Intersection Theory Class 1: Ravi Vakil
Intersection Theory Class 1: Ravi Vakil
RAVI VAKIL
C ONTENTS
1. Welcome! 1
2. Examples 2
3. Strategy 5
1. W ELCOME !
Scheduling. We are tentatively considering switching the times of the course to some-
thing like Mondays and Wednesdays 9:30-10:50.
About the subject. We’ll be using Fulton’s book Intersection Theory. There are currently
copies available at the bookstore, and it’s about $45, which is cheap. I should point out
that it’s one of the few paperbacks you’ll see in that Springer series; the author is one of
the few people with both the stature and the interest to force Springer to put out a cheaper
paperback version.
Intersection theory deals loosely with the following sort of problem: intersect two
things of some codimension, get something of expected codimension. We’ll basically
construct something that looks like homology, and later something that looks like coho-
mology.
In algebraic geometry: can deal with singular things. Also over other fields. Even in
holomorphic category, can get more refined information: 2 points on elliptic curve are the
same in homology, but not in “Chow”.
Before the 1970’s, the field was a mess — there were a great deal of painful ad hoc
constructions that people used. Fulton and MacPherson understood the right way to
Date: Monday, September 27, 2004.
1
describe it, and the result is in this book. In particular, the first 8 chapters are the heart of
the subject, and subsume volumes and volumes of earlier work. Each of the subsequent
chapters is a different important application; although not every application is important
to everyone, every application is important to someone.
I’d like to suggest two readings for you for next day, and they are both light. First, read
the MathReview for this book, to get a sense of the context in which these ideas appeared.
www.mathscinet.org. Second, take a look at the introduction.
What you need to know: ideally: a first course in algebraic geometry, such as much
of Hartshorne Chapter II, plus flatness. However: we can make do, if you’re willing to
work. The reason he can move so quickly is because of the power embedded in certain
algebro-geometric ideas. So we’re not going to be able to avoid notions such as flatness,
or Cartier divisors. It doesn’t matter if you haven’t seen Chern classes before; you’ll get a
definition here.
2. E XAMPLES
Before I dive into the subject, let me start with some examples.
The first will get across some notion of intersection multiplicity. Also, if you haven’t seen
schemes or varieties before, this will get your feet wet. Admittedly, I’ll let you dip in your
toe here, and throw you in the deep end. (Talk to me!)
Picture of parabola x = y2. Project it to t-line. Call the plane A2. (x, y) → x = t.
Algebra: K[x, y] surjects into K[x, y]/(x − y2). Affine schemes correspond to rings (the
categories are the “same” with the arrows reversed). I’m being agnostic about my field.
(I’ll try to call it K throughout the course.) Closed immersions correspond to surjections.
Ideals correspond to closed subschemes. K[t]. Map of rings in the opposite direction
t 7→ x.
Let’s intersect this with t = 1, or equivalently x = 1. K[x, y]/(x − 1). Intersecting these
schemes corresponds to taking the union of the deals. (Unions of schemes correspond to
intersections of ideals.)
∼ K[y]/(y2 − 1) =
K[x, y]/(x − y2, x − 1) = ∼ (K[y]/(y + 1)) ⊕ (K[y]/(y − 1))
(the latter by the Chinese remainder theorem; here I’m assuming char K 6= 2). Notice how
we can see this in the picture. Great, the line meets the parabola at 2 points. If I replace 1
by something else nonzero, then I still get 2 points (assuming K is algebraically closed!).
Complication 1 is not important: for most of this course I’ll assume the field is alge-
braically closed. But for those of you willing to think about nonalgebraically closed fields,
like Q, consider
∼ Q[y]/(y2 − 2).
Q[x, y]/(x − y2, x − 2) =
2
But now y2 − 2 is irreducible! So we get a single point, but we want it to count for 2. What
to do?
Complication 2 is more serious, but you’ve seen it before: what happens when you in-
tersect with x = 0? Then even if you only care about the complex numbers, you definitely
only get 1 point. In this case, you also want to say that the multiplicity is 2. Here’s the
algebra:
Q[x, y]/(x − y2, x) =
∼ Q[y]/(y2).
(In complication 1, the ring is a domain; in the second case it isn’t; the ring has a nilpo-
tent y. Those of you who have seen the geometry will know how to draw this.)
Here’s a picture that might (or might not) help you deal with complication 1. (Give a
triple cover picture.)
2.1. Fundamental theorem of algebra. Given a nonzero polynomial f(x) of degree n, the
number of zeros, counted appropriately, is n.
2.2. Bezout’s Theorem. Given two curves in P2 of degree d and e with no common
components, they meet in de points, counted properly. More generally, given n curves
in Pn of degree d1, . . . , dn, such that their intersection is zero-dimensional, they meet in
d1 · · · dn points, counted properly.
3
We haven’t proved this, but these things were known long ago, certainly before the
twentieth century. But here’s an example of why people got nervous.
Let X1 and X2 be two random planes in P4. They meet in a point. We can even give them
co-ordinates. Say the coordinates on P4 are [v; w; x; y; z]. We’ll say that X1 corresponds to
w = x = 0 and X2 corresponds to y = z = 0. Then X1 ∩ X2 meet at the point [1; 0; 0; 0; 0].
(Remember how projective space works: [1; 0; 0; 0; 0] = [17; 0; 0; 0; 0].)
Let X be the union X1 ∪ X2 in P4. Then this reasonably has degree 2. Let P be third
random plane. Then P meets X1 in one point, X2 in another, and misses their intersection,
to it meets X in 2 points. So far so good. If we move P around we should always get 2 (so
long as dim X ∩ P = 0).
But: we get something strange if we put P through X1 ∩ X2: we’ll get 3. Here’s the
calculation. We’ll work locally on projective space, on the open set where v 6= 0, so we
can set v = 1. Coordinates on this 4-space are given by w, x, y, z. The open set corresponds
to the ring K[w, x, y, z]. (You can let K = C.)
X1 corresponds to the ideal (w, x), and X2 corresponds to the ideal (y, z). So X1 ∪ X2
corresponds to (w, x) ∩ (y, z) (Ask.) = (wy, wz, xy, xz).
Let’s say P is given by w = y, x = z. This indeed meets X in one point. (Where does it
meet X1? w = y, x = z, w = x = 0, so w = x = y = z = 0. Where does it meet X2? similar.)
But let’s put our machine to work, and work out the scheme-theoretic intersection.
Ideal:
This was very alarming; Serre figured out what to do. I want to write down the answer,
but modern intersection theory bypasses this, so you shouldn’t worry about it much —
this may even begin to give you a warm feeling in your stomach for the new version of
the subject.
Suppose you wanted to intersect two things of complementary dimension in An, cor-
responding to K[x1, . . . , xn]/I1 and K[x1, . . . , xn]/I2 respectively. (Let R = K[x1, . . . , xn] for
convenience.) Interpretation of old formula
R R
dimk ⊗ .
I1 I2
There is something that could go to the left that would make this look like a long exact
sequence in cohomology:
4
/ Tor2(M, A) / Tor2(M, B) / Tor2(M, C) /
X R R
(−1)i dimk Tori( , )
I1 I2
and he proved it. Magically, in all of our previous examples, all the “higher” Tor’s van-
ished. (“Cohen-Macaulay”.) But we were just lucky.
3. S TRATEGY
Here’s the strategy we’re going to use. Here are things that homology satisfies in
“usual” topology in “good circumstances”. We have cycles in homology, and cycle classes,
which are cycles modulo homotopy.
Homology satisfies lots of other things. (In order to make this precise, Robert suggests:
allow locally finite chains.)
We’ll define our version of homology groups, which we’ll call Chow groups, using this.
6
INTERSECTION THEORY CLASS 2
RAVI VAKIL
C ONTENTS
1. Last day 1
2. Zeros and poles 2
3. The Chow group 4
4. Proper pushforwards 4
The new times starting next week will be Mondays 9–10:50 and Wednesdays 10–10:50.
So there will be a class on Friday.
Looking over today’s notes, I realize that what will be newest and most disconcerting
for those who haven’t seen schemes is the fact that we can localize at the generic point
of a subvariety X of a scheme Y. What this means is that we are considering the ring
of rational functions defined in a neighborhood of the generic point of X in Y; in other
words, they are defined on a dense open subset of X. This is indeed a ring (you can add
and multiply). The dimension of this ring is the difference of the dimensions of X and Y (or
more precisely dimensions of X and “Y near X”). Recall that the points of Y correspond to
irreducible subvarieties of Y; the “old-fashioned” (“before schemes”) points are the closed
points in the Zariski topology. So what are the points of Spec OX,Y, or equivalently, what
are the prime ideals of the ring OX,Y? They are the irreducible subvarieties of Y containing
X. The maximal ideal of this local ring corresponds to X itself.
1. L AST DAY
1.1. Examples. I showed you some examples. For example: Parabola x = y2 projected
to t-line. Q[t] 7→ Q[x, y]/(x − y2) via t → x. (I’m letting my field be Q for the moment.)
Intersecting parabola with a vertical line x = α. We get the scheme
Spec Q[x, y]/(x − y2, x − α) =
∼ Spec Q[y]/(y2 − α)
1
which is length 2 over the base field Q. If α = 1, we get 2 points:
Q[y]/(y2 − α) = ∼ (K[y]/(y + 1)) ⊕ (K[y]/(y − 1))
If α = 0, we get 1 point, with multiplicity 2:
Q[y]/(y2)
has only one maximal ideal. If α = 2, we get 1 point with multiplicity 1, but this point has
“degree 2 over Q”; the residue field is a degree 2 extension of Q.
1.2. Strategy. We’re going to define Chow groups of a variety X as cycles modulo
“homotopy” (called rational equivalence). Dimension k cycles are easy: they are dimension
k subvarieties of X. More subtle is rational equivalence.
Just to be clear before we start: throughout this course we’ll work over a field, to be
denoted K. We’ll consider schemes X that are sometimes called algebraic schemes over K.
They are schemes of finite type over K. This means that you get them by gluing together
a finite number of affine schemes of the form Spec K[x1, . . . , xn]/I. Mild generalization
of algebraic variety. All morphisms between algebraic schemes are separated and of finite
type. In this language, a variety is a reduced irreducible algebraic scheme. We’ll end up
localizing schemes: this leads to the notation of “essentially of finite type” = localizations
of schemes/rings of finite type.
Given a rational function on an irreducible variety X, I’ll define its order of pole or zero
along a codimension 1 variety. (A rational function is a(n algebraic) function on some dense
(Zariski-)open set. )
If X is generically nonsingular=smooth along Weil divisor, then “the same thing will
work”. More precicsely, in this case the local ring along the subvariety is dimension 1,
with m/m2 = 1, i.e. it is a discrete valuation ring, which I’ll assume you’ve seen.
Discrete valuation rings are certain local rings (A, m). Here are some characterizations:
If generator of m is π, then the ideals are all of the form (πn) or 0. The corresponding
scheme as 2 points; it is “the germ of a smooth curve”.
Examples: K[x, y], localized along divisor x = 0. We get rational functions of the form
f(x, y)/g(x, y) where x is not a factor of g. This is a local ring, and it is a DVR! Given any
rational function, you can tell me the order of poles or zeros. (Ask: (x2 − 3y)/(x2 + x4y)?)
Then this also works if x is replaced by some other irreducible polynomial, e.g. x2 − 3y.
This is nice and multiplicative.
So what if X is singular along that divisor (dim m/m2 > 1)? Example: y2 = x3 − x2, the
rational function y/x.
Exercise. Consider y/x on y2 = x3. What is the order of this pole/zero? (This will be
homework, due date TBA.)
Patch 1: If V is a Weil divisor, and r is a rational function that gives an element of the
local ring OV,X, then define
ordV (r) = dimK OV,X/(r) .
(What it means to be in the local ring, intuitively: at a general point of V it is defined.
More precisely: there is an open set meeting V — not necessarily containing it — where
the rational function is an actual function. For example, x/y on Spec K[x, y] is defined
near the generic point of x = 0. Language of generic points.) Then given a general rational
function, f, we can always write it as f = r1/r2, where r1 and r2 lie in the local ring.
(But we need to check that if we write f as a fraction in two different ways, then the
answer is the same. That’s true. More on that in a minute.)
Technical problem: If you have a dimension 1 local ring (A, m) with quotient field K, then
A isn’t necessarily a K-vector space. Z(p), pZ(p). (Exercise: Find an example in character-
istic 0.)
Better:
ordV (r) = lOV,X (OV,X/(r)) .
Recall “length” is the one more than the length of the longest series of nested modules
you can fit in a row. So the “length” of a vector space over K is its dimension.
Facts about facts. (I will pull facts out of Fulton’s appendix as black boxes. But if you
take a look at the appendix, you’ll see that these results are very easy. The vast majority
of proofs in A.1–A.5 are no longer than a few lines. With the exception of the section on
determinantal identities — which we likely won’t use in this course — I think almost no
proof is longer than half a page. He even has a crash course in algebraic geometry in
Appendix B.)
3
Fact: finiteness of zeros and poles (Appendix B.4.3). For a given r, there are only a
finite number of Weil divisors V where ordV (r) 6= 0.
3. T HE C HOW GROUP
Let X be anPalgebraic scheme (again: finite type over field K). Recall: A k-cycle is a finite
formal sum ni[Vi], ni ∈ Z. A cycle is positive if all ni ≥ 0, some ni > 0. (I forgot to
mention this.) Call this Zk[X], the group of k-cycles.
X
Zk[X] = ni[Vi], ni ∈ Z .
(You can probably see where I’m going to go with this.) Define
(Say visually.)
Note: this definition doesn’t care about any nonreduced structure on X: Ak[X] ≡ Ak[Xred].
4. P ROPER PUSHFORWARDS
Next day we’ll see that rational equivalence pushes forward under proper maps. First:
Finite morphisms are projective, hence proper. A morphism is finite if for each affine
open U = Spec S, f−1(U) is affine = Spec R, and the corresponding map of rings S → R is a
finite ring extension, i.e. R is a finitely generated S-module (which is stronger than a finitely
generated S-algebra!). Example: parabola double-covering line. (How to recognize: finite
implies each point of target has finite number of preimages. Reverse implication isn’t
true. finite = proper plus this property.) Another example: closed immersion.
4
Finite, projective, and proper morphisms are preserved by base change: if f is one of
them, then f 0 is too in the following fiber diagram:
f0 /
W X
f
Y / Z
5
INTERSECTION THEORY CLASS 3
RAVI VAKIL
C ONTENTS
1. Last day 1
2. Proper, projective, finite 2
3. Proper pushforwards 3
The new times will be Mondays 9–10:50 and Wednesdays 10–10:50. Because this is an
advanced course, I won’t have office hours; I’m happy to talk about it at any time. My
210 office hours are MW2:05–3 in case you want a specific time when I’ll definitely be in
my office.
1. L AST DAY
I should have been clearer on what I meant by “numbers of zeros and poles of a rational
function r ∈ R(X) along a Weil divisor (codimension 1 subvariety).” I meant the function
ordV (r). I defined it as follows. If r is actually defined at the generic point of V, we have
ordV (r) = lOV,X (OV,X/(r))
and then we define additively for quotients of two such: ordV (r/s) = ordV (r) − ordV (s).
Recall “length” is the one more than the length of the longest series of nested modules
you can fit in a row, so the “length” of a vector space over K is its dimension.
The algebraic fact from Fulton shows that this function is well-defined. In doing the
following exercise, use this definition.
is the group of k-cycles. A cycle is positive if all ni ≥ 0, some ni > 0. (I may have forgotten
to say this.)
Date: Friday, October 1, 2004.
1
The homotopies, or “rational equivalences” among k-cycles, were generated as follows.
For any (k + 1)-dimensional subvariety W of X, and any nonzero rational function r ∈
R(W)∗, define a K-cycle on X by
X
[div(r)] = ordV (r)[V].
This generates a subgroup Ratk X, the subgroup of cycles rationally equivalent to 0.
2.1. Proper, projective, finite morphisms. Crash course in proper morphisms: A mor-
phism f : X → Y is said to be proper if it is separated (true in our case of algebraic schemes),
of finite type (true in our case), and universally closed. (Closed: takes closed sets to closed
sets. Universally closed: for any Y 0 → Y, X ×Y Y 0 → Y 0 is closed.)
Some pictures: f : A1 → Spec K is not proper. f is certainly separated and of finite type
and closed, so what’s the problem? Consider the fibered diagram:
P 1 × A1 / A1
f
P1 / Spec K
The projection on the left isn’t closed: consider the graph of A1 ,→ P1.
First approximation of how to think of proper morphisms, if you are a complex geome-
ter: fibers are compact in the analytic topology. Warning: A1 ,→ P1 isn’t proper (it isn’t
closed), so I need to say something a bit more refined.
Key examples: projective morphisms are proper. As I said last day, a morphism f : X → Y
is projective if Y can be covered by opens such that on each open U, f−1(X) ×Y U → U
factors f−1(X) ×Y U ,→ Pk × U → U where the left morphism is a closed immersion.
Finite morphisms are projective, hence proper. A morphism is finite if for each affine
open U = Spec S, f−1(U) is affine = Spec R, and the corresponding map of rings S → R
is a finite ring extension, i.e. R is a finitely generated S-module (which is stronger than
a finitely generated S-algebra!). I’ll repeat the example from last time: parabola double-
covering line. (How to recognize: finite implies each point of target has finite number of
preimages. Reverse implication isn’t true. finite = proper plus this property.) Another
example: closed immersion.
Third (important) example: normalization (in good cases, such as those we’ll consider).
This requires a theorem in algebra! Normalization of an affine algebraic scheme Spec R
is Spec R̃, where R̃ is the normalization of R in its function field. Normalization of an
algebraic scheme X in general is obtained by gluing. (Theorem: this is possible, and also
independent of what affine cover you take of X.)
2
Crash course in normalization: given a variety W, define its normalization as follows.
If A is affine, let à be its integral closure in its function field R(A) = R(W). We have
Spec à → Spec A. Do this for every open affine set of W. Fact: they all glue together.
The result is called the normalization. Fact: The normalization map W̃ → W is finite
(algebra fact, Hartshorne Thm I.3.9A), hence proper. Hence: normalizations are regular in
codimension 1. (Proof: all local rings are integrally closed; in particular true for dimension
1 rings = codimension 1 subvarieties; hence any dimension 1 local rings (A, m) is a discrete
valuation ring, which (thanks to an earlier crash course) satisfies dimA/m m/m2 = 1, which
is the definition of nonsingularity.
