Patrick Lei

All notes

Minimal Model Program

Talks by Joaquín Moraga, Zoom, Spring 2021.

Minimal Model Program Learning Seminar
Spring 2021
Notes by Patrick Lei
Lectures by Joaquín Moraga

Zoom University

Disclaimer

These notes were taken during the seminar using the vimtex package of the editor neovim. Any errors are mine and not the speakers’. In addition, my notes are picture-free (but will include commutative diagrams) and are a mix of my mathematical style and that of the lecturers. If you find any errors, please contact me at plei@math.columbia.edu.

Also, I stopped attending the seminar after we finished the MMP in dimension 3. The seminar has continued onto multiplier ideals, and will later cover Birkar-Cascini-Hacon-McKernan and finally even more advanced topics in birational geometry.

Acknowledgements

I would like to thank Joaquín Moraga for organizing the seminar, finding mistakes in the notes, and more. I would also like to thank Nicolás Vilches for providing notes from the time I was unable to attend.

1 Overview

The goal of the minimal model program is to classify smooth projective complex varieties X⊆ℙn. Let TX be the tangent byndle and ΩX be the cotangent bundle. Then ωX=ΩXn is called the canonical bundle, and can be written as 𝒪X⁢(KX) for some Cartier divisor KX.

Question 1.0.1.

Can we understand the geometry of X using numerical properties of KX?

For a curve C⊆X and a line bundle ℒ on X, then ℒ.C=degC⁡(i∗⁢ℒ). Then KX is ample (resp. antiample) if KX.C>0 (resp <0) for all curves C⊆X. Similarly, KX is numerically trivial if KX.C=0 for all curves C⊆X.

Definition 1.0.2.

We say that X is Fano if KX is antiample, Calabi-Yau if KX is numerically trivial, and canonically polarized if KX is ample.

Example 1.0.3.

If C is a Fano curve, then C≃ℙ1. If C is a CY curve, then C is an elliptic curve. If C is a canonically polarized curve, then g⁡(C)≥2.

Example 1.0.4.

Let X⊆ℙn be a smooth hypersurface of degree d. Then by the adjunction formula, we have KX≃(Kℙn+X)|X≃(d−n−1)⁢H|. Therefore, X is Fano if d≤N, CY if d=n+1, and canonically polarized if d≥n+2.

Remark 1.0.5.

If E is an elliptic curve, then E×ℙ1 has KX.C=0 for some curves and KX.C<0 for others.

Now we consider various properties of different varieties:

Table 1.1: Properties of varieties of different classes
Fano CY Canonically polarized
π1 trivial ? generally infinite
Automorphisms linear algebraic groups ? finite groups
Birational automorphisms monstrous ? finite groups
Geometry simple geometry ? complicated, rich
Arithmetic a lot of ℚ-points ? ℚ-points in a proper closed

We have said nothing about Calabi-Yaus, but of course by the Beauville-Bogomolov decomposition, we can reduce to pure Calabi-Yaus (hi,0=0 for 0<i<dimX), hyperkählers, and abelian varieties up to taking a finite cover.

Now let x be a closed point on X. Then there is a variety Blx⁡X→X that is an isomorphism away from x where the fiber above x is an exceptional divisor E parameterizing tangent directions at x.

Example 1.0.6.

Consider points p1,…,pn,…∈ℙ2. Then if we blow up these points in a sequence, we obtain a sequence of varieties X1,…,X2,…,Xn,… Over ℙ2∖{p1,…,pi−1}, the morphism Xi→ℙ2 is an isomorphism. For i≠j, clearly Xi is not isomorphic to Xj (they have different Picard ranks), but they are birational. We say that X1∼birX2 if they have isomorphic dense open subsets.

Now we can state the goal of the Minimal Model Program. If X is projective and has “mild singularities” the goal is to prove that there exists a birational map π:X⇢X′ and a fibration (φ∗⁢𝒪X′=𝒪Z and positive dimensional general fiber) X′→𝜑Z such that one of the following holds:

  1. 1.

    F is Fano;

  2. 2.

    F is Calabi-Yau;

  3. 3.

    Z=Spec⁡ℂ and X′ is canonically polarized.

The way we will construct this birational morphism is by studying the geometry of curves on X which intersect KX negatively. If KX.C<0 under some hypotheses (extremity on N⁢E⁢(X)), we can find φC:X→X1 contracting precisely the curves which are numerically equivalent to a positive multiple of C.

  1. 1.

    If the curves numerically equivalent to a positive multiple of C cover X, then φC has positive-dimensional fibers, is a contraction, and the general fiber F is Fano. This is called a Mori fiber space.

  2. 2.

    If the curves numerically equivalent to a positive multiple of C cover a divisor on X, then we say that φC:X→X1 is a divisorial contraction and thus ρ⁡(X1)=ρ⁡(X)−1. X1 still has nice singularities, so we can iterate this process.

  3. 3.

    The last case is called a small contraction. The curves which are numerically equivalent to a positive multiple of C cover a set of codimension at least 2. In this case, X1 may have very bad singularities (by this, we mean that KX1 is not ℚ-Cartier). We construct a new birational morphism φC+:X+→X1 which contracts KX+-positive curves.

Another type of surgery is a flip, which changes a locus of codimension at least 2. For example, consider D=p1∗⁢𝒪⁢(1)⊗p2∗⁢𝒪⁢(r) on ℙ1×ℙ1. Then write

Then X is (locally) a cone over ℙ1×ℙ1, so KX is not ℚ-Cartier. Therefore, we can blow up the vertex, and the exceptional divisor is E≃ℙ1×ℙ1. Now the ℙ1×ℙ1 can be collapsed onto each of the two factors, so we obtain a birational map π:X1⇢X1+. Here, if C,C+ are the resulting curves, we have KX1.C<0 and KX1+.C+>0.

E≃ℙ1×ℙ1KX1.C<0X1CKX1+.C+>0X1+C+KX is not ℚ-Cartierflipπ bir. map
Figure 1.1: A Flip

This gives us the question:

Question 1.0.7.

Do flips always exist?

Now either the algorithm given by iterating this process always terminates or it continues infinitely. We are fine once we reach the Mori fiber space case, and there are only ρ⁡(X)−1 divisorial contractions, so the only possible problem is that there is an infinite sequence of flips.

Conjecture 1.0.8 (Termination of flips).

This algorithm always terminates after finitely many flips with either a Mori fiber space Xn→Z or a variety Xn such that KXn.C≥0 for every curve C (in other words, KXn is nef).

Conjecture 1.0.9 (Abundance).

X has mild singularities and KX is nef. Then |m⁢KX| is basepoint-free for some m≫0.

If this is true, and X→𝜑X1 contracts all KX-trivial curves, then either

  1. 1.

    The general fiber has positive dimension. In this case, KF≡0.

  2. 2.

    dimX=dimX1. Then X→X1 is birational and X1 is canonically polarized.

Therefore, the goal of the MMP is achieved if we can solve the conjectures of existence of flips, termination of flips, and abundance. Existence of flips was proved by Birkar, Cascini, Hacon, and McKernan in 2006 and termination is known in dimension at most 3 and in some cases in dimension 4. Finally, abundance is known in dimension at most 3.

2 MMP in Dimension 3

2.1 Rational Curves

Proposition 2.1.1 (Bend and Break).

Let X be proper and C be a smooth proper curve. Let p∈C and g0:C→X be nonconstant. Next, let 0∈D be a pointed curve and G:C×D→X such that

  1. 1.

    G|C×{0}=g0.

  2. 2.

    G⁡({p}×D)=g0⁢(p).

  3. 3.

    G|C×{t} is different from g0 for general t.

All of these imply that this is a nontrivial deformation of g0 fixing p. Then there exists g1:C→X and Z=∑ai⁢Zi a union of rational curves such that (g0)∗⁢C is algebraically equivalent to (g1)∗⁢(C)+Z and g0⁢(p)∈⋃iZi. In particular, there exists a rational curve through g0⁢(p).

Proof.

First, compactify D and let G¯:C×D¯⇢X be the rational map. This map is undefined at {p}×D¯ by the rigidity lemma, so let S be the normalization of the graph of G¯. So we have a map π:S→C×D¯ and write GS:S→.

Then we define h:S→C×D¯→D¯. Then there exist d∈C×D¯ such that π is not an isomorphism over d. Then we know that h−1⁢(d)=C′+E where C′ is a birational transform of C and E is π-exceptional. Then we set g1:C→X to be the restriction of GS to C′ and Z=GS⁢(E).

By a lemma of Abhyankar, we know that E is a union of rational curves, and then by the Lüroth theorem, we know that Z is a union of rational curves and

□

Lemma 2.1.2 (Abhyankar).

Let X have mild singularities and Y→𝜋X be a proper birational morphism. For any x∈X, either π−1⁢(x) is a point or is covered by rational curves.

Here is an intuitive image of the bend-and-break process:

SEC×DDXZxx∈Z∪g1∗⁢C′
Figure 2.1: Bend and Break
Proposition 2.1.3 (Bend and Break II).

Let X be a projective variety and g0:ℙ1→X be a nonconstant morphism. Let D be a smooth pointed curve and G:ℙ1×D→X such that

  1. 1.

    G|ℙ1×{0D}=g0;

  2. 2.

    G⁡({0}×D)=g0⁢(0),G⁡({∞}×D)=g0⁢(∞);

  3. 3.

    G⁡(ℙ1×D) is a surface.

Then (g0)∗⁢ℙ1 is algebraically equivalent either to a reducible curve or a multiple curve.

Proof.

Let S be a ℙ1-bundle containing ℙ1×D and consider the rational map G~:S⇢X. Then we can resolve the basepoints to obtain G~:S~→S and induct on ρ⁡(S~/S)≕ρ.

Case 1: ρ=0:

Consider the sections C0,C∞ at 0 and ∞. Then let H is ample on X and then we see that (G~∗⁢H)2>0 and (C0⋅G~∗⁢H)=(C∞⋅G~∗⁢H)=0 by the projection formula. By the Hodge index theorem, we see that C02<0,C∞2<0 (because G~∗⁢H,C0,C∞ are linearly independent). But then we know that ρ⁡(S)=2, which is a contradiction.

