φ0 (x)((4 + m2 )φ−1 )(y)
ZU×U
= ((4y + m2 )G(x, y))φ0 (x)φ−1 (y)
U×U
Z
= φ0 (x)φ−1 (x)
U
= d2 (φ0 φ−1 ).
On the first line, the element φ0 φ−1 lives in Sym2 , as does its image under d1 .
On the fourth line, we have used the fact that 4y + m2 applied to G(x, y) is the
delta-distribution on the diagonal.
It is also immediate that ∂G commutes with d2 . Thus, if we define W(α) =
e~∂G (α), we have
(d1 + d2 )W(α) = W(d1 α).
In other words, there is an R[~]-linear cochain isomorphism
WU : Obscl (U)[~] → Obsq (U),
where the domain is the complex Obscl (U)[~] with differential d1 and the codomain
is the same graded vector space with differential d1 + d2 . This isomorphism holds
for every open U.
Note that W is not a map of prefactorization algebras. Thus, W induces a fac-
torization product (i.e., structure maps) on classical observables that quantizes the
original factorization product. Let us denote this quantum product by ?~ , whereas
the original product on classical observables will be denoted by ·. If U, V are dis-
joint open subsets of M with α ∈ Obscl (U), β ∈ Obscl (V), we have

α ?~ β = e−~∂G e~∂G α · e~∂G β .
This product is an analog of the Moyal formula for the product on the Weyl algebra.
6.3. Beyond the free scalar field. We have described this construction for the
case of a free scalar field theory. This construction can be readily generalized to
the case of an arbitrary free theory. Suppose we have such a theory on a manifold
M, with space of fields E (M) and differential d. Instead of a Green’s function, we
require a symmetric and continuous linear operator G : (E (M)[1])⊗2 → R such
that
Gd(e1 ⊗ e2 ) = he1 , e2 i
where h−, −i is the pairing on E (M), which is part of the data of a free field theory.
(When M is compact and H ∗ (E (M)) = 0, the propagator of the theory satisfies
this property. For M = Rn , we can generally construct such a G from the Green’s
function for the Laplacian.) In this context, the operator e~∂G is, in the terminology
of Costello (2011b), the renormalization group flow operator from scale zero to
scale ∞.
This construction often makes it easier to analyze the structure of the quantum
observables. We will use it, for instance, to understand the correlation functions of
some chiral conformal field theories in the next chapter.
Example: We can also use it to provide another proof of Lemma 5.4.1, which as-
serts that Abelian Chern-Simons knows about linking number. Recall that for two
disjoint knots K and K 0 , we introduced smoothed versions, µκ and µκ0 , of the cur-
rents with support along the knots.
It is easy to check that, up to a constant, the 2-form
3
1 X
ω= (−1)i (xi − yi ) d(x1 − y1 ) ∧ d(x[
i − yi ) ∧ d(x3 − y3 )
|x − y|3 i=1
on R3 × R3 acts as a Green’s function for the de Rham complex on R3 . It is the

pullback of the standard volume form on S 2 along the projection
(x, y) 7→ (x − y)/|x − y|
from the configuration space of two distinct points in R3 to S 2 . (Indeed, ω is the
standard propagator, as discussed in Costello (n.d.b) or Kontsevich (1994).) Using
the ?~ formula from above, one can see that the cohomology class of a product
observable µκ µκ0 is precisely the integral formula for the Gauss linking number.
To be explicit, we see that
Z
µκ ?~ µκ0 = µκ µκ0 + ~ µκ (x)µκ0 (y)ω(x, y).
x,y∈R3
Our tubular embeddings, κ : T → R3 and κ0 : T → R3 , allow us to rewrite the

integral as an integral over T × T C × C × R4 . The product C × C of two circles
parametrize K × K 0 , and the last four coordinates parametrize small shifts of K × K 0
within their tubular neighborhoods. Fixing such a shift, we are integrating ω over
K × K 0 , and hence obtain the linking number `(K, K 0 ) because ω is a pullback
of the volume form on S 2 . When we integrate over the shifts, we are integrating
µT (x) ∧ µT (y), which integrates to 1 by construction. (We have not checked the
overall sign, of course.) ^
7. Correlation functions
Suppose we have a free field theory on a compact manifold M, with the prop-
erty that H ∗ (E (M)) = 0. As an example, consider the massive scalar field theory
on M where
4+m2
E = C∞
M −−−−→ C M .
∞
Our conventions are such that the eigenvalues of 4 are non-negative. Adding a
non-zero mass term m2 gives an operator with no zero eigenvalues, so that this
complex has no cohomology, by Hodge theory.
In this situation, the dg Lie algebra Eb(M) we constructed above has coho-
mology spanned by the central element ~ in degree 1. It follows that there is an
isomorphism of R[~]-modules
H ∗ (Obsq (M)) R[~].
Let us normalize this isomorphism by asking that the element 1 in
Sym0 (Eb[1]) = R ⊂ C∗ (Eb[1])
is sent to 1 ∈ R[~].
7.0.1 Definition. In this situation, we define the correlation functions of the free
theory as follows. For a finite collection U1 , . . . , Un ⊂ M of disjoint opens and a
closed element Oi ∈ Obsq (Ui ) for each open, we define
hO1 · · · On i = [O1 · · · On ] ∈ H ∗ (Obsq (M)) = R[~].
Here O1 · · · On ∈ Obsq (M) denotes the image under the structure map
O
Obsq (Ui ) → Obsq (M)
i
of the tensor product O1 ⊗ · · · ⊗ On .
In other words, our normalization gives a value in R[~] for each cohomology
class, and we use this to read off the “expected value” of the product observable.
The map W constructed in the previous section allows us to calculate correla-
tion functions. Since we have a non-zero mass term and M is compact, there is a
unique Green’s function for the operator 4 + m2 .
7.0.2 Lemma (Wick’s lemma). Let U1 , . . . , Un be pairwise disjoint open sets. Let
αi ∈ Obscl (Ui ) = Sym(Cc∞ (Ui )[1] ⊕ Cc∞ (Ui ))
be classical observables, and let
W(αi ) = e~∂G αi ∈ Obsq (Ui )
be the corresponding quantum observables under the isomorphism W of cochain
complexes. Then
hW(α1 ) · · · W(αn )i = W −1 (W(α1 ) · · · W(αn )) (0) ∈ R[~].
7. CORRELATION FUNCTIONS 107
On the right hand side, W(α1 )·W(α2 ) indicates the product in the algebra Sym(C ∞ (M)[1]⊕
C ∞ (M))[~]. The symbol (0) indicates evaluating a function on C ∞ (M)[1]⊕C ∞ (M)[~]
at zero, i.e., taking its component in Sym0 .
Proof. The map W is an isomorphism of cochain complexes between Obsq (U)
and Obscl (U)[~] for every open subset U of M. As above, let ?~ denote the fac-
torization product on Obscl [~] corresponding, under the isomorphism W, to the
factorization product on Obsq . That is, we apply W to each input, multiply in
Obsq , and then apply W −1 to the output: explicitly, we have

α1 ?~ · · · ?~ αn = e−~∂G (e~∂G α1 ) · · · (e~∂G αn ) .
Since W gives an isomorphism of prefactorization algebras between Obsq and

(Obscl [~], ?~ ), the correlation function for the observables W(αi ) is the cohomol-
ogy class of α1 ?~ · · · ?~ αn in Obscl (M)[~]. The map Obscl (M)[~] → R[~] sending
an observable α to α(0) is an R[~]-linear cochain map inducing an isomorphism on
cohomology, and it sends 1 to 1. There is a unique such map (unique up to cochain
homotopy). Therefore,
hW(α1 ) · · · W(αn )i = (α1 ?~ · · · ?~ αn ) (0),
as claimed.
This statement does not resemble the usual statement in a physics text, so we
unpack it in an example to clarify the relationship with other versions of Wick’s
lemma.
As an example, let us suppose that we have two linear observables α1 , α2 of
cohomological degree 0, defined on open sets U, V. Thus, α1 ∈ Cc∞ (U) and α2 ∈
Cc∞ (V). As these are linear observables, the second-order operator ∂G does nothing,
so W(αk ) = αk . But the product is quadratic so we have
W −1 (α1 α2 ) = α1 α2 − ~∂G (α1 α2 ).
As Z
∂G (α1 α2 ) = G(x, y)α1 (x)α2 (y),
M×M
it follows that
Z
hW(α1 )W(α2 )i = −~ G(x, y)α1 (x)α2 (y),
M×M
which is what a physicist would write down as the expected value of two linear
observables.
Remark: We have set things up so that we are computing the functional integral
against the measure eS /~ dµ ,where S is the action functional and dµ is the (non-
existent) “Lebesgue measure” on the space of fields. Physicists often use the mea-
sures e−S /~ or eiS /~ . By changing ~ (e.g., ~ 7→ −~), one can move between these
different conventions. ^.
8. Translation-invariant prefactorization algebras

In this section we will analyze in detail the notion of translation-invariant
prefactorization algebras on Rn . Most theories from physics possess this property.
For one-dimensional field theories, one often uses the phrase “time-independent
Hamiltonian” to indicate this property, and in this section we will explain, in ex-
amples, how to relate the Hamiltonian formalism of quantum mechanics to our
approach.
8.1. The definition. We now turn to the definition of a translation-invariant

prefactorization algebra. It is a special case of the definition of an equivariant
prefactorization algebra in Section 7, but we go through the definition here in some
detail.
Remark: We use the term “translation-invariance” here in place of translation-
equivariance. The motivation is that a field theory on Rn whose action functional
is translation-invariant will produce a translation-equivariant prefactorization alge-
bra. We hope this slight inconsistency of terminology causes no confusions. ^
If U ⊂ Rn and x ∈ Rn , let
τ x U = {y : y − x ∈ U}
denote the translate of U by x.
8.1.1 Definition. A prefactorization algebra F on Rn is discretely translation-
invariant if we have isomorphisms
τ x : F (U) F (τ x U)
for all x ∈ Rn and all open subsets U ⊂ Rn . These isomorphisms must satisfy a few
conditions. First, we require that τ x ◦ τy = τ x+y for every x, y ∈ Rn . Second, for all
disjoint open subsets U1 , . . . , Uk in V, the diagram
τx
F (U1 ) ⊗ · · · ⊗ F (Uk ) / F (τ x U1 ) ⊗ · · · ⊗ F (τ x Uk )
τx
F (V) / F (τ x V)
commutes. (Here the vertical arrows are the structure maps of the prefactorization
algebra.)
Example: Consider the prefactorization algebra of quantum observables of the free

scalar field theory on Rn , as defined in Section 2.6. This theory has as its complex
of fields
4
E = C∞ −
→ C ∞ [−1] .
By definition, Obsq (U) is the Chevalley-Eilenberg chains of a −1-shifted central
extension Ec (U) of Ec (U), with cocycle defined by φ φ where φk denotes the
R
b 0 1
field in cohomological degree k.

8. TRANSLATION-INVARIANT PREFACTORIZATION ALGEBRAS 109
This Heisenberg algebra is defined using only the Riemannian structure on Rn ,

and it is therefore automatically invariant under all isometries of Rn . In particular,
the resulting prefactorization algebra is discretely translation-invariant. ^
We are interested in a refined version of this notion, where the structure maps
of the prefactorization algebra depend smoothly on the position of the open sets. It
is a bit subtle to talk about “smoothly varying an open set,” and in order to do this,
we introduce some notation.
Recall the notion of a derivation of a prefactorization algebra on a manifold M.
8.1.2 Definition. A degree k derivation of a prefactorization algebra F is a collec-
tion of maps DU : F (U) → F (U) of cohomological degree k for each open subset
U ⊂ M, with the property that for any finite collection U1 , . . . , Un ⊂ V of disjoint
opens and an element αi ∈ F (Ui ) for each open, the derivation acts by a Leibniz
rule on the structure maps:
1 ,...,U n 1 ,...,U n
X
DV mUV (α1 , . . . , αn ) = (−1)k(|α1 |+···+|αi−1 |) mU
V (α1 , . . . , DUi αi , . . . , αn ),
i
where the sign arises from the usual Koszul sign rule.
Example: Let us consider, again, the prefactorization algebra of the free scalar field
on Rn . Observables on U are Chevalley-Eilenberg chains of the Heisenberg algebra
Ebc (U).
Let X denote a Killing vector field on Rn (i.e., X is an infinitesimal isometry).
For example, we could take X to be a translation vector field ∂x∂ j . Then the Heisen-
berg dg Lie algebra has a derivation where φk 7→ Xφk , for k = 0, 1, and the central
element ~ is annihilated. (Recall that φk is notation for an element of the copy of
Cc∞ (U) situated in degree k.) Because X is a Killing vector field on Rn , it commutes
with the differential on the Heisenberg algebra and satisfies the equality
Z Z
(Xφ )φ + φ0 (Xφ1 ) = 0.
0 1
Hence, X is a derivation of dg Lie algebras.

By naturality, X extends to an endomorphism of Obsq (U) = C∗ (Ebc (U)). This
endomorphism defines a derivation of the prefactorization algebra Obsq of observ-
ables of the free scalar field theory. ^
All the derivations together Der∗ (F ) form a differential graded Lie algebra.
The differential is defined by (dD)U = [dU , DU ], where dU is the differential on
F (U). The Lie bracket is defined by [D, D0 ]U = [DU , D0U ].
The concept of derivation allows us to talk about the action of a dg Lie algebra
g on a prefactorization algebra F . Such an action is simply a homomorphism of
differential graded Lie algebras g → Der∗ (F ).
Next, we introduce some notation that will help us describe the smoothness
conditions for a discretely translation-invariant prefactorization algebra.
Let U1 , . . . , Uk ⊂ V be disjoint open subsets. Let Disj(U1 , . . . , Uk | V) ⊂ (Rn )k
be the set of k-tuples (x1 , . . . , xk ) such that the translates τ x1 U1 , . . . , τ xk Uk are still
disjoint and contained in V. It parametrizes the way we can move the open sets
without causing overlaps. Let us assume that Disj(U1 , . . . , Uk | V) has non-empty
interior, which happens when the closure of the Ui are disjoint and contained in V.
Let F be any discretely translation-invariant prefactorization algebra. For each
(x1 , . . . , xk ) ∈ Disj(U1 , . . . , Uk | V), we have a multilinear map obtained as a com-
position
m x1 ,...,xk : F (U1 ) × · · · × F (Uk ) → F (τ x1 U1 ) × · · · × F (τ xk Uk ) → F (V),
where the second map arises from the inclusion τ x1 U1 t · · · t τ xk Uk ,→ V.
8.1.3 Definition. A discretely translation-invariant prefactorization algebra F is
smoothly translation-invariant if the following conditions hold.
((i)) The map m x1 ,...,xk above depends smoothly on (x1 , . . . , xk ) ∈ Disj(U1 , . . . , Uk |
V).
((ii)) The prefactorization algebra F is equipped with an action of the Abelian
Lie algebra Rn of translations. If v ∈ Rn , we will denote the corre-
sponding action maps by
d
: F (U) → F (U).
dv
We view this Lie algebra action as an infinitesimal version of the global
translation invariance.
((iii)) The infinitesimal action is compatible with the global translation in-
variance in the following sense. For v ∈ Rn , let vi ∈ (Rn )k denote the
vector (0, . . . , v, . . . , 0), with v placed in the i-slot and 0 in the other
k − 1 slots. If αi ∈ F (Ui ), then we require that
!
d d
m x ,...,x (α1 , . . . , αk ) = m x1 ,...,xk α1 , . . . , αi , . . . , αk .
dvi 1 k dv
When we refer to a translation-invariant prefactorization algebra without fur-
ther qualification, we will always mean a smoothly translation-invariant prefactor-
ization algebra.
Example: We have already seen that the prefactorization algebra of the free scalar
field theory on Rn is discretely translation invariant and is also equipped with an
action of the Abelian Lie algebra Rn by derivations. It is easy to verify that this
prefactorization algebra is smoothly translation-invariant. ^
Example: This example is a special case of example 7.3.3.
Let F be the cohomology of the prefactorization algebra of observables of the
free scalar field theory on R with mass m. We have seen in Section 3 that the
corresponding associative algebra A is the Weyl algebra, generated by p, q, ~ with
commutation relation [p, q] = ~.
It is straightforward to verify that F is a locally-constant, smoothly translation
invariant prefactorization algebra on R, valued in vector spaces. As the action of R
on A is smooth, it differentiates to an infinitesimal action of the Lie algebra R on A
∂
by derivations. The basis element ∂x of R becomes a derivation H of A, called the
Hamiltonian. The Hamiltonian hereis given by H(a) = 2~1
[p2 − m2 q2 , a]. ^
8.2. An operadic reformulation. Next, we will explain how to think of the
structure of a translation-invariant prefactorization algebra on Rn in more operadic
terms. This description has a lot in common with the En algebras familiar from
topology.
Let r1 , . . . , rk , s ∈ R>0 . Let
Discsn (r1 , . . . , rk | s) ⊂ (Rn )k
be the (possibly empty) open subset consisting of k-tuples x1 , . . . , xk ∈ Rn with the
property that the closures of the balls Bri (xi ) are all disjoint and contained in Bs (0).
Here, Br (x) denotes the open ball of radius r around x.
8.2.1 Definition. Let Discsn be the R>0 -colored operad in the category of smooth
manifolds whose space of k-ary morphisms Discsn (r1 , . . . , rk | s) between ri , s ∈
R>0 is described above.
Note that a colored operad is the same thing as a multicategory. An R>0 -

colored operad is thus a multicategory whose set of objects is R>0 .
The essential data of the colored operad structure on Discsn is the following.
We have maps
◦i : Discsn (r1 , . . . , rk | ti )× Discsn (t1 , . . . , tm | s)
→ Discsn (t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tm | s).
This map is defined by inserting the outgoing ball (of radius ti ) of a configuration
x ∈ Discsn (r1 , . . . , rk | ti ) into the ith incoming ball of a point y ∈ Discsn (t1 , . . . , tk |
s). These maps satisfy the natural associativity and commutativity properties of a
multicategory.
Next, let F be a translation-invariant prefactorization algebra on Rn . Let
Fr = F (Br (0))
denote the cochain complex that F assigns to a ball of radius r. This notation is rea-
sonable because translation invariance gives us an isomorphism between F (Br (0))
and F (Br (x)) for any x ∈ Rn .
The structure maps for a translation-invariant prefactorization algebra yield,
for each configuration p ∈ Discsn (r1 , . . . , rk | s), a multiplication operation
m[p] : Fr1 × · · · × Frk → F s .
The map m[p] is a smooth multilinear map of differentiable spaces; and further-
more, this map depends smoothly on p.
These operations make the complexes Fr into an algebra over the R>0 -colored
operad Discsn (r1 , . . . , rk | s), valued in the multicategory of differentiable cochain
complexes. In addition, the complexes Fr are endowed with an action of the
Abelian Lie algebra Rn . This action is by derivations of the Discsn -algebra F
compatible with the action of translation on Discsn , as described above.
Now, let us unravel explicitly what it means to be such a Discsn algebra.
The first property is that, for each configuration p ∈ Discsn (r1 , . . . , rk | s),
the map m[p] is a multilinear map of cohomological degree 0, compatible with
differentials.
Second, let N be a manifold and let fi : N → Frdi i be smooth maps into the
space Frdi i of elements of degree di . The smoothness properties of the map m[p]
mean that the map
N × Discsn (r1 , . . . , rk | s) → Fs
(x, p) 7 → m[p]( f1 (x), . . . , fk (x))
is smooth. Thus, we can work with smooth families of multiplications.
Next, note that a permutation σ ∈ S k gives an isomorphism
σ : Discsn (r1 , . . . , rk | s) → Discsn (rσ(1) , . . . , rσ(k) | s).
We require that for each p ∈ Discsn (r1 , . . . , rk | s) and collection of elements
αi ∈ Fri ,
m[σ(p)](ασ(1) , . . . , ασ(k) ) = m[p](α1 , . . . , αk ).
Finally, we require that the maps m[p] are compatible with composition, in
the following sense. For any configurations p ∈ Discsn (r1 , . . . , rk | ti ) and q ∈
Discsn (t1 , . . . , tl | s) and for collections αi ∈ Fri and β j ∈ Ft j , we require that
m[q](β1 , . . . , βi−1 , m[p](α1 , . . . , αk ), βi+1 , . . . , βl )
= m[q ◦i p](β1 , . . . , βi−1 , α1 , . . . , αk , βi+1 , . . . , βl ).
In addition, the action of Rn on each F r is compatible with these multiplication
maps, in the manner described above.
8.3. A co-operadic reformulation. Let us give one more equivalent way of
rewriting these axioms, which will be useful when we discuss the holomorphic
context. These alternative axioms will say that the spaces C ∞ (Discsn (r1 , . . . , rk |
s)) form an R>0 -colored co-operad when we use the appropriate completed tensor
product. Since we know how to tensor a differentiable vector space with the space
of smooth functions on a manifold, it makes sense to talk about an algebra over
this colored co-operad in the category of differentiable cochain complexes.
We can rephrase the smoothness axiom for the product map
m[p] : Fr1 ⊗ · · · ⊗ Frk → F s ,
as p varies in Discsn (r1 , . . . , rk | s), as follows. For any differentiable vector space
V and smooth manifold M, we use the notation V⊗C ∞ (M) interchangeably with the
notation C ∞ (M, V); both indicate the differentiable vector space of smooth maps
M → V. The smoothness axiom states that the map above extends to a smooth
map of differentiable spaces
µ(r1 , . . . , rk | s) : Fr1 × · · · × Frk → F s ⊗ C ∞ (Discsn (r1 , . . . , rk | s)).
In general, if V1 , . . . , Vk , W are differentiable vector spaces and if X is a smooth
manifold, let
C ∞ (X, HomDVS (V1 , . . . , Vk | W))
denote the space of smooth multilinear maps V1 × · · · × Vk → C ∞ (X, W).
Note that there is a natural gluing map

◦i : C ∞ (X, HomDVS (V1 , . . . , Vk | Wi )) × C ∞ (Y, HomDVS (W1 , . . . , Wl | T ))
→ C ∞ (X × Y, HomDVS (W1 , . . . , Wi−1 , V1 , . . . , Vk , Wi+1 , . . . , Wl | T )).
With this notation in hand, there are elements
µ(r1 , . . . , rk | s) ∈ C ∞ Discs(r1 , . . . rk | s), HomDVS (Fr1 , . . . , Frk | F s )

with the following properties.

((i)) The map µ(r1 , . . . , rk | s) is closed under the natural differential, arising
from the differentials on the cochain complexes Fri .
((ii)) If σ ∈ S k , then
σ∗ µ(r1 , . . . , rk | s) = µ(rσ(1) , . . . , rσ(k) | s)
where
σ∗ : C ( Discs(r1 , . . . rk | s), HomDVS (Fr1 , . . . , Frk | F s ))
∞
→ C ∞ (Discs(rσ(1) , . . . rσ(k) | s), HomDVS (Frσ(1) , . . . , Frσ(k) | F s ))

is the natural isomorphism.
((iii)) As before, let
◦i : Discsn (r1 , . . . , rk | ti )× Discsn (t1 , . . . , tm | s)
→ Discsn (t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tm | s).
denote the gluing map. Then we require that
◦∗i µ(t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tl ) = µ(r1 , . . . , rk | ti ) ◦i µ(t1 , . . . , tl | s).
These elements equip the Fr with the structure of an algebra over the colored co-
operad, as stated earlier.
8.4. The free scalar field. Let us write down explicit formulas for these prod-
uct maps, as a Discsn algebra, in the case of the free scalar field theory. We have
seen in Section 6 that the choice of a Green’s function G leads to an isomorphism
of cochain complexes WU : Obscl (U)[~] → Obsq (U)[~] for every open subset U.
This isomorphism allows us to transfer the product (or structure map) in the pref-
actorization algebra Obsq to a deformed product ?~ in the prefactorization algebra
Obscl [~], defined by
α ?~ β = WUtV
−1
(WU (α) · WV (β)),
where U and V are disjoint open sets, α ∈ Obscl (U)[~], β ∈ Obscl (V)[~], and the
dot · indicates the structure map in the prefactorization algebra Obsq .
This isomorphism leads to a completely explicit description of the product
maps
µrs1 ,...,rk : Fr1 ⊗ · · · ⊗ Frk → C ∞ (P(r1 , . . . , rk | s), F s )
discussed above, in the case that F arises from the prefactorization algebra of
quantum observables of a free scalar field theory, or equivalently from the prefac-
torization algebra (Obscl [~], ?~ ).
For example, on R2 , let α1 , α2 be compactly supported smooth functions on

discs Bri (0) of radii ri around 0. Let us view each αi as a cohomological degree 0
element of
Fri = Obscl (Bri (0))[~] = Sym(Cc∞ (Bri (0))[1] ⊕ Cc∞ (Bri (0)))[~].
Let τ x αi denote the translate of αi to an element of Cc∞ (Bri (x)).
Then, for x1 , x2 ∈ R2 such that the Bri (xi ) are disjoint and contained in Bs (0),
we have
µrs1 ,r2 (α1 , α2 ) = τ x1 α1 ?~ τ x2 α2
Z
= (τ x1 α1 ) · (τ x2 α2 ) − ~ α1 (u1 + x1 )α2 (u2 + x2 ) log |u1 − u2 | .
u1 ,u2 ∈R2
Here, we used the Green’s function G(x, y) = − 2π

1
log |x − y|.
9. States and vacua for translation invariant theories

We develop notions such as state and vacuum in this setting of translation-
invariant prefactorization algebras. Our running example is the free scalar field.
As should be clear from their construction, the prefactorization algebra of ob-
servables only encodes the local relationships between the observables. Long-
range and global aspects of a physical situation appear in a different way, and this
notion of state provides one method by which to introduce them.The examples here
exhibit the role of fixing constraints on the asymptotic or boundary behavior of the
fields.
Remark: We do not pursue here a comparison with classic axioms for quantum field
theory, such as the Wightman or Osterwalder-Schrader axioms, but the familiar
reader will recognize the relationship. Both Kazhdan’s lectures in Deligne et al.
(1999) and Glimm and Jaffe (1987) cover such material and much more. ^
9.0.1 Definition. Let F be a smoothly translation-invariant prefactorization alge-
bra on Rn , over a ring R. (In practice, R is R, C or R[[~]], C[[~]]).
A state for F is a smooth linear map h−i : H ∗ (F (Rn )) → R.
A state h−i is translation invariant if it commutes with the action of both the
group Rn and of the infinitesimal action of the Lie algebra Rn , where Rn acts triv-
ially on R.
Recall that “smooth” means the following. The ring R itself is a differen-
tiable vector space: for instance, the real numbers R as a topological vector space
corresponds to the sheaf C ∞ on the site of smooth manifolds. As F (Rn ) is a dif-
ferentiable vector space, it is also a sheaf on the site of smooth manifolds, and
H ∗ (F (Rn )) denotes the cohomology sheaf on the site of smooth manifolds. Unrav-
eling the structures in the definition above, we find that for each smooth manifold
M and each element α ∈ H ∗ (C ∞ (M, F (Rn )), the element hαi is a smooth R-valued
function on M.
9. STATES AND VACUA FOR TRANSLATION INVARIANT THEORIES 115
A state h−i allows us to define correlation functions of observables (even if the

complex of fields is not acyclic). If Oi ∈ F (Bri (0)) are observables, then we can
construct τ xi Oi ∈ F (Bri (xi )). If the configuration (x1 , . . . , xk ) satisfies the condition
that the discs Bri (xi ) are disjoint — that is, (x1 , . . . , xk ) ∈ P(r1 , . . . , rk | ∞) — then
we define the correlation function
hO1 (x1 ) · · · On (xn )i
by applying the state h−i to the product observable τ x1 O1 · · · τ xn On ∈ F (Rn ). Be-
cause we have a smoothly translation invariant prefactorization algebra and the
state is a map of differentiable cochain complexes, we see that hO1 (x1 ) · · · On (xn )i
is a smooth function of the points xi , i.e., an element of C ∞ (P(r1 , . . . , rk | ∞), R).
Further, this function is invariant under simultaneous translation of all the points xi .
9.0.2 Definition. A translation-invariant state h−i is a vacuum if it satisfies the
cluster decomposition principle: in the situation above, for any two observables
O1 ∈ F (Br1 (0)) and O2 ∈ F (Br2 (0)), we have
hO1 (0) O2 (x)i − hO1 (0)i hO2 (0)i → 0 as x → ∞.
A vacuum is massive if hO1 (0) O2 (x)i − hO1 (0)i hO2 (0)i tends to zero exponentially
fast.
Example: This extended example examines vacua of the free scalar field with mass
m.
We have seen that the choice of a Green’s function G for the operator 4 + m2
leads to an isomorphism of cochain complexes Obscl (U)[~] Obsq (U) for every
open subsets U ⊂ Rn . This produces an isomorphism of prefactorization algebras
if we endow Obscl [~] with a deformed factorization product ?~ , defined using the
Green’s function.
If m > 0, there is a unique Green’s function G of the form G(x, y) = f (x − y),
where f is a distribution on Rn that is smooth away from the origin and tends
to zero exponentially fast at infinity. For example, if n = 1, the function f is
f (x) = 2m
1 −m|x|
e .
If m = 0, we already know

 4π1d/2 Γ(d/2 − 1) |x − y|2−n if n , 2

G(x, y) = 

− 1 log |x − y|

2π if n = 2
is the canonically-defined Green’s function.
It is standard to interpret the Green’s function as the two-point correlation func-
tion hδ x δ0 i. Hence, we see that this two-point function exhibits the necessary be-
havior: for m > 0, it decays exponentially fast, and it decays slowly in the massless
case. Strictly speaking, this observable is not in our prefactorization algebra, as
it is distributional, but it does exist at the cohomological level. Alternatively, we
could work with smeared versions.
Because Obscl (Rn ) is, as a cochain complex, the symmetric algebra on the
4+m2
complex Cc∞ (Rn )[1] −−−−→ Cc∞ (Rn ), there is a map from Obscl (Rn ) to R that is
the identity on Sym0 and sends Sym>0 to 0. (In other words, we have a natural
augmentation map.) This map extends to an R[~]-linear cochain map
h−i : Obscl (Rn )[~] → R[~].
Clearly, this map is translation invariant and smooth. Thus, because we have a
cochain isomorphism between Obscl (Rn )[~] and Obsq (Rn ), we have produced a
translation-invariant state.
9.0.3 Lemma. For m > 0, this state is a massive vacuum. For m = 0, this state is a
vacuum if n > 2; for n ≤ 2, it does not satisfy the cluster decomposition principle.
Proof. Let F1 ∈ Cc∞ (D(0, r1 ))⊗k1 and F2 ∈ Cc∞ (D(0, r2 ))⊗k2 . We will view
F1 , F2 as observables in Obscl (Bri (0)) by using the natural map from Cc∞ (U)⊗k to
the coinvariants Symk Cc∞ (U). Let τc F2 be the translation of F2 by c, where c is
sufficiently large so that the discs D(0, r1 ) and D(c, r2 ) are disjoint. Explicitly, τc F2
is represented by the function F2 (y1 −c, . . . , yk2 −c). We are interested in computing
the expected value hF1 (τc F2 )i.
We gave above an explicit formula for the quantum factorization product ?~
on Obscl [~]. In this case, it states that F1 ?~ τc F2 is given by
min(k
X 1 ,k2 ) X Z Z r
Y
~ r
F1 (x)F2 (τc y) G(xik , y jk ),
r=0 1≤i1 <···<ir ≤k1 x∈Rnk1 y∈Rnk2 k=1
1≤ j1 <···< jr ≤k2
where x = (x1 , . . . , xk1 ) and τc y = (y1 − c, . . . , yk2 − c). Note that after performing
the integral on the right hand side, we are left with a function of the k1 + k2 − 2r
copies of Rn that we have not integrated over. This function is then viewed as an
observable in Symk1 +k2 −2r Cc∞ (Rn ).
If k1 , k2 > 0, then hF1 i = 0 and hF2 i = 0. Further, hF1 (τc F2 )i selects the
constant term in the expression for F1 ?~ τc F2 . There is only a non-zero constant
term if k1 = k2 = k. In that case, the constant term hF1 , τc F2 i is
Z
~k F1 (x1 , . . . , xk )F2 (y1 − c, . . . , yk − c)G(x1 , y1 ) . . . G(xk , yk ).
xi ,yi ∈Rn
To check whether the cluster decomposition principle holds, we need to check

whether or not
hF1 (τc F2 )i = hF1 (τc F2 )i − hF1 i hF2 i
tends to zero as c → ∞. If m > 0, we know that G(x, y) tends to zero exponentially
fast as x − y → ∞. This implies immediately that we have a massive vacuum in this
case, because the correlation function involves integrating against powers of G.
If m = 0, the Green’s function G(x, y) tends to zero like the inverse of a poly-
nomial as long as n > 2. In this case, we have a vacuum. If n = 1 or n = 2, then
G(x, y) does not tend to zero, so we don’t have a vacuum.
This lemma simply reduces the full condition on a vacuum down to the Green’s
function, and hence to the usual perspective. ^
9. STATES AND VACUA FOR TRANSLATION INVARIANT THEORIES 117
Example: We can give a more abstract construction for the vacuum of associated
to a massive scalar field theory on Rn . It illuminates how boundary or asymptotic
conditions on fields relate to observables.
4+m2
Let E = C ∞ −−−−→ C ∞ be, as before, the complex of fields. Let Ebc be the
Heisenberg central extension Ebc = Ec ⊕ R · ~ where ~ has degree 1. We defined the
complex of observables on U as the Chevalley-Eilenberg chain complex of Ebc (U).
4+m2
Let ES (Rn ) be the complex S(Rn ) −−−−→ S(Rn ), where S(Rn ) is the space of
Schwartz functions on Rn . (Recall that a function is Schwartz if it tends to zero at
∞ faster than the reciprocal of any polynomial.)
A Heisenberg dg Lie algebra can be defined using ES (Rn ) instead R of Ec (R ).
n
We let EbS (Rn ) be ES (Rn ) ⊕ R · ~, with bracket defined by [φ0 , φ1 ] = ~ φ0 φ1 . Here

φk denotes a Schwartz function of cohomological degrees k = 0 or 1. This bracket
makes sense because the product of any two Schwartz function is Schwartz and
Schwartz functions are integrable.
Schwartz functions have a natural topology, so we will view them as being a
convenient vector space. Since the topology is nuclear Fréchet, a result discussed in
Appendix B tells us that the tensor product b ⊗β in the category of convenient vector
spaces coincides with the completed projective tensor product b ⊗π in the category
of nuclear Fréchet spaces. A result of Grothendieck (1952) tells us that
⊗S(Rm ) = S(Rn+m ),
S(Rn )b
and similarly for Schwartz sections of vector bundles.
This allows us to define the Chevalley-Eilenberg chain complex
q def
ObsS (Rn ) = C∗ (EbS (Rn ))
of the Heisenberg algebra based on Schwartz functions; as usual, we use the tensor
⊗β when defining the symmetric algebra.
product b
There’s a map
q
Obsq (Rn ) → ObsS (Rn )
given by viewing a compactly supported function as a Schwartz function. The
Schwartz observables are well-behaved, as we will show, because zero is the only
Schwartz solution to the equations of motion.
q
9.0.4 Lemma. The cohomology of ObsS (Rn ) is R[~], concentrated in degree zero.
Proof. The complex
4+m2
ES (Rn ) = S(Rn ) −−−−→ S(Rn )
has no cohomology. We can see this using Fourier duality: the Fourier transform
is an isomorphism on the space of Schwartz functions,Pand the Fourier dual of the
operator 4 + m2 is the operator p2 + m2 , where p2 = p2i and pi are coordinates
P 2
on the Fourier dual Rn . (Note that our convention is that the Laplacian is − ∂x∂ i .)
The operator of multiplication by p2 + m2 is invertible on the space of Schwartz
functions, just because if f is a Schwartz function then so is (p2 + m2 )−1 f , as
we assumed m , 0. The inverse is a smooth linear map, so that this complex is

smoothly homotopy equivalent to the zero complex.
The fact that the complex ES (Rn ) has no cohomology implies immediately that
the Chevalley-Eilenberg chain complex of EbS (Rn ) has the same cohomology as the
Chevalley-Eilenberg chain complex of the Abelian Lie algebra R~[−1]. But that
complex is R[~], as desired.
Thus, we have a translation-invariant state
q
h−i : H ∗ (Obsq (Rn )) → H ∗ (ObsS (Rn )) = R[~].
This state is characterized uniquely by the fact that it is defined on Schwartz ob-
servables. Since the state constructed more explicitly above also has this property,
we see that these two states coincide, so that this state is a massive vacuum. ^
CHAPTER 5
Holomorphic field theories and vertex algebras
This chapter serves two purposes. On the one hand, we develop several exam-
ples that exhibit how to understand the observables of a two-dimensional theory
from the point of view of prefactorization algebras and how this approach recovers
standard examples of vertex algebras. On the other hand, we provide a precise def-
inition of a prefactorization algebra on Cn whose structure maps vary holomorphi-
cally, much as we defined translation-invariant prefactorization algebras in Section
8. We then give a proof that when n = 1 and the prefactorization algebra possesses
a U(1)-action, we can extract a vertex algebra. For n > 1, the structure we find is
a higher-dimensional analog of a vertex algebra. (We leave a detailed analysis of
the structure present in the higher-dimensional case for future work). Such higher-
dimensional vertex algebras appear, for example, as the prefactorization algebra of
observables of partial twists of supersymmetric gauge theories.
0.1. Vertex algebras and holomorphic prefactorization algebras on C. In

mathematics, the notion of a vertex algebra is a standard formalization of the ob-
servables of a two-dimensional chiral field theory. In the version of the axioms we
consider, we do not ask for the Virasoro algebra to act. Before embarking on our
own approach, we recall the definition of a vertex algebra and various properties as
given in Frenkel and Ben-Zvi (2004).
0.1.1 Definition. Let V be a vector space. An element a(z) = n∈Z an z−n in
P
End V[[z, z−1 ]] is a field if, for each v ∈ V, there is some N such that a j v = 0
for all j > N.
Remark: The usage of the term “field” in the theory of vertex operators often pro-
vokes confusion. In this book, the term field is used to refer to a configuration in a
classical field theory: for example, in a scalar field theory on a manifold M, a field
is an element of C ∞ (M). The term “field” as used in the theory of vertex algebras
is not an example of this usage. Instead, as we will see below, a “field” in a vertex
algebra corresponds to an observable with support at a point of the surface; it can
be understand as a special kind of operator. ^
0.1.2 Definition. [Definition 1.3.1, Frenkel and Ben-Zvi (2004)] A vertex algebra
is the following data:
• a vector space V over C (the state space);
• a nonzero vector |0i ∈ V (the vacuum vector);
• a linear map T : V → V (the shift operator);
119
120 5. HOLOMORPHIC FIELD THEORIES AND VERTEX ALGEBRAS
• a linear map Y(−, z) : V → End V[[z, z−1 ]] sending every vector v to a

field (the vertex operation);
subject to the following axioms:
• (vacuum axiom) Y(|0i, z) = idV and Y(v, z)|0i ∈ v + zV[[z]] for all
v ∈ V;
• (translation axiom) [T, Y(v, z)] = ∂z Y(v, z) for every v ∈ V and T |0i =
0;
• (locality axiom) for any pair of vectors v, v0 ∈ V, there exists a nonneg-
ative integer N such that (z − w)N [Y(v, z), Y(v0 , w)] = 0 as an element
of End V[[z±1 , w±1 ]].
The vertex operation is best understood in terms of the following intuition. The
vector space V represents the set of pointwise measurements one can make of the
fields, and one should imagine labeling each point z ∈ C by a copy of V, which
we’ll denote Vz . Moreover, in a disc D containing the point z, the measurements
at z are a dense subspace of the measurements one can make in D (this is a feature
of chiral theories). We’ll denote the observables on D by VD . The vertex operation
is a way of combining pointwise measurements. Let D be a disc centered on the
origin. For z , 0, we can multiply observables to get a map
z
D Yz : V0 ⊗ Vz → VD .
0
This vertex operator should vary holomorphically in z ∈ D \ {0}. In other words,

we should get something with properties like the formal definition Y(−, z) above.
(This picture clearly resembles the “pair of pants” product from two-dimensional
topological field theories, although we’ve shrunk the two “incoming” boundary
circles to points.)
An appealing aspect of our approach to observables is that this intuition be-
comes explicit and rigorous. Our procedure provides structure maps from disjoint
discs into a larger disc (and also describes how to “multiply” observables supported
in more complicated opens). Moreover, it gives the structure maps in a coordinate-
free way. By choosing a coordinate z on C, we recover the usual formulas for
vertex algebras.
This relationship is seen in the main theorem in this chapter, which we now
state. The theorem connects vertex algebras with a certain class of prefactorization
algebras on C. The prefactorization algebras of interest are holomorphically trans-
lation invariant prefactorization algebra. We will give a precise definition of what
this means shortly, but here is a rough definition. Let F be such a prefactorization
5. HOLOMORPHIC FIELD THEORIES AND VERTEX ALGEBRAS 121
algebra. Let Fr denote the cochain complex F (D(0, r)) that F assigns to a disc
of radius r. Recall, as we explained in Section 8 of Chapter 4, if F is smoothly
translation invariant, then we have an operator product map
µ : Fr1 ⊗ · · · ⊗ Frn → C ∞ (Discs(r1 , . . . , rk | s), F s ),
where Discs(r1 , . . . , rk | s) refers to the open subset of Ck consisting of points
z1 , . . . , zk such that the discs of radius ri around zi are all disjoint and contained in
the disc of radius s around the origin. If F is holomorphically translation invariant,
then µ lifts to a cochain map
µ : Fr1 ⊗ · · · ⊗ Frn → Ω0,∗ (Discs(r1 , . . . , rk | s), F s ),
where on the right hand side we have used the Dolbeault complex of the complex
manifold Discs(r1 , . . . , rk | s). We also require some compatibility of these lifts
with compositions, which we will detail later.
In other words, being holomorphically translation invariant means that the op-
erator product map is holomorphic (up to homotopy) in the location of the discs.
0.1.3 Theorem. Let F be a holomorphically translation invariant prefactorization
algebra on C. Let F be equivariant under the action of S 1 on C by rotation, and
let Frk denote the weight k eigenspace of the S 1 action on the complex Fr . Assume
that for every r < s, the extension map Frk → F sk associated to the inclusion
D(0, r) ⊂ D(0, s) is a quasi-isomorphism. Finally, we need to assume that the S 1
action on each Fr satisfies a certain technical “tameness” condition.
Then the vector space M
VF = H ∗ (Frk )
k∈Z
has the structure of a vertex algebra. The vertex algebra structure map
YF : VF ⊗ VF → VF [[z, z−1 ]]
is the Laurent expansion of operator product map
H ∗ µ : H ∗ (Frk11 ) ⊗ H ∗ (Frk22 ) → Hol(Discs(r1 , r2 | s), H ∗ (F s )).
On the right hand side, Hol denotes the space of holomorphic maps.
In other words, the intuition that the vertex algebra structure map is the opera-
tor product expansion is made precise in our formalism.
This result should be compared with the classic result of Huang (1997), who
relates vertex algebras with chiral conformal field theories at genus 0, in the sense
used by Segal. As we have seen in Section 4, the axioms for prefactorization
algebras are closely related to Segal’s axioms. Our axioms for a holomorphically
translation invariant field theory are similarly related to Segal’s axioms for a two-
dimensional chiral field theory. Although our result is closely related to Huang’s,
it is a little different because of the technical differences between a prefactorization
algebra and a Segal-style chiral conformal field theory.
One nice feature of our definition of holomorphically translation-invariant pref-
actorization algebra is that it makes sense in any complex dimension. The structure
present on the cohomology of a higher-dimensional holomorphically translation-

invariant prefactorization algebra is a higher dimensional version of the axioms of
a vertex algebra. This structure is discussed briefly in Costello and Scheimbauer
(2015).
Another nice feature of our approach is that our general construction of field
theories allows one to construct vertex algebras by perturbation theory, starting
with a Lagrangian. This method should lead to the construction of many interesting
vertex algebras, and of their higher dimensional analogs.
0.2. Organization of this chapter. We will start this chapter by stating and
proving our main theorem. The first thing we define, in Section 1, is the notion
of a holomorphically translation invariant prefactorization algebra on Cn for every
n ≥ 1. Holomorphic translation invariance guarantees that the operator products
are all holomorphic. We will then show to construct a vertex algebra from such an
object in dimension 1.
The rest of this chapter is devoted to analyzing examples. In Section 3, we
discuss the prefactorization algebra associated to a very simple two-dimensional
chiral conformal field theory: the free βγ system. We will show that the vertex
algebra associated to the prefactorization algebra of observables of this theory is
an object called the βγ vertex algebra in the literature. Then, in Section 4, we
will construct a prefactorization algebra encoding the affine Kac-Moody algebra.
This prefactorization algebra again encodes a vertex algebra, which is the standard
Kac-Moody vertex algebra.
1. Holomorphically translation-invariant prefactorization algebras

In this section we will analyze in detail the notion of translation-invariant pref-
actorization algebras on Cn . On Cn we can ask for a translation-invariant pref-
actorization algebra to have a holomorphic structure; this implies that all struc-
ture maps of the prefactorization algebra are (in a sense we will explain shortly)
holomorphic. There are many natural field theories where the corresponding pref-
actorization algebra is holomorphic: for instance, chiral conformal field theories
in complex dimension 1, and minimal twists of supersymmetric field theories in
complex dimension 2, as described in Costello (2011b).
1.1. The definition. We now explain what it means for a (smoothly) translation-
invariant prefactorization algebra F on Cn to be holomorphically translation-invariant.
For this definition to make sense, we require that F is defined over C: that is, the
vector spaces F (U) are complex vector spaces and all structure maps are complex-
linear.
Recall that such a prefactorization algebra has, as part of its structure, an action
of the real Lie algebra R2n = Cn by derivations. This action is as a real Lie algebra.
Since F is defined over C, the action extends to an action of the complexified
translation Lie algebra R2n ⊗R C. We will denote the action maps by
∂ ∂
, : F (U) → F (U).
∂zi ∂z j
1. HOLOMORPHICALLY TRANSLATION-INVARIANT PREFACTORIZATION ALGEBRAS 123
1.1.1 Definition. A translation-invariant prefactorization algebra F on Cn is holo-

morphically translation-invariant if it is equipped with derivations ηi : F → F of
cohomological degree −1, for i = 1 . . . n, with the following properties:
∂
dηi = ∈ Der(F )
∂zi
[ηi , η j ] = 0,
∂
" #
ηi , = 0.
∂zi
Here, d refers to the differential on the dg Lie algebra Der(F ).
We should understand this definition as saying that the vector fields ∂z∂ i act
homotopically trivially on F . In physics terminology, the operators ∂x∂ i , where xi
are real coordinates on Cn = R2n , are related to the energy-momentum tensor. We
are asking that the components of the energy momentum tensor in the zi directions
are exact for the differential on observables (in physics, this might be called “BRST
exact”). This is rather similar to a phenomenon that sometimes appears in the study
of topological field theory, in which a topological theory depends on a metric, but
the variation of the metric is exact for the BRST differential. See Witten (1988) for
discussion.
1.2. An operadic reformulation. Now we will interpret holomorphically translation-

invariant prefactorization algebras in the language of R>0 -colored operads. When
we work in complex geometry, it is better to use polydiscs instead of balls, as is
standard in complex analysis.
Thus, if z ∈ Cn , let
PDr (z) = {w ∈ Cn | |wi − zi | < r for 1 ≤ i ≤ n}
denote the polydisc of radius r around z. Let
PDiscsn (r1 , . . . , rk | s) ⊂ (Cn )k
denote the set of z1 , . . . , zk ∈ Cn with the property that the closures of the polydiscs
PDri (zi ) are disjoint and contained in the polydisc PD s (0).
The spaces PDiscsn (r1 , . . . , rk | s) form a R>0 -colored operad in the category of
complex manifolds. We will explain here why a holomorphically translation invari-
ant prefactorization algebra provides an algebra over this colored operad. Proposi-
tion 1.3.1 will provide a precise version and proof of this argument.
Now, let F be a holomorphically translation-invariant prefactorization algebra
on Cn . Let Fr denote the differentiable cochain complex F (PDr (0)) associated to
the polydisc of radius r. For each p ∈ PDiscsn (r1 , . . . , rk | s), we have a map
m[p] : Fr1 × · · · × Frk → F s .
This map is smooth, multilinear, and compatible with the differential. Further, this
map varies smoothly with p.
The fact that F is a holomorphically translation-invariant prefactorization al-

gebra means that these maps are equipped with extra structure: we have derivations
η j of F that make the derivations ∂z∂ j homotopically trivial.
For i = 1, . . . , k and j = 1, . . . , n, let zi j , zi j refer to coordinates on (Cn )k . We
thereby obtain coordinates on the open subset
PDiscsn (r1 , . . . , rk | s) ⊂ (Cn )k .
Thus, we have operations
∂
m[p] : Fr1 × · · · × Frk → F s .
∂zi j
obtained by differentiating the operation m[p], which depends smoothly on p, in
the direction zi j .
Let m[p] ◦i η j denote the operation
m[p] ◦i η j : Fr1 × · · · × Frk → Fs
α1 × · · · × αk 7 → (−1)|α1 |+···|αi−1 | m[p](α1 , . . . , η j αi , . . . , αk ),
where the sign arises from the usual Koszul rule.
The axioms of a (smoothly) translation invariant prefactorization algebra tell
us that
∂ ∂
m[p] = m[p] ◦i ,
∂zi j ∂z j
∂
where ∂z j is the derivation of the prefactorization algebra F . This equality, together
∂
with the fact that [d, ηi ] = ∂z j , tells us that
∂ h i
m[p] = d, m[p] ◦i η j
∂zi j
holds. Hence the product map m[p] is holomorphic in p, up to a homotopy given
by ηi .
1.3. A cooperadic reformulation. In the smooth case, we saw that we could

also describe the structure as that of an algebra over a R>0 -colored cooperad built
from smooth functions on the spaces Discsn (r1 , . . . , rk | s). In this section we
will see that there is an analogous story in the complex world, where we use the
Dolbeault complex of the spaces PDiscsn (r1 , . . . , rk | s).
Because the spaces PDiscsn (r1 , . . . , rk | s) form a colored operad in the cate-
gory of complex manifolds, their Dolbeault complexes form a colored cooperad in
the category of convenient cochain complexes. We use here that the contravariant
functor sending a complex manifold to its Dolbeault complex, viewed as a conve-
nient vector space, is symmetric monoidal:
Ω0,∗ (X × Y) = Ω0,∗ (X)b
⊗β Ω0,∗ (Y)
where b⊗β denotes the symmetric monoidal structure on the category of convenient
vector spaces.
Explicitly, the colored cooperad structure is given as follows. The operad struc-
ture on the complex manifolds PDiscsn (r1 , . . . , rk | ti ) is given by maps
◦i : PDiscsn (r1 , . . . , rk | ti ) × PDiscsn (t1 , . . . , tm | s)

→ PDiscsn (t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tm | s).
We let
◦∗i : Ω0,∗ (PDiscsn (t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tm | s))

→ Ω0,∗ (PDiscsn (r1 , . . . , rk | ti ) × PDiscsn (t1 , . . . , tm | s))
be the corresponding pullback map on Dolbeault complexes.

The prefactorization algebras we are interested in take values in the category
of differentiable vector spaces. We want to say that if F is a holomorphically trans-
lation invariant prefactorization algebra on Cn , then it defines an algebra over the
colored cooperad given by the Dolbeault complex of the spaces PDiscsn . A priori,
this does not make sense, because the category of differentiable vector spaces is
not a symmetric monoidal category; only its full subcategory of convenient vector
spaces has a tensor structure. However, in order to make this definition, all we
need to be able to do is to tensor differentiable vector spaces with the Dolbeault
complexes of complex manifolds, and this we know how to do.
Recall from Appendix B that for any manifold X, C ∞ (X) defines a differen-
tiable vector space (in fact, a commutative algebra in DVS). If V is any differ-
entiable vector space, then there is a differentiable vector space C ∞ (X, V) whose
value on a manifold M is C ∞ (M × X, V). If X is a complex manifold, then Ω0,∗ (X)
is a differentiable cochain complex, for the same reasons. We define Ω0,∗ (X, V) as
the tensor product
Ω0,∗ (X, V) := Ω0,∗ (X) ⊗C ∞ (X) C ∞ (X, V).
as sheaves of C ∞ (X)-modules on the site of smooth manifolds. This is a reasonable

thing to do as Ω0,∗ (X) is a projective module over C ∞ (X). Note also that the fact
that V is a differentiable vector space, and not just a sheaf of C ∞ -modules on the
site of smooth manifolds, means that differential operators on X act on C ∞ (X, V).
This is what allows us to extend the differential ∂ on the Dolbeault complex Ω0,∗ (X)
to an operator on Ω0,∗ (X, V).
The Dolbeault complex with coefficients in V is functorial both for maps f ∗ :
Ω (X) → Ω0,∗ (Y) that arise from holomorphic maps f : Y → X and also for
0,∗
arbitrary smooth maps between differentiable vector spaces. Since the cooperad
structure maps in our colored dg cooperad Ω0,∗ (PDiscsn ) arise from maps of com-
plex manifolds, it makes sense to ask for a coalgebra over this cooperad in the
category of differentiable vector spaces.
1.3.1 Proposition. Let F be a holomorphically translation-invariant
prefactoriza-

tion algebra on Cn . Then F defines an algebra over the cooperad Ω0,∗ (PDiscsn )
n∈N
in differentiable cochain complexes.
PDt (0)
PDr (0)
µ∂
0 PD s (z) Fr × F s Ω0,∗ (PDiscs(r, s | t), Ft )
Figure 1. Configurations of discs and the operator product map
More precisely, the product maps

m[p] : Fr1 × · · · × Frk → F s
for p ∈ PDiscsn (r1 , . . . , rk | s) lift to multilinear maps
µ∂ (r1 , . . . , rk | s) : Fr1 × · · · × Frk → Ω0,∗ (PDiscsn (r1 , . . . , rk | s), F s ))

that are compatible with differentials and that satisfy the properties needed to de-
fine a coalgebra over a cooperad.
In other words, we have closed elements
µ∂ (r1 , . . . , rk | s) ∈ Ω0,∗ (PDiscsn (r1 , . . . , rk | s), HomDVS (Fr1 , . . . , Frk | F s ))

satisfying the following properties.
((i)) The element µ∂ (r1 , . . . , rk | s) is closed under the natural differential on
Ω0,∗ (PDiscsn (r1 , . . . , rk | s), HomDVS (Fr1 , . . . , Frk | F s )),
which incorporates the Dolbeault differential as well as the internal
differentials on the complexes Fri , F s . Explicitly, the differential

∂
(dF + ∂)µ (r1 , . . . , rk | s) ( f1 , . . . , fk )
is given by the sum
±µ∂ (r1 , . . . , rk | s)( f1 , . . . , dF fi , . . . , fk ).

X
((ii)) Let σ ∈ S k . As in the smooth case,
σ∗ µ∂ (r1 , . . . , rk | s) = µ∂ (rσ(1) , . . . , rσ(k) | s)

where σ∗ is induced by the isomorphism of complex manifolds
σ : PDiscsn (r1 , . . . , rk | s) → PDiscsn (rσ(1) , . . . , rσ(k) | s).
((iii)) For all 1 ≤ i ≤ m,
◦∗i µ∂ (t1 , . . . , ti−1 , r1 , . . . , rk , ti+1 , . . . , tm | s)

= µ∂ (t1 , . . . , tm | s) ◦i µ∂ (r1 , . . . , rk | ti ).
That is, we have associativity of composition.
Remark: For similar axiom systems in the context of topological field theory, the
interested reader should consult, for example, Getzler (1994) and Costello (2007).
In Segal (2004) there is a related system of axioms for chiral conformal field theo-
ries. This construction is also closely related to the construction of “descendants”
in the physics literature on topological field theory (see, for example, Witten (1988,
1992)).
Here is a brief explanation of the relationship. Suppose one has a field theory
that depends on a Riemannian metric, but where the variation of the original action
functional with respect to the metric is exact for the BRST operator (which cor-
responds to the differential on observables in our language). A metric-dependent
functional making the variation BRST exact can be viewed as a one-form on the
moduli space of metrics. Higher homotopies yield forms on the moduli space of
metrics, or, in two dimensions, on the moduli of conformal classes of metrics (i.e.,
on the moduli of Riemann surfaces).
In our approach, because we are working with holomorphic instead of topolog-
ical theories, we find elements of the Dolbeault complex of the appropriate moduli
spaces of complex manifolds. In the simple situation considered here, these moduli
spaces are the spaces PDiscsn .
For the factorization algebras associated to holomorphic twists of supersym-
metric field theories, these elements of the Dolbeault complex of PDiscsn come
from the operator product of supersymmetric descendents of BPS operators. ^
Proof of the proposition. Giving a smooth multilinear map
φ : Fr1 × · · · × Frk → Ω0,∗ (PDiscsn (r1 , . . . , rk | s), F s )
compatible with the differentials is equivalent to giving an element of
φ ∈ Ω0,∗ PDiscsn (r1 , . . . , rk | s), HomDVS (Fr1 , . . . , Frk | F s )

that is closed with respect to the differential ∂ + dF , where dF refers to the natu-
ral differential on the differentiable cochain complex of smooth multilinear maps
HomDVS (Fr1 , . . . , Frk | F s ).
We will produce the desired element
µ∂ (r1 , . . . , rk | s) ∈ Ω0,∗ PDiscsn (r1 , . . . , rk | s), HomDVS (Fr1 , . . . , Frk | F s

starting from the operations

µ0 (r1 , . . . , rk | s) ∈ C ∞ (PDiscs(r1 , . . . , rk | s), HomDVS (Fr1 , . . . , Frk | s))
provided by the hypothesis that F is a smoothly translation-invariant prefactoriza-
tion algebra.
First, we need to introduce some notation. Recall that

PDiscsn (r1 , . . . , rk | s) ⊂ Cn k

is an open (possibly empty) subset. Thus,

Ω0,∗ (PDiscsn (r1 , . . . , rk | s)) = Ω0,0 (PDiscsn (r1 , . . . , rk | s)) ⊗ C[dzi j ]
where the dzi j are commuting variables of cohomological degree 1, with i = 1, . . . , k
and j = 1, . . . , n. We let ∂(dz∂ i j ) denote the graded derivation that removes dzi j .
As before, let η j : Fr → Fr denote the derivation that cobounds the derivation
∂
∂z j . We can compose any element
(†) α ∈ Ω0,∗ (PDiscn (r1 , . . . , rk | s), Hom(Fr1 , . . . , Frk | s))

with η j acting on Fri , to get
α ◦i η j ∈ Ω0,∗ (PDiscn (r1 , . . . , rk | s), Hom(Fr1 , . . . , Frk | s)).
We use the shorthand notations:
li j (α) = α ◦i η j ,
∂
ri j (α) = α ◦i ,
∂z j
where ∂z∂ j refers to the derivation acting on Fr j .
Note that
[dF , li j ] = ri j ,
where dF is the differential on the graded vector space (†) above arising from the
differentials on the spaces Fri , F s . Further, the operators li j , ri j all commute with
each other in the graded sense.
In what follows, for concision, we will write µ∂ instead of µ∂ (r1 , . . . , rk | s),
and similarly for µ0 .
The cochains µ∂ are defined by
µ∂ = exp(−
X
dzi j li j )µ0 .
i, j
Here dzi j li j denotes the operation of wedging with dzi j after applying the operator
li j . Note that these two operators graded-commute.
Next, we need to verify that (∂ + dF )µ∂ = 0, where ∂ + dF is the differential
on graded vector space (†) above that arises by combining the ∂ operator with the
differential arising from the complexes Fri , F s .
We use the following identities:
X X
[dF , dzi j li j ] = dzi j ri j ,
i, j i, j
∂ 0
ri j µ0 = µ ,
∂zi j
dF µ0 = 0.
2. A GENERAL METHOD FOR CONSTRUCTING VERTEX ALGEBRAS 129
The second and third identities are part of the axioms for a smoothly translation-
invariant prefactorization algebra.
These identities allows us to calculate that
(∂ + dF )µ∂ = (∂ + dF ) exp(−
X
dzi j li j )µ0
X X X
= − exp − dzi j li j dzi j ri j µ0 + exp − dzi j li j (∂ + dF )µ0
∂
X X !
= exp − dzi j li j − dzi j + ∂ µ0
∂zi j
= 0.
Thus, µ∂ is closed.
It is straightforward to verify that the elements µ∂ are compatible with compo-
sition and with the symmetric group actions.
2. A general method for constructing vertex algebras

In this section we will prove that the cohomology of a holomorphically trans-
lation invariant prefactorization algebra on C with a compatible circle action gives
rise to a vertex algebra. Together with the central theorem of the second volume,
which allows one to construct prefactorization algebras by obstruction theory start-
ing from the Lagrangian of a classical field theory, this gives a general method to
construct vertex algebras.
2.1. The circle action. Recall from Section 7 the definition of a smoothly G-
equivariant prefactorization algebra on a manifold M with smooth action of a Lie
group G. The case of interest here is the action of the isometry group S 1 n C of C
acting on C itself.
2.1.1 Definition. A holomorphically translation-invariant prefactorization algebra
F on C with a compatible S 1 action is a smoothly S 1 n R2 -invariant prefactoriza-
tion algebra F , defined over the base field of complex numbers, together with an
extension of the action of the complex Lie algebra
LieC (S 1 n R2 ) = C ∂θ , ∂z , ∂z ,

where ∂θ is a basis of LieC (S 1 ), to an action of the dg Lie algebra

C ∂θ , ∂z , ∂z ⊕ C{η},

where η is of cohomological degree −1 and the differential is

dη = ∂z .
In this dg Lie algebra, all commutators involving η vanish except for
[∂θ , η] = −η.
Note that, in particular, F is a holomorphically translation invariant prefactor-

ization algebra on C.
Our prefactorization algebras take values in differentiable cochain complexes.
Here we will work with such vector spaces over the complex numbers; in other
words, every differentiable vector space E discussed here will assign a complex
vector space E(M) to a manifold M, the restriction maps are complex-linear, and
so on.
By their very definition, we know how to differentiate “smooth maps to a differ-
entiable vector space.” Thus, if M is a complex manifold and E is a differentiable
cochain complex, we have, for every open U ⊂ M, a cochain map
∂ : C ∞ (U, E) → Ω0,1 (U, E).
These maps form a map of sheaves of cochain complexes on M. Taking the coho-
mology sheaves of E, we obtain a graded differentiable vector space H ∗ (E) and we
have a map of sheaves of graded vector spaces
∂ : C ∞ (U, H ∗ (E)) → Ω0,1 (U, H ∗ (E)).
The kernel of this map defines a sheaf on M, whose sections we denote by Hol(U, H ∗ (E)).
We call these the holomorphic sections of H ∗ (E) on M.
2.1.1. The theorem on vertex algebras will require an extra hypothesis re-
garding the S 1 -action on the prefactorization algebra. Given any compact Lie
group G, we will formulate a concept of tameness for a G-action on a differen-
tiable vector space. We will require that the S 1 -action on the spaces Fr in our
factorization algebra is tame.
Note that for any compact Lie group G, the space D(G) of distributions on G
is an algebra under convolution. Since spaces of distributions are naturally differ-
entiable vector spaces, and the convolution product
∗ : D(G) × D(G) → D(G)
is smooth, D(G) forms an algebra in the category of differentiable vector spaces.
There is a map
δ : G → D(G)
sending an element g to the δ-distribution at g. It is a smooth map and a homomor-
phism of monoids.
2.1.2 Definition. Let E be a differentiable vector space and G be a compact Lie
group. A tame action of G on E is a smooth action of the algebra D(G) on E.
Note that this is, in particular, an action of G on E via the smooth map of groups
G → D(G)× sending g to δg .
If E is a differentiable cochain complex, a tame action is an action of D(G) on
E which commutes with the differential on E.
Let us now specialize to the case when G = S 1 , which is the case relevant
for the theorem on vertex algebras. For each integer k, there is an irreducible
representation ρk of S 1 given by the function ρk : λ 7→ λk . (Here we use λ to
denote a complex number by viewing S 1 as a subset of C. ) We can view ρk as a
map of sheaves on the site of smooth manifolds from the sheaf represented by the
manifold S 1 to the differentiable vector space C ∞ , which assigns complex-valued
ρk of the
smooth functions to each manifold. It lifts naturally to a representation e
algebra D(S 1 ), defined by

ρk (φ) = hφ, ρk i ,
e
where h−, −i indicates the pairing between distributions and functions. We will
abusively denote this representation by ρk , for simplicity.
If E is a differentiable vector space equipped with such a smooth action of
D(S 1 ), we use Ek ⊂ E to denote the subspace on which D(S 1 ) acts by ρk . We call
this the weight k eigenspace for the S 1 -action on E.
In the algebra D(S 1 ), the element ρk , viewed as a distribution on S 1 , is an
idempotent. If we denote the action of D(S 1 ) on E by ∗, then the map ρk ∗ − : E →
E defines a projection from E onto Ek . We will denote this projection by πk .
Remark: Of course, all this holds for a general compact Lie group where instead
of these ρk , we use the characters of irreducible representations. ^
2.2. The statement of the main theorem. Now we can state the main theo-
rem of this section.
2.2.1 Theorem. Let F be a unital S 1 -equivariant holomorphically translation in-
variant prefactorization algebra on C valued in differentiable vector spaces. As-
sume that, for each disc D(0, r) around the origin, the action of S 1 on F (D(0, r) is
tame. Let Fk (D(0, r)) be the weight k eigenspace of the S 1 action on F (D(0, r)).
We make the following additional assumptions.
((i)) Assume that, for r < r0 , the structure map
Fk (D(0, r)) → Fk (D(0, r0 ))
is a quasi-isomorphism.
((ii)) For k 0, the vector space H ∗ (Fk (D(0, r)) is zero.
((iii)) For each k and r, we require that H ∗ (Fk (D(0, r)) is isomorphic, as a
sheaf on the site of smooth manifolds, to a countable sequential colimit
of finite-dimensional graded vector spaces.
Let Vk = H ∗ (Fk (D(0, r)), and let V =
L
k∈Z Vk . (This space is independent of
r by assumption.) Then V has the structure of a vertex algebra, determined by the
structure maps of F .
The construction is functorial: it will be manifest that a map of factorization
algebras respecting all the equivariance conditions produces a map between the
associated vertex algebras.
We give the proof in Section 2.4 below. We spell out the data of the vertex
algebra — the vacuum vector, translation operator, vertex operation, and so on —
as we construct it. First, though, we discuss issues around applying the theorem.
Remark: If V is not concentrated in cohomological degree 0, then it will have the
structure of a vertex algebra valued in the symmetric monoidal category of graded
vector spaces. That is, the Koszul rule of signs will apear in the axioms. ^
Remark: We will often deal with prefactorization algebras F equipped with a com-
plete decreasing filtration F i F , so that F = lim F /F i F . In this situation, to con-
struct the vertex algebra we need that the properties listed in the theorem hold on
each graded piece Gri F . This condition implies, by a spectral sequence, that they
hold on each F /F i F , allowing us to construct an inverse system of vertex algebras
associated to the prefactorization algebra F /F i F . The inverse limit of this system
of vertex algebras is the vertex algebra associated to F . ^
2.3. Using the theorem. The conditions of the theorem are always satisfied
in practice by prefactorization algebras arising from quantizing a holomorphically
translation invariant classical field theory. We describe in great detail the example
of the free βγ system in Section 3.
There is also the problem of recognizing the vertex algebra produced by such
a prefactorization algebra. Thankfully, there is a useful “reconstruction” theorem
that provides simple criteria to uniquely construct a vertex algebra given “genera-
tors and relations.” We will exploit this theorem to verify that we have recovered
standard vertex algebras in our examples later in this chapter.
2.3.1 Theorem (Reconstruction, Theorem 4.4.1, Frenkel and Ben-Zvi (2004)).
Let V be a complex vector space equipped with a nonzero vector |0i, an endomor-
phism T , a countable ordered set {aα }α∈S of vectors, and fields
aα (z) = aα(n) z−n−1 ∈ End(V)[[z, z−1 ]]

X
n∈Z
such that
((i)) for all α, aα (z)|0i = aα + O(z);
((ii)) T |0i = 0 and [T, aα (z)] = ∂z aα (z) for all α;
((iii)) all fields aα (z) are mutually local;
((iv)) V is spanned by the vectors
aα( j11 ) · · · aα( jmm ) |0i
with the ji < 0.

