MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS)

LECTURES BY RAVI VAKIL

NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

CONTENTS
About these notes 1
1. Introduction 2
2. The category of presheaves on C (aka the functor category) 4
3. Adding in the topology: Grothendieck topologies and sheaves 5
4. Toward stacks: things that are not quite sheaves 10
5. Lecture Five 15
6. Fiber products and representable morphisms 16
7. Lecture Six 19
8. Geometry 21
9. Lecture Seven 22
10. Lecture Eight 25
11. Relationships among Topologies 28
12. Local Representability and Moduli of Curves 32
13. Lecture Eleven 36
14. Lecture Twelve 38
15. Lecture Thirteen 41
16. Lecture Fourteen 43
17. Lecture Fifteen 47
18. Lecture Sixteen 51
19. Lecture Seventeen 55
20. Lecture Eighteen 60
21. Lecture Nineteen 63

ABOUT THESE NOTES

These notes are live-texed by Tony and Dan, and hopefully will be edited by people
participating in the course, including Ravi.
The content is intended to be in logical order. Things were supposed to be presented
in logical order, but that never happens; so the material is not in chronological order
(where it has been sufficiently edited), and it is not in logical order (where it hasn’t been
sufficiently edited).
Date: June 7, 2024.
1
 2 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Participants in the class can edit the file. Others can see the results at this link
https://www.overleaf.com/read/fytmvtjbqyth
(which is read-only). So please share this link, not the editing link.

FIGURE 1. Part of the hierarchy of the abstract nonsense involved in the

world of stacks

category fibered in groupoids

t +
pre-stack presheaf/functor

 
stack / sheaf o separated presheaf

 
Artin stack locally representable sheaf

 
DM stack "geometric space"


algebraic space


scheme
An arrow from A to B means "B is an example of A". The bottom, starting with Artin
stack on the left and locally representable sheaf on the right, is "geometry", and the
top is "abstract nonsense"

1. INTRODUCTION
Lecture 1,
1.1. Why stacks? This course is about a new object in algebraic geometry: (algebraic)
stacks. The theory of stacks was motivated by certain geometric examples, and we will
try to motivate this theory starting from these examples, which include:
• Orbifolds.
• “Moduli spaces” that are not representable, such as moduli spaces of curves.
• Quotients by group actions. These include spaces like
– BG, from toplogy. We want to make sense of statements like
dim BGL(2) = −4.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 3

– More generally, if G is a group acting a scheme X , we want to be able to

make sense of a quotients [X /G ].
• Understanding why the j -line is “wrong” for classifying elliptic curves. (It is
really a “coarse” version of a stack.)

1.2. Geometric spaces. We are interested in geometric spaces such as :

• balls in Rn or Cn ,
• real or complex manifolds,
• complex-analytic varieties,
• varieties,
• schemes,
• algebraic spaces
These are all topological spaces, equipped with sheaves of functions. We denote such
by a ringed space (X , OX ). Morphisms are continuous maps (of topological spaces) X →
Y , which pulls backs functions, i.e. together with a map
OY → π∗ OX .
But that won’t be relevant (for quite some time).
I want to collect some facts about your geometric category that we will find useful.
First, there are various types of morphisms that are "nice" in various ways, notably:
• they are preserved by composition (this requires only morphisms)
• they are preserved by pullback
• they are local in nature (this requires the topology)
Another key feature is that there is a type of morphism called "open embedding",
that is "nice" in the sense described just above. Because these will later not be "embed-
dings", we need another name for them. Unfortunately, because "open map" already
has a meaning, "open morphism" is bad terminology. Any ideas? For now, I’ll call them
o-morphisms.
A key feature is that we know which morphisms are open embeddings, and we know
how to recognize covers. In the following discussion, take G to be your favorite cate-
gory of geometric spaces.
Exercise 1.2.1. What is G ?

1.3. Moduli spaces. The notion of moduli space is a central notion in algebraic geome-
try. We’re going to give a precise definition of (fine) moduli spaces shortly. However, it’s
worth remarking that the “ancients” recognized moduli spaces, even before they knew
how to properly define them — for instance, the Grassmannian G(k , n ).
Example 1.3.1. The moduli space Mg is a hypothetical space equipped with a univer-
sal family Cg → Mg such that given any family of genus g “curves” X → S , there is a
map f : S → Mg such that f ∗ Cg ' X .
We can formalize this as follows. Given a functor F : Schemeso p p → Sets, a (fine)
moduli space for F is a space MF such that F (−) ' Mor(−, Mg ) (an isomorphism of
functors). This object MF is unique up to isomorphism, if it exists.
Example 1.3.2. The Hilbert functor and Quot functor have moduli spaces.
4 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

2. THE CATEGORY OF PRESHEAVES ON C (AKA THE FUNCTOR CATEGORY )

Let C be a category. (Later, it will be your favorite category G .) Informally, Yoneda’s
Lemma says that “we know an object through maps to it”.
For every Y ∈ C , we have a functor Mor→Y : (C o p p → Sets defined on objects by

X 7→ Mor(X , Y ).

and defined on morphisms by (X ← Z ) 7→ (Mor(X , Y ) → Mor(Z , Y )).

From C we can built the "functor category" PSh(C ), whose objects are contravariant
functors
F : C → Sets
and the morphisms F → G are the natural transformations. There is a (covariant) func-
tor
Y o : C → Fun(C )
sending Y 7→ Mor(−, Y ).
I prefer to call them the presheaves on C , for a number of reasons. Some of them: (1)
these will be the ’things’ we care about, so having them at the level of ’thing’ rather than
’map’ makes things cleaner. (2) it makes the covariance/contravariance unconfusing.
Admittedly, it is weird having the notion of presheaves without any topology in the
picture. But that’s okay — a topology will soon turn up.
Exercise 2.0.1 (Yoneda Lemma (version I like best)). Suppose X ∈ C , and F ∈ PSh(C ).
Give a bijection
M o r (Y o (X ), F ) ↔ F (X )
that is functorial in both X ∈ C and F ∈ Psh(C ).
In particular, by taking F = Y o (Y ) for Y ∈ C , we have shown:

Theorem 2.0.2 (Yoneda Lemma). The functor Y o is full and faithful.

But we get a bit more than Yoneda; we see that we can recover the category P Sh (C )
(i.e. the morphisms and objects) by simply understanding maps from elements of C
(and their functoriality).
Definition 2.0.3. A presheaf F ∈ PSh(C ) is representable if it is isomorphic to Y o (Y ),
for some Y ∈ C .
We thus think of C as being a subcategory of PSh(C ).

2.1. Fibered products, and representable morphisms.

Exercise 2.1.1. Show that PSh(C ) has fibered products. (This boils down to the fact
that sets have fibered products.)
Exercise 2.1.2. Assuming that C has fibered products, show that the fibered product
of representable presheaves/functors is representable.
We can thus define representable morphism.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 5

Definition 2.1.3. A morphism F → G in PSh(G ) is representable if the base change

against a map hX → G from a representable functor hX is representable:
h? /F

 
hX /G

Moreover, any "quality" of morphism in C that behaves well under pullback can be
extended to morphisms in PSh(C ).
For example, we have a notion of o-morphism in PSh(G ), and also open cover.
Exercise 2.1.4. Show that
Mgno
,1
aut


Mgno aut

is a representable morphism of presheaves (on the category of schemes). (possible

homework)
(This is something that I am thinking about having someone write up.)
Exercise 2.1.5. Let X , Y ∈ G and F ∈ PSh(G ). Suppose we have maps
Y
b

X /F
a

The fibered product is a sheaf denoted IsomF (X , Y ) . Show that X ×F Y (T ) is the set
f g
of pairs (T − → Y ) such that f ∗ a = g ∗ b ∈ F (T ).
→ X ,T −

3. ADDING IN THE TOPOLOGY: GROTHENDIECK TOPOLOGIES AND SHEAVES

3.1. (Not-so-)Toy example. We’re going to describe how to “discover” the category of
complex manifolds, by a process analogous to the one we’ll eventually use to define
algebraic stacks.
Let C be the category of open subsets of Cn , with morphisms being holomorphic
maps. We know that there is a Yoneda embedding
C ,→ PSh(C ).
Let D be the category of complex manifolds. The inclusion C ,→ D factors the Yoneda
embedding
C → D → PSh(C ).
Exercise 3.1.1. Show that D → PSh(C ) is a fully faithful embedding. (This is the Yoneda
result mentioned above, which is an easy consequence of the Yoneda-I-like.)
Question. Can we characterize the image of D in PSh(C )?
 6 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Similarly, one could imagine describing schemes as a full subcategory of PSh(Ringsopp ).

The kind of stacks we will be most interested in are understood in the same way.
3.2. Motivation: Sheaves on a topological space. First we review the notion of a sheaf
on a toplogical space. Let X be a topological space. The category OpenX of open sub-
sets of X has objects the open subsets, and morphisms the inclusions.
Definition 3.2.1. A presheaf on X is a contravariant functor
F : OpenX → Sets.
A presheaf F is a sheaf if it satisfies “identity” and “gluing” axioms. Sheaves on X form
a category Shv(X ).
More generally, we want to talk about sheaves on categories. Suppose we have a
(contravariant) functor F : C → Sets ∈ Fun(C opp ). We say that F is a sheaf if it sat-
isfies identity and gluing axioms. To make sense of this, we need to have a notion of
“open embedding” and a notion of “cover”. We then define the morphisms of sheaves
as morphisms as presheaves.
In particular, Sh(C ) is a full subcategory of PSh(C ) (by definition!).
3.3. Grothendieck topologies. A Grothendieck topology is a notion of a “topology on
a category”. What does this mean?
We need to specify two things:
• a distinguished class of morphisms (analogous to “open embeddings”). We are
temporarily calling them o-morphisms
• a notion of when a collection of “open embeddings” forms a “cover”.
The “covers” satisfying some natural axioms:
• Isomorphisms are covers.
• Covers of covers are covers.
• Pullbacks of covers are covers. (Reminder: we don’t need all pullbacks to ex-
ist in G — for example, the category of polydiscs and holomorphic maps is not
closed under pullback. What we do need is that if f : X → Y is an “open embed-
ding”, then the pullback of f by any morphism Y 0 → Y exists and is an “open
embedding”.)
Lecture 2 A “sheaf on G ” is then a presheaf that has the “gluing” and “identity” axioms.
3.4. Examples (and non-examples) of sheaves. An important property of your cat-
egory G is that “maps glue”, in some sense. (Our earlier statement that “morphisms
in G glue” means exactly that the representable functors in PreShv(G ) are sheaves.)
This meaning of this statement includes the specification of a collection of morphisms
which are declared to be “open embeddings”. This means we have a (not full!) subcat-
egory G open , with the same objects but whose morphisms are the “open embeddings”.
So we can think of spaces as being (special cases of ) sheaves. In fact, I think this is a
central input/axiom we need — that objects of G are sheaves.
We now have this picture:

G ,→ Sh(G ) ,→ PSh(G ).
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 7

where each "inclusion" is faithful.

Here are some (other) examples of sheaves:
(1) Mor→X should be a sheaf (stated just above).
(2) For G the category of polydiscs in Cn , complex manifolds are sheaves. (The
coverings are open coverings. There is a subtlety here, which is that not all pul-
backs exist. Exercise: check that everything is okay.)
V3
(3) The “structure sheaf” Γ (O ), the sheaf of “3-forms” Γ ( Ω).
(4) The Grassmannian G (k , n ), viewed as the functor classifying subbundles of rank
k in a trivial bundle of rank n, is a sheaf.
(5) Mgno aut , the functor classifying families of genus g curves with no automor-
phisms. Exercise: think through why this is a sheaf. (What is the significance of
having no automorphisms?) possible homework problem
(6) The Hilbert functor, classifying families of subvarieties of an ambient variety
with fixed Hilbert polynomial, is a sheaf.
Now here are some non-examples of sheaves.
(1) The functor sending X to vector bundles of rank 2 on X . This is not a sheaf,
since it is not separated: if vector bundles are isomorphic locally, they are not
necessarily isomorphic globally. (In fact, it takes care even to make this into a
functor.)
(2) The “quotient presheaf” [An \ 0]/[Gm ]. Exercise: why isn’t this a sheaf? Can you
describe its sheafification “explicitly”? [Hint: it is something very familiar!]
(3) The quotient presheaf of sets [pt /Gm ] is again the sheaf [pt] (in particular, it is
already a sheaf ). Based on topological intuition, we might hope that it is “B Gm ”.
Exercise: think about how you could modify the situation to achieve this. [Hint:
this might motivate the need for groupoids.]
(4) Suppose we have a finite group G acting freely on X . Define X /G = ([X ]/[G ])sh .
This may fail to be representable.
(5) The Isom functor. Given two maps F1 → X and F2 → X in your favorite category
G , we can define a functor on GX sending

π: Z → X

to the set of isomorphisms

π∗ F1 ' π∗ F2 .

Exercise: Is this a sheaf? (Yes!)

Exercise 3.4.1. Show that Sh(G ) has fibered products (regardless of whether G does).
(possible homework problem; likely doing "separated presheaves" first)

3.5. Back to complex manifolds. Let’s return to our toy example, with G being the cat-
egory of open subsets of Cn , and the standard Grothendieck topology.
8 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

We can refine our earlier observation, realizing the category D of complex manifolds
as a full subcategory of sheaves on C .
G / Sh(G ) / PSh(G )
O

"
D
The complex manifolds are "cut out" in sheaves by some geometric condition.
Example 3.5.1. What’s a sheaf that’s not representable by a complex analytic “thing”?
Answer: differential forms

3.6. Geometric conditions on functors. We’ve described a hierarchy

G ,→ Sh(G ) ,→ PSh(G )
and we want to think about ways of describing a class of geometric objects extending
G.
Here was our toy example. If we take G to be the category of polydisks in Cn , with
holomorphic maps, we want to recover the category of complex manifolds. We decided
that complex manifolds should be in Shv(G ). But that isn’t enough to characterize the
complex manifolds. The last piece of the puzzle is the condition: a complex manifold
is a sheaf on G that is “locally representable".
Recall that a morphism F → G in PSh(G ) is representable if the base change against
a map hX → G from a representable functor hX is representable:
h? /F

 
hX /G

Definition 3.6.1. A subfunctor U ⊂ X is a representable map of functors which is point-

wise a monomorphism in G . An open subfunctor is a subfunctor which base changes
to an open embedding on representable objects.
Exercise 3.6.2. An open subfunctor of a representable functor is represented by an
open subthing. (This is kind of a tautology; or at least it has a one-line proof.)
I’m not sure this is useful — "opens" won’t be "subs" pretty soon.
Recall: If “P ” a property of maps in G that is stable under base change, then it makes
sense to talk about when a representable map of functors has property “P ”. Namely, we
declare a representable F → G to have property P if the base change with any hX → G
has property P .
h? /F

P
 
hX /G

Furthermore, if “P ” is also local on the source in G then one can formulate a notion
for a locally representable morphism of functors to have “P ”.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 9

Definition 3.6.3. A presheaf is locally representable if it admits a cover by open sub-

functors which are representable.
If you apply this construction to polydisks, you get complex manifolds. If you apply
it to the opposite category of commutative rings (i.e. the category of affine schemes),
you get schemes.
Exercise 3.6.4. Show that locally representable functors are sheaves.
3.7. Good stuff by Dan that Ravi thinks should be moved back. Let’s see if we can
work out concretely what some of these conditions on morphisms of (pre)sheaves look
like.
First of all, what is a map from an element of G to a sheaf S ? It is a map of functors
from Mor(·, X ) to S (·), so for each object Y over X , there is an element of S and these
are connected in the evident way. In fact, taking id : X → X , we get an element s of
S (X ) such that for any f : Y → X , the image of f is S (f )(s ), i.e. f ∗ (s ). In other words:
Proposition 3.7.1. For any object M of PSh(G ) and object X of G , there is a bijection
MorPSh(G ) (X , M ) ' M (X )
and this is natural in X and M .
Now, let’s unwind the definition of an open subfunctor a little: let i : M 0 → M be
a map. A map X → M is "the same as" an element s of M (X ), and M 0 ×M X is the
functor sending Y ∈ G to the set of
{(f , s 0 ) | f : Y → X , s 0 ∈ M 0 (Y ), i Y (s 0 ) = f ∗ (s )}
Now, if M 0 is representable, i.e. M 0 = M 0 ∈ G , and M 0 → M corresponds to t ∈ M (M 0 ),
this says that M 0 ×M X is the functor sending Y to the set
{(f , g ) | f : Y → X , g : Y → M 0 , g ∗ (t ) = f ∗ (s )}
If i is representable, there is some Z ∈ G with maps a : Z → X , b : Z → M 0 such that for
any f : Y → X , g : Y → M 0 such that g ∗ (t ) = f ∗ (s ), there is a unique map (f , g ) : Y → Z
such that f = a ◦ (f , g ), g = b ◦ (f , g ). In particular, taking Y = Z , (f , g ) = id, we see that
a ∗ (s ) = b ∗ (t ). If i is a representable open subfunctor, this says that a : Z → X is an open
embedding for any such X .
We can rewrite this as saying that:
Proposition 3.7.2. If M 0 is an object of G and M 0 → M is the morphism corresponding
to an element t ∈ M (M 0 ), M 0 is an open subfunctor of M iff for any X , s ∈ M (X ), there
is an open embedding i : U → X and a map b : U → M 0 such that i ∗ (s ) = b ∗ (t ). Fur-
thermore, for any f : Y → X , g : Y → M 0 , if g ∗ (t ) = f ∗ (s ), f factors uniquely through
i.
In the common (but not universal - think of the étale topology!) case where the open
embeddings of G are monomorphisms, this can be put in a slightly more transparent
version:
M 0 is an open subfunctor of M iff for any X , s ∈ M (X ), there is an open subset U ⊆ X
and a map b : U → M such that s |U = b ∗ (t ), and for any f : Y → X , the image of f is
contained in U iff f ∗ (s ) = g ∗ (t ) for some map g : Y → M 0 .
 10 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Finally, M is locally representable if there is some set M i consisting of objects of G

and elements t ∈ M (M i ) such that for any X , s ∈ M (X ), there is an open cover X = ∪i Ui
Lecture 3 such that s |Ui is uniquely the pullback of t by some map bi : Ui → M i .

4. TOWARD STACKS: THINGS THAT ARE NOT QUITE SHEAVES

4.1. A digression on 2-categories. When working stacks, one needs to replace cate-
gories with “2-categories”. This can be disorienting at first, so we give a discussion.
Before getting into this: we needn’t ever define 2-categories, and indeed we won’t. We
just need to know how to talk about maps between categories (these are functors, aka
1-morphisms), and maps between maps.
• A set is what we might call a 0-category. It has some objects, which we might call
“0-objects”. There are no relations (morphisms) between the objects, except
that we can recognize when two objects are the same. Thus,“0-same” is the
notion of equality.
• The category of sets is a 1-category, which is what you’re probably used to think-
ing of as a category. It has some objects, namely sets (these are the “0-objects”).
It has morphisms, which we might think of as 1-objects, between the objects.
The notion of “1-same” is now bijection. There are injections and surjections,
and fibered products.
• The category of categories is a 2-category. It has some 0-objects, namely the
categories. It has 1-objects, namely functors between categories. And now it
has “2-objects” as well, namely natural transformations between functors. In
particular, the notion of 2-sameness between 1-objects (i.e. functors) is a nat-
ural isomorphism, not equality.

