The Intrinsic Normal Cone 6 9 K. Behrend and B. Fantechi 9 1 January 13, 1996 n a J 5 Abstract 1 We suggestaconstructionofvirtualfundamental classesof certain 1 types of moduli spaces. v 0 1 Contents 0 1 0 0 Introduction 2 6 Notations and Conventions . . . . . . . . . . . . . . . . . . . . . . 5 9 / m 1 Cones and Cone Stacks 5 o Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 e Abelian Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 g - Exact Sequences of Cones . . . . . . . . . . . . . . . . . . . . . . . 8 g E-Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 l a Cone Stacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 : v Xi 2 Stacks of the Form h1/h0 15 r The General Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 15 a Application to Schemes . . . . . . . . . . . . . . . . . . . . . . . . 19 3 The Intrinsic Normal Cone 22 Normal Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 The Intrinsic Normal Cone . . . . . . . . . . . . . . . . . . . . . . 27 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 4 Obstruction Theory 32 The Intrinsic Normal Sheaf as Obstruction . . . . . . . . . . . . . 32 Obstruction Theories . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Obstructions for Small Extensions . . . . . . . . . . . . . . . . . . 36 1 5 Obstruction Theories and Fundamental Classes 37 Virtual Fundamental Classes . . . . . . . . . . . . . . . . . . . . . 37 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 6 Examples 45 The Basic Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Fibers of a Morphism between Smooth Stacks . . . . . . . . . . . . 45 Moduli Stacks of Projective Varieties . . . . . . . . . . . . . . . . . 46 Spaces of Morphisms . . . . . . . . . . . . . . . . . . . . . . . . . . 47 7 The Relative Case 48 Bivariant Theory for Artin Stacks . . . . . . . . . . . . . . . . . . 48 The Relative Intrinsic Normal Cone . . . . . . . . . . . . . . . . . 49 0 Introduction Moduli spaces in algebraic geometry often have an expected dimension at each point, which is a lower bound for the dimension at that point. For instance, the moduli space of smooth, complex projective n-dimensional varieties with ample canonical class has expected dimension h1(V,T ) − V h2(V,T ) at a point [V]. In general, the expected dimension will vary with V the point; however, in some significant cases it will stay constant on con- nected components. In the previous example, this is the case if n ≤ 2, for then the expected dimension is −χ(V,T ). In some cases the dimen- V sion coincides with the expected dimension, in others it does so under some genericity assumptions. However, it can happen that there is no way to get a space of the expected dimension; it is also possible that special cases with bigger dimension are easier to understand and to deal with than the generic case. When we have a moduli space X which has a well-defined expected dimension, it can beusefulto beable to constructin its Chow ringa class of theexpecteddimension. ThemainexampleswehaveinmindareDonaldson theory (with X the moduli space of torsion-free, semi-stable sheaves on a surface) and the Gromov-Witten invariants (with X the moduli space of stable maps from curves of genus g to a fixed projective variety). In this paper we deal with the problem of defining such a class in a very general set-up; the construction is divided into two steps. First, given any Deligne-Mumford stack X, we associate to it an alge- braic stack C over X of pure dimension zero, its intrinsic normal cone. X 2 This has nothing to do with X being a moduli space; it is just an intrin- sic invariant, whose structure is related to the singularities of X (see for instance Proposition 3.12). Then, we define the concept of an obstruction theory and of a perfect obstruction theory for X. To say that X has an obstruction theory means, very roughly speaking, that we are given locally on X an (equivalence class of) morphisms of vector bundles such that at each point the kernel of the induced linear map of vector spaces is the tangent space to X, and the cokernelisaspaceofobstructions. Usually,ifX isamodulispacethenithas an obstruction theory, and if this is perfect then the expected dimension is