, ,

How to construct irreducible holomorphic symplectic manifolds? Except for the variety of Beauville and Donagi [4] all known examples arise from moduli spaces of semistable sheaves on a K3 or abelian surface.

For every element

v

in the Mukai lattice

H^{even} (X, Z)

of a polarised K3 or abelian surface

(X, H)

there is an associated moduli space

M_{v}

that parametrises polystable sheaves

E

with Mukai vector

v = v (E) : = c h (E) \sqrt{td (X)}

. If

H

and

v

are chosen to the effect that no strictly semistable sheaves exist, i.e. every semistable sheaf is automatically stable, then

M_{v}

is a projective holomorphically symplectic manifold due to Mukai [21] .

In the opposite case,

M_{v}

is singular and one may ask whether

M_{v}

at least admits a projective symplectic resolution. This question has been raised and successfully answered in two cases by O’Grady [23, 24] , leading to two new deformation classes of irreducible holomorphic symplectic manifolds.

In this paper we give a complete answer to O’Grady’s question for general ample divisors

H

and moduli spaces whose expected dimension

2 + 〈 v, v 〉

\geq 4

. The answer depends essentially only on the divisibility of the Mukai vector

v \in H^{even} (X, Z)

and the dimension of the moduli space. We may write

v = m v_{0}

with a primitive Mukai vector

v_{0} = (r, c, a)

and a multiplicity

m \in N

. Suppose for simplicity that

r > 0

, and let

H

denote a

v

–general ample divisor. Then every semistable sheaf

E

with Mukai vector

v (E) = v_{0}

is stable, and a necessary and sufficient condition for the existence of

E

is that

c \in NS (X)

and that

〈 v_{0}, v_{0} 〉 \geq - 2

There are five principal cases to distinguish:

1) If

〈 v_{0}, v_{0} 〉 = - 2

, then Mukai has shown that

M_{v_{0}}

consists of a single point

[E_{0}]

only. As the expected dimension of the moduli space

M_{m v_{0}}

is negative for

m > 0

there are no stable sheaves in this case, and it follows by induction that any semistable sheaf must be of the form

E_{0}^{\oplus m}

. Hence

M_{v}

is a single point as well.

2) If

〈 v_{0}, v_{0} 〉 = 0

, the moduli space

M_{v_{0}}

is again a K3 surface or an abelian surface if

X

is K3 or abelian due to beautiful results of Mukai. It turns out that any semistable sheaf

E

with

v (E) = m v_{0}

S

-equivalent to a direct sum

E = E_{1} \oplus \dots \oplus E_{m}

with stable sheaves

[E_{i}] \in M_{v_{0}}

. It follows that

M_{v} = S^{m} (M_{v_{0}})

. Thus the moduli spaces are singular in codimension 2, but admit symplectic resolutions in terms of the Hilbert scheme

{Hilb}^{m} (M_{v_{0}}) \to M_{v}

3) Assume now that

〈 v_{0}, v_{0} 〉 \geq 2

. Due to the combined efforts of many authors, with important steps taken by Mukai, Huybrechts, O’Grady and Yoshioka, one finally has the following result [27] :

M_{v_{0}}

is a smooth symplectic variety that is deformation equivalent to

{Hilb}^{1 + \frac{1}{2} 〈 v_{0}, v_{0} 〉} (X)

, if

X

is a K3-surface, and to

{Pic}_{0} (X) \times {Hilb}^{\frac{1}{2} 〈 v_{0}, v_{0} 〉}

, if

X

is an abelian surface.

Assume in addition that

m \geq 2

. The main result of this article implies that one has to further distinguish the following two cases: 4) Let

〈 v_{0}, v_{0} 〉 = 2

and

m = 2

. The moduli spaces

M_{K 3} (2; 0, 4)

and

M_{A b} (2; 0, 2)

studied by O’Grady [23, 24] and Rapagnetta [25] fall into this class. The moduli space

M_{v}

has dimension

10

, its singular locus has codimension 2 and is in fact isomorphic to

S^{2} M_{v_{0}}

. As shown in [18] , the symplectic desingularisations constructed by O’Grady exist for all Mukai vectors in this class and can be obtained by blowing-up the reduced singular locus.

5) In all other cases our main result states:

Theorem A — If either

m \geq 2

and

〈 v_{0}, v_{0} 〉 > 2

m > 2

and

〈 v_{0}, v_{0} 〉 \geq 2

, then

M_{m v_{0}}

is a locally factorial singular symplectic variety.

As an immediate application one obtains:

Theorem B — Under the hypotheses of Theorem A,

M_{m v_{0}}

does not admit a proper symplectic resolution.

Under some technical hypotheses theorems A and B hold as well for semistable torsion sheaves (see the main text). Partial results for Theorem B in the case

m = 2

have been obtained previously by two of us [17] , and, independently and with different methods, by Kiem and Choy [15, 16] .

We note that our approach is rather general; our main technical result, Proposition 3.5 , is essentially a linear-algebraic fact. Therefore, we expect that results similar to Theorem A and B might hold in other situations with similar geometry – in particular, for the moduli spaces of flat connections on an algebraic curve. In fact, Proposition 3.5 is a statement about quiver varieties of H. Nakajima [22] , although the numerical data corresponding to our quivers are specifically excluded from consideration in [22] (which is not surprising as one of the results of [22] is that any quiver variety considered there does admit a symplectic resolution). Thus our approach and our Proposition 3.5 might be used wherever one finds quiver varieties of the same type.

Acknowledgements: The moment when Grothendieck’s theorem on factoriality was shown to us by Duco van Straten turned around our approach to the problem.

We thank him as well as Stefan Bauer, Daniel Huybrechts and Joseph Le Potier for many helpful discussions.

Many authors have worked on moduli of sheaves on K3 and abelian surfaces ever since the seminal work of Mukai. The most general results for our purposes have been obtained by Yoshioka [27] . We refer to the textbook [13] and Yoshioka’s paper for further references and general information on semistable sheaves and their moduli spaces.

2 Notation and conventions, plan of the paper

2.1 The underlying surface.

Throughout this article

X

will denote a complex projective K3 or abelian surface with a fixed symplectic structure, i.e. an isomorphism

H^{2} (X, O_{X}) \sim = C

, and a fixed ample divisor

H

The even integral cohomology

H^{even} (X, Z)

is equipped with a pairing

〈 v, w 〉 : = - \int_{X} v w^{\lor},

where

w^{\lor} = (- 1)^{i} w

for

w \in H^{2 i} (X, Z)

. Following Mukai we associate to each coherent sheaf

E

its Mukai vector

v (E) : = ch (E) \sqrt{td (X)} \in H^{even} (X, Z) .

The Hilbert polynomial of

E

with respect to an ample divisor

H

can be expressed in terms of its Mukai vector as follows:

χ (E \otimes O_{X} (m H)) = - 〈 v (E), v (O_{X} (- m H)) 〉 = : P_{v} (m) .

2.2 Semistable sheaves.

Stability or semistability of a coherent sheaf is defined with respect to a fixed ample divisor

H

. We let

M_{v}

denote the moduli space of semistable sheaves with Mukai vector

v

. Closed points of

M_{v}

are in natural bijection with polystable sheaves

E

. Points corresponding to stable sheaves form a – possibly empty – open subset

M_{v}^{s} \subset M_{v}

Semistable sheaves may have two–, one– or zero–dimensional support. Stability in the first case was defined by Maruyama and Gieseker, the generalisation to pure sheaves of arbitrary dimension is due to Simpson. In the rest of the paper we exclude once for all the case of zero-dimensional sheaves as being well-known: if the Mukai vector is

v = (0, 0, a)

then

M_{v} \sim = S^{a} X

, the symmetric product of

X

, and the Hilbert-Chow morphism

{Hilb}^{a} (X) \to M_{v}

provides a symplectic resolution.

2.3 General assumptions.

An element

v_{0} \in H^{even} (X, Z)

is primitive if it is not an integral multiple of another lattice element. Given a non-trivial element

v \in H^{even} (X, Z)

we may always decompose it as

v = m v_{0}

with a primitive element

v_{0}

and a multiplicity

m \in N

. Throughout this article we assume that

v_{0} = (r_{0}, c_{0}, a_{0})

has the following properties:

(*)

{\begin{matrix} Either r_{0} > 0 and c_{0} \in NS (X), \\ or r_{0} = 0, c_{0} \in NS (X) is effective, and a_{0} \neq 0; \\ 〈 v_{0}, v_{0} 〉 \geq 2 . \end{matrix}

The results of this paper suggest to distinguish systematically between the following three cases for a Mukai vector

v

satisfying assumptions

(*)

(A) $m = 1$ .
(B) $m = 2$ and $〈 v_{0}, v_{0} 〉 = 2$ .
(C) $m \geq 3$ , or $m = 2$ and $〈 v_{0}, v_{0} 〉 \geq 4$ .