4.2. An example (that we will use to motivate the definition). Here is a functor which
is not a sheaf, but is “like a sheaf”. This is the quintessential example of a thing which
we want eventually to be a “stack”.
Consider your favorite category G , and the site GX for some X ∈ G for your favoriate
Grothendieck topology.
Consider the functor assigning to U → X the “collection” of vector bundles on F (U ).
What do we mean by “collection”? One option would be to mean “isomorphism classes
of vector bundles”. Then this defines a presheaf, but it isn’t separated (all vector bun-
dles are locally isomorphic). In addition, we cannot glue isomorphism classes of vector
bundles. A second attempt might be to mean “vector bundles not up to isomorphism”.
Then we can “glue” given the additional data of isomorphisms over “intersections”, but
we don’t even really have a functor.
We are grasping for the notion of groupoid, but we don’t want to get into this yet.
Given a map f : X → Y , and a vector bundle on W on Y , we cannot specify “the
pullback” f ∗ W . However, we can define what it means for a vector bundle V on X to
be “a pullback”, which is that it satisfies the universal property (this specifies it up to
unique isomorphism).
So we can define a category structure on “vector bundles on G ”, in which
• the objects are vector bundles V /X for X ∈ G , and
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 11

• the morphisms

(V /X ) → (W /Y )
∼
“are” isomorphisms V − → π∗ W , where π∗ W is any choice of vector bundle on X
satisfying the universal property. (The set of morphisms can be defined up to
canonical bijection in this case, using the adjunction between π∗ and π∗ .)
Gluing. Now
S let’s discuss gluing objects in this category. Suppose we have an open
cover X = Ui and vector bundles Vi /Ui .
• For every pair Ui ,U j we get two vector bundles Vi |Ui ×X U j and V j |Ui ×X U j . We don’t
ask that these are the same, since the category of vector bundles on Ui ×X U j is
a 1-category, where the notion of sameness is an isomorphism. So we ask for a
choice of isomorphism

∼
τi j : Vi |Ui ×X U j −
→ V j |Ui ×X U j .

• Consider the triple overlaps Ui ×U j ×Uk . We have several vector bundles here,
and several isomorphisms. In the category of vector bundles on Ui ×U j ×Uk the
hom sets are just sets, so it makes sense to ask for equality, and so we can ask:

τi j ◦ τ j k = τi k .

• We also want to encode a notion that this gluing process produces a “unique”
vector bundle, in the appropriate sense. We capture this by saying that for any
V , V 0 /X , the functor U 7→ Isom(V |U , V 0 |U ) is a sheaf. Again, this is allowed
because the hom sets in the category of vector bundles on a space are just sets.
Okay, let’s now start making definition.

4.3. Prestacks. We now introduce a notion of prestack on G , which generalizes the no-
tion of a presheaf on G . This is also known as a category fibered in groupoids. This will
be a category M , equipped with a a covariant functor

π: M → G .

Suppose we have an arrow f : X → Y in G . A morphism V /X → W /Y is a pullback if

for any triangle in G

'/
X Y
f
 12 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

and map U /T → W /Y , there exists a unique morphim U /T → V /X lying over the

triangle:
U

∃!
V /( W


T

 (/ 
X Y
f

Example 4.3.1. The following are examples of categories fibered in groupoids:

• The category of vector bundles over objects in G , with morphisms defined as
discussed in the previous example.
• Mg , the category of families of relative curves of genus g over objects in G .
• Any presheaf F on G gives a category fibered in groupoids. The category con-
sists of pairs (X ∈ G , V ∈ F (X )) with morphisms being “pullback arrows”.
This suggests that if we want to discover stacks, then we cannot be content with
presheaves of sets - we need presheaves of categories. In other words, we embed G in
into the category of “2-presheaves” on G . The point of categories fibered in groupoids
(CFG) is to make sense of this notion.
Lecture 4, Jan 19 ’17
4.4. Recap at the start of lecture 4. We’ve been thinking about embedding our favorite
geometric category G inside the category PShv(G ), as the representable functors. If we
furthermore have a (Grothendieck) topology on G , then we can make sense of sheaves.
• If G is the category of affine schemes (with the Zariski topology), then the full
category of schemes can be recovered as the locally representable sheaves on G .
• If G is the category of open balls in complex affine space (with the topology of
open embeddings), then the full category of complex manifolds can be recov-
ered as the locally representable sheaves on G .
To get to stacks, we have to go outside this box. This was motivated by the examples
Mg and B G . What we need is a notion of “presheaves of categories”. (We only really
need to broaden our horizon from sets to groupoids.) However, these aren’t literally
presheaves, since we don’t have a “pullback functor”. We started discussing a specific
case: presheaves of groupoids, which we called “categories fibered in groupoids”.

4.5. Categories fibered in groupoids. Now G is the category of schemes. The functor
Mg assigns to any scheme S the families of relative curves X → S of genus g .
A category fibered in groupoids over G is a category M with a functor
π: M → G
satisfies some axioms, which we now discuss.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 13

A pullback of an object W /Y in M with respect to an arrow X → Y in G is an arrow

V /X → W /Y in M .
• (A1) Pullbacks exist, and satisfy the universal property: Suppose we have an
arrow f : X → Y in G . A morphism V /X → W /Y is a pullback if for any triangle
in G
T

'/
X Y
f

and map U /T → W /Y , there exists a unique morphim U /T → V /X lying over

the triangle:
U

∃!
V /( W


T

 (/ 
X Y
f

Exercise 4.5.1. Show that it is equivalent to only require the property for T = X ,
and the map T → X being the identity.1
• (A2) All morphisms of M are pullbacks.
Morphisms of categories fibered in groupoids will just be functors, commuting with
the projection maps:
M / M0

}
G
Note that the commutativity is demanding that two functors M → G are the same.

4.6. 2-Yoneda. We now want to formulate what it means for something in CFG(G )
to be “2-representable”. We have a functor G → PShv(G ). Presheaves of sets can be
thought of equivalently as categories fibered in sets, which embed in categories fibered
in groupoids.
Exercise 4.6.1. Show that the functor PShv(G ) → CFG(G ) is fully faithful? (Is this actu-
ally true? What does this even mean?)
1According to nlab, this weaker notion of pullback is equivalent to the one given only under the additional
assumption that the composition of "weak pullbacks" is a "weak pullback". In the cases we care about,
the axiom (A2) holds, so we don’t need to worry about this.
14 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

4.7. 2-fibered products (first attempt). Let X , Y , Z be categories fibered in groupoids

over G . Suppose we have maps
Y


X /Z

We want to produce a CFG “X ×Z Y ”. Of course, we will only be able to produce it

up to equivalence. We can try to describe it by a universal property. If W is another
(test) CFG, given morphisms W → X and W → Y plus a natural isomorphism between
W → X → Z and W → Y → Z

(
B: Y

 
X /Z

then W admits a factorization through a functor W → X ×Z Y which is unique up to

unique natural transformation.

$ )/
X ×Z Y 6> Y

  
X /Z

Let’s be clear about what’s going on here. The triangle

$ )/
X ×Z Y Y

commutes up to natural isomorphism, and similarly with the triangle

$
X ×Z Y

 
X
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 15

These natural isomorphisms are conditions, i.e. they are not given. However, the natu-
ral transformation
W

(
B: Y

 
X /Z
is given, and the natural transformation
X ×Z Y /Y
6>

 
X /Z
is also given, and we demand that the pullback of the second one to W agrees with the
first one.
Exercise 4.7.1. Is the above discussion correct?

5. LECTURE FIVE
Recall our set-up. For a geometric category G , there is a presheaf category PSh(G ),
also known as Fun(G ), consisting of the contravariant functors from G to Set. The
Yoneda embedding gives a full and faithful embedding from G to PSh(G ), and the im-
age of this map is called the category of "representable presheaves".
We then categorified this setup to obtain a notion of a "category-valued presheaf".
To make this notion precise without talking about pseudo-functors, we use the notion
of a prestack (warning: the terminology prestack here is not consistent with the us-
age in other places - the more common, but less intuitive, term is category fibered in
groupoids). This is an object which consists of a category M with a functor π from M
to G such that every morphism in M is a "pullback".
The category consisting of these objects is called PSt(G ), and it is a 2-category. Its
1-morphisms consist of functors F : M → M 0 such that π0 ◦ F = π (strictly), and its
2-morphisms are natural transformations between such functors F1 , F2 which project
to the identity on G . In other words, this data consists of, for each object m of M , a
morphism ηm : F1 (m ) → F2 (m ) of M 0 , such that π0 (ηm ) = idπ0 (m ) .
5.1. "2-Yoneda". We obtain a natural embedding from PSh(G ) to PSt(G ) by making
any presheaf a "category fibered in sets". In other words, if F is a presheaf, the objects
of the category F st which map to an object X of G are the elements of F (X ), and for
each arrow π : X → Y of G , there is exactly one arrow from F (π)(s ) to s for each s ∈
F (X ), and this arrow projects to π. In particular, we can do this for the representable
presheaves MorG (·, X ), which we identify (using Yoneda) with X .
What are the morphisms from X to M in PSt(G )? Remember that this is a category,
since PSt(G ) is a 2-category. The objects of this category, i.e. the 1-morphisms from X
to M correspond to the set of objects of M which project to X . We can refer to this as
16 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

M (X ), as it is the "category-valued sheaf" M evaluated at X . The 2-morphisms in this

category are exactly the morphisms of M which project to the identity morphism of X ,
which we can think of as the "automorphisms" of the objects over X .
∼
Proposition 5.1.1. We have an equivalence of categories MorPSt(G ) (X , M ) −
→ M (X )

Proof. Let F be a morphism from X to M . The objects of the (prestack associated to) X
which project to some Y in G are exactly the elements of MorG (Y , X ). Thus, the identity
morphism idX : X → X is a (canonically defined) object of (the prestack associated to)
X , and F maps this to some object of M which projects to X . Thus, we have a nice,
natural map from the object set of MorPSt(G ) (X , M ) to the object set of M (X ). Given a
2-morphism η between F : X → M and F 0 : X → M , ηidX is, by definition, a morphism
in M from F (idX ) to F 0 (idX ) which projects to idX . So, we obtain a well-defined, natural
functor from MorPSt(G ) (X , M ) to M (X ).
In order to see that this functor is an equivalence of categories, there is a "pseudo-
inverse" which sends an object s of M (X ) to the morphism from X to M which sends
π ∈ MorG (Y , X ) to "the" pullback π∗ (s ) ∈ M (Y ). This is non-uniquely defined, so this
only works "up to isomorphism". 
Exercise 5.1.2. Fill in the details of this proof, and clarify where exactly a non-canonical
choice is required.
Exercise 5.1.3. State and prove a precise sense in which this is natural in both X and
M.
Now, we want to show that the embedding PSh(G ) → PSt(G ) is, in an appropriate
sense "full and faithful", which can be thought of as the "2-Yoneda Lemma". This
means that:

Proposition 5.1.4. For any presheaves X , Y , we have an equivalence of categories

∼
MorPSh(G ) (X , Y ) −
→ MorPSt(G ) (X , Y )

The object on the left is a set, but we can always regard it as a discrete category by
making it the object set of a category with only identify morphisms. (A category is dis-
crete iff it is equivalent to one of this form).
Exercise 5.1.5. Prove the proposition.
Using this result, we can work in PSt(G ) without fear of "losing any information"
that existed in PSh(G ). This also allows us to make sure the notion of "representable
pre-stack" makes sense:
Definition 5.1.6. A representable pre-stack is a pre-stack M which is isomorphic to
the image of an object of G under the composition of the Yoneda maps.

6. FIBER PRODUCTS AND REPRESENTABLE MORPHISMS

Now, when we were working in the category of presheaves, we used the notion of a
representable morphism to extend geometric notions such as open and closed immer-
sions to the level of presheaves. So, we need to work out something analogous here.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 17

Example 6.0.1. Take G to be the category of schemes. Let M3 be the pre-stack classi-
fying curves of genus 3, M3,1 the pre-stack classifying curves of genus 3 with 1 marked
point. (i.e. these are fibered categories with the fiber over some X in G given by the cat-
egory of flat families of the given object over X with isomorphisms, and other arrows
given by pullback maps). There is certainly a forgetful functor from M3,1 to M3 given
by forgetting the marking, and this commutes with the maps to G .
Now, there is a scheme X (we can take X = P14 \ ∆ for ∆ the discriminant locus)
representing the locus of smooth plane quartics. This admits a map X → M3 , which
as we saw above is the same as an element of M3 (X ). This is the universal family of
smooth plane quartics over X .
Y / M3,1

 
X / M3
There is certainly a commuting square with a functor Y which classifies families of "a
point in P2 and smooth quartics through this point". With the correct notion of 2-fiber
product (see discussion below), we can see that Y is the 2-fiber product of X and M3,1
over M3 . Y is representable exactly by the universal family of smooth plane quartics
over X .
Exercise 6.0.2. Prove this. Possible homework problem.
There is still confusion about the appropriately 2-categorical definition of fiber prod-
ucts in PSt(G ). That is, what universal property should they have?
Let X , Y , Z be objects of PSt(G ) which form a diagram:
Y
G

X
F /Z
Then, whatever the fiber product X ×Z Y is, it should be an object of PSt(G ) such that
there is an equivalence of categories, natural in W between Mor(W , X ×Z Y ) and the
category whose objects consist of triples (A, B , η) of maps A : W → X , B : W → Y ,
and 2-isomorphisms η : F ◦ A → G ◦ B . The morphisms of this category are morphisms
(A, B , η) → (A 0 , B 0 , η0 ) consisting of 2-morphisms εA : A → A 0 , εB : B → B 0 which are
"compatible" in the sense that the following diagram commutes.
η
F ◦A / G ◦B

F ◦εA G ◦εB
 η0 
F ◦ A0 / G ◦ B0
Note that εA , εB must actually be 2-isomorphisms in our setting, since W , X , Y are not
just categories, but are objects of PSt(G ).2
2Section 3.4.9 of (whatever old edition I have of ) Olsson’s book works out the details of this. In particular,
the equivalence of categories between Mor(W , X ×Z Y ) and the "category of triples" is an isomorphism,
and also the "category of triples" is the "2-categorical fiber product of groupoids" Mor(W , X ) ×Mor(W ,Z )
18 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Exercise 6.0.3. The 2-category of prestacks admits 2-fiber products: in other words,
verify that the 2-fiber product as categories of two prestacks satisfies the axioms of a
prestack (i.e. the existence of pullbacks).
Remark 6.0.4. It should be possible to simplify this discussion somewhat: what we said
above can be stated in any 2-category and in particular in the 2-category of categories.
However, the 2-category we are considering has the pleasant feature that its objects and
morphisms are defined relative to the 1-category G . So, we should be able to use only
representable W in the previous definition, which simplifies things considerably since
we can understand maps from a representable prestack via Proposition 5.1.1.
It should be the case that a prestack M satisfies the universal property of the fiber
product X ×Z Y iff for each object X of G , there is an equivalence of categories, natural
in X (what exactly this means is unclear - perhaps it means that precomposing with a
map Y → X is "the same as" "the" pullback from M (X ) to M (Y )), from M (X ) to the
category of triples (A, B , η) where A is an object of X (X ), B is an object of Y (X ), and η
is an isomorphism from G (A) to F (B ). A morphism from (A, B , η) to (A 0 , B 0 , η0 ) is then a
pair of maps εA : A → A 0 , εB : B → B 0 such that F (εB ) ◦ η = η0 ◦ G (εA ).
Exercise 6.0.5. Try to prove this?
With this definition in hand, we can define a representable morphism.
Definition 6.0.6. Let F : M 0 → M be a morphism of prestacks. We say that F is repre-
sentable if for every object X of G and any morphism from (the representable prestack
associated to) X to M , the fiber product M 0 ×M X is representable.
As with presheaves, given any property of morphism which is preserved under pull-
back, we can extend this to a property of representable morphisms of prestacks.
Definition 6.0.7. A morphism M 0 → M is discrete if for any presheaf Y and any map
Y → N , the fiber product Y ×M M 0 is a presheaf.
Exercise 6.0.8. Given a prestack M and presheaves X , Y , the fiber product X ×M Y is
a presheaf. In other words, maps from presheaves are always discrete.
Example 6.0.9. The previous exercise is more intuitive in the special case of repre-
sentable presheaves. If X , Y are objects of G with morphisms to M , this is the same
thing as specifying objects of a of M (X ), b of M (Y ). So a morphism from an element
Z of G to X ×M Y is a pair of maps f : Z → X , g : Z → Y together with a choice of
isomorphism from f ∗ (a ) to g ∗ (b ). We can call this prestack IsomM (X , Y ).
Exercise 6.0.10. (a) Let X , Y be objects of G , with s : X → M , t : Y → M , and ∆ :
M → M × M the diagonal morphism. Then we can identify the fiber product of:
X ×X

∆

M / M ×M

with IsomM (X , Y ).
(b) The same exercise, but with X , Y presheaves rather than objects of G .

Mor(W , Y ).
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 19

Note: in general, we can go from statements about representable presheaves to state-

ments about presheaves by replacing our geometric category G with the geometric cat-
egory PSh(G ).
Next up, we will add topology to the discussion: just as the category of sheaves on G
sit between the category of representable presheaves and the category of presheaves,
the category of stacks will sit between the category of representable prestacks and the
category of prestacks. Lecture 6, Jan 26 ’17

7. LECTURE SIX
When we discuss anything in a categorical world, what do we need to define?
• We need to be able to recognize what a morphism is.
• We need a class of "things" for morphisms to go between, but these are really
just 0-morphisms.
• Beyond that, we need a way to discuss notions which resemble "surjectivity"
and "injectivity" (of course, we use other words for these).
• To take "operations" in our category, we need some version of "fibered prod-
ucts" (i.e. so we can take categorical limits - colimits for some reason are not as
vital).
• We need something resembling the Yoneda morphism, but which often will go
beyond just the Yoneda lemma itself - i.e. we need a way to understand every-
thing there is to know about objects in terms of some sort of functions on them.
Can we say more precisely what we mean regarding Yoneda? We have an embedding
G → PSh(G ). The Yoneda lemma tells us that this is a full and faithful embedding, but
we actually have more. For any object in PSh(G ), we can understand it entirely in terms
of maps from (the Yoneda image of ) objects of G into it; in this setting, the precise
statement is that Mor(X , F ) ' F (X ), and the bijection is natural with respect to both
morphisms between X and other objects of G and with respect to morphisms between
F and other objects of PSh(G ).
In our setting, we go beyond these fundamental categorical notions to discuss geo-
metric ideas. It turns out that the primitive notion here is the notion of a (Grothendieck)
topology on the geometric category G , which consists of the data of open morphisms
between objects of G (we can state this in terms of a subcategory containing all of the
objects and only some of the morphisms) and the data of specified collections of open
morphisms fi : Ui → X with common codomain X which are said to cover X . This data
suffices to define the notion of a sheaf on G , which is a special kind of presheaf on G
such that objects of F (X ) are precisely compatible systems of objects in F (Ui ).
The "miracle" property of the Yoneda morphism from G to PSh(G ) is that repre-
sentable presheaves are sheaves, which is a form of "descent". (This is a much deeper
and more miraculous fact when our topology is something like the étale, fppf, or smooth
topology on the category of schemes than the ordinary Zariski or analytic topology).
Aside: which fiber products need to exist in G ? We definitely need the fiber product
of an open morphism and any other morphism to exist in order to satisfactorily define
sheaves. We can possibly avoid using the others, but the category of sheaves on G will
always have fiber products in any case.
20 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Sitting inside the category of sheaves is the category of locally representable sheaves,
which we defined before. Another "miracle" is that if you replace the category G with
the category of locally representable sheaves on G (with a canonically induced topol-
ogy), the resulting categories of presheaves and sheaves are canonically identified!
Now, we have extended this hierarchy one level higher, to pre-stacks (recall that for
us, this means the same thing as categories fibered in groupoids). We have a Yoneda
embedding from the category of presheaves to the 2-category of prestacks, which is full
and faithful in the appropriate sense. Additionally, something we want here is that the
improved version of Yoneda we mentioned above should continue to hold: prestacks
are entirely determined by maps from representable prestacks into them. We (essen-
tially) proved these properties in the last class.
These are a higher analogue of presheaves - what is the analogue of sheaves? Intu-
itively, since a prestack is a "category-valued presheaf", it should be a presheaf that
"satisfies the gluing property", i.e. the fiber category over an object X should be en-
tirely determined by the fiber categories over a cover fi : Ui → X . More precisely, the
objects of this category should satisfy gluing, and the isomorphisms need to be a sheaf.
Let’s spell this out:
Definition 7.0.1. A prestack M on G is a stack if the following properties hold:
(a) For any objects X , Y of G , given any objects α ∈ M (X ), β ∈ M (Y ), the isomorphism
presheaf Isom(α, β ) should be a sheaf. This presheaf is the prestack fiber product
X ×M Y , with α : X → M , β : Y → M (recall that we can identify maps X → M
with M (X )).3 By the lemma below, X ×M Y is a presheaf.4
(b) "Objects glue". What does this mean? For a cover fi : Ui → X in G , let si ∈ M (Ui ) be
a family of objects over each Ui . If this is "compatible on the intersections", saying
that these objects glue means that there is a (unique up to unique isomorphism)
object s in M (X ) and pullback maps fei : si → s in M which lie over fi .
What does it mean to be compatible on the intersections? Well, for each Ui ,U j ,
there are pullback morphisms si j → si , s j i → s j lying over the morphisms Ui j → si ,
Ui j → s j of the fiber product Ui j : Ui ×X U j . Compatibility means that we have M -
isomorphisms ϕi j : si j → s j i lying over idUi j , which satisfy the cocycle condition that
(the pullbacks to Ui ×X U j ×X Uk of ) ϕ j k ◦ ϕi j = ϕi k (note that we really do mean
equality - M is a 1-category so we can only talk about whether morphisms are the
same or different).
Remark 7.0.2. Note that the compatibility condition on the "intersections" has real
content even when i = j (even just at the level of sheaves!). For example, consider a
nontrivial covering morphism U → X . (These exist in the geometric categories where
3The identification going from objects of M (X ) to morphisms X → M is not canonically defined, but
different choices give canonically isomorphic morphisms to M and thus canonically isomorphic Isom
sheaves.
4Some authors, including Olsson, use the term "prestack" to refer to categories fibered in groupoids such
that this axiom is true. However, our usage of "prestack" is more in line with the terminology for sheaves:
what we call a prestack is analogous to a presheaf and what others call a prestack is more analogous to the
notion of a separated presheaf (i.e. a presheaf where the identity axiom is true but perhaps not the gluing
axiom).
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 21

