Homological Algebra

Yuri Berest1
Fall 2013 Spring 2014

Notes taken by Daniel Miller and Sasha Patotski at Cornell University

ii

Contents
1 Standard complexes in Geometry
1
Complexes and cohomology . . . . . . . . . . . . . . . .
2
Simplicial sets and simplicial homology . . . . . . . . . .
2.1
Motivation . . . . . . . . . . . . . . . . . . . . .
2.2
Definitions . . . . . . . . . . . . . . . . . . . . .
2.3
Geometric realization . . . . . . . . . . . . . . .
2.4
Homology and cohomology of simplicial sets . . .
2.5
Applications . . . . . . . . . . . . . . . . . . . .
2.6
Homology and cohomology with local coefficients
3
Sheaves and their cohomology . . . . . . . . . . . . . . .
3.1
Presheaves . . . . . . . . . . . . . . . . . . . . .
3.2
Definitions . . . . . . . . . . . . . . . . . . . . .
3.3
Kernels, images and cokernels . . . . . . . . . . .
3.4
Germs, stalks, and fibers . . . . . . . . . . . . . .
3.5
Coherent and quasi-coherent sheaves . . . . . . .
3.6
Motivation for sheaf cohomology . . . . . . . . .
3.7
Sheaf cohomology . . . . . . . . . . . . . . . . .
3.8
Applications . . . . . . . . . . . . . . . . . . . .

3.9
Cech
cohomology . . . . . . . . . . . . . . . . . .
2 Standard complexes in algebra
1
Group cohomology . . . . . . . . . . . . . . .
1.1
Definitions and topological origin . . .
1.2
Interpretation of H1 (G, A) . . . . . . .
1.3
Interpretation of H2 (G, A) . . . . . . .
2
Hochschild (co)homology . . . . . . . . . . .
2.1
The Bar complex . . . . . . . . . . . .
2.2
Differential graded algebras . . . . . .
2.3
Why DG algebras? . . . . . . . . . . .
2.4
Interpretation of bar complex in terms
2.5
Hochschild (co)homology: definitions .
2.6
Centers and Derivations . . . . . . . .
iii

1
. . . . . . . . . . . . 1
. . . . . . . . . . .
3
. . . . . . . . . . .
3
. . . . . . . . . . .
5
. . . . . . . . . . .
6
. . . . . . . . . . .
8
. . . . . . . . . . .
9
. . . . . . . . . . .
10
. . . . . . . . . . .
12
. . . . . . . . . . .
12
. . . . . . . . . . .
12
. . . . . . . . . . .
13
. . . . . . . . . . .
14
. . . . . . . . . . .
16
. . . . . . . . . . .
17
. . . . . . . . . . .
19
. . . . . . . . . . . . 21
. . . . . . . . . . .
23

. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
of DG algebras
. . . . . . . . .
. . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

25
25
25
26
27
29
29
30
33
34
35
37

2.7
Extensions of algebras . . . . . . . . . .
2.8
Crossed bimodules . . . . . . . . . . . .
2.9
The characteristic class of a DG algebra
Deformation theory . . . . . . . . . . . . . . . .
3.1
Motivation . . . . . . . . . . . . . . . .
3.2
Formal deformations . . . . . . . . . . .
3.3
Deformation theory in general . . . . . .
3.4
The Gerstenhaber bracket . . . . . . . .
3.5
Stasheff construction . . . . . . . . . . .
3.6
Kontsevich Formality Theorem . . . . .
3.7
Deformation theory in algebraic number

. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
theory

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
37
.
38
.
40
. . 41
. . 41
.
42
.
44
.
49
.
50
.
52
.
53

3 Category theory
1
Basic category theory . . . . . . . . . . . . . . . . . .
1.1
Definition of categories . . . . . . . . . . . . . .
1.2
Functors and natural transformations . . . . .
1.3
Equivalences of categories . . . . . . . . . . . .
1.4
Representable functors and the Yoneda lemma
1.5
Adjoint functors . . . . . . . . . . . . . . . . .
1.6
Limits and colimits . . . . . . . . . . . . . . . .
2
Special topics in category theory . . . . . . . . . . . .
2.1
Brief introduction to additive categories . . . .
2.2
Center of a category and Bernstein trace . . . .
2.3
Morita theory . . . . . . . . . . . . . . . . . . .
2.4
Recollement (gluing) of abelian sheaves . . . .
2.5
Kan extensions . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

57
. . . . . . . . .
57
. . . . . . . . .
57
. . . . . . . . .
59
. . . . . . . . . . 61
. . . . . . . . .
62
. . . . . . . . .
66
. . . . . . . . .
68
. . . . . . . . . . 71
. . . . . . . . . . 71
. . . . . . . . .
72
. . . . . . . . .
76
. . . . . . . . .
78
. . . . . . . . .
79

4 Classical homological algebra

1
Abelian categories . . . . . . . . . . . . . . . .
1.1
Additive categories . . . . . . . . . . . .
1.2
Non-additive bimodules . . . . . . . . .
1.3
Abelian categories . . . . . . . . . . . .
1.4
Complexes in abelian categories . . . . .
1.5
Exact functors . . . . . . . . . . . . . .
1.6
Adjointness and exactness . . . . . . . .
2
Finiteness conditions . . . . . . . . . . . . . . .
2.1
AB5 categories . . . . . . . . . . . . . .
2.2
Grothendieck categories . . . . . . . . .
2.3
Inductive closure of an abelian category
2.4
Finiteness conditions . . . . . . . . . . .
3
Classical derived functors . . . . . . . . . . . .
3.1
Injectives and injective envelopes . . . .
3.2
Canonical constructions on complexes .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

iv

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

83
.
83
.
83
.
85
.
87
. . 91
.
95
.
96
.
99
.
99
. . 101
. 102
. 104
. 106
. 106
. 108

3.3
3.4
3.5
3.6
3.7

Classical definition of classical derived functors . .

-functors . . . . . . . . . . . . . . . . . . . . . . . .
Main properties of resolutions . . . . . . . . . . . . .
Definition of classical derived functor via -functors
Examples of derived functors . . . . . . . . . . . . .

5 Derived categories
1
Localization of categories . . . . . .
1.1
Motivation . . . . . . . . . .
1.2
Definition of localization . . .
1.3
Calculus of fractions . . . . .
1.4
Ore localization . . . . . . . .
2
Localization of abelian categories . .
2.1
Serre quotients . . . . . . . .
2.2
Injective envelopes . . . . . .
2.3
Localizing subcategories . . .
3
Derived categories . . . . . . . . . .
3.1
Definition and basic examples
3.2
The homotopy category . . .
3.3
Verdier theorem . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

. . 111
. 113
. 115
. 117
. . 121

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

125
. 125
. 125
. 126
. 128
. 130
. 134
. 134
. 137
. 138
. 139
. 139
. . 141
. 143

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

145
145
145
146
148
152
152
153
153
154
155
156
160
160
. 161
162
163
164
165
165
167
168

6 Triangulated categories
1
The basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1
Definitions . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Examples of triangulated categories . . . . . . . . . .
1.3
Basic properties of triangulated categories . . . . . . .
2
Further properties . . . . . . . . . . . . . . . . . . . . . . . .
2.1
Abstract cone and octahedron axiom . . . . . . . . . .
2.2
The homotopy category is triangulated (need proofs!!)
2.3
Localization of triangulated categories . . . . . . . . .
2.4
Exact functors . . . . . . . . . . . . . . . . . . . . . .
2.5
Verdier quotients . . . . . . . . . . . . . . . . . . . . .
2.6
Exact categories . . . . . . . . . . . . . . . . . . . . .
3
t-structures and the recollement . . . . . . . . . . . . . . . . .
3.1
Motivation and definition . . . . . . . . . . . . . . . .
3.2
The core of a t-structure and truncation functors . . .
3.3
Cohomological functors . . . . . . . . . . . . . . . . .
3.4
t-exact functors . . . . . . . . . . . . . . . . . . . . . .
3.5
Derived category of t-structure . . . . . . . . . . . . .
3.6
Gluing t-structures . . . . . . . . . . . . . . . . . . . .
3.7
Examples of gluing . . . . . . . . . . . . . . . . . . . .
3.8
Recollement for triangulated categories . . . . . . . .
3.9
Example: topological recollement . . . . . . . . . . . .
v

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

3.10

Example: algebraic recollement . . . . . . . . . . . . . . . . . . . . .

168

7 Applications of homological algebra

1
Perverse sheaves and quivers . . . . . . . . . . . . . . . . . . . . .
1.1
Stratifications . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Perversity function and perverse sheaves . . . . . . . . . . .
1.3
Middle extension . . . . . . . . . . . . . . . . . . . . . . . .
1.4
Constructible complexes . . . . . . . . . . . . . . . . . . . .
1.5
Refining stratifications . . . . . . . . . . . . . . . . . . . . .
1.6
Gluing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Matrix factorizations . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
The classical exact equivalences . . . . . . . . . . . . . . . .
2.3
The stabilization of the residue field . . . . . . . . . . . . .
2.4
An enlargement of the matrix factorization category . . . .
2.5
Homotopy theory of DG categories . . . . . . . . . . . . . .
2.6
Derived Morita Equivalence . . . . . . . . . . . . . . . . . .
2.7
Hochschild cohomology . . . . . . . . . . . . . . . . . . . .
3
Some physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1
Physical motivation . . . . . . . . . . . . . . . . . . . . . .
3.2
Cobordism . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Topological quantum field theories . . . . . . . . . . . . . .
3.4
Chern-Simons . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5
Atiyah-Segal axioms . . . . . . . . . . . . . . . . . . . . . .
4
Perfectoid rings, almost mathematics, and the cotangent complex .
4.1
Perfectoid fields . . . . . . . . . . . . . . . . . . . . . . . . .
4.2
Almost mathematics . . . . . . . . . . . . . . . . . . . . . .
4.3
The cotangent complex . . . . . . . . . . . . . . . . . . . .
5
Nearby and Vanishing Cycle Functors . . . . . . . . . . . . . . . .
5.1
Construction . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2
Monodromy . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3
Gluing problems . . . . . . . . . . . . . . . . . . . . . . . .
5.4
Monadology . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5
Reflection functors . . . . . . . . . . . . . . . . . . . . . . .
6
Abstract Nearby Cycles functor . . . . . . . . . . . . . . . . . . . .
6.1
The setup . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2
Multiplicative Preprojective Algebra Associated to a Quiver

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

171
. 171
. 171
172
172
174
175
176
177
177
177
183
184
186
189
. 191
195
195
197
198
199
199
199
199
. 201
203
204
204
205
206
208
210
212
212
214

A Miscellaneous topics
1
Characteristic classes of representations (after
2
Generalized manifolds . . . . . . . . . . . . .
3
Quivers and path algebras . . . . . . . . . . .
3.1
Basic definitions . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

217
217
220
222
222

vi

Quillen)
. . . . .
. . . . .
. . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3.2
3.3
3.4

Path algebra of a quiver . . . . . . . . . . . . . . . . . . . . . . . . .

The structure of the path algebra . . . . . . . . . . . . . . . . . . . .
Representations of quivers . . . . . . . . . . . . . . . . . . . . . . . .

B Exercises
1
Standard complexes in Algebra and Geometry . . . . . . .
2
Classical homological algebra . . . . . . . . . . . . . . . .
3
Residues and Lie cohomology . . . . . . . . . . . . . . . .
3.1
Commensurable subspaces . . . . . . . . . . . . . .
3.2
Traces . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Residues . . . . . . . . . . . . . . . . . . . . . . . .
3.4
Interpretation in terms of Lie algebra cohomology
3.5
Adeles and residues on algebraic curves . . . . . .

vii

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

223
224
225
233
233
233
235
235
236
237
238
238

viii

Chapter 1

Standard complexes in Geometry

In this chapter we introduce basic notions of homological algebra such as complexes and
cohomology. Moreover, we give a lot of examples of complexes arising in different areas
of mathematics giving different cohomology theories. For instance, we discuss simplicial
(co)homology, cohomology of sheaves, group cohomology, Hochschild cohomology, differential
graded (DG) algebras and deformation theory.
Throughout the chapter we will use language of category theory. All the necessary
categorical definitions are reviewed in the first section of the Chapter 3.

Complexes and cohomology

Definition 1.0.1. A chain complex C is a sequence of abelian groups together with group
homomorphisms
C :

dn+1

n
Cn+1 Cn
Cn1

such that dn dn+1 = 0 for all n Z.

Definition 1.0.2. A cochain complex is a sequence
C :

dn

dn+1

C n C n+1 C n+2

such that dn+1 dn = 0 for all n.

One usually calls d the differential, or a boundary operator in the case of a chain complex.
Also, one often leaves out the subscripts / superscripts of d, writing d2 = 0.
Remark 1.0.3. (C n , dn ) is a cochain complex if and only if (Cn = C n , dn = dn ) is a
chain complex.
Definition 1.0.4. If (Cn , dn ) is a chain complex, then the n-th homology of C is
Hn (C) = Ker(dn )/ Im(dn+1 )
1

If (C n , dn ) is a cochain complex, then the n-th cohomology of C is

Hn (C) = Ker(dn+1 )/ Im(dn )
L
L
One writes H (C) = n Hn (C) and H (C) = n Hn (C). One calls H (C) the homology
and H (C) the cohomology of the complex C. If H (C) = 0 or H (C) = 0, one says that C
is acyclic (or an exact complex).
To define a morphism of complexes, we will work only with cochain complexes for
simplicity. A morphism f : C D of complexes is a sequence f n : C n Dn such that
Cn


dn
C

/ C n+1

fn

Dn

dn
D

f n+1

/ D n+1

commutes, that is f n+1 dnC = dnD f n . Note that a morphism of complexes f : C D

induces a morphism H (f ) : H (C) H (D) by letting Hn (f ) : Hn (C) Hn (D) send
the coset [c] to [f n (c)]. The definition of a morphism of complexes ensures that H (f ) is
well-defined.
Definition 1.0.5. Given f, g : C D , we say that f g (f and g are homotopic)
if there is a sequence of homomorphisms {hn : C n Dn1 }nZ such that f n g n =
n
n+1 dn for all n Z:
dn1
C
D h +h

/ C n1

g n1

f n1

 }
/ D n1

n1
dC

hn

n1
dD

/ Cn

gn

fn

  }
/ Dn

dn
C

hn+1

/ C n+1

g n+1

dn
D

f n+1


/ D n+1

Lemma 1.0.6. If f, g : C D are homotopic, then H (f ) = H (g).

Proof. Indeed, if c Ker(dnC ), then f n (c) = g n (c) + d(h(c)).
Corollary 1.0.7. Suppose f : C D and g : D C are such that g f idC and
f g idD , then f and g induce mutually inverse isomorphisms between H (C) and
H (D).
Definition 1.0.8. In this latter case we say that C is homotopy equivalent to D . We
call themaps f and g homotopy equivalences.
Definition 1.0.9. A morphism f : C D is called a quasi-isomorphism if Hn (f ) : Hn (C)
Hn (D) is an isomorphism for each n.
2

Example 1.0.10. Every homotopy equivalence is a quasi-isomorphism.

Remark 1.0.11. The converse is not true.
We are now ready to give a formal definition of a derived category.
Definition 1.0.12. Let A be an abelian category (e.g. the category of abelian groups). Let
Com(A) be the category of complexes over A. The (unbounded) derived category of A is the
(abstract) localisation of Com(A) at the class of all quasi-isomorphisms.