Finite, projective, and proper morphisms are preserved by base change: if f is one of
them, then f 0 is too in the following fiber diagram:
f0 /
W X
f
Y / Z
(They are also preserved by composition: f, g proper implies g ◦ f is too. Ditto for
projective and finite.)
3. P ROPER PUSHFORWARDS
3.1. For any subvariety V of X, let W = f(V) be the image; it is closed (image of closed
is closed for proper morphisms). I want to define f∗ V. If dim W < dim V, define f∗ V =
0. Otherwise, R(V) is a finite field extension of R(W) (both are field extensions of K of
transcendence dgree dim V). Set
deg(V/W) = [R(V) : R(W)].
f g
Note: If X / Y / Z , then (g ◦ f)∗ = g∗ f∗ .
E-mail address: [email protected]
3
INTERSECTION THEORY CLASS 4
RAVI VAKIL
C ONTENTS
1. Proper pushforwards 1
2. Flat pullback 3
3. Parsimonious definition of Chow groups 7
4. Things Rob will tell you about on Wednesday 7
4.1. Excision exact sequence 7
2. In no more than half a page, explain why Bezout’s Theorem for plane curves is true
(i.e. explicate Fulton’s Example 1.4.1). Feel free to assume that F is irreducible.
You can also get a “bye” for two weeks of homework by (at some point in the future)
explaining to me the “rational equivalence pushes forward under proper morphisms”
result (Prop. 1.4).
We’re in the process of seeing that cycles (proper) pushforward and (flat) pullback, and
that rational equivalences do to.
We need a lot of algebra to set ourselves up. This will decrease in later chapters.
Also, Rob will give Wednesday’s class; he’ll end Chapter 1 and start Chapter 2.
1. P ROPER PUSHFORWARDS
1.1. For any subvariety V of X, let W = f(V) be the image; it is closed (image of closed
is closed for proper morphisms). I want to define f∗ V. If dim W < dim V, define f∗ V =
0. Otherwise, R(V) is a finite field extension of R(W) (both are field extensions of K of
transcendence degree dim V). Set
deg(V/W) = [R(V) : R(W)].
1
In the complex case, this degree is what you think it is: it’s the number of preimages of
a general point. In positive characteristic, this needn’t be true; K[tp] → K[t] gives a map
of schemes that is degree p but is one-to-one on points.
f g
Note: If X / Y / Z , then (g ◦ f)∗ = g∗ f∗ .
Example: the parabola example (what happens to points, and to the entire parabola).
Normalization.
I’m not going to prove this; I’ll only point out that we can reduce this statement to
something simpler:
This is a really natural reduction. We need only to prove it for a generator of rational
equivalence, which involves X 0 ,→ X of dimension k + 1, and α = div(r) for r ∈ K(X 0).
We can just work on X 0 instead. We can also replace Y by f(X), because this construction
doesn’t care about anything else.
Bonus 1. We can now define the degree of a dimension 0 cycle class (= cycle mod rational
equivalence) on something proper over K. (Definition: complete = proper over K. This is a
common word, but I may try to avoid it.)
P P
Definition. If α = nPP is a zero-cycle on X, define the degree of α to be nP deg[P/K]
(the sum of the degree extensions). Example: Spec Q[x]/(x2 + 2) over Spec Q, there is one
point, that counts for 2.
Bonus 2. Recall that we were annoyed at having to working out ordOV,X (r) for r ∈ R(X),
and needing to use lengths, and not what we know about DVR’s. This theorem tells us
we don’t have to. We could pull r back to the normalization X̃ of X, which is regular in
codimension 1. We work out how it vanishes on all the divisors mapping to V. (There are
a finite number, by finiteness of normalization, which I said earlier.)
Exercise. Check your answer to ord(y/x) at (0, 0) on y2 = x3, i.e. k[x, y]/(y2 − x3) by
pulling it back to the normalization, which is k[t], given by t 7→ (t2, t3).
2. F LAT PULLBACK
Idea “flat morphisms are nice”. They are more general than fibrations, but have all the
same properties.
Examples of flat morphisms: the map of the parabola to the line is one. Reason: k[x] is
a flat k[x2]-module, as it is a free k[x2]-module.
Goal: flat pullbacks exist. In other words, we’ll define out how cycles pullback, and
then we’ll check that rational equivalences pull back to rational equivalences.
You can see why we don’t like dimensional jumping. But it’s interesting that we don’t
mind degenerations as in the family of nodal curves, or in the parabola example.
Example: k[x, y]/(y2(x + y)3). The length of the local rings at the two generic points are
2 and 3 respectively.
What’s wrong with that? Well, we don’t know that (gf)∗ = f∗ g∗ . Example (picture
omitted in notes): branched double cover of a branched double cover. Fortunately we get
4 both ways. But does this work in general?
Lemma (pulling back fundamental classes). If f : X → Y is flat, then for any equidi-
mensional subscheme Z of Y, f∗ [Z] = [f−1(Z)]. In other words, the pullback of a funda-
mental class of a scheme is the fundamental class of the pullback of a scheme.
This makes us happy, because schemes pullback nicely; f−1g−1(Z) = (gf)−1(Z). Thus
pullbacks are functorial.
g0
X0 / X
f0 f
g
Y0 / Y
4
be a fibered square, with g flat and f proper (so g 0 flat and f 0 proper). Then f∗0 g 0∗ α =
g∗ f∗ α.
This is on the level of cycles. We don’t yet know that we can flat-pullback cycle classes.
Proof also by direct algebra. Reduce first to the case where X and Y are varieties, and
f is surjective. Here are the reductions: it suffices to do this for a generator of ZkX, so
α = [V] where V is a variety. f(V) is also a variety (remember f is proper, hence f(closed)
is closed). Base change the entire square by f(V) → Y. Then we can assume f(V) = Y.
Next base change the upper arrow by V → X:
X 0 ×X V / V
g0
X0 / X
f0 f
g
Y0 / Y
We’ll next show that rational equivalences flat-pullback to rational equivalences. Hence
we’ll have shown that we have flat pullback of Chow groups.
(An effective Cartier divisor is a subscheme locally cut out by a single function that is not
a zero-divisor.)
This is certainly reasonable! (Draw a picture, when D doesn’t have a component along
the intersection of two of the Xi’s.) I omitted this explanation in class due to time.
Proof. One checks this along each Weil divisor V of X. Immediately reduces to algebra.
Let me get us to the algebra. We’ll check that each codimension one subvariety V of X
appears with the same multiplicity on both sides of the equation. We reduce to the local
situation: let A be the local ring of X along V, and a ∈ A a local equation for D. The
minimal prime ideals pi in A correspond to the irreducible components Xi of X which
contain V.
mi = lApi (Api ). The multiplicity of [V] in [D] is lA(A/aA). The multiplicity of [V] in
[Di] is lA/pi (A/(pi + aA)). So we want to show:
X
lA(A/aA) = milA/pi (A/pi + aA).
(Draw picture.)
Before I get into it, notice that flatness is already in the picture here: each Vi → P1 is
flat. We’ll see that [Vi(0)] is the flat pullback of 0, and ditto for ∞.
Proof. (I only roughly sketched this proof in class.) If there are such subvarieties, then
α ∼ 0: Certainly the classes on X × P1 are each rationally equivalent to 0 by the definition
of rational equivalence. The projection X × P1 → X is proper (because P1 → pt is proper,
and properness is preserved by base change).
Now for the other direction. We need to show this for a generator of rational equiva-
lence on X, so there is a subvariety W of dimension k + 1 in X, and a rational function on
W r ∈ R(W)∗. This gives a rational map W 99K P1. Let V be the closure of the graph of
this rational map, so V ⊂ W × P1 ,→ X × P1. (The generic point of W maps to the generic
point of P1, so the same is true of V.) V maps birationally and properly onto W. That
morphism is degree 1. If f is the induced morphism to P1, then div(r) = p∗ [div(f)] by our
big theorem on proper pushforwards, which in turn equals [V(0)] − [V(∞)].
Proof. (I did not give this proof in class.) We may deal with a generator of rational equiv-
alence. Thanks to the geometric lemma, we can take our generator to be of the form
α = [V(0)] − [V(∞)].
cl. imm. p
(f × 1)−1(V) / 1 /X
W =< X × Pflat prop.
<<
<<
<< flat flat flat f
<< q
<<h cl. imm. / 1 /Y
<< V Y × Pflat prop.
<< ooo
oo
<< g oooo
<<
wooo
P1
We have a cycle α that is rationally equivalent to 0 on Y. It is the proper pushforward of
[g−1(0)] − [g−1(∞)] from V. When we pull back this class from Y to X, we want to see that
it is rationally equivalent to 0. But by our lemma showing that proper pushforwards and
flat pullbacks commute, that’s the same as pulling back to W, and pushing forward to X.
The pullback to W is [h−1(0)] − [h−1(∞)].
6
We feel like we’re done: we just push this forward to X, and that should be it. But: W
may not be a variety (it may be reducible and nonreduced), so we don’t (yet) know that
this class
P is rationally equivalent to 0. This is why we need our algebraic lemma. Let
[W] = mi[Wi]. Since
[h−1 −1
i (0)] − [hi (∞)] = div(hi)]
P
is ratinoally equivalent to 0, it suffices to verify that [h−1(P)] = mi[h−1
i (P)] (and then
plug in P = 0 and ∞). And that’s precisely what the algebraic lemma tells us.
I’d promised earlier that Chow groups would satisfy 3 conditions: (a) 0 would be ratio-
nally equivalence to ∞ in P1. (b) They would satisfy flat pullbacks. (c) They would satisfy
proper pushforward.
We’ve shown this. Now note that these three things define Chow groups. Translation:
anything satisfying these three things is a quotient of the Chow group, so the Chow group
is the “minimal” thing satisfying these three conditions. To prove this, all we have to do
is show that if W is a (k + 1)-dimensional subvariety of X, and r is a rational function on
W, then div(r) is forced to be 0. W 99Kr P1. Pullback (0) − (∞). Pushforward by closed
immersion W → X.
Something else to point out: the divisor of zeros and poles of a rational function r on a
variety W is easy to understand if W is regular in codimension 1. It was a pain otherwise.
Here’s an alternate way of computing it.
Pull back the function r to W̃. Do the calculation there. Then take proper pushforward.
4.1. Excision exact sequence. Proposition. Let Y be a closed subscheme of X, and let
U = X − Y. Let i : Y ,→ X be the closed immersion (proper!) and j : U → X be the open
immersion (flat!). Then
i∗ j∗
AkY / AkX / AkU / 0
is exact for all k.
Proof omitted.
Exercise: Example 1.9.3 (a). Show that Ak(Pn) is generated by the class of a k-dimensional
linear space. (Hint: use the excision exact sequence.)
Proposition.
(a) if α ∼ 0 then α × β ∼ 0. There are exterior products AkX ⊗ AlY → Ak+l(X × Y).
(b) If f and g are proper, then so is f × g, and (f × g)∗ (α × β) = f∗ α × g∗ β. Hence
exterior product respects proper pushforward.
(c) If f and g are flat of relative dimensions m and n, (so f × g is flat of relative dimen-
sion m + n), then
(f × g)∗ (α × β) = f∗ α × g∗ β.
Hence exterior product respects flat pullback.
E-mail address: [email protected]
8
INTERSECTION THEORY CLASS 6
RAVI VAKIL
C ONTENTS
1. Divisors 2
1.1. Crash course in Cartier divisors and invertible sheaves (aka line bundles) 3
1.2. Pseudo-divisors 3
2. Intersecting with divisors 4
2.1. The first Chern class of a line bundle 5
2.2. Gysin pullback 6
1
Definition. Y → X is an affine bundle of rank n over X if there is an open covering ∪Uα
∼ Ui × An → Ui. This is a flat morphism.
of X such that f−1(Ui) =
Homework: Example 1.9.3 (a). Show that Ak(Pn) is generated by the class of a k-
dimensional linear space. (Hint: use the excision exact sequence.)
Proposition.
(a) if α ∼ 0 then α × β ∼ 0. There are exterior products AkX ⊗ AlY → Ak+l(X × Y).
(b) If f and g are proper, then so is f × g, and (f × g)∗ (α × β) = f∗ α × g∗ β. Hence
exterior product respects proper pushforward.
(c) If f and g are flat of relative dimensions m and n, (so f × g is flat of relative dimen-
sion m + n), then
(f × g)∗ (α × β) = f∗ α × g∗ β.
1. D IVISORS
There are three related concepts of divisors: Weil divisors, Cartier divisors, and (a con-
cept local to intersection theory) pseudodivisors.
The notion of Cartier divisor looks more unusual when you first see it. A Cartier divisor
is defined by data (Uα, fα) where the Uα form an open covering of X and fα are non-zero
functions in R(Uα) = R(X), subject to the condition that fα/fβ is a unit (regular, nowhere
vanishing function) on the intersection Uα ∩ Uβ. There is an equivalence classes of Cartier
divisors.
The rational functions are called local equations. Local equations are defined up to mul-
tiplication by a unit.
Baby Example: X = Uα = A1 − {1}, local equation 1/t2. Another local equation for the
same Cartier divisor: (t − 1)/t2.
2
1.1. Crash course in Cartier divisors and invertible sheaves (aka line bundles). (See
Appendix B.4 for an even faster introduction!)
{ Cartier divisors } o / { inv. sheaves w. nonzero rat’l sec. }/inv. funcs.Γ (X, OX
∗
)
Given a Cartier divisor (Uα, fα), here’s how you produce an invertible sheaf F . I need
to tell you F (U). F (U) = {(gα ∈ OX ∗
(Uα ∩ U))α = R(X) : gαfα ∈ OX(Uα ∩ U)), gαfα =
gβfβ ∈ OX(U ∩ Uα ∩ Uβ)}. You can check that this is indeed a sheaf, and it is locally trivial:
check for U = Uα that F (Uα) consists of rational functions gα on Uα such that gαfα is a
regular function. Thus the gα are all of the form O(Uα)/fα (regular functions divided by
fα), and thus as a O(Uα)-module, it is isomorphic to O(Uα) itself.
Baby Example: X = Uα = A1 − {1}, fα = 1/t2. The rational functions on X are K[t, 1/(t −
1)]. The module corresponding to this Cartier divisor is K[t, 1/(t − 1)]t2, which is clearly
isomorphic to K[t, 1/(t − 1)].
A Cartier divisor is effective if the fα are all regular functions (“have no poles”). Thus
we can add to that square above: { effective Cartier divisors } correspond to invertible
sheaves with nonzero sections.
A Cartier divisor is principal if it is the divisor of a rational function i.e. div(r) where
r ∈ R(X)∗ . Two Cartier divisors differing by a principal Cartier divisor give rise to the
same invertible sheaf.
Rob told you that the Cartier divisor form an abelian group Div(X). When you mod
out by the subgroup of principal Cartier divisors, you get the group of invertible sheaves
Pic X.
The support of a Cartier divisor D, denoted |D|, is the union of all subvarieties Z of X
such that the local equation for D in the ring OZ,X is not a unit. This is a closed algebraic
subset of X of pure codimension one.
Notice: invertible sheaves pull back, but Cartier divisors don’t necessarily. (Give an
example.)
We have a map from Cartier divisors to Weil divisors. Linear equivalence of Cartier di-
visors rational equivalence of Weil divisors, hence this map descends to Pic X → Adim X−1X.
Hence given any pseudo-divisor D, we get a Weil divisor class in An−1X. But we can
do better. Given a pseudo-divisor D, we get a Weil divisor class [D] ∈ An−1(|D|).
We will now define our first intersections, that with Cartier divisors, or more generally
pseudo-divisors. Let D be a pseudo-divisor on a scheme X. We define D · [V] where V is
a k-dimensional subvariety. D · [V] := [j∗ D] where j is the inclusion V ,→ X. This lies in
Ak−1V ∩ |D|. Hence we can do this with any finite combination of varieties. Note that we
get a map ZkX → Ak−1X, but we’re asserting more: we’re getting classes not just on X,
but on subsets smaller than X.
Proposition 2.3.
Example: Intersection of two curves in P2, C1 and C2. We get a number. Old-fashioned
intersection theory (Hartshorne V): O(C1)|C2 gives you a number.
This tells you a bit more: the class has “local contributions” from each connected com-
ponent of the intersection.
4
Excess intersection can happen! Example: A line meeting itself.
Remark: This proves some of the things Fulton said about Bezout in the first chapter.
Here’s a natural question: if you intersect two effective Cartier divisors, then if you
reverse the order of intersection, you had better get the same thing!
D · [D 0] = D 0 · [D]?
Big Theorem 2.4 Let D and D 0 be Cartier divisors on an n-dimensional variety X. Then
D · [D 0 ] = D 0 · [D] in An−2(|D| ∩ |D 0|).
2.1. The first Chern class of a line bundle. If L is a line bundle on X, we define “c1(L)∩”.
If V is a subvariety, then write the restriction of L to C as OV (C) for some Cartier divisor
C. Then define c1(L) ∩ [V] = [C]. (C is well-defined up to linear equivalence, so this makes
sense in Adim V−1V ,→ Adim V−1X.) Extend this by linearity to define c1(L)∩ : ZkX → Ak−1X.
Proposition 2.5.
5
We’ll prove this next day.
Proposition.
Proof next day; although in fact you may be able to see how all but (d) comes from
what we’ve said earlier today. (Part (d) comes from something we discussed earlier, but
I’ll leave that for next time.)
E-mail address: [email protected]
6
INTERSECTION THEORY CLASS 7
RAVI VAKIL
C ONTENTS
Here’s where we are. We have defined divisors of 3 sorts: Weil divisors, Cartier divi-
sors, and pseudo-divisors (L, Z, s).
I’d like to make something more explicit than I have. An effective Cartier divisor on a
scheme is a closed subscheme locally cut out by one function, and that function is not a
zero-divisor. (Translation: the zero-set does not contain any associated points.) Pic X =
group of line bundles = Cartier divisors modulo linear equivalence = Cartier divisors
modulo principal Cartier divisors. We get a map from Cartier divisors to Weil divisors
that descends to Pic X → Adim X−1X.
Proposition 2.3.
1
(c) (projection formula) Let D be a pseudo-divisor on X, f : X 0 → X a proper mor-
phism, α a k-cycle on X 0 , and g the morphism from f−1(|D|) ∩ |α| to |D| ∩ f(|α|)
induced by f. Then
g∗ (f∗ D · α) = D · f∗ (α) in Ak−1(|D| ∩ f(|α|)).
(d) (commutes with flat base change) Let D be a pseudo-divisor on X, f : X 0 → X a flat
morphism of relative dimension n, α a k-cycle on X, and g the induced morphism
from f−1(|D| ∩ |α|) to |D| ∩ |α|. Then
f∗ D · f∗ α = g∗ (D · α) in Ak+n−1(f−1(|D| ∩ |α|).