2.4 General ample divisors.

The significance of

(*)

lies in the fact that one has the notion of a

v

–general ample divisor

H

: there is a systems of hyperplanes in the ample cone of

X

, called

v

–walls, that is countable but locally finite for torsion free sheaves ([13] , ch. 4C) and finite for torsion sheaves ([27] , sec. 1.4.) with the following property: if

H

v

–general, i.e. if

H

is not contained in any

v

–wall, then for every direct summand

E^{'}

of a polystable sheaf

E

with

v (E) = v

one has

v (E^{'}) \in Q v (E)

Let

H

be a

v_{0}

–general ample divisor and consider the following assertions:

\begin{matrix} (* *) & M_{v_{0}} is non-empty. \end{matrix}

\begin{matrix} (* * *) & M_{v_{0}} is irreducible. \end{matrix}

Yoshioka shows in [27] , Thm 0.1 and Thm 8.1, that

(*)

implies

(* *)

and

(* * *)

except when

X

is a K3 surface,

r_{0} = 0

and

c_{0}

is not ample. Moreover he has communicated to us an unpublished note that fills this gap, so that

(* *)

and

(* * *)

are consequences of

(*)

. An essential technique in Yoshioka’s work is the deformation of the underlying surface; the arguments are rather involved. For the irreducibility part

(* * *)

we give a new and direct proof, based on an old and beautiful idea of Mukai, see Theorem 4.1 .

2.5 Elements of the construction of moduli spaces.

We need to recall some basic elements of the construction machinery of moduli spaces of sheaves following the approach of Simpson [26] (see also [13] , ch. 4). Let

v

be a Mukai vector satisfying

(*)

and let

P_{v}

denote the corresponding Hilbert polynomial. Choose a sufficiently large integer

k = k (v)

and put

N = P_{v} (k)

ℋ : = O_{X} (- k H)^{\oplus N}

. Then there is a closed subscheme

R \subset {Quot}_{X, H} (ℋ, P)

with the following property: a closed point

[q : ℋ \to E] \in R

is stable or semistable with respect to the canonical

PGl (N)

–action and the corresponding linearisation of the determinant line bundle on

R

if and only if

q

induces an isomorphism

C^{N} \to H^{0} (X, E (k H))

and if

E

is stable or semistable, respectively. Let

R^{s} \subset R^{s s} \subset R

denote the open subsets of stable and semistable points, respectively. Then

R^{s s} / / PGl (N)) \sim = M_{v} and R^{s} / / PGl (N) \sim = M_{v}^{s} .

Let

π : R^{s s} \to M_{v}

denote the quotient map. The orbit of a point

[q : ℋ \to E]

is closed in

R^{s s}

if and only if

E

is polystable. In that case, the stabiliser subgroup of

[q]

PGl (N)

is canonically isomorphic to

PAut (E) = Aut (E) / C^{*}

. Moreover, by Luna’s slice theorem there is a

PAut (E)

–invariant subscheme

S \subset R^{s s}

, containing

[q]

, such that the canonical morphisms

(PGl (N) \times S) / / PAut (E) \to R^{s s} and S / / PAut (E) \to M

are etale. The Zariski tangent space

T_{[q]} S

is isomorphic to

{Ext}^{1} (E, E)

2.6 Local description.

The completion of the local ring

O_{S, [q]}

has the following deformation theoretic description:

Let

C [{Ext}^{1} (E, E)]

denote the ring of polynomial functions on

{Ext}^{1} (E, E)

and let

A : = C [{Ext}^{1} (E, E)]^{\land}

denote its completion at the maximal ideal

M

of functions vanishing at

0

. There is a trace map

tr : {Ext}^{2} (E, E) \to H^{2} (O_{X})

. We denote its kernel by

{Ext}^{2} (E, E)_{0}

. The automorphism group

Aut (E)

naturally acts on

{Ext}^{1} (E, E)

and

{Ext}^{2} (E, E)_{0}

by conjugation. Since the scalar multiples of the identity act trivially we actually have an action of the projective automorphism group

PAut (E) = Aut (E) / C^{*}

. There is a linear map

κ : {Ext}^{2} (E, E)_{0}^{*} ⟶ C [{Ext}^{1} (E, E)]^{\land},

the so-called Kuranishi map, with the following properties:

1. $κ$ is $PAut (E)$ –equivariant.
2. Let $I$ be the ideal generated by the image of $κ$ . Then there are isomorphisms of complete rings ${\hat{O}}_{S, [q]} \sim = A / I and {\hat{O}}_{M_{v}, [E]} \sim = (A / I)^{PAut (E)} .$
3. For every linear form $φ \in {Ext}^{2} (E, E)_{0}^{*}$ one has, for $e \in {Ext}^{1} (E, E)$ , $κ (φ) (e) = \frac{1}{2} φ (e \cup e) + higher order terms in e .$

2.7 Passage to the normal cone.

Let

J \subset C [{Ext}^{1} (E, E)]

denote the ideal generated by the image of the quadratic part of

κ

κ_{2} : {Ext}^{2} (E, E)_{0}^{*} ⟶ S^{2} {Ext}^{1} (E, E)^{*}, φ \mapsto (e \mapsto \frac{1}{2} φ (e \cup e))

Then

J

is the ideal of the null-fibre

F = μ^{- 1} (0)

of the morphism

μ : {Ext}^{1} (E, E) ⟶ {Ext}^{2} (E, E)_{0}, μ (e) = \frac{1}{2} (e \cup e) .

The ideals

I \subset C [{Ext}^{1} (E, E)]^{\land}

and

J \subset C [{Ext}^{1} (E, E)]

are related as follows.

The graded ring

gr A

associated to the

m

-adic filtration on

A = C [{Ext}^{1} (E, E)]^{\land}

is canonically isomorphic to

C [{Ext}^{1} (E, E)]

. For any ideal

a \subset A

let

in (a) \subset gr A

denote the ideal generated by the leading terms

in (f)

of all elements

f \in a

. Then property 3 of the Kuranishi–map says that

J \subset in (I) .

Hence there is the following chain of inequalities:

\begin{matrix} \begin{matrix} dim (F) & = & dim (gr A) / J \\ \geq & dim (gr A) / in (I) = dim gr (A / I) = dim (A / I) \\ \geq & dim {Ext}^{1} (E, E) - dim {Ext}^{2} (E, E)_{0}, \end{matrix} \end{matrix}

(2.1)

where the last inequality comes from the fact that

A

is regular of dimension

= dim {Ext}^{1} (E, E)

and

I

is generated by

dim {Ext}^{2} (E, E)_{0}

elements.

We need to describe

μ

in greater detail; the resulting description is similar to Nakajima’s construction of the so-called quiver varieties [22] . Write

\begin{matrix} E =^{s} \oplus_{i = 1} W_{i} \otimes E_{i} \end{matrix}

(2.2)

with pairwise non-isomorphic stable sheaves

E_{i}

and vector spaces

W_{i}

of dimension

n_{i}

. Let

W_{i j} : = Hom (W_{i}, W_{j})

and

V_{i j} : = {Ext}^{1} (E_{i}, E_{j})

. Then

End (E) = \oplus_{i} W_{i i}, {Ext}^{1} (E, E) = \oplus_{i, j} W_{i j} \otimes V_{i j}, {Ext}^{2} (E, E) = \oplus_{i} W_{i i} .

The automorphism group

Aut (E) = \prod_{i} Aut (W_{i}) \sim = \prod_{i} Gl (n_{i}) = : G (n)

acts on

{Ext}^{1} (E, E)

by conjugation on the first factor in each direct summand. By Serre-Duality, the pairing

V_{i j} \otimes V_{j i} \to C, e \otimes e^{'} \mapsto tr (e^{'} \cup e)

is non-degenerate and antisymmetric. This yields a symplectic form

ω

{Ext}^{1} (E, E)

such that

W_{i j} \otimes V_{i j}

and

W_{a b} \otimes V_{a b}

are perpendicular, unless

i = b

and

j = a

, in which case

ω : (W_{i j} \otimes V_{i j}) \otimes (W_{j i} \otimes V_{j i}) ⟶ C, ω (A \otimes e, A^{'} \otimes e^{'}) = tr (A^{'} A) tr (e^{'} \cup e) .

Moreover, the quadratic map

μ : {Ext}^{1} (E, E) \to {Ext}^{2} (E, E)_{0}

is given by

\begin{matrix} μ (\sum_{i j} \sum_{k} A_{i j}^{k} \otimes e_{i j}^{k}) = \sum_{i j} \sum_{k, ℓ} A_{i j}^{k} A_{j i}^{ℓ} tr (e_{i j}^{k} e_{j i}^{ℓ}) . \end{matrix}

(2.3)

2.8 Strategy

In general we do not know how to compute the Kuranishi map explicitly. However, the explicit description of the quadratic part

μ

given above allows for a detailed study of the fibre

F : = μ^{- 1} (0) \subset {Ext}^{1} (E, E)

. The passage from

κ

μ

corresponds to the passage from the local ring

{\hat{O}}_{S, [q]}

to the coordinate ring

O_{F}

of its tangent cone.

In section 3 we show that under certain hypotheses the fibre

F

is an irreducible normal complete intersection which is, in case (C), regular in codimension

\leq 3

and state consequences for the local rings

O_{R^{s s}, [q]}

of points

[q]

in closed orbits of

R^{s s}

Section 4 contains a basic irreducibility result for moduli spaces of sheaves on a K3 or abelian surface.

In section 5 it is proved — under the hypothesis that the ample divisor is

v

–general — that the moduli space

M

is a non-empty irreducible normal variety of expected dimension, and that it is locally factorial in case (C). As an application we show in section 6 that in case (C) the moduli space does not admit a symplectic resolution.

