AN INTRODUCTION INTO (MOTIVIC) DONALDSON–THOMAS THEORY SVENMEINHARDT Abstract. Theaimofthepaperistoprovidearathergentleintroductioninto 6 Donaldson–Thomastheoryusingquiverswithpotential. Thereadershouldbe 1 familiarwithsomebasicknowledgeinalgebraicorcomplexgeometry. Thetext 0 containsmanyexamplesandexercisestosupporttheprocessofunderstanding 2 themainconcepts andideas. n a J 8 Contents 1 1. Introduction 1 ] G 2. The problem of constructing a moduli space 3 A 2.1. Moduli spaces 3 2.2. Stability conditions 7 . h 2.3. Moduli stacks 8 t a 3. Quiver representations and their moduli 14 m 3.1. Quivers and C-linear categories 14 [ 3.2. Quiver moduli spaces and stacks 15 4. From constructible functions to motivic theories 19 1 4.1. Constructible functions 19 v 4.2. Motivic theories for schemes 21 1 3 4.3. Motivic theories for quotient stacks 24 6 5. Vanishing cycles 27 4 5.1. Vanishing cycles for schemes 27 0 5.2. Vanishing cycles for quotient stacks 34 . 1 6. Donaldson–Thomas theory 36 0 6.1. Definition and main results 36 6 6.2. Examples 39 1 : 6.3. The Ringel–Hall algebra 43 v 6.4. Integration map 45 i X 6.5. The wall-crossing identity 46 r References 47 a 1. Introduction The theory ofDonaldson–Thomasinvariantsstartedaround2000withthe seminal workofR.Thomas[33]. Heassociatedintegerstothosemodulispacesofsheaveson a compact Calabi–Yau 3-fold which only contain stable sheaves. After some years, K.Behrendrealizedin[1]thatthesenumbers,originallywrittenas“integrals”over algebraiccycles or characteristicclasses,canalso be obtainedby anintegralovera constructiblefunction, the so-calledBehrendfunction, withrespectto the measure given by the Euler characteristic. This new point of view did not only extend the theory to non-compact moduli spaces but revealed also the “motivic nature” of this new invariant. It has also been realized that quivers with potential provide 1 2 SVENMEINHARDT another class of examples to which Donaldson–Thomas theory applies. Starting around2006,D.Joyce[11],[12],[13],[14],[15],[16]andY.Song[17]extendedthethe- ory using all kinds of “motivic functions” to produce (possibly rational) numbers evenin the presence of semistable objects which is the genericsituation when clas- sifying objects inabeliancategories. Aroundthe same time, M.KontsevichandY. Soibelman [19],[21],[20] independently proposed a theory producing even motives, some sort of refined “numbers”, instead of simple numbers, also in the presence of semistable objects. The technical difficulties occurring in their approach disap- pear in the special situation of representations of quivers with potential. The case of zero potential has been intensively studied by M. Reineke in a series of papers [28],[29],[30]. Despite some computations of motivic or evennumericalDonaldson– Thomas invariants for quivers with potential (see [2],[7],[5],[25]), the true nature ofDonaldson–Thomasinvariantsfor quiverwithpotentialremainedmysteriousfor quite some time. A full understanding has been obtained recently and is the con- tent of a series of papers [4],[6],[24],[23]. ThepresenttextaimsatgivingagentleintroductiontoDonaldson–Thomastheory in the case of quiver with potential. We have two reasons for our restriction to quivers. Firstly, so-called orientation data will not play any role, and secondly, we donotneedtotouchderivedalgebraicgeometry. Apartfromthis, manyimportant ideas and concepts are already visible in the case of quiver representations, and since the theory is fully understood, we belief that this is a good starting point for your journey towards an understanding of Donaldson–Thomas theory. There are more survey articles available focusing on different aspects of the theory. (see [17],[20],[32]) Let us give a short outline of the paper. The next section starts very elementary by discussing the problem of classifying objects. The objects which are of inter- est to us form an abelian category although many ideas of section 2 also apply to “non-linear” moduli problems. We study in detail the difficulties arising from the construction of moduli spaces and develop slowly the concept of a (moduli) stack. Although the theory of stacks is rather rich and complicated, we can restrict our- selves to quotient stacks throughout this paper. Hence, a good understanding of a quotient stacks is inevitable. We try