**Przeglądaj wersję html pliku:**

### Algebraic Geometry (ang)

Algebraic Geometry

Andreas Gathmann

Notes for a class taught at the University of Kaiserslautern 2002/2003

C ONTENTS 0. Introduction 0.1. What is algebraic geometry? 0.2. Exercises 1. Afﬁne varieties 1.1. Algebraic sets and the Zariski topology 1.2. Hilbert’s Nullstellensatz 1.3. Irreducibility and dimension 1.4. Exercises 2. Functions, morphisms, and varieties 2.1. Functions on afﬁne varieties 2.2. Sheaves 2.3. Morphisms between afﬁne varieties 2.4. Prevarieties 2.5. Varieties 2.6. Exercises 3. Projective varieties 3.1. Projective spaces and projective varieties 3.2. Cones and the projective Nullstellensatz 3.3. Projective varieties as ringed spaces 3.4. The main theorem on projective varieties 3.5. Exercises 4. Dimension 4.1. The dimension of projective varieties 4.2. The dimension of varieties 4.3. Blowing up 4.4. Smooth varieties 4.5. The 27 lines on a smooth cubic surface 4.6. Exercises 5. Schemes 5.1. Afﬁne schemes 5.2. Morphisms and locally ringed spaces 5.3. Schemes and prevarieties 5.4. Fiber products 5.5. Projective schemes 5.6. Exercises 6. First applications of scheme theory 6.1. Hilbert polynomials 6.2. B´ zout’s theorem e 6.3. Divisors on curves 6.4. The group structure on a plane cubic curve 6.5. Plane cubic curves as complex tori 6.6. Where to go from here 6.7. Exercises 1 1 6 8 8 11 13 16 18 18 21 23 27 31 32 35 35 39 40 44 47 50 50 54 57 63 67 71 74 74 78 80 82 86 89 92 92 96 101 104 108 112 117

7. More about sheaves 7.1. Sheaves and sheaﬁﬁcation 7.2. Quasi-coherent sheaves 7.3. Locally free sheaves 7.4. Differentials 7.5. Line bundles on curves 7.6. The Riemann-Hurwitz formula 7.7. The Riemann-Roch theorem 7.8. Exercises 8. Cohomology of sheaves 8.1. Motivation and deﬁnitions 8.2. The long exact cohomology sequence 8.3. The Riemann-Roch theorem revisited 8.4. The cohomology of line bundles on projective spaces 8.5. Proof of the independence of the afﬁne cover 8.6. Exercises 9. Intersection theory 9.1. Chow groups 9.2. Proper push-forward of cycles 9.3. Weil and Cartier divisors 9.4. Intersections with Cartier divisors 9.5. Exercises 10. Chern classes 10.1. Projective bundles 10.2. Segre and Chern classes of vector bundles 10.3. Properties of Chern classes 10.4. Statement of the Hirzebruch-Riemann-Roch theorem 10.5. Proof of the Hirzebruch-Riemann-Roch theorem 10.6. Exercises References Index

120 120 127 131 133 137 141 144 147 149 149 152 155 159 162 163 165 165 171 176 181 185 188 188 191 194 200 203 209 211 212

0.

Introduction

1

0. I NTRODUCTION

In a very rough sketch we explain what algebraic geometry is about and what it can be used for. We stress the many correlations with other ﬁelds of research, such as complex analysis, topology, differential geometry, singularity theory, computer algebra, commutative algebra, number theory, enumerative geometry, and even theoretical physics. The goal of this section is just motivational; you will not ﬁnd deﬁnitions or proofs here (and probably not even a mathematically precise statement).

0.1. What is algebraic geometry? To start from something that you probably know, we can say that algebraic geometry is the combination of linear algebra and algebra: • In linear algebra, we study systems of linear equations in several variables. • In algebra, we study (among other things) polynomial equations in one variable. Algebraic geometry combines these two ﬁelds of mathematics by studying systems of polynomial equations in several variables. Given such a system of polynomial equations, what sort of questions can we ask? Note that we cannot expect in general to write down explicitly all the solutions: we know from algebra that even a single complex polynomial equation of degree d > 4 in one variable can in general not be solved exactly. So we are more interested in statements about the geometric structure of the set of solutions. For example, in the case of a complex polynomial equation of degree d, even if we cannot compute the solutions we know that there are exactly d of them (if we count them with the correct multiplicities). Let us now see what sort of “geometric structure” we can ﬁnd in polynomial equations in several variables. Example 0.1.1. Probably the easiest example that is covered neither in linear algebra nor in algebra is that of a single polynomial equation in two variables. Let us consider the following example: Cn = {(x, y) ∈ C2 ; y2 = (x − 1)(x − 2) · · · (x − 2n)} ⊂ C2 , where n ≥ 1. Note that in this case it is actually possible to write down all the solutions, because the equation is (almost) solved for y already: we can pick x to be any complex number, and then get two values for y — unless x ∈ {1, . . . , 2n}, in which case we only get one value for y (namely 0). So it seems that the set of equations looks like two copies of the complex plane with the two copies of each point 1, . . . , 2n identiﬁed: the complex plane parametrizes the values for x, and the two copies of it correspond to the two possible values for y, i.e. the two roots of the number (x − 1) · · · (x − 2n). This is not quite true however, because a complex non-zero number does not have a distinguished ﬁrst and second root that could correspond to the ﬁrst and second copy of the complex plane. Rather, the two roots of a complex number get exchanged if you run around the origin once: if we consider a path x = r eiϕ for 0 ≤ ϕ ≤ 2π and ﬁxed r > 0

around the complex origin, the square root of this number would have to be deﬁned by √ iϕ √ x = re 2 which gives opposite values at ϕ = 0 and ϕ = 2π. In other words, if in Cn we run around one of the points 1, . . . , 2n, we go from one copy of the plane to the other. The way to draw this topologically is to cut the two planes along the lines [1, 2], . . . , [2n − 1, 2n], and to glue the two planes along these lines as in this picture (lines marked with the same letter are to be identiﬁed):

2

Andreas Gathmann

C

1

A B

2

3

C D

4

5

E F

6

glue

B A

1

2

3

D C

4

5

F E

C

6

To make the picture a little nicer, we can compactify our set by adding two points at inﬁnity, in the same way as we go from C to C∞ by adding a point ∞. If we do this here, we end up with a compact surface with n − 1 handles:

add points at infinity

Such an object is called a surface of genus n − 1; the example above shows a surface of genus 2. Example 0.1.2. Example 0.1.1 is a little “cheated” because we said before that we want to ﬁgure out the geometric structure of equations that we cannot solve explicitly. In the example however, the polynomial equation was chosen so that we could solve it, and in fact we used this solution to construct the geometric picture. Let us see now what we can still do if we make the polynomial more complicated. What happens if we consider Cn = {(x, y) ∈ C2 ; y2 = f (x)} ⊂ C2 , with f some polynomial in x of degree 2n? Obviously, as long as the 2n roots of f are still distinct, the topological picture does not change. But if two of the roots approach each other and ﬁnally coincide, this has the effect of shrinking one of the tubes connecting the two planes until it ﬁnally reduces to a “singular point” (also called a node), as in the following picture on the left:

glue

=

0.

Introduction

3

Obviously, we can view this as a surface with one handle less, where in addition we identify two of the points (as illustrated in the picture on the right). Note that we can still see the “handles” when we draw the surface like this, just that one of the handles results from the glueing of the two points. Example 0.1.3. You have probably noticed that the polynomial equation of example 0.1.2 could be solved directly too. Let us now consider Cd = {(x, y) ∈ C2 ; f (x, y) = 0} ⊂ C2 , where f is an arbitrary polynomial of degree d. This is an equation that we certainly cannot solve directly if f is sufﬁciently general. Can we still deduce the geometric structure of C? In fact, we can do this with the idea of example 0.1.2. We saw there that the genus of the surface does not change if we perturb the polynomial equation, even if the surface acquires singular points (provided that we know how to compute the genus of such a singular surface). So why not deform the polynomial f to something singular that is easier to analyze? Probably the easiest thing that comes into mind is to degenerate the polynomial f of degree d into a product of d linear equations 1 , . . . , d : Cd = {(x, y) ∈ C2 ;

1 (x, y) · · · d (x, y) = 0}

⊂ C2 ,

This surface should have the same “genus” as the original Cd . It is easy to see what Cd looks like: of course it is just a union of d lines. Any two of them intersect in a point, and we can certainly choose the lines so that no three of them intersect in a point. The picture below shows Cd for d = 3 (note that every line is — after compactifying — just the complex sphere C∞ ).

What is the genus of this surface? In the picture above it is obvious that we have one loop; so if d = 3 we get a surface of genus 1. What is the general formula? We have d spheres, and every two of them connect in a pair of points, so in total we have d connections. But 2 d − 1 of them are needed to glue the d spheres to a connected chain without loops; only the remaining ones then add a handle each. So the genus of Cd (and hence of Cd ) is d d −1 − (d − 1) = . 2 2 This is commonly called the degree-genus formula for plane curves. Remark 0.1.4. One of the trivial but common sources for misunderstandings is whether we count dimensions over C or over R. The examples considered above are real surfaces (the dimension over R is 2), but complex curves (the dimension over C is 1). We have used the word “surface” as this ﬁtted best to the pictures that we have drawn. When looking at the theory however, it is usually best to call these objects curves. In what follows, we always mean the dimension over C unless stated otherwise. Remark 0.1.5. What we should learn from the examples above: • Algebraic geometry can make statements about the topological structure of objects deﬁned by polynomial equations. It is therefore related to topology and differential geometry (where similar statements are deduced using analytic methods).

4

Andreas Gathmann

• The geometric objects considered in algebraic geometry need not be smooth (i.e. they need not be manifolds). Even if our primary interest is in smooth objects, degenerations to singular objects can greatly simplify a problem (as in example 0.1.3). This is a main point that distinguishes algebraic geometry from other “geometric” theories (e.g. differential or symplectic geometry). Of course, this comes at a price: our theory must be strong enough to include such singular objects and make statements how things vary when we degenerate from smooth to singular objects. In this regard, algebraic geometry is related to singularity theory which studies precisely these questions. Remark 0.1.6. Maybe it looks a bit restrictive to allow only algebraic (polynomial) equations to describe our geometric objects. But in fact it is a deep theorem that for compact objects, we would not get anything different if we allowed holomorphic equations too. In this respect, algebraic geometry is very much related (and in certain cases identical) to complex (analytic) geometry. The easiest example of this correspondence is that a holomorphic map from the Riemann sphere C∞ to itself must in fact be a rational map (i.e. the quotient of two polynomials). Example 0.1.7. Let us now turn our attention to the next more complicated objects, namely complex surfaces in 3-space. We just want to give one example here. Let S be the cubic surface S = {(x, y, z) ; 1 + x3 + y3 + z3 − (1 + x + y + z)3 = 0} ⊂ C3 . As this object has real dimension 4, it is impossible to draw pictures of it that reﬂect its topological properties correctly. Usually, we overcome this problem by just drawing the real part, i.e. we look for solutions of the equation over the real numbers. This then gives a real surface in R3 that we can draw. We should just be careful about which statements we can claim to “see” from this incomplete geometric picture. The following picture shows the real part of the surface S:

In contrast to our previous examples, we have now used a linear projection to map the real 3-dimensional space onto the drawing plane. We see that there are some lines contained in S. In fact, one can show that every smooth cubic surface has exactly 27 lines on it (see section 4.5 for details). This is another sort of question that one can ask about the solutions of polynomial equations, and that is not of topological nature: do they contain curves with special properties (in this case lines), and if so, how many? This branch of algebraic geometry is usually called enumerative geometry. Remark 0.1.8. It is probably surprising that algebraic geometry, in particular enumerative geometry, is very much related to theoretical physics. In fact, many results in enumerative geometry have been found by physicists ﬁrst. Why are physicists interested e.g. in the number of lines on the cubic surface? We try to give a short answer to this (that is necessarily vague and incomplete): There is a branch of theoretical physics called string theory whose underlying idea is that the elementary

0.

Introduction

5

particles (electrons, quarks,. . . ) might not be point-like, but rather one-dimensional objects (the so-called strings), that are just so small that their one-dimensional structure cannot be observed directly by any sort of physical measurement. When these particles move in time, they sweep out a surface in space-time. For some reason this surface has a natural complex structure coming from the underlying physical theory. Now the same idea applies to space-time in general: string theorists believe that spacetime is not 4-dimensional as we observe it, but rather has some extra dimensions that are again so small in size that we cannot observe them directly. (Think e.g. of a long tube with a very small diameter — of course this is a two-dimensional object, but if you look at this tube from very far away you cannot see the small diameter any more, and the object looks like a one-dimensional line.) These extra dimensions are parametrized by a space that sometimes has a complex structure too; it might for example be the complex cubic surface that we looked at above. So in this case we’re in fact looking at complex curves in a complex surface. A priori, these curves can sit in the surface in any way. But there are equations of motion that tell you how these curves will sit in the ambient space, just as in classical mechanics it follows from the equations of motion that a particle will move on a straight line if no forces apply to it. In our case, the equations of motion say that the curve must map holomorphically to the ambient space. As we said in remark 0.1.6 above, this is equivalent to saying that we must have algebraic equations that describe the curve. So we are looking at exactly the same type of questions as we did in example 0.1.7 above. Example 0.1.9. Let us now have a brief look at curves in 3-dimensional space. Consider the example C = {(x, y, z) = (t 3 ,t 4 ,t 5 ) ; t ∈ C} ⊂ C3 . We have given this curve parametrically, but it is in fact easy to see that we can give it equally well in terms of polynomial equations: C = {(x, y, z) ; x3 = yz, y2 = xz, z2 = x2 y}. What is striking here is that we have three equations, although we would expect that a one-dimensional object in three-dimensional space should be given by two equations. But in fact, if you leave out any of the above three equations, you’re changing the set that it describes: if you leave out e.g. the last equation z2 = x2 y, you would get the whole z-axis {x = y = 0} as additional points that do satisfy the ﬁrst two equations, but not the last one. So we see another important difference to linear algebra: it is not true that every object of codimension d can be given by d equations. Even worse, if you are given d equations, it is in general a very difﬁcult task to ﬁgure out what dimension their solution has. There do exist algorithms to ﬁnd this out for any given set of polynomials, but they are so complicated that you will in general want to use a computer program to do that for you. This is a simple example of an application of computer algebra to algebraic geometry. Remark 0.1.10. Especially the previous example 0.1.9 is already very algebraic in nature: the question that we asked there does not depend at all on the ground ﬁeld being the complex numbers. In fact, this is a general philosophy: even if algebraic geometry describes geometric objects (when viewed over the complex numbers), most methods do not rely on this, and therefore should be established in purely algebraic terms. For example, the genus of a curve (that we introduced topologically in example 0.1.1) can be deﬁned in purely algebraic terms in such a way that all the statements from complex geometry (e.g. the degree-genus formula of example 0.1.3) extend to this more general setting. Many geometric questions then reduce to pure commutative algebra, which is in some sense the foundation of algebraic geometry.

6

Andreas Gathmann

Example 0.1.11. The most famous application of algebraic geometry to ground ﬁelds other than the complex numbers is certainly Fermat’s Last Theorem: this is just the statement that the algebraic curve over the rational numbers C = {(x, y) ∈ Q2 ; xn + yn = 1} ⊂ Q2 contains only the trivial points where x = 0 or y = 0. Note that this is very different from the case of the ground ﬁeld C, where we have seen in example 0.1.3 that C is a curve of genus n−1 2 . But a lot of the theory of algebraic geometry applies to the rational numbers (and related ﬁelds) as well, so if you look at the proof of Fermat’s theorem (which you most probably will not understand) you will notice that it uses e.g. the concepts of algebraic curves and their genus all over the place, although the corresponding point set C contains only some trivial points. So, in some sense, we can view (algebraic) number theory as a part of algebraic geometry. Remark 0.1.12. With this many relations to other ﬁelds of mathematics (and physics), it is obvious that we have to restrict our attention in this class to quite a small subset of the possible applications. Although we will develop the general theory of algebraic geometry, our focus will mainly be on geometric questions, neglecting number-theoretic aspects most of the time. So, for example, if we say “let k be an algebraically closed ﬁeld”, feel free to read this as “let k be the complex numbers” and think about geometry rather than algebra. Every now and then we will quote results from or give applications to other ﬁelds of mathematics. This applies in particular to commutative algebra, which provides some of the basic foundations of algebraic geometry. So unless you want to take commutative algebra as a black box that spits out a useful theorem from time to time (which is possible but not recommended), you should get some background in commutative algebra while learning algebraic geometry. Some knowledge about geometric objects occurring in other ﬁelds of mathematics (manifolds, projective spaces, differential forms, vector bundles, . . . ) is helpful but not necessary. We will develop these concepts along the way as we need them. 0.2. Exercises. Note: As we have not developed any theory yet, you are not expected to be able to solve the following problems in a mathematically precise way. Rather, they are just meant as some “food for thought” if you want to think a little further about the examples considered in this section. Exercise 0.2.1. What do we get in example 0.1.1 if we consider the equation Cn = {(x, y) ∈ C2 ; y2 = (x − 1)(x − 2) · · · (x − (2n − 1))} ⊂ C2 instead? Exercise 0.2.2. (For those who know something about projective geometry:) In example 0.1.3, we argued that a polynomial of degree d in two complex variables gives rise to a surface of genus d−1 . In example 0.1.1 however, a polynomial of degree 2n gave us a 2 surface of genus n − 1. Isn’t that a contradiction? Exercise 0.2.3. (i) Show that the space of lines in Cn has dimension 2n − 2. (Hint: use that there is a unique line through any two given points in Cn .) (ii) Let S ⊂ C3 be a cubic surface, i.e. the zero locus of a polynomial of degree 3 in the three coordinates of C3 . Find an argument why you would expect there to be ﬁnitely many lines in S (i.e. why you would expect the dimension of the space of lines in S to be 0-dimensional). What would you expect if the equation of S has degree less than or greater than 3?

0.

Introduction

7

Exercise 0.2.4. Let S be the speciﬁc cubic surface S = {(x, y, z) ; x3 + y3 + z3 = (x + y + z)3 } ⊂ C3 . (i) Show that there are exactly 3 lines contained in S. (ii) Using the description of the space of lines of exercise 0.2.3, try to ﬁnd an argument why these 3 lines should be counted with multiplicity 9 each (in the same way as e.g. double roots of a polynomial should be counted with multiplicity 2). We can then say that there are 27 lines on S, counted with their correct multiplicities. (Remark: It is actually possible to prove that the number of lines on a cubic surface does not depend on the speciﬁc equation of the surface. This then shows, together with this exercise, that every cubic surface has 27 lines on it. You need quite a lot of theoretical background however to make this into a rigorous proof.) Exercise 0.2.5. Show that if you replace the three equations deﬁning the curve C in example 0.1.9 by (i) x3 = y2 , x5 = z2 , y5 = z4 , or (ii) x3 = y2 , x5 = z2 , y5 = z3 + ε for small but non-zero ε, the resulting set of solutions is in fact 0-dimensional, as you would expect it from three equations in three-dimensional space. So we see that very small changes in the equations can make a very big difference in the result. In other words, we usually cannot apply numerical methods to our problems, as very small rounding errors can change the result completely. Exercise 0.2.6. Let X be the set of all complex 2 × 3 matrices of rank at most 1, viewed as a subset of the C6 of all 2 × 3 matrices. Show that X has dimension 4, but that you need 3 equations to deﬁne X in the ambient 6-dimensional space C6 .

8

Andreas Gathmann

1. A FFINE VARIETIES

A subset of afﬁne n-space An over a ﬁeld k is called an algebraic set if it can be written as the zero locus of a set of polynomials. By the Hilbert basis theorem, this set of polynomials can be assumed to be ﬁnite. We deﬁne the Zariski topology on An (and hence on any subset of An ) by declaring the algebraic sets to be the closed sets. Any algebraic set X ⊂ An has an associated radical ideal I(X) ⊂ k[x1 , . . . , xn ] that consists of those functions that vanish on X. Conversely, for any radical ideal I there is an associated algebraic set Z(I) which is the common zero locus of all functions in I. If k is algebraically closed, Hilbert’s Nullstellensatz states that this gives in fact a one-to-one correspondence between algebraic sets in An and radical ideals in k[x1 , . . . , xn ]. An algebraic set (or more generally any topological space) is called irreducible if it cannot be written as a union of two proper closed subsets. Irreducible algebraic sets in An are called afﬁne varieties. Any algebraic set in An can be decomposed uniquely into a ﬁnite union of afﬁne varieties. Under the correspondence mentioned above, afﬁne varieties correspond to prime ideals. The dimension of an algebraic set (or more generally of a topological space) is deﬁned to be the length of the longest chain of irreducible closed subsets minus one.

1.1. Algebraic sets and the Zariski topology. We have said in the introduction that we want to consider solutions of polynomial equations in several variables. So let us now make the obvious deﬁnitions. Deﬁnition 1.1.1. Let k be a ﬁeld (recall that you may think of the complex numbers if you wish). We deﬁne afﬁne n -space over k, denoted An , to be the set of all n-tuples of elements of k: An := {(a1 , . . . , an ) ; ai ∈ k for 1 ≤ i ≤ n}. The elements of the polynomial ring k[x1 , . . . , xn ] :={polynomials in the variables x1 , . . . , xn over k} ={∑ aI xI ; aI ∈ k}

I

(with the sum taken over all multi-indices I = (i1 , . . . , in ) with i j ≥ 0 for all 1 ≤ j ≤ n) deﬁne functions on An in the obvious way. For a given set S ⊂ k[x1 , . . . , xn ] of polynomials, we call Z(S) := {P ∈ An ; f (P) = 0 for all f ∈ S} ⊂ An the zero set of S. Subsets of An that are of this form for some S are called algebraic sets. By abuse of notation, we also write Z( f1 , . . . , fi ) for Z(S) if S = { f1 , . . . , fi }. Example 1.1.2. Here are some simple examples of algebraic sets: (i) (ii) (iii) (iv) (v) Afﬁne n-space itself is an algebraic set: An = Z(0). / The empty set is an algebraic set: 0 = Z(1). Any single point in An is an algebraic set: (a1 , . . . , an ) = Z(x1 − a1 , . . . , xn − an ). Linear subspaces of An are algebraic sets. All the examples from section 0 are algebraic sets: e.g. the curves of examples 0.1.1 and 0.1.3, and the cubic surface of example 0.1.7.

Remark 1.1.3. Of course, different subsets of k[x1 , . . . , xn ] can give rise to the same algebraic set. Two trivial cases are: (i) If two polynomials f and g are already in S, then we can also throw in f + g without changing Z(S). (ii) If f is in S, and g is any polynomial, then we can also throw in f · g without changing Z(S).

1.

Afﬁne varieties

9

Recall that a subset S of a commutative ring R (in our case, R = k[x1 , . . . , xn ]) is called an ideal if it is closed both under addition and under multiplication with arbitrary ring elements. If S ⊂ R is any subset, the set (S) = { f1 g1 + · · · + fm gm ; fi ∈ S, gi ∈ R} is called the ideal generated by S; it is obviously an ideal. So what we have just said amounts to stating that Z(S) = Z((S)). It is therefore sufﬁcient to only look at the cases where S is an ideal of k[x1 , . . . , xn ]. There is a more serious issue though that we will deal with in section 1.2: a function f 2 has the same zero set as any of its powers f i ; so e.g. Z(x1 ) = Z(x1 ) (although the ideals 2 ) are different). (x1 ) and (x1 We will now address the question whether any algebraic set can be deﬁned by a ﬁnite number of polynomials. Although this is entirely a question of commutative algebra about the polynomial ring R = k[x1 , . . . , xn ], we will recall here the corresponding deﬁnition and proposition. Lemma and Deﬁnition 1.1.4. Let R be a ring. The following two conditions are equivalent: (i) Every ideal in R can be generated by ﬁnitely many elements. (ii) R satisﬁes the ascending chain condition: every (inﬁnite) ascending chain of ideals I1 ⊂ I2 ⊂ I3 ⊂ · · · is stationary, i.e. we must have Im = Im+1 = Im+2 = · · · for some m. If R satisﬁes these conditions, it is called Noetherian. Proof. (i) ⇒ (ii): Let I1 ⊂ I2 ⊂ · · · be an inﬁnite ascending chain of ideals in R. Then I := ∪i Ii is an ideal of R as well; so by assumption (i) it can be generated by ﬁnitely many elements. These elements must already be contained in one of the Im , which means that Im = Im + 1 = · · · . (ii) ⇒ (i): Assume that there is an ideal I that cannot be generated by ﬁnitely many elements. Then we can recursively construct elements fi in I by picking f1 ∈ I arbitrary and fi+1 ∈ I\( f1 , . . . , fi ). It follows that the sequence of ideals ( f1 ) ⊂ ( f1 , f2 ) ⊂ ( f1 , f2 , f3 ) ⊂ · · · is not stationary. Proposition 1.1.5. (Hilbert basis theorem) If R is a Noetherian ring then so is R[x]. In particular, k[x1 , . . . , xn ] is Noetherian; so every algebraic set can be deﬁned by ﬁnitely many polynomials. Proof. Assume that I ⊂ R[x] is an ideal that is not ﬁnitely generated. Then we can deﬁne a sequence of elements fi ∈ I as follows: let f0 be a non-zero element of I of minimal degree, and let fi+1 be an element of I of minimal degree in I\( f0 , . . . , fi ). Obviously, deg fi ≤ deg fi+1 for all i by construction. For all i let ai ∈ R be the leading coefﬁcient of fi , and let Ii = (a0 , . . . , ai ) ⊂ R. As R is Noetherian, the chain of ideals I0 ⊂ I1 ⊂ · · · in R is stationary. Hence there is an m such that am+1 ∈ (a0 , . . . , am ). Let r0 , . . . , rm ∈ R such that am+1 = ∑m ri ai , and consider the i=0 polynomial f = fm+1 − ∑ xdeg fm+1 −deg fi ri fi .

i=0 m

We must have f ∈ I\( f0 , . . . , fm ), as otherwise the above equation would imply that fm+1 ∈ ( f0 , . . . , fm ). But by construction the coefﬁcient of f of degree deg fm+1 is zero, so deg f < deg fm+1 , contradicting the choice of fm+1 . Hence R[x] is Noetherian.

10

Andreas Gathmann

In particular, as k is trivially Noetherian, it follows by induction that k[x1 , . . . , xn ] is. We will now return to the study of algebraic sets and make them into topological spaces. Lemma 1.1.6. (i) If S1 ⊂ S2 ⊂ k[x1 , . . . , xn ] then Z(S2 ) ⊂ Z(S1 ) ⊂ An . T S (ii) If {Si } is a family of subsets of k[x1 , . . . , xn ] then i Z(Si ) = Z( i Si ) ⊂ An . n. (iii) If S1 , S2 ⊂ k[x1 , . . . , xn ] then Z(S1 ) ∪ Z(S2 ) = Z(S1 S2 ) ⊂ A In particular, arbitrary intersections and ﬁnite unions of algebraic sets are again algebraic sets. Proof. (i) and (ii) are obvious, so let us prove (iii). “⊂”: If P ∈ Z(S1 ) ∪ Z(S2 ) then P ∈ Z(S1 ) or P ∈ Z(S2 ). In particular, for any f1 ∈ S1 , f2 ∈ S2 , we have f1 (P) = 0 or f2 (P) = 0, so f1 f2 (P) = 0. “⊃”: If P ∈ Z(S1 ) ∪ Z(S2 ) then P ∈ Z(S1 ) and P ∈ Z(S2 ). So there / / / are functions f1 ∈ S1 and f2 ∈ S2 that do not vanish at P. Hence f1 f2 (P) = 0, so P ∈ / Z(S1 S2 ). Remark 1.1.7. Recall that a topology on any set X can be deﬁned by specifying which subsets of X are to be considered closed sets, provided that the following conditions hold: / (i) The empty set 0 and the whole space X are closed. (ii) Arbitrary intersections of closed sets are closed. (iii) Finite unions of closed sets are closed. Note that the standard deﬁnition of closed subsets of Rn that you know from real analysis satisﬁes these conditions. A subset Y of X is then called open if its complement X\Y is closed. If X is a topological space and Y ⊂ X any subset, Y inherits an induced subspace topology by declaring the sets of the form Y ∩ Z to be closed whenever Z is closed in X. A map f : X → Y is called continuous if inverse images of closed subsets are closed. (For the standard topology of Rn from real analysis and the standard deﬁnition of continuous functions, it is a theorem that a function is continuous if and only if inverse images of closed subsets are closed.) Deﬁnition 1.1.8. We deﬁne the Zariski topology on An to be the topology whose closed sets are the algebraic sets (lemma 1.1.6 tells us that this gives in fact a topology). Moreover, any subset X of An will be equipped with the topology induced by the Zariski topology on An . This will be called the Zariski topology on X. Remark 1.1.9. In particular, using the induced subspace topology, this deﬁnes the Zariski topology on any algebraic set X ⊂ An : the closed subsets of X are just the algebraic sets Y ⊂ An contained in X. The Zariski topology is the standard topology in algebraic geometry. So whenever we use topological concepts in what follows we refer to this topology (unless we specify otherwise). Remark 1.1.10. The Zariski topology is quite different from the usual ones. For example, on An , a closed subset that is not equal to An satisﬁes at least one non-trivial polynomial equation and has therefore necessarily dimension less than n. So the closed subsets in the Zariski topology are in a sense “very small”. It follows from this that any two nonempty open subsets of An have a non-empty intersection, which is also unfamiliar from the standard topology of real analysis. Example 1.1.11. Here is another example that shows that the Zariski topology is “unusual”. The closed subsets of A1 besides the whole space and the empty set are exactly the ﬁnite sets. In particular, if f : A1 → A1 is any bijection, then f is a homeomorphism. (This

1.

Afﬁne varieties

11

last statement is essentially useless however, as we will not deﬁne morphisms between algebraic sets as just being continuous maps with respect to the Zariski topology. In fact, this example gives us a strong hint that we should not do so.) 1.2. Hilbert’s Nullstellensatz. We now want to establish the precise connection between algebraic sets in An and ideals in k[x1 , . . . , xn ], hence between geometry and algebra. We have already introduced the operation Z(·) that takes an ideal (or any subset of k[x1 , . . . , xn ]) to an algebraic set. Here is an operation that does the opposite job. Deﬁnition 1.2.1. For a subset X ⊂ An , we call I(X) := { f ∈ k[x1 , . . . , xn ] ; f (P) = 0 for all P ∈ X} ⊂ k[x1 , . . . , xn ] the ideal of X (note that this is in fact an ideal). Remark 1.2.2. We have thus deﬁned a two-way correspondence algebraic sets in An −→ ←−

Z I

ideals in k[x1 , . . . , xn ]

.

We will now study to what extent these two maps are inverses of each other. Remark 1.2.3. Let us start with the easiest case of algebraic sets and look at points in An . Points are minimal algebraic sets, so by lemma 1.1.6 (i) they should correspond to maximal ideals. In fact, the point (a1 , . . . , an ) ∈ An is the zero locus of the ideal I = (x1 −a1 , . . . , xn − an ). Recall from commutative algebra that an ideal I of a ring R is maximal if and only if R/I is a ﬁeld. So in our case I is indeed maximal, as k[x1 , . . . , xn ]/I ∼ k. However, = for general k there are also maximal ideals that are not of this form, e.g. (x2 + 1) ⊂ R[x] (where R[x]/(x2 + 1) ∼ C). The following proposition shows that this cannot happen if k = is algebraically closed, i.e. if every non-constant polynomial in k[x] has a zero. Proposition 1.2.4. (Hilbert’s Nullstellensatz (“theorem of the zeros”)) Assume that k is algebraically closed (e.g. k = C). Then the maximal ideals of k[x1 , . . . , xn ] are exactly the ideals of the form (x1 − a1 , . . . , xn − an ) for some ai ∈ k. Proof. Again this is entirely a statement of commutative algebra, so you can just take it on faith if you wish (in fact, many textbooks on algebraic geometry do so). For the sake of completeness we will give a short proof here in the case k = C that uses only some basic algebra; but feel free to ignore it if it uses concepts that you do not know. A proof of the general case can be found e.g. in [Ha] proposition 5.18. So assume that k = C. From the discussion above we see that it only remains to show that any maximal ideal m is contained in an ideal of the form (x1 − a1 , . . . , xn − an ). As C[x1 , . . . , xn ] is Noetherian, we can write m = ( f1 , . . . , fr ) for some fi ∈ C[x1 , . . . , xn ]. Let K be the subﬁeld of C obtained by adjoining to Q all coefﬁcients of the fi . We will now restrict coefﬁcients to this subﬁeld K, so let m0 = m ∩ K[x1 , . . . , xn ]. Note that then m = m0 · C[x1 , . . . , xn ], as the generators fi of m lie in m0 . Note that m0 ⊂ K[x1 , . . . , xn ] is a maximal ideal too, because if we had an inclusion m0 m0 K[x1 , . . . , xn ] of ideals, this would give us an inclusion m m C[x1 , . . . , xn ] by taking the product with C[x1 , . . . , xn ]. (This last inclusion has to be strict as intersecting it with K[x1 , . . . , xn ] gives the old ideals m0 m0 back again.) So K[x1 , . . . , xn ]/m0 is a ﬁeld. We claim that there is an embedding K[x1 , . . . , xn ]/m0 → C. To see this, split the ﬁeld extension K[x1 , . . . , xn ]/m0 : Q into a purely transcendental part L : Q and an algebraic part K[x1 , . . . , xn ]/m0 : L. As K[x1 , . . . , xn ]/m0 and hence L is ﬁnitely generated over Q whereas C is of inﬁnite transcendence degree over Q, there is an embedding L ⊂ C. Finally, as K[x1 , . . . , xn ]/m0 : L is algebraic and C algebraically closed, this embedding can be extended to give an embedding K[x1 , . . . , xn ]/m0 ⊂ C.

12

Andreas Gathmann

Let ai be the images of the xi under this embedding. Then fi (a1 , . . . , an ) = 0 for all i by construction, so fi ∈ (x1 − a1 , . . . , xn − an ) and hence m ⊂ (x1 − a1 , . . . , xn − an ). Remark 1.2.5. The same method of proof can be used for any algebraically closed ﬁeld k that has inﬁnite transcendence degree over the prime ﬁeld Q or F p . Corollary 1.2.6. Assume that k is algebraically closed. (i) There is a 1:1 correspondence {points in An } ←→ {maximal ideals of k[x1 , . . . , xn ]} given by (a1 , . . . , an ) ←→ (x1 − a1 , . . . , xn − an ). (ii) Every ideal I k[x1 , . . . , xn ] has a zero in An . Proof. (i) is obvious from the Nullstellensatz, and (ii) follows in conjunction with lemma 1.1.6 (i) as every ideal is contained in a maximal one. Example 1.2.7. We just found a correspondence between points of An and certain ideals of the polynomial ring. Now let us try to extend this correspondence to more complicated algebraic sets than just points. We start with the case of a collection of points in A1 . (i) Let X = {a1 , . . . , ar } ⊂ A1 be a ﬁnite algebraic set. Obviously, I(X) is then generated by the function (x − a1 ) · · · (x − ar ), and Z(I(X)) = X again. So Z is an inverse of I. (ii) Conversely, let I ⊂ k[x] be an ideal (not equal to (0) or (1)). As k[x] is a principal ideal domain, we have I = ( f ) for some non-constant monic function f ∈ k[x]. Now for the correspondence to work at all, we have to require that k be alge/ braically closed: for if f had no zeros, we would have Z(I) = 0, and I(Z(I)) = (1) would give us back no information about I at all. But if k is algebraically closed, we can write f = (x − a1 )m1 · · · (x − ar )mr with the ai distinct and mi > 0. Then Z(I) = {a1 , . . . , ar } and therefore I(Z(I)) is generated by (x − a1 ) · · · (x − ar ), i.e. all exponents are reduced to 1. Another way to express this fact is that a function √ is in I(Z(I)) if and only if some power of it lies in I. We write this as I(Z(I)) = I, where we use the following deﬁnition. Deﬁnition 1.2.8. For an ideal I ⊂ k[x1 , . . . , xn ], we deﬁne the radical of I to be √ I := { f ∈ k[x1 , . . . , xn ] ; f r ∈ I for some r > 0}. √ (In fact, this is easily seen to be an ideal.) An ideal I is called radical if I = I. Note that the ideal of an algebraic set is always radical. The following proposition says that essentially the same happens for n > 1. As it can be guessed from the example above, the case Z(I(X)) is more or less trivial, whereas the case I(Z(I)) is more difﬁcult and needs the assumption that k be algebraically closed. Proposition 1.2.9. (i) (ii) (iii) If X1 ⊂ X2 are subsets of An then I(X2 ) ⊂ I(X1 ). For any algebraic set X ⊂ An we have Z(I(X)) = X. If √k is algebraically closed, then for any ideal I ⊂ k[x1 , . . . , xn ] we have I(Z(I)) = I.

Proof. (i) is obvious, as well as the “⊃” parts of (ii) and (iii). (ii) “⊂”: By deﬁnition X = Z(I) for some I. Hence, by (iii) “⊃” we have I ⊂ I(Z(I)) = I(X). By 1.1.6 (i) it then follows that Z(I(X)) ⊂ Z(I) = X.

1.

Afﬁne varieties

13

(iii) “⊂”: (This is sometimes also called Hilbert’s Nullstellensatz, as it follows easily from proposition 1.2.4.) Let f ∈ I(Z(I)). Consider the ideal J = I + ( f t − 1) ⊂ k[x1 , . . . , xn ,t]. This has empty zero locus in An+1 , as f vanishes on Z(I), so if we require f t = 1 at the same time, we get no solutions. Hence J = (1) by corollary 1.2.6 (i). In particular, there is a relation 1 = ( f t − 1)g0 + ∑ fi gi ∈ k[x1 , . . . , xn ,t] for some gi ∈ k[x1 , . . . , xn ,t] and fi ∈ I. If t N is the highest power of t occurring in the gi , then after multiplying with f N we can write this as f N = ( f t − 1)G0 (x1 , . . . , xn , f t) + ∑ fi Gi (x1 , . . . , xn , f t) where Gi = f N gi is considered to be a polynomial in x1 , . . . , xn , f t. Modulo f t − 1 we get f N = ∑ fi Gi (x1 , . . . , xn , 1) ∈ k[x1 , . . . , xn , f t]/( f t − 1). But as the map k[x1 , . . . , xn ] → k[x1 , . . . , xn , f t]/( f t − 1) is injective, this equality holds in fact in k[x1 , . . . , xn ], so f N ∈ I. Corollary 1.2.10. If k is algebraically closed, there is a one-to-one inclusion-reversing correspondence between algebraic sets in An and radical ideals in k[x1 , . . . , xn ], given by the operations Z(·) and I(·). (This is also sometimes called the Nullstellensatz.) Proof. Immediately from proposition 1.2.9 and lemma 1.1.6 (i). From now on up to the end of section 4, we will always assume that the ground ﬁeld k is algebraically closed. √ Remark 1.2.11. Even though √ radical I of an ideal I was easy to deﬁne, it is quite the difﬁcult to actually compute I for any given ideal I. Even worse, it is already quite difﬁcult just to check whether I itself is radical or not. In general, you will need non-trivial methods of computer algebra to solve problems like this. 1.3. Irreducibility and dimension. The algebraic set X = {x1 x2 = 0} ⊂ A2 can be written as the union of the two coordinate axes X1 = {x1 = 0} and X2 = {x2 = 0}, which are themselves algebraic sets. However, X1 and X2 cannot be decomposed further into ﬁnite unions of smaller algebraic sets. We now want to generalize this idea. It turns out that this can be done completely in the language of topological spaces. This has the advantage that it applies to more general cases, i.e. open subsets of algebraic sets. However, you will want to think only of the Zariski topology here, since the concept of irreducibility as introduced below does not make much sense in classical topologies. Deﬁnition 1.3.1. (i) A topological space X is said to be reducible if it can be written as a union X = X1 ∪ X2 , where X1 and X2 are (non-empty) closed subsets of X not equal to X. It is called irreducible otherwise. An irreducible algebraic set in An is called an afﬁne variety. (ii) A topological space X is called disconnected if it can be written as a disjoint union X = X1 ∪ X2 of (non-empty) closed subsets of X not equal to X. It is called connected otherwise. Remark 1.3.2. Although we have given this deﬁnition for arbitrary topological spaces, you will usually want to apply the notion of irreducibility only in the Zariski topology. For example, in the usual complex topology, the afﬁne line A1 (i.e. the complex plane) is reducible because it can be written e.g. as the union of closed subsets A1 = {z ∈ C ; |z| ≤ 1} ∪ {z ∈ C ; |z| ≥ 1}.

14

Andreas Gathmann

In the Zariski topology however, A1 is irreducible (as it should be). In contrast, the notion of connectedness can be used in the “usual” topology too and does mean there what you think it should mean. Remark 1.3.3. Note that there is a slight inconsistency in the existing literature: some authors call a variety what we call an algebraic set, and consequently an irreducible variety what we call an afﬁne variety. The algebraic characterization of afﬁne varieties is the following. Lemma 1.3.4. An algebraic set X ⊂ An is an afﬁne variety if and only if its ideal I(X) ⊂ k[x1 , . . . , xn ] is a prime ideal. Proof. “⇐”: Let I(X) be a prime ideal, and suppose that X = X1 ∪ X2 . Then I(X) = I(X1 ) ∩ I(X2 ) by exercise 1.4.1 (i). As I(X) is prime, we may assume I(X) = I(X1 ), so X = X1 by proposition 1.2.9 (ii). “⇒”: Let X be irreducible, and let f g ∈ I(X). Then X ⊂ Z( f g) = Z( f ) ∪ Z(g), hence X = (Z( f ) ∩ X) ∪ (Z(g) ∩ X) is a union of two algebraic sets. As X is irreducible, we may assume that X = Z( f ) ∩ X, so f ∈ I(X). Example 1.3.5. (i) An is an afﬁne variety, as I(An ) = (0) is prime. If f ∈ k[x1 , . . . , xn ] is an irreducible polynomial, then Z( f ) is an afﬁne variety. A collection of m points in An is irreducible if and only if m = 1. (ii) Every afﬁne variety is connected. The union of the n coordinate axes in An is always connected, although it is reducible for n > 1. A collection of m points in An is connected if and only if m = 1. As it can be expected, any topological space that satisﬁes a reasonable ﬁniteness condition can be decomposed uniquely into ﬁnitely many irreducible spaces. This is what we want to show next. Deﬁnition 1.3.6. A topological space X is called Noetherian if every descending chain X ⊃ X1 ⊃ X2 ⊃ · · · of closed subsets of X is stationary. Remark 1.3.7. By corollary 1.2.10 the fact that k[x1 , . . . , xn ] is a Noetherian ring (see proposition 1.1.5) translates into the statement that any algebraic set is a Noetherian topological space. Proposition 1.3.8. Every Noetherian topological space X can be written as a ﬁnite union X = X1 ∪ · · · ∪ Xr of irreducible closed subsets. If one assumes that Xi ⊂ X j for all i = j, then the Xi are unique (up to permutation). They are called the irreducible components of X. In particular, any algebraic set is a ﬁnite union of afﬁne varieties in a unique way. Proof. To prove existence, assume that there is a topological space X for which the statement is false. In particular, X is reducible, hence X = X1 ∪ X1 . Moreover, the statement of the proposition must be false for at least one of these two subsets, say X1 . Continuing this construction, one arrives at an inﬁnite chain X X1 X2 · · · of closed subsets, which is a contradiction as X is Noetherian. To show uniqueness, assume that we have two decompositions X = X1 ∪ · · · ∪ Xr = S S X1 ∪ · · · ∪ Xs . Then X1 ⊂ i Xi , so X1 = (X1 ∩ Xi ). But X1 is irreducible, so we can assume X1 = X1 ∩ X1 , i.e. X1 ⊂ X1 . For the same reason, we must have X1 ⊂ Xi for some i. So X1 ⊂ X1 ⊂ Xi , which means by assumption that i = 1. Hence X1 = X1 is contained in both decompositions. Now let Y = X\X1 . Then Y = X2 ∪ · · · ∪ Xr = X2 ∪ · · · ∪ Xs ; so proceeding by induction on r we obtain the uniqueness of the decomposition.

1.

Afﬁne varieties

15

Remark 1.3.9. It is probably time again for a warning: given an ideal I of the polynomial ring, it is in general not easy to ﬁnd the irreducible components of Z(I), or even to determine whether Z(I) is irreducible or not. There are algorithms to ﬁgure this out, but they are computationally quite involved, so you will in most cases want to use a computer program for the actual calculation. Remark 1.3.10. In the same way one can show that every algebraic set X is a (disjoint) ﬁnite union of connected algebraic sets, called the connected components of X. Remark 1.3.11. We have now seen a few examples of the correspondence between geometry and algebra that forms the base of algebraic geometry: points in afﬁne space correspond to maximal ideals in a polynomial ring, afﬁne varieties to prime ideals, algebraic sets to radical ideals. Most concepts in algebraic geometry can be formulated and most proofs can be given both in geometric and in algebraic language. For example, the geometric statement that we have just shown that any algebraic set can be written as a ﬁnite union of irreducible components has the equivalent algebraic formulation that every radical ideal can be written uniquely as a ﬁnite intersection of prime ideals. Remark 1.3.12. An application of the notion of irreducibility is the deﬁnition of the dimension of an afﬁne variety (or more generally of a topological space; but as in the case of irreducibility above you will only want to apply it to the Zariski topology). Of course, in the case of complex varieties we have a geometric idea what the dimension of an afﬁne variety should be: it is the number of complex coordinates that you need to describe X locally around any point. Although there are algebraic deﬁnitions of dimension that mimics this intuitive one, we will give a different deﬁnition here that uses only the language of topological spaces. Finally, all these deﬁnitions are of course equivalent and describe the intuitive notion of dimension (at least over C), but it is actually quite hard to prove this rigorously. The idea to deﬁne the dimension in algebraic geometry using the Zariski topology is the following: if X is an irreducible topological space, then any closed subset of X not equal to X must have dimension (at least) one smaller. (This is of course an idea that is not valid in the usual topology that you know from real analysis.) Deﬁnition 1.3.13. Let X be a (non-empty) irreducible topological space. The dimension / of X is the biggest integer n such that there is a chain 0 = X0 X1 · · · Xn = X of irreducible closed subsets of X. If X is any Noetherian topological space, the dimension of X is deﬁned to be the supremum of the dimensions of its irreducible components. A space of dimension 1 is called a curve, a space of dimension 2 a surface. Remark 1.3.14. In this deﬁnition you should think of Xi as having dimension i. The content of the deﬁnition is just that there is “nothing between” varieties of dimension i and i + 1. Example 1.3.15. The dimension of A1 is 1, as single points are the only irreducible closed subsets of A1 not equal to A1 . We will see in exercise 1.4.9 that the dimension of A2 is 2. Of course, the dimension of An is always n, but this is a fact from commutative algebra that we cannot prove at the moment. But we can at least see that the dimension of An is not less than n, because there are sequences of inclusions A0 A1 ··· An of linear subspaces of increasing dimension. Remark 1.3.16. This deﬁnition of dimension has the advantage of being short and intuitive, but it has the disadvantage that it is very difﬁcult to apply in actual computations. So for the moment we will continue to use the concept of dimension only in the informal way as we have used it so far. We will study the dimension of varieties rigorously in section 4, after we have developed more powerful techniques in algebraic geometry.

16

Andreas Gathmann

Remark 1.3.17. Here is another application of the notion of irreducibility (that is in fact not much more than a reformulation of the deﬁnition). Let X be an irreducible topological space (e.g. an afﬁne variety). Let U ⊂ X be a non-empty open subset, and let Y X be a closed subset. The fact that X cannot be the union (X\U) ∪ Y can be reformulated by saying that U cannot be a subset of Y . In other words, the closure of U (i.e. the smallest closed subset of X that contains U) is equal to X itself. Recall that an open subset of a topological space X is called dense if its closure is equal to the whole space X. With this wording, we have just shown that in an irreducible topological space every non-empty open subset is dense. Note that this is not true for reducible spaces: let X = {x1 x2 = 0} ⊂ A2 be the union of the two coordinate axes, and let U = {x1 = 0} ∩ X be the open subset of X consisting of the x1 -axis minus the origin. Then the closure of U in X is just the x1 -axis, and not all of X. 1.4. Exercises. In all exercises, the ground ﬁeld k is assumed to be algebraically closed unless stated otherwise. Exercise 1.4.1. Let X1 , X2 ⊂ An be algebraic sets. Show that (i) I(X1 ∪ X2 ) = I(X1 ) ∩ I(X2 ), (ii) I(X1 ∩ X2 ) = I(X1 ) + I(X2 ). Show by example that taking the radical in (ii) is in general necessary, i.e. ﬁnd algebraic sets X1 , X2 such that I(X1 ∩ X2 ) = I(X1 ) + I(X2 ). Can you see geometrically what it means if we have inequality here? Exercise 1.4.2. Let X ⊂ A3 be the union of the three coordinate axes. Determine generators for the ideal I(X). Show that I(X) cannot be generated by fewer than 3 elements, although X has codimension 2 in A3 . Exercise 1.4.3. In afﬁne 4-dimensional space A4 with coordinates x, y, z,t let X be the union of the two planes X = {x = y = 0} and X = {z = x − t = 0}.

Compute the ideal I = I(X) ⊂ k[x, y, z,t]. For any a ∈ k let Ia ⊂ k[x, y, z] be the ideal obtained by substituting t = a in I, and let Xa = Z(Ia ) ⊂ A3 . Show that the family of algebraic sets Xa with a ∈ k describes two skew lines in A3 approaching each other, until they ﬁnally intersect transversely for a = 0. Moreover, show that the ideals Ia are radical for a = 0, but that I0 is not. Find the √ elements in I0 \I0 and interpret them geometrically.

2 Exercise 1.4.4. Let X ⊂ A3 be the algebraic set given by the equations x1 − x2 x3 = x1 x3 − x1 = 0. Find the irreducible components of X. What are their prime ideals? (Don’t let the simplicity of this exercise fool you. As mentioned in remark 1.3.9, it is in general very difﬁcult to compute the irreducible components of the zero locus of given equations, or even to determine if it is irreducible or not.)

Exercise 1.4.5. Let A3 be the 3-dimensional afﬁne space over a ﬁeld k with coordinates x, y, z. Find ideals describing the following algebraic sets and determine the minimal number of generators for these ideals. (i) The union of the (x, y)-plane with the z-axis. (ii) The union of the 3 coordinate axes. (iii) The image of the map A1 → A3 given by t → (t 3 ,t 4 ,t 5 ). Exercise 1.4.6. Let Y be a subspace of a topological space X. Show that Y is irreducible if and only if the closure of Y in X is irreducible.

1.

Afﬁne varieties

17

Exercise 1.4.7. (For those of you who like pathological examples. You will need some knowledge on general topological spaces.) Find a Noetherian topological space with inﬁnite dimension. Can you ﬁnd an afﬁne variety with inﬁnite dimension? Exercise 1.4.8. Let X = {(t,t 3 ,t 5 ) ; t ∈ k} ⊂ A3 . Show that X is an afﬁne variety of dimension 1 and compute I(X). Exercise 1.4.9. Let X ⊂ A2 be an irreducible algebraic set. Show that either • X = Z(0), i.e. X is the whole space A2 , or • X = Z( f ) for some irreducible polynomial f ∈ k[x, y], or • X = Z(x − a, y − b) for some a, b ∈ k, i.e. X is a single point. Deduce that dim(A2 ) = 2. (Hint: Show that the common zero locus of two polynomials f , g ∈ k[x, y] without common factor is ﬁnite.)

18

Andreas Gathmann

2. F UNCTIONS ,

MORPHISMS , AND VARIETIES

If X ⊂ An is an afﬁne variety, we deﬁne the function ﬁeld K(X) of X to be the quotient ﬁeld of the coordinate ring A(X) = k[x1 , . . . , xn ]/I(X); this can be thought of as the ﬁeld of rational functions on X. For a point P ∈ X the local ring OX,P is the subring of K(X) of all functions that are regular (i.e. well-deﬁned) at P, and for U ⊂ X an open subset we let OX (U) be the subring of K(X) of all functions that are regular at every P ∈ U. The ring of functions that are regular on all of X is precisely A(X). Given two ringed spaces (X, OX ), (Y, OY ) with the property that their structure sheaves are sheaves of k-valued functions, a set-theoretic map f : X → Y determines a pull-back map f ∗ from k-valued functions on Y to k-valued functions on X by composition. We say that f is a morphism if f is continuous and f ∗ OY (U) ⊂ OX ( f −1 (U)) for all open sets U in Y . In particular, this deﬁnes morphisms between afﬁne varieties and their open subsets. Morphisms X → Y between afﬁne varieties correspond exactly to k-algebra homomorphisms A(Y ) → A(X). In complete analogy to the theory of manifolds, we then deﬁne a prevariety to be a ringed space (whose structure sheaf is a sheaf of k-valued functions and) that is locally isomorphic to an afﬁne variety. Correspondingly, there is a general way to construct prevarieties and morphisms between them by taking afﬁne varieties (resp. morphisms between them) and patching them together. Afﬁne varieties and their open subsets are simple examples of prevarieties, but we also get more complicated spaces as e.g. P1 and the afﬁne line with a doubled origin. A prevariety X is called a variety if the diagonal ∆(X) ⊂ X × X is closed, i.e. if X does not contain “doubled points”.

2.1. Functions on afﬁne varieties. After having deﬁned afﬁne varieties, our next goal must of course be to say what the maps between them should be. Let us ﬁrst look at the easiest case: “regular functions”, i.e. maps to the ground ﬁeld k = A1 . They should be thought of as the analogue of continuous functions in topology, or differentiable functions in real analysis, or holomorphic functions in complex analysis. Of course, in the case of algebraic geometry we want to have algebraic functions, i.e. (quotients of) polynomial functions. Deﬁnition 2.1.1. Let X ⊂ An be an afﬁne variety. We call A(X) := k[x1 , . . . , xn ]/I(X) the coordinate ring of X. Remark 2.1.2. The coordinate ring of X should be thought of as the ring of polynomial functions on X. In fact, for any P ∈ X an element f ∈ A(X) determines a polynomial map X → k (usually also denoted by f ) given by f → f (P): • this is well-deﬁned, because all functions in I(X) vanish on X by deﬁnition, • if the function f : X → k is identically zero then f ∈ I(X) by deﬁnition, so f = 0 in A(X). Note that I(X) is a prime ideal by lemma 1.3.4, so A(X) is an integral domain. Hence we can make the following deﬁnition: Deﬁnition 2.1.3. Let X ⊂ An be an afﬁne variety. The quotient ﬁeld K(X) of A(X) is called the ﬁeld of rational functions on X. Remark 2.1.4. Recall that the quotient ﬁeld K of an integral domain R is deﬁned to be the set of pairs ( f , g) with f , g ∈ R, g = 0, modulo the equivalence relation ( f , g) ∼ ( f , g ) ⇐⇒ f g − g f = 0.

2.

Functions, morphisms, and varieties

19

f An element ( f , g) of K is usually written as g , and we think of it as the formal quotient of two ring elements. Addition of two such formal quotients is deﬁned in the same way as you would expect to add fractions, namely

fg +gf f f , + := g g gg and similarly for subtraction, multiplication, and division. This makes K(X) into a ﬁeld. In the case where R = A(X) is the coordinate ring of an afﬁne variety, we can therefore think of elements of K(X) as being quotients of polynomial functions. We have to be very careful with this interpretation though, see example 2.1.7 and lemma 2.1.8. Now let us deﬁne what we want to mean by a regular function on an open subset U of an afﬁne variety X. This is more or less obvious: a regular function should be a rational function that is well-deﬁned at all points of U: Deﬁnition 2.1.5. Let X ⊂ An be an afﬁne variety and let P ∈ X be a point. We call

OX,P :=

f ; f , g ∈ A(X) and g(P) = 0 ⊂ K(X) g

the local ring of X at the point P. Obviously, this should be thought of as the rational functions that are regular at P. If U ⊂ X is a non-empty open subset, we set

OX (U) :=

\

P∈U

OX,P .

This is a subring of K(X). We call this the ring of regular functions on U. Remark 2.1.6. The set mX,P := { f ∈ A(X) ; f (P) = 0} of all functions that vanish at P is an ideal in A(X). This is a maximal ideal, as A(X)/mX,P ∼ k, the isomorphism being = evaluation of the polynomial at the point P. With this deﬁnition, OX,P is just the localization of the ring A(X) at the maximal ideal mX,P . We will explain in lemma 2.2.10 where the name “local” (resp. “localization”) comes from. Example 2.1.7. We have just deﬁned regular functions on an open subset of an afﬁne variety X ⊂ An to be rational functions, i.e. elements in the quotient ﬁeld K(X), with certain properties. This means that every such function can be written as the “quotient” of two elements in A(X). It does not mean however that we can always write a regular function as the quotient of two polynomials in k[x1 , . . . , nn ]. Here is an example showing this. Let X ⊂ A4 be the variety deﬁned by the equation x1 x4 = x2 x3 , and let U ⊂ X be the open subset of all points in X where x2 = 0 or x4 = 0. The function x1 is deﬁned at all x2 x3 points of X where x2 = 0, and the function x4 is deﬁned at points of X where x4 = 0. By the equation of X, these two functions coincide where they are both deﬁned; in other words x1 x3 = ∈ K(X) x2 x4 by remark 2.1.4. So this gives rise to a regular function on U. But there is no representation of this function as a quotient of two polynomials in k[x1 , x2 , x3 , x4 ] that works on all of U — we have to use different representations at different points. As we will usually want to write down regular functions as quotients of polynomials, we should prove a precise statement how regular functions can be patched together from different polynomial representations: Lemma 2.1.8. The following deﬁnition of regular functions is equivalent to the one of deﬁnition 2.1.5: Let U be an open subset of an afﬁne variety X ⊂ An . A set-theoretic map ϕ : U → k is called regular at the point P ∈ U if there is a neighborhood V of P in U such that there are

20

Andreas Gathmann

f (Q) g(Q)

polynomials f , g ∈ k[x1 , . . . , xn ] with g(Q) = 0 and ϕ(Q) = regular on U if it is regular at every point in U.

for all Q ∈ V . It is called

Proof. It is obvious that an element of the ring of regular functions on U determines a regular function in the sense of the lemma. Conversely, let ϕ : U → A1 be a regular function in the sense of the lemma. Let P ∈ U f (Q) be any point, then there are polynomials f , g such that g(Q) = 0 and ϕ(Q) = g(Q) for all

f points Q in some neighborhood V of P. We claim that g ∈ K(X) is the element in the ring of regular functions that we seek. In fact, all we have to show is that this element does not depend on the choices that we made. So let P ∈ U be another point (not necessarily distinct from P), and suppose f f that there are polynomials f , g such that g = g on some neighborhood V of P . Then f g = g f on V ∩ V and hence on X as V ∩ V is dense in X by remark 1.3.17. In other f f words, f g − g f ∈ I(X), so it is zero in A(X), i.e. g = g ∈ K(X).

Remark 2.1.9. An almost trivial but remarkable consequence of our deﬁnition of regular functions is the following: let U ⊂ V be non-empty open subsets of an afﬁne variety X. If ϕ1 , ϕ2 : V → k are two regular functions on V that agree on U, then they agree on all of V . This is obvious because the ring of regular functions (on any non-empty open subset) is a subring of the function ﬁeld K(X), so if two such regular functions agree this just means that they are the same element of K(X). Of course, this is not surprising as open subsets are always dense, so if we know a regular function on an open subset it is intuitively clear that we know it almost everywhere anyway. The interesting remark here is that the very same statement holds in complex analysis for holomorphic functions as well (or more generally, in real analysis for analytic functions): two holomorphic functions on a (connected) open subset U ⊂ Cn must be the same if they agree on any smaller open subset V ⊂ U. This is called the identity theorem for holomorphic functions — in complex analysis this is a real theorem because there the open subset V can be “very small”, so the statement that the extension to U is unique is a lot more surprising than it is here in algebraic geometry. Still this is an example of a theorem that is true in literally the same way in both algebraic and complex geometry, although these two theories are quite different a priori. Let us compute the rings OX (U) explicitly in the cases where U is the complement of the zero locus of just a single polynomial. Proposition 2.1.10. Let X ⊂ An be an afﬁne variety. Let f ∈ A(X) and X f = {P ∈ X ; f (P) = 0}. (Open subsets of this form are called distinguished open subsets.) Then

OX (X f ) = A(X) f :=

g ; g ∈ A(X) and r ≥ 0 . fr

In particular, OX (X) = A(X), i.e. any regular function on X is polynomial (take f = 1). Proof. It is obvious that A(X) f ⊂ OX (X f ), so let us prove the converse. Let ϕ ∈ OX (X f ) ⊂ K(X). Let J = {g ∈ A(X) ; gϕ ∈ A(X)}. This is an ideal in A(X); we want to show that f r ∈ J for some r. For any P ∈ X f we know that ϕ ∈ OX,P , so ϕ = h with g = 0 in a neighborhood of P. g In particular g ∈ J, so J contains an element not vanishing at P. This means that the zero locus of the ideal I(X) + J ⊂ k[x1 , . . . , xn ] is contained in the set {P ∈ X ; f (P) = 0}, or in other words that Z(I(X) + J) ⊂ Z( f ). By proposition 1.2.9 (i) it follows that I(Z( f )) ⊂ I(Z(I(X) + J)). So f ∈ I(Z(I(X) + J)), where f ∈ k[x1 , . . . , xn ] is a representative of f . Therefore f r ∈ I(X) + J for some r by the Nullstellensatz 1.2.9 (iii), and so f r ∈ J.

2.

Functions, morphisms, and varieties

21

Remark 2.1.11. In the proof of proposition 2.1.10 we had to use the Nullstellensatz again. In fact, the statement is false if the ground ﬁeld is not algebraically closed, as you can see from the example of the function x21 that is regular on all of A1 (R), but not polynomial. +1 Example 2.1.12. Probably the easiest case of an open subset of an afﬁne variety X that is not of the form X f as in proposition 2.1.10 is the complement U = C2 \{0} of the origin in the afﬁne plane. Let us compute OC2 (U). By deﬁnition 2.1.5 any element ϕ ∈ OC2 (U) ⊂ f C(x, y) is globally the quotient ϕ = g of two polynomials f , g ∈ C[x, y]. The condition that we have to satisfy is that g(x, y) = 0 for all (x, y) = (0, 0). We claim that this implies that g is constant. (In fact, this follows intuitively from the fact that a single equation can cut down the dimension of a space by only 1, so the zero locus of the polynomial g cannot only be the origin in C2 . But we have not proved this rigorously yet.) We know already by the Nullstellensatz that there is no non-constant polynomial that has empty zero locus in C2 , so we can assume that g vanishes on (0, 0). If we write g as g(x, y) = f0 (x) + f1 (x) · y + f2 (x) · y2 + · · · + fn (x) · yn , this means that f0 (0) = 0. We claim that f0 (x) must be of the form xm for some m. In fact: • if f0 is the zero polynomial, then g(x, y) contains y as a factor and hence the whole x-axis in its zero locus, • if f0 contains more than one monomial, f0 has a zero x0 = 0, and hence g(x0 , 0) = 0. So g(x, y) is of the form g(x, y) = xm + f1 (x) · y + f2 (x) · y2 + · · · + fn (x) · yn . Now set y = ε for some small ε. As g(x, 0) = xm and all fi are continuous, the restriction g(x, ε) cannot be the zero or a constant polynomial. Hence g(x, ε) vanishes for some x, which is a contradiction. So we see that we cannot have any denominators, i.e. OC2 (U) = C[x, y]. In other words, a regular function on C2 \{0} is always regular on all of C2 . This is another example of a statement that is known from complex analysis for holomorphic functions, known as the removable singularity theorem. 2.2. Sheaves. We have seen in lemma 2.1.8 that regular functions on afﬁne varieties are deﬁned in terms of local properties: they are set-theoretic functions that can locally be written as quotients of polynomials. Local constructions of function-like objects occur in many places in algebraic geometry (and also in many other “topological” ﬁelds of mathematics), so we should formalize the idea of such objects. This will also give us an “automatic” deﬁnition of morphisms between afﬁne varieties in section 2.3. Deﬁnition 2.2.1. A presheaf F of rings on a topological space X consists of the data: • for every open set U ⊂ X a ring F (U) (think of this as the ring of functions on U), • for every inclusion U ⊂ V of open sets in X a ring homomorphism ρV,U : F (V ) → F (U) called the restriction map (think of this as the usual restriction of functions to a subset), such that / • F (0) = 0, • ρU,U is the identity map for all U, • for any inclusion U ⊂ V ⊂ W of open sets in X we have ρV,U ◦ ρW,V = ρW,U . The elements of F (U) are usually called the sections of F over U, and the restriction maps ρV,U are written as f → f |U .

22

Andreas Gathmann

A presheaf F of rings is called a sheaf of rings if it satisﬁes the following glueing property: if U ⊂ X is an open set, {Ui } an open cover of U and fi ∈ F (Ui ) sections for all i such that fi |Ui ∩U j = f j |Ui ∩U j for all i, j, then there is a unique f ∈ F (U) such that f |Ui = fi for all i. Remark 2.2.2. In the same way one can deﬁne (pre-)sheaves of Abelian groups / k-algebras etc., by requiring that all F (U) are objects and all ρV,U are morphisms in the corresponding category. Example 2.2.3. If X ⊂ An is an afﬁne variety, then the rings OX (U) of regular functions on open subsets of X (with the obvious restriction maps OX (V ) → OX (U) for U ⊂ V ) form a sheaf of rings OX , the sheaf of regular functions or structure sheaf on X. In fact, all deﬁning properties of presheaves are obvious, and the glueing property of sheaves is easily seen from the description of regular functions in lemma 2.1.8. Example 2.2.4. Here are some examples from other ﬁelds of mathematics: Let X = Rn , and for any open subset U ⊂ X let F (U) be the ring of continuous functions on U. Together with the obvious restriction maps, these rings F (U) form a sheaf, the sheaf of continuous functions. In the same way we can deﬁne the sheaf of k times differentiable functions, analytic functions, holomorphic functions on Cn , and so on. The same deﬁnitions can be applied if X is a real or complex manifold instead of just Rn or Cn . In all these examples, the sheaves just deﬁned “are” precisely the functions that are considered to be morphisms in the corresponding category (for example, in complex analysis the morphisms are just the holomorphic maps). This is usually expressed in the following way: a pair (X, F ) where X is a topological space and F is a sheaf of rings on X is called a ringed space. The sheaf F is then called the structure sheaf of this ringed space and usually written OX . Hence we have just given afﬁne varieties the structure of a ringed space. (Although being general, this terminology will usually only be applied if F actually has an interpretation as the space of functions that are considered to be morphisms in the corresponding category.) Remark 2.2.5. Intuitively speaking, any “function-like” object forms a presheaf; it is a sheaf if the conditions imposed on the “functions” are local. Here is an example illustrating this fact. Let X = R be the real line. For U ⊂ X open and non-empty let F (U) be the ring of constant (real-valued) functions on U, i.e. F (U) ∼ R for all U. Let ρV,U for U ⊂ V = be the obvious restriction maps. Then F is obviously a presheaf, but not a sheaf. This is because being constant is not a local property; it means that f (P) = f (Q) for all P and Q that are possibly quite far away. For example, let U = (0, 1) ∪ (2, 3). Then U has an open cover U = U1 ∪U2 with U1 = (0, 1) and U2 = (2, 3). Let f1 : U1 → R be the constant function 0, and let f2 : U2 → R be the constant function 1. Then f1 and f2 trivially agree / on the overlap U1 ∩U2 = 0, but there is no constant function on U that restricts to both f1 and f2 on U1 and U2 , respectively. There is however a uniquely deﬁned locally constant function on U with that property. In fact, it is easy to see that the locally constant functions on X do form a sheaf. Remark 2.2.6. If F is a sheaf on X and U ⊂ X is an open subset, then one deﬁnes the restriction of F to U, denoted F |U , by (F |U )(V ) = F (V ) for all open subsets V ⊂ U. Obviously, this is again a sheaf. Finally, let us see how the local rings of an afﬁne variety appear in the language of sheaves. Deﬁnition 2.2.7. Let X be a topological space, P ∈ X, and F a (pre-)sheaf on X. Consider pairs (U, ϕ) where U is an open neighborhood of P and ϕ ∈ F (U) a section of F over U. We call two such pairs (U, ϕ) and (U , ϕ ) equivalent if there is an open neighborhood V of

2.

Functions, morphisms, and varieties

23

P with V ⊂ U ∩U such that ϕ|V = ϕ |V . (Note that this is in fact an equivalence relation.) The set of all such pairs modulo this equivalence relation is called the stalk FP of F at P, its elements are called germs of F . Remark 2.2.8. If F is a (pre-)sheaf of rings (or k-algebras, Abelian groups, etc.) then the stalks of F are rings (or k-algebras, Abelian groups, etc.). Remark 2.2.9. The interpretation of the stalk of a sheaf is obviously that its elements are sections of F that are deﬁned in an (arbitrarily small) neighborhood around P. Hence e.g. on the real line the germ of a differentiable function at a point P allows you to compute the derivative of this function at P, but none of the actual values of the function at any point besides P. On the other hand, we have seen in remark 2.1.9 that holomorphic functions on a (connected) complex manifold are already determined by their values on any open set, so germs of holomorphic functions carry “much more information” than germs of differentiable functions. In algebraic geometry, this is similar: it is already quite obvious that germs of regular functions must carry much information, as the open subsets in the Zariski topology are so big. We will now show that the stalk of OX at a point P is exactly the local ring OX,P , which ﬁnally gives a good motivation for the name “local ring”. Lemma 2.2.10. Let X be an afﬁne variety and P ∈ X. The stalk of OX at P is OX,P . Proof. Recall that OX (U) ⊂ OX,P ⊂ K(X) for all P ∈ U by deﬁnition. Therefore, if we are given a pair (U, ϕ) with P ∈ U and ϕ ∈ OX (U), we see that ϕ ∈ OX,P determines an element in the local ring. If we have another equivalent pair (U , ϕ ), then ϕ and ϕ agree on some V with P ∈ V ⊂ U ∩U by deﬁnition, so they determine the same element in OX (V ) and hence in OX,P .

f Conversely, if ϕ ∈ OX,P is an element in the local ring, we can write it as ϕ = g with polynomials f , g such that g(P) = 0. Then there must be a neighborhood U of P on which g is non-zero, and therefore the (U, ϕ) deﬁnes an element in the stalk of OX at P.

2.3. Morphisms between afﬁne varieties. Having given the structure of ringed spaces to afﬁne varieties, there is a natural way to deﬁne morphisms between them. In this section we will allow ourselves to view morphisms as set-theoretic maps on the underlying topological spaces with additional properties (see lemma 2.1.8). Deﬁnition 2.3.1. Let (X, OX ) and (Y, OY ) be ringed spaces whose structure sheaves OX and OY are sheaves of k-valued functions (in the case we are considering right now X and Y will be afﬁne varieties or open subsets of afﬁne varieties). Let f : X → Y be a set-theoretic map. (i) If ϕ : U → k is a k-valued (set-theoretic) function on an open subset U of Y , the composition ϕ ◦ f : f −1 (U) → k is again a set-theoretic function. It is denoted by f ∗ ϕ and is called the pull-back of ϕ. (ii) The map f is called a morphism if it is continuous, and if it pulls back regular functions to regular functions, i.e. if f ∗ OY (U) ⊂ OX ( f −1 (U)) for all open U ⊂ Y . Remark 2.3.2. Recall that a function f : X → Y between topological spaces is called continuous if inverse images of open subsets are always open. In the above deﬁnition (ii), the requirement that f be continuous is therefore necessary to formulate the second condition, as it ensures that f −1 (U) is open, so that OX ( f −1 (U)) is well-deﬁned. Remark 2.3.3. In our context of algebraic geometry OX and OY will always be the sheaves of regular maps constructed in deﬁnition 2.1.5. But in fact, this deﬁnition of morphisms is used in many other categories as well, e.g. one can say that a set-theoretic map between complex manifolds is holomorphic if it pulls back holomorphic functions to holomorphic functions. In fact, it is almost the general deﬁnition of morphisms between ringed spaces —

24

Andreas Gathmann

the only additional twist in the general case is that if f : X → Y is a continuous map between arbitrary ringed spaces (X, OX ) and (Y, OY ), there is no a priori deﬁnition of the pull-back map OY (U) → OX ( f −1 (U)). In the case above we solved this problem by applying the settheoretic viewpoint that gave us a notion of pull-back in our special case. In more general cases (e.g. for schemes that we will discuss later in section 5) one will have to include these pull-back maps in the data needed to deﬁne a morphism. We now want to show that for afﬁne varieties the situation is a lot easier: we actually do not have to deal with open subsets, but it sufﬁces to check the pull-back property on global functions only: Lemma 2.3.4. Let f : X → Y be a continuous map between afﬁne varieties. Then the following are equivalent: (i) f is a morphism (i.e. f pulls back regular functions on open subsets to regular functions on open subsets). (ii) For every ϕ ∈ OY (Y ) we have f ∗ ϕ ∈ OX (X), i.e. f pulls back global regular functions to global regular functions. (iii) For every P ∈ X and every ϕ ∈ OY, f (P) we have f ∗ ϕ ∈ OX,P , i.e. f pulls back germs of regular functions to germs of regular functions. Proof. (i) ⇒ (ii) is trivial, and (iii) ⇒ (i) follows immediately from the deﬁnition of OY (U) and OX ( f −1 (U)) as intersections of local rings. To prove (ii) ⇒ (iii) let ϕ ∈ OY, f (P) be g the germ of a regular function on Y . We write ϕ = h with g, h ∈ A(Y ) = OY (Y ) and ∗ g and f ∗ h are global regular functions in A(X) = O (X), hence h( f (P)) = 0. By (ii), f X ∗g f ∗ ϕ = f ∗ h ∈ OX,P , since we have h( f (P)) = 0. f Example 2.3.5. Let X = A1 be the afﬁne line with coordinate x, and let Y = A1 be the afﬁne line with coordinate y. Consider the set-theoretic map f : X → Y, x → y = x2 .

We claim that this is a morphism. In fact, by lemma 2.3.4 (ii) we just have to show that f pulls back polynomials in k[y] to polynomials in k[x]. But this is obvious, as the pull-back of a polynomial ϕ(y) ∈ k[y] is just ϕ(x2 ) (i.e. we substitute x2 for y in ϕ). This is still a polynomial, so it is in k[x]. Example 2.3.6. For the very same reason, every polynomial map is a morphism. More precisely, let X ⊂ Am and Y ⊂ An be afﬁne varieties, and let f : X → Y be a polynomial map, i.e. a map that can be written as f (P) = ( f1 (P), . . . , fn (P)) with fi ∈ k[x1 , . . . , xm ]. As f then pulls back polynomials to polynomials, we conclude ﬁrst of all that f is continuous. Moreover, it then follows from lemma 2.3.4 (ii) that f is a morphism. In fact, we will show now that all morphisms between afﬁne varieties are of this form. Lemma 2.3.7. Let X ⊂ An and Y ⊂ Am be afﬁne varieties. There is a one-to-one correspondence between morphisms f : X → Y and k-algebra homomorphisms f ∗ : A(Y ) → A(X). Proof. Any morphism f : X → Y determines a k-algebra homomorphism f ∗ : OY (Y ) = A(Y ) → OX (X) = A(X) by deﬁnition. Conversely, if g : k[y1 , . . . , ym ]/I(Y ) → k[x1 , . . . , xn ]/I(X) is any k-algebra homomorphism then it determines a polynomial map f = ( f1 , . . . , fm ) : X → Y as in example 2.3.6 by fi = g(yi ), and hence a morphism.

2.

Functions, morphisms, and varieties

25

Example 2.3.8. Of course, an isomorphism is deﬁned to be a morphism f : X → Y that has an inverse (i.e. a morphism such that there is a morphism g : Y → X with g ◦ f = idX and f ◦ g = idY ). A warning is in place here that an isomorphism of afﬁne varieties is not the same as a bijective morphism (in contrast e.g. to the case of vector spaces). For example, let X ⊂ A2 be the curve given by the equation x2 = y3 , and consider the map f : A1 → X,

A1

t → (x = t 3 , y = t 2 ).

x2= y3 f

This is a morphism as it is given by polynomials, and it is bijective as the inverse is given by x if (x, y) = (0, 0), f −1 : X → A1 , (x, y) → y 0 if (x, y) = (0, 0). But if f was an isomorphism, the corresponding k-algebra homomorphism k[x, y]/(x2 − y3 ) → k[t], x → t 3 and y → t 2

would have to be an isomorphism by lemma 2.3.7. This is obviously not the case, as the image of this algebra homomorphism contains no linear polynomials. Example 2.3.9. As an application of morphisms, let us consider products of afﬁne varieties. Let X ⊂ An and Y ⊂ Am be afﬁne varieties with ideals I(X) ⊂ k[x1 , . . . , xn ] and I(Y ) ⊂ k[y1 , . . . , ym ]. As usual, we deﬁne the product X ×Y of X and Y to be the set X ×Y = {(P, Q) ∈ An × Am ; P ∈ X and Q ∈ Y } ⊂ An × Am = An+m . Obviously, this is an algebraic set in An+m with ideal I(X ×Y ) = I(X) + I(Y ) ⊂ k[x1 , . . . , xn , y1 , . . . , ym ] where we consider k[x1 , . . . , xn ] and k[y1 , . . . , ym ] as subalgebras of k[x1 , . . . , xn , y1 , . . . , ym ] in the obvious way. Let us show that it is in fact a variety, i.e. irreducible: Proposition 2.3.10. If X and Y are afﬁne varieties, then so is X ×Y . Proof. For simplicity, let us just write x for the collection of the xi , and y for the collection of the yi . By the above discussion it only remains to show that I(X × Y ) is prime. So let f , g ∈ k[x, y] be polynomial functions such that f g ∈ I(X ×Y ); we have to show that either f or g vanishes on all of X ×Y , i.e. that X ×Y ⊂ Z( f ) or X ×Y ⊂ Z(g). So let us assume that X × Y ⊂ Z( f ), i.e. there is a point (P, Q) ∈ X × Y \Z( f ) (where P ∈ X and Q ∈ Y ). Denote by f (·, Q) ∈ k[x] the polynomial obtained from f ∈ k[x, y] by plugging in the coordinates of Q for y. For all P ∈ X\Z( f (·, Q)) (e.g. for P = P) we must have Y ⊂ Z( f (P , ·) · g(P , ·)) = Z( f (P , ·)) ∪ Z(g(P , ·)). As Y is irreducible and Y ⊂ Z( f (P , ·)) by the choice of P , it follows that Y ⊂ Z(g(P , ·)). This is true for all P ∈ X\Z( f (·, Q)), so we conclude that (X\Z( f (·, Q)) × Y ⊂ Z(g). But as Z(g) is closed, it must in fact contain the closure of (X\Z( f (·, Q)) × Y as well, which is just X × Y as X is irreducible and X\Z( f (·, Q)) a non-empty open subset of X (see remark 1.3.17). The obvious projection maps πX : X ×Y → X, (P, Q) → P and πY : X ×Y → Y, (P, Q) → Q are given by (trivial) polynomial maps and are therefore morphisms. The important main property of the product X ×Y is the following:

26

Andreas Gathmann

Lemma 2.3.11. Let X and Y be afﬁne varieties. Then the product X × Y satisﬁes the following universal property: for every afﬁne variety Z and morphisms f : Z → X and g : Z → Y , there is a unique morphism h : Z → X ×Y such that f = πX ◦ h and g = πY ◦ h, i.e. such that the following diagram commutes: Z

h g

f

" X ×Y X

πX

%

πY

/ Y

In other words, giving a morphism Z → X × Y “is the same” as giving two morphisms Z → X and Z → Y . Proof. Let A be the coordinate ring of Z. Then by lemma 2.3.7 the morphism f : Z → X is given by a k-algebra homomorphism f˜ : k[x1 , . . . , xn ]/I(X) → A. This in turn is determined by giving the images f˜i := f˜(xi ) ∈ A of the generators xi , satisfying the relations of I (i.e. F( f˜1 , . . . , f˜n ) = 0 for all F(x1 , . . . , xn ) ∈ I(X)). The same is true for g, which is determined by the images gi := g(yi ) ∈ A. ˜ ˜ Now it is obvious that the elements f˜i and gi determine a k-algebra homomorphism ˜ k[x1 , . . . , xn , y1 , . . . , ym ]/(I(X) + I(Y )) → A, which determines a morphism h : Z → X ×Y by lemma 2.3.7. To show uniqueness, just note that the relations f = πX ◦ h and g = πY ◦ h imply immediately that h must be given set-theoretically by h(P) = ( f (P), g(P)) for all P ∈ Z. Remark 2.3.12. It is easy to see that the property of lemma 2.3.11 determines the product X × Y uniquely up to isomorphism. It is therefore often taken to be the deﬁning property for products. Remark 2.3.13. If you have heard about tensor products before, you will have noticed that the coordinate ring of X ×Y is just the tensor product A(X) ⊗ A(Y ) of the coordinate rings of the factors (where the tensor product is taken as k-algebras). See also section 5.4 for more details. Remark 2.3.14. Lemma 2.3.7 allows us to associate an afﬁne variety up to isomorphism to any ﬁnitely generated k-algebra that is a domain: if A is such an algebra, let x1 , . . . , xn be generators of A, so that A = k[x1 , . . . , xn ]/I for some ideal I. Let X be the afﬁne variety in An deﬁned by the ideal I; by the lemma it is deﬁned up to isomorphism. Therefore we should make a (very minor) redeﬁnition of the term “afﬁne variety” to allow for objects that are isomorphic to an afﬁne variety in the old sense, but that do not come with an intrinsic description as the zero locus of some polynomials in afﬁne space: Deﬁnition 2.3.15. A ringed space (X, OX ) is called an afﬁne variety over k if (i) X is irreducible, (ii) OX is a sheaf of k-valued functions, (iii) X is isomorphic to an afﬁne variety in the sense of deﬁnition 1.3.1. Here is an example of an afﬁne variety in this new sense although it is not a priori given as the zero locus of some polynomials in afﬁne space: Lemma 2.3.16. Let X be an afﬁne variety and f ∈ A(X), and let X f = X\Z( f ) be a distinguished open subset as in proposition 2.1.10. Then the ringed space (X f , OX |X f ) is isomorphic to an afﬁne variety with coordinate ring A(X) f .

2.

Functions, morphisms, and varieties

27

Proof. Let X ⊂ An be an afﬁne variety, and let f ∈ k[x1 , . . . , xn ] be a representative of f . Let J ⊂ k[x1 , . . . , xn ,t] be the ideal generated by I(X) and the function 1 − t f . We claim that the ringed space (X f , OX |X f ) is isomorphic to the afﬁne variety Z(J) = {(P, λ) ; P ∈ X and λ =

1 f (P) }

⊂ An+1 .