3 Symplectic reduction

3.1 The symplectic momentum map.

Let

U

be a smooth affine algebraic variety over

C

endowed with a symplectic form

ω

. Let

G

be a reductive group that acts on

U

preserving

ω

. This action induces an infinitesimal action of the Lie algebra

g

G

, i.e. a homomorphism of Lie algebras

g \to Γ (U, T_{U})

. We denote the vector field corresponding to

A \in g

x \in U

A_{x}

A momentum map for the action is a

G

-equivariant morphism

μ : U \to g^{*}

with the property that

d μ_{x} (ξ) (A) = ω (ξ, A_{x})

for all

x \in U

and

ξ \in T_{x} U

. If a momentum map exists, it is unique up to an additive constant in

(g^{*})^{G}

Let

μ : U \to g^{*}

be a momentum map with null-fibre

F : = μ^{- 1} (0)

Lemma 3.2. — Let

x \in F

be a point with stabiliser subgroup

H \subset G

Then the image of

d μ_{x} : T_{x} U \to g^{*}

(g / h)^{*} = h^{⊥}

, where

h \subset g

denotes the Lie algebra of

H

. In particular, if

H

is finite then

d μ_{x}

has maximal rank and

F

is regular at

x

of dimension

dim (U) - dim (G)

Proof. The image $d μ_{x}$ annihilates $A \in g$ if and only if $ω (ξ, A_{x}) = 0$ for all $ξ \in T_{x} U$ , i.e. if $A_{x}$ is perpendicular to $T_{x} U$ with respect to $ω$ . As $ω$ is non-degenerate, this is equivalent to saying that $A_{x}$ vanishes, hence is a tangent vector to the stabiliser subgroup $H$ . □

Lemma 3.3. — Let

μ : U \to g^{*}

be a momentum map with null-fibre

F

Let

Z \subset F

be the closed subset of points with non-finite stabiliser group.

Let

d = dim U - dim g

1. If $dim (Z) \leq d - 1$ , then $F$ is a reduced complete intersection of dimension $d$ .
2. If $dim (Z) \leq d - 2$ , then $F$ is normal.

Proof. Every irreducible component of $F$ must have dimension $\geq d$ since $F$ is cut out by $dim g$ equations. By Lemma 3.2 , $F$ has dimension $d$ in each point $x \in F \ Z$ .
If $dim Z < d$ , then $F \ Z$ is dense in $F$ and every irreducible component has precisely dimension $d$ . Hence $F$ is a complete intersection and in particular Cohen-Macaulay ([2] , Cor. III 4.5). Since $F \ Z$ is smooth, $F$ in addition satisfies condition $R_{0}$ and is therefore reduced ([2] , Prop. VII 2.2). If in addition $dim (Z) \leq d - 2$ , then $F$ is regular in codimension 1 and normal by Serre’s criterion ([2] , Cor. VII 2.13). □

3.4 The key estimate.

We want to apply the lemma to the following particular situation, that arises in the study of local rings of the moduli space of sheaves.

Set-up: Let

W_{1}, \dots, W_{s}

be a sequence of vector spaces,

s \geq 1

. The dimensions

n_{i} = dim (W_{i})

form the components of a vector

n \in N_{0}^{s}

. Furthermore, let

W_{i j} = Hom (W_{i}, W_{j})

. There is a natural symmetric pairing

W_{i j} \otimes W_{j i} \to C, (A, B) \mapsto tr (B A) .

Moreover, let

V_{i j}

1 \leq i, j \leq s

, be vector spaces, equipped with non-degenerate pairings

ω_{i j} : V_{i j} \otimes V_{j i} \to C,

that are skew-symmetric in the sense that

ω_{i j} (e, e^{'}) = - ω_{j i} (e^{'}, e)

. Then the vector space

U (n) : = \oplus_{i, j} W_{i j} \otimes V_{i j}

carries a natural symplectic form

ω

with the property that

W_{i j} \otimes V_{i j}

is perpendicular to all

W_{a b} \otimes V_{a b}

(a, b) \neq (j, i)

and

ω : (W_{i j} \otimes V_{i j}) \otimes (W_{j i} \otimes V_{j i}) ⟶ C, (A \otimes e) \otimes (A^{'} \otimes e^{'}) \mapsto tr (A^{'} A) ω_{i j} (e, e^{'}) .

In the following arguments the vector spaces

V_{i j}

are fixed and chosen once for all, whereas the sequence of vector spaces

W_{i}

can be replaced by appropriate subspaces etc. We will argue by induction over the dimension vector

n

as an element in the monoid

N_{0}^{s}

. Most objects defined below will therefore be indexed by

n

, like the space

U (n)

above, even if this is not quite accurate as they really depend on the spaces

W_{i}

The group

G (n) = \prod_{i} Aut (W_{i})

acts on

U (n)

by conjugation on the first factors in the decomposition. The subgroup of scalars

C^{*} \subset G (n)

acts trivially. Let

P G (n) : = G (n) / C^{*}

. The action of

P G (n)

U (n)

preserves the symplectic structure. The moment map for the action is

\begin{matrix} μ (n) : U (n) & ⟶ & p g (n)^{*} ≃ Ker (\oplus_{i} gl (n_{i}) - - - \to tr C), \end{matrix}

\begin{matrix} \sum_{i, j, k} A_{i j}^{(k)} v_{i j}^{(k)} & \mapsto & \sum_{k, ℓ} \sum_{i, j} A_{i j}^{(k)} A_{j i}^{(ℓ)} tr (v_{i j}^{(k)} \cup v_{j i}^{(ℓ)}) \end{matrix}

Let

F (n) : = μ (n)^{- 1} (0) \subset U (n)

denote the null-fibre of the moment map. The structure of

F (n)

depends only on

n

and the dimensions

d_{i j} : = dim (V_{i j})

. Let

D

denote the matrix

(d_{i j})

and let

a : = min {d_{i j} - 2 δ_{i j}}

Proposition 3.5. — Assume that

a \geq 2

. Then

F (n)

is an irreducible normal complete intersection of dimension

d : = n^{t} (D - I) n + 1

. Moreover,

F (n)

is regular in codimension

\leq 3

with the possible exception of the two cases

1. $n = (1, 1)$ , $d_{12} = 2$ , and
2. $n = (2)$ , $d_{11} = 4$ .

Proof. 1. Since

dim (U (n)) = \sum_{i, j} n_{i} n_{j} d_{i j}

and since the range of

μ

has dimension

\sum_{i} n_{i} n_{i} - 1

, the expected dimension of

F (n)

d = \sum_{i, j} n_{i} n_{j} d_{i j} - \sum_{i} n_{i} n_{i} + 1 = n^{t} (D - I) n + 1 .

Also,

F (n)

is a cone and hence connected. By Lemma 3.3 , it suffices to show that the locus

Z

of points in

F (n)

with non-trivial stabiliser in

P G (n)

has dimension

\leq d - 4

in general and

\leq d - 3

in the two exceptional cases. This will be done by induction on the dimension vector

n \in N_{0}^{s}

The induction starts with

n = (0, \dots, 1, \dots, 0)

, in which case the statement is trivial. So let

n \in N_{0}^{s}

be an arbitrary element and assume that the proposition holds for all

n^{'} \in N_{0}^{s}

such that

0 < \sum_{i} n_{i}^{'} < \sum_{i} n_{i}

2. We can analyse

Z

as follows: Let

g \in G (n)

g / \in C^{*}

, and consider the corresponding fixed point locus

F (n)^{g}

. The image

G (n) F (n)^{g}

of the morphism

φ : G (n) \times F (n)^{g} \to F (n)

(g^{'}, x) \mapsto g^{'} x

, consists of all points

y \in F

whose stabiliser subgroup

G (n)_{y}

contains an element conjugate to

g

. Suppose that

H \subset G (n)

is a subgroup that stabilises the fixed point set

F (n)^{g}

. Then we can bound the dimension of the fibres of

φ

dim (H)

. It follows that

dim (G (n) F (n)^{g}) \leq dim F (n)^{g} + dim G (n) - dim H

. In the following we will describe a finite set of elements

g

such that

Z

is covered by the corresponding sets

G (n) F (n)^{g}

and such that for each

g

one has

dim (G (n) F (n)^{g}) \leq d - 3

\leq d - 4

. This gives the desired bound for

dim (Z)

3. Let

g = (g_{1}, \dots, g_{s}) \in G (n)

g / \in C^{*}

. We distinguish three cases: 3.1. Case:

g

is semisimple. For each

λ \in C

consider the eigenspaces

W_{i} (λ) \subset W_{i}

g_{i}

, and let

n_{i} (λ) = dim (W_{i} (λ))

n (λ) = (n_{i} (λ))_{i}

. Then

n = \sum_{λ} n (λ)

. There is a decomposition

U (n)^{g} = \oplus_{λ} U (n (λ)), U (n (λ)) = \oplus_{i j} Hom (W_{i} (λ), W_{j} (λ)) \otimes V_{i j} .

Moreover, the restriction of the momentum map to the fixed point locus splits into a product of momentum maps for each

U (n (λ))

μ (n) |_{U (n)^{g}} = \prod_{λ} μ (n (λ)), μ (n (λ)) : U (n (λ)) ⟶ p g (n (λ))^{*} .

It follows that

F (n)^{g} = \prod_{λ} F (n (λ)) with F (n (λ)) = μ (n (λ))^{- 1} (0) .