to illustrate this concept by giving impor- tant examples. We should mention that only very little knowledge of algebraic or complexgeometryis needed. Inmanycases,youcaneasilyreplace“schemes”with “varieties” or “complex manifolds”. Section 3 provides the backgroundon quivers andtheir representations. The point of view taken here is that quivers are the categorical (noncommutative) analogue of polynomial algebras in ordinary commutative algebra. In other words, they are a useful tool for practical computations when dealing with linear categories, but at the end of the day the result should only depend on the linear category and not its presentation as a quotient of the path category of a quiver by some ideal of relations. The relations important in this paper are given by noncommutative partial derivatives of a so-called potential. ThenexttwosectionsprovidethelanguageandtheframeworktoformulateDonaldson– Thomas theory in section 6. We start in section 4 with the concept of “motivic theories”. The best example the reader should have in mind are constructible functions. It should be clear that constructible functions can be pulled back and multiplied. Using fiberwise integrals with respect to the Euler characteristic, we can even push forward constructible functions. Moreover, every locally closed subscheme/subvariety/submanifolddeterminesaconstructiblefunction,namelyits characteristicfunction. Inanutshell,amotivictheoryisjustageneralizationofthis associating to every scheme X an abelian group R(X) of “functions” on X which AN INTRODUCTION INTO (MOTIVIC) DONALDSON–THOMAS THEORY 3 can be pulled back, pushed forward and multiplied. Moreover, to every locally closed subscheme in X there is a “characteristic function” in R(X) such that the characteristic function of a disjoint union is the sum of the characteristic function of its summands. Its is this property what makes a theory of generalizedfunctions “motivic”. Asusualinalgebraicgeometry,theterm“function”shouldbeusedwith some care. Everyfunction onsaya complex variety X determines a usualfunction from the set of points in X to the coefficient ring R(point) of our theory, but this is not a one-one correspondence. In section 5 we introduce vanishing cycles. We do not assume that the reader is familiarwithanytheoryofvanishingcycles. Asintheprevioussection,avanishing cycle is just an additional structure on motivic theories formalizing the properties of ordinary classical vanishing cycles. The Behrend function mentioned at the be- ginning of this introduction provides a good example of a vanishing cycle on the theory of constructible functions. In fact, we will construct in a functorial way two vanishing cycles associated to a givenmotivic theory. The first construction is rather stupid, but the second one essentially covers all known nontrivialexamples. At the end of sections 4 and 5 we extend motivic theories and vanishing cycles to quotient stacks as quotient stacks arise naturally in moduli problems. There is a way to circumvent stacks in Donaldson–Thomas theory by considering framed ob- jects, but we belief that the usual approachof using stacks is more conceptual and should be known by anyone who wants to understand Donaldson–Thomas theory seriously. In the last section 6 we finally introduce Donaldson–Thomas functions and invari- ants. After stating the main results, we consider many examples to illustrate the theory. Finally, we develop some tools used in Donaldson–Thomas theory such as Ringel–Hall algebras, an important integration map and the celebrated wall- crossing formula. The reader will realize shortly that the text contains tons over exercises and ex- amples. Most of the exercises are rather elementary and require some elementary computations and standard arguments. Nevertheless, we suggest to the reader to do them carefully in orderto getyour hands on the subject andto obtaina feeling about the objects involved. There is a lot of materialin this text which is not part ofthe standardgraduatecoursesatuniversities,andifyouarenotalreadyfamiliar withthe subjectyoucertainlyneedsome practiceaswe cannotprovidea deepand lengthy discussion of the material presented here. Acknowledgments. The paper is an expandedversionof a couple of lectures the author hasgivenin collaborationwith Ben Davisonat KIASin February2015. He is more than grateful to Michel van Garrel and Bumsig Kim for giving him the opportunitytovisitthiswonderfulplace. Alotofworkonthispaperhasalsobeen done atthe University of Hong Kong,where the author gaveanother lecture series on Donaldson–Thomas theory. The author wants to use the opportunity to thank MattYoungforbeingsuchawonderfulhost. HealsowantstothankJanManschot for keeping up the pressure to finish this paper and for offering the opportunity to publishthepaper. Finally,theauthorisverygratefultoMarkusReinekeforgiving him as much support as possible. 