Proceedings of the International Congress of Mathematicians

Hyderabad, India, 2010

A Geometric Construction of the

Witten Genus, I

Kevin Costello∗

Abstract
I describe how the Witten genus of a complex manifold X can be seen from
a rigorous analysis of a certain two-dimensional quantum field theory of maps
from a surface to X.

Mathematics Subject Classification (2010). 58J26, 81T40

Keywords. Elliptic genera, quantum field theory

1. Introduction
This paper will describe an application of my work on the foundations of quan-
tum field theory (much of it joint with Owen Gwilliam) to topology. I will show
how consideration of certain two-dimensional quantum field theories – called
holomorphic Chern-Simons theories – leads to a geometric construction of the
Witten genus.
Usually the Witten genus is defined by its q-expansion. In the construction
presented here, however, we find directly a function on the moduli space of
(suitable decorated) elliptic curves. It is only after careful calculation that we
can compute the q-expansion of this function and identify it with the Witten
class.
Hopefully, this construction will give some hints about the mysterious geo-
metric origins of elliptic cohomology.
I am very grateful to Owen Gwilliam, Mike Hopkins, Josh Shadlen, Stefan
Stolz and Peter Teichner for many helpful conversations about the material in
this paper.

∗ Department of Mathematics, Northwestern University, Evanston, Illinois, United States

of America. E-mail: costello@math.northwestern.edu.

A Geometric Construction of the Witten Genus, I 943

2. Hochschild Homology and the Todd Class

Before turning to elliptic cohomology and the Witten class, I will describe the
analog of my construction for the Todd class.
The most familiar way in which the Todd class occurs is, of course, in the
Grothendieck-Riemann-Roch theorem. Let me recall the statement. Let X be a
smooth projective variety, and let E be an algebraic vector bundle on X. Then,
the Grothendieck-Riemann-Roch theorem states that
X Z
i i
(−1) dim H (X, E) = Td(T X) ch(E).
X

Another (and closely related) way in which the Todd class appears is in
the study of deformation quantization. There is a rich literature on algebraic
and non-commutative analogs of the index theorem: see [Fed96, BNT02]. Much
of this literature concerns index-type statements on quantizations of general
Poisson manifolds. For the purposes of this paper, we are only interested in
the relatively simple case when we are quantizing the cotangent bundle of a
complex manifold X.
Let Diff X denote the algebra of differential operators on X. Let Diff ~X denote
the sheaf of algebras on X over the ring C[~] obtained by forming the Rees
algebra of the filtered algebra Diff X . Explicitly,

Diff ~X ⊂ Diff X ⊗C[~]

is the subalgebra consisting of those finite sums

X
~i D i

where Di is a differential operator of order at most i. Thus, Diff ~X is a C[~]

algebra whose specialization to ~ = 0 is the commutative algebra OT ∗ X of
functions on the cotangent bundle of X. When specialized to a non-zero value
of ~, Diff ~X is just Diff X .

2.1. The theorem we are interested in states that the Todd class of X appears
when one computes the Hochschild homology of the algebra Diff ~X . The index
theorem concerns, ultimately, traces of differential operators. Since HH(Diff ~X )
is the universal recipient of a trace on the algebra Diff ~X , it is perhaps not so
surprising that the Todd class should appear in this context.

2.2. Recall that the Hochschild-Kostant-Rosenberg theorem gives a quasi-

isomorphism
IHKR : HH(OX ) ∼ = Ω−∗ (X).
Here HH(OX ) refers to the sheaf of Hochschild chains of OX , and Ω−∗ (X) refers
to the algebra of forms of X, with reversed grading. Applied to the cotangent
944 Kevin Costello

bundle of X, the Hochschild-Kostant-Rosenberg theorem gives an isomorphism

IHKR : HH(OT ∗ X ) ∼
= Ω−∗ (T ∗ X).

The algebra Diff ~X is a deformation quantization of OT ∗ X . We will see that the

Todd genus appears when we study how HH(OT ∗ X ) changes when we replace
OT ∗ X by Diff ~X .

2.3. Before we state the theorem, we need some notation. Let π ∈

Γ(T ∗ X, ∧2 T (T ∗ X)) denote the canonical Poisson tensor on T ∗ X. Let

Lπ : Ωi (T ∗ X) → Ωi−1 (T ∗ X)

denote the operator of Lie derivative with respect to π. Thus, if iπ is contraction

by π,
Lπ = [iπ , ddR ].
Note that Lπ makes Ω−∗ (T ∗ X) into a cochain complex; the cohomology of this
complex is called Poisson homology.
Let
Td(X) ∈ H 0 (X, Ω−∗ (X)) = ⊕H i (X, Ωi )
be the Todd class of X. Note that the reversal of grading in the de Rham
complex means that Td(X) is an element of cohomological degree 0.
The first statement of the theorem is as follows.