we allow non-monic maps to be open embeddings, and we will see shortly that we
should allow such behavior even in the Zariski topology). Then the "descent data" for
this cover consists of an object s of M (U ) and an isomorphism from p1∗ (s ) to p2∗ (s ) with
pi : U ×X U → U the canonical projections such that the cocycle condition is satisfied
by the various pullbacks to U ×X U ×X U .
Lemma 7.0.3. If G , F are presheaves on G with α : G → M , β : F → M , the prestack
fiber product G ×M F over a prestack M is a presheaf.
Proof. This means that we need to show that there are no nontrivial morphisms in
F ×M G which lie over an identity morphism in G . Recall that for H ∈ G , there is a
natural isomorphism of categories from Mor(H , F ×M G ) with (F ×M G )(H ), the fiber
category over H (i.e. the category of objects over H and morphisms over idH ). We need
to show that the former category is discrete (meaning it is equivalent to a category with
no non-identity morphisms).
Remember that we constructed an explicit candidate for the 2fiber2product (lolDan).
It is the category over G whose objects over some X in G are triples (a , b , ϕ) with a ∈
F (X ), b ∈ G (X ), ϕ : α(a ) → β (b ) an isomorphism in M lying over idX . It admits func-
tors to F,G given by mapping (a , b , ϕ) to a or to b . The morphisms of this category
from (a , b , ϕ) to (a 0 , b 0 , ϕ 0 ) which lie over some π : X → Y consist of the data of a pair
(f , g ) of morphisms of F,G respectively, f : a → a 0 , g : b → b 0 lying over π such that
β (g ) ◦ ϕ 0 = ϕ 0 ◦ α(f ).
Now, for any (a , b , ϕ), (a 0 , b 0 , ϕ 0 ) lying over some X , a morphism lying over the iden-
tity idX must be a pair (f , g ) of morphisms in F,G lying over idX . But since F,G are
presheaves, there are no nontrivial morphisms lying over the identity, so f , g must be
identity morphisms. 
We’ve seen that the category of prestacks over G has fiber products. In the world of
sheaves, it turns out that given two sheaves F , G , their presheaf fiber product is a sheaf,
and is the fiber product in the category of sheaves. The same thing happens with stacks
- this essentially boils down to proving that if M , M 0 , N are stacks, their prestack fiber
product is a stack.
Exercise 7.0.4. Do this. Possible homework problem. Most likely by doing this with
all the intermediate 2-categories.
Recall how we show that the category of schemes admits fiber products by showing
that affine schemes admit fiber products, and then everything glues by a completely
general functorial argument. Similarly, we can show that locally representable stacks
have fiber products if G does (but perhaps not in general). (We didn’t say this earlier,
but the same thing works for locally representable sheaves).
Exercise 7.0.5. Prove the above statements.

8. GEOMETRY
Example 8.0.1. We will define the stack BZ2 . As a fibered category, the fiber over an
object X is the category of locally trivial Z2 -covers of X , where "locally" means locally
with respect to the given topology. (a Z2 -cover of X is a covering space f : Y → X with a
22 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

group of X -automorphisms isomorphic to Z2 , and a locally trivial one is one such that
Y = ∪i Ui with f |Ui an isomorphism onto its image).
It’s not too hard to verify that this is a stack. Is it locally representable? This would
mean that there are G -objects M i with specified maps of stacks from M i to BZ2 classi-
fying Z2 -bundles Ei . Then, for any X and any given Z2 bundle ϕ : Y → X , there is an
open cover X = ∪i Ui such that the restriction ϕi of ϕ to Ui is classified uniquely by a
map fi from Ui to M i , i.e. a map fi such that fi ∗ (Ei ) ' ϕi .
In our case, we can take M to be a point, since we specified that the map is locally triv-
ial. This means that any X has an "open cover" f : X 0 → X such that f ∗ (ϕ) is classified
by a map to a point (since it is the pullback of the trivial bundle), and a map g : Z → X
factors uniquely through f iff the pullback g ∗ (ϕ) is trivial. Indeed, we can take X 0 = Y ,
f : X 0 → X as ϕ : Y → X , which is surjective and a local isomorphism. However, this
map is not an open cover in the Zariski topology! (the map isn’t injective). Thus, we
need to improve the Zariski topology to include maps which are (source) locally open
embeddings.

The property that matters here is that we can check openness locally: for a map f :
X → Y , if there is an open cover X = ∪i Ui and f |Ui , then f is an open map.

9. LECTURE SEVEN
Here are some words we don’t actually need to define:

• "2-category" - we don’t need to specify what a 2-category actually is, because

the only relevant 2-categories for us will be the categories of prestacks or stacks,
where the 2-categorical structure comes for free from the underlying categories.
• "Fibered category" - we don’t need the general structure of a fibered category,
but just the notion of prestack (which others call "categories fibered in groupoids").
• "Pre-topology" - for now, we only use the "covering" definition of Grothendieck
topologies and we don’t need to use sieves, so no confusion could arise between
these (almost identical in any case!) concepts.
• "Site" - we use "geometric category".

We can add one additional category to our chain of improvements of our geometric
category G : this is the category PSh+ (G ) of separated presheaves, which sits between
Sh(G ) and PSh(G ). These are the presheaves such that the identity axiom holds but
the gluing axiom does not have to. In this notation, we could refer to the category of
sheaves as Psh++ (G ).
Inspired by this, we will change the notation a bit to eliminate our non-compliance
with the "official" terminology: the issue is that others use the term "prestack" to refer
to categories fibered in groupoids such that the Isom presheaf is a sheaf. We will call
these objects St− (G ), denoting things that are like stacks, but which don’t necessarily
have to satisfy the gluing axiom. This sits inside St−− (G ), where the Isom presheaves
are required to be separated presheaves, and finally the category formally known (by
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 23

us) as "prestacks" is St−−− (G ) or CFG(G ).

G / Rep(G ) / LocRep(G ) / Sh(G ) / PSh+ (G ) / PSh(G )

#     
RepSt(G ) / LocRepSt(G ) / St(G ) / St− (G ) / St−− (G ) / St−−− (G )

Exercise 9.0.1. Prove that if M is a presheaf, then its Isom presheaves are always sep-
arated presheaves, but that the Isom presheaves are sheaves if and only if M is sepa-
rated. In other words, PSh(G ) is a "subset" of St−− (G ), and PSh+ (G ) is the "intersection"
of PSh(G ) with St− (G ).
Of course, it is immediate to see that a presheaf is a stack if and only if it is a sheaf,
locally representable as a presheaf if and only if it is as a stack, and representable as a
presheaf if and only if it is as a stack.
With all this setup in place, let’s talk a bit more about what we mean by the general-
ized Yoneda lemma in the setting of stalks. This is the following:
Lemma 9.0.2 (2-Yoneda). If X is an object of G and F is an object of PSt−−− (G ), the
∼
map Mor(X , F ) −→ F (X ) which takes a morphism Φ to Φ(idX ) (which is an object over
X of the category fibered in groupoids associated to X ) is an equivalence of categories.
Note that this implies that the map from PSh(G ) to PSt(G ) is fully faithful, because
both sides of this equivalence are discrete, so we get a bijection on the (unique isomor-
phism classes of ) maps.
Remember that last time, we had to replace the Zariski topology by the improved
Zariski topology (and we have to do something similar with the analytic topology). This
means that we declared
Proposition 9.0.3. As subcategories of the categories of presheaves on G , the category
of schemes (which is entirely insensitive to the topology), the category of sheaves with
respect to the Zariski and improved Zariski topologies are identical.
Exercise 9.0.4. Is this true for locally representable sheaves? (We saw that this was not
true for locally representable stacks)
Exercise 9.0.5. Is this true for stacks? (i.e. if something is a stack with respect to the old
Zariski topology, is it a stack with respect to the new one?)
Remember that we discussed the classifying stack of a group G to be the stack B G
which, as a functor on G , takes an object X to the category of G -locally trivial G -bundles
with isomorphisms.
Exercise 9.0.6. Show that if Y → X is a locally trivial G -bundle, then the following
diagram is a 2-fiber product:
Y / {pt}

 
X / BG
In general, we can use this sort of construction to take quotients by finite groups.
24 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Exercise 9.0.7. For X an object of G and G a finite group, define X /G as a stack and
verify that it satisfies reasonable quotient properties. B G is the case when X is a point.
This should be locally representable, at least in the improved topologies.
Here’s another example:
Example 9.0.8. We’ll work in the improved analytic topology with G complex polydiscs.
Let M1,1 be the stack of elliptic curves, i.e. genus 1 curves with one marked point. We
want to show that this is locally representable. This means we need to find some repre-
sentable open substack of M1,1 . It will turn out to be a bit easier to use M1,1 , the stack
of marked genus 1 curves which are either smooth or have nodal singularities.
Here’s a cool way to construct a local cover. Choose a "generic" pencil of cubics in
P2 with a chosen base point. Since this is a family of marked genus 1 curves over P1 ,
we obtain a map from P1 to M1,1 (a generic pencil has at worst nodal singularities).
This map has "degree 24", but the j -invariant map gives a degree 12 map to P1 - this
j -invariant map factors through a map from M1,1 to P1 has "degree 1/2".
Now, we can factor P1 → M1,1 through the P8 which parametrizes cubic curves in
P2 . Given some X → M1,1 which classifies a family of marked cubic curves over X , we
will show that the fiber product over M1,1 of X with P8 is representable. Then, we know
that the category of analytic spaces admits fiber products, so the base change to P1 of
this - which agrees with the base change of X with P1 over M1,1 - is representable.
Note that since locally representable analytic spaces are analytic spaces, so we can
do this locally on X .
How should we interpret things like the "good old-fashioned" j -line (i.e. the com-
plex projective line classifying j -invariants of elliptic curves)? This falls in the following
framework:
Definition 9.0.9. Given a locally representable stack M in LocRepSt(G ), we say that a
locally representable sheaf M is a course moduli space for M if:
• It is the "best approximation" to M in the sense that any map from M to a
locally representable sheaf X factors uniquely through M .
• It is a bijection on points, i.e. maps from the terminal object of G . (Right now,
we’re thinking of G as being the category of complex polydiscs and holomor-
phic maps or - equivalently - of complex-analytic spaces, so a "point" of X is
a map from the one-point space into X . We’ll have to be a little more careful
about what points we consider when we make this definition for schemes).5
Note: in the category of schemes, complex analytic spaces, etc., locally representable
means representable, so the course moduli space here is actually representable.
Remark 9.0.10. Why are we assuming things are locally representable? Well, in the
traditional cases, we want the course moduli space to be a scheme, manifold, analytic
space, etc., which means exactly that it is a locally representable sheaf. If we are hoping
to "approximate" a stack by something that is locally representable, we should start
5The first axiom is already sufficient to uniquely characterize X , so the second is a "bonus" axiom to make
sure that we have the properties we expect. It’s possible that something would satisfy the first but not the
second axiom.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 25

with a stack that is already locally representable. We could make the same definition,
dropping locally representability, for a "coarse" sheaf associated to a general stack.
However, at least in some settings of interest, there aren’t often good maps taking a
sheaf to a locally representable sheaf, or a stack to a locally representable stack. For ex-
ample, an algebraic space is (a certain kind of ) a locally representable étale sheaf on the
category of schemes. Since the étale topology has more covers than the Zariski topol-
ogy (it "is finer than" or "refines" the Zariski topology), any étale sheaf on the category
of schemes is in particular a Zariski sheaf in the category of schemes. Remember that
a locally representable Zariski sheaf is a scheme. Thus, if we had a nice functor taking
Zariski sheaves to locally representable Zariski sheaves, this would in particular give a
nice "associated scheme" map from algebraic spaces to schemes. It turns out that there
really is no such thing.

10. LECTURE EIGHT

We’ll fix a notational convention (for now) to fix the conflict we discussed last time
with the various notions of "pre-stacks". We will use the term pseudo-functor, with
notation PsFun(G ) to refer to the "category valued presheaves" or "categories fibered
in groupoids" or "pre-stacks" [in our previous usage]. Then we have three improve-
ments of this category: PsFun+ (G ), PsFun++ (G ), PsFun+++ (G ) = St(G ), analogous to
the improvements Sh(G ) ⊆ PSh+ (G ) ⊆ PSh(G ), where the middle term refers to sepa-
rated presheaves. These are presheaves which satisfy the identity axiom but not the glu-
ing axiom. For stacks, there are two intermediate notion: PsFun+ (G ) refers to pseudo-
functors plus the condition that the Isom presheaf for any two objects is separated.
In other words, isomorphisms that agree locally agree globally. Any presheaf satisfies
this axiom because in that case isomorphisms, if they exist, are unique. If the Isom
presheaves satisfy the gluing axiom, meaning that if two objects are isomorphic locally
by isomorphisms compatible on the intersections, they are isomorphic globally. The
presheaves that satisfy this property are exactly the separated presheaves, because in
that case it says exactly that objects that are locally equal are globally equal. Finally,
there is the category of stacks St(G ) = PsFun+++ (G ). Here, objects must glue. A presheaf
is a stack iff it is a sheaf, since elements of a presheaf glue iff the objects of the corre-
sponding stack glue (the cocycle condition is vacuous for presheaves since there is at
most one isomorphism between any two objects).
Note: A lot of the discussion for the rest of this section will not be entirely precise,
and we will fill in more of the details in the next section.
Let’s write down some names for our various topologies:
• The classical topology on complex analytic varieties (equivalently, on complex
polydiscs): covering families are jointly surjective holomorphic embeddings.
• The "improved" version of the complex analytic topology, which we will call
complex étale. The category is the category of complex analytic varieties, and
covering families are jointly surjective holomorphic functions which are source-
locally open embeddings.
• There’s an intermediate topology between these two: we allow disjoint unions
of open embeddings. These define exactly the same category of sheaves as the
26 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

classical topology, but in this case the sheaf condition for self-intersections is
nontrivial. In other words, if U = tUi → X is an open cover, U ×X U = ti , j Ui ∩
U j , so all the action happens here!
We would like to make sure that the complex étale topology defines the same categories
of sheaves as the classical complex topology. Since all covers in the intermediate com-
plex topology are covers in the complex étale topology, anything which is a complex
étale sheaf is an intermediate complex sheaf. We need to show the converse. In order
to do this, it suffices to check that any complex étale cover can be refined to an inter-
mediate complex cover. Then, gluing data for the étale cover restricts to gluing data on
the refinement. Since étale covers are necessarily locally open embeddings, if Y → X
is an étale cover, we can take an open cover Y = ∪i Yi and consider the composed map
ti Yi → X , which is an intermediate complex cover.
Definition 10.0.1. Another topology we can consider on the category of complex an-
alytic varieties is the smooth topology. The covers will be jointly surjective families of
submersions, which are maps that locally on the source look like U × D → U for D a
polydisc.
This also comes in an intermediate version, where we allow covers Y → X with Y =
ti Ui × Di such that Ui is an open cover of X , with maps Ui × Di → Ui on each factor.
As in the étale case, a smooth cover can be refined to an intermediate smooth cover,
which immediately shows that the categories of sheaves for the two versions of this
topology are the same.
More interestingly, does the complex étale topology have the same sheaves as the
complex smooth topology? Since the smooth topology has more open sets than the
étale topology, a sheaf in the smooth topology is necessarily an étale sheaf. Conversely,
it suffices to show that an étale sheaf is a sheaf in the intermediate smooth topology.
Consider an intermediate smooth cover tUi × Di → X . Then in particular, tUi → X
is an intermediate étale cover, and choosing any maps Ui → Ui × Di factors this later
cover through tUi × Di → X .
Exercise 10.0.2. (The Yoneda image of ) a complex variety is a sheaf in the complex
étale topology.
Note that we’ve shown that the categories of sheaves for the classical complex topol-
ogy, the intermediate complex topology, and the complex étale topology are all the
same, so it suffices to verify any one of these! Can you prove the exercise for the étale
case directly? What about the smooth case?
Exercise 10.0.3. Is the notion of locally representable sheaf the same in these different
topologies?
These are generally distinct for the case of stacks: for example, B(Z/2) is étale-locally
representable, since the cover by a point is étale, but not classically-locally representable.
In addition, BGLn is smooth-locally representable but not étale-locally representable:
the cover by a point is a smooth cover of positive relative dimension.
On the other hand, we have the following:
Exercise 10.0.4. The notion of stack (as well as the notions PsFun+ and PsFun++ ) are
the same in the classical, intermediate classical, and étale complex topologies.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 27

Theorem 10.0.5. In this setting, M1,1 is an orbifold (meaning an étale-locally repre-

sentable stack) in the complex étale topology.
The harder part of this theorem is saying that it is locally representable - the pseudo-
functor giving flat families of any class of varieties is a stack. This is essentially the
statement that for a complex analytic variety T , flat Ui -spaces X i → Ui with coherent
(i.e. satisfying the cocycle condition) Ui ×T U j -isomorphisms between the restrictions
of Ui and U j can be glued to a flat family X → T . (Since taking fibers is local, this family
has the right sort of fibers).6
Exercise 10.0.6. Prove the previous statement, and in particular, that M1,1 is a stack.
Aside: how do we define M1,1 in the complex analytic world? We need to make sure
that the notion of flat makes sense. It does, because complex analytic variety are locally
ringed spaces, so a morphism f : Y → X is flat iff for each point y of Y , OY ,y is a flat
module over OX , f (x ) . (Remember that Serre defined flatness in order to prove GAGA!).
This is all we need to define M1,1 , since we know what a genus 1 curve is in the analytic
setting. To extend this to the definition of M1,1 , we need to say what a stable curve is,
which in this setting is just genus 1 curve with nodal singularities. This is even easier
to define in the complex analytic setting, since we have an actual open subset which
is analytically isomorphic to an open neighborhood of the origin in the standard node
C[x , y ]/(x y = 0).
Now, we need to show that M1,1 is locally representable. We cover it with the stack
N which classifies marked genus 1 curves (C , p ) together with 3 chosen linearly inde-
pendent sections s1 , s2 , s3 in the three-dimensional vector space Γ (C , OC (3p )). This is
the same thing as specifying a surjective map from OC⊕3 to OC (3p ), which is in turn the
same thing as specifying an embedding i of C to P2 such that i ∗ OP2 (1) ' O (3p ).
Exercise 10.0.7. Find a good invariant way to specify what the "fiber" N (B ) is for a
complex variety B is.
Now, N is representable by P8 \∆, which is the moduli space of smooth cubic curves
in P2 which pass through the origin. (Here, ∆ is the discriminant locus). In other words,
for any complex analytic variety B , the maps Mor(B , P8 \ ∆) is naturally isomorphic to
N (B ), by pulling back the universal family. Next, we need to verify that N → M1,1 is
a smooth cover of stacks (in other words, it is representable and the pullback to any
complex analytic variety is a smooth surjection).
Can we get a nice étale cover of M1,1 from this? The space P8 \ ∆ is a dimension 8
smooth cover of M1,1 of positive relative dimension. To get an étale cover of M1,1 , we
claim that we can take a general "slice" of P8 \∆. More precisely, if we take a "sufficiently
general" Cartier divisor D ⊆ P8 \ ∆, D → M1,1 is still a smooth cover, and its relative
dimension is one less. Keep doing this until we get an étale cover. The statement that
6This argument subtly fails for schemes. The issue is that if we work in the étale topology, we can’t nec-
essarily glue flat families together and get a scheme. However, this is sort of the wrong thing to ask for,
since we can (almost tautologically) glue flat families together to get a flat family over an étale-locally rep-
resentable sheaf, a.k.a. an algebraic space. This does work for complex-analytic variety because, as we’ve
seen, the locally representable sheaves in the complex étale topology are the same as locally representable
sheaves in the classical complex topology, which are - basically by definition - complex analytic variety.
28 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

a sufficiently general divisor is still smooth over M1,1 is some sort of Bertini argument,
but it’s a little bit more subtle to show that the map is still surjective. Because we’re in
the complex analytic setting, it suffices to show that it’s still surjective on points (maps
from Spec C) and we can argue topologically using the fact that flat maps are open.
Why does this argument not show that the category of étale-locally representable
stacks and smooth-locally representable stacks are the same? This is not true: for ex-
ample, BGL n is smooth-locally representable but not étale-locally representable. The
issue is implicit in the "slicing" argument: the smooth cover {pt} → BGL n cannot be
"sliced". The intuition here is that {pt} is a dimension 0 space, but the smooth map
from this to BGL n is still smooth of relative dimension dim GLn , so the dimension of
BGLn is negative. Since étale maps are smooth of relative dimension 0 and any actual
complex analytic variety has strictly non-negative dimension, we cannot find an étale
cover of BGLn . What we’ve actually used is that we can only slice a positive-dimensional
space, so we’ve shown:

Proposition 10.0.8. A locally representable smooth stack of non-negative dimension is

étale-locally representable.