2. The problem of constructing a moduli space 2.1. Moduli spaces. Let us start by recalling the generalidea of a moduli space. Dependingonthesituation,mathematiciansaretryingtoclassifyobjectsofvarious types. The general pattern is the following. There is some set (or class) of objects 4 SVENMEINHARDT and isomorphisms between two objects. Such a structure is called a groupoid. A groupoid is a category with every morphism being an isomorphism. If the set of objects has cardinality one, a groupoid is just a group. The other extreme is a groupoidsuch that every morphismis the identity morphism of some object. Such groupoids are in one-to-one correspondence with ordinary sets. Hence, a groupoid interpolates between sets and groups. There are two main sources of groupoids. Example 2.1. Let X be a topological space. The fundamental groupoid π (X) 1 is the groupoid having the points of X as objects, and given two points x,y X, ∈ the set of morphisms from x to y is the set of homotopy classes of paths from x to y. Fixing a base point x X, the usual fundamental group π (X,x) is just the 1 ∈ automorphism group of x considered as an object in the groupoid π (X). Denote 1 by π (X)the setofpath connectedcomponents,i.e. the set ofobjects in π (X)up 0 1 to isomorphism. Example 2.2. Given a category , one can considerthe subcategory so( ) of all C I C isomorphismsin . Thus, so( ) is a groupoid,and / denotes the setof objects ∼ C I C C in up to isomorphism. C These two examples are related as follows. To every (small) category one can con- struct a topological space X - the classifying space of - such that π (X ) = C C 1 C ∼ so( ) and π (X )= / . 0 C ∼ I C C Let us come back to the classification problem, say of objects in up to isomor- C phism. The problem is to describe the set / . If it is discrete in a reasonable ∼ C sense, one tries to find a parameterization by less complicated (discrete) objects. This applies for instance to the classification of semisimple algebraic groups or fi- nite dimensional representations of the latter. In many other situations, / is ∼ C uncountable, and one wants to put a geometric structure on the set / to obtain ∼ C a “moduli space”. However, if for instance / has the cardinality of the field of ∼ C complex numbers, one can always choose a bijection / =M to the set of points C ∼ ∼ of any complex variety or manifold M of dimension greater than zero. Pulling back the geometric structure of M along this bijection, we can equip / in many ∼ C different(non-isomorphic)wayswitha structureofa complexmanifold. Hence,we should ask: Question: Is there a natural geometric structure on / ? What does “natural” ∼ C actually mean? There is a very beautiful idea of what “natural” should mean, and which applies to many situations. Assume there is a notion of a family of objects in over C some “base” scheme/variety/(complex) manifold S, i.e. some object on S which has “fibers” over s S, and these fibers should be objects in . ∈ C Example2.3. GivenaC-algebraA,afamilyoffinitedimensionalA-representations is a (holomorphic) vector bundle V on S and an C-algebra homomorphisms A → Γ(S, nd(V)) from A into the algebra of sections of the endomorphism bundle of E V. Example 2.4. Given a scheme/variety/manifold X over C and some parameter space S, a family of coherent sheaves on X parametrized by S is just a coherent sheafEonS X whichisflatoverS. Thelatterconditionensuresthattakingfibers × andpull-backsoffamiliesbehaveswell. IfE isafamilyofzerodimensionalsheaves onX,i.e.ifthe projectionp:Supp(E) S has zero-dimensionalfibers, flatnessof → E over S is equivalent to the requirement that E is locally free over S. Using the coherenceofE oncemore,onecanshowthatp:Supp(E) S isafinitemorphism → AN INTRODUCTION INTO (MOTIVIC) DONALDSON–THOMAS THEORY 5 and if X = SpecA is affine, E is completely determined by the vector bundle V := p E on S together with a C-algebra homomorphism A Γ(S, nd(V)). ∗ → E From that perspective, the previous example can be seen as a non-commutative version,namelyfamiliesofzerodimensionalsheavesonthe non-commutativeaffine scheme SpecA for A being a C-algebra. Example 2.5. A G-homogeneous space with respect to some (algebraic) group G is a scheme P with a right G-action such that P = G as varieties with right ∼ G-action, where G acts on G by right multiplication. A (locally trivial) family of G-homogeneous spaces over S is defined as a principal G-bundle on S. Once a family is given, by taking the “fiber” over s S we get an object in ∈ C and, hence, a point in / . Varying s S, we end up with a map u : S / . ∼ ∼ C ∈ → C Moreover,we see that the pull-back of a family on S along a morphism f :S′ S → induces a morphism u′ : S′ / such that u′ = u f. Coming back to the ∼ → C ◦ question formulated above, we can now be more precise by asking: Question: Is there a structure of a variety or scheme on / such that for ev- ∼ C ery family of objects over any S, the induced map S / is a morphism of ∼ → C schemes? If so, is there any way to get back the family by knowing only the mor- phism S / ? ∼ →C If the first question has a positive answer, we call = ( / ,scheme structure) ∼ M C a coarse moduli space for . If the second part of the question is also true, we C should be able to (re)construct a “universal” family on by considering the map M id : . Moreover, given a map u′ : S′ such that u′ = u f for M → M → M ◦ some map f : S′ S, the family on S′ should be the pull-back of the family on → S associated to u by uniqueness. As every morphism u : S has an obvious → M u id factorization S , we finally see that every family on S must be the −→ M −→ M pull-back of the “universal” family on . In such case, we call a fine moduli M M space. Example2.6. Let =VectCbethecategoryoffinitedimensionalC-vectorspaces. C A (locally trivial) family of finite dimensional vector spaces is just a vector bundle onsomeparameterspaceS. Asavectorspaceisclassifiedbyitsdimension,wecan put the simplest scheme structure on VectC/∼ = N by thinking of N as a disjoint union of countably many copies of SpecC. Given a vector bundle V, we obtain a well-definedmorphismS N mapping s S to the copyofSpecC indexedbythe → ∈ dimension of the fiber V of V at s. The scheme N is a course moduli space, but s apart from the zero dimensional case, it can never be a fine moduli space. Indeed, there isanobviousandessentiallyunique vectorbundle onN inducing the identity map N N, but a vector bundle on S can never be the pull-back of the one on N → unless it is constant. Thus, N is not a fine moduli space. These are also bad news for our previous examples concerning representations of an algebra A or sheaves on a variety X. Indeed, for A=C or X =SpecC, we are back in the classification problem of finite dimensional C-vector spaces. Example 2.7. Similartothepreviousexample,weseethattheclassificationprob- lem for homogeneous G-spaces has only a coarse moduli space given by SpecC. Thereareseveralstrategiesto overcomethe difficulty ofconstructingafine moduli space. Example 2.8 (rigidfamilies). Onepossibilityistorigidifyfamilies ofobjects. For example, instead of considering all vector bundles we could also restrict ourselves 6 SVENMEINHARDT to constant vector bundles. In this particular case, N is even a fine moduli space. However,inmanysituations onewants toglue families togetherto formfamilies of objects on bigger spaces. This is incompatible with the concept of rigidity, and we will not follow this path. Example 2.9 (weaker equivalence). Instead of classifying objects up to isomor- phism, we could allow weaker equivalences. For example, we could identify to families V(1) and V(2) (over S) of vector spaces or representations of an algebra A if there is a line bundle L on S such that V(2) = V(1) L. By doing this, we ⊗OS can always replace a rank one bundle with the trivial rank one bundle. Hence, the moduli space SpecC of one-dimensional vector spaces is a fine moduli space. Example 2.10(projectivization). Similartofamiliesofvectorspacesofdimension r, one could look at locally trivial families of projective spaces