Commun. Math. Phys.

164, 525-562 (1994) C o m m u n i c a t i o n s ΪΠ

Mathematical
Physics

Gromov-Witten Classes, Quantum Cohomology,

and Enumerative Geometry
M. Kontsevichι, Yu. Manin 2
Max-Planck-Institut fur Mathematik, Gottfried-Claren-Strasse 26, D-53225 Bonn, Germany
1
E-mail: [email protected]
2
E-mail: [email protected]

Received: 11 April 1994

Abstract: The paper is devoted to the mathematical aspects of topological quantum

field theory and its applications to enumerative problems of algebraic geometry.
In particular, it contains an axiomatic treatment of Gromov-Witten classes, and a
discussion of their properties for Fano varieties. Cohomological Field Theories are
defined, and it is proved that tree level theories are determined by their correlation
functions. Application to counting rational curves on del Pezzo surfaces and projective
spaces are given.

1. Introduction

Let V be a projective algebraic manifold.

Methods of quantum field theory recently led to a prediction of some numerical
characteristics of the space of algebraic curves in V, especially of genus zero,
eventually endowed with a parametrization and marked points. It turned out that an
appropriate generating function Φ whose coefficients are these numbers has a physical
meaning ("potential," or "free energy"), and its analytical properties can be guessed
with such a precision that it becomes uniquely defined. In particular, when V is a
Calabi-Yau manifold, Φ conjecturally describes a variation of Hodge structure of the
mirror dual manifold in special coordinates (see contributions in [Y, Ko, Ma2]) which
identifies Φ as a specific combination of hypergeometric functions.
In this paper, we use a different tool, the so called "associativity" relations, or
WDVV-equations (see [W, D]), in order to show that for Fano manifolds these
equations tend to be so strong that they can define Φ uniquely up to a choice of
a finite number of constants. (For Calabi-Yau varieties these equations hold as well,
but they do not constrain Φ to such extent.)
Mathematically, this formalism is based upon the theory of the Gromov-Witten
classes. In our setup, they form a collection of linear maps ljn^ : H*(V, Q)® n —•
H*(Mgn,Q) that ought to be defined for all integers g > 0,n + 2g - 3 > 0,
and homology classes β e H2(V, Z) and are expected to satisfy a series of formal
526 M. Kontsevich, Yu. Manin

properties as well as geometric ones. (Here Mg n is the coarse moduli space of stable
curves of genus g with n marked points.)
In Sect. 2 of this paper, we compile a list of these formal properties, or "axioms"
(see Subsect. 2.2.0-2.2.8), and explain the geometric intuition behind them (2.3.0-
2.3.8). This is an elaboration of Witten's treatment [W].
Unfortunately, the geometric construction of these classes to our knowledge has
not been given even for V — P 1 . The most advanced results were obtained for g = 0
by the techniques of symplectic geometry going back to M. Gromov (see [R, RT]),
but they fall short of the complete picture. In 2.4 we sketch an algebro-geometric
approach to this problem based upon a new notion of stable map due to one of us
(M. K.)
The axiomatic treatment of Sect. 2 in principle opens a way to prove this existence
formally, at least for some Fano varieties V and g = 0. This is the content of Sect. 3
and the Reconstruction Theorem 3.1, which basically says that Gromov-Witten classes
in certain situations can be recursively calculated. However the equations determining
these classes form a grossly overdetermined family, so that checking compatibility
at each step presents considerable algebraic difficulties. The Second Reconstruction
Theorem 8.8 shows that it suffices to check this compatibility for codimension zero
classes. This allows one to extend the construction of [RT] from codimension zero to
all tree level GW-classes.
The subject matter of Sect. 4 is the beautiful geometric picture encoded in the
potential function Φ constructed with the help of zero-codimensional Gromov-
Witten classes of genus zero. Namely, over a convergence subdomain M C H*(V)
(the cohomology space being considered as a linear supermanifold) Φ induces the
following structures:
a) A structure of the (super)commutative associative algebra with identity on the
(fibers of the) tangent bundle &Ή*(V) depending on the point 7 e H*(V). The
fibers ^H*(Y) were called by Vafa "quantum cohomology rings" of V.
b) A flat connection on 3^H*(V) which was used by B. Dubrovin [D] in order to
show that the associativity equations constitute a completely integrable system.
c) An extended connection on ^H*(V) lifted to H*(V) x P 1 and its partial Fourier
transform which may define a variation of Hodge structure.
We show that the axioms for the Gromov-Witten classes imply all the properties
of Φ postulated in [D].
Together, Sects. 2 and 4 can be considered as a pedagogical attempt to present the
formalism of correlation functions of topological sigma-models in a form acceptable
for mathematicians with algebro-geometric background.
A more ambitious goal of our treatment is to define a framework for the
conjectural interpretation of H*(V) as an extended moduli space (see [Ko] and
Witten's contribution to [Y]).
In Sect. 5 we discuss examples. Since from the enumerative geometry viewpoint
the logic of this discussion is somewhat convoluted, we try to describe it here.
Assuming the existence of the relevant Gromov-Witten classes we calculate the
potential Φ and give the recursive formulas for its coefficients whenever feasible.
Assuming in addition that these classes can be constructed and/or interpreted along
the lines of Sect. 2, we state the geometric meaning of these numbers.
On the other hand, the potential Φ can be directly defined by using the (numerical
version of the) Reconstruction Theorem. Then the redundancy of the associated
equations translates into a family of strange number-theoretical identities. In principle,
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 527

they can be also checked directly, without recourse to the geometric context in which
they arose. Until this is done, they remain conjectural. We discuss del Pezzo surfaces
from this angle (cf. also [I]).
The last three sections are devoted to a description of a less constrained structure
of Cohomological Field Theory. Roughly speaking, we forget about the dependence
of our theory on the target manifold V, and retain only its part dealing with moduli
spaces. In Sect. 6, we give two definitions of a CohFT and prove their equivalence.
One is modelled upon the axiomatics of Gromov-Witten classes, another is based
upon (a version of) operads.
This formalism is used in Sect. 7 for a description of the cohomology of moduli
spaces of genus zero. Keel in [Ke] described its ring structure in terms of generators,
the classes of boundary divisors, and relations between them. We need more detailed
understanding of linear relations between homology classes of boundary strata of
any codimension, and derive from Keel's result a complete system of such relations.
(E. Getzler informed us that he and R. Dijkgraaf obtained similar results).
Finally, in Sect. 8 we prove the second Reconstruction Theorem, which allows
us, in particular, to classify Cohomology Field Theories via solutions of WDVV-
equations, and to formally prove the existence of GW-classes e.g., for projective
spaces. This theorem can be viewed as an instance of a general principle that a
quantum field theory can be completely recovered from the collection of its Green
functions.
The authors are grateful to the Max-Planck-Institut fiir Mathematik in Bonn for
the stimulating atmosphere in which this work was done.

2. Gromov-Witten Classes

2.1. Setup. Let V be a projective algebraic manifold over C with canonical class Kv.
Denote by B C H2(V, Z) the semigroup consisting of homology classes β such
(L.β) > 0 for all Kahler L.
In what follows, we will often consider cohomology classes as represented by
differential forms, and then write e.g. §cx(T(V)) instead of (—K V .C). Cup product
c
is denoted Λ.
2.2. Definition. A system (resp. tree level system) of Gromov-Witten (GW) classes for
V is a family of linear maps
J : Θn
£n,/3 #*(^,Q) -* H\Mgn,Q) (2.1)
defined for all g > 0, n > 0, n -f 2g — 3 > 0 (resp. g = 0, n > 3) and satisfying the
following axioms.
2.2.0. Effectivity. lvgnβ = 0forβ(£B.

2.2.1. Sn-Covariance. The symmetric group Sn acts upon H*(V, Q)® n (considered
as super space via Z mod 2 grading) and upon Mg n via renumbering of marked points.
The maps I^n β must be compatible with this action.

2.2.2. Grading. For ^γ e H\ put |7| = i. The map l^nβ must be homogeneous of
degree 2(Kv.β) + (2g - 2)dim c V, that is
n
I = 5Z |7il + 2(Kv.β) + (2^ - 2)dimc V . (2.2)
2=1
528 M. Kontsevich, Yu. Manin

Before stating the remaining axioms, let us introduce the following terminology.
Call a GW-class basic if it corresponds to the least admissible values of (n, β) that
is, belongs to the following list:
h i ® 7 2 ® 7 3 ); ^i,/3(7); ^ O ) , 9 > 2, (2.3)
where 1 is the canonical generator of H*(V, Q)®° = Q. Call a class new if it is not
basic, and if among its (homogeneous) arguments ηi there are none with |^γ| = 0 or
2. Finally, call the number
2(3<7 - 3 + n) - I J£ n ϊ / 3 (7i ® .. <8> 7 n )l
the codimension of the class (recall that dim c M ^ n = 3# — 3 + n). The classes of
codimension zero are especially important and are expected to express the number of
solutions of some counting problems (see 2.3 below). Instead of such a class Γ£n β
we will often consider the corresponding number (ljnβ) defined by

< C Λ ® ® 7n) = / #nl/3(7l ® ® Ίn) (2-4)

Mg,n

Notice the following facts:

a) For g = 0, all basic classes have codimension zero, because dim c M o 3 = 0.
b) For a non-vanishing class, we have from (2.2):
n
1
(~Kv.β) - n - ( g - l)dimc V < - ^ ( | 7 i | - 2)

< (-if v ./3) - (p - l)(dim c V-3), (2.5)

with the second equality sign for codimension zero classes. If the class is new,
and Hι(V) = 0, the middle term of (2.5) must be non-negative, and even > n,
if H3(V) = 0. Hence for Kv = 0,dim c V = 3, (e.g., Calabi-Yau threefolds) there
are no non-vanishing new classes. For Fano varieties, this inequality bounds n if β
is fixed, and β if n is fixed.
The next two axioms partially reduce the calculation of GW-classes to that of basic
ones and new ones.
2.2.3. Fundamental Class. Let βy G H°(V) be the identity in the cohomology ring
(i.e. its dual homology class is the fundamental class [V]). If the following class is
not basic, we have

C/3(7i ® • ® 7n-i ® 4 ) = <#n-i,/?(7i ® ®7n-i), (2-6)

where π n : Mg n —> M^ n _ 1 is the map forgetting the n-th section. In particular,
(2.6) cannot be of codimension zero unless it vanishes.
In addition, for basic classes with argument e\r we have
f 0, if β φ 0;
WΛ/J>(7I ® 7 2 ® ey) = < / 7 l Λ 72, if /3 = 0. (2-7)
Iv
2.2.4. Divisor. If J ^ β(ηλ ® . . . (E>7n) is a non-basic codimension zero class with the
last argument 7 n , I7J = 2, then

'n> = / 7 n * (Jp,n-l,)9)(7l ® ® 7 n -l> '

β
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 529

More generally, for non-necessarily codimension zero classes,

n» = / 7n ' ^ n - l ^ l ®'"® 7n-l)

2.2.5. Mapping to Point. By this catchword we describe the situation when β = 0.

For 0 = 0, the situation is simple: the only non-vanishing classes must have
n
= 2dim c V, and

;O
Mθ,n' W

where e ^ is the identity in ίf*(M 0 > n ).

n
For 0 = 1 , the non-vanishing classes have ]C I7J = 0 or 2 a n d therefore are lifts
ί=ι
of the following two basic ones:

