Вы находитесь на странице: 1из 22

Chapter 9

Cohomology and Spectral


Sequences

This appendix gives a short but intense introduction to cohomology and spectral
sequences, two powerful but sometimes intimidating topics. It is included mainly
for the convenience of the reader who is a non-specialist, so contains basic definitions and theorems (sometimes without proof) as well as illustrative examples.
In particular, after reading the chapter, we hope the reader will feel competent to
compute
Higher derived functors (cohomology from a new viewpoint).
Spectral sequences.
Comprehensive treatments of the material here (and proofs) may be found in Eisenbud [11], Hartshorne [19], and Weibel [?] Throughout this section, the ring R is a
commutative, finitely generated Calgebra, and the sheaf OX is the sheaf of regular
functions on a complex variety X.

0. Homological basics: Complexes and Resolutions


A sequence of Rmodules and homomorphisms
C :

j+2

/ Mj+1

j+1

/ Mj

/ Mj1

j1

is a complex (or chain complex) if


imj+1 ker j .
The sequence is exact at Mj if imj+1 = ker j ; a complex which is exact everywhere is called an exact sequence. The j th homology module of C is defined
381

9. Cohomology and Spectral Sequences

382

as:
Hj (C ) = ker j /imj+1 .
The following complex is ubiquitous:
Example 0.1. [Koszul complex] Let S = C[x1 , .P
. . , xn ], with f1 , . . . , fm S.
Set V = S m with basis {e1 , . . . , em }, and let f = m
i=1 fi ei . The sequence:
0

/S

dn

/ 1 (V )

dn1

/ 2 (V )

d1

/ n (V )
=S

/0.

with dj () = f is easily checked to be a complex; it is called the Koszul


complex on {f1 , . . . , fm }.

Example 0.2. sheaf theoretic version of something not exact as sequence of modules, but exact at sheaf level, in particular Toric with supp in B.

0.1. Maps of complexes, Snake Lemma, long exact sequence.


Definition 0.3. If A and B are complexes, then a morphism of complexes is a
i

family of homomorphisms Ai Bi making the diagram below commute:


A:
B:

/ Ai+1

i+1

i+1
i+1
Bi+1

/ Ai

/ Ai1

/ Bi

i1

i1
i1
Bi1

Lemma 0.4 (Induced Map on Homology). A morphism of complexes induces a


map on homology.
Proof. To show that i induces a map Hi (A) Hi (B), take ai Ai with
i (ai ) = 0. Since the diagram commutes,
0 = i1 i (ai ) = i i (ai )
Hence, i (ai ) is in the kernel of i , so we obtain a map ker i Hi (B). If
ai = i+1 (ai+1 ), then
i (ai ) = i i+1 (ai+1 ) = i+1 i+1 (ai+1 ),
so takes the image of to the image of , yielding a map Hi (A) Hi (B).

When do two morphisms of complexes induce the same map on homology?


Definition 0.5. If A and B are complexes, and , are morphisms of complexes,
i
then and are homotopic if there exists a family of homomorphisms Ai Bi+1
such that for all i, i i = i+1 i + i1 i . Notice that need not commute
with and .
Theorem 0.6. Homotopic maps induce the same map on homology.

0. Homological basics: Complexes and Resolutions

383

Proof. It suffices to show that if i = i+1 i + i1 i then induces the zero map
on homology. But if ai Hi (A), then since i (ai ) = 0,
i (ai ) = i+1 i (ai ) + i1 i (ai ) = i+1 i (ai ) im(),
so i (ai ) = 0 in Hi (B).

Lemma 0.7 (The Snake Lemma). For a commutative diagram of Rmodules with
exact rows
a1

A1
f1

/ B1

f3

b1

/0

/ A3

f2

a2

/ A2

b2

/ B2

/ B3

then there exists an exact sequence:


ker f1

/ ker f2

/ ker f3

/ cokerf1

/ cokerf2

/ cokerf3 .

Definition 0.8. A short exact sequence of complexes is a commuting diagram:


0

A :

B :

C :


/ A2


/ A1


/ A0

/0


/ B2


/ B1


/ B0

/0


/ C2


/ C1


/ C0

/0

where the columns are exact and the rows are complexes.
In Exercise 0.1 youll prove the snake lemma, and in Exercise 0.2 youll combine it with induction to show:
Theorem 0.9 (Long Exact Sequence in Homology). A short exact sequence of
complexes yields a long exact sequence in homology:

/ Hn+1 (C)

/ Hn (A)

/ Hn (B)

/ Hn (C)

A proof using spectral sequences appears in 3.