Pr−1. The tran- P sition functions between local trivializations are regular functions with values in Aut(Pr−1)=PGL(r). Every vector bundle V of rank r provides such a bundle by taking :=P(V),thebundleoflinesorhyperplanesinV. TwovectorbundlesV(1), P V(2) define isomorphic bundles P(V(1))=P(V(2)) if and only if V(2) =V(1) L ∼ ⊗OS for some line bundle L on S, providing the bridge to the previous example. How- ever, not every Pr−1-bundle can be realized as P(V) for some vector bundle V P on S. Given a Pr−1-bundle , there is an associated locally trivial bundle of P P E C-algebras isomorphic to EndC(Cr) ∼= MatC(r,r). Conversely, every locally trivial bundle of C-algebrasisomorphicto MatC(r,r) defines an associatedPr−1-bundle E E asthetransitionfunctionsof mustbeinAut(MatC(r,r))=PGL(r). Thus,we P E have an equivalence of categories between locally trivial Pr−1-bundles and locally trivialMatC(r,r)-bundles. If the Pr−1-bundle is givenby P(V) for a vector bun- P dle V of rank r, then P(V) = nd(V). Given a C-algebraA, we can study families E E givenbyalocallyfreePr−1-bundle orequivalentlyalocallyfreeMatC(r,r)-bundle P and a homomorphism of algebras A Γ(S, ). If A = C, there is only a fine E → E moduli space for r = 1 as every P0-bundle must be constant. If the algebra A is more complicated, there are also fine moduli spaces for r >1, but only for objects which are simple in a suitable sense. For A= C there are no simple vector spaces of dimension r >1. As we have seen, the construction of fine moduli spaces can only be done in a few cases and severe restrictions. But even if we were only interested in coarse moduli spaces, a standard problem will occur as the following example shows. Example 2.11. Instead of looking at representations of A=C, we enter the next level of complexity by looking at finite dimensional representations of A=C[z]. A one-dimensionalrepresentationV isdetermined bythe valueofz in EndC(V)∼=C. In other words, a coarse moduli space is given by the complex affine line A1. Still, we have to face the problemdiscussedbefore that a line bundle onS with z acting by multiplication with a fixed number c C could almost never be the pull-back ∈ of a universal family under the constant map S A1 mapping s S to c A1. → ∈ ∈ Let us ignore the problem of finding a fine moduli space and continue with two- dimensional representations. Consider the trivial rank 2 bundle on S =A1 with z acting via the nilpotent matrix 1 s 0 1 (cid:18) (cid:19) in the fiber over s S = A1. The representations for s = 0 are all isomorphic ∈ 6 to each other, and our “classifying map” u : S to a coarse moduli space 2 → M of rank 2 representations must be constant on S 0 . For s = 0 we ob- 2 M \{ } tain a different representation and u(0) must be another point in if the latter 2 M AN INTRODUCTION INTO (MOTIVIC) DONALDSON–THOMAS THEORY 7 parametrizes isomorphism types. However, such a discontinuous map u:S 2 →M cannot exist, and we have to abandon the idea of finding a coarse moduli space parameterizing isomorphism classes. One can show that a “reasonable” coarse moduli space is given by the GIT-quotient MatC(2,2)//GL(2) which is realized as SpecC[MatC(2,2)]GL(2) ∼= A2 and similar for higher ranks. The classifying map S A2 will map s S to the unordered pair of eigenvalues of the z-action in → ∈ the fiber over s. Such an unordered pair of eigenvalues is determined by the sum (corresponding to the trace) and the product (corresponding to the determinant) of the eigenvalues and similar for higher ranks. Therefore, will parametrize un- M ordereddirectsums ofone-dimensionalrepresentations. Inotherwords,bypassing from / to ,weidentifyeachrepresentationwiththe(unordered)directsumof ∼ C M its simple Jordan–Ho¨lderfactors. Representations having the same Jordan–Ho¨lder factors, i.e. corresponding to the same point in , are often called S-equivalent1. M Let us summarize the lessons we have learned in the previous examples: (1) Constructing coarse moduli spaces has only a chance if we do not param- etrize objects up to isomorphism but up to the weaker S-equivalence. In otherwords,classifyingobjects up to isomorphismis only possible for sim- ple objects, i.e. objects without subobjects. (2) Theconstructionofauniversalfamilyonthemodulispaceofsimpleobjects mightonlyworkifweidentifytwofamiliesunderaweakerequivalence(twist with a line bundle) or pass to some projectivization. We suggest to the reader to check these statements in the previous examples. 