By induction, we have

dim (F (n)^{g}) = \sum_{λ} dim (F (n (λ)) = {\sum_{λ}}^{'} (n (λ)^{t} (D - I) n (λ) + 1),

where

\sum^{'}

indicates that we only sum over all

λ

with

n (λ) \neq 0

Next,

U (n)^{g}

is stabilised by

H = \prod_{λ} G (n (λ))

, a subgroup in

G (n)

of codimension

n^{t} n - \sum_{λ} n (λ)^{t} n (λ)

. We obtain the following upper bound for the dimension of

G (n) F (n)^{g}

dim (G (n) F (n)^{g}) \leq {\sum_{λ}}^{'} (n (λ)^{t} (D - 2 I) n (λ) + 1) + n^{t} n .

Note that

ν : = | {λ | n (λ) \neq 0} | \geq 2

, since

g / \in C^{*}

. The difference of

dim (G (n) F (n)^{g})

to the expected dimension of

F (n)

is therefore bounded below by

\begin{matrix} Δ & : = & (n^{t} (D - 2 I) n + 1) - {\sum_{λ}}^{'} (n (λ)^{t} (D - 2 I) n (λ) + 1) \end{matrix}

\begin{matrix} = & \sum_{λ \neq μ} n (λ)^{t} (D - 2 I) n (μ) - (ν - 1) \end{matrix}

\begin{matrix} \geq & 2 ν (ν - 1) - (ν - 1) = (2 ν - 1) (ν - 1) \geq 3 . \end{matrix}

Clearly,

Δ \geq 4

for

ν \geq 3

. Assume that

ν = 2

, say with the distinct eigenvalues

λ

and

λ^{'}

. Then

Δ \geq 2 \sum_{i, j} n_{i} (λ^{'}) n_{j} (λ^{''}) (d_{i j} - 2 δ_{i j}) - 1 \geq 2 a \sum_{i} n_{i} (λ^{'}) \sum_{i} n_{i} (λ) - 1 .

Thus

Δ = 3

implies

a = 2

and

\sum_{i} n_{i} (λ) = 1 = \sum_{i} n_{i} (λ^{'})

. Hence there are only the following exceptional cases:

1. $s = 1$ , $n = 2$ , $d_{11} = 2 + 2 δ_{11} = 4$ , or
2. $s = 2$ , $n = (1, 1)$ and $d_{12} = d_{21} = 2$ .

If a point

f \in F (n)

is fixed by a semisimple element, it is also fixed by a whole subtorus

T \subset G (n)

. Up to a conjugation, there is only a finite number of such subtori

T_{i} \subset G (n)

. Choosing an element

g_{i} \in T_{i}

in each of these subtori, we see that the union of all sets

G (n) F (n)^{g}

g

semisimple, is covered by the finite union of all sets

G (n) F (n)^{g_{i}}

3.2. Case:

g

is unipotent. We may write

g = 1 + h

, with a non-zero nilpotent element

h = (h_{1}, \dots, h_{s}) \in \oplus_{i} End (W_{i})

. Let

K_{i}^{(ℓ)} : = ker (h_{i}^{ℓ}) \subset W_{i}

and

m_{i} (ℓ) : = dim K_{i}^{(ℓ)}

for all

ℓ \in N_{0}

. There is a filtration

0 = K_{i}^{(0)} \subset K_{i}^{(1)} \subset \dots = W_{i} .

For each level

ℓ > 0

we choose a graded complement

W_{i}^{(ℓ)}

h K_{i}^{(ℓ + 1)} + K_{i}^{(ℓ - 1)}

K^{(ℓ)}

and let

n_{i}^{(ℓ)} = dim W_{i}^{(ℓ)}

. (We note that this is an instance of the so-called Jacobson-Morozov-Deligne filtration associated to a nilpotent element, see [8,1.6] ; the spaces

W_{i}^{(ℓ)}

are the primitive subspaces with respect to an

s l_{2}

-triple containing

h

.) Suppose that

A = (A_{i j}) \in \oplus_{i j} Hom (W_{i}, W_{j})

commutes with

h

. Then

A_{i j}

is completely determined by its value on the spaces

W_{i}^{(ℓ)}

ℓ \in N

, and conversely, any value of

A_{i j} : W_{i}^{(ℓ)} \to K_{j}^{(ℓ)}

can be prescribed. The composition with the canonical projection

K_{j}^{(ℓ)} \to W_{j}^{(ℓ)}

defines a homomorphism

A_{i j}^{(ℓ)} : W_{i}^{(ℓ)} ⟶ W_{j}^{(ℓ)}

, and the map

Φ : {(\oplus_{i j} Hom (W_{i}, W_{j}))}^{g} ⟶ \oplus_{ℓ} (\oplus_{i j} Hom (W_{i}^{(ℓ)}, W_{j}^{(ℓ)})), (A_{i j}) \mapsto (A_{i j}^{(ℓ)}),

is a ring homomorphism. Let

Φ_{V} : U (n)^{g} = {(\oplus_{i, j} W_{i j} \otimes V_{i j})}^{g} ⟶ \oplus_{ℓ} U (n^{(ℓ)}) = \oplus_{ℓ} \oplus_{i, j} Hom (W_{i}^{(ℓ)}, W_{j}^{(ℓ)}) \otimes V_{i j}

be analogously defined. Then

Φ_{V} (F (n)^{g}) \subset \prod_{ℓ} F (n^{(ℓ)})

, and the fibres of

Φ_{V}

have dimension

\sum_{ℓ} {n^{(ℓ)}}^{t} D (m^{(ℓ)} - n^{(ℓ)})

. By induction, this yields the bound

\begin{matrix} dim (F (n)^{g}) & \leq & \sum_{ℓ} dim (F (n (ℓ)) + dim (ker (Φ_{V})) \end{matrix}

\begin{matrix} = & {\sum_{ℓ}}^{'} ({n^{(ℓ)}}^{t} (D - I) n^{(ℓ)} + 1) + \sum_{ℓ} {n^{(ℓ)}}^{t} D (m^{(ℓ)} - n^{(ℓ)}), \end{matrix}

where

\sum^{'}

signifies summation over all

ℓ

with

n^{(ℓ)} \neq 0

. Moreover, the centraliser

H \subset G (n)

g

is an open subset in

{(\oplus_{i} End (W_{i}))}^{g} \sim = \oplus_{ℓ} \oplus_{i} Hom (W_{i}^{(ℓ)}, K_{i}^{(ℓ)})

and therefore has dimension

dim (H) = \sum_{ℓ} {n^{(ℓ)}}^{t} m^{(ℓ)}

. Connecting these pieces of information we obtain

\begin{matrix} dim (G (n) F (n)^{g}) & \leq & dim (F (n)^{g}) + dim (G (n)) - dim (H) \end{matrix}

\begin{matrix} \leq & {\sum_{ℓ}}^{'} ({n^{(ℓ)}}^{t} (D - I) n^{(ℓ)} + 1) + \sum_{ℓ} {n^{(ℓ)}}^{t} D (m^{(ℓ)} - n^{(ℓ)}) \end{matrix}

\begin{matrix} + n^{t} n - \sum_{ℓ} {n^{(ℓ)}}^{t} m^{(ℓ)} . \end{matrix}

The difference of the last expression to the expected dimension of

F (n)

Δ : = [n^{t} (D - I) n - n^{t} n + 1] - {\sum_{ℓ}}^{'} [{n^{(ℓ)}}^{t} (D - I) m^{(ℓ)} - {n^{(ℓ)}}^{t} n^{(ℓ)} + 1] .

Note that the two bracketed expressions are not quite symmetric to each other due to the presence of

m^{(ℓ)}

instead on

n^{(ℓ)}

. We can get rid of

n

and

m^{(ℓ)}

due to the relations

m^{(ℓ)} = \sum_{k} n^{(k)} min {k, ℓ}, n = \sum_{k} n^{(k)} k,

and can rewrite the bound

Δ

in terms of the

n^{(k)}

as follows:

Δ = \sum_{ℓ, k} {n^{(k)}}^{t} (D - I) n^{(ℓ)} (k ℓ - min {k, ℓ}) - \sum_{k, ℓ} {n^{(k)}}^{t} n^{(ℓ)} k ℓ + {\sum_{k}}^{'} ({n^{(k)}}^{t} n^{(k)} - 1) + 1

Reorganise the sum in collecting those terms that contain

n^{(1)}

\begin{matrix} Δ = 1 + [- 1 + 2 {n^{(1)}}^{t} \sum_{k \geq 2} ((k - 1) (D - 2 I) - I) n^{(k)}] + {\sum_{k \geq 2}}^{'} ({n^{(k)}}^{t} n^{(k)} - 1) \end{matrix}

\begin{matrix} + \sum_{k, ℓ \geq 2} {n^{(k)}}^{t} ((k ℓ - min {k, ℓ}) (D - 2 I) - min {k, ℓ} I) n^{(ℓ)} \end{matrix}

Here the second summand

[\dots]

appears only if

n^{(1)} \neq 0

. Note that there always is at least one index

k \geq 2

with

n^{(k)} \neq 2

, since

h \neq 0

. This shows that all summands in the last expression for

Δ

are non-negative.

The minimal contribution of a non-zero vector

n^{(k)}

k \neq 2

, to

Δ

k ((k - 1) a - 1) {(\sum_{i} n_{i}^{(k)})}^{2} \geq 2 .

Thus we always have

Δ \geq 3

, and even better:

Δ \geq 4

in all cases except

a = 2, n^{(2)} = (1), n^{(k)} = 0 for all k \neq 2 .

In this case

s = 1

n = (2)

, and

d_{11} = 4

, which is the same exceptional case as before.