/ Hn1 (A)

9. Cohomology and Spectral Sequences

384

0.2. Projective and injective modules.


Definition 0.10. A module P is projective if it possesses a universal lifting prop
erty: For any Rmodules G and H, given a homomorphism P H and sur

jection G H, there exists a homomorphism making the diagram below


commute:
P
~
~~
~

~
~~

~~ ~
/H
G

/0

A (left) R-module M is free if M is isomorphic to a direct sum of copies of


the (left) Rmodule R. Free modules are projective.

Definition 0.11. A module I is injective if given a homomorphism H I and

injection H G, there exists a homomorphism making the diagram below


commute:
> IO
~~
~~
~
~
~~
~~
Ho
Go

Projective and Injective modules will come to the forefront in the next section,
which describes derived functors.
0.3. Resolutions. Given an Rmodule M , there exists a projective module surjecting onto M ; for example, take a free module with a generator for each element
of M . This yields an exact sequence:
d

0
P0
M 0.

The map d0 has a kernel, so the process can be iterated, producing an exact sequence (possibly infinite) of free modules, terminating in M .
Definition 0.12. A projective resolution for an Rmodule M is an exact sequence
of projective modules
d

2
1
P2
P1
P0 , with coker(d1 ) = M.

Notice there is no uniqueness property; for example we could set P4 = P4 R


and P3 = P3 R, and define a map P4 P3 which is the identity on the R
summands, and the original map on the Pi summands. In the category of R
modules the construction above shows that projective resolutions always exist.
Surprisingly this is not the case for sheaves of OX -modules. In fact ([19], Exercise
III.6.2), for X = P1 there is no projective object surjecting onto OX .

1. Functors and Derived Functors

385

Definition 0.13. An injective resolution for an Rmodule M is an exact sequence


of injective modules
d

0
1
I0
I1
I2 , with ker(d0 ) = M.

While it is not obvious that injective resolutions exist, it can be shown (see,
e.g. [?]) that in both the category of Rmodules and the category of sheaves of
OX modules, every object does have an injective resolution.
Exercises for 0.
0.1. Prove the snake lemma.
0.2. Prove the existance of the long exact sequence in homology.
0.3. Prove the existance of an injective resolution for Rmodules, when R = Z (see
Hungerford, divisibility section).

1. Functors and Derived Functors


In this section we describe the construction of derived functors, focusing on Exti
(in the category of R-modules) and H i (in the category of sheaves of OX modules).
For brevity we call these two categories our categories. Working in our categories keeps things concrete and lets us avoid introducing too much terminology,
while highlighting the most salient features of the constructions, most of which apply in much more general contexts. For proofs and a detailed discussion, see [11]
or [?].
1.1. Categories and Functors. Recall that a category is a class of objects, along
with morphisms between the objects, satisfying certain properties: composition of
morphisms is associative, and identity morphisms exist.
Definition 1.1. Suppose B and C are categories. A functor F is a function from
B to C , taking objects to objects and morphisms to morphisms, preserving identity
morphisms and compositions. If
b

1
2
B1
B2
B3

is a sequence of objects and morphisms in B, then


F is covariant if applying F yields a sequence of objects and morphisms
in C of the form:
F (b1 )

F (b2 )

F (B1 ) F (B2 ) F (B3 ).


F is contravariant if applying F yields a sequence of objects and morphisms in C of the form:
F (b2 )

F (b1 )

F (B3 ) F (B2 ) F (B1 ).

9. Cohomology and Spectral Sequences

386

A functor is additive if it preserves addition of homomorphisms; this property


will be necessary in the construction of derived functors.
Example 1.2. The global sections functor is covariant: given a sequence of OX
modules
f
g
M1
M2
M3 ,
taking global sections yields a sequence
(M1 ) (M2 ) (M3 ).

Definition 1.3. Let F be a functor from B to C , with B and C categories of


modules over a ring. Let
b1

/ B1

b2

/ B2

/0

/ B3

be a short exact sequence. F is leftexact if either


F is covariant, and the sequence
F (b1 )

/ F (B1 )

/ F (B2 )

F (b2 )

/ F (B3 ) is exact, or

F is contravariant, and the sequence


/ F (B3 )

F (b2 )

/ F (B2 )

F (b1 )

/ F (B1 ) is exact.