2
2.1

Simplicial sets and simplicial homology

Motivation

There are a number of complexes that appear quite algebraic, but whose construction
involves topology.
Definition 2.1.1. The geometric n-dimensional simplex is the topological space
(
)
n
X
n+1
n = (x0 , . . . , xn ) R
:
xi = 1, xi > 0
i=0

For example, 0 is a point, 1 is an interval, 2 is an equilateral triangle, 3 is a filled

tetrahedron, etc. We will label the vertices of n as e0 , . . . , en .
Definition 2.1.2. A (geometric) complex K = {Si }iI is a union of geometric simplices
Si in RN of varying dimensions such that the intersection Si Sj of any two simplices is a
face of each simplex.
Definition 2.1.3. A polyhedron is a space which is homeomorphic to a geometric complex.
The choice of such a homeomorphism is usually called a triangulation. Clearly, triangulations are highly non-canonical. Both the n-sphere and the n-ball are polyhedra.
Remark 2.1.4. The triangulation of a space is a finitistic way of defining the space, similar
to defining groups or algebras by a finite list of generators and relations.
Remark 2.1.5. In fact, one can triangulate groups, algebras, modules, and objects in
any category.
If X is a geometric complex (or a polyhedron) we can associate to X the following chain
complex:
M
Cn (X) =
Zi
i X
dim(i )=n

dn : Cn Cn1 , i 7

n
X
k=0

(1)k k ik

In the above definition of dn each simplex i is equipped with an orientation (i.e. choice
of an ordering of its vertices). Then ik denotes the simplex {e0 , . . . , ek , . . . , en }, and k = +1
or 1 depending on the sign of the permutation that maps the sequence {e0 , . . . , ek , . . . , en }
to the sequence of vertices of ik determined by its orientation.
As an exercise, show that dn dn1 = 0.
L
Theorem 2.1.6. The homology groups H (X) =
Hn (X) of the complex C = (Cn , dn )
are independent of the choice of triangulation and orientation of simplices.
Proof. See any book on algebraic topology.
It follows that the Hn (X) are invariants of X as a topological space.
Geometric intuition A homology cycle c Hn (X) can be viewed as n-dimensional chains
(n-cycles) modulo the equivalence relation c c0 if there exists an (n + 1)-cycle of which c
and c0 are the boundary.
Definition 2.1.7.
1. A simplicial set is a family of sets X = {Xn }n>0 and a family of
maps {X(f ) : Xn Xm }, one for each non-decreasing function f : [m] [n], where
[n] = {0, . . . , n}, satisfying
X(id) = id
X(f g) = X(g) X(f )
2. A map of simplicial sets : X Y is a family of maps {n : Xn Yn }n>0 such that
for all f : [m] [n]:
Xn
X(f )

Xm

/ Yn


Y (f )

/ Ym

commutes.
Remark 2.1.8. A simplicial set is just a contravariant functor from the simplicial category
to the category of sets Set, and a map of simplicial sets is just a natural transformation of
functors. The simplicial category has finite sets [n] as objects, and non-decreasing functions
as morphisms. We denote the category ofsimplicial sets by Set.
Definition 2.1.9. Let X be a simplicial set. The geometric realization of X is
,

a
|X| =
(n Xn )

n=0

where the equivalence relation is defined by (s, x) (t, y) if, for (s, x) n Xn and
(t, y) m Xm , there exists f : [m] [n] non-decreasing
such that y = X(f )x and
`
t = f s. We give |X| the weakest topology such that (n Xn )  |X| is continuous.
4

2.2

Definitions

Recall that we defined simplicial set as a family X = {Xn }n>0 of sets and a family of maps
X(f ) : Xn Xm , one for each non-decreasing map f : [m] [n], such that X(id) = id and
X(f g) = X(g)X(f ) when the compositions are defined. This can be rephrased more
conceptually using the simplicial category.
Definition 2.2.1. The simplicial category has as objects all finite well ordered sets.
That is, Ob = {[n] = {0 < 1 < < n}}. Morphisms are order-preserving maps (i.e.
i 6 j f (i) 6 f (j)).
A simplicial set is just a contravariant functor X : Set. Thus the category of

simplicial sets is just the category Set = Set . There are two distinguished classes of
maps in :
in : [n] , [n + 1]

06i6n

ji :

06j 6n+1

[n + 1]  [n]

called the face maps in and degeneracy maps jn . They are defined by
(
k
if k < i
in (k) =
k+1 if k > i

(
k
if k 6 j
jn (k) =
k 1 if k > j

Theorem 2.2.2. Any morphism f Hom ([n], [m]) can be decomposed in a unique way as
f = i1 i2 ir j1 js
such that m = n s + r and i1 6 6 ir and j1 6 6 js .
The proof of this theorem is a little technical, but a few examples make it clear what is
going on.
Example 2.2.3. Let f : [3] [1] be {0, 1 7 0; 2, 3 7 1}. One can easily check that
f = 11 22 .
Corollary 2.2.4. For any f Hom ([n], [m]), there is a unique factorization
[n]

/ [m]
O

?

[k]

Corollary 2.2.5. The category can be presented by {i } and {j } as generators with the
following relations:
j i = i j

i<j

j i = i j+1

i j1
j i = id

i1 j

i6j

(1.1)

if i < j
if i = j or i = j + 1
if i > j + 1

Corollary 2.2.6. Giving a simplicial set X = {Xn }n>0 is equivalent to giving a family of
sets {Xn } equipped with morphisms in : Xn Xn1 and sni : Xn Xn+1 satisfying
i j = j i

i<j

si sj = sj+1 si

sj1 i
j si = id

sj i1

i6j

(1.2)

if i < j
if i = j or i = j + 1
if i > j + 1

The relation between (1.1) and (1.2) is given by in = X(in1 ) and sni = X(in ).
P
Consider the n-dimensional geometric simplex n = {(x0 , . . . , xn ) Rn+1
xi = 1}.
>0 :
For aPnon-empty subset I [n], define the I-th face of n by eI = {(x0 , . . . , xn )
n :
iI xi = 1}. In particular, if I = {i}, then the I-th face of n is just the i-th vertex
ei = (0, . . . , 1, . . . , 0).
It is more convenient to parametrize faces by maps f : [m] [n] for m 6 n with
Im(f ) = I.
Example 2.2.7. Let I = {0, 1, 3} [3]. The corresponding map f : [2] [3] is just
{0 7 0, 1 7 1, 2 7 3} = 23 .
In general, given f : [m] [n], the corresponding f : m n is defined to be the
restriction of the linear map Rm+1 Rn+1 sending ei to ef (i) .

2.3

Geometric realization

Recall that given any simplicial set X = {Xn }, we defined the geometric realization of X as
|X| =

.
(n Xn )

n=0

where (s, x) m Xm is equivalent to (y, t) n Xn if there is f : [m] [n] such that

y = X(f )(x) and t = f (s).
To any triangulated space, we can associate a simplicial set. Let X be a triangulated
space. We define the gluing data of X as follows:
6

1. Let X(n) be the set of all n-simplices in X;

2. For each f : [m] [n] define the gluing maps X(f ) : X(n) X(m) so that the fibre
of X(f ) over an m-simplex x X(m) consists of exactly all n-simplices in X(n) which
have x as a common face in X.
Now, with this gluing data we can associate the simplicial set
X := {Xm , X(f )} Set
defined as follows. First we can define Xm by


Xm := (x, g) | x X(k) , g Surj ([m], [k])
Suppose f Hom ([n], [m]). For (x, g) Xm consider the composition g f : [n]
[m] [k]. By Corollary 2.2.4 we can factorize g f as g f = with Surj ([n], [l])
and Inj ([l], [k]). Then define X(f ) to be
X(f )(x, g) = (X()x, ) Xn
It is straightforward to check that X(id) = id and X(f 0 f ) = X(f ) X(f 0 ).
Theorem 2.3.1. The geometric realization |X | is homotopically equivalent to X.
Proof. (sketch) By definition, Xn consists of all pairs x
= (x, g) with x X(m) and
g : [n]  [m] Mor(). Define
:

n Xn

n=0

m X(m)

m=0

by (s, x
) 7 (g (s), x) m X(m) . Clearly if (s, x
) (s0 , x
0 ) then (s, x
) = (s0 , x
0 ).
Hence induces a continuous map : |X| X.The homotopically inverse map is induced
by the map

a
a
:
m X(m)
n Xn
m=0

n=0

defined by (s, x) 7 (s, (x, id[n] )).

Example 2.3.2 (Simplicial model of the circle S 1 ). The simplest simplicial model for the
circle S 1 is a simplicial set S1 which is generated by two non-trivial cells: one in dimension
0 (the basepoint ) and one in dimension 1 which we will denote . The face maps on are
given by d0 () = d1 () = . But we also need to introduce an element s0 () in S11 . Similarly,
at the level n, the set Sn1 has n + 1 elements:
Sn1 = {sn0 (), sn1 sn2 . . . si1 . . . s0 (), i = 1, 2, . . . , n}
This is enough because of the relations between di and Sj . Elements in Sn1 are in natural
bijection with the (additive) group Z/(n + 1)Z:
S01 = {}, S11 = {s0 (), }, S21 = {s20 (), s1 (), s0 ()}, . . .
Simplicial set S 1 is a special kind of simplicial set, called a cyclic set (A.Connes). Such
sets give rise to cyclic homology. We will discuss this type of homology later in the course.
7

2.4

Homology and cohomology of simplicial sets

Let X = {Xn } be a simplicial set. Recall that for each n Z and for each fixed abelian
group A, we defined
(
0
if n < 0
Cn (X, A) =
.
AXn = A ZXn otherwise
The differential dn : Cn Cn1 is defined by
X

a(x) x 7

xXn

a(x)

xXn

a(x)

n
X
(1)i X(in1 )x

i=0
n
X

(1)i in x

i=0

xXn

In other words, we define differential dn : Cn Cn1 to be dn =

n
P

(1)i i .

i=0

Dually, we define C n (X, A) = {functions Xn A}, and the differential dn : C n C n+1

by
n+1
X

(dn f )(x) =

(1)i f (X(in )x)

i=0
n+1
X

(1)i f (in+1 x)

i=0

Theorem 2.4.1. The objects Cn (X, A) and C n (X, A) are actually complexes, i.e. dn1 dn =
0 and dn+1 dn = 0.
"
#
n
P
j
(1) j =
Proof. Lets check that dn1 dn = 0. We have dn1 dn = dn1
j=0
n n1
P
P
(1)i+j i j . Then we can split this sum into two parts and use the relations (1.2)
j=0 i=0

(actually, only the first one of these relations) to get

n n1
X
X
dn1 dn =
(1)i+j i j
j=0 i=0

(1)i+j i j +

i<j

i+j

(1)

j1 i +

ij1

(1)i+j i j

ij

X
X

X
(1)i+j i j
ij

i0 +j 0 +1

(1)

i0 j 0

i0 j 0 +

X
ij

=0
8

(1)i+j i j

Remark 2.4.2. For any category C we can define simplicial objects in C as functors op C.
If category C is abelian (for example abelian groups Ab, or vector spaces Vect, or modules
Mod(R) over some algebra), then we can define homology of a simplicial object X op C as
the homology P
of complex X, where the differential dn : Xn Xn1 is defined by the same
formula dn = (1)i X(i ) as before.
Then the above definition of homology of a simplicial set coincides with homology of
free
the simplicial abelian group S : op Ab defined by the composition op Sets A Ab,
where freeA sends a set X to the abelian group A ZX.
Dually we can define cosimplicial objects in C as functors C. Again, if C is abelian,
then we can define cohomology of a cosimplicial object Y in CPas the cohomology of a
complex Y where the differential is defined by the formula dn = (1)i Y (i ).

2.5

Applications

Example 2.5.1 (Singular (co)homology). Let X be a topological space. A (singular) nsimplex of X is a continuous map : n X. Put, for n > 0, Xn = HomTop (n , X) =
{singular n-simplices in X}. For f Hom ([m], [n]), define X(f ) : Xm Xn by X(f )() =
f .
If A is an abelian group, then we can define singular homology and cohomology of X by

Hsing
(X, A) = H (X, A) and Hsing (X, A) = H (X, A) respectively.
Note that if X has some extra structure (e.g. is a C -manifold or a complex manifold)
then it is often convenient to take simplices compatible with that structure.
Example 2.5.2 (Nerve of a covering). Let X be a topological space, U = {U }I a
covering of X. Define
Xn = {(0 , . . . , n ) I n+1 : U0 Un 6= }.
For f Hom ([m], [n]), the morphism X(f ) : Xn Xm is given by (0 , . . . , n ) 7

(f (0) , . . . , f (n) ). The Cech

(co)homology of X with respect to U is
(U, A) = H (X, A)
H
(U, A) = H (X, A)
H
Example 2.5.3 (Classifying space of a group). Let G be a group. Define a simplicial
set BG by (BG)n = Gn = G G (n-fold product). For f Hom ([m], [n]), define
BG(f ) : Gn Gm by (g1 , . . . , gn ) 7 (h1 , . . . , hm ) where

gj if f (i 1) 6= f (i)

hi = f (i1)<j6f (i)

e
otherwise
G
We call |BG| the classifying space of G. Group (co)homology is defined as H (G, A) =

Hsing
(BG, A), H (G, A) = Hsing (BG, A).
9

Example 2.5.4 (Hochschild homology). Let A be an algebra over a field k and M be Abimodule. Then we can form a simplicial module C (A, M ) by setting Cn (A, M ) = M An
and defining face maps and degeneracy maps as follows:
d0 (m, a1 , . . . , an ) = (ma1 , a2 , . . . , an )
di (m, a1 , . . . , an ) = (m, a1 , . . . , ai ai+1 , . . . , an ), i = 1, . . . , n 1
dn (m, a1 , . . . , an ) = (an m, a1 , . . . , an1 )
sj (m, . . . , an ) = (m, a0 , . . . , aj , 1, aj+1 , . . . , an )
Homology HH (A, M ) := H(C (A, M )) of this simplicial module is called Hochschild
homology of A with coefficients in bimodule M . We will consider this cohomology theory in
more details in section 2.
What is a (co)homology theory? A (co)homology theory should be a function of
two arguments: H(X, A ), where X is a nonabelian argument, and A is an object in
some abelian category. For example, X could be a topological space, algebra, group etc.
and usually A will be a sheaf, (bi)module, representation etc. The modern perspective is
that we should fix X and think of H(X, ) as a functor from some abelian category to
abelian groups. More formally, we have some non-abelian (that is, arbitrary) category C,
and an additive category A over C fibred in abelian categories. For example, we can consider
C = Top, and the fiber of A over Top being Sh(X).

2.6

Homology and cohomology with local coefficients

L
Recall that given a simplicial P
set X we defined Cn (X, A) = xXn Ax. That is, elements
a Cn (X, A) are of the form xX a(x) x with a(x) A. What if we allowed the a(x) to
live in different abelian groups? That is exactly what we will try to do!
Definition 2.6.1. A homological system of coefficients for X consists of
1. a family of abelian groups {Ax }xXn one for each simplex x Xn
2. a family of group homomorphisms {A (f, x) : Ax AX(f )x }xXn ,f : [m][n]
satisfying
1. A (id, x) = idAx for all x Xn
2. the following diagram commutes:
Ax
A (f g,x)

A (f,x)

/ AX(f )x


AX(f g)x

A (g,X(f )x)

AX(g)X(f )x

that is, A (f g, x) = A (g, X(f )x)A (f, x). (this is a cocycle condition).
10

Definition 2.6.2. Given (X , A ), define

M

Cn (X, A ) =

Ax x

xXn

We define a differential dn : Cn (X, A ) Cn1 (X, A ) by

X n+1
X
(1)i A (in1 , x) (a(x)) in x

!
X

dn

a(x) x

xXn i=0

xXn

One can check that the cocycle condition forces d2 = 0.