Inductive step:

Consider the diagram

Then r is the first blowup in S~→S∋P and y∈D will be a point such that P∈q−1⁢(y). Let F1 be the exceptional divisor of r and F2 be the strict transform of q−1⁢(y) in S′. Then F1,F2 intersect at a point Q. But then G¯′ is a morphism around F2. But then (g0)∗⁢ℙ1∼G~∗⁢((q∘r)∗⁢(y)), which is reduced and irreducible. If G¯ is not defined at Q≠P, then

which is a contradiction and thus G¯ is defined at Q≠P. Then if G¯′ is not defined at Q0, then after blowing up Q0, we see that (q∘r)∗⁢(y) must contain a component of multiplicity at least 2. Now contracting F2, we have the desired result by induction on ρ. □

This tells us that to produce rational curves, we simply need to deform them with enough fixed points and use bend and break. But now we need to actually find rational curves.

Theorem 2.1.4.

Let X be smooth and projective and −KX be ample. For every x∈X, there exists a rational curve C through x such that

Proof.

Choose some curve C⊆X through x. Then the space of deformations of C on X fixing x has dimension at least

We have several cases:

  1. 1.

    If g⁡(C)=0, then we are done.

  2. 2.

    If g⁡(C)=1, then we can replace f with the composition by an endomorphism of large degree n, then we see that

    whenever n is sufficiently large

  3. 3.

    Assume g⁡(C)≥2. Then there are no endomorphisms of high degree, so assume X,C are defined over ℤ. Then let Xp,Cp be the reduction to F¯p. Now we apply the Frobenius map Fp, which has degree p. By generic flatness, we know that (fp)∗⁢Cp.KXp,g⁡(Cp),χ⁢(TX|Cp) are the same for almost all p, so by the same argument as in the genus 1 case, we see have a rational curve Ap on Xp for almost all p. By bend and break II, we can find a rational curve of the desired degree.

    Then we use the fact that if a statement holds for all p large enough, then it holds for the complex numbers, and we obtain a curve. This is analogous to the idea that if Z⊆ℙℤn, then if the image of π:Z→Spec⁡ℤ contains a Zariski-dense subset, then it contains the generic point. □

Theorem 2.1.5.

Let X be a smooth projective variety and let H be ample on X. Assume there exists C′⊆X such that −(C′.KX)>0. Then there exists a rational curve E such that dimX+1≥−(E.KX)>0 and

Theorem 2.1.6 (Cone Theorem).

Let X be smooth and projective. Then there exist countably many curves Cr⊆X such that 0<−KX.Ci≤dimX+1 and

Proof.

Choose Ci with 0<−(C.KX)≤dimX+1 and let W be the closure of NE¯K≥0+∑iℝ≥0⁢[Ci]. Now choose D positive on W∖{0} and negative somewhere on NE¯⁢(X). Let H be ample and

This means that H+μ⁢D is nef. Then let Z∈NE¯⁢(X) with (H+μ⁢D).Z=0 and KX.D<0. Let Zk be a sequence of curves approximating Z. Then we see that

is obtained by Zk0. Now we will replace Zk with rational curves Ei⁡(k) such that dimX+1≥−Ei⁡(k).KX>0 and

Because Ei⁡(k).D≥0, we have

Fixing M≫0 such that M⁢H+KX is ample, then we see that (M⁢H+KX).Ei⁡(k)>0, so

Taking k→∞,μ′→μ, we see that

a contradiction. □

Example 2.1.7.

Suppose KX is not nef. Then there are no rational curves on X. For example, there are no rational curves of an abelian variety.

2.2 Singularities of the MMP

Consider pairs (X,Δ) such that X is normal quasiprojective and KX+Δ is ℚ-Cartier. These are called log pairs.

Definition 2.2.1.

Let π:Y→X is a resolution of singularities, E⊆Y is exceptional, and π∗⁢(KX+D)+∑Ei has simple normal crossings. Define the log discrepancy

Definition 2.2.2.

We say that (X,Δ) is

  1. 1.

    terminal if aE⁢(X,Δ)>1 for every exceptional E over Y.

  2. 2.

    canonical if aE⁢(X,Δ)≥1 for every exceptional E over Y.

  3. 3.

    Kawamata log terminal if aE⁢(X,Δ)>0 for every E.

  4. 4.

    log canonical if aE⁢(X,Δ)≥0 for every E.

Now if X is smooth projective and KX is pseudoeffective, then there exists X⇢Xter such that KXter is nef. By abundance, KXter is semiample. Then there is a morphism Xter→X𝔠⁢𝔞⁢𝔫 such that KXcan is ample. Then terminal singularities are those that may appear in the terminal model, and canonical singularities are those that may appear on the canonical model.

Recall the adjunction formula: If (X,D) is log smooth, then KX+D|D∼KX. Usually, (X,D) is log canonical but not klt. But then aD⁢(X,D)=0 but aE⁢(X,D)>0 for every E≠D. Terminal singularities are the smallest category of singularities that we need to understand to run the minimal model program. On the other hand, log canonical is the largest class of singularities in which we can expect the MMP to work.

Example 2.2.3 (Examples of klt singularities).

Both cone singularities and quotient singularities are klt.

Proposition 2.2.4 (Cones).

Let (X,Δ) be a log pair and A an ample Cartier divisor on X. Then define

Then C⁡(X,A) is

  1. 1.

    terminal if and only if r⁢A∼ℚKX+Δ with r<−1 and (X,Δ) terminal;

  2. 2.

    canonical if and only if r⁢a∼ℚkx+δ with r≤−1 and (X,Δ) canonical;

  3. 3.

    klt if and only if r⁢a∼ℚkx+δ with r<0 and (X,Δ) is klt;

  4. 4.

    log canonical if and only if r⁢a∼ℚkx+δ with r≤0 and (X,Δ) is log canonical.

In particular, the cone over a Fano is klt, the cone over a Calaby-Yau is log canonical, and cones over canonically polarized varieties are terrible.

Example 2.2.5.

Consider the cone Cn over a rational normal curve of degree n. Then resolving Yn→Cn, we see that the exceptional En≃ℙ1 and

and therefore aEn⁢(Cn)=2n.

Consider E⊆ℙ3 and elliptic curve. Then π∗⁢(KCE)=KYE+E, so aE⁢(CE)=0.

Now consider G⊆G⁢Ln a finite group. Then ℂn/G=Spec⁡ℂ⁢[x1,…,xn]G has klt singularities.

Example 2.2.6.

In dimension 2, we have

  1. 1.

    terminal is equivalent to smooth;

  2. 2.

    canonical is equivalent to ADE;

  3. 3.

    klt is equivalent to quotient singularity;

  4. 4.

    log canonical is “equivalent” to quotient and/or elliptic cone.

Example 2.2.7.

In dimension 3, terminal singularities are classified as quotients of hypersurface singularities, also known as hyperquotient singularities. Then are given by actions of finite groups G acting on hypersurface singularities of the form {x2+y2+f⁡(z,w)=0}. Even for canonical singularities, we have no idea what they look like.

Example 2.2.8.

In dimension 4, there are examples of 4-fold terminal singularities with analytic embedding dimension n for every n (Kollar, 2010). By contrast, terminal singularities in dimension 3 all have analytic embedding dimension 4.

Theorem 2.2.9 (Prokhorov, Xu, 2019).

Any klt singularity deforms to a klt cone singularity. For x∈X, there exists a flat morphism 𝒳→𝜑𝔸1 such that φ⁡(𝔸1∖0)∼(𝔸1∖0)×X and φ−1⁢(0)=X0 is a klt cone singularity. This is just a deformation to the normal cone.

We will apply the following philosophy:

Any theorem for smooth projective varieties should work with klt singularities.

Example 2.2.10.

Let KX=0. By Beauville-Bogomolov (1970s), there exists a cover X←Y, where Y is a product of abelian varieties, irreducible Calabi-Yau varieties, and hyperkählers. There is an analogue for klt singularities that was proved by Druel, Campana,…in 2020.

Example 2.2.11.

Let X be smooth projective and −KX be nef. Then

where Yi is strict Calabi-Yau, Sk is hyperkahler, and Z is rationally connected. There is a version for klt singularities in progress.

We will now discuss localization of singularities. Suppose X is a variety with terminal singularities and Z⊆X a subvariety of codimension 1. Then Spec⁡𝒪X,Z has terminal singularities if and only if Spec⁡𝒪X,Z is a smooth local ring, and this is equivalent to smoothness at the generic point of Z. In particular, if X is terminal, the singularities must appear in codimension at least 3.

2.3 Vanishing

Let C⊆X be a KX-negative curve. Then if KX∼ℚE≥0 for some E effective, we see that KX.C=E.C<−, so C⊆E. Now if (X,D) is a log smooth pair, we have an exact sequence

If H1⁢(X,𝒪X⁢(KX))=0, then H0⁢(KX+D)↠H0⁢(KD). This tells us that vanishing theorems help us find sections of line bundles.

Theorem 2.3.1 (Kodaira Vanishing).

Let X be smooth projective and ℒ be ample. Then Hi⁢(X,ℒ−1)=0 for all i<dimX.

Sketch of Proof.

Let s∈H0⁢(X,ℒm) and D=(s=0) be smooth. Then we have 𝒪X→𝑠ℒm. Now we have

Setting Z=Spec⁡⨁i=0m−1ℒ−i with projection p:Z→X, we note that if X,D are smooth, then Z is smooth. Consider a morphism τ:Hi⁢(Z,ℂZ)↠Hi⁢(Z,𝒪Z) and its pushforward p∗⁢τ:Hi⁢(X,p∗⁢𝒞Z)↠Hi⁢(X,p∗⁢𝒪Z). Now we consider the surjection

Now we use the result that ℂ⁡[ζr]↪ℒ−r factors through ℂ⁡[ζr]↪ℒ−r⁢(−k⁢D)↪ℒ−r. By Serre, we see that Hi⁢(X,ℒ−(r+m⁢k))=0 for arbitrary k. □

Theorem 2.3.2 (KV vanishing).

Let X be a smooth projective complex and ℒ be a line bundle on X such that ℒ≡M+∑ai⁢Di, where M is a big and nef ℚ-divisor, ∑Di is a snc divisor, and 0≤ai<1,ai∈ℤ. Then Hi⁢(X,ℒ−1)=0 for i<dimX.