A similar definition applies for right exactness; a functor F is said to be exact


if it is both left and right exact, which is synonymous with saying that F preserves
exact sequences.
1.2. Derived Functors. The construction of derived functors is motivated by the
following question: if F is a left exact, contravariant functor and
/ B1

b1

/ B2

b2

/0

/ B3

is a short exact sequence, then what is the cokernel of F (b1 )? For example, if N is
a fixed Rmodule, then the functor HomR (, N ) is a functor of exactly this type.
Definition 1.4. Let B be the category of modules over a ring, and let F be a left
exact, contravariant, additive functor from B to itself. If M B, then there exists
a projective resolution P for M .

/ P2

d2

/ P1

d1

/ P0

Applying F to P yields a complex:


0

/ F (P0 )

F (d1 )

/ F (P1 )

F (d2 )

/ F (P2 )

/ .

1. Functors and Derived Functors

387

The right derived functors Ri F (M ) are defined as


Ri F (M ) = H i (F (P )).
Theorem 1.5. Ri F (M ) is independent of the choice of projective resolution, and
has the following properties:
R0 F (M ) = F (M ).
If M is projective then Ri F (M ) = 0 if i > 0.
A short exact sequence
b

1
2
B : 0 B1
B2
B3 0

gives rise to a long exact sequence


Rj1 F (B3 )

Rj1 (F (b2 ))

/ Rj1 F (B2 )

Rj1 (F (b1 ))

/ Rj1 F (B1 )
i
i
i
i
iiii
iiii
i
i
i
iiii
iiii j1
i
i
i
i
iij i(Fi (b ))
tiiR
Rj (F (b1 ))
2
j
/ Rj F (B1 )
/ Rj F (B2 )
R F (B3 )
i
i
i
iiii
iiii
i
i
i
iiii
iiii j
i
i
i
ii
iiii
tiij+1
Rj+1 (F (b1 ))
R
(F (b2 ))
/ Rj+1 F (B1 )
/ Rj+1 F (B2 )
Rj+1 F (B3 )

of derived functors, where the connecting maps are natural: given another
short exact sequence C and map from B to C , the obvious diagram
involving the Ri F commutes.
The proof is really not bad but long to write out, and we refer to Proposition
A3.17 of [11] for details. There are four possible combinations of variance and
exactness; the type of resolution used to compute the derived functors of F is
given below:
F
covariant contravariant
lef t exact injective
projective
right exact projective
injective
In the next sections, we study some common derived functors.
1.3. Ext. Let R be a ring, and suppose
b

1
2
B : 0 B1
B2
B3 0

is a short exact sequence of Rmodules, with N some fixed Rmodule. Applying


HomR (, N ) to B yields an exact sequence:
c

1
2
0 HomR (B3 , N )
HomR (B2 , N )
HomR (B1 , N ),

9. Cohomology and Spectral Sequences

388

with c1 () 7 b2 and c2 () 7 b1 ; HomR (, N ) is left exact and contravariant.


Definition 1.6. ExtiR (, N ) is the ith right derived functor of HomR (, N )
Given Rmodules M and N , to compute ExtiR (M, N ), we must find a projective resolution for M
d2
d1
P2
P1
P0 ,
and compute the homology of the complex
0 Hom(P0 , N ) Hom(P1 , N ) Hom(P2 , N )
Example 1.7. Let R = C[x, y, z], M = R/hxy, xz, yzi, and suppose N R1 .
Applying HomR (, R1 ) to the projective (indeed, free) resolution of M
z z
7
y
0 7
5
h
i
xy xz yz
0
x
3
2
R(2) R
0 R(3)
3

2
6
6
4

simply means dualizing the module and transposing the differentials, so Exti (R/I, R)
is:
3
2
xy
3
2
7
6
6 xz 7
z
y
0
5
4
5
4
#
"
z 0 x
yz
Hi 0 R R(2)3 R(3)2 0
Thus, Ext2 (R/I, R) is the cokernel of the last map, and it is easy to check that
Ext0 (R/I, R) = Ext1 (R/I, R) = 0.

For a fixed Rmodule M , applying HomR (M, ) to B yields an exact sequence:


0

/ HomR (M, B1 )

c1

/ HomR (M, B2 )

c2

/ HomR (M, B3 ) ,

with c1 () 7 b1 and c2 () 7 b2 ; HomR (, M ) is left exact and covariant.