Theorem 2.3.1 (Fedosov [Fed96], Bressler-Nest-Tsyan [BNT02]). There is a

natural quasi-isomorphism of cochain complexes


HH(Diff ~X ) ' Ω−∗ (T ∗ X)[~], ~Lπ

sending 1 ∈ HH(Diff ~X ) to

Td(X) ∈ RΓ(X, Ω−∗ (X)).

2.4. This is a rather weak formulation of the theorem, because both sides in
the quasi-isomorphism are simply cochain complexes. There is a refined version
which identifies a certain algebraic structure present on both sides. It will take
a certain amount of preparation to state this refined version.
The operator Lπ is an order two differential operator with respect to the
natural product on Ω−∗ (T ∗ X). We will let {−, −}π denote the Poisson bracket
on Ω−∗ (T ∗ X) of cohomological degree 1 defined by the standard formula

{a, b}π = Lπ (ab) − (Lπ a)b − (−1)|a| aLπ b.

The bracket {−, −}π is of cohomological degree 1, and satisfies the standard
Leibniz rule. Further, Lπ is a derivation for the bracket {−, −}π .
A Geometric Construction of the Witten Genus, I 945

Theorem 2.4.2. There is a quasi-isomorphism of cochain complexes

HH(Diff ~X ) ' Ω−∗ (T ∗ X)[~], ~Lπ + ~{log Td(X), −} .

The isomorphism in this theorem is related to that of the previous formu-

lation by conjugating by Td(X).

2.5. The isomorphism appearing in this second formulation is the one that is
compatible with an additional algebraic structure. The structure is that of an
algebra over a certain operad, introduced by Beilinson and Drinfeld [BD04].
Definition 2.5.3. A Beilinson-Drinfeld algebra A is a flat graded C[~] module
endowed with the following structures.
1. A commutative unital product.
2. A Poisson bracket {−, −} of cohomological degree 1.
3. A differential D : A → A of cohomological degree 1, satisfying D2 = 0
and D1 = 0, such that

D(ab) = (Da)b + (−1)|a| a(Db) + ~{a, b}.

The complex HH(Diff ~X ) is endowed with the structure of Beilinson-

Drinfeld (or BD) algebra in a natural way.
The complex Ω−∗ (T ∗ X)[~]) also has the structure of BD algebra, with prod-
uct the ordinary wedge product of forms. The bracket is {−, −}π , and the
differential ~Lπ + ~{log Td(X), −}.
Proposition 2.5.4. The quasi-isomorphisms


HH(Diff ~X ) ' Ω−∗ (T ∗ X)[~], ~Lπ + ~{log Td(X), −} .

is a quasi-isomorphism of BD algebras.
In this lecture I will state a generalization of this result, in which the Witten
class appears in place of the Todd class.

3. Factorization Algebras
Hochschild homology, K-theory and the Todd genus are all intimately con-
cerned with the concept of associative algebra. In order to understand the
Witten genus, one needs to consider a richer algebraic structure called a factor-
ization algebra (or more precisely, a translation-invariant factorization algebra
on the complex plane C).
Factorization algebras can be defined on any smooth manifold: they can
be viewed as a “multiplicative” analog of a cosheaf. In the algebro-geometric
946 Kevin Costello

context, factorization algebras were first considered by Beilinson and Drinfeld

[BD04].
In this section, I will give the formal definition of a factorization algebra,
and state a theorem (from [CG10]) which allows one to construct factoriza-
tion algebras using the machinery of perturbative renormalization developed in
[Cos10b].
The approach to constructing factorization algebras developed in [CG10] is
a quantum field theoretic analog of the deformation quantization approach to
quantum mechanics. Thus, a classical field theory yields a commutative factor-
ization algebra (I will define what this means shortly). Quantizing a classical
field theory amounts to replacing this commutative factorization algebra by a
plain factorization algebra. Just like the Todd genus of a complex manifold
X appears when one considers the deformation quantization of the cotangent
bundle T ∗ X, we will see that the Witten genus arises when we consider the
quantization of a commutative factorization algebra associated to a classical
field theory whose fields are maps from a Riemann surface to T ∗ X.

3.1. The definition of a factorization algebra is rather straightforward to give.

Definition 3.1.5. Let M be a manifold. A factorization algebra F on M con-
sists of the following data.
1. For every open set U ⊂ M , a cochain complex of topological vector spaces,
F(U ).
2. If U1 , . . . , Uk are disjoint open sets in M , all contained in a larger open
set V , a continuous linear map

F(U1 ) ⊗ · · · ⊗ F(Uk ) → F(V )

(where we use the completed projective tensor product).