Consider the projection map π : Z(J) → X given by π(P, λ) = P. This is a morphism with 1 image X f and inverse morphism π−1 (P) = (P, f (P) ), hence π is an isomorphism. It is obvious that A(Z(J)) = A(X) f . Remark 2.3.17. So we have just shown that even open subsets of afﬁne varieties are themselves afﬁne varieties, provided that the open subset is the complement of the zero locus of a single polynomial equation. Example 2.1.12 shows however that not all open subsets of afﬁne varieties are themselves isomorphic to afﬁne varieties: if U ⊂ C2 \{0} we have seen that OU (U) = k[x, y]. So if U was an afﬁne variety, its coordinate ring must be k[x, y], which is the same as the coordinate ring of C2 . By lemma 2.3.7 this means that U and C2 would have to be isomorphic, with the isomorphism given by the identity map. Obviously, this is not true. Hence U is not an afﬁne variety. It can however be covered by two open subsets {x = 0} and {y = 0} which are both afﬁne by lemma 2.3.16. This leads us to the idea of patching afﬁne varieties together, which we will do in the next section. 2.4. Prevarieties. Now we want to extend our category of objects to more general things than just afﬁne varieties. Remark 2.3.17 showed us that not all open subsets of afﬁne varieties are themselves isomorphic to afﬁne varieties. But note that every open subset of an afﬁne variety can be written as a ﬁnite union of distinguished open subsets (as this is equivalent to the statement that every closed subset of an afﬁne variety is the zero locus of ﬁnitely many polynomials). Hence any such open subset can be covered by afﬁne varieties. This leads us to the idea that we should study objects that are not afﬁne varieties themselves, but rather can be covered by (ﬁnitely many) afﬁne varieties. Note that the following deﬁnition is completely parallel to the deﬁnition 2.3.15 of afﬁne varieties (in the new sense). Deﬁnition 2.4.1. A prevariety is a ringed space (X, OX ) such that (i) X is irreducible, (ii) OX is a sheaf of k-valued functions, (iii) there is a ﬁnite open cover {Ui } of X such that (Ui , OX |Ui ) is an afﬁne variety for all i. As before, a morphism of prevarieties is just a morphism as ringed spaces (see deﬁnition 2.3.1). An open subset U ⊂ X of a prevariety such that (U, OX |U ) is isomorphic to an afﬁne variety is called an afﬁne open set. Example 2.4.2. Afﬁne varieties and open subsets of afﬁne varieties are prevarieties (the irreducibility of open subsets follows from exercise 1.4.6). Remark 2.4.3. The above deﬁnition is completely analogous to the deﬁnition of manifolds. Recall how manifolds are deﬁned: ﬁrst you look at open subsets of Rn that are supposed to form the patches of your space, and then you deﬁne a manifold to be a topological space that looks locally like these patches. In the algebraic case now we can say that the afﬁne varieties form the basic patches of the spaces that we want to consider, and that e.g. open subsets of afﬁne varieties are spaces that look locally like afﬁne varieties.

28

Andreas Gathmann

As we deﬁned a prevariety to be a space that can be covered by afﬁne opens, the most general way to construct prevarieties is of course to take some afﬁne varieties (or prevarieties that we have already constructed) and patch them together: Example 2.4.4. Let X1 , X2 be prevarieties, let U1 ⊂ X1 and U2 ⊂ X2 be non-empty open subsets, and let f : (U1 , OX1 |U1 ) → (U2 , OX2 |U2 ) be an isomorphism. Then we can deﬁne a prevariety X, obtained by glueing X1 and X2 along U1 and U2 via the isomorphism f : • As a set, the space X is just the disjoint union X1 ∪ X2 modulo the equivalence relation P ∼ f (P) for all P ∈ U1 . • As a topological space, we endow X with the so-called quotient topology induced by the above equivalence relation, i.e. we say that a subset U ⊂ X is open if and only if i−1 (U) ⊂ X1 and i−1 (U) ⊂ X2 are both open, with i1 : X1 → X and 1 2 i2 : X2 → X being the obvious inclusion maps. • As a ringed space, we deﬁne the structure sheaf OX by

OX (U) = {(ϕ1 , ϕ2 ) ; ϕ1 ∈ OX1 (i−1 (U)), ϕ2 ∈ OX2 (i−1 (U)), 2 1

ϕ1 = ϕ2 on the overlap (i.e. f ∗ (ϕ2 |i−1 (U)∩U ) = ϕ1 |i−1 (U)∩U )}.

2 2 1 1

It is easy to check that this deﬁnes a sheaf of k-valued functions on X and that X is irreducible. Of course, every point of X has an afﬁne neighborhood, so X is in fact a prevariety. Example 2.4.5. As an example of the above glueing construction, let X1 = X2 = A1 , U1 = U2 = A1 \{0}. • Let f : U1 → U2 be the isomorphism x → 1 . The space X can be thought of as x A1 ∪ {∞}: of course the afﬁne line X1 = A1 ⊂ X sits in X. The complement X\X1 is a single point that corresponds to the zero point in X2 ∼ A1 and hence = 1 to “∞ = 0 ” in the coordinate of X1 . In the case k = C, the space X is just the Riemann sphere C∞ .

8 X2 f X1 0 0 glue X2 f X1 glue 8 X X

We denote this space by P1 . (This is a special case of a projective space; see section 3.1 and remark 3.3.7 for more details.) • Let f : U1 → U2 be the identity map. Then the space X obtained by glueing along f is “the afﬁne line with the zero point doubled”:

Obviously this is a somewhat weird space. Speaking in classical terms (and thinking of the complex numbers), if we have a sequence of points tending to the zero, this sequence would have two possible limits, namely the two zero points. Usually we want to exclude such spaces from the objects we consider. In the theory of manifolds, this is simply done by requiring that a manifold satisﬁes the socalled Hausdorff property, i.e. that every two distinct points have disjoint open neighborhoods. This is obviously not satisﬁed for our space X. But the analogous deﬁnition does not make sense in the Zariski topology, as non-empty open

2.

Functions, morphisms, and varieties

29

subsets are never disjoint. Hence we need a different characterization of the geometric concept of “doubled points”. We will do this in section 2.5. Example 2.4.6. Let X be the complex afﬁne curve X = {(x, y) ∈ C2 ; y2 = (x − 1)(x − 2) · · · (x − 2n)}. We have already seen in example 0.1.1 that X can (and should) be “compactiﬁed” by adding two points at inﬁnity, corresponding to the limit x → ∞ and the two possible values for y. Let us now construct this compactiﬁed space rigorously as a prevariety. To be able to add a limit point “x = ∞” to our space, let us make a coordinate change x = 1 , so that the equation of the curve becomes ˜ x y2 x2n = (1 − x)(1 − 2x) · · · (1 − 2nx). ˜ ˜ ˜ ˜ If we make an additional coordinate change y = ˜

y xn ,

this becomes

y2 = (1 − x)(1 − 2x) · · · (1 − 2nx). ˜ ˜ ˜ ˜ In these coordinates we can add our two points at inﬁnity, as they now correspond to x = 0 ˜ (and therefore y = ±1). ˜ Summarizing, our “compactiﬁed curve” of example 0.1.1 is just the prevariety obtained by glueing the two afﬁne varieties X = {(x, y) ∈ C2 ; y2 = (x − 1)(x − 2) · · · (x − 2n)} ˜ and X = {(x, y) ∈ C2 ; y2 = (1 − x)(1 − 2x) · · · (1 − 2nx)} ˜ ˜ ˜ ˜ ˜ ˜ along the isomorphism ˜ f :U → U, ˜ f −1 :U → U, (x, y) → (x, y) = ˜ ˜ (x, y) → (x, y) = ˜ ˜ 1 y , , x xn 1 y ˜ , , x xn ˜ ˜

˜ ˜ where U = {x = 0} ⊂ X and U = {x = 0} ⊂ X. ˜ Of course one can also glue together more than two prevarieties. As the construction is the same as in the case above, we will just give the statement and leave its proof as an exercise: Lemma 2.4.7. Let X1 , . . . , Xr be prevarieties, and let Ui, j ⊂ Xi be non-empty open subsets for i, j = 1, . . . , r. Let fi, j : Ui, j → U j,i be isomorphisms such that

−1 (i) fi, j = f j,i , (ii) fi,k = f j,k ◦ fi, j where deﬁned.

Then there is a prevariety X, obtained by glueing the Xi along the morphisms fi, j as in example 2.4.4 (see below). Remark 2.4.8. The prevariety X in the lemma 2.4.7 can be described as follows: • As a set, X is the disjoint union of the Xi , modulo the equivalence relation P ∼ fi, j (P) for all P ∈ Ui, j . • To deﬁne X as a topological space, we say that a subset Y ⊂ X is closed if and only if all restrictions Y ∩ Xi are closed. • A regular function on an open set U ⊂ X is a collection of regular functions ϕi ∈ OXi (Xi ∩U) that agree on the overlaps.

30

Andreas Gathmann

Condition (ii) of the lemma gives a compatibility condition for triple overlaps: consider three patches Xi , X j , Xk that have a common intersection. Then we want to identify every point P ∈ Ui, j with fi, j (P) ∈ U j,k , and the point fi, j (P) with f j,k ( fi, j (P)) (if it lies in U j,k ). So the glueing map fi,k must map P to the same point f j,k ( fi, j (P)) to get a consistent glueing. This is illustrated in the following picture:

Xi Ui,j Ui,k P fi,k Uk,i fj,k Uk,j Xk fi,j Uj,i Uj,k Xj X glue P

Let us now consider some examples of morphisms between prevarieties. Example 2.4.9. Let f : P1 → A1 be a morphism. We claim that f must be constant. In fact, consider the restriction f |A1 of f to the open afﬁne subset A1 ⊂ P1 . By deﬁnition the restriction of a morphism is again a morphism, so f |A1 must be of the form x → y = p(x) for some polynomial p ∈ k[x]. Now consider the second patch of P1 with coordinate x = 1 . ˜ x Applying this coordinate change, we see that f |P1 \{0} is given by x → p( 1 ). But this must ˜ x ˜ be a morphism too, i.e. p( 1 ) must be a polynomial in x. This is only true if p is a constant. ˜ x ˜ In the same way as prevarieties can be glued, we can patch together morphisms too. Of course, the statement is essentially that we can check the property of being a morphism on afﬁne open covers: Lemma 2.4.10. Let X,Y be prevarieties and let f : X → Y be a set-theoretic map. Let {U1 , . . . ,Ur } be an open cover of X and {V1 , . . . ,Vr } an afﬁne open cover of Y such that f (Ui ) ⊂ Vi and ( f |Ui )∗ OY (Vi ) ⊂ OX (Ui ). Then f is a morphism. Proof. We may assume that the Ui are afﬁne, as otherwise we can replace the Ui by a set of afﬁnes that cover Ui . Consider the restrictions fi : Ui → Vi . The homomorphism fi∗ : OY (Vi ) = A(Vi ) → OX (Ui ) = A(Ui ) is induced by some morphism Ui → Vi by lemma 2.3.7 which is easily seen to coincide with fi . In particular, the fi are continuous, and therefore so is f . It remains to show that f ∗ takes sections of OY to sections of OX . But if V ⊂ Y is open and ϕ ∈ OY (V ), then f ∗ ϕ is at least a section of OX on the sets f −1 (V ) ∩ Ui . Since OX is a sheaf and the Ui cover X, these sections glue to give a section in OX ( f −1 (V )). Example 2.4.11. Let f : A1 → A1 , x → y = f (x) be a morphism given by a polynomial f ∈ k[x]. We claim that there is a unique extension morphism f˜ : P1 → P1 such that f˜|A1 = f . We can assume that f = ∑n ai xi is not constant, as otherwise the result is trivial. It is then i=1 obvious that the extension should be given by f˜(∞) = ∞. Let us check that this deﬁnes in fact a morphism. We want to apply lemma 2.4.10. Consider the standard open afﬁne cover of the domain P1 by the two afﬁne lines V1 = P1 \{∞} and V2 = P1 \{0}. Then U1 := f˜−1 (V1 ) = A1 , and f˜|A1 = f is a morphism. On the other hand, let U2 := f˜−1 (V2 )\{0}. Consider the coordinates x = 1 and y = 1 on U2 and V2 , respectively. In these coordinates, the map f˜ is ˜ x ˜ y given by xn ˜ y= n ˜ ; ˜ ∑i=1 ai xn−i

2.

Functions, morphisms, and varieties

31

in particular x = 0 maps to y = 0. So by deﬁning f˜(∞) = ∞, we get a morphism f˜ : P1 → P1 ˜ ˜ that extends f by lemma 2.4.10. 2.5. Varieties. Recall example 2.4.5 (ii) where we constructed a prevariety that was “not Hausdorff” in the classical sense: take two copies of the afﬁne line A1 and glue them together on the open set A1 \{0} along the identity map. The prevariety X thus obtained is the “afﬁne line with the origin doubled”; its strange property is that even in the classical topology (for k = C) the two origins do not have disjoint neighborhoods. We will now deﬁne an algebro-geometric analogue of the Hausdorff property and say that a prevariety is a variety if it has this property. Deﬁnition 2.5.1. Let X be a prevariety. We say that X is a variety if for every prevariety Y and every two morphisms f1 , f2 : Y → X, the set {P ∈ Y ; f1 (P) = f2 (P)} is closed in Y . Remark 2.5.2. Let X be the afﬁne line with the origin doubled. Then X is not a variety — if we take Y = A1 and let f1 , f2 : Y → X be the two obvious inclusions that map the origin in Y to the two different origins in X, then f1 and f2 agree on A1 \{0}, which is not closed in A1 . On the other hand, if X has the Hausdorff property that we want to characterize, then two morphisms Y → X that agree on an open subset of Y should also agree on Y . One can make this deﬁnition somewhat easier and eliminate the need for an arbitrary second prevariety Y . To do so note that we can deﬁne products of prevarieties in the same way as we have deﬁned products of afﬁne varieties (see example 2.3.9 and exercise 2.6.13). For any prevariety X, consider the diagonal morphism ∆ : X → X × X, P → (P, P). The image ∆(X) ⊂ X × X is called the diagonal of X. Of course, the morphism ∆ : X → ∆(X) is an isomorphism onto its image (with inverse morphism being given by (P, Q) → P). So the space ∆(X) is not really interesting as a new prevariety; instead the main question is how ∆(X) is embedded in X × X: Lemma 2.5.3. A prevariety X is a variety if and only if the diagonal ∆(X) is closed in X × X. Proof. It is obvious that a variety has this property (take Y = X × X and f1 , f2 the two projections to X). Conversely, if the diagonal ∆(X) is closed and f1 , f2 : Y → X are two morphisms, then they induce a morphism ( f1 , f2 ) : Y → X × X by the universal property of exercise 2.6.13, and {P ∈ Y | f1 (P) = f2 (P)} = ( f1 , f2 )−1 (∆(X)) is closed. Lemma 2.5.4. Every afﬁne variety is a variety. Any open or closed subprevariety of a variety is a variety. Proof. If X ⊂ An is an afﬁne variety with ideal I(X) = ( f1 , . . . , fr ), the diagonal ∆(X) ⊂ A2n is an afﬁne variety given by the equations fi (x1 , . . . , xn ) = 0 for i = 1, . . . , r and xi = yi for i = 1, . . . , n, where x1 , . . . , xn , y1 , . . . , yn are the coordinates on A2n . This is obviously closed, so X is a variety by lemma 2.5.3. If Y ⊂ X is open or closed, then ∆(Y ) = ∆(X) ∩ (Y ×Y ); i.e. if ∆(X) is closed in X × X then so is ∆(Y ) in Y ×Y . Example 2.5.5. Let us illustrate lemma 2.5.3 in the case of the afﬁne line with a doubled origin. So let X1 = X2 = A1 , and let X be the prevariety obtained by glueing X1 to X2 along the identity on A\{0}. Then X × X is covered by the four afﬁne varieties X1 × X1 , X1 × X2 ,

32

Andreas Gathmann

X2 × X1 , and X2 × X2 by exercise 2.6.13. As we glue along A1 \{0} to obtain X, it follows that the space X × X contains the point (P, Q) ∈ A1 × A1 • once if P = 0 and Q = 0, • twice if P = 0 and Q = 0, or if P = 0 and Q = 0, • four times if P = 0 and Q = 0.

Xx X ∆(X)

In particular, X × X contains four origins now. But the diagonal ∆(X) contains only two of them (by deﬁnition of the diagonal we have to take the same origin in both factors). So on the patch X1 × X2 , the diagonal is given by {(P, P) ; P = 0} ⊂ X1 × X2 = A1 × A1 , which is not closed. So we see again that X is not a variety. 2.6. Exercises. Exercise 2.6.1. An algebraic set X ⊂ A2 deﬁned by a polynomial of degree 2 is called a conic. Show that any irreducible conic is isomorphic either to Z(y − x2 ) or to Z(xy − 1). Exercise 2.6.2. Let X,Y ⊂ A2 be irreducible conics as in exercise 2.6.1, and assume that X = Y . Show that X and Y intersect in at most 4 points. For all n ∈ {0, 1, 2, 3, 4}, ﬁnd an example of two conics that intersect in exactly n points. (For a generalization see theorem 6.2.1.) Exercise 2.6.3. Which of the following algebraic sets are isomorphic over the complex numbers? (a) A1 (b) Z(xy) ⊂ A2 (c) Z(x2 + y2 ) ⊂ A2 (d) Z(y2 − x3 − x2 ) ⊂ A2 (e) Z(x2 − y3 ) ⊂ A2 (f) Z(y − x2 , z − x3 ) ⊂ A3 Exercise 2.6.4. Let X be an afﬁne variety, and let G be a ﬁnite group. Assume that G acts on X, i.e. that for every g ∈ G we are given a morphism g : X → X (denoted by the same letter for simplicity of notation), such that (g ◦ h)(P) = g(h(P)) for all g, h ∈ G and P ∈ X. (i) Let A(X)G be the subalgebra of A(X) consisting of all G-invariant functions on X, i.e. of all f : X → k such that f (g(P)) = f (P) for all P ∈ X. Show that A(X)G is a ﬁnitely generated k-algebra. (ii) By (i), there is an afﬁne variety Y with coordinate ring A(X)G , together with a morphism π : X → Y determined by the inclusion A(X)G ⊂ A(X). Show that Y can be considered as the quotient of X by G, denoted X/G, in the following sense: (a) π is surjective. (b) If P, Q ∈ X then π(P) = π(Q) if and only if there is a g ∈ G such that g(P) = Q. (iii) For a given group action, is an afﬁne variety with the properties (ii)(a) and (ii)(b) uniquely determined? (iv) Let Zn = {exp( 2πik ) ; k ∈ Z} ⊂ C be the group of n-th roots of unity. Let Zn act n on Cm by multiplication in each coordinate. Show that C/Zn is isomorphic to C for all n, but that C2 /Zn is not isomorphic to C2 for n ≥ 2. Exercise 2.6.5. Are the following statements true or false: if f : An → Am is a polynomial map (i.e. f (P) = ( f1 (P), . . . , fm (P)) with fi ∈ k[x1 , . . . , xn ]), and. . .

2.

Functions, morphisms, and varieties

33

(i) X ⊂ An is an algebraic set, then the image f (X) ⊂ Am is an algebraic set. (ii) X ⊂ Am is an algebraic set, then the inverse image f −1 (X) ⊂ An is an algebraic set. (iii) X ⊂ An is an algebraic set, then the graph Γ = {(x, f (x)) | x ∈ X} ⊂ An+m is an algebraic set. Exercise 2.6.6. Let f : X → Y be a morphism between afﬁne varieties, and let f ∗ : A(Y ) → A(X) be the corresponding map of k-algebras. Which of the following statements are true? (i) (ii) (iii) (iv) If P ∈ X and Q ∈ Y , then f (P) = Q if and only if ( f ∗ )−1 (I(P)) = I(Q). f ∗ is injective if and only if f is surjective. f ∗ is surjective if and only if f is injective. f is an isomorphism if and only if f ∗ is an isomorphism.

If a statement is false, is there maybe a weaker form of it which is true? Exercise 2.6.7. Let X be a prevariety. Show that: (i) X is a Noetherian topological space, (ii) any open subset of X is a prevariety. Exercise 2.6.8. Let (X, OX ) be a prevariety, and let Y ⊂ X be an irreducible closed subset. For every open subset U ⊂ Y deﬁne OY (U) to be the ring of k-valued functions f on U with the following property: for every point P ∈ Y there is a neighborhood V of P in X and a section F ∈ OX (V ) such that f coincides with F on U. (i) Show that the rings OY (U) together with the obvious restriction maps deﬁne a sheaf OY on Y . (ii) Show that (Y, OY ) is a prevariety. Exercise 2.6.9. Let X be a prevariety. Consider pairs (U, f ) where U is an open subset of X and f ∈ OX (U) a regular function on U. We call two such pairs (U, f ) and (U , f ) equivalent if there is an open subset V in X with V ⊂ U ∩U such that f |U = f |U . (i) Show that this deﬁnes an equivalence relation. (ii) Show that the set of all such pairs modulo this equivalence relation is a ﬁeld. It is called the ﬁeld of rational functions on X and denoted K(X). (iii) If X is an afﬁne variety, show that K(X) is just the ﬁeld of rational functions as deﬁned in deﬁnition 2.1.3. (iv) If U ⊂ X is any non-empty open subset, show that K(U) = K(X). Exercise 2.6.10. If U is an open subset of a prevariety X and f : U → P1 a morphism, is it always true that f can be extended to a morphism f˜ : X → P1 ? Exercise 2.6.11. Show that the prevariety P1 is a variety. Exercise 2.6.12. (i) Show that every isomorphism f : A1 → A1 is of the form f (x) = ax + b for some a, b ∈ k, where x is the coordinate on A1 . (ii) Show that every isomorphism f : P1 → P1 is of the form f (x) = ax+b for some cx+d a, b, c, d ∈ k, where x is an afﬁne coordinate on A1 ⊂ P1 . (For a generalization see corollary 6.2.10.) (iii) Given three distinct points P1 , P2 , P3 ∈ P1 and three distinct points Q1 , Q2 , Q3 ∈ P1 , show that there is a unique isomorphism f : P1 → P1 such that f (Pi ) = Qi for i = 1, 2, 3. (Remark: If the ground ﬁeld is C, the very same statements are true in the complex analytic category as well, i.e. if “morphisms” are understood as “holomorphic maps” (and P1 is the Riemann sphere C∞ ). If you know some complex analysis and have some time to kill, you may try to prove this too.)

34

Andreas Gathmann

Exercise 2.6.13. Let X and Y be prevarieties with afﬁne open covers {Ui } and {V j }, respectively. Construct the product prevariety X × Y by glueing the afﬁne varieties Ui × V j together. Moreover, show that there are projection morphisms πX : X ×Y → X and πY : X × Y → Y satisfying the “usual” universal property for products: given morphisms f : Z → X and g : Z → Y from any prevariety Z, there is a unique morphism h : Z → X ×Y such that f = πX ◦ h and g = πY ◦ h.

3.

Projective varieties

35

3. P ROJECTIVE VARIETIES

Similarly to the afﬁne case, a subset of projective n-space Pn over k is called a projective algebraic set if it can be written as the zero locus of a (ﬁnite) set of homogeneous polynomials. The Zariski topology on Pn is the topology whose closed sets are the projective algebraic sets. The concepts of irreducibility and dimension are purely topological and extend therefore immediately to subsets of projective space. We prove a projective version of the Nullstellensatz and make projective varieties into ringed spaces that are varieties. The main property of projective varieties distinguishing them from afﬁne varieties is that (over C in the classical topology) they are compact. In terms of algebraic geometry this translates into the statement that if f : X → Y is a morphism between projective varieties then f (X) is closed in Y .

3.1. Projective spaces and projective varieties. In the last section we have studied varieties, i.e. topological spaces that are locally isomorphic to afﬁne varieties. In particular, the ability to glue afﬁne varieties together allowed us to construct compact spaces (over the ground ﬁeld C) like e.g. P1 , whereas afﬁne varieties themselves are never compact unless they are a single point (see exercise 3.5.6). Unfortunately, the description of a variety in terms of its afﬁne patches is often quite inconvenient in practice, as we have seen already in the calculations in the last section. It would be desirable to have a global description of the spaces that does not refer to glueing methods. Projective varieties form a large class of “compact” varieties that do admit such a uniﬁed global description. In fact, the class of projective varieties is so large that it is not easy to construct a variety that is not (an open subset of) a projective variety. To construct projective varieties, we need to deﬁne projective spaces ﬁrst. Projective spaces are “compactiﬁcations” of afﬁne spaces. We have seen P1 already as a compactiﬁcation of A1 ; general projective spaces are an extension of this construction to higher dimensions. Deﬁnition 3.1.1. We deﬁne projective n -space over k, denoted Pn , to be the set of all one-dimensional linear subspaces of the vector space kn+1 . Remark 3.1.2. Obviously, a one-dimensional linear subspace of kn+1 is uniquely determined by a non-zero vector in kn+1 . Conversely, two such vectors a = (a0 , . . . , an ) and b = (b0 , . . . , bn ) in kn+1 span the same linear subspace if and only if they differ only by a common scalar, i.e. if b = λa for some non-zero λ ∈ k. In other words, Pn = {(a0 , . . . , an ) ; ai ∈ k, not all ai = 0}/ ∼ with the equivalence relation (a0 , . . . , an ) ∼ (b0 , . . . , bn ) if ai = λbi for some λ ∈ k\{0} and all i. This is often written as Pn = (kn+1 \{0})/(k\{0}), and the point P in Pn determined by (a0 , . . . , an ) is written as P = (a0 : · · · : an ) (the notation [a0 , . . . , an ] is also common in the literature). So the notation (a0 : · · · : an ) means that the ai are not all zero, and that they are deﬁned only up to a common scalar multiple. The ai are called the homogeneous coordinates of the point P (the motivation for this name will become obvious in the course of this section). Example 3.1.3. Consider the one-dimensional projective space P1 . Let (a0 : a1 ) ∈ P1 be a point. Then we have one of the following cases:

36

Andreas Gathmann

a1 (i) a0 = 0. Then P can be written as P = (1 : a) with a = a0 ∈ k. Obviously (1 : a) = (1 : b) if and only if a = b, i.e. the ambiguity in the homogeneous coordinates is gone if we ﬁx one of them to be 1. So the set of these points is just A1 . We call a1 a = a0 the afﬁne coordinate of the point P; it is uniquely determined by P (and not just up to a multiple as for the homogeneous coordinates). (ii) a0 = 0, and therefore a1 = 0. There is just one such point that we can write as (0 : 1).

So P1 is just A1 with one point added. This additional point (0 : 1) can be thought of as a 1 “point at inﬁnity”, as you can see from the fact that its afﬁne coordinate is formally 0 . So 1 as in example 2.4.5 (i). we arrive at the same description of P Remark 3.1.4. There is a completely analogous description of Pn as An with some points added “at inﬁnity”: let P = (a0 : · · · : an ) ∈ Pn be a point. Then we have one of the following cases:

ai (i) a0 = 0. Then P = (1 : α1 : · · · : αn ) with αi = a0 for all i. The αi are the afﬁne coordinates of P; they are uniquely determined by P and are obtained by setting a0 = 1. So the set of all P with a0 = 0 is just An . (ii) a0 = 0, i.e. P = (0 : a1 : · · · : an ), with the ai still deﬁned only up to a common scalar. Obviously, the set of such points is Pn−1 ; the set of all one-dimensional linear subspaces of An . We think of these points as points at inﬁnity; the new twist compared to P1 is just that we have a point at inﬁnity for every one-dimensional linear subspace of An , i.e. for every “direction” in An . So, for example, two lines in An will meet at inﬁnity (when compactiﬁed in Pn ) if and only if they are parallel, i.e. point in the same direction. (This is good as it implies that two distinct lines always intersect in exactly one point.)

Usually, it is more helpful to think of the projective space Pn as the afﬁne space An compactiﬁed by adding some points (parametrized by Pn−1 ) at inﬁnity, rather than as the set of lines in An+1 . Remark 3.1.5. In the case k = C, we claim that Pn is a compact space (in the classical topology). In fact, let S2n+1 = {(a0 , . . . , an ) ∈ Cn+1 ; |a0 |2 + · · · + |an |2 = 1} be the unit sphere in Cn+1 = R2n+2 . This is a compact space as it is closed and bounded, and there is an obvious surjective map S2n+1 → Pn , (a0 , · · · , an ) → (a0 : · · · : an ). As images of compact sets under continuous maps are compact, it follows that Pn is also compact. Remark 3.1.6. In complete analogy to afﬁne algebraic sets, we now want to deﬁne projective algebraic sets to be subsets of Pn that can be described as the zero locus of some polynomials in the homogeneous coordinates. Note however that if f ∈ k[x0 , . . . , xn ] is an arbitrary polynomial, it does not make sense to write down a deﬁnition like Z( f ) = {(a0 : · · · : an ) ; f (a0 , . . . , an ) = 0},

2 because the ai are only deﬁned up to a common scalar. For example, if f (x0 , x1 ) = x1 − x0 then f (1, 1) = 0 but f (−1, −1) = 0, although (1 : 1) and (−1 : −1) are the same point in P1 . To get rid of this problem we have to require that f be homogeneous, i.e. that all of its monomials have the same (total) degree d. This is equivalent to the requirement

f (λx0 , . . . , λxn ) = λd f (x0 , . . . , xn ) for all λ,

3.

Projective varieties

37

so in particular we see that f (λx0 , . . . , λxn ) = 0 ⇐⇒ f (x0 , . . . , xn ) = 0, i.e. the condition that a homogeneous polynomial in the homogeneous coordinates vanishes is indeed well-deﬁned. Deﬁnition 3.1.7. For every f ∈ k[x0 , . . . , xn ] let f (d) denote the degree-d part of f , i.e. f = ∑ f (d) with f (d) homogeneous of degree d for all d. Lemma 3.1.8. Let I ⊂ k[x0 , . . . , xn ] be an ideal. The following are equivalent: (i) I can be generated by homogeneous polynomials. (ii) For every f ∈ I we have f (d) ∈ I for all d. An ideal that satisﬁes these conditions is called homogeneous. Proof. (i) ⇒ (ii): Let I = ( f1 , . . . , fm ) with all fi homogeneous. Then every f ∈ I can be written as f = ∑i ai fi for some ai ∈ k[x0 , . . . , xn ] (which need not be homogeneous). Restricting this equation to the degree-d part, we get f (d) = ∑i (ai )(d−deg fi ) fi ∈ I. (ii) ⇒ (i): Any ideal can be written as I = ( f1 , . . . , fm ) with the fi possibly not being (d) homogeneous. But by (ii) we know that all fi are in I too, so it follows that I is generated (d) by the homogeneous polynomials fi . Remark 3.1.9. Note that it is not true that every element of a homogeneous ideal I is a homogeneous polynomial: we can always add two polynomials of I to get another element of I, even if they do not have the same degree. With the exception of the homogeneity requirement, the following constructions are now completely analogous to the afﬁne case: Deﬁnition 3.1.10. Let I ⊂ k[x0 , . . . , xn ] be a homogeneous ideal (or a set of homogeneous polynomials). The set Z(I) := {(a0 : · · · : an ) ∈ Pn ; f (a0 , . . . , an ) = 0 for all f ∈ I} is called the zero locus of I; this is well-deﬁned by remark 3.1.6. Subsets of Pn that are of the form Z(I) are called algebraic sets. If X ⊂ Pn is any subset, we call I(X) :=the ideal generated by { f ∈ k[x0 , . . . , xn ] homogeneous ; f (a0 , . . . , an ) = 0 for all (a0 : · · · : an ) ∈ X} ⊂ k[x0 , . . . , xn ] the ideal of X; by deﬁnition this is a homogeneous ideal. If we want to distinguish between the afﬁne zero locus Z(I) ⊂ An+1 and the projective zero locus Z(I) ⊂ Pn of the same (homogeneous) ideal, we denote the former by Za (I) and the latter by Z p (I). Remark 3.1.11. A remark that is sometimes useful is that every projective algebraic set can be written as the zero locus of ﬁnitely many homogeneous polynomials of the same d d degree. This follows easily from the fact that Z( f ) = Z(x0 f , . . . , xn f ) for all homogeneous polynomials f and every d ≥ 0. Example 3.1.12. Let L ⊂ An+1 be a linear subspace of dimension k + 1; it can be given by n − k linear equations in the coordinates of An+1 . The image of L under the quotient map (An+1 \{0})/(k\{0}) = Pn , i.e. the subspace of Pn given by the same n − k equations (now considered as equations in the homogeneous coordinates on Pn ) is called a linear subspace of Pn of dimension k. Once we have given projective varieties the structure of varieties, we will see that a linear subspace of Pn of dimension k is isomorphic to Pk . For

38

Andreas Gathmann

example, a line in P3 (with homogeneous coordinates x0 , x1 , x2 , x3 ) is given by two linearly independent equations in the xi . One example is the line {x2 = x3 = 0} = {(a0 : a1 : 0 : 0) ; a0 , a1 ∈ k} ⊂ P3 , which is “obviously isomorphic” to P1 . Example 3.1.13. Consider the conics in A2

2 X1 = {x2 = x1 } and X2 = {x1 x2 = 1}

˜ of exercise 2.6.1. We want to “compactify” these conics to projective algebraic sets X1 , ˜2 in P2 . Note that for a projective algebraic set we need the deﬁning polynomials to be X homogeneous, which is not yet the case here. On the other hand, we have an additional coordinate x0 that you can think of as being 1 on the afﬁne space A2 ⊂ P2 . So it is obvious that we should make the deﬁning equations homogeneous by adding suitable powers of x0 : consider 2 2 ˜ ˜ X1 = {x0 x2 = x1 } and X2 = {x1 x2 = x0 } ˜ in P2 . Then the restriction of Xi to the afﬁne space A2 ⊂ P2 is just given by Xi for i = 1, 2. ˜i the projective completion of Xi ; it can be done in the same way for all afﬁne We call X varieties (see exercise 3.5.3). ˜ Let us consider X1 ﬁrst. The points that we add at inﬁnity correspond to those where x0 = 0. It follows from the deﬁning equation that x1 = 0 as well; and then we must necessarily have x2 = 0 as the coordinates cannot be simultaneously zero. So there is only one point added at inﬁnity, namely (0 : 0 : 1). It corresponds to the “vertical direction” in 2 A2 , which is the direction in which the parabola x2 = x1 goes off to inﬁnity (at both ends actually). ˜ For X2 , the added points have again x0 = 0. This means that x1 x2 = 0, which yields the two points (0 : 1 : 0) and (0 : 0 : 1) in P2 : we added two points at inﬁnity, one corresponding to the “horizontal” and one to the “vertical” direction in A2 . This is clear from the picture below as the hyperbola x1 x2 = 1 extends to inﬁnity both along the x1 and the x2 axis.

x2 X1 x2 X2 x1 x1

˜ ˜ Note that the equations of X1 and X2 are exactly the same, up to a permutation of the coordinates. Even if we have not given projective varieties the structure of varieties yet, ˜ ˜ it should be obvious that X1 and X2 will be isomorphic varieties, with the isomorphism being given by exchanging x0 and x1 . Hence we see that the two distinct types of conics in A2 become the same in projective space: there is only one projective conic in P2 up to isomorphism. The difference in the afﬁne case comes from the fact that some conics “meet inﬁnity” in one point (like X1 ), and some in two (like X2 ). Proposition 3.1.14. (i) If I1 ⊂ I2 are homogeneous ideals in k[x0 , . . . , xn ] then Z(I2 ) ⊂ Z(I1 ). T S (ii) If {Ii } is a family of homogeneous ideals in k[x0 , . . . , xn ] then i Z(Ii ) = Z( i Ii ) ⊂ Pn . (iii) If I1 , I2 ⊂ k[x0 , . . . , xn ] are homogeneous ideals then Z(I1 ) ∪ Z(I2 ) = Z(I1 I2 ) ⊂ Pn .

3.

Projective varieties

39

In particular, arbitrary intersections and ﬁnite unions of algebraic sets are again algebraic sets. Proof. The proof is the same as in the afﬁne case (proposition 1.1.6). Deﬁnition 3.1.15. We deﬁne the Zariski topology on Pn to be the topology whose closed sets are the algebraic sets (proposition 3.1.14 tells us that this gives in fact a topology). Moreover, any subset X of Pn (in particular any algebraic set) will be equipped with the topology induced by the Zariski topology on Pn . This will be called the Zariski topology on X. Remark 3.1.16. The concepts of irreducibility and dimension introduced in section 1.3 are purely topological ones, so they apply to projective algebraic sets (or more generally to any subset of Pn ) as well. They have the same geometric interpretation as in the afﬁne case. Irreducible algebraic sets in Pn are called projective varieties. As in the afﬁne case (see lemma 1.3.4) a projective algebraic set X is irreducible if and only if its ideal I(X) is a prime ideal. In particular, Pn itself is irreducible. 3.2. Cones and the projective Nullstellensatz. We will now establish a correspondence between algebraic sets in Pn and homogeneous radical ideals in k[x0 , . . . , xn ], similar to the afﬁne case. This is quite straightforward; the only twist is that there is no zero point (0 : · · · : 0) in Pn , and so the zero locus of the radical homogeneous ideal (x0 , . . . , xn ) is empty although the ideal is not equal to (1). So we will have to exclude this ideal from our correspondence, which is why it is sometimes called the irrelevant ideal. As we want to use the results of the afﬁne case for the proof of this statement, let us ﬁrst establish a connection between projective algebraic sets in Pn and certain afﬁne algebraic sets in An+1 . Deﬁnition 3.2.1. An afﬁne algebraic set X ⊂ An+1 is called a cone if it is not empty, and if we have for all λ ∈ k (x0 , . . . , xn ) ∈ X ⇒ (λx0 , . . . , λxn ) ∈ X.

If X ⊂ Pn is a projective algebraic set, then C(X) := {(x0 , . . . , xn ) | (x0 : · · · : xn ) ∈ X} ∪ {0} is called the cone over X (obviously this is a cone). Remark 3.2.2. In other words, a cone is an algebraic set in An+1 that can be written as a (usually inﬁnite) union of lines through the origin. The cone over a projective algebraic set X ⊂ Pn is the inverse image of X under the projection map An+1 \{0} → (An+1 \{0})/(k\{0}) = Pn , together with the origin. Example 3.2.3. The following picture shows an example of a (two-dimensional) cone C(X) in A3 over the (one-dimensional) projective variety X in H = P2 :

x0 H X L2 L1

/A3

C(X)

40

Andreas Gathmann

(C(X) consists only of the “boundary” of the cone, not of the “interior”.) Note that C(X) contains the two lines L1 and L2 , which correspond to “points at inﬁnity” of the projective space P2 . Lemma 3.2.4. (i) Let I k[x0 , . . . , xn ] be a homogeneous ideal. If X = Z p (I) ⊂ Pn , then C(X) = Za (I) ⊂ An+1 . (ii) Conversely, if X ⊂ Pn is a projective algebraic set and I(X) ⊂ k[x0 , . . . , xn ] is its homogeneous ideal, then I(C(X)) = I(X). In other words, there is a one-to-one correspondence between projective algebraic sets in Pn and afﬁne cones in An+1 , given by taking the zero locus of the same homogeneous ideal (not equal to (1)) either in Pn or in An+1 . Proof. This is obvious from the deﬁnitions. Using this lemma, it is now very simple to derive a projective version of the Nullstellensatz: Proposition 3.2.5. (“The projective Nullstellensatz”) (i) If X1 ⊂ X2 are algebraic sets in Pn then I(X2 ) ⊂ I(X1 ). (ii) For any algebraic set X ⊂ Pn we have Z p (I(X)) = X. (iii) For any homogeneous ideal I ⊂ k[x0 , . . . , xn ] such that Z p (I) is not empty we have √ I(Z p (I)) = I. (iv) For any homogeneous ideal I ⊂ k[x0 , . . . , xn ] such that Z p (I) is empty we have √ either I = (1) or I = (x0 , . . . , xn ). In other words, Z p (I) is empty if and only if (x0 , . . . , xn )r ⊂ I for some r. Proof. The proofs of (i) and (ii) are literally the same as in the afﬁne case, see proposition 1.2.9. (iii): Let X = Z p (I). Then √ I(Z p (I)) = I(X) = I(C(X)) = I(Za (I)) = I by lemma 3.2.4 and the afﬁne Nullstellensatz of proposition 1.2.9 (iii). (iv): If Z p (I) is empty, then Za (I) is either empty or just the origin. So corollary 1.2.10 √ k tells us that I = (1) or I = (x0 , . . . , xn ). In any case, this means that xi i ∈ I for some ki , so (x0 , . . . , xn )k0 +···+kn ⊂ I. Theorem 3.2.6. There is a one-to-one inclusion-reversing correspondence between algebraic sets in Pn and homogeneous radical ideals in k[x0 , . . . , xn ] not equal to (x0 , . . . , xn ), given by the operations Z(·) and I(·). Proof. Immediately from proposition 3.2.5. 3.3. Projective varieties as ringed spaces. So far we have deﬁned projective varieties as topological spaces. Of course we want to make them into ringed spaces and ﬁnally show that they are varieties in the sense of deﬁnitions 2.4.1 and 2.5.1. So let X ⊂ Pn be a projective variety. First of all we have to make X into a ringed space whose structure sheaf is a sheaf of k-valued functions. The construction is completely analogous to the afﬁne case discussed in section 2.1. Deﬁnition 3.3.1. The ring S(X) := k[x0 , . . . , xn ]/I(X) is called the homogeneous coordinate ring of X.

3.

Projective varieties

41

Remark 3.3.2. In contrast to the afﬁne case, the elements of S(X) do not deﬁne functions on X, because the homogeneous coordinates are only determined up to a common scalar. Rather, to get well-deﬁned functions, we have to take quotients of two homogeneous polynomials of the same degree d, because then a rescaling of the homogeneous coordinates by a factor λ ∈ k\{0} gives a factor of λd in both the numerator and denominator, so that it cancels out: Deﬁnition 3.3.3. Let S(X)(d) := { f (d) ; f ∈ S(X)} be the degree-d part of S(X). Note that this is well-deﬁned: if f ∈ I(X) then f (d) = 0 by lemma 3.1.8. We deﬁne the ﬁeld of rational functions to be f ; f , g ∈ S(X)(d) and g = 0 . K(X) := g By remark 3.3.2, the elements of K(X) give set-theoretic functions to the ground ﬁeld k wherever the denominator is non-zero. Now as in the afﬁne case set \ f OX,P := ∈ K(X) ; g(P) = 0 and OX (U) := OX,P g P∈U for P ∈ X and U ⊂ X open. It is easily seen that this is a sheaf of k-valued functions. Remark 3.3.4. In the same way as for afﬁne varieties (see exercise 2.6.9) one can show that the function ﬁeld K(X) deﬁned above agrees with the deﬁnition for general varieties. Remark 3.3.5. Note that OX (X) = k, i.e. every regular function on all of X is constant. This follows trivially from the description of K(X): if the function is to be deﬁned everywhere g must be a constant. But then f has to be a constant too as it must have the same degree as g. A (slight) generalization of this will be proved in corollary 3.4.10. Proposition 3.3.6. Let X be a projective variety. Then (X, OX ) is a prevariety. Proof. We need to ﬁnd an open afﬁne cover of X. Consider the open subset X0 = {(a0 : · · · : an ) ∈ X ; a0 = 0} = X ∩ An (where An ⊂ Pn as in remark 3.1.4). If X = Z( f1 , . . . , fr ) with fi ∈ k[x0 , . . . , xn ] homogeneous, set gi (x1 , . . . , xn ) = fi (1, x1 , . . . , xn ) ∈ k[x1 , . . . , xn ] and deﬁne Y = Z(g1 , . . . , gr ) ⊂ An . We claim that there is an isomorphism F : X ∩ An → Y, (a0 : · · · : an ) → a1 an ,..., a0 a0 .

In fact, it is obvious that a set-theoretic inverse is given by F −1 : Y → X ∩ An , (a1 , . . . , an ) → (1 : a1 : · · · : an ). Moreover, F is a morphism because it pulls back a regular function on (an open subset of) Y of the form an p( a1 , . . . , a0 ) p(a1 , . . . , an ) a0 to , q(a1 , . . . , an ) q( a1 , . . . , an ) a0 a0 which is a regular function on X ∩ An as it can be rewritten as a quotient of two homogeneous polynomials of the same degree (by canceling the fractions in the numerator and denominator). In the same way, F −1 pulls back a regular function on (an open subset of) X ∩ An p(a0 , . . . , an ) p(1, a1 , . . . , an ) to , q(a0 , . . . , an ) q(1, a1 , . . . , an ) which is a regular function on Y . So F is an isomorphism.

42

Andreas Gathmann

In the same way we can do this for the open sets Xi = {(x0 : · · · : xn ) ∈ X ; xi = 0} for i = 0, . . . , n. As the xi cannot be simultaneously zero, it follows that the Xi form an afﬁne cover of X. So X is a prevariety. Remark 3.3.7. Following the proof of proposition 3.3.6, it is easy to see that our “new” deﬁnition of P1 agrees with the “old” deﬁnition of example 2.4.5 (i) by glueing two afﬁne lines A1 . Remark 3.3.8. Proposition 3.3.6 implies that all our constructions and results for prevarieties apply to projective varieties as well. For example, we know what morphisms are, and have deﬁned products of projective varieties. We have also deﬁned the ﬁeld of rational functions for prevarieties in exercise 2.6.9; it is easy to check that this deﬁnition agrees with the one in deﬁnition 3.3.3. Although this gives us the deﬁnition of morphisms and products, we would still have to apply our glueing techniques to write down a morphism or a product. So we should ﬁnd a better description for morphisms and products involving projective varieties: Lemma 3.3.9. Let X ⊂ Pn be a projective variety (or an open subset of a projective variety). Let f1 , . . . , fm ∈ k[x0 , . . . , xn ] be homogeneous polynomials of the same degree in the homogeneous coordinates of Pn , and assume that for every P ∈ X at least one of the fi does not vanish at P. Then the fi deﬁne a morphism f : X → Pm , P ∈ X → ( f0 (P) : · · · : fm (P)). Proof. First of all note that f is well-deﬁned set-theoretically: we have assumed that the image point can never be (0 : · · · : 0); and if we rescale the homogeneous coordinates xi we get ( f0 (λx0 : · · · : λxn ) : · · · : fm (λx0 : · · · : λxn )) = (λd f0 (x0 : · · · : xn ) : · · · : λd fm (x0 : · · · : xn )) = ( f0 (x0 : · · · : xn ) : · · · : fm (x0 : · · · : xn )), where d is the common degree of the fi . To check that f is a morphism, we want to use lemma 2.4.10, i.e. check the condition on an afﬁne open cover. So let {Vi } be the afﬁne open cover of Pm with Vi = {(y0 : · · · : ym ) ; yi = 0}, and let Ui = f −1 (Vi ). Then in the afﬁne f coordinates on Vi the map f |Ui is given by the quotients of polynomials fij for j = 0, . . . , n with j = i, hence gives a morphism as fi (P) = 0 on Ui . So f is a morphism by lemma 2.4.10. Remark 3.3.10. It should be noted however that not every morphism between projective varieties can be written in this form. The following example shows that this occurs already in quite simple cases. For a more precise statement see lemma 7.5.14. Example 3.3.11. By lemma 3.3.9, the map f : P1 → P2 , (s : t) → (x : y : z) = (s2 : st : t 2 ) is a morphism (as we must have s = 0 or t = 0 for every point of P1 , it follows that s2 = 0 or t 2 = 0; hence the image point is always well-deﬁned). Let X = f (P1 ) be the image of f . We claim that X is a projective variety with ideal I = (xz − y2 ). In fact, it is obvious that f (P1 ) ⊂ Z(I). Conversely, let P = (x : y : z) ∈ Z(I). As xz − y2 = 0 we must have x = 0 or z = 0; let us assume without loss of generality that x = 0. Then (x : y) ∈ P1 is a point that maps to (x2 : xy : y2 ) = (x2 : xy : xz) = (x : y : z). It is now easy to show that f : P1 → X is in fact an isomorphism: the inverse image −1 : X → P1 is given by f f −1 (x : y : z) = (x : y) and f −1 (x : y : z) = (y : z).

3.

Projective varieties

43

Note that at least one of the two points (x : y) and (y : z) is always well-deﬁned; and if they are both deﬁned they agree because of the equation xz = y2 . By lemma 3.3.9 both equations determine a morphism where they are well-deﬁned; so by lemma 2.4.10 they glue to give an inverse morphism f −1 . Note that f −1 is a (quite simple) morphism between projective varieties that cannot be written globally in the form of lemma 3.3.9. Summarizing, we have shown that f is an isomorphism: the curve {xz = y2 } ⊂ P2 is isomorphic to P1 . This example should be compared to exercise 2.6.1 and example 3.1.13. It is a special case of the Veronese embedding of 3.4.11. Finally, let us analyze the isomorphism f geometrically. Let Q = (1 : 0 : 0) ∈ X, and let L ⊂ P2 be the line {x = 0}. For any point P = (a : b : c) = Q there is a unique line PQ through P and Q with equation yc = zb. This line has a unique intersection point PQ ∩ L with the line L, namely (0 : b : c). If we identify L with P1 in the obvious way, we see that the above geometric construction gives us exactly f −1 (P) = PQ ∩ L. We say that f −1 is the projection from Q to L.

Q X P L

f −1 (P)

Example 3.3.12. Consider Pn with homogeneous coordinates x0 , . . . , xn , and Pm with homogeneous coordinates y0 , . . . , ym . We want to ﬁnd an easy description of the product Pn × Pm . Let PN = P(n+1)(m+1)−1 be projective space with homogeneous coordinates zi, j , 0 ≤ i ≤ n, 0 ≤ j ≤ m. There is an obviously well-deﬁned set-theoretic map f : Pn × Pm → PN given by zi, j = xi y j . Lemma 3.3.13. Let f : Pn × Pm → PN be the set-theoretic map as above. Then: (i) The image X = f (Pn × Pm ) is a projective variety in PN , with ideal generated by the homogeneous polynomials zi, j zi , j − zi, j zi , j for all 0 ≤ i, i ≤ n and 0 ≤ j, j ≤ m. (ii) The map f : Pn ×Pm → X is an isomorphism. In particular, Pn ×Pm is a projective variety. (iii) The closed subsets of Pn × Pm are exactly those subsets that can be written as the zero locus of polynomials in k[x0 , . . . , xn , y0 , . . . , ym ] that are bihomogeneous in the xi and yi . The map f is called the Segre embedding. Proof. (i): It is obvious that the points of f (Pn × Pm ) satisfy the given equations. Conversely, let P be a point in PN with coordinates zi, j that satisfy the given equations. At least one of these coordinates must be non-zero; we can assume without loss of generality that it is z0,0 . Let us pass to afﬁne coordinates by setting z0,0 = 1. Then we have zi, j = zi,0 z0, j ; so by setting xi = zi,0 and y j = z0, j we obtain a point of Pn × Pm that is mapped to P by f . (ii): Continuing the above notation, let P ∈ f (Pn × Pm ) be a point with z0,0 = 1. If f (xi , y j ) = P, it follows that x0 = 0 and y0 = 0, so we can assume x0 = 1 and y0 = 1 as the xi and y j are only determined up to a common scalar. But then it follows that xi = zi,0 and y j = z0, j ; i.e. f is bijective.

44

Andreas Gathmann

The same calculation shows that f and f −1 are given (locally in afﬁne coordinates) by polynomial maps; so f is an isomorphism. (iii): It follows by the isomorphism of (ii) that any closed subset of Pn × Pm is the zero locus of homogeneous polynomials in the zi, j , i.e. of bihomogeneous polynomials in the xi and y j (of the same degree). Conversely, a zero locus of bihomogeneous polynomials can always be rewritten as a zero locus of bihomogeneous polynomials of the same degree in the xi and yi by remark 3.1.11. But such a polynomial is obviously a polynomial in the zi, j , so it determines an algebraic set in X ∼ Pn × Pm . = Example 3.3.14. By lemma 3.3.13, P1 × P1 is (isomorphic to) the quadric surface X = {(z0,0 : z0,1 : z1,0 : z1,1 ) ; z0,0 z1,1 = z1,0 z0,1 } ⊂ P3 . by the isomorphism P1 × P1 → X, ((x0 : x1 ), (y0 : y1 )) → (x0 y0 : x0 y1 : x1 y0 : x1 y1 ). In particular, the “lines” P1 × P and P × P1 in P1 × P1 where the ﬁrst or second factor is constant are mapped to lines in X ⊂ P3 . We can see these two families of lines on the quadric surface X:

IP 1

≅ IP 1

X in IP 3

Corollary 3.3.15. Every projective variety is a variety. Proof. We have already seen in proposition 3.3.6 that every projective variety is a prevariety, so by lemma 2.5.3 and lemma 2.5.4 it only remains to be shown that the diagonal ∆(Pn ) ⊂ Pn × Pn is closed. We can describe this diagonal as ∆(Pn ) = {((x0 : · · · : xn ), (y0 : · · · : yn )) ; xi y j − x j yi = 0 for all i, j}, because these equations mean exactly that the matrix x0 y0 x1 y1 ··· ··· xn yn

has rank (at most 1), i.e. that (x0 : · · · : xn ) = (y0 : · · · : yn ). In particular, it follows by lemma 3.3.13 (iii) that ∆(Pn ) ⊂ Pn × Pn is closed. 3.4. The main theorem on projective varieties. The most important property of projective varieties is that they are compact in the classical topology (if the ground ﬁeld is k = C). We have seen this already for projective spaces in remark 3.1.5, and it then follows for projective algebraic sets as well as they are closed subsets (even in the classical topology) of the compact projective spaces. Unfortunately, the standard deﬁnition of compactness does not make sense at all in the Zariski topology, so we need to ﬁnd an alternative description. One property of compact sets is that they are mapped to compact sets under continuous maps. In our language, this would mean that images of projective varieties under a morphism should be closed. This is what we want to prove.

3.

Projective varieties

45

Remark 3.4.1. Note ﬁrst that this property deﬁnitely does not hold for afﬁne varieties: consider e.g. the afﬁne variety X = {(x, y) ; xy = 1} ⊂ A2 and the projection morphism f : X → A1 , (x, y) → x. The image of f is A1 \{0}, which is not closed in A1 . In fact, we can see from example 3.1.13 why it is not closed: the “vertical point at inﬁnity”, which would map to x = 0 ∈ A1 and make the image closed, is missing in the afﬁne variety X.

y X

x f (X) = A1 \{0}

To prove the above mentioned statement we start with a special case (from which the general one will follow easily). Theorem 3.4.2. The projection map π : Pn × Pm → Pn is closed, i.e. if X ⊂ Pn × Pm is closed then so is π(X). Proof. Let X ⊂ Pn × Pm be an algebraic set. By lemma 3.3.13 (iii) we can write X as the zero locus of polynomials f1 (x, y), . . . , fr (x, y) bihomogeneous in the coordinates xi of Pn and yi of Pm (where we use the short-hand notation fi (x, y) for fi (x0 , . . . , xn , y0 , . . . , ym )). By remark 3.1.11 we may assume that all fi have the same degree d in the yi . Let P ∈ Pn be a ﬁxed point. Then P ∈ π(X) if and only if the common zero locus of the polynomials fi (P, y) in y is non-empty in Pm , which by proposition 3.2.5 is the case if and only if (y0 , . . . , ym )s ⊂ ( f1 (P, y), . . . , fr (P, y)) (∗) for all s ≥ 0. As (∗) is obvious for s < d, it sufﬁces to show that for any s ≥ d, the set of all P ∈ Pn satisfying (∗) is closed, as π(X) will then be the intersection of all these sets and therefore closed as well. Note that the ideal (y0 , . . . , ym )s is generated by the m+s monomials of degree s in the m yi , which we denote by Mi (y) (in any order). Hence (∗) is not satisﬁed if and only if there are polynomials gi, j (y) such that Mi (y) = ∑ j gi, j (y) f j (P, y) for all i. As the Mi and f j are homogeneous of degree s and d, respectively, this is the same as saying that such relations exist with the gi, j homogeneous of degree s − d. But if we let Ni (y) be the collection of all monomials in the yi of degree s − d, this is in turn equivalent to saying that the collection of polynomials {Ni (y) f j (P, y) ; 1 ≤ i ≤ m+s−d , 1 ≤ j ≤ r} generates the whole vector m space of polynomials of degree s. Writing the coefﬁcients of these polynomials in a matrix A = As (P), this amounts to saying that A has rank (at least) m+s . Hence (∗) is satisﬁed m if and only if all minors of A of size m+s vanish. But as the entries of the matrix A are m homogeneous polynomials in the coefﬁcients of P, it follows that the set of all P satisfying (∗) is closed. Remark 3.4.3. Let us look at theorem 3.4.2 from an algebraic viewpoint. We start with some equations fi (x, y) and ask for the image of the projection map (x, y) → x, which can be written as {x ; there is a y such that fi (x, y) = 0 for all i}.

46

Andreas Gathmann

In other words, we are trying to eliminate the variables y from the system of equations fi (x, y) = 0. The statement of the theorem is that the set of all such x can itself be written as the solution set of some polynomial equations. This is sometimes called the main theorem of elimination theory. Corollary 3.4.4. The projection map π : Pn ×Y → Y is closed for any variety Y . Proof. Let us ﬁrst show the statement for Y ⊂ Am being an afﬁne variety. Then we can regard Y as a subspace of Pm via the embedding Am ⊂ Pm (Y is neither open nor closed ˜ in Pm , but that does not matter). Now if Z ⊂ Pn × Y is closed, let Z ⊂ Pn × Pm be the ˜ projective closure. By theorem 3.4.2, π(Z) is closed in Pm , where π is the projection morphism. Therefore ˜ ˜ π(Z) = π(Z ∩ (Pn ×Y )) = π(Z) ∩Y is closed in Y . If Y is any variety we can cover it by afﬁne open subsets. As the condition that a subset is closed can be checked by restricting it to the elements of an open cover, the statement follows from the corresponding one for the afﬁne open patches that we have just shown. Remark 3.4.5. Corollary 3.4.4 is in fact the property of Pn that captures the idea of compactness (as we will see in corollary 3.4.7). Let us therefore give it a name: we say that a variety X is complete if the projection map π : X × Y → Y is closed for every variety Y . (You can think of the name “complete” as coming from the geometric idea that it contains all the “points at inﬁnity” that were missing in afﬁne varieties.) So corollary 3.4.4 says that Pn is complete. Moreover, any projective variety Z ⊂ Pn is complete, because any closed set in Z ×Y is also closed in Pn ×Y , so its image under the projection morphism to Y will be closed as well. Remark 3.4.6. We have just seen that every projective variety is complete. In fact, whereas the converse of this statement is not true, it is quite hard to write down an example of a complete variety that is not projective. We will certainly not meet such an example in the near future. So for practical purposes you can usually assume that the terms “projective variety” and “complete variety” are synonymous. Corollary 3.4.7. Let f : X → Y be a morphism of varieties, and assume that X is complete. Then the image f (X) ⊂ Y is closed. Proof. We factor f as f : X → X × Y → Y , where Γ = (idX , f ) (the so-called graph morphism), and π is the projection to Y . We claim that Γ(X) = {(P, f (P)) ; P ∈ X} ⊂ X ×Y is closed. To see this, note ﬁrst that the diagonal ∆(Y ) ⊂ Y ×Y is closed as Y is a variety. Now Γ(X) is just the inverse image of ∆(Y ) under the morphism ( f , idY ) : X ×Y → Y ×Y , and is therefore also closed. As X is complete, it follows that f (X) = π(Γ(X)) is closed. Corollary 3.4.8. Let X ⊂ Pn be a projective variety that contains more than one point, and / let f ∈ k[x0 , . . . , xn ] be a non-constant homogeneous polynomial. Then X ∩ Z( f ) = 0. Proof. Assume that the statement is false, i.e. that f is non-zero on all of X. Let P, Q ∈ X be two distinct points of X and choose a homogeneous polynomial g ∈ k[x0 , . . . , xn ] of the same degree as f such that g(P) = 0 and g(Q) = 0. Let F : X → P1 be the morphism deﬁned by P → ( f (P) : g(P)); this is well-deﬁned as f is non-zero on X by assumption. By corollary 3.4.7 the image F(X) is closed in P1 . Moreover, F(X) is irreducible as X is. Therefore, F(X) is either a point or all of P1 . But by assumption (0 : 1) ∈ F(X), so / F(X) must be a single point. But this is a contradiction, as F(P) = ( f (P) : g(P)) = (1 : 0) and F(Q) = ( f (Q) : g(Q)) = (1 : 0) by the choice of g.

Γ π

3.

Projective varieties

47

Remark 3.4.9. Again this statement is false for afﬁne varieties: consider e.g. X = {x = / 0} ⊂ A2 and f = x − 1, then X ∩ Z( f ) = 0 although X is a line (and therefore contains more than one point). This example worked because in A2 we can have parallel lines. In P2 such lines would meet at inﬁnity, so the intersection would be non-empty then. Corollary 3.4.10. Every regular function on a complete variety is constant. Proof. Let f : X → A1 be a regular function on a complete variety X. Consider f as a morphism to P1 that does not assume the value ∞. In particular, f (X) P1 , hence it is a single point by corollary 3.4.7. Example 3.4.11. (This is a generalization of example 3.3.11 and exercise 3.5.2.) Let fi (x0 , . . . , xn ), 0 ≤ i ≤ N = n+d − 1 be the set of all monomials in k[x0 , . . . , xn ] of degree n i i d, i.e. of the monomials of the form x00 · · · xnn with i0 + · · · + in = d. Consider the map F : Pn → PN , (x0 : · · · : xn ) → ( f0 : · · · : fN ).

d d By lemma 3.3.9 this is a morphism (note that the monomials x0 , . . . , xn , which cannot be simultaneously zero, are among the fi ). So by corollary 3.4.7 the image X = F(Pn ) is a projective variety. We claim that F : X → F(X) is an isomorphism. All we have to do to prove this is to ﬁnd an inverse morphism. This is not hard: we can do this on an afﬁne open cover, so d let us consider the open subset where x0 = 0 (and therefore x0 = 0). We can then pass to

afﬁne coordinates and set x0 = 1. The inverse morphism is then given by xi =

1 ≤ 1 ≤ n. The morphism F is therefore an isomorphism and thus realizes Pn as a subvariety of PN . This is usually called the degree-d Veronese embedding. Its importance lies in the fact that degree-d polynomials in the coordinates of Pn are translated into linear polynomials when viewing Pn as a subvariety of PN . An example of this application will be given in corollary 3.4.12. The easiest examples are the degree-d embeddings of P1 , given by P1 → Pd , (s : t) → (sd : sd−1t : sd−2t 2 : · · · : t d ). The special cases d = 2 and d = 3 are considered in example 3.3.11 and exercise 3.5.2. Note that by applying corollary 3.4.7 we could conclude that F(X) is a projective variety without writing down its equations. Of course, in theory we could also write down the equations, but this is quite messy in this case. Corollary 3.4.12. Let X ⊂ Pn be a projective variety, and let f ∈ k[x0 , . . . , xn ] be a nonconstant homogeneous polynomial. Then X\Z( f ) is an afﬁne variety. Proof. We know this already if f is a linear polynomial (see the proof of proposition 3.3.6). But by applying a Veronese embedding of degree d, we can always assume this. 3.5. Exercises. Exercise 3.5.1. Let L1 and L2 be two disjoint lines in P3 , and let P ∈ P3 \(L1 ∪ L2 ) be a point. Show that there is a unique line L ⊂ P3 meeting L1 , L2 , and P (i.e. such that P ∈ L / and L ∩ Li = 0 for i = 1, 2). Exercise 3.5.2. Let C ⊂ P3 be the “twisted cubic curve” given by the parametrization P1 → P3 (s : t) → (x : y : z : w) = (s3 : s2t : st 2 : t 3 ). Let P = (0 : 0 : 1 : 0) ∈ P3 , and let H be the hyperplane deﬁned by z = 0. Let ϕ be the projection from P to H, i.e. the map associating to a point Q of C the intersection point of the unique line through P and Q with H.

d−1 xi x0 d x0

for

48

Andreas Gathmann

(i) Show that ϕ is a morphism. (ii) Determine the equation of the curve ϕ(C) in H ∼ P2 . = (iii) Is ϕ : C → ϕ(C) an isomorphism onto its image? Exercise 3.5.3. Let I ⊂ k[x1 , . . . , xn ] be an ideal. Deﬁne I h to be the ideal generated by { f h ; f ∈ I} ⊂ k[x0 , . . . , xn ], where f h (x0 , . . . , xn ) := x0

deg( f )

·f

x1 xn ,..., x0 x0

denotes the homogenization of f with respect to x0 . Show that: (i) I h is a homogeneous ideal. (ii) Z(I h ) ⊂ Pn is the closure of Z(I) ⊂ An in Pn . We call Z(I h ) the projective closure of Z(I). h h (iii) Let I = ( f1 , . . . , fk ). Show by an example that I h = ( f1 , . . . , fk ) in general. (Hint: You may consider (again) the twisted cubic curve of exercise 3.5.2.) Exercise 3.5.4. In this exercise we will make the space of all lines in Pn into a projective variety. Fix n ≥ 1. We deﬁne a set-theoretic map ϕ : {lines in Pn } → PN with N = n+1 − 1 as follows. For every line L ⊂ Pn choose two distinct points P = 2 (a0 : · · · : an ) and Q = (b0 : · · · : bn ) on L and deﬁne ϕ(L) to be the point in PN whose homogeneous coordinates are the n+1 maximal minors of the matrix 2 a0 b0 in any ﬁxed order. Show that: (i) The map ϕ is well-deﬁned and injective. (ii) The image of ϕ is a projective variety that has a ﬁnite cover by afﬁne spaces A2(n−1) (in particular, its dimension is 2(n − 1)). It is called the Grassmannian G(1, n). Hint: recall that by the Gaussian algorithm most matrices (what does this mean?) are equivalent to one of the form 1 0 0 a2 1 b2 ··· ··· an bn ··· ··· an bn ,

for some ai , bi . (iii) G(1, 1) is a point, G(1, 2) ∼ P2 , and G(1, 3) is the zero locus of a quadratic equa= tion in P5 . Exercise 3.5.5. Let V be the vector space over k of homogeneous degree-2 polynomials in three variables x0 , x1 , x2 , and let P(V ) ∼ P5 be its projectivization. = (i) Show that the space of conics in P2 can be identiﬁed with an open subset U of P5 . (One says that U is a “moduli space” for conics in P2 and that P5 is a “compactiﬁed moduli space”.) What geometric objects can be associated to the points in P5 \U? (ii) Show that it is a linear condition in P5 for the conics to pass through a given point in P2 . More precisely, if P ∈ P2 is a point, show that there is a linear subspace L ⊂ P5 such that the conics passing through P are exactly those in U ∩ L. What happens in P5 \U, i.e. what do the points in (P5 \U) ∩ L correspond to? (iii) Prove that there is a unique conic through any ﬁve given points in P2 , as long as no three of them lie on a line. What happens if three of them do lie on a line?

3.

Projective varieties

49

Exercise 3.5.6. Show that an afﬁne variety over C is never compact in the classical topology unless it is a single point. (Hint: Given an afﬁne variety X ⊂ An , show that the image of X under the projection map An → A1 onto the ﬁrst coordinate is either a point or an open subset (in the Zariski topology) of A1 . Conclude that an afﬁne variety with more than one point is never bounded, i.e. is never contained in a ball {(z1 , . . . , zn ) ; |z1 |2 + · · · + |zn |2 ≤ R2 } ⊂ Cn , and therefore not compact.) Exercise 3.5.7. Let G(1, n) be the Grassmannian of lines in Pn as in exercise 3.5.4. Show that: (i) The set {(L, P) ; P ∈ L} ⊂ G(1, n) × Pn is closed. (ii) If Z ⊂ G(1, n) is any closed subset then the union of all lines L ⊂ Pn such that L ∈ Z is closed in Pn . (iii) Let X,Y ⊂ Pn be disjoint projective varieties. Then the union of all lines in Pn intersecting X and Y is a closed subset of Pn . It is called the join J(X,Y ) of X and Y . Exercise 3.5.8. Recall that a conic is a curve in P2 that can be given as the zero locus of an irreducible homogeneous polynomial f ∈ k[x0 , x1 , x2 ] of degree 2. Show that for any 5 given points P1 , . . . , P5 ∈ P2 in general position, there is a unique conic passing through all the Pi . This means: there is a non-empty open subset U ⊂ P2 × · · · × P2 such that there is a unique conic through the Pi whenever (P1 , . . . , P5 ) ∈ U. (Hint: By mapping a conic 2 2 2 {a0 x0 + a1 x1 + a2 x2 + a3 x0 x1 + a4 x0 x2 + a5 x1 x2 = 0} to the point (a0 : · · · : a5 ) ∈ P5 , you can think of “the space of all conics” as an open subset of P5 .)

50

Andreas Gathmann

4. D IMENSION

We have already introduced the concept of dimension of a variety. Now we develop some methods that allow to compute the dimension of most varieties rigorously. We show that the dimension of An and Pn is n. The dimension of a variety equals the dimension of any of its non-empty open subsets. Every irreducible component of the zero locus of a single function on an afﬁne or projective variety X has dimension dim X − 1. Two varieties are called birational if they contain isomorphic open subsets. As a large class of examples of birational varieties we construct the blow-up of an afﬁne variety in a subvariety or an ideal. We study in detail the case of blowing up a single point P in a variety X. In this case, the exceptional hypersurface is the tangent cone CX,P . For any point P in a variety X, the tangent space TX,P is the linear space dual to M/M 2 , where M ⊂ OX,P is the maximal ideal. The point P is called a smooth point of X if TX,P = CX,P , i.e. if X “can be approximated linearly” around P. Smoothness can easily be checked by the Jacobi criterion. As an application of the theory developed so far, we show that every smooth cubic surface X has exactly 27 lines on it. We study the conﬁguration of these lines, and show that X is isomorphic to P2 blown up in 6 suitably chosen points.

4.1. The dimension of projective varieties. Recall that in section 1.3 we have introduced the notion of dimension for every (Noetherian) topological space, in particular for every variety X: the dimension dim X of X is the largest integer n such that there is a chain of irreducible closed subsets of X / 0 = X0 X1 ··· Xn = X. For simplicity of notation, in what follows we will call this a longest chain in X. While this deﬁnition is quite simple to write down, it is very difﬁcult to use in practice. In fact, we have not even been able yet to compute the dimensions of quite simple varieties like An or Pn (although it is intuitively clear that these spaces should have dimension n). In this section, we will develop techniques that allow us to compute the dimensions of varieties rigorously. Remark 4.1.1. We will start our dimension computations by considering projective varieties. It should be said clearly that the theory of dimension is in no way special or easier for projective varieties than it is for other varieties — in fact, it should be intuitively clear that the dimension of a variety is essentially a local concept that can be computed in the neighborhood of any point. The reason for us to start with projective varieties is simply that we know more about them: the main theorem on projective varieties and its corollaries of section 3.4 are so strong that they allow for quite efﬁcient applications in dimension theory. One could as well start by looking at the dimensions of afﬁne varieties (and most textbooks will do so), but this requires quite some background in (commutative) algebra that we do not have yet. Remark 4.1.2. The main idea for our dimension computations is to compare the dimensions of varieties that are related by morphisms with various properties. For example, if f : X → Y is a surjective morphism, we would expect that dim X ≥ dimY . If f : X → Y is a morphism with ﬁnite ﬁbers, i.e. such that f −1 (P) is a ﬁnite set for all P ∈ Y , we would expect that dim X ≤ dimY . In particular, if a morphism both is surjective and has ﬁnite ﬁbers, we expect that dim X = dimY . Example 4.1.3. The standard case in which we will prove and apply the idea of comparing dimensions is the case of projections from a point. We have already seen such projections in example 3.3.11 and exercise 3.5.2; let us now consider the general case.

4.

Dimension

51

Let X Pn be a projective variety, and let P ∈ Pn be a point that is not in X. By a change of coordinates we can assume that P = (0 : · · · : 0 : 1). Let H ∼ Pn−1 ⊂ Pn be a = linear subspace of codimension 1 that does not contain P; again by a change of coordinates we can assume that H = {xn = 0}. We deﬁne a projection map π : X → H from P as follows: for every point Q ∈ X let π(Q) be the intersection point of the line PQ with H. (Note that this is well-deﬁned as Q = P by assumption.)

IP n P

Q

π (Q)

H ≅ IP n −1

This is in fact a morphism: if Q = (a0 : · · · : an ) ∈ X, the line PQ is given parametrically by PQ = {(λa0 : · · · : λan−1 : λan + µ) ∈ Pn ; (λ : µ) ∈ P1 }. The intersection point of this line with H is obviously the point (a0 : · · · : an−1 : 0), which is well-deﬁned by the assumption that Q = P. Hence the projection π is given in coordinates by π : X → Pn−1 , (a0 : · · · : an ) → (a0 : · · · : an−1 ). In particular, this is a polynomial map and therefore a morphism. Note that projections always have ﬁnite ﬁbers: by construction, the inverse image π−1 (Q) of a point Q ∈ H must be contained in the line PQ ∼ P1 , but it must also be = an algebraic set and cannot contain the point P, hence it must be a ﬁnite set. Note also that we can repeat this process if the image of X is not all of Pn−1 : we can then project π(X) from a point in Pn−1 to Pn−2 , and so on. After a ﬁnite number of such projections, we arrive at a surjective morphism X → Pm for some m that is the composition of projections as above. In particular, as this morphism is surjective and has ﬁnite ﬁbers, we expect dim X = m. This is the idea that we will use for our dimension computations. Let us start with some statements about dimensions that are not only intuitively clear but actually also easy to prove. Lemma 4.1.4. / (i) If 0 = X0 · · · Xn = X is a longest chain in X then dim Xi = i for all i. (ii) If Y X is a closed subvariety of the variety X then dimY < dim X. (iii) Let f : X → Y be a surjective morphism of projective varieties. Then every longest / / chain 0 = Y0 · · · Yn in Y can be lifted to a chain 0 = X0 · · · Xn in X (i.e. the Xi are closed and irreducible with f (Xi ) = Yi for all i). In particular, dim X ≥ dimY . Proof. (i): It is obvious that dim Xi ≥ i. If we had dim Xi > i there would be a longer chain / in Xi than 0 = X0 · · · Xi . This chain could then be extended by the X j for j > i to a chain in X that is longer than the given one. / / (ii): We can extend a longest chain 0 = Y0 Y1 · · · Yn = Y in Y to a chain 0 = Y0 Y1 · · · Yn = Y X in X which is one element longer. (iii): We prove the statement by induction on n = dimY ; there is nothing to show if n = 0. Otherwise let Z1 , . . . , Zr ⊂ X be the irreducible components of f −1 (Yn−1 ), so that f (Z1 )∪ · · · ∪ f (Zr ) = Yn−1 . Note that Yn−1 is irreducible and the f (Zi ) are closed by corollary

52

Andreas Gathmann

3.4.7, so one Zi must map surjectively to Yn−1 . Applying the induction hypothesis to the / restriction f |Zi : Zi → Yn−1 we get dim Zi ≥ dimYn−1 = n − 1, so there is a chain 0 = X0 · · · Xn−1 = Zi . Extending this chain by X at the end, we thus obtain a chain in X of length n lying over the given chain in Y . Lemma 4.1.5. Let X Pn be a projective variety, and assume without loss of generality that P = (0 : · · · : 0 : 1) ∈ X. / (i) Any homogeneous polynomial f ∈ k[x0 , . . . , xn ] satisﬁes a relation of the form f D + a1 f D−1 + a2 f D−2 + · · · + aD = 0 in S(X) = k[x0 , . . . , xn ]/I(X) for some D > 0 and some homogeneous polynomials ai ∈ k[x0 , . . . , xn−1 ] that do not depend on the last variable xn . (ii) Let π : X → Pn−1 be the projection from P as in example 4.1.3. If Y ⊂ X is a closed subvariety such that π(Y ) = π(X) then Y = X. Remark 4.1.6. Before we prove this lemma let us give the idea behind these statements. In (i), you should think of f as being a polynomial containing the variable xn , while the ai do not. So for given values of x0 , . . . , xn−1 the relation in (i) is a non-zero polynomial equation in xn that therefore allows only ﬁnitely many values for xn on X. As the projection from P is just given by dropping the last coordinate xn , the statement of (i) is just that this projection map has ﬁnite ﬁbers. We have argued in remark 4.1.1 that we then expect the dimension of π(X) to be less than or equal to the dimension of X. To show this we will want to take a longest chain in X and project it down to π(X). It is obvious that the images of the elements of such a chain in X are again closed subvarieties in π(X), but it is not a priori obvious that a strict inclusion Xi Xi+1 translates into a strict inclusion π(Xi ) π(Xi+1 ). This is exactly the statement of (ii). Proof. (i): Let d be the degree of f . Consider the morphism

d d ˜ π : X → Pn , (x0 : · · · : xn ) → (y0 : · · · : yn ) := (x0 : · · · : xn−1 : f (x0 , . . . , xn ))

˜ (which is well-deﬁned since P ∈ X). The image of π is closed by corollary 3.4.7 and is / therefore the zero locus of some homogeneous polynomials F1 , . . . , Fr ∈ k[y0 , . . . , yn ]. Note that / Z(y0 , . . . , yn−1 , F1 , . . . , Fr ) = 0 ⊂ Pn ˜ because the Fi require the point to be in the image π(X), while the x0 , . . . , xn−1 do not vanish simultaneously on X. So by the projective Nullstellensatz of proposition 3.2.5 (iv) it follows that some power of yn is in the ideal generated by y0 , . . . , yn−1 , F1 , . . . , Fr . In other words, yD = n

n−1 i=0

∑ gi (y0 , . . . , yn ) · yi

˜ in S(π(X)) = k[y0 , . . . , yn ]/(F1 , . . . , Fr )

˜ for some D. Substituting the deﬁnition of π for the yi thus shows that there is a relation f D + a1 f D−1 + a2 f D−2 + · · · + aD = 0 in S(X)

for some homogeneous ai ∈ k[x0 , . . . , xn−1 ]. (ii): Assume that the statement is false, i.e. that Y X. Then we can pick a homogeneous polynomial f ∈ I(Y )\I(X) ⊂ k[x0 , . . . , xn ] of some degree d that vanishes on Y but not on X. Now pick a relation as in (i) for the smallest possible value of D. In particular we then have aD = 0 in S(X), i.e. aD ∈ I(X). But we have chosen f such that f ∈ I(Y ), therefore / the relation (i) tells us that aD ∈ I(Y ) as well.

4.

Dimension

53

It follows that aD ∈ I(Y )\I(X). But note that aD ∈ k[x0 , . . . , xn−1 ], so aD is a function on Pn−1 that vanishes on π(Y ) but not on π(X), in contradiction to the assumption. Corollary 4.1.7. Let X Pn be a projective variety, and assume without loss of generality that P = (0 : · · · : 0 : 1) ∈ X. Let π : X → Pn−1 be the projection from P as in example 4.1.3. / Then dim X = dim π(X). / / Proof. Let 0 = X0 · · · Xr = X be a longest chain in X. Then 0 = Y0 · · · Yr = Y with Yi = π(Xi ) is a chain in π(X): note that the Yi are closed by corollary 3.4.7, irreducible as they are the images of irreducible sets, and no two of them can coincide by lemma 4.1.5. It follows that dim π(X) ≥ dim X. But also dim π(X) ≤ dim X by lemma 4.1.4 (iii), so the statement follows. Corollary 4.1.8. The dimension of Pn is n. Proof. By lemma 4.1.4 (ii) we know that dim P0 < dim P1 < dim P2 < dim P3 < · · · . (∗) Moreover, we have seen in example 4.1.3 that every projective variety X can be mapped surjectively to some Pn by a sequence of projections from points; it then follows that dim X = dim Pn by corollary 4.1.7. In other words, every dimension that occurs as the dimension of some projective variety must occur already as the dimension of some projective space. But combining (∗) with lemma 4.1.4 (i) we see that every non-negative integer occurs as the dimension of some projective variety — and therefore as the dimension of some projective space. So in (∗) we must have dim Pn = n for all n. Proposition 4.1.9. Let X ⊂ Pn be a projective variety, and let f ∈ k[x0 , . . . , xn ] be a nonconstant homogeneous polynomial that does not vanish identically on X. Then dim(X ∩ Z( f )) = dim X − 1. Remark 4.1.10. Note that in the statement of this proposition X ∩ Z( f ) may well be reducible; the statement is then that there is at least one component that has dimension dim X − 1 (and that no component has bigger dimension). We will prove a stronger statement, namely a statement about every component of X ∩ Z( f ), in corollary 4.2.5. Proof. Let m = dim X. After applying a Veronese embedding of degree deg f as in example 3.4.11 we can assume that f is linear. Now construct linear functions f0 , . . . , fm and algebraic sets X0 , . . . , Xm+1 ⊂ X inductively as follows: Let X0 = X and f0 = f . For i ≥ 0 let Xi+1 = Xi ∩ Z( fi ), and let fi+1 be any linear form such that (i) fi+1 does not vanish identically on any component of Xi+1 , and (ii) fi+1 is linearly independent from the f1 , . . . , fi . It is obvious that (i) can always be satisﬁed. Moreover, (ii) is automatic if Xi+1 is not empty (as f1 , . . . , fi vanish on Xi+1 ), and easy to satisfy otherwise (as then (i) is no condition). Applying lemma 4.1.4 (ii) inductively, we see that no component of Xi has dimension bigger than m − i. In particular, Xm+1 must be empty. Hence the linear forms f0 , . . . , fm do not vanish simultaneously on X; so they deﬁne a morphism π : X → Pm . As the fi are linear and linearly independent, π is up to a change of coordinates the same as fi = xi for 0 ≤ i ≤ m, so it is just a special case of a continued projection from points as in example 4.1.3. In particular, dim π(X) = dim X = m by corollary 4.1.7. By lemma 4.1.4 (ii) it then follows that π(X) = Pm , i.e. π is surjective. Now suppose that every component of X1 = X ∩ Z( f ) has already dimension at most m−2, then by the above inductive argument already Xm is empty and the forms f0 , . . . , fm−1 do not vanish simultaneously on X. But this means that (0 : · · · : 0 : 1) ∈ π(X), which / contradicts the surjectivity of π.

