Patrick Lei

All notes

Good moduli spaces, positivity, and rationality

Lectures by Maksym Fedorchuk, Boston College, Spring 2026.

Abstract.

This course will be an introduction to the theory of good moduli spaces (in the sense of Alper). We will generalize filtrations from a vector space to an algebraic variety, define theta-reductivity, S-completeness, and unpuncturedness of inertia. Examples will include good moduli spaces for Deligne-Mumford stable curves, and K-semistable Fanos, with the focus on the cases where there is a positive tautological line bundle and the good moduli space is particularly simple, i.e., rational.

1 Moduli of rational curves

We will begin by interpolating between M¯0,n (which is a scheme) and the stack (ℙ1)n⫽PGL⁡(2) (which has a good moduli space by definition) via Hassett spaces M¯0,A. There is a single smooth rational curve, which is ℙ1={x=0}⊆ℙx,y,z2.

Viewing the curve inside ℙ2, we may (see Figure 1) consider the family

𝔸t1t≠0t=0t≠0Smooth Fiber (x⁢y=t2)Smooth Fiber (x⁢y=t2)Singular Fiber (x⁢y=0)Singular PointFamily: x⁢y−t2⁢z2=0In Affine Patch (z=1):x⁢y=t2The map π:𝒳→𝔸t1is a fibration withnodal degeneration at t=0.
Figure 1. The family {x⁢y−t2⁢z2=0} over 𝔸1

This morphism is a (flat) isotrivial degeneration, and will be the first example of a test configuration. What this means is that there is an action of 𝔾m on 𝔛 by the formula

such that 𝔛→𝔸1 is equivariant. More generally, any tree of rational curves joined by nodes is also a rational curve (as in it has genus 0).

Remark 1.1.

A reduced rational curve is a “tree” of ℙ1’s meeting in m-fold rational points, where each singularity is locally analytically isomorphic to the union of the coordinate axes in 𝔸m.

In fact, the definition of a rational curve tells about what kinds of singularities it has, and then there is a classification of curve singularities. Another viewpoint is that if we degenerate a rational normal curve in ℙm to a cone over a hyperplace section, we get a union of m lines meeting at a point in exactly the way described above.

We will now consider the functor 𝒰~0,n of n-pointed smooth rational curves. In particular, a family 𝔛→B must have n sections σ1,…,σn:B→𝔛, which are not required to be disjoint. We may now consider the larger functor ℳ~0,n of n-pointed reduced rational curves with sections σi avoiding the singularities of the fibers. Unfortunately, this is extremely non-separated, so we will consider stability conditions.

Define the functor M¯0,n of n-pointed rational curves with at worst nodal singularities such that all sections σi⁢(B) are disjoint and tha4t the relative dualizing sheaf

is relatively ample over B. One reason we consider this (and that it makes sense) is that the m-fold rational point is Gorenstein if and only if m≤2. This stability condition was introduced by Deligne and Mumford. In fact, the condition on the relative dualizing sheaf is equivalent to enforcing that the automorphism group of each fiber is finite, but is easier to modify for obvious reasons. From either perspective, it is clear that we must have a tree of ℙ1’s with at least three special points on each component. One example can be seen in Figure 2.

CcenterCleftCrightx1x2x3x4x5x6x7Configuration in M¯0,7Dual Graph: Tree (Genus 0)Stability: Each component has ≥3special points (nodes + marks).
Figure 2. An example of a stable curve of genus 0 with marked points

One way to turn a family with an m-fold rational point into a nodal one is to blow up the singular point. The remaining conditions are non-disjoint sections, which can be separated by finitely many blowups. Finally, the unstable components can be contracted by the basic theory of surfaces (they have self-intersection either −1 or −2). In particular, we have sketched a proof that M¯0,n is universally closed. In fact, it is also separated, but we will omit the proof here.

Suppose we have a family π:𝔛→B in ℳ~0,n. Then we have n+1 line bundles on B where n of them are given by

(this is fine even in the non-Gorenstein case because the sections avoid the singularities) and the last is given by

which counts singularities. For example, if B is a smooth projective curve, then

Theorem 1.2.

The line bundle

is ample on M¯0,n.

Remark 1.3.

Projectivity of M¯0,n can be proved in more synthetic ways, for example via the explicit blowup construction of Keel.

1.1 Hassett moduli spaces.

There is a modification of this, which is the functor M¯0,A, where

Here, we consider n-pointed rational curves which are at worst nodal, but where the sections σi1,…,σik can meet if and only if

and the condition on the dualizing sheaf is that ω𝔛/B⁢(∑iαi⁢σi) is ample. Projectivity of this is proven using similar methods to M¯0,n.

More specifically, the functor is given by

This functor is actually representable by a scheme because the three conditions imply that there are no automorphisms.

Example 1.4.

If A=(1,…,1) or more generally if the sum of any two weights is strictly greater than 1, then M¯0,A=M¯0,n.

Example 1.5.

If A=(1,ε,…,ε) such that 1−(n−1)⁢ε>2 and ε≤1n−2, then the combinatorics imply that all marked points lie on the same component. If we call the heavy marked point ∞, then all of the light markings lie in 𝔸1. By translation invariance, we may assume that

and therefore we obtain

Remark 1.6.

There exists a morphism M¯0,n→M¯0,A given (when the base is a curve) by changing the weights and contracting all of the (−1)-curves in the fibers where the total weight of the sections has weight at most 1.

Example 1.7.

If we set A=(ε,…,ε) such that ε≈2n+ε′, then the curve must be a ℙ1 with n points such that we cannot have ⌈2n⌉ colliding. More concretely, if n=2⁢m+1 and ε=1m, then we have a ℙ1 with 2⁢m+1 marked points such that at most m can collide. In particular, we have

where the linearization is the tensor product of the 𝒪⁡(1) on each factor. For an example when m=7, see Figure 3.

ℙ1Triple{x1,x2,x3}w=1Pair{x4,x5}x6x7Space: M¯0,(ε7), ε=13Topology: All stable curves are irreducible (smooth).Reason: Splitting the curve requires Wleft>1 and Wright>1.Total weight is 73. If Wleft≥43 (min 4 pts),then Wright≤1, which is unstable.Coincidences: Up to 3 points may coincide (3×13≤1).
Figure 3. An object in M¯0,(310,310,310,310,310,310,310)

1.2 Semistable objects.

Now we will define a new moduli space

by relaxing the ampleness condition. This can happen if we have sections whose weights sum to exactly 1.

Example 1.8.

If n=2⁢m and A=(1m,…,1m), then we have

This is not a scheme and does not even have a coarse moduli space because the curve where the first m markings collide and the last m markings collide has an entire 𝔾m of automorphisms. For an example when n=4, see Figure 4.

ℙ1𝔾m Action (Dilation)z↦λ⁢z{x1,x2}(Coincident) {x3,x4}(Coincident) Configuration: 4 points, weights wi=1/2.Geometry: Smooth curve (ℙ1) with 2 effective marked locations.Automorphisms: Fixing 2 points on ℙ1 leaves 1 degree of freedom(scaling), hence Aut ≅𝔾m.Note: This is strictly semistable if ∑wi=2.
Figure 4. A strictly semistable object in M¯0,12∗.

The Hassett moduli spaces have a wall-and chamber-structure, where for example, if A=(α,…,α) is Sn-symmetric, then M¯0,A is constant for α∈(1k+1,1k), but on the walls we only have M¯0,A∗. The main observation is that if we write M¯0,α for M¯0,(α,…,α), then there are open immersions

If we consider the complement M¯0,1k∗∖M¯0,1k−ε, then it consists of curves which contain an irreducible component which is a tree in the dual graph and contains k marked points, so this part of the moduli space has a factor which describes the rest of the curve and one which is [𝔸k−2/𝔾m]. More precisely, we have

This actually allows us to construct a test configuration, where if we have ℙ1×𝔸1 with a bunch of sections, we can blow up a point on the central fiber and then contract the central (−1)-curve with all of the sections.

Exercise 1.9.

Figure out what M¯0,1k∗∖M¯0,1k+ε is.

For now, we will denote this picture by ℳ−⊂ℳ⊃ℳ+. Then in fact we actually have

where the second factor comes from the place where k of the sections collide. In this picture, we can degenerate out a new rational tail which has only the k colliding sections (which has a ℂ×-action), which is another example of a test configuration, as in Figure 5.

C1 (Stable)C2 (Unstable)Nodex1x2⋯xk−1xkxk+1=⋯=x2⁢k
Figure 5. A curve that we bubble out

Later, when we discuss good moduli spaces, there is a diagram

where the rightward arrow in the bottom row is a divisorial contraction.

In fact, there is a series of divisorial contractions

and another series of divisorial contractions

1.3 Divisors on M¯0,n.

There are boundary divisors ΔI,J where I⊔J={1,…,n} where the curve degenerates into two components, one with the markings from I and the other with the markings from J. It is clear that

We will denote their union by Δ and call it the total boundary.

It is easy to see that M¯0,n∖Δ≕M0,n consists only of smooth curves and is given by

This has no Picard group, so the entire Picard group of M¯0,n is generated by the boundary divisors. For example, M¯0,3 is a point, M¯0,4=ℙ1, and M¯0,5 is ℙ1×ℙ1 blown up at three points.

Exercise 1.10.

Compute the Picard rank of M¯0,n.

If we consider a 1-dimensional base B, then basic the theory of surfaces tells us that ψi|B=−σi2, and if we blow down all rational tails in 𝒞 to form a new ℙ1-bundle 𝒟→B which has a section σ1 with self-intersection −r and a number of other sections σ~i with self-intersection +r. If we now blow up the other collision points, then we see that

In particular, because M¯0,n is smooth, we have derived Keel’s relation, which is that for all i≠j, then

These in fact generate all relations.

We now consider the morphism fn+1:M¯0,n+1→M¯0,n given by deleting the last marked point and stabilizing. This is in fact the universal family over M¯0,n, so in particular it has n disjoint sections σ1,…,σn. From this perspective, we can construct M¯0,5 by blowing up the points on ℙ1×ℙ1 where the diagonal intersects the three horizontal sections corresponding to 0, 1, and ∞. If we blow down the four (−1)-curves corresponding to the diagonal section and the three fibers that we blew up, the family becomes the pencil of conics passing through four points on ℙ2 and the reducible fibers are in fact the three degenerate members (see Figure 6). In particular, we see that Pic⁡(M¯0,5)=ℤ5.

