Mit 726

MIT OpenCourseWare
http://ocw.mit.edu
18.726 Algebraic Geometry

Spring 2009
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

Introduction
In this lecture, Ill give a bit of an overview of what we will be doing this semester, and
in particular how it will dier from 18.725. We will start in earnest (with the rudiments of
category theory) in the next lecture.
Where we were, and where we need to go
In 18.725, we studied the notion of an abstract algebraic variety over an algebraically closed
eld. This combines a lot of the commutative algebra developed in the early 20th century
(largely to explain the geometric reasoning of the masters of the Italian school) with Weils
fundamental idea to glue ane algebraic varieties in the same way that one glues local charts
together to build manifolds. So whats left?
We would like to deal with phenomena of nonreducedness, for instance as it emerges
under degenerations. One of the key ideas of the Italian school for understanding
things like the geometry of the moduli space of curves was to notice that if you have
a family of algebro-geometric objects dened in terms of a parameter t, then the
behavior of a particular member of the family is sometimes much simpler than that of
a general member. For instance, for a general t, the elliptic curve y 2 = x3 + tx + t does
not have a rational parametrization, but it does in the special case t = 0. One can
often understand something about the general member of the family by rst analyzing
a special member, then guring out how the information you are looking for gets
transmitted back to the general member via the degeneration.
However, degenerations of algebraic varieties are not always best viewed as algebraic
varieties. For example, if t = 0, then the homogeneous polynomial y 2 tx2 in x, y, z
(over, say, the complex numbers) denes a pair of lines. The degeneration at t = 0,
however, is the single line y = 0, because the equations y 2 = 0 and y = 0 dene the
same variety. In order for the degeneration to preserve the degree of the curve, we need
to remember that it is y 2 rather than y which denes this line. That is, the function
y on the variety should be nilpotent, a possibility that is not aorded by the category
of algebraic varieties.
We would like to work over elds which are not algebraically closed. The restriction to
algebraically closed elds was originally needed to make things like Bezouts theorem
work. However, at the end of the day, we are sometimes interested in solving polyno
mials over non-algebraically closed elds. For instance, the elliptic curves y 2 = x3 + tx
dened by rational numbers t are all isomorphic as algebraic varieties over the complex
numbers. However, they have rather dierent arithmetic behaviors; for instance, the
curve for t = 1 has only nitely many rational points, whereas the curve for t = 73 has
innitely many.
1
Weil had an answer for this point: he suggested embedding ones given base eld in
a large algebraically closed eld, called a universal domain. However, Weils answer
looks like a mistake in hindsight, because it is not suciently functorial ; see below.
We would also like to work over (commutative, unital) rings, not just elds. For
instance, already in Weils work the question of reduction mod p arises, but cannot be
addressed while working over elds.
Even in the context of varieties, one often wants to work over a base which is not a
eld. For instance, the theory of elliptic surfaces is largely thought of by viewing these
surfaces as (relative) elliptic curves over a base curve.
Theres more, but enough for now.
Paradigm shift 1: sheaves
At the time, one might have expected that the future development of algebraic geometry
would proceed as a natural descent from Weils 1946 Foundations, with more bells and
whistles attached to extend generality. However, just as the theory of epicycles to explain the
motion of planets was thrown into disrepute by two paradigm shifts (Galileos heliocentricity
and Keplers elliptic orbits), two paradigm shifts rendered Weils foundations a dead end in
the development of algebraic geometry. (Most material written in that language has since
appeared in modern terminology; what remains untranslated is as intelligible to the modern
reader as Chaucers Middle English.)
The rst of these shifts can be attributed to Serre, who introduced the notion of sheaves
into algebraic geometry. These are the sort of objects dened by descriptions like take all
continuous functions on all open subsets of a topological space, or take all dierentiable
functions on all open subsets of a smooth manifold. The latter example is particularly
helpful to keep in mind: it is possible to have two dierent smooth manifolds which are
isomorphic as topological spaces (e.g., to R4 , or to a seven-dimensional sphere), but not as
smooth manifolds. That is, the underlying topological space does not carry enough infor
mation to detect nonisomorphism of smooth manifolds. However, the sheaf of dierentiable
functions does carry enough information.
Sheaves were originally introduced by Cartan in order to simplify and extend the theory
of complex analytic geometry. It is Serre who recognized their place in modern algebraic
geometry, by observing (among other things) that they give you a natural way to add nilpo
tents. In my example of the lines y = 0 versus y 2 = 0 in the (x, y)-plane, it will turn out
that (in the category of schemes) the underlying sets of these two objects will be the same,
but the sheaves of regular functions will dier.
However, it will take us some time before we can relate sheaves to algebraic geometry.
We will rst have to take some time to discuss topological spaces equipped with rings of
interesting functions, giving rise to the notion of a locally ringed space. This notion includes
many familiar things: topological spaces, topological manifolds, smooth manifolds, and even
abstract algebraic varieties.
2
But what we really want to include into this category is the prime spectrum of an arbitrary
(commutative) ring. Recall that over an algebraically closed eld, by the Nullstellensatz
there is a bijection between the points of an ane algebraic variety and the maximal ideals
of its ring of regular functions. For a general ring, Zariski suggested to instead look at the
set of prime ideals, i.e., the prime spectrum of the ring; that way, any map of rings would
correspond to a map (contraction) on prime ideals in the opposite direction.
The fundamental theorem of schemes is that this set carries the natural structure of
a sheaf of rings. In other words, the prime spectrum of a ring can be viewed as a locally
ringed space. With that (nontrivial) fact in hand, we will be ready to glue prime spectra
together to manufacture arbitrary schemes.
Paradigm shift 2: functors
The second paradigm shift that stood between Weil and modern algebraic geometry is mostly
due to Grothendieck, though it is of a piece with the formalist view of mathematics pro
pounded by the Bourbaki school of French mathematicians in the middle of the 20th century.
It is to conceive of algebraic geometry in the language of categories and functors. Roughly
speaking, a category is the collection of all mathematical objects of a given type, equipped
with the maps between those objects that preserve the distinguishing structures. The key ex
ample to keep in mind is the category of all rings, together with all homomorphisms between
rings.
At rst, it may seem rather a bad idea to deal with categories; for one thing, they cannot
be viewed as sets due to some annoying paradoxes in set theory (such as Russells paradox).
But once you get past such considerations, dealing with categories is not so hard, and in fact
they appear everywhere around you.
Here is where categories appear naturally in algebraic geometry. Say P1 , . . . , Pm are
polynomials in the variables x1 , . . . , xn over a ring R. Then for any ring S equipped with a
homomorphism R S, it makes sense to consider the set
{(x1 , . . . , xn ) S n : P1 (x1 , . . . , xn ) = = Pm (x1 , . . . , xn ) = 0}
of S-valued solutions to the system of equations P1 = = Pm = 0. One should thus avoid
linking these polynomials to a single set of points, but instead view them as a scheme for
converting rings into sets of solutions. This gives a natural example of a functor between two
categories, i.e., a rule for converting objects of one category into objects of the other, and
for converting morphisms between two objects of the rst category into morphisms between
the image objects of the second category. (In our example, we are converting R-algebras
into sets.)
One benet of this point of view is that it naturally distinguishes, for instance, the zero
loci of y and y 2 : they give the same sets when we plug in an algebraically closed eld k, but
not when we plug in a ring such as k[]/(2 ).
That benet by itself is not so signicant, as it still doesnt really prove that category
theory is good for anything other than formulating simple statements in complicated lan
3
guage. What makes category theory so useful, and how we will exploit it in our work, is that
it lets you formalize certain types of reasoning by analogy that mathematicians would like
to engage in all the time, but which is sometimes dicult. One key example in the context
of schemes is the notion of a product. Given two mathematical objects X and Y , how should
one dene their product X Y ? When X and Y are given as sets carrying some extra
structure (e.g., groups, rings, etc.), the correct answer is to take the Cartesian product of
the underlying sets and then somehow cook up a good structure on that.
From the point of view of category theory, though, the right way to answer this question is
to specify a universal property that should be satised by the product. Namely, the product
X Y should have the following properties.
(a) It should come with projection maps 1 : X Y X and 2 : X Y Y .
(b) Given any object Z, the function taking a map f : Z X Y to the pair of compo
sitions 1 f : Z X and 2 g : Z Y should be a bijection.
This does not by itself actually construct products; indeed, some categories may not always
admit product objects according to this denition. However, it does give a characterization
of how a correct denition of a product object should behave. In fact, its okay to come up
with two dierent denitions as long as they both satisfy the universal property; the eect
is that there will be canonical identications between the two types of projects.
We will use this particular example to construct products in the category of schemes.
There, we will discover that the product of two schemes does not have underlying set equal
to the Cartesian products of the underlying sets!
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Category theory (updated 8 Feb 09)
Were going to use the language of category theory freely. Fortunately, its easy to learn
because it corresponds naturally to the way you (hopefully) already think about mathemat
ical objects. (I could give a reference, but in fact you should be ne just looking these things
up in Wikipedia.)
Warning: set-theoretic diculties
Category theory is a bit tricky because it tries to deal with objects like the ??? of all sets,
or all rings, or whatnot. Russells paradox shows that there is in fact no set of all sets.
Namely, if there were, it would have a subset consisting of those sets U for which U
/ U.
But that would then be a set V , and if V V then V
/ V and vice versa.
The way around this is to tamper with the axioms of set theory slightly, by introducing
the notion of a class. A class is something which behaves just like a set whose members are
sets, except that there is no power axiom; there is not guaranteed to be a class consisting
of all subclasses of a given class. Unless, that is, your class consists just of the elements of
some actual set. (You might then ask what kind of object is the ??? of all classes. Never
mind that for now.) We also assume there is a class of all sets, called the universe.
Except for the power axiom, you may perform operations on classes like you do with
sets. For instance, given a class C and a logical statement P depending on a single set,
you can form the subclass of C consisting of all elements for which P is true. You can also
form Cartesian products indexed by sets, although Ill hardly ever do this except for nite
products. (There is also an axiom of choice at the class level.)
A class is small if its elements are in bijection with some set.
Categories, and examples
A category C consists of the following data.

A class of objects, denoted Obj(C).
For each ordered pair of objects (X, Y ), a set of morphisms, denoted Hom(X, Y ). (You
may think of this as an element of the Cartesian product of two copies of C and one
copy of the universe.)
For each ordered triple of objects (X, Y, Z), a function : Hom(Y, Z) Hom(X, Y )
Hom(X, Z), called composition, which satises the following properties.
The associative law : given an ordered quadruple of objects (X, Y, Z, W ), the two
ways to compose Hom(Z, W ) Hom(Y, Z) Hom(X, Y ) to Hom(X, W ) give the
same answer.
1
The identity law : for each object X, there must exist a morphism id X Hom(X, X)
which is an identity under composition on either side. Note that id X is forced to
be unique by this condition.
(I have a habit of calling morphisms arrows because they are usually pictorially represented
as such.)
This denition is meant to capture many, if not all, basic types of structured mathematical
objects. Examples:
The category of sets, denoted Set, where Hom(X, Y ) is all functions from X to Y .
The category of topological spaces, denoted Top, where Hom(X, Y ) is all continuous
functions from X to Y .
The category of (commutative, unital) rings, denoted Ring, where Hom(X, Y ) is all
ring homomorphisms from X to Y .
The category of topological rings, denoted TopRing, where Hom(X, Y ) is all continuous
ring homomorphisms from X to Y .
The category of modules over a xed ring R, denoted ModR , where Hom(X, Y ) is all
R-module homomorphisms from X to Y .
And so forth. Ill leave to your imagination the denitions of some more categories for which
I might need names later: Ab (abelian groups), Grp (groups), TopGp (topological groups).
However, there are other things that can be viewed as categories. An important example:
given any partially ordered set S, make a category in which the objects are the elements of
S, and there is exactly one morphism from X to Y if X Y and none otherwise.
Important special case of the previous one: given a topological space S, we can make a
category in which the objects are the open subsets of S, and the morphisms are the inclusions
of one open subset into another.
Another example comes from algebra. Given a group, we can make a category with
only one object X, in which Hom(X, X) is the group and the composition law is the group
operation.
Heres a more interesting example along the lines of the previous one. Given a topological
space S, make a category in which the objects are the points of S, and the morphisms from
X to Y are the continuous functions f : [0, 1] S with f (0) = X and f (1) = Y . Dene the
composition g f for g Hom(Y, Z) and f Hom(X, Y ) to be the function h : [0, 1] S
with
f (2x)
x [0, 1/2]
h(x) =
g(2x 1) x [1/2, 1].
This is a special case of turning a groupoid (something which is like a group except that
objects can only be composed if they satisfy a matching condition) into a category. This
example comes from the fundamental groupoid of a topological space.
2
Interlude: is versus does
The rigorous formulation of category theory exposes a dark secret of mathematics: objects
in a category are rarely ever equal. For instance, we all think we agree on what the ring Z
is, but if we all sat down and wrote down set-theoretic denitions, probably no two of them
would exactly match. The point is that we conceive of Z, and of most mathematical objects
in general, not in terms of what they literally are as sets, but by how they work, and in
particular how they relate to other mathematical objects.
The solution for this suggested by category theory is to characterize interesting mathe
matical objects using universal properties. For instance, the ring Z is characterized by the
fact that it is an initial object in the category of rings: for every ring Y , there is a unique
morphism from Z to Y . Any two objects with this property are uniquely isomorphic.
Here are a few other arrow-theoretic properties that can be used for this purposes. Ill
talk more about universal properties later.
Y Obj(C) is a nal object in C if for any X Obj(C), there is a unique morphism
from X to Y . An object which is both initial and nal is a terminal object.
A morphism f Hom(X, Y ) is a monomorphism if for any g, h Hom(W, X), if
f g = f h, then g = h. In the category of sets (and many other examples), f is a
monomorphism if and only if f is injective.
A morphism f Hom(X, Y ) is an epimorphism if for any g, h Hom(Y, Z), if g f =
h f , then g = h. In the category of sets (and many other examples), f is an
epimorphism if and only if f is surjective. But beware of surprises: for example,
the morphism Z Q of rings is an epimorphism (and also a monomorphism).
A morphism f Hom(X, Y ) is an isomorphism if it has a two-sided inverse. This im
plies that it is a monomorphism and an epimorphism, but not conversely (see previous
example).
Functors and natural transformations
Functors can be thought of as functions between categories. A covariant functor from a

category C1 to a category C2 consists of:
A function F from Obj(C1 ) to Obj(C2 ).
For each pair (X, Y ) of Obj(C1 ), a function FX,Y : Hom(X, Y ) Hom(F (X), F (Y )),
such that F commutes with composition and F carries idX to idF (X) .
A contravariant functor works the same way except that FX,Y carries Hom(X, Y ) to Hom(F (Y ), F (X)),
that is, it reverses the sense of the morphisms. You can turn it into a covariant functor by
replacing one of the two categories with its opposite category, in which all morphisms are
reversed; for simplicity, let us just talk about covariant functors for the moment.
This point is actualized by the notion of a natural transformation. Given two functors
F1 , F2 from C1 to C2 , a natural transformation of F1 to F2 consists of, for each X Obj(C1 ), a
morphism X : F1 (X) F2 (X) such that for every morphism f Hom(X, Y ), the diagram
F1 (X)
F1 (f )
F1 (Y )
F2 (X)
Y
F2 (f )
F2 (Y )
is commutative (that is, if you trace around both ways you get the same answer). Natural
transformations can be composed; one with an inverse (equivalently, in which the morphisms
X are all isomorphisms) is called a natural isomorphism. For instance, the functors taking
ordered triples (M1 , M2 , M3 ) of modules over a ring R to
(M1 R M2 ) R M3
and
M1 R (M2 R M3 )
are naturally isomorphic.
Other properties of functors
A functor is faithful if the maps FX,Y are injective. Typical examples of these are forgetful
functors, in which you start with a category of objects carrying a lot of structure, and the
functor strips o some structure. E.g., take groups to their underlying sets, or take rings to
their additive groups, or take topological groups to their underlying topological spaces.
The analogues of injectivity and surjectivity for functors are:
A functor is fully faithful if the maps FX,Y are bijective. A typical example is the
inclusion of a full subcategory (i.e., take some of the objects, and all of the morphisms
between the chosen objects).
A functor is essentially surjective if every object in C2 is isomorphic to an object of the
form F (X) for some X Obj(C1 ).
A functor is an equivalence of categories if it is fully faithful and essentially surjective.
This is equivalent to the existence of a quasi-inverse functor, i.e., one for which the
compositions in both directions are naturally isomorphic to the relevant identities.
A typical example from last semester: take the category of ane algebraic varieties over
an algebraically closed eld k. The functor computing regular functions is an equivalence
between this category and (the opposite category of) nitely generated k-algebras which are
reduced (have no nilpotent elements). One of the goals of schemes is to set up a similar
equivalence between some sort of geometric objects and the category of all commutative
unital rings.
6 Representable functors, Yonedas lemma, and uni

versal properties
An individual object in a category casts a sort of shadow on the entire category, via the
notion of representable functors. For a xed object X in a category C, let hX be the functor
from C to Set such that hX (Y ) = Hom(X, Y ), and the image of f Hom(Y, Z) under hX
carries Hom(X, Y ) to Hom(Y, Z) via postcomposition with f .
It turns out that any natural transformation from hX to any other functor F : C Set is
determined by specifying the image of the special element idX of Hom(X, X) = hX (X), and
conversely any such choice induces a natural transformation from hX to F . This is Yonedas
lemma; proof is left as an (easy) homework problem.
An arbitrary functor F : C Set is representable if it is naturally isomorphic to hX
for some X. By Yonedas lemma, if X and Y represent the same functor, then they are
isomorphic in a natural way (i.e., one compatible with the action of the functor).
In practice, this is usually interpreted as saying that an object of a category determined
by a universal mapping property is unique up to unique isomorphism (or up to natural
isomorphism). Here is an example of this which will help explain why categorical thinking
is so helpful when dealing with schemes. For objects X, Y in a category C, an (absolute)
product of X and Y is an object Z equipped with maps Z X and Z Y , with the
following universal mapping property. Given any object W and morphisms W X and
W Y , there must be a unique morphism W Z such that the diagram
W
commutes. The product is unique in the sense that if Z is an other object equipped with
morphisms Z X and Z Y satisfying the mapping property, there is a unique isomor
phism Z Z making everything commute.
In any normal category, in which objects are sets equipped with some extra structure
(e.g., groups, topological groups), products exist and can be written as Cartesian products
with some appropriate extra structure. But in general, products need not exists, and even
if they do they might look weird. Case in point: suppose we tried to make a theory of
abstract algebraic varieties over the non-algebraically closed eld Q, in which the points are
Galois orbits of points over Q. (This is close to what will happen with schemes, except that
there will be some more points.) Then in the ane line, we have a variety consisting of
the single orbit {i, i}. The product of this with itself will then consist of the two orbits
{(i, i), (i, i)} and {(i, i), (i, i)}.
Limits and colimits
The universal mapping properties we will consider can all be wrapped into the following
framework. Let C, D be two categories. A diagram on C of type D is just a functor from D
to C.
Fix a diagram F : D C. Let D be the category formed from D by adding one extra
object I with a unique morphism to every object in D (and the obvious composition law).
Now look at extensions of F to functors D C; that is, you have to add one object X
of C and maps X F (Y ) for each Y D which commute with the maps coming from
the diagram. A limit of F is a universal set of such data, i.e., any other extension factors
uniquely through this one. My example of a product is the case where D consists of two
objects and no morphisms.
Dene colimits as limits in the opposite category. For example, the co-analogue of the
product is the coproduct. In Set, the product is the Cartesian product, while the coproduct
is the disjoint union.
Important special case: a directed set is a partially ordered set in which any two elements
have a common upper bound. (I.e., for any x, y, there is some z with x z, y z. A diagram
from a directed set into some category C is called a direct system; a colimit of a direct system
is called a direct limit, or an inductive limit, in C. (It should be called a direct/inductive
colimit. Sorry about that.) For example, take the natural numbers under divisibility; then
the direct limit of the abelian groups n1 Z is the group Q.
A diagram from the opposite of a directed set into some category C is called an inverse
system; a colimit of an inverse system is called an inverse limit (or projective limit). For
example, view the nonnegative integers as a partially ordered set using the reverse of the
usual ordering. Then for any ring R, the inverse limit of the rings R[x]/(xn ) is the ring Rx
of formal power series. (A similar example is the p-adic numbers.)
Adjoint functors
One other notion that comes up a lot is that of an adjoint pair of functors, which you might
like to think of as category-theoretic analogues of a linear operator and its transpose.
Let C, D be categories. A pair of functors F : C D and F : D C form an adjoint
pair if we can form bijections
HomC (F X, Y ) HomD (X, F Y )
which are functorial in X and Y (imagine the diagrams yourself). In this relationship, F is
the left adjoint and F is the right adjoint.
The notation was chosen because the adjoint pairs we will use correspond to operations of
promotion and demotion between two categories, one of which has more structured objects
than the other. Here is a typical example. Let F : Ab Set be the forgetful functor on
abelian groups. Let F : Set Ab be the functor carrying a set S to the free abelian group
generated by S. Then F and F form an adjoint pair.
6
Another important example for us: let R S be a homomorphism of rings. Let F :

ModR ModS be the functor M M R S. Let F : ModS ModR be the functor given
by restriction of scalars from S to R. Then F and F form an adjoint pair.
We can of course compose F and F both ways, and we dont in general get the identity,
or even something naturally isomorphic to the identity. We do get something interesting,
though. The identity map on F X corresponds to a morphism X F F X, while the
identity map on F Y corresponds to a morphism F F Y Y . These morphisms are called
adjunction morphisms. For example, in the previous example, for X an R-module, X
F F X = X
R S is the map x x 1. For Y an S-module, Y R S Y is the map

i yi si
i y i si .
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Sheaves (updated 12 Feb 09)
We are now ready to introduce the basic building block in the theory of schemes, the
notion of a sheaf. See also: Hartshorne II.1, EGA 1 0.3. (The latter means: look in the
Chapitre 0 section of EGA volume 1.) The base reference for this bit of EGA is Godement,
Theorie des Faisceaux.
Note that Hartshorne assumes all sheaves take values in the category of abelian groups,
that being the case of most interest in algebraic geometry. I will only impose that restriction
in the next lecture.
Presheaves
Fix a category C, e.g., sets or abelian groups. Given a topological space X, let X be the
category of open sets of X. A presheaf on X with values in C is a contravariant functor
F : X C. the category of open sets of X to C. In other words, to specify a sheaf F on X,
you must specify:
(a) for each open subset U of X an element F (U) C;
(b) for each inclusion V U of open subsets of X a morphism ResU,V = ResU,V (F ) :
F (U ) F (V ) (called restriction), such that:
(i) for each open subset U of X, ResU,U = idF (U ) ;

(ii) for each series of inclusions W V U of open subsets of X, we have ResV,W ResU,V =
ResU,W within Hom(U, W ).
There seems to be some confusion over whether it is required that F () is required to be
a nal object of C; Hartshornes two characterizations of presheaves disagree on this point
(because the denition of a functor doesnt include this condition). Fortunately, this doesnt
have any serious consequences; the denition of a sheaf will stamp out this ambiguity. (EGA
avoids this issue by omitting the denition of a presheaf entirely!)
We will typically use this denition in cases where C carries a forgetful functor to Set.
In that case, it makes sense to speak of the elements of F (U) for U an open subset of X;
we call these elements the sections of F on U. For V U an inclusion of open sets, and
s F (U ), we often write s|V instead of ResU,V (s).
The restriction of a presheaf F on X to an open subset U of X is dened in the obvious
fashion. It is denoted F |U . It is also called the induced presheaf of F on U.
If F1 , F2 : X C are both presheaves on a topological space X with values in a category
C, a morphism F1 F2 of presheaves is a natural transformation of functors from F1 to
F2 , i.e., a collection of maps F1 (U) F2 (U) compatible with restrictions.
Sheaves
Here is an example of a set-valued presheaf F: take another topological space Y , and put
F (U ) = HomTop (U, Y ) (the continuous functions from U to Y ) with restriction being the
usual restriction of functions. This example has a special feature not implied by the denition
of a presheaf: a continuous function can be specied locally. In other words, for any index
set I, if {Vi }iI is a family of open sets with union U, then on one hand, each element of
F (U ) is uniquely determined by its restrictions to all of the Vi ; and on the other hand, any
family of elements of F (Vi ) which agree on the overlaps of the Vi gives a section over U .
This is formalized by the notion of a sheaf. A sheaf on X with values in C is a presheaf
with the following property (called the sheaf axiom).
Axiom (Sheaf axiom). For any index set I, for any family of open sets {Vi }iI which form
a cover of the open set U, the object F (U) is the limit of the diagram formed by the F (Vi )
for i I, the F (Vi Vj ) for i, j I, and the arrows ResVi ,Vi Vj for i, j I.
Let us make this explicit in case C = Set. Dene I, U, Vi as in the sheaf axiom.
(i) If s1 , s2 F(U) is such that s1 |Vi = s2 |Vi for all i, then s1 = s2 . (If C = Ab, we can
just check this for s2 = 0.)
(ii) Suppose we are given for each i I, an element si F(Vi ) such that for each i, j I,
si |Vi Vj = sj |Vi Vj . Then there exists an element s F(U) such that s|Vi = si for each
i. (The element s is unique by (i).)
We dene restriction of sheaves, and morphisms of sheaves, by copying the denitions
from presheaves.
Some examples of sheaves:
On a manifold, the continuous functions to some xed topological space. Special
example: if you take a target space C equipped with the discrete topology, you get
whats called the locally constant sheaf associated to C.
On a dierentiable manifold, the dierentiable functions.
On a complex manifold, the holomorphic functions.
On an abstract algebraic variety over an algebraically closed eld, the regular functions,
or the dierential forms.
These all come from a class of objects called locally ringed spaces, which we will discuss later.
Although sheaves can be dened to take values in an arbitrary category, we will only
be interested in cases where the category consists of objects with well-dened elements, and
all the glueing is determined by the elements. So to keep things simple, let me drop in a
hypothesis that I would like to keep in place from now on. (With only limits, Grothendieck
calls this hypothesis (E). However, well want the colimits in order to talk about stalks later.)
2
Hypothesis (E). Assume hereafter that all sheaves under discussion take values in a xed
category C which admits a forgetful functor to Set that reects small limits and colimits.
That is, all small (indexed by sets) limits exist, and their formation commutes with passage
to Set.
For example, C could be Set itself. It could also be any one of the usual algebraic
categories: Ab, Grp, Ring, ModR for a ring R, etc. Under this hypothesis, the sheaf axiom
for C is exactly as for Set, so a presheaf is a sheaf if and only if it becomes a sheaf after
composing with the forgetful functor. We can thus forget the extra structure of C when
checking basic facts about sheaves.
A typical bad example is Top; the basic problem is that the image of a morphism under
the forgetful functor can be an isomorphism even if the original morphism is not. That is, a
continuous bijection of topological spaces need not be a homeomorphism.
Here is a trick for dealing with bad cases: given a presheaf F on X, for each object
Y C, let FY be the presheaf on X with values in Set dened by U Hom(Y, F(U )). Then
F is a sheaf if and only if each FY is a sheaf.
Dening sheaves on a basis
It is very often convenient not to have to explicitly specify the sections of a sheaf on every
open subset, but simply on a basis of open sets. Recall that a basis (of open sets) in a
topological space X is a collection of open sets such that every open set can be written as a
union of elements of the basis.
Let X be a topological space, and let X be the category of open sets of X. Let B be
a basis of X, and let B be the full subcategory of X with Obj(B) = B. (That is, keep all
of the morphisms.) A presheaf on X specied on B is a contravariant functor from B to C.
A sheaf on X specied on B is a presheaf F on X specied on B, such that F satises the
following modied sheaf axiom.
Axiom (Sheaf axiom for a basis). For any index set I, for any U B and any family of
open sets {Vi }iI in B which form a cover of U, we can choose a covering {Wijk }kJi,j of each
Vi Vj such that the object F(U ) is the limit of the diagram formed by the F(Vi ) for i I,
the F (Wijk ) for i, j I and k Ji,j , and the arrows ResVi ,Wijk for i, j I and k Ji,j .
For example, suppose B is a basis in which the intersection of any two basic opens is a
basic open; Ravi Vakil calls this a nice basis, so I will too. For a nice basis, this follows from
the sheaf axiom applied to coverings of basic opens by other basic opens, because you just
take the trivial covering of Vi Vj by itself. (The niceness condition is satised in most of
our examples.)
Lemma (Basis lemma). Any sheaf on X specied on B extends uniquely to a sheaf on X.
Similarly, any morphism between two sheaves on X specied on B extends to a morphism of
sheaves on X.
3
In other words, the restriction functor from sheaves on X to sheaves on X specied on

B is an equivalence of categories.
Proof. Let F be the presheaf dened by taking F(U ) to be the limit of the diagram formed
by the F (V ) (and the restriction maps) for all basic opens V contained in U. If U is a basic
open, then the construction comes with a map F (U) F(U) which denes a morphism
of presheaves specied on B. Also, the limit property also denes the restriction maps
ResU,V : F (U) F (V ) whenever V U are arbitrary opens, since F (U) maps to F(W )
for any basic open W contained in V . By a similar argument, any morphism F G of
presheaves induces a morphism F G .
What is left to check that on one hand the map F (U) F(U) is an isomorphism, and
on the other hand F satises the sheaf axiom. We leave these as exercises.
As a corollary, we learn how to glue sheaves together.
Corollary. Let I be an index set and let {Ui }iI be an open cover of X. Suppose we are
given the following data.
(a) For each i I, a sheaf Fi on Ui with values in C.
(b) For each i, j I, an isomorphism ij : Fi |Ui Uj
= Fj |Ui Uj , satisfying the following

conditions.
(i) For each i I, ii is the identity morphism on Fi .
(ii) For each i, j, k I, we have jk ij = ik as morphisms of sheaves on Ui Uj Uk .
(This is called the cocycle condition, for reasons to be discussed later.)
Then there exist a sheaf F on X and isomorphisms i : F|Ui
= Fi for each i I, such that
for each i, j I, ij i = j . Moreover, F is unique up to unique isomorphism (in a sense
to be interpreted by the reader).
You might describe this by saying that a sheaf of sheaves is a sheaf. In fact, this is the
same sort of data needed to glue, say, topological spaces.
Proof. Suppose we are in the happy situation where whenever an open set U of X belongs
to both Vi and Vj , we have a literal equality Fi (U) = Fj (U) and the map ij between these
two is the identity morphism. (Note that the cocycle condition is automatically valid here.)
Then we can apply the basis lemma, where B is the (nice) basis consisting of those open
sets U contained in Ui for at least one index i.
The trouble is that as usual, objects in a category are usually not equal. However,
using the cocycle condition we can force them to become equal as follows. Dene a functor
F : B C as follows. For U B, pick an index i = i(U) such that U Ui , and put
F (U ) = Fi (U). For an inclusion V U of elements of B, put i = i(U) and j = i(V ), so that
V is contained in both Ui and Uj . Dene ResU,V (F) as the composition of the restriction
map ResU,V (Fi ) : Fi (U) Fi (V ) with the map ij : Fi (V ) Fj (V ). The cocycle condition
4
then implies that these restriction maps are associative, so they dene a presheaf F specied
on B. The fact that each Fi is a sheaf implies that F is a sheaf specied on B, so it extends
to a sheaf.
Stalks
An important source of information about sheaves is given by looking at their behavior in

the neighborhood of a point, as follows.
First let us recall something about direct limits. (Warning: I had the terminology slightly
wrong when I introduced this in the category theory lecture. The notes have been corrected.)
A directed set is a partially ordered set in which any two elements have an upper bound (but
not necessarily a least upper bound). A direct system in a category C is a covariant functor
F : P C with P a directed set. If the colimit exists, it is called the direct limit of the
system.
Before using this notion for much, it might be helpful to make it explicit in the case of
sets. (The case of abelian groups, which we also use, works the same way.) In this case, the
direct limit is formed by starting with the union of F (S) over all S P , then identifying the
elements x F (S) and y F (T ) if there exist arrows f : S U and g : T U in P such
that F (f )(x) = F (g)(y). A typical example is the formation of the fraction eld Frac(R) of
an integral domain R, as the direct limit of the rings R[x]/(xf 1) over all nonzero f R.
Here the poset is the nonzero elements of R ordered under divisibily, and the map from
R[x]/(xf 1) to R[x]/(xf g 1) takes x to xg.
Now let F be a presheaf on the topological space X, and let x X be any point. View
the open subsets of X containing x as a partially ordered set Px under reverse inclusion.
They then form a directed set, and the direct limit of the functor F : Px C is called the
stalk of F, denoted Fx .
The elements of a stalk (which exist because we assumed (E)) are typically called germs.
If s is a section of a sheaf on an open set containing x, we write sx for the germ of s at x.
Example: the stalk of the sheaf of real-valued continuous functions consists of germs of
real-valued continuous functions. Two continuous functions dened on open subsets of X
containing a point x determine the same germ at x if and only if they coincide on some open
subset containing x.
We can make a similar construction for the other functions on manifolds examples
above. Beware that in these examples, the germ of a function at a point carries much more
information than the value at that point. Extreme example: two holomorphic functions
dened on a connected complex manifold have the same germ at a single point if and only
if they coincide (because of analytic continuation!).
One variant well need a bit later: given any subset Z of X, not necessarily a single point,
we can similarly take the direct limit of F(U) over all open subsets U of X containing Z.
We call this the stalk of X at Z.
Stalks and morphisms
Stalks can be used to detect lots of interesting properties of sheaves, particularly in relation
to morphisms. Throughout this section, let : F1 F2 be a morphism of sheaves on a
topological space X.
Lemma. Consider the following conditions.
(a) For each x X, the map x : F1,x F2,x is injective/surjective/bijective.
(b) For each open subset U of X, the map (U ) : F1 (U) F2 (U) is injective/surjective/bijective.
Then (b) implies (a) in all cases, while (a) implies (b) in the injective and bijective cases.
Proof. Suppose (a). Let Yi be the product of F
i,x over all x U. Then the sheaf axiom
implies that the map Fi (U) Yi carrying s to x sx is injective. This gives injectivity in
(b). (This is a toy example of the construction of the espace etale of a sheaf; I asked more
about it on Problem Set 1.)
If x is bijective for all x, then for any section t F2 (U) and any x U, there is an
open neigborhood V = Vx of x on which t coincides with the image under of some section
sx F1 (Vx ). For y U also, the restrictions of sx and sy to F1 (Vx Vy ) have the same image
under (namely the restriction of t to F2 (Vx Vy )), so they coincide by what we proved in
the previous paragraph. We can thus invoke the sheaf axiom to assemble s F1 (U) with
(s) = t. so surjectivity/bijectivity in (b) is an easy consequence.
Suppose (b). The surjectivity aspect is more or less obvious, so we only check the
injectivity aspect. Suppose we are given two elements of F1,x with the same image in F2,x .
We can represent these by sections s1 , s2 of F1 on some open neighborhood of x. In fact, we
can take them on the same open neighborhood U . Their images are sections of F2 which
have the same image in F2,x . That means that we can replace U by some smaller open
neighborhood V so that (s1 ) and (s2 ) coincide in F2 (V ). But then s1 = s2 in F1 (V ), so
(a) holds.
We dene a morphism of sheaves to be injective/surjective/bijective if it has the cor
responding property on stalks. By the previous lemma, bijective is the same as being an
isomorphism (in the sense of having an inverse).
The disturbing thing is of course the failure of the implication from (a) to (b) in the
surjectivity case. Yes, a morphism of sheaves can be surjective without being surjective on
sections! What is true is: if is surjective and U is an open in X, then for each s F2 (U),
we can cover U with open subsets Vi such that ResU,Vi (s) is in the image of (Vi ) for each
i. The trouble is that you may not be able to choose elements of the F1 (Vi ) which can be
glued.
Here is a familiar example. Put X = C \ {0}. Let F1 be the sheaf of holomorphic
functions on X. Let F2 be the sheaf of nowhere vanishing holomorphic functions on X. Let
: F1 F2 be the map taking f : U C to exp f . Then is surjective because the
logarithm of a nonzero holomorphic function exists locally, but not globally: the function
z F2 (X) is not in the image of (X).
6
Sheacation
If we x a topological space X and a category C, there is an obvious forgetful functor from

sheaves on X with values in C to presheaves on X with values in C. If you properly digested
the notion of an adjoint functor, you should be asking whether this forgetful functor occurs
as the right adjoint in an adjoint pair. It does!
Let F : X C be a presheaf on X with values in C. Dene
another presheaf F + on X
as follows.
For U X open, take F + (U) to be the subset of xU Fx consisting of elements
s = x sx with the following property: for each x U, there exists an open neighborhood
V of x in U and a section t F (V ) such that sy = ty for all y V . From the denition,
it is easy to check that F + is a sheaf and that its stalk Fx+ is canonically isomorphic to Fx .
We call F + the sheacation of F; its construction is functorial in F.
Proposition. The functor F F + from presheaves on X to sheaves on X, and the forgetful
functor from sheaves on X to presheaves on X, form an adjoint pair.
Proof. Exercise.
Direct and inverse image
Let f : X Y be a continuous map. For F a sheaf on X, the formula

(f F)(V ) = F(f 1 (V ))
obviously denes a sheaf f F on Y . It is called the direct image of F.
Now let G be a sheaf on Y . Dene a presheaf f1 G on X as follows: for U open in X,
let (f1 G)(U) be the stalk of G at f (U), i.e., the direct limit of G(V ) over open sets V X
containing f (U). This is general not a sheaf; its sheacation is called the inverse image of
G, denoted f 1 G.
Proposition. The functors f 1 and f form an adjoint pair.
Proof. Exercise.
You might wonder why I didnt use the notation f for the inverse image. That is because
I will need that notation later for a dierent functor, dened for a morphism of ringed spaces.
Using the inverse image, we can dene the restriction of F to an arbitrary subset Z of
X, as the sheaf i1 F for i : Z X the inclusion map (with Z given the subspace topology).
If Z = {x}, this coincides with the stalk Fx (exercise).
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

More on abelian sheaves
We now specialize the discussion of sheaves to the situation where the target category
consists of abelian groups. At the end, Ill explain how to generalize to the case of a target
which is an abelian category.
Abelian groups
Assume until I say otherwise that C = Ab. (At the end, well generalize to the case where
C can be any abelian category.) Let me rst set some notation and terminology about
morphisms of abelian groups themselves.
For f : A B a morphism of abelian groups,
ker(f ) = {x A : f (x) = 0}
im(f ) = {f (x) : x A}
coker(f ) = A/ im(f ) = {y + im(f ) : y B}.
A (nite or innite) sequence
Ai1 Ai Ai+1
in Ab is exact if for each i,
im(Ai1 Ai ) = ker(Ai+1 Ai ).
If we only have the weaker assertion that
im(Ai1 Ai ) ker(Ai+1 Ai )
(i.e., the composition Ai1 Ai Ai+1 is zero) for each i, we say that the sequence is a
complex.
Here are some useful facts about exact sequences; their proofs are fun exercises in what
is sometimes called diagram chasing. Remember that in Ab, monomorphism equals injective
and epimorphism equals surjective (so mono plus epi equals iso, which is not true in an
arbitrary category).
Lemma (Five lemma). Let
A0
A1
f0
A2
f1
B0
f2
B1
A3
B2
be a commuting diagram in C with exact rows.