54

Andreas Gathmann

4.2. The dimension of varieties. After having exploited the main theorem on projective varieties as far as possible, let us now study the dimension of more general varieties. We have already remarked that the dimension of a variety should be a local concept; in particular the dimension of any open subvariety U of a variety X should be the same as that of X. This is what we want to prove ﬁrst. Proposition 4.2.1. Let X be a variety, and let U ⊂ X be a non-empty open subset of X. Then dimU = dim X. / Proof. “≤”: Let 0 = U0 U1 · · · Un = U be a longest chain in U. If Xi denotes the / closure of Ui in X for all i, then 0 = X0 · · · Xn = X is a chain in X. “≥”: We will prove this in several steps. / Step 1: Let 0 = X0 · · · Xn = X be a longest chain in X, and assume that X0 ⊂ U. / Then set Ui = Xi ∩U for all i; we claim that 0 = U0 · · · Un = U is a chain in U (from which it then follows that dimU ≥ dim X). In fact, the only statement that is not obvious here is that Ui = Ui+1 for all i. So assume that Ui = Ui+1 for some i. Then Xi+1 = (Xi+1 ∩U) ∪ (Xi+1 ∩ (X\U)) = (Xi ∩U) ∪ (Xi+1 ∩ (X\U)) = Xi ∪ (Xi+1 ∩ (X\U)), where the last equality follows from Xi ∩(X\U) ⊂ Xi+1 ∩(X\U). But this is a contradiction to Xi+1 being irreducible, as Xi is neither empty nor all of Xi+1 . So we have now proven the proposition in the case where the element X0 of a longest chain in X lies in U. Step 2: Let X be a projective variety. Then we claim that we can always ﬁnd a longest / chain 0 = X0 · · · Xn (with n = dim X) such that X0 ⊂ U. We will construct this chain by descending recursion on n, starting by setting Xn = X. So assume that Xi Xi+1 / · · · Xn = X has already been constructed such that Xi ∩ U = 0. Pick any non-constant homogeneous polynomial f that does not vanish identically on any irreducible component of Xi \U. By proposition 4.1.9 there is a component of Xi ∩ Z( f ) of dimension i − 1; call / this Xi−1 . We have to show that Xi−1 ∩ U = 0. Assume the contrary; then Xi−1 must be contained in Xi \U. But by the choice of f we know that Xi−1 is not a whole component of Xi \U, so it can only be a proper subset of a component of Xi \U. But by lemma 4.1.4 (ii) the components of Xi \U have dimension at most i − 1, and therefore proper subsets of them have dimension at most i − 2. This is a contradiction to dim Xi−1 = i − 1. Combining steps 1 and 2, we have now proven the proposition if X is a projective va¯ riety. Of course the statement then also follows if X is an afﬁne variety: let X be the projective closure of X as in exercise 3.5.3, then by applying our result twice we get ¯ dimU = dim X = dim X. / Step 3: Let X be any variety, and let 0 = X0 · · · Xn = X be a longest chain in X. Let V ⊂ X be an afﬁne open neighborhood of the point X0 ; then dimV = dim X by step 1. In the same way we can ﬁnd an afﬁne open subset W of U such that dimW = dimU. As / V ∩W = 0, it ﬁnally follows from steps 1 and 2 that dim X = dimV = dim(V ∩W ) = dimW = dimU.

In particular, as every variety can be covered by afﬁne varieties, this proposition implies that it is sufﬁcient to study the dimensions of afﬁne varieties. Let us ﬁrst prove the afﬁne equivalent of proposition 4.1.9. Example 4.2.2. (i) As An is an open subset of Pn , it follows by corollary 4.1.8 that dim An = n.

4.

Dimension

55

(ii) As Am+n is an open subset of Pn ×Pm , it follows by (i) that dim(Pn ×Pm ) = n+m. (iii) Let f ∈ k[x1 , . . . , xn ] be a non-constant polynomial. We claim that Z( f ) ⊂ An has ¯ dimension n − 1. In fact, let X ⊂ Pn be the projective closure of Z( f ); by propo¯ sition 4.1.9 there is a component Y of X of dimension n − 1. As the homogenized ¯ polynomial f does not contain x0 as a factor, X cannot contain the whole “inﬁnity ¯ locus” Pn \An ∼ Pn−1 . So the part of X in the inﬁnity locus has dimension at most = ¯ n − 2; in particular the component Y of X has non-empty intersection with An . In other words, Z( f ) ⊂ An has dimension n − 1. (iv) Let f ∈ k[x1 , . . . , xn ] be as in (iii); we claim that in fact the dimension of every irreducible component of Z( f ) ⊂ An is n − 1: in fact, as k[x1 , . . . , xn ] is a unique factorization domain, we can write f as a product f1 · · · fr of irreducible polynomials, so that the decomposition of Z( f ) into its irreducible components is Z( f1 ) ∪ · · · ∪ Z( fr ). Now we can apply (iii) to the fi separately to get the desired result. (v) The corresponding statements to (iii) and (iv) are true for the zero locus of a homogeneous polynomial in Pn as well (the proof is the same). By (iv) and (v), there is a one-to-one correspondence between closed subvarieties of An (resp. Pn ) of dimension n − 1 and non-constant irreducible polynomials in k[x1 , . . . , xn ] (resp. non-constant homogeneous polynomials in k[x0 , . . . , xn ]). Varieties that are of this form are called hypersurfaces; if the degree of the polynomial is 1 they are called hyperplanes. Remark 4.2.3. Next we want to prove for general afﬁne varieties X ⊂ An that the dimension of (every component of) X ∩ Z( f ) is dim X − 1. Note that this does not follow immediately from the projective case as it did for X = An in example 4.2.2 (iii) or (iv): (i) As for example 4.2.2 (iii), of course we can still consider the projective closure ¯ X of X in Pn and intersect it with the zero locus of the homogenization of f ; but proposition 4.1.9 only gives us the existence of one component of dimension ¯ ¯ dim X − 1 in X ∩ Z( f ). It may well be that there is a component of X ∩ Z( f ) n \An , in which case we get that is contained in the “hyperplane at inﬁnity” P no information about the afﬁne zero locus X ∩ Z( f ). As an example you may 2 consider the projective variety X = {x0 x2 = x1 } ⊂ P2 and f = x1 : then X ∩Z( f ) = (1 : 0 : 0) ∪ (0 : 0 : 1) contains a point (0 : 0 : 1) at inﬁnity as an irreducible component. (ii) As for example 4.2.2 (iv), note that a factorization of f as for An is simply not possible in general. For example, in the case just considered in (i), Z( f ) intersects X in two points, but there is no decomposition of the linear function f into two factors that vanish on only one of the points. Nevertheless the idea of the proof is still to use projections from points: Proposition 4.2.4. Let X ⊂ An be an afﬁne variety, and let f ∈ k[x1 , . . . , xn ] be a nonconstant polynomial that does not vanish identically on X. Then dim(X ∩ Z( f )) = dim X − / 1 (unless X ∩ Z( f ) = 0). Proof. We prove the statement by induction on n (not on dim X!); there is nothing to show for n = 0. If X = An the statement follows from example 4.2.2 (iv), so we can assume that X An . ¯ Let X be the projective closure in Pn ; we can assume by an afﬁne change of coordi¯ ¯ nates that P = (0 : · · · : 0 : 1) ∈ X. Consider the projection π : X → Pn−1 from P as in / ¯ example 4.1.3. Obviously, we can restrict this projection map to the afﬁne space An ⊂ Pn given by x0 = 0; we thus obtain a morphism π : X → π(X) that is given in coordinates by ¯ ¯ (a1 , . . . , an ) → (a1 , . . . , an−1 ). Note that π(X) is closed in An , as π(X) = π(X) ∩ An .

56

Andreas Gathmann

By lemma 4.1.5 (i) applied to the function xn we see that there is a relation

D D−1 p(xn ) := xn + a1 xn + · · · aD = 0

in A(X)

(∗)

for some D > 0 and some ai ∈ k[x1 , . . . , xn−1 ] that do not depend on xn . Let K be the ﬁeld k(x1 , . . . , xn−1 ) of rational functions in n − 1 variables. Set V = K[xn ]/p(xn ); by (∗) this is D−1 a D-dimensional vector space over K (with basis 1, xn , . . . , xn ). Obviously, every polynomial g ∈ k[x1 , . . . , xn ] deﬁnes a vector space homomorphism g : V → V (by polynomial multiplication), so we can talk about its determinant det g ∈ K. Moreover, it is easy to see that det g ∈ k[x1 , . . . , xn−1 ], as the deﬁnition of the determinant does not use divisions. Note also that det g = gD if g ∈ k[x1 , . . . , xn−1 ]. Now go back to our original problem: describing the zero locus of the given polynomial f on X. We claim that π(X ∩ Z( f )) = π(X) ∩ Z(( f ) ∩ k[x1 , . . . , xn−1 ]) ⊃ π(X) ∩ Z(det f ) (in fact there is equality, but we do not need this). The ﬁrst equality is obvious from the deﬁnition of π. To prove the second inclusion, note that by the Nullstellensatz it sufﬁces to show that ( f ) ∩ k[x1 , . . . , xn−1 ] ⊂ (det f ). So let g ∈ ( f ) ∩ k[x1 , . . . , xn−1 ]; in particular g = f · b for some b ∈ k[x1 , . . . , xn ]. It follows that gD = det g = det f · det b ∈ (det f ), i.e. g ∈ (det f ), as we have claimed. The rest is now easy: dim(X ∩ Z( f )) = dim π(X ∩ Z( f )) ≥ dim(π(X) ∩ Z(det f )) by corollary 4.1.7 and proposition 4.2.1 by the inclusion just proven

= dim π(X) − 1 by the induction hypothesis = dim X − 1 by corollary 4.1.7 and proposition 4.2.1 again. The opposite inequality follows trivially from lemma 4.1.4 (ii). It is now quite easy to extend this result to a statement about every component of X ∩ Z( f ): Corollary 4.2.5. Let X ⊂ An be an afﬁne variety, and let f ∈ k[x1 , . . . , xn ] be a non-constant polynomial that does not vanish identically on X. Then every irreducible component of X ∩ Z( f ) has dimension dim X − 1. Proof. Let X ∩ Z( f ) = Z1 ∪ · · · ∪ Zr be the decomposition into irreducible components; we want to show that dim Z1 = dim X − 1. Let g ∈ k[x1 , . . . , xn ] be a polynomial that vanishes on Z2 , . . . , Zr but not on Z1 , and let U = Xg = X\Z(g). Then U is an afﬁne variety by lemma 2.3.16, and U ∩ Z( f ) has only one component Z1 ∩ U. So the statement follows from proposition 4.2.4 together with proposition 4.2.1. Remark 4.2.6. Proposition 4.2.1 and especially corollary 4.2.5 are the main properties of the dimension of varieties. Together they allow to compute the dimension of almost any variety without the need to go back to the cumbersome deﬁnition. Here are two examples: Corollary 4.2.7. Let f : X → Y be a morphism of varieties, and assume that the dimension of all ﬁbers n = dim f −1 (P) is the same for all P ∈ Y . Then dim X = dimY + n. Proof. We prove the statement by induction on dimY ; there is nothing to show for n = 0 (i.e. if Y is a point).

4.

Dimension

57

By proposition 4.2.1 we can assume that Y ⊂ Am is an afﬁne variety. Let f ∈ k[x1 , . . . , xm ] be any non-zero polynomial in the coordinates of Am that vanishes somewhere, but not everywhere on Y , let Y ⊂ Y be an irreducible component of Y ∩ Z( f ), and let X = f −1 (Y ). Then it follows by corollary 4.2.5 and the induction hypothesis that dim X = dim X + 1 = dimY + n + 1 = dimY + n. Example 4.2.8. (i) For any varieties X, Y we have dim(X × Y ) = dim X + dimY (apply corollary 4.2.7 to the projection morphism X ×Y → X). (ii) Combining corollary 4.2.7 with proposition 4.2.1 again, we see that it is actually sufﬁcient that f −1 (P) is non-empty and of the same dimension for all P in a non-empty open subset U of Y . Corollary 4.2.9. Let X and Y be afﬁne varieties in An . Then every irreducible component of X ∩Y ⊂ An has dimension at least dim X + dimY − n. Proof. Rewrite X ∩ Y as the intersection of X × Y with the diagonal ∆(An ) in An × An . The diagonal is given by the zero locus of the n functions xi − yi for 1 ≤ i ≤ n, where x1 , . . . , xn , y1 , . . . , yn are the coordinates of An × An . By corollary 4.2.5, every component of the intersection of an afﬁne variety Z with the zero locus of a non-constant function has dimension at least equal to dim Z − 1 (it is dim Z if f vanishes identically on Z, and dim Z − 1 otherwise). Applying this statement n times to the functions xi − yi on X ×Y in An × An we conclude that every component of X ∩Y has dimension at least dim(X ×Y ) − n = dim X + dimY − n. Remark 4.2.10. (For commutative algebra experts) There is another more algebraic way of deﬁning the dimension of varieties that is found in many textbooks: the dimension of a variety X is the transcendence degree over k of the ﬁeld of rational functions K(X) on X. Morally speaking, this deﬁnition captures the idea that the dimension of a variety is the number of independent coordinates on X. We have not used this deﬁnition here as most propositions concerning dimensions would then have required methods of (commutative) algebra that we have not developed yet. Here are some ideas that can be used to show that this algebraic deﬁnition of dimension is equivalent to our geometric one: • If U ⊂ X is a non-empty open subset we have K(U) = K(X), so with the algebraic deﬁnition of dimension it is actually trivial that dimU = dim X. • It is then also obvious that dim An = tr deg k(x1 , . . . , xn ) = n. • Let π : X → π(X) be a projection map as in the proof of proposition 4.2.4. The relation (∗) in the proof can be translated into the fact that K(X) is an algebraic ﬁeld extension of K(π(X)) (we add one variable xn , but this variable satisﬁes a polynomial relation). In particular, these two ﬁelds have the same transcendence degree, translating into the fact that dim π(X) = dim X. 4.3. Blowing up. We have just seen in 4.2.1 that two varieties have the same dimension if they contain an isomorphic (non-empty) open subset. In this section we want to study this relation in greater detail and construct a large and important class of examples of varieties that are not isomorphic but contain an isomorphic open subset. Let us ﬁrst make some deﬁnitions concerning varieties containing isomorphic open subsets. We will probably not use them very much, but they are often found in the literature. Deﬁnition 4.3.1. Let X and Y be varieties. A rational map f from X to Y , written f : X Y , is a morphism f : U → Y (denoted by the same letter) from a non-empty open

58

Andreas Gathmann

subset U ⊂ X to Y . We say that two such rational maps f : U → Y and g : V → Y with U,V ⊂ X are the same if f = g on U ∩V . A rational map f : X Y is called dominant if its image is dense in Y , i.e. if f is given by a morphism f : U → Y such that f (U) contains a non-empty open subset of Y . If f :X Y and g : Y Z are rational maps, and if f is dominant, then the composition g◦ f : X Z is a well-deﬁned rational map. A birational map from X to Y is a rational map with an inverse, i.e. it is a (dominant) rational map f : X Y such that there is a (dominant) rational map g : Y X with g ◦ f = idX and f ◦ g = idY as rational maps. Two varieties X and Y are called birational if there is a birational map between them. In other words, X and Y are birational if they contain an isomorphic non-empty open subset. We will now construct the most important examples of birational morphisms (resp. birational varieties), namely blow-ups. Construction 4.3.2. Let X ⊂ An be an afﬁne variety, and let f0 , . . . , fr ∈ k[x1 , . . . , xn ] be polynomial functions that do not vanish identically on X. Then U = X\Z( f0 , . . . , fr ) is a non-empty open subset of X, and there is a well-deﬁned morphism f : U → Pr , P → ( f0 (P) : · · · : fr (P)). Now consider the graph Γ = {(P, f (P)) ; P ∈ U} ⊂ X × Pr which is isomorphic to U (with inverse morphism (P, Q) → P). Note that Γ is in general not closed in X × Pr , because the points in X\U where ( f0 : · · · : fr ) is ill-deﬁned as a point in Pr are “missing”. ˜ The closure of Γ in X × Pr is called the blow-up of X in ( f0 , . . . , fr ); we denote it by X. r , and it is irreducible as Γ is; so it is a closed subvariety of It is a closed subset of X × P ˜ ˜ X × Pr . In particular, there are projection morphisms π : X → X and p : X → Pr . Note that ˜ both contain U as a dense open subset, so X and the blow-up X have the same ˜ X and X dimension. Let us now investigate the geometric meaning of blow-ups. Example 4.3.3. If r = 0 in the above notation, i.e. if there is only one function f0 , the ˜ ˜ ˜ blow-up X is isomorphic to X. In fact, we then have X ⊂ X × P0 ∼ X, so X is the smallest = closed subvariety containing U. Example 4.3.4. Let X = A2 with coordinates x0 , x1 , and let f0 = x0 , f1 = x1 . Then the blow-up of X in ( f0 , f1 ) is a subvariety of A2 × P1 . The morphism (x0 , x1 ) → (x0 : x1 ) is well-deﬁned on U = X\{(0, 0)}; so on this open subset the graph is given by Γ = {((x0 , x1 ), (y0 : y1 )) ; x0 y1 = x1 y0 } ⊂ U × P1 . The closure of Γ is now obviously given by the same equation, considered in A2 × P1 : ˜ X = {((x0 , x1 ), (y0 : y1 )) ; x0 y1 = x1 y0 } ⊂ A2 × P1 . The projection morphisms to X = A2 and P1 are obvious. Note that the inverse image of a point P = (x0 , x1 ) ∈ X\{(0, 0)} under π is just the single point ((x0 , x1 ), (x0 : x1 )) — we knew this before. The inverse image of (0, 0) ∈ X however is P1 , as the equation x0 y1 = x1 y0 imposes no conditions on y0 and y1 if (x0 , x1 ) = (0, 0). To give a geometric interpretation of the points in π−1 (0, 0) let us ﬁrst introduce one more piece of notation. Let Y ⊂ X be a closed subvariety that has non-empty intersection ˜ ˜ with U. As U is also a subset of X, we can consider the closure of Y ∩ U in X. We call this the strict transform of Y . Note that by deﬁnition the strict transform of Y is just the ˜ blow-up of Y at ( f0 , . . . , fr ); so we denote it by Y .

4.

Dimension

59

Now let C ⊂ X = A2 be a curve, given by the equation

i g(x0 , x1 ) = ∑ ai, j x0 x1 = a0,0 + a1,0 x0 + a0,1 x1 + a1,1 x0 x1 + · · · . j i, j

Assume that a0,0 = 0, i.e. that C passes through the origin in A2 , and that (a1,0 , a0,1 ) = (0, 0), so that C has a well-deﬁned tangent line at the origin, given by the linearization ˜ a1,0 x0 + a0,1 x1 = 0 of g. Let us compute the strict transform C. Of course, the points ˜ ((x0 , x1 ), (y0 : y1 )) of C satisfy the equation

2 2 a1,0 x0 + a0,1 x1 + a1,1 x0 x1 + a2,0 x0 + a0,2 x1 + · · · = 0. (∗) ˜ But it is not true that C is just the common zero locus in A2 × P1 of this equation together with x0 y1 = x1 y0 , because this common zero locus contains the whole ﬁber π−1 (0, 0) ∼ P1 = ˜ — but C has to be irreducible of dimension 1, so it cannot contain this P1 . In fact, we have forgotten another relation: on the open set where x0 = 0 and x1 = 0 we can multiply (∗) y0 with x0 ; using the relation y0 = y1 we get x0 x1

a1,0 y0 + a0,1 y1 + a1,1 y0 x1 + a2,0 x0 y0 + a0,2 x1 y1 + · · · = 0. ˜ This equation must then necessarily hold on the closure C too. Restricting it to the origin (x0 , x1 ) = (0, 0) we get a1,0 y0 + a0,1 y1 = 0, which is precisely the equation of the tangent ˜ line to C at (0, 0). In other words, the strict transform C of C intersects the ﬁber π−1 (0, 0) precisely in the point of P1 corresponding to the tangent line of C in (0, 0). In this sense we can say that the points of π−1 (0, 0) correspond to tangent directions in X at (0, 0). The following picture illustrates this: we have two curves C1 , C2 that intersect at the ˜ ˜ origin with different tangent directions. The strict transforms C1 and C2 are then disjoint ˜ on the blow-up X.

X ~ C1 π −1 (0,0) ~ C2 ~ X π C2

C1

Let us now generalize the results of this example to general blow-ups. Note that in the example we would intuitively say that we have “blown up the origin”, i.e. the zero locus of the functions f0 , . . . , fr . In fact, the blow-up construction depends only on the ideal generated by the fi : Lemma 4.3.5. The blow-up of an afﬁne variety X at ( f0 , . . . , fr ) depends only on the ideal I ⊂ A(X) generated by f0 , . . . , fr . We will therefore usually call it the blow-up of X at the ideal I. If I = I(Y ) for a closed subset Y ⊂ X, we will also call it the blow-up of X in Y . Proof. Let ( f0 , . . . , fr ) and ( f0 , . . . , fs ) be two sets of generators of the same ideal I ⊂ A(X), ˜ ˜ and let X and X be the blow-ups of X at these sets of generators. By assumption we have relations in A(X) fi = ∑ gi, j f j and f j = ∑ g j,k fk .

j k

˜ ˜ We want to deﬁne a morphism X → X by sending (P, (y0 : · · · : yr )) to (P, (y0 : · · · : ys )), where y j = ∑k g j,k (P)yk . First of all we show that this deﬁnes a morphism to X × Ps , i.e. that the y j cannot be simultaneously zero. Note that the relation fi = ∑ j,k gi, j g j,k fk

60

Andreas Gathmann

implies by lemma 4.3.6 (i) that yi = ∑ j,k gi, j g j,k yk on X. So if y j = ∑k g j,k yk = 0 then also yi = ∑ j gi, j y j = 0, which is a contradiction. ˜ Hence we have deﬁned a morphism X → X × Ps . By construction it maps the open ˜ ˜ ˜ ˜ subset X\Z( f0 , . . . , fr ) ⊂ X to X\Z( f0 , . . . , fs ) ⊂ X , so it must map its closure X to X ˜ ˜ ˜ ˜ as well. By the same arguments we get an inverse morphism X → X, so X and X are isomorphic. ˜ Let us now study the variety X itself, in particular over the locus Z( f0 , . . . , fr ) where ˜ π : X → X is not an isomorphism. ˜ Lemma 4.3.6. Let X ⊂ An be an afﬁne variety, and let X be the blow-up of X at the ideal I = ( f0 , . . . , fr ). Then: ˜ (i) The blow-up X is contained in the set {(P, (y0 : · · · : yr )) ; yi f j (P) = y j fi (P) for all i, j = 0, . . . , r} ⊂ X × Pr . (ii) The inverse image π−1 (Z( f0 , . . . , fr )) is of pure dimension dim X − 1. It is called the exceptional hypersurface. Proof. (i): By deﬁnition we must have (y0 : · · · : yr ) = ( f0 (P) : · · · : fr (P)) on the non˜ empty open subset X\Z(I) ⊂ X. So these equations must be true as well on the closure of ˜ this open subset, which is X by deﬁnition. (ii): It is enough to prove the statement on the open subset where yi = 0, as these open ˜ subsets for all i cover X. Note that on this open subset the condition fi (P) = 0 implies f j (P) = 0 for all j by the equations of (i). So the inverse image π−1 (Z( f0 , . . . , fr )) is ˜ given by one equation f j = 0, and is therefore of pure dimension dim X − 1 = dim X − 1 by corollary 4.2.5. Example 4.3.7. In example 4.3.4, X = A2 has dimension 2, and the exceptional hypersurface was isomorphic to P1 , which has dimension 1. ˜ Remark 4.3.8. The equations in lemma 4.3.6 (i) are in general not the only ones for X. Note that they do not impose any conditions over the zero locus Z( f0 , . . . , fr ) at all, so that it would seem from these equations that the exceptional hypersurface is always Pr . This must of course be false in general just for dimensional reasons (see lemma 4.3.6 (ii)). In fact, we can write down explicitly the equations for the exceptional hypersurface. We will do this here only in the case of the blow-up of (the ideal of) a point P, which is the most important case. By change of coordinates, we can then assume that P is the origin in An . For any f ∈ k[x1 , . . . , xn ] we let f in be the “initial polynomial” of f , i.e. if f = ∑i f (i) is the splitting of f such that f (i) is homogeneous of degree i, then f in is by deﬁnition equal to the smallest non-zero f (i) . If I ⊂ k[x1 , . . . , xn ], we let I in be the ideal generated by the initial polynomials f in for all f ∈ I. Note that I in is by deﬁnition a homogeneous ideal. So its afﬁne zero locus Za (I in ) ⊂ An is a cone, and there is also a well-deﬁned projective zero locus Z p (I in ). By exercise 4.6.8, the exceptional hypersurface of the blowup of an afﬁne variety X ⊂ An in the origin is precisely Z p (I(X)in ). (The proof of this statement is very ˜ similar to the computation of C in example 4.3.4.) Let us ﬁgure out how this can be interpreted geometrically. By construction, I(X)in is obtained from I(X) by only keeping the terms of lowest degree, so it can be interpreted as an “approximation” of I(X) around zero, just in the same way as the Taylor polynomial approximates a function around a given point. Note also that Za (I(X)in ) has the same dimension as X by lemma 4.3.6 (ii). Hence we can regard Za (I(X)in ) ⊂ An as the cone that approximates X best around the point P. It is called the tangent cone of X in P and ˜ denoted CX,P . The exceptional locus of the blow-up X of X in P is then the “projectivized

4.

Dimension

61

tangent cone”, i.e. it corresponds to “tangent directions” in X through P, just as in example 4.3.4. Example 4.3.9. Here are some examples of tangent cones. (i) Let X = {(x, y) ; y = x(x − 1)} ⊂ A2 . The tangent cone of X in P = (0, 0) is given by keeping only the linear terms of the equation y = x(x − 1), i.e. CX,P = {(x, y) ; y = −x} is the tangent line to X in P. Consequently, the exceptional ˜ hypersurface of the blow-up of X in P contains only one point. In fact, X is isomorphic to X in this case: note that on X, the ideal of P is just given by the single function x, as (y − x(x − 1), x) = (x, y). So we are blowing up at f0 = x ˜ only. It follows then by example 4.3.3 that X = X. (ii) Let X = {(x, y) ; y2 = x2 + x3 } ⊂ A2 . This time there are no linear terms in the equation of X, so the tangent cone in P = (0, 0) is given by the quadratic terms CX,P = {(x, y) ; y2 = x2 }, i.e. it is the union of the two tangent lines y = x and y = −x to X in P (see the picture below). The exceptional hypersurface of the blow-up of X in P therefore contains exactly two points, one for every tangent direction in P. In other words, the two local branches of X around P get separated in the blow-up. Note that we cannot apply the argument of (i) here that ˜ X should be isomorphic to X: the ideal of P cannot be generated on X by one function only. While it is true that the zero locus of (x, y2 − x2 − x3 ) is P, the ideal (x, y2 − x2 − x3 ) = (x, y2 ) is not equal to I(P) = (x, y) — and this is the important point. In particular, we see that the blow-up of X in an ideal I really does depend on the ideal I and not just on its zero locus, i.e. on the radical of I. (iii) Let X = {(x, y) ; y2 = x3 } ⊂ A2 . This time the tangent cone is CX,P = {y2 = 0}, ˜ i.e. it is only one line. So for X the point P ∈ X is replaced by only one single ˜ point again, as in (i). But in this case X and X are not isomorphic, as we will see in 4.4.7.

y X P x CX,P X P CX,P x CX,P P y y X x

(i)

(ii)

(iii)

Remark 4.3.10. Let X be any variety, and let Y ⊂ X be a closed subset. For an afﬁne open ˜ ˜ cover {Ui } of X, let Ui be the blow-up of Ui in Ui ∩Y . It is then easy to see that the Ui can ˜ be glued together to give a blow-up variety X. In what follows, we will only need this in the case of the blow-up of a point, where the construction is even easier as it is local around the blown-up point: let X be a variety, and ˜ let P ∈ X be a point. Choose an afﬁne open neighborhood U ⊂ X of P, and let U be the ˜ by glueing X\P to U along the common open subset ˜ blow-up of U in P. Then we obtain X U\P. In particular, this deﬁnes the tangent cone CX,P to X at P for any variety X: it is the afﬁne cone over the exceptional hypersurface of the blow-up of X in P. This sort of glueing currently works only for blow-ups at subvarieties, i.e. for blow-ups at radical ideals. For the general construction we would need to patch ideals, which we do not know how to do at the moment. Note however that it is easy to see that for projective varieties, the blow-up at a homogeneous ideal can be deﬁned in essentially the same way as for afﬁne varieties: let X ⊂ Pn be a projective variety, and let Y ⊂ X be a closed subset. If f0 , . . . , fr are homogeneous

62

Andreas Gathmann

generators of I(Y ) of the same degree, the blow-up of X in Y is precisely the closure of Γ = {(P, ( f0 (P) : · · · : fr (P)) ; P ∈ U} ⊂ X × Pr in X × Pr (this is easily checked on the afﬁne patches fi = 0). Example 4.3.11. The following property of blow-ups follows trivially from the deﬁnitions, yet it is one of their most important properties. Let X ⊂ An be an afﬁne variety, and let f0 , . . . , fr be polynomials that do not vanish identically on X. Note that the morphism f : P → ( f0 (P) : · · · : fr (P)) to Pr is only welldeﬁned on the open subset U = X\Z( f0 , . . . , fr ) of X. In general, we can not expect that this morphism can be extended to a morphism on all of X. But we can always extend it “after blowing up the ideal ( f0 , . . . , fr ) of the indeterminacy locus”, i.e. there is an extension f˜ : ˜ ˜ X → Pr (that agrees with f on the open subset U), namely just the projection from X ⊂ X × Pr → Pr . So blowing up is a way to extend morphisms to bigger sets on which they would otherwise be ill-deﬁned. The same is true for projective varieties and the construction at the end of remark 4.3.10. Let us consider a concrete example of this idea in the next lemma and the following remark: Lemma 4.3.12. P1 × P1 blown up in one point is isomorphic to P2 blown up in two points. Proof. We know from example 3.3.14 that P1 × P1 is isomorphic to the quadric surface Q = {(x0 : x1 : x2 : x3 ) ; x0 x3 = x1 x2 } ⊂ P3 . ˜ Let P = (0 : 0 : 0 : 1) ∈ Q, and let Q ⊂ P3 × P2 be the blow-up of Q in the ideal I(P) = (x0 , x1 , x2 ). ˜ On the other hand, let R1 = (0 : 1 : 0), R2 = (0 : 0 : 1) ∈ P2 , and let P2 ⊂ P2 × P3 be the blow-up of P2 in the ideal I = (y2 , y0 y1 , y0 y2 , y1 y2 ). Note that this is not quite the ideal 0 I(R1 ∪ R2 ) = (y0 , y1 y2 ), but this does not matter: the blow-up is a local construction, so let us check that we are doing the right thing around R1 . There is an open afﬁne neighborhood around R1 given by y1 = 0, and on this neighborhood the ideal I is just (y2 , y0 , y0 y2 , y2 ) = 0 (y0 , y2 ), which is precisely the ideal of R1 . The same is true for R2 , so the blow-up of P2 in I is actually the blow-up of P2 in the two points R1 and R2 . Now we claim that an isomorphism is given by ˜ ˜ f : Q → P2 , ((x0 : x1 : x2 : x3 ), (y0 : y1 : y2 )) → ((y0 : y1 : y2 ), (x0 : x1 : x2 : x3 )). In fact, this is easy to check: obviously, f is an isomorphism P2 × P3 → P3 × P2 , so we ˜ ˜ ˜ ˜ only have to check that f maps Q to P2 , and that f −1 maps P2 to Q. Note that it sufﬁces ˜ to check this away from the blown-up points: f −1 (P2 ) is a closed subset of P3 × P2 , so if ˜ it contains a non-empty open subset U ⊂ Q (e.g. Q minus the exceptional hypersurface), it must contain all of Q. ˜ But this is now easy to check: on Q we have x0 x3 = x1 x2 and (y0 : y1 : y2 ) = (x0 : x1 : x2 ) (where this is well-deﬁned), so in the image of f we get the correct equations

2 2 (x0 : x1 : x2 : x3 ) = (x0 : x0 x1 : x0 x2 : x0 x3 ) = (x0 : x0 x1 : x0 x2 : x1 x2 ) = (y2 : y0 y1 : y0 y2 : y1 y2 ) 0 ˜ 2 . Conversely, on P2 we have (x0 : x1 : x2 : x3 ) = (y2 : y0 y1 : ˜ for the image point to lie in P 0 y0 y2 : y1 y2 ) where deﬁned, so we conclude x0 x3 = x1 x2 and (y0 : y1 : y2 ) = (x0 : x1 : x2 ).

Remark 4.3.13. The proof of lemma 4.3.12 is short and elegant, but not very insightful. Let us try to understand geometrically what is going on. As in the proof, we think of P1 × P1 as the quadric Q = {(x0 : x1 : x2 : x3 ) ; x0 x3 = x1 x2 } ⊂ P3 . Consider the projection π from P to P2 , given in coordinates by π(x0 : x1 : x2 : x3 ) = (x0 : x1 : x2 ). We have considered projections from points before, but so far the projection point

4.

Dimension

63

P was always assumed not to lie on the given variety Q. This is not the case here, and consequently π is only well-deﬁned on Q\P. To construct π(P) we would have to take “the line through P and P” and intersect it with a given P2 ⊂ P3 that does not contain P. Of course this is ill-deﬁned. But there is a well-deﬁned line through P and any point P near P which we can intersect with P2 . It is obvious that π(P) should be the limit of these projection points when P tends to P. The line P P will then become a tangent line to Q. But Q, being two-dimensional, has a one-parameter family of tangent lines. This is why π(P) is ill-deﬁned. But we also see from this discussion that blowing up P on Q, i.e. replacing it by the set of tangent lines through P, will exactly resolve the indeterminacy. ˜ We have thus constructed a morphism Q = P1 × P1 → P2 by projection from P. If there is an inverse morphism, it is easy to see what it would have to look like: pick a point R ∈ P2 ⊂ P3 . The points mapped to R by π are exactly those on the line PR not equal to P. In general, this line intersects the quadric Q in two points, one of which is P. So there is exactly one point on Q which maps to R. This reasoning is false however if the whole line PR = P1 lies in Q. This whole line would then be mapped to R, so that we cannot have an isomorphism. But of course we expect again that this problem can be taken care of by blowing up R in P2 , so that it is replaced by a P1 that can then be mapped one-to-one to PR. There are obviously two such lines PR1 and PR2 , given by R1 = (0 : 1 : 0) and R2 = (0 : 0 : 1). If you think of Q as P1 × P1 again, these lines are precisely the “horizontal” and “vertical” lines P1 × {point} and {point} × P1 passing through P. So we would expect that ˜ π can be made into an isomorphism after blowing up R1 and R2 , which is what we have shown in lemma 4.3.12.

Q P’ P

R1

π (P’)

R2

IP 2

4.4. Smooth varieties. Let X ⊂ An be an afﬁne variety, and let P ∈ X be a point. By a change of coordinates let us assume that P = (0, . . . , 0) is the origin. In remark 4.3.8 we have deﬁned the tangent cone of X in P to be the closed subset of An given by the initial ideal of X, i.e. the “local approximation” of X around P given by keeping only the terms of the deﬁning equations of X of minimal degree. Let us now make a similar deﬁnition, but where we only keep the linear terms of the deﬁning equations. Deﬁnition 4.4.1. For any polynomial f ∈ k[x1 , . . . , xn ] denote by f (1) the linear part of f . For an ideal I ⊂ k[x1 , . . . , xn ] denote by I (1) = { f (1) ; f ∈ I} the vector space of all linear parts of the elements of I; this is by deﬁnition a vector subspace of the n-dimensional space k[x1 , . . . , xn ](1) of all linear forms {a1 x1 + · · · + an xn ; ai ∈ k}. The zero locus is then a linear subspace of An . It is canonically dual (as a vector (1) /I (1) , since the pairing space) to k[x1 , . . . , xn ] k[x1 , . . . , xn ](1) /I (1) × Z(I (1) ) → k, is obviously non-degenerate. ( f , P) → f (P) Z(I (1) )

64

Andreas Gathmann

Now let X ⊂ An be a variety. By a linear change of coordinates, assume that P = (0, . . . , 0) ∈ X. Then the linear space Z(I(X)(1) ) is called the tangent space to X at P and denoted TX,P . Remark 4.4.2. Let us make explicit the linear change of coordinates mentioned in the deﬁnition. If P = (a1 , . . . , an ) ∈ X, we need to change coordinates from the xi to yi = xi −ai . By a (purely formal) Taylor expansion we can rewrite any polynomial f ∈ k[x1 , . . . , xn ] as f (x1 , . . . , xn ) = f (P) + ∑

i

∂f (P) · yi + (terms at least quadratic in the yi ), ∂xi

so we see that the tangent space TX,P to any point P = (a1 , . . . , an ) ∈ X is given by the equations ∂f ∑ ∂xi (P) · (xi − ai ) = 0 i for all f ∈ I(X). Here is an alternative description of the tangent space. For simplicity, we will assume again that the coordinates have been chosen such that P = (0, . . . , 0). Lemma 4.4.3. Let X ⊂ An be a variety, and assume that P = (0, . . . , 0) ∈ X. Then k[x1 , . . . , xn ](1) /I(X)(1) = M/M 2 , where M = {ϕ ; ϕ(P) = 0} ⊂ OX,P is the maximal ideal in the local ring of X at P. Proof. Recall that f ; f , g ∈ A(X), g(P) = 0 , g

OX,P =

and therefore M=

f ; f , g ∈ A(X), f (P) = 0, g(P) = 0 . g

There is an obvious homomorphism k[x1 , . . . , xn ](1) /I(X)(1) → M/M 2 of k-vector spaces. We will show that it is bijective. f Injectivity: Let f ∈ k[x1 , . . . , xn ](1) be a linear function. Then 1 is zero in OX,P if and only if it is zero in A(X), i.e. if and only if f ∈ I(X). Surjectivity: Let ϕ = Set

f g

∈ M. Without loss of generality we can assume that g(P) = 1. ϕ =∑

i

∂ϕ (P) · xi , ∂xi

which is obviously an element of k[x1 , . . . , xn ](1) . We claim that ϕ − ϕ ∈ M 2 . In fact, g(ϕ − ϕ ) = f − g ∑

i ∂f ∂g ∂xi (P)g(P) − ∂xi (P) f (P) g(P)2

xi

= f −g ∑

i

∂f (P) xi ∂xi

i

≡ f − g(P) ∑ = f −∑

i

∂f (P) xi ∂xi

(mod M 2 )

(as g − g(P) and xi are in M)

∂f (P) xi ∂xi (mod M 2 ) (as this is the linear Taylor expression for f ).

≡0 So ϕ = ϕ in M/M 2 .

4.

Dimension

65

Remark 4.4.4. In particular, this lemma gives us a more intrinsic deﬁnition of the tangent space TX,P : we can say that TX,P is the dual of the k-vector space M/M 2 , where M is the maximal ideal in the local ring OX,P . This alternative deﬁnition shows that the tangent space TX,P (as an abstract vector space) is independent of the chosen embedding of X in afﬁne space. It also allows us to deﬁne the tangent space TX,P for any variety X (that is not necessarily afﬁne). Let us now compare tangent spaces to tangent cones. Remark 4.4.5. Let X be an afﬁne variety, and assume for simplicity that P = (0, . . . , 0) ∈ X. For all polynomials f ∈ k[x1 , . . . , xn ] vanishing at P, linear terms are always initial. Hence the ideal generated by I(X)(1) is contained in the ideal I(X)in deﬁning the tangent cone (see remark 4.3.8). So the tangent cone CX,P ⊂ An is contained in the tangent space TX,P ⊂ An . In particular, we always have dim TX,P ≥ dimCX,P = dim X. Summarizing, we can say that, in studying the local properties of X around P, the tangent cone has the advantage that it always has the “correct” dimension dim X, whereas the tangent space has the advantage that it is always a linear space. We should give special attention to those cases when both notions agree, i.e. when X “can be approximated linearly” around P. Deﬁnition 4.4.6. A variety X is called smooth at the point P ∈ X if TX,P = CX,P , or equivalently, if the tangent space TX,P to X at P has dimension (at most) dim X. It is called singular at P otherwise. We say that X is smooth if it is smooth at all points P ∈ X; otherwise X is singular. Example 4.4.7. Consider again the curves of example 4.3.9: (i) X = {y = x(x − 1)} ⊂ A2 , (ii) X = {y2 = x2 + x3 } ⊂ A2 , (iii) X = {y2 = x3 } ⊂ A2 . In case (i), the tangent space is {y = −x} ⊂ A2 and coincides with the tangent cone: X is smooth at P = (0, 0). In the cases (ii) and (iii), there are no linear terms in the deﬁning equations of X. So the tangent space of X at P is all of A2 , whereas the tangent cone is one-dimensional. Hence in these cases X is singular at P. In case (iii) let us now consider the blow-up of X in P = (0, 0). Let us ﬁrst blow up the ambient space A2 in P; we know already that this is given by ˜ A2 = {((x, y), (x : y )) ; xy = x y} ⊂ A2 × P1 . ˜ So local afﬁne coordinates of A2 around the point ((0, 0), (1 : 0)) are (u, v) ∈ A2 , where u= y x and v = x

so that ((x, y), (x : y )) = ((v, uv), (1 : u)). In these local coordinates, the equation y2 = x3 of the curve X is given by (uv)2 = v3 . The exceptional hypersurface has the local equation v = 0, so away from this hypersurface the curve X is given by the equation v = u2 . By ˜ deﬁnition, this is then also the equation of the blow-up X. ˜ is smooth, although X was not. We say So we conclude ﬁrst of all that the blow-up X that the singularity P ∈ X got “resolved” by blowing up. We can also see that the blow-up of the curve (with local equation v = u2 ) is tangent to the exceptional hypersurface (with local equation v = 0). All this is illustrated in the following picture (the blow-up of A2 is the same as in example 4.3.4):

66

Andreas Gathmann

~ X π −1 (0,0) π

X

It can in fact be shown that every singularity can be “resolved” in a similar way by successively blowing up the singular locus. The good thing about smoothness is that is very easy to check: Proposition 4.4.8. (i) (Afﬁne Jacobi criterion) Let X ⊂ An be an afﬁne variety with ideal I(X) = ( f1 , . . . , fr ), and let P ∈ X be a point on X. Then X is smooth at P if and only ∂ if the rank of the r × n “Jacobi matrix” ∂xfij (P) is (at least) n − dim X. (ii) (Projective Jacobi criterion) Let X ⊂ Pn be a projective variety with ideal I(X) = ( f1 , . . . , fr ), and let P ∈ X be a point on X. Then X is smooth at P if and only if ∂ the rank of the r × n Jacobi matrix ∂xfij (P) is (at least) n − dim X. In particular, if the rank is r (the number of functions) then X is smooth of dimension n − r. Proof. (i): By remark 4.4.2, the linearization of the functions fi around the point P = ∂ (a1 , . . . , an ) is given by ∑ j ∂xfij (P) · (xi − ai ). By deﬁnition, X is smooth at P if these functions deﬁne a linear subspace of An of dimension (at most) dim X, i.e. if and only if the linear subspace of k[x1 , . . . , xn ](1) spanned by the above linearizations has dimension (at least) n − dim X. But the dimension of this linear space is exactly the rank of the matrix whose entries are the coefﬁcients of the various linear function. (ii): This follows easily by covering the projective space Pn by the n + 1 afﬁne spaces {xi = 0} ∼ An , and applying the criterion of (i) to these n + 1 patches. = Remark 4.4.9. Note that a matrix has rank less than k if and only if all k × k minors are zero. These minors are all polynomials in the entries of the matrix. In particular, the locus of singular points, i.e. where the Jacobi matrix has rank less than n − dim X as in the proposition, is closed. It follows that the set {P ∈ X ; X is singular at P} ⊂ X is closed. In other words, the set of smooth points of a variety is always open. One can show that the set of smooth points is also non-empty for every variety (see e.g. [H] theorem I.5.3). Hence the set of smooth points is always dense. Example 4.4.10. (i) For given n and d, let X be the so-called Fermat hypersurface

d d X = {(x0 : · · · : xn ) ; x0 + · · · + xn = 0}. d−1 Then the Jacobi matrix has only one row, and the entries of this row are d xi for i = 0, . . . , n. Assuming that the characteristic of the ground ﬁeld is zero (or at least not a divisor of d), it follows that at least one of the entries of this matrix is

4.

Dimension

67

non-zero at every point. In other words, the rank of the Jacobi matrix is always 1. Therefore X is smooth by proposition 4.4.8. (ii) Let X be the “twisted cubic curve” of exercise 3.5.2 X = {(s3 : s2t : st 2 : t 3 ) ; (s : t) ∈ P1 }. We have seen earlier that X can be given by the equations

2 2 X = {(x0 : x1 : x2 : x3 ) ; x1 − x0 x2 = x2 − x1 x3 = x0 x3 − x1 x2 = 0}.

So the Jacobi matrix is given by −x2 2x1 0 −x3 x3 −x2

−x0 2x2 −x1

0 −x1 . x0

By proposition 4.4.8, X is smooth if and only if the rank of this matrix is 2. (We know already that the rank cannot be bigger than 2, which is also easily checked directly). 2 The 2 × 2 minor given by the last two rows and the ﬁrst two columns is x3 . 2 The 2 × 2 minor given by last two rows and the ﬁrst and last column is x1 x3 = x2 . 2 and x2 . These cannot all be simultaneSimilarly we ﬁnd 2 × 2 minors that are x1 0 ously zero; hence X is smooth. (Of course we have known this before, since X is just the degree-3 Veronese embedding of P1 (see example 3.4.11. In particular, X is isomorphic to P1 and therefore smooth.) Remark 4.4.11. The Jacobi criterion of proposition 4.4.8 gives us a direct connection to complex analysis. Assume that we are given r holomorphic functions on Cn (e.g. polynomials), and that the matrix of the derivatives of the fi has rank n − dim X at a point P, where X is the zero locus of the fi . Assume for simplicity that the square matrix ∂ fi of size n − dim X is invertible. Then the inverse function ∂x j (P)

1≤i≤n−dim X,dim X< j≤n

theorem states that the coordinates xdim X+1 , . . . , xn are locally around P determined by the other coordinates x1 , . . . , xdim X . I.e. there is a neighborhood U of P in Cn (in the classical topology!) and holomorphic functions gdim X+1 , . . . , gn of x1 , . . . , xdim X such that for every P = (x1 , . . . , xdim X ) ∈ U the functions fi vanish at P if and only if xi = gi (x1 , . . . , xdim X ) for i = dim X + 1, . . . , n. So the zero locus of the fi is “locally the graph of a holomorphic map” given by the gi . In other words, smoothness in algebraic geometry means in a sense the same thing as differentiability in analysis: the geometric object has “no edges”. Note however that the inverse function theorem is not true in the Zariski topology, because the open sets are too big. For example, consider the curve X = {(x, y) ; f (x, y) = y − x2 = 0} ⊂ C2 . Then ∂ f = 0 say at the point P = (1, 1) ∈ X. Consequently, in complex ∂x analysis x can be expressed locally in terms of y around P: it is just the square root of y. But any non-empty Zariski open subset of X will contain pairs of points (x, x2 ) and (−x, x2 ) for some x, so the inverse function theorem cannot hold here in algebraic geometry. 4.5. The 27 lines on a smooth cubic surface. As an application of the theory that we have developed so far, we now want to study lines on cubic surfaces in P3 . We have already mentioned in example 0.1.7 that every smooth cubic surface has exactly 27 lines on it. We now want to show this. We also want to study the conﬁguration of these lines, and show that every smooth cubic surface is birational to P2 . The results of this section will not be needed later on. Therefore we will not give all the proofs in every detail here. The goal of this section is rather to give an idea of what can be done with our current methods.

68

Andreas Gathmann

First let us recall some notation from exercise 3.5.4. Let G = G(1, 3) be the Grassmannian variety of lines in P3 . This is a 4-dimensional projective variety. In this section we will use local afﬁne coordinates on G: if L0 ∈ G is the line in P3 (with coordinates x0 , . . . , x3 ) given by the equations x2 = x3 = 0 (of course every line is of this form after a linear change of coordinates), then there is an open neighborhood A4 ⊂ G of L0 in G given by sending a point (a, b) := (a2 , b2 , a3 , b3 ) ∈ A4 to the line through the points (1, 0, a2 , a3 ) and (0, 1, b2 , b3 ). The cubic surfaces in P3 are parametrized by homogeneous polynomials of degree 3 in x0 , x1 , x2 , x3 up to scalars, which is a 19-dimensional projective space P19 . A cubic surface given by the equation fc := ∑α cα xα = 0 (in multi-index notation, so α runs over all quadruples of indices (α0 , α1 , α2 , α3 ) with αi ≥ 0 and ∑i αi = 3) corresponds to the point in P19 with homogeneous coordinates c = (cα ). We denote the corresponding cubic surface by Xc = { fc = 0}. To study lines in cubic surfaces, we consider the so-called incidence correspondence M := {(L, X) ; L ⊂ X} ⊂ G × P19 consisting of all pairs of a line and a cubic such that the line lies in the cubic. Let us start by proving some facts about this incidence correspondence. Lemma 4.5.1. With the above notation, the incidence correspondence M has an open cover by afﬁne spaces A19 . In particular, M is a smooth 19-dimensional variety. Proof. In the coordinates (a, b, c) = (a2 , a3 , b2 , b3 , cα ) as above, the incidence correspondence M is given by the equations (a, b, c) ∈ M ⇐⇒ s (1, 0, a2 , a3 ) + t (0, 1, b2 , b3 ) ∈ Xc for all s,t ⇐⇒

∑ cα sα0 t α1 (s a2 + t b2 )α2 (s a3 + t b3 )α3 = 0 for all s,t

α i

⇐⇒ : ∑ sit 3−i Fi (a, b, c) = 0 for all s,t ⇐⇒ Fi (a, b, c) = 0 for 0 ≤ i ≤ 3. Note that the Fi are linear in the cα . Moreover, ci,3−i,0,0 occurs only in Fi for i = 0, . . . , 3, and it occurs there with coefﬁcient 1. So these equations can be written as ci,3−i,0,0 = Gi (a, b, c) for i = 0, . . . , 3, where the Gi depend only on those cα where α2 > 0 or α3 > 0. Therefore the variety A4 × P15 (with coordinates a2 , a3 , b2 , b3 , and all cα with α2 > 0 or α3 > 0) is isomorphic to an open subvariety of M, with the isomorphism given by the equations ci,3−i,0,0 = G(a, b, c). It follows that M has an open cover by afﬁne spaces A4 × A15 = A19 . Lemma 4.5.2. Again with notations as above, let (a, b, c) ∈ M be a point such that the ,F ) corresponding cubic surface Xc is smooth. Then the 4 × 4 matrix ∂(F0 ,F1 ,F2 ,b3 ) is invertible. ∂(a ,a ,b

2 3 2 3

Proof. After a change of coordinates we can assume for simplicity that a = b = 0. Then ∂ ∂ (∑ sit 3−i Fi )|(0,0,c) = fc (s,t, s a2 + t b2 , s a3 + t b3 )|(0,0,c) ∂a2 i ∂a2 =s ∂ fc (s,t, 0, 0). ∂x2

∂ fc ∂x2 (s,t, 0, 0), ∂ fc ∂x3 (s,t, 0, 0). ∂Fi ∂(a,b) (0, 0, c).

The (s,t)-coefﬁcients of this polynomial are the ﬁrst row in the matrix other rows are obviously s

∂ fc ∂x3 (s,t, 0, 0),

The

t

and t

So if the matrix

4.

∂Fi ∂(a,b) (0, 0, c)

Dimension

69

were not invertible, there would be a relation (λ2 s + µ2t) ∂ fc ∂ fc (s,t, 0, 0) + (λ3 s + µ3t) (s,t, 0, 0) = 0 ∂x2 ∂x3

∂ fc ∂x2 (s,t, 0, 0) and (x0 , x1 , 0, 0) ∈ P3 such

identically in s,t, with (λ2 , µ2 , λ3 , µ3 ) = (0, 0, 0, 0). But this means that

∂ fc ∂x3 (s,t, 0, 0) have a common linear ∂f ∂f that ∂xc (P) = ∂xc (P) = 0. But as 2 3

factor, i.e. there is a point P =

the line L0 lies in the cubic fc , we must have fc =

∂f ∂f x2 · g2 (x0 , x1 , x2 , x3 ) + x3 · g3 (x0 , x1 , x2 , x3 ) for some g2 , g3 . Hence ∂xc (P) = ∂xc (P) = 0 0 1 also, which means that P is a singular point of the cubic Xc . This is a contradiction to our assumptions.

Remark 4.5.3. By remark 4.4.11, lemma 4.5.2 means that locally (in the classical topology) around any point (a, b, c) ∈ M such that Xc is smooth, the coordinates a2 , a3 , b2 , b3 are determined uniquely in M by the cα . In other words, the projection map π : M → P19 is a local isomorphism (again in the classical topology!) around such a point (a, b, c) ∈ M. So the local picture looks as follows:

M

π IP 19

As the number of lines in a given cubic Xc is just the number of inverse image points of c ∈ P19 under this projection map, it follows that the number of lines on a smooth cubic surface is independent of the particular cubic chosen. Theorem 4.5.4. Every smooth cubic surface X ⊂ P3 contains exactly 27 lines. Proof. We have just argued that the number of lines on a smooth cubic surface does not depend on the surface, so we can pick a special one. We take the surface X given by 3 3 3 3 the equation f = x0 + x1 + x2 + x3 = 0 (which is smooth in characteristic not equal to 3). Up to a permutation of coordinates, every line in P3 can be written x0 = a2 x2 + a3 x3 , x1 = b2 x2 + b3 x3 . Substituting this in the equation f yields the conditions a3 + b3 = −1, 2 2 a3 + b3 3 3 2 a2 a3 = −1, = −b2 b3 , 2 2 a2 a3 = −b2 b2 . 3 (1) (2) (3) (4)

Assume that a2 , a3 , b2 , b3 are all non-zero. Then (3)2 /(4) gives a3 = −b3 , while (4)2 /(3) 2 2 yields a3 = −b3 . This is obviously a contradiction to (1) and (2). Hence at least one of the 3 3 a2 , a3 , b2 , b3 must be zero. Assume without loss of generality that a2 = 0. Then b3 = 0 and a3 = b3 = −1. This gives 9 lines by setting a3 = −ωi and b2 = −ω j for 0 ≤ i, j ≤ 2 and ω 3 2 a third root of unity. So by allowing permutations of the coordinates we ﬁnd that there are

70

Andreas Gathmann

exactly the following 27 lines on X: x0 + x1 ωi = x2 + x3 ω j = 0, 0 ≤ i, j ≤ 2, x0 + x2 ωi = x1 + x3 ω j = 0, 0 ≤ i, j ≤ 2, x0 + x3 ωi = x1 + x2 ω j = 0, 0 ≤ i, j ≤ 2. Remark 4.5.5. We will now study to a certain extent the conﬁguration of the 27 lines on a cubic surface, i.e. determine which of the lines intersect. Consider the special cubic X of the proof of theorem 4.5.4, and let L be the line L = {x0 + x1 = x2 + x3 = 0} in X. Then we can easily check that L meets exactly 10 of the other lines in X, namely x0 + x1 ωi = x2 + x3 ω j = 0, (i, j) = (0, 0) x0 + x2 = x1 + x3 = 0, x0 + x3 = x1 + x2 = 0. The same is true for every other line in X. In fact, the statement is also true for every smooth cubic surface, and not just for the special one that we have just considered. The proof of this is very similar to the proof above that the number of lines on a smooth cubic surface does not depend on the particular cubic chosen. Now let L1 and L2 be two disjoint lines on a smooth cubic surface X. We claim that there are exactly 5 lines on X that intersect both L1 and L2 . To show this, one can proceed in the same way as above: check the statement directly on a special cubic surface, and then show that it must then be true for all other smooth cubic surfaces as well. Proposition 4.5.6. Any smooth cubic surface in P3 is birational to P2 . Proof. By remark 4.5.5 there are two disjoint lines L1 , L2 ⊂ X. The following mutually inverse rational maps X L1 × L2 and L1 × L2 X show that X is birational to P1 × P1 2: and hence to P “X L1 × L2 ”: By exercise 3.5.1, for every point P not on L1 or L2 there is a unique line L(P) in P3 through L1 , L2 and P. Take the rational map P → (L1 ∩ L(P), L2 ∩ L(P)) that is obviously well-deﬁned away from L1 ∪ L2 . “L1 × L2 X”: Map any pair of points (P, Q) ∈ L1 × L2 to the third intersection point of X with the line PQ. This is well-deﬁned whenever PQ is not contained in X. Proposition 4.5.7. Any smooth cubic surface in P3 is isomorphic to P1 × P1 blown up in 5 (suitably chosen) points, or equivalently, to P2 blown up in 6 (suitably chosen) points. Proof. We will only sketch the proof. Let X be a smooth cubic surface, and let f : X L1 × L2 ∼ P1 × P1 be the rational map as in the proof of proposition 4.5.6. = First of all we claim that f is actually a morphism. To see this, note that there is a different description for f : if P ∈ X\L1 , let H be the unique plane in P3 that contains L1 and P, and let f2 (P) = H ∩ L2 . If one deﬁnes f1 (P) similarly, then f (P) = ( f1 (P), f2 (P)). Now if the point P lies on L1 , let H be the tangent plane to X at P, and again let f2 (P) = H ∩L2 . Extending f1 similarly, one can show that this extends f = ( f1 , f2 ) to a well-deﬁned morphism X → P1 × P1 on all of X. Now let us investigate where the inverse map P1 × P1 X is not well-deﬁned. As already mentioned in the proof of proposition 4.5.6, this is the case if the point (P, Q) ∈ L1 × L2 is such that PQ ⊂ X. In this case, the whole line PQ ∼ P1 will be mapped to (P, Q) = by f , and it can be checked that f is actually locally the blow-up of this point. By remark 4.5.5 there are exactly 5 such lines PQ on X. Hence f is the blow-up of P1 × P1 at 5 points.

4.

Dimension

71

By lemma 4.3.12 it then follows that f is also the blow-up of P2 in 6 suitably chosen points. Remark 4.5.8. It is interesting to see the 27 lines on a cubic surface X in the picture where one thinks of X as a blow-up of P2 in 6 points. It turns out that the 27 lines correspond to the following curves that we all already know (and that are all isomorphic to P1 ): • the 6 exceptional hypersurfaces, • the strict transforms of the 6 = 15 lines through two of the blown-up points, 2 • the strict transforms of the 6 = 6 conics through ﬁve of the blown-up points (see 5 exercise 3.5.8). In fact, it is easy to see by the above explicit description of the isomorphism of X with the blow-up of P2 that these curves on the blow-up actually correspond to lines on the cubic surface. It is also interesting to see again in this picture that every such “line” meets 10 of the other “lines”, as mentioned in remark 4.5.5: • Every exceptional hypersurface intersects the 5 lines and the 5 conics that pass through this blown-up point. • Every line through two of the blown-up points meets – the 2 exceptional hypersurfaces of the blown-up points, – the 4 = 6 lines through two of the four remaining points, 2 – the 2 conics through the four remaining points and one of the blown-up points. • Every conic through ﬁve of the blown-up points meets the 5 exceptional hypersurfaces at these points, as well as the 5 lines through one of these ﬁve points and the remaining point. 4.6. Exercises. Exercise 4.6.1. Let X,Y ⊂ Pn be projective varieties. Show that X ∩ Y is not empty if dim X + dimY ≥ n. On the other hand, give an example of a projective variety Z and closed subsets X,Y ⊂ Z / with dim X + dimY ≥ dim Z and X ∩Y = 0. (Hint: Let H1 , H2 be two disjoint linear subspaces of dimension n in P2n+1 , and consider X ⊂ Pn ∼ H1 ⊂ P2n+1 and Y ⊂ Pn ∼ H2 ⊂ P2n+1 as subvarieties of P2n+1 . Show that the = = join J(X,Y ) ⊂ P2n+1 of exercise 3.5.7 has dimension dim X + dimY + 1. Then construct X ∩Y as a suitable intersection of J(X,Y ) with n + 1 hyperplanes.) Exercise 4.6.2. (This is a generalization of corollary 4.2.7). Let f : X → Y be a morphism of varieties. Show that there is a non-empty open subset U of Y such that every component of the ﬁber f −1 (P) has dimension dim X − dimY for all P ∈ U. (Hint: You can assume X ⊂ An and Y ⊂ Am to be afﬁne. By considering the graph (P, f (P)) ∈ An+m , reduce to the case where f : An+1 → An is the projection map.) Exercise 4.6.3. Let f : X → Y be a morphism of varieties, and let Z ⊂ X be a closed subset. Assume that f −1 (P) ∩ Z is irreducible and of the same dimension for all P ∈ Y . Use exercise 4.6.2 to prove that then Z is irreducible too. (This is a quite useful criterion to check the irreducibility of closed subsets.) Show by example that the conclusion is in general false if the f −1 (P) ∩ Z are irreducible but not all of the same dimension. Exercise 4.6.4. Let X be a variety, and let Y ⊂ X a closed subset. For every element in an ˜ open afﬁne cover {Ui } of X, let Vi = Ui ∩ Y , and let Ui be the blow-up of Ui at Vi . Show

72

Andreas Gathmann

˜ ˜ that the spaces Ui can be glued together to give a variety X. (This variety is then called the blow-up of X at Y .) Exercise 4.6.5. A quadric in Pn is a projective variety in Pn that can be given as the zero locus of a quadratic polynomial. Show that every quadric in Pn is birational to Pn−1 . Exercise 4.6.6. Show that for four general lines L1 , . . . , L4 ⊂ P3 , there are exactly two lines in P3 intersecting all the Li . (This means: the subset of G(1, 3)4 of all (L1 , . . . , L4 ) such that there are exactly two lines in P3 intersecting L1 , . . . , L4 is dense. You may want to use the result of exercise 3.5.4 (iii) that G(1, 3) is a quadric in P5 .) Exercise 4.6.7. Let P1 = (1 : 0 : 0), P2 = (0 : 1 : 0), P3 = (0 : 0 : 1) ∈ P2 , and let U = P2 \{P1 , P2 , P3 }. Consider the morphism f : U → P2 , (a0 : a1 : a2 ) → (a1 a2 : a0 a2 : a0 a1 ). (i) Show that there is no morphism F : P2 → P2 extending f . ˜ (ii) Let P2 be the blow-up of P2 in the three points P1 , P2 , P3 . Show that there is an ˜ ˜ isomorphism f˜ : P2 → P2 extending f . This is called the Cremona transformation. Exercise 4.6.8. Let X ⊂ An be an afﬁne variety. For every f ∈ k[x0 , . . . , xn ] denote by f in the initial terms of f , i.e. the terms of f of the lowest occurring degree (e.g. if f = 2 2 x2 + 3x1 x3 − x2 x3 then the lowest occurring degree in f is 2, so the initial terms are the 2 terms of degree 2, namely f in = x2 + 3x1 x3 ). Let I(X)in = { f in ; f ∈ I(X)} be the ideal of the initial terms in I(X). ˜ Now let π : X → X be the blow-up of X in the origin {0} = Z(x1 , . . . , xn ). Show that the exceptional hypersurface π−1 (0) ⊂ Pn is precisely the projective zero locus of the homogeneous ideal I(X)in . Exercise 4.6.9. Let X ⊂ An be an afﬁne variety, and let P ∈ X be a point. Show that the coordinate ring A(CX,P ) of the tangent cone to X at P is equal to ⊕k≥0 I(P)k /I(P)k+1 , where I(P) is the ideal of P in A(X). Exercise 4.6.10. Let X ⊂ An be an afﬁne variety, and let Y1 ,Y2 X be irreducible, closed ˜ subsets, no-one contained in the other. Let X be the blow-up of X at the (possibly non˜ radical, see exercise 1.4.1) ideal I(Y1 ) + I(Y2 ). Then the strict transforms of Y1 and Y2 on X are disjoint. Exercise 4.6.11. Let C ⊂ P2 be a smooth curve, given as the zero locus of a homogeneous polynomial f ∈ k[x0 , x1 , x2 ]. Consider the morphism ϕC : C → P2 , P → ∂f ∂f ∂f (P) : (P) : (P) . ∂x0 ∂x1 ∂x2

The image ϕC (C) ⊂ P2 is called the dual curve to C. (i) Find a geometric description of ϕ. What does it mean geometrically if ϕ(P) = ϕ(Q) for two distinct points P, Q ∈ C ? (ii) If C is a conic, prove that its dual ϕ(C) is also a conic. (iii) For any ﬁve lines in P2 in general position (what does this mean?) show that there is a unique conic in P2 that is tangent to these ﬁve lines. (Hint: Use exercise 3.5.8.) Exercise 4.6.12. Resolve the singularities of the following curves by subsequent blow-ups of the singular points. This means: starting with the given curve C, blow up all singular points of C, and replace C by its strict transform. Continue this process until the resulting curve is smooth. Also, describe the singularities that occur in the intermediate steps of the resolution process.

4.

Dimension

73

(i) C = {(x, y) ; x2 − x4 − y4 = 0} ⊂ A2 , (ii) C = {(x, y) ; y3 − x5 = 0} ⊂ A2 , (iii) C = {(x, y) ; y2 − xk = 0} ⊂ A2 , k ∈ N. Exercise 4.6.13. Show that “a general hypersurface in Pn is smooth”. In other words, for n+d any given d we can consider P( d )−1 as the “space of all hypersurfaces of degree d in Pn ”, by associating to any hypersurface { f (x0 , . . . , xn ) = 0} ⊂ Pn with f homogeneous of degree d the projective vector of all n+d coefﬁcients of f . Then show that the subset of d n+d P( d )−1 corresponding to smooth hypersurfaces is non-empty and open. Exercise 4.6.14. (This is a generalization of exercises 3.5.8 and 4.6.11 (iii).) For i = 0, . . . , 5, determine how many conics there are in P2 that are tangent to i given lines and in addition pass through 5 − i given points.

74

Andreas Gathmann

5. S CHEMES

To any commutative ring R with identity we associate a locally ringed space called Spec R, the spectrum of R. Its underlying set is the set of prime ideals of R, so if R is the coordinate ring of an afﬁne variety X over an algebraically closed ﬁeld, then Spec R as a set is the set of non-empty closed irreducible subvarieties of X. Moreover, in this case the open subsets of Spec R are in one-to-one correspondence with the open subsets of X, and the structure sheaves of Spec R and X coincide via this correspondence. A morphism of locally ringed spaces is a morphism of ringed spaces that respects the maximal ideals of the local rings. Locally ringed spaces of the form Spec R are called afﬁne schemes; locally ringed spaces that are locally of the form Spec R are called schemes. Schemes are the fundamental objects of study in algebraic geometry. Prevarieties correspond exactly to those schemes that are reduced, irreducible, and of ﬁnite type over an algebraically closed ﬁeld. For any two morphisms of schemes X → S and Y → S there is a ﬁber product X ×S Y ; this is a scheme such that giving morphisms Z → X and Z → Y that commute with the given morphisms to S is “the same” as giving a morphism Z → X ×S Y . If X and Y are prevarieties over k and we take S = Spec k, we get back our old notion of the product X ×Y of prevarieties. For any graded ring R there is a scheme Proj R whose points are the homogeneous prime ideals of R that do not contain the irrelevant ideal. This construction generalizes our earlier construction of projective varieties; if R is the homogeneous coordinate ring of a projective variety X over an algebraically closed ﬁeld then Proj R “is” just the projective variety X.

5.1. Afﬁne schemes. We now come to the deﬁnition of schemes, which are the main objects of study in algebraic geometry. The notion of schemes extends that of prevarieties in a number of ways. We have already met several instances where an extension of the category of prevarieties could be useful: • We deﬁned a prevariety to be irreducible. Obviously, it makes sense to also consider reducible spaces. In the case of afﬁne and projective varieties we called them algebraic sets, but we did not give them any further structure or deﬁned regular functions and morphisms of them. Now we want to make reducible spaces into full-featured objects of our category. • At present we have no geometric objects corresponding to non-radical ideals in k[x1 , . . . , xn ], or in other words to coordinate rings with nilpotent elements. These non-radical ideals pop up naturally however: e.g. we have seen in exercise 1.4.1 that intersections of afﬁne varieties correspond to sums of their ideals, modulo taking the radical. It would seem more natural to deﬁne the intersection X1 ∩ X2 of two afﬁne varieties X1 , X2 ⊂ An to be a geometric object associated to the ideal I(X1 ) + I(X2 ) ⊂ k[x1 , . . . , xn ]. This was especially obvious when we discussed blow-ups: blowing up X1 ∩ X2 in An “separates” X1 and X2 (if none of these two ˜ ˜ sets is contained in the other), i.e. their strict transforms X1 and X2 are disjoint ˜ n , but this is only true if we blow-up at the ideal I(X1 ) + I(X2 ) and not at its in A radical (see exercise 4.6.10). • Recall that by lemma 2.3.7 and remark 2.3.14 we have a one-to-one correspondence between afﬁne varieties over k and ﬁnitely generated k-algebras that are domains, both modulo isomorphism. We have just seen that we should drop the condition on the k-algebra to be a domain. We can go even further and also drop the condition that it is ﬁnitely generated — then we would expect to arrive at “inﬁnite-dimensional” objects. Moreover, it turns out that we do not even need a k-algebra to do geometry; it is sufﬁcient to start with any commutative ring with

5.

Schemes

75

identity, i.e. we do not have to have a ground ﬁeld. This can be motivated by noting that most constructions we made with the coordinate ring of a variety — deﬁning the structure sheaf, setting up correspondences between points and maximal ideals, and so on — actually only used the ring structure of the coordinate ring, and not the k-algebra structure. All these generalizations are included in the deﬁnition of a scheme. Note that they apply already to afﬁne varieties; so we will start by deﬁning an afﬁne scheme to be “an afﬁne variety generalized as above”. Later we will then say that a scheme is an object that looks locally like an afﬁne scheme, just as we did it in the case of prevarieties. We are now ready to construct from any ring R (which will always mean a commutative ring with identity) an afﬁne scheme, which will be a ringed space and which will be denoted Spec R, the spectrum of R. Deﬁnition 5.1.1. Let R be a ring (commutative with identity, as always). We deﬁne Spec R to be the set of all prime ideals of R. (As usual, R itself does not count as a prime ideal, but (0) does if R is a domain.) We call Spec R the spectrum of R, or the afﬁne scheme associated to R. For every p ∈ Spec R, i.e. p ⊂ R a prime ideal, let k(p) be the quotient ﬁeld of the domain R/p. Remark 5.1.2. Let X = Spec R be an afﬁne scheme. We should think of X as the analogue of an afﬁne variety, and of R as the analogue of its coordinate ring. Remark 5.1.3. Any element f ∈ R can be considered to be a “function” on Spec R in the following sense: for p ∈ Spec R, denote by f (p) the image of f under the composite map R → R/p → k(p). We call f (p) the value of f at the point p. Note that these values will in general lie in different ﬁelds. If R = k[x1 , . . . , xn ]/I(X) is the coordinate ring of an afﬁne variety X and p is a maximal ideal (i.e. a point in X), then k(p) = k and the value of an element f ∈ R as deﬁned above is equal to the value of f at the point corresponding to p in the classical sense. If p ⊂ R is not maximal and corresponds to some subvariety Y ⊂ X, the value f (p) lies in the function ﬁeld K(Y ) and can be thought of as the restriction of the function f to Y . Example 5.1.4. (i) If k is a ﬁeld, then Spec k consists of a single point (0). (ii) The space Spec C[x] (that will correspond to the afﬁne variety A1 over C) contains a point (x − a) for every a ∈ A1 , together with a point (0) corresponding to the subvariety A1 . (iii) More generally, if R = A(X) is the coordinate ring of an afﬁne variety X over an algebraically closed ﬁeld, then the set Spec R contains a point for every closed subvariety of X (as subvarieties correspond exactly to prime ideals). This afﬁne scheme Spec R will be the analogue of the afﬁne variety X. So an afﬁne scheme has “more points” than the corresponding afﬁne variety: we have enlarged the set by throwing in an additional point for every closed subvariety Y of X. This point is usually called the generic point (or general point) of Y . In other words, in the scheme corresponding to an afﬁne variety with coordinate ring R we will have a point for every prime ideal in R, and not just for every maximal ideal. These additional points are sometimes important, but quite often one can ignore this fact. Many textbooks will even adopt the convention that a point of a scheme is always meant to be a point in the old geometric sense (i.e. a maximal ideal). (iv) In contrast to (ii), the afﬁne scheme Spec R[x] contains points that are not of the form (x − a) or (0), e.g. (x2 + 1) ∈ Spec R[x]. (v) The afﬁne scheme Spec Z contains an element for every prime number, and in addition the generic point (0).

76

Andreas Gathmann

So far we have deﬁned Spec R as a set. This is not particularly interesting, so let us move on and make Spec R into a topological space. This is done in the same way as for afﬁne varieties. Deﬁnition 5.1.5. Let R be a ring. For every subset S ⊂ R, we deﬁne the zero locus of S to be the set Z(S) := {p ∈ Spec R ; f (p) = 0 for all f ∈ S} ⊂ Spec R, where f (p) is the value of f at p as in remark 5.1.3. (Obviously, S and (S) deﬁne the same zero locus, so we will usually only consider zero loci of ideals.) Remark 5.1.6. By the deﬁnition of the value of an element f ∈ R at a point p ∈ Spec R, we can also write the deﬁnition of the zero locus as Z(S) = {p ∈ Spec R ; f ∈ p for all f ∈ S} = {p ∈ Spec R ; p ⊃ S}. Lemma 5.1.7. Let R be a ring. (i) If {Ii } is a family of ideals of R then i Z(Ii ) = Z(∑i Ii ) ⊂ Spec R. (ii) If I1 , I2 ⊂ R then Z(I1 ) ∪ Z(I2 ) = Z(I1 I2 ) ⊂ Spec R. √ √ (iii) If I1 , I2 ⊂ R then Z(I1 ) ⊂ Z(I2 ) if and only if I2 ⊂ I1 .

T

Proof. The proof is literally the same as in the case of afﬁne algebraic sets. Hence we can deﬁne a topology on Spec R by taking the subsets of the form Z(S) as the closed subsets. In particular, this deﬁnes the notions of irreducibility and dimension for Spec R, as they are purely topological concepts. Remark 5.1.8. Note that points p in Spec R are not necessarily closed: in fact, {p} = Z(p) = {q ∈ Spec R ; q ⊃ p}. This is equal to {p} only if p is maximal. Hence the closed points of Spec R correspond to the points of an afﬁne variety in the classical sense. The other points are just generic points of irreducible closed subsets of Spec R, as already mentioned in example 5.1.4. Example 5.1.9. The motivation for the name “generic point” can be seen from the following example. Let k be an algebraically closed ﬁeld, and let R = Spec k[x1 , x2 ] be the afﬁne scheme corresponding to A2 . Consider Z(x2 ) ⊂ Spec R, which “is” just the x1 -axis; so its complement Spec R\Z(x2 ) should be the set of points that do not lie on the x1 -axis. But note that the element p = (x1 ) is contained in Spec R\Z(x2 ), although the zero locus of x1 , namely the x2 -axis, does intersect the x1 -axis. So the geometric way to express the fact that (x1 ) ∈ Spec R\Z(x2 ) is to say that the generic point of the x2 -axis does not lie on the x1 -axis. Remark 5.1.10. Let R be a ring, let X = Spec R, and let f ∈ R. As in the case of afﬁne varieties, we call X f := X\Z( f ) the distinguished open subset associated to f . Note that any open subset of X is a (not necessarily ﬁnite) union of distinguished open subsets. This is often expressed by saying that the distinguished open subsets form a base of the topology of X. Now we come to the deﬁnition of the structure sheaf of Spec R. Recall that in the case of an afﬁne variety X, we ﬁrst deﬁned the local ring OX,P of the functions regular at a point P ∈ X to be the localization of A(X) at the maximal ideal corresponding to P, and then said that an element in OX (U) for an open subset U ⊂ X is a function that is regular at every point P ∈ U. We could accomplish that in the case of varieties just by intersecting the local rings OX,P , as they were all contained in the function ﬁeld K(X). But in the case of a general afﬁne scheme Spec R the various local rings Rp for p ∈ Spec R do not lie inside

5.

Schemes

77

some big space, so we cannot just take their intersection. The way around this problem is to say that an element in OX (U) (for X = Spec R and U ⊂ X open) is given by a collection of elements in the various local rings Rp for all p ∈ U, and require that these elements can locally be written as quotients of elements of R (recall that we had a similar condition for afﬁne varieties in lemma 2.1.8): Deﬁnition 5.1.11. Let R be a ring, and let X = Spec R. For every open subset U ⊂ X we deﬁne OX (U) to be

OX (U) := {ϕ = (ϕp )p∈U with ϕp ∈ Rp for all p ∈ U

such that “ϕ is locally of the form

f g

for f , g ∈ R”}

= {ϕ = (ϕp )p∈U with ϕp ∈ Rp for all p ∈ U such that for every p ∈ U there is a neighborhood V in U and f , g ∈ R with g ∈ q and ϕq = /

f g

∈ Rq for all q ∈ V .}

As the conditions imposed on the elements of OX (U) are local, it is easy to verify that this deﬁnes a sheaf OX on X = Spec R. The ﬁrst thing to do is to check that this sheaf has the properties that we expect from the case of afﬁne varieties (see deﬁnition 2.1.5, remark 2.1.6, and proposition 2.1.10). Proposition 5.1.12. Let R be a ring and X = Spec R. (i) For any p ∈ X the stalk OX,p of the sheaf OX is isomorphic to the local ring Rp . (ii) For any f ∈ R, the ring OX (X f ) is isomorphic to the localized ring R f . In particular, OX (X) = R. Proof. (i): There is a well-deﬁned ring homomorphism ψ : OX,p → Rp , (U, ϕ) → ϕp . We have to show that ψ is a bijection. ψ is surjective: Any element of Rp has the form

f g f g

with f , g ∈ R and g ∈ p. The function /

f is well-deﬁned on Xg , so (Xg , g ) deﬁnes an element in OX,p that is mapped by ψ to the given element. ψ is injective: Let ϕ1 , ϕ2 ∈ OX (U) for some neighborhood U of p, and assume that (ϕ1 )p = (ϕ2 )p . We have to show that ϕ1 and ϕ2 coincide in a neighborhood of p, so that they deﬁne the same element in OX,p . By shrinking U if necessary we may assume that f ϕi = gii on U for i = 1, 2, where fi , gi ∈ R and gi ∈ p. As ϕ1 and ϕ2 have the same image / in Rp , it follows that h( f1 g2 − f2 g1 ) = 0 in R for some h ∈ p. Therefore we also have / f1 f = g2 in every local ring Rq such that g1 , g2 , h ∈ q. But the set of such q is the open set / g1 2 Xg1 ∩ Xg2 ∩ Xh , which contains p. Hence ϕ1 = ϕ2 on some neighborhood of p, as required. (ii): There is a well-deﬁned ring homomorphism g g ψ : R f → OX (X f ), r → r f f

(i.e. we map

g fr

to the element of OX (X f ) that assigns to any p the image of

g1 f r1 ) g2 f r2 ),

g fr

in Rp ).

ψ is injective: Assume that ψ( = ψ( i.e. for every p ∈ X f there is an element h ∈ p such that h(g1 f r2 − g2 f r1 ) = 0. Let I ⊂ R be the annihilator of g1 f r2 − g2 f r1 , then we / / have just shown that I ⊂ p, as h ∈ I but h ∈ p. This holds for every p ∈ X f , so Z(I) ∩ X f = 0, / or in other words Z(I) ⊂ Z( f ). By lemma 5.1.7 (iii) this means that f r ∈ I for some r, so f r (g1 f r2 − g2 f r1 ) = 0, hence fgr11 = fgr22 in R f . ψ is surjective: Let ϕ ∈ OX (X f ). By deﬁnition, we can cover X f by open sets Ui on which ϕ is represented by a quotient gii , with fi ∈ p for all p ∈ Ui , i.e. Ui ⊂ X fi . As the open / f

78

Andreas Gathmann

subsets of the form Xhi form a base for the topology of X, we may assume that Ui = Xhi for some hi . We want to show that we can assume fi = hi . In √ as Xhi ⊂ X fi , i.e. by taking fact, complements we get Z( fi ) ⊂ Z(hi ), and therefore hi ∈ fi by lemma 5.1.7 (iii). Hence hr = c fi , so gii = cgi . Replacing hi by hr (as Xhi = Xhr ) and gi by cgi we can assume that i i f hr i i X f is covered by open subsets of the form Xhi , and that ϕ is represented by gii on Xhi . h Next we prove that X f canT actually be covered by ﬁnitely many such Xhi . Indeed, X f ⊂ S i Xhi if and only if Z( f ) ⊃ i Z(hi ) = Z(∑(hi )). By lemma 5.1.7 (iii) this is equivalent to saying that f r ∈ ∑(hi ) for some r. But this means that f r can be written as a ﬁnite sum f r = ∑ bi hi . Hence we can assume that we have only ﬁnitely many hi . g g On Xhi ∩Xh j = Xhi h j , we have two elements hii and h jj representing ϕ, so by the injectivity proven above it follows that gii = h jj in Rhi h j , hence (hi h j )n (gi h j − g j hi ) = 0 for some n. h As we have only ﬁnitely many hi , we may pick one n that works for all i, j. Now replace gi by gi hn and hi by hn+1 for all i, then we still have ϕ represented by gii on Xhi , and moreover i i h gi h j − g j hi = 0 for all i, j. Now write f r = ∑ bi hi as above, which is possible since the Xhi cover X f . Let g = ∑ bi gi . Then for every j we have gh j = ∑ bi gi h j = ∑ bi hi g j = f r g j ,

i i g

so

f g

=

hj gj

on Xh j . Hence ϕ is represented on X f by

g fr

∈ R f , i.e. ψ is surjective.