^ ° (2.9)

= 2 : /y,,0(7) = f / ^ O O Λ 7^ 4., ,, (2.10)

where χ(V) = Σ (-1)%(V), e2Ή = CX{&{1)) (recall that Mιx ~ P 1 ), and

is a certain characteristic class of the tangent bundle of V whose explicit description
we postpone to 2.4.2.
Finally if g > 2, then / ^ n 0 ~ 0 unless dim c V < 3, and non-vanishing classes
again can be described explicitly.
n n n n
2.2.6. Splitting. Fix g{, g2 and nj, n2 such that g — gx +025 — i + 2 ' i^9ι~^ —
0. Fix also two complementary subsets S = Sι^S2 of {1,... ,n}, |5J = ni. Denote
by φs : Mgunι+ι x M^ 2 n 2 + 1 —» M^ n the canonical map which assigns to marked
curves (C^x^,... ,x ( ^ + 1 ),i = 1,2, their union C^ U C 2 , with ^ ^ j identified to
x[ . The remaining points are then renumbered by { 1 , . . . , n} in such a way that their
relative order is kept intact, and points on Ci are numbered by Sτ.
Finally, choose a homogeneous basis {Δa\a = 1, . . . , £ ) } of iJ*(y, Q) and put

V
The Splitting Axiom now reads:

where ε(5) is the sign of the permutation induced by S on {7.} with odd
530 M. Kontsevich, Yu. Manin

Notice that ]Γ Δa <g> Δbgab is the class of the diagonal in H*(V x V).
a,b
The sum in (2.11) is finite because of the Effectivity Axiom 2.2.0.
2.2.7. Genus Reduction. Denote by ψ : Mg_{ n + 2 —* Mgn the map corresponding
to glueing together the last two marked points. Then

2.2.8. Motivic Axiom. The maps / J n ^ are induced by certain correspondences in the
Chow rings:
ClnφeC\VnxMgn). (2.13)
We recall that c e C*(V x W) induces the map
[c]), (2.14)
where πv, πw are projections of V x W to V, VF and [c] is the cohomology class
of c.
2.3. Comments to the Definition. When Gromov-Witten classes are considered in the
literature in an algebraic context, it is usually assumed that V is either Fano (—Kv
ample), or Calabi-Yau (Kv = 0) so that (—Kv.β) > 0 is automatically satisfied
for all algebraic homology classes. However, nothing in the formalism forces us to
postulate it. Notice that for manifolds of general type (2.5) implies vanishing of all
ΐζn β with large enough β.
The simplest example of a tree level system of GW-classes on V is:
JO, if β φ 0,
^n,^(7i Ίn) = S ( J7 l Λ . . . Λ 7 r Λ e°Ή otherwise . (215)
n
IV J °'
In addition, any system of GW-classes posesses an obvious scaling transformation
(if we allow to extend coefficients Q to C): I^n β ι-» e(β)I^n β, where e : B —> C*
is a semigroup homomoφhism. If we put e(β) = exp(—t(ω.β)) for a Kahler class ω,
then the scaling of any initial tree level system tends to (2.15) as t —» oo. In terms of
quantum cohomology (see 4.5 below) (2.15) gives rise to the classical cohomology
ring, whereas I^nβ supply the instanton corrections.
Intuitively, one can imagine the geometry behind these corrections ί^n β as
follows. For ητ e ίf*(V, Z), choose some generic representatives Γτ of dual homology
classes in V. Consider an appropriate space of triples {/, C ; x l t . . . , xn}, where C is
a curve of genus g with n marked points xi, and / : C —> V is an algebraic map
such that /*([C]) = β and f(x%) e Γ% for all i. The projection of this space to Mgn,
under some genericity conditions, must be the cycle dual to the cohomology class
Iv
As we have already remarked, the most powerful known constructions of the
classes J j n β leave the domain of algebraic or even complex geometry, in order
to satisfy the necessary genericity assumptions. The whole subject seems to belong
rather to symplectic topology: cf. [R]. To our knowledge, only a part of the picture
of Def. 2.2 is at the moment rigorously established even in this wider context.
We will use the naive picture described above in order to motivate the expected
properties of GW-classes.
2.3.0. Effectivity. The meaning of this axiom is obvious.
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 531

2.3.1. Grading. The condition (2.2) expresses the following genericity assumptions.
Firstly, the space of maps / : C —• V landing at β must have the complex dimension
of its first order infinitesimal approximation at a point, that is H°(Cy f*(.9y)) which
in turn must be unobstructed and coincide with χ(C, /*(^y)) By Riemann-Roch,
= (-Kv.β) + (1 - g)άimcV.
Secondly, when we constrain f(C) to intersect all Γ/s, this diminishes (real)
dimension by ^ dim R Γi.
i
From this discussion it is clear that zero-codimensional classes, or rather numbers
(lYn β) morally count curves constrained by incidence conditions to such a degree
that only a finite number of such curves occur "generically." For instance, the number
of curves of given degree d on P 2 passing through 3d — 1 fixed points in general
position and having additional (d— l)(d — 2)/2 double points elsewhere must actually
coincide with (io*j3d_1 dβo)(e®3d~])> where β0 is the homology class of a line, and
e the dual cohomology class of a point. However, in more complex situations such
naive counting may be totally misleading.
We nevertheless use this language, but it should not be taken too literally.
2.3.2. Sn-Covariance. The meaning of this Axiom is obvious. Notice that codimen-
sion zero classes are simply 5 n -symmetric.
2.3.3. Fundamental Class. The dual cycle to e°v is V. Therefore, the l.h.s. of (2.6)
imposes no constraints on the n-th point of C. The r.h.s. expresses this in terms of
moduli space.
2.3.4. Divisor. If I^n^(ji 0 . . . ® 7 n ) is zero-codimensional class, the same is true
for / ^ n _ ! β(ri\ 0 . . . (8) 7 n _i) Hence the l.h.s. of (2.4) (resp. the integral at r.h.s.)
counts the number of marked curves passing through Γx,...,Γn (resp. Γx,..., Γn_λ).
But the two problems differ only by the additional presence of the n-th point in the
l.h.s. which may be chosen among intersection points of f(C) and Γn. Their number

2.3.5. Mapping to Point. If β = f*([C]) = 0, then / maps C to a point which is

constrained to belong to Γx Π . . . Π Γn\ otherwise the curve and the marked points on
it are arbitrary. This justifies the axiom (2.8) for g = 0.
Unfortunately, for g > 1 this reasoning is too naive, and dealing with this
very degenerate situation requires some sophistication. In 2.4.4 below we sketch an
argument giving simultaneously (2.8)-(2.10) and formulas for g > 2.
2.3.6. Splitting. J n the picture_described above, the l.h.s. of (2.11) can be represented
by a cycle on M + 1 x Mg2 n 2 + 1 which is the sum of cycles corresponding to
various splittings β = βλ + β2. When βx,β2 are chosen, we must consider pairs of
maps /j : Cλ -> V x V and f2 : C2 -» V x V such that τr2 o fλ{C{) and π{ o f2(C2)
are points, /^([CJ) = ^_-(/?J, with the incidence conditions described by τr*(ΓJ,
the partition 5, and the additional relation /iθ*4 1) +1 ) = f^xψ). On V x V1 the latter
can be expressed by intersecting with diagonal. This leads to the r.h.s. of (2.11).
2.3.7. Genus Reduction. A similar reasoning motivates (2.12).
2.3.8. Motίvic Axiom. For g = 0, and at least V with tangent sheaf ,9γ generated
by global sections, one can try to construct C§n β directly as follows. Consider the
532 M. Kontsevich, Yu. Manin

scheme of maps Map^P 1 , V) such that /^([P 1 ]) = β. Construct the morphism

n (V^\n v Man fP* \ί\ v T/ n v λ/f

-~ Z™ ^» . f Λ / f (rγ \ f (rγ Λ /ϋ 1 rf rp \\
u
n ^ l ' ***' ^ n ' J / vj V^l^? 5 J v^n^' ^ ' 1' ' ' * ' ^ n "

(here (P 1 )^ is (P 1 ) 7 1 with deleted diagonals.) The closure of the image of an is a

candidate for C^n β.
Generally, this construction is inadequate; but see 2.4 below for a more refined
version which hopefully works for certain V.
Anyway, if one takes the Motivic Axiom for granted, then all the other axioms
can and should be stated directly in terms of C^n β. We will show how to do it for
the Splitting Axiom leaving the remaining ones to the reader.
n n

For a product Y[ Wi and a subset T C {1,..., n}, denote by τr τ : Y[ Wτ —>

Yl Wi the projection. Keeping the notation of 2.2.6, consider the following corre-
spondences (Δ = the diagonal class in C*(V x V)):
C{ = (iάyn xφs)*Cln,β € c*(yn x Mgunι+ι x Mg2tn2+1),
r - V nr*

Then

where t £ Sn+4 is the obvious reshuffling of factors.

We did not start with data (2.14) for two reasons. First, at present, when
constructions of GW-classes rely upon Gromov's symplectic methods (which actually
work for non-algebraic manifolds), (2.14) looks unnecessarily restrictive. Second,
using axioms like the Splitting Axiom directly in terms of correspondences would
entail a very clumsy notation, especially in the next section.
2.4. Construction Project. Fix an algebraic manifold V as above.
2.4.1. Definition. A stable map (to V) is a structure (C; {x{1..., xn}, /) consisting
of the following data.
a) (C;Xj,..., xn), or simply C, is a connected reduced curve with n > 0 pairwise
distinct marked non-singular points and at most ordinary double singular points.
b) / : C —> V is a map having no non-trivial first order infinitesimal automorphisms,
identical on V and ( x 1 ? . . . ,x n ). This means that every component of C of genus 0
(resp. 1) which is contracted by f must have at least 3 (resp. I) special (i.e., marked
or singular) points on its normalization.
For an algebraic cohomology class β e H2(V, Z), consider the stack M (V, β) of
stable maps to V of n-marked curves of arithmetical genus g such that /*([C]) = β
for any point ( ( C ; x l 7 . . . , x n ) , / ) of this stack. We expect that this stack is proper
and separated.
If n + 2g - 3 > 0, there is a map M (Vyβ) --> Vn x Mg^n consecutively
contracting the non-stable components of C. For g = 0, it is useful to extend this
map putting M o —point for n < 2.
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 533

2.4.2. Definition. V is called convex if for any stable map f : C —> V of genus zero
ι
we have H (C
Examples, a) Generalized flag spaces G/P are convex.
<8>n
b) More generally, if for some n > 0, ^ is generated by global sections, V is
convex.
c) Although in general Fano manifolds are not convex (look at exceptional curves
on del Pezzo surfaces), it is conceivable that indecomposable Fano manifolds of
sufficiently large anticanonical degree are.
We expect that M o (V, β) is a smooth stack ("an orbifold") whenever V is convex.
2 A3. Construction. For a convex V, denote by C^n β the image of the fundamental
class of M 0 > n (V,β) in C*(Vn x M O n ) .
2.4A. Hope. For convex V, {C^ β] is a tree level motivic system of GW-classes.
The main property to be checkedjs the Splitting Axiom. It must follow fern
the natural structure at infinity of M o n(V, β) which is parallel to that of M o n
(stratification according to the degeneration graph of a curve).
If g > 1 and/or V is not convex, the fundamental class of Mg n (V, β) is "too
big," and Cjn β must be defined as image of a characteristic class of an appropriate
"obstruction complex" on this moduli space. We illustrate the arising complications
on the "mapping to point" example.
2.4.5. Mapping to Point. By Definition 2.4.1, we have