P1P2P3P4
Figure 6. Pencil of conics through four points in general position on ℙ2. The three degenerate conics correspond to pairs of lines through these points (dashed).

Another way to see this is to consider the sequence

and use this to compute the Picard rank, which is

A more precise description is that

The blowup construction (really the one that starts with ℙn−3) proves that the Picard group also has no torsion.

One way to find a basis for the Picard group is the following. We will call ΔI⊔J non-adjacent if I is not an arc when we arrange 1,…,n in a circle.

Proposition 1.11.

Non-adjacent boundary divisors form a basis of Pic⁡(M¯g,n).

Proof.

We use Keel’s relations to rewrite every divisor as a linear combination of non-adjacent ones, and then we check that the number of non-adjacent ones is exactly the Picard rank.

For example, if we have I={1,2}, we will use the relations for ψ1+ψ2, ψ3+ψn (with a plus sign), ψ1+ψn, and ψ2+ψ3 (with a minus sign), then the left hand side vanishes while the only adjacent boundary divisor remaining is the one with I={1,2}. The general case of this is left as an exercise, as is counting the non-adjacent boundary divisors. □

We have discussed the fact that M¯0,n is a projective variety. We know it is proper, but we would like to produce an ample line bundle. One candidate is

The first step to proving that this is ample is to sum all of Keel’s relations, which gives us

Because |I|⁢|J|≥2⁢(n−2)>n−1 (here n≥4), we see that

where all coefficients are positive.

We will now use the ampleness criterion and show that (ψ−Δ).B>0 for any irreducible proper curve B↪M¯0,n. If B intersects the interior nontrivially, then by definition it has to intersect the boundary with positive degree. The other case is when B↪ΔI⊔J lives inside one of the boundary divisors.

In this case, we have

Restricting ψ−Δ, we obtain

so we induct using the fact that on M¯0,4≃ℙ1, ψ−Δ=𝒪⁡(1).

2 Geometric invariant theory

2.1 Groupoids and coarse moduli.

We will now recall the Yoneda embedding, which is the fully faithful functor

Definition 2.1.

A scheme G is a group scheme if the functor hG factors through the embedding 𝖦𝗋𝗉↪𝖲𝖾𝗍.

Example 2.2.
  1. (1)

    There is the group scheme 𝔾m=S×ℤ⁡[x,x−1] which sends a scheme T to H0⁢(T,𝒪T)×.

  2. (2)

    There is the group scheme 𝔾a=S×ℤ⁡[x] which sends a scheme T to H0⁢(T,𝒪T) with addition.

Definition 2.3.

A scheme G is a group scheme if there exist morphisms

satisfying the usual group axioms:

  1. (1)

    The diagram

    commutes (associativity);

  2. (2)

    The diagram

    commutes (inverse axiom).

  3. (3)

    The diagram

    commutes (identity axiom).

Example 2.4.

If k is algebraically closed, then a (Zariski)-closed subgroup of GLn⁡(k) is a group scheme over k, and is known as a linear algebraic group.

Definition 2.5.

An algebraic group over k is a smooth group scheme over k, and if it is affine, we call it linear.

Definition 2.6.

An action of an algebraic group G on a scheme X is a morphism σ:G×X→X such that the diagram

commutes and the diagram

commutes.

On k-points over an algebraically closed field, this says that g⁡(g′⁢(x))=(g⁢g′)⁢(x) and that e⋅x=x which are the usual group action axioms.

Construction 2.7 (Action groupoid).

Suppose that G acts on X. Define R≔G×X, s≔pr2:R→X (source morphism), and t≔σ:R→X (target morphism). Then set j≔(s,t):R→X×X. There is a morphism

giving composition. Finally there is the morphism

and the reflexivity morphism

This construction gives a groupoid in schemes called the action groupoid associated to the action of G on X.

More generally, if we replace G×X by an arbitrary R, we obtain the general notion of a groupoid in the category of schemes. Our goal will be to prove the following theorems.

Theorem 2.8.

Let j:R→X×X be a flat groupoid such that the stabilizer j−1⁢(ΔX)→ΔX≃X is finite. Then there exists an algebraic space “quotient” X/R and a quotient morphism q:X→X/R which is a uniformly categorical and geometric quotient. Moreover, if j is finite, then X/R is separated.

Definition 2.9.

Let σ:G×X→X be a group action. Then ϕ:X→Y is a geometric quotient if the square in the diagram

commutes and the dashed arrow is unique for any commutative diagram where the ψ exists. The categorical quotient is uniform if it is preserved by flat base change and universal if it is preserved by arbitrary base change.

Naively, we may form a quotient pre-sheaf

where x∼x′ if there exists g∈G⁡(T) such that g⋅x=x′. However, this is not a sheaf in general (and is definitely not a scheme), so we need to improve our constructions.

2.2 Good quotients.

Let S=Spec⁡k for an algebraically closed field k and G be an algebraic group over k. Suppose that X is a scheme with an action

of G. Now suppose we have another action

on another scheme X′.

Definition 2.10.

A morphism f:X→X′ is G-invariant if the diagram

commutes.

In the finite-type case, this reduces to the equation

for all k-points g∈G⁡(k) and x∈X⁡(k).

Definition 2.11.

If G acts trivially on Y, then a G-equivariant f:X→Y is G-invariant.

In particular, f has to collapse all orbits of G to points.

Definition 2.12.

A morphism ϕ:X→Y is a categorical quotient of X by G if it is G-invariant and for any other G-invariant morphism ψ, there exists a unique g:y→y′ such that ψ=g∘ϕ, or in other words the dashed arrow in

exists and is unique.

Exercise 2.13.

If X has property 𝒫 being one of normal, irreducible, reduced, connected, or locally factorial, and if ϕ:X→Y is a categorical quotient, then Y also has 𝒫.

Remark 2.14.

If G acts on X and U⊂X is a G-invariant open subset, then the group G⁡(k) acts on 𝒪X⁢(U), which is a k-algebra.

Definition 2.15.

Suppose that G acts on X. A morphism π:X→Y is a good quotient if

  1. (1)

    It is G-invariant;

  2. (2)

    π is affine and surjective;

  3. (3)

    If U⊂Y is open, the morphism of rings 𝒪Y⁢(U)→𝒪X⁢(π−1⁢(U))G is an isomorphism;

  4. (4)

    If W1 and W2 are disjoint G-invariant closed subsets in X, then π⁡(W1)∩π⁡(W2)=∅ and the two images are closed.

Proposition 2.16.

Every good quotient is a categorical quotient.

Proof.

Suppose that π:X→Y is a good quotient and ϕ:X→Z is G-invariant. We will construct a morphism g:Y→Z filling in the dashed arrow in

Let {Vi} be a cover of Z by affine open subschemes.

The first claim we will prove is that there exists an open Ui⊂Y such that π−1⁢(Ui)=ϕ−1⁢(Vi). If x∈Vi is a closed point and we set W=Z∖Vi, then ϕ−1⁢(x) and ϕ−1⁢(W) are disjoint G-invariant closed subsets of X. This is true for all x∈Vi, so we see that ϕ−1⁢(Vi)=π−1⁢(Ui) for some Ui⊂X. Because Y∖Ui=π⁡(ϕ−1⁢(W)) is closed, Ui must be open.

We now treat Ui as a subset of Y. We need to construct a morphism gi:Ui→Vi=Spec⁡Ai. There is a map

which is the desired morphism. These morphisms glue, so we are done. □

We will denote good quotients by X⫽G.

Lemma 2.17.

If π:X→X⫽G is a good quotient, then for all closed y∈X⫽G, the fiber π−1⁢(y) contains a unique closed G-orbit.

Proof.

Consider the orbit G⋅x⊂π−1⁢(y) of the smallest dimension. Then G⋅x¯∖G⋅x is a union of orbits, so it must be empty. □

Corollary 2.18.

The set X⫽G⁡(k) is in bijection with the set of closed orbits in X.

Definition 2.19.

A good quotient is geometric if all orbits of closed points are closed. Equivalently, X⫽G⁡(k) is in bijection with the set of G⁡(k)-orbits in X⁡(k). A third equivalent definition is that G⋅x1¯∩G⋅x2¯=∅ for all x1,x2∈X⁡(k) in disjoint orbits.

Lemma 2.20.

Suppose that f:X→Y is finite and G-equivariant. Then G⋅x is closed if and only if G⋅f⁡(x) is closed.

Proof.

Note that equivariance implies that G⋅f⁡(x)=f⁡(G⋅x). If G⋅x is closed, then so is G⋅f⁡(x) by finiteness. If G⋅x is not closed and there exists x′∈G⋅x¯∖G⋅x but G⋅f⁡(x) is closed, then dimG⋅x′<dimG⋅x. The sequence of containments

implies that G⋅f⁡(x′)=G⋅f⁡(x), contradicting the dimension inequality. □

Definition 2.21.

Let X be a G-scheme. A G-invariant subset U⊆X is saturated if for all x∈U and y∈X such that G⋅x¯∩G⋅y¯≠∅, then y∈U.

Example 2.22.

Suppose we have f1,…,fn∈𝒪X⁢(X)G. Then the open set

is saturated.

Exercise 2.23.

Suppose that π:X→X⫽G is a good quotient. Then a G-invariant subset U⊆X is saturated if and only if U=π−1⁢(π⁡(U)).

One major statement we want to prove is the following.

Theorem 2.24 (Affine GIT).

Suppose that G is a linearly reductive group acting on an affine scheme X=Spec⁡R. Then

is a universal good quotient.

Definition 2.25.

A group scheme G is linearly reductive if every finite-dimensional rational representation of G is completely reducible.

Remark 2.26.

In characteristic zero, a group G is reductive (equivalently, linearly reductive), if and only if BG=[∙/G]→∙ is a good moduli space.

Definition 2.27.

Let V be a finite-dimensional G-representation. Consider the vector space

By complete reducibility, there is a direct sum decomposition V=VG⊕V′, so we define a map

by projecting onto the first factor. This is clearly G-equivariant, surjective, and idempotent, and is called the Reynolds operator.

By Schur’s lemma, the diagram

commutes.

By gluing Reynolds operators on finite-dimensional G-representations together, we obtain a Reynolds operator on infinite-dimensional rational representations, and in particular on k-algebras R as

Remark 2.28.

Note that RG is a k-subalgebra of R.

Let G be a linear algebraic group. Then if G=Spec⁡S (note here that S is a Hopf algebra), an action

of G on an affine scheme Spec⁡R corresponds to a coaction

Composing this with with g∗:S→k corresponding to g∈G⁡(k), we obtain an action

This defines a k-linear action

Definition 2.29.