11. RELATIONSHIPS AMONG TOPOLOGIES

We’re basically done with the technicalities about general sheaves and stacks! We
will now discuss a few facts about topologies, and then move to the algebro-geometric
setting where we will consider stacks on the category of schemes.
To keep things concrete without worrying about some of the distracting technical
difficulties of working with schemes, here’s our main motivating example to keep in
mind:
Example 11.0.1. Our geometric category will consist of complex analytic varieties (per-
haps even allowing nilpotents, although some problems may arise in that case). For
each object X , we specify the additional data of a specified class of maps U → X which
are open sets, as well as which collections of these cover X . Here, U → X is considered
open if and only if it is the inclusion of an open subset of X in the classical/analytic
topology, and a family Ui → X is said to cover X if and only if it is jointly surjective.7 In
addition, we have two modified versions:
• In the complex étale topology, the open "subsets" are the maps which are source-
locally isomorphisms.
• In the complex smooth topology, the open "subsets" are the maps which are
source-locally of the form U × D − U , where D is a polydisc.
In both of these cases, when we say "locally", we mean locally with respect to the clas-
sical topology. Additionally, the covers are the jointly surjective families of open "sub-
sets".
How do we compare sheaves on the different complex topologies? In general:
7For all of the topologies we will actually consider, the covers are the jointly surjective families of open
"subsets", but we do not want to make this an axiom - one reason is that we don’t need to discuss "under-
lying sets".
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 29

Proposition 11.0.2. On a geometric category G , if one topology τ0 is finer (i.e. has more
open subsets) than the other topology τ, then any τ0 -sheaf (resp. τ0 -separated presheaf )
is a τ sheaf (resp. τ-separated presheaf ). The same statement is true for stacks as well
as the intermediate notions between stacks and pseudo-functors/categories fibered in
groupoids.
Proof. In order to verify the various sheaf/stack axioms on τ, we have to verify that
some statement holds for every τ-cover of an object X . But since every τ-cover is a
τ0 -cover, the sheaf/stack axioms for τ0 immediately imply that these hold. 
This implies that every complex smooth sheaf is a complex étale sheaf, and every
complex étale sheaf is a complex classical sheaf, and the same is true for stacks and
any of the intermediate notions.
One property that some, but not all, of our topologies have is that if {Ui → X }i ∈I
is a cover by some family, then we can form the disjoint union ti ∈I Ui → X . This is
still a cover in the complex étale or smooth topologies but not the classical topology.
However, we have the following:
Proposition 11.0.3. For any topology τ, if we extend τ to τ0 , the topology "generated by"
maps of the form ti ∈I Ui → X where {Ui → X } is a cover in τ, the resulting categories
of sheaves and separated presheaves and the corresponding stack notions are exactly the
same.
Proof. The sheaf/stack axioms with respect to the one-element cover ti ∈I Ui → X has
exactly the same content as the sheaf/stack axioms with respect to {Ui → X }. 
Using the above proposition, we will freely interchange between the complex classi-
cal topology and the modified version of the topology which also allows disjoint unions
of open sets.
Remark 11.0.4. On the other hand, notions such as "locally representable" are differ-
ent! This is what went wrong when we tried to discuss BG in the Zariski topology.
Thus, in order to verify sheaf/stack axioms, it suffices to consider covers consisting
of a single element. From now on, we will do this.
In addition, we have a criterion for showing that in some cases, passing to finer
topologies does not change the classes of separated presheaves or sheaves. This comes
down to the following lemma.
Proposition 11.0.5. Given a presheaf F on X , and a diagram in G which factors as:
V

~
U


X
Then the following statements are true:
30 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

(i) If F satisfies the identity axiom with respect to V → X , then it satisfies the identity
axiom with respect to U → X , no matter what sort of map V → U is.
(ii) If F satisfies the identity axiom with respect to V → U and V → X , and F satisfies
the gluing axiom with respect to V → X , then it satisfies the gluing axiom with
respect to U → X .

Proof. (i) Let s , s 0 be elements of F (X ) such that s |U = s 0 |U . Then by the fact that F
is a presheaf, s |V = (s |U )|V = (s 0 |U )|V = s 0 |V . Since F satisfies the identity axiom
with respect to V → X , this implies that s |V = s 0 |V .
(ii) Let s ∈ F (U ), and assume that p1∗ (s ) = p2∗ (s ) ∈ F (U ×X U ) with p1 , p2 the canonical
projections. Then consider t := ι ∗ (s ) with ι : V → U . If π1 , π2 : V ×X V → V are the
canonical projections, then we have ι ◦ πi = pi ◦ (ι × ι) for i = 1, 2. So, by the fact
that F is a presheaf,

π∗1 (t ) = π∗1 (ι ∗ (s )) = (ι × ι)∗ (p1∗ (s )) = (ι × ι)∗ (p2∗ (s )) = π∗2 (t )

By the gluing axiom for V → X , there is an element se of F (X ) such that se|V = t .

But by the presheaf property, se|V = ι ∗ (e
s |U ) = t . Thus, ι ∗ (e
s |U ) = ι ∗ (s ), so by the
identity axiom for V → U , s = se|U . Thus, F satisfies the gluing axiom for V → X .


Note that we do not require the maps to be covers in any topology: this is useful
because it allows us to verify the sheaf axioms with respect to certain particularly nice
covers, even when these do not form a topology.
In particular, this statement allows us to prove the following:

Proposition 11.0.6. The categories of sheaves in the complex classical and complex étale
topology are the same.

Proof. By definition, if U → X is a complex étale cover, then there is some open cover
U = ∪i Ui with Ui → X an open embedding in the classical topology. So, taking U to be a
complex étale cover and V to ti Ui with the Ui as above, we see that V → X and V → U
are classical covers (modified to allow disjoint unions), so they satisfy the identity and
gluing axioms. Thus, the above proposition applies. 

More generally:

Proposition 11.0.7. If τ, τ0 are two topologies on a category G with τ0 finer (more open
sets/covers) than τ, but any τ0 -cover can be τ-refined to a τ-cover, then the categories of
sheaves in the two topologies are the same.
Here, "τ-refined" means that a τ0 -cover U → X admits a τ-cover V → U such that
V → X is a τ-cover.

Another application lets us check the sheaf axioms in the complex smooth topology
only on particularly simple smooth covers:

Proposition 11.0.8. If a complex étale (or equivalently, complex classical) sheaf F sat-
isfies a sheaf axioms for covers of the form π1 : X × B → X , then it is a smooth sheaf.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 31

Proof. If U → X is a smooth cover, there is an étale cover V → U and an étale cover

X 0 → X such that V ' X 0 × B (this assumes the relative dimension is constant, but this
doesn’t matter since we can work on each connected component separately). Since F
is an étale sheaf, it satisfies the sheaf axioms for V → U and X 0 → X . By assumption, it
satisfies the sheaf axioms for V → X 0 , so it satisfies the sheaf axioms for V → X . Now,
we can apply the Proposition. 
Finally, we have:
Proposition 11.0.9. A presheaf F is separated with respect to the complex étale topology
or classical topology if and only if it is separated with respect to the complex smooth
topology.
Proof. A smooth cover U → X admits an étale cover U 0 → U such that U 0 = tUi × B
and Ui is an open cover of X . Let V = ti Ui ; this is an étale cover of X . Then by choosing
arbitrary sections Ui → Ui × B , V → X factors as V → U 0 → U → X , and we can apply
the Proposition to this V ,U . Since V → U is not a cover, we only get the identity axiom,
not the gluing axiom. 
Since the stack axioms regarding the Isom sheaf are stating that a particular presheaf
is a sheaf, we can apply Proposition 11.0.5 as above. In particular, we have:
Proposition 11.0.10. Fixing a category fibered in groupoids M over the category of com-
plex analytic variety, the Isom sheaves of M are separated with respect to the complex
smooth topology if and only if they are separated with respect to the complex étale and
classical topologies. They are sheaves in the complex étale topology if and only if they are
with respect to the classical topology.
Exercise 11.0.11. Show that the condition that objects of a given category fibered in
groupoids glue satisfies something similar to Proposition 11.0.5, and state what this
implies for the various topologies on the category of complex analytic varieties.
Proposition 11.0.12. The structure sheaf O : X 7→ Γ (X , O ) is a sheaf in the complex clas-
sical topology.
Proof. This is cheap, because O is represented by Ga (C) = A1C . 
As in the category of schemes, we can define various categories of sheaves of OX -
modules. (Defining O -modules can be done formally as "module objects" in the cate-
gory of sheaves on X ). This allows us to define the categories of quasicoherent sheaves,
finite type quasicoherent sheaves, finitely presented quasicoherent sheaves, and co-
herent sheaves. (These are all distinct in general, and it’s a difficult theorem that O is
coherent).
It turns out that more is true:
Proposition 11.0.13. The structure sheaf O is also a sheaf in the complex smooth topol-
ogy.
Note that separated presheaves in the classical and smooth topologies are the same,
so we know that O is a separated presheaf in the smooth topology. Now, we need to
32 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

show that it satisfies gluing. Consider a ball B in CN and the smooth projection map
π : X × B − X . We need to show that any section s ∈ Γ (X × B , O ) such that the two
pullbacks of s to (X × B ) ×X (X × B ) ' X × B × B agree comes from a section in Γ (X , OX ).
But the condition that the two pullbacks are the same means exactly that s is constant
on the fibers of π, so it gives a well-defined holomorphic function on X .
Now, this shows that O is a sheaf with respect to the smooth topology.
Corollary 11.0.14. The representable presheaf associated to any complex analytic vari-
ety is a sheaf in the smooth topology.
Because O is a sheaf in the smooth topology, and it is represented by A1C , the latter is a
sheaf in the smooth topology. Taking products, any affine space is a sheaf in the smooth
topology. Taking closed subsets defined by equations, any affine complex variety is a
sheaf in the smooth topology. Finally, since being a sheaf is a local property, this means
that any complex analytic space is a sheaf in the smooth topology.
We have a little more, which we will use to show that various categories fibered in
groupoids which are defined in terms of morphisms of analytic varieties satisfy at least
the stack axioms about the Isom sheaf.
Exercise 11.0.15. Given analytic varieties X , Y over a base B , the presheaf of mor-
phisms between X and Y is a sheaf in the smooth topology. Additionally, the same
is true for the presheaf of isomorphisms (this follows easily from the previous sentence
because we can apply it to the inverse of a given map).

12. LOCAL REPRESENTABILITY AND MODULI OF CURVES

For stacks to be "geometric objects", we need to verify more than the descent prop-
erty: they should also be locally representable, meaning that they admit an "open" cover
with respect to one of the topologies.
Why do we care about whether stacks are locally representable? Well, we can use
this notion to get well-defined interpretations of various geometric properties. For ex-
ample, we obtain a well-defined set of points, which are just maps from the terminal
object Spec C. We can put an actual topology on this set, because it is locally given
by an actual analytic variety, and we can represent the "gluing" by taking a topological
quotient. One amazing use of points - in the complex analytic setting - is that we can
check whether a map is a cover on points.
We can also define notions such as dimension, smoothness, and tangent space. If
a sheaf is étale-locally representable, it has non-negative dimension and the dimen-
sion is equal to the dimension of an étale cover by an analytic variety. However, if it
is only smoothly-locally representable, it may have negative dimension. For example,
consider B GL2 = [{pt}]/ GL2 ; this has dimension −4! Here, the tangent space has "neg-
ative" components - it is formally the difference of a 0-dimensional space and a posi-
tive dimensional space of "infinitesimal automorphisms". We can play a similar game
with M0,3 → M0 ; the former is a point, since a projective line with three marked points
has no nontrivial automorphisms. The cover has relative dimension 3, so M0 has di-
mension −3, which represents the fact that there is a 3-dimensional space of automor-
phisms of an unmarked projective line.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 33

In addition to local representability, another important property to consider stacks

as "geometric" objects is that they have a "well-behaved" diagonal: Recall that ear-
lier, we constructed complex analytic varieties as locally representable sheaves on the
category of affine complex analytic varieties with holomorphic maps, with the ordi-
nary analytic topology on them. Similarly, smooth manifolds can be defined as locally
representable sheaves on the category of open balls in Rn with smooth maps between
them, with respect to the Euclidean topology. However, this definition only produces
manifolds without the Hausdorff condition, which is usually included in the definition
of ’manifold’. How can we recognize this condition here?
It turns out that a manifold is Hausdorff if and only if the diagonal map M → M × M
is a closed embedding (remember that this can be defined in the sheaf-y sense by saying
that the pullback of M → M × M to any object of the base category - in this case, any
open ball in Rn - is a closed embedding).
The diagonal morphism M → M × M is always defined, and we can ask for various
properties of it. The most general is that it is a (locally) representable morphism. Given
that property, we can ask for further "geometric" properties of it - in analogy with the
corresponding case for schemes, we might ask for the diagonal to be quasi-compact
(in which case M is quasi-separated) or a closed immersion (in which case M is sepa-
rated).
Next, we want to apply these ideas to the moduli of curves.
Theorem 12.0.1. In the complex étale topology, Mg is a smooth-locally representable
stack for g > 1.
Proof. The fact that it is a category fibered in groupoids is easy, because relative genus g
curves pullback to relative genus g curves along any morphism. In addition, the Isom
presheaf is a sheaf: to see this, note that since it is a presheaf, by the above results it
suffices to show that it is a sheaf for the classical topology. In this case, this is easy,
since we can glue local isomorphisms.
We should be able to show that objects of Mg glue in the étale topology using the fact
that they glue in the classical topology, by refining an arbitrary étale cover to a (non-
injective) classical cover. See Exercise 11.0.11. We would like to be able to show that
Mg is a stack in the smooth topology as well. Note that since the Isom presheaf is an
étale sheaf, it is automatically a separated presheaf in the smooth topology. The fact
that it is a sheaf should follow from Exercise 11.0.15.

Exercise 12.0.2. Check that the Isom presheaf is actually a smooth sheaf. Do the ob-
jects of Mg glue in the smooth topology?

Since g > 1, the canonical sheaf ωC /k of any genus g curve is ample, and furthermore
ω⊗3
C /k
is very ample. By Riemann-Roch, it has 5g − 5 global sections, so its complete
linear system embeds C in P5g −6 . We will use this fact to cover Mg by the moduli space
of smooth genus g curves which are 3-canonically embedded in P5g −6 .
Inside the Hilbert scheme (which is an algebraic thing, but this is okay because an-
alytic subvarieties of projective space are algebraic, by Chow’s theorem) of genus g
curves embedded in of P5g −6 , there is an open subvariety X 0 which parametrizes smooth
34 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

connected curves. This has a universal family π0 : F0 → X 0 of smooth connected

curves. This family has a relative canonical bundle ω0 which restricts to the canoni-
cal bundle of each fiber. A fiber F is 3-canonically embedded iff ω−3 F ⊗ OF (−1) ' (ω0 ⊗
−3

OF0 (1))|F is trivial. Let L be the line bundle ω0 ⊗ OF0 (1). Then, there is a closed sub-
−3

variety X g ,3K of X 0 and a line bundle E on X g ,3k such that π∗ (E ) = L |π−1 (X 1 ) (this follows
from the first proposition in Chapter III of Mumford’s Abelian Varieties). In particular,
the restriction of L to any fiber of X g ,3K is trivial, so these fibers are 3-canonically em-
bedded. Moreover, X g ,3K is universal for this property, so Fg ,3K := π−1 (X 1 ) → X g ,3K is a
universal family of 3-canonically embedded smooth connected genus g curves.
We claim that X g ,3K → Mg is a PGL(5g − 5)-torsor. To see this, consider the fiber
product B ×Mg X g ,3K for any B → Mg with B an analytic variety, and B → Mg the map
classifying a family F → B . The maps to this fiber product are the same as pairs of
maps f : Y → B , g : Y → X g ,3K and isomorphisms between f ∗ F and g ∗ of the universal
family over X g ,3K . This is the same thing as a map f : Y → B and a family over Y of
3-canonical embeddings of the fibers of f ∗ F into P5g −6 .
On f ∗ F , there is a relative canonical line bundle f ∗ (ωF /B ) which restricts on the
fibers to the canonical line bundles. In order to get the 3-canonical embedding on the
fibers of F → B , we just need to specify 5g − 5 sections of f ∗ (ω⊗3 F /B
), i.e. elements of
Γ (f ∗ F, f ∗ (ω⊗3
F /B
)) = Γ (Y , f ∗ (π∗ (ω⊗3
F /B
))) which generate the restriction to every fiber.
Thus, the fiber product can be identified, by the universal property of projective
space, with the projective bundle P(π∗ ω⊗3 F /B
). Since this is smooth over B and B , F were
arbitrary, this shows that H → Mg is a smooth map. The statement that this smooth
map is a smooth cover says exactly that every smooth connected genus g curve can be
3-canonically embedded into projective space. Since g > 1, this works in the algebraic
category, and it also works in the analytic category since all compact Riemann surfaces
are algebraic. 
In order to deal with g = 0, 1 and also to find compactified versions of Mg for any g ,
we define some more precise versions of Mg :
Definition 12.0.3. • For n > 2g −2, Mg ,n is the pseudo-functor/category fibered
in groupoids of flat families of smooth connected genus g curves with n marked
points. (In a family, marked points means sections). For n ≤ 2g −2, this is empty!
• For n > 2g − 2, Mg ,n is the pseudo-functor/category fibered in groupoids of
flat families of connected genus g curves with n marked points which are ei-
ther smooth or nodal (i.e. semistable) and are Deligne-Mumford stable. The
latter condition means that every fiber that contains irreducible components
isomorphic to P1 has at least 3 "special" points, where a point is special if it is
the intersection of two irreducible components or if it is marked. For n ≤ 2g −2,
this is empty!
• For any n , Mg ,n is the pseudo-functor/category fibered in groupoids of any flat
families of connected genus g curves which are either smooth or nodal.
Proposition 12.0.4. All versions of Mg ,n are stacks in the complex étale topology.
Due to Exercise 11.0.15, the Isom presheaf is a sheaf, even in the smooth topology.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 35

But we do need to verify that objects glue. Because all of the genus, stability, etc. con-
ditions are fiber-wise, we just need to verify that the total spaces of the families glue as
analytic varieties. In the classical topology, this is straightforward. The statement for
the étale topology will follow from Exercise 11.0.11.
In fact, more is true:

Proposition 12.0.5. All versions of Mg ,n are stacks even in the complex smooth topology.