Proposition 2.3.3.

Let X be quasiprojective and normal, D Cartier, and m>0 a positive integer. Suppose Y→𝑝X is finite and D′ Cartier such that p∗⁢D∼M⁢D′. If X is smooth and ∑Fj is simple normal crossing, then Y is smooth and ∑p∗⁢Fj has simple normal crossings.

Lemma 2.3.4.

Let Y→X be finite. Then 𝒪X→f∗⁢𝒪Y splits. If ℱ is a coherent sheaf on X, then ℱ is a direct summand of f∗⁢f∗⁢ℱ. Finally, Hi⁢(X,ℱ) is a direct summand of Hi⁢(Y,f∗⁢ℱ).

Sketch of KV Vanishing.

Consider ∑ai⁢Di and write a1=b/m for some m≥0. Then consider p1:X1→X such that p∗⁢D1∼m⁢D. Then Hi⁢(X,ℒ−1) is a direct summand of Hi⁢(X1,p1∗⁢ℒ−1). Then p1∗⁢D1 is a section of 𝒪X⁢(m⁢D), so we can apply the index cover to obtain X2→p2X with X2 smooth, p2∗⁢(Di) is smooth, and ∑p2∗⁢p1∗⁢Di has simple normal crossings. Then

so

Choosing j=b, we see that Hi⁢(X1,p1∗⁢ℒ−1) is a direct summand of Hi⁢(X2,p2∗⁢p1∗⁢ℒ−1⁢(b⁢D)). Therefore,

But now we have reduced the number of components of Di, so by induction, the cohomology of the pullback vanishes. Now we need to consider M. We know M is big and nef, so M∼ℚA+E, where A is ample and E is effective. Then there exists f:Y→X projective and birational such that f∗⁢ℒ=A+E, where A is ample and E has simple normal crossings with E=∑ai⁢Ei,0≤ai<1. Now let H⊆X be an ample divisor. Then

But then f∗⁢ℒ⁢(r⁢H)=(A+r⁢f∗⁢H)+E, where A+r⁢f∗⁢H is ample. But now we know that

for k>0, so

and therefore Rj⁢f∗⁢ωY=0 for j>0. But now setting r=0, we see that

Theorem 2.3.5.

Let (X,Δ) be a proper klt pair with N a ℚ-Cartier divisor. Suppose that N=M+Δ, where M is big and nef. Then Hi⁢(X,𝒪X⁢(−N))=0 for i<dimX. Equivalently, Hn−i⁢(X,KX+N)=0 for n−i>0.

Remark 2.3.6.

This philosophy fails for log canonical singularities.

2.4 Cone Theorems

Note that Kollar-Mori use outdated notation. What they call klt is actually sub-klt and what they call klt with Δ≥0 we call klt. This still confuses people today, so let’s blame the person with the Fields medal.

Theorem 2.4.1 (Non-vanishing).

Let X be proper with (X,Δ) sub-klt1 and D be a nef Cartier divisor. Assume a⁢D−(KX+Δ) is big and nef for some a≥0. Then for m≫0, we have

Theorem 2.4.2 (Basepoint-freeness).

Let X be proper and (X,Δ) be klt. Suppose D is a nef Cartier divisor. Assume a⁢D−(KX+Δ) is big and nef for some a≥0. Then for m≫0, the linear sustem |m⁢D| is basepoint-free.

Theorem 2.4.3 (Rationality).

Let X be proper and (X,Δ) be klt. Suppose that KX+Δ is not nef, a⁡(KX+Δ) is Cartier, and H is a nef and big Cartier. Define

Then r is rational and its denominator is at most a⁡(dimX+1).

Theorem 2.4.4 (Cone Theorem).

Let (X,Δ) be a projective klt pair.

  1. 1.

    There are countably many Ci⊆X such that 0<−(KX+Δ).Ci≤2⁢dimX and

  2. 2.

    For any H ample and ε>0, we have

  3. 3.

    If F⊆N⁢E¯⁢(X) is extremal and (KX+Δ)-negative, then there exists a contraction morphism contF:X→Z such that C⊆X is mapped to a point if and only if [C]∈F.

  4. 4.

    Let contF:X→Z be as above and ℒ be a line bundle on X such that ℒ.F=0. Then there exists ℒZ on Z such that ℒ≃contF∗⁢ℒZ.

The logical structure is this: We use non-vanishing to find sections, then use various techniques (Kodaira vanishing) to lift enough sections to get basepoint-freeness, then we prove rationality by studying linear systems of the form |p⁢H+q⁢KX|, and finally we get the cone theorem as a formal consequence of convex geometry. However, we will follow the order of Kollar-Mori.

Remark 2.4.5.

The entire discussion will be carried out using klt singularities. However, these results hold when (X,Δ) is log canonical at the cost of replacing nef and big with ample and at the cost of using significantly more machinery that was developed in the last 15 years.

Proof of basepoint-freeness.

By non-vanishing, we know that H0⁢(X,m⁢D)≠0 for m≫0. If B⁡(s) is the base locus of |s⁢D|, it suffices to prove that for Bs=B⁡(m)≠0. Next, we may consider a log resolution

Now we may perturb KY such that

is ample. Therefore f∗⁢|m⁢D|=|μ|+∑rj⁢Fj, so ∑rj⁢Fj is the fixed part. Therefore Bs=⋃{f⁡(Fj)∣rj>0} and f−1⁢Bs⁢|m⁢D|=Bs⁢|m⁢f∗⁢D|.

We want to prove that there exists Fj with rj≥0 such that for all b≫0, Fj is not contained in Bs⁢|b⁢f∗⁢D|. Let b>0 be an integer, c>0 be rational, and b>c⁢m+a. Then we define

But now we see that the first term is is nef, the second term is basepoint-free, and the final two terms form an ample divisor. Therefore N⁡(b,c)is ample. By Kodaira vanishing, we see that H1⁢(Y,⌈N⁡(b,c)⌉+KY)=0 and ⌈N⁡(b,c)⌉=b⁢f∗⁢D+∑⌈−c⁢rj+aj−pj⌉⁢Fj−KY. Now

where A≥0 is effective and F=Fk′ is prime. Therefore

and we have the exact sequence

Therefore we have a surjection H0⁢(Y,b⁢f∗⁢D+⌈A⌉)↠H0⁢(F,(b⁢f∗⁢D+⌈A⌉)|F) for b≥c⁢m+a. Now ⌈A⌉ is f-exceptional, so

Now by non-vanishing, H0⁢(F,(b⁢f∗⁢D+⌈A⌉)|F)≠0, so H0⁢(Y,b⁢f∗⁢D+⌈A⌉) has a section not vanishing on F. Because ⌈A⌉ is f-exceptional,

by the negativity lemma. Here, if 0≤E∼b⁢f∗⁢D+⌈A⌉, then E−⌈A⌉∼b⁢f∗⁢D∼ℚ,X0, so if E−⌈A⌉≥0, then f∗⁢(E−⌈A⌉)=f∗⁢E=0. However, F must be disjoint from E, so we have found a section E of b⁢f∗⁢D disjoint from F. Therefore f∗⁢E is a section of b⁢D disjoint from W. □

Lemma 2.4.6 (Negativity lemma).

Let h:Z→Y be birational and proper between normal varieties and −B be h-nef. Then B is effective if and only if h∗⁢B≥0. If B is effective, either h−1⁢(Y)⊆supp⁡B or h−1⁢(Y)∩supp⁡B=∅.

Theorem 2.4.7.

Let (X,Δ) be a proper klt pair and KX+Δ be big and nef. Then the graded ring

is finitely-generated over ℂ.

This result holds even after dropping the big and nef assumption and was proved by Birkar, Cascini, Hacon, and McKernan in 2006.

Conjecture 2.4.8 (Abundance).

Let (X,Δ) be projective and klt. If KX+Δ is nef, then it is semiample.

Conjecture 2.4.9 (Effectivity, folklore).

Let (X,Δ) be projective klt. If KX+Δ is pseudoeffective (in the closure of the effective cone), then KX+Δ is effective.

2.4.1 Proof of the Cone Theorem

We will now prove the cone theorem. This relies on the following result from convex geometry.

Theorem 2.4.10.

Let Nℤ⊆Nℚ⊆Nℝ and N⁢E¯⊆Nℝ be a closed strictly convex clone. Let K∈Nℚ∗ such that (K.C)<0 for some C∈N⁢E¯. Assume there exists α⁡(K)∈ℤ≥0 such that for all H∈Nℤ∗ with H>0 on N⁢E¯∖{0} and that

is rational of the form U/α⁡(K). Then

with ξi∈Nℤ with (ξi.K)<0 and such that ℝ≥0⁢[ξi] do not accumulate in KX<0.

Let H be ample and Cartier. Suppose L is nef and define FL=L⟂∩N⁢E¯. Then for n∈ℤ≥0, we can set

Then rK⁢(n,H)∈ℤ>0 is non-decreasing with respect to n. Now if ξ∈FL∖N⁢E¯K≥0, then

Therefore rL⁢(n,H) is bounded above, integral, and non-decreasing, so this sequences stabilizes for n large enough to rL⁢(H). Now define the divisor

so FD(nL.H)⊆N⁢E¯K<0∪{0}. To prove this, let ξ∈FD⁡(n⁢L,H) with ξ∉FL. Then we know

For n′≫n, we have

so ξ∉FD⁡(n′⁢L,H). Because L is nef, FD⁡(n′⁢L,H)⊊FD⁡(n⁢L,H). If FD⁡(n′⁢L,H)⊆FL, then we stop. If not, we can iterate the above process to decrease dimFD⁢(n′⁢L,H) again, so the desired result eventually holds. Now 0≠FD(nL.H)⊆FL holds up to replacing n with a large multiple.

Now we will show that for some H, dimFD(nL.H)<dimFL. If Hi is a basis for FL∗, the linear functions

cannot all vanish, so dimFD⁡(n⁢L,Hi)<dimFL for some i. Now we can reduce to FL′⊆FL of dimension 1. This implies that N⁢E¯ and N⁢E¯K≥0+∑dimFL=1FL have the same closure.