G⁡(k) acts rationally on R if every element of R lies in a finite-dimensional G⁡(k)-invariant subspace of R.

Lemma 2.30.

The action G⁡(k)×R→R coming form the action of G on Spec⁡R is rational.

Idea of proof.

For any r∈R, consider the coaction

Then r lies in the span of the ri, which is finite-dimensional and G⁡(k)-invariant. □

This clarifies what we meant by gluing Reynolds operators together. By complete reducibility, we can write R=RG⊕R′.

Claim 2.31.

R′ is an RG-module and the direct sum decomposition holds at the level of RG-modules.

Proof.

We need to prove that RG⁢R′⊆R′. Choose a finite-dimensional subrepresentation V⊆R′ and r∈R′. We want to prove that r⁢V⊆R′. Now consider the sequence

Because V is a direct sum of nontrivial representations and RG is a trivial representation, the composition is zero by Schur’s lemma. □

Now consider the Cartesian diagram

Corollary 2.32.

For all RG-algebras T, we have

  1. (1)

    A canonical isomorphism (T⊗RGR)G≅T.

  2. (2)

    Suppose that {Ii} are G-invariant ideals of R. Then

Proof.
  1. (1)

    Because R≃RG⊕R′ as RG-modules, we have

    Applying the Reynolds operator, we see that

  2. (2)

    We use the fact that ρ⁡(∑Ii)=∑ρ⁡(Ii). □

Proof of Theorem 2.24.

The morphism is clearly G-invariant and affine. We will begin with surjectivity. Choose T=k, so we are looking at a k-point of Spec⁡RG. We only need to prove that k⊗RGR is nonzero, but we already know that (k⊗RGR)G=k is nonzero. In particular, this means that every fiber is nonempty.

Now let Spec⁡T be an open affine in Spec⁡RG. Now we see that

and this holds for all open affines, so it extends to an isomorphism

of sheaves.

Now let Wi=V⁡(Ii) be disjoint G-invariant closed subsets of X. Then I1 and I2 are G-invariant ideals, and in particular

This implies that

so the closures are disjoint. Finally, if we choose W2=π−1⁢(pt), we see that π⁡(W1) is closed. □

Now if we have a G-invariant ideal I⊆R, the sequence

is also exact. In particular, if we have a G-invariant closed subscheme Z↪X, then Z→π⁡(Z) is also a good quotient and the diagram

commutes.

Lemma 2.33.

If X=Spec⁡R is Noetherian, then so is Y=X⫽G=Spec⁡RG.

Proof.

If W⊆Y is closed, then π⁡(π−1⁢(W))=W, so ascending chains of ideals downstairs correspond to ones upstairs. Here, we use the fact that

and thus (R⁢J)G=J for any ideal J⊆RG. □

Proposition 2.34.

If R is a finitely-generated k-algebra, then so is RG. Equivalently, if X is finite type, then so is X⫽G.

Proof.

We first consider the case that

is a graded ring with R0=k and G preserves the grading. In this case, RG⊆R is also graded and is Noetherian, and is thus finitely generated.

In the second case, G acts rationally on R. Suppose that R is generated by x1,…,xn. Then the generators are contained in some finite-dimensional G-subrepresentation V⊆R, so conclude using the diagram

to show that RG is finitely-generated. □

2.3 Numerical criteria for stability.

Now suppose that σ:G×X→X is a group action. Consider the morphism

If jx:Spec⁡k→X is the inclusion of a point, then consider the Cartesian diagram

This defines a morphism ψx:G→X and G⋅x=Im(ψx) is the orbit of x.

Definition 2.35.

The action σ is called:

  1. (1)

    closed if all maps ψx are closed;

  2. (2)

    separated if ψ itself is closed;

  3. (3)

    proper if ψ is proper;

  4. (4)

    free if ψ is a closed embedding.

If we consider the Cartesian diagram

then in fact the fibers of IX→X are the stabilizers of points in X.

Lemma 2.36.

If σ is proper, then all ψx are proper. This further implies that all orbits are closed and all stabilizers are proper.

Remark 2.37.

There are examples where the converse is false, even when all orbits are closed and stabilizers are trivial.

Proof.

Use the valuative criterion of properness. □

Definition 2.38 (1-PS).

A 1-parameter subgroup of an algebraic group G is a morphism

Example 2.39.

If G=GLn, then one example of a 1-parameter subgroup is the set of matrices of the form

in some basis.

Proposition 2.40.

An action of a reductive group on a finite type and separated1 X over k is proper if and only if for every nontrivial 1-parameter subgroup

the induced action of 𝔾m on X is proper.

Theorem 2.41 (Iwahori).

For a 1-PS λ:𝔾m→G and a DVR R with fraction field K, denote by ⟨λ⟩K the K-point

induced by sending t to a uniformizer π of R. Then

is a union of double cosets.

Theorem 2.42.

If X=Spec⁡R is finite type and G is linearly reductive, then the action is proper if and only if all stabilizers are finite.

Remark 2.43.

Note that finite stabilizers imply that all orbits are closed.

Corollary 2.44.

If X⫽G=X/G is a geometric good quotient by a reductive group, then the action is proper.

Remark 2.45.

If π:X→Y is G-invariant and surjective then the following are equivalent:

  1. (1)

    Im(ψ)=X×YX⊂X×X;

  2. (2)

    All fibers of π are orbits.

In particular, if this holds and Y is separated, then the diagram

is Cartesian, so ψ is separated.

Proof of Proposition 2.40.

We apply the valuative criterion. Let R be a DVR with fraction field K. Consider the diagram

Then we know that

by assumption. Also consider the Iwahori decomposition

for ψi∈G⁡(R). Now consider a 1-PS λ:𝔾m→G. Then we compute

Therefore the morphism 𝔾m×X→X×X is proper, so ⟨λ⟩∈𝔾m⁢(R)⊂G⁡(R). □

Proposition 2.46.

Suppose that 𝔾m acts on X such that the action is induced from an action of 𝔾m on ℙkn and a locally closed 𝔾m-equivariant embedding X↪ℙkn. Then the action is proper if and only if all stabilizers are finite.

Proof of Theorem 2.42.

If a good quotient exists and all orbits are closed with finite stabilizers, then X⫽G=X/G=Spec⁡RG is geometric. Therefore, σ is separated, which implies that all 1-PS actions are proper, and thus that σ is proper. □

Proof of Proposition 2.46.

𝔾m=Spec⁡k⁡[t,t−1] acts on ℙn=Proj⁡k⁡[x0,…,xn] by the xi↦tαi⁢xi for αi∈ℤ. Choose R=k⁢⟦t⟧ and K=k⁡((t)). Now the calculation with the diagram

becomes very explicit. We assume that ϕ⁢ζ,ζ∈X⁡(R)⊂ℙn⁡(R). Write ζ=[a0:⋯:an]∈ℙn⁡(R) for ai∈R. Then we see that

where ϕ=A∈K×=𝔾m⁢(K). Because (ϕ⁢ζ,ϕ)∈Im(ψ), so is (ϕ⁢ζ⁢(0),ζ⁡(0)) by separatedness. Therefore, (ϕ⁢ζ)⁢(0)=ϕ0⁢(ζ⁡(0)) for some ϕ0∈𝔾m⁢(k)=k×. Now there exists some ϕ~∈𝔾m⁢(R) such that ϕ~⁢(0)=ϕ0. Replace ζ∈X⁡(R) by ζ~=ϕ~−1⁢ζ∈X⁡(R). Therefore,

and thus (ϕ⁢ζ~,ζ~)∈X⁡(R)×X⁡(R) satisfies

Now write

Clearly, we have a~i⁢(0)=Aαi⁢ai~⁢(0), so if a~i≠0 and αi≠0 , then A is a unit. But the point

has finite stabilizer under the action with weights α0,…,αn, so the desired i exists. □

Example 2.47 (Non-proper action, Mumford).

Let G=SL⁡(2) and

The action will be induced from the natural action of SL⁡(2) on ℂ2.

Exercise 2.48.

There is a map

which is SL⁡(2)-invariant with π−1⁢(π⁡(x))=SL⁡(2)⋅x. Moreover, all stabilizers are trivial.

There is a locus

which is closed. Then we see that

which is not closed. Therefore, the orbit space map cannot be any kind of good quotient.

This example is morally the same the example of the action of 𝔾m⋊ℤ2 on 𝔸2∖0, where 𝔾m acts with weights 1 and −1 and ℤ2 acts by swapping the coordinates. This has a map to 𝔸1 which separates orbits, but this is still not a good quotient because even though the diagonal Δ of (𝔸2∖0)×(𝔸2∖0) is closed, Ψ⁡(Δ) is not closed.

In the language of stacks, in fact thre is an isomorphism

where ℤ2 acts by swapping the two origins and x↦−x. This stack does not have a good moduli space which is a scheme, but it does have a good moduli space which is a non-separated algebraic space by Keel-Mori.

Example 2.49 (Hironaka).

There is a proper but non-projective threefold X with a free action of 𝔾m which is not proper, and the quotient is not a scheme but is an algebraic space.

2.4 Linearizations and GIT.

Let X be finite type and separated over an algebraically closed field k, ℒ be a line bundle on X, and σ:G×X→X be an action.

Definition 2.50.

A G-linearization of ℒ is an action

on the total space of ℒ such that the diagram

commutes and the induced map

is a morphism of line bundles over G×X.

Example 2.51.

If G acts on V rationally by some morphism G→GL⁡(V), then the induced action on ℙ⁡(V) comes with a linearization of 𝒪ℙ⁡(V)⁢(1), which comes from the fact that the total space of 𝒪ℙ⁡(V)⁢(1) is the blowup of V at the origin, and the action on V lifts to an action on the blowup (since we blew up a fixed point).

Definition 2.52.

Let X be finite type over an algebraically closed field k, G act on X, and ℒ be a G-linearized line bundle on X. A point x∈X⁡(k) is semistable if there exists s∈H0⁢(X,ℒn) for some positive integer n such that

  1. (1)

    s is G-invariant under the natural action of G on H0⁢(X,ℒn) induced by the linearization;

  2. (2)

    Xs=X∖𝕍⁡(s) is affine;

  3. (3)

    x∈Xs.

Furthermore, x is stable if the action on Xs is closed.

Remarks 2.53.
  1. (1)

    We see that the semistable locus Xss⊂X is open.

  2. (2)

    If ℒ is ample, then Xs is automatically affine.

Theorem 2.54 (Geometric invariant theory).