1
A4
f3
f4
B3
B4
(a) If f1 and f3 are monomorphisms and f0 is an epimorphism, then f2 is a monomor

phism.
(b) If f1 and f3 are epimorphisms and f4 is a monomorphism, then f2 is an epimorphism.
Proof. Exercise.
Lemma (Snake lemma). Let
0
A1
A2
f1
f2
B2
f3
B1
A3
B3
be a short exact sequence. Then there exists a canonical homomorphism : ker(f3 )

coker(f1 ) (the connecting homomorphism) such that
0 ker(f1 ) ker(f2 ) ker(f3 ) coker(f1 ) coker(f2 ) coker(f3 ) 0

is exact, where all the maps other than are the obvious ones induced by the diagram.
Proof. Here is what is supposed to be: given a3 ker(f3 ), lift it to a2 A2 , then apply f2
to get b2 B2 . Since the diagram commutes, b2 must map to zero in B3 , so it lifts to b1 in
B1 . Declare (a3 ) = b1 .
It remains to show that is well-dened and is a homomorphism, and that the claimed
sequence is exact. These are left as exercises.
Corollary (Short ve lemma). Let
0
A1
A2
f1
B1
A3
f2
f3
B2
B3
be a commuting diagram in C with exact rows. Then f2 is a monomorphism/epimorphism if

and only if f1 and f3 both are.
Exact functors
For C1 = C2 = Ab, a covariant functor F : C1 C2 is additive if it commutes with addition

of morphisms. Any additive functor sends complexes to complexes (because the property of
the composition of two maps being zero is preserved), but not necessarily exact sequences
to exact sequences. Hence the following denitions.
We say F is left exact if for any exact sequence
0 A1 A2 A3
2
the sequence
0 F (A1 ) F (A2 ) F (A3 )
is exact. We say F is right exact if for any exact sequence
A1 A2 A3 0
the sequence
F (A1 ) F (A2 ) F (A3 ) 0
is exact. We say F is exact if it is both left exact and right exact; equivalently, for any exact
sequence
0 A1 A2 A3 0
the sequence
0 F (A1 ) F (A2 ) F (A3 ) 0
It in turn implies that any exact sequence of any length goes into another exact sequence
under F . (Ill try avoid using these notions for contravariant functors, since there is a
left/right ambiguity.)
Examples:
For any given X C, the covariant functor Hom(X, ) is left exact.
For any given X C, the covariant functor X is right exact.
Many left/right exact functors arise from the following proposition.
Proposition. Suppose the covariant functors f : C1 C2 and f : C2 C1 form an adjoint
pair. Then f is right exact and f is left exact.
Proof. Exercise.
Abelian sheaves
Let F be a sheaf on a topological space X with values in C = Ab. A subsheaf of F is what

you think: take a subset of the sections on each open so that you still have a sheaf. The
quotient of F by a subsheaf G is a bit trickier: take the presheaf U 7 F (U)/G(U), then
sheafy. Note that the stalk at x is indeed Fx /Gx .
Given a morphism : F G of sheaves, the presheaf U 7 ker((U)) is a sheaf; we call
it the kernel of . The presheaves U 7 im((U)) and U 7 coker((U)) are not in general
sheaves; their sheacations are called the image and cokernel of .
Proposition. For x X, we have ker()x = ker(x ), im()x = im(x ), and coker()x =
coker(x ). Consequently,
im()
= F / ker(),
G/ im().
coker() =
3
Proof. Exercise.
Using these, we extend the notion of exactness to a sequence of sheaves; its equivalent
to dene it using sheaves or stalks, but not using sections.
Let ShC (X) be the category of sheaves on X with values in C. We dene the global
sections functor (, X) : ShC (X) C by the formula
(F , X) = F (X).
(No set-theoretic diculties here: X is a small category, so sheaves on X with values in C
do form a class.)
Proposition. The global sections functor is left exact.
Proof. Exercise.
The failure of the global sections functor to be right exact will give rise to the notion of
sheaf cohomology later.
Abelian categories
Everything I dened above can be generalized to the case where C is what is called an abelian
category, i.e., a category which captures the useful properties of abelian groups.
First, let me give an ad hoc denition which will suce for our purposes. A nice abelian
category is an additive category in which all limits and colimits exist, together with a forgetful
functor to Ab which preserves limits and colimits.
Next, lets gure out what the correct abstract denition shoul be. We rst write down
the denition of an preadditive category (which I called an additive category by mistake on
Problem Set 1). That is a category C equipped with the structure of an abelian group on
each set Hom(X, Y ), over which composition is distributive.
We next dene an additive category. The key notion is that direct sum and direct product
coincide for a nite collection of abelian groups. We should thus require the existence of
biproducts: that is, for any X1 , . . . , Xn Obj(C), there must exist an object X equipped
with maps i : X Xi and i : Xi X, such that X is both a product (using the i ) and
a coproduct (using the i ), and the sum 1 1 + + n n is the identity on X. (Exercise:
this exists as soon as you have nite products.)
Since the empty biproduct exists, an additive category has a terminal (initial and nal)
object, which we call the zero object and label 0. In an additive category, we can dene a
kernel of the morphism f : X Y to be a limit of the diagram
X
Y
4
i.e., an object W plus a morphism g : W X such that f g = 0, and any other morphism
h : V X for which f h = 0 factors uniquely through g. Similarly, a cokernel of f is a
colimit of
X
Y
To get a preabelian category, we insist that every morphism admit a kernel and cokernel
(which as usual are only unique up to unique isomorphism). To get an abelian category, we
insist that every monomorphism be the kernel of its cokernel, and every epimorphism be the
cokernel of its kernel.
The Freyd-Mitchell embedding theorem asserts that at least for every small abelian cat
egory C, we can construct an exact and fully faithful functor F : C ModR for a not
necessarily commutative ring R (where ModR now means left modules). This lets you prove
theorems about abelian categories by reducing to situations where objects really do have
elements.
The main dierence between my nice abelian categories and true abelian categories is
that I want all limits and colimits to exist. This is a bit strong for some purposes, but since
I need limits anyway to work with sheaves, its not so strange.
Anyway, the point here is that if you start with a (nice) abelian category C, for any
topological space X, the category Sh C (X) is again a (nice) abelian category. This follows by
assembling various homework exercises.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Schemes
We next introduce locally ringed spaces, ane schemes, and general schemes. References:
Hartshorne II.2, Eisenbud-Harris I.1, EGA 1.1.
Ringed and locally ringed spaces
A ringed space is a topological space X equipped with a sheaf OX on X with values in

Ring (called the structure sheaf ). This denition isnt so useful because it doesnt force the
topology to have much to do with the ring structure; for instance, any ring can be viewed
as a ringed space on a one-element topological space.
A more useful notion is that of a locally ringed space. This is a ringed space in which for
each x X, the stalk OX,x of OX at x is a local ring, i.e., a ring with a unique maximal
ideal mX,x . (The zero ring is not a local ring!)
For example, suppose X is a manifold and let OX be the sheaf of real-valued continuous
functions. We check that (X, OX ) forms a locally ringed space. Given x X, let mX,x be
the ideal of OX,x consisting of germs of functions taking the value 0 at x. This is clearly
an ideal, and the quotient OX,x /mX,x is certainly contained in R. Since X is a manifold,
the quotient is nonzero, so mX,x is indeed a maximal ideal of OX,x . To check that it is the
unique maximal ideal, it suces to check that any f OX,x not contained in mX,x is a unit
in OX,x . For such an f , f (x) is some nonzero real number, so we can nd an open subinterval
I R such that f (x) belongs to I but 0 does not. Represent f by a continuous function
on some open subset U of X containing x, which Ill also call f . The key point is that by
continuity, V = f 1 (I) is again an open subset of X containing x, and f takes nonzero
values everywhere on V . Hence there exists a multiplicative inverse g of f on V , which is
necessarily continuous.
Similarly, a smooth manifold, complex manifold, or abstract algebraic variety equipped
with the obvious sheaf is a locally ringed space.
For any x X, the quotient OX,x /mX,x is a eld. We denote it by (x) and call it the
residue eld of x. In the aforementioned examples, the residue elds of all of the points of
x are the same (either R, C, or a prescribed algebraically closed eld), but that will not be
the case for schemes!
Ill talk about morphisms of (locally) ringed spaces later. For the moment, let me at least
point out that an isomorphism of (locally) ringed spaces is what you think: a homeomor
phism of topological spaces and corresponding bijections of sections which commute with
restriction.
The prime spectrum of a ring
The notion of a locally ringed space is a suciently broad generalization of manifolds that
it admits a meaningful functor from the category of arbitrary (commutative unital) rings.
1
This gives rise to the concept of an ane scheme; to dene this, we must rst recall the
construction of the prime spectrum of a ring. See the exercises for lots of examples.
Let R be an arbitrary ring. Following Zariski, we dene the prime spectrum of R,
denoted Spec(R), to be the set of prime ideals of R. (An ideal p of R is prime if R/p
is an integral domain. The zero ring is not an integral domain, so the trivial ideal is not
prime.) For a general ring, this is a better idea than using only maximal ideals because a
ring homomorphism : R S induces a map Spec(S) Spec(R) taking p S to 1 (p).
(The latter is prime because induces an injective map R/1 (p) S/p, so the source is
an integral domain.) By contrast, may not carry maximal ideals of S to maximal ideals
of R; for instance, consider : Z Q.
Again following Zariski, we equip the set Spec(R) with the Zariski topology, in which the
closed sets have the form
V (I) = {p Spec(R) : I p}
for I an ideal of R. This is indeed a topology because
V (I) V (J) = V (I J) = V (IJ)
V (Ii ) = V
Ii .
i
We will use a special basis of open sets for this topology: the distinguished open sets, of
the form
D(f ) = {p Spec(R) : f
/ p}
for f an element of R. Note that this basis is nice in the sense that the intersection of
any two distinguished opens D(f ) and D(g) is again a distinguished open, namely D(f g).
Note also that for : R S a homomorphism, the induced map Spec(S) Spec(R) is
continuous because the inverse image of D(f ) is D((f )).
Lemma. Any distinguished open D(f ) of Spec(R) is quasicompact for the Zariski topology.
In particular, Spec(R) = D(1) itself is quasicompact.
Proof. It is enough to prove that any covering of D(f ) by distinguished open subsets admits
a nite subcover. If the sets D(fi ) cover D(f ) (for i running over some arbitrary index
set), then the radical of (f ) is contained in the radical of the ideal generated by the f i . In
particular, some power of f is in the ideal generated by the fi . But that means that we can
write f as a nite R-linear combination of the fi , so those D(fi ) already cover D(f ).
For example, if k is an algebraically closed eld, then Spec k[x] consists of one point of
the form (x a) for each a k, plus a point corresponding to the prime ideal (0). The
latter is an example of a generic point, a point whose closure is equal to the entire space in
question. For the analogous picture of Spec k[x, y], see Hartshorne Example 2.3.4.
A presheaf of rings
We now specify a presheaf of rings on X = Spec(R), but only on the distinguished open
subsets. To do this, we must do a bit of work to clean up their description, to account for
the fact that prime ideals dont see the dierence between an element of a ring and a power
of that element.
Lemma. For f, g R, we have D(f ) D(g) if and only if some power of f is a multiple
of g.
Proof. Note that D(f ) = D(f n ) for any positive integer n. Hence if f n is a multiple of g
for some n, then D(f ) = D(f n ) is contained in D(g). Conversely, suppose D(f ) D(g),
or in other words, V (g) V (f ). Recall that the radical of the ideal (g) is the intersection
of the prime ideals containing (g). Since V (g) V (f ), it follows that the radical of (f ) is
contained in the radical of (g), so in particular f belongs to the radical of (g). That is, some
power of f is a multiple of g, as desired.
A multiplicative subset of R is a subset closed under multiplication. For example, Sf =
{1, f, f 2 , f 3 , . . . } is a multiplicative subset. A multiplicative subset S is saturated if for any
x R such that some power of x equals an element of S times a unit, we have in fact
x S. For any multiplicative subset S of R, there is a unique saturated multiplicative
subset S containing it, formed in the obvious fashon. By the previous lemma, we now have
the following.
Corollary. For f, g R, we have D(f ) = D(g) if and only if Sf = Sg .
Given any multiplicative subset S of R, there is a unique initial object among the Ralgebras in which each element of S has a multiplicative inverse. It is called the localization
of R at S, denoted S 1 R. We can construct it as the polynomial ring in one variable xf
for each f S, modulo the relations xf f 1. Note that there is a canonical isomorphism
S1 R
= S 1 R since they both satisfy the same universal property. In particular, we can
write
Sf1 R
= R[x]/(xf 1).
From now on, write Rf instead of Sf1 R.
Let D be the set of distinguished open subsets of X = Spec R. Dene a presheaf of rings
OX on X specied on D as follows. First put
OX (D(f )) = Rf ;
this is well-dened by the previous corollary. Then note that given an inclusion D(g) D(f ),
we have Rf Rg , so the universal property of localization gives a canonical homomorphism
Rf Rg . If you want to write this more concretely (but less canonically), apply the lemma
above to write f n = gh for some positive integer n, identify OX (D(f )) = R[x]/(xf 1) and
OX (D(g)) = R[y]/(yg 1), and take the R-algebra homomorphism
R[x]/(xf 1) R[y]/(yg 1),
3
x f n1 hy.
The fundamental theorem of ane schemes
We are now ready to prove what I call the fundamental theorem of ane schemes. I dont
know whether its appearance in EGA 1 is its rst.
Theorem 1. The presheaf OX on X = Spec R specied on D satises the sheaf axiom
for coverings of distinguished opens by other distinguished opens. Consequently, it extends
uniquely to a sheaf of rings on Spec R.
While were at it, though, we may as well prove something stronger which we will need
later. This proof is basically the same one used to compute the regular functions on an ane
algebraic variety. It may also be thought of as an enhancement of the Chinese remainder
theorem; indeed, the latter is an immediate corollary (exercise).
of abelian groups on X specied
Theorem 2. Let M be an R-module. Dene a presheaf M
satises the sheaf axiom for coverings
on D by the formula D(f ) M R Rf . Then M
of distinguished opens by other distinguished opens. Consequently, it extends uniquely to a
sheaf on Spec R.
Proof. By replacing R with Rf , we may reduce to checking the sheaf axiom for a cover
of X itself by some distinguished open subsets D(fi ). We rst verify that the map M
i M R Rfi is injective, as follows. Suppose m M belongs to the kernel of this map.

Then the annihilator of m is an ideal of R which cannot be contained in any prime ideal p
of R, or else we would have p D(fi ) for some i, and the image of m in M R Rfi would be
nonzero. Thus 1 m = 0, so m = 0.
This proves the rst half of the sheaf axiom; we must now check the glueing property.
For this, we remember that X is quasicompact, so we may reduce to checking for a nite
cover. Say D(f1 ), . . . , D(fn ) cover X. Suppose that some D(fi ) cover D(f ), and that we
h
are given elements mi /fihi M R Rfi such that mi /fihi and mj /fj j have the same image
in Rfi fj . Since there are only nitely many fi , we may take the nonnegative integers hi to
be equal to a common value h. For each i, j, we then have
(fi fj )gij (fih mj fjh mi ) = 0
for some nonnegative integers gij . By rechoosing the mi , we can force gij = 0 for all i, j,
that is, we now have literal equalities
fih mj = fjh mi .
Since D(fih ) = D(fi ), the D(fih ) again cover X, so the fih generate the unit ideal. We may
now pick g1 , . . . , gn R such that g1 f1h + + gn fnh = 1. Put
m = g 1 m1 + + g n mn .
We then have
fih m =
fih gj mj =
fjh gj mi = mi ,
so m is an element of M restricting to mi /fih for each i. This completes the proof of the
glueing property, so we are done.
4
Schemes
From now on, we view X = Spec(R) as a ringed space with structure sheaf OX as dened
above. Note that for any prime ideal p of R, the stalk OX,p is canonically isomorphic to the
local ring Rp (the localization of R at the multiplicative set R \ p). Hence Spec(R) is in fact
a locally ringed space.
At this point, we make schemes out of prime spectra by glueing, just as we would make
manifolds out of open subspaces of Rn . We dene an ane scheme to be any locally ringed
space X isomorphic to Spec(R) for some ring R; note that the ring R is uniquely determined
by the fact that
(Spec(R), OSpec(R) ) = R
(from the previous theorem). A scheme is a locally ringed space in which each point has an
open neighborhood isomorphic to an ane scheme.
Warning: if X = Spec(R) is an ane scheme, each distinguished open subset D(f ) is
an ane scheme, namely Spec(Rf ) (exercise). By construction, these form a basis of open
sets. However, it is possible for there to be an open subset U of X such that (U, O X |U ) is
isomorphic to an ane scheme but U is not distinguished. (Counterexample to appear as
an exercise.)
Schemes by glueing
We often specify nonane schemes using glueing data. For instance, if X1 and X2 are two
schemes admitting open subsets U1 , U2 which are isomorphic as locally ringed spaces, we can
glue along this isomorphism to get a third scheme X. For more than two schemes, though,
we must add a cocycle condition to keep the glueing maps consistent. Here is how that
works.
Let us rst specify glueing data for sets. Let (Xi )iI be a collection of sets. For each pair
(i, j) I I, let Uij be an open subset of Xi , and suppose that Uii = Xi . Let ij : Uij Uji
be an isomorphism, and suppose that ii = idXi . Suppose also that for i, j, k I, ij restricts
to an isomorphism of Uij Uik with Uji Ujk , and the cocycle condition
ik = jk ij
holds on Uij Uik . (In particular, ji = 1
ij .)
We would like to identify the Xi with subsets of a single set X in such a way that Uij
identies with Xi Xj and ij
identies with the identity map on Xi Xj . To do this, rst
form the disjoint union X = iI Xi . Then dene a binary relation on X as follows: for
xi Xi and xj Xj , we say that xi xj if xi Uij , xj Uji , and ij (xi ) = xj . The glueing
conditions guarantee that this is an equivalence relation, so we may form the quotient X of
X by ; this gives the desired glueing. (Exercise: reformulate this denition in terms of a
limit construction.)
We next specify glueing data for topological spaces. Set notation as above, except that
each Uij must be an open subset of Xi , and each ij must be a homeomorphism. Using the
5
glueing construction for sets, identify the Xi with subsets of a single set X. We may then
use the topologies on the Xi as a basis for a topology on X; in particular, Xi is open in X.
We must still check, however, that the given topology on Xi coincides with the subspace
topology from X (it is only obvious that the subspace topology is ner). Suppose x i Xi
and V is an open neighborhood of xi in X. There then exists some j such that xi Xj and
V contains an open neighborhood of xi for the topology on Xj . Since xi Xi Xj = Uji and
the latter is open in Xj , V Uji also contains an open neighborhood of xi for the topology
on Xj . Since ij is a homeomorphism, V Uji = V Uij contains an open neighborhood of
xi for the topology on Xi . This proves the claim.
We next specify glueing data for (locally) ringed spaces. Set notation as above, except
that each Xi now carries a structure sheaf OXi , and each ij is an isomorphism of (locally)
ringed spaces. Using the glueing construction for topological spaces, identify the Xi with
open subsets of a single topological space X. By the glueing property for sheaves, we now
obtain a sheaf of rings OX , so X may be viewed as a ringed space. Moreover, for x Xi , we
have a canonical identication of OX,x with OXi ,x ; hence if each Xi is a locally ringed space,
so is X.
We nally specify glueing data for schemes. This is the easy part: set notation as above,
except that each Xi is a scheme. Then it is evident that X is also locally isomorphic to an
ane scheme, so X is a scheme! (This part also works for manifolds and the like.)
Examples of glueing
Glueing can be a force for both good and evil. Lets start with good.
Start with any ring R. For i = 0, . . . , n, put
Xi = Spec R[x0 /xi , . . . , xi1 /xi , xi+1 /xi , . . . , xn /xi ].
Dene the distinguished open subset
Uij = D(xj /xi ) Xi .
Then there is an obvious isomorphism of Uij with Uji given by identifying xk /xi with
(xk /xj )(xj /xi ). It is easy to check that the cocycle condition is satised, so we get a scheme
PnR , the projective space over R. (An alternate construction of projective space uses graded
rings. More on this later.)
Now for the evil. Let k be an algebraically closed eld. Let X1 and X2 be two copies of
Spec k[x]. We may glue these on the open sets obtained by removing the point x = 0 (i.e.,
the distinguished opens D(x)) to get a rather unpleasant object; it is a line with a doubled
point.
We would like to formulate a condition that rules out such pathologies. In topology,
the Hausdor condition does the job, but that wont work for schemes. We need a more
category-theoretical notion, which will be provided once we dene separatedness.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Morphisms of schemes (updated 20 Feb 09)
We next introduce morphisms of locally ringed spaces and schemes. Same references as
the previous handout.
Missing remark from last time: when EGA was written, what we now call a scheme was
called a prescheme. EGAs scheme is what we will call a separated scheme.
To dene morphisms of ringed spaces, I need the direct and inverse image functors for
sheaves. I had these on a previous handout, but didnt discuss them in class, so let me
review.
Let f : X Y be a continuous map. For F a sheaf on X, the formula
(f F )(V ) = F (f 1(V ))
obviously denes a sheaf f F on Y . It is called the direct image of F .
Now let G be a sheaf on Y . Dene a presheaf f1 G on X as follows: for U open in X,
let (f1 G)(U) be the stalk of G at f (U), i.e., the direct limit of G(V ) over open sets V X
containing f (U). This is general not a sheaf; its sheacation is called the inverse image of
G, denoted f 1 G. (The notation f is reserved for something else; see below.)
Proposition. The functors f 1 and f form an adjoint pair.
Proof. Exercise.
Using the inverse image, we can dene the restriction of F to an arbitrary subset Z of
X, as the sheaf i1 F for i : Z X the inclusion map (with Z given the subspace topology).
If Z = {x}, this coincides with the stalk Fx (exercise).
Morphisms of (locally) ringed spaces
Let (X, OX ) and (Y, OY ) be ringed spaces. A morphism of ringed spaces from X to Y
consists of a continuous map f : X Y plus a homomorphism f : OY f (OX ) of
sheaves of rings on Y .
For example, if X and Y are manifolds, then f acts as follows. Given an open subset
U of Y , we must specify a homomorphism from OY (U) to f (OX )(U) = OX (f 1 (U)). This
homomorphism is pullback by f ; that is, it takes a continuous function g : U R to the
continuous function g f : f 1 (U) R.
In this example, both X and Y are locally ringed spaces. Moreover, the homomorphism
f has an important property not implied by the denition of a morphism of ringed spaces:
if g : U R vanishes at some point y Y , then g f vanishes at any point x X for which
f (x) = y. More generally, the value of g at y equals the value of g f at x.
1
This is such an important property that we build it into the denition of a morphism
of locally ringed spaces. If (X, OX ) and (Y, OY ) are locally ringed spaces, a morphism of
locally ringed spaces is a morphism of the underlying ringed spaces such that for each point
x X mapping to y Y , the induced homomorphism OY,y OX,x of local rings is a local
homomorphism, that is, the inverse image of mX,x is mY,y .
Morphisms of schemes
A morphism between two schemes is simply a morphism between the underlying locally
ringed spaces. E.g., for U X open, restricting X to U gives a locally ringed space which
is again a scheme, and the inclusion is a morphism of schemes. Such a U is called an open
subscheme of X.
It may be a surprise that merely requiring morphisms of schemes to preserve the locally
ringed space structure has the expected eect.
Theorem 1. For A and B two rings, the set of morphisms (f, f ) : Spec(A) Spec(B) of
locally ringed spaces corresponds bijectively to the set of ring homomorphisms f : B A,
where (f, f ) goes to the map
f (Spec(A)) : (Spec(B), OSpec(B) ) = B (Spec(B), f (OSpec(A) )) = A.
In fact, something more general is true.
Theorem 2. Let LocRingSp be the category of locally ringed spaces. For any locally ringed
space (X, OX ) and any ring A, there is a natural bijection
HomLocRingSp ((X, OX ), Spec(A)) HomRing (A, (X, OX ))
obtained by taking global sections. In other words, the functors Spec and (, O ) from the
category of rings to the opposite category of locally ringed spaces form an adjoint pair.
Proof. We rst dene the inverse map. Given a map f : A (X, OX ) and a point
x X, let f (x) be the point of Spec(A) corresponding to the inverse image of mX,x under
the composition A (X, OX ) OX,x .
To see that f is continuous, it is enough to check that the inverse image of a distinguished
open subset D(g) is open. But this inverse image consists of the points x X where
f (g) (X, OX ) has a nonzero value, and this is indeed open. Better yet, if x X is a
point where f (g) has a nonzero value, then (since OX,x is a local ring) g has a multiplicative
inverse in OX,x , and so has a multiplicative inverse everywhere in some open neighborhood
of X. As a corollary, we observe that g has a multiplicative inverse everywhere on f 1 (D(g))
(since the local inverses are unique, and hence must glue).
Now that f is known to be continuous, we can dene f . It is sucient to dene it on
distinguished opens, that is, we must specify
f (D(g)) : (D(g), OSpec(A) ) = Ag (D(g), f(OX )) = (f 1 (D(g)), OX ).
2
To do this, write any h Ag as a/g i with a A and i Z0 . We can then map a to

(X, OX ) and then by restriction to (f 1 (D(g)), OX ). By the previous paragraph, g maps
to a unit in (f 1 (D(g)), OX ), so we can compute (unambiguously) the image of a/g i.
It is clear that if we start with a ring homomorphism, then pass to locally ringed spaces,
then return, we get back the original ring homomorphism. The hard part is to check that if
we start with a morphism of locally ringed spaces on the left, then go to the right and come
back, we get back the morphism we started with. What we need the extra condition for is
to see that the underlying map on topological spaces is reproduced; once that holds, we get
the equality of homomorphisms of ring sheaves by comparing them on stalks.
Here is a simple example to illustrate why we need morphisms of locally ringed spaces,
rather than ringed spaces. Let R be a discrete valuation ring with fraction eld K, e.g.,
R = Zp and K = Qp . Then Spec(K) consists of a single point (0), while Spec(R) consists
of two points (0) and mR (the maximal ideal) with the rst being closed and the second
not. The inclusion R K of rings corresponds to a map of locally ringed spaces sending
the unique point of Spec(K) to the point (0). However, there is also a map of ringed spaces
sending Spec(K) to the point mR and again using R K to dene the map on structure
sheaves. This is not a morphism of locally ringed spaces because the map R K on stalks
is not a local homomorphism. (For the good morphism, the map on stalks is just the identity
map K K.)
Some strange morphisms to schemes
Given any locally ringed space (X, OX ), we can use the previous theorem to construct a
canonical morphism
(X, OX ) Spec((X, OX ))
(this is basically an adjunction morphism). This in itself may not be so useful, because X
may have very few global functions (e.g., the Riemann sphere with the sheaf of holomorphic
functions). On the other hand, if X contains enough global functions to separate points
(i.e., if for any x, y X we can nd f (X, OX ) with f mX,x but f
/ mX,y ), then the
canonical homomorphism is injective.
For instance, if X is an ane algebraic variety, this gives a map from X to a scheme. It
turns out that this map is a bijection from X to the closed point to the resulting scheme,
and in fact gives an embedding of the category of varieties into the category of schemes; see
Hartshorne Proposition II.2.6 and the related exercise.
Another example occurs when X is a suciently small complex analytic manifold (e.g.,
a Stein space). Such examples will occur when we talk about analytication of complex
algebraic varieties and Serres GAGA principle.
One other funny but useful example: for X any scheme and x X, we can construct
a morphism Spec(OX,x ) X by taking any open ane neighborhood U of x in X and
performing adjunction on the ring map (U, OX ) OX,x . The result does not depend on
U; it carries the closed point of OX,x (the point corresponding to the maximal ideal of the
local ring) to x.
Fibre products
Recall that a bre product of the morphisms Y X and Z X in any category is a limit
of the diagram
Y
Z
X
i.e., a nal object among objects mapping to Y and Z making the diagram commute.
Well construct bre products of schemes in a moment. First, we observe how bre
products interact with passage to open subschemes.
Lemma. Suppose f : Y X and g : Z X are morphisms of schemes such that the bre
product Y X Z exists. Let 1 : Y X Z Y , 2 : Y X Z Z be the induced maps. Let
T X, U Y, V Z be open subsets such that f (U), g(V ) T , viewed as subschemes.
Then
11 (U) 21 (V ),
viewed as a subscheme of Y X Z, is a bre product of U T and V T . (In particular,
the construction does not depend on T .)
Proof. Suppose S U and S V are morphisms such that S U T and S V T
f
g
agree.
Then S U Y X and S V Z X agree, so S factors uniquely
f
through Y X Z. Now writing S U T X as S Y X Z 1 Y X shows
that the image of S in Y X Z lands in 11 (U); similarly, it lands in 21 (V ). So we get a

map S 11 (U) 21 (V ); conversely, any such map can be composed with the inclusion
11 (U) 21 (V ) X , so the above argument shows that the map is unique.
With this, it is easy to check the existence of bre products.
Theorem 3. All bre products exist in the category of schemes.
Proof. The easy part is when X = Spec(A), Y = Spec(B), Z = Spec(C) are all ane. In
that case, the tensor product B A C is a bre coproduct in the category of rings, using the
maps 1 : B B A C and 1 : C B A C.
To get the general case, we apply the previous lemma twice. First, if X is ane, then we
can cover Y and Z with open anes and use the previous lemma to glue the bre products.
Second, once we know bre products exist when X is ane, we can cover X itself with open
anes (and cover Y and Z with the inverse images of these) and use the lemma again.
As noted earlier, this notion of product behaves a bit strangely on the level of sets. For
instance, Spec R and Spec C both contain a single point, but Spec C Spec R Spec C consists
of two points!
4
The functor of points
The previous example illustrates that the set of points of a scheme does not really reect our
geometric intuition, derived largely from our experience with varieties, about the behavior
of points on geometric objects. A good conceptual workaround for this is the functor of
points.
Given two schemes S and X, the set of S-valued points of X, denoted X(S), is simply
the set Hom(S, X) of morphisms of schemes. If S = Spec R, we may write X(R) instead of
X(S). For instance, for any ring R, dene the ane space AnR = Spec R[x1 , . . . , xn ]. Then
for any ring R,
AnZ (R) = AnR (R) = Rn .
A more telling example is the bre product. If Y X and Z X are morphisms, then
Hom(S, Y X Z) = Hom(S, Y ) Hom(S,X ) Hom(S, Z),
where the right side denotes the usual bre product in the category of sets, i.e., you take
pairs of morphisms from S, one to Y and one to Z, which agree when you pass to morphisms
from S to X.
If we x X, we may view X() as a functor on the category of schemes. There is an
appropriate sense in which it is a sheaf on that category, but never mind that for now. The
one thing you might want to take away from this is that if X is covered by open anes Ui ,
a morphism S X may not land in any one of the Ui . For instance, PnZ (S) is not just
obtained by taking the union of the R-valued points of the distinguished open subsets. For
instance, the identity morphism PnZ PnZ doesnt occur this way. You can even have this
problem for ring-valued points: there is a natural map Spec Z[x0 , . . . , xn ] PnZ which does
not factor through a distinguished open. (See exercises.)