Now we need to show that the FL do not accumulate in K<0. This is a formal argument in linear algebra, so we skip it. Next, we need to prove that

In the limit as ε→0, we produce countably many FL (with a formal argument that is omitted here).

The next step is to prove that if F⊆N⁢E¯⁢(X) is a (KX+Δ)-negative face, then there exists a nef Cartier divisor D such that FD=F. Let ⟨F⟩ be the linear span of F and V⊆N⁢(X)∗ be the set of linear functions vanishing on ⟨F⟩. Because the generators of F are defined over ℚ, then V is also defined over ℚ. Take ε>0 small enough such that KX+Δ+ε⁢H is negative on F. Because F is extremal, we know ⟨F⟩∩N⁢E¯⁢(X)=F. Therefore,

is a closed strictly convex cone intersecting ⟨F⟩ at the origin. We also note that N⁢E¯=WF+F, so we can find a lattice point p∈V such that (p=0)⊇⟨F⟩ and (p=1)∩WF=0. Therefore we can find a Cartier divisor D which gives a supporting function of F⊆N⁢E⁢(X).

Now by assumption, −(KX+Δ) is positive on F. This means that m⁢D−(KX+Δ) is strictly positive on N⁢E¯⁢(X)∖{0} for m≫0, so |m⁢D| is basepoint-free. Now let gF:x→Z be the contraction associated by the Stein factorization to the linear system |m⁢D|. Because gF is not an isomorphism, it constracts some curve C. Similarly to teh smooth case, we may assume that

Finally, we prove that any line bundle ℒ on X such that ℒ.F=0 descends to Z, which means there exists a line bundle ℒZ on Z such that ℒ=gF∗⁢ℒZ. Now let D be a Cartier divisor supporting F. We know WF⊆N⁢E¯⁢(X) and that gF is defined by m⁢D¯. Therefore, both m⁢D and (m+1)⁢D are pullbacks of Cartier divisors on Z. Write m⁢D=gF∗⁢D1,(m+1)⁢D=gF∗⁢(D2), and therefore we see that D=(m+1)⁢D−m⁢D=gF∗⁢(D2−D1). This implies that D is the pullback of a Cartier divisor on Z. If ℒ.F=0, then ℒ+m⁢D also supports F, so ℒ+m⁢D=gF∗⁢MZ for some Cartier divisor MZ of Z. We simply set ℒZ=𝒪Z⁢(MZ−D1). □

2.4.2 Nonvanishing and Rationality

First we will prove rationality assuming non-vanishing and basepoint-freeness and then we will prove rationality.

Lemma 2.4.11.

Let Y be a smooth projective variety and D1,…,Dn be Cartier divisors. Then let A be a normal crossing divisor with ⌈A⌉≥0. Define

Assume that for certain ui, ∑ui⁢Di is nef and ∑ui⁢Di+A−KY is ample. Then P⁡(u1,…,un) is a nonzero polynomial of degree at most dimY.

Proof.

For m≫0, the sub ∑m⁢ui⁢Di+A−KY is still ample. Then we know Hi⁢(∑m⁢ui⁢Di+⌈A⌉)=0 for i>0 by KV vanishing. By non-vanishing, we know h0⁢(∑m⁢ui⁢Di+⌈A⌉)≠0 and thus χ⁢(∑m⁢ui⁢Di+⌈A⌉)≠0 and thus P⁡(m⁢u1,…,m⁢un)≠0. □

Lemma 2.4.12.

Let P⁡(x,y)≠0 be a polynomial of degree at most n. Assume P vanishes for all sufficiently large integral solutions of 0<a⁢y−r⁢x<ε for a∈ℤ>0 and ε∈ℝ>0. Then r is rational and in reduced form it has denominator at most a⁡(n+1)/ε.

This is a purely arithmetic fact, so proof is omitted.

Proof of Rationality.

First, we reduce to the case in which H is basepoint-free. Define

By basepoint-freeness, we know that |H′| is basepoint-free. For m≫c≫d>0, we know r⁡(H)=r⁡(H′)+m⁢d⁢am⁢c and thus rationality of r⁡(H) is equivalent to r⁡(H′). If the denominator of r⁡(H′) divides v, then the denominator of r⁡(H) divides m⁢c⁢v. Replacing H with H′, now H is basepoint-free.

Now we need to study the base locus L⁡(p,q), which is the base locus of |p⁢H+q⁢a⁢(KX+Δ)|. Then we know L⁡(p,q)=X if and only if |p⁢H+q⁢a⁢(KX+Δ)|=∅. If p,q are large enough to be in the strip between a⁢y−r⁢x=0,a⁢y−r⁢x=ε, then L⁡(p,q) stabilizes. To see this, the x⁢H direction is semiample and then the base locus stabilizes by Noetherian induction to some L0. Now define I⊆ℤ×ℤ to be the set of (p,q) such that 0<a⁢q−r⁢p<ε and L⁡(p,q)=L0. Now I contains arbitrary large lattice points.

Next, we will define the polynomial P⁡(x,y) and prove that it does not vanish. Define p:Y→X to be a log resolution of (X,Δ). Then if D1=p∗⁢H,D2=p∗⁢(a⁡(KX+Δ)),KY=p∗⁢(KX+Δ)+A where ⌈A⌉≥0 is p-exceptional, then P⁡(x,y)≔χ⁡(x⁢D1+y⁢D2+⌈A⌉) is a polynomial of degree at most dimY=dimX=n. Then note that if y=0,x≫0, D1 is big and nef, so P≠0. Furthermore, we know that

From now on we will assume that r is not rational.

Now we show that L0≠X. If 0<a⁢y−r⁢x<1, then

is big and nef. Thus Hi⁢(Y,x⁢D1+y⁢D2+⌈A⌉)=0 for i>9. For (p,q) large enough we know P⁡(p,q)≠0 by the first lemma, so h0⁢(Y,p⁢D1+q⁢D2+⌈A⌉)≠0. But this implies that |p⁢H+q⁢a⁢(KX+Δ)|≠∅, so L0≠∅.

We show that L⁡(p′,q′)⊊L0 for (p′,q′) large in the strip. This will lead to a contradiction. Fixing (p,q)∈I, let f:Y→(X,Δ) be a log resolution satisfying:

  1. 1.

    The divisor f∗⁢(p⁢H+(q⁢a−1)⁢(KX+Δ))−∑pj⁢Fj is ample;

  2. 2.

    KY≡f∗⁢(KX+Δ)+∑aj⁢Fj for aj>−1.

  3. 3.

    f∗⁢|p⁢H+q⁢a⁢(KX+Δ)|=|L|+∑rj⁢Fj, where |L| is the movable part and ∑rj⁢Fj is the fixed part.

Then we can choose c>0 and pj>0 such that

where F is prime and ⌈A′⌉≥0 and A′ does not contain F in its support. Now F maps to a component B of L⁡(p,q)=f⁢(⋃rj>0Fj). Now define

We can choose (p′,q′) with a⁢q′−r⁢p′<a⁢q−r⁢p. Then (q′−(1+c)⁢q)⁢a<r⁡(p′−(1+c)⁢p), so

is nef. We conclude that N⁡(p′,q′) is ample because c⁢L is nef and the second term in the sum is ample. Therefore,

By the adjunction formula, we know

Applying the lemmas, we conclude that

and thus H0⁢(Y,f∗⁢(p′⁢H+q′⁢a⁢(KX+Δ))) contains a section Γ≥0 not vanishing at F. Running the same argument using the negativity lemma implies that Γ actually is disjoint from F and thus 0≤f∗⁢Γ∼|p′⁢H+q′⁢a⁢(KX+Δ)| is a section disjoint from B=f⁡(F)⊆L0. Therefore L⁡(p′,q′)⊊L0 and thus r is rational.

Now we need to control the denominator of r. Assume the denominator is larger than the constant given by the second lemma. Setting ε=1, we choose (p,q) large with 0<a⁢q−r⁢p<1. Then we have P⁡(p,q)=h0⁢(Y,p⁢D1+q⁢D2+⌈A⌉)≠0, so |p⁢H+q⁢a⁢(KX+Δ)|≠∅ for all (p,q)∈I. Now choose (p,q) such that a⁢q−r⁢p is the maximum, equal to dv. Then we can show that χ=h0≠0 for (f∗⁢(p′⁢H+q′⁢a⁢(KX+Δ))+⌈A′⌉)|F. Then there exist (p′,q′) large enough in 0<a⁢q′−r⁢p′<1 with ε=1 and a⁢q′−r⁢p′<dv=a⁢q−r⁢p. Running the same argument from the previous paragraph gives us the desired conclusion. □

Proof of non-vanishing.

First we will reduce to the case where X is smooth and a⁢D−(KX+Δ) is ample. Choose f:X′→X be a projective resolution and suppose f∗⁢(KX+Δ)−KX′+Δ′ and (X′,Δ′) is a sub-klt pair. Then we know a⁢f∗⁢D−(KX′+Δ′)=f∗⁢(a⁢D−(KX+Δ)) is nef and big, so a⁢f∗⁢D−(KX′+Δ′)−F is ample, so (X′,Δ′+F) is sub-klt. Writing Δ′′=Δ′+F, we see that f∗⁢(Δ′′)≤Δ and

Replacing X,Δ with (X′,Δ′′), we have the desired reduction.

Next, we need to rule out the case that D is numerically trivial. By KV vanishing, we know

Thus we may assume that D is not numerically trivial.

Now we show that there exists q0 such that if x∈X is not in the support of Δ, then for q≥q0 we can find M⁡(q)≡(q⁢D−(KX+Δ)) with multiplicity greater than 2⁢dimX. For A ample and e>0, we have De⁢Ad−e≥0, so we conclude that

Because a⁢D−(KX+Δ) is ample, then (a⁢D−(KX+Δ))d−1=C+eff, where C is a curve satisfying C⋅D>0. Therefore, (q⁢D−(KX+Δ))d→∞ as q→∞. Now

Thus if M⁡(q,e)∈|e⁡(q⁢D−(KX+Δ))|, imposing that M⁡(q,e) has multiplicity greater than 2⁢d⁢e at x imposes at most edd!⁢(2⁢d)d+O⁡(ed−1) conditions. As q→∞, then we know (q⁢D+(KX+Δ))d>(2⁢d)d, so for q large enough, some section satisfies the condition. Now we define M⁡(q)≔M⁡(q,e)/e, so M⁡(q)∈|q⁢D−(KX+Δ)| has multiplicity at least 2⁢d at x.