As in the semisimple case, the union of all sets

G (n) F (n)^{g}

g

unipotent, is covered by a finite number of such sets. In fact, this is even easier to see: up to conjugation there are only finitely many different nilpotent elements

h

and hence only finitely many different subschemes

G (n) F (n)^{1 + h} \subset Z

3.3. Case:

g \in G (n) \ C^{*}

arbitrary. Consider the multiplicative Jordan decomposition

g = s u

, where

s

is semisimple,

u

is unipotent and

s

and

u

commute. Any endomorphism that commutes with

g

also commutes with

s

and

u

. This implies that

F (n)^{g} \subset F (n)^{s} \cap F (n)^{u}

, so that the general case is covered by 3.1. and 3.2. above. □

3.6 Return from the normal cone

Let

v_{0} \in H^{even} (X, Z)

be a primitive Mukai vector satisfying

(*)

. Let

v = m v_{0}

for some multiplicity

m \in N

. We keep the notation introduced earlier.

Proposition 3.7. — Let

H

be an arbitrary ample divisor. Let

E = \oplus_{i = 1}^{s} E_{i}^{\oplus n_{i}}

be a polystable sheaf whose stable direct summands

E_{i}

satisfy the condition

\begin{matrix} v (E_{i}) \in N v_{0} \end{matrix}

(3.1)

Consider a point

[q : ℋ \to E] \in R^{s s}

and a slice

S \subset R^{s s}

to the orbit of

[q]

as above. Then

O_{S, [q]}

is a normal complete intersection domain of dimension

dim {Ext}^{1} (E, E) - dim {Ext}^{2} (E, E)_{0} = 1 + \sum_{i, j} n_{i} (dim {Ext}^{1} (E_{i}, E_{j}) - δ_{i j}) n_{j},

that has property

R_{3}

in all cases except the following two:

1. $s = 1$ , $n_{1} = 2$ , $dim {Ext}^{1} (E_{1}, E_{1}) = 4$ ,
2. $s = 2$ , $n_{1} = n_{2} = 1$ , $dim {Ext}^{1} (E_{1}, E_{2}) = 2$ .

Proof. Recall the notation introduced in sections 2.6 and 2.7 . By Proposition 3.5 ,

F = μ^{- 1} (0) = Spec (gr A / J)

is a normal complete intersection variety of dimension

\begin{matrix} dim (F) & = & 1 + \sum_{i, j} n_{i} (dim {Ext}^{1} (E_{i}, E_{j}) - δ_{i j}) n_{j} \end{matrix}

\begin{matrix} = & dim {Ext}^{1} (E, E) - dim {Ext}^{2} (E, E)_{0} . \end{matrix}

Therefore, we must have equality at all places in inequality 2.1 . Furthermore, since

F = Spec (gr A / J)

is reduced and irreducible, the equality of dimensions implies

J = in (I)

. It follows that

gr ({\hat{O}}_{S, [q]}) = gr (A / I) = gr A / in (I) = Γ (F, O_{F})

is a normal complete intersection. In particular,

gr ({\hat{O}}_{S, [q]})

is Cohen-Macaulay, hence satisfies

S_{k}

for all

k \in N

. Unless we are in the two exceptional cases,

gr ({\hat{O}}_{S, [q]})

is smooth in codimension 3. Now remark that

gr ({\hat{O}}_{S, [q]}) = gr (O_{S, [q]})

([1] , 10.22) and then use the following proposition which shows that

O_{S, [q]}

itself is a normal complete intersection which, unless we are in the two exceptional cases, satisfies

R_{3}

. □

Proposition 3.8. — Let

B

be a noetherian local ring with maximal ideal

m

and residue field

B / m \sim = C

. Let

gr B

denote the graded ring associated to the

m

-adic filtration of

B

. Then

dim (B) = dim (gr B)

, and if

gr B

is an integral domain or normal or a complete intersection then the same is true for

B

. Moreover if

gr B

satisfies

R_{k}

and

S_{k + 1}

for some

k \in N

then

B

satisfies

R_{k}

Proof. The assertion about integrality and normality is Krull’s theorem (see [20] (17.D) Thm 34). The assertions about complete intersections and the property $R_{k}$ are due to Cavaliere and Niesi ([6] , Theorems 3.4 and 3.13). □

Lemma 3.9. — The assumption 3.1 in Proposition 3.7 is satisfied in any of the following two situations:

1. $H$ is $v$ –general.
2. $E = E_{0}^{\oplus m}$ for some stable sheaf $E_{0}$ with $v (E_{0}) = v_{0}$ .

The exceptions of Proposition 3.7 are met in case (B) only, i. e. if

〈 v_{0}, v_{0} 〉 = 2

and

m = 2

Proof. Under the assumption that $H$ is $v$ –general one has $v (E_{j}) = r_{j} v_{0}$ for some $r_{j} \in N$ and all direct summands $E_{j}$ of $E$ . Then $dim {Ext}^{1} (E_{i}, E_{j}) = r_{i} r_{j} 〈 v_{0}, v_{0} 〉 \geq 2$ .
Thus Proposition 3.7 applies. □

Proposition 3.10. — 1. Let

H

be a

v

–general ample divisor. Then

R^{s s}

is normal and locally a complete intersection of dimension

〈 v, v 〉 + 1 + N^{2}

In case (C) it has property

R_{3}

and hence is locally factorial.

2. Suppose that

E = E_{0}^{\oplus m}

for some stable sheaf

E_{0}

with

v (E_{0}) = v_{0}

Let

H

be an arbitrary ample divisor. In case (C), there is an open neighbourhood

U

[E] \in M_{v}

such that

π^{- 1} (U) \subset R^{s s}

is locally factorial of dimension

〈 v, v 〉 + 1 + N^{2}

Proof. 1. Let $[q] \in R^{s s}$ be a point with closed orbit, and let $S \subset R^{s s}$ be a $PAut (E)$ –equivariant subscheme as in subsection 2.5 . By Lemma 3.9 and Proposition 3.7 , the local ring $O_{S, [q]}$ is a normal complete intersection that has property $R_{3}$ in case (C). But being normal or locally a complete intersection or having property $R_{k}$ are open properties [EGA IV 19.3.3, 6.12.9]. Hence there is an open neighbourhood $U$ of $[q]$ in $S$ that is normal, locally a complete intersection, and has property $R_{3}$ in case (C). The natural morphism $PGl (N) \times S \to R^{s s}$ is smooth. Therefore every closed orbit in $R^{s s}$ has an open neighbourhood that is normal, locally a complete intersection, and has property $R_{3}$ in case (C). Finally, every $PGl (N)$ –orbit of $R^{s s}$ meets such an open neighbourhood. It follows that $R^{s s}$ is normal, locally a complete intersection. In case (C), $R^{s s}$ is regular in codimension 3 and hence locally factorial due to the following theorem of Grothendieck. 2. The second assertion follows analogously. □

Theorem 3.11. (Grothendieck [12] Exp. XI Cor. 3.14) — Let

B

be noetherian local ring. If

B

is a complete intersection and regular in codimension

\leq 3

, then

B

is factorial.

4 A basic irreducibility result

The following theorem generalises a beautiful result of Mukai [21] .

Theorem 4.1. — Let

X

be a projective K3 or abelian surface with an ample divisor

H

. Let

M_{v}

be the moduli space of semistable sheaves associated to a vector

v \in H^{even} (X, Z)

. Suppose that

Y \subset M_{v}

is a connected component parametrising stable sheaves only. Then

M_{v} = Y

Proof. 1. Since all points in

Y

correspond to stable sheaves,

Y

is smooth of expected dimension

dim (Y) = 2 + 〈 v, v 〉

. Fix a point

[F] \in Y

and suppose that there is a point

[G] \in M_{v} \ Y

. We shall exploit a beautiful old idea of Mukai [21] : assume for a moment that there were a universal family

F \in Coh (Y \times X)

. Let

p : Y \times X \to Y

and

q : Y \times X \to X

be the projections. We may then compare the relative Ext-sheaves

{Ext}_{p}^{∙} (q^{*} F, F)

and

{Ext}_{p}^{∙} (q^{*} G, F)

. Since

F

and

G

are numerically equal on

X

, the same is true for the classes of the Ext-sheaves according to the Grothendieck-Riemann-Roch theorem. This will lead to a contradiction.