Let G be reductive and (X,ℒ) be a finite-type scheme over k with a G-linearized line bundle. Then the semistable locus

is open and there exists a good quotient

and an open Y~⊂Y such that π−1⁢(Y~)=Xs⁢(ℒ) and Xs⁢(ℒ)→Y~ is geometric. Moreover, there exists an ample line bundle ℳ on Y and positive integer N such that

Proof sketch.

Write

which requires replacing ℒ by ℒ⊗N. Here, we have

for si∈H0⁢(X,ℒ)H. By definition, each Xsi=Spec⁡Γ⁡(Xsi,𝒪) is affine. Applying affine GIT, we obtain good quotients

Our goal is now to glue the Ui. If we consider the function sisj, it is a G-invariant function on Xsj, so it descends to a function σi⁢j on Uj. Now (Uj)σi⁢j is the categorical quotient of Xsi∩Xsj, but so is (Ui)σj⁢i, so they are canonically isomorphic via

These satisfy the cocycle condition because everything is a canonical isomorphism between categorical quotients, so the Ui glue to a scheme. The {σi⁢j} give a 1-cocycle defining a line bundle ℳ such that π∗⁢ℳ=ℒ. Ampleness follows from the fact that the distinguished opens Ui are affine. □

Theorem 2.55.

If X is projective and ℒ is ample, then

is a projective scheme and π∗⁢𝒪Y⁢(1)=ℒ⊗N.

Remark 2.56.

In this case, we have

3 Stacks

3.1 Quotient stacks.

Let X be a G-scheme for an affine group scheme G.

Definition 3.1.

Define the quotient stack

to be the contravariant (pseudo)functor defined by

Now we will work over a field k and require G to be reductive. We will need to make sense of everything in the above definition.

Definition 3.2.

A morphism π:E→B is a fibration with fiber F if there exists a surjective étale (equivalently, smooth) morphism B′→B such that the diagram

is a Cartesian diagram.

Definition 3.3.

Suppose that F and E are G-schemes. We say that a fibration E→B with fiber F is G-equivariant if σ in the previous definition is G-equivariant.

Definition 3.4.

A morphism π:E→B is a principal G-bundle if it is a G-equivariant fibration with fiber G, where G acts on itself by left multiplication.

Lemma 3.5 (G reductive).

Suppose that π:E→B is a principal G-bundle. Then

  1. (1)

    B≃E/G is a good geometric quotient;

  2. (2)

    G acts freely on E in the stronger sense that

    is a closed embedding.

Proposition 3.6 (Converse).

If G acts freely on X and the quotient X/G exists as a scheme, then X→X/G is a principal G-bundle.

Since π:E→B is affine, we can glue affine GITs to get a good quotient E/G=BGIT. This produces the diagram

where the bottom square is Cartesian. This implies that BGIT→B is an isomorphism because this is local on the target.

An alternative proof is to consider the diagram

Because G acts freely on E′, Ψ′ is a closed embedding, and therefore so is Ψ. Then we can deduce things going in a direction that is not the one we usually go in.

Proposition 3.7.

Suppose G acts freely on X and a good geometric quotient π:X→X/G exists. Then π is a principal G-bundle.

Exercise 3.8.

Suppose Y is an affine G-scheme and X→B is a principal G-bundle. Then Y×Spec⁡kX/G→X/G≅B is a G-equvariant fibration with fiber Y (here, G acts on the first factor Y).

Exercise 3.9.

Let f:X→Y be a G-morphism of principal G-bundles. Then we have a Cartesian diagram

Hint: consider the diagram

and use the next exercise to conclude that the dashed arrow is an isomorphism.

Exercise 3.10.

A G-equivariant morphism of principal G-bundles over some base is an isomorphism.

Example 3.11.

Let X be a G-scheme with G affine. Define the quotient stack

to be the contravariant (pseudo)functor given by

Note that this is a groupoid by the last exercise. The morphisms in this category are f∗:[X/G]⁢(U)→[X/G]⁢(U′) for morphisms f:U′→U, and are given by

We also have a natural isomorphism

but do not have equality on the nose. This boils down to the fiber product being unique up to unique isomorphism.

Exercise 3.12.

If X is a principal G-bundle, then [X/G] is representable by X/G. Hint: consider the Cartesian diagram

and show that [X/G]⁢(U)=𝖬𝗈𝗋⁡(U,B).

Let H<G be a closed subgroup and Y be an H-scheme with a good quotient. Then H acts on G×Y by h⋅(g,y)=(g⁢h−1,h⁢y), and this action commutes with the G-action on G×Y. Therefore, G acts on

by the formula

In addition, the quotients become

To see this, suppose X=G×Y=Spec⁡A is affine. Then we have

In particular, we see that every H-quotient is a G-quotient.

Remark 3.13.

If Y↪X where X is a G-scheme, we have a natural map

whose image is G⁢Y.

Let X and Y be G-schemes admitting good quotients.

Definition 3.14.

A G-morphism f:X→Y is strongly étale if in the diagram

the morphisms πX and f induce an isomorphism X≃(X⫽G)×Y⫽GY and f/G is étale.

Example 3.15.

Consider ℂ××ℂ×⊂ℂ××ℂ⊂ℂ×ℂ. Consider the action of G=ℂ× by λ⋅(x,y)=(λ⁢x,λ⁢y). The quotient of ℂ×ℂ is a point, the quotient of ℂ××ℂ is ℂ=Spec⁡ℂ⁡[1y], and the quotient of ℂ×ℂ× is ℂ×. In the diagram

we see that the square on the left is Cartesian, but the square on the right is not. Thus ℂ××ℂ×↪ℂ× is strongly étale, but ℂ××ℂ↪ℂ×ℂ is not.

Lemma 3.16.

If f is strongly étale, then

  1. (1)

    f is étale;

  2. (2)

    f⁡(X) is a saturated open in Y, or namely,

  3. (3)

    For all closed points x¯∈X⫽G, we have πX−1⁢(x¯)≃πY−1⁢(f/G⁡(x¯)).

  4. (4)

    The morphism f|G⋅x:G⋅c→G⋅f(x) is an isomorphism for all x∈X⁡(k) and G⋅x is closed in X if and only if G⋅f⁡(x) is closed in Y. Also, we have Gx≃Gf⁡(x).

  5. (5)

    X is normal at x.

Lemma 3.17 (Luna’s fundamental lemma).

Suppose f:X→Y is a G-morphism of affine G-schemes. Suppose that x∈X⁡(k) is a closed point such that

  1. (1)

    T=G⋅x is closed;

  2. (2)

    The morphism f|T:T→Y is injective (equivalently, Gx≃Gf⁡(x));

  3. (3)

    f is étale at x.

Then there exists a saturated affine open U⊂X such that V=f⁡(U) is a saturated open affine in Y, T⊂U, and f|U:U→V is strongly étale.

Theorem 3.18 (Luna’s étale slice theorem).

Suppose X is an affine G-scheme and x∈X⁡(k) is such that G⋅x is closed. Then there exists a Gx-invariant locally closed affine V⊂X with x∈V such that U≔G⋅V is a G-saturated open affine in X and the morphism

is strongly étale. Moreover, if x∈X is a smooth point, then we have a strongly étale morphism

which is Gx-equivariant such that η⁡(x)=0.

In other words, there is a diagram

where all of the squares are Cartesian.

Corollary 3.19.

Suppose X is an affine G-scheme. Let X0⊂X be the subset of points such that

  1. (1)

    G⋅x is closed;

  2. (2)

    Gx≃1.

Then X0 is a saturated open and X0→X0⫽G is a principal G-bundle.

Proof.

Let x∈X0 such that Gx≃1 and G⋅x is closed. By Luna’s étale slice theorem, there is a strongly étale morphism

and covering X0 by such U gives the desired result. □

Corollary 3.20.

Suppose X is a smooth affine G-scheme. Then XG is smooth.

Proof.

Let x∈XG. Then G⋅x={x} is closed and Gx=G. By Luna’s étale slice theorem, we have

This gives us a correspondence

of étale morphisms, and TX,xG is a vector space, and hence smooth, so everything else is smooth. □

Lemma 3.21.

Let X be affine and x∈X be such that Gx is reductive. Then there exists an étale Gx-morphism ϕ:X→TX,x such that ϕ⁡(x)=0.

Proof.

Let X=Spec⁡A with x corresponding to a maximal ideal 𝔪. Then of course we have

There is a section 𝔪/𝔪2→𝔪⊂A which then lifts to an algebra morphism. This is clearly an isomorphism on tangent spaces, so it is étale. □

Lemma 3.22.

Every affine finite-type G-scheme X can be embedded as a G-stable closed subscheme of an affine space with G-action.

Proof.

If X=Spec⁡A and A is finitely generated, then we choose G-invariant generators x1,…,xn of A. □

Proof sketch of Luna’s étale slice theorem.

Suppose that x∈X has a closed orbit G⋅x. Then there is a closed immersion

and thus the orbit is affine. Matsushima’s criterion states that if G is reductive and H⊂G is a closed subgroup, then G/H is affine if and only if H is reductive. Therefore, Gx is reductive. By Lemma 3.21, there is an étale Gx-morphism ϕ:X→TX,x such that ϕ⁡(x)=0. Assume now that x∈X is a smooth point. Then we can decompose

where Nx is the normal space to G⋅x at x. Let Y=ϕ−1⁢(Nx). This is a closed subscheme of X and is smooth at x. We now consider the morphism

Then in fact G×GxY is smooth at z (which checks part 4 of Luna’s fundamental lemma). Now, note that

is a principal Gx-bundle. We also compute Gz=G(1G,x)¯≃Gx because g⋅(1G,x)¯=(1G,x)¯ if and only if (g,g⁢x)∈Gx⋅(1G,x), which is equivalent to g∈Gx. Therefore, ψ is étale at z. This implies that G⋅z is closed in G×GxY because Gx⋅x is closed in Y.

We can now apply Luna’s fundamental lemma to ψ:G×GxY→X to obtain a saturated open U⊂G×GxY such that

is strongly étale. Now consider the diagram

Because U is a saturated open in G×GxY, there exists a saturated open V⊂Y such that U=G×GxV. Therefore, we obtain a strongly étale morphism G×GxV→X. This finishes the proof in the case when x∈X is a smooth point. The general case follows by Lemma 3.22. □

One application of Luna’s étale slice theorem is that the quotient stack 𝔛=[X/G] is locally isomorphic to [V/Gx] near a closed point x∈𝔛 (at least at a smooth point).