Thus, to compute the derived functors of HomR (, M ), on a module N , we must
find an injective resolution of N :
I0
then compute
"
Hi 0

/ Hom(I 0 , M )

/ I1

/ I2

/ Hom(I 1 , M )

/ Hom(I 2 , M )

Using spectral sequences (see next section), it is possible to show that that Exti (M, N )
can be regarded as the ith derived functor of either HomR (, N ) or HomR (M, ).

1. Functors and Derived Functors

389

1.4. The global sections functor. Let X be a variety, and suppose B is a coherent
OX module. As we saw in Chapter 7, the global sections functor is left exact
and covariant. Hence, to compute Li (B), we take an injective resolution of B:
/ I1

I0

/ I2

then compute
"

Hi 0

/ (I 0 )

/ (I 1 )

/ (I 2 )

In Example 1.7 we wrote down an explicit free resolution and computed the Ext
modules. Unfortunately, the general construction for injective resolutions produces
very complicated objects. For example, if R is a polynomial ring, then the smallest
injective module in which the residue field can be included is infinitely generated.

It is not obvious that there is a relation between the Cech


cohomology which
appeared in Chapter 7 and the derived functors of defined above. At the end of
this chapter, well see that there is a map
i (U , F ) H i (X, F ),
H
and use spectral sequences to show that with certain conditions on U this is an
isomorphism. The upshot is that many key facts about Cech cohomology (for example, the fact that a short exact sequence of sheaves gives rise to a long exact
sequence in cohomology) follow automatically from the derived functor machinery!
1.5. Acyclic objects. The last concept we need in order to work with derived functors is the notion of an acyclic object.
Definition 1.8. Let F be a leftexact, covariant functor. An object A is acyclic for
F if Ri F (A) = 0 for all i > 0. An acyclic resolution of M is an exact sequence
A0

d0

/ A1

d1

/ A2

d2

where the Ai are acyclic, and M = ker(d0 ).


The reason acyclic objects are important is that a resolution of acyclic objects
is good enough to compute higher derived functors; in other words we have an
alternative to using resolutions by projective or injective objects.
Theorem 1.9. Let M be a coherent OX module, and
A0

/ A1

a acyclic resolution of M . Then


"
Ri (M ) = H i 0

/ (A 0 )

/ A2

/ (A 1 )

/ (A 2 )

9. Cohomology and Spectral Sequences

390

Proof. First, break the resolution into short exact sequences:


0 DD

DD
"

/ A0
CCC
C!

/M

<0
zz
z
z

M C
CCC
{=
!
{{{
/ A2
/ A1
CCC
{=
C!
{{{

< M DD
DD
zz
zz
"

/ A3
=
{
{
{{

< M DD
DD
zz
zz
"

M
{=
{{{

=0
{{
{
{
/

Since the A i are acyclic for , applying to the short exact sequence
0 M A 0 M0 0
yields an exact sequence
0 (M ) (A 0 ) (M 0 ) R1 (M ) 0
Now apply the snake lemma to the middle two columns of the (exact, commutative)
diagram below.
0


/ (M )


/ (A 0 )


/ (M 0 )

/ R1 (M )

/0


/ (M )


/ (A 0 )


/ (A 1 )

/ (A 1 )/(A 0 )

/0


/ (M 1 )

/ (M 1 )

This yields a right exact sequence


0 R1 (M ) (A 1 )/(A 0 ) (M 1 ) (A 1 )/ ker(d1 ),
d1

where (A 1 ) (A 2 ). Hence,
"
1

R (M ) = H

/ (A 0 )

/ (A 1 )

/ (A 2 )

In Exercise 1.2 youll show that iterating this process yields the theorem.