Proof. Due to Proposition 11.0.15, we just need to check that objects glue. So, it suffices
to show that if a family of curves (with the appropriate adjectives on the curves) on
X i × B → X i for X = ∪i X i a (classical) cover of X and for B a ball satisfies the descent
conditions, then it gives a well-defined family over X .
We can take sections X i → X i × B , and we get a pulled-back family of curves over
each X i . We need to check that the descent data is now satisfied for these families, so
that we can glue over the cover X = ∪i X i . Then, we need to check that the resulting
family over X actually pulls back to the given family over X i × B . But this last condition
is just saying that the family is "constant" over X i × B , which comes from the original
descent data. 

Now, it is time to show that the various versions of Mg are locally representable.
More precisely:

Proposition 12.0.6. Mg ,n and Mg ,n are locally representable in the complex étale topol-
ogy, and Mg ,n are locally representable in the complex smooth topology.
aut−free
In addition, we define Mgaut−free
,n and Mg ,n to be the subfunctors of Mg ,n , Mg ,n
respectively such that the curves in the fibers have no automorphisms. (The stability
conditions already implies they have only finitely many). These are locally representable
in the classical topology.

The proof for Mg , g ≥ 2 extends fairly directly to a proof for Mg ,n , n > 2g −2 (and the
statement for Mg ,n , n ≤ 2g − 2 is trivial). The role of the canonical sheaf K C is replaced
Pn
by K C + i =1 σi , with σi the marked points. Since n > 2g −2 = − deg K C , this always has
Pn
positive degree, so it is ample (as C is smooth). In addition 3(K C + i =1 σi ) is always
very ample as well (for genus 0, this is the statement that any invertible sheaf of positive
degree on P1 is very ample, and for genus 1, this is the statement that an invertible sheaf
of degree at least 3 is very ample).Then the rest of the proof carries through as above,
covering Mg ,n by the locus in the appropriate Hilbert scheme of smooth connected
Pn
genus g curves which are embedded by 3(K C + i =1 σi ). Since the marked points in
a family are sections over the base, this is also the restriction to the fiber of a globally
defined sheaf.
We can extend this to Mg ,n - allowing nodes - by removing the word "smooth" before
"genus g curve" everywhere and replacing it with "at worst nodal singularities". We
need to show:
• There is an open subset of the Hilbert scheme parametrizing the curves with at
worst nodal singularities.
36 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Pn
• There is some N such that N (ωC + i =1 σi ) is always very ample when C is a
stable curve, and ωC is the dualizing sheaf.
• The Deligne-Mumford stable locus (meaning that the locus where all P1 com-
ponents have at least three marked points) inside the Hilbert scheme is locally
closed, so that it is representable.
There are some technical details to worry about here, involving coherent duality. They
are addressed, for example, in the Deligne-Mumford paper The irreducibility of the
space of curves of given genus, Publ. Math. IHES 36, 75-109.
Finally, for Mg ,n , we can find a smooth cover of this by Mg ,n 0 for any n 0 > 2g − 2, so
a smooth atlas for the latter gives a smooth atlas for the former.
We still need to prove the stronger statements about the stable and automorphism-
free versions being respectively étale or classically locally representable.

13. LECTURE ELEVEN

We still "owe" a few results from the last class, regarding Mg . We’ll state these in
terms of schemes, but the arguments should go through for analytic varieties as well.
Theorem 13.0.1 (From Mumford’s Abelian Varieties, pp. 89). Suppose we have a flat,
finitely presented, and proper map π : Z → Y which is O -connected, meaning that OY →
π∗ OZ is an isomorphism. Then for any line bundle L on Z , there is a unique locally
closed subscheme Y 0 → Y such that:
• There exists a line bundle M 0 on Y 0 and an isomorphism π∗ (M 0 ) ' L 0 where
L 0 := L |Z Y 0 . In fact, M 0 is the pushforward (πY 0 )∗ L 0 , and the map π∗Y 0 M 0 →
L 0 is the canonical map π∗Y 0 (πY 0 ) ∗ L 0 → L 0 .
• A map f : W → Y factors through Y 0 iff there is a line bundle K on W and an
isomorphism π∗ (f ∗ K ) → L
If the fibers are (geometrically) irreducible, this is in fact a closed subscheme.
Proof. The statement is local on Y , so we can immediately reduce to the case that Y =
Spec A is affine. We will show Y 0 is a closed subscheme of an open neighborhood of
the set of points p in Y such that L |Z p is trivial, which implies that Y 0 is locally closed.
(If L |Z p is nontrivial, then p certainly cannot be in Y 0 !).
Now, we use Mumford’s cohomology and base change trick. We can compute the
pushforward of L universally with some complex 0 → A ⊕r0 → A ⊕r1 → · · · . Here, "uni-
versally", means that for any scheme X → Y = Spec A, the pullback of this complex to
X computes the pushforward of the base change πX : X ×Y Z → X . This complex is
produced using the Čech complex, which does compute the pushforward, but is horri-
bly infinite. In order to get around this, we use the fact that R p π∗ L is always coherent
(or finitely presented in the non-noetherian case) and then apply some homological
algebra to get the complex of finite free modules.
Aside: why does this trick work in the complex analytic category? The Čech complex
still computes sheaf cohomology, so the thing to be careful about is the finiteness result.
In fact, it is still true that the higher direct images of a coherent sheaf via a proper map
are coherent, but the proof is much harder than in the scheme case. The theorem is
due to Grauert - see the book Coherent Analytic Sheaves by Grauert and Remmert.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 37

Now, we can take the transpose of the first few terms of the complex to get an exact
sequence:
A ⊕r1 → A ⊕r0 → M → 0
By right-exactness of the tensor product, we can tensor this with B for any A-algebra
B:
B ⊕r1 → B ⊕r0 → M ⊗A B → 0
and this is still an exact sequence. Next, we can apply the left-exact contravariant func-
tor HomB (·, B ) to get an exact sequence:

0 → HomB (M ⊗A B , B ) → B ⊕r0 → B ⊕r1

Thus, the kernel of B ⊕r0 → B ⊕r1 can be canonically identified with HomB (M ⊗A B , B ) '
HomA (M , B ). By the universality of this complex, we then get that α∗ β ∗ L ' Hom
å A (M , B ).
Let m be the maximal ideal of A corresponding to the closed point p . By assumption,
L |Z p is trivial, so 1 = h 0 (π∗ (p ), L |Z p ) = dimA/m (HomA (M , A/m)) = dimA/m (M /mM ),
taking B = A/m above. By Nakayama’s Lemma, M m is generated over A m by a single
element. We can spread this out because M is finitely presented, so by taking a smaller
open neighborhood of p , we may assume that M ' A/I for some ideal I . We will take
Y 0 = V (I ). Then, taking B = A/I , we see that πB ∗ L ' HomA/I (A/I ⊗A A/I , A/I ) ' B is
a trivial line bundle on Spec B . This suffices, because πY 0 : Z Y 0 → Y 0 is O -connected,
which means that since (πY 0 )∗ (L 0 ) is a line bundle, the canonical map of line bundles
π∗Y 0 (πY 0 )∗ (L 0 ) → L 0 is surjective and therefore an isomorphism. This essentially fol-
lows from Nakayama’s lemma, because surjectivity can be checked on the fibers. This
means that we can assume that (πY 0 )∗ (L 0 ) ' OY 0 ' (πY 0 )∗ (OZ Y 0 ), and we need to check
that the canonical map plays nicely with these identifications. This is spelled out in
more detail in Ravi’s notes.
Now, how do we see that when the fibers are (geometrically) irreducible, we get an
actual closed subscheme? To see that this is necessary, consider a family of curves over
P1 which is generically irreducible and whose special fiber is reducible. Then, take the
line bundle associated to an irreducible component C of the special fiber, which is a
Cartier divisor. Away from this irreducible component, this line bundle is trivial (and
its canonical inclusion in OX is an isomorphism). However, it is not trivial on the special
fiber; we can check that its restriction to another irreducible component of the special
fiber meeting C has nonzero degree.
We will use the valuative criterion for a locally closed subscheme to be closed. In the
algebraic setting, this says that we just need to check that for any smooth Spec OC ,c . In
the analytic setting, this says that we need to check that a map from a punctured ana-
lytic disc extends to the disc. By the universal property we proved for Y 0 , this amounts
to showing that for any curve C in Y such that L on the fibers over C are generically
trivial, then L is also trivial on the special fiber.

Remark 13.0.2. Warning - the valuative criterion is not true in the complex analytic set-
ting. See http://mathoverflow.net/questions/194358/a-weak-analytic-version-of-the-valua
for a counterexample.
38 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

We have to additionally assume that the total space Z is normal. We have a rational
map s : O → L , i.e. a rational section s of L , which is an isomorphism away from
0. Then since the special fiber F0 is a Cartier divisor, s has an order of vanishing m on
F0 . Now, we can twist the line bundle on the base by OC (−m [0]), and apply Hartogs’
Lemma.
Here’s another way to see this: at least in the case that the fibers are curves, if L |Z p
is nontrivial and Z p is integral, then h 0 (Z p , LZ p ) = 0. Then, by upper semicontinuity
of fiber cohomology (which follows from the same Mumford trick), there is an open
neighborhood of p where this is the case. So, when all of the fibers are (geometrically?)
integral curves, we’re done in any case.
Note: does this extend beyond the case of curves? By flatness, it should at least be
the case that LZ p is numerically trivial. Can a nonzero effective divisor be numerically
trivial? 
Another missing point from our proof on the local representability of the moduli
stack of curves is showing that the locus in the Hilbert scheme of genus g curves with
at worst nodal singularities in P5g −6 is open. We already know that the locus of smooth
curves is open, so it suffices to show that the locus of curves with nodal singularities is
an open subscheme of the locus of singular curves.
So, let C → B be a flat family of (pure) relative dimension 1 such that every fiber
is singular. We need to give a more formal definition for what a node is. If m is the
maximal ideal of a node in a curve C , m/m2 has dimension 2, and m2 /m3 has dimension
2. These are the minimum dimensions for these among singular curves, so by upper
semicontinuity, there is an open subset of B where these are the dimensions. There
is a canonical map from Sym2 (m/m2 ) to m2 /m3 . The former is a 3-dimensional vector
space, and m2 /m3 is a 2-dimensional vector space, so there is a kernel of dimension 1.
We can think of the map from Sym2 (m/m2 ) to m2 /m3 as a quadratic form on the space
m/m2 , and we can define a node to be a curve such that this quadratic form does not
have any repeated roots. This is cut out by the complement of the vanishing locus of
the discriminant, so it is open.
We have a problem: consider k [x , y ]/(x y , x 10 ). We can check that this satisfies the
above properties, but we do not want to consider it to be a node. It turns out that re-
quiring the fibers to be reduced is sufficient.
Next, we will discuss the Picard functor. Let C be a curve and p a (rational) point.
Then for any base B , the Picard functor parametrizes the line bundles L on C ×B along
with a trivialization of L |p ×B

14. LECTURE T WELVE

Today we’d like to construct the Picard stack. We remain “suspiciously” in the ana-
lytic setting.
Consider the following two fibered categories.
Definition 14.1. Let C be a Riemann surface, and let p be a point on C . Let Pic(C , p )
be the category fibered in groupoids sending a base B to line bundles L on C × B with
an isomorphism L |p ×B ' Op ×B .
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 39

Definition 14.2. Let C be a Riemann surface. Let Pic(C ) denote the category fibered
in groupoids sending a base B to the set of isomorphism classes of line bundles L on
C × B.
Remark 14.0.1. We have maps between Pic(C , p ) and Pic(C ). The map in the forward
direction just forgets the isomorphism over p . For the map in the reverse direction, let
π : C × B → B be the projection. Then the map sends an line bundle L to L ⊗π∗ L |∨p ×B
together with the isomorphism to the trivial sheaf on p × B . fill in what this isomor-
phism is, exactly

14.1. Verifying the picard functors are stacks.

Aaron: This subsection
was quite poorly written
Definition 14.3. We say a functor F is determined by its values on points if the following by me, perhaps it would
holds. For any space X , let WX be the disjoint union of all points. Then, the map F (X ) → be helpful to rewrite it
more precisely with actual
F (WX ) is injective. symbols? This is well ex-
plained in Olsson’s book,
Lemma 14.1.1. Sections of Isom of the two picard functors defined above are determined chapter 4.2.9 - descent for
morphisms with a section
by their values on points.
Proof. Whether a map of line bundles is an isomorphism of line bundles can be checked
on each fiber. 
Lemma 14.1.2. We have that Isom for both the functors Pic(C ) and Pic(C , p ) is a sheaf
in both the complex étale and complex smooth topologies.
Proof. First, it’s easy to see isom is a sheaf in the complex étale topology by pulling back
to a cover in the analytic topology.
Next, let’s check Isom is a sheaf in the smooth topology. All separated presheaves in
the complex étale topology are separated presheaves in the complex smooth topology.
To check it’s a sheaf, the main issue is gluability. Let’s check it in the smooth topology.
This is similar to the above. For this, we can check on a refinement by balls, and then
from this construct an étale cover of the base which surjects onto each ball by coving
a ball U := B × X by its slices p × X for each p ∈ B . Call this union of slices V . This
is then an étale cover of X , and since we know it for étale covers we know it for these
covers by balls. There’s a small issue in that the map from the functor applied to U to
the functor applied to V may not be injective. But, it is injective on objects constant on
fibers. Then, one can use that Isom is determined by its values on points, 14.1.1. This
is enough to guarantee that Isom is a sheaf. 
Lemma 14.1.3. Both picard functors are stacks in the complex étale topology.
Proof. Suppose we have an étale cover with gluing data. We can refine it to a complex
analytic cover. We can then descend using the complex analytic topology, and recover
our original descent data in the complex étale cover by pulling back. To check we recov-
ered our original data, we can check after pulling back further to the complex analytic
cover. 
Lemma 14.1.4. Both picard functors are stacks in the smooth topology.
Proof. Suppose we have a cover with descent data. We can pass to a refinement by
balls. We can then pull back to a complex etale cover as above, and this will descend by
40 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

the previous lemma. We then pull this back to our cover by balls. We now get another
collection of descent data which agrees with our original collection. Let U → X be
our cover by balls. We have two elements s , t ∈ F (U ), where F is our desired picard
functor. We have anisomorphism between s and t coming from an isomorphism in
U × U restricting to q × U .
We want to check the descent data is isomorphic. this means checking a certain
commuting square on F (U ×U ). That is, we want to check we have a commuting square

pr∗2 s pr∗2 t
(14.1)

pr∗1 s pr∗1 t

We want to check this square commutes. To do this, we check on U ×U ×U , which we

omit.

Exercise 14.1.5. Complete the verification that the picard functor is a stack.


Lemma 14.1.6. The stack Pic(C , p ) in the smooth topology is actually a setoid (coming
from a category fibered in sets).
Proof. If we have a line bundle with a trivialization at p and an isomorphism hom(L , L )
we claim the isomorphism must be the identity. To see this, note that hom(L , L ) ' O .
Those isomorphisms preserving the section at p correspond to the sections of O . Since
the global sections of O are 1-dimensional. Preserving the isomorphism at p ensures
that it is the identity. 

14.2. Local Representability. Let’s show Pic(C , p ) is locally representable.

Lemma 14.2.1. The stack Pic(C , p ) = Picd (C , p ).
`

Proof. The degree of a line bundle is locally constant in families. 

Lemma 14.2.2. We have an isomorphism Pic0 (C , p ) ' Picd (C , p ).
Proof. Tensor with any line bundle of degree d . 
Proposition 14.4. For d sufficiently high, Picd is locally representable in the smooth
topology.
Proof. Take d more than 2g . Then, h 0 (C , L ) = d −g +1 and h 1 (C , L ) = 0. Consider H ,
the connected component of the hilbert scheme of curves embedded by the complete
linear system in Pd −g . Let U denote the universal family and consider the trivial family
C × H . Assume further that we restrict to the locus where U → H is smooth.
Construct the relative hilbert scheme U ×C mapping to H . Take the hilbert scheme
of subschemes of U × C which are isomorphisms onto both projections. Call this H 0 .
We then have H 0 and a family over it parameterizing embeddings of C → Pd −g . We will
then have that H 0 / PGL ' Pic(C , p ). This is the desired smooth cover. 
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 41

15. LECTURE THIRTEEN

15.1. Showing Pic(C , p ) is a complex analytic space, for C a curve. Recall that we’re
trying to show Pic(C , p ) and Pic(C ) are representable analytic spaces. We’d also like to
show there is a space of all curves.
We have a map Hilb0 → Picd (C , p ) for large d . This is a PGL torsor, and hence a
smooth cover.
Proposition 15.1. There exists an étale cover of Pic(C , p ), hence a complex cover.
Proof. Choose any invertible sheaf L on C . The complete linear system determines an
inclusion C → PN . The action of PGL on this curve determines the map PGL → Hilb0
which is a fiber product diagram

PGL / Hilb0 (15.1)

 
pt / Pic(C , p ).

By the slicing criterion for smoothness on the source, we can slice the source until we
get down to an étale map. This is possible because there are only finitely many auto-
morphisms (in fact, there are none for Pic(C , p )).
More precisely we use the following:
Exercise 15.1.1. Suppose π : X → Y is smooth of relative dimension n and p ∈ X . Sup-
pose f ∈ OX (X ) so that f |π−1 (π(p )) is nonzero in m/m2 in the fiber. That is, V (f )|π−1 (π(p ))
is smooth of dimension n − 1 near p in the fiber.
Then, there is a neighborhood U of p ∈ X so that V (f )|U is smooth of relative di-
mension n − 1. Hint: Smooth means that a map is locally like spec of some polynomial
ring modulo some equations satisfying the Jacobian criterion. If we add one equation
it drops the rank of the Jacobian by 1 at p , and hence in some neighborhood of p .
So long as PGL → Hilb0 is unramified there is some neighborhood of PGL which is
unramified. So, if we then replace PGL by a cover, we can then slice. Unramified means
the function pulls back to something nonzero in m/m2 .
For verifying unramified-ness, we need one fact about curves embedded in projec-
tive space. This uses the following:
Theorem 15.2. The maps
C , p , L → Pn

are precisely deformations of the identity in PGL. Here, the functor is more precisely


C , p , L × Spec C[ε] PN
(15.2)

Spec C[ε]

with an isomorphism π−1 (O (1)) ' L .