3.2 General stacks.

Every stack is going to be a category fibered in groupoids. We begin by discussing some examples.

Example 3.23.

Every contravariant functor (presheaf) F:𝖲𝖼𝗁Sop→𝖲𝖾𝗍 defines a category fibered in groupoids 𝒢F over 𝖲𝖼𝗁S whose objects are simply

and whose morphisms are given by

The reason it is called a category fibered in groupoids is that there is a morphism to 𝖲𝖼𝗁S and each p−1⁢(B)=F⁡(B) is a set (which is the simplest example of a groupoid).

Example 3.24.

Consider the moduli problem ℳg, which is given by

Morphisms are given by Cartesian diagrams

and over the identity arrow of B we clearly only have isomorphisms.

Example 3.25.

If Z is a scheme, its functor of points F=hZ is a category fibered in groupoids.

If we have a category fibered in groupoids p:𝒞→𝖲𝖼𝗁S, we have the Isom functor

for any two objects X,Y∈𝒞⁡(B).

Example 3.26.

If 𝒞=𝒞F for a functor F:𝖲𝖼𝗁Sop→𝖲𝖾𝗍, then

Definition 3.27.

A stack is a category 𝒞 fibered in groupoids over 𝖲𝖼𝗁S if

  1. (1)

    For all B∈𝖲𝖼𝗁S, the Isom functor 𝖨𝗌𝗈𝗆⁡(X,Y) is a sheaf on the big étale site of B;

  2. (2)

    Descent is effective in the étale topology for all B. More precisely, given a surjective étale cover B′=⨆Bi→B, xi∈𝒞⁡(Bi), and isomorphisms ϕi⁢j:xi|Bi×BBj≃xj|Bi×BBj satisfying the cocycle condition ϕi⁢j∘ϕj⁢k=ϕj⁢k on triple overlaps Bi×BBj×BBk, there exists x∈𝒞⁡(B) such that xi≃x|Bi.

Equivalently, it is a sheaf of groupoids on (𝖲𝖼𝗁S)e´⁢t.

Definition 3.28.

The big {étale, smooth, fppf, Zariski, etc} site is the category 𝖲𝖼𝗁S with the Grothendieck topology in which covers are S-morphisms which are a disjoint union of {étale, smooth, fppf, open immersions, etc} morphisms and jointly surjective. This is denoted by (𝖲𝖼𝗁S)e´⁢t,smooth,fppf,etc.

Definition 3.29.

A sheaf on (𝖲𝖼𝗁S)τ is a presheaf F satisfying the sheaf condition that

is an equalizer for all τ-covers U′→U.

Example 3.30.

Because schemes are sheaves in the fpqc topology, every scheme is a stack.

Now suppose that 𝔛, 𝔜, and ℨ are stacks with morphisms f:𝔛→𝒵 and g:𝔜→𝒵. We can form the fiber product

The main point is that this fiber product is again a stack.

Definition 3.31.

A morphism of stacks f:𝔛→𝔜 is representable by schemes if for every scheme B and morphism B→𝔜, the fiber product 𝔛×𝔜B is a scheme.

Example 3.32.

The morphism X→[X/G] given by the action

is representable by schemes because for any

giving a morphism B→[X/G], the diagram

is Cartesian.

Definition 3.33.

An algebraic space is a sheaf F of sets on (𝖲𝖼𝗁X)e´⁢t such that

  1. (1)

    The diagonal morphism ΔF:F→F×F is representable by schemes (this implies that any morphism from a scheme is representable by schemes because B×FC=(B×C)×F×FF, and the two conditions are in fact equivalent);

  2. (2)

    There exists a surjective étale morphism U→F where U is a scheme.

Definition 3.34.

A stack 𝔛 is Deligne-Mumford (resp. algebraic) if

  1. (1)

    The diagonal morphism Δ𝔛:𝔛→𝔛×𝔛 is representable by algebraic spaces;2

  2. (2)

    There exists a surjective étale (resp. smooth) morphism U→𝔛 where U is a scheme.

Example 3.35.

The quotient stack [X/G] is an algebraic stack and the morphism X→[X/G] is a smooth atlas (at least in characteristic zero) and an algebraic space if G is finite and all stabilizers are trivial. Without the assumption on stabilizers, it is a Deligne-Mumford stack.

Because étale descent is effective for principal G-bundles, we will not discuss descent. The reason that X→[X/G] is representable is because the square

is Cartesian, and we see that E is a scheme. Here, if we have morphisms D→B and D→X, the fact that X×[X/G]X≃G×X implies that E×BD≅G×D and gives us a morphism D→E, which is the desired result. The morphism X→[X/G] is also clearly smooth and surjective, so it is a smooth atlas.

Example 3.36.

If we define ℳg to send a scheme B to the set of isomorphism classes of flat families 𝒞→B of genus g smooth curves over B, then it is not a stack. For example, consider a genus g curve C with a nontrivial involution and an étale double cover B′→B. Then we can consider the quotient X of C×B′ by the diagonal action of ℤ2, and the resulting diagram

is Cartesian, but X≄C×B in general. This shows that the Isom functor 𝖨𝗌𝗈𝗆⁡(C×B,X) is not a sheaf and that descent is not effective.

However, if we allow the total space to be an algebraic space, then if we consider the diagram

then taking the pushout in the category of algebraic spaces gives us a family 𝒞 over B filling in the diagram.

Example 3.37.

Let X be the affine line with doubled origin and ℤ2 act by switching the two origins and exchanging x↔−x for x≠0. Then [X/ℤ2] is a sheaf on 𝖲𝖼𝗁e´⁢t (and in fact an algebraic space) but is not a scheme because the two origins do not lie in a single affine.

Example 3.38.

Consider two copies of ℙ1 which are glued along two nodes. This configuration admits two inclusions f,g into ℙ2 (where one becomes a line and the other a conic), and there is a pushout ℙ2∪f=gℙ2, which is not projective. There is an involution induced by switching the two copies of ℙ1 and the two nodes, and the stack [X/ℤ2] is not a scheme because the two nodes do not lie in a single affine, but is an algebraic space.

Remark 3.39.

If we have a presentation U→𝔛, then we can replace the condition that the diagonal is representable by requiring that the presentation is representable.

3.3 Equivalence relations.

It is easier to deal with morphisms of schemes than morphisms of stacks, so we will give another way to think about stacks.

Definition 3.40.

A pre-relation on a scheme U over S is a morphism j:R→U×SU where R is a scheme. It is a relation if j is a monomorphism (equivalently, Δ:R→R×U×UR is an isomorphism, or the morphism is injective on T-points for all T).

A pre-equivalence (pre-)relation is a (pre-)relation such that the image of R⁡(T) in U⁡(T)×U⁡(T) is an equivalence relation for all T. An equivalence relation is a pre-equivalence relation.

Example 3.41.

Let G×X→𝜎X be a group action and

Then Ψ is a pre-equivalence pre-relation and is an equivalence relation if all stabilizers are trivial.

Definition 3.42.

A groupoid scheme over S is a quintuple (U,R,s,t,c) and

  1. (1)

    An S-morphism j:R→(t,s)U×SU;

  2. (2)

    An S-morphism c:R×t,U,sR→R

such that j⁡(T) and c⁡(T) define a groupoid on U⁡(T).

Exercise 3.43.

If (U,R,s,t,c) is a groupoid scheme, then j:R→U×U is a pre-equivalence pre-relation.

Exercise 3.44.

Every equivalence relation j:R→U×U extends uniquely to a groupoid scheme (U,R,s,t,c). Hint: define c using the Yoneda lemma.

Example 3.45.

Let f:U→𝔛 be a presentation of an algebraic space 𝔛. Then if we set R=U×𝔛U, the morphism j:R→U×U is an étale equivalence relation. Here, the set

is obviously an equivalence relation.

Our goal now is to discuss quotients by pre-relations j:R→U×U. If T is a scheme, we get an induced equivalence relation ∼T on U⁡(T). In particular, we obtain a presheaf on (𝖲𝖼𝗁/S)e´⁢t given by

Definition 3.46.

The quotient U/R is the sheafification of this presheaf.

Remark 3.47.

If we consider U and R as sheaves, then the diagram

is a coequalizer in the category of sheaves.

Remark 3.48.

For any surjection of sheaves on any site q:ℱ→𝒢, the diagram

is both a pullback and a pushout.

If we apply this to the diagram

then the map R→U×U/RU is surjective as a map of sheaves.

Theorem 3.49.

Let U be a scheme and j:R→U×U be an étale equivalence relation. Then the sheaf U/R is an algebraic space and q:U→U/R is a representable étale surjective morphism.

Proposition 3.50.

Suppose F:(𝖲𝖼𝗁/S)e´⁢top→𝖲𝖾𝗍 is an algebraic space with an étale atlas f:U→F. Write R=U×FU. Then

  1. (1)

    The map j:R→U×U is an equivalence relation;

  2. (2)

    The maps s,t=pr1,pr2 are étale and surjective;

  3. (3)

    The diagram R⇉U→F is a coequalizer, and in particular F≃U/R.

Proof.

Note that the set

defines an equivalence relation. Because objects in F have no automorphisms, we see that R=U×FU→U×U is a monomorphism, so R is an equivalence relation.

The fact that s and t are étale is just the fact that the property of being étale is stable under base change.

Finally, the property of being a coequalizer is a property of any surjection of sheaves. □

Definition 3.51.

Let 𝔛 be an algebraic (resp. DM) stack. A morphism ϕ:𝔛→X to an algebraic space X is a good (resp. coarse) moduli space if ϕ is quasicompact and

  1. (1)

    ϕ∗:𝖰𝖢𝗈𝗁⁡(𝔛)→𝖰𝖢𝗈𝗁⁡(X) is exact (namely ϕ is cohomologically affine);

  2. (2)

    We have ϕ∗⁢𝒪𝔛≃𝒪X.

Example 3.52.

If G is reductive, then the map

is a good moduli space in characteristic zero. The first follows from the fact that taking invariants is exact, and the second follows from GIT.

Theorem 3.53.

Suppose 𝔛=[U/G] is a quotient stack where U is normal and G is reductive. If all stabilizers are finite, then 𝔛 is a Deligne-Mumford stack and has a coarse moduli space.

For example, ℳ¯g is of this form (where U is a suitable locus inside a Hilbert scheme of ℙN and G=PGLN+1), so this theorem applies to it.

Proof.

By Luna’s étale slice theorem, there exists a strongly étale morphism V≃V0×GxG→𝜓U where V0 is affine and locally closed in U, ψ⁡(V) is saturated, and G⋅x⊂V.

