Complex Analysis on Riemann Surfaces

Math 213b Harvard University

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Maps between Riemann surfaces . . . . . . . . . . . . . . . . 6
3 Sheaves and analytic continuation . . . . . . . . . . . . . . . 13
4 Algebraic functions . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Holomorphic and harmonic forms . . . . . . . . . . . . . . . . 26
6 Cohomology of sheaves . . . . . . . . . . . . . . . . . . . . . . 33
7 Cohomology on a Riemann surface . . . . . . . . . . . . . . . 42
8 Riemann-Roch . . . . . . . . . . . . . . . . . . . . . . . . . . 46
9 The MittagLeer problems . . . . . . . . . . . . . . . . . . 51
10 Serre duality . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
11 Maps to projective space . . . . . . . . . . . . . . . . . . . . . 59
12 Line bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
13 Curves and their Jacobians . . . . . . . . . . . . . . . . . . . 80
14 Hyperbolic geometry . . . . . . . . . . . . . . . . . . . . . . . 97
15 Uniformization . . . . . . . . . . . . . . . . . . . . . . . . . . 107
1 Introduction
Scope of relations to other elds.
1. Topology: genus, manifolds. Algebraic topology, intersection for on
H
1
(X, Z),
_
.
2. 3-manifolds. (a) Knot theory of singularities. (b) Isometries of H
3
and
Aut

C. (c) Deformations of M
3
and M
3
.
3. 4-manifolds. (M, ) studied by introducing J and then pseudo-holomorphic
curves.
4. Dierential geometry: every Riemann surface carries a conformal met-
ric of constant curvature. Einstein metrics, uniformization in higher
dimensions. String theory.
1
5. Algebraic geometry: compact Riemann surfaces are the same as alge-
braic curves. Intrinsic point of view: x
2
+y
2
= 1, x = 1, y
2
= x
2
(x+1)
are all the same curve. Moduli of curves.
1
(/
g
) is the mapping
class group.
6. Arithmetic geometry: Genus g 2 implies X(Q) is nite. Other
extreme: solutions of polynomials; C is an algebraically closed eld.
7. Number theory. Automorphic forms and theta functions. Let q =
exp(2iz) on H. Then f(q) =

q
n
2
is an automorphic form and
f(q)
k
=

a
n
(k)q
n
where a
n
(k) is the number of ways to represent n
as a sum of k ordered squares.
8. Complex geometry: Sheaf theory; several complex variables; Hodge
theory. The Jacobian Jac(X)

= C
g
/ determines an element of
H
g
/ Sp
2g
(Z): arithmetic quotients of bounded domains.
9. Dynamics: unimodal maps exceedingly rich, can be studied by com-
plexication: Mandelbrot set, Feigenbaum constant, etc.
Denition and examples of Riemann surfaces. A Riemann surface is
a connected, Hausdor topological space X equipped with an open covering
U
i
and a collection of homeomorphisms f
i
: U
i
C such that there exist
analytic maps g
ij
satisfying
f
i
= g
ij
f
j
on U
ij
= U
i
U
j
.
We put the denition into this form because it suggest that (g
ij
) is a
sort of 1-coboundary of the 0-chain (f
i
). This equation will recur in the
denition of sheaf cohomology.
The Riemann sphere. The simplest compact Riemann surface is

C =
Cwith charts U
1
= C and U
2
=

C0 with f
1
(z) = z and f
2
(z) = 1/z.
Alternatively,

C

= P
1
is the space of lines in C
2
:
P
1
= (C
2
0)/C

.
The isomorphism is given by [Z
0
: Z
1
] z = Z
1
/Z
0
.
Note that we have many natural (inverse) charts C P
1
given by f(t) =
[ct + d, at + b]. The image omits a/c. Transitions between these charts are
given by Mobius transformations.
2
In fact the full automorphism group of

C is SL
2
(C). This means every
automorphism lifts to a linear map on C
2
.
The round sphere. The round metric on

C is given by 2[dz[/(1 + [z[
2
)
and preserved exactly by
SU(2) =
__
a b
b a
_
: [a[
2
+[b[
2
= 1
_

= S
3
.
Alternatively, if we represent points z
i
in P
1
by unit vectors v
i
in C
2
, then
the spherical distance satises:
cos
2
(d(z
1
, z
2
)/2) = [v
1
, v
2
[
2
.
For example, d(1, e
i
) = , and if we set v
1
= (1, 1)/

2, v
2
= (1, e
i
)/

2,
then
[v
1
, v
2
[
2
= [1 +e
i
[
2
/4 = (1 + cos )/2 = cos
2
(/2).
Q. What is a circle? A. A circle is the zero locus of a Hermitian form
on C
2
of signature (1, 1); for example, the unit circle is the zero locus of
[Z
0
[
2
[Z
1
[
2
. This makes it clear that circles are sent to circles by Aut

C.
The Euclidean and hyperbolic planes. The other two simply-connected
Riemann surfaces are C and H, with automorphism groups B = AN and
SL
2
(R).
Note that P
1
= G/B where G = SL
2
(C); B is called a Borel subgroup
because the associated homogeneous space is a projective variety.
Miraculously, the full automorphism of H preserves the hyperbolic metric
[dz[/y. On the unit disk

= H the automorphism group becomes
SU(1, 1) =
__
a b
b a
_
: [a[
2
[b[
2
= 1
_
,
and the metric becomes 2[dz[/(1 [z[
2
). This 3-manifold is homeomorphic
to a solid torus.
Hyperbolic distance in . As in the spherical case, one can use the inner
product underlying the Hermitian form [Z
0
[
2
[Z
1
[
2
to compute distances
in the hyperbolic metric on the disk. We nd:
cosh
2
(d(z
1
, z
2
)/2) = [v
1
, v
2
[
2
.
For example, d(0, r) =
_
r
0
2 dx/(1x
2
) = 2 tanh
1
(r), and hence cosh
2
(d(0, r)/2) =
(1 r
2
)
1
, which agrees with the inner product squared between the unit
vectors v
0
= (1, 0) and v
r
= (1, r)/

1 r
2
.
3
Doubly-connected Riemann surfaces. The Riemann surfaces with

1
(X)

= Z are given by C

and
A(r) = z : r < [z[ < 1,
up to isomorphism.
In terms of covering spaces, we have C

= C/Z,

= H/Z, and
A(r)

= H/
Z
,
where > 1 and r = exp(2
2
/ log ).
The covering map f : H A(r) can be factored as z log(z), z z
and then z e
z
, where = 2i/ log . In other words, f(z) = z

with
purely imaginary.
Warning. It is not always true that the quotient of a planar domain by a
free group action with discrete orbits is a Hausdor surface! However:
Theorem 1.1 If is a discrete group of isometries acting freely, then the
action is properly discontinuous and X covers the quotient manifold X/.
Triplyconnected Riemann surfaces, etc. Any nitelyconnected pla-
nar surface is equivalent to

C with a nite number of points and round disks
removed. In particular, such a surface depends on only nitely many moduli.
The double of a surface with g+1 circular boundaries is a closed Riemann
surface of genus g, with a real symmetry. In fact the space of such planar
regions is isomorphic to the space of Riemann surfaces dened over R with
the maximal number of real components.
More constructions of Riemann surfaces.
1. Polygons in C,

C, H, glued together by isometries or real analytic
maps. Polyhedra in R
3
. Billiards.
2. Riemannian 2-manifolds: hypersurfaces in R
3
(requires isothermal co-
ordinates or harmonic functions).
3. Quotients X/, Aut(X). Covering spaces.
4. The group (2) SL
2
(Z); isometric for hyperbolic metric; quotient is
triply-punctured sphere.
5. Branched covers. p
n
: by p
n
(z) = z
n
. Elliptic curves as covers
of P
1
branched over four points.
4
6. Algebraic curves. A curve C C
2
dened by f(x, y) = 0 is smooth
df ,= 0 along C. Example: f(x, y) = y
2
p(x) is smooth i p has no
multiple roots. Example: y
2
= x
3
, y
2
= x
2
(x + 1).
7. Symmetric examples. Square with opposite sides identied yields a
curve E

= C/Z[i] of genus one. The quotient E/(z z) is a sphere
with four distinguished points B. The symmetry z iz gives a sym-
metry of the conguration B, xing two points and interchanging two
others. Thus we can take B = 0, , 1.
8. A regular Euclidean octagon with opposite sides identied has an 8-
fold automorphism r : X X. The quotient X/(r
4
)

=

C is branched
over six points B. The action of r gives an order 4 symmetry of B
that xes 2 points and cyclically permutes the remaining 4. Thus X
is dened by y
2
= x(x
4
1).
Note: if one is worried about the vertices of the octagon, one can take
a regular hyperbolic octagon with interior angles of 45

.
9. Genus 2 with 48 symmetries. The curve X dened by y
2
= x(x
4
1)
can also be regarded as the degree 2 cover of P
1
branched over the ver-
tices of a regular octahedron. The latter solid has symmetry group S
4
.
The element x ix in S
4
lifts to the automorphism (x, y) (ix,

iy)
in Aut(X), showing that Aut(X) =

S
4
is actually the unique nontriv-
ial central extension of S
4
. Geometrically, Aut(X) is the preimage of
S
4
in SU(2).
10. In fact the regular octagon gives the same surface of genus two. In-
deed, we can map the regular octagon to the square by z z
2
(up
to Riemann mappings), and then map the square conformally to com-
plement in

C of the rays the 4th roots of unity to the origin, sending
the center of the square to innity. Because of its folding, this map
respects the identications on the sides of the octagon and gives an
explicit degree two map of the octagon surface to

C, branched over
the vertices of a octahedron.
This map gives nothing more than the quotient of the octagon surface
by a 180

rotation. Its xed points (the Weierstrass points) correspond

to the vertices and edge-midpoints of the octagon, and its center.
5
2 Maps between Riemann surfaces
Let f : X Y be a nonconstant map. Then f is locally modeled on z z
m
.
We write m(f, p) for the multiplicity of f at p. It follows that:
1. The map f is open and discrete.
2. f satises the maximum principle.
3. If X is compact, then f(X) = Y .
4. An analytic function on a compact Riemann surface is constant.
Removable singularities. An isolated singularity of a bounded analytic
function is removable. Consequently:
1. A bounded analytic function on a Riemann surface of nite type (com-
pact with a nite number of points removed) is constant.
2. A bounded analytic function on C is constant. (Proof: it extends to

C.)
3. Every polynomial of degree d 1 has a root in C.
Covering maps. Let : X Y be a covering space of a Riemann surface
Y . There is a unique complex structure on X such that is holomorphic.
The space X is determined up to isomorphism over Y by the subgroup
H

=
1
(X, p)
1
(Y, q).
The deck group is dened by = Deck(X/Y ) Aut(X) is the group
of automorphisms such that = . We say X/Y is normal (Galois,
regular) if the deck group acts transitively on the bers of .
In general Deck(X/Y ) = N(H)/H where N(H) is normalizer of H in
G. To see this think of Y as

Y /G: then a deck transformation lifts to a g
in G satisfying gh
1
x = h
2
gx, so it descends to X =

Y /H. This requires
gh
1
g
1
= h
2
and thus g N(H).
The key property of covering space is an algebraic solution to the lifting
problem for f : Z Y .
Branched coverings. Now let : X Y be a general nonconstant map
between Riemann surfaces. Let C = x : mult(f, x) > 1, let B = f(C)
and let

B =
1
(B). Let X

= X

B and Y

= Y B.
We say is a branched covering if B is discrete, : X

is a
covering map, and for any small loop around a single point of B, every
component of
1
() is compact.
6
Theorem 2.1 A branched covering X is uniquely determined, up to iso-
morphism over Y , by a discrete set B and a subgroup H G =
1
(Y, p).
Any subgroup that meets each peripheral subgroup of G with nite index
determines such a covering.
As before a covering is normal (Galois, regular) if Deck(X/Y ) acts tran-
sitivity on bers; this is equivalent to X

/Y

being normal.
Theorem 2.2 Any properly discontinuous subgroup Aut(X) yields a
quotient Riemann surface Y and a branched covering map : X Y with
Deck(X/Y ) = .
Example. The innite dihedral group Aut(C) corresponds to the
branched covering map cos : C C/

= C. Aside from ordinary covering
spaces, the only other examples for C are those coming from elliptic curves
with automorphisms, i.e. branched covers C E = C/

C = E/G,
where G Aut(E) is a nite group.
Proper maps. Let f : X Y be a proper, nonconstant map between
Riemann surfaces. That is, assume K compact implies f
1
(K) compact.
Then:
1. f is closed: i.e. E closed implies f(E) closed. (This requires only local
connectivity of the base Y .)
2. f is surjective. (Since it is also open).
3. If D X is discrete, so is f(D). (Since f(D) meets any compact set
K in a nite set, namely the image of D f
1
(K).)
4. In particular, the branch locus B(f) Y is discrete.
5. f
1
(q) is nite for all q Y . (Since f is discrete.)
6. For any neighborhood U of f
1
(q) there exists a neighborhood V of q
whose preimage is contained in U. (Since f(XU) is closed and does
not contain q.)
7. If f is a proper local homeomorphism, then it is a covering map. (This
only requires that Y is locally compact.)
Proof. A local homeomorphism is discrete, so given q Y we can
choose neighborhoods U
i
of its preimages p
1
, . . . , p
n
such that f : U
i

V
i
is a homeomorphism. Let V be a neighborhood of q such that
f
1
(V ) is contained in

U
i
. Then f evenly covers V .
7
8. For any q Y with f
1
(q) = p
1
, . . . , p
n
, there exists a disk V
containing q such that f
1
(V ) =

U
i
with U
i
a disk, p
i
U
i
and
f
i
= f[U
i
satises f
i
: (U
i
, p
i
) (V, q) is conjugate to z z
d
i
on .
(Here one can use the Riemann mapping theorem to prove that if
g : is given by g(z) = z
m
, and V is a simply-connected
neighborhood of z = 0, then the pullback of g to V is also conjugate
to z z
m
.)
9. The function

f(p)=q
m(f, p) is independent of q. It is called the degree
of f.
Properness and branched covers.
Theorem 2.3 Any proper map is a branched covering. A branched covering
is proper i it has nite degree, i.e. the cardinality of one (and hence every)
ber is nite, i.e. X

has nite degree.

Proof. For the rst statement use the fact that a proper map has a degree
set of branch values, and a proper local homeomorphism is a covering. For
the second, use the fact that a nite covering map is proper.
Extension of proper maps. Let Y

be the complement of a discrete set D

in a Riemann surface Y , and let f : X

be a proper map. Then there

is a unique way to complete X and f to obtain a proper map f : X Y .
Examples.
1. Any rational map f :

C

C is a proper branched covering; it is a
covering map i it is a Mobius transformation. An entire function is
proper i it is a polynomial.
2. Any proper map of the disk to itself is given by a Blaschke product,
f(z) = exp(i)
d

1
z a
i
1 a
i
z

3. The map f :

C

C given by f(z) = z
d
is a degree d, proper, regular
branched cover with B = 0, and Deck(X/Y ) = Z/d. Its restric-
tion to a map C

is a covering map. (As a map C

C, it is
not a covering map.)
8
4. The map f(z) = z
d
+1/z
d
is a regular branched cover, with deck group
the dihedral group generated by z 1/z and z exp(2i/d)z.
5. A basic example of an irregular cover is given by f(z) = z
3
3z. This
map has branched points at 1 and branch values B = 2; we have

B = f
1
(B) = 1, 2, and f : C

B CB is an irregular degree
three cover. It corresponds to the abbaab triple cover of a bouquet
of circles.
6. The map f(z) = z
3
: H C is not proper, even though it is a
local homeomorphism. Its extension to H is proper, but not a local
homeomorphism.
7. The map f(z) = e
z
: C C

is a regular covering map, but not

proper. Its deck group Z is generated by z z + 2i.
8. The map f(z) = tan(z) : C f(C)

C is a covering map. In fact
tan(z) = g(e
2iz
) where g(z) = i(z1)/(z+1). Thus f(C) =

Ci.
9. The map f(z) = cos(z) : C C is a regular branched covering. Its
branch values are B = 1, and its branch points are

B = Z. The
deck group is the innite dihedral group generated by z z and
z z + 2. Note that cos(z) = g(e
iz
) where g(z) = (z + 1/z)/2.
The Riemann-Hurwitz formula. A Riemann surface X is of nite type
if it is obtained from a compact Riemann surface by deleting a nite number
of points. In this case (X) = 2 2g n where g is the genus and n is the
number of points removed.
Theorem 2.4 If f : X Y is a branched covering map between Riemann
surfaces of nite type, then
(X) = d(Y )

X
(mult(f, x) 1).
Since (X) = 2 2g(X), this formula also relates the genera of X and
Y .
Example: Let f :

C

C be a map of degree d. The f has 2d2 critical
points.
Example: The only compact Riemann surfaces admitting self maps of
degree d > 1 are

C and C/. The latter only admit self covering maps.
(Second proof: if f : C C is a lift of a self-map of C/, then f

(z) is
a holomorphic, doubly-periodic function, hence constant.)
9
Belyis Theorem. The Riemann surfaces that arise as coverings of

C
branched over B = 0, 1, are especially interesting.
Theorem 2.5 A compact Riemann surface X is dened over a number eld
i it can be presented as a branched cover of

C, branched over just 3 points.
We rst need to explain what it means for X to be dened over a number
eld. For our purposes this means there exists a branched covering map
f : X

C with B(f) consisting of algebraic numbers. Alternatively, X
can be described as the completion of a curve in C
2
dened by an equation
f(x, y) = 0 with algebraic coecients.
Grothendieck wrote that this was the most striking theorem he had heard
since at age 10, in a concentration camp, he learned the denition of a circle
as the locus of points equidistant from a given center.
Proof. I. Suppose X is dened over a number eld, and f : X

C is given
with B(f) algebraic. We will show there exists a polynomial p :

C

C such
that B(p) 0, 1, and p(B(f)) 0, 1, . This suces, since
B(p f) = B(p) p(B(f)).
We may assume , B(f). Let deg(B) denote the maximal degree over
Q of the points of B, and let z B(f) have degree deg(B). Then there
exists a polynomial p Q[z] of degree d such that p(z) = 0. Moreover
deg(B(p)) d 1 since the critical values of B are the images of the zeros
of p

(z). Thus
B

= B(p) p(B)
has fewer points of degree deg(B) over Q. Iterating this process, we can
reduce to the case where B Q.
Now comes a second beautiful trick. Consider the polynomial p(z) =
Cz
a
(1 z)
b
. This polynomial has critical points at 0, 1 and w = a/(a +b).
By choosing the value of C Q correctly, we can arrange that p(w) = 1
and hence B(p) 0, 1, . Thus if w B is rational, we can choose p
so that B

= p(B) B(p) has fewer points outside 0, 1, . Thus we can

eventually eliminate all such points.
II. Let X be a Riemann surface presented as a branched covering f :
X

C with B(f) = 0, 1, . We need a nontrivial fact: there exists a
second g /(X) and a conite set X

X such that (f, g) : X

C
2
is
an immersion with image the zero locus V
p
of a polynomial p(x, y) C[x, y].
This polynomial has the property that the projection of (the normaliza-
tion of) V
p
to the rst coordinate which is just f is branched over
10
just 0, 1 and . The set of all such polynomials of given degree is an alge-
braic subset W of the space of coecients C
N
for some large N. This locus
W is dened by rational equations. Thus the component of W contain-
ing our given p also contains a polynomial q with coecients in a number
eld. But V
q
and V
p
are isomorphic, since they have the same branch locus
and the same covering data under projection to the rst coordinate. Thus
X

= V
p

= V
q
is dened over a number eld.
Corollary 2.6 X is dened over a number eld i X can be built by gluing
together nitely many unit equilateral triangles.
Proof. If X can be built in this way, one can 2-color the barycentric
subdivision and hence present X as a branched cover of the double of a
30-60-90 triangle. The converse is clear, since the sphere can be regarded as
the double of an equilateral triangle.
Example: A square torus can be built by gluing together 8 isosceles right
triangles. But these can also be taken to be equilateral triangles, with the
same result! This is because the double of any two triangles is the same, as
a Riemann surface.
Trees and Belyi polynomials. Let us say p : C C is a Belyi polynomial
if p(0) = 0, p(1) = 1 and the critical values of p are contained in 0, 1.
Then p
1
[0, 1] is a topological tree T C. We adopt the convention that
any inverse image of 0 or 1 is a vertex of three tree, and that the vertices of
T at 0 and 1 are labelled.
It is a consequence of the uniformization theorem that every tree T
embedded in C, with two vertices labelled 0 and 1 (and an even number of
vertices between them), determines a unique Belyi polynomial.
For example, a star with d endpoints corresponds to p(z) = z
d
. A
segment with 2 interval vertices marked 0 and 1 corresponds to the cubic
polynomial p(z) = 3z
2
2z
3
. To see this, we note that p

(z) = az(1 z)
since the two vertices are internal, and use the normalization p(0) = 0 and
p(1) = 1 to solve for a and then for p(z).
Describing compact Riemann surfaces and their meromorphic func-
tions. Let X be a compact Riemann surface, and let f : X

C be a mero-
morphic function of degree d 1. Then (X, f) determines the following
data:
A nite set B = B(f)

C; and
11
A monodromy map :
1
(

C B, y) S
d
.
To obtain the latter, choose y Y

=

CB and label the points in f
1
(y)
as x
1
, . . . , x
d
. Then path lifting gives a natural action of
1
(Y

, y) on the
ber over y, and hence a representation :
1
(Y

, y) S
d
. Explicitly,
()(i) = j if a lift of starting at x
i
terminates at x
j
.
Let X

= Xf
1
(B). Then f : X

is a covering map, and clearly

1
(X

, x
i
)

=
1
(Y

, y) : ()(i) = i.
To be even more explicit, let us choose loops
1
, . . . ,
n
based at y and en-
circling the points of B = (b
1
, . . . , b
n
) in order. Then
1
(Y

, y) is generated
by these elements, with the single relation

1

n
= e.
Thus is uniquely determined by the permutations
i
= (
i
), which must
satisfy two properties:
The product
1

n
= e in S
d
; and
The group G =
1
, . . . ,
n
is a transitive subgroup of S
d
.
The conjugacy class of
i
in S
d
determines a partition p
i
of d, which is
independent of choices. This partition records how the sheets of X

come
together over b
i
.
Theorem 2.7 The pair (X, f) is uniquely determined by the data (B,
1
, . . . ,
n
)
above, and any data satisfying the product and transitivity conditions arises
from some (X, f).
Proof. The covering space f : X

is determined, up to isomorphism
over Y , by the preimage of G S
d1
in
1
(Y

, y); lling in the branch

points, we obtain the desired compact Riemann surface X and meromorphic
function f.
Moduli of Riemann surfaces. Note that if we x the genus of X and
the degree of f, then the cardinality of B is also bounded. On the other
hand, one can show that X admits a meromorphic whose degree is bounded
in terms of g. Thus Riemann surfaces of genus g depend on only nitely
many parameters. A more precise bound will be computed later.
12
Hurwitz problem. Here is an unsolved problem. Let f : X

C be a
branched covering of degree d with critical values B = (b
i
)
n
1
. Let p
i
be the
partition of d corresponding to the ber over b
i
. What partitions (p
1
, . . . , p
n
)
can be so realized?
This is really a problem in topology or group theory. We have to lift
each p
i
to an element g
i
S
d
in the conjugacy class specied by p
i
, in such
a way that g
1
g
n
= e.
Example 1: For n = 3 and d = 4, the partitions 2+2, 2+2, 3+1 cannot
be realized, since any 2 involutions in A
4
commute.
Example 2: If
1
, . . . ,
n1
lie in A
n
, then so must
n
. So for example,
the partitions of d = 5 given by 2 +2 +1, 3 +1 +1 and 2 +1 +1 +1 cannot
coexist.
For more details, see [EKS].
Have we seen all compact Riemann surfaces? Yes! A basic result,
the Riemann existence theorem, to be proved later, states:
Theorem 2.8 Every compact Riemann surface admits a nonconstant mero-
morphic function. In fact, the elements of /(X) separate points.
3 Sheaves and analytic continuation
Presheaves and sheaves. A presheaf of abelian groups on X is a functor
T(U) from the category of open sets in X, with inclusions, to the category
of abelian groups, with homomorphisms. (This functor is covariant, i.e. it
reverses the directions of arrows.)
It is a sheaf if (I) elements f T(U) are determined by their restrictions
to an open cover U
i
, and (II) any collection f
i
T(U
i
) with f
i
= f
j
on U
ij
for all i, j comes from an f T(U).
These axioms say we have an exact sequence
0 T(U)

T(U
i
)