Remark 5.1.13. Note that a regular function is in general no longer determined by its values on points. For example, let R = k[x]/(x2 ) and X = Spec R. Then X has just one point (x). On this point, the function x ∈ R = OX (X) takes the value 0 = x ∈ (k[x]/(x2 ))/(x) = k. In particular, the functions 0 and x have the same values at all points of X, but they are not the same regular function. 5.2. Morphisms and locally ringed spaces. As in the case of varieties, the next step after deﬁning regular functions on an afﬁne scheme is to deﬁne morphisms between them. Of course one is tempted to deﬁne a morphism f : X → Y between afﬁne schemes to be a morphism of ringed spaces as in deﬁnition 2.3.1, but recall that for this deﬁnition to work we needed a notion of pull-back f ∗ of regular functions. In the case of varieties we got this by requiring that the structure sheaves be sheaves of k-valued functions, so that a settheoretic pull-back exists. But this is not possible for schemes, as we do not have a ground ﬁeld, and the values ϕ(p) of a regular function ϕ lie in unrelated rings. Even worse, we have seen already in example 5.1.13 that a regular function is not determined by its values on points. The way out of this dilemma is to make the pull-back maps f ∗ : OY (U) → OX ( f −1 (U)) part of the data required to deﬁne a morphism. Hence we say that a morphism f : X → Y between afﬁne schemes is given by a continuous map f : X → Y between the underlying ∗ topological spaces, together with pull-back maps f ∗ = fU : OY (U) → OX ( f −1 (U)) for every open subset U ⊂ Y . Of course we need some compatibility conditions among the ∗ ∗ fU . The most obvious one is compatibility with the restriction maps, i.e. fV ◦ ρU,V = ∗ . But we also need some sort of compatibility between the f ∗ and the ρ f −1 (U), f −1 (V ) ◦ fU U ∗ continuous map f . To explain this condition, note that the maps fU give rise to a map between the stalks

∗ fP : OY, f (P) → OX,P , (U, ϕ) → ( f −1 (U), f ∗ ϕ)

for every point P ∈ X (this is easily seen to be well-deﬁned). These stalks are local rings, call their maximal ideals mY, f (P) and mX,P , respectively. Now the fact that f maps P

5.

Schemes

79

to f (P) should be reﬂected on the level of the pull-back maps f ∗ by the condition that ∗ ( fP )−1 (mX,P ) = mY, f (P) . This leads to the following deﬁnition. Deﬁnition 5.2.1. A locally ringed space is a ringed space (X, OX ) such that at each point P ∈ X the stalk OX,P is a local ring. The maximal ideal of OX,P will be denoted by mX,P , and the residue ﬁeld OX,P /mX,P will be denoted k(P). A morphism of locally ringed spaces from (X, OX ) to (Y, OY ) is given by the following data: • a continuous map f : X → Y , ∗ • for every open subset U ⊂ Y a ring homomorphism fU : OY (U) → OX ( f −1 (U)),

∗ ∗ such that fV ◦ ρU,V = ρ f −1 (U), f −1 (V ) ◦ fU for all V ⊂ U ⊂ Y (i.e. the f ∗ are compatible with ∗ )−1 (m ∗ the restriction maps) and ( fP X,P ) = mY, f (P) , where the f P : OY, f (P) → OX,P are the maps induced on the stalks, as explained above. We will often omit the index of the various pull-back maps f ∗ if it is clear from the context on which spaces they act. A morphism of afﬁne schemes is a morphism as locally ringed spaces.

The following proposition is the analogue of lemma 2.3.7. It shows that deﬁnition 5.2.1 was “the correct one”, because it gives us ﬁnally what we want. Proposition 5.2.2. Let R, S be rings, and let X = Spec R and Y = Spec S the corresponding afﬁne schemes. There is a one-to-one correspondence between morphisms X → Y and ring homomorphisms S → R. Proof. If ψ : S → R is a ring homomorphism, we deﬁne a map f : X → Y by f (p) = ψ−1 (p). For every ideal I ⊂ S it follows that f −1 (Z(I)) = Z(ψ(I)), so f is continuous. For each p ∈ Spec R, we can localize ψ to get a homomorphism of local rings ψp : OY, f (p) = Sψ−1 (p) → Rp = OX,p satisfying the condition ψ−1 (mX,p ) = mY, f (p) . By deﬁnition of the structure p sheaf, this gives homomorphisms of rings f ∗ : OY (U) → OX ( f −1 (U)), and by construction ∗ fp = ψp , so we get a morphism of locally ringed spaces. If f : X → Y is a morphism, we get a ring homomorphism f ∗ : S = OY (Y ) → OX (X) = R by proposition 5.1.12 (ii). By the above this again determines a morphism g : X → Y . We leave it as an exercise to check that the various compatibility conditions imply that f = g. Example 5.2.3. Let X = Spec R be an afﬁne scheme. If I ⊂ R is an ideal, then we can form the afﬁne scheme Y = Spec(R/I), and the ring homomorphism R → R/I gives us a morphism Y → X. Note that the prime ideals of R/I are exactly the ideals p ⊂ R with p ⊃ I, so the map Y → X is an inclusion with image Z(I). So we can view Y as an afﬁne “closed subscheme” of X. For a precise deﬁnition of this concept see example 7.2.10. Now let Y1 = Spec(R/I1 ) and Y2 = Spec(R/I2 ) be closed subschemes of X. We deﬁne the intersection scheme Y1 ∩Y2 in X to be Y1 ∩Y2 = Spec R/(I1 + I2 ). For example, let X = Spec C[x1 , x2 ], Y1 = Spec C[x1 , x2 ]/(x2 ), Y2 = Spec C[x1 , x2 ]/(x2 − 2 x1 + a2 ) for some a ∈ C. Then the intersection scheme Y1 ∩Y2 is Spec C[x1 ]/((x1 − a)(x1 + a)). For a = 0 we have C[x1 ]/((x1 −a)(x1 +a)) ∼ C[x1 ]/(x1 −a)×C[x1 ]/(x1 +a) ∼ C×C, = = so Y1 ∩ Y2 is just the disjoint union of the two points (a, 0) and (−a, 0) in C2 . For a = 0 2 however we have Y1 ∩ Y2 = Spec C[x1 ]/(x1 ), which has only one point (0, 0). But in all cases the ring C[x1 ]/((x1 − a)(x1 + a)) has dimension 2 as a vector space over C. We say that Y1 ∩Y2 is a “scheme of length 2”, which consists either of two distinct points of length 1 each, or of one point of length (i.e. multiplicity) 2. Note also that there is always a unique line in A2 through Y1 ∩Y2 , even in the case a = 0 where the scheme has only one geometric point. This is because the scheme Y1 ∩ Y2 = Spec C[x1 , x2 ]/(x2 , (x1 − a)(x1 + a)) is a subscheme of the line L = Spec C[x1 , x2 ]/(c1 x1 +

80

Andreas Gathmann

c2 x2 ) if and only if (c1 x1 + c2 x2 ) ⊂ (x2 , (x1 − a)(x1 + a)), which is the case only if c1 = 0. 2 In particular, the x1 -axis is the only line in A2 that contains Spec C[x1 , x2 ]/(x2 , x1 ). One can therefore think of this scheme as “the origin together with a tangent direction along the x1 -axis”.

x2 Y2 Y1 −a a x1 Y1 a =0 x2 Y2 x1

a =0

Example 5.2.4. Again let Y1 = Spec(R/I1 ) and Y2 = Spec(R/I2 ) be closed subschemes of of the afﬁne scheme X = Spec R. Note that for afﬁne varieties the ideal of the union of two closed subsets equals the intersection of their ideals (see exercise 1.4.1 (i)). So scheme-theoretically we just deﬁne the union Y1 ∪Y2 to be Spec R/(I1 ∩ I2 ). The following lemma is the scheme-theoretic analogue of lemma 2.3.16. Lemma 5.2.5. Let X = Spec R be an afﬁne scheme, and let f ∈ R. Then the distinguished open subset X f is the afﬁne scheme Spec R f . Proof. Note that both X f and Spec R f have the description {p ∈ X ; f ∈ p}. So it only / remains to be checked that the structure sheaves on X f and Spec R f agree. Now let g ∈ R and consider the distinguished open subset X f g = (Spec R f )g . By proposition 5.1.12 (ii) we have

OX f (X f g ) = OX (X f g ) = R f g

and OSpec R f ((Spec R f )g ) = (R f )g = R f g . So the rings of regular functions are the same for X f and Spec R f on every distinguished open subset. But every open subset is the intersection of such distinguished opens, so the rings of regular functions must be the same on every open subset. 5.3. Schemes and prevarieties. Having deﬁned afﬁne schemes and their morphisms, we can now deﬁne schemes as objects that look locally like afﬁne schemes — this is in parallel to the deﬁnition 2.4.1 of prevarieties. Deﬁnition 5.3.1. A scheme is a locally ringed space (X, OX ) that can be covered by open subsets Ui ⊂ X such that (Ui , OX |Ui ) is isomorphic to an afﬁne scheme Spec Ri for all i. A morphism of schemes is a morphism as locally ringed spaces. Remark 5.3.2. From the point of view of prevarieties, it would seem more natural to call the objects deﬁned above preschemes, and then say that a scheme is a prescheme having the “Hausdorff” property, i.e. a prescheme with closed diagonal (see deﬁnition 2.5.1 and lemma 2.5.3). This is in fact the terminology of [M1], but nowadays everyone seems to adopt the deﬁnition that we gave above, and then say that a scheme having the “Hausdorff property” is a separated scheme. From our deﬁnitions we see that prevarieties are in a sense special cases of schemes — if we have an afﬁne variety X = Z(I) ⊂ An with I ⊂ k[x1 , . . . , xn ] an ideal, the scheme Spec A(X) corresponds to X (where A(X) = k[x1 , . . . , xn ] is the coordinate ring of X); and any glueing along isomorphic open subsets that can be done in the category of prevarieties can be done equally well for the corresponding schemes. Hence we would like to say that every prevariety is a scheme. In the strict sense of the word this is not quite true however,

5.

Schemes

81

because the topological space of a scheme contains a point for every irreducible closed subset, whereas the topological space of a prevariety consists only of the geometric points in the classical sense (i.e. the closed points). But of course there is a natural way to consider every prevariety as a scheme, by throwing in additional generic points for every irreducible closed subset. We give the precise statement and leave its proof as an exercise: Proposition 5.3.3. Let k be an algebraically closed ﬁeld, and let X be a prevariety over k. Let Xsch be the space of all non-empty closed irreducible subsets of X. Then Xsch is a scheme in a natural way. The open subsets of X correspond bijectively to the open subsets of Xsch , and for every open subset U of X (which can then also be considered as an open subset of Xsch ) we have OXsch (U) = OX (U). Every morphism X → Y of prevarieties over k extends to a morphism Xsch → Ysch of schemes in a natural way. Let us now investigate the properties of schemes that arise from prevarieties in this way. As we have mentioned already, the glueing of schemes from afﬁne schemes is exactly the same as that of prevarieties from varieties. Hence the special properties of schemes that come from prevarieties can already be seen on the level of afﬁne schemes. We have also seen above that in an afﬁne scheme Spec R the ring R corresponds to what is the coordinate ring A(X) of an afﬁne variety. Moreover we know by remark 2.3.14 that the coordinate ring of an afﬁne variety is a ﬁnitely generated k-algebra that is a domain. So we have to write down conditions on a scheme that reﬂect the property that its local patches Spec R are not made from arbitrary rings, but rather from ﬁnitely generated k-algebras that are domains. Deﬁnition 5.3.4. Let Y be a scheme. A scheme over Y is a scheme X together with a morphism X → Y . A morphism of schemes X1 , X2 over Y is a morphism of schemes X1 → X2 such that / X2 X1 @@ ~ @@ ~ @@ ~~ @ ~~ ~ Y commutes. If R is a ring, a scheme over R is a scheme over Spec R. A scheme X over Y is said to be of ﬁnite type over Y if there is a covering of Y by open afﬁne subsets Vi = Spec Bi such that f −1 (Vi ) can be covered by ﬁnitely many open afﬁnes Ui, j = Spec Ai, j , where each Ai, j is a ﬁnitely generated Bi -algebra. In particular, a scheme X over a ﬁeld k is of ﬁnite type over k if it can be covered by ﬁnitely many open afﬁnes Ui = Spec Ai , where each Ai is a ﬁnitely generated k-algebra. A scheme X is called reduced if the rings OX (U) have no nilpotent elements for all open subsets U ⊂ X. Now it is obvious what these conditions mean for an afﬁne scheme Spec R: • Spec R is a scheme over k if and only if we are given a morphism k → R, i.e. if R is a k-algebra. Moreover, a morphism Spec R → Spec S is a morphism of schemes over k if and only if the corresponding ring homomorphism S → R is a morphism of k-algebras. • Spec R is of ﬁnite type over k if and only if R is a ﬁnitely generated k-algebra. • Spec R is reduced and irreducible if and only if f · g = 0 in R implies f = 0 or g = 0, i.e. if and only if R is a domain. To see this, assume that f · g = 0, but f = 0 and g = 0. If f and g are the same up to a power, then R is not nilpotent-free, so Spec R is not reduced. Otherwise, we get a decomposition of Spec R into two proper closed subsets Z( f ) and Z(g), so Spec R is not irreducible.

82

Andreas Gathmann

As glueing afﬁne patches is allowed for varieties in the same way as for schemes, we get the following result: Proposition 5.3.5. Let k be an algebraically closed ﬁeld. There is a one-to-one correspondence between prevarieties over k (and their morphisms) and reduced, irreducible schemes of ﬁnite type over k (and their morphisms). Hence, from now on a prevariety over k will mean a reduced and irreducible scheme of ﬁnite type over k. Remark 5.3.6. As in the case of prevarieties, schemes and morphisms of schemes can (almost by deﬁnition) be glued together. As for glueing schemes lemma 2.4.7 holds in the same way (except that one may now also glue inﬁnitely many patches Xi , and the isomorphic open subsets Ui, j ⊂ Xi and U j,i ⊂ X j can be empty, which might give rise to disconnected schemes). A morphism from the glued scheme X to some scheme Y can then be given by giving morphisms Xi → Y that are compatible on the overlaps in the obvious sense. The following generalization of proposition 5.2.2 is an application of these glueing techniques. Proposition 5.3.7. Let X be any scheme, and let Y = Spec R be an afﬁne scheme. Then there is a one-to-one correspondence between morphisms X → Y and ring homomorphisms R = OY (Y ) → OX (X). Proof. Let {Ui } be an open afﬁne cover of X, and let {Ui, j,k } be an open afﬁne cover of Ui ∩ U j . Then by remark 5.3.6 giving a morphism f : X → Y is the same as giving morphisms fi : Ui → Y such that fi and f j agree on Ui ∩U j , i.e. such that fi |Ui, j,k = f j |Ui, j,k for all i, j, k. But as the Ui and Ui, j,k are afﬁne, by proposition 5.2.2 the morphisms fi and fi |Ui, j,k correspond exactly to ring homomorphisms OY (Y ) → OUi (Ui ) = OX (Ui ) and OY (Y ) → OUi, j,k (Ui, j,k ) = OX (Ui, j,k ), respectively. Hence a morphism f : X → Y is the same as a collection of ring homomorphisms fi∗ : OY (Y ) → OX (Ui ) such that the compositions ρUi ,Ui, j,k ◦ fi∗ : OY (Y ) → OX (Ui, j,k ) and ρU j ,Ui, j,k ◦ f j∗ : OY (Y ) → OX (Ui, j,k ) agree for all i, j, k. But by the sheaf axiom for OX , this is exactly the data of a ring homomorphism OY (Y ) → OX (X). Remark 5.3.8. By the above proposition, every scheme X admits a unique morphism to Spec Z, determined by the natural map Z → OX (X). More explicitly, on points this map is given by associating to every point P ∈ X the characteristic of its residue ﬁeld k(P). In particular, if X is a scheme over C (or any ground ﬁeld of characteristic 0 for that matter) then the morphism X → Spec Z maps every point to the zero ideal (0). 5.4. Fiber products. In example 2.3.9 and exercise 2.6.13 we deﬁned the product X ×Y for two given prevarieties X and Y by giving the product set X × Y a suitable structure of a ringed space. The idea of this construction was that the coordinate ring A(X × Y ) should be A(X) ⊗ A(Y ) if X and Y are afﬁne (see remark 2.3.13), and then to globalize this construction by glueing techniques. The characteristic property of the product X ×Y was that giving a morphism to it is equivalent to giving a morphism to X and a morphism to Y (see lemma 2.3.11 and exercise 2.6.13). Now we want to do the same thing for schemes. More generally, if X and Y are two schemes over a third scheme S (i.e. if morphisms f : X → S and g : Y → S are given) we want to construct the so-called ﬁber product X ×S Y , that should na¨vely correspond to ı the points (x, y) ∈ X × Y such that f (x) = g(y). As in the case of prevarieties this will be done by ﬁrst constructing this product in the afﬁne case, and then glueing these products together to obtain the ﬁber product of general schemes. We start by deﬁning ﬁber products using the characteristic property mentioned above.

5.

Schemes

83

Deﬁnition 5.4.1. Let f : X → S and g : Y → S be morphisms of schemes. We deﬁne the ﬁber product X ×S Y to be a scheme together with “projection” morphisms πX : X ×S Y → X and πY : X ×S Y → Y such that the square in the following diagram commutes, and such that for any scheme Z and morphisms Z → X and Z → Y making a commutative diagram with f and g there is a unique morphism Z → X ×S Y making the whole diagram commutative: Z " X ×S Y X

πX

πY

% / Y

g

f

/ S

Let us ﬁrst show that the ﬁber product is uniquely determined by this property: Lemma 5.4.2. The ﬁber product X ×S Y is unique if it exists. (In other words, if F1 and F2 are two ﬁber products satisfying the above characteristic property, then F1 and F2 are canonically isomorphic.) Proof. Let F1 and F2 be two ﬁber products satisfying the characteristic property of the deﬁnition. In particular, F2 comes together with morphisms to X and Y . As F1 is a ﬁber product, we get a morphism ϕ : F2 → F1 F2

ϕ

F1 X

# / Y

g

f

/ S

so that this diagram commutes. By symmetry, we get a morphism ψ : F1 → F2 as well. The diagram F1

ϕ◦ψ

F1 X

# / Y

g

f

/ S

is then commutative by construction. But the same diagram is commutative too if we replace ϕ ◦ ψ by the identity morphism. So by the uniqueness part of the deﬁnition of a ﬁber product it follows that ϕ ◦ ψ is the identity. Of course ψ ◦ ϕ is then also the identity by symmetry. So F1 and F2 are canonically isomorphic. Remark 5.4.3. The following two properties of ﬁber products are easily seen from the deﬁnition:

84

Andreas Gathmann

(i) If S ⊂ U is an open subset, then X ×S Y = X ×U Y (morphisms from any Z to X and Y commuting with f and g are then the same regardless of whether the base scheme is S or U). (ii) If U ⊂ X and V ⊂ Y are open subsets, then the ﬁber product

−1 U ×S V = π−1 (U) ∩ πY (V ) ⊂ X ×S Y X

is an open subset of the total ﬁber product X ×S Y . Now we want to show that ﬁber products always exist. We have already mentioned that in the afﬁne case, ﬁber products should correspond to tensor products in commutative algebra. So let us deﬁne the corresponding tensor products ﬁrst. Deﬁnition 5.4.4. Let R be a ring, and let M and N be R-modules. For every m ∈ M and n ∈ N let m ⊗ n be a formal symbol. We let F be the “free R-module generated by the symbols m ⊗ n”, i.e. F is the R-module of formal ﬁnite linear combinations F=

∑ ri (mi ⊗ ni ) ; ri ∈ R, mi ∈ M, ni ∈ N

i

.

Now we deﬁne the tensor product M ⊗R N of M and N over R to be the R-module F modulo the relations (m1 + m2 ) ⊗ n = m1 ⊗ n + m2 ⊗ n, m ⊗ (n1 + n2 ) = m ⊗ n1 + m ⊗ n2 , r(m ⊗ n) = (rm) ⊗ n = m ⊗ (rn) for all m, mi ∈ M, n, ni ∈ N, and r ∈ R. Obviously, M ⊗R N is an R-module as well. Example 5.4.5. (i) Let k be a ﬁeld. Then k[x] ⊗k k[y] = k[x, y], where the isomorphism is given by k[x] ⊗k k[y] → k[x, y], f (x) ⊗ g(y) → f (x) · g(y) and k[x, y] → k[x] ⊗k k[y],

∑ ai, j xi y j → ∑ ai, j (xi ⊗ y j ).

i, j i, j

(ii) Let R be a ring, and let I1 and I2 be ideals. Then R/I1 and R/I2 are R-modules, and we have R/I1 ⊗R R/I2 = R/(I1 + I2 ). In fact, the isomorphism is given by R/I1 ⊗R R/I2 → R/(I1 + I2 ), r1 ⊗ r2 → r1 · r2 and R/(I1 + I2 ) → R/I1 ⊗R R/I2 , r → r(1 ⊗ 1) = (r ⊗ 1) = (1 ⊗ r). (iii) If M is any R-module, then M ⊗R R = R ⊗R M = M. Remark 5.4.6. It is easy to see that the tensor product of modules satisﬁes the following characteristic property (which is exactly the same as that of deﬁnition 5.4.1, just with all the arrows reversed): Let R, M, and N be rings, and assume that we are given ring homomorphisms f : R → M and g : R → N (that make M and N into R-modules). Then for every ring A and homomorphisms M → A and N → A making a commutative diagram with f and g there is a unique

5.

Schemes

85

ring homomorphism M ⊗R N → A making the whole diagram commutative: AU lb M ⊗R N o O M o

N O

g

f

R

where M → M ⊗R N and N → M ⊗R N are the obvious maps m → m ⊗ 1 and n → 1 ⊗ n. In fact, if a : M → A and b : N → A are the two ring homomorphisms, then M ⊗R N → A is given by m ⊗ n → a(m) · b(n). Using the tensor product of modules, we can now construct the ﬁber product of schemes. Lemma 5.4.7. Let f : X → S and g : Y → S be morphisms of schemes. Then there is a ﬁber product X ×S Y . Proof. First assume that X, Y , and S are afﬁne schemes, so X = Spec M, Y = Spec N, and S = Spec R. The morphisms X → S and Y → S make M and N into R-modules by proposition 5.2.2. We claim that Spec(M ⊗R N) is the ﬁber product X ×S Y . Indeed, giving a morphism Z → Spec(M ⊗R N) is the same as giving a homomorphism M ⊗R N → OZ (Z) by proposition 5.3.7. By remark 5.4.6, this is the same as giving homomorphisms M → OZ (Z) and N → OZ (Z) that induce the same homomorphism on R, which again by proposition 5.3.7 is the same as giving morphisms Z → X and Z → Y that give rise to the same morphism from Z → S. Hence Spec(M ⊗R N) is the desired product. Now let X, Y and S be general schemes. Cover S by open afﬁnes Si , then cover f −1 (Si ) and g−1 (Si ) by open afﬁnes Xi, j and Yi,k , respectively. Consider the ﬁber products Xi, j ×Si Yi,k that exist by the above tensor product construction. Note that by remark 5.4.3 (i) these will then be ﬁber products over S as well. Now if we have another such product Xi , j ×S Yi ,k , both of them will contain the (unique) ﬁber product (Xi, j ∩ Xi , j ) ×S (Yi,k ∩ Yi ,k ) as an open subset by remark 5.4.3 (ii), hence they can be glued along these isomorphic open subsets. It is obvious that the ﬁnal scheme X ×S Y obtained by glueing the patches satisﬁes the deﬁning property of a ﬁber product. Example 5.4.8. Let X and Y be prevarieties over a ﬁeld k. Then the scheme-theoretic ﬁber product X ×Spec k Y is just the product prevariety X ×Y considered earlier. In fact, this follows from remark 2.3.13 in the afﬁne case, and the glueing is done in the same way for prevarieties and schemes. Consequently, we will still use the notation X ×Y to denote the ﬁber product X ×Spec k Y over Spec k. Note however that for general schemes X and Y one also often deﬁnes X ×Y to be X ×Spec Z Y (see remark 5.3.8). For schemes over k, X ×Spec k Y and X ×Spec Z Y will in general be different (see exercise 5.6.10), so one has to make clear what is meant by the notation X ×Y . Example 5.4.9. Let Y1 → X and Y2 → X be morphisms of schemes that are “inclusion morphisms”, i.e. the Yi might be open subsets of X, or closed subschemes as in example 5.2.3. Then Then Y1 ×X Y2 is deﬁned to be the intersection scheme of Y1 and Y2 in X and is usually written Y1 ∩Y2 . For example, if X = Spec R, Y1 = Spec R/I1 , and Y2 = Spec R/I2 as in example 5.2.3, then Y1 ∩ Y2 is Spec R/(I1 + I2 ), which is consistent with example 5.4.5 (ii).

86

Andreas Gathmann

Example 5.4.10. Let Y be a scheme, and let P ∈ Y be a point. Let k = k(P) be the residue ﬁeld of P. Then there is a natural morphism Spec k → Y that maps the unique point of Spec k to P and pulls back a section ϕ ∈ OY (U) (with P ∈ U) to the element in k(P) determined by the composition of maps OY (U) → OY,P → k(P). Now let X → Y be a morphism. Then the ﬁber product X ×Y Spec k (with the morphism Spec k → Y constructed above) is called the inverse image or ﬁber of X → Y over the point P ∈ Y (hence the name “ﬁber product”). As an example, consider the morphism X = A1 → Y = A1 given by x → y = x2 . Over C C the point 0 ∈ Y the ﬁber is then Spec(C[x]⊗C[y] C), where the maps are given by y ∈ C[y] → x2 ∈ C[x] and y ∈ C[y] → 0 ∈ C. This tensor product is equal to C[x]/(x2 ), so the ﬁber over 0 is the double point Spec C[x]/(x2 ); it is a non-reduced scheme and therefore different from the set-theoretic inverse image of 0 as deﬁned earlier for prevarieties.

X

0

0

Y

Example 5.4.11. Continuing the above example, one might want to think of a morphism X → Y as some sort of ﬁbered object, giving a scheme X ×Y Spec k(P) for every point P ∈ Y . (This is analogous to ﬁbered objects in topology.) Now let f : Y → Y be any morphism. Then the ﬁber product X = X ×Y Y has a natural projection morphism to Y , and its ﬁber over a point P ∈ Y is equal to the ﬁber of X → Y over the point P ∈ Y . This is usually called a base extension of the morphism X → Y . (It corresponds to e.g. the pull-back of a vector bundle in topology.)

X´ X

Y´

Y

5.5. Projective schemes. We know that projective varieties are a special important class of varieties that are not afﬁne, but still can be described globally without using glueing techniques. They arise from looking at homogeneous ideals, i.e. graded coordinate rings. A completely analogous construction exists in the category of schemes, starting with a graded ring and looking at homogeneous ideals in it. Deﬁnition 5.5.1. Let R be a graded ring (think of the homogeneous coordinate ring S(X) L of a projective variety X), i.e. a ring together with a decomposition R = d≥0 R(d) into abelian groups such that R(d) · R(e) ⊂ R(d+e) . An element of R(d) is called homogeneous of degree d. An ideal I ⊂ R is called homogeneous if it can be generated by homogeneous L elements. Let R+ be the ideal d>0 R(d) . We deﬁne the set Proj R to be the set of all homogeneous prime ideals p ⊂ R with R+ ⊂ p (compare this to theorem 3.2.6; R+ corresponds to the “irrelevant ideal” (x0 , . . . , xn ) ⊂ k[x0 , . . . , xn ]). If I ⊂ R is a homogeneous ideal, we deﬁne Z(I) = {p ∈ Proj R ; p ⊃ I} to be the zero locus of I.

5.

Schemes

87

The proof of the following lemma is the same as in the case of afﬁne or projective varieties: Lemma 5.5.2. Let R be a graded ring. (i) If {Ii } is a family of homogeneous ideals of R then i Z(Ii ) = Z(∑i Ii ) ⊂ Proj R. (ii) If I1 , I2 ⊂ R are homogeneous ideals then Z(I1 ) ∪ Z(I2 ) = Z(I1 I2 ) ⊂ Proj R.

T

In particular, we can deﬁne a topology on Proj R by taking the subsets of the form Z(I) for some I to be the closed sets. Of course, the next thing to do is to deﬁne a structure of (locally) ringed space on Proj R. This is in complete analogy to the afﬁne case. Next we have to deﬁne the rings of regular functions on Proj R. This is a mixture of the case of afﬁne schemes and projective varieties. We will more or less copy deﬁnition 5.1.11 for afﬁne schemes, keeping in mind that in the projective (i.e. homogeneous) case our functions should locally be quotients of homogeneous elements of R of the same degree. Deﬁnition 5.5.3. Let R be a graded ring, and let X = Proj R. For every p ∈ Proj R, let R(p) = f ; g ∈ p and f , g ∈ R(d) for some d / g

be the ring of degree zero elements of the localization of R with respect to the multiplicative system of all homogeneous elements of R that are not in p. (Of course, this will correspond to the local ring at the point p, see proposition 5.5.4 below.) Now for every open subset U ⊂ X we deﬁne OX (U) to be

OX (U) := {ϕ = (ϕp )p∈U with ϕp ∈ R(p) for all p ∈ U

such that “ϕ is locally of the form

f g

for f , g ∈ R(d) for some d”}

= {ϕ = (ϕp )p∈U with ϕp ∈ R(p) for all p ∈ U such that for every p ∈ U there is a neighborhood V in U and f , g ∈ R(d) for some d with g ∈ q and ϕq = /

f g

∈ R(q) for all q ∈ V .}

It is clear from the local nature of the deﬁnition of OX (U) that OX is a sheaf. Proposition 5.5.4. Let R be a graded ring. (i) For every p ∈ Proj R the stalk OX,p is isomorphic to the local ring R(p) . (ii) For every homogeneous f ∈ R+ , let X f ⊂ X be the distinguished open subset X f := X\Z( f ) = {p ∈ Proj R ; f ∈ p}. / These open sets cover X, and for each such open set we have an isomorphism of locally ringed spaces (X f , OX |X f ) ∼ Spec R( f ) , where = R( f ) = g ; g ∈ R(r·deg f ) fr

is the ring of elements of degree zero in the localized ring R f . In particular, Proj R is a scheme. Proof. (i): There is a well-deﬁned homomorphism

OX,p → R(p) , (U, ϕ) → ϕ(p).

The proof that this is an isomorphism is the same as in the afﬁne case (see proposition 5.1.12 (i). (ii): Let p ∈ X be a point. By deﬁnition, R+ ⊂ p, so there is a f ∈ R+ with f ∈ p. But / then p ∈ X f ; hence the open subsets of the form X f cover X.

88

Andreas Gathmann

Now ﬁx f ∈ R+ ; we will deﬁne an isomorphism ψ : X f → Spec R( f ) . For any homogeneous ideal I ⊂ R, set ψ(I) := (I R f ) ∩ R( f ) . In particular, restricting this to prime ideals gives a map of sets X f → Spec R( f ) , which is easily seen to be a bijection. Moreover, if I ⊂ R is any ideal then ψ(p) ⊃ ψ(I) if and only if p ⊃ I, so ψ : X f → Spec R( f ) is a homeomorphism. Note also that for p ∈ X f the local rings g OProj R,p = R(p) = ; g and h homogeneous of the same degree, h ∈ p / h and

OSpec R( f ) ,ψ(p) = (R( f ) )ψ(p)

= g/ f r ; g and h homogeneous of degrees r · deg f and s · deg f , h ∈ p / h/ f s

are isomorphic for f ∈ p. This gives rise to isomorphisms between the rings of regular / functions OX f (U) and OSpec R( f ) (U) (as they are by deﬁnition made up of the local rings). Example 5.5.5. If k is an algebraically closed ﬁeld, then by construction Proj k[x0 , . . . , xn ] is the scheme that corresponds to projective n-space Pn over k. More generally, the scheme k associated to a projective variety X is just Proj S(X), where S(X) = k[x0 , . . . , xn ]/I(X) is the homogeneous coordinate ring of X. Of course, scheme-theoretically we can now also consider schemes that are of the form Proj k[x0 , . . . , xn ]/I where I is any homogeneous ideal of the polynomial ring. This allows projective “subschemes of Pn ” that are not necessarily irreducible or reduced. Let us turn this into a deﬁnition. Deﬁnition 5.5.6. Let k be an algebraically closed ﬁeld. A projective subscheme of Pn is k a scheme of the form Proj k[x0 , . . . , xn ]/I for some homogeneous ideal I. As mentioned above, every projective variety is a projective subscheme of Pn . However, the category of projective subschemes of Pn is bigger because it contains schemes that are reducible (e.g. the union of the coordinate axes in the plane Proj k[x0 , x1 , x2 ]/(x1 x2 )) or 2 non-reduced (e.g. the double point Proj k[x0 , x1 ]/(x1 )). As in the case of projective varieties, we now want to make precise the relation between projective subschemes of Pn and homogeneous ideals in k[x0 , . . . , xn ]. Note that the existence of the irrelevant ideal (x0 , . . . , xn ) implies that this correspondence is not one-toone: the example Proj k[x0 , . . . , xn ]/( f ) = Proj k[x0 , . . . , xn ]/( f x0 , . . . , f xn ) of remark 3.1.11 works for schemes as well. ¯ Deﬁnition 5.5.7. Let I ⊂ S = k[x0 , . . . , xn ] be a homogeneous ideal. The saturation I of I is deﬁned to be m ¯ I = {s ∈ S ; xi · s ∈ I for some m and all i}. ¯ Example 5.5.8. If I = ( f x0 , . . . , f xn ) then I = ( f ). So in this case the saturation removes the ambiguity of the ideal associated to a projective subscheme of Pn . We will now show that this is true in general: Lemma 5.5.9. Let I, J ⊂ S = k[x0 , . . . , xn ] be homogeneous ideals. Then ¯ (i) I is a homogeneous ideal. ¯ (ii) Proj S/I = Proj S/I. ¯ = Proj S/J¯ if and only if I = J. ¯ ¯ (iii) Proj S/I (d) = I (d) for d ¯ (iv) I 0. Here and in the following we say that a statement holds for d 0 if and only if it holds for large enough d, i.e. if and only if there is a number D ≥ 0 such that the statement holds for all d ≥ D.

5.

Schemes

89

m ¯ Proof. (i): Let s ∈ I any (possibly non-homogeneous) element. Then by deﬁnition xi ·s ∈ I m for some m and all i. As I is homogeneous, it follows that the graded pieces xi · s(d) are (d) ∈ I for all i. Hence I is ¯ ¯ in I as well for all d. Therefore, by deﬁnition, it follows that s homogeneous. (ii): As the open afﬁnes Ui := {xi = 0} ⊂ Pn cover Pn , it sufﬁces to show that Ui ∩ ¯ ¯ Proj S/I = Ui ∩ Proj S/I. But this is obvious as I|xi =1 = Ixi =1 . (iii): The direction “⇒” is trivial. For “⇐” it sufﬁces to show that the saturated ideal ¯ ¯ I can be recovered from the projective scheme X = Proj S/I alone. Thinking of projective ¯ should just be “the ideal I(X) of X”, i.e. the ideal of functions vanishing on X. varieties, I Now the elements of S do not deﬁne functions on X, but after setting one xi equal to 1 they ¯ do deﬁne functions on X ∩Ui . Hence we can recover I from X as ¯ I = {s ∈ S ; s|xi =1 = 0 on X ∩Ui for all i}

(note that the right hand side depends only on the scheme X and not on its representation as Proj S/I for a certain I. ¯ ¯ (iv): The inclusion I (d) ⊂ I (d) is obvious (for all d) as I ⊂ I. So we only have to show ¯(d) ⊂ I (d) for d 0. that I ¯ First of all note that I is ﬁnitely generated; let f1 , . . . , fm be (homogeneous) generators. ¯ Let D1 be the maximum degree of the fi . Next, by deﬁnition of I there is a number D2 such that xd · fi ∈ I for all 0 ≤ j ≤ n, 1 ≤ i ≤ m, and d ≥ D2 . Set D = D1 + (n + 1)D2 . j ¯ Now let f ∈ I (d) be any homogeneous element in the saturation of degree d ≥ D. We can write f as ∑i ai fi , with the ai homogeneous of degree at least (n + 1)D2 . This degree bound implies that every monomial of ai contains at least one x j with a power of at least D2 . But then this power multiplied with fi lies in I by construction. So it follows that ai fi ∈ I for all i, and therefore f ∈ I (d) . Deﬁnition 5.5.10. If X is a projective subscheme of Pn , we let I(X) be the saturation of any ideal I ⊂ k[x0 , . . . , xn ] such that X = Proj k[x0 , . . . , xn ]/I. (This is well-deﬁned by lemma 5.5.9 (iii) and generalizes the notion of the ideal of a projective variety to projective subschemes of Pn .) We deﬁne S(X) to be k[x0 , . . . , xn ]/I(X). As usual, we call I(X) the ideal of X and S(X) the homogeneous coordinate ring of X. Corollary 5.5.11. There is a one-to-one correspondence between projective subschemes of Pn and saturated homogeneous ideals in k[x0 , . . . , xn ], given by X → I(X) and I → k Proj k[x0 , . . . , xn ]/I. 5.6. Exercises. Exercise 5.6.1. Find all closed points of the real afﬁne plane A2 . What are their residue R ﬁelds? Exercise 5.6.2. Let f (x, y) = y2 − x2 − x3 . Describe the afﬁne scheme X = Spec R/( f ) set-theoretically for the following rings R: (i) R = C[x, y] (the standard polynomial ring), (ii) R = C[x, y](x,y) (the localization of the polynomial ring at the origin), (iii) R = C[[x, y]] (the ring of formal power series). Interpret the results geometrically. In which of the three cases is X irreducible? Exercise 5.6.3. For each of these cases below give an example of an afﬁne scheme X with that property, or prove that such an X does not exist: (i) X has inﬁnitely many points, and dim X = 0. (ii) X has exactly one point, and dim X = 1. (iii) X has exactly two points, and dim X = 1.

90

Andreas Gathmann

(iv) X = Spec R with R ⊂ C[x], and dim X = 2. Exercise 5.6.4. Let X be a scheme, and let Y be an irreducible closed subset of X. If ηY is the generic point of Y , we write OX,Y for the stalk OX,ηY . Show that OX,Y is “the ring of rational functions on X that are regular at a general point of Y ”, i.e. it is isomorphic to / the ring of equivalence classes of pairs (U, ϕ), where U ⊂ X is open with U ∩ Y = 0 and ϕ ∈ OX (U), and where two such pairs (U, ϕ) and (U , ϕ ) are called equivalent if there is / an open subset V ⊂ U ∩U with V ∩Y = 0 such that ϕ|V = ϕ|V . (In particular, if X is a scheme that is a variety, then OX,ηX is the function ﬁeld of X as deﬁned earlier. Hence the stalks of the structure sheaf of a scheme generalize both the concepts of the local rings and the function ﬁeld of a variety.) Exercise 5.6.5. Let X be a scheme of ﬁnite type over an algebraically closed ﬁeld k. Show that the closed points of X are dense in every closed subset of X. Conversely, give an example of a scheme X such that the closed points of X are not dense in X. Exercise 5.6.6. Let X = {(x, y, z) ∈ C3 ; xy = xz = yz = 0} be the union of the three coordinate lines in C3 . Let Y = {(x, y) ∈ C2 ; xy(x − y) = 0} be the union of three concurrent lines in C2 . Are X and Y isomorphic as schemes? (Hint: Deﬁne and compute the tangent spaces of X and Y at the origin.) Exercise 5.6.7. Let X ⊂ P3 the complex cubic surface

3 X = {(x0 : x1 : x2 : x3 ) ; x0 = x1 x2 x3 }.

(i) Show that X is singular. (ii) Let M ⊂ G(1, 3) be the subset of the Grassmannian of lines in P3 that corresponds to all lines in P3 that lie in X. By writing down explicit equations for M, show that M has the structure of a scheme in a natural way. (iii) Show that the scheme M contains exactly 3 points, but that it has length 27 over C, i.e. it is of the form M = Spec R with R a 27-dimensional C-algebra. Hence in a certain sense we can say that even the singular cubic surface X contains exactly 27 lines, if we count the lines with their correct multiplicities. Exercise 5.6.8. Let k be an algebraically closed ﬁeld. An n-fold point (over k) is a scheme of the form X = Spec R such that X has only one point and R is a k-algebra of vector space dimension n over k (i.e. X has length n). Show that every double point is isomorphic to Spec k[x]/(x2 ). On the other hand, ﬁnd two non-isomorphic triple points over k, and describe them geometrically. Exercise 5.6.9. Show that for a scheme X the following are equivalent: (i) X is reduced, i.e. for every open subset U ⊂ X the ring OX (U) has no nilpotent elements. (ii) For any open subset Ui of an open afﬁne cover {Ui } of X, the ring OX (Ui ) has no nilpotent elements. (iii) For every point P ∈ X the local ring OX,P has no nilpotent elements. Exercise 5.6.10. Show that A2 C A1 ×Spec Z A1 . C C

2 2 Exercise 5.6.11. Let X = Z(x1 x2 + x1 x2 x3 ) ⊂ A3 , and denote by πi the projection to the C i-th coordinate. Compute the scheme-theoretic ﬁbers Xxi =a = π−1 (a) for all a ∈ C, and i determine the set of isomorphism classes of these schemes.

Exercise 5.6.12. Let X be a prevariety over an algebraically closed ﬁeld k, and let P ∈ X be a (closed) point of X. Let D = Spec k[x]/(x2 ) be the “double point”. Show that the tangent space TX,P to X at P can be canonically identiﬁed with the set of morphisms D → X that map the unique point of D to P.

5.

Schemes

91

(In particular, this gives the set of morphisms D → X with ﬁxed image point P ∈ X the structure of a vector space over k. Can you see directly how to add two such morphisms, and how to multiply them with a scalar in k ?) Exercise 5.6.13. Let X be an afﬁne variety, let Y be a closed subscheme of X deﬁned by ˜ the ideal I ⊂ A(X), and let X be the blow-up of X at I. Show that: L ˜ (i) X = Proj( d≥0 I d ), where we set I 0 := A(X). ˜ (ii) The projection map X → X is the morphism induced by the ring homomorphism 0→L d. I d≥0 I ˜ ˜ (iii) The exceptional divisor of the L blow-up, i.e. the ﬁber Y ×X X of the blow-up X → X over Y , is isomorphic to Proj( d≥0 I d /I d+1 ). Exercise 5.6.14. Let X = Spec R and Y = Spec S be afﬁne schemes. Show that the disjoint union X Y is an afﬁne scheme with X Y = Spec(R × S), where as usual R × S = {(r, s) ; r ∈ R, s ∈ S} (with addition and multiplication deﬁned componentwise).

92

Andreas Gathmann

6. F IRST APPLICATIONS OF SCHEME THEORY

To every projective subscheme of Pn we associate the Hilbert function hX : Z → k Z, d → dimk S(X)(d) . For large d the Hilbert function is a polynomial in d of degree dim X, the so-called Hilbert polynomial χX . We deﬁne (dim X)! times the leading coefﬁcient of χX to be the degree of X; this is always a positive integer. For zero-dimensional schemes the degree is just the number of points in X counted with their scheme-theoretic multiplicities. The degree is additive for unions of equidimensional schemes and multiplicative for intersections with hypersurfaces (B´ zout’s theorem). e We give some elementary applications of B´ zout’s theorem for plane curves. e Among others, we give upper bounds for the numbers of singularities of a plane curve and the numbers of loops of a real plane curve. A divisor on a curve C is just a formal linear combination of points on C with integer coefﬁcients. To every polynomial or rational function on C we can associate a divisor, namely the divisor of “zeros minus poles” of the polynomial or function. The group of all divisors modulo the subgroup of divisors of rational functions is called the Picard group PicC of C. We show that the degree-0 part of PicC is trivial for C = P1 , whereas it is bijective to C itself if C is a smooth plane cubic curve. This deﬁnes a group structure on such cubic curves that can also be interpreted geometrically. In complex analysis, plane cubic curves appear as complex tori of the form C/Λ, where Λ is a rank-2 lattice in C. Finally, we give a short outlook to the important parts of algebraic geometry that have not been covered yet in this class.

6.1. Hilbert polynomials. In this section we will restrict our attention to projective subschemes of Pn over some ﬁxed algebraically closed ﬁeld. Let us start by deﬁning some numerical invariants associated to a projective subscheme of Pn . Deﬁnition 6.1.1. Let X be a projective subscheme of Pn . Note that the homogeneous coork dinate ring S(X) is a graded ring, and that each graded part S(X)(d) is a ﬁnite-dimensional vector space over k. We deﬁne the Hilbert function of X to be the function hX : Z → Z d → hX (d) := dimk S(X)(d) . (Note that we trivially have hX (d) = 0 for d < 0 and hX (d) ≥ 0 for d ≥ 0, so we will often consider hX as a function hX : N → N.) Example 6.1.2. Let X = Pn be projective space itself. Then S(X) = k[x0 , . . . , xn ], so the Hilbert function hX (d) = d+n is just the number of degree-d monomials in n+1 variables n

x0 , . . . , xn . In particular, note that hX (d) = (d+n)(d+n−1)···(d+1) is a polynomial in d of degree n! 1 n with leading coefﬁcient n! (compare this to proposition 6.1.5). Example 6.1.3. Let us now consider some examples of zero-dimensional schemes.

(i) Let X = {(1 : 0), (0 : 1)} ⊂ P1 be two points in P1 . Then I(X) = (x0 x1 ). So a d d basis of S(X)(d) is given by {1} for d = 0, and {x0 , x1 } for d > 0. We conclude that 1 for d = 0, hX (d) = 2 for d > 0. (ii) Let X = {(1 : 0 : 0), (0 : 1 : 0), (0 : 0 : 1)} ⊂ P2 be three points in P2 that are not on a line. Then I(X) = (x0 x1 , x0 x2 , x1 x2 ). So in the same way as in (i), a basis of

6.

First applications of scheme theory

93

d d d S(X)(d) is given by {1} for d = 0 and {x0 , x1 , x2 } for d > 0. Therefore

hX (d) =

1 3

for d = 0, for d > 0.

(iii) Let X = {(1 : 0), (0 : 1), (1 : 1)} ⊂ P1 be three collinear points. Then I(X) = 2 2 (x0 x1 (x0 − x1 )). The relation x0 x1 = x0 x1 allows us to reduce the number of x0 in i x j provided that i ≥ 2 and j ≥ 1. So a basis of S(X)(d) is given by a monomial x0 1 d−1 d d {1} for d = 0, {x0 , x1 } for d = 1, and {x0 , x0 x1 , x1 } for d > 1. Hence 1 for d = 0, hX (d) = 2 for d = 1, 3 for d > 1. It is easy to see that we get the same result for three collinear points in P2 . So comparing this with (ii) we conclude that the Hilbert function does not only depend on the scheme X up to isomorphism, but also on the way the scheme is embedded into projective space. 2 (iv) Let X ⊂ P1 be the “double point” given by the ideal I(X) = (x0 ). A basis of (d) is given by {1} for d = 0 and {x xd−1 , xd } for d > 0, so it follows that S(X) 0 1 1 hX (d) = 1 2 for d = 0, for d > 0.

just as in (i). So the double point “behaves like two separate points” for the Hilbert function. So we see that in these examples the Hilbert function becomes constant for d large enough, whereas its initial values for small d may be different. We will now show that this is what happens in general for zero-dimensional schemes: Lemma 6.1.4. Let X be a zero-dimensional projective subscheme of Pn . Then (i) X is afﬁne, so equal to Spec R for some k-algebra R. (ii) This k-algebra R is a ﬁnite-dimensional vector space over k. Its dimension is called the length of X and can be interpreted as the number of points in X (counted with their scheme-theoretic multiplicities). (iii) hX (d) = dimk R for d 0. In particular, hX (d) is constant for large values of d. Proof. (i): As X is zero-dimensional, we can ﬁnd a hyperplane that does not intersect X. Then X = X\H is afﬁne by proposition 5.5.4 (ii). (ii): First assume that X is irreducible, i.e. it consists of only one point (but may have a non-trivial scheme structure). By a change of coordinates we can assume that √ point this is the origin in An . If X = Spec k[x1 , . . . , xn ]/I we then must have (x1 , . . . , xn ) = I by the d Nullstellensatz. It follows that xi ∈ I for some d and all i. Consequently, every monomial of degree at least D := d · n lies in I (as it must contain at least one xi with a power of at least d). In other words, k[x1 , . . . , xn ]/I has a basis (as a vector space over k) of polynomials of degree less than D. But the space of such polynomials is ﬁnite-dimensional. (iii): Note that I(X) is simply the homogenization of I. Conversely, I is equal to I(X)|x0 =1 . So for d ≥ D an isomorphism S(d) → R as vector spaces over k is given by (k[x0 , . . . , xn ]/I(X))(d) → k[x1 , . . . , xn ]/I, and the inverse k[x1 , . . . , xn ]/I → (k[x0 , . . . , xn ]/I(X))(d) , f → f h · x0

d−deg f

f → f |x0 =1

94

Andreas Gathmann

where f h denotes the homogenization of a polynomial as in exercise 3.5.3 (note that the second map is well-deﬁned as k[x1 , . . . , xn ]/I has a basis of polynomials of degree less than D). We will now discuss the Hilbert function of arbitrary projective subschemes of Pn (that are not necessarily zero-dimensional). Proposition 6.1.5. Let X be a (non-empty) m-dimensional projective subscheme of Pn . Then there is a (unique) polynomial χX ∈ Z[d] such that hX (d) = χX (d) for d 0. Moreover, (i) The degree of χX is m. (ii) The leading coefﬁcient of χX is

1 m!

times a positive integer.

Remark 6.1.6. As the Hilbert polynomial is deﬁned in terms of the Hilbert function for large d, it sufﬁces to look at the graded parts of I(X) (or S(X)) for d 0. So by lemma 5.5.9 (iv) we do not necessarily need to take the saturated ideal of X for the computation of the Hilbert polynomial. We have as well that χX (d) = dimk (k[x0 , . . . , xn ]/I)(d) for d 0 for any homogeneous ideal I such that X = Proj k[x0 , . . . , xn ]/I. Proof. We will prove the proposition by induction on the dimension m of X. The case m = 0 follows from lemma 6.1.4, so let us assume that m > 0. By a linear change of coordinates we can assume that no component of X lies in the hyperplane H = {x0 = 0}. Then there is an exact sequence of graded vector spaces over k

0 0 −→ k[x0 , . . . , xn ]/I(X) −→ k[x0 , . . . , xn ]/I(X) −→ k[x0 , . . . , xn ]/(I(X) + (x0 )) −→ 0.

·x

(if the ﬁrst map was not injective, there would be a homogeneous polynomial f such that f ∈ I(X) but f x0 ∈ I(X). We would then have X = (X ∩ Z( f )) ∪ (X ∩ H). But as no irre/ ducible component lies in H by assumption, we must have X = X ∩ Z( f ), in contradiction to f ∈ I(X)). Taking the d-th graded part of this sequence (and using remark 6.1.6 for the / ideal I(X) + (x0 )), we get hX∩H (d) = hX (d) − hX (d − 1). for large d. By the induction assumption, hX∩H (d) is a polynomial of degree m − 1 for 1 large d whose leading coefﬁcient is (m−1)! times a positive integer. We can therefore write

m−1

hX∩H (d) =

i=0

∑ ci

d i

for d

0

d i

for some constants ci , where cm−1 is a positive integer (note that 1 degree i in d with leading coefﬁcient i! ). We claim that

m−1

is a polynomial of

hX (d) = c +

i=0

∑ ci

d +1 i+1

for d

0

for some c ∈ Z. In fact, this follows by induction on d, as hX (d) = hX∩H (d) + hX (d − 1)

m−1 m−1 d d + c + ∑ ci i i+1 i=0

=

i=0

∑ ci

m−1 i=0

= c+

∑ ci

d +1 . i+1

6.

First applications of scheme theory

95

The statement of proposition 6.1.5 motivates the following deﬁnition: Deﬁnition 6.1.7. Let X be a projective subscheme of Pn . The degree deg X of X is deﬁned to be (dim X)! times the leading coefﬁcient of the Hilbert polynomial χX . (By proposition 6.1.5, this is a positive integer.) Example 6.1.8. (i) If X is a zero-dimensional scheme then deg X is equal to the length of X, i.e. to “the number of points in X counted with their scheme-theoretic multiplicities”. (ii) deg Pn = 1 by example 6.1.2. (iii) Let X = Proj k[x0 , . . . , xn ]/( f ) be the zero locus of a homogeneous polynomial. We claim that deg X = deg f . In fact, taking the d-th graded part of S(X) = k[x0 , . . . , xn ]/ f · k[x0 , . . . , xn ] we get hX (d) = dimk k[x0 , . . . , xn ](d) − dimk k[x0 , . . . , xn ](d−deg f ) d +n d − deg f + n − n n 1 ((d + n) · · · (d + 1) − (d − deg f + n) · · · (d − deg f + 1)) = n! deg f n−1 d + lower order terms. = (n − 1)! = Proposition 6.1.9. Let X1 and X2 be m-dimensional projective subschemes of Pn , and assume that dim(X1 ∩ X2 ) < m. Then deg(X1 ∪ X2 ) = deg X1 + deg X2 . Proof. For simplicity of notation let us set S = k[x0 , . . . , xn ]. Note that X1 ∩ X2 = Proj S/(I(X1 ) + I(X2 )) and X1 ∪ X2 = Proj S/(I(X1 ) ∩ I(X2 )). So from the exact sequence 0 → S/(I(X1 ) ∩ I(X2 )) → S/I(X1 ) ⊕ S/I(X2 ) → S/(I(X1 ) + I(X2 )) → 0 f we conclude that hX1 (d) + hX2 (d) = hX1 ∪X2 (d) + hX1 ∩X2 (d) for large d. In particular, the same equation follows for the Hilbert polynomials. Comparing only the leading (i.e. d m ) coefﬁcient we then get the desired result, since the degree of χX1 ∩X2 is less than m by assumption. Example 6.1.10. Let X be a projective subscheme of Pn . We call g(X) := (−1)dim X · (χX (0) − 1) the (arithmetic) genus of X. The importance of this number comes from the following two facts (that we unfortunately cannot prove yet with our current techniques): (i) The genus of X is independent of the projective embedding, i.e. if X and Y are isomorphic projective subschemes then g(X) = g(Y ). See section 6.6.3 and exercise 10.6.8 for more details. (ii) If X is a smooth curve over C, then g(X) is precisely the “topological genus” introduced in example 0.1.1. (Compare for example the degree-genus formula of example 0.1.3 with exercise 6.7.3 (ii).) Remark 6.1.11. In general, the explicit computation of the Hilbert polynomial hX of a projective subscheme X = Proj k[x0 , . . . , xn ]/I from the ideal I is quite complicated and requires methods of computer algebra. → (f, f) ( f , g) → f −g

96

Andreas Gathmann

6.2. B´ zout’s theorem. We will now prove the main property of the degree of a projective e variety: that it is “multiplicative when taking intersections”. We will prove this here only for intersections with hypersurfaces, but there is a more general version about intersections in arbitrary codimension (see e.g. cite Ha theorem 18.4). Theorem 6.2.1. (B´ zout’s theorem) Let X be a projective subscheme of Pn of positive e dimension, and let f ∈ k[x0 , . . . , xn ] be a homogeneous polynomial such that no component of X is contained in Z( f ). Then deg(X ∩ Z( f )) = deg X · deg f . Proof. The proof is very similar to that of the existence of the Hilbert polynomial in proposition 6.1.5. Again we get an exact sequence 0 −→ k[x0 , . . . , xn ]/I(X) −→ k[x0 , . . . , xn ]/I(X) −→ k[x0 , . . . , xn ]/(I(X) + ( f )) −→ 0 from which it follows that χX∩Z( f ) = χX (d) − χX (d − deg f ). But we know that χX (d) = deg X m d + cm−1 d m−1 + terms of order at most d m−2 , m!

·f

where m = dim X. Therefore it follows that χX∩Z( f ) = deg X m (d − (d − deg f )m ) + cm−1 (d m−1 − (d − deg f )m−1 ) m! + terms of order at most d m−2 deg X = · m deg f · d m−1 + terms of order at most d m−2 . m!

We conclude that deg(X ∩ Z( f )) = deg X · deg f . Example 6.2.2. Let C1 and C2 be two curves in P2 without common irreducible components. These curves are then given as the zero locus of homogeneous polynomials of degrees d1 and d2 , respectively. We conclude that deg(C1 ∩C2 ) = d1 · d2 by B´ zout’s thee orem. By example 6.1.8 (i) this means that C1 and C2 intersect in exactly d1 · d2 points, if we count these points with their scheme-theoretic multiplicities in the intersection scheme C1 ∩C2 . In particular, as these multiplicities are always positive integers, it follows that C1 and C2 intersect set-theoretically in at most d1 · d2 points, and in at least one point. This special case of theorem 6.2.1 is also often called B´ zout’s theorem in textbooks. e Example 6.2.3. In the previous example, the scheme-theoretic multiplicity of a point in the intersection scheme C1 ∩C2 is often easy to read off from geometry: let P ∈ C1 ∩C2 be a point. Then: (i) If C1 and C2 are smooth at P and have different tangent lines at P then P counts with multiplicity 1 (we say: the intersection multiplicity of C1 and C2 at P is 1). (ii) If C1 and C2 are smooth at P and are tangent to each other at P then the intersection multiplicity at P is at least 2. (iii) If C1 is singular and C2 is smooth at P then the intersection multiplicity at P is at least 2. (iv) If C1 and C2 are singular at P then the intersection multiplicity at P is at least 3. The key to proving these statements is the following. As the computation is local around P we can assume that the curves are afﬁne in A2 , that P = (0, 0) is the origin, and that the

6.

First applications of scheme theory

97

two curves are given as the zero locus of one equation C1 = { f1 = 0} C2 = { f2 = 0} where f1 = a1 x + b1 y + higher order terms, where f2 = a2 x + b2 y + higher order terms.

If both curves are singular at the origin, their tangent space at P must be two-dimensional, i.e. all of A2 . This means that a1 = b1 = a2 = b2 = 0. It follows that 1, x, and y are three linearly independent elements in k[x, y]/( f1 , f2 ) (whose spectrum is by deﬁnition the intersection scheme). So the intersection multiplicity is at least 3. In the same way, we get at least 2 linearly independent elements (the constant 1 and one linear function) if only one of the curves is singular, or both curves have the same tangent line (i.e. the linear parts of their equations are linearly dependent). Example 6.2.4. Consider again the twisted cubic curve in P3 C = {(s3 : s2t : st 2 : t 3 ) ; (s : t) ∈ P1 }

2 2 = {(x0 : x1 : x2 : x3 ) ; x1 − x0 x2 = x2 − x1 x3 = x0 x3 − x1 x2 = 0}.

We have met this variety as the easiest example of a curve in P3 that cannot be written as the zero locus of two polynomials. We are now able to prove this statement very easily using B´ zout’s theorem: assume that I(C) = ( f , g) for some homogeneous polynomials f e and g. As the degree of C is 3 by exercise 6.7.2, it follows that deg f ·deg g = 3. This is only possible if deg f = 3 and deg g = 1 (or vice versa), i.e. one of the polynomials has to be linear. But C is not contained in a linear space (its ideal does not contain linear functions). In particular we see that C cannot be the intersection of two of the quadratic polynomials given above, as this intersection must have degree 4. In fact,

2 2 Z(x1 − x0 x2 , x2 − x1 x3 ) = C ∪ {x1 = x2 = 0}

in accordance with B´ zout’s theorem and proposition 6.1.9 (note that {x1 = x2 = 0} is a e line and thus has degree 1). Let us now prove some corollaries of B´ zout’s theorem. e Corollary 6.2.5. (Pascal’s theorem) Let X ⊂ P2 be a conic (i.e. the zero locus of a homogeneous polynomial f of degree 2). Pick six points A, B,C, D, E, F on X that form the vertices of a hexagon inscribed in X. Then the three intersection points of the opposite edges of the hexagon (i.e. P = AB ∩ DE, Q = BC ∩ EF, and R = CD ∩ FA) lie on a line.

X A B C F E D R Q

P

Proof. Consider the two reducible cubics X1 = AB ∪ CD ∪ EF and X2 = BC ∪ DE ∪ FA, and let f1 = 0 and f2 = 0 be the equations of X1 and X2 , respectively. In accordance with B´ zout’s theorem, X1 and X2 meet in the 9 points A, B,C, D, E, F, P, Q, R. e Now pick any point S ∈ X not equal to the previously chosen ones. Of course there are λ, µ ∈ k such that λ f1 + µ f2 vanishes at S. Set X = Z(λ f1 + µ f2 ); this is a cubic curve too. Note that X meets X in the 7 points A, B,C, D, E, F, S, although deg X · deg X = 6. We conclude by B´ zout’s theorem that X and X have a common component. For degree e reasons the only possibility for this is that the cubic X is reducible and contains the conic X as a factor. Therefore X = X ∪ L, where L is a line.

98

Andreas Gathmann

Finally note that P, Q, R lie on X as they lie on X1 and X2 . Therefore P, Q, R ∈ X ∪ L. But these points are not on X, so they must be on the line L. Corollary 6.2.6. Let C ⊂ P2 be an irreducible curve of degree d. Then C has at most singular points.

d−1 2

Remark 6.2.7. For d = 1 C must be a line, so there is no singular point. A conic is either irreducible (and smooth) or a union of two lines, so for d = 2 the statement is obvious too. For d = 3 the corollary states that there is at most one singular point on an irreducible curve. In fact, the projectivization of the singular cubic afﬁne curve y2 = x2 + x3 is such an example with one singular point (namely the origin). Proof. Assume the contrary and let P1 , . . . , P(d−1)+1 be distinct singular points of C. More2 over, pick arbitrary further distinct points Q1 , . . . , Qd−3 on C (we can assume d ≥ 3 by 2 remark 6.2.7). We thus have a total of d−1 + 1 + d − 3 = d2 − d − 1 points Pi and Q j . 2 2 We claim that there is a curve C of degree d − 2 that passes through all Pi and Q j . In fact, the space of all homogeneous degree-(d − 2) polynomials in three variables is a d 2 -dimensional vector space over k, so the space of hypersurfaces of degree d − 2 is a projective space PN of dimension N = d − 1, with the coefﬁcients of the equation as the 2 homogeneous coordinates. Now the condition that such a hypersurface passes through a given point is obviously a linear condition in this PN . As N hyperplanes in PN always have a non-empty intersection, it follows that there is a hypersurface passing through any 2 N given points. But N = d − 1 = d2 − d − 1 is precisely the number of points we have. 2 2 (Compare this argument to exercise 3.5.8 and the parametrization of cubic surfaces at the beginning of section 4.5.) Now compute the degree of the intersection scheme C ∩ C . By B´ zout’s theorem, it e must be degC · degC = d(d − 2). Counting the intersection points, we see that we have the d − 3 points Qi , and the d−1 + 1 points Pj that count with multiplicity at least 2 as 2 they are singular points of C (see example 6.2.3). So we get deg(C ∩C ) ≥ (d − 3) + 2 d −1 + 1 = d 2 − 2d + 1 > degC · degC . 2

By B´ zout’s theorem it follows that C and C must have a common component. But C is e irreducible of degree degC > degC , so this is impossible. We thus arrive at a contradiction and conclude that the assumption of the existence of d−1 + 1 singular points was false. 2 The following statement about real plane curves looks quite different from corollary 6.2.6, yet the proof is largely identical. Note that every smooth real plane curve consists of a certain number of connected components (in the classical topology); here are examples with one real component (the left two curves) and with two real components (the right curve):

x2 4

+ y2 − 4 = 0

y2 − x2 − x4 − 1 = 0

3

y2 − x2 − x4 + 1 = 0

3

We want to know the maximum number of such components that a real smooth curve of degree d can have. One way of constructing curves with many components is to start with a singular curve, and then to deform the equation a little bit to obtain a smooth curve. The

6.

First applications of scheme theory

99

following example starts with a reducible quartic curve and deforms it into a smooth curve with two and four components, respectively.

( x4 + y2 − 4)(x2 + y4 − 4) = 0

2

2

( x4 + y2 − 4)(x2 + y4 − 4) = 1

2

2

( x4 + y2 − 4)(x2 + y4 − 4) = −1

2

2

As in the complex case, it is more convenient to pass to the projective plane P2 instead R of A2 . This will add points at inﬁnity of the curves so that every component becomes R a loop (i.e. it has no ends). For example, in the two cubic curves above one point each is added to the curves, so that the components extending to inﬁnity become a loop. We are therefore asking for the maximum number of loops that a projective smooth real plane curve of degree d can have. There is an extra topological twist in P2 that we have not encountered before. As usual, R we construct P2 by taking A2 (which we will draw topologically as an open disc here) R R and adding a point at inﬁnity for every direction in A2 . This has the effect of adding a R boundary to the disc (with the boundary point corresponding to the point at inﬁnity). But note that opposite points of the boundary of the disc belong to the same direction in A2 R and hence are the same point in P2 . In other words, P2 is topologically equivalent to a R R closed disc with opposite boundary points identiﬁed:

2 IPIR

B A A

B identify

It is easy to see that this is a non-orientable surface: if we start with a small circle and move it across the boundary of the disc (i.e. across the inﬁnity locus of P2 then it comes R out with opposite orientation:

A D C D A C D A A B C B C B

Consequently, we have two different types of loops. A “type 1 loop” is a loop such that its complement has only one component (which is topologically a disc). A “type 2 loop” is a loop such that its complement has two components (an “interior” and “exterior” of the loop). It is interesting to note that of these two components one is a disc, and the other is a M¨ bius strip. o

100

Andreas Gathmann

A

A Type 1 loop Type 2 loop

(Those of you who know some algebraic topology will note that the homology group H1 (P2 ) is isomorphic to Z/2Z; so the two types of curves correspond to the two elements R of Z/2Z.) With these prerequisites at hand, we can now prove the following statement (modulo some topology statements that should be intuitively clear): Corollary 6.2.8. (Harnack’s theorem) A real smooth curve in P2 of degree d has at most R d−1 2 + 1 loops. Remark 6.2.9. A line (d = 1) has always exactly one loop. A non-empty conic (d = 2) is a hyperbola, parabola, or ellipse, so in every case the number of loops is 1. For d = 3 the corollary gives a maximum number of 2 loops, and for d = 4 we get at most 4 loops. We have just seen examples of these numbers of loops above. One can show that the bound given in Harnack’s theorem is indeed sharp, i.e. for every d one can ﬁnd smooth real curves of degree d with exactly d−1 + 1 loops. 2 Proof. Assume that the statement is false, so that there are d−1 + 2 loops in a smooth 2 real plane curve C. Note that any two type 1 loops must intersect (which is impossible for a smooth curve), so there can be at most one type 1 loop. Hence assume that the ﬁrst d−1 + 1 loops are of type 2, and pick one point P1 , . . . , P(d−1)+1 on each of them. By 2 2 remark 6.2.9 we can assume that d ≥ 3, so pick d − 3 further distinct points Q1 , . . . , Qd−3 on the last loop (which can be of any type). We thus have a total of d−1 + 1 + d − 3 = 2 − d − 1 points Pi and Q j . 2 As in the proof of corollary 6.2.6 there is a curve C of degree d − 2 that passes through e all Pi and Q j . Compute the degree of the intersection scheme C ∩C . By B´ zout’s theorem, it must be degC · degC = d(d − 2). Counting the intersection points, we see that we have the d − 3 points Qi , and the d−1 + 1 points Pj that count with multiplicity at least 2 as 2 every type 2 loop divides the real projective plane in an interior and exterior region; so if C enters the interior of a type 2 loop it must exit it again somewhere. (It may also be that C is tangent to the loop or singular at the intersection point, but in this case the intersection multiplicity must be at least 2 too.)

P1 P4 C’ P2 Q1 P3 Q2

d2 2

Q3

So we get deg(C ∩C ) ≥ (d − 3) + 2 d −1 + 1 = d 2 − 2d + 1 > degC · degC . 2

6.

First applications of scheme theory

101

By B´ zout’s theorem it follows that C and C must have a common component. But C is e irreducible of degree degC > degC , so this is impossible. We thus arrive at a contradiction and conclude that the assumption of the existence of d−1 + 2 loops was false. 2 Corollary 6.2.10. Every isomorphism f : Pn → Pn is linear, i.e. it is of the form f (x) = A·x, where x = (x0 , . . . , xn ) and A is an invertible (n + 1) × (n + 1) matrix with elements in the ground ﬁeld. Proof. Let H ⊂ Pn be a hyperplane, and let L ⊂ Pn be a line not contained in H. Of course, H ∩ L is scheme-theoretically just one reduced point. As f is an isomorphism, f (H) ∩ f (L) must also be scheme-theoretically one reduced point, i.e. deg( f (H) ∩ f (L)) = 1. As degrees are always positive integers, it follows by B´ zout’s theorem that deg f (H) = e deg f (L) = 1. In particular, f maps hyperplanes to hyperplanes. Applying this to all hyperplanes {xi = 0} in turn, we conclude that f maps all coordinate functions xi to linear functions, so f (x) = A · x for some scalar matrix A. Of course A must be invertible if f has an inverse. 6.3. Divisors on curves. B´ zout’s theorem counts the number of intersection points of a e projective curve with a hypersurface. For example, if C ⊂ P2 is a plane cubic then the intersection of C with any line consists of 3 points (counted with their scheme-theoretic multiplicities). But of course not every collection of three points on C can arise this way, as three points will in general not lie on a line. So by reducing the intersections of curves to just the number of intersection points we are losing information about the possible conﬁgurations of intersection schemes. In contrast, we will now present a theory that is able to keep track of the conﬁgurations of (intersection) points on curves. Deﬁnition 6.3.1. Let C ⊂ Pn be a smooth irreducible projective curve. A divisor on C is a formal ﬁnite linear combination D = a1 P1 + · · · + am Pm of points Pi ∈ C with integer coefﬁcients ai . Obviously, divisors can be added and subtracted. The group of divisors on C is denoted DivC. Equivalently, DivC is the free abelian group generated by the points of C. The degree deg D of a divisor D = a1 P1 + · · · + am Pm is deﬁned to be the integer a1 + · · · + am . Obviously, the degree function is a group homomorphism deg : DivC → Z. Example 6.3.2. Divisors on a curve C can be associated to several objects: (i) Let Z ⊂ Pn be a zero-dimensional projective subscheme of Pn , and let P1 , . . . , Pm be the points in Z. Each of these points comes with a scheme-theoretic multiplicity ai (the length of the component of Z at Pi ) which is a positive integer. If the points Pi are on C, then a1 P1 + · · · am Pm is a divisor on C which we denote by (Z). It is called the divisor associated to Z. (ii) Let f ∈ k[x0 , . . . , xn ] be a homogeneous polynomial such that C is not contained in Z( f ). Then C ∩ Z( f ) is a zero-dimensional subscheme of Pn whose points lie in C, so by (i) there is an associated divisor (C ∩ Z( f )) on C. It is called the divisor of f and denoted ( f ); we can think of it as the zeros of f on C counted with their respective multiplicities. By B´ zout’s theorem, the number of such zeros is e deg( f ) = degC · deg f . (iii) Note that the intersection scheme C ∩ Z( f ) and therefore the divisor ( f ) do not change if we add to f an element of the ideal I(C). Hence there is a well-deﬁned divisor ( f ) for every non-zero f ∈ S(C)(d) . (iv) Assume that C ⊂ P2 , and that C = Z( f ) ⊂ P2 is another (not necessarily irreducible) curve that does not contain C as a component. Then the divisor ( f ) is also called the intersection product of C and C and denoted C ·C ∈ DivC.

102

Andreas Gathmann

Lemma 6.3.3. Let C ⊂ Pn be a smooth irreducible curve, and let f , g ∈ S(C) be non-zero homogeneous elements in the coordinate ring of C. Then ( f g) = ( f ) + (g). Proof. Let ( f g) = a1 P1 + · · · + am Pm . It is obvious that set-theoretically the zeros of f g are the union of the zeros of f and g, so f and g vanish at most at the points Pi . Let ( f ) = b1 P1 + · · · + bm Pm and (g) = c1 P1 + · · · cm Pm . We have to show that ai = bi + ci for all i = 1, . . . , m. Fix a certain i and choose an afﬁne open subset U = Spec R ⊂ C that contains Pi , but no other zero of f g. Then by deﬁnition we have ai = dimk R/( f g), bi = dimk R/( f ), and ci = dimk R/(g). The statement now follows from the exact sequence 0 −→ R/( f ) −→ R/( f g) −→ R/(g) −→ 0.

·g ·1

Deﬁnition 6.3.4. Let C ⊂ Pn be a smooth irreducible curve, and let ϕ ∈ K(C) be a non-zero f rational function. By deﬁnition we can write ϕ = g for some non-zero f , g ∈ S(C)(d) . We deﬁne the divisor of ϕ to be (ϕ) = ( f ) − (g) (this is well-deﬁned by lemma 6.3.3). It can be thought of as the zeros minus the poles of the rational function. Remark 6.3.5. Note that the divisor of a rational function always has degree zero: if ϕ = with f , g ∈ S(C)(d) , then deg(ϕ) = deg( f ) − deg(g) = d degC − d degC = 0 by B´ zout’s theorem. e Example 6.3.6. Let C = P1 , and consider the two homogeneous polynomials f (x0 , x1 ) = x0 x1 and g(x0 , x1 ) = (x0 − x1 )2 . Then ( f ) = P1 + P2 with P1 = (1 : 0) and P2 = (0 : 1), f and (g) = 2P3 with P3 = (1 : 1). The quotient g deﬁnes a rational function ϕ on P1 with (ϕ) = P1 + P2 − 2P3 . We have deg( f ) = deg(g) = 2 and deg(ϕ) = 0 (in accordance with remark 6.3.5). Remark 6.3.7. By lemma 6.3.3, the map K(C)\{0} → DivC that sends every rational function ϕ to its divisor (ϕ) is a group homomorphism, if we regard K(C)\{0} as an abelian group under multiplication. In particular, the subset of DivC of all divisors of the form (ϕ) is a subgroup of DivC. Deﬁnition 6.3.8. The Picard group (or divisor class group) PicC of C is deﬁned to be the group DivC modulo the subgroup of all divisors of the form (ϕ) for ϕ ∈ K(C)\{0}. If f ∈ S(C)(d) , we will usually still write ( f ) for the divisor class in PicC associated to f . Two divisors D1 and D2 are said to be linearly equivalent if D1 − D2 = 0 ∈ PicC, i.e. if they deﬁne the same divisor class. Remark 6.3.9. By remark 6.3.5, the degree function deg : DivC → Z passes to a group homomorphism deg : PicC → Z. So it makes sense to talk about the degree of a divisor class. We deﬁne Pic0 C ⊂ PicC to be the group of divisor classes of degree 0. Remark 6.3.10. The divisor group DivC is a free (and very “big”) abelian group and therefore not very interesting. In contrast, the divisor class group PicC has quite a rich structure that we want to study now in some easy examples. Lemma 6.3.11. Pic P1 ∼ Z (with an isomorphism being the degree homomorphism). In = other words, on P1 all divisors of the same degree are linearly equivalent.

f g

6.

First applications of scheme theory

103

Proof. Let D = a1 P1 + · · · am Pm be a divisor of degree zero, i.e. a1 + · · · + am = 0. We have to show that D is the divisor of a rational function. In fact, assume the Pi have homogeneous coordinates (xi : yi ); then ϕ = ∏(x yi − y xi )ai

i=1 m

is a rational function such that (ϕ) = D. Let us now move on to more complicated curves. We know already that smooth conics in P2 are isomorphic to P1 , so their Picard group is isomorphic to the integers too. Let us therefore consider cubic curves in P2 . We will compute PicC and show that it is not isomorphic to Z (thereby showing that cubic curves are not isomorphic to P1 ). Let us prove a lemma ﬁrst. Lemma 6.3.12. Let C = Z( f ) ⊂ P2 be a smooth cubic curve, and let C = Z(g) with g ∈ k[x0 , x1 , x2 ](d) be another curve that does not have C as a component. Assume that “three points of C ∩ C lie on a line”, i.e. that C · C contains three points P1 , P2 , P3 (that need not be distinct) such that there is a line L = Z(l) with C · L = P1 + P2 + P3 . Then there is a polynomial g ∈ k[x0 , x1 , x2 ](d−1) such that g = l · g in S(C). Proof. By B´ zout’s theorem we have C · L = P1 + · · · + Pd for some points Pi (that need e not be distinct, but they must contain the ﬁrst three given points P1 , P2 , P3 ). Let a ∈ k[x0 , x1 , x2 ](d−3) be a homogeneous polynomial such that Z(a) · L = P4 + · · · + Pd (it is obvious that this can always be found). Then Z(a f ) · L = P1 + · · · Pd too. Now pick any point Q ∈ L distinct from the Pi . As g and a f do not vanish at Q, we can ﬁnd a λ ∈ k such that g + λa f vanishes at Q. It follows that g + λa f vanishes on L at least at the d + 1 points P1 , . . . , Pd , Q. So it follows by B´ zout’s theorem that Z(g + λa f ) contains e the line L, or in other words that g + λa f = lg for some g . Passing to the coordinate ring S(C) = k[x0 , x1 , x2 ]/I(C) we get the desired result. Proposition 6.3.13. Let C ⊂ P2 be a smooth cubic curve, and let P, Q be distinct points on C. Then P − Q = 0 in PicC. In other words, there is no rational function ϕ ∈ K(C)\{0} with (ϕ) = P − Q, i.e. no rational function that has exactly one zero which is at P, and exactly one pole which is at Q. Remark 6.3.14. It follows from this proposition already that a smooth plane cubic curve is not isomorphic to P1 (as the statement of the proposition is false for P1 by lemma 6.3.11). Proof. Assume the contrary. Then there is a positive integer d and homogeneous polynomials f , g ∈ S(C)(d) such that (i) There are points P1 , . . . , P3d−1 and P = Q such that ( f ) = P1 + · · · + P3d−1 + P and (g) = P1 + · · · + P3d−1 + Q

f (hence (ϕ) = P − Q for ϕ = g ). (ii) Among the P1 , . . . , P3d−1 there are at least 2d − 1 distinct points. (If this is not the case in the ﬁrst place, we can replace f by f · l and g by g · l some linear function l that vanishes on C at three distinct points that are not among the Pi . This raises the degree of the polynomials by 1 and the number of distinct points by 3, so by doing this often enough we can get at least 2d − 1 distinct points.)

Pick d minimal with these properties. If d = 1 then ( f ) = P1 + P2 + P and (g) = P1 + P2 + Q, so both f and g deﬁne the unique line through P1 and P2 (or the tangent to C at P1 if P1 = P2 ). In particular, it follows that P = Q as well, which is a contradiction. So we can assume that d > 1. We can rearrange the Pi such that P2 = P3 , and such that P1 = P2 if there are any equal points among the Pi .

104

Andreas Gathmann

Now consider curves given by linear combinations λ f + µg. These curves will intersect C at least in the points P1 , . . . , P3d−1 (as Z( f ) and Z(g) do). Note that for any point R ∈ C we can adjust λ and µ so that (λ f + µg)(R) = 0. Such a curve will then have intersection divisor P1 + · · · + P3d−1 + R with C. In other words, by passing to linear combinations of f and g we can assume that the last points P and Q in the divisors of f and g are any two points we like. We choose P to be the third intersection point of P1 P2 with C, and Q to be the third intersection point of P1 P3 with C. By lemma 6.3.12, it now follows that f = l · f and g = l · g in S(C) for some linear functions l and l that have intersection divisors P1 + P2 + P and P1 + P3 + Q with C. Hence ( f ) = P4 + · · · + P3d−1 + P3 and (g ) = P4 + · · · + P3d−1 + P2

Note that these f and g satisfy (i) for d replaced by d − 1, as P2 = P3 by assumption. Moreover, f and g satisfy (ii) because if there are any equal points among the Pi at all, then by our relabeling of the Pi there are only two distinct points among P1 , P2 , P3 , so there must still be at least 2(d − 1) − 1 distinct points among P4 , . . . , P3d−1 . This contradicts the minimality of d and therefore proves the proposition. Corollary 6.3.15. Let C be a smooth cubic curve, and let P0 ∈ C be a point. Then the map C → Pic0 C, is a bijection. Proof. The map is well-deﬁned and injective by proposition 6.3.13. We will show that it is surjective. Let D = P1 + · · · + Pm − Q1 − · · · − Qm be any divisor of degree 0. If m > 1 let P be the third intersection point of P1 P2 with C, and let Q be the third intersection point of Q1 Q2 with C. Then P1 + P2 + P and Q1 + Q2 + Q are both the divisors of linear forms on C. The quotient of these linear forms is a rational function whose divisor P1 + P2 + P − Q1 − Q2 − Q is therefore 0 in PicC. It follows that D = P3 + · · · + Pm + Q − Q3 − · · · − Qm − P. We have thus reduced the number m of (positive and negative) points in D by 1. Continuing this process, we can assume that m = 1, i.e. D = P − Q for some P, Q ∈ C. Now let P be the third intersection point of PP0 with C, and let Q be the third intersection point of P Q with C. Then P + P + P0 = P + Q + Q in PicC as above, so D = P − Q = Q − P0 , as desired. 6.4. The group structure on a plane cubic curve. Let C ⊂ P2 be a smooth cubic curve. Corollary 6.3.15 gives a canonical bijection between the variety C and the abelian group Pic0 C, so between two totally different mathematical objects. Using this bijection, we can give C a group structure (after choosing a base point P0 as in the corollary) and Pic0 C the structure of a smooth projective variety. We should mention that Pic0 C can be made into a variety (the so-called Picard variety) for every smooth projective curve C; it is in general not isomorphic to C however. (If C is not P1 one can show that the map P → P − P0 of corollary 6.3.15 is at least injective, so we can think of C as a subvariety of the Picard variety.) In contrast, the statement that C can be made into an abelian group is very special to cubic plane curves (or to be precise, to curves of genus 1). Curves of other types do not admit such a group structure. Example 6.4.1. Let us investigate the group structure on C geometrically. If P and Q are two points on C (not necessarily distinct), we denote by ϕ(P, Q) the third point of intersection of the line PQ with C, i.e. the unique point of C such that P + Q + ϕ(P, Q) is linearly equivalent to the divisor of a linear function. We will denote the group structure P → P − P0

6.

First applications of scheme theory

105

on C by ⊕, to distinguish it from the addition of points in DivC or PicC. Consequently, we write P for the inverse of P, and n P for P ⊕ · · · ⊕ P (n times). Of course, the zero element of the group structure on C is just P0 . By construction, P ⊕ Q is the unique point of C such that (P − P0 ) + (Q − P0 ) = (P ⊕ Q) − P0 in PicC, i.e. P + Q = (P ⊕ Q) + P0 . Now let R = ϕ(P, Q). Then P + Q + R = (P ⊕ Q) + P0 + R ∈ PicC, so P ⊕ Q = ϕ(R, P0 ) = ϕ(ϕ(P, Q), P0 ). In other words, to construct the point P ⊕ Q we draw a line through P and Q. Then we draw another line through the third intersection point R of this line with C and the point P0 . The third intersection point of this second line with C is P ⊕ Q (see the picture below on the left). Similarly, to construct P we are looking for a point such that (P − P0 ) + (( P) − P0 ) = 0, so P + ( P) = 2P0 . In the same way as above we conclude P = ϕ(ϕ(P0 , P0 ), P). In other words, to construct the inverse P we draw the tangent to C through P0 . Then we draw another line through the (scheme-theoretic) third intersection point R of this tangent with C and the point P. The third intersection point of this second line with C is P:

C P0 R QP P Q P P0 P R C

Of special geometric importance are the (tangent) lines that meet C in a point with multiplicity (at least) 3. In analogy with the real analysis case such points will be called inﬂection points: Deﬁnition 6.4.2. Let C ⊂ P2 be a smooth curve. A point P ∈ C is called an inﬂection point of C if the tangent line to C at P intersects C in P with multiplicity at least 3. Such a tangent line is then called a ﬂex.

flex C

inflection point

For cubic curves C, any line intersects C in three points, so P ∈ C is a ﬂex if and only if 3P is the divisor of a linear function. Let us ﬁrst prove that there are some inﬂection points on every smooth cubic curve. Lemma 6.4.3. Let C = Z( f ) ⊂ P2 be a smooth curve of degree d. Then h = det ∂2 f ∂xi ∂x j

0≤i, j≤2

is a homogeneous polynomial of degree 3(d − 2). (It is called the Hessian polynomial of C. The corresponding curve H = Z(h) ⊂ P2 is called the Hessian curve of C.) Then P ∈ C is an inﬂection point of C if and only if P ∈ H.