This implies that the morphism [V0/Gx]≃[V/G]→[U/G] is strongly étale. Note that [V0/Gx] has a coarse moduli space Spec⁡H0⁢(V0,𝒪V0)Gx. Therefore, there exists a strongly étale morphism 𝔜→𝔛 where 𝔜 has a coarse moduli space Y. Now consider the fiber product 𝔜×𝔛𝔜, which has a coarse moduli space Y2. Then there are two morphisms s,t:Y2→Y which are affine and étale, so the left square in the diagram

is Cartesian. Finally, Y2⇉Y is an étale equivalence relation, so we can form its quotient X, which is an algebraic space. □

Remark 3.54.

The theorem also applies in the case where G is reductive and all stabilizers are reductive.

Theorem 3.55 (Grothendieck).

Let f:T′→T be an fppf morphism and X′→T′ be separated and locally quasi-finite, then descent is effective.

We have discussed descent data and whether it is effective, so we will now give a precise definition of these notions.

Definition 3.56.

Let T′→T be a morphism of schemes. A descent datum is a scheme X′/T′ and ϕ:X′×TT′→T′×TX′ satisfying the cocycle condition that the diagram

commutes. It is effective if X′=X×TT′ where X→T and ϕ is the canonical isomorphism.

Lemma 3.57.

Let f:T′→T be an étale surjective morphism of schemes. Suppose that ℱ is a sheaf on (𝖲𝖼𝗁/T)e´⁢t. Suppose that f∗⁢ℱ is a scheme. Then X′=f∗⁢ℱ defines a descent datum relative to f. Moreover, this descent datum is effective if and only if ℱ is a scheme.

Consider the diagram

Here, note that ϕ comes from the fact that the two morphisms T′×TT′→T coincide.

Proof of Theorem 3.49.

For simplicity, we will take U to be quasi-compact. Then there exists a surjective étale morphism U′→U where U′ is affine. Consider the Cartesian diagram

Then j′:R′→U′×U′ is an equivalence relation and U′/R′≃U/R as sheaves. Therefore, we may assume that U is affine.

Now because j:R→U×U is an étale equivalence relation, then j is a monomorphism and thus Δj is a closed embedding (in fact an isomorphism) and hence j is separated. Therefore, we may assume that R is also separated.

We now need to prove that q:U→U/R is representable and étale (since it is clearly surjective). First, note that the diagram

is Cartesian. Now consider a Cartesian diagram

There is some surjective étale ϕ:T′→T such that there is a lift a′:T′→U such that q∘a′=a∘ϕ, so it fills in as

We now claim that ℱ′≃R×t,U,a′T′. Here, note that

This implies that ℱ′ is a scheme, and hence ℱ is a scheme. □

Now let 𝔛=U/R be an algebraic space. Consider the Cartesian diagram

In particular, we can descend properties of j to properties of the diagonal of 𝔛. For example, to show that the diagonal is representable, consider the Cartesian cube

Because j is quasi-finite and separated (it is a monomorphism) and ℱ′ is a scheme, we see that Δ is representable. Now consider the Cartesian diagram

defining the inertia stack (which is not interesting for an algebraic space). Therefore, we can see that properties of j correspond to properties of the inertia stack. One important property is that the diagram

is Cartesian. One upshot is that we can define properties of an algebraic stack using properties of the diagonal (or the inertia stack).

Definition 3.58.

An algebraic stack 𝔛 is separated if Δ𝒳 is proper. For an algebraic space, this is equivalent to being a closed embedding.

If 𝔛 is an algebraic stack, then we have a smooth atlas q:U→𝔛, and we can form the Cartesian diagram

where j:R→U×𝔛U defines a pre-equivalence pre-relation. If we have such a pre-equivalence pre-relation, we can define the groupoid

where U⁡(T) are objects and R⁡(T) are arrows, which defines a prestack. The sheafification of this prestack is the stack 𝔛.

Definition 3.59.

The condition for 𝔛 to be separated is equivalent to the condition that j is proper.

One upshot of the diagram defining the inertia stack is that if 𝔛 is separated, then 𝖨𝗌𝗈𝗆T⁢(a,a) is a proper group algebraic space over T.

4 Moduli spaces

4.1 Some basic situations.

Theorem 4.1 (Keel-Mori).

If 𝔛 is an algebraic stack with finite inertia, then 𝔛 has a good moduli space ϕ:𝔛→X where X is an algebraic space. Furthermore, if 𝔛 is separated, then X is separated.

Example 4.2.

When 𝔛=[X/G], then Δ𝔛 is separated (proper, finite, etc.) if and only if j=(σ,pr2):G×X→X×X is. These correspond to 𝔛 being an unknown condition, separated, DM with finite inertia, etc.

Example 4.3 (Mumford).

In Example 2.47, the stack 𝔛=[X/SL2] is an algebraic space because the stabilizers are trivial but is not separated because the action is not proper.

Example 4.4.

Let U={D∈Sym4⁡(ℙ1)||Supp⁡(D)|≥3} and let 𝔛=[U/PGL⁡(2)]. This stack has quasi-finite inertia, but the generic stabilizer is ℤ2×ℤ2 and the automorphism group of a point corresponding to a divisor with a double point is ℤ2, so the inertia is not proper (hence not finite). In particular, 𝔛 has no coarse (or good) moduli space.

In fact, we have 𝔛⊂𝔛GIT=[Sym4⁡(ℙ1)ss/PGL⁡(2)], and the latter stack has a good moduli space X, but 𝔛 does not have unpunctured inertia, so it has no good moduli space despite 𝔛→X being bijective on geometric points.

Example 4.5.

The stack ℳg has finite inertia when g≥2. The inertia is quasi-finite by Hurwitz’s theorem that |Aut⁡(C)|≤84⁢(g−1) when C is a smooth curve. Note that a morphism π:Spec⁡K→I𝔛 is the data of a smooth curve 𝒞K over Spec⁡K and an automorphism of 𝒞K. The diagram

corresponds to a lift to a family 𝒞R. Lifting to a morphism Spec⁡R→I𝔛 corresponds to extending the automorphism to the central fiber. Any family of genus g smooth curves has no (−1)-curves, and therefore the birational map 𝒞R⇢𝒞R induced by the automorphism of 𝒞K is an automorphism of 𝒞R. This implies that the inertia is proper, and hence finite, so ℳg has a coarse moduli space.

We will now list some properties of good moduli spaces.

  1. (1)

    The good moduli space ϕ is initial from 𝔛 to algebraic spaces;

  2. (2)

    X⁡(k) is in bijection with closed k-points of 𝔛 for all algebraically closed fields k;

  3. (3)

    If 𝔛 is DM, then X is a coarse moduli space.

Theorem 4.6.

Suppose that 𝔛 has a good moduli space and f:𝔜→𝔛 is a representable affine morphism of algebraic stacks. Then 𝔜 has a good moduli space.

The idea here is that Spec𝒪X⁡(ϕ∗⁢f∗⁢𝒪𝔜)≕Y serves as the good moduli space for 𝔜.

Definition 4.7.

Let f:𝔛→𝔜 be a morphism of finite type algebraic stacks over an algebraically closed field k. Given a closed point x∈𝔛⁡(k), f is strongly étale at x if

  1. (1)

    f is étale at x;

  2. (2)

    f⁡(x) is closed;

  3. (3)

    The induced map of stabilizer groups Gx→Gf⁡(x) is an isomorphism.

Lemma 4.8 (Luna’s fundamental lemma).

Let f:𝔛→𝔜 be a morphism of algebraic stacks with good moduli spaces X and Y.

  1. (1)

    f is strongly étale if and only if the map f¯:X→Y is étale and the diagram

    is Cartesian.

  2. (2)

    If f is strongly étale at a closed x∈𝔛⁡(k), then there exists an open 𝔘⊂𝔛 containing x such that f|𝔘:𝔘→𝔜 is strongly étale.

Note that this is the same as the previous version of Luna’s fundamental lemma in the case of quotient stacks.

Proof.

By a result of Alper, if ϕ:𝔛→X is a good moduli space and x∈𝔛⁡(k) is a closed point, then 𝒪^X,ϕ⁡(x)≃𝒪^𝔛,xGx (we will not define the local ring for a stack here). Because f is stabilizer preserving, we have Gx≃Gf⁡(x). Because f is formally étale, we have 𝒪^𝔜,f⁡(x)≃𝒪^𝔛,x. Taking invariants gives us 𝒪^X,x≃𝒪^Y,f⁡(x). This implies that if f is strongly étale, then f¯ is étale.

Now consider the diagram

Note that X is also a good moduli space of 𝔜×YX by a result of Alper. Note also that f is separated, and therefore so is g:𝔛→𝔜×YX. Also, g is étale because f is. By Zariski’s main theorem for stacks, we have a factorization

where 𝔛→ℨ is an open immersion and ℨ→𝔜×YX is finite. Note also that g sends closed points to closed points and is stabilizer preserving. Note that ℨ has a good moduli space, so we have a diagram

However, both X→Z and X→Z are isomorphisms. Because g preserves closed points, then 𝔛→ℨ must also preserve closed points. However, the only way for an open immersion to induce an isomorphism on good moduli spaces is if there is a closed point in 𝔛 which is not closed in ℨ, which is impossible because then g would not preserve closed points. This implies that 𝔛≃ℨ.

The final step is to note that Z=Spec𝒪X⁡(π∗⁢ϕ∗⁢𝒪ℨ)=Spec𝒪X⁡𝒪X⊕π∗⁢ℒ for some additional factor ℒ making ϕ∗⁢𝒪ℨ≃𝒪𝔜×YX⊕ℒ. This factor must disappear, so ϕ is an isomorphism. □

In the proof, we implicitly used the following lemma.

Lemma 4.9.

Let ℨ have a good moduli space and 𝔘⊆ℨ be an open immersion. Then 𝔘⊆ℨ is strongly étale if and only if 𝔘 is saturated.

Definition 4.10.

Let 𝔛 be an algebraic stack of finite type over an algebraically closed field k and x∈𝔛⁡(k) be a closed point. Then 𝔛 has a strongly étale local quotient presentation at x if

  1. (1)

    Gx is linearly reductive;

  2. (2)

    There exists f:[Spec⁡A/Gx]→𝔛 sending a closed point w↦x with A a finitely generated k-algebra such that:

  3. (3)

    f is strongly étale (at w is enough).

Theorem 4.11.