Then, using the formula
1
Y(aα( j11 ) · · · aα( jmm ) |0i, z) = : ∂ 1 aα1 (z) · · · ∂z m aαm (z) :
− j −1 − j −1
(− j1 − 1)! · · · (− jm − 1)! z
to define a vertex operation, we obtain a well-defined and unique vertex algebra
(V, |0i, T, Y) such that Y(aα , z) = aα (z).
Here : a(z)b(w) : denotes the normally ordered product of fields, defined as
: a(z)b(w) := a(z)+ b(w) + b(w)a(z)−
where
X X
a(z)+ = an zn and a(z)− = an zn .
n≥0 n<0
Normal ordering eliminates various “divergences” that appear in naively taking

products of fields.
2.4. The proof of the main theorem. The strategy of the proof is as follows.
We will analyze the structure on V given to us by the axioms of a translation-
invariant prefactorization algebra. The prefactorization product will become the
operator product expansion or state-field map. The locality axiom of a vertex alge-
bra will follow from the associativity axioms of the prefactorization algebra.
2.4.1. The vector space structure. Let us begin by describing important fea-
tures of V and its natural completion V.
Because our prefactorization algebra F is translation invariant, the cochain
complex F (D(z, r)) associated to a disc of radius r is independent of z. We there-
fore use the notation F (r) for F (D(z, r)) and F (∞) to denote F (C).
Let Fk (r) denote the weight k space of the S 1 action on F . The S 1 action on
each F (r) extends to an action of D(S 1 ). The projection map F (r) → Fk (r) onto
the weight k space induces a map at the level of cohomology that we also denote
πk : H ∗ (F (r)) → H ∗ (Fk (r)) = Vk .
It is a map of differentiable vector spaces, splitting the natural inclusion. Fur-
thermore, the extension map H ∗ (Fk (r)) → H ∗ (Fk (r0 )) associated to the inclusion
D(0, r) ,→ D(0, r0 ) is the identity on V.
By assumption, the differentiable vector space Vk = H ∗ (Fk (r)) is the colimit of
its finite-dimensional subspaces. This means that a section of Vk on the manifold
M is given, locally on M, by a smooth map (in the ordinary sense) to a finite
dimensional subspace of Vk .
By construction there is a canonical map
M
V= Vk → H ∗ (F (r))
k
for every r > 0. We emphasize that V need not be the limit limr→0 H ∗ (F (r)).
Instead, it is an algebraically tractable vector space mapping into that limit (so to
speak, the finite sums of modes).
We introduce now a natural partner to V into which all the H ∗ (F (r)) map. Let
Y
V= Vk ,
k
where the product is taken in the category of differentiable vector spaces. There is
a canonical map
Y
πk : H ∗ (F (r)) → V
k
given by the product of all the projection maps. We emphasize that V need not
be the colimit colimr→∞ H ∗ (F (r)). Instead, it is an algebraically tractable vector
space receiving a map from that colimit. Note that V is a completion of V in the
sense that we now allow arbitrary, infinite sums of the modes.
2.4.2. The operator product. Let us start to analyze the structure on V given
by the prefactorization algebra structure on F . Our approach will be to use instead
the structure as an algebra over the cooperad Ω0,∗ (PDiscs1 ). (Since a 1-dimensional
polydisc is simply a disc, we will simply write disc from hereon.) We will call the
structure maps for this algebra the operator product. Recall that we use
µ∂ (r1 , . . . , rk | ∞) : F (r1 ) × · · · × F (rk ) → Ω0,∗ (Discs(r1 , . . . , rk | ∞), F (∞))

to denote such a multilinear cochain map.
We want to use this operator product to produce a map
mz1 ,...,zk : V ⊗k → V
for every configuration (z1 , . . . , zk ) of k ordered distinct points in C. This map will
vary holomorphically over that configuration space.
Note that for any complex manifold X there is a truncation map

(−)0 : H ∗ Ω0,∗ (X, F (∞)) → Hol(X, H ∗ (F (∞))).
If α ∈ Ω0,∗ (X, F (∞)) is a cocycle, this map is defined by first extracting the com-
ponent α0 of α in Ω0,0 (X, F (∞)). The fact that α is closed means that α0 is closed
for the differential DF on F (∞) and that ∂α0 is exact. Thus, the cohomology class
[α0 ] is a section of F (∞) on X, and ∂[α0 ] = 0 so that [α0 ] is holomorphic.
Hence, at the level of cohomology, the operator product produces a smooth
multilinear map
mz1 ,...,zk : H ∗ (F (r1 )) × · · · × H ∗ (F (rk )) → Hol(Discs(r1 , . . . , rk | ∞), H ∗ (F (∞))).
Here zi indicate the positions of the centers of the discs in Discs(r1 , . . . , rk | ∞).
Consider what happens if we shrink the discs. If ri0 < ri , then
Discs(r1 , . . . , rk | ∞) ⊂ Discs(r10 , . . . , rk0 | ∞).
We also have the extension map F (r0 ) → F (r), from the inclusion D(0, r0 ) ,→
D(0, r). Note that the following diagram commutes:
H ∗ (F (r10 )) ⊗ · · · ⊗ H ∗ (F (rk0 )) / Hol(Discs(r0 , . . . , r0 | ∞), H ∗ (F (∞)))
1 k

H ∗ (F (r1 )) ⊗ · · · ⊗ H ∗ (F (rk )) / Hol(Discs(r1 , . . . , rk | ∞), H ∗ (F (∞))).
Because V maps to limr→0 H ∗ (F (r)), the operator product, when restricted to V,

gives a map
∗ (F (∞))
mzH1 ,...,zk
: V ⊗k → lim Hol(Discs(r, . . . , r | ∞), H ∗ (F (∞))),
r→0
but
lim Hol(Discs(r, . . . , r | ∞), H ∗ (F (∞))) = Hol(Conf k (C), H ∗ (F (∞))),
r→0
where Conf k (C) is the configuration space of k ordered distinct points in C. Note
that if the zi lie in a disc D(0, r), this map provides a map
(r))
mzH∗(F
1 ,...,zk
: V ⊗k → Hol(Conf k (D(0, r)), H ∗ (F (r))).
∗
(F (∞))
Composing the prefactorization product map mzH1 ,...,z k
with the map
Y Y
πn : H ∗ (F (∞)) → V = Vn
n n
gives a map
Y
mz1 ,...,zk : V ⊗k → Hol(Conf k (C), V) = Hol(Conf k (C), Vn ).
n
This map does not involve the spaces F (r) anymore, only the space V and its
natural completion V. If there is potential confusion, we will refer to this version
of the operator product map by mVz1 ,...,zk instead of just mz1 ,...,zk .
The vertex operator will, of course, be constructed from the maps mz1 ,...,zk .
Consider the map
Y
mz,0 : V ⊗ V → Hol(C× , Vn ),
n
where we have restricted the map mVz,w to the locus where w = 0. Since each space
Vn is a discrete vector space (i.e., a colimit of finite dimensional vector spaces), we
can form the ordinary Laurent expansion of an element in Hol(C× , Vn ) to get a map
Lz mVz,0 : V ⊗ V → V[[z, z−1 ]].
It has the following important property.
2.4.1 Lemma. The image of Lz mVz,0 is in the subspace V((z)).
Proof. The map mVz,0 is S 1 -equivariant, where S 1 acts on V and C× in the

evident way. Therefore, so is Lz mVz,0 . Since every element in V ⊗ V is in a finite
sum of the S 1 -eigenspaces, the image of Lz mVz,0 is in the subspace of V[[z, z−1 ]]
spanned by finite sums of eigenvectors. An element of V[[z, z−1 ]] is in the weight
k eigenspace of the S 1 action if it is of the form
X
zk−n vn ,
n
where vn ∈ Vn . Since Vn = 0 for n 0, every such element is in V((z)).
2.4.3. The vertex algebra structure. Let us now define the structures on V that
will correspond to the vertex algebra structure.
((i)) The vacuum element |0i ∈ V: By assumption, F is a unital prefac-
torization algebra. Therefore, the commutative algebra F (∅) has a unit
element |0i. The prefactorization structure map F (∅) → F (D(0, r))
for any r gives an element |0i ∈ F (r). This element is automatically
S 1 -invariant, and therefore in V0 .
((ii)) The translation map T : V → V: The structure of a holomorphi-
∂
cally translation prefactorization algebra on F includes a derivation ∂z
corresponding to infinitesimal translation in the (complex) direction z.
The fact that F has a compatible S 1 action means that, for all r, the
∂
map ∂z maps Fk (r) to Fk−1 (r). Therefore, passing to cohomology, it
∂ ∂
becomes a map ∂z : Vk → Vk−1 . Let T : V → V be the map that is ∂z
on Vk .
((iii)) The state-field map Y : V → End(V)[[z, z−1 ]] : We let
Y(v, z)(v0 ) = Lz mz,0 (v, v0 ) ∈ V((z)).
Note that Y(v, z) is a field in the sense used in the axioms of a vertex
algebra, because Y(v, z)(v0 ) has only finitely many negative powers of
z.
It remains to check the axioms of a vertex algebra. We need to verify:
((i)) Vacuum axiom: Y(|0i , z)(v) = v.
((ii)) Translation axiom:
∂
Y(T v1 , z)(v2 ) = Y(v1 , z)(v2 ).
∂z
and T |0i = 0.
((iii)) Locality axiom:
(z1 − z2 )N [Y(v1 , z1 ), Y(v2 , z2 )] = 0
for N sufficiently large.
The vacuum axiom follows from the fact that the unit |0i ∈ F (∅), viewed as an
element of F (D(0, r)), is a unit for the prefactorization product.
The translation axiom follows immediately from the corresponding axiom of
prefactorization algebras, which is built into our definition of holomorphic transla-
tion invariance: namely,
∂
mz,0 (v1 , v2 ) = mz,0 (∂z v1 , v2 ).
∂z
The fact that T |0i = 0 follows from the fact that the derivation ∂z of F gives a
derivation of the commutative algebra F (∅), which must therefore send the unit to
zero.
It remains to prove the locality axiom. Essentially, it follows from the associa-
tivity property of prefactorization algebras.
2.5. Proof of the locality axiom. Let us observe some useful properties of the
operator product maps
mz1 ,...,zk : V ⊗k → Hol(Conf k (C), V).
Firstly, the map mz1 ,...,zk is S 1 -equivariant, where we use the diagonal S 1 action on
V ⊗k on the left hand side and on the right hand side we use the rotation action of S 1
on Conf k (C) coupled to the S 1 action on V coming from F ’s S 1 -equivariance. The
operator product is also S k -equivariant, where S k acts on V ⊗k and on Conf k (C) in
the evident way. It is also invariant under translation, in the sense that for arbitrary
vi ∈ V,
mz1 +λ,...,zk +λ (v1 , . . . , vk ) = τλ mz1 ,...,zk (v1 , . . . , vk ) ∈ V

where τλ denotes the action of C on V that integrates the infinitesimal translation

action of the Lie algebra C.
We write X
Lz mz,0 (v1 , v2 ) = zk mk (v1 , v2 )
k
where mk (v1 , v2 ) ∈ V. As we have seen, mk (v1 , v2 ) is zero for k 0.
The key proposition in proving the locality axiom is the following bit of com-
plex analysis.
2.5.1 Proposition. Let Ui j ⊂ Conf k (C) be the open subset of configurations for
which z j − zi < z j − zl for l , i, j.
Then, for (z1 , . . . , zk ) ∈ Ui j , we have the following identity in Hol(Ui j , V):
X
mz1 ···zk (v1 , . . . , vk ) = z j ···zk (v1 , . . . , vi−1 , m (vi , v j ), . . . , v
(zi − z j )n mz1 ···b n
bj , . . . , vk ).
In particular, the sum on the right hand side converges. (In this expression, the
hats zbj and vbj indicate that we skip these entries.)
When k = 3, this identity allows us to expand mz1 ,z2 ,z3 in two different ways.
We find that
P
 (z2 − z3 )k mz1 ,z3 (mk (v1 , v2 ), v3 ) if |z2 − z3 | < |z1 − z3 |

mz1 ,z2 ,z3 (v1 , v2 , v3 ) = 

 (z − z )k m (v , mk (v , v )) if |z − z | > |z − z | .
P
z2 ,z3

1 3 2 1 3 2 3 1 3
This formula should be compared to the “associativity” property in the theory of

vertex algebras (see, e.g., Theorem 3.2.1 of Frenkel and Ben-Zvi (2004)).
Proof. By symmetry, we can reduce to the case when i = 1 and j = 2.
Since the operator product is induced from the structure maps of H ∗ F , the
argument will proceed by analyzing the operator product for discs of small radius
and then showing that the relevant property extends to the case where the radii go
to zero.
Fix a configuration (z1 , . . . , zk ) ∈ Conf k (C). There is an ε > 0 such that the
closures of every disc D(zi , ε) are disjoint. As we are interested in the region U12 ,
we can find δ > ε (possibly after shrinking ε) such that D(z1 , ε) ⊂ D(z2 , δ) and
the closure of D(z2 , δ) is disjoint from the closure of D(zm , ε) for m > 2. There
is an open neighborhood U 0 of the configuration (z1 , . . . , zk ) in U12 where these
conditions hold.
For configurations in U 0 , the axioms of a prefactorization algebra tell us that
the following associativity condition holds:
(†) mz1 ,z2 ,...,zk (v1 , . . . , vk ) = mz2 ,z3 ,...,zk (mz1 ,z2 (v1 , v2 ), v3 , . . . , vk ).
Here, we view mz1 ,z2 (v1 , v2 ) as an element of H ∗ (F (D(z2 , δ))) that depends holo-
morphically on points (z1 , z2 ) that lie in the open set where D(z1 , ε) is disjoint from
D(z2 , ε) and contained in D(z2 , δ). By translating z2 to 0, we identify H ∗ (F (D(z2 , δ)))
with H ∗ (F (δ)). This associativity property is an immediate consequence of the ax-
ioms of a prefactorization algebra.
By taking ε and δ sufficiently small, we can cover U12 by sets of the form U 0 .
Hence, locally on U12 , we have the associativity we need. We will continue to
work with our open U 0 described above.
We now assume, without loss of generality, that each vi is homogeneous of

weight |vi | under the S 1 action on V.
Let πk : H ∗ (F (δ)) → Vk be the projection onto the eigenspace Vk . Recall that
the algebra D(S 1 ) of distributions on S 1 acts on F (δ) and so on H ∗ (F (δ)). If this
action is denoted by ∗, and if λ ∈ S 1 ⊂ C× denotes an element of the circle viewed
as a complex number, then
πk ( f ) = λ−k ∗ f.
Note that S 1 -equivariance of the operator product map means that we can write the
Laurent expansion of mz,0 (v1 , v2 ) as
X X
mz,0 (v1 , v2 ) = zk mk (v1 , v2 ) = π|v1 |+|v2 |−k mz,0 (v1 , v2 ).
k k
That is, 1 , v2 ) is the projection of mz,0 (v1 , v2 ) onto the weight |v1 + v2 | − k
zk mk (v
eigenspace of V.
We want to show that
X
(‡) (z1 − z2 )n mz2 ,z3 ,...,zk (mn (v1 , v2 ), . . . , vk ) = mz1 ,...,zk (v1 , . . . , vk )
n
where the zi lie in our open U 0 . This identity is equivalent to showing that
X
mz2 ,z3 ,...,zk (πn mz1 ,z2 (v1 , v2 ), . . . , vk ) = mz1 ,...,zk (v1 , . . . , vk ).
n
Indeed, the Laurent expansion of mz1 ,z2 (v1 , v2 ) and the expansion in terms of eigenspaces
of the S 1 action on V differ only by a reordering of the sum.
Fix v1 , . . . , vk ∈ U 0 . Define a map
Φ : D(S 1 ) → Hol(U 0 , V)
α 7→ mz2 ,z3 ,...,zk (α ∗ mz1 ,z2 (v1 , v2 ), v3 , . . . , vk ).
Here α ∗ − refers to the action of D(S 1 ) on H ∗ (F (δ)). Note that Φ is a smooth map.
Let δ1 denote the delta-function on S 1 with support at the identity 1. The
associativity identity (†) of prefactorization algebras implies that
Φ(δ1 ) = mz1 ,...,zk (v1 , . . . , vk ).
The point here is that δ1 ∗ is the identity on H ∗ (F (δ)).
To prove the identity (‡), it now suffices to prove that
 
X n 
Φ  λ  = Φ(δ1 ).
n∈Z
with its natural topology the sum n∈Z λn converges to δ1 , as it

D(S 1 )
P
In the space
is simply the Fourier expansion of the delta-function. So, to prove the proposition,
it suffices to prove that Φ is continuous, and not just smooth, where the spaces
D(S 1 ) and Hol(U12 δ , V) are endowed with their natural topologies. (The topology
on Hol(Uiδj , V) is given by saying that a sequence converges if its projection to each

Vk converges uniformly, with all derivatives, on compact sets of Uiδj ).)
The spaces D(S 1 ) and Hol(Uiδj , V) both lie in the essential image of the func-
tor from locally-convex topological vector spaces to differentiable vector spaces.
A result of Kriegl and Michor (1997) tells us that smooth linear maps between
topological vector spaces are bounded. Therefore, the map Φ is a bounded linear
map of topological vector spaces.
In Lemma 5.1.1 in Appendix B we show, using results of Kriegl and Michor
(1997), that the space of compactly supported distributions on any manifold has
the bornological property, meaning that a bounded linear map from it to any topo-
logical vector space is the same as a continuous linear map. It follows that Φ is
continuous, thus completing the proof.
As a corollary, we find the following.
2.5.2 Corollary. For v1 , . . . , vk ∈ V, mz1 ,...,zk (v1 , . . . , vk ) has finite order poles on
every diagonal in Conf k (C). That is, for some N sufficiently large, the function
Y
( (zi − z j )N )mz1 ,...,zk (v1 , . . . , vk )
i< j
extends to an element of Hol(Ck , F).

Proof. This corollary follows immediately from the previous proposition.
We are finally ready to prove the locality axiom.
2.5.3 Proposition. The locality axiom holds: for any v1 , v2 ∈ V,
(z1 − z2 )N [Y(v1 , z1 ), Y(v2 , z2 )] = 0
for N 0.
Proof. For any holomorphic function F(z1 , . . . , zk ) of variables z1 , . . . , zk ∈ C× ,
we let Lzi F denote the Laurent expansion of F in the variable zi . This expansion
converges when |zi | < z j for all j. It can be defined by fixing the values of z j with
j , i, then viewing F as a function of zi on the punctured disc where 0 < |zi | <
min j,i z j , and taking the usual Laurent expansion. We can also define iterated
Laurent expansions. For example, if F is a function of z1 , z2 ∈ C× , we define
1 , z2 ]]
Lz2 Lz1 F ∈ C[[z±1 ±1
by first taking the Laurent expansion with respect to z1 , yielding a series in z1 whose
coefficients are holomorphic functions of z2 ∈ C× , and then applying the Laurent
expansion with respect to z2 to each of the coefficient functions of the expansion
with respect to z1 .
Recall that we define
Y(v1 , z)(v2 ) = Lz mz,0 (v1 , v2 ) ∈ V((z)).
We define mk (v1 , v2 ) so that
X
Lz mz,0 (v1 , v2 ) = zk mk (v1 , v2 ).
k
Note that, by definition,

Y(v1 , z1 )Y(v2 , z2 )(v3 ) = Lz1 mz1 ,0 v1 , Lz2 mz2 ,0 (v2 , v3 )

X
= Lz 1 zk2 mz1 ,0 (v1 , mn (v2 , v3 )) ∈ V[[z±1
1 , v2 ]].
±1
k
Proposition 2.5.1 tells us that
X
mz1 ,z2 ,0 (v1 , v2 , v3 ) = zn2 mz1 ,0 v1 , mn (v2 , v3 )

n
as long as |z2 | < |z1 |. Thus,
Y(v1 , z1 )Y(v2 , z2 )(v3 ) = Lz1 Lz2 mz1 ,z2 ,0 (v1 , v2 , v3 ).
Similarly,
Y(v2 , z2 )Y(v1 , z1 )(v3 ) = Lz2 Lz1 mz2 ,z2 ,0 (v2 , v1 , v3 )
= Lz2 Lz1 mz1 ,z2 ,0 (v1 , v2 , v3 ).
Therefore,
[Y(v1 , z1 ), Y(v2 , z2 )](v3 ) = Lz1 Lz2 − Lz2 Lz1 mz1 ,z2 ,0 (v1 , v2 , v3 ).

Since Lz1 and Lz2 are maps of C[z1 , z2 ] modules, we have
(z1 − z2 )N Lz1 Lz2 − Lz2 Lz1 mz1 ,z2 ,0 (v1 , v2 , v3 )

= Lz1 Lz2 − Lz2 Lz1 (z1 − z2 )N mz1 ,z2 ,0 (v1 , v2 , v3 )

Finally, we know that for N sufficiently large, (z1 − z2 )N z1N z2N mz1 ,z2 ,0 has no poles,
and so extends to a function on C2 . It follows from the fact that partial derivatives
commute that
Lz1 Lz2 − Lz2 Lz1 (z1 − z2 )N z1N z2N mz1 ,z2 ,0 (v1 , v2 , v3 ) = 0.

Since Laurent expansion is a map of C[z1 , z2 ]-modules, and since z1 , z2 act invert-
1 , z2 ]], the result follows.
ibly on C[[z±1 ±1
Remark: In the vertex algebra literature, a heuristic justification of the locality ax-
iom is frequently given by unpacking the consequences of pretending that Y(v1 , z1 )Y(v2 , z2 )(v3 )
and Y(v2 , z2 )Y(v1 , z1 )(v3 ) arise as expansions of a holomorphic function of z1 , z2 ∈
C× in the regions when |z1 | < |z2 | and |z2 | < |z1 |. Our approach makes this idea
rigorous. ^
3. The βγ system and vertex algebras

This section focuses on one of the simplest holomorphic field theories, the free
βγ system. Our goal is to study it just as we studied the free particle in Section
2. Following the methods developed there, we will construct the prefactorization
algebra for this theory, show that it is holomorphically translation-invariant, and fi-
nally show that the associated vertex algebra is what is known in the vertex algebra
literature as the βγ system. Along the way, we will compute the simplest operator
product expansions for the theory using purely homological methods.
3. THE βγ SYSTEM AND VERTEX ALGEBRAS 141

3.1. The βγ system. Let M = C and let E = Ω0,∗ M ⊕ Ω M , ∂ be the Dolbeault
1,∗
complex resolving holomorphic functions and holomorphic 1-forms as a sheaf on

M. Following the convention of physicists, we denote by γ an element of Ω0,∗ and
by β an element of Ω1,∗ . The pairing h−, −i is
h−, −i : Ec ⊗ Ec → R C,
(γ0 + β0 ) ⊗ (γ1 + β1 ) 7→ C γ0 ∧ β1 + β0 ∧ γ1 .
Thus we have the data of a free BV theory. The action functional for the theory is
Z
S (γ, β) = hγ + β, ∂(γ + β)i = 2 β ∧ ∂γ
M
The Euler-Lagrange equation is simply ∂γ = 0 = ∂β. One should think of E as the

“derived space of holomorphic functions and 1-forms on M.”
Remark: Note that this theory is well-defined on any Riemann surface, and one can
study how it varies over the moduli space of curves. In fact, there are many variants
of this theory. Let Σ be a Riemann surface and V a holomorphic vector bundle on
Σ. Define a free BV theory on Σ with fields E = Ω0,∗ (V) ⊕ Ω1,∗ (V∨ ) and with
pairing given by “fiberwise evaluate duals and then integrate.”
For instance, if one adds d copies of E from above (equivalently, tensor E
with Cd ) and lets S d be the d-fold sum of the action S on each copy, then the
Euler-Lagrange equations for S d picks out “holomorphic maps γ from M to Cd
and holomorphic sections β of Ω1M (γ∗ T Cd ).” ^
3.2. The quantum observables of the βγ system. To construct the quantum
observables, following Section 2, we start by defining a certain graded Heisenberg
Lie algebra and then take its Chevalley-Eilenberg complex for Lie algebra homol-
ogy.
For each open U ⊂ C, we set
H(U) = Ω0,∗
c (U) ⊕ Ωc (U) ⊕ (C~)[−1],
1,∗
where C~ is situated in cohomological degree 1. The Lie bracket is simply

Z
[µ, ν] = ~ µ ∧ ν,
U
so H is a central extension of the abelian dg Lie algebra given by all the Dolbeault
forms (with ∂ as differential).
The prefactorization algebra Obsq of quantum observables assigns to each open
U ⊂ C, the cochain complex C∗ (H(U)), which we will write as

Obsq (U) = Sym Ω1,∗ c (U)[1] ⊕ Ωc (U)[1] [~], ∂ + ~∆ .
0,∗
The differential has a component ∂ arising from the underlying cochain complex
of H and a component arising from the Lie bracket, which we’ll denote ∆. It is the
BV Laplacian for this theory.
Below we will unpack the information Obsq that encodes by examining some
simple open sets and the cohomology H ∗ Obsq on those open sets. As usual, the
meaning of a complex is easiest to garner through its cohomology. The main theo-
rem of this section is that the vertex algebra extracted from Obsq is the well-known
βγ vertex algebra.
3.3. An isomorphism of vertex algebras. Our goal is to demonstrate that the

vertex algebra constructed by Theorem 2.2.1 is from the quantum observables of
the βγ-system is isomorphic to a vertex algebra considered in the physics literature
called the βγ vertex algebra.
3.3.1. The βγ vertex algebra. We follow Frenkel and Ben-Zvi (2004), notably
chapters 11 and 12, to make the relationship clear. Let W denote the space of
polynomials C[an , a∗m ] generated by variables an , a∗m n < 0 and m ≤ 0.
3.3.1 Definition. The βγ vertex algebra has state space W, vacuum vector 1, trans-
lation operator T the map
ai 7→ −iai−1 ,
a∗i 7→ −(i − 1)a∗i−1 ,
and the vertex operator

X X ∂
Y(a−1 , z) = an z−1−n + ∗ z
−1−n
n<0 n≥0
∂a −n
and
X X ∂
Y(a∗0 , z) = a∗n z−n − z−n .
n≤0 n>0
∂a−n
By the reconstruction theorem 2.3.1, these determine a vertex algebra.
The main theorem of this section is the following.

3.3.2 Theorem. Let V~=2πi denote the vertex algebra constructed from quantum
observables of the βγ system, specialized to ~ = 2πi. There is an S 1 -equivariant
isomorphism of vertex algebras
V~=2πi W.
The circle S 1 acts on W by giving ai , a∗j weights i, j respectively.

The proof of this theorem breaks up into a few stages. First, we need to verify
that we can apply Theorem 2.2.1. In particular, we need to verify that Obsq is
holomorphically translation invariant and S 1 -equivariant. We then need to compute
the weight spaces for the S 1 action. Finally, we need to demonstrate that the vertex
operator arising from Obsq agrees with that of the βγ vertex algebra.
These arguments will be divided over several subsections. The techniques we
develop here make it possible to prove theorems of this flavor for many other field
theories on C whose action functional is holomorphic in flavor.
3.4. The criteria to obtain a vertex algebra. The βγ field theory is mani-
∂
festly translation-invariant, so it remains to verify that the action of ∂z is homotopi-
cally trivial. Consider the operator
d
η= ,
d(dz)
which acts on the space of fields E . The operator η maps Ω1,1 to Ω1,0 and Ω0,1 to
Ω0,0 . Then we see that
d
[∂ + ~∆, η] = ,
dz
so that the action of d/dz is homotopically trivial, as desired.
We would like to apply the result of Theorem 2.2.1, which shows that a holo-
morphically translation invariant prefactorization algebra with certain additional
conditions gives rise to a vertex algebra. We need to check the following condi-
tions.
((i)) The prefactorization algebra must have an action of S 1 covering the
action on C by rotation.
((ii)) For every disc D(0, r) ⊂ C, including r = ∞, the S 1 action on Obsq (D(0, r))
must extend to an action of the algebra D(S 1 ) of distributions on S 1 .
q
((iii)) If Obsk (D(0, r)) denotes the weight k eigenspace of the S 1 action, then
we require that the extension map
q q
H ∗ (Obsk (D(0, r))) → H ∗ (Obsk (D(0, s)))
is an isomorphism for r < s.
q
((iv)) Finally, we require that the space H ∗ (Obsk (D(0, r))) is a discrete vector
space, that is, a colimit of finite dimensional vector spaces.
The first condition is obvious in our example: the S 1 action arises from the natural
action of S 1 on Ω0,∗
c (C) and Ωc (C). The second condition is also easy to check: if
1,∗
f ∈ Cc∞ (D(0, r)) then the expression

Z
z 7→ φ(λ) f (λz)
λ∈S 1
makes sense for any distribution φ ∈ D(S 1 ), and it defines a continuous and hence
smooth map
ρ : D(S 1 ) × Cc∞ (D(0, r)) → Cc∞ (D(0, r)).
To check the remaining two conditions, we need to analyze H ∗ (Obsq (D(0, r))) more
explicitly.
3.5. Analytic preliminaries. In Section 6 we showed that if a free field theory
possesses a Green’s function for the differential defining the elliptic complex of
fields, then there is an isomorphism of differentiable cochain complexes
Obscl (U)[~] Obsq (U)
for any open set U. In our example, we want to understand H ∗ (Obsq (D(0, r))) as
a differentiable cochain complex and also its decomposition into eigenspaces for
the action of S 1 . This result about the Green’s function shows that it suffices to
understand the cohomology of the corresponding complex of classical observables

(see equation (3.6.1) for the Green’s function). We will describe the classical ob-
servables in this subsection.
Note that the classical observables are particularly simple to analyze because
computing their cohomology breaks up into a collection of easier problems. To be
explicit, we recall that

Obscl (U) = (Sym Ω1,∗
c (U)[1] ⊕ Ωc (U)[1] , ∂),
0,∗
and ∂ preserves the symmetric powers. Hence, we can analyze each symmetric
power
c (U)[1] ⊕ Ωc (U)[1] , ∂)
(Symn Ω1,∗ 0,∗
separately. The nth symmetric power can then be studied as an elliptic complex on
the complex manifold U n .
3.5.1. Recollections. We remind the reader of some facts from the theory of
several complex variables (references for this material are Gunning and Rossi
(1965), Forster (1991), and Serre (1953)). We then use these facts to describe
the cohomology of the observables.
3.5.1 Proposition. Every open set U ⊂ C is Stein Forster (1991). As the product
of Stein manifolds is Stein, every product U n ⊂ Cn is Stein.
Remark: Behnke and Stein (1949) proved that every noncompact Riemann surface
is Stein, so the arguments we develop here extend farther than we exploit them. ^
We need a particular instance of Cartan’s theorem B about coherent analytic
sheaves. See Gunning and Rossi (1965).
3.5.2 Theorem (Cartan’s Theorem B). Let X be a Stein manifold and let E be a
holomorphic vector bundle on X. Then,
(
0, k,0
H (Ω (X, E), ∂) =
k 0,∗
Hol(X, E), k = 0,
where Hol(X, E) denotes the holomorphic sections of E on X.
We now use a corollary noted by Serre (1953); it is a special case of the Serre
duality theorem. (Nowadays, people normally talk about the Serre duality theorem
for compact complex manifolds, but in Serre’s original paper he proved it for non-
compact manifolds too, under some additional hypothesis that will be satisfied on
Stein manifolds).
Note that we use the Fréchet topology on Hol(X, E), obtained as a closed sub-
space of C ∞ (X, E). We let E ! be the holomorphic vector bundle E ∨ ⊗ KX where KX
is the canonical bundle of X.
3.5.3 Corollary. For X a Stein manifold of complex dimension n, the compactly-
supported Dolbeault cohomology is
(
0, k,n
H k (Ω0,∗ (X, E), ∂) =
c (Hol(X, E ! ))∨ , k = n,
where (Hol(X, E ! ))∨ denotes the continuous linear dual to Hol(X, E ! ).

Proof. The Atiyah-Bott lemma (see Lemma D) shows that the inclusion
0,∗
c (X, E), ∂) ,→ (Ωc (X, E), ∂)
(Ω0,∗
is a chain homotopy equivalence. (Recall that the bar denotes “distributional sec-
0,k
tions.”) As Ωc (X, E) is the continuous linear dual of Ω0,n−k (X, E ! ), it suffices to
prove the desired result for the continuous linear dual complex.
Consider the acyclic complex
i ∂
0 → Hol(X, E) ,→ C ∞ (X, E) → Ω0,1 (X, E) → · · · → Ω0,n (X) → 0.
Our aim is to show that the linear dual of this complex is also acyclic. Note that this
is a complex of Fréchet spaces. The result is then a consequence of the following
lemma.
3.5.4 Lemma. If V ∗ is an acyclic cochain complex of Fréchet spaces, then the dual
complex (V ∗ )∨ is also acyclic.
Proof. Let di : V i → V i+1 denote the differential. We need to show that the
sequence
(V i+1 )∨ → (V i )∨ → (V i−1 )∨
is exact in the middle. That is, we need to show that if α : V i → C is a continuous
linear map, and if α ◦ di−1 = 0, then there exists some β : V i+1 → C such that
α = β ◦ di .
Note that α is zero on Im di−1 = Ker di , so that α descends to a linear map
α : V i / Im di−1 → C.
Since the complex is acyclic, Im di−1 = Ker di as vector spaces. However, it is not
automatically true that they are the same as topological vector spaces, where we
view Im di−1 as a quotient of V i and Ker di as a subspace of V i . Here is where we
use the Fréchet hypothesis: the open mapping theorem holds for Fréchet spaces,
and it tells us that any surjective map between Fréchet spaces is open. Since Ker di
is a closed subspace of V i , it is a Fréchet space. The map di−1 : V i−1 → Ker di is
surjective and therefore open. It follows that Im di−1 = Ker di as topological vector
spaces.
From this, we see that our α : V i → C descends to a continuous linear func-
tional on Ker di+1 . As it is a closed subspace of V i+1 , the Hahn-Banach theorem
tells us that it extends to a continuous linear functional on V i+1 .
3.5.2. The setting of differentiable vector spaces. These lemmas allow us to
understand the cohomology of classical observables just as a vector space. Since
we treat classical observables as a differentiable vector space, however, we are
really interested in its cohomology as a sheaf on the site of smooth manifolds. It
turns out (perhaps surprisingly) that for X a Stein manifold of dimension n, the
isomorphism
c (X, E)) = Hol(X, E )
! ∨
H n (Ω0,∗
from above is not an isomorphism of sheaves on the site of smooth manifolds.

(Here we define a smooth map from a manifold M to Hol(X, E ! )∨ to be a contin-
uous linear map Hol(X, E) → C ∞ (M), as discussed in Appendix B.) Thankfully,
there is a different and very useful description of compactly-supported Dolbeault
cohomology as a differentiable vector space.
We will present this description for polydiscs, although it works more gener-
ally. If 0 < r1 , . . . , rn ≤ ∞, let Dr ⊂ C be the disc of radius r, and let
Dr1 ,...,rn = Dr1 × · · · × Drk ⊂ Cn
be the corresponding polydisc. We can view Dr as an open subset in P1 .

In general, if X is a complex manifold and C ⊂ X is a closed subset, we define
Hol(C) = colim Hol(U)

C⊂U
to be the germs of holomorphic functions on C. (The colimit is over opens U

containing C.) The space Hol(C) has a natural structure of differentiable vector
space, where we view Hol(U) as a differentiable vector space and take the colimit
in the category of differentiable vector spaces. The same definition holds for the
space Hol(C, E) of germs on C of holomorphic sections of a holomorphic vector
bundle on E.
We have the following theorem, describing compactly supported Dolbeault co-
homology as a differentiable vector space. Let O(−1) denote the holomorphic line
bundle on P1 consisting of functions vanishing at ∞ ∈ P1 . Let O(−1)n denote the
line bundle on (P1 )n consisting of functions that vanish at infinity in each variable.
3.5.5 Theorem. For a polydisc Dr1 ,...,rn ⊂ Cn , we have a natural isomorphism of
differentiable vector spaces
H n (Ωc0,∗ (Dr1 ,...,rn )) Hol((P1 \ Dr1 ) × · · · × (P1 × Drn ), O(−1)n ).
Further, all other cohomology groups of Ω0,∗

c (Dr1 ,...,rn ) are zero as differentiable
vector spaces.
This isomorphism is invariant under holomorphic symmetries of the polydisc
Dr1 ,...,rk , under the actions of S 1 by rotation in each coordinate, and under the action
of the symmetric group (when the ri are all the same).
Before we prove this theorem, we need a technical result.
3.5.6 Proposition. For any complex manifold X and holomorphic vector bundle
E on X and for any manifold M, the cochain complex C ∞ (M, Ω0,∗ (X, E)) is a fine
resolution of the sheaf on M × X consisting of smooth sections of the bundle π∗X E
that are holomorphic along X.
Further, if we assume that H i (Ω0,∗ (X, E)) = 0 for i > 0, then

C ∞ (M, Hol(X, E)) if i = 0,

H i (C ∞ (M, Ω0,∗ (X, E))) = 

0
 if i > 0.
Proof. The first statement follows from the second statement. Locally on X,
the Dolbeault-Grothendieck lemma tells us that the sheaf Ω0,∗ (X, E) has no higher
cohomology, and the sheaves C ∞ (M, Ω0,i (X, E)) are certainly fine.
To prove the second statement, consider the exact sequence
0 → Hol(X, E) → Ω0,0 (X, E) → Ω0,1 (X, E) · · · → Ω0,n (X, E) → 0.
It is an exact sequence of Fréchet spaces. Proposition 3 of section I.1.2 in Grothendieck
(1955) tells us that the completed projective tensor product of nuclear Fréchet
spaces is an exact functor, that is, takes exact sequences to exact sequences. An-
other result of Grothendieck (1952) tells us that for any complete locally con-
vex topological vector space F, C ∞ (M, F) is naturally isomorphic to C ∞ (M)b ⊗π F,
where b ⊗π denotes the completed projective tensor product. The result then fol-
lows.
Proof of the theorem. We need to produce an isomorphism of differentiable
vector spaces, i.e., an isomorphism of sheaves. In other words, we need to find an
isomorphism between their sections on each manifold M,
c (Dr1 ,...,rn ))) C (M, Hol((P \ Dr1 ) × . . . (P × Drn ), O(−1) )),
C ∞ (M, H n (Ω0,∗ ∞ 1 1 n
and this isomorphism must be natural in M. We also need to show that the sections
c (Dr1 ,...,rn ))) of the other cohomology sheaves, with i < n, are zero.
C ∞ (M, H i (Ω0,∗
The first thing we prove is that, for any complex manifold X and holomorphic
vector bundle E on X, and any open subset U ⊂ X, there is an exact sequence
(‡) 0 → C ∞ (M, Ω0,∗
c (U, E)) → C (M, Ω (X, E)) → C (M, Ω (X \ U, E)) → 0
∞ 0,∗ ∞ 0,∗
of the sections on any smooth manifold M.

We will do this by working with these differentiable cochain complexes Ω0,∗ (X, E)
simply as sheaves on M, not on the whole site of smooth manifolds. To start, we
will prove that we have an exact sequence of sheaves on M. As these complexes
are built out of topological vector spaces, they are fine sheaves on M and hence
their global sections will provide the exact sequence (‡).
Pick an exhausting family {Ki } of compact subsets of U. Define
res

Ω0,∗
Ki (U, E) = Ker Ω0,∗
(U, E) −
−→ Ω 0,∗
(U \ Ki , E) .
This vector space coincides with Ω0,∗

Ki (X, E), since a section that vanishes outside
of Ki in X must vanish outside of U. The sequence
0 → Ω0,∗
Ki (X, E) → Ω (X, E) → Ω (X \ Ki , E)
0,∗ 0,∗
is therefore exact, but it is not necessarily exact on the right.

Consider the colimit of this sequence as i goes to ∞ in the category of sheaves
on M. We then obtain an exact sequence
0 → Ω0,∗
c (U, E) → Ω (X, E) → Ω (X \ U, E)
0,∗ 0,∗
as sheaves on M. Now, we claim that this sequence is exact on the right, as sheaves.
The point is that, locally on M, every smooth function on M × (X \ U) extends to
a smooth function on some M × (X \ Ki ) and by applying a bump function that is 1
on a neighbourhood of M × (X \ U) and zero on the interior of Ki , we can extend it

to a smooth function on M × X.
Let’s apply this exact sequence to the case when X = (P1 )n , U = Dr1 ,...,rn and
E = O(−1)n . Note that in this case E is trivialized on U, so that we find, on taking
cohomology of the Dolbeault complexes, a long exact sequence of sheaves on M:
c (Dr1 ,...,rn )) → H (Ω ((P ) , (O(−1)) ))
n
· · · → H i (Ω0,∗ i 0,∗ 1 n
→ H i (Ω0,∗ ((P1 )n \ Dr1 ,...,rn , O(−1)n )) → · · · .

Note that the Dolbeault cohomology of (P1 )n with coefficients in O(−1)n van-
ishes. Hence, the middle term in this exact sequence vanishes by the preceding
proposition, and so we get an isomorphism
C ∞ (M, H i (Ω0,∗ ∞
c (Dr1 ,...,rn ))) C (M, H
i−1
(Ω0,∗ ((P1 )n \ Dr1 ,...,rn , O(−1)n ))).
For the purposes of the theorem, we now need to compute the right hand side. The
complex appearing on the right hand side is the colimit, as ε → 0, of
C ∞ (M, Ω0,∗ ((P1 )n \ Dr1 −ε,...,rn −ε , O(−1)n )).
Here D indicates the closed disc.
We now have a sheaf on M × (P1 )n \ Dr1 −ε,...,rn −ε that sends an open set U × V
to C ∞ (U, Ω0,∗ (V, O(−1)n )). It is a cochain complex of fine sheaves. We are in-
terested in the cohomology of global sections. We can compute this sheaf co-
homology using the local-to-global spectral sequence associated to a Čech cover
of (P1 )n \ Dr1 −ε,...,rn −ε . Consider the cover {Ui }1≤i≤n given by
Ui = P1 × · · · × (P1 \ Dri −ε ) × · · · × P1 .

We take the corresponding cover of M × (P1 )n \ Dr1 −ε,...,rn −ε given by the opens
{M × Ui }1≤i≤n .
Note that, for k < n, we have
H ∗ (Ω0,∗ (Ui1 ,...,ik , O(−1)n )) = 0,
where Ui1 ,...,ik denotes the intersection of all the Ui j . The previous proposition
implies that we also have
H ∗ (C ∞ (M, Ω0,∗ (Ui1 ,...,ik , O(−1)n ))) = 0.
The local-to-global spectral sequence then tells us that we have a natural isomor-
phism identifying
H ∗ (C ∞ (M, Ω0,∗ ((P1 )n \ Dr1 −ε,...,rn −ε , O(−1)n )))
with
C ∞ (M, Hol(U1,...,n , O(−1)n ))[−n].
That is, all cohomology groups on the left hand side of this equation are zero,
except for the top cohomology, which is equal to the vector space on the right.
Note that
U1,...,n = (P1 \ Dr1 −ε ) × · · · × (P1 \ Drn −ε ).
Taking the colimit as ε → 0, combined with our previous calculations, gives the
desired result. Note that this colimit must be taken in the category of sheaves on
M. We are also using the fact that sequential colimits commute with formation of
cohomology.
3.6. A description of observables on a disc. This analytic discussion allows

us to understand both classical and quantum observables of the βγ system. Let us
first define some basic classical observables.
3.6.1 Definition. On any disc D(x, r) centered at the point x, let cn (x) denote the
linear classical observable
1
cn (x) : γ ∈ Ω0,0 (D(x, r)) 7→ (∂nz γ)(x).
n!
Likewise, for n > 0, let bn (x) denote the linear functional
1
bn (x) : β dz ∈ Ω1,0 (D(x, r)) 7→ (∂n−1 β)(x).
(n − 1)! z
These observables descend to elements of H 0 (Obscl (D(x, r))).
Let us introduce some notation to deal with products of these. For a multi-
index K = (k1 , . . . , kn ) or L = (l1 , . . . , lm ), we let
bK (x) = bk1 (x) · · · bkn (x),
cL (x) = cl1 (x) · · · clm (x).
Of course these products make sense only if ki > 0 for all i. Note that under the
natural S 1 action on H 0 (Obscl (D(x, r))), the elements bK (x) and cL (x) are of weight
− |K| and − |L|, where |K| = i ki and similarly for L.
P
3.6.2 Lemma. These observables generate the cohomology of the classical ob-
servables in the following sense.
((i)) For i , 0 the cohomology groups H i (Obscl (D(x, r))) vanish as differ-
entiable vector spaces.
((ii)) The monomials {bK (x)cL (x)}|K|+|L|=n form a basis for the weight n space
H 0 (Obscl (D(x, r)))n . Further, they form a basis in the sense of differ-
entiable vector spaces, meaning that any smooth section
f : M → Hn0 (Obscl (D(x, r)))
can be expressed uniquely as a sum
X
f = fKL bK (x)cL (x),
|K|+|L|=n
where fKL ∈ C ∞ (M) and locally on M all but finitely many of the fKL
are zero.
((iii)) More generally, any smooth section
f : M → H 0 (Obscl (D(x, r)))
can be expressed uniquely as a sum

X
f = fKL bk (x)cL (x)
that is convergent in a natural topology on C ∞ (M, H 0 (Obscl (D(x, r)))),

with the coefficient functions fKL satisfying the following properties:
((a)) Locally on M, fk1 ,...,kn ,l1 ,...,lm = 0 for n + m 0.
((b)) For fixed n and m, the sum
X
−kn −l1 −1
fk1 ,...,kn ,l1 ,...,lm z−k 1
1 · · · zn w1 · · · w−l
1
m −1
K,L
is absolutely convergent when |zi | ≥ r and w j ≥ r, in the natural

topology on C ∞ (M).
This lemma gives an explicit description of C ∞ (M, H 0 (Obscl (D(x, r)))).
Proof. We will set x = 0. By definition, Obscl (D(0, r)) is a direct sum, as

differentiable cochain complexes,
M
Obscl (U) = Ω0,∗
c (D(0, r) n
, E n
)[n] ,
Sn
n
where E is the holomorphic vector bundle O ⊕ K on C.

It follows from Theorem 3.5.5 that, as differentiable vector spaces,
H i (Ω0,∗ (D(0, r)n , E n ) = 0
unless i = n, and that
H n (Ω0,∗ (D(0, r)n , E n ) = Hol((P1 \ D(0, r)), (O(−1) ⊕ O(−1)dz)n ).
Theorem 3.5.5 was stated only for the trivial bundle, but the bundle E is trivial on
D(x, r). It is not, however, S 1 -equivariantly trivial. The notation O(−1)dz indicates
how we have changed the S 1 -action on O(−1).
It follows that we have an isomorphism, of S 1 -equivariant differentiable vector
spaces
M
H 0 (Obscl (D(0, r))) H 0 ((P1 \ D(0, r))n , (O(−1) ⊕ O(−1)dz)n ))S n .
n
Under this isomorphism, the observable bK (0)cL (0) goes to the function
1
z−k1 dz1 · · · z−kn −l1 −1
n dzn w1 · · · w−l
m
m −1
.
(2πi)n+m 1
Everything in the statement is now immediate. On H 0 (Obscl (D(0, r))) we use the
topology that is the colimit of the topologies on holomorphic functions on (P1 \
D(0, r − ε))n as ε → 0.
3.6.3 Lemma. We have
H ∗ (Obsq (U)) = H ∗ (Obscl (U))[~]
as S 1 -equivariant differentiable vector spaces. This isomorphism is compatible

with maps induced from inclusions U ,→ V of open subsets of C. It is not compat-
ible with the prefactorization product map.
Proof. This lemma follows, as explained in Section 6, from the existence of a
Green’s function for the ∂ operator, namely
dz1 − dz2
(3.6.1) G(z1 , z2 ) = ∈ E (C) ⊗ E (C)
z1 − z2
where E denotes the complex of fields of the βγ system.
We will use the notation bK (x)cL (x) for the quantum observables on D(x, r)
that arise from the classical observables discussed above, using the isomorphism
given by this lemma.
3.6.4 Corollary. The properties listed in Lemma 3.6.2 also hold for quantum ob-
servables. As a result, all the conditions of Theorem 2.2.1 are satisfied, so that the
structure of prefactorization algebra leads to a C[~]-linear vertex algebra structure
on the space M
V= H ∗ (Obsq (D(0, r)))n
n
where H ∗ (Obsq (D(0, r)))n indicates the weight n eigenspace of the S 1 action.
Proof. The only conditions we have not yet checked are
((i)) the inclusion maps
H ∗ (Obsq (D(0, r)))n → H ∗ (Obsq (D(0, s)))n
for r < s are quasi-isomorphisms, and
((ii)) the differentiable vector spaces Vn = H ∗ (Obsq (D(0, r)))n are countable
colimits of finite-dimensional vector spaces in the category of differen-
tiable vector spaces.
Both of these conditions follow immediately from the analog of lemma 3.6.2 that
applies to quantum observables.
3.7. The proof of Theorem 3.3.2. We now finish the proof of the main theo-
rem. We have shown that we obtain some vertex algebra, because we have verified
the criteria to apply Theorem 2.2.1. What remains is to exhibit an explicit isomor-
phism with the βγ vertex algebra.
Proof. Note that V~=2πi is the polynomial algebra on the generators bn , cm
where n ≥ 1 and m ≥ 0. Also bn , cm have weights −n, −m respectively, under
the S 1 -action. We define an isomorphism V~=2πi to W by sending bn to a−n and cm
to a∗−m , and extending it to be an isomorphism of commutative algebras. By the
reconstruction theorem, it suffices to calculate Y(b1 , z) and Y(c0 , z).
By the way we defined the vertex algebra associated to the prefactorization
algebra of quantum observables in Theorem 2.2.1, we have
Y(b1 , z)(α) = Lz mz,0 (b1 , α) ∈ V~=2πi ((z)),
where Lz is denotes Laurent expansion, and

mz,0 : V~=2πi ⊗ V~=2πi → V ~=2πi
is the map associated to the prefactorization product coming from the inclusion
of the disjoint discs D(z, r) and D(0, r) into D(0, ∞) (where r can be taken to be
arbitrarily small).
We are using the Green’s function for the ∂ operator to identify classical and
quantum observables. Let us recall how the Green’s function leads to an explicit
formula for the prefactorization product.
Let E denote the sheaf on C of fields of our theory, so that
E (U) = Ω0,∗ (U, O ⊕ K).
It is the sheaf of smooth sections of a graded bundle E on C. Let E ! = E [1] denote
the sheaf of smooth sections of the vector bundle E ∨ ⊗ ∧2 T ∗ , so that a section is
a fiberwise linear functional on E valued in 2-forms. Let E denote the sheaf of
distributional sections and Ec denote compactly-supported sections.
The propagator, or Green’s function, is
dz1 ⊗ 1 − 1 ⊗ dz2
P= .
2πi(z1 − z2 )
It is an element of
E (C)b
⊗π E (C) = D(C2 , E E),
where D denotes the space of distributional sections.
We can also view at as a symmetric and smooth linear map
(†) P : Ec! (C)b
⊗β Ec! (C) = CC∞ (C2 , (E ! )2 ) → C.
⊗β denotes the completed bornological tensor product on the category of
Here b
convenient vector spaces.
Recall that we identify
Obscl (U) = Sym(Ec! (U)) = Sym(Ec (U)[1]),
where the symmetric algebra is defined using the completed tensor product on the
category of convenient vector spaces.
From P we construct an order two differential operator on this symmetric al-
gebra:
∂P : Obscl (U) → Obscl (U).
This operator is characterized by the fact that it is a smooth (or, equivalently, con-
tinuous) order two differential operator, vanishes on Sym≤1 , and on Sym2 is deter-
mined by the map in (†).
For example, if x, y ∈ U, then
1 ∂i−1 ∂ j 1
∂P (bi (x)c j (y)) = .
(i − 1)! j! ∂i−1 x ∂ j y 2πi(x − y)
To compute this, note that bi (x)c j (y) is quadratic and we can pick a representative
for this observable in Ec! (C)b
⊗β Ec! (C). We then apply P to obtain the number on the
right hand side. Explicitly, bi (x) and c j (y) are derivatives of delta-functions and we
apply them to the Green’s function.
We can identify Obsq (U) as a graded vector space with Obscl (U)[~]. The
quantum differential is d = d1 + ~d2 , a sum of two terms, where d1 is the dif-
ferential on Obscl (U) and d2 is the differential arising from the Lie bracket in the
shifted Heisenberg Lie algebra whose Chevalley-Eilenberg chain complex defines
Obsq (U).
It is easy to check that
[~∂P , d1 ] = d2 .
This identity follows immediately from the fact that (1, 1)-current ∂P on C2 is the
delta-current on the diagonal.
As we explained in Section 6, we get an isomorphism of cochain complexes
W : Obscl (U)[~] 7→ Obsq (U)
.
α 7→ e~∂P α
Further, for U1 , U2 disjoint opens in V, the prefactorization product map
?~ : Obscl (U1 )[~] × Obscl (U2 )[~] → Obscl (V)[~],
arising from that on Obsq under the identification W, is given by the formula

α ?~ β = e−~∂P e~∂P α · e~∂P β .
Here · refers to the commutative product on classical observables.
Let us apply this formula to α = b1 (z) and β in the algebra generated by
bi (0) and c j (0). First, note that since b1 (z) is linear, ~∂P b1 (z) = 0. Note also
that [∂P , b1 (z)] commutes with ∂P . Thus, we find that

b1 (z) ?~ β = e−~∂P b1 (z)e~∂P β
= b1 (z)β − [~∂P , b1 (z)]β.
Note that [∂P , b1 (z)] is an order one operator, and so a derivation. So it suffices to
calculate what it does on generators. We find that
1 1 ∂j 1
!
[∂P , b1 (z)]c j (0) =
2πi j! ∂ j w z − w w=0
1 − j−1
= z ,
2πi
and
[∂P , b1 (z)]b j (0) = 0.
In other words,
∞
1 X − j−1 ∂
[∂P , b1 (z)] = z .
2πi j=0 ∂c j (0)
Note as well that for |z| < r, we can expand the cohomology class b1 (z) in H 0 (Obscl (D(0, r))
as a sum
∞
X
b1 (z) = bn+1 (0)zn .
n=0
Indeed, for a classical field γ ∈ Ω1hol (D(0, r)) satisfying the equations of motion,
we have
b1 (z)(γ) = γ(z)
∞
X 1
= zn γ(n) (0)
n=0
n!
∞
X
= zn bn1 (0)(γ).
n=0
Putting all these computations together, we find that, for β in the algebra generated
by c j (0), bi (0), we have
∞ ∞

X ~ X ∂ 
b1 (z) ?~ β =  bn+1 (0)zn + z−m−1  β ∈ H 0 (Obscl (D(0, r))[~].
n=0
2πi m=0 ∂cm (0)
Thus, if we set ~ = 2πi, we see that the operator product on the space V~=2πi
matches the one on W if we sent bn (0) to a−n and cn (0) to a∗−n . A similar calculation
of the operator product with c0 (z) completes the proof.
4. Kac-Moody algebras and factorization envelopes

In this section, we will construct a holomorphically translation invariant pref-
actorization algebra whose associated vertex algebra is the affine Kac-Moody ver-
tex algebra. This construction is an example of the twisted prefactorization enve-
lope construction, which also produces the prefactorization algebras for free field
theories (see Section 6). As we explained there, it is our version of the chiral enve-
lope construction from Beilinson and Drinfeld (2004).
Our work here recovers the Heisenberg vertex algebra, the free fermion vertex
algebra, and the affine Kac-Moody vertex algebras. These methods can be applied,
however, to any dg Lie algebra with an invariant pairing, so there is a plethora of
unexplored prefactorization algebras provided by this construction.
4.1. The context. The input data is the following:

• a Riemann surface Σ ;
• a Lie algebra g (for simplicity, we stick to ordinary Lie algebras like
sl2 );
• a g-invariant symmetric pairing κ : g⊗2 → C.
From this data, we obtain a cosheaf on Σ,
gΣ : U 7→ (Ω0,∗
c (U) ⊗ g, ∂),
where U denotes an open in Σ. Note that gΣ is a cosheaf of dg vector spaces

and merely a precosheaf of dg Lie algebras. When κ is nontrivial (though not
necessarily nondegenerate), we obtain a −1-shifted central extension on each open:
gΣκ : U 7→ (Ω0,∗
c (U) ⊗ g, ∂) ⊕ C · c,
4. KAC-MOODY ALGEBRAS AND FACTORIZATION ENVELOPES 155
where C denotes the locally constant cosheaf on Σ and c is a central element of

cohomological degree 1. The bracket is defined by
Z !
1
[α ⊗ X, β ⊗ Y]κ := α ∧ β ⊗ [X, Y] − ∂α ∧ β κ(X, Y)c,
2πi U
with α, β ∈ Ω0,∗ (U) and X, Y ∈ g. (These constants are chosen to match with the
use of κ for the affine Kac-Moody algebra below.)
Remark: Every dg Lie algebra g has a geometric interpretation as a formal moduli
space Bg, a theme we develop further in Volume 2. This dg Lie algebra gΣ (U) in
fact possesses a natural geometric interpretation: it describes “deformations with
compact support in U of the trivial G-bundle on Σ.” Equivalently, it describes the
moduli space of holomorphic G-bundles on U which are trivialized outside of a
compact set. For U a disc, it is closely related to the affine Grassmannian of G.
The affine Grassmannian is defined to be the space of algebraic bundles on a formal
disc trivialized away from a point, whereas our formal moduli space describes G-
bundles on an actual disc trivialized outside a compact set.
The choice of κ has the interpretation of a line bundle on the formal moduli
problem BgΣ (U) for each U. In general, −1-shifted central extensions of a dg Lie
algebra g are the same as L∞ -maps g → C, that is, as rank-one representations.
Rank-one representations of a group are line bundles on the classifying space of
the group. In the same way, rank-one representations of a Lie algebra are line
bundles on the formal moduli problem Bg. ^
As explained in Section 6, we can form the twisted prefactorization envelope
of gΣ .
Concretely, this prefactorization algebra assigns to an open subset U ⊂ Σ,
the complex
F κ (U) = C∗ (gΣκ (U))
= (Sym(Ω0,∗
c (U) ⊗ g[1])[c], dCE ),
where c now has cohomological degree 0 in the Lie algebra homology complex.
It is a prefactorization algebra in modules for the algebra C[c], generated by the
central parameter. We should therefore think of it as a family of prefactorization
algebras depending on the central parameter c.
Remark: Given a dg Lie algebra (g, d), we interpret C∗ g as the “distributions with
support on the closed point of the formal space Bg.” Hence, our prefactoriza-
tion algebras F κ (Σ) describes the κ-twisted distributions supported at the point in
BunG (Σ) given by the trivial bundle on Σ.
This description is easier to understand in its global form, particularly when Σ
is a closed Riemann surface. Each point of P ∈ BunG (Σ) has an associated dg Lie
algebra gP describing the formal neighborhood of P. This dg Lie algebra, in the
case of the trivial bundle, is precisely the global sections over Σ of gΣ . For a nontriv-
ial bundle P, the Lie algebra gP is also global sections of a natural cosheaf, and we
can apply the enveloping construction to this cosheaf to obtain a prefactorization
algebras. By studying families of such bundles, we recognize that our construction
C∗ gP should recover differential operators on BunG (Σ). When we include a twist κ,
we should recover κ-twisted differential operators. When the twist is integral, the
twist corresponds to a line bundle on BunG (Σ) and the twisted differential operators
are precisely differential operators for that line bundle.
It is nontrivial to properly define differential operators on the stack BunG (Σ),
and we will not attempt it here. At the formal neighbourhood of a point, however,
there are no difficulties and our statements are rigorous. ^
4.2. The main result. Note that if we take our Riemann surface to be C, the
prefactorization algebra F κ is holomorphically translation invariant because the
∂ ∂
derivation ∂z is homotopically trivial, via the homotopy given by ∂dz . It follows
that we are in a situation where we might be able to apply Theorem 2.2.1. The
main result of this section is the following.
Theorem. The holomorphically translation invariant prefactorization algebra F κ
on C satisfies the conditions of Theorem 2.2.1, and so defines a vertex algebra.
This vertex algebra is isomorphic to the affine Kac-Moody vertex algebra.
Before we can prove this statement, we of course need to describe the affine
Kac-Moody vertex aigebra.
Recall that the Kac-Moody Lie algebra is the central extension of the loop
algebra Lg = g[t, t−1 ],
0 → C · c →b gκ → Lg → 0.
As vector spaces, we have gκ = g[t, t ] ⊕ C · c, and the Lie bracket is given by the
b −1
formula
I !
[ f (t) ⊗ X, g(t) ⊗ Y]κ := f (t)g(t) ⊗ [X, Y] + f ∂g κ(X, Y)
for X, Y ∈ g and f, gH ∈ C[t, t−1 ]. Here, c has cohomological degree 0 and is

central. The notation denotes an algebraic H version of integration around the unit
circle (aka the residue pairing), and so tn dt = 2πiδn,−1 . In particular, we have
t ∂t = 2πimδm+n,0 .
H
n m
Observe that g[t] is a sub Lie algebra of the Kac-Moody algebra. The vacuum
module W for the Kac-Moody algebra is the induced representation from the trivial
rank one representation of g[t]. The induction-restriction adjunction provides a
natural map C → W of g[t]-modules, where C is the trivial representation of g[t].
Let |0i ∈ W denote the image of 1 ∈ C.
There is a useful, explicit description of W as a vector space. Consider the sub
Lie algebra C · c ⊕ t−1 g[t−1 ] of b
gκ , which is complementary to g[t]. The action of
this Lie algebra on the vacuum element |0i ∈ W gives a canonical isomorphism
U(C · c ⊕ t−1 g[t−1 ]) W
as vector spaces.
The vacuum module W is also a C[c] module in a natural way, because C[c] is
inside the universal enveloping algebra of b
gκ .
4.2.1 Definition. The Kac-Moody vertex algebra is defined as follows. It is a vertex
algebra structure over the base ring C[c] on the vector space W. (Working over
the base ring C[c] simply means all maps are C[c]-linear.) By the reconstruction
theorem 2.3.1, to specify the vertex algebra structure it suffices to specify the state-
field map on a subset of elements of W that generate all of W (in the sense of
the reconstruction theorem). The following state-field operations define the vertex
algebra structure on W.
((i)) The vacuum element |0i ∈ W is the unit for the vertex algebra, that is,
Y(|0i , z) is the identity.
((ii)) If X ∈ t−1 g ⊂ bgκ , we have an element X |0i ∈ W. We declare that
X
Y(X |0i , z) = Xn z−1−n
n∈Z
where Xn = tn X ∈ b
gκ , and we are viewing elements of b
gκ as endomor-
phisms of W.
4.3. Verification of the conditions to define a vertex algebra. We need to

verify that F κ satisfies the conditions listed in Theorem 2.2.1 guaranteeing that we
can construct a vertex algebra. Our situation here is entirely parallel to that of the
βγ system, so we will be brief. The first thing to check is that the natural S 1 action
on F κ (D(0, r)) extends to an action of the algebra D(S 1 ) of distributions on the
circle, which is easy to see by the same methods as with βγ.
Next, we need to check that, if Fnκ (D(0, r)) denotes the weight n eigenspace for
the S 1 -action, then the following properties hold.
((i)) The inclusion Fnκ (D(0, r)) → Fnκ (D(0, s)) for r < s is a quasi-isomorphism.
((ii)) For n 0, the cohomology H ∗ (Flκ (D(0, r)) vanishes as a sheaf on the
site of smooth manifolds.
((iii)) The differentiable vector spaces H ∗ (Fnκ (D(0, r)) are countable sequen-
tial colimits of finite-dimensional vector spaces in DVS.
Note that
F κ (D(0, r)) = Sym Ω0,∗

c (D(0, r), g)[1] ⊕ C · c
with differential dCE the Chevalley-Eilenberg differential for gCκ . Give F κ (D(0, r)
an increasing filtration, by degree of the symmetric power. This filtration is com-
patible with the action of S 1 and of D(S 1 ). In the associated graded, the differential
is just that from the differential ∂ on Ω0,∗
c (D(0, r)).
It follows that there is a spectral sequence of differentiable cochain complexes
H ∗ Gr∗ Flκ (D(0, r)) ⇒ H ∗ Flκ (D(0, r)) .