constantonX. Oncewearegivenanobstructiontheory, withtheadditional (technical) assumption that it admits a global resolution, we can define a virtual fundamental class of the expected dimension. An application of the results of this work is contained in a paper [3] by the first author. There Gromov-Witten invariants are constructed for any genus, any target variety and the axioms listed in [4] are verified. We now give a more detailed outline of the contents of the paper. In the first section we recall what we need about cones and we introduce the notion of cone stacks over a Deligne-Mumford stack X. These are Artin stacks which are locally the quotient of a cone by a vector bundle acting on it. We call a cone abelian if it is defined as SpecSymF, where F is a coherent sheaf on X. Every cone is contained as a closed subcone in a minimal abelian one, which we call its abelian hull. The notions of being abelian and of abelian hull generalize immediately to cone stacks. In the second section we construct, for a complex E• in the derived cate- goryD(O )whichsatisfiessomesuitableassumptions(whichwecallCondi- X tion (⋆), see Definition 2.3), an associated abelian cone stack h1/h0((E•)∨). In particular the cotangent complex L• of X satisfies Condition (⋆), so we X can definethe abelian cone stack N := h1/h0((L• )∨), the intrinsic normal X X sheaf. The name is motivated in the third section, where N is constructed X more directly as follows: ´etale locally on X, embed an open set U of X in a smooth scheme W, and take the stack quotient of the normal sheaf (viewed as abelian cone) N by the natural action of T | . One can glue these U/W W U abelian cone stacks together to get N . The intrinsic normal cone C is the X X closed subcone stack of N defined by replacing N by the normal cone X U/W C in the previous construction. U/W In the fourth section we describe the relationship between the intrinsic normal sheaf of a Deligne-Mumford stack X and the deformations of affine 3 X-schemes, showing in particular that N carries obstructions for such de- X formations. With this motivation, we introduce the notion of obstruction theory for X. This is an object E• in the derived category together with a morphism E• → L• , satisfying Condition (⋆) and such that the induced X map N → h1/h0((E•)∨) is a closed immersion. X Anobstruction theoryE• iscalled perfectifE = h1/h0((E•)∨)issmooth over X. So we have a vector bundle stack E with a closed subcone stack C , and to define the virtual fundamental class of X with respect to E• we X simply intersect C with the zero section of E. This construction requires X Chow groups for Artin stacks, which we do not have at our disposal. There are several ways around this problem. We choose to assume that E• is globally given by a homomorphism of vector bundles F−1 → F0. Then C X gives riseto a cone C in F = F−1∨ and weintersect C withthe zero section 1 of F . 1 Another approach, suggested by Kontsevich [11], is via virtual structure sheaves (see Remark 5.4). Thedrawback of that approach is that it requires a Riemann-Roch theorem for Deligne-Mumford stacks, for which we do not know a reference. In the sixth section we give some examples of how this construction can be applied in some standard moduli problems. We consider the following cases: a fiber of a morphism between smooth algebraic stacks, the scheme of morphisms between two given projective schemes, a moduli space for Gorenstein projective varieties. In the seventh section we give a relative version of the intrinsic normal cone and sheaf C and N for a morphism X → Y with unramified X/Y X/Y diagonal of algebraic stacks; we are mostly interested in the case where Y is smooth and pure-dimensional, which preserves many good properties of the absolute case (e.g., C is pure-dimensional). This is not needed in this X/Y paper, but will be applied by the first author to give an algebraic definition of Gromov-Witten classes for smooth projective varieties. ThestartingpointforthisworkwasatalkbyJunLiattheAMSSummer Institute on Algebraic Geometry, Santa Cruz 1995, where he reported on joint work in progress with G. Tian. Their construction, in the complex analytic context, is based on the existence of the Kuranishi map; by using it they define, under suitable