(e) If D is a pseudodivisor on X whose line bundle OX(D) is trivial, and α is a k-cycle
on X, then
D · α = 0 in Ak−1(|α|).
Proof. (a) This follows from the definition; D · α is linear in the second argument because
it was defined by linearity and D · [V] for V a subvariety! Hence for the rest of the proof
we can assume α = [V].
(b) Recall the definition of D · [V]: We pull the pseudo-Cartier divisor D back to V. We
take any Cartier divisor giving that pseudo-divisor (let me sloppily call this
P D as well).
We then take the Weil divisor corresponding to that Cartier divisor: D 7→ W ordW(D).
This latter is a group homomorphism.
D can be chosen to be some Cartier divisor on X = f(V). Note that f∗ D is also Cartier:
the support of D doesn’t contain the generic point of F(V) hence f∗ D doesn’t contain
the generic point of V. Then we want to prove: f∗ [f∗ D] = deg(X 0 /X)[D]. This is a local
question on X, so we can assume D = div(r) for some rational function on X. Then we
have:
f∗ [div(f∗ r)] = [div(N(f∗r))]
(came up in discussion of why proper pushforwards exist)
0
= div(rdeg(X /X))
(definition of norm)
= (deg X 0 /X)[div r].
(d) (Skipped for the sake of time) Again we can assume V = X, so D is represented by a
Cartier divisor. We want to prove [f∗ D] = f∗ [D] as cycles on X. Both sides are additive, so
we need only prove it for the case where D is effective. But we’ve shown earlier (Lemma
2
in Section 2 of Class 4 notes, Oct. 4) that fundamental classes of subschemes behave well
with respective to flat pullbacks, so we’re done.
(e) We may assume again that α = [V], V = X, and D is a Cartier divisor on X. We know
that D is principal. Then we want to show that [D] = 0 in Ak−1(X). This follows from
the fact that there is a group homomorphism from the group of Cartier divisors modulo
linear equivalence (i.e. modulo principal divisors) to the group of Weil divisors modulo
linear equivalence (the latter is Ak−1(X)).
Big Theorem 2.4. Let D and D 0 be Cartier divisors on an n-dimensional variety X. Then
D · [D 0 ] = D 0 · [D] in An−1(|D| ∩ |D 0|).
Proof soon.
Proposition 2.5.
3
Proof. (a) follows from a corollary to our big theorem that I stated last time: If D is a
pseudo-divisor on X, α a k-cycle on X which is rationally equivalent to 0. Then D · α = 0
in Ak−1(|D|).
(b) follows from another corollary to the big theorem that I stated last day: If D and D 0
are pseudo-divisors on a scheme X. Then for any k-cycle α on X, D · (D 0 · α) = D 0 · (D · α)
in Ak−2(|D| ∩ |D 0 | ∩ |α|).
3. G YSIN PULLBACK
Proposition.
So the real problem is what to do if D and D 0 have a common component. We’ll deal
with this by induction on this:
(D, D 0) := max{ordV (D) ordV (D 0) : codim(V, X) = 1}.
Note that we know the result when = 0.
4.1. Crash course in blowing up. I’m going to repeat this next time, in more detail. Let
X be a scheme, and I ⊂ OX a sheaf of ideals on X. (Technical requirement automatically
satisfied in our situation: I should be a coherent sheaf, i.e. finitely generated.) Here is
the “universal property” definition of blowing-up. Then the blow-up of OX along I is a
morphism π : X̃ → X satisfying the following universal property. f−1IOX̃ (the “inverse
ideal sheaf”) is an invertible sheaf of ideals, i.e. an effective Cartier divisor, called the
exceptional divisor. (Alternatively: the scheme-theoretic pullback of the subscheme O/I
is a closed subscheme of X̃ which is (effective) Cartier, and this is called the exceptional
(Cartier) divisor E.) If f : Z → X is any morphism such that (f−1I)OZ is an invertible sheaf
of ideals on Z (i.e. the pullback of O/I is an effective Cartier divisor), then there exists a
unique morphism g : Z → X̃ factoring f.
g
Z >_ _ _/ X̃
>
>> f
>> π
>>
X
In other words, if you have a morphism to X, which, when you pull back the ideal I, you
get an effective Cartier divisor, then this factors through X̃ → X.
As with all universal property statements, any two things satisfying the universal prop-
erty are canonically isomorphic.
Theorem: Blow-ups exist. The proof is by construction: show that Proj⊕d≥0 I d satisfies
the universal property. (See Hartshorne II.7, although his presentation is opposite.)
I’m going to start next day by discussing three examples: (i) blowing up a point in
the plane, (ii) blowing up along an effective Cartier divisor, and (iii) blowing up X along
itself.
E-mail address: [email protected]
5
INTERSECTION THEORY CLASS 8
RAVI VAKIL
C ONTENTS
Our goal now is to prove the key result of Chapter 2. It’s not impressive in and of itself,
but we used it to do a lot of other things.
Big Theorem 2.4. Let D and D 0 be Cartier divisors on an n-dimensional variety X. Then
D · [D 0 ] = D 0 · [D] in An−1(|D| ∩ |D 0|).
Last time, I discussed the case where D and D 0 have no common components, so |D| ∩
|D 0 | is codimension 2. I didn’t prove it, but argued that it boils down to algebra. So the
real problem is what to do if D and D 0 have a common component.
The proof involves an extremely clever use of blowing up. Given the background of
the people in this class, I’ve had to make some decisions as to what arguments to include,
and I think I’d most like to give you some feeling for blowing up, and then to outline the
proof, rather than getting into the gory details.
1.1. Crash course in blowing up. Last time I began to talk about blowing up. Let X
be a scheme, and I ⊂ OX a sheaf of ideals on X. (Technical requirement automatically
satisfied in our situation: I should be a coherent sheaf, i.e. finitely generated.) Here is
the “universal property” definition of blowing-up. Then the blow-up of OX along I is a
morphism π : X̃ → X satisfying the following universal property. f−1IOX̃ (the “inverse
ideal sheaf”) is an invertible sheaf of ideals, i.e. an effective Cartier divisor, called the
exceptional divisor. (Alternatively: the scheme-theoretic pullback of the subscheme O/I
is a closed subscheme of X̃ which is (effective) Cartier, and this is called the exceptional
(Cartier) divisor E.) If f : Z → X is any morphism such that (f−1I)OZ is an invertible sheaf
Date: Monday, October 18, 2004.
1
of ideals on Z (i.e. the pullback of O/I is an effective Cartier divisor), then there exists a
unique morphism g : Z → X̃ factoring f.
g
Z >_ _ _/ X̃
>
>> f
>> π
>>
X
In other words, if you have a morphism to X, which, when you pull back the ideal I, you
get an effective Cartier divisor, then this factors through X̃ → X.
As with all universal property statements, any two things satisfying the universal prop-
erty are canonically isomorphic.
Theorem: Blow-ups exist. The proof is by construction: show that Proj⊕d≥0 I d satisfies
the universal property. (See Hartshorne II.7, although his presentation is opposite.)
Example 1. The “typical” first example is the blow-up of the plane at a point, Bl0A2.
Let X = {(p ∈ A2, ` line in plane through p and 0 }. Note that (i) X is smooth (it is an A1-
bundle = total space of a line bundle over the P1 parametrizing the possible `), (ii) it has a
map π to A2, (iii) π is an isomorphism away from p, and π−1p = ∼ P1. This P1 is codimension
1 on a smooth space, hence an effective Cartier divisor. Fact: This satisfies the universal
property, hence is a blow-up. More generally, if you blow up a point on a smooth surface,
the same story happens. More generally still, if you blow up a smooth variety X along a
smooth subvariety V of codimension k, you get something that is isomorphic away from
V, and the preimage of V is a Pk−1-bundle over V; it is the projectivized normal bundle
(i.e. points of the exceptional divisor E correspond to points of V along with a line in the
normal bundle to V in X.)
Example 2. If you blow up X along an effective Cartier divisor D, then nothing changes.
(X, D) → X already satisfies the universal property, tautologically.
Example 3. If you blow up X along itself, it disappears. For example, consider X = A1,
and I = 0. Then there is no way to pullback this ideal sheaf and get a Cartier divisor,
which is codimension 1. Well, there is one way: via the morphism ∅ → X.
Example 3a. If you blow up X along one of its components, the component is blown
away (disappears), and the rest will be affected too (blown up along their intersection
with the old component).
Fun Example 4. Consider the cone, and blow it up along a line. The line is not a Cartier
divisor, as we showed last day. Hence the blow-up does something. Moreover, it does
nothing away from cone point. It turns out that this does indeed smooth out the cone! (It
does the same thing as blowing up the cone point by itself.)
2
Remark. If X is a variety and Y 6= X, then X̃ → X is birational.
1.2. Back to the proof. D and D 0 are two Cartier divisors, cut out locally by a single
equation. Let D ∩ D 0 be the intersection scheme of D and D 0 . Let π : X̃ → X be the
blow-up of X along D ∩ D 0, and let E = π−1(D ∩ D 0) be the exceptional divisor. The local
equations for π∗ D and π∗ D 0 are divisible by the local equation for E. Translation: D and
D 0 both lie in the ideal sheaf of D ∩ D 0, hence their pullback lies in the (Cartier) ideal sheaf
of E. Hence we can write equalities of Cartier divisors:
π∗ D = E + C, π∗ D 0 = E + C 0 .
Let
Note that we know the result when = 0. We’re going to work by induction on .
Omitted Lemma. (a) C and C 0 are disjoint. (This is a special case of Hartshorne Exercise
II.7.12.) (b) If (D, D 0) > 0, then (C, E), (C 0, E) < (D, D 0).
Proof is omitted. But caution: something very interesting is going on here. I’ll give
three examples to show you this. First, suppose L1, L2, and L3 are three general lines in
P2. If D = L1 and D 0 = L1 +L3, then D∩D 0 = L1, and the blow-up does nothing. However,
E = L1, and then C = ∅ and C 0 = L3.
Case D and D 0 effective. We do this by induction on (D, D 0). The case = 0 is already
done (or more precisely, assumed!), as described earlier. If (D, D 0) > 0, then blow up
X along D ∩ D 0. Then the omitted lemma asserts that the theorem holds for (E, C 0) and
(C, E). The theorem also holds for (E, E) stupidly (clearly E · [E] = E · [E]), and also for
(C, C 0 ) for different stupid reasons (C · [C 0 ] = 0 = C 0 · [C]). So the above lemma completes
this proof.
General case. Blow up X along the ideal sheaf of denominators of D 0. Then the pairs
(π∗ D, C) and (π∗ D, E) are covered by the previous case, so we’re done by the Lemma.
In the next chapter, we’re going to generalize the notion of the first Chern class of a
line bundle to the notion of an arbitrary Chern class on an arbitrary vector bundle. These
Chern classes will have similar properties to those you may have seen elsewhere, but we
get at them in a strangely backwards way, by defining Segre classes first. The generating
function for Segre classes will be inverse to that of Chern classes.
When you look through this chapter, you’ll note that only a very small portion of it
consists of propositions and theorems. The rest is full of useful examples.
2.1. Segre classes of vector bundles. Let E be a vector bundle of rank e+1 on an algebraic
scheme X. Let P = PE be the Pe-bundle of lines on E, and let p = pE : P → X be the
projection. Note that it is both flat and proper (explain).
The line bundle O(1). On P there is a canonically defined line bundle, called the tau-
tological bundle, denoted O(−1) or OE(−1). For any point of P, I’ll need to give you a
4
one-dimensional vector space in some natural way. But each point of P corresponds to a
line of E.
Define O(1) as the dual of O(−1), and let O(n) be O(1)⊗n (with the obvious convention
if n is nonpositive).
Remark. On Pe, there is a line bundle / invertible sheaf O(1), and indeed OE(1) restricts
to each of the fibers to give O(1). But this doesn’t determine the class OE(1). Indeed, if
I pull back any line bundle on X to P, I get a line bundle trivial on each of the fibers, so
OE(1) ⊗ L has this property for any invertible sheaf L.
Proof. In this case PE = X, and OE(−1) = E so OE(1) = E∨, hence s1(E) ∩ α = c1(OE(1) ∩
α) = −c1(E) ∩ α.
Segre class Theorem. (a) for all α ∈ AkX, (i) si(E) ∩ α = 0 for i < 0, and (ii) s0(E) ∩ α = α.
(b) (commutativity) If E and F are vector bundles on X, and α ∈ AkX, then for all i, j,
si(E) ∩ (sj(F) ∩ α) = sj(F) ∩ (si(E) ∩ α).
(d) (Segre classes behave well with respect to flat pullback) If f : X 0 → X is flat, E a
vector bundle on X, α ∈ A∗ X
si(f∗ E) ∩ f∗ α = f∗ (si(E) ∩ α).
5
Corollary. The flat pullback p∗ : AkX → Ak+e(PE) is a split monomorphism: by (a) (ii), an
inverse is β 7→ p∗ (c1(OE(1))e ∩ β).
(d) Exercise.
(a) We may assume that α = [V]. Then by (c), using the (proper) closed immersion
V ,→ X, we may assme X = V. Then for i < 0, si(E) ∩ [V] ∈ Adim V−iX = 0, so (i) is done.
Similarly,
s0(E) ∩ [V] = p∗ (c1(OPE(1))e ∩ [P]) = m[V]
for some m. We will show that m = 1. We can check this on an open set of V, so restrict
to an open set where E is a trivial bundle. Then P = PE = X × Pe, and O(1) has sections
whose zero scheme is X × Pe−1. Then c1(O(1)) ∩ [X × Pe] = [X × Pe−1] (from earlier theorem
on c1 of an effective Cartier divisor). Repeat this e times to get the desired result.
6
INTERSECTION THEORY CLASS 9
RAVI VAKIL
C ONTENTS
I have one update from last time, and this is aimed more at the experts. Rob pointed
out that there was no reason that we know that a Cartier divisor can be expressed as
a difference (or quotient) of effective Cartier divisors. More precisely, a Cartier divisor
can be described cohomologically as follows. Let X be a scheme. We have a sheaf O∗
of invertible functions. There is another sheaf K∗ that are things that locally look like
quotients of a function by a nonzerodivisor. (If X is a variety, then K∗ is the constant
sheaf with K∗ (U) = R(X) for all U.) Then I informally described Cartier divisors of X
as determined by certain data: there is an open cover of X by open sets Ui; we have
an element of K∗ for each Ui; and on Ui ∩ Uj the quotient of the two elements of K∗
∗
corresponding to i and j is an element of OX . We then mod out by an equivalence relation
that I was careless about defining. This definition translates to the more compact notation:
Cartier divisors are global sections of the (quotient) sheaf K∗ /OX∗
. (More generally, we get
∗ ∗
a sheaf of Cartier divisors K /OX.) The description I gave was the Cech description of a
quotient sheaf. This drives home the point that any Cartier divisor is locally the quotient
of two effective Cartier divisors, but not necessarily globally. I don’t know of any specific
examples of a Cartier divisor that is not the quotient/difference of two effective Cartier
divisors, and I would like to see one.
Fulton is then proving that even though we don’t know for sure that any Cartier divisor
D on X is a difference of two effective divisors, we can construct a proper surjective mor-
phism π : X̃ → X such that π∗ D is Cartier, and a difference of effective Cartier divisors.
Moreover, he tells us what to do: define a closed subscheme by taking the “ideal sheaf of
denominators” of the Cartier divisor, and blow it up. The example I said I’d like to see
corresponds to the question: find a scheme and a Cartier divisor where this ideal sheaf of
denominators is not Cartier. I’m still a bit perplexed; it seems to me that it should always
be Cartier, as “Cartier-ness” is a local condition, and locally every Cartier divisor is prin-
cipal. (I’m assuming, as we are throughout this course, that all schemes are essentially of
finite type, so the Cech description certainly holds.)
1
As an aside: when you see that Cartier divisors are global sections of a quotient sheaf,
you should immediately be curious about the corresponding long exact sequence of co-
homology.
0 → O∗ → K∗ → K∗ /O∗ → 0
gives us
0 → H0(X, O∗ ) → H0(X, K∗ ) → H0(X, K∗ /O∗ ) → H1(X, O∗) → H1(X, K∗ ) = 0.
The right term is 0 because K is a flasque (=flabby) sheaf. All the other terms have obvious
meanings too. The image of H0(X, K∗ ) is the set of principal Cartier divisors. (An element
of H0(X, O∗ ) gives a trivial principal divisor.) H1(X, OX ∗
) = Pic X. So this shows that
∼ Cartier divisors modulo principal divisors.
Pic X =
1.1. Segre classes of vector bundles. Let E be a vector bundle of rank e+1 on an algebraic
scheme X. Let P = PE be the Pe-bundle of lines on E, and let p = pE : P → X be the
projection. Note that it is both flat and proper (explain).
Define homomorphisms
si(E)∩ : AkX → Ak−iX
e+i
by α 7→ p∗ (c1(O(1)) ∗
∩ p α). Note that this indeed maps from AkX → Ak−iX.
Segre class Theorem. (a) for all α ∈ AkX, (i) si(E) ∩ α = 0 for i < 0, and (ii) s0(E) ∩ α = α.
(b) (commutativity) If E and F are vector bundles on X, and α ∈ AkX, then for all i, j,
si(E) ∩ (sj(F) ∩ α) = sj(F) ∩ (si(E) ∩ α).
(d) (Segre classes behave well with respect to flat pullback) If f : X 0 → X is flat, E a
vector bundle on X, α ∈ A∗ X
si(f∗ E) ∩ f∗ α = f∗ (si(E) ∩ α).
(a) and (c) proved last time. (d) Exercise. Before proving (b), let me mention some
useful consequences.
Corollary. The flat pullback p∗ : AkX → Ak+e(PE) is a split monomorphism: by (a) (ii), an
inverse is β 7→ p∗ (c1(OPE(1))e ∩ β).
(We’re going to use this soon in the proof of the splitting principle, so don’t forget this.
We’ll also soon see that AmPE =∼ ⊕e Am−iX. The inclusion Am−iX ,→ AmPE will be given
i=0
e−i
by c1(OPE(1))i ∩ β. The projection will be given by: “cap with c1(OPE and push forward.)
2
Corollary. It makes sense to multiply by various polynomials in Segre classes of various
bundles, by the commutativity part (b).