The analytic results we proved in Section 3.5 concerning compactly

supported Dol-
∗ κ

∗
beault cohomology immediately imply that H Gr Fl (D(0, r)) satisfy properties
(1) − (3) above. It follows that these properties also hold for H ∗ (Flκ (D(0, r)).
4.4. Proof of the theorem. Let us now prove that the vertex algebra associ-
ated to Fκ is isomorphic to the Kac-Moody vertex algebra. The proof will be a
little different than the proof of the corresponding result for the βγ system.
We first prove a statement concerning the behaviour of the prefactorization

algebra Fκ on annuli. Consider the radial projection map
ρ : C× → R>0 ,
z 7→ |z| .
We define the pushforward prefactorization algebra ρ∗ Fκ on R>0 that assigns to
any open subset U ⊂ R>0 , the cochain complex Fκ (ρ−1 (U)). In particular, this
prefactorization algebra assigns to an interval (a, b), the space Fκ (A(a, b)), where
A(a, b) indicates the annulus of those z with a < |z| < b. The product map for ρ∗ F
associated to the inclusion of two disjoint intervals in a larger one arises from the
product map for F from the inclusion of two disjoint annuli in a larger one.
Recall from example 1.1 that any associative algebra A gives rise to a prefac-
torization algebra A f act on R which assigns to the interval (a, b) the vector space A
and whose product map is the multiplication in A. The first result we will show is
the following.
4.4.1 Proposition. There is an injective map of C[c]-linear prefactorization alge-
bras on R>0
gκ ) f act → H ∗ (ρ∗ F κ )
i : U(b
whose image is a dense subspace.
This map i is controlled by the following property. Observe that, for every
open subset U ⊂ C, the subspace of linear elements
κ
Ω0,∗
c (U, g)[1] ⊕ C · c ⊂ C ∗ (Ωc (U, g) ⊕ C · c[1]) = F (U)
0,∗
is, in fact, a subcomplex. Applying this fact to U = ρ−1 (I) for an interval I ⊂ R>0
and taking cohomology, we obtain a linear map
−1 κ
H 1 (Ω0,∗ 0
c (ρ (I))) ⊗ g) ⊕ C · c → H (F (I)).
Note that we have a natural identification
c (ρ (I))) = Ωhol (ρ (I)) ,
−1 −1 ∨
H 1 (Ω0,∗ 1
where ∨ indicates continuous linear dual. Thus, the linear elements of H ∗ F κ ,

which generate the factorization algebra in a certain sense that we’ll identify below,
are g-valued linear functionals on holomorphic 1-forms.
The following notation will be quite useful. Fix a circle {|z| = r} where
r ∈ I. Performing a contour integral around this circle defines a linear function
on Ω1hol (ρ−1 (I)), and so an element of H 1 (Ω0,∗
c (ρ (I))). Denote it by φ(1). Like-
−1
wise, for each integer n, performing a contour integral against zn around this circle
c (ρ (I))) that we call φ(z ).
defines an element of H 1 (Ω0,∗ −1 n
The map
gκ → H ∗ (F κ (ρ−1 (I)))
i :b
constructed by the theorem factors through the map
i: gκ → H 1 (Ω0,∗
b −1
c (ρ (I))) ⊗ g ⊕ C · c
n
Xz 7→ φ(zn ) ⊗ X .
c 7→ c
Note that Cauchy’s integral theorem ensures that, in a sense, it does not matter
what circle we use for the contour integrals that define φ. Hence, this map should
produce a map from the universal enveloping algebra of b gκ , viewed as a locally
∗ κ
constant prefactorization algebra, to H F .
Proof. In Chapter 3, Section 4, we showed how the universal enveloping alge-
bra of any Lie algebra a arises as a prefactorization envelope. Let U f act (a) denote
the prefactorization algebra on R that assigns to an interval I, the cochain complex
U(a) f act (I) = C∗ (Ω∗c (I, a)).
We showed that the cohomology of U(a) f act is locally constant and corresponds to
the ordinary universal enveloping algebra Ua.
Let us apply this construction to a = b
gκ . To prove the theorem, we will produce
a map of prefactorization algebras on R>0
gκ ) f act → ρ∗ F κ .
i : U(b
Since both sides are defined as the Chevalley-Eilenberg chains of local Lie alge-
bras, it suffices to produce such a map at the level of dg Lie algebras.
We will produce such a map in a homotopical sense. To be explicit, we will
introduce several precosheaves of Lie algebras and quasi-isomorphisms
' '
(†) Ω∗c ⊗b
gκ −→ L1 L01 −→ L2 .
In the middle, between L1 and L01 , there will be a cochain homotopy equivalence.
At the level of cohomology, we then obtain an isomorphism

gκ = H ∗ (Ω∗c ⊗b
b gκ ) −→ H ∗ L2
that proves the theorem.
Let us remark on a small but important point. These precosheaves will satisfy
a stronger condition that ensures the functor C∗ of Chevalley-Eilenberg chains pro-
duces a prefactorization algebra. Consider the symmetric monoidal category whose
underlying category is dg Lie algebras but whose symmetric monoidal structure is
given by direct sum as cochain complexes. (The coproduct in dg Lie algebras is not
this direct sum.) All four precosheaves from the sequence (†) are prefactorization
algebras valued in this symmetric monoidal category. In general, a precosheaf L
of dg Lie algebras is such a prefactorization algebra if it has the property that that
for any two disjoint opens U, V in W, the elements in L(W) coming from L(U)
commute with those coming from L(V).
We now describe these prefactorization dg Lie algebras.
Let L1 be the prefactorization dg Lie algebras on R that assigns to an interval
I the dg Lie algebra

L1 (I) = Ω∗c (I) ⊗ g[z, z−1 ] ⊕ C · c[−1].
Thus, L1 is a central extension of Ω∗c ⊗ g[z, z−1 ]. The cocycle defining the central
extension on the interval I is
Z ! I !
αXzn ⊗ βYzm 7→ α ∧ β κ(X, Y) zn ∂z zm ,
I
where α, β ∈ Ω∗c (U).

Note that there is a natural map of such prefactorization dg Lie algebras
j : Ω∗c ⊗b
gκ → L1
that is the identity on Ω∗c ⊗ g[z, z−1 ] but on the central element is given by
Z !
j(αc) = α c,
I
where α ∈ Ω∗c (I).
This map is clearly a quasi-isomorphism when I is an interval.
Hence, we have produced the first map in the sequence (†).
It follows immediately that the map
gκ ) f act = C∗ (Ω∗c ⊗b
C∗ j : U(b g κ ) → C ∗ L1
is a map of prefactorization algebras that is a quasi-isomorphism on intervals.
Therefore, the cohomology prefactorization algebra of C∗ L1 assigns to an inter-
val U(bgκ ), and the prefactorization product is just the associative product on this
algebra.
Let L2 be the prefactorization dg Lie algebra ρ∗ gCκ . In other words, it assigns
to an interval I the dg Lie algebra

L2 (I) = Ω0,∗
c (ρ −1
(I)) ⊗ g ⊕ C · c[−1],
with the central extension inherited from gCκ . Hence,
C ∗ L2 = ρ ∗ F κ ,
so that the last element in the sequence (†) is what we need for the theorem.
Finally, we define L01 to be a central extension of Ω∗c ⊗ g[z, z−1 ], but where the
cocycle defining the central extension is
∂
Z ! I ! Z !
αXz ⊗ βYz 7→
n m
α ∧ β κ(X, Y) z ∂z z + πκ(X, Y)δn+m,0
n m
αr β ,
I I ∂r
∂
where the vector field r ∂r acts by the Lie derivative on the form β ∈ Ω∗c (I). It is
easy to verify that this cochain is closed and so defines a central extension. Note
that L01 looks like L1 except with a small addition to the cocycle giving the central
extension.
Remark: This central extension, as well as the ones defining L1 and L2 , are lo-
cal central extensions of local dg Lie algebras in the sense of Definition 6.3.1 of
Chapter 3. This concept is studied in more detail in Volume 2. ^
To prove the proposition, we will do the following.
((i)) Prove that L1 and L01 are homotopy equivalent prefactorization dg Lie
algebras.
((ii)) Construct a map of precosheaves of dg Lie algebras L01 → L2 .
For the first point, note that the extra term
∂
Z
πκ(X, Y)δn+m,0 αr β
U ∂r
in the coycle for L01 is an exact cocycle. It is cobounded by the cochain

Z
πκ(X, Y)δn+m,0 αrι ∂ β,
∂r
U
where ι ∂ indicates contraction. The fact that this expression cobounds follows
∂r
from the Cartan homotopy formula for the Lie derivative of vector fields on differ-
ential forms. Since the cobounding cochain is also local, L1 and L01 are homotopy
equivalent as prefactorization dg Lie algebras.
Now we will produce the desired map Φ : L01 → L2 . We use the following
notation:
• r denotes both the coordinate on R>0 and also the radial coordinate in
C× ,
• θ the angular coordinate on C× , and
• X will denote an element of g.
We will view b gκ as g[z, z−1 ] ⊕ C · c.
On an open I ⊂ R>0 , the map Φ is
f (r)Xzn 7→ f (r)zn X,
1 iθ
f (r)dr Xzn 7→ n
2 e f (r)z dz X,
c 7→ c,
with f ∈ Ω0c (U). It is a map of precosheaves but we need to verify it is a map of dg
∂
Lie algebras. Compatibility with the differential follows from the formula for ∂z in
polar coordinates:
∂ ∂ 1 ∂
!
1
= eiθ − .
∂z 2 ∂r ir ∂θ
Only the central extension might cause incompatibility with the Lie bracket.
Consider the following identity:
Z ! Z
1 iθ
f (r)z ∂ z g(r) e dz = 2πimδn+m,0
n m
f (r)g(r)dr
ρ−1 (I) 2 I
∂
Z
+ πδn+m,0 f (r)r g(r)dr,
I ∂r
where I ⊂ R>0 is an interval. The expression on the right hand side gives the central
extension term in the Lie bracket on L01 , whereas that on the left is the central
extension term for L2 applied to the elements Φ( f (r)Xzn ) and Φ(g(r)dr Yzm ).
Applying Chevalley chains, we get a map of prefactorization algebras
C∗ Φ : C∗ L01 → C∗ L2 = ρ∗ Fκ .
Since C∗ L01 is homotopy equivalent to C∗ L1 (although we have not given an equiv-
alence explicitly), we get the desired map of prefactorization algebras on R>0 :
gκ ) f act → H ∗ (ρ∗ Fκ ).
U(b
The analytical results about compactly supported Dolbeault cohomology in Section
3.5 imply immediately this map has dense image.
Finally, we check that this map has the properties discussed just after the state-
ment of the theorem. Let I ⊂ R>0 be an interval. Then, under the isomorphism

→ H ∗ (C∗ (L01 (I))),
gκ ) −
i : U(b
the element Xzn is represented by the element
f (r)zn Xdr ∈ L01 (U)[1],
where f has compact support and is chosen so that f (r)dr = 1.
R
To check the desired properties, we need to verify that if α is a holomorphic

1-form on ρ−1 (I), then
Z I
α f (r)zn 21 eiθ dz = αzn .
ρ−1 (I) |z|=1
But this identity follows immediately from Stokes theorem and the equation
f (r) 21 eiθ dz = ∂(h(r)),
Rr
where h(r) = ∞
f (r)dr.
This theorem shows how to relate observables on an annulus to the universal
enveloping algebra of the affine Kac-Moody Lie algebra. Recall that the Kac-
Moody vertex algebra is the vacuum representation equipped with a vertex algebra
structure. The next result will show that the vertex algebra associated to the pref-
actorization algebra F κ is also a vertex algebra structure on the vacuum represen-
tation.
More precisely, we will show the following.
4.4.2 Proposition. Let
H ∗ (Fnκ (D(0, ε))
M
V=
n
be the cohomology of the direct sum of the weight n eigenspaces of the S 1 action
on Fκ (D(0, ε)). Let A(r, r0 ) be the annulus {r < |z| < r0 }.
The map
gκ ) → H ∗ (F κ (A(r, r0 ))
i : U(b
gκ ) on V. Moreover,
constructed in the previous proposition induces an action of U(b
there is a unique isomorphism of U(b gκ )-modules from V to the vacuum module W,
which sends the unit observable 1 ∈ V to the vacuum element |0i ∈ W.
Proof. If ε < r < r0 , then the prefactorization product gives a map
F κ (D(0, ε)) × F κ (A(r, r0 )) → F κ (D(0, r0 ))
of cochain complexes. Passing to cohomology, and using the relationship between
U(bgκ ) and F κ (A(r, r0 )), we get a map
gκ ) → H ∗ (F κ (D(0, r0 ))).
V ⊗ U(b
Note that on the left hand side, every element is a finite sum of elements in S 1 -
eigenspaces. Since the map is S 1 -equivariant, its image is the subspace in H ∗ (F (D(0, r0 )))
that consists of finite sums of S 1 -eigenvectors. This subspace is V. We therefore

find a map
(†) V ⊗ U(b
gκ ) → V.
To see that it makes V a module over U(b gκ ), combine the following observations.
First, the map i : U(bgκ ) → H ∗ (F κ (A(r, r0 )) takes the asosciative product on U(b
gκ )
to the prefactorization product map
H ∗ (F κ (A(r, r0 )) ⊗ H ∗ (F κ (A(s, s0 )) → H ∗ (F κ (A(r, s0 ))
for r < r0 < s < s0 . Second, the diagram of prefactorization product maps
F κ (D(0, ε)) ⊗ F κ (A(r, r0 )) ⊗ F κ (A(s, s0 )) / F κ (D(0, ε)) ⊗ F κ (A(r, s0 ))

F κ (D(0, r0 )) ⊗ F κ (A(s, s0 )) / F κ (D(0, s0 ))
commutes.
We need to show that this action identifies V with the vacuum representation
W. The putative map from W to V sends the vacuum element |0i to the unit ob-
servable 1 ∈ V. To show that this map is well-defined, we need to show that 1 ∈ V
is annihilated by the elements zn X ∈ b
gκ for n ≥ 0.
The unit axiom for prefactorization algebras implies that the following diagram
commutes:
U(bgκ ) / H ∗ (F κ (A(s, s0 )) .

V / H ∗ (F κ (D(0, s0 ))
The left vertical arrow is given by the action of U(b gκ ) on the unit element 1 ∈ V, and
the right vertical arrow is the map arising from the inclusion A(r, r0 ) ⊂ D(0, r0 ). The
bottom right arrow is the inclusion into the direct sum of S 1 -eigenspaces, which is
injective.
The proof of Theorem 4.4.1 gives an explicit representative for the element
Xzn ∈ b gκ in F κ (A(s, s0 )). Namely, let f (r) be a function supported in the interval
(s, s ) and with f (r)dr = 1. Then Xzn is represented by
R
0
1 iθ n
2 e f (r)dzz X ∈ Ω0,1 0
c (A(s, s )) ⊗ g.
We will show that this element is exact when it is viewed as an element of Ω0,1 0
c (D(0, s ))⊗
n
g, so that Xz vanishes when going right and then down in the square above.
Set Z r
h(r) = f (t)dt,
∞
∂
so h(r) = 0 for r and h(r) = 1 for r < s. Also,
s0 ∂r h = f . The polar-coordinate
∂
representation of ∂z tells us that
∂h(r)zn X = 12 eiθ f (r)dzzn X.

Thus, we have shown that the elements Xzn ∈ b gκ , where n ≥ 0, act by zero on the
element 1 ∈ V. We therefore have a unique map of U(b gκ -modules from W to V
sending |0i to 1.
It remains to show that this map is an isomorphism. Note that every object
we are discussing is filtered. The universal enveloping algebra U(b gκ ) is filtered, as
usual, by setting F i U(b gκ ) to be the subspace spanned by products of ≤ i elements
of b
gκ . (This filtration is inherited from the tensor algebra.) Similarly, the space
F κ (U) = C∗ (Ω0,∗
c (U, g) ⊕ C · c[−1])
is filtered by saying that F i F κ (U) is the subcomplex C≤i . All the maps we have
been discussing are compatible with these increasing filtrations.
The associated graded of U(b gκ ) is the symmetric algebra of b
gκ , and the associ-
ated graded of F κ (U) is the appropriate symmetric algebra on Ω0,∗ c (U, g)[1] ⊕ C · c
in differentiable cochain complexes. Upon taking associated graded, the maps
gκ ) → Gr H ∗ F κ (A(s, s0 )),
Gr U(b
Gr H ∗ (F κ (A(s, s0 )) → Gr H ∗ F κ (D(0, s0 ))
are maps of commutative algebras. It follows that the map

gκ ) → Gr V ⊂ H ∗ Sym Ω0,∗
Gr U(b c (D(0, s), g)[1] ⊕ C · c
is a map of commutative algebras. Moreover, Gr V is the direct sum of the S 1 -

eigenspaces in the space on the right side of this map. The direct sum of the S 1 -
eigenspaces in H 1 (Ω0,1 (D(0, s)) is naturally identified with z−1 C[z−1 ]. Thus, we
find the associated graded of the map U(b gκ ) to V is a map of commutative algebras

Sym g[z, z−1 ] [c] → Sym z−1 g[z−1 ] [c].
We have already calculated that on the generators of the commutative algebra, it
arises from the natural projection map
g[z, z−1 ] ⊕ C · c → z−1 g[z−1 ] ⊕ C · c.
It follows immediately that the map Gr W → Gr V is an isomorphism, as desired.

In order to complete the proof that the vertex algebra associated to the prefac-
torization algebra F κ is isomorphic to the Kac-Moody vertex algebra, we need to
identify the operator product expansion map
Y : V ⊗ V → V((z)).
Recall that this map is defined, in terms of the prefactorization algebra F κ , as
follows. Consider the map
mz,0 : V ⊗ V → H ∗ (F κ (D(0, ∞))
defined by restricting the prefactorization product map
H ∗ (F κ (D(z, ε)) × H ∗ (F κ (D(0, ε)) → H ∗ (F κ (D(0, ∞))
to the subspace
V ⊂ H ∗ (F κ (D(z, ε)) = H ∗ (F κ (D(0, ε)).
Composing the map mz,0 with the map

H ∗ (F κ (D(0, ∞)) → V =
Y
Vk
k
given by the product of the projection maps to the S 1 -eigenspaces, we get a map
mz,0 : V ⊗ V → V.
This map depends holomorphically on z. The operator product map is obtained as
the Laurent expansion of mz,0 .
Our aim is to calculate the operator product map and identify it with the vertex
operator map in the Kac-Moody vertex algebra. We use the following notation. If
X ∈ g, let Xi = zi X ∈ b gκ on V by the symbol ·. Then we
gκ . We denote the action of b
have the following.
4.4.3 Proposition. For all v ∈ V,
X
mz,0 (X−1 · 1, v) = z−i−1 (Xi · v) ∈ V
i∈Z
where the sum on the right hand side converges.
Before we prove this proposition, let us observe that it proves our main result.
4.4.4 Corollary. This isomorphism of U(b gκ )-modules from the vacuum represen-
tation W to V is an isomorphism of vertex algebras, where V is given the vertex
algebra structure arising from the prefactorization algebra F κ , and W is given the
Kac-Moody vertex algebra structure defined in 4.2.1.
Proof. This corollary follows immediately from the reconstruction theorem
2.3.1.
Proof of the proposition. The inclusion of a disc into an annulus induces a
structure map of the prefactorization algebra. Each element v of V thus determines
an element of the prefactorization algebra on that annulus, and hence a map from
V to V. The formula will follow by applying our earlier analysis of the structure
maps.
As before, we let Xi ∈ b gκ denote Xzi for X ∈ g. As we explained in the
discussion following Theorem 4.4.1, we can view the element
X−1 · 1 ∈ V
as being represented at the cochain level by X times the linear functional
Ω1hol (D(0, s)) → R C
α 7→ |z|=r z−1 α.
Cauchy’s theorem tells us that this linear functional sends h(z)dz, when h is holo-
morphic, to 2πih(0).
Now fix z0 ∈ A(s, s0 ), and let
ιz0 : V → H ∗ (F κ (A(s, s0 )))
denote the map arising from the restriction to V of the structure map
H ∗ (F κ (D(0, ε))) = H ∗ (F κ (D(z0 , ε))) → H ∗ (F κ (A(s, s0 )))
arising from the inclusion of the disc D(z0 , ε) into the annulus A(s, s0 ). It is clear
from the definition of mz0 ,0 and the axioms of a prefactorization algebra that the
following diagram commutes:
mz0 ,0
V ⊗ V ⊗ Id /V .
O
ιz0

H ∗ (F κ (A(s, s0 )) ⊗ V / H ∗ (F κ (D(0, ∞)))
Here the bottom right arrow is the restriction to V of the prefactorization structure
map
H ∗ (F κ (A(s, s0 )) ⊗ H ∗ (F κ (D(0, ε)) → H ∗ (F κ (D(0, ∞)).
It therefore suffices to show that
∈ H ∗ (F κ (A(s, s0 )))
X
ιz0 (X−1 · 1) = Xi z−i−i
0
i∈Z
gκ as elements of H 0 (F κ (A(s, s0 ))) via the map

where we view Xi ∈ b
gκ ) → H 0 (F κ (A(s, s0 )))
U(b
constructed in Theorem 4.4.1.
It is clear from the construction of F κ that ιz0 (X−1 · 1) is in the image of the
natural map
κ
Ω0,∗ 0 0
c (A(s, s )) ⊗ g → F (A(s, s )).
Recall that the cohomology of Ω0,∗ 0
c (A(s, s )) is the linear dual of the space of holo-
morphic 1-forms on the annulus A(s, s0 ). The element ιz0 (X−1 · 1) can thus be
represented by the continuous linear map
Ω1hol (D(0, s)) → R g
α 7→ |z|=r
z−1 α X.
Similarly, the elements Xi ∈ H ∗ (F κ (A(s, s0 ))) are represented by the linear maps
Z !
i
h(z)dz 7→ z h(z)dz X.
|z|=r
It remains to show that, for all holomorphic functions h(z) on the annulus A(s, s0 ),
we have I !
X
2πih(z0 ) = i
z0 z −i−1
h(z)dz .
i∈Z |z|=s+ε
This equality is proved using Cauchy’s theorem. We know
I I
h(z) h(z)
2πih(z0 ) = dz − dz.
|z|=s+ε z − z0 |z|=s0 −ε z − z0
Expanding (z−z0 )−1 in the regions when |z| < |z0 | (relevant for the first integral) and
when |z| > |z0 | (relevant for the second integral) gives the desired expression.
Part 3
Factorization algebras
CHAPTER 6
Factorization algebras: definitions and constructions
Our definition of a prefactorization algebra is closely related to that of a pre-

cosheaf or of a presheaf. Mathematicians have found it useful to refine the axioms
of a presheaf to those of a sheaf: a sheaf is a presheaf whose value on a large open
set is determined, in a precise way, by values on arbitrarily small subsets. In this
chapter we describe a similar “descent” axiom for prefactorization algebras. We
call a prefactorization algebra satisfying this axiom a factorization algebra.
After defining this axiom, our next task is to verify that the examples we have
constructed so far, such as the observables of a free field theory, satisfy it. This we
do in sections 5 and 6.
Philosophically, our descent axiom for factorization algebras is important: a
prefactorization algebra satisfying descent (i.e., a factorization algebra) is built
from local data, in a way that a general prefactorization algebra need not be. How-
ever, for many practical purposes, such as the applications to field theory, this
axiom is often not essential.
A reader with little taste for formal mathematics could thus skip this chapter
and the next and still be able to follow the rest of this book.
1. Factorization algebras
A factorization algebra is a prefactorization algebra that satisfies a local-to-
global axiom. This axiom is the analog of the gluing axiom for sheaves; it ex-
presses how the values on big open sets are determined by the values on small
open sets. We thus begin by reviewing the notion of a (co)sheaf, with more back-
ground and references available in Appendix 4. Next, we introduce Weiss covers,
which are the type of covers appropriate for factorization algebras. Finally, we give
the definition of factorization algebras,
1.1. Sheaves and cosheaves. In order to motivate our definition of factoriza-

tion algebra, let us recall the sheaf axiom and then work out its dual, the cosheaf
axiom.
Let M be a topological space and let Opens(M) denote the poset category of
open subsets of a space M. A presheaf A on a topological space M with values in
a category C is a functor from Opens(M)op to C. Given an open cover {Ui | i ∈ I}
of an open set U ⊂ M, there is a canonical map
Y
A(U) → A(Ui )
i
169
170 6. FACTORIZATION ALGEBRAS: DEFINITIONS AND CONSTRUCTIONS
given by taking the product of the structure maps A(U) → A(Ui ). A presheaf A
is a sheaf if for every U and for every open cover of U, this map equalizes the pair
of maps Y Y
A(Ui ) ⇒ A(Ui ∩ U j )
i i, j
given by the two inclusions Ui ∩ U j → Ui and Ui ∩ U j → U j . That is, A(U) is
the limit of that diagram. In words, this condition means that a section s of A on
U can be described as sections si on each element of the cover that agree on the
overlaps, i.e., si |Ui ∩U j = s j |Ui ∩U j .
We now dualize this definition. A precosheaf Φ on M with values in a category
C is a functor from Opens(M) to C. It is a cosheaf if for every open cover of U, the
canonical map given by taking the coproduct of the structure map Φ(Uk ) → Φ(U),
a
Φ(Uk ) → Φ(U),
k
coequalizes the pair of maps
a a
Φ(Ui ∩ U j ) ⇒ Φ(Uk )
i, j k
given by the two inclusions Ui ∩ U j → Ui and Ui ∩ U j → U j . That is, Φ(U) is the

colimit of that diagram. Thus, a cosheaf with values in C can be understood as a
sheaf with values in Cop .
In this book, our target categories are linear in nature, so that C is the category
of vector spaces or of dg vector spaces or of cochain complexes with values in some
additive category, such as differentiable vector spaces. In such a setting, we can
write the cosheaf axiom in an equivalent but slightly different way. A precosheaf
Φ on M with values in, say, vector spaces is a cosheaf if, for every open cover of
any open set U, the sequence
M M
Φ(Ui ∩ U j ) → Φ(Uk ) → Φ(U) → 0
i, j k
is exact, where the first map is the difference of the structure maps for the two
inclusions. In the remainder of this book, we will phrase our gluing axioms in this
style, as we will always restrict to linear target categories.
1.2. Weiss covers. For factorization algebras, we require our covers to be fine
enough that they capture all the “multiplicative structure” — the structure maps —
of the underlying prefactorization algebra. In fact, a factorization algebra will be a
cosheaf with respect to this modified notion of cover.
1.2.1 Definition. Let U be an open set. A collection of open sets U = {Ui | i ∈ I}
is a Weiss cover of U if for any finite collection of points {x1 , . . . , xk } in U, there is
an open set Ui ∈ U such that {x1 , . . . , xk } ⊂ Ui .
The Weiss covers define a Grothendieck topology on Opens(M), the poset cate-
gory of open subsets of a space M. We call it the Weiss topology of M. (This notion
1. FACTORIZATION ALGEBRAS 171
was introduced in Weiss (1999) and it is explained nicely and further developed in
Boavida de Brito and Weiss (2013).)
Remark: A Weiss cover is certainly a cover in the usual sense, but a Weiss cover
typically contains an enormous number of opens. It is a kind of “exponentiation”
of the usual notion of cover, because a Weiss cover is well-suited to studying all
configuration spaces of finitely many points in U. For instance, given a Weiss cover
U of U, the collection
{Uin | i ∈ I}
provides a cover in the usual sense of U n ⊂ M n for every positive integer n. ^
Example: For a smooth n-manifold M, there are several simple ways to construct
a Weiss cover for M. One construction is simply to take the collection of open sets
in M diffeomorphic to a disjoint union of finitely many copies of the open n-disc.
Alternatively, the collection of opens {M \ q | q ∈ M} is a Weiss cover of M. One
can also produce a Weiss cover of a manifold with countably many elements. Fix
a Riemannian metric on M. Pick a collection of points {qn ∈ M | n ∈ N} such that
the union of the unit “discs” D1 (qn ) = {q ∈ M | d(qn , q) < 1} covers M. (Such
an open may not be homeomorphic to a disc if the injectivity radius of qn is less
than 1.) Consider the collection of all “discs” D = {D1/m (qn ) | m ∈ N, n ∈ N}.
The collection of all finite, pairwise disjoint union of elements of D is a countable
Weiss cover. ^
The examples above suggest the following definition.
1.2.2 Definition. A cover U = {Uα } of M generates the Weiss cover V if every open
V ∈ V is given by a finite disjoint union of opens Uα from U.
1.3. Strict factorization algebras. The value of a factorization algebra on U

is determined by its behavior on a Weiss cover, just as the value of a cosheaf on an
open set U is determined by its value on any cover of U. Throughout, the target
multicategory C will be linear in nature (e.g., vector spaces or cochain complexes
of differentiable vector spaces).
1.3.1 Definition. A prefactorization algebra on M with values in a multicategory
C is a factorization algebra if it has the following property: for every open subset
U ⊂ M and every Weiss cover {Ui | i ∈ I} of U, the sequence
M M
F (Ui ∩ U j ) → F (Uk ) → F (U) → 0
i, j k
is exact. That is, F is a factorization algebra if it is a cosheaf with respect to the

Weiss topology.
A factorization algebra is multiplicative if, in addition, for every pair of disjoint
open sets U, V ∈ M, the structure map
mU,V
UtV : F (U) ⊗ F (V) → F (U t V)
is an isomorphism.
To summarize, a multiplicative factorization algebra satisfies two conditions:

an axiom of (co)descent and a multiplicativity axiom. The descent axiom says that
it is a cosheaf with respect to the Weiss topology, and the multiplicativity axiom
says that its value on finite collections of disjoint opens factors into a tensor product
of the values on each open.
Remark: It is typically not easy to verify directly that a prefactorization algebra is a
factorization algebra. In the locally constant case discussed in Section 4, we discuss
a theorem that exhibits many more examples, including the prefactorization algebra
on the real line arising from an associative algebra. Later in this chapter, we exhibit
a large class of examples explicitly, arising from the enveloping construction for
local Lie algebras. The next chapter provides analogs of methods familiar from
sheaf theory, such as pushforwards and pullbacks and extension from a basis. ^
1.4. The Čech complex and homotopy factorization algebras. If our target
multicategory has a notion of weak equivalence — such as quasi-isomorphism
between cochain complexes — then there is a natural modification of the notion
of cosheaf. It arises from the fact that one should work with a corrected notion
of colimit that takes into account that some objects are weakly equivalent, even if
there is no strict equivalence between them. In other words, we need to work with
the notion of colimit appropriate to ∞-categories. We will use the term homotopy
colimit for this notion. (Appendix 5 contains an extensive discussion in the context
relevant to us: a category of cochain complexes in an additive category.) One then
revises the cosheaf gluing axiom by using homotopy colimits, leading to the notion
of a homotopy cosheaf. After describing these, we define homotopy factorization
algebras.
Remark: Our target category will always be unbounded cochain complexes with
values in some additive category C, and we want to treat quasi-isomorphic cochain
complexes as equivalent. In other words, we want to treat Ch(C) as an ∞-category,
and there are several approaches are, in a precise sense, equivalent. When C is
a Grothendieck abelian category — as is the case for differentiable vector spaces
DVS — this situation is well-developed, and much of Appendix 5 is devoted to
understanding it. We discuss there a stable model category, a pretriangulated dg
category, a simplicially-enriched category, and a stable quasicategory associated
to Ch(C), and how to transport constructions and statements between all these ap-
proaches. (We rely heavily upon and summarize the approach in Lurie (n.d.b).) In
light of the results explained there, here we will write down explicit cochain com-
plexes that describe the homotopy colimits of the diagrams we need. The reader
familiar with higher categories will recognize how to generalize these definitions
but we do not pursue such generality here. Work in that direction can be found in
Ayala and Francis (2015) or Chapter 5 of Lurie (n.d.b). ^
Let us provide the definition of a homotopy cosheaf. (See Appendix 4.3 for
more discussion.) Let Φ be a precosheaf on M with values in cochain complexes
in an additive category C. Let U = {Ui | i ∈ I} be a cover of some open subset U of
M. Consider the cochain complex with values in cochain complexes whose −kth
1. FACTORIZATION ALGEBRAS 173
term is M
Φ(U j0 ∩ · · · ∩ U jk ),
j0 ,..., jk ∈I
ji all distinct
and note that this term has an “internal” differential arising from Φ. The “external”
differential is the alternating sum of the structure maps arising from inclusion of
k + 1-fold intersections into k-fold intersections. The Čech complex of U with
coefficients in Φ is the totalization of this double complex:
 
∞
 
M  M 
Č(U, Φ) = 
 Φ(U j0 ∩ · · · ∩ U jk )[k] .
 
k=0  j0 ,..., jk ∈I
ji all distinct
There is a natural map from the 0th term (of the double complex) to Φ(U) given by
the sum of the structure maps Φ(Ui ) → Φ(U). Thus, the Čech complex possesses
a natural map to Φ(U). We say that Φ is a homotopy cosheaf if the natural map
from the Čech complex to Φ(U) is a quasi-isomorphism for every open U ⊂ M and
every open cover of U.
Remark: This Čech complex is simply the dual of the Čech complex for sheaves
of cochain complexes. Note that this Čech complex is also the normalized cochain
complex arising from a simplicial cochain complex, where we evaluate Φ on the
simplicial space U• associated to the cover U. In more sophisticated language, the
Čech complex describes the homotopy colimit of Φ on this simplicial diagram. ^.
We can now define our main object of interest. Let C be a multicategory whose
underlying category is a Grothendieck abelian category. Then there is a natural
multicategory whose underlying category is Ch(C), the category of cochain com-
plexes in C in which the weak equivalences are quasi-isomorphisms.
1.4.1 Definition. A homotopy factorization algebra is a prefactorization algebra
F on X valued in Ch(C), with the property that for every open set U ⊂ X and Weiss
cover U of U, the natural map
Č(U, F ) → F (U)
is a quasi-isomorphism. That is, F is a homotopy cosheaf with respect to the Weiss
topology.
If C is a symmetric monoidal category, then we can formulate a multiplicativity

condition, as follows.
1.4.2 Definition. A homotopy factorization algebra F is multiplicative if for every
pair U, V of disjoint open subsets of X, the structure map
F (U) ⊗ F (V) → F (U t V)
Remark: In light of the weak equivalences, the strict notion of multiplicative fac-
torization algebra is not appropriate for the world of cochain complexes. Whenever
we refer to a factorization algebra in cochain complexes, we will mean a homotopy

factorization algebra. ^
Because we will primarily work with differentiable cochain complexes, we
introduce the following terminology.
1.4.3 Definition. A differentiable factorization algebra is a factorization algebra
valued in the multicategory of differentiable cochain complexes.
Remark: Although the multiplicativity axiom does not directly extend to the setting
of multicategories, we can examine the bilinear structure map
mU,V
UtV ∈ Ch(DVS)(F (U), F (V) | F (U t V))
for two disjoint opens U and V, and we can ask about the image of mU,V UtV . For
almost all the examples constructed in this book, the differentiable vector spaces
that appear are convenient vector spaces, which have a natural topology. In our
constructions, one can typically see that the image of the structure map is dense
with respect to this natural topology on F (U t V), which is a cousin of the mul-
tiplicativity axiom. In practice, though, it is the descent axiom that is especially
important for our purposes. ^
1.5. Weak equivalences. The notion of weak equivalence, or quasi-isomorphism,
on cochain complexes induces such a notion on homotopy factorization algebras.
1.5.1 Definition. Let F, G be homotopy factorization algebras valued in Ch(C), as
above. A map φ : F → G of homotopy factorization algebras is a weak equivalence
if, for all open subsets U ⊂ M, the map F(U) → G(U) is a quasi-isomorphism of
cochain complexes.
We now provide an explicit criterion for checking weak equivalences, using

the notion of a factorizing basis (see Definition 2.1.1).
1.5.2 Lemma. A map F → G between differentiable factorization algebras is a
weak equivalence if and only if, for every factorizing basis U of X and every U in
U, the map
F(U) → G(U)
is a weak equivalence.
Proof. For any open subset V ⊂ X, let UV denote the Weiss cover of V gener-
ated by all open subsets in U that lie in V. By the descent axiom, the map
Č(UV , F) → F(V)
is a weak equivalence, and similarly for G. Thus, it suffices to check that the map
Č(UV , F) → Č(UV , G)
is a weak equivalence, for the following reason. Every Čech complex has a natural
filtration by number of intersections. We thus obtain of spectral sequences from
2. FACTORIZATION ALGEBRAS IN QUANTUM FIELD THEORY 175
the map of Čech complexes. The fact that the maps

F(U1 ∩ · · · ∩ Uk ) → G(U1 ∩ · · · ∩ Uk )
are all weak equivalences implies that we have a quasi-isomorphism on the first
page of the map of spectral sequences, so our original map is a quasi-isomorphism.

1.6. The (multi)category structure on factorization algebras. Every fac-
torization algebra is itself a prefactorization algebra. Thus we define the category
of factorization algebras FA(X, C) to be the full subcategory whose objects are the
factorization algebras.
We have also described in Section 1.5 the symmetric monoidal category of
prefactorization algebras. In brief, the tensor product of two prefactorization alge-
bras is given by taking the tensor product of values at each open:
(F ⊗ G)(U) = F(U) ⊗ G(U),
and we simply define the structure maps as the tensor product of the structure maps.
If the tensor product commutes (separately in each variable) with colimits over
simplicial diagrams (i.e., geometric realizations), then the category of factorization
algebras inherits this symmetric monoidal structure structure. The key point is
to ensure that F ⊗ G satisfies the descent axiom, which follows quickly from the
hypothesis on the tensor product. As we will not use this symmetric monoidal
structure on factorization algebras, we do not develop this observation.
2. Factorization algebras in quantum field theory

We have seen in Section 3 how prefactorization algebras appear naturally when
one thinks about the structure of observables of a quantum field theory. It is natural
to ask whether the local-to-global axiom that distinguishes factorization algebras
from prefactorization algebras also has a quantum field-theoretic interpretation.
The local-to-global axiom we posit states, roughly speaking, that all observ-
ables on an open set U ⊂ M can be built up as sums of observables supported on ar-
bitrarily small open subsets of M. To be concrete, let us consider a Weiss cover Uε
of M, built out of all open discs in M of radius smaller than ε. Applied to this Weiss
cover, our local-to-global axiom states that any observable O ∈ Obs(U) can be writ-
ten as a sum of observables of the form O1 O2 · · · Ok , where Oi ∈ Obs(Dδi (xi )) and
x1 , . . . , xk ∈ M.
By taking ε to be very small, we see that our local-to-global axiom implies that
all observables can be written as sums of products of observables whose support
arbitrarily close to being point-like in U.
This assumption is physically reasonable: most of the observables (or oper-
ators) considered in quantum field theory textbooks are supported at points, so it
might make sense to restrict attention to observables built from these.
However, more global observables are also considered in the physics literature.
For example, in a gauge theory, one might consider the observable which measures
the monodromy of a connection around some loop in the space-time manifold.
How would such observables fit into the factorization algebra picture?
The answer reveals a key limitation of our axioms: the concept of factorization
algebra is only appropriate for perturbative quantum field theories.
Indeed, in a perturbative gauge theory, the gauge field (i.e., the connection) is
taken to be an infinitesimally small perturbation A0 + δA of a fixed connection A0 ,
which is a solution to the equations of motion. There is a well-known formula (the
time-ordered exponential) expressing the holonomy of A0 + δA as a power series in
δA, where the coefficients of the power series are given as integrals over Lk , where
L is the loop which we are considering. This expression shows that the holonomy
of A0 +δA can be built up from observables supported at points (which happen to lie
on the loop L). Thus, the holonomy observable will form part of our factorization
algebra.
However, if we are not working in a perturbative setting, this formula does not
apply, and we would not expect (in general) that the prefactorization algebra of
observables satisfies the local-to-global axiom.
3. Variant definitions of factorization algebras

In this section we sketch variations on the definition of (pre)factorization alge-
bra used in this book. Our primary goal is to provide a bridge to the next section,
where we discuss locally constant factorization algebras, which are well-developed
in the topology literature using an approach along the lines sketched here. The
variations we describe undoubtedly admit improvements and modifications; they
should not be taken as definitive. We hope and expect that approaches with this
flavor may nonetheless play a role in future work.
The first observation is that many constructions of (pre)factorization algebras
make sense on a large class of manifolds and not just on a fixed manifold. As
an example, consider the factorization envelope of a dg Lie algebra g. Under a
smooth open embedding f : M → N, the natural map f∗ : Ω∗c (M) → Ω∗c (N) of
extending by zero preserves the wedge product. Hence, given a dg Lie algebra g,
we obtain a natural map f∗ ⊗ idg : Ω∗c (M) ⊗ g → Ω∗c (N) ⊗ g of dg Lie algebras.
Thus, the pushforward f∗ Ug of the factorization envelope on M is isomorphic to the
factorization envelope on N, when restricted to the image of M. In other words, this
factorization envelope construction naturally lives on a category of n-manifolds,
not just on a fixed n-manifold.
We can formalize this kind of structure as follows.
3.0.1 Definition. Let Embn denote the category whose objects are smooth n-manifolds
and whose morphisms are open embeddings. It possesses a symmetric monoidal
structure under disjoint union.
Then we introduce the following variant of the notion of a prefactorization

algebra. (In Section 3.4 below, we explain the appropriate local-to-global axiom.)
3.0.2 Definition. A prefactorization algebra on n-manifolds with values in a sym-
metric monoidal category C is a symmetric monoidal functor from Embn to C.
3. VARIANT DEFINITIONS OF FACTORIZATION ALGEBRAS 177
Observe the following property of such a functor F : the space of embeddings

Embn (M, M) acts on the value F (M). In particular, the diffeomorphism group of
M is a subset of Embn (M, M), and so F (M) has an action of the diffeomorphism
group. Thus a prefactorization algebra F is sensitive to the topology of smooth
manifolds; it is not a trivial generalization of the notion we’ve already developed.
As an example, for any dg Lie algebra g, the factorization envelope Un g : M 7→
C∗ (Ω∗c (M) ⊗ g) provides an a prefactorization algebra on n-manifolds.
3.1. Varying in smooth families. It would be nice, for instance, to let the
diffeomorphism group act smoothly, in some sense, and not just as a discrete group.
Within the framework we’ve developed, there is a natural approach, parallel to
our definition of a smoothly equivariant prefactorization algebra (see Section 7.3).
Note that embeddings define a natural sheaf of sets on the site Mfld of smooth
manifolds: given n-manifolds M, N, the sheaf Emb(M, N) assigns to the manifold
X the subset
{ f ∈ C ∞ (X × M, N) | ∀x ∈ X, f (x, −) is an embedding of M in N}.
In other words, it gives smooth families over X of embeddings of M into N.
3.1.1 Definition. Let Embn denote the category enriched over Sh(Mfld) whose
objects are smooth n-manifolds and in which Embn (M, N) is the sheaf Emb(M, N).
It possesses a symmetric monoidal structure under disjoint union.
If we work with a multicategory enriched over Sh(Mfld), such as differentiable

vector spaces DVS, we can define a smooth prefactorization algebra as an enriched
multifunctor (or map of enriched colored operads). This definition captures a sense
in which we have smooth families of structure maps. Our primary example — the
factorization envelope — provides such a smooth prefactorization algebra because
Ω∗ (M) on a fixed manifold M defines a differentiable cochain complex.
3.2. Other kinds of geometry. This kind of construction works very gener-
ally. For instance, if we are interested in complex manifolds, we could work in the
following setting.
3.2.1 Definition. Let Holn denote the category whose objects are complex n-manifolds
and whose morphisms are open holomorphic embeddings. It possesses a symmetric
monoidal structure under disjoint union.
3.2.2 Definition. A prefactorization algebra on complex n-manifolds with values

in a symmetric monoidal category C is a symmetric monoidal functor from Holn to
C.
Again, there is a natural example arising from a factorization envelope con-

struction. Compactly supported Dolbeault forms Ω0,∗
c provide a functor from Holn
to nonunital commutative dg algebras. Hence, tensoring with a dg Lie algebra g
provides a symmetric monoidal functor from Holn to the category of dg Lie alge-
bras, equipped with direct sum of underlying cochain complexes as a symmetric
0,∗
monoidal structure. Thus, the factorization envelope Uhol
n g : M 7→ C ∗ (Ωc (M) ⊗ g)
provides a prefactorization algebra on complex n-manifolds.
The Kac-Moody factorization algebras (see example 6.3) provide an example
on Hol1 (i.e., for Riemann surfaces). In Williams (n.d.), Williams constructs the
Virasoro factorization algebras on Hol1 . These examples extend naturally to Holn
for any n, although the choices of central extension become more interesting and
complicated. (Indeed, one finds extensions as L∞ algebras and not just dg Lie
algebras.)
It should be clear that one can enrich this category Holn over Sh(Mfld) and
hence to talk about smooth families of such holomorphic embeddings. (Or, indeed,
one can enrich it over a site of all complex manifolds.) Thus, one can talk, for
instance, about smooth prefactorization algebras on complex n-manifolds.
In general, let G denote some kind of local structure for n-manifolds, such
a Riemannian metric or complex structure or orientation. In other words, G is
a sheaf on Embn . A G-structure on an n-manifold M is a section G ∈ G(M).
There is a category EmbG whose objects are n-manifolds with G-structure (M, G M )
and whose morphisms are G-structure-preserving embeddings, i.e., embeddings
f : M ,→ N such that f ∗G N = G M . This category is fibered over EmbG . One can
then talk about prefactorization algebras on G-manifolds.
3.3. Application to field theory. This notion is quite useful in the context of
field theory. One often studies a field theory that makes sense on a large class of
manifolds. Indeed, physicists often search for action functionals that depend only
on a particular geometric structure. For instance, a conformal field theory only
depends on a conformal class of metric, and so the solutions to the equations of
motion form a sheaf on the site of conformal manifolds. For example, the free
βγ system makes sense on all conformal 2-manifolds, indeed, its solutions form
a sheaf on Hol1 . Similarly, the physicist’s examples of a topological field theory
only depend on the underlying smooth manifold (at least at the classical level). The
solutions to the equations of motion for Abelian Chern-Simons theory (see Section
5) make sense on the site of oriented 3-manifolds. (More generally, Chern-Simons
theory for gauge group G makes sense on the site of oriented 3-manifolds with a
principal G-bundle.) Finally, the free scalar field theory with mass m makes sense
on the site of Riemannian n-manifolds.
Since most of the constructions with field theories in this book have exploited
the enveloping factorization algebra construction, the classical and quantum ob-
servables we have constructed are typically prefactorization algebras on G-manifolds
for some G. This universality of these quantizations — that these prefactorization
algebras of quantum observables simultaneously work for all manifolds with some
geometric structure — illuminates why our constructions typically recover stan-
dard answers: these are the answers that work generally.
For interacting theories, the focus of Volume 2, the quantizations are rarely so
easily obtained or described. On a fixed manifold, a quantization of the prefactor-
ization algebra of classical observables might exist, but it might not exist for all
manifolds with that structure. More explicitly, quantization proceeds via Feynman
3. VARIANT DEFINITIONS OF FACTORIZATION ALGEBRAS 179
diagrammatics and renormalization, and hence it involves some explicit analysis

(e.g., the introduction of counterterms). It is often difficult to quantize in a way
that varies nicely over the collection of all G-manifolds. There are situations where
quantizations exist over some category of G-manifolds, however. For instance, Li
and Li (2016) construct a version of the topological B-model in this formalism and
show that a quantization exists for all smooth oriented 2-manifolds after choosing
a holomorphic volume form on the target complex manifold X.
3.4. Descent axioms. So far we have only discussed a variant of the notion of
prefactorization algebra. We now turn to the local-to-global axiom in this context.
3.4.1 Definition. A Weiss cover of a G-manifold M is a collection of G-embeddings
{φi : Ui → M}i∈I such that for any finite set of points x1 , . . . , xn ∈ M, there is some
i such that {x1 , . . . , xn } ⊂ φi (Ui ).
With this definition in hand, we can formulate the natural generalization of our
earlier definition.
3.4.2 Definition. A factorization algebra on G-manifolds is a symmetric monoidal
functor F : EmbG → Ch(DVS) that is a homotopy cosheaf in the Weiss topology.
Let’s return to our running example. Our arguments in Section 6 show that
the factorization envelope of a dg Lie algebra using the de Rham complex is a
factorization algebra on Embn , and similarly the factorization enveloped using the
Dolbeault complex is a factorization algebra on Holn .
Remark: These definitions, indeed the spirit of this variation on factorization alge-
bras, are inspired by conversations with and the work of John Francis and David
Ayala. See Ayala and Francis (2015) for a wonderful overview of factorization
algebras in the setting of topological manifolds, along with deep theorems about
them. We note that they use the term “factorization homology,” rather than fac-
torization algebra, as they view the construction as a symmetric monoidal analog,
for n-manifolds, of ordinary homology. In this picture, the value on a disc defines
the coefficients for a homology theory on n-manifolds. (We discuss factorization
homology in the next section.) In further work Ayala et al. (n.d.b), they have devel-
oped a generalization of factorization techniques where the coefficients are (∞, n)-
categories. One payoff of their results is a proof of the Cobordism Hypothesis,
and hence an explanation of functorial topological field theories via factorization-
theoretic thinking. (See Lurie (2009b) for the canonical reference on the Cobor-
dism Hypothesis.) ^
Remark: The quantizations mentioned above produce factorization algebras and
not just prefactorization algebras. Hence, their values on big manifolds can be
computed from their values on smaller manifolds, via the local-to-global axiom. It
is interesting to compare this approach to understanding the global behavior of a
quantum field theory — by a kind of open cover — with the approach in functo-
rial field theory, in the style of Atiyah-Segal-Lurie, where a manifold is soldered
together from manifolds with corners. The Ayala-Francis work shows that the
factorization approach undergirds the cobordism approach, for topological field

theories; the Cobordism Hypothesis is a consequence of their approach. Moving
in a geometric direction, Dwyer, Stolz, and Teichner have suggested a method of
using a factorization algebra on G-manifolds to build a functorial field theory on a
category of G-cobordisms, appropriately understood. ^
4. Locally constant factorization algebras

There is a family of operads that plays an important role in algebraic topol-
ogy, known as the “little n-dimensional discs” operads, that were introduced by
Boardman and Vogt (1973). The little n-discs operad is the operad in topologi-
cal spaces whose k-ary operations is the space of embeddings of a disjoint union
of k n-dimensional discs into a single n-dimensional disc via radial contraction
and translation. (In other words, the embedding is completely specified by saying
where the center of each disc goes and what its radius is.) We will use the term En
operad for any topological operad weakly equivalent to the little n-discs, follow-
ing a standard convention. An algebra over an En operad is called an En algebra.
It has, by definition, families of k-fold multiplication maps parametrized by the
configurations of k n-dimensional discs inside a single disc.
If M is an n-dimensional manifold, then prefactorization algebras on M locally
bear a strong resemblance to En algebras. After all, a prefactorization algebra
prescribes a way to combine an element sitting on each of k distinct discs into
an element for a big disc containing all k discs. In fact, En algebras form a full
subcategory of factorization algebras on Rn , as we now indicate.
4.0.1 Definition. A factorization algebra F on an n-manifold M is locally constant
if for each inclusion of open discs D ⊂ D0 , then the map F (D) → F (D0 ) is a quasi-
isomorphism.
A central example is the locally constant factorization algebra A f act on R given

by an associative algebra A. (Recall example 1.1.)
Lurie has shown the following vast extension of this example. (See section
5.4.5 of Lurie (n.d.b), particularly Theorem 5.4.5.9.)
4.0.2 Theorem. There is an equivalence of (∞, 1)-categories between En algebras
and locally constant factorization algebras on Rn .
We remark that Lurie (and others) uses a different gluing axiom than we do.
A careful comparison of the different axioms and a proof of their equivalence (for
locally constant factorization algebras) can be found in Matsuoka (n.d.). We em-
phasize that higher categories are necessary for a rigorous development of these
constructions and comparisons, although we will not emphasize that point in our
discussion below, which is expository in nature.
Remark: This theorem, together with our work in these two books, builds an ex-
plicit bridge between topology and physics. Topological field theories in the physi-
cist’s sense, such as perturbative Chern-Simons theory, produce locally constant
factorization algebras, which can then be viewed as En algebras and analyzed using
4. LOCALLY CONSTANT FACTORIZATION ALGEBRAS 181
the powerful machinery of modern homotopy theory. Conversely, intuition from

physics about the behavior of topological field theories suggests novel construc-
tions and examples to topology. Much of the amazing resonance between operads,
homological algebra, and string-theoretic physics over the last few decades can be
understood from this perspective. ^
4.1. Motivation from topology. Many examples of En algebras arise natu-
rally from topology, such as labelled configuration spaces (as discussed in the work
of Segal Segal (1973), McDuff McDuff (1975), Bödigheimer Bödigheimer (1987),
Salvatore Salvatore (2001), and Lurie Lurie (n.d.b)). We will discuss an important
example, that of mapping spaces.
Recall that (in the appropriate category of spaces) there is an isomorphism
Maps(U t V, X) Maps(U, X) × Maps(V, X). This fact suggests that we might fix
a target space X and define a prefactorization algebra by sending an open set U to
Maps(U, X). This construction almost works, but it is not clear how to “extend” a
map f : U → X from U to a larger open set V ⊃ U. By working with “compactly-
supported” maps, we solve this issue.
Fix (X, p) a pointed space. Let F denote the prefactorization algebra on M
sending an open set U to the space of compactly-supported maps from U to (X, p).
(Here, “ f is compactly-supported” means that the closure of f −1 (X−p) is compact.)
Then F is a prefactorization algebra in the category of pointed spaces. (Composing
with the singular chains functor gives a prefactorization algebra in chain complexes
of abelian groups, but we will work at the level of spaces.)
Let’s consider for a moment the case when the open set is an open disc D ⊂
Rn = M. There is a weak homotopy equivalence
F(D) ' Ωnp X
between the space of compactly supported maps D → X and the n-fold based loop
space of X, based at the point p. (We are using the topologist’s notation Ωnp X for the
n-fold loop space; hopefully, this use does not confuse due to the standard notation
for the space of n-forms.)
To see this equivalence, note that a compactly supported map f : D → X
extends uniquely to a map from the closed D, sending the boundary ∂D to the base
point p of X. Since Ωnp X is defined to be the space of maps of pairs (D, ∂D) →
(X, p), we have constructed the desired map from F(D) to Ωnp X. It is easily verified
that this map is a homotopy equivalence.
Note that this prefactorization algebra is locally constant: if D ,→ D0 is an
inclusion of open discs, then the map F(D) → F(D0 ) is a weak homotopy equiva-
lence.
Note also that there is a natural isomorphism
F(U1 t U2 ) = F(U1 ) × F(U2 )
if U1 , U2 are disjoint opens. If D1 , D2 are disjoint discs contained in a disc D3 , the
prefactorization structure gives us a map
F(D1 ) × F(D2 ) → F(D3 ).
These maps correspond to the standard En structure on the n-fold loop space.
For a particularly nice example, consider the case where the source manifold
is M = R. The structure maps of F then describe the standard product on the space
Ω p X of based loops in X. At the level of components, we recover the standard
product on π0 Ω p X = π1 (X, p).
This prefactorization algebra F does not always satisfy the gluing axiom. How-
ever, Salvatore Salvatore (2001) and Lurie Lurie (n.d.b) have shown that if X is
sufficiently connected, this prefactorization algebra is in fact a factorization alge-
bra.
4.2. Relationship with Hochschild homology. The direct relationship be-
tween En algebras and locally constant factorization algebras on Rn raises the
question of what the local-to-global axiom means from an algebraic point of view.
Evaluating a locally constant factorization algebra on a manifold M is known as
factorization homology or topological chiral homology.
The following result, for n = 1, is striking and helpful.
4.2.1 Theorem. For A an E1 algebra (e.g., an associative algebra), there is a weak
equivalence
A f act (S 1 ) ' HH∗ (A),
where A f act denotes the locally constant factorization algebra on R associated to
A and HH∗ (A) denotes the Hochschild homology of A.
Here HH∗ (A) ' A ⊗LA⊗A A means any cochain complex quasi-isomorphic to
the usual bar complex (i.e., we are interested in more than the mere cohomology
groups). In particular, this theorem says that the cohomology of the global sections
on S 1 of A f act equals the Hochschild homology of the E1 algebra A.
This result is one of the primary motivations for the higher dimensional gener-
alizations of factorization homology. It has several proofs in the literature, depend-
ing on choice of gluing axiom and level of generality (for instance, one can work
with algebra objects in more general ∞-categories). See Theorem 3.9 of Ayala and
Francis (2015) or Theorem 5.5.3.11 of Lurie (n.d.b); a barehanded proof for dg
algebras can be found in Section 4.3 of Gwilliam (2012).
There is a generalization of this result even in dimension 1. Note that there
are more general ways to extend A f act from the real line to the circle, by allowing
“monodromy.” Let σ denote an automorphism of A. Pick an orientation of S 1 and
fix a point p in S 1 . Let (A f act )σ denote the prefactorization algebra on S 1 such
that on S 1 − {p} it agrees with A f act but where the structure maps across p use the
automorphism σ. For instance, if L is a small interval to the left of p, R is a small
interval to the right of p, and M is an interval containing both L and R, then the
structure map is
A⊗A → A
a ⊗ b 7→ a ⊗ σ(b)
where the leftmost copy of A corresponds to L and so on. It is natural to view the
copy of A associated to an interval containing p as the A − A bimodule Aσ where
A acts as the left by multiplication and the right by σ-twisted multiplication.
4. LOCALLY CONSTANT FACTORIZATION ALGEBRAS 183
4.2.2 Theorem. There is a weak equivalence (A f act )σ (S 1 ) ' HH∗ (A, Aσ ).

Remark: In Section 1.2, we use this theorem to analyze the factorization homology
of the quantum observables of the massive free scalar field on the circle. ^
Beyond the one-dimensional setting, some of the most useful insights into the
meaning of factorization homology arise from its connection to the cobordism
hypothesis and fully extended topological field theories. See section 4 of Lurie
(2009b) for an overview of these ideas, and Scheimbauer (n.d.) for development
and detailed proofs. As mentioned in Remark 3.4, Ayala et al. (n.d.b) extends fac-
torization homology by allowing coefficients in (∞, n)-categories, which leads to
richer invariants. There is closely related work by Morrison and Walker (2012)
using the formalism of blob homology.
4.3. Factorization envelopes. The connection between locally constant fac-

torization algebras and En algebras also provides a beautiful universal property to
the factorization envelope of a dg Lie algebra g,
Un g : M 7→ C∗ (Ω∗c (M) ⊗ g),
where M runs over the category Embn of smooth n-manifolds and embeddings.
This enveloping construction provides a functor from dg Lie algebras to En alge-
bras in the category of cochain complexes over a characteristic zero field. (Here
we view Un g as an En algebra, by Lurie’s theorem.) On the other hand, there is
a “forgetful” functor from this category of En algebras in cochain complexes to
the category of dg Lie algebras. For n = 1, this functor is essentially the familiar
functor from dg associative algebras to dg Lie algebras given by only remembering
the commutator bracket and not the full multiplication. For n > 1, such a functor
exists due to the formality of the En operad: as an operad in cochain complexes,
En ' H ∗ (En ) = Pn , and there is clearly a forgetful functor from Pn algebras to dg
Lie algebras.
The factorization envelope provides the left adjoint for this forgetful functor:
for any map of Lie algebras g → A, where A is an En algebra viewed as simply a
Lie algebra, then there is a unique map Un g → A of En algebras, and conversely. To
actually state this kind of result, we need to use ∞-categories, of course. Knudsen
(n.d.) has provided a proof of this universal property using factorization techniques.
4.3.1 Theorem. Let k be a field of characteristic zero. Let C be a stable, k-linear,
presentably symmetric monoidal ∞-category. There is an adjunction
Un : AlgLie (C) AlgEn (C) : F
such that for any X ∈ C, FreeEn X ' Un FreeLie (Σn−1 X).