The functor of points doesnt by itself prove much of anything; for instance, it doesnt tell
you how to construct the bre product. However, it can be used to suggest certain natural
denitions, e.g., the denition of a group scheme. See exercises.
Zen and the art of base change
Although the bre product is a symmetric construction in the two factors, in algebraic
geometry we will often use it in an asymmetric fashion. Namely, for f : Y X a morphism
and g : Z X another morphism, we refer to f g : Y X Z Z as the base change of f
by g. Geometrically, if you imagine f as giving a family of geometric objects parametrized
by X, f g describes the pullback of this family to Z.
When we start dening properties of morphisms (next lecture), we will be particularly
concerned with their behavior under base change. Typical questions:
(a) Is the property stable under arbitrary base change? If not, how about base changes
where the base change morphism is itself restricted in some way?
5
(b) Does the property descend down a suitable base change? E.g., if g is surjective, does
f g having the property imply the same for f ?
In particular, if (b) is true whenever g : i Ui X is a covering of X by open subschemes,

we say that the property is local on the target. For instance, by our lemma about base change,
the property of being injective/surjective on points is local on the target.
Well give many examples of properties of morphisms later. Here are two to use as a
mental model. For f : Y X a morphism of schemes with X = Spec(A) ane, we say that
f is ane if Y = Spec(B) is also ane. I claim that this property satises the following
condition.
(i) Let f : Y X be a morphism with X ane. Let D(g1 ), . . . , D(gn ) be a nite
covering of X by distinguished open subsets. Then f is ane if and only if the induced
morphisms Y X D(gi) D(gi ) are all ane.
This follows from the assigned Hartshorne exercise II.2.17, since the gi generate the unit
ideal in (Y, OY ). (Well prove something much stronger in the next lecture.)
It can be deduced from this (exercise) that for f : Y X an arbitrary morphism, the
following are equivalent.
(a) For a single open ane cover {Ui }iI of X, each induced morphism Y X Ui Ui is
ane.
(b) For each open ane cover {Ui }iI of X, each induced morphism Y X Ui Ui is ane.
(In other words, for every open ane U X, the induced morphism Y X U U is
ane.)
If these hold, we say that f itself is ane.
In this case, we have an extra condition that is evidently satised.
(ii) Let f : Y X be a morphism with X ane, which is ane. Then for any morphism
g : Z X with Z also ane, f g : Y X Z Z is also ane.
In this case, (a) and (b) are equivalent to this condition (exercise again).
(c) For every morphism g : Z X with Z ane, f g : Y X Z Z is nite.

Moreover, these equivalent conditions are stable under arbitrary base change.
Here is another important example. For f : Y X a morphism of schemes with
X = Spec(A) ane, we say that f is nite if Y = Spec(B) is also ane and B is nite as
an A-module. I claim this satises (i) and (ii). For (i), we already know that Y = Spec(B)
is ane. Suppose B A Agi is generated as an Agi -module by some nite set of elements.
Since each element is an element of B divided by a power of gi , we can generate B A Agi
as an Agi -module with a nite subset of B itself. These subsets together generate a nite
A-submodule B such that (B/B ) A Agi = 0 for each i. That is, the sheaf corresponding
to the A-module B/B is zero; but by a theorem from last time, this forces B/B = 0.
For (ii), note that if Z = Spec(C), then Y X Z = Spec(B A C) and B A C is nite
as a C-module: use a set of generators of B as an A-module.
6
Back to schemes for a moment
The strategy I just introduced can be used to establish properties of schemes, not just
morphisms, using a trick: to dene a property P of schemes, say that a morphism f : Y X
has property P if and only if f is an isomorphism and X has property P .
For instance, an ane scheme is reduced if its corresponding ring A has no nilpotent
elements. This holds if each local ring of A is reduced (exercise), so in particular (i) holds.
This allows us to extend the denition of reducedness to arbitrary schemes, and the resulting
condition holds if and only if each local ring of the scheme is reduced.
Another approach is to recall that each scheme admits a unique morphism to Spec(Z),
and extract properties of schemes from properties of this morphism. Trivial example: X is
ane if and only if X Spec(Z) is ane.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Sheaves of modules (updated 27 Feb 09)
Having discussed sheaves of sets, abelian groups, and rings, we now consider sheaves of
modules over a locally ringed space, with an emphasis on the situation for schemes. Well
then use what we know to talk about closed immersions and separated morphisms. (Many
more properties of morphisms to follow in the next lectures!) References: Hartshorne II.4,
II.5; EGA I 0.4, I 0.5, I 1.1.
Sheaves of modules
Let (X, OX ) be a ringed space. A sheaf of OX -modules is a sheaf of sets F such that for
each U X open, the group of sections F (U) is equipped with a structure of a OX (U)
module, in a fashion compatible with restriction. We sometimes call such a thing simply an
OX -module. A morphism of two OX -modules F G is a morphism of sheaves which gives
a OX (U)-module homomorphism F (U) G(U) for each U X open.
Fun fact: for F a sheaf of OX -modules, there is a natural bijection between
Hom(OX , F ) F (X).
If F is a sheaf of OX -modules, then for each x X, the stalk Fx inherits a structure
of a OX,x -module. Using that, we can talk about submodules, quotient modules, kernels,
cokernels, images, and the like; in fact, these agree with the corresponding notions in the
category of sheaves of abelian groups.
I
An OX -module is free if it is isomorphic to a direct sum OX
of copies of OX . This
property cannot be checked locally; if OX is only locally isomorphic to a free OX -module,
we say it is locally free. More on locally free sheaves later.
One new operation is the tensor product, but it requires some care. If F , G are two
OX -modules, the presheaf
U F (U) OX G(U)
may not be a sheaf. Form its sheacation, and call that F OX G; it has the expected
arrow-theoretic behavior. (I might forget the OX sometimes when it is not ambiguous.)
If f : (X, OX ) (Y, OY ) is a morphism of ringed spaces, and F is a OX -module, then f F

may naturally be viewed as a f OX -module. Using the map f : OY f OX , we may also
give f F the structure of an OY -module. We call this again the direct image of F .
On the other hand, if G is a sheaf of OY -modules, then f 1 G is an f 1 OY -module.
Adjointness turns f into a homomorphism f 1 OY OX , so we can form the tensor product
G f 1 OY OX
1
and get a OX -module. We notate this f G and call it the (module-theoretic) inverse image
of G under f .
Again, f and f are adjoint in the obvious fashion. Statement and proof left to the
reader.
Quasicoherent sheaves of modules
Since thinking about ane schemes is supposed to be equivalent to thinking about rings (after
all, the two categories are equivalent!), we would like our thinking about sheaves of modules
on ane schemes to be equivalent to thinking about modules over rings. We even know
what the functors realizing this equivalence should be: on an ane scheme X = Spec(R),
(which we already proved is a
we should go from R-modules to OX -modules via M M
sheaf!), and back via F (X, F ).
While it is clear that the composition one way is naturally isomorphic to the identity
functor on R-modules, the other way fails because there are too many OX -modules. For
instance, for R = k[x] with k an algebraically closed eld, we can make an OX -module F by
declaring
0
(0)
/U
F (U) =
k[x] (0) U
and putting in the obvious restriction maps. This sheaf has the same global sections as OX
itself but is clearly not the same!
The x for this is a bit heavy-handed: we simply declare that we only want sheaves
which locally come from a module over a ring. For schemes, it is clear what this means:
we want F to have the property that for each x X, there is an open ane neighborhood
for some A-module M. For locally ringed spaces,
U = Spec(A) of x in X such that F |U = M
we must be a bit more careful: we want F to locally be given as the cokernel of some module
I
J
homomorphism OX
|U OX
|U between free OX -modules.
It still remains to check that this gives what we expect for ane schemes. I call this the
third fundamental theorem of ane schemes.
Theorem 1. Let F be a quasicoherent sheaf of OX -modules for X = Spec(R), and put

M = (X, F ). Then the natural homomorphism M
= F of OX -modules is in fact an
isomorphism. In other words, the category of quasicoherent OX -modules on Spec(R) is
equivalent to the category of modules on R.
Proof. The claim is equivalent to the fact that for each prime ideal p of R, the natural map
p Fp is a bijection. Since F is quasicompact, we can nd a distinguished open
Mp
=M
for some Rf -module N. The map M
F induces a map
D(f ) of X on which F
= N
|D(f ) F |D(f )
Taking global sections gives a homomorphism Mf
Mf
= M
= N.
= N of
Rf -modules.
We check injectivity of Mf N. Suppose m/f h Mf maps to zero in N. For each
prime ideal q of R, we can nd a distinguished open neighborhood D(g) of q in X such that
2
F |D(g)
= P for some Rg -module P . Now Pf
= Ng
= Mf g as Rf g -modules since they all
give rise to the same sheaf. Hence the image of m/f h in Pf is zero, so the image of m in
P = (D(g), F ) is killed by some power of f . We conclude that for some nonnegative integer
j, f j m restricts to the zero section of F on D(g). Since this holds for each q, and we only
need nitely many D(g) to cover X (since X is quasicompact), there exists a nonnegative
integer j such that f j m = 0 in (X, F ) = M. Hence m/f h represents the zero element in
Mf .
We check surjectivity of Mf N. Let n N be any class. Cover X with nitely
many D(gi) on each of which F |D(gi ) is represented by a module Pi over Rgi . Then FD(f gi )
is represented by (Pi )gi . Hence for some j, f j n is the restriction to D(f gi) of a section si
of F over D(gi ). We may enlarge j so that it works for all i; then si sj represents the
zero section of F over D(f gigj ). By the previous paragraph, we can nd some k such that
f k (si sj ) is the zero section of F over D(gi gj ). Ergo, for some very large j, the f j n give
sections si of F over D(gi) which glue to give a section of F on X itself.
We now have Mf
= N, so Mp
= Np
= Fp. Since p X was arbitrary, this proves
M = F.
Relative Spec, ideal sheaves, and closed immersions
Let X be a scheme, and let F be a quasicoherent sheaf of OX -modules which also carries an
OX -algebra structure (or for short, a quasicoherent OX -algebra ). Then for each open ane
subscheme U of X, we can form the scheme Spec (U, F ), and these glue. We thus obtain
a scheme Y = Spec F which comes with a morphism Y X. This is called the relative
spectrum of F .
One important class of examples of relative spectra are closed immersions. (For more
examples, see exercises.) An ideal sheaf on a locally ringed space X is a quasicoherent
subsheaf of OX . For I an ideal sheaf, the quotient OX /I is a quasicoherent OX -algebra.
Let Z be its relative Spec; any map arising as such a map f : Z X is called a closed
immersion.
Let us see why this name is tting in the case of schemes. Say X = Spec R. Then an
ideal sheaf corresponds to an ideal I of R, and Z = Spec(R/I). (Notice that this means
that any closed immersion is an ane and nite morphism!) The points of Z are in bijection
with the points p X where the stalks Rp and Ip dier, which is precisely the vanishing set
V (I). But there can be many dierent closed immersions with the same underlying set! For
instance, in Spec k[x, y], the ideals (x) and (x2 ) dene the same closed set but not the same
closed immersion.
Beware that algebraic geometers have the habit of calling the source of a closed immersion
a closed subscheme of X even though its not really a subscheme of X in any precise sense.
But this isnt so misleading because the map Z X is indeed a monomorphism. And there
is something comforting in the thought that (looking at our previous example) the x, y-plane
contains a doubled line dened by x2 = 0, which in turn contains the undoubled line
dened by x = 0.
3
Finally, note that if f : Z X is a closed immersion, the dening ideal sheaf can be
recovered as the kernel of f : OX f OZ . In fact, following Hartshorne, you may dene
a closed immersion as a map of schemes f : Z X which induces a homeomorphism of Z
with a closed subset of X, such that f : OX f OZ is surjective. This has the advantage
that it is clear that the property of being a closed immersion is local on the target (though
you can check it the other way too, since the ideal sheaf is uniquely determined).
Separated schemes and morphisms
We are now ready to introduce the analogue of the Hausdor property for schemes. However,
it is more natural to introduce it for morphisms of schemes.
A morphism f : Y X is separated if the diagonal morphism : Y Y X Y is a
closed immersion. Since the formation of , and the property of it being a closed immersion,
are local on the target, so is the notion of being separated.
For instance, if X = Spec k for k a eld, and Y consists of two copies of Spec k[x] glued
along Spec k[x, x1 ], then Y X Y is the ane plane with two copies of each axis, and four
copies of the origin. The image of contains two copies of the origin, but its closure contains
all four. Hence Y X is not separated.
Lemma. A morphism f : Y X is separated if and only if the image of the diagonal
: Y Y X Y is a closed subset of Y X Y .
Proof. See Hartshorne, Corollary II.4.2. It relies on another useful (but easy) fact: any ane
morphism is separated.
We say a scheme itself is separated if its unique morphism to Spec(Z) is separated. This
means that by at Spec(Z) is itself separated, but this seems reasonable enough, especially
in light of the following.
Theorem 2. Let X be a separated scheme. Then the intersection of any two open ane
subschemes of X is again ane.
Proof. Exercise. (Beware that the converse is false. The nonseparated example above also
satises this condition.)
There is a valuative criterion for separatedness, but it is hardly ever useful (because
it involves arbitrary valuation rings, which can be rather nasty); see Hartshorne Theorem
II.4.3.
Separatedness and base change
Theorem 3. Separatedness is stable under base change.
Proof. Let f : Y X be a separated morphism, and let Z X be an arbitrary morphism.

We are supposed to check that the diagonal
Y X Z (Y X Z) Z (Y X Z)
is a closed immersion. On the other hand, since closed immersions are stable under base
change, we already know that
Y X Z (Y X Y ) X Z
is a closed immersion. It thus suces to identify the two right sides in a way commuting
with the arrows.
The way to see this is to use the functor of points: given W X, we must identify the
maps into the two right hand sides in a natural way. The maps to (Y X Y ) X Z give
pairs of maps W Y X Y and W Z which agree on X, then (by splitting the rst bre
product) gives triples of maps W Y , W Y , and W Z which all agree on X.
The maps to (Y X Z) Z (Y X Z) similarly give pairs of maps W (Y X Z),
W Y X Z which agree on Z. Splitting again, we get quadruples W Y , W Z,
W Y , W Z which agree on X, but moreover the two maps W Z must be the same
map. We thus identify with the previous description.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

More properties of morphisms (updated 5 Mar 09)
Note that nite presentation is not discussed in EGA 1; see EGA 4.1 instead.
More about separated morphisms
Lemma. The composition of closed immersions is a closed immersion.

Proof. Let f : X Y and g : Y Z be closed immersions. Since the property of being
a closed immersion is local on the base, we may assume Z = Spec(A) is ane. Then
Y = Spec(B) for B a quotient of A, so X = Spec(C) for C a quotient of B. Hence C is
a quotient of A, proving the claim. (A similar argument shows that a composition of nite
morphisms is nite.)
Lemma.
(a) Any closed immersion is separated.
(b) A composition of separated morphisms is separated.

(c) Separatedness is stable under base change.
(d) A product of separated morphisms is separated.
(e) If f : X Y and g : Y Z are morphisms, g f is separated, and g is separated,
then f is separated.
(f ) If f : X Y is separated, then fred : Xred Yred is separated.
Proof. We know (a) because closed immersions are ane and ane morphisms are separated.
We know (c) from the previous handout. Parts (d)-(f) follow once we also have (b); see
exercises.
It remains to prove (b). Let f : X Y and g : Y Z be separated morphisms. Then
X Y X maps to X Z X; in fact, this morphism is the base change of the closed immersion
: Y Y Z Y by f f : X Z X Y Z Y . (To check this: use functor-of-points to
reduce to the analogous assertion for sets. This can be checked with Z equal to a singleton
set, so we just want to know that for a morphism of sets X Y , the bre product of Y
and X X over Y Y equals X Y X. This is obvious.) Hence X Y X X Z X
is a closed immersion. Since the composition of closed immersions is a closed immersion
(previous lemma), we nd that X X Y X X Z X is a closed immersion.
Quasicompact morphisms
A morphism f : Y X with X ane is quasicompact if Y is quasicompact as a topological

space. This denition satises the strong collater (exercise), so we get a notion which is local
on the base and stable under base change.
Exercise. Any ane morphism is quasicompact.
1
Finite type and nite presentation
Let A be a ring. Recall that an A-algebra B is nitely generated if it is of the form

A[x1 , . . . , xn ]/I for some nonnegative integer n and some ideal I of A[x1 , . . . , xn ]. If I can be
chosen to be a nitely generated ideal, we say that B is nitely presented; this is of course
automatic if A is noetherian (as it will be in most of our examples).
Let f : Y X be a morphism of schemes with X = Spec(A) ane. We say f is locally
of nite type/presentation if Y is a union of open subschemes, each of the form Spec(B) with
B a nitely generated/presented A-algebra. If only nitely many such open subschemes are
needed, we say f is of nite type/presentation. These satisfy the strong collater (exercise).
If f : Y X is of nite type, we sometimes say that Y is of nite type over X. Similarly
for the other denitions.
Obvious: any nite morphism, including any closed immersion, is of nite type.
Exercise. A morphism f : Y X is of nite type/presentation if and only if it is quasicompact and locally of nite type/presentation.
Algebraic varieties
We can now give a scheme-theoretic rendition of the theory of abstract algebraic varieties,
in the sense of 18.725. (But see below.)
Let k be an algebraically closed eld. An ane variety is a locally ringed space dened by
some data of the following form. Pick a nonnegative integer n and an ideal I of k[x1 , . . . , xn ],
and put X = V (I). Equip X with the Zariski topology, i.e., take a basis of open sets of
the form D(g) = {x X : g(x) = 0} for g k[x1 , . . . , xn ]. Dene a regular function on
an open subset U of X to be a function h : U k such that for each x U, there exist
f, g k[x1 , . . . , xn ] and a nonnegative integer m such that g vanishes nowhere on U while
g m h f vanishes identically on U . Then the regular functions on U form a sheaf.
In the context of schemes, we interpret X to be the set of maximal ideals in Spec(A)
for A = (k[x1 , . . . , xn ]/I)red , equipped with the structure of a locally ringed space given by
restriction from Spec(A).
Now recall that an abstract algebraic variety is a locally ringed space covered by ane
varieties.
Theorem 1. The category of abstract algebraic varieties over the algebraically closed eld
k is equivalent to the category of schemes which are reduced and locally of nite type over
Spec(k).
Proof. Exercise. The key point is to check that if X = Spec(A) and Y = Spec(B) for A, B
two reduced nitely generated k-algebras, then the morphisms from X to Y are the same
as the morphisms of the corresponding algebraic varieties. But that is because they both
correspond to ring homomorphisms B A.
Beware that there is no universal denition of algebraic varieties, because everyone seems
to prefer to add additional hypotheses. For instance, Hartshorne (see Chapter I) forces
his varieties to be separated (as often do I). Some authors also force their varieties to be
irreducible, i.e., not admitting two disjoint open subschemes. And so on.
Proper morphisms
We would like to have an algebraic analogue of the notion of a compact algebraic variety
over the complex numbers. For this, we introduce the notion of properness.
A morphism f : Y X of schemes is proper if it is separated, of nite type, and
universally closed. The latter means that any base change of f is a closed map of topological
spaces (i.e., carries closed sets to closed sets); this condition comes from the notion of a
proper map of topological spaces (see exercises). Since these properties are all local on the
base and stable under base change (the last one by at), properness is also.
The denition of properness is rather hard to check. One easy case: a closed immersion
is separated (because its ane), of nite type (obvious), and universally closed (because
any base change is still a closed immersion, so has closed image), so is proper. Besides this
example, and the following slightly fancier example...
Exercise. Any nite morphism (including any closed immersion) is proper.
... all examples of properness will ultimately be extracted from the following theorem.
Theorem 2. The morphism f : P
nZ Spec Z is proper.
Hartshorne proves this using the valuative criterion for properness (under a somewhat
mysterious noetherian hypothesis). Ill ultimately prove this following EGA, but I need to
wait until the next lecture so I can say more about projective spaces in the interim. I will
point out now that the fact that f is of nite type is evident from the glueing construction,
and the separatedness may be obtained by describing the diagonal : P
nZ P
nZ Spec Z PnZ
explicitly (exercise).
As for separated morphisms, we have some properties.
Lemma.
(a) Any closed immersion is proper.
(b) A composition of proper morphisms is proper.

(c) Properness is stable under base change.
(d) A product of proper morphisms is proper.
(e) If f : X Y and g : Y Z are morphisms, g f is proper, and g is separated, then
f is proper.
(f ) If f : X Y is proper, then fred : Xred Yred is proper.
3
Proof. Again, (d)-(f) follow from (a)-(c). We already observed (a) and (c). To check (b), we
already checked that separatedness composes. Finite type composes by an argument similar
to the proof that closed immersions compose. Universal closedness composes because a
composition of closed maps of topological spaces is again closed.
Corollary. Any morphism f : X Y that factors as a closed immersion of X into P
nY =
P
nZ Spec Z Y followed by the projection PnY Y is proper.
The converse is not true even over C, as there are compact algebraic varieties which are
not closed subvarieties of any projective space. See the appendices to Hartshorne for an
example. One can often deal with these using Chows lemma, about which more later.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Projective morphisms, part 1 (updated 3 Mar 08)
We now describe projective morphisms, starting over an ane base.
Proj of a graded ring
The construction of Proj of a graded ring was assigned as an exercise; let me now recall the
result of that exercise.
Let S =
n=0 Sn be a graded ring, i.e., a ring such that each Sn is closed under addition,
and Sm Sn Sm+n . An element of Sn is said to be homogeneous of degree n; the elements of
S0 form a subring of S, and each Sn is an S0 -module. (One could also dene a graded ring
to allow negative degrees; on the few occasions where Ill need that construction, Ill call it
a graded ring with negative degrees.) Let S + denote the ideal
n=1 Sn .
Let Proj S be the set of all homogeneous prime ideals of S not containing S+ . For each
positive integer n and each f Sn , we may view the localization Sf as a graded ring with
negative degrees, by placing g/f k in degree m kn whenever g Sm . We may then identify
the set
D(f ) = {p Proj S : f
/ p}
with Spec Sf,0 , where Sf,0 is the degree zero subring of Sf . These glue to equip Proj S with
the structure of a scheme (note that D(f ) D(g) = D(f g)). In the case S = A[x0 , . . . , xn ]
where each of x0 , . . . , xn is homogeneous of degree 1, this simply produces the projective
space PnA .
Any morphism S T of graded rings induces a morphism Proj T Proj S of schemes.
For example, we say an ideal I of S is homogeneous if as abelian groups we have
I =
n=0 (I Sn ).
In other words, if we split each element of I into homogeneous components, the components
themselves belong to I. Then S/I may also be viewed as a graded ring, the projection
S S/I induces a morphism Proj S/I Proj S, and this morphism is a closed immersion
(as we see immediately by checking on a D(f )).
Beware that the scheme Proj S does not by itself determine the graded ring S. For
instance, omitting S1 gives another graded ring with the same Proj. Well come back to this
point later.
More generally, if M =
n= is a graded S-module, i.e., Sm Mn Mm+n for all m, n,
on Proj S by doing so on each D(f ) (using
we can convert M into a quasicoherent sheaf M
the degree-zero subset of Mf ) and then glueing. For a converse, see below.
The sheaf O(1)
For S a graded ring, n a nonnegative integer, and M a graded S-module, let M(n) denote
the shifted module
M(n)i = Mn+i .
Let OX (n) be the quasicoherent sheaf on X = Proj S dened by the graded module S(n).
In particular, OX (0) = OX . More generally, for any quasicoherent sheaf F of OX -modules,
put F (n) = F OX OX (n).
Lemma. Suppose that S is generated by S1 as an S0 -algebra. Then the sheaves OX (n)
on Proj S are locally free of rank 1, and OX (m) OX OX (n) is canonically isomorphic to
OX (m + n).
Proof. See Hartshorne, Proposition II.5.12.
Note: a quasicoherent sheaf F on a locally ringed space X which is locally free of rank
1 is also called an invertible sheaf. That is because there is a unique sheaf F such that
F OX F
= OX , the dual of X (exercise). In this case, the dual of OX (n) is in fact OX (n).
This gives us an explanation for what x0 , . . . , xn are on the projective space Proj A[x0 , . . . , xn ]:
they are global sections not of the sheaf OX , but rather of the sheaf OX (1).
Theorem 1. Suppose that S is nitely generated by S1 as an S0 -algebra. Then each quasi for a canonical choice of M.
coherent sheaf on Proj S can be written as M
The nitely generated hypothesis is needed to ensure that Proj S is quasicompact; we
will impose it pretty consistently hereafter.
Proof. Let F be a quasicoherent sheaf on M. Then the module we want is
(F ) = nZ (X, F (n)),
where
F (n) = F OX OX (n).
For the rest of the proof, see Hartshorne, Proposition II.5.15.
Beware that this does not imply that S =
n=0 (X, OX (n)) in general. For a stupid
example, take S = A[x], in which case the sheaves OX (n) are all free and so (X, OX (n)) =
6 0
even when n < 0. For less stupid examples, see Hartshorne exercise II.5.14. However, the
following is true.
Lemma. Let n 1 be an integer. For S = A[x0 , . . . , xn ] with the usual grading (by total
degree), we have
S =
n=0 (X, OX (n)).
Proof. Exercise, or see Hartshorne Proposition II.5.13.
2
Closed subschemes of projective spaces
Proposition. For n 1, any closed immersion into PnA is dened by some homogeneous
ideal of A[x0 , . . . , xn ].
Proof. In fact, there is a canonical way to pick out the ideal. Let I be the ideal sheaf dening
the closed immersion; then (I) is an ideal of (OX ), but we already identied the latter
with S = A[x0 , . . . , xn ]. (This identication uses the fact that S is nitely generated by S1
as an S0 -algebra, in order to invoke the previous theorem. In fact, it is part of the proof of
that theorem; see Hartshorne Proposition II.5.13.)
In general, there may be multiple homogeneous ideals dening the same closed subscheme
of PnA . If we start with an ideal I, pass to the closed subscheme, then use the previous
proposition to get back, we get the saturation of I, namely, the set of all elements f
A[x0 , . . . , xn ] such that xj0 f, . . . , xjn f I for some nonnegative integer j. We thus obtain a
one-to-one correspondence between closed subschemes of PnA and saturated (equal to their
saturation) homogeneous ideals.
Corollary. For n 1, let I be a homogeneous ideal of S = A[x0 , . . . , xn ]. The following
conditions are equivalent.
(a) The subscheme of PnA dened by I is empty.
(b) The saturation of I equals S + .
(c) For some n0 , we have Sn I for all n n0 .
Proof. We just proved the equivalence of (a) and (b). It is clear that (c) implies (b). Let
us check that (b) implies (c). Given (b), each f {x0 , . . . , xn } has the property that
xj0 f, . . . , xjn f I for some j. In particular, we have xj0 , . . . , xjn I for some j. This in turn
implies S(n+1)(j1)+1 I since each monomial of degree (n + 1)(j 1) + 1 is divisible by one
of xj0 , . . . , xjn (pigeonhole principle!).
Projective implies proper
We are now ready to complete the proof that f : PnZ Spec Z is proper. Recall that
the missing step was to show that f is universally closed, i.e., for any scheme X, the map
PnX X is closed. It is enough to do this in case X = Spec A is ane. Moreover, we may
assume n 1, as the case n = 0 is stupid (because f is an isomorphism).
Let Z be a closed subset of PnX , suppose z X is not in the image of Z, and put k = (z).
We must exhibit an open neighborhood U of x in X such that Z PnU = . Let I =
n=0 In
be the saturated homogeneous ideal in S = A[x0 , . . . , xn ] dening Z. Then I A k denes
the empty subscheme of Proj k[x0 , . . . , xn ], but may not be saturated. Nonetheless, for some
m, we have that In A k = Sn A k, and so (Sn /In ) A k = 0.
3
Since Sn /In is a nitely generated A-module, by Nakayamas lemma, (Sn /In ) A Ap = 0

for p the prime ideal of A dening z. Again since Sn /In is nitely generated, we have
(Sn /In ) A Ag = 0 for some g A \ p. Then z D(g) and D(g) is disjoint from the image
of Z, proving the claim.
What is a projective morphism?
Several authors (Hartshorne, Eisenbud-Harris) dene a morphism f : Y X to be projective

n
if it is the composition of a closed immersion Y PnX with the projection PX
for some
nonnegative integer n. This denition is evidently stable under base change, but it is not
local on the base! Better to say that such a morphism is globally projective, and to say that f
is locally projective if each x X admits an open neighborhood U such that f : Y X U U
is globally projective.
This is not such a serious distinction in practice, as globally projective equals locally
projective if X is not too large. For instance, this occurs if X is itself globally quasiprojective
over an ane scheme. (A morphism is globally quasiprojective if it factors as an open
immersion followed by a globally projective morphism. Again, this is stable under base
change but not local on the base; the version where we force locality on the base is a
quasiprojective morphism.)
The denition of projective given in EGA is in fact somewhere between locally and glob
ally projective. More on that later. (Warning: Eisenbud-Harris claim that locally projective
and projective are the same. They arent, but counterexamples are rather pathological.)
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Projective morphisms, part 2 (updated 7 Mar 09)
I particularly recommend Eisenbud-Harris for the material in this section; they give a
complete description of the relationship between the two descriptions of a blowup (using
charts versus relative Proj).
Relative Proj
Let X be a scheme. Let S =

n=0 Sn be a graded quasicoherent OX -algebra. For each open
ane U in X, we can form the morphism Proj S(U) U; these glue to give a morphism
Proj S X. The object Proj S is called the relative Proj of S.
We will hereafter assume that S1 is nitely generated (remember that this is a local
notion), and that S is (locally) generated by S1 as a S0 -algebra. We also assume that S0 is a
quotient of OX ; Hartshorne assumes that in fact S0 = OX , but Id rather not do that. Pick
(m+1)
an open ane subset U of X and a surjection OU
S1 |U ; we then obtain a surjection
(m+1)
Sym OU
S|U , where
(m+1)
Sym OU
(m+1)
n
=
n=0 Sym OU
(m+1)
This in turn gives a closed immersion Proj S Proj Sym OU

, and the latter is nothing
n
but the projective space PU . Consequently, Proj S X is locally projective.
We say that a morphism f : Y X is projective if and only if it occurs as a relative Proj.
This implies locally projective and is implied by globally projective, but does not coincide
with either.
Following EGA, we write P(F) for Proj Sym F whenever F is a nitely generated qua
sicoherent OX -module.
Very ample sheaves
An immersion is a morphism f : Y X of schemes which on topological spaces is an

isomorphism of Y with a locally closed subset (i.e., a closed subset of an open subset) of X,
such that for each y Y mapping to x X, the map f : OX,x OY,y is surjective. Any
composition of closed immersions and open immersions is an immersion; conversely, if f is
an immersion, then it can be written as a closed immersion followed by an open immersion.
(Let U be an open subset of X in which Y is closed; then f factors uniquely through the
open immersion U X, and the resulting map Y U satises Hartshornes denition of
a closed immersion.)
Let f : Y X be a morphism. A quasicoherent sheaf F on Y is very ample relative to f
if there exists an immersion Y P(S1 ) for some nitely generated quasicoherent OX -module
S1 , under which F occurs as the pullback of O(1). Unlike the denition of projectivity, this
notion is indeed local on the base.
1
Lemma. The morphism f : Y X is projective if and only if f is proper and there exists
a very ample sheaf relative to f .
Proof. See Hartshorne, Remark II.5.16.1.
The very ample sheaf pulled back from O(1) can be used to retrieve the morphism to
P(S1 ). Namely, if S1 is globally nitely generated, any set of generators pull back to sections
of the pullback of O(1), and those dene a morphism to projective space. See Hartshorne
Theorem II.7.1.
Blowups
Here is a neat class of examples of relative Proj. Let I be a nitely generated ideal sheaf on
the scheme X, and put Y = P(I). We call Y the blowup of X along I.
For example, say X = Spec k[x, y] for k a eld, and let I be the ideal sheaf dened by
(x, y). Over U = D(x) D(y), we have an isomorphism I|U
= OU , so Y X U U is an
isomorphism. But the bre over the origin looks like a projective line with homogeneous
coordinates x, y.
The blowup dened by I carries less information than I itself. For instance, for any
locally principal ideal sheaf, the blowup dened by I is the identity.
Here is a special property of the blowup. For f : Y X a morphism and I an ideal sheaf
on X, we may compose the inclusion I OX with f : OX f (OY ) and then perform
adjunction to get f I OY . The image is an ideal sheaf on Y , called the inverse image
ideal sheaf of I under f .
Theorem 1. If f : Y X is the blowup dened by the nitely generated ideal sheaf I on
X, then the inverse image ideal sheaf of I on Y is locally principal.
Proof. Recall that Y = Proj S for Sn = Symn I. In this notation, the inverse image ideal
sheaf of I on Y is simply OY (1), which is locally free. This proves the claim.
In fact, f is universal for this property: any morphism Z X such that the inverse
image ideal sheaf of I on Z is locally principal factors uniquely through f (Hartshorne,
Proposition II.7.14).
More concrete description of the standard example: the blowup of Spec k[x, y] at (x, y)
is covered by the two charts
Spec k[x, y/x]
Spec k[y, x/y]
glued along Spec k[x, y, x/y, y/x]. In fact, any blowup can be described analogously: the
blowup of Spec A along I = (r0 , . . . , rm ) is covered by m + 1 charts, a typical one of which is
Spec A[t1 , . . . , tm ](t1 r0 r1 , . . . , tm r0 rm ).
The point is that the inverse image ideal sheaf is supposed to become locally principal, so you
must force one of the generators r0 , . . . , rm to divide into the other ones, and the dierent
2
choices for which generator will divide into the others produces the dierent charts. (Explicit
description of the other charts and the glueing is left to the reader.)
A blowup is a special example of a modication. The latter is a morphism f : Y X
of schemes which is proper, surjective, and birational (i.e., its restriction to an open dense
subset of X is an isomorphism). In the case of a blowup dened by an ideal sheaf, we
get an isomorphism over the complement of the closed set dened by the ideal. In fact,
under certain circumstances, every modication can be written as a blowup; see Hartshorne
Theorem II.7.17. The catch is that the ideal sheaf is not unique; for example, on Spec k[x, y],
the ideals (x, y) and (x2 , xy, y 2) dene the same blowup.
Chows lemma
One use of modications is to turn proper schemes (over some base) into projective schemes.
Theorem 2 (Chows lemma). Let f : X S be a morphism of nite type. Assume that
either S is noetherian, or S is quasicompact and X has nitely many irreducible components.
Then there exists a quasiprojective S-scheme X and a projective surjective morphism f :
X X which restricts to an isomorphism over some open U X such that f 1 (U)
= U is
dense in X . Moreover, if X is reduced/irreducible/integral, we can ensure that X is also.
See EGA 2, Lemme 5.6.1, or for a weaker result, Hartshorne exercise II.4.10.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