assumptions, a pure-dimensional cone in some bundle and get classes of the expected dimension by intersecting with the zero section. Our construction owes its existence to theirs; we started by trying to understand and reformulate their results in an algebraic way, and found 4 stackstobeaconvenient, intrinsiclanguage. Inouropiniontheintroduction of stacks is very natural, and it seems almost surprising that the intrinsic normal cone was not defined before. We find it important to separate the construction of the cone, which can be carried out for any Deligne-Mumford stack, from its embedding in a vector bundle stack. We work completely in an algebraic context; of course the whole paper could be rewritten without changes over the category of analytic spaces. Acknowledgments. This work was started in the inspiringatmosphere of the Santa Cruzconference. A significant part of it was done duringthe authors’ stay at the Max-Planck-Institut fu¨r Mathematik in Bonn, to which both authors are grateful for hospitality and support. The second author is a member of GNSAGA of CNR. Notations and Conventions Unless otherwise mentioned, we work over a fixed ground field k. An algebraic stack is an algebraic stack over k in the sense of [1] or [12]. Unless mentioned otherwise, we assumeall algebraic stacks (in particular all algebraic spaces and all schemes) to be quasi-separated and locally of finite type over k. ADeligne-Mumfordstack isanalgebraicstackinthesenseof[5],inother words an algebraic stack with unramified diagonal. For a Deligne-Mumford stack X we denote by X the big fppf-site and by X the small ´etale site of fl ´et X. The associated topoi of sheaves are denoted by the same symbols. Recall that a complex of sheaves of modules is of perfect amplitude con- tained in [a,b], where a,b ∈ Z, if, locally, it is isomorphic (in the derived category) to a complex Ea → ... → Eb of locally free sheaves of finite rank. 1 Cones and Cone Stacks Cones To fix notation we recall some basic facts about cones. Let X be a Deligne-Mumford stack. Let S = MSi i≥0 be a graded quasi-coherent sheaf of O -algebras such that S0 = O , S1 X X is coherent and S is generated locally by S1. Then the affine X-scheme 5 C = SpecS is called a cone over X. A morphism of cones over X is an X- morphism inducedby a graded morphismof graded sheaves of O -algebras. X A closed subcone is the image of a closed immersion of cones. If C 2 ↓ C −→ C 1 3 is a diagram of cones over X, the fibered product C × C is a cone over 1 C3 2 X. Every cone C → X has a section 0 : X → C, called the vertex of C, and an A1-action (or a multiplicative contraction onto the vertex), that is a morphism γ :A1×C −→ C such that 1. (1,id) C −→ A1×C idց ↓γ C commutes, 2. (0,id) C −→ A1×C 0 ց ↓γ C commutes, 3. A1×A1×C i−d→×γ A1×C m×id ↓ ↓γ A1×C −γ→ C commutes, where m : A1×A1 → A1 is multiplication, m(x,y) = xy. The vertex of C is induced by the augmentation S → S0, the A1-action is given by the grading of S. In fact, the morphism S → S[x] giving rise to γ maps s ∈ Si to sxi. Note that a morphism of cones is just a morphism respecting 0 and γ. 6 Abelian Cones If F is a coherent O -module we get an associated cone X C(F) = SpecSym(F). For any X-scheme T we have C(F)(T) = Hom(F ,O ), T T so C(F) is a group scheme over X. We call a cone of this form an abelian cone. A fibered product of abelian cones is an abelian cone. If E is a vector bundle over X, then E = C(E∨), where E is the coherent O -module of X sections of E and E∨ its dual. Any coneC = Spec Si is canonically aclosed subconeof anabelian Li≥0 cone A(C) = SpecSymS1, called the associated abelian cone or the abelian hull of C. Theabelian hull is a vector bundleif and only if S1 is locally free. Any morphism of cones φ : C → D induces a morphism A(φ) : A(C) → A(D), extending φ. Thus A defines a functor from cones to abelian cones called abelianization. Note that φ is a closed immersion if and only if A(φ) is. Lemma 1.1 A cone C over X is a vector bundle if and only if it is smooth over X. Proof. Let C = Spec Si, and assume that C → X has constant Li≥0 relative dimension r. Then S1 = 0∗Ω is a rank r vector bundle. C is a C/X closed subcone of A(C) = (S1)∨, hence by dimension reasons C = A(C). 