3. These maps must satisfy an evident compability condition, which says that
different ways of composing these maps yield the same answer.
4. Finally, we need a locality axiom, saying that every element of F(V ) can
be built from elements of F(U ) for arbitrarily small open subsets U of V .
Let V ⊂ M , and let V = {Vi | i ∈ I} be an open cover of V . Then, we
require that

F(V ) = hocolimU1 ,...,Un F (U1 ) ⊗ · · · ⊗ F (Un )

where U1 , . . . , Un are disjoint subsets of V , each of which is contained in

some Vj .
If U1 , U2 are disjoint subsets, then a particular case of this axiom says
that
F(U1 q U2 ) = F(U1 ) ⊗ F(U2 ).
A Geometric Construction of the Witten Genus, I 947

This definition is reminiscent of that of an En algebra. In fact, Jacob Lurie

has shown the following [Lur09].

Proposition 3.1.6. There is an equivalence of (∞, 1)-categories between the

category of En algebras, and the category of factorization algebras F on Rn with
the additional property that if B ⊂ B 0 are balls, the map

F(B) → F(B 0 )

is a quasi-isomorphism.

In another direction, what we call a factorization algebra is the C ∞ analog of

a definition introduced by Beilinson and Drinfeld [BD04]. Beilinson and Drinfeld
introduced the notion of chiral algebra in order to give a geometric formulation
of the axioms of a vertex algebra. In particular, every vertex algebra yields a
chiral algebra on the complex line C, and one can turn this into a factorization
algebra on C (considered as a Riemann surface).

3.2. As our first example of a factorization algebra, let us see how a differential
graded associative algebra A gives rise to a translation-invariant factorization
algebra FA on R.
We will define the value of FA on the open intervals of R; the value of FA
on more complicated open subsets is formally determined by this data.
Let −∞ ≤ a < b ≤ ∞, and let (a, b) be the corresponding (possibly infinite)
open interval in R. We set
FA ((a, b)) = A.

If (a, b) ⊂ (c, d), then the map

FA ((a, b)) → FA ((c, d))

is the identity map on A.

If −∞ ≤ a1 < b1 < a2 < b2 < · · · < an < bn ≤ ∞, then the intervals (ai , bi )
are disjoint. Part of the data of a factorization algebra is thus a map

FA ((a1 , b1 )) ⊗ · · · ⊗ FA ((an , bn )) → FA ((a1 , bn )).

Once we identify each FA ((ai , bi )) with A, this map is the n-fold product map

A⊗n → A
α1 ⊗ · · · ⊗ αn 7→ α1 · α2 · · · · · αn .

The value of FA on any other open subset of R is determined from this data
by the axioms of a factorization algebra.
948 Kevin Costello

4. Descent and Factorization Homology

In this paper, we are only interested in translation-invariant factorization alge-
bras on C. In this section, we will see that associated to such a factorization
algebra F, and to an elliptic curve E, equipped with a never-vanishing volume
element ω, one can define the factorization homology
F H(E, F).
Factorization homology is the analog, in the world of factorization algebras, of
Hochschild homology.
As motiviation, I will first explain how the Hochschild homology groups of
an associative algebra A can be viewed as the factorization homology of the
translation-invariant factorization algebra FA on R associated to A.

4.1. Factorization algebras satisfy a gluing axiom. Suppose that our manifold
M is written as a union M = U ∪ V of two open subsets. If F is a factorization
algebra on an open subset U ⊂ M , and if G is a factorization algebra on V , and
if
φ : F |U ∩V → G |U ∩V
is an isomorphism of factorization algebras on U ∩ V , then we can construct a
factorization algebra H on M , whose restriction to U is F and whose restriction
to V is G.
Similarly, factorization algebras satisfy descent. Suppose that a discrete
group G acts properly discontinuously on a manifold M , and suppose that F e is
e
a G-equivariant factorization algebra on M . Then, F descends to a factorization
algebra F on the quotient M/G.

4.2. Since we will be using the descent property extensively, it is worth ex-
plaining how one constructs the descended factorization algebra F. Let us
choose an open cover of M/G by connected and simply connected open subsets
ei ⊂ M which map homeomorphically onto
{Ui }. Let us choose open subsets U
Ui .
Then, if V ⊂ M/G is an open subset which lies in some Ui , we set
e Ve )
F(V ) = F(

where Ve ⊂ U
fi is the lift of V .
If V ⊂ Ui ∩ Uj , then the fact that the factorization algebra F e on M is
G-equivariant implies that F(V ) is independent of the lift we choose.
If V ⊂ M/G is an arbitrary open subset of M/G, then we set
e Ve1 ) ⊗ · · · F(
F(V ) = hocolimV1 ,...,Vn F( e Ven )

where the homotopy colimit is over open subsets V1 , . . . , Vn ⊂ M/G each of

which lies inside one of the subsets Uj .
A Geometric Construction of the Witten Genus, I 949