Here FreeLie : C → AlgLie (C) denotes the free Lie algebra functor adjoint
to the forgetful functor AlgLie (C) → C, and likewise FreeEn : C → AlgLie (C)
denotes the free En algebra functor adjoint to the forgetful functor AlgEn (C) → C.
As is conventional in topology, and hence in much work on stable ∞-categories,
Σ : C → C denotes the “suspension functor” sending X to the pushout 0

`
X 0. In
the setting of cochain complexes, suspension is given by a shift.
4.4. References. For a deeper discussion of factorization homology than we’ve

given, we recommend Ayala and Francis (2015) to start, as it combines a clear
overview with a wealth of applications. A lovely expository account is Ginot
(2015). Locally constant factorization algebras already possess a substantial lit-
erature, as they sit at the nexus of manifold topology and higher algebra. See, for
instance, Ayala et al. (n.d.a), Ayala et al. (n.d.b), Francis (2013), Ginot et al. (n.d.b),
Ginot et al. (n.d.a), Horel (n.d.), Lurie (n.d.b), Matsuoka (n.d.), and Morrison and
Walker (2012). Some striking applications to representation theory, particularly
quantum character varieties, can be found in Ben Zvi et al. (n.d.).
5. Factorization algebras from cosheaves

The goal of this section is to describe a natural class of factorization algebras.
The factorization algebras that we construct from classical and quantum field the-
ory will be closely related to the factorization algebras discussed here.
The main result of this section is that, given a nice cosheaf of vector spaces or
cochain complexes F on a manifold M, the functor
Sym F : U 7→ Sym(F(U))
is a factorization algebra. It is clear how this functor is a prefactorization algebra
(see example 1.1); the hard part is verifying that it satisfies the local-to-global
axiom. The examples in which we are ultimately interested arise from cosheaves F
that are compactly supported sections of a vector bundle, so we focus on cosheaves
of this form.
We begin by providing the definitions necessary to state the main result of this
section. We then state the main result and explain its role for the rest of the book.
Finally, we prove the lemmas that culminate in the proof of the main result.
5.1. Preliminary definitions.

5.1.1 Definition. A local cochain complex on M is a graded vector bundle E on
M (with finite rank), whose smooth sections will be denoted by E , equipped with a
differential operator d : E → E of cohomological degree 1 satisfying d2 = 0.
Recall the notation from Section 5. For E be a local cochain complex on M

and U an open subset of M, we use E (U) to denote the cochain complex of smooth
sections of E on U, and we use Ec (U) to denote the cochain complex of compactly
supported sections of E on U. Similarly, let E (U) denote the distributional sections
on U and let E c (U) denote the compactly supported distributional sections of E on
U. In Appendix B, it is shown that these four cochain complexes are differentiable
cochain complexes in a natural way.
We use E ! = E ⊗ Dens M to denote the appropriate dual object. We give E ! the
differential that is the formal adjoint to d on E. Note that, ignoring the differential,
5. FACTORIZATION ALGEBRAS FROM COSHEAVES 185
E c (U) is the continuous dual to E ! (U) and that Ec (U) is the continuous dual to
!
E (U).
5.1.2 Lemma. A local cochain complex (E , d) is a sheaf of differentiable cochain
complexes. In fact, it is also a homotopy sheaf. Similarly, compactly supported
sections (Ec , d) is a cosheaf of differentiable cochain complexes, as well as a ho-
motopy cosheaf.
Proof. We describe the cosheaf case as the sheaf case is parallel.
In Section 7.2, we showed that precosheaf Eck given the degree k vector space
is a cosheaf with values in DVS. Thus we know that Ec , without the differential, is
a cosheaf of graded differentiable vector spaces. As colimits of cochain complexes
(note: not homotopy hocolimits) are computed degreewise, we see that (Ec , Q) is a
cosheaf of differentiable cochain complexes. Identical arguments apply to E c .
The homotopy cosheaf condition is only marginally more difficult. Suppose we
are interested in an open U and a cover U for U. Recall that the Čech complex is the
totalization of a double complex; we will view d as the vertical differential and the
Čech differential as horizontal, so that the double complex is concentrated in the
left half plane. We will work with the “augmented” double complex in which Ec (U)
is added as the degree 1 column. The contracting cochain homotopy constructed in
Section 7.2 applies to the cosheaf Eck of degree k elements. Hence this augmented
double complex is acyclic in the horizontal direction. The usual “staircase” or
“zigzag” then ensures that the whole double complex is quasi-isomorphic to the
rightmost column.
The factorization algebras we will discuss are constructed from the symmetric
! !
algebra on the vector spaces Ec (U) and E c (U). Note that, since E c (U) is dual to
d E !c (U) as the algebra of formal power series on E (U).
E (U), we can view Sym
Thus, we often write
d E !c (U) = O(E (U)),
Sym
because we view this algebra as the space of “functions on E (U).”
Note that if U → V is an inclusion of open sets in M, then there are natural
maps of commutative dg algebras
Sym Ec (U) → Sym Ec (V),
d Ec (U) → Sym
Sym d Ec (V),
Sym E c (U) → Sym E c (V),
d E c (U) → Sym
Sym d E c (V).
Thus, each of these symmetric algebras forms a precosheaf of commutative alge-

bras, and thus a prefactorization algebra. We denote these prefactorization algebras
by Sym Ec and so on.
5.2. The main theorem. The main result of this section is the following.
5.2.1 Theorem. We have the following parallel results for vector bundles and local
cochain complexes.
((i)) Let E be a vector bundle on M. Then
((a)) Sym Ec and Sym E c are strict (non-homotopical) factorization
algebras valued in the category of differentiable vector spaces,
and
d Ec and Sym
((b)) Sym d E c are strict (non-homotopical) factorization
algebras valued in the category of differentiable pro-vector spaces.
((ii)) Let E be a local cochain complex on M. Then
((a)) Sym Ec and Sym E c are homotopy factorization algebras valued
in the category of differentiable cochain complexes, and
d Ec and Sym
((b)) Sym d E c are homotopy factorization algebras valued
in the category of differentiable pro-cochain complexes.
Proof. Let us first prove the strict version of the result. To start with, consider
the case of Sym Ec . We need to verify the local-to-global axiom.
Let U be an open set in M and let U = {Ui | i ∈ I} be a Weiss cover of U. We
need to prove that Sym∗ Ec (U) is the cokernel of the map
M M
Sym(Ec (Ui ∩ U j )) → Sym(Ec (Ui ))
i, j∈I i∈I
that sends a section on Ui ∩U j to its extension to a section of Ui minus its extension

to U j . This map preserves the decomposition of Sym Ec (U) into symmetric powers:
a section of Symk (Ec (Ui ∩ U j )) extends to a section of Symk (Ec (Ui )). Thus, it
suffices to show that

Symm Ec (U) = coker ⊕i, j∈I Symm (Ec (Ui ∩ U j )) → ⊕i∈I Symm Ec (Ui ) ,
for all m.
Now, observe that
Ec (U)b⊗π m = Ecm (U m )
where Ecm is the cosheaf on U m of compactly supported smooth sections of the
vector bundle E m , which denotes the external product of E with itself m times.
Thus it is enough to show that
 
M M 
Ecm (U m ) = coker  Ecm (Ui ∩ U j )m → Ecm Uim  .
i, j∈I i∈I
Our cover U is a Weiss cover. This means that, for every finite set of points
x1 , . . . , xm ∈ U, we can find an open Ui in the cover U containing every x j . This
implies that the subsets of U m of the form (Ui )m , where i ∈ I, cover U m . Further,
(Ui )m ∩ (U j )m = (Ui ∩ U j )m .
The desired isomorphism now follows from the fact that Ecm is a cosheaf on M m .
The same argument applies to show that Sym Ec! is a factorization algebra.
In the completed case, essentially the same argument applies, with the subtlety
6. FACTORIZATION ALGEBRAS FROM LOCAL LIE ALGEBRAS 187
(see Section 4) that, when working with pro-cochain complexes, the direct sum is
completed.
For the homotopy case, the argument is similar. Let U = {Ui | i ∈ I} be a
Weiss cover of an open subset U of M. Let F = Sym Ec denote the prefactoriza-
tion algebras we are considering (the argument below will apply when we use the
completed symmetric product or use E c instead of Ec ). We need to show that the
map
Č(U, F ) → F (U)
is an equivalence, where the left hand side is equipped with the standard Čech
differential.
Let F m (U) = Symm Ec . Both sides above split as a direct sum over m, and the
map is compatible with this splitting. (If we use the completed symmetric product,
this decomposition is as a product rather than a sum.)
We thus need to show that the map
M
Symm Ec Ui1 ∩ · · · ∩ Uin [n − 1] → Symm (Ec (U))

i1 ,...,in
is a weak equivalence.
For i ∈ I, we get an open subset Uim ⊂ U m . Since U is a Weiss cover of U,
these open subsets form a cover of U m . Note that
Uim1 ∩ · · · ∩ Uimn = Ui1 ∩ · · · ∩ Uin m .

Note that Ec (U)⊗m can be naturally identified with Γc (U m , E m ) (where the tensor
product is the completed projective tensor product).
Thus, to show that the Čech descent axiom holds, we need to verify that the
map M
Γc Uim1 ∩ · · · ∩ Uimn , E m [n − 1] → Γc (V m , E m )
i1 ,...,in ∈I
is a quasi-isomorphism. The left hand side above is the Čech complex for the
cosheaf of compactly supported sections of E m on V m . Standard partition of unity
arguments show that this map is a weak equivalence. (See Section 4.4 for discus-
sion and references.)
6. Factorization algebras from local Lie algebras

We just showed that for a local cochain complex E, the prefactorization algebra
d Ec is, in fact, a factorization algebra. What this construction says is that the
Sym
“functions on E ,” viewed as a space, satisfy a locality condition on the manifold
M over which E lives. We can reconstruct functions on E (M) from knowing about
functions on E with very small support on M. But E is a simple kind of space, as
it is linear in nature. (We should remark that E is a simple kind of derived space
because it is a cochain complex.) We now extend to a certain type of nonlinear
situation.
In Section 6, we introduced the notion of a local dg Lie algebra. Since a lo-
cal cochain complex is an Abelian local dg Lie algebra, we might hope that the
Chevalley-Eilenberg cochain complex C ∗ L of a local Lie algebra L also forms

a factorization algebra. As we explain in Volume 2, a local dg Lie algebra can
be interpreted as a derived space that is nonlinear in nature. In this setting, the
Chevalley-Eilenberg cochain complex are the “functions” on this space. Hence,
if C ∗ L is a factorization algebra, we would know that functions on this nonlinear
space L can also be reconstructed from functions localized on the manifold M.
In Section 6, we also constructed an important class of prefactorization alge-
bras: the factorization envelope of a local dg Lie algebra. Again, in the preceding
section, we showed Sym Ec is a factorization algebra, which we can view as the fac-
torization envelope of the Abelian local Lie algebra E [−1]. Thus, we might expect
that the factorization envelope of a local Lie algebra satisfies the local-to-global
axiom.
We will now demonstrate that both these Lie-theoretic constructions are fac-
torization algebras.
6.0.1 Theorem. Let L be a local dg Lie algebra on a manifold M. Then the pref-
actorization algebras
UL :U 7→ C∗ (Lc (U))
OL :U 7→ C ∗ (L(U))
are factorization algebras.
Remark: The argument below applies, with very minor changes, to a local L∞ al-
gebra, a modest generalization we introduce later. ^
Proof. The proof is a spectral sequence argument, and we will reuse this idea
throughout the book (notably in proving that quantum observables form a factor-
ization algebra).
We start with the factorization envelope. Note that for any dg Lie algebra g,
the Chevalley-Eilenberg chains C∗ g have a natural filtration F n = Sym≤n (g[1])
compatible with the differential. The first page of the associated spectral sequence
is simply the cohomology of Sym(g[1]), where g is viewed as a cochain complex
rather than a Lie algebra (i.e., we extend the differential on g[1] as a coderivation
to the cocommutative coalgebra Sym(g[1])).
Consider the Čech complex of UL with respect to some Weiss cover U for an
open U in M. Applying the filtration above to each side of the map
Č(U, UL) → UL(U),
we get a map of spectral sequences. On the first page, this map is a quasi-isomorphism
by Theorem 5.2.1. Hence the original map of filtered complexes is a quasi-isomorphism.
Now we provide the analogous argument for OL. For any dg Lie algebra g, the
Chevalley-Eilenberg cochains C ∗ g have a natural filtration F n = Sym d ≥n (g∨ [−1])
compatible with the differential. The first page of the associated spectral sequence
d ∨ [−1]), where we view g∨ [−1] as a cochain
is simply the cohomology of Sym(g
complex and extend its differential as a derivation to the completed symmetric
algebra.
6. FACTORIZATION ALGEBRAS FROM LOCAL LIE ALGEBRAS 189
Consider the Čech complex of OL with respect to some Weiss cover U for an
open U in M. Applying the filtration above to each side of the map
Č(U, OL) → OL(U),
we get a map of spectral sequences. On the first page, this map is a quasi-isomorphism
by Theorem 5.2.1. Hence the original map of filtered complexes is a quasi-isomorphism.

CHAPTER 7
Formal aspects of factorization algebras
In this chapter, we describe some natural constructions with factorization alge-

bras. For example, we show how factorization algebras, like sheaves, push forward
along maps of spaces. We We also describe ways in which a factorization algebra
can be reconstructed from local data.
1. Pushing forward factorization algebras

A crucial feature of factorization algebras is that they push forward nicely.
Let M and N be topological spaces admitting Weiss covers and let f : M → N
be a continuous map. Given a Weiss cover U = {Uα } of an open U ⊂ N, let
f −1 U = { f −1 Uα } denote the preimage cover of f −1 U ⊂ M. Observe that f −1 U
is Weiss: given a finite collection of points {x1 , . . . , xn } in f −1 U, the image points
{ f (x1 ), . . . , f (xn )} are contained in some Uα in U and hence f −1 Uα contains the x j .
1.0.1 Definition. Given a factorization algebra F on a space M and a continuous
map f : M → N, the pushforward factorization algebra f∗ F on N is defined by
f∗ F (U) = F ( f −1 (U)).
Note that for the map to a point f : M → pt, the pushforward factorization
algebra f∗ F is simply the global sections of F . We also call this the factorization
homology of F on M.
2. Extension from a basis

Let X be a topological space, and let U be a basis for X, that is closed under
taking finite intersections. It is well-known that there is an equivalence of cate-
gories between sheaves on X and sheaves that are only defined for open sets in the
basis U. In this section we will prove a similar statement for factorization algebras.
In this section, we require topological spaces to be Hausdorff.
2.1. Preliminary definitions. We begin with a paired set of definitions.

2.1.1 Definition. A factorizing basis U = {Ui }i∈I for a space X is a basis for the
(usual) topology of X with the following properties:
((i)) for every finite set {x1 , . . . , xn } ⊂ X, there exists i ∈ I such that {x1 , . . . , xn } ⊂
Ui ;
((ii)) if Ui and U j are disjoint, then Ui ∪ U j is in U;
((iii)) Ui ∩ U j ∈ U for every Ui and U j in U.
191
192 7. FORMAL ASPECTS OF FACTORIZATION ALGEBRAS
As an example, for X = Rn , consider the collection of all opens that are a

disjoint union of finitely many convex open sets.
Note that a factorizing basis provides a Weiss cover for every open U. Given
any finite set {x1 , . . . , xn } ⊂ U, fix a collection of pairwise disjoint opens Vi ⊂ U
such that xi ∈ Vi . (Here we use that X is Hausdorff.) Then there is an element of
the basis Ui0 ⊂ Vi with xi ∈ Ui0 . The disjoint union U10 t · · · t Un0 ⊂ U is in U and
contains {x1 , . . . , xn } ⊂ U. Thus, the collection of opens in U contained in U form
a Weiss cover for U.
Let C denote a symmetric monoidal Grothendieck abelian category.
2.1.2 Definition. Given a factorizing basis U, a U-prefactorization algebra consists
of the following:
((i)) for every Ui ∈ U, an object F (Ui ) ∈ C;
((ii)) for every finite tuple of pairwise disjoint opens Ui0 , . . . , Uin all con-
tained in U j — with all these opens from U — a structure map
F (Ui0 ) ⊗ · · · ⊗ F (Uin ) → F (U j ));
((iii)) the structure maps are equivariant and associative.
In other words, a U-prefactorization algebra F is like a prefactorization algebra,
except that F (U) is only defined for sets U in U.
A U-factorization algebra is a U-prefactorization algebra with the property
that, for every U in U and every Weiss cover V of U consisting of open sets in
U, the natural map
'
Č(V, F ) −
→ F (U)
is a weak equivalence, where Č(V, F ) denotes the Čech complex described earlier
(Section 1).
Note that we have not required that F (U) ⊗ F (V) → F (U t V) is an equiva-

lence, so that we are focused here on factorization algebras without the multiplica-
tivity condition. The extension construction will rely on having a lax symmetric
monoidal functor from simplicial cochain complexes to cochain complexes via an
Eilenberg-Zilber-type shuffle map. That is, we want the classical formula to hold
in this setting. (Those familiar with higher categories will recognize that a similar
proof works with any symmetric monoidal ∞-category whose symmetric monoidal
product preserves sifted colimits.)
2.2. The statement. In this section we will show that any U-factorization al-
gebra on X extends to a factorization algebra on X. This extension is unique up to
quasi-isomorphism.
For an open V ⊂ X, we use UV to denote the collection of all the opens in U
contained inside V.
Let F be a U-factorization algebra. Let us define a prefactorization algebra
ext(F ) on X by
ext(F )(V) = Č(UV , F ),
2. EXTENSION FROM A BASIS 193
for each open V ⊂ X. Conversely, if G is a factorization algebra on X, let res(G)

denote its restriction to U-factorization algebra. Our main result here is then the
following.
2.2.1 Proposition. For any U-factorization algebra F , its extension ext(F ) is a
factorization algebra on X such that res(ext(F )) is naturally isomorphic to F .
Conversely, if G is a factorization algebra on X, then ext(res(G)) is naturally quasi-
isomorphic to G.
Thus, we will prove that there is a kind of equivalence of categories
res
+
PreFA(X, C)k PreFA(U, C)
ext
between the category PreFA(XC) of factorization algebras on X and the category

PreFA(U, C) of U-factorization algebras. The subtlety here is that ext(res(F )) will
not be isomorphic to F but just weakly equivalent.
Remark: With these definitions in hand, it should be clear why we can recover a
factorization algebra on X from just a factorization algebra on a factorizing basis U.
The first property of U ensures that the “atomic” structure maps factor through the
basis. (Compare with how an ordinary cover means that the “atomic” restriction
maps of a sheaf factor through the cover: every sufficiently small neighborhood of
any point is contained in some open in the cover.) The second property ensures
that we know how to multiply within the factorizing basis. In particular, we have
the maps
F (Ui0 ) ⊗ · · · ⊗ F (Uin ) → F (Ui0 ∪ · · · ∪ Uin )
and
F (Ui0 ∪ · · · ∪ Uin ) → F (Uk ),
for any opens in our basis. Associativity thus determines the multiplication map
F (Ui0 ) ⊗ · · · ⊗ F (Uin ) → F (Uk ).
In other words, any multiplication map factors through a unary structure map and a
multiplication map arising from disjoint union. Finally, the third property ensures
that we know F on the intersections that appear in the gluing condition. ^
2.3. The proof. Let U be a factorizing basis and F a U-factorization algebra.
We will construct a factorization algebra ext(F ) on X in several stages:
• we give the value of ext(F ) on every open V in X,
• we construct the structure maps and verify associativity and equivari-
ance under permutation of labels,
• we verify that ext(F ) satisfies the gluing axiom.
Finally, it will be manifest from our construction how to extend maps of U-factorization
algebras to maps of their extensions.
We use the following notations. Given a simplicial cochain complex A• (so
each An is a cochain complex), let C(A• ) denote the totalization of the double
complex obtained by taking the unnormalized cochains. Let C N (A• ) denote the
totalization of the double complex by taking the normalized cochains. Finally, let
shAB : C N (A• ) ⊗ C N (B• ) → C N (A• ⊗ B• )
denote the Eilenberg-Zilber shuffle map, which is a lax symmetric monoidal func-
tor from simplicial cochain complexes to cochain complexes. See Theorem 1.2.3
for some discussion.
2.3.1. Extending values. For an open V ⊂ X, recall UV denotes the Weiss
cover of V given by all the opens in U contained inside V. We then define
ext(F )(V) := Č(UV , F ).
This construction provides a precosheaf on X.
2.3.2. Extending structure maps. The Čech complex for the factorization glu-
ing axiom arises as the normalized cochain complex of a simplicial cochain com-
plex, as should be clear from its construction. We use Č(V, F )• to denote this
simplicial cochain complex, so that
Č(V, F ) = C N (Č(V, F )• ),
with F a factorization algebra and V a Weiss cover.
We construct the structure maps by using the simplicial cochain complexes
Č(V, F )• , as many properties are manifest at that level. For instance, the unary
maps ext(F )(V) → ext(F )(W) arising from inclusions V ,→ W are easy to un-
derstand: UV is a subset of UW and thus we get a map between every piece of the
simplicial cochain complex.
We now explain in detail the map
mVV 0 : ext(F )(V) ⊗ ext(F )(V 0 ) → ext(F )(V ∪ V 0 ),
where V ∩ V 0 = ∅. Note that knowing this map, we recover every other multiplica-
tion map
ext(F )(V) ⊗ ext(F )(V 0 ) / ext(F )(W)
6
mVV 0
)
ext(F )(V ∪ V 0 )
by postcomposing with a unary map.
The n-simplices Č(UV , F )n is the direct sum of the complexes
F (Ui0 ∩ · · · ∩ Uin ),
with each Uik in V. The n-simplices of the tensor product of simplicial cochain
complexes Č(UV , F )• ⊗Č(UV 0 , F )• are precisely the levelwise tensor product Č(UV , F )n ⊗
Č(UV 0 , F )n , which breaks down into a direct sum of terms
F (Ui0 ∩ · · · ∩ Uin ) ⊗ F (U j0 ∩ · · · ∩ U jn )
with the Uik ’s in V and the U jk ’s in V 0 . We need to define a map
mVV 0 ,n : Č(UV , F )n ⊗ Č(UV 0 , F )n → Č(UV∪V 0 , F )n
for every n, and we will express it in terms of these direct summands.
2. EXTENSION FROM A BASIS 195
Now
(Ui0 ∩ · · · ∩ Uin ) ∪ (U j0 ∩ · · · ∩ U jn ) = (Ui0 ∪ U j0 ) ∩ · · · ∩ (Uin ∪ U jn ).
Thus we have a given map

F (Ui0 ∩ · · · ∩ Uin ) ⊗ F (U j0 ∩ · · · ∩ U jn ) → F (Ui0 ∪ U j0 ) ∩ · · · ∩ (Uin ∪ U jn ) ,
because F is defined on a factorizing basis. The right hand term is one of the direct
summands for Č(UV∪V 0 , F )n . Summing over the direct summands on the left, we
obtain the desired levelwise map.
Finally, the composition
C N (Č(U ∩ V, F )• ) ⊗ C N (Č(UV 0 , F )• )
sh

C N Č(U ∩ V, F )• ⊗ Č(UV 0 , F )•
C N (mVV 0 ,• )

C N (Č(UV∪V 0 , F )• )
gives us mVV 0 .
A parallel argument works to construct the multiplication maps from n disjoint
opens to a bigger open.
The desired associativity and equivariance are clear at the level of the simplicial
cochain complexes Č(UV , F ), since they are inherited from F itself.
2.3.3. Verifying gluing. We have constructed ext(F ) as a prefactorization al-
gebra, but it remains to verify that it is a factorization algebra. Thus, our goal is
the following.
2.3.1 Proposition. The extension ext(F ) is a cosheaf with respect to the Weiss
topology. In particular, for every open subset W and every Weiss cover W, the
complex Č(W, ext(F )) is quasi-isomorphic to ext(F )(W).
As our gluing axiom is simply the axiom for cosheaf — but using a funny class
of covers — the standard refinement arguments about Čech homology apply. We
now spell this out.
For W an open subset of X and W = {W j } j∈J a Weiss cover of W, there are two
associated covers that we will use:
(a) UW = {Ui ⊂ W | i ∈ I} and
(b) UW = {Ui | ∃ j such that Ui ⊂ W j }.
The first is just the factorizing basis of W induced by U, but the second consists of
the opens in U subordinate to the cover W. Both are Weiss covers of W.
To prove the proposition, we break the argument into two steps and exploit the
intermediary Weiss cover UW .
2.3.2 Lemma. There is a natural quasi-isomorphism

'
f : Č(W, ext(F )) → Č(UW , F ),
for every Weiss cover W of an open W.
Proof of lemma. This argument boils down to combinatorics with the covers.
Some extra notation will clarify what’s going on. In the Čech complex for the
cover W, for instance, we run over n + 1-fold intersections W j0 ∩ · · · ∩ W jn . We will
denote this open by W~j , where ~j = ( j0 , . . . , jn ) ∈ J n+1 , with J the index set for W.
Since we are only interested in intersections for which all the indices are pairwise
n+1 denote this subset of J n+1 .
distinct, we let Jd
First, we must exhibit the desired map f of cochain complexes. The source
complex Č(W, ext(F )) is constructed out of F ’s behavior on the opens Ui . Explic-
itly, we have MM
Č(W, ext(F )) = Č(UW~j , F )[n].
n≥0 ~j∈ Jd
n+1
Note that each term F (U~i ) appearing in this source complex appears only once
in the target complex Č(UW , F ). Let f send each such term F (U~i ) to its unique
image in the target complex via the identity. This map f is a cochain map: it clearly
respects the internal differential of each term F (U~i ), and it is compatible with the
Čech differential by construction.
Second, we need to show f is a quasi-isomorphism. We will show this by
imposing a filtration on the map and showing the induced spectral sequence is a
quasi-isomorphism on the first page.
We filter the target complex Č(UW , F ) by
MM
F nČ(UW , F ) := F (U~i )[k].
k≤n ~i∈Id
k+1
Equip the source complex Č(W, ext(F )) with a filtration by pulling this filtration
back along f . In particular, for any U~i = Ui0 ∩ · · · ∩ Uin , the preimage under f
consists of a direct sum over all tuples W~j = W j0 ∩ · · · ∩ W jm such that every Uik
is a subset of W~j . Here, m can be any nonnegative integer (in particular, it can be
bigger than n).
Consider the associated graded complexes with respect to these filtrations. The
source complex has
 
 
MM M
Gr Č(UW , F ) = C[m] .
 
F (U~i )[n] ⊗ 
n≥0 ~i∈Id ~j∈ Jd
 
n+1 m+1 such that U ⊂W ∀k
ik ~j
The rightmost term (after the tensor product) corresponds to the chain complex for
a simplex — here, the simplex is infinite-dimensional — and hence is contractible.
In consequence, the map of spectral sequences is a quasi-isomorphism.
We now wish to relate the Čech complex on the intermediary UW to that on UW .
3. PULLING BACK ALONG AN OPEN IMMERSION 197
2.3.3 Lemma. The complexes Č(UW , F ) and Č(UW , F ) are quasi-isomorphic.

Proof of lemma. We will produce a roof
' '
Č(UW , F ) ← Č(UW , extW (F )) → Č(UW , F ).
Recall that UW is a factorizing basis for W. Then F , restricted to W, is a UW -
factorizing basis. Hence, for any V ∈ UW and any Weiss cover V ⊂ UW of V, we
have
'
Č(V, F ) → F (V).
For any V ∈ UW , let UW |V = {Ui ∈ U | Ui ⊂ V ∩ W j for some j ∈ J}. Note that this
is a Weiss cover for V. We define
extW (F )(V) := Č(UW |V , F ).
By construction, the natural map
(2.3.1) extW (F )(V) → F (V)
Thus we have a quasi-isomorphism
'
Č(UW , extW (F )) → Č(UW , F ),
by using the natural map (2.3.1) on each open in the Čech complex for UW .
The map
'
Č(UW , extW (F )) → Č(UW , F )
arises by mimicking the construction in the preceding lemma.
3. Pulling back along an open immersion

Factorization algebras do not pull back along an arbitrary continuous map, at
least not in a simple way. On the other hand, when U is an open subset of a space
X, a factorization algebra F on X clearly restricts to a factorization algebra on
U. We would like to generalize this observation to open immersions f : Y → X
and to have a pullback operation producing a factorization algebra f ∗ F on Y from
a factorization algebra on X. In particular, we would like to have pullbacks to
covering spaces.
For simplicity (and because it encompasses all the examples we care about),
let Y and X be smooth manifolds of the same definition and let f : Y → X be a
local diffeomorphism, i.e., for any y ∈ Y there is some open neighborhood Uy such
that f |Uy is a diffeomorphism from Uy to f (Uy ).
We also restrict to multiplicative factorization algebras, a restriction that will
be visible in the construction of the pullback construction f ∗ .
Equip X with a factorizing basis as follows. Fix a Riemannian metric on X and
let C = {Ci }i∈I be the collection of all strongly convex open sets in X. (A set U ⊂ M
is strongly convex if for any two points in U, the unique minimizing geodesic in M
between the points is wholly contained in U and no other geodesic connecting the
points in is U.) As every sufficiently small ball is strongly convex, C forms a basis
for the ordinary topology of X. Note that the intersection of two strongly convex
opens is another strongly convex open. Let C denote the Weiss cover generated by
C, which is thus a factorizing basis for X.
There is a closely associated basis on Y. Let C0 denote the collection of all
connected opens U in Y such that f (U) is in C. Let C0 be the factorizing basis of Y
generated by C0 .
Given F a multiplicative factorization algebra on X, let FC denote the asso-
ciated multiplicative factorization algebra on the basis C. We now define a multi-
plicative factorization algebra f ∗ FC0 on the basis C0 as follows. To each U ∈ C0 ,
let f ∗ FC0 (U) = F ( f (U)). To a finite collection of disjoint opens V1 , . . . , Vn in C0 ,
we assign
f ∗ FC0 (V1 t · · · t Vn ) = f ∗ FC0 (V1 ) ⊗ · · · ⊗ f ∗ FC0 (Vn ),
the tensor product of the values on those components, by multiplicativity. For U in
C0 (and thus convex) and V in C0 such that V ⊂ U, equip f ∗ FC0 with the obvious
structure map
f ∗ FC0 (V) = F (V) → F (U) = f ∗ FC0 (U).
By multiplicativity, this determines the structure map for any inclusion in the fac-
torizing basis C0
t j V j ⊂ ti Ui ,
where each Ui ∈ C0 . This inclusion is the disjoint union over i of the inclusions
Ui ∩ (t j V j ) ⊂ Ui ,
and we have already specified the structure map for each i. Let f ∗ F denote the
factorization algebra on Y obtained by extending f ∗ FC0 from the basis C0 .
Note that although we used a choice of Riemannian metric in making the con-
struction, the factorization algebra f ∗ F is unique up to isomorphism. It is enough
to exhibit a map locally and the pullbacks are locally isomorphic on Y.
4. Descent along a torsor

Let G be a discrete group acting on a space X. Recall the definition of discretely
G-equivariant prefactorization algebra (see Definition 7.1.1).
4.0.1 Proposition. Let G be a discrete group acting properly discontinuously on
X, so that X → X/G is a principal G-bundle. Then there is an equivalence of
categories between G-equivariant multiplicative factorization algebras on X and
multiplicative factorization algebras on X/G.
Proof. If F is a factorization algebra on X/G, then f ∗ F is a G-equivariant
factorization algebra on X.
Conversely, let F be a G-equivariant factorization algebra on F . Let Ucon
be the open cover of X/G consisting of those connected sets where the G-bundle
X → X/G admits a section. Let U denote the factorizing basis for X/G generated
by Ucon . We will define a U-factorization algebra F G by defining
F G (U) = F (σ(U)),
4. DESCENT ALONG A TORSOR 199
where σ is any section of the G-bundle π−1 (U) → U.

Because F is G-equivariant, F (σ(U)) is independent of the section σ chosen.
Since U is a factorizing basis, F G extends canonically to a factorization algebra on
X/G.
CHAPTER 8
Factorization algebras: examples
1. Some examples of computations

The examples at the heart of this book have an appealing aspect: for these
examples, it is straightforward to compute global sections – the factorization ho-
mology – because we do not need to use the gluing axiom directly. Instead, we can
use the theorems in the preceding sections to compute global sections via analysis.
By analogy, consider how de Rham cohomology, which exploits partitions of unity
and analysis, compares to the derived functor of global sections for the constant
sheaf.
In this section, we compute the global sections of several examples we’ve al-
ready studied in the preceding chapters.
1.1. Enveloping algebras. Let g be a Lie algebra and let gR denote the local
Lie algebra Ω∗R ⊗ g on the real line R. Recall Proposition 4.0.1 which showed that
the factorization envelope UgR recovers the universal enveloping algebra Ug.
This factorization algebra UgR is also defined on the circle.
1.1.1 Proposition. There is a weak equivalence
UgR (S 1 ) ' C∗ (g, Ugad ) ' HH∗ (Ug),

where in the middle we mean the Lie algebra homology complex for Ug as a g-
module via the adjoint action
v · a = va − av,
where v ∈ g and a ∈ Ug and va denotes multiplication in Ug.
The second equivalence is a standard fact about Hochschild homology (see,
e.g., Loday (1998)). It arises by constructing the natural map from the Chevalley-
Eilenberg complex to the Hochschild complex, which is a quasi-isomorphism be-
cause the filtration (inherited from the tensor algebra) induces a map of spectral
sequences that is an isomorphism on the E1 page.
Note one interesting consequence of this theorem: the structure map for an in-
terval mapping into the whole circle corresponds to the trace map Ug → Ug/[Ug, Ug].
Proof. The wonderful fact here is that we do not need to pick a Weiss cover
and work with the Čech complex. Instead, we simply need to examine the cochain
complex C∗ (Ω∗ (S 1 ) ⊗ g).
201
202 8. FACTORIZATION ALGEBRAS: EXAMPLES
One approach is to use the natural spectral sequence arising from the filtration
F k = Sym≤k . The E1 page is C∗ (g, Ug), and one must verify that there are no
further pages.
Alternatively, recall that the circle is formal, so Ω∗ (S 1 ) is quasi-isomorphic to
its cohomology H ∗ (S 1 ) as a dg algebra. Thus, we get a homotopy equivalence of
dg Lie algebras
Ω∗ (S 1 ) ⊗ g ' g ⊕ g[−1],
where, on the right, g acts by the (shifted) adjoint action on g[−1]. Now,
C∗ (g ⊕ g[−1]) C∗ (g, Sym g),
where Sym g obtains a g-module structure from the Chevalley-Eilenberg homology
complex. Direct computation verifies this action is precisely the adjoint action of g
on Ug (we use, of course, that Ug and Sym g are isomorphic as vector spaces).
1.2. The free scalar field in dimension 1. Recall from Section 3 that the
Weyl algebra is recovered from the factorization algebra of quantum observables
Obsq for the free scalar field on R. We know that the global sections of Obsq on
a circle should thus have some relationship with the Hochschild homology of the
Weyl algebra, but we will see that the relationship depends on the ratio of the mass
to the radius of the circle.
For simplicity, we restrict attention to the case where S 1 has a rotation-invariant
metric, radius r, and total length L = 2πr. Let S 1L denote this circle R/ZL. We also
work with C-linear observables, rather than R-linear.
1.2.1 Proposition. If the mass m = 0, then H k Obsq (S 1L ) is C[~] for k = 0, −1 and
vanishes for all other k.
If the mass m satisfies mL = in for some integer n, then H k Obsq (S 1L ) is C[~]
for k = 0, −2 and vanishes for all other k.
For all other values of mass m, H 0 Obsq (S 1L ) C[~] and all other cohomology
groups vanish.
Hochschild homology with monodromy (recall Theorem 4.2.2) provides an
explanation for this result. The equations of motion for the free scalar field locally
have a two-dimensional space of solutions, but on a circle the space of solutions
depends on the relationship between the mass and the length of the circle. In the
massless case, a constant function is always a solution, no matter the length. In
the massive case, there is either a two-dimensional space of solutions (for certain
imaginary masses, because our conventions) or a zero-dimensional space. Viewing
the space of local solutions as a local system, we have monodromy around the
circle, determined by the Hamiltonian.
When the monodromy is trivial (so mL ∈ iZ), we are simply computing the
Hochschild homology of the Weyl algebra. Otherwise, we have a nontrivial auto-
morphism of the Weyl algebra.
Note that this proposition demonstrates, in a very primitive example, that the
factorization homology can detect spectral properties of the Hamiltonian. The
proof will exhibit how the existence of solutions to the equations of motion (in
1. SOME EXAMPLES OF COMPUTATIONS 203
this case, a very simple ordinary differential equation) affects the factorization ho-
mology. It would be interesting to develop further this kind of relationship.
Proof. We directly compute the global sections, in terms of the analysis of the
local Lie algebra
2
!
∞ 1 4+m
L = C (S L ) −−−−→ C (S L )[−1] ⊕ C~
∞ 1
where ~ has cohomological degree 1 and the bracket is

Z
[φ , φ ] = ~
0 1
φ0 φ1
S 1L
with φk a smooth function with cohomological degree k = 0 or 1. We need to

compute C∗ (L).
Fourier analysis allows us to make this computation easily. We know that the
exponentials eikx/L , with k ∈ Z, form a topological basis for smooth functions on
S 1 and that
k2
!
(4 + m )e
2 ikx/L
= 2 +m e 2 ikx/L
.
L
Hence, for instance, when m2 L2 is not a square integer, it is easy to verify that the
complex
4+m2
C ∞ (S 1L ) −−−−→ C ∞ (S 1L )[−1]
has trivial cohomology. Thus, H ∗ L = C~ and so H ∗ (C∗ L) = C[~], where in the
Chevalley-Eilenberg complex ~ is shifted into degree 0.
When m = 0, we have
H ∗ L = Rx ⊕ Rξ ⊕ R~,
where x represents the constant function in degree 0 and ξ represents the constant
function in degree 1. The bracket on this Lie algebra is
[x, ξ] = ~L.
Hence H 0 (C∗ L) C[~] and H −1 (C
∗ L) C[~], and the remaining cohomology
groups are zero.
When m = in/L for some integer n, we see that the operator 4 + m2 has two-
dimensional kernel and cokernel, both spanned by e±inx/L . By a parallel argument
to the case with m = 0, we see H 0 (C∗ L) C[~] H −2 (C∗ L) and all other groups
are zero.
1.3. The Kac-Moody factorization algebras. Recall from example 6.3 and
Section 4 that there is a factorization algebra on a Riemann surface Σ associated to
every affine Kac-Moody Lie algebra. For g a simple Lie algebra with symmetric
invariant pairing h−, −ig , we have a shifted central extension of the local Lie algebra
Ω0,∗
Σ ⊗ g with shift Z
ω(α, β) = hα, ∂βig .
Σ
Let Uω gΣ denote the twisted factorization envelope for this extended local Lie al-
gebra.
1.3.1 Proposition. Let g denote the genus of a closed Riemann surface Σ. Then
H ∗ (Uσω g(Σ)) H ∗ (g, Ug⊗g )[c]
where c denotes a parameter of cohomological degree zero.

Proof. We need to understand the Lie algebra homology
C∗ (Ω0,∗ (Σ) ⊗ g ⊕ Cc),
whose differential has the form ∂ + dLie . (Note that in the Lie algebra, c has coho-
mological degree 1.)
Consider the filtration F k = Sym≤k by polynomial degree. The purely an-
alytic piece ∂ preserves polynomial degree while the Lie part dLie lowers poly-
nomial degree by 1. Thus the E1 page of the associated spectral sequence has
Sym(H ∗ (Σ, O) ⊗ g ⊕ Cc) as its underlying graded vector space. The differential
now depends purely on the Lie-theoretic aspects. Moreover, the differential on
further pages of the spectral sequence are zero.
In the untwisted case (where there is no extension), we find that the E1 page is
precisely H ∗ (g, Ug⊗g ), by computations directly analogous to those for Proposition
1.1.1.
In the twisted case, the only subtlety is to understand what happens to the
extension. Note that H 0 (Ω0,∗ (Σ) is spanned by the constant functions. The pairing
ω vanishes if a constant function is an input, so we know that the central extension
does not contribute to the differential on the E1 page.
An alternative proof is to use the fact the Ω0,∗ (Σ) is homotopy equivalent to
its cohomology as a dg algebra. Hence we can compute C∗ (H ∗ (Σ, O) ⊗ g ⊕ Cc)
instead.
We can use the ideas of sections 6 and 7 to understand the “correlation func-
tions” of this factorization algebra Uω gΣ . More precisely, given a structure map
Uω gΣ (V1 ) ⊗ · · · ⊗ Uω gΣ (Vn ) → Uω gΣ (Σ)
for some collection of disjoint opens V1 , . . . , Vn ⊂ Σ, we will provide a method for

describing the image of this structure map. It has the flavor of a Wick’s formula.
Let bOc denote the cohomology class of an element O of Uω gΣ . We want to
encode relations between cohomology classes.
−1
As Σ is closed, Hodge theory lets us construct an operator ∂ , which vanishes
on the harmonic functions and (0,1)-forms but provides an inverse to ∂ on the com-
plementary spaces. (We must make a choice of Riemannian metric, of course, to
do this.) Now consider an element a1 · · · ak of cohomological degree 0 in Uω gΣ (Σ),
where each a j lives in
Ω0,1 (Σ) ⊗ g ⊕ Cc.
1. SOME EXAMPLES OF COMPUTATIONS 205
Then we have
−1 −1
−1
(∂ + dLie ) (∂ a1 )a2 · · · ak = ∂(∂ a1 )a2 · · · ak + dLie (∂ a1 )a2 · · · ak
k
−1 X −1
= ∂(∂ a1 )a2 · · · ak + [∂ a1 , a j ]a2 · · · abj · · · ak
j=2
k Z
X −1
+ ∂ a1 , ∂a j c a2 · · · abj · · · ak .
j=2 Σ g
−1
When a1 is in the complementary space to the harmonic forms, we know ∂∂ a1 =
a1 so we have the relation
k
X −1
0 = ba1 · · · ak c + b[∂ a1 , a j ]a2 · · · abj · · · ak c
j=2
k Z
X −1
+ ∂ a1 , ∂a j bc a2 · · · abj · · · ak c,
j=2 Σ g
which allows us to iteratively reduce the “polynomial” degree of the original term
ba1 · · · ak c (from k to k − 1 or less, in this case).
This relation looks somewhat complicated but it is easy to understand for k
small. For instance, in k = 1, we see that
0 = ba1 c
whenever a1 is in the space orthogonal to the harmonic forms. We get
Z
−1 −1
0 = ba1 a2 c + b[∂ a1 , a2 ]c + ∂ a1 , ∂a2 bcc.
Σ g
for k = 2.
1.4. The free βγ system. In Section 3 we studied the local structure of the
free βγ theory; in other words, we carefully examined the simplest structure maps
for the factorization algebra Obsq of quantum observables on the plane C. It is
natural to ask about the global sections on a closed Riemann surface.
There are many ways, however, to extend this theory to a Riemann surface Σ.
Let L be a holomorphic line bundle on Σ. Then the L-twisted free βγ system has
fields
E = Ω0,∗ (Σ, L) ⊕ Ω1,∗ (Σ, L∨ ),
where L∨ denotes the dual line bundle. The −1-symplectic pairing on fields is to
apply pointwise the evaluation pairing between L and L∨ and then to integrate
the resulting density. The differential is the ∂ operator for these holomorphic line
bundles.
q
Let ObsL denote the factorization algebra of quantum observables for the L-
twisted βγ system. Locally on Σ, we know how to understand the structure maps:
pick a trivialization of L and employ our work from Section 3.
q
1.4.1 Proposition. Let bk = dim H k (Σ, L). Then H ∗ ObsL (Σ) is a rank-one free
module over C[~] and concentrated in degree −b0 − b1 .
Proof. As usual, we use the spectral sequence arising from the filtration F k =
Sym≤k on Obsq . The first page just depends on the cohomology with respect to ∂,
so the underlying graded vector space is

Sym H ∗ (Σ, L)[1] ⊕ H ∗ (Σ, L∨ ⊗ Ω1hol )[1] [~],
where Ω1hol denotes the holomorphic cotangent bundle on Σ. By Serre duality, we

know this is the symmetric algebra on a +1-symplectic vector space, concentrated
in degrees 0 and −1 and with dimension b0 + b1 in each degree. The remaining
differential in the spectral sequence is the BV Laplacian for the pairing, and so
the cohomology is spanned by the maximal purely odd element, which has degree
−b0 − b1 .
2. Abelian Chern-Simons theory and quantum groups

In this section, we analyze Abelian Chern-Simons theory from a point of view
complementary to our analysis in Section 5. It is standard in mathematics to inter-
pret Chern-Simons theory in terms of a quantum group: line operators in Chern-
Simons theory are representations of the quantum group, and expected values of
line operators (which provide invariants of knots in three-manifolds) can be con-
structed using the representation theory of the quantum group.
In the language of physics, the connection between Chern-Simons theory and
the quantum group is explained as follows. In any topological field theory, one
expects to be able to define a category whose objects are line operators and whose
morphisms are interfaces between line operators. More explicitly, we have in mind
fixing a line in space and every line operator in this category is supported on a line
parallel to the fixed line. In a three-dimensional TFT, we thus have two directions
orthogonal to the fixed direction, and we can imagine taking the operator product
of line operators in this plane. This OPE is expected to give the category of line
operators the structure of a braided monoidal category. For Chern-Simons theory,
the expectation from theoretical physics is that the braided monoidal category of
line operators is the category of representations of a quantum group.
Our primary goal in this section is to implement this idea in the language of
factorization algebras, but developed in the particular example of Abelian Chern-
Simons theory. This example demonstrates, in a substantially simpler context,
many of the ideas and techniques used to construct the Yangian from a non-Abelian
4-dimensional gauge theory in Costello (n.d.c). (The gauge theory yielding the
Yangian should be viewed as a variant of non-Abelian Chern-Simons theory that
lives C × R2 . The theory is holomorphic in the C direction and topological along
the R2 directions.)
The construction here uses deep theorems about En algebras, Koszul duality,
and Tannakian reconstruction, which take us outside the central focus of this text.
So we will explain the structure of the argument and provide references to the
2. ABELIAN CHERN-SIMONS THEORY AND QUANTUM GROUPS 207
relevant literature, but we will not give a complete proof. The interested reader is
encouraged to consult Costello (n.d.c) for more.
2.0.1. Overview of this section. Before we can explain the argument, we need
to formulate the statement. Thus, we begin by explaining how every En algebra —
q
and in particular ObsCS , which is an E3 algebra — has an associated ∞-category of
modules with a natural En−1 -monoidal structure. This terminology means that in
this category of modules, there is a space of ways to tensor together modules, and
this space is homotopy equivalent to the configuration space of points in Rn−1 . In
q
our situation, n = 3 so the ∞-category of modules for ObsCS will be E2 -monoidal,
which is the ∞-categorical version of being braided monoidal.
With these ideas about modules in hand, we can give a precise version of our
claim and provide a strategy of proof. In particular, we explain how these modules
q
for ObsCS are equivalent to the category of representations of a quantum group.
We finish this section by connecting the concept of module for an E3 algebra
to that of a line operator defined in a more physical sense. This will justify our
q
interpretation of the E2 -monoidal category of left modules of ObsCS as being the
category of line operators for Abelian Chern-Simons theory.
Remark: The kinds of assertions and constructions we want to make here rely on
methods and terminology from higher category theory and homotopical algebra.
Formalizing our ideas here would be a nontrivial endeavor, and orthogonal to our
primary goals in this book. Hence we will use the language of higher categories
here quite loosely, although our expressions can be made precise. In particular, in
this section only, we will sometimes simply use “category” in place of ∞-category,
except where the distinction becomes important. We will also use terminology like
“functor” as if working with ordinary categories. ^
2.1. Modules for En algebras. Even in classical algebra, there are different
notions of module for an associative algebra: left modules, right modules, and
bimodules. If we view associative algebras as 1-dimensional in nature, it makes
us suspect that for an n-dimensional algebra — such as an En algebra — there
might be a proliferation of distinct notions reasonably called “modules,” and that
suspicion is true. (See Ayala et al. (n.d.a) for examples.) Nevertheless, in the
context here, we will only need the notion of left (equivalently, right) modules.
The reason is that every En algebra A has an “underlying” E1 algebra A, which
is unique up to equivalence (we explain this “forgetful functor” momentarily).
Thus we can associate to A its category of left modules, which we will denote
LModA to emphasize that we mean left modules (and not bimodules). As we will
explain, this category LModA is naturally equipped with an En−1 -monoidal struc-
ture.
More generally, every En algebra has an underlying Ek algebra for all 0 ≤ k ≤
n. This fact can be seen in two distinct ways: via the operadic definition or via the
factorization algebra picture. They are both geometric in nature.
We start with the operadic approach. Fix a linear inclusion of ι : Rk ,→ Rn . We
will use ι : (x1 , . . . , xk ) 7→ (x1 , . . . , xk , 0, . . . , 0). Then if Dkr (p) denotes the k-disc
of radius r in Rk centered at p, then Dnr (ι(p)) ∩ ι(Rk ) = ι(Dkr (p)), where Dnr (ι(p))
denotes the n-disc of radius r in Rn centered at p. In this way, ι induces a map

ι(m) : Ek (m) → En (m) from the configuration space of m disjoint little k-discs
to the configuration space of m disjoint little n-discs. Altogether, we get a map
ι : Ek → En of topological operads. Algebras over operads pull back along maps
of operads, so for A ∈ AlgEn , we obtain ι∗ A ∈ AlgEk .
One can rephrase this from the point of view of factorization algebras. We will
need a construction of factorization algebras that we have not used before. If M is
a manifold, F a factorization algebra on M, and X ⊂ M a submanifold (possibly
with boundary), we can define a factorization algebra i∗X F by setting
i∗X F (U) = lim F (V)

V⊃U
where the limit is taken over those opens V in X which contain U. The factorization
product on F defines one on i∗X F .
If F is locally constant, this construction is particularly nice: in that case, the
limit is eventually constant so that i∗X F (U) is F (V) for some sufficiently small open
V containing U.
If R ⊂ Rn is a line, we can restrict a locally constant factorization algebra on
n
R to one on R in this way, to produce an associative algebra. This restriction
procedure is how one describes the E1 algebra underlying an Ek algebra in the
language of factorization algebras.
2.1.1. Modules via geometry. We now want to examine the type of monoidal
structure on LModA arising from an En algebra structure on A. Again, one can use
results about operads and categories or one can use factorization algebras. Here we
begin with the factorization approach.
Let
Hn = {(x1 , . . . , xn ) ∈ Rn | xn ≥ 0}
denote the “upper half space.” Let ∂Hn denote the wall {xn = 0}. By “disc in H” we
will mean the intersection D ∩ Hn for any open disc D ⊂ Rn . Since we are always
working in Hn for the moment, we will use an open U to denote its intersection
with Hn .
Let A be a locally constant factorization algebra on Rn associated to the En
algebra A. Given M ∈ LModA and a point p ∈ ∂Hn living on the wall {xn = 0},
we construct a factorization algebra M p on Hn by extending from the following
factorization algebra on a basis. For an open U not containing p, we set M p (U) =
A(U), and for any disc D in Hn containing p, we set M p (D) = M. For any disjoint
union U t D with p ∈ D, we set M p (U t D) = A(U) ⊗ M. The structure maps are
easy to specify. If U t D ⊂ D0 , then U ⊂ D0 \ D, and on this hemispherical shell,
A(D0 \ D) ' A. Since we know how A acts on M, we have the composition
M p (U) ⊗ M p (D) = A(U) ⊗ M → A(D0 \ D) ⊗ M → A ⊗ M → M.
Now extend M p from this basis to all opens.

To help visualize this factorization algebra, here is a picture in dimension 2.
U2
A⊗A⊗M
U1
M
V2 V1
p
To each of the opens U1 and U2 , M p assigns a copy of A. To the half-discs V1 and

V2 , M p assigns the left A-module M.
Remark: It can help to use the following pushforward to understand the structure
maps. The distance-from-p function d p : x 7→ d(p, x) sends Hn to R≥0 . Pushing
forward M p along d p produces the factorization algebra encoding A on R>0 with
M supported at 0, as discussed in Section 3. This pushforward makes clear how
the action on M of A, evaluated on complicated opens in Hn , factors through the
action of A. This pushforward encodes precisely the action of A on M with which
we began. ^
The same type of procedure allows us to assign a factorization algebra behav-
ing like A in the bulk but with modules “inserted” at points on the boundary. Let
p1 , . . . , pk be distinct points on Hn . Let M1 , . . . , Mk be objects of LModA . Abu-
sively, let ~p denote this collection {(Mi , pi )}. We construct a factorization algebra
M ~p as follows. For any open U disjoint from the points {pi }, set M ~p (U) = A(U).
For any disc D containing exactly one distinguished point pi , set M ~p (D) = Mi .
For any disc D0 containing U t D, with D and D0 containing pi but not other
distinguished point p j , we provide the structure map as above. In other words,
in a sufficiently small neighborhood of pi , M ~p becomes M pi as described above.
Extend to all opens to produce M ~p .
In the picture below, we use M p to denote the left A-module associated with
the point p and Mq to denote the left A-module associated with the point q.
V3 A ⊗ M p ⊗ Mq
U
M p ⊗A Mq
V1 V2
p q
To the half-disc V1 , M p,q assigns M p ; to the half-disc V2 , M p,q assigns Mq ; and

to the open U, M pq assigns A. The value on the big half-disc V3 must receive
maps from all these values separately and all intermediary opens (e.g., through a
half-disc containing V2 and U but not V1 ), so its value must be some version of
“M p ⊗A Mq .”
From such an M ~p , we can recover a left A-module in a very simple way. Con-
sider the projection map π : Hn → R≥0 sending (x1 , . . . , xn ) to xn . Then π∗ M ~p
returns A on any interval not containing the boundary point 0, but it has interesting
value on an interval containing 0.
In this way, we produce a left A-module for any configuration of distinct
points {p1 , . . . , pk } in ∂Hn Rn−1 and any corresponding k-tuple of left A-modules
{M1 , . . . , Mk }. In other words, we have sketched a process that can lead to making
LModA an En−1 algebra in some category of categories.
2.1.2. Modules via operads. Developing the picture sketched above into a for-
mal theorem would be somewhat non-trivial. Fortunately, Lurie has proved the
theorem we need in Lurie (n.d.b), using the operadic approach to thinking about
En algebras. We begin by stating the relevant result from his work and then gloss
the essential argument (although we hide a huge amount of technical machinery
developed to make such an argument possible).
We need some notation to state the result. If C is a symmetric monoidal ∞-
category, then C is itself an E∞ algebra in the ∞-category Cat∞ of ∞-categories,
equipped with the Cartesian symmetric monoidal structure (i.e., the product of cat-
egories). Hence, it makes sense to talk about left module categories over C, which
we denote LModC (Cat∞ ). Let Cat∞ (∆) denote the subcategory of ∞-categories
possessing geometric realizations and whose functors preserve geometric realiza-
tions.
2.1.1 Theorem (Theorem 4.8.5.5 and Corollary 5.1.2.6, Lurie (n.d.b)). Let C
denote a symmetric monoidal ∞-category possessing geometric realizations and
whose tensor product preserves colimits separately in each variable. Then there is
a fully faithful symmetric monoidal functor
LMod− : AlgE1 (C) → LModC (Cat∞ (∆))
sending an E1 algebra A to LModA . This functor determines a fully faithful functor
AlgEn (C) → AlgEn−1 (LModC (Cat∞ (∆))).
This result might seem quite technical but it is a natural generalization of more
familiar constructions from ordinary algebra. One might interpret the first part as
saying that
((i)) there is a big category whose objects are categories with an action of
the symmetric monoidal category C,
((ii)) the category of left modules over an algebra in C gives an object in this
big category, and
((iii)) there is a way of “tensoring” categories of modules such that
LModA LModB ' LModA⊗B .
Tensoring categories might not be familiar, but it is suggested by generalizing the
construction of the tensor product of ordinary vector spaces: given a product of
categories D × D0 , we might hope there is a category D D0 receiving a bifunctor
from D × D0 such that all other bifunctors from D × D0 factor through it. Note that
we have a natural bifunctor LModA × LModB → LModA⊗B .
To understand the second part of the theorem, it helps to know about Dunn
additivity: it is equivalent to give an object A ∈ C the structure of an En algebra as
it is to give A first the structure of an E1 algebra and then the structure of an En−1
algebra in the category of E1 algebras. In other words, there is an equivalence of
categories
AlgEn−1 (AlgE1 (C)) ' AlgEn (C).
(See Theorem 5.1.2.2 of Lurie (n.d.b).) By induction, to give A an En algebra is to
specify n compatible E1 structures on A.
Geometrically, this assertion should seem plausible. If one has an En structure
on A, then each way of stacking discs along a coordinate axes in Rn provides an
E1 multiplications. Hence, we obtain n different E1 multiplications on A. In the
other direction, knowing how to multiply along these axes, one should be able to
reconstruct a full En structure.
The theorem then says that we get a natural functor
LMod
AlgEn (C) ' AlgEn−1 (AlgE1 (C)) −−−−→ AlgEn−1 (LModC (Cat∞ (∆))),
where the first equivalence is Dunn additivity and the second functor exists because
LMod− is symmetric monoidal. This composed arrow is what allows one to turn
an En algebra into an En−1 monoidal category.
2.2. The main statement and the strategy of proof. Let ACS denote the E3
q
algebra associated to the locally constant factorization algebra ObsCS of quantum
observables for Abelian Chern-Simons theory with the rank 1 Abelian Lie algebra.
Since we will do everything over the field of complex numbers, we will identify
this Lie algebra with C. By the theory described above, the category LModACS is
an E2 -monoidal category. Recall that E2 -monoidal categories are the ∞-categorical
version of braided monoidal categories. There is another braided monoidal cate-
gory intimately connected with Chern-Simons theory: representations of the quan-
tum group. We want to relate these two categories.
In the case of interest, the quantum group is a quantization of the universal
enveloping algebra of the Abelian Lie algebra C. Typically, a quantum group de-
pends on a parameter often denoted q or ~. The size of this parameter is related to
the choice of an invariant pairing on the Lie algebra. In the Abelian case, multi-
plication by scalars is a symmetry of the Lie algebra that also scales the invariant
pairing. Thus, all values of the quantum parameter are equivalent, and we find only
a single quantum group.
It will be convenient to consider a completed version of the universal envelop-
ing algebra. Thus, our Abelian quantum group will be U(C) b = C[[t]]. The Hopf
algebra structure will be that on the universal enveloping algebra. Thus, it is com-
mutative and co-commutative and the variable t is primitive. The only interesting
structure we find is the R-matrix, given by the formula
R = et⊗t ∈ U(C)
b b ⊗U(C).
b
This R-matrix gives us a quasi-triangular Hopf algebra (in a completed sense).