The complex dimension of this space is bigger than the expected one by g dim c V :=
G. This is precisely the rank of the locally free obstruction sheaf J 7 " ^ on M o (V, β)
whose geometric fiber at the point [/], f(C) = v e V, is
H\C, Γi^v)) = H\C, (9C) (g) TVV .
Denote by & n the locally free sheaf Rιπ^& on Mg n , where π is the projection of
the universal curve. We have ^ ( 1 ) = ^ n EiJ?y.
Notice that Wgn is simply the pullback of one of the basic sheaves ^Q3 =
0, Wxλ, Sg 0 , for g = 0,1, > 2 respectively.
Consider now the Euler class cG0f^) and denote by Pι,p2 the two projections of
M n x V. An intuitive argument appealing to our desire to pass to an unobstructed
situation suggests the following definition:
^n,o(7i ® ® 7 n ) = Vu^G^ ^ ^ ) Λ p 2 *( 7 l Λ . . . Λ 7 j ) .
We hope to develop this construction in a separate publication.
We will end our general discussion of GW-classes with the following two
constructions.
2.5. Tensor Product. Let {I^niβ}, i = 1,2, be two full (or only tree level) GW-
systems for VVV2, H1^) = 6. Put

== < n A
(7{U ® ® 7i υ ) Λ / 9 V2 ni/32 (7ί 2> ® • ® 7 f ) , (2.16)
where ε is the sign of the relevant permutation of odd-dimensional arguments.
534 M. Kontsevich, Yu. Manin

2.5.7. Claim. (2.15) is a full (resp. tree level) system of GW-clases for Vλ x V2 which
is called the tensor product of given systems.
In fact, one easily checks all axioms, including the refined version of the Mapping
ι
to Point Axiom. Probably, condition H (Vτ) = 0 is irrelevant.
Notice that even if one is interested only in codimension zero classes of the tensor
product, one has to know all classes of the factors. In the tree level setting, they can
be in turn be calculated from the codimension zero classes of the factors, but in a
highly non-trivial way. In fact, we have:
2.5.2. Proposition Let a tree level system of GW -classes I^n^ be given for V. Then
it can be uniquely reconstructed from its codimension zero subsystem.
Proof. In fact, consider a class of codimension > 1 with n > 4. The Splitting
Axiom (2.11) allows one to calculate its restrictions to all boundary components
of the moduli space (corresponding to lesser values_of n). It remains to show that
ΠsKεrφ*s = H2n-6(M0n). In fact, let ds e H2(M0n) be the dual class of the
boundary component corresponding to the partition S. The whole cohomology ring
is generated by these classes (see [Ke]). On the other hand, φs*φ*s(ά) ~ a Λ ^s f° r
any class α. And if a Φ 0, \a\ < 2(n — 3), then by Poincare duality there exists some
non-constant monomial d in ds such that / a Λ d Φ 0. Hence a φ Γ)s Ker φ*s.

Summarizing, we may say that the tensor product can be defined on the codimen-
sion zero subsystems, but there are no simple formulas for doing it.
2.5.3. Cusp Classes. Generally, let us call cusp classes those elements of H*(Mgn)
which vanish on all boundary divisors of this moduli space. For g > 1, non-trivial
cusp classes may exist (e.g. the Ramanujan class for g = 1, n = 11). It would be
interesting to have examples of GW-classes with non-trivial cusp components.
2.6. Restricted GW-Systems. Let C* be an intersection theory such that a given
GW-system {I^n β] can be represented by C-correspondences. We have C * ( F n ) C
H*(V)®n. We will say that the maps I^nβ restricted to C*(Vn) form the restricted
GW-system. Slightly elaborating the discussion of the Splitting Axiom in 2.2.8, one
can convince oneself that all the axioms restricted to ΘnC*(Vn) make sense and can
be stated entirely in terms of this restriction, without appealing to extra cohomology
classes like Δa in (2.11) and (2.12).
This is useful for those enumerative problems where we want to consider incidence
conditions stated in terms of algebraic cycles only.

3. First Reconstruction Theorem

3.1. Theorem. Let V be a manifold for which a tree level system of GW-classes
eχ
itf,n,β} i<sts'
If H*(V) as a ring is generated by iί 2 (V), then {I^nβ} can be uniquely
reconstructed starting with the following system of codimension zero basic classes:

^73) I (-K y ./5)<2dim c y+l;

3 Ί

Σ\Ίι\=2{-Kv.β) + 2άimcV- | 7 3 | = 2 [ . (3.1)

Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 535

3.1.1. Comments. We may and usually will choose I7J from the elements of a fixed
basis of H*(V). Then, if V is Fano, (3.1) is a finite set because the degree of β
is also bounded. For instance, if V = P n , (3.1) is satisfied only by β = 0 and
β = class of a line. The β = 0 case is settled by (2.8). For the line, (3.1) gives
{7i572>73} = {e 2 n ,e 2 r \e 2 }, where e2i = q ί ^ l ) ) * . Since e2n is the dual class of a
point, one can imaginatively say that all enumerative problems about rational curves
in P n eventually reduce to counting the number of lines passing through two points.
On the other hand, for Calabi-Yau varieties with Kv = 0 (3.1) does not restrict
2
β at all. Besides, H (V) almost never generates H*(V). Nevertheless, Theorem 3.1
does say something about this case as well.
a) The algebraic (or Hodge) part of cohomology may be generated by H2, and the
corresponding restricted GW-system can be reconstructed from (3.1).
b) Then Theorem 3.1 says that all tree level classes with algebraic arguments can
be reconstructed if one knows all /^-contributions to the triple quantum intersection
indices. This information is conjecturally supplied by the Mirror family.
3.2. Proof. It will consist of several reduction steps.
3.2.1. Step 1. Every class I^n β{ηλ 0 . . . ® ηn) of codimension > 1 with n > 4 can
be reconstructed from classes with lesser values of n.
The proof of Proposition 2.5.2 shows this.
It remains to deal with codimension zero classes, that is, numbers

2
(-KV'^ +2 d i m c y + 2(n - 3). (3.2)

We start with some preliminaries.

3.2.2. Quadratic Relations. Fix some {β\Ίγ, • - 7JV}» N > 4, and four pairwise
distinct indices {i,j,k,l} C { l , . . . , i V } . Assume that iXNβ(Π\ 0 ••• 0 7 ^ ) n a s
topological codimension two. Take the cup product of this class with the following
linear relation between the generators ds established in [Ke]:

{ijSkl} {ikTjl}

where Σ m e a n s t h a t w e s u m o v e r a l l p a r t i t i o n s S s u c h t h a t i ,j G S j , k , l G S 2 ,
{*jS7c/}
or vice versa.
Calculate the degrees of all summands using (2.11). We get the following
fundamental system of quadratic relations among codimension zero classes:
Λ v
V^ V S^ F(qMτ \ / ( (\?\
fc i
/ , / , Z_j v^Λ o,|51| + l, i 0 1 / V ^
{ijSkl} βλ+β2=β a,b V \reSι

= Σ ΣΣ
{ikTjl} β\+β2=β a,b
/ / \\
(3.3)
536 M. Kontsevich, Yu. Manin

Now, define a partial order on pairs (/3,n), β £ B, n > 3, by setting

09, n) > (β\ n'), iff either β = β'+ β"', /?', /?" G B, /?" ?έ 0, or /? = /?', n > n'.
Observe that the highest order terms enter in (3.3) linearly. In fact, for these terms
we have either βx — β or β2 = β. The complementary class, with β2 = 0 (resp. /^ =
0), can be non-zero only if | 5 ? | = 2 or \T2\ = 2 (resp. ^ | = 2 or \Tλ | = 2): see (2.8).
Hence there are four possibilities: Sx = {i,j}\S2 — {k, l}\Tλ = {i, k}\T2 — {j,l}
Let us look, say, at the first group of highest terms:

We have by (2.8):

(^3,o>(7t ® 7, ® ^ α ) = / 7< A 7,

Since (/Q^V-I /?) ^s (poly)linear, we can rewrite (3.4) as

Using analogs of (3.5) for all four groups of highest order terms we can finally write
(3.3) as
Λ
i 7, ® ί )

Λ7fc j
= a quadratic combination of lower order terms. (3.6)
r
3.2.3. Step 2. Every class I^n β( γι ® . . . Θ 7 n ) of codimension zero with n > 4 can
be reconstructed from basic classes (with n = 3).
In fact, it suffices to calculate numbers (I^n β)(ηλ <S> 0 7 n ) for n > 4 and
2dimc(y)>|7l|>...>|7n|>4
(if | τ j = 2, we can apply the Divisor Axiom to reduce n). We will now for the first
time use the assumption that H*(V) is generated by H2(V) and write Ίn = Y^lδi/\δ'i
i
for some δti 6f{ with | ^ | = 2. Clearly, it suffices to treat the case ηn = <SΛ<5', \δ'\ = 2.
Apply the construction of 3.2.2 to the codimension two class

and indices {z, j , fc, /} = {1,2, n,n+ 1}. Relation (3.6) becomes
J y
± ( o n,/?)(7i Λ 7 2 (8) 73 0 . . . 0 7 n -i 0 « 0 ί

,/ ^ 7 n - l ® «)
= a quadratic combination of lower terms . (3.7)
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 537

<
Now, the second summand in (3.7) is our initial class (I^n β)(jι 8>>. ®7 n _i ®7 n )
The first and the third summands are lifted from MQ n_ι because of the Divisor Axiom.
Finally, in the fourth summand the last argument is_of lesser dimension than in the
initial class: \δ\ < \jn\. If \δ\ = 2, it is lifted from Mo n_ι; if \δ\ > 2, we can repeat
the same trick applying it to this summand. In a finite number of iterations, we will
reduce n.
3.2.4. Step 3. Every basic class I^3 β can be calculated via those with restrictions
(3.1).
In fact, if |7 3 | > 4 in I^3 β(ηγ ® η2 (g> 7 3 ), β φ 0, (the β = 0 case is given by
(2.8)), then we can apply to this class the reduction procedure described above and
diminish |7 3 |. The remaining conditions follow from the Grading Axiom.