4.3. This descent property implies that any translation-invariant factorization

1
algebra F on R descends to a factorization algebra FS on S 1 = R/Z. We will
let 1
F H(S 1 , F) = FS (S 1 )
1
denote the complex of global sections of the factorization algebra FS on S 1 .
We will refer to the complex F H(S 1 , F) as the factorization homology complex
of S 1 with coefficients in F.
Lemma 4.3.7. Let FA denote the factorization algebra on R associated to
a differential graded associated algebra A. Then, there is a natural quasi-
isomorphism
F H(S 1 , FA ) ' HH(A)
between the factorization homology complex of S 1 with coefficients in FA , and
the Hochschild complex of A.
Proof. If one analyzes the descent prescription described above, one sees that

F H(S 1 , FA ) = hocolimI1 ,...,In A⊗n

where the homotopy colimit is over disjoint unordered intervals in S 1 . The maps
in this homotopy colimit just arise form multiplication in A. One sees that a
complex which looks like the ordinary cyclic bar complex emerges from this
procedure. In [Lur09] it is proven that the result of this homotopy colimit is
indeed homotopy equivalent to the cyclic bar complex.
4.4. If λ ∈ R>0 , let Sλ1 be the quotient of R by the lattice λZ. If F is a
translation-invariant factorization algebra on R, then we can descend F to a
factorization algebra on Sλ1 , and thus define factorization homology F H(Sλ1 , F).
When λ = 1, this coincides with the definition given above. In principle, there
is no reason that F H(Sλ1 , F) should be independent of λ.
If we use the factorization algebra FA arising from an associative algebra
A, then all the factorization homology complexes F H(Sλ1 , FA ) are canonically
isomorphic. This is because the factorization algebra FA on R is not only trans-
lation invariant but also dilation invariant.
4.5. As I mentioned earlier, the factorization algebras relevant to the Witten
genus are translation-invariant factorization algebras on C. Let F be such a
factorization algebra. Let E be an elliptic curve equipped with a volume element
ω. We will write E as a quotient C/Λ of C by a lattice Λ, in such a way that
form ω on E pulls back to the volume form dz on C.
Since F is translation-invariant, it is in particular invariant under Λ. Thus,
F descends to a factorization algebra FE on E. We define the factorization
homology complex of E with coefficients in F by

F H(E, F) = FE (E).

Thus, F H(E, F) is the global sections of FE on E.

950 Kevin Costello

Thus, there is an analog of the Hochschild homology groups for every elliptic
curve E with volume element ω.

5. Main Theorem
The main theorem states that the Witten class of a complex manifold X arises
when one considers the factorization homology of a certain sheaf (on X) of
translation-invariant factorization algebras on C. Before I state this theorem, I
need to recall the definition of the Witten class.

5.1. Let E be an elliptic curve, and let ω be a translation-invariant volume

element on E. The Witten class

Wit(X, E, ω) ∈ RΓ(X, Ω−∗ (X)) = ⊕H i (X, Ωi (X))

is a cohomology class, defined as follows.

Let
X
E2k (E, ω) = λ−2k
λ∈Λ

be the Eisenstein series of the marked elliptic curve (E, ω). Here, as before,
we are writing E as the quotient of C by a lattice Λ, in such a way that ω
corresponds to dz.
The Witten class of X is defined by
 
X (2k − 1)! 
Wit(X, E, ω) = exp E2k (E, ω)ch2k (T X) .
 (2πi)2k 
k≥2

If τ is in the upper half-plane, let (Eτ , ωτ ) denote the elliptic curve associ-
ated to the lattice generated by (1, τ ), with volume form ωτ corresponding to
dz. Then, the Witten class has the property that

lim Wit(X, Eτ , ωτ ) = e−c1 (TX )/2 Td(T X).

τ →i∞

This follows from the identities

lim E2k (Eτ , ωτ ) = 2ζ(2k)

τ →i∞
X  
x2k x x
2ζ(2k) = log −
2k(2πi)2k 1 − e−x 2
k≥1

where ζ is the Riemann zeta function.

A Geometric Construction of the Witten Genus, I 951

5.2. Now we can state the theorem.

Theorem 5.2.8. Let X be a complex manifold, equipped with a trivializa-

~
tion of the second Chern character ch2 (T X). Then, there is a sheaf DX,ch
of translation-invariant factorization algebras on C, over the algebra C[~], such
that, for every elliptic curve E with volume element ω, there is a natural iso-
morphism of BD algebras


~
F H(E, DX,ch ) ' Ω−∗ (T ∗ X)[~], ~Lπ + ~{log Wit(X, E, ω), −} .

Alternatively, there is an isomorphism of cochain complexes

~
F H(E, DX,ch ) ' Ω−∗ (T ∗ X)[~], ~Lπ .

sending
1 → Wit(X, E, ω).

The factorization algebra appearing in this theorem is an analytic avatar

of the chiral differential operators constructed by Gorbounov, Mailkov and
Schechtman [GMS00]. Note that in their work, the q-expansion of the Witten
genus appears as the character of the algebra of chiral differential operators.
The way the Witten genus appears in this paper is somewhat different, and
has the advantage that we see the Witten genus directly as a function on the
moduli space of elliptic curves, and not just as a q-expansion.