It is a standard result in the theory of quantum groups that the category of

modules over a quasi-triangular Hopf algebra is a braided monoidal category. We
will let Rep(U(C))
b denote the category of finite-dimensional representations of our
quantum group. Such a finite-dimensional representation is the same as a vector
space V with a nilpotent endomorphism.
We would like to compare Rep(U(C))
b with a braided monoidal category arising
from our E3 -algebra ACS . The category of left modules over this E3 algebra is an
E2 -monoidal ∞-category. To construct a braided monoidal category, we need to
turn this ∞-category into an ordinary category, and we need this ordinary category
to be equivalent as a monoidal category to Rep(U(C)).
b
2.2.1. Identifying the category of interest. To isolate the classical category we
need, the following technical proposition will be useful.
2.2.1 Proposition. Let us view ACS as an E2 algebra via the map of operads from
E2 to E3 . Then there is a canonical equivalence of E2 algebras from ACS to the
commutative algebra C[ε], where the parameter ε is of degree 1. We view C[ε] as
an E2 algebra via the map from E2 to E∞ .
Since the proof is both technical and unilluminating, we will defer it to the end
of this section.
With this proposition in hand, we can understand the category of modules over
ACS much better. Since ACS is equivalent as an E2 algebra to C[ε], the monoidal
category LModACS is equivalent to the monoidal category LModC[ε] . We will iso-
late in this monoidal ∞-category, a subcategory whose homotopy category is equiv-
alent to Rep(U(C)).
b
Let us start by indicating why there is any reason to expect a relationship be-
tween such categories. A first observation is that Koszul duality of associative
algebras interchanges exterior and symmetric algebras. Indeed, C[ε] is an exte-
rior algebra on one variable, and C[[t]] is a (completed) symmetric algebra on
one variable, and they are Koszul dual to one another. Alternatively, C[[t]] is the
(completed) universal enveloping algebra of the abelian Lie algebra C, C[ε] is the
Chevalley-Eilenberg cochains of the same Lie algebra, and Koszul duality inter-
changes the universal enveloping algebras and the Chevalley-Eilenberg cochains
of a given Lie algebra.
It is then well-known that Koszul duality induces an equivalence between (ap-
propriate) categories of modules. In the case at hand, this relationship works as
follows. If V is a finite-dimensional representation of U(C),
b in which the generator
t acts by a nilpotent endomorphism ρ(t) of V, we define a dg module Φ(V) for C[ε]
by setting
Φ(V) = V ⊗ C[ε], d = ερ(t).
More formally, since V is a module for the Lie algebra C, we can take the Chevalley-
Eilenberg cochains C ∗ (C, V) of C with coefficients in V. This complex is a dg mod-
ule for C ∗ (C) = C[ε]. Let LMod0 (C[ε]) be the full subcategory of the ∞-category
of C[ε]-modules given by the essential image of the functor Φ.
One can describe this category in a different way. There is a functor

Aug : LMod(C[ε]) → dgVect
M 7→ M ⊗LC[ε] C.
We call this functor Aug because it is constructed from the augmentation map
C[ε] → C sending ε → 0. Notice that Aug is a monoidal functor, since the map
C[ε] → C is a map of E2 algebras. The category LMod0 (C[ε]) then consists of
those modules M for which Aug(M) has finite-dimensional cohomology concen-
trated in degree 0. (In other words, it is the homotopy fiber product of Aug and the
inclusion of finite-dimensional vector space Vect f d into dgVect.) Since, as we have
seen, ACS is equivalent to C[ε] as an E2 algebra, and so also as an E1 algebra, we
can view this category as being a subcategory of LMod(ACS ).
2.2.2. The main result. The classical category we want to compare with Rep(U(C))
b
is the homotopy category Ho(LMod0 (ACS )). Note that, since LMod0 (ACS ) is an
E2 -monoidal category, its homotopy category is a braided monoidal category.
Our main theorem is the following.
2.2.2 Theorem. Equip Rep(U(C))
b with the structure of braided monoidal category
by the R-matrix R = et⊗t for the Abelian quantum group. The functor
Φ : Rep(U(C))
b → Ho(LMod0 (ACS ))
is an equivalence of braided monoidal categories, where
The interesting part of this theorem is that it replaces the somewhat-mysterious
E3 algebra ACS with the much more explicit and concrete quasi-triangular Hopf
algebra C[[t]] with R-matrix et⊗t .
2.2.3. The argument. We will break the proof up into several lemmas.
2.2.3 Lemma. The functor
Φ : Rep(U(C))
b → Ho(LMod0 (ACS ))
is an equivalence of monoidal categories.
This equivalence is as ordinary categories, not ∞-categories.
Proof. Since ACS is equivalent as an E2 algebra to C[ε], the monoidal product
on the category of C[ε]-modules is given by tensoring over C[ε]. Let us check that
Φ is monoidal. If V, W are representations of U(C)b = C[[t]], on which t acts by
ρV (t), ρW (t), respectively, then t acts on V ⊗ W by ρV (t) ⊗ 1 + 1 ⊗ ρW (t). There is
an isomorphism
Φ(V) ⊗C[ε] Φ(W) = (C[ε] ⊗ V) ⊗C[ε] (C[ε] ⊗ W) → C[ε] ⊗ (V ⊗ W) = Φ(V ⊗ W)
that takes the differential on the left hand side, coming from the tensor product of
the differentials on Φ(V) and Φ(W), to the differential on Φ(V ⊗ W). This natural
isomorphism makes Φ monoidal. Next, let us check that Φ is an equivalence of cat-
egories. By the definition of LMod0 (ACS ), the functor Φ is essentially surjective.
It is easy to verify that Φ is full and faithful.
2.2.4 Lemma. The equivalence of the previous lemma transfers the braided monoidal
category structure on Ho(LMod0 (ACS )) to a braided monoidal category on Rep(U(C)).
b
This braided monoidal category structure is realized by some R-matrix
R ∈ U(C)
b b ⊗U(C),
b
giving U(C)
b the structure of quasi-triangular Hopf algebra.
Proof. Let us sketch how one can see this using the Tannakian formalism,
before giving a more concrete approach to constructing the quasi-triangular Hopf
algebra structure.
The category Ho(LMod0 (ACS ) is a braided monoidal category. There is a
functor
Aug : Ho(LMod0 (ACS )) → Vect
sending a module M to M ⊗C[ε] C, using the equivalence of E2 algebras between
ACS and C[ε]. The augmentation map C[ε] → C is a map of E2 algebras, so this
augmentation functor is monoidal. Under the equivalence
Rep(U(C))
b ' Ho(LMod0 (ACS ))
the functor Aug becomes the forgetful functor, which takes a representation of
U(C)
b and forgets the U(C)
b action.
Let us recall how the Tannakian formalism works for braided monoidal cat-
egories. (See, for example, Etingof et al. (2015).) Suppose we have a braided
monoidal category C with a monoidal (but not necessarily braided monoidal) func-
tor F : C → Vect. Then, by considering the automorphisms of the fibre functor
F, one can construct a quasi-triangular Hopf algebra whose representations will be
our original braided monoidal category C, and where the functor F is the forget-
ful functor. (One needs some additional technical hypothesis on C and F for this
construction to work.)
We are in precisely the situation where this formalism applies. The braided
monoidal category is Ho(LMod0 (ACS ), and the fibre monoidal functor is the func-
tor Aug given by the augmentation of ACS as an E2 algebra. Since we already
know that Ho(LMod0 (ACS )) is equivalent to representations of the Hopf algebra
U(C),
b in such a way that the fibre functor becomes the forgetful functor, the Hopf
algebra produced by the Tannakian story is U(C).
b We therefore find that the braid-
ing is described by some R-matrix on this Hopf algebra.
Since this approach is so abstract, let us describe how to see concretely the
R-matrix on U(C).
b
The equivalence with Ho(LMod0 (ACS )) gives us a braided monoidal struc-
ture on representations of U(C).
b For any two representations V, W of U(C),
b this
braiding gives an isomorphism V ⊗ W W ⊗ V. There is already such an iso-
morphism, just of vector spaces, since vector spaces are a symmetric monoidal
category. Composing these two isomorphisms gives, for all representations V, W

of U(C),
b an isomorphism RV,W : V ⊗ W − → V ⊗ W. It is an isomorphism of vector
spaces and not necessarily of U(C)-modules,
b and it is natural in V and W.
The functor
Rep(U(C))
b × Rep(U(C))
b → Vect
(V, W) 7→ V ⊗ W
b b⊗2 module U(C)
is represented by the right U(C) b b ⊗U(C),
b i.e., the functor is given by
tensoring with this right module. The natural automorphism RV,W of this functor is
thus represented by an automorphism of this right module, which must be given by
left multiplication with some element
R ∈ U(C)
b b ⊗U(C).
b
This element is the universal R-matrix.
Now, the axioms of a braided monoidal category translate directly into the
properties that R must satisfy to define a quasi-triangular Hopf algebra structure on
U(C).
b
So far, we have seen that we can identify the braided monoidal category of rep-
resentations of U(C)
b with the braided monoidal category Ho LMod0 (ACS ), where
we equip U(C)
b with some R-matrix. It remains to show that there is essentially
only one possible R-matrix.
2.2.5 Proposition. Every R-matrix for U(C)
b satisfying the axioms defining a quasi-
triangular Hopf algebra is of the form
R = exp(c(t ⊗ t)) ∈ U(C)
b b ⊗U(C)
b
where c is a complex number.
Proof. Recall that an R-matrix of a Hopf algebra A is an invertible element R
of A ⊗ A such that
∆op (x) = R∆(x)R−1 for all x ∈ A,
∆ ⊗ idA (R) = R13 R23 ,
idA ⊗∆(R) = R13 R12 ,
where ∆ denotes the coproduct on A, ∆op denotes ∆ postcomposed with the flip
operation, and R13 , for example, denotes
X
ai ⊗ 1 ⊗ bi ∈ A ⊗ A ⊗ A
i
where R = i ai ⊗ bi ∈ A ⊗ A.
P
For us, due to the completion of the tensor product, an element R will live
⊗C[[t2 ]]. As C[[t]] is cocommutative, we see that the first equation will
in C[[t1 ]]b
be trivially satisfied by any invertible element. The only conditions are then the
remaining equations. As we are working with a deformation of the usual symmetric
braiding on C[[t]]-modules, an R-matrix is of the form
1 ⊗ 1 + c1 t1 ⊗ 1 + c2 1 ⊗ t2 + · · · .
To simplify the problem further, note that, because C[[t1 , t2 ]] is a commutative

algebra, we can take the logarithm of R. If we let R denote this logarithm, then the
equations then become additive:
∆ ⊗ idA (R) = R13 + R23 ,
idA ⊗∆(R) = R13 + R12 .
Writing R = i x(i) ⊗ y(i) with x(i) ∈ C[[t1 ]] and y(i) ∈ C[[t2 ]], we see that the first
P
equation becomes
X X
∆(x(i) ) ⊗ y(i) = (x(i) ⊗ 1 + 1 ⊗ x(i) ) ⊗ y(i) ,
i i
which forces each x(i) to be primitive. The other equation likewise forces each y(i)
to be primitive. Hence R = ct1 ⊗ t2 for some constant c and so R = exp(ct1 ⊗ t2 ).
The constant c is the only freedom available in choosing an R-matrix.
Thus, we have shown the desired equivalence of braided monoidal categories,
where U(C)
b is equipped with an R-matrix of the form exp(c(t ⊗ t)) for some value
of the constant c. This constant cannot be zero, as this would force ACS to be a
commutative E3 algebra, and ACS is not commutative. Any two non-zero values
of the constant c are related by scaling t, and so all the braided monoidal categories
are equivalent. This completes the proof of Theorem 2.2.2.
2.3. Line operators. In this section, we will explain a general definition of a

Wilson line operator in the language of factorization algebras, and we will attempt
to match our definition with ideas prevalent in the physics literature. We will then
show that, in the case of Abelian Chern-Simons theory, the category of Wilson line
operators is the category we considered above.
2.3.1. The basic idea of defects. Let us first explain how line operators (and
more general defects) are understood from the physics point of view. Suppose we
have a field theory on a manifold M, and X ⊂ M is a submanifold. A defect for
our field theory on X is a way of changing the field theory on X while leaving it
the same outside X. Often, one can realize a defect by coupling a theory living
on X to our original theory living on M. In the case that X is a one-dimensional
submanifold, we call such a defect a line operator (or line defect).
For example, if K : S 1 → R3 is a knot, we could couple a free fermion on S 1
to Abelian Chern-Simons theory on R3 . If A is the field of Chern-Simons theory
and ψ : S 1 → C the fermionic field on S 1 , we can couple the theories with the
action Z Z
S K (A, ψ) = ψ(θ)(d + K A)ψ(θ) +
∗
A ∧ dA.
S1 R3
The quantization of this field theory, which we will not develop here, should pro-
q q
duce a factorization algebra of observables ObsCS ,K that agrees with ObsCS on
opens disjoint from K but that on opens intersecting K also includes observables
on ψ. A small tubular neighborhood K b of K inside R3 is diffeomorphic to K × R2 ,
q
so we can pushforward ObsCS ,K from this neighborhood to S 1 . The underlying
E1 algebra AK should look something like ACS tensored with a Clifford algebra
(the observables for the uncoupled free fermion). Thus the line defect for Abelian
q q
Chern-Simons theory produces an interesting modification ObsCS ,K of ObsCS near
K; we will view this factorization algebra as a model for what a line defect should
be in the setting of factorization algebras.
By picking different field theories supported on K and coupling them to Chern-
Simons theory in the bulk, we obtain other factorization algebras. One can do this,
for instance, by varying the coupling constants in the example above or adding
more interactions terms. One can similarly imagine picking a 2-dimensional sub-
manifold and coupling a field theory on it to the 3-dimensional bulk.
2.3.2. Defects via factorization algebras. Let us now explain in the language
of factorization algebras how to couple a field theory to a quantum-mechanical
system living on a line. Suppose we have a factorization algebra F on Rn , which we
assume here is locally constant. Let us fix a copy of R ⊂ Rn . Let us place a locally-
constant factorization algebra G on R, which we think of as the observables of a
quantum-mechanical system. We will view G as a homotopy-associative algebra or
E1 algebra or A∞ algebra, but we will tend to write formulas as if everything were
strictly associative. (One can always replace an A∞ algebra by a weakly equivalent
but strictly associative dg algebra.)
Let i∗ G denote the pushforward of G to a factorization algebra on Rn . Note
that i∗ G takes value C on any open that does not intersect R, and it takes value
G(i−1 (U)) on any open which does intersect R.
Before we define a coupled factorization algebra, we need to understand the
situation where the field theory on Rn and the field theory on R are uncoupled. In
this case, the factorization algebra of observables for the system is simply F ⊗ i∗ G.
We are interested in deforming this uncoupled situation.
2.3.1 Definition. A coupling of our factorization algebra F to the factorization
algebra G living on R is an element α ∈ F |R ⊗ G of cohomological degree 1
satisfying the Maurer-Cartan equation
dα + 12 [α, α] = dα + α ? α = 0,
where ? is the product in F |R ⊗ G viewed as an associative algebra.
Here, by F |R , we mean the factorization algebra on R obtained by restricting

F (e.g., by taking a small tubular neighborhood of this copy of R and then pushing
forward to R). We can view F |R as a homotopy-associative algebra. The tensor
product F |R ⊗ G is then also a homotopy-associative algebra.
Remark: Bear in mind the following:
((i)) It is common to model homotopy associative algebras by A∞ alge-
bras. If one takes this route, then one should use the L∞ version of
the Maurer-Cartan equation.
((ii)) We are always imagining a situation in which our line defect is a small
deformation of a decoupled system. One can capture this requirement
by including a formal parameter, viewed as a coupling constant, next

to α in this definition, but we will not do so. ^
Given such a coupling α, we can deform the factorization algebra F |R ⊗ G by
simply adding [α, −] to the differential of F |R ⊗ G. In other words, the twisted
differential dα is given by
dα (x) = dx + [α, x] = dx + α ? x + x ? α.
We denote this factorization algebra on R by F |R ⊗α G.

We can also deform F ⊗ i∗ G using this data. On R, it is simply F |R ⊗α G. Away
from R, we change nothing and the factorization algebra is just given by F . It is
not completely obvious that this construction makes sense, but we will see in the
following lemma that it does.
2.3.2 Lemma. This construction defines a factorization algebra F ⊗α i∗ G on Rn
deforming F ⊗ i∗ G.
We call this factorization algebra a line operator, and below we will relate it
explicitly to the standard example of a Wilson observable.
Proof. We will sketch a proof using technology developed in the literature on

En algebras. For simplicity, let us suppose that R is a straight line in Rn .
Let us identify Rn = R × Rn−1 and use coordinates (x, v) where x ∈ R and
v ∈ Rn−1 . Our chosen copy of R is {v = 0}. Define a projection
π : Rn \ R → R × R>0
(x, v) 7→ (x, kvk).
The pushforward π∗ F will be a locally constant factorization algebra on the two-

dimensional manifold R × R>0 , and so an E2 algebra.
We are interested in extensions of π∗ F into a factorization algebra on the half-
space R × R≥0 , which is constructible with respect to the stratification where the
two strata are R × 0 and R × R>0 . Such an extension will be an E1 algebra on R × 0,
together with some compatibility between this E1 algebra and the E2 algebra π∗ F .
We would like the E1 algebra on R × 0 to be F |R ⊗α G.
The work of Ayala, Francis, and Tanaka Ayala et al. (n.d.a) tell us that such an
extension of π∗ F is the same as a constructible factorization algebra on Rn that is
constructible with respect to the stratification where the strata are Rn \ R and R, and
which on Rn \ R is F and on R is F |R ⊗α G.
One can ask, what compatibility is required to glue the E2 algebra π∗ F on
R×R>0 to the E1 algebra F |R ⊗α G? The answer is that we need to give a morphism
of E2 algebras from π∗ F to the Hochschild cochain complex of F |R ⊗α G. Now,
since the deformation of F |R ⊗ G given by [α, −] is an inner deformation, it does
not change the Hochschild cochain complex (up to weak equivalence). Thus, the
gluing data we need is the same as in the case when α = 0. But when α = 0, we
already have the desired factorization algebra on Rn , namely F ⊗ i∗ G.
Remark: It might seem restrictive to only consider deformations of F |R ⊗G that are

inner (i.e., given by an element α) rather than arbitrary deformations of the algebra
structure. The last paragraph of the proof indicates one reason why we define
“coupling” via inner deformations, and we provide some further motivations in the
next section. ^
2.3.3. Motivating this construction. There are two natural questions remain-
ing from our discussion of line operators. Firstly, in the case of Chern-Simons
theory, how do we match up line operators with representations of the quantum
group? And secondly, why does our notion of “coupling,” Definition 2.3.1, match
up with the physical idea of coupling a quantum mechanical system to our field
theory?
Let us address the first point. In the factorization algebra approach, a one-
dimensional topological field theory is specified by an associative algebra, viewed
as the algebra of operators (or observables) for the field theory. In other approaches
to (topological) field theory Lurie (2009b), however, one is required to have a
Hilbert space as well. A one-dimensional topological field theory is then speci-
fied by a finite-dimensional vector space V, which is the Hilbert space attached to
a point. The algebra of operators is the matrix algebra End(V). It is thus natural
to consider coupling a field theory to a topological quantum mechanical system of
Atiyah-Segal-Lurie type, where the algebra of operators is End(V).
If we do this, we find that a coupling is a Maurer-Cartan element in F |R ⊗
End(V). But such an element also provides a deformation of F |R ⊗ V as a free left
F |R -module into a non-trivial projective left module. In this way, we have sketched
why the following lemma is true.
2.3.3 Lemma. The following are equivalent:
((i)) specifying a projective rank k left module for the En algebra F , and
((ii)) coupling a locally constant factorization algebra F on Rn to the trivial
quantum mechanical system on R ⊂ Rn whose algebra of operators is
the matrix algebra glk .
For Chern-Simons theory, this class of line operators matches up nicely with
the usual Wilson observables. As a particularly simple example, let’s return to
Abelian Chern-Simons theory. In that case, we know that F |R is equivalent to the
algebra C[ε], where ε has cohomological degree one. Every degree one element in
C[ε] ⊗ glk is of the form εT , with T ∈ glk , and every εT is a Maurer-Cartan element
since the differential d here is trivial and ε2 = 0. A deformation of the action of
C[ε] on C[ε] ⊗ V, of course, amounts to changing how ε acts, and we are free to
change its action by left multiplication from the free action
mε : M + εN 7→ εM
to any linear map of the form
eε : M + εN 7→ εAM,
m
where A ∈ glk . Now set T = A − 1.
Next, let us explain how to interpret our concept of “coupling” from the point
of view of physics. To do this, we need to say a little bit about local operators (that
is, operators supported at a point) from the point of view of factorization algebras.
This topic will be treated in more detail in volume 2. Here, we will be brief.
In any field theory on Rn whose observables are given by a factorization algebra
F , there is a co-stalk F x of F for each point x ∈ Rn . This co-stalk is the limit
lim x∈U F (U) of the spaces of observables on neighbourhoods U of x. The co-stalk
F x is a way to define local observables at x for the field theory: by definition an
element of F x is an observable contained in F (U) for every neighbourhood U of
x. For a general field theory, this limit might produce something hard to describe
(and one might also worry about the difference between the limit and the homotopy
limit). For a topological field theory, however, the limit is essentially constant, so
that we can identify F x with F (D) for any disc D around x.
As x varies in Rn , the space F x forms a vector bundle, typically of infinite rank.
This vector bundle has a flat connection, reflecting the fact that we can differentiate
local operators. Because F x may be infinite-dimensional, one can not in general
solve the parallel transport equation from the flat connection. Thus, for a general
factorization algebra, F x should be thought of as a D-module. For a topological
field theory, however, the bundle F x is a local system, meaning one can solve the
parallel transport equation. The equivalence F x ' Fy given by parallel transport
' '
arises by the quasi-isomorphisms F x −→ F (D) ←− Fy for any disc D containing
both x and y.
Let us consider, as before, a line R ⊂ Rn , with a field theory on Rn whose
observables are described by a factorization algebra F and with a field theory on
R whose observables are described by a factorization algebra G. Let us assume
that both factorization algebras are locally constant. For x ∈ R, we can consider
the space F x ⊗ G x , which is the tensor product of local operators of our two field
theories at x ∈ R ⊂ Rn . As we have seen, the space F x ⊗ G x forms, as x varies,
a vector bundle with flat connection on R. Let us denote this vector bundle by
Floc ⊗ Gloc . Since Floc ⊗ Gloc is a local system, there is a quasi-isomorphism
F |R (D) ⊗ G(D) ' Ω∗ (R, Floc ⊗ Gloc )
for any disc D ⊂ R.
From a physics point of view, we might expect that to couple the system spec-
ified by F to that specified by G, we need to give a Lagrangian density on the real
line R. In this case, such a Lagrangian would be a one-form on R with values in
the bundle of local observables Floc ⊗ Gloc , viewed as observables on the combined
fields for the two field theories.
Such a Lagrangian density does appear naturally in our earlier mathematical
definition. Under the isomorphism
F |R (D) ⊗ G(D) ' Ω∗ (R, Floc ⊗ Gloc ),
a Maurer-Cartan element α on the left hand side corresponds to a sum
S α = S α(0) + S α(1) ,
where S α(0) is a function along R with values in the degree 1 component of the
vector bundle of local observables and S α(1) is a 1-form with values in the degree 0
component of the local observables. The term S α(1) is thus a Lagrangian density. To
better understand what it means, we need to go a bit further.
Under the correspondence above, the Maurer-Cartan equation for α becomes
the equation
h i
ddR S α(0) + dF ⊗G S α(0) + dF ⊗G S α(1) + S α(0) , S α(1) = 0,
where dF ⊗G is the differential on Floc ⊗ Gloc . This equation is essentially a version
of the master equation in the Batalin-Vilkovisky formalism, but this relationship is
manifest only with factorization algebras produced by BV quantization, which is
the subject of Volume 2. Solving the equation perturbatively, however, we can get
the flavor of the relationship.
Consider the situation where we try to solve for S α iteratively in some formal
variable c: (0) (0)
(1) (1)
S α = c S α,1 + S α,1 + c2 S α,2 + S α,2 + ··· .
At the first stage, we work modulo c2 . We then want to solve the equation
(0) (1)
ddR S α,1 + dF ⊗G S α,1 = 0.
(Note that for grading reasons, we always need dF ⊗G S α(0) = 0.) And we are in-
terested in solutions modulo the image of the operator ddR + dF ⊗G on elements in
Ω∗ (R, Floc ⊗Gloc ) of cohomological degree 0; this space is the first-order version of
gauge equivalence between solutions of the Maurer-Cartan equation. Unraveling
these conditions, we see that we care about a Lagrangian density in Floc ⊗ Gloc up
to a total derivative. This data is what one expects from physics. Continuing onto
higher powers of c, one finds that our algebraic definition for coupling the quan-
tum mechanical system G to the field theory F matches with what one expects
from physics.
In sum, for any topological field theory whose observables are given by a fac-
torization algebra F on Rn , this discussion indicates that the category of projec-
tive left modules for F captures the category of Wilson line operators for the n-
dimensional field theory described by F . These categories both admit interesting
ways to “combine” or “fuse” objects: the left F -modules have an En−1 -monoidal
structure, by Lurie’s work, and the line defects can be fused, equipping them with
an En−1 -monoidal structure as well. (For a physically-oriented introduction to the
relationship between defects and higher categories, see Kapustin (2010).) We as-
sert that these agree: when one examines the physical constructions, they behave as
the homotopical algebra suggests. For some discussion of these issues, see Costello
(2014).
2.3.4. The take-away message. Let us apply this discussion to Abelian Chern-
Simons theory, where we have an E3 algebra ACS . The left modules that we find
in our discussion above — projective modules of rank n — match exactly with
the category LMod0 (ACS ) we introduced earlier. If we define the E2 -monoidal
category of line operators of Abelian Chern-Simons theory to be the E2 -monoidal
category LMod0 (ACS ), we find that our theorem relating Abelian Chern-Simons
theory to the quantum group takes the following satisfying form.
Theorem. The following are equivalent as braided monoidal categories:
((i)) the homotopy category of the category of Wilson line operators of
Abelian Chern-Simons theory
((ii)) the category of representations of the quantum group for the Abelian
Lie algebra C.
In this case, the identification between braided monoidal structures can be seen
explicitly. We have already shown that the factorization algebra encodes the Wil-
son loop observables and recovers the Gauss linking number, but it is these rela-
tions that are used in providing the braiding on the line operators. (See, e.g., the
discussion in Kapustin and Saulina (2011).)
2.4. Proof of a technical proposition. Let us now give the proof of the fol-
lowing proposition, whose proof we skipped earlier.
2.4.1 Proposition. View ACS as an E2 algebra via the map of operads from E2 to
E3 . There is a canonical equivalence of E2 algebras from ACS to the commutative
algebra C[ε], where the parameter ε is of degree 1. We view C[ε] as an E2 algebra
via the map of operads from E2 to E∞ .
Proof. Recall that ACS is a twisted factorization envelope. Explicitly, it as-
signs to an open subset U ⊂ R3 , the cochain complex
C∗ (Ω∗c (U)[1] ⊕ C · ~[−1]) ⊗C[~] C~=1 .
The Lie bracket on the dg Lie algebra Ω∗c (U)[1] ⊕ C · ~[−1] is given by the shifted
central extension of the Abelian Lie algebra Ω∗c (U)[1] for the cocycle
Z
α ⊗ β 7→ α ∧ β.
We are interested in the restriction of this twisted enveloping factorization algebra

to a plane R2 ⊂ R3 . Recall that

i∗R2 ACS (U) = lim ACS (V),
V⊂U
where the limit is taken over those opens V in R3 that contain our chosen open
U ⊂ R2 . Note that since ACS is locally constant, this limit is eventually constant.
If we identify R3 = R2 × R, we can identify

i∗R2 ACS (U) = ACS (U × R).
In other words, we can identify i∗R2 ACS with the push-forward π∗ ACS where π :
R3 → R2 is the projection. This identification is as factorization algebras.
There is a quasi-isomorphism, for all U ⊂ R2 ,
Ω∗c (U)[−1] ' Ω∗c (U × R).
Concretely, we can define this Rcochain map as follows. Choose a function f on R

with compact supprt such that f (t)dt = 1. The quasi-isomorphism sends a form
ω ∈ Ω∗c (U) to ω ∧ f (t)dt, where t indicates the coordinate on R.
The factorization algebra π∗ ACS is the twisted factorization algebra of the pre-
cosheaf of Abelian dg Lie algebras π∗ Ω∗c [1]. To an open subset U ⊂ R2 , this
Abelian Lie algebra assigns Ω∗c (U × R)[1]. The cocycle defining the central exten-
sion is given by the formula
Z
[α, β] = ~ α ∧ β.
U×R
We can restrict this cocycle to one on the sub-cosheaf Ω∗c (U), where we use the
map Ω∗c (U) → Ω∗c (U × R)[1] discussed above. We find that it is zero, since for
ω1 , ω2 ∈ Ω∗c (U)[1], we have
Z
ω1 f (t)dt ∧ ω2 f (t)dt = 0.
U×R
It follows that we have a quasi-isomorphism of precosheaves of Lie algebras on

R2 :
C · ~[−1] ⊕ Ω∗c (U) ' C · ~[−1] ⊕ Ω∗c (U × R)[1].
On the right hand side, the central extension is the one defined above, and on the
left hand side it is zero.
By applying the twisted factorization envelope construction to both sides, we
find that that π∗ ACS is quasi-isomorphic to the un-twisted factorization envelope
of the precosheaf of Abelian dg Lie algebras Ω∗c on R2 . Since this Lie algebra is
Abelian, and there is no central extension, the result is a commutative factorization
algebra. The corresponding E2 algebra must therefore come from a commutative
algebra.
Let us now identify this E2 algebra explicitly. Note that, for every open U ⊂
R2 , there is a cochain map
Ω∗c (U) → C[−2]
given by integration over U. This map is a quasi-isomorphism if U is a disc. It also
respects the cosheaf operations, and so gives a map of cosheaves on R2
Ω∗c → C[−2]
to the constant cosheaf C[−2]. The constant cosheaf C assigns to any open U ⊂
R2 the space H0 (U, C), which is a coproduct of a copy of C for each connected
component of U.
Applying the factorization envelope construction, we find a map of prefactor-
ization algebras
C∗ (Ω∗c ) = Sym Ω∗c [1] → Sym C[−1].
This map is a quasi-isomorphism when restricted to disjoint unions of discs. It
therefore identifies the E2 algebra with the commutative algebra Sym C[−1] =
C[ε].
Remark: There is another approach to this lemma that is less direct but applicable
more broadly, by using the machinery of deformation theory.
There is a Hochschild complex for En algebras (in cochain complexes over a
field of characteristic zero) akin to the Hochschild cochain complex for associative
algebras. Furthermore, given an En algebra A, there is an exact sequence of the
form
Def(A) → HHE∗ n (A) → A
relating the En Hochschild complex of A and the complex describing En algebra
deformations of A (see Francis (2013)). First-order deformations of the En algebra
structure are given by the n + 1st cohomology group of the complex Def(A). The
funny shift is due to the fact that the En deformation complex is an En+1 algebra
and the underlying dg Lie algebra is obtained by shifting to turn the Poisson bracket
into an unshifted Lie bracket. Compare, for instance, the fact that HH 2 (A) classifies
first-order deformations of associative algebras.
Moreover, there is an analog of the Hochschild-Kostant-Rosenberg theorem
for En algebras: for A a commutative dg algebra A, it identifies the En -deformation
complex for A with the shifted polyvector fields SymA (TA [−n]). Here TA denotes
the tangent complex of A. (See Calaque and Willwacher (2015) for the full state-
ment and a proof.)
For us, the classical observables of Abelian Chern-Simons are equivalent to
the commutative dg algebra C[ε], where ε has cohomological degree 1, because it
is the factorization envelope of an Abelian Lie algebra. The E2 HKR theorem then
tells us that
HHE∗ 2 (C[ε]) ' C [ε, ∂/∂ε] ,
where ε and ∂/∂ε have cohomological degree 1. There are no degree 3 elements
of our deformation complex, so the E2 deformation must be trivial. Hence ACS
is a trivial deformation as an E2 algebra. Note, by contrast, that there is a one-
dimensional space of first-order deformations of C[ε] as an E3 algebra because
HHE∗ 3 (C[ε]) is C [ε, ∂/∂ε], with ∂/∂ε now in degree 2. ^
APPENDIX A
Background
We use techniques from disparate areas of mathematics throughout this book

and not all of these techniques are standard knowledge, so here we provide a terse
introduction to
• simplicial sets and simplicial techniques;

• operads, colored operads (or multicategories), and algebras over col-
ored operads;
• differential graded (dg) Lie algebras and their (co)homology;
• sheaves, cosheaves, and their homotopical generalizations;
• elliptic complexes, formal Hodge theory, and parametrices,
along with pointers to more thorough treatments. By no means does the reader
need to be expert in all these areas to use our results or follow our arguments. She
just needs a working knowledge of this background machinery, and this appendix
aims to provide the basic definitions, to state the results relevant for us, and to
explain the essential intuition.
We do assume that the reader is familiar with basic homological algebra and ba-
sic category theory. For homological algebra, there are numerous excellent sources,
in books and online, among which we recommend the complementary texts by
Weibel (1994) and Gelfand and Manin (2003). For category theory, the standard
reference Mac Lane (1998) is adequate for our needs; we also recommend the
series Borceux (1994a).
Remark: Our references are not meant to be complete, and we apologize in advance
for the omission of many important works. We simply point out sources that we
found pedagogically oriented or particularly accessible. ^
0.1. Reminders and Notation. We overview some terminology and notations
before embarking on our expositions.
For C a category, we often use x ∈ C to indicate that x is an object of C. We
typically write C(x, y) to the denote the set of morphisms between the objects x
and y, although occasionally we use HomC (x, y). The opposite category Cop has
the same objects but Cop (x, y) = C(y, x).
Given a collection of morphisms S in C, a localization of C with respect to
S is a category C[S −1 ] and a functor q : C → C[S −1 ] satisfying the following
conditions.
((i)) For every morphism s ∈ S , its image q(s) is an isomorphism.

225
226 A. BACKGROUND
((ii)) For any functor F : C → D such that F(s) is an isomorphism for every
s ∈ S , there is a functor F : C[S −1 ] → C and a natural isomorphism
τ : F ◦ F ⇒ F.
((iii)) For every category D, the functor q∗ : Fun(C[S −1 ], D) → Fun(C, D)
sending G to G ◦ q is full and faithful.
These ensure that the localization is unique up to equivalence of categories, if it
exists. (We will not discuss size issues in this book.) The localization is the cate-
gory in which every s ∈ S becomes invertible. A morphism f : C[S −1 ](x, y) can
be concretely understood as a “zig zag”
' '
x ← z1 → z2 ← · · · → y
where the arrows going to the left (the “wrong way”) are all in S and the arrows
going to the right are arbitrary morphisms from C.
We typically apply localization in the setting of a category C with a class W
of weak equivalences. We call W a class of weak equivalences if W contains all
isomorphisms in C and satisfies the 2-out-of-3 property, which states that given
morphisms f ∈ C(x, y) and g ∈ C(y, z) such that two morphisms in the set { f, g, g ◦
f } are in W, then all three are in W. As examples, consider
((i)) any category with isomorphisms as weak equivalences,
((ii)) the category of cochain complexes with cochain homotopy equiva-
lences as weak equivalences,
((iii)) the category of cochain complexes with quasi-isomorphisms as weak
equivalences, or
((iv)) the category of topological spaces with weak homotopy equivalences
as weak equivalences.
We often denote the localization C[W −1 ] by Ho(C, W), or just Ho(C), and call it
the “homotopy category.”
1. Simplicial techniques
Simplicial sets are a combinatorial substitute for topological spaces, so it should
be no surprise that they can be quite useful. On the one hand, we can borrow intu-
ition for them from algebraic topology; on the other, simplicial sets are extremely
concrete to work with because of their combinatorial nature. In fact, many con-
structions in homological algebra are best understood via their simplicial origins.
Analogs of simplicial sets (i.e., simplicial objects in other categories) are useful as
well.
In this book, we use simplicial sets in two ways:
• when we want to talk about a family of QFTs or space of parametrices,
we will usually construct a simplicial set of such objects instead; and
• we accomplish some homological constructions by passing through
simplicial sets (e.g., in constructing the extension of a factorization
algebra from a basis).
After giving the essential definitions, we state the main theorems we use.
1. SIMPLICIAL TECHNIQUES 227
1.0.1 Definition. Let ∆ denote the category whose objects are totally ordered finite
sets and whose morphisms are non-decreasing maps between ordered sets. We
usually work with the skeletal subcategory whose objects are
[n] = {0 < 1 < · · · < n}.
A morphism f : [m] → [n] then satisfies f (i) ≤ f ( j) if i < j.
We will relate these objects to geometry below, but it helps to bear in mind the
following picture. The set [n] corresponds to the n-simplex 4n equipped with an
ordering of its vertices, as follows. View the n-simplex 4n as living in Rn+1 as the
set  

 X 

, , . . . , = .
 
(x x x ) x 1 and 0 ≤ x ≤ 1 ∀ j
 

 0 1 n j j 

 
 j 
Identify the element 0 ∈ [n] with the 0th basis vector e0 = (1, 0, . . . , 0), 1 ∈ [n]
with e1 = (0, 1, 0, . . . , 0), and k ∈ [n] with the kth basis vector ek . The ordering
on [n] then prescribes a path along the edges of 4n , starting at e0 , then going to e1 ,
and continuing till the path ends at en .
Every map f : [m] → [n] induces a linear map f∗ : Rm+1 → Rn+1 by setting
f∗ (ek ) = e f (k) . This linear map induces a map of simplices f∗ : 4m → 4n .
There are particularly simple maps that play an important role throughout the
subject. Note that every map f factors into a surjection followed by an injection.
We can then factor every injection into a sequence of coface maps, namely maps
of the form
fk : [n] → ( [n + 1]
i, i≤k .
i 7→
i + 1, i > k
Similarly, we can factor every surjection into a sequence of codegeneracy maps,
namely maps of the form
dk : [n] → ( [n − 1]
i, i≤k .
i 7→
i − 1, i > k
The names face and degeneracy fit nicely with the picture of the geometric sim-
plices: a coface map corresponds to a choice of n-simplex in the boundary of the
n + 1-simplex, and a codegeneracy map corresponds to “collapsing” an edge of the
n-simplex to project the n-simplex onto an n − 1-simplex.
We now introduce the main character.
1.0.2 Definition. A simplicial set is a functor X : ∆op → S et, often denoted X• .
The set X([n]), often denoted Xn , is called the “set of n-simplices of X.” A map of
simplicial sets F : X → Y is a natural transformation of functors. We denote this
category of simplicial sets by sSet.
Let’s quickly examine what the coface maps tell us about a simplicial set X.
For example, by definition, a map fk : [n] → [n+1] in ∆ goes to X( fk ) : Xn+1 → Xn .
228 A. BACKGROUND
We interpret this map X( fk ) as describing the kth n-simplex sitting as a “face on the
boundary” of an n + 1-simplex of X. A similar interpretation applies to the dk .
1.1. Simplicial sets and topological spaces. When working in a homotopical

setting, simplicial sets often provide a more tractable approach than topological
spaces themselves. In this book, for instance, we describe “spaces of field theories”
as simplicial sets. Below, we sketch how to relate these two kinds of objects.
One can use a simplicial set X• as the “construction data” for a topological
space: each element of Xn labels a distinct n-simplex 4n , and the structure maps of
X• indicate how to glue the simplices together. In detail, the geometric realization
is the quotient topological space
 
a
|X• | =  Xn × 4  / ∼
n

n
under the equivalence relation ∼ where (x, s) ∈ Xm × 4m is equivalent to (y, t) ∈

Xn × 4n if there is a map f : [m] → [n] such that X( f ) : Xn → Xm sends y to x and
f∗ (s) = t.
1.1.1 Lemma. Under the Yoneda embedding, [n] defines a simplicial set
∆[n] : [m] ∈ ∆op 7→ ∆([m], [n]) ∈ S ets.
The geometric realization of ∆[n] is the n-simplex 4n (more accurately, it is home-
omorphic to the n-simplex).
In general, every (geometric) simplicial complex can be obtained by the geo-
metric realization of some simplicial set. Thus, simplicial sets provide an efficient
way to study combinatorial topology.
One can go the other way, from topological spaces to simplicial sets: given
a topological space X, there is a simplicial set Sing X, known as the “singular
simplicial set of S .” The set of n-simplices (Sing X)n is simply Top(4n , X), the set
of continuous maps from the n-simplex 4n into X. The structure maps arise from
pulling back along the natural maps of simplices.
1.1.2 Theorem. Geometric realization and the singular functor form an adjunc-
tion
| − | : sSet Top : Sing
between the category of simplicial sets and the category of topological spaces.
Thus, when we construct a simplicial set of BV theories T • , we obtain a topo-
logical space |T • |.
This theorem suggests as well how to transport notions of homotopy to simpli-
cial sets: the homotopy groups of a simplicial set X• are the homotopy groups of its
realization |X• |, and maps of simplicial sets are homotopic if their realizations are.
We would then like to say that these functors | − | and Sing make the categories of
simplicial sets and topological spaces equivalent, in some sense, particularly when
it comes to questions about homotopy type. One sees immediately that some care
must be taken, since there are topological spaces very different in nature from sim-
plicial or cell complexes and for which no simplicial set could provide an accurate
description. The key is only to think about topological spaces and simplicial sets
up to the appropriate notion of equivalence.
Remark: It is more satisfactory to define these homotopy notions directly in terms
of simplicial sets and then to verify that these match up with the topological no-
tions. We direct the reader to the references for this story, as the details are not
relevant to our work in the book. ^
We say a continuous map f : X → Y of topological spaces is a weak equiva-
lence if it induces a bijection between connected components and an isomorphism
πn ( f, x) : πn (X, x) → πn (Y, f (x)) for every n > 0 and every x ∈ X. Let Ho(Top),
the homotopy category of Top, denote the category of topological spaces where
we localize at the weak equivalences. There is a concrete way to think about
this homotopy category. For every topological space, there is some CW complex
weakly equivalent to it, under a zig zag of weak equivalences; and by the White-
head theorem, a weak equivalence between CW complexes is in fact a homotopy
equivalence. Thus, Ho(Top) is equivalent to the category of CW complexes with
morphisms given by continuous maps modulo homotopy equivalence.
Likewise, let Ho(sSet) denote the homotopy category of simplicial sets, where
we localize at the appropriate notion of simplicial homotopy. Then Quillen proved
the following wonderful theorem.
1.1.3 Theorem. The adjunction induces an equivalence
| − | : Ho(sSet) Ho(Top) : Sing
between the homotopy categories. (In particular, they provide a Quillen equiva-
lence between the standard model category structures on these categories.)
This theorem justifies the assertion that, from the perspective of homotopy
type, we are free to work with simplicial sets in place of topological spaces. In
addition, it helpful to know that algebraic topologists typically work with a better
behaved categories of topological spaces, such as compactly generated spaces.
Among simplicial sets, those that behave like topological spaces are known as
Kan complexes or fibrant simplicial sets. Their defining property is a simplicial
analogue of the homotopy lifting property, which we now describe explicitly. The
horn for the kth face of the n-simplex, denoted Λk [n], is the subsimplicial set of
∆[n] given by the union of the all the faces ∆[n − 1] ,→ ∆[n] except the kth. (As a
functor on ∆, the horn takes the [m] to monotonic maps [m] → [n] that do not have
k in the image.) A simplicial set X is a Kan complex if for every map of a horn
Λk [n] into X, we can extend the map to the n-simplex ∆[n]. Diagrammatically, we
can fill in the dotted arrow
Λk [n] /
=X

∆[n]
230 A. BACKGROUND
to get a commuting diagram. In general, one can always find a “fibrant replace-
ment” of a simplicial set X• (e.g., by taking Sing |X• | or via Kan’s Ex∞ functor)
that is weakly equivalent and a Kan complex.
1.2. Simplicial sets and homological algebra. Our other use for simplicial
sets relates to homological algebra. We always work with cochain complexes, so
our conventions will differ from those who prefer chain complexes. For instance,
the chain complex computing the homology of a topological space is concentrated
in non-negative degrees. We work instead with the cochain complex concentrated
in non-positive degrees. (To convert, simply swap the sign on the indices.)
1.2.1 Definition. A simplicial abelian group is a simplicial object A• in the cate-
gory of abelian groups, i.e., a functor A : ∆op → Ab.
By composing with the forgetful functor U : Ab → Set that sends a group to

its underlying set, we see that a simplicial abelian group has an underlying sim-
plicial set. To define simplicial vector spaces or simplicial R-modules, one simply
replaces abelian group by vector space or R-module everywhere. All the work
below will carry over to these settings in a natural way.
There are two natural ways to obtain a cochain complex of abelian groups
(respectively, vector spaces) from a simplicial abelian group. The unnormalized
chains CA• is the cochain complex
A|m| , m ≤ 0
(
(CA• )m =
0, m > 0
with differential
d : (CA• )m → (CA• )m+1
P|m| k
a 7→ k=0 (−1) A• ( fk )(a),
where the fk run over the coface maps from [|m|−1] to [|m|]. The normalized chains
NA• is the cochain complex
|m|−1
\
(NA• )m = ker A• ( fk ),
k=0
where m ≤ 0 and where the fk run over the coface maps from [|m| − 1] to [|m|]. The
differential is A( f|m| ), the remaining coface map. One can check that the inclusion
NA• ,→ CA• is a quasi-isomorphism (in fact, a cochain homotopy equivalence).
Example: Given a topological space X, its singular chain complex C∗ (X) arises as a
composition of three functors in this simplicial world. (Because we prefer cochain
complexes, the singular chain complex is, in fact, a cochain complex concentrated
in nonpositive degrees.) First, we make the simplicial set Sing X, which knows
about all the ways of mapping a simplex into X. Then we apply the free abelian
group functor Z− : S ets → AbG ps levelwise to obtain the simplicial abelian group
Z Sing X : [n] 7→ Z(Top(4n , X)).
Then we apply the unnormalized chains functor to obtain the singular chain com-
plex
C∗ (X) = CZ Sing X.
In other words, the simplicial language lets us decompose the usual construction
into its atomic components. ^
It is clear from the constructions that we only ever obtain cochain complexes
concentrated in nonpositive degrees from simplicial abelian groups. In fact, the
Dold-Kan correspondence tells us that we are free to work with either kind of
object — simplicial abelian group or such a cochain complex — as we prefer.
Let Ch≤0 (Ab) denote the category of cochain complexes concentrated in non-
positive degrees, and let sAbG ps denote the category of simplicial abelian groups.
1.2.2 Theorem (Dold-Kan correspondence). The normalized chains functor
N : sAb → Ch≤0 (Ab)
is an equivalence of categories. Under this correspondence,
πn (A• ) H −n (NA• )
for all n ≥ 0, and simplicial homotopies go to chain homotopies.
Throughout the book, we will often work with cochain complexes equipped
with algebraic structures (e.g., commutative dg algebras or dg Lie algebras). Thank-
fully, it is well-understood how the Dold-Kan correspondence intertwines with the
tensor structures on sAb and Ch≤0 (Ab).
Let A• and B• be simplicial abelian groups. Then their tensor product (A ⊗ B)•
is the simplicial abelian group with n-simplices
(A ⊗ B)n = An ⊗Z Bn .
There is a natural transformation
∇A,B : CA ⊗ CB → C(A ⊗ B),
known as the Eilenberg-Zilber map or shuffle map, which relates the usual tensor
product of complexes with the tensor product of simplicial abelian groups. As
CA• ⊗ CB• is not isomorphic to C(A ⊗ B)• , it is not a strong monoidal functor but
instead a lax monoidal functor. The Eilenberg-Zilber map is, however, always a
quasi-isomorphism, and so it preserves weak equivalences.
1.2.3 Theorem. The unnormalized chains functor and the normalized chains func-
tor are both lax monoidal functors via the Eilenberg-Zilber map.
Thus, with a little care, we can relate algebra in the setting of simplicial mod-
ules with algebra in the setting of cochain complexes.
1.3. References. Friedman (2012) is a very accessible and concrete introduc-
tion to simplicial sets, with lots of intuition and pictures. Weibel (1994) explains
clearly how simplicial sets appear in homological algebra, notably for us, the Dold-
Kan correspondence and the Eilenberg-Zilber map. As usual, Gelfand and Manin
232 A. BACKGROUND
(2003) provides a nice complement to Weibel. The expository article Goerss and
Schemmerhorn (2007) provides a lucid and quick discussion of how simplicial
methods relate to model categories and related issues. For the standard, modern
reference on simplicial sets and homotopy theory, see the thorough and clear Go-
erss and Jardine (2009).
2. Operads and algebras

An operad is a way of describing the essential structure underlying some class
of algebraic objects. For instance, there is an operad Ass that captures the essence
of associative algebras, and there is an operad Lie that captures the essence of Lie
algebras. An algebra over an operad is an algebraic object with that kind of struc-
ture: a Lie algebra is an algebra over the operad Lie. Although their definition can
seem abstract and unwieldy at first, operads provide an efficient language for think-
ing about algebra and proving theorems about large classes of algebraic objects. As
a result, they appear throughout mathematics.
A colored operad is a way of describing a collection of objects that interact al-
gebraically. For example, there is a colored operad that describes a pair consisting
of an algebra and a module over that algebra. Another name for a colored operad
is a symmetric multicategory, which emphasizes a different intuition: it is a gener-
alization of the notion of a category in which we allow maps with multiple inputs
and one output.
In the book, we use these notions in several ways:
• we capture the algebraic essence of the Batalin-Vilkovisky notion of a
quantization via the Beilinson-Drinfeld operad;
• the notion of a prefactorization algebra — perhaps the central notion
in the book — is an algebra over a colored operad made from the open
sets of a topological space; and
• we define colored operads that describe how observables vary under
translation and that generalize the notion of a vertex algebra.
The first use has a different flavor than the others, so we begin here by focusing on
operads with a linear flavor before we introduce the general formalism of colored
operads. We hope that by being concrete in the first part, the abstractions of the
second part will not seem arid.
2.1. Operads. In the loosest sense — encompassing Lie, associative, com-

mutative, and more — an algebra is a vector space A with some way of combining
elements multilinearly. Typically, we learn first about examples determined by a
binary operation µ : A ⊗ A → A such that
• µ has some symmetry under permutation of the inputs (e.g., µ(a, b) =
−µ(b, a) for a Lie algebra), and
• the induced 3-ary operations µ ◦ (µ ⊗ 1) and µ ◦ (1 ⊗ µ) satisfy a linear
relation (e.g., associativity or the Jacobi identity).
2. OPERADS AND ALGEBRAS 233
But we recognize that there should be more elaborate notions involving many dif-
ferent n-ary operations required to satisfy complicated relations. As a basic exam-
ple, a Poisson algebra has two binary operations.
Before we give the general definition of an operad in (dg) vector spaces, we
explain how to visualize such algebraic structures. An n-ary operation τ : A⊗n → A
is pictured as a rooted tree with n labeled leaves and one root.
a1 a2 ... an
µ(a1 , . . . , an )
For us, operations move down the page. To compose operations, we need to spec-
ify where to insert the output of each operation. We picture this as stacking rooted
trees. For example, given a binary operation µ, the composition µ ◦ (µ ⊗ 1) corre-
sponds to the tree
a1 a2 a3
µ 1
µ(µ(a1 , a2 ), a3 )
whereas µ ◦ (1 ⊗ µ) is the tree

a1 a2 a3
1 µ
µ(a1 , µ(a2 , a3 ))
with the first µ on the other side.

We also allow permutations of the inputs, which rearranges the inputs. Thus
the vector space of n-ary operations {τ : A⊗n → A} has an action of the permutation
group S n . We want the permutations to interact in the natural way with composi-
tion.
We now give the formal definition. We will describe an operad in vector spaces
over a field K of characteristic zero (e.g., C or R) and use ⊗ to denote ⊗K . It is
straightforward to give a general definition of an operad in an arbitrary symmetric
monoidal category. It should be clear, for instance, how to replace vector spaces
with cochain complexes.
234 A. BACKGROUND
2.1.1 Definition. An operad O in vector spaces consists of

(i) a sequence of vector spaces {O(n) | n ∈ N}, called the operations,
(ii) a collection of multilinear maps
 
n
X 
◦n;m1 ,...,mn : O(n) ⊗ (O(m1 ) ⊗ · · · ⊗ O(mn )) → O  m j  ,
j=1
called the composition maps,

(iii) a unit element η : K → O(1).
This data is equivariant, associative, and unital in the following way.
(1) The n-ary operations O(n) have a right action of S n .
(2) The composition maps are equivariant in the sense that the diagram
below commutes,
σ⊗σ−1
/
O(n) ⊗ (O(m1 ) ⊗ · · · ⊗ O(mn )) O(n) ⊗ (O(mσ(1) ) ⊗ · · · ⊗ O(mσ(n) ))
◦ ◦
P σ(mσ(1) ,...,mσ(n) ) P
/O

n n
O j=1 m j j=1 m j
where σ ∈ S n acts as a block permutation on the

Pn
j=1 m j inputs, and
the diagram below also commutes,
id ⊗(τ1 ⊗···⊗τn ))
(
O(n) ⊗ (O(m1 ) ⊗ · · · ⊗ O(mn )) O(n) ⊗ (O(mσ(1) ) ⊗ · · · ⊗ O(mσ(n) ))
◦ ◦
P τ1 ⊕···⊕τn P
/O

n n
O j=1 m j j=1 m j
where each τ j is in S m j and τ1 ⊕ · · · ⊕ τn denotes the blockwise permu-

tation in S Pnj=1 m j .
(3) The composition maps are associative in the following sense. Let n,
m1 , . . . , mn , `1,1 , . . . , `1,m1 , `2,1 , . . . , `n,mn be positive integers, and set
Pm
M = nj=1 m j , L j = i=1j ` j,i , and
P
n
X X
N= Lj = ` j,k .
i=1 ( j,k)∈M
Then the diagram
n
N mj
N
!
O(n) ⊗ O(m j ) ⊗ O(` j,k )
j=1 k=1
shuffle
id ⊗(⊗ j ◦) n
N N mj
n N
O(n) ⊗ O(m j ) ⊗ O(` j,k )
j=1 j=1 k=1
n
N
O(n) ⊗ O(L j ) ◦⊗id
j=1
mj
n N
N
O(M) ⊗ O(` j,k )
◦ j=1 k=1
O(N)
commutes.
(4) The unit diagrams commute:

'
O(n) ⊗ K⊗n / O(n) ⊗ O(1)⊗n / O(n)
id ⊗η ⊗n ◦
and

&
K ⊗ O(n) / O(1) ⊗ O(n) / O(n) .
η⊗id ◦
A map of operads f : O → P is a sequence of linear maps
{ f (n) : O(n) → P(n)}
interwining in the natural way with all the structure of the operads.
Example: The operad Com describing commutative algebras has Com(n) K, with
the trivial S n action, for all n. This is because there is only one way to multiply n
elements: even if we permute the inputs, we have the same output. ^
236 A. BACKGROUND
Example: The operad Ass describing associative algebras has Ass(n) = K[S n ], the
regular representation of S n , for all n. This is because the product of n elements
only depends on their left-to-right ordering, not on a choice of parantheses. We
should have exactly one n-ary product for each ordering of n elements. ^
Remark: One can describe operads via generators and relations. The two examples
above are generated by a single binary operation. In Ass, there is a relation between
the 3-ary operations generated by that binary operation — the associativity relation
— as already discussed. For a careful treatment of this style of description, we
direct the reader to the references. ^
We now explain the notion of an algebra over an operad. Our approach is
modeled on defining a representation of a group G on a vector space V as a group
homomorphism ρ : G → GL(V). Given a vector space V, there is an operad EndV
that contains all imaginable multilinear operations on V and how they compose,
just like GL(V) contains all linear automorphisms of V.
2.1.2 Definition. The endomorphism operad EndV of a vector space V has n-ary
operations EndV (n) = Hom(V ⊗n , V) and compositions
P
◦n;m1 ,...,mn : O(n) ⊗ (O(m1 ) ⊗ · · · ⊗ O(mn )) → O nj=1 m j
µn ⊗ (µm1 ⊗ · · · ⊗ µmn ) 7→ µn ◦ (µm1 ⊗ · · · ⊗ µmn )
are simply composition of multilinear maps.
2.1.3 Definition. For O an operad and V a vector space, an algebra over O is a

map of operads ρ : O → EndV .
For example, an associative algebra A is given by specifying a vector space A

and a linear map µ : A⊗2 → A satisfying the associativity relation. This data is
equivalent to specifying an operad map from Ass to EndA .
Alternatively, an algebra V over O is a collection of equivariant maps
ρ(n) : O(n) ⊗ V ⊗n → V
compatible with the structure maps of O. These maps arise from our definition
of an algebra by the hom-tensor adjunction: a map in Hom(O(n), Hom(V ⊗n , V))
yields a map in Hom(O(n) ⊗ V ⊗n , V).
Dually, we consider the notion of a coalgebra over an operad. The coendo-
morphism operad coEndV of a vector space V has n-ary operations
coEndV (n) = Hom(V, V ⊗n ),
and compositions are the obvious multilinear maps. A coalgebra over O is a map
of operads ρ : O → coEndV . Alternatively, it is a collection of equivariant maps
O(n) ⊗ V → V ⊗n with natural compatibilities.
Remark: Dual to these definitions is the notion of a cooperad, which intuitively
encodes the ways something can decompose into constituent pieces. One can then
speak about algebras and coalgebras over a cooperad. ^
2.1.1. References. For a very brief introduction to operads, see the “What is”
column by Stasheff (2004). The article Vallette (2014) provides a nice motivation
and overview for linear operads and their relation to homotopical algebra. For a
systematic treatment with an emphasis on Koszul duality, see Loday and Vallette
(2012). In Costello (n.d.a), there is a description of the basics emphasizing a dia-
grammatic approach: an operad is a functor on a category of rooted trees. Finally,
the book Fresse (n.d.) is wonderful.
2.2. The P0 and BD operads. Two linear operads play a central role in this
book and in the Batalin-Vilkovisky formalism. In physics, there is a basic division
into classical and quantum, and the BV formalism also breaks down along these
lines. The operad controlling the classical BV formalism is the P0 operad, encod-
ing Poisson-zero algebras, equivalently commutative dg algebras with a compatible
Poisson bracket of cohomological degree 1. The operad controlling the quantum
BV formalism is the BD operad, encoding Beilinson-Drinfeld algebras.
2.2.1. P0 . We will define the shifted Poisson operads Pk in terms of their al-
gebras. These operads live in the category of graded vector spaces over any field
of characteristic zero.
2.2.1 Definition. A Pk algebra is a cochain complex A that possesses a commu-
tative product · : A ⊗ A → A of cohomological degree 0 and a Lie bracket
{−, −} : A[k − 1] ⊗ A[k − 1] → A[k − 1] on the shifted complex A[k − 1] sat-
isfying the Poisson relation: for every a ∈ A, the graded linear map {a, −} is a
derivation for the commutative product.
A P1 algebra is precisely an ordinary Poisson dg algebra. The strange-looking

indexing that we use is due to the following fact. By taking homology, the Ek op-
erad in spaces provides an operad in graded vector spaces whose n-ary operations
are H∗ (Ek (n)). This operad is Pk when k ≥ 2, essentially by work of Fred Cohen.
In Volume 2 we will develop the classical and quantum BV formalisms in great
detail, but we can already give a quick motivation for the appearance of P0 alge-
bras. As explained in Section 1, the key motivating example for the BV formalism
is the divergence complex of a measure. The classical limit of the divergence com-
plex is the polyvector fields on a manifold equipped with a derivation so as to form
a commutative dg algebra. We call this complex the functions on the “derived criti-
cal locus” (see Definition 1.1.1). The crucial observation here is that the polyvector
fields possess a natural shifted Poisson bracket known as the Schouten bracket. It
is simply the Lie bracket of vector fields extended antilinearly and compatibly with
the action of vector fields on smooth functions.
2.2.2. BD. Just as we find the P0 operad by examining the classical limit of
divergence complexes and extracting the formal properties, the notion of the BD
operad arises by extracting formal properties of the divergence complex.
2.2.2 Definition. The BD operad is an operad in dg modules over R[[~]], where
~ has cohomological degree 0, whose underlying graded operad agrees with the
P0 operad (tensored up to R[[~]]). The differential on the binary operations · (in
238 A. BACKGROUND
degree 0) and {−, −} in degree 1 satisfy the following relation,

d(·) = ~{−, −},
and determine the differential on the n-ary operations.
Let’s unpack what that condition means for a BD algebra. Let (A, dA ) be a
cochain complex. To equip it with a BD algebra structure is to give a map ρ of dg
operads from BD to EndA . Forgetting the differentials, we see that the underlying
graded vector space A] is thus a P0 algebra. Now let’s examine what it means that
the map ρ(2) : BD(2) → Hom(A⊗A, A) is a cochain map. Let • denote the commu-
tative product on A, which is the element ρ(2)(·), and let {−, −} denote the Poisson
bracket on A, which is the element ρ(2)({−, −}). Then, since ρ(2) commutes with
the differentials, we must have that
dHom (ρ(2)(·)) = dA ◦ • − • ◦ (dA ⊗ 1 + 1 ⊗ dA )
equals
ρ(2)(d(·)) = ~{−, −}.
In other words, we need
dA (x • y) − (dA x) • y − x • (dA y) = ~{x, y}
for any x, y ∈ A. This equation is precisely the standard relation in the Batalin-
Vilkovisky formalism.
We should remark that this operad is quite similar to an operad known as the
Batalin-Vilkovisky operad, except that cohomological degrees are different. In the
mathematics literature, the BV operad typically refers to an operad related to the
little two-dimensional discs operad E2 . There is the potential for confusion here,
because the BV quantization procedure is related not to the E2 operad, but to the
E0 operad. (In the original physics literature, where these relations were first intro-
duced, everything was Z/2-graded and so it is impossible to distinguish between
these structures.) The role of this BD operad was first recognized in Beilinson and
Drinfeld (2004), and so we named it after them.
2.3. Colored operads aka multicategories. We now introduce a natural si-
multaneous generalization of the notions of a category and of an operad. The es-
sential idea is to have a collection of objects that we can “combine” or “multiply.”
If there is only one object, we recover the notion of an operad. If there are no ways
to combine multiple objects, then we recover the notion of a category.
We will give the symmetric version of this concept, just as we did with operads.
Let (C, ) be a symmetric monoidal category, such as (S ets, ×) or (VectK , ⊗K ).
We require C to have all reasonable colimits.
2.3.1 Definition. A multicategory (or colored operad) M over C consists of
(i) a collection of objects (or colors) ObM,
(ii) for every n + 1-tuple of objects (x1 , . . . , xn | y), an object
M(x1 , . . . , xn | y)
in C called the maps from the x j to y,
(iii) a unit element η x : 1C → M(x | x) for every object x.

(iv) a collection of composition maps in C

M(x1 , . . . , xn | y) M(x11 , . . . , x1m1 | x1 ) · · · M(xn1 , . . . , xnmn | xn )
→ M(x11 , . . . , xnmn | y),
(v) for every n + 1-tuple (x1 , . . . , xn | y) and every permutation σ ∈ S n a
morphism
σ∗ : M(x1 , . . . , xn | y) → M(xσ(1) , . . . , xσ(n) | y)
in C.
This data satisfies conditions of associativity, units, and equivariance directly anal-
ogous to that of operads. For instance, given σ, τ ∈ S n , we require
σ∗ τ∗ = (τσ)∗ ,
so we have an analog of a right S n action on maps out of n objects. Each unit η x
is a two-sided unit for composition in M(x | x).
2.3.2 Definition. A map of multicategories (or functor between multicategories)

F : M → N consists of
(i) an object F(x) in N for each object x in M, and
(ii) a morphism
F(x1 , . . . , xn | y) : M(x1 , . . . , xn | y) → N(F(x1 ), . . . , F(xn ) | F(y))
in the category (C, ) for every tuple (x1 , . . . , xn | y) of objects in M
such that the structure of a multicategory is preserved (namely, the units, associa-
tivity, and equivariance).
There are many familiar examples.

Example: Let B be an ordinary category, so that each collection of morphisms
B(x, y) is a set. Then B is a multicategory over the symmetric monoidal category
(S ets, ×) where B(x | y) = B(x, y) and B(x1 , . . . , xn | y) = ∅ for n > 1. ^
Similarly, a K-linear category is a multicategory over the symmetric monoidal
category (VectK , ⊗K ) with no compositions between two or more elements.
Example: An operad O, in the sense of the preceding subsection, is a multicategory
over the symmetric monoidal category (VectK , ⊗K ) with a single object ∗ and
O(∗, . . . , ∗ | ∗) = O(n).
| {z }
n
Replacing VectK with another symmetric monoidal category C, we obtain a defini-
tion for operad in C. ^
Example: Every symmetric monoidal category (C, ) has an underlying multicate-
gory C with the same objects and with maps
C(x1 , . . . , xn | y) = C(x1 · · · xn , y).
240 A. BACKGROUND
When C = VectK , these are precisely all the multilinear maps. ^

For every multicategory M, one can construct a symmetric monoidal enve-
lope SM by forming the left adjoint to the “forgetful” functor from symmetric
monoidal categories to multicategories. An object of SM is a formal finite se-
quence [x1 , . . . , xm ] of colors xi , and a morphism f : [x1 , . . . , xm ] → [y1 , . . . , yn ]
consists of a surjection φ : {1, . . . , m} → {1, . . . , n} and a morphism f j ∈ M({xi }i∈φ−1 ( j) |y j )
for every 1 ≤ j ≤ n. The symmetric monoidal product in SM is simply concatena-
tion of formal sequences.
Finally, an algebra over a colored operad M with values in N is simply a
functor of multicategories F : M → N. When we view O as a multicategory
and use the underlying multicategory VectK , then F : O → VectK reduces to an
algebra over the operad O as in the preceding subsection.
2.3.1. References. For a readable discussion of operads, multicategories, and
different approaches to them, see Leinster (2004). Beilinson and Drinfeld (2004)
develop pseudo-tensor categories — yet another name for this concept — for ex-
actly the same reasons as we do in this book. Lurie (n.d.b) provides a thorough
treatment of colored operads compatible with ∞-categories.
3. Lie algebras and Chevalley-Eilenberg complexes

Lie algebras, and their homotopical generalization L∞ algebras, appear through-
out both volumes in a variety of contexts. It might surprise the reader that we never
use their representation theory or almost any aspects emphasized in textbooks on
Lie theory. Instead, we primarily use dg Lie algebras as a convenient language for
describing deformations of some structure, which is a central theme of Volume 2.
For the purposes of Volume 1, we need to describe standard homological construc-
tions with dg Lie algebras. For the definition of L∞ algebras and the relationship
with derived geometry, see Volume 2.
3.1. A quick review of homological algebra with ordinary Lie algebras.
Let K be a field of characteristic zero. (We always have in mind K = R or C.) A
Lie algebra over K is a vector space g with a bilinear map [−, −] : g ⊗K g → g such
that
• [x, y] = −[y, x] (antisymmetry) and
• [x, [y, z]] = [[x, y], z] + [y, [x, z]] (Jacobi rule).
The Jacobi rule is the statement that [x, −] acts as a derivation. A simple example
is the space of n × n matrices Mn (K), usually written as gln in this context, where
the bracket is
[A, B] = AB − BA,
the commutator using matrix multiplication. Another classic example is given by
Vect(M), the vector fields on a smooth manifold M, via the commutator bracket of
derivations acting on smooth functions.
A module over g, or representation of g, is a vector space M with a bilinear
map ρ : g ⊗K M → M such that
ρ(x ⊗ ρ(y ⊗ m)) − ρ(y ⊗ ρ(x ⊗ m)) = ρ([x, y] ⊗ m).
3. LIE ALGEBRAS AND CHEVALLEY-EILENBERG COMPLEXES 241
Usually, we will suppress the notation ρ and simply write x·m or [x, m]. Continuing
with the examples from above, the matrices gln acts on Kn by left multiplication,
so Kn is naturally a gln -module. Analogously, vector fields Vect(M) act on smooth
functions C ∞ (M) as derivations, and so C ∞ (M) is a Vect(M)-module.
There is a category g−mod whose objects are g-modules and whose morphisms
are the natural structure-preserving maps. To be explicit, a map f ∈ g − mod(M, N)
of g-modules is a linear map f : M → N such that [x, f (m)] = f ([x, m]) for every
x ∈ g and every m ∈ M.
Lie algebra homology and cohomology arise as the derived functors of two
natural functors on the category of g-modules. We define the invariants as the
functor
(−)g : g − mod → VectK
M 7→ Mg
where M g = {m | [x, m] = 0 ∀x ∈ g}. (A nonlinear analog is taking the fixed points
of a group action on a set. )The coinvariants is the functor
(−)g : g − mod → VectK
M 7→ Mg
where Mg = M/gM = M/{[x, m] | x ∈ g, m ∈ M}. (A nonlinear analog is taking
the quotient, or orbit space, of a group action on a set, i.e., the collection of orbits.)
To define the derived functors, we rework our constructions into the setting of
modules over associative algebras so that we can borrow the Tor and Ext functors.
The universal enveloping algebra of a Lie algebra g is
Ug = Tens(g)/(x ⊗ y − y ⊗ x − [x, y])
where Tens(g) = ⊕n≥0 g⊗n denotes the tensor algebra of g. Note that the ideal by
which we quotient ensures that the commutator in Ug agrees with the bracket in g:
for all x, y ∈ g,
x · y − y · x = [x, y],
where · denotes multiplication in Ug. It is straightforward to verify that there is an
adjunction
U : LieAlgK AssAlgK : Forget,
where Forget(A) views an associative algebra A over K as a vector space with
bracket given by the commutator of its product. As a consequence, we can view
g − mod as the category of left Ug-modules, which we’ll denote Ug − mod, without
harm.
Now observe that K is a trivial g-module for any Lie algebra g: x · k = 0 for all
x ∈ g and k ∈ K. Moreover, K is the quotient of Ug by the ideal (g) generated by g
itself, so that K is a bimodule over Ug. It is then straightforward to verify that
Mg = K ⊗Ug M and M g = HomUg (K, M)
for every module M.
3.1.1 Definition. For M a g-module, the Lie algebra homology of M is
H∗ (g, M) = TorUg
∗ (K, M),
242 A. BACKGROUND
and the Lie algebra cohomology of M is

H ∗ (g, M) = Ext∗Ug (K, M).
Notice that H 0 and H0 recover invariants and coinvariants, respectively. There

are concrete interpretations of the lower (co)homology groups, such as that H 1 (g, M)
describes “outer derivations” (i.e., derivations modulo inner derivations) and and
H 2 (g, M) describes extensions of g as a Lie algebra by M.
There are standard cochain complexes for computing Lie algebra (co)homology,
and their generalizations will play a large role throughout the book. The key is to
find an efficient, tractable resolution of K as a Ug module. We use again that K is
a quotient of Ug to produce a resolution:
'
· · · → ∧n g ⊗K Ug → · · · → g ⊗K Ug → Ug −→ K,
where the final map is given by the quotient and the remaining maps are of the
form
n
X
(y1 ∧ · · · ∧ yn )⊗(x1 · · · xm ) 7→ (−1)n−k (y1 ∧ · · · ybk · · · ∧ yn ) ⊗ (yk · x1 · · · xm )
k=1
X
− (−1) j+k−1 ([y j , yk ] ∧ y1 ∧ · · · ybj · · · ybk · · · ∧ yn ) ⊗ (x1 · · · xm ),
1≤ j<k≤n
where the hat ybk indicates removal. As this is a free resolution of K, we use it to
compute the relevant Tor and Ext groups: for coinvariants we have
K ⊗LUg M ' (· · · → ∧n g ⊗K M → · · · → g ⊗K M → M)
and for invariants we have
R HomUg (K, M) ' (M → g∨ ⊗K M → · · · → ∧n g∨ ⊗K M → · · · ),
where g∨ = HomK (g, K) is the linear dual. These resolutions were introduced by
Chevalley and Eilenberg and so their names are attached.
3.1.2 Definition. The Chevalley-Eilenberg complex for Lie algebra homology of
the g-module M is
C∗ (g, M) = (SymK (g[1]) ⊗K M, d)
where the differential d encodes the bracket of g on itself and on M. Explicitly, we
have
X
d(x1 ∧ · · · ∧ xn ⊗ m) = (−1) j+k [x j , xk ] ∧ x1 ∧ · · · xbj · · · xbk · · · ∧ xn ⊗ m
1≤ j<k≤n
n
X
+ (−1)n− j x1 ∧ · · · xbj · · · ∧ xn ⊗ [x j , m].
j=1
We often call this complex the Chevalley-Eilenberg chains.