More properties of schemes (updated 9 Mar 09)
Ive now spent a fair bit of time discussing properties of morphisms of schemes. How
ever, there are a few properties of individual schemes themselves that merit some discussion
(especially for those of you interested in arithmetic applications); here are some of them.
Reduced schemes
I already mentioned the notion of a reduced scheme. An ane scheme X = Spec(A) is

reduced if A is a reduced ring (i.e., A has no nonzero nilpotent elements). This occurs if
and only if each stalk Ap is reduced. We say X is reduced if it is covered by reduced ane
schemes.
Lemma. Let X be a scheme. The following are equivalent.
(a) X is reduced.
(b) For every open ane subsheme U = Spec(R) of X, R is reduced.
(c) For each x X, OX,x is reduced.
Proof. A previous exercise.
Recall that any closed subset Z of a scheme X supports a unique reduced closed subscheme, dened by the ideal sheaf I which on an open ane U = Spec(A) is dened by the
intersection of the prime ideals p Z U. See Hartshorne, Example 3.2.6.
Connected schemes
A nonempty scheme is connected if its underlying topological space is connected, i.e., cannot
be written as a disjoint union of two open sets. (The empty scheme is not connected.)
Lemma. The scheme X is connected if and only if the idempotent elements of (X, O X )
(i.e., the solutions of e = e2 ) are 0 and 1.
Proof. If X is a disjoint union of open sets U and V , then we can construct an idempotent
e = 0, 1 by taking the pullback of 0 along U Spec Z and the pullback of 1 along V
Spec Z. Conversely, if e (X, OX ) is an idempotent, then its value at each x X is either
0 or 1; the sets where the two values occur are closed and form a partition of X, so X is
disconnected.
In many reasonable cases, X can be written as a disjoint union of connected open subschemes; these are then called the connected components of X.
Irreducible schemes
A nonempty scheme is irreducible if its underlying topological space is irreducible, i.e., can
not be written as a union of two proper closed subsets, i.e., does not contain two disjoint
nonempty open subsets. (The empty scheme is not irreducible.) Note that a nonempty open
subscheme of an irreducible scheme is still irreducible.
Lemma. The nonempty ane scheme X = Spec(A) is irreducible if and only if the nilradical
of A is a prime ideal (i.e., every zero divisor of A is nilpotent).
Proof. Note that X is irreducible if and only if the intersection of any two nonempty open
subsets is nonempty. It is of course enough to check the intersection of two distinguished
opens D(f ), D(g). They are nonempty if and only if f and g are not nilpotent; the inter
section D(f g) is nonempty if and only if f g is not nilpotent. Hence X is irreducible if and
only if the nilradical of A is prime.
Handy fact: the spectrum of a local ring is irreducible, because the maximal ideal belongs
to every closed subset.
A generic point of a topological space is a point belonging to every nonempty open subset.
Lemma. If X is irreducible, then X has a unique generic point.
Proof. If X = Spec(A), then the nilradical of A is the unique generic point. In general, if
X is irreducible and U = Spec(A) is a nonempty open ane, then any generic point of X is
also a generic point of U. Conversely, if U is the unique generic point of U (which exists
because U is forced to be irreducible), then there cannot be an open ane subset V of X
omitting , as then V U would have to be empty (since it is an open subset U missing the
generic point of U ), a contradiction.
Integral schemes
A nonempty scheme is integral if it is irreducible and reduced. (The empty scheme is not
irreducible.)
Lemma. Put X = Spec(A). Then the following are equivalent.
(a) X is integral.
(b) A is an integral domain. (The zero ring is not an integral domain.)
(c) X is connected and each local ring OX,x is an integral domain.
Proof. The only nontrivial implication is (c) = (a). Suppose (c); note that it implies that
X is reduced. Choose f A. Let U be the set of x X such that f has nonzero image in
OX,x ; then U is open (previously assigned exercise).
2
We claim that X \ U is also open. To see this, pick x X \ U corresponding to a prime

ideal p of A. Since f maps to zero in Ap , there must exist g A \ p for which f g = 0. That
equality in turn implies that D(g), which contains p, is in fact contained in X \ U. Since
each point of X \ U has an open neighborhood contained in X \ U, we conclude that X \ U
is open.
Since X is connected, it follows that U equals either X or the empty set. In the latter
case, f belongs to the nilradical of A, and so equal 0 because X is reduced.
We conclude that if f, g A are nonzero, their images in each Ap are nonzero. Hence f g
also has nonzero image in each Ap , and so must be nonzero. This proves (a).
Normal schemes
A scheme X is normal if for each x X, the local ring OX,x is an integral domain and is
integrally closed in its eld of fractions.
Lemma. Suppose X = Spec(A) is connected. Then X is normal if and only if A is an
integral domain which is integrally closed in its eld of fractions.
Proof. If A is an integral domain which is integrally closed in its eld of fractions, then so is
each localization of A (see Atiyah-Macdonald, Proposition 5.12), so X is normal. Conversely,
suppose X is connected and normal. By the previous lemma, A is an integral domain.
It remains to check that an integral domain is integrally closed (in its eld of fractions) if
and only if its localization at each prime ideal has this property. This follows from the easy
fact that A is the intersection of the Ap .
The construction of the integral closure of a domain can be sheaed. (Note: a dominant
morphism is one with dense image.)
Theorem 1. Let X be an integral scheme. Then the category of dominant morphisms
X with X
normal has a nal element.
X
Proof. Exercise.
The nal element is called the normalization of X. Under normal circumstances, the
X is nite, but there are pathological counterexamples unless one imposes
morphism X
some hypotheses.
One attempt is the notion of a Nagata ring. We say an integral domain R is N-1 if the
integral closure of R in Frac(R) is nite as an R-module. We say R is N-2 if for any nite
extension L of Frac(R), the integral closure of R in L is nite as an R-module. We say a
general ring R is a Nagata ring if R is noetherian and R/p is N-2 for any prime ideal p of
R. (Without the noetherian hypothesis, I think this is what is called a universally Japanese
ring in EGA. My denition is from Matsumura, Commutative Algebra, 31.) The point is
that the Nagata property is stable under many natural operations: localizations, quotients,
passing to a nitely generated ring extension, certain types of completion, etc.
3
Dimension and codimension
The dimension of a scheme X is the length of the longest chain Z0 Z1 Zn of

distinct irreducible closed subsets of X (keeping in mind that the numbering starts at 0).
The dimension of an ane scheme X = Spec(A) is the same as the Krull dimension, since
irreducible closed sets of X correspond to prime ideals of A.
The codimension of an irreducible closed subset Z of X is the length of the longest chain
Z0 Z1 Zn of distinct irreducible closed subsets of X for which Z0 = Z. We can
similarly dene the codimension of one irreducible closed subset inside another.
These notions can behave badly even for the spectrum of a noetherian ring (Hartshorne,
Caution 3.2.8). Again, we need to impose more hypotheses before working with these in any
detail; the best way to do this is work with the class of excellent schemes. More on those
later.
Regular schemes
Let A be a local ring with maximal ideal m and residue eld k = A/m. The cotangent space
of A is the k-vector space m/m2 ; its dual is called the tangent space of A.
Suppose A is noetherian. Then it is a nontrivial theorem from commutative algebra (e.g.,
Matsumura 12) that
dimk (m/m2 ) dim A.
If equality holds, we say that A is regular.
We say that a scheme X is regular at a point x if OX,x is a regular local ring, and simply
regular if it is regular everywhere. For instance, if X is a scheme of nite type over a eld
k, then X is regular if and only if the corresponding variety is nonsingular everywhere. For
another example, Spec Z and Spec Z[x] are both regular. We will give a relative version of
nonsingularity later (the notion of a smooth morphism).
Excellent rings and schemes
A quasiexcellent ring is a noetherian ring R with the following properties.

(a) For any prime ideal p of R and any homomorphism R K with K a eld, the ring
p A K is regular.
R
(b) Any integral domain A which is nite as an R-algebra is generically regular, i.e., there
exists a A nonzero such that Aa is regular.
An excellent ring is a quasiexcellent ring R with the following additional property.
(c) The ring R is universally catenary. That is, for any nonnegative integer n and any
two prime ideals p1 p2 of R[x1 , . . . , xn ], any two maximal chains of prime ideals of
R[x1 , . . . , xn ] starting with p1 and p2 have the same length.
4
The class of excellent rings is introduced by Grothendieck in EGA IV part 3 (see 7.8). It
includes some natural examples (elds, Z, complete local rings, and the series in Cx1 , . . . , xn
convergent in a neighborhood of the origin) and is stable under nice operations (localiza
tion, completion, quotient, polynomial ring). These rings have lots of useful properties: for
instance, they are Nagata rings.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Flat morphisms and descent (updated 11 Mar 09)
Hartshorne only treats atness after cohomology (so see III.9) and doesnt talk about
descent at all. The EGA reference for atness is EGA IV, part 2, 2. Im not sure if descent
is discussed at all in EGA, so I gave references to SGA 1 instead.
Flat sheaves and at morphisms
Let f : Y X be a morphism and let F be a quasicoherent OY -module. We say F is at

relative to f if for each point y Y with f (y) = x, if we use the map f : OX,x OY,y to
view F as a OX,x -module, then that module is at in the usual sense. (The usual sense is
that an R-module M is at if tensoring with it is exact, not just right exact.) If this holds
at a particular y, we say F is at at y relative to f .
Two special cases:
If Y = X, we say that F is a at OX -module; it is equivalent to saying that tensoring
with F is an exact functor on quasicoherent OX -modules. For instance, any locally
free OX -module is at.
If F = OY , we say that f is a at morphism. For example, any open immersion is at.
Note that if F is a at OY -module and f is a at morphism, then F is at relative to f .
Note that also that atness is local on the source, not just on the target, and stable under
base change.
is a
Lemma. Let X = Spec(R) be an ane scheme, and let M be an R-module. Then M
at OX -module if and only if M is a at R-module.
Proof. This should be a familiar fact from commutative algebra: M is at over R if and only
if Mp is at over Rp for each prime ideal p. For completeness, I include the proof here.
Suppose rst that M is at. Let p be an ideal and let N P be an injection of
Rp -modules. We may then view N, P as R-modules and identify
Mp R N = M p Rp N
and similarly for P . Since localization is at, Rp is a at R-algebra, so Mp is at not just
over Rp but also over R. Hence Mp N Mp P is injective, so Mp is at over Rp .
Suppose next that Mp is at over Rp for each p. If N P is an injection of R-modules,
we must check that M N M P is still injective. Localizing gives Mp Np Mp Pp
(since localization commutes with tensor product), which is injective because Mp is at.
Corollary. Let A B be a homomorphism of rings. Then Spec(B) Spec(A) is at if
and only if B is at as an A-module.
1
Proof. The statement that Spec(B) Spec(A) is at says that for each q Spec(B)
mapping to p Spec(A), the morphism Ap Bq is at. This follows from A B being
at because the localization Bp Bq is at. Conversely, suppose that this holds. Let
N P be an injection of A-modules. Then for each prime ideal p of A, we may view
Bp A N Bp A P as a morphism of Bp -modules. For each prime ideal q of B over p,
tensoring with Bq over Bp simply gives Bq A N Bq A P . This is injective because
A Ap is at always and Ap Bq is at by hypothesis.
Applying the previous lemma over Bp , we may now deduce that Bp A N Bp A P is
injective. That is, Bp is at over A, or equivalently over Ap . Applying the previous lemma
over A, we deduce that B is at over A.
The notion of atness, while useful (especially when we study cohomology), is geomet
rically somewhat mysterious. For projective morphisms, one can give a geometric interpre
tation in terms of Hilbert polynomials; more on that later. In the interim, you may wish to
chew on the following examples. (See Eisenbud-Harris II.3.4 for more examples.)
Let k be an algebraically closed eld. The morphism
Spec k[x, t]/(x2 t) Spec k[t]
is at. If the characteristic of k is not 2, then the bres above points t = 0 are pairs of
distinct points whereas the bre above t = 0 is the doubled origin in Spec k[x].
The morphism
Spec k[x, t]/(x2 t2 ) Spec k[t]
is also at, but the source is not normal. If we replace the source by its normalization, we
get two copies of the ane line mapping to one ane line, and this is also at.
Hartshorne gives the example of the family of cubic curves in A3 given as parametric
equations in u by
x = u2 1, y = u3 u, z = tu.
If we eliminate u and make sure the result is at over Spec k[t], we get
Spec k[x, y, z, t]/(t2 (x + 1) z 2 , tx(x + 1) yz, xz ty, y 2 x2 (x + 1)) Spec k[t].
The bre over t = 0 is supported on the plane curve y 2 = x2 (x + 1), z = 0 but is not a
subscheme of the plane z = 0 in Spec k[x, y, z]: the local ring at the origin contains the
nonzero nilpotent element z.
Here are some deep results about atness. For this one, see EGA 4, part 2, Theoreme
2.4.6
Theorem 1. Let f : X Y be a morphism which is at and locally of nite presentation.
Then f is universally open, i.e., any base change of f is an open map (the image of any
open set is open) on topological spaces.
For this one, see SGA 1, Expose IV, Theorème 6.10 or EGA 4, part 3, 11.1.1.
Theorem 2. Let f : Y X be a morphism of nite type, with X locally noetherian, and
let F be a quasicoherent OY -module. The set of y Y at which F is at relative to f is an
open subset of U .
2
Faithfully at morphisms and descent
A morphism which is both at and surjective is faithfully at. For instance, if Spec(B)
Spec(A) is a morphism of ane schemes, then this morphism is faithfully at if and only if
B is faithfully at in the usual sense, i.e., B is at over A, and for any A-module M , the
map M M A B of A-modules is injective.
Faithfully at morphisms are important because of their role in descent, the process of
undoing a base change. Here is a typical example.
Let f : Y X be a morphism. Let 1 , 2 : Y X Y Y be the canonical projections.
The category of descent data for quasicoherent sheaves relative to f is dened as follows. A
descent datum is a quasicoherent OY -module F equipped with an isomorphism : 1 F
2 F , satisfying the following cocycle condition. Let 1 , 2 , 3 : Y X Y X Y Y be

the canonical projections. Use rst to identify 1 F with 2 F , then 2 F with 3 F. The
resulting isomorphism 1 F 3 F must coincide with the one obtained directly by applying
to the rst and third factors.
A morphism of two descent data is a morphism F G of the underlying OY -modules,
such that the induced morphisms 1 F 1 G and 2 F 2 G commute with the isomor
phisms . There is no extra cocycle condition.

In general, there is a functor from quasicoherent OX -modules to descent data taking E
to f E, and dening in the obvious manner.
Theorem 3 (Faithfully at descent). Let f : Y X be a faithfully at, quasicompact
morphism. Then the natural functor from quasicoherent OX -modules to descent data for
quasicoherent sheaves dened by f is an equivalence of categories.
The reference for this is SGA 1, Expose VIII, section 1. However, the proof there is
written in a somewhat cryptic manner; we will see a somewhat simplied proof in the
exercises.
Note that faithfully at descent for quasicoherent sheaves includes as a special case Galois
descent: if L/K is a nite Galois extension of elds, and V is an L-vector space equipped
with a semilinear action of Gal(L/K), then V has a basis of invariant elements. (The usual
proof uses Noethers nonabelian generalization of Hilberts Theorem 90, i.e., the fact that
the rst Galois cohomology set of Gal(L/K) acting on GLn (L) is trivial.)
Armed with faithfully at descent for quasicoherent sheaves, one can now establish de
scent for various properties of morphisms. (Some of these can be found in EGA 4, part 2.)
For example:
Theorem 4. Let f : Y X be a morphism, and let g : Z X be a faithfully at
quasicompact morphism. Then f is of nite type if and only if the base change of f by g is
of nite type.
Proof. Suppose rst that X = Spec(A), Y = Spec(B), Z = Spec(C) are all ane, and that
the base change of f by g is of nite type. Then B is the direct limit of its nitely generated
A-subalgebras Bi , and so B A C is the direct limit of the nitely generated C-subalgebras
3
Bi A C. By hypothesis, B A C is nitely generated as a C-algebra; each generator can

itself be written in terms of nitely many elements of B and C. Hence B A C can be
generated over C by nitely many elements of B, and so must occur as one of the B i A C.
For that index i, the fact that the inclusion Bi B is an isomorphism follows from the fact
that Bi A C B A C is an isomorphism because C was assumed to be faithfully at over
A.
To nish, we must show that if the base change of f by g is quasicompact, then f is
quasicompact. We may assume X is ane, as then is Z because g was required to be
quasicompact, as then is Y X Z by hypothesis. Let {Ui } be an open ane cover of Y .
By hypothesis, the open cover {Ui X Z} of Y X Z admits a nite subcover. Since those
Z X is surjective, the corresponding Ui must then cover Y . Hence Y is a union of nitely
many anes, hence quasicompact.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Dierentials
See Hartshorne II.8.
The module of K
ahler dierentials
Let A B be a homomorphism of rings. The module of K

ahler dierentials of B over
A is a B-module B/A equipped with an A-linear derivation d : B B/A (an A-linear
homomorphism satisfying the Leibniz rule d(xy) = x dy + y dx for x, y B; note that
this forces d(a) = 0 for a A), with the following universal property: for any B-module
M and any A-linear derivation : B M, factors uniquely through d via a B-linear
homomorphsism B/A M .
There are two standard ways to construct B/A . One is to form the B-module generated
by symbols db for b B, modulo the necessary relations:
(a) d(b1 b2 ) b1 db2 b2 db1 for b1 , b2 B;
(b) d(b1 + b2 ) = d(b1 ) + d(b2 ) for b1 , b2 B;
(c) d(a) = 0 for a A.
This obviously has the desired universal property. The other is to let I be the kernel of
the multiplication map B A B B, and put B/A = I/I 2 equipped with the map d(b) =
b 1 1 b. This evidently gives an A-linear derivation. Given a derivation : B M,
view B M as a B-algebra in by setting m1 m2 = 0 for all m1 , m2 M. Then the formula
b1 b2 (b1 b2 , b1 (b2 ))
induces a ring homomorphism B A B B M under which I maps to M, so I 2 maps
to 0 and we get a B-linear map I/I 2 M. Composing with d easily gives back . The
uniqueness of the factorization follows by observing that
x y = xy 1 x(y 1 1 y)
so the image of d generates I (and hence I/I 2 ) as a B-module.
For instance, if B = A[x1 , . . . , xn ], then B/A is freely generated by dx1 , . . . , dxn . Also,
if k is an algebraically closed eld and A is a reduced quotient of k[x1 , . . . , xn ], then the
Jacobian criterion can be interpreted as saying that A corresponds to a nonsingular variety
over k if and only if A/k is locally free as an A-module.
For another example, if A is a eld and B is a nite eld extension, then B/A = 0 if
and only if B is separable over A.
The sheaf of K
ahler dierentials
Let f : Y X be a morphism. For each open ane subset U = Spec(A) of X and each
open ane subset V = Spec(B) of f 1 (U), form the module B/A . We would like these
to form the sections of a sheaf Y /X , but checking the glueing property directly from this
denition is a bit awkward.
Fortunately, our second construction of the module of Kahler dierentials suggests a
global denition of the sheaf Y /X . Well explain this rst in case f is separated. In that
case, : Y Y X Y is a closed immersion; let I be the corresponding ideal sheaf on
Y X Y . We then put
Y /X = (I/I 2 ).
But what if f is not separated? In that case, we still claim that is an immersion; this
follows from the proof of Hartshorne Corollary II.4.2. Then gives rise to an ideal sheaf not
on Y X Y , but on some open subscheme containing the image of ; we may then proceed
as in the separated case.
Useful properties of :
The formation of Y /X commutes with base change as follows. If g : Z X is another
morphism, then Y X Z/Z can be identied canonically with the pullback of Y /X along
the projection Y X Z Y (Hartshorne, Proposition II.8.10).
If f : Z Y and g : Y X are morphisms, then there is a natural exact sequence
f Y /X Z/X Z/Y 0
(Hartshorne, Proposition II.8.11).
If f : Y X is a morphism, and j : Z Y is the closed immersion dened by the
ideal sheaf I on Y , then there is a natural exact sequence of sheaves on Z:
j (I/I 2 ) j (Y /X ) Z/X 0
(Hartshorne, Proposition II.8.12).
Let A be a ring, and let f : Y = PnA X = Spec A be the natural morphism. We
then have a short exact sequence
0 Y /X OX (1)(n+1) OY 0
(Hartshorne, Theorem II.8.13).
As in the ane case, a variety X over a eld k is nonsingular if and only if X/k is locally
free. Since X/k is necessarily nitely generated (deduce this from the case of ane space),
there is always an open dense subset U of X which is nonsingular over k.
Suppose X is nonsingular of dimension n (on each component). Then we call the sheaf
X/k = n X/k the canonical sheaf on X; it is locally free of rank 1. As the name suggests,
2
the canonical sheaf is an omnipresent object in the study of the geometry of varieties; we
will see it in the Riemann-Roch theorem, and more generally in Serre duality, but it is also
a central player in modern birational geometry, as in the following very hard theorem.
Theorem (Bircar-Cascini-Hacon-McKernan, Siu). Let X be a smooth projective irreducible
variety over C. Then the ring
n
(X, X/k
)
n=0
is nitely generated as a C-algebra.
Smooth, unramied, and

etale morphisms
Let f : Y X be a morphism of schemes. For each morphism g : X X with X ane,

and each closed subscheme Z of X dened by a nilpotent ideal of O(X ), we have a canonical
map
HomX (X , Y ) HomX (Z, Y ).
If this map is always injective/surjective/bijective, we say that f is formally unramied/smooth/etale.
We drop the formally if f is also locally of nite presentation. These properties have all the
expected behaviors (local on the base, stable under base change, descendable down faithfully
at quasicompact morphisms).
The denition above is given in terms of an innitesimal lifting property. There are more
practical characterizations in terms of dierentials; some of these will be exercises. (See
EGA IV, part 4, section 17.)
Proposition. The morphism f is formally unramied if and only if Y /X = 0.
Proposition. If f is locally of nite presentation, then f is etale if and only if f is at and
unramied.
Proposition. If f is locally of nite presentation, then f is smooth if and only if f is at
and for each x X, the bre f 1 (x) is geometrically regular over (x). (That is, for k an
algebraic closure of (x), f 1 (x) Spec (x) k is regular.)
For example, the projective space PnX is smooth over X.
The dierence between regular and geometrically regular shows up only when the eld
(x) is imperfect. For instance, put = Fp (x), X = Spec and Y = Spec Fp (x1/p ) =
Spec [y]/(y p x). Then Y is a regular scheme, but its base change to an algebraic closure
k of is
Spec k[y]/(y p x) = Spec k[y]/((y x1/p )p ),
which is not regular. For a slightly less trivial example, see Hartshorne exercise III.10.1.
The notion of an etale morphism is an algebro-geometric analogue of the concept of a
covering space in topology. As such, it forms the basis for one of the most successful notions
of cohomology in algebraic geometry, that of etale cohomology. I probably wont have time
to say more than a few words about that at the end of the course.
3
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Divisors, linear systems, and projective embeddings (updated 1 Apr 09)
We conclude the rst half of the course by translating into the language of schemes some
classical notions related to the concept of a divisor. This will serve to explain (in part) why
we will be interested in the cohomology of quasicoherent sheaves.
In order to facilitate giving examples, I will mostly restrict to locally noetherian schemes.
See Hartshorne II.6 for divisors, and IV.1 for Riemann-Roch.
Weil divisors
Introduce Hartshornes hypothesis (*): let X be a scheme which is noetherian, integral,

separated, and regular in codimension 1. The latter means that for each point x X whose
local ring OX,x has Krull dimension 1, that local ring must be regular.
Lemma. Let A be a noetherian local ring of dimension 1. Then the following are equivalent.
(a) A is regular.
(b) A is normal.
(c) A is a discrete valuation ring.
(This is why normalizing a one-dimensional noetherian ring produces a regular ring.)
Warning: for a noetherian integral domain, normal implies regular in codimension 1 but
not conversely. You have to add Serres condition S2: for a A, every associated prime of
the principal ideal (a) has codimension 1 when a is not a zerodivisor, and has codimension
0 when a = 0.
A prime (Weil) divisor on X is a closed integral (irreducible and reduced) subscheme of
codimension 1. A formal Z-linear combination of prime divisors is called a Weil divisor. If
only nonnegative coecients are used, we say the divisor is eective.
For example, let K(X) be the function eld of X, i.e., the local ring of X at its generic
point. (This equals Frac(O(U)) for any nonempty open ane subscheme U of X.) For
f K(X) nonzero, we can dene a principal divisor associated to f as follows. For each
prime divisor Z on X, let Z be the generic point of Z. Then OX,Z is a discrete valuation
ring; let vZ be the valuation. Now dene the divisor
(f ) =
vZ (f )Z;
Z
this makes sense because only nitely many vZ (f ) are nonzero. (Thats because f restricts
to an invertible regular function on some nonempty open subscheme U of X, and vZ (f ) = 0
whenever Z 6 X U.)
Let Div X be the group of Weil divisors of X. The principal divisors form a subgroup
(since (f ) + (g) = (f g)); the quotient by this subgroup is called the divisor class group of
1
X, denoted Cl X. For example, if X = Spec(A) with A a Dedekind domain, then Div X is

the group of fractional ideals, and Cl X is the ideal class group. We say two divisors which
dier by a principal divisor are linearly equivalent.
There are a number of examples in Hartshorne. One of my favorites is that of an elliptic
curve; here is a summary. Let k be an algebraically closed eld (for starters). Let P (x, y, z)
k[x, y, z] be a homogeneous polynomial of degree 3 dening a nonsingular subvariety C of P2k .
Pick a point O C(k). There is a surjective map Div X Z mapping each prime divisor
P to 1, called the degree. This map factors through Cl X because each principal divisor has
degree 0. The kernel of the degree map Cl X Z is generated by (P ) (O) for P C(k).
In fact it is equal to the set of such elements: given P, Q C, we rst draw the line through
P, Q in P2k and nd its third intersection point R with C. We then draw the line through R
and O in P2k and nd its third intersection point S with C. Then
(P ) + (Q) + (R) (R) + (S) + (O),
so
(P ) (O) + (Q) (O) (S) (O).
Cartier divisors
When the scheme X is not regular, there is a more restrictive notion of divisors that turns
out to be more useful in many cases.
Let K be the locally constant sheaf associated to the function eld K(X). A Cartier
divisor on X is a section of the sheaf K(X)/O . Using the construction of principal divisors,
we obtain a map from Cartier divisors to Weil divisors: if the Cartier divisor is represented
on some open subset U of X by the rational function f K(X), then the Weil divisor
we get should agree with (f ) when restricted to U (i.e., only keep the components of those
prime divisors meeting U). This map is injective if X is normal, because an integrally closed
noetherian domain is the intersections of its localizations at minimal prime ideals.
Proposition (Hartshorne, Proposition II.6.11). Suppose X is locally factorial (i.e., each
local ring OX,x is a unique factorization domain). Then the previous map is an isomorphism.
(In particular, this holds if X is regular, because a regular local ring is factorial by a not-so
easy theorem of commutative algebra.)
Example: if X = Spec k[x, y, z]/(xy z 2 ), the ideal (x, z) denes a Weil divisor which is
not a Cartier divisor.
Again, there is an obvious notion of a principal Cartier divisor, namely one dened by
a single element of K(X). The group of Cartier divisors modulo principal divisors is called
the Cartier class group of X, denoted CaCl X.
The Picard group
The Cartier class group is usually the same as the Picard group, namely the group of
invertible sheaves on X under the tensor product. Namely, if D is a Cartier divisor on X,
let L(D) be the subsheaf of K such that
L(D)(U) = {f K(X) : ((f ) + (D))|U 0}.
Assuming that X is normal, this is locally free of rank 1, hence an invertible sheaf. This gives
a homomorphism from Cartier divisors to the Picard group, which we see kills the principal
divisors. The resulting homomorphism is always injective, even without any hypotheses on
X (Hartshorne, Corollary II.6.14) but may not be surjective; however, it is surjective if X
is integral (Hartshorne, Proposition II.6.15).
Note that if D is eective, then the function 1 denes a global section of L(D). Since L
is locally principal, we can locally identify L with OX ; when we do so, the subsheaf of L(D)
generated by 1 goes into correspondence with an ideal sheaf of OX , which doesnt depend
on any choices. This ideal sheaf denes D as a closed subscheme. In other words, D is the
zero locus of a certain section of L(D). More generally, even if D is eective, we can view D
as the zero locus of a meromorphic section of L(D) (meaning a zero locus of L(D) OX KX ),
and indeed the zero locus of any meromorphic section of L(D) is linearly equivalent to D.
Linear systems
Suppose X is an integral separated scheme of nite type over a eld k (which need not be
algebraically closed). Let L be an invertible sheaf on X. A linear system dened by L is the
set of zero loci of some k-linear subspace H of H 0 (X, L). If we take the entire space, that is
called the complete linear system dened by L.
We can attempt to use the elements of H to dene a map X Pnk , where n = dimk (H)1.
This might fail to give a morphism because H may have a base point, i.e., a point in the
intersection of all of the divisors in the linear system. In fact, we get a morphism X Pnk
if and only if H has no base points.
Suppose now that k is algebraically closed, and that X is one-dimensional, projective,
irreducible, and nonsingular (i.e., a curve). Consider the complete linear system associated
to L(D) for some divisor D.
(a) We get a map X Pnk if and only if for each closed point x X, we have dimk H 0 (X, L(D
x) = dimk H 0 (X, L(D)) 1. (In other words, there must be a section of L(D) not
vanishing at x.)
(b) The map in (a) is injective as a map of sets if and only if for each pair of distinct closed
points x, y X, we have dimk H 0 (X, L(D x y)) = dimk H 0 (X, L(D)) 2. (In other
words, there must be a section of L(D) vanishing at x but not at y, and vice versa.)
(c) The map in (b) is a closed immersion if and only if for each closed point x X, we
have dim H 0 (X, L(D 2x)) = dimk H 0 (X, L(D)) 2. (In other words, there must be
a section of L(D) not vanishing at x, and a section vanishing to exact order 1 at x.)
(Condition (c) is needed to ensure that the tangent space at x embeds into the tangent space
at the image of x. See Remark 7.8.2.)
Since we would like to know under what circumstances X embeds into a projective space,
we would like to be able to compute at least the dimension of H 0 (X, L(D)) for each divisor
D. This quest is greatly abetted by the Riemann-Roch theorem, more on which next time.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Divisors on curves and Riemann-Roch (updated 31 Mar 09)
We continue the discussion of divisors but now restricted to curves. Again, see IV.1 for
Riemann-Roch and IV.2 for Riemann-Hurwitz.
The Riemann-Roch theorem
Again, let X be a (projective, irreducible, nonsingular) curve over an algebraically closed eld
k. Since X is one-dimensional, the canonical sheaf X/k coincides with the sheaf of Kahler
dierentials X/k . By a canonical divisor, I mean a divisor K dened by any meromorphic
section of X/k . (This means that a canonical divisor is in fact not canonical in any sense.
Sorry about that.)
As in the elliptic curve example, there is a homomorphism Div X Z sending (P ) to 1
for each P X(k), and this factors through Cl X because any principal divisor has degree
0 (Hartshorne, Corollary II.6.10).
Write l(D) as shorthand for dimk (X, L(D)). The following theorem will be proved later
using properties of sheaf cohomology (particularly Serre duality), but in the meantime we
will see (in this lecture and in the next problem set) how it tells us many useful things that
have no overt relationship to cohomology.
Theorem (Riemann-Roch). There exists a nonnegative integer g = g(X) with the following
property. For any divisor D and any canonical divisor K,
l(D) l(K D) = deg(D) + 1 g.
Corollary. The integer g in Riemann-Roch can be identied as
g = l(K) = dimk (X, X/k ).
Proof. Take D = 0. Then l(D) = 1 because any global regular function on a curve (or
indeed on any projective variety) is constant. This forces l(K) = g.
The quantity l(K) is called the genus of K, or more precisely the geometric genus. In case
k = C, this will end up matching the topological genus of the Riemann surface associated
to X.
Corollary. The integer g in Riemann-Roch can also be identied by the formula
deg(K) = 2g 2.
Proof. Apply Riemann-Roch with D = K to obtain (by the previous corollary)
g 1 = l(K) l(0) = deg(K) + 1 g.
Corollary. If deg(D) > 2g 2, or deg(D) = 2g 2 and D K, then

l(D) = deg(D) + 1 g g 1.
Proof. If deg(D) = 2g2, then deg(K D) = 0. If f K(X) nonzero satises (f )+K D
0, we must have equality because the left side has degree 0. Thus l(K D) is only nonzero
if K D.
If deg(D) > 2g 2, then deg(K D) < 0. In this case, (f ) + K D has negative degree
and so cannot be eective, so l(K D) = 0 no matter what.
Corollary. For g 2, for any divisor D of degree at least 2g 1, the complete linear system
associated to D denes a closed immersion of D into a projective space.
The canonical (almost) embedding
The canonical embedding is the map to projective space dened by the complete linear
system associated to a canonical divisor K. The name suggests that it is always a closed
immersion, but this is only almost true; there are a few exceptions in low genus (for which
see the exercises).
Lemma. For any point P and any divisor D, we have
l(D) l(D + P ) l(D) + 1.
Consequently, l(D) deg(D) + 1.
Proof. We have an exact sequence of sheaves
0 L(D) L(D + (P )) E 0
where E is the quotient of OX by the ideal sheaf dening P . So clearly l(D) l(D + P ).
On the other hand, taking global sections yields a short exact sequence
0 (X, L(D)) (X, L(D + (P ))) (X, E)
and the last term is one-dimensional over k, so we get l(D + P ) l(D) + 1.
Proposition. The canonical embedding is a closed immersion if and only if X is not hyperelliptic.
Proof. The special cases g = 2, 3 are discussed in the problem set, so Ill only sketch the
general argument. Put D = (P ) + (Q) for P, Q X(k) not necessarily distinct. We need to
check whether we always have
l(K D) = l(K) 2 = g 2.
2
By Riemann-Roch,
l(K D) = l(D) + g 3
so we have an embedding if and only if l(D) = 0 for any eective D of degree 2; but a failure
of that denes a two-to-one map to P1 , in which case X is hyperelliptic. (Strictly speaking,
we should also check for D of degree 1, but its esay to see that if such D has l(D) > 0, then
there exists a rational function on X with a single pole, which gives a a degree 1 map to P 1 .
That is, X
= P1 .)
The canonical embedding, and variants of it (e.g., using higher multiples of a canonical
divisor) are key tools for studying the moduli space of curves of a given genus. This is
almost a scheme Mg which represents the functor taking schemes to families of curves of
genus g, except that this functor is not quite representable. It becomes representable in
the category of Deligne-Mumford stacks, which extend schemes in much the same way that
orbifolds extend manifolds (by allowing quotients by nite group actions).
The Riemann-Hurwitz formula
Let f : X Y be a nite separable morphism of curves (i.e., the induced eld extension
k(X)/k(Y ) is separable). The ramication divisor of f is dened as
R=
length(X/Y )P (P ),
P X(k)
where as usual X/Y is the module of Kahler dierentials.