#


2. Spectral Sequences

391

Exercises for 1.
1.1. Prove that the derived functors do not depend on choice of resolution. The key to this
is to construct a homotopy between resolutions, and appeal to Theorem 1.6.
1.2. Complete the proof of Theorem 1.9 by replacing the sequence
0 M A 0 M0 0
with
0 M i1 A i M i 0

2. Spectral Sequences
Spectral sequences are a fundamental tool in algebra and topology; at first glance,
they can seem quite confusing. In this brief overview, we describe a specific type
of spectral sequence, state the main theorem, and illustrate the use of spectral sequences by several examples.
2.1. Total complex of double complex.
Definition 2.1. A first quadrant double complex is a commuting diagram, where
each row and each column is a complex:

d03

d13

P02 o

12

P12 o

d02

22

P22 o

11

P11 o

d01

21

P21 o

d11

P10 o

31

d21

10

32

d22

P00 o

d12

P01 o

d23

20

P20 o

30

For each antidiagonal, define a module


M
Pm =
Pij .
i+j=m

We may define maps

m
Pm
Pm1

via
Dm (cij ) = dij (cij ) + (1)m ij (cij ).
Thus, Dm1 Dm (a) = dd(a) + (a) (d(a) d(a)). The fact that each row
and each column are complexes implies that (a) = 0 and dd(a) = 0. The
commutativity of the diagram implies that d(a) = d(a), and so D 2 = 0.

9. Cohomology and Spectral Sequences

392

Definition 2.2. The total complex Tot(P ) associated to a double complex Pij is
the complex (P , D ) defined above.
Definition 2.3. A filtration of a module M is a chain of submodules
0 Mn Mn1 M1 M0 = M
A filtration has an associated graded object gr(M ) = Mi /Mi+1 . The main
theorem concerning the spectral sequence of a double complex describes two different filtrations of the homology of the associated single complex. To describe
these filtrations, we need to follow two different paths through the double complex.

2.2. The vertical filtration. For a double complex as above, we first compute
homology with respect to the vertical differentials, yielding the following diagram:
ker(d02 )/im(d03 ) o

ker(d01 )/im(d02 ) o

P00 /im(d01 ) o

ker(d12 )/im(d13 ) o

12

ker(d11 )/im(d12 ) o

11

P10 /im(d11 ) o

10

22

21

20

ker(d22 )/im(d23 ) o

ker(d21 )/im(d22 ) o

P20 /im(d21 ) o

These objects are renamed as follows:


1
vertE02

1
vertE01

1
vertE00

12

1
vertE12

11

1
vertE11

10

1
vertE10

22

1
vertE22

32

21

1
vertE21

31

20

1
vertE20

30

The vertical arrows disappeared after computing homology with respect to d, and
the horizontal arrows reflect the induced maps on homology from the original diagram. Now, compute the homology of the diagram above, with respect to the
2 represents
horizontal maps. For example, the object vertE11

11
21
1
1
1
1
ker(vertE11

vertE01 )/im(vertE21 vertE11 )

2. Spectral Sequences

393

The resulting modules may be displayed in a grid:


2
vertE02

2
vertE12

2
vertE22

2
vertE01

2
vertE11

2
vertE21

2
vertE00

2
vertE10

2
vertE20

Although it appears at first that there are no maps between these objects, the crucial
2 to E 2
observation is that there is a map d2i,j from Eij
i2,j+1 . This knights move
is constructed just like the connecting map appearing in the snake lemma. The
diagram above (with differentials added) is thus:
2
vertE02

2
vertE01

2
vertE00

2
vertE22

2
vertE21

vertE12
iTTT
TTTT
TTT
TTTT
TTT
d221
TTT
T

vertE11
iTTT
TTTT
TTT
TTT
TTTT
d220
TTT
T
2
vertE10

2
vertE20

So we may compute homology with respect to this differential. The homology at


3 ; it is now the case (but far
position (i, j) is labeled, as one might expect, vert Eij
3 to
3
from intuitive) that there is a differential d3i,j taking vert Eij
vert Ei3,j+2 :
3
vertE02

3
vertE01

3
vertE00

E3

E3

E3

vert 32
vert 22
vert 12
gPPP
PPP
PPP
PPP
PPP
PPP
PPP
3
3
E3
vertE31
vertE11
PPvert
PPP 21
d330
PPP
PPP
PPP
PP
3
vertE10

3
vertE20

3
vertE30

r to
r
The process continues, with drij mapping vert Eij
vert Eir,j+r1 . One thing that
is obvious is that since the double complex lies in the first quadrant, eventually
the differentials in and out at position (i, j) must be zero, so that the module at
. For example, it is easy to see that
position (i, j) stabilizes; it is written vert Eij

2
vert E10 = vert E10 , while vert E20 6= vert E20 but vert E20 = vert E20 .