Proof of (b). It won’t be surprising how we get commutativity. Consider the fibered
square:
QA
q0 }} AA p0
}} AA
}} AA
}~ } A
PE B PF
BB }}
BB }}
p BBB }
} q
}~ }
X
where p and q are the projections; all morphisms are projective bundles. Let f + 1 be the
rank of F (and as usual e + 1 is the rank of E). Then:
si(E) ∩ (sj(F) ∩ α) = p∗ (c1(OPE(1))e+i ∩ p∗ (q∗ (c1(OPF(1))f+j ∩ q∗ α)))
(left side of desired equality)
∗
= p∗ (c1(OPE(1))e+i ∩ q 0 ∗ (p 0 (c1(OPF(1))f+j ∩ q∗ α)))
(pr. pushforwards and fl. pullbacks “commute”)
∗ ∗ ∗
= p∗ q∗0 (c1(q 0 OPE(1))e+i ∩ (p 0 c1(OPF(1))f+j ∩ p 0 q∗ α))
(proj. form. and flat pull. behaves well w.r.t. c1)
∗ ∗ ∗
= q∗ p∗0 (p 0 c1(OPF(1))f+j ∩ (c1(q 0 OPE(1))e+i ∩ p 0 q∗ α))
(prop. pushforwards commute, and c1’s commute)
= (then go backwards to get desired result)
(Hint: Identify PE with P(E ⊗ L), with universal subbundle OPE(−1) ⊗ p∗ L. Then sp(E ⊗
L) ∩ α = p∗ ((c1(OPE(1)) − c1(p∗ L))e+p ∩ p∗ α).)
1.2. Chern classes. We now define Chern classes. Define the Segre power series st(E) to
be the generating function of the si:
∞
X
st(E) = si(E)ti = 1 + s1(E)t + s2(E)t2 + · · · .
i=0
Define the Chern power series (soon to be Chern polynomial!) as the inverse of st(E):
∞
X
ct(E) = ci(E)ti = 1 + c1(E)t + c2(E)t2 + · · · .
i=0
ct(E)st(E) = 1.
3
Hence c0(E) = 1, c1(E) = −s1(E), c2(E) = s1(E)2 − s2(E), . . . ,
cn(E) = −s1(E)cn−1(E) − s2(E)cn−2(E) − · · · − sn(E).
Note: The old-fashioned definition of c1(L) agrees with the new definition of c1(L), by the
last part of the previous Theorem.
Chern class Theorem. The Chern classes satisfy the following properties.
(a) (vanishing) For all bundles E on X, and all i > rank E, ci(E) = 0.
(c) (projection formula, i.e. Chern classes behave well with respect to proper pushfor-
ward) Let E be a vector bundle on X, f : X 0 → X a proper morphism. Then
f∗ (ci(f∗ E) ∩ α) = ci(E) ∩ f∗ (α)
for all cycles α on X 0 and all i.
(d) (Chern classes behave well with respect to flat pullback) Let E be a vector bundle
on X, f : X 0 → X a flat morphism. Then
ci(f∗ E) ∩ f∗ α = f∗ (ci(E) ∩ α)
for all cycles α on X, and all i.
(b), (c), and (d) follow from the Segre class theorem above. I explained (f) last time.
Thus we have to show (a) and (e). I’ll set up the right way of thinking about (a) and (e),
and then prove them next day.
Splitting principle. This uses a very nice (and very important) construction, the splitting
principle. It is not true that the every vector bundle splits into a direct sum of line bundles.
However, the splitting principle in essence tells that we can pretend that every vector
bundle splits, not into a direct sum, but into a nice filtration.
4
Injectivity shows that if we can show some equality involving Chern classes on the pull-
back to X 0 , then it will imply the same equality downstairs on X.
The construction is pretty simple: it will be a tower of projective bundles. Recall that
we showed earlier today that if F is any vector bundle on Y, and g : PF → Y, then g∗ :
AkX 7→ Ak+ePE is an injection, so we’ll get (1) immediately. We’ll constructive the tower of
projective bundles inductively on the rank of E. If r = 1, we’re already done. Otherwise,
let g : PE → X. On PE, we can split off the tautological subline bundle.
0 / OPE(−1) / g∗ E / Q / 0
PE
g
X
Here Q is the quotient bundle of rank one less than that of E.
Thus we’ve shown how to split a single vector bundle. Clearly we can split any finite
number of vector bundles in this way as well.
Lemma. Assume that E is filtered with line bundle quotients L1, . . . , Lr. Let s be a section
of E, and let Z be the closed subset of X where s vanishes. Then for any k-cycle α on X,
there is a (k − r)-cycle β on Z with
Yr
c1(Li) ∩ α = β
i=1
in Ak−rX. (Even better, we will see that we will get equality in Ak−r(Z): we have pinned
down (or “localized”) this class even further.) In particular, if s is nowhere zero, then
Q r
i=1 c1(Li) = 0. (Recall r = rank E.)
I suggested that people browse through the many examples in this chapter, including
the Chern character and the Todd class.
E-mail address: [email protected]
5
INTERSECTION THEORY CLASS 10
RAVI VAKIL
C ONTENTS
1. Last time 1
2. Chern classes 1
2.1. Fun with the splitting principle 4
2.2. The Chern character and Todd class 6
2.3. Looking forward to next day: Rational equivalence on bundles 6
1. L AST TIME
Corollary to Segre class theorem. The flat pullback p∗ : AkX → Ak+e(PE) is a split
monomorphism: by (a) (ii), an inverse is β 7→ p∗ (c1(OPE(1))e ∩ β).
2. C HERN CLASSES
We then defined Chern classes. Define the Segre power series st(E) to be the generating
function of the si. Define the Chern power series (soon to be Chern polynomial!) as the
inverse of st(E).
We’re in the process of proving parts of the Chern class theorem. Left to do:
Chern class Theorem. The Chern classes satisfy the following properties.
(a) (vanishing) For all bundles E on X, and all i > rank E, ci(E) = 0.
1
P
of vector bundles on X, then ct(E) = ct(E 0 ) · ct(E 00 ), i.e. ck(E) = i+j=k ci(E
0
)cj(E 00 ).
Notation. The Chern classes and Segre classes of all vector bundles determine a ring of
operators on Chow groups. I won’t give this ring a name (or I may tentatively call it the
Segre-Chern ring); later we will define a ring A∗ X of operators, in which these Chern and
Segre classes will lie.
Splitting principle. I introduced the splitting principle, which tells that we can pretend that
every vector bundle splits, not into a direct sum, but into a nice filtration.
f∗ E = Er ⊃ Er−1 ⊃ · · · ⊃ E1 ⊃ E0 = 0.
Injectivity shows that if we can show some equality involving Chern classes on the pull-
back to X 0 , then it will imply the same equality downstairs on X.
I should have said explicitly: we’ve shown how to split a single vector bundle. But
clearly we can split any finite number of vector bundles in this way as well.
Lemma. Assume that E is filtered with line bundle quotients L1, . . . , Lr. Let s be a section
of E, and let Z be the closed subset of X where s vanishes. Then for any k-cycle α on X,
there is a (k − r)-cycle class β on Z (i.e. an element of Ak−rZ) with
r
Y
c1(Li) ∩ α = β
i=1
in Ak−rX. (Even better, we will see that we will get equality in Ak−r(Z): we have pinned
down (or “localized”) this class even further.) In particular, if s is nowhere zero, then
Q r
i=1 c1(Li) = 0. (Recall r = rank E.)
Proof. For simplicity of exposition, let me show you how this works for r = 2. We have
0 → L1 → E → L2 = 0. The section s of E induces a section s of L2. If Y is the zero scheme
of s, then (L2, Y, s) is a pseudodivisor D2 on X. Let j : Y ,→ X be the closed immersion.
Intersecting with D2 gives a class D2 · α in Ak−1Y such that c1(L2) ∩ α = j∗ (D2 · α). By the
projection formula (“proper pushforward behaves with respect to c1”):
The bundle L1Y = j∗ E has a section, induced by s, whose zero set is Z. So c1(j∗ L1) ∩ (D2 ·
α) ∈ Ak−2Z as desired.
We’ll now unwind this to get the result. Let ζ = c1(OPE(1)) for convenience. Let σi be
the ith symmetric function in c1(L1), . . . , c1(Lr). Let σ̃i be the ith symmetric function in
c1(p∗L1), . . . , c1(p∗Lr).
We know that c1(p∗ Li ⊗OPE(1)) = c1(p∗ Li)+c1(OPE(1)) = c1(p∗ Li)+ζ. Hence we know:
ζr + σ̃1ζr−1 + · · · + σ̃r = 0.
(We feel like turning ζ into 1/t and using injectivity. That’s in spirit what we’ll do.) Mul-
tiply by ζi−1 for some i. Pick any α ∈ A∗ X, and cap the equation with p∗ α. Then pushfor-
ward:
p∗ (ζe+i ∩ p∗ α) + p∗ (σ̃1ζe+i−1 ∩ p∗ α) + · · · + p∗ (σ̃rζi−1 ∩ p∗ α) = 0.
Thus these are Segre classes:
(1) si(E) ∩ α + σ1si−1(E) ∩ α + · · · + σrsi−r(E) ∩ α = 0.
Multiply this by a formal variable ti, and add up over all i to get:
(1 + σ1t + · · · + σrtr)st(E) = 0.
Oops, that wasn’t quite right! Equation (1) holds for i > 0, so in fact
(1 + σ1t + · · · + σrtr)st(E) = constant.
But that constant is 1. Thus by the definition of ct(E), we get our desired result: ct(E) =
1 + σ1t + · · · + σrtr.
I’m now finally ready to prove (a) and (e) of theQChern class theorem. It suffices to prove
(a) assuming that E is filtered. But then ct(E) = ri=1(1 + c1(Li)t) is clearly a polynomial
of degree at most r — we’ve proved (a).
3
(e) is also easy. Given an exact sequence of vector bundles as in the statement, pullback
to a flat f : X 0 → X so that both the (pullback of the) kernel E 0 and the (pullback of the)
cokernel E 00 split into line bundles. Then the pullback of E also splits. Thus by the lemma,
ct(f∗ E) = ct(f∗ E 0 )ct(f∗ E 00 ).
(Reason: we can interpret each factor c1(i∗ OPN (1)) as intersecting with a randomly chosen
hyperplane.)
2.1. Fun with the splitting principle. Thanks to the splitting principle, given the Chern
classes of a vector bundle, you can find the Chern classes of other related vector bundles.
The way I think about it: imagine that the Chern polynomial factors (even though it
doesn’t!). Imagine that the bundle splits (even though it doesn’t!).
Example 1: Dual bundle. Suppose E is a vector bundle, and E∨ is the dual bundle. Then
ci(E∨) = (−1)ici(E). (Reason: ct(E) = c−t(E). The reason for this in turn is that if you
assume that E is filtered (which we may do by the splitting principle) then E∨ is filtered
too. Do you see why?
Example 2: Tensor products. I’ll do a specific example, in the hope that you’ll see the
general pattern. Suppose E and F are rank 2 bundles. Then E ⊗ F is a rank 4 bundle. We
can compute its Chern classes in terms of those of E and F. Suppose E has Chern roots
e1 and e2, and suppose F has Chern roots f1 and f2. (Translation: assume that both E
and F can be filtered. Let e1 and e2 be the line bundle quotients of the filtration of E, and
similarly for f1 and f2.) Thus from
1 + c1(E)t + c2(E)t2 = (1 + e1t)(1 + e2t)
we get e1 + e2 = c1(E) and e2 = c2(E), and similarly for F. Then
ct(E ⊗ F) = (1 + (e1 + f1)t)(1 + (e1 + f2)t)(1 + (e2 + f1)t)(1 + (e2 + f2)t)
= 1 + (2e1 + 2e2 + 2f1 + 2f2)t + · · ·
= 1 + (2c1(E) + 2c1(F))t + · · ·
from which we get c1(E ⊗ F) = 2c1(E) + 2c1(F), and similarly we can compute formulae
for higher Chern classes of E ⊗ F.
4
To justify that first equality for ct(E ⊗ F), we need to give a filtration of E ⊗ F using the
filtrations of E and F. I’ll leave that for you.
Example 3: Exterior powers. I’ll again do a specific example to illustrate a general princi-
ple. Suppose E is rank 3, with Chern roots e1, e2, e3. In other words, as assume we have
a specific filtration of E. The ∧2E is also rank 3, with Chern roots e1 + e2, e1 + e3, e2 + e3.
Again, we do this by producing a filtration of ∧2E induced by that filtration on E.
Thus we can find the Chern classes of ∧2E in terms of those of E. We know e1 +e2 +e3 =
c1(E), e1e2 + e2e3 + e3e1 = c2(E), and e1e2e3 = c3(E). Thus
In general, if E is rank n and we want to compute the Chern classes of ∧kE, the roots are
sums of k distinct Chern roots of E.
Exercise: if E is rank n, then you can check that ∧nE = det E. Show that c1(E) =
c1(det E). This gives a different interpretation of c1 of a vector bundle — as c1 of the
determinant bundle.
Exercise: what about symmetric powers? If E is rank 2, can you compute the Chern
classes of Sym4 E?
Homework (due Nov. 1.) Suppose E is a bundle of rank r on a scheme X, p is the projection
PE → X, and ζ = c1(OPE(1)). Show that ζr+c1(p∗ E)ζr−1+· · ·+cr(p∗ E) = 0. (Hint: consider
the exact sequence of vector bundles on PE: 0 → OPE(−1) → p∗ E → Q → 0.)
For convenience let, H = c1(OPn (1)). Hence ct(TPn ) = (1+Ht)n+1. (Note that deg cn(TPn ) =
n + 1, which is indeed the topological Euler characteristic of Pn.)
0 → TY → TX|Y → N → 0.
∼ OX(Y)).
(N =
Suppose next that X = Pn, and Y is a degree d hypersurface. Let H denote the restriction
of c1(OPn (1)) to Y. (Equivalently, it is c1 of the pullback of OPn (1) to Y: we’ve shown that
c1 commutes with any pullback.) Then as operators on A∗ Y, we get
You can use this to compute the topological Euler characteristic of a hypersurface, or
inductively, of a complete intersection. (Fun exercise: use this to work out the genus of a
degree d plane curve.)
5
2.2. The Chern character and Todd class. The Chern character ch is defined by ch(E) =
P r αi 0 00
i=1 e . Then if 0 → E → E → E → 0 is a short exact sequence of vector bundles,
ch(E) = ch(E 0) + ch(E 00 ). (You should immediately see the corresponding long exact
sequence!) Also, ch(E ⊗ E 0 ) = ch(E)ch(E 0).
Q
The Todd class is defined by td(E) = ri=1 Q(αi) where
X ∞
x 1 Bk 2k
Q(x) = −x
= 1 + x + (−1)k−1 x .
1−e 2 k=1
(2k)!
The first two cases are relative straightforward: if n = 1, then this corresponds to curves
in planes; the only way for a genus 1 curve to lie in P2 is if it is degree 3.
It can sit in P4, but we’ll see that it can only sit as a degree 10 hypersurface, and there is
a famous such example called the Horrocks-Mumford abelian variety.
2.3. Looking forward to next day: Rational equivalence on bundles. I stated a couple
of things that we’ll do on Wednesday.
Definition: Gysin pullback by zero section of a vector bundle. Let s = sE denote the
zero section of a vector bundle E. s is a morphism from X to E with π ◦ s = idX. By part (a)
of the Chern class theorem allows us to define Gysin homomorphisms s∗ : AkE → Ak−rX,
r = rank E, by s∗ (β) := (π∗ )−1(β).
E-mail address: [email protected]
7
INTERSECTION THEORY CLASS 11
RAVI VAKIL
C ONTENTS
Pe i ∗
θE : ⊕αi 7→ i=0 c1(OPE(1)) p αi.
Proof. Here’s the plan: π∗ surjective, θE surjective, θE injective, π∗ injective. So the proof is
a delicate interplay between E and PE.
We’ll make repeated use of something Rob stated, from the end of the first Chapter: the
“excision exact sequence”. Suppose X is a scheme, U an open set, and Z the complement
(a closed subset). Then the following sequence is exact:
AkZ → AkX → AkU → 0.
I’ll now show surjectivity of π∗ and θE. First reduction: it suffices to deal with the case
where E is the trivial bundle. Proof by the induction on the dimension of X. Here’s the π∗
argument:
Date: Wednesday, October 27, 2004.
1
Let U be a dense open set where E is trivial. Then its complement Y is of dimension
strictly smaller than X.
A∗ Y / A∗ X / A∗ U / 0
−1
A∗ (π Y) / A∗ E / A∗ (π−1U) / 0
?
0 0 0
The two horizontal rows are exact. By the inductive hypothesis, the left column is exact.
We’re assuming we know the result for trivial vector bundles, so the right column is also
exact. Then the central vertical row is exact, by a quick diagram chase.
The same argument works for θE. Here’s the exact sequence:
A∗ Y / A∗ X / A∗ U / 0
⊕A∗ (p−1Y) / ⊕A∗ PE / ⊕A∗ (p−1U) / 0
?
0 0 0
So let’s show surjectivity of π∗ and θE in the case where E is a trivial bundle. I’ll show
both by induction on the rank of E. In the case where the rank is 0, both are clearly
surjective. (In fact, π∗ is tautologically an isomorphism, and PE is the empty set, and the
left side of θE is the empty direct sum!)
The surjectivity of π∗ in the trivial bundle was shown in Chapter 1, so for the sake of
time I’ll omit it. (The atomic statement that needs to be shown: AkX → Ak+1(X × A1) is
surjective. Then by induction AkX → Ak+1(X × An) is surjective.)
`
Recall that P(E ⊕ 1) = PE E, where PE is a closed subset and E is an open subset; let
i : PE ,→ P(E ⊕ 1) be the closed immersion, and j : E ,→ P(E ⊕ 1) be the open immersion.
(In fact PE is a Cartier divisor, in class OP(E⊕1)(1); this was one of my definitions of O(1).)
Let q be the morphism P(E ⊕ 1) → X. The excision exact sequence gives us:
i∗
AkPE / AkP(E ⊕ 1) /A E
8 k
/ 0
O rr
π∗ rrr
q∗ rr
rrr
Ak−rX
You may feel like drawing an arrow Ak−rX → AkP, but that’s not right; the morphism is
of course Ak−rX → Ak−1P, as the fiber dimension of AkPE → Ak−r is r − 1.
2
Remark: For any α ∈ A∗ X, c1(OP(E⊕1)(1)) ∩ q∗ α = i∗ p∗ α. Reason: I’ll show this for any
cycle α ∈ Z∗ X. Then we can interpret the left side as pulling the cycle back to P(E ⊕ 1),
and intersecting with the Cartier divisor PE. But that’s exactly the same as the right side.