2.2. Stability conditions. Eventhoughthesetofobjectsin uptoisomorphism C mightbe verylarge,the setof(isomorphismclassesof)simple objectscanbe quite small, even finite. Thus, the “coarse”moduli space would not deliver much insight into the set of isomorphism types in . However, there is a simple but clever idea C to overcome this problem. Instead of looking at , we should “scan” by means C C of a collection ( ) of “small” full subcategories . An object which µ µ∈T µ C C ⊆ C might be far away from being simple or semisimple (direct sum of simples) can become semisimple or even simple in . By doing this, we can distinguish many µ C S-equivalent objects either because they live in different subcategories or they live in the same subcategory but have different Jordan–Ho¨lder filtrations taken in µ C . This brilliant idea is the essence of the concept of stability conditions. The µ C following definition is due to Tom Bridgeland. However, there are more general definitions of stability conditions. Definition 2.12. (1) A central charge on a noetherian abelian category is a function Z on the C set of objects in with values in H := rexp(√ 1φ) C r 0,φ + C { − ∈ | ≥ ∈ (0,π] such that Z(E)= 0 implies E = 0 and Z(E) = Z(E′)+Z(E′′) for } every short exact sequence 0 E′ E E′′ 0. → → → → (2) Given a central charge Z, we call an object E semistable if ∈C argZ(E′)) argZ(E) for all subobjects E′ E. ≤ ⊂ (3) For µ ( ,+ ] we denote with the full subcategory of all semistable µ ∈ −∞ ∞ C objects E of slope cot(argZ(E)) = µ and the zero object. It turns out − that is an abelian subcategory of (cf. Exercise 3.12). µ C C 1The“S”in“S-equivalent”referstosemisimple,i.e.sumsofsimples,andshouldnotbeconfused withournotationofabaseofafamily. 8 SVENMEINHARDT (4) A simple object in is called stable. We assume that every semistable µ C object of slope µ has a Jordan–H¨older filtration with stable subquotients of thesameslope. Semisimple objects of , i.e. sumsof stableobjects of slope µ C µ, are called polystable. (5) Every object E in has a unique filtration 0 E E ... E = 1 2 n C ⊂ ⊂ ⊂ ⊂ E, the Harder–Narasimhan filtration, with semistable quotients E /E of i i−1 strictly decreasing slopes. Example 2.13 (The r-Kroneckerquiver). Letus illustrate this idea with a simple example. Considertheabeliancategoryofr-tuplesx¯oflinearmapsx :V V for i 1 2 → 1 i r between finite dimensional vector spaces V ,V . Choosing two complex 1 2 ≤ ≤ numbers ζ ,ζ H , we get a central charge by putting Z(x¯) = ζ dimV + 1 2 + 1 1 ∈ ζ dimV . Assume first that arg(ζ )=arg(ζ ). Then, all objects are semistable of 2 2 1 2 thesameslopeµ= cot(argζ ),andwehavetofacethe oldproblems. Choosefor 1 − instancedimV =dimV =1. Theisomorphismtypeofsuchobjectsisdetermined 1 2 bythechoiceofrcomplexnumbersx ,...,x uptorescalingby(g ,g ) GL(V ) 1 r 1 2 1 ∈ × GL(V ) = C∗ C∗ via g x g−1. As the diagonal group (g,g) g C∗ acts 2 × 1 i 2 { | ∈ } trivially, we have the take the GIT quotient of Ar by C∗ which is just a point as SpecC[x ,...,x ]C∗ = SpecC. This corresponds to the fact that all objects 1 r have the same Jordan–Ho¨lder factors x = 0 : V 0 and x = 0 : 0 V . i 1 i 2 → → 0 Thus, all objects are S-equivalent to “V V ”= V V . If argζ > argζ , 1 2 1 2 2 1 ⊕ −→ non of our objects with dimV = dimV = 1 are semistable as the central charge 1 2 ζ of the subobject 0 : 0 V has a bigger argument than the central charge 2 2 → ζ +ζ of our given object. If, however, argζ < argζ , all objects except for the 1 2 2 1 semisimple V V corresponding to x = 0 for all 1 i r are semistable of 1 2 i slope µ= e(⊕ζ +ζ )/ m(ζ +ζ ), and evenstable. T≤he m≤oduli space ζ1,ζ2 = −ℜ 1 2 ℑ 1 2 M(1,1) Ar 0 /C∗ = Pr−1 parameterizing isomorphism classes of simple objects in of µ \{ } C dimensionvector(dimV ,dimV )=(1,1)is evena fine moduli space ifwe identify 1 2 two families of r-tuples of line bundle morphisms x : V V on a parameter i 1 2 → spaceS assoonasthey become isomorphicaftertwisting V andV withsome line 1 2 bundle L. Note that coarse moduli spaces parameterizing S-equivalence classes of objects in mightnotexistforallcentralcharges,butonecanshowtheexistenceforgeneric µ C central charges and reasonable abelian categories. We should also keep in mind that we paid a price for getting a refined version of S-equivalence, namely S-equivalence in subcategories. Indeed, coarse moduli spaces of (S-equivalence classes of) semistable objects can only “see” semistable objectsbutnoobjectswithanon-trivialHarder–Narasimhanfiltration. Hence,the construction of (coarse) moduli spaces remains unsatisfying. 