9. Cohomology and Spectral Sequences

394

2.3. Main theorem. The main theorem is that the E terms of a spectral sequence from a first quadrant double complex are related to the homology of the
total complex.
L
, then we say that a spectral sequence of
Definition 2.4. If gr(M )m
Eij
i+j=m

the filtered object M converges, and write

Er M
Theorem 2.5. For the filtration of Hm (Tot) obtained by truncating columns of the
double complex,
M

vert Eij Hm (Tot).


i+j=m

As with the long exact sequence of of derived functors, the proof is not bad,
but lengthy, so we refer to [?] or [11] for details. In the previous section, we first
computed homology with respect to the vertical differential d. If instead we first
compute homology with respect to the horizontal differential , then the higher
differentials are:
05

05

d01505

0055d
00525
00 55d3
0 55
55
55
55


As before, for r 0, the source and target are zero, so the homology at position
, and we have:
(i, j) stabilizes. The resulting value is denoted hor Eij
Theorem 2.6. For the filtration of Hm (Tot) obtained by truncating rows of the
double complex,
M

hor Eij Hm (Tot).


i+j=m

For a first quadrant double complex (the only kind that will be of interest to
us), the above two theorems tell us that
M
M

vert Eij Hm (Tot).


hor Eij Hm (Tot) and
i+j=m

i+j=m

Because the filtrations for the horizontal and vertical spectral sequence are different, it is often the case that for one of the spectral sequences the E terms stabilize
very early (perhaps even vanishing). So the main idea is to play off the two different filtrations against each other. This is illustrated in the next example.

2. Spectral Sequences

395

Example 2.7. We prove Theorem 0.9 via spectral sequences. Let 0 C2


C1 C0 0 be a short exact sequence of complexes:
0

C2 : 0 o

C02 o

C1 : 0 o

C01 o

C0 : 0 o

C00 o

C12 o

C22 o

C11 o

C21 o

C10 o

C20 o

Since the columns are exact, it is immediate that for all (i, j)
1
vert Eij

= vert Eij
=0

By Theorem 2.5, we conclude Hm (Tot) = 0 for all m. For the horizontal filtration
1
2
hor Eij = Hi (Cj ) if j {1, 2, 3}, and 0 otherwise. For E

j=1
ker(Hi (C1 ) Hi (C2 ))
2
ker(Hi (C2 ) Hi (C3 ))/im(Hi (C1 ) Hi (C2 )) j = 2
hor Eij =

coker(Hi (C2 ) Hi (C3 ))


j = 3.

2
2
The d2 differential is zero for the middle row, and maps horEi,2
horEi+1,0 :
2
horEi,2

horE

i+1,2
88
88
88
88 d
2
88 2horE 2
horEi,1
i+1,1
88
88
88


2
horEi,0

2
horEi+1,0

2 =

So horEi,1
horEi,1 , while
3
horEi,2

2
2
= horEi,2
= ker(horEi,2
horEi+1,0
)

and
3
horEi,0

2
2
= horEi,0
= coker(horEi,2
horEi+1,0
)

By Theorem 2.6,
Hm (Tot) =

i+j=m

hor Eij

9. Cohomology and Spectral Sequences

396

From the vertical spectral sequence, Hm (Tot) = 0, so all the terms


vanish. Working backwards, we see this means

hor Eij

must

2
0 = horEi,1
= ker(Hi (C2 ) Hi (C3 ))/im(Hi (C1 ) Hi (C2 )),

hence Hi (C1 ) Hi (C2 ) Hi (C3 ) is exact, and


ker(Hi (C1 ) Hi (C2 )) coker(Hi+1 (C2 ) Hi+1 (C3 ))

which yields the long exact sequence in homology.


Exercises for 2.

2.1. Tensor product is right exact and covariant. Prove that the ith left derived functor of
R N is isomorphic to the ith left derived functor of N R as follows: Let
p2

p1

q2

q1

P2 P1 P0
be a projective resolution for M and
Q2 Q1 Q0
be a projective resolution for N . Form the double complex

d13

d03


P0 Q2 o

12

d02


P1 Q2 o

d23

22

d12


P0 Q1 o

11

d01


P1 Q1 o

10


P1 Q0 o

32

d22

21

d11


P0 Q0 o


P2 Q2 o


P2 Q1 o

31

d21

20


P2 Q0 o

30

ij

dij

with differentials Pi Qj Pi1 Qj defined by a b 7 pi (a) b, and Pi Qj


Pi Qj1 defined by a b 7 a qj (b).
(a) Show that for the vertical filtration, the E 1 terms are
(
Pi N j = 0
1
vert Eij =
0
j 6= 0.
Now explain why vert E 2 = vert E , and these terms are:
(
Hi (P N ) = T ori (M, N ) j = 0
2
vert Eij =
0
j=
6 0.
(b) Show that for the horizontal filtration
(
Hj (M Q ) = T orj (N, M ) i = 0
2
hor Eij =
0
i=
6 0.