Proposition. We have
KX f KY + R.
Proof. (Compare Hartshorne Proposition IV.2.3.) Note that
0 f Y /k X/k X/Y 0
is exact; this follows from properties of Kahler dierentials except for the injectivity on the
left. But that we can check at generic points, where it follows because k(X) is separable
over k(Y ).
We can then tensor with X/k to obtain another exact sequence
0 (f Y /k ) X/k OX X/Y X/k 0.
However, X/Y is supported on nitely many points, so it is isomorphic to its twist by X/k .
So we really have an isomorphism
(f Y /k ) X/k
= OX /X/Y .
We thus get an equality of associated divisors; these are f KY KX on the left and R on
the right.
3
Using Riemann-Roch, we deduce the Riemann-Hurwitz formula.

Proposition. We have
2g(X) 2 = (deg(f ))(2g(Y ) 2) + deg(R),
where deg(f ) is the degree of f (i.e., the degree of the eld extension k(X)/k(Y )).
Moreover, the contribution of P X(k) can sometimes be computed very simply.
Namely, put Q = f (P ), and pick t k(Y ) which generates mY,Q ; then f (t) generates
meX,P for some nonnegative integer e. We call e = eP the ramication index of P . Then
length(X/Y )P eP 1,
with equality if and only if f is tamely ramied, i.e., eP is not divisible by the characteristic
of k.
In case k = C, the Riemann-Hurwitz formula has a topological meaning: the quantity
2 2g(X) turns out to compute the Euler characteristic of the associated Riemann surface.
The Euler characteristic (computed using homology, or compactly supported cohomology)
is an additive invariant of a topological space. If the map f were unramied, then we would
have deg(R) = 0 and the space X would have Euler characteristic equal to deg(f ) times
that of Y . Otherwise, one must subtract eP 1 for each point P with eP > 1, because you
get X from an unramied cover of Y by removing eP dierent points from the bre (each of
which has Euler characteristic 1) and adding one point back in.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Homological algebra (updated 8 Apr 09)
We now enter the second part of the course, in which we use cohomological methods to
gain further insight into the theory of schemes. To start with, let us recall some of the basics
of homological algebra. The original reference for derived functors is the book Homological
Algebra of Cartan and Eilenberg, and for cohomological functors is Grothendiecks article
Sur quelques points dalgèbre homologique; however, any good modern book on homolog
ical algebra (e.g., Weibel, An Introduction to Homological Algebra) should suce. (It is
worth keeping in mind Langs suggested exercise in homological algebra: take any book on
homological algebra, read the statements of the theorems, and prove them all yourself.)
Abelian categories
We saw once before the notion of an abelian category. This is a category C in which each
homset has the structure of an abelian group in a manner compatible with composition,
with some additional restrictions designed to make things well-behaved. Lets recall some
of these. First of all, there must exist biproducts, i.e., for any nonnegative integer n and
any objects X1 , . . . , Xn in C, there must exist an object Y and morphisms i : Xi Y and
i : Y Xi for i = 1, . . . , n such that
Y is the product of the Xi (using the i ) and the
coproduct of the Xi (using the i ), and ni=1 i i = 1.
Also, each morphism must have a kernel and a cokernel. A kernel of the morphism
f : X Y to be a limit of the diagram
X
We write Ker(f ) for the domain of a kernel. Similarly, a cokernel of f is a colimit of
We write Coker(f ) for the codomain of a cokernel.

Finally, we insist that every monomorphsm be the kernel of its cokernel, and every
epimorphism be the cokernel of its kernel.
Examples:
1. Ab, the category of abelian groups.
2. ModR , the category of modules over a ring. We can drop our running commutativity
hypothesis if we choose to work with, say, left modules.
1
3. The category of sheaves on a xed topological space with values in another abelian
category.
I recommend just thinking about the case of abelian groups. The Freyd-Mitchell embedding
theorem implies that most things you prove about an abelian category can be deduced from
the case of abelian groups, where you can use diagram-chasing arguments.
Complexes and exact sequences
Throughout this section, all objects are in a particular abelian category C.

A sequence of morphisms
di1
di
C i1 C i C i+1
is a complex if the composition of any two of the arrows is zero, i.e., di di1 = 0 for all
i. Note that I number the objects so that the arrows point in the increasing direction;
this is called a cohomological grading. If I numbered things the other way, I would have a
homological grading. I will mostly talk about the cohomological grading because that is what
is most convenient for algebraic geometry. (In a homological grading, you usually write with
subscripts instead of superscripts, i.e., di : Ci Ci1 .)
The i-th cohomology of a complex C , denoted hi (C ), is dened as
ker(di )
.
im(di1 )
hi (C ) =
We say that C is exact if hi (C ) = 0 for all i.

A morphism of complexes f : C D is a commutative diagram
C i1
di1
Ci
f i1
D i1
di1
di
C i+1
fi
Di
f i+1
di
D i+1
With this denition, we obtain a category of complexes with values in C; this is again an
abelian category (exercise).
Any morphism f : C D induces maps
f i : hi (C ) hi (D )
for each i. We say f is a quasi-isomorphism (or quasiisomorphism, but Ill spare you the
doubled vowel) if each f i is an isomorphism; for example, this occurs if f is homotopic to
the zero map in the following sense. Given two maps f , g : C D , we say that f and g
are homotopic if there exist a sequence of maps
k i : C i D i1
2
such that
k i+1 di + di1 k i = f g;
this is obviously an equivalence relation. It is an exercise to show that this implies that f
and g induce the same maps hi (C ) hi (D ). (The collection of maps k i are called a chain
homotopy between f and g.) Important: the fact that a morphism is a quasi-isomorphism is
not stable under applying functors, but the fact that two morphisms are homotopic is stable
under applying functors because it is arrow-theoretic. (This should remind you of the fact
that a sequence being exact is not stable under applying functors, but it being a complex is
stable.)
The homology functors dont quite capture as much information as possible, just as
passing from a ltered object to its associated graded object loses information. A better
construction is that of the derived category of complexes with values in C; in this construction,
one formally inverts all quasi-isomorphisms. This is not completely straightforward, and I
wont talk about it more just now.
The long exact sequence in cohomology
Let
0 C D E 0
be a short exact sequence of complexes, i.e., a diagram
..
.
..
.
D i1
C i1
di1
Ci
Di
Ei
di
C i+1
D i+1
..
.
di1
di
E i1
di1
..
.
di
E i+1
..
.
..
.
in which the rows are exact, and the columns are complexes. As was shown in a previous
exercise, this leads to a long exact sequence
i1
hi1 (C ) hi1 (D ) hi1 (E ) hi (C ) hi (D ) hi (E )

in which the maps h (C ) h (D ) and h (D ) hi1 (E ) are the obvious induced ones,
and the maps i are the connecting homomorphisms. (Recall the denition of i : given an
3
element x in E i1 representing a class in hi1 (E ), use exactness in the row to lift x to

y D i1 . Then the image of di1 (y) in E i equals di1 (x) = 0, so di1 (y) lifts to z C i .
The image of di (z) in D i+1 equals di (di1 (y)) = 0, so z represents a class in hi (C ). The fact
that this class is well-dened independent of choices, and that the resulting map i makes
the long sequence exact, were part of the earlier exercise.)
Cohomological functors
Let F : C1 C2 be an additive covariant functor between abelian categories. Recall that F

is left exact if for any exact sequence
0 A1 A2 A3
the sequence
0 F (A1 ) F (A2 ) F (A3 )
is exact. The functor is right exact if for any exact sequence
A1 A2 A3 0
the sequence
F (A1 ) F (A2 ) F (A3 ) 0
is exact. The functor is exact if it is both left exact and right exact; equivalently, for any
exact sequence
0 A1 A2 A3 0
the sequence
0 F (A1 ) F (A2 ) F (A3 ) 0
is exact. This implies that F preserves exact sequences of any length.
Many interesting functors in mathematics are left or right exact but not exact. For
example, for C an abelian category and X an object, the functor Hom(X, ) carrying Y to
Hom(X, Y ) is left exact. (We saw this previously for ModR but it holds in general.) We
would like to be able to quantify the failure of a functor to be exact; our ability to do this is
aided by the presence of objects on which the functor behaves well. For instance, in ModR ,
the functor X R behaves badly on a general exact sequence. However, if
0 Y1 Y2 Y3 0
is a short exact sequence in which Y3 is a at R-module, then it can be shown that
0 X Y1 X Y2 X Y3 0
is again exact. For instance, this holds if Y3 is a free R-module.
Assume now that F is a left exact functor. The idea now is to replace the single bad
object X rst with the complex 0 X 0 , then with a quasi-isomorphic complex
0 X0 X1
of good objects. If we can lift short exact sequences of maps to short exact sequences of these
resolving complexes, we can then use the long exact sequence in cohomology to quantify the
failure of right exactness. Namely, our short exact sequence
0ABC0
will be replaced by a short exact sequence of complexes
0 A B C 0.
If we have chosen the good objects well, then
0 F (A ) F (B ) F (C ) 0
will still form a short exact sequence of complexes, and its long exact sequence in homology
0
0 h0 (F (A )) h0 (F (B )) h0 (F (C )) h1 (F (A ))
will tell us something useful. What we really want is that h0 (F (A )) = A and so forth, so
that this long exact sequence lls in the gap left at the right end of the exact sequence
0 F (A) F (B) F (C).
To quantify this notion, we dene a cohomological functor (or -functor ) between abelian
categories C1 and C2 to be a sequence of functors
T i : C1 C2
(i = 0, 1, . . . )
plus for each short exact sequence 0 A B C 0 in C1 a morphism i : T i (C)

T i+1 (A) functorial in the sequence (Ill let you draw the diagram), such that the sequence
0
0 T 0 (A) T 0 (B) T 0 (C) T 1 (A) T 1 (B) T 1 (C) T 2 (A)

is exact. A cohomological functor is universal if given any other cohomological functor U and
a natural transformation f 0 : T 0 U 0 , there is a unique sequence of natural transformations
f i : T i U i starting with f 0 which commute with the i . Given T 0 , any two extensions of
it to a universal cohomological functor are naturally isomorphic.
This notion does not become useful without a criterion for checking whether a cohomo
logical functor is universal. Here is one. A functor F : C1 C2 between abelian categories
is eaceable if for any A C1 , there is a monomorphism u : A B with F (u) = 0. I like
to think of this in the following way. Most of the time, we deal with functors which are
kind of monotonic, in the sense that under some appropriate hypothesis, the bigger the
input object into the functor, the bigger the output object. Eaceable functors are quite the
opposite!
5
Theorem (Grothendieck). Let T : C1 C2 be a cohomological functor such that T i is

eaceable for each i > 0. Then T is universal.
Proof. Heres how to construct the natural transformation from T i to U i . Given an object
A and an index i > 0 such that we know the existence and uniqueness of the natural
transformation for indices less than i, choose a monomorphism u : A B with T i (u) = 0.
Then form the long exact sequence 0 A B C 0, apply both cohomological
functors, and use the equality u = 0 to truncate the upper one:
T i1 (A)
T i1 (B)
T i1 (C)
i1
T i (A)
U i1 (A)
U i1 (B)
U i1 (C)
i1
U i (A)
An easy diagram chase shows that there is a unique arrow T i (A) U i (A) making the
diagram commute. It remains to check that:
the arrow T i(A) U i (A) does not depend on the choice of u;
these arrows form a natural transformation.
We leave these verications as an exercise.
A typical case is when each object A C1 admits a monomorphism u : A B in which B
is acyclic for T , that is, T i (B) = 0 for i > 0. These objects are good in the sense considered
above.
Theorem (Acyclic resolution theorem). Let T : C1 C2 be a universal cohomological
functor. Given J C1 , suppose 0 A0 A1 is a complex in C1 with each A acyclic,
h0 (A )
= J, and hi (A ) = 0 for i > 0. (That is, this complex is an acyclic resolution of J.)
Then for each i 0, there is an isomorphism T i (h0 (A ))
= hi (T 0 (A )) which is functorial in
the input data.
Derived functors
We are now ready to make some universal cohomological functors. Unfortunately, we are in
a bit of a jam: we would like to dene them using acyclic resolutions, but the denition of
an acyclic object depends on the denition of the cohomological functor. We get out of this
vicious circle by identifying some objects which are always acyclic.
An object X in an abelian category C is injective if the functor Hom(, X) : C op Ab is
exact. Since this functor is already left exact, it is enough to require something weaker: if
0 Y Z is a monomorphism, then for any morphism Y X we can nd some morphism
Z X tting into the diagram:
0
X
6
For instance, in Ab, an object X is injective if and only if it is divisible, i.e., the multiplication
by-n maps for each positive integer n are all surjective. You might be more familiar with the
dual notion: an object X in an abelian category C is projective if the functor Hom(X, ) :
C Ab is exact. In ModR , any free module is projective; in fact, a module is projective if
and only if it is a direct summand of a free module.
Lemma. Any short exact sequence
0IBC0
with I injective is split, i.e., there exists an arrow C B such that C B C is an
isomorphism.
Proof. Apply the denition of injectivity to the monomorphism I B and the arrow I I
to get a map B I such that I B I is the identity. Then the kernel of B I will be
isomorphic to C.
We once again hit a distinction between non-arrow-theoretic and arrow-theoretic condi
tions; while the property of being a short exact sequence is not preserved under an arbitrary
additive functor, the property of being split short exact is. That is because a splitting of
0 A B C 0 species a pair of endomorphisms e1 , e2 : B B whose sum is B,
namely B A B and B C B, and conversely these endomorphisms determine the
sequence.
Proposition. Let T be a cohomological functor such that T i is eaceable for i > 0 (so in
particular it is universal). Then for any injective object I, T i (I) = 0 for i > 0.
Proof. Choose a monomorphism u : I B with T i (u) = 0, then form the short exact
sequence
0 I B C 0.
Since this sequence splits, the resulting sequences
0 T j (I) T j (B) T j (C) 0
are exact for all j. Consequently, the connecting homomorphism i1 : T i1 (C) T i (I) is
zero. On the other hand, the morphism T j (I) T j (B) is just T j (u), which is also zero. So
T i (u)
i1
the exactness of the sequence T i1 (C) T i (I) T i (B) forces T i (I) = 0.

This more or less forces us into the following denition. We say that the category C has
enough injectives if for any object X there exists a monomorphism X I with I injective.
Then any universal cohomological functor can be computed using injective resolutions. On
the other hand, given an object X, we can always nd an injective resolution; better yet,
given any morphism X Y and an injective resolution of X, we can nd an injective
resolution of Y and a morphism inducing X Y on cohomology. This suggests that we
dene the right derived functors of a left exact functor F by saying for any object X, if I is
an injective resolution of X, put
Ri F (X) = hi (F (I )).
7
Theorem. Assume that C has enough injectives. Then the previous denition gives a welldened cohomological functor, which is eaceable and hence universal.
The eaceability is obvious from the fact that injectives are acyclic under this denition
(if X is injective, use 0 X 0 as the injective resolution). The hard part, or
rather the easy but tedious part, is to check that what you are writing down is really a
well-dened cohomological functor in the rst place. This is so tedious I wont even make
you do it as an exercise; rather, Ive just asked you to list which compatibilities need to be
checked in the rst place, which is already a nontrivial eort.
Examples
Here are some possibly familiar examples of derived functors. Some of these admit reasonable
explicit computations; see exercises.
For X ModR , X is a right exact covariant functor from ModR to ModR , hence a left
op
i
exact covariant functor from Modop
R to ModR . The derived functors are called Tor (X, ).
Proposition. For X ModR , the following are equivalent.
(a) X is at.
(b) Tori (X, Y ) = 0 for any i > 0 and any Y ModR .
(c) Tor1 (X, Y ) = 0 for any Y ModR .
Proof. Given (a), the functor X is exact, so its derived functors are zero, proving (b).
Given (b), (c) is trivial. Given (c), for any short exact sequence 0 A B C 0, we
get a long exact sequence
0 X A X B X C Tor1 (X, A) = 0
so X A is exact, proving (a).
This is of course a totally general argument: if F is a left exact covariant functor, then
F is exact i Ri F = 0 identically for all i > 0 i R1 F = 0 identically.
Given that the tensor product is symmetric, one would like to identify Tori (X, Y ) with
Tori (Y, X). However, the denition of Tor is asymmetric, so this takes a bit of thinking
(which Ill do using the dual language of projective resolutions and homology and lower
indices, but you can switch back if you like). Before starting, note that at least the fact that
Tori (X, Y ) is functorial in X (not just in Y ) is clear from the universal property of universal
cohomological functors.
Let P and Q be projective resolutions of X and Y , respectively. Then we have a double

complex
..
..
..
.
.
.
P1 Q1
P0 Q1
P1 Q0
P0 Q0
P1 Y
P0 Y
Q1
X Q0
X Y
in which the homology of the bottom row computes Tori (X, Y ), the homology of the right
column computes Torj (Y, X), and the other rows and columns are exact.
It is now a diagram chase to check that we have canonical isomorphisms Tori (Y, X)
=
Tori (X, Y ). For instance, say I start with a class in Tor1 (X, Y ) represented by x X Q1 .
Lift x to P0 Q1 , then push to P0 Q0 . The result maps to 0 in X Q0 , so lifts to P1 Q0 ;
push to P1 Y to get a class in Tor1 (Y, X). (This is really an example of a spectral sequence;
more on those a bit later.)
Corollary. Let 0 A1 A2 A3 0 be an exact sequence of R-modules with A3 at.
Then for any R-module M, 0 M A1 M A2 M A3 0 is again exact.
Proof. We have a long exact sequence
Tor1 (M, A1 ) M A1 M A2 M A3 0
but the left term can be identied with Tor1 (A1 , M), which vanishes because A1 is at.
The example of Tor is particularly important in algebraic geometry because of Serres
intersection multiplicity formula. Let X be a regular excellent scheme, let Y, Z be two
integral closed subschemes dened by the ideal sheaves I, J , and let x be the generic point
of a component of Y Z. The nave intersection multiplicity of Y and Z at x is
OY Z,x = OX,x /(IJ )x ,
and this gives the correct answer when dim(X) = 2, dim(Y ) = dim(Z) = 1 (meaning the
answer that makes Bezouts theorem work) but not in general. Serre found that the right
multiplicity is
(1)i lengthOX,x ToriOZ,x (OX,x /Ix , OX,x /Jx ).

i
It was an open question for a long time to give a geometric interpretation of the Tor
contributions in this formula; such an interpretation was recently provided by Jacob Lurie
using derived algebraic geometry. (Roughly speaking, one replaces rings by certain topo
logical rings before applying Spec; the intersection multiplicity then appears as the Euler
characteristic of the derived schematic intersection.)
A similar example occurs using the bifunctor Hom, except that it is really a bifunctor
from C op C. (Here C can be any abelian category, not just ModR .) Anyway, the right
derived functors of Hom(X, ) are called Exti (X, ), and they also occur as derived functors
of Hom(, Y ) by the double complex argument (with arrows appropriately reversed).
One more important example: if G is a group (considered with the discrete topology, if
you must), let Z[G] be the group algebra of G with coecients in Z, i.e., additively the direct
sum gG Z[g] with Z-linear multiplication characterized by [g][h] = [gh]. Then the covariant
functor G : ModZ[G] ModZ computing G-invariants is left exact; its derived functors are
called group cohomology and denoted H (G, M). The covariant functor G : ModZ[G] ModZ
computing G-coinvariants (i.e., M maps to the quotient of M by g(m) m for all g G and
m M) is right exact; its derived functors are called group homology and denoted H (G, M).
These are actually special cases of the previous example, namely
ModZ[G]
H (G, M) = ExtiModZ[G] (Z, M),
H (G, M) = Tori
(More generally, one could replace Z with an arbitrary ring.)
10
(Z, M).
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009
Sheaf cohomology (updated 13 Apr 09)
In the previous lecture, we discussed the construction of derived functors for left exact
additive functors out of on an abelian category that has enough injectives. In this lecture,
we specialize to the case of the global sections functor for sheaves on a locally ringed space,
and thus obtain the denition of sheaf cohomology.
Having enough injectives
I thought I assigned this as homework, but apparently not, so here is the proof.
Lemma. The category Ab has enough injectives.
Proof. It has been assigned as an exercise that an abelian group G is injective if and only if
it is divisible, i.e., if the multiplication by n maps are surjective for all positive integers n.
It remains to show that every group G is isomorphic to a subgroup of a divisible abelian
group. For instance, write G = F/H where F is a free abelian group, then embed G into
(F Z Q)/H. If you want something more canonical, take F to be the free abelian group
generated by the elements of G, with the map G F taking each g G to the generator
of F indexed by g (a/k/a the adjunction morphism for the forgetful functor Ab Set).
There isnt quite as nice an argument for ModR because we dont have as simple a
description of the injective modules. One proof that ModR has enough injectives is assigned
as an exercise; another will be given using Grothendiecks criterion later in this lecture.
Categories of sheaves have enough injectives
Let X be a locally ringed space, let C be an abelian category, and let D be the category
of sheaves on X with values in C; then D is again an abelian category. However, in order
to use the denition of derived functors, we need to know that D has enough injectives,
i.e., that for any object A D, there exists a monomorphism A I with I injective. I
should certainly assume that C itself has enough injectives; but then how can we go about
constructing injective objects in D?
One method is to try to identify the injective objects in D, but that is a bit dicult,
even for C = Ab. Another method is to construct a large enough class of injective objects
using skyscraper sheaves. Let x X be a point and let G be an object of C. We may then
view G as a sheaf on the one-point topological space {x}; the skyscraper sheaf at x with
values in G, denoted ix (G) is the direct image of G along {x} X. Its sections are G on
any open set containing x and 0 otherwise; its stalks are G at all points in the closure of x
and 0 elsewhere.
If we assume that C has colimits, then we can use the adjointness property between direct
and inverse image to assert that
HomShC (X) (F, ix (G)) = HomC (Fx , G).
In particular, if G is injective in C, then ix (G) is injective in ShC (X). (Remember that this
means that Hom(, ix (G)) is an exact functor.)
If we assume that C also has arbitrary products, it becomes easy to guess how to embed
an arbitrary sheaf F into an injective: for each x X, use the hypothesis that
C has enough
injectives to construct a monomorphism Fx Gx , and then F embeds into xX ix (Gx ).
Namely, for U X open, the map
F(U )
ix (Gx ) (U) =
Gx =
Fx
xX
xU
xU
takes a section s to the tuple (sx ) of its germs. This is a monomorphism by the sheaf axiom.
Moreover, an arbitrary product of injective objects is injective.
In fact, something even stronger is true, and the proof is similar; see Hartshorne, Propo
sition III.3.2. (This reproduces the previous statement by taking the sheaf of rings to be a
constant sheaf.)
Proposition. Let (X, OX ) be a ringed space. Then the category of sheaves of OX -modules
has enough injectives.
Beware that if X is a locally ringed space, it does not follow that the category of qua
sicoherent sheaves of OX -modules has enough injectives. (However, this is true for ane
schemes because ModR has enough injectives.)
More on having enough injectives
One can also establish that the category of sheaves has enough injectives using a very general
criterion introduced by Grothendieck in Sur quelques points...
Theorem. Let C be an abelian category satisfying the following conditions.
(a) C admits arbitrary (small) direct sums.
(b) Suppose we are given a monomorphism X Y in C, a totally ordered set I, and an
increasing family of subobjects Yi of Y indexed by i I. (This last means that we are
given a monomorphism Yi Y for each i I, and a monomorphism Yi Yj for each
i, j in I with i j, such that Yi Yj Y agrees with Yi Y .) Then inside Y ,
Yi X =
(Yi X) .
i
In other words, forming the direct limit of the Yi commutes with taking the bred product
with X over Y . (The direct limits on both sides exist by (a).)
2
(c) There exists an object U C such that for any monomorphism X Y which is not an
epimorphism, the map Hom(U, X) Hom(U, Y ) is also not an epimorphism. (That
is, there is a map U Y not factoring through X. Grothendieck calls U a generator
of C.) Also, the class of isomorphism classes of monomorphisms into U is small (this
is automatic if C admits a forgetful additive functor to Ab).
Then C has enough injectives.
Before proving this, I should point out that these conditions are suciently weak that
they are satised by ModR . Namely, (a) and (b) are obvious, while (c) holds by taking
U = R because then Hom(U, ) coincides with the forgetful functor to abelian groups. (It is
also possible to prove more directly that ModR has enough injectives, but never mind.)
I should also check a bit more carefully that these conditions are satised by the category
of sheaves of abelian groups on a locally ringed space. To check (a), note that if Fi is a family
of sheaves on X, then we may construct the direct sum by taking the sheacation of the
presheaf U i Fi (U). We may check (b) stalkwise. To check (c), we take U to be the
direct sum over open subsets V X of the pushforward jV (ZV ) of the constant sheaf on
V with values in Z. The point is that for any sheaf G,
Hom
jV (ZV ), G =
Hom(jV (ZV ), G)
V
Hom(ZV , G|V )
(V, G).
You can also use a direct sum over points, as in the previous section.
Lemma. Under the conditions of the theorem, an object M C is injective if and only if
for any monomorphism V U into the generator, every morphism V M extends to a
morphism U M .
Proof. Exercise.
Proof of the theorem. We make a rst approximation to the desired construction as follows.
Let M C be any object. Let I(M) be the set of isomorphism classes of pairs (T, t), where
T U is a monomorphism and t : T M is a morphism. Consider the map
(T,t)I(M ) T M U I(M )
in which the factor of T coming from a pair (T, t) maps to M via T , maps to the (T, t)-th
factor of U I(M ) via the monomorphism T U, and maps to the other factors of U I(M ) via the
zero map. Let M U I(M ) C(M ) be the cokernel of that map, and let f (M) : M C(M)
be the composition of this with the injection of M into the rst factor of M U I(M ) . One
checks using (b) that this is a monomorphism.
3
By construction, we have a monomorphism f (M) : M I(M) such that for any

monomorphism T U and any morphism T M, we can extend T M I(M)
to a morphism T I(M ). This doesnt quite solve our problem because M = I(M). The
trick is to repeat this construction using transnite induction. Namely, start with M0 = 0.
For any nonlimit ordinal i, put Mi+1 = f (Mi ); for any limit ordinal, let Mi be the direct
limit of Mj over j < i. There must then be a least ordinal k such that the cardinality of k is
strictly greater than the cardinality of the number of isomorphism classes of monomorphisms
into U . Then for any morphism T Mk , the sequence of inverse images of the Mj in T for
j < k must stabilize; that is, T maps into Mj for some Mj . Then this extends to a map of
U into Mj+1 , so Mk satises the condition of the previous lemma.
4 Sheaf cohomology for topological spaces and ringed

spaces
Let C be an abelian category admitting arbitrary products and colimits, and having enough
injectives. We have just shown that for any topological space X, Sh C (X) also has enough
injectives. We may now dene the sheaf cohomology functors H i : ShC (X) C to be the
right derived functors of the left exact functor (X, ) : ShC (X) C. In particular, H 0 (X, F )
is just another notation for F(X) or (X, F).
If (X, OX ) is a ringed space, we can also dene derived functors of (X, ) directly on the
category of sheaves of OX -modules. The fact that these coincide with the H i requires some
justication, but its not hard. One way to see it is to note that the H i , when restricted to
the category of OX -modules, return O(X)-modules, then argue that these are an eaceable
cohomological functor and so coincide with the derived functors.
Another argument is to use some acyclic objects which are not injective, remembering
that we may use resolutions with these objects to compute derived functors. Here is a cheap
supply of acyclic objects. A sheaf F on X is asque (or abby) if for any inclusion V U of
open sets, the restriction map F (U ) F(V ) is surjective. For instance, if X is an irreducible
topological space, then any constant sheaf is asque. (Reminder: for C C, the constant
sheaf C X on any space X is the sheacation of the constant presheaf U C.) However,
if X = R with the usual topology then the sections of C X on X are C but on R \ {0} are
C C, so C X is not asque unless C = 0.
Lemma. For any ringed space (X, OX ), any injective OX -module is asque. In particular
(by taking OX = ZX ), any injective sheaf of abelian groups on X is asque.
Proof. (Compare Hartshorne, Lemma III.2.4.) Let I be an injective OX -module. For any
open subset U of X, let OU denote the extension by zero of OX |U to X, i.e., the sheacation
of the presheaf assigning V to OX (V ) if V U and 0 otherwise. Note that it has stalks OX,x
for x U and 0 otherwise. (This diers from the direct image under the inclusion U X,
which has nonzero sections on any open set meeting V .)
For V U an inclusion of open sets, we get a monomorphism OV OU of sheaves of

OX -modules. Since I is injective, this gives a surjection Hom(OU , I) Hom(OV , I). But
Hom(OU , I) = I(U) and Hom(OV , I) = I(V ), so I is asque.
Proposition. Let F be a asque sheaf of abelian groups on a topological space X. Then
H i (X, F ) = 0 for all i > 0.
Proof. The argument is a classic example of dimension shifting. Embed F into an injective
sheaf I, and put G = I/F . Using the fact that F is asque, we nd (exercise)
0 H 0 (X, F ) H 0 (X, I) H 0 (X, G) 0
is exact. Using this, the long exact sequence in cohomology associated to
0 F I G 0,
and the fact that I is acyclic, we nd that H 1 (X, F) = 0 and
H i (X, F)
= H i1 (X, G)
(i > 1).
Since F is asque, and I is injective and hence asque by the previous lemma, it follows
that G is asque (exercise). Hence by the induction hypothesis, we may deduce H i (X, F)
=
H i1 (X, G) = 0 for i > 1.
Sheaf cohomology and topological cohomology
If you know some topology, you might appreciate the following relationship between sheaf
cohomology and the usual cohomology of topological spaces. (If not, pretend that the coho
mology of the constant sheaf ZX is the denition of topological cohomology of a space X,
then skip directly to the next section.)
Theorem. Let X be a locally contractible topological space. Then the sheaf cohomology
of X with coecients in the constant sheaf ZX is canonically isomorphic to the singular
cohomology of X.
Recall that X is contractible if there is a continuous map f : X [0, 1] X with
f (x, 0) = x for all x X, and f (x, 1) = f (y, 1) for all x, y X; it is locally contractible
if each point has a basis of contractible neighborhoods. For instance, all manifolds and
CW-complexes are locally contractible.
The singular n-chains in X, collectively denoted Cn (X), are formal nite Z-linear com
binations of continuous maps : Tn X, where Tn denotes the standard n-simplex.
The boundary map : Cn (X) Cn1 (X) takes each simplex to its signed boundary,
i.e., if Tn has vertices e0 , . . . , en , then for i = 0, . . . , n, you take (1)i times the restric
tion to the subsimplex omitting ei . These form a homologically graded complex; putting
5
C n (X) = HomZ (Cn (X), Z) gives the singular n-cochains, which form a cohomologically
graded complex.
Let C n (X) be the sheacation of the presheaf U C n (U); it is straightforward to check
that in fact C n (X) is asque. Using the hypothesis that X is locally contractible (so that we
can check exactness on stalks by running over a basis of contractible neighborhoods), one
checks that
0 C 0 (X) C 1 (X)
is a resolution of ZX . We may thus compute H i (ZX ) by computing global sections of this
complex.
It remains to check that the natural map
C (X) (X, C (X))
is a quasi-isomorphism of complexes. To see this, let us x an open cover {Ui } of X, and let
D (X) be the set of singular cochains only dened on simplices contained in some Ui . One
then reduces to the following assertion.
Lemma. The restriction C (X) D (X) is a homotopy equivalence, with a quasi-inverse
dened as follows. Given a cochain in D (X), extend to a cochain on X by mapping each
simplex not contained in some Ui to 0.
This is a standard if tedious calculation; see Spaniers Algebraic Topology.
Cech
cohomology
From the previous section, we know that if X is a contractible topological space, then Z X
is an acyclic sheaf (because the singular cohomology of X vanishes). This can be used to
compute the cohomology of X in terms of the combinatorics of a good cover, i.e., an open
cover {Ui } of X in which each nite intersection is contractible. (You may have read about
this in Bott and Tu, Dierential Forms in Algebraic Topology.) We will use the same idea
later in order to compute the cohomology of quasicoherent sheaves on schemes.
Let X be a topological space, and let U = {Ui }iI be an open cover of X (i.e., each x X
appears in only nitely many Ui ). For convenience, let us assume the set I is equipped with
a total ordering (this helps straighten out some sign conventions). For each nite subset J
of I, put UJ = iJ Ui , with the convention that U = X.
Let F be a sheaf of abelian groups on X. We dene the Cech

complex of F dened by
j
the open cover {Ui } as follows. For j 0, let C (U, F) be the direct product of (F, UJ )
over all (j + 1)-element subsets J of I. The dierential dj : C j (U, F) C j+1 (U, F) is dened
as follows: for = (J ) C j (U, F ), we have
j
d ()J =
j+1
(1)k ResUJ{ik },J (J{ik } )
k=0
J = {i0 ij+1 }.
For instance, if there are only two open sets U1 and U2 , then you have
0 (F , U1 ) (F, U2 ) (F, U1 U2 ) 0
where the nontrivial map is the dierence between the two restrictions. The signs were
rigged up to make sure that this is indeed a complex: the point is that if you pull ij and ik
out of a set J in on order and multiply the two resulting signs, you get the opposite sign as
if you pulled them out in the opposite order.
It is an easy exercise to check that this gives a complex, and continues to do so if you
insert (X, F ) in front (with the individual restriction maps to C 0 (U, F).
It is convenient to also work with a sheaer analogue of this construction. Let Cj (U, F)
be the direct product of jJ F|UJ over all (j + 1)-element subsets J of I, where jJ : UJ X
is the inclusion. The global sections of this are just C j (U, F).
Lemma. The complex
0 F C0 (U, F) C1 (U, F)
is exact.
Proof. (Compare Hartshorne Lemma III.4.2.) It suces to check exactness on stalks. Pick a
point x X; we may then replace X by some Ui containing x. In this case, we can construct
an explicit chain homotopy k between the identity map and the zero map. Its action can be
described as follows: given a j-cochain = (J ), you make a (j 1)-cochain by identifying
J with a section of UJ\{i} whenever i J, and discarding the J whenever i
/ J. To do
this correctly, you need to add some signs; Ill leave this to the Hartshorne reference.
(U, F ) = h (C (U, F )). These do not form a cohomological functor if we x
We write H
the choice of U. As noted in Hartshorne Caution 4.0.2, this is clear for the trivial cover of X
by itself because the global sections functor is not exact. However, they do at least give the
right answer in the asque case. (They also give the correct answer in degree 0 no matter
what the cover, by the sheaf axiom!)
i (U, F) = 0 for i > 0.
Lemma. If F is asque, then H
Proof. In the resolution
0 C0 (U, F) C1 (U, F)
of F , each term is again asque and hence acyclic for sheaf cohomology. If we then take
global sections and compute cohomology of the resulting complex, on one hand we just get
i (U, F ). On the other hand, by the acyclic resolution theorem, we are also computing
H
H i (X, F ), which vanishes for i > 0.
On the other hand, suppose V is a renement of U, i.e., a new covering {Vj }jJ equipped
with a map : J I of index sets such that Vj U(j) for all j J. Then we get a
restriction morphism
(U, F) H
(V, F).
H
7
Using renements, the coverings of X form a direct system, so (since we are working with
abelian groups, which admit colimits) we can form the direct limit
(X, F ) = lim H
(U, F).
H
Under certain circumstances, we can show that this computes sheaf cohomology. This wont
cover the case of schemes, but well deal with that separately later.
Theorem. Suppose that X is paracompact, i.e., X is Hausdor and every open covering
(X, F) form a cohomological functor which
renes to a locally nite subcovering. Then the H
is eaceable, hence universal, hence canonically isomorphic to H i (X, F). In particular, for
(U, F) H (X, F) functorial in F.
any particular covering U, we obtain a morphism H
Proof. Since X is paracompact, we need only take the direct limit over locally nite coverings.
In that case, the functors
F lim C (U, F)
are exact (exercise). Given that, we may apply them to a short exact sequence and then take
the long exact sequence in cohomology to get the connecting homomorphisms. Eaceability
holds because each F embeds into a sheaf which is injective, hence asque, hence acyclic for
(X, ) by an earlier lemma.
H
All well and good, but what we really want to know is, when can we use the Cech
complex associated to a particular complex U to compute the cohomology of F ? Here is a
useful answer in practice. We say the cover U is good for F if for each J, F|UJ is acyclic.
(No hypothesis on X needed.)
(U, F) H (X, F) are iso
Theorem (Leray). If U is good for F, then the morphisms H
morphisms. That is, the Cech complex C (U, F) computes the sheaf cohomology of F.
Proof. As in the proof that Cech