Definition 2.6.3. A cohomological system of coefficients is
1. a family of abelian groups {Bx }xXn
2. a family of group homomorphisms {B(f, x) : BX(f )x Bx }xXn ,f : [m][n]
satisfying
1. B(id, x) = idBx
2. B(f g, x) = B(f x)B(g, X(f )x)
Definition 2.6.4. Given a cohomological system (X , B), define
(
C n (X, B) =

functions f : Xn

Bx

xXn

We define dn : C n (X, B) C n+1 (X, B) by

(dn f )(x) =

(1)i B(in , x) f (in+1 x)

Example 2.6.5. The system of constant coefficients is Ax = A for all x Xn , with

A (f, x) = idA . One can verify that C (X, A ) = C (X, A).
Remark 2.6.6. The notion of a system of coefficients can be defined much more succinctly.
The category of simplicial sets is, as noted, just Set, i.e. the category
R Psh() of presheaves
R
on . For X Psh(), consider the category of elements over X, X. Objects of X
are pairs (n, x) where x Xn , and a morphism (n, x) (m, y) is just a nondecreasing map
f : [n] [m] such that X(f

= x. One can readily check that the category of coefficient
R )(y)
systems on X is AbPsh X .
11

Sheaves and their cohomology

3.1

Presheaves

Sheaves were originally considered by J. Leray.

Let X be a topological space. Define Open(X) to be the category of open sets in X.
That is, Ob (Open(X)) = {open sets in X}, and
(

if U 6 V
HomOpen(X) (U, V ) =
U , V if U V
Here and elsewhere, U V means that U is not necessarily a proper subset of V . We will
use notation U ( V if U is proper subset of V .
Definition 3.1.1. Let X be a topological space. A presheaf on X with values in a category
C is a contravariant functor F : Open(X) C.
Common categories are C = Set, Grp, Ring, . . . . Elements s F (U ) are called sections
of F over U , and F (X) is the set of global sections. One often writes (U, F ) instead of
F (U ), and thinks of (U, ) as a functor on F . From the definition, we see that for U V
we have maps VU : F (V ) F (U ); these are called the restriction maps from V to U . Since
F is a functor, these satisfy:
1. U
U = id for all open U .
V
W
2. if U V W , then W
U = U V .

There is an obvious notion of a morphism between presheaves namely a morphism

: F G is just a natural transformation from F to G . That is, = {(U ) : F (U )
G (U )}, and for U V open, the following diagram commutes:
F (V )
V
U

F (U )

3.2

(V )

(U )

/ G (V )


V
U

/ G (U )

Definitions

Definition 3.2.1. A presheaf F is a sheaf if given any open U X, any open cover
S
U

U = U , and {s F (U )} such that U

U U (s ) = U U (s ), there exists a unique
s F (U ) such that U
U (s) = s for all .
Remark 3.2.2. Sheaves are continuous functors, in the sense that they map colimits (in
Open(X)) to limits in Set. That is, the diagram
F (U V )

/ F (U ) F (V )

is exact, i.e. is an equalizer.

12

// F (U V )

(1.3)

We will mostly deal with abelian sheaves, that is sheaves of abelian groups. The following
are all examples of sheaves.
c (U ) = Hom
Example 3.2.3. Let X be a topological space. Set OX
Top (U, C).
diff
c
If X is a differentiable manifold, we can define OX OX , the sheaf of differentiable
diff (U ) be the ring of C -functions U C.
functions by letting OX
an O diff to
If X is a complex analytic manifold, e.g. X = P1 (C), then we can define OX
X
be the sheaf of holomorphic functions.
If we go even further and stipulate that X is an algebraic variety over C, then we can
alg
define OX
to be the sheaf of regular functions.

All the above sheaves are often called structure sheaves. Indeed, smooth manifolds,
analytic manifolds. . . can be defined to be topological spaces along with a sheaf of rings
satisfying certain properties.
For abelian sheaves, (1.3) is exact in the usual sense:
0

/ F (U V )

/ F (U ) F (V )

/ F (U V )

V U V and the second is (s, t) 7 U

V
where the first map is U
U
V
U V (s) U V (t).
Two basic problems in sheaf theory are the following:

Example 3.2.4 (Extensions of sections). Given a sheaf F and a section s F (U ) for

some open U , does there exist s F (V ) for some V U , such that VU (
s) = s? For
an , f F (U ) a holomorphic function. Then for every z U , there
example, let F = OCP
0
1
P
exists a neighborhood Uz0 U such that f (z) = n>0 an (z z0 )n converges for
P all z Uz0 .
For example, we could take the Riemann zeta-function defined by (s) = n>1 ns on
U = {s C : Re(s) > 1}. One can prove that can be extended analytically to all of
C \ {1}. There is a large class of similar functions (for example, Artin L-functions and the
L-functions of more general Galois representations) for which existence of analytic extensions
to all of C is an open problem.
Example 3.2.5 (Riemann-Roch). Compute (X, F ) for a given F . An important example
is when X is a compact Riemann surface and F = 1 is the sheaf of 1-forms. In this case,
dimC H0 (X, 1 ) = g, the genus of X. More generally, the Riemann-Roch theorem says that
if L is an invertible sheaf on X, then dim H0 (L ) dim H0 (1 L 1 ) = deg L g + 1.
Since H0 ( L 1 ) = H1 (L ) by Serre duality, we can write this as (L ) = deg L g + 1.

3.3

Kernels, images and cokernels

Definition 3.3.1. If F and G are sheaves, a morphism of sheaves : F G is just a

morphism of their underlying presheaves.
Consider the category AbSh(X) of abelian sheaves. This is clearly an additive category,
so the notions of kernel / image make sense. We can define, for : F G , two presheaves
K (U ) = Ker ((U ))
I 0 (U ) = Im ((U ))
13

The hope would be that K and I 0 are the category-theoretic kernel / image of .
Lemma 3.3.2. With the above notation, K is a sheaf, but I 0 is not a sheaf in general.
S
Proof. Given open U X, U = U an open cover, and s K (U ) such that
U

U
U U (s ) = U U (s ), since F is a sheaf, there exists a unique s F (U ) such that
U
U (s) = s for all . We need to show that s is actually in K . Since is a morphism, for
any we have
U
U
U ((U )s) = (U )(U (s)) = (U )(s ) = 0
since each s K (U ). Since G is also a sheaf, this force (U )(s) = 0, hence s K (U ).
To show that I 0 is not in general a sheaf, we give a counterexample. Let X = C = C\{0}
an O an given by f 7 f 0 = df . One can check
with the analytic topology. Consider : OX
X
dz
that for all x X, there exists a neighborhood Ux X with x Ux such that (Ux ) :
+
P
an (U ) O an (U ) is surjective. But, the equation df = g for g =
an ),
OX
an z n (X, OX
x
x
X
dz
n=

has a global solution if and only if a1 = 0, which implies that the function z1 is not in the
image Im ((X)) although z1 Im ((Ux )) for any x X. This violates the sheaf axiom.
The moral of this is that I 0 needs to be redefined in order to be a sheaf.
Definition 3.3.3. Given a morphism of sheaves : F G , define Im() by
Im()(U ) = {s G (U ) : x U, Ux U : U
Ux (s) Im (Ux )}
It is a good exercise to show that Im() actually is a sheaf, and is moreover the categorytheoretic image of . If one is more ambitious, it is not especially difficult to show that
AbSh(X) is an abelian category.

3.4

Germs, stalks, and fibers

Let X be a topological space, F an (abelian) presheaf on X.

Definition 3.4.1. Let x X. A germ of sections of F at x is an equivalence class of pairs
(s, U ), where U X is an open neighborhood of x and s F (U ). We say that (s, U ) and
V
(t, V ) are equivalent if there exists W U V such that U
W (s) = W (t).
One checks easily that what we have defined actually is an equivalence relation. The
stalk (or fiber) of F at x X is the set Fx of all equivalence classes of pairs (s, U ) with
x U , s F (U ). More formally,
Fx = lim F (U )

U 3x

Note that for any x U , there is a canonical map U

x : F (U ) Fx given y s 7 sx = [(s, U )].
14

Remark 3.4.2. There is an even more abstract characterization of Fx . Let E = Sh(X),

the category of sheaves of sets on X. This is a topos, i.e. it has all finite limits and colimits,
exponentials, and a subobject classifier. See [MLM92] for the definition of exponentials
and subobject classifiers. For arbitrary topoi X , S, one says that a geometric morphism
x : X S is a pair (x , x ) where x : S X and x : X S form an adjoint pair (with
x a x ) and x commutes with finite limits. (Since x is a left-adjoint, it already commutes
with all colimits.)
For X an arbitrary topos, call a geometric point a geometric morphism x : Set X .
Let |X | denote the class of geometric points of X . If X is a sober topological space (every
irreducible closed subset has a unique generic point) then there is a natural bijection
X | Sh(X)| that sends x X to the pair (x , x ) where x F = Fx and
(
S if x U
x S = Sx : U 7
otherwise
Topoi of the form X = Sh(X) have one very nice property: a morphism f : F G in X is
an isomorphism in X if and only if x f is an isomorphism for all x |X |. Such topoi are
said to have enough points.
It is possible to give |X | a topology in a canonical way. The functor X 7 Sh (|X |) can
be characterized as an adjoint for details, see [Hak72]
`
Definition 3.4.3. Let F be a presheaf on X. The total space of F is Et(F ) = xX Fx .
For s F (U ), define Et(F )(s) = {sx }xU F. We put the coarsest topology on Et(F )
such that each Et(F )(s) is open.
It is an easy consequence of the definitions that the projection map : Et(F )  X is
continuous (in fact, it is a local homeomorphism).
Example 3.4.4. Let f : Y X be any continuous map of topological spaces. We can
define the sheaf f on X of continuous local sections of f , i.e.
X U 7 f (U ) = {s HomTop (U, Y ) : f s = idU }
It is easy to check that f is a sheaf (without any hypotheses on f ).
Definition 3.4.5. Let F be a presheaf on X. Define the sheafification of F as F + = ,
where : Et(F ) X is the canonical projection.
There is a canonical morphism of presheaves F : F F + , called the sheafification
map. One defines F (U ) : F (U ) F + (U ) by s 7 (u 7 su ). If F is already a sheaf, then
: F F + is an isomorphism.
Remark 3.4.6. Let Sh(X) and PSh(X) be the categories of sheaves and presheaves on X.
Let : Sh(X) PSh(X) be the natural inclusion. One can characterize sheafification by
saying that it is the left-adjoint to . In diagrams:
()+ : PSh(X)  Sh(X) :
15

that is, HomSh(X) (F + , G ) ' HomPSh(X) (F , G ). The map F is the one induced by idF + ,
when we set G = F + .
Definition 3.4.7. For sheaves F and G with an embedding G , F define the quotient
F /G as sheafification of the presheaf U 7 F (U )/G (U ).

3.5

Coherent and quasi-coherent sheaves

In this section we will introduce some basic types of sheaves (finitely generated, coherent,etc.)
that we will widely use later.
Let (X, O) be a topological space with the structure sheaf O of continuous functions
on it.
Definition 3.5.1. A sheaf F on X of O-modules is called finitely generated (or of finite
type) if every point x X has an open neighbourhood U such that there is a surjective
morphism of restricted sheaves
O n |U  F |U , n N
In other words, locally such sheaf is generated by finite number of sections. That is, for any
x X and small enough open U 3 x, for any V U the abelian group F (V ) is finitely
generated as a module over O(V ).
Example 3.5.2. The structure sheaf O itself is of finite type, as well as O n .
Example 3.5.3. If F is finitely generated, then any quotient F /G and any inverse image
1 (F ) will be finitely generated.
Proposition 3.5.4. Suppose F is finitely generated. Suppose for some point x X and
open U 3 x the images of sections s1 , . . . , sn F (U ) in Fx generate the stalk Fx . Then
there exists an open subset V U s.t. the images of s1 , . . . , sn F (U ) in Fy generate Fy
for all y V .
Proof. Since F is finitely generated, there is some V 0 U and t1 , . . . , tm F (V 0 ) such
that t1 , . . . , tm generate Fy for anyP
y V 0 . Since Fx is also generated by s1 , . . . , sn , we
can express ti in terms of sj : ti =
aij sj , where aij Ox . There are finitely many aij ,
and since they are germs, there is a small open neighbourhood V 00 s.t. aij are actually
restrictions of some a
ij F (V 00 ). If we now take V = V 0 V 00 , sections si generate Fy for
any y V .
Corollary 3.5.5. If F is of finite type and Fx = 0 for some x X, then F |V = 0 for
some small open neighbourhood V of x.
Definition 3.5.6. A sheaf F is called quasi-coherent if it is locally presentable, i.e. for
every x X there is an open U X containing x s.t. there exist an exact sequence
O I |U O J |U F |U 0,
where I and J may be infinite, i.e. if F is locally the cokernel of free modules. If both I
and J can be chosen to be finite then F is called finitely presented.
16

Definition 3.5.7. A sheaf F is called coherent if it is finitely generated and for every open
U X and every finite n N, every morphism O n |U F |U of O|U -modules has a finitely
generated kernel.
Example 3.5.8. If X is Noetherian topological space, i.e. such that any chain V1 V2 . . .
of closed subspaces stabilizes, then the structure sheaf OX is coherent.
Example 3.5.9. The sheaf of complex analytic functions on a complex manifold is coherent.
This is a hard theorem due to Oka, see [Oka50].
Example 3.5.10. The sheaf of sections of a vector bundle on a scheme or a complex analytic
space is coherent.
Example 3.5.11. If Z is a closed subscheme of a scheme X, the sheaf IZ of all regular
functions vanishing on Z is coherent.
Lemma 3.5.12. If X is Noetherian then F is of finite type if and only if F is finitely
presented, if and only if F is coherent.
Lemma 3.5.13. For coherent and quasi-coherent sheaves the two out of three property
holds. Namely, if there is a short exact sequence
0

/F

/G

/H

/0

and two out of three sheaves F , G , H are coherent (resp. quasi-coherent), then the third
one is also coherent (resp. quasi-coherent).
Theorem 3.5.14. On affine variety (or better affine scheme) X with affine algebra of
functions A = OX (X) the global section functor gives equivalence of categories Qcoh(X)
Mod(A). Moreover, restriction of to coh(X) Qcoh(X) gives equivalence of categories
coh(X) fgMod(A), where fgMod(A) Mod(A) is a full subcategory of finitely generated
A-modules.
Proof. The inverse functor is given by tilde-construction. For details see, for example,
[EH00].

3.6

Motivation for sheaf cohomology

Let X be a topological space. Recall that a presheaf (of abelian groups) on X S

is a

contravariant functor F : Open(X) Ab. The presheaf F is a sheaf if whenever U = Ui

is an open cover, the sequence
0

/ F (U )

F (Ui )

F (Ui Uj )

i,j



Uj
Ui
where the first map is s 7 (U
.
Ui (s))i and the second is (si )i 7 Ui Uj (s) Ui Uj (t)
i,j

17

Warning. Some textbooks only require that this sequence be exact for finite open covers.
This does not yield the same notion of a sheaf. For example, let X = Cn and F be the
sheaf of bounded continuous C-valued functions. Then F satisfies the sheaf axiom for all
finite covers, but it is easily seen that F is not a sheaf.
In the previous section, we defined kernels and images of sheaves. This enables us to
define an exact sequence of sheaves. In particular, we can consider short exact sequences
/K

/F

/G

/0

(1.4)

The global section functor (X, ) : Sh(X) Ab is left exact. That is, if we apply
(X, ) to the sequence (1.4), then
0

/ K (X)

/ F (X)

/ G (X)

is exact. It can be proved directly.