2. In general, there is no universal family, but the following construction will be sufficient:

Lemma 4.2. — There is a smooth projective variety

Y^{'}

that parametrises a family

F

of stable sheaves on

X

with Mukai vector

v

such that the classifying morphism

f : Y^{'} \to Y

is surjective, generically finite, and etale over a neighbourhood of

[F]

Proof. Let $R^{'} : = Y \times_{M_{v}} R^{s s}$ . Then $R^{'} \to Y$ is a $PGl (N)$ –principal fibre bundle, locally trivial in the etale topology. Moreover, there is a universal epimorphism $O_{R^{'}} ⊠ ℋ \to F^{'}$ . We form the quotient $P : = (P^{N - 1} \times R^{'}) / / Gl (N)$ . Then $P$ is a smooth projective variety, and the natural morphism $P \to Y$ is locally a product in the etale topology with fibres isomorphic to $P^{N - 1}$ . The center $C^{*} \subset Gl (N)$ acts trivially on the family $O_{P^{N - 1}} (- 1) ⊠ F^{'}$ . Therefore, this sheaf descends to a family $F_{P}$ on $P \times X$ . Let $L$ be a very ample line bundle on $P$ . Choose a linear subspace $Z \subset P (H^{0} (P, L))$ of codimension $N - 1$ in such a way that $Y^{'} : = Z \cap P$ is smooth and $f : Y^{'} \to Y$ is etale over a neighbourhood of $[F]$ . Finally, let $F : = F_{P} |_{Y^{'} \times X}$ . □

3. Let

f : Y^{'} \to Y

and

F

be chosen as in the lemma and let

p : Y^{'} \times X \to Y^{'}

and

q : Y^{'} \times X \to X

denote the two projections. Moreover, let

f^{- 1} ([F]) = {p_{1}, \dots, p_{n}}

G

represents a point in

M \ Y

and hence is not isomorphic to any of the stable sheaves

E

[E] \in Y

, one has

Hom (G, E) = 0 = {Ext}^{2} (G, E)

for all

[E] \in Y

It follows that

{Ext}_{p}^{0} (q^{*} G, F)

and

{Ext}_{p}^{2} (q^{*} G, F)

vanish and that

W : = {Ext}_{p}^{1} (q^{*} G, F)

is a locally free sheaf on

Y^{'}

of rank

〈 v, v 〉 = dim (Y) - 2

G

is replaced by

F

the situation gets more complicated as the dimension of the Ext-groups jumps on the fibre

T

. There is a complex of locally free

O_{Y^{'}}

–sheaves

\begin{matrix} 0 ⟶ A^{0} - - \to α A^{1} - - \to β A^{2} ⟶ 0 \end{matrix}

(4.1)

with the property that

{Ext}_{p_{_{S}}}^{i} (t_{_{X}}^{*} q^{*} G, t_{X}^{*} F) \sim = h^{i} (t^{*} (A^{∙}))

for every base change

\begin{matrix} \begin{matrix} S \times X & - - \to t_{X} & Y^{'} \times X & - \to q & X \\ p_{S} ↓ & p ↓ \\ S & - \to t & Y^{'} . \end{matrix} \end{matrix}

(4.2)

Lemma 4.3. — The degeneracy locus of

α

and

β

is the union of the reduced points

p_{1}, \dots, p_{n}

. Moreover,

rk α (p_{i}) = rk A^{0} - 1

and

rk β (p_{i}) = rk A^{2} - 1

for

i = 1, \dots, n

Proof. For all $[E] \in Y$ , $E ≁ = F$ , one has $Hom (F, E) = 0 = {Ext}^{2} (F, E)$ . This implies that $α$ and $β$ have maximal rank on $Y^{'} \ {p_{1}, \dots, p_{n}}$ . Moreover, $Hom (F, F) = C = {Ext}^{2} (F, F)$ , and this gives the second assertion of the lemma. It remains to show that the degeneracy locus is reduced. Recall that tangent vectors in $T_{[F]} Y$ correspond bijectively to elements $γ \in {Ext}^{1} (F, F)$ . Let $F_{γ}$ be the infinitesimal extension of $F$ over $Spec C [ɛ]$ corresponding to $γ$ . The extension
$\begin{matrix} 0 ⟶ F - \to ɛ F_{γ} ⟶ F ⟶ 0 \end{matrix}$ (4.3)
induces a long exact sequence $⟶ {Ext}^{i} (F, F) ⟶ {Ext}_{Spec C [ɛ]}^{i} (F \otimes C [ɛ], F_{γ}) ⟶ {Ext}^{i} (F, F) - - \to \partial {Ext}^{i + 1} (F, F) ⟶$ where the boundary operator is given by $\partial (e) = γ \cup e$ . Now $γ \cup - : C = End (F, F) \to {Ext}^{1} (F, F)$ is clearly injective, and $γ \cup - : {Ext}^{1} (F, F) \to {Ext}^{2} (F, F)$ is surjective since the symplectic form on ${Ext}^{1} (F, F)$ is non-degenerate. It follows that ${Ext}^{0} (F, F) \sim = {Ext}_{Spec C [ɛ]}^{0} (F \otimes C [ɛ], F_{γ})$ and ${Ext}_{Spec C [ɛ]}^{2} (F \otimes C [ɛ], F_{γ}) \sim = {Ext}^{2} (F, F)$ . If the degeneracy locus of $α$ resp. $β$ were not reduced, the corresponding $Ext$ groups should be bigger than $C$ for at least one $γ$ . The calculation shows that this is not the case. □

4. Let

σ : Z \to Y

denote the blow-up of

Y

[F]

with exceptional divisor

D

and similarly

φ : Z^{'} \to Y^{'}

the blow-up of

Y^{'}

in all points

p_{i}

with corresponding exceptional divisors

D_{i}

\begin{matrix} \begin{matrix} D & \subset & Z & \leftarrow - g & Z^{'} & \supset & D_{i} \\ ↓ & σ ↓ & φ ↓ & ↓ \\ [F] & \in & Y & \leftarrow - - f & Y^{'} & ∋ & p_{i} \end{matrix} \end{matrix}

(4.4)

According to the lemma, the degeneracy locus of both

φ^{*} (α)

and

φ^{*} (β)

is precisely the smooth divisor

D^{'} = D_{1} \cup \dots \cup D_{n}

. Therefore these maps factor as follows:

\begin{matrix} φ^{*} A^{0} \subset {A^{'}}^{0} - - \to α^{'} φ^{*} A^{1} - - \to β^{'} {A^{'}}^{2} \subset φ^{*} A^{2}, \end{matrix}

(4.5)

where

{A^{'}}^{0}

and

{A^{'}}^{2}

are locally free,

α^{'}

and

β^{'}

are homomorphisms of maximal rank. Moreover, the line bundles

L : = φ^{*} A^{2} / {A^{'}}^{2}

and

M : = {A^{'}}^{0} / φ^{*} A^{0}

D^{'}

are characterised by the canonical isomorphisms

L \otimes O_{D^{'}} \sim = {Ext}_{D^{'}}^{2} (q^{*} F, O_{D} ⊠ F |_{f^{- 1} ([F]) \times X}) \sim = {Ext}^{2} (F, F) \otimes_{C} O_{D^{'}}

and

{Tor}_{1}^{O_{Z^{'}}} (M, O_{D^{'}}) \sim = {Ext}_{D^{'}}^{0} (q^{*} F, O_{D} ⊠ F |_{f^{- 1} ([F]) \times X}) \sim = Hom (F, F) \otimes_{C} O_{D^{'}},

implying

\begin{matrix} L \sim =^{n} \oplus_{i = 1} O_{D_{i}} and M \sim =^{n} \oplus_{i = 1} O_{D_{i}} (D_{i}) . \end{matrix}

(4.6)

5. Let

W^{'}

denote the middle cohomology of the complex

0 ⟶ {A^{'}}^{0} - - \to α^{'} φ^{*} A^{1} - - \to β^{'} {A^{'}}^{2} ⟶ 0 .

W^{'}

is locally free of rank

dim Y - 2

. We obtain the following equation of Chern classes in

H^{*} (Z^{'}, Z)

\begin{matrix} φ^{*} c (A^{1} - A^{0} - A^{2}) = c (W^{'} + M - L) . \end{matrix}

(4.7)

On the other hand, as

c (F) = c (G)

H^{*} (X, Z)

, the Grothendieck-Riemann-Roch Theorem yields the following identity in

H^{*} (Y^{'}, Z)

\begin{matrix} c (A^{1} - A^{0} - A^{2}) = c ({Ext}_{p}^{∙} (q^{*} F, F)) = c ({Ext}_{p}^{∙} (q^{*} G, F)) = c (W) . \end{matrix}

(4.8)

Combining 4.7 and 4.8 , we conclude that

\begin{matrix} c (W^{'}) = φ^{*} c (W) \cdot c (L - M) \in H^{*} (Z^{'}, Z) \end{matrix}

(4.9)

Moreover,

c (L - M) =^{n} \prod_{i = 1} \frac{c (O_{D_{i}})}{c (O_{D_{i}} (D_{i}))} =^{n} \prod_{i = 1} \frac{1}{c (O_{Z^{'}} (- D_{i})) c (O_{Z^{'}} (D_{i}))} = 1 + \sum_{k = 1}^{\infty} \sum_{i = 1}^{n} D_{i}^{2 k} .

The product of any cohomology class in

H^{*} (Y^{'}, Z)

of positive degree with any of the classes

D_{i}

is zero. It follows that

c_{2 k} (W^{'}) = φ^{*} c_{2 k} (W) + \sum_{i = 1}^{n} D_{i}^{2 k} for all k > 0 .

The key point now is that both

W

and

W^{'}

are vector bundles of rank

dim (Y) - 2

, so that the Chern classes

c_{dim (Y)} (W)

and

c_{dim (Y)} (W^{'})

vanish (cf. [19] , Lemma 4).

We get the contradiction

0 = \sum_{i = 1}^{n} D_{i}^{dim (Y)} = - n .

This finishes the proof of Theorem 4.1 . □

Theorem 4.4. — Let

v_{0}

be a primitive Mukai vector satisfying condition

(*)

and

(* *)

. Let

v = m v_{0}

and let

H

be a

v

–general ample divisor. Then

M_{v}

is a normal irreducible variety of dimension

2 + 〈 v, v 〉

This theorem is due to Yoshioka [28] in the case of torsion free sheaves. Using the local information obtained in Proposition 3.10 , the basic irreducibility result of Theorem 4.1 , we can give a simple direct proof.