6. Factorization Algebras from Quantum Field

Theory
A factorization algebra is the algebraic structure satisfied by the observables of
a quantum field theory. In [CG10] we prove a theorem allowing one to construct
factorization algebras using the techniques of perturbative renormalization. The
~
factorization algebra DX,ch encoding the Witten genus will be constructed by
quantizing a certain two-dimensional quantum field theory, called holomorphic
Chern-Simons theory.
Before I discuss this particular quantum field theory, let me explain, heuris-
tically, why one would expect the observables of a quantum field theory to form
a factorization algebra. Suppose we have a quantum field theory (whatever that
is) on a manifold M . Then, for every open subset U ⊂ M , we would expect the
set of observables on U – that is, the set of measurements that can be made by
an observer in the open subset U – to form a vector space, which we call F(U ).
If U ⊂ V , then an observable on U will, in particular, be an observable on
V , so that we get a map F(U ) → F(V ).
If U1 and U2 are disjoint, we would expect that all obervables on U1 qU2 are
obtained by taking the product of an observable on U1 with one on U2 . Thus,
952 Kevin Costello

we would expect that

F(U1 q U2 ) = F(U1 ) ⊗ F(U2 ).

Together, these maps give F the structure of a factorization algebra.

6.1. The idea that the observables of a quantum field theory form a factor-
ization algebra is compatible with two familiar examples.
Quantum mechanics is a quantum field theory on the real line R. The obser-
ables for quantum mechanics form an associative algebra. Associative algebras
are a particular class of factorization algebras on R. In [CG10], we show that the
factorization algebra associated to the free field theory on R is, in fact, an E1
algebra; specifically, it is the familiar Weyl algebra of observables of quantum
mechanics.
A second well-understood example is conformal field theory. The observables
of conformal field theory on C form a vertex algebra; and, as we have seen,
vertex algebras are a special class of factorization algebra C.

6.2. Let me now briefly state the results of [CG10] and [Cos10b], allowing one
to construct factorization algebras.
In [Cos10b], I gave a definition of a quantum field theory on a manifold M ,
using a synthesis between Wilson’s concept of a low-energy effective field theory
and the Batalin-Vilkovisky formalism for quantizing gauge theories. Further, I
developed techniques (based on the machinery of perturbative renormalization)
allowing one to construct such quantum field theories from Lagrangians.
Many quantum field theories of physical and mathematical interest, such as
Chern-Simons theory and Yang-Mills theory, can be put in this framework.
The most succinct way to state the main construction of [CG10] is as follows.

Theorem 6.2.9. Any quantum field theory in the sense of [Cos10b], on a

manifold M , yields a factorization algebra on M .

We have seen that factorization algebras satisfy a descent property: if a

discrete group G acts properly discontinuously on a manifold M , then a G-
e on M descends to the quotient M/G. Quan-
equivariant factorization algebra F
tum field theories in the sense of [Cos10b] satisfy a similar descent property,
and the construction of a factorization algebra from a quantum field theory is
compatible with descent.

7. Deformation Quantization in Quantum Field

Theory
In this section, I will explain a little about how one associates a factorization
algebra to a classical or quantum field theory. We will see that the procedure
A Geometric Construction of the Witten Genus, I 953

of quantizing a classical field theory can be interpreted in algebraic terms as a

kind of deformation quantization in the world of factorization algebras.
The observables of a classical mechanical system form a commutative al-
gebra, whereas the observables of a quantum mechanical system are only an
associative algebra. We should view this commutativity as being an extra struc-
ture present on the observables of a classical system.
There is a similar story in field theory: the observables of a classical field
theory on a manifold M have an extra structure, that of a commutative factor-
ization algebra.
Factorization algebras form a symmetric monoidal category: if F, G are fac-
torization algebras, then we can define a factorization algebra F ⊗ G by the
formula
(F ⊗ G) (U ) = F(U ) ⊗ G(U )
for an open subset U ⊂ M .

Definition 7.0.10. A commutative factorization algebra is a commutative al-

gebra in the category of factorization algebras.

Thus, a commutative factorization algebra assigns to every U ⊂ X a com-

mutative algebra F(U ), and to inclusion maps U ,→ V , a map of commutative
algebras.

7.1. The main object of interest in a classical field theory is the space of
solutions to the Euler-Lagrange equation. If U ⊂ M is an open set, let EL(U )
be this space. Sending U 7→ EL(U ) defines a sheaf of formal spaces on M .
This sheaf of solutions to the Euler-Lagrange equations can be encoded in
the structure of a commutative factorization algebra. If U ⊂ M is an open
subset, we will let O(EL(U )) denote the space of functions on EL(U ).
Sending U 7→ O(EL(U )) defines a commutative factorization algebra: if
U1 , . . . , Un are disjoint open subsets of Un+1 , there is a restriction map

EL(Un+1 ) → EL(U1 ) × · · · × EL(Un ).