Next, consider a log resolution of (X,Δ+M⁡(q)) that dominates Blx⁡X. Then

  1. 1.

    KY≡f∗⁢(KX+Δ)+∑bj⁢Fj, bj>−1;

  2. 2.

    f∗⁢(a⁢D−(KX+Δ))−∑pj⁢Fj is ample for 0<pj≪1;

  3. 3.

    f∗⁢(M⁡(q))=∑rj⁢Fj where F0 corresponds to the exceptional divisor of the blowup at x.

Now we will perturb the coefficients and lift from lower dimension. Set

This is ample as long as c≤12 and b≥a+c⁡(q−a). Because we can choose b arbitrarily large, the second condition is always achievable. But now because x∉Supp⁡(Δ), we know b0=d−1 and r0>2⁢d, so c<1+(d−1)−p12⁢d<12. Therefore we can write N⁡(b,c)=b⁢f∗⁢D+A−F−KY, so

Because N⁡(b,c) is ample, then

so H0⁢(X,b⁢D−⌊Δ⌋)≠0 as long as H0⁢(F,(b⁢f∗⁢D+⌈A⌉)|F)≠0. But by the non-vanishing theorem in lower dimension, this is true. □

2.4.3 The Relative Case

We will state a relative version of the cone theorem. The proof is essentially the same as in the absolute case.

Theorem 2.4.13 (Relative Cone Theorem).

Let X→𝜑Z be a projective contraction of varieties over an algebraically closed field 𝕜 of characteristic 0. Let (X,Δ) be a klt pair.

  1. 1.

    There are countably many curves Cj⊆X such that φ⁡(Cj)=pt, 0<−(KX+Δ).Cj<2⁢dimX, and

  2. 2.

    For any ε>0 and φ-ample H, we have

  3. 3.

    Let F⊆N⁢E¯⁢(X/Z) be a (KX+Δ)-negative extremal face. Then there is a unique contraction X→contFY commuting with the structure morphisms to Z such that C⊆X is mapped to a point if and only if [C]∈F.

  4. 4.

    Let ℒ be a line bundle on X such that ℒ.C=0 for every curve C with [C]∈F. Then there exists a line bundle ℒY on Y such that ℒ=contF∗⁢ℒY.

Remark 2.4.14.

So far everything is a “formal” consequence of Kodaira vanishing and resolution of singularities (along with various classical arguments).

Remark 2.4.15.

Bhatt and Lurie proved a version of the Riemann-Hilbert correspondence in positive characteristic. Bhatt proved the Cohen-Macaulayness of the integral closure of an excellent Noetherian domain. Using these techniques and results, the Minimal Model Program has recently been generalized in two different directions:

  1. 1.

    In dimension 3 in mixed characteristic (essentially over Spec⁡ℤ) by Bhatt-Ma-Patakfalvi-Schwede-Tucker-Waldron-Witaszek.

  2. 2.

    In characteristic 0, most of the MMP works over an excellent ℚ-scheme by Murayama-Lyu.

Of course one may wonder why we need to run the MMP relative over a base. Consider a smooth projective family 𝒳→ℂ∗ such that K𝒳 is ample over ℂ∗. We want a good compactification, but naively compactifying gives us an arbitrary central fiber that may not even be normal. Using the resolution of singularities, we may consider a log resolution. This has many components and K𝒳 is not ample over ℂ in general. Then we can try to run the MMP over the base, and we expect a commutative diagram

such that K𝒳¯ is nef over the base and the singularities of 𝒳¯0 are slc. Here, slc means that the normalization is log canonical and has nodal singularities at codimension 1 points.

We may also use the relative MMP to study singularities. Here, let z∈Z and consider a log resolution X→𝜑Z. Then we write φ∗⁢(KZ)=KX+Δ. We may then perturb the coefficients of Δ:

  • •

    If c>1, we may decrease to c=1;

  • •

    If c<0, we can increase to c=ε>0;

This obtains a new boundary B with the same support as Δ. If we run the MMP for KX+B over Z, we obtain a partial resolution of singularities which has the singularities of the minimal model program. Therefore, by studying the exceptional divisors of the previous partial resolution and the singularities of the MMP, we can understand the singularities of Z∋z.

2.5 Running the Minimal Model Program

2.5.1 Flipping Contractions and Flips

Definition 2.5.1.

A map X→𝜑W is a flipping contraction if X is k⁢l⁢t and ℚ-factorial, ρ⁡(X/W)=1, and φ is a small birational contraction, and −KX is ample over W.

Remark 2.5.2.

When we work with log pairs we assume −(KX+Δ) is ample over W.

Remark 2.5.3.

W is never ℚ-factorial. Even worse, KW is not ℚ-Cartier.

Definition 2.5.4.

Let X→𝜑X be a flipping contraction. We say that X⁢⇢𝜋⁢X+ is a flip if it is a small birational map, KX++Δ+ is ℚ-Cartier, and there is a projective morphism φ+:X+→W such that KX++Δ+ is ample over W.

Lemma 2.5.5.

Let f:X⇢Y be a small birational map between normal varieties and let D be a Weil divisor on X. Then H0⁢(𝒪X⁢(D))≃H0⁢(𝒪Y⁢(f∗⁢D)).

Lemma 2.5.6.

Let X→𝜑W be a flipping contraction and let X⁢⇢𝜋⁢X+ be a flip. Then ρ⁡(X)=ρ⁡(X+) and X+ is ℚ-factorial.

Proof.

Let D+ be a divisor on X+ and D on X be the pushforward. We will find r such that R⋅(D+r⁡(KX+Δ))=0, where R is the extremal ray defining the flipping contraction. We know X is ℚ-factorial, so m⁡(D+r⁡(KX+Δ)) is Cartier for m≫0. Then we see that

for some Cartier divisor DW on W. Then

is Cartier, so D+ is ℚ-Cartier. For equality of ρ, we prove that π∗ induces an isomorphism between Weil divisors modulo linear equivalence. □

Lemma 2.5.7.

Let X→𝜑Y be a projective contraction between normal varieties with ρ⁡(X/Y)=1. Assume that the exceptional locus of φ contains a divisor. Then φ is the contraction of a unique irreducible divisor.

Proof.

Suppose there are two divisors E1,E2. We can find Ci covering Ei with Ci.Ei<0. Then we can find a such that E1+a⁢E2≡Y0. We will show that a is positive. Assume that C1 does not intersect E2. Then C1.(E1+a⁢E2)=C1.E1<0. Choosing a general C1 inside E1, we may assume E2.C1>0. Therefore C1.E1+a⁢E2.C1=0, so a=−C1.E1E2.C1>0. But E=E1+a⁢E2≕E is an effective divisor which is contracted, so it must be covered by E-negative curves, which gives a contradiction. We conclude that E1 must be the only component. □

Proposition 2.5.8.

Let φ:X→W be a flipping contraction for (X,Δ) klt. Then the flip exists if and only if

is a finitely-generated 𝒪Z-algebra. If this is the case, then

Proof.

Assume the flip

exists. We know that π is small, so

Moreover, KX++Δ+ is ample over W, so

Now assume that A≔⨁m≥0φ∗⁢𝒪X⁢(m⁡(KX+Δ)) is a finitely-generated 𝒪W-algebra and define X+=Proj⁡A. Then X⁢⇢𝜋 is an isomorphism in codimension 1. It could happen that the indeterminancy locus is a divisor. We know φ is an isomorphism over X∖Ex⁡(φ). Then A is just a sum of copies of the structure sheaf on X∖Ex⁡(φ). Therefore, X+⁢⇢π−1⁢X is an isomorphism over X∖Ex⁡(φ) and suppose D is mapped to a higher codimension cycle by φ+. Then

for some m>0. Because E is exceptional over W, we have

We have a natural inclusion φ∗+⁢𝒪X+⁢(t)⁢(E)↪𝒪W⁢(t⁢m⁢(KX+φ∗⁢Δ)), so there are no contracted divisors. Thus π is small. Then the property that ρ⁡(X/W)=ρ⁡(X+/W)=1 is by Lemma 2.5.6. □

2.5.2 Finite generation of the canonical ring

Conjecture 2.5.9.

Let X→Z be a projective morphism and (X,Δ) be a klt pair. Then

is a finitely-generated 𝒪Z-algebra.

Remark 2.5.10.

If X is a smooth projective variety, then ⨁m≥0H0⁢(X,𝒪X⁢(m⁢KX)) is finitely generated over the base field k.

2.5.3 Running the MMP

Now we may finally discuss the process of actually running the Minimal Model Program.

  1. 1.

    Let (Xi,Δi) be a klt pair and Xi→Z be a projective morphism. If KXi+Δi is nef over Z, then we stop and call this a minimal model over Z. If KXi+Δi is not nef over Z, we consider an extremal ray R in N⁢E¯⁢(Xi/Z) which is (KXi+Δi)-negative.

  2. 2.

    Let Xi→W be the contraction define by R.

    1. (a)

      We have dimW<dimXi, −KXi is ample over W, and the general fiber is klt. Hence the general fiber is klt Fano. In this case, we stop the algorithm and call this a Mori fiber space.

    2. (b)

      We have dimW=dimXi and Xi→W contains a divisor in its exceptional locus. This is a divisorial contraction. Then W is ℚ-factorial and ρ⁡(W/Z)=ρ⁡(X/Z)−1. Write Xi+1≔W and Δi+1≔f∗⁢(Δi). Return to step 1.

    3. (c)

      We have dimXi=dimW and Xi→W is a small birational map. We find the flip X⁢X+𝜋 and define Xi+1=X+ and Δi+1=π∗⁢Δi. We know that Xi+1 is ℚ-factorial provided that Xi is ℚ-factorial and ρ⁡(Xi/Z)=ρ⁡(Xi+1/Z). Return to step 1.

Remark 2.5.11.

Using the negativity lemma, we can prove that (Xi+1,Δi+1) is klt in step 2b.