Remark 3.6.1. Another way to prove that (X, ) : Sh(X) Ab is left exact is to notice
that (X, ) is right adjoint to the functor F : Ab Sh(X) that sends any abelian group
G to the constant G-valued sheaf on X. Then we can use the general fact that right adjoint
functor is left exact.
However, the morphism F (X) G (X) on the right may not be surjective (i.e. may
not be right exact).
Example 3.6.2. Let X be any compact connected Riemann surface (e.g. P1 (C)). Let
an , the sheaf of holomorphic functions defined earlier. For = {x , . . . , x } X, we
F = OX
1
n
define a sheaf G by
M
G (U ) =
C [xi ]
xi U

P
where [xi ] is a formal basis element. Define : F G by (U )f = xU f (xi ) [xi ]. Set
K = Ker(). Taking global sections, we obtain (G ) = Ck , and, by Louivilles theorem,
an ) = C. If k > 1, then it is certainly not possible for (F ) (G ) to be
(F ) = (X, OX
surjective, even though it is trivial to check that is surjective (at the level of sheaves).
Given an exact sequence 0 K F G 0, we can apply the global sections
functor to obtain an exact sequence 0 (X, K ) (X, F ) (X, G ). The main idea
of classical homological algebra is to canonically construct groups Hi (X, F ) that extend the
exact sequence on the right:
0 (K ) (F ) (G ) H1 (K ) H1 (F ) H1 (G ) H2 (K )
Here, and elsewhere, we will write (F ) and Hi (F ) for (X, F ) and Hi (X, F ) when X is
clear from the context.
18

3.7

Sheaf cohomology

In this section we will define sheaf cohomology using the classical Godement resolution, and

compare it to Cech
cohomology.
Definition 3.7.1. A sheaf F is called flabby if for all U X, the restriction map
X
U : F (X)  F (U ) is surjective.
Note that if F is flabby, then the sequence
0

/ F (U U 0 )

/ F (U ) F (U 0 )

/ F (U U 0 )

/0

is exact on the right. Indeed, given any r F (U U 0 ), we can choose r F (X) with
U U 0 (
r) = r. Then (X
r), 0) maps to r.
U (
Lemma 3.7.2. Let 0 K F G 0 be an exact sequence of sheaves, and assume
K is flabby. Then 0 K (X) F (X) G (X) 0 is also exact.
Proof. Recall that the surjectivity of : F G implies that (X) : F (X) G (X)
is locally surjective. That is, for all t G (X) and for all x X, there exists an open
neighborhood U of x and s F (U ) such that X
U (t) = (U )(s). Assume now that a
given t G (X) lifts to local sections s F (U ) and s0 F (U 0 ). Put (U )s = X
and
U (t)
0 agree over U U 0 , i.e. U
U0
0 ),
(U 0 )s0 = X
(t).
If
it
happens
that
s
and
s
(s)
=

(s
U0
U U 0
U U 0
then we can glue s and s0 along U U 0 . Unfortunately s and s0 do not always agree over
U0
0
U U 0 . However, if we let r = U
U U 0 (s) U U 0 (s ), then
0

U
0
(U U 0 )(r) = U
U U 0 ((U )(s)) U U 0 ((U )(s))
0

X
U
X
= U
U U 0 U (t) U U 0 U 0 (t)

=0
In other words, r K (U U 0 ). Since K is flabby, there exists r K (X) such that
X
r) = r. Using r, we can correct s0 by replacing it with s00 = s0 + X
r) F (U 0 ).
U (
U U 0 (
Now
0

00
U
0
U
X
U
r)
U U 0 (s ) = U U 0 (s ) + U U 0 U 0 (
0

0
X
= U
r)
U U 0 (s ) + U U 0 (

= U
U U 0 (s)
0

U = s and U U (
thus there exists s F (U U 0 ) such that U
s) = s00 . By (transfinite)
U
U0
induction, the result follows.

Lemma 3.7.3. There are enough flabby sheaves. More precisely, for any abelian sheaf
F , there is a (functorial) embedding  : F C 0 (F ), where C 0 (F ) is flabby.
Q
Proof. Define C 0 (F )(U ) = xU Fx and (U )(s) = (sx )xU , where sx = U
x (s) is the germ
0
of s at x. It is not difficult to check that C (F ) is actually a flabby sheaf.
19

The assignment F 7 C 0 (F ) is actually an exact functor C 0 : Sh(X) Sh(X).

Let C 1 (F ) = C 0 (Coker ) = C 0 (C 0 (F )/F ). There is a canonical map d0 : C 0 (F )
C 0 (F )/F , C 0 (C 0 (F )/F ) = C 1 (F ). This procedure can be iterated: assume we have
defined (C k (F ), dk1 ) for all k 6 n. Then set


C n+1 (F ) = C 0 C n F /dn1 C n1 F
dn = C n  C n F /dn1 C n1 F , C 0 (C n /dn1 C n1 ) = C n+1
Definition 3.7.4. The complex
d0

d1

C (F ) : 0 0 C 0 (F ) C 1 (F )
together with the morphism of complexes  : F C (F ), where we regard F as the complex
0 0 F 0 , is called the Godement resolution of F .
Note that by construction, Hn (C (F )) = 0 for all n > 0, and H0 (C (F )) ' F .
Equivalently, we can say that  is a quasi-isomorphism. Applying (X, ) termwise to
C (F ), we get a new complex (X, C (F )) which may not be acyclic.
Definition 3.7.5. The cohomology of X with coefficients in F is Hn (X, F ) = Hn (C (F )(X)).
It follows immediately from the definition that Hn (X, F ) = 0 if n < 0, and that
H0 (X, F ) ' (X, F ) = F (X) canonically.
Theorem 3.7.6. Given any short exact sequence of sheaves
0

/K

/F

/G

/0

(1.5)

there is a long exact sequence, functorial in (1.5)

/ Hn (X, K )

/ Hn (X, F )

/ Hn (X, G )

/ Hn+1 (X, K )

Proof. We know that 0 K F G 0 is exact. It is not hard to show that C is an

exact functor, so 0 (K ) C (F ) C (G ) 0 is also exact. Since C takes sheaves
to complexes of flabby sheaves, Lemma 3.7.2 shows that 0 C (K )(X) C (F )(X)
C (G )(X) 0 is exact. By Theorem 3.7.7 it follows that Hn (X, K ) Hn (X, F )
Hn (X, G ) Hn+1 (X, K ) is exact.
Theorem 3.7.7 (main theorem of homological algebra). If 0 K F G 0 is
a termwise exact sequence of complexes of abelian groups, there is a natural exact sequence
in cohomology

/ Hn (K )

/ Hn (F )

/ Hn (G )

/ Hn+1 (K )

where all the maps but are the obvious induced ones, and is canonically constructed.
20

Definition 3.7.8. A sheaf F is acyclic if Hn (X, F ) = 0 for all n > 0.

By Lemma 3.7.2, flabby sheaves are acyclic. It is a good exercise to show that if X is an
irreducible topological space, then any constant sheaf on X is flabby, hence acyclic. Recall
that X is reducible if there is some decomposition X = X1 X2 , where the Xi are nonempty
proper closed subsets.
Remark 3.7.9. In classical topology, many interesting invariants of a space X appear as
H (X, Z), where here Z represents the constant sheaf. In algebraic geometry, algebraic
varieties (equipped with the Zariski topology) are generally irreducible, so this construction
is completely useless. There are two ways to fix this. One is to use the etale topology (or
some other Grothendieck topology). Alternatively, one can replace constant sheaves by
(quasi-)coherent sheaves. The latter idea is due to Serre.

3.8

Applications

Recall that a sheaf homomorphism : F G is an epimorphism if we can locally lift

sections of G to F . The obstructions to lifting global sections live in the first sheaf
cohomology group H1 (X, F ).
Consider the projective space Pnk over
Example 3.8.1 (M. Noethers AF+BG Theorem).


n+1
n
some field k. Recall that Pk = Ak \ {0} /k as a set. Explicitly, elements of Pnk are
equivalence classes of tuples (x0 , . . . , xn ) 6= 0, where (x0 , . . . , xn ) (x0 , . . . , xn ) for all
k . We will write (x0 : : xn ) for the equivalence class of (x0 , . . . , xn ) in Pnk . Let
: An+1
\ {0}  Pnk be the canonical projection. For any integer m Z, define a sheaf
k
OPn (m) by
OPn (m)(U ) = {regular functions on 1 (U ) that are homogeneous of degree m}
Note that OPn (0) = OPn , the structure sheaf of Pn . For each m, OPn (m) is a sheaf of OPn modules. For any sheaf F of OPn -modules, we can define F (m) = F OPn OPn (m). It is easy
to check that multiplication induces an isomorphism OPn (m) OPn OPn (m0 ) OPn (m + m0 ).
As an application, let C1 and C2 be curves in P2k given by


Ci = V (Fi ) = (x0 : x1 : x2 ) P2 : Fi (x0 : x1 : x2 ) = 0
where the Fi are homogeneous polynomials of degrees, say, deg F1 = m and deg F2 = n. For
simplicity, we will assume that C1 and C2 intersect transversely. Let C be another curve that
passes through all intersection points of the curves C1 and C2 . We dont assume that the
Ci are smooth. Write C = V (F ) for some homogeneous polynomial F . Then Max Noether
proved that F = A1 F1 + A2 F2 for some homogeneous polynomials Ai . We will prove this
using sheaves.
Proof. Let X = P2 , and let I be the ideal sheaf of C1 C2 . One has I (U ) = {a OX (U ) :
a(c) = 0c C1 C2 }. We can define I (k) = I OP2 OP2 (k) for any k Z. So (X, I (k))
21

is the set of homogeneous forms of degree k that vanish on C1 C2 . There is the following
exact sequence, which is actually a locally free resolution of I (k).
0

/ O 2 (k m n)
P

/ O 2 (k m) O 2 (k n)
P
P

/ I (k)

/0

Here, (c) = (F2 c, F1 c) and (a, b) = F1 a + F2 b. The theorem we are trying to prove
simply asserts the surjectivity of on global sections. Taking global sections, we get
0

/ (O(k m n))

/ (O(k m)) (O(k n))

/ (I (k))

/ H1 (O(k m n))

We will see that H1 (Pn , OPn (k)) = 0 for all n > 1 and all k Z. This yields the result.
Example 3.8.2 (Exponential Sequence). Let X be a complex analytic manifold, for example

a Riemann surface. Let OX be the structure sheaf of X. Let OX

be the sheaf of holomorphic
functions X C = C \ {0}, with the natural multiplicative structure. Then there is an
exact sequence
0

/Z

/ OX exp / O
X

/1

where Z is the constant sheaf and exp(s) = e2is . The obstructions to lifting sections from

OX
to OX lie in the first cohomology group H1 (X, Z), which does not vanish in general.
In fact, in cohomology we get an exact sequence
...

/ H1 (X, Z)

/ H1 (X, OX )

/ H1 (X, O )
X

/ H2 (X, Z)

/ H2 (X, OX )

/ ...

(1.6)

If X is complete (i.e. compact in this case), then the global sections H0 (X, OX
) is C , and

the map H0 (X, OX

) H1 (X, Z) will actually be zero, so we can put 0 at the beginning of
(1.6).
From the geometric point of view the most important is the term in the middle which is

denoted by Pic(X) = H1 (X, OX

). The elements of this group classify (up to isomorphism)
all invertible sheaves, i.e. locally free sheaves of rank = 1 on X. All such sheaves are
actually sheaves of sections of holomorphic line bundles on X. By definition, for a line
bundle [L] Pic(X) we call the element c1 (L) = ([L]) H2 (X, Z) the Chern class of L.
Assume now that X is complete. Then right hand side of (1.6) gives the famous HodgeLefschetz Theorem which asserts that an integral cohomology class c H2 (X, Z) represents
Chern class c1 (L) of some line bundle L if and only if c vanishes in H2 (X, OX ).
Next, let Pic0 (X) = Ker() Pic(X). From (1.6) we see that Pic0 (X) ' H1 (X, OX )/ H1 (X, Z)
which is the quotient of finite dimensional vector space over C modulo lattice of finite rank.
The natural complex structure on H1 (X, OX ) descends to Pic0 (X) making it an analytic
variety called the Picard variety of X. A deeper fact is that if X is an analytification of a
projective algebraic variety then so is Pic0 (X).
The quotient group NS(X) := Pic(X)/ Pic0 (X) is called the Neron-Severi group. Since
NS(X) embeds into H2 (X, Z) via , it is finitely generated.
22

3.9

Cech
cohomology

Recall that if X is a topological space, U = {U }I an open covering, we defined a simplicial

set (the nerve of U) by


Xn = (0 , . . . , n ) I n+1 : U0 Un 6=
and for f Hom ([m], [n]):
X(f ) : Xn Xm ,

(0 , . . . , n ) 7 (f (0) , . . . , f (m) )

Given any abelian sheaf F , we define a cohomological system of coefficients for X , by

Bx = F0 ...n = F (U0 Un )
and B(f, x) : BX(f )x Bx as restriction maps
U

F (f, (0 , . . . , n )) : F (Uf (0) Uf (m) )

U
f (0)
f (m)
U
n
0

/ F (U U )
n
0

Definition 3.9.1. The Cech

cohomology of U with coefficients in F is

H(U,
F ) := H (X , B)
Definition 3.9.2. A covering U is called F -acyclic if Hi (U0 Un , F ) = 0 for all
0 , . . . , n I and i > 0.
The following result allows one to establish acyclicity of some coverings.
Theorem 3.9.3 (H.Cartans criterion). Let A be a class of open subsets of a topological
space X such that
(a) A is closed under finite intersections, i.e.
U1 , . . . , Un A U1 Un A
(b) A contains arbitrary small open subsets, i.e. for any open U there is V ( U such that
V A.
Suppose that for any U A and A-covering U = {Ui } of U , Hi (U, F ) for all i > 0.
Then any A-covering is F -acyclic. In particular, for any A-covering of the space X there is
isomorphism

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

23

24

Chapter 2

Standard complexes in algebra

1

Group cohomology

1.1

Definitions and topological origin

Recall that given a (discrete) group G, we define (BG)n = Gn , the n-fold cartesian product of
G with itself. For f : [m] [n], we define BG(f ) : Gn Gm by (g1 , . . . , gn ) 7 (h1 , . . . , hm ),
where
(Q
f (i1)<j6f (i) gj if f (i 1) 6= f (i)
hi =
1
otherwise
This is a simplicial set. Next, given a representation of G in an abelian group A (i.e. a
homomorphism G Aut(A)), define a cohomological system of coefficients:
Bx = A

for all x BG

B(f, x)(a) = ha
where for x = (g1 , . . . , gn ) Gn , we set h =

Qf (0)
j=1

gj if f (0) 6= 0, and h = 1 otherwise.