3. Spectral Sequences and Derived Functors

397

(c) Put everything together to conclude that


T orm (M, N ) =

vert Eij

gr(Hm (Tot))

hor Eij

T orm (N, M )

i+j=m

i+j=m

2.2. Prove that Exti (M, N ) can be regarded as the ith derived functor of either HomR (, N )
or HomR (M, ). The method is quite similar to the proof above, except for this one, youll
need both projective and injective resolutions.

3. Spectral Sequences and Derived Functors


In this last section, well see how useful the machinery of spectral sequences in
yielding theorems about derived functors. To do this, we first define resolutions of
complexes. Note that sometimes our differentials on the double complex go up
and right instead of down and left, so the higher differentials change accordingly.

3.1. Resolution of a complex. Suppose


/ A0

A:0

/ A1

/ A2

is a complex, either of Rmodules or of sheaves of OX modules. An injective


resolution of A is a double complex:
O

I O 02

/ I 12
O

/ I 22
O
/

I O 01

/ I 11
O

/ I 21
O
/

I 00

/ I 10

/ I 20
/

satisfying the following properties (djk denotes the horizontal differential at (j, k)).
The complex is exact.
Each column I i is an injective resolution of A i .
ker(djk ) is an injective summand of I jk .

9. Cohomology and Spectral Sequences

398

The last condition implies that im(dj,k ) is also injective. This yields a Hodge
decomposition:
/ im(dj1,k )

/ ker(dj,k )

/ H j,k
d

/0

im(dj1,k )
It follows that we may decompose the sequence
I j1,k

dj1,k

/ I j,k

dj,k

/ I j+1,k

as:
/ im(dj2,k )

/ H j1,k

/ im(dj1,k )

/ im(dj,k )

/ H j,k

/ im(dj1,k )

/ im(dj,k )

/ H j+1,k

/ im(dj+1,k )

An inductive argument (Exercise 3.1) shows that in a category with enough injective objects, injective resolutions of complexes always exist.

3.2. Grothendieck spectral sequence. One of the most important spectral sequences is due to Grothendieck, and relates the higher derived functors of a pair of
functors F ,G, and their composition F G.
Theorem 3.1. Suppose that F is a left exact, covariant functor from C1 C2
and G is a left exact, covariant functor from C2 C3 , where the Ci are one of
our categories. If A C1 has an F acyclic resolution A such that F (A i ) is
Gacyclic then
Ri G(Rj F (A)) Ri+j GF (A)
Proof. Take an injective resolution I , for the complex
0

/ F (A 0 )

/ F (A 1 )

/ F (A 2 )

3. Spectral Sequences and Derived Functors

399

Apply G to I , . It follows from the construction above that a row of the double
complex G(I , ) has the form:
G(0)

/ G(im(dj2,k ))

G(0)

/ G(H j1,k )

/ G(im(dj,k ))

/ G(H j,k )

G(1)

/ G(im(dj1,k ))

L
L

/ G(im(dj1,k ))

G(1)

G(0)

G(0)

/ G(im(dj,k ))

/ G(H j+1,k )

/ G(im(dj+1,k ))

Hence,
1
hor Eij

= G(H i,j )

By construction, H i,j is the j th object in an injective resolution for the ith cohomology of F (A ). Since A was an F acyclic resolution for A, the ith cohomology
is exactly Ri F (A), so that
"
#
2
hor Eij

= H j 0 G(H i,0 ) G(H i,1 ) G(H i,2 )

= Rj G(Ri F (A))

Next, we turn to the vertical filtration. We have the double complex


O

G(I 02 )

/ G(I 12 )
O

/ G(I 22 )
O
/

G(I 01 )
O

/ G(I 11 )
O

/ G(I 21 )
O
/

G(I 00 )

/ G(I 10 )

/ G(I 20 )

Since I ij is an injective resolution of F (A i ),


"

Rj G(F (A i )) = H j 0 G(I i0 ) G(I i1 ) G(I i2 ) .