The possible outcomes for the MMP are either a minimal model, which assuming abundance maps to the canonical model, or a Mori fiber space.

2.6 More about singularities

2.6.1 Surface singularities

Theorem 2.6.1.
  1. 1.

    A point x∈X is a surface klt singularity if and only if is the quotient of 0∈ℂ2 be a finite subgroup of G⁢L2⁢(ℂ).

  2. 2.

    A point x∈X is a canonical surface singularity if and only if it is the quotient of 0∈ℂ2 by a finite subgroup of S⁢L2⁢(ℂ).

  3. 3.

    A point x∈X is a terminal surface singularity if and only if it is a smooth point.

The idea of the proof is as follows: If KX is ℚ-Cartier, we can take its index one cover. This is a finite Galois morphism Y→𝜋X that is quasi-étale (étale in codimension 1) such that KY is a Cartier divisor. Then we can write X=Y/G. Because Y is klt and KY is Cartier, its log discrepancies are positive integers, so it is canonical. Therefore it suffices to study canonical singularities.

For canonical singularities, we obtain the following:

Theorem 2.6.2.

Let x∈X be a canonical surface singularity. Then x∈X has embedding dimension 3. Moreover, up to analytic change of coordinates, the following is a complete list of possible singularities:

Type A:

For n>0, the An singularity has equation x2+y2+zn+1=0 and dual graph

Type D:

For n≥4, the Dn singularity has equation x2+y2⁢z+zn−1=0 and dual graph

Type E:

The E6 singularity has equation x2+y3+z4=0 and dual graph

The E7 singularity has equation x3+y3+y⁢z3=0 and dual graph

The E8 singularity has equation x2+y3+z5 and dual graph

The idea of the proof is to study the dual graph of the resolution and use the Weierstrass preparation theorem to write down the equations.

2.6.2 Geography of singularities

The geography of singularity types with implications looks like this:

sing withgood dualizingcomplexsing of the MMP.sing related toderived AG.less rigid.symplecticquotientsorbifolds∥TerminalCanonicalkltdltlcrationalDu BoisCohen–Macaulay+ GorensteinKovács 1999Kollár–Kovács 2010+ GorensteinHypersurfaceComplete intersectionsGorenstein
Figure 2.2: Map of singularities

We will now define some of these singularity classes and prove some of the implications.

Definition 2.6.3.

Let (R,𝔪) be a Noetherian local ring and N be a finite R-module. Write dimN=dimsupp⁡N. Then N is called Cohen-Macaulay of one of the following equivalent conditions holds:

  1. 1.

    There exists x1,…,xr∈𝔪, where r=dimN such that xi is not a zero divisor in N/(x1,…,xi−1)⁢N for all i. Here, x1,…,xr is called a N-regular sequence.

  2. 2.

    If x1,…,xr∈𝔪 for r=dimN and dimN/(x1,…,xr)⁢M=0, then x1,…,xr is an N-regular sequence.

A coherent sheaf ℱ on a scheme X is Cohen-Macaulay if ℱx is Cohen-Macaulay over 𝒪X,x for every x∈X. Finally, a scheme X is Cohen-Macaulay if its structure sheaf 𝒪X is Cohen-Macaulay.

There is a condition of Serre, which is slightly weaker than being Cohen-Macaulay. We say that a ring R satisfies the properties

(Rk):

For all 𝔭∈Spec⁡A such that ht⁡(𝔭)≤k, then A𝔭 is regular.

(Sk):

For all 𝔭∈Spec⁡A, depth⁡(A𝔭)≥inf(k,ht⁡(𝔭)).

Examples 2.6.4.

Normality is equivalent to R1 and S2, so all normal surfaces are Cohen-Macaulay. If R is Cohen-Macaulay and a group G acts on R, then RG is Cohen-Macaulay by Hochster-Roberts. This means that quotient singularities are Cohen-Macaulay. The rings k⁡[x]/x2 and k⁡[t2,t3] are Cohen-Macaulay. However, the scheme X=Spec⁡k⁡[x,y]/(x2,x⁢y) is not Cohen-Macaulay at the origin.

Definition 2.6.5.

Let Y be a variety over a field of characteristic 0 and X→𝑓Y be a resolution of singularities. We say that f is rational if

  1. 1.

    f∗⁢𝒪X=𝒪Y (here, Y is normal).

  2. 2.

    The higher derived pushforwards are given by Ri⁢f∗⁢𝒪X=0 for i>0.

We say that Y has rational singularities if every resolution is rational.

Equivalently, the morphism 𝒪Y→R⁢f∗⁢𝒪X is a quasi-isomorphism in the derived category.

Examples 2.6.6.

All singularities of the form An are rational. If X→Y=An, then f∗⁢(KY)=KX, so Y is rational. However, the cone over an elliptic curve is not rational.

Definition 2.6.7.

A point y∈Y is symplectic if Y is normal at y and there exists a symplectic form ω on Yreg which extends to the exceptional divisor over y on any resolution.

Definition 2.6.8.

Let X⊆Y be an embedding of a scheme into a regular scheme and Z→Y be a log resolution of X which is an isomorphism outside X. Let E be the reduced preimage of X in Z. Then we say X has Du Bois singularities if 𝒪X→R⁢π∗⁢𝒪E is a quasi-isomorphism.

Remark 2.6.9.

Du Bois singularities commonly appear in Hodge theory.

Remarks 2.6.10.
  • •

    Serre duality for Cohen-Macaulay sheaves

  • •

    If H⊆X is Cartier and H is Cohen-Macaulay, then so is X.

Proposition 2.6.11.

Let Y be a variety over a field of characteristic 0 and f:X→Y be a resolution. The following are equivalent:

  1. 1.

    f is rational.

  2. 2.

    Y is Cohen-Macaulay and f∗⁢ωX=ωY.

Proof.

Let Y be projective and D be an ample Cartier divisor on Y. Then Hi⁢(X,ωX⁢(r⁢f∗⁢D))=0 for r>0,r>0. By Serre duality, we have Hn−i⁢(X,𝒪X⁢(−r⁢f∗⁢D))=0 for i>0 and r>0. Then the Leray spectral sequence

Assuming that f is rational, we then obtain

Now we will prove that Hi⁢(Y,𝒪Y⁢(−r⁢D))=0 for i<n and r≫0 only if Y is Cohen-Macaulay.2 Let H⊆|r′⁢D| be a general element. We obtain the exact sequence

By vanishing, we see that Hi⁢(H,𝒪H⁢(−r⁢D)) for i<n−1 and r≫0. Therefore by induction, H is Cohen-Macaulay and thus Y is Cohen-Macaulay. Taking the case of i=0, we obtain

for all r>0. This implies that f∗⁢ωX=ωY by a result in Hartshorne.

Now assume that Y is Cohen-Macaulay and f∗⁢ωX=ωY. We will induct on the dimension, and CM curves are smooth, so we proceed to the inductive step. We will prove that Ri⁢f∗⁢𝒪X is supported in some 0-dimensional locus. Let H⊆Y be general and H′=f−1⁢H, so f:H′→H is a resolution. Then

By induction, we see that 𝒪H⊗Ri⁢f∗⁢𝒪X=Ri⁢f∗⁢𝒪H′=0 is trivial outside a zero dimensional set. Therefore H0⁢(Y,Rq⁢f∗⁢𝒪X⁢(−r⁢D))=0 for p,q>0 or if p<n and q=0. By the spectral sequence, we obtain H0⁢(Y,Rq⁢f∗⁢𝒪X⊗𝒪Y𝒪Y⁢(−r⁢D))=0 for q<n−1 and

This implies that Rq⁢f∗⁢𝒪X=0 for q<n−1. When q=n, α is the dual to the morphism

because ωY≃f∗⁢ωX. Thus α is an isomorphism and thus Rn⁢f∗⁢𝒪X=0. □

Lemma 2.6.12.

Let (X,Δ) be a klt pair, H be basepoint free, and Hg∈|H| be a general element. Then (H,δH) is klt.

Lemma 2.6.13.

Let Y→𝑓X be finite with KY+ΔY=f∗⁢(KX+Δ). Then if both (X,Δ),(Y,ΔY) are log pairs, then (X,Δ) is klt (resp. log canonical) if and only if (Y,ΔY) is klt (resp. log canonical).

The idea of the proof is to use the Riemann-Hurwitz formula on a log resolution to compare the discrepancies and observe that aE⁢(X,Δ)=r⁢aE⁢(Y,ΔY) for some positive integer r (a ramification index).

Theorem 2.6.14 (Elkik 1981).

If (X,Δ) is dlt, then X has rational singularities.

Proof.

Write KY≡f∗⁢(KX+Δ)+A−B, where Y→X is a log resolution, Supp⁡B⊆Ex⁡(f), and ⌊A⌋=0. By KV vanishing, Rj⁢f∗⁢𝒪Y⁢(⌈B⌉)=0 for j>0. Now let ℒ be an ample Cartier divisor on X. WE have a commutative diagram

Therefore we have

and Hi⁢(𝒪Y⁢(−r⁢f∗⁢ℒ))=0 for i<n and r>0 by KV vanishing. We want to conclude that Hi⁢(𝒪X⁢(−r⁢ℒ))=0 for r>0,i<n. Using vanishing and the spectral sequence, we see that β is an isomorphism. But then by vanishing again, the top left corner is 0, and thus everything in the diagram must vanish. Therefore X is Cohen-Macaulay.

Now if we take j=0, we obtain an injection Hn⁢(𝒪X⁢(−r⁢ℒ))↪Hn⁢(𝒪Y⁢(−r⁢f∗⁢ℒ)). By Serre duality, the map H0⁢(Y,ωY⁢(r⁢f∗⁢ℒ))=H0⁢(X,f∗⁢ωY⊗𝒪X⁢(r⁢ℒ))↠H0⁢(X,ωX⊗𝒪X⁢(r⁢ℒ)) is a surjection, and surjectivity for r≫0 implies that f∗⁢ωY↠ωX, and they are isomorphic because they are both rank 1 reflexive sheaves. Therefore X has rational singularities. □

Proposition 2.6.15.

If X is Gorenstein (this means KX is Cartier), then rational is equivalent to canonical.

Proof.