Replacing the map of spaces by the corresponding map of algebras of functions

yields the desired structure of commutative factorization algebra.

7.2. In the familiar deformation quantization story, the algebra of observables

of a classical mechanical system is a commutative algebra endowed with an extra
structure, namely a Poisson bracket. This extra structure is what tells us that
the commutative algebra “wants” to deform into an associative algebra.
There is a similar picture in the world of factorization algebras: the commu-
tative factorization algebra associated to a classical field theory is endowed with
an extra structure, which makes it “want” to deform into a plain factorization
algebra.
954 Kevin Costello

Ordinary Poisson algebras interpolate between commutative algebras and

associative (or E1 ) algebras. For us, the object describing the observables of a
quantum field theory is not an E1 algebra in a symmetric monoidal category;
instead, it is an E0 algebra. An E0 algebra in vector spaces is simply a vector
space with an element. An E0 algebra in any symmetric monoidal category is an
object of this category with a map from the unit object. An E0 algebra in the
symmetric monoidal category of factorization algebras is simply a factorization
algebra, as every factorization algebra is equipped with a unit.
Thus, the analog of the Poisson operad we are searching for is an operad
that interpolates between the commutative operad and the E0 operad. Such an
operad was constructed by Beilinson and Drinfeld [BD04]; we will call it the
BD operad 1 .

Definition 7.2.11. Let P0 be the graded operad over C generated by a commu-

tative and associatve product, ∗, and a Poisson bracket {−, −} of cohomological
degree +1.
Let BD denote the differential graded operad over the ring C[~] which, as a
graded operad, is simply P0 ⊗ C[~], but which is equipped with differential

d∗ = ~{−, −}.

If we specialize to ~ = 0, we find the BD operad becomes the operad P0 . If

we specialize to ~ = 1, however, the BD operad becomes the E0 operad. Thus,
we find that the operad P0 bears the same relationship to the operad E0 as the
usual Poisson operad bears to the associative operad E1 .

Definition 7.2.12. A Poisson factorization algebra on M is a P0 algebra in

the category of factorization algebras on M . A quantization of a Poisson factor-
ization algebra Fcl is a BD algebra Fq in the category of factorization algebras
on M , together with an isomorphism

Fq ⊗C[~] C ∼
= Fcl

of Poisson factorization algebras.

7.3. Now we can restate the main results of [CG10].

Theorem 7.3.13. Every classical field theory on M gives rise to a Poisson

factorization algebra on M . A quantization of this classical field theory (in the
sense of [Cos10b]) gives rise to a quantization of this Poisson factorization
algebra.

1 Beilinson and Drinfeld called this operad the Batalin-Vilkovisky operad. However, in

the literature, the Batalin-Vilkovisky operad has, unfortunately, come to refer to a different
object.
A Geometric Construction of the Witten Genus, I 955

What I mean by a classical field theory on M is detailed in [Cos10b], but

it is something rather familiar. There is a space of fields, which is taken to
be the space E of sections of some vector bundle E on M , or more generally
some space of maps M → N to some other manifold N . In addition, there is
an action functional S : E → R (or to C), which is taken to be the integral
of some Lagrangian density. When dealing with theories with gauge symmetry,
this basic picture needs to be modified by the introduction of fields which
possess a cohomological degree. This more sophisticated picture is known as
the Batalin-Vilkovisky formalism.
In [Cos10b, CG10] we always work in the Batalin-Vilkovisky formalism.
Thus, our space of classical fields is equipped with a symplectic form of coho-
mological degree −1. The fact that the factorization algebra of observables of a
classical field theory is equipped with a Poisson bracket of degree +1 is simply
a version of the familiar statement that the algebra of functions on a symplectic
manifold has a natural Poisson bracket.

8. Holomorphic Chern-Simons Theory

As we have seen, when we work in the BV formalism, the space of classical
fields is a (typically infinite dimensional) differential graded manifold equipped
with a symplectic form of cohomological degree −1. The action functional is
a secondary object in this approach. The differential on the space of fields
preserves the symplectic form, and thus, at least locally, is given by Poisson
bracket with some Hamiltonian function S, of cohomological degree zero. This
function is the classical action.
In the paper [AKSZ97], Alexandrov, Kontsevich, Schwartz and Zabronovsky
introduced a beautiful and general method for constructing classical field the-
ories in the BV formalism. Many quantum field theories studied in mathe-
matics arise from the AKSZ construction. For example, Chern-Simons theory,
Rozansky-Witten theory and the Poisson σ model all fit very naturally into
this framework.
For us, the relevance of the AKSZ construction is that the classical field
theory related to the Witten genus arises most naturally from the AKSZ con-
struction.
Before I introduce the AKSZ construction, we need some notation.