Definition 1.1.1. With the above notation, we define C (G, A) = C (BG, B), and define
the cohomology of G with coefficients in A to be H (G, A) = H [C(G, A)].
dn

Explicitly, C (G, A) has C 0 (G, A) = A, and for n > 1, C n (G, A) = HomSet (Gn , A), with
: C n C n+1 defined by

n
X
(df )(g1 , . . . , gn+1 ) = g1 f (g2 , . . . , gn+1 )+ (1)i+1 f (g1 , . . . , gi gi+1 , . . . , gn+1 )+(1)n+1 f (g1 , . . . , gn )
i=1

There is a topological interpretation of H (G, A) if A has trivial G-action. Suppose G

acts continuously on a topological space X. Let Y = G\X be the orbit space, with the
quotient topology, and let : X Y be the projection map.
25

Theorem 1.1.2. If X is a contractible space, G acts freely on X (so is a principal

G-bundle over Y ), then Hn (G, A) = Hn (Y, A), where in the second term A is viewed as the
constant sheaf on Y .
Proof (sketch). Under our assumptions, we have 1 (Y ) ' G and i (Y ) = 0 for i > 2.
Moreover, X Y is a universal cover for Y . From topology, we know that all spaces
with 1 (Y ) = G and >1 (Y ) = 0 are homotopy equivalent. They are often denoted by
K(G, 1), and called the first Eilenberg-Mac Lane space of G. We know that |BG| satisfies
the conclusions of the theorem, so the proof is complete.
Remark 1.1.3. For the data (G, A) we can define a homological system of coefficients A
with Ax = A and A (f, x) : A A by a 7 h1 a, where x = (g1 , . . . , gn ) and h is the same
as above.
Definition 1.1.4. Let the notation be as above. The homology of G with coefficients in A
is H (G, A) = H (C(BG, A )).

1.2

Interpretation of H1 (G, A)

We would like to interpret H1 (G, A) and H2 (G, A) in terms of more familiar objects. Recall
that an extension of G by N is an exact sequence of groups:
/N

/E

/G

/1

We say that two extensions are equivalent if there is a commutative diagram:

/N

/G

/1

/G

/1

/ E0

/N

1
i

/E

An extension 1 N
E
G 1 is split if it splits on the right, i.e. there is a
homomorphism s : G E such that s = idG .
Lemma 1.2.1. Let A be an abelian group with is also a G-module. Then any split extension
of G by A is equivalent to the canonical one:
0

/A

/AoG

/G

/ 1.

Recall that AoG, the semidirect product of G and A, is AG as a set, with (a, g)(b, h) =
(a + gb, gh). The first cohomology group H1 (G, A) classifies splittings up to A-conjugacy.
Every splitting s : G A o G is of the form g 7 (dg, g), where d is some map G A. The
fact that s is a group homomorphism forces (dg, g) (dh, h) = (dg + gdh, gh) = (d(gh), gh).
Thus we need d(gh) = dg + gdh. It would be natural to write dg h + g dh, but A is not
a G-bimodule. If it were, then this condition would require d : G A to be a derivation.
26

Definition 1.2.2. A map d : G A is called a derivation on G with coefficiens in A if

d(gh) = dg + gdh for all g, h G.
Again, if A is a G-bimodule, then we require the Leibniz rule to hold, i.e. d(gh) =
dg h + g dh.
Definition 1.2.3. Two sections s1 , s2 : G A o G are said to be A-conjugate if there is
an a A such that
s2 (g) = (a)s1 (g)(a)1
for all g G.
It is a good exercise to check that Z 1 (G, A) = {f C 1 (G, A) : d1 f = 0} is the set of
derivations d : G A, i.e. Z 1 (G, A) = Der(G, A). If we write s1 g = (d1 g, g), s2 g = (d2 g, g),
then the definition of A-conjugacy means that for some a, we have d2 g d1 g = g a a
for all g. Once again, if A were a G-bimodule, we would want d2 g d1 g = g a a g, i.e.
d2 d1 = [, a].
Definition 1.2.4. A derivation d : G A is called inner if dg = ag g for some a A.
It is a good exercise to check that if B 1 (G, A) := Im(d0 ), then B 1 (G, A) is exactly the
set of inner derivations.
Theorem 1.2.5. The set of splittings of the canonical extension of G by A up to A-conjugacy
is in natural bijection with H1 (G, A).

1.3

Interpretation of H2 (G, A)

Theorem 1.3.1. The set of equivalence classes of extensions of G by A is in natural bijection

with H2 (G, A).
Proof. Recall that H2 (G, A) = Z 2 (G, A)/B 2 (G, A), where
Z 2 (G, A) = {f : G2 A : g1 f (g2 , g3 ) + f (g1 , g2 g3 ) = f (g1 , g2 ) + f (g1 g2 , g3 )}
We will interpret this as a kind of associativity condition. Given a cocycle f Z 2 (G, A),
we define an extension of G by A explicitly as follows. We have
0

/ A of G

/G

/0

where A of G = A G as a set, and (a1 , g1 ) (a2 , g2 ) = (a1 + g1 a2 + f (g1 , g2 ), g2 g2 ). One

can check that the associativity
(a1 , g1 ) ((a2 , g2 ) (a3 , g3 )) = ((a1 , g1 ) (a2 , g2 )) (a3 , g3 )
27

is equivalent to g1 f (g2 , g3 ) + f (g1 , g2 g3 ) = f (g1 , g2 ) + f (g1 g2 , g3 ). In addition, for (0, 1)

to be the identity element in A of G, we need to impose the normalisation condition
f (g, 1) = 0 = f (1, g) for all g G. Thus we have a map

extensions 0 A
E
G1
{normalized 2-cocycles}
with a (set-theoretic) normalized

section s : G E s.t. s(1) = 1

where a normalized 2-cocycle f maps to A of G along with the section s : G A of G

given by g 7 (0, g). The inverse of this map associates to an extension E with normalized
section s the map


f (g1 , g2 ) = i1 s(g1 )s(g2 )s(g1 g2 )1
As an exercise, check that choosing a different section s corresponds to changing f by a
2-boundary.
S
Example 1.3.2 (Cyclic groups). Let G = Z/2, and let G act on X = S = n>1 S n by
reflection. Then X/G = Y = RP , and from topology we know that 1 (Y ) = Z/2 and
i (Y ) = 0 for i > 2. Thus RP = K(Z/2, 1), and thus

Hp (Z/2, Z) = Hp (RP , Z) = 0

Z/2

if p = 0
if p 0 (mod 2) and p > 2
otherwise

On the other hand, H2 (Z/2, Q) = 0 for all p > 1.

Algebraically, let G = Z/n, and consider the complex of Z[G]-modules
0
where G = hti and N (1) =

/Z

Pn1
i=0

/ Z[G] 1t / Z[G]

/Z

/0

ti . This gives us an infinite resolution

/ Z[G]

/ Z[G]

/ Z[G] 1t / Z

/0

One can check that this is a projective resolution of Z as a Z[G]-modules, and it yields

Z
Hp (Z/n, Z) = 0

Z/n

if p = 0
if p 0 (mod 2) and p > 2
otherwise
28

Hochschild (co)homology

2.1

The Bar complex

For the rest of this section, let k be a field, and A be a associative, unital k-algebra. Also, let
M be an A-bimodule (also called a two-sided module), i.e. we have (am)b = a(mb). Define
enveloping algebra of A by Ae = A k Ao , where A denotes the opposite algebra of A. It
is easy to see that the category of A-bidmodules is equivalent to the categories of left and
right Ae -modules. Indeed, we define
(a b )m = amb
m(a b ) = bma
Example 2.1.1. Consider M = Ae as a module over itself. It is naturally a Ae -bimodule
in two different (commuting) ways. We can compute explicitly:
(a b )(x y) = ax b y = ax yb

outer structure

(x y)(a b ) = xa yb = xa by

inner structure

e A to be the complex
Consider the multiplication map m : A A A. Define B
[

e A :=
B

/ A3

/ A2

/A

/0

where b : A(n+1) An is given by

b(a0 , . . . , an ) =

n1
X

(1)i a0 ai ai+1 an

i=1

It is an easy exercise to check that b2 = 0.

Definition 2.1.2. Let Bn A = A An A. The bar complex of A is the complex of
A-bimodules (with outer structure):
B A =

/ B2 A

/ B1 A

/ B0 A

/0

Write m : B A A, where A is regarded as a complex supported in degree zero. This

is actually a morphism of complexes because m b0 = 0 by associativity. We call B A
the bar resolution of A as a A-bimodule.
Lemma 2.1.3. The morphism m : B A A is a quasi-isomorphism.
e A is exact. We use the fact that if the identity on B
e A
Proof. It is equivalent to say that B
e A is quasi-isomorphic to A. So we want to construct maps
is homotopic to zero, then B
n
(n+1)
hn : A A
such that id = b0 h + h b0 . Define
hn (a1 an ) = 1 a1 an
29

We now compute
0

(h b )(a0 an ) =

n1
X

(1)i 1 a0 ai ai+1 an

i=0

(b0 h)(a0 an ) = 1 a0 a1 an +

n1
X

(1)i+1 1 a0 ai ai+1 an

i=0

It follows that (h b0 + b0 h)(a0 an ) = a0 an , as desired.

2.2

Differential graded algebras

Definition 2.2.1. A chain differential graded (DG) algebra (resp. cochain DG algebra)
over a field k is a Z-graded k-algebra, equipped with a k-linear map d : A A1 (resp.
d : A A+1 ) such that
1. d2 = 0
2. d(ab) = (da)b + (1)|a| adb for all a, b A with a homogeneous.
Here |a| denotes the degree of a in A . We call the second L
requirement the graded Leibniz
rule. (Recall that a graded algebra is a direct sum A = iZ Ai such that 1 A0 and
Ai Aj Ai+j .)
A DG algebra A is called non-negatively graded if Ai = 0 for all i < 0. In addition, if
A0 = k then A is called connected. We let DGAk denote the category of all DG k-algebras,
and DGA+
k denote the full subcategory of DGAk consisting of non-negatively graded DG
algebras.
Example 2.2.2 (Trivial DG algebra). An ordinary associative algebra A can be viewed as
DG algebra with differential d = 0 and grading A0 = 0, Ai = 0 for i =
6 0. Hence the category
Algk of associative algebras over k can be identified with a full subcategory of DGAk .
Example 2.2.3 (Differential forms). Let A be a commutative k-algebra. The
L de Rham
algebra of A is a non-negatively graded commutative DG algebra (A) = n0 n (A)
defined as follows. First, we set 0 (A) = A and take 1 (A) to be the A-module of K
ahler
differentials. By definition, 1 (A) is generated by k-linear symbols da for all a A (so
d(a + b) = da + db for , k) with the relation
d(ab) = a(db) + b(da), a, b A.
It is easy to show that 1 (A) is isomorphic (as an A-module) to the quotient of A A
modulo the relations ab c a bc + ca b = 0 for all a, b, c A. Then we define n (A)
using the exterior product over A by
V
n (A) := nA 1 (A)
30

Thus n (A) is spanned by the elements of the form a0 da1 dan , which are often
denoted simply by a0 da1 . . . dan and called differential forms of degree n.
The differential d : n n+1 is defined by
d(a0 da1 . . . dan ) := da0 da1 . . . dan
The product on the space (A) is given by the formula
(a0 da1 . . . an ) (b0 db1 . . . bm ) = a0 b0 da1 . . . dan db1 . . . dbm
This makes (A) a differential graded algebra over k. If X is a complex variety and
A = O(X) is the algebra of regular functions, then (A) = (X), where (X) is
the algebra of regular differential forms on X. However, if M is a smooth manifold and
A = C (M ), then the natural map (A) (M ) is not an isomorphism. Indeed, David
Speyer pointed out that if f, g are algebraically independent in A, then df and dg are linearly
independent in (A). (see the discussion before Theorem 26.5 in [Mat89]). Since ex and 1
are algebraically independent, d(ex ) and d(1) = dx are linearly independant over A = C (R)
in 1 (A). But certainly d(ex ) = ex d(1) in 1 (R).
Example 2.2.4 (Noncommutative differential forms). The previous example can be generalized to all associative (not necessarily commutative) algebras.
Suppose A is an (associative) algebra over a field k. First we define noncommutative
K
ahler differentials 1nc (A) as the kernel of multiplication map m : A A A:
0

/ 1 (A)
nc

/AA

/A

/0

So 1nc (A) is naturally an A-bimodule. Then we can define DG algebra of noncommutative

differential forms nc (A) as the tensor algebra T (1nc (A)):
nc (A) := T (1nc (A)) = A 1nc (A) 1nc (A)2 . . .
Differential d on nc (A) is completely defined by the derivation
d0 : A 1nc (A) d0 (a) = a 1 1 a Ker(m) = 1nc (A)
0
Indeed, there exists unique differential d : nc (A) +1
nc (A) of degree 1 that lifts d .
Explicitly it can be defined by the following formula:

d(a0 a
1 a
n ) = 1 a
0 a
1 a
n
Here for a A we denote by a
the element d0 (a) = a 1 1 a 1nc (A).
Actually, there is more conceptual way of defining Kahler differential and noncommutative
forms. Consider the functor Der(A, ) : A-bimod Sets associating to a bimodule M the set
of all derivations Der(A, M ). This functor is representable. Precisely, we have the following
31

Proposition 2.2.5. For every A-bimodule M there exists canonical isomorphism

Der(A, M ) ' HomA-bimod (1nc (A), M )
For any DG algebra B let B0 be its 0-component. This is just an ordinary algebra.
Then noncommutative differential forms can be also described by the following universal
property.
Proposition 2.2.6. For any associative k-algebra A and any k-algebra B there is natural
isomorphism
HomDGA+ (nc (A), B) ' HomAlgk (A, B0 ).
k

This proposition says essentially that the functor nc () : Algk DGA+

k is left adjoint to
+
the forgetful functor ()0 : DGAk Algk . For more details on noncommutative differential
forms see paper by Cuntz-Quillen [CQ95], or lecture notes by Ginzburg [Gin05].
Definition 2.2.7. If A is any graded algebra and d : A A is a derivation of A we say
that d is even or odd if one of the following holds:
d(ab) = (da)b + a(db)

(even)

d(ab) = (da)b + (1)|a| adb

(odd)

Lemma 2.2.8. Any derivation d (even or odd) is uniquely determined by its values on the
generators of A as a k-algebra. In other words, if S A is a generating set and d1 (s) = d2 (s)
for all s S, then d1 = d2 .
Proof. Apply iteratively the Leibniz rule.
Corollary 2.2.9. If d : A A is an odd derivation and d2 (s) = 0 for all s in some
generating set of A, then d2 = 0 on all of A.
Proof. If d is an odd derivation, then d2 is an even derivation. Indeed, d2 = 12 [d, d]+ , or
explicitly
d2 (ab) = d((da)b + (1)|a| adb)
= (d2 a)b + (1)|da| dadb + (1)|a| dadb + (1)|a|+|a| ad2 b
= (d2 a)b + a(d2 b)
The result follows now from the previous lemma.
Definition 2.2.10. If (A , d) is a DG algebra define the set of cycles in A to be
Z (A, d) := {a A : da = 0}
Notice that Z (A, d) is a graded subalgebra of A.
32

Moreover, define the set of boundaries B (A, d) to be

B (A, d) = {b A : b = da for some a A}
Then B (A, d) is a two-sided graded ideal in Z (A). Thus the quotient
H (A) = Z (A)/B (A)
is a graded algebra, called the homology algebra of A. For trivial reasons, the differential is
zero on H (A). This gives us a functor from the category of DG algebras to the category of
graded algebras.

2.3

Why DG algebras?