(That’s basically how we defined c1 of a line bundle!)
(the last step by using the remark). Thus we see that θE⊕1 is surjective.
If the αi are not all zero, then let k be the largest integers with αk 6= 0. Then
e
X
p∗ (c1(OPE(1))e−k ∩ c1(OPE(1))i ∩ p∗ αi) = αk
i=0
−1
A∗ (π Y) / A∗ E / A∗ (π−1U) / 0
0 0 0
3
Now we’ll do it by induction on the rank. So we want to show that AkX ,→ Ak+1X ×
A1 ,→ Ak+2X × A2 ,→ · · · : we just need to show the rank 1 case. Suppose α ∈ AkX and
π∗ α ∈ Ak+1(X × A1) = 0. Consider q∗ α ∈ Ak+1(X × P1). As θE is an isomorphism, we have
q∗ α = i∗ (p∗ α0 + c1(OP1 (1)) ∩ p∗ α1).
(One nice thing about P1 is that OP1 (1)2 = 0: the intersection of two distinct points is
empty!) Using our remark:
q∗ α = c1(OP1 (1)) ∩ q∗ α0 + c1(OP1 (1))2 ∩ p∗ α1.
(Thus by injectivity of θE (which is uniqueness of α0 and α1) we have α1 = 0.) But the
first part of the Segre class theorem stated that if we take a class α downstairs, pull it back
to a projective bundle, and cap it with the right number of O(1)’s (corresponding to the
projective bundle), and push it forward, we’ll get α again. Hence
α = q∗ (c1(OP1 (1)) ∩ q∗ α)
= c1(OP1 (1))2 ∩ q∗ α0 + c1(OP1 (1))3 ∩ p∗ α1
= 0
4
INTERSECTION THEORY CLASS 12
RAVI VAKIL
C ONTENTS
Proof. Our plan was to prove this in the following order: π∗ surjective, θE surjective, θE
injective, π∗ injective. The proof is a delicate interplay between E and PE. We had done all
but the last step, and we had reduced the last step to the case where E is a trivial bundle,
i.e. we wanted to show that A∗ X ,→ A∗ (X × Ar). By induction, we needed to deal with the
case where E had rank 1.
Date: Monday, November 1, 2004.
1
We repeatedly used the “excision exact sequence”. Suppose X is a scheme, U an open
set, and Z the complement (a closed subset). Then the following sequence is exact:
AkZ → AkX → AkU → 0.
A`construction we used throughout the proof was the following: Note that P(E ⊕ 1) =
PE E, where PE is a closed subset and E is an open subset; let i : PE ,→ P(E ⊕ 1) be the
closed immersion, and j : E ,→ P(E ⊕ 1) be the open immersion. (In fact PE is a Cartier
divisor, in class OP(E⊕1)(1); this was one of my definitions of O(1).) Let q be the morphism
P(E ⊕ 1) → X. The excision exact sequence gives us:
i∗
AkPE / AkP(E ⊕ 1) /A E
8 k
/ 0
O rr
π∗ rrr
q∗ rr
r
rr
Ak−rX
1.1. Intersecting with the zero-section of a vector bundle. We can already intersect with
the zero-section of a line bundle (i.e. an effective Cartier divisor); we get a map AkX →
Ak−1D, which we’ve called the Gysin pullback.
Definition: Gysin pullback by zero section of a vector bundle. Let s = sE denote the
zero section of a vector bundle E. s is a morphism from X to E with π ◦ s = idX. By part (a)
of the Chern class theorem allows us to define Gysin homomorphisms s∗ : AkE → Ak−rX,
r = rank E, by s∗ (β) := (π∗ )−1(β).
This ability to intersect with zero sections of vector bundles will be the basis for many
important future constructions.
You should think of this as intersecting with the zero-section of a vector bundle. This
should be a codimension r intersection. In fact there is “excess” intersection — the actual
intersection is codimension 0 — but there is a class of the right dimension.
Proposition. Let β ∈ AkE, and let β be any element of Ak(P(E ⊕ 1)) which restricts to β
in AkE. Then s∗ (β) = q∗ (cr(Q) ∩ β) where q is the projection from P(E ⊕ 1) to X, and Q is
the universal (rank r) quotient bundle of q∗ (E ⊕ 1).
Proof omitted (but is in book, and isn’t too long). Note that cr is the “top” Chern class.
In the case when X is a smooth subvariety of a smooth variety, C is the normal bundle.
More generally, if Y is arbitrary, then X is a local complete intersection (hereafter lci) in Y
(what Fulton calls a regular imbedding) if it is scheme-theoretically cut out by r equations,
where r is the codimension of X in Y. (Example 1: X is a smooth subvariety of a smooth
variety. Example 2: any Cartier divisor. Example 3: the union of the x and y axes in A3.) If
X is a regular imbedding (=lci) in X, then X still has a normal bundle, defined as follows:
if I is the ideal sheaf cutting out X, then I/I 2 is a vector bundle of rank r. This is the
conormal bundle, and its dual is the normal bundle. (Warning: in differential geometry, if
X ,→ Y, then X has a tubular neighborhood that looks like the normal bundle. In algebraic
geometry, there are no such small neighborhoods, but in some sense it is even worse: in
example 3, the total space Y = A3 is smooth, but the total space of the normal bundle — a
vector bundle over a nodal curve — is singular.)
2.2. Cones. I’ll now define cone. Let X be a scheme, and let S· = ⊕i≥0Si be a sheaf of
graded OX-algebras. Assume OX → S0 is surjective, S1 is coherent, and S· is generated
(as an algebra) by S1. This sounds complicated, but it isn’t. It is defined so you can take
Proj(S· ), and that this makes sense, and has a line bundle O(1).
Here’s how it works: over any affine open set Spec R of X, S· is a graded R-algebra,
generated in degree 1. Then we can take Proj of this graded R-algebra. The fact that
the algebra is generated in degree 1 (by R1 say) means that we have a surjective map of
graded rings
Symi R1 → ⊕iRi
3
which, upon applying Proj, becomes
X 0 ,→ X × P(R1)∨
where P(R1)∨ is an honest projective bundle. So the morphism X 0 → X is projective
and has a line bundle called O(1). You can do this over each affine, and glue the result
together, and the O(1)’s also glue together.
Example 1: say let E be a vector bundle, and Si = Symi(E∨). Then Proj S· = PE.
Example 3: The blow-up can be described in this way, and it will be good to know
this. Suppose X is a subscheme of Y, cut out by ideal sheaf I. (In our situation where
all schemes are finite type, I is a coherent sheaf.) Then let S· = ⊕iI i, where I is the ith
power of the ideal I. (I 0 is defined to be OX.) Then BlX Y = ∼ Proj S· . A short calculation
shows that the exceptional divisor class is O(−1). The exceptional divisor turns out to be
Proj ⊕I n/I n+1. (Note that this is indeed a graded sheaf of algebras.) As ⊕I n → ⊕I n/I n+1
is a surjective map of rings, this indeed describes a closed subscheme of the blow-up.
(Remember this formula — it will come up again soon!)
Now I’ll finally define cone. Let S· be a sheaf of graded OX-algebras as before. Then
C = Spec(S· ) is a cone. (We can construct Spec(S· ) of a sheaf of algebras in the same way
as we can construct Proj; in fact it is a logically prior construction.)
` ·
`
Remember
` that P(E⊕1) = E PE. The direct generalization is: Proj(S [z]) = C Proj(S· ) =
Spec S· Proj(S· ). The argument is just the same. The right term is a Cartier divisor in
class OProj(S· [z])(1).
2.3. Segre class of a cone. The Segre class of a cone C on X, denoted s(C), is the class in
A∗ X defined by the formula
!
X
i
s(C) = q∗ c1(O(1)) ∩ [Proj(C ⊕ 1)] .
i≥0
This is very much the same definition as for vector bundles, except in the vector bundle
case we get operators on Chow groups. In this case we get elements of Chow groups
themselves: we are capping with a fundamental class!
Proposition (a) If E is a vector bundle on X, then s(E) = c(E)−1 ∩ [X], where c(E) is
the total Chern class of X, r = rank(E). c(E) = 1 + c1(E) + · · · + cr(E). (I would write
s(E) = s(E) ∩ [X], but the two uses of s(E) are confusing!) This is basically our definition
of Segre/Chern classes.
Example. For any cone C, s(C ⊕ 1) = s(C). (In the language of Dan’s talk last week, the
Segre class of a cone depends on its stable equivalence class.)
2.4. The Segre class of a subscheme. Let X be a closed subscheme of a scheme Y (not
necessarily lci).
I told you that I/I 2 is the conormal bundle of a local complete intersection subscheme.
In general, it is the conormal sheaf.
P n n+1
Consider ∞ n=0 I /I . (Recall that Proj of this sheaf gives us the exceptional divisor
of the blow-up.) Define the normal cone C = CXY by
∞
X
C = Spec I n/I n+1.
n=0
Define the Segre class of X in Y as the Segre class of the normal cone:
s(X, Y) = s(CXY) ∈ A∗ X.
(a) If f is proper, Y irreducible, and f maps each irreducible component of Y 0 onto Y then
Let me point out why I find this a remarkable result. X 0 is a priori some nasty scheme;
even if it is nice, its codimension in Y 0 isn’t necessarily the same as the codimension of
X in Y. The argument is quite short, and shows that what we’ve proved already is quite
sophisticated.
I will give the proof next time. Today I gave most of the proof, by describing the dia-
gram around which everything revolves.
Let me assume that Y 0 is irreducible. (It’s true in general, and I may deal with the
general case later.)
5
Let me first write the diagram on the board, and then explain it.
OProj(C0 ⊕1)(1) = G∗ OProj(C⊕1)(1)
UUUU
UUUU
UUUU
UUUU
U*
OProj(C⊕1)(1) Proj(C 0 ⊕ 1)Cartier div.
/ Bl 0 (Y 0 × A1)
X ×0
UUUU rr
UUUU
UUUU rrrr G F
UUrUrU
r r U*
Proj(C ⊕ 1) Cartier div.
r
rr 1
X×0(Y × A )
rrq0 / Bl
rr
rr qq
rr qqq
rrr q q
rr qq
yrr qqq
0
X qq
q q
qq qqq
g q
qqq
xqqqq
X
Explanation: We blow up Y × A1 along X × 0, and similarly for Y 0 and X 0 . The exceptional
divisor of BlX×0(Y × A1) is Proj(C ⊕ 1), and similarly for Y 0 and X 0 . The universal property
of blowing up Y × A1 shows that there exists a unique morphism G from the top excep-
tional divisor to the bottom. Moreover, by construction, the exceptional divisor upstairs
is the pullback of the exceptional divisor downstairs (that’s the statement about the two
O(1)’s in the diagram). Let q be the morphism from the exceptional divisor Proj(C ⊕ 1)
to X, and similarly for q 0 . That square commutes: q ◦ G = g ◦ q 0 (basically because that
morphism G was defined by the universal property of blowing up).
We’ll finish the proof next time (and I’ll describe this diagram once again).
E-mail address: [email protected]
6
INTERSECTION THEORY CLASS 13
RAVI VAKIL
C ONTENTS
1. Where we are: Segre classes of vector bundles, and Segre classes of cones 1
2. The normal cone, and the Segre class of a subvariety 3
3. Segre classes behave well with respect to proper and flat morphisms 3
We first defined Segre class of vector bundles over an arbitrary scheme X. If E is a vector
bundle, we get an operator on class on X. We define it by projectivizing E, so we have
a flat and proper morphism PE → X, pulling back α to PE, capping with O(1) a certain
number of times, and pushing forward.
Hence we get si(E)∩ : AkX → Ak−iX, and for example we checked the non-immediate
fact that s0(E) is the identity. (Recall s0 involved pulling back, capping with precisely
rank E − 1 copies of O(1), and then pushing forward.) Note that sk(E) = sk(E ⊕ 1), as the
Whitney product formula gives s(E ⊕ 1) = s(E)s(1) = s(E).
We want to generalize this to cones. Here again is the definition of a cone on a scheme
X. Let S· = ⊕i≥0Si be a sheaf of graded OX-algebras. Assume OX → S0 is surjective, S1 is
coherent, and S· is generated (as an algebra) by S1. Then you can define Proj(S· ), which
has a line bundle O(1). Proj(S· ) → X is a projective (hence proper) morphism, but it isn’t
necessarily flat! (Draw a picture, where the cone has components of different dimension.)
Flat morphisms have equidimensional fibers, and cones needn’t have this.
A couple of important points, brought out by Joe and Soren. I’ve been imprecise with
terminology. Although one often sees phrases such as “the cone is C = Spec(S· )”, we lose
a little information this way; the cone should be defined to be the graded sheaf S· . The
∗
sheaf can be recovered from CXY along with the action of the multiplicative group OX ;
the nth graded piece is the part of the algebra where the multiplicative group acts with
weight n.
Example 1: say let E be a vector bundle, and Si = Symi(E∨). Then Proj S· = PE. Example
2: Say T i = Symi(E∨ ⊕ 1) = Si ⊕ Si−1z, so (better) T · = S· [z]. Then Proj T · = PE. Example 3:
Date: Wednesday, November 3, 2004.
1
` `
Proj(S· [z]) = C Proj(S· ) = Spec S· Proj(S· ). The argument is just the same. The right
term is a Cartier divisor in class OProj(S· [z])(1). Example 4: The blow-up can be described
in this way, and it will be good to know this. Suppose X is a subscheme of Y, cut out by
ideal sheaf I. (In our situation where all schemes are finite type, I is a coherent sheaf.)
Then let S· = ⊕iI i, where I is the ith power of the ideal I. (I 0 is defined to be OX.) Then
BlX Y =∼ Proj S· . A short calculation shows that the exceptional divisor class is O(−1).
The exceptional divisor turns out to be Proj ⊕I n/I n+1. (Note that this is indeed a graded
sheaf of algebras.) As ⊕I n → ⊕I n/I n+1 is a surjective map of rings, this indeed describes
a closed subscheme of the blow-up. (Remember this formula — it will come up again
soon!)
So the same construction of Segre classes of vector bundles doesn’t work: there is no
flat pullback to Proj(S· ). So what do we do?
Idea (slightly wrong): We can’t pull classes back to Proj(S· ). But there is a natural class
up there already: the fundamental class. So we define
?
X
s(C) = q∗ ( c1(O(1))i ∩ [Proj C])
i≥0
where q is the morphism Proj C → X. Instead, as Segre class of vector bundles are stable
with respect to adding trivial bundles, we define
X
s(C) := q∗ ( c1(O(1))i ∩ [Proj(C ⊕ 1)])
i≥0
where q is the morphism Proj(C ⊕ 1) → X. Why is adding in this trivial factor the right
thing to do? Partial reason: if C is the 0 cone, i.e. Si = 0 for i > 0, then Proj C is empty,
but Proj C ⊕ 1 is not; we get different answers. But if you add more 1’s, you will then get
the same answer: s(C ⊕ 1 ⊕ · · · ⊕ 1) = s(C).
Proposition. (a) If E is a vector bundle on X, then s(E) = c(E)−1 ∩ [X], where c(E) is
the total Chern class of X, r = rank(E). c(E) = 1 + c1(E) + · · · + cr(E). (I would write
s(E) = s(E) ∩ [X], but the two uses of s(E) are confusing!) This is basically our definition
of Segre/Chern classes.
Let X be a closed subscheme of a scheme Y (not necessarily lci = local complete inter-
section), cut out by ideal sheaf I.
Define the Segre class of X in Y as the Segre class of the normal cone:
s(X, Y) = s(CXY) ∈ A∗ X.
If X is regularly imbedded (=lci) in Y, then the definition of s(X, Y) is
s(X, Y) = s(N) ∩ [X] = c(N)−1 ∩ [X].
The following geometric picture will come up in the central construction in intersection
(the deformation to the normal cone). X × A1 ,→ Y × A1. Then blow up X × 0 in Y × A1.
The ideal sheaf of X × 0 is I[t], where t is the coordinate on A1. Thus the normal cone to
X × 0 in Y × A1 is CXY[t]. Hence the exceptional divisor is Proj(CXY[t]) (draw a picture).
Inside it is the Cartier divisor t = 0, which is Proj(CXY).
3. S EGRE CLASSES BEHAVE WELL WITH RESPECT TO PROPER AND FLAT MORPHISMS
(a) If f proper, Y irreducible, and f maps each irreducible component of Y 0 onto Y then
g∗ (s(X 0 , Y 0 )) = deg(Y 0 /Y)s(X, Y).
Let me repeat why I find this a remarkable result. X 0 is a priori some nasty scheme;
even if it is nice, its codimension in Y 0 isn’t necessarily the same as the codimension of
X in Y. The argument is quite short, and shows that what we’ve proved already is quite
sophisticated.
3
As a special case, this result shows that Segre classes have a fundamental birational
invariance: if f : Y 0 → Y is a birational proper morphism, and X 0 = f−1X, then s(X 0 , Y 0 )
pushes forward to s(X, Y).
Proof. Let me assume that Y 0 is irreducible. (It’s true in general, and I may deal with the
general case later.)
Let me first write the diagram on the board, and then explain it.
Now f∗ [Y 0 × A1] = d[Y × A1] (where I am sloppily using the name f for the morphism
Y 0 × A1 → Y × A1). This is computed on a dense open set, so blow-up doesn’t change this
fact:
Now we’ve shown that proper pushforward commutes with intersecting with a
(pseudo-)Cartier divisor. Hence
We immediately have:
Corollary. With the same assumptions as the proposition, if X 0 is regular imbedded (=lci) in
Y 0 , with normal bundle N 0 , then
To see why the first part might matter: Suppose X ,→ Y is a very nasty closed immer-
sion. Then blow up Y along X, to get Y 0 with exceptional divisor X 0 . Then X 0 is regularly
imbedded (lci) in Y 0 — it is a Cartier divisor! This is the content of the next corollary.
5
Corollary. Let X be a open closed subscheme of a variety Y. Let Ỹ be the blow-up of Y
along X, X̃ = PC the exceptional divisor, η : X̃ → X the projection. Then
X
s(X, Y) = (−1)k−1η∗ (X̃k)
k≥1
X
= η∗ (c1(O(1))i ∩ [PC])
i≥0
In that first equation, the term X̃k should be interpreted as the kth self intersection of
the Cartier divisor X̃, also known as the exceptional divisor.