The Chevalley-Eilenberg complex for Lie algebra cohomology of the g-module
M is
C ∗ (g, M) = (SymK (g∨ [−1]) ⊗K M, d)
3. LIE ALGEBRAS AND CHEVALLEY-EILENBERG COMPLEXES 243
where the differential d encodes the linear dual to the bracket of g on itself and on
M. Fixing a linear basis {ek } for g and hence a dual basis {ek } for g∨ , we have
X X
d(ek ⊗ m) = − ek ([ei , e j ])ei ∧ e j ⊗ m + ek ∧ el ⊗ [el , m]
i< j l
and we extend d to the rest of the complex as a derivation of cohomological degree

1 (i.e., use the Leibniz rule repeatedly to reduce to the case above). We often call
this complex the Chevalley-Eilenberg cochains.
When M is the trivial module K, we simply write C∗ (g) and C ∗ (g). It is impor-
tant for us that C ∗ (g) is a commutative dg algebra and that C∗ (g) is a cocommutative
dg coalgebra. This property suggests a geometric interpretation of the Chevalley-
Eilenberg complexes: under the philosophy that every commutative algebra should
be interpreted as the functions on some “space,” we view C ∗ (g) as “functions on
a space Bg” and C∗ (g) as ”distributions on Bg.” Here we interpret the natural pair-
ing between the two complexes as providing the pairing between functions and
distributions. This geometric perspective on the Chevalley-Eilenberg complexes
motivates the role of Lie algebras in deformation theory, as we explain in the fol-
lowing section.
3.1.1. References. Weibel (1994) contains a chapter on the homological alge-
bra of ordinary Lie algebras, of which we have given a gloss. In Lurie (n.d.a), Lurie
gives an efficient treatment of this homological algebra in the language of model
and infinity categories.
3.2. dg Lie algebras. We now quickly extend and generalize homologically

the notion of a Lie algebra. Our base ring will now be a commutative algebra R
over a characteristic zero field K, and we encourage the reader to keep in mind
the simplest case: where R = R or C. Of course, one can generalize the setting
considerably, with a little care, by working in a symmetric monoidal category (with
a linear flavor); the cleanest approach is to use operads.
3.2.1 Definition. A dg Lie algebra over R is a Z-graded R-module g such that
(1) there is a differential
d d d
· · · → g−1 → g0 → g1 → · · ·
making (g, d) into a dg R-module;
(2) there is a bilinear bracket [−, −] : g ⊗R g → g such that
• [x, y] = −(−1)|x||y| [y, x] (graded antisymmetry),
• d[x, y] = [dx, y] + (−1)|x| [x, dy] (graded Leibniz rule),
• [x, [y, z]] = [[x, y], z] + (−1)|x||y| [y, [x, z]] (graded Jacobi rule),
where |x| denotes the cohomological degree of x ∈ g.
In other words, a dg Lie algebra is an algebra over the operad Lie in the cate-
gory of dg R-modules. In practice — and for the rest of the section — we require
the graded pieces gk to be projective R-modules so that we do not need to worry
about the tensor product or taking duals.
244 A. BACKGROUND
Here are several examples.

(a) We construct the dg analog of gln . Let (V, dV ) be a cochain complex
over K. Let End(V) = ⊕n Homn (V, V) denote the graded vector space
where Homn consists of the linear maps that shift degree by n, equipped
with the differential
dEnd V = [dV , −] : f 7→ dV ◦ f − (−1)| f | f ◦ dV .
The commutator bracket makes End(V) a dg Lie algebra over K.
(b) For M a smooth manifold and g an ordinary Lie algebra (such as su(2)),
the tensor product Ω∗ (M) ⊗R g is a dg Lie algebra where the differential
is simply the exterior derivative and the bracket is
[α ⊗ x, β ⊗ y] = α ∧ β ⊗ [x, y].
We can view this dg Lie algebra as living over K or over the commuta-
tive dg algebra Ω∗ (M). This example appears naturally in the context
of gauge theory.
(c) For X a simply-connected topological space, let g−nX = π1+n (X) ⊗Z Q
and use the Whitehead product to provide the bracket. Then gX is a dg
Lie algebra with zero differential. This example appears naturally in
rational homotopy theory.
The Chevalley-Eilenberg complexes of Lie algebras admit immediate analogues
for dg Lie algebras. The formulas are more complicated, due to the internal dif-
ferential of a dg Lie algebra g and the signs arising from the brackets, which now
depend on the cohomological degree of the inputs. But the constructions are paral-
lel. One starts by providing the natural resolution of K as a Ug-module: view the
resolution given for an ordinary Lie algebra as a chain complex of chain complexes
(i.e., as a double complex) and then take the totalization. Then one uses this res-
olution to compute the derived hom and tensor products over Ug. For a thorough
reference on dg Lie algebras and homological algebra with them, see Félix et al.
(2001).
4. Sheaves, cosheaves, and their homotopical generalizations

Sheaves appear throughout geometry and topology because they capture the
idea of gluing together local data to obtain something global. Cosheaves are an
equally natural construction that are nonetheless used less frequently. We will
give a very brief discussion of these ideas. As we always work with sheaves and
cosheaves of a linear nature, we give definitions in that setting.
4.1. Sheaves.
op
4.1.1 Definition. A presheaf of vector spaces on a space X is a functor F : OpensX →
VectK where OpensX is the category encoding the partially ordered set of open sets
in X (i.e., the objects are open sets in X and there is a map from U to V exactly
when U ⊂ V).
4. SHEAVES, COSHEAVES, AND THEIR HOMOTOPICAL GENERALIZATIONS 245
In other words, a presheaf F assigns a vector space F (U) to each open U and
a restriction map resV⊃U : F (V) → F (U) whenever U ⊂ V. Bear in mind the
following two standard examples. The constant presheaf F = K has K(U) = K
(hence it assigns the same vector space to every open) and its restriction map is
always the identity. The presheaf of continuous functions C X0 assigns the vector
space C X0 (U) of continuous functions from U to R (or C, as one prefers) and the
restriction map resV⊃U consists precisely of restricting a continuous function from
V to a smaller open U.
A sheaf is a presheaf whose value on an open is determined by its behavior on
smaller opens.
4.1.2 Definition. A sheaf of vector spaces on a space X is a presheaf F such that
for every open U and every cover U = {Vi }i∈I of U, we have
 

Y Y 
F (U) → lim  F (Vi ) ⇒ F (Vi ∩ V j ) ,
i∈I i, j∈I
where the map out of F (U) is the product of the restriction maps for the inclusion
of each Vi into U and where, in the limit diagram, the top arrow is restriction for
the inclusion of Vi ∩ V j into Vi and the bottom arrow is restriction for the inclusion
of Vi ∩ V j into V j .
This gluing condition says that an element s ∈ F (U), called a section of F on

U, is given by sections on the cover, (si ∈ F (Vi ))i∈I , that agree on the overlapping
opens Vi ∩ V j . It captures in a precise way how to reconstruct F on an open in
terms of data on a cover.
It is a good exercise to verify that C X0 is always a sheaf and that K is not a sheaf
on a disconnected space.
In this book, our spaces are nearly always smooth manifolds, and most of our
sheaves arise in the following way. Let E → X be a vector bundle on a smooth
manifold. Let E denote the presheaf where E (U) is the vector space of smooth
sections E|U → U on the open set U. It is quick to show that E is a sheaf.
4.2. Cosheaves. We now discuss the dual notion of a cosheaf.
4.2.1 Definition. A precosheaf of vector spaces on a space X is a functor G :
OpensX → Vect. A cosheaf is a precosheaf such that for every open U and every
cover U = {Vi }i∈I of U, we have
 
a a 
colim  G(Vi ∩ V j ) ⇒ G(Vi ) → G(U),
i, j∈I i∈I
where the map into G(U) is the coproduct of the extension maps from the Vi to U
and where, in the colimit diagram, the top arrow is extension from Vi ∩ V j to Vi
and the bottom arrow is extension from Vi ∩ V j to V j .
The crucial example of a cosheaf (for us) is the functor Ec that assigns to the
open U, the vector space Ec (U) of compactly-supported smooth sections of E on
246 A. BACKGROUND
U. If U ⊂ V, we can extend a section s ∈ Ec (U) to a section extU⊂V (s) ∈ Ec (V) on

V by setting it equal to zero on V \ U.
A closely related example is the cosheaf of compactly-supported distributions
dual to the sheaf E of smooth sections of a bundle. When E denotes the sheaf of
fields for a field theory, this cosheaf describes the linear observables on the fields
and, importantly for us, organizes them by their support.
4.3. Homotopical versions. Often, we want our sheaves or cosheaves to take

values in categories of a homotopical nature. For instance, we might assign a
cochain complex to each open set, rather than a mere vector space. In this homo-
topical setting, one typically works with a modified version of the gluing axioms
above. The modifications are twofold:
((i)) we work with all finite intersections of the opens in the cover (i.e., not
just overlaps Vi ∩ V j but also Vi1 ∩ · · · ∩ Vin ), and
((ii)) we use the homotopy limit (or colimit).
The first modification is straightforward — and familiar to anyone who has seen a
Čech complex — but the second is more subtle. (We highly recommend Dugger
(n.d.) for an introduction and development of these notions.)
In our main examples, we give explicit complexes that encode the relevant
information.
For completeness’ sake, we develop the homotopical version explicitly. A pre-
cosheaf with values in dg vector spaces is a functor G : OpensX → dgVect. Given
a cover cover U = {Vi }i∈I of an open U, the Čech complex Č(U, G) is the totalization
of the following double cochain complex: take the cochain complex, with values
in dg vector spaces, whose −nth term is
M
G(∩nj=0 Vi j ),
~i∈I n+1
which has an internal differential inherited from G, and whose “external” differen-
tial is the alternating sum of the structure maps for the inclusion of an n + 1-fold
intersection into an n-fold intersection. We write
 
∞ M
M 
Č(U, G) = Tot⊕  G(∩nj=0 Vi j )[n] ,
n=0 ~i∈I n+1
as a compact description. Observe that there is a canonical map from Č(U, G) to

G(U) given by the sum of the structure maps from G(Ui ) to G(U). (In other words,
one can extend the double complex by inserting G(U) as the first term.)
4.3.1 Definition. A homotopy cosheaf with values in dg vector spaces is a pre-
cosheaf G : OpensX → dgVect such that for every open U and every cover
U = {Vi }i∈I of U, the natural map
'
Č(U, G) → G(U),
We say that a homotopy cosheaf G satisfies Čech (co)descent.

Remark: In our definition above, we use implicitly the fact that the Čech complex
actually provides a representative of the homotopy colimit. For dg vector spaces
(or, more generally, cochain complexes of modules over a ring), one can deduce
this assertion from standard results in homotopical algebra. We provide an argu-
ment in the more general setting of cochain complexes in a Grothendieck abelian
category in Section 5. ^
4.4. Partitions of unity and the Čech complex. We review here why compactly-
supported sections of a vector bundle form both a cosheaf and a homotopy cosheaf.
Our arguments are simple modifications of the standard arguments for why smooth
sections of a vector bundle form both a sheaf and a homotopy sheaf. (See, for
instance, Bott and Tu (1982).)
4.4.1 Definition. A partition of unity on a smooth manifold M subordinate to an
open cover {Ui }i∈I is a collection {ρi }i∈I of nonnegative smooth functions such that
(1) every point x ∈ M has a neighborhood in which all but finitely many ρi
vanish, P
(2) the sum i ρi = 1, and
(3) supp ρi ⊂ Ui for all i ∈ I.
For every open cover, there is a partition of unity subordinate to it.

Let π : V → M be a smooth vector bundle. Let Vc denote the precosheaf
that assigns to the open U ⊂ M the vector space Vc (U) of compactly-supported
smooth sections of V on U. Note that Vc (U) is a module for C ∞ (U). Partitions of
unity then make it easy to verify the cosheaf gluing condition.
4.4.2 Lemma. Vc is a cosheaf.
Proof. Let {Ui }i∈I be an open cover of an open U ⊂ M. Let exti : Vc (Ui ) →
Vc (U) denote the extension-by-zero maps defining the precosheaf structure of Vc .
Then the map
X M
exti : Vc (Ui ) → Vc (U)
i i
induces a map
 
M M 
(†) extI : colim  Vc (Ui ∩ U j ) ⇒ Vc (Ui ) → Vc (U)
i, j∈I i∈I
since the extension from Ui ∩ U j to U factors injectively through both Ui and U j .

To show that Vc is a cosheaf, we need to show that the map extI is an isomor-
phism. A choice of partition of unity allows us to split the map extI , and thus it
becomes straightforward to verify the isomorphism.
Fix a partition of unity {ρi }i∈I subordinate to the cover. Then we obtain a map
ρ̃i : Vc (U) → Vc (Ui )

248 A. BACKGROUND
sending v to ρi v. (Note that here we use the fact that if the support of a section on
U is contained in Ui , it is in the image of exti .) The map
X M
ρ̃i : Vc (U) → Vc (Ui )
i i
has the property that
   
X  X 
 exti  ◦  ρ̃i  = idVc (U)
i i
by the definition of a partition of unity. P
We now want to show that this map i ρ̃i induces an inverse to extI . Note that
the colimit appearing in (†) is a quotient of ⊕i Vc (Ui ). If we denote this quotient
space by Q, then we have an exact sequence
M f M
Vc (Ui ∩ U j ) → Vc (Ui ) → Q → 0,
i, j i
where f denotes the difference P of the two maps in the colimit appearing in (†).
Consider the composition of i ρ̃i followed by the projection to A. To show it is a
left inverse to extI , we
P will P show that every element v in ⊕i Vc (Ui ) is in the same
equivalence class as i ρ̃i ◦ i exti (v), i.e., they map to the same element in Q.
It suffices to verify this for a particularly simple class of elements that span
⊕i Vc (Ui ). Fix an index k ∈ I P v = (vi ) ∈ ⊕i Vc (Ui ) have the property that
and let P
vi = 0 if i , k. Let ṽ = (ṽi ) be i ρ̃i ◦ i exti (v). Then ṽi = ρi vk . We want to show
that v and ṽ map to the same element in Q, so we need their difference to live in
image of f . Consider the element w = (wi j ) in ⊕i, j Vc (Ui ∩U j ) where w Pik = ρi vk for
all i and wi j = 0 for j , k. Then f (w) = ( f (w)i ) is given by f (w)i = j (wi j − w ji )
and so f (w)i = ρi vk for i , k and
X
f (w)k = ρk vk − ρ j v k = ρk v k − v k .
j
Hence f (w) = ṽ − v. As this holds for any k, we know it holds for all vectors v.
The argument above is interesting because it shows not only that Vc is a
cosheaf but exhibits an explicit decomposition of a section into sections on the
cover, by using a partition of unity. In fact, the argument can be carried further.
4.4.3 Lemma. Let U = {Ui }i∈I be an open cover of an open U ⊂ M, and let {ρi }i∈I
be a partition of unity subordinate to the cover. There is a cochain homotopy
equivalence between Č(U, Vc ) and Vc (U) where the cochain map
X
σ= extUi ⊂U : Č(U, Vc ) → Vc (U)
i
is determined by the extension maps. Explicitly, σ vanishes on an element of
Vc (Ui0 ∩ · · · ∩ Uin ) for n > 0 and it sends an element of Vc (Ui ) to its extension by
zero.
This lemma immediately implies the following.
4.4.4 Corollary. Vc is a homotopy cosheaf.

Proof of lemma. Recall that a cochain map induces a cochain homotopy equiv-
alence if and only if the mapping cone admits a contracting cochain homotopy. We
will thus show instead that the “augmented” cochain complex given by the cone
of σ admits a contracting homotopy K. (Those who prefer simplicial construc-
tions will recognize that we are giving a simplicial cochain complex with extra
degeneracies.)
A degree −n element v of Č(U, Vc ) can be expressed in terms of its components
(v~i )~i∈I n+1 , where each v~i ∈ Vc (Ui0 ∩ · · · ∩ Uin ). The cone is obtained by putting
Vc (U) in degree 1 and making σ the differential from degree 0 to degree 1.
The contraction K sends a section v ∈ Vc (U) (i.e., a degree 1 element) to
K(v) = (K(v)i )i∈I where K(v)i = ρi v. (Note that this is precisely the map ρ̃ from the
proof that Vc is a cosheaf.) To a degree −n element v = (v~i )~i∈I n+1 , the contraction
K assigns
K(v)~j = ρ j0 v~j>0 ,
where ~j ∈ I n+2 and ~j>0 = ( j1 , . . . , jn+1 ). (Note that this is a natural extension of the
construction of the element w.) Recall that for ~i ∈ I n+1 , we have
n+1
X X
d(v)~i = (−1) j vc j (k,~i)
j=0 k∈I
where c j (k,~i) = (i0 , . . . , i j−1 , k, i j , . . . , in ). (There is no internal differential here to

worry about, just the Čech differential.) Then we find, with a little manipulation,
that
 
X X n Xn  X
((dK + Kd)v)~i = ρi0  (−1) vc j (k,~i>0 ) +

 j
(−1) vc j (k,~i)>0  +
j
ρc0 (k,~i) vc0 (k,~i) .
k∈I j=0 j=1 k∈I
On the right hand side, the term multiplied by ρi0 vanishes, and the remaining term
simplifies as  
X X 
ρc0 (k,~i) vc0 (k,~i) =  ρk  v~i = v~i .
k∈I k∈I
Hence [d, K] = id.
These arguments are minor modifications of those in Bott and Tu (1982). In
section 8, just after Proposition 8.5, they construct the homotopy equivalence we
use in the case of the de Rham complex. It is quick to see that the construction ap-
plies to any smooth graded vector bundle. In the case of the de Rham complex, they
explicitly provide the cochain homotopy as the “Collating Formula,” Proposition
9.5.
4.4.1. References. Nearly any modern book on differential or algebraic ge-
ometry contains an introduction to sheaves. See, for instance, Griffiths and Harris
(1994) or Ramanan (2005). For a clear exposition (and more) of cosheaves, see
Curry (n.d.). The classic text Bott and Tu (1982) explores in depth the sheaf of
smooth sections of a vector bundle, but typically via the example of the de Rham
250 A. BACKGROUND
complex. For systematic development of sheaves on sites and sheaves of modules

over a sheaf of rings, see Kashiwara and Schapira (2006) or Stacks Project Authors
(2016).
5. Elliptic complexes
Classical field theory involves the study of systems of partial differential equa-
tions (or generalizations), which is an enormously rich and sophisticated subject.
We focus in this book on a tractable and well-understood class of PDE that appear
throughout differential geometry: elliptic complexes. Here we will spell out the
basic definitions and some examples that suffice for our work in this book.
5.1. Ellipticity. We start with a local description of differential operators be-

fore homing in on elliptic operators. Let U be an open set in Rn . A linear differen-
tial operator is an R-linear map L : C ∞ (U) → C ∞ (U) of the form
aα (x)∂α f (x),
X
L( f ) =
α∈Nn
where we use the multi-index α = (α1 , . . . , αn ) to efficiently denote

∂α1 ∂αn
∂α = α1 · · · ,
∂x1 ∂xnαn
where the coefficients aα are smooth functions,
P and where only finitely many of
the aα are nonzero functions. We call |α| = j α j the order of the index, and thus
the order of L is the maximum orderP|α| among the nonzero coefficients aα . (For
example, the order of the Laplacian j ∂2 /∂x2j is two.) The principal symbol (or
leading symbol) of a kth order differential operator L is the “fiberwise polynomial”
aα (x)ik ξα
X
σL (ξ) =
|α|=k
obtained by summing over the indices of order k and replacing the partial deriv-
ative ∂/∂x j by variables iξ j , where i is the usual square root of one. (It is the
standard convention to include the factor of i, due to the role of Fourier transforms
in motivating many of these constructions and definitions.) It is natural to view the
principal symbol as a function on the cotangent bundle T ∗ U that is a homogeneous
polynomial of degree k along the cotangent fibers, where ξ j is the linear functional
dual to dx j . The principal symbol controls the qualitative behavior of L. It also be-
haves nicely under changes of coordinates, transforming as a section of the bundle
Symk (T U ) → U.
Remark: Elsewhere in the text, we talk about differential operators in a more ab-
stract, homological setting. For instance, we say the BV Laplacian is a second-
order differential operator. Our use of the differential operators in this homologi-
cal setting is inspired by their use in analysis, but the definitions are modified, of
course. For instance, we view the odd directions as geometric in the BV formalism,
so the BV Laplacian is second-order. ^
5. ELLIPTIC COMPLEXES 251
It is straightforward to extend these notions to a smooth manifold and to smooth

sections of vector bundles on that manifold. Let E → X be a rank m vector bun-
dle and let F → X be a rank n vector bundle. Let E denote the smooth global
sections of E and let F denote the smooth global sections of F. A differential
operator from E to F is an R-linear map L : E → F such that for a local choice
of coordinates on X and trivializations of E and F,
aα (x)∂α s j (x),
X
ij
(Ls)i =
α∈Nn
where s = (s1 , . . . , sm ) is a section of E (U) C ∞ (U)m on a small neighborhood
U in X, Ls = ((Ls)1 , . . . , (Ls)n ) is a section of F (U) C ∞ (U)n on that small
ij
neighborhood, and where the aα (x) are smooth functions. In other words, locally
we have a matrix-valued differential operator. The principal symbol σL of a kth
order differential operator L defines a section of the bundle Symk (T X )⊗E ∨ ⊗F → X.
Thus, by pulling back along the canonical projection π : T ∗ X → X, we can view the
principal symbol as a map of vector bundles σL : π∗ E → π∗ F over the cotangent
bundle T ∗ X.
We now introduce one of our key definitions.
5.1.1 Definition. An elliptic operator L : E → F is a differential operator whose
principal symbol σL : π∗ E → π∗ F is an isomorphism of vector bundles on T ∗ X \ X,
the cotangent bundle with the zero section removed.
This definition says that the principal symbol is an invertible linear operator
after evaluating at any nonzeroP covector. As an example, consider the Laplacian
on Rn : its principal symbol is ξ2j , which only vanishes when all the ξ j are zero.
5.2. Functional analytic consequences. Ellipticity is purely local and is an
easy property to check in practice. Globally, it has powerful consequences, of
which the following is the most famous.
5.2.1 Theorem. For X a closed manifold (i.e., compact and boundaryless), an
elliptic operator Q : E → F is Fredholm, so that its kernel and cokernel are
finite-dimensional vector spaces.
Thus, an elliptic operator on a closed manifold is invertible up to a finite-
dimensional “error.” Moreover, there is a rich body of techniques for constructing
these partial inverses, especially for classical operators such as the Laplacian.
The notions from above extend naturally to cochain complexes. A differential
complex on a manifold X is a Z-graded vector bundle ⊕n E n → X (with finite total
rank) and a differential operator Qn : E n → E n+1 for each integer n such that
Qn+1 ◦ Qn = 0. We typically denote this by (E , Q). There is an associated principal
symbol complex (π∗ E, σQ ) on T ∗ X by taking the principal symbol of each operator
Qn .
5.2.2 Definition. An elliptic complex is a differential complex whose principal
symbol complex is exact on T ∗ X \ X (i.e., the cohomology of the symbol complex
vanishes away from the zero section of the cotangent bundle).
252 A. BACKGROUND
Every elliptic operator Q : E → F defines a two-term elliptic complex

Q
··· → 0 → E → F → 0 → ··· .
The other standard examples of elliptic complexes are the de Rham complex (Ω∗ (X), d)
of a smooth manifold and the Dolbeault complex (Ω0,∗ (X), ∂) ¯ of a complex mani-
fold.
The analog of the Fredholm result above is the following, sometimes known
as the formal Hodge theorem (by analogy to the Hodge theorem, which is usually
focused on the de Rham complex).
5.2.3 Theorem. Let (E , Q) be an elliptic complex on a closed manifold X. Then
the cohomology groups H k (E , Q) are finite-dimensional vector spaces.
In fact, the proof does something better: it constructs a continuous cochain
homotopy equivalence between the elliptic complex and its cohomology. Below,
we will explain some ingredients of the proof, emphasizing aspects that play a role
in our constructions of observables.
5.3. Parametrices. In proving these theorems, one constructs a “partial in-
verse” to Q with nice, geometric properties. We will not give a full definition here,
because we do not want to delve into pseudodifferential operators, but will state
the important properties. We will invoke the analysis as a black box.
Recall that every continuous linear operator F : E → F between smooth
sections of a vector bundles possesses a Schwartz kernel, KF , a section of the
bundle F (E ∨ ⊗ Dens) on X × X that is smooth along the first copy of X (i.e.,
in the F direction) and distributional along the second copy of X (i.e., in the E
direction). Then Z
F(s)(x) = KF (x, y)s(y).
y∈X
The notation is a suggestive way of writing the composition
F⊗id id ⊗ ev
E −−−−→ F b
⊗π E ∗b
⊗π E −−−−−→ F ,
s 7→ F⊗s 7→ F(s)
where b ⊗π denotes the completed projective tensor product, which is the natural
tensor to use in this context. This notation emphasizes the formal similarity to
matrix multiplication: the position y provides an index for a kind of “basis,” s(y)
denotes the coefficient for the vector s of that basis element, KF (x, y) describes the
matrix coefficient, and we sum (i.e., integrate) over a common index y to compute
the multiplication by the linear operator F.)
If the kernel KF is a smooth over all of X × X, then F is called a smoothing
operator. For X a closed manifold, a smoothing operator F is compact (when
viewed as an operator between Sobolev space completions of the smooth sections).
For Q : E → F an elliptic operator, a parametrix is an operator P : F → E
satisfying
((i)) idE −PQ = S for some smoothing operator S ,
((ii)) idF −QP = T for some smoothing operator T , and
5. ELLIPTIC COMPLEXES 253
((iii)) the Schwartz kernel of P is smooth away from the diagonal X ⊂ X × X.

Thus, a parametrix P for Q is a partial, or approximate, inverse to Q with a partic-
ularly nice Schwartz kernel. The theory of pseudodifferential operators provides a
construction of such a parametrix for any elliptic operator.
Armed with a parametrix P, we deduce that Q is Fredholm, as follows. First,
note that the Fredholm operators are stable under compact perturbations: if T is
Fredholm and C is compact, then T + C is also Fredholm. In particular, id +C is
always Fredholm. Second, because idE −S is Fredholm, we see that Ker(PQ) is
finite-dimensional and hence its subspace Ker(Q) is finite-dimensional. Similarly,
we deduce that Coker(Q) is finite-dimensional.
5.4. Formal Hodge theory. We want to explain how to generalize the parametrix
approach to elliptic complexes. The essential idea to make a retraction of the ellip-
tic complex E onto its cohomology H ∗ E , viewed as a complex with zero differen-
tial:
π /
η
8E o ι
H∗E .
Recall that a retraction means that ι includes the cohomology as a subcomplex, π

projects away everything but cohomology, and η is a degree −1 map on E , such
that these maps satisfy
idH ∗ E = π ◦ ι and idE −ι ◦ π = [Q, η] = Qη + ηQ.
Thus, a retraction is a homological generalization of a partial inverse.

We want the retractions for elliptic complexes to have two important properties.
First, we require all the maps above to be continuous with respect to the usual
Fréchet topologies. Second, the operator C = ι ◦ π should be a cochain map that is
smoothing, which ensures that idE −C is Fredholm.
The existence of such a retraction implies Theorem 5.2.3, for the following rea-
son. The operator C : E → E sends any cocycle s to a distinguished representative
C(s) of its cohomology class. In fact, C 2 = C, so it is a projection operator whose
image is isomorphic to the cohomology of E . Moreover, C annihilates exact cocy-
cles (since π does) and elements that are not cocycles. As idE −C is an idempotent
Fredholm operator, we know the kernel of this operator is finite-dimensional, and
so the image of C is finite-dimensional.
Finding the necessary retraction η exploits the existence of parametrices for
elliptic operators. In the situations relevant to this book, the general idea is simple.
First, one finds a differential operator Q∗ of degree −1 such that the commutator
D = [Q, Q∗ ] is a generalized Laplacian, meaning its principal symbol looks like
that of a Laplacian. (One way to find Q∗ is to pick inner products on the bundles
E j (hermitian, if complex bundles) and a Riemannian metric on X. This provides
an adjoint to Q, i.e., a differential operator Q∗ of degree −1.) In our setting, we
require the existence of such a Q∗ . This operator D is Fredholm by Theorem 5.2.1,
so we know it has finite-dimensional kernel.
254 A. BACKGROUND
Second, let G denote the parametrix for D (where G is for “Green’s function”).
Set η = Q∗G. We need to show it a contracting homotopy with the desired proper-
ties. Let
id −DG = S and id −GD = T,
where S and T are smoothing endomorphisms of cohomological degree 0. Note
that
QD = QQ∗ Q = DQ,
so
G(QD)G = G(DQ)G =⇒ (GQ)(id −S ) = (id −T )(QG).
Hence we find GQ = QG − U, where U = QS − T QG is smoothing because Q is
a differential operator. In consequence,
[Q, η] = QQ∗G + Q∗GQ
= QQ∗G + Q∗ QG − Q∗ U
= DG − Q∗ U
= id −(S + Q∗ U).
The final term in parantheses is smoothing. We have thus verified that η has the
desired properties.
5.5. References. As usual, Atiyah and Bott (1967) explain beautifully the es-
sential ideas of elliptic complexes and pseudodifferential techniques and show how
to use them efficiently. For an accessible development of the analytic methods in
the geometric setting, we recommend Wells (2008). The full story and much more
is available in the classic works of Hörmander (2003).
APPENDIX B
Functional analysis
1. Introduction
The goal of this appendix is to introduce several types of vector spaces and
explain how they are related. In the end, most of the vector spaces we work with
— which are built out of smooth or distributional sections of vector bundles —
behave nicely in whichever framework one chooses to use, but it is important to
have a setting where abstract constructions behave well. In particular, we will
do homological algebra with infinite-dimensional vector spaces, and that requires
care. Below, we introduce the underlying “functional analysis” that we need (i.e.,
we describe here just the vector spaces and discuss the homological issues in a
separate appendix). For a briefer overview, see Section 5.
There are four main categories of vector spaces that we care about:
• LCTVS, the category of locally convex Hausdorff topological vector
spaces,
• BVS, the category of bornological vector spaces,
• CVS, the category of convenient vector spaces, and
• DVS, the category of differentiable vector spaces.
The first three categories are vector spaces equipped with some extra structure,
like a topology or bornology, satisfying some list of properties. The category DVS
consists of sheaves of vector spaces on the site of smooth manifolds, equipped with
some extra structure that allows us to differentiate sections (hence the name).
The main idea is that DVS provides a natural place to compare and relate vec-
tor spaces that arise in differential geometry and physics. As a summary of the
relationships between these categories, we have the following diagram of functors:
incβ
BVS / LCTVS .
c∞ di fβ di ft
%
CVS / DVS
di fc
All the functors into DVS preserve limits. The functors incβ and c∞ out of BVS
are left adjoints. The functors into DVS all factor as a right adjoint followed by the
functor di fβ . For example, di ft : LCTVS → DVS is the composition di fβ ◦ born,
where the bornologification functor born : LCTVS → BVS is the right adjoint to
the inclusion incβ : BVS → LCTVS.
255
256 B. FUNCTIONAL ANALYSIS
For our work, it is also important to understand multilinear maps between vec-
tor spaces (in the categories mentioned above). Of course, it is pleasant to have
a tensor product that represents bilinear maps. In the infinite-dimensional setting,
however, the tensor product is far more complicated than in the finite-dimensional
setting, with many different versions of the tensor product, each possessing various
virtues and defects. We will discuss certain natural tensor products that appear on
the categories above and how the functors intertwine with them.
2. Differentiable vector spaces

We begin with a sheaf-theoretic approach, introducing the category of differen-
tiable vector spaces. This category will provide a setting in which we can compare
other approaches.
2.0.1 Definition. The smooth site Mfld will denote the category where an object
is a smooth manifold, a morphism is a smooth map, and a cover f : X → Y is a
surjective local diffeomorphism.
Note that “manifold” means Hausdorff, second-countable, and without bound-

ary.
Our basic objects of study will be sheaves on this site. Thus, from the be-
ginning, everything is built to vary nicely in families over smooth manifolds. For
instance, a smooth vector space is a sheaf of vector spaces on Mfld.
Remark: As we develop the relevant facts about differentiable vector spaces in this
section, it is important to bear in mind some useful facts about (pre)sheaves on a
site. For a development of these assertions and much more, see Chapters 17 and 18
of Kashiwara and Schapira (2006) or Chapter 7 of Stacks Project Authors (2016).
Suppose C is a category containing limits and colimits, and let X be a site.
Then the inclusion of sheaves on X with values in C into presheaves is a right
adjoint whose left adjoint is the sheafification functor. When C is the category of
sets Set, then sheafification also preserves finite limits, in addition to all colimits.
Suppose that there is a “forgetful functor F : C → Set that is faithful, reflects
isomorphisms, and commutes with limits and filtered colimits. (For example, con-
sider the usual forgetful functor from vector spaces to sets.) Then a presheaf V
with values in C is a sheaf with values in C if and only if F(V) is a sheaf with
values in Set, and sheafification commutes with F. These results make it possible
to work easily with “sheaves with algebraic structure” as simply sheaves of sets
whose openswise algebraic structure intertwines with the structure maps. ^
Our focus here will be on smooth vector spaces with some extra structure that
makes our main constructions possible.
2.0.2 Definition. Let C ∞ denote the smooth vector space that assigns C ∞ (X) to
each smooth manifold X. A C ∞ -module is a smooth vector space V equipped
with the structure of a module over C ∞ . That is, it is equipped with a map of
sheaves · : C ∞ × V → V, called the scalar multiplication map, satisfying the usual
compatibilities with the addition map + : V × V → V.
2. DIFFERENTIABLE VECTOR SPACES 257
A map of C ∞ -modules is map of sheaves that respects the structure as a C ∞ -

module. We denote the category of C ∞ -modules by ModC ∞ .
As suggestive notation, we will typically denote the vector space V(X) of sec-
tions on X of V by C ∞ (X, V), to emphasize that we think of these as the smooth
functions from X to the smooth vector space V. A map φ : V → W then con-
sists of a map φX : C ∞ (X, V) → C ∞ (X, W) for every manifold X such that for
every smooth map f : X → Y, we have fW∗ ◦ φY = φX ◦ fV∗ , where we use
fV∗ : C ∞ (Y, V) → C ∞ (X, V) to denote the natural pullback map.
Remark: Given a presheaf F on Mfld that is a module over C ∞ (in the category of
presheaves of vector spaces), the sheafification is also naturally a C ∞ -module, and
this sheafification can be computed as simply a presheaf of sets or vector spaces.
Moreover, the forgetful functor from C ∞ -modules to smooth vector spaces is right
adjoint to the tensor product C ∞ ⊗−, i.e., base-change behaves as expected from or-
dinary algebra in this sheaf-theoretic context. See Kashiwara and Schapira (2006)
or Stacks Project Authors (2016) for more. ^
All the natural vector bundles in differential geometry provide examples of
C ∞ -modules. For instance, consider differential forms. Let Ωk denote the sheaf
that assigns Ωk (X) to each smooth manifold X. Given a C ∞ -module V, we define
the k-forms with values in V as
Ωk (−, V) := Ωk ⊗C ∞ V,
so that on a manifold X, we have
Ωk (X, V) = Ωk (X) ⊗C ∞ (X) C ∞ (X, V).
With this definition in hand, we can define our main object of interest.
2.0.3 Definition. A differentiable vector space is a C ∞ -module V equipped with a
flat connection
∇X,V : C ∞ (X, V) → Ω1 (X, V)
for every smooth manifold X such that pullback commutes with the connections,
f ∗ ◦ ∇Y,V = ∇X,V ◦ f ∗ ,
for every smooth map f : X → Y.
To say that ∇X,V is a connection means that it satisfies the Leibniz rule,
∇X,V ( f · v) = (d f )v + f ∇X,V v,
where f ∈ C ∞ (X) and v ∈ C ∞ (X, V). To say that it is flat means that the curvature
F(∇X,V ) = (∇X,V )2 : C ∞ (X, V) → Ω2 (X, V)
vanishes.
The flat connection ∇X,V thus allows us to differentiate sections of V on X. If
X is a vector field on X, we define
X(v) := X, ∇X,V v ∈ C ∞ (X, V),
where h−, −i denotes the C ∞ (X)-pairing between vector fields and 1-forms
h−, −i : T (X) × Ω1 (X, V) → C ∞ (X, V).
Moreover, because the curvature vanishes, the Lie bracket of vector fields goes to
the commutator:
[X, Y](v) = X(Y(v)) − Y(X(v)).
Thus we can take as many derivatives as we would like.
This ability to do calculus with differentiable vector spaces is crucial to many
of our constructions. See, for instance, the sections on translation-invariant factor-
ization algebras, holomorphically translation-invariant factorization algebras, and
equivariant factorization algebras.
2.0.4 Definition. A map of differentiable vector spaces φ : V → W is a map of
C ∞ -modules such that
∇X,W ◦ φX = (idΩ1 (X) ⊗φX ) ◦ ∇X,V
for every smooth manifold X.
We denote the category of differentiable vector spaces by DVS. We denote the
vector space of morphisms from V to W by DVS(V, W)
In the remainder of this section, we will introduce a large class of examples

of differentiable vector spaces, which we use throughout the book. We will also
discuss categorical properties of DVS, such as the existence of limits and colimits.
Finally, we describe the natural multicategory structure arising from multilinear
maps.
2.1. Differentiable vector spaces from sections of a vector bundle. Let M

be a smooth manifold and p : E → M a smooth vector bundle. Let Γ(M, E) denote
the vector space of smooth sections of E. Let Γc (M, E) denote the vector space
of compactly-supported smooth sections of E. There is a natural way to construct
differentiable vector spaces whose value on a point recovers these familiar vector
spaces.
2.1.1 Definition. For a smooth manifold X, let π M : X × M → M denote the
projection map. Let C ∞ (X, E (M)) denote the vector space of smooth sections of
π∗M E over X × M. This smooth vector space E (M) is naturally a C ∞ -module, as
C ∞ (X, E (M)) is a module over C ∞ (X).
We say s ∈ C ∞ (X, E (M)) has proper support over X if the composition
πX
Supp(s) ,→ X × M −→ X
is a proper map. Let C ∞ (X, Ec (M)) denote the vector space of smooth sections with
proper support over X. Then Ec (M) is a C ∞ -module.
We still need to equip these sheaves with flat connections to make them differ-
entiable vector spaces. There is a natural choice, due to the fact that the pullback
bundle π∗M E is trivial in the X-direction.
2.1.2 Definition. Let X be a smooth manifold. Equip the pullback bundle π∗M E on
X × M with the natural flat connection along the fibers of the projection map π M
(i.e., we only differentiate in X-directions). We thus obtain a map
∇X,E : Γ(X × M, π∗M E) → Γ(X × M, T X∗ π∗M E),
or equivalently a map
∇X,E : C ∞ (X, E (M)) → Ω1 (X, E (M)).
This map defines a flat connection on E (M) and hence gives it the structure of a
differentiable vector space.
As this flat connection does not increase support, it preserves the subspace
Ec (M) of sections with proper support over X, and it gives Ec (M) the structure of
a differentiable vector space.
Remark: Typically, one works with Γ(M, E) as a complete locally convex topolog-
ical vector space, via the Fréchet topology. Below, in Section 3, we will explain a
result of Kriegl and Michor that shows that every locally convex topological vector
space V naturally produces a differentiable vector space. Moreover, Γ(M, E) goes
to E under this functor. Hence, the differentiable vector space E (M) arises from
the standard topology on Γ(M, E). (An identical comment applies to Γc (M, E).) ^
2.1.1. We are also interested in distributional sections of a vector bundle p :
E → M. We will show that these also form a differentiable vector space, after
setting up the preliminaries about distributions.
Let D(M) denote the vector space of distributions on the smooth manifold
M. That is, D(M) is the continuous linear dual to the vector space Cc∞ (M). Let
Dc (M) denote the vector space of compactly supported distributions on M, i.e., the
continuous linear dual of C ∞ (M). This space is also a C ∞ (M)-module.
2.1.3 Definition. Let D(M) denote the C ∞ -module whose smooth sections C ∞ (X, D(M))
on the manifold X are the continuous linear maps from Cc∞ (M) to C ∞ (X).
Similarly, let Dc (M) denote the C ∞ -module whose smooth sections on the man-
ifold X, denoted C ∞ (X, Dc (M)), are the continuous linear maps from C ∞ (M) to
C ∞ (X).
Note that when X is a point, we recover the usual notion of a distribution. The
definition above arises by asking that a smooth map φ from X to D(M) correspond
to a smooth family of distributions {φ x } x∈X on M. If one evaluates on f ∈ Cc∞ (M),
then {φ x ( f )} x∈X should be a smooth function X.
We now equip these C ∞ -modules with a natural flat connection.
2.1.4 Definition. The vector fields TX on X act in a natural way on the vector
space C ∞ (X, D(M)) by
X · φ : f 7→ X(φ( f )),
where f ∈ Cc∞ (M), φ ∈ C ∞ (X, D(M)), and X ∈ TX . This action is compatible

with the Lie bracket of vector fields, and hence it makes C ∞ (X, D(M)) into a rep-
resentation of TX . This data is equivalent to equipping C ∞ (X, D(M)) with a flat
connection and hence the structure of a differentiable vector space.
This action clearly preserves the subspace C ∞ (X, Dc (M)), and hence gives it
the structure of a differentiable vector space.
2.1.2. Let p : E → M be a smooth vector bundle. We wish to construct a

differentiable vector space of distributional sections of E.
Note that D(M) is a module over C ∞ (M). The vector space of distributional
sections of E is
Γ(M, E) ⊗C ∞ (M) D(M),
using the algebraic tensor product. We are simply allowing sections of E with
distributional coefficients. We extend this object to a C ∞ -module by using the
same logic as we used for D(M).
Recall that E denotes the differentiable vector space associated to the smooth
sections Γ(M, E).
2.1.5 Definition. Let E (M) denote the C ∞ -module where
C ∞ (X, E (M)) := C ∞ (X, E (M)) ⊗C ∞ (X×M) C ∞ (X, D(M)).
We call this sheaf the distributional sections of the vector bundle E. Similarly, the
compactly-supported distributional sections of E is the C ∞ -module E¯c (M) where
C ∞ (X, E c (M)) := C ∞ (X, Ec ) ⊗C ∞ (X, Cc∞ (M)) C ∞ (X, Dc (M)).
Note that setting X to be a point, we recover the original notion of the distribu-
tional sections.
There is a natural flat connection on E (M) and E c (M) arising from a natural
action of the vector fields TX on these spaces, just as in the case of D(M) and
Dc (M). Hence, these are differentiable vector spaces.
2.1.3. Holomorphic vector bundles and sections. Let M be a complex mani-
fold and p : E → M a holomorphic vector bundle. The holomorphic sections also
form a differentiable vector space; we will show this in the simplest case, as the
general case is completely parallel.
2.1.6 Definition. For M a complex manifold and X a smooth manifold, let f ∈
C ∞ (X, O(M)) denote a smooth function on the product manifold X × M such that
f x (m) := f (x, m) is holomorphic on M for every x ∈ X.
Note that C ∞ (−, O(M)) is a subsheaf of C ∞ (−, C ∞ (M)), using the definition
of the differentiable vector space C ∞ (M) from above. It is straightforward to see
that the flat connection on C ∞ (X, C ∞ (M)) preserves the subspace C ∞ (X, O(M)) for
every X. Hence C ∞ (−, O(M)) is a differentiable vector space that we will denote
O(M).
Below, in Proposition 3.0.6, we explain how to obtain a differentiable vector

space from each locally convex topological vector space. The construction there
applied to O(M) coincides with this construction.
2.2. Stalks and local properties. Let V be a differentiable vector space. Re-
stricting to Rn , we obtain a sheaf V|Rn on Rn . We can thus define the n-dimensional
stalk of V:
Stalkn (V) = colim0∈U⊂Rn V(U).
The colimit above is taken over open subsets of Rn containing the origin and is
computed in the category of vector spaces.
Note that the stalk of V at a point in any manifold can be defined in the same
way, but the stalk at a point in a n-dimensional manifold is the same as the stalk
at the origin in Rn . Because of this “uniformity” of differentiable vector spaces,
a local property (i.e., a property verified stalkwise) only needs to be checked in
countably many cases. As an example, consider how convenient the standard cri-
terion for the exactness of a sequence becomes in our setting.
2.2.1 Lemma. A sequence of differentiable vector spaces
0→A→B→C→0
is exact if and only if, for all n, the sequence
0 → Stalkn (A) → Stalkn (B) → Stalkn (C) → 0
of vector spaces is exact.
More generally, it is a standard fact that the functor of taking stalks preserves
colimits and finite limits. The following corollary will be useful for us.
2.2.2 Corollary. A map f : A → B of differentiable vector spaces is an isomor-
phism if and only if, for all n, the map Stalkn ( f ) : Stalkn (A) → Stalkn (B) is an
isomorphism.
2.3. Categorical properties. The category DVS is well-behaved in the sense
that all the usual constructions make sense. For example, we have the following.
2.3.1 Lemma. DVS contains all finite products.
Q Given a finite collection {Vα }α∈A of differentiable vector spaces, the

Proof.
product α∈A Vα is the functor
 
 Y  Y ∞
∞
X 7→ C X, Vα  = C (X, Vα ),
α∈A α∈A
where on the right hand side we use the product as vector spaces.
Note that finite products coincide with finite coproducts in C ∞ -modules, and
the tensor product commutes with coproducts so that
 
Y  Y 1
Ω ⊗C ∞ 
1
Vα  Ω ⊗C ∞ Vα
α∈A α∈A
for A a finite set. Hence the product of the connections for each Vα provides the
desired connection
   
 Y   Y 
∞
∇X, Vα : C X,
Q Vα  → Ω X,
1
Vα  .
α∈A α∈A
It is straightforward to check that this construction satisfies the universal property
of a product.
Note in the proof that we used crucially the tensoring with Ω1 commutes with
finite products of C ∞ -modules. There is no reason to expect tensoring commutes
with infinite products. Below we give an indirect proof that DVS contains infinite
products as well.
We now provide kernels.
2.3.2 Lemma. For any map φ : V → W of differentiable vector spaces, the kernel,
given by
ker φ : X 7→ ker φ(X) : C ∞ (X, V) → C ∞ (X, W) ,

is a differentiable vector space by restricting the connection ∇X,V to ker φ(X).

Proof. The definition of ker φ clearly provides a C ∞ -module, since φ(X) is
C ∞ (X)-linearfor every smooth manifold X. It is also manifestly the kernel in the
category of C ∞ -modules.
As φ commutes with the flat connection, we see that if v ∈ ker φ(X) ⊂ C ∞ (X, V),
then
φ(∇X,V v) = ∇X,W (φ(v)) = 0,
so hX, ∇X,V vi = 0 for any vector field X on X.
2.3.3 Corollary. DVS contains all finite limits.
Proof. Every finite limit can be constructed out of finite products and equaliz-
ers. Since we are in a linear setting, we can construct an equalizer as a kernel.
Note that one reason this situation for limits is simple is that limits of sheaves
are simply computed objectwise (i.e., separately on each manifold X ∈ Mfld), and
we know how to compute limits of vector spaces.
Colimits are a bit more complicated.
2.3.4 Lemma. DVS contains all coproducts.
Proof. Given a collection {Vα }α∈A of differentiable vector spaces, the coprod-
L
uct α∈A Vα is the sheafification of the functor
M
X 7→ C ∞ (X, Vα ),
α∈A
where on the right hand side we use the coproduct as vector spaces. (We use
the notation ⊕ for the coproduct, as usual for vector spaces.) This sheaf is the
coproduct in C ∞ -modules, since the presheaf defined above is the coproduct in
presheaves of C ∞ -modules and sheafification is a left adjoint.
The connection
M M
∇X,L Vα : C ∞ (X, Vα ) → Ω1 (X, Vα )
α∈A α∈A
is the unique connection that restricts to ∇X,Vα for each Vα . It is manifest that any
cocone in DVS for this collection {Vα } factors through the differentiable vector
space just constructed.
Remark: For finite coproducts, the presheaf formula is already a sheaf and provides
the coproduct in C ∞ -modules. The natural connection makes it the coproduct of
differentiable vector spaces. ^
2.3.5 Lemma. For any map φ : V → W of differentiable vector spaces, the coker-
nel coker φ is given by sheafifying the presheaf
] φ : X 7→ coker φ(X) : C ∞ (X, V) → C ∞ (X, W) ,
coker
This C ∞ module is a differentiable vector space by a natural connection induced
from the connection ∇X,W .
] φ. (Recall Re-
Proof. Let Cφ denote the sheafification of the presheaf coker
mark 2: the sheafification can be as a presheaf of sets or vector spaces or C ∞ -
modules, as they all provide the same object, after suitable forgetting of module
structure.) This sheaf Cφ provides the cokernel in C ∞ -modules.
We need to produce the flat connection on Cφ . As Ω1 is a projective C ∞ -
module, it is a flat C ∞ -module so that
1⊗φ 1⊗qφ
Ω1 ⊗C ∞ V −→ Ω1 ⊗C ∞ W −→ Ω1 ⊗C ∞ Cφ −→ 0
is exact. As flat connections are C-linear but not C ∞ -linear (i.e., they are maps of
smooth vector spaces but not of C ∞ -modules), the commutative diagram
1⊗φ 1⊗qφ
Ω1 ⊗OC ∞ V / Ω1 ⊗C ∞ W / Ω1 ⊗C ∞ Cφ /0
O
∇V ∇W
φ qφ
V /W / Cφ /0
lives in the category of smooth vector spaces. Since smooth vector spaces form
an abelian category, there is naturally a C-linear map fCφ from Cφ to Ω1 ⊗C ∞ Cφ
inherited from the connection on W. We need to show that fCφ is a flat connection,
but this assertion follows from the fact that φ intertwines the flat connections on V
and W.
2.3.6 Corollary. DVS contains all colimits. Hence it is a cocomplete category.
Proof. Every colimit can be constructed out of coproducts and coequalizers.
Since we are in a linear setting, we can construct an coequalizer as a cokernel.
In fact, DVS contains all limits and is hence complete, but the argument is
indirect. We first prove that DVS is locally presentable, and by Corollary 1.28 of
Adámek and Rosický (1994), we know that every locally presentable category is
complete.
2.3.7 Proposition. DVS is locally presentable.
2.3.8 Corollary. DVS contains all limits, and so it is a complete category.
The proof of this proposition depends on some technical results in category
theory, rather orthogonal from the rest of this book, so we will not give expository
background. Our main reference are Adámek and Rosický (1994) and Kashiwara
and Schapira (2006).
Proof of Proposition 2.3.7. As we know that DVS is cocomplete, it suffices to
prove that DVS is accessible (see Corollary 2.47 of Adámek and Rosický (1994)).
Our strategy is to present DVS as constructed out of known accessible categories
by operations that produce accessible categories. To be more precise, there is a 2-
category ACC of accessible categories, and it possesses certain “limits” understood
in the correct 2-categorical sense. In particular, ACC is closed under “inserters”
and “equifiers.” Below, we will define and use precisely these constructions.
First, we must show that our initial inputs are accessible. Recall that the mod-
ule sheaves for a sheaf of rings on a site is presentable and hence accessible. (See,
for instance, Theorem 18.1.6 of Kashiwara and Schapira (2006).) Thus the cat-
egory of smooth vector spaces Vect sm – i.e., sheaves of vector spaces on the site
Mfld – is accessible. Likewise, the category ModC ∞ of C ∞ -modules is accessible.
Consider the following two functors:
F : ModC ∞ → Vect sm ,
the forgetful functor, and
G := G ◦ Ω1 ⊗C ∞ : ModC ∞ → Vect sm ,
which sends a C ∞ -module V to the 1-forms valued in V, Ω1 ⊗C ∞ , and then views
it as a smooth vector space. The forgetful functor F preserves filtered colimits by
Theorem 18.1.6 of Kashiwara and Schapira (2006); hence F is accessible. The
functor Ω1 ⊗C ∞ − is a left adjoint from ModC ∞ to itself and hence preserves all col-
imits, so the composite G also preserves filtered colimits and is hence accessible.
The inserter of F and G is the following category, denoted Ins(F, G). An
object of Ins(F, G) is a morphism f : F(V) → G(V) with V an object in ModC ∞ .
A morphism in Ins(F, G) is a commuting square
f
F(V) / G(V)
F(φ) G(φ)
f0
F(V 0 ) / G(V 0 )
with φ : V → V 0 a morphism in ModC ∞ . For our situation, this inserter amounts to

a choice of “pre-connection” ∇ : V → Ω1 ⊗C ∞ V on a C ∞ -module V. It is accessible
because the inserter of two accessible functors is accessible (see Theorem 2.72 of
Adámek and Rosický (1994)).
A pre-connection ∇ is a connection if it satisfies the Leibniz rule:

∇( f · s) = d f ⊗ s + f · ∇s
for f a smooth function and s a section. This equation can be used to produce the
category of C ∞ -modules with connection, as follows.
Consider the following two functors:
F 0 : Ins(F, G) → Vect sm ,
the functor which sends a pre-connection ∇ : V → Ω1 ⊗C ∞ V to the smooth vector
space C ∞ ⊗ F(V), where this tensor product is in Vect sm ; and
G0 : Ins(F, G) → Vect sm ,
which sends ∇ to F(Ω1 ⊗C ∞ V), the underlying smooth vector space of the 1-forms
valued in V. Note that the forgetful functor P : Ins(F, G) → ModC ∞ is itself ac-
cessible (see Remark 2.73 of Adámek and Rosický (1994)). These functors F 0 , G0
are accessible because they are compositions of P with the accessible functors F,
G, and tensoring with C ∞ .
Consider the following two natural transformations α, β : F 0 ⇒ G0 . The trans-
formation α sends a pre-connection (V, ∇) to the map of smooth vector spaces
α(∇) : C ∞ ⊗ F(V) → F(Ω1 ⊗C ∞ V)
.
f ⊗s 7→ ∇( f · s) − (d f ⊗ s + f · ∇s)
The transformation β sends a pre-connection (V, ∇) to the map of smooth vector
spaces
β(∇) : C ∞ ⊗ F(V) → F(Ω1 ⊗C ∞ V)
.
f ⊗s 7→ 0
Note that ∇ is a connection precisely if α(∇) = β(∇), which is the Leibniz rule.
The equifier of F 0 , G0 , α, and β, denoted Eq(F 0 , G0 , α, β), is the full subcate-
gory of Ins(F, G) consisting of objects ∇ on which α(∇) = β(∇). In other words,
Eq(F 0 , G0 , α, β) is the category of connections. By Lemma 2.76 of Adámek and
Rosický (1994), we see that the equifier of accessible functors is accessible.
The category of C ∞ -modules with flat connections can likewise we constructed
as an equifier, since a connection is flat exactly when its curvature is zero. Here the
functors are F 00 , G00 : Eq(F 0 , G0 , α, β) → Vect sm , where F 00 (∇ : V → Ω1 ⊗C ∞ V) =
F(V) the underlying smooth vector space of V and G00 (∇) = F(Ω2 ⊗C ∞ V). These
are accessible functors, for reasons identical to the preceding constructions. The
natural transformations α0 , β0 : F 00 ⇒ G00 are α0 (∇) = Ω∇ , the curvature of the
connection, and β0 (∇) = 0. The equifier then amounts to those connections with
zero curvature.
2.4. Tensoring over C ∞ . Standard constructions on vector spaces, such as

tensor product and formation of internal hom spaces, carry over to vector bundles,
and also to vector bundles with connection. We now spell out how those construc-
tions work in DVS.
Let V, W be C ∞ -modules. Then we define a C ∞ -module V⊗C ∞ W by sheafifying

the presheaf X 7→ C ∞ (X, V) ⊗C ∞ (X) C ∞ (X, W). When V, W are differentiable, we
equip this tensor product with the flat connection
∇X,V⊗W := ∇X,V ⊗ IdW + IdV ⊗∇X,W ,
which is the usual formula for connections.
Recall now the following standard construction. Let V, W be C ∞ -modules. Let
HomC ∞ (V, W) denote the “sheaf Hom”: to each manifold X, we assign
C ∞ (X, HomC ∞ (V, W)) := HomC ∞ (X) (C ∞ (X, V), C ∞ (X, W)),
the vector space of C ∞ (X)-linear maps between C ∞ (X, V) and C ∞ (X, W). This
presheaf is, in fact, a sheaf, and it is naturally a C ∞ -module. If V, W are differen-
tiable, then HomC ∞ (V, W) is naturally a differentiable vector space, with
(X · φ)(v) = X(φ(v)) − φ(X(v)),
where X is a vector field on X, φ is a section of C ∞ (X, HomC ∞ (V, W)), and v is a
section of C ∞ (X, V). By definition, this connection is flat. In order to emphasize
that it has a differentiable structure, we use HomDVS to denote this HomC ∞ .
This symmetric monoidal structure, however, does not have several properties
that we desire. For example, “global sections” of HomDVS (V, W), namely its value
on a point C ∞ (∗, HomDVS (V, W)), is not DVS(V, W). Instead, it is ModC ∞ (V, W),
merely the smooth maps and not the connection-respecting maps. This is the
“wrong answer” for certain purposes. We explain a resolution of this issue in Sec-
tion 6 below.
As another example of a “defect” of this monoidal structure, we would like
C (M) ⊗C∞ C ∞ (N) C ∞ (M × N) in DVS, but it is not true. Consider evaluating
∞
both sheaves on a point ∗: the sections of the tensor product is

C ∞ (∗, C ∞ (M) ⊗C∞ C ∞ (N)) = C ∞ (M) ⊗R C ∞ (N)
and that is not equal to C ∞ (M × N). The usual remedy to this issue is to take the ap-
propriate completion of this algebraic tensor product, using natural topologies on
these spaces. Alternatively, relying on ideas from the next sections of this appen-
dix, one can note that these vector spaces come from bornological vector spaces,
and that their tensor product in DVS agrees with the tensor product in BVS. But we
know C ∞ (X)b ⊗βC ∞ (Y) C ∞ (X × Y), which is the completion of the bornological
tensor product to a convenient vector space. The bornological tensor product needs
to be completed to obtain what we want. Hence a strategy coherent with our em-
phasis on DVS is to “complete” the tensor product ⊗C ∞ and develop “convenient
differentiable vector spaces,” but we will not pursue this construction here.
2.5. Multilinear maps and the multicategory structure. There is a natural
way to form “many-to-one” maps between ordinary vector spaces: consider the
vector space of multilinear maps φ : V1 ×· · ·×Vn → W. Multilinear maps compose
naturally, by feeding the output of one multilinear map into one of the inputs of the
next. Altogether, this rich structure is formalized in the notion of a multicategory
(see 2.3). We have a similar story for differentiable vector spaces.
3. LOCALLY CONVEX TOPOLOGICAL VECTOR SPACES 267
2.5.1 Proposition. DVS forms a multicategory in which

DVS(V, . . . , Vn | W)
denotes the collection of multilinear maps of differentiable vector spaces. A multi-
linear map
φ : V1 × · · · × Vn → W
∞
is a multilinear map of C -modules compatible with the flat connections: for each
manifold X,
(1) given sections fi ∈ C ∞ (X, Vi ), φ( f1 , . . . , fn ) ∈ C ∞ (X, W), and
(2) if X is a vector field on X, then
n
X
X(φ( f1 , . . . , fn )) = φ( f1 , . . . , X( fi ), . . . , fn ).
i=1
3. Locally convex topological vector spaces

This section will be brief, and it aims simply to explain a result of Kriegl and
Michor (1997) that shows how locally convex Hausdorff topological vector spaces
provide differentiable vector spaces. In other words, we explain how a very large
and important class of topological vector spaces fits inside our preferred approach.
3.0.1 Definition. Let LCTVS denote the category where an object is a locally con-
vex Hausdorff topological vector space and a morphism is a continuous linear map.
Remark: In the remainder of this section, we will use topological vector space to
mean locally convex Hausdorff topological vector space, for simplicity. ^
The key idea of Kriegl and Michor is simple and compelling: the notion of a
smooth curve is well-behaved, so build up everything from that notion.
In particular, it is possible to define smooth curves γ : R → V in a topological
vector space in a very simple way, in the style of elementary calculus. One then
defines a smooth map f : V → W between topological vector spaces to be a
map that sends smooth curves in V to smooth curves in W. This notion extends
naturally to define smooth maps from a finite-dimensional smooth manifold X into
a topological vector space V. Hence, we obtain a smooth vector space (i.e., a sheaf
on Mfld) from V: to the manifold X, we assign the set of smooth maps C ∞ (X, V).
Remark: In fact, a map f : M → V is smooth in this sense if and only if it is
smooth in the usual sense. See Lemma 3.0.7 below. ^
3.0.2 Definition. Let V ∈ LCTVS. A curve γ : R → V is differentiable if its
derivative
γ(t + s) − γ(t)
γ0 (t) := lim
s→0 s
exists at t ∈ R for all t.
A curve γ is smooth if all iterated derivatives exist.
With this definition in hand, we introduce the notion of a smooth map between
topological vector spaces, following Kriegl and Michor.
3.0.3 Definition. Let V, W ∈ LCTVS. A (not necessarily linear) function f : V →

W is smooth if for each smooth curve γ : R → V, the composition f ◦ γ : R → W
is smooth.
As Kriegl and Michor explain, this notion arises by taking seriously the per-
spective of variational calculus.
Remark: We note that not every smooth function between topological vector spaces,
in this sense, is necessarily continuous. As Kriegl and Michor point out, this a
priori unappealing mismatch is not so strange: smoothness is about nonlinear phe-
nomena, and the usual topology on a topological vector space is focused on linear
phenomena. (For many nice topological vector spaces, however, a smooth function
is continuous.) ^
In particular, for an open U ⊂ Rn and V a topological vector space, we say that
f : U → V is smooth if any smooth curve γ : R → U goes to a smooth curve
f ◦ γ : R → V. For X a smooth manifold, a function f : X → V is smooth if
its restriction to every chart is smooth. Hence the notion is local on X, so that we
obtain the following definition.
3.0.4 Definition. Let V ∈ LCTVS. We define a C ∞ -module by
X ∈ Mfldop 7→ { f : X → V | f smooth}.
We denote this functor by C ∞ (−, V).
Let modt : LCTVS → ModC ∞ denote the functor that sends V to C ∞ (−, V). It
has the following useful property.
3.0.5 Lemma. The functor modt preserves limits.
Proof. The forgetful functor from sheaves of C ∞ -modules to sheaves of sets
preserves limits, because it is a right adjoint. Likewise, the forgetful functor from
sheaves of sets to presheaves of sets preserves limits, since it is a right adjoint. But
a limit of presheaves is computed objectwise, since it is simply a limit in a functor
category. We thus merely need to show that
C ∞ (X, modt (lim V)) C ∞ (X, lim modt (V))
I I
for every smooth manifold X.