T(U
ij
).
Here U =

U
i
is an open cover, the index ij is considered to be ordered,
U
ij
= U
i
U
j
, and the maps are given by f f[U
i
and (f
i
) (f
i
f
j
)[U
ij
.
Example: applying (I) to the empty cover of the empty set, we see
T() = (0). Thus the presheaf that assigns a xed, nontrivial group G to
every open set is not a sheaf.
Examples of presheaves and sheaves. The sheaves (, (

, O and /, of
rings of continuous, smooth, holomorphic and meromorphic functions. The
multiplicative group sheaves O

and /

.
13
If G is a nontrivial abelian group, T(U) = G for U nonempty and T() =
(0) is a presheaf. But it is not a sheaf: T(U
1
U
2
) ,= GG.
To rectify this we can dene T(U) to be the additive group of locally
constant maps f : U G. This is now a sheaf; it is often denoted simply
by G. (E.g. R, C, Q, S
1
, Z.)
The presheaf T(U) = O(U)/C(U) is not a sheaf. For example, local
logarithms do not assemble.
Pushforward. We note that sheaves naturally push forward under contin-
uous maps, i.e. we can dene

(T)(U) = T(
1
(U)).
We can also pull back a sheaf under a local homeomorphism.
In particular, homeomorphisms between spaces give isomorphisms of
sheaves.
Structure. From a modern perspective, a Riemann surface can be dened
as a pair (X, O
X
) consisting of a connected Hausdor topological space and
a sheaf that is locally isomorphic to (C, O
C
). Similarly denitions work for
smooth manifolds. For example: a homeomorphism f : X Y between
Riemann surfaces is holomorphic i it sends O
X
to O
Y
.
Stalks. Let T be a presheaf.
The stalk T
x
is the direct limit of T(U) over the (directed) system of
open sets containing x X. It can be described directly as the disjoint
union of these groups modulo f
1
f
2
if they have a common restriction
near x. (Alternatively, T
x
= (T(U))/N where N is generated by elements
of the form (f[U
1
) (f[U
2
), f T(U
1
U
2
).)
There is a natural map T(U) T
x
for any neighborhood U of x. We
let f
x
denote the image of f T(U) under this map. We have f
x
= 0 i
there is a neighborhood V of x such that f[V = 0.
Example: O
a

= Cz
a
for any local chart (uniformizer) z
a
: U C,
z
a
(a) = 0. Note that O
a
is a local ring with residue eld C, given by
evaluation at z
a
= 0.
Theorem 3.1 Let T be a sheaf. Then f T(U) is zero i f
x
= 0 for all
x U.
Espace etale. The espace etale [T[ of a presheaf is the disjoint union
of the stalks T
x
, with a base for the topology given by sets of the form
[U, f] = f
x
: x U. It comes equipped with a natural projection
p : [T[ X which is a local homeomorphism.
14
Example. For a (discrete) group G, we have [G[ = GX with the product
topology.
We say T satises the identity theorem if whenever U is open and con-
nected, f, g T(U) and f
x
= g
x
for some x U, then f = g. Examples: O
and /.
Theorem 3.2 Assume X is Hausdor and locally connected, and T satis-
es the identity theorem. Then [T[ is Hausdor.
Proof. We need to nd disjoint neighborhoods centered at a pair of distinct
germs f
x
and g
y
. This is easy if x ,= y, since X is Hausdor. Otherwise,
choose a connected open set U containing x = y, and small enough that
both f
x
and g
x
are represented by f, g T(U). Then the identity theorem
implies the neighborhoods [U, f] and [U, g] are disjoint, as required.
Structure of |O|. Since analytic functions satisfy the identity theorem,
[O[ is Hausdor. There is a unique complex structure on [O[ such that
p : [O[ X is an analytic local homeomorphism.
There is also a natural map F : [O[ C given by F(f
x
) = f(x). This
map is analytic. However [O[ is never connected, since the stalks O
x
are
uncountable.
Local rings. The construction of the function F on [O[ appears ad hoc. It
can be made more understanble by noting that O
p
is a local ring; that is, it
has a unique maximal ideal m
p
consisting of germs which vanish at p. Then
F(f
x
) is really the image of f
x
in the residue eld k
x
= O
x
/m
x
. For this
space to be canonically identied with C, it is probably best to regard O as
a sheaf of C-algebras.
If we play the same game with Spec Z, then k
p
= F
p
and k
0
= Q. Thus
an integer n Z can be regarded as a section n : Spec Z [O[, with
n(p) = nmodp.
Path lifting. We now recall some results from the theory of covering spaces.
Let p : X Y be a local homeomorphism between Hausdor spaces. Let
f : I = [0, 1] Y be a path, and x X a point such that p(x) = f(0). A
lifting of f based at x is a path F : I X such that f = pF and F(0) = x.
Theorem 3.3 A lifting is unique if it exists.
Proof. If we have two liftings, F
1
and F
2
, the set of t I such that
F
1
(t) = F
2
(t) is open since p is a local homeomorphism, closed since X is
Hausdor, and nonempty since F
1
(0) = F
2
(0) = x. By connectedness it is
the whole interval.
15
Now let f
s
(t) be a homotopy of paths parameterized by s [0, 1], such
that f
s
(0) = y
0
and f
s
(1) = y
1
are constant. Suppose for every s there is a
lift F
s
of f
s
based at x
0
. We then have:
Theorem 3.4 (Monodromy Theorem) The terminus F
s
(1) = x
1
is in-
dependent of s. Moreover F
s
(t) is a lift of f
s
(t) as a function on I I. In
particular, F
0
(t) and F
1
(t) are homotopic rel their endpoints.
Non-unique path lifting. Let T be the space of continuous functions
on R. Consider the functions f(x) = 0 and g(x) = max(0, x). The germs
of these functions determine a subspace Y of [T[. The space p : Y R
has 1 points over (, 0) and two points over [0, ). In particular the two
solutions to p(y) = 0 cannot be separated by disjoint open sets, since these
sets always meet over the negative real axis.
In this case f
a
= g
a
for any a < 0, and the path [a, 1] R can be lifted
to [T[ in two dierent ways, both starting at this point.
Structure of |C
k
|. Suppose : [0, 1] X with (0) = a and (1) = b.
Even though path lifting is not unique in [C
0
[, if two lifts satisfy
1
(t) =
2
(t)
on [0, 1] then f
i
=
i
(1) satises f
1
(b) = f
2
(b). Similarly, in [C
k
[, the rst k
derivaties of f
i
at b must also agree. We can think of [C
k
[ as combing the
lifts so they are tangent to higher and higher order when they diverge. In
the limit of C

we have tangency to innite order, but divergence is still

possible. Finally for real analytic functions /, there is no divergence, and
[/[ becomes Hausdor.
Analytic continuation. Let O be the sheaf of analytic functions on X =
C. Let : [0, 1] C be path with a = (0) and b = (1). Let f
a
be the
germ of an analytic function at z = a, i.e. let f O
a
.
We say f
b
O
b
is obtained from f
a
by analytic continuation along if
there are analytic functions g
i
on balls B
i
containing (t
i
), i = 0, . . . , n such
that:
1. 0 = t
0
< t
1
< < t
n
= 1;
2. B
i
meets B
i+1
, g
i
and g
i+1
agree there, and [t
i
, t
i+1
] B
i
B
i+1
;
and
3. g
0
= f
a
and g
1
= f
b
.
Example: For a = 0, f
a
(z) =

z
n
/n = log(1/(1z)) can be analytically
continued to any point in C

, in many ways; the various possibilities for f

b
(z)
dier by multiplies of 2i.
16
Observation: analytic continuation along is equivalent to path-lifting
to the space p : [O[ C.
Corollary 3.5 An analytic continuation is unique if it exists.
Corollary 3.6 (Monodromy theorem) If analytic continuation from a
to b is possible along a family of paths
s
, s I, then the result f
b
is always
the same.
Maximal analytic continuation (spreads). If f : X Y is holomor-
phic and f(a) = b then there is a natural pullback map f

: O
b
O
a
given
by composition with f. If mult(f, a) = 1 then f

is an isomorphism and
hence there is also a pushforward f

: O
a
O
b
. (Remark: by summing
over sheets, pushforward can also be dened at general points.)
To keep track of the potential multi-valuedness, we dene an analytic
continuation of f
a
O
a
, a C, to be a pointed Riemann surface (Y, b)
endowed with an analytic local homeomorphism p : (Y, b) (C, a) and an
F O(Y ) such that p

F
b
= F
a
.
We say an analytic continuation (Y, b) is maximal (universal might be a
better terminology) if, for any other analytic continuation (Y

, b

), there is a
unique (analytic) local homeomorphism (Y

, b

) (Y, b) making the natural

diagrams commute.
Theorem 3.7 Every germ of an analytic function f
a
has a unique maximal
analytic continuation, obtained by taking Y to be the connected component
of [O[ containing f
a
, and restricting the natural maps F : [O[ C and
p : [O[ C to Y .
Proof. The map y p

(F
y
) gives a local homeomorphism of any analytic
continuation into [O[, which is then tautologically contained in the maximal
one described above.
Examples.
1. The maximal domain of f(z) =

z
n!
is the unit disk.
2. For f(z) =

z
n
, the maximal analytic continuation in C is : C
1 C with (z) = z and F(z) = 1/(1 z). Evidentally f((z)) =
1/(1 z) = F(z).
17
3. For f(z) =

z
n+1
/(n + 1), the maximal analytic continuation is :
C C 1 with (z) = 1 e
z
and F(z) = z. Check: f(z) =
log(1 z) so f((z)) = log(e
z
) = z = F(z).
4. Let f(z) =
_
q(z) where q(z) = 4z
3
+ az + b. Then its maximal
analytic continuation is given by Y = E E[2] with E = C/, with
p : Y C given p(z) = (z), and with F(z) =

(z).
Remark. One can replace the base C of analytic continuation with any
other Riemann surface X. Note however that the maximal analytic con-
tinuation may become larger under an inclusion X X

.
4 Algebraic functions
We will develop two main results.
Theorem 4.1 Let : X Y be a holomorphic branched covering of degree
d. Then /(X)//(Y ) is an algebraic eld extension of degree d.
To see the degree is exactly d we need to know that /(X) separates the
points of X, i.e. we will appeal to the Riemann Existence Theorem.
Theorem 4.2 Let K//(Y ) be a eld extension of degree d. Then there is
a unique degree d branched covering : X Y such that K

= /(X) over
/(Y ).
Moreover /(X)//(Y ) is Galois i X/Y is Galois, in which case there
is a natural isomorphism
Gal(/(X)//(Y ))

= Deck(X/Y ).
Putting these results together, we nd for example that X /(X)
establishes an equivalence between (i) the category of nite-sheeted branched
coverings of

C, and (ii) the category of nite eld extensions of C(x).
Symmetric functions. The proof will implicitly use the idea of symmetric
functions. Recall that the elementary symmetric functions s
i
of f
1
, . . . , f
d
are related to the coecients c
i
Z[f
1
, . . . , f
d
] of the polynomial
P(T) =
d

1
(T f
i
) = T
d
+c
1
T
d1
+ +c
d
(4.1)
18
by s
i
= (1)
i
c
i
. (Thus s
1
=

f
i
, s
2
=

i<j
f
i
f
j
and s
d
=

f
i
.) Clearly
these polynomials s
i
lie in Z[f
i
]
S
d
and in fact they generate the ring of
invariant polynomials [La, IV.6].
Separation of points. We also observe the following consequence of the
fact that /(X) separates points.
Proposition 4.3 Let A X be a nite subset of a compact Riemann sur-
face. Then there exists a meromorphic function f : X

C such that f[A is
injective.
Proof. Let V /(X) be a nite-dimensional space of functions that
separate the points of A = a
1
, . . . , a
n
. Then for any i ,= j, the subspace
V
ij
V where f(a
i
) ,= f(a
j
) has codimension at least one. It follows that
V ,=

V
ij
, and hence a typical f V is injective on A.
Proof of Theorem 4.1. Let : X

be the unbranched part of the

degree d branched covering : X Y , and let f /(Y ). By deleting
more points if necessary, we can assume f O(X

).
Given y Y

, dene a degree d polynomial in T by

P
y
(T) =

(x)=y
(T f(x)).
Cleary P
y
(T) is globally well-dened on Y

.
We now check its local properties. Since is a covering map over Y

, we
can choose local sections s
i
: U X

such that P
y
(T) =

(T f(s
i
(y)))
on U. This shows that the coecients of P
y
(T) are analytic functions of
y. On the other hand, the denition of P
y
makes no references to these
local section, so we get a globally well-dened polynomial P O(Y

)[T]
satisfying P(f) = 0.
Near any puncture of Y over which f takes nite values, the coeccients
of P
y
(T) are bounded, so they extend across the puncture. At any puncture
where f has a pole, we can choose a local coordinate z on Y and n > 0
such that z
n
f has no poles. Then there is a monic polynomial Q(T) with
holomorphic coecients such that Q(z
n
f) = 0 locally, constructed just as
above. On the other hand, P(T) = z
nd
Q(z
n
T), since both sides are monic
polynomials with the same zeros. Thus the coecients of P(T) extend to
these punctures as well, yielding meromorphic functions on all of Y .
In other words, we have shown there is a degree d polynomial P
/(Y )[T] such that P(f) = 0. It follows that deg(/(Y )//(X)) is at most
d (e.g. by the theory of the primitive element).
19
To see the degree is exactly d, choose a function f /(Y ) which
seperates the d points in a ber over y
0
Y . Then P(f) = 0 for a
polynomial of degree d as above. If P factors Q(T)R(T), then Q(f)R(f) = 0
in /(Y ). Since Y is connected, /(Y ) is a eld and one of these terms
vanishes identically say Q(f) = 0. Then the values of f on the ber
over y
0
are zeros of the polynomial Q
y
0
(T). Since deg Q < d, this is a
contradiction.
Compositums of elds. It is well-known in eld theory that if K
1
, K
2
are two nite extensions of k, then there exists an extension L containing
both the compositum of K
1
and K
2
. An analogous fact holds for Riemann
surfaces.
Theorem 4.4 Let f
i
: X
i


C, i = 1, 2 be a pair of Riemann surfaces
presented as branched coverings of

C. Then there is a third Riemann surface
Y dominating them both.
More precisely, there is a Riemann surface Y and a pair of holomorphic
maps g
i
: Y X
i
such that f
1
g
1
= f
2
g
2
.
Proof. Let B = B(f
1
) B(f
2
). Then we can regard X
i
as the branched
covering of

C determined by a subgroup of nite index H
i

1
(

C B).
Now let Y be the branched covering determined by H
1
H
2
.
The eld /(Y ) is a compositum of /(X
1
) and /(X
2
).
Note that Y is not quite canonical, because we must choose basepoints
to make the construction. However Y can be regarded as an irreducible
component of the ber product
X
1

C
X
2
= (x
1
, x
2
) X
1
X
2
: f
1
(x
1
) = f
2
(x
2
),
which is canonical.
Local algebraic functions. Now let us go backwards from a polynomial
P(T) of degree d to analytic functions (f
i
)
d
1
. We rst work locally.
Let P(T) be a monic polynomial of degree d with coecients c
i
(z) in
the local ring O
a
, a Y .
Theorem 4.5 If P(T, a) = T
d
+ c
1
(a)T
d1
+ + c
d
(a) has simple zeros,
then there exist f
i
O
a
such that P(T) =

(T f
i
).
20
Proof. This is simply the statement that the zeros of P(T, z) vary holo-
morphically as its coecients do. To make this precise, let w
1
, . . . , w
d
be
the zeros of P(T, a), and let
1
, . . . ,
d
be the boundaries of disjoint disks in
C, centered at these zeros. We can then take:
f
i
(z) =
1
2i
_

i
P

(, z) d
P(, z)

Here f
i
(z) is analytic in z so long as P(, z) never vanishes on
i
. This is
true for P(, a), so by continuity it remains true in a small neighborhood of
z = a.
Resultant and discriminant. To go further, it is useful to recall the
resultant R(f, g).
Let K be a eld, and let f, g K[T] be nonzero polynomials with
deg(f) = d and deg(g) = e. Recall that K[T] is a PID and hence a UFD.
We wish to determine if f and g have a common factor, say h, of degree 1 or
more. In this case f = hf
1
, g = hg
1
and hence g
1
f f
1
g = 0. Conversely,
if we can nd nonzero r and s with deg(r) < e and deg(s) < d such that
rf +sg = 0, then f and g have a common factor.
The existence of such r and s is the same as a linear relation among the
elements (f, xf, . . . , x
e1
f, g, xg, . . . , x
d1
g), and hence it can be written as
a determinant R(f, g) which is simply a polynomial in the coecients of f
and g. We have R(f, g) = 0 i f and g have a common factor.
The discriminant D(f) = R(f, f

) is nonzero i f has simple zeros.

Example: the cubic discriminant. Let f(x) = x
3
+ ax + b, so f

(x) =
3x
2
+a. Then D(f) = R(f, f

) is given by:
det
_
_
_
_
_
_
_
_
_
1 0 a b 0
0 1 0 a b
3 0 a 0 0
0 3 0 a 0
0 0 3 0 a
_
_
_
_
_
_
_
_
_
= 4a
3
+ 27b
2
.
The number is only well-dened up to sign. For the usual denition, D(f) =

(x
i
x
j
)
2
, one gets 4a
3
27b
2
.
From a eld extension to a branched cover. Now let K be a degree
d eld extension of /(Y ). By the theorem of the primitive element, there
exists a monic irreducible polynomial P(T) with coecients c
i
/(Y ) such
that K

= /(Y )[T]/(P(T)).
21
By irreducibility, the discriminant D(P) /(Y ) is not identically zero.
Let Y

Y be the complement of the zeros and poles of D(P), and of the

poles of the coecients c
i
.
Let X

[O
Y
[ be the set of germs f
a
such that P(f
a
, a) = 0. By the
preceding local result, X

is a degree d covering space of Y

. The tautolog-
ical map f : X

C sending f
a
to f
a
(a) is analytic. We can complete X

to a branched covering : X Y , and extend f to a meromorphic function

on X using Riemanns removable singularities theorem.
By construction, P(f) = 0.
We must show that X is connected. If not, one of its connected com-
ponents X
0
/Y has degree d
0
< d. But then f[X
0
satises a polynomial
Q /(Y )[T] of degree d
0
< d, which is a factor of P. This contradicts
irreducibility of P.
It remains to consider the Galois group. Suppose X/Y is Galois as a
branched covering. Since P
a
(T) has distinct zeros for some a Y , the group
Deck(X/Y ) maps injectively into Gal(/(X)//(Y )); indeed, only the iden-
tity stabilizes the function f. By degree considerations, this map is surjective
as well. Similarly, if /(X)//(Y ) is Galois, then G = Gal(/(X)//(Y ))
permutes the roots of the polynomial P(T). These correspond to the sheets
of X, so we get a map G Deck(X/Y ) which is an isomorphism again by
degree considerations.
The Riemann surface X can be regarded as a completion of the maximal
analytic continuation of f
a
, for any germ f
a
O
a
(Y ) satisfying P(f
a
, a) = 0
at a point where the discriminant of P(T) is not zero.
Construction as a curve. An alternative to the construction of : X
Y is the following. Fix P(T) /(Y )[T], irreducible, and as before let Y

be the locus where the discriminant is nonzero and where the coecients of
P(T) are holomorphic. Then dene
X

= (x, y) : P
y
(x) = 0 C Y

.
Let (F, ) be projections of X

to its two coordinates. Then : X

is
a covering map, and F : X

C is an analytic function satisfying P(F) = 0;

and the remainder of the construction carries through as before.
Examples. For any polynomial p(z) C[z] which is not a square (i.e.
which at least one root of odd order), we can form the Riemann surface
: X

C corresponding to adjoining f(z) =
_
p(z) to K = C(z) = /(

C).
When p has 2n or 2n1 simple zeros, the map is branched over 2n points
(including innity in the latter case), and hence X has genus n1. Note that
the polynomial P(T) = T
2
p(z) has discriminant Res(P, P

) = 4p(z).
22
Puiseux series.
Avec les series de Puiseux,
Je marche comme sur des oeufs.
Il sensuit que je les fuis
Comme un poltron que je suis.
A. Douady, 1996
(Variante: Et me refugie dans la nuit.)
Let K = /
p
be the local eld of a point p on a Riemann surface; it
is isomorphic to the eld of convergent Laurent series

n
a
i
z
i
in a local
parameter z O
p
with z(p) = 0.
Theorem 4.6 Every algebraic extension of /
p
of degree d is of the form
L = K[], where
d
= z.
Proof. Let P(T) K[T] be the degree d irreducible polynomial for a
primitive element f L (so L = K(f)).) Let Y = [z[ < r be a small
neighborhood of P. Then we can nd an r > 0 such that the coecients c
i
of
P(T) are well-dened on Y , only have poles at z = 0, and the discriminant
D(P)[Y vanishes only at z = 0 as well.
Note that P(T) remains irreducible as a polynomial in /(Y )[T]. Thus
it denes a degree d branched covering X Y , branched only over the
origin z = 0. But there is also an obvious branched covering Y
d
Y given
by
d
= z. Since
1
(Y

)

= Z has a unique subgroup of index d, we
have Y
d

= X over Y .
Corollary 4.7 Any irreducible, degree d polynomial P(T) = 0 with coe-
cients in /
p
has a solution of the form
f(z) =

n
a
i
z
i/d
.
In particular, any polynomial is locally solvable by radicals, i.e. its
roots can be expressed in the form f
i
(z) = g
i
(z
1/d
), where g
i
(z) is analytic.
Equivalent, after the change of variables z =
d
, the polynomial P(T) factors
into linear terms.
Example. For P(T) = T
3
+T
2
z = 0, we have a solution
f(z) = z
1/2

z
2
+
5z
3/2
8
z
2
+
231z
5/2
128

23
Why is this only of degree 1/2? Because the original polynomial has two
distinct roots when z = 0: it is reducible over /
0
!
This calculation can be carried out recursively by setting = z
1/2
and
observing that
f =
_
1 f
3
/
2
=
_
1
f
3
2
2

f
6
8
4

_
.
Note: it is a general fact that only the primes dividing d appear in the
denominators of the power series for (1 x)
1/d
.
Riemann surfaces Number elds
K = /(

C) = C(z) K = Q
A = C[z] A = Z
p C pZ prime ideal
uniformizer z
p
= (z p) C[z] uniformizer p Z
order of vanishing of f(z) at p power of p dividing n Z
A
p
= O
p
=

0
a
n
z
n
p
A
p
= Z
p
=

a
n
p
n

m
p
= z
p
O
p
m
p
= pZ
p
residue eld k = A
p
/m
p
= C residue eld k = A
p
/m
p
= F
p
value f(p), f C[z] value nmodp, n Z
L = /(X), : X

C extension eld L/Q
B = f holomorphic on X
0
=
1
(C) B = integral closure of A in L
P X
0
: (P) = p prime P lying over p
B
P
m
p
= m
e
P
; e = ramication index
k

= B
P
/m
P
= C k

= B
P
/m
P
= F
p
f ; f = residue degree
Table 1. A brief dictionary.
Valuations on

C. A discrete valuation on a eld K is a surjective homo-
morphism v : K

Z such that
v(f +g) minv(f), v(g).
The homomorphism property means
v(fg) = v(f) +v(g).
24
It is also convenient to dene v(0) = +.
Example: for K = C(z) = K(

C), the order of zero (or pole) a rational

function at a given point p

C gives a point valuation v
p
(f).
Perhaps the most important property of a valuation is that:
v(f +g) = min(v(f), v(g)) provided v(f) ,= v(g). (4.2)
For example, adding a constant cannot destroy a pole, while it always de-
stroys a zero. To see (4.2), just note that if v(f) > v(g) then:
v(g) min(v(f +g), v(f)) = v(f +g) min(v(f), v(g)) = v(g).
(Here the rst minimimum canot be v(f) because v(f) > v(g)).
Theorem 4.8 Every valuation on K(