4. Potential, Associativity Relations, and Quantum Cohomology

4.1. Setup. Let M be a supermanifold endowed with a tensor g of rank two and a
tensor A of rank three. To fix our sign conventions, it is convenient to choose a (local)
coordinate system {xa} which defines the basis da = d/dxa of vector fields and the
basis dxa of 1-forms. Our tensors then have components gab and Acah.
Generally, x denotes the Z/2Z-degree of x. To simplify notation, in superscripts
we replace xa by a so that e.g. {-iγ^b^cχd becomes ( _ i ) α b + c d . Hopefully, this
will not lead to a confusion.
We want g and A to be even, i.e. gab = ( - l ) α + 6 , Acab = ( - l ) α + 6 + c . The sign
conventions about the de Rham complex are: d is odd, and Ω*M is the symmetric
algebra of ΩιM.
4.2. Pairing, Multiplication and Connection (Dubrovin's Formalism). We use gab in
order to define an even ^^-pairing on the tangent sheaf ^M :

We will always assume it to be symmetric (gba = (— l)abgab) and non-degenerate, so

that the inverse matrix (gab) = (gaι>)~1 exists. In addition, we will assume that gab
are constant, so that it defines a flat metric in a flat coordinate system (for non-flat
coordinates, cf. [D]).
We use Acah in order to construct, firstly, a structure of ^ M -algebra on ,TM with
multiplication o :
J2Λ, (4.2)
and secondly, a family of connections on 3^ depending on a even parameter λ and
defined by the covariant differential

Vx(db):=\J2dxaAcab®dc, (4.3)
a,c

or equivalently, by covariant derivatives

Acabdc = λ(da o db). (4.4)

We will now consecutively impose some relations upon A,g, and interpret them
both in terms of multiplications and connections.
538 M. Kontsevich, Yu. Manin

4.2.1. Commutativity^/Vanishing Torsion.

VαAc, Ala = (-l)abAcab. (4.5)

In view of (4.2), this means supercommutativity of (!7~M, o). From (4.4) it follows that

Vα,6, VλA(δ6) = (-ir6VλA(9α). (4.6)

4.2.2. Associativity/Flatness.
e d
Vα, b, c, d, Σ A abA ec = (_ l)-(^) Σ KΛa > (4 7a)

Vα, 6, c, d, 0 d A£6 = ( - l ) α d a α A 5 6 . (4.7b)

Using (4.2) one checks that (4.7a) is equivalent to the associativity relations
(da o db) o dc = da o (5 6 o <9C).
Using (4.3) one checks that (4.7a) and (4.7b) together are equivalent to V^ = 0.
More precisely, if one puts symbolically V λ = V o + λΛ, then (4.7b) is V 0 (A) = 0,
and (4.7a) is equivalent to [A, A] — 0 if one assumes (4.5) or (4.6).
4.2.3. Frobenίus Algebra. Put Aabc := ΣAlbgec. The next relation we impose is:
a

Vα,6,c, Aabc = (-l)a^c)Abca. (4.8)

Together with (4.6), this means that Aabc is 53-invariant (in the sense of superalgebra).
In terms of multiplication, (4.8) reads

that is, the scalar product is invariant wrt multiplication. In terms of connection, (4.8)
reads [use (4.4) and (4.1)]:

<V λ ϊ β β (0 6 )Λ> = ( - D α 6 ( 5 6 , Vλβa(dc)). (4.9)

4.2.4. Identity. Assume that the coordinate vector field dQ is even, and

V6,c, Acob = δbc (4.10a)

or equivalently,
Aobc = gbc- (4.10b)

According to (4.2), this means that d0 is identity in {£ΓM, o). According to (4.3), this
is equivalent to

or more suggestively,

This also means that

(4.11)

4.2.5. Potential. An even (local) function Φ on M is called a. potential for the (A, # ) -
structure, if
Ma^c, Aabc = dadbdcΦ. (4.12)
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 539

Such a potential always exists locally. On the other hand, for any function Φ and
the tensor of its derivatives Aahc1 the S^-invariance, in particular (4.8), is automatic.
If we then define Acah by ^Λabegec, then (4.5) is also automatic. If in addition gec
e
are constant, (4.7b) follows.
The crucial associativity relations (4.7a) then become a remarkable system of
quadratic differential equations called WDVV-equations in [D]:
ef
ΣdadbdeΦ g dfdcddΦ
ef

= (-\)«**ΣdbdβeΦ g*fdfdaddΦ. (4.13)

e/

We will now show how to derive a potential (A, g)-structure from a tree level
system of GW-classes.
4.3. GW-Potential. Let V be a manifold equipped with a system of tree level GW-
classes. We will actually use only the numbers {I^nβ)(Ί\ ® . ® 7 n ) with properties
postulated in Sect. 2.
We will consider H*(V, C) as a linear superspace Z/2Z-graded by 7 : = |7|mod 2,
and as a supermanifold which we then denote Hv. Our potential Φω will depend on
a choice of a class α; € H2(V, C) whose real part lies in the Kahler cone. We first
define Φω as a formal sum depending on a variable point 7 G Hv :

n>3 ^

To make sense of this expression, choose a basis {Δa} of ίf*(V,C), write the

generic point as 7 = ^ £aΔa, xa — Δa, and define the metric by Poincare duality:
a=0

(da,db)=gab:= J ΔaΛΔb. (4.15)

v

Since {I^n^)(Ji 0 . . . 0 7 n ) is simply 5 n -invariant when all 72 are even, we have

D '

where ε ( n 0 , . . . , n D ) = ε = ±1 is the sign acquired in the supercommutative algebra

D DO
S(H*(V))[x°, ...,xD] after reshuffling Π (xaΔa)Ua = ε Π Δ™a Π ( ^ α ) n α

There are several natural convergence assumptions that can be made about (4.14).
A. Vn > 3, there exists only finitely many effective β satisfying the grading condition
(3.2) for the zero—codimensional classes.
540 M. Kontsevich, Yu. Manin

This is the case of Fano manifolds. If this hypothesis is satisfied, (4.14) can be
α
interpreted as a formal series in x ,
-Jα,
ωKΊ)
2-< 2-< no!...nD!
υ υ
no+...+nD>3 β
n D
x (%nβ)(Δ® ° ® ... ® < " ) ( x ° Γ o . . . ( ^ ) " o , (4.17)
because each interior sum Σ is effectively finite.
β
B. The previous condition is not satisfied, but each Σ m ί/ze r.h.s. of (4Λ7) converges
β
for each n, at least when ω has a sufficiently large real Kάhler part
(Conjecturally, this is the case for Calabi-Yau manifolds.)
Then again, (4.17) is a well-defined formal series.
C. A or B is satisfied, and in addition Φω(j) converges in a subdomain M of Hv
(which may depend on ω) as a function of {xa}.
We expect this to be generally true, because of the conjectural exponential growth
estimates for — {lXn β)
The following formal calculations can be easily justified in each of these contexts.
4.4. Proposition. Let 7 = 7' + η2 + 7 0 , where \η2\ = 2, |7°| = 0, and η' = the sum
of components of dimension φ 0,2. Then
a) ΦJa' + 7 2 + 7°) = Φω(Ί' + 7 2 ) + a function quadratic in {ηf + η2).
b) Ifx° is the coefficient of η at Δo = e°v (identity in H*(V)) so that 7 0 = x°Δ0, we
have
d0dbdcΦJl) = 9bc- (4.18)
c) We have
Φω{ηr + 7 2 ) = Φω-Ί2(rt') + a function quadratic in 7, η'. (4.19)
n r 2 0n / 2 i / 2 j
Proof a) 7® = ( 7 + 7 ) + Σ ( 7 + 7 ) 0 7 ° Θ ( 7 + 7 ) plus terms containing
ι+j=n-l
0
7 at least twice. From the Fundamental Class Axiom it follows that {I^n β) = 0
for n > 4, if βy is among the arguments. Hence the contribution of 7 0 to Φω(η)
is restricted to the terms n = 3 in (4.17), and they are (no more than) quadratic in

b) We can now calculate dQdbdcΦω taking into account only the n = 3 terms:

This vanishes if β φ 0 and is gbc otherwise: see (2.7).

c) We have

p>3,q>0 β
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 541

(p ί/y _|_ *y \ rr: Λ > 6 (

n !
n>3 /?

e- β fω
=ΣΣ p+q>3 β
- P
'^'

Finally, because of the Divisor Axiom, for p > 3, q > 0, we have

4.5. Theorem-Definition. The tensors gab and dadbdcΦω define a potential Dubrovin
structure satisfying all the properties (4.1)—(4.11), with d0 as identity.
The fibers ofJ^ endowed with multiplication o are called the quantum cohomology
rings ofV associated with the tree level system of GW-classes Iv.
Proof It remains only to check the relations (4.13). Let us calculate the l.h.s. of (4.13)
using (4.17). The terms with fixed e, / are:

x e β2
Σ °°7~^^Zn2,

1
n>6,/? β{+β2=β nι+n2=n

X ( / ί ^ , Λ. ) ( 7 < g > ( n i ~ 3 ) (g) Z\^ (g) Z\L 0 Z

x (C,
Rewriting similarly the r.h.s., we see that (4.13) is equivalent to the family of identities

Σ Σ Σ( n e,/ ^ ' -
N c
\ O,n2,P2' / a i
-6
, -3
2 =/3 e,f

X {Cl!/3l)(7®(ni-3) ® A ® A: ® Δ
e)9ef
x {/0Vn2,/32)(^/ Θ Δa Θ A* ® 7 Θ ( n 2 "" 3 ) )
Obviously, this is a particular case of (3.3).
Remark A polarization argument shows that, vice versa, (4.13) implies all the
quadratic relations (3.3).
We will show now that the grading conditions imply an additional symmetry of
the (A,#)-structure with potential Φω.
542 M. Kontsevich, Yu. Manin

4.6. Scaling. If {7^ . . . ηn} = {Δa with multiplicity n o , a = 0 , . . . , £)}, then the
grading condition (3.2) for non-vanishing summands of Φω becomes
D
a(\Δ
Σna(\Δ a\ a
\ - 2) --ijc^V) = 2(dim
2 / Cι(V) = 2(dimccV-3).
V - 3). (4.20)
α=0 β

This leads to the following flows on our geometric data.

4.6.1. Scaling Transformation of the Potential.

φ μ_>e2(dimc V-3)tφ ^

[See (4.17).]
4.6.2. Scaling Transformation of the Dubrovin Structure. Since Aabc is the tensor of
the third derivatives of Φω, Proposition 4.4 allows one to replace the flow shift of
ω by the reverse shift of η1 — Σ xbΔb by —2cι(T(V))t. Defining the numbers
{ξh| |Z\bl=2}by \Δb\=2

we obtain the following flow: ω ι-» ω, and

χa ^ e(\Δa\-2)tχa ι^

Xb h-> X 6 - 2 £ 6 t , = 2;
|Zi έ (4.21)
j (2dimcV-|z\ α |-|^b|-|^c|)t 4 f4 2T>
α6c abc ' yr.z,z,j
Since the Poincare pairing is invariant, and gab φ 0 only for | Z\α | -f | Z\61 = 2dim c V,
from these basic formulas we get furthermore
a { Λa 2)t a
β ^ e(2-\Δa\)tQ dx ^ e \ \~ dx \

and finally, from (4.3),

v v e2t v v
λ *~* o + ( λ - o) (4.23b)

We now introduce an extended supermanifold Hv = f ί v x P 1 , where P 1 is the

completion of the affine line with coordinate λ, and extend the flow (4.21) to Hv by

λ h-» e2tλ (4.24)

which is another form of (4.23b).

From now on, ω can and will be assumed fixed.
4.6.3. Infinitesimal Form. The flow (4.21) can be written as exp(ίJί), where X is the
following even vector field on Hv :

X = Y^(\Δa\-2)xada-2
\Δb\=2
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 543

v
In view of (4.24), its natural extension to H is

(4.26)

v v
Let π : H —> H be the projection, and ^ — τ r * ( ^ y ) . For a local section d of
3^v, we denote by <9 its lift to J^". We will now extend V λ to a connection on jΓ\
4.7. Proposition. Put

λ
(4.27)
Vγ(βa):=(2-\Δa\)da.

Then V is a flat connection on ^Γ.