Let A be a k-vector space equipped with an (arbitrary) bilinear product A A A , or

equivalently a linear map : A A A, (x, y) 7 xy . Assume that dimk A < . Then,
we have the commutative diagram
A
can

/ (A A) o

Tk (A )

! d

A A


(2.1)

can

/ Tk (A )

In this diagram, : A (AA) is the linear map dual to , the map A A (AA) is
given by f g 7 [x y 7 f (x)g(y)] and it is an isomorphism because A is finite-dimensional.
By Lemma 2.2.8, any linear map A Tk (A ) determines a derivation d : T (A ) T (A ):
precisely, there is a unique d : T (A ) T (A ) such that
(1) d|A =
(2) deg(d) = +1
(3) d satisfies the graded Leibniz rule
Conversely, if d : T (A ) T (A ) satisfies (2) (3), then restricting d|A : A A A
and dualizing d : [A A ] ' A A A we get a linear mapping A a A.
Thus, if A is finite-dimensional, giving a bilinear map A A A is equivalent to giving
a derivation of degree 1 on T (A ).
Remark 2.3.1. For notational reasons, one usually takes = d, so that : A (A A)
is given by ()(x y) = (xy), A , x, y A.
Lemma 2.3.2. The map : A A A is associative if and only if 2 = 0 on T (A ).
33

Proof. Take any A and x, y, z A. Then we have

(()) (x y z) = ()(xy z) (1)()(x yz)
= ()(x yz) ()(xy z)
= (x(yz)) + ((xy)z)
= ((xy)z x(yz)) .
Hence is associative iff 2 = 0 on A iff 2 = 0 on T (A ). To get the last iff we used
Corollary 2.2.9.
To sum up, giving a finite-dimensional associative k-algebra is equivalent to giving a
free connected DG algebra which is generated by finitely many elements in degree 1. It is
therefore natural (and for many purposes, useful) to think of all finitely generated free DG
algebras, including the ones having generators in degree 1, as a categorical closure of the
finite-dimensional associative algebras.

2.4

Interpretation of bar complex in terms of DG algebras

Example 2.4.1. Define Ahi = A k k[], where is an indeterminate. Here A k B denotes

the coproduct in the category of (not necessarily commutative) k-algebras, which is given by
the free product of algebras. Assume |a| = 0 for all a A, and suppose || = 1. This makes
Ahi a graded algebra whose elements look like
a1 n1 a2 n2 ak
Since n = 111 . . . 1, any element in Ahi can be written as a1 a2 ak , i.e. is a
e A via
separator (or bar, if we write ab as a | b). We can identify Ahi with B
a1 a2 an 7 a1 an
en1 A = An and a1 a2 an also has degree
This actually is degree-preserving because B
n 1. Define the differential on Ahi by
a A

da = 0,
d = 1

This makes Ahi a DG algebra. Notice, that since d(a) = 0 and |a| = 0 for a A, then d
e A, b0 ). Indeed, we have
is A-linear. DG algebra Ahi is isomorphic as a complex to (B
d(a0 a2 . . . an ) = a0 d()a1 a2 . . . an a0 d(a1 . . . an )
= a0 a1 a2 . . . an a0 a1 d()a2 . . . an + a0 a1 d(a2 . . . an ) =
= a0 a1 a2 . . . an a0 a1 a2 . . . an + a0 a1 d(a2 . . . an )
= ...
n
X
=
(1)i a0 . . . ai ai+1 . . . an ,
i=0

34

which exactly maps to the differential b0 (a0 a2 an ) via the identification map .
Notice that 1 = d, so 1 = 0 in H (Ahi), hence H (Ahi) = 0. Under our identification of
e A with Ahi, the homotopy h is just u 7 u. Indeed, we can check that for all u Ahi,
B
(dh + hd)(u) = (du + (1)1 du) + du = 1 u du = u
Given an A-bimodule M , define M Ae B A to be the complex
M Ae A(n+2)
Note that M Ae A(n+2) ' M Ae Ae An ' M k An via the map
m Ae (a0 an+1 ) 7 an+1 ma0 (a1 an )
The induced differential b : M k An M k An1 turns out to be
m1
X

(1)i ma1 ai ai+1 an +(1)n an ma1 an1

m(a1 an ) 7 ma1 a2 an +

i=1

2.5

Hochschild (co)homology: definitions

Definition 2.5.1. The Hochschild homology of A with coefficients in M is

HH (A, M ) = H (M Ae B A)
To define Hochschild cohomology we need the notion of the morphism complex.
Definition 2.5.2. Let A be a ring (or k-algebra), and (M , dM ), (N , dN ) two complexes
of left A-modules. Set
M
HomA (M, N ) =
HomA (M, N )n ,
nZ

where
HomA (M, N )n = {f HomA (M, N ) : f (Mi ) Ni+n for all i Z}
is the set of A-module homomorphisms M N of degree n.
Warning In general, HomA (M, N ) 6= HomA (M, N ), i.e. not every A-module map f :
M N can be written as a sum of homogeneous maps.
L
Example 2.5.3. Let A = k be a field, N = k and M = V = nZ Vn a graded k-vector
space such that dim Vn > 1 for all n. Let f : V k be such that f (Vn ) 6= 0 for infinitely
many n. Then f 6 Homk (V, k). Why?
35

Exercise Prove that if M is a finitely generated (as A-module) then HomA (M, N ) =
HomA (M, N ).
Definition 2.5.4. For A a ring and M, N chain complexes over A, define
dHom : HomA (M, N )n HomA (M, N )n1
by
f 7 dN f (1)n f dM
Note that this is well-defined because (dN f )(Mi ) dN (Ni+n ) Ni+n1 and (f
dM )(Mi ) f (Mi1 ) Ni1+n . We claim that d2Hom = 0. Indeed, we have
d2Hom (f ) = dN (dN f (1)n f dM ) (1)n1 (dN f (1)n f dM ) dM
= d2N f (1)n dN f dM + (1)nN f dM + (1)2n1 f d2M
=0
Definition 2.5.5. Let A be a k-algebra, M an A-bimodule. The Hochschild cochain complex
of M is
C n (A, M ) = HomAe (B A, M )n
where M is viewed as a left Ae -module via (a b )m = amb, and we view M as a complex
concentrated in degree zero.
Explicitly, we have
C n (A, M ) = HomAe (A(n+2) , M )
= HomAe (An+2 , M )
= Homk (An , M )
A map f Homk (An , M ) is identified with : A(n+2) M , where (a0 an+1 ) =
a0 f (a1 , . . . , an )an+1 . The differential is dnHom = (1)n+1 b0 , or in terms of f : An M ,
n
X
(d f )(a1 , . . . , an+1 ) = a1 f (a2 , . . . , an+1 )+ (1)i f (a1 , . . . , ai ai+1 , . . . , an+1 )+(1)n1 f (a1 , . . . , an )an+1
n

i=1

There are many different interpretations of Hochschild cohomology we will concentrate

on extensions of algebras and deformation theory. Hochschild homology is related to de
Rham algebras, and can be used to compute the cohomology of free loop spaces. It is also
useful in studying the representation theory of preprojective algebras of graphs.
36

2.6

Centers and Derivations

Example 2.6.1 (Center). Unpacking the definition, we get C 0 (A, M ) = Homk (A0 , M ) =
Homk (k, M ) = M . The differential d0 : M C 1 (A, M ) = Homk (A, M ) sends m to the
function
d0 (m)(a) = am am
We have HH0 (A, M ) = Ker(d0 ) = {m M : am ma = 0} = Z(M ), the center of the
bimodule M .
Example 2.6.2 (Derivations). We have d1 : C 1 (A, M ) C 2 (A, M ), defined by
(d1 f )(a1 a2 ) = af (a2 ) f (a1 a2 ) + f (a1 )a2
One checks that Ker(d1 ) = Derk (A, M ) = {f Homk (A, B) : f (ab) = af (b) + f (a)b}. The
map d0 : M C 1 (A, M ) sends m to the inner derivation adm : a 7 [a, m]. Thus we have
an exact sequence
0

/ Z(M )

/M

ad

/ Derk (A, M )

/ HH1 (A, M )

/0

In other words, we have

HH1 (A, M ) = Derk (A, M )/ InnDerk (A, M ).

2.7

Extensions of algebras

Let k be a field, and let A be a k-algebra.

Definition 2.7.1. An extension of A is just a surjective k-algebra homomorphism : R  A.
Equivalently, we can write a short exact sequence
0

/M

/R

/A

/0

where M = Ker() is a two-sided ideal in R. We call the extension R  A a nilpotent

extension if M is a nilpotent ideal of degree n > 1, i.e. M n = 0 in R. An abelian extension
of A is a nilpotent extension of A of degree 2.
Lemma 2.7.2. If : R  A is an abelian extension with M = Ker(), then M is
canonically an A-bimodule.
Proof. Choose a k-linear section s : A R of . We can do this because k is a field. We
then define a map A M A M by
a m b 7 s(a)ms(b) = a m b
To see that this map is well-defined, first lets check that (a1 a2 ) m = a1 (a2 m) for
all a1 , a2 A. Indeed, we have (s(a1 a2 ) s(a1 )s(a2 )) = a1 a2 s(a1 )s(a2 ) = 0. So
s(a1 a2 ) s(a1 )s(a2 ) Ker() = M . Since M 2 = 0,
(a1 a2 ) m a1 (a2 m) = (s(a1 a2 ) s(a1)s(a2 )) m = 0
37

Finally, if s0 : A R is another section of , the fact that (s s0 ) = 0 implies s(a) s0 (a)

M for all a, whence (s(a) s0 (a)) m = 0, i.e. s(a)m = s0 (a)m.
Note that we could have defined A M M by a m 7 rm for any r with (r) = a,
without using the existence of a section. Thus, the lemma will be true if we replace k by
any commutative ring.
Reversing the logic, we fix A and M .
Definition 2.7.3. An abelian extension of A by M is an extension : R  A where
Ker() ' M as an A-bimodule.
A trivial example is M o A, which is M A as a k-vector space, and which has
multiplication
(m1 , a1 ) (m2 , a2 ) = (m1 a2 + a1 m2 , a1 a2 )
We say that two extensions E, E 0 of A by M are equivalent if there is a commutative diagram
(as in the case of group cohomology):
/M

/E


/0

/A

/0

/ E0

/M

/A

Theorem 2.7.4. There is a natural bijection



equivalence classes of abelian
2
HH (A, M ) '
extensions of A by M
Proof. Essentially as in the group case, the bijection is induced by the map C 2 (A, M )
E xt(A, M ) assigning to a 2-cochain f : A A M a k-algebra M of A of the form M A
with multiplication defined by
(m1 , a1 ) f (m2 , a2 ) = (m1 a2 + a1 m2 + f (a1 , a2 ), a1 a2 )
The key point is that the product f is associative if and only if f is a Hochschild 2-cocycle,
i.e. d2 f = 0. Moreover, two algebras M of A and M og A give equivalent extensions of A if
and only if f g is a Hochschild coboundary. (Check this!)

2.8

Crossed bimodules

Definition 2.8.1. A crossed bimodule is a DG algebra C with Cn = 0 for all n 6= 0, 1.

So as a complex, C is

/0

/ C1

/ C0

/0

Explicitly, C0 is an algebra, C1 is a bimodule over C0 , and C12 = 0. The Leibniz rule implies
that for all a C0 , b C1 , we have (ab) = a(b) and (ba) = (b)a, i.e. : C1 C0 is
38

a homomorphism of C0 -bimodules. For any b1 , b2 C1 , because b1 b2 = 0, the Leibniz rule

implies (b1 )b2 = b1 (b2 ). So we could have defined a crossed bimodule to be a bimodule C1
over C0 together with a C0 -bimodule map : C1 C0 satisfying (b1 )b2 = b1 (b2 ).
Remark 2.8.2. If C is a crossed bimodule, we can define the structure of an algebra on
C1 by b1 b2 = (b1 ) b2 , and with this structure is an algebra homomorphism.
Remark 2.8.3. Let XBimod be the category of crossed bimodules, which is a full subcategory
+
+
of DGA+
k . The inclusion functor i : XBimod DGAk has left adjoint functor : DGAk
XBimod that assignes to any DG algebra
d2

/ C1

d1

/ C0

/0]

/ coker(d2 )

d1

/ C0

/0]

/ C2

C = [

correspondent cross-bimodule defined by

(C ) = [ 0

C0 ) be a crossed bimodule, and let A = H0 (C) = Coker(),

Lemma 2.8.4. Let C = (C1
M = H1 (C) = Ker(). Then A is a k-algebra and M is canonically an A-bimodule.
We have the exact sequence
0

/M

/ C1

/ C0

/A

/0

Definition 2.8.5. A crossed extension of A by M is a crossed bimodule C with H0 (C) = A

and H1 (C) = M .
We say that two crossed extensions C , C0 are equivalent if there is an isomorphism of
DG algebras : C C0 inducing the identity on A and M .
Let X Ext(A, M ) denote the set of equivalence classes of crossed extensions of A by M .
Theorem 2.8.6. Let k be a field, A a k-algebra, and M an A-bimodule. Then there is a
natural bijection X Ext(A, M ) ' HH3 (A, M ).
Proof. We will define the map : X Ext(A, M ) HH3 (A, M ). Given an extension
E =[0

/M

/ C1

/ C0

/A

/0]

choose splittings s : A C0 and q : Im() C1 of and . Define g : A A C0

by g(a b) = q(s(ab) s(a)s(b)). Since is a morphism of algebras, g = 0 implies
s(ab) s(a)s(b) Ker(), so g is well-defined. We can define E : A3 C1 by
E (a1 a2 a3 ) = s(a1 )g(a2 g3 ) g(a1 a2 a3 ) + g(a1 a2 a3 ) g(a1 a2 )s(a3 )
Note that E = 0. Since is a bimodule map over C0 and q = 1, the image of E is
contained in Ker() = Im(i). We leave it as an exercise to show that i1 E is a Hochschild
3-cocycle whose class in HH3 (A, M ) is independent of the choice of s and q.
39

2.9

The characteristic class of a DG algebra

L
Let A = ( p0 Ap , d) be a DG algebra. Consider the graded vector spaces
C1 := Coker(d)0 [1] C0 := Ker(d)
Note that C0 is a graded subalgebra of A while C1 is a graded C0 -bimodule. The
differential d on A induces a graded map
: C1 C0

(2.2)

which makes (2.2) a graded cross-bimodule. The cokernel of is the algebra H (A), while
the kernel of is the H (A)-bimodule whose underlying (graded) vector space is H1 (A)[1].
The right multiplication on H1 (A)[1] is given by the usual multiplication in H (A), while
the left multiplication is twisted by a sign:
a
s(
x) = (1)|a| s(
ax
),
where a
H (A) is homogeneous, x
H1 (A)[1].
Definition 2.9.1. By Theorem 2.8.6 the crossed bimodule
0

/ H (A)

/ H1 (A)[1]

/0

represents an element A HH3 (H (A), H1 (A)[1]), which is called the characteristic class
of A.
This class is secondary (co)homological invariant of A. It is naturally related to Massey
triple products. In more detail, let
E := [ 0

/M

/ C1

/ C0

/B

/0]

be a crossed extension of an algebra B by M . Given a, b, c B such that ab = bc = 0 we

define Massey triple product ha, b, ci M/(aM + M c) as follows. Choose a k-linear section
s : b C0 so that s = idB , and let q : Im() C1 be a section of so that q = idIm() .
Since ab = 0 we have s(a) s(b) Ker so we can take q(s(a) s(b)) C1 . Similarly, since
bc = 0 we may define q(s(b)s(c)) C1 . Now, consider the element
{a, b, c} := s(a)q(s(b) s(c)) q(s(a) s(b))s(c) C1
Since {a, b, c} = 0 we see that {a, b, c} M . We define
ha, b, ci := {a, b, c} M/aM + M c,
where {. . . } denotes the residue class modulo aM + M c. The class ha, b, ci is independent
of the choice of sections s and q. It only depends on the class of (C , ) in HH3 (B, M ) and
the elements a, b, c B. In fact, ha, b, ci can be computed from HH3 (B, M ) by
ha, b, ci = E (a, b, c),
40

where E is the Hochschild 3-cocycle associated to the crossed extension E in the proof of
Theorem 2.8.6.
Finally, if the crossed extension E comes from a DG algebra A , i.e. A = [E] in
HH3 (H (A), H1 (A)[1]), then we recover the classical definition of triple Massey products
for homology classes a, b, c H (A) of a GD algebra (see [GM03] for details).
Remark 2.9.2. One useful application of characteristic classes of DG algebras is concerned
with realizability of modules in homology:
, we say
Given a DG algebra A with homology H (A) and a graded H (A)-module M
. Here, by DG
is realizable if there is a DG module M over A such that H (M ) ' M
that M
module we mean a graded module M over the DG algebra A endowed with a differential
dM : M M1 satisfying d2M = 0 and dM (am) = dA (a) m + (1)|a| a dM (m).
It turns out that the characteristic class A of the DG algebra A provides a single
. In particular, if A = 0, then any graded H (A)-module is
obstruction to realizability of M
realizable. For details, see [BKS03].