C) is a point valuation.
Proof. Let v : K

Z be a valuation. We claim v vanishes on the constant

subeld C K. Indeed, if f K

has arbitrarily large roots f

1/n
, then
v(f
1/n
) = v(f)/n 0 and thus v(f) = 0.
Next suppose v(z a) > 0 for some a. Then, v(z a) > v(a b) for all
b ,= a. Thus 4.2 implies v(z b) = 0 for all b ,= 0. Since a rational function
is a product of linear terms and their reciprocals, this shows v is a multiple
of the point valuation at a. But v maps surjectively to Z, so we must have
v(z a) = 1 and hence v = v
a
.
Now suppose v(a) 0 for all a. Since v is surjective, v(z a) < 0 for
some a. But then v(z b) = v(z a) for all b, by (4.2) again. It follows
that v(z a) = 1 for all a and hence v = v

gives the order of vanishing

at innity.
Valuations on Riemann surfaces. Using the Riemann existence theo-
rem, it can be shown that the valuations on /(X) correspond to the points
of X, for any compact Riemann surface X. This provides a natural way
to associate X canonically to the abstract eld (C-algebra) /(X). By
pushing valuations forward, we can then associate to a nite eld extension
/(Y ) /(X) a canonical map f : X Y , which turns out to be holo-
morphic. Of course we have already seen how to do this directly, by writing
/(X) = /(Y )[T]/P(T), looking at the discriminant of P(T), etc.
Local rings and elds. For a summary of connection, see Table 1. Ex-
ample: Let L = Q(

D), where D Z is square-free. Then (forgetting the

prime 2) we have ramication over the primes p[D. At these primes we have
B
P

= Z
p
[p
1/2
], just as for Puiseux series. For primes not dividing D, either:
25
(i) T
2
D is irreducible modp, and there is a unique P over p;
or
(ii) D is a square mod p, and there are two primes P
1
and P
2
over p.
In case (i) the prime p behaves like a circle (closed string?) rather than a
point, and P like a double cover of p.
In general when P/p has ramication index e and residue degree f, it can
be thought of roughly as modeled on the map (z, w) (z
e
, w
f
) of S
1
to itself.
5 Holomorphic and harmonic forms
The cotangent space. Let p X be a point on a Riemann surface, and
let m
p
(

p
be the ideal of smooth, complex-valued functions vanishing at
p. Let m
2
p
be the ideal generated by products of element in m
p
.
Theorem 5.1 The ideal m
2
p
consists exactly of the smooth functions all of
whose derivatives vanish at p.
Proof. (Cf. [Hel, p.10]). Clearly the derivatives of fg, f, g m
p
, all vanish.
For the converse, let us work locally at p = 0 C, and suppose f m
0
.
Let f
1
and f
2
denote the partial derivatives of f with respect to x and y.
Then we have:
f(x, y) =
_
1
0
d
dt
f(tx, ty) dt = x
_
1
0
f
1
(tx, ty) dt +y
_
1
0
f
2
(tx, ty) dt
= xg
1
(x, y) +yg
2
(x, y)
where g
i
(0, 0) = f
i
(0, 0). Thus g
1
, g
2
m
p
if both derivatives of f vanish,
which implies f m
2
p
.
Corollary 5.2 The vector space T
(1)
p
= m
p
/m
2
p
is isomorphic to C
2
.
Note: Forster skirts this points (p.60) by dening m
2
p
to consist of the
functions vanishing to order 2.
We refer to T
(1)
p
as the complexied cotangent space of the real surface
X at p. The exterior dierential of a function can then be regarded as the
map dened by
(df)
p
= [f(z) f(p)] m
p
/m
2
p
.
26
The subspace dO
p
is T
(1,0)
p
= T

p
X

= C, the complex (or holomorphic)
tangent space. Its complex conjugate is T
(0,1)
. In a local holomorphic
coordinate z = x +iy, we have:
T
(1)
p
= C dx C dy = C dz C dz = T
(1,0)
p
T
(0,1)
p
.
The dual of T

p
is T
p
, the complex tangent space to X at p.
Exterior dierentials. We have a natural splitting d = + where:
df =
df
dz
dz +
df
dz
dz = f +f,
and
d(f dz +g dz) =
_

df
dz
+
dg
dz
_
dz dz.
(Recall that dz dz = 2i dx dy.) Note that one 1-forms we have:
d(f dz) = (f) dz and d(g dz) = (g) dz.
Holomorphic forms. A function f : X C is holomorphic if f = 0.
A (1, 0)-form on X is holomorphic if the following equivalent conditions
hold:
1. d = 0; i.e. is closed.
2. Locally = df with f holomorphic.
3. Locally = (z) dz with (z) holomorphic.
4. = 0.
The space of all such forms is denoted (X).
Examples: (

C) = 0; (C/) = C dz; dz/z (C

).
Harmonic forms and periods. A function u : X C is harmonic if
u = 0; equivalently, if u is a holomorphic 1-form.
A 1-form on X is harmonic if the following equivalent conditions hold:
1. = = 0 (in particular, is closed);
2. Locally = du with u harmonic;
3. Globally = + with (X) and (X).
27
The space of all such forms is denoted 1
1
(X). By the last condition we
have 1
1
(X) = (X) (X). In particular, every holomorphic 1-form is
harmonic.
Example: dx and dy are harmonic 1-forms on X = C/.
The period map associates to the homomorphism
_

on
1
(X).
Theorem 5.3 Let X be a compact Riemann surface of genus g. Then the
period map
1
1
(X) Hom(
1
(X), C)

= C
2g
is injective.
Proof. Any form in the kernel can be expressed globally as = du where
u is harmonic and hence constant on the compact Riemann surface
X.
Corollary 5.4 The space (X) has dimension g, and the space 1
1
(X)
has dimension 2g.
Example: the torus. For X = C/ we have (X) = C with = dz,
and the period map sends
1
(X)

= Z
2
to .
Example: the regular octagon. Here is concrete example of a compact
Riemann surface X and a holomorphic 1-form (X). Namely let X =
Q/ where Q is a regular octagon in the plane, with vertices at the 8th
roots of unity, and identies opposite edges. Since translations preserve
dz, the form = dz/ is well-dened on X. It has a zero of order two
at the single point p X coming from the vertex. The edges of Q form a
system of generators for
1
(X), and the periods
_

are given by
i+1

i
where is a primitive 8th root of unity.
Meromorphic forms and the residue theorem. These are forms ex-
pressed locally as = (z) dz, with (z) meromorphic.
The residue of at p X is dened by Res
p
() = a
1
if locally (z) =

a
n
z
n
dz, where z(p) = 0; equivalently, by Res
p
() = (2i)
1
_

where
is a small loop around p (inside of which has only one pole).
Theorem 5.5 If X is compact, then

p
Res
p
() = 0.
By considering df/f = d log f, this gives another proof of:
28
Corollary 5.6 If f : X

C is a meromorphic function on a compact
Riemann surface, then f has the same number of zeros as poles (counted
with multiplicity).
(The rst proof was that

f(x)=y
mult(f, x) = deg(f) by general consider-
ations of proper mappings.)
Degree of the canonical bundle. Since the ratio
1
/
2
of any two
meromorphic 1-forms is a meromorphic function (so long as
2
,= 0), we
have:
Theorem 5.7 The degree of a meromorphic 1-form on a compact Rie-
mann surface that is, the dierence between the number of zeros and the
number of poles is independent of the form.
Example: meromorphic 1-forms on

C have degree 2, i.e. they have 2
more poles than zeros. The case of hyperelliptic curves suggests the following
result:
Theorem 5.8 For any compact Riemann surface of genus g, we have dim(X) =
g; and any meromorphic 1-form on X has 2g 2 more zeros than poles.
Proof. This is an easy consequence of the RiemannHurwitz theorem. Let
f : X

C be a meromorphic function of degree d > 0 whose branch locus
does not include . Let = f

(dz). Then over , has a total of 2d

poles. Each critical point of f, on the other hand, creates a simple zero.
Thus
deg() = # cps of f, counted with multiplicities[ 2d = (X) = 2g 2.
Example: hyperelliptic Riemann surfaces.
Theorem 5.9 Let B C be a nite set of cardinality 2n, and let p(T) =

B
(T b). Let : X

C be the unique 2-fold covering of

C branched over
B. Then X has genus g = n 1, and the forms

i
=
z
i
dz
_
p(z)
, i = 0, 1, . . . , n 2
form a basis for (X). In particular dim(X) = g.
29
Proof. To check this, we begin by investigating
0
. Note that if z = w
2
then
dz = 2wdw. Thus the pullback of dz to X has simple zeros at the branch
points of , which lie over the zeros of p(z). Also
_
p(z), as a meromorphic
function on X, has simple zeros in the same locations. These zeros cancel
when we form the quotient dz/
_
p(z), and thus
0
is holomorphic except
possibly over the two unbranched points p
1
, p
2
X lying over z = . But
now dz has a pole of order 2 at z = , while 1/
_
p(z) has a zero of order
n at z = . We conclude:
The form
0
is holomorphic on X, with zeros of order n 2 at
p
1
and p
2
and nowhere else.
In particular, the degree of a meromorphic 1-form on X is 2n 4 = 2g 2.
Since z
i
has a pole of order i at z = , it follows that the g forms
i
on X
are holomorphic for i n 2.
Note: periods. When B = r
1
, . . . , r
2n
R, the periods of
0
can be
expressed in terms of the integrals
_
r
i+1
r
i
dx
_
(x r
1
) (x r
2n
)

We can also allow one point of B to become innity. An example of a period

that can be determined explicitly comes from the square torus:
_
1
0
dx
x(x
2
1)
=
2i

(5/4)
(3/4)

A more general type of period is:
(3) =

n
3
=
_
0<x<y<z<1
dxdy dz
(1 x)yz

See [KZ] for much more on periods.

Appendix: The Hodge star operator. Let V be an n-dimensional real
vector space with an inner product v
1
, v
2
. Choose an orthonormal basis
e
1
, . . . , e
n
for V . Then the wedge products e
I
= e
i
1
e
i
k
provide an
orthonormal basis for
k
V . The Hodge star operator :
k
V
nk
V is
the unique linear map satisfying
e
I
= e
J
where e
I
e
I
= e
1
e
2
e
n
.
30
Here J are the indices not occurring in I, ordered so the second equation
holds. More generally we have:
v w = v, we
1
e
2
e
n
.
and thus v w = w v. Since v v = (1)
k(nk)
v v, we have

2
= (1)
k(nk)
on
k
V . Equivalently,
2
= (1)
k
when n is even, and

2
= 1 when n is odd.
Now let (M, g) be a compact Riemannian manifold. We can then try
to represent each cohomology class by a closed form minimizing
_
M
, .
Formally this minimization property implies:
d = d = 0, (5.1)
using the fact that a minimizer satises
_
M
d, =
_
M
(d) =
_
M
d = 0
for all smooth (k 1)-forms . Thus we call harmonic if (5.1) holds, and
let 1
k
(M) denote the space of all harmonic k-forms.
Theorem 5.10 (Hodge) There is a natural isomorphism 1
k
(M)

= H
k
DR
(M).
Adjoints. The adjoint dierential d

: c
k
(M) c
k1
(M) is dened so
that:
d, = , d

,
where , =
_
M
. It is given by
d

() = d
for a suitable choice of sign, since:
d, =
_
d =
_
d =
_
(d ).
Since
2
= 1 on an odd-dimensional manifold, in that case we have d

=
d. For a k-form on an even dimensional manifold, we have instead:
d

= (1)
k
d .
Here we have used the fact that
2
= (1)
k1
on n (k 1) forms such as
d .
Generalized Hodge theorem. Once the adjoint d

is in play, the argu-

ments of the Hodge theorem give a complete picture of all smooth k-forms
on M.
31
Theorem 5.11 The space of smooth k-forms has an orthogonal splitting:
c
k
(M) = dc
k1
(M) 1
k
(M) d

c
k+1
(M).
The Laplacian. Once we have a metric we can combine d and d

to obtain
the Hodge Laplacian
: c
k
(M) c
k
(M),
dened by
= (dd

+d

d).
Theorem 5.12 A form on a compact manifold is harmonic i = 0.
Proof. Clearly d = d = 0 implies = 0. Conversely if = 0 then:
0 =
_
M
,
=
_
M
d, d +d

, d

and so d = d = 0 as well.
Note that on functions these denitions give
f = d

df = d df,
independent of the dimension of M. This satises
_
f, f 0, but diers
by a sign from the usual Euclidean Laplacian. (For example on S
1
we have
_
ff

=
_
[f

[
2
0 for the usual Laplacian.)
Riemann surfaces. Now suppose M has even dimension n = 2k. The
Hodge star on the middle-dimensional k-forms is then conformally invariant.
Thus it makes sense to talk about harmonic k-forms when only a conformal
structure is present.
In particular, the Hodge star is canonical for 1-forms on a Riemann
surface, and can be expressed by dx = dy and dy = dx for a local
coordinate with z = x +iy. The pair of conditions d = d = 0 are then
the same as the pair of conditions = = 0 for a 1-form.
Geometrically, for to be closed means that the foliation dened by
Ker admits a transverse invariant measure. The orthogonal foliation, de-
ned by , also admits such a measure i is harmonic.
Example: the level sets of Re f(z) and Imf(z) give the foliations associ-
ated to the form = du, u = Re f(z). The case f(z) = z +1/z in particular
32
gives foliations by confocal ellipses and hyperboli, with foci 1 coming from
the critical points of f.
Laplacian on Riemann surfaces. A harmonic function is one which
satises f = 0; equivalently, d df = 0. Since the Hodge star is natural
for 1-forms on a Riemann surface, the harmonic functions are conformally
invariant. For the same reason, harmonic 1-forms are conformally natural:
the 2-form d = 0 is independent of the conformal factor when is a
1-form on a 2-manifold. In local coordinates z = x +iy we have explicitly:
dx = dy and dy = dx.
We also obtain a conformally natural Laplacian sending functions to
2-forms (or measures); it is given by
f = d df = (f
xx
+f
yy
) dx dy
= i( )( +)f = 2if
zz
dz dz.
(Note that f
zz
= (1/4)(f
xx
+ f
yy
) and dz dz = 2i dx dy, so indeed
equality holds.)
On the other hand, the spectrum of the Laplacian on functions a Riemann
surface with a metric depends very much on the metric. This is because
we must divide by the volume form to get a map sending functions to
functions.
Since it is natural to take functions and forms on a Riemann surface to
be complex valued, the complexied Hodge star includes composition with
complex conjugation. This insures, for example, that
_
M
f f =
_
M
[f[
2
dV.
With this convention, (i) = i , and thus
dz = (dx +i dy) = dy +i dx = idz,
and dz = i dz.
6 Cohomology of sheaves
Maps of sheaves; exact sequences. A map between sheaves is always
specied at the level of open sets, by a family of compatible morphisms
T(U) ((U). A map of sheaves induces maps T
p
(
p
between stalks.
33
We say T ( is injective, surjective, an isomorphism, etc. i T
p
(
p
has
the same property for each point p.
We say a sequence of sheaves / B ( is exact at B if the sequence
of groups
/
p
B
p
(
p
is exact, for every p.
The exponential sequence. As a prime example: on any Riemann surface
X, the sequence of sheaves
0 Z O O

0
is exact. But it is only exact on the level of stalks! For every open set the
sequence
0 Z(U) O(U) O

(U)
is exact, but the nal arrow need not be surjective. (Consider f(z) = z
O

(C

); it cannot be written in the form f(z) = exp(g(z)) with g O(C

).)
More generally, we have:
Theorem 6.1 The global section functor is left exact. That is, for any
short exact sequence of sheaves, 0 / B ( 0, the sequence of
global sections
0 /(U) B(U) ((U)
is also exact.
Proof. Most of what needs to be veried follows from the fact that we have
a commutative diagram
0
//
/(U)
//

B(U)
//

((U)

0
//

/
p
//

B
p
//

(
p
where the bottom row is exactly and the vertical arrows are injective. The
one step remaining is to show that if b B(U) maps to zero in ((U), then
it is in the image of /(U). This is true on the level of stalks: for every
p U there is a neighborhood U
p
of p and a unique a
p
/(U
p
) such that
a
p
= b[U
p
. But then a
p
a
q
= 0 on U
pq
so these local solutions piece together
to show b /(U).
34
Sheaf cohomology is the derived functor which measures the failure of
exactness to hold on the right.

Cech: the nerve of a covering. A precursor to sheaf cohomology is

Cech
cohomology. The idea here is that any open covering U = (U
i
) of X has an
associated simplicial complex that is an approximation to the topology of X.
The simplices in this complex are simply ordered nite sequence of indices
I such that

I
U
i
,= .
This works especially well if we require that all the multiple intersections
are connected. Note that this is equivalent to requiring that Z(U
I
)

= Z
whenever U
I
,= .
Cochains, cocycles and coboundaries. Now suppose we also have a
sheaf T in play (for classical

Cech cohomology, this sheaf is just Z). We
then put weights on our simplices and dene the space of q-cochains by:
C
q
(U, T) =

|I|=q+1
T(U
I
).
Here I ranges over ordered sets of indices (i
0
, . . . , i
q
), and
U
I
= U
i
0
U
iq
.
Examples: a 0-cochain is the data f
i
T(U
i
); a 0-cochain is the data
g
ij
T(U
i
U
j
); etc.
Next we dene a boundary operator
: C
q
(U, T) C
q+1
(U, T)
by setting f = g where, for q = 0:
g
ij
= f
j
f
i
;
for q = 1:
g
ijk
= f
jk
f
ik
+f
ij
,
and more generally
g
I
=
q

0
(1)
j
f
I
j
where I
j
= (i
0
, i
1
, . . . ,

i
j
, . . . , i
q+1
). When two indices are eliminated, they
come with opposite sign, so
2
= 0.
35
The kernel of is the group of cocycles Z
q
(U, T), its image is the group
of coboundaries B
q
(U, T), and the qth cohomology group of T relative to
the covering U is dened by:
H
q
(U, T) = Z
q
(U, T)/B
q
(U, T).
Example: H
0
. A 0-cocycle (f
i
) is a coboundary i f
j
f
i
= g
ij
= 0 for
all i and j. By the sheaf axioms, this happens i f
i
= f[U
i
, and thus:
H
0
(U, T) = T(X).
Example: H
1
. A 1-cocycle g
ij
satises g
ii
= 0, g
ij
= g
ji
and
g
ij
+g
jk
= g
ik
on U
ijk
. It is a coboundary if it can be written in the form g
ij
= f
i
f
j
.
Example on S
1
. Let T = Z. Suppose we cover S
1
with n 3 intervals
U
1
, . . . , U
n
such that U
ij
is an interval when i and j are consecutive, and
otherwise U
ij
is empty. Then the space of 1-cocycles in simply Z
n
, but
g
ij
= f
i
f
j
i

g
i,i+1
= 0. Thus H
1
(U, Z) = Z.
Renement. Whenever V = (V
i
) is a ner covering than U = (U
i
), we can
choose a renement map on indices such that V
i
U
i
. Once is specied,
it determines maps T(U
I
) T(V
I
), and hence chain maps giving rise to a
homomorphism
H
q
(U, T) H
q
(V, T).
Theorem 6.2 The renement map H
q
(U, T) H
q
(V, T) is independent
of .
Denition. We dene the cohomology of X with coecients in T by:
H
q
(X, T) = lim

H
q
(U, T),
where the limit is taken over the system of all open coverings, directed by
renement.
Theorem 6.3 The renement map H
1
(U, T) H
1
(V, T) is injective.
Proof. Suppose we are given coverings (U
i
) and (V
i
) with V
i
U
i
. Let
g
ij
be a 1-cocycle for the covering (U
i
) that becomes trivial for (V
i
). That
means there exist f
i
T(V
i
) such that
g
i,j
= f
i
f
j
36
on V
ij
.
Our goal is to nd h
i
T(U
i
) so g
ij
= h
i
h
j
. Note that on V
ij
U
k
we have:
g
i,k
+g
k,j
= f
i
f
j
,
and thus we can dene
h
k
= f
i
g
i,k
= f
j
g
j,k
consistently throughout U
k
. We then have, on U
kl
V
i
,
h
k
h
l
= f
i
g
i,k
f
i
+g
i,l
= g
kl
as desired.
Failure for q 2. Renement is not injective for q 2 in general. For
example, let X be the 1-skeleton of a 3-simplex and let | = (U
1
, . . . , U
4
) an
open covering obtained by thickening each of its triangular faces. Then the
nerve of | is a simplicial 2-sphere and hence H
2
(X, |; Z) = Z even though
H
2
(X; Z) = 0.
Note that | has a renement 1 with no 3-fold intersections.
Theorem 6.4 (Leray) If U is acyclic (H
q
(U
I
, T) = 0 for all q > 0) then
H
q
(X, T)

= H
q
(U, T).
If just H
1
(U
i
, T) = 0 for every i, U can still be used to compute H
1
.
Vanishing theorem by dimension. Using the existence of ne coverings
where the (n + 2)-fold intersections are all empty, we have:
Theorem 6.5 For any n-dimensional space X and any sheaf T, H
p
(X, T) =
0 for all p > n.
Vanishing theorems for smooth functions, forms, etc. Let T be the
sheaf of C

functions on a (paracompact) manifold X, or more generally a

sheaf of modules over C

. We then have:
Theorem 6.6 The cohomology groups H
q
(X, T) = 0 for all q > 0.
37
Proof for q = 1. To indicate the argument, we will show H
1
(U, T) = 0
for any open covering U = (U
i
). We will use the fact that there exists a
partition of unity
i
C

(X) subordinate to U
i
: that is, a set of functions
with K
i
= supp
i
U
i
, such that K
i
forms a locally nite covering of X
and

i
(x) = 1 for all x X.
Let g
ij
Z
1
(U, T) be a 1-cocycle. Then g
ii
= 0, g
ij
= g
ji
and g
ij
+
g
jk
= g
ik
. Our goal is to write g
ij
= f
j
f
i
(or f
i
f
j
).
How will we ever get from g
ij
, which is only dened on U
ij
, a function
f
i
dene on all of U
i
? The central observation is that:

j
g
ij
, extended by 0, is smooth on U
i
.
This is because supp
j
g
ij
K
j
U
i
is closed as a subset of U
i
(even though
it might not be closed as a subset of X.) Thus we can dene:
f
i
=

k
g
ik
;
and then:
f
i
f
j
=

k
(g
ik
g
jk
) =

k
(g
ik
+g
kj
) =

k
g
ij
= g
ij
.
Proof for q = 2. Let f
ab
=

x
g
abx
. Then
(f)
abc
= f
bc
f
ac
+f
ab
=

x
(g
bcx
g
acx
+g
abx
) =

x
g
abc
= g
abc
.
The exact cohomology sequence; deRham cohomology. We can now
explain how sheaf cohomology is used to capture global aspects of analytic
problems that can be solved locally.
Let c
p
denote the sheaf of smooth p-forms on a manifold X. Suppose
c
1
(X) is closed; then locally = df for f c
0
(X). When we can we
nd a global primitive for ?
To solve this problem, let U = (U
i
) be an open covering of X by disks.
Then we can write = df
i
on U
i
. On the overlaps, g
ij
= f
i
f
j
satises
dg
ij
= 0, i.e. it is a constant function. Moreover we obviously have g
ij
+g
jk
=
g
ik
, i.e. g
ij
is an element of Z
1
(U, C).
38
Now we may have chosen our f
i
wrong to t together, since f
i
is not
uniquely determined by the condition df
i
=
i
; we can always add a constant
function c
i
. But if we replace f
i
by f
i
+ c
i
, then g
ij
will change by the
coboundary c
i
c
j
. Thus we can conclude:
= df i [g
ij
] = 0 in H
1
(X, C) .
The exact cohomology sequence. The conceptual theorem underlying
the preceding discussion is the following:
Theorem 6.7 Any short exact sequence of sheaves on X,
0 / B ( 0,
gives rise to a long exact sequence
0 H
0
(X, /) H
0
(X, B) H
0
(X, ()
H
1
(X, /) H
1
(X, B) H
1
(X, ()
H
2
(X, /) H
2
(X, B) H
2
(X, ()
on the level of cohomology.
Note: for any open set U, we get an exact sequence
0 /(U) B(U) ((U), (6.1)
as can be checked using the sheaf axioms. Surjectivity of the maps B
x
(
x
implies that for any c ((X), there is an open covering (U
i
) and b
i
B(U
i
)
mapping to c.
To obtain the connecting homomorphism

: H
0
(X, () H
1
(X, /),
we use the exactness of (6.1) to write b
i
b
j
as the image of a
ij
. The resulting
cocycle [a
ij
] H
1
(X, /) is the image

(c).
Sheaves and deRham cohomology. Let :
p
denote the sheaf of closed
p-forms. Recall that the deRham cohomology groups of X are given by:
H
p
DR
(X) = :
p
(X)/dc
p1
(X).
The Poincare Lemma asserts that
H
p
DR
(R
n
) = 0
39
for all p and n.
Proof. The cases p = 0 and p = 1 are easy, since R
n
is connected and
simplyconected. The general case is proved by induction on n. We can
regard R
n+1
as a bundle over R
n
with projection and zero section s. The
idea is to construct an operator K on forms on R
n+1
such that
I