Now observe that a map from a manifold X into a product of locally convex
topological vector spaces is smooth if and only if it it is smooth into each com-
ponent separately (i.e., if its projection to each component is smooth). Hence,
products are preserved. Similarly, a map from X into a kernel ker(φ : V → W) in
LCTVS is smooth if and only if it is smooth into V and the composition into W
is the constant zero map. As limits are generated by kernels and products, we are
done.
But this functor C ∞ (−, V) is even better. By construction, these maps C ∞ (X, V)
are differentiable, so we find that C ∞ (−, V) is naturally a differentiable vector
3. LOCALLY CONVEX TOPOLOGICAL VECTOR SPACES 269
space. To be explicit, for each smooth map f : X → V, there is a derivative

T f : T X → V ⊕ V, by the definition of “smooth.” Hence the flat connection is
∇X : C ∞ (X, V) → Ω1 (X, V)
,
f 7→ d f = π2 ◦ T f
where π2 : V ⊕ V → V denotes projection onto the second summand and π1 ◦ T f =
f . Note that this construction is precisely the usual construction of the natural flat
connection on a trivialized vector bundle.
3.0.6 Proposition. There is a faithful functor di ft : LCTVS → DVS sending V to
C ∞ (−, V), equipped with its natural flat connection. It preserves limits.
Proof. We already described the construction that produces the functor, so
it remains to check that di ft preserves limits. This result is not formal because
the forgetful functor U from DVS to ModC ∞ does not preserve all limits. (One
problem is that the same C ∞ -module might admit multiple flat connections, and
hence distinct objects in DVS might have the same image.) But modt = U ◦ di ft ,
so the situation simplifies.
Given a diagram V : I → LCTVS, we know that di ft (limI V) is simply

C ∞ −, lim V = modt lim V lim modt ◦ V,
I I I
equipped with its canonical flat connection. Since every map of differentiable vec-
tor spaces is also C ∞ -linear, the map from limI di ft ◦ V to the diagram di ft ◦ V
factors through di ft (limI V). Hence limI di ft ◦ V di ft (limI V).
This functor di ft factors through the category of bornological vector spaces
(and in consequence is not full, as we will see). Thus, in Section 4, we will de-
scribe the bornological vector space associated to every complete locally convex
topological vector space and how every bornological vector space provides a dif-
ferentiable vector space.
Finally, we elaborate on Remark 3.
3.0.7 Lemma (Kriegl and Michor (1997), Lemma 3.14). For U ⊂ Rn an open
set and V ∈ LCTVS, a function f : U → V is smooth in the sense of Definition
3.0.3 if and only if all iterated partial derivatives ∂µ f /∂xµ exist and are smooth.
This lemma is due to Boman.
3.1. Categorical properties.
3.1.1 Lemma. LCTVS admits all limits.
Proof. We show the existence of products and of kernels. Together, these
guarantee the existence of limits.
Let {Vα | α ∈ A} be a collection in LCTVS. Consider the product of the Vα as
topological spaces, i.e., equipped with the usual product topology. As any product
of convex sets is convex, the topology on this topological product space is also
generated by convex sets. One can check that the induced scalar multiplication and
vector addition are continuous with respect to this topology.
Kernels are similar. Being the kernel of a continuous linear map is a closed
condition, and a closed linear subspace of a locally convex vector space is a locally
convex vector space.
3.1.2 Lemma. LCTVS admits all colimits.
Proof. We show the existence of cokernels and coproducts.
Cokernels are straightforward. For a continuous linear map φ : V → W, the
cokernel is W/φ(V), the quotient of W by the closure of the image of φ. We are
forced to use the closure as we work with Hausdorff spaces.
Coproducts are a little trickier. The underlying vector space is simply the direct
sum, but we equip it with the “diamond” topology. (We learned this term from the
helpful online lecture notes of Paul Garrett.) We construct a collection of convex
neighborhoods of the origin out of convex neighborhoods of the constituents as
follows.
Let iα : Vα ,→
L
α∈A Vα denote the canonical inclusion. Given a choice of
convex neighborhood Uα of Vα for every α ∈ A, take the convex hull of the union
S
α∈A iα (U α ). Declare such a convex hull to be an open. We take the vector space
topology induced by all such opens.
4. Bornological vector spaces

There is another natural approach to taming infinite-dimensional vector spaces,
via the notion of a bornology. Instead of specifying a topology on V, one specifies
the bounded subsets of V.
4.0.1 Definition. A bornology on a vector space V (over a field K = R or C) is a
collection of subsets B such that
(1) B covers V, i.e., V = B∈B B,
S
(2) B is closed under inclusions, i.e., if B ∈ B and B0 ⊂ B, then B0 ∈ B,
(3) B is closed under finite unions, i.e., if B1 , . . . , Bn ∈ B, then B1 ∪ · · · ∪
Bn ∈ B.
Moreover, these bounded sets are compatible with the vector space structure:
(4) B is closed under translation, i.e., if v ∈ V and B ∈ B, then v + B ∈ B,
(5) B is closed under dilation, i.e., if λ ∈ K and B ∈ B, then λ · B ∈ B,
(6) B is closed under the formation of balanced hulls, i.e., if B, B0 ∈ B,
then B + B0 ∈ B.
A topological vector space V has an associated bornology, as follows. We say

B ⊂ V is bounded if for every open set U ⊂ V containing the origin, there is a real
number λ > 0 such that B ⊂ λU. When V is locally convex, a set B is bounded if
and only if every continuous semi-norm of V is bounded on B. Note that distinct
topologies may have the identical associated bornologies.
We are only interested in such bornologies, so we work with the following
category.
4. BORNOLOGICAL VECTOR SPACES 271
4.0.2 Definition. A bornological vector space (V, B) is a vector space whose bornol-
ogy is obtained from some locally convex Hausdorff topology on V.
A linear map f : V → W is bounded if the image of every bounded set in V is
a bounded set in W.
We denote by BVS the category where an object is a bornological vector space
and a morphism is a bounded linear map.
Note that a continuous linear map is always bounded, but the converse is false
in general. (Consider an infinite-dimensional Banach space V, and let Vw denote
the underlying vector space equipped with the weak topology. The identity map
id : Vw → V is bounded but not continuous.)
Consider the functor incβ : BVS → LCTVS that assigns to a bornological vec-
tor space the finest locally convex topology with the same underlying bounded sets.
By Lemma 4.2 of Kriegl and Michor (1997), this functor embeds BVS as a full sub-
category of LCTVS. In particular, a linear map f : V → W between bornological
vector spaces is bounded if and only if f : incβ V → incβ W is continuous.
4.0.3 Corollary. The functor incβ : BVS ,→ LCTVS preserves all colimits.
Proof. LCTVS contains all colimits. Given a diagram δ : D → BVS, let
colim incβ ◦ δ denote its colimit in LCTVS. If there exists a colimit colim δ in
BVS, then we know there is a canonical map
colim incβ ◦ δ → incβ colim δ.
Hence, it suffices to show that colim incβ ◦ δ is in the image of incβ to obtain the
claim.
For each object d ∈ D, we have a continuous linear map
incβ δ(d) → colim incβ ◦ δ,
which is thus a bounded linear map. Hence, we know that this map factors through
the underlying vector space of colim incβ ◦ δ, but now equipped with the finest lo-
cally convex topology with the same bounded sets. Hence, the colimit colim incβ ◦δ
is in the image of incβ .
By the very definition of incβ , we see that there is a natural right adjoint.
4.0.4 Corollary. Consider the functor
born : LCTVS → BVS
sending a locally convex Hausdorff topological vector space to its underlying bornolog-
ical vector space. This functor born is right adjoint to incβ .
It is important to know that it is the bornology of a topological vector space
that matters for smoothness, not the topology itself. The following two results from
Kriegl and Michor (1997) make this assertion precise.
4.0.5 Lemma (Kriegl and Michor (1997), Corollary 1.8). For V ∈ LCTVS, a
curve γ : R → V is smooth if and only if γ : R → incβ (born(V)) is smooth.
4.0.6 Lemma (Kriegl and Michor (1997), Corollary 2.11). A linear map L :
V → W of locally convex vector spaces is bounded if and only if it maps smooth
curves in V to smooth curves in W.
In consequence, we have the following result.
4.0.7 Proposition. The functor
di fβ : BVS → DVS
V 7→ di ft (incβ (V))
embeds BVS as a full subcategory of DVS. It preserves all limits.
Note that BVS is therefore equivalent to the essential image of LCTVS in DVS.
Proof. Only the assertion about limits remains to be proved. Note that born
preserves limits, so that since BVS is a full subcategory of LCTVS, we can com-
pute a limit in LCTVS and then apply born. As di ft preserves limits and only cares
about the underlying bornology, we are finished.
4.1. Categorical properties. We have seen that incβ realizes BVS as a reflec-
tive subcategory of LCTVS. Thus, we know the following.
4.1.1 Lemma. BVS admits all limits and colimits.
Proof. LCTVS contains all limits, and born : LCTVS → BVS is a right
adjoint and hence preserves limits. A colimit of BVS is computed by applying
born to the colimit computed in LCTVS.
4.2. Multilinear maps and the bornological tensor product. As in any cat-
egory of linear objects, we can discuss multilinear maps. We can also ask if mul-
tilinear maps out of a tuple of vector spaces is co-represented by a vector space,
called the tensor product. For bornological vector spaces, there is such a natural
tensor product with all the properties we love about the tensor product of finite-
dimensional vector spaces.
4.2.1 Definition. Let BVS(V1 , . . . , Vn | W) denote the set of all bounded multilin-
ear maps from the set-theoretic product of bornological vector spaces V1 × · · · × Vn
to the bornological vector space W. Equip it with the bornology of uniform con-
vergence on bounded sets. This bornology makes BVS(V1 , . . . , Vn | W) into a
bornological vector space.
The following result is the first step to showing that BVS forms a closed sym-
metric monoidal category.
4.2.2 Proposition (Kriegl and Michor (1997), Proposition 5.2). There are natu-
ral bornological isomorphisms
BVS(V1 , . . . , Vn+m | W) BVS (V1 , . . . , Vn | BVS(Vn+1 , . . . , Vn+m | W)) .
Hence, BVS forms a multicategory where multimorphisms are themselves bornolog-
ical vector spaces and composition is bounded.
5. CONVENIENT VECTOR SPACES 273
In BVS, there is a natural bornology to put on the algebraic tensor product

V ⊗alg W that co-represents bilinear maps. Simply put, we equip V ⊗alg W with the
finest locally convex topology such that the canonical map V × W → V ⊗alg W is
bounded. We denote this bornological vector space by V ⊗β W.
4.2.3 Theorem (Kriegl and Michor (1997), Theorem 5.7). The bornological ten-
sor product V ⊗β − : BVS → BVS is left adjoint to the Hom-functor BVS(V, −) :
BVS → BVS.
In particular, we have the following bornological isomorphisms:
BVS(V ⊗β W, U) BVS(V, W | U) BVS(V, BVS(W, U)),
V ⊗β R V,
V ⊗β W W ⊗β V,
(V ⊗β W) ⊗β U V ⊗β (W ⊗β U).
Moreover, ⊗β preserves all colimits separately in each variable.
To summarize, we have the following.
4.2.4 Proposition. BVS is a closed symmetric monoidal category via ⊗β . This
symmetric monoidal structure agrees with the multicategory structure on BVS in-
duced from the embedding incβ : BVS → DVS.
Proof. What remains is to show that the multicategory structure on BVS agrees
with that on DVS. Lemma 5.5 of Kriegl and Michor (1997) show that a multilinear
map of bornological vector spaces is bounded if and only if it is smooth, and so
this multicategory structure agrees with the induced multicategory structure from
DVS.
5. Convenient vector spaces

The power of calculus often involves the interplay of differentiation with inte-
gration, and we will find it convenient to focus on a class of bornological vector
spaces where integration along curves is well-behaved. As integration involves,
speaking casually, infinite sums, we can view the existence of integrals as a kind
of completeness property. We use the notion developed and applied to great effect
by Kriegl and Michor in Kriegl and Michor (1997).
5.0.1 Definition. A bornological vector space V ∈ BVS is c∞ -complete if one of
the following equivalent conditions holds:
(a) For any smooth curve γ : R → V, there is a smooth curve Γ : R → V
such that Γ0 = γ. We call Γ an antiderivative.
(b) A curve γ : R → V is smooth if and only if for every bounded linear
functional λ : V → R, the composition λ ◦ γ is smooth.
A c∞ -complete bornological vector space will be called a convenient vector
space. We denote by CVS the full subcategory of BVS whose objects are convenient
vector spaces.
Several other equivalent definitions of c∞ -complete are given in Theorem 2.14

of Kriegl and Michor (1997).
The following result ensures that the inclusion functor incc : CVS → BVS is a
right adjoint.
5.0.2 Theorem (Kriegl and Michor (1997), Theorem 2.15). The full subcategory
CVS ⊂ BVS is closed under limits.
We denote the left adjoint by c∞ : BVS → CVS. It sends a bornological vector
space to its c∞ completion.
As di fβ preserves limits, we see that the composition
di fc : CVS → DVS
V 7→ di fβ ◦ incc (V)
also preserves limits. Moreover, because CVS is a full subcategory of BVS and
BVS is a full subcategory of DVS, we have the following.
5.0.3 Lemma. The functor di fc : CVS → DVS embeds CVS as a full subcategory
that is closed under limits.
We note that while di fc does not preserve all colimits, it does preserve some.
5.0.4 Proposition. The functor di fc : CVS → DVS preserves countable coprod-
ucts and sequential colimits of closed embeddings.
Proof. A countable coproduct is a special case of a sequential colimit of closed
embeddings, so we will prove only the latter property.
Given a sequence
V1 → V2 → · · ·
of closed embeddings, let V denote colim Vi . We need to show that
η : colim di ft (Vi ) → di ft (V)
is an equivalence of differentiable vector spaces. By Corollary 2.2.2, it is enough

to show that this natural transformation is an isomorphism on stalks, as these are
sheaves. In other words, the problem is local and we can restrict our attention to
the neighborhood of the origin in each Euclidean space Rn .
First, by Lemma 3.8 of Kriegl and Michor (1997), because Vi is a closed sub-
space of V, a curve in Vi is smooth if and only if the composition into V is smooth.
Hence, Vi ,→ V is a smooth map. As C ∞ (X, Vi ) ,→ C ∞ (X, V) for each i, we see
that the map ηX above is an inclusion for any manifold X.
Now we want to show that each smooth map f : Rn → V factors, in some
sufficiently small neighborhood of the origin, through some Vi . Pick an open set
0 ∈ U ⊂ Rn whose closure U is compact. As f : Rn → V is continuous by
Lemma 3.0.7, f (U) is also compact and hence bounded. By Result 52.8 of Kriegl
and Michor (1997), a subset of V is bounded if and only if it is a bounded subset
of some Vi . Hence f restricted to U maps into this Vi .
5. CONVENIENT VECTOR SPACES 275
5.1. Examples from differential geometry. Let us explain some further ex-
amples. Let E be a vector bundle on a manifold M, and, as before, let E , Ec , E , E c
refer to sections of E which are smooth, compactly supported, distributional, or
compactly-supported and distributional. All of these spaces have natural topologies
and so can be viewed as bornological vector spaces, and they are all c∞ -complete.
We explained in Section 2.1 how to view these vector spaces as differentiable.
It is easy to check that the smooth structure discussed there is the same as the
one that arises from the topology. Indeed, the Serre-Swan theorem tells us that any
vector bundle is a direct summand of a trivial vector bundle, so we can reduce to the
case that E is trivial. Then results of Grothendieck, summarized in Grothendieck
(1952), allow one to describe smooth maps to these various vector spaces using
the theory of nuclear vector spaces; in this way we arrive at the description given
earlier.
5.1.1 Lemma. For any vector bundle E on a manifold M, the standard nuclear
topologies on E c , E and Ec are bornological. It follows that these spaces are
convenient (because completeness in the locally-convex sense is stronger than c∞ -
completeness).
Proof. Because every vector bundle is a summand of a trivial one, it suffices to
prove the statement for the trivial vector bundle. According to Kriegl and Michor
(1997), 52.29, the strong dual of a Fréchet Montel space is bornological. The space
C ∞ (M) of smooth functions on a manifold is Fréchet Montel, because every nu-
clear space is Schwartz (see pages 579-581 of Kriegl and Michor (1997)) and every
Fréchet Schwarz space is Montel. Thus the strong dual of C ∞ (M) is bornological,
as desired.
Next, we will see that any Fréchet space is bornological. This property follows
immediately from proposition 14.8 of Trèves (1967) (see also the corollary on the
following page). It follows that C ∞ (M) is bornological, and that the same holds for
C K∞ (M) for any compact subset K ⊂ M.
Since bornological spaces are closed under formation of colimits, the same
holds for Cc∞ (M).
5.2. Multilinear maps and the completed tensor product. We can take the
bornological tensor product of any two convenient vector spaces V and W, but
V ⊗β W is rarely convenient itself. The completion of a tensor product c∞ (V ⊗β W) is
convenient, however, and equips CVS with a natural symmetric monoidal structure.
We denote this completed tensor product by b ⊗β . That is,
⊗β W := c∞ (V ⊗β W).
Vb
(Strictly speaking, we should write c∞ (incc (V) ⊗β incc (W)) to make clear where
everything lives.)
5.2.1 Lemma. The multicategory structure on CVS induced by the functor di fc :
CVS → DVS is represented by b
⊗β .
Proof. We have
⊗β W, U) = CVS(c∞ (incc (V) ⊗β incc (W)), U)

CVS(V b
BVS(incc (V) ⊗β incc (W), incc (U))
DVS(di fβ (incc (V)), di fβ (incc (W)) | di fβ (incc (U)))
DVS(di fc (V), di fc (W) | di fc (U)).
⊗β encodes multilinear maps of the underlying differentiable

Hence we see that b
vector spaces.
⊗β , CVS forms a closed symmetric monoidal

But the situation is even better. Via b
category, as we now explain.
5.2.2 Lemma. For any two convenient vector spaces V, W, the bornological vector
space BVS(V, W) is a convenient vector space.
Note that BVS(V, W) = CVS(V, W) as CVS is a full subcategory, by definition.
Proof. Recall that C ∞ (V, W) denotes the set of functions that sends smooth
curves in V to smooth curves in W. We equip it with the initial topology such that
for any smooth γ : R → V, the pullback γ∗ : C ∞ (V, W) → C ∞ (R, W) is continuous.
This topology makes C ∞ (V, W) into a locally convex topological vector space.
Then BVS(V, W) ⊂ C ∞ (V, W) is a closed set because f ∈ C ∞ (V, W) is linear if
and only if
0 = (ev x + λevy − ev x+λy ) f = f (x) + λ f (y) − f (x + λy)
for every x, y ∈ V and λ ∈ R. As a closed linear subspace of a convenient vector

space is convenient (because CVS is closed under limits), it suffices to show that
C ∞ (V, W) is convenient.
By Lemma 3.7 of Kriegl and Michor (1997), the set of smooth curves C ∞ (R, W)
is a convenient vector space, as W is convenient. By Lemma 3.11 of Kriegl and
Michor (1997), the space C ∞ (V, W) is the limit of C ∞ (R, W) over all the smooth
curves γ : R → V, where the morphisms in the diagram category are given by
pullback along the reparametrizations γ1 = c∗ γ0 , with c ∈ C ∞ (R, R). (This result
should be plausible because a function f : V → W is smooth precisely when it
sends smooth curves into V to smooth curves in W.) Putting these facts together
with Theorem 5.0.2, we see that C ∞ (V, W) is convenient.
Thanks to Lemma 5.2.2, we can prove the following.

5.2.3 Lemma. The functor c∞ : BVS → CVS is symmetric monoidal.
6. THE RELEVANT ENRICHMENT OF DVS OVER ITSELF 277
Proof. Let V, W ∈ BVS and U ∈ CVS. Then

CVS(c∞ (V ⊗β W), U) BVS(V ⊗β W, incc U)
BVS(V, BVS(W, incc U))
BVS(V, incc CVS(c∞ W, U))
CVS(c∞ V, CVS(c∞ W, U))
BVS(c∞ V, BVS(c∞ W, incc U))
BVS(c∞ V ⊗β c∞ W, incc U)
CVS(c∞ (c∞ V ⊗β c∞ W), U)
CVS(c∞ V b
⊗β c∞ W, U).
In moving from the fourth to the fifth line, we use the fact that CVS(X, Y) =
BVS(X, Y) for all X, Y ∈ CVS, since it is a full subcategory of BVS.
One nice consequence of these observations is the following.

⊗β commutes with all
5.2.4 Corollary. The category CVS admits all colimits, and b
colimits separately in each variable.
Proof. As the category BVS admits all colimits, we obtain the colimit of a dia-
gram D → CVS by computing the colimit first in BVS and then applying the com-
pletion functor c∞ , which preserves colimits because it is a left adjoint. Completion
is also symmetric monoidal. As ⊗β preserves all colimits, we are finished.
Here is a summary of the main result.

⊗β .
5.2.5 Proposition. CVS is a closed symmetric monoidal category via b
Moreover, this symmetric monoidal structure agrees with the multicategory
structure induced by the functor di fc : CVS → DVS.
In the next section, we show that the inner hom in CVS is compatible with
the correct self-enrichment of DVS. In other words, there is a notion of the dif-
ferentiable vector space of differentiable maps between convenient vector spaces,
developed below, and it agrees with the differentiable vector space of convenient
maps between convenient vector spaces.
6. The relevant enrichment of DVS over itself

Our goal in this section is to enhance DVS to a category in which there is a
differentiable vector space of maps between any two differentiable spaces V and
W. Here is the first important result of this section, whose proof appears in the first
two subsections.
6.0.1 Theorem. There exists a category DVS, enriched over DVS, whose objects
are differentiable vector spaces and whose morphism-sheaves HomDVS (V, W) sat-
isfy
C ∞ (X, HomDVS (V, W)) = DVS(V, W(X × −))
for any manifold X. In particular, the value on a point C ∞ (∗, HomDVS (V, W))
equals DVS(V, W).
We use this enriched category in our definition of equivariant prefactorization
algebras (see section 7) and of holomorphically translation-invariant prefactoriza-
tion algebras (see 1).
Another compelling feature of this category is that it matches nicely with the
internal hom in CVS, the category of convenient vector spaces, which is the source
of most of the vector spaces that we actually work with.
6.0.2 Lemma. For any manifold M and convenient vector spaces V, W, there is an
isomorphism
C ∞ (M, CVS(V, W)) = CVS(V, C ∞ (M, W)),
and it is natural in M. This identification is compatible with the flat connection
possessed by the C ∞ (M) modules on each side. Hence
CVS(V, W) HomDVS (V, W)
as differentiable vector spaces.
The proof appears in the third subsection.
6.1. Constructions with families. To start, we want to construct, for every
manifold X, a sheaf on Mfld that returns, intuitively speaking, “families over X of
horizontal sections of the differentiable vector space V.” To do this, we need to
specify what we mean by such families.
6.1.1 Definition. For a manifold X and a differentiable vector space V, let C∞ (X, V)
denote the presheaf sending a manifold Σ to C ∞ (Σ × X, V).
In other words, C∞ (X, V) is the composite of the product functor −×X followed
by evaluating the sheaf V.
6.1.2 Lemma. For any manifold X and any V ∈ DVS, the presheaf C∞ (X, V) is
canonically a differentiable vector space.
Proof. It is straightforward to see that C∞ (X, V) is a sheaf. Any cover of Σ
pulls back to a cover over Σ × X along the projection πΣ : Σ × X → Σ, and since V
is a sheaf, we can then reconstruct the value of V on Σ × X using this cover.
Likewise, as C ∞ (Σ, C∞ (X, V)) = C ∞ (Σ × X, V) is a C ∞ (Σ × V)-module, it is a
C (Σ)-module via the pullback map π∗Σ : C ∞ (Σ) → C ∞ (Σ × X).
∞
It remains to produce the natural flat connection. By definition, we have a flat

connection
∇Σ×X,V : C ∞ (Σ × X, V) → Ω1 (Σ × X, V),
but we need to produce a flat connection
∇Σ,C∞ (X,V) : C ∞ (Σ, C∞ (X, V)) → Ω1 (Σ, C∞ (X, V)).
Observe that on Σ × X, the cotangent bundle T Σ×X∗ naturally splits as a direct sum
πX T X ⊕ πΣ T Σ , where πX : Σ × X → X and πΣ : Σ × X → Σ are the natural projection
∗ ∗ ∗ ∗
maps. In consequence,
Ω1 (Σ × X) Ω1 (X) ⊗C ∞ (X) C ∞ (Σ × X) ⊕ Ω1 (Σ) ⊗C ∞ (Σ) C ∞ (Σ × X),
where, for example, C ∞ (Σ × X) is a module over C ∞ (X) via the projection πX :
Σ × X → X. Hence we have a decomposition
Ω1 (Σ × X, V) Ω1 (Σ × X) ⊗C ∞ (Σ×X) C ∞ (Σ × X, V)
Ω1 (X) ⊗C ∞ (X) C ∞ (Σ × X, V) ⊕ Ω1 (Σ) ⊗C ∞ (Σ) C ∞ (Σ × X, V)
Ω1 (X, C∞ (Σ, V)) ⊕ Ω1 (Σ, C∞ (X, V)).
Define ∇Σ,C∞ (X,V) as the composition of ∇Σ×X,V followed by projection to Ω1 (Σ, C∞ (X, V)).
On Σ × X, we know that ∇Σ×X,V is a connection on Σ × X and hence satisfies the
Leibniz rule, so the projection onto Ω1 (Σ, C∞ (X, V)) also satisfies the Leibniz rule.
Likewise, the curvature of ∇Σ,C∞ (X,V) over Σ is the projection of the curvature of
∇Σ×X,V to Ω2 (Σ, C∞ (X, V)), and hence zero.
We also need a slight generalization of the mapping space C∞ (X, V). Note that
C ∞ (X) provides an algebra object in DVS and that C∞ (X, V) is a module sheaf over
C ∞ (X).
6.1.3 Definition. For E → X a vector bundle on a smooth manifold and V a
differentiable vector space, let C∞ (X, E ⊗ V) denote the differentiable vector space
E (X) ⊗C ∞ (X) C∞ (X, V).
The value of this sheaf on a manifold Σ is
C ∞ (Σ, C∞ (X, E ⊗ V)) = C ∞ (Σ, E (X)) ⊗C ∞ (Σ,C ∞ (X)) C ∞ (Σ, C∞ (X, V))
= Γ(Σ × X, π∗X E) ⊗C ∞ (Σ×X) C ∞ (Σ × X, V).
We use Ω1 (X, V) to denote C∞ (X, T X∗ ⊗ V).
Tensoring over C ∞ (X) is well-behaved here as E (X) is finite rank and projec-
tive over C ∞ (X).
We now prove a key result for constructing the enriched hom HomDVS .
6.1.4 Lemma. For any differentiable vector spaces V and for any manifold X,
there is a natural map of sheaves
∇C∞ (X,V) : C∞ (X, V) → Ω1 (X, V)
making C∞ (X, V) a C ∞ (X)-module sheaf with a flat connection.
Proof. The flat connection on C∞ (X, V) arises from the composition
∇X×Σ,V
C ∞ (Σ, C∞ (X, V)) = C ∞ (Σ × X, V) −−−−−→ Ω1 (Σ × X, V) → Ω1 (Σ, C∞ (X, V)),
∗
where the final projection map arises from the splitting T Σ×X π∗X T X∗ ⊕ π∗Σ T Σ∗ . This
∞
flat connection ∇Σ,C (X,V)) is C (X)-linear, since it only takes derivatives in the
∞
Σ-directions. This construction is functorial in Σ.

We now show that the flat connection on V induces a natural map of sheaves
∇C∞ (X,V) : C∞ (X, V) → Ω1 (X, V).
This map is a flat connection for C∞ (X, V) viewed as a C ∞ (X)-module in ModC∞ . It
is not a map in DVS because it anticommutes with the flat connections on C∞ (X, V)
and Ω1 (X, V). Nonetheless, as we will see, this property is enough to obtain the
lemma.
Let Σ be an arbitrary manifold. The splitting T Σ×X
∗ π∗X T X∗ ⊕ π∗Σ T Σ∗ produces
a connection in the X-direction by projection. More explicitly, the composition of
the connection over Σ × X followed by projection onto the first summand,
∇X×Σ,V
C ∞ (Σ × X, V) −−−−−→ Ω1 (Σ × X, V) → Γ(Σ × X, π∗X T X∗ ) ⊗C ∞ (Σ×X) C ∞ (Σ × X, V),
gives a map
∇C∞ (X,V) : C ∞ (Σ, C∞ (X, V)) → C ∞ (Σ, Ω1 (X, V)) = C ∞ (Σ, C∞ (X, T X∗ ⊗ V)).
Note that this map is C ∞ (Σ)-linear since it only takes derivatives in the X-direction.
Moreover, this construction is clearly natural in Σ and thus defines a map as C ∞ -
modules.
This map ∇C∞ (X,V) is a flat connection with respect to C ∞ (X), viewed as a
C -module. On Σ × X, we know that ∇Σ×X,V is a connection on Σ × X and hence
∞
satisfies the Leibniz rule, so the projection onto C ∞ (Σ, C∞ (X, T X∗ ⊗V)) also satisfies
the Leibniz rule. Likewise, the curvature of ∇C∞ (X,V) over Σ is the projection of the
curvature of ∇Σ×X,V to C ∞ (Σ, Ω2 (X, V)), and hence zero. In short, we know that
∇Σ×X,V = ∇Σ,C∞ (X,V) + ∇C∞ (X,V) ,
because they both arise as ∇Σ×X,V followed by projections onto the two summands
of Ω1 (Σ × X). As ∇Σ×X,V is flat, we obtain the for each summand separately.
Note that this decomposition also implies that
0 = ∇Σ,Ω1 (X,V) ◦ ∇C∞ (X,V) + ∇C∞ (X,V) ◦ ∇Σ,C∞ (X,V) ,
justifying our claim that ∇C∞ (X,V) anticommutes with the flat connections and hence
is not a map of differentiable vector spaces.
6.2. The enriched hom. We now introduce the appropriate notion of “fami-
lies of maps from V to W.”
6.2.1 Definition. For V and W differentiable vector spaces, let HomDVS (V, W) de-
note the following presheaf. The sections of the presheaf HomDVS (V, W) on a man-
ifold X are
HomDVS (V, W)(X) = DVS(V, C∞ (X, W)).
Note that, by definition, a section F assigns to each manifold Σ a map
FΣ : C ∞ (Σ, V) → C ∞ (Σ, C∞ (X, W)) = C ∞ (Σ × X, W),
and these maps FΣ intertwine with pullbacks along maps φ : Σ → Σ0 . To a smooth
map f : Y → X, the presheaf HomDVS (V, W) assigns the natural pullback
f ∗ : DVS(V, C∞ (X, W)) → DVS(V, C∞ (Y, W)),
where for each manifold Σ,

( f ∗ F)Σ = (idΣ × f )∗W ◦ FΣ
with F ∈ DVS(V, C∞ (X, W)) and (idΣ × f )∗W : C ∞ (Σ × X, W) → C ∞ (Σ × Y, W) the
pullback map for the sheaf W.
Note that by definition, HomDVS (V, W)(∗) = DVS(V, W), so our definition has
one property necessary to provide an enrichment of DVS. We now see that it
naturally lifts to an element DVS.
6.2.2 Lemma. For any V, W ∈ DVS, the presheaf HomDVS (V, W) is a differentiable
vector space.
Proof. First, we explain why HomDVS (V, W) is a sheaf. Let U = {Ui } be a
cover of the manifold X in the site Mfld. Then, for any manifold Σ, there is a
natural product cover {Σ × Ui } of Σ × X, and
 
Y Y 
C ∞ (Σ × X, W) lim  C ∞ (Σ × Ui , W) ⇒ C ∞ (Σ × (Ui ∩ U j ), W) ,
i i, j
as it is a sheaf. Hence, to give a map FΣ : C ∞ (Σ, V)

→ C ∞ (Σ × X, W) is equivalent
(i)
to giving maps FΣ : C ∞ (Σ, V) → C ∞ (Σ × Ui , W) for every Ui such that for every i
and j, the postcompositions with pullback to the intersections Σ × (Ui ∩ U j ) agree.
As this equivalence is natural in Σ, we see that
 
Y Y 
DVS(V, C∞ (X, W)) lim  DVS(V, C∞ (Ui , W)) ⇒ DVS(V, C∞ (Ui ∩ U j , W)) ,
i i, j
which is the sheaf condition.

Now let’s see why HomDVS (V, W) is a C ∞ -module. Observe that for any
manifolds Σ and X, the projection map Σ × X → X produces an algebra map
C ∞ (X) → C ∞ (Σ × X), and so for any differentiable vector space W, there is a
canonical C ∞ (X)-module structure on C ∞ (Σ × X, W). Given f ∈ C ∞ (X), let m f de-
note the multiplication-by- f operator on C ∞ (Σ × X, W). Consider now an element
F ∈ DVS(V, C∞ (X, W)). For each Σ, the map FΣ : C ∞ (Σ, V) → C ∞ (Σ × X, W) is
C ∞ (Σ)-linear and commutes with differentiation by vector fields on Σ. (We are only
interested here in the flat connection along Σ of C ∞ (Σ, C∞ (X, W)).) The composite
m f ◦ FΣ is also C ∞ (Σ)-linear and commutes with differentiation by vector fields
on Σ, since m f is constant in the Σ-direction. As m f is also natural in Σ, we can
define f · F by ( f · F)Σ = m f ◦ FΣ . This construction makes DVS(V, C∞ (X, W)) a
C ∞ (X)-module, and as the construction is natural in X, we see that HomDVS (V, W)
is a C ∞ -module.
Finally, recall from Lemma 6.1.4 that there is a natural flat connection
∇C∞ (X,W) : C∞ (X, W) → Ω1 (X, W),
which induces a map
∇C∞ (X,V)∗ : DVS(V, C∞ (X, W)) → DVS(V, Ω1 (X, W))
by post-composition: F 7→ ∇C∞ (X,W) ◦ F. We want, however, a flat connection

∇X : DVS(V, C∞ (X, W)) → Ω1 (X) ⊗C ∞ (X) DVS(V, C∞ (X, W)),
so we begin by showing that there is an isomorphism

µX : Ω1 (X) ⊗C ∞ (X) DVS(V, C∞ (X, W)) −
→ DVS(V, Ω1 (X, W)),
where µX (α ⊗ F)Σ sends v ∈ C ∞ (Σ, V) to α ⊗ FΣ (v).
As we are working with sheaves, we can work locally in X, so it suffices to
check only for X = Rn . In that situation, we fix a frame {dx1 , . . . , dxn } for Ω1 (X),
and hence obtain an isomorphism
Ω1 (X) C ∞ (X)⊕n ,
which induces an isomorphism
Ω1 (X, W) C∞ (X, W)⊕n .
We then see that
Ω1 (X) ⊗C ∞ (X) DVS(V, C∞ (X, W)) DVS(V, C∞ (X, W))⊕n
DVS(V, C∞ (X, W)⊕n )
DVS(V, Ω1 (X, W)),
and direct inspection identifies this composite as µX . An analogous argument
shows that
Ω2 (X) ⊗C ∞ (X) DVS(V, C∞ (X, W)) DVS(V, Ω2 (X, W)).
The composite µ−1 ∞
X ◦∇C (X,V)∗ thus provides a natural flat connection on DVS(V, C (X, W)).
∞

We want to show that HomDVS provides differentiable vector spaces with the
structure of a category. Thus, we need to explain how to compose such “families
of maps;” the definition is natural but somewhat elaborate upon first encounter.
Let U, V, W be differentiable vector spaces, and let X be a manifold. There is a
natural composition
•X : DVS(V, C∞ (X, W)) × DVS(U, C∞ (X, V)) → DVS(U, C∞ (X, W))
constructed as follows. Let G ∈ DVS(V, C∞ (X, W)) and F ∈ DVS(U, C∞ (X, V)),
and recall that FΣ : C ∞ (Σ, U) → C ∞ (Σ × X, V) denote how F acts on sections for
the manifold Σ. For each input manifold Σ, set
(G •X F)Σ = (idΣ ×∆X )∗ ◦ GΣ×X ◦ FΣ : C ∞ (Σ, V) → C ∞ (Σ × X, W),
where ∆X : X → X × X is the diagonal map. Observe that for any map φ : Σ → Σ0 ,
we have
(G •X F)Σ ◦ φ∗U = φ∗W ◦ (G •X F)Σ0 ,
as each constituent of (G •X F)Σ commutes with pullbacks. Hence, •X is well-
defined. Moreover, by construction, •X is C ∞ (X)-bilinear, and •X respects the flat
connections, since its constituents do.
6.2.3 Definition. Let U, V, W be differentiable vector spaces. The composition in

families
• : HomDVS (V, W) × HomDVS (U, V) → HomDVS (U, W)
is a bilinear map of differentiable vector spaces where the map on sections over a
manifold X is
•X : DVS(V, C∞ (X, W)) × DVS(U, C∞ (X, V)) → DVS(U, C∞ (X, W)).
To see that this operation • is well-defined, observe that for any map f : Y → X
of manifolds, we find
( f ∗G) •Y ( f ∗ F) = f ∗ ◦ (G •X F)
as each constituent plays naturally with pullbacks such as f ∗ : DVS(U, C∞ (X, W)) →
DVS(U, C∞ (Y, W)). Hence • defines a map of presheaves, in fact, a bilinear map
of differentiable vector spaces.
6.2.4 Lemma. This composition operation is associative:
− • (− • −) = (− • −) • −.
Proof. We need to show that for a manifold X and morphisms F ∈ DVS(T, C∞ (X, U)),
G ∈ DVS(U, C∞ (X, V)), and H ∈ DVS(V, C∞ (X, W)),
H •X (G •X F) = (H •X G) •X F.
Fix an input manifold Σ. Observe that (H •X (G •X F))Σ is given by
(idΣ ×∆X )∗ ◦ HΣ×X ◦ ((idΣ ×∆X )∗ ◦ GΣ×X ◦ FΣ )
and that ((H •X G) •X F)Σ is given by
(idΣ ×∆X )∗ ◦ ((idΣ×X ×∆X )∗ ◦ H(Σ×X)×X ◦ GΣ×X ) ◦ FΣ .
Using associativity of (ordinary) composition, we see that it suffices to show
HΣ×X ◦ (idΣ ×∆X )∗ = (idΣ×X ×∆X )∗ ◦ H(Σ×X)×X ,
but this equality holds because H is a map of sheaves on manifolds, and hence
intertwines the pullback of sections.
In consequence, we obtain the desired result.
6.2.5 Theorem. There is a category DVS whose objects are differentiable vector
spaces and in which the differentiable vector space of morphisms from V to W is
HomDVS (V, W). As
C ∞ (∗, HomDVS (V, W)) = DVS(V, W),
we recover the category DVS via the global sections functor.
It is also possible to enrich the multicategory structure of DVS. We define
HomDVS (V1 , . . . , Vn |W) as the differentiable vector space whose sections on X are
C ∞ (X, HomDVS (V1 , . . . , Vn |W)) = DVS(V1 , . . . , Vn |C∞ (X, W)),
which is analogous to our definition of the enriched hom.
6.3. Compatibility with CVS. Most of the differentiable vector spaces we

work with in this book appear naturally as convenient vector spaces. Thus, we are
often interested in describing a convenient vector space as a differentiable vector
space, i.e., in describing smooth maps to a convenient vector space. Since CVS
is self-enriched due to its closed symmetric monoidal structure, we can ask how
CVS(V, W), viewed as a differentiable vector space, compares to HomDVS (V, W),
where V, W are convenient. The answer is the best we can hope for.
6.3.1 Lemma. For any manifold M and convenient vector spaces V, W there is an
isomorphism
C ∞ (M, CVS(V, W)) = CVS(V, C ∞ (M, W)),
and it is natural in M. This identification is compatible with the flat connections
along M on the C ∞ (M)-modules for each side. Hence
CVS(V, W) HomDVS (V, W)
as differentiable vector spaces.
Proof. Let us begin by verifying the final assertion — the enriched homs in
CVS and DVS agree — assuming the first isomorphism. Recall the exponential
law: there is an isomorphism of convenient vector spaces
C ∞ (N, C ∞ (M, W)) C ∞ (N × M, W)
for any manifolds N, M and convenient vector space W. (See Theorem 3.12 of
Kriegl and Michor (1997).) Hence we have an isomorphism of differentiable vector
spaces
di fcC ∞ (M, W) C∞ (M, di fc W),
where di fc is the fully faithful functor embedding CVS into DVS. (We have sup-
pressed that notation so far.) Thus for every manifold M and every V, W ∈ CVS,
we have
C ∞ (M, CVS(V, W)) CVS(V, C ∞ (M, W))
DVS(di fc V, di fcC ∞ (M, W))
DVS(di fc V, C∞ (M, di fc W))
C ∞ (M, HomDVS (di fc V, di fc W)),
and so di fc CVS(V, W) HomDVS (di fc V, di fc W) as differentiable vector spaces.
We now turn to proving the main isomorphism.
Lemma 3.7 of Kriegl and Michor (1997) states that the set of smooth curves
C ∞ (R, W) is a convenient vector space when W is convenient. By Lemma 3.11
of Kriegl and Michor (1997), the space C ∞ (V, W) is the limit of C ∞ (R, W) over
all the smooth curves γ : R → V, where the morphisms in the diagram category
are given by pullback along the reparametrizations γ1 = c∗ γ0 , with c ∈ C ∞ (R, R).
(This result should be plausible because a function f : V → W is smooth precisely
when it sends smooth curves into V to smooth curves in W.) Putting these facts
together with the fact that CVS is closed under limits (see theorem 5.0.2,) we see
that C ∞ (V, W) is convenient.
7. VECTOR SPACES ARISING FROM DIFFERENTIAL GEOMETRY 285
Let us now analyze the smooth maps from a manifold M to C ∞ (V, W). We
will repeatedly use the fact that the functor sending a convenient vector space E to
C ∞ (M, E) preserves limits. (This claim follows from combining Lemma 3.8 with
Lemma 3.11 of Kriegl and Michor (1997) and using the fact that limits commute.)
We compute
C ∞ (M, C ∞ (V, W)) = C ∞ (M, lim C ∞ (R, W))
γ∈C ∞ (R,V)
= lim C ∞ (M, C ∞ (R, W))
γ∈C ∞ (R,V)
= lim C ∞ (M × R, W)
γ∈C ∞ (R,V)
= lim C ∞ (R, C ∞ (M, W))
γ∈C ∞ (R,V)
= C (V, C (M, W)).
∞ ∞
(Alternatively, one can simply cite the exponential law, Theorem 3.12 of Kriegl
and Michor (1997).) Now the linear maps are a closed subset of the smooth maps
CVS(V, W) ⊂ C ∞ (V, W)
cut out by linear equations that enforce linearity of the smooth maps. (See the proof
of lemma 5.2.2.) It follows that a smooth map from M to CVS(V, W) is a smooth
map from M to C ∞ (V, W) that is pointwise in CVS(V, W). Similarly an element of
C ∞ (V, C ∞ (M, W)) is linear if for each m ∈ M the corresponding map from V to W
is linear. It follows that
C ∞ (M, CVS(V, W)) = CVS(V, C ∞ (M, W))
as desired.
7. Vector spaces arising from differential geometry

We describe here some aspects of the vector spaces that feature in our work,
namely, sections of vector bundles on manifolds.
7.1. Comparison with tensor product of nuclear spaces. We have shown
that the tensor products on BVS and CVS encode precisely the multilinear maps of
the associated differentiable vector spaces. We have avoided, however, discussing
how these tensor products compare to the various flavors of tensor products for
topological vector spaces. Unfortunately, there are no simple general statements.
Thankfully, for the kinds of vector spaces that we will work with, things are much
better behaved.
Recall that the vector space of smooth sections of a vector bundle is typically
equipped with a Fréchet topology. (That is, it is a complete locally convex topo-
logical vector space whose topology is metrizable.)
7.1.1 Proposition (Kriegl and Michor (1997), Proposition 5.8). If V, W ∈ LCTVS
are metrizable, then
V ⊗π W = V ⊗β W,
where ⊗π denotes the projective tensor product.
To be more pedantic, we might write that

incβ (born(V) ⊗β born(W)) = V ⊗π W,
although we will leave these functors implicit from hereon.
As a corollary, we have the following crucial result.
⊗π W = V b
7.1.2 Corollary. For V, W Fréchet spaces, V b ⊗β W.
Here it is important to note that a Fréchet space is convenient, by Theorem
4.11 of Kriegl and Michor (1997).
Proof. By Theorem 2.14 of Kriegl and Michor (1997), a locally convex vector
space is convenient if and only if it is Mackey complete.
By Result 52.23 of Kriegl and Michor (1997), we know that in a metrizable
locally convex vector space, the convergent sequences coincide with the Mackey-
convergent sequences. Hence, the usual metric space completion agrees with the
Mackey-completion.
Taking the metric space completion of V ⊗π W is thus equivalent to the com-
pletion to a convenient vector space. Because V ⊗π W = V ⊗β W, we obtain the
result.
We are now in a position to prove a result that is crucial for our constructions.
For p : E → M a smooth vector bundle, let E (M) denote the space of smooth
sections, which we will momentarily view as a Fréchet space in LCTVS or as a
convenient vector space, compatibly via our functors. Let E ! denote the vector
bundle E ∨ ⊗ Dens M , obtained by fiberwise tensor product of the fiberwise linear
dual E ∨ with the density line bundle Dens M . The fiberwise evaluation pairing
between E and E ∨ induces a pairing
h−, −i : Ec! (M) × E (M) → R R
(λ, s) 7→ M ev(λ, s),
since fiberwise evaluation produces a compactly supported smooth density on M.
Extending this pairing to the compactly supported distributional sections of E ! , we
!
see that the compactly distributional sections E c (M) are canonically isomorphic to
the strong dual E (M)∗ of E (M).
7.1.3 Proposition. Let E1 (M1 ), . . . , En (Mn ), and F (N) denote smooth sections of
vector bundles (possibly on different manifolds). For concision, we now suppress
the manifolds from the notation.
Then we have an isomorphism of differentiable vector spaces:
! !
⊗β En , F )) di ft (E 1,cb
⊗β · · · b
di fc (CVS(E1b ⊗π E n,cb
⊗π · · · b ⊗π F ).
Proof. We use the preceding corollary. For Fréchet spaces, we know the com-
pleted projective tensor product agrees with the completed bornological tensor
product, so we have another isomorphism
⊗β En , F ) CVS(E1b
⊗β · · · b
CVS(E1b ⊗π En , F ).
⊗π · · · b
7. VECTOR SPACES ARISING FROM DIFFERENTIAL GEOMETRY 287
For complete nuclear spaces V, W, we know

LCTVS(V, W) V ∗b
⊗π W,
where V ∗ denotes the strong dual. Hence, we have
⊗π . . . b
CVS(E1b ⊗π En , F ) (E1b ⊗ π E n )∗ b
⊗π · · · b ⊗π F
E1∗b ⊗π En∗b
⊗π · · · b ⊗π F .
!
For smooth sections of a vector bundle E , we know E ∗ = E c .
Let us now explain an application of this result that we use repeatedly in this
book. Viewing E as a vector space, it is natural to try to define the algebra of
functions on E as something like the symmetric algebra Sym E ∗ on the dual to
E . We need to be careful about what we mean by taking the symmetric algebra,
however, so we give our preferred definition.
7.1.4 Definition. For E the convenient vector space of smooth sections of a vector
bundle E, the uncompleted algebra of functions on E is the convenient vector space
∞
M
CVS(E b⊗β n , R)S n .
n=0
The subscript indicates that one takes the coinvariants with respect to the natural
action of the symmetric group S n on the n-fold tensor product.
This vector space is convenient. “Uncompleted” here refers to using the direct
sum rather than product; it is the uncompleted symmetric algebra — namely, poly-
nomials — rather than the completed symmetric algebra — namely, formal power
series.
We will often work with the completed algebra of functions
∞
Y
O(E ) := CVS(E b⊗β n , R)S n ,
n=0
particularly in the setting of interacting field theories.

Thanks to our work above, we know that we could also work with the com-
pleted projective tensor product instead of the completed bornological tensor prod-
uct.
7.1.5 Corollary. We have
∞
Y !
O(E ) (E c )b⊗π n
Sn
n=0
as convenient vector spaces

In Section 5.7, we provide an alternative description of these algebras using
the enriched hom HomDVS . See, in particular, Lemma 5.7.1.
7.2. Sections of a vector bundle as (co)sheaves of differentiable vector

spaces. Let M be a manifold and E a vector bundle on M. Recall the following
notations:
• E (M) denotes the vector space of smooth sections of E over M,
• Ec (M) denotes the vector space of compactly supported smooth sec-
tions of E over M,
• E (M) denotes the vector space of distributional sections of E over M,
and
• E c (M) denotes the vector space of compactly supported distributional
sections of E over M.
We can view these spaces as living in LCTVS, BVS, CVS, or DVS.
We now discuss what happens as one varies over the opens of M.
op
7.2.1 Lemma. The functor E : Opens M → DVS sending U to E (U) is both a
op
sheaf and a homotopy sheaf. Likewise the functor E : Opens M → DVS sending U
to E (U) is both a sheaf and a homotopy sheaf.
The functor Ec : Opens M → DVS sending U to Ec (U) is both a cosheaf and a
homotopy cosheaf. Likewise the functor E c : Opens M → DVS sending U to E c (U)
is both a cosheaf and a homotopy cosheaf.
As explained in the proof, one can also view these as (co)sheaves with values
in CVS.
Proof. We explain the precosheaf cases; the sheaf cases are parallel.
In Section 4.4, we verified explicitly the gluing axiom. In particular, we used
a partition of unity for a cover of an open U to produce a map from the compactly
supported sections on an open U to the colimit diagram for the cover (or Čech
complex, in the homotopy cosheaf case). The maps we used are maps of convenient
vector spaces, since they consist either of addition or of multiplication by a smooth
function. As CVS is a full subcategory of DVS, these yield maps in DVS as well.
Hence, we obtain cochain homotopy equivalences in DVS as well.
We want to emphasize now why it was so useful that we provided cochain ho-
motopy equivalences in Section 4.4 and not just a quasi-isomorphism. The functor
di fc preserves cochain homotopy equivalences. By contrast, it need not preserve
quasi-isomorphism, since taking cohomology involves taking a cokernel, which
need not to be preserved by di fc .
APPENDIX C
Homological algebra in differentiable vector spaces
1. Introduction
In the study of field theories, one works with vector spaces of an analytical
nature, like the space of smooth functions or distributions on a manifold. To use
the Batalin-Vilkovisky formalism, we need to perform homological algebra in this
setting. The standard approach to working with objects of this nature is to treat
them as topological vector spaces, but it is not obvious how to set up a well-behaved
version of homological algebra with topological vector spaces.
Our approach here breaks the problem into two steps. First, our cochain com-
plexes are constructed out of very nice topological vector spaces that are already
convenient vector spaces. Hence, we view them as cochain complexes of conve-
nient vector spaces, since CVS is a better-behaved category (for our purposes) than
LCTVS. Second, we apply the functor di fc to view them as cochain complexes of
differentiable vector spaces; as CVS is a full subcategory of DVS, nothing dras-
tic has happened. The benefit, however, is that DVS is a Grothendieck abelian
category, so that standard homological algebra applies immediately.
1.1. Motivation. There are a few important observations to make about this
approach.
Recall that di fc does not preserve cokernels, so that di fc need not preserve
cohomology. Given a complex C ∗ in CVS, the cohomology group H k C ∗ is a cok-
ernel computed in CVS. Hence H k (di fcC ∗ ) could be different from di fc (H k C ∗ ).
In consequence, di fc need not preserve quasi-isomorphisms. We will view quasi-
isomorphisms as differentiable cochain complexes as the correct notion and avoid
discussing quasi-isomorphisms as convenient cochain complexes.
The functor di fc does preserve cochain homotopy equivalences, however. Thus,
certain classical results – such as the Atiyah-Bott lemma (see Section D) or the use
of partitions of unity (see Section 4.4) – play a crucial role for us. They establish
explicit cochain homotopy equivalences for convenient cochain complexes, which
go to cochain homotopy equivalences of differentiable cochain complexes. Later
constructions, such as the observables of BV theories, involve deforming the differ-
entials on these differentiable cochain complexes. In almost every situation, these
deformed cochain complexes are filtered in such a way that the first page of the
spectral sequence is the original, undeformed differential. Thus we can leverage
the classical result in the new deformed situation.
289
290 C. HOMOLOGICAL ALGEBRA IN DIFFERENTIABLE VECTOR SPACES
In short, many constructions in geometry and analysis are already done in CVS
(or cochain complexes thereof), but the homological algebra is better done in DVS.
We should explain why it’s better to do the algebra in DVS.
As discussed in Appendix B, convenient vector spaces can be understood via
their associated sheaves on the site of smooth manifolds. (In particular, morphisms
between convenient vector spaces are precisely the maps as such sheaves of vector
spaces.) In other words, doing convenient linear algebra is equivalent to thinking
about smooth families of points in the convenient spaces.
For instance, in thinking about a differential operator D : E → F, one can think
about smoothly varying the inputs and studying the associated family of outputs.
It is natural then to understand the kernel by thinking about smooth families of
solutions to Dφ = 0. As di fc preserves kernels, this way of thinking agrees with
simply picking out the vector subspace annihilated by D.
These ways of thinking differ when it comes to the cokernel. In DVS, one
studies the cokernel of D by thinking about when two smooth families in F differ
by a family of outputs from D. The cokernel in DVS just amounts to studying such
equivalence relations locally, i.e., in very small smooth families. In other words,
we patch together equivalence classes; we need to sheafify the cokernel as pre-
cosheaves. This approach reflects our initial impulse to study something by vary-
ing inputs and outputs. From this perspective, the cokernel in CVS is constructed
by a more opaque procedure. The cokernel in DVS maps to the cokernel in CVS,
viewed as a differentiable vector space, so the two approaches communicate.
Remark: For another approach to these issues, see Wallbridge (n.d.), where a monoidal
combinatorial model category structure is put on the category Ch(CVS) of un-
bounded chain complexes of convenient vector spaces. ^
1.2. Outline of the appendix. After establishing foundational properties of
DVS, we develop some homological techniques that we will use repeatedly through-
out the text. Most are simple reworkings of standard results of homological alge-
bra with sheaves. First, we will discuss spectral sequences, which we often use to
verify some map of differentiable cochain complexes is a quasi-isomorphism. Sec-
ond, we discuss the category of differentiable pro-cochain complexes, which is a
somewhat technical definition but which plays a crucial role for observables of in-
teracting theories. Finally, we will discuss homotopy colimits and explicit methods
for constructing them, which we use in our definition of factorization algebras.
2. Linear algebra and homological algebra in DVS

The category DVS behaves formally like a category of modules, as we establish
in the first part. Thus, it is no surprise that cochain complexes in DVS is also
straightforward, which we explain in the second part.
2.1. Linear algebra. In Section 2.3, we showed that DVS is complete and
cocomplete. Going further, we showed DVS is a locally presentable category. With
a little more work, we obtain the following.
2.1.1 Proposition. The category DVS is abelian.
2. LINEAR ALGEBRA AND HOMOLOGICAL ALGEBRA IN DVS 291
Proof. The category of C ∞ -modules is abelian, as it is the category of modules

for a sheaf of rings on a site. (See, for instance, Theorem 18.1.6 of Kashiwara and
Schapira (2006).) As map of differentiable vector spaces φ : V → W is C ∞ -linear,
it possesses a kernel iφ : Kφ → V and a cokernel qφ : W → Cφ in the category
of C ∞ -modules. We have already shown that these sheaves naturally inherit flat
connections and hence provide the desired kernel and cokernel in differentiable
vector spaces.
It remains to verify that there is a natural isomorphism from the coimage of φ
(i.e., the cokernel of ıφ ) to the image of φ (i.e., the kernel of qφ ). In the abelian
category of C ∞ -modules, there is such an isomorphism, and because φ intertwines
the connections, this map also intertwines the connections, and hence provides a
map of differentiable vector spaces.
To do homological algebra, however, we typically want an abelian category

with much stronger properties, first enumerated by Grothendieck (1957). As we
already know that DVS possesses limits and colimits, it remains to check the ex-
actness of filtered colimits. (This property already ensures that standard formulas
from homological algebra satisfy ∞-categorical universal properties to define our
desired homotopy colimits. See Proposition 5.2.2.)
2.1.2 Lemma. Let I denote a filtered category. Given an exact sequence of I-
diagrams
f g
0 → A → B → C → 0,
with diagrams A, B, C : I → DVS, the colimits also form an exact sequence
0 → colim A → colim B → colim C → 0.

I I I
In other words, DVS is an AB5 abelian category.

Proof. As in the previous proof, it is convenient first to “forget” down to
the category of C ∞ -modules, which is an AB5 abelian category Kashiwara and
Schapira (2006). Thus, in this category, filtered colimits are exact. The underlying
C ∞ -modules of the colimits in DVS are simply those computed in C ∞ -modules, so
it remains to check that the C ∞ -morphisms arising from the colimit in the category
of C ∞ -modules
0 → colim A → colim B → colim C → 0.

I I I
respect the flat connections. This, however, is almost immediate.

For example, for any manifold X and any section of s ∈ colimI B(X), there is
some i ∈ I such that s arises from an element si ∈ Bi (X). The map gi intertwines
the flat connections on Bi and Ci , so we know that ∇Ci (gi si ) = gi (∇Bi si ). Moreover,
this equality holds for the images s j ∈ B j for all j ≥ i. Hence, it will hold for the
section s itself, by the defining property of the connection on the colimits colim B
and colim C.
2.2. Homological algebra in DVS. To summarize the results so far, we have

shown the following.
2.2.1 Theorem. The category DVS is a Grothendieck abelian category.
This result makes it possible to deploy standard arguments from homological
algebra. In fact, Grothendieck’s original work, aimed at unifying homological
algebra for modules over a ring and sheaves of abelian groups on a topological
space, singled out these properties as particularly useful, as they guarantee the
existence of injective objects. For those readers familiar with the basics of sheaf
cohomology, this theorem ensures that our homological constructions behave as
they expect. For those readers who would like deeper treatments, we point to
Chapter 14 of Kashiwara and Schapira (2006), which treats the unbounded derived
category of a Grothendieck abelian category, and to section 1.3 of Lurie (n.d.b),
which studies the stable ∞-category associated to unbounded chain complexes in
a Grothendieck abelian category.
3. Spectral sequences
We will often use versions of spectral sequence arguments in the category of
differentiable complexes.
3.0.1 Definition. A filtered differentiable cochain complex is a sequence of differ-
entiable cochain complexes
· · · → Fn−1C → FnC → Fn+1C → · · ·
where each map Fn−1C → FnC is a monomorphism (explicitly, a monomorphism
in each cohomological degree). A map of filtered differentiable cochain complexes
is a sequence of cochain maps { fn : FnC → Fn D}n∈Z that commute with the maps
of each sequence.
Typically, one views these cochain complexes FnC as subcomplexes of a given

cochain complex C. We will always identify C as colimn FnC, the differentiable
cochain complex that is colimn (FnC)k in cohomological degree k.
The following condition is essential for taking advantage of spectral sequences.
3.0.2 Definition. A filtered differentiable cochain complex is complete if C
limn C/FnC along the canonical map induced by the quotient maps C → C/FnC.
Recall the following useful theorem. (For a proof and further discussion, see
Eilenberg and Moore (1962).)
3.0.3 Theorem (Eilenberg-Moore Comparison Theorem). Let f : C → D be a
map of complete filtered cochain complexes in an AB4 abelian category. Suppose
for each integer n, there exists an integer p such that F pC n = 0 = F p Dn (i.e., in
each cohomological degree n, the filtration is bounded below).
If there is a natural number r such that the induced map between the rth pages
of the spectral sequences
pq pq pq
fr : E r C → E r D
3. SPECTRAL SEQUENCES 293
is an isomorphism for all p, q, then f : C → D is a quasi-isomorphism.

As DVS is Grothendieck, we can (and do throughout this book) invoke the
Eilenberg-Moore comparison theorem.
For ease of later use, we state an immediate corollary.
3.0.4 Lemma. Let C, D be filtered differentiable cochain complexes in which FnC =
0 = Fn D for n < 0. Let f : C → D be a map of filtered differentiable cochain
complexes.
If there is a natural number r such that the map fr between the rth pages of the
spectral sequences is an isomorphism of differentiable cochain complexes, then f
is a quasi-isomorphism of differentiable cochain complexes.
In particular, the following two cases show up several times in the text:
((i)) If the maps Gr fn : FnC/Fn−1C → Fn D/Fn−1 D are quasi-isomorphisms
for all n, then f is a quasi-isomorphism.
((ii)) If the maps fn : FnC → Fn D are quasi-isomorphisms for all n, then f
These two cases amount to the assertion that the E2 or E1 pages are isomorphisms.
Remark: There is another approach to proving these kinds of results that is, in some
sense, more concrete. To verify that a map of differentiable cochain complexes is a
quasi-isomorphism, it is enough to exhibit the quasi-isomorphism on stalks. Taking
stalks, however, sends a differentiable cochain complex to a dg vector space. Ap-
plying these observations to the situation of a map of filtered differentiable cochain
complexes, we see that we can invoke the Eilenberg-Moore comparison theorem
for dg vector spaces at the level of stalks. ^
We have a similar statement for inverse systems, but only under a stronger
hypothesis. A tower is a sequence
vn+1 vn
· · · → Vn+1 −→ Vn −→ Vn−1 → · · ·
where each map vn is an epimorphism (explicitly, an epimorphism in each coho-
mological degree). There is an associated filtration for the limit V = limn Vn by
taking the kernel of the canonical map
Fn V = ker (V → Vn ) .
We can thus hope to use this filtration to learn something about V.
3.0.5 Lemma. Let V• , W• be two towers of differentiable cochain complexes such
that
((i)) the towers are eventually constant: there is some integer n such that
epimorphisms v p and w p are isomorphisms for all p ≥ n;
((ii)) the towers are bounded: there is some integer m such that Vq = 0 = Wq
for all q ≤ m.
If f• : V• → W• is a map of sequences that induces a quasi-isomorphism Ker vn →
Ker wn for all n, then the map lim f : limn Vn → limn Wn is a quasi-isomorphism.
Proof. Under the hypothesis that the towers are eventually constant, we know
that the associated filtrations F• V and F• W are bounded below. Under the hypothe-
sis that they are bounded, the associated filtrations are exhaustive: V = colimn Fn V
and W = colimn Fn W.
The map f• then induces a map of filtered differentiable cochain complexes b f :
V → W. As Fn+1 V/Fn V Ker vn and likewise for the filtration on W, the spectral
sequence argument above implies that b f is a quasi-isomorphism. But lim f = bf , so
we are finished.
4. Differentiable pro-cochain complexes

Most of the examples of factorization algebras in this book take values not in
the category of cochain complexes but in the category of complete filtered cochain
complexes. (For free field theories, it is natural to work just with differentiable
cochain complexes, but we approach interacting theories via formal geometry, aka
perturbative methods, and hence work with completed symmetric algebras.) Some
straightforward modifications of certain constructions thus appear, such as colim-
its. This section is devoted to spelling out some important aspects of the category
of complete filtered differentiable cochain complexes that appears in this book.
4.0.1 Definition. A differentiable pro-cochain cochain complex is a differentiable
cochain complex V with a filtration
· · · ,→ Fn V ,→ Fn+1 V ,→ · · · ,→ F0 V = V
such that the canonical map V → limn V/Fn V is a quasi-isomorphism.
A map of differentiable pro-cochain complexes is a cochain map f : V →
W that preserves the filtration. Such a map is a filtered weak equivalence if the
associated graded map Gr f : Gr V → Gr W is a quasi-isomorphism: for each n,
the map Grn f : Fn V/Fn−1 V → Fn W/Fn−1 W is a quasi-isomorphism.
4.1. Colimits. Colimits are not just given by colimits as differentiable cochain
complexes: one must complete the “naive” colimit. To be explicit, there is an
inclusion functor of differentiable pro-cochain complexes into negatively-filtered
differentiable cochain complexes (i.e., filtered objects such that Fn V = F0 V for all
positive n). This inclusion functor is right adjoint to a “completion functor” that
sends a filtration
· · · ,→ Fn V ,→ Fn+1 V ,→ · · · ,→ F0 V = V
V = limn V/Fn V equipped with the filtration Fn b
to its completion b V = ker(b V →
V/Fn V). Note that completion is “idempotent” in the sense that a complete filtered
cochain complex is isomorphic to its completion.
We define the completed colimit of a diagram of differentiable pro-cochain
complexes as the completion of the colimit in the category of filtered differentiable
cochain complexes. To emphasize the role of completion, we will sometimes write
[ For example, we have the following.
colim.
4. DIFFERENTIABLE PRO-COCHAIN COMPLEXES 295
4.1.1 Definition. Let a set A index a collection of differentiable pro-cochain com-

plexes {Vα | α ∈ A}. The completed direct sum is the limit
M M
Vα = lim Vα /Fn Vα ,
d
α∈A n α∈A
where on the right hand side we use the direct sum of differentiable cochain com-
plexes.
4.2. Stalks. The n-dimensional stalk of a differentiable pro-cochain complex

Stalkn (V) = colim
[ 0∈U⊂Rn V(U)
satisfies Stalkn (V)/F i Stalkn (V) = Stalkn (V/F i V).