106

Andreas Gathmann

Proof. By a linear change of coordinates we can assume that P = (1 : 0 : 0) and that the i j k tangent line to C at P is L = {x2 = 0}. Let f = ∑i+ j+k=d ai, j,k x0 x1 x2 . In inhomogeneous coordinates (x0 = 1) the restriction of f to L is

i f (1, x1 , 0) = ∑ ad−i,i,0 x1 . i=0 d

As f passes through P and is tangent to L there, f |L (x1 ) must have a zero of order at least 2 at P, so ad,0,0 = ad−1,1,0 = 0. Now note that

∂2 f 2 (P) = d(d − 1)ad,0,0 , ∂x0 ∂2 f ∂x0 ∂x2 (P) = (d − 1)ad−1,0,1 , ∂2 f ∂x0 ∂x1 (P) = (d − 1)ad−1,1,0 , ∂2 f (P) = 2ad−2,2,0 . ∂x2

1

So the Hessian polynomial at P has the form 0 0 0 2ad−2,2,0 h(P) = det (d − 1)ad−1,0,1 ∗ In the same way, note that (

(d − 1)ad−1,0,1 . ∗ ∗

∂f ∂f ∂f , , )(P) = (dad,0,0 , ad−1,1,0 , ad−1,0,1 ) = (0, 0, ad−1,0,1 ), ∂x0 ∂x1 ∂x2

which must be a non-zero vector by the Jacobian criterion of proposition 4.4.8 (ii) as C is smooth at P. So ad−1,0,1 = 0, and therefore h(P) = 0 if and only if ad−2,2,0 = 0. This is the case if and only if f |L (x1 ) vanishes to order at least 3 at P, i.e. if and only if P is an inﬂection point. Corollary 6.4.4. Every smooth cubic curve in P2 has exactly 9 inﬂection points. Proof. By lemma 6.4.3 the inﬂection points of C are precisely the points of C ∩ H ⊂ P2 , where H is the Hessian curve of C. But by B´ zout’s theorem, deg(C ∩ H) = d · 3(d − 2) = 9 e for d = 3. So we only have to check that every point in C ∩ H occurs with intersection multiplicity 1. Let us continue with the notation of the proof of lemma 6.4.3, and assume that P is an inﬂection point, so that a3,0,0 = a2,1,0 = a1,2,0 = 0. We will show that the Hessian curve H is smooth at P and has a tangent line different from that of C (i.e. its tangent line is not L = {x2 = 0}. Both statements follow if we can prove that h(1, x1 , x2 ) contains the 2 monomial x1 with a non-zero coefﬁcient, i.e. that h contains the monomial x0 x1 with a non-zero coefﬁcient. But note that 0 0 2a2,0,1 x0 + a1,1,1 x1 , 0 6a0,3,0 x1 ∗ h(x2 = 0) = det 2a2,0,1 x0 + a1,1,1 x1 ∗ ∗

2 so the x0 x1 -coefﬁcient of h is −24a2 a0,3,0 . The corollary now follows from the following 2,0,1 two observations:

(i) the Jacobian matrix of f at P is (3a3,0,0 , a2,1,0 , a2,0,1 ). As C is smooth this matrix must have rank 1 by proposition 4.4.8 (ii). But a3,0,0 and a2,1,0 are zero already, so a2,0,1 = 0. 3 (ii) We know already that f |L = a0,3,0 x1 . As L cannot be a component of C, it follows that a0,3,0 = 0.

6.

First applications of scheme theory

107

Remark 6.4.5. If C is a smooth curve of degree d in P2 , we would still expect from B´ zout’s e theorem that C has 3d(d − 2) inﬂection points. This is indeed the “general” number, but for d > 3 it may occur that C and its Hessian H do not intersect at all points with multiplicity 1, so that there are fewer than 3d(d − 2) inﬂection points. Lemma 6.4.6. Let C ⊂ P2 be a smooth cubic curve, and choose an inﬂection point P0 as the zero element of the group structure on C. Then a point P ∈ C is an inﬂection point if and only if 3 P = P0 . In particular, there are exactly 9 3-torsion points in PicC, i.e. 9 points P ∈ C such that 3 P = P0 . Proof. Assume that P0 is an inﬂection point, i.e. 3P0 is the divisor of a linear function on C. Then P is an inﬂection point if and only if 3P is the divisor of a linear function too, which is the case if and only if 3P − 3P0 = 3(P − P0 ) is the divisor of a rational function (a quotient of two linear functions). This in turn is by deﬁnition the case if and only if 3 P = P0 . It then follows by corollary 6.4.4 that there are exactly 9 3-torsion points in PicC. Corollary 6.4.7. Let C ⊂ P2 be a smooth cubic curve. Then any line through two inﬂection points of C passes through a third inﬂection point of C. Proof. Choose an inﬂection point P0 ∈ C as the zero element for the group structure on C. Now let P and Q be two inﬂection points, and let R = ϕ(P, Q) be the third intersection point of PQ with C. Then P + Q + R is the divisor of a linear function and hence equal to 3P0 in PicC. It follows that 3(R − P0 ) = 3(2P0 − P − Q) = 3(P0 − P) + 3(P0 − Q) = 0 ∈ PicC. So 3 R = P0 , i.e. R is an inﬂection point by lemma 6.4.6. Example 6.4.8. There is an interesting application of the group structure on a cubic curve to cryptography. The key observation is that “multiplication is easy, but division is hard”. More precisely, assume that we are given a speciﬁc cubic curve C and a zero point P0 ∈ C for the group structure. (For practical computations one will usually do this over a ﬁnite ﬁeld to avoid rounding errors. The group structure exists in these cases too by exercise 6.7.10.) Then: (i) Given any point P and a positive integer n, the point n P can be computed quickly, even for very large n (think of numbers with hundreds of digits): (a) By repeatedly applying the operation P → P ⊕ P, we can compute all points 2k P for all k such that 2k < n. (b) Now we just have to add these points 2k P for all k such that the k-th digit in the binary representation of n is 1. This computes the point n P in a time proportional to log n (i.e. in a very short time). (ii) On the other hand, given a point P and a positive integer n, it is essentially impossible to compute a point Q such that n Q = P (assuming that such a point exists). This is not a mathematically precise statement; there is just no algorithm known to exist that can perform the “inverse” of the multiplication P → n P in shorter time than a simple trial-and-error approach. Of course, if the ground ﬁeld is large and C contains enough points, this is practically impossible. In the same way, given two points P and Q on C, there is no way to ﬁnd the (smallest) number n such that n Q = P except trying out all integers in turn. Again, if n has hundreds of digits this is of course practically impossible. Using this idea, assume that Alice wants to send a secret message to Bob. We can think of this message as just a number N (every message can be converted into a sequence of numbers, of course). There is an easy way to achieve this if they both know a secret

108

Andreas Gathmann

key number N0 : Alice just sends Bob the number N + N0 in public, and then Bob can reconstruct the secret N by subtracting the key N0 from the transmitted number N + N0 . Any person who observed the number N + N0 in transit but does not know the secret key N0 is not able to reconstruct the message N. The problem is of course that Alice and Bob must ﬁrst have agreed on a secret key N0 , which seems impossible as they do not have a method for secure communication yet. This is where our cubic curve can help. Let us describe a (simpliﬁed) version of what they might do. Alice and Bob ﬁrst (publicly) agree on a ground ﬁeld, a speciﬁc cubic curve C, a zero point P0 ∈ C, and another point P ∈ C. Now Alice picks a secret (very large) integer a, and Bob picks a secret integer b. They are not telling each other what their secret numbers are. Instead, Alice computes a P and sends (the coordinates of) this point to Bob. In the same way, Bob computes b P and sends this point to Alice. Now the point ab P can be used as a secret key number N0 : (i) Alice got the information about b P from Bob and knows her own secret number a, so she can compute ab P = a (b P). (ii) In the same way, Bob knows ab P = b (a P). (iii) The only information that Alice and Bob exchanged was the data of the cubic curve chosen, P, a P, and b P. But we have just noted that there is no practical way to reconstruct a and b from this information, so anybody else will not be able to determine the secret key ab P from this data. 6.5. Plane cubic curves as complex tori. We will now restrict our attention to the ground ﬁeld k = C and see how smooth plane cubic curves arise in complex analysis in a totally different way. We will only sketch most arguments; more details can be found e.g. in [K] section 5.1 (and many other books on complex analysis). Let U ⊂ C be an open set in the classical topology. Recall that a (set-theoretic) function f : U → C is called holomorphic at z0 ∈ U if it is complex differentiable at z0 , i.e. if the limit f (z) − f (z0 ) f (z) := lim z→z0 z − z0 exists. A function f : U\{z0 } → C is called meromorphic if there is a number n ∈ Z and a holomorphic function f˜ : V → C in a neighborhood V of z0 in U such that f (z) = (z − z0 )n · f˜(z) and f˜(z0 ) = 0 on V . Note that the number n is then uniquely determined; it is called the order of f at z0 and denoted ordz0 f . If n > 0 we say that f (z) has a zero of order n at z0 . If n < 0 we say that f (z) has a pole of order −n at z0 . A function that is meromorphic at z0 is holomorphic at z0 if and only if its order is non-negative. Example 6.5.1. Any regular function on A1 (i.e. any polynomial in z) is a holomorphic C function on C. Similarly, any rational function ϕ on A1 is a meromorphic function on C. C The notion of zeros and poles of ϕ as a meromorphic function agrees with our old one of deﬁnition 6.3.4, so the multiplicity of a point z ∈ C in the divisor of ϕ is precisely the order of ϕ at z. Conversely, there are holomorphic (resp. meromorphic) functions on C that are not regular (resp. rational), e.g. f (z) = ez . Remark 6.5.2. Although the deﬁnition of holomorphic, i.e. complex differentiable functions is formally exactly the same as that of real differentiable functions, the behavior of the complex and real cases is totally different. The most notable differences that we will need are: (i) Every holomorphic function is automatically inﬁnitely differentiable: all higher derivatives f (k) exist for k > 0 and are again holomorphic functions.

6.

First applications of scheme theory

109

(ii) Every holomorphic function f is analytic, i.e. it can be represented locally around every point z0 by its Taylor series. The radius of convergence is “as large as it can be”, i.e. if f is holomorphic in an open ball B around z0 , then the Taylor series of f at z0 converges and represents f at least on B. Consequently, a meromorphic function f of order n at z0 can be expanded in a Laurent series as f (z) = ∑k≥n ck (z − z0 )k . The coefﬁcient c−1 of this series is called the residue of f at z0 and denoted resz0 f . (iii) (Liouville’s theorem) Every function f that is holomorphic and bounded on the whole complex plane C is constant. (iv) (Identity theorem) Let f and g be holomorphic functions on a connected open subset U ⊂ C. If f and g agree on any open subset V ⊂ U then they agree on U. By (ii) this is e.g. the case if their Taylor series agree at some point in U. One should compare this to the algebro-geometric version of remark 2.1.9. (v) (Residue theorem) If γ is a closed (positively oriented) contour in C and f is a meromorphic function in a neighborhood of γ and its interior that has no poles on γ itself, then

Z

γ

f (z) dz = 2πi ∑ resz0 f (z),

z0

with the sum taken over all z0 in the interior of γ (at which f has poles). In particular, if f is holomorphic then this integral vanishes. In this section we will study a particular meromorphic function on C associated to a lattice. Let us describe the construction. Fix once and for all two complex numbers ω1 , ω2 ∈ C that are linearly independent over R, i.e. that do not lie on the same real line in C through the origin. Then the subset Λ = Zω1 + Zω2 = {mω1 + nω2 ; m, n ∈ Z} ⊂ C is called a lattice in C. Obviously, the same lattice in C can be obtained by different choices of ω1 and ω2 . The constructions that we will make in this section will only depend on the lattice Λ and not on the particular choice of basis ω1 , ω2 .

Im z

ω2

Re z

ω1

Proposition and Deﬁnition 6.5.3. Let Λ = Zω1 + Zω2 be a lattice in C. There is a meromorphic function ℘(z) on C deﬁned by ℘(z) = 1 1 1 + ∑ − 2 2 2 z (z − ω) ω ω∈Λ\{0} .

It is called the Weierstraß ℘-function. It has poles of order 2 exactly at the lattice points. Proof. It is a standard fact that an (inﬁnite) sum of holomorphic functions is holomorphic at z0 provided that the sum converges uniformly in a neighborhood of z0 . We will only sketch the proof of this convergence: let z0 ∈ C\Λ be a ﬁxed point that is not in the lattice. Then every summand is a holomorphic function in a neighborhood of z0 . The expansions

110

Andreas Gathmann

of these summands for large ω are 1 1 1 − = 2 (z0 − ω)2 ω2 ω 1 z0 − 1 = 3 + terms of order at least 1 − z0 ω ω

1 ω4

so the summands grow like ω3 . Let us add up these values according to the absolute value of ω. As the number of lattice points with a given absolute value (approximately) 1 1 equal to N grows linearly with N, the ﬁnal sum behaves like ∑N N · N 3 = ∑N N 2 , which is convergent. Note that the sum would not have been convergent without subtraction of the constant 1 1 in each summand, as then the individual terms would grow like ω2 and therefore the ω2 1 ﬁnal sum would be of the type ∑N N , which is divergent. Remark 6.5.4. It is a standard fact that in an absolutely convergent series as above all manipulations (reordering of the summands, term-wise differentiation) can be performed as expected. In particular, the following properties of the ℘-function are obvious: (i) The ℘-function is an even function, i.e. ℘(z) = ℘(−z) for all z ∈ C. In particular, its Laurent series at 0 contains only even exponents. −2 (ii) Its derivative is ℘ (z) = ∑ω∈Λ (z−ω)3 . It is an odd function, i.e. ℘ (z) = −℘ (−z). In particular, its Laurent series at 0 contains only odd exponents. It has poles of order 3 exactly at the lattice points. (iii) The ℘-function is doubly periodic with respect to Λ, i.e. ℘(z0 ) =℘(z0 +ω) for all z0 ∈ C and ω ∈ Λ. To show this note ﬁrst that it is obvious from (ii) that ℘ (z0 ) = ℘ (z0 + ω). Now integrate ℘ (z) along the closed contour γ = γ1 + γ2 + γ3 + γ4 shown in this picture:

z0 γ1 γ2 γ3 ω _ 2 ω z 0+ ω

γ4 −ω _ −ω 2 0

Of course, the result is 0, since ℘ is an integral of ℘ . But also the integral along γ2 cancels the integral along γ4 as ℘ (z) is periodic. The integral along γ3 is equal to ℘(− ω ) −℘( ω ) and hence vanishes too as ℘(z) is an even function. So we 2 2 conclude that

Z

0=

γ1

℘ (z) dz = ℘(z0 + ω) −℘(z0 ),

i.e. ℘(z) is periodic with respect to Λ too. Lemma 6.5.5. The ℘-function associated to a lattice Λ satisﬁes a differential equation ℘ (z)2 = c3℘(z)3 + c2℘(z)2 + c1℘(z) + c0 for some constants ci ∈ C that depend on Λ. Proof. By remark 6.5.4 (ii) ℘ (z)2 is an even function with a pole of order 6 at 0. Hence its Laurent series around 0 is a−6 a−4 a−2 ℘ (z)2 = 6 + 4 + 2 + a0 + terms of order z>0 z z z for some constants a−6 , a−4 , a−2 ∈ C. The functions ℘(z)3 , ℘(z)2 , ℘(z), and 1 are also even, and they have poles of order 6,4,2, and 0, respectively. Hence there are constants c−6 , c−4 , c−2 , c0 ∈ C such that the series of the linear combination f (z) := ℘ (z)2 − c3℘(z)3 − c2℘(z)2 − c1℘(z) − c0

6.

First applications of scheme theory

111

has only positive powers of z. We conclude that f (z) is holomorphic around 0 and vanishes at 0. By the identity theorem of remark 6.5.2 (iv) it then follows that f = 0 everywhere. Remark 6.5.6. An explicit computation shows that the coefﬁcients ci in lemma 6.5.5 are given by 1 1 c3 = 4, c2 = 0, c1 = −60 ∑ , c0 = −140 ∑ . 4 ω ω6 ω∈Λ\{0} ω∈Λ\{0} Proposition 6.5.7. Let Λ ⊂ C be a given lattice, and let C ⊂ P2 be the cubic curve C

2 3 2 2 3 C = {(x0 : x1 : x2 ) ; x2 x0 = c3 x1 + c2 x1 x0 + c1 x1 x0 + c0 x0 }

for the constants ci ∈ C of lemma 6.5.5. Then there is a bijection Φ : C/Λ → C, z → (1 : ℘(z) : ℘ (z)). Proof. As℘(z) and℘ (z) are periodic with respect to Λ and satisfy the differential equation of lemma 6.5.5, it is clear that Φ is well-deﬁned. (Strictly speaking, for z = 0 we have to note that ℘(z) has a pole of order 2 and ℘ (z) has a pole of order 3, so ℘(z) = fz(z) and 2 ℘ (z) = g(z) locally around 0 for some holomorphic functions f , g around 0 that do not z3 vanish at 0. Then (1 : ℘(0) : ℘ (0)) = (z3 : z f (z) : g(z))|z=0 = (0 : 0 : 1), so Φ is well-deﬁned at 0 too.) Now let (x0 : x1 : x2 ) ∈ C be a given point; we will show that it has exactly one inverse image point under Φ. By what we have just said this is obvious for the “point at inﬁnity” (0 : 0 : 1), so let us assume that we are not at this point and hence pass to inhomogeneous coordinates where x0 = 1. We will ﬁrst look for a number z ∈ C such that ℘(z) = x1 . To do so, consider the integral

Z

℘ (z)

γ ℘(z) − x1

dz

over the boundary of any “parallelogram of periodicity” as in the following picture:

Im z

Re z

The integrals along opposite sides of the parallelogram vanish because of the periodicity of ℘ and ℘ , so the integral must be 0. So by the residue theorem of remark 6.5.2 (v) we get ℘ (z) . (∗) 0 = ∑ resz0 ℘(z) − x1 z ∈C/Λ

0

Now note that if F(z) is any meromorphic function of order n around 0 then res0 F (z) nan zn−1 + · · · = res0 = n, F(z) an zn + · · ·

so we conclude from (∗) that ∑z0 ∈C/Λ ordz0 (℘(z) − x1 ) = 0: the function ℘(z) − x1 has as many zeros as it has poles in C/Λ, counted with multiplicities. (This is a statement

112

Andreas Gathmann

in complex analysis corresponding to remark 6.3.5.) As ℘(z) has a pole of order 2 in the lattice points, it thus follows that there are exactly two points z1 , z2 ∈ C/Λ such that ℘(z) = x1 . Since the ℘-function is an even function, these two points are obviously negatives of each other. Now as ℘ is an odd function, it follows that ℘ (z1 ) = −℘ (z2 ). So if we specify ℘(z) and ℘ (z) there is exactly one point z ∈ C/Λ leading to the given image point in C. Remark 6.5.8. We are again in a similar situation as in corollary 6.3.15: we have a bijection between a group C/Λ and a variety C. In fact, one can show that the group structure of C/Λ is precisely the same as that of Pic0 C, so we have just rediscovered our old group structure on a plane cubic curve. But the group structure is a lot more obvious in this new picture: e.g. the n-torsion points of C are easily read off to be 1 (iω1 + jω2 ) ; 0 ≤ i, j < n . n In particular, there are exactly n2 points P ∈ C such that n P = 0, in accordance with exercise 6.7.11 and lemma 6.4.6. It should be said however that the bijection of proposition 6.5.7 differs from that of corollary 6.3.15 in that both C/Λ and C can independently be made into a complex manifold (which you should roughly think of as a variety whose structure sheaf consists of holomorphic functions instead of just polynomial functions). The map Φ of the above proposition is then an isomorphism between these two complex manifolds. Remark 6.5.9. The topology of a plane cubic curve becomes very clear from proposition 6.5.7: it is just a parallelogram with opposite sides identiﬁed, i.e. a torus. This agrees with our earlier statements that a smooth plane cubic curve has genus 1, and that the genus should be thought of as the number of “holes” in the (real) surface. 6.6. Where to go from here. After having discussed some basic algebraic geometry we now want to sketch which important parts of the general theory are still missing in our framework. Example 6.6.1. Intersection theory. Let X ⊂ Pn be a projective variety of dimension r, and let X1 , . . . , Xr ⊂ Pn be r hypersurfaces. If the hypersurfaces are in sufﬁciently general position, the intersection X1 ∩ · · · ∩ Xr ∩ X will be zero-dimensional. B´ zout’s theorem then e tells us that the intersection consists of exactly deg X1 · · · · · deg Xr · deg X points, counted with multiplicities. There is obvious room for generalizations here. Assume that we do not have r hypersurfaces X1 , . . . , Xr , but rather closed subvarieties X1 , . . . , Xs of X whose codimensions in X add up to r. If these subvarieties are in sufﬁciently general position then we still expect the intersection X1 ∩ · · · ∩ Xs ∩ X to be zero-dimensional. So we can still ask for the number of points in the intersection and expect a ﬁnite answer. If X = Pr is projective space itself, then the answer is still just deg X1 · · · · · deg Xs : in Pr the degree is multiplicative when taking intersections. For general X the situation is a lot more subtle though — there is no single number that can be associated to any subvariety of X and that is just multiplicative with respect to intersections. This is easy to see: if e.g. X = P1 × P1 and we consider the three 1-dimensional subvarieties of X X1 = P1 × {0}, X2 = P1 × {1}, X3 = {0} × P1 ,

then X1 ∩ X2 is empty, so if there were numbers associated to X1 and X2 whose product gives the number of intersection points (namely zero), then one of these two numbers (say for X1 ) must obviously itself be zero. But then the product of the numbers for X1 and X3 would also be zero, although X1 and X3 intersect in precisely one point.

6.

First applications of scheme theory

113

It turns out however that there is a ﬁnite collection of numbers that can be associated to any subvariety of X such that the number of points in X1 ∩ · · · ∩ Xs is given by an explicit multilinear form in these collections of numbers. For example, in the P1 × P1 case above, curves (like X1 , X2 , X3 given above) are characterized by their bidegree (i.e. the bidegree of the deﬁning equation). In our example, the bidegrees of X1 , X2 , and X3 are (1, 0), (1, 0), and (0, 1), respectively. Two curves of bidegrees (d1 , e1 ) and (d2 , e2 ) then intersect in exactly d1 e2 + d2 e1 points. Setting up a corresponding theory for any variety X is the object of intersection theory. It is essentially a well-established theory that can be set up both in algebraic geometry and (for the ground ﬁeld C) topology. In the latter case it is a part of algebraic topology. In both cases the theory allows you to answer most questions concerning numbers of intersection points quite effectively (and without the need for computer algebra techniques). Intersection theory is used in one form or the other in virtually every geometric ﬁeld of mathematics. Example 6.6.2. Sheaves and vector bundles. Let us illustrate the idea behind vector bundles by an example. In section 4.5 we have shown that every smooth cubic surface in P3 has exactly 27 lines on it. We did this by ﬁrst proving that the number of lines does not depend on the particular cubic chosen, and then calculating the number for a speciﬁc cubic for which the answer happened to be directly computable. Now let us consider a slightly more difﬁcult setting. Let X ⊂ P4 be a (3-dimensional) smooth hypersurface of degree 5. We will see momentarily that we again expect there to be a ﬁnite number of lines in X. So again we ask for the number of such lines. Compared to the cubic surface case it is still true that the answer does not depend on the particular quintic hypersurface chosen. There is no speciﬁc quintic any more however for which we can read off the answer by simply writing down all the lines explicitly. So we need to apply a different technique to obtain the answer. As before, we ﬁrst consider again the Grassmannian variety G(1, 4) of lines in P4 (see exercise 3.5.4). The dimension of G(1, 4) is 6. Now deﬁne the set E := {(L, f ) ; L ∈ G(1, 4), f is a homogeneous polynomial of degree 5 on L ∼ P1 }, = so elements of E are pairs of a line in P4 and a quintic equation on this line. There is an obvious projection map π : E → G(1, 4) given by forgetting f . We claim that E is a variety in a natural way. In fact, as in exercise 3.5.4 consider the open subset U ⊂ G(1, 4) isomorphic to A6 (with coordinates a2 , b2 , a3 , b3 , a4 , b4 ) where the line L ∈ U can be represented by the matrix 1 0 0 a2 1 b2 a3 b3 a4 b4 . (1)

For every such line we can obviously take x0 and x1 as homogeneous coordinates on L ∼ = i 5−i P1 , so every quintic equation on L is of the form ∑i ci x0 x1 for some c0 , . . . , c5 . Then π−1 (U) can obviously be thought of as a 12-dimensional afﬁne space with coordinates a2 , a3 , a4 , b2 , b3 , b4 , c0 , . . . , c5 . As E can be covered by these spaces, it is a 12-dimensional variety. Note that the ﬁbers π−1 (L) for L ∈ G(1, 4) are all 6-dimensional vector spaces, namely the spaces of degree-5 homogeneous polynomials on L. They are not just 6-dimensional afﬁne spaces but rather linear afﬁne spaces in the sense that it is meaningful to add two polynomials on L, and to multiply them with a scalar. So two points in E that map to the same base point in G(1, 4) can be “added”, just by summing up their coordinates ci . In contrast, it does not make much sense to add the coefﬁcients ai and bi in two matrices as in (1), as the resulting line is not related to the two original lines in any obvious way. So

114

Andreas Gathmann

although the coordinates a2 , a3 , a4 , b2 , b3 , b4 in U live in an afﬁne space A6 , it does not make sense to think of this A6 as a vector space. Note also that E is not just the direct product of G(1, 4) with a constant 6-dimensional vector space k[x0 , x1 ](5) , as the coordinates that we can use on the line L vary with the line. Only the ﬁbers of π are all 6-dimensional vector spaces. We say that E is a vector bundle of rank 6 on G(1, 4). Now let us return to our original question: to count the lines on X. Let f ∈ k[x0 , . . . , x4 ](5) be the polynomial whose zero locus is X. There is an obvious morphism σ : G(1, 4) → E, L → (L, f |L ) (2) such that π◦σ = idG(1,4) . Such a morphism is called a section of E: it assigns to every point L in the base G(1, 4) an element in the vector space π−1 (L) “sitting over” L. Note that this can indeed be thought of as a section in the sheaf-theoretic sense: suppose that we have an open cover {Ui } of G(1, 4) and morphisms σi : Ui → π−1 (Ui ) such that π ◦ σi = idUi (i.e. on every Ui we associate to any point L ∈ Ui an element in the vector space π−1 (L)). If σi = σ j on Ui ∩U j for all i, j, then there is obviously a global section σ : U → E that restricts to the σi on the Ui . In other words, we can think of the vector bundle E as a sheaf, with E(U) (in the sense of deﬁnition 2.2.1) being the space of all morphisms σ : U → π−1 (U) such that π ◦ σ = idU . Finally, return to our speciﬁc section σ in (2). As the ﬁbers of π are vector spaces, there is also a well-deﬁned zero section σ0 : G(1, 4) → E, L → (L, 0).

Obviously, a line L lies in the quintic hypersurface X if and only if f |L = 0, i.e. if and only if σ(L) = σ0 (L). So the number of lines we are looking for is simply the number of intersection points of σ(G(1, 4)) and σ0 (G(1, 4)). As these are both 6-dimensional varieties in the 12-dimensional variety E, we expect a ﬁnite number of such intersection points, showing that we expect a ﬁnite number of lines in X. Their number is now given by intersection theory methods as explained in example 6.6.1. It can be computed explicitly and the result turns out to be 2875. (To mention the corresponding keywords: we need the 6th Chern class of the vector bundle E on G(1, 4), and the result can be obtained using Schubert calculus, i.e. the intersection theory on the Grassmannian G(1, 4).) Another example of a vector bundle on a smooth r-dimensional variety X is the tangent bundle: it is just the rank-r vector bundle whose ﬁber over a point P ∈ X is the tangent space TX,P . The dual vector bundle (i.e. the rank-r bundle whose ﬁber over a point P ∈ X is the dual vector space to TX,P ) is called the cotangent bundle and denoted ΩX,P . It can be thought of as the vector bundle of differential forms on X. Any operations that can be done with vector spaces can be done with vector bundles as well, just by performing the corresponding operation in every ﬁber. So there are e.g. direct sums of vector bundles, tensor products, symmetric products, exterior products, and so on. If X is a smooth r-dimensional variety, the r-th exterior power Λr ΩX of the cotangent bundle is called the canonical bundle and denoted KX . Obviously it is a vector bundle of rank 1: such bundles are called line bundles. Its importance (and name) stems from the fact that it is canonically given for any smooth variety X. Vector bundles (and corresponding sheaves) occur in almost any branch of algebraic geometry, as well as in topology and differential geometry. Example 6.6.3. Sheaf cohomology. Let X be a variety, and let E be a vector bundle on X. By the remark above, (global) sections σ : X → E can be added and multiplied with a scalar, so the space of global sections is in fact a vector space over the ground ﬁeld k. It is denoted H 0 (X, E).

6.

First applications of scheme theory

115

As an example, let X ⊂ P2 be a curve, and let n be an integer. For an open subset U ⊂ X deﬁne f E(U) = ; f , g ∈ S(X) homogeneous with deg f − deg g = n, g(P) = 0 for all P ∈ U . g These data form a sheaf E that can be thought of as the sheaf of regular “functions” ϕ(x0 , x1 , x2 ) on X that satisfy ϕ(λx0 , λx1 , λx2 ) = λn ϕ(x0 , x1 , x2 ) under rescaling of the homogeneous coordinates. An element in the ﬁber of E over a point P is then just given by a number in k that rescales with λn . So E is a line bundle. We will usually denote it by O (n). For n = 0 we obviously just get the ordinary structure sheaf O . The spaces H 0 (X, O (n)) of sections are easily written down: H 0 (X, O (n)) = S(X)(n) 0 for n ≥ 0, for n < 0.

In particular, their dimensions (usually denoted h0 (X, O (n))) are just the values hX (n) of the Hilbert function. So the Hilbert function can be thought of as the dimension of the space of global sections of a line bundle O (n). In our study of Hilbert polynomials we have seen that Hilbert functions and polynomials are usually computed using exact sequences (of graded vector spaces). In the same way, the spaces of sections H 0 (X, E) are usually computed using exact sequences of vector bundles. For example, if Y is a smooth subvariety of a smooth variety X, then there is an exact sequence of vector bundles on X 0 → TY → TX |Y → NY /X → 0, where NY /X is the normal bundle of Y in X — it is by deﬁnition simply the vector bundle whose ﬁbers are the normal spaces TX,P /TY,P . The sequence is then exact by deﬁnition (i.e. it is exact locally at every ﬁber). This does not mean however that the spaces of global sections necessarily form an exact sequence 0 → H 0 (Y, TY ) → H 0 (Y, TX |Y ) → H 0 (Y, NY /X ) → 0. In fact one can show that one always gets an exact sequence 0 → H 0 (Y, TY ) → H 0 (Y, TX |Y ) → H 0 (Y, NY /X ), but exactness need not be preserved in the last term: a surjective map E → F of vector bundles need not give rise to a surjective map H 0 (X, E) → H 0 (X, F) of global sections. An example is easily found: consider the morphism of vector bundles

2 2 O ⊕ O → O (2), (ϕ1 , ϕ2 ) → x0 ϕ1 + x1 ϕ2

on P1 . This is obviously surjective in every ﬁber — for every point P = (x0 : x1 ) ∈ P1 at least one of the coordinates is non-zero, so by picking suitable ϕ1 (P) and ϕ2 (P) we can get 2 2 any number for x0 ϕ1 (P) + x1 ϕ2 (P). But the corresponding morphism of global sections H 0 (P1 , O ⊕ O ) → H 0 (P1 , O (2)) cannot be surjective simply for dimensional reasons, as the dimensions of these vector spaces are 2 and 3, respectively. It turns out however that there are canonically deﬁned cohomology groups H i (X, E) for i > 0 and every vector bundle E (in fact even for more general sheaves) such that every exact sequence 0 → E1 → E2 → E3 → 0 of the bundles gives rise to an exact sequence of cohomology groups

0 → H 0 (X, E1 ) → H 0 (X, E2 ) → H 0 (X, E3 ) → H 1 (X, E1 ) → H 1 (X, E2 ) → H 1 (X, E3 ) → H 2 (X, E1 ) → · · · .

116

Andreas Gathmann

So every such sequence of vector bundles gives rise to a relation between the (dimensions of the) cohomology groups: if we set hi (X, E) = dim H i (X, E) and χ(X, E) = ∑(−1)i hi (X, E)

i

then χ(X, E2 ) = χ(X, E1 ) + χ(X, E3 ). It can be shown that the sums in the deﬁnition of χ(X, E) are always ﬁnite. In fact, the higher cohomology groups vanish in many cases anyway (there are a lot of so-called “vanishing theorems”), so that the above long sequence between the cohomology groups is usually by far not as complicated as it seems to be here. The problem of computing these numbers hi (X, E) (or rather χ(X, E)) is solved by the Riemann-Roch theorem: expressed in simple terms this theorem states that χ(X, E) can always be computed using the intersection-theoretic data of the vector bundle (namely the Chern classes mentioned above in example 6.6.2). It is an explicit multilinear function in these Chern classes that is usually easily computable. In particular, χ(X, O (n)) turns out to be a polynomial in n — it is just the Hilbert polynomial of X. There is a vanishing theorem that implies hi (X, O (n)) = 0 for i > 0 and n 0, so we arrive at our old characterization of the Hilbert polynomial as the polynomial that agrees with the Hilbert function for large n. In particular, we see that the arithmetic genus of a variety (see example 6.1.10) is just (−1)dim X (χ(X, O ) − 1), which obviously does not depend on the embedding of X in projective space. The easiest case of the Riemann-Roch theorem is that of line bundles on smooth curves. If E is a line bundle on a curve X (e.g. a bundle of the form O (n) if X is projective), we can associate to it: (i) intersection-theoretic data: given a (rational) section of E, how many zeros and poles does this section have? This number is called the degree of E. For example, the degree of O (n) on a plane curve of degree d is d · n, as every global section of O (n) (i.e. a polynomial of degree n) vanishes on X at d · n points by B´ zout’s e theorem. (ii) cohomological data: how many sections of E are there? Ideally we would like to know h0 (X, E), but the Riemann-Roch theorem will only give us χ(X, E) = h0 (X, E) − h1 (X, E). The Riemann-Roch theorem then states that χ(X, E) = deg E + 1 − g, where g is the genus of the curve X. For example, for X = P1 we get χ(X, O (n)) = n+1−0, which is indeed the Hilbert polynomial of P1 . Example 6.6.4. Moduli spaces. We have now met several instances already where it proved useful to make the set of all geometric objects of a certain type into a scheme (or maybe a variety): (i) The Grassmannian G(1, n) is a variety that can be thought of as the set of all lines in Pn . (ii) The afﬁne space AN = k[x0 , . . . , xn ](d) (with N = n+d ) can be thought of as the d set of all degree-d hypersurfaces in Pn . (iii) The vector bundle E of example 6.6.2 can be thought of as the set of pairs (L, f ), where L is a line in P4 and f is a quintic polynomial on L.

6.

First applications of scheme theory

117

Schemes whose points describe geometric objects in this sense are called moduli spaces. So we say e.g. that G(1, n) is the moduli space of lines in Pn . There are many other moduli spaces one may want to consider. The most prominent ones are: (i) moduli spaces of curves (with a ﬁxed given genus), (ii) moduli spaces of projective subschemes of Pn with a ﬁxed given Hilbert polynomial (the so-called Hilbert schemes), (iii) moduli spaces of vector bundles over a given variety, but you can try to give more or less every set of geometric objects a scheme structure. Such a scheme structure may or may not exist, and it may or may not behave nicely. Moduli spaces come into play when you want to consider families of geometric objects, e.g. families of varieties. For example, a family of lines in Pn over a base scheme B is simply a morphism f : B → G(1, n) to the moduli space of lines. This assigns to every point of B a line in Pn in a continuously varying way (as a morphism is given by continuous functions). For example, if the ground ﬁeld is C and you have a sequence of points Pi in B converging to a point P ∈ B, then we get a corresponding sequence of lines f (Pi ) in Pn that converges to f (P). We can thus talk about convergence, limits, or “small deformations” of the objects for which we have a moduli space. Deformations are often a powerful tool to make complicated objects into easier ones. For example, in example 0.1.3 we computed the genus of a plane curve by deforming it into a union of lines, for which the genus could be read off easily. Example 6.6.5. Classiﬁcation theory. Closely related to the study of moduli spaces is the desire to “classify all algebraic varieties” (or other objects occurring in algebraic geometry). For smooth curves the result is quite easy to state: (i) Every smooth curve has a genus (see e.g. example 0.1.1 and 6.1.10) that is a non-negative integer. (ii) The moduli space of all smooth curves of a given genus g is an irreducible projective variety (with only mild singularities). Its dimension is 0 for g = 0, 1 for g = 1, and 3g − 3 for g > 1. So this result says that curves are characterized by one discrete invariant, namely its genus. Once the genus is ﬁxed, every curve of this genus can be deformed continuously into any other curve of the same genus. In contrast, curves cannot be deformed into each other if their genera are different. For higher-dimensional varieties the situation is a lot more complicated. As above, one ﬁrst looks for discrete invariants, i.e. “integers that can be associated to the variety in a natural way” and that are invariant under deformation. In a second step, one can then ask for the dimension (and other properties) of the moduli space of varieties with the given ﬁxed discrete invariants. Examples of discrete invariants are: (i) the dimension (of course), (ii) cohomological or intersection-theoretic properties of the tangent bundle and related bundles, e.g. hi (X, TX ), hi (X, ΩX ), the Chern classes of the tangent bundle, ... (iii) the genus (−1)dim X (χ(X, O ) − 1), (iv) various intersection-theoretic data, e.g. the collection of numbers and the multilinear functions describing intersection products as in example 6.6.1. For surfaces, this classiﬁcation problem is solved, but the result is quite complicated. For higher-dimensional varieties, the problem is still largely unsolved. 6.7. Exercises.

118

Andreas Gathmann

Exercise 6.7.1. Let X be a collection of four distinct points in some Pn . What are the possible Hilbert functions hX ? Exercise 6.7.2. Compute the Hilbert function and the Hilbert polynomial of the “twisted cubic curve” C = {(s3 : s2t : st 2 : t 3 ) ; (s : t) ∈ P1 } ⊂ P3 . Exercise 6.7.3. Let X ⊂ Pn be a projective scheme with Hilbert polynomial χ. As in example 6.1.10 deﬁne the arithmetic genus of X to be g(X) = (−1)dim X · (χ(0) − 1). (i) Show that g(Pn ) = 0. (ii) If X is a hypersurface of degree d in Pn , show that g(X) = d−1 . In particular, if n 1 C ⊂ P2 is a plane curve of degree d, then g(C) = 2 (d − 1)(d − 2) (compare this to example 0.1.3). (iii) Compute the arithmetic genus of the union of the three coordinate axes Z(x1 x2 , x1 x3 , x2 x3 ) ⊂ P3 . Exercise 6.7.4. For N = (n + 1)(m + 1) − 1 let X ⊂ PN be the image of the Segre embedding Pn × Pm → PN . Show that the degree of X is n+m . n Exercise 6.7.5. Let X be an ellipse in the real plane R2 , and let P be a given point on X. Using only a ruler with no markings, construct the tangent line to X at P. (In other words: start with a piece of paper which has only the ellipse X and the marked point P ∈ X on it. The only thing you are now allowed to do is to repeatedly draw straight lines through two points that have already been constructed (the point P, intersection points of previously drawn curves, or arbitrarily chosen points). No measuring of lengths or angles is permitted. Give an algorithm that ﬁnally allows you to draw the tangent line to X at P this way.) Exercise 6.7.6. Let C ⊂ Pn be an irreducible curve of degree d. Show that C is contained in a linear subspace of Pn of dimension d. Exercise 6.7.7. Let X and Y be subvarieties of Pn that lie in disjoint linear subspaces of k Pn . Recall from exercises 3.5.7 and 4.6.1 that the join J(X,Y ) ⊂ Pn of X and Y is deﬁned k k to be the union of all lines PQ with P ∈ X and Q ∈ Y . L (i) Show that S(J(X,Y ))(d) ∼ S(X)(i) ⊗k S(X)( j) . =

i+ j=d

(ii) Show that deg J(X,Y ) = deg X · degY . Exercise 6.7.8. Let C1 = { f1 = 0} and C2 = { f2 = 0} be afﬁne curves in A2 , and let k P ∈ C1 ∩ C2 be a point. Show that the intersection multiplicity of C1 and C2 at P (i.e. the length of the component at P of the intersection scheme C1 ∩C2 ) is equal to the dimension of the vector space OA2 ,P /( f1 , f2 ) over k. Exercise 6.7.9. Let C1 ,C2 ⊂ P2 be distinct smooth cubic curves, and assume that C1 and C2 intersect in 9 (distinct) points P1 , . . . , P9 . Prove that every cubic curve passing through P1 , . . . , P8 also has to pass through P9 . Can you ﬁnd a stronger version of this statement that applies in the case that the intersection multiplicities in C1 ∩C2 are not all equal to 1 ? Exercise 6.7.10. Let C be a smooth cubic curve of the form C = {(x : y : z) ; y2 z = x3 + axz2 + bz3 } ⊂ P2 k for some given a, b ∈ k. (It can be shown that every cubic can be brought into this form by a change of coordinates.) Pick the point P0 = (0 : 1 : 0) as the zero element for the group structure on C. For given points P1 = (x1 : y1 : 1) and P2 = (x2 : y2 : 1) compute explicitly the coordinates of the inverse P1 and of the sum P1 ⊕ P2 . Conclude that the group structure on C is well-deﬁned even if k is not necessarily algebraically closed.

6.

First applications of scheme theory

119

Exercise 6.7.11. Let C ⊂ P2 be a smooth cubic curve, and let P ∈ C be an inﬂection point C of C. Show that there are exactly 4 tangents of C that pass through P. Conclude that there are exactly 4 divisor classes D in PicC such that 2D = 0. Exercise 6.7.12. Let C ⊂ P2 be a smooth cubic curve, and let P, Q ∈ C be two points. Show that there is an isomorphism f : C → C with f (P) = Q. Is this isomorphism unique? Exercise 6.7.13. Check that the cubic curve C ⊂ P2 deﬁned by a lattice Λ ⊂ C as in C proposition 6.5.7 is smooth. Exercise 6.7.14. Using the complex analysis methods of section 6.5, reprove the statement of proposition 6.3.13 that there is no rational function ϕ on a smooth plane complex cubic curve C with divisor (ϕ) = P − Q if P and Q are two distinct points on C. Exercise 6.7.15. Let C ⊂ P2 be a smooth cubic curve arising from a lattice Λ ⊂ C. Show C 0 that the group structure of PicC is isomorphic to the natural group structure of C/Λ. Exercise 6.7.16. Let Λ ⊂ C be a lattice. Given a point z ∈ C/Λ and any n ∈ Z, it is obviously very easy to ﬁnd a point w ∈ C/Λ such that n · w = z (in the group structure of C/Λ). Isn’t this a contradiction to the idea of example 6.4.8?

120

Andreas Gathmann

7. M ORE ABOUT SHEAVES

We present a detailed study of sheaves on a scheme X, in particular sheaves of OX modules. For any presheaf F on X there is an associated sheaf F that describes “the same objects as F but with the conditions on the sections made local”. This allows us to deﬁne sheaves by constructions that would otherwise only yield presheaves. We can thus construct e.g. direct sums of sheaves, tensor products, kernels and cokernels of morphisms of sheaves, as well as push-forwards and pull-backs along morphisms of schemes. A sheaf of OX -modules is called quasi-coherent if it is induced by an R-module on every afﬁne open subset U = Spec R of X. Almost all sheaves that we will consider are of this form. This reduces local computations regarding these sheaves to computations in commutative algebra. A quasi-coherent sheaf on X is called locally free of rank r if it is locally iso⊕r morphic to OX . Locally free sheaves are the most well-behaved sheaves; they correspond to vector bundles in topology. Any construction and theorem valid for vector spaces can be carried over to the category of locally free sheaves. Locally free sheaves of rank 1 are called line bundles. For any morphism f : X → Y we deﬁne the sheaf of relative differential forms ΩX/Y on X relative Y . The most important case is when Y is a point, in which case we arrive at the sheaf ΩX of differential forms on X. It is locally free of rank dim X if and only if X is smooth. In this case, its top alternating power Λdim X ΩX is a line bundle ωX called the canonical bundle. On a smooth projective curve it has degree 2g − 2, where g is the genus of the curve. On every smooth curve X the line bundles form a group which is isomorphic to the Picard group Pic X of divisor classes. A line bundle together with a collection of sections that do not vanish simultaneously at any point determines a morphism to projective space. If f : X → Y is a morphism of smooth projective curves, the Riemann-Hurwitz formula states that the canonical bundles of X and Y are related by ωX = f ∗ ωY ⊗ OX (R), where R is the ramiﬁcation divisor. For any smooth projective curve X of genus g and any divisor D the Riemann-Roch theorem states that h0 (D) − h0 (KX − D) = deg D + 1 − g, where h0 (D) denotes the dimension of the space of global sections of the line bundle O (D) associated to D.

7.1. Sheaves and sheaﬁﬁcation. The ﬁrst thing we have to do to discuss the more advanced topics mentioned in section 6.6 is to get a more detailed understanding of sheaves. Recall from section 2.2 that we deﬁned a sheaf to be a structure on a topological space X that describes “function-like” objects that can be patched together from local data. Let us ﬁrst consider an informal example of a sheaf that is not just the sheaf of regular functions on a scheme. Example 7.1.1. Let X be a smooth complex curve. For any open subset U ⊂ X, we have seen that the ring of regular functions OX (U) on U can be thought of as the ring of complexvalued functions ϕ : U → C, P → ϕ(P) “varying nicely” (i.e. as a rational function) with P. Now consider the “tangent sheaf” TX , i.e. the sheaf “deﬁned” by TX (U) = {ϕ = (ϕ(P))P∈U ; ϕ(P) ∈ TX,P “varying nicely with P”} (of course we will have to make precise what “varying nicely” means). In other words, a section ϕ ∈ TX (U) is just given by specifying a tangent vector at every point in U. As an example, here is a picture of a section of TP1 (P1 ):

7.

More about sheaves

121

TX,P P φ(P)

As the tangent spaces TX,P are all one-dimensional complex vector spaces, ϕ(P) can again be thought of as being speciﬁed by a single complex number, just as for the structure sheaf OX . The important difference (that is already visible from the deﬁnition above) is that these one-dimensional vector spaces vary with P and thus have no canonical identiﬁcation with the complex numbers. For example, it does not make sense to talk about “the tangent vector 1” at a point P. Consequently, there is no analogue of “constant functions” for sections of the tangent sheaf. In fact, we will see in lemma 7.4.15 that every global section of TP1 has two zeros, so there is really no analogue of constant functions. (In the picture above, the north pole of the sphere is a point where the section of TP1 would be ill-deﬁned if we do not choose a section in which the lengths of the tangent vectors approach zero towards the north pole.) Hence we have seen that the tangent sheaf of P1 is a sheaf that is not isomorphic to the structure sheaf OP1 although its sections are given locally by “one complex number varying nicely”. (We should mention that the above property of P1 is purely topological: there is not even a continuous nowhere-zero tangent ﬁeld on the unit ball in R3 . This is usually called the “hairy ball theorem” and stated as saying that “you cannot comb a hedgehog (i.e. a ball) without a bald spot”.) Let us now get more rigorous. Recall that a presheaf of rings F on a topological space X was deﬁned to be given by the data: • for every open set U ⊂ X a ring F (U), • for every inclusion U ⊂ V of open sets in X a ring homomorphism ρV,U : F (V ) → F (U) called the restriction map, such that / • F (0) = 0, • ρU,U is the identity map for all U, • for any inclusion U ⊂ V ⊂ W of open sets in X we have ρV,U ◦ ρW,V = ρW,U . The elements of F (U) are then called the sections of F over U, and the restriction maps ρV,U are written as f → f |U . The space of global sections F (X) is often denoted Γ(F ). A presheaf F of rings is called a sheaf of rings if it satisﬁes the following glueing property: if U ⊂ X is an open set, {Ui } an open cover of U and fi ∈ F (Ui ) sections for all i such that fi |Ui ∩U j = f j |Ui ∩U j for all i, j, then there is a unique f ∈ F (U) such that f |Ui = fi for all i. In other words, sections of a sheaf can be patched from compatible local data. The same deﬁnition applies equally to categories other than rings, e.g. we can deﬁne sheaves of Abelian groups, k-algebras, and so on. For a ringed space (X, OX ), e.g. a scheme, we can also deﬁne sheaves of OX -modules in the obvious way: every F (U) is required to be an OX (U)-module, and these module structures have to be compatible with

122

Andreas Gathmann

the restriction maps in the obvious sense. For example, the tangent sheaf of example 7.1.1 on a curve X is a sheaf of OX -modules: “sections of the tangent sheaf can be multiplied with regular functions”. Example 7.1.2. Let X ⊂ PN be a projective variety over an algebraically closed ﬁeld k, and L let S(X) = S = d≥0 S(d) be its homogeneous coordinate ring. For any integer n, let K(n) be the n-th graded piece of the localization of S at the non-zero homogeneous elements, i.e. f K(n) = ; f ∈ S(d+n) , g ∈ S(d) for some d ≥ 0 and g = 0 . g Now for any P ∈ X and open set U ⊂ X we set

OX (n)P =

f ∈ K(n) ; g(P) = 0 g

and OX (n)(U) =

\

P∈U

OX (n)P .

For n = 0 this is precisely the deﬁnition of the structure sheaf, so OX (0) = OX . In general, OX (n) is a sheaf of OX -modules whose sections can be thought of as “functions” of degree n in the homogeneous coordinates of X. For example: (i) Every homogeneous polynomial of degree n deﬁnes a global section of OX (n). (ii) There are no global sections of OX (n) for n < 0. (iii) In P1 with homogeneous coordinates x0 , x1 , we have 1 ∈ OP1 (−1)(U) x0 for U = {(x0 : x1 ) ; x0 = 0}. Note that on the distinguished open subset Xxi (where xi are the coordinates of PN ) the sheaf OX (n) is isomorphic to the structure sheaf OX : for every open subset U ⊂ Xxi the maps ϕ OX (U) → OX (n)(U), ϕ → ϕ · xin and OX (n)(U) → OX (U), ϕ → n xi give an isomorphism, hence OX (n)|Xxi ∼ OX |Xxi . So OX (n) is locally isomorphic to the = structure sheaf, but not globally. (This is the same situation as for the tangent sheaf of a smooth curve in example 7.1.1.) The sheaves O (n) on a projective variety (or more generally on a projective scheme) are called the twisting sheaves. They are probably the most important sheaves after the structure sheaf. If we want to deal with more general sheaves, we certainly need to set up a suitable category, i.e. we have to deﬁne morphisms of sheaves, kernels, cokernels, and so on. Let us start with some simple deﬁnitions. Deﬁnition 7.1.3. Let X be a topological space. A morphism f : F1 → F2 of presheaves of abelian groups (or rings, sheaves of OX -modules etc.) on X is a collection of group homomorphisms (resp. ring homomorphisms, OX (U)-module homomorphisms etc.) fU : F1 (U) → F2 (U) for every open subset U ⊂ X that commute with the restriction maps, i.e. the diagram

F1 (U)

F2 (U) is required to be commutative.

fU

ρU,V

/ F1 (V ) / F2 (V )

fV

ρU,V

7.

More about sheaves

123

Example 7.1.4. If X ⊂ PN is a projective variety and f ∈ k[x0 , . . . , xN ] is a homogeneous polynomial of degree d, we get morphisms of sheaves of OX -modules

OX (n) → OX (n + d), ϕ → f · ϕ

for all n. Deﬁnition 7.1.5. If f : X → Y is a morphism of topological spaces and F is a sheaf on X, then we deﬁne the push-forward f∗ F of F to be the sheaf on Y given by f∗ F (U) = F ( f −1 (U)) for all open subsets U ⊂ Y . Example 7.1.6. By deﬁnition, a morphism f : X → Y of ringed spaces comes equipped with a morphism of sheaves OY → f∗ OX . This is exactly given by the data of the pull-back morphisms OY (U) → OX ( f −1 (U)) for all open subsets U ⊂ Y (see deﬁnition 5.2.1). Deﬁnition 7.1.7. Let f : F1 → F2 be a morphism of sheaves of e.g. Abelian groups on a topological space X. We deﬁne the kernel ker f of f by setting (ker f )(U) = ker( fU : F1 (U) → F2 (U)). We claim that ker f is a sheaf on X. In fact, it is easy to see that ker f with the obvious restriction maps is a presheaf. Now let {Ui } be an open cover of an open subset U ⊂ X, and assume we are given ϕi ∈ ker(F1 (Ui ) → F2 (Ui )) that agree on the overlaps Ui ∩U j . In particular, the ϕi are then in F1 (Ui ), so we get a unique ϕ ∈ F1 (U) with ϕ|Ui = ϕi as F1 is a sheaf. Moreover, f (ϕi ) = 0, so ( f (ϕ))|Ui = 0 by deﬁnition 7.1.3. As F2 is a sheaf, it follows that f (ϕ) = 0, so ϕ ∈ ker f . What the above argument boils down to is simply that the property of being in the kernel, i.e. of being mapped to zero under a morphism, is a local property — a function is zero if it is zero on every subset of an open cover. So the kernel is again a sheaf. Remark 7.1.8. Now consider the dual case to deﬁnition 7.1.7, namely cokernels. Again let f : F1 → F2 be a morphism of sheaves of e.g. Abelian groups on a topological space X. As above we deﬁne a presheaf coker f by setting (coker f )(U) = coker( fU : F1 (U) → F2 (U)) = F2 (U)/ im fU . Note however that coker f is not a sheaf. To see this, consider the following example. Let X = A1 \{0}, Y = A2 \{0}, and let i : X → Y be the inclusion morphism (x1 ) → (x1 , 0). Let i# : OY → i∗ OX be the induced morphisms of sheaves on Y of example 7.1.6, and consider the presheaf coker i# on Y . Look at the cover of Y by the afﬁne open subsets U1 = {x1 = 0} ⊂ Y and U2 = {x2 = 0} ⊂ Y . Then the maps 1 1 , x2 = OY (U1 ) → OX (U1 ∩ X) = k x1 , x1 x1 1 and k x1 , x2 , = OY (U2 ) → OX (U2 ∩ X) = 0 x2 k x1 , are surjective, hence (coker i# )(U1 ) = (coker i# )(U2 ) = 0. But on global sections the map k[x1 , x2 ] = OY (Y ) → OX (X) = k x1 , 1 x1

is not surjective, hence (coker i# )(Y ) = 0. This shows that coker i# cannot be a sheaf — the zero section on the open cover {U1 ,U2 } has no unique extension to a global section on Y. What the above argument boils down to is simply that being in the cokernel of a morphism, i.e. of being a quotient in F2 (U)/ im fU , is not a local property — it is a question about ﬁnding a global section of F2 on U that cannot be answered locally.

124

Andreas Gathmann

Example 7.1.9. Here is another example showing that quite natural constructions involving sheaves often lead to only presheaves because the constructions are not local. Let X ⊂ PN be a projective variety. Consider the tensor product presheaf of the sheaves OX (1) and OX (−1), deﬁned by (OX (1) ⊗ OX (−1))(U) = OX (1)(U) ⊗OX (U) OX (−1)(U). As OX (1) describes “functions” of degree 1 and OX (−1) “functions” of degree −1, we expect products of them to be true functions of pure degree 0 in the homogeneous coordinates of X. In other words, the tensor product of OX (1) with OX (−1) should just be the structure sheaf OX . However, OX (1) ⊗ OX (−1) is not even a sheaf: consider the case X = P1 and the open subsets U0 = {x0 = 0} and U1 = {x1 = 0}. On these open subsets we have the sections 1 x0 ⊗ ∈ (OX (1) ⊗ OX (−1))(U0 ) x0 1 and x1 ⊗ ∈ (OX (1) ⊗ OX (−1))(U1 ). x1 Obviously, both these local sections are the constant function 1, so in particular they agree on the overlap U0 ∩U1 . But there is no global section in OX (1)(X) ⊗OX (X) OX (−1)(X) that corresponds to the constant function 1, as OX (−1) has no non-zero global sections at all. The way out of this trouble is called sheaﬁﬁcation. This means that for any presheaf

F there is an associated sheaf F that is “very close” to F and that should usually be

the object that one wants. Intuitively speaking, if the sections of a presheaf are thought of as function-like objects satisfying some conditions, then the associated sheaf describes the same objects with the conditions on them made local. In particular, if we look at F locally, i.e. at the stalks, then we should not change anything; it is just the global structure that changes. We have done this construction quite often already without explicitly saying so, e.g. in the construction of the structure sheaf of schemes in deﬁnition 5.1.11. Here is the general construction: Deﬁnition 7.1.10. Let F be a presheaf on a topological space X. The sheaﬁﬁcation of F , or the sheaf associated to the presheaf F , is deﬁned to be the sheaf F such that

F (U) := {ϕ = (ϕP )P∈U with ϕP ∈ FP for all P ∈ U

such that for every P ∈ U there is a neighborhood V in U and a section ϕ ∈ F (V ) with ϕQ = ϕQ ∈ FQ for all Q ∈ V .} (For the notion of the stalk FP of a presheaf F at a point P ∈ X see deﬁnition 2.2.7.) It is obvious that this deﬁnes a sheaf. Example 7.1.11. Let X ⊂ AN be an afﬁne variety. Let OX be the presheaf given by

OX (U) = ϕ : U → k ; there are f , g ∈ k[x1 , . . . , xN ] with g(P) = 0

and ϕ(P) =

f (P) g(P)

for all P ∈ U

for all open subsets U ⊂ X, i.e. the “presheaf of functions that are (globally) quotients of polynomials”. Then the structure sheaf OX is the sheaﬁﬁcation of OX , i.e. the sheaf of functions that are locally quotients of polynomials. We have seen in example 2.1.7 that in general OX differs from OX , i.e. it is in general not a sheaf. Example 7.1.12. If X is a topological space and F the presheaf of constant real-valued functions on X, then the sheaﬁﬁcation of F is the sheaf of locally constant functions on X (see also remark 2.2.5). The sheaﬁﬁcation has the following nice and expected properties:

7.

More about sheaves

125

Lemma 7.1.13. Let F be a presheaf on a topological space X, and let F be its sheaﬁﬁcation. (i) The stalks FP and FP agree at every point P ∈ X. (ii) If F is a sheaf, then F = F . Proof. (i): The isomorphism between the stalks is given by the following construction: • An element of FP is by deﬁnition represented by (U, ϕ), where U is an open neighborhood of P and ϕ = (ϕQ )Q∈U is a section of F over U. To this we can associate the element ϕP ∈ FP . • An element of FP is by deﬁnition represented by (U, ϕ), where ϕ ∈ F (U). To this we can associate the element (ϕQ )Q∈U in F (U), which in turn deﬁnes an element of FP . (ii): Note that there is always a morphism of presheaves F → F given by F (U) → F (U), ϕ → (ϕP )P∈U . Now assume that F is a sheaf; we will construct an inverse morphism F → F . Let U ⊂ X be an open subset and ϕ = (ϕP )P∈U ∈ F (U) a section of F. For every P ∈ U the germ ϕP ∈ FP is represented by some (V, ϕ) with ϕ ∈ F (V ). As P varies over U, we are thus getting sections of F on an open cover of U that agree on the overlaps. As F is a sheaf, we can glue these sections together to give a global section in F (U). Using sheaﬁﬁcation, we can now deﬁne all the “natural” constructions that we would expect to be possible: Deﬁnition 7.1.14. Let f : F1 → F2 be a morphism of sheaves of e.g. Abelian groups on a topological space X. (i) The cokernel coker f of f is deﬁned to be the sheaf associated to the presheaf coker f . (ii) The morphism f is called injective if ker f = 0. It is called surjective if coker f = 0. (iii) If the morphism f is injective, its cokernel is also denoted F2 /F1 and called the quotient of F2 by F1 . (iv) As usual, a sequence of sheaves and morphisms · · · → Fi−1 → Fi → Fi+1 → · · · is called exact if ker(Fi → Fi+1 ) = im(Fi−1 → Fi ) for all i. Remark 7.1.15. Let us rephrase again the results of deﬁnition 7.1.7 and remark 7.1.8 in this new language: (i) A morphism f : F1 → F2 of sheaves is injective if and only if the maps fU : F1 (U) → F2 (U) are injective for all U. (ii) If a morphism f : F1 → F2 of sheaves is surjective, this does not imply that all maps fU : F1 (U) → F2 (U) are surjective. (The converse of this is obviously true however: if all maps fU : F1 (U) → F2 (U) are surjective, then coker f = 0, so coker f = 0.) This very important fact is the basis of the theory of cohomology, see chapter 8. Example 7.1.16. Let X = P1 with homogeneous coordinates x0 , x1 . Consider the mork phism of sheaves f : OX (−1) → OX given by the linear polynomial x0 (see example 7.1.4). We claim that f is injective. In fact, every section of OX (−1) over an open subset of X g(x ,x ) has the form h(x0 ,x1 ) for some homogeneous polynomials g, h with deg g − deg h = −1. But 0 1 gx0 g f ( h ) = h is zero on an open subset of X if and only if g = 0 (note that we are not asking

126

Andreas Gathmann

for zeros of gx0 , but we are asking whether this function vanishes on a whole open subset). h As this means that g itself is zero, we see that the kernel of f is trivial, i.e. f is injective. h We have seen already in example 7.1.2 that f is in fact an isomorphism when restricted to U = X\{P} where P := (0 : 1). In particular, f is surjective when restricted to U. However, f is not surjective on X (otherwise it would be an isomorphism, which is not true as we already know). Let us determine its cokernel. First we have to compute the cokernel presheaf coker f . Consider an open subset U ⊂ X. By the above argument, (coker f )(U) = 0 if P ∈ U. So assume that P ∈ U. Then we / have an exact sequence of OX (U)-modules 0 →

OX (−1)(U) → OX (U) →

g h

k

→ 0

→

ϕ=

gx0 h

g h

→ ϕ(P)

as the functions in the image of OX (−1)(U) → OX are precisely those that vanish on P. So we have found that k if P ∈ U, (coker f )(U) = 0 if P ∈ U. / It is easily veriﬁed that coker f is in fact a sheaf. It can be thought of as the ground ﬁeld k “concentrated at the point P”. For this reason it is often called a skyscraper sheaf and denoted kP . Summarizing, we have found the exact sequence of sheaves of OX -modules

0 0 → OX (−1) → OX → kP → 0.

·x

Example 7.1.17. Let F1 , F2 be two sheaves of OX -modules on a ringed space X. Then we can deﬁne the direct sum, the tensor product, and the dual sheaf in the obvious way: (i) The direct sum F1 ⊕ F2 is the sheaf of OX -modules deﬁned by (F1 ⊕ F2 )(U) = F1 (U) ⊕ F2 (U). (It is easy to see that this is a sheaf already, so that we do not need sheaﬁﬁcation.) (ii) The tensor product F1 ⊗ F2 is the sheaf of OX -modules associated to the presheaf U → F1 (U) ⊗OX (U) F2 (U). (iii) The dual F1∨ of F1 is the sheaf of OX -modules associated to the presheaf U → F1 (U)∨ = HomOX (U) (F1 (U), OX (U)). Example 7.1.18. Similarly to example 7.1.16 consider the morphism f : OX (−2) → OX of sheaves on X = P1 given by multiplication with x0 x1 (instead of with x0 ). The only k difference to the above example is that the function x0 x1 vanishes at two points P0 = (0 : 1), P1 = (1 : 0). So this time we get an exact sequence of sheaves

0 0 → OX (−2) →1 OX → kP0 ⊕ kP1 → 0,

·x x

where the last morphism is evaluation at the points P0 and P1 . The important difference is that this time the cokernel presheaf is not equal to the cokernel sheaf: if we consider our exact sequence on global sections, we get 0 → Γ(OX (−2)) → Γ(OX ) → k ⊕ k, where Γ(OX (−2)) = 0, and Γ(OX ) are just the constant functions. But the last morphism is evaluation at P and Q, and constant functions must have the same value at P and Q. So the last map Γ(OX ) → k ⊕ k is not surjective, indicating that some sheaﬁﬁcation is going on. (In example 7.1.16 we only had to evaluate at one point, and the corresponding map was surjective.)

7.

More about sheaves

127

Example 7.1.19. On X = PN , we have OX (n)⊗ OX (m) = OX (n+m), with the isomorphism given on sections by f1 f2 f1 f2 ⊗ → . g1 g2 g1 g2 Similarly, we have OX (n)∨ = OX (−n), as the OX (U)-linear homomorphisms from OX (n) to OX are precisely given by multiplication with sections of OX (−n). 7.2. Quasi-coherent sheaves. It turns out that sheaves of modules are still too general objects for many applications — usually one wants to restrict to a smaller class of sheaves. Recall that any ring R determines an afﬁne scheme X = Spec R together with its structure ˜ sheaf OX . Hence one would expect that any R-module M determines a sheaf M of OX modules on X. This is indeed the case, and almost any sheaf of OX -modules appearing in practice is of this form. For computations, this means that statements about this sheaf ˜ M on X are ﬁnally reduced to statements about the R-module M. But it does not follow from the deﬁnitions that a sheaf of OX -modules has to be induced by some R-module in this way (see example 7.2.3), so we will say that it is quasi-coherent if it does, and in most cases restrict our attention to these quasi-coherent sheaves. If X is a general scheme, we accordingly require that it has an open cover by afﬁne schemes Spec Ri over which the sheaf is induced by an Ri -module for all i. ˜ Let us start by showing how an R-module M determines a sheaf of modules M on X = Spec R. This is essentially the same construction as for the structure sheaf in deﬁnition 5.1.11. Deﬁnition 7.2.1. Let R be a ring, X = Spec R, and let M be an R-module. We deﬁne a ˜ sheaf of OX -modules M on X by setting ˜ M(U) := {ϕ = (ϕp )p∈U with ϕp ∈ Mp for all p ∈ U such that “ϕ is locally of the form

m r

for m ∈ M, r ∈ R”}

= {ϕ = (ϕp )p∈U with ϕp ∈ Mp for all p ∈ U such that for every p ∈ U there is a neighborhood V in U and m ∈ M, r ∈ R with r ∈ q and ϕq = /

m r

∈ Mq for all q ∈ V }.

˜ It is clear from the local nature of the deﬁnition that M is a sheaf. The following proposition corresponds exactly to the statement of proposition 5.1.12 for structure sheaves. Its proof can be copied literally, replacing R by M at appropriate places. Proposition 7.2.2. Let R be a ring, X = Spec R, and let M be an R-module. ˜ (i) For every p ∈ X the stalk of M at p is Mp . ˜ ˜ (ii) For every f ∈ R we have M(X f ) = M f . In particular, M(X) = M. Example 7.2.3. The following example shows that not all sheaves of OX -modules on X = ˜ Spec R have to be of the form M for some R-module M. 1 , and let F be the sheaf associated to the presheaf Let X = Ak U→

OX (U) if 0 ∈ U, /

0 if 0 ∈ U.

with the obvious restriction maps. Then F is a sheaf of OX -modules. The stalk F0 is zero, whereas FP = OX,P for all other points P ∈ X. Note that F has no non-trivial global sections: if ϕ ∈ F (X) then we obviously must have ϕ0 = 0 ∈ F0 , which by deﬁnition of sheaﬁﬁcation means that ϕ is zero in some neighborhood of 0. But as X is irreducible, ϕ must then be the zero function. Hence it follows

128

Andreas Gathmann

˜ that F (X) = 0. So if F was of the form M for some R-module M, it would follow from proposition 7.2.2 (ii) that M = 0, hence F would have to be the zero sheaf, which it obviously is not. Deﬁnition 7.2.4. Let X be a scheme, and let F be a sheaf of OX -modules. We say that F is quasi-coherent if for every afﬁne open subset U = Spec R ⊂ X the restricted sheaf F |U ˜ is of the form M for some R-module M. Remark 7.2.5. It can be shown that it is sufﬁcient to require the condition of the deﬁnition only for every open subset in an afﬁne open cover of X (see e.g. [H] proposition II.5.4). In other words, quasi-coherence is a local property. Example 7.2.6. On any scheme the structure sheaf is quasi-coherent. The sheaves OX (n) are quasi-coherent on any projective subscheme of PN as they are locally isomorphic to the structure sheaf. In the rest of this section we will show that essentially all operations that you can do with quasi-coherent sheaves yield again quasi-coherent sheaves. Therefore almost all sheaves that occur in practice are in fact quasi-coherent. Lemma 7.2.7. Let R be a ring and X = Spec R. (i) For any R-modules M, N there is a one-to-one correspondence ˜ ˜ {morphisms of sheaves M → N} ↔ {R-module homomorphisms M → N}. (ii) A sequence of R-modules 0 → M1 → M2 → M3 → 0 is exact if and only if the ˜ ˜ ˜ sequence of sheaves 0 → M1 → M2 → M3 → 0 is exact on X. ˜ ⊕ N = (M ⊕ N)˜. ˜ (iii) For any R-modules M, N we have M ˜ ˜ (iv) For any R-modules M, N we have M ⊗ N = (M ⊗ N)˜. ˜ ∨ = (M ∨ )˜. (v) For any R-module M we have (M) In particular, kernels, cokernels, direct sums, tensor products, and duals of quasi-coherent sheaves are again quasi-coherent on any scheme X. ˜ ˜ Proof. (i): Given a morphism M → N, taking global sections gives an R-module homomorphism M → N by proposition 7.2.2 (ii). Conversely, an R-module homomorphism M → N gives rise to morphisms between the stalks Mp → Np for all p, and therefore by deﬁnition ˜ ˜ determines a morphism M → N of sheaves. It is obvious that these two operations are inverse to each other. (ii): By exercise 7.8.2, exactness of a sequence of sheaves can be seen on the stalks. Hence by proposition 7.2.2 (i) the statement follows from the algebraic fact that the sequence 0 → M1 → M2 → M3 → 0 is exact if and only if 0 → (M1 )p → (M2 )p → (M3 )p → 0 is for all prime ideals p ∈ R. (iii), (iv), and (v) follow in the same way as (ii): the statement can be checked on the stalks, hence it follows from the corresponding algebraic fact about localizations of modules. Example 7.2.8. Let X = P1 and P = (0 : 1) ∈ X. The skyscraper sheaf kP of example 7.1.16 is quasi-coherent by lemma 7.2.7 as it is the cokernel of a morphism of quasicoherent sheaves. One can also check explicitly that kP is quasi-coherent: if U0 = {x0 = 0} = P1 \{P} and U1 = {x1 = 0} = Spec k[x0 ] ∼ A1 then kP |U0 = 0 (so it is the sheaf = associated to the zero module) and kP |U1 ∼ M where M = k is the k[x0 ]-module with the = ˜ module structure k[x0 ] × k → k ( f , λ) → f (0) · λ.

7.

More about sheaves

129

Proposition 7.2.9. Let f : X → Y be a morphism of schemes, and let F be a quasi-coherent sheaf on X. Assume moreover that every open subset of X can be covered by ﬁnitely many afﬁne open subsets (this should be thought of as a technical condition that is essentially always satisﬁed — it is e.g. certainly true for all subschemes of projective spaces). Then f∗ F is quasi-coherent on Y . Proof. Let us ﬁrst assume that X and Y are afﬁne, so X = Spec R, Y = Spec S, and F = ˜ M for some R-module M. Then it follows immediately from the deﬁnitions that f∗ F = (M as an S-module)˜, hence push-forwards of quasi-coherent sheaves are quasi-coherent if X and Y are afﬁne. In the general case, note that the statement is local on Y , so we can assume that Y is afﬁne. But it is not local on X, so we cannot assume that X is afﬁne. Instead, cover X by afﬁne opens Ui , and cover Ui ∩U j by afﬁne opens Ui, j,k . By our assumption, we can take these covers to be ﬁnite. Now the sheaf property for F says that for every open set V ⊂ Y the sequence 0 → F ( f −1 (V )) →

M

i

F ( f −1 (V ) ∩Ui ) →

M

i, j,k

F ( f −1 (V ) ∩Ui, j,k )

is exact, where the last map is given by (. . . , si , . . . ) → (. . . , si |Ui, j,k − s j |Ui, j,k , . . . ). This means that the sequence of sheaves on Y 0 → f∗ F →

M

i

f∗ (F |Ui ) →

M

i, j,k

f∗ (F |Ui, j,k )

is exact. But as we have shown the statement already for morphisms between afﬁne schemes and as ﬁnite direct sums of quasi-coherent sheaves are quasi-coherent, the last two terms in this sequence are quasi-coherent. Hence the kernel f∗ F is also quasi-coherent by lemma 7.2.7. Example 7.2.10. With this result we can now deﬁne (and motivate) what a closed embedding of schemes is. Note that for historical reasons closed embeddings are usually called closed immersions in algebraic geometry (in contrast to differential geometry, where an immersion is deﬁned to be a local embedding). We say that a morphism f : X → Y of schemes is a closed immersion if (i) f is a homeomorphism onto a closed subset of Y , and (ii) the induced morphism OY → f∗ OX is surjective. The kernel of the morphism OY → f∗ OX is then called the ideal sheaf IX/Y of the immersion. Let us motivate this deﬁnition. We certainly want condition (i) to hold on the level of topological spaces. But this is not enough — we have seen that even isomorphisms cannot be detected on the level of topological spaces (see example 2.3.8), so we need some conditions on the structure sheaves as well. We have seen in example 5.2.3 that a closed immersion should be a morphism that is locally of the form Spec R/I → Spec R for some ideal I ⊂ R. In fact, this is exactly what condition (ii) means: assume that we are in the afﬁne case, i.e. X = Spec S and Y = Spec R. As OY and f∗ OX are quasi-coherent (the former by deﬁnition and the latter by proposition 7.2.9), so is the kernel of OY → f∗ OX by lemma 7.2.7. So the exact sequence 0 → IX/Y → OY → f∗ OX → 0 comes from an exact sequence of R-modules 0→I→R→S→0

130

Andreas Gathmann

by lemma 7.2.7 (ii). In other words, I ⊂ R is an ideal of R, and S = R/I. So indeed the morphism f is of the form Spec R/I → Spec R and therefore corresponds to an inclusion morphism of a closed subset. Example 7.2.11. Having studied push-forwards of sheaves, we now want to consider pullbacks, i.e. the “dual” situation: given a morphism f : X → Y and a sheaf F on Y , we want to construct a “pull-back” sheaf f ∗ F on X. Note that this should be “more natural” than the push-forward, as sheaves describe “function-like” objects, and for functions pull-back is more natural than push-forward: given a “function” ϕ : Y → k, there is set-theoretically a well-deﬁned pull-back function ϕ ◦ f : X → k. In contrast, a function ϕ : X → k does not give rise to a function Y → k in a natural way. Let us ﬁrst consider the afﬁne case: assume that X = Spec R, and Y = Spec S, so that the morphism f corresponds to a ring homomorphism S → R. Assume moreover that the sheaf F on Y is quasi-coherent, so that it corresponds to an S-module M. Then M ⊗S R is a well-deﬁned R-module, and the corresponding sheaf on X should be the pull-back f ∗ F . Indeed, if e.g. M = S, i.e. F = OY , then M ⊗S R = S ⊗S R = R, so f ∗ F = OX : pull-backs of regular functions are just regular functions. This is our “local model” for the pull-back of sheaves. To show that this extends to the global case (and to sheaves that are not necessarily quasi-coherent), we need a different description though. So assume now that X, Y , and F are arbitrary. The ﬁrst thing to do is to deﬁne a sheaf of abelian groups on X from F . This is more complicated than for the push-forward constructed in deﬁnition 7.1.5, because f (U) need not be open if U is. We let f −1 F be the sheaf on X associated to the presheaf U → limV ⊃ f (U) F (V ), where the limit is taken over all open subsets V with f (U) ⊂ V ⊂ Y . This notion of limit means that an element in limV ⊃ f (U) F (V ) is given by a pair (V, ϕ) with V ⊃ f (U) and ϕ ∈ F (V ), and that two such pairs (V, ϕ) and (V , ϕ ) deﬁne the same element if and only if there is an open subset W with f (U) ⊂ W ⊂ V ∩V such that ϕ|W = ϕ |W . This is the best we can do to adapt deﬁnition 7.1.5 to the pull-back case. It is easily checked that this construction does what we want on the stalks: we have ( f −1 F )P = F f (P) for all P ∈ X. Note that f −1 F is obviously a sheaf of ( f −1 OY )-modules, but not a sheaf of OX modules. (This corresponds to the statement that in the afﬁne case considered above, M is an S-module, but not an R-module.) We have seen in our afﬁne case what we have to do: we have to take the tensor product over f −1 OY with OX (i.e. over S with R). In other words, we deﬁne the pull-back f ∗ F of F to be f ∗ F = f −1 F ⊗ f −1 OY OX , which is then obviously a sheaf of OX -modules. As this construction restricts to the one given above if X and Y are afﬁne and F quasi-coherent, it also follows that pull-backs of quasi-coherent sheaves are again quasi-coherent. It should be stressed that this complicated limit construction is only needed to prove the existence of f ∗ F in the general case. To compute the pull-back in practice, one will almost always restrict to afﬁne open subsets and then use the tensor product construction given above. Example 7.2.12. Here is a concrete example in which we can see again why the tensor product construction is necessary in the construction of the pull-back. Consider the morphism f : X = P1 → Y = P1 given by (s : t) → (x : y) = (s2 : t 2 ). We want to compute the pull-back sheaf f ∗ OY (1) on X. As we already know, local sections of OY (1) are of the form g(x,y) , with g and h homoh(x,y) geneous such that deg g − deg h = 1. Pulling this back just means inserting the equations

7.

More about sheaves

131

x = s2 and y = t 2 of f into this expression; so the sheaf f −1 OY (1) has local sections

g(s2 ,t 2 ) , h(s2 ,t 2 )

where now deg(g(s2 ,t 2 )) − deg(h(s2 ,t 2 )) = 2. But note that these sections do not even describe a sheaf of OX -modules: if we try to t multiply the section s2 with the function s (i.e. a section of OX ) on the open subset where s = 0, we get st, which is not of the form g(s2 ,t 2 ) . We have just seen the solution to this h(s ,t ) problem: consider the tensor product with OX . So sections of f ∗ OY (1) are of the form g(s2 ,t 2 ) g (s,t) ⊗ h(s2 ,t 2 ) h (s,t) with deg(g(s2 ,t 2 )) − deg(h(s2 ,t 2 )) = 2 and deg g − deg h = 0. It is easy to see that this describes precisely all expressions of the form g (s,t) with deg g − deg h = 2, so the result h (s,t) we get is f ∗ OY (1) = OX (2). In the same way one shows that f ∗ OY (n) = OX (dn) for all n ∈ Z and any morphism f : X → Y between projective schemes that is given by a collection of homogeneous polynomials of degree d. We have seen now that most sheaves occurring in practice are in fact quasi-coherent. So when we talk about sheaves from now on, we will usually think of quasi-coherent sheaves. This has the advantage that, on afﬁne open subsets, sheaves (that form a somewhat complicated object) are essentially replaced by modules, which are usually much easier to handle. 7.3. Locally free sheaves. We now come to the discussion of locally free sheaves, i.e. sheaves that are locally just a ﬁnite direct sum of copies of the structure sheaf. These are the most important and best-behaved sheaves one can imagine. Deﬁnition 7.3.1. Let X be a scheme. A sheaf of OX -modules F is called locally free of rank r if there is an open cover {Ui } of X such that F |Ui ∼ OUi for all i. Obviously, every = ⊕r locally free sheaf is also quasi-coherent. Remark 7.3.2. The geometric interpretation of locally free sheaves is that they correspond to “vector bundles” as known from topology — objects that associate to every point P of a space X a vector bundle. For example, the “tangent sheaf” of P1 in example 7.1.1 is such a vector bundle (of rank 1). Let us make this correspondence precise. A vector bundle of rank r on a scheme X over a ﬁeld k is a k-scheme F and a kmorphism π : F → X, together with the additional data consisting of an open covering {Ui } of X and isomorphisms ψi : π−1 (Ui ) → Ui × Ar over Ui , such that the automorphism k ψi ◦ ψ−1 of (Ui ∩ U j ) × Ar is linear in the coordinates of Ar for all i, j. In other words, j the morphism π : F → X looks locally like the projection morphism U × Ar → U for k sufﬁciently small open subsets U ⊂ X.

2 2

F

ψi

A Ui

r

π Ui X

132

Andreas Gathmann

We claim that there is a one-to-one correspondence {vector bundles π : F → X of rank r} ↔ {locally free sheaves F of rank r on X} given by the following constructions: (i) Let π : F → X be a vector bundle of rank r. Deﬁne a sheaf F on X by

F (U) = {k-morphisms s : U → F such that π ◦ s = idU }.