= (Kd dK).
The existence of such an operator implies any closed form on R
n+1
is
cohomologous to one pulled back from R
n
. Since closed forms on R
n
are
exact, the proof is complete.
The operator K is constructed as follows. Let (t, x) denote coordinates
on RR
n
, and let denote a product of dx
1
, . . . , dx
n
. Then K(f(t, x)) = 0,
while
K(f(t, x) dt ) = F(t, x),
where F(t, x) =
_
t
0
f(t, x) dt.
The verication that K gives a chain homotopy is straightforward. For
example, dK +Kd should be the identity on a form like f dt. And indeed,
dK(f dt, dt) = d(F) = f dt + (d
x
F),
while
Kd(f dt) = K((d
x
f) dt = (d
x
F).
The sign enters because we must move dt in front of d
x
f. On a form like
f(x, t) we should get (f(t, x) f(0, x)), and indeed dK kills this form
while
Kd(f(t, x)) = K(f

(t, x) dt + (d
x
f)) = (f(t, x) f(0, x)).
The deRham theorem. Since every manifold looks locally like R
n
, the
Poincare Lemma implies we have an exact sequence of sheaves:
0 :
p1
c
p1
d
:
p
0.
Now let p = 1. Then :
p1
= C. By examining the associate long exact
sequence we nd:
Theorem 6.8 For any manifold X, we have H
1
DR
(X)

= H
1
(X, C).
40
More generally, using all the terms in the exact sequence and all values
of p, we nd:
Theorem 6.9 For any manifold X, we have
H
p
DR
(X)

= H
1
(X, :
p1
)

= H
2
(X, :
p2
)

= H
p
(X, :
0
) = H
p
(X, C).
Corollary 6.10 The deRham cohomology groups of homeomorphic smooth
manifolds are isomorphic.
(In fact one can replace homeomorphic by homotopy equivalent here.)
Finiteness. Now it is easy to prove that H
p
DR
(R
n
) = 0 for all p > 0. We
thus have, by taking a Leray covering:
Theorem 6.11 For any compact manifold, the cohomology groups H
p
(X, C)
are nite.
Periods revisited. Finally we mention the fundamental group:
Theorem 6.12 For any connected manifold X, we have
H
1
(X, C)

= H
1
DR
(X)

= Hom(
1
(X), C).
Proof. We have
_

df = 0 for all closed loops , so the period map is

well-dened; if has zero periods then f(q) =
_
q
p
is also well-dened and
satises df = . To prove surjectivity, take a (Leray) covering of X by
geodesically convex sets, and observe that (i) every element of
1
(X) is rep-
resented by a 1-chain and (ii) every 1-boundary is a product of commutators.
Remarks. One can dene the period map H
1
(X, C) Hom(
1
(X), C)
directly, using : S
1
X to obtain from H
1
(X, C) a class ()
H
1
(S
1
, C)

= C. For more exotic topological spaces, however, this map need
not be an isomorphism: e.g. the topologists sine curve is a compact,
connected space X with
1
(X) = 0 but H
1
(X, Z) = Z.
41
7 Cohomology on a Riemann surface
On a Riemann surface we have the notion of holomorphic functions and
forms. Thus in addition to the sheaves c
p
we have the important sheaves:
O the sheaf of holomorphic functions; and
the sheaf of holomorphic 1-forms.
Let h
i
(T) = dimH
i
(X, T). We will show that a compact Riemann surface
satises:
h
0
(O) = 1, h
0
() = g,
h
1
(O) = g, h
1
() = 1.
The symmetry of this table is not accidental: it is rather the rst instance
of Serre duality, which we will also prove.
We remark that a higher cohomology group, like H
1
(X, O), is a sort of
place holder, like paper money. It has no actual worth until it is exchanged
for something else. Serre duality makes such an exchange possible.
The Dolbeault Lemma. Just as the closed forms can be regarded as
the subsheaf Ker d c
p
, the holomorphic functions can be regarded as the
subsheaf Ker (

= c
0
. So we must begin by studying the operator.
We begin by studying the equation df/dz = g L
1
(C). An example is
given for each r > 0 by:
f
r
(z) =
_
1/z if [z[ > r,
z/r
2
if [z[ r,
which satises g
r
= df/dz = (1/r
2
)
B(0,r)
(z). In particular
_
g
r
= is
independent of r, which suggests the distributional equation:
d
dz
1
z
=
0
.
Using this fundamental solution (and the fact that dxdy = (i/2)dz dz),
we obtain:
Theorem 7.1 For any g C

c
(C), a solution to the equation df/dz = g is
given by:
f(z) = g
1
z
=
i
2
_
C
g(w)
z w
dw dw.
42
Proof. It is enough to check df/dz at z = 0, where we nd:
2i
df
dz
(0) =
_
C
dg
dz
_
1
z
_
dz dz =
_
C
d
_
g(z) dz
z
_
= lim
r0
_
CB(0,r)
d
_
g(z) dz
z
_
= lim
r0
_
S
1
(0,r)
g(z)
dz
z
= 2i g(0).
The penultimate minus sign comes from the back that S
1
(0, r) acquires a
negative orientation as (C B(0, r)).
Theorem 7.2 For any g (

(), there is an f C

() with df/dz = g.
Proof. Write g =

g
n
where each g
n
is smooth and compactly supported
outside the disk D
n
of radius 1 1/n. Solve df
n
/dz = g
n
. Then f
n
is
holomorphic on D
n
. Expanding it in power series, we can nd holomorphic
functions h
n
on the disk such that [f
n
h
n
[ < 2
n
n D
n
. Then f =

(f
n
h
n
) = limF
n
converges uniformly, and F F
i
is holomorphic on D
n
for all i > n, so the convergence is also C

.
Corollary 7.3 For any g (

(), there is an f C

() with f = g.
Proof. First solve dh/dz = g, then df/dz = h. Then (f/4) = d
2
f/dz dz =
g.
Remarks. The same results hold with replaced by C. It is easy to solve
the equation df/dz = g on C when g(z, z) is a polynomial: simply formally
integrate with respect to the z variable.
In higher dimensions, = g should be thought of as a (0, 1)-form, and
this equation can be solved for g only when = 0.
Dolbeault cohomology. Let us dene, for any Riemann surface X,
H
0,1
(X) =
c
0,1
(X)
c
0
(X)

The preceding results show the sequence of sheaves:

0 O c
0

c
0,1
0
is exact. Consequently we have:
43
Theorem 7.4 On any Riemann surface X, we have H
1
(X, O)

= H
0,1
(X).
Thus the Dolbeault lemma can be reformulated as:
Theorem 7.5 The unit disk satises H
1
(, O)

= H
0,1
() = 0. The same
is true for the complex plane.
Corollary 7.6 We have H
1
(

C, O) = 0.
Proof. Let U
1
U
2
be the usual covering by U
1
= C and U
2
=

C 0.
By the preceding result, H
1
(U
i
, O) = 0. Thus by Lerays theorem, this
covering is sucient for computing H
1
: H
1
(

C, O) = H
1
(U, O). Suppose
g
12
O(U
12
) = O(C

) is given. Then g
12
(z) =

a
n
z
n
. Splitting this
Laurent series into its positive and negative parts, we obtain f
i
O(U
i
)
such that g
12
= f
2
f
1
.
Similarly, we dene
H
1,1
(X) =
c
1,1
(X)
c
1,0
(X)

Theorem 7.7 On any Riemann surface X we have H

1
(X, )

= H
1,1
(X).
Proof. Consider the exact sequence of sheaves
0 c
1,0

c
1,1
0.
Corollary 7.8 The dimension of H
1
(X, ) is 1.
Proof. The map
_
X
on H
1,1
(X) has image C.
We will later see that equality holds.
The Dolbeault isomorphism. Using the fact that
2
= 0 one can dened
the Dolbeault cohomology groups for general complex manifolds, and prove
using sheaf theory the following variant of the deRham theorem:
Theorem 7.9 For any compact complex manifold X, we have H
p,q

(X)

=
H
q
(X,
p
).
44
Here
q
is the sheaf of holomorphic (q, 0)-forms.
Finiteness. Recall that we already know dim(X) g, by period con-
siderations; in particular, (X) is nite-dimensional. We remark that a
more robust proof of this nite-dimensionality can be given by endowing
(X) with a reasonable normal e.g. ||
2
=
_
X
[[
2
and then ob-
serving that the unit ball is compact. (The same proof applies to show
dimO(X) < , without using the maximum principle. More generally the
space of holomorphic sections of a complex line bundle over a compact space
is nite-dimensional.)
Serre duality, special case. We can now use this niteness to show
niteness of cohomology groups. (An alternative proof, again based on
norms, is given in Forster.)
Theorem 7.10 On any compact Riemann surface X, we have H
1
(X, O)

=
(X). In particular, H
1
(X, O)

is nite-dimensional.
We let g
a
= dim(X), the arithmetic genus of X. We will eventually
see that g
a
= g = the topological genus.
Proof. By Dolbeault we have
H
1
(X, O)

= H
0,1
(X) = c
(0,1)
(X)/c
0
(X).
We claim c
0
(X) is closed in c
(0,1)
(X) in the C

topology. If not, there

is a sequence f
n
c
0
(X) with f
n
in the C

topology, such that f

n
has no convergent subsequence in c
0
(X). Since bounded sets in c
0
(X) are
compact, the latter condition implies for some k 0, |f
n
|
C
k . From
this we will obtain a contradiction.
Dividing through by the C
k
-norm, we can arrange that |f
n
|
C
k = 1 in
c
0
/C and f
n
0. Taking a bump function in a chart, we have
(f
n
) = (f
n
) + ()f
n
.
Since f
n
is bounded in C
k
and f
n
tends to zero, the right-hand side is
bounded in C
k
. But since the solution to the equation is given by con-
volution with 1/z, a smoothing operator, we nd that f
n
is precompact
in C
k
. Thus we can pass to a C
k
convergent subsequence, f
n
g. Then
g = 0 so g is constant. But then |f
n
|
C
k 0 in c
0
(X)/C, a contradiction.
Since c
0
(X) is closed, we have
H
1
(X, O)

= W

= (c
0
(X))

(c
0,1
(X))

= T
1,0
(X).
But any (1, 0)-current W satises
_
(f) = 0 for all smooth f, and
thus = 0, which implies is holomorphic and thus W = (X).
45
Corollary 7.11 We have natural isomorphisms H
1,0
(X)

= (X) and H
0,1
(X)

=
(X).
Proof. By denition, H
1,0
(X) consists of -closed (1, 0) forms, which means
it is isomorphic to (X).
As for H
0,1
(X), we have a natural map (X) H
0,1
(X). On this
subspace, the pairing with (X) given by
_
is obviously perfect, since
(i/2)
_
> 0. Thus is isomorphic to H
0,1
(X).
Taking stock. In preparation for Riemann-Roch, we now know H
1
(X, O)

=
(X), g
a
= dim(X) g, and dimH
1
(X, ) 1 because of the residue
map.
8 Riemann-Roch
One of the most basic questions about a compact Riemann surface X is:
does there exist a nonconstant holomorphic map f : X

C? But as we
have seen already in the discussion of the eld /(X), it is desirable to ask
more: for example, do the meromorphic functions on X separate points?
Even better, does there exist a holomorphic embedding X P
n
for some
n?
To answer these questions we aim to determine the dimension of the
space of meromorphic functions with controlled zeros and poles. This is the
Riemann-Roch problem.
Divisors. Let X be a compact Riemann surface. The group of divisors
Div(X) is the free abelian group generated by the points of X. A divisor is
sum D =

a
P
P, where a
P
Z and a
P
= 0 for all but nitely many P X.
We say D 0 if a
P
0 for all P. A divisor is eective if D 0. Any divisor
can be written unique as a dierence of eective divisors, D = D
+
D

.
The degree of a divisor, deg(D) =

a
P
, denes a homomorphism deg :
Div(X) Z.
Note: one can also dene a sheaf by Div(U) =

a
P
P where the
sum is locally nite. This is the sheaf canonically generated by the presheaf
/

/O

.
Associated to any meromorphic function f ,= 0 is a divisor of degree zero
recording its zeros and poles:
(f) =

P
ord(f, P) P.
46
The divisors that arise in this way are said to be principal. Note that:
(fg) = (f) + (g),
so (f) denes a homomorphism from /

(X) into Div

0
(X). Note also that
the topological degree of f is given by deg(f)
+
.
The sheaf O
D
consists of meromorphic functions f such that (f)+D 0.
For example, O
nP
(X) is the vector space of meromorphic functions on X
with poles of order at most n at p.
Isomorphisms between sheaves. If D E = (f) is principal, we say D
and E are linearly equivalent. Then the map h hf gives an isomorphism
of sheaves:
O
D
= O
E+(f)

= O
E
,
because
(hf) +E = (h) + (f) +E = (h) +D.
The divisor K = () of a nonzero meromorphic 1-form is dened similarly.
Once we have such a canonical divisor, we get an isomorphism
O
K

=
by h h, because h is holomorphic i
(h) = (h) +K 0
i h O
K
.
Riemann-Roch Problem. The Riemann-Roch problem is to calculate or
estimate
h
0
(D) = dimH
0
(X, O
D
) = dimO
D
(X).
Example: if X is a complex torus, we have h
0
(nP) = 1, 1, 2, 3, . . .. This can
be explained by the fact that Res
P
(f dz) = 0.
It is a general principle that Euler characteristics are more stable than
individual cohomology groups, and we dene:
(T) =

(1)
q
h
q
(T).
We then have:
Theorem 8.1 (Riemann-Roch, Euler characteristic version) For any
divisor D, we have
(O
D
) = h
0
(O
D
) h
1
(O
D
) = deg D g
a
+ 1.
47
Here g
a
= dim(X) is the arithmetic genus. (Note: this bound is not
optimal. The optimal bound, coming from BrillNoether theory, is given by
(g + 3)/2 rounded down. We will later see this bound explicitly for genus
g 3.)
For the proof we will use:
Theorem 8.2 If 0 / B ( 0 is an exact sequence of sheaves with
nite-dimensional cohomology groups on a nite-dimensional space X, then
we have:
(B) = (A) +(C).
Proof. Let the homomorphisms in dimension p for the resulting long exact
sequence be denoted by
p
,
p
and
p
. We then have:
(/) =

(1)
p
(dimKer
p
+ dimIm
p
),
(B) =

(1)
p
(dimKer
p
dimIm
p
), and
(() =

(1)
p
(dimKer
p
+ dimIm
p
).
By exactness we have:
0 =

(1)
p
(dimIm
p
dimKer
p
),
0 =

(1)
p
(dimIm
p
+ dimKer
p
), and
0 =

(1)
p
(dimIm
p
dimKer
p+1
),
and thus (/) (B) +(() = 0.
Skyscrapers. The skyscraper sheaf C
P
is given by O(U) = C if P U,
and O(U) = 0 otherwise. For any divisor D, we have the exact sequence:
0 O
D
O
D+P
C
P
0, (8.1)
where the nal map records the leading coecient of the polar part of f at
P. It is easy to see (e.g. by taking ne enough coverings, without P in any
multiple intersections):
Theorem 8.3 We have H
p
(X, C
P
) = 0 for all p > 0. In particular (C
P
) =
h
0
(C
P
) = 1.
48
Theorem 8.4 The cohomology groups H
p
(X, O
D
) are nite-dimensional
for p = 0, 1 and vanish for all p 2.
Proof. We have already seen the result is true for D = 0: the space
H
1
(X, O)

= (X)

is nite-dimensional, and using the Dolbeault sequence,

one can show H
p
(X, O) = 0 for all p 2. The result for general D then
follows induction, using the skyscraper sheaf.
Proof of Riemann-Roch. Since h
1
(O) = dim(X) = g
a
, the formula
is correct for the trivial divisor. Using (8.1) and additivity of the Euler
characteristic short exact sequences, we nd:
(O
D+P
) = (O
D
) +(C
P
) = (O
D
) + 1,
which implies Riemann-Roch for an arbitrary divisor D.
Existence of meromorphic functions and forms. A more general form
of Serre duality will lead to a more useful formulation of Riemann-Roch, but
we can already deduce several useful consequences.
Theorem 8.5 Any compact Riemann surface admits a nonconstant map
f : X P
1
with deg(f) g
a
+ 1.
Proof. We have dimH
0
(X, O
D
) deg D g
a
+ 1. Once deg D > g
a
this
gives dimH
0
(O
D
) > 1 = dimC.
Corollary 8.6 Any surface with g
a
= 0 is isomorphic to P
1
.
Proof. It admits a map to P
1
of degree g
a
+ 1 = 1.
Corollary 8.7 Any compact Riemann carries a nonzero meromorphic 1-
form.
Proof. Take = df.
Corollary 8.8 Canonical divisors K = () exist, and satisfy O
K

= .
The isomorphism is given by f f.
Arithmetic and topological genus; degree of canonical divisors;
residues.
Corollary 8.9 The degree of any canonical divisor is 2g 2, where g is the
topological genus of X.
Proof. Apply Riemann-Hurwitz to compute deg(df).
49
Corollary 8.10 The topological and arithmetic genus of X agree: we have
g = g
a
; and dimH
1
(X, ) = 1.
Proof. Apply Riemann-Roch to a canonical divisor K: we get
h
0
(K) h
1
(K) = g
a
h
1
(K) = 1 g
a
+ deg(K) = 2g g
a
1,
or in other words:
2g
a
h
1
(K) = 2g 1.
Now we know g
a
g and h
1
(K) 1, so the only way equality can hold is if
g
a
= g and h
1
(K) = 1.
For equality to hold in the calculation above we need h
1
(K) = 1, i.e.:
Corollary 8.11 We have dim(H
1
(X, )) = 1.
Theorem 8.12 (Hodge theorem) On a compact Riemann surface every
class in H
1
DR
(X, C) is represented by a harmonic 1-form. More precisely we
have
H
1
DR
(X) = (X) (X) = H
1,0
(X) H
0,1
(X) = 1
1
(X).
Proof. We already know the harmonic forms inject into deRham cohomol-
ogy, by considering their periods; since the topological and arithmetic genus
agree, they also surject.
The space of smooth 1-forms. On a compact Riemann surface, the full
Hodge theorem
c
1
(X) = dc
0
(X) 1
1
(X) d

c
2
(X)
becomes the statement:
c
1
(X) = ( +)c
0
(X) ((X) (X)) ( )c
0
(X).
We have now proved this statement. Indeed, it suces to show c
0
(X)
c
0
(X) spans the complement of the harmonic forms. But the isomorphism
H
1,0
(X)

= (X) gives c
1,0
= (X) c
0
(X), and similarly for c
0,1
.
Theorem 8.13 A smooth (1, 1) form lies in the image of the Laplacian
i
_
X
= 0.
Proof. We must show = f. Since the residue map on H
1,1
(X) is
an isomorphism, we can write = where is a (1, 0)-form. Since
(X)

= H
0,1
(X), we can nd a holomorphic 1-form such that = f
for some smooth function f. Then f = , and hence f = = .
50
Remark: isothermal coordinates. The argument we have just given also
proves the Hodge theorem for any compact, oriented Riemannian 2-manifold
(X, g). To see this, however, we need to know that every Riemannian metric
is locally conformally at; i.e. that one can introduce isothermal coordi-
nates to make X into a Riemann surface with g a conformal metric.
9 The MittagLeer problems
In this section we show that the residue theorem is the only obstruction to
the construction of a meromorphic 1-form with prescribed principal parts.
The residue map on H
1
(X, ) will also play a useful role in the discussion
of Serre duality.
We then solve the traditional Mittag-Leer theorem for meromorphic
functions, and use it to prove the RiemannRoch theorem for eective divi-
sors.
The residue map. Since d = on c
1,0
, the residue map
Res() =
1
2i
_
X

gives a canonical map

Res : H
1
(X, )

= H
1,1
(X) C.
(We will later see that this map is an isomorphism.)
Now suppose we have an element = [
ij
] H
1
(X, ) that can be
expressed in the form
ij
=
i

j
, with
i
/
(1)
(U
i
). We then refer to
(
i
) as Mittag-Leer data, and as a Mittag-Leer coboundary. We can
then associate to quantity
R() =

p
Res
p
(
i
).
Note that if
ij
=
i

j
then there exists a global meromorphic form given
locally by =
i

i
. Since

Res
p
() = 0, the value of R() is well-dened.
Example. For any point P X we can choose a local coordinate z : U
1

=
, set U
2
= X P, and dene
1
= dz/z and
2
= 0. Then
12
= dz/z
on U
12
, and = [
ij
] satises R() = 1.
Theorem 9.1 If = [
i

j
] is a Mittag-Leer coboundary, then
Res() =

Res
p
(
i
).
51
Proof. Using the fact that H
1
(X, c
1,0
) = 0, we can write the cocycle
= (
ij
) as a coboundary in two ways: we have

ij
=
i

j
= g
i
g
j
with g
i
c
1,0
(U
i
). Then = dg
i
is globally well-dened and Res() =
_
X
.
Now let X

X be the complement of very small disks around the

poles of the
i
. Then on X

, we have a global smooth 1-form h = g

i

i
on U
i
(these denitions agree on the overlap by the above). Since
i
is
holomorphic, we have dh = dg
i
= , and thus:
2i Res() =
_
X

_
X

dh =
_
X

h 2i

Res
p
(
i
).
In the last step we have taken into account the fact that X

is oriented
to give loops that go negatively around the points p, and that
_
X

g
i
0
since g
i
is smooth at p.
Mittag-Leer for 1-forms. The Mittag-Leer problem is to construct
meromorphic 1-form with prescribed principal parts.
The problem does not always have a solution. For example, if just a
single simple pole is specied, then X would have to have genus 0 for f to
exist.
The data is conveniently given by
i
/
(1)
(U
i
) such that = ()
ij
lies in Z
1
(), i.e. such that
ij
=
j

i
is holomorphic. Then the problem
of constructing a global with the same principal parts reduces to showing
that () is a coboundary in H
1
(X, ).
We are thus lead to consider that cohomology group. Since we have
shown that h
1
(K) = 1 (above), we have:
Theorem 9.2 There exists a meromorphic 1-form with prescribed principal
parts i the sum of its residues is equal to zero.
Proof. The principal parts are specied by a Mittag-Leer 0-cochain
i

/
(1)
(U
i
) with boundary

j
=
ij
(U
ij
);
and the class = [
ij
] H
1
(X, ) is equal to zero i Res() = 0, in
which case
ij
=
i

j
with
i
(U
i
), and
i
+
i
= denes a global
meromorphic 1-form with the prescribed principal parts.
52
Corollary 9.3 The Mittag-Leer problem for 1-forms has a solution if and
only if the sum of the residues of the principal parts is zero.
Corollary 9.4 Given any pair of distinct points p
1
, p
2
X, there is a mero-
morphic 1-form with simple poles of residues (1)
i
at p
i
and no other
singularities.
This elementary dierential of the third kind is unique up to the addi-
tion of a global holomorphic dierential. Example: the form dz/z works for
0,

C.
Corollary 9.5 For any p X and n 2 there exists a meromorphic 1-
form with a pole of order n at p (but vanishing residue) and no other
singularities.
This is an elementary dierential of the second kind.
Currents and residues. Here is another formulation of the Mittag-Leer
problem for 1-forms. Using sheaves of distributions and currents, we have
an exact sequence
0 T
1,0

T
1,1
0,
which shows the isomorphism H
1
(X, )

= C can be computed using currents
instead of smooth forms. Then a current in T
1,1
representing the Mittag-
Leer cocycle
i

j
is given by = d
i
, which is supported just at the
poles of
i
and satises
1
2i
_
=

p
Res
p
(
i
).
Thus
i
= 0 i Res(
i
) = 0.
Mittag-Leer for functions. Given a nite set of points p
i
X, and
the Laurent tails
f
i
(z) =
b
n
z
n
+ +
b
1
z
of meromorphic functions f
i
in local coordinates near P
i
, when can we nd
a global meromorphic function f on X with the given principal parts?
suppose we
Equivalently, given f
i
/(U
i
), we want to determine when f
i

H
1
(X, O) is a coboundary. By the case of Serre duality we have already
proven, H
1
(X, O) is isomorphic to (X)

, so there is a natural pairing be-

tween (f
i
) H
1
(X, O) and (X).
53
Recalling from Theorem 9.1 that (up to a constant) the pairing dening
Serre duality is given by
f
i
, = Res(f
i
),
we then have:
Corollary 9.6 The Mittag-Leer problem specied by (f
i
) has a solution
i

Res
p
(f
i
) = 0
for every (X).
This proves the geometric Riemann-Roch theorem (in which h
1
(D) is
replaced by h
0
(KD)) for eective divisors. For example, suppose D = nP.
Then an element f O
D
(X) is determined by its Laurent tail
f(z) =
b
n
z
n
+ +
b
1
z
+b
0
in a coordinate system where z(p) = 0. The set of (b
i
) that can arise is
determined by the constraint Res(f) = 0 for all (X). But the
residue vanishes trivially if belongs to
nP
(X). Thus the number of
constraints on (b
0
, . . . , b
n
) is dim(X) dim
nP
(X) = g h
0
(K nP),
and we nd
h
0
(nP) = n + 1 g +h
0
(K nP)
in agreement with Riemann-Roch.
10 Serre duality
Let /
(1)
(X) denote the space of meromorphic 1-forms on X. It is a 1-
dimensional vector space over the eld /(X). We dene:

D
(X) = /
(1)
(X) : () +D 0.
The sheaf
D

= O
D+K
is dened similarly. The goal of this section is to
present:
Theorem 10.1 (Serre duality) For any divisor D, we have a canonical
isomorphism
H
1
(X, O
D
)

=
D
(X).
54
This result allows us to eliminate h
1
entirely from the statement of
Riemann-Roch, and obtain:
Theorem 10.2 (Riemann-Roch, geometric version) For any divisor D
on a compact Riemann surface X, we have
h
0
(D) = deg D g + 1 +h
0
(K D).
Proof of Serre duality: analysis. A proof of Serre duality can be given
using the elliptic regularity, just as we did for the case (X)

= H
1
(X, O)

.
For this one should either consider forms that are locally (smooth) + (mero-
morphic), or consider smooth sections of the line bundle dened by a divisor,
which we will elaborate later.
Proof of Serre duality: dimension counts. We rst show

D
(X) H
1
(X, O
D
)

is injective. For this, given

D
(X) choose a small neighborhood U
0
centered at at point P with U
0
outside outside the support of D and (),
on which we have a local coordinate where = dz. Then set f
0
= 1/z on
U
0
and f
1
= 0 on Then set U
1
= XP. We then nd f
i
H
1
(X, O
D
)

,
and
Res(f
i
) = 1,
so is nontrivial in the dual space.
This shows h
0
(K D) h
1
(D), or h
0
(K D) h
1
(D) 0. Now we
sum the following two applications of Riemann-Roch:
h
0
(K D) h
1
(K D) = 1 g + deg(K D)
h
0
(D) h
1
(D) = 1 g + deg(D)
to get
(h
0
(D) h
1
(K D)) + (h
0
(K D) h
1
(D)) = 0
(using the fact that deg(K) = 2g 2). Since both terms on the left are
0, they must vanish, and this gives Serre duality. (I am grateful to Levent
Alpoge for pointing out this last step in the argument.)
Proof of Serre duality: using just niteness. Finally, here is an
argument that just uses niteness of h
1
(O). In particular it shows that
niteness implies H
1
(X, O)

= (X)

. It is based loosely on Forster.