Proof. By Proposition 3.10 , $R^{s s}$ is normal. As a GIT-quotient of a normal scheme, $M_{v}$ is also normal. If $m = 1$ , all points in $M_{v} = M_{v_{0}}$ correspond to stable sheaves and hence $M_{v}$ is smooth. By Theorem 4.1 , $M_{v_{0}}$ is irreducible. By $(* *)$ , $M_{v_{0}}$ is non-empty.
Assume now that $m \geq 2$ and that the assertion of the theorem has been proved for all moduli spaces $M_{m^{'} v_{0}}$ , $1 \leq m^{'} < m$ . For any decomposition $m = m^{'} + m^{''}$ with $1 \leq m^{'} \leq m^{''}$ , consider the morphism $φ (m^{'}, m^{''}) : M_{m^{'} v_{0}} \times M_{m^{''} v_{0}} ⟶ M_{m v_{0}}, ([E^{'}], [E^{''}]) \mapsto [E^{'} \oplus E^{''}],$ and let $Y (m^{'}, m^{''}) \subset M_{v}$ denote its image. The subschemes $Y (m^{'}, m^{''})$ , $1 \leq m^{'} \leq m^{''}$ , are the irreducible components of the strictly semistable locus of $M_{v}$ .
Since all $Y (m^{'}, m^{''})$ are irreducible by induction and intersect in the points of the form $[E_{0}^{\oplus m}]$ , $[E_{0}] \in M_{v_{0}}$ , the strictly semistable locus is connected. Since $M_{v}$ is normal, the connected components are irreducible. In particular, there is exactly one component that meets the strictly semistable locus. Theorem 4.1 excludes the possibility of a component that does not meet the strictly semistable locus. □

5 Factoriality of moduli spaces

Proposition 5.1. — Let

v_{0}

be a primitive Mukai vector satisfying

(*)

Let

v = m v_{0}

for some

m \in N_{0}

. Assume that

– either $E = E_{0}^{\oplus m}$ , for some $E_{0}$ stable with $v (E_{0}) = v_{0}$ , and $H$ is arbitrary,
– or $E$ is arbitrary polystable with $v (E) = v$ , and $H$ is $v$ –general.

Assume further that case (C) applies. Then

M_{v}

is locally factorial at

[E]

if and only if the isotropy subgroup

PGl (N)_{[q]} \sim = PAut (E)

of any point

[q]

in the closed orbit in

π^{- 1} ([E]) \subset R^{s s}

acts trivially on the fibre

L ([q])

for every

PGl (N)

-linearised line bundle

L

on an invariant open neighbourhood of the orbit of

[q]

Proof. This is Drezet’s Theoreme A [10] . In Drezet’s situation the Quot scheme $R^{s s}$ is smooth. However, all his arguments go through under the weaker hypothesis that $R^{s s}$ is locally factorial in a $PGl (N)$ –equivariant open neighbourhood of the closed orbit in the fibre $π^{- 1} ([E])$ . But this is true under the given hypothesis due to Proposition 3.10 □

Corollary 5.2. — Let

E_{0}

be a stable sheaf with Mukai vector

v (E_{0}) = v_{0}

satisfying

(*)

and assume that

v = m v_{0}

satisfies (C). Then

M_{v}

is locally factorial at

[E_{0}^{\oplus m}]

Proof. The isotropy subgroup of any point $[q]$ in the closed orbit in $π^{- 1} ([E_{0}^{\oplus m}]) \subset R^{s s}$ is isomorphic to $PGl (m)$ and therefore has no non-trivial characters. Hence the action of $PGl (m)$ on $L ([q])$ is necessarily trivial (notations as in Proposition 5.1 ). □

Theorem 5.3. — Let

v_{0}

be a primitive Mukai vector satisfying

(*)

and

(* *)

. Assume that

v = m v_{0}

m \in N

, satisfies (C) and let

H

be a

v

–general ample divisor. Then

M_{v}

is locally factorial.

Proof. Let [ E ] ∈ M v be an arbitrary point that is represented by the polystable sheaf E = ⊕ i = 1 s E i ⊕ n i , and let [ q : ℋ → E ] be a point in the closed orbit in π − 1 ( [ E ] ) ⊂ R s s . Since H is v –general, the Mukai vectors of the stable direct summands E i have the form v ( E i ) = m i v 0 , m i ∈ N , ∑ i = 1 s m i n i = m . We repeat the construction in section 2.5 for each of the Mukai vectors m i v 0 , i = 1 , … , s . Note that we can choose a sufficiently large integer k that works for all Mukai vectors simultaneously. Let P i ( z ) = − m i 〈 v 0 , v ( O X ( − z H ) 〉 , N i = P i ( k ) and ℋ i = O X ( − k H ) ⊕ N i . Then N = ∑ i n i N i and ℋ = ⊕ i ℋ i ⊕ n i . Moreover there are parameter spaces R i s s ⊂ Quot X , H ( ℋ i , P i ) with PGl ( N i ) -actions and quotient maps π i : R i s s → M m i v 0 . Finally there is a canonical map Φ : ∏ i R i s s ⟶ R s s , ( [ ℋ i → F i ] ) i ↦ [ ℋ = ⊕ i ℋ i ⊕ n i → ⊕ i F i ⊕ n i ] . Let Z denote the image of Φ . It has the following properties:
- $∙$ By Theorem 4.4 , the moduli spaces $M_{m_{i} v_{0}}$ are irreducible. It follows that the schemes $R_{i}^{s s}$ and $Z$ are irreducible, too.
- $∙$ $Z$ contains the point $[q]$ and as well a point $[q^{'} : ℋ \to E_{0}^{\oplus m}]$ for some stable sheaf $E_{0}$ with $v (E_{0}) = v_{0}$ .
- $∙$ The group $G : = (\prod_{i} Gl (n_{i})) / C^{*} \subset PGl (N)$ fixes $Z$ pointwise. It equals the stabiliser subgroup of $[q]$ and is contained in the stabiliser subgroup of $[q^{'}]$ .
Now let L be a PGl ( N ) –linearised line bundle on R s s . The group G acts on L | Z with a locally constant character, which must in fact be constant, since Z is connected.
Moreover, the action is trivial at the point $[q^{'}]$ according to the proof of Corollary 5.2 . Thus the character is trivial everywhere on $Z$ and in particular at $[q]$ . According to Drezet’s criterion (Proposition 5.1 ), $M_{v}$ is locally factorial at $[E]$ . □

Remark 5.4. It is also known that the moduli space of semi-stable torsion free sheaves on the projective plane is locally factorial by the work of Drezet [9] . However it may be false for other surfaces as has been observed by Le Potier: the moduli space

M_{P^{1} \times P^{1}} (2, 0, 2)

is not locally factorial at the point represented by

O (1, - 1) \oplus O (- 1, 1)

(see [10] , p. 106).

6 Symplectic resolutions

Let

v_{0}

be a primitive Mukai vector satisfying

(*)

and

(* *)

. Let

v = m v_{0}

and let

H

be a

v

–general divisor. Recall that the following three cases are possible:

(A) $m = 1$ .
(B) $m = 2$ and $〈 v_{0}, v_{0} 〉 = 2$ .
(C) $m \geq 3$ , or $m = 2$ and $〈 v_{0}, v_{0} 〉 \geq 4$ .

In case (A) the moduli space

M_{v}

consists only of stable sheaves. It is irreducible and smooth of dimension

2 + 〈 v, v 〉

. Mukai [21] has defined a symplectic structure on

M_{v}

Proposition 6.1. — Assume that

m \geq 2

. The singular locus

M_{v, sing}

M_{v}

is non-empty and equals the semistable locus. The irreducible components of

M_{v, sing}

correspond to integers

m^{'}

1 \leq m^{'} \leq m / 2

, and have codimension

2 m^{'} (m - m^{'}) 〈 v_{0}, v_{0} 〉 - 2

, respectively. In particular,

codim M_{v, sing} = 2

in case (B) and

\geq 4

in case (C).

Proof. Recall the varieties

Y (m^{'}, m^{''})

introduced in the proof of Theorem 4.4 . The union of the

Y (m^{'}, m^{''})

is the strictly semistable locus. The maps

φ (m^{'}, m^{''}) : M_{m^{'} v_{0}} \times M_{m^{''} v_{0}} \to Y (m^{'}, m^{''})

are finite and surjective, hence

\begin{matrix} codim (Y (m^{'}, m^{''})) & = & 2 + m^{2} 〈 v_{0}, v_{0} 〉 - (2 + {m^{'}}^{2} 〈 v_{0}, v_{0} 〉) - (2 + {m^{''}}^{2} 〈 v_{0}, v_{0} 〉) \end{matrix}

\begin{matrix} = & 2 m^{'} m^{''} 〈 v_{0}, v_{0} 〉 - 2 . \end{matrix}

Clearly, the codimension 2 is attained only if

m^{'} = m^{''} = 1

and

〈 v_{0}, v_{0} 〉 = 2

, which is case (B). As

M_{v}

is smooth in all stable points, it remains to show that the strictly semistable points are really singular. For this it suffices to show that

M_{v}

is singular at a generic point

[E = E^{'} \oplus E^{''}] \in Y (m^{'}, m^{''})

, where

E^{'}

and

E^{''}

are stable sheaves with

v (E^{'}) = m^{'} v_{0}

and

v (E^{''}) = m^{''} v_{0}

. In this case,

PAut (E) \sim = C^{*}

{Ext}^{2} (E, E)_{0} \sim = C

, and the Kuranishi map

{Ext}^{2} (E, E)_{0} \to C [{Ext}^{1} (E, E)]^{\land}

is completely described by an invariant function

f \in C [{Ext}^{1} (E, E)]^{\land}

. It follows, that

{\hat{O}}_{M_{v}, [E]} \sim = {(C [{Ext}^{1} (E, E)]^{\land})}^{C^{*}} / (f) .