2.3. Moduli stacks. There is, however, an alternative way to overcome all the problems seen in the previous examples. Following this approach, one can con- struct a fine moduli “space” with a universal family parameterizing all objects - not only simple or stable ones - up to isomorphism. According to the conserva- tive law of mathematical difficulties, we also have to pay a price for getting such a beautifulsolutionofourmoduliproblem. Itishiddenintheword“space”. Infact, we have to leaveour comfort zone of varieties or schemes and have to dive into the universe of more general spaces known as “Artin stacks”. Recall that a scheme X is uniquely characterized by its set-valued functor h : X S Mor(S,X) of points. We have seen many set-valued functors before while 7→ studying moduli problems. The general pattern was the following. We considered set-valued contravariantfunctors F :S families of objects in / and a fine ∼ 7−→{ C} AN INTRODUCTION INTO (MOTIVIC) DONALDSON–THOMAS THEORY 9 modulispacewouldbeascheme suchthatF =h ,whileacoarsemodulispace M ∼ M is a scheme together with a mapF h whichis universalwith respectto all M M → maps F h of functors. One possibility of generalizing the concept of a space X → is to consider set-valued functors satisfying similar properties like the functor h . X Note that if one has a collection of morphisms U X defined on open subsets U i i → covering S such that the maps agree on overlaps,one can glue the maps to form a global morphism S X. This sheaf property should also be satisfied by a general → set-valued functor to be a reasonable generalization of a scheme. Such set-valued functors are also often called “spaces”. A generalized space is called algebraic if it can be written as the “quotient” X/ of a scheme X by an (´etale) equivalence ∼ relation. In other words, algebraic spaces are not to far away from schemes and many results for schemes can be generalized to algebraic spaces. In our situation of forming moduli spaces, this is still not the right approach to take, but shows alreadyintotherightdirection. Indeed,the problemsarisingintheconstructionof universalfamiliesarerelatedtothepresenceof(non-trivial)automorphisms. Thus, we should take automorphisms and isomorphisms more seriously into account. Recallthat a set with isomorphismsbetween points wasjust a groupoidstudied at the beginning of this section. Hence, instead of looking at set-valued functors on the category of schemes, we should consider groupoid-valued contravariant func- tors. These functors should satisfy some gluing property which looks a bit more complicated than in the set-theoretic context. The best idea of remembering the gluingpropertyisbylookingatanexamplewhichis-asbefore-the babyexample for all Artin stacks. Example2.14. Considerthegroupoid-valuedfunctorVectwhichmapsanyscheme S to the groupoidof vector bundles (the objects) and isomorphisms between them (the morphisms). By pulling back vector bundles along morphisms f :S′ S, we → get indeed a contravariant functor.2 Given two vector bundles V,V′ and an open cover U =S of S together with isomorphisms3 φ :V V′ on the open ∪i∈I i i |Ui → |Ui subsets U suchthat they agreeafter restrictionto the overlaps,i.e. φ =φ i i|Uij j|Uij with U = U U , we can always find a unique global isomorphism φ : V V′ ij i j ∩ → such that φ = φ . On the other hand, if we have vector bundles V on U i |Ui i i and isomorphisms φ : V V such that the only possible composition ij i|Uij → j|Uij V V V V of their restrictions to the triple overlaps i|Uijk → j|Uijk → k|Uijk → i|Uijk U = U U U is the identity (cocycle condition), one can use the transition ijk i j k ∩ ∩ isomorphisms φ to glue the V together, i.e. there is a vector bundle V on S and ij i a family of isomorphisms φ : V V such that the only possible composition i |Ui → i V V V V of their restrictions with φ is the identity. This |Uij → i|Uij → j|Uij → |Uij ij was the gluing property for isomorphisms and objects, and if we replace the word “vector bundle” with “object”, we get the general form of the gluing property for a groupoid-valued functor. Definition 2.15. A stack is a groupoid-valued contravariant functor4 on the cate- gory of schemes satisfying the gluing property for isomorphisms and objects as seen in Example 2.14 In that perspective, a stack is like a (generalized) space with set-valued functors replaced with groupoid-valued functors. 