We begin with some useful qualitative dimension counts that follow im-
mediately from the fact that h
0
(D) = 0 if deg(D) < 0, h
1
(D) 0 and the
fact that

= O
K
.
55
Theorem 10.3 For deg(D) > 0, we have
dimO
D
(X) deg(D) +O(1),
dim
D
(X) deg(D) +O(1), and
dimH
1
(X, O
D
) = deg(D) +O(1).
Pairings. Next we couple the product map
O
D

D

together with the residue map Res : H
1
(X, ) C to obtain a map
H
1
(X, O
D
)
D
(X) H
1
(X, ) C,
or equivalently a natural map

D
(X) H
1
(X, O
D
)

.
This map explicitly sends to the linear functional dened by
() = Res().
Using the long exact sequence associated to equation (8.1) we obtain:
Theorem 10.4 The inclusion O
D
O
D+P
induces a surjection:
H
1
(X, O
D
) H
1
(X, O
D+P
) 0.
Corollary 10.5 We have a natural surjective map:
H
1
(X, O
D
) H
1
(X, O
E
) 0
for any E D.
Put dierently, for E D we get a surjective map H
1
(X, O
E
)
H
1
(X, O
D
) and thus an injective map on the level of duals. For organi-
zational convenience we take the direct limit over increasing divisors and
set
V (X) = lim
D+
H
1
(X, O
D
)

.
Clearly
D
(X) maps into V (X) for every D, so we get a map /
(1)
(X)
V (X).
56
Theorem 10.6 The natural map /
(1)
(X) V (X) is injective. Moreover
a meromorphic form maps into H
1
(X, O
D
)

i
D
(X).
Proof. We have already shown injectivity in the previous proof. Now for
the second statement. If is in H
1
(X, O
D
)

, then it must vanish on all

coboundaries for this group. Suppose however (k+1) = ord
P
() < D(P)
for some P. Then k D(P), so in the construction above we can arrange
that = 0 in H
1
(X, O
D
). This is a contradiction. Thus ord
P
() D(P)
for all P, i.e.
D
(X).
Completion of the proof of Serre duality. Note that both /
(1)
(X)
and V (X) are vector spaces over the eld of meromorphic functions /(X).
(Indeed the former vector space is one-dimensional, generated by any mero-
morphic 1-form.)
Given H
1
(X, O
D
)

V (X), we must show is represented by a

meromorphic 1-form . (By the preceding result, will automatically lie in

D
(X).)
The proof will be by a dimension count. We note that for n 0, we
have
dimH
1
(X, O
DnP
)

= n +O(1).
On the other hand, this space contains
D+nP
(X) as well as O
nP
(X) .
Both of these spaces have dimension bounded below by n+O(1). Thus for n
large enough, they meet in a nontrivial subspace. This means we can write
f = where f is a nonzero meromorphic function and is a meromorphic
form. But then = /f is also a meromorphic form, so we are done!
We can now round out the discussion by proving some results promised
above.
Theorem 10.7 For any divisor D we have
H
0
(X, O
D
)

= H
1
(X,
D
)

.
Proof. We have
H
0
(X, O
D
)

= H
0
(X,
DK
)

= H
1
(X, O
KD
)

= H
1
(X,
D
)

57
Corollary 10.8 We have H
1
(X, O
D
) = 0 as soon as deg(D) > deg(K) =
2g 2.
Proof. Because then H
1
(X, O
D
)

=
D
(X) = 0.
Corollary 10.9 If deg(D) > 2g 2, then h
0
(D) = 1 g + deg(D).
Corollary 10.10 We have H
1
(X, /) = H
1
(X, /
(1)
) = 0.
Proof. Any representative cocycle (f
ij
) for a class in H
1
(X, /) can be re-
garded as a class in H
1
(X, O
D
) for some D of large degree. But H
1
(X, O
D
) =
0 once deg(D) is suciently large. Thus (f
ij
) splits for the sheaf O
D
, and
hence for /.
Corollary 10.11 Every element of H
1
(X, O
D
) can be represented as a Mittag-
Leer coboundary, g
ij
= f
i
f
j
with meromorphic (f
i
). Similarly for
H
1
(X,
D
).
Mittag-Leer for 1-forms, revisited. Here is another proof of the
Mittag-Leer theorem for 1-forms. Suppose we consider all possible prin-
cipal parts with poles of order at most n
i
> 0 at P
i
, and let D =

n
i
P
i
.
The dimension of the space of principal parts is then n =

n
i
= deg D.
In addition, the principal part determines the solution to the Mittag-Leer
problem up to adding a holomorphic 1-form. That is, the solutions lie in the
space
D
(X), and the map to the principal parts has (X) as its kernel.
Thus the dimension of the space of principal parts that have solutions
is:
k = dim
D
(X) dim(X).
But by Riemann-Roch we have
dim
D
(X) = h
0
(K +D) = h
0
(D) + deg(K +D) g + 1
= 2g 2 + deg D g 1 = g + deg D 1 = g +n 1,
so the solvable data has dimension k = n1. Thus there is one condition on
the principal parts for solvability, and that condition is given by the residue
theorem.
58
11 Maps to projective space
In this section we explain the connection between the sheaves O
D
, linear
systems and maps to projective space.
Projective space. Let V be an (n + 1)-dimensional vector space over C.
The space of lines (one-dimensional subspaces) in V forms the projective
space
PV = (V 0)/C

.
It has the structure of a complex n-manifold. The subspaces S V give
rise to planes PS PV ; when S has codimension one, PS is a hyperplane.
The dual projective space PV

parameterizes the hyperplanes in PV , via

the correspondence V

Ker() = S V .
We can also form the quotient space W = V/S. Any line L in V that
is not entirely contained in S projects to a line in W. Thus we obtain a
natural map
: (PV PS) P(V/S).
All the analytic automorphisms of projective space come from linear
automorphisms of the underlying vector space: that is,
Aut(PV ) = GL(V )/C

= PGL(V ).
We let P
n
= PC
n+1
with homogeneous coordinates [Z] = [Z
0
: : Z
n
]. It
satises
Aut(P
n
) = PGL
n+1
(C).
The case n = 1 gives the usual identication of automorphisms of

C with
Mobius transformations.
Projective varieties. The zero set of a homogeneous polynomial f(Z)
denes an algebraic hypersurface V (f) P
n
. A projective algebraic variety
is the locus V (f
1
, . . . , f
n
) obtained as an intersection of hypersurfaces.
The ratio
F(Z) = f
1
(Z)/f
0
(Z) = [f
0
(Z) : f
1
(Z)]
of two homogeneous polynomials of the same degree denes a meromorphic
function
F : P
n
P
1
.
Its values are undetermined on the subvariety V (f
0
, f
1
), which has codimen-
sion two if f
0
and f
1
are relatively prime.
The Hopf bration. By considering the unit sphere in C
n+1
, we obtain the
Hopf bration : S
2n+1
CP
n
with bers S
1
. This shows that projective
59
space is compact. Moreover, for n = 1 the bers of are linked circles in
S
3
, and generates
3
(S
2
)

= Z.
Ane space. The locus A
n
= P
n
V (Z
0
) is isomorphic to C
n
with
coordinates (z
1
, . . . , z
n
) = Z
i
/Z
0
, while V (Z
0
) itself is a hyperplane; thus
we have
P
n

= C
n
P
n1
.
By permuting the coordinates, we get a covering of P
n
by n+1 ane charts.
Any ordinary polynomial p(z
i
) has a unique homogeneous version P(Z
i
)
of the same maximal degree, such that p(z
i
) = P(1, z
1
, . . . , z
n
). Thus any
ane variety V (p
1
, . . . , p
m
) has a natural completion V (P
1
, . . . , P
m
) P
m
.
This variety is smooth if it is a smooth submanifold in each ane chart.
Examples: Curves in P
2
.
1. The ane curve x
2
y
2
= 1 meets the line at innity in two points
corresponding to its two asymptotes; its homogenization X
2
Y
2
= Z
2
is smooth in every chart. In (y, z) coordinates it becomes 1 y
2
= z
2
,
which means the line at innity (z = 0) in two points.
2. The ane curve dened by p(x, y) = y x
3
= 0 is smooth in C
2
, but
its homogenization P(X, Y, Z) = Y Z
2
X
3
denes the curve z
2
= x
3
in the ane chart where Y ,= 0, which has a cusp.
3. The space of homogeneous polynomials of degree d in P
n
has dimension
given by
dimP
d
(C
n+1
) =
_
d +n
n
_

This can be seen by inserting n markers into a list of d symbols, and

turning all the symbols up to the rst marker into Z
0
s, then the next
stretch into Z
1
s, etc. Thus the space of curves of degree d in P
2
is
itself a projective space P
N
, with N = d(d + 3)/2. E.g. there is a
5-dimensional space of conics, a 9-dimensional space of cubics and a
14-dimensional space of quartics.
The degree of a plane curve. If f is irreducible, then V (f) meets a
typical line in exactly d = deg(f) points. Thus the degree of f is a visible
property of V (f).
If f = f
1
f
n
is a product of distinct irreducibles, then V (f) =

V (f
i
).
These are also important examples of curves of degree d. For example, d
distinct lines form a curve of degree d.
60
The normalization of a singular curve. Every irreducible homogeneous
polynomial f on P
2
determines a compact Riemann surface X together with
a generically injective map : X V (f). The Riemann surface X is called
the normalization of the (possibly singular) curve V (f).
To construct X, use projection from a typical point P P
2
V (f) to
obtain a surjective map : V (f) P
1
. After deleting a nite set from
domain and range, including all the singular points of V (f), we obtain an
open Riemann surface X

= V (f) C and a degree d covering map

: X

(P
1
B).
As we have seen, there is a canonical way to complete X

and to a compact
Riemann surface X and a branched covering : X P
1
. It is then easy to
see that the isomorphism
: X

V (f) C P
2
extends to a holomorphic map : X P
2
with image V (f).
Examples: cusps and nodes. The cuspidal cubic y
2
= x
3
is normalized
by : P
1
P
2
given by (t) = (t
2
, t
3
).
The nodal cubic y
2
= x
2
(1x) is also interesting to uniformize. The idea
is to project from the node at (x, y) = (0, 0), i.e. let us use the parameter
t = y/x or equivalently set y = tx. Then we nd
t
2
x
2
= x
2
(1 x)
and so (x, y) = (1 t
2
, t(1 t
2
) give a parameterization of this curve.
Linear systems. Two divisors D, E are linearly equivalent if DE = (f)
is principal; then O
D

= O
E
.
The complete linear system [D[ determined by D is the collection of
eective divisors E linearly equivalent to D. These are exactly the divisors
of the form E = (f) +D, where f H
0
(X, O
D
). The divisor only depends
on the line determined by f. We have a natural bijection
[D[

= PO
D
(X).
Technically [D[ is a subset of the space of all divisors on X. Equivalent
divisors determine the same linear system, i.e.
[E[ = [D[ Div
d
(X)

= X
(d)
.
We emphasize that all divisors in [D[ are eective and of the same degree.
61
Linear systems and basepoints. A general linear system is simply a
subspace L

= P
r
[D[. It corresponds to a subspace S H
0
(X, O
D
).
The base locus of a linear system is the largest divisor B 0 such that
E B for all E L. A linear system is basepoint free if B = 0.
Geometric linear systems. Let X P
n
be a smooth nondegenerate
curve, and let H be a hyperplane. Then D = H X is an eective divisor
called a hyperplane section. Any two such divisors are linearly equivalent,
because H
1
H
2
= (f) where f : P
n
P
1
is a rational map of degree one
(a ratio of linear forms), and f[X is a meromorphic function.
Thus the set of hyperplane sections determines a linear system L [D[.
It is basepoint free because for any P X there is an H disjoint from P.
On the other hand, if we take only those hyperplanes passing through
P X, then we get a smaller linear system whose base locus is P itself.
Maps to P
n
. Let S /(X) be a linear subspace of meromorphic functions
of dimension n + 1. We have a natural map

S
: X PS

given by
S
(x)(f) = f(x). More concretely, if we choose a basis for S then
the map is given by

S
(x) = [f
0
(x) : : f
n
(x)].
The linear functional x f(x) in S

makes sense at most points of X; the

only potential problems come from zeros and poles. To handle these, we
simply choose a local coordinate z near x with z(x) = 0, and dene

S
(x)(f) = (z
p
f(z))[
z=0
,
where p is chosen so all values are nite and so at least one value is nonzero.
Note that the map
S
is nondegenerate: its image is contained in no
hyperplane.
Note also that if we replace S with fS, where f /

(X), then the map

S
remains the same. Thus we can always normalize so that 1 S.
Dierent perspectives on maps to projective space. The following
objects are in natural bijection:
1. Nondegenerate holomorphic maps f : X P
n
of degree d, with a
distinguished hyperplane H, up to automorphisms of P
n
;
2. Subspaces S /(X) of dimension (n + 1) with 1 S, and with
generic elements of degree d; and
62
3. Basepoint-free linear systems L of degree d and dimension n, together
with a distinguished divisor D L.
Here are some of the relations between them.
1 = 2. We can normalize so that H = (Z
0
), and then f
i
= Z
i
/Z
0
when restricted to X span the desired space of meromorphic functions S.
1 = 3. We let D = X H and let L [D[ be the collection of all
hyperplane sections.
2 = 1. We have seen that S canonically determines a map to PS

.
Any function f S then determines a hyperplane in the image, and we let
H be the hyperlane determined by the constant function 1 S.
2 = 3. To dene the linear system L, we rst let D = (f)

be the
polar divisor of a generic f S. This is the same as setting D = max(f
i
)

over a basis for S. Then we let

L = E = (f) +D : f S [D[.
Since the maximum is achieved, this linear system is basepoint free, and
since 1 S we have D L.
3 = 2. We let S = f : (f) = E D, E L. Then D corresponds
to the constant function.
3 = 1. For each x X we get a hyperplane in [D[ by considering the
set of E containing x. This gives a natural map X P
n
= [D[

, and points
of D span a hyperplane H P
n
.
Embeddings and smoothness for complete linear systems. We now
study the complete linear system [D[, and the associated map

D
: X PH
0
(X, O
D
)

in more detail.
We say O
D
is generated by global sections if for each x X we have a
global section f H
0
(X, O
D
) such that the stalk O
D,x
, which is an O
x
-
module, is generated by f; that is, O
D,x
= O
x
f.
Theorem 11.1 The following are equivalent.
1. [D[ is basepointfree.
2. h
0
(D P) = h
0
(D) 1 for all P X.
3. O
D
is generated by global sections.
4. For any x X there is an f O
D
(X) such that ord
x
(f) = D(x).
63
Example. Note that h
0
(nP) = 1 for n = 0 and = g for n = 2g 1. Thus
(for large genus) there must be values of n such that h
0
(nP) = h
0
(nP P).
In this case, D = nP is not globally generated.
Here is a detailed proof of the equivalence between (1) and (3) above.
Theorem 11.2 If D is basepointfree, then [D[ consists of the hyperplane
sections
1
D
(H).
Proof. Let f
0
, f
1
, . . . , f
n
be a basis for O
D
(X). Then the map to P
n
is
given locally near p X by

D
(z) = [z
d
f
0
(z) : z
d
f
1
(z) : : z
d
f
n
(z)]
where z is a local coordinate, z(p) = 0, and d is the maximum order of pole
at p of the f
i
(z).
Since O
D
is globally generated, we have d = D(p). Now p belongs
to the hyperplane at innity H = (Z
0
= 0), determined by coordinates
[Z
0
: : Z
n
] on P
n
, if and only if z
d
f
0
(p) = 0. But this means exactly
that ord
p
(f
0
) > D(p). Therefore the divisor E = (f
0
) +D coincides with

1
D
(H) (where the preimage is counted with multiplicity).
Conversely, if E = (f)+D 0, then f can be taken as a basis of element
of O
D
(X), determining in turn a hyperplane section giving E.
Corollary 11.3 If [D[ is base-point free and
D
is an embedding, then

D
(X) is a curve of degree deg D.
Theorem 11.4 Let : X P
n
be a map of X to projective space with
the image not contained in a hyperplane. Then =
D
, where D is the
divisor of a hyperplane section, [D[ is a base-point free linear system, and
where : P
N
P
n
is projection from a linear subspace PS disjoint from

D
(X).
Proof. Since hyperplanes can be moved, the linear system [D[ is base-point
free, and all hyperplane sections are linearly equivalent to D. Thus can
be regarded as the natural map from X to PS

, where S is a subspace of
H
0
(X, O
D
), so can be factored through the map
D
to PH
0
(X, O
D
)

.
64
Theorem 11.5 Let [D[ be base-point free. Then for any eective divisor
P = P
1
+. . . +P
n
on X, the linear system
P +[D P[ [D[
consists exactly of the hyperplane sections E =
D
(X) H passing through
(P
1
, . . . , P
n
). In particular, the dimension of the space of such hyperplanes
is dim[D P[.
Proof. A hyperplane section E [D[ passes through P i E P 0 i
E = P +E

where E

[D P[.
Theorem 11.6 If h
0
(DP Q) = h
0
(D) 2 for any P, Q X, then [D[
provides a smooth embedding of X into projective space.
Proof. This condition says exactly that the set of hyperplanes passing
through
D
(P) and
D
(Q) has dimension 2 less than the set of all hyper-
planes. Thus
D
(P) ,=
D
(Q), so
D
is 1 1. The condition on (D2P)
says that the set of hyperplanes containing
D
(P) and

D
(P) is also 2
dimensions less, and thus
D
is an immersion. Thus
D
is a smooth embed-
ding.
Theorem 11.7 The linear system [D[ is base-point free if deg D 2g, and
gives an embedding into projective space if deg D 2g + 1.
Proof. Use the fact that if deg D > deg K = 2g 2, then we have h
0
(D) =
deg D g + 1, which is linear in the degree.
Corollary 11.8 Every compact Riemann surface embeds in projective space.
Remark. By projecting we get an embedding of X into P
3
and an immer-
sion into P
2
.
Not every Riemann surface can be embedded into the plane! In fact a
smooth curve of degree d has genus g = (d 1)(d 2)/2, so for example
there are no curves of genus 2 embedded in P
2
.
Examples of linear systems.
65
1. Genus 0. On P
1
, we have O
D

= O
E
if d = deg(D) = deg(E). This
sheaf is usually referred to as O(d); it is well-dened up to isomor-
phism, [D[ consists of all eective divisors of degree d. Note that
O
d
(P
1
) is the d + 1-dimensional space of polynomials of degree d.
The corresponding map
D
: P
1
P
d
is given in ane coordinates by

D
(t) = (t, t
2
, . . . , t
d
). Its image is the rational normal curve of degree
d. Particular cases are smooth conics and the twisted cubic.
2. Genus 1. On X = C/ with P = 0, the linear system [P[ is not base-
point free, but [2P[ is, and [3P[ gives an embedding into the plane,
via the map z ((z),

(z)).
Recall that D =

m
i
P
i
is a principal divisor on X i deg(D) = 0 and
e(D) =

m
i
P
i
in C/ is zero. Thus 3Q [3P[ i e(3(P Q)) = 0.
This shows:
A smooth cubic curve X in P
2
has 9 ex points, correspond-
ing to the points of order 3 in the group law on X.
The canonical map. Note that the embedding of a curve X of genus 1 into
P
2
by the linear series [3P[ breaks the symmetry group of the curve: since
Aut(X) acts transitively, the 9 exes of Y =
3P
(X) are not intrinsically
special.
For genus g 2 on the other hand there a more natural embedding
one which does not break the symmetries of X given by the canonical
linear system [K[.
The linear system [K[ corresponds to the map

K
: X P(X)

= P
g1
given by (x) = [
1
(x), . . . ,
g
(x)], where
i
is a basis for (X).
Theorem 11.9 The linear system [K[ is base-point free.
Proof. Since X is not isomorphic to P
1
, we have h
0
(P) = h
0
(0) = 1. Thus
h
0
(K P) = h
0
(P) + deg(K P) g + 1 = h
0
(K) 1.
In other words, for any P X we have (P) ,= 0 for some (X).
66
Theorem 11.10 Either [K[ gives an embedding of X into P
g1
, or X is
hyperelliptic.
Proof. Suppose [K[ does not given an embedding. Then h
0
(KP Q) >
h
0
(K) 2 for some P, Q X, and thus h
0
(P + Q) > 1. Thus there exists
meromorphic function f with polar divisor P + Q, providing a degree two
map f : X P
1
.
Now suppose X is hyperelliptic. Then there is a degree two holomorphic
map f : X P
1
branched over the zeros of a polynomial p(z) of degree
2g + 2. A basis for the holomorphic 1-forms on X is given by

i
=
z
i
dz
_
p(z)
for i = 0, . . . , g 1. That is,
i
(x) = f(x)
i

0
. It follows that the canonical
map : X P
g1
is given by
f(x) = [
i
(x)] = [
0
(x)f(x)
i
] = [f(x)
i
].
In other words, the canonical map factors as = f, where : P
1
P
g1
is the rational normal curve of degree g 1.
Canonical curves of genus two. We now describe more geometrically
the canonical curves of genus two, three and four.
Theorem 11.11 Any Riemann surface X of genus 2 is hyperelliptic, and
any degree two map of X to P
1
agrees with the canonical map (up to Aut P
1
).
Proof. In genus 2, we have deg K = 2g 2 = 2, so the canonical map
: X P
g1
= P
1
already presents X as a hyperelliptic curve. If f :
X P
1
is another such map, with polar divisor P + Q, then we have
h
0
(P + Q) = 2 = h
0
(K P Q) + 2 2 + 1; thus there exists an with
zeros just at P, Q and therefore P +Q is a canonical divisor.
Corollary 11.12 The moduli space of curves of genus two is isomorphic to
the 3-dimensional space /
0,6
of isomorphism classes of unordered 6-tuples
of points on P
1
. Thus /
2
is nitely covered by C
3
D, where D consists
of the hyperplanes x
i
= 0, x
i
= 1 and x
i
= x
j
.
67
Remark. Here is a topological fact, related to the fact that every curve
of genus two is hyperelliptic: if S has genus two, then the center of the
mapping-class group Mod(S) is Z/2, generated by any hyperelliptic involu-
tion. (In higher genus the center of Mod(S) is trivial.)
Canonical curves of genus 3. Let X be a curve of genus 3. Then
X is either hyperelliptic, or its canonical map realizes it as a smooth plane
quartic. We will later see that, conversely, any smooth quartic is a canonical
curve (we already know it has genus 3). This shows:
Theorem 11.13 The moduli space of curves of genus 3 is the union of the
5-dimensional space /
0,8
and the 6-dimensional moduli space of smooth
quartics, (P
14
D)/ PGL
3
(C).
Note: a smooth quartic curve that degenerates to a hyperelliptic one be-
comes a double conic. The eight hyperelliptic branch points can be thought
of as the intersection of this conic with an innitely near quartic curve.
Symmetric curves of genus 3. A particularly symmetric curve of genus
3 is given by the Klein quartic
X
3
Y +Y
3
Z +Z
3
X = 0.
Its symmetry group G