Deformation theory

The main reference for the main part of this section is the survey by Bertrand Keller [Kel03].

3.1

Motivation

In classical mechanics, one starts with the phase space, which is a symplectic manifold (e.g.
the cotangent bundle T X). The ring of smooth functions C (M ) has extra structure: the
Poisson bracket {, } : C (M ) C (M ) C (M ), and the Hamiltonian H C (M ).
Locally, the equations of motion (i.e. the Hamilton equations) are, for coordinates pi , qi M :
pi = {H, pi }
qi = {H, qi }
where f =

df
dt .

Example 3.1.1. Let X = Rn , M = T X = R2n , with coordinates (q1 , . . . , qn , p1 , . . . , pn ),

where we think of the qi as space coordinates and the pi as momentum coordinates. Let
F, G C (Rn ). Then the Poisson bracket is

n 
X
F G G F
{F, G} =

pi qi
pi qi
i=1

Note that {pi , pj } = 0, {qi , qi } = 0, and {pi , qj } = ij . In this context, the Hamilton
equations are
H
qi
H
qi =
pi

pi =

41

One has to choose an H; one common example represents n harmonic oscillators:


n 
X
1 2
2 2
H=
p + i qi
2 i
i=1

In quantum mechanics, we replace the Hamiltonian function H with a differential

operator. In this context, instead of C (M ) we have a non-commutative algebra and we
replace {, } with the commutator [, ].
For example, we have the Heisenberg relation [
p, q] = i~ij , where pi = i~ q i , and this
relation lives in the ring C (Rn )[ q i , . . . , qn ].
(needs more details on quantum mechanics)

3.2

Formal deformations

Let k be a commutative ring (later a field of characteristic zero). Let A be a fixed unital
associative k-algebra. Write kt = kJtK for the ring of formal power series in t, and let
At = AJtK = A k kJtK. Elements of At look like
X
u=
an tn
, an A.
n>0

Let m : A A A be the multiplication map a b 7 ab.

Definition 3.2.1. Let k be a commutative ring, A an associative unital k-algebra. A formal
kt At At such
deformation (or star product) on A is a continuous kt -linear map : At
that the following diagram commutes:
At At


t70

AA

/ At


t70

/A

denotes completed tensor product, i.e. At

kt At = lim A[t]/tn k[t] A[t]/tm .
Here,

Basically, we are thinking of At as a topological kt -algebra, and taking tensor product in the
category of topological kt -algebras. By continuity, is determined uniquely by its restriction
to AA. For all a, b A, we can write ab = ab+B1 (a, b)t+B2 (a, b)t2 + +Bn (a, b)tn + ,
where the Bi : A A A are bilinear maps and B0 = m.
Let Gt = Autkt (At ) be the group of all kt -linear automorphisms of At such that
such that : At At satisfies (a) a (mod t) for all a A. In other words, we have
(a) = a + 1 (a)t + 2 (a)t2 + where the i are k-linear.
Definition 3.2.2. Two products and 0 are equivalent if there exists Gt such that
(u v) = (u) 0 (v) for all u, v At .
Definition 3.2.3. A Poisson bracket on A is a k-bilinear map {, } : A A A such
that for al a, b, c A
42

1. {a, b} = {b, a} (skew-symmetry)

2. {{a, b}, c} + {{b, c}, a} + {{c, a}, b} = 0 (Jacobi identity)
3. {a, bc} = b{a, c} + {a, b}c (Leibniz rule)
(not necessarily
Lemma 3.2.4. Let A be a commutative k-algebra, and let be
Pan associative
n
commutative) formal deformation of (A, m). Write =
n>0 Bn t . Define {, } :
A A A by {a, b} = B1 (a, b) B1 (b, a). Then
1. {, } is a Poisson bracket on A
2. {, } depends only on the equivalence class of .
Proof. We can define h, i : At At At by hu, vi = 1t [u, v] = uvuv
(mod t). This
t
induces a bracket {, } on A, and one can check that {a, b} = B1 (a, b) b1 (b, a) is indeed
a Poisson bracket.
Suppose 0 via some , i.e. (a b) = (a) 0 (b) for all a, b A. Then
(a b) = ab + (1 (a, b) + B1 (a, b))t + O(t2 )
(a) 0 (b) = ab + (a1 (b) + 1 (a)b + B10 (a, b)) + O(t2 )
The equality (a b) = (a) 0 (b) yields B1 (a, b) + 1 (a, b) = B10 (a, b) + a1 (b) + 1 (a)b.
This implies B1 (a, b) B1 (b, a) = B10 (a, b) B10 (b, a), i.e. {, } is independent of the
equivalence class of .
Remark 3.2.5. Our definition of {, } for a noncommutative ring is a bit superfluous. It
turns out that any Poisson bracket on any (possibly noncommutative) prime ring A is just
{, } = [, ] for some Frac (Z(A)). This is a theorem of Farkas-Letzter, see [FL98].
Exercise Show that any associative star product on At is unital, and that for any , there
exists 0 such that 10 = 1A .
For any associative k-algebra A, define a map
A : {star products on A}/

/ {Poisson brackets on A}

by A () = {, } . A major question is: is A surjective?

Theorem 3.2.6 (Kontsevich 1997). Let M be a smooth manifold over k = R. Let A =
C (M ) be the ring of smooth functions on M . Then A is surjective. More precisely, for
any M , there exists a canonical (up to equivalence) section A to A .
Example 3.2.7 (Moyal-Weyl). Let M = R2 with the standard bracket
{F, G} =

F G F G

.
x y
y x
43

Then A is given by

X
n F n G tn
F t G =
xn y n n!
n=0

It is not obvious that t is associative.

If M = U Rn is open and A = C (M ), then any Poisson bracket on A is of the
following form:


X
G F
F G
ij

{F, G} =
(x1 , . . . , xn )
xi xj
xi xj
i<j

It means that there are (unique) smooth functions ij C (M ), 1 i < j n such that


X

ij
{, } =

xi xj
i<j

The functions (ij ) are actually components of a tensor field of type (2, 0) on M which is
called Poisson bivector.

3.3

Deformation theory in general

The general idea is any deformation problem is controlled by a dg Lie algebra. A bit
more precisely, suppose we hav a category A and an object A Ob A. We can define a
deformation functor Def A (A, ) : R Set, where R is some category of test commutative
algebras (or cocommutative coalgebras). For example, R could be the category of artinian
local algebras over a field (which can be of characteristic p > 0). The functor Def A (A, )
assigns to some R the set of all deformations of A parameterized by Spec(R), modulo
equivalence. We say that a dgla LA controls this deformation problem if there is a natural
isomorphism
Def A (A, R) ' MC(LA , R)
where MC(LA , R) is the set of Maurer-Cartan elements in LA R.
Assume from now on that k is a field of characteristic zero. Let A = Algk be the category
of associative unital k-algebras. Let R = Artk be the category of local Artinian k-algebras
with residue field k. That is, R Ob R if and only if R is a local commutative k-algebra
with finite-dimensional maximal ideal m R, such that R/m = k. This clearly implies
mn = 0 for all n  0. A good example is R = k[t]/(tn ).
Given R R, write AR for A k R.
Definition 3.3.1. An R-deformation of A is an associative R-linear map : AR R AR
AR such that the following diagram commutes:
AR R AR


eR e
R
m

AA
44

/ AR


eR

/A

where
eR = 1 R and R : R R/m is the canonical projection.
We say that two R-deformations and 0 are equivalence if there is an R-module
isomorphism g : AR AR such that
AR


/ AR


commutes, and such that g(u v) = g(u) 0 g(v) for all u, v AR . Note that by R-linearity,
: AA Am
is determined by its restriction to A A. Moreover, is determined by e
because = m + e
.
Definition 3.3.2. The deformation functor Def(A, ) : R Set is given by
Def(A, R) = {R-deformations of A}/equivalence.
Definition 3.3.3. If R = k[t]/t, then R-deformations of A are called infinitesimal.
Lemma 3.3.4. There is a natural bijection Def(A, k[t]/t2 ) = HH2 (A, A).
Proof. By definition, for R = k[t]/t2 , an R-deformation is determined by
: A A A[t]/t2 = A k[t]/t2
which will be of the form a b 7 ab + B1 (a, b)t, where B1 : A A A is some k-linear
map. The associativity of is equivalent to B1 being a Hochshield 2-cocycle. Indeed,
(a b) c = (ab)c + (B1 (a, b)c + B1 (ab, c))t
a (b c) = a(bc) + (aB1 (b, c) + B1 (a, bc))t
Since A is associative, the two are equal exactly when
(d2Hoch B1 )(a, b, c) = aB1 (b, c) B1 (ab, c) + B1 (a, bc) B1 (a, b)c = 0
Moreover, if , 0 are equivalent, then there is g : A A[t]/t2 such that g(a b) = g(a) 0 g(b),
which is equivalent to
B10 (a, b) B1 (a, b) = ag1 (b) g(ab) + g(a)b
Thus 0 if and only if B10 B1 = dg1 .
Remark 3.3.5. One might hope that Def Algk (A, k[t]/tn ) ' HHn (A, A). Unfortunately, this
is not true in general if n > 3. However, it is true that Def Alg (k) (A, k[t]/tn ) = HHn (A, A),
where |t| = 2n, and Alg (k) is the category of A -algebras (also called strongly homotopyassociative algebras) over k. Note that A -algebras are not in general associative.
45

Definition 3.3.6.LLet k be a field. A differential graded Lie algebra over k is a Z-graded

vector space L = nZ Ln with a bracket [, ] : L L L that is homogeneous of degree
zero (that is, [Lp , Lq ] Lp+q for all p, q) such that
1. [x, y] = (1)|x||y| [y, x] for all homogeneous x, y
2. [x, [y, z]] = [[x, y], z] + (1)|x||y| [y, [x, z]] (the Jacobi identity)
Moreover, there is a differential d : L L+1 such that d2 = 0 and d[x, y] = [dx, y] +
(1)|x| [x, dy].
L0

If L is a dgla, then L0 is an honest Lie algebra. There is a canonical representation of

(the adjoint representation) ad : L0 End(L ) given by ad(x)y = [x, y].

Definition 3.3.7. Let L be a differential graded Lie algebra. The space of Maurer-Cartan
elements in L is


1
MC(L) = x L1 : dx + [x, x] = 0
2
If L1 is finite-dimensional, then MC(L) is actually a variety (in fact an intersection of
quadrics in L1 ). For x MC(L), set


Tx MC(L) = v L0 : dv + [x, v] = 0 .
This is precisely the Zariski tangent space to MC(L). The action of L0 on L1 fixes MC(L).
If L0 is a nilpotent Lie algebra, we can define the reduced Maurer-Cartan space to be
MC(L) = MC(L)/ exp(L0 ).
We will see that there exists a dgla LA such that Def(A, R) ' MC(LA mR ).
We continue to assume that k is a field of characteristic zero. Let L = L be a dgla over
k. The Maurer-Cartan space


1
1
MC(L) = x L : dx + [x, x] = 0
2
can be regarded as a subscheme of L1 , where L1 is viewed as affine space over k. We would
like to compute the Zariski tangent space of MC(L). Recall that if X is an affine variety (or
scheme) over an algebraically closed field k, and x X is a closed point, then there is (by
the Nulstellensatz) a k-algebra homomorphism : O(X) k with Ker() = mx .
Definition 3.3.8. The Zariski tangent space of X at x is
Tx X = Der(O(X), x) = { : O(X) k : (f g) = f g + gf }
There is a canonical identification of Tx X with (mx /m2x ) . For more details, see any
good book on algebraic geometry.
46

Example 3.3.9. If X = V = Spec (Sym(V )) is a finite dimensional vector space over k,

then O(X) = k[V ], and there is a canonical isomorphism V Tx V given by


d
v 7 v,x : f 7 f (x + vt)
dt
t=0
Let f : X Y be a morphism of affine schemes. (Since X and Y are affine, it is
equivalent to give a homomorphism f : O(Y ) O(X).)
Definition 3.3.10. For y Y define the scheme-theoretic fiber of f at y by
f 1 (y) = Spec (O(X)/f (my )O(X)) .
The differential of f at x X is dfx : Tx X Tf (x) Y given by 7 f .
One can check that there is a natural isomorphism Tx f 1 (y) = Ker(dfx ).
Example 3.3.11. Let X = V , Y = W where V, W are finite dimensional vector spaces over
k viewed as affine schemes. Let f : V W be a morphism. For x V , y = f (x) W , one
can verify that there is an exact sequence
0

/ Tx f 1 (y)

/ Tx V

dfx

/ Ty W

Under the identifications Tx V = V , Ty W = W , the kernel Tx f 1 (y) is




d 
v V :  f (x + vt) = 0
dt t=0
Now we can redefine the Maurer-Cartan space MC(L), viewing it as a scheme-theoretic
fiber. From now on assume that L is locally finite-dimensional, i.e. dimk Lp < for all
p. Define f : L1 L2 by x 7 dx + 12 [x, x]. This is a morphism, and we can redefine the
Maurer-Cartan scheme to be the scheme-theoretic fiber f 1 (0). For x MC(L), we have




d
1

Tx MC(L) = v L : f (x + vt)
=0
dt
t=0
We can compute
d
(dfx )t =
dt



1
d(x + vt) + [x + vt, x + vt] 
2
t=0

= dv + [x, v].
This allows us to make the following definition even if L1 is not locally finite-dimensional.
Definition 3.3.12. The (Zariski) tangent space to MC(L) at x is
Tx MC(L) = {v L1 : dx v = 0}.
where dx = d + adx .
47

Lemma 3.3.13. For all x MC(L),

1. d2x = 0
2. The assignment L0 3 7 (x 7 dx ) Der(MC(L)) defines a Lie algebra homomorphism. In particular, dx Tx MC(L).
Proof. Part 2 follows from 1 trivially. To see that part 1 is true, we assume dx + 12 [x, x] = 0,
and compute
d2x = (d + adx )(d + adx )
= (d + adx )(d + [x, ])
= dx + d[x, ] + [x, d] + [x, [x, ]]
= [dx, ] [x, d] + [x, d] + [x, [x, ]]
1
= [[x, x], ] + [x, [x, ]]
2
=0
The last equality follows from Jacobi identity:
[x, [x, ]] = [[x, x], ] [x, [x, ]]
1
= [[x, x], ]
2

From here on out, assume that L0 is a nilpotent Lie algebra, i.e. for all L0 , the
endomorphism ad End(L0 ) is nilpotent. Moreover, we assume that the adjoint action
of L0 on L1 is nilpotent. Consider the group Aff(L1 ) of affine linear transformations of L1 .
This is just the semidirect product GL(L1 ) n L1 . Let aff(L1 ) be the Lie algebra of Aff(L1 ).
We have aff(L1 ) = gl(L1 ) n L1 , and there is an anti-homomorphism L0 aff(L1 ) given by
7 dx = d + [x, ]. Exponentiation gives an anti-homomorphism exp(L0 ) Aff(L1 ),
which yields a right action of exp(L0 ) on L1 . This action restricts to an action of exp(L0 )
on MC(L).
Definition 3.3.14. MC(L) = MC(L)/ exp(L0 )
Remark 3.3.15. For x MC(L), we can consider the orbit Ox = x exp(L0 ) MC(L). It
turns out that Tx MC(L)/Tx Ox = Ker(dx )/ Im(dx ) = H1 (L, dx ).
Let R = Art be the category of local commutative k-algebras with finite-dimensional
(hence nilpotent) maximal ideal and residue field k. For any dgla L and for any R R with
maximal ideal m, the Lie algebra L m is nilpotent.
Definition 3.3.16. MC(L, R) = MC(L mR ) and MC(L, R) = MC(L mR ).
48

One main theorem of this section is the following.