Now, the assumption that the F (A i ) are Gacyclic forces Rj G(F (A i )) to vanish,
for all j > 0! Hence, the cohomology of a column of the double complex above
vanishes, except at position zero. In short
(
GF (A i ) j = 0
1
vert Eij =
0
j 6= 0.

9. Cohomology and Spectral Sequences

400

Thus,

vert Eij

2
vert Eij

(
Ri+j GF (A)
=
0

j=0
j 6= 0.


Applying Theorem 2.5 and Theorem 2.6 concludes the proof.

3.3. Comparing cohomology theories. Our final application of spectral sequences

will be to relate the higher derived functors of to the Cech


cohomology. For any
f

map Y X and sheaf F on Y , the pushforward is defined via:


f F (V ) = F (f 1 (V ))
If Ip denotes a p + 1-tuple {i0 < i1 < < ip } and UIp = Ui0 Uip , then
i

applying this to the inclusion UIp X gives a sheaf theoretic version of the Cech
complex.
Y
i (F |UIp ).
C p (U , F ) =
Ip

In Exercise 3.2 youll show that


gives a resolution of F , and if F is injective,
so are the C . Taking this as given, we then have:
Lemma 3.2. For an open cover U , there is a map
i (U , F ) H i (X, F )
H
Proof. Take an injective resolution I for F . By injectivity, we get
0

/F

/ C0

/ C1

 }

I0

Iterating the construction gives a map of complexes C I , which by Lemma 0.4


yields a map on cohomology.

Theorem 3.3. Let U be an open cover such that for any Ip ,
H i (UIp , F ) = 0, for all i 1.
Then
i (U , F ) = H i (X, F ).
H
Proof. Take an injective resolution I for F . The hypothesis that H i (UIp , F ) =
0, i > 0 implies that the sequence
0

/ F (UIp )

/ I 0 (UI )
p

/ I 1 (UI )
p

is exact. Then as in the construction of the sheaf-theoretic Cech


complex, we obtain

a Cech
complex built out of the direct product of these, which is by construction

3. Spectral Sequences and Derived Functors

401

a resolution (depicted below) of the Cech


complex for F . The bottom row is
included for clarity, it is not part of the complex.
O

C 0 (U , I 2 )

/ C 1 (U , I 2 )
O

/ C 2 (U , I 2 )
O
/

C 0 (U , I 1 )

/ C 1 (U , I 1 )
O

/ C 2 (U , I 1 )
O
/

C 0 (U , I 0 )
O

/ C 1 (U , I 0 )
O

/ C 2 (U , I 0 )
O
/

C 0 (U , F )

/ C 1 (U , F )

/ C 2 (U , F )

Applying , since H i (UIp , F ) = 0 for i > 0,


(
(C i (U , F ))
1
E
=
vert ij
0

j=0
j 6= 0.

Thus, E 2 = E , and since (C i (U , F )) = C i (U , F )


(
i (U , F ) j = 0
H
2
E
=
vert ij
0
j 6= 0.
For the horizontal filtration, since the C i (U , I j ) are injective,
(
(I j ) i = 0
1
E
=
hor ij
0
i 6= 0.
and thus

(
H j ((I )) i = 0
2
hor Eij =
0
i=
6 0.
But this is the usual derived functor cohomology. The result follows from Theorem 2.5 and Theorem 2.6.


Exercises for 3.
3.1. Prove that in a category with enough injective objects, injective resolutions of complexes always exist.
3.2. Show that C is a resolution of F , as follows. By working at the level of stalks, show
that there is a morphism of complexes
k

C i (U , F )p C i1 (U , F )p
such that (di1 k + kdi ) is the identity. Conclude by applying Theorem 0.6. Finally, show
that if F is injective, then so are the sheaves C i (U , F ). If you get stuck, see [19], III.4.

9. Cohomology and Spectral Sequences

402

3.3. Let Y X be a continuous map between topological spaces, with A a sheaf of


abelian groups on Y .
(a) Show that pushforward f is left exact and covariant, so associated to A are objects
Rj f (A)).
(b) Use Theorem 3.1 to obtain the Leray spectral sequence:
H i (Rj f (A)) H i+j (A)
i (U, F ) = 0 for all open sets U U . Show that then
3.4. Suppose U is a Leray cover: H
i
i

H (U , F ) = H (X, F ).