= PSL
2
(Z/7) has order 168 and gives as quotient
the (2, 3, 7)-orbifold. A subgroup of order 21 is easily visible in the form
above: we can cyclically permutate the coordinates and also send (X, Y, Z)
to (
2
X, Y,
4
Z) where
7
= 1. More geometrically, the Klein quartic is
tiled by 24 regular 7-gons. It is easily shown that 1/42 = (S
2
(2, 3, 7))
is the largest possible Euler characteristic for a hyperbolic orbifold, which
implies:
Theorem 11.14 (Hurwitz) For any compact Riemann surface of genus g
we have [ Aut(X)[ 84(g 1).
Equality is achieved in the case above.
The Fermat quartic, dened by
X
4
+Y
4
= Z
4
,
is somewhat similar; its symmetry group is PSL
2
(Z/8), which has order 96;
the quotient is the (2, 3, 8) orbifold, of Euler characteristic 1/24, and this
curve is tiled by 12 octagons. In this case the full symmetry group is easily
68
visible: there is an S
3
coming from permutations of the coordinates, and a
(Z/4)
2
action coming from multiplication of X and Y by powers of

1.
Canonical curves of genus 4. Now let X P
3
be a canonical curve of
genus 4 (in the non-hyperelliptic case). Then X has degree 6.
Theorem 11.15 X is the intersection of an irreducible quadric and cubic
hypersurface in P
3
.
Proof. The proof is by dimension counting again. There is a natural lin-
ear from Sym
2
((X)) into H
0
(X, O
2K
). Since dim(X) = 4, the rst
space has dimension
_
3+2
2
_
= 10, while the second has dimension 3g
3 = 9 by Riemann-Roch. Thus there is a nontrivial quadratic equation
Q(
1
, . . . ,
4
) = 0 satised by the holomorphic 1-forms on X; equivalent X
lies on a quadric. The quadric is irreducible because X does not lie on a
hyperplane.
Carrying out a similar calculation for degree 3, we nd dimSym
3
((X)) =
_
3+3
3
_
= 20 while dimH
0
(X, O
3K
) = 5g 5 = 15. Thus there is a 5-
dimensional space of cubic relations satised by the (
i
). In this space, a
4-dimensional subspace is accounted for by the product of Q with an arbi-
trary linear equation. Thus there must be, in addition, an irreducible cubic
surface containing X.
We will later see that the converse also holds.
Dimension counts for linear systems. Here is another perspective on
the preceding proof. In intersection of surfaces of degree d in P
3
with X
gives a birational map between projective spaces,
[dH[ [dK[.
Now note that in general a linear map : A B between vector spaces
gives a birational map
: PA PB
which is projection fromPC where C = Ker . Then dimAdimB dimC
and thus
dimPC dimPAdimPB 1.
In the case at hand PC corresponds to the linear system S
d
(X) of surfaces
of degree d containing X. Thus we get:
dimS
d
(X) dim[dH[ dim[dK[ 1.
69
For d = 2 this gives
dimS
2
(X) 9 (3g 4) 1 = 9 8 1 = 0
which shows there is a unique quadric Q containing X. For d = 3 we get
dimS
3
(X) 19 (5g 6) 1 = 19 14 1 = 4.
Within S
3
(X) we have Q+[H[ which is 3-dimensional, and thus there must
be a cubic surface C not containing Q in S
3
(X) as well.
This cubic is not unique, since we can move it in concert with Q+H.
Special divisors. An eective divisor D is special if h
0
(K D) > 0, i.e. if
there is a holomorphic 1-form ,= 0 vanishing at D.
In terms of the canonical map : X P
g1
, a divisor D is special i
(D) lies in a hyperplane H (determined by ). (Moreover, the index of
speciality, i(D) = h
0
(K D), is one more than the dimension of the space
of hyperplanes passing through D.)
Special divisors of degree g. The case of divisors of degree g is partic-
ularly interesting. Geometrically we see there exist plenty of such divisors
note that [H (X)[ = 2g 2, so a given hyperplane determines many
such divisors. On the other hand, g typical points on (X) do not span a
hyperplane, so these divisors really are special.
Proposition 11.16 If g 2 then there exist special divisors of degree g.
Theorem 11.17 An eective divisor D of degree g is special if and only if
there is a nonconstant meromorphic function on X with (f) +D 0.
Proof. By Riemann-Roch, we have h
0
(D) = i(D) + deg D g + 1 =
i(D) + 1 > 1 i i(D) > 0 i D is special.
Corollary 11.18 If g 2 then X admits a meromorphic function of degree
g.
Example: genus 3. A curve of genus 3 either admits a map to P
1
of
degree two, or it embeds as a curve X P
2
of degree 4. In the latter case,
projection to P
1
from any P X gives a map of degree g = 3.
The Wronskian and Weierstrass points. Now we focus on divisors of
the form D = gP. We say P is a Weierstrass point if gP is special, i.e. if
70
there is a meromorphic function f : X P
1
with a pole of order at least
one and at most g at P, and otherwise holomorphic.
Example: there are no Weierstrass points on a Riemann surface of genus
1. The branch points of every hyperelliptic surface of genus g 2 are
Weierstrass points.
To have H(X) = D = gP, the hyperplane H should contain not just
P but the appropriate set of tangent directions at P, namely
((P),

(P),

(P), . . . ,
(g1)
(P)).
For these tangent (g 1) tangent directions to span a (g 2)-dimensional
plane H through (P), there must be a linear relation among them; that is,
the Wronskian determinant W(P) must vanish.
In terms of a basis for and a local coordinate z at P, the Wronskian
is given by
W(z) = det
_
d
j

i
dz
j
_
,
where j = 0, . . . , g 1 and i = 1, . . . , g.
We can see directly the vanishing of the Wronskian is equivalent to gP
being special.
Theorem 11.19 The Wronskian vanishes at P i there is a holomorphic
1-form ,= 0 with a zero at P of order at least g.
Proof. The determinant vanishes i there is a linear combination of the
basis elements
i
whose derivatives through order (g 1) vanish at P.
The quantity W = W(z) dz
N
turns out to be independent of the choice
of coordinate, where N = 1 + 2 + . . . + g = g(g + 1)/2. This value of N
arises because the jth derivative of a 1-form behaves like dz
j+1
.
Thus W(z) is a section of O
NK
, so its number of zeros is deg NK =
N(2g 2) = (g 1)g(g + 1). This shows:
Theorem 11.20 Any Riemann surface of genus g has (g1)g(g+1) Weier-
strass points, counted with multiplicity.
Weierstrass points of a hyperelliptic curve. These correspond to the
branch points of the hyperelliptic map : X P
1
, since the projective
normal curve P
1
P
g1
has no exes.
Weierstrass points in genus 3. The Weierstrass points on a smooth
quartic correspond to exes; there are 2 3 4 = 24 of them in general. For
71
example, on the Fermat curve x
4
+ y
4
= 1, there are 12 exes altogether,
each of multiplicity 2. Of these, 8 lie in the ane plane, and arise when one
coordinate vanishes and the other is a 4th root of unity.
At the exes we have h
0
(3P) = 2. How can one go from a ex P to a
degree 3 branched covering f : X P
1
? We can try projection f
P
from P,
but in general the line L tangent to X at P will meet X in a fourth point
Q. Thus f
P
will have a double pole at P and a simple pole at Q.
Instead, we project from Q! Then the line L through P and Q has
multiplicity 3 at P, giving a triple order pole there.
Flexes of plane curves. In general, if C is dened by F(X, Y, Z) = 0,
then the exes of C are the locus where both F and the Hessian H of F
vanish. For the Fermat curve, we have F(X, Y, Z) = X
4
+ Y
4
+ Z
4
and
H = 1728(XY Z)
2
. On a smooth curve of degree d the number of exes is
3d(d 2).
The dimension of moduli space M
g
: Riemanns count. What is the
dimension of /
g
? We know the dimension is 0, 1 and 3 for genus g = 0, 1
and 2 (using 6 points on P
1
for the last computation).
Here is Riemanns heuristic. Take a large degree d g, and consider
the bundle T
d
/
g
whose bers are meromorphic functions f : X P
1
of degree d. Now for a xed X, we can describe f T
d
(X) by rst giving
its polar divisor D 0; then f is a typical element of H
0
(X, O
D
). (The
parameters determining f are its principal parts on D.) Altogether with
nd
dimT
d
(X) = d +h
0
(D) = 2d g + 1.
On the other hand, f has b critical points, where
(X) = 2 2g = 2d b,
so b = 2d + 2g 2. Assuming the critical values are distinct, they can be
continuously deformed to determine new branched covers (X

, f

). Thus the
dimension of the total space is given by
b = dimT
d
= dimT
d
(X) + dim/
g
= 2d + 2g 2 = 2d g + 1 + dim/
g
,
and thus dim/
g
= 3g 3. This dimension is in fact correct.
On the other hand, the space of hyperelliptic Riemann surfaces clearly
satises
dim1
g
= 2g + 2 dimAut P
1
= 2g 1,
since such a surface is branched over 2g +2 points. Thus for g > 2 a typical
Riemann surface is not hyperelliptic.
72
Tangent space to M
g
. As an alternative to Riemanns count, we note
that the tangent space to the deformations of X is H
1
(X, ), where

=

is the sheaf of holomorphic vector elds on X. By Serre duality and

Riemann-Roch, we have
dimH
1
(X, ) = dimH
1
(X, O
K
) = h
0
(2K) = 4g 4 g + 1 = 3g 3.
Serre duality also shows H
1
(X, ) is naturally dual to the space of holomor-
phic quadratic dierentials Q(X).
Plane curves again. The space of homogeneous polynomials on C
n+1
of
degree d has dimension N =
_
n+d
n
_
. Thus the space of plane curves of degree
d, up to automorphisms of P
2
, has dimension
N
d
=
_
2 +d
d
_
9.
We nd
N
d
=
_

_
3 = dimAut P
1
for d = 2,
1 = dim/
1
for d = 3,
6 = dim/
3
for d = 4,
12 < dim/
6
= 15 for d = 5.
Thus most curves of genus 6 cannot be realized as plane curves. In a sense
made precise by the Theorem below, there is no way to simply parameterize
the moduli space of curves of high genus:
Theorem 11.21 (Harris-Mumford) For g suciently large, /
g
is of
general type.
In fact g 24 will do.
12 Line bundles
Let X be a complex manifold. A line bundle : L X is a 1-dimensional
holomorphic vector bundle.
This means there exists a collection of trivializations of L over charts U
i
on X, say L
i

= U
i
C. Viewing L in two dierent charts, we obtain clutching
data g
ij
: U
i
U
j
C

such that (x, y

j
) L
j
is equivalent to (x, y
i
) L
i
i
g
ij
(x)y
j
= y
i
. This data satises the cocycle condition g
ij
g
jk
= g
ik
.
73
In terms of charts, a holomorphic section s : X L is encoded by
holomorphic functions s
i
= y
i
s(x) on U
i
, such that
s
i
(x) = g
ij
(x)s
j
(x).
Examples: the trivial bundle X C; the canonical bundle
n
T

X. Here
the transition functions are g
ij
= 1 for the trivial bundle and g
ij
= 1/ det D(
i

1
j
) for the canonical bundle, with charts
i
: U
i
C
n
.
In detail, on a Riemann surface X, with local coordinates z
i
: U
i
C,
a section of the canonical bundle is locally given by
i
= s
i
(z) dz
i
; it must
satisfy s
i
(z) dz
i
= s
j
(z) dz
j
, so s
i
= (dz
j
/dz
i
)s
j
.
Tensor powers. From L we can form the line bundle L

= L
1
, and more
generally L
d
, with transition functions g
d
ij
.
A line bundle is trivial if it admits a nowhere-vanishing holomorphic
section (which then provides an isomorphism between L and X C). Such
a section exists i there are s
i
O

(U
i
) such that s
i
/s
j
= g
ij
, i.e. i g
ij
is
a coboundary.
Thus line bundles up to isomorphism over X are classied by the coho-
mology group H
1
(X, O

).
Sections and divisors. Now consider a divisor D on a Riemann surface
X. Then we can locally nd functions s
i
/(U
i
) with (s
i
) = D. From
this data we construction a line bundle L
D
with transition functions g
ij
=
s
i
/s
j
. These transitions functions are chosen so that s
i
is automatically a
meromorphic section of L
D
; indeed, a holomorphic section if D is eective.
Theorem 12.1 The sheaf of holomorphic sections / of L = L
D
is isomor-
phic to O
D
.
Proof. Choose a meromorphic section s : X L with (s) = D. (On a
compact Riemann surface, s is well-dened up to a constant multiple.) Then
a local section t : U L is holomorphic if and only if the meromorphic
function f = t/s satises
(t) = (fs) = (f) +D 0
on U, which is exactly the condition that f O
D
. Thus the map f fs
gives an isomorphism between O
D
(U) and /(U).
74
Line bundles on Riemann surfaces. Conversely, it can be shown that
every line bundle L on a Riemann surface admits a non-constant meromor-
phic section, and hence L = L
D
for some D. More precisely, if / is the sheaf
of sections of L one can show (see e.g. Forster, Ch. 29):
Theorem 12.2 The group H
1
(X, /) is nite-dimensional.
Corollary 12.3 Given any P X, there exists a meromorphic section
s : X L with a pole of degree 1+h
1
(/) at P and otherwise holomorphic.
Corollary 12.4 Every line bundle has the form L

= L
D
for some divisor
D.
From the point of view of sheaf theory, we have
0 O

Div H
1
(X, O

) H
1
(X, /

) 0,
and since every line bundle is represented by a divisor, we nd:
Corollary 12.5 The group H
1
(X, /

) = 0.
Divisors and line bundles in higher dimensions. On complex man-
ifolds of higher dimension, we can similarly construct line bundles from
divisors. First, a divisor is simply an element of H
0
(/

/O

); this means
it is locally a formal sum of analytic hypersurfaces, D =

(f
i
). Then the
associated transition functions are g
ij
= f
i
/f
j
as before, and we nd:
Theorem 12.6 Any divisor D on a complex manifold X determines a line
bundle L
D
X and a meromorphic section s : X L with (s) = D.
Failure of every line bundle to admit a nonzero section. However
in general not every line bundle arises in this way. For example, there exist
complex 2-tori M = C
2
/ with no divisors but with plenty of line bundles
(coming from characters :
1
(M) S
1
).
Degree. The degree of a line bundle, deg(L), is the degree of the divisor of
any meromorphic section.
In terms of cohomology, the degree is associated to the exponential se-
quence: we have
. . . H
1
(X, Z) H
1
(X, O) H
1
(X, O

) H
2
(X, Z)

= Z.
75
This allows one to dene the degree or rst Chern class, c
1
(L) H
2
(X, Z),
for a line bundle on any complex manifold.
Projective space. For projective space, we have H
1
(P
n
, C) = 0, and in
fact H
1
(P
n
, O) = H
0,1
(P
n
) = 0. (This follows from the Hodge theorem,
which implies that
H
n
(X, C) =
p+q=n
H
p,q
(X).
This is nontrivial for the case of P
n
, n 2, because elements of H
0,1
are
represented by -closed forms, rather than simply all (0, 1)-forms as in the
case of a Riemann surface).
It follows that line bundles on projective space are classied by their
degree: we have
0 H
1
(P
n
, O

) H
2
(P
n
, Z)

= Z.
We let O(d) denote the (sheaf of sections) of the unique line bundle of degree
d. It has the property that for any meromorphic section s H
0
(P
n
, O(d)),
the divisor D = (s) represents d[H] H
2
(P
n
, Z), where H

= P
n1
is a
hyperplane.
Example: the tautological bundle. Let P
n
be the projective space of
C
n+1
with coordinates Z = (Z
0
, . . . , Z
n
). The tautological bundle P
n
has, as its ber over p = [Z], the line
p
= C Z C
n+1
. Its total space is
C
n+1
with the origin blown up.
To describe in terms of transition functions, let U
i
= (Z
i
,= 0) P
n
.
Then we can use the coordinate Z
i
itself to trivialize [U
i
; in other words,
we can map [U
i
to U
i
C the map
(Z
0
, . . . , Z
n
) ([Z
0
: : Z
n
], Z
i
).
(Here the origin must be blown up.) Then clearly the transition functions
are given simply by g
ij
= Z
i
/Z
j
, since they satisfy
Z
i
= g
ij
Z
j
.
Theorem 12.7 There is no nonzero holomorphic section of the tautological
bundle.
Proof. A section gives a map s : P
n
C
n+1
which would have to be
constant because P
n
is compact. But then the constant must be zero, since
this is the only point in C
n+1
that lies on every line through the origin.
76
As a typical meromorphic section, we can dene s(p) to be the intersec-
tion of
p
with the hyperplane Z
0
= 1. In other words,
s([Z
0
: Z
1
: : Z
n
]) = (1, Z
1
/Z
0
, . . . , Z
n
/Z
0
).
Then s
i
= Z
i
/Z
0
. Notice that this section is nowhere vanishing (since Z
i
has no zero on U
i
), but it has a pole along the divisor H
0
= (Z
0
).
Thus we have

= O(1). Similarly,

= O(1).
Homogeneous polynomials. Note that the coordinates Z
i
are sections
of O(1). Indeed, any element in V

naturally determines a function on the

tautological bundle over PV , linear on the bers, and hence a section of the
dual bundle. Similarly we have:
Theorem 12.8 The space of global sections of O(d) over P
n
can be natu-
rally identied with the homogeneous polynomials of degree d on C
n+1
.
One can appeal to Hartogs extension theorem to show all global sections
have this form.
Corollary 12.9 The hypersurfaces of degree d in projective space are ex-
actly the zeros of holomorphic sections of O(d).
The canonical bundle. To compute the canonical bundle of project space,
we use the coordinates z
i
= Z
i
/Z
0
, i = 1, . . . , n to dene a nonzero canonical
form
= dz
1
dz
n
on U
0
. To examine this form in U
1
, we use the coordinates w
1
= Z
0
/Z
1
,
w
i
= Z
i
/Z
1
, i > 1; then z
1
= 1/w
1
and z
i
= w
i
/w
1
, i > 1, so we have
= d(1/w
1
) d(w
2
/w
1
) d(w
n
/w
1
) = (dw
1
dw
n
)/w
n+1
1
.
Thus () = (n1)H
0
and thus the canonical bundle satises K

= O(n
1) on P
n
.
The adjunction formula.
Theorem 12.10 Let X Y be a smooth hypersurface inside a complex
manifold. Then the canonical bundles satisfy
K
X

= (K
Y
L
X
)[X.
77
Proof. We have an exact sequence of vector bundles on X:
0 TX TY TY/ TX = NX 0,
where NX is the normal bundle. Now (NX)

is the sub-bundle of T

Y [X
spanned by 1-forms that annihilate TX. If X is dened in charts U
i
by
f
i
= 0, then g
ij
= f
i
/f
j
denes L
X
. On the other hand, df
i
is a nonzero
holomorphic section of (NX)

. The 1-forms df
i
[X, however, do not t to-
gether on overlaps to form a global section of (NX)

. Rather, on X we have
f
j
= 0 so
df
i
= d(g
ij
f
j
) = g
ij
df
j
.
This shows (df
i
) gives a global, nonzero section of (NX)

L
X
, and hence
this bundle is trivial on X.
On the other hand K
Y
= K
X
(NX)

, by taking duals and determi-

nants. Thus K
X
= K
Y
NX = K
Y
L
X
.
Smooth plane curves. Using the adjunction formula plus Riemann-Roch
we can obtain some interesting properties of smooth plane curves X P
2
of degree d.
Theorem 12.11 Let f : X P
n
be a holomorphic embedding. Then f is
given by a subspace of sections of the line bundle L X, where L = f

O(1).
Proof. The divisors of section of O(1) are hyperplanes.
Theorem 12.12 Every smooth plane curve of degree d has genus g = (d
1)(d 2)/2.
Proof. We have K
X

= K
P
2 L
X
= O(d 3). Any curve Z of degree d 3
is the zero set of a section of O(d 3) and hence restricts to the zero set of
a holomorphic 1-form on X. Thus we nd 2g 2 = d(d 3).
Corollary 12.13 Every smooth quartic plane curve X is a canonical curve.
Proof. We have K
X

= O(d 3) = O(1), which is the linear system that
gives the original embedding of X into P
2
.
78
Next note that the genus g(X) = (d 1)(d 2)/2 coincides with the
dimension of the space of homogeneous polynomials on C
3
of degree d 3.
This shows:
Theorem 12.14 Every eective canonical divisor on X has the form K =
X Y , where Y is a curve of degree d 3.
Theorem 12.15 Any n + 1 distinct points in P
2
impose independent con-
dition on curves of degree n.
Proof. Choose Y to be the union of n random lines through the rst k n
points. Then Y is an example of a curve through the rst k points that
does not pass through the k +1st. This shows that adding the k +1st point
imposes an additional condition on Y .
Theorem 12.16 A smooth curve X of degree d > 1 admits a nonconstant
map to P
1
of degree d 1, but none of degree d 2.
Proof. For degree d 1, simply projection from a point on X. For the
second assertion, suppose f : X P
1
has degree e d2. Let E X be a
generic ber of f. Then the d 2 points E impose independent conditions
on the space of curves Y degree d 3. Consequently
h
0
(K E) = g deg E.
By Riemann-Roch we then have:
h
0
(E) = 1 g + deg(E) h
0
(K E) = 1,
so [E[ does not provide a map to P
1
.
Hypersurfaces in products of projective spaces. Here are two further
instances of the adjunction theorem.
Theorem 12.17 Every smooth degree 6 intersection X of a quadric Q and
a cubic surface C is a canonical curve in P
3
.
Proof. It can be shown that Q and C are smooth and transvere along X
(this is nontrivial). We then have K
Q

= K
P
3 L
Q
and thus
K
X

= K
Q
L
C

= K
P
3 L
Q
L
C

= O(4 + 2 + 3) = O(1).
79
Theorem 12.18 Every smooth (d, e) curve on Q = P
1
P
1
has genus g =
(d 1)(e 1).
Proof. It is easy to see that K
XY
= K
X
K
Y
. Thus K
Q
= O(2, 2).
Therefore 2g 2 = C K
C
and K
C
= O(d 2, e 2), so 2g 2 = d(e 2) +
e(d 2), which implies the result.
Canonical curves of genus 5. By a calculation with Riemann-Roch, one
can show that any canonical curve X P
4
is an intersection of 3 quadrics.
Conversely, a complete intersection of 3 quadrics is a canonical curve, be-
cause its canonical bundle is O(5 + 2 + 2 + 2) = O(1).
In general, the canonical bundle of a complete intersection X of hyper-
surfaces of degrees d
i
in P
n
is given by
K
X

= O(n 1 +

d
i
).
K3 surfaces. Manifolds with trivial canonical bundle are often interesting
in higher dimensions they are called Calabi-Yau manifolds.
For Riemann surfaces, K
X
is trivial i X is a complex torus. For 2-
dimensional manifolds, complex tori also have trivial canonical, but they
are not the only examples. Another class is provided by the K3 surfaces,
which by denition are simply-connected complex surfaces with K
X
trivial.
Example:
Theorem 12.19 Every smooth surface of degree 4 in X = P
3
and of degree
(2, 2, 2) in X = P
1
P
1
P
1
is a K3 surface.
Proof. Here K
X
= O(3) or K
X
= O(2, 2, 2), so the canonical bundle
is trivial. By the Lefschetz hyperplane theorem, a hypersurface in a simply-
connected complex 3-manifold is always itself simply-connected.
13 Curves and their Jacobians
We now turn to the important problem of classifying line bundles on a Rie-
mann surface X; equivalently, of classifying divisors modulo linear equiva-
lence.
The Jacobian. Recall that a holomorphic 1-form on X is the same thing
as a holomorphic map f : X C well-dened up to translation in C. If
80
the periods of f happen to generate a discrete subgroup of C, then we
can regard f as a map to C/. However the periods are almost always
indiscrete. Nevertheless, we can put all the 1-forms together and obtain a
map to C
g
/.
Theorem 13.1 The natural map H
1
(X, Z) (X)

has as its image a

lattice

= Z
2g
.
Proof. If not, the image lies in a real hyperplane dened, for some nonzero
(X), by the equation Re () = 0. But then all the periods of Re
vanish, which implies the harmonic form Re = 0.
The Jacobian variety is the quotient space Jac(X) = (X)