Now

C^{*}

acts on the four summands of

{Ext}^{1} (E, E) = {Ext}^{1} (E^{'}, E^{'}) \oplus {Ext}^{1} (E^{'}, E^{''}) \oplus {Ext}^{1} (E^{''}, E^{'}) \oplus {Ext}^{1} (E^{''}, E^{''})

with weights

0

1

- 1

, and

0

. It follows that

{Ext}^{1} (E, E) / / C^{*} = {Ext}^{1} (E^{'}, E^{'}) \times C \times {Ext}^{1} (E^{''}, E^{''}),

where

C \subset M (d, C)

is the cone of matrices of rank

\leq 1

and

d = dim {Ext}^{1} (E^{'}, E^{''}) = m^{'} m^{''} 〈 v_{0}, v_{0} 〉 \geq 2 .

Since the quotient of a singular local ring by a non-zero divisor cannot become regular,

{\hat{O}}_{M_{v}, [E]}

is singular. □

Theorem 6.2. — Suppose that

v

belongs to case (C). Then

M_{v}

is a locally factorial symplectic variety of dimension

2 + 〈 v, v 〉

. The singular locus is non-empty and has codimension 4. All singularities are symplectic, but there is no open neighbourhood of a singular point in

M_{v}

that admits a projective symplectic resolution.

Proof. We have already seen that $M_{v}$ is a locally factorial variety. Mukai [21] constructed a non-degenerate 2-form on $M_{v}^{s}$ . This form is closed even if $M_{v}^{s}$ is not projective ([13] Prop. 10.3.2). By Flenner’s theorem [11] this form extends to any resolution of the singularities of $M_{v}$ . Hence the singularities are symplectic in the sense of Beauville [5] . Now let $[E] \in M_{v}$ be a singular point and let $U \subset M_{v}$ be an open neighbourhood of $[E]$ . A projective symplectic resolution of $U$ is a projective resolution $σ : U^{'} \to U$ of the singularities of $U$ such that the restriction of the symplectic form on $M_{v}^{s}$ to $U^{reg}$ extends to a symplectic form on $U^{'}$ . In such a case the morphism $σ$ would have to be semismall according to a result of Kaledin, [14] Lemma 2.11. As the singular locus of $U$ has codimension $\geq 4$ according to Proposition 6.1 , the exceptional locus of $σ$ has codimension $\geq 2$ in $U^{'}$ . On the other hand $O_{M_{v}, [E]}$ is factorial by Theorem 5.3 . This implies that the exceptional locus must be a divisor (see [7] no. 1.40 p. 28). □

Remark 6.3. — 1) The completion of a factorial local ring is not factorial in general. The local rings of the moduli spaces of type (C) provide nice examples of this phenomenon. Pushing the arguments in the previous proof a bit further, one sees that

{\hat{O}}_{M_{v}, [E]} \sim = C [{Ext}^{1} (E^{'}, E^{'}) \oplus {Ext}^{1} (E^{''}, E^{''})]^{\land} \hat{\otimes} B,

where

B

is the completed coordinate ring of the cone

C_{0} \subset C \subset M (d, C)

of traceless matrices of rank

\leq 1

, with

d \geq 4

. But

{\hat{O}}_{M_{v}, [E]}

cannot be factorial: the vertex of

C_{0}

is an isolated singularity of codimension

\geq 6

, and there are two small symplectic resolutions

T^{*} P ({Ext}^{1} (E^{'}, E^{''})) \to C_{0} \leftarrow T^{*} P ({Ext}^{1} (E^{''}, E^{'}))

. We see that in this case

O_{M, [E]}

is factorial due to Theorem 5.3 , but

{\hat{O}}_{M, [E]}

is not. Geometrically, what happens is this: an irreducible Weil divisor becomes reducible after completion; while the whole thing still is a Cartier divisor, some of its newly acquired irreducible components need not be.

2) On the other hand, for polystable sheaves

E_{0}^{\oplus m}

with

E_{0}

stable and

v (E_{0})

satisfying

(*)

, the completed local ring

{\hat{O}}_{M_{v}, [E_{0}^{\oplus m}]}

is factorial. In fact, the proof of proposition 3.7 shows that

{\hat{O}}_{S, [q]}

is factorial. Moreover, the stabiliser is isomorphic to

PGl (m)

hence has no non-trivial characters. Under these conditions one can show that the invariant ring

({\hat{O}}_{S, [q]})^{PGl (m)} ≃ {\hat{O}}_{M_{v}, [E_{0}^{\oplus m}]}

is also factorial.

References

M. Atiyah, I. Macdonald, Introduction to Commutative Algebra, Addison-Wesley.
A. Altman, S. Kleiman, Introduction to Grothendieck Duality Theory. Lecture Notes in Mathematics 146. Springer Verlag 1970.
A. Beauville, Varietes Kahleriennes dont la premiere classe de Chern est nulle. J. Differential Geom. 18 (1983), 755 – 782.
A. Beauville, Arnaud, R. Donagi, La variete des droites d’une hypersurface cubique de dimension 4. C. R. Acad. Sci., Paris, Ser. I 301 (1985), 703 – 706.
A. Beauville, Symplectic singularities. Invent. math. 139 (2000), 541 – 549.
M. P. Cavaliere, G. Niesi, On Serre’s conditions in the form ring of an ideal. J. Math. Kyoto Univ. 21 (1981) 537 – 546.
O. Debarre, Higher dimensional algebraic geometry. Universitext Springer Verlag 2001.
P. Deligne, La conjecture de Weil, II. Inst. Hautes E’tudes Sci. Publ. Math. No. 52 (1980), 137–252.
J.-M. Drezet, Groupe de Picard des varietes de modules de faisceaux semi-stables sur $P_{C}^{2}$ . Annales de l’institut Fourier, 38 no. 3 (1988)
J.-M. Drezet, Points non factoriels des varietes de modules de faisceaux semi-stables sur une surface rationnelle. J. reine angew. Math. 413 (1991), 99 – 126.
H. Flenner, Extendability of differential forms on non-isolated singularities. Invent. math. 94 (1988), 317 – 326.
A. Grothendieck, Cohomologie locale des faisceaux coherents et Theoremes de Lefschetz locaux et globaux. Seminaire de Geometrie Algebrique du Bois-Marie 1962 (SGA 2). North-Holland Publishing Company Amsterdam 1968.
D. Huybrechts, M. Lehn, The Geometry of Moduli Spaces of Sheaves. Aspects of Mathematics E 31. Vieweg Verlag 1997.
D. Kaledin, Symplectic resolutions from the Poisson point of view. To appear in: J. reine angewandte Math. math.AG/0310186.
Y. Kiem, On the existence of a symplectic desingularisation of some moduli spaces of sheaves on a K3 surface. To appear in Comp. Math.
Y. Kiem, J. Choy, Nonexistence of crepant resolution of moduli space of sheaves on a K3 surface.
D. Kaledin, M. Lehn, Local structure of hyperkahler singularities in O’Grady’s examples. math.AG/0405575.
M. Lehn, Ch. Sorger, La singularite de O’Grady. math.AG/0504182.
E. Markman, Generators of the cohomology ring of moduli spaces of sheaves on symplectic surfaces. J. reine angew. Math. 544 (2002), 61 – 82
H. Matsumura, Commutative Algebra. W. A. Benjamin 1970.
S. Mukai, Symplectic structure of the moduli space of sheaves on an abelian or K3 surface. Invent. math. 77 (1984), 101 – 116.
H. Nakajima, Instantons on ALE spaces, quiver varieties, and Kac-Moody-algebras. Duke Math. J. 76 (1994), 365 – 416.
K. O’Grady, Desingularized moduli spaces of sheaves on a K3, J. reine angew. Math. 512 (1999), 49 – 117.
K. O’Grady, A new six-dimensional irreducible symplectic variety. J. Algebraic Geom. 12 (2003), 435 – 505.
A. Rapagnetta, Topological invariants of O’Grady’s six dimensional irreducible symplectic variety. math.AG/0406026.
C. Simpson, Moduli of representations of the fundamental group of a smooth projective variety I. Publ. Math. IHES 79 (1994), 47 – 129.
K. Yoshioka, Moduli Spaces of stable sheaves on abelian surfaces. Mathem. Ann. 321 (2001), 817 – 884.
K. Yoshioka, Twisted stability and Fourier-Mukai Transform. Comp. Math. 138 (2003), 261 – 288.

Dmitry Kaledin, Independent University of Moscow, B. Vlassievski per. 11, Moscow, 119002, Russia E-mail address : kaledin@mccme.ru Manfred Lehn, Fachbereich Physik, Mathematik und Informatik, Johannes Gutenberg–Universitat Mainz, D-55099 Mainz, Germany E-mail address : lehn@mathematik.uni-mainz.de Christoph Sorger, Laboratoire de Mathematiques Jean Leray (UMR 6629 du CNRS), Universite de Nantes, 2, Rue de la Houssiniere, BP 92208, F-44322 Nantes Cedex 03, France E-mail address : christoph.sorger@univ-nantes.fr

D. Kaledin, M. Lehn, and Ch. Sorger