2Strictly speaking, we only get a pseudofunctor as g∗◦f∗ is only equivalent to (f ◦g)∗, but wewillignorethistechnical problemasonecanalwaysresolveit. 3Wewillalwaysdenote thepull-backalonganinclusionU ֒→S ofanopensubsetby|U. 4Again, weignorethe fact that g∗◦f∗ mightonlybe equivalent to(f ◦g)∗ forapair S′′ −→g S′−→f S ofcomposablemorphisms. 10 SVENMEINHARDT Exercise 2.16. Thinking of a set as a special groupoid with no nontrivial isomor- phisms, show that every generalized space is a stack. Exercise 2.17. Fix a C-algebra A. Show that the functor A-Rep associating to every scheme S the groupoid of vector bundles V with algebra homomorphisms A Γ(S, nd(V))(theobjects)andisomorphismsofvectorbundlescompatiblewith → E thealgebrahomomorphisms (themorphisms) isastack. Provethesameforbundles of matrix algebras and algebra homomorphisms A Γ(S, ) as in Example 2.10. E → E Exercise 2.18. Fix a scheme/variety/manifold X over C. Show that the functor CohX associating to every scheme S the groupoid of coherent sheaves E on S X × flatoverS (theobjects)andisomorphisms betweenthem(themorphisms)isastack. Exercise 2.19. Fix an algebraic group G. Show that the functor SpecC/G as- sociating to every scheme the groupoid of principal G-bundles (the objects) and isomorphisms between them (the morphisms) is a stack. Example 2.20. Thefollowingexampleisageneralizationofthepreviousexercise. Fix analgebraicgroupGandaschemeX witha(right)G-action. Thereis astack X/G associating to every scheme S the groupoid of pairs (P S,m : P X), → → whereP S isaprincipalG-bundleandm:P X isaG-equivariantmap,with → → morphismsbeinggivenbyG-bundleisomorphismsu:P P′satisfyingm′ u=m. → ◦ The pull-back along a morphism f : S′ S is given by (S′ P S′,m pr ). → ×S → ◦ P The morphism m : P X can also be interpreted as a section of the X-bundle → P X S. The stack X/Gis calledthe quotientstack of X with respect to the G × → G-action. When is comes to locally trivial families, there is some choice involved, namely the choice of the underlying (Grothendieck) topology. Intuitively, one would start with the Zariski topology, but the ´etale or even the smooth topology have their advantages, too. In fact, the quotient stack X/G defined above is usually taken with respect to the smooth or, equivalently, ´etale topology. However, for so-called “special” groups G like GL(n) we could equivalently take the Zariski topology as every´etale locally trivial principal G-bundle is then already Zariski locally trivial. Notice that PGL(d) is not special and we should better take the ´etale topology when it comes to principal PGL(d)-bundles and quotient stacks X/PGL(d). Definition2.21. A 1-morphism (or morphism for short) from astackF toastack F′ is a natural transformation η : F F′, i.e. a family of functors η : F(S) S → → F′(S) compatible with pull-backs along f : S′ S up to equivalence of functors. → In other words, the functors F′(f) η and η F(f) from F(S) to F′(S′) are S S′ ◦ ◦ equivalent. A 2-morphism α:η η′ between 1-morphisms is an invertible natural → transformation α : η η′ for every scheme S, compatible with pull-backs. In S S → S particular, given two stacks F,F′, we get a groupoid of morphisms Mor(F,F′) with 1-morphisms being the objects and 2-morphisms being the morphisms. Hence, the category of stacks is a 2-category. Thinking of a set as being a groupoid having only identity morphisms, we can associate to every scheme X a contravariant functor h : S Mor(S,X). As we X 7→ canglue morphisms, h is indeed a stack. The following lemma is veryimportant. X Lemma 2.22 (Yoneda-Lemma). The covariant functor h:X h from schemes X 7→ to stacks provides a full embedding of the category of schemes into the (2-)category of stacks. Moreover, there is an equivalence of groupoids Mor(h ,F) = F(X) for X ∼ every scheme X and every stack F, natural in X and F. Exercise 2.23. Try to prove the Yoneda-Lemma.