cohomology vanishes for asque sheaves, it would suce
to show that the resolution
0 C0 (U, F) C1 (U, F)
is acyclic. Unfortunately, we cant directly conclude this from the fact that each F |UJ is
acyclic, because the direct image jJ functor need not be exact.
So instead, we argue by dimension-shifting. The claim is evident for i = 0 by the sheaf
axiom. Given the claim for all indices less than i, embed F into an injective sheaf I, and
let G be the quotient:
0 F I G 0.
On each UJ , F and I are acyclic, so G is as well by the long exact sequence in cohomology.
Moreover, we have short exact sequences
0 (UJ , F) (UJ , I) (UJ , G) H 1 (UJ , F ) = 0.
8
This means that not only does this short exact sequence of sheaves give rise to a long exact
i (U, ) (because
sequence for the H i (X, ), it also gives rise to a long exact sequence for the H
we get a short exact sequence of Cech complexes). We thus have a commuting diagram with
exact rows:
H
H
H
i1 (U, I)
i1 (U, G)
i (U, F)
i (U, I)
H
H i1 (X, I)
H i1 (X, G)
H i (X, F)
H i (X, I)
in which the corners are zero (because injective implies asque, which implies both the
ordinary and Cech

cohomologies vanish). So we transfer our question about F at index i to
a question about G at index i 1, which we know by the induction hypothesis.
This has practical applications outside of algebraic geometry: you can now use good
covers to compute the singular cohomology of ordinary topological spaces! The analogue of
this in algebraic geometry will appear next, when we start computing the cohomology of
quasicoherent sheaves; the analogue of contractible open subsets in the topological case will
turn out to be the ane schemes, on which quasicoherent sheaves will be acyclic.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Cohomology of quasicoherent sheaves (updated 25 Apr 09)
There is one more fundamental theorem about ane schemes! Here it is.
The theorem, and a bogus proof
Lets start with the statement of the fourth fundamental theorem of ane schemes.
Theorem. Let X be an ane scheme and let F be a quasicoherent sheaf on X. Then
H i (X, F ) = 0 for i > 0, that is, F is acyclic for sheaf cohomology.
Here is how not to prove this theorem.
Bogus proof of the theorem. Put X = Spec A. From the earlier fundamental theorems of
for some A-module M. Since ModA has
ane schemes, we know we can write F = M
enough injectives, we can nd a monomorphism M I with I an injective A-module. Put
N = I/M. Again by the earlier fundamental theorems of ane schemes, we know that the
exact sequence
0 M I I/M 0
of A-modules is precisely what you get by taking global sections of the exact sequence
0.
I I/M
0M
)
So in the long exact sequence in cohomology, the connecting homomorphism into H 1 (X, M
) is
is zero. On the other hand, H i(X, I) = 0 for all i > 0 since I is injective, so H 1 (X, M
so we may prove the
)
forced to be zero. Moreover, for i > 1, H i(X, M
= H i1 (X, I/M),
higher vanishing by dimension shifting.
Whats wrong with this proof? The problem is that while the injectivity of I in ModA
implies the injectivity of I in the category of quasicoherent OX -modules, it does not imply
injectivity in the category of arbitrary OX -modules, or of sheaves of abelian groups on X.
In particular, it is unclear whether injectivity of I implies that I is asque. One can at least
show that I is basically asque, in that the restriction from (X, I) = I to (D(f ), I) = If
is surjective, but this isnt really enough to do anything useful.
There are two ways to x this. One way (used in Hartshorne, and also in the book by
Ueno that I recommended earlier) is to restrict attention to noetherian rings, and then prove
that an injective module does indeed give rise to a asque sheaf. The other way (used in
EGA) is to compute with Cech

cohomology instead of sheaf cohomology, so that you can
deal only with nite covers by distinguished opens. Thats what Ill do here.
First, however, I should mention that there is an easy argument to show that H 1 vanishes.
The following is close to the third fundamental theorem of ane schemes; see Hartshorne
Proposition II.5.6 for the proof. (Since I wont use this to prove the theorem, you may
instead view it as a corollary of the theorem.)
1
Lemma. Let X = Spec A be an ane scheme. Let

0 F1 F F2 0
be an exact sequence of OX -modules such that F1 is quasicoherent (but dont assume anything
about the other two). Then the sequence
0 (X, F1 ) (X, F ) (X, F2 ) 0
is exact.
This implies that the connecting homomorphism H 0 (X, F2 ) H 1 (X, F1 ) is zero, so
H (X, F1 ) injects into H 1 (X, F ). If we then choose F to be injective, we deduce H 1(X, F1 ) =
0.
1
Applications
Before proving the theorem, let me mention some corollaries. First, from the Cech
cohomol
ogy discussion, we deduce the following.
Corollary. Let X be a scheme and let U = {Ui }iI be an open cover of X. Suppose that for
each nite subset J I, the intersection UJ = jJ Uj is ane. Then for any quasicoherent
sheaf F on X, the sheaf cohomology of F is computed by the Cech

cohomology for the cover
U; that is,
i(U, F ).
H i (X, F ) = H
Recall that inside a separated scheme, the intersection of any two open anes is again
ane. We thus have the following; Ill illustrate next time by using this to compute the
cohomology of the sheaves O(n) on projective space.
Corollary. Let X be a separated scheme and let U = {Ui }iI be an open ane cover of X.
Then for any quasicoherent sheaf F on X,
i(U, F ).
H i (X, F ) = H
Here is an even more specialized corollary, which in itself is not so useful. I mention it
because I will prove this directly and use it as a lemma to prove the whole theorem.
Corollary. Let A be a ring and choose f1 , . . . , fn A which generate the unit ideal. Let
U be the open cover of X = Spec A by D(fi ) for i = 1, . . . , n. Then for any A-module M,
0 (U, M
) = M and H
i (U, M
) = 0 for i > 0.
H
A correct proof of the theorem
Following Grothendieck (and I think Serre before him, in the context of varieties), we will
prove the last corollary rst, and then use that to prove the theorem. So our rst order of
business is to show that the complex
) C 1 (U, M
)
0 M C 0 (U, M
is exact. Remember that this complex came from the complex of sheaves
C0 (U, M
) C1 (U, M
)
0M
by taking global sections. We proved in the Cech

cohomology lecture that this sequence
of sheaves is exact (by computing at stalks). Moreover, each of the constituent sheaves is
quasicoherent, for the following reason. Each sheaf in the sequence equals a direct sum of
|U ) for U an intersection of some of the Uj . In particular,
sheaves each of the form jU (M
each such intersection has the form D(g) for some g A. But this sheaf is simply the
quasicoherent sheaf associated to the A-module Mg .
Since we have an exact sequence of quasicoherent sheaves, taking global sections gives us
an exact sequence of A-modules. This proves the corollary. So now we know that for any
nite cover U of Spec A by distinguished opens,
0 (U, M
) = M,
H
i (U, M
) = 0 (i > 0).
H
The same holds if we take the direct limit over nite covers by distinguished opens. However,
this gives the same result as taking the direct limit over all open covers because any cover
can be rened to a nite cover by distinguished opens. We conclude that
0 (X, M
) = M,
H
i (X, M
) = 0 (i > 0);
H
although the theorem that says that the direct limit Cech
cohomology also computes sheaf
cohomology doesnt apply (because X is not Hausdor), one can still show that this implies
) = M,
H 0 (X, M
) = 0 (i > 0)
H i (X, M
using the following theorem of Cartan, applied with B being the collection of distinguished
open anes.
Theorem (Cartan). Let X be a topological space. Let B be a basis of X closed under
i (U, F ) = 0 for
pairwise intersections. Let F be a sheaf of abelian groups on X such that H
i (X, F ) is naturally isomorphic to H i (X, F ) for all i 0.
all U B. Then H
We will prove this in the next section. It can also be proved using spectral sequences;
see the optional handout.
Comparison of Cech
and sheaf cohomology
Before proving Cartans theorem, here is a lemma which generalizes a fact we already know
about asque sheaves.
Lemma. Let X be a topological space. Let F be a sheaf of abelian groups on X such that
1 (X, F ) = 0. Then for any short exact sequence
H
0F GH0
of sheaves,
0 (X, F ) (X, G) (X, H) 0
is exact.
Proof. (proof suggested by Fucheng Tan) We need only check surjectivity on the right. Let
s (X, H) be any section; let U = {Ui }iI be an open cover of X such that for each i I,
s|Ui lifts to a section ti (Ui , G). For i, j I, put
uij = ti |Ui Uj tj |Ui Uj (Ui Uj , G).
Since uij has zero image in (Ui Uj , G), we may also view as an element of (Ui Uj , F ).
With this convention, we see that the uij form a Cech

1-cocycle of F for the open cover U.
Before proceeding, note that there is a natural way to replace the above data for one
cover U with a renement V = {Vj }jJ . Namely, the renement comes by denition with a
map : J I such that Vj U(j) for each j. To pass from U to V:
replace the collection of the ti for i I with the collection of the t(j) |Vj for j I;
replace the collection of the tij for i, j I with the collection of the u(k)(l) |Vk Vj for
k, l I.
To avoid excess notation, we will speak of replacing U by a renement which will also be
labeled U.
1 (X, F ) = 0 by hypothesis, we can replace U by a renement in such a way that
Since H
uij become a Cech

1-coboundary. This means that there are sections vi (Ui , F ) such that
vi |Ui Uj vj |Ui Uj = uij
(i, j I).
For i I, we now form

wi = ti vi (Ui , G).
These sections have the property that on one hand, the image of wi in (Ui , H) equals s|Ui
(since vi , having come from F , maps to zero in H), and on the other hand,
wi |Ui Uj wj |Ui Uj = (ti |Ui Uj vi |Ui Uj ) (tj |Ui Uj vj |UiUj )
= (ti |Ui Uj tj |Ui Uj ) (vi |Ui Uj vj |UiUj )
= uij uij = 0.
Hence the wi glue to a section w (X, G) lifting s, as desired.
4
Proof of Cartans theorem. The claim is true for i = 0 because of the sheaf axiom. We use
this as a basis for induction on i, using dimension shifting. Assume that for some i > 0, the
claim is true for every value less than i. Choose a short exact sequence
0F GH0
with G asque. By the previous lemma, for any U B,
0 (U, F ) (U, G) (U, H) 0
is exact. Let U = {Ui }iI be an open cover of X by basic open sets. Since B is closed under
pairwise intersections, it follows that
0 C . (U, F ) C . (U, G) C . (U, H) 0
is an exact sequence of complexes. Since every open cover can be rened to an open cover
by basic opens, taking direct limits over all covers is the same as taking direct limits over
basic open covers, which means that
0 C . (X, F ) C . (X, G) C . (X, H) 0
is again an exact sequence of complexes. The same holds if we replace X by any basic open
set U, by the same reasoning.
We now take the long exact sequence in cohomology associated to this short exact se
quence of complexes, and compare it to the long exact sequence in sheaf cohomology. We
get the diagram
H
i1 (X, G)
H i1 (X, G)
i1(X, H)
H
i(X, F )
H
i1
i(X, G)
H
(X, H)
H (X, F )
H (X, G)
i (X, G) = H i (X, G) = 0 because G is asque. If we replace X by

We rst notice that H
i1 (U, H)
a basic open set U and then look at the top sequence, we see that for i > 1, H
is sandwiched between two zero groups, so it is also zero. That is, H also satises the
hypothesis of the theorem.
We may now argue by dimension shifting as follows. The rst vertical arrow is an
isomorphism (for i = 1 this holds by the sheaf axiom, otherwise both groups vanish), the
second vertical arrows is an isomorphism by the induction hypothesis, and the fourth vertical
arrow is the zero map between zero groups. The ve lemma thus implies that the third arrow
is an isomorphism.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Cohomology of projective space (updated April 14)
Using Cech
cohomology, we now give Serres computation of the cohomology of the
twisting sheaves on projective space, and deduce from it Serres niteness theorem. See also
Hartshorne III.5 (except that you can ignore the last paragraph at the bottom of page 227,
which is irrelevant).
Finitely generated sheaves
Let (X, OX ) be a locally ringed space. An OX -module F is nitely generated if for each
point x X, there exist an open subset U of X containing x, a nonnegative integer n, and
n
a surjection OX
|U F |U of OX -modules. (In other words, F is locally generated by nite
collections of sections.)
Lemma. Let A be a ring and let M be an A-module. Then M is nite as an A-module if
is nitely generated.
and only if the quasicoherent sheaf M
is nitely generated. Conversely,
Proof. It is clear that niteness of M implies that M
is nitely generated. We can then cover X with distinguished opens Xi = D(fi )
suppose M
|X = M
f is generated by nitely many sections mij /f hj . Since X is
on each of which M
i
i
i
quasicompact, we need only nitely many i, and the resulting mij generate M because they
do so stalkwise.
Beware that Hartshorne says a sheaf is coherent if it is quasicoherent and nitely gen
erated. This is not the correct denition, but it does agree with the correct denition on
a locally noetherian scheme, which is the only place he ever uses it. Ill give the proper
denition of a coherent sheaf later.
Odds and ends
We will make frequent use of the following fact.

Lemma. Let f : Z X be a closed immersion of locally ringed spaces, and let F be a sheaf
of abelian groups on Z. Then there are canonical isomorphisms
H i (Z, F ) H i(X, f F )
(i 0).
Proof. We already know that f preserves asqueness (previous exercise). We can also see
that f is exact by its behavior on stalks:
Fx x Z
(f F )x =
0
x
/ Z.
1
Finally, we note that we have a canonical isomorphism at the level of H 0 . If we then

start with a asque resolution of F , we can either push forward (obtaining another asque
resolution) and then take global sections, or take global sections directly, to obtain the same
complexes, and in particular the same cohomology.
This will very often allow us to carry out induction arguments on dimension, by setting
up a short exact sequence of sheaves on X in which one member is the direct image of a
sheaf on Z.
Lemma. Let X be a noetherian topological space, and let (Fj ) be a direct system of abelian
sheaves. Then the natural functoriality maps
lim H (X, Fj ) H (X, lim Fj )
(coming from the maps Fj lim Fj ) are isomorphisms.
Proof. Fix the index set for the direct system. Then both sides are cohomological functors
on the category of direct systems of abelian sheaves for that index set, and the noetherian
hypothesis forces the H 0 terms to match (earlier exercise), so it suces to check that they
are both eaceable. For that, it suces to observe that there is a functorial way to embed
any sheaf of abelian groups on X into a asque sheaf; for instance, see Hartshorne exercise
II.1.16(e).
The result
Before stating the theorem, I need to make one general observation. Let F , G be sheaves
of OX -modules on a scheme X, with F quasicoherent and at. I claim there is a natural
homomorphism
H 0 (X, F ) OX H r (X, G) H r (X, F OX G)
of (X, OX )-modules for any nonnegative integer r. It comes from the facts that:
both sides are cohomological functors in G (using the fact that F is at on the right
side, and the fact that H 0 (X, F ) is at over O(X) on the left side);
the left side is eaceable and hence universal;
there is a natural map for r = 0.
We can also compute this map using Cech

complexes.
Theorem (Serre). Let A be any ring, x an integer r 1, and put X = PrA and S =
A[x0 , . . . , xn ].
(a) The natural map
S
H 0 (X, OX (n))
nZ
is an isomorphism of graded S-modules.

2
(b) For 0 < i < r and n Z, we have H i (X, OX (n)) = 0.

(c) We have H r (X, OX (r 1))
= A.
(d) The natural A-bilinear map
H 0 (X, OX (n)) H r (X, OX (n r 1)) H r (X, OX (r 1))
=A
is a perfect pairing of nitely generated free A-modules, for each n Z. (That is, each
side is isomorphic to Hom of the other side into A; in particular, they have the same
rank.)
(e) For i > r, we have H i (X, OX (n)) = 0.
Proof of the theorem. It is enough to compute the cohomology of the sheaf F = nZ OX (n)
and keep track of the grading by n. (This can be seen by applying the previous lemma, but
this forces a noetherian hypothesis. But it is also clear without such a hypothesis, because
the Cech
cohomology computation we are about to do commutes with the direct sums.)
For starters, recall that we checked part (a) earlier; see for instance Hartshorne Proposition
II.5.13.
We use the obvious Cech

resolution by the D+ (xi ) for i = 0, . . . , r. Since the complex
vanishes above degree r, we immediately get (e). (This also follows from a general theorem of
Grothendieck; see Hartshorne Theorem III.2.7.) To compute H r (X, F ), we need the cokernel
of
r
r1
d
:
Sx0 xk1 xk+1 xr Sx0 xr .
k=0
View Sx0 xr =
as the free A-module generated by the monomials xe00 xerr
with e0 , . . . , er Z. The image under dr1 of the k-th factor of the product is precisely the
span of the monomials with ek 0. Hence H r (X, F ) is the free A-module generated by
the xe00 xerr with e0 , . . . , er < 0, graded by degree. In particular, in degree r 1, we see
1
exactly one monomial x1
0 xr , proving (c).
To see (d), we must make the pairing explicit. First, note that there is nothing to check
if n < 0, since then H 0 (X, OX (n)) = Sn = 0 obviously, and H 0 (X, OX (n r 1)) = 0
because there are no monomials of degree greater than r 1 with all exponents negative.
So assume hereafter n 0. If we identify H 0 (X, OX (n)) with the A-span of the monomials
in x0 , . . . , xr (with nonnegative powers) of degree n, then the pairing with H r (X, F ) (the
1
A-span of the monomials of degree n r 1 in x1
0 , . . . , xr ) is to simply multiply together
1
1
and throw away everything except the term x0 xr . This implies (c).
It remains to prove (b), which we do by induction on r. The base case is r = 1, for which
there is nothing to check because 0 < i < r is impossible. Before running the induction, we
note that if we localize the Cech

complex by inverting xr , we get the corresponding Cech
complex on the open set D+ (xr ), which is ane. So the localized Cech
complex must be
acyclic since it computes the cohomology of a quasicoherent sheaf on an ane scheme. On
the other hand, localizing in xr is exact, so it commutes with taking cohomology; that is,
A[x
0 , . . . , xr ]
the localization H i (X, F )xr = 0 for i > 0. In other words, every element of H i (X, F ) is
annihilated by some power of xr .
It thus suces to show that for 0 < i < r, multiplication by any power of xr is injective
on H i(X, F ); it also suces to check multiplication by xr itself. Look at the exact sequence
x
0 S(1) r S S/(xr ) 0
of graded S-modules. Writing H
for the hyperplane xr = 0 and j : H X for the
= Pr1
A
inclusion, this sheaes to
0 F (1) F nZ (j OH )(n) 0.
Lets take the long exact sequence in homology. In degree 0 we get back our original sequence,
so the connecting homomorphism into H 1 (X, F (1)) is zero. That (which holds even in
the base case) plus the induction hypothesis, which implies that H i (X, n (j OH )(n)) =
H i (H, n OH (n)) = 0 for 0 < i < r 1, gives us the bijectivity of multiplication by xr on
H i (X, F ) for i = 0, . . . , r 2 and injectivity for i = r 1. This is enough to get H i (X, F ) = 0
for 0 < i < r.
Finiteness of cohomology on projective schemes
Using the previous calculation, Serre was able to derive a powerful niteness and vanish
ing theorem for cohomology on projective schemes. First, we need another result of Serre
(Hartshorne II.5.17 except without the noetherian hypothesis).
Theorem. Let A be a ring, let X PrA be a closed immersion for some r 1, and let
OX (1) be the pullback of the twisting sheaf. Let F be a nitely generated quasicoherent sheaf
on X. Then there exists an integer n0 such that for all n n0 , F (n) is generated by a nite
number of global sections.
Proof. By replacing F by its direct image, we reduce to the case X = PrA itself. For
i for some nitely generated module Mi over Bi =
i = 0, . . . , r, we have F |D+ (xi ) = M
A[x0 /xi , . . . , xr /xi ]. For any s Mi , for n n0 for some n0 depending on s, xni s is a section
of F (n). For n0 large enough, we can lift a set of generators of each Mi to sections of F (n)
whenever n n0 ; this proves the claim.
Corollary. With notation as in the previous theorem, we obtain a surjection m
i=1 O(n) F
for some n Z.
For this I need a noetherian hypothesis, but only until we dene coherent sheaves.
Theorem (Serre). Let A be a noetherian ring, let X PrA be a closed immersion for some
r 1, and let OX (1) be the pullback of the twisting sheaf. Let F be a nitely generated
quasicoherent sheaf on X.
4
(a) The A-modules H i (X, F ) are nitely generated for i 0.

(b) There exists an integer n0 (depending on F ) such that for each i > 0 and n n0 ,
H i (X, F (n)) = 0.
Proof. We proceed by descending induction on i. For i > r, we have H i (X, F ) = 0 because
X admits a good over by at most r + 1 open anes.
By the previous corollary, we can write F as a quotient of some sheaf E which is a direct
sum of twisting sheaves. Let G be the kernel:
0GE F 0
Thanks to the noetherian hypothesis, we may conclude that G is also nitely generated. The
long exact sequence in cohomology gives:
H i (X, E) H i(X, F ) H i+1(X, G)
Given the claim for i + 1, we know that the right term is nitely generated as an A-module.
The left term is a sum of things of the form H i(X, OX (n)) for various n Z, and we already
computed those and saw that they were nitely generated as A-modules. Again since A is
noetherian, we can conclude that the middle module is nitely generated. This proves (a).
To get (b), twist by n and then again take the long exact sequence in cohomology:
H i (X, E(n)) H i (X, F (n)) H i+1 (X, G(n))
For n large, the right module vanishes by the induction hypothesis, while the left module
vanishes by the explicit calculation, so the middle group vanishes.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Hilbert polynomials and atness (revised 17 Apr 09)
See Hartshorne III.9 again.
Hilbert polynomials
Let k be a eld (not necessarily algebraically closed). Let j : X Prk be a closed immersion
for some r 1. Write OX (1) for the inverse image by j of the twisting sheaf O(1). Let F
be a nitely generated quasicoherent sheaf on X.
The Euler characteristic of F is dened as
(X, F ) =
(1)i dimk H i(X, F );
i=0
we know from Serres niteness theorem that each summand is nite, and we also know that
there are no terms in dimension greater than r. So this is indeed a well-dened integer.
Lemma. The Euler characteristic is additive in short exact sequences; that is, if
0F GH0
is exact, then
(X, G) = (X, F ) + (X, H).
Proof. Exercise in the long exact sequence in cohomology.
Corollary. If
0 F1 Fn 0
is an exact sequence of nitely generated quasicoherent sheaves, then
n
(1)i (X, Fi ) = 0.
i=1
Theorem. There exists a polynomial P (z) Q[z] such that

(X, F (n)) = P (n)
(n Z).
Moreover, the degree of P is at most the dimension of X.

Proof. By replacing F by j F , we may reduce to the case X = Prk . Also, changing the base
eld doesnt change any of the dimensions (e.g., by looking at Cech

cohomology; this is a
special case of the at base change theorem), so we may assume k is algebraically closed.
We proceed by induction on the dimension of the support of F . If that support is empty
(i.e., F is the zero sheaf), then obviously P (z) = 0 works.
1
Otherwise, form an exact sequence

x
0 G F (1) r F H 0
and note that G and H have support of lower dimension than F provided that we ensure
that the hyperplane xr = 0 does not contain any component of the support of F . (We can
arrange this given that k is algebraically closed; see exercises. In fact, k innite would be
sucient.) By the induction hypothesis, we know that (Prk , F (n)) (Prk , F (n 1)) is a
polynomial in n of degree at most dim(Supp F ) 1. It is an elementary exercise in algebra
to then see that (Prk , F (n)) is a polynomial in n of degree at most dim(Supp F ).
The polynomial P (n) is called the Hilbert polynomial of the sheaf F ; in case F = OX , we
call it the Hilbert polynomial of the scheme X itself. Note that by Serres vanishing theorem,
for some n0 , we have
P (n) = dimk H 0 (X, F )
(n n0 );
this was the original denition of the Hilbert polynomial.

For example, the Hilbert polynomial of Prk itself is P (n) = n+r
. For another example,
n
the Hilbert polynomial of the subscheme Spec k[x]/(x2 ) of P1k is P (n) = 2, which is the
same as the Hilbert polynomial of a scheme consisting of two distinct reduced points. This
is suggestive, because this scheme can indeed be written as a at limit of pairs of distinct
points.
Flatness and Hilbert polynomials
The Hilbert polynomial can be used to give the following numerical criterion for atness.
(The locally noetherian hypothesis is important; I think one can replace integral by con
nected and reduced.)
Theorem. Let T be an integral (locally) noetherian scheme. Let X PrT be a closed subscheme. Let F be a coherent sheaf on X. For each t T , let Pt Q[z] be the Hilbert
polynomial of the pullback of F to the bre Xt viewed as a subscheme of Pr(t) (where
(t) = OT,t /mT,t is the residue eld of the point t). Then F is at relative to X T
if and only if Pt is constant as a function of t.
In particular, X itself is at over T if and only if the Hilbert polynomial of Xt is constant
as t varies. This gives us a way to check whether a morphism is at which we were sorely
lacking before.
Proof. (Compare Hartshorne Theorem III.9.9, or EGA III 7.9.) We rst note that we can
reduce to the case X = PrT by replacing F with its direct image. We next note that it suces
to consider the case where T = Spec A for A a local integral noetherian ring.
We then show that F is at over T if and only if H 0 (X, F (m)) is nite free over A for
m suciently large. On one hand, if F is at over T , then so are all the terms in the sheafy
2

Cech
resolution of F (m) for the usual open cover U (since open immersions are at). Taking
global sections, we see that the terms of the exact sequence
0 H 0 (X, F (m)) C 0 (U, F (m)) C r (U, F (m)) 0
are all at except possibly for the rst term. This then implies atness of H 0 (X, F (m))
(exercise). Since its also nitely generated over A by Serres niteness theorem, it is free.
On the other hand, if we pick m0 such that H 0(X, F (m)) is nite free over A for m m0 ,
for
then we can recover F as M
M=
H 0 (X, F (m)).
mm0
Since M is at, so is F .
We now need to show that H 0 (X, F (m)) is nite free for m large if and only if Pt is
constant. I claim that this follows by checking
H 0 (Xt , Ft (m)) = H 0 (X, F (m)) A (t)
for m large (even if I dont prove this uniformly in t). Namely, if H 0 (X, F (m)) is nite free
over A for m m0 , then for each t, for m large, I nd that Pt equals P for the generic
point of T . On the other hand, if Pt is the same for the generic point and the closed point,
then I can make m large enough to work for both, and obtain nite freeness of H 0 (X, F (m)).
To check
H 0 (Xt , Ft (m)) = H 0(X, F (m)) A (t),
we may reduce to the case where t is the closed point by replacing A with OT,t . Since A is
noetherian, we can nd a short exact sequence
An A (t) 0
of A-modules. We can then tensor with F to get an exact sequence; it follows (exercise) that
H 0 (X, F (m)n ) H 0 (X, F (m)) H 0 (X, Ft (m)) 0
is exact for m large. I can pull out the direct sum, and then we read o what we want.
Hilbert schemes
It turns out that there is a universal family of closed subschemes of projective space with a
xed Hilbert polynomial.
Theorem (Grothendieck). Fix a eld k and an integer r. Let P (z) Q[z] be a polynomial.
There exists a noetherian scheme H over Spec k and a closed subscheme X of PrH which is
at with Hilbert polynomial P (z), such that for any noetherian scheme T and any closed
subscheme Y of PrT which is at with Hilbert polynomial P (z), there is a unique morphism
T H such that Y = X H T as closed subschemes of PrT
= PrH H T .
For instance, one can show that a closed
of Prk is a d-dimensional plane if and
n+subscheme
d1
only if it has Hilbert polynomial P (n) =
. The parameter scheme in this case is the
n
r
Grassmannian of d-dimensional planes in Pk .
3
Hilbert polynomials, degree, and dimension
Some of the basic information contained in the Hilbert polynomial of a scheme is the follow
ing.
Lemma. Let P (z) be the Hilbert polynomial of a closed subscheme X of Pnk .
(a) We have deg(P ) = dim(X).
(b) Put d = dim(X). For any d-dimensional plane P Pnk such that dim(X P ) = 0,
the length of X P is d! times the leading coecient of P . (This length is called the
degree of X.)
Proof. We may assume k is algebraically closed. We rst need to know that for a generic d
dimensional plane P (i.e., one chosen outside some closed subscheme of the Grassmannian),
we have dim(X P ) = 0. This follows from the fact that as long as X 6= , for a generic
hyperplane H, we have dim(X H) < dim(X) (exercise).
Put F = j OX for j : X Pnk the given closed immersion. For H a hyperplane with
dim(X H) < dim(X), we have an exact sequence
0 F (1) F G 0
where G is the direct image of the structure sheaf of X H. If P (z) is the Hilbert polynomial
of X, it follows that the Hilbert polynomial of X H is P (z) P (z 1). From this, both
claims follows.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009
Spectral sequences and Cech

cohomology
We explain the construction (or rather, one particular construction) of spectral sequences,
enough to explain how they are used as part of the computation of the sheaf cohomology of
quasicoherent sheaves on ane schemes using Cech

cohomology.
I continue to recommend Bott and Tu, Dierential Forms in Algebraic Topology as a
good reference for spectral sequences.
Exact couples
It is handy to start with the following bit of homological algebra. An exact couple is a
circular exact sequence
A
For instance, given an exact sequence 0 A A B 0, we get an exact couple by

taking k = 0. A more typical example: given an exact sequence of complexes
0 A A B 0,
we get an exact couple involving the total cohomologies i hi (A ) and i hi (B ) using the
long exact sequence in cohomology.
From an exact couple we obtain a derived exact couple
A
as follows.
Dene d : B B as d = j k. Then d d = j k j k = 0 because k j = 0, so I
can dene the cohomology B = h(B) = ker(d)/ im(d).
Put A = im(i).
We now have an obvious map i : A A induced by i.
We now claim that there is a well-dened map j : A B sending i(a) to the class
of j(a) for any a A. To make sense of this, we rst note that j(a) ker(d) because
j k j = 0. We next note that if i(a) = 0, then a = k(b) for some b B by exactness,
so j(a) = k(j(b)) = d(b).
1
We now claim there is a well-dened map k : B A induced by k. That is,

if b ker(d), k should carry the class of k to k(b); this belongs to im(a) because
(j k)(b) = 0, so k(b) = i(a) for some a A by exactness. This is well-dened:
It is a routine exercise in diagram chasing to verify that this is again exact.
Filtered complexes and double complexes
Let C be a cohomologically graded complex in nonnegative degrees. A ltration on C is a

decreasing sequence of subcomplexes
C = Fil0 C Fil1 C .
The associated graded complex is
Gri C = Fili C / Fili+1 C .
For instance, suppose D p,q is a double complex, with diererentials dp in the p-direction
and dq in the q-direction. We form a single complex
Ck =
D p,q
p+q=k
with derivation dp + (1)p dq . (The alternating sign is needed to ensure that this is actually
a complex.) We then obtain a ltration on C by setting
Fili C k =
D p,q .
p+q=k,pi
The spectral sequence of a ltered complex
Given a ltered complex C , there are two interesting invariants one can consider. Perhaps
the most natural one is the cohomology h (C ), equipped with the decreasing ltration
Fili h (C ) = im(h (Fili C )).
However, in practice this will usually be something complicated. A less complicated in
variant will be the cohomology of the graded complex h (Grp C ). This is a rather crude
approximation to the cohomology of the total complex; it turns out that there is a sequence
of renements that give closer and closer approximations. These constitute the spectral
sequence associated to the ltered complex.
To start with, take the exact sequence of complexes
0
Filp+1 C
Filp C
Grp C 0.
pZ
pZ
pZ
Identifying the rst two members by shifting indices, then taking the long exact sequence in
cohomology, we get an exact couple
A1
i1
k1
E1
A
1
j1
in which E1 = pZ h (Grp C ). By repeatedly extracting derived exact couples, we get a

sequence of exact couples
Ah
ih
kh
Eh
A
h
jh
for h = 1, 2, . . . . The spectral sequence here is specically the sequence of groups Eh equipped
with the square-zero endomorphisms dh = jh kh . Note that Eh+1 is just the cohomology
of Eh for dh ; the mysterious part is where the next map dh+1 comes from. (The terms
in this sequence are often called the sheets, or pages, of the spectral sequence. The visual
signicance of these metaphors may become more clear in the next section.)
Without any additional hypotheses, the spectral sequence does not say much. But under
certain circumstances, the Eh converge to something useful. Namely, suppose that the
complex C comes not only with a ltration but with a grading C = q Cq .
Theorem. Suppose that for each q, the induced ltration on Cq has only nitely many
distinct steps. Then the spectral sequence converges, in the sense that for each q, the q
th graded piece of Eh stabilizes for h large. If we let E denote the sum of the stable
graded pieces, then E is canonically isomorphic to the associated graded group of the ltered
cohomology Fili h (C ).
Note that we still dont quite manage to compute the ltered cohomology, but but only
its graded pieces. Still, that information itself is often very very useful. (It is sometimes said
that the spectral sequence abuts to the ltered cohomology.)
Proof. See Bott and Tu, Theorem 14.6.
The spectral sequence of a double complex
Let us see how this works in the specic example of a double complex. (Im just going to
state the result; see Bott and Tu for the derivation.) Let D p,q be a double complex, and
let C be the associated ltered single complex. It is customary to draw pictures in this
orientation:
..
.
..
.
..
.
D 0,2
D 1,2
D 2,2
D 0,1
D 1,1
D 2,1
D 0,0
D 1,0
D 2,0
without any arrows (at least for now).