E-mail address: [email protected]
6
INTERSECTION THEORY CLASS 14
RAVI VAKIL
C ONTENTS
1. Where we are: Segre classes of vector bundles, and Segre classes of cones 1
1.1. Segre classes of cones 1
2. What the “functoriality of Segre classes of subschemes” buys us 2
2.1. The multiplicity of a variety along a subvariety 2
3. Deformation to the normal cone 3
3.1. The construction 3
4. Specialization to the normal cone 5
4.1. Gysin pullback for local complete intersections 6
4.2. Intersection products on smooth varieties! 6
1.1. Segre classes of cones. Once again, the definition of a cone on a scheme X. Let
S· = ⊕i≥0Si be a sheaf of graded OX-algebras. Assume OX → S0 is surjective, S1 is coher-
ent, and S· is generated (as an algebra) by S1. I’m happy calling this the cone. C = Spec S· .
Proj(S· ) has a line bundle O(1). (The “underline” under Spec and Proj is meant to dis-
tinguish the “sheafy” version from the usual version of these constructions.) Define the
Segre class
X
s(C) := q∗ ( c1(O(1))i ∩ [Proj(C ⊕ 1)])
i≥0
1
(a) If f proper, Y irreducible, and f maps each irreducible component of Y 0 onto Y then
g∗ (s(X 0 , Y 0 )) = deg(Y 0 /Y)s(X, Y).
As a special case, this result shows that Segre classes have a fundamental birational
invariance: if f : Y 0 → Y is a birational proper morphism, and X 0 = f−1X, then s(X 0 , Y 0 )
pushes forward to s(X, Y).
Corollary. With the same assumptions as the proposition, if X 0 is regular imbedded (=lci) in
Y 0 , with normal bundle N 0 , then
g∗ (c(N 0)−1 ∩ [X 0 ]) = deg(Y 0 /Y)s(X, Y).
If X ⊂ Y is also regularly imbedded, with normal bundle N, then
g∗ (c(N 0)−1 ∩ [X 0 ]) = deg(Y 0 /Y)(c(N)−1 ∩ [X]).
To see why the first part might matter: Suppose X ,→ Y is a very nasty closed immer-
sion. Then blow up Y along X, to get Y 0 with exceptional divisor X 0 . Then X 0 is regularly
imbedded (lci) in Y 0 — it is a Cartier divisor! This is the content of the next corollary.
In that first equation, the term X̃k should be interpreted as the kth self intersection of
the Cartier divisor X̃, also known as the exceptional divisor. In other words, it should be
interpreted as meaning the second line.
2.1. The multiplicity of a variety along a subvariety. We’ll now define them multiplicity
of a scheme Y along a subvariety X. (As a special case, this will define the multiplicity of
a variety at a closed (=old-fashioned) point. That special case is a fundamental commuta-
tive algebra notion due to Samuel.) If the general point of X is a smooth point of Y, we’ll
get 1. Definition: s(X, Y) ∈ A∗ X. Then s(X, Y) = eXY[X] + lower order terms. eXY is the
multiplicity.
2
Useful exercise: What is the multiplicity of (0, 0) in the cusp y2 − x3? Here the char-
acteristic is not 2 or 3. (Answer: 2. Hint: blow this up. The blow-up is Spec k[t] →
Spec k[x, y]/(y2 − x3) given by t 7→ (t2, t3).)
Example. If codim(X, Y) = n > 0, define the multiplicity eXY as follows. Let q be the
projection Proj(C ⊕ 1) → X and p be the projection Proj C → X.
eXY[X] = q∗ (c1(1))n ∩ [Proj(C ⊕ 1)]
= p∗ (c1(O(1))n−1 ∩ [Proj C]
= (−1)n−1p∗ (X̃n)
Here X̃ is the exceptional divisor of the blow-up.
Useful exercise: See that this works in for the cusp point (the previous useful exercise).
Note that as a vector space k[x, y]/(y2 − x3) = ⊕n≥0,n6=1ktn; note that the n = 2 term is kx,
the n = 3 term is ky, the n = 4 term is kx2, the n = 5 term is kxy, and the n = 6 term is
kx3 = ky2.
We next come to the central construction. There’s not much for us to do here, as we’ve
built up all the necessary machinery, and even seen the construction.
Here is the main goal. Suppose X → Y is a closed immersion of schemes. The idea is
that C = CXY “looks like Y near X”; it “is like a tubular neighborhood”. But it is nicer than
Y near X; in particular it is a cone.
I’ll try to give you an intuitive idea for what this means. (Try it.)
3.1. The construction. Here’s how we do it. Let me set some notation. If W ,→ Z is a
closed immersion, recall that BlW Z is the blow-up of Z along W. For the purposes of the
next few lectures, let EWZ be the exceptional divisor, and let IWZ be the ideal sheaf. Then
recall:
Suppose now that X ,→ Y is a closed immersion. (Other notation: when Fulton says
“imbedding”, we will say “closed immersion”.) Let’s blow up Y × P1 along X × 0 and see
what we get. (Here let t be a coordinate on P1. Notational caution: Fulton prefers to blow
up X × ∞.) We certainly have a morphism to P1:
BlX×0(Y × P1) → Y × P1 → P1.
Away from t = 0, the blow-up doesn’t do anything: BlX×0 Y|t6=0 = Y × (P1 − 0).
So what is the fiber over t = 0? I claim it is the union of two things, that we can
identify. One “piece” is BlX Y, with exceptional (Cartier) divisor EXY. The other piece is
Proj(CXY ⊕ 1); this has a Cartier divisor “P(CXY ⊕ 1)” = CXY ⊕ PCXY = ∼ EXY. We glue
these two pieces together along EXY.
I want to convince you that we really get these two pieces. If you’ve never seen this
before, I want to convince you that we get those two pieces, and you can be happy with
that. At the end I’ll explain how to verify that we get nothing else.
Consider the morphism BlX×0 Y × P1 → Y × P1. Away from the X × 0 on the target, this
is an isomorphism. The exceptional divisor is
EX×0 Y × P1 = Proj ⊕ (IX×0Y × P1)n/(IX×0Y × P1)n+1
Let’s next see the piece BlX Y. Translation: we want a morphism BlX Y to BlX×0(Y × P1)
that lies in the scheme-theoretic fiber t = 0, and we want this to be a closed immersion.
I will just show you that the morphism exists; as usual we use the universal property.
Consider the map BlX Y → Y × P1 obtained via BlX Y → Y × 0 ,→ Y × P1. The pullback of
X × 0 is an effective Cartier divisor EXY. Thus by the universal property of blowing-up,
we get a morphism BlX Y → BlX×0(Y × P1).
So I’ve given you an indication that we see both the projective completion of the normal
cone, and BlX Y, in the central fiber (t = 0) of BlX×0(Y × P1). How would you show that
this is all we get, and that they are glued together along EXY? This is a local question,
so we can take Y = Spec A, and X = Spec A/I. Then the question becomes completely
explicit: Y × A1 = Spec A[t]. (We can work locally in P1 as well.) Then BlX×0(Y × P1)
locally is Proj ⊕(I, t)n. We are interested in the fiber over t = 0, so we mod out by t:
ρ−1(0) = Proj ((⊕(I, t)n) / (t (⊕(I, t)n))) .
We want to show that this is the union of
EX×0(Y × P1) = Proj ⊕(I, t)n/(I, t)n+1
4
and
BlXY = Proj ⊕In
glued along
EXY = Proj ⊕ In/In+1 .
Consider ordered pairs of elements of the second and third graded rings, that are required
to give the same element in the fourth graded ring. Show that this ring is the same as the
first graded ring. Finally, realize that this algebraic statement is precisely the geometric
statement you want to prove. (I’m not going to give the details.)
Let me now do it. Let M◦ = BlX×0(Y × P1) − BlX Y. A picture is helpful here. Away
from 0, M◦ is still Y 1
`× A . Over 0, the big blow-up was the projectivized completion of the
normal cone CXY PCXY glued to BlX Y along EXY = PCXY. We’re throwing out BlX Y,
so the central fiber is now just the normal cone CXY. So we have really deformed Y to the
normal cone. Hence this scheme M◦ is often called the “deformation to the normal cone”.
Let i : C ,→ M◦ be the closed immersion of the normal cone, and let j : Y × (P1 − 0) ,→ M◦
be the open immersion of the complement.
Now we have shown i∗ i∗ : Ak+1C → AkC is the same as capping with c1 of the normal
(line) bundle to the divisor C in M◦ . But in this case the normal line bundle is trivial: it is
the pullback of the normal bundle to t = 0 in P1. Thus i∗ i∗ = 0. Hence Ak+1M◦ → AkC
descends to a map Ak+1(Y × A1) → AkC, and hence we get a map σ : AkY → AkC, which
is what we wanted! Here’s the final diagram:
i∗ j∗
Ak+1CK / Ak+1M◦ / Ak+1(Y × A1) / 0
KK o O
KKK ∴ oooo
K i∗ o o ∼
i i∗ =0 KKK%
∗ oo
ow oo
AkC o ∴σ
AkY.
Remark We could define this morphism more explicitly as follows. Define σ : ZkY → ZkC
by σ([V]) = [CV∩XV] where V is a subvariety of Y. (Extend this to ZkY by linearity.) Note
5
that CV∩XV ,→ CXY, so this makes sense. Proposition. This descends to the morphism
AkY → AkC I just defined. Sketch of proof. In the bottom row of that last big diagram, it
suffices to verify that [V] 7→ [CV∩XV]. Hence it suffices to show that in the “southwest”
morphism in the big diagram (marked “∴”), [V × A1] maps to [CV∩XV]. We take the
subvariety V × A1 ,→ Y × (P1 − 0), take its closure in M◦ , and intersect with the Cartier
divisor (t = 0) = C. We can do this explicitly locally on Y, using Y = Spec A, X = Spec A/I,
etc.; I’ll omit this since I don’t think we’ll need this fact.
4.1. Gysin pullback for local complete intersections. We already had defined the Gysin
pullback or Gysin homomorphism in the case where Y is a vector bundle over X. This
extends it to when “Y looks like a vector bundle over X”. Notice that the two definitions
agree; one needs to check that the normal cone to a the zero section of a vector bundle
is the vector bundle itself (which is true). Also, we showed that the Gysin pullback for
vector bundles satisfied all sorts of nice properties; if we show that σ satisfies these nice
properties too, then we’ll know it for Gysin pullbacks to local complete intersections.
Note also: If Y is purely n-dimensional, notice that i∗ [Y] = [X]. Because σ[Y] = [C], and
s∗N[C]
= [X].
I should probably be a bit clearer about that first map, which might reasonably be called
. (You can see a discussion in Chapter 1 if you want.) Here’s what we need: consider
the map
×
ZpX ⊗ ZqY / Zp+q(X × Y)
defined on varieties by [V] × [W] = [V × W], and defined generally by linearity. (We’ll
take X = Y, but we might as well do this in some generality.)
6
1. Linear Systems
Possible problems:
(1) Does a choice of coordinate on our line bundle L alter the map?
Nope, since projective space is nice under scalar action
(2) Projective space isn’t supposed to have a point [0; . . . ; 0], so
what do we do when x ∈ ker si for each si ?
This is a serious issue.
The good news about our second problem is that the offending set
is as nice as we could want: it is naturally the (closed) subscheme
of X which is cut out by the section si . This subscheme is the base
locus of V , which we’ll write as B. So we at least have a rational map
ϕ : X − B → Pr .
Example. Let X = P2 , and L = OP2 (2). Some sections: J :=
{x2 , xy, y 2, xz, yz, z 2 }. What’s the base locus? It’s nothing! Yeah!
So we get a bona fide map P2 → P5 .
Example. Let’s stick with our scheme and line bundle, but pick out
some new sections: R := {x2 , xy, y 2, xz, yz}. Will we get lucky and
not have a base locus again? No. This time B = [0; 0; 1]. So we get a
rational map P2 99K P4 defined away from [0; 0; 1].
We’re not happy with just having a rational map, since we think we
should be able to ‘fill in’ the missing information.
Example. Let’s return to the previous example for a minute, and see
what’s happening near the base locus. We’ll let [αt; βt; 1] be a point
near [0; 0; 1] (of course, we won’t let α = β = 0 yet). Where does this
point go?
[α2 t2 ; αβt2; βt2 ; αt; βt] = [α2 t; αβt; β 2t; α; β].
It seems that as we approach [0; 0; 1] along the line [αt; βt; 1], the map
is taking us to [0; 0; 0; α; β]. Hmm...seems that we’re getting an idea of
what happens to a tangent to our base locus...what does that make us
think of?
1
2
In fact, there is a way to extend the map we’ve been thinking about
to the blowup of X along B, X̃ = BlB X.
/
E X̃ = BlBKX
KK f
π
KK
KK
KK
K%
/ X _ _ _ _ _/ Pr
B
How can we get our hands on f ? We’re going to pull back the line
bundle L to BlB X and tweak it remove the base locus. Fact: the
bundle π ∗ L − E has no base locus. Now we’ll use this new line bundle
to map to Pr , and all will be well.
Example. Suppose our variety is P1 , and we’ve chosen the bundle
OP1 (2), with sections {x2 , xy}. This gives a map P1 99K P1 away from
[0; 1], though we ‘should’ know how to fix this map to make it a bona
fide map. How do we resolve this? Let’s follow the formula above: we’ll
blow up P1 along [0; 1] (this will just give us [0; 1] ,→ P1 back again,
since [0; 1] is cartier and cut out by x = 0), and twist our sections by
−E, which in this case means divide by x.
[0; 1] / P1 ?
?? [x;y]
??
π ??
[x2 ;xy]?
[0; 1] / P1 _ _ _/ P2
Useful reminder: We saw in class last day that, using our language
of the day,
X
s(B, X) = (−1)k−1 π∗ (E k ).
k≥1
Since this example has no base locus, we see that the degree of our
map to P5 is 4.
Example. Let’s return to the second example. Since the base locus
is [0; 1] we see that s(B, X) = [pt]. So our (extended) map has degree
4 − 1 = 3.
[yz;xz;xy]
Example. Cremona: Let’s try P2 −→ P2 . The base locus is the
subscheme yz = xz = xy = 0, otherwise known as the three reduced
points [1; 0; 0], etc. Each has segre class of a point, so that we get the
degree of the map 4 − 3 = 1. It’s a birational morphism! What’s the
inverse? Itself.
4
2. Intersection Product
In this section I’m going to tie up some of the loose ends Ravi left for
me on Monday. Recall that we’re interested in finding a map Ap (X) ⊗
×
Aq (X) → Ap+q−n (X). The things left for me are to clear up that there’s
×
a map Zk (X) ⊗ Zl (Y ) → Zk+l (X × Y ) and that this map gives us a
map on cycle classes.
For the first, we define the map by giving its action on subvarieties
and extending by linearity. We take [W ]×[V ] 7→ [W ×V ]. The studious
student asks why this product lands in the right cycle class.
dim W ×k V = dim (W ×k V ) ×k k̄ = dim (W ×k k̄) ×k̄ (V ×k k̄)
= dim W ×k k̄ + dim V ×k k̄ = dim W + dim V.
Here we have used two exercises from Chap 3 of Hartshorne and the
following diagram
/o /o o/ /o o/ o/ /o o/ /o o/ / V ×k k̄
9y 9y
9y 9y
9 y
< | |< <|
y 9
y 9y |< <|
# |
W ×k k̄ o / o / o / o / o / / o o / o / o / / o o / / o o / / k̄
W ×k V /V
qq xx
qqqq xxx
q x
q xx
x qq
q #/ x{ x
W k
• If α ∼ 0 or β ∼ 0, then α × β ∼ 0.
• The product f × g of proper (resp., flat) maps is again proper
(resp., flat), and (f ×g)∗ α×β = f∗ α×g∗ β (resp., (f ×g)∗ α×β =
f ∗ α × g ∗β).
RAVI VAKIL
C ONTENTS
1. Where we are 1
1.1. Deformation to the normal cone 1
1.2. Gysin pullback for local complete intersections 2
1.3. Intersection products on smooth varieties 3
2. Intersection Products 3
3. Refined Gysin homomorphisms 5
1. W HERE WE ARE
We’ve covered a lot of ground so far. I want to remind you that we’ve essentially de-
fined a very few things, and spent all our energy on showing that they behave well with
respect to each other. In particular: proper pushforward, flat pullback, c· , s· , s· (X, Y).
Gysin pullback for divisors; intersecting with pseudo-divisors. Gysin pullback for 0-
sections of vector bundles.
1.1. Deformation to the normal cone. This is the central construction. Suppose X → Y is
a closed immersion of schemes.
If W ,→ Z is a closed immersion, recall that BlW Z is the blow-up of Z along W. For the
purposes of the next few lectures, let EWZ be the exceptional divisor, and let IWZ be the
ideal sheaf. Then recall:
Date: Monday, November 15, 2004.
1
• BlW Z = Proj ⊕(IWZ)n
• EWZ = Proj ⊕(IWZ)n/(IWZ)n+1
• EWZ ,→ BlWZ is a closed immersion, and describes EWZ as an effective Cartier
divisor, in fact in class OProj ⊕(IW Z)n (1). The closed immersion is visible at the level
of graded algebras.
We throw out BlX Y: let M◦ = BlX×0(Y × P1) − BlX Y. (A picture is helpful here.) Away
from 0, M◦ is still Y ×A1. Over 0, we see the normal cone CXY. So we have really deformed
Y to the normal cone. Let i : C ,→ M◦ be the closed immersion of the normal cone, and let
j : Y × (P1 − 0) ,→ M◦ be the open immersion of the complement.
The argument from last week was slick enough that I’m going to repeat it (quickly).
Consider the following diagram:
i∗ j∗
Ak+1C / Ak+1M◦ / Ak+1(Y × A1) / 0
O
Gysin map for divisors i∗ ∼
AkC AkY.
The top row is the excision exact sequence. The right column is flat pullback and is an
isomorphism, as flat pullback to the total space of a line bundle is always an isomorphism.
The left column is the Gysin pullback map to divisors.
We have shown i∗ i∗ : Ak+1C → AkC is the same as capping with c1 of the normal
(line) bundle to the divisor C in M◦ . (Reminder for future use: if i : W ,→ Z is the
closed immersion of W into a vector bundle over W, as the zero section, then the map
i∗ i∗ : A∗ W → A∗ W is capping with the top Chern class of the vector bundle.) In this case
the normal line bundle is trivial: it is the pullback of the normal bundle to t = 0 in P1.
Thus i∗ i∗ = 0. Hence Ak+1M◦ → AkC descends to a map Ak+1(Y × A1) → AkC, and hence
we get a map σ : AkY → AkC, which is what we wanted! The final diagram:
i∗ j∗
Ak+1C / Ak+1M◦ / Ak+1(Y × A1) / 0
KK o O
KKK ∴ oooo
KKK i∗ o ∼
∗
i i∗ =0 KK% ooo
wooo
AkC o ∴σ
AkY.