Proof. Clearly, (9 6 , <9c)-components of the curvature vanish because of (4.7a). It
remains to check that
V [ak ,y](2α) = [ V v V y ] ( a o ) . (4.28)

We have
[db,Y]xc = (\Δc\-2)δbc,

so that
[db,Y] = (\Δb\-2)db,

and
V[a6,κ](5α) = M\Δb\ - 2) (4-29)

On the other hand, in view of (4.27), (4.23),

so that

which agrees with (4.29).

544 M. Kontsevich, Yu. Manin

4.8. Horizontal Sections. Consider an even local section ψ = ^2Ψada of 3Γ. From
(4.27) we find

=Σ
a

a
Hence ψ is horizontal iff

V6,α, d 6 ^ α = -λ]Γ</Λ4^, (4.30)

c

Vα, yψ α + (2 - IΛJ)ψ α = 0 (4.31)

But

Replacing here the 9-derivatives of φ by the r.h.s. of (4.30), we get finally the
equation governing the λ-dependence of the horizontal sections:

c e \Δb\=2 e

+ (2-\Δa\)ψa=0,
or else

c \Λb\=2
a
+ {2-\Δa\)ψ = 0. (4.32)
Equation (4.32) has two singular points of which λ = 0 is regular, but λ = oc is
irregular one.
Therefore we make a formal Fourier transform (of 3f as a relative ^-module in
λ-direction)

αμ ό\

Denote the resulting ^-module ^ .

In order to describe it explicitly, it is convenient to pass from the language of
connections to that of quantum multiplication.
For any 7 e Hv, 7 = ΣxcΔc, put

Kin) := E ^Mί> - E<l AJ " 2)xMα = -Kv - %iri) + 2 7 , (4.33)

|4 b |=2 α

where <ξ is the Euler grading operator multiplying A by |/A|. Denote by B{η) the
operator of quantum multiplication by if (7) <zί ίΛ^ /?ί?mί 7 that is, in 3^.
We will identify B(η) with the matrix describing its action from the left to the
column vector (Λa), and consider it also as a matrix acting from the right on the row
of coordinates of a section (φa) of 3Γ. Similar convention applies to C
S.
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 545

v
4.9. Proposition. IfBQy) is semisimple in a subdomain ofΐί ^ is an isomonodromy
1
deformation of a holonomic ^-module on P with < dim H*(V)+ 1 regular singular
points parametrized by this subdomain.

Proof Equation (4.32) becomes in J^ :

r
e

On the other hand,

\Δb\=2 a c a

Comparing these expressions, one sees that the Fourier transform of (4.32) can be
written in the form
dφ 1
-^(B(Ί) - μE) = - φ^, (4.34)
oμ 2

where E is the identity matrix. One easily sees now that singularities of 17~ consist of
the spectrum of B{η) and infinity, and that they are regular when B{η) is semisimple.
This finishes the proof.
4.10. Schlessinger and Painlevέ Equations. If we want now to understand the analytic
properties of Φ, or rather the associated structure constants Acah, as a function of 7
using (4.34), we have to solve the following problems.
a) Find an appropriate point of Hv, where B(η) is computable and semisimple.
We will see in a moment that for Fano varieties 7 = 0 is the first choice.
Denote by M a realization of the moduli space of isomonodromy deformations of
the relevant ^-module constructed by Malgrange.
b) Calculate the (partial) map q : Hv —> M inducing Jf'.
This map can be considered as an analog, or a version, of the mirror map discovered
in the context of Calabi-Yau manifolds, especially if the following problem can be
solved affirmatively.
c) Identify (4.34) as an equation for the horizontal sections of the canonical connec-
tion of a variation of Hodge structure.
In the classical language, the compatibility conditions imposed upon Acah by (4.34)
are called Schlessinger equations.
2
When there are only 4 singularities (the only non-trivial example being V = P ),
the isomonodromy deformations are governed by the Painleve VI equation, which
is obtained from the corresponding Schlessinger equation by a change of variables
whenever the monodromy group can be reduced to SL(2). In this sense, the quantum
cohomology of P 2 furnishes a solution of this equation, whereas other V lead to
generalized Painleve equations.
4.11. Quantum Multiplication at 7 = 0. From the discussion in 4.3, 4.4, and 2.2, one
sees that for Fano manifolds the structure constants at 7 = 0 are defined by a finite
set of basic Gromov-Witten numbers:
ωΛβ
Aα6c(0) = ί Δa Λ Δb Λ Δc + £ e- * (I&ιβ)(Δa ®Δb® Δc), (4.35)
546 M. Kontsevich, Yu. Manin

where the summation is taken over β with

dim
(-Kv.β) = \ (\Δa\ + IΛI + IAJ) - c V < 2dimc V. (4.36)
n
In 5.3-5.4 below, we will calculate these coefficients for P and show that B(0) is
indeed semisimple with simple spectrum.

5. Examples

5.7. The Structure of Φ. In this section, we discuss in more detail some classes of
manifolds V. We always tacitly assume that at least tree level GW-classes for V
exist, and that the potential Φ constructed from them satisfies one of the convergence
hypotheses of 4.3.
We start with (4.17) and take into account Proposition 4.4 in order to drop
the redundant terms, in particular those that are no more than quadratic in 7. Our
conventions are:
a) ω = 0 (because this can be achieved by shifting 7).
imc
b) D = r + 1, Δo = ey, Δr+ι = ev = the dual class of a point.
X
We also assume that H (V) = 0 so that 2 < \Δa\ < 2dim c V-2 for a = 1,..., r.
a
The coordinates are renamed: 7 = xΔQ + Σ V ^a + ^ Ά +i
From the proof of Proposition 4.4 one knows that only the (β — 0, n — 3)-term in
(4.17) depends on x, and is | / 7 Λ 7 Λ 7 := \ (7 s ) [see (2.8)]. The (/? = 0, n > 3)-
terms all vanish because of the grading condition (4.20) combined with (2.8). So we
start with

x (C^X^Γ1 Θ ® Δf^)(yιΓ • • • (yΎrzn^ , (5.1)

and the grading condition for non-vanishing terms
r+l
Σ na(\Δa\ - 2) = 2(-Kv.β) + 2(dimc V - 3). (5.2)
0=1

1 ι
Dimension 1. Let V = P , r = 0, /3 = d[F ], d > 1. From (5.2) it follows that
(d > 2)-terms vanish. In view of the Divisor Axiom,

for all n > 3, and this must be 1, which is the number of the automorphisms of P 1
fixing three points. So finally we find a (conditional) answer:
5.1.1. Proposition.

Φ P V ° + ze2) = ± χ 2 z + ez-l-z-^. (5.4)

In the following calculations, we will be omitting quadratic, linear, and constant
terms of Φ without changing its notation.
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 547

2
Dimension 2. Here {Δx,..., Δr} form a basis of H (V); (5.2) is equivalent to
nr+ι=(-Kv.β)-l:=k(β)-l, (5.5)
so that we must sum only over β with k(β) > 1.
5.1.2. Proposition. For surfaces V we have up to terms no more than quadratic in 7
zk{β)-\

Σ^y3T)i^ ) (5 6 )
δ

Here for k{β) > 4 we put

(^Mβ)^,β)fir). (5.7)
For k(β) < 3, the definition of N(β) is given below: see (5.9).
Proof If (-Kv.β) > 4, then n r + 1 > 3 [see (5.5)] so that the contribution of β to
(5.1) in view of the Divisor Axiom takes form

z
Mβ)-ι _ i
- \)\ ^nx\.. .nr\

fc(/3)-l

For 1 < k{β) < 3 the calculation is only slightly longer. The actual contribution
of β is given by the same formula as the first expression in (5.8), but this time with
summation taken over n{ -f .. + nr > 4 — k(β). First, this sum lacks the terms of
total degree < 2 in y%£, but they are negligible. Second, in order to represent this
sum as the last expression in (5.8), we are bound to put

N(0) •= {1°>n

For fixed n^ the r.h.s. is well defined if (β.Δj) φ 0 for all i. One can secure this by
choosing an appropriate basis (eventually depending on β). The result does not depend
on ni. In fact, one can reach any point (nι,... nr) in the set Σ n% ^ 4 — k(β), ni > 0,
from any other point, without ever leaving this set, by adding and subtracting Γs from
coordinates. In view of the Divisor Axiom, these steps multiply the numerator and
the denominator of (5.9) by the same amount.
We expect that N(β) counts the number of rational curves in the homology class
β passing through k(β) — 1 points, at least in unobstructed problems.
Dimension 3. In this dimension, Calabi-Yau manifolds make their first appearance,
and we consider their potentials. Since Kv = 0, (5.2) shows that na -φ 0 only for
|Z\ α | = 2. Therefore, we may and will disregard the other elements of the basis of
H*(V), and in this subsection denote by {Ax,..., Δr} a basis of H2(V).
548 M. Kontsevich, Yu. Manin

5.1.3. Proposition. For a threefold V with Kv = 0, we have, up to terms of degree

< 2 in 7,
v ψ Ί)
Φ = - (Y) + > N(β)e , (5.10)

where
)(Δ nι
f
v
o n β

for any n = n{ + .. .+nr > 3 and any basis ofH2{V) such that (β.Δ^ φ Ofor all i.
The proof does not differ much from the previous one.
Let us guess the geometric meaning of N(β) £ Q restricting ourselves to the case
r = dim H2(V) = rkPic(F) = 1. Let Δx be the ample generator of H2(V), β0
the effective generator of JΪ 2 (V,Z) with (βo.Δλ) = 1, and /3 = d/30. If ΛΓ(d) e Z is
the "geometric" number of unparametrized rational curves in the class dβ0, then the
number of primitively parametrized curves with three marked points landing in β and
incident to three fixed cycles dual to Ax must be d?N(d).
According to [AM], the parametrizations of degree m with three marked points
must be counted with multiplicity m~ 3 . Hence we expect that

Σ ^ (5.12)
a
k/d

which can also be taken as a formal definition of numbers iV(fc) via GW-classes.
Miraculously, all classes appear to be integral. Rewriting (5.10) in this situation, we
get for y = y\
Σ (5.13)
k>\
yVίl
where Li3(z) = Σ ~r

Projective Spaces. Let V = P r + 1 , Δi = q ^ l ) ) * , | ^ | = 2i, r > 1. Put for

r+l
Σ na{a - 1) = (r + 2)d + r - 2 (this is (5.2)),
a=2

jV \f/λ Θ n l 6?) (5?) Λnr+1\

2
,... ,nr+1) = ,, (5.14)

where n = n t + . . . + n r + 1 > 3. The r.h.s. of (5.14) being independent on n 1 ?

for n 2 + . . . + n r + 1 > 3 one can take n1 = 0. Again, (5.14) must be the number
of rational curves of degree d in P r + 1 intersecting na hyperplanes of codimension
α, a = 2,... ,r + 1.
A version of previous calculations now gives:
5.1.4. Proposition.

r+1 .
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 549

5.2. Enumerative Predictions. We start with projective plane: (5.6) for r = 1. Put

φ(y^ z) = Φ (7) (7 ) = > . N(a) e y, iv(l) = I.