Suppose 𝔛 is an algebraic stack of finite type over an algebraically closed field with affine diagonal. Suppose also that

  1. (1)

    Every closed point has a strongly étale local quotient presentation;

  2. (2)

    Every {x}¯ has a good moduli space.

Then 𝔛 has a good moduli space.

Example 4.12.

Consider the diagram

This is Cartesian, and the left vertical arrow is smooth, so 𝔛=[X/G] has affine diagonal if G is affine.

Remark 4.13.

If 𝔛 is a Deligne-Mumford stack, then {x}¯={x}, and therefore the second condition in the theorem is automatically satisfied.

Proposition 4.14 (Alper-Kresch).

Every quotient stack [X/G] where G is a linear algebraic group and X is a separated and finite type scheme3 over an algebraically closed field has a étale local quotient presentation at every closed point with linearly reductive stabilizer.

Corollary 4.15.

If 𝔛=[X/G] is as above and is Deligne-Mumford, then 𝔛 has:

  1. (1)

    étale local quotient presentations at all closed points;

  2. (2)

    {x}¯={x} for all x∈𝔛⁡(k).

This does not quite imply that 𝔛 has a good moduli space, but it is close. The only missing ingredient is the following result.

Proposition 4.16.

Suppose that 𝔛 has finite inertia. Then 𝔛 has an étale local quotient presentation at a closed point x if and only if it has a strongly étale local quotient presentation at x.

Recall here the diagram

Proof.

Let f:𝔚=[Spec⁡A/Gx]→𝔛 be an étale local quotient presentation at x. Note that f sends closed points to closed points because all points in both 𝔚 and 𝔛 are closed.

Lemma 4.17.

f is stabilizer preserving in an open neighborhood x∈𝔘⊆𝔚, where 𝔘=[Spec⁡B/Gx] such that Spec⁡B⊆Spec⁡A is a Gx-invariant open (which is clearly saturated).

Note that it is enough to find any 𝔘 where f is stabilizer-preserving because we can just shrink to make it affine. Consider the diagram

Denote the induced map h:I𝔚→𝔚×𝔛I𝔛. These are both finite over 𝔚, so our goal is to show that h is an open immersion. Now note that f:𝔚→𝔛 is étale, so we can consider the diagram

which is Cartesian. Because Δf is an open immersion (this follows from f being unramified), so is h.

Now consider the diagram

where ℨ=𝔚×𝔛×𝔛𝔛. By the cancellation property, we see that

Remark 4.18.

We can extend the definition of a groupoid scheme to that of a groupoid algebraic space.

Example 4.19.

Consider f:U→B and then set R=U×BU. We can also consider U×B×BU, and for example the diagram

is Cartesian (as are all other similar diagrams). We can think of this as

To recognize a general groupoid scheme or algebraic space, we can just chase the properties of this picture.

Proof of Theorem 4.11.

For every closed point x∈𝔛⁡(k), it in particular has a reductive stabilizer, and a strongly étale and affine

Note that Spec⁡AxGx is a good moduli space for [Spec⁡Ax/Gx], so we can consider

which is surjective, strongly étale, and affine, and has a good moduli space. As a remark, by GIT, 𝔛′ satisfies both both conditions in the theorem.

Proposition 4.20.

Suppose that 𝔛′→𝔛 is strongly étale where 𝔛′ and 𝔛 both satisfy the condition that every {x}¯ has a good moduli space. Then for all 𝔚→𝔛, the base change 𝔛′×𝔛𝔚→𝔚 is strongly étale.

Proof.

Let z∈𝔛′×𝔛𝔚 is closed. We will name the morphisms by

Replace 𝔛′ by g′⁢(z)¯, 𝔚 by f′¯⁢(z), and 𝔛 by 𝔤⁡(𝔣′⁢(𝔷)). Now both 𝔛 and 𝔛′ have good moduli space given by a point, and the diagram

is Cartesian. This implies that f and f′ are isomorphisms. □

Returning to the proof of the theorem, set X′′≔𝔛′×𝔛𝔛′⇉𝔛′. Note that both pr1 and pr2 are also strongly étale by the proposition. Set 𝔛′′′=𝔛′×𝔛𝔛′×𝔛𝔛′. Note that the pri⁢j are also strongly étale. Because 𝔛′ has a good moduli space and all pri and pri⁢j are affine, 𝔛′′ and 𝔛′′′ have good moduli spaces. We obtain a diagram

where the bottom row is an étale groupoid algebraic space. We will prove that it is in fact an étale equivalence relation. Once we have this, we can form the quotient X=X′/X′′, which is an algebraic space, and the map X′→X is étale and surjective, so X is a good moduli space for 𝔛.

The intuition is that 𝔛′=[Z/G] and 𝔛=[Y/G] and f is induced by a G-morphism f:Z→Y that sends closed orbits to closed orbits and is stabilizer-preserving on closed orbits. Consider the morphism

Note that 𝔛′×𝔛𝔛′=[Z×YZ/G], because taking the quotient of the diagram

by G preserves the property of being Cartesian.4

Consider two closed orbits G⋅(a,b) and G⋅(c,d) in Z×YZ. Note that in Y, we have f⁡(a)=f⁡(b) and f⁡(c)=f⁡(d). Suppose they map to the same point in X′×X′. But then we have G⋅a=G⋅c and G⋅b=G⋅d in Z. This implies that a=g1⁢c and b=g2⁢d for some g1,g2∈G. This implies that g1⁢f⁢(c)=f⁡(a)=f⁡(b)=g2⁢f⁢(d)=g2⁢f⁢(c), and therefore g2−1⁢g1 stabilizes f⁡(c). Because f is stabilizer-preserving on closed orbits, g2−1⁢g1 stabilizes c. In particular, we see that g2−1⁢a=c and therefore g2−1⁢(a,b)=(c,d). Thus they define the same point in [Z×YZ/G]. □

4.2 Keel-Mori.

Our next goal is to prove the following.

Theorem 4.21.

If 𝔛 is an algebraic stack with finite inertia, then it has a coarse (also good) moduli space π:𝔛→X such that

  1. (1)

    π is initial among morphisms to algebraic spaces;

  2. (2)

    |𝔛⁡(k)|=X⁡(k) for all algebraically closed fields k;

  3. (3)

    π is proper and quasi-finite;

  4. (4)

    The formation of π is invariant under flat base change and is in particular a universal homeomorphism.

Our first reduction is to write 𝔛=⋃𝔛i, where 𝔛i⊂𝔛 is open such that there exist good moduli spaces 𝔛i→Xi satisfying the four conditions, and then glue the Xi together in the Zariski topology. By universality, we have Ui⁢j=Uj⁢i and the cocycle condition is automatically satisfied. The second reduction is the following.

Lemma 4.22 (Grothendieck).

Let 𝔛 be an algebraic stack with quasi-finite diagonal (quasi-finite inertia is enough). Then there exists an open cover of 𝔛 by open substacks such that each open substack admits a flat, quasi-finite, finitely presented cover by a quasi-projective scheme.

The next reduction is that if 𝔛 has finite inertia and admits a flat, quasi-finite, finitely presented cover by a quasi-projective scheme, then there exists a representable morphism 𝔘′→𝔛 which is strongly étale, quasicompact, and surjective, where 𝔘′ is an algebraic stack and has a finite locally free cover by a scheme. By a standard argument, it suffices to show that 𝔘′ has a good moduli space.

Definition 4.23.

A morphism Spec⁡B→Spec⁡A is finite locally free if B is a free A-module of finite rank.

In conclusion, it suffices to show that a groupoid space [R⇉U] where s and t are finite locally free has a good moduli space. In the first case, we will consider the case where U=Spec⁡B is affine, where B is a finite type A-algebra.

Definition 4.24.

Define the subring

Example 4.25.

If R=G×Spec⁡B for a finite group G acting on B, then BR=BG is the ring of invariants.

Theorem 4.26.
  1. (1)

    BR is a finite type A-algebra;

  2. (2)

    B is a finite BR-algebra;

  3. (3)

    The finite surjective morphism Spec⁡B=U→Q≔BR is a good moduli space for the groupoid scheme R⇉U which produces the stack [R/U];

  4. (4)

    For every flat affine Q′→Q, [R×QQ′⇉U×QQ′]→Q is a good moduli space.

Proof of the fourth statement.

Consider the exact sequence

of BR-modules. If Q′=Spec⁡B0′, then tensoring with B0′ is exact. This proves the fourth statement. □

Recall that if ϕ:Spec⁡C→Spec⁡B is finite flat, then C≃BN as a B-module. For any g∈C, mg:C→C given by multiplication by g is B-linear, and we define the norm

Here are some properties of the norm:

  1. (1)

    Nm is multiplicative;

  2. (2)

    For all b∈B⊂C, we have Nm⁡(b)=bN.

  3. (3)

    If we have a Cartesian diagram

    then f∗⁢Nmϕ⁡(g)=Nmϕ′⁡((f′)∗⁢g).

Choose generators 1,x2,…,xn of B as an A-algebra. Then there is a faithfully flat base change

Set B¯=B⁡[y1,…,yn], U¯=Spec⁡B¯, and R¯=R×ASpec⁡A¯. Therefore, we get a new finite flat groupoid R¯⇉U¯. Set

Lemma 4.27.

In fact, we have Nmt⁡(s∗⁢ϕ)∈BR⁢[y1,…,yn]⊂B¯.

Proof.

By definition of B¯R¯ we need to prove that s∗⁢Nmt⁡(s∗⁢ϕ)=t∗⁢Nmt⁡(s∗⁢ϕ). Consider the diagram

Then we see that

Because R¯⇉U¯ is a groupoid, there is a morphism

which is an isomorphism over R¯ via pr1. Now we have the diagram

This shows that

This gives us

Let D⊆BR be the A-algebra generated by the coefficients of F. Using the sequence of inclusions A⊆D⊆BR⊆B, we can use Artin-Tate to conclude that D is finitely generated as long as we know that B is finite over D. Note that F⁡(y1,0,…,0)=Nmt⁡(s∗⁢y1)=y1deg⁡(t), so if we consider the diagram

we see that multiplication by y1 is diagonalizable, and thus F is monic in y1. We claim now that F⁡(xi,0,…,0,−1,0,…,0)=0 and consequently that xi is integral over D. To prove this claim, we have e:U¯→R¯ such that t∘e=idU¯. This gives the diagram

and we see that t∗:B¯↪O⁡(R¯) has the contraction e∗. Therefore, we see that 𝒪⁡(R¯)=B¯⊕L where L≃Bdeg⁡(t)−1 as a B-module. We now compute

because t∘e=s∘e=1U.