(This is called the sheaf of sections of F.) Note that this has a natural structure of a sheaf of OX -modules (over every point in X we can multiply a vector with a scalar — doing this on an open subset means that we can multiply a section in F (U) with a regular function in OX (U)). Locally, on an open subset U on which π is of the form U × Ar → U, we k obviously have

F (U) = {k-morphisms s : U → Ar }, k

so sections are just given by r independent functions. In other words, F |U is ⊕r isomorphic to OU . So F is locally free by deﬁnition. (ii) Conversely, let F be a locally free sheaf. Take an open cover {Ui } of X such that ⊕r there are isomorphisms ψi : F |Ui → OUi . Now consider the schemes Ui × Ar and k glue them together as follows: for all i, j we glue Ui × Ar and U j × Ar on the k k common open subset (Ui ∩U j ) × Ar along the isomorphism k (Ui ∩U j ) × Ar → (Ui ∩U j ) × Ar , k k (P, s) → (P, ψi ◦ ψ−1 ). j

Note that ψi ◦ ψ−1 is an isomorphism of sheaves of OX -modules and therefore j linear in the coordinates of Ar . k It is obvious that this gives exactly the inverse construction to (i). Remark 7.3.3. Let π : F → X be a vector bundle of rank r, and let P ∈ X be a point. We call π−1 (P) the ﬁber of F over P; it is an r-dimensional vector space. If F is the corresponding locally free sheaf, the ﬁber can be realized as i∗ F where i : P → X denotes the inclusion morphism (note that i∗ F is a sheaf on a one-point space, so its data consists only of one k-vector space (i∗ F )(P), which is precisely the ﬁber FP ). Lemma 7.3.4. Let X be a scheme. If F and G are locally free sheaves of OX -modules of rank r and s, respectively, then the following sheaves are also locally free: F ⊕ G (of rank r + s), F ⊗ G (of rank r · s), and F ∨ (of rank r). If f : X → Y is a morphism of schemes and F is a locally free sheaf on Y , then f ∗ F is a locally free sheaf on X of the same rank. (The push-forward of a locally free sheaf is in general not locally free.) Proof. The proofs all follow from the corresponding statements about vector spaces (or free modules over a ring): for example, if M and N are free R-modules of dimension r and s respectively, then M ⊕ N is a free R-module of dimension r + s. Applying this to an open ⊕r ⊕s ˜ ˜ afﬁne subset U = Spec R in X on which F and G are isomorphic to OU = M and OU = N gives the desired result. The statement about tensor products and duals follows in the same way. As for pull-backs, we have already seen that f ∗ OY = OX , so f ∗ F will be of the form O ⊕r (U) on the inverse image f −1 (U) ⊂ X of an open subset U ⊂ Y on which F is of the f −1

⊕r form OU .

Remark 7.3.5. Lemma 7.3.4 is an example of the general principle that any “canonical” construction or statement that works for vector spaces (or free modules) also works for vector bundles. Here is another example: recall that for any vector space V over k (or any free module) one can deﬁne the n-th symmetric product SnV and the n-th alternating

7.

More about sheaves

133

product ΛnV to be the vector space of formal totally symmetric (resp. antisymmetric) products v1 · · · · · vn ∈ SnV and v1 ∧ · · · ∧ vn ∈ ΛnV. If V has dimension r, then SnV and ΛnV have dimension More precisely, if {v1 , . . . , vr } is a basis of V , then

n+r−1 n

and

r n

, respectively.

{vi1 · · · · · vin ; i1 ≤ · · · ≤ in } is a basis of SnV , and {vi1 ∧ · · · ∧ vin ; i1 < · · · < in } is a basis of ΛnV . Using the same construction, we can get symmetric and alternating products Sn F and Λn F on X for every locally free sheaf F on X of rank r. They are locally free sheaves of ranks n+r−1 r and n , respectively. n Here is an example of a linear algebra lemma that translates directly into the language of locally free sheaves: Lemma 7.3.6. Let 0 → U → V → W → 0 be an exact sequence of vector spaces of dimensions a, a + b, and b, respectively. Then Λa+bV = ΛaU ⊗ ΛbW . Proof. Denote the two homomorphisms by i : U → V and p : V → W . Then there is a canonical isomorphism ΛaU ⊗ ΛbW → Λa+bV (u1 ∧ · · · ∧ ua ) ⊗ (w1 ∧ · · · ∧ wb ) → i(u1 ) ∧ · · · ∧ i(ua ) ∧ p−1 (w1 ) ∧ · · · ∧ p−1 (wb ). The key remark here is that the p−1 (wi ) are well-deﬁned up to an element of U by the exact sequence. But if the above expression is non-zero at all, the u1 , . . . , ua must form a basis of U, so if we plug in any element of U in the last b entries of the alternating product we will get zero. Therefore the ambiguity in the p−1 (wi ) does not matter and the above homomorphism is well-deﬁned. It is obviously not the zero map, and it is then an isomorphism for dimensional reasons (both sides are one-dimensional vector spaces). Corollary 7.3.7. Let 0 → F1 → F2 → F3 → 0 be an exact sequence of locally free sheaves of ranks a1 , a2 , a3 on a scheme X. Then Λa2 F2 = Λa1 F1 ⊗ Λa3 F3 . Proof. Immediately from lemma 7.3.6 using the above principle. 7.4. Differentials. We have seen in proposition 4.4.8 that (formal) differentiation of functions is useful to compute the tangent spaces at the (closed) points of a scheme X. We now want to introduce this language of differentials. The idea is that the various tangent spaces TP for P ∈ X should not just be independent vector spaces at every point, but rather come from a global object on X. For example, if X is smooth over C, so that it is a complex manifold, we know from complex geometry that X has a cotangent bundle whose ﬁber at a point P is just the cotangent space, i.e. the dual of the tangent space, at P. We want to give an algebro-geometric analogue of this construction. So let us ﬁrst deﬁne the process of formal differentiation. We start with the afﬁne case. Deﬁnition 7.4.1. Let f : X = Spec R → Y = Spec S be a morphism of afﬁne schemes, corresponding to a ring homomorphism S → R. We deﬁne the R-module ΩR/S , the module of relative differentials, to be the free R-module generated by formal symbols {dr ; r ∈ R}, modulo the relations: • d(r1 + r2 ) = dr1 + dr2 for r1 , r2 ∈ R, • d(r1 r2 ) = r1 dr2 + r2 dr1 for r1 , r2 ∈ R, • ds = 0 for s ∈ S.

134

Andreas Gathmann

Example 7.4.2. Let S = k be a ﬁeld and R = k[x1 , . . . , xn ], so that we consider the morphism f : An → pt. Then by the relations in ΩR/k , which are exactly the rules of differentiation k

∂ with the elements of k being the “constant” functions, it follows that d f = ∑i ∂xfi dxi for all f ∈ k[x1 , . . . , xn ]. So ΩR/k is just the free R-module generated by the symbols dx1 , . . . , dxn . Again let S = k, but now let R = k[x1 , . . . , xn ]/( f1 , . . . , fm ) be the coordinate ring of an afﬁne variety. By the same calculation as above, ΩR/S is still generated as an R-module by dx1 , . . . , dxn , but the relations fi give rise to relations d fi = 0 in ΩR/S . It is easy to see that these are all relations in ΩR/S , so we have

ΩR/S = (Rdx1 + · · · + Rdxn )/(∑

i

∂fj dxi , j = 1, . . . , m). ∂xi

In particular, if X = Spec R, k is algebraically closed, and P ∈ X is a closed point of X corresponding to a morphism R → k, then by deﬁnition 4.4.1 we see that ∂fj ΩR/S ⊗R k = dx1 , . . . , dxn /(∑ (P)dxi , j = 1, . . . , m) i ∂xi

∨ is just the dual TX,P of the tangent space to X at P.

Example 7.4.3. If Y is not a point, then the difference in the module of differentials is just that all elements of S (i.e. all differentials that come from Y ) are treated as “constants”. So then ΩR/S can be thought of as “the differentials on X modulo pull-backs of differentials on Y ”. We will probably not need this very often. Of course, if f : X → Y is a morphism of general (not necessarily afﬁne) schemes, we want to consider the relative differentials of every restriction of f to afﬁne opens of X and Y , and glue them together to get a quasi-coherent sheaf ΩX/Y . To do this, we have to give a different description of the relative differentials, as the construction given above does not glue very well. Lemma 7.4.4. Let S → R be a homomorphism of rings. Consider the map δ : R ⊗S R → R given by δ(r1 ⊗ r2 ) = r1 r2 and let I ⊂ R ⊗S R be its kernel. Then I/I 2 is an R-module that is isomorphic to ΩR/S . Proof. The R-module structure of I/I 2 is given by r · (r1 ⊗ r2 ) := rr1 ⊗ r2 = r1 ⊗ rr2 , where the second equality follows from rr1 ⊗ r2 − r1 ⊗ rr2 = (r1 ⊗ r2 ) · (r ⊗ 1 − 1 ⊗ r) ∈ I · I if r1 ⊗ r2 ∈ I. Deﬁne a map of R-modules ΩR/S → I/I 2 by dr → 1 ⊗ r − r ⊗ 1. Now we construct its inverse. The R-module E := R ⊕ ΩR/S is a ring by setting (r1 ⊕ dr1 ) · (r2 ⊕ dr2 ) := r1 r2 ⊕ (r1 dr2 + r2 dr1 ). It is easy to check that the map R × R → E given by (r1 , r2 ) → (r1 r2 , r1 dr2 ) is an S-bilinear ring homomorphism, hence gives rise to a map g : R ⊗S R → E. As g(I) ⊂ ΩR/S by deﬁnition and g(I 2 ) = 0, this induces a map I/I 2 → ΩR/S . It is easy to see that this is in fact the inverse of the map ΩR/S → I/I 2 given above. Remark 7.4.5. It is easy to translate this lemma into the language of schemes: let X = Spec R and Y = Spec S, so that the ring homomorphism S → R corresponds to a map X → Y . Then Spec R ⊗S R = X ×Y X, and δ : R ⊗S R → R corresponds to the diagonal morphism X → X ×Y X. Hence I ⊂ R ⊗S R is the ideal of the diagonal ∆(X) ⊂ X ×Y X. This motivates the following construction. Deﬁnition 7.4.6. Let f : X → Y be a morphism of schemes. Let ∆ : X → X ×Y X be the diagonal morphism, and let I = I∆(X)/X×Y X be its ideal sheaf. Then the sheaf of relative differentials ΩX/Y is deﬁned to be the sheaf ∆∗ (I/I 2 ) on X. If X is a scheme over a ﬁeld k and Y = Spec k is a point, then we will usually write ΩX/Y as ΩX .

7.

More about sheaves

135

Remark 7.4.7. Here we assume that the diagonal morphism ∆ is a closed immersion, which is the case if the schemes in question are separated (this is the analogue of lemma 2.5.3 for schemes). We will always assume this here to avoid further complications. Remark 7.4.8. It should be stressed that deﬁnition 7.4.6 is essentially useless for practical computations. Its only use is to show that a global object ΩX/Y exists that restricts to the old deﬁnition 7.4.1 on afﬁne open subsets. For applications, we will always use deﬁnition 7.4.1 and example 7.4.2 on open subsets. Remark 7.4.9. The sheaf ΩX/Y is always quasi-coherent: on afﬁne open subsets it restricts to the sheaf associated to the module ΩR/S constructed above. Remark 7.4.10. Any morphism f : X → Y of schemes over a ﬁeld induces a morphism of sheaves f ∗ ΩY → ΩX on X that is just given by dϕ → d( f ∗ ϕ) = d(ϕ ◦ f ) for any function ϕ on Y . Proposition 7.4.11. An n-dimensional scheme X (of ﬁnite type over an algebraically closed ﬁeld, e.g. a variety) is smooth if and only if ΩX is locally free of rank n. (Actually, this is a local statement: P ∈ X is a smooth point of X if and only if ΩX is (locally) free in a neighborhood of P.) Proof. One direction is obvious: if ΩX is locally free of rank n then its ﬁbers at any point ∨ P, i.e. the cotangent spaces TX,P , have dimension n. By deﬁnition this means that P is a smooth point of X. Now let us assume that X is smooth (at P). As the proposition is of local nature we can assume that X = Spec R with R = k[x1 , . . . , xr ]/( f1 , . . . , fm ). By example 7.4.2 we then have ∂fj ∨ TX,P = dx1 , . . . , dxr /(∑ (P)dxi , j = 1, . . . , m). i ∂xi As this vector space has dimension n, we know that the matrix of differentials D(P) = ( ∂ fli (P)) at the point P has rank r − n. Without loss of generality we can assume that the ∂x submatrix of D given by the ﬁrst r − n columns and rows has non-zero determinant. This ∨ means that dxr−n+1 , . . . , dxr form a basis of TX,P . But the condition for a determinant to be non-zero is an “open condition”, i.e. the set on which it is satisﬁed is open. In other words, there is a neighborhood U of P in X such that the submatrix of D(Q) given by the ﬁrst r − n columns and rows has non-zero determinant ∨ for all Q ∈ U. Consequently, the differentials dxr−n+1 , . . . , dxr generate TX,Q for all Q ∈ U. ∨ is at most n. But the opposite inequality dim T ∨ ≥ n In particular, the dimension of TX,Q X,Q is always true; so we conclude that the differentials dxr−n+1 , . . . , dxr actually form a basis of the cotangent space at all points Q ∈ U. So ΩX |U = OU dxr−n+1 ⊕ · · · ⊕ OU dxr , i.e. ΩX is locally free. Remark 7.4.12. There is a similar statement for any quasi-coherent sheaf F . It says that: (i) The dimension of the ﬁbers is an upper semi-continuous function. This means that if the dimension of the ﬁber of F at a point P is n, then it is at most n in some neighborhood of P. (ii) If the dimension of the ﬁbers is constant on some open subset U, then F |U is locally free. The idea of the proof of this statement is very similar to that of proposition 7.4.11.

136

Andreas Gathmann

Deﬁnition 7.4.13. Let X be a smooth n-dimensional scheme over an algebraically closed ﬁeld. The dual bundle Ω∨ of the cotangent bundle is called the tangent bundle and is X denoted TX . It is a locally free sheaf of rank n. The top exterior power Λn ΩX of the cotangent bundle is a locally free sheaf of rank 1; it is called the canonical bundle ωX of X. Remark 7.4.14. The importance of the cotangent / canonical bundles stems from the fact that these bundles are canonically deﬁned (hence the name) for any smooth scheme n. This gives e.g. a new method to show that two varieties are not isomorphic: if we have two varieties whose canonical bundles have different properties (say their spaces of global sections have different dimensions), then the varieties cannot be isomorphic. As an example, let us now compute the cotangent / tangent / canonical bundles of some easy varieties. Lemma 7.4.15. The cotangent bundle of Pn is determined by the exact sequence 0 → ΩPn → O (−1)⊕(n+1) → O → 0. (This sequence is usually called the Euler sequence.) Consequently, the tangent bundle ﬁts into the dual exact sequence 0 → O → O (1)⊕(n+1) → TPn → 0, and the canonical bundle is ωPn = O (−n − 1). Proof. We know already from example 7.4.2 that the cotangent bundle ΩPn is generated on the standard open subsets Ui = {xi = 0} ∼ An by the differentials d( x0 ), . . . , d( xn ) of the = xi xi x afﬁne coordinates. Therefore the differentials d( x ij ), where deﬁned, generate all of ΩPn . By the rules of differentiation we have to require formally that d xi xj = x j dxi − xi dx j . x2 j

Note that the dxi are not well-deﬁned objects, as the xi are not functions. But if we formally let the symbols dx0 , . . . , dxn be the names of the generators of O (−1)⊕(n+1) , the morphism of sheaves 1 xi xi ΩPn → O (−1)⊕(n+1) , d → · dxi − 2 · dx j xj xj xj is obviously well-deﬁned and injective. It is now easily checked that the sequence of the lemma is exact, with the last morphism given by

O (−1)⊕(n+1) → O , dxi → xi .

The sequence for the tangent bundle is obtained by dualizing. The statement about the canonical bundle then follows from corollary 7.3.7. Lemma 7.4.16. Let X ⊂ Pn be a smooth hypersurface of degree d, and let i : X → Pn be the inclusion morphism. Then the cotangent bundle ΩX is determined by the exact sequence 0 → OX (−d) → i∗ ΩPn → ΩX → 0. Consequently, the tangent bundle is determined by the exact sequence 0 → TX → i∗ TPn → OX (d) → 0, and the canonical bundle is ωX = OX (d − n − 1).

7.

More about sheaves

137

Proof. We claim that the exact sequence is given by 0 → OX (−d) → i∗ ΩPn → ΩX → 0 ϕ → d( f · ϕ), dϕ → d(ϕ|X ), where f is the equation deﬁning X. In fact, the second map is just the usual pull-back of differential forms as in remark 7.4.10 (which is just a restriction in this case). It is surjective g because functions on X are locally of the form h for some homogeneous polynomials g and h of the same degree, so they are locally obtained by restricting a function on Pn to X. It is not an isomorphism though, because we have the identity f = 0 on X. Consequently, differentials dϕ are zero when restricted to X if and only if ϕ contains f as a factor. This explains the ﬁrst map of the above sequence. As in the previous lemma, the statements about the tangent and canonical bundles are obtained by dualizing and applying corollary 7.3.7, respectively. Remark 7.4.17. In general, if i : X → Y is a closed immersion of smooth schemes over a ﬁeld, there is an injective morphism TX → i∗ TY of sheaves on X. In other words, at points in X the tangent spaces of X are just subspaces of the tangent spaces of Y . The quotient TY,P /TX,P is called the normal space, and consequently the quotient bundle NX/Y = i∗ TY /TX is called the normal bundle. This is the same construction as in differential geometry. Thus lemma 7.4.16 just tells us that the normal bundle of a degree-d hypersurface in Pn is NX/Pn = OX (d). Example 7.4.18. Let us evaluate lemma 7.4.16 in the simplest cases, namely for curves X ⊂ P2 of low degrees d. (i) d = 1: A linear curve in P2 is just isomorphic to P1 . We get ΩX = ωX = O (1 − 2 − 1) = O (−2) by lemma 7.4.16. This is consistent with lemma 7.4.15 for n = 1. (ii) d = 2: We know from example 3.3.11 that a smooth plane conic is again just isomorphic to P1 by means of a quadratic map f : P1 → X ⊂ P2 . Our formula of lemma 7.4.16 gives ωX = OX (2 − 2 − 1) = OX (−1). By pulling this back via f we obtain ωX = OP1 (−2) by example 7.2.12. So by applying the isomorphism to case (i) we get the same canonical bundle back — which has to be the case, as the cotangent bundle is canonically deﬁned and cannot change with the embedding in projective space. (iii) d = 3: Here we get ωX = O (3 − 2 − 1) = O , i.e. the canonical bundle is simply isomorphic to the sheaf of regular functions. We can understand this from our representation in proposition 6.5.7 of cubic curves as complex tori of the form C/Λ for some lattice Λ ⊂ C. If z is the complex coordinate on C, note that the differential form dz is invariant under shifts in Λ, as d(z + a) = dz for all a ∈ C. Therefore dz descends to a global differential form on X = C/Λ without zeros or poles. It follows that we have an isomorphism OX → ωX given by ϕ → ϕ · dz. 7.5. Line bundles on curves. We now want to specialize even further and consider vector bundles of rank 1 (also called “line bundles”, because their ﬁbers are just lines) on smooth curves. This section should be compared to section 6.3 where we considered divisors on such curves. We will show that divisor classes and line bundles are essentially the same thing. Recall that the group Pic X of divisor classes on a smooth curve X has a group structure in a natural way. So let us ﬁrst make the set of all line bundles on X into a group as well. In fact, this can be done for any scheme: Deﬁnition 7.5.1. Let X be a scheme. A line bundle on X is a vector bundle (i.e. a locally free sheaf) of rank 1. We denote the set of all line bundles on X by Pic X. This set has a

138

Andreas Gathmann

natural structure of Abelian group, with multiplication given by tensor products, inverses by taking duals, and the neutral element by the structure sheaf. We will now restrict our attention to smooth curves. To set up a correspondence between line bundles and divisors, we will have to deﬁne the divisor of a (rational) section of a line bundle. This is totally analogous to the divisor of a rational function in deﬁnition 6.3.4. Deﬁnition 7.5.2. Let L be a line bundle on a smooth curve X, and let P ∈ X be a point. Assume that we are given a section s ∈ L (U) of L on some neighborhood U of P. As L is a line bundle, there is an isomorphism ψ : L |U → OU (possibly after shrinking U). The order of vanishing ordP s of the section s at P is deﬁned to be the order of vanishing of the regular function ψ(s) at P. Remark 7.5.3. Note that this deﬁnition does not depend on the choice of ψ: if ψ : L |U → OU is another isomorphism, then the composition ψ ◦ ψ−1 : OU → OU is an isomorphism of the structure sheaf, which must be given by multiplication with a function ϕ that is nowhere zero (in particular not at P). So we have an equation of divisors (ψ (s)) = (ψ ψ−1 ψ(s)) = (ϕ · ψ(s)) = (ϕ) + (ψ(s)) = (ψ(s)), which shows that ordP s is well-deﬁned. Deﬁnition 7.5.4. Let L be a line bundle on a smooth curve X. A rational section of L over U is a section of the sheaf L ⊗OX KX , where KX denotes the “sheaf of rational functions” whose value at every open subset U ⊂ X is just K(X). In other words, a rational section of a line bundle is given by an ordinary section of the line bundle, possibly multiplied with a rational function. Now let P ∈ X be a point, and let s be a rational section of L in a neighborhood of P. With the same isomorphism ψ as in deﬁnition 7.5.2, the order ordP s of s at P is deﬁned to be the order of the rational function ψ(s) at P. (This is well-deﬁned for the reason stated in remark 7.5.3.) If s is a global rational section of L , we deﬁne the divisor (s) of s to be (s) =

P∈X

∑ ordP s · P ∈ Div X.

Example 7.5.5. Let X = P1 with homogeneous coordinates x0 , x1 . (i) Consider the global section s = x0 x1 of OX (2). It vanishes at the points P = (0 : 1) and Q = (1 : 0) with multiplicity 1 each, so (s) = P + Q. (ii) The divisor of the global rational section s = x10 of OX (−1) is (s) = −P. To show that Pic X ∼ Pic X for smooth curves we need the following key lemma (which = is the only point at which smoothness is needed). Lemma 7.5.6. Let X be a curve (over some algebraically closed ﬁeld), and let P ∈ X be a smooth point. Then there is a function ϕP in a neighborhood of P such that (i) ϕP vanishes at P with multiplicity 1, i.e. its divisor contains the point P with multiplicity 1. (ii) ϕP is non-zero at all points distinct from P. Proof. We can assume that X = Spec R is afﬁne, with R = k[x1 , . . . , xr ]/( f1 , . . . , fm ) being the coordinate ring of X. As P is a smooth point of X, its cotangent space ∂fj ∨ TX,P = dx1 , . . . , dxr /(∑ (P) dxi for all j) i ∂xi is one-dimensional. Let ϕP be any linear function such that dϕP generates this vector space. Then ϕP vanishes at P with multiplicity 1 by construction. We can now pick a neighborhood of P such that ϕP does not vanish at any other point.

7.

More about sheaves

139

Remark 7.5.7. If the ground ﬁeld is C and one thinks of X as a complex one-dimensional manifold, one can think of the function ϕP of lemma 7.5.6 as a “local coordinate” of X around P, i.e. a function that gives a local isomorphism of X with C, with P mapping to 0 ∈ C. Note however that this is not true in the algebraic category, as the Zariski open subsets are too big. We are now ready to prove the main proposition of this section. Deﬁnition 7.5.8. A divisor D = ∑P aP P on a smooth curve X is called effective (written D ≥ 0) if aP ≥ 0 for all P. Proposition 7.5.9. Let X be a smooth curve. Then there is an isomorphism of Abelian groups Pic X → Pic X L → (s) for any rational section s of L . Its inverse is given by Pic X → Pic X D → O (D), where O (D) is the line bundle deﬁned by

O (D)(U) = {ϕ ∈ K(X) ; (ϕ) + D ≥ 0 on U}.

Proof. We have to check a couple of things: (i) If L is a line bundle, then there is a rational section s of L : This is obvious, as L is isomorphic to O on an open subset of X. So we can ﬁnd a section of L on this open subset (corresponding to the constant function 1). This will be a rational section of L on all of X. (ii) The divisor class (s) of a rational section s of L does not depend on the choice of s: If we have another section s , then the quotient ss will be a rational function, which has divisor class zero by deﬁnition of Pic X. So (s) = ( ss ·s ) = ( ss )+(s ) = (s ) in Pic X. (iii) If D is a divisor then O (D) is actually a line bundle: let P ∈ X be a point and choose a neighborhood U of P such that no point of U\P is contained in D. Let n be the coefﬁcient of P in D. Then an isomorphism ψ : O (D) → O on U is given by multiplication with ϕn , where ϕP is the function of lemma 7.5.6. In fact, P a rational function ϕ in K(X) is by deﬁnition a section of O (D) if and only if ordP ϕ + n ≥ 0, which is the case if and only if ϕ · ϕn is regular at P. P (iv) If the divisors D and D deﬁne the same element in Pic X then O (D) = O (D ): By assumption we have D − D = (ϕ) in Pic X for some rational function ϕ. Obviously, this induces an isomorphism O (D) → O (D ) through multiplication with ϕ. We have now shown that the maps stated in the proposition are well-deﬁned. Let us now check that the two maps are inverse to each other. (v) Pic X → Pic X → Pic X: Let s0 be a rational section of a line bundle L , and consider O ((s0 )) = {ϕ ∈ K(X) ; (ϕ) + (s0 ) ≥ 0}. We have an isomorphism s L → O ((s0 )), s → . s0 (vi) Pic X → Pic X → Pic X: The (constant) rational function 1 deﬁnes a rational section of O (D). To determine its order at a point P we have to apply the local isomorphism with O constructed in (iii): the order of this rational section at P is just the order of 1 · ϕn , which is n. This is exactly the multiplicity of P in D, so P the divisor of our section is precisely D.

140

Andreas Gathmann

Finally, we have to check that the map is a homomorphism of groups. But this is clear: if s and s are rational sections of L and L , respectively, then ss is a rational section of L ⊗ L , and (ss ) = (s) + (s ). Hence tensor products of line bundles correspond to addition of divisors under our correspondence. Deﬁnition 7.5.10. Let X be a smooth curve. From now on we will identify line bundles with divisor classes and call both groups Pic X. In particular, this deﬁnes the degree of a line bundle (to be the degree of the associated divisor class). The divisor class associated to the canonical bundle ωX is denoted KX ; it is called the canonical divisor (class). Example 7.5.11. We have seen in lemma 6.3.11 that Pic P1 = Z, i.e. there is exactly one divisor class in every degree. Consequently, there is exactly one line bundle for every degree n, which is of course just O (n). On the other hand, if X ⊂ P2 is a smooth cubic curve we know from corollary 6.3.15 that Pic X consists of a copy of X in every degree. So on a cubic curve there are (many) more line bundles than just the bundles of the form O (n). Remark 7.5.12. The correspondence of proposition 7.5.9 allows us to deﬁne the pull-back f ∗ D of a divisor class D on Y for any (surjective) morphism of smooth curves f : X → Y : it is just given by pulling back the corresponding line bundle. In fact, we can even deﬁne a pull-back f ∗ D for any divisor D ∈ DivY that induces this construction on the corresponding divisor class: let P ∈ X be any point, and let Q = f (P) be its image, considered as an element of DivY . Then the subscheme f −1 (Q) of X has a component whose underlying point is P. We deﬁne the ramiﬁcation index eP of f at P to be the length of this component subscheme. In more down to earth terms, this means that we take a function ϕQ as in lemma 7.5.6 that vanishes at Q with multiplicity 1, and deﬁne eP to be the order of vanishing of the pull-back function f ∗ ϕQ = ϕQ ◦ f at P. The ramiﬁcation index has a simple interpretation in complex analysis: in the ordinary topology the curves X and Y are locally isomorphic to the complex plane, so we can pick local coordinates z on X around P and w on Y around Q. Every holomorphic map is now locally of the form z → w = uzn for some n ≥ 1 and an invertible function u (i.e. a function that is non-zero at P). The number n is just the ramiﬁcation index deﬁned above. It is 1 if and only if f is a local isomorphism at P in complex analysis. We say that f is ramiﬁed at P if n = eP > 1, and unramiﬁed at P otherwise.

P f Y X P X f Y

Q eP =1

Q eP =2

If we now consider a point Q as an element of DivY , we simply deﬁne f ∗Q =

P: f (P)=Q

∑

eP · P

and extend this by linearity to obtain a homomorphism f ∗ : DivY → Div X. In other words, f ∗ D is just obtained by taking the inverse image points of the points in D with the appropriate multiplicities. Using the correspondence of proposition 7.5.9 it is now easily checked that the induced map f ∗ : PicY → Pic X on the Picard groups agrees with the pull-back of line bundles.

7.

More about sheaves

141

2 2 Example 7.5.13. Let f : X = P1 → Y = P1 be the morphism given by (x0 : x1 ) → (x0 : x1 ). ∗ (1 : 0) = 2 · (1 : 0) and f ∗ (1 : 1) = (1 : 1) + (1 : −1) as divisors in X. Then f

As an application of line bundles, we will now see how they can be used to describe morphisms to projective spaces. This works for all schemes (not just curves). Lemma 7.5.14. Let X be a scheme over an algebraically closed ﬁeld. There is a one-toone correspondence line bundles L on X together with global sections s0 , . . . , sr ∈ Γ(X, L ) such that: {morphisms f : X → Pr } ←→ for all P ∈ X there is some si with si (P) = 0 Proof. “←−”: Given r + 1 sections of a line bundle L on X that do not vanish simultaneously, we can deﬁne a morphism f : X → Pr by setting f (P) = (s0 (P) : · · · : sr (P)). Note s that the values si (P) are not well-deﬁned numbers, but their quotients s ij (P) are (as they are sections of L ⊗ L ∨ = O , i.e. ordinary functions). Therefore f (P) is a well-deﬁned point in projective space. “−→”: Given a morphism f : X → Pr , we set L = f ∗ OPr (1) and si = f ∗ xi , where we consider the xi as sections of O (1) (and thus the si as sections of f ∗ O (1)). Remark 7.5.15. One should regard this lemma as a generalization of lemma 3.3.9 where we have seen that a morphism to Pr can be given by specifying r + 1 homogeneous polynomials of the same degree. Of course, this was just the special case in which the line bundle of lemma 7.5.14 is O (d). We had mentioned already in remark 3.3.10 that not all morphisms are of this form; this translates now into the statement that not all line bundles are of the form O (n). 7.6. The Riemann-Hurwitz formula. Let X and Y be smooth projective curves, and let f : X → Y be a surjective morphism. We want to compare the sheaves of differentials on X and Y . Note that every projective curve admits a surjective morphism to P1 : by deﬁnition it sits in some Pn to start with, so we can ﬁnd a morphism to P1 by repeated projections from points not in X. So if we know the canonical bundle of P1 (which we do by lemma 7.4.15: it is just OP1 (−2)) and how canonical bundles transform under morphisms, we can at least in theory compute the canonical bundles of every curve. Deﬁnition 7.6.1. Let f : X → Y be a surjective morphism of smooth projective curves. We deﬁne the ramiﬁcation divisor of f to be R = ∑P∈X (eP − 1) · P ∈ Div X, where eP is the ramiﬁcation index of f at P deﬁned in remark 7.5.12. So the divisor R contains all points at which f is ramiﬁed, with appropriate multiplicities. Proposition 7.6.2. (Riemann-Hurwitz formula) Let f : X → Y be a surjective morphism of smooth projective curves, and let R be the ramiﬁcation divisor of f . Then KX = f ∗ KY +R (or equivalently ωX = f ∗ ωY ⊗ OX (R)) in Pic X. Proof. Let P ∈ X be any point, and let Q = f (P) be its image point. Choose local functions ϕP and ϕQ around P (resp. Q) that vanish at P (resp. Q) with multiplicity 1 as in lemma 7.5.6. Then by the deﬁnition of the ramiﬁcation index we have f ∗ ϕQ = u · ϕeP P for some local function u on X with no zero or pole at P. Now pick a global rational section α of ωY . If its divisor (α) contains the point Q with multiplicity n, we can write locally α = v · ϕn dϕQ , Q

142

Andreas Gathmann

where v is a local function on Y with no zero or pole at Q. Inserting these equations into each other, we see that f ∗ α = f ∗ v · ( f ∗ ϕn )d( f ∗ ϕQ ) = un f ∗ v · ϕneP · (ϕPp du + ue p ϕeP −1 dϕP ). Q P P This vanishes at P to order neP + eP − 1. Summing this over all points P ∈ X we see that the divisor of f ∗ α is f ∗ (α) + R. As KX = ( f ∗ α) and f ∗ KY = f ∗ (α), the proposition follows. We will now study the same situation from a topological point of view (if the ground ﬁeld is C). Then X and Y are two-dimensional compact manifolds. For such a space X, we say that a cell decomposition of X is given by writing X as a ﬁnite disjoint union of points, (open) lines, and discs. This decomposition should be “nice” in a certain topological sense, e.g. the boundary points of every line in the decomposition must be points of the decomposition. It takes some work to make this deﬁnition (and the following propositions) bullet-proof. We do not want to elaborate on this, but only remark that every “reasonable” decomposition that one could think of will be allowed. For example, here are three valid decompositions of the Riemann sphere P1 : C

e

(i)

(ii)

(iii)

(In (i), we have only one point (the north pole), no line, and one “disc”, namely P1 minus the north pole). We denote by σ0 , σ1 , σ2 the number of points, lines and discs in the decomposition, respectively. So in the above examples we have (σ0 , σ1 , σ2 ) = (1, 0, 1), (2, 2, 2), and (6, 8, 4), respectively. Of course there are many possible decompositions for a given curve X. But there is an important number that is invariant: Lemma 7.6.3. The number σ0 − σ1 + σ2 depends only on X and not on the chosen decomposition. It is called the (topological) Euler characteristic χ(X) of X. Proof. Let us ﬁrst consider the case when we move from one decomposition to a “ﬁner” one, i.e. if we add points or lines to the decomposition. For example, in the above pictures (iii) is a reﬁnement of (ii), which is itself a reﬁnement of (i). Note that every reﬁnement is obtained by applying the following steps a ﬁnite number of times: (i) Adding another point on a line: In this case we raise σ0 and σ1 by 1, so the alternating sum σ0 − σ1 + σ2 does not change (see the picture below).

add a point

add a line

(ii) Adding another line in a disc: In this case we raise σ1 and σ2 by 1, so the alternating sum σ0 − σ1 + σ2 again does not change (see the picture above).

7.

More about sheaves

143

So we conclude that the alternating sum σ0 − σ1 + σ2 does not change under reﬁnements. But it is easily seen that any two decompositions have a common reﬁnement (which is essentially given by taking all the points and lines in both decompositions, and maybe add more points where two such lines intersect. For example, the common reﬁnement of decomposition (ii) above and the same decomposition rotated clockwise by 90 degrees would be (iii)). It follows that the alternating sum is independent of the decomposition. We have already noted in example 0.1.1 that a smooth complex curve is topologically a (real) closed surface with a certain number g of “holes”. The number g is called the genus of the curve. Let us compute the topological Euler characteristic of such a curve of genus g: Lemma 7.6.4. The Euler characteristic of a curve of genus g is equal to 2 − 2g. Proof. Take e.g. the decomposition illustrated in the following picture:

It has 2g + 2 points, 4g + 4 lines, and 4 discs, so the result follows. Let us now compare the Euler characteristics of two curves X and Y if we have a morphism f : X → Y : Lemma 7.6.5. Let f : X → Y be a morphism of complex smooth projective curves. Let n be the number of inverse image points of any point of Y under f . As in proposition 7.6.2 let R be the ramiﬁcation divisor of f . Then −χ(X) = −n · χ(Y ) + deg R. Proof. Choose “compatible” decompositions of X and Y , i.e. loosely speaking decompositions such that the inverse images of the points / lines / discs of the decomposition of Y are (ﬁnite) unions of points / lines / discs of the decomposition of X, and such that all points / lines / discs of the decomposition of X arise in this way. Moreover, we require that all ramiﬁcation points of f are points of the decomposition of X. (It is easily seen that this can always be achieved.) Denote by σX , σX , σX the number of points / lines / discs of the 0 1 2 decomposition of X, and similarly for Y . As every point of Y that is not the image of a ramiﬁcation point has n inverse images under f , it follows that σX = n σY and σX = n σY . We do not have σX = n σY however: if P 1 1 2 2 0 0 is a ramiﬁcation point, i.e. eP > 1, then f is locally eP -to-one around P, i.e. P counts for eP in n σY , whereas it is actually only one point in the decomposition of X. Hence we have to 0 subtract eP − 1 for any ramiﬁcation point P from n σY to get the correct value of σX . This 0 0 means that σX = n σY − deg R and hence −χ(X) = −n χ(Y ) + deg R. 0 0 Corollary 7.6.6. Let X be a (complex) smooth projective curve. Then deg KX = 2g − 2. Proof. As we have already remarked, any such curve X admits a surjective morphism f to P1 by projection. Using that deg KP1 = −χ(P1 ) = −2 (by lemma 7.4.15 and lemma 7.6.4)

144

Andreas Gathmann

and applying lemma 7.6.5 together with the Riemann-Hurwitz formula 7.6.2, we see that deg KX = −χ(X). The result therefore follows from lemma 7.6.4. 7.7. The Riemann-Roch theorem. As in the last section let X be a smooth projective curve of genus g over an algebraically closed ﬁeld. For any line bundle L we want to compute the dimensions of the vector spaces Γ(L ) of global sections of L . We will denote this dimension by h0 (L ) (the reason for this notation will become obvious when we discuss cohomology in chapter 8). By abuse of notation we will also write h0 (D) instead of h0 (O (D)) for any divisor D. We should remark that this is a classical question that was one of the ﬁrst problems studied in algebraic geometry: given a smooth projective curve X (resp. a compact onedimensional complex manifold), points P1 , . . . , Pr ∈ X, and numbers a1 , . . . , ar ≥ 0, what is the dimension of the space of rational (resp. meromorphic) functions on X that have poles of order at most ai at the points Pi and are regular (resp. holomorphic) everywhere else? In our language, this just means that we are looking for the number h0 (a1 P1 + · · · + ar Pr ). Example 7.7.1. Let D be a divisor on X with negative degree. Recall that sections of O (D) are just rational functions ϕ on X such that (ϕ) + D is effective. Taking degrees, this certainly implies that deg(ϕ) + deg D ≥ 0. But deg(ϕ) = 0 by remark 6.3.5 and deg D < 0 by assumption, which is a contradiction. Hence we conclude that h0 (D) = 0 if deg D < 0: there are no global sections of O (D) in this case. Example 7.7.2. Let L be the line bundle OX (n) for some n ∈ Z. Recall that sections of L f are of the form g with f and g homogeneous such that deg f − deg g = n. Now for global sections g must be a constant function (otherwise we would have a pole somewhere), so we conclude that Γ(L ) is simply the n-th graded piece of the homogeneous coordinate ring S(X).In other words, h0 (L ) is by deﬁnition equal to the value hX (n) of the Hilbert function introduced in section 6.1. We have seen in proposition 6.1.5 that hX (n) is equal to a linear polynomial χX (n) in n for n 0. Moreover, the linear coefﬁcient of χX (n) is the degree of OX (n), and the constant coefﬁcient is 1 − g by deﬁnition of g (see example 6.1.10). So we conclude that h0 (D) = deg D + 1 − g if D is the divisor class associated to a line bundle OX (n) for n 0.

Theorem 7.7.3. (Riemann-Roch theorem for line bundles on curves) Let X be a complex smooth projective curve of genus g. Then for any divisor D on X we have h0 (D) − h0 (KX − D) = deg D + 1 − g. Proof. Step 1. Recall that for any point P ∈ X and any divisor D we have the exact “skyscraper sequence” by exercise 7.8.4 0 → O (D) → O (D + P) → kP → 0 where the last morphism is given by evaluation at the point P. From this we get an exact sequence of global sections 0 → Γ(O (D)) → Γ(O (D + P)) → C (where the last map is in general not surjective, see example 7.1.18). Therefore h0 (D + P) − h0 (D) is either 0 or 1. If we denote the left hand side of the Riemann-Roch theorem by χ(D) = h0 (D) − h0 (KX − D), we conclude that χ(D + P) − χ(D) = (h0 (D + P) − h0 (D)) + (h0 (KX − D) − h0 (KX − D − P)) is either 0, 1, or 2. (Of course, what we want to prove is that χ(D + P) − χ(D) is always equal to 1.)

7.

More about sheaves

145

Step 2. We want to rule out the case that χ(D + P) − χ(D) = 2. For this we actually have to borrow a theorem from complex analysis. So assume that h0 (D + P) − h0 (D) = 1 and h0 (KX − D) − h0 (KX − D − P) = 1. The fact that h0 (D + P) − h0 (D) = 1 means precisely that there is a global section ϕ of OX (D + P) that is not a global section of OX (D), i.e. that ϕ is a rational section of OX (D) that has a simple pole at P and is regular at all other points. Similarly, there is a global section α of OX (KX − D) that is not a global section of OX (KX − D − P). In other words, α is a global section of ωX ⊗ L ∨ that does not vanish at P. By multiplication we see that ϕ·α is a rational section of L ⊗ (ωX ⊗ L ∨ ) = ωX that has a simple pole at P and is regular at all other points. In other words, ϕ · α is a global rational differential form with just a single pole which is of order 1. But this is a contradiction to the residue theorem of complex analysis: the sum of the residues of any rational (or meromorphic) differential form on a compact Riemann surface is zero, but in our case we have ∑Q∈X resQ (ϕ · α) = resP (ϕ · α) = 0. Step 3. We claim that χ(D) ≥ deg D + 1 − g for all divisors D. Note that we can choose points P1 , . . . , Pr such that D + P1 + · · · + Pr is precisely the intersection divisor of X with a certain number n of hyperplanes: for every point in D we just choose a hyperplane through that point and add all other intersection points with X to the Pi . This then means that O (D + P1 + · · · + Pr ) = O (n). By possibly adding more intersection points of X with hyperplanes we can make n arbitrarily large. So by example 7.7.2 we ﬁnd that h0 (D + P1 + · · · + Pr ) = deg D + r + 1 − g. Moreover, if n (and thus r) is large enough we see by example 7.7.1 that h0 (KX − D − P1 − · · · − Pr ) = 0 and therefore χ(D + P1 + · · · + Pr ) = deg D + r + 1 − g. But by step 2 we know that subtracting a point from the divisor will decrease χ(·) by 0 or 1. If we apply this r times to the points P1 , . . . , Pr we conclude that χ(D) ≥ (deg D + r + 1 − g) − r, as we have claimed. Step 4. Replacing D by KX − D in the inequality of step 3 yields −χ(D) = h0 (KX − D) − h0 (D) ≥ deg KX − deg D + 1 − g = − deg D − 1 + g as deg KX = 2g − 2 by corollary 7.6.6. Combining the two inequalities of steps 3 and 4 proves the theorem. Remark 7.7.4. If D is the divisor associated to the line bundle O (n) (for any n), note that χ(D) is just the value χX (n) of the Hilbert polynomial. So for these line bundles we can reinterpret our main proposition 6.1.5 about Hilbert polynomials as follows: the difference between hX (n) and χX (n) is simply h0 (ωX ⊗ OX (−n)). As this vanishes for large n by degree reasons, it follows that hX (n) = χX (n) for large n. Example 7.7.5. Setting D = 0 in the Riemann-Roch theorem yields h0 (KX ) = g. This gives an alternate deﬁnition of the genus of a smooth projective curve: one could deﬁne the genus of such a curve as the dimension of the space of global differential forms. This deﬁnition has the advantage that it is immediately clear that it is well-deﬁned and independent of the projective embedding (compare this to example 6.1.10). Remark 7.7.6. In general one should think of the Riemann-Roch theorem as a formula to compute h0 (D) for any D, modulo an “unwanted” correction term h0 (KX − D). In many applications one can make this correction term vanish, e.g. by making the degree of D large enough so that deg(KX − D) becomes negative.

146

Andreas Gathmann

Remark 7.7.7. There are numerous generalizations of the Riemann-Roch theorem. In fact, there are whole books on Riemann-Roch type theorems. Let us mention some of the generalizations without proof: (i) The requirement that the ground ﬁeld be C is not essential. The very same statement holds over any algebraically closed ground ﬁeld (the proof has to be changed though at step 2 where we invoked complex analysis). (ii) The requirement that the curve be projective is not essential either, it only needs to be complete (i.e. “compact”). (iii) Instead of a line bundle one can take a vector bundle: if F is any vector bundle on X of rank r then h0 (F ) − h0 (ωX ⊗ F ∨ ) = deg Λr F + r(1 − g) (see example 10.4.7). (iv) There are versions of the Riemann-Roch theorem for singular curves as well. (Note that in the singular case we do not have a canonical bundle, so one needs a new idea here.) (v) There are also versions of the Riemann-Roch theorem for varieties of dimension bigger than 1 (see theorem 10.4.5). (vi) Finally, the same theorem can be proven (with the same proof actually) in complex analysis, where h0 (D) then denotes the dimension of the space of meromorphic functions with the speciﬁed zeros and poles. As the resulting dimension does change we conclude that on a projective smooth complex curve every meromorphic function is in fact rational. This is an example of a very general result that says that complex analysis essentially reduces to algebraic geometry in the projective case (in other words, we “do not gain much” by allowing holomorphic functions instead of rational ones in the ﬁrst place). As an application of the Riemann-Roch theorem let us consider again morphisms to projective spaces. Let X be a smooth projective curve, and let D be a divisor on X. Let s0 , . . . , sr be a basis of the space Γ(O (D)) of global sections of O (D). Then we have seen in lemma 7.5.14 that we get a morphism X → Pr , P → (s0 (P) : · · · : sr (P)) provided that the sections si do not vanish simultaneously at any point. Using the RiemannRoch theorem we can now give an easy criterion when this is the case. Note ﬁrst however that picking a different basis of section would result in a morphism that differs from the old one simply by a linear automorphism of Pr . Thus we usually say that the divisor D (or its associated line bundle) determines a morphism to Pr up to automorphisms of Pr . Proposition 7.7.8. Let X be a smooth projective curve of genus g, and let D be a divisor on X. (i) If deg D ≥ 2g then the divisor D determines a morphism X → Pr as above. (ii) If deg D ≥ 2g + 1 then moreover this morphism is an embedding (i.e. an isomorphism onto its image). Proof. (i): By what we have said above we simply have to show that for every point P ∈ X there is a global section s ∈ Γ(O (D)) that does not vanish at P. By the degree condition we know that deg(KX − D) ≤ 2g − 2 − 2g < 0 and deg(KX − D+P) ≤ 2g−2−2g+1 < 0. So by example 7.7.1 we get from the Riemann-Roch theorem that h0 (D) = deg D + 1 − g and h0 (D − P) = (deg D − 1) + 1 − g. In particular we have h0 (D) − h0 (D − P) = 1, i.e. there is a section s ∈ Γ(O (D)) that is not a section of O (D − P), i.e. that does not vanish at P.

7.

More about sheaves

147

(ii): The idea of the proof is the same as in (i). However, as we have not developed enough powerful techniques yet to prove that a morphism has an inverse, we will restrict ourselves to proving that the morphism determined by D is bijective. So let P and Q be distinct points of X. To prove that they are mapped to different points it sufﬁces to show that there is a section s ∈ Γ(O (D)) with s(P) = 0, s(Q) = 0: the morphism R → (s(R) : s (R) : · · · ) then maps P to a point with the ﬁrst coordinate 0, while the ﬁrst coordinate is non-zero for the image point of Q. To ﬁnd this section s, simply apply the argument of (i) to D − P and the point Q: we get h0 (D − P) − h0 (D − P − Q) = 1, i.e. there is a section s ∈ Γ(O (D − P)) that is not a section of O (D − P − Q), i.e. it is a section of O (D) that vanishes at P but not at Q. Example 7.7.9. If X is a smooth projective curve of genus g ≥ 2 we get a canonical embedding X → Pr into a projective space (up to automorphisms by Pr ) by taking the morphism associated to the divisor 3KX . This follows by part (ii) of proposition 7.7.8 as 3(2g − 2) ≥ 2g + 1 if g ≥ 2. By remark 7.7.7 (ii) the same is true for any complete (i.e. “compact”) curve that is not necessarily given initially as a subvariety of projective space. 7.8. Exercises. Exercise 7.8.1. Let F be a presheaf on a topological space X, and let F be its sheaﬁﬁcation as in deﬁnition 7.1.10. Show that (i) There is a natural morphism θ : F → F . (ii) Any morphism from F to a sheaf factors uniquely through θ. Exercise 7.8.2. Let f : F → G be a morphism of sheaves of abelian groups on a topological space X. Show that f is injective / surjective / an isomorphism if and only if all induced maps fP : FP → GP on the stalks are injective / surjective / isomorphisms. Exercise 7.8.3. Let f : F1 → F2 be a morphism of locally free sheaves on a scheme X over a ﬁeld k. Let P ∈ X be a point, and denote by (Fi )P the ﬁber of the vector bundle Fi over P, which is a k-vector space. Are the following statements true or false: (i) If F1 → F2 is injective then the induced map (F1 )P → (F2 )P is injective for all P ∈ X. (ii) If F1 → F2 is surjective then the induced map (F1 )P → (F2 )P is surjective for all P ∈ X. Exercise 7.8.4. Prove the following generalization of example 7.1.16: If X is a smooth curve over some ﬁeld k, L a line bundle on X, and P ∈ X a point, then there is an exact sequence 0 → L (−P) → L → kP → 0, where kP denotes the “skyscraper sheaf” kP (U) = k if P ∈ U, 0 if P ∈ U. /

Exercise 7.8.5. If X is an afﬁne variety over a ﬁeld k and F a locally free sheaf of rank r on X, is then necessarily F ∼ OX ? = ⊕r Exercise 7.8.6. Let X be a scheme, and let F be a locally free sheaf on X. Show that (F ∨ )∨ ∼ F . Show by example that this statement is in general false if F is only quasi= coherent but not locally free. Exercise 7.8.7. Figure out what exactly goes wrong with the correspondence between line bundles and divisor classes on a curve X if X is singular. Can we still associate a divisor to any section of a line bundle? Can we still construct a line bundle from any divisor?

148

Andreas Gathmann

Exercise 7.8.8. What is the line bundle on Pn × Pm leading to the Segre embedding Pn × Pm → PN by the correspondence of lemma 7.5.14? What is the line bundle leading to the degree-d Veronese embedding Pn → PN ? Exercise 7.8.9. Show that any smooth projective curve of genus 2. . . (i) can be realized as a curve of degree 5 in P3 , (ii) admits a two-to-one morphism to P1 . How many ramiﬁcation points does such a morphism have? Exercise 7.8.10. Let X be a smooth projective curve, and let P ∈ X be a point. Show that there is a rational function on X that is regular everywhere except at P.

8.

Cohomology of sheaves

149

8. C OHOMOLOGY OF SHEAVES

For any quasi-coherent sheaf F on a scheme X we construct the cohomology groups ˇ H i (X, F ) for i ≥ 0 using the Cech complex associated to an afﬁne open cover of X. We show that the cohomology groups do not depend on the choice of afﬁne open cover. The cohomology groups H i (X, F ) vanish for i > 0 if X is afﬁne, and in any case for i > dim X. For any short exact sequence of sheaves on X there is an associated long exact sequence of the corresponding cohomology groups. If L is a line bundle of degree at least 2g−1 on a smooth projective curve of genus g then the cohomology group H 1 (X, L ) is zero. Using this “vanishing theorem” we reprove the Riemann-Roch theorem in a cohomological version. Comparing this to the old version yields the equality dim H 0 (KX − D) = dim H 1 (D) for any divisor D, which is a special case of the Serre duality theorem. As an application we can now deﬁne the genus of a possibly singular curve to be dim H 1 (X, OX ). We compute the cohomology groups of all line bundles on projective spaces. As a consequence, we obtain the result that the cohomology groups of coherent sheaves on projective schemes are always ﬁnite-dimensional vector spaces, and that H i (X, F ⊗ OX (d)) = 0 for all i > 0 and d 0.

8.1. Motivation and deﬁnitions. There are numerous ways to motivate the theory of cohomology of sheaves. Almost all of them are based on the observation that “the functor of taking global sections of a sheaf is not exact”, i.e. given an exact sequence of sheaves of Abelian groups 0 → F1 → F2 → F3 → 0 on a scheme (or topological space) X, by taking global sections we get an exact sequence 0 → Γ(F1 ) → Γ(F2 ) → Γ(F3 ) of Abelian groups in which the last map Γ(F2 ) → Γ(F3 ) is in general not surjective. We have seen one example of this in example 7.1.18. Here is one more example: Example 8.1.1. Let X ⊂ Pn be a smooth hypersurface of degree d with inclusion morphism i : X → Pn . We know from lemma 7.4.15 that the cotangent sheaf of Pn ﬁts into an exact sequence of vector bundles 0 → ΩPn → O (−1)⊕(n+1) → O → 0. Pulling this sequence back by i and taking global sections, we see that we have an exact sequence 0 → Γ(i∗ ΩPn ) → Γ(OX (−1)⊕(n+1) ) → · · · . But OX (−1) has no global sections, so we conclude that i∗ ΩPn has no global sections either. Now consider the exact sequence of lemma 7.4.16 0 → OX (−d) → i∗ ΩPn → ΩX → 0, from which we deduce the exact sequence 0 → Γ(OX (−d)) → Γ(i∗ ΩPn ) → Γ(ΩX ). We have just seen that the ﬁrst two groups in this sequence are trivial. But Γ(ΩX ) is not trivial in general (e.g. for a cubic curve in P2 we have ΩX = OX and thus Γ(ΩX ) = k). Hence the last map in the above sequence of global sections cannot be surjective in general. We have however already met a case in which the induced map on global sections is ˜ exact: if X = Spec R is an afﬁne scheme and Fi = Mi for some R-modules Mi are quasicoherent sheaves on X then by lemma 7.2.7 (ii) the sequence 0 → F1 → F2 → F3 → 0

150

Andreas Gathmann

is exact if and only if the sequence 0 → Γ(F1 ) → Γ(F2 ) → Γ(F3 ) → 0 is exact (note that Γ(Fi ) = Mi by proposition 7.2.2 (ii)). We have mentioned already that essentially all sheaves occurring in practice are quasi-coherent, so we will assume this from now on for the rest of this chapter. The conclusion is that we know that taking global sections is an exact functor if the underlying scheme is afﬁne. The goal of the theory of cohomology is to extend the global section sequence to the right for all schemes X in the following sense: for any (quasi-coherent) sheaf F on X we will deﬁne natural cohomology groups H i (X, F ) for all i > 0 satisfying (among other things) the following property: given any exact sequence 0 → F1 → F2 → F3 → 0 of sheaves on X, there is an induced long exact sequence of cohomology groups

0 → Γ(F1 ) → Γ(F2 ) → Γ(F3 ) → H 1 (X, F1 ) → H 1 (X, F2 ) → H 1 (X, F3 ) → H 2 (X, F1 ) → · · · .

If X is an afﬁne scheme then H i (X, F ) = 0 for all i > 0, so that we arrive again at our old result that the sequence of global sections is exact in this case. Let us now give the deﬁnition of these cohomology groups. There are various ways to ˇ deﬁne these groups. In these notes we will use the approach of so-called Cech cohomology. This is the most suitable approach for actual applications (but maybe not the best one from ˇ a purely theoretical point of view). The idea of Cech cohomology is simple: we have seen above that the global section functor is exact (i.e. does what we ﬁnally want) if X is an afﬁne scheme. So if X is any scheme we will just choose an afﬁne open cover {Ui } of X and consider sections of our sheaves on these afﬁne open subsets and their intersections. Deﬁnition 8.1.2. Let X be a scheme, and let F be a (quasi-coherent) sheaf on X. Fix an afﬁne open cover {Ui }i∈I of X, and assume for simplicity that I is an ordered set. For all p ≥ 0 we deﬁne the Abelian group C p (F ) =

i0 <···<i p

∏

F (Ui0 ∩ · · · ∩Ui p ).

In other words, an element α ∈ CP (F ) is a collection α = (αi0 ,...,i p ) of sections of F over all intersections of p + 1 sets taken from the cover. These sections can be totally unrelated. For every p ≥ 0 we deﬁne a “boundary operator” d p : C p (F ) → C p+1 (F ) by (d p α)i0 ,...,i p+1 =

p+1 k=0

∑ (−1)k αi0 ,...,ik−1 ,ik+1 ,...,i p+1 |Ui0 ∩···∩Ui p+1 .

Note that this makes sense as the αi0 ,...,ik−1 ,ik+1 ,i p+1 are sections of F on Ui0 ∩ · · · ∩Uik−1 ∩ Uik+1 ∩ · · · ∩Ui p+1 , which contains Ui0 ∩ · · · ∩Ui p+1 as an open subset. By abuse of notation we will denote all these operators simply by d if it is clear from the context on which C p (F ) they act. Lemma 8.1.3. Let F be a sheaf on a scheme X. Then d p+1 ◦ d p : C p (F ) → C p+2 (F ) is the zero map for all p ≥ 0.

8.

Cohomology of sheaves

151

Proof. This statement is essentially due to the sign in the deﬁnition of dα: for every α ∈ C p (F ) we have (d p+1 d p α)i0 ,...,i p+2 = =

p+2

k=0 p+2 k−1

∑ (−1)k (dα)i0 ,...,ik−1 ,ik+1 ,...,i p+2

k=0 m=0 p+2 p+2

∑ ∑ (−1)k+m αi0 ,...,im−1 ,im+1 ,...,ik−1 ,ik+1 ,...,i p+2

+

k=0 m=k+1

∑ ∑

(−1)k+m−1 αi0 ,...,ik−1 ,ik+1 ,...,im−1 ,im+1 ,...,i p+2

=0 (omitting the restriction maps). We have thus deﬁned a sequence of Abelian groups and homomorphisms C0 (F ) −→ C1 (F ) −→ C2 (F ) −→ · · · such that d p+1 ◦ d p = 0 at every step. Such a sequence is usually called a complex of Abelian groups. The maps d p are then called the boundary operators. Deﬁnition 8.1.4. Let F be a sheaf on a scheme X. Pick an afﬁne open cover {Ui } of X and consider the associated groups C p (F ) and homomorphisms d p : C p (F ) → C p+1 (F ) for p ≥ 0. We deﬁne the p-th cohomology group of F to be H p (X, F ) = ker d p / im d p−1 with the convention that C p (F ) and d p are zero for p < 0. Note that this is well-deﬁned as im d p−1 ⊂ ker d p by lemma 8.1.3. If X is a scheme over a ﬁeld k then the cohomology groups will be vector spaces over k. The dimension of the cohomology groups H i (X, F ) as a k-vector space is then denoted hi (X, F ). Remark 8.1.5. The deﬁnition of the cohomology groups as it stands depends on the choice of the afﬁne open cover of X. It is a very crucial (and non-trivial) fact that the H i (X, F ) actually do not depend on this choice (as we have already indicated by the notation). It is ˇ the main disadvantage of our Cech approach to cohomology that this independence is not obvious from the deﬁnition. There are other constructions of the cohomology groups (for example the “derived functor approach” of [H] chapter III) that never use such afﬁne open covers and therefore do not face this problem. On the other hand, these other approaches ˇ are essentially useless for actual computations. This is why we have given the Cech approach here. We will prove the independence of our cohomology groups of the open cover in section 8.5. For now we will just assume this independence and rather discuss the properties and applications of the cohomology groups. Example 8.1.6. The following examples follow immediately from the deﬁnition and the assumption of remark 8.1.5: (i) For any X and F we have H 0 (X, F ) = Γ(F ). In fact, we have H 0 (X, F ) = ker(d 0 : C0 (F ) → C1 (F )) by deﬁnition. But an element α ∈ C0 (F ) is just given by a section αi ∈ F (Ui ) for every element of the open cover, and the map d 0 is given by (αi − α j )|Ui ∩U j . By the sheaf axiom this is zero for all i and j if and only if the αi come from a global section of F . Hence H 0 (X, F ) = Γ(F ). (In particular, our deﬁnition of h0 (L ) in section 7.7 is consistent with our current deﬁnition of h0 (X, L ).) (ii) If X is an afﬁne scheme then H i (X, F ) = 0 for i > 0. In fact, if X is afﬁne we can pick the open cover consisting of the single element X, in which case the groups Ci (F ) and hence the H i (X, F ) are trivially zero for i > 0.

d0 d1 d2

152

Andreas Gathmann

(iii) If X is a projective scheme of dimension n then H i (X, F ) = 0 whenever i > n. In fact, by proposition 4.1.9 we can pick homogeneous polynomials f0 , . . . , fn / such that X ∩ Z( f0 , . . . , fn ) = 0. We thus get an open cover of X by the n + 1 subsets X\Z( fi ) which are all afﬁne by proposition 5.5.4. Using this open cover for the deﬁnition of the cohomology groups, we see that the Ci (F ) and hence the H i (X, F ) are zero for i > n. Note that the same is true for any scheme that can be covered by n + 1 afﬁne open subsets. Note that for (i) we did not need the independence of the cohomology groups of the open cover, but for (ii) and (iii) we did. In fact, the last two statements are both highly non-trivial theorems about cohomology groups. They only follow so easily in our setup because we assumed the independence of the cover. Example 8.1.7. Let X = P1 and F = O . By example 8.1.6 (i) we know that H 0 (P1 , O ) ∼ k = is simply the space of (constant) global regular functions, and by part (iii) we know that H i (P1 , O ) = 0 for i > 1. So let us determine H 1 (P1 , O ). To compute this cohomology group let us pick the obvious afﬁne open cover Ui = {xi = 0} for i = 0, 1. Then C1 (O ) = O (U0 ∩U1 ) = = f

a b x0 x1 m n x0 x1 a b x0 x1

; f homogeneous of degree a + b ; m + n = a + b and m, n, a, b ≥ 0 .

Of course the condition m + n = a + b implies that we always have m ≥ a or n ≥ b. So every such generator is regular on at least one of the open subsets U0 and U1 . It follows that every such generator is in the image of the boundary map d 0 : C0 (O ) = O (U0 ) × O (U1 ) → O (U0 ∩U1 ), (α0 , α1 ) → α1 − α0 |U0 ∩U1 .

Consequently H 1 (P1 , O ) = 0 by deﬁnition of the cohomology groups. Example 8.1.8. In the same way as in example 8.1.7 let us now compute the cohomology group H 1 (P1 , O (−2)). With the same notations as above we have now C1 (O (−2)) = O (−2)(U0 ∩U1 ) = = f ; f homogeneous of degree a + b − 2 a b x0 x1 m n x0 x1 ; m+n = a+b−2 . a b x0 x1

The condition m + n = a + b − 2 implies that m ≥ a − 1 or n ≥ b − 1. If one of these n xm x1 inequalities is strict, then the corresponding generator x0 xb is regular on U0 or U1 and is a therefore zero in the cohomology group H 1 (P1 , O (−2)) as above. So we are only left with the function x01x1 where neither inequality is strict. As C2 (O (−2)) = 0 and so the boundary operator d 1 is trivial, we conclude that H 1 (P1 , O (−2)) is one-dimensional, with the function x01x1 as a generator. 8.2. The long exact cohomology sequence. The main property of the cohomology groups is that they solve the problem of ﬁnding an exact sequence of sections associated to a short exact sequence of sheaves:

0 1

8.

Cohomology of sheaves

153

Proposition 8.2.1. Let 0 → F1 → F2 → F3 → 0 be an exact sequence of sheaves on a (separated) scheme X. Then there is a canonical long exact sequence of cohomology groups 0 → H 0 (X, F1 ) → H 0 (X, F2 ) → H 0 (X, F3 ) → H 1 (X, F1 ) → H 1 (X, F2 ) → H 1 (X, F3 ) → H 2 (X, F1 ) → · · · . Proof. Consider the diagram of Abelian groups and homomorphisms ··· / C p−1 (F1 ) / C p (F1 )

d f d

··· / C p−1 (F2 ) / C p (F2 )

d g d

··· / C p−1 (F3 ) / C p (F3 )

d d

0

f

g

/ 0

0

f

g

/ 0

0

/ C p+1 (F1 )

d

/ C p+1 (F2 )

d

/ C p+1 (F3 )

d

/ 0

···

···

···

The columns of this diagram are complexes (i.e. d ◦ d = 0 at all places) by lemma 8.1.3. We claim that the rows of this diagram are all exact: by lemma 7.2.7 (ii) and what we have said in section 8.1 we know that the sequences 0 → F1 (U) → F2 (U) → F3 (U) → 0 are exact on every afﬁne open subset U of X. But the intersection of two (and hence ﬁnitely many) afﬁne open subsets of X is again afﬁne as U ∩V = ∆X ∩ (U ×V ) is a closed subset of an afﬁne scheme U ×V (where ∆X ⊂ X × X denotes the diagonal of X). As the C p (Fi ) are made up from sections on such open subsets, the claim follows. Moreover, note that all squares in this diagram are commutative by construction. The statement of the proposition now follows from a basic lemma of homological algebra: Lemma 8.2.2. Any short exact sequence of complexes ··· / C p−1 / Cp

d f

··· / D p−1 / Dp

d g

··· / E p−1 / Ep

d

0

f

g

/ 0

0

/ 0

d f

d g

d

0

/ C p+1 ···

d

/ D p+1 ···

d

/ E p+1 ···

d

/ 0

154

Andreas Gathmann

(i.e. the C p , D p , E p are Abelian groups, the diagram commutes, the rows are exact and the columns are complexes) gives rise to a long exact sequence in cohomology · · · → H p−1 (E) → H p (C) → H p (D) → H p (E) → H p+1 (C) → · · · where H p (C) = ker(C p → C p+1 )/ im(C p−1 → C p ), and similarly for D and E. Proof. The proof is done by pure “diagram chasing”. We will give some examples. (i) Existence of the morphisms ψ : H p (C) → H p (D): let α ∈ H p (C) be represented by an element in C p (which we denote by the same letter by abuse of notation). Then dα = 0 ∈ C p+1 . Set ψ(α) = f (α). Note that dψ(α) = f (dα) = 0, so ψ(α) is a well-deﬁned cohomology element. We still have to check that this deﬁnition does not depend on the representative chosen in C p . So if α = dα for some α ∈ C p−1 (so that α = 0 in H p (C)) then ψ(α) = f (dα ) = d f (α ) (so that ψ(α) = 0 in H p (D)). (ii) The existence of the morphisms H p (D) → H p (E) follows in the same way: they are simply induced by the morphisms g. (iii) Existence of the morphisms φ : H p (E) → H p+1 (C): The existence of these “connecting morphisms” is probably the most unexpected part of this lemma. Let α be a (representative of a) cohomology element in E p , so that dα = 0. As g : D p → E p is surjective, we can pick a β ∈ D p such that g(β) = α. Consider the element dβ ∈ D p+1 . We have g(dβ) = dg(β) = dα = 0, so dβ is in fact of the form f (γ) for a (unique) γ ∈ C p+1 . We set φ(α) = γ. We have to check that this is well-deﬁned: (a) dγ = 0 (so that γ actually deﬁnes an element in cohomology): we have f (dγ) = d f (γ) = d(dβ) = 0 as the middle column is a complex, so dγ = 0 as the f are injective. (b) The construction does not depend on the choice of β: if we pick another β with g(β ) = α then g(β − β ) = 0, so β − β = f (δ) for some δ ∈ C p as the p-th row is exact. Now if γ and γ are the elements such that f (γ) = dβ and f (γ ) = β then f (γ − γ ) = d(β − β ) = d f (δ) = f (dδ). As f is injective we conclude that γ − γ = dδ, so γ and γ deﬁne the same element in H p+1 (C). (c) If α = dα for some α ∈ E p−1 (so that α deﬁnes the zero element in cohomology) then we can pick an inverse image β with g(β ) = α as g is surjective. For β we can then take dβ . But then dβ = d(dβ ) = 0 as the middle column is a complex, so the resulting element in H p+1 (C) is zero. Summarizing, we can say that the morphism H p (E) → H p+1 (C) is obtained by going “left, down, left” in our diagram. We have just checked that this really gives rise to a well-deﬁned map. We have now seen that there is a canonical sequence of morphisms between the cohomology groups as stated in the lemma. It remains to be shown that the sequence is actually exact. We will check exactness at the H p (D) stage only (i.e. show that ker(H p (D) → H p (E)) = im(H p (C) → H p (D)) and leave the other two checks (at H p (C) and H p (E)) that are completely analogous as an exercise. im(H p (C) → H p (D)) ⊂ ker(H p (D) → H p (E)): Let α ∈ H p (D) be of the form α = f (β) for some β ∈ H p (C)). Then g(α) = g( f (β)) = 0 as the p-th row is exact. ker(H p (D) → H p (E)) ⊂ im(H p (C) → H p (D)): Let α ∈ H p (D) be a cohomology element (i.e. dα = 0) such that g(α) = 0 in cohomology, i.e. g(α) = dβ for some E p−1 . As g is surjective we can pick an inverse image γ ∈ D p−1 of β. Then g(α − dγ) = g(α) − g(dγ) = g(α) − dβ = 0,

8.

Cohomology of sheaves

155

so there is a δ ∈ C p such that f (δ) = α − dγ as the p-th row is exact. Note that δ deﬁnes an element in H p (C) as f (dδ) = d(α − dγ) = 0 − 0 = 0 and thus dδ = 0 as f is injective. Moreover, f (δ) = α in H p (D) by construction, so α ∈ im(H p (C) → H p (D)). Example 8.2.3. Consider the exact sequence of sheaves on X = P1

0 1 0 −→ O (−2) −→ O −→ kP ⊕ kQ −→ 0

·x x

from example 7.1.18, where P = (0 : 1) and Q = (1 : 0), and the last map is given by evaluation at P and Q. From proposition 8.2.1 we deduce an associated long exact cohomology sequence 0 → H 0 (X, O (−2)) → H 0 (X, O ) → H 0 (X, kP ⊕ kQ ) → H 1 (X, O (−2)) → H 1 (X, O ) → · · · . Now H 0 (X, O (−2)) = 0 by example 7.7.1 and H 1 (X, O ) = 0 by example 8.1.7. Moreover, H 0 (X, O ) is just the space of global (constant) functions, H 0 (X, kP ⊕ kQ ) is isomorphic to k × k (given by specifying values at the points P and Q), and H 1 (X, O (−2)) = x01x1 is 1-dimensional by example 8.1.8. So our exact sequence is just 0 → k → k × k → k → 0. We can actually also identify the morphisms. The ﬁrst morphism in this sequence is a → (a, a) as it is the evaluation of the constant function a at the points P and Q. The second morphism is given by the “left, down, left” procedure of part (iii) of the proof of lemma 8.2.2 in the following diagram: 0 / C0 (O (−2)) / C1 (O (−2)) / C0 (O ) / C1 (O ) / C0 (kP ⊕ kQ ) / C1 (kP ⊕ kQ ) / 0

0

/ 0

Starting with any element (a, b) ∈ C0 (kP ⊕ kQ ) we can ﬁnd an inverse image in C0 (O ) = O (U0 ) × O (U1 ) (with Ui = {xi = 0}, namely the pair of constant functions (b, a) (as P ∈ U1 and Q ∈ U0 ). Going down in the diagram yields the function a − b ∈ O (U0 ∩ U1 ) by the deﬁnition of the boundary operator. Recalling that the morphism from O (−2) to O is given a−b by multiplication with x0 x1 , we ﬁnd that x0 x1 is the element in C1 (O (−2)) that we were looking for. In terms of the basis vector x01x1 of H 1 (X, O (−2)) this function has the single coordinate a − b. So in this basis our exact cohomology sequence becomes 0 → k → k×k → k → 0 a → (a, a) (a, b) → a − b,

which is indeed exact. 8.3. The Riemann-Roch theorem revisited. Let us now study the cohomology groups of line bundles on smooth projective curves in some more detail. So let X be such a curve, and let L be a line bundle on X. Of course by example 8.1.6 (i) and (iii) the only interesting cohomology group is H 1 (X, L ). We will show that this group is trivial if L is “positive enough”: Proposition 8.3.1. (Kodaira vanishing theorem for line bundles on curves) Let X be a smooth projective curve of genus g, and let L be a line bundle on X such that deg L ≥ 2g − 1. Then H 1 (X, L ) = 0. Proof. We compute H 1 (X, L ) using our deﬁnition of cohomology groups. So let U0 ⊂ X be an afﬁne open subset of X. It must be of the form X\{P1 , . . . , Pr } for some points Pi on X. Now pick any other afﬁne open subset U1 ⊂ X that contains the points Pi . Then U1

156

Andreas Gathmann

is of the form X\{Q1 , . . . , Qs } with Pi = Q j for all i, j. So we have an afﬁne open cover X = U0 ∪U1 . By deﬁnition we have H 1 (X, L ) = L (U0 ∩U1 )/(L (U0 ) + L (U1 )). Note that L (U0 ∩U1 ) is precisely the space of rational sections of L that may have poles at the points Pi and Q j , and similarly for L (U0 ) and L (U1 ). In other words, to prove the proposition we have to show that any rational section α of L with poles at the Pi and Q j can be written as the sum of two rational sections α0 and α1 , where α0 has poles only at the Pi and α1 only at the Q j . So let α be such a rational section. It is a global section of L ⊗ OX (a1 P1 + · · · + ar Pr + b1 Q1 + · · · + bs Qs ) for some ai , b j ≥ 0. Let us assume that a1 ≥ 1. Note that then the degree of the line bundle ωX ⊗ L ∨ ⊗ OX (−a1 P1 − · · · − ar Pr ) is at most −2 by assumption and corollary 7.6.6. Hence by the Riemann-Roch theorem 7.7.3 (and example 7.7.1) it follows that h0 (L ⊗ OX (a1 P1 + · · · + ar Pr )) = deg L + a1 + · · · + ar + 1 − g. In the same way we get h0 (L ⊗ OX ((a1 − 1)P1 + · · · + ar Pr )) = deg L + a1 − 1 + a2 + · · · + ar + 1 − g. We conclude that h0 (L ⊗ OX (a1 P1 + · · · + ar Pr )) − h0 (L ⊗ OX ((a1 − 1)P1 + · · · + ar Pr )) = 1. So we can pick a rational section α0 in Γ(L ⊗ OX (a1 P1 +· · ·+ar Pr )) that is not in Γ(h0 (L ⊗ OX ((a1 − 1)P1 + · · · + ar Pr ))), i.e. a section that has a pole of order exactly a1 at P1 . Now α and α0 are both sections of the one-dimensional vector space Γ(L ⊗ OX (a1 P1 + · · · + ar Pr ))/Γ(L ⊗ OX ((a1 − 1)P1 + · · · + ar Pr )), and moreover α0 is not zero in this quotient. So by possibly multiplying α0 with a constant scalar we can assume that α − α0 is in Γ(L ⊗ OX ((a1 − 1)P1 + · · · + ar Pr )). Note now that α0 has poles only at the Pi , whereas the total order of the poles of α − α0 at the Pi is at most a1 + · · · + ar − 1. Repeating this process we arrive after a1 + · · · + ar steps at a rational section α0 with poles only at the Pi such that α1 := α − α0 has no poles any more at the Pi . This is precisely what we had to construct. Remark 8.3.2. As in the case of the Riemann-Roch theorem there are vast generalizations of the Kodaira vanishing theorem, e.g. to higher-dimensional spaces. One version is the following: if X is a smooth projective variety then H i (X, ωX ⊗ OX (n)) = 0 for all i > 0 and n > 0. Note that in the case of a smooth curve this follows from our version of proposition 8.3.1, as deg(ωX ⊗ OX (n)) = 2g − 2 + 1 ≥ 2g − 1. In general cohomology groups “tend to be zero quite often”. There are many so-called vanishing theorems that assert that certain cohomology groups are zero under some conditions that can hopefully easily be checked. We will prove one more vanishing theorem in theorem 8.4.7 (ii). Of course, the advantage of vanishing cohomology groups is that they break up the long exact cohomology sequence of proposition 8.2.1 into smaller pieces. Using our Kodaira vanishing theorem we can now reprove the Riemann-Roch theorem in a “cohomological version”. In analogy to the notation of section 7.7 let us denote H 1 (X, OX (D)) also by H 1 (D) for any divisor D, and similarly for h1 (D). Corollary 8.3.3. (Riemann-Roch theorem for line bundles on curves, second version) Let X be a smooth projective curve of genus g. Then for any divisor D on X we have h0 (D) − h1 (D) = deg D + 1 − g.

8.

Cohomology of sheaves

157

Proof. From the exact skyscraper sequence 0 → OX (D) → OX (D + P) → kP → 0 for any point P ∈ X we get the long exact sequence in cohomology 0 → H 0 (D) → H 0 (D + P) → k → H 1 (D) → H 1 (D + P) → 0 by proposition 8.2.1. Taking dimensions, we conclude that χ(D + P) − χ(D) = 1, where χ(D) := h0 (D) − h1 (D). It follows by induction that we must have h0 (D) − h1 (D) = deg D + c for some constant c (that does not depend on D). But by our ﬁrst version of the RiemannRoch theorem 7.7.3 we have h0 (D) − h0 (KX − D) = deg D + 1 − g. So to determine the constant c we can pick a divisor D of degree at least 2g − 1: then h1 (D) vanishes by proposition 8.3.1 and h0 (KX − D) by example 7.7.1. So we conclude that c = 1 − g, as desired. Remark 8.3.4. Comparing our two versions of the Riemann-Roch theorem we see that we must have h0 (ωX ⊗ L ∨ ) = h1 (L ) for all line bundles L on a smooth projective curve X. In fact, this is just a special case of the Serre duality theorem that asserts that for any smooth n-dimensional variety X and any locally free sheaf F there are canonical isomorphisms ∼ H i (X, F ) = H n−i (X, ωX ⊗ F ∨ )∨ for all i = 0, . . . , n. Unfortunately, these isomorphisms cannot easily be written down. There are even more general versions for singular varieties X and more general sheaves F . We refer to [H] section III.7 for details. Note that our new version of the Riemann-Roch theorem can be used to deﬁne the genus of singular curves: Deﬁnition 8.3.5. Let X be a (possibly singular) curve. Then the genus of X is deﬁned to be the non-negative integer h1 (X, OX ). (This deﬁnition is consistent with our old ones as we can see by setting L = OX in corollary 8.3.3.) Let us investigate the geometric meaning of the genus of singular curves in two cases: Example 8.3.6. Let C1 , . . . ,Cn be smooth irreducible curves of genera g1 , . . . , gn , and de˜ ˜ note by C = C1 ∪ · · · ∪ Cn their disjoint union. Now pick r pairs of points Pi , Qi ∈ C that ˜ are all distinct, and denote by C the curve obtained from C by identifying every Pi with the corresponding Qi for i = 1, . . . , r. Curves obtained by this procedure are called nodal curves. To compute the genus of the nodal curve C we consider the exact sequence 0 → OC → ⊕n OCi → ⊕r kPi → 0 i=1 i=1 where the last maps ⊕n OCi → kPi are given by evaluation at Pi minus evaluation at Qi . i=1 The sequence just describes the fact that regular functions on C are precisely functions on ˜ C that have the same value at Pi and Qi for all i. By proposition 8.2.1 we obtain a long exact cohomology sequence 0 → H 0 (C, OC ) → ⊕n H 0 (Ci , OCi ) → k⊕r → H 1 (C, OC ) → ⊕n H 1 (Ci , OCi ) → 0. i=1 i=1 Taking dimensions, we get 1 − n + r − h1 (C, OC ) + ∑i gi = 0, so we see that the genus of C is ∑i gi + r + 1 − n. If C is connected, note that r + 1 − n is precisely the number of “loops” in the graph of C. So the genus of a nodal curve is the sum of the genera of its components plus the number of “loops”. This ﬁts well with our topological interpretation of the genus given in examples 0.1.2 and 0.1.3.

158

Andreas Gathmann

P1 C1

Q1 C2

C1

C2 C3

C1

C genus = g 1 + g 2 genus = g 1 + g 2 + g 3 +1 genus = g 1 +1

Proposition 8.3.7. Let X ⊂ P2 be a (possibly singular) curve of degree d, given as the zero locus of a homogeneous polynomial f of degree d. Then the genus of X is equal to 1 2 (d − 1)(d − 2). Proof. Let x0 , x1 , x2 be the coordinates of P2 . By a change of coordinates we can assume that the point (0 : 0 : 1) is not on X. Then the afﬁne open subsets U0 = {x0 = 0} and U1 = {x1 = 0} cover X. So in the same way as in the proof of proposition 8.3.1 we get H 1 (X, OX ) = OX (U0 ∩U1 )/(OX (U0 ) + OX (U1 )).

d Moreover, the equation of f must contain an x2 -term, so the relation f = 0 can be used to restrict the degrees in x2 of functions on X to at most d − 1. Hence we get

OX (U0 ∩U1 ) =

and

i x2 k x0 x1 j

i x2 k x0 x1 j

; 0 ≤ i ≤ d − 1 and i = j + k

OX (U0 ) =

; 0 ≤ i ≤ d − 1, k ≤ 0, and i = j + k

(and similarly for OX (U1 )). We conclude that H 1 (X, OX ) =

i x2 k x0 x1 j

; 0 ≤ i ≤ d − 1, j > 0, k > 0, and i = j + k .

To compute the dimension of this space note that for a given value of i (which can run from 0 to d − 1) we get i − 1 choices of j and k (namely (1, i − 1), (2, i − 2), . . . , (i − 1, 1)). So 1 the total dimension is h1 (X, OX ) = 1 + 2 + · · · + (d − 2) = 2 (d − 1)(d − 2). Remark 8.3.8. The important point of proposition 8.3.7 is that the genus of a curve is constant in families: if we degenerate a smooth curve into a singular one (by varying the coefﬁcients in its equation) then the genus of the singular curve will be the same as the genus of the original smooth curve. This also ﬁts well with our idea in examples 0.1.2 and 0.1.3 that we can compute the genus of a plane curve by degenerating it into a singular one, where the result is then easy to read off. Remark 8.3.9. Our second (cohomological) version of the Riemann-Roch theorem is in fact the one that is needed for generalizations to higher-dimensional varieties. If X is an n-dimensional projective variety and F a sheaf on X then the generalized Riemann-Roch theorem mentioned in remark 7.7.7 (v) will compute the Euler characteristic χ(X, F ) := ∑ (−1)i hi (X, F )

i=0 n

8.

Cohomology of sheaves

159

in terms of other data that are usually easier to determine than the cohomology groups themselves. 8.4. The cohomology of line bundles on projective spaces. Let us now turn to higherdimensional varieties and compute the cohomology groups of the line bundles OX (d) on the projective space X = Pn . Proposition 8.4.1. Let X = Pn , and denote by S = k[x0 , . . . , xn ] the graded coordinate ring of X. Then the sheaf cohomology groups of the line bundles OX (d) on X are given by: (i) H 0 (X, OX (d)) = S as graded k-algebras, Ld∈Z n ∼ (ii) d∈Z H (X, OX (d)) = S as graded k-vector spaces, where S = S with the grading given by Sd = S−n−1−d . (iii) H i (X, OX (d)) = 0 whenever i = 0 and i = n.

L

Remark 8.4.2. By splitting up the equations of (i) and (ii) into the graded pieces one obtains the individual cohomology groups H i (X, OX (d)). So for example we have hn (X, OX (d)) = h0 (X, OX (−n − 1 − d)) =

−d−1 n

0

if d ≤ −n − 1, if d > −n − 1.