d=l

We have the following simple

v
5.2.1. Claim. The associativity condition for the potential Φ is equivalent to the
single equation

which in turn is equivalent to the recursive relation

( 5 17)
l(^tzΐ)Qizt)}
uniquely defining N(d) and φ.
This discovery made by M. Kontsevich was the starting point for this paper. The
first values of N(d), starting with d = 2, are 1,12,620,87304,26312976,14616808192.
From 3.1 and 3.1.1 it follows that a similar uniqueness result holds for any
projective space: in the notation of (5.14), (5.15), we have
5.2.2. Claim. The associativity relations together with the initial condition
JV(1; 0 , . . . , 0,2) = 1 uniquely define all N(d; n 2 , . . . , n r + 1 ) and the potential ΦpT .
Here, however, the compatibility of the associativity relations must be established
either geometrically (via a construction of GW-classes), or by number theoretic
and combinatorial means. We will now look at some of the identities implied by
associativity for del Pezzo surfaces.
Del Pezzo Surfaces. Let V = Vr be a del Pezzo surface that can be obtained by
blowing up 0 < r < 8 points (in sufficiently general position) of P 2 . A choice of
such a representation π : Vr —>• P 2 allows one to identify Pic(y r ) with Z r + 1 via

L = aΛ — b{l{ — . . . - brlr *-> (α, bλ,... br),

where A — π*(c1(<^)(l))) and li — inverse image of the i-th blown point. Under this
identification, the intersection index becomes ((α,fc^.ία',^)) = ao! — Σ\b'%, and
—K v — (3; 1,..., 1) so that (—KV.L) = 3a - ]Π bί. The cone of effective classes B
is generated by its indecomposable elements A for r = 0, A — lx and lx for r = 1,
and all exceptional classes for r > 2. (Recall that I is exceptional iff (I2) = — 1 and
(—Kv.β) = 1; for more details see [Mai].)
This allows us to rewrite (5.6) as an explicit sum over B.
r

Writing 7 = xe° + yΛ + ze - Σ y1!^ ΦVr(η) = - (7 3 ) + ^(7), we can easily

A

check the following generalization of 5.2.1:

5.2.3. Claim, a) One of the associativity relations reads
r r
2 \~~^ 2 \~~^
Ψzzz ~ Ψyyz ~~ / j LPyyιz ~ ^yyy^Pyzz ' / j Ψyyy%Ψyxzz '
550 M. Kontsevich, Yu. Manin

b) This relation is equivalent to the following recursive formula for the coefficients
N(β) [see (5.6)7:

N(β)= Σ N(βι)N(β2)(βι.β2)(Λ.βι)

'(-Kv.β)-4\ ίia
J(-Kv.β)-A

The initial conditions for (5.18) consist of the list of values of N(β) for all
indecomposable elements of B and β's with (—Kv.β) < 3.
The redundancy of the associativity relations is reflected here in the presence of
A which depends on the choice of π : Vr —> P 2 . The number cr of such choices for
r = 1,...,8 is respectively 1,1,2,5,2 4 ,2 3 .3 2 ,2 6 .3 2 ,2 7 .3 3 .5. In fact, the symmetry
group Wr of the configuration of exceptional classes acts upon the set of associativity
relations, and cr = |W r |/r!, the denominator corresponding to the renumberings of
blown points.
Question. Is it true that the linear span of all relations (5.18) for various choices
of π contains all the associativity relations, at least for larger values of r?
5.2.4. Quadric. The quadric V = P 1 x P 1 is the last del Pezzo surface. Here
Pic(F) = Z 2 , — Kv = (2,2), and all associativity relations were written explicitly in
[I]. In self-explanatory notation
2α+26-l 2

and the associativity relations together with initial conditions -/V(0,1) = iV(l,0) = 1
imply the following recursive definition of N(a, b) in the effective cone a > 0, b > 0 :

N(a, b)= Σ N(
<ai» b0N(a2i ^ X α A + α2^i)δ2

2a + 2b - 4 \ ( 2a + 2b - 4
ax'

The remaining relations are:

N(a,b) = N(b,a), (5.20)
v-^ ( 2a-\-2b — 3 \ ? 9
2abN(a,b) =
X 7 /
> A^(αl,61)A/"(α9,69)ί / v
Λ7 , αί(αiθ 9 — aώΛbn, (5.21)
/ V / 7 V
/ j ^ i " l Z ' Z l ' Λ _ . I O L 1 / 1 1 Z Z 1 Z ^
04+0,2=0,

aN(a,b)= Σ Nia^b^Nia^bJ^^^λa^bl-^al), (5.22)

x a\[{a2 + b2-
Question. Can one deduce (5.20)-(5.23) directly from (5.19)?
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 551

5.2.6. Nonsingular Rational Curves. Consider an effective class β with

pa(β) : = (β.β + Kv)/2 + 1 = 0 , i.e.
(-Kv.β) = d>0, (β.β) = d-2.
Any irreducible curve in this class must be nonsingular rational, so that passing
through points imposes only linear conditions. Thus we may expect that N(β) — 1
for such a class. This was observed numerically on cubic surfaces V6 by C. Itzykson.
Question. Can one deduce directly from (5.19) that N(β) — 1 whenever pa(β) = 0?
Notice that there are infinitely many such classes on any Vr with r > 1. The
simplest family is: r = 1, β = nΛ — (n — \)lx projecting into rational curves of degree
n with one (n - l)-ple point on P 2 .
5.5. Quantum Multiplication in i ί * ( P n ) at 7 = 0. Choose as above Δa = c 1 ( ^ ( l ) ) α ,
\Δa\ = 2a, 0 < a < n. Calculate Aabc with the help of (4.35). (We now drop
the restriction ω = 0. Equivalently, we can say that we calculate the quantum
multiplication with ω — 0 but at all points of the subspace H2 C H*.) The
β Φ 0 terms in (4.35) do not vanish only for β = class of a line, 1 < a,b,c <
n, α + 6 + c = n + l . ?ut q = e~fωAβ. The compatibility of WDVV-equations
implies that (I0^β)(Δa<g> Δb<g> Δc) = 1 in this range; this agrees also with geometric
interpretation. Finally, gab = δa+b n. Putting this all together, we obtain:
5.3.1. Proposition. (Z\ 1 ) oα = Δa for 0 < a < n, and qΔ0 for a = n + 1.
Since Δo is the identity with respect to quantum multiplication, we see that
iJ* (Pn)| — C[x]/(x n + 1 — q) (5.24)
n
This formula was heuristically obtained for P in many papers, and was generalized
by Batyrev for toric varieties, and by Givental and Kim for flag spaces.
We now see however, that (5.24) and these generalizations describe only a subspace
of quantum deformations parametrized by H2.
5.4. ^-Module 3^ at 7 = 0. We can now easily write for V = P n , Eq. (4.34) at
7 = 0. In fact, the matrix B(0) describes the quantum multiplication by — Kv —
(n + l)Δx in the basis (Δa). Therefore (4.34) reads:
/ -μ 0 0 ... 0 (n+l)<?\ /dφ°/dμ\ ( 0 \
n+1 -μ 0 0 0 dφι/dμ

\ 0 0 0 ... n + 1 ~μ \dφn/dμ/ \nφn

The finite singular points are (n + l)qι/n.

6. Cohomological Field Theory

6.1. Definition. A two-dimensional cohomological field theory (CohFT) with coef-
ficient field K consists of the following data:
a) A iί-linear superspace (of fields) A, endowed with an even non-degenerate
pairing.
b) A family of even linear maps (correlators)
n
9 > '
H*(M.n,K),
defined for all g > 0 and n + 2g - 3 > 0.
552 M. Kontsevich, Yu. Manin

These data must satisfy the following axioms:

6.7.2. Sn-Covariance.
6.1.2. Splitting. In order to state this axiom, we retain all notation of 2.2.6 with the
ab
following minimal changes: {Δa} denotes a basis of A, Δ = "Σlg Δa ® Δb is the
Casimir element of the pairing, and all mention of V and /3's is omitted. Then the
axiom reads:

Ίj
® Δa )gab ® /„,„,,, I Δb ® [ (X) 7 7 ]) • (6.1)
a,b

6.1.3. Genus Reduction. This is (2.12), with V and β omitted.

6.2. Remarks, a) We will mostly assume A finite-dimensional. However, A can also
be graded with finite-dimensional components, or an object of a if-linear tensor
category, etc.
b) We will be mostly concerned with tree level CohFT. Correlators for such a theory
must be given only for g = 0, n > 3, and the Genus Reduction Axiom is irrelevant.
6.3. Example: GW-Theories. Any system of GW-classes for a manifold V satisfying
appropriate convergence assumptions gives rise to the following cohomological field
theory depending on a class ω as in 4.3: A = H*(V, C) with Poincare pairing,

The series in β can also be treated as a formal one. Equivalently, we can put
A = 0 77*(V, if) and work with 73-graded objects.
βeB

6.4. Operations on Field Theories, a) Tensor product of two cohomological field

theories can be defined as in 2.5.
b) If a group G acts upon the space of fields A of a tree level CohFT preserving
scalar product and / n , a new tree level CohFT can be obtained by restricting all
maps to the space of invariants AG.
c) Following Witten ([W]), one can define an infinite-dimensional family of defor-
mations of any CohFT. The parameter space will be the formal neighborhood of zero
in the vector superspace A ® C[[x]], where x is an even formal variable. The space
of fields A and its scalar product are kept undeformed.
oo
For a point a = Σaixι, az e A, of the parameter space, define the new correla-
tors by °

fc=0

>VΛ Λc^:+J^
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 553

Here xi denotes the i-th marked point of the universal curve C, πk is the projection
Mg n+k —> Mg n forgetting the last k points, π^ is the direct image in cohomology
for the proper map π of smooth stacks (orbifolds).
6.5. Mirror Symmetry. Physicists believe that to each Calabi-Yau manifold V of
sufficiently big Kahler volume one can associate two different cohomological field
theories, called A- and B-models in [W].
A-model depends only on the cohomology class [ω] of the symplectic form and
remains invariant when one deforms the complex structure of V.
In the infinite volume limit A-model should approximate the GW-model from
6.3, taking its existence for granted. One expects (see [BCOV]) that the difference
between any correlators in A- and GW-models is a cohomology class of a moduli
space Poincare dual to a homology class supported on the boundary.
B-model should depend only on the complex structure of V via the universal
infinitesimal variation of its Hodge structure and, possibly, on some additional data.
The space of fields in the B-model must be 0 Hq(/\V3^) with grading (p+q) mod 2.
VΆ
The quantum multiplication in the computed examples is given by the symbol of
iterated canonical connection on the cohomology space whose definition requires a
choice of the global volume form on V.
Mirror symmetry ought interchange A- and B-models of dual varieties.
It would be important to have a treatment of A- and B-models axiomatizing their
dependence on the geometry of V.
In the remaining part of this section, we introduce an operadic formalism for
description of tree level CohFT's. Our framework is similar to that of [V, BG, GiK].
6.6. Trees. We will formally introduce trees describing combinatorial structure of a
marked stable curve of arithmetical genus 0. Their vertices correspond to components,
and edges to special points.
6.6.1. Definition. A (stable) tree τ is a collection of finite sets VT (vertices), Eτ
(interior edges), Tr (exterior edges, or tails), and two boundary maps b : Tτ —> Vτ
(every tail has one end vertex), and b : Eτ —> {unordered pairs of distinct vertices}
(every interior edge has exactly two vertices).
The geometric realization of r must be connected and simply connected. Every
vertex must belong to at least three edges, exterior and/or interior (stability).
6.6.2. Definition. A morphism of trees f : r —> σ is a collection of three maps (notice
arrow directions)

v v
Jv ' τ σ >J ' -1-σ ±
τ^ J ' -^σ ^r ?