42 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Proof. Because we are fixing C , p , and L , the only thing that can vary in a family is the
choice of global sections of L defining the map to Pn . We can identify this with the
tangent space at the identity to PGL. 
This finishes the proof that Pic(C , p ) is étale locally representable. 
More generally, consider a smooth map of relative dimension k , X → Y with X a
space and Y a C-smooth stack. Then if we have a pullback diagram:
F /X

 
{pt} /Y

with F → X unramified. Then:

Proposition 15.1.2. There exists a section f ∈ OX (U ) such that V (f )|F is smooth of rel-
ative dimension k − 1 near p . Then V (f ) is a smooth cover of relative dimension k − 1 of
Y near p .
We may apply this to the Picard functor to show that the sheaf Pic(C , p ) is locally
representable in the complex étale topology. Since it is a sheaf (i.e. the fiber categories
are equivalent to sets), this also implies that it is locally representable in the complex
topology. Thus, Pic(C , p ) is a complex-analytic space. Indeed, since the Picard functor
naturally takes values in abelian groups, it is actually a "group variety", so it admits a
"multiplication" morphism m : Pic(C , p ) × Pic(C , p ) → Pic(C , p ) satisfying the usual
axioms.
Now, for d sufficiently large, there is a surjection Symd C − Picd (C , p ) mapping d
points p1 , . . . , pd to O (p1 + · · · + pd ). The fibers are all isomorphic to Pd −g . This implies
that Picd (C , p ) is actually a compact (smooth) complex manifold of dimension g , since
Symd (C ) is a compact (smooth) complex manifold of dimension d .
We can actually do this much more generally. Let X be a projective (complex) variety
which is (O )-connected (i.e. H 0 (X , OX ) ' C), with a point p . Just as before, we can
produce a category fibered in groupoids Pic(X , p ), and this is a stack in the complex
topology. It is even a sheaf.
Our goal is to produce a C-smooth cover, and then as before use this to deduce that
there exists a C-étale cover and then a classical cover. We will do this "near" some L ∈
Pic(X , p ). Let H be an ample line bundle on X . Then, replacing L with H ⊗n ⊗ L , we
may assume that it has no higher cohomology and is generated by its global sections,
so |L | gives a closed embedding from X to PN for some N .
Given a family of line bundles Lb over a base B such that L is the fiber over a point
b0 ∈ B , there is an open neighborhood U of b0 such that Lb is very ample iff b ∈ U .
Semicontinuity of cohomology shows that there is an open set on which the higher co-
homology vanishes and on which Lb is generated by its global sections. Then, we can
show similarly that the locus where Lb "separates points and tangent vectors" is also
open, i.e. the locus where the map defined by the complete linear system of Lb is a
closed embedding. To do this, the hypotheses already imply that Lb defines a mor-
phism (i.e. it is base-point free), and this gives a map π : X B → PN × B . Then, Lb
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 43

"separates points and tangent vectors" iff the coherent sheaf π∗ OX has rank 1 (and the
higher derived pushforward sheaves vanish).
Thus, there is an open subfunctor of Pic(X , p ) parametrizing families of line bundles
Lb such that Lb0 ' L for some b0 and such that Lb is very ample for all b .
Now, consider the scheme Mor(X , PN ), which can be constructed as an open neigh-
borhood of the Hilbert scheme of X × PN which parametrizes graphs of morphisms.
There is an open neighborhood in this which parametrizes non-degenerate maps, mean-
ing that the image is not contained in a hyperplane (the locus of degenerate maps is the
locus where ∧N +1 f ∗ (O (1)) is nonzero). This scheme covers Pic(X , p ) near L , by map-
ping f : X → PN to f ∗ (O (1)), and it is a smooth PGL-torsor.
How can we estimate the dimension of Pic(X , p )? We can show that the dimension
of the tangent space to Mor(X , PN ) at a map f is
h 0 (f ∗ TPn ) = dimC Hom(f ∗ Ω1PN , OX ) = dimC Hom(Ω1PN , f∗ OX ) = dimC DerC (OPN , f∗ OX )
by identifying the tangent space to the moduli scheme Mor(X , PN ) with the set of fami-
lies of maps from X ×C C[ε]/ε2 to PN which restrict to f on the special fiber, and finally
using a local computation to identify the set of these families with the set of derivations.
To compute h 0 (f ∗ TPn ), the Euler exact sequence gives:
0 → OPN → OPN (1)N +1 → TPN → 0
Then, we can pull this back to X to get:
0 → OX → L N +1 → f ∗ TPN → 0
Taking cohomology:
0 → H 0 (OX ) → H 0 (L )N +1 → H 0 (f ∗ TPN ) → H 1 (OX ) → H 1 (L )N +1 = 0
(remember that we assumed that L has no higher cohomology by tensoring with a
high enough power of a very ample line bundle).
Taking dimensions gives:
h 0 (f ∗ TPN ) = h 1 (OX ) + (N + 1)h 0 (L ) − 1
= h 1 (OX ) + (N + 1)(N + 1) − 1 = h 1 (OX ) + dim PGLN +1
Thus, since Mor(X , PN ) → Pic(X , p ) is a PGLN +1 -torsor, the dimension of Pic(X , p ) is at
most h 1 (OX ).

16. LECTURE FOURTEEN

We return to moduli of curves. Remember that we showed the stack Mg ,n of n -
marked genus g stable curves is a Deligne-Mumford stack. Recall that part of the defi-
nition of stability is that 2g − 2 + n > 0, so for smaller n (i.e. g = 1, n = 0 and g = 0, n =
0, 1, 2), the stack is empty. If we get rid of the stability condition, what can we say?
We have a stack Mg ,n which parametrizes families of connected "pre-stable" (syn-
onymously, semistable) curves of genus g with n marked points. This means that we
require the curves to have at worst nodal singularities, but we do not impose any fur-
ther stability condition (i.e. we may have P1 components without 3 special points).
44 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

The inclusion Mg ,n ⊆ Mg ,n is an open embedding, but for small n the former is just
the empty set!
We claim that:

Proposition 16.0.1. Mg ,n is locally representable in the complex smooth topology.

We will show that there is a representable open subfunctor around any C-point of
Mg ,n , and later make an argument that this implies we have an open covering (i.e. that
there are "enough" points). So, choose some curve (C , p1 , . . . , pn ) ∈ Mg ,n . Choose some
additional smooth points qi such that L := O (pi , q j ) is ample, for m points total. Take
a big tensor power K so that L is very ample, h i (L ) = 0 for i > 0, and h 0 (L ) = N + 1.
Then, L defines an embedding of C into PN such that p1 , . . . , pn are in general linear
position.
This gives a point of Hilb0 , the (analytified) Hilbert scheme of connected, (at worst)
nodal curves Z ⊆ PN which go through m points [1 : 0 : · · · : 0], [0 : 1 : · · · : 0], . . . with
O (1)|Z ' O (pi , qi )⊗K .
We may identify the quotient of this Hilbert "scheme" by the appropriate automor-
phism group (this group is just the stabilizer of the first n points in PGL) with the open
substack U ⊆ Mg ,m . This functor parametrizes m -marked semistable curves C , σ1 , . . . , σm
such that O (σ1 , . . . , σm ) is ample. Then, we have a smooth map Mg ,m → Mg ,n given
by "forgetting" σn+1 , . . . , σm , and the restriction of this map to the representable open
substack U contains our original curve C , p1 , . . . , pn in its image.
Why does it suffice, when checking local representability, to show that every C-point
can be put inside a representable open set? Formally, to say that Mg ,n = ∪i Ui , with
the Ui representable open subfunctors, we need to show that for any base B and map
B → Mg ,n , the pullbacks Bi := B ×Mg ,n Ui cover B . But, since B is a complex analytic
space, it suffices to show that for any b ∈ B , with b a C-point, there is some i such that
b ∈ Bi . But this is equivalent to saying that b → Mg ,n factors through Ui for some i ,
which is exactly what we showed.

Corollary 16.0.2. The stack parametrizing genus g curves with n marked points which
admit no automorphisms is a locally representable sheaf in the complex étale topology,
and therefore it is a complex analytic space.

Here’s something somewhat strange/surprising: Mg ,n is proper, but it is an open

substack of Mg ,n !
Now, what is the dimension of Mg ? For any (semistable) curve C , we have:

Proposition 16.0.3. dim(def(C )) − dim(inf.aut.(C )) = 3g − 3.

Another way we can see this is to consider the normal bundle, when C is smooth.
Taking some embedding C ,−→ PN , we have:
0 → TC → i ∗ TPN → NC /PN → 0
This gives, in cohomology:

0 → H 0 (TC ) → H 0 (i ∗ TPN ) → H 0 (NC /PN ) → H 1 (TC ) → H 1 (i ∗ TPN ) → H 1 (NC /PN ) → 0

 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 45

By the Euler exact sequence, we have:

0 → OC → OC (1)N +1 → i ∗ TP N → 0
so if we take the degree of the embedding large enough, as we may do, we can ensure
that h 1 (i ∗ TP N ) = 0. Then, we have:
0 → H 0 (TC ) → H 0 (i ∗ TPN ) → H 0 (NC /PN ) → H 1 (TC ) → 0
Deformation theory lends interpretations to all of these groups as "tangent spaces" to
appropriate moduli stacks. These tangent spaces to a moduli stack M are Zariski tan-
gent spaces, which may be identified with maps C[ε]/ε2 → M = M (C[ε]/ε2 ) which re-
strict to a given C-point mod ε. By directly evaluating the moduli functors on C[ε]/ε2 ,
we can see that when C is smooth:
Theorem 16.0.4. For a smooth curve C , we have the following interpretations of coho-
mology groups:
• H 0 (TC ) = HomOC (Ω1C , OC ) = DerC (OC , OC ) may be identified with the space of
"infinitesimal automorphisms" of C , i.e. automorphisms of C ×C C[ε]/ε2 which
restrict to the identity mod ε.
• Similarly, H 0 (i ∗ TPN ) may be identified with the space of infinitesimal deforma-
tions of the morphism i : C ,−→ PN , i.e. the space of morphisms C ×C C[ε]/ε2 →
PN restricting to i mod ε.
• H 0 (NC /PN ) = HomOC (i ∗ (IC /IC2 ), OC ) = HomOPN (IC /IC2 , i ∗ OC ) may be identi-
fied with the space of infinitesimal deformations of the closed subscheme [C ] ⊆
PN , i.e. the space of closed subschemes of PN ×C C[ε]/ε2 which are flat over
C[ε]/ε2 and restrict to [C ] mod ε. This is the Zariski tangent space to the (an-
alytification of the) Hilbert scheme of PN .
• H 1 (TC ), when viewed as the space of Čech 1-cocycles for the sheaf TC = Hom(Ω1C , OC )
mod coboundaries, may be identified with the space of infinitesimal deforma-
tions of C , i.e. the space of flat families F → C[ε]/ε2 such that the special fiber of
F is C .8
Remark 16.0.5. All of these statements generalize more or less immediately to arbitrary
closed immersions X → Y with X , Y smooth. If we relax the smoothness assumptions
to X → Y being an lci immersion, the statements remain true after replacing the se-
quence
0 → TX → i ∗ TY → NX /Y → 0
with
0 → i ∗ IY /IY2 → i ∗ (Ω1Y ) → Ω1X → 0

and taking the long exact sequence for the functor Hom(·, OC ). This has the effect of
replacing the cohomology groups with Ext groups. Note: I’m not sure how to see that
Ext1 (Ω1X , OX ) is the space of infinitesimal deformations of X . Is this true?
8This goes by noting that the underlying topological space of F is the same as that of C , so the problem is
about deforming sheaves of algebras on this space. Then, we show that such deformations are necessarily
trivial on affine subsets of C and identify the possible choices of gluing data on overlaps with C-derivations
from ΩC to ΩC .
46 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Returning to the proposition, we see that the dimension of the space of infinitesimal
deformations of C is h 1 (TC ) and the dimension of the space of infinitesimal automor-
phisms of C is h 0 (TC ), so we have, by Riemann-Roch:

dim(def(C )) − dim(inf.aut.(C )) = h 1 (TC ) − h 0 (TC ) = −χ(TC )

= −χ(OC ) − deg(TC )
= (g − 1) − (2 − 2g )
= 3g − 3

Thus, we get the desired dimension in the smooth case. Something similar works in the
semistable case but requires a little bit more work, being sure to use the appropriate
form of coherent duality in Riemann-Roch, etc.
Now, this implies:

Proposition 16.0.6. At any C-point C , the dimension of Mg is at most 3g − 3.

In particular, for g = 0, the dimension of M0 is at most −3! Indeed, the local dimen-
sion at a point of any smoothly locally representable stack is at most the dimension of
the tangent space minus the dimension of the infinitesimal automorphism group.
In order to get the other inequality, we may use the space of degree d branched cov-
erings from a genus g curve to P1 . Fixing a curve C of genus g , the set of degree d
branched covers from C to P1 can be identified with the set of degree d line bundles L
on C with two specified linearly independent local sections up to scaling, or more pre-
cisely, the set of degree d line bundles L with a specified one-dimensional subspace
of P(H 0 (C , L )). If we pick d ≥ 2g − 1, h 0 (L ) = d − g + 1, so the dimension of the space
of choices of one-dimensional subspaces of P(H 0 (C , L )) is 2(d − g + 1) − 1. Then, we
can use the fact that the dimension of Pic(C , p ) is at most g , so the dimension of Pic(C )
is at most g + 1, and we get that the dimension of the space of degree d covers of P1 is
at most dim Mg + g + 2(d − g + 1).
On the other hand, we can explicitly compute this dimension, using Riemann uni-
formization. Basically, once we fix the branch locus B ⊆ P1 , we know that the inverse
image of P1 \B locally looks like P1 , and we only need to specify the gluing information.
Working out these details correctly - being careful about what happens in the nodal case
- shows that the dimension of the space of degree d covers of P1 is exactly 6g − 2d − 5.
Plugging this into the previous inequality, we have:

6g − 2d − 5 ≤ dim Mg + g + 2(d − g + 1)

so we get dim Mg ≥ 3g − 3, and thus:

Theorem 16.0.7. At any C-point, the dimension of Mg is 3g − 3.

We can use this to understand the boundary of Mg ,n . This stack has dimension 3g −
3 + n . The boundary may be stratified by "topological type", i.e. the number of nodes
allowed. It turns out that the dimension of the boundary component parametrizing
genus g curves with n marked points with k nodes is 3g − 3 + n − k .
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 47

17. LECTURE FIFTEEN

We’re finally going to see how these notions work in the world of algebraic geometry.
To do this, we will need to define our geometric category and Grothendieck topology
so that we can get the appropriate notions of sheaves and stacks. Remember the setup:
we have a geometric category G and a topology on G which allows us to define

 / Rep(G )   / Sh(G )   / PSh(G )

G

The game will be to find a topology on G such that the topology is fine enough to allow
a lot of interesting locally representable sheaves and stacks, but not so fine that it is too
difficult to verify that something is a sheaf.

Proposition 17.0.1. Fix a geometric category G and a topology on it. If M is a locally

representable stack, then M → M × M is a locally representable morphism.

This means that if X ∈ G is an object and X → M × M is a morphism, the fiber

product X ×M ×M M → X is locally representable. How does this work? Let U → M
be an open cover with U ∈ G (maybe this is a family of morphisms, but it’s harmless to
think of it as just one object).
Then, we get a representable open cover U × U → M × M . Pulling this back to M
via the diagonal shows that U ×M U → M is a representable morphism.
and pulling back the diagonal gives the map U ×M U → U × U .
Now we have some generalities on descent. We have three levels of gluing:

• The identity axiom for the Isom sheaves: if isomorphisms between two objects
are equal on an open cover, they are equal. In other words, the notion of equality
glues. This is "−1-gluing"; we are just talking about equality.
• The gluing axiom for the Isom sheaves: isomorphisms on an open cover glue.
Here, the notion of isomorphism glues. This is "0-gluing"; we are talking about
sheaves, i.e. functors valued in sets.
• The gluing axiom for the objects: locally defined objects glue. This is "1-gluing".

Note that there are more compatibility conditions at each step: in the −1-gluing, there
are none; in the 0-gluing, there is just the notion of equality on intersections; in the
1-gluing, there is the notion of isomorphism on intersections together with the cocycle
condition on triple intersections.

Exercise 17.0.2. Can you make the "−1-gluing" statement more precise? If a sheaf is a
"1-functor", i.e. a functor valued in sets, and a stack is a "2-functor", i.e. a (pseudo-)
functor valued in categories, what is a 0-functor? It should be valued in {T, F }.

We stated this in slightly different language earlier:

48 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Lemma 17.0.3. If we have a diagram

V

~
U


V
such that V → X satisfies n -gluing and V → U satisfies n −1-gluing, then U → X satisfies
n-gluing.
There is a "simplicial" way to do this: if U → X is a cover, we can form the complex
X o π
U o π1 ,π2
U ×X U o U ×X U ×X U
π1 ,π2 ,π3

If we assume that U → X has a section, a lot of descent arguments become easier. One
formal way we can do this is to ramp up the sequence to:
XO o UO o U ×X U o U ×X U ×X U
π π1 ,π2 O π1 ,π2 ,π3
π

U o π1 ,π2
U ×X U o π1 ,π2 ,π3
U ×X U ×X U

Then, U ×X U → U has a section. Since the second complex is a "cover" of the first, it
often suffices to work with it.
Now, let’s do some algebraic geometry! The relevant geometric categories G are:
• Affine schemes.
• Schemes.
• Algebraic spaces.
Then, there are a number of useful topologies: Zariski, étale, smooth, fppf, etc.
Remark 17.0.4. Warning! "fppf" is a misnomer! The admissible maps are locally finite
presentation and flat, not necessarily actually finite presentation.
These are not totally distinct notions; they can be built from one another using the
tools we’ve seen. In particular, each is the category of locally representable sheaves
on the category above it with an appropriate topology. If we take G 0 = LocRep(G ) to
be the category of locally representable sheaves on G , the diagram of types of sheaves
becomes:
 / LocRep(G )   / Sh(G )   / PSh(G )
G _

=
 z
G0
 / LocRep(G 0 )   / Sh(G 0 )   / PSh(G 0 )

We can see that Sh(G ) ' Sh(G 0 ) and LocRep(G ) ' LocRep(G 0 ). This is because we can
cover any object in G 0 with objects in G , so a sheaf on G 0 is entirely determined by
its restriction to G , and a sheaf on G canonically extends to G 0 (we can think of G as
a "basis" in the sense of topology for G 0 ). For locally representable, we just use the
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 49

fact that if something has an open cover by objects of G 0 , we can cover each of these
objects by elements of G , so we get a G -cover. However, this statement is not true for
the categories of presheaves: think of the functor X 7→ H 1 (X , OX ). This vanishes on all
affine schemes but does not vanish on all schemes. The reason is that presheaves don’t
care about the topology, so local representability is useless.
In particular, taking G to be the category of affine schemes in the Zariski topology,
G 0 is the category of schemes. This discussion shows that sheaves on the category of
schemes are the same as sheaves in the category of affine schemes.
In order to make this make any sense, it is crucial that affine schemes are sheaves in
the Zariski topology. One of Grothendieck’s miracles is to prove that the same is true in
the flat topology:
Theorem 17.0.5. Affine schemes are sheaves in the flat topology.
First, we’ll show:
Theorem 17.0.6. Quasicoherent sheaves are sheaves in the flat topology.
Consider a faithfully flat map of rings A → B (so Spec B → Spec A is a faithfully flat
cover). A quasicoherent sheaf on Spec A is the same thing as an A-module M . We have
a complex:
0 → M → M ⊗A B → M ⊗A B ⊗A B
The 0-gluing condition says that this complex is exact.
Note: we can get the theorem for affine schemes immediately from the statement for
quasicoherent sheaves. From the module statement, we get that the following is exact:
0 → A → B → B ⊗A B
Descent for maps of affine schemes comes down to the statement that if f : R → B is a
map such that the two compositions R → B → B ⊗A B (i.e. the compositions via id ⊗1
and 1 ⊗ id) agree, then we can factor f through A → B . But this follows immediately
from the exactness of the above sequence.
Now, let’s prove the module statement. First, assume that there is a section B → A.
This allows us to create a null-homotopy from this complex to itself, which shows that
it is exact. However, we can actually reduce to this case: by faithful flatness, if
0 → B → B ⊗ B → (B ⊗ B ) ⊗ B
is exact, then the original sequence is exact. But here, B → B ⊗ B is a faithfully flat map
of rings which has a section, so we are done.
Note that the following implies that affine schemes are sheaves in the étale topology,
etc. because all of these topologies are refinements of the flat topology. So, we have:
 / Shfppf (Aff.Sch.) 

/ PSh(Aff.Sch.)
Aff.Sch.

Between affine schemes and fppf sheaves, there are various notions of locally repre-
sentable:
• A scheme is an fppf sheaf on the category of affine schemes which is locally
representable in the Zariski topology.
50 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

• An algebraic space is an fppf sheaf on the category of affine schemes which

is locally representable in the étale topology. By the above discussion about
sheaves on the category of representable sheaves, this is the same category as
the category of étale-locally representable sheaves on the category of schemes.
If we work out the notion of representable morphism in these categories, a representable
morphism in the category of schemes is an affine morphism. For algebraic spaces, if we
view them as sheaves on the category of affine schemes, the representable morphisms
are also the affine morphisms. However, if we view them as sheaves on the categories
of schemes, the representable morphisms are called schematic. (Note: any morphism
of schemes is schematic because fiber products of schemes are always schemes!)
Remark 17.0.7. The usual definition of algebraic space includes the criterion that the
diagonal map F → F × F is schematic. We’re omitting it for now and waiting to see
when it becomes relevant.
Note that every scheme is an algebraic space, but we will see that the converse is
false. We can also define the notion of "smooth-algebraic spaces", which are locally
representable for the smooth topology, and we could define analogous notions for any
topology between the Zariski and fppf topologies.
We can define:
Definition 17.0.8. On an algebraic space X (or an fppf sheaf which is locally repre-
sentable in one of the other topologies), a quasicoherent sheaf is a sheaf which is iso-
morphic to one of the form M Ýi on an affine cover Ui = Spec A i → X , with M i an A i -
module.
Note that this is exactly the same definition as in the case of schemes! This is be-
cause affine schemes also form perfectly good local models for algebraic spaces, but
the gluing information is more serious. This all works because quasicoherent sheaves
are automatically fppf sheaves, as we saw above. This means that M f is actually a qua-
sicoherent sheaf on Spec A in any topology inside fppf.
Why does nobody care about "smooth-algebraic" spaces? Well, they’re actually just
the same thing as algebraic spaces!
We need the following:
Definition 17.0.9. A representable morphism G → H of sheaves on the category of
affine schemes is surjective iff for each affine scheme Spec A and any morphism Spec A →
H , the fiber product G ×H Spec A → Spec A is a surjective morphism of affine schemes.
Exercise 17.0.10. Verify that for a representable morphism of algebraic spaces to be
surjective, it suffices to consider the case when A is a separably closed field.
Since we know what surjective means for schemes as well, we can make the same
definition in case the morphism is only representable by schemes rather than by affine
schemes. To go further, we can extend this to a definition for a map between any locally
representable sheaves in some topology. First, if Y is representable, and Y → F is a
morphism, we can pull back the representable cover U → F to Y to get a morphism of
schemes U ×F Y → Y . Then we require this to be surjective - in order for this definition
to make sense, we need to verify that it does not depend on the choice of cover U .
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 51

Finally, we say that F → G is surjective with both F,G locally representable iff U →
F → G is surjective with U → F a representable open cover. Then we need to verify
that this does not depend on the choice of U .