An immediate consequence is thus the following.
4.2.1 Lemma. A map f : V → W of differentiable pro-cochain complexes is a fil-
tered weak equivalence if and only if the maps Stalkn ( f ) : Stalkn (V) → Stalkn (W)
are filtered weak equivalences for every n.
4.3. Spectral sequences. We have analogs of the results from Section 3.
4.3.1 Lemma. Let V• , W• be two towers of differentiable pro-cochain complexes
such that
((i)) the towers are eventually constant: there is some integer n such that
epimorphisms v p and w p are isomorphisms for all p ≥ n;
((ii)) the towers are bounded: there is some integer m such that Vq = 0 = Wq
for all q ≤ m.
If f• : V• → W• is a map of sequences that induces a quasi-isomorphism Ker vn →
Ker wn for all n, then the map lim f : limn Vn → limn Wn is a filtered weak equiva-
lence of differentiable pro-cochain complexes.
Proof. This result follows immediately from Lemma 3.0.5.
4.4. Multilinear maps. Differentiable pro-cochain complexes form a multi-

category, just like differentiable cochain complexes.
4.4.1 Definition. Let V1 , . . . , Vk , W be differentiable pro-cochain complexes. In the
multicategory of differentiable pro-cochain complexes, an element of Hom(V1 , . . . , Vk ; W)
is a multilinear map of differentiable cochain complexes
Φ : V1 × · · · × Vk → W = lim W/F i W
that preserves filtrations: if vi ∈ F ri (Vi ), then Φ(v1 , . . . , vk ) ∈ F r1 +···+rk W.
4.5. Working over a differentiable dg ring. The category of differentiable

cochain complexes is a multicategory. Thus, we can talk about commutative dif-
ferentiable dg algebras R: it is a differentiable cochain complex R and a bilinear
map m ∈ Hom(R, R; R), defining the multiplication, that satisfies the axioms of
a commutative algebra. Similarly, a commutative differentiable pro-algebra is a
commutative algebra in the multicategory of differentiable pro-cochain complexes.
In either context, we can define an R-module M to be a differentiable (pro-

)cochain complex equipped with an action of the commutative differentiable (pro-
)algebra R, in the obvious way. We say a map M → M 0 is a weak equivalence if
it is a weak equivalence (as defined above) in the category of differentiable (pro-
)cochain complex. In either context, we say a sequence of R-modules 0 → M1 →
M2 → M3 → 0 is exact if it is exact in the category of differentiable (pro-)cochain
complexes.
The category of modules over a differentiable (pro-)dg algebra R is, as above,
a multicategory. In either case, the multi-maps
HomR (M1 , . . . , Mn ; N)
are the multi-maps in the category of differentiable (pro-)cochain complexes whose
underlying multilinear map M1 × · · · × Mn → N are R-multilinear.
5. Homotopy colimits
A factorization algebra is, in particular, a cosheaf with respect to the Weiss
topology. When working with precosheaves in cochain complexes, the conceptu-
ally correct version of the cosheaf gluing axiom uses the homotopy colimit rather
than the usual colimit. Explicitly, a precosheaf F of cochain complexes is a cosheaf
if for any open set V and any Weiss cover U = {Ui }i∈I of V, the canonical map
hocolimČU F → F (V)
is a quasi-isomorphism, where the left hand side denotes the homotopy colimit
over the Čech cover of U.
In our situation, the homotopy colimit of a diagram F : I → Ch(DVS) can
be described by a familiar cochain complex, as we explain below. In the case of
the cosheaf gluing axiom, we recover precisely the formula that appears in the
definition of a factorization algebra (see Section 1).
We begin by discussing a general situation that includes Ch(DVS) as a special
case. In particular, we explain the homotopical universal property characterizing a
homotopy colimit and then state a theorem providing an explicit construction of a
homotopy colimit in this situation. In the next subsection, we give the proof of this
theorem. Finally, we treat the case of differentiable pro-cochain complexes.
Remark: The Čech complex we use is long-established in homological algebra and
easy to motivate. The reader satisfied with it should probably look no further, as the
arguments below are highly technical. The goal of this section, though, is to ensure
compatibility between our work in this book and the perspective emphasized by
higher category theory. It bridges an odd gap in the literature between homological
and homotopical algebra. ^
5.1. Reminder on homotopy colimits. Let A denote an abelian category, and
let A = Ch(A) denote the category of unbounded cochain complexes in A. Let I
be a category and let F : I → A be a functor, which we will call a diagram of
cochain complexes (or I-diagram to emphasize the shape of the diagram).
5. HOMOTOPY COLIMITS 297
Let’s start by reviewing the definition of a colimit in order to motivate the

definition of a homotopy colimit. There are several different definitions that char-
acterize the same notion.
Note that any cochain complex C provides a “constant” diagram p∗C : I → A
given by pullback along the constant functor p : I → {∗}: for every i, p∗C(i) = C,
and every map in this diagram is the identity. There is thus a “constant diagram”
functor
p∗ : A → Fun(I, A)
given by pullback along p : I → {∗}. Given a diagram F : I → A, a cocone for
F is a natural transformation α : F ⇒ p∗C for some cochain complex C. In other
words, it is the data of a map from every F(i) to C such that for every arrow i → j
in I, the associated composition F(i) → F( j) → C equals the map F(i) → C. A
colimit for F is an initial cocone, if it exists. If the colimit exists, it is unique up to
isomorphism. Every cocone of F thus factors uniquely through the colimit of F.
In the case of the category A, every diagram admits a colimit. (There are
categories where only certain diagrams have a colimit.) We can, in fact, construct
a functor “take the colimit”
colim : Fun(I, A) → A
I
as the left adjoint top∗ .
The universal property satisfied by the colimit of a diagram
ensures that we have the adjunction. This kind of characterization of the colimit
is global in the sense that it works with all I-diagrams simultaneously, rather than
being local, like the cocone definition, which only involves a fixed diagram F.
In the setting of homological algebra, this notion of colimit is not appropri-
ate because we want to consider cochain complexes that are quasi-isomorphic as
equivalent, even though not every quasi-isomorphism is invertible. This consid-
eration leads to two different notions of homotopy colimit, a global and a local
one. The construction we give will be global in nature, because the ∞-category of
cochain complexes is nice enough that homotopy colimits exist for all diagrams.
In a moment, we will sketch the (local) definition of colimit used in the setting of
∞-categories, and we will provide an explicit construction of it in Theorem 5.4.6.
There is, however, a more concrete notion of homotopy colimit requiring no higher
category theory that we now describe. (This notion is a shadow in the homotopy
category of an ∞-categorical notion.)
5.1.1. Since we want to view quasi-isomorphisms as equivalences, we want
to work with the localization of A with respect to the quasi-isomorphisms, which
we call the homotopy category of A and denote Ho(A). Likewise, we want to
consider two I-diagrams F and G as equivalent if there is a natural transformation
η : F ⇒ G such that η(i) : F(i) → G(i) is a quasi-isomorphism for all i. Let
Ho Fun(I, A) denote the homotopy category of I-diagrams.
Observe that the constant diagram functor p∗ preserves weak equivalences and
hence induces a functor Ho(p∗ ) : Ho(A) → Ho Fun(I, A).
5.1.1 Definition. The homotopy colimit over I is the left adjoint to Ho(p∗ ), if it
exists. The homotopy colimit of a diagram F : I → A is denoted hocolimI F.
In other words, the homotopy colimit of a diagram satisfies, at the level of ho-
motopy categories, a version of adjunction in the global definition of colimit. It
is the initial cochain complex, up to quasi-isomorphism, among all cochain com-
plexes, up to quasi-isomorphism, that receive a map from the diagram, up to quasi-
isomorphism of diagrams.
5.1.2. In Section 5.3 below, we describe the ∞-category of cochain com-
plexes in several ways, as a model category and as a quasi-category, following
Lurie. Given a diagram F : I → C in an ∞-category, there is an ∞-category of
cocones for F, which Lurie denotes as CF/ . A cocone is essentially a homotopy co-
herent version of the notion of cocone given in ordinary categories. A colimit for F
is then an initial object in CF/ , which means that any cocone admits a contractible
space of maps from the colimit. (Having a contractible space of morphisms is the
∞-categorical replacement of the notion of unique map.) This ∞-categorical col-
imit is often called the homotopy colimit to distinguish it from the 1-categorical
colimit, when there is the possibility of confusion. See section 1.2.13 of Lurie
(2009a) for the beginning of a systematic discussion.
5.2. A cochain-level construction. We will now explain how to construct the

homotopy colimit when A is a Grothendieck abelian category. In fact, this con-
struction produces an explicit cochain complex from an actual diagram of cochain
complexes. It is a functor from diagrams to A that preserves weak equivalences and
that induces the homotopy colimit at the level of homotopy categories. In short, we
are in the best possible situation.
Let F : I → A be a diagram. Consider the double cochain complex C(F)∗,∗
with M
C I (F)−p,q = F(σ(0))q ,
σ:[p]→I
where [p] denotes the category of the totally ordered set {0 < · · · < p} and σ varies
over all functors from [p] to I, the vertical differential (in the q-grading) is simply
the internal differentials of the complexes F(i), and the horizontal differential is
dhor : C I (F)−p,q → C I (F)1−p,q
Pp
i=0 (−1) (F(σ ◦ δi (0)))σ◦δi ,
(x)σ 7→ i
where δi : [p − 1] → [p] is the order-preserving injection that avoids the element

i ∈ [p]. (Note that this double complex is concentrated in the “left” half-space.)
5.2.1 Definition. The cone of F is ConeI (F) = Tot⊕ (C I (F)), the totalization (using
the direct sum) of the double complex defined above.
Remark: When I = ∆op , ConeI (F) is precisely the direct sum totalization of the
double complex that the Dold-Kan correspondence produces out of the simplicial
cochain complex F. ^
Our goal is to prove the following result (indeed, its ∞-categorical strengthen-
ing).
5.2.2 Theorem. For a Grothendieck abelian category A,

hocolimI F = Ho(ConeI )(F).
In other words, the functor ConeI : Fun(I, A) → A induces, at the level of homo-
topy categories, an adjunction
Ho(ConeI ) : Ho Fun(I, A) Ho(A) : Ho(p∗ ),
where the right adjoint Ho(p∗ ) comes from p∗ , the “constant diagram” functor
given by pullback along p : I → {∗}.
We learned the proof given below from Rune Haugseng, who explained how
standard arguments with simplicial model categories admit a modest modification
to include this case. His approach is stronger than the theorem stated above: it is
a statement at the level of ∞-categories. (Any mistakes in the presentation below
are ours. For an account of a more general statement, see Haugseng (n.d.).)
The proof of this theorem is quite technical, using ideas and results from higher
category theory that are not used elsewhere in this book, so we will not give ex-
pository background. Our main references will be Lurie (2009a) and Lurie (n.d.b),
but the essential ingredients from higher categories are due to Bousfield and Kan
(1972), Dwyer and Kan (1980b), and Dwyer and Kan (1980a).
Remark: Lecture notes of Hörmann Hörmann (n.d.) describe an alternative, very
concrete proof of this theorem, requiring only that A is an AB4 abelian category.
The proof that he outlines in the problem sets relies on the existence of certain
spectral sequences and exhibits explicitly the units and counits of the adjunction at
the level of homotopy categories. ^
5.3. Getting oriented. For A a Grothendieck abelian category, let A denote
the ordinary (i.e., discrete) category Ch(A) of unbounded cochain complexes in A.
Let W denote the collection of quasi-isomorphisms in A as the appropriate notion
of weak equivalences. We will describe here an ∞-categorical refinement of the
earlier notion of the derived category, which is simply the localization A[W −1 ],
following Lurie (n.d.b). In fact, Lurie discusses two different but equivalent ways
to construct this ∞-category, each of which will be useful in the proof.
5.3.1. Via the dg nerve. The following theorem gives us a good grip on the
situation.
5.3.1 Proposition (Lurie (n.d.b), Proposition 1.3.5.3). There is a left proper com-
binatorial model structure on A in which a map f : M → N is
• a weak equivalence if it is a quasi-isomorphism,
• a cofibration if in every degree k, the map f k : M k → N k is a monomor-
phism in A, and
• a fibration if it satisfies the right lifting property with respect to every
acyclic cofibration.
Note that every cochain complex is automatically cofibrant.
Following Lurie, we define the derived ∞-category of A, denoted D(A), to be
Ndg (A◦ ), the differential graded nerve of the subcategory A◦ of fibrant objects of
this model category A. As Lurie shows in Proposition 1.3.5.9, D(A) is a stable

∞-category. It also satisfies a universal property: D(A) is the ∞-categorical local-
ization of A at the quasi-isomorphisms. We now explain this property precisely.
Recall first the ∞-categorical version of the localization of a category, given
in definition 1.3.4.1 of Lurie (n.d.b). Let W be a collection of morphisms in an
∞-category C. A functor q : C → C0 between ∞-categories exhibits C0 as the ∞-
category obtained by inverting W if, for any other ∞-category C00 , precomposition
with q induces a fully faithful embedding q∗ : Fun(C0 , C00 ) → Fun(C, C00 ) whose
essential image consists of the functors from C to C00 that send a morphism in W to
an equivalence in C00 . For C an ordinary category and W a collection of morphisms
in C, we use C[W −1 ] to denote the ∞-category obtained by inverting W in the
∞-category N(C), the nerve of C.
The inclusion functor A◦ ,→ A induces a functor Ndg (A◦ ) → Ndg (A). The
composite functor
D(A) → Ndg (A◦ ) → Ndg (A)
admits a left adjoint that we will denote by L.
5.3.2 Proposition (Lurie (n.d.b), Proposition 1.3.5.15). The composite functor
L
N(A) →
− Ndg (A) −
→ D(A)
exhibits D(A) as the ∞-category obtained by inverting the quasi-isomorphisms.
In particular, the homotopy category of D(A) is equivalent to the usual derived
category.
5.3.2. Via Dold-Kan. There is a different procedure for extracting an ∞-category
from A that is perhaps better-known than the dg nerve construction. First, the sub-
category of fibrant objects A◦ admits a natural enrichment over cochain complexes
of abelian groups, by the usual “hom-complexes.” Next, apply Dold-Kan to ob-
tain a simplicially enriched category A◦∆ . Finally, take the homotopy coherent (or
simplicial) nerve N(A◦∆ ).
We now make more precise those aspects of this construction that we use in
the proof.
Let Ab denote the category of abelian groups, and let B be an AB4 abelian cat-
egory (i.e., B possesses infinite coproducts and hence is a cocomplete Ab-enriched
category). We say that B is tensored over Ab to mean that the following construc-
tion is well-defined. For any abelian group M and object A ∈ B, define M ⊗ A ∈ B
by the property that there is an isomorphism of abelian groups
B(M ⊗ A, B) Ab(M, B(A, B)),
natural in B. This object M ⊗ A exists, as can be seen in a few steps. First, this
“tensor product” clearly exists for finite-generately free abelian groups via finite
direct sum in B, so Zn ⊗ A = A⊕n (this holds in any abelian category). As B
possesses infinite coproducts, we also have the tensor product with any infinitely-
generated free abelian group. Second, any abelian group M is the cokernel of a
morphism between free abelian groups f : F1 → F2 , so define M ⊗ A as the
cokernel of f ⊗ id : F1 ⊗ A → F2 ⊗ A. (Note that this morphism f ⊗ id makes sense
since there is a canonical map Z → B(A, A) that defines what “multiplication by

an integer” in A means.)
Let Ch = Ch(Ab) denote the category of cochain complexes of abelian groups
where morphisms are cochain maps. The category A = Ch(A) has a natural en-
richment over Ch that we will denote ACh . Between any two complexes A, B ∈ A,
there is a cochain complex of abelian groups ACh (A, B) where a degree n element
is a levelwise linear map f = ( f m : Am → Bm+n )m∈Z and where the differential
sends f to dB ◦ f − (−1)m f ◦ dA . In parallel with the argument above, we see that
A = Ch(A) is tensored over Ch: for every cochain complex M ∈ Ch, there is a
natural isomorphism
A(M ⊗ A, B) Ch(M, ACh (A, B)).
We now use the Dold-Kan correspondence to obtain a natural relationship with

simplicial sets.
Recall that every simplicial set X has an associated cochain complex N∗ X =
N• ZX, by applying the Dold-Kan correspondence after taking the level-wise free
abelian group. By Dold-Kan, we can thus produce a natural simplicially enriched
category A∆ : for K a simplicial set, we have
sSet(K, A∆ (A, B)) Ch(N∗ K, ACh (A, B)).
As A is Grothendieck and hence presentable, we know that A∆ is actually tensored

over simplicial sets, so that it is meaningful to write K ⊗ A, with K ∈ sSet and
A ∈ A.
Thus we have a model category A with a natural simplicial enrichment A∆ . (It
does not form a simplicial model category, though. See Warning 1.3.5.4 of Lurie
(n.d.b).) Note that in this enrichment, the hom-spaces are always simplicial abelian
groups and hence Kan complexes. Thus, A∆ is a fibrant simplicial category.
5.3.3. The equivalence. Proposition 1.3.1.17 of Lurie (n.d.b) states that for a
dg category C, there is an equivalence of ∞-categories between its dg nerve Ndg (C)
and N(C∆ ), the homotopy coherent nerve of it simplicial enrichment. Hence, we
see that D(A) = Ndg (A◦ ) is equivalent to N(A◦∆ ).
Propositions 1.3.5.13 and 1.3.5.15 imply that D(A) is an ∞-category obtained
by inverting the quasi-isomorphisms W. Hence we see that D(A) ' N(A∆ )[W −1 ]
as well.
5.4. The ∞-categorical version of the theorem. Our goal is to find a way of
describing colimits in the ∞-category D(A) in a concrete way. We will see that
the construction ConeI from above provides this colimit.
5.4.1. As an example for the kind of result we want, we review the following
construction of the homotopy limit of a diagram of simplicial sets, since we will
use it in our proof.
The twisted arrow category Tw(C) of a category C has objects the morphisms
in C but a morphism from f ∈ C(x, y) to f 0 ∈ C(x0 , y0 ) is a commuting diagram
X / X0 .
f f0

Yo Y0
Note that there is a natural functor from Tw(C) to C × Cop sending an object to its
source and target. Given a functor F : C × Cop → D, the end of F is the limit of
the composition Tw(C) → C × Cop → D. (The coend is the colimit.)
A cosimplicial simplicial set is a functor X • : ∆ → sSet. A key example is
∆ whose n-cosimplex is precisely the simplicial set ∆[n]. It is a kind of fattened
•
model of a point.
Recall that the totalization of a cosimplicial object X • : ∆ → sSet is the end of
sSet∆ (∆• , X • ), where sSet∆ will denote the internal hom of sSet. This construction
is dual to the more familiar geometric realization. (Geometric realization assem-
bles a space by attaching cells, so it is a colimit. Totalization assembles a space as
a tower of extensions, so it is a limit.)
5.4.1 Definition. Let F : I → sSet be a diagram of simplicial sets. The cosimpli-
cial cobar construction B̌• F is the cosimplicial simplicial set whose n-cosimplices
are Y
B̌n F = F(σ(n)),
σ:[n]→I
where σ runs over all functors from the poset [n] into I. The cobar construction
B̌(F) is the totalization of B̌• F.
Remark: The cosimplicial cobar construction keeps track, in a very explicit sense,
of all the ways that a given object F(i) of the diagram is mapped into. It remem-
bers every finite sequence of arrows whose terminus is F(i). The cobar construction
then looks at ways of coherently mapping a “fat point” into this enormous arrange-
ment. ^
Here is the statement at the level of homotopy categories.
5.4.2 Proposition (Bousfield and Kan (1972), XI.8.1). Let F : I → sSet be a
diagram such that F(i) is a Kan complex for every i ∈ I (i.e., F is objectwise
fibrant). Then B̌(F) is a representative for the homotopy limit holimI F in Ho(sSet).
There is a corresponding ∞-categorical statement. Let S denote the ∞-category
of spaces, which is N(sSet◦ ), i.e., the homotopy coherent nerve of sSet◦ , the cate-
gory of Kan complexes, which is the subcategory of fibrant objects in sSet. As an
immediate corollary of the proceeding proposition and Theorem 4.2.4.1 of Lurie
(2009a), we have the following.
5.4.3 Proposition. Let F : I → sSet◦ be a diagram. Then B̌(F) is a limit of the
associated diagram in spaces S.
5.4.2. We will parallel this development to show that our cone construction
ConeI is a kind of bar construction and hence a colimit in ∞-categories.
Above, we used the fact that a diagram of simplicial sets naturally produced a
cosimplicial simplicial set, and we could then talk about mapping the “fat point”
into this cosimplicial cobar construction. Here, we will start with a diagram of
cochain complexes in A and then produce a simplicial cochain complex in A. To
do a bar construction, we need to find a replacement for the “fat point” ∆• in the
setting of cochain complexes.
We will use 4n to denote the cochain complex N∗ ∆[n] ∈ Ch, the natural
cochain complex associated to the simplicial set ∆[n] by the Dold-Kan correspon-
dence. There is thus a natural cosimplicial cochain complex 4• that provides our
“fat point.” Given a simplicial cochain complex X• : ∆op → A = Ch(A), we thus
have a functor
X• ⊗ 4• : ∆op × ∆ → A,
using the fact that A is tensored over Ch. The realization of X• is the colimit of the
composite functor Tw(∆op ) → ∆op × ∆ → A.
5.4.4 Definition. Let F : I → A be a diagram in A. The simplicial bar construction
B• F is the simplicial object in A whose n-simplices are
M
Bn F = F(σ(0)),
σ:[n]→I
where σ runs over all functors from the poset [n] into I. The bar construction B(F)
is the realization of B• F.
It is a combinatorial exercise to verify the following.

5.4.5 Lemma. For a diagram F : I → A, there is an isomorphism B(F)
ConeI (F).
Our theorem can then be stated as follows.
5.4.6 Theorem. Let F : I → A be a diagram of unbounded cochain complexes in
the Grothendieck abelian category A. Then the bar construction B(F) is a colimit
of the associated diagram in D(A).
Taking the underlying homotopy category of D(A), we obtain Theorem 5.2.2.
5.5. The proof of Theorem 5.4.6. Let A∆ be the natural simplicial enrich-
ment of A, as described in Section 5.3.2. We have seen that the derived ∞-category
D(A) is equivalent to N(A∆ )[W −1 ], with W the quasi-isomorphisms. By Theorem
4.2.4.1 of Lurie (2009a), we can compute the colimit in D(A) by computing the ho-
motopy colimit in the simplicially enriched category A∆ with quasi-isomorphisms
as the weak equivalences.
To show that B(F) is the homotopy colimit in A∆ , it suffices to show that for
every fibrant object Z ∈ A, the space A∆ (B(F), Z) is a homotopy limit in sSet of
the diagram
A∆ (F(−), Z) : I → sSet .
Indeed, this property characterizes the homotopy colimit of F. (See Remark A.3.3.13
of Lurie (2009a).)
Observe that for any simplicial set K, we have natural isomorphisms
sSet(K, A∆ (B(F), Z)) A(K ⊗ B(F), Z)

limop A(K ⊗ 4• ⊗ B• F, Z)
Tw(∆ )
sSet(K, limop A∆ (4• ⊗ B• F, Z)).
Tw(∆ )
The Yoneda lemma then implies we have a natural isomorphism
A∆ (B(F), Z) limop A∆ (4• ⊗ B• F, Z).

Tw(∆ )
We now analyze the right hand side in more detail.

The Alexander-Whitney and Eilenberg-Zilber maps provide a cochain homo-
topy equivalence
ACh (4n ⊗ Bm F, Z) Ch(4n , A∆ (Bm F, Z))
for every m and n. Applying Dold-Kan thus produces a simplicial homotopy equiv-
alence
A∆ (∆[n] ⊗ Bm F, Z) sSet∆ (∆[n], A∆ (Bm F, Z))
for every m and n. Taking the limit over the twisted arrow category, we obtain a
simplicial homotopy equivalence
A∆ (B(F), Z) limop sSet∆ (∆• , A∆ (B• F, Z)),

Tw(∆ )
and hence a weak equivalence of simplicial sets. The right hand side is isomorphic
to the cobar construction,
lim sSet∆ (∆• , A∆ (B• F, Z)) B̌(F),

Tw(∆op )
which computes the homotopy limit in sSet by Lemma 5.4.2.
5.6. Homotopy colimits in differentiable pro-cochain complexes. We often

work as well with diagrams of differentiable pro-cochain complexes. The construc-
tion ConeI makes sense in this setting, so long as one replaces the direct sum total-
ization by its natural completion (i.e., use the completed direct sum rather than the
direct sum). We want it to satisfy the homotopical universal property that ensures
this cone is a homotopy colimit.
Let A denote a Grothendieck abelian category and A = Ch(A). We will de-
scribe homotopy colimits of complete filtered objects in A. Again, our argument is
quite technical. We start with some preliminaries and then deduce the result from
our earlier work.
Remark: These arguments arose in conversations with Dmitri Pavlov. ^
5.6.1. Sequences, filtrations, and completions. Let Z≤0 denote the category
whose objects are the nonpositive integers {n ≤ 0} with a single morphism n → m if
n ≤ m and no morphism if n > m. Let Seq denote the functor category Fun(Z≤0 , A).
We call an object
· · · X(n) → X(n + 1) → · · · → X(0)
in Seq a sequence. Equip Seq with the projective model structure: a map of se-
quences f : X → Y is
• a weak equivalence if f (n) : X(n) → Y(n) is a quasi-isomorphism for
every n, and
• a fibration if f (n) : X(n) → Y(n) is a fibration for every n.
Note that we are working with levelwise weak equivalences, which we denote WL .
The following observation gives a nice explanation for the role of filtered
cochain complexes: they are the cofibrant sequences.
5.6.1 Lemma. A sequence X is cofibrant if and only if every map X(n) → X(n + 1)
is a cofibration (i.e., a monomorphism in every cohomological degree).
Hence, we introduce the following definition.
5.6.2 Definition. Let A f il denote the category Seqc of cofibrant objects, i.e., the
filtered cochain complexes.
We want to work with a different notion of weak equivalence than WL . We say

a map of sequences f : X → Y is a filtered weak equivalence if the induced maps
hocofib(X(n) → X(n + 1)) → hocofib(Y(n) → Y(n + 1))
are quasi-isomorphisms for every n. Here “hocofib” is the homotopical version of
the cokernel, and it can be computed using the standard cone construction of ho-
mological algebra. This condition is the homotopical version of saying we have
a quasi-isomorphism on the associated graded, since we are taking the homo-
topy quotient rather than the naive quotient. On the filtered cochain complexes,
it agrees with our earlier notion of a filtered weak equivalence, since the homotopy
cofiber agrees with the cokernel. (This observation gives an explanation for why
one should work with filtered complexes: one can do the naive computations.)
Let WF denote the collection of filtered weak equivalences. We want to get
a handle on the ∞-category A f il [WF−1 ], which can be called the filtered derived
∞-category of A.
The complete filtered cochain complexes will be the crucial tool for us. Recall
that a filtered cochain complex X ∈ A f il is complete if the canonical map X(0) →
limn X(0)/X(n) is a quasi-isomorphism. (We apologize for the slight change of
notation from Fn X to X(n), but this notation makes clearer the comparison with
sequences, which is useful in this context.)
5.6.3 Definition. Let Ac f il denote the category of completed filtered cochain com-
plexes in A (i.e., the pro-cochain complexes in A, in analogy with the terminology
of Section 4).
The inclusion functor Ac f il ,→ A f il is right adjoint to a “completion functor”

that sends a filtration
X = (· · · ,→ X(n) ,→ X(n + 1) ,→ · · · ,→ X(0))
to its completion X b where X(0)b = limn X(0)/X(n) equipped with the filtration
X(n) = ker(X → X(0)/X(n)). Note that completion is “idempotent” in the sense
b b
that a complete filtered cochain complex is isomorphic to its completion.
5.6.4 Lemma. The inclusion (Ac f il , WF ) ,→ (A f il , WF ) induces an equivalence be-
tween their Dwyer-Kan simplicial localizations, and hence we have an equivalence
Ac f il [WF−1 ] ' A f il [WF−1 ]
between the ∞-categories presented by these categories with weak equivalences.
Proof. Under this inclusion-completion adjunction, we see that filtered weak
equivalences are preserved. Moreover, for every filtered cochain complex X, the
canonical map X → X b is a filtered weak equivalence. By Corollary 3.6 of Dwyer
and Kan (1980a), these have weakly equivalent hammock localizations, which im-
plies that the standard simplicial localizations are weakly equivalent.
5.6.2. The theorem.
5.6.5 Definition. The completed cone construction Cone [ I (F) of a diagram F : I →
⊕
Ac f il is given by Tot (C I (F)), the totalization (via the completed direct sum) of the
b
double complex C I (F) of Section 5.2.
5.6.6 Theorem. Let F : I → Ac f il be a diagram. The completed cone construction

[ I (F) is a model for the colimit of F in the ∞-category Ac f il [W −1 ].
Cone F
Proof. We can view this diagram as living in A f il (and hence also the category
Seq), and by Lemma 5.6.4, it suffices to compute the colimit in A f il [WF−1 ]. We
will thus verify the (uncompleted) cone construction provides a model for this col-
imit. Completing this cone will provide a complete filtered cochain complex that
is filtered weak equivalent, and hence also a model for the desired colimit.
We compute this colimit in filtered cochain complexes by using model cate-
gories. Let SeqL denote Seq equipped with the projective model structure, so the
weak equivalences are WL . Observe that WL ⊂ WF as collections of morphisms
inside Seq. Then the left Bousfield localization at WF , which we will denote SeqF ,
has the same cofibrations, by definition. Hence, if one wants to compute homotopy
colimits in SeqF , they agree with the homotopy colimits in SeqL . (More explicitly,
the projective model structures on Fun(I, SeqF ) and Fun(I, SeqL ) have the same
cofibrant replacement functor, so homotopy colimits can be computed in either
category.) In SeqL , our computation becomes simple: a homotopy colimit of di-
agram in such a functor category can be computed objectwise, so we can apply
Theorem 5.4.6 objectwise. This construction naturally produces ConeI with the
natural filtration.
APPENDIX D
The Atiyah-Bott Lemma
Atiyah and Bott (1967) shows that for an elliptic complex (E , Q) on a compact
closed manifold M, with E the smooth sections of a Z-graded vector bundle, there
is a homotopy equivalence (E , Q) ,→ (E , Q) into the elliptic complex of distribu-
tional sections. The argument follows from the existence of parametrices for el-
liptic operators. This result was generalized to the non-compact case in Tarkhanov
(1987).
Let M be a smooth manifold, not necessarily compact. Let (E , Q) be an elliptic
complex on M. Let E denote the complex of distributional sections of E . We will
endow both E and E with their natural topologies.
0.0.1 Lemma (Lemma 1.7, Tarkhanov (1987)). The natural inclusion (E , Q) ,→
(E , Q) admits a continuous homotopy inverse.
The continuous homotopy inverse
Φ:E →E
is given by a kernel KΦ ∈ E ! ⊗ E with proper support. The homotopy S : E → E
is a continuous linear map with
[Q, S ] = Φ − Id .
!
The kernel KS for S is a distribution, that is, an element of E ⊗ E with proper
support.
Proof. We will reproduce the proof in Tarkhanov (1987). Choose a metric on

E and a volume form on M. (If E is a complex vector bundle, use a Hermitian
metric.) Let Q∗ be the formal adjoint to Q, which is a degree −1 differential op-
erator, and form the graded commutator D = [Q, Q∗ ]. This operator D is elliptic
on each space E i , the cohomological degree i part of E . Thus, by standard results
in the theory of pseudodifferential operators, there is a parametrix P for D. The
!
kernel KP is an element of E ⊗ E , and the corresponding operator P : Ec → E is
an inverse for D up to smoothing operators.
By multiplying KP by a smooth function on M × M that is 1 in a neighborhood
of the diagonal, we can assume that KP has proper support. Hence, we can extend
P to a map P : E → E that is still a parametrix: thus P ◦ D and D ◦ P both differ
from the identity by smoothing operators.
The homotopy S is now defined by S = Q∗ P.
307
308 D. THE ATIYAH-BOTT LEMMA
Note that the homotopy inverse Φ : E → E and the homotopy S : E → E

produce a cochain homotopy equivalence in DVS, after applying the functor di ft :
LCTVS → DVS.
The following corollary is quite useful in the study of free field theories.
Let Ec denote the cosheaf that assign to an open U ⊂ M the cochain com-
plex of differentiable vector spaces (Ec (U), Q), namely the compactly supported
smooth sections on U of the graded vector bundle E. Note that, somewhat abu-
sively, we are viewing these as differentiable vector spaces rather than topologi-
cal vector spaces. Similarly, let E c denote the cosheaf sending U to (E c (U), Q),
namely the compactly supported distributional sections on U of the graded vec-
tor bundle E, viewed as a differentiable cochain complex. (See Section 7.2 and
Lemma 5.1.2 for the proof that we have a cosheaf.)
0.0.2 Lemma. The canonical map Ec → E c is a quasi-isomorphism of cosheaves
of differentiable cochain complexes.
Proof. The Atiyah-Bott-Tarkhanov lemma assures us that Ec (U) → E c (U) is
a quasi-isomorphism on every open U.
Bibliography
Adámek, Jiřı́, and Rosický, Jiřı́. 1994. Locally presentable and accessible cate-
gories. London Mathematical Society Lecture Note Series, vol. 189. Cambridge
University Press, Cambridge.
Atiyah, M., and Bott, R. 1967. A Lefschetz Fixed Point Formula for Elliptic Com-
plexes: I. Ann. of Math. (2), 86(2), 374–407.
Ayala, David, and Francis, John. 2015. Factorization homology of topological
manifolds. J. Topol., 8(4), 1045–1084.
Ayala, David, Francis, John, and Tanaka, Hiro Lee. Factorization homology for
stratified manifolds. Available at http://arxiv.org/abs/1409.0848.
Ayala, David, Francis, John, and Rozenblyum, Nick. Factorization homology from
higher categories. Available at http://front.math.ucdavis.edu/1504.
04007.
Behnke, Heinrich, and Stein, Karl. 1949. Entwicklung analytischer Funktionen auf
Riemannschen Flächen. Math. Ann., 120, 430–461.
Beilinson, Alexander, and Drinfeld, Vladimir. 2004. Chiral algebras. American
Mathematical Society Colloquium Publications, vol. 51. Providence, RI: Amer-
ican Mathematical Society.
Ben Zvi, David, Brochier, Adrien, and Jordan, David. Integrating quantum groups
over surfaces: quantum character varieties and topological field theory. Avail-
able at http://arxiv.org/abs/1501.04652.
Boardman, J. M., and Vogt, R. M. 1973. Homotopy invariant algebraic structures
on topological spaces. Lecture Notes in Mathematics, Vol. 347. Springer-Verlag,
Berlin-New York.
Boavida de Brito, Pedro, and Weiss, Michael. 2013. Manifold calculus and homo-
topy sheaves. Homology Homotopy Appl., 15(2), 361–383.
Bödigheimer, C.-F. 1987. Stable splittings of mapping spaces. Pages 174–187 of:
Algebraic topology (Seattle, Wash., 1985). Lecture Notes in Math., vol. 1286.
Springer, Berlin.
Borceux, Francis. 1994a. Handbook of categorical algebra. 1. Encyclopedia of
Mathematics and its Applications, vol. 50. Cambridge: Cambridge University
Press. Basic category theory.
Borceux, Francis. 1994b. Handbook of categorical algebra. 2. Encyclopedia of
Mathematics and its Applications, vol. 51. Cambridge University Press, Cam-
bridge. Categories and structures.
Bott, Raoul, and Tu, Loring W. 1982. Differential forms in algebraic topology.
Graduate Texts in Mathematics, vol. 82. Springer-Verlag, New York-Berlin.
309
310 Bibliography
Bousfield, A. K., and Kan, D. M. 1972. Homotopy limits, completions and local-
izations. Lecture Notes in Mathematics, Vol. 304. Springer-Verlag, Berlin-New
York.
Calaque, Damien, and Willwacher, Thomas. 2015. Triviality of the higher formal-
ity theorem. Proc. Amer. Math. Soc., 143(12), 5181–5193.
Costello, Kevin. The A∞ operad and the moduli space of curves. Available at
http://arxiv.org/abs/math/0402015.
Costello, Kevin. Renormalisation and the Batalin-Vilkovisky formalism. Available
at http://arxiv.org/abs/0706.1533.
Costello, Kevin. Supersymmetric gauge theory and the Yangian. Available at
http://arxiv.org/abs/1303.2632.
Costello, Kevin. 2007. Topological conformal field theories and Calabi-Yau cate-
gories. Adv. Math., 210(1), 165–214.
Costello, Kevin. 2010. A geometric construction of the Witten genus, I. In: Pro-
ceedings of the International Congress of Mathematicians, (Hyderabad, 2010).
Costello, Kevin. 2011a. A geometric construction of the Witten genus, II. Available
at http://arxiv.org/abs/1112.0816.
Costello, Kevin. 2011b. Renormalization and effective field theory. Mathematical
Surveys and Monographs, vol. 170. American Mathematical Society, Provi-
dence, RI.
Costello, Kevin. 2013. Notes on supersymmetric and holomorphic field theories in
dimensions 2 and 4. Pure Appl. Math. Q., 9(1), 73–165.
Costello, Kevin. 2014. Integrable lattice models from four-dimensional field the-
ories. Pages 3–23 of: String-Math 2013. Proc. Sympos. Pure Math., vol. 88.
Amer. Math. Soc., Providence, RI.
Costello, Kevin, and Li, Si. 2011. Quantum BCOV theory on Calabi-Yau man-
ifolds and the higher genus B-model. Available at http://arxiv.org/abs/
1201.4501.
Costello, Kevin, and Scheimbauer, Claudia. 2015. Lectures on mathematical as-
pects of (twisted) supersymmetric gauge theories. Pages 57–87 of: Mathemati-
cal aspects of quantum field theories. Math. Phys. Stud. Springer, Cham.
Curry, Justin. Sheaves, Cosheaves and Applications. Available at http://arxiv.
org/abs/1303.3255.
Deligne, Pierre, Etingof, Pavel, Freed, Daniel S., Jeffrey, Lisa C., Kazhdan, David,
Morgan, John W., Morrison, David R., and Witten, Edward (eds). 1999. Quan-
tum fields and strings: a course for mathematicians. Vol. 1, 2. American Mathe-
matical Society, Providence, RI; Institute for Advanced Study (IAS), Princeton,
NJ. Material from the Special Year on Quantum Field Theory held at the Insti-
tute for Advanced Study, Princeton, NJ, 1996–1997.
Dugger, Daniel. A primer on homotopy colimits. Available at http://pages.
uoregon.edu/ddugger/.
Dwyer, W. G., and Kan, D. M. 1980a. Calculating simplicial localizations. J. Pure
Appl. Algebra, 18(1), 17–35.
Dwyer, W. G., and Kan, D. M. 1980b. Simplicial localizations of categories. J.
Pure Appl. Algebra, 17(3), 267–284.
Bibliography 311
Eilenberg, Samuel, and Moore, John C. 1962. Limits and spectral sequences.
Topology, 1, 1–23.
Etingof, Pavel, Gelaki, Shlomo, Nikshych, Dmitri, and Ostrik, Victor. 2015. Ten-
sor categories. Mathematical Surveys and Monographs, vol. 205. American
Mathematical Society, Providence, RI.
Félix, Yves, Halperin, Stephen, and Thomas, Jean-Claude. 2001. Rational homo-
topy theory. Graduate Texts in Mathematics, vol. 205. Springer-Verlag, New
York.
Forster, Otto. 1991. Lectures on Riemann surfaces. Graduate Texts in Mathematics,
vol. 81. New York: Springer-Verlag. Translated from the 1977 German original
by Bruce Gilligan, Reprint of the 1981 English translation.
Francis, John. 2013. The tangent complex and Hochschild cohomology of En -
rings. Compos. Math., 149(3), 430–480.
Freed, Daniel S. 1994. Higher algebraic structures and quantization. Comm. Math.
Phys., 159(2), 343–398.
Frenkel, Edward, and Ben-Zvi, David. 2004. Vertex algebras and algebraic curves.
Second edn. Mathematical Surveys and Monographs, vol. 88. Providence, RI:
American Mathematical Society.
Fresse, Benoit. Homotopy of Operads and the Grothendieck-Teichmüller
group. Available at http://math.univ-lille1.fr/˜fresse/
OperadHomotopyBook/.
Friedman, Greg. 2012. Survey article: an elementary illustrated introduction to
simplicial sets. Rocky Mountain J. Math., 42(2), 353–423.
Gelfand, Sergei I., and Manin, Yuri I. 2003. Methods of homological algebra.
Second edn. Springer Monographs in Mathematics. Berlin: Springer-Verlag.
Getzler, E. 1994. Batalin-Vilkovisky algebras and two-dimensional topological
field theories. Comm. Math. Phys., 159(2), 265–285.
Ginot, Grégory. 2015. Notes on factorization algebras, factorization homology and
applications. Pages 429–552 of: Mathematical aspects of quantum field theories.
Math. Phys. Stud. Springer, Cham.
Ginot, Grégory, Tradler, Thomas, and Zeinalian, Mahmoud. Derived Higher
Hochschild cohomology, Brane topology and centralizers of En -algebra maps.
Available at http://arxiv.org/abs/1205.7056.
Ginot, Grégory, Tradler, Thomas, and Zeinalian, Mahmoud. Derived Higher
Hochschild Homology, Topological Chiral Homology and Factorization Alge-
bras. Available at http://arxiv.org/abs/1011.6483.
Glimm, James, and Jaffe, Arthur. 1987. Quantum physics. Second edn. Springer-
Verlag, New York. A functional integral point of view.
Goerss, Paul, and Schemmerhorn, Kristen. 2007. Model categories and simplicial
methods. Pages 3–49 of: Interactions between homotopy theory and algebra.
Contemp. Math., vol. 436. Providence, RI: Amer. Math. Soc.
Goerss, Paul G., and Jardine, John F. 2009. Simplicial homotopy theory. Modern
Birkhäuser Classics. Basel: Birkhäuser Verlag. Reprint of the 1999 edition
[MR1711612].
312 Bibliography
Grady, Ryan, Li, Qin, and Li, Si. BV quantization and the algebraic index. Avail-
Griffiths, Phillip, and Harris, Joseph. 1994. Principles of algebraic geometry. Wi-
ley Classics Library. New York: John Wiley & Sons Inc. Reprint of the 1978
original.
Grothendieck, A. 1952. Résumé des résultats essentiels dans la théorie des produits
tensoriels topologiques et des espaces nucléaires. Ann. Inst. Fourier Grenoble,
4, 73–112 (1954).
Grothendieck, Alexander. 1957. Sur quelques points d’algèbre homologique.
Tôhoku Math. J. (2), 9, 119–221.
Grothendieck, Alexandre. 1955. Produits tensoriels topologiques et espaces
nucléaires. Mem. Amer. Math. Soc., 1955(16), 140.
Gunning, Robert C., and Rossi, Hugo. 1965. Analytic functions of several complex
variables. Englewood Cliffs, N.J.: Prentice-Hall Inc.
Gwilliam, Owen. 2012. Factorization Algebras and Free Field Theories. ProQuest
LLC, Ann Arbor, MI. Thesis (Ph.D.)–Northwestern University.
Gwilliam, Owen, and Grady, Ryan. 2014. One-dimensional Chern-Simons theory
and the Â genus. Algebr. Geom. Topol., 14(4), 2299–2377.
Gwilliam, Owen, and Johnson-Freyd, Theo. How to derive Feynman diagrams
for finite-dimensional integrals directly from the BV formalism. Available at
Hatcher, Allen. Notes on Basic 3-Manifold Topology. Available at http://www.
math.cornell.edu/˜hatcher/.
Haugseng, Rune. Weakly simplicial model categories and homotopy colimits.
Available upon request.
Horel, Geoffroy. Factorization homology and calculus à la Kontsevich Soibelman.
Available at http://arxiv.org/abs/1307.0322.
Hörmander, Lars. 2003. The analysis of linear partial differential operators.
I. Classics in Mathematics. Berlin: Springer-Verlag. Distribution theory
and Fourier analysis, Reprint of the second (1990) edition [Springer, Berlin;
MR1065993 (91m:35001a)].
Hörmann, Fritz. Homotopy Limits and Colimits in Nature: A Motivation
for Derivators. Available at http://home.mathematik.uni-freiburg.de/
hoermann/holimnature.pdf.
Huang, Yi-Zhi. 1997. Two-dimensional conformal geometry and vertex operator
algebras. Progress in Mathematics, vol. 148. Boston, MA: Birkhäuser Boston
Inc.
Johansen, A. 1995. Twisting of N = 1 SUSY gauge theories and heterotic topo-
logical theories. Internat. J. Modern Phys. A, 10(30), 4325–4357.
Kapustin, Anton. 2010. Topological field theory, higher categories, and their ap-
plications. Pages 2021–2043 of: Proceedings of the International Congress of
Mathematicians. Volume III. Hindustan Book Agency, New Delhi.
Kapustin, Anton, and Saulina, Natalia. 2011. Topological boundary conditions in
abelian Chern-Simons theory. Nuclear Phys. B, 845(3), 393–435.
Bibliography 313
Kashiwara, Masaki, and Schapira, Pierre. 2006. Categories and sheaves.

Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of
Mathematical Sciences], vol. 332. Springer-Verlag, Berlin.
Knudsen, Ben. Higher enveloping algebras. Available at http://www.math.
northwestern.edu/˜knudsen/.
Kontsevich, Maxim. 1994. Feynman diagrams and low-dimensional topology.
Pages 97–121 of: First European Congress of Mathematics, Vol. II (Paris,
1992). Progr. Math., vol. 120. Basel: Birkhäuser.
Kriegl, Andreas, and Michor, Peter W. 1997. The convenient setting of global
analysis. Mathematical Surveys and Monographs, vol. 53. Providence, RI:
American Mathematical Society.
Leinster, Tom. 2004. Higher operads, higher categories. London Mathematical
Society Lecture Note Series, vol. 298. Cambridge: Cambridge University Press.
Li, Qin, and Li, Si. 2016. On the B-twisted topological sigma model and Calabi–
Yau geometry. J. Differential Geom., 102(3), 409–484.
Loday, Jean-Louis. 1998. Cyclic homology. Second edn. Grundlehren der Mathe-
matischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
vol. 301. Berlin: Springer-Verlag. Appendix E by Marı́a O. Ronco, Chapter 13
by the author in collaboration with Teimuraz Pirashvili.
Loday, Jean-Louis, and Vallette, Bruno. 2012. Algebraic operads. Grundlehren
der Mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences], vol. 346. Heidelberg: Springer.
Lurie, Jacob. Derived Algebraic Geometry X: Formal Moduli Problems. Available
at http://math.mit.edu/˜lurie/papers.
Lurie, Jacob. Higher Algebra. Available at http://math.mit.edu/˜lurie/
papers.
Lurie, Jacob. 2009a. Higher topos theory. Annals of Mathematics Studies, vol.
170. Princeton University Press, Princeton, NJ.
Lurie, Jacob. 2009b. On the classification of topological field theories. Pages 129–
280 of: Current developments in mathematics, 2008. Int. Press, Somerville,
MA.
Mac Lane, Saunders. 1998. Categories for the working mathematician. Second
edn. Graduate Texts in Mathematics, vol. 5. New York: Springer-Verlag.
Matsuoka, Takuo. Descent and the Koszul duality for factorization algebras. Avail-
McDuff, Dusa. 1975. Configuration spaces of positive and negative particles.
Topology, 14, 91–107.
Morrison, Scott, and Walker, Kevin. 2012. Blob homology. Geom. Topol., 16(3),
1481–1607.
Ramanan, S. 2005. Global calculus. Graduate Studies in Mathematics, vol. 65.
Providence, RI: American Mathematical Society.
Salvatore, Paolo. 2001. Configuration spaces with summable labels. Pages 375–
395 of: Cohomological methods in homotopy theory (Bellaterra, 1998). Progr.
Math., vol. 196. Birkhäuser, Basel.
314 Bibliography
Scheimbauer, Claudia. On fully extended topological field theories. Ph.D. thesis

at ETH Zurich.
Segal, Graeme. 1973. Configuration-spaces and iterated loop-spaces. Invent.
Math., 21, 213–221.
Segal, Graeme. 2004. The definition of conformal field theory. Pages 421–577 of:
Topology, geometry and quantum field theory. London Math. Soc. Lecture Note
Ser., vol. 308. Cambridge: Cambridge Univ. Press. First cirulated in 1988.
Segal, Graeme. 2010. Locality of holomorphic bundles, and locality in quantum
field theory. Pages 164–176 of: The many facets of geometry. Oxford Univ.
Press, Oxford.
Serre, Jean-Pierre. 1953. Quelques problèmes globaux relatifs aux variétés de
Stein. Pages 57–68 of: Colloque sur les fonctions de plusieurs variables, tenu à
Bruxelles, 1953. Georges Thone, Liège.
Stacks Project Authors, The. 2016. Stacks Project. http://stacks.math.
columbia.edu.
Stasheff, Jim. 2004. What is . . . an operad? Notices Amer. Math. Soc., 51(6),
630–631.
Tarkhanov, N.N. 1987. On Alexander duality for elliptic complexes. Math. USSR
Sbornik, 58(1).
Trèves, François. 1967. Topological vector spaces, distributions and kernels. New
York: Academic Press.
Vallette, Bruno. 2014. Algebra + homotopy = operad. Pages 229–290 of: Sym-
plectic, Poisson, and noncommutative geometry. Math. Sci. Res. Inst. Publ., vol.
62. Cambridge Univ. Press, New York.
Wallbridge, James. Homotopy theory in a quasi-abelian category. Available at
Weibel, Charles A. 1994. An introduction to homological algebra. Cambridge
Studies in Advanced Mathematics, vol. 38. Cambridge: Cambridge University
Press.
Weiss, Michael. 1999. Embeddings from the point of view of immersion theory. I.
Geom. Topol., 3, 67–101 (electronic).
Wells, Jr., Raymond O. 2008. Differential analysis on complex manifolds. Third
edn. Graduate Texts in Mathematics, vol. 65. New York: Springer. With a new
appendix by Oscar Garcia-Prada.
Williams, Brian. The Virasoro vertex algebra and factorization algebras on
Riemann surfaces. Available at http://www.math.northwestern.edu/
˜bwill/.
Witten, Edward. 1988. Topological quantum field theory. Comm. Math. Phys.,
117(2), 014, 21 pp. (electronic).
Witten, Edward. 1992. Mirror manifolds and topological field theory. Pages 120–
158 of: Essays on mirror manifolds. Int. Press, Hong Kong.
Index
BD algebra, 86, 237 g G , see also prefactorization algebra,

Disj
BD operad, 237 equivariant
C ∗ (g), 243 Čech complex, 173, 246
C∗ (g), 242
Discsn , 111
En algebra, 12, 111, 180, 206
as locally constant factorization algebra, action functional, 19, 34
180 as differential on observables, 80
En -monoidal category, 210 algebra of functions on sections of a vector
P0 algebra, 86, 237 bundle, 287
P0 operad, 237 associative algebra
R-matrix, 211 in quantum mechanics, 5
BVS, 255 Atiyah-Bott lemma, 78, 307
CVS, 255
Batalin-Vilkovisky formalism, 19, 86, 237
DVS, 255
bornological vector space, 270
self-enrichment of, 277
bornology, 270
∆, 227
braided monoidal category, 206
E , 56, 258 BV formalism, see also Batalin-Vilkovisky
Ec , 56, 258 formalism
LCTVS, 255
S Disj, 41 canonical quantization, see also
βγ system, 140 quantization, canonical
factorization homology of, 205 central extension, see also local Lie algebra
E , 57, 260 Chern-Simons theory, see also field theory,
E c , 57, 260 Abelian Chern-Simons
D, 259 chiral algebra, 9
chiral conformal field theory, see also field
C ∞ -module, 52, 256
theory, chiral conformal
distributional sections of a vector bundle,
chiral homology, 9
260
cluster decomposition principle, 115
distributions, 259
cobordism, 10, 11, 179
from a topological vector space, 268 Cobordism Hypothesis, 179
holomorphic sections of a holomorphic colimit
vector bundle, 260 homotopy, see also homotopy colimit
smooth sections of a vector bundle, 258 compactification, 92
Disj, 39 cone, 298
DisjG , see also prefactorization algebra, completed, 306
equivariant convenient vector space, 56, 273
Mfld, 52, 256 sections of vector bundles as, 275
PDiscs, 123 tensor product of, 275
315
316 INDEX
cooperad, 236 Kac-Moody, 64, 204

correlation function, 12, 106, 115 limitations from point of view of quantum
correlator, 31 field theory, 176
correspondence principle, 5 local-to-global axiom, 169
cosheaf, 245 locally constant, 11, 180
gluing axiom, 169 multiplicative, 171
homotopy, 172, 246 multiplicative homotopy, 173
cosimplicial cobar construction, 302 on a site, 176
coupling, 217 on manifolds with geometric structure G,
cover 176
Weiss, 170 pushforward, 191
relationship with En algebras, 180
defect, 206, 216
weak equivalence of, 174
point, 43
with respect to a factorizing basis, 192
dense inclusion, 55
factorization envelope, 62, 154, 187
derived critical locus, 74
twisted, 64, 85
descendant, 127
factorization homology, 179, 182
descent along a torsor, 198
βγ system, 205
differentiable cochain complex, 54
of Kac-Moody factorization algebra, 204
complete filtered, 292
of quantum observables of free scalar field
filtered, 292
theory, 202
differentiable pro-cochain complex, 294
completed direct sum, 295 of universal enveloping algebra, 201
differentiable vector space, 53, 257 relationship with Hochschild homology,
distributions, 259 182
map of, 258 factorizing basis, 191
self-enrichment, 277 field theory
smooth sections of a vector bundle, 259 Abelian Chern-Simons, 94, 105, 206
stalk of, 261 Abelian Yang-Mills, 81
differential complex, 251 Atiyah-Segal, 99
divergence complex, 73 chiral conformal, 121, 127
divergence operator, 21, 24, 73 free, 80
free βγ system, see also βγ system
elliptic complex, 251 free scalar, 20, 22, 81
elliptic operator, 251 functorial, 99
enveloping symmetric monoidal category, interacting, 33
240 quantum, 20
equivariant prefactorization algebra, see also topological, 96
prefactorization algebra, equivariant translation invariance of free scalar, 109
expectation value, 19, 107, 115 vacuum for free scalar field theory, 115
expected value, 19 free scalar field theory, see also field theory,
free scalar
factorization algebra, 169, 171
functorial field theory, 10, 179
Beilinson-Drinfeld, 9
category of, 175
computations, 201 Gaussian integral, 20
differentiable, 174 Gaussian measure, 20
equivariant, 198 Green’s function, 31, 83, 102
extension from a basis, 191 free scalar field, 115
from cosheaf, 184
from local Lie algebra, 187 Hamiltonian, 88, 108, 111
homotopy, 173 Hamiltonian formalism, 108
in a multicategory, 173 Hochschild cohomology for En algebras, 224
in quantum field theory, 175 Hochschild homology
INDEX 317
relationship with factorization homology, classical, 29, 76

182 classical as ~ → 0 limit of quantum, 29
Hochschild-Kostant-Rosenberg (HKR) quantum, 25
theorem, 224 relationship between quantum and
holomorphically translation invariant classical, 29, 101
prefactorization algebras, 122 shifted Poisson structure on classical, 83
homotopy colimit, 297 structure maps as Moyal product on
quantum, 104
interacting field theory, see also field theory, operad, 234
interacting BD, 237
interaction term, 34 P0 , 237
algebra over, 236
Kac-Moody algebra, see also Lie algebra,
colored, 238
Kac-Moody
endomorphism, 236
Koszul duality, 206
operator product expansion, 121
Lie algebra, 240 ordinal category, 227
acting via derivations on a prefactorization
parametrix, 252
algebra, 69
partition of unity, 247
Chevalley-Eilenberg complex, 242
path integral, 19
cohomology, 242
precosheaf, 245
coinvariants, 241
prefactorization algebra, 37, 40, 41
differential graded (dg), 243
bimodule as a defect, 43
enveloping algebra as prefactorization
category of, 41
algebra, 49
classical observables as, 76
Heisenberg, 75, 85
derivation of, 69, 109
homology, 241
differentiable, 55
invariants, 241
encoding standard quantum mechanics, 45
Kac-Moody, 156
enveloping, see also factorization
module over, 240
envelope
representation, 240
equipped with action of a Lie algebra, 69
universal enveloping algebra of, 49, 241
equivariant, 66
line operator, 206, 216
examples from topology, 181
linking number, 100, 105
holomorphically translation-invariant, 123
local cochain complex, 184
holomorphically translation-invariant with
local Lie algebra, 62
a compatible S 1 action, 129
central extension, 64
locally constant, 42
local-to-global axiom, see also factorization
morphism of, 41
algebra, local-to-global axiom
motivating example of, 2
locally constant factorization algebra, see
multicategory of, 42
also factorization algebra, locally
on a site, 176
constant
on manifolds with geometric structure G,
measurement, 3 176
Moyal product, 104 physical interpretation of, 7
multicategory, 238 pushforward, 92, 96
map of, 239 smoothly equivariant, 69
symmetric monoidal envelope, 240 smoothly translation-invariant, 110
structure on observables, 26
observable, 3 translation-invariant, 108
as cocycle in cochain complex, 8 unital, 38, 40
support of, 78 with respect to a factorizing basis, 192
Wilson loop, 100 presheaf, 244
observables propagator, 31, 102
318 INDEX
pushforward, see also prefactorization vacuum, 115

algebra, pushforward, see also free scalar field theory, 115
factorization algebra, pushforward, 191 massive, 115
vector space
quantization bornological, see also bornological vector
as enveloping algebra of Heisenberg Lie space
algebra, 75 convenient, see also convenient vector
Batalin-Vilkovisky, see also space
Batalin-Vilkovisky formalism locally convex topological, 267
canonical, 92 vertex algebra, 119, 119
quantum field theory, 20 βγ, 142
comparison of formalizations, 9 arising from factorization algebra, 129
Osterwalder-Schrader axioms, 12 arising from Lagrangian field theory, 122
Segal’s axioms, 10 field in, 119
Wightman axioms, 12 Kac-Moody, 156
quantum group, 206 reconstruction theorem, 132
quantum mechanics, 87 relationship with operator product
as a prefactorization algebra, 45 expansion, 121
as motivating prefactorization algebras, 2 vertex operator, 120
measurement in, 3
Weiss cover
Schwartz functions, 117 examples, 171
sheaf, 245 Weiss topology, 170
simplicial abelian group, 230 Weyl algebra, 87
simplicial bar construction, 303 rigidity of, 88
simplicial set, 227 Wick’s formula, 20, 32
site of manifolds, 52, 256 Wick’s lemma, 20, 106
Wilson loop, see also observable, Wilson
smoothly equivariant prefactorization
loop
algebra, see also prefactorization
algebra, smoothly equivariant Yang-Mills field theory, see also field theory,
stalk, 261 Abelian Yang-Mills
state, 114
translation-invariant, 114
vacuum, 115
symmetric monoidal envelope, 240
Tannakian reconstruction, 206

tensor product
bornological, 273
relationship between completed projective
and convenient, 286
relationship between projective and
bornological, 286
topological chiral homology, 182
topological field theory, 11, see also field
theory, topological
translation invariance
of free scalar field theory, 109
translation invariance vs. translation
equivariance, 108
universal enveloping algebra, see also Lie

algebra, universal enveloping algebra of

Costello

Uploaded by

Copyright:

Available Formats

Costello

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Costello

Uploaded by

Copyright:

Available Formats

Factorization Algebras in Quantum Field

Kevin Costello and Owen Gwilliam

Part 1. Prefactorization algebras 17

Chapter 2. From Gaussian measures to factorization algebras 19

Chapter 3. Prefactorization algebras and basic examples 37

Part 2. First examples of field theories and their observables 71

Chapter 4. Free field theories 73

7. Correlation functions 106

Chapter 5. Holomorphic field theories and vertex algebras 119

Part 3. Factorization algebras 167

Chapter 6. Factorization algebras: definitions and constructions 169

Chapter 7. Formal aspects of factorization algebras 191

Chapter 8. Factorization algebras: examples 201

Appendix A. Background 225

Appendix B. Functional analysis 255

Appendix C. Homological algebra in differentiable vector spaces 289

Thus, Acl is recovered in the ~ → 0 limit, i.e., the classical limit.

This quantization theorem is proved using the physicists’ techniques of perturba-

1. The motivating example of quantum mechanics

1.3. Measurements in quantum mechanics. The notion of measurement is

(The precise meaning of these terms is detailed in Section

2. A preliminary definition of prefactorization algebras

Remark: A prefactorization algebra resembles a precosheaf, except that we tensor

determined by the observables on small open sets. The notion of a factorization

3. Prefactorization algebras in quantum field theory

3.1. Interpretation of the prefactorization algebra axioms. Suppose that

3.2. The cochain complex of observables. In the approach to quantum field

Although the cohomology groups H n (Obs(U)) where n ≥ 1 do not have a clear

4. Comparisons with other formalizations of quantum field theory

n from a Segal field theory. Given an open subset U ⊂ M whose boundary is a

4.4. Correlation functions. In some classic approaches to the axiomatics of

5. Overview of this volume

From Gaussian measures to factorization algebras

This chapter serves as a kind of second introduction, demonstrating how the

where x1 , . . . , xn are points in M. Note that there is an observable

1. Gaussian integrals in finite dimensions

Then, such a textbook would go on to define similar infinite-dimensional Gauss-

(This formula is an exercise in applying Cartan’s magic formula: LV = [d, ιV ].

for all polynomial vector fields V, because

and then we apply Stokes’ lemma. By changing basis of Rn to diagonalize A, one

1.0.1 Lemma. The quotient map

2. Divergence in infinite dimensions

g c (C ∞ (U)) can be written as an finite sum of monomials of

3. The prefactorization structure on observables

P(C ∞ (V)) ⊗ P(C ∞ (V)) → P(C ∞ (V))

does not descend to a map

H 0 (Obs(V)) ⊗ H 0 (Obs(V)) → H 0 (Obs(V)).

If U1 , U2 are disjoint open subsets of the open V ⊂ M, then we have a map

P(U1 ) ⊗ P(U2 ) → P(V)

Proof. Let U1 , U2 be disjoint open subsets of M, both contained in an open V.

4. From quantum to classical

of operators over R[~] connecting H 0 (Obsq (M)) to H 0 (Obscl (M)). It is straight-

h−i : H 0 (Obsq (U1 )) ⊗ · · · ⊗ H 0 (Obsq (Un )) → H 0 (Obsq (M))  R

We should compare this definition with what happens in finite dimensions. If

Note that this result is exactly the standard result in physics.

6. Further results on free field theories

Consider an interacting scalar field theory on a Riemannian manifold M. For

If φ0 ∈ Dc (U) is a compactly-supported distribution, then for any local func-

Now, M f φ no longer make sense, because we are trying to multiply distributions.

Vectc (C ∞ (U)) → P(C ∞ (U))

h−i : H 0 (Obsq (U1 )) ⊗ · · · ⊗ H 0 (Obsq (Un )) → H 0 (Obsq (M)) R

Ec! (U) Cc∞ (U)[1] → Cc∞ (U) ,