with the following properties:

a) fv is surjective, fι and fe are injective.
b) Ifvγ,v2 are ends of an edge e' ofr, then either fv(v{) = fυ(v2)i or fv(vτ) are ends
of an edge e" of σ; we say that e' covers this edge, and we must then have e' = fe(e").
c)Ifvfe VT is such a vertex that fυ(υ') is the end oft" e Γ r , then v' is the end of
f\t").
In other words, / contracts interior edges from Er \ fe(Eσ) and tails from
Tτ \ /*(Γ σ ), and is one-to-one on the remaining edges. We will denote by f(e)
the image of a non-contracted edge.
554 M. Kontsevich, Yu. Manin

The composition of morphisms is the composition of maps. In this way, trees form
a category.
6.6.3. Flags and Dimension. A pair {edge, one end of it} is called a flag. For a tree
r, we denote by Fτ the set of its flags, and by Fτ(υ) the set of flags ending in vertex
v. We have \FT\ = 2\ET\ + \Tr\.
The dimension of r is defined by
3
dim τ:=Σ ( l * » l - ) = ΆK\ + \TT\ ~ 3\VT\. (6.3)

5.4. Glueing. Let (T^,^), i = 1,2, be two pairs consisting each of a tree and its
tail. Their glueing (tλ to £2) produces a pair (r, e) consisting of a tree and its interior
edge:
(r,e) ^ ( / η , ^ ) * ^ , ^ ) .
Formally:

T τ = (T τ i J j T T 2 ) \ {ί,,ί 2 }, 6(e) = {b(tx),b{t2)} •

This operation is functorial in the following sense: for two morphisms / t : τ t —> σ4
not contracting ί i ; we have a self explanatory morphism

h*h (T-i.ti) * (72, ί 2 ) - " (^i,/i(ii)) * (σ 2 ,/ 2 (ί 2 )).

Finally, F τ = F T ) ]J ^
6.7. Products of Families. Assume that we work in a monoidal category where
products of families of objects are associative and commutative in a functorial way.
Then we can use a convenient formalism (spelled out e.g., by Deligne) of products
of families indexed by finite sets and functorial with respect to the maps of such sets.
In this sense, we will use notation like f| Az, (^) A%1 A®F (for a constant family
Ai = A), etc. it*1 ^F
In the same vein, if | F | > 3, we will denote by M0F the moduli space of stable
curves of genus zero with \F\ marked points indexed by F.
6.8. From Trees to Moduli Spaces. In this subsection, we define a functor
^/S : {trees} —• {algebraic manifolds}
(ground field is arbitrary).
6.8.1. Objects. Put

vevr
We have dim ^M(τ) = dim r.
This space parametrizes a family of (generally reducible) stable rational curves
C(r) with marked points indexed by Tτ. The dual graph of a generic (but not arbitrary)
curve of this family is (canonically identified with) r. To describe it, consider a point
x — (xv) e ^#(τ), xv e M0^Fτ(v), and let C(xυ) be the fiber of a universal curve at
this point. If υι,v2 bound an edge e of r, C(xv) contains a point y{vi^ e) marked by
the flag (υ^e). Identify y(yx,e) with y(v2,e) in the disjoint union \J C(xυ) for all
e. This will be C(τ)(x). ^Fr
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 555

Clearly, its remaining special points are marked by Tτ so that we have a canonical
morphism (closed embedding) J%(τ) —> M o T r . This is a special case of morphisms
defined below.
6.8.2. Morphisms. Any morphism of trees / : r —•» σ contracting no tails induces a
closed embedding yM(τ) —> J%>(σ). To construct it, identify Tτ = Tσ = T by means
of /*, and denote by p the one-vertex tree with tails T. Clearly, ^#(p) = M o Γ ,
and by universality, we have embeddings of ^#(σ) and ^S(τ) into ^#(p). In this
embedding, ,M>{σ) C ^ # ( τ ) which is the seeked for moφhism.
Any moφhism of one-vertex trees contracting tails induces the forgetful morphism
of the respective moduli spaces (see e.g. [Ke]).
The general construction of a moduli space morphism corresponding to a morphism
of trees can be obtained by combining these two cases: embed <J£(τ) into M0Tτ,
J%>{σ) into M o τ , and restrict the forgetful map onto ^#(τ).
6.8.3. Glueing. If (τ,e) = ( r ^ ^ ) * (τ 2 ,t 2 ), we have canonically (i/*^#( ) :=

^(r1)x^(τ2),
( j
E\M(τ) = H*y£(τ{) ® H\Jζ{τ2). '
6.9. From Γrees ίo Tensors. Let A be a linear superspace over K with a Casimir
element Z\, as in 6.1.
6.9.7. Objects. For a tree r, put ^?(r) = A®Fτ. We will show that this construction
is (contravariant) functorial with respect to pure contractions that is, morphisms of
trees contracting no tails.
6.9.2. Morphisms. Let / : r —» σ be a pure contraction, F^ the set of non-contracted
flags, and E^ the set of contracted edges. Each such edge gives rise to a pair of
contracted flags. Therefore, we have two identifications:

® (A ® ^ ) E -
the second being defined only up to switches in A <g> A factors. Since the Casimir
element is invariant with respect to the switch, we can unambiguously set

(6.6)

6.9.3. Glueing. Obviously, for (r, e) as in 6.8.3, we have canonically

^(r) = ^(r1)(8)^(r2). (6.7)
We see that with respect to pure contractions and glueing, ,sβ has the same
properties as H*^M from 6.8.
We can now state our new definition.
6.10. Definition. An operadic tree level CohFT is a morphism of functors compatible
with glueing
I: Λ->H*J%.
In other words, it consists of a family of maps indexed by trees
556 M. Kontsevich, Yu. Manin

such that for any pure contraction f : τ —>• σ we have:

and for any τ glued from τx,τ2>

I(τ) = / ( T J ) <8> I(τ2): ^(T) — ~&{j{) <S> ^ ( τ 2 ) —> H*.sM(τ{) <g) ϋ " * ^ # ( τ 2 ) . (6.9)

<5.ii. Claim. The two definitions of a tree level CohFT are equivalent.
Proof (Sketch). Notice that a two-vertex tree with tails { 1 , . . . , n} is the same as
a partition 5 = (Sι,S2) of {1,... ,n}. Now, given an operadic tree level CohFT,
restrict it to the following subclass of trees: Tτ = {1,... ,n} for some n, \VT\ = 1
or 2. One easily checks that morphisms φs from 2.2.6 are induced by non-trivial
pure contractions in this subclass, and that (6.7) restricted to it becomes (6.1), and
5 n -covariance corresponds to the functorality with respect to bijective maps of tails.
Conversely, given a tree level CohFT in the sense of 6.1, we first rewrite it
as a fragment of an operadic CohFT as above, and then reconstruct the whole
operadic CohFT using glueing and decomposition of pure contractions into products
of morphisms contracting exactly one edge each.

7. Homology of Moduli Spaces

7.1. Additive Generators. If Tτ = { 1 , . . . , n}, we will call r an n-tree. A morphism

of n-trees r —» σ identical on tails will be called n-morphism. If such a moφhism
exists, it is unique. Let ρn be a one-vertex n-tree. Then ^S{pn) = M o . For any
n-tree r, there exists a unique n-contraction r —> pn. Let dτ e H^(MQ ) be the
homology class of ^ # ( r ) corresponding to this contraction. It depends only on the
n-isomoφhism class of r. The manifolds ^ # ( r ) embedded into each other in this
way will be called strata.
Lemma 7.1.1. dτ span H^(MOn) (over any coefficient ring).
Proof. Easy induction by n, as in the proof of 2.5.2.
7.2. Linear Relations. Choose a system R = (r, {i,j, fe, Z}, v), where r is an n-tree,
1 < 2, j , £;, Z < n are its pairwise distinct tails, and v G Vr is such a vertex that paths
from v to z, j ,fc,Z start with pairwise distinct edges e ι ? e^ , efe, e ; respectively (some of
these edges may be tails themselves).
Consider all n-contractions τf —> r which contract exactly one edge onto the vertex
1
υ and satisfy the following condition: lifts to τ of e ΐ ? e^ on the one hand, and efc, ez
on the other, are incident to different vertices of the contracted edge. Below we will
denote by {ijr'kl} the summation over n-isomoφhism classes of such contractions,
R being fixed.
7.2.1. Lemma. For any R, we have

/ a^f = > α_// (/.IJD

Proof Consider a moφhism of r contracting all edges and tails except of i, j , fc, Z. It
induces the forgetful moφhism ^€(τ) —> M o r ϊ7 fcn = P 1 . Two fibers over boundary
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 557

1
divisors of the latter moduli space are represented by the cycles Σ ^(j ) and
X) J&(τ") respectively. {τjτ'ki}
{ιkτ"3l}

7.3. Theorem. Relations (7A)R span the space of all linear relations between dT.
7.3.1. Lemma [Ke]. As an algebra, H* := H*(M0 ) is generated by the boundary
divisorial cohomology classes Ds indexed by unordered partitions S o/{l,..., n} into
two parts Sx, S2 of cardinality > 2 and satisfying the following generating relations:

D D
Σ s= Σ τ v Vvki
{ijSkl} {ikTjl}

and
DSDT =0 (7.3)

if four sets Sz Π Tj are pairwise distinct and non-empty. (In this case we will call S
and T incompatible.)
In 3.2.2, Ds were denoted ds whereas here we reserve lower case letters for
homology classes. Classes Ds are dual to the homology classes dσ, where σ run over
n-trees with two vertices, and (7.2) is a consequence of (7.1).
Denote now by H^ the linear space generated by the symbols [dσ] subject to all
relations (ΊΛ)R, where σ, r run over all n-isomorphism classes of n-trees.
There is an obvious surjective map a : H^ —• H* :

a([dσ]) := the cohomology class dual to dσ . (7.4)

7.3.2. Main Lemma. H* can be endowed with a structure of cyclic H* -module

generated by [dpn] := 1 so that the map

b: H* ->H^ b(h) = h 1 (7.5)

is surjective.
Comparing (7.4) and (7.5) we see that dimfί^ = dim if* and both α and b are
linear isomorphisms. Theorem 7.3 follows.
Proof of 7.3.2 (Sketch). Every interior edge e of an n-tree σ determines a partition
5(r, e) of its tails. Put

The stratum ^ # ( r ) is the transversal intersection of pairwise compatible boundary

divisors in ^

Therefore we devise the action of ϋ P upon H^ in such a way that

DsΛ = [dτ] (7.6)

making obvious the surjectivity of (7.5). If we now want to define any product
Ds.[dτ], (7.6) forces us to consider three cases.
558 M. Kontsevich, Yu. Manin

Case 1. S is incompatible with some T G S*(σ). Then we put

Ds.[dσ] = 0. (7.7)
Case 2. S is compatible with all T e S^{σ) but S £ ¥'{σ). Then we check that
y{σ) U {5} = 9\τ) for a unique r, and put
D5.[dσ] =[dτ]. (7.8)
Case 3. S G . ^ ( T ) . This case concentrates all the difficulties. Let us start with a two-
vertex σ, y{σ) — {S}. Choose i,j G 5 1 ?fc,ZG 5 2 and apply D 5 to the relation
between homology classes dual to (7.2) fez. We are forced to put

(7 9 )
ΣX],
w h e r e r r u n s o v e r a l l t r e e s w i t h 9 " { j ) = { S , S f } , Sf ^ S, i , j e S[, k , l G S'2, a n d
then to check that, modulo postulated relations, the r.h.s. of (7.8) does not depend on
the choice of i,j,k,l.
Now, consider the r.h.s. of (7.9) as a divisorial cohomology class in H*(yS(σ)) =
H*(Mo^s)®H*(MOβ). It can be naturally represented as a sum of two components
Dx ® 1 + 1 <g> D2. If dτ is represented by ^ # ( r ) C ^ ( σ ) , we have (r, e) =
( T J , ^ ) * (T2,t2)> where e is the lift of the unique edge of σ to r. The action of
Z^ upon d r is assumed to be inductively defined, which determines the action of Ds
upon dτ.
It remains to check that these prescriptions are compatible with (7.1)—(7.3).
This is a tedious but straightforward verification which we omit.