(Note that the equality of these two dimensions is consistent with the Serre duality theorem of remark 8.3.4, since ωX = OX (−n − 1) by lemma 7.4.15.) Proof. (i) is clear from example 8.1.6 (i). (ii): Let {Ui } for 0 ≤ i ≤ n be the standard afﬁne open cover of X, i.e. Ui = {xi = 0}. We will prove the proposition for all d together by computing the cohomology of the quasiL coherent graded sheaf FX = d∈Z OX (d) while keeping track of the grading (note that cohomology commutes with direct sums). This is just a notational simpliﬁcation. Of course we have Ui0 ,...,ik = {xi0 · · · · · xik = 0}. So F (Ui0 ,...,ik ) is just the localization Sxi0 ···xik . It follows that the sequence of groups Ck (FX ) reads

∏ Sxi0 → ∏ Sxi0 xi1 → · · · → ∏ Sx0 ···x j−1 x j+1 ···xn → Sx0 ···xn .

i0 i0 <i1 j

(∗)

Looking at the last term in this sequence, we compute that H n (X, F ) = coker(∏ Sx0 ···x j−1 x j+1 ···xn → Sx0 ···xn )

j

= = =

j j x00 · · · xnn j j x00 · · · xnn

j ; ji ∈ Z / x00 · · · xnn ; ji ≥ 0 for some i

j

; ji < 0 for all i

1 −1 k[x−1 , . . . , xn ], x0 · · · xn 0

so up to a shift of deg x0 · · · xn = n + 1 these are just the polynomials in xi with non-positive exponents. This shows (ii). (iii): We prove this by induction on n. There is nothing to show for n = 1. Let H = {xn = 0} ∼ Pn−1 be a hyperplane. Note that there is an exact sequence of sheaves on X = 0 → OX (d − 1) → OX (d) → OH (d) → 0 for all d, where the ﬁrst map is given by multiplication with xn , and the second one by setting xn to 0. Taking these sequences together for all d ∈ Z we obtain the exact sequence

n 0 → F (−1) → F → FH → 0

·x

160

Andreas Gathmann

where we set F (−1) = F ⊗ OX (−1). From the associated long exact cohomology sequence and the induction hypothesis we get the following exact sequences:

0 → H 0 (X, F (−1)) → H 0 (X, F ) → H 0 (H, FH ) → H 1 (X, F (−1)) → H 1 (X, F ) → 0, 0 → H i (X, F (−1)) → H i (X, F ) → 0 0→H

n−1

for 1 < i < n − 1,

(X, F (−1)) → H

n−1

(X, F ) → H

n−1

(H, FH ) → H n (X, F (−1)) → H n (X, F ) → 0.

So ﬁrst of all we see that H i (X, F (−1)) ∼ H i (X, F ) for all 1 < i < n − 1. We claim that = this holds in fact for 1 ≤ i ≤ n − 1. To see this for i = 1 note that the ﬁrst exact sequence above starts with

n 0 → k[x0 , . . . , xn ] → k[x0 , . . . , xn ] → k[x0 , . . . , xn−1 ] → · · · ,

·x

which is obviously exact on the right, so it follows that H 1 (X, F (−1)) ∼ H 1 (X, F ). A = similar analysis of the third exact sequence above, using the explicit description of the proof of part (ii), shows that H n−1 (X, F (−1)) ∼ H n−1 (X, F ). So we see that the map = ·xn H i (X, F (−1)) → H i (X, F ) is an isomorphism for all 1 ≤ i ≤ n − 1. (Splitting this up into the graded parts, this means that H i (X, OX (d − 1)) ∼ H i (X, OX (d)) for all d, i.e. the = cohomology groups do not depend on d. We still have to show that they are in fact zero.) ˇ Now localize the Cech complex (∗) with respect to xn . Geometrically this just means that we arrive at the complex that computes the cohomology of F on Un = {xn = 0}. As Un is an afﬁne scheme and therefore does not have higher cohomology groups by example 8.1.6 (ii), we see that H i (X, F )xn = H i (Un , F |Un ) = 0.

k So for any α ∈ H i (X, F ) we know that xn · α = 0 for some k. But we have shown above that multiplication with xn in H i (X, F ) is an isomorphism, so α = 0. This means that H i (X, F ) = 0, as desired.

Example 8.4.3. As a consequence of this computation we can now of course compute the cohomology groups of all sheaves on Pn that are made up of line bundles in some way. Let us calculate the cohomology groups H i (X, ΩX ) as an example. By the Euler sequence of lemma 7.4.15 0 → ΩPn → O (−1)⊕(n+1) → O → 0 we get the long exact cohomology sequence 0 → H 0 (ΩPn ) → H 0 (O (−1))⊕(n+1) → H 0 (O ) → H 1 (ΩPn ) → H 1 (O (−1))⊕(n+1) → H 1 (O ) → H 2 (ΩPn ) → · · · . By proposition 8.4.1 the cohomology groups of O (−1) are all zero, while the cohomology groups H i (O ) are zero unless i = 0, in which case we have h0 (O ) = 1. So we conclude that hi (Pn , ΩPn ) = 1 0 if i = 1, otherwise.

As an application of our computation of the cohomology groups of line bundles on projective spaces, we now want to prove in the rest of this section that the cohomology groups of certain “ﬁnitely generated” quasi-coherent sheaves on projective schemes are always ﬁnite-dimensional. Let us ﬁrst deﬁne what we mean by this notion of ﬁnite generation. Deﬁnition 8.4.4. Let X be a scheme. A sheaf F on X is called coherent if for every afﬁne open subset U = Spec R ⊂ X the restricted sheaf F |U is the sheaf associated to a ﬁnitely generated R-module in the sense of deﬁnition 7.2.1.

8.

Cohomology of sheaves

161

Remark 8.4.5. Except for the ﬁnite generation condition this deﬁnition is precisely the same as for quasi-coherent sheaves. Consequently, our results that essentially all operations that one can do with quasi-coherent sheaves yield again quasi-coherent sheaves carry over to coherent sheaves without much change. To show that the cohomology groups of coherent sheaves on projective schemes are ﬁnite-dimensional we need an auxiliary lemma ﬁrst. Lemma 8.4.6. Let X be a projective scheme over a ﬁeld, and let F be a coherent sheaf on X. Then there is a surjective morphism OX (−d)⊕n → F for some d and n. Proof. Let X ⊂ Pr = Proj k[x0 , . . . , xr ] and consider the standard afﬁne open subsets Ui = ˜ Spec Ri ⊂ X given by xi = 0. As F is coherent, F |Ui is of the form Mi , where Mi is a ﬁnitely generated Ri -module. Let si,1 , . . . , si,ki be generators. Then the si, j deﬁne sections of F over Ui , and their germs generate the stalk of F at every point of Ui . The si, j do not need to extend to global sections of F , but we will now show that after d d multiplying with xi for some d we get global sections si, j · xi ∈ Γ(F ⊗ OX (d)). As X\Ui is covered by the afﬁne open subsets Uk for k = i, it is sufﬁcient to show that we can extend si, j to all Uk in this way. But F (Uk ) = Mk and F (Ui ∩Uk ) = (Mk )xi by proposition 7.2.2 (ii), so si, j ∈ F (Ui ∩ Uk ) ∈ (Mk )xi obviously gives an element in F (Uk ) = Mk after multiplying with a sufﬁciently high power of xi . Hence we have shown that for some d we get global sections si, j ∈ Γ(F ⊗ OX (d)) that generate the stalk of F ⊗ OX (d) at all points of X. So these sections deﬁne a surjective morphism O → F ⊗ OX (d)⊕n (where n is the total number of sections chosen) and hence a surjective morphism OX (−d)⊕n → F . Theorem 8.4.7. Let X be a projective scheme over a ﬁeld, and let F be a coherent sheaf on X. (i) The cohomology groups H i (X, F ) are ﬁnite-dimensional vector spaces for all i. (ii) We have H i (X, F ⊗ OX (d)) = 0 for all i > 0 and d 0. Proof. Let i : X → Pr be the inclusion morphism. As i∗ F is coherent by proposition 7.2.9 (or rather its analogue for coherent sheaves) and the cohomology groups of F and i∗ F agree by deﬁnition, we can assume that X = Pr . We will prove the proposition by descending induction on i. By example 8.1.6 (iii) there is nothing to show for i > r. By lemma 8.4.6 there is an exact sequence 0 → R → OX (−d)⊕n → F → 0 for some d and n, where R is a coherent sheaf on X by lemma 7.2.7. Tensoring with OX (e) for some e ∈ Z and taking the corresponding long exact cohomology sequence, we get · · · → H i (X, OX (e − d)⊕n ) → H i (X, F ⊗ OX (e)) → H i+1 (X, R ⊗ OX (e)) → · · · . (i): Take e = 0. Then the vector space on the left is always ﬁnite-dimensional by the explicit computation of proposition 8.4.1, and the one on the right is ﬁnite-dimensional by the induction hypothesis. Hence H i (X, F ) is ﬁnite-dimensional as well. (ii): For e 0 the group on the left is zero again by the explicit calculation of proposition 8.4.1, and the one on the right is zero by the induction hypothesis. Hence H i (X, F ⊗ OX (e)) = 0 for e 0. Remark 8.4.8. Of course the assumption of projectivity is essential in theorem 8.4.7, as for example H 0 (A1 , OA1 ) = k[x] is not ﬁnite-dimensional as a vector space over k.

162

Andreas Gathmann

For a more interesting example, consider X = A2 \{(0, 0)} and compute H 1 (X, OX ). Using the afﬁne open cover X = U1 ∪U2 with Ui = {xi = 0} for i = 1, 2, we get H 1 (X, OX ) = OX (U1 ∩U2 )/(OX (U1 ) + OX (U2 ))

i i = x1 x2 ; i, j ∈ Z / x1 x2 ; j ≥ 0 or i ≥ 0 i = x1 x2 ; i, j < 0 , j j j

which is not ﬁnite-dimensional. So we conclude that X is not projective (which is obvious). But we have also reproven the statement that X is not afﬁne (see remark 2.3.17), as otherwise we would have a contradiction to example 8.1.6 (ii). 8.5. Proof of the independence of the afﬁne cover. To make our discussion of sheaf cohomology rigorous it remains to be proven that the cohomology groups as of deﬁnition 8.1.4 do not depend on the choice of afﬁne open cover. So let us go back to the original deﬁnitions 8.1.2 and 8.1.4 that (seem to) depend on this choice. For simplicity let us assume that all afﬁne covers involved are ﬁnite. Lemma 8.5.1. Let F be a quasi-coherent sheaf on an afﬁne scheme X. Then H i (X, F ) = 0 for all i > 0 and every choice of afﬁne open cover {Ui }. ˇ Proof. Let us deﬁne a “sheaﬁﬁed version” of the Cech complex as follows: we set

C p (F ) =

i0 <···<i p

∏

i∗ F |Ui0 ∩···∩Ui p

where i : Ui0 ∩ · · · ∩ Ui p → X denotes the various inclusion maps. Then the C p (F ) are quasi-coherent sheaves on X by proposition 7.2.9. Their spaces of global sections are Γ(C p (F )) = C p (F ) by deﬁnition. There are boundary morphisms d p : C p (F ) → C p+1 (F ) deﬁned by the same formula as in deﬁnition 8.1.2, giving rise to a complex

C 0 (F ) → C 1 (F ) → C 2 (F ) → · · · .

(∗)

Note that it sufﬁces to prove that this sequence is exact: as taking global sections of quasicoherent sheaves on afﬁne schemes preserves exact sequences by proposition 7.2.2 (ii) it then follows that the sequence C0 (F ) → C1 (F ) → C2 (F ) → · · · is exact as well, which by deﬁnition means that H i (X, F ) = 0 for i > 0. The exactness of (∗) can be checked on the stalks. So let P ∈ X be any point, and let U j be an afﬁne open subset of the given cover that contains P. We deﬁne a morphism of stalks of sheaves at P k−1 k k : CP → CP , α → kα by (kα)i0 ,...,i p−1 = α j,i0 ,...,i p−1 , where we make the following convention: if the indices j, i0 , . . . , i p−1 are not in sorted order and σ ∈ S p+1 is the permutation such that σ( j) < σ(i0 ) < · · · < σ(i p−1 ) then by α j,i0 ,...,i p−1 we mean (−1)σ · ασ( j),σ(i0 ),...,σ(i p−1 ) .

k k We claim that kd + dk : CP → CP is the identity. In fact, we have

(dkα)i0 ,...,i p = αi0 ,...,i p − ∑ (−1)k α j,i0 ,...,ik−1 ,ik+1 ,...,i p

k=1

p

and (kdα)i0 ,...,i p = from which the claim follows.

k=1

∑ (−1)k α j,i0 ,...,ik−1 ,ik+1 ,...,i p

p

8.

Cohomology of sheaves

163

Finally we can now prove that the sequence (∗) is exact at any point P: we know already that im d k−1 ⊂ ker d k as d k ◦ d k−1 = 0. Conversely, if α ∈ ker d k , i.e. dα = 0, then α = (kd + dk)(α) = d(kα), i.e. α ∈ im d k−1 . Lemma 8.5.2. Let F be a quasi-coherent sheaf on a scheme X. Pick an afﬁne open cover ˜ U = {U1 , . . . ,Uk }. Let U0 ⊂ X be any other afﬁne open subset, and denote by U the afﬁne open cover {U0 , . . . ,Uk }. Then the cohomology groups determined by the open covers U ˜ and U are the same. ˇ Proof. Let C p (F ) and H p (X, F ) be the groups of Cech cycles and the cohomology groups ˜ p (F ) and H p (X, F ) the corresponding groups for the ˜ for the cover U , and denote by C ˜ cover U . ˜ ˜ Note that there are natural morphisms C p (F ) → C p (F ) and H p (X, F ) → H p (X, F ) given by “forgetting the data that involves the open subset U0 ”, i.e. by (αi0 ,...,i p )0≤i0 <i1 <···<i p ≤k → (αi0 ,...,i p )1≤i0 <i1 <···<i p ≤k . ˜ ˜ ˜ More precisely, an element α ∈ C p (F ) can be thought of as a pair α = (α, α0 ), where p (F ) is given by α 0 p−1 (U , F | ) is given by ˜ α ∈C i0 ,...,i p = αi0 ,...,i p (for i0 > 0) and α ∈ C 0 U0 ˜ ˜ α00 ,...,i p−1 = α0,i0 ,...,i p−1 . Moreover, d α = 0 if and only if i dα = 0 ˜ (these are the equations (d α)i0 ,...,i p+1 = 0 for i0 > 0) and α|U0 − dα0 = 0 (2) ˜ (these are the equations (d α)i0 ,...,i p+1 = 0 for i0 = 0). ˜ We have to show that the morphism H p (X, F ) → H p (X, F ) is injective and surjective. ˜ (i) H p (X, F ) → H p (X, F ) is surjective: Let α ∈ H p (X, F ) be a cohomology cycle, ˜ i.e. dα = 0. We have to ﬁnd an α0 ∈ C p−1 (U0 , F |U0 ) such that α = (α, α0 ) satis˜ ﬁes d α = 0, i.e. by (2) such that dα0 = α|U0 . But d(α|U0 ) = (dα)|U0 = 0, so by lemma 8.5.1 α|U0 = dα0 for some α0 . ˜ ˜ ˜ (ii) H p (X, F ) → H p (X, F ) is injective: Let α ∈ H p (X, F ) be a cohomology cycle ˜ (i.e. d α = 0) such that α = 0 ∈ H p (X, F ), i.e. α = dβ for some β ∈ C p−1 (F ). ˜ ˜ ˜ We have to show that α = 0 ∈ H p (X, F ), i.e. we have to ﬁnd a β = (β, β0 ) ∈ ˜ ˜ p−1 (F ) such that d β = α. By (2) this means that we need β|U − dβ0 = α0 . ˜ C 0 But d(β|U0 − α0 ) = α|U0 − α|U0 = 0, so by lemma 8.5.1 there is a β0 such that β|U0 − α0 = dβ0 . Corollary 8.5.3. The cohomology groups of quasi-coherent sheaves on any scheme do not depend on the choice of open afﬁne cover. Proof. Let F be a quasi-coherent sheaf on a scheme X. Let U = {U1 , . . . ,Uk } and U = {U1 , . . . ,Um } be two afﬁne open covers of X. Then the cohomology groups H i (X, F ) determined by U are the same as those determined by U ∪ U by (a repeated application of) lemma 8.5.2, which in turn are equal to those determined by U by the same lemma. 8.6. Exercises. Exercise 8.6.1. Let X be a smooth projective curve. For any point P ∈ X consider the exact skyscraper sequence of sheaves on X 0 → ωX → ωX ⊗ OX (P) → kP → 0 as in exercise 7.8.4. Show that the induced sequence of global sections is not exact, i.e. the last map Γ(ωX ⊗ OX (P)) → Γ(kP ) is not surjective. (1)

164

Andreas Gathmann

Exercise 8.6.2. Complete the proof of lemma 8.2.2, i.e. show that the sequence of morphisms of cohomology groups · · · → H p−1 (E) → H p (C) → H p (D) → H p (E) → H p+1 (C) → · · · associated to an exact sequence of complexes 0 → C → D → E → 0 is exact at the H p (C) and H p (E) positions. Exercise 8.6.3. Compute the cohomology groups H i (P1 × P1 , p∗ OP1 (a) ⊗ q∗ OP1 (b)) for all a, b ∈ Z, where p and q denote the two projection maps from P1 × P1 to P1 . Exercise 8.6.4. Give an example of a smooth projective curve X and line bundles L1 , L2 on X of the same degree such that h0 (X, L1 ) = h0 (X, L2 ). Exercise 8.6.5. Let X ⊂ Pr be a complete intersection of dimension n ≥ 1, i.e. it is the scheme-theoretic zero locus of r − n homogeneous polynomials. Show that X is connected. (Hint: Prove by induction on n that the natural map H 0 (Pr , OPr (d)) → H 0 (X, OX (d)) is surjective for all d ∈ Z.)

9.

Intersection theory

165

9. I NTERSECTION THEORY

A k-cycle on a scheme X (that is always assumed to be separated and of ﬁnite type over an algebraically closed ﬁeld in this section) is a ﬁnite formal linear combination ∑i ni [Vi ] with ni ∈ Z, where the Vi are k-dimensional subvarieties of X. The group of k-cycles is denoted Zk (X). A rational function ϕ on any subvariety Y ⊂ X of dimension k + 1 determines a cycle div(ϕ) ∈ Zk (X), which is just the zeroes of ϕ minus the poles of ϕ, counted with appropriate multiplicities. The subgroup Bk (X) ⊂ Zk (X) generated by cycles of this form is called the group of k-cycles that are rationally equivalent to zero. The quotient groups Ak (X) = Zk (X)/Bk (X) are the groups of cycle classes or Chow groups. They are the main objects of study in intersection theory. The Chow groups of a scheme should be thought of as being analogous to the homology groups of a topological space. A morphism f : X → Y is called proper if inverse images of compact sets (in the classical topology) are compact. Any proper morphism f gives rise to push-forward homomorphisms f∗ : A∗ (X) → A∗ (Y ) between the Chow groups. On the other hand, some other morphisms f : X → Y (e.g. inclusions of open subsets or projections from vector bundles) admit pull-back maps f ∗ : A∗ (Y ) → A∗ (X). If X is a purely n-dimensional scheme, a Weil divisor is an element of Zn−1 (X). ∗ ∗ In contrast, a Cartier divisor is a global section of the sheaf KX /OX . Every Cartier divisor determines a Weil divisor. On smooth schemes, Cartier and Weil divisors agree. On almost any scheme, Cartier divisors modulo linear equivalence correspond exactly to line bundles. We construct bilinear maps Pic X × Ak (X) → Ak−1 (X) that correspond geometrically to taking intersections of the divisor (a codimension-1 subset of X) with the k-dimensional subvariety. If one knows the Chow groups of a space and the above intersection products, one arrives at B´ zout style theorems that allow to compute the e number of intersection points of k divisors on X with a k-dimensional subspace.

9.1. Chow groups. Having discussed the basics of scheme theory, we will now start with the foundations of intersection theory. The idea of intersection theory is the same as that of homology in algebraic topology. Roughly speaking, what one does in algebraic topology is to take e.g. a real differentiable manifold X of dimension n and an integer k ≥ 0, and consider formal linear combinations of real k-dimensional submanifolds (with boundary) on X with integer coefﬁcients, called cycles. If Zk (X) is the group of closed cycles (those having no boundary) and Bk (X) ⊂ Zk (X) is the group of those cycles that are boundaries of (k +1)dimensional cycles, then the homology group Hk (X, Z) is the quotient Zk (X)/Bk (X). There are (at least) two main applications of this. First of all, the groups Hk (X, Z) are (in contrast to the Zk (X) and Bk (X)) often ﬁnitely generated groups and provide invariants of the manifold X that can be used for classiﬁcation purposes. Secondly, there are intersection products: homology classes in Hn−k (X, Z) and Hn−l (X, Z) can be “multiplied” to give a class in Hn−k−l (X, Z) that geometrically corresponds to taking intersections of submanifolds. Hence if we are for example given submanifolds Vi of X whose codimensions T sum up to n (so that we expect a ﬁnite number of points in the intersection i Vi ), then this number can often be computed easily by taking the corresponding products in homology. Our goal is to establish a similar theory for schemes. For any scheme of ﬁnite type over a ground ﬁeld and any integer k ≥ 0 we will deﬁne the so-called Chow groups Ak (X) whose elements are formal linear combinations of k-dimensional closed subvarieties of X, modulo “boundaries” in a suitable sense. The formal properties of these groups Ak (X) will be similar to those of homology groups; if the ground ﬁeld is C you might even want to think of the Ak (X) as being “something like” H2k (X, Z), although these groups are usually different. But there is always a map Ak (X) → H2k (X, Z) (at least if one uses the “right” homology theory, see [F] chapter 19 for details), so you can think of elements in the Chow

166

Andreas Gathmann

groups as something that determines a homology class, but this map is in general neither injective nor surjective. Another motivation for the Chow groups Ak (X) is that they generalize our notions of divisors and divisor classes. In fact, if X is a smooth projective curve then A0 (X) will be by deﬁnition the same as Pic X. In general, the deﬁnition of the groups Ak (X) is very similar to our deﬁnition of divisors: we consider the free Abelian groups Zk (X) generated by the k-dimensional subvarieties of X. There is then a subgroup Bk (X) ⊂ Zk (X) that corresponds to those linear combinations of subvarieties that are zeros minus poles of rational functions. The Chow groups are then the quotients Ak (X) = Zk (X)/Bk (X). To make sense of this deﬁnition, the ﬁrst thing we have to do is to deﬁne the divisor of a rational function (see deﬁnition 6.3.4) in the higher-dimensional case. This is essentially a problem of commutative algebra, so we will only sketch it here. The important ingredient is the notion of the length of a module. Remark 9.1.1. (For the following facts we refer to [AM] chapter 6 and [F] section A.1.) Let M be a ﬁnitely generated module over a Noetherian ring R. Then there is a so-called composition series, i.e. a ﬁnite chain of submodules 0 = M0 M1 ··· Mr = M (∗) ∼ such that Mi /Mi−1 = R/pi for some prime ideals pi ∈ R. The series is not unique, but for any prime ideal p ⊂ R the number of times p occurs among the pi does not depend on the series. The geometric meaning of this composition series is easiest explained in the case where R is an integral domain and M = R/I for some ideal I ⊂ R. In this case Spec M is a closed subscheme of the irreducible scheme Spec R (see examples 5.2.3 and 7.2.10). The prime ideals pi are then precisely the ideals of the irreducible (and maybe embedded) components of Spec M, or in other words the prime ideals associated to all primary ideals in the primary decomposition of I. The number of times p occurs among the pi can be thought of as the “multiplicity” of the corresponding component in the scheme. For example, if I is a radical ideal (so Spec M is reduced) then the pi are precisely the ideals of the irreducible components of Spec M, all occurring once. We will need this construction mainly in the case where I = ( f ) ⊂ R is the ideal generated by a single (non-zero) function. In this case all irreducible components of Spec M have codimension 1. If p ⊂ R is a prime ideal corresponding to any codimension-1 subvariety of Spec R we can consider a composition series as above for the localized module Mp over Rp . As the only prime ideals in Rp are (0) and pRp (corresponding geometrically to Spec R and Spec M, respectively) and f does not vanish identically on Spec M, the only prime ideal that can occur in the composition series of Mp is pRp . The number of times it occurs, i.e. the length r of the composition series, is then called the length of the module Mp over Rp , denoted lRp (Mp ). It is equal to the number of times p occurs in the composition series of M over R. By what we have said above, we can interpret this number geometrically as the multiplicity of the subvariety corresponding to p in the scheme Spec R/( f ), or in other words as the order of vanishing of f at this codimension-1 subvariety. We should rephrase these ideas in terms of general (not necessarily afﬁne) schemes. So let X be a scheme, and let V ⊂ X be a subvariety of codimension 1. Note that V can be considered as a point in the scheme X, so it makes sense to talk about the stalk OX,V of the structure sheaf OX at V . If U = Spec R ⊂ X is any afﬁne open subset with nonempty intersection with V then OX,V is just the localized ring Rp where p is the prime ideal corresponding to the subvariety V ∩U of U (see proposition 5.1.12 (i)). So if f ∈ OX,V is a local function around V then its order of vanishing at the codimension-1 subvariety V is simply the length lOX,V (OX,V /( f )). To deﬁne the order of a possibly rational function ϕ on X we just have to observe that the ﬁeld of fractions of the ring OX,V is equal to the ﬁeld of

9.

Intersection theory

167

f rational functions on X. So we can write ϕ as g for some f , g ∈ OX,V and simply deﬁne the order of ϕ at V to be the difference of the orders of f and g at V .

With these prerequisites we can now deﬁne the Chow groups in complete analogy to the Picard group of divisor classes in section 6.3. For the rest of this section by a scheme we will always mean a scheme of ﬁnite type over some algebraically closed ﬁeld (that is not necessarily smooth, irreducible, or reduced). A variety is a reduced and irreducible (but not necessarily smooth) scheme. Deﬁnition 9.1.2. Let X be a variety, and let V ⊂ X be a subvariety of codimension 1, and set R = OX,V . For every non-zero f ∈ R ⊂ K(X) we deﬁne the order of f at V to be the f integer ordV ( f ) := lR (R/( f )). If ϕ ∈ K(X) is a non-zero rational function we write ϕ = g with f , g ∈ R and deﬁne the order of ϕ at V to be ordV (ϕ) := ordV ( f ) − ordV (g). To show that this is well-deﬁned, i.e. that ordV the exact sequence

·b f g f g

= ordV

whenever f g = g f , one uses

0 → R/(a) → R/(ab) → R/(b) → 0 and the fact that the length of modules is additive on exact sequences. From this it also follows that the order function is a homomorphism of groups ordV : K(X)∗ := K(X)\{0} → Z. Example 9.1.3. Let X = A1 = Spec k[x] and let V = {0} ⊂ X be the origin. Consider the function ϕ = xr for r ≥ 0. Then R = OX,V = k[x](x) , and R/(x) ∼ k. So as R/(xr ) = {a0 + = a1 x + · · · + ar−1 xr−1 } has vector space dimension r over k we conclude that ord0 (xr ) = r, as expected. By deﬁnition, we then have the equality ord0 (xr ) = r for all r ∈ Z. Deﬁnition 9.1.4. Let X be a scheme. For k ≥ 0 denote by Zk (X) the free Abelian group generated by the k-dimensional subvarieties of X. In other words, the elements of Zk (X) are ﬁnite formal sums ∑i ni [Vi ], where ni ∈ Z and the Vi are k-dimensional (closed) subvarieties of X. The elements of Zk (X) are called cycles of dimension k. For any (k + 1)-dimensional subvariety W of X and any non-zero rational function ϕ on W we deﬁne a cycle of dimension k on X by div(ϕ) = ∑ ordV (ϕ)[V ] ∈ Zk (X),

V

called the divisor of ϕ, where the sum is taken over all codimension-1 subvarieties V of W . Note that this sum is always ﬁnite: it sufﬁces to check this on a ﬁnite afﬁne open cover {Ui } of W and for ϕ ∈ OUi (Ui ), where it is obvious as Z(ϕ) is closed and Ui is Noetherian. Let Bk (X) ⊂ Zk (X) be the subgroup generated by all cycles of the form div(ϕ) for all W ⊂ X and ϕ ∈ K(W )∗ as above. We deﬁne the group of k-dimensional cycle classes to be the quotient Ak (X) = Zk (X)/Bk (X). These groups are usually called the Chow groups of X. Two cycles in Zk (X) that determine the same element in Ak (X) are said to be rationally equivalent. L L We set Z∗ (X) = k≥0 Zk (X) and A∗ (X) = k≥0 Ak (X). Example 9.1.5. Let X be a scheme of pure dimension n. Then Bn (X) is trivially zero, and thus An (X) = Zn (X) is the free Abelian group generated by the irreducible components of X. In particular, if X is an n-dimensional variety then An (X) ∼ Z with [X] as a generator. = In the same way, Zk (X) and Ak (X) are trivially zero if k > n. Example 9.1.6. Let X be a smooth projective curve. Then Z0 (X) = Div X and A0 (X) = Pic X by deﬁnition. In fact, the 1-dimensional subvariety W of X in deﬁnition 9.1.4 can only be X itself, so we arrive at precisely the same deﬁnition as in section 6.3.

168

Andreas Gathmann

Example 9.1.7. Let X = {x1 x2 = 0} ⊂ P2 be the union of two projective lines X = X1 ∪ X2 that meet in a point. Then A1 (X) = Z[X1 ] ⊕ Z[X2 ] by example 9.1.5. Moreover, A0 (X) ∼ Z = is generated by the class of any point in X. In fact, any two points on X1 are rationally equivalent by example 9.1.6, and the same is true for X2 . As both X1 and X2 contain the intersection point X1 ∩ X2 we conclude that all points in X are rationally equivalent. So A0 (X) ∼ Z. = Now let P1 ∈ X1 \X2 and P2 ∈ X2 \X1 be two points. Note that the line bundles OX (P1 ) and OX (P2 ) (deﬁned in the obvious way: OX (Pi ) is the sheaf of rational functions that are regular away from Pi and have at most a simple pole at Pi ) are not isomorphic: if i : X1 → X is the inclusion map of the ﬁrst component, then i∗ OX (P1 ) ∼ OP1 (1), whereas = i∗ OX (P2 ) ∼ OP1 . So we see that for singular curves the one-to-one correspondence between = A0 (X) and line bundles no longer holds. Example 9.1.8. Let X = An . We claim that A0 (X) = 0. In fact, if P ∈ X is any point, pick a line W ∼ A1 ⊂ An through P and a linear function ϕ on W that vanishes precisely = at P. Then div(ϕ) = [P]. It follows that the class of any point is zero in A0 (X). Therefore A0 (X) = 0. Example 9.1.9. Now let X = Pn ; we claim that A0 (X) ∼ Z. In fact, if P and Q are any two = distinct points in X let W ∼ P1 ⊂ Pn be the line through P and Q, and let ϕ be a rational = function on W that has a simple zero at P and a simple pole at Q. Then div(ϕ) = [P] − [Q], i.e. the classes in A0 (X) of any two points in X are the same. It follows that A0 (X) is generated by the class [P] of any point in X. On the other hand, if W ⊂ X = Pn is any curve and ϕ a rational function on W then we have seen in remark 6.3.5 that the degree of the divisor of ϕ is always zero. It follows that the class n · [P] ∈ A0 (X) for n ∈ Z can only be zero if n = 0. We conclude that A0 (X) ∼ Z = with the class of any point as a generator. Example 9.1.10. Let X be a scheme, and let Y ⊂ X be a closed subscheme with inclusion morphism i : Y → X. Then there are canonical push-forward maps i∗ : Ak (Y ) → Ak (X) for any k, given by [Z] → [Z] for any k-dimensional subvariety Z ⊂ Y . It is obvious from the deﬁnitions that this respects rational equivalence. Example 9.1.11. Let X be a scheme, and let U ⊂ X be an open subset with inclusion morphism i : U → X. Then there are canonical pull-back maps i∗ : Ak (X) → Ak (U) for any k, given by [Z] → [Z ∩ U] for any k-dimensional subvariety Z ⊂ X. This respects rational equivalence as i∗ div(ϕ) = div(ϕ|U ) for any rational function ϕ on a subvariety of X. Remark 9.1.12. If f : X → Y is any morphism of schemes it is an important part of intersection theory to study whether there are push-forward maps f∗ : A∗ (X) → A∗ (Y ) or pull-back maps f ∗ : A∗ (Y ) → A∗ (X) and which properties they have. We have just seen two easy examples of this. Note that neither example can be reversed (at least not in an obvious way): (i) if Y ⊂ X is a closed subset, then a subvariety of X is in general not a subvariety of Y , so there is no pull-back morphism A∗ (X) → A∗ (Y ) sending [V ] to [V ] for any subvariety V ⊂ X. (ii) if U ⊂ X is an open subset, there are no push-forward maps A∗ (U) → A∗ (X): if U = A1 and X = P1 then the class of a point is zero in A∗ (U) but non-zero in A∗ (P1 ) by examples 9.1.8 and 9.1.9. We will construct more general examples of push-forward maps in section 9.2, and more general examples of pull-back maps in proposition 9.1.14.

9.

Intersection theory

169

Lemma 9.1.13. Let X be a scheme, let Y ⊂ X be a closed subset, and let U = X\Y . Denote the inclusion maps by i : Y → X and j : U → X. Then the sequence

∗ Ak (Y ) → Ak (X) → Ak (U) → 0

i

j∗

is exact for all k ≥ 0. The homomorphism i∗ is in general not injective however. Proof. This follows more or less from the deﬁnitions. If Z ⊂ U is any k-dimensional ¯ ¯ subvariety then the closure Z of Z in X is a k-dimensional subvariety of X with j∗ [Z] = [Z]. ∗ is surjective. So j If Z ⊂ Y then Z ∩ U = 0, so j∗ ◦ i∗ = 0. Conversely, assume that we have a cycle ∑ ar [Vr ] ∈ Ak (X) whose image in Ak (U) is zero. This means that there are rational functions ϕs on (k + 1)-dimensional subvarieties Ws of U such that ∑ div(ϕs ) = ∑ ar [Vr ∩ U] on U. Now the ϕs are also rational functions on the closures of Ws in X, and as such their divisors can only differ from the old ones by subvarieties Vr that are contained in X\U = Y . We conclude that ∑ div(ϕs ) = ∑ ar [Vr ] − ∑ br [Vr ] on X for some br . So ∑ ar [Vr ] = i∗ ∑ br [Vr ]. As an example that i∗ is in general not injective let Y be a smooth cubic curve in X = P2 . If P and Q are two distinct points on Y then [P] − [Q] = 0 ∈ A0 (Y ) = Pic X by proposition 6.3.13, but [P] − [Q] = 0 ∈ A0 (X) ∼ Z by example 9.1.9. = Proposition 9.1.14. Let X be a scheme, and let π : E → X be a vector bundle of rank r on X (see remark 7.3.2). Then for all k ≥ 0 there is a well-deﬁned surjective pull-back homomorphism π∗ : Ak (X) → Ak+r (E) given on cycles by π∗ [V ] = [π−1 (V )]. Proof. It is clear that π∗ is well-deﬁned: it obviously maps k-dimensional cycles to (k + r)dimensional cycles, and π∗ div(ϕ) = div(π∗ ϕ) for any rational function ϕ on a (k + 1)dimensional subvariety of X. We will prove the surjectivity by induction on dim X. Let U ⊂ X be an afﬁne open subset over which E is of the form U × Ar , and let Y = X\U. By lemma 9.1.13 there is a commutative diagram Ak (Y ) Ak+r (E|Y ) / Ak (X)

π∗

/ Ak (U) / Ak+r (U × Ar )

/ 0

/ Ak+r (E)

/ 0

with exact rows. A diagram chase (similar to that of the proof of lemma 8.2.2) shows that in order for π∗ to be surjective it sufﬁces to prove that the left and right vertical arrows are surjective. But the left vertical arrow is surjective by the induction assumption since dimY < dim X. So we only have to show that the right vertical arrow is surjective. In other words, we have reduced to the case where X = Spec R is afﬁne and E = X × Ar is the trivial bundle. As π then factors as a sequence E = X × Ar → X × Ar−1 → · · · → X × A1 → X we can furthermore assume that r = 1, so that E = X × A1 = Spec R[t]. We have to show that π∗ : Ak (X) → Ak (X × A1 ) is surjective. So let V ⊂ X × A1 be a (k + 1)-dimensional subvariety, and let W = π(V ). There are now two cases to consider: • dimW = k. Then V = W × A1 , so [V ] = π∗ [W ]. • dimW = k + 1. As it sufﬁces to show that [V ] is in the image of the pull-back map Ak (W ) → Ak+r (W × A1 ) we can assume that W = X. Consider the ideal I(V ) ⊗R K ⊂ K[t], where K = K(W ) denotes the quotient ﬁeld of R. It is not the unit ideal as otherwise we would be in case (i). On the other hand K[t] is a principal ideal domain, so I(V )⊗R K is generated by a single polynomial ϕ ∈ K[t].

170

Andreas Gathmann

Considering ϕ as a rational function on X × A1 we see that the divisor of ϕ is precisely [V ] by construction, plus maybe terms of the form ∑ ai π∗ [Wi ] for some Wi ⊂ X corresponding to our tensoring with the ﬁeld of rational functions K(X). So [V ] = π∗ (∑ ai [Wi ]) (plus the divisor of a rational function), i.e. [V ] is in the image of π∗ .

Remark 9.1.15. Note that the surjectivity part of proposition 9.1.14 is obviously false on the cycle level, i.e. for the pull-back maps Zk (X) → Zk (E): not every subvariety of a vector bundle E over X is the inverse image of a subvariety in X. So this proposition is an example of the fact that working with Chow groups (instead of with the subvarieties themselves) often makes life a little easier. In fact one can show (see [F] theorem 3.3 (a)) that the pull-back maps π∗ : Ak (X) → Ak+r (E) are always isomorphisms. Corollary 9.1.16. The Chow groups of afﬁne spaces are given by Ak (An ) = Z for k = n, 0 otherwise.

Proof. The statement for k ≥ n follows from example 9.1.5. For k < n note that the homomorphism A0 (An−k ) → Ak (An ) is surjective by proposition 9.1.14, so the statement of the corollary follows from example 9.1.8. ∼ Corollary 9.1.17. The Chow groups of projective spaces are Ak (Pn ) = Z for all 0 ≤ k ≤ n, with an isomorphism given by [V ] → degV for all k-dimensional subvarieties V ⊂ Pn . Proof. The statement for k ≥ n follows again from example 9.1.5, so let us assume that k < n. We prove the statement by induction on n. By lemma 9.1.13 there is an exact sequence Ak (Pn−1 ) → Ak (Pn ) → Ak (An ) → 0. n ) = 0 by corollary 9.1.16, so we conclude that A (Pn−1 ) → A (Pn ) is surWe have Ak (A k k jective. By the induction hypothesis this means that Ak (Pn ) is generated by the class of a k-dimensional linear subspace. As the morphism Zk (Pn−1 ) → Zk (Pn ) trivially preserves degrees it only remains to be shown that any cycle ∑ ai [Vi ] that is zero in Ak (Pn ) must satisfy ∑ ai degVi = 0. But this is clear from B´ zouts theorem, as deg div(ϕ) = 0 for all e rational functions on any subvariety of Pn . Remark 9.1.18. There is a generalization of corollary 9.1.17 as follows. Let X be a scheme / that has a stratiﬁcation by afﬁne spaces, i.e. X has a ﬁltration by closed subschemes 0 = X−1 ⊂ X0 ⊂ · · · ⊂ Xn = X such that Xk \Xk−1 is a disjoint union of ak afﬁne spaces Ak for all k. For example, X = Pn has such a stratiﬁcation with ak = 1 for 0 ≤ k ≤ n, namely / 0 ⊂ P0 ⊂ P1 ⊂ · · · ⊂ Pn = X. We claim that then Ak (X) is isomorphic to Zak modulo some (possibly trivial) subgroup, where Zak is generated by the classes of the closures of the ak copies of Ak mentioned above. We will prove this by induction on dim X, the case of dimension 0 being obvious. In fact, consider the exact sequence of lemma 9.1.13 Ak (Xn−1 ) → Ak (X) → ⊕an Ak (An ) → 0. i=1 / Note that Xn−1 itself is a scheme with a ﬁltration 0 = X−1 ⊂ X0 ⊂ · · · ⊂ Xn−1 as above. So it follows that: (i) For k < n we have Ak (An ) = 0, so Ak (X) is generated by Ak (Xn−1 ). Hence the claim follows from the induction hypothesis. (ii) For k ≥ n we have Ak (Xn−1 ) = 0, so An (X) ∼ ⊕an Ak (An ) is generated by the = i=1 classes of the closures of the an copies of An in X\Xn−1 .

9.

Intersection theory

171

This proves the claim. In fact, one can show that Ak (X) is always isomorphic to Zak if X has a stratiﬁcation by afﬁne spaces as above (see [F] example 1.9.1). In particular, this construction can be applied to compute the Chow groups of products of projective spaces and Grassmannian varieties (see exercise 3.5.4). Remark 9.1.19. Using Chow groups, B´ zout’s theorem can obviously be restated as fole lows: we have seen in corollary 9.1.17 that Ak (Pn ) ∼ Z for all k ≤ n, with the class of a = k-dimensional linear subspace as a generator. Using this isomorphism, deﬁne a product map An−k (Pn ) × An−l (Pn ) → An−k−l (Pn ), (a, b) → ab for k + l ≤ n. This “intersection pairing” has the following property: if X,Y ⊂ Pn are two subvarieties that intersect in the expected dimension (i.e. codim(X ∩Y ) = codim X + codimY ) then [X ∩ Y ] = [X] · [Y ]. So “intersections of subvarieties can be performed on the level of cycle classes”. As we have mentioned in the introduction to this section, the existence of such intersection pairing maps between the Chow groups will generalize to arbitrary smooth varieties. It is one of the most important properties of the Chow groups. 9.2. Proper push-forward of cycles. We now want to generalize the push-forward maps of example 9.1.10 to more general morphisms, i.e. given a morphism f : X → Y of schemes we will study the question under which conditions there are induced push-forward maps f∗ : Ak (X) → Ak (Y ) for all k that are (roughly) given by f∗ [V ] = [ f (V )] for a k-dimensional subvariety V of X. Remark 9.2.1. We have seen already in remark 9.1.12 (ii) that there are no such pushforward maps for the open inclusion A1 → P1 . The reason for this is precisely that the point P = P1 \A1 is “missing” in the domain of the morphism: a rational function on A1 (which is then also a rational function on P1 ) may have a zero and / or pole at the point P which is then present on P1 but not on A1 . As the class of P is not trivial in the Chow group of P1 , this will change the rational equivalence class. Therefore there is no well-deﬁned push-forward map between the Chow groups. Another example of a morphism for which there is no push-forward for Chow groups is the trivial morphism f : A1 → pt: again the class of a point is trivial in A0 (A1 ) but not in A0 (pt). In contrast, the morphism f : P1 → pt admits a well-deﬁned push-forward map f∗ : A0 (P1 ) ∼ Z → A0 (pt) ∼ Z sending the class of a point in P1 to the class of a point in = = pt. These counterexamples can be generalized by saying that in general there should be no points “missing” in the domain of the morphism f : X → Y for which we are looking for a push-forward f∗ . For example, if Y is the one-pointed space, by “no points missing” we mean exactly that X should be compact (in the classical topology), i.e. complete in the sense of remark 3.4.5. For general Y we need a “relative version” of this compactness (resp. completeness) condition. Morphisms satisfying this condition are called proper. We will give both the topological deﬁnition (corresponding to “compactness”) and the algebraic deﬁnition (corresponding to “completeness”). Deﬁnition 9.2.2. (Topological deﬁnition:) A continuous map f : X → Y of topological spaces is called proper if f −1 (Z) is compact for every compact set Z ⊂ Y . (Algebraic deﬁnition:) Let f : X → Y be a morphism of “nice” schemes (separated, of ﬁnite type over a ﬁeld). For every morphism g : Z → Y from a third scheme Z form the ﬁber diagram / X X ×Y Z Z

f g

/ Y.

f

172

Andreas Gathmann

The morphism f is said to be proper if the induced morphism f is closed for every such g : Z → Y , i.e. if f maps closed subsets of X ×Y Z to closed subsets of Z. This condition is sometimes expressed by saying that f is required to be “universally closed”. Remark 9.2.3. Note that the two deﬁnitions look quite different: whereas the topological deﬁnition places a condition on inverse images of (compact) subsets by some morphism, the algebraic deﬁnition places a condition on images of (closed) subsets by some morphism. Yet one can show that for varieties over the complex numbers the two deﬁnitions agree if we apply the topological deﬁnition to the classical (not the Zariski) topology. We will only illustrate this by some examples below. Note however that both deﬁnitions are “obvious” generalizations of their absolute versions, i.e. properness of a map in topology is a straightforward generalization of compactness of a space, whereas properness of a morphism in algebraic geometry is the expected generalization of completeness of a variety (see remark 3.4.5). In particular, if Y = pt is a point then the (trivial) morphism f : X → pt is proper if and only if X is complete (resp. compact). Example 9.2.4. If X is complete (resp. compact) then any morphism f : X → Y is proper. We will prove this both in the topological and the algebraic setting: (i) In topology, let Z ⊂ Y be a compact subset of Y . In particular Z is closed, hence so is the inverse image f −1 (Z) as f is continuous. It follows that f −1 (Z) is a closed subset of a compact space X, hence compact. (ii) In algebra, the ﬁber product X ×Y Z in deﬁnition 9.2.2 is isomorphic to the closed subscheme p−1 (∆Y ) of X × Z, where p = ( f , g) : X × Z → Y ×Y and ∆Y ⊂ Y ×Y is the diagonal. So if V ⊂ X ×Y Z is any closed subset, then V is also closed in X × Z, and hence its image in Z is closed as X is complete. This is the easiest criterion to determine that a morphism is proper. Some more can be found in exercise 9.5.5. Example 9.2.5. Let U ⊂ X be a non-empty open subset of a (connected) scheme X. Then the inclusion morphism i : U → X is not proper. This is obvious for the algebraic deﬁnition, as i is not even closed itself (it maps the closed subset U ⊂ U to the non-closed subset U ⊂ X). In the topological deﬁnition, let Z ⊂ X be a small closed disc around a point P ∈ X\U. Its inverse image i−1 (Z) = Z ∩U is Z minus a closed non-empty subset, so it is not compact. Example 9.2.6. If f : X → Y is proper then every ﬁber f −1 (P) is complete (resp. compact). Again this is obvious for the topological deﬁnition, as {P} ⊂ Y is compact. In the algebraic deﬁnition let P ∈ Y be a point, let Z be any scheme, and form the ﬁber diagram Z × f −1 (P) Z

f

/ f −1 (P) / P

/ X

f

/ Y.

If f is proper then by deﬁnition the morphism f is closed for all choices of P and Z. By deﬁnition this means exactly that all ﬁbers f −1 (P) of f are complete. The converse is not true however: every ﬁber of the morphism A1 → P1 is complete (resp. compact), but the morphism is not proper. Remark 9.2.7. It turns out that the condition of properness of a morphism f : X → Y is enough to guarantee the existence of well-deﬁned push-forward maps f∗ : Ak (X) → Ak (Y ). To construct them rigorously however we have to elaborate further on our idea that f∗ should map any k-dimensional cycle [V ] to [ f (V )], as the following two complications can occur:

9.

Intersection theory

173

(i) The image f (V ) of V may have dimension smaller than k, so that f (V ) does not deﬁne a k-dimensional cycle. It turns out that we can consistently deﬁne f∗ [V ] to be zero in this case. (ii) It may happen that dim f (V ) = dimV and the morphism f is a multiple covering map, i.e. that a general point in f (V ) has d > 1 inverse image points. In this case the image f (V ) is “covered d times by V ”, so we would expect that we have to set f∗ [V ] = d · [ f (V )]. Let us deﬁne this “order of the covering” d rigorously: Proposition 9.2.8. Let f : X → Y be a morphism of varieties of the same dimension such that f (X) is dense in Y . Then: (i) K(X) is a ﬁnite-dimensional vector space over K(Y ). Its dimension is called the degree of the morphism f , denoted deg f . (One also says that K(X) : K(Y ) is a ﬁeld extension of dimension [K(X) : K(Y )] = deg f .) (ii) The degree of f is equal to the number of points in a general ﬁber of f . (This means: there is a non-empty open set U ⊂ Y such that the ﬁbers of f over U consist of exactly deg f points.) (iii) If moreover f is proper then every zero-dimensional ﬁber of f consists of exactly deg f points if the points are counted with their scheme-theoretic multiplicities. Proof. (i): We begin with a few reduction steps. As the ﬁelds of rational functions do not change when we pass to an open subset, we can assume that X ⊂ An and Y ⊂ Am are afﬁne. Next, we factor the morphism f : X → Y as f = π ◦ γ with γ : X → Γ ⊂ X × Y the graph morphism and π : X ×Y → Y the projection. As γ is an isomorphism it is sufﬁcient to show the statement of the proposition for the projection map π. Finally, we can factor the projection π (which is the restriction of the obvious projection map An+m → Am to X × Y ) into n projections that are given by dropping one coordinate at a time. Hence we can assume that X ⊂ An+1 and Y ⊂ An , and prove the statement for the map π : X → Y that is the restriction of the projection map (x0 , . . . , xn ) → (x1 , . . . , xn ) to X. In this case the ﬁeld K(X) is generated over K(Y ) by the single element x0 . Assume that x0 ∈ K(X) is transcendental over K(Y ), i.e. there is no polynomial relation of the form

d−1 d Fd x0 + Fd−1 x0 · · · + F0 = 0,

(∗)

for Fi ∈ K(Y ) and Fd = 0. Then for every choice of (x1 , . . . , xn ) ∈ Y the value of x0 in X is not restricted, i.e. the general ﬁber of f is not ﬁnite. But then dim X > dimY in contradiction to our assumption. So x0 ∈ K(X) is algebraic over K(Y ), i.e. there is a relation d−1 (∗). It follows that K(X) is a vector space over K(Y ) with basis {1, x0 , . . . , x0 }. (ii): Continuing the proof of (i), note that on the non-empty open subset of Y where all Fi are regular and Fd is non-zero every point in the target has exactly d inverse image points (counted with multiplicity). Restricting the open subset further to the open subset where the discriminant of the polynomial (∗) is non-zero, we can in fact show that there is an open subset of Y on which the inverse images of f consist set-theoretically of exactly d points that all count with multiplicity 1. (iii): We will only sketch this part, using the topological deﬁnition of properness. By (ii) there is an open subset U ⊂ Y on which all ﬁbers of f consist of exactly n points. Let P ∈ Y be any point, and choose a small closed disc ∆ ⊂ U ∪ {P} around P. If ∆ is small enough then the inverse image f −1 (∆\{P}) will be a union of d copies of ∆\{P}. As f is proper, the inverse image f −1 (∆) has to be compact, i.e. all the holes in the d copies of ∆\{P} have to be ﬁlled in by inverse image points of P. So the ﬁber f −1 (P) must contain at least d points (counted with multiplicities). But we see from (∗) above that every ﬁber contains at most d points unless it is inﬁnite (i.e. all Fi are zero at P). This shows part (iii).

174

Andreas Gathmann

We are now ready to construct the push-forward maps f∗ : Ak (X) → Ak (Y ) for proper morphisms f : X → Y . Construction 9.2.9. Let f : X → Y be a proper morphism of schemes. Then for every subvariety Z ⊂ X the image f (Z) is a closed subvariety of dimension at most dim Z. On the cycle level we deﬁne homomorphisms f∗ : Zk (X) → Zk (Y ) by f∗ [Z] = [K(Z) : K( f (Z))] · [ f (Z)] if dim f (Z) = dim Z, 0 if dim f (Z) < dim Z.

By proposition 9.2.8 this is well-deﬁned and corresponds to the ideas mentioned in remark 9.2.7. Remark 9.2.10. By the multiplicativity of degrees of ﬁeld extensions it follows that the push-forwards are functorial, i.e. (g ◦ f )∗ = g∗ f∗ for any two morphisms f : X → Y and g : Y → Z. Of course we want to show that these homomorphisms pass to the Chow groups, i.e. give rise to well-deﬁned homomorphisms f∗ : Ak (X) → Ak (Y ). For this we have to show by deﬁnition that divisors of rational functions are pushed forward to divisors of rational functions. Theorem 9.2.11. Let f : X → Y be a proper surjective morphism of varieties, and let ϕ ∈ K(X)∗ be a non-zero rational function on X. Then f∗ div(ϕ) = 0 if dimY < dim X div(N(ϕ)) if dimY = dim X

in Z∗ (Y ), where N(ϕ) ∈ K(Y ) denotes the determinant of the endomorphism of the K(Y )vector space K(X) given by multiplication by ϕ (this is usually called the norm of ϕ). Proof. The complete proof of the theorem with all algebraic details is beyond the scope of these notes; it can be found in [F] proposition 1.4. We will only sketch the idea of the proof here. Case 1: dimY < dim X (see the picture below). We can assume that dimY = dim X − 1, as otherwise the statement is trivial for dimensional reasons. Note that we must have f∗ div(ϕ) = n · [Y ] for some n ∈ Z by example 9.1.5. So it only remains to determine the number n. By our interpretation of remark 9.2.7 (ii) we can compute this number on a general ﬁber of f by counting all points in this ﬁber with the multiplicity with which they occur in the restriction of ϕ to this ﬁber. In other words, we have n = ∑P: f (P)=Q ordP (ϕ| f −1 (Q) ) for any point Q ∈ Y over which the ﬁber of f is ﬁnite. But this number is precisely the degree of ϕ| f −1 (Q) on the complete curve f −1 (Q), which must be zero. (Strictly speaking we have only shown this for smooth projective curves in remark 6.3.5, but it is true in the general case as well. The important ingredient is here that the ﬁber is complete.) Case 2: dimY = dim X (see the picture below). We will restrict ourselves here to showing the stated equation set-theoretically, i.e. we will assume that ϕ is (locally around a ﬁber) a regular function and show that f (Z(ϕ)) = Z(N(ϕ)), where Z(·) denotes as usual the zero locus of a function. Note ﬁrst that we can neglect the ﬁbers of f that are not ﬁnite: these ﬁbers can only lie over a subset of Y of codimension at least 2 (otherwise the non-zero-dimensional ﬁbers would form a component of X for dimensional reasons, in contrast to X being irreducible). So as f∗ div(ϕ) is a cycle of codimension 1 in Y these higher-dimensional ﬁbers cannot contribute to the push-forward.

9.

Intersection theory

175

ord φ =2 ord φ =−1

2 −1 −1

P1 X P2 P3

X

n=2−1−1=0 f Y Q Case 1

f Y Q Case 2

Now let Q ∈ Y be any point such that the ﬁber f −1 (Q) is ﬁnite. Then f −1 (Q) consists of exactly d = [K(X) : K(Y )] points by proposition 9.2.8 (iii). Let us assume for simplicity that all these points are distinct (although this is not essential), so f −1 (Q) = {P1 , . . . , Pd }. The space of functions on this ﬁber is then just kd , corresponding to the value at the d points. In this basis, the restriction of the function ϕ to this ﬁber is then obviously given by the diagonal matrix with entries ϕ(P1 ), . . . , ϕ(Pd ), so its determinant is N(ϕ)(Q) = ∏d ϕ(Pi ). i=1 Now it is clear that Q ∈ f (Z(ϕ)) ⇐⇒ there is a Pi over Q with ϕ(Pi ) = 0 ⇐⇒ Q ∈ Z(N(ϕ)). We can actually see the multiplicities arising as well: if there are k points among the Pi where ϕ vanishes, then the diagonal matrix ϕ| f 1 (Q) contains k zeros on the diagonal, hence its determinant is a product that contains k zeros, so it should give rise to a zero of order k, in accordance with our interpretation of remark 9.2.7 (ii). Corollary 9.2.12. Let f : X → Y be a proper morphism of schemes. Then there are welldeﬁned push-forward maps f∗ : Ak (X) → Ak (Y ) for all k ≥ 0 given by the deﬁnition of construction 9.2.9. Proof. This follows immediately from theorem 9.2.11 applied to the morphism from a (k + 1)-dimensional subvariety of X to its image in Y . Example 9.2.13. Let X be a complete scheme, and let f : X → pt be the natural (proper) map. For any 0-dimensional cycle class α ∈ A0 (X) we deﬁne the degree of α to be the integer f∗ α ∈ A0 (pt) ∼ Z. This is well-deﬁned by corollary 9.2.12. More explicitly, if = α = ∑i ni [Pi ] for some points Pi ∈ X then deg α = ∑i ni . ˜ Example 9.2.14. Let X = P2 be the blow-up of P2 with coordinates (x0 : x1 : x2 ) in the point P = (1 : 0 : 0), and denote by E ⊂ X the exceptional hypersurface. In this example we will compute the Chow groups of X using remark 9.1.18. Note that P2 has a stratiﬁcation by afﬁne spaces as A2 ∪ A1 ∪ A0 . Identifying A0 with ˜ P and recalling that the blow-up P2 is obtained from P2 by “replacing the point P with a line P1 ” we see that X has a stratiﬁcation A2 ∪ A1 ∪ A1 ∪ A0 . By remark 9.1.18 it follows that the closures of these four strata generate A∗ (X). More precisely, these four classes are [X] ∈ A2 (X), [L] ∈ A1 (X) where L is the strict transform of a line in P2 through P, the exceptional hypersurface [E] ∈ A1 (X), and the class of a point in A0 (X). It follows immediately that A2 (X) ∼ Z and A0 (X) ∼ Z. Moreover we see that A1 (X) is generated by = = [L] and [E]. We have already stated without proof in remark 9.1.18 that [L] and [E] form in fact a basis of A1 (X). Let us now prove this in our special case at hand. So assume that there is a relation n[L] + m[E] = 0 in A1 (X). Consider the following two morphisms:

176

Andreas Gathmann

(i) Let π : X → P2 be the projection to the base of the blow-up. This is a proper map, and we have π∗ [L] = [H] and π∗ [E] = 0 where [H] ∈ A1 (P2 ) is the class of a line. So we see that 0 = π∗ (0) = π∗ (n[L] + m[E]) = n[H] ∈ A1 (P2 ), from which we conclude that n = 0. (ii) Now let p : X → P1 be the morphism that is the identity on E, and sends every point Q ∈ X\E to the unique intersection point of E with the strict transform of the line through P and Q. Again this is a proper map, and we have p∗ [L] = 0 and p∗ [E] = [P1 ]. So again we see that 0 = p∗ (0) = p∗ (n[L] + m[E]) = m[P1 ] ∈ A1 (P1 ), from which we conclude that m = 0 as well. Combining both parts we see that there is no non-trivial relation of the form n[L]+m[E] = 0 in A1 (X). Now let [H] be the class of a line in X that does not intersect the exceptional hypersurface. We have just shown that [H] must be a linear combination of [L] and [E]. To compute x1 which one it is, consider the rational function x0 on X. It has simple zeros along L and E, and a simple pole along H (with coordinates for L and H chosen appropriately). So we conclude that [H] = [L] + [E] in A1 (X). 9.3. Weil and Cartier divisors. Our next goal is to describe intersections on the level of Chow groups as motivated in the beginning of section 9.1. We will start with the easiest case, namely with the intersection of a variety with a subset of codimension 1. To put it more precisely, given a subvariety V ⊂ X of dimension k and another one D ⊂ X of codimension 1, we want to construct an intersection cycle [V ] · [D] ∈ Ak−1 (X) with the property that [V ] · [D] = [V ∩ D] if this intersection V ∩ D actually has dimension k − 1. Of course these intersection cycles should be well-deﬁned on the Chow groups, i.e. the product cycle [V ] · [D] ∈ Ak−1 (X) should only depend on the classes of V and D in A∗ (X). Example 9.3.1. Here is an example showing that this is too much to hope for in the generality as we stated it. Let X = P2 ∪P1 P2 be the union of two projective planes glued together along a common line. Let L1 , L2 , L3 ⊂ X be the lines as in the following picture.

L3 L1 P Q L2

Their classes in A1 (X) are all the same since A1 (X) ∼ Z by remark 9.1.18. But note that = L1 ∩ L2 is empty, whereas L1 ∩ L3 is a single point P. But 0 = [P] ∈ A0 (X), so there can be no well-deﬁned product map A1 (X) × A1 (X) → A0 (X) that describes intersections on this space X. The reason why this construction failed is quite a subtle one: we have to distinguish between codimension-1 subspaces and spaces that can locally be written as the zero locus of a single function. In general the intersection product exists only for intersections with spaces that are locally the zero locus of a single function. For most spaces this is the same thing as codimension-1 subspaces, but notably not in example 9.3.1 above: neither of the three lines Li can be written as the zero locus of a single function on X: there is a

9.

Intersection theory

177

(linear) function on the vertical P2 that vanishes precisely on L1 , but we cannot extend it to a function on all of X that vanishes at the point Q but nowhere else on the horizontal P2 . (We can write the Li as the zero locus of a single function on a component of X, but this is not what we need.) So for intersection-theoretic purposes we have to make a clear distinction between codimension-1 subspaces and spaces that are locally the zero locus of a single function. Let us make the corresponding deﬁnitions. Deﬁnition 9.3.2. Let X be a scheme. (i) If X has pure dimension n a Weil divisor on X is an element of Zn−1 (X). Obviously, the Weil divisors form an Abelian group. Two Weil divisors are called linearly equivalent if they deﬁne the same class in An−1 (X). The quotient group An−1 (X) is called the group of Weil divisor classes. ∗ (ii) Let KX be the sheaf of rational functions on X, and denote by KX the subsheaf of invertible elements (i.e. of those functions that are not identically zero on any ∗ component of X). Note that KX is a sheaf of Abelian groups, with the group ∗ structure given by multiplication of rational functions. Similarly, let OX be the sheaf of invertible elements of OX (i.e. of the regular functions that are nowhere ∗ zero). Note that OX is a sheaf of Abelian groups under multiplication as well. In ∗ is a subsheaf of K ∗ . fact, OX X ∗ ∗ A Cartier divisor on X is a global section of the sheaf KX /OX . Obviously, the Cartier divisors form an Abelian group under multiplication, denoted Div X. In analogy to Weil divisors the group structure on Div X is usually written additively however. A Cartier divisor is called linearly equivalent to zero if it is induced ∗ by a global section of KX . Two Cartier divisors are linearly equivalent if their difference (i.e. quotient, see above) is linearly equivalent to zero. The quotient ∗ ∗ ∗ group Pic X := Γ(KX /OX )/Γ(KX ) is called the group of Cartier divisor classes. Remark 9.3.3. Let us analyze the deﬁnition of Cartier divisors. There is an obvious exact sequence of sheaves on X

∗ ∗ ∗ ∗ 0 → OX → KX → KX /OX → 0.

Note that these are not sheaves of OX -modules, so their ﬂavor is slightly different from the ones we have considered so far. But it is still true that we get an exact sequence of global sections ∗ ∗ ∗ ∗ 0 → Γ(OX ) → Γ(KX ) → Γ(KX /OX ) that is in general not exact on the right. More precisely, recall that the quotient sheaf ∗ ∗ ∗ ∗ KX /OX is not just the sheaf that is KX (U)/OX (U) for all open subsets U ⊂ X, but rather the ∗ ∗ sheaf associated to this presheaf. Therefore Γ(KX /OX ) is in general not just the quotient ∗ )/Γ(O ∗ ). Γ(KX X ∗ ∗ To unwind the deﬁnition of sheaﬁﬁcation, an element of Div X = Γ(KX /OX ) can be ∗ (U )/O ∗ (U ) repregiven by a (sufﬁciently ﬁne) open covering {Ui } and elements of KX i i X ϕ ∗ sented by rational functions ϕi for all i such that their quotients ϕ ij are in OX (Ui ∩U j ) for all i, j. So a Cartier divisor is an object that is locally a (non-zero) rational function modulo a nowhere-zero regular function. Intuitively speaking, the only data left from a rational function if we mod out locally by nowhere-zero regular functions is the locus of its zeros and poles together with their multiplicities. So one can think of Cartier divisors as objects that are (linear combinations of) zero loci of functions. A Cartier divisor is linearly equivalent to zero if it is globally a rational function, just the same as for Weil divisors. From cohomology one would expect that one can think of the ∗ quotient group Pic X as the cohomology group H 1 (X, OX ). We cannot say this rigorously ∗ because we have only deﬁned cohomology for quasi-coherent sheaves (which OX is not).

178

Andreas Gathmann

But there is a more general theory of cohomology of arbitrary sheaves of Abelian groups ∗ on schemes, and in this theory the statement that Pic X = H 1 (X, OX ) is correct. Lemma 9.3.4. Let X be a purely n-dimensional scheme. Then there is a natural homomorphism Div X → Zn−1 (X) that passes to linear equivalence to give a homomorphism Pic X → An−1 (X). In other words, every Cartier divisor (class) determines a Weil divisor (class). Proof. Let D ∈ Div X be a Cartier divisor on X, represented by an open covering {Ui } of X and rational functions ϕi on Ui . For any (n − 1)-dimensional subvariety V of X deﬁne the order of D at V to be ordV D := ordV ∩Ui ϕi , where i is an index such that Ui ∩ V = ϕ / 0. This does not depend on the choice of i as the quotients ϕ ij are nowhere-zero regular functions, so the orders of ϕi and ϕ j are the same where they are both deﬁned. So we get a well-deﬁned map Div X → Zn−1 (X) deﬁned by D → ∑V ordV D · [V ]. It is obviously a homomorphism as ordV (ϕi · ϕi ) = ordV ϕi + ordV ϕi . It is clear from the deﬁnition that a Cartier divisor that is linearly equivalent to zero, i.e. a global rational function, determines a Weil divisor in Bn−1 (X). Hence the homomorphism passes to linear equivalence. Lemma 9.3.5. Let X be a smooth projective curve. Then Cartier divisors (resp. Cartier divisor classes) on X are the same as Weil divisors (resp. Weil divisor classes). In particular, our deﬁnition 9.3.2 (ii) of Div X and Pic X agrees with our earlier one from section 6.3. Proof. The idea of the proof is lemma 7.5.6 which tells us that every point of X is locally the scheme-theoretic zero locus of a single function, hence a Cartier divisor. To be more precise, let ∑n ai Pi ∈ Z0 (X) be a Weil divisor. We will construct a Cartier i=1 divisor D ∈ Div X that maps to the given Weil divisor under the correspondence of lemma 9.3.4. To do so, pick an open neighborhood Ui of Pi for all i = 1, . . . , n such that (i) Pj ∈ Ui for j = i, and / (ii) there is a function ϕPi on Ui such that div ϕPi = 1 · Pi on Ui (see lemma 7.5.6). Moreover, set U = X\{P1 , . . . , Pn }. Then we deﬁne a Cartier divisor D by the open cover {U,U1 , . . . ,Un } and the rational functions (i) 1 on U, (ii) ϕaii on Ui . P Note that these data deﬁne a Cartier divisor: no intersection of two elements of the open cover contains one of the points Pi , and the functions given on the elements of the open cover are regular and non-vanishing away from the Pi . By construction, the Weil divisor associated to D is precisely ∑n ai Pi , as desired. i=1 Example 9.3.6. In general, the map from Cartier divisors (resp. Cartier divisor classes) to Weil divisors (resp. Cartier divisor classes) is neither injective nor surjective. Here are examples of this: (i) not injective: This is essentially example 9.1.7. Let X = X1 ∪ X2 be the union of two lines Xi ∼ P1 glued together at a point P ∈ X1 ∩ X2 . Let Q be a point on = X1 \X2 . Consider the open cover X = U ∪V with U = X\Q and V = X1 \P. We deﬁne a Cartier divisor D on X by choosing the following rational functions on U and V : the constant function 1 on U, and the linear function on V ∼ A1 that = has a simple zero at Q. Note that the quotient of these two functions is regular and nowhere zero on U ∩V , so D is well-deﬁned. Its associated Weil divisor [D] is [Q].

9.

Intersection theory

179

By symmetry, we can construct a similar Cartier divisor D whose associated Weil divisor is the class of a point Q ∈ X2 \X1 . Now note that the Cartier divisor classes of D and D are different (because D − D is not the divisor of a rational function), but their associated Weil divisors [Q] and [Q ] are the same by example 9.1.7. (ii) not surjective: This is essentially example 9.3.1. The classes [Li ] of this example are Weil divisors but not Cartier divisors. Another example on an irreducible space X is the cone

2 2 2 X = {x3 = x1 + x2 } ⊂ P3 .

x3 L1 L2 x2 x1

Let L1 = Z(x2 , x1 + x3 ) and L2 = Z(x2 , x1 − x3 ) be the two lines as in the picture. We claim that there is no Cartier divisor on X corresponding to the Weil divisor [L1 ]. In fact, if there was such a Cartier divisor, deﬁned locally around the origin by a function ϕ, we must have an equality of ideals

2 2 2 (x1 + x2 − x3 , ϕ) = (x2 , x1 + x3 )

in the local ring OP3 ,0 . This is impossible since the right ideal contains two linearly independent linear parts, whereas the left ideal contains only one. But note that the section x2 of the line bundle OX (1) deﬁnes a Cartier divisor div(x2 ) on X whose associated Weil divisor is [L1 ] + [L2 ], and the section x1 + x3 deﬁnes a Cartier divisor whose associated Weil divisor is 2[L1 ]. So [L1 ] and [L2 ] are not Cartier divisors, whereas [L1 ] + [L2 ], 2[L1 ], and 2[L2 ] are. In particular, there is in general no “decomposition of a Cartier divisor into its irreducible components” as we have it by deﬁnition for Weil divisors. There is quite a deep theorem however that the two notions agree on smooth schemes: Theorem 9.3.7. Let X be a smooth n-dimensional scheme. Then Div X ∼ Zn−1 (X) and = Pic X ∼ An−1 (X). = Proof. We cannot prove this here and refer to [H] remark II.6.11.1.A for details. One has to prove the analogue of lemma 7.5.6, i.e. that every codimension-1 subvariety of X is locally the scheme-theoretic zero locus of a single function. This is a commutative algebra statement as it can be shown on the local ring of X at the subvariety. (To be a little more precise, the property of X that we need is that its local rings are unique factorization domains: if this is the case and V ⊂ X is an subvariety of codimension 1, pick any non-zero (local) function f ∈ OX,V that vanishes on V . As OX,V is a unique factorization domain we can decompose f into its irreducible factors f = f1 · · · fn . Of course one of the fi has to vanish on V . But as fi is irreducible, its ideal must be the ideal of V , so V is locally the zero locus of a single function. The problem with this is that it is almost impossible to check that a ring (that one does not know very well) is a unique factorization domain. So one uses the result from commutative algebra that every regular local ring (i.e. “the local ring of a scheme at a smooth point”) is a unique factorization

180

Andreas Gathmann

domain. Actually, we can see from the above argument that it is enough that X is “smooth in codimension 1”, i.e. that its set of singular points has codimension at least 2 — or to express it algebraically, that its local rings OX,V at codimension-1 subvarieties V are regular.) Example 9.3.8. Finally let us discuss the relation between divisors and line bundles as observed for curves in section 7.5. Note that we have in fact used such a correspondence already in example 9.3.6 where we deﬁned a Cartier divisor by giving a section of a line bundle. The precise relation between line bundles and Cartier divisors is as follows. Lemma 9.3.9. For any scheme X there are one-to-one correspondences {Cartier divisors on X} ↔ {(L , s) ; L a line bundle on X and s a rational section of L } and {Cartier divisor classes on X} ↔ {line bundles on X that admit a rational section}. Proof. The proof of this is essentially the same as the correspondence between divisor classes and line bundles on a smooth projective curve in proposition 7.5.9. Given a Cartier divisor D = {(Ui , ϕi )} on X, we get an associated line bundle O (D) by taking the subsheaf 1 of OX -modules of KX generated by the functions ϕi on Ui . Conversely, given a line bundle with a rational section, this section immediately deﬁnes a Cartier divisor. The proof that the same correspondence holds for divisor classes is the same as in proposition 7.5.9. Remark 9.3.10. We should note that almost any line bundle on any scheme X admits a rational section. In fact, this is certainly true for irreducible X (as the line bundle is then isomorphic to the structure sheaf on a dense open subset of X by deﬁnition), and one can show that it is true in most other cases as well (see [H] remark 6.14.1 for more information). Most books actually deﬁne the group Pic X to be the group of line bundles on X. Summarizing our above discussions we get the following commutative diagram:

together with a iii rational section iii iii

line bundles O

iiii tiiii / Cartier divisor classes Pic X Cartier divisors Div X Weil divisors Zn−1 (X) where (i) the bottom row (the Weil divisors) exists only if X is purely n-dimensional, (ii) the upper right vertical arrow is an isomorphism in most cases, at least if X is irreducible, (iii) the lower vertical arrows are isomorphisms at least if X is smooth (in codimension 1). Remark 9.3.11. Although line bundles, Cartier divisor classes, and Weil divisor classes are very much related and even all the same thing on many schemes (e.g. smooth varieties), note that their “functorial properties” are quite different: if f : X → Y is a morphism then for line bundles and Cartier divisors the pull-back f ∗ is the natural operation, whereas for Weil divisors (i.e. elements of the Chow groups) the push-forward f∗ as in section 9.2 is more natural. In algebraic topology this can be expressed by saying that Weil divisors correspond to homology cycles, whereas Cartier divisors correspond to cohomology cycles. On nice spaces this is the same by Poincar´ duality, but this is a non-trivial statement. The e / Weil divisor classes An−1 (X)

9.

Intersection theory

181

natural operation for homology (resp. cohomology) is the push-forward (resp. pull-back). Intersection products are deﬁned between a cohomology and a homology class, yielding a homology class. This corresponds to our initial statement of this section that intersection products of Chow cycles (“homology classes”) with divisors will usually only be welldeﬁned with Cartier divisors (“cohomology classes”) and not with Weil divisors. 9.4. Intersections with Cartier divisors. We are now ready to deﬁne intersection products of Chow cycles with Cartier divisors, as motivated in the beginning of section 9.3. Let us give the deﬁnition ﬁrst, and then discuss some of its features. Deﬁnition 9.4.1. Let X be a scheme, let V ⊂ X be a k-dimensional subvariety with inclusion morphism i : V → X, and let D be a Cartier divisor on X. We deﬁne the intersection product D ·V ∈ Ak−1 (X) to be D ·V = i∗ [i∗ OX (D)], where OX (D) is the line bundle on X associated to the Cartier divisor D by lemma 9.3.9, i∗ denotes the pull-back of line bundles, [i∗ OX (D)] is the Weil divisor class associated to the line bundle i∗ OX (D) by remark 9.3.10 (note that V is irreducible), and i∗ denotes the proper push-forward of corollary 9.2.12. Note that by deﬁnition the intersection product depends only on the divisor class of D, not on D itself. So using our deﬁnition we can construct bilinear intersection products Pic X × Zk (X) → Ak−1 (X), D, ∑ ai [Vi ] → ∑ ai (D ·Vi ). If X is smooth and pure-dimensional (so that Weil and Cartier divisors agree) and W is a codimension-1 subvariety of X, we denote by W · V ∈ Ak−1 (X) the intersection product D ·V , where D is the Cartier divisor corresponding to the Weil divisor [W ]. Example 9.4.2. Let X be a smooth n-dimensional scheme, and let V and W be subvarieties of dimensions k and n − 1, respectively. If V ⊂ W , i.e. if dim(W ∩ V ) = k − 1, then the intersection product W · V is just the cycle [W ∩ V ] with possibly some scheme-theoretic multiplicities. In fact, in this case the Weil divisor [W ] corresponds by remark 9.3.10 to a line bundle OX (W ) together with a section f whose zero locus is precisely W . By deﬁnition of the intersection product we have to pull back this line bundle to V , i.e. restrict the section f to V . The cycle W · V is then the zero locus of f |V , with possibly schemetheoretic multiplicities if f vanishes along V with higher order. As a concrete example, let C1 and C2 be two curves in P2 of degrees d1 and d2 , respectively, that intersect in ﬁnitely many points P1 , . . . , Pn . Then the intersection product C1 ·C2 ∈ A0 (P2 ) is just ∑i ai [Pi ], where ai is the scheme-theoretic multiplicity of the point Pi in the intersection scheme C1 ∩C2 . Using that all points in P2 are rationally equivalent, i.e. that A0 (P2 ) ∼ Z is generated by the class of any point, we see that C1 · C2 is just the = B´ zout number d1 · d2 . e Example 9.4.3. Again let X be a smooth n-dimensional scheme, and let V and W be subvarieties of dimensions k and n − 1, respectively. This time let us assume that V ⊂ W , so that the intersection W ∩ V = V has dimension k and thus does not deﬁne a (k − 1)dimensional cycle. There are two ways to interpret the intersection product W · V in this case: (i) Recall that the intersection product W · V depends only on the divisor class of W , not on W itself. So if we can replace W by a linearly equivalent divisor W such that V ⊂ W then the intersection product W · V is just W · V which can now be constructed as in example 9.4.2. For example, let H ⊂ P2 be a line and assume that we want to compute the intersection product H · H ∈ A0 (P2 ) ∼ Z. = The intersection H ∩ H has dimension 1, but we can move the ﬁrst H to a different line H which is linearly equivalent to H. So we see that H · H = H · H = 1, as

182

Andreas Gathmann

H ∩ H is just one point. Note however that it may not always be possible to ﬁnd such a linearly equivalent divisor that makes the intersection have the expected dimension. (ii) If the strategy of (i) does not work or one does not want to apply it, there is also a different description of the intersection product for which no moving of W is necessary. Let us assume for simplicity that W is smooth. By the analogue of remark 7.4.17 for general hypersurfaces the bundle i∗ OX (W ) (where i : V → X is the inclusion morphism) is precisely the restriction to V of the normal bundle NW /X of W in X. By deﬁnition 9.4.1 the intersection product W · V is then the Weil divisor associated to this bundle, i.e. the locus of zeros minus poles of a rational section of the normal bundle NW /X restricted to V .

V=W P1 P2 X

W. V

= [ P 1 ] +[ P 2 ]

Note that we can consider this procedure as an inﬁnitesimal version of (i): the section of the normal bundle describes an “inﬁnitesimal deformation” of W in X, and the deformed W meets V precisely in the locus where the section vanishes. Proposition 9.4.4. (Commutativity of the intersection product) Let X be an n-dimensional variety, and let D1 , D2 be Cartier divisors on X with associated Weil divisors [D1 ], [D2 ]. Then D1 · [D2 ] = D2 · [D1 ] ∈ An−2 (X). Proof. We will only sketch the proof in two easy cases (that cover most applications however). For the general proof we refer to [F] theorem 2.4. Case 1: D1 and D2 intersect in the expected dimension, i.e. the locus where the deﬁning equations of both D1 and D2 have a zero or pole has codimension 2 in X. Then one can show that both D1 · [D2 ] and D2 · [D1 ] is simply the sum of the components of the geometric intersection D1 ∩D2 , counted with their scheme-theoretic multiplicities. In other words, if V ⊂ X is a codimension-2 subvariety and if we assume for simplicity that the local deﬁning equations f1 , f2 for D1 , D2 around V are regular, then [V ] occurs in both intersection products with the coefﬁcient lA (A/( f1 , f2 )), where A = OX,V is the local ring of X at V . Case 2: X is a smooth scheme, so that Weil and Cartier divisors agree on X. Then it sufﬁces to compare the intersection products W ·V and V ·W for any two (n − 1)-dimensional subvarieties V,W of X. But the two products are obviously equal if V = W , and they are equal by case 1 if V = W . Corollary 9.4.5. The intersection product passes to rational equivalence, i.e. there are well-deﬁned bilinear intersection maps Pic X × Ak (X) → Ak−1 (X) determined by D · [V ] = [D ·V ] for all D ∈ Pic X and all k-dimensional subvarieties V of X. Proof. All that remains to be shown is that D · α = 0 for any Cartier divisor D if the cycle α is zero in the Chow group Ak (X). But this follows from proposition 9.4.4, as for any rational function ϕ on a (k + 1)-dimensional subvariety W of X we have D · [div(ϕ)] = div(ϕ) · [D] = 0 (note that div(ϕ) is a Cartier divisor on W that is linearly equivalent to zero).

9.