Proposition 17.0.11. A locally representable sheaf in the smooth topology is locally rep-
resentable in the étale topology.

Proof. Let U → F be a smooth cover, with U an affine scheme. Then U ×F U → U is

a smooth cover of U . If we take a point Spec k → U , we want to show that the fiber
product (U ×F U ) ×U Spec k → Spec k is étale. 

18. LECTURE SIXTEEN

Remember that given a geometric category G with a topology, we can define the cat-
egory of locally representable sheaves in that category. These are sheaves with an open
cover (with respect to the topology) by representable sheaves. If we form sheaves with
respect to the category of locally representable sheaves, we obtain the same categories
as we started with.
In our current case, G is either the category of affine schemes, or the category of all
schemes, with the étale topology. Since schemes are étale sheaves which are Zariski-
locally representable, they define the same notion of étale-locally representable sheaf:
these are called algebraic spaces. This means that we can use the category of affine
schemes, schemes, or algebraic spaces, all with the étale topology as our geometric
category and this does not change the notion of sheaf. (Although notions such as rep-
resentable morphism do change!)

Lemma 18.0.1. If f : X → Y is a smooth cover of a scheme Y (X is also a scheme), then

there is an étale cover U → Y which factors through X .

Proof. The local structure theorem for étale morphisms says that there are Zariski cov-
ers Y = ∪i Yi , f −1 Yi = ∪ j X i j such that f |X i j : X i j → Yi factors as an étale morphism to
AnYi followed by the projection AnYi → Yi . Then there is a section si j : Yi → AnYi and we
can take the base change of the étale map X i j → Yi with respect to this section to get
an étale map Ui j → Yi , with Ui j = X i j ×AnY Yi . Then we can take U = ti , j Ui j . Since
i
X → Y is surjective, t j X i j → AnYi → Yi is surjective. To make sure that t j Ui j → Yi is
surjective, note that we can take Yi to be affine, and thus quasicompact. For any point
y ∈ Yi , there is some j such that X i j → Yi hits y . Then, since the image of X i j is open in
AnYi , the intersection of this image with the fiber Ank (y ) is a non-empty open subvariety.
Thus, we just need to choose some si j whose base change to k (y ) meets this subvari-
ety. Assuming that Yi ' Spec A i is affine, a choice of section si j is the same thing as
choosing n elements of A i , so we should be able to choose these elements so that their
images in k (y ) land in an arbitrary open set.

Proposition 18.0.2. Any sheaf F in the étale topology is automatically a sheaf in the
smooth topology.
52 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Proof. From the lemma, if U → X is a smooth cover, there is an étale cover V → X such
that the following diagram commutes:
V

w
U


X
Let W = V ×X U : the map W → U is a base change of the étale cover V → X , so it is
also an étale cover; also, W → V is a base change of a smooth cover, so it is a smooth
cover. In addition, W → V has a section.
The sheaf exact sequences are related as follows (we do not yet know that they are
exact).
0 0

 
0 / F (X ) / F (U ) / F (U ×X U )

   
0 / F (V ) / F (W ) / F (W ×V W )

 
F (V ×X V ) F (W ×U W )

 
0 0
Exactness on the third row follows because we have a section V → W , and exactness of
the vertical sequences follows because V → X and W → U are is étale covers. Then we
conclude by chasing the diagram. 
Here’s an uncomfortable fact which we have been avoiding: usually, the definition
of an algebraic space X includes the condition that the diagonal map X → X × X is
representable by schemes. In other words, for any scheme T and map T → X × X , the
fiber product T ×X ×X X is a scheme. (It is étale-locally representable automatically, just
because it is a map of étale-locally representable sheaves, i.e. algebraic spaces).
Fortunately, this is actually a consequence:
Proposition 18.0.3. An algebraic space, defined as an étale-locally representable sheaf
on the category of (affine) schemes automatically has a representable diagonal.
Unfortunately, the proof uses Zariski’s main theorem and is somewhat involved; it is
proved in Olsson’s book.
We want to be able to extend all the ordinary properties of maps of schemes (smooth
of relative dimension n, flat, locally finite type, separated, etc.) to algebraic spaces. Any
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 53

decent property (i.e. a property which is étale-local on the base and preserved under
pullback) can be extended to the class of representable morphisms (here, representable
means representable by a scheme; a morphism is representable by an affine scheme
only if it is an affine morphism, which is far too restrictive).
How does the extension to arbitrary morphisms of algebraic spaces go? If the notion
is étale-local on the source as well as the target, this is not so hard. If f : X → Y is a
morphism of algebraic spaces, we can take an étale cover U → Y . Then we have the
fiber product X ×Y U → U , and we can ask for this to satisfy the property. So far, this
notion makes sense just because the property is target-local. Then, because it is source-
local, we can take an étale cover V → X ×Y U and ask for V → U to satisfy the property.
So this works for smooth, flat, etc.
Some more "topological" notions are easy too: since the notion of a "cover" has sur-
jectivity built in, it’s not so hard. Since it is target-local, we can assume the target is a
scheme. Then we have a map X → S with S a scheme and X an algebraic space. We
say that it is surjective iff the composition U → X → S is surjective for some étale cover
U → X . Similarly, quasi-compact morphisms can be defined, again using target-local-
ness to reduce to the case that X → S is a morphism with S a scheme.
What is a closed sub-algebraic-space of an algebraic space, without referring to points?
Let X be an algebraic space. A closed subset Y ⊆ X consists of the data of a closed sub-
set SY ⊆ S for every scheme S → X , compatible with pullback via a map T → S over
X . A closed immersion is defined to be a representable morphism Y → X such that the
pullback to any scheme is a closed immersion. We can define an equivalence relation
on these to get the notion of a closed sub-algebraic-space.
Actually, we can define "subset" similarly: a "subset" of an algebraic space X is the
data of subset of S for any S → X , compatible with the operation of inverse image. If
f : X → Y is a morphism of algebraic spaces and T → Y is a map from a scheme, we
can define the image of a subset of X on T as the image with respect to T ×X Y → T .
Something non-local but suitably functorial like separated is not hard to extend: we
require that the diagonal is a closed immersion. What about proper? Well, we have fi-
nite type (locally finite type + quasicompact) and separated, so we just need the notion
of a closed morphism, and this gives us universally closed automatically.
So what is a closed morphism? Well, we can literally say that the image of a closed
subset is closed.
Remark 18.0.4. We have to check that the various compatibilities and equivalences
between these notions which holds in the category of schemes also holds for algebraic
spaces. This can all be done, but sometimes requires some serious commutative alge-
bra to verify that things are étale-local.
Now, at least for comfort, we want to be able to define points. Already in the world of
varieties over an algebraically closed field, there are a lot of different notions: we have
rational points, points of the underlying set, and (various notions of ) geometric points.
These can all be different! However, they are all special cases of the notion of "maps
from Spec k for k an algebraically closed field". This has way too many points (i.e.
there’s a proper class of these), but worrying about set theory is no fun anyway. We can
apply various equivalence relations to get down to a more reasonable class of things if
54 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

we really want to make sure we get e.g. a sober topological space. Certainly anything
that can be checked by looking at all points in the ordinary sense can be checked by
looking at all points in this sense.
Definition 18.0.5. A geometric point of an algebraic space X is a morphism from Spec k
to X , with k algebraically closed.
One harmless equivalence relation we can impose which has some hope of making
this actually a set rather than a proper class is to say that Spec k → X and Spec L → X
define the same geometric point if there is a mutual extension K ⊇ k , L such that the
compositions of these maps with the maps Spec K → Spec k , Spec K → Spec L agree.
Remark 18.0.6. This is easier to work with than the notion of all field-valued points,
because taking étale covers induces separable field extensions on residue fields, so ar-
bitrary field-valued points are not sufficiently local in the étale topology. On the other
hand, nothing is lost if we work with just separably closed fields - these definitions are
essentially equivalent.

Proposition 18.0.7. Let X be a quasiseparated algebraic space. Then any point f :

Spec K → X factors through a monomorphism Spec k → X , with k a not necessarily
algebraically closed field.

What happens if we work over the complex numbers? Let X be a locally finite type
algebraic space over C. We can "analytify" this to get an analytic space over C, in the
sense we defined earlier of "locally representable sheaf in the complex étale topology".
We saw that this is equivalent to being a locally representable sheaf in the classical (an-
alytic) complex topology. In addition, the underlying set of this complex analytic space
will be the same as the set of C-maps from Spec C to X .
To build this thing, we first analytify affine schemes which are finite type over C.
These are all of the form Spec C[X 1 , . . . , X N ]/(F1 , . . . , FN ), i.e. an affine variety contained
in CN . We can apply the ordinary Euclidean topology and analytic structure here. Thus,
if G denotes the geometric category of complex analytic varieties with the complex
étale topology (i.e. the topology of analytic-local isomorphisms) and G 0 is the geomet-
ric category of affine finite type C-schemes with the étale topology, we obtain a functor
F from G 0 to G . Then, the analytification map will just be the pullback via F of a lo-
cally representable sheaf on G 0 . (Note that the pullback functor F −1 goes from Sh(G 0 )
to Sh(G ); the reason for this is that if G 0 , G were the categories of open sets of a topo-
logical space, a functor from G 0 to G is induced by a map of topological spaces in the
other direction).
Exercise 18.0.8. Define pullback rigorously, and verify that under this definition, the
pullback of a locally representable sheaf with respect to this functor F is locally repre-
sentable.
The relevant property of F is that it is continuous: if f : U → X is an étale cover in
G 0 , then F (f ) : F (U ) → F (X ) is an open cover in the complex étale topology on G . Fi-
nally, since sheaves and locally representable sheaves are the same in this topology as
the classical analytic topology, we see that F actually takes algebraic spaces to analytic-
locally representable sheaves, which are already analytic spaces. The analytification of
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 55

an algebraic space X should be the same thing as the disjoint union of the analytifica-
tions of a cover by affine schemes, mod the gluing relation defined by X .
Why do we care about algebraic spaces? The main "problem" with the category of
schemes is that it does not admit nearly enough "colimits". In other words, schemes
often are not closed by gluing along subschemes, quotients, etc. It turns out that alge-
braic spaces allow us to perform many more of these operations.
One important example of this is:

Proposition 18.0.9. If G is a finite group acting freely on any scheme X , we have a quo-
tient X /G as an algebraic space.

To prove this, we can first define X /G as an étale sheaf as the sheafification of the
presheaf given by (X /G )(T ) = X (T )/G (T ), then verify that X → X /G is an étale cover.
The key point is that the finiteness of G and freeness of the action
When X = Spec A is affine, we can define X /G as Spec A G , which is an affine scheme
and is the categorical quotient with decent geometric properties. We can extend this to
get a scheme quotient when X is quasiprojective, because this implies there is always
a cover by G -invariant affine open sets. (The quotient is not obviously quasiprojective
itself ). But when X is not quasiprojective, this generally does not work: Hironaka fa-
mously gave an example of a proper non-singular threefold with a Z2 action that does
not admit a quotient map to a scheme. On the other hand, this quotient is indeed an
algebraic space!
Here’s another example: consider E × P1 with E a genus 1 curve. There is a map
from P1 to the nodal cubic curve C which identifies two points p1 , p2 . To get a family of
genus 1 curves over C , we can identify the fiber over p1 with the fiber over p2 with an
infinite-order translation in E . We can show that this is not a scheme.
Note that this implies that the naive definition of M1,0 does not give a stack in the
étale topology! More precisely, this is the pseudo-functor on the category of schemes
which takes a scheme S to the category of flat families T → S such that T is a scheme
and the fibers are genus 1 curves.

19. LECTURE SEVENTEEN

19.1. Recap. We have now reached algebraic geometry land. We have built things up
and we have to figure out what tweaks we need to get the algebro-geometric results we
want.
Just to recall the definition: an algebraic space is a locally representable sheaf in the
étale topology.
Today we will focus on developing more properties algebraic spaces.
Remark 19.1.1. Recall we previously saw that the diagonal is locally representable. In
fact, this means that the diagonal is representable, though this takes more work.
One could also ask that the diagonal be quasi-compact (as was initially asked by
Knutson) meaning X is quasi-separated. This is not part of the definition, but we shall
sometimes assume it.
56 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Example 19.1.2. An useful (prototypical) example of a non-quasi-separated algebraic

space is evidenced by A1 /Z, by translation, which has non-quasi-compact diagonal,
since the base change of the diagonal along the cover A1 → A1 /Z is not quasi-compact.
This is also an example of an algebraic space which is not a scheme.

19.2. Étale equivalence relations: Definitions and examples.

Definition 19.1. Let U and R be schemes with two maps s , t : R → U . These two maps
(s ,t )
form an étale equivalence relation if R −−→ U × U is a monomorphism so that
(1) the projections s , t : R → U are both étale
(2) for any scheme T , the inclusion
R (T ) → U (T ) × U (T )
is an equivlence relation on U (T ), (meaning that (a , a ) lies in the image, if (a , b )
lies in the image then (b , a ) does, and if (a , b ) and (b , c ) lie in the image, then
so do (a , c ).

Remark 19.2.1. Here is the motivation for the above definition. Suppose we have an
algebraic space F . Consider a representable étale cover U → F by a scheme. Then,
take R := U ×F U so that we have a fiber square

R U
(19.1)

U F.
Exercise 19.2.2. Verify that the two projections R → U form an étale equivalence rela-
tion.
Note that by the magic diagram we have a fiber square
(s ,t )
U ×F U U ×U
(19.2)

F F × F.
So, we see that algebraic spaces yield étale equivalence relations. We would like to
go the other direction. That is, we would like to start with an étale equivalence relation
and produce an algebraic space.

Definition 19.2. Given an étale equivalence relation U /R define a presheaf

T 7→ U (T )/R (T )
by which we mean we quotient U (T ) by the equivalence relation defined by R (T ) →
U (T ) × U (T ).
Exercise 19.2.3. Verify this presheaf already satisfies the identity axoim.
Define U /R to be the sheafification of this presheaf.
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 57

Exercise 19.2.4. Make precise the following:

(U /R ) (T ).
0
Say we have a cover T → T with a map T → U .

T 0 ×T T 0 R

(19.3)
T0 U

T
We have two maps T ×T T . This gives a map T 0 ×T T 0 → U ×U , and we want to know
0 0

whether this lifts to R .

T 0 ×T T 0 R
(19.4)

U ×U .
Exercise 19.2.5. Suppose U and R are affine and flat and s , t : R → U is finite étale.
Suppose R → U ×U is an étale equivalence relation. Then, show that U /R is an affine
scheme and verifies the universal property of quotients for maps to schemes. Hint: Say
U = Spec A, R = Spec B , then construct U /R as the fibered product Spec A ×B A. Note:
This is not the tensor product, it is the fibered product of rings, whose elements are
(a 1 , a 2 ) mapping to the same element of B .
Example 19.2.6. Suppsoe F is an algebraic space and a representable chart U → F .
We have seen taking R := U ×F U , we have an étale equivalence relation
R U
(19.5)

U F
with all maps étale surjective. We obtain a map of sheaves U /R → F , coming from the
map on presheaves.
Exercise 19.2.7. Verify this map U /R → F is an isomorphism. Hint: To check this is
an isomorphism, we can check locally. Check it on the cover U → F , after which it is
clearly an isomorphism.

19.3. Using étale equivalence relations. Here are two goals we would like to embark
on:
Goal 19.3. Given an étale equivalence relation,
R U
(19.1)

U
 58 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

we want to show the quotient is an algebraic space, meaning U → U /R is a repre-

sentable étale cover.

Goal 19.4. In algebraic spaces, any map from a scheme X to F is representable. In

other words, the diagonal is representable.

The two above goals both follow from the following theorem (with the possible issue
that we don’t know that U → U /R is étale.

Theorem 19.5. Given an étale equivalence relations

R U
(19.2)

any map from a scheme Y to U /R is representable.

we will fill this in next Proof. 

week

This theorem will also imply (with quite some further work) the following useful fact.

Theorem 19.6. Suppose X is a quasi-separated algebraic space. Then there is a dense

open embedding by a scheme.

Proof. Start with

R U
(19.3)

U X.
For any U 0 ⊂ U , we have
R0 R
(19.4)

U 0 ×U 0 U ×U .
Aaron: I don’t think this The quotient U 0 /R 0 → X has a map which is an open sub algebraic space of X .
is correct. In particular, it
is not clear why the quo- Now, write U = ∪i Ui0 with Ui0 each open affines. It suffices to verify that for each Ui0
tient will be a subspace.
If U 0 is not the preimage
separately, we can find a dense open whose image in X is a scheme. So, we can restrict
of a subspace in the al- to the case that U 0 ⊂ U is a single affine open which is, in particular, dense.
gebraic space, we cannot
have any hope that this Since the diagonal is quasi-compact, we see R 0 → U 0 is quasi-compact. It is also
quotient will be an open étale. Therefore, by Zariski’s main theorem, we obtain a factorization
subspace. I believe one
needs to construct this U 0
more carefully. f
R0 Z0
g
possibly replace this by (19.5)
just spreading out h
U0
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 59

with Z 0 → U 0 finite. Then, take U 00 ⊂ U 0 defined to be U 0 \ h (Z 0 − f (R 0 )). Let R 00 :=

f −1 (U 00 ), or pictorially
U 00 R 00
(19.6)

U0 R 0.
Then, R 00 → U 00 is a finite morphism. It is also flat because the map began as an étale
map.
After this, it follows R 00 → U 00 is flat and we can further shrink U 00 to assume it is
affine, in which case R 00 is also affine, because the morphism is finite (hence affine).
Then, in this case the quotient R 00 /U 00 is in fact a scheme, by 19.2.5. 

19.4. Quasi-coherent sheaves. Let X be an algebraic space. Then, we have defined a

quasi-coherent sheaf on X as a quasi-coherent sheaf on Y for each scheme Y with a
map to X , compatible with pullbacks between schemes.
Exercise 19.4.1. Verify that if X is a scheme, then this notion of a quasi-coherent sheaf,
described above, agrees with the usual notion (using descent for quasi-coherent shaves).
If X is an algebraic space and π : X → Y is algebraic space, then we obtain a push-
forward map π∗ : Q C o hX → Q C o hY .
Remark 19.4.2. More over, any construction which behaves well with respect to flat
pullback applies.
For example, blowing up behaves well (if you have a blow up and base change along
a flat map is the blow up along the preimage) Stein factorization works, since Stein
factorizations can be constructed flat locally.

19.5. Comments on cohomology and preview of coming attractions. Cohomology of

OX modules makes sense as well. We would like to have derived functors. We need to
have enough injectives. It turns out there are enough injectives, but we’ll defer this to
next class.
For locally ringed spaces, we could constructed enough injectives by using stalks at
every point, and constructing injectives over each point, then taking the product over
all points. So, we need to figure out the right notion of points.
But, once we have checked there are enough injectives, one can construct cohomol-
ogy as R i Γ and derived pushforward R i π∗ . Further, one can verify the Leray spectral
sequence still works and Cech cohomology equals derived functor cohomology, using
the same proof.
We’ll also prove Chow’s lemma for quasi-separated algebraic spaces (which is why
we needed algebraic space to have a dense open scheme in the quasi-separated case.)
We’ll also see Grothendieck’s coherence theorem, showing that R π∗ F for F coher-
ent, is also coherent.
If X → Y is flat and Y is affine, say Y = Spec B , then using Mumford’s trick. Then,
over Spec B , we can produce a complex of finite free modules that universally com-
pute all pushforwards. Then, one can deduce Grauert’s theorem, cohomology and base
60 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

change, the fact that euler characteristic is constant in proper flat families. So, to de-
duce all of these things work, we’ll need to prove Grothendieck’s coherence theorem,
for which we will use Chow’s lemma, for which we will use that quasi-separated alge-
braic spaces have a dense open subscheme.