We know canonical implies dlt implies rational. Now suppose that X is rational and Gorenstein and let Y→𝜋X be a resolution. Then π∗⁢(KX)=KY+E−F, where E,F are integral and effective with no common support. To prove that X is canonical, we need to prove that E=0. If we push-forward KY+E−F and E≠0, then we get associated primes on the image of E, and this contradicts π∗⁢ωY=ωX. Therefore E is trivial, so X is canonical. □

Proposition 2.6.16.

Symplectic singularities are rational and Gorenstein.

Proof.

Let π:Y→X be a resolution. If φ be a 2-form, then φr generates the line bundle ωXreg. The fact that π∗⁢φ extends to a regular holomorphic form implies that π∗⁢φr extends. Thus π∗⁢ωY=π∗⁢(π∗⁢φr)=φr=ωX. In addition, we see that π∗⁢ωX=ωY−F. □

We can characterize the condition of being Gorenstein as follows: A local ring (R,𝔪) is Gorenstein if and only if there exists a regular sequence x1,…,xr such that R/(x1,…,xr) is Gorenstein and 0-dimensional. By Nakayama’s lemma, this implies that R/(x1,…,xi)⁢R is Gorenstein for every i.

Lemma 2.6.17.

Let (0∈X) be an index 1 canonical threefold singularity and 0∈H⊆X be a general hyperplane section. Then either (0∈H) is a Du Val singularity or an elliptic singularity.

The idea of the proof is that if H′→𝜋H is a resolution, then π∗⁢ωH′=m⁢ωH. If m=1, we have a Du Val singularity, and if m>1, then we have an elliptic singularity.

Theorem 2.6.18.

All terminal threefold singularities of index 1 are cDV (compound Du Val), which means a one-parameter deformation of a Du Val singularity.

Example 2.6.19.

Let G⊆G⁢Ln⁢(ℂ) be a finite group. We know ℂn/G is a quotient singularity and thus klt. If G⊆S⁢Ln⁢(ℂ), then ℂn/G is symplectic.

Example 2.6.20.

If x∈X is a cone and is symplectic, then it is isomorphic to a Lie group quotient by the smallest nonzero nilpotent orbit.

2.7 Duality Theory

Given a projective variety X, we want a sheaf ωX such that Hi⁢(X,ℱ)=Hn−i⁢(X,ℋ⁢o⁢m⁢(ℱ,ωX))∨ for every Cohen-Macaulay sheaf. Recall that Serre duality establishes this duality for Hn⁢(ℙn,ℱ)=Hom⁡(ℱ,Kℙn) for any quasicoherent sheaf. Moreover, if ℱ is locally free, then

Definition 2.7.1.

A dualizing sheaf ωX is a coherent sheaf with a surjection

such that for all coherent ℱ, TrX induces an isomorphism

The pair ωX,TrX is unique if it exists.

Proposition 2.7.2.

Let f:X→Y be a finite morphism, ℱ∈Coh⁡(X),𝒢∈Coh⁡(Y). Let f!𝒢≔ℋom𝒪Y(f∗𝒪X,𝒢),

  1. 1.

    There exists an isomorphism f∗ℋom(ℱ,f!𝒢)=ℋom𝒪Y(f∗ℱ,𝒢). In particular, there is a natural isomorphism HomX(ℱ,f!𝒢)=HomY(f∗ℱ,𝒢).

Proof.

Assume X=Spec⁡A,Y=Spec⁡B and ℱ=M~,𝒢=N~. Then we have reduced to the identity

but this is obvious. Simply write ψ⁡(m)⁢(a)=ϕ⁡(a⁢m) for some ϕ∈HomB⁡(M,N). □

Proof.

Let f:X→Y be a finite morphism of proper schemes of pure dimension m. If ωY exists, then ωX exists and ωX=f!ωY. □

Proof.

We simply compute

Here, finiteness gives us Ri⁢f∗⁢ℱ=0 for i>0 and thus the Leray spectral sequence applies. □

Corollary 2.7.3.

The sheaf ωX exists and is an S2-sheaf for projective varieties X over k.

Proof.

We always have a finite surjective morphism X→ℙn (embed in large projective space, then project down). Then ωℙn exists, so ωX does also. Also ωX is S2 if and only if f∗⁢ωX is, but f∗⁢ωX=ℋ⁢o⁢m⁢(locally free). □

Corollary 2.7.4.

If X is a projective scheme of pure dimension m over k, ℱ a coherent sheaf such that supp⁡ℱ has pure dimension m, then

  1. 1.

    If ℱ is Cohen-Macaulat, so is ℋ⁢o⁢m⁢(ℱ,ωX). The converse holds if ℱ is S2.

  2. 2.

    If X is S2, then 𝒪X is Cohen-Macaulay if and only if ωX is.

Proof.

Let f:X→ℙm be surjective and finite. Then ℱ is Cohen-Macaulay if and only if f∗⁢ℱ is locally free, so ℋ⁢o⁢m𝒪X⁢(ℱ,ωX) is Cohen-Macaulay if and only if

and f∗⁢ℱ is locally free if and only if ℱ is S2. □

Theorem 2.7.5.

Let X be a projective scheme of pure dimension m, ℱ a Cohen-Macaulay sheaf such that supp⁡ℱ has pure dimension. Then

Proof.

As before, let f:X→ℙm be finite and surjective. We know that Hi⁢(X,ℱ)=Hi⁢(ℙm,f∗⁢ℱ). Then f∗⁢ℱ is locally free, so by Serre duality, we have

and therefore we obtain

Proposition 2.7.6.

Let X be a projective Cohen-Macaulay scheme of pure dimension m, D⊂X a Cartier divisor. Then ωD=ωX⁢(D)⊗𝒪D.

Proposition 2.7.7.

Let X be a normal projective variety of dimension m. Then ωX=𝒪X⁢(KX).

These two facts tive the classical adjunction formula.

Proposition 2.7.8.

Let X be smooth and D⊂X be a smooth Cartier divisor. Then (KX+D)|D=KD.

Next, we want to compare singularities of (X,B+S) with (S,B|S).

Remark 2.7.9.

If S is Cartier in codimension 2, then (KX+B+S)|S=KS+B|S. If there are quotient singularities in codimension 2, then (KX+B+S)|S=KS+B|S+∑Pi(1−1mi)⁢Pi, where Pi is a codimension 2 point of X, orbifold of order mi.

The idea of adjunction is that if (X,B+S) has nice singularities, then so does (S,B|S). The convsers is called the inverse of adjunction.

Proposition 2.7.10.

Let X be normal and S be a normal Weil divisor that is Cartier in codimension 2. Let Z⊆X be closed and B=∑bi⁢Bi≥0. Assume that KX+S+B is ℚ-Cartier. Then

Proof.

Let f:Y→X be a log-resolution of (X,S+B) and write S′=f∗−1⁢S and f∗−1⁢(S+B) is smooth. By doing more blowups, we may assume that if Ei ie exceptional such that Ei∩S′≠∅, then the center of Ei is contained in S. We have

and KS′=(KY+S′)|S′, and therefore

Now S′ is distinct from f∗−1⁢B, so if Ei∩S′≠∅, then Ei is f-exceptional and the center of Ei is contained in S. this shows that every log discrepancy that happens in S′→S also happens in Y→X in such a way that the center lies inside S. □

Therefore, if (X,B+S) is plt, then (S,B|S) is klt. The inverse of adjunction is more difficult. Suppose that (S,B|S) has nice singularities, but (X,B+S) has bad singularities. In particular, there exist exceptional E over X such that aE⁢(X,B+S)<0. This contradicts the philosophy that in a Fano object, the locus where it’s more singular is contracted. The theorem that realizes this is called Kollàr-Shokurov connectedness.

Definition 2.7.11.

Let (X,B) be a log-canonical. We say that Z⊆X is a log-canonical center if there exists E over X with aE⁢(X,B)=0 and cE⁢(X)=Z. We denote by l⁢c⁢c⁢(X,B) the union of all log-canonical centers. This is a closed subset.

Theorem 2.7.12 (Kollaàr-Shokurov connectedness).

Let f:X→Z be projective, (X,B) be log-canonical, and −(KX+B) be ample over Z. Then l⁢c⁢c⁢(X,B) is connected over Z.

Example 2.7.13.

Let ℙ1→pt be the structure morphism and note that deg⁡Kℙ1=−2. The coefficients of the dropped points must have sum stricly less than 2.

Remark 2.7.14.

Big and nef suffices in place of ample.

Theorem 2.7.15.

Let g:Y→X be a proper birational mapping with Y smooth and X normal. Let D=∑di⁢Di be a smooth normal crossings ℚ-divisor on Y such that g∗⁢D≥0 and −(KY+D) is g-nef. Write

Then supp⁡F=supp⁡⌊F⌋ is connected in any neighborhood of any fiber of g.

Remark 2.7.16.

A field K is C1 if every homogeneous f⁡(x0,…,xn)∈K⁡[x0,…,xn] of degree at most n has a nontrivial zero. A field is pseudo-algebraically closed if every geometrically integral k-variety has a k-point.

Conjecture 2.7.17 (Ax).

Every pseudo-algebraically closed field of characteristic 0 is C1.

This is true and proved by Kollàr.

Proof of Theorem.

Note that ⌈−A⌉−⌊F⌋=KY−g∗⁢(KX+D)+{A}+{F}. By KV vanishing, we have R1⁢g∗⁢𝒪Y⁢(⌈−A⌉−⌊F⌋)=0. Then we have the exact sequence

so g∗⁢𝒪Y⁢(⌈−A⌉)↠g∗⁢𝒪⌊F⌋⁢(⌈−A⌉). Now ⌈−A⌉ is g-exceptional, so g∗⁢𝒪Y⁢(⌈−A⌉)=𝒪X⁢g∗⁢D≥0. Writing ⌊F⌋=F1∪F2 is a neighborhood of h−1⁢(x), we obtain

This will give a surjection 𝒪X,x→𝒪⌊F⌋,x, which is not true by connectedness. □

Theorem 2.7.18.

Let X be normal and S⊆X be a normal Weil divisor Cartier in codimension 2. Let B≥0 be a ℚ-divisor such that KX+S+B is ℚ-Cartier. Then

  1. 1.