/H
1
(X, Z),
the cycles embedded via periods.
Theorem 13.2 Given any basepoint P X, there is a natural map
P
:
X Jac(X) given by
P
(Q) =
_
Q
P
.
We will later show this map is an embedding, and thus Jac(X), roughly
speaking, makes X into a group.
Example: the pentagon. Let X be the hyperelliptic curve dened by
y
2
= x
5
1. Geometrically, X can be obtained by gluing together two regular
pentagons. Cleary X admits an automorphism T : X X of order 5. Using
the pentagon picture, one can easily show there is a cycle C H
1
(X, Z) such
that its ve images T
i
(C) span H
1
(X, Z)

= Z
4
. This means H
1
(X, Z) = AC
is a free, rank one A-module, where A = Z[T]/(1 +T +T
2
+T
3
+T
4
).
Now T

acts on (X). Since X/T has genus zero, T has no invariant

forms. Thus we can choose a basis (
1
,
2
) for (X) such that
T

i
=
i

i
where
i
is a primitive 5th root of unity.
Let us scale these
i
so
_
C

i
= 1. Let
: H
1
(X, Z) C
2
be the period map, dened by
(B) =
__
B

1
,
_
B

2
_
.
81
Since _
T(B)
=
_
B
T

,
we have
(TB) =
_

1
0
0
2
_
(B).
Since H
1
(X, Z) = Z[T] C, we nd that C
2
is simply the image of the
ring Z[T] under the ring homomorphism that sends T to (
1
,
2
).
Since is a lattice, we cannot have
2
=
1
=
4
1
, nor can we have

2
=
1
. Thus
2
=
2
1
or
3
1
. In the latter case we can interchange the
eigenforms to obtain the former case. This nally shows:
Theorem 13.3 The Jacobian of the hyperelliptic curve y
2
= x
5
1 is iso-
morphic to C
2
/, where is the ring Z[T]/(1+T +T
2
+T
3
+T
4
) embedded
in C
2
by
T (,
2
),
and is a primitive 5th root of unity.
This is an example of a Jacobian variety with complex multiplication.
The Picard group. Let Pic(X) denote the group of all line bundles on
X. Since every line bundle admits a meromorphic section, there is a natural
isomorphic between Pic(X) and Div(X)/(/

(X)), where /

(X) is the
group of nonzero meromorphic functions, mapping to principal divisors in
Pic(X). Under this isomorphism, the degree and a divisor and of a line
bundle agree.
The Abel-Jacobi map : Div
0
(X) Jac(X) is dened by

Q
i
P
i
_
() =

_
Q
i
P
i
.
Because of the choice of path from P
i
to Q
i
, the resulting linear functional
is well-dened only modulo cycles on X.
For example: given any basepoint P X, we obtain a natural map
f : X Jac(X) by f(Q) = (QP) =
_
Q
P
.
One of the most basic results regarding the Jacobian is:
Theorem 13.4 The map establishes an isomorphism between Jac(X) and
Pic
0
(X) = Div
0
(X)/(principal divisors).
82
The proof has two parts: Abels theorem, which asserts that Ker() coin-
cides with the group of principal divisors, and the Jacobi inversion theorem,
which asserts that is surjective.
Theorem 13.5 (Abels theorem) A divisor D is principal i D =

Q
i

P
i
and

_
Q
i
P
i
= 0
for all (X), for some choice of paths
i
joining P
i
to Q
i
.
Proof in one direction. Suppose D = (f). We can assume after multiply-
ing f by a scalar, that none of its critical values are real. Let = f
1
([0, ]).
Then we have

_
Q
i
P
i
=
_

=
_

0
f

() = 0
since f

() = 0, being a holomorphic 1-form on

C. (To see this, suppose
locally f(z) = z
d
. Then f

(z
i
dz) = 0 unless z
i
dz is invariant under rotation
by the dth roots of unity. This rst happens when i = d 1, in which case
z
d1
dz = (1/d)d(z
d
), so the pushforward is proportional to dz.)
The curve in its Jacobian. Before proceeding to the proof of Abels
theorem, we derive some consequences.
Given P X, dene

P
: X Jac(X)
by
P
(Q) = (Q P). Note that with respect to a basis
i
for (X), the
derivative of
P
(Q) in local coordinates is given by:
D
P
(Q) = (
1
(Q), . . . ,
g
(Q)).
This shows:
Theorem 13.6 The canonical map X P(X)

is the Gauss map of

P
.
Theorem 13.7 For genus g 1, the map
P
: X Jac(X) is a smooth
embedding.
Proof. If Q P = (f), then f : X P
1
has degree 1 so g = 0. Since [K[
is basepoint-free, there is a nonzero-holomorphic 1-form at every point, and
hence D
P
,= 0.
83
Theorem 13.8 (Jacobi) The map Div
0
(X) Jac(X) is surjective. In
fact, given (P
1
, . . . , P
g
) X
g
, the map
: X
g
Jac(X)
given by
(Q
1
, . . . , Q
g
) =
_

Q
i
P
i
_
=
__
Q
i
P
i

j
_
is surjective.
Proof. It suces to show that det D ,= 0 at some point, so the im-
age is open. To this end, just note that d/dQ
i
= (
j
(Q
i
)), and thus
det D(Q
1
, . . . , Q
g
) = 0 if and only if there is an vanishing simultane-
ously at all the Q
i
, i.e. i (Q
i
) lies on a hyperplane under the canonical
embedding. For generic Q
i
s this will not be the case, and hence det D ,= 0
almost everywhere on X
g
.
Corollary 13.9 We have a natural isomorphism:
Pic
0
(X) = Div
0
(X)/(/

(X))

= Jac(X).
Bergman metric. We remark that the space (X), and hence its dual,
carries a natural norm given by:
||
2
2
=
_
X
[(z)[
2
[dz[
2
.
This induces a canonical metric on Jac(X), and hence on X itself.
This metric ulimately comes from the intersection pairing or symplectic
form on H
1
(X, Z), satisfying a
i
b
j
=
ij
.
Abels theorem, proof I: The -equation. (Cf. Forster.) For the
converse, we proceed in two steps. First we will construct a smooth solution
to (f) = D; then we will correct it to become holomorphic.
Smooth solutions. Let us say a smooth map f : X

C satises (f) = D
if near P
i
(resp. Q
i
) we have f(z) = zh(z) (resp. z
1
h(z)) where h is a
smooth function with values in C

, and if f has no other zeros or poles.

Note that for such an f, the distributional logarithmic derivative satises
log f =
f
f
+

(p
i
q
i
),
84
where q
i
and p
i
are functions (in fact currents), locally given by log z,
and f/f is smooth.
Inspired by the proof in one direction already given, we rst construct
a smooth solution of (f) = D which maps a disk neighborhood U
i
of
i
dieomorphically to a neighborhood V of the interval [0, ].
Lemma 13.10 Given any arc joining P to Q on X, there exists a smooth
solution to (f) = QP satisfying
1
2i
_
X

f
f
=
_
Q
P

for all (X), where the integral is taken along .

Proof. First suppose P and Q are close enough that they belong to a single
chart U, and is almost a straight line. Then we can choose the isomorphism
f : U V

C so that f(P) = 0, f(Q) = and f() = [0, ] (altering
by a small homotopy rel endpoints). This f is already holomorphic on U,
and it sends U to a contractible loop in C

. Thus we can extend f to a

smooth function sending X U into C

, which then satises (f) = D.

Now note that z admits a single-valued logarithm on the region

C[0, ],
and thus log f(z) has a single-valued branch on Y = X . Thus f/f =
log f is an exact form on Y . However as one approaches from dierent
sides, the two branches of log f dier by 2i. Applying Stokes theorem, we
nd: _
X
log f = 2i
_

.
To handle the case of well-separated P and Q, simply break up into
many small segments and take the product of the resulting fs.
Taking the product of the solutions for several pairs of points, and using
additivity of the logarithmic derivative, we obtain:
Corollary 13.11 Given an arc
i
joining P
i
to Q
i
on X, there exists a
smooth solution to (f) =

Q
i
P
i
satisfying
1
2i
_
X

f
f
=

for all (X).

85
From smooth to holomorphic. To complete the solution, it suces to
nd a smooth function g such that fe
g
is meromorphic. Equivalently, it
suces to solve the equation g = f/f. (Note that f/f is smooth even
at the zeros and poles of f, since near there f = z
n
h where h ,= 0.) Since
H
0,1
(X)

= (X)

, such a g exists i
1
2i
_

f
f
=

_
Q
i
P
i
= 0
for all (X). This is exactly the hypothesis of Abels theorem.
Abels theorem, proof II: Symplectic forms. Our second proof makes
the connection with the Jacobian and the symplectic form on H
1
(X, C) more
transparent. (Cf. Lang, Algebraic Functions).
The symplectic form on H
1
. We begin with some topological remarks.
First, recall that a surface of genus g admits a basis for H
1
(X, Z) of the
form (a
i
, b
i
), i = 1, . . . , g, such that a
i
b
i
= 1 and all other products vanish.
In other words, the intersection pairing makes H
1
(X, Z) into a unimodular
symplectic space.
This symplectic form on H
1
(X, Z) gives rise to one on the space of peri-
ods, H
1
(X, C), dened by:
[, ] =

(a
i
)(b
i
) (b
i
)(a
i
).
In matrix form, we can associate a vector of periods (
a
,
b
) C
2g
to any
homomorphism , and then we have
[, ] = (
a
,
b
)
_
0 I
I 0
__

b
_
=

a

b

a

b
_

=
a

b

b

a
.
Theorem 13.12 Under the period isomorphism between H
1
DR
(X) and H
1
(X, C),
we have _
X
= [, ].
Proof. Cut X along the (a
i
, b
i
) curves to obtain a surface F with boundary,
on which we can write = df. Then we have
_
X
=
_
F
(df) =
_
F
f .
86
If we write F =

a
i
+b

i
a

i
b
i
, then
f[a

i
f[a
i
= (b
i
) and f[b

i
f[b
i
= (a
i
).
On the other hand, is the same along corresponding edges of F. Therefore
we nd
_
=

(a
i
)(b
i
) (b
i
)(a
i
).
Poincare duality. Note that for any cocycle N H
1
(X, Z), there is a dual
cycle C H
1
(X, Z) such that
[N, ] =
_
C

for all H
1
(X, C). In fact, since
[N, ] =

N(a
i
)(b
i
) N(b
i
)(a
i
),
we can simply take
C =

N(a
i
)b
i
N(b
i
)a
i
.
Since the intersection form is unimodular, this construction gives a natural
isomorphism
H
1
(X, Z)

= H
1
(X, Z).
Recognizing holomorphic 1-forms. Our second remark uses the Rie-
mann surface structure. Namely, we note that a class H
1
(X, C) can be
represented by a holomorphic 1-form i
[, ] = 0 (X).
This is because H
1
(X, C)

= (X) (X).
Construction of df/f. To try to construct f with (f) = D, we rst
construct a candidate for = df/f.
Theorem 13.13 For any divisor with deg D = 0, there exists a meromor-
phic dierential with only simple poles such that

Res
P
() P = D.
87
Proof. By Mittag-Leer for 1-forms, exists because the sum of its
residues is zero.
Alternatively, by Riemann-Roch, for any P, Q X we have dimH
0
(K+
P + Q) > dimH
0
(K). Thus there exists a meromorphic 1-form with a
simple pole at one of P or Q. By the residue theorem, has poles at both
points with opposite residues. Scaling proves the Theorem for QP, and
a general divisor of degree zero is a sum of divisors of this form.
Now if we could arrange that the periods of are all in the group 2iZ,
then f(z) = exp
_
would be a meromorphic function with (f) = D.
(Compare Mittag-Leers proof of Weierstrasss theorem on functions
with prescribed zeros.)
Periods of . As we have just seen, D determines a meromorphic form
with simple poles and
Res() =

P
Res
P
() P = D.
However is not uniquely determined by this condition it can be modied
by a form in (X). Furthermore, the periods of are not well dened
when we move a path across a pole of , it changes by an integral multiple
of 2i.
What we do obtain, canonically from D, is a class
[] H
1
(X, C)/(2iH
1
(X, Z) + (X)),
where (X) is embedded by its periods. And this class is zero i there is a
meromorphic function such that (f) = D.
Reciprocity. To make everything better dened, we now choose a sym-
plectic basis (a
i
, b
i
) as before, and cut X open along this basis to obtain a
region F. We also arrange that the poles of lie inside F. Then the periods
of become well-dened, by integrating along these specic representatives
a
i
, b
i
of a basis for H
1
(X, Z).
Now any (X) can be written as = df on F. On the other hand,
= 0 in F, at least if we exclude small loops around the poles of .
Integration around these loops gives residues. Thus we nd
0 =
_
F
=
_
F
f 2i

Res(fP
i
) + Res(fQ
i
)
= [, ] 2i

_
Q
i
P
i
,
88
where the path of integration is taken in F.
Now suppose (D) = 0. Then the sum above vanishes modulo the
periods of . In other words, there exists a cycle C such that
[, ] = 2i
_
C
= 2i[, N]
for some N H
1
(X, C). But this means [, N] = 0 for all , and hence
N is represented by a holomorphic 1-form. Modifying by this form,
we obtain [] = 2iN, and thus the periods of the are integral multiples
of 2i. Then f = exp(
_
) satises (f) = D.
Natural subvarieties of the Jacobian. Just as we can embed X into
Jac(X) (or more naturally into Pic
1
(X)), we can map X
(k)
= X
k
/S
k
to
Jac(X) (or into Pic
k
(X)). By what we have just seen, the bers of this map
consists of linearly equivalent divisors, and the dimensions of the bers are
predicted by Riemann-Roch. In particular, we have:
Theorem 13.14 The bers of the natural map X
(k)
Pic
k
(X) are projec-
tive spaces.
The images of these maps in Jac(X) (which are well-dened up to trans-
lation) are usually denoted W
k
; in particular, we obtain an important divi-
sors W
g1
in this way, while W
g
= Jac(X).
Mordells Conjecture.
Theorem 13.15 Suppose X has genus g 2. Then, given any nitely
generated subgroup A Jac(X), the set X A is nite.
This theorem is in fact equivalent to Mordells conjecture (Faltings the-
orem), which states that X(K) is nite for any number eld K.
The exponential sequence, the Picard group and the Jacobian. An
alternative description of the Jacobian is via the exponential sequence which
leads to the exact sequence
H
1
(X, Z) H
1
(X, O) H
1
(X, O

) H
2
(X, Z) 0.
Under the isomorphisms H
1
(X, Z)

= H
1
(X, Z) by cup product, H
1
(X, O)

=
(X)

by Serre duality, and H

2
(X, Z)

= Z by degree, we obtain the isomor-
phism
Pic
0
(X) = Ker(H
1
(X, O) H
2
(X, Z))

= (X)

/H
1
(X, Z) = Jac(X).
89
The norm map. We note that if f : X Y is a nonconstant holomor-
phic map, we have a natural map on divisors and hence an induced map
Jac(X) Jac(Y ) and even a pushforward map on line bundles, Pic(X)
Pic(Y ). We also have a norm map N : /

(X) /(Y )

, given by taking
the product over the bers. Note that:
D D

= (g) = f(D) f(D

) = (N(g)).
For this to be correct, the points of f(D) must be taken with multiplicity
according to the branching of f. In particular, deg f(P) is not generally a
continuous function of P X.
Recalling that a divisor is represented by a cochain g
i
/

(U
i
) with
(g)
ij
O

(U
ij
), the pushforward of divisors can be dened by the local
norm map as well.
The same construction works for equidimensional maps between complex
manifolds of higher dimension. In this case f(D) = 0 if the image of f is a
lower-dimensional variety.
The Siegel upper half-space. The Siegel upper half-space is the space H
g
of symmetric complex g g matrices P such that ImP is positive-denite.
The space H
g
is the natural space for describing the g-dimensional com-
plex torus Jac(X), just as H is the natural space for describing the 1-
dimensional torus C/.
To describe the Jacobian via H
g
, we need to choose a symplectic basis
(a
i
, b
i
) for H
1
(X, Z).
Theorem 13.16 There exists a unique basis
i
of (X) such that
i
(a
j
) =

ij
.
Proof. To see this we just need to show the map from (X) into the space
of a-periods is injective (since both have dimension g. But suppose the
a-periods of vanish. Then the same is true for the (0, 1)-form , which
implies
_
[(z)[
2
dz dz =
_
=

(a
i
)(b
i
) (b
i
)(a
i
) = 0,
and thus = 0.
Denition. The period matrix of X with respect to the symplectic basis
(a
i
, b
i
) is given by

ij
=
_
b
j

i
=
i
(b
j
).
90
Theorem 13.17 The matrix is symmetric and Im is positive-denite.
Proof. To see symmetry we use the fact that for any i, j:
0 =
_

i

j
=

i
(a
k
)
j
(b
k
)
i
(b
k
)
j
(a
k
) =
ij

jk

ik

jk
=
ji

ij
.
Similarly, we have
i
2
_
=
_
[(z)[
2
[dz[
2
0.
Thus, if we let
Q
ij
=
i
2
_

i

j
=
i
2

i
(a
k
)
j
(b
k
)
i
(b
k
)
j
(a
k
)
=
i
2
(
ji

ij
) = Im
ij
,
then

Q
ij
c
i
c
j
=
_

[c
i
[
2
[
i
[
2
,
and thus Im is positive-denite.
The symplectic group Sp
2g
(Z) now takes over for SL
2
(Z), and two com-
plex tori are isomorphic as principally polarized abelian varieties if and only
if there are in the same orbit under Sp
2g
(Z).
With respect to the basis (b
1
, . . . , b
g
, a
1
, . . . , a
g
) the symplectic group
acts on the full matrix of a and b periods by
_
A B
C D
__

I
_
=
_
A +B
C +D
_
.
We must then change the choice of basis for (X) to make the a-periods,
C +D into the identity matrix; and we nd
g() = (A +B)(C +D)
1
,
which shows Sp
2g
(Z) acts on H
g
by non-commutative fractional linear trans-
formations.
A non-Jacobian. We note that for g 2 there are period matrices which
do not arise from Jacobians. The simplest example comes from the rank 2
diagonal matrices

ij
=
_
t
1
0
0 t
2
_
91
with t
i
H (which, incidentally, show there is a copy of H
g
in H
g
). It is
essential, here, that the matrix above is with respect to a symplectic base.
To see this matrix cannot arise, suppose in fact
ij
= Jac(X). Note that
Per(
1
) = ZZt
1
=
1
and
_
a
2

1
=
_
b
2

1
= 0. Thus by integrating
1
we
get a map f : X E = C/
1
such that f

(dz) =
1
. But then
_
X
[f

(dz)[
2
=
i
2
_
X

1

1
= (i/2)(
1
(b
1
)
1
(b
1
)) = Im(b
1
) =
_
E
[dz[
2
,
and hence deg(f) = 1, which is impossible for a holomorphic map between
Riemann surfaces of dierent genera.
In fact A = C
2
/Z
2
Z
2
can be realized as the Jacobian of a stable curve,
E
1
E
2
. And as a complex torus, A can be realized as a Jacobian in fact,
in genus two, a complex torus is a Jacobian unless it splits symplectically as
a product.
A dimension count shows that in genus g 4, there are principally
polarized Abelian varieties which cannot be realized by stable curves.
The innitesimal Torelli theorem. A beautiful computation of Ahlfors
gives the explicit derivative of the map /
g
/
g
provided by X Jac(X).
More precisely, the coderivative of this map sends the cotangent space at a
given periodic matrix (
ij
) to the cotangent space to /
g
at X, which is
naturally the space Q(X) of holomorphic quadratic dierentials on X.
The map is given simply by
d
ij
=
i

j
Q(X).
Now we have the following basic fact:
Theorem 13.18 (Max Noether) The natural map (X)(X) Q(X)
is surjective i X is not hyperelliptic.
Corollary 13.19 (Local Torelli theorem) Away from the hyperelliptic
locus, the map /
g
/
g
is an immersion.
Noethers theorem can be rephrased as a property of the canonical curve
X P
g1
the complete linear series of quadrics in P
g1
restricts to a
complete linear series on X. In fact, any canonical curve is projectively
normal see Griths and Harris.
Ahlfors formula: sketch of the proof. A new complex structure on
X is specied by a (small) Beltrami dierential = (z)dz/dz. Then the
complex cotangent space is generated at each point, not by dz, but by
dz

= (1 +) dz = dz +(z)dz.
92
In this way we obtain a new Riemann surface X

. We let
i
denote the
basis for (X) with
i
(a
j
) =
ij
, and similarly for

i
. We have

i
=

i
(z)(dz +(z)dz).
Let k = sup[[ which is assumed to be small. Since
i

i
has vanishing
a-periods, this form has self-intersection number zero in cohomology. Thus
0 =
i
2
_
X
(
i

i
) (
i

i
).
This yields
_
[
i

i
[
2
[dz[
2
=
_
[[
2
[

i
[
2
[dz[
2
,
which implies

i
=
i
+O(k).
A simple computation also gives

ij

ij
=
_
X

i
=
_
X

j
+O(k
2
).
Thus the variation in
ij
is given, in the limit, by pairing with
i

j
.
Theta functions. The theory of -functions allows one to canonically
attach a divisor
A = C
g
/(Z
g
Z
g
)
to the principally polarized Abelian variety A determined by H
g
. Using
the fact that Im 0, we dene the entire -function : C
g
C by
(z) =

nZ
g
exp(2in, z) exp(in, n).
The zero-set of is -invariant, and descends to a divisor on A.
For example, when g = 1 and H, we obtain:
(z) =

Z
exp(2inz) exp(in
2
).
Clearly (z + 1) = (z), and we have
(z +) = (z) exp(2iz i).
93
A useful notation for g = 1 is to set q = exp(i) and = exp(2iz);
then
() =

nZ
q
n
2

n
,
and we have
(q
2
) = (q)
1
().
Zeros and poles. It turns out that has a unique zero on X = C/ZZ =
C

/q
2Z
. Indeed, we have
(q) =

q
n
2
+n
(1)
n
= q
1/4

q
(n+1/2)
2
(1)
n
= 0,
since the terms q
m
2
and q
(m)
2
, m Z + 1/2, occur with opposite signs.
A general meromorphic function on X can be expressed as a suitable
product of translates of functions. Indeed,
f
a
() =
n

1
(q/a
i
)
has zeros at a
i
, and saties
f
a
(q
2
) = (q
2
)
n
n

1
(a
i
).
Thus if

n
1
a
i
=

n
1
b
i
, the function
F() =
f
a
(z)
f
b
(z)
is meromorphic on X, with zeros at (a
i
) and poles at (b
i
). The condition
that the products agree guarantees that the divisor of (F) maps to zero
under the Abel-Jacobi map.
Theta functions for g 2. A convenient notation for functions in
higher genus is the following: we set
= (
1
, . . . ,
g
) = (exp 2iz
i
) (C

)
g
,
q = q
ij
= exp(i
ij
) exp(iH
g
),
n = (n
1
, . . . , n
g
) Z
g
,

n
=

n
i
i
, and
q
n
2
=

i,j
q
n
i
n
j
ij
.
94
Then we set
() =

nZ
g
q
n
2

n
,
and nd for all k Z
g
:
(q
2k
) =

q
n
2
q
2nk

n
= q
k
2

q
(n+k)
2

n+k
= q
k
2

k
().
Noting that (z
i
+

m
i
+

ij
k
j
) = q
2k
, this shows that = () is a
divisor on A = C
g
/(Z
g
Z
g
).
Theorem 13.20 (Riemann) We have = W
g1
+ for some Jac(X),
where W
g1
Jac(X) is the image of X
g1
under the Abel-Jacobi map.
In particular, for g = 2 the divisor is isomorphic to X itself. Note
however that Theta makes sense even when A is not a Jacobian.
The Torelli theorem.
Theorem 13.21 Suppose (Jac(X), ) and (Jac(Y ),