2 If E and F are abelian cones over X, then any morphism of cones φ : E → F is a morphism of X-group schemes. If E and F are vector bundles, then φ is a morphism of vector bundles. Example If X → Y is a closed immersion with ideal sheaf I, then MIn/In+1 n≥0 is a sheaf of O -algebras and X CX/Y = SpecMIn/In+1 n≥0 is a cone over X, called the normal cone of X in Y. The associated abelian cone N = SpecSymI/I2 is also called the normal sheaf of X in Y. X/Y More generally, any local immersion of Deligne-Mumford stacks has a normal cone whose abelian hull is its normal sheaf (see [14], definition 1.20). 7 Exact Sequences of Cones Definition 1.2 A sequence of cone morphisms i 0−→ E −→ C −→ D −→ 0 is exact if E is a vector bundle and locally over X there is a morphism of cones C → E splitting i and inducing an isomorphism C → E×D. Remark Given a short exact sequence 0 −→ F′ −→ F −→ E −→0 of coherent sheaves on X, with E locally free, then 0 −→ C(E) −→ C(F′) −→ C(F) −→ 0 is exact, and conversely (see [6], Example 4.1.7). Lemma 1.3 Let C → D be a smooth, surjective morphism of cones, and let E = C × X; then the sequence D,0 0−→ E −→ C −→ D −→ 0 is exact. Proof. Write C = Spec Si, D = Spec S′i. We start by proving that L L 0−→ E −→ A(C) −→ A(D) −→ 0 is exact. By base change we may assume S′i = 0 for i ≥ 2. The cone E = SpecSymE is a vector bundle because it is smooth. On the other hand, E = Spec (Si/S′1Si−1). As C → D is smooth and surjective, S1 → S′1 is L injective. So we get an exact sequence 0−→ S1 −→ S′1 −→ E −→ 0. To complete the proof, remark that C → A(C)× D is a closed im- A(D) mersion, and both these schemes are smooth of the same relative dimension over C. 2 8 E-Cones If E is a vector bundle and d : E → C a morphism of cones, we say that C is an E-cone, if C is invariant under the action of E on A(C). We denote the induced action of E on C by E ×C −→ C (ν,γ) 7−→ dν +γ . A morphism φ from an E-cone C to an F-cone D (or a morphism of vector bundle cones) is a commutative diagram of cones d E −→ C φ ↓ ↓ φ d F −→ D. If φ : (E,d,C) → (F,d,D) and ψ : (E,d,C) → (F,d,D) are morphisms, we call them homotopic, if there exists a morphism of cones k : C → F, such that 1. kd= ψ−φ, 2. dk = ψ−φ. Here the second condition is to be interpreted as saying that φ+dk = ψ. (More precisely, we say that k is a homotopy from φ to ψ.) Remark A sequence of cone morphisms with E a vector bundle i 0−→ E −→ C −→ D −→ 0 is exact if and only if C is an E-cone, C → D is surjective, and the diagram σ E ×C −→ C p ↓ ↓φ φ C −→ D is cartesian, where p is the projection and σ the action. 9 Proposition 1.4 Let (C,0,γ) and (D,0,γ) be algebraic X-spaces with sec- tions and A1-actions and let φ : C → D be an A1-equivariant X-morphism, which is smooth and surjective. Let E = C × X. Then C is an E-cone D,0 over X if and only if D is a cone over X. Moreover, C is abelian (a vector bundle) if and only if D is. Proof. Let us first assume that C is an abelian cone, C = SpecSymF. The morphism E → C gives rise to F → E∨, where E is the coherent O - X modules of sections of E. Note that F → E∨ is an epimorphism, since E → C is injective. Let G be the kernel, so that 0−→ G −→ F −→ E∨ −→ 0 is a short exact sequence. Then 0−→ E −→ C −→ C(G) −→ 0 is a short exact sequence of abelian cones over X, so D ∼= C(G) and so D is an abelian cone. In general, C ⊂ A(C) is defined by a homogeneous sheaf of ideals I ⊂ SymS1,whereS = Si andC = SpecS. LetF = S1 andletG asabovebe L the kernel of F → E∨. Let J = I∩SymG, which is a homogeneous sheaf of ideals in SymG, so C′ = SpecSymG/J is a cone over X. By construction, C′ is the scheme theoretic image of C in C(G). Hence C′ is the quotient of C by E and so C′ ∼= D and D is a cone. Now for the converse. The claim is local in X. So since D is affine over X we may assume that C = D×E as X-schemes with A1-action. Then we are done. 2 Cone Stacks Let X be, as above, a Deligne-Mumford stack over k. We need to define the 2-category of algebraic stacks with A1-action over X. Definition 1.5 Let C be an algebraic stack over X, together with a section 0 : X → C. An A1-action on (C,0) is given by a morphism of X-stacks γ :A1×C −→ C and three 2-isomorphisms θ , θ and θ between the 1-morphisms in the 1 0 γ following diagrams. 10