Lemma 4.28.

There is a bijection (U/R)⁢(k)=Q⁡(k), where Q=Spec⁡BR.

Proof.

Consider the diagram

If f∈BR, then f⁡(q⁡(a))=s∗⁢(f)⁢(r) and f⁡(q⁡(b))=t∗⁢(f)⁢(r), where (a,b)=(s⁡(r),t⁡(r)). By invariance, we see that these two values are equal. In particular, the map (U/R)⁢(k)→Q⁡(k) is surjective for all algebraically closed fields k.

Now suppose that q⁡(a)=q⁡(a′) for geometric points a,a′ of U. Then F⁡(y1,…,yn)=Nmt⁡(∑yi⁢s∗⁢(xi)), and in particular at a we see that

Because all xi∈BR, we see that

Now consider the cartesian diagram

which can be extended to

Therefore, we obtain

and similarly

Because the two are equal, there exists b∈t−1⁢(a) and b′∈t−1⁢(a′) such that xi⁢(s⁡(b))=xi⁢(s⁡(b′)) for all xi. Because the xi generate B, we have s⁡(b)=s⁡(b′). Therefore, we see that

and therefore a∼a′. □

Now consider j:R→X×X, where s,t are flat and j is quasi-finite. We will find an étale open U at every closed geometric point x∈X such that R|U and U are separated, s and t are flat and quasi-finite, and x is fixed by R|U, namely that t⁡(s−1⁢(x))={x}.

Lemma 4.29.

The algebraic space j−1⁢(ΔX) is separated if and only if j is separated.

Proof.

Recall that there is a morphism e:X→R, so we get e:X→SX≔j−1⁢(ΔX).

Suppose that j′:SX→X is separated.

Exercise 4.30.

Suppose that f:Z→X is separated and s:X→Z is a section. Then s is a closed embedding.

By the exercise, e is a closed embedding. We now claim that the horizontal arrows in the diagram

are isomorphisms. Here, the top arrow is defined by r↦(r,e⁡(s⁡(r))), which is clearly a bijection on points, and the bottom arrow is defined by (r1,r2)↦(r1,i⁡(r1)∘r2), which is also a bijection on points. □

Lemma 4.31.

Let X and R of finite type and let (s,t) be flat with j quasi-finite with j−1⁢(X)→X separated. Suppose that x∈X is a geometric point. TO prove that x∈X has a étale neighborhood U such that U/R|U has a good categorical quotient, we can assume:

  1. (1)

    X and R are separated;

  2. (2)

    s,t are quasi-finite and flat;

  3. (3)

    x∈X is fixed by R, namely that t⁡(s−1⁢(x))={x}.

We can take an affine étale neighborhood of X, so we may assume that X=Spec⁡B is separated. By the previous lemma, j is separated, so R is separated and quasi-finite over X×X (which is a scheme), so R is also a scheme. Let F=s−1⁢(X). We claim that F is Cohen-Macaulay along j−1⁢(x,x).

Note that t:F→t⁡(F)¯ is quasi-finite, so

is finite. Because the Cohen-Macaulay locus in F is open and nonempty, there exists w∈F such that F is Cohen-Macaulay along t−1⁢(t⁡(w)). Consider h:F→s−1⁢(t⁡(w)) given by r↦r∘i⁡(w). Then let x′=t⁡(w)∼Rx. We see that s−1⁢(x′) is Cohen-Macaulay along j−1⁢(x′,x′).

Our next claim is the following. Suppose that s−1⁢(x) is Cohen-Macaulay along j−1⁢(x,x)=s−1⁢(x)∩t−1⁢(x). Let W be a closed subscheme defined by d=dimx(F) hypersurfaces whose intersection with O¯⁢(x)=t⁡(s−1⁢(x))¯=t⁡(F)¯ is 0-dimensional at x. Shrinking X in the Zariski topology, we can asume theoretically that

Consider the embedding

Then the composition p:t−1⁢(W)↪R→𝑠X is flat and quasi-finite.

Consider the diagram

where every square is Cartesian. Now note that

which is finite. Note that F is Cohen-Macaulay along F∩t−1⁢(H1)∩⋯∩t−1⁢(Hd), so because 0=dimF∩t−1⁢(H1)∩⋯∩t−1⁢(Hd)=dimF−d, we see that t−1(H1)|F,…,t−1(Hd)|F is a regular sequence, so by a lemma from commutative algebra, t−1⁢(W) is flat over X.

Lemma 4.32.

Let ϕ:(A,𝔪)→(B,𝔫) is a local homomorphism of Noetherian local rings and E=B/𝔪⁢B. Choose x1,…,xn∈𝔫. The following are equivalent:

  1. (1)

    x1,…,xn is a regular sequence in B and B/(x1,…,xn) is flat over A;

  2. (2)

    x¯1,…,x¯n is a regular sequence in E and B is flat over A.

Lemma 4.33 (Keel-Mori).

Let j:R→X×X be a pre-equivalence relation such that s and t are open. Suppose that g:W→X is such that p:S−1⁢(W)→X is surjective in the quasi-finite flat topology. Then X/R=W/R|W.

The upshot is that from flat s,t and quasi-finite j, we have reduced to flat and quasi-finite s,t.

Definition 4.34.

A flat groupoid R⇉X of schemes is split at x∈X if R=P⊔R2 where P is open and closed such that

  1. (1)

    P⇉X is a groupoid;

  2. (2)

    P⇉X is flat and finite;

  3. (3)

    j−1⁢(x,x)⊂P.

Proposition 4.35.

For a quasi-flat groupoid of schemes R⇉X and all x∈X, there is an étale affine (W,w)→(X,x) neighborhood of x such that R|W⇉W is split at w.

Proof.

Consider the diagram

Note that s is separated and quasi-finite and is therefore quasi-projective, so there exists a Hilbert scheme HilbR/X→X given by the functor

where the morphism to X is locally of finite type and separated. Denote this Hilbert scheme by H and HR=R×XH. Denoting the universal family by 𝒵, we have a diagram

Choose some t=(Zt,x)∈H and define W↪H to be the closed subscheme (see next lemma) of t∈H such that eH⁢(t)∈Zt.

Lemma 4.36.

Suppose p:X→Y is a morphism of schemes and F and G are two closed subschemes of X such that p|F:F→X is finite and flat. Then

is a closed subscheme of Y.

Proof.

This is local on Y, so we can shrink Y so that p∗⁢𝒪F is finite and free. Then we have

and therefore Gy↪Fy if and only if ℐFy⊆ℐGy if and only if ℐFy∩IGy=IFx. In addition, p∗⁢𝒪F↠p∗⁢𝒪F∩G because p|F is finite flat. Therefore, the sequence

is exact, where h is given by an s×t 𝒪Y-valued matrix ei⁢j. Then I=𝕍⁡(ei⁢j)↪Y. □

Now let x∈X⁡(k) be a geometric point. Now consider

Choose Z0↪s−1⁢(x) to be any collection of points containing e⁡(x). Now choose W=(Z0,w)∈H. We claim that g:H→X is étale at w and therefore W=H in a neighborhood of w.

To see this, by construction k⁡(x)=k⁡(w) by construction, so we need to show that g is formally étale. For any local Artinian pointed scheme (S,y) we want to show that

is a singleton, where we use the fact that H→X is formally étale.

We need to lift maps S→X to H. Consider the diagram

Note that |RS|top=|s−1⁢(x)|top=|Z0|top, so we take 𝒵S to be the open and closed subscheme of R supported on |Z0|top. Then 𝒵S×Sx=Z0, the map 𝒵S→S is flat and proper, and any 𝒵S′⊂RS with 𝒵S′×Sx=Z0 is a subscheme of 𝒵S. Finally, eS↪𝒵S, so a lift does exist. Uniqueness of the lift follows from the universal property of the Hilbert scheme.

W represents the functor which sends a scheme to the set of isomorphism classes of diagrams

such that e⁡(S)⊆Z. Consider the diagram

Our goal is to find a finite flat subgroupoid of R|W. There is a P↪W×XR which is finite flat over W. We need to find a map tP:P→W which gives a map P→R|W and a groupoid structure on P. We will define this tP on S-points for any scheme S. For y:S→P, this is a diagram

such that e⁡(S)⊆Z. Then the corresponding point in W⁡(S) is

where composition is taken inside the groupoid Rt⁡(a).

This P is given by the functor which sends a scheme S to the set of subschemes

such that e⁡(S)⊆Z. Also, note that W sends a scheme S to the set of subschemes Z↪s−1⁢(x) flat and proper over S such that x∈X⁡(S) and e⁡(S)⊆Z. On S-points, note that

Now note that

The target map is given by

Now we send

Note that

so P is closed under composition. The inverse is given by (Z,a)↦(Z∘a−1,a−1) and the identity is given by Z↦(Z,e), so P is a closed subgroupoid of R|W. By shrinking W further, we can also assume that P is open. □

We return to the big picture. Begin with j:R→X×X a flat groupoid with 𝒮≔j−1⁢(ΔX)→X finite. We first reduced to the case where j is quasi-finite, and then we reduced to the case where s and t are quasi-finite and flat and R and X are separated schemes. Then there exists an affine étale W→X such that R|W⇉W is split, which means that

  1. (1)

    R|W=P⊔R2 is a union of disjoint open and closed subschemes;

  2. (2)

    P⇉W is finite and flat;

  3. (3)

    P contains the inertia, namely 𝒮⊂P.

Remark 4.37.

Note that W/P=Spec⁡AP, where W=Spec⁡A.

The next step is the bootstrap. Suppose R⇉X is a quasi-finite and flat split groupoid of separated (even affine) schemes. Now consider the diagram

We obtain groupoids P⇉X and P×(s,t)R×(s,t)P⇉R which are finite and flat. Then X′≔X/P is a good moduli space by the finite flat affine case. Now suppose R′′ is a good moduli space for the second groupoid. Then s,t induce an étale equivalence relation R′′⇉X′ and X′/R′′=X/R. We will conclude using the finite quotient lemma.

Lemma 4.38 (Finite quotient lemma).

Suppose that j:R→X×X is a finite flat groupoid of separated schemes with j a closed immersion. Then X/R exists as a scheme.

Proof.

Consider the diagram

This defines a map q:X→Hilb⁡(X) given by q⁡(x)=t⁡(s−1⁢(x)). This is clearly R-invariant, so let Q be its scheme-theoretic image. Then X/R=Q. □