    (X,S+B) is plt near S if and only if (S,B|S) is klt.

  2. 2.

    Assume B is ℚ-Cartier and S it klt. Then (X,S+B) is log canonical near S if and only if (S,B|S) is log canonical.

Proof.

Take g:Y→X a log resolution of (X,S+B) and KY+D=g∗⁢(KX+B+S). Write S′=g∗−1⁢S and F=S′+F′. Then F′ contains a component of codimensino at least 1. By adjunction, we have

so (K,B+S) is plt near S if and only if F′∩g−1⁢(S)=∅ and (S,B|S) is klt if and only if (F′)≥0∩S′=∅. By Kollàr-Shokurov connectedness, given x∈S, there exists Vx⊆X such that (S′∪(F′)≥1)∩g−1⁢(Vx) is connected. Here F′∩g−1⁢(Vx)=∅. Moving x around S finishes the argument. □

We will now consider adjunction for higher codimension centers. We know (X,S+B) is snc if and only if (X,B|S) is. For (X,B), let Z be a log-canonical center on (X,B). If Z has codimension 1 in X, do adjunction to Z and repeat. If Z has codimension at least 2, define (Z,BZ) such that the singularities of (X,B) around Z can be compared to those of (Z,BZ).

This is indeed possible. First blow up enough so that g maps E to Z. Write g∗⁢(KX+B)=KY+BY+E. Then we have

and this is trivial over Z. We want to write KE+BE=g∗⁢(KZ+BZ+MZ), where BZ measures the singularities of the fiber MZ, the variation of moduli.

2.8 Terminal threefold MMP

Let X be a ℚ-factorial terminal projective threefold and X→𝑓W be a flipping contraction. Suppose ρ⁡(X/W)=1, −KX is ample over W, and X has terminal singularities and is smooth in codimension 2. First, we want to construct a rational map π which is an isomorphism in codimension 1 such that ρ⁡(X+/W)=1, KX+ is ample over W, and X+ has terminal singularities. Also, X+ is unique and coincides with

provided that this ring is a finitely generated 𝒪W-algebra. Denote this ring by R⁡(X).

Proposition 2.8.1.

The ring R⁡(X) is finitely generated as an 𝒪W-algebra if and only if it is finitely generated locally over W (and this is even true locally analytically).

For a geometric picture, let X be a terminal threefold and suppose f contracts several curves. Then W is not ℚ-factorial in general. In 1988, Mori proved that these curves can be contracted one-by-one in the analytic category (in the algebraic category, this is clearly impossible by the previous discussion). Now our terminal threefold flipping contractions lead us to studying extremal neighborhoods. Here, we consider an analytic neighborhood of the fiber above w∈W. We may assume that (W,w) is a rational singularity. Now what happens if our extremal neighborhood is smooth?

Proposition 2.8.2.

Let X⊇C=ℙ1 be an extremal neighborhood. Then 𝒪C⁢(KX)=𝒪C⁢(−1), ℐC/ℐC2≃𝒪C⊕𝒪C⁢(1), and |−KX| has a smooth member.

Proof.

We know that KX.C=−1 because KX.C<0 and H1⁢(𝒪C⁢(KX))=0. Now we have an exact sequence

and thus we have an isomorphism ⋀2(ℐC/ℐC2)≃𝒪⁡(KX)⊗𝒪C⁢(−K). Taking the degree, we conclude that degℐC/ℐC2=(−KX.C)−degKC=1. Now if we tensor with 𝒪C⁢(−1), we obtain the exact sequence

Because H1⁢(ℐC/ℐC2⊗𝒪C⁢(−1))=0, we have ℐC/ℐC2⊗𝒪C⁢(−1)≃𝒪C⁢(−1)⊕𝒪C.

Now let x∈C and (D,x) be a smooth divisor on the germ (X,x). Then D extends naturally to a divisor D′ of X. We need to show that D′∈|−KX|. Here, we know that Pic⁡X=ℤ and the isomorphism is induced by ℒ↦ℒ.C. Thus D′∼−KX. □

Corollary 2.8.3.

In the above case, X is the blowup of a smooth threefold (W,w) along a smooth curve C0∋w.

Corollary 2.8.4.

Let X→W be a terminal threefold flipping contraction. For every w∈W, f−1⁢(W) contains a singularity.

By the classification of terminal threefold singularities, for any extremal neighborhood, there are at most three singular points and there is a nice divisor D passing through some of them.

Theorem 2.8.5 (Mori 1988).

Let X⊇C=ℙ1 be an extremal neighborhood. Then one of the following holds for |−a⁢KX| (a=1,2):

  1. 1.

    Either |−KX| has a member D with Du Val singularities, or

  2. 2.

    |−2⁢KX| has a member D such that the double cover Z of X branched over X with only Du Val singularities.

Because D is Du Val, we know (X,D) is purely log terminal. Now the finite generation of the ring R⁡(X) is implied by the finite generation of the ring

This is now a problem about projective surfaces, and apparently this is doable by hand. Here, D is a log-Calabi Yau surface with at most three singular points. Now we conclude the existence of flips for terminal threefolds.

Now we need to prove termination of flips.

Theorem 2.8.6.

An arbitrary sequence of three-dimensional extremal canonical (KX+Δ)-flips is finite.

Lemma 2.8.7.

Let ϕ:X⇢X′ be a (KX+Δ)-flip of a three=dimensional canonical pair (X,Δ), where Δ=∑ai⁢Di. Let C′⊆X′ be a flipped curve and EC′ be the exceptional divisor obtained by blowing up C′. Then X′ is smooth along C′ and

where the multiplicities are of D′ along C′.

Proof.

Because C′ is a flipped curve, X′ is smooth along the generic point of C′. Indeed, X′ is terminal along ηC′, so it is smooth. If there is a nonterminal valuation with center on C′, then there is a noncanonical valuation on (X,Δ). □

We will now define the difficulty function. Let (X,Δ=∑ai⁢Di) be a canonical pair with Di pairwise distinct prime divisors. Set a=max⁡{ai} and S≔∑ai⁢ℤ≥0⊆ℚ. Also write

For example, with compound Du Val singularities, we have d⁡(c⁢D⁢v)=r.

Remark 2.8.8.

The difficulty function measures the number of nonterminal valuations.

We are now ready to prove termination of canonical threefold flips.

Proof of termination of flips.

Write Δ=∑ai⁢Di, where a1≤⋯≤ak. Then in the sequence of flips, we have

If k=0, then d⁡(Xk−1,0)>d⁡(Xj,0). Therefore, after finitely many flips, d⁡(Xℓ,0)=0 and thus there are no more flips.

Now assume that k>0. Then we know d⁡(Xj,Δj) is nondecreasing. If Cj is the flipped curve for ϕj−1, assume it is contained in Dkj. Then ak<1 and d⁡(Xj−1,Δj−1)>d⁡(Xj,Δj). Therefore for j≫0, Dkj contains no flipped curves.

Now denote by D¯kj the normalization of Dkj. Then D¯kj−1→D¯kj is a birational morphism. The exceptional curves of D¯kj→D¯kℓ for ℓ>k are linearly independent. At some point, we have D¯kj≃D¯kℓ for ℓ≫j. This means that both the flipping and flipped curves are disjoint from Dkj. Thus C.Dkj=0, so now flips for (X,Δ=∑i=1kai⁢Di) induce flips on (X,Δ′=∑i=1k−1ai⁢Dk), and by induction, these flips stop. □

Conjecture 2.8.9 (Abundance).

If X is klt and KX is nef, then KX is semiample.

This is proved for terminal threefolds by Kawamata. Now we have settled the minimal model program for terminal threefolds. Existence and termination of flips were proved in 1988 by Mori and later in a more general case by Kollár-Shokurov, and abundance was proved by Kawamata in the 1990s. A historical sketch of more recent developments is as follows:

Table 2.1: Historical Sketch since 2000
Existence of Flips Termination of Flips
In the 2000s Kollár-Shokurov proved existence for (X,Δ) log-canonical threefolds In 2004, Alexeev-Hacon-Kawamata proved termination for (X,Δ) fourfolds with −(KX+Δ) effective.
In 2005, Hacon and McKernan proved flips exist in dimension n provided the MMP works in dimension n−1. In 2000, Fujino proved termination of flips for (X,Δ) a terminal fourfold.
In 2006, Birkar-Cascini-Hacon-McKernan proved existence of flips for (X,Δ) klt. They also proved termination of flips for (X,Δ) klt and KX+Δ big.
In 2010, Hacon-Xu and Birkar proved exietence of flips for (X,Δ) log canonical. In 2018, Joaquín proved termination of flips for (X,Δ) a log canonical fourfold with KX+Δ pseudoeffective.

Termination is unknown for klt uniruled fourfolds and in dimension at least 5. Abundance is known for log canonical pairs in dimension 3 and log canonical pairs in dimension 4 with X uniruled. It is unknown for log canonical pairs with KX+Δ pseudoeffective. Some people believe that abundance should be tackled using different techniques from the minimal model program.

2.9 Applications to Singularities

Conjecturally, the MMP contracts or flips the locus Bs⁡(KX)=Bs−⁢(KX). This is known in dimension 3 and follows from termination and abundance. Recall that if D⊆X and A is an ample divisor on X, then

This is a countable union of algebraic varieties. Then (someone) showed the existence of a divisor on a certain blowup of ℙ3 whose Bs− is a countable union of curves. Here, the dynamics of what happens lead to something non-algebraic.

The first application we will give is terminalization. Let (X,Δ) be a klt pair. Then there exists a projective birational morphism Y→X such that Y is terminal and extracts exactly the divisors with aE⁢(X,Δ)∈(−1,0].

To prove this, let (Z,ΔZ)→𝜑(X,Δ) be a log resolution of X. Then we know φ∗⁢(KX+Δ)=KZ+ΔZ. This is not necessarily a boundary, so write ΔZ′ for the result of killing all negative coefficients. By negativity, we see that Bs−⁢(KZ+ΔZ′) contains all divisors with aE⁢(X,Δ)>0. Now we run the minimal model program for KZ+ΔZ′, and this contracts all of these divisors. Because (Z,ΔZ′) is terminal, it remains terminal after we run the minimal model program.