1.2. Gysin pullback for local complete intersections. We already had defined the Gysin
pullback or Gysin homomorphism in the case where Y is a vector bundle over X: AkY →
Ak−dX. We now extend it to when “Y looks like a vector bundle over X”: when X is a
local complete intersection inside Y. Define the Gysin pullback i∗ : AkY → Ak−d as the
composition
σ s∗N
AkY / AkN / Ak−dX
2
where s∗N is the old Gysin morphism for vector bundles. We’re going to generalize this
further soon!
I showed that the two definitions agree, by observing that the normal cone to a the zero
section of a vector bundle is the vector bundle itself (which is true). Also, we showed
earlier that the Gysin pullback for vector bundles satisfied all sorts of nice properties; if
we show that σ satisfies these nice properties too, then we’ll know it for Gysin pullbacks
to local complete intersections.
Note also: If Y is purely n-dimensional, notice that i∗ [Y] = [X]. Because σ[Y] = [C], and
s∗N[C]
= [X].
I concluded with:
I should probably be a bit clearer about that first map, which might reasonably be called
. (You can see a discussion in Chapter 1 if you want.) Here’s what we need: consider
the map
×
ZpX ⊗ ZqY / Zp+q(X × Y)
defined on varieties by [V] × [W] = [V × W], and defined generally by linearity. (We’ll
take X = Y, but we might as well do this in some generality.) We want this to descend to
the level of Chow classes:
(This is Prop. 1.10 (a) in the book.) Likely exercise: finish this proof.
2. I NTERSECTION P RODUCTS
We’re now ready to discuss the last chapter in the core of the book, on intersection
products. We’ll define the intersection product, and then we’ll verify that it has a host of
properties. This verification will involve lots of diagram-chasing and symbol-pushing, so
I’m going to try to concentrate on helping you keep your eye on the big picture.
What we know so far: proper pushforward, flat pullback, s· of cones, e.g. s· (X, Y), c· .
Gysin pullbacks have gotten more and more complicated: i) X ,→ Y as a divisor. More
3
generally
V
V .
loc. comp. int.
X / Y
Here’s the context in which we’ll work. i : X ,→ Y will be a local complete intersection
of codimension d. Y is arbitrarily horrible. Suppose V is a scheme of pure dimension k,
with a map f : V → Y. Here I am not assuming V is a closed subscheme of Y. Then define
W to be the closed subscheme of V given by pulling back the equations of X in Y:
cl. imm./
W V
g f
cl. imm. /
X Y
(notice definiton of g). We’ll define the intersection product X · V ∈ Ak−dW. (We’ll most
obviously care about the case where V ,→ Y, but you’ll see that this more general case will
be handy too!)
Let’s check that CWV ,→ g∗ NXY: The ideal sheaf I of X in Y generates the ideal sheaf J
of W in V, hence there is a surjection
Algebraic fact (black box from appendix): as V is purely k-dimensional scheme, CWV
is also. Then we may define X · V as I said we would: X · V = s∗ CW/V .
Proposition. If ξ is the universal quotient bundle of rank d on P(g∗ NX/Y ⊕ 1), and
q : P(g∗ NX/Y ⊕ 1) → W is the projection, then
X · V = q∗ (cd(ξ) ∩ [P(CW/V ⊕ 1)]).
4
Proof. Let C = CW/V .
cl. imm. /
C N
E
open imm. open imm.
` cl. imm. / `
C PC N PN
= s =
cl. imm./
P(C ⊕ 1) P(N ⊕ 1)
ppp
q
ppppp
wpppp
W
We want to take the cone C and intersect it with the zero section s of the vector bundle
N (the top row of this diagram). We we can do this on the second row of the diagram.
Recall that we proved: if β ∈ AkN and β ∈ Ak(P(N ⊕ 1)) which restricts to β. Then
s∗ β = q∗ (cr(ξ) ∩ β) where ξ is the universal (rank r) quotient bundle of q∗ (N ⊕ 1). Then
we’re done.
Proposition. X · V = {c(g∗NX/Y) ∩ s(W, V)}k−d. (Here {·}k−d means “take the dimension
k − d piece of ·.)
Proof omitted.
X /Y
As before, the normal cone C 0 = CX0 Y 0 is a closed subcone of g∗ NXY. Define the refined
Gysin homomorphism i! (pronounced i shriek, which is what people sometimes do when
they first hear about this) as the composition:
σ s∗
AkY 0 / AkC 0 / AkN / Ak−dX 0 .
Note what we can now do: we used to be able to intersect with a local complete intersec-
tion of codimension d. Now we can intersect in a more general setting.
We’ll next show that these homomorphisms behave well with respect to everything
we’ve done before.
E-mail address: [email protected]
6
INTERSECTION THEORY CLASS 17
RAVI VAKIL
C ONTENTS
1. Where we are 1
!
1.1. Refined Gysin homomorphisms i 2
1.2. Excess intersection formula 4
2. Local complete intersection morphisms 6
1. W HERE WE ARE
We defined the Gysin pullback i! and a rather general intersection product. Let i : X ,→
Y be a local complete intersection of codimension d. Y is arbitrarily horrible. Suppose V
is a scheme of pure dimension k, with a map f : V → Y. Here I am not assuming V is a
closed subscheme of Y. Then define W to be the closed subscheme of V given by pulling
back the equations of X in Y:
cl. imm.//
W V
g f
cl. imm. //
X Y
(notice definition of g).
1
Proposition. If ξ is the universal quotient bundle of rank d on P(g∗ NX/Y ⊕ 1), and
q : P(g∗ NX/Y ⊕ 1) → W is the projection, then
X · V = q∗ (cd(ξ) ∩ [P(CW/V ⊕ 1)]).
and
Proposition. X · V = {c(g∗NX/Y) ∩ s(W, V)}k−d. (Here {·}k−d means “take the dimension
k − d piece of ·.)
X0 // Y0
X // Y
As before, C 0 = CX0 Y 0 ,→ g∗ NXY. Define the refined Gysin homomorphism i! as the compo-
sition:
σ s∗
AkY 0 // AkC 0 // AkN // Ak−dX 0 .
Note what we can now do: we used to be able to intersect with a local complete intersec-
tion of codimension d. Now we can intersect in a more general setting.
We’ll next show that these homomorphisms behave well with respect to everything
we’ve done before. These are all important, but similar to what we’ve done before, so I’ll
state the various results. I’ll just sporadically give proofs.
X0 ∩ V // V
h
X0 // Y0
g
X // Y
Then
i![V] = c(g∗ NX/Y) ∩ h∗ s(X 0 ∩ V, V).
Reason we like this: we already know Chern and Segre classes behave well. So we can
reduce calculations about i! to things we’ve already proved. Reason for fact: Calculate
2
X · V using
X0 ∩ V // V
g◦h
X // Y
using the projection formula. We now have to show that this really gives i![V]. (Fulton
uses this second version as the original definition.) Omitted.
Refined Gysin commutes with proper pushforward and proper pullback. Consider the
fiber diagram
i00 //
X 00 Y 00
q p
i0
X0 // Y0
g f
i
X // Y
(a) If p is proper and α ∈ AkY 00 , then i!p∗ (α) = q∗ (i!α) in Ak−dX 0 . (Caution: i! means
two different things here!)
(b) If p is flat of relative dimension n, and α ∈ AkY 0 , then i!p∗ (α) = q∗ (i!α) in
Ak+n−dX 00 .
It suffices to verify that g∗ NXY =∼ NX0 Y 0 . Reason: If I and I 0 are the respective ideal
sheaves, the canonical epimorphism g∗ (I/I 2) → I 0 /(I 0 )2 must be an isomorphism. (De-
tails omitted. X is locally cut out in Y by d equations. X 0 is cut out in Y 0 by (the pullbacks
of) the same d equations.)
3
1.2. Excess intersection formula. Consider the same fiber diagram as before
i00 //
X 00 Y 00
q p
i0
X0 // Y0
g f
i0
X // Y
where now i is still a locally closed intersection of codimension d, and i 0 is also a locally
closed intersection, of possibly different dimension d 0 . Let N and N 0 be the two normal
bundles; as before we have a canonical closed immersion N 0 ,→ g∗ N. Let E = g∗ N/N 0 be
the quotient vector bundle, of rank d = d − d 0 .
For any α ∈ AkY 00 , note that i!(α) and (i 0)!(α) differ in dimension by e. What is their
relationship? Answer:
Excess intersection formula. For any α ∈ AkY 00 , i!(α) = ce(q∗ E) ∩ (i 0)!(α) in Ak−dX 00 .
Immediate corollary. Specialize to the case where the top row is the same as the middle
row, and i 0 is an isomorphism:
∼ //
X0 i0
Y0
g
i
X // Y
Then i!α = cd(g∗ N) ∩ α. Specialize again to X 0 = Y 0 = X to get the self-intersection
formula: i∗ i∗ α = cd(N) ∩ α.
X // Y
i
a fiber square, and F a vector bundle on Y 0 . Then for all α ∈ AkY 0 and all m ≥ 0,
i!(cm(F) ∩ α) = cm(i 0 F) ∩ i!α
∗
in Ak−d−m(X 0 )
Proof omitted.
Proof (long!) omitted. Idea: by blowing up to reduce to the case of divisors, as we did
when we showed that the intersection of two divisors was independent of the order of
intersection, long ago.
Functoriality.
The refined Gysin homomorphisms for a composite of locally closed intersections is the
composite of the refined Gysin homomorphisms of the factors.
(a) Assume that i is a locally closed intersection of codimension d, and that p and pi
are flat of relative dimensions n and n − d. Then i 0 is a locally closed intersection
of codimension d, p 0 and p 0i 0 are flat, and for α ∈ AkZ 0 ,
(p 0 i 0 )∗ α = i!p 0 α
∗
in Ak+n−dX 0 .
(b) Assume that i is a locally closed intersection of codimension d, p is smooth of
relative dimension n, and pi is locally closed intersection of codimension d − n.
Then for all α ∈ AkZ 0 ,
(pi)!α = i!(p 0 α)
∗
in Ak+n−dX 0 .
5
Short proof, omitted.
i1 p1
Because (a) seems surprising, and the roof is short, I’ll give it to you. If X // P1 // Y
is another factorization of f, compare them both with the diagonal:
;; P1 ??
vvv ?? p1
vv ??
vvv ??
(i,i1 ) v
X // P ×Y PH1 // Y.
??
HH
HH ~~~
HH ~
HH ~~ p
H$$ ~~
P
6
Use the second functoriality proposition (b).
7
INTERSECTION THEORY CLASS 18
RAVI VAKIL
C ONTENTS
1. Last day 1
2. Towards Grothendieck-Riemann-Roch 2
2.1. The Chern character and the Todd class 2
2.2. The Grothendieck groups K0X and K0X 3
3. Statement of the theorem 4
3.1. Why you should care 4
4. Toward a proof 6
5. Grothendieck-Riemann-Roch for Pn → pt 6
1. L AST DAY
We defined the Gysin pullback i! in a rather general circumstance. I have only a few
additional comments to make. Recall that a morphism f : X → Y is a local complete inter-
section morphism if f can be factored as a local complete intersection followed by a smooth
morphism.
I don’t know why one wouldn’t more generally think of factorizations into a local com-
plete intersection followed by a flat morphism.
I gave you a few examples as to why you might care about such morphisms. Here is
another. If X and Y are smooth then any morphism between them is an lci morphism.
Reason: factor it into
X ,→ X × Y → Y.
1
2. TOWARDS G ROTHENDIECK -R IEMANN -R OCH
I’m going to first explain the terminology behind the statement, then give the statement.
I will then give some examples to show you that the statement is in fact very powerful.
And finally, I hope to sketch a proof in an important special case,
2.1. The Chern character and the Todd class. Suppose E is a rank n vector bundle. Let
α1, . . . , αn be the Chern roots of the vector bundle, so α1 + · · · + αn = c1(E), etc. Define
r
X
ch(E) = exp(αi)
i=1
For comparison, the Chern polynomial is multiplicative in exact sequences; the Chern
character is additive.
Exercise. For tensor products of vector bundles, ch(E ⊗ E 0 ) = ch(E) · ch(E 0 ). I don’t
think
P this is true for coherent sheaves in general, but haven’t checked. (I would expect
i 0 0
i≥0 ch(Tor (E, E )) = ch(E) · ch(E ).)
where
X ∞
x 1 Bk 2k
Q(x) = −x
= 1 + x + (−1)k−1 x .
1−e 2 k=1
(2k)!
The first few terms are
1 1 1
td(E) = 1 + c1 + (c21 + c2) + c1c2
2 12 24
1
−c41 + 4c12c2 + 3c22 + c1c3 − c4 + · · ·
+
720
0 00
If 0 → E → E → E → 0 is exact, then
td(E) = td(E 0 ) td(E 00 ).
Like the Chern polynomial, it is multiplicative in exact sequences.
2
2.2. The Grothendieck groups K0X and K0X. If you went to Dan Ramras’ K-theory talks,
you will have seen these.
The Grothendieck group of vector bundles K0X on X is the group generated by vector bun-
dles, modulo the relations on exact sequences [E] = [E 0 ] + [E 00 ]. Vector bundles pull back
to vector bundles, and exact sequences of vector bundles pull back to exact sequences of
vector bundles, so if a morphism f : X → Y induces a homomorphism f∗ : K0X → K0Y.
However, vector bundles seldom pushforward to vector bundles.
The Grothendieck group of coherent sheaves K0X on X is the group generated by coherent
sheaves, modulo the same relations on exact sequences. Bad news: coherent sheaves pull
back to coherent sheaves, but exact sequences of coherent sheaves don’t pull back to exact
sequences of coherent sheaves. So we don’t have a pullback map f∗ : K0X → K0Y. Good
news: we can make sense of pushforwards; if f : X → Y is a proper morphism, then
coherent sheaves pushforward to coherent sheaves (see Hartshorne). Bad news: exact
sequences don’t push forward to exact sequences: If
0 → F 0 → F → F 00 → 0
is an exact sequence on X, then we only get left exactness of pushforwards:
0 → f∗ F 0 → f∗ F → f∗ F 00 .
Good news: we can extend this to a long exact sequence:
(See Hartshorne for more on these “higher direct image sheaves. They can be defined
as follows: Rif∗ F is the sheaf associated to the presheaf U → Hi(f−1(U), F ).
From now on, X will be smooth, so K0X = K0X, so I’ll just call this group K(X).
(I should point out where all the intersections take place, and where the pushforwards
take place!)
This can be generalized further to singular schemes, but this is enough generality for
now.
3.1. Why you should care. Before we get into proving it, let me first try to convince you
how powerful it is. I’ll first show that it gives you old-fashioned Riemann-Roch. (I won’t
try to convince you why you should care about Riemann-Roch for curves — that is a
whole lecture in itself, or more!)
Let’s apply this to Y a point, X a smooth curve, and α a line bundle L. Then we get
h0(X, L) − h1(X, L) = · · ·
On the right side, we have
1 1 1
f∗ ((1 + c1(L))(1 + c1(T )) = f∗ (1 + c1(L) − c1(K)) = deg(c1(L) − c1(K)).
2 2 2
Recall that c1(K) = −c1(T ) = 2g − 2. Thus the right side is d − g + 1.
4
That’s the baby-est case. Let’s make things more interesting. We’ll keep Y a point and
X a nonsingular curve, and now α is the class of a vector bundle V of rank r. Then we get
1 r
h0(X, V) − h1(X, V) = f∗ ((r + c1(V))(1 + c1(T )) = f∗ (c1(V) + c1(T )) = d + r(1 − g).
2 2
Let’s generalize further; now V is a coherent sheaf, of “rank” r (rank at the generic
point). The same formula holds!
Next let’s go to the case where X is now a smooth surface, Y a point, and to keep things
calm, let’s make α the class of a line bundle L. Then the left side is
which is Riemann-Roch for surfaces, which you can read about in Hartshorne chapter V.
More generally still, if X is a smooth surface, and E is a vector bundle, and Y is still a
point, we get
Z
χ(X, E) = ch(E) · td(TX).
X
We have reproved the Hirzebruch-Riemann-Roch theorem. And this also works for coher-
ent sheaves.
What about if Y is not a point? I’ll describe why you care somewhat philosophically.
Suppose you have a nice morphism X → Y, interpreted as “nice family” (say of smooth
surfaces). Say you have a vector bundle on the family. On each of the elements of the
family (the fibers of the morphism), you have a vector bundle; let’s say to make things nice
that for every element of the family, this vector bundle has vanishing higher cohomology.
Then h0(V) is constant, as h0(V) = χ(V), and χ(V) is constant on connected families.
Thus for each point of the base Y, you have a vector space of some rank h0(V). You
should expect this to glue together into a vector bundle, and indeed it does: f∗ V. (Again,
to make this precise requires Hartshorne chapter III or its equivalent.) Which vector bundle
do you get? For example, what are its Chern classes? Grothendieck-Riemann-Roch will
answer this for you!
So let me emphasize: you’re going to see a proof of GRR that will not be too bad; and as
a special case you’ll get old-fashioned Riemann-Roch for curves. I think the difficulty of
this proof is comparable to the difficulty of building up the machinery behind the “usual”
proof of Riemann-Roch in the algebraic category; so you may as well get a much more
powerful result for the same price.
5
4. TOWARD A PROOF
I’ll prove this in the case where X → Y factors through X ,→ Pn × Y → Y where the
first is a closed immersion. (This is a projective morphism in the sense of Hartshorne,
and a special case of a projective morphism according to other people, such as EGA. I
don’t want to get into this.) This isn’t such an outrageous assumption; for example, if X
is projective, then X ,→ Pn, and then X ,→ Pn × Y.
f g
Lemma. Given X / Z / Y . Suppose GRR holds for f and g. Then it holds for g ◦ f.
Proof. This has been cooked up to be easy! (That Grothendieck is quite a tricky guy!)
The pushforward of ch(α) td(TX) by f is ch(f∗ α) td(TZ), by GRR for f. The pushforward
of this in turn is ch(g∗ f∗ α) td(TY), by GRR for g. But then we have GRR for g ◦ f: (g ◦
f)∗ (ch(α) td(TX)) = ch(g∗ f∗ α) td(TY).
So our strategy is clear. We’re going to prove GRR for closed immersions X ,→ Y, and
we’ll prove it for Pn × Y → Y.
First we show:
Lemma. K0(Pm) is generated by the classes of line bundles [O(n)], without any restriction
on n.