8. Second Reconstruction Theorem

8.1. Definition. An abstract tree level system of correlation functions (ACF) over a
coefficient field K consists of a pair (A, Δ) as in Defi 6.1 and a family of even linear
maps
Yn : A®n -> K, n>3 (8.1)
satisfying the following axioms:
8.1.1. Sn-lnvariance.
8.1.2. Coherence. In notation of 6.1 and (3.3) it reads: for any pairwise distinct
1 < i, jf, A;, I < n,

Σ ) ( (
a,b \reSι ) \ \seS2

= Σ{ikTjl} a,b
( \rETi

8.2. Example. For a tree level CohFT / = (A, A, / 0 n ) , put

(^o,n>(7i ® ® 7 n ) '•= f io >n (7i ® • 0 7 n )

M
0,n
Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 559

These polynomials are called correlation functions of /. Their ^-symmetry is

obvious, and (8.2) follows from the Splitting Axiom 6.1 in the same way as (3.3) in
the context of GW-classes. The following main result of this section shows that these
examples essentially exhaust ACFs.
8.3. Theorem. Any tree level ACF consists of correlation functions of a unique tree
level CohFT.
8.3.1. Remark. Starting with any ACF, one can construct a potential

n>3

and check that the differential equations (4.13) are equivalent to the coherence
relations (8.2). In this sense, WDVV equations are equivalent to tree level CohFTs.
We start a proof of 8.3 with some preliminaries.
8.4. Correlation Functions on Trees. In this context, we will be considering tensors
jg>(τ) := A®Tτ (8.4)
Fr
rather than J&{j) — A® . This construction is obviously functorial with respect to
pure contractions. More important is its behaviour with respect to glueing and cutting.
If (r, e) = (TJ, t{) * (τ 2 , t 2 ), we say that (τt,tτ) are obtained by cutting τ across e.
One can easily generalize this notion for any subset of edges E C ET instead of {e}.
For instance, cutting r across all edges results in a set of one-vertex trees, stars of
vertices υ G Vτ. Formally, a star p(υ,τ) has υ as its vertex and Fτ(v) as its tails.
Let τ\E be the set of trees obtained from r by cutting it across all e G E. Tails of
any σ G τ\E consist of some tails of r and some "half-edges" of r, each edge in E
giving rise to two tails of the latter type. Therefore, we can construct a well defined
map

JΘ(E) : JB{τ) = A®Tτ -> ( g ) J8(σ) = A®{]1Tσ) 9* A®Tτ ®(A® A)E (8.5)
σer\E

which tensor multiplies any element of J?(τ) by ZAΘJ£; G (A 0 A)^. (Compare this to
(6.6).)
8.4.1. Lemma. For any system of Sn-symmetric polynomials Yn : A®n —>• K, there
exists a unique extension to trees
Y(r) : Mr) -> AT
with the following properties:
a) If pn is one-vertex tree with tails {1,..., n}, ί/i£ft V(p n ) = Y^
b) For α^j r and any E C Er, we have

Y(τ)= ( (g) Y(σ)\ ojg{E). (8.6)

\σ£τ\E /

c) y ( r ) are compatible with tree isomorphisms.

Proof. Put

Y(r):= 0 y(j)o^(£ τ )=0y(p(ί; J τ))oi(ί! τ ) ) (8.7)

σeτ|£τ
560 M. Kontsevich, Yu. Manin

and Y(p(υ,τ)) = Y\Fτ(V)\> Then a) follows by definition from the ^-symmetry of

Yn, and (8.6) becomes a corollary of the associativity of tensor products.
As an example of (8.6), let r = τs be a tree with vertices vuv2 and tails Si ending
at i-th vertex. Put Y^ = Y(p(υi,τ)). Cutting r across its edge we get

ses2

seS2
We turn now to correlation functions.
8.5. Lemma. Let {Yn} be correlation functions of a CohFT /, {Y(τ)} their extension
to trees which we will call operadic correlation functions.
For a tree r, denote by f : r —>• p a maximal pure contraction identical on tails,
and by φ : ^M{τ) —» ^M(p) the corresponding embedding. Then

Y(r)= j φ*(I(j>)). (8.9)

Proof From (6.4), we know that

Write the relation (6.8) for / taking in account the following identifications:

vevτ

Thus, applying in addition (8.7), we see that the following two functions A®Tτ —> K
coincide:

j I(τ) o Λ?(/) = ί (g) J I{p(v, τ))\ τ ) = Y(τ)

veVτ
.Mir)
Mir) \ .Mρ(v,τ
.Mρ(v,τ)) J
and
J H\JZ(f)oI(p)= J φ\I

8.5.1. Corollary. Let fa: τa -* p be a family of maximal pure contractions defining

strata ,^Cα in <J&(p) = Mo n whose homology classes da satisfy a linear relation

a
IfY(r) are operadic correlation functions of a CohFT /, they satisfy the identity

ΣaaY(τa) = 0. (8.10)

We will say that (8.10) is correlated with R.

Gromov-Witten Classes, Quantum Cohomology, and Enumerative Geometry 561

For example, the Coherence Axiom 8.1.2 together with (8.6) means that ACF
satisfy all equations correlated with Keel's linear relations between boundary divisors.
The central observation is that this implies the following stronger statement:
8.6. Lemma. Any ACF satisfies all the equations correlated with linear relations
between strata homology classes.
Proof. Clearly, it suffices to treat relations ( 7 . 1 ) ^ , where R = ( r , {i, j , k, l}^v) as
in (7.2). Consider the star p — ρ(v,τ) and its four tails i.j,kJ corresponding to
ee q,ej,e
q
,ej,ek,et. Write the Keel relation in H^(.M(p(υ.τ)) and the correlated identity
(8.10):

{Γjp'kϊ} {ϊkpffjl\

where the sums are taken over two-vertex trees with tails F .
There is a natural bijection between summands in (8.11) and (ΊΛ)R.
Denote by E the set of all interior edges of r incident to v excepting e i ? e , ek, eι.
Cut r across these edges, and denote by T the set of resulting trees excepting the star
of v. According to (8.6), for the terms of l.h.s. in (8.11) and (1 Λ)R corresponding to
each other, we have

Y(τ') = Yip') ® (®) Y(σ) o .

\σeτj
and similarly for the r.h.s.
Hence from (8.11) it follows that

Y(τ')=
{ijτ'kl} {ikτ"jl}

This identity is correlated with (JΛ)R.

8.7. Proof of Theorem 8.3. Start with an ACF Yn. It suffices to reconstruct an
"economy class" CohFT defined by 6.1 rather than the full fledged operadic one.
Therefore, in this section, as in 7.1, we may and will consider only n-trees and
n-contractions.
First, construct I(pn) — / O r ? . From (8.9) we know the integrals of I(pn) over
all tree strata, whose homology classes generate H^(.y/^(pn))^ and by Lemma 8.6,
these integrals extend to a linear functional on H^(../€{pn)). By Poincare duality, this
uniquely defines 70 n.
Second, check (6.1). It suffices to verify that integrals of both sides over any
stratum . /€{τ) contained in the relevant boundary divisor .J£(js) coincide. But this
is a particular case of (8.6). In fact, if r = r^, this is exactly (8.8). Generally, one
can assume that (r, e) = ( η , ^ ) * (τ2,£2)> TT] = Sγ U { ί j , TT2 = S2 U {t2}, and
apply (8.6) to this glueing.
In the context of GW-classes, we can now more or less formally deduce the
following version of the Reconstruction Theorem 8.3.
8.8. Theorem. Assume that for a manifold V, a system of maps

is given satisfying all the conditions that are imposed by Axioms 2.2.0-2.2.6 on the
tree level codimension zero GW-classes.
562 M. Kontsevich, Yu. Manin

Then there exists a unique tree level system of GW-classes I^n β such that
v v
(I )= Y
To prove it, we first reconstruct the relevant CohFT, and then check the Axioms
involving the geometry of V. We leave the details to an interested reader.
Finally, results announced in [RT] give codimension zero tree level GW-classes
and therefore all tree level GW-classes for semi-positive symplectic manifolds.

References

[AM] Aspinwall, P.S., Morrison, D.R.: Topological field theory and rational curves. Preprint
OUTP-91-32P, 1991.
[BCOV] Bershadsky, M., Cecotti, S., Ooguri, H., Vafa, C : Kodaira-Spencer theory of gravity and
exact results for quantum string amplitudes. Commun. Math. Phys. (to appear)
[BG] Beilinson, A., Ginzburg, V.: Infinitesimal structure of moduli spaces of G-bundles. Int.
Math. Res. Notices, Duke Math. J. 4, 63-74 (1992)
[D] Dubrovin, B.: Integrable systems in topological field theory. Nucl. Phys. B379, 627-689
(1992)
[GK] Getzler, E., Kaprano, M.M.: Cyclic operads and cyclic homology. Preprint, 1994
[GiK] Ginzburg, V.A., Kapranov, M.M.: Koszul duality for operads. Preprint, 1993
[I] Itzykson, C : Counting rational curves on rational surfaces. Preprint Saclay T94/001
[Ke] Keel, S.: Intersection theory of moduli spaces of stable n-pointed curves of genus zero.
Trans. AMS 330, 545-574 (1992)
[Ko] Kontsevich, M.: A^-algebras in mirror symmetry. Bonn MPI Arbeitstagung talk, 1993
[Mai] Manin, Yu.: Cubic forms: algebra, geometry, arithmetic. Amsterdam: North Holland, 1974
[Ma2] Manin, Yu.: Problems on rational points and rational curves on algebraic varieties. To be
published in Surveys of Diff. Geometry
[R] Ruan, Y.: Topological sigma model and Donaldson type invariants in Gromov theory.
Preprint MPI, 1992
[RT] Ruan, Y., Tian, G.: Mathematical theory of quantum cohomology. Preprint, 1993
[V] Voronov, A.: Topological field theories, string backgrounds, and homotopy algebras.
Preprint, 1993
[W] Witten, E.: Two-dimensional gravity and intersection theory on moduli space. Surveys in
Diff. Geom. 1, 243-310 (1991)
[Y] Yau, S.-T.: ed. Essays on mirror manifolds. Hong Kong: International Press Co., 1992.

Communicated by A. Jaffe