Definition 8.0.14. A differential graded manifold is a smooth manifold X

equipped with a sheaf OX of differential graded commutative algebras over
C, with the property that OX is locally isomorphic as a graded algebra to
∞
CX [[x1 , . . . , xn ]], where xi are formal variables of cohomological degree di ∈ Z.
∞
In this definition, CX refers to the sheaf of complex-valued smooth functions
on X. One can talk about geometric structures – such as Poisson or symplectic
structures – on a differential graded manifold.
956 Kevin Costello

If X is a smooth manifold, we will let XdR denote the dg manifold whose

underlying smooth manifold is X, and whose sheaf of functions is the complex-
ified de Rham complex Ω∗X of X. If X is a complex manifold, we will let X∂
denote the dg manifold whose underlying smooth manifold is X, and whose
sheaf of functions is the Dolbeault complex Ω0,∗
X .

8.1. Now we can explain the AKSZ construction. Suppose we have a compact
differential graded manifold M , equipped with volume element of cohomological
degree k. Let N be a differential graded manifold with a symplectic form of
cohomological degree l. Then, the infinite-dimensional differential graded man-
ifold Maps(M, N ) acquires a symplectic form of cohomological degree l − k.
If f : M → N is a map, then the tangent space to Maps(M, N ) at f is

Tf Maps(M, N ) = Γ(M, f ∗ T N ).

We define a pairing on Tf Maps(M, N ) by the formula

Z
hα, βi = hα, βiN .
M
R ∞
Since the integration map M : C (M ) → R is of cohomological degree −k,
and the symplectic pairing on T N is of cohomological degree m, the pairing on
Tf Maps(M, N ) is of cohomological degree m − k. The case of interest in the
Batalin-Vilkovisky formalism is when m − k = −1.
There is a variation of this construction which applies when the source
manifold M is non-compact. In this situation, the space Maps(M, N ) has a
natural integrable distribution given by the subspace

Γc (M, f ∗ T N ) ⊂ Tfc Maps(M, N ) ⊂ Tf Maps(M, N )

consisting of compactly supported tangent vector fields. In this situation, in-

stead of having a symplectic pairing on Tf Maps(M, N ), we only have one on the
distribution Tfc Maps(M, N ). The action functional, instead of being a closed
one-form on Maps(M, N ), is a closed one-form on the leaves of the foliation.

8.2. There are two broad classes of AKSZ theories which are commonly con-
sidered. These are the theories of Chern-Simons type, and the theories of holo-
morphic Chern-Simons type.
The two classes of theories are distinguished by the nature of the source dg
manifold M . In theories of Chern-Simons type, the source differential graded
ringed space is XdR , where X is an oriented manifold. The orientation on X
gives rise to a volume element on XdR of cohomological degree dim(X).
The target manifolds for Chern-Simons theories of dimension k are dg sym-
plectic manifolds of dimension k − 1. For example, perturbative Chern-Simons
theory arises when we take the target to be the dg manifold whose underly-
ing manifold is a point, and whose algebra of functions is the algebra C ∗ (g)
A Geometric Construction of the Witten Genus, I 957

of cochains on a semi-simple Lie algebra g. The Killing form endows this dg

manifold with a symplectic form of cohomological degree 2. This theory is per-
turbative, because maps MdR → C ∗ (g) are the same as connections on the
trivial principal G bundle which are infinitesimally close to the trivial connec-
tion.
Non-perturbative Chern-Simons theory arises from a genearlized form of the
AKSZ construction whcih takes the stack BG as the target manifold. Vector
bundles on BG are the same as G-modules; the tangent bundle of BG is the
adjoint module g[1]. The Killing form on g is G-equivariant, and so gives rise
to a symplectic form on BG of cohomological degree 2.
Rozansky-Witten theory also arises from this framework. Let X∂ be a holo-
morphic symplectic manifold. Let us work over the base ring C[q, q −1 ] where q
is a parameter of degree −2. Then the symplectic form q −1 ω on X∂ is of coho-
mological degree 2, and so we can define a 3-dimensional Chern-Simons type
theory. The fields of this theory are maps MdR → X∂ , where M is a 3-manifold,
and everything takes place over the base ring C[q, q −1 ].
Another example is the Poisson σ-model of [Kon03, CF00]. Here, the source
is ΣdR where Σ is a smooth surface. The target is the differential graded man-
ifold T ∗ [1]N , whose underlying smooth manifold is N , and whose algebra of
functions is Γ(N, ∧∗ T N ). The Schouten-Nijenhuis bracket {−, −} endows this
dg manifold with a symplectic form of cohomological degree 1. The differential
on Γ(N, ∧∗ T N ) is given by bracketing with the Poisson tensor π.