Proof. I will need some machinery we have not developed. How much extra you will
need to consider as a “black box” will depend on how much you already know. Let F be
any coherent sheaf. Our goal is to get a resolution of F by direct sums of line bundles:
By an earlier statement, we need only show that for any coherent sheaf F , we can find a
surjection ⊕ji=1O(n) → F , because then we can iterate this, and at some point we will get
a 0.
Here’s how this implies the theorem. Twisting this by O(N) we get:
m+1
0 → O(N) → ⊕m+1O(N+1) → ⊕( 2 ) O(N+2) → · · ·⊕m+1O(N+m) → O(N+m+1) → 0.
This expresses [O(N + m)] in terms of the classes of the m + 1 smaller line bundles. Simi-
larly, it expresses [O(N)] in terms of the classes of the m + 1 larger line bundles. Thus by
using this repeatedly, any line bundle can be expressed in terms of O, O(1), . . . , O(m).
Aside: you also get some interesting algebra out of this. Apply the Chern polynomial
to this exact sequence. You get
m+1
Y
(1 + iH)(−1) ( ) ≡ 1 (mod Hm+1)
i m+1
i
i=0
Theorem. GRR is true for Pm → pt for the line bundles O(n) (0 ≤ n ≤ m). Hence GRR is
true for Pm → pt.
Proof. Now p∗ [O(n)] = χ(Pm, O(n)). Now we can compute the cohomology groups
of O(n) by hand, and we find that hi(O(n)) = 0 for n ≥ 0. Thus χ(Pm, O(n)) =
h0(Pm, O(n)). And this corresponds to the vector space of degree n polynomials with
m + 1 variables. This turns out to be n+m
m
.
Next, we’ll show that knowing the result for Pm → pt will imply the result for Pm ×Y →
Y.
E-mail address: [email protected]
8
INTERSECTION THEORY CLASS 19
RAVI VAKIL
C ONTENTS
Today I’m going to try to finish the proof of Grothendieck-Riemann-Roch in the case
of projective morphisms from smooth varieties to smooth varieties. We’ll see that we’re
essentially going to prove it more generally for projective lci morphisms.
Recall the definition of the Chern character and Todd class. Suppose F is a coherent
sheaf. Let α1, . . . , αnPbe the Chern roots of the vector bundle, so α1 + · · · + αn = c1(F ),
etc. Define ch(F ) = ri=1 exp(αi) This is additive on exact sequences. For vector bundles,
we have ch(E ⊗ E 0 ) = ch(E) · ch(E 0 ).
Q
The Todd class td(E) of a vector bundle is defined by td(E) = ri=1 Q(αi) where
X∞
x 1 Bk 2k
Q(x) = −x
=1+ x+ (−1)k−1 x .
1−e 2 k=1
(2k)!
It is multiplicative in exact sequences.
We defined the Grothendieck groups K0X and K0X. They are vector bundles, respec-
tively coherent sheaves, modulo the relation [E] = [E 0 ] + [E 00 ]. We have a pullback on
K0: f∗ : P
K0X → K0Y. K0X is a ring: [E] · [F] = [E ⊗ F]. We have a pushforward on K0:
f∗ [F ] = i≥0(−1)i[Rif∗ F ].
1
The Chern character map descends to K(X): ch : K(X) → A(X)Q . This does not commute
with proper pushforward; Grothendieck-Riemann-Roch explains how to fix this.
1.1. New facts. Here are some useful facts, that I didn’t mention last time. We have
an excision exact sequence for K0: If Z ,→ X is a closed immersion, and U is the open
complement, we have an excision exact sequence
The proof is similar to our proof for Chow; this is Hartshorne Exercise II.6.10(c).
∼ K(Y).
Similarly, we have K0(A1 × Y) =
Last time I showed: Lemma. The group K0(Pm) is generated by the classes [OPm (n)],
with 0 ≤ n ≤ m.
I’d like to do it differently today. Instead, I’ll show it is generated by the classes
[OPm (−n)], with 0 ≤ n ≤ m.
` ` `
Using the excision exact sequence for K-theory, and Pm = A0 A1 · · · Am, we get
inductively: K0(Pm) is generated by n + 1 things: [OP0 ], [OP1 ], . . . , [OPm ].
Important philosophy behind Riemann-Roch: K(Pm) and A∗ (Pm) are both m-dimensional
vector spaces; Chern character provides an isomorphism between them. Multiplying by
the Todd class provides a “better” isomorphism between them.
Interesting exercise: how do you make sense of this when X and Y are singular? For
example, what if X → Y is a smooth morphism, we get ch(f∗ α)· = f∗ (ch(α) · td(TX/Y))
where X/Y is the relative tangent bundle. As another example, what if X → Y is a com-
plete intersection? Then TX and TY don’t make sense, but NX/Y is a vector bundle, and
then ch(f∗ α) · td(NX/Y) = f∗ (ch(α)). Combining these two, you can now make sense of
GRR in the case when f is an lci morphism (i.e. closed immersion followed by a smooth
morphism).
Hence the strategy is now to show GRR for Y × Pm → Y, and for closed immersions.
Proof. We showed last time that this is true in the case where Y is a point. Consider the
following diagram.
τY ⊗τPm
K(Y) ⊗ K(Pm) / A(Y)Q ⊗ A(Pm)Q
× ×
τY ×τPm
K(Y × Pm) / A(Y × Pm)Q
f∗ f∗
τY
K(Y) / A(Y)Q
(I won’t be using anything special about Pm now.) We want to show that the bottom
square commutes.
Note that the top square commutes. Reason: TY×Pm = p∗1TY ⊕ p∗2TPm (where p1 and p2
are the projections) from which td(TY×Pm ) = td(p∗1TY) × td(p∗2TPm ).
So it suffices to show that the big rectangle commutes. But it does because we’ve al-
ready shown that GRR holds for Pm → pt.
3
2.1. GRR for a special case of closed immersions f : X → Y = P(N ⊕ 1). Suppose f is a
closed immersion into a projective completion of a normal bundle. Let d = rank N. We
want to prove GRR for a vector bundle E. As the vector bundles generate K(X), this will
suffice.
This example comes the closest to telling me why the Todd class wants to be what it is.
Let p : Y = P(N ⊕ 1) → X be the projection. Let OY(−1) be the tautological line bundle
on Y = P(N ⊕ 1). Then as in previous lectures we have a tautological exact sequence of
vector bundles on Y:
0 → OY(−1) → p∗ (N ⊕ 1) → Q → 0
where Q is the universal quotient bundle. (Recall that f∗ Q = NX/Y.) Here is something
you have to think through, although we’ve implicitly used it before. We have a natural
section of p∗ (Q ⊕ 1), the 1. This gives a section s of Q. This section vanishes precisely
(scheme-theoretically) along X. In particular, for any α ∈ A(Y), f∗ (f∗ α) = cd(Q) · α . (This
was one of our results about the top Chern class. f∗ f∗ will knock the degree down by d,
and we found that this operator was the same as capping with the top Chern class.)
s∨
(1) 0 / ∧dQ∨ / ··· / ∧2Q∨ / Q∨ / OY / f∗ OX / 0.
Proof. Rather than proving this precisely, I’ll do a special case, to get across the main
idea. This in fact becomes a proof, once the “naturality” of my argument is established.
Suppose Y = Spec k[x1, . . . , xn], so OY = k[x1, . . . , xn] (a bit sloppily) and X = ~0 ⊂ Y. Then
let’s build a resolution of OX. We start with
OY → OX → 0.
We have a big kernel obviously: the ideal sheaf of OX. So our next step is:
OYx1 ⊕ OY x2 · · · ⊕ OYxn → OY → OX → 0.
We still have a kernel; (−x2x1, x1x2, 0, · · · , 0) is in the kernel, for example. So our next step
is:
OYx1x2 ⊕ · · · ⊕ OY xn−1xn →
(We need to check that we’ve surjected onto the kernel! But that’s not hard; you can try
to prove that yourself.) And the pattern continues. We get:
0 → OYx1 · · · xn → ⊕n ^i · · · xn → · · · → ⊕n
i=1OY x1 · · · x i=1OY xi → OY → OX → 0.
Therefore
d
X
ch f∗ [E] = (−1)p ch(∧pQ∨) · ch(p∗ E).
p=0
Lemma.
d
X
(−1)p ch(∧pQ∨) = cd(Q) · td(Q)−1.
p=0
Proof. This is remarkably easy. Let α1, . . . , αd be the ChernProots of Q. Then the Chern
P P − αi ···αi
roots of ∧pQ∨ are − αi1 · · · αip . Hence ch(∧pQ∨) = e 1 p from which
d
X d
X X
(−1)p ch(∧pQ∨) = (−1)p e−αi1 · · · e−αip
p=0 p=0
d
Y
= (1 − e−αi )
i=1
d
Y 1 − e−αi
= (α1 · · · αd)
αi
i=1
= cd(Q) · · · td(Q)−1.
Hence
ch f∗ [E] = cd(Q) td(Q)−1 · ch(p∗ E)
= f∗ (f∗ td(Q)−1 · f∗ ch(p∗E)) (using cd(Q) ∩ β = f∗ (f∗ β), see 1st par of Section 2)
= f∗ (td(NX/Y)−1 ch(E)) (using f∗ Q = NX/Y, f∗ p∗ E = E)
= f∗ (td(TX)f∗ td(TY)−1 ch(E))
= td(TY)f∗ (td(TX) ch(E)) (projection formula)
as desired!
This ends the proof of GRR for a closed immersion of X into the projective completion
of a normal bundle.
5
2.2. GRR for closed immersions in general. Suppose f : X → Y is a closed immersion.
We’ll prove GRR in this case; again, we need only to consider a generator of K(X), a vector
bundle E on X.
{0} / P 1 o {∞}
Define F (above), p : X × P1 → X. Resolve p∗ E on M:
0 → Gn → Gn−1 → · · · → G0 → F∗ (p∗ E) → 0.
Both X × P1 and M are flat over P1 (recall that dominant morphisms from irreducible
varieties to a smooth curve are always flat), so restriction of these exact sequences to
the fibers M0 and M∞ (also known as tensoring with the structure sheaves of the fibers)
preserves exactness.
∼ M0 ,→ M, j∞ : BlX Y ∪ P(N ⊕ 1) = M∞ ,→ M, k : P(N ⊕ 1) ,→ M,
Let j0 : Y =
l : BlX Y ,→ M.
6
INTERSECTION THEORY CLASSES 20 AND 21: BIVARIANT
INTERSECTION THEORY
RAVI VAKIL
C ONTENTS
In this final week of class, I’ll describe bivariant intersection theory, covering much of
Chapter 20. Again, you should notice that given chapters 1 through 6, we can comfortably
jump into chapter 20.
X / Y.
f
Recall that the final fundamental intersection construction we came up with was the
following. Suppose f is a local complete intersection of codimension d (or more generally
a local complete intersection morphism). Then we defined
f! : AkY 0 → Ak−dX 0
1
for all Y 0 → Y. These Gysin pullbacks were well-behaved in all ways, and in particular
compatible with proper pushforward, flat pullback, and intersection products.
An earlier example was that of a flat pullback; if f if flat of relative dimension n, then
f 0 is too, and we got f∗ : AkY 0 → Ak−nX 0 , which again behaves well with respect to
everything else.
We’ll now generalize this notion. Define a bivariant class for any f (not just lci) as fol-
lows. It is a collection of homomorphisms AkY 0 → Ak−pX 0 for all Y 0 → Y, all k, again
compatible with pushforward, pullback, and intersection products. We’ll call the group
of such things Ap(f : X → Y).
We’ll see that the group A−k(X → pt) will be (canonically) isomorphic to AkX . We’ll
see that Ak(id : X → X) is a ring, which Fulton calls the cohomology group; I might call the
resulting ring the Chow ring. We’ll denote this by AkX .
The ring structure is a product of the form ApX ⊗ AqX → Ap+qX. We’ll define more
generally
Ap(f : X → Y) ⊗ Aq(g : Y → Z) → Ap+q(g ◦ f : X → Z) .
We’ll prove Poincare duality when X is smooth: A∗ X =∼ A∗ X (as rings — recall we defined
a ring structure on the latter). We’ll define proper pushforward and pullback operations
for bivariant groups. Basically, they’ll behave the way you’d expect from homology and
cohomology. This will give a cap product A∗ X × A∗ X → A∗ X. Alarming fact: This ring is
apparently not known to be commutative in general, because the argument requires reso-
lution of singularities. (It is known to be commutative in characteristic 0, and for smooth
things in positive characteristic, and a few more things.) I think it should be possible to
show that the ring is commutative in general using technology not available when this
theory was first developed, using Johan de Jong’s “alteration theorem” in positive char-
acteristic. If you would like to patch this hole, then come talk to me.
Okay, let’s get started. Today I’ll outline the results, and prove a few things; Wednesday
I’ll prove some more things.
2. P RECISE STATEMENTS
2
for all g : Y 0 → Y, and all k, compatible with proper pushforwards, flat pullbacks, and in-
tersection products. I’ll make that precise in a moment, by stating 3 conditions explicitly.
But first I want to show you that you’ve seen this before in several circumstances.
So we’re creating a machine that in some sense incorporates most things we’ve done so
far.
f00
(1) X 00 / Y 00 ,
h0 h
X0 / Y0
f0
g0 g
X / Y
f
in Ak−pX 0 .
(k+n)
(h∗ α) = h 0 c(k)
∗
cgh g α
in Ak+n−pX 00 .
3
(C3): If g : Y 0 → Y, h : Y 0 → Z 0 are morphisms, and i : Z 00 → Z 0 is a local complete
intersection of codimension e, and one forms the diagram
(2) X 00 / Y 00 / Z 00
f00 h0
i00 i0 i
X0 / Y0 / Z0
f0 h
g0 g
X / Y
f
2.1. Basic operations and properties. Here are some basic operations on bivariant Chow
groups A∗ (X → Y).
X / Y / Z.
f g
(P3): Pullback (not necessarily flat!!): Given f : X → Y, g : Y1 → Y, form the fiber square
f1
(4) X1 / Y1
X / Y.
f
Here are seven more axioms, which can also be easily verified.
(Apr) Associativity of products. If c ∈ A(X → Y), d ∈ A(Y → Z), e ∈ A(Z → W), then
(c · d) · e = c · (d · e) ∈ A(X → W).
(Aprpb) Product and pullback commute. If c ∈ A(f : X → Y), d ∈ A(Y → Z), and g : Z1 → Z
is a morphism, form the fiber diagram
f0
X1 / Y1 / Z1
g0 g
X / Y / Z.
f
Then
∗
g∗ (c · d) = g 0 (c) ·∗ (d) ∈ A(X1 → Z1).
5
3. P ROVING THINGS
φ : A−p(X → S) → Ap(X)
given by c 7→ c([S]).
Proof. We will define the inverse morphism. Given a ∈ Ap(X), define a bivariant class
ψ(a) ∈ A−p(X → S) as follows: for any Y → S, and any α ∈ AkY, define
(Here a × α is the exterior product.) Since exterior products are compatible with proper
pushforward, flat pullback, and intersections, ψ(a) is a bivariant class.
X ×S V / V
cl. imm.
X ×S Y / Y
X / S
3.1. The Chow (“cohomology”) ring. Define ApX := Ap(id : X → X). We have a cup
product. We also have an element 1 ∈ A0X, which acts as the identity. We have a cap
product
∩ : ApX × AqX → Aq−pX
determined by
Ap(X → X) × A−q(X → S) → A−(q−p)(X → S)
which makes A∗ X into a left A∗ X-module. All of this follows formally from our axioms.
6
4. T HINGS YOU MIGHT WANT TO BE TRUE
4.1. Poincare duality. Theorem. Let Y be a smooth, purely n-dimensional scheme (va-
riety) — not necessarily proper (compact).
(a) The canonical homomorphism ∩[Y] : ApY → An−pY is an isomorphism.
(b) The ring structure on A∗ Y is compatible with that defined on A∗ Y earlier. More gen-
erally, if f : X → Y is a morphism, β ∈ A∗ Y, α ∈ A∗ X, then the class f∗ (β) ∩ α ∈ A∗ X
coincides with that constructed earlier.
is an isomorphism.
f /
X Y
γ δ
q
f0
X ×Z Y / Y ×Z Y / Y
p0 p g
X / Y / Z
f g
where δ is the diagonal map, and p and q are the first and second projections. Here γ is
the “graph” of the morphism X → Y over Z. Define
L : Ap−n(gf : X → Z) → AP(X → Y)
by L(c) = [γ] · g∗ (c). Notice that γ and δ is a local complete intersection of codimension
n, with f 0 ∗ [δ] = [γ]. (This requires a check in the case of γ.)
Notice that p 0 ◦ γ : X → X and q ◦ δ : Y → Y are both the identity morphisms (on X and
Y respectively).
7
Let’s verify that L and “multiplication by [g]” are inverse homomorphisms. (This is
easier to understand if you see someone pointing at diagrams!) First,
L(c) · [g] = [γ] · (g∗[c] · [g]) (axiom (Apr))
= [γ] · [p 0] · c (axiom (C2))
= [p 0 ◦ γ] · c = 1 · c = c (axiom (Apr))
Second,
∗
L(c · [g]) = f 0 [δ] · p∗ (c) · g∗ [g] (axioms (Aprpb), (Apr))
= (p ◦ δ)∗ (c) · [δ][q] (axiom (C2))
= c · [δ ◦ q] = c · 1 = c (axiom (Apr)).
4.2. Chern classes commute with all bivariant classes. Put another way, any operation
which commutes with proper pushforward, pullback, and intersections, automatically
commutes with Chern classes. Precisely:
Proof. Recall our definition of Chern classes. They are certain polynomials in Segre classes.
Segre classes are defined using operations of the form α 7→ p∗ (c1(O(1))i ∩ p∗ α), and since
c commutes with p∗ and p∗ , we just have to show that c commutes with c1(L)∩, where
L is a line bundle on Y 0 . We may assume α = [V]. Because c commutes with proper
pushforward, we may assume V = Y 0 . Let L = O(D), D a Cartier divisor on V.
4.3. Bivariant classes vanish in dimensions that you’d expect them to. Proposition. Let
f : X → Y be a morphism. Let m = dim Y, and let n be the largest dimension of any fiber
f−1y, y ∈ Y.
Ap(f : X → Y) = 0 if p < −n or p > m.
(Think about why this is what you’d expect!) In particular, for any X, ApX = 0 unless
0 ≤ p ≤ dim X.
(Proof omitted.)
E-mail address: [email protected]