Let me redraw this picture writing
standing for (1)p dq :
..
.
E0p,q for D p,q , and drawing in the vertical arrows

..
.
..
.
E00,2
E01,2
E02,2
E00,1
E01,1
E02,1
E00,0
E01,0
E02,0
Taking cohomology here gives you exactly E1 . A quick diagram chase shows that the next
dierential is precisely the one induced by dp :
..
.
E10,2
..
.
..
.
E11,2
E12,2
E10,1
E11,1
E12,1
E10,0
E11,0
E12,0
Taking cohomology gives the next sheet E2 . But what is the next dierential? Again, Ill
just state the answer. Each element of E2 is represented by an element of b for which for
some c,
dq (b) = 0,
dp (b) = (1)p+1 dq (c).
4
The next dierential carries this class to dp (c), which turns out to be well-dened.
0
?
c
That is, our next page should be drawn like this:

..
.
..
.
..
.
E20,2
E 1,2
E22,2
E20,1
E2
E22,1
E20,0
E2
E22,0
1,1
1,0
The pattern continues: we have

dr : Erp,q Erp+r,qr+1
and we can explicitly see the stabilization, since we get an increasingly large bottom left
p,q
corner with no arrows to or from anyplace other than 0. Let E
denote the stable values;
then the associated graded complex of the ltered total cohomology has k-th step given by
p,q
E
.
p+q=k
Spectral sequences and Cech

cohomology
Here is how spectral sequences make quick work of the comparison theorem between Cech
and sheaf cohomology, in the form needed for algebraic geometry. Let X be a topological
space, and let F be a sheaf of abelian groups on X. Let I be a asque resolution of F .
Take the double complex
D p,q = C p (X, I q ) = lim C p (U, I q ).
The trick here is that there are two dierent ways to run the spectral sequence construction
from a double complex, depending on how you orient the diagram. As written, we rst take
Cech
cohomology, and then take cohomology of whatever that yields:
p,q
p (X, I q )
E1a
=H
p,q
p (X, I )).
E2a
= hq (H
p,q
Note that E1a

= 0 for p > 0 because the Cech
cohomology of a asque sheaf is zero, whereas
p,0
p,q
p,q
p,q
E1a = (X, F ). Thus E2a = 0 for p > 0, and in fact E2a
= Ea
for all p, q. Since we only
have one term along each antidiagonal, we actually get much more than usual: we really
have computed the cohomology of the total complex, and it is the E20a,q = H q (X, F ).
Now lets run the spectral sequence with the roles of p and q reversed. This time, I take
cohomology in the q-direction rst, so I start with
q,p
E1b
= hq (C p (X, I )).
This is a rather strange object, but we can repackage it in a useful way by noting that the
functor I C p (X, I) preserves exact sequences of presheaves, i.e., sequences of presheaves
where the sections over any open give an exact sequence. That means that working with
presheaves, I can commute the cohomology computation across the C p . Ill take advantage
of this by dening the presheaf Hq by
Hq (U) = H q (I (U)) = H q (U, F ),
so that
q,p
E1b
= C p (X, Hq )
p (X, Hq )
E q,p = H
2b
interpreted as the Cech

complex associated to a presheaf (dened using the same formula as
q,p
for sheaves). This spectral sequence must converge to some term Eb
giving graded pieces
of the total cohomology, which we already identied as the sheaf cohomology of F itself.
This isnt useful as an abstract method for dealing with Cech

cohomology. However,
it is just the thing I need to prove the theorem that I need to nish the argument that
quasicoherent sheaves on ane schemes are acyclic.
Theorem. Let X be a topological space equipped with a nice basis B (i.e., a basis closed
under pairwise intersections; we need not assume X B). Let F be a sheaf of abelian
i(U, F ) = 0 for all i > 0 and all U B. Then there are natural
groups on X such that H
i
(X, F ) H i (X, F ) for all i 0.
isomorphisms H
Proof. The natural maps come from the fact that if I is an injective resolution of F , then
the Cech
resolution C (X, F ) admits a map into I which is a quasi-isomorphism, and is
well-determined up to a chain homotopy. (This is similar to the homework problem about
injective resolutions of complexes; see PS 8, problem 7.)
6
To prove the theorem, it suces to check for X equal to an open in B, as then the Leray
theorem asserts that we can compute sheaf cohomology using any cover by elements of B,
and any open cover renes to such. So assume hereafter X B.
We induct on i, the case i = 0 being an easy consequence of the sheaf axiom. Say we
know that
H j (U, F ) = 0
(0 < j < i, U B).
q,p
Then the spectral sequence Eb from above has E2b
= 0 for 0 < q < i. By staring at the
spectral sequence, we see that the terms with q + p = i must already be stable, so the total
i-th cohomology must just be
i (X, H0 ) = H
i(X, F ).
E20b,i = H
Since we also know that the total cohomology is H i(X, F ), we obtain the desired isomor
phism.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

GAGA (updated 30 Apr 2009)
We now discuss a classic theorem of algebraic geometry, Serres GAGA, which exposes
a tight relationship between algebraic geometry over the complex numbers and complex
analytic geometry. By far the best reference for this is Serres original paper Geometrie
algebrique et geometrie analytique. (Thanks to Bjorn Poonen for reporting some errors,
which have now been corrected.)
Coherent sheaves
In order to discuss GAGA, I need to talk about coherent sheaves not just on schemes, but
on analytic spaces. In fact, the notion is well-dened on any locally ringed space.
Let (X, OX ) be a locally ringed space. We say a sheaf F is coherent if F is nitely
generated, and for any open subset U of X, any nonnegative integer n, and any homomor
n
phism h : OX
|U F |U , the kernel of h is itself nitely generated. Warning! I originally
only required this for h surjective, but I dont think that is enough. (Important note: we
dont require the kernel to be generated by nitely many sections over all of U.) This is
stronger than saying that F is nitely presented, in which case we only require that one
such surjection h must have this property. In particular, OX itself need not be coherent.
However, if X is locally noetherian, then all nitely generated quasicoherent sheaves are
in fact coherent. This follows from the following result.
Theorem. Let A be a noetherian ring, put X = Spec A, let V be an open subset of X, and
let F be an OX |V -module. Then the following are equivalent.
(a) F is coherent.
(b) F is nitely generated and quasicoherent.
for some nitely generated A-module M.
(c) We have F = M
Im only going to show this for V = X, as this is the only case I need. For the general
case, see EGA 1, Theorème 1.5.1.
Proof. Even without a noetherian hypothesis, it is obvious that (a) implies (b), and we
checked (b) implies (c) in a previous lecture.
To check that (c) implies (a) under the noetherian hypothesis, note that the claim is
n
local, so it suces to check that the kernel of a homomorphism OX
|D(f ) F |D(f ) is
nitely generated. It is represented by a homomorphism Anf Mf of Af -modules, but Af is
noetherian since A is. Hence the kernel of the homomorphism, being a submodule of a nitely
generated Af -module, is itself a nitely generated Af -module (because A is noetherian).
Lemma. Let
0 F1 F F2 0
be a short exact sequence of quasicoherent sheaves on a locally ringed space X. Then if any
two of F , F1 , F2 are coherent, so is the third.
Proof. Exercise.
Beware that it is not obvious that the inverse image of a coherent sheaf is coherent, since
the dening condition involves looking at all open subsets.
Analytication of coherent sheaves
In order to state the GAGA theorems, we use the fact that there is a morphism of locally
ringed spaces
r Pr ,
h
: P
C
C
where the left side is the projective r-space over C viewed as a complex manifold (or a
complex analytic variety, on which more later). Where does this morphism come from?
Well give a functorial answer later, but for now Ill do something more direct.
For each i = 0, . . . , r, put Xi = D+ (xi ) PrC .
This space is an ane n-space over C
i be the complex analytic ane r-space with the same
with coordinates xj /xi for j =
i; let X
coordinates. There is an obvious map
i , O );
(Xi , OXi ) = C[x0 /xi , . . . , xr /xi ] (X
Xi
by adjunction, this gives us a morphism
i Xi
X
r includes
of locally ringed spaces. These glue to give the morphism I described. Note that P
C
only some of the points of PrC (namely the closed points), but gives them a ner topology
(the analytic topology rather than the Zariski topology). This is consistent with the fact
r Pr is continuous.
that the map P
C
C
What is nice about viewing the analytication process this way is that we can apply
r to Pr . For instance, for
operations dened on locally ringed spaces uniformly to both P
C
C
r .
any quasicoherent sheaf F on PrC , the pullback h F is a quasicoherent sheaf on P

C
Lemma (Cartan). For any coherent sheaf F on PrC , h F is coherent.
Proof. Recall that there exists a surjection O(n)m F for some integers m, n. It thus
suces to show that h O(n) is coherent. Since coherence is a local property, it is enough to
show that the structure sheaf on complex analytic ane n-space is coherent (as a module
over itself). This follows from the fact that each local ring of this space is noetherian.
I wont give a complete proof of this here, but the basic idea is as follows. Let C{x1 , . . . , xr }
be the ring of formal power series which converge in some neighborhood of the origin; this is
2
the ring we are trying to prove is noetherian. We proceed by induction on r, the case r = 0
being trivial. The key to the induction step is the Weierstrass preparation theorem, which im
plies that any element of C{x1 , . . . , xr } equals a unit times an element of C{x1 , . . . , xr1 }[xr ].
Since that ring is noetherian by the induction hypothesis plus the Hilbert basis theorem, we
deduce that C{x1 , . . . , xr } is too. For a proof of the Weierstrass preparation theorem, see for
example the rst few pages of Griths and Harris, Principles of Algebraic Geometry.
We also need the following relationship between analytic and algebraic stalks.
r , the morphism f : OPr ,z Or is at. That is, the morphism
Lemma. For any z P
C
PC ,z
C
r
r
h
: P P is at.
C
Proof. Let t1 , . . . , tn be local (algebraic) coordinates at z. Then we have a completion morphism g : OPr ,z Ct1 , . . . , tn . Both g and g f are faithfully at because they are maps
C
from noetherian local rings into their completions for their maximal ideals. This easily yields
atness of f .
Corollary. The functor h from quasicoherent sheaves on PrC to quasicoherent sheaves on
r is exact.
P
C
The rst GAGA theorem
Note that for any quasicoherent sheaf F on PrC , there is always a natural morphism
r , h F )
H i (PrC , F ) H i(P
C
by pulling back along h. More concretely, you may view an algebraic Cech
cocycle as an
analytic one.
Theorem (GAGA, part 1). For any coherent sheaf F on PrC , the natural morphism
r , h F )
H i (PrC , F ) H i(P
C
is an isomorphism for each i 0.
In order to prove this, we need a mechanism for computing sheaf cohomology on analytic
spaces. Here it is, presented as a black box.
Theorem (Cartan). For any nonempty subset J of {0, . . . , r} and any coherent sheaf F on
j , H i (U, F ) = 0 for i > 0.
U = jJ X
The key point is that U is a Stein manifold. This also holds if U is the analytication
of any ane scheme of nite type over C (which Ill leave to you to dene). By Lerays
theorem, this gives the following corollary.
r , we may compute sheaf cohomology using the

Corollary. For any coherent sheaf F on P
C
Cech
complex associated to the cover U = {X0 , . . . , Xr }. In particular, the i-th cohomology
vanishes for i > r.
With this, the proof is parallel to that of Serres niteness theorem.
Proof of GAGA (part 1). We rst prove the claim for F = O by an explicit Cech
cohomology
0 r
calculation (exercise); note that the computation H (PC , O) = C comes down to the fact
that any bounded entire function on Cn is constant, which reduces to Liouvilles theorem.
(By the way, this makes it clear that the theorem is completely false if we replace PrC with,
say, the ane space ArC . More on this later.)
We next deal with the cases F = O(n) for n Z, using the exact sequence
x
0 O(n 1) r O(n) OH (n) 0

for H
the hyperplane xr = 0. By induction on r, and comparing long exact sequences
= Pr1
C
in cohomology, we can infer all of the cases from the case n = 0.
Finally, we treat the general case by descending induction on i (as in the proof that Cech
cohomology for a good cover computes sheaf cohomology). Build an exact sequence
0GE F 0
in which E is a direct sum of twisting sheaves. Note that applying h is exact, so we get
an exact sequence on the analytic side. Then twist and compare long exact sequences in
cohomology after twisting:

H i (PrC , E(n))
H i(PrC , F (n))
r , h E(n))
H (P
C
i
H i+1 (PrC , G(n))
H i+1 (Pr , E(n))
r , h F (n))
H (P
C
i
i+1
r , h G(n))
(P
C
i+1
r , h E(n))
(P
C
The second GAGA theorem
We now know that algebraic coherent sheaves preserve their cohomology under pullback to
the analytic side. We next show that they also preserve their morphisms.
Theorem (GAGA, part 2). Let F , G be coherent sheaves on PrC . Then the natural map
HomOPr (F , G) HomOPr (h F , h G)
C
is an isomorphism.
Using the ve lemma, we get the desired result.
Proof. In general, for sheaves of OX -modules F and G, let H om(F , G) be the presheaf
H om(F , G)(U) = HomOU (F |U , GU ).
This is in fact a sheaf, called the sheaf Hom from F to G. Its global sections are just
HomOX (F , G). (I should really write H om OX (F , G) with the subscript OX , but never mind
for now.)
Note that there is a natural map
h H om(F , G) H om(h F , h G)
r , given by viewing an algebraic morphism over a Zariski open subset of
of sheaves on P
C
r . We claim this map is an
PrC as an analytic morphism over the corresponding subset of P
C
isomorphism; this will imply the theorem by taking global sections of this isomorphism, then
applying the rst GAGA theorem.
r a natural
We check the isomorphism on stalks. Using coherence, we have for each z P
C
identication
H om(F , G)z = Hom(Fz , Gz )
and similarly on the analytic side. Put
=
Or ;
R
P ,z
R = OPrC ,z ,
is at over R. By that atness plus the lemma below

a lemma from earlier states that R
(and the fact that R is noetherian), we have a natural identication
= Hom(Fz R R,
Gz R R).
Hom(Fz , Gz ) R R
This yields the claim.
Lemma. Let R be a noetherian ring. Let S be a at R-algebra. Then for any R-modules
M, N, the natural map
HomR (M, N) R S HomS (M R S, N R S)
is a bijection.
Proof. Since R is noetherian, I can nd an exact sequence
F1 F0 M 0
where F0 , F1 are nite free R-modules. Then we get a diagram
HomR (F1 , N) R S
HomR (F0 , N) R S
HomS (F1 R S, N R S)
HomR (M, N) R S
HomS (F0 R S, N R S)
HomS (M
R S, N R S)
with exact rows (the exactness in the rst row requiring the atness of S over R). Since the
second and third vertical arrows are isomorphisms, so is the rst by the ve lemma.
5
The third GAGA theorem
We next try to classify the coherent sheaves on the analytic projective space. We need one
more black box.
r , the spaces H i(P
r , F ) are nite
Theorem (Cartan, Serre). For F a coherent sheaf on P
C
C
dimensional over C for all i 0.
Sketch of proof. Equip the Cech

cocycles for the usual open cover with the topology of
i is
uniform convergence on compact subsets. Then restrict to a cover in which each X
i . Using this, one sees that for the induced
replaced by an open subset with closure inside X
i r
topology on H (PC , F ), the identity map is compact, which is only possible if this vector
space is nite dimensional over C.
r is the pullback under h of a unique
Theorem (GAGA, part 3). Every coherent sheaf on P
C
coherent sheaf on PrC .
The uniqueness follows from the second GAGA theorem. To prove existence, we induct
r , we extend the twisting
on r, the case r = 0 being trivial. For F a coherent sheaf on P
C
notation from the algebraic case by writing
F (n) = F h O(n).
Lemma. Assume the third GAGA theorem in dimensions up to r 1. For any coherent
r and any z P
r , there exists an integer n0 (depending on F and z) such that
sheaf F on P
C
C
for n n0 , F (n)z is generated by global sections of F (n).
Proof. Choose xr H 0 (PrC , O(1)) vanishing at z, and let E be the hyperplane xr = 0. We
then have the usual exact sequence
x
0 O(1) r O OE 0
of algebraic coherent sheaves. Tensoring with F , we have an exact sequence
F (1) F FE 0
where FE denotes the pushforward of the restriction to E. Let G be the kernel on the left
side. Twisting, we get
0 G(n) F (n 1) F (n) FE (n) 0.
Split this into short exact sequences:
0 G(n) F (n 1) H 0
0 H F (n) FE (n) 0
and then take long exact sequences in cohomology:

r , F (n 1)) H 1 (P
r , H) H 2 (P
r , G(n))
H 1 (P
C
C
C
1 r
1 r
1 r
H (P , H) H (P , F (n)) H (P , FE (n)).
C
Note that G and FE are supported on E, so by the induction hypothesis, they both come
r , G(n))
from algebraic coherent sheaves. It follows that for n large enough, the terms H 2 (P
C
r , FE (n)) both vanish. We thus obtain inequalities
and H 1 (P
C
r , F (n 1)) dimC H 1 (P
r , H) dimC H 1 (P
r , F (n))
dimC H 1 (P
C
C
C
r , F (n))
for n large. By the previous Cartan theorem, the terms of the sequence dimC H 1 (P
C
are all nite; we just showed that they are nonincreasing for n large enough. They thus
eventually reach a constant value for n large enough!
In particular, for n large, the previous inequalities all become equalities. Backing up the
second of the two long exact sequences, we see that
r , F (n)) H 0 (P
r , FE (n))
H 0(P
C
C
must be surjective for n large.
r , FE (n)) generates (FE )z . By
Again since FE is known to be algebraic, for n large, H 0 (P
C
r , F (n)) also generates F (n)z .
a quick Nakayamas lemma argument, for such n, H 0 (P
C
Corollary. Assume the third GAGA theorem in dimensions up to r 1. For any coherent
r , there exists an integer n0 (depending only on F ) such that for any n n0
sheaf F on P
C
r , F (n)z is generated by global sections of F (n).
and any z P
C
Proof. For a single n, if the claim holds for a single z, it holds in a neighborhood of that z;
moreover, by multiplying these sections by monomials in x0 , . . . , xr , we infer the claim for
r is compact, we may nd a single n0 such
all larger n in the same neighborhood. Since P
C
r and each n n0 .
that F (n)z is generated by global sections of F (n) for each z P
C
r . By the previous corollary, for some
Proof of the theorem. Let F be a coherent sheaf on P
C
r , F (n)), which by Cartans
n, each stalk of F (n) is generated by the space of sections H 0(P
C
theorem is nite dimensional over C. We thus obtain a surjection h O(n)m F for some
m, n. Applying the same argument to the kernel of this map (which is again coherent), we
get an exact sequence
F1 F2 F 0
in which each Fi is a direct sum of pullbacks of twisting sheaves. In particular, the Fi
are algebraic; by the second GAGA theorem, the morphism between them is also algebraic.
We may then form the algebraic cokernel, whose analytication is isomorphic to F , as
desired.
More analytication
One can state the GAGA theorems more generally, but rst we need to discuss analytication
of spaces other than projective space. We rst specify the target category: a locally ringed
space (X, OX ) is a complex analytic space if each point x admits a neighborhood U and an
immersion : U Cn for some n. This is not the same as a complex manifold because we
allow singularities, and for that matter nonreducedness (so these shouldnt be called complex
analytic varieties either). Let AnSp denote the category of complex analytic spaces.
We would like a process for turning schemes locally of nite type over C into complex
analytic spaces in a natural way. It is easy to say what we want to have happen in local
coordinates: if X = Spec C[z1 , . . . , zn ]/(f1 , . . . , fm ), we want to take the subspace Z of Cn
on which f1 = = fm = 0, equipped with the quotient of OCn by the coherent ideal sheaf
generated by f1 , . . . , fm (or rather, its inverse image on Z).
However, if one works this way, one has to check independence from coordinates. This is
doable but annoying (its like Hartshornes Proposition II.2.14 comparing certain schemes to
varieties). There is a more functorial description of analytication introduced by Grothendieck;
see SGA I, expose XII, Theorème-Denition 1.1.
Theorem. Let X be a scheme locally of nite type over C. The functor
Y HomLocRingSp (Y, X)
from AnSp to Set is represented by an analytic space X an ; that is, there are natural isomor
phisms
HomLocRingSp (Y, X) HomAnSp (Y, X an ).
Moreover, X an has underlying set X(C), and the morphism X an X induces isomorphisms
of completed local rings, and so is at.
You could interpret this as saying that the inclusion functor from analytic spaces to
locally ringed spaces has a partial right adjoint.
Sketch of proof. One rst shows that the class of schemes for which the theorem holds is
closed under forming open subschemes, closed subschemes, and products, by mirroring these
constructions on the analytic side. It then suces to check the theorem for X = A1C ; this
amounts to observing that giving a map Y X is the same (by the adjunction property of
ane schemes) as specifying a map C[t] (Y, OY ), which in turn is the same as specifying
the image of t. That is, Hom(Y, X) is naturally isomorphic to (Y, OY ). On the other
hand, if we view C as an analytic space in the obvious fashion, then we may again identify
Hom(Y, C) naturally with holomorphic functions on Y , i.e., with (Y, OY ). This proves the
claim for ane space.
This paradigm extends to other categories derived from schemes. For instance, for k an
algebraically closed eld, separated reduced schemes of nite type over k admit varietica
tions, thus reproducing the class of abstract algebraic varieties and giving a stronger version
of Hartshorne Proposition II.2.14.
8
Extension to projective and proper schemes
In terms of the analytication functor, we can now extend the GAGA theorems as follows.
Theorem (GAGA for projective schemes). Let X be a closed subscheme of PrC for some
r 1. Let h : X an X be the analytication morphism.
(a) For any coherent sheaf F on X, the natural morphism
H i (X, F ) H i(X an , h F )
is an isomorphism.
(b) For any coherent sheaves F , G on X, the natural morphism
HomOX (F , G) HomOX an (h F , hG)
is an isomorphism.
(c) Every coherent sheaf on X an is isomorphic to h F for a unique coherent sheaf F on
X.
We saw earlier that already (a) is totally false for X = ArC , so some sort of completeness
is necessary. Grothendieck noticed that it suces to assume X is proper over C; this reduces
to the projective case using Chows lemma (exercise).
Applications
The GAGA theorem has applications too numerous to count, so Ill just mention a few (see
SGA 1, expose XII for more). The following was proved before GAGA by Chow, but is an
immediate corollary.
r is the analytication of a closed
Corollary (Chow). Any closed analytic subvariety of P
C
algebraic subvariety.
Another application is the following.
Theorem. Let X be a smooth proper scheme over C. Then
H p (X, qX/C ) = H p (X an , qX an )
(p, q 0).
This can be used to show that the hypercohomology of the algebraic de Rham complex
coincides with the hypercohomology of the analytic de Rham complex. (If F : C1 C2
is a left exact additive functor of abelian categories with C1 having enough injectives, the
hypercohomology of a complex C is dened by forming a quasi-isomorphism C I with
the I all injective, and taking hi (F (I )). More on this construction a bit later.) This in turn
can be combined with some more results on the analytic/topological side (the Dolbeault and
de Rham theorems, respectively) to show that algebraic de Rham cohomology computes the
usual topological Betti numbers of a smooth variety over C.
Here is another application by Grothendieck. See SGA 1, expose XII again.
.X/C
Theorem (Grothendieck). Let X be a smooth proper scheme over C. Then any nite
covering space map Y X an (of topological spaces) corresponds to a nite etale cover of X
in the category of schemes.
One can dene the etale fundamental group of a scheme X as, roughly speaking, the
automorphism group of a maximal inverse system of connected nite etale covers of X. For
instance, if X = Spec F with F a eld, this gives the absolute Galois group of F . (To make
this more precise, one must x a choice of a basepoint just as in the topological case.) The
previous theorem implies that for a smooth proper scheme over C, the etale fundamental
group is the pronite completion of the usual topological fundamental group, i.e., the inverse
limit of its nite quotients. For instance, for an elliptic curve, the topological fundamental
group is Z Z, while the pronite completion is
Z
(Zp Zp ),
=
p
where Zp denotes the p-adic integers.

Corollary. Let K be a number eld and let X be a smooth proper scheme over K. Then the
pronite completion of the fundamental group of (X K C)an does not depend on the choice
of the embedding K C.
This might not be so surprising until I tell you that Serre exhibited examples in which
the topological fundamental group does depend on the choice of the embedding! (Serres
example is a rather articial construction using elliptic curves with complex multiplications.
There are some more natural examples due to one of our postdocs, Junecue Suh.)
The following is an example of a rather large class of results from SGA 1. See the exercises
for an example involving properness.
Theorem. Let f : X Y be a morphism of schemes locally of nite type over C. Then f
is separated if and only if f an : X an Y an is separated. In particular, X is separated if and
only if X an is Hausdor.
Analogues
I know of at least two analogues of GAGA, though there may be more.

One is formal GAGA, in which one passes from a scheme to its formal completion
along a closed subscheme.
The other is rigid GAGA, which is like ordinary GAGA except that one works over a
complete nonarchimedean eld, and uses Tates notion of rigid analytic geometry (or
Berkovichs notion of nonarchimedean analytic geometry instead of complex analytic
geometry.
10
I suppose there is also an instance of GAGA for passing from reduced separate schemes
of nite type over an algebraically closed eld to abstract algebraic varieties. But that
one is neither surprising nor useful.
11
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Serre duality for projective space
Ext groups
For R a ring and M, N ModR , I dened Exti (M, N) as the image of N under the i-th
right derived functor of HomR (M, ). This makes sense because HomR (M, ) is a left exact
covariant functor from ModR to Ab (it actually goes to ModR but never mind that). I also
remarked that it can be viewed as the image of M under the i-th right derived functor of
HomR (, N), provided we view this as a functor on Modop
R.
For the category ModX of sheaves of OX -modules on a ringed space X, we can imitate
the rst construction pretty directly, except that we have to choose between the normal
Hom and the sheaf Hom. Let Exti (F , ) be the right derived functors of Hom(F , ), and let
Ext i (F , ) be the right derived functors of Hom(F , ).
For example, there is a natural isomorphism
Exti (OX , F )
= H i(X, F )
because these are derived functors of the naturally isomorphic functors Hom(OX , F )
=
0
H (X, F ). On the other hand, Hom(OX , F ) is the identity functor, so
Ext 0 (OX , F )
= F,
Ext i (OX , F ) = 0 (i > 0).
Lemma. Let I be an injective OX -module. Then for any open subset U of X, I|U is an
injective OU -module.
Proof. Let j : U X be the inclusion. We must show that given a monomorphism F G,
any map F I|U extends to G. Let j denote extension by zero, so that j F has the
same stalks as F over U and zero stalks elsewhere. (Sections are the same as F over
opens contained in U and zero elsewhere.) By looking at stalks, j F j G is still a
monomorphism. Moreover, we have a map j I|U I by adjunction, and the resulting
composition j F j I|U I extends to j G I. Restricting back to U gives the desired
map G I|U .
Corollary. For any open subset U of X, there are natural isomorphisms
Ext i (F , G)|U
= Ext i (F |U , G|U ).
In particular, the right side is a sheaf; e.g., Ext i (F , G) = 0 for i > 0 whenever F is locally
free of nite rank.
Proof. Both sides are cohomological functors in G whose higher terms vanish on injectives (by
the previous lemma in the case of the right side), hence are eaceable and thus universal.
Corollary. For I an injective OX -module, the functors

Hom(, I),
Hom(, I)
are exact.
Proof. This is true for Hom by the denition of injectivity. For Hom, use the lemma.
Proposition. For F an OX -module, Exti (, F ) and Ext i (, F ) are cohomological functors on
Modop
X.
Proof. Let I be an injective resolution of F . Given a short exact sequence
0E GH0
in ModX , we obtain the long exact sequence by taking Hom or Hom into I , yielding a short
exact sequence of complexes (by the previous corollary), and then taking the long exact
sequence of cohomology groups. One does need to check independence from the choice of
the resolution, but this is similar to other arguments weve done before, so I wont bore you
with it. (The summary: by a pushout construction, it suces to compare I and J when
there is a quasi-isomorphism I J . You then get a morphisms of short exact sequences
which is a quasi-isomorphism on each term, etc.)
Unfortunately, we cant check that Exti (, F ) and Ext i (, F ) are eaceable, or construct
them as derived functors, because ModX need not have enough projectives (exercise). How
ever, we can still use certain acyclic resolutions to compute.
Proposition. Suppose that
L1 L 0 F 0
is an exact sequence in ModX , where each Li is locally free of nite rank. (We say the L
form a locally free resolution of F .) Then there is a isomorphism
Ext i (F , G)
= hi (Hom(L , G))
which is functorial both in G and in the resolution of F .
Proof. Since Li is locally free of nite rank, Ext 1 (I, ) always vanishes, so Hom(Li , ) is an
exact functor (even though Hom(Li |U , ) is not exact on any open U). From this we see that
the right side is a cohomological functor: given a short exact sequence 0 G1 G G2
0, the sequence
0 Hom(L , G1 ) Hom(L , G) Hom(L , G2 ) 0
is again exact, so admits a long exact sequence in cohomology.
We now have that both sides of the desired isomorphism are cohomological functors in G
whose higher terms vanish on injectives, so are eaceable. Hence they are both universal.
2
Note that locally free resolutions are much easier to write down in practice than injective
resolutions. For instance, if X = Pnk for k a eld, and F is coherent, then Serres theorem
gives a surjection E F where E is a direct sum of twisting sheaves. Repeated application
gives not just a locally free resolution but a free resolution!
Proposition. For any coherent sheaves F , G on Pnk for k a eld, Ext i (F , G) is again coherent.
Proof. We just argued that a free resolution L of F exists. By the previous proposition, we
need only check that the hi (Hom(L , G)) are coherent. This is true because the L and G
are coherent, so the Hom(L , G) are too.
Lemma. For F , G, L ModX with L locally free of nite rank, there are canonical isomor
phisms
Exti (F L, G)
= Exti (F , L G)
and
Ext i (F L, G)
= Ext i (F , L G)
= Ext i (F , G) L .
Proof. Again, check that everything is an eaceable cohomological functor of G and that
things match at i = 0. (See Hartshorne Proposition III.6.7.)
Final note: you may be wondering what the relationship is between Ext and Ext. It
comes from the following general fact: if F and G are left exact functors such that F G
makes sense, then there is a spectral sequence relating the derived functors of F , the derived
functors of G, and the derived functors of F G. (See Godements book for details.) In our
case, given a sheaf F , take
F = H 0 (X, ), G = Hom X (F , ), F G = HomX (F , ).
Duality on projective space
For the rest of this lecture, we work over a eld k, but it need not be algebraically closed.
Theorem (Duality on projective space). Put X = Pnk . Let F be a coherent sheaf on X.
Recall that H n (X, OX (n 1)) is one-dimensional over k.
(a) The map
HomX (F , OX (n 1)) H n (X, F ) H n (X, OX (n 1))
is a perfect pairing of nite dimensional k-vector spaces (i.e., it identies each space
with the Hom of the other into the target).
(b) For V a k-vector space, put
V = Homk (V, H n (X, OX (n 1)).
3
For each i 0, there is a natural isomorphism

Exti (F , OX (n 1)) H ni (X, F )
which for i = 0 reproduces (a), and which is compatible with short exact sequences.
Proof. For (a), we have a natural morphism
Hom(F , OX (n 1)) H n (X, F )
of left exact covariant functors on Modop
X , which we claim is an isomorphism. In case F =
OX (m), we want a natural isomorphism
H 0 (X, OX (m n 1))
= Hom(H n (X, OX (m)), H n (X, OX (n 1)))
and this is exactly what we got from Serres calculation. Likewise, we already have the
isomorphism when F is a direct sum of twisting sheaves.
In general, we can write an exact sequence
E1 E0 F 0
in ModX with E0 , E1 both direct sums of twisting sheaves. Since the things we are computing
are left exact on Modop
X , this exact sequence turns into a diagram with exact rows:
Hom(F , OX (n 1))
Hom(E0 , OX (n 1))
Hom(E1 , OX (n 1))
H n (X, F )
H n (X, E0 )
H n (X, E1 )
The ve lemma gives the desired isomorphism.

For (b), we have two cohomological functors on the category of coherent OX -modules
which agree at index 0. We need only check that they are both eaceable. For this, given
F coherent, we can write it as a quotient of E = OX (q)m for any suciently large q. So
all we need to do is check that for any given i > 0, both Exti (OX (q), OX (n 1)) and
H ni(X, OX (q)) vanish for q large. The second statement is true by Serres calculation; so
is the rst because Exti (OX (q), OX (n 1))
= H i (X, OX (q n 1)).
Dierentials and duality
This is not really the right way to view the duality theorem, because it does not generalize
well. To x this, we reintroduce the sheaf X/k of Kahler dierentials on X = Pnk , and its
top exterior power X , the canonical sheaf.
Lemma. For X = Pnk , the sheaf X is isomorphic to OX (n 1).
4
Proof. This can be seen using the exact sequence

0 X/k OX (1)n+1 OX 0
of sheaves on X, where the middle term corresponds to the sheaf ni=0 S(1)ei , the right term
corresponds to S = k[x0 , . . . , xn ], and the map S(1)n+1 S takes ei to xi (Hartshorne,
Theorem 8.13). This gives exact sequences
0 iX/k ik OX (1)n+1 i1
X/k 0
for all i. For i = n + 1, this becomes an isomorphism OX (n 1) nX/k because n+1
X/k = 0.
One can also see this more directly by writing down a global generator of X (n + 1). For
instance, dene H 0 (D+ (x0 xn ), X ) by the formula
xn+1
0
d(x1 /x0 ) d(xn /x0 )
x0 xn
n
1
i dxn .
=
(1)i xi dx0 dx
x0 xn i=0
The rst line shows that x0 xn generates X (n + 1) over D+ (x0 ); it also shows that
performing an automorphism of X which swaps two of x1 , . . . , xn only changes by a sign.
The second line shows that the same is true of the automorphism of X which swaps x0 and
xn . Hence x0 xn generates X (n + 1) over D+ (xi ) for i = 1, . . . , n.
Warning: Hartshorne Remark 7.1.1 claims that , viewed as a Cech

n-cocycle, is invariant
under coordinate changes. However, we just contradicted this by showing that itself
changes sign when you swap two coordinates. What is really happening is that if T : Pnk Pnk
is the linear automorphism dened by the matrix A, in the sense that
T (xj ) =
Aij xi
(i, j = 0, . . . , n),
i
then
T (x0 xn ) = det(A)x0 xn .
In any case, we can use X in place of OX (n1) in the statement of the duality theorem
on projective space. Next time, Ill talk about how this can be generalized to other schemes
over k.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Dualizing sheaves and Riemann-Roch (updated 6 May 09)
In this lecture, we introduce dualizing sheaves for projective schemes over a eld, then
use them to derive the Riemann-Roch theorem for curves. Throughout, let k be a eld (not
necessarily algebraically closed), let j : X P = PN
k be a closed immersion with X of
dimension n, and let OX (1) be the corresponding twisting sheaf.
Dualizing sheaves
For V a k-vector space, let V denote the dual space Homk (V, k). A dualizing sheaf for X
is a coherent sheaf X
equipped with a trace morphism t : H n (X, X
) k, such that for all
coherent sheaves F on X, the composition
t
HomX (F , X
) H n (X, F ) H n (X, X
)k
of the natural pairing with the trace morphism induces an isomorphism

HomX (F , X
) = H n (X, F ).
By interpreting this in terms of representing a certain functor, we see that a dualizing sheaf
is unique up to unique isomorphism if it exists.
Theorem. There exists a dualizing sheaf for X.
This also holds when X is proper, but I wont give the proof in this course (see the
references at the end of Hartshorne III.7).
The previous theorem is not so useful unless one can identify the dualizing sheaf explicitly.
This is tricky in general, but can be done well in the smooth case.
Theorem. Suppose that X is smooth and irreducible over k. Then the canonical sheaf X
is a dualizing sheaf.
Application to Riemann-Roch
Modulo the previous two theorems, we can already deduce Riemann-Roch for curves. Sup
pose in this section that k is algebraically closed, and that X is smooth over k, irreducible,
and of dimension 1.
For any divisor D on X, the identication of the canonical sheaf X
= X/k with the
dualizing sheaf X gives us an isomorphism

H 0 (X, X L(D))
= HomX (L(D), X )
)
= HomX (L(D), X
1
= H (X, L(D)) .
1
This tells us several useful things. First, the genus g = g(X), which is typically dened as
dimk H 0 (X, X ), is also equal to dimK H 1 (X, O). Second, the desired statement of RiemannRoch is now
?
deg(D) + 1 g = dimk H 0 (X, L(D)) dimk H 0 (X, X L(D))

= dimk H 0 (X, L(D)) dimk H 1 (X, L(D))
= (X, L(D)).
Third, Riemann-Roch does indeed hold for D = 0 (by the previous two assertions).
To nish the proof, it is enough to show that the Riemann-Roch equality for a given
divisor D is equivalent to its truth for the divisor D + (Q) for any closed point Q X(k).
(With that in hand, we can walk from 0 to any other divisor by adding or subtracting points.)
So let us see how much both sides of the Riemann-Roch equality change when we add the
point Q. On one hand, obviously
(deg(D + (Q)) + 1 g) (deg(D) + 1 g) = 1.
On the other hand, we have a short exact sequence
0 L(D) L(D + (Q)) OQ 0
where OQ denotes the skyscraper sheaf k at the point Q. Since Euler characteristics add in
short exact sequences,
(X, L(D + (Q))) (X, L(D)) = (X, OQ ) = 1.
Hence Riemann-Roch for D is equivalent to Riemann-Roch for D + (Q).
Construction of the dualizing sheaf
We now go back and construct dualizing sheaves following the argument in Hartshorne (but
eshing out some details which he leaves opaque). Recall that we already know the duality
theorem for X = P , with the dualizing sheaf being the canonical sheaf P . The plan is to
manufacture a dualizing sheaf on X out of P , using Serre duality for P . That tells us that
if we x an isomorphism H N (P, P )
= k of k-vector spaces, then for any coherent sheaf F
on X,
H n (X, F ) = H n (P, j F )
= ExtPN n (j F , P ) .
So we are reduced to nding a sheaf X

on X for which there is a functorial isomorphism
n

HomX (F , X
) = ExtN
(j F , P ).
P
(We then get the required trace map H n (X, X

) k by tracing the identity element of
HomX (X
, X
) through the identications.)
One might imagine that this isomorphism comes from an isomorphism of sheaves
?