8.3. Let us now discuss holomorphic Chern-Simons theory, which is the only
quantum field theory we will be concerned with in this paper. In holomorphic
Chern-Simons theory, the source dg manifold is X∂ , where X is a complex man-
ifold equipped with a never-vanishing holomorphic volume element ω (thus, X
is a Calabi-Yau manifold). This volume form can be thought of as a volume ele-
ment on X∂ of cohomological degree dimC (X). Integration against this volume
element is simply the map

Ω0,dimC (X) (X) → C

Z
α 7→ ω ∧ α.
X

Theories of holomorphic Chern-Simons type on Calabi-Yau manifolds of

complex dimension k can thus be constructed from dg symplectic manifolds
with a symplectic form of cohomological degree k − 1.
In this paper, we are only interested in one-dimensional holomorphic Chern-
Simons theories. In these theories, the source dg manifold is Σ∂ , where Σ is a
Riemann surface equipped with a never-vanishing holomorphic volume form.
The target is Y∂ , where Y is a holomorphic symplectic manifold. (The holo-
morphic symplectic form can be thought of as a dg symplectic form on Y∂ ).
958 Kevin Costello

8.4. We can now give a more precise statement of the theorem relating elliptic
cohomology and the Witten genus.
Theorem 8.4.15. Let X be a complex manifold. Then,
1. The obstruction to quantizing the holomorphic Chern-Simons theory
whose fields are maps C∂ → (T ∗ X)∂ is
ch2 (T X) ∈ H 2 (X, Ω2cl (X))
where Ω2cl (X) is the sheaf of closed holomorphic 2-forms on X.
2. If this obstruction vanishes (or, more precisely, is trivialized), then we
can quantize holomorphic Chern-Simons theory to yield a factorization
algebra on C with values in quasi-coherent sheaves on XdR × Spec C[~].
~
We will call this factorization DX,ch .
3. If E is an elliptic curve, then there is a quasi-isomorphism of BD algebras
in quasi-coherent sheaves on XdR × C[~]


~
F H(E, DX,ch ) ' Ω−∗ (T ∗ X)[~], ~Lπ + ~{log Wit(X, E)−} .

~
8.5. Recall that the factorization homology complex F H(E, DX,ch ) is defined
E,~
by first constructing a factorization algebra DX,ch on E, using the descent
property of factorization algebras; and then taking global sections.
Quantum field theories in the sense of [Cos10b] have a descent property
similar to that satisfied by factorization algebras, and the construction of a
factorization algebra from a quantum field theory is compatible with descent.
The quantum field theory on an elliptic curve E which arises by descent from
holomorphic Chern-Simons theory on C is simply holomorphic Chern-Simons
theory on E.
~
Thus, one can interpret the factorization homology group F H(E, DX,ch ) in
terms of holomorphic Chern-Simons theory on the elliptic curve E. From this
~
point of view, F H(E, DX,ch ) is the cochain complex of global observables for
the holomorphic Chern-Simons theory of maps E → T ∗ X.
This theorem is proved using the Wilsonian approach to quantum field the-
ory developed in [Cos10b]. The result is then translated into the language of
factorization algebras. The proof appears in [Cos10a].

References
[AKSZ97] M. Alexandrov, M. Kontsevich, A. Schwarz and O. Zabronovsky, The Ge-
ometry of the master equation and topological field theory, Internat. J.
Modern Phys. 12(7), 1405–1429 (1997), hep-th/9502010.
[BD04] A. Beilinson and V. Drinfeld, Chiral algebras, volume 51 of American
Mathematical Society Colloquium Publications, American Mathematical
Society, Providence, RI, 2004.
A Geometric Construction of the Witten Genus, I 959

[BNT02] P. Bressler, R. Nest and B. Tsyagn, Riemann-Roch theorems via deforma-

tion quantization, I, Adv. Math. 167(1), 1–25 (2002), math.AG/9904121.
[CF00] A. Cattaneo and G. Felder, A path-integral approach to the Kontsevich
quantization formula, Comm. Math. Phys. 212, 591–611 (2000).
[CG10] K. Costello and O. Gwilliam, Factorization algebras in perturbative quan-
tum field theory, (2010).
[Cos10a] K. Costello, A geometric construction of the Witten genus, II, (2010).
[Cos10b] K. Costello, Renormalization and effective field theory, (2010),
http://www.math.northwestern.edu/~costello/.
[Fed96] B. Fedosov, Deformation quantization and index theory, Akademie Verlag,
1996.
[GMS00] V. Gorbounov, F. Malikov and V. Schechtman, Gerbes of chiral differential
operators, Math. Res. Lett. 7(1), 55–66 (2000).
[Kon03] M. Kontsevich, Deformation quantization of Poisson manifolds, Lett.
Math. Phys. 66(3), 157–216 (2003), q-alg/0709040.
[Lur09] J. Lurie, Derived algebraic geometry VI: Ek algebras., (2009),
http://math.mit.edu/~lurie/papers/DAG-VI.pdf.