) are isomorphic as
(principally) polarized complex tori (Abelian varieties). Then X is isomor-
phic to Y .
Sketch of the proof. Using the polarization, we can reconstruct the divisor
W
g1
Jac(X) up to translation, using -functions.
Now the tangent space at any point to Jac(X) is canonically identied
with (X)

, and hence tangent hyperplanes in Jac(X) give hyperplanes in

P(X), the ambient space for the canonical curve. (For convenience we will
assume X is not hyperelliptic.) In particular, the Gauss map on the smooth
points of W
g1
denes a natural map
: W
g1
P(X)

= P
g1
.
The composition of with the natural map C
g1
W
g1
simply sends g1
points P
i
on the canonical curve to the hyperplane H(P
i
) they (generically)
span.
Clearly this map is surjective. Since W
g1
and P
g1
have the same
dimension, is a local dieomorphism at most points. The branch locus
corresponds to the hyperplanes that are tangent to the canonical curve X
P
g1
. Thus from (Jac(X), ) we can recover the collection of hyperplanes
X

tangent to X P
g1
.
Finally one can show geometrically that X

= Y

implies X = Y . The
idea is that, to each point P X we have a (g 3) dimensional family of
95
hyperplanes H
P
X

containing the tangent line T

P
(X). These hyper-
planes must all be tangent to Y as well; but the only reasonable way this
can happen is if T
P
(X) = T
Q
(Y ) for some Q on Y .
Now for genus g > 3 one can show a tangent line meets X in exactly one
point; thus Q is unique and we can dene an isomorphism f : X Y by
f(P) = Q. (For example, in genus 4 the canonical curve is a sextic in P
3
.
If a tangent line L P
3
were to meet 2P and 2Q, then the planes through
L would give a complementary linear series of degree 2, so X would be
hyperelliptic.) For g = 3 there are in general a nite number of bitangents
to the quartic curve X P
2
, and away from these points we can dene f,
then extend. (For details see Griths and Harris).
Explicit Torelli theorem in genus 2. In the case of genus two, the
Riemann surface X can be reconstructed directly from its period matrix
as the zero locus of the associated theta function. As remarked above, the
Gauss map then gives the canonical map from X to P
1
. To give X as
an algebraic curve it suces to determine the six branch points B of the
canonical map. But these are nothing more than the images under the
Gauss map of the six odd characteristics. In brief, we nd:
Theorem 13.22 If g(X) = 2, then the 6 branch points of the canonical map
X P
1
correspond to the values of d at the 6 odd theta characteristics.
Spin structures in genus two. A spin structure on X is the choice of
a square-root of the canonical bundle. Equivalently it is a divisor class [D]
such that 2D K. The parity of a spin structure is dened to be the parity
of h
0
(D).
The spin structures on X lie naturally in Pic
g1
(X) and form a homo-
geneous space for Jac(X)[2]

= F
2g
2
. The parity is, in fact, determined on
these points by the polarization of the Jacobian.
In the case of genus 2, it is easy to see that there are 6 odd spin structures,
one for each Weierstrass point P
i
X. In fact the dierences P
i
P
j
of
Weierstrass points span Jac(X)[2].
To see this more clearly, rst note that 2P
i
2P
j
K for any i and j.
Thus if we take any (
i
) F
6
2
with

i
= 0 mod 2, and lift to integers such
that

i
= 0, then
() =
6

i
P
i
Jac(X)[2]
96
is well-dened. We also note that D = [5P1

6
2
P
i
] is a princpal divisor;
indeed, D = (y) where y
2
= f(x) and deg f = 5. Thus (1, 1, 1, 1, 1, 1) = 0.
Taking the quotient of the space of (
i
) with

i
= 0 mod 2 by subspace
(1, 1, 1, 1, 1, 1), we obtain Jac(X)[2]

= F
4
2
.
14 Hyperbolic geometry
(The B-side.)
Elements of hyperbolic geometry in the plane. The hyperbolic metric
is given by = [dz[/y in H and = 2[dz[/(1 [z[
2
) in .
Thus d(i, iy) = log y in H, and d(0, x) = log(x + 1)/(x 1) in . Note
that (x + 1)/(x 1) maps (1, 1) to (0, ).
An important theorem for later use gives the hyperbolic distance from
the origin to a hyperbolic geodesic which is an arc of a circle of radius r:
sinhd(0, ) =
1
r

To see this, let x be the Euclidean distance from 0 to . Then we have, by

algebra,
sinhd(0, x) =
2x
1 x
2
(14.1)
On the other hand, we have by a right triangle with sides 1, r and x + r.
Thus 1 +r
2
= (x +r)
2
which implies 2x/(1 x
2
) = 1/r.
A more intrinsic statement of this theorem is that for any point p H
and geodesic H, we have
sinhd(p, ) = cot(/2),
where is the visual angle subtended by as seen from p.
Area of triangles and polygons. The area of an ideal triangle is . The
area of a triangle with interior angles (0, 0, ) is . From these facts one
can see the area of a general triangle is given by the angle defect:
T(, , ) = .
To see this, one extends the edges of T to rays reaching the vertices of an
ideal triangle I; then we have
T(, , ) = I T( ) T( ) T( )
which gives for the area.
97
Another formulation is that the area of a triangle is the sum of its exterior
angles minus 2. In this form the formula generalizes to polygons.
Right quadrilaterals with an ideal vertex.
Theorem 14.1 Let Q be a quadrilateral with edges of lengths (a, b, , )
and interior angles (/2, /2, /2, 0). Then we have
sinh(a) sinh(b) = 1.
Proof. First we make a remark in Euclidean geometry: let Q

be an ideal
hyperbolic quadrilateral, centered at the origin, with sides coming from cir-
cles of Euclidean radii (r, R, r, R). Then rR = 1.
Indeed, from this picture we can construct a right triangle with right-
angle vertex 0, with hypotenuse of length r +R, and with altitude from the
right-angle vertex of 1. By basic Euclidean geometry of similar triangles, we
nd rR = 1
2
= 1.
Now cut Q

into 4 triangles of the type Q. Then we have sinh(a) = 1/r

and sinh(b) = 1/R, by (14.1). Therefore sinh(a) sinh(b) = 1.
Right hexagons. For a right-angled hexagon H, the excess angle is
6(/2) 2 = , and thus area(H) = .
Theorem 14.2 For any a, b, c > 0 there exists a unique right hexagon with
alternating sides of lengths (a, b, c).
Proof. Equivalently, we must show there exist disjoint geodesics , , in
H with a = d(, ), b = d(, ) and c = d(, ). This can be proved by
continuity.
Normalize so that is the imaginary axis, and choose any geodesic at
distance a from . Then draw the parallel line L of constant distance b
from , on the same side as . This line is just a Euclidean ray in the upper
half-plane. For each point p L
p
there is a unique geodesic
p
tangent to
L
p
at p, and consequently at distance b from .
Now consider f(p) = d(
p
, ). Then as p moves away from the juncture
of and L, f(p) decreases from to 0, with strict monotonicity since
p
L
separates from
q
. Thus there is a unique p such that f(p) = c.
98
L

p
Figure 2. Right hexagons.
Doubling H along alternating edges, we obtain a pair of pants P. Thus
area(P) = 2.
Corollary 14.3 Given any triple of lengths a, b, c > 0, there exists a pair of
pants, unique up to isometry, with boundary components of lengths (a, b, c).
Pairs of pants decomposition.
Theorem 14.4 Any essential simple loop on a compact hyperbolic surface
X is freely homotopic to a unique simple geodesic. Any two disjoint simple
loops are homotopic to disjoint simple geodesics.
Corollary 14.5 Let X be a compact surface of genus g. Then X can be
cut along 3g 3 simple geodesics into 2g 2 pairs of pants. In particular,
we have
area(X) = 2[(X)[.
Parallels of a geodesic. There is a nice parameterization of the geodesic
[z[ = 1 in H: namely
(t) = tanh t +i sech t.
We have |

(t)| = 1 in the hyperbolic metric.

Now given a closed simple geodesic on X, let C(, r) be a parallel
curve at distance r from . Then we have:
L(C(), r) = L() cosh(r).
99
Indeed, let can be covered by the imaginary axis iR
+
in H. Then C(, r)
is a ray from 0 to which passes through (r). Thus the Euclidean slope
of C(, r) is the same as that of the vector (x, y) = (tanh t, sech t). Thus
projection along Euclidean horizontal lines from C(, r) to contracts by
a factor of y/
_
x
2
+y
2
= sech(t). Therefore C(, r) is longer than by a
factor of cosh(t).
The collar lemma.
Theorem 14.6 Let and be disjoint simple geodesics on a compact Rie-
mann surface, of lengths a and b respectively. Dene A and B by sinh(a/2) sinh(A) =
1 and sinh(b/2) sinh(B) = 1. Then the collars of widths A and B about
and are disjoint.
Proof. We can assume that and are part of a pants decomposition of X,
which reduces the result to the case where and are two cus of a pants
P. By the Schwarz lemma, we can assume the lengths of the boundaries of
P are (a, b, 0).
Now cut P along a simple loop that begins and ends at its ideal
boundary component, i.e. the cu of length zero. Then the components of
P = are doubles of quadrilaterals with one ideal vertex and the remaining
angles /2. The quadrilateral meeting has nite sides of lengths a/2
and A satisfying sinh(a/2) sinh(A) = 1, and similarly for the quadrilateral
containing . Thus these collars are disjoint.
Boundary of a collar.
Theorem 14.7 The length of each component of the boundary of the stan-
dard collar around with L() = a satises
L(C(, r))
2
= a
2
cosh
2
(r) =
a
2
1 + sinh
2
(a/2)
4
as a 0. Thus the length of each component of the collar about tends to
2 as L() 0.
Proof. Apply the preceding formulas.
Check. The limiting case is the triply-punctured sphere, the double of
an ideal triangle T H with vertices (1, 1, ). The collars limit to the
horocycles given by the circles of radius 1 resting on 1 together with the
horizontal line segment H at height 2 running from 1 +2i to +1 +2i. We
have L(H) = 1, so upon doubling we obtain a collar boundary of length 2.
100
Corollary 14.8 The collars about short geodesics on a compact hyperbolic
surface cover the thin part of the surface.
Proof. Suppose x X lies in the thin part that is, suppose there is a
short essential loop through x. Then is homotopic to a closed geodesic ,
which is necessarily simple. But since is short, we see by the result above
that must lie in the collar neighborhood of .
Thick-thin decomposition. There is a universal constant r > 0 such that
any compact Riemann surface of genus g can be covered by a collection of
O(g) balls B(x
i
, r) and O(g) standard collars about short geodesics.
Bers constant.
Theorem 14.9 There exists a constant L
g
such that X admits a pants
decomposition with no cu longer than L
g
.
Theorem 14.10 We can take L
g
= O(g), but there exist examples requiring
at least one curve of length > C

g > 0.
A nite-to-one map to M
g
.
Theorem 14.11 For each trivalent graph of G with b
1
(G) = g, there is a
nite-to-one map

G
: ((0, L
g
] S
1
)
3g3
/
g
,
sending (r
i
,
i
) to the surface obtained by gluing together pants with cus of
lengths r
i
and twisting by
i
, using pants and cus corresponding to vertices
and edges of G.
The union of the images of the maps
G
is all of /
g
.
Corollary 14.12 (Mumford) The function L : /
g
R sending X to
the length L(X) of its shortest geodesic is proper.
The Laplacian. Let M be a Riemannian manifold. The Laplace operator
: C

0
(M) C

0
(M) is dened so that
_
M
[f[
2
=
_
M
ff,
both integrals taken with respect to the volume element on M.
101
For example, on R we nd f = d
2
f/dx
2
by integrating by parts.
Similarly on R
n
we obtain
f =

d
2
f
dx
2
i

Note this is the negative of the traditional Laplacian.

In terms of the Hodge star we can write
_
f, f =
_
df df =
_
f d df =
_
f( d df) dV,
and therefore we have
f = d df.
For example, on a Riemann surface with a conformal metric (z)[dz[, we
have dx = dy, dy = dx, and
f =
2
(z)
_
d
2
f
dx
2
+
d
2
f
dy
2
_
.
As a particular case, for = [dz[/y on H we see y

= (1)y

, showing
that y

is an eigenfunction of the hyperbolic Laplacian.

The heat kernel. Let X be a compact hyperbolic surface. Enumerat-
ing the eigenvalues and eigenfunctions of the Laplacian, we obtain smooth
functions satisfying

n
=
n

n
,

n
0. The heat kernel K
t
(x, y) is dened by
K
t
(x, y) =

e
nt

n
(x)
n
(y),
The heat kernel is the fundamental solution to the heat equation. That
is, for any smooth function f on X, the solution to the heat equation
df
t
dt
= f
t
with initial data f
0
= f is given by f
t
= K
t
f. Indeed, if f(x) =

a
n

n
(x)
then
f
t
(x) = K
t
f =

a
n
e
nt

n
(x)
clearly solves the heat equation and has f
0
(x) = f(x).
Note also that formally, convolution with K
t
is the same as the operator
exp(), which acts by exp(
n
) on the
n
-eigenspace.
102
Brownian motion. The heat kernel can also be interpreted using diu-
sion; namely, K
t
(x, ) denes a probability measure on X that gives the
distribution of a Brownian particle x
t
satisfying x
0
= x.
For example, on the real line, the heat kernel is given by
K
t
(x) =
1

4t
exp(x
2
/(4t)).
Also we have K
s+t
= K
s
K
t
, as bets a Markov process.
To check this, note that K
t
solves the heat equation, and that
_
K
t
= 1
for all t. Thus K
t
f f as t 0, since K
t
concentrates at the origin.
In terms of Brownian motion, the solution to the heat equation is given
by f
t
(x) = E(f
0
(x
t
)), where x
t
is a random path with x
0
= x.
The trace. The trace of the heat kernel is the function
Tr K
t
=
_
X
K
t
(x, x) =

e
nt
.
It is easy to see that the function Tr K
t
determines the set of eigenvalues
n
and their multiplicities.
Length spectrum and eigenvalue spectrum.
Theorem 14.13 The length spectrum and genus of X determine the eigen-
values of the Laplacian on X.
Proof. The proof is based on the trace of the heat kernel. Let k
t
(x, y)
denote the heat kernel on the hyperbolic plane H; it satises k
t
(x, y) = k
t
(r)
where r = d(x, y). Then for X = H/ we have
K
t
(x, y) =

k
t
(x, y),
where we regard K
t
as an equivariant kernel on H.
Working more intrinsically on X, we can consider the set of pairs (x, )
where is a loop in
1
(X, x). Let
x
() denote the length of the geodesic
representative of based at x. Then we have:
K
t
(x, x) =

k
t
(
x
()).
Let /(X) denote the space of nontrivial free homotopy classes of maps
: S
1
X. For each /(X) we can build a covering space p : X

X
corresponding to
1
(X).
103
The points of X

correspond naturally to pairs (x, ) on X with freely

homotopic to . Indeed, given x

in X

, there is a unique homotopy class

of loop

through x

that is freely homotopic to , and we can set (x, ) =

(p(x), p(

)). Conversely, given (x, ), from the free homotopy of to

we obtain a natural homotopy class of path joining x to , which uniquely
determines the lift x

of x to X

.
For x

, let r(x

) denote the length of the unique geodesic through

x

that is freely homotopic to . Then we have

x
() = r(x

). It follows that
Tr K
t
=
_
X
K
t
(x, x) =
_
X
k
t
(0) +

L(X)
_
X
r(x

).
But
_
X
k
t
(0) = area(X)k
t
(0) depends only on the genus of X by Gauss-
Bonnet, and the remaining terms depend only on the geometry of X

. Since
the geometry of X

is determined by the length of , we see the length spec-

trum of X determines the trace of the heat kernel, and hence the spectrum
of the Laplacian on X.
Remark. Almost nothing was used about the heat kernel in the proof.
Indeed, the length spectrum of X determine the trace of any kernel K(x, y)
on X derived from a kernel k(x, y) on H such that k(x, y) depends only on
d(x, y).
Remark. In fact the genus is determined by the length spectrum.
Isospectral Riemann surfaces.
Theorem 14.14 There exist a pair of compact hyperbolic Riemann surfaces
X and Y , such that the length spectrum of X and Y agree (with multiplici-
ties), but X is not isomorphic to Y .
Isospectral subgroups. Here is a related problem in group theory. Let G
be a nite group, and let H
1
, H
2
be two subgroups of G. Suppose [H
1
C[ =
[H
2
C for every conjugacy class C in G. Then are H
1
and H
2
conjugate
in G?
The answer is no in general. A simple example can be given inside
the group G = S
6
. Consider the following two subgroups inside A
6
, each
isomorphic to (Z/2)
2
:
H
1
= e, (12)(34), (12)(56), (34)(56),
H
2
= e, (12)(34), (13)(24), (14)(23).
104
Note that the second group actually sits inside A
4
; it is related to the sym-
metries of a tetrahedron.
Now conjugacy classes in S
n
correspond to permutations of n, i.e. cycle
structures of permutations. Clearly [H
i
C[ = 3 for the cycle structure
(ab)(cd), and [H
i
C[ = 0 for other conjugacy classes (except that of the
identity). Thus H
1
and H
2
are isospectral. But they are not conjugate
(internally isomorphic), because H
1
has no xed-points while H
2
has two.
Construction of isospectral manifolds.
Theorem 14.15 (Sunada) Let X Z be a nite regular covering of com-
pact Riemannian manifolds with deck group G. Let Y
i
= X/H
i
, where H
1
and H
2
are isospectral subgroups of G. Then Y
1
and Y
2
are also isospectral.
Proof. For simplicity of notation we consider a single manifold Y = X/G
and assume the geodesics on Z are discrete, as is case for a negatively curved
manifold. Every closed geodesic on Y lies over a closed geodesic on Z.
Fixing a closed geodesic on Y , we will show the set of lengths / of
geodesics on Y lying over depends only on the numbers n
C
= H C for
conjugacy classes C in G.
For simplicity, assume has length 1. Let
1
, . . . ,
n
denote the com-
ponents of the preimage of on X. Let S
i
G be the stabilizer of
i
. The
subgroups S
i
ll out a single conjugacy class in G, and we have S
i

= Z/m
where nm = [G[. Each loop
i
has length m.
Let k be the index of H S
i
in S
i
. Then k is the length of
i
/H in Y .
Moreover, the number of components
j
in the orbit H
i
is [H[/[HS
i
[ =
[H[[S
i
[/k, and of course all these components descend to a single loop on
X/H. Thus the number of times the integer k occurs in / is exactly
[/(k)[ =
kA
k
m[H[
,
where
A
k
= [i : [S
i
: S
i
H] = k[.
Thus to determine /, it suces to determine the integers A
k
.
For example, let us compute A
1
, the number of i such that we have
S
i
H. Now H contains S
i
if and only if H contains a generator g
i
of S
i
.
We can choose the g
i
s to ll out a single conjugacy class C, since the groups
S
i
are all conjugate. Then the proportion of is satisfying S
i
H is exactly
[H C[/[C[, and therefore
A
1
=
n[H C[
[C[

105
An important point here: it can certainly happen that S
i
= S
j
even
when i ,= j. For example if G is abelian, then all the groups S
i
are the
same. But the number of i such that S
i
is generated by a given element
g C is a constant, independent of g. Thus the proportion of S
i
generated
by an element of H is still [H C[/[C[.
Now for d[m, let C
d
be the dth powers of the elements in C. Then the
subgroups of index d in the S
i
s are exactly the cyclic subgroups generated
by elements g C
d
. Again, the correspondence is not exact, but constant-
to-one; the number of i such that g S
i
is independent of g C
d
. Thus
the proportion of S
i
s such that H S
i
contains a subgroup of index d is
exactly [H C
d
[/[C
d
[, which implies:

k|d
A
k
=
n[H C
d
[
[C
d
[

From these equations it is easy to compute A
k
.
Cayley graphs. The spaces in Sunadas construction do not have to be
Riemannian manifolds. For example, we can take Z to be a bouquet of
circles. Then X is the Cayley group of G, and Y
1
and Y
2
are coset graphs
on which G acts. The coset graphs Y
1
and Y
2
also have the same length
spectrum!
A small example. Let G = (Z/8)

Z/8 be the ane group of A = Z/8,

i.e. the group of invertible maps f : A A of the form f(x) = ax +b. Let
H
1
= x, 3x, 5x, 7x,
H
2
= x, 3x + 4, 5x + 4, 7x.
Then the subgroups H
1
and H
2
are isospectral.
In both cases, the coset space Y
i
= G/H
i
can be identied with Z/8;
that is, Z/8 H
i
= G.
To make associated graphs, Y
1
and Y
2
, we take x + 1, 3x, 5x as gener-
ators for G. Note that 3x and 5x have order 2. Then the coset graph Y
1
is
an octagon, coming from the generator x + 1, with additional (unoriented,
colored) edges joining x to 3x and 5x. Similarly, Y
2
is also an octagon, but
now the colored edges join x to the antipodes of 3x and 5x, namely 3x + 4
and 5x + 4.
These graphs are isospectral. In counting the number of loops, it is
important to regard the graphs as covering spaces. For this it is best to
replace each colored edge which is not a loop by a pair of parallel edges with
106
Figure 3. Isospectral graphs
opposite arrows. Each colored loop should be replaced by a single oriented
edge. Then the graphs become covering spaces of the bouquet of 3 circles,
and the number of loops of length n is the same for both graphs.
Not isometric. Using short geodesics, we can arrange Z such that one
can reconstruct the action of G on G/H
i
from the intrinsic geometry of Y
i
.
Then Y
1
and Y
2
are isometric i H
1
and H
2
are conjugate. So in this way
we obtain isospectral, but non-isometric, Riemann surfaces.
15 Uniformization
Theorem 15.1 Every simply connected Riemann surface is isomorphic to

C, C or H.
Poincare series. We will now show that the uniformization theorem im-
plies the Riemanne existence theorem; for example, it implies that every
compact Riemann surface can be presented as a branched cover of

C.
For surfaces of genus 0 this is clear. For genus one, uniformization implies
X =

C/ and then the construction of the Weierstrass -function completes
the proof.
For surfaces of higher genus we have X = /. Here the idea is the
following: given a meromorphic form = (z) dz
k
on the unit disk, we
generate a -invariant form by summation (just as one would do for the
-function):
107
() =

.
More explicitly, this form is given by
() =

((z))(cz +d)
2k
dz
k
,
since

(z) = (cz +d)

2
when (z) = (az +b)/(cz +d) is normalized so that
ad bc = 1.
There is no chance that this sum will converge when is a nonzero
function. However, when k = 2, [[ is naturally an area form, and hence
_
X
[()[
_

[[.
This shows the sum converges (and denes a meromorphic form on X) so
long as
_
[[ < . In fact, it is enough that
_
[[ is nite outside a compact
set K .
A second problem is that () might be zero. Indeed, if =

,
then the sum will telescope and give zero. As in the case of the function,
we can insure the sum is nonzero by introducing a pole. That is, if a
projects to p X, then

p
=
_
dz
2
z a
_
gives a meromorphic quadratic dierential with a simple pole at p and else-
where holomorphic on X. It follows that for p ,= q,
f
pq
=
p
/
q
denes a nonconstant meromorphic function on X; indeed, f
pq
(p) = and
f
pq
(q) = 0.
The multiplicities of these zeros and poles is not quite clear; for example,

q
might vanish at p, and then f
pq
has at least a double order pole at p.
But if p is close enough to q, then
q
(p) ,= 0, and hence f
pq
provides a local
chart near p.
With a little more work we have established:
Corollary 15.2 Every Riemann surface carries a nonconstant meromor-
phic function, indeed, enough such functions to separate points.
108
Quasiconformal geometry Here is another approach to the uniformiza-
tion theorem. For any on

C with ||

< 1, there exists a quasiconformal

map f :

C

C with complex dilatation .
Evidentally we can uniformize at least one Riemann surface X
g
of genus
g, e.g. using a regular hyperbolic 4g-gon. Now take any other surface Y
of the same genus. By topology, there is a dieomorphism f : X
g
Y .
Pulling back the complex structure to X
g
and lifting to the universal cover,
we obtain by qc conjugacy a Fuchsian group uniformizing Y .
References
[EKS] A. Edmonds, R. Kulkarni, and R. E. Stong. Realizability of branched
coverings of surfaces. Trans. Amer. Math. Soc. 282(1984), 773790.
[Hel] S. Helgason. Dierential Geometry, Lie Groups, and Symmetric
Spaces. Academic Press, 1978.
[KZ] M. Kontsevich and D. Zagier. Periods. In Mathematics Unlimited
2001 and Beyond, pages 771808. Springer-Verlag, 2001.
[La] S. Lang. Algebra. Addison-Wesley, 1984.
109