Theorem 3.3.17. 1. Let f : L1 L2 be a quasi-isomorphism of dg Lie algebras. That is,
f is a morphism of graded Lie algebras, f d1 = d2 f , and f : H (L1 ) H (L2 ) is an
isomorphism. For any R R, f induces a natural bijection.
MC(L1 , R)

/ MC(L2 , R).

2. For any x MC(L1 , R) f also induces a quasi-isomorphism

(L1 m, dx )

/ (L2 m, d).

Any homomorphism f : L1 L2 gives a natural map f : MC(L1 ) MC(L2 ). The

nontrivial part of the theorem is that f induces a bijection on reduced Maurer-Cartan
spaces.

3.4

The Gerstenhaber bracket

Let A be an associative k-algebra, where kLis a field of characteristic zero. Recall that the
p
p
Hochschild complex of A is C (A, A) =
pZ C (A, A), where C (A, A) = 0 for p < 0,
p
p
p
p+1
C (A, A) = Homk (A , A) for p > 0, and d : C C
is
p1
X
(1) (df )(a0 , . . . , ap ) = a0 f (a1 , . . . , ap )+ (1)i f (a0 , . . . , ai ai+1 , . . . , ap )+(1)p1 f (a0 , . . . , ap1 )ap
p

i=0

Definition 3.4.1. Define an insertion operation : C p C q C p+q1 by

(f g)(a1 , . . . , ap+q1 ) =

p
X

(1)i(q1) f (a1 , . . . , ai , g(ai+1 , . . . , ai+1 ), ai+q+1 , . . . , ap+q1 ).

i=0

Notice that the operation is not associative in general. We can use to define the
following bracket.
Definition 3.4.2. The G-bracket (Gerstenhaber bracket) [, ]G : C p C q C p+q1 is
defined by
[f, g]G = f g (1)|f ||g| g f .
Let LAS (A) = C (A, A)[1]. That is, LpAS = C p+1 (A, A). The Gerstenhaber bracket
induces a bracket [, ]G : LpAS LqAS Lp+q
AS of degree zero.
Lemma 3.4.3 (Gerstenhaber). (LAS (A), [, ]G , d) is a dg Lie algebra.
Proof. This can be checked directly, though checking the (super) Jacobi directly in this way
is quite tedious. However, it follows from the following.
Since is not associative we define the associator
A(f, g, h) = (f g) h f (g h)
One can check that A is (super)symmetric in g and h, i.e. A(f, g, h) = (1)(|g|1)(|h|1) A(f, h, g).
This symmetry formally implies the Jacobi identity.
49

Remark 3.4.4. The Lie algebra (LAS (A), [, ]G ) only depends on A as a vector space.
The multiplication on A enters in this picture as follows. We have m C 2 (A, A) = L1AS (A),
and it is easy to check that for all f C p (A, A), we have [m, f ] = df , where d is the
Hochschild differential. We can rewrite this as d = adm .
We define the cup-product on C (A, A) by
(f ` g)(a1 , . . . , ap+q ) = f (a1 , . . . , ap )g(ap+1 , . . . , ap+q )
where f C p , g C q . It is easy to check that d(f ` g) = df ` g + (1)|f | f ` dg.
By the lemma 3.4.3, (HH (A, A)[1], [, ]G ) is a graded Lie algebra. In fact, (HH (A, A), [, ]H , `)
is a (graded commutative) Gerstenhaber algebra.
Definition 3.4.5. A Gerstenhaber algebra G is a graded commutative algebra with product
and bracket {, } : Gp Gq Gp+q1 such that
1. (G [1], {, }) is a graded Lie algebra
2. {a b, c} = a {b, c} + (1)(|a|1)|b| {a, b} c
Example 3.4.6 (Gestenhaber algebra from differential geometry). Let M be
Vpa smooth
manifold over R, and let T M be the tangent
P bundle of M . Set p (M ) = (M,L T M ). In
coordinates, an element of p (M ) looks like 1 p . We can set (M ) = p>0 p (M ).
This is a graded commutative algebra with respect to the exterior product . The algebra
p (M ) has a Schouten bracket {, }S : p q p+q1 , defined by
{1 p , 1 , q } =

p X
q
X

(1)i+j [i , j ] 1 bi 1 bj q

i=1 j=1

This algebra ( (M ), , {, }) is a Gerstenhaber algebra.

Theorem 3.4.7 (Gerstenhaber). (HH (A, A), [, ]G , `) is a commutative Gerstenhaber
algebra.

3.5

Stasheff construction

The following is a construction by Jim Stasheff, 1993. Recall that an associative algebra is a
vector space A with a map m : A A A such that the following diagram commutes:
AAA


m1

1m
m

AA

/AA


/A

Thus it is very natural to define a dual objects to algebras to be coalgebras, defined as

follows.
50

Definition 3.5.1. A (coassociative) coalgebra over k is a vector space C with a coproduct

: C C C such that the following diagram commutes:

/C C

C C

/C C C

Recall that d : A A is a derivation if d m = m (1 d + d 1). This makes it natural

to make the following definition.
Definition 3.5.2. A linear map D : C C is called a coderivation if it satisfies the
coLeibniz rule, i.e.
D = (1 D + D 1) .
c
Starting
L with annalgebra A, one can define a coalgebra C = T (A[1]), which as a k-vector
space is n>1 A[1] . There is a natural coproduct on C, given by

(x1 , . . . , xn ) =

n1
X

(x1 , . . . , xi ) (xi+1 , . . . , xn )

i=1

where we write x1 , . . . , xn for x1 xn . It is essentially trivial that (C, ) is a coassociative

coalgebra. Note that for an algebra A the space Der(A) has a natural bracket [D1 , D2 ] =
D1 D2 D2 D1 . Let Coder(C) be the space of all graded coderivations of C. This is a
graded Lie algebra with bracket [D1 , D2 ] = D1 D2 (1)|D1 ||D2 | D2 D1 .
For a vector space V , if A = T V with embedding i : V , T V , then the natural map
i : Der(T V ) Hom(V, T V ) is an isomorphism. Similarly, let p : T c (A[1])  A[1] be the
canonical projection. Then the natural map
p : Coder (T c (A[1])) Hom (T c A[1], A[1]) ' C (A, A)[1]
is an isomorphism. One can check that the induced bracket on C (A, A)[1] is nothing but
the Gerstenhaber bracket.
The coalgebra (T c A[1], ) is called the bar construction of A. A number of complicated
constructions on algebras actually come from simple constructions on the bar construction.
Theorem 3.5.3. The dg Lie algebra LAS (A) = C (A, A)[1] controls the deformations of A.
Precisely, there is a functorial bijection for all R R:
Def(A, R) ' MC(LAS (A) mR ).
Proof. Given R R, then the corresponding star-product : AR R AR AR is determined
by R-linearity to its restriction B : A R A m, i.e. a b = ab + B(a, b). One has
B C 2 (A, A) m = L1AS (A) m. A key point is that is associative if and only if B
satisfies the Maurer-Cartan equation, i.e.
1
dB + [B, B]G = 0
2
0
0
Moreover, if and only if B and B are in the same orbit in MC(LAS (A) m).
51

3.6

Kontsevich Formality Theorem

Recall that for an algebra A, (LAS (A) = (C (A, A)[1], [, ]G ) is a dg Lie algebra that
controls deformations of A in the sense that
MC(LAS (A), R)) ' Def(A, R)
for all R R.
Definition 3.6.1. Two dg Lie algebras L and L0 are called homotopy equivalent if there is
a sequence of dg Lie algebras L1 , . . . , Ln with quasi-isomorphisms
L = L0 L1 Ln = L0
(the arrows can be in either direction).
Lemma 3.6.2 (Goldman-Milson). Let L and L0 be dg Lie algebras. The following are
equivalent:
1. L and L0 are homotopy equivalent
2. There exists L0 such that that there are quasi-isomorphisms
L0

L0

3. There is an L homomorphism L L0
We think of L0 from the Goldman-Milson lemma as being a generalized morphism
from L to L0 .
Let A = C (M ) be the ring of functions on a smooth manifold M . Recall that in
Lemma 3.2.4 we have constructed a natural map A from formal deformations of A to
Poisson brackets on A. Kontsevich proved that A is surjective. There is a dg Lie algebra
LPois (A) that controls Poisson deformations of A with the trivial bracket. Kontsevichs
theorem can be interpreted as saying that LAS (A) and LPois (A) homotopy equivalent (i.e.
there is an L -morphism LAS (A) LPois (A). The existence of this homotopy is known as
the formality theorem. To be more precise, we need the following definitions.
Definition
V2 3.6.3 (Chevally-Eilenberg (co)homology). Let g be a Lie algebra with bracket
[, ] :
g g. Let V be a g module (i.e. there is a homomorphism of Lie algebras
g End V ). The Chevally-Eilenberg complex C (g, V ) has C p (g, V ) = 0 if p < 0, and
C p (g, V ) = Homk (
52

Vp

g, V )

with differential
(1)p df (X0 , . . . , Xp+1 ) =



ci , . . . , X
cj , . . . , Xp+1
(1)i+j+1 f [xi , xj ], X1 , . . . , X

i<j

X
ci , . . . , Xp+1 )
+
(1)i Xi f (X1 , . . . , X
i

This definition is motivated by the following theorem:

Theorem 3.6.4 (H. Cartan). Let G be a compact connected Lie group with Lie algebra g.
Then H (G, R) ' H (g, R), where the action of g on R is trivial.
As before, let M be a smooth manifold and A = C (M ). Let {, } be a Poisson
bracket on A, and denote by g the Lie algebra (A, {, }). Consider g as a g-module via
the adjoint action g End g, where 7 ad : x 7 [, x]. Let L = C (g, g)[1]. There is an
analogue of the Gerstenhaber bracket on L. We set
X
f g =
sgn()f (g(X(1) , . . . , X(q) ), Xq+1 , . . . , Xp+q+1 )
sSq

Define [f, g]C = f g = (1)(|f |1)(|g|1) g f . Then (LCE , [, ]CE ) is a dg Lie algebra.
Theorem 3.6.5 (Kontsevichs formality theorem). For a smooth manifold M and A =
C (M ), gA = (A, {, } = 0), there is a homotopy equivalence
LAS (A) LCE (gA )
Its very interesting to study the space of L equivalences between LAS (A) and LCE (gA ).
d . The group
This group admits a faithful action of the Grothendieck-Teichm
uller group GT

d
d
GT contains Gal(Q/Q) as a subgroup, and Grothendieck conjecture that GT = Gal(Q/Q).
d on L -equivalences between LAS (A) and LCE (gA )
There is a conjecture that the action of GT
is simply transitive!

3.7

Deformation theory in algebraic number theory

Another source of motivation for deformation theory comes from algebraic number theory.
Let k be a field of characteristic not 2 or 3. Recall that an elliptic curve over a field k is the
subset of P2k given by a homogeneous equation
y 2 z = x3 + axz 2 + bz 3
where a, b k are such that = 16(4a3 + 27b2 ) 6= 0. Let E be an elliptic curve. One can
show using the Riemann-Roch theorem that E naturally has the structure of an abelian
variety (projective group variety) with unit (0 : 1 : 0) in projective coordinates. For each
integer N > 5, there is a smooth projective curve X0 (N ) over Q that parameterizes elliptic
53

schemes with level N structure. Complex-analytically, X0 (N ) is a compactification of

h/0 (N ), where h = {z C : <(z) > 0} is the upper half plane and



a b
0 (N ) =
SL(2, Z) : d 0 (mod N )
c d
One says that an elliptic curve E/Q is modular if there non-constant rational map X0 (N )
E for some N . Equivalently, E is modular if there is a surjection J0 (N )  E, where
J0 (N ) is the Jacobian of X0 (N ). The Taniyama-Shimura conjecture claims that all elliptic
curves over Q are modular. One can show that Fermats Last Theorem follows from the
Taniyama-Shimura conjecture. Andrew Wiles proved the Taniyama-Shimura conjecture by
creating a mod-p Galois representation and studying its deformations! More precisely, if
E is an elliptic curve with discriminant and p - , the p-adic Tate module of E is the
Zp -module
' H1 (E , Zp )
Tp E = lim E[pn ](Q)

et
Q

As a Zp -module, Tp ' Z2
p , so the action of GQ = Gal(Q/Q) on Tp gives us a representation
E,p : GQ GL(2, Zp )
On the other hand, each modular form f of level N gives rise to an ideal If End(J0 (N )),
and the quotient J0 (N )/If is an elliptic curve. We denote the representation J0 (N )/If ,p by
f,p . It is a theorem that E is modular if and only if E,p is isomorphic to f,p for some
modular form f .
For a special class of elliptic curves, it was already known that the mod p representation
E,p : GQ GL(2, Zp )  GL(2, Fp )
was modular. What Andrew Wiles did is consider the category R whose objects are finite
local Zp -algebras with residue field Fp , and define the functor DE,p : R Set by
DE,p (R) = { : GQ GL(2, R) : E,p

(mod mR )} /

Here and 0 are equivalent if they are conjugate by an element of Ker(GL(2, R)

GL(2, Fp )). It is a theorem that DE,p is representable in the sense that there exists a
profinite Zp -algebra RE,p such that
DE,p (R) ' HomZp -TopAlg (RE,p , R)
functorially in R. Wiles then considered a deformation functor DE classifying a special
class of lifts of E,p . This functor is also (pro-) representable, with representing algebra
RE . There is another deformation functor classifying modular lifts of E,p . This functor
is also representable by a ring TE,p . It was known that E,p was modular for a specific
p; this gave a homomorphism TE,p RE . Wiles proved that this homomorphism is
actually an isomorphism (i.e. that R = T ) and thus that E,p corresponds with a modular
representation GQ GL(2, Zp ), from which it follows that E is modular.
54

There are many other places in number theory where one attempts to classify deformations
of object X0 defined over Fp . One does this by defining a functor assigning to each R R
some class of lifts of X to R, and then hoping that the deformation functor is representable.
Unlike the situation of this course, where the fact that our deformation functor is representable
(by a dgla) is trivial, in the number-theoretic context it is often very difficult to show that a
given deformation problem is representable.

55

56

Chapter 3

Category theory
1

Basic category theory

1.1

Definition of categories

Definition 1.1.1. A category C consists of

a class Ob C of objects (written X, Y, Ob C)
a class Mor C of morphisms (written ( : X Y ) HomC (X, Y ))
composition maps Hom(Y, Z) Hom(X, Y ) Hom(X, Z) (written (f, g) 7 f g)
satisfying the following axions
Mor C uniquely determines X, Y with Hom(X, Y )
for all X Ob C there is a distinguished idX Mor C such that f = f idX , g = idX g
whenever defined
composition is