Intersection theory

183

Remark 9.4.6. Obviously we can now iterate the process of taking intersection products with Cartier divisors: if X is a scheme and D1 , . . . , Dm are Cartier divisors (or divisor classes) on X then there are well-deﬁned commutative intersection products D1 · D2 · · · Dm · α ∈ Ak−m (X) for any k-cycle α ∈ Ak (X). If X is an n-dimensional variety and α = [X] is the class of X we usually omit [X] from the notation and write the intersection product simply as D1 · D2 · · · Dm ∈ An−m (X). If m = n and X is complete, the notation D1 · D2 · · · Dm is moreover often used to denote the degree of the 0-cycle D1 · D2 · · · Dm ∈ A0 (X) (see example 9.2.13) instead of the cycle itself. If a divisor D occurs m times in the intersection product we will also write this as Dm . Example 9.4.7. Let X = P2 . Then Pic X = A1 (X) = Z · [H], and the intersection product is determined by H 2 = 1 (“two lines intersect in one point”). In the same way, H n = 1 on Pn . ˜ Example 9.4.8. Let X = P2 be the blow-up of P2 in a point P. By example 9.2.14 we have Pic X = Z[H] ⊕ Z[E], where E is the exceptional divisor, and H is a line in P2 not intersecting E. The strict transform L of a line in P2 through P has class [L] = [H] − [E] ∈ Pic X. The intersection products on X are therefore determined by computing the three prod/ ucts H 2 , H · E, and E 2 . Of course, H 2 = 1 and H · E = 0 (as H ∩ E = 0). To compute E 2 we use the relation [E] = [H] − [L] and the fact that E and L meet in one point (with multiplicity 1): E 2 = E · (H − L) = E · H − E · L = 0 − 1 = −1. By our interpretation of example 9.4.3 (ii) this means that the normal bundle of E ∼ P1 in = X is OP1 (−1). In particular, this normal bundle has no global sections. This means that E cannot be deformed in X as in the picture of example 9.4.3 (ii): one says that the curve E is rigid in X. We can consider the formulas H 2 = 1, H · E = 0, E 2 = −1, together with the existence of the intersection product Pic X × Pic X → Z as a B´ zout style theorem for the blow-up e ˜ X = P2 . In the same way, we get B´ zout style theorems for other (smooth) surfaces and e even higher-dimensional varieties. Example 9.4.9. As a more complicated example, let us reconsider the question of exercise 4.6.6: how many lines are there in P3 that intersect four general given lines L1 , . . . , L4 ⊂ P3 ? Recall from exercise 3.5.4 that the space of lines in P3 is the smooth four-dimensional Grassmannian variety X = G(1, 3) that can be described as the set of all rank-2 matrices a0 b0 a1 b1 a2 b2 a3 b3

modulo row transformations. By the Gaussian algorithm it follows that G(1, 3) has a stratiﬁcation by afﬁne spaces X4 , X3 , X2 , X2 , X1 , X0 (where the subscript denotes the dimension and the stars denote arbitrary complex numbers) 1 0 0 ∗ 1 ∗ X4 ∗ ∗ 1 ∗ 0 ∗ 0 0 1 ∗ X3 0 0 1 ∗ 0 0 0 1 X1 1 ∗ ∗ 0 0 0 0 1 X2 0 0 0 1 0 0 X0 0 1

0 0

1 0 ∗ 0 1 ∗ X2

184

Andreas Gathmann

If we denote by σ4 , . . . , σ0 the classes in A∗ (X) of the closures of X4 , . . . , X0 , we have seen in remark 9.1.18 that A∗ (X) is generated by the classes σ4 , . . . , σ0 . These classes actually all have a geometric interpretation: (i) σ4 = [X]. (ii) σ3 is the class of all lines that intersect the line {x0 = x1 = 0} ⊂ P3 . Note that this is precisely the zero locus of a0 b1 − a1 b0 . In particular, if L ⊂ P3 is any other line then the class σL of all lines in P3 meeting L is also a quadratic function q 3 in the entries of the matrix that is invariant under row transformations (in fact a 2 × 2 minor in a suitable choice of coordinates of P3 ). The quotient a0 b1 −a1 b0 is q then a rational function on X whose divisor is σ3 − σL . It follows that the class 3 σL does not depend on L. So we can view σ3 as the class that describes all lines 3 intersecting any given line in P3 . (iii) σ2 is the class of all lines passing through the point (0 : 0 : 0 : 1). By an argument similar to that in (ii) above, we can view σ2 as the class of all lines passing through any given point in P3 . (iv) σ2 is the class of all lines that are contained in a plane (namely in the plane x0 = 0 for the cycle X2 given above). (v) σ1 is the class of all lines that are contained in a plane and pass through a given point in this plane. (vi) σ0 is the class of all lines passing through two given points in P3 . ∼ Hence we see that the intersection number we are looking for is just σ4 ∈ A0 (X) = Z — 3 the number of lines intersecting any four given lines in P3 . So let us compute this number. Step 1. Let us compute σ2 ∈ A2 (X), i.e. class of all lines intersecting two given lines 3 L1 , L2 in P3 . We have seen above that it does not matter which lines we take, so let us choose L1 and L2 such that they intersect in a point P ∈ P3 . A line that intersects both L1 and L2 has then two possibilities: (i) it is any line in the plane spanned by L1 and L2 , (ii) it is any line in P3 passing through P. As (i) corresponds to σ2 and (ii) to σ2 we see that σ2 = σ2 + σ2 . To be more precise, 3 we still have to show that σ2 contains both X2 and X2 with multiplicity 1 (and not with a 3 higher multiplicity). As an example, we will show that σ2 contains σ2 with multiplicity 3 1; the proof for σ2 is similar. Consider the open subset X4 ⊂ G(1, 3); it is isomorphic to an afﬁne space A4 with coordinates a2 , a3 , b2 , b3 . On this open subset, the space of lines intersecting the line {x0 = x2 = 0} is given scheme-theoretically by the equation b2 = 0, whereas the space of lines intersecting the line {x0 = x3 = 0} is given scheme-theoretically by the equation b3 = 0. The scheme-theoretic intersection of these two spaces (i.e. the product σ2 ) is then given by b2 = b3 = 0, which is precisely the locus of lines through the 3 point (0 : 1 : 0 : 0) (with multiplicity 1), i.e. the cycle σ2 . Step 2. In the same way we compute that (i) σ3 · σ2 = σ1 (lines meeting a line L and a point P are precisely lines in the plane spanned by L and P passing through P), (ii) σ3 · σ2 = σ1 (lines meeting a line L and contained in a plane H are precisely lines in the plane H passing through the point H ∩ L), (iii) σ3 · σ1 = σ0 . So we conclude that σ4 = σ2 (σ2 + σ2 ) = 2σ3 σ1 = 2, 3 3 i.e. there are exactly two lines in P3 meeting four other general given lines.

9.

Intersection theory

185

We should note that similar decompositions into afﬁne spaces exist for all Grassmannian varieties, as well as rules how to intersect the corresponding Chow cycles. These rules are usually called Schubert calculus. They can be used to answer almost any question of the form: how many lines in Pn satisfy some given conditions? Finally, let us prove a statement about intersection products that we will need in the next section. It is based on the following set-theoretic idea: let f : X → Y be any map of sets, and let V ⊂ X and W ⊂ Y be arbitrary subsets. Then it is checked immediately that f ( f −1 (W ) ∩V ) = W ∩ f (V ). This relation is called a projection formula. There are projection formulas for many other morphisms and objects that can be pushed forward and pulled back along a morphism. We will prove an intersection-theoretic version here. Lemma 9.4.10. Let f : X → Y be a proper surjective morphism of schemes. Let α ∈ Ak (X) be a k-cycle on X, and let D ∈ PicY be a Cartier divisor (class) on Y . Then f∗ ( f ∗ D · α) = D · f∗ α ∈ Ak−1 (Y ). Proof. (Note that this is precisely the set-theoretic intersection formula from above, together with the statement that the scheme-theoretic multiplicities match up in the right way.) By linearity we may assume that α = [V ] for a k-dimensional subvariety V ⊂ X. Let W = f (V ), and denote by g : V → W the restriction of f to V . Then the left hand side of the equation of the lemma is g∗ [g∗ D ], where D is the Cartier divisor on W associated to the line bundle OY (D)|W . The right hand side is [K(V ) : K(W )] · [D ] by construction 9.2.9, with the convention that [K(V ) : K(W )] = 0 if dimW < dimV . We will prove that these expressions actually agree in Zk−1 (W ) for any given Cartier divisor D . This is a local statement (as we just have to check that every codimension-1 subvariety of W occurs on both sides with the same coefﬁcient), so passing to an open subset we can assume that D is the divisor of a rational function ϕ on W . But then by theorem 9.2.11 the left hand side is equal to g∗ div(g∗ ϕ) = div N(g∗ ϕ) = div(ϕ[K(V ):K(W )] ) = [K(V ) : K(W )] · div(ϕ), which equals the right hand side. 9.5. Exercises. Exercise 9.5.1. Let X ⊂ Pn be a hypersurface of degree d. Compute the Chow group An−1 (Pn \X). Exercise 9.5.2. Compute the Chow groups of X = Pn × Pm for all n, m ≥ 1. Assuming that there are “intersection pairing homomorphisms” An+m−k (X) × An+m−l (X) → An+m−k−l (X), (α, α ) → α · α

such that [V ∩W ] = [V ] · [W ] for all subvarieties V,W ⊂ X that intersect in the expected dimension, compute these homomorphisms explicitly. Use this to state a version of B´ zout’s e theorem for products of projective spaces. Exercise 9.5.3. (This is a generalization of example 9.1.7.) If X1 and X2 are closed subschemes of a scheme X show that there are exact sequences Ak (X1 ∩ X2 ) → Ak (X1 ) ⊕ Ak (X2 ) → Ak (X1 ∪ X2 ) → 0 for all k ≥ 0.

186

Andreas Gathmann

Exercise 9.5.4. Show that for any schemes X and Y there are well-deﬁned product homomorphisms Ak (X) × Al (Y ) → Ak+l (X ×Y ), [V ] × [W ] → [V ×W ]. If X has a stratiﬁcation by afﬁne spaces as in remark 9.1.18 show that the induced homomorphisms M Ak (X) × Al (Y ) → Am (X ×Y )

k+l=m

are surjective. (In general, they are neither injective nor surjective). Exercise 9.5.5. Prove the following criteria to determine whether a morphism f : X → Y is proper: (i) The composition of two proper morphisms is proper. (ii) Properness is “stable under base change”: if f : X → Y is proper and g : Z → Y is any morphism, then the induced morphism f : X ×Y Z → Z is proper as well. (iii) Properness is “local on the base”: if {Ui } is any open cover of Y and the restrictions f | f −1 (Ui ) : f −1 (Ui ) → Ui are proper for all i then f is proper. (iv) Closed immersions (see 7.2.10) are proper. Exercise 9.5.6. Let f : P1 → P1 be the morphism given in homogeneous coordinates by 2 2 (x0 : x1 ) → (x0 : x1 ). Let P ∈ P1 be the point (1 : 1), and consider the restriction f˜ : 1 \{P} → P1 . Show that f is not proper, both with the topological and the algebraic ˜ P deﬁnition of properness. Exercise 9.5.7. For any n > 0 compute the Chow groups of P2 blown up in n points. Exercise 9.5.8. Let k be an algebraically closed ﬁeld. In this exercise we will construct an example of a variety that is complete (i.e. compact if k = C) but not projective. 2 Consider X = P3 and the curves C1 = {x3 = x2 − x1 = 0} and C2 = {x3 = x0 x2 − x1 = 0} in X. Denote by P1 = (1 : 0 : 0 : 0) and P2 = (1 : 1 : 1 : 0) their two intersection points. ˜ ˜ ˜ Let X1 → X be the blow-up at C1 , and let X1 → X1 be the blow-up at the strict trans˜1 → X the projection map. Similarly, let π2 : X2 → X be ˜ form of C2 . Denote by π1 : X the composition of the two blow-ups in the opposite order; ﬁrst blow up C2 and then the ˜ ˜ strict transform of C1 . Obviously, X1 and X2 are isomorphic away from the inverse im−1 age of {P1 , P2 }, so we can glue π1 (X\{P1 }) and π−1 (X\{P2 }) along the isomorphism 2 π−1 (X\{P1 , P2 }) ∼ π−1 (X\{P1 , P2 }) to get a variety Y . This variety will be our example. = 2 1 From the construction there is an obvious projection map π : Y → X. (i) Show that Y is proper over k. (ii) For i = 1, 2 we know that Ci is isomorphic to P1 . Hence we can choose a rational function ϕi on Ci with divisor P1 −P2 . Compute the divisor of the rational function ϕi ◦ π on the variety π−1 (Ci ), as an element in Z1 (Y ). (iii) From (ii) you should have found two irreducible curves D1 , D2 ⊂ Y such that [D1 ] + [D2 ] = 0 ∈ A1 (Y ). Deduce that Y is not a projective variety. Exercise 9.5.9. Let X be a smooth projective surface, and let C, D ⊂ X be two curves in X that intersect in ﬁnitely many points. (i) Prove that there is an exact sequence of sheaves on X 0 → OX (−C − D) → OX (−C) ⊕ OX (−D) → OX → OC∩D → 0. (ii) Conclude that the intersection product C · D ∈ Z is given by the formula C · D = χ(X, OX ) + χ(X, OX (−C − D)) − χ(X, OX (−C)) − χ(X, OX (D)) where χ(X, F ) = ∑i (−1)i hi (X, F ) denotes the Euler characteristic of the sheaf F.

9.

Intersection theory

187

(iii) Show how the idea of (ii) can be used to deﬁne an intersection product of divisors on a smooth complete surface (even if the divisors do not intersect in dimension zero).

188

Andreas Gathmann

10. C HERN CLASSES

For any vector bundle π : F → X of rank r on a scheme X we deﬁne an associated projective bundle p : P(F) → X whose ﬁbers p−1 (P) are just the projectivizations of the afﬁne ﬁbers π−1 (P). We construct natural line bundles OP(F) (d) on P(F) for all d ∈ Z that correspond to the standard line bundles O (d) on projective spaces. As in the case of vector bundles there are pull-back homomorphisms A∗ (X) → A∗ (P(F)) between the Chow groups. For a bundle as above we deﬁne the i-th Segre class si (F) : A∗ (X) → A∗−i (X) by si (F) · α = p∗ (Dr−1+i · p∗ α), where DF denotes the Cartier divisor associated to the F line bundle OP(F) (1). The Chern classes ci (F) are deﬁned to be the inverse of the Segre classes. Segre and Chern classes are commutative; they satisfy the projection formula for proper push-forwards and are compatible with pull-backs. They are multiplicative on exact sequences. Moreover, ci (F) = 0 for i > r. The top Chern class cr (F) has the additional geometric interpretation as the zero locus of a section of F. Using the technique of Chern roots one can compute the Chern classes of almost any bundle that is constructed from known bundles in some way (e.g. by means of direct sums, tensor products, dualizing, exact sequences, symmetric and exterior products). The Chern character ch(F) and Todd class td(F) are deﬁned to be certain polynomial combinations of the Chern classes of F. The Hirzebruch-Riemann-Roch theorem states that ∑i hi (X, F) = deg(ch(F)·td(TX )) for any vector bundle F on a smooth projective scheme X. We study some examples and applications of this theorem and give a sketch of proof.

10.1. Projective bundles. Recall that for any line bundle L on a variety X there is a Cartier divisor on X corresponding to L that in turn deﬁnes intersection homomorphisms Ak (X) → Ak−1 (X). These homomorphisms can be thought of as intersecting a k-cycle on X with the divisor of any rational section of L . We now want to generalize this idea from line bundles to vector bundles. To do so, we need some preliminaries on projective bundles ﬁrst. Roughly speaking, the projective bundle P(E) associated to a vector bundle E of rank r on a scheme X is simply obtained by replacing the ﬁbers (that are all isomorphic to Ar ) by the corresponding projective spaces Pr−1 = (Ar \{0})/k∗ . Let us give the precise deﬁnition. Deﬁnition 10.1.1. Let π : F → X be a vector bundle of rank r on a scheme X (see remark 7.3.2). In other words, there is an open covering {Ui } of X such that (i) there are isomorphisms ψi : π−1 (Ui ) → Ui × Ar over Ui , (ii) on the overlaps Ui ∩U j the compositions ψi ◦ ψ−1 : (Ui ∩U j ) × Ar → (Ui ∩U j ) × Ar j are linear in the coordinates of Ar , i.e. they are of the form (P, x) → (P, Ψi, j x) where P ∈ U, x = (x1 , . . . , xr ) ∈ Ar , and the Ψi, j are r × r matrices with entries in OX (Ui ∩U j ). Then the projective bundle P(F) is deﬁned by glueing the patches Ui × Pr−1 along the same transition functions, i.e. by glueing Ui × Pr−1 to U j × Pr−1 along the isomorphisms (Ui ∩U j ) × Pr−1 → (Ui ∩U j ) × Pr−1 , (P, x) → (P, Ψi, j x)

for all i, j, where P ∈ Ui ∩U j and x = (x1 : · · · : xr ) ∈ Pr−1 . We say that P(F) is a projective bundle of rank r − 1 on X.

10.

Chern classes

189

Note that in the same way as for vector bundles there is a natural projection morphism p : P(F) → X that sends a point (P, x) to P. In contrast to the vector bundle case however the morphism p is proper (which follows easily from exercise 9.5.5). Example 10.1.2. Let X = P1 , and let F be the vector bundle (i.e. locally free sheaf) OX ⊕ OX (−1) on X. Then P(F) is a projective bundle of rank 1 on X, so it is a scheme of ˜ dimension 2. We claim that P(F) is isomorphic to the blow-up P2 of the projective plane in a point P. In fact, this can be checked directly: by deﬁnition 10.1.1 P(F) is obtained by glueing two copies U1 ,U2 of A1 × P1 along the isomorphism (A1 \{0}) × P1 → (A1 \{0}) × P1 , ˜ On the other hand, P2 is given by ˜ P2 = {((x0 : x1 : x2 ), (y1 : y2 )) ; x1 y2 = x2 y1 } ⊂ P2 × P1 (see example 4.3.4). Now an isomorphism is given by ˜ U1 ∼ A1 × P1 → P2 , (z, (x1 : x2 )) → ((x1 : zx2 : x2 ), (z : 1)), = ˜ U2 ∼ A1 × P1 → P2 , (z, (x1 : x2 )) → ((x1 : x2 : zx2 ), (1 : z)) = (note that this is compatible with the glueing isomorphism above). ˜ ˜ To see geometrically that P2 is a projective bundle of rank 1 over P1 let p : P2 → ∼ P1 be the projection morphism onto the exceptional divisor as of example 9.2.14 E= (ii). The ﬁbers of this morphism are the strict transforms of lines through P, so they are all isomorphic to P1 . ∼ Remark 10.1.3. If F is a vector bundle and L a line bundle on X then P(F) = P(F ⊗ L). In fact, tensoring F with L just multiplies the transition matrices Ψi, j of deﬁnition 10.1.1 with a scalar function, which does not affect the morphism as the xi are projective coordinates. Example 10.1.4. Let p : P(F) → X be a projective bundle over a scheme X, given by an open cover {Ui } of X and transition matrices Ψi, j as in deﬁnition 10.1.1. In this example we want to construct line bundles OP(F) (d) for all d ∈ Z on P(F) that are relative versions of the ordinary bundles OPr−1 (d) on projective spaces. The construction is simple: on the patches Ui × Pr−1 of P(F) we take the line bundles f OPr−1 (d). On the overlaps Ui ∩U j these line bundles are glued by ϕ → ϕ◦Ψi, j , where ϕ = g is (locally) a quotient of homogeneous polynomials f , g ∈ k[x1 , . . . , xr ] with deg f − deg g = d. Note that the ϕ ◦ Ψi, j satisﬁes the same degree conditions as the Ψi, j are linear functions. Summarizing, we can say that sections of the line bundle OP(F) (d) are locally given by quotients of two polynomials which are homogeneous in the ﬁber coordinates and whose degree difference is d. Construction 10.1.5. Again let p : P(F) → X be a projective bundle over a scheme X, given by an open cover {Ui } of X and transition matrices Ψi, j . Consider the vector bundle p∗ F on P(F). It is given by glueing the patches Ui × Pr−1 × Ar along the isomorphisms (Ui ∩U j ) × Pr−1 × Ar → (Ui ∩U j ) × Pr−1 × Ar , (P, x, y) → (P, Ψi, j x, Ψi, j y), where x = (x1 : · · · : xr ) are projective coordinates on Pr−1 , and y = (y1 , . . . , yr ) are afﬁne coordinates on Ar . Now consider the subbundle S of p∗ F given locally by the equations xi y j = x j yi for all i, j = 1, . . . , r, i.e. the subbundle of p∗ F consisting of those (y1 , . . . , yr ) that are scalar multiples of (x1 : · · · : xr ). Obviously, S is a line bundle on P(F) contained in p∗ F. Geometrically, the ﬁber of S over a point (P, x) ∈ P(F) is precisely the line in the ﬁber FP whose projectivization is the point x. The line bundle S ⊂ p∗ F is called the tautological subbundle on P(F). 1 (z, (x1 : x2 )) → ( , (x1 : zx2 )). z

190

Andreas Gathmann

We can actually identify the subbundle S in the language of example 10.1.4: we claim that S is isomorphic to OP(F) (−1). In fact, an isomorphism is given by

OP(F) (−1) → S, ϕ → (yi = ϕ · xi ),

where ϕ is (locally) the quotient of two polynomials homogeneous in the xi of degree difference −1. It is obvious that the ϕ · xi are then quotients of two polynomials homogeneous in the xi of the same degree, so that the yi are well-deﬁned. Example 10.1.6. One place where projective bundles occur naturally is in blow-ups. Re˜ call from construction 4.3.2 that the blow-up X of an afﬁne variety X ⊂ An at a subvariety Y ⊂ X with ideal I(Y ) = ( f1 , . . . , fr ) is deﬁned to be the closure of the graph Γ = {(P, ( f1 (P) : · · · : fr (P))) ; P ∈ X\Y } ⊂ X × Pr−1 . The exceptional hypersurface of the blow-up must be contained in Y × Pr−1 , which has dimension dimY +r −1. So if Y has dimension dim X −r (which is the expected dimension as its ideal has r generators) then the exceptional hypersurface must be all of Y × Pr−1 for dimensional reasons. Let us now sketch how this construction can be generalized to blow-ups of arbitrary (not necessarily afﬁne) varieties X in a subvariety Y . For simplicity let us assume that there are r line bundles L1 , . . . , Lr on X together with global sections si ∈ H 0 (X, Li ) such that Y is scheme-theoretically the zero locus s1 = · · · = sr = 0. Then the straightforward generalization of the above construction is to deﬁne the blow-up of X in Y to be the closure of the graph Γ = {(P, (s1 (P) : · · · : sr (P)) ; P ∈ X\Y } ⊂ P(L1 ⊕ · · · ⊕ Lr ). As above, if Y has codimension r in X then the exceptional hypersurface of the blow-up is the projective bundle P((L1 ⊕ · · · ⊕ Lr )|Y ) over Y . Now recall from remark 7.4.17 and example 9.4.3 (ii) that the normal bundle of a smooth codimension-1 hypersurface Y in a smooth variety X that is given as the zero locus of a section of a line bundle L is just the restriction of this line bundle L to Y . If we iterate this result r times we see that the normal bundle of a smooth codimension-r hypersurface Y in a smooth variety X that is given as the zero locus of sections of r line bundles L1 , . . . , Lr is just (L1 ⊕ · · · ⊕ Lr )|Y . Combining this with what we have said above we conclude that the exceptional hypersurface of the blow-up of a smooth variety X in a smooth variety Y is just the projectivized normal bundle P(NY /X ) over Y . This is a relative version of our earlier statement that the exceptional hypersurface of the blow-up of a variety in a smooth point is isomorphic to the projectivized tangent space at this point. In the above argument we have used for simplicity that the codimension-r subvariety Y is globally the zero locus of r sections of line bundles. Actually we do not need this. We only need that Y is locally around every point the zero locus of r regular functions, as we can then make the above construction locally and ﬁnally glue the local patches together. Using techniques similar to those in theorem 9.3.7 one can show that every smooth subvariety Y of codimension r in a smooth variety X is locally around every point the zero locus of r regular functions. So it is actually true in general that the exceptional hypersurface of the blow-up of X in Y is P(NY /X ) if X and Y are smooth. Finally, in analogy to the case of vector bundles in proposition 9.1.14 let us discuss pull-back homomorphisms for Chow groups induced by projective bundles. Lemma 10.1.7. Let F be a vector bundle on a scheme X of rank r +1, and let p : P(F) → X be the associated projective bundle of rank r. Then there are pull-back homomorphisms p∗ : Ak (X) → Ak+r (P(F)), [V ] → [p−1 (V )] for all k, satisfying the following compatibilities with our earlier constructions:

10.

Chern classes

191

(i) (Compatibility with proper push-forward) Let f : X → Y be a proper morphism, and let F be a vector bundle of rank r + 1 on Y . Form the ﬁber diagram P( f ∗ F)

p f

/ P(F)

p

X

f

/ Y.

Then p∗ f∗ = f∗ p ∗ as homomorphisms Ak (X) → Ak+r (P(F)). (ii) (Compatibility with intersection products) Let F be a vector bundle of rank r + 1 on X, and let D ∈ Pic X be a Cartier divisor (class). Then p∗ (D · α) = (p∗ D) · (p∗ α) in Ak+r−1 (P(F)) for every k-cycle α ∈ Ak (X). Proof. (i): Let V ⊂ X be a k-dimensional subvariety. Then p−1 ( f (V )) = f (p −1 (V )) =: W , and both p∗ f∗ [V ] and f∗ p ∗ [V ] are equal to d · [W ], where d is the generic number of inverse image points of f (resp. f ) on f (V ) (resp. p−1 ( f (V )). (ii): Let α = [V ] for a k-dimensional subvariety V ⊂ X. On V the Cartier divisor D is given by a line bundle L . If ϕ is any rational section of L then the statement follows from the obvious identity p∗ div(ϕ) = div(p∗ ϕ). Remark 10.1.8. We have now constructed pull-back morphisms for Chow groups in three cases: (i) inclusions of open subsets (example 9.1.11), (ii) projections from vector bundles (proposition 9.1.14), (iii) projections from projective bundles (lemma 10.1.7). These are in fact special cases of a general class of morphisms, called ﬂat morphisms, for which pull-back maps exist. See [F] section 1.7 for more details. 10.2. Segre and Chern classes of vector bundles. Let X be a scheme, and let F be a vector bundle of rank r on X. Let p : P(F) → X be the projection from the corresponding projective bundle. Note that we have the following constructions associated to p: (i) push-forward homomorphisms p∗ : Ak (P(F)) → Ak (X) since p is proper (see corollary 9.2.12), (ii) pull-back homomorphisms p∗ : Ak (X) → Ak+r−1 (P(F)) by lemma 10.1.7, (iii) a line bundle OP(F) (1) on P(F) by example 10.1.4 (the dual of the tautological subbundle). We can now combine these three operations to get homomorphisms of the Chow groups of X that depend on the vector bundle F: Deﬁnition 10.2.1. Let X be a scheme, and let F be a vector bundle of rank r on X. Let p : P(F) → X be the projection map from the associated projective bundle. Assume for simplicity that X (and hence P(F)) is irreducible (see below), so that the line bundle OP(F) (1) corresponds to a Cartier divisor DF on P(F). Now for all i ≥ −r + 1 we deﬁne Segre class homomorphisms by the formula si (F) : Ak (X) → Ak−i (X), α → si (F) · α := p∗ (Dr−1+i · p∗ α). F

Remark 10.2.2. We will discuss some geometric interpretations of Segre classes (or rather some combinations of them) later in proposition 10.2.3 (i) and (ii), proposition 10.3.12, and remark 10.3.14. For the moment let us just note that every vector bundle F gives rise to these homomorphisms si (F) that look like intersections (hence the notation si (F) · α) with

192

Andreas Gathmann

some object of codimension i as they decrease the dimension of cycles by i. (In algebraic topology the Segre class si (F) is an object in the cohomology group H 2i (X, Z).) Note also that the condition that X be irreducible is not really necessary: even if OP(F) (1) does not determine a Cartier divisor on P(F) it does so on every subvariety of P(F), and this is all we need for the construction of the intersection product (as we intersect with a cycle in P(F) which is by deﬁnition a formal linear combination of subvarieties). Proposition 10.2.3. Let X and Y be schemes. (i) For any vector bundle F on X we have • si (F) = 0 for i < 0, • s0 (F) = id. (ii) For any line bundle L on X we have si (L) · α = (−1)i Di · α for i ≥ 0 and all α ∈ A∗ (X), where D is the Cartier divisor class associated to the line bundle L. (iii) (Commutativity) If F1 and F2 are vector bundles on X, then si (F1 ) · s j (F2 ) = s j (F2 ) · si (F1 ) as homomorphisms Ak (X) → Ak−i− j (X) for all i, j (where the dot denotes the composition of the two homomorphisms). (iv) (Projection formula) If f : X → Y is proper, F is a vector bundle on Y , and α ∈ A∗ (X), then f∗ (si ( f ∗ F) · α) = si (F) · f∗ α. (v) (Compatibility with pull-back) If f : X → Y is a morphism for which a pull-back f ∗ : A∗ (Y ) → A∗ (X) exists (see remark 10.1.8), F is a vector bundle on Y , and α ∈ A∗ (Y ), then si ( f ∗ F) · f ∗ α = f ∗ (si (F) · α).

Proof. (i): Let V ⊂ X be a k-dimensional subvariety. By construction we can represent si (F) · [V ] by a cycle of dimension k − i supported in V . As Zk−i (V ) = 0 for i < 0 and Zk (V ) = [V ] we conclude that si (F) = 0 for i < 0 and s0 (F) · [V ] = n · [V ] for some n ∈ Z. The computation of the multiplicity n is a local calculation, so we can replace X by an open subset and thus assume that F is a trivial bundle. In this case P(F) = X × Pr−1 and r−1 DF is a hyperplane in Pr−1 . So Dr−1 is a point in Pr−1 , i.e. DF · p∗ [V ] = [V × {pt}] and F hence s0 (F) · [V ] = [V ]. (ii): If L is a line bundle then P(L) = X and p is the identity. Hence the statement follows from the identity OP(L) (−1) = L. The proofs of (iii), (iv), and (v) all follow from the various compatibilities between push-forward, pull-back, and intersection products. As an example we give the proof of (iv), see [F] proposition 3.1 for the other proofs. For (iv) consider the ﬁber square P( f ∗ F)

p f

/ P(F) / Y

p

X

f

10.

Chern classes

193

and denote the Cartier divisors associated to the line bundles OP(F) (1) and OP( f ∗ F) (1) by DF and DF , respectively. Then f∗ (si ( f ∗ F) · α) = f∗ p∗ (DF r−1+i · p ∗ α) = p∗ f∗ (DF r−i+1 · p ∗ α) = = p∗ f∗ (( f ∗ DF )r−i+1 · p ∗ α) r−i+1 p∗ (DF · f∗ p ∗ α) r−i+1 p∗ (DF · p∗ f∗ α) by deﬁnition 10.2.1 by remark 9.2.10 as DF = f ∗ DF by lemma 9.4.10 by lemma 10.1.7 (i) by deﬁnition 10.2.1.

= = si (E) · f∗ α

Corollary 10.2.4. Let F be a vector bundle on a scheme X, and let p : P(F) → X be the projection. Then p∗ : A∗ (P(F)) → A∗ (X) is surjective and p∗ : A∗ (X) → A∗ (P(F)) is injective. Proof. By proposition 10.2.3 (i) we have α = s0 (F) · α = p∗ (Dr−1 · p∗ α) F for all α ∈ A∗ (X), so p∗ is surjective. The same formula shows that α = 0 if p∗ α = 0, so p∗ is injective. By proposition 10.2.3 (iii) any polynomial expression in the Segre classes of some vector bundles acts on the Chow groups of X. Although the Segre classes are the characteristic classes of vector bundles that are the easiest ones to deﬁne, some others that are polynomial combinations of them have nicer properties and better geometric interpretations. Let us now deﬁne these combinations. Deﬁnition 10.2.5. Let X be a scheme, and let F be a vector bundle of rank r on X. The total Segre class of F is deﬁned to be the formal sum s(F) = ∑ si (F) : A∗ (X) → A∗ (X).

i≥0

Note that: (i) All si (F) can be recovered from the homomorphism s(F) by considering the graded parts. (ii) Although the sum over i in s(F) is formally inﬁnite, it has of course only ﬁnitely many terms as Ak (X) is non-zero only for ﬁnitely many k. (iii) The homomorphism s(F) is in fact an isomorphism of vector spaces: by proposition 10.2.3 (i) it is given by a triangular matrix with ones on the diagonal (in the natural grading of A∗ (X)). By (iii) it makes sense to deﬁne the total Chern class of F c(F) = ∑ ci (F)

i≥0

to be the inverse homomorphism of s(F). In other words, the Chern classes ci (F) are the unique homomorphisms ci (F) : Ak (X) → Ak−i (X) such that s(F) · c(F) = (1 + s1 (F) + s2 (F) + · · · ) · (c0 (F) + c1 (F) + c2 (F) + · · · ) = id .

194

Andreas Gathmann

Explicitly, the ﬁrst few Chern classes are given by c0 (F) = 1, c1 (F) = −s1 (F), c2 (F) = −s2 (F) + s1 (F)2 , c3 (F) = −s3 (F) + 2s1 (F)s2 (F) − s1 (F)3 . Proposition 10.2.3 translates directly into corresponding statements about Chern classes: Proposition 10.2.6. Let X and Y be schemes. (i) For any line bundle L on X with associated Cartier divisor class D we have c(L) · α = (1 + D) · α. In other words, ci (L) = 0 for i > 1, and c1 (L) is the homomorphism of intersection with the Cartier divisor class associated to L. By abuse of notation, the Cartier divisor class associated to L is often also denoted c1 (L). (ii) (Commutativity) If F1 and F2 are vector bundles on X, then ci (F1 ) · c j (F2 ) = c j (F2 ) · ci (F1 ) for all i, j. (iii) (Projection formula) If f : X → Y is proper, F is a vector bundle on Y , and α ∈ A∗ (X), then f∗ (ci ( f ∗ F) · α) = ci (F) · f∗ α. (iv) (Pull-back) If f : X → Y is a morphism for which a pull-back f ∗ : A∗ (Y ) → A∗ (X) exists, F is a vector bundle on Y , and α ∈ A∗ (Y ), then ci ( f ∗ F) · f ∗ α = f ∗ (ci (F) · α). Proof. (i): This follows from proposition 10.2.3, since (1 − D + D2 − D3 ± · · · )(1 + D) = 1. (ii), (iii), (iv): All these statements follow from the corresponding properties of Segre classes in proposition 10.2.3, taking into account that the Chern classes are just polynomials in the Segre classes. 10.3. Properties of Chern classes. In this section we will show how to compute the Chern classes of almost any bundle that is constructed from other known bundles in some way (e.g. by means of direct sums, tensor products, dualizing, exact sequences, symmetric and exterior products). We will also discuss the geometric meaning of Chern classes. The most important property of Chern classes is that they are multiplicative in exact sequences: Proposition 10.3.1. Let 0 → F → F → F → 0 be an exact sequence of vector bundles on a scheme X. Then c(F) = c(F ) · c(F ). Proof. We prove the statement by induction on the rank of F . Step 1: rank F = 1. We have to show that s(F ) · [V ] = c(F ) · s(F) · [V ] for all kdimensional subvarieties V ⊂ X. Consider the diagram P = P(F |V ) NNN NNN NN' p

i

V

/ P(F|V ) = P ppp ppp p xpp

10.

Chern classes

195

Then c(F ) · s(F) · [V ] = c(F ) · p∗ ((1 + DF + D2 + · · · ) · [P]) F = c(F ) · p∗ (s(OP (−1)) · [P]) = (1 + c1 (F )) · p∗ (s(OP (−1)) · [P]) = p∗ ((1 + c1 (p∗ F )) · s(OP (−1)) · [P])

by deﬁnition 10.2.1 by proposition 10.2.3 (ii) by proposition 10.2.6 (i) by proposition 10.2.6 (iii).

On the other hand, we have a bundle map OP (−1) → p∗ F → p∗ F on P, which by construction fails to be injective exactly at the points of P . In other words, P in P is the (scheme-theoretic) zero locus of a section of the line bundle p∗ F ⊗ OP (−1)∨ . So we get s(F ) · [V ] = p∗ (s(OP (−1)) · [P ]) = p∗ i∗ (s(i∗ OP (−1)) · [P ]) = p∗ (s(OP (−1)) · i∗ [P ]) = p∗ (s(OP (−1)) · (c1 (p∗ F ) − c1 (OP (−1))) · [P]). Subtracting these two equations from each other, we get c(F ) · s(F) · [V ] − s(F ) · [V ] = p∗ (s(OP (−1)) c(OP (−1)) [P]) = p∗ [P] = 0 for dimensional reasons. Step 2: rank F > 1. Let Q = P(F ∨ ) with projection map q : Q → X, and let L∨ ⊂ ∗ F ∨ be the universal line bundle. Then we get a commutative diagram of vector bundles q on Q with exact rows and columns 0 0 0 / q∗ F / q∗ F / F ˜ / q∗ F L 0 0 / F ˜ / q∗ F L 0 / 0 / 0

˜ ˜ ˜ for some vector bundles F and F on Q with rank F = rank F − 1. Recall that we want to prove the statement that for any short exact sequence of vector bundles the Chern polynomial of the bundle in the middle is equal to the product of the Chern polynomials of the other two bundles. In the above diagram we know that this is true for the columns by step 1 and for the top row by the inductive assumption; hence it must be true for the bottom row as well. So we have shown that c(q∗ F) = c(q∗ F ) · c(q∗ F ). It follows that q∗ c(F) = q∗ (c(F ) · c(F )) by proposition 10.2.6 (iv), and ﬁnally that c(F) = c(F ) · c(F ) as q∗ is injective by corollary 10.2.4. Remark 10.3.2. Of course proposition 10.3.1 can be split up into graded parts to obtain the equations ck (F) = ∑ ci (F ) · c j (F )

i+ j=k

196

Andreas Gathmann

for all k ≥ 0 and any exact sequence 0 → F → F → F → 0 of vector bundles on a scheme X. Note moreover that by deﬁnition the same relation s(F) = s(F ) · s(F ) then holds for the Segre classes. Example 10.3.3. In this example we will compute the Chern classes of the tangent bundle TX of X = Pn . By lemma 7.4.15 we have an exact sequence of vector bundles on X 0 → OX → OX (1)⊕(n+1) → TX → 0. Moreover proposition 10.2.6 (i) implies that c(OX ) = 1 and c(OX (1)) = 1 + H, where H is (the divisor class of) a hyperplane in X. So by proposition 10.3.1 it follows that c(TX ) = c(OX (1))n+1 /c(OX ) = (1 + H)n+1 , i.e. ck (TX ) =

n+1 k

· H k (where H k is the class of a linear subspace of X of codimension k).

Remark 10.3.4. Note that proposition 10.3.1 allows us to compute the Chern classes of any bundle F of rank r on a scheme X that has a ﬁltration 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fr−1 ⊂ Fr = F by vector bundles such that the quotients Li := Fi /Fi−1 are all line bundles (i.e. Fi has rank i for all i). In fact, in this case a recursive application of proposition 10.3.1 to the exact sequences 0 → Fi−1 → Fi → Li → 0 yields (together with proposition 10.2.6 (i)) c(F) = ∏(1 + Di )

i=1 r

where Di = c1 (Li ) is the divisor associated to the line bundle Li . Unfortunately, not every vector bundle admits such a ﬁltration. We will see now however that for computations with Chern classes we can essentially pretend that such a ﬁltration always exists. Lemma 10.3.5. (Splitting construction) Let F be a vector bundle of rank r on a scheme X. Then there is a scheme Y and a morphism f : Y → X such that (i) f admits push-forwards and pull-backs for Chow groups (in fact it will be an iterated projective bundle), (ii) the push-forward f∗ is surjective, (iii) the pull-back f ∗ is injective, (iv) f ∗ F has a ﬁltration by vector bundles 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fr−1 ⊂ Fr = f ∗ F such that the quotients Fi /Fi−1 are line bundles on Y . In other words, “every vector bundle admits a ﬁltration after pulling back to an iterated projective bundle”. Proof. We construct the morphism f by induction on rank F. There is nothing to do if rank F = 1. Otherwise set Y = P(F ∨ ) and let f : Y → X be the projection. Let L∨ ⊂ f ∗ F ∨ be the tautological line bundle on Y . Then we have an exact sequence of vector ˜ ˜ bundles 0 → F → f ∗ F → L → 0 on Y , where rank F = rank F − 1. Hence by the inductive ˜ assumption there is a morphism f : Y → Y such that f ∗ F has a ﬁltration (Fi ) with line bundle quotients. If we set f = f ◦ f it follows that we have an induced ﬁltration of f ∗ F on Y ˜ 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fr−1 = f ∗ F ⊂ f ∗ F

10.

Chern classes

197

with line bundle quotients. Moreover, f∗ is surjective and f ∗ is injective, as this is true for f by the inductive assumption and for f by corollary 10.2.4. Construction 10.3.6. (Splitting construction) Suppose one wants to prove a universal identity among Chern classes of vector bundles on a scheme X, e.g. the statement that ci (F) = 0 whenever i > rank F (see corollary 10.3.7 below). If the identity is invariant under pull-backs (which it essentially always is because of proposition 10.2.6 (iv)) then one can assume that the vector bundles in question have ﬁltrations with line bundle quotients. More precisely, pick a morphism f : Y → X as in lemma 10.3.5. We can then show the identity for the pulled-back bundle f ∗ F on Y , using the ﬁltration. As the pull-back f ∗ is injective and commutes with the identity we want to show, the identity then follows for F on X as well. (This is the same argument that we used already at the end of the proof of proposition 10.3.1.) Corollary 10.3.7. Let F be a vector bundle of rank r on a scheme X. Then ci (F) = 0 for all i > r. Proof. By the splitting construction 10.3.6 we can assume that F has a ﬁltration with line bundle quotients Li , i = 1, . . . , r. But then c(F) = ∏r (1+c1 (Li )) by remark 10.3.4, which i=1 obviously has no parts of degree bigger than r. Remark 10.3.8. This vanishing of Chern classes beyond the rank of the bundle is a property that is not shared by the Segre classes (see e.g. proposition 10.2.3 (ii)). This is one reason why Chern classes are usually preferred over Segre classes in computations (although they carry the same information). Remark 10.3.9. The splitting construction is usually formalized as follows. Let F be a vector bundle of rank r on a scheme X. We write formally c(F) = ∏(1 + αi ).

i=1 r

There are two ways to think of the α1 , . . . , αr : • The αi are just formal “variables” such that the k-th elementary symmetric polynomial in the αi is exactly ck (F). So any symmetric polynomial in the αi is expressible as a polynomial in the Chern classes of F in a unique way. • After having applied the splitting construction, the vector bundle F has a ﬁltration with line bundle quotients Li . Then we can set αi = c1 (Li ), and the decomposition c(F) = ∏r (1 + αi ) becomes an actual equation (and not just a formal one). i=1 The αi are usually called the Chern roots of F. Using the splitting construction and Chern roots, one can compute the Chern classes of almost any bundle that is constructed from other known bundles by standard operations: Proposition 10.3.10. Let X be a scheme, and let F and F be vector bundles with Chern roots (αi )i and (α j ) j , respectively. Then: (i) (ii) (iii) (iv) F ∨ has Chern roots (−αi )i . F ⊗ F has Chern roots (αi + α j )i, j . Sk F has Chern roots (αi1 + · · · + αik )i1 ≤···≤ik . Λk F has Chern roots (αi1 + · · · + αik )i1 <···<ik .

Proof. (i): If F has a ﬁltration 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fr = F with line bundle quotients Li = Fi /Fi−1 , then F ∨ has an induced ﬁltration 0 = (F/Fr )∨ ⊂ (F/Fr−1 )∨ ⊂ · · · ⊂ (F/F0 )∨ = F ∨ ∨ with line bundle quotients Li . (ii): If F and F have ﬁltrations 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fr = F and 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fs = F

198

Andreas Gathmann

with line bundle quotients Li := Fi /Fi−1 and Li := Fi /Fi−1 , then F ⊗ F has a ﬁltration 0 = F0 ⊗ F ⊂ F1 ⊗ F ⊂ · · · ⊂ Fr ⊗ F = F ⊗ F with quotients Li ⊗ F . But Li ⊗ F itself has a ﬁltration 0 = Li ⊗ F0 ⊂ Li ⊗ F1 ⊂ · · · ⊂ Li ⊗ Fs = Li ⊗ F with quotients Li ⊗ L j , so the result follows. (iii) and (iv) follow in the same way. Example 10.3.11. The results of proposition 10.3.10 can be restated using Chern classes instead of Chern roots. For example, (i) just says that ci (F ∨ ) = (−1)i ci (F). It is more difﬁcult to write down closed forms for the Chern classes in the cases (ii) to (iv). For example, if F = L is a line bundle, then c(F ⊗ L) = ∏(1 + (αi + α )) = ∑(1 + c1 (L))r−i ci (F)

i i

where r = rank F. So for 0 ≤ p ≤ r we have c p (F ⊗ L) = ∑

p

i=0

r−i ci (F) c1 (L) p−i . p−i

Also, from part (iv) it follows immediately that c1 (F) = c1 (Λr F). As a more complicated example, assume that F is a rank-2 bundle on a scheme X and let us compute the Chern classes of S3 F. Say F has Chern roots α1 and α2 , so that c1 (F) = α1 + α2 and c2 (F) = α1 α2 . Then by part (iii) a tedious but easy computation shows that c(S3 F) = (1 + 3α1 )(1 + 2α1 + α2 )(1 + α1 + 2α2 )(1 + 3α2 ) = 1 + 6c1 (F) + 10c2 (F) + 11c1 (F)2 + 30c1 (F)c2 (F) + 6c1 (F)3 + 9c2 (F)2 + 18c1 (F)2 c2 (F). Splitting this up into graded pieces one obtains the individual Chern classes, e.g. c4 (S3 F) = 9c2 (F)2 + 18c1 (F)2 c2 (F). Now that we have shown how to compute Chern classes let us discuss their geometric meaning. By far the most important property of Chern classes is that the “top Chern class” of a vector bundle (i.e. cr (F) if r = rank F) is the class of the zero locus of a section: Proposition 10.3.12. Let F be a vector bundle of rank r on an n-dimensional scheme X. Let s ∈ Γ(F) be a global section of F, and assume that its scheme-theoretic zero locus Z(s) has dimension n − r (as expected). Then [Z(s)] = cr (F) · [X] ∈ An−r (X). Proof. We will only sketch the proof; for details especially about multiplicities we refer to [F] section 14.1. We prove the statement by induction on r. Applying the splitting principle we may assume that there is an exact sequence 0→F →F →L→0 (∗)

of vector bundles on X, where L is a line bundle and rank F = rank F − 1. Now let s ∈ Γ(X, F) be a global section of F as in the proposition. Then s induces (i) a section l ∈ Γ(X, L), and (ii) a section s ∈ Γ(Z(l), F ) (i.e. “s is a section of F on the locus where the induced section on L vanishes”).

10.

Chern classes

199

Let us assume that l is not identically zero, and denote by i : Z(l) → X the inclusion morphism. Note that then i∗ [Z(s )] = cr−1 (F) · [Z(l)] by the induction hypothesis, and [Z(l)] = c1 (L) · [X] as the Weil divisor associated to a line bundle is just the zero locus of a section. Combining these results we get [Z(s)] = i∗ [Z(s )] = cr−1 (F) · c1 (L) · [X]. But applying proposition 10.3.1 to the exact sequence (∗) we get cr (F) = cr−1 (F ) · c1 (L), so the result follows. Remark 10.3.13. Proposition 10.3.12 is the generalization of our old statement that the ﬁrst Chern class of a line bundle (i.e. the divisor associated to a line bundle) is the zero locus of a (maybe rational) section of that bundle. In contrast to the line bundle case however, it is not clear that a section of the vector bundle exists that vanishes in the right codimension. This is why proposition 10.3.12 cannot be used as a deﬁnition for the top Chern class. Remark 10.3.14. There are analogous interpretations for the intermediate Chern classes ck (F) that we state without proof: let F be a vector bundle of rank r on a scheme X. Let s1 , . . . , sr+1−k be global sections of X, and assume that the (scheme-theoretic) locus Z ⊂ X where the sections s1 , . . . , sr+1−k are linearly dependent has codimension k in X (which is the expected codimension). Then [Z] = ck (F) · [X] ∈ A∗ (X). (For a proof of this statement see [F] example 14.4.1). Two special cases of this property are easy to see however: (i) In the case k = r we are reduced to proposition 10.3.12. (ii) In the case k = 1 the locus Z is just the zero locus of a section of Λr F, so we have [Z] = c1 (Λr F) = c1 (F) (the latter equality is easily checked using proposition 10.3.10 (iv)). Example 10.3.15. As an example of proposition 10.3.12 let us recalculate that there are 27 lines on a cubic surface X in P3 (see section 4.5). To be more precise, we will not reprove here that the number of lines in X is ﬁnite; instead we will assume that it is ﬁnite and just recalculate the number 27 under this assumption. Let G(1, 3) be the 4-dimensional Grassmannian variety of lines in P3 . As in construction 10.1.5 there is a tautological rank-2 subbundle F of the trivial bundle C4 whose ﬁber over a point [L] ∈ G(1, 3) (where L ⊂ P3 is a line) is precisely the 2-dimensional subspace of C4 whose projectivization is L. Dualizing, we get a surjective morphism of vector bundles (C4 )∨ → F ∨ that corresponds to restricting a linear function on C4 (or P3 ) to the line L. Taking the d-th symmetric power of this morphism we arrive at a surjective morphism Sd (C4 )∨ → Sd F ∨ that corresponds to restricting a homogeneous polynomial of degree d on P3 to L. Now let X = { f = 0} be a cubic surface. By what we have just said the polynomial f determines a section of S3 F ∨ whose set of zeros in G(1, 3) is precisely the set of lines that lie in X (i.e. the set of lines on which f vanishes). So assuming that this set is ﬁnite we see by proposition 10.3.12 that the number of lines in the cubic surface X is the degree of the cycle c4 (S3 F ∨ ) on G(1, 3). To compute this number note that by example 10.3.11 we have c4 (S3 F ∨ ) = 9c2 (F ∨ )2 + 18c1 (F ∨ )2 c2 (F ∨ ), so that it remains to compute the numbers c2 (F ∨ )2 and c1 (F ∨ )2 c2 (F ∨ ). There are general rules (called “Schubert calculus”) how to compute such intersection products on Grassmannian varieties, but in this case we can also compute the result directly in a way similar to that in example 9.4.9: (i) By exactly the same reasoning as above, c2 (F ∨ ) = c2 (S1 F ∨ ) is the locus of all lines in P3 that are contained in a given plane.

200

Andreas Gathmann

(ii) The class c1 (F ∨ ) = c1 (Λ2 F ∨ ) is (by deﬁnition of the exterior product, see also remark 10.3.14) the locus of all lines L ⊂ P3 such that two given linear equations f1 , f2 on P4 become linearly dependent when restricted to the line. This means that f1 |L and f2 |L must have their zero at the same point of L. In other words, L intersects Z( f1 , f2 ), which is a line. In summary, c1 (F ∨ ) is just the class of lines that meet a given line in P3 . Using these descriptions we can now easily compute the required intersection products: c2 (F ∨ )2 is the number of lines that are contained in two given planes in P3 , so it is 1 (the line must precisely be the intersection line of the two planes). Moreover, c1 (F ∨ )2 c2 (F ∨ ) is the number of lines intersecting two given lines and lying in a given plane, i.e. the number of lines through two points in a plane, which is 1. Summarizing, we get that the number of lines on a cubic surface is c4 (S3 F ∨ ) = 9c2 (F ∨ )2 + 18c1 (F ∨ )2 c2 (F ∨ ) = 9 · 1 + 18 · 1 = 27. Remark 10.3.16. The preceding example 10.3.15 shows very well how enumerative problems can be attacked in general. By an enumerative problem we mean that we want to count the number of curves in some space with certain conditions (e.g. lines through two points, lines in a cubic surface, plane conics through 5 points, and so on). Namely: (i) Find a complete (resp. compact) “moduli space” M whose points correspond to the curves one wants to study (in the above example: the Grassmannian G(1, 3) that parametrizes lines in P3 ). (ii) Every condition that one imposes on the curves (passing through a point, lying in a given subvariety, . . . ) corresponds to some intersection-theoretic cycle on M — a divisor, a combination of Chern classes, or something else. (iii) If the expected number of curves satisfying the given conditions is ﬁnite then the intersection product of the cycles in (ii) will have dimension 0. As M is complete the degree of this zero-cycle is a well-deﬁned integer. It is called the virtual solution to the enumerative problem. Note that this number is well-deﬁned even if the actual number of curves satisfying the given conditions is not ﬁnite. (iv) It is now a different (and usually more difﬁcult, in any case not an intersectiontheoretic) problem to ﬁgure out whether the actual number of curves satisfying the given conditions is ﬁnite or not, and if so whether they are counted in the intersection product of (iii) with the scheme-theoretic multiplicity 1. If this is the case then the solution of (iii) is said to be enumerative (and not only virtual). For example, we have shown in section 4.5 that the number 27 computed intersectiontheoretically in example 10.3.15 is actually enumerative for any smooth cubic surface X. 10.4. Statement of the Hirzebruch-Riemann-Roch theorem. As a ﬁnal application of Chern classes we will now state and sketch a proof of the famous Hirzebruch-RiemannRoch theorem that is a vast and very useful generalization (yet still not the most general version) of the Riemann-Roch theorem (see section 7.7, in particular remark 7.7.7). As usual the goal of the Riemann-Roch type theorems is to compute the dimension h0 (X, F ) of the space of global sections of a sheaf F on a scheme X, in the case at hand of a vector bundle on a smooth projective scheme X. As we have already seen in the case where X is a curve and F a line bundle there is no easy general formula for this number unless you add some “correction term” (that was −h1 (X, F ) in the case of curves). The same is true in higher dimensions. Here the Riemann-Roch theorem will compute the Euler characteristic of F : Deﬁnition 10.4.1. Let F be a coherent sheaf on a projective scheme X. Then the dimensions hi (X, F ) = dim H i (X, F ) are all ﬁnite by theorem 8.4.7 (i). We deﬁne the Euler

10.

Chern classes

201

characteristic of F to be the integer χ(X, F ) := ∑ (−1)i hi (X, F ).

i≥0

(Note that the sum is ﬁnite as

hi (X, F

) = 0 for i > dim X.)

The “left hand side” of the Hirzebruch-Riemann-Roch theorem will just be χ(X, F ); this is the number that we want to compute. Recall that there were many “vanishing theorems”, e.g. hi (X, F ⊗ OX (d)) = 0 for i > 0 and d 0 by theorem 8.4.7 (ii). So in the cases when such vanishing theorems apply the theorem will actually compute the desired number h0 (X, F ). The “right hand side” of the Hirzebruch-Riemann-Roch theorem is an intersectiontheoretic expression that is usually easy to compute. It is a certain combination of the Chern (resp. Segre) classes of the bundle F (corresponding to the locally free sheaf F ) and the tangent bundle TX of X. These combinations will have rational coefﬁcients, so we have to tensor the Chow groups with Q (i.e. we consider formal linear combinations of subvarieties with rational coefﬁcients instead of integer ones). Deﬁnition 10.4.2. Let F be a vector bundle of rank r with Chern roots α1 , . . . , αr on a scheme X. Then we deﬁne the Chern character ch(F) : A∗ (X) ⊗ Q → A∗ (X) ⊗ Q to be ch(F) = ∑ exp(αi )

i=1 r

and the Todd class td(F) : A∗ (X) ⊗ Q → A∗ (X) ⊗ Q to be td(F) = ∏ αi , i=1 1 − exp(−αi )

r

where the expressions in the αi are to be understood as formal power series, i.e. 1 1 exp(αi ) = 1 + αi + α2 + α3 + · · · i 2 6 i and αi 1 1 = 1 + αi + α2 + · · · . 1 − exp(−αi ) 2 12 i Remark 10.4.3. As usual we can expand the deﬁnition of ch(F) and td(F) to get symmetric polynomials in the Chern roots which can then be written as polynomials (with rational coefﬁcients) in the Chern classes ci = ci (F) of F. Explicitly, 1 1 ch(F) = r + c1 + (c2 − 2c2 ) + (c3 − 3c1 c2 + 3c3 ) + · · · 1 2 6 1 1 1 2 1 and td(F) = 1 + c1 + (c1 + c2 ) + c1 c2 + · · · . 2 12 24 Remark 10.4.4. If 0 → F → F → F → 0 is an exact sequence of vector bundles on X then the Chern roots of F are just the union of the Chern roots of F and F . So we see that ch(F) = ch(F ) + ch(F ) and td(F) = td(F ) · td(F ). We can now state the Hirzebruch-Riemann-Roch theorem: Theorem 10.4.5. (Hirzebruch-Riemann-Roch theorem) Let F be a vector bundle on a smooth projective variety X. Then χ(X, F) = deg(ch(F) · td(TX )) where deg(α) denotes the degree of the dimension-0 part of the (non-homogeneous) cycle α.

202

Andreas Gathmann

Before we sketch a proof of this theorem in the next section let us consider some examples. Example 10.4.6. Let F = L be a line bundle on a smooth projective curve X of genus g. Then χ(X, L) = h0 (X, L) − h1 (X, L). On the right hand side, the dimension-0 part of ch(L) · td(TX ), i.e. its codimension-1 part, is equal to

1 deg(ch(L) · td(TX )) = deg((1 + c1 (L))(1 + 2 c1 (TX )))

by remark 10.4.3 by corollary 7.6.6

= deg(c1 (L) −

1 2 c1 (ΩX ))

1 = deg L − 2 (2g − 2) = deg L + 1 − g,

so we are recovering our earlier Riemann-Roch theorem of corollary 8.3.3. Example 10.4.7. If F is a vector bundle of rank r on a smooth projective curve X then we get in the same way h0 (X, F) − h1 (X, F) = deg(ch(F) · td(TX )) 1 = deg((r + c1 (F))(1 + c1 (TX ))) 2 = deg c1 (F) + r(1 − g). Example 10.4.8. Let L = OX (D) be a line bundle on a smooth projective surface X corresponding to a divisor D. Now the dimension-0 part of the right hand side has codimension 2, so the Hirzebruch-Riemann-Roch theorem states that h0 (X, L) − h1 (X, L) + h2 (X, L) = deg(ch(F) · td(TX )) = deg 1 1 + c1 (L) + c1 (L)2 2 1 1 1 + c1 (TX ) + (c1 (TX )2 + c2 (TX )) 2 12

1 K 2 + c2 (TX ) = D · (D − KX ) + X . 2 12 Note that:

2 (i) The number χ(X, OX ) = X 12 X is an invariant of X that does not depend on the line bundle. The Hirzebruch-Riemann-Roch theorem implies that it is always 2 an integer, i.e. that KX + c2 (TX ) is divisible by 12 (which is not at all obvious from the deﬁnitions). (ii) If X has degree d and L = OX (n) for n 0 then h1 (X, L) = h2 (X, L) = 0 by 2 = dn2 , so we get theorem 8.4.7 (ii). Moreover we then have D

K 2 +c (T )

K 2 + c2 (TX ) d 2 1 n + (H · KX ) · n + X 2 2 12 where H denotes the class of a hyperplane (restricted to X). In other words, we have just recovered proposition 6.1.5 about the Hilbert function of X. Moreover, we have identiﬁed the non-leading coefﬁcients of the Hilbert polynomial in terms of intersection-theoretic data. h0 (X, OX (n)) = Example 10.4.9. The computation of example 10.4.8 works for higher-dimensional varieties as well: let X be a smooth projective N-dimensional variety of degree d and consider the line bundle L = OX (n) on X for n 0. We see immediately that the codimension-N part of ch(OX (n)) · td(TX ) is a polynomial in n of degree N with leading coefﬁcient 1 d N c1 (L)N = n , N! N!

10.

Chern classes

203

which reproves proposition 6.1.5 (for smooth X). Moreover, we can identify the other coefﬁcients of the Hilbert polynomial in terms of intersection-theoretic expressions involving the characteristic classes of the tangent bundle of X. Example 10.4.10. Let F = OX (d) be a line bundle on X = Pn . Then we can compute both sides of the Hirzebruch-Riemann-Roch theorem explicitly and therefore prove the theorem in this case: As for the left hand side, proposition 8.4.1 implies that h0 (X, OX (d)) = n+d if d ≥ 0, n n hn (X, O (d)) = (−1)n −d−1 χ(X, OX (d)) = (−1) if d ≤ −n − 1, X n 0 otherwise. Note that this means in fact in all cases that χ(X, OX (d)) = n+d . n

As for the right hand side let us ﬁrst compute the Todd class of TX . By the Euler sequence 0 → OX → OX (1)⊕(n+1) → TX → 0 of lemma 7.4.15 together with the multiplicativity of Chern classes (see proposition 10.3.1) we see that the Chern classes (and hence the Todd class) of TX are the same as those of OX (1)⊕(n+1) . But the Chern roots of the latter bundle are just n + 1 times the class H of a hyperplane, so it follows that td(TX ) = H n+1 . (1 − exp(−H))n+1

As the Chern character of OX (d) is obviously exp(dH) we conclude that the right hand side of the Hirzebruch-Riemann-Roch theorem is the H n -coefﬁcient of H n+1 exp(dH) . (1 − exp(−H))n+1 But this is equal to the residue resH=0 exp(dH) dH, (1 − exp(−H))n+1

1 1−x

which we can compute using the substitution x = 1 − exp(−H) (so exp(H) = dH 1 dx = 1−x ): resH=0 exp(dH) (1 − x)−d−1 dH = resx=0 dx. (1 − exp(−H))n+1 xn+1 −d − 1 n+d = n n

and

This number is equal to the xn -coefﬁcient of (1 − x)−d−1 , which is simply (−1)n

in agreement with what we had found for the left hand side of the Hirzebruch-RiemannRoch theorem above. So we have just proven the theorem for line bundles on Pn . 10.5. Proof of the Hirzebruch-Riemann-Roch theorem. Finally we now want to give a very short sketch of proof of the Hirzebruch-Riemann-Roch theorem 10.4.5, skipping several subtleties from commutative algebra. The purpose of this section is just to give an idea of the proof, and in particular to show why the rather strange-looking Todd classes come into play. For a more detailed discussion of the proof or more general versions see [F] chapter 15.

204

Andreas Gathmann

The proof of the theorem relies heavily on certain constructions being additive (or otherwise well-behaved) on exact sequences of vector bundles. Let us formalize this idea ﬁrst. Deﬁnition 10.5.1. Let X be a scheme. The Grothendieck group of vector bundles K ◦ (X) on X is deﬁned to be the group of formal ﬁnite sums ∑i ai [Fi ] where ai ∈ Z and the Fi are vector bundles on X, modulo the relations [F] = [F ] + [F ] for every exact sequence 0 → F → F → F → 0. (Of course we then also have ∑r (−1)i [Fi ] = 0 for every exact i=1 sequence 0 → F1 → F2 → · · · → Fr → 0.) Example 10.5.2. Deﬁnition 10.5.1 just says that every construction that is additive on exact sequences passes to the Grothendieck group. For example: (i) If X is projective then the Euler characteristic of a vector bundle (see deﬁnition 10.4.1) is additive on exact sequences by the long exact cohomology sequence of proposition 8.2.1. Hence the Euler characteristic can be thought of as a homomorphism of Abelian groups χ : K ◦ (X) → Z, χ([F]) = χ(X, F).

(ii) The Chern character of a vector bundle is additive on exact sequences remark 10.4.4. So we get a homomorphism ch : K ◦ (X) → A∗ (X) ⊗ Q, ch([F]) = ch(F).

(It can in fact be shown that this homomorphism gives rise to an isomorphism K ◦ (X) ⊗ Q → A∗ (X) ⊗ Q if X is smooth; see [F] example 15.2.16(b). We will not need this however in our proof.) (iii) Let X be a smooth projective variety. For the same reason as in (ii) the right hand side of the Hirzebruch-Riemann-Roch theorem gives rise to a homomorphism τ : K ◦ (X) → A∗ (X) ⊗ Q, τ(F) = ch(F) · td(TX ).

In particular, by (i) and (iii) we have checked already that both sides of the HirzebruchRiemann-Roch theorem are additive on exact sequences (which is good). So to prove the theorem we only have to check it on a set of generators for K ◦ (X). To use this to our advantage however we ﬁrst have to gather more information about the structure of the Grothendieck groups. We will need the following lemma of which we can only sketch the proof. Lemma 10.5.3. Let X be a smooth projective scheme. Then for every coherent sheaf F on X there is an exact sequence 0 → Fr → Fr−1 → · · · → F0 → F → 0 where the Fi are vector bundles (i.e. locally free sheaves). We say that “every coherent sheaf has a ﬁnite locally free resolution”. Moreover, if X = Pn then the Fi can all be chosen to be direct sums of line bundles OX (d) for various d. Proof. By a repeated application of lemma 8.4.6 we know already that there is a (possibly inﬁnite) exact sequence · · · → Fr → · · · → F1 → F0 → F → 0. Now one can show that for an n-dimensional smooth scheme the kernel K of the morphism Fr−1 → Fr−2 is always a vector bundle (see [F] B.8.3). So we get a locally free resolution 0 → K → Fr−1 → Fr−2 → · · · → F0 → F → 0 as required.

10.

Chern classes

205

If X = Pn with homogeneous coordinate ring S = k[x0 , . . . , xn ] then one can show that a coherent sheaf F on X is nothing but a graded S-module M (in the same way that a coherent sheaf on an afﬁne scheme Spec R is given by an R-module). By the famous Hilbert syzygy theorem (see [EH] theorem III-57) there is a free resolution of M 0→

M

i

Sn,i → · · · →

M

i

S1,i →

M

i

S0,i → M → 0

where each S j,i is isomorphic to S, with the grading shifted by some constants a j,i . This means exactly that we have a locally free resolution 0→ of F . Corollary 10.5.4. The Hirzebruch-Riemann-Roch theorem 10.4.5 is true for any vector bundle on Pn . Proof. By lemma 10.5.3 (applied to X = Pn and a vector bundle F ) the Grothendieck group K ◦ (Pn ) is generated by the classes of the line bundles OPn (d) for d ∈ Z. As we have already checked the Hirzebruch-Riemann-Roch theorem for these bundles in example 10.4.10 the statement follows by the remark at the end of example 10.5.2. Remark 10.5.5. To study the Hirzebruch-Riemann-Roch theorem for general smooth projective X let i : X → Pn be an embedding of X in projective space and consider the following diagram: K ◦ (X) A(X) ⊗ Q

τ i∗

M

i

OX (an,i ) → · · · →

M

i

OX (a1,i ) →

M

i

OX (a0,i ) → F → 0

/ K ◦ (Pn ) / A(Pn ) ⊗ Q

τ

χ

/ Z _ / Q.

i∗

deg

Let us ﬁrst discuss the right square. The homomorphisms χ and τ are explained in example 10.5.2, and deg denotes the degree of the dimension-0 part of a cycle class. The Hirzebruch-Riemann-Roch theorem for Pn of corollary 10.5.4 says precisely that this right square is commutative. Now consider the left square. The homomorphism τ is as above, and the i∗ in the bottom row is the proper push-forward of cycles of corollary 9.2.12. We have to explain the pushforward i∗ in the top row. Of course we would like to deﬁne i∗ [F] = [i∗ F] for any vector bundle F on X, but we cannot do this directly as i∗ F is not a vector bundle but only a coherent sheaf on Pn . So instead we let 0 → Fr → Fr−1 → · · · → F0 → i∗ F → 0 be a locally free resolution of the coherent sheaf i∗ F on Pn and set i∗ : K ◦ (X) → K ◦ (Pn ), i∗ ([F]) =

r

(∗)

k=0

∑ (−1)k [Fk ].

One can show that this is indeed a well-deﬁned homomorphism of groups (i.e. that this deﬁnition does not depend on the choice of locally free resolution), see [F] section B.8.3. But in fact we do not really need to know this: we do know by the long exact cohomology sequence applied to (∗) that χ(X, F) =

k=0

∑ (−1)k χ(Pn , Fk ),

r

206

Andreas Gathmann

so it is clear that at least the composition χ ◦ i∗ does not depend on the choice of locally free resolution. The Hirzebruch-Riemann-Roch theorem on X is now precisely the statement that the outer rectangle in the above diagram is commutative. As we know already that the right square is commutative, it sufﬁces therefore to show that the left square is commutative as well (for any choice of locally free resolution as above), i.e. that

k=0

∑ (−1)k ch(Fk ) · td(TPr ) = i∗ (ch(F) · td(TX )).

r

As the Todd class is multiplicative on exact sequences by remark 10.4.4 we can rewrite this using the projection formula as

k=0

∑ (−1)k ch(Fk ) = i∗ td(NX/Pn ) .

r

ch(F)

Summarizing our ideas we see that to prove the general Hirzebruch-Riemann-Roch theorem it sufﬁces to prove the following proposition (for Y = Pn ): Proposition 10.5.6. Let i : X → Y be a closed immersion of smooth projective schemes, and let F be a vector bundle on X. Then there is a locally free resolution 0 → Fr → Fr−1 → · · · → F0 → i∗ F → 0 of the coherent sheaf i∗ F on Y such that

r

k=0

∑ (−1)k ch(Fk ) = i∗ td(NX/Y )

ch(F)

in A∗ (Y ) ⊗ Q. Example 10.5.7. Before we give the general proof let us consider an example where both sides of the equation can be computed explicitly: let X be a smooth scheme, E a vector bundle of rank r on X, and Y = P(E ⊕ OX ). The embedding i : X → Y is given by X = P(0⊕ OX ) → P(E ⊕ OX ). In other words, X is just “the zero section of a projective bundle”. The special features of this particular case that we will need are: (i) There is a projection morphism p : Y → X such that p ◦ i = id. (ii) X is the zero locus of a section of a vector bundle on Y : consider the exact sequence 0 → S → p∗ (E ⊕ OX ) → Q → 0 (∗) on Y , where S is the tautological subbundle of construction 10.1.5. The vector bundle Q (which has rank r) is usually called the universal quotient bundle. Note that we have a global section of p∗ (E ⊕ OX ) by taking the point (0, 1) in every ﬁber (i.e. 0 in the ﬁber of E and 1 in the ﬁber of OX ). By deﬁnition of S the induced section s ∈ Γ(Q) vanishes precisely on P(0 ⊕ OX ) = X. (iii) Restricting (∗) to X (i.e. pulling the sequence back by i) we get the exact sequence 0 → i∗ S → E ⊕ OX → i∗ Q → 0 (∗) on X. Note that the ﬁrst morphism is given by λ → (0, λ) by construction, so we conclude that i∗ Q = E. (iv) As X is given in Y as the zero locus of a section of Q, we see from example 10.1.6 that the normal bundle of X in Y is just NX/Y = i∗ Q = E. Let us now check proposition 10.5.6 in this case. Note that away from the zero locus of s there is an exact sequence 0 → OY → Q → Λ2 Q → Λ3 Q → · · · → Λr−1 Q → Λr Q → 0

·s ∧s ∧s ∧s

10.

Chern classes

207

of vector bundles (which follows from the corresponding statement for vector spaces). Dualizing and tensoring this sequence with p∗ F we get the exact sequence 0 → p∗ F ⊗ Λr Q∨ → p∗ F ⊗ Λr−1 Q∨ → · · · → p∗ F ⊗ Q∨ → p∗ F → 0 again on Y \Z(s) = Y \X. Let us try to extend this exact sequence to all of Y . Note that the last morphism p∗ F ⊗ Q∨ → p∗ F is just induced by the evaluation morphism s : Q∨ → OY , so its cokernel is precisely the sheaf (p∗ F)|Z(s) = i∗ F. One can show that the other stages of the sequence remain indeed exact (see [F] B.3.4), so we get a locally free resolution 0 → p∗ F ⊗ Λr Q∨ → p∗ F ⊗ Λr−1 Q∨ → · · · → p∗ F ⊗ Q∨ → p∗ F → i∗ F → 0 on Y . (This resolution is called the Koszul complex.) So what we have to check is that

k=0

∑ (−1)k ch(p∗ F ⊗ Λk Q∨ ) = i∗ td(i∗ Q) .

r

ch(F)

But note that i∗ ch(p∗ F) ch(p∗ F)cr (Q) ch(F) = · i∗ [X] = td(i∗ Q) td(Q) td(Q)

by the projection formula and proposition 10.3.12. So by the additivity of Chern characters it sufﬁces to prove that

k=0

∑ (−1)k ch(Λk Q∨ ) = td(Q) .

r

r

cr (Q)

But this is easily done: if α1 , . . . , αr are the Chern roots of Q then the left hand side is

k=0

∑ (−1)k

r

i1 <···<ik

∑

exp(−αi1 − · · · − αik ) = ∏(1 − exp(−αi )) = α1 · · · αr · ∏

i=1

1 − exp(−αi ) , αi i=1

r

which equals the right hand side. It is in fact this formal identity that explains the appearance of Todd classes in the Hirzebruch-Riemann-Roch theorem. Using the computation of this special example we can now give the general proof of the Hirzebruch-Riemann-Roch theorem. Proof. (of proposition 10.5.6) We want to reduce the proof to the special case considered in example 10.5.7. Let i : X → Y be any inclusion morphism of smooth projective varieties. We denote by M be the blow-up of Y × P1 in X × {0}. The smooth projective scheme M comes together with a projection morphism q : M → P1 . Its ﬁbers q−1 (P) for P = 0 are all isomorphic to Y . The ﬁber q−1 (0) however is reducible with two smooth components: one of them (the exceptional hypersurface of the blow-up) is the projectivized normal bundle of X × {0} in ˜ Y × P1 by example 10.1.6, and the other one is simply the blow-up Y of Y in X. We are particularly interested in the ﬁrst component. As the normal bundle of X × {0} in Y × P1 is NX/Y ⊕ OX this component is just the projective bundle P := P(NX/Y ⊕ OX ) on X. Note that there is an inclusion of the space X × P1 in M that corresponds to the given inclusion X ⊂ Y in the ﬁbers q−1 (P) for P = 0, and to the “zero section inclusion” X ⊂ P(NX/Y ⊕ OX ) = P as in example 10.5.7 in the ﬁber q−1 (0). The following picture illustrates the geometric situation.

208

Andreas Gathmann

X X P= IP ( N X /Y + O X ) X M

Y

~ Y

Y

q

0

IP 1

The idea of the proof is now simply the following: we have to prove an equality in the Chow groups, i.e. modulo rational equivalence. The ﬁbers q−1 (0) and q−1 (∞) are rationally equivalent as they are the zero resp. pole locus of a rational function on the base P1 , so they are effectively “the same” for intersection-theoretic purposes. But example 10.5.7 shows that the proposition is true in the ﬁber q−1 (0), so it should be true in the ﬁber q−1 (∞) as well. To be more precise, let F be a sheaf on X as in the proposition. Denote by pX : X ×P1 → X the projection, and by iX : X × P1 → M the inclusion discussed above. Then iX ∗ p∗ F is X a coherent sheaf on M that can be thought of as “the sheaf F on X in every ﬁber of q”. By lemma 10.5.3 we can choose a locally free resolution 0 → Fr → Fr−1 → · · · → F0 → iX ∗ p∗ F → 0 X (1)

on M. Note that the divisor [0] − [∞] on P1 is equivalent to zero by example 9.1.9. So it follows that

k=0

8

∑ (−1)k ch(Fi ) · q∗ ([0] − [∞]) = 0

r

˜ in A∗ (M)⊗Q. Now by deﬁnition of the pull-back we have q∗ [0] = [Y ]+[P] and q∗ [∞] = [Y ], so we get the equality ˜ ∑ (−1)k ch(Fi |Y˜ ) · [Y ] + ∑ (−1)k ch(Fi |P ) · [P] = ∑ (−1)k ch(Fi |Y ) · [Y ]

k=0 k=0 r r r

(2)

k=0

˜ in A∗ (M) ⊗ Q. But note that the restriction to Y of the sheaf iX ∗ p∗ F in (1) is the zero sheaf X 1 ∩ Y = 0 in M. So the sequence ˜ / as X × P 0 → Fr |Y → · · · → F1 |Y → F0 |Y → 0 ˜ ˜ ˜ is exact, which means that the ﬁrst sum in (2) vanishes. The second sum in (2) is precisely ch(F) td(N ) · [X] by example 10.5.7. So we conclude that

X/Y

k=0

∑ (−1)k ch(Fi |Y ) · [Y ] = td(NX/Y ) · [X]

r

ch(F)

in A∗ (M) ⊗ Q. Pushing this relation forward by the (proper) projection morphism from M to Y then gives the desired equation. This completes the proof of the Hirzebruch-Riemann-Roch theorem 10.4.5.

10.

Chern classes

209

Remark 10.5.8. Combining proposition 10.5.6 with remark 10.5.5 we see that we have just proven the following statement: let f : X → Y be a closed immersion of smooth projective schemes, and let F be a coherent sheaf on X. Then there is a locally free resolution 0 → Fr → Fr−1 → · · · → F0 → f∗ F → 0 of the coherent sheaf f∗ F on Y such that

k=0

∑ (−1)k ch(Fk ) · td(TY ) = f∗ (ch(F) · td(TX )) ∈ A∗ (Y ) ⊗ Q.

ch( f∗ F) · td(TY ) = f∗ (ch(F) · td(TX )).

r

This is often written as In other words, “the push-forward f∗ commutes with the operator τ of example 10.5.2 (iii)”. It is the statement of the Grothendieck-Riemann-Roch theorem that this relation is actually true for any proper morphism f of smooth projective schemes (and not just for closed immersions). See [F] section 15 for details on how to prove this. The Grothendieck-Riemann-Roch theorem is probably one of the most general Riemann-Roch type theorems that one can prove. The only further generalization one could think of is to singular schemes. There are some such generalizations to mildly singular schemes; see [F] section 18 for details. 10.6. Exercises. Exercise 10.6.1. Let X = P1 , and for n ∈ Z let Fn be the projective bundle Fn = P(OX ⊕ OX (n)). Let p : Fn → X be the projection morphism. The surfaces Fn are called Hirzebruch surfaces. (i) Show that F0 ∼ P1 × P1 , and Fn ∼ F−n for all n. = = (ii) Show that all ﬁbers p−1 (P) ⊂ Fn for P ∈ X are rationally equivalent as 1-cycles on Fn . Denote this cycle by D ∈ A1 (Fn ). n (iii) Now let n ≥ 0. Show that the global section (1, x0 ) of OX ⊕ OX (n) (where x0 , x1 are the homogeneous coordinates of X) determines a morphism s : X → Fn . Denote by C ∈ A1 (Fn ) the class of the image curve s(X). (iv) Again for n ≥ 0, show that A0 (Fn ) ∼ Z and A1 (Fn ) = Z · [C] ⊕ Z · [D]. Compute = the intersection products C2 , D2 , and C · D, arriving at a B´ zout style theorem for e the surfaces Fn . Exercise 10.6.2. Let F and F be two rank-2 vector bundles on a scheme X. Compute the Chern classes of F ⊗ F in terms of the Chern classes of F and F . Exercise 10.6.3. Let F be a vector bundle of rank r on a scheme X, and let p : P(F) → X be the projection. Prove that

r−1 Dr + DF · p∗ c1 (F) + · · · + p∗ cr (F) = 0, F

where DF is the Cartier divisor associated to the line bundle OP(F) (1). Exercise 10.6.4. Let X ⊂ P4 be the intersection of two general quadric hypersurfaces. (i) Show that one expects a ﬁnite number of lines in X. (ii) If there is a ﬁnite number of lines in X, show that this number is 16 (as usual counted with multiplicities (which one expects to be 1 for general X)). Exercise 10.6.5. A circle in the plane P2 is deﬁned to be a conic passing through the two C points (1 : ±i : 0). Why is this called a circle? How many circles are there in the plane that are tangent to

210

Andreas Gathmann

(i) (ii) (iii) (iv)

three circles two circles and a line one circle and two lines three lines

in general position? (Watch out for possible non-enumerative contributions in the intersection products you consider.) If you are interested, try to ﬁnd out the answer to the above questions over R (and the “usual” deﬁnition of a circle). Exercise 10.6.6. Let X ⊂ P4 be a smooth quintic hypersurface, i.e. the zero locus of a homogeneous polynomial of degree 5. (i) Show that one expects a ﬁnite number of lines in X, and that this expected number is then 2875. 5 5 (ii) Show that the number of lines on the special quintic X = {x0 + · · · + x4 = 0} is not ﬁnite. This illustrates the fact that the intersection-theoretic computations will only yield virtual numbers in general. (In fact one can show that the number of lines on a general quintic hypersurface in P4 is ﬁnite and that the computation of (i) then yields the correct answer.)

2 Exercise 10.6.7. Let X = P1 × P1 . Compute the number KX + c2 (TX ) directly and check that it is divisible by 12 (see example 10.4.8).

Exercise 10.6.8. Let X and Y be isomorphic smooth projective varieties. Use the Hirzebruch-Riemann-Roch theorem 10.4.5 to prove that the constant coefﬁcients of the Hilbert polynomials of X and Y agree, whereas the non-constant coefﬁcients will in general be different.

References

211

R EFERENCES

[AM] [D] [Do] [EGA] [EH] [F] [G] [Gr] [GP] [H] [Ha] [K] [M1] [M2] [Ma] [R] [S1] [S2] [S3] [V] M. Atiyah, I. Macdonald, Introduction to commutative algebra, Addison-Wesley (1969). O. Debarre, Introduction a la g´ om´ trie alg´ brique, notes for a class taught at the University Louis e e e ` Pasteur in Strasbourg, available at http://www-irma.u-strasbg.fr/˜debarre/. I. Dolgachev, Introduction to algebraic geometry, notes for a class taught at the University of Michigan, available at http://www.math.lsa.umich.edu/˜idolga/lecturenotes.html. A. Grothendieck, J. Dieudonn´ , El´ ments de g´ om´ trie alg´ brique, Publications Math´ matiques IHES e ´e e e e e (various volumes). D. Eisenbud, J. Harris, The geometry of schemes, Springer Graduate Texts in Mathematics 197 (2000). W. Fulton, Intersection theory, Springer-Verlag (1984). A. Gathmann, Algebraic geometry, notes for a class taught at Harvard University (1999-2000), available at http://www.math.ias.edu/˜andreas/pub/. G.-M. Greuel, Introduction to algebraic geometry, notes for a class taught at the University of Kaiserslautern, Mathematics International Lecture Notes, University of Kaiserslautern (1997–1998). G.-M. Greuel, G. Pﬁster, A Singular introduction to commutative algebra, Springer-Verlag (2002). R. Hartshorne, Algebraic geometry, Springer Graduate Texts in Mathematics 52 (1977). J. Harris, Algebraic geometry, Springer Graduate Texts in Mathematics 133 (1992). F. Kirwan, Complex algebraic curves, London Mathematical Society Student Texts 23, Cambridge University Press (1992). D. Mumford, The red book of varieties and schemes, Springer Lecture Notes in Mathematics 1358 (1988). D. Mumford, Algebraic geometry I — complex projective varieties, Springer Classics in Mathematics (1976). H. Matsumura, Commutative algebra, W. A. Benjamin Publishers (1970). M. Reid, Undergraduate algebraic geometry, London Mathematical Society Student Texts 12, Cambridge University Press (1988). I. Shafarevich, Basic algebraic geometry, Springer Study Edition (1977). I. Shafarevich, Basic algebraic geometry I — Varieties in projective space, Springer-Verlag, Berlin (1994). I. Shafarevich, Basic algebraic geometry II — Schemes and complex manifolds, Springer-Verlag, Berlin (1994). R. Vakil, Introduction to algebraic geometry, notes for a class taught at MIT (1999), available at http://math.stanford.edu/˜vakil/725/course.html.

Note: This is a very extensive list of literature of varying usefulness. Here is a short recommendation which of the references you might want to use for what: • For a general reference on the commutative algebra background, see [AM]. • For commutative algebra problems involving computational aspects, see [GP]. • For motivational aspects, examples, and a generally “fairy-tale” style introduction to the classical theory of algebraic geometry (no schemes) without much theoretical background, see [Ha], or maybe [S1] and [S2]. • For motivations and examples concerning scheme theory, see [EH], or maybe [S1] and [S3]. • For a good book that develops the theory, but largely lacks motivations and examples (especially in chapters II and III), see [H]. You should not try to read the “hard-core” parts of this book without some motivational background. • For intersection theory and Chern classes the best reference is [F]. • For the ultimate reference (“if it is not proven there, it must be wrong”), see [EGA]. Warning: this is unreadable if you do not have a decent background in algebraic geometry yet, and it is close to being unreadable even if you do.

212

Andreas Gathmann

I NDEX

genus of a curve, 2