20. LECTURE EIGHTEEN

Now, we know what an algebraic space is, and we want to make sure that some of the
basic theory of schemes carries over. Recall that this takes place over the geometric cat-
egory G of schemes, and that algebraic spaces are étale-locally representable sheaves
on G with the étale topology. This definition always implies that the diagonal mor-
phism is locally representable.
There are a few variations of the definition in use:
• Some authors require algebraic spaces to have quasi-compact diagonal, so that
they are quasi-separated. Note that not all schemes are quasi-separated!
• Often, the condition that the diagonal morphism is representable (not just lo-
cally representable) is included.
Actually, it is true that the second condition is a consequence of our definition, us-
ing the fact that our definition easily implies that the diagonal is locally representable.
Perhaps this condition is not actually very important, but some useful properties do
require it. For example, we want:
Proposition 20.0.1. "Reasonable" (i.e. quasi-separated) algebraic spaces X to have a
dense Zariski-open subscheme.
Proof. We want to consider an algebraic space X as an étale equivalence relation:

R
s /U

t
 
U /X
Recall the definition: we say that this setup is an étale equivalence relation if:
• s , t are étale morphisms.
• s × t : R → U × U is a monomorphism.
• For any T , R (T ) ,−→ U (T ) × U (T ) is an equivalence relation.
∼
We can recover X just from U and R , because there is an isomorphism U /R − → X,
where U /R is constructed as a sheaf by the sheafification of the naive presheaf quotient
(U /R )pre (i.e. (U /R )pre (T ) := U (T )/R (T )).
For any open subset U 0 ⊆ U , we get an induced equivalence relation:
R0 / U0


U0
and this induces an open embedding (U 0 /R 0 ) ,−→ X (this is an open embedding ba-
sically because étale morphisms are open). Now, with this in mind, we can pick an
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 61

affine open subset U 0 of U and R 0 is a finite flat equivalence relation on U 0 . Then,

we can directly construct this quotient as a scheme (i.e. because rings and therefore
affine schemes have all limits and colimits) and verify that it is the algebraic space quo-
tient. 
Here is a useful application:
Proposition 20.0.2. Over a field k , every proper regular algebraic space curve is a pro-
jective scheme curve.
("curve" for algebraic space means a finite type, integral, separated, dimension 1
algebraic space over k ).
Proof. By the previous proposition, we have a dense open subscheme U , which is then
a smooth curve. This admits a quasi-finite map to P1 , and we use the curve-to-projective
extension theorem to extend this to C , which forces C to be a (projective) scheme.  Aaron: This proof seems
to be seriously flawed -
Recall that a quasi-coherent sheaf on an algebraic space can be defined as a compat- where is properness used?
In particular, there do ex-
ible system of A-modules M for each morphism Spec A → X . (The same thing works ist non-proper 1 dimen-
sional algebraic spaces
for schemes!). over fields which are not
schemes, see example
Exercise 20.0.3. Show that we can verify quasicoherence just by looking at étale maps 5.3.3 in Olsson’s stacks
Spec A → X . book.
A different proof can
Given a morphism π : X → Y of algebraic spaces, we can define a pullback functor be given using Chow’s
lemma - namely let X
π∗ on the category of quasi-coherent sheaves, and a pushforward functor π∗ on the cat- be our given curve and
X 0 → X be the map
O
egory of -modules. The pullback is easy: if Spec A → X is a morphism, then so is the let from chows lemma. Take
composition Spec A → X → Y , so the data of a quasicoherent sheaf on Y automatically a schematic étale 0cover
Y → X and let Y be
specifies the A-module corresponding to this morphism. Compatibility with pullback the pullback. Show that
Y → Y 0 is an isomor-
follows from compatibility on Y . For pushforward, given a morphism Spec A → Y , phism.
we get a base change Spec A ×Y X → X , and we can pull back the sheaf from X to But if the above proof is
intended, it should surely
Spec A ×Y X . Spec A ×Y X may not be affine, but we may cover it with affines. be placed after Chow’s
It is an interesting exercise to show the analogous properties/constructions for schemes,lemma.
defined as Zariski-locally representable sheaves on the category of affine schemes. If
U , V are two affine opens of a scheme S , U ∩ V = U ×S V may not be affine, so if we
want to verify compatibility on the intersections of something defined purely in terms
of affine schemes, we have to pass to an affine cover of U ×S V . The same arguments
work for algebraic spaces!
Proposition 20.0.4. If π is qcqs, then π∗ takes quasi-coherent sheaves to quasi-coherent
sheaves.
Proof. If X , Y are both affine, this is obvious. The problem is clearly local on Y , so
we can let Y = Spec B , and let Ui be a finite affine cover of X (using the fact that π
is quasicompact). Then for each i , j , let Ui j k be a finite affine open cover of Ui ∩ U j
(this uses the fact that π is quasi-separated, i.e. the diagonal morphism X → X ×Y X is
quasi-compact). Then we get an exact sequence by the sheaf axioms:
M M
F (Ui j k ) → F (Ui ) → F (X ) → 0
where we can take direct sums instead of direct products because they are finite.
62 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Now, let Spec B 0 → Y be an étale morphism into Y . If we tensor the above exact
sequence of B -modules with the flat B -module B 0 , we get an exact sequence of the
same form because tensor product commutes with finite direct sums. This shows us
that the pushforward is F
á (X ) (viewing F (X ) as a B -module), so it is quasi-coherent.

Next up, we want to discuss cohomology. A key fact here is that the abelian cate-
gory of OX -modules has enough injectives. We can make the ordinary Hartshorne-ish
proof work here, but we need to play with points to do this. Alternatively, we can use
Grothendieck’s very general criterion from the Tohoku paper (i.e. that there are coprod-
ucts and filtered colimits that play nicely with exact sequences, and that there is a set -
as opposed to proper class - of generators).
This lets us immediately extend the derived functor definitions of cohomology as
well as higher derived pushforwards and to obtain some of the basic properties.
More interestingly, we can extend the notion of Čech cohomology, at least in the qcqs
setting. Let π : X → Spec R be a quasi-compact and separated morphism and F be a
quasicoherent sheaf on X . We define Ȟ i (X , F ) as an R -module: because π is quasi-
compact, we may choose a finite open cover U = {Ui } of X . Then by separated-ness,
the intersection Ui j ,Ui j k , etc. are all affine. Then we can define the Čech complex:
Y Y Y
0→ F (Ui ) → F (Ui j ) → F (Ui j k )
and everything is an R -module. We can do this for an arbitrary quasi-compact sep-
arated morphism π : X → Y to get a Čech derived pushforward sheaf on Y , because
everything plays nicely with étale localization (because flat base change preserves the
above exact sequence due to finiteness).
We have:
Proposition 20.0.5. When π is quasi-compact and separated, then Čech cohomology
agrees with the derived functor cohomology.
If you use the right sort of proof (i.e. using the Čech-to-derived spectral sequence) in
the case of schemes, it extends directly. Note that the consideration of Čech cohomol-
ogy is not just for historical comfort - it is very hard to prove much of anything without
it!
Remark 20.0.6. If our sheaves are not quasi-coherent, then more problems can arise!
This is because there is generally no acyclic Zariski-open cover, so the intervention of
the étale topology changes things more. This is fortunate, because we want étale co-
homology of things like constant sheaves of abelian groups to look very different than
the Zariski version.
Our goal is to show that if π : X → Y is a proper morphism of (locally noetherian)
algebraic spaces, and F is a coherent sheaf on X , then R i π∗ F is coherent. To prove
this very important theorem, we want to be able to reduce to the case of a projective
morphism of schemes (note that even in the case of proper, non-projective schemes,
this is a valuable strategy).
To do this, we need:
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 63

Theorem 20.0.7 (Chow’s Lemma for Algebraic Spaces). Suppose that X is a reduced
separated finite type algebraic space over a (noetherian) base S . Then there is a quasipro-
jective scheme X 00 and a proper birational map X 00 → X .
Proof. Choose an affine Zariski-open scheme U ⊆ X . There is a locally closed embed-
ding e : U → Pn , and this lets us define a projective compactification U ⊆ Pn just by
taking the schematic image (i.e. taking the closure).
We have a diagram:
X0 / X × Pn /U


X
Here, X 0 is the closure of the graph of e in X × Pn .
X 0 is an algebraic space, and X 0 → X is proper and birational. Now, there is a scheme
V which is étale over X 0 , and so we get a map V → U . By the ordinary scheme Chow’s
lemma, there is a proper birational map V 0 → V such that the composition V 0 → V →
U is quasiprojective.
Now what? We can blow up U so that V 00 := V 0 ×U BlU → BlU is flat and quasipro-
jective.
To see why we can do this, first assume that V 0 → U is projective. This is flat over
a dense open subscheme U0 ⊆ U , so we have a closed immersion V00 = V 0 ×U U0 ,−→
PUn
0
= PSn ×S U0 , which is flat over U0 . Thus, we get a morphism from U0 to the Hilbert
scheme of S , or in other words we get a rational map U ¹¹Ë Hilb. The desired blowup
of U is the closure of the graph of this rational map; because this thing maps into the
Hilbert scheme, we can pull back the universal family to get V 00 , and this must be flat.
On the other hand, V 00 → BlU is quasi-finite and flat, so it is quasi-affine (this uses
Zariski’s Main Theorem). This shows that V 00 is quasi-projective. 

21. LECTURE NINETEEN

What are the remaining IOU’s? Some results that we have discussed, but have not
completely proved yet, include:
• Chow’s Lemma for algebraic spaces.
• Grothendieck’s Theorem on the coherence of the proper derived pushforward
of a coherent sheaf.
We sketched an incomplete proof of Chow’s Lemma last time. Using this, the archi-
tecture of the proof of the coherence statement is just like the proof of the correspond-
ing statement for proper (not necessarily projective) morphisms of schemes. Essen-
tially, when f : X 0 → Spec R is a proper morphism of algebraic spaces, we use Chow’s
lemma to find a projective fe : X 00 → Spec R which factors through f via a birational
morphism X 00 → X 0 . Then, we want to relate R p f∗ F to pushforwards of sheaves from
X 00 , and this is done via a standard noetherian induction/dévissage argument.
Similarly, this statement allows us to extend Mumford’s result on cohomology and
base change. Let f : X → S be a proper morphism and F an f -flat coherent sheaf on X .
64 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

Then, there is a complex of finite rank free sheaves on Spec R which universally com-
putes the higher derived pushforwards of F . Essentially, this complex is constructed
in a suitably canonical manner from the higher derived pushforwards of F , and the
only serious input necessary is the coherence theorem. This result then easily gives us
cohomology and base change, upper semicontinuity of the h i , etc.
Let’s return to the somewhat subtle point earlier about what exactly the definition of
algebraic space is. The "cheap" definition is that an algebraic space is an étale-locally
representable sheaf on the category of schemes with the étale topology. It turns out
that the condition that the diagonal is representable by a scheme is a consequence of
this definition, but it is not entirely obvious.
Next up, we want to define the appropriate notions of stacks on the category of schemes.
Definition 21.0.1. An algebraic stack is a smooth-locally representable stack on the
category of schemes with the smooth topology. A Deligne-Mumford stack is an étale-
locally representable algebraic stack.
There is the additional criterion that the diagonal ∆ : X → X × X is representable
by algebraic spaces. It is not too hard to see that this follows from the other proper-
ties. The stack axioms easily imply that ∆ is a relative sheaf, so it remains to see that
it is étale-locally representable. But this comes from the fact that X is smooth-locally
representable, so ∆ is at least smooth-locally representable. Then, for sheaves the two
conditions are the same so ∆ is a relative algebraic space.
Remark 21.0.2. There are various equivalent versions of this definition in different
sources: the key fact is that if a category fibered in groupoids admits a smooth cover
by a scheme, then it satisfies descent with respect to the smooth topology iff it satisfies
descent with respect to the étale (or fppf, etc.) topology.
Here is a useful and widely occurring definition for algebraic stacks:
Definition 21.0.3. If X is an algebraic stack, then the inertia stack of X is the fiber prod-
uct
I X := X ×X ×X X → X
i.e. the fiber product of the diagonal of X with itself.
The morphism I X → X is representable by algebraic spaces because it is the base
change of the diagonal morphism X → X × X , and this is representable by algebraic
spaces since X is an algebraic stack. In fact, I X → X is a "relative group algebraic space",
in the sense that for any scheme T and any object x ∈ X (T ), for the morphism T → X
corresponding to x , I X ×X T = T ×X ×X X = Isom(x , x ) = Aut(x ).
It turns out that this is actually even a relative scheme, which follows the (non-trivial)
fact that a group-algebraic-space is automatically a group scheme.
Just as in the case of algebraic spaces, we may define things like smooth, locally finite
type, locally finite presentation, surjective, etc. for representable morphisms.
Remark 21.0.4. When discussing algebraic stacks, the word "representable" without
modification typically refers to representable by algebraic spaces, and if a morphism is
representable by schemes, it is called "strongly representable".
If we view schemes as sheaves on the category of affine schemes, a morphism is rep-
resentable iff it is affine. Then, the analogous condition to requiring an algebraic space
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 65

or stack to have a representable diagonal is the condition that a scheme has an affine
diagonal. This property sits strictly between the property of being separated and qua-
siseparated. (Because closed immersions are affine and affine morphisms are quasi-
compact).
In general, for any scheme X , X → X × X is a locally closed immersion. Is something
similar true for algebraic spaces?
The properties of the diagonal for algebraic stacks can be a bit more complicated to
state, because the diagonal will generally not be a monomorphism. We define:
Definition 21.0.5. A morphism π : X → Y of algebraic stacks is separated iff ∆X /Y is
proper, and quasi-separated if ∆X /Y is qcqs.
Why do these definitions make sense? Well, the whole issue with stacks is that the
diagonal in general will not be a monomorphism, unlike in the case of schemes or alge-
braic spaces. (The presence of non-trivial automorphisms prevents monicity). Then,
a proper monomorphism is a closed immersion, so at least this restricts to the correct
definition when X , Y are schemes.
These notions make sense because the diagonal is a relative algebraic space, and
notions such as proper and qcqs are pullback-stable. So the condition that e.g. ∆X /Y :
X → X ×Y X is proper means that for any algebraic space T → X ×Y X , the base change
of ∆X /Y to T is a proper morphism of algebraic spaces, and we know how to define
these.
We will mention without proof a few useful criteria for algebraic stacks to be Deligne-
Mumford (resp. algebraic spaces), formalizing the idea that these notions are stacks
with "0-dimensional" (resp. no) automorphisms:
Proposition 21.0.6. An algebraic stack whose automorphism groups are trivial is an
algebraic space.
Theorem 21.0.7. An algebraic stack is Deligne-Mumford if ∆ is formally unramified,
meaning that Ω∆ = 0. This means informally that the stack has no ’infinitesimal auto-
morphisms’.
As we did for algebraic spaces, we may define quasicoherent sheaves in the category
of algebraic stacks:
Definition 21.0.8. If X is an algebraic stack, a quasicoherent sheaf on X is a compatible
(via pullback) system of A-modules for each Spec A → X
This allows us to make some constructions involving quasicoherent sheaves right
away, such as relative Spec and Proj, blowup, etc.
How do we extend the notion of sheaf cohomology to algebraic stacks? It turns out
that the categories of sheaves of abelian groups and of OX -modules always have enough
injectives, so we may define their derived functors as usual. On the other hand, perhaps
the category of quasi-coherent sheaves does not.
Just as in the case of algebraic spaces, we can extend the notion of Čech cohomology
for a quasi-compact morphism X → Spec A.
Here is a subtle issue, pointed out by Martin Olsson. The lisse-étale site sits in be-
tween the big and small étale sites of X . The big site consists of all morphisms Y → X ,
66 LECTURES BY RAVI VAKIL NOTES BY TONY FENG, DAN DORE, AND AARON LANDESMAN

and covers are étale maps Y 0 → Y over X ; the small site requires Y → X to be étale
itself; the lisse-étale site requires Y → X to be smooth. This is the correct setting for
various pullback functors to be defined - strange errors can occur if we are not careful
about which site we are considering at a given time.
Remark 21.0.9. Caution: for schemes, H i (X , F ) = 0 for i > dim X and any sheaf F .
This property extends, with a minor modification to the proof, to algebraic spaces. On
the other hand, it is false for general algebraic stacks.
For example, the cohomology of a sheaf on the stack B G = {pt}/G is the same thing
of group cohomology. (i.e. a sheaf of abelian groups on B G is the same thing as an
abelian group with a G -action). Certainly this can be nontrivial in positive degrees,
despite the fact that the dimension of B G is at most 0.
How do we define proper morphisms? This is certainly desirable, so we can make
sense of statements like "Mg ,n is the compactification of Mg ,n ".
Let X be an algebraic stack and Y a scheme. This is said to be closed if the image of
any closed substack is closed. (A closed substack is a morphism which is representable
by closed immersions of schemes). Using this, we can at least characterize universally
closed morphisms of stacks:
Definition 21.0.10. A (not necessarily representable) morphism of stacks X → Y is
universally closed if for any scheme Y → Y , the base change X ×Y Y → Y is closed in
the above sense.
Now, since we already understand separated and finite-type morphisms, we have:
Definition 21.0.11. π : X → Y is proper if π is separated, finite type, and universally
closed.
It is not hard to verify that with this definition, proper morphisms are local on the
target and closed under composition and base change.
A somewhat harder statement is the fact that:
Proposition 21.0.12. If X , Y , Z are algebraic stacks and Y → Z is separated, then
X → Y is proper iff the composition X → Y → Z is proper.
Another important notion is that of a course moduli space, which in some sense is
the "closest approximation" to an algebraic stack by an algebraic space:
Definition 21.0.13. If M is an algebraic stack, a course moduli space of M is a mor-
phism M → M with M an algebraic space such that:
• Any morphism from M to an algebraic space factors through M .
• M → M induces an isomorphism on geometric points.
Why should we care? The first reason that comes to mind is historical or psycho-
logical: in the old days before stacks, Mumford got a Fields medal for constructing the
course moduli space M g ,n ! It is also psychologically appealing to know that (in some
cases), a stack can be approximated by an algebraic space. But of course, we should
not care about a notion just because it makes us feel better or just to pay attention to
history.
But the actual reason why we care is that, in favorable circumstances, the course
moduli space allows us to prove things about the stack using tools that only work for
 MATH 245: TOPICS IN ALGEBRAIC GEOMETRY (STACKS) 67

algebraic spaces. This is helpful because we have a good criterion for course moduli
spaces to actually exist:
Theorem 21.0.14 (Keel-Mori). If X is of local finite presentation over a locally noether-
ian base scheme S with finite diagonal, then there exists a course moduli space X → X .
Furthermore, if X is separated or proper, then X is as well. Additionally, this con-
struction is preserved by flat base change (i.e. if Y → X is a flat morphism of algebraic
spaces, then X ×X Y → Y is a course moduli space).
Here is another useful feature:
Proposition 21.0.15. Pushforward along course moduli morphisms is an exact functor
on the categories of quasi-coherent sheaves.
This means that for the purposes of sheaf cohomology, it suffices to work in the cat-
egory of algebraic spaces.
Using course moduli spaces, we can show:
Theorem 21.0.16 (Chow’s Lemma for Deligne-Mumford stacks). Let X be a Deligne-
Mumford stack with a proper morphism X → S . Then, there is a proper surjective mor-
phism X → X such that X → S is projective.
As in the case of algebraic spaces or proper non-projective morphisms of schemes,
we may use this lemma to prove things like:
• The higher direct images of coherent sheaves under proper morphisms are co-
herent.
• There is a valuative criterion for properness.
A final note is that all of our discussion in the world of analytic stacks carries over
without much change to the algebraic setting. Thus, we can show things like the state-
ment that Mg ,n is a Deligne-Mumford stack using essentially the same proof. On the
other hand, occasionally there are complications due to the fact that étale morphisms
of schemes are not Zariski-local isomorphisms, so more care needs to be taken for cer-
tain arguments.