Hom X (F , X
) = Ext PN n (j F , P )
by taking global sections. Taking F = OX in this hypothetical isomorphism suggests the

right denition:
X
= j Ext PN n (j OX , P ).
We would like to get back from this to the claimed isomorphism

HomX (F , X
) = ExtPN n (j F , P ).
by rst forming the canonical identication

Hom X (F , j Hom P (j OX , ))
= Hom P (j F , )
(local version: for A a ring, I an ideal, M ModA/I , N ModA , we identify HomA (M, N)
=
HomA/I (M, HomA (A/I, N))), then evaluating the resulting derived functors at P , then
taking global sections. This is complicated by the fact that in the second step, Hom X (F , )
is not exact, and in the third step, taking global sections is not exact.
To straighten these things out, we need to know more about the relationship between the
sheaf Ext and the global Ext. For starters, here is one thing I can say using Serre vanishing.
(See Hartshorne Proposition III.6.9.)
Proposition. Let F and G be coherent sheaves on X. Then there exists an integer q0
depending on F and G, such that for every q q0 , we have
i
ExtiX (F , G(q))
(F , G)(q)).
= (X, Ext X
Proof. This holds for i = 0 without any restriction on q. For F locally free, the right side is
zero for i > 0, whereas the left side vanishes for n large enough by Serres vanishing theorem.
The general case then follows by a dimension shifting argument; see Hartshorne Proposition
III.6.9.
Next, I must recall a theorem which I skipped over earlier.
Theorem (Grothendieck). For any F ShAb (X), H i (X, F ) = 0 for i > n.
Proof. This holds with X replaced by any noetherian topological space of dimension n. The
argument is a rather elaborate dimension-shifting argument; see Hartshorne Theorem III.2.7.
(See also Hartshorne exercise III.4.8(d), which is enough for this discussion.)
Corollary. For any coherent sheaf F on X, we have Ext iP (j F , P ) = 0 for i < N n.
Proof. Put Fi = Ext iP (j F , P ). On one hand, for q large,
(P, Fi(q)) = ExtiP (j F , P (q))
= H N i(P, j F (q))
by Serre duality for P . For i < N n, H N i (P, j F (q)) = 0 by the theorem. Hence
(P, Fi(q)) = 0 for q large. On the other hand, since Fi is coherent, for q large, Fi (q) is
generated by global sections. This forces Fi (q) = 0 for q large, whence Fi = 0.
3
At this point, we can nish with the following argument; compare Hartshorne Lemma
III.7.4. (Once again, there is a spectral sequence hiding behind this, but never mind.)
Take an injective resolution I of P , so we can compute Ext (j F , P ) as the cohomology
of Hom P (j F , I ), and similarly for Ext and Hom. But using the canonical identication
from earlier, if we write J = j Hom P (j OX , I ), we can also compute Ext (j F , P ) as the
cohomology of Hom X (F , J ), and similarly for Ext and Hom. So now what we need to
know is that
?
N n
Hom X (F , X
)=h
(Hom X (F , J ))
and similarly with straight Homs.
However, the sheaves J are injective OX -modules. (Local version: if A is a ring, I
an ideal, and I ModA is injective, then HomA (A/I, M) is an injective A/I-module; this
uses the previous local identication.) By the previous corollary, the complex J (whose
cohomology computes Ext (j OX , P )) is acyclic in degrees up to N n 1. We can then
split it into two complexes of injectives J1 , J2 , where J1 is exact and only has terms in
degrees up to N n, and J2 only has terms in degrees at least N n (exercise).
Since J1 is a bounded complex of injectives, it splits into a series of split short exact
sequences; thus it stays exact no matter what left exact functors you apply to it. So we can
replace J by J2 for the purposes of computing derived functors, i.e., what we need to prove
is reduced to
?
N n
Hom X (F , X
)=h
(Hom X (F , J2 ))
and similarly for straight Hom. But J2 only starts in degree N n, and Hom and Hom are
left exact, so we have
n
Ext N
(j F , P )
= hN n (Hom X (F , J2 ))
P
= Hom X (F , hN n (J ))
2
= Hom X (F , hN n (Hom X (OX , J2 )))
= Hom X (F , Ext N n (j OX , P ))
= Hom X (F , )
X
and similarly
ExtPN n (j F , P )
).
= hN n (HomX (F , J2 ))
= HomX (F , X
That completes the proof that
X
= Ext PN n (j OX , P )
is a dualizing sheaf for X.
Calculation of the dualizing sheaf for smooth schemes
To nish the proof of Riemann-Roch, we must still show that we can take X
= X when
X is smooth over k. Fortunately, this is a local problem.
Theorem. Suppose that X is a local complete intersection in P . Let I be the ideal sheaf of
X. Then there is a canonical isomorphism
Ext rP (j OX , P )
= P j OX r (I/I 2 ) .
The local complete intersection condition asserts that I is locally generated by N n
elements; this is true for X smooth basically by the Jacobian criterion. See Hartshorne
Theorem II.8.17. The fact that the right side gives X comes from the exact sequence
0 I/I 2 P/k j OY j Y /k
by taking exterior powers; see Hartshorne Proposition II.8.20. The stated theorem itself is
proved by computing in local coordinates; see Hartshorne Theorem III.7.11
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Cohen-Macaulay schemes and Serre duality
In this lecture, we extend Serre duality to Cohen-Macaulay schemes over a eld. As in the
previous lecture, let k be a eld (not necessarily algebraically closed), let j : X P = PN
k
be a closed immersion with X of dimension n, and let OX (1) be the corresponding twisting
sheaf.
Cohen-Macaulay schemes and duality
Let X
denote a dualizing sheaf on X; remember that this choice includes a trace map
n
H (X, X
) k. We then obtain natural functorial maps
i : ExtiX (F , X
) H ni(X, F )
because both sides are cohomological functors on the opposite category of coherent sheaves
on X, and the one on the left is eaceable because it vanishes on direct sums of twisting
sheaves. By the denition of a dualizing sheaf, 0 is always an isomorphism.
Theorem. The following conditions are equivalent.
(a) The scheme X is equidimensional (each irreducible component has dimension n) and
Cohen-Macaulay.
(b) The maps i are isomorphisms for all i 0 and all coherent sheaves F on X.
This is of course meaningless if I dont tell you what a Cohen-Macaulay scheme is. For
the moment, suce to say that a scheme is Cohen-Macaulay if and only if each of its local
rings is a Cohen-Macaulay ring. That already has content, because then the theorem says
that (b) is equivalent to a local condition on X, which is far from obvious.
Ill also point out that a regular local ring is always Cohen-Macaulay. This implies the
following.
Corollary. If X is smooth over k, then i is an isomorphism for all i 0 and all coherent
sheaves F on X.
Proof of the duality theorem, part 1
Even without knowing what a Cohen-Macaulay scheme is, we can at least start working to
prove that condition (b) is equivalent to a local condition on X. Let us start by relating (b)
to two global vanishing assertions.
Lemma. The following conditions are equivalent to (b).
(c) For any locally free coherent sheaf F on X, for q suciently large, we have H i (X, F (q)) =
0 for all i < n.
(c ) For q suciently large, we have H i (X, OX (q)) = 0 for all i < n.
Note that condition (c) is a sort of opposite to Serres vanishing theorem, which gives
the vanishing of H i (X, F (q)) for i > 0 and q suciently large.
Proof. Given (b), for any locally free coherent sheaf F on X, we have
ni

H i(X, F (q)) = ExtX
(F (q), X
)
= Extni
X (OX , F X (q))
= H ni(X, F X
(q))
and this vanishes for n i > 0 and q large by Serres vanishing theorem. Thus (b) implies
(c).
It is clear that (c) implies (c ). Given (c ), it follows that H ni(X, ) is eaceable for all
i > 0 since we can cover F with a direct sum of twisting sheaves. Hence i is the natural map
between two universal cohomological functors, hence is an isomorphism. Thus (c ) implies
(b).
We next reformulate this in local terms, using Serre duality on P .
Lemma. The following condition is equivalent to (b).
i
(d) For all i < n, Ext N
(j OX , P ) = 0.
P
i
Remember that no matter what X is, we have Ext N
(j OX , P ) = 0 for i > n: we
P
proved this in the course of constructing the dualizing sheaf X

.
Proof. By Serre duality on P (and choosing an isomorphism H n (P, P )

= k), we may identify
H i (X, OX (q))
= H i(P, j OX (q))
= ExtPN i (j OX , P (q)).
i
So (c) is equivalent to the condition that for q suciently large, ExtN
(j OX , P (q)) = 0
P
for all i < n. Recall from earlier that for q large,
i
ExtPN i (j OX , P (q)) = (P, Ext PN i (j OX , P (q))) = (P, Ext N
(j OX , P )(q)).
P
i
i
Since Ext N
(j OX , P ) is coherent, (P, Ext N
(j OX , P )(q)) vanishes for q suciently
P
P
i
large if and only if Ext N
(j
O
,
)
=
0.
X
P
P
Condition (d) can be rewritten as follows.

(e) For each point x X, if A = OP,x and I is the ideal of A dening X at x, then for all
i
i < n, ExtN
A (A/I, A) =
0.
Proof. This translates directly from (d) once we remember that P is locally free of rank 1
on P .
This is almost the local condition we are seeking, except that it still refers to the position
of X within P .
2
The Cohen-Macaulay condition
To get rid of the dependence of our duality condition on the relative geometry of X within
P , we need some more sophisticated commutative algebra.
Proposition. Let A be a regular local ring and let M be a nitely generated A-module. Then
for any nonnegative integer n, the following are equivalent.
(a) We have Exti (M, A) = 0 for all i > n.
(b) For any A-module N, we have Exti (M, N) = 0 for all i > n.
(c) There exists a projective resolution 0 Ln L1 L0 M 0 of M at
length at most n.
Proof. See Hartshorne Proposition III.6.10A (and associated Matsumura reference) and ex
ercise III.6.6.
The smallest integer for which this holds is called the projective dimension of M (if it
exists), denoted pdA (M). For instance, M is projective if and only if pdA (M) = 0.
For M a module over a ring A, a regular sequence is a sequence x1 , . . . , xn of elements
of A such that for i = 1, . . . , n, xi is not a zerodivisor on M/(x1 , . . . , xi1 )M. For A a local
ring, the depth of M is the maximal length of a regular sequence with all xi in the maximal
ideal of A.
Proposition. For A a regular local ring and M an A-module,
pdA (M) + depthA (M) = dim(A).
Proof. See Hartshorne Proposition III.6.12A (and associated Matsumura reference).
We can nally give a local equivalent to condition (b) from the duality theorem. Recall
that our last equivalent (e) said that for each x X, for A = OP,x and I the ideal of A
N i
dening X at x, ExtA
(A/I, A) = 0 for all i < n. This is equivalent to pdA (A/I) N n,
and hence to depthA (A/I) n. The trick is that if M is an A/I-module, then depthA (M) =
depthA/I (M). Thus we have the following.
(f ) For each point x X, if B = OX,x , then depthB (B) n.
On the other hand, we always have depthB (B) dim(B) n, so it is equivalent to
require depthB (B) = dim(B) = n.
This condition depthB (B) = dim(B) is in fact the denition of a Cohen-Macaulay lo
cal ring B. Any regular local ring is Cohen-Macaulay, since we can use generators of the
cotangent space as a regular sequence. But the Cohen-Macaulay condition is much more
permissive; for instance, any local complete intersection is Cohen-Macaulay.
3
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009

Higher Riemann-Roch
In this lecture, we discuss some higher-dimensional versions of the Riemann-Roch theo
rem: the Riemann-Roch theorem for surfaces, the Hirzebruch-Riemann-Roch theorem, and
the Grothendieck-Riemann-Roch theorem. For the rst, see Hartshorne V.1; for the others,
see Chapter 15 of Fultons Intersection Theory (a well-deserved winner of the Steele Prize
for mathematical exposition).
Surfaces
Let X be a smooth irreducible projective surface over an algebraically closed eld k. Let K
be a canonical divisor on X. As in the case of curves, the Riemann-Roch theorem combines
an input from Serre duality with an Euler characteristic calculation.
The input from Serre duality is that for any divisor D,
H 2 (X, L(D)).
H 0 (X, L(D) X ) =
We can thus write the Euler characteristic (X, L(D)) as
dimk H 0 (X, L(D)) dimk H 1 (X, L(D)) + dimk H 0 (X, L(K D)).
Unfortunately, we cant do much with the term dimk H 1(X, L(D)) other than give it a name:
its called the superabundance of D. However, we do at least know that it is nonnegative,
and this turns out to be surprisingly useful.
The Euler characteristic calculation is made as follows. Write D as the dierence between
two eective divisors C E with no common components. We then have exact sequences
0 L(C E) L(C) L(C) OE 0,
0 OX L(C) L(C) OC 0.
By additivity of , we get
(X, L(C E)) = (X, OX ) + (C, L(C)) (E, L(C)).
The rst term we are happy to leave alone since it depends only on X. The other two are
calculated using intersection theory on the surface X. For instance, the term (E, L(C))
equals C E + 1 gE , where gE is the genus and C E is the length of the scheme-theoretic
intersection C X E (this amounts to Riemann-Roch on the curve E).
The term (C, L(C)) is a bit trickier: it is C C + 1 GC where C C = C 2 is the
self-intersection of C. That can be dened as C C if C is linearly equivalent to a divisor C
having no common components with C, but that is not always possible. In fact, the correct
denition is to force the intersection pairing to be bilinear, and this sometimes involves
letting C 2 take negative values. For instance, if you blow up P 2 at a point, the exceptional
divisor has self-intersection 1. (This is a general pattern; one can in fact blow down any
curve isomorphic to P1 with self-intersection 1.)
1
Moreover, one can write the genera of C and E in terms of the canonical divisor K,
basically using Riemann-Roch again:
C (C + K) = 2gC 2,
E (E + K) = 2gE 2.
So
1
(X, L(D)) = D (D K) + (X, OX ).
2
As in the case of curves, this is useful for many calculations involving the geometry of
surfaces, such as the Hodge index theorem and the Nakai-Moishezon criterion. These in
turn gure in the classication of surfaces (which you should read about in Hartshorne if
you are interested in Abhinavs work).
Theorem (Hodge index theorem). Fix a projective embedding of X, and let H be a divisor
with L(H)
= OX (1). Then for any divisor D such that D H = 0, we have D 2 0. (This
also holds if H is ample, i.e., some positive multiple of H comes from an OX (1).)
Theorem (Nakai-Moishezon criterion). A divisor D on X is ample if and only if D 2 > 0
and D C > 0 for all irreducible curves C on X.
Hirzebruchs generalization
Hirzebruch noticed that the Euler characteristic aspect of Riemann-Roch could be general
ized to handle arbitrary vector bundles on arbitrary smooth varieties over an algebraically
closed eld k. Let me state his result and then explain what it means.
Theorem (Hirzebruch). Let X be a smooth proper scheme over k. Let F be a locally free
coherent sheaf on X. Then
(X, F ) =
ch(F ) td(TX ).
X
Here TX is the tangent bundle of X, i.e., the dual to the bundle X of Kahler dierentials
(which is also called the cotangent bundle).
The Chern character ch is a certain map from coherent sheaves on X to a certain group
of cycles on X. The latter are formal Q-linear combinations of subschemes of X modulo a
relation of rational equivalence. You should imagine this as generalizing the function taking
a line bundle L on a curve C to (the equivalence class of) the divisor of a nonzero rational
section of L.
The group of cycles is graded by codimension, and forms a commutative ring under the
(appropriately dened) intersection pairing with the identity
being the class of X itself in
codimension 0. The Chern character is usually split up as d cd () with cd being the bit in
codimension d; for F locally free of rank 1, we always have
cd (F ) =
1
c1 (F )d.
d!
2
The Todd class td is another such map on coherent sheaves, which I wont try to construct
here, except to give the characterizing identity: for F locally free of rank d,
td(F )
(1)p ch(p F ) = cd (F ).
p=0
. The point is that it depends only on X, not on F .

The Chern character and the Todd class are both examples of characteristic classes
of vector bundles, which originally appeared in algebraic topology as tools for classifying
manifolds. For instance, Milnor uses them to construct dierentiable manifolds which are
homeomorphic but not dieomorphic to the 7-sphere, the so-called exotic 7-spheres. See
Milnor and Stashe,
Characteristic Classes for an introduction.
Oh, and X means use intersection theory (which is a pretty complicated thing to dene,
as evidenced by the length of Fultons book), keep only the zero-dimensional part, and count
points.
Grothendiecks generalization
In characteristic fashion, Grothendieck noticed that one can make a relative version of the
Hirzebruch-Riemann-Roch theorem. Also, one can drop the locally free condition.
Theorem (Grothendieck). Let f : X Y be a proper morphism of smooth schemes over
an algebraically closed eld k. Then for any coherent sheaf F on X,
ch(f F ) td(TY ) = f (ch(F ) td(TX )).
One has to dene direct image for cycles; I wont try here.
It should be noted that already our formulation of Hirzebruchs statement is Grothendiecks;
the original statement was made in the language of topology. One byproduct of this work
is the development of K-theory, which is now a frequently occurring construction in both
algebraic topology and algebraic geometry.
MIT OpenCourseWare
http://ocw.mit.edu

Spring 2009
Etale
cohomology (updated 13 May 09)
In this lecture, we give a hint of the theory of etale cohomology. Standard references:
Milne, Etale
Cohomology (he also has some more accessible lecture notes online at jmilne.org);
Tamme, Introduction to Etale

Cohomology; Freitag-Kiehl, Etale
Cohomology and the Weil
Conjectures. You might want to read Hartshorne Appendix C rst for an overview. Also,
note that there is a rogue volume of SGA called SGA 4 1/2, written mostly by Deligne
after the fact, which gives a surprisingly legible (albeit in French) account of this stu.
Since this is the last lecture of the course, I would like to take the opportunity to thank
you, the participants, for all the hard work you put in on the problem sets, and especially
for all your feedback on the notes. If you have further questions about algebraic geometry,
from the general to the specic, I would be happy to discuss them!
Motivation: the Weil conjectures
Let X be a variety over a nite eld Fq . Weil predicted that the zeta function of X, dened
as an Euler product
X (T ) =
(1 T deg(xFq ) )1
x
over the closed points of X, could always be interpreted as the power series expansion of a
rational function of T ; this analogizes the analytic continuation of the Riemann zeta function.
For instance, for X = P1 ,
1
X (T ) =
.
(1 T )(1 qT )
Weil also predicted analogues of the functional equation of the zeta function, and the Rie
mann hypothesis. For instance, for X an elliptic curve, Hasse proved that
X (T ) =
1 aT + qT 2
(1 T )(1 qT )
for some a Z. This expression has the symmetry property that

X (q 1 /T ) = X (T ).
(This example is a bit lucky; more generally, you might be o by a factor of q a T b for some
a, b Z. For X of pure dimension n, you should compare X (T ) with X (q n /T ).) Hasse
also proved that
|a| 2 q,
or equivalently, the numerator polynomial 1 aT + qT 2 has complex roots of norm q 1/2 .
Weil also noticed that the degrees of the factors in the zeta function appeared to have
topological meaning. Namely, if X is obtained from a smooth proper scheme over some
1
arithmetic ring (i.e., a localization of the ring of integers in a number eld) by reduction
modulo a prime, then the degrees of the factors in X (T ) correspond to the Betti numbers
of (X C)an . For example, the degrees of the factors 1 T, 1 aT + qT 2 , 1 qT in the
elliptic curve case match the Betti numbers 1, 2, 1 of a genus 1 Riemann surface.
Weil proved analogues of all these assertions for arbitrary curves, and (based on some
evidence from Fermat hypersurfaces) conjectured analogues for higher dimensional varieties.
More precisely, he predicted the existence of a cohomology theory H i() for varieties over
Fq , taking values in nite dimensional vector spaces over a eld K of characteristic zero, in
which the number of Fq -rational points (i.e., the xed points of the q-power Frobenius map)
could be computed using an analogue of the Lefschetz xed point formula in topology:
2 dim(X)
#X(F ) =
qn
(1)i Trace(Fqn , H i(X)).
i=0
This immediately implies rationality of X (T ). Symmetry should follow from a form of

Poincare duality, i.e., a perfect pairing
H i (X) H 2 dim(X)i (X) H 2 dim(X) (X) K.
The Riemann hypothesis is not quite as purely formal a consequence, since it is basically a
nonnegativity condition, whereas K need not have anything to do with R. But never mind
that for now.
Curves
For curves, Weil proved his conjectures by constructing an algebraic group associated to a
curve C, called the Jacobian variety J(C). Over C, this gives a complex torus which had
been constructed by Abel-Jacobi using abelian integrals.
For a prime not equal to the characteristic of Fq , and a positive integer n, the group
J(C)(Fq )[n ] of geometric n -torsion points is abstractly isomorphic to (Z/n Z)2g , for g the
genus of C. The absolute Galois group of Fq acts by (Z/n Z)-module endomorphisms. If we
take the inverse limit over n, we get a Z -module T (J(C)) equipped with an action of the
absolute Galois group; it is nowadays called the Tate module of C. (For instance, if C is an
elliptic curve, then J(C) = C.)
This gives the H 1 (or really its dual) in a good cohomology theory. The symmetry comes
from the Tate pairing. The Riemann hypothesis can be deduced using the Hodge index
theorem, which gives a nonnegativity (or really a nonpositivity) assertion for the intersection
pairing on C Fq C.
Aside: a noncohomological proof, using only Riemann-Roch and some clever estimates,
was found later by Stepanov (and simplied by Bombieri). Good reference: Lorenzinis
Invitation to Arithmetic Geometry.
Why
etale?
One might think that coherent sheaf cohomology, as we have developed in this course,
might be useful against the Weil conjectures. However, it has several problems: it lives in
characteristic p rather than characteristic 0 (so it can only aspire to prove rationality mod p,
rather than integrally), and its dimensions do not match Betti numbers. For instance, sheaf
cohomology on a scheme of dimension n only goes up to index n, rather than 2n.
Grothendieck realized that one might get around this by trying to make an analogue of
topological cohomology in which etale maps play the role of local homeomorphisms. For
instance, recall one of the consequences of GAGA: for a smooth proper variety X over C,
every nite covering space map comes from a unique nite etale cover of X. Thus the
pronite completion of the topological fundamental group can be recovered as an inverse
limit of Galois groups of these etale covers.
Perhaps a better justication for considering etale covers is the following. For X a
complex analytic variety and x X, the local ring OX,x , while not complete, is henselian:
the conclusion of Hensels lemma still holds. (That is, given a polynomial over OX,x , any
simple root of the reduction modulo the maximal ideal lifts uniquely to a root.) This is
not true for schemes, though. A related geometric statement is that if f : Y X is an
etale morphism of schemes, and x X is a point, then there is no way to draw disjoint
open neighborhoods of the points of f 1 (x), so you cannot view the etale map as a local
homeomorphism.
Topology revisited
In order to combine the ideas about etale covers with sheaf cohomology, Grothendieck had
to take the apparently drastic step of modifying the notion of a topology on a space. But
in retrospect, this isnt such a strange modication to make. After all, presheaves on a
topological space X are nothing more than contravariant functors on the category X of open
sets. Why not state all the sheaf axioms in terms of the structure of that category?
Grothendieck realized that stating the sheaf axiom really only requires knowing what
an open cover is, leading to the following denition. Let C be a category admitting bre
products. A Grothendieck topology consists of the following data. For each X C, you must
tell me which collections of morphisms {Ui X}iI are coverings of X. This prescription
must satisfy some hypotheses.
Any isomorphism X Y is by itself a cover of Y .
For any Y X, if {Ui X} is a cover, then {Ui X Y Y } is a cover. That is,
open covers can be restricted to open subsets.
If {Ui X} is a cover, and for each i {Vij Ui } is a cover, then {Vij X} is also a
cover. That is, covering each open in a cover gives a cover.
(Strictly speaking, this is a Grothendieck pretopology because it only gives you the analogue
of a basis for a topology. You should really throw in all coverings generated by these too.)
A category equipped with a Grothendieck topology is called a site. For instance, the big
etale site of a scheme S is the category of all S-schemes, in which coverings are collections of
etale morphisms which form a set-theoretic cover. That is, {Ui X} is a cover if and only
if each Ui is etale and the union of their images is X. (If you only bother keeping objects
which are themselves etale over S, you get the small etale site.)
There are many other useful Grothendieck topologies that occur frequently in algebraic
geometry. These include the fppf topology (dèlement plat de presentation nie = faithfully
at of nite presentation), the fpqc topology (dèlement plat et quasicompact = faithfully
at quasicompact), the smooth topology, the at topology, the syntomic topology (at and
locally of nite presentation), the Nisnevich topology (etale, but each point must be covered
by a point with the same residue eld), etc. There are also useful examples where you start
with a usual topological space but use only some of the available open covers; this occurs in
the denition of rigid analytic spaces (i.e., analytic spaces over a nonarchimedean complete
eld like Qp ).
Anyway, once you know what a Grothendieck topology is, you can dene a sheaf of
abelian groups (say) on it. Namely, you want a contravariant functor F from your category
to Ab, such that for any cover {Ui X}, we have an exact sequence
0 F (X)
F (Ui )
F (Ui X Uj )
i
i,j
where the last map computes a section on F (Ui X Uj ) as the restriction from Ui minus the
restriction from Uj . For instance, in most reasonable cases, the structure sheaf F (X) = OX
is a sheaf.
There is also a notion of sheacation but this is complicated by the fact that we dont
have points with with to dene stalks. No matter: what are points anyway but decreasing
families of open sets? One can make an articial denition of points in that fashion; this
brings one dangerously close to the notion of a topos, which I will skip over entirely. (Roughly
speaking, a topos is the category of sheaves on a site with values in a given category, like
sets or abelian groups.)
Etale
cohomology in practice
We can now dene sheaf cohomology on any site with a nal object as the derived functors
of global sections, meaning sections over the nal object. (One can x this even if there is no
nal object, by taking a compatible family of sections over every element of the site. Yeesh.)
However, its not so straightforward to compute etale cohomology of a scheme X with
coecients in a sheaf F . On one hand, writing down etale cochains is not a problem: you
specify an etale cover of X and then some sections on each element of the cover. Writing
down cocycles isnt that much harder: you have to write down another etale cover on which
4
you can check that the dierential of your cochain vanishes. The hard part is, given a
cochain, how do you tell whether it is zero or not?
Despite this complication, one can prove quite a lot. For instance, if you start with a
quasicoherent sheaf F on a scheme X, you get a sheaf on its big and small etale sites by
setting the sections over an open i : U X to be i F . But this is a boring example,
because the resulting sheaf cohomology turns out to agree with usual sheaf cohomology on
the Zariski site (i.e., what we already know).
What makes the etale site fun is that you get strange new sheaves, much more akin
to the locally constant sheaves in topology, and their cohomology is quite interesting. For
instance, you can make a locally constant sheaf associated to any (pro)nite abelian group (by
sheafying the constant presheaf), and this gives you something with topological meaning.
Theorem. Let X be a smooth proper scheme over C. Then for any prime , the cohomology
of the etale locally constant sheaf associated to the -adic integers Z computes the topological
Betti numbers of X.
The fun comes when you start with a scheme over an arithmetic base, like Q. If you
extend the base to Q and then take etale cohomology with coecients in Z , the result
carries an action of the absolute Galois group of Q. E.g., for an elliptic curve, the rst etale
cohomology is (dual to) the -adic Tate module, i.e., the inverse limit of the -power torsion
groups viewed as a Galois representation.
Back to the Weil conjectures
Let X be a smooth proper scheme over the nite eld Fq . Pick any prime =
q. For each
n
positive integer n, we can consider the locally constant etale sheaf Z/ ZX on X. Let Z X be
the inverse limit of these; this is not the same as the locally constant etale sheaf generated
by Z . (E.g., in the example of the elliptic curve, that is because the -power torsion is not
dened over a nite extension of the base eld.)
Nonetheless, Z is a good sheaf to work with. (It is an example of a sheaf which is lisse,
or smooth if you prefer to translate from the French.) We will be interested in working with
the
i
H i (X) = Het
(X Fq Fq , Z ) Z Q ,
which is a collection of Q -vector spaces. These turn out (with some eort) to be nite
dimensional over Q , and carry a Lefschetz trace formula. This proves rationality of the zeta
function.
Aside: rationality had already been proved by Dwork around 1960 using p-adic analytic
methods, but not using cohomology. Nowadays, though, Dworks proof has been reinter
preted in terms of a dierent Weil cohomology, called rigid cohomology, taking values in a
p-adic eld. (Remember that = p is excluded in etale cohomology, because this case be
haves badly. For instance, an elliptic curve over an algebraically closed eld of characteristic
p has at most p points killed by p, not p2 .)
5
Returning to etale cohomology, there is also a Poincare duality which implies symmetry.
The Riemann hypothesis, of course, is more subtle; Grothendieck had predicted it would
follow from a suitable analogue of the Hodge index theorem, which was one of his standard
conjectures. This analogue is still open; instead, Deligne proved the Riemann hypothesis by
a rather clever combination of ideas, including an algebro-geometric variant of the Rankin
squaring argument from classical modular forms. Laumon later gave a similar but techni
cally simpler proof by adding the use of a cohomological Fourier transform. (These proofs
are largely independent of which Weil cohomology you are using. In particular, with some
eort they can be transposed into rigid cohomology.)

Mit 726

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Mit 726

Încărcat de

Drepturi de autor:

Formate disponibile

MIT OpenCourseWare

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

Where we were, and where we need to go

Paradigm shift 1: sheaves

Paradigm shift 2: functors

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

Warning: set-theoretic diculties

Categories, and examples

A category C consists of the following data.

Interlude: is versus does

from X to Y . An object which is both initial and nal is a terminal object.

A morphism f Hom(X, Y ) is a monomorphism if for any g, h Hom(W, X), if

f g = f h, then g = h. In the category of sets (and many other examples), f is a

monomorphism if and only if f is injective.

A morphism f Hom(X, Y ) is an epimorphism if for any g, h Hom(Y, Z), if g f =

h f , then g = h. In the category of sets (and many other examples), f is an

epimorphism if and only if f is surjective. But beware of surprises: for example,

the morphism Z Q of rings is an epimorphism (and also a monomorphism).

A morphism f Hom(X, Y ) is an isomorphism if it has a two-sided inverse. This im

Functors and natural transformations

Functors can be thought of as functions between categories. A covariant functor from a

are naturally isomorphic.

Other properties of functors

6 Representable functors, Yonedas lemma, and uni

Limits and colimits

Another important example for us: let R S be a homomorphism of rings. Let F :

R S is the map x x 1. For Y an S-module, Y R S Y is the map

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

F (U ) F (V ) (called restriction), such that:

(i) for each open subset U of X, ResU,U = idF (U ) ;

Dening sheaves on a basis

In other words, the restriction functor from sheaves on X to sheaves on X specied on

(b) For each i, j I, an isomorphism ij : Fi |Ui Uj

= Fj |Ui Uj , satisfying the following

An important source of information about sheaves is given by looking at their behavior in

Stalks and morphisms

If we x a topological space X and a category C, there is an obvious forgetful functor from

Direct and inverse image

Let f : X Y be a continuous map. For F a sheaf on X, the formula

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

be a commuting diagram in C with exact rows.

(a) If f1 and f3 are monomorphisms and f0 is an epimorphism, then f2 is a monomor

be a short exact sequence. Then there exists a canonical homomorphism : ker(f3 )

0 ker(f1 ) ker(f2 ) ker(f3 ) coker(f1 ) coker(f2 ) coker(f3 ) 0

be a commuting diagram in C with exact rows. Then f2 is a monomorphism/epimorphism if

For C1 = C2 = Ab, a covariant functor F : C1 C2 is additive if it commutes with addition

Let F be a sheaf on a topological space X with values in C = Ab. A subsheaf of F is what

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

Ringed and locally ringed spaces

A ringed space is a topological space X equipped with a sheaf OX on X with values in

The prime spectrum of a ring

The fundamental theorem of ane schemes

i M R Rfi is injective, as follows. Suppose m M belongs to the kernel of this map.

18.726 Algebraic Geometry

18.726: Algebraic Geometry (K.S. Kedlaya, MIT, Spring 2009)

Direct and inverse image

Morphisms of (locally) ringed spaces

To do this, write any h Ag as a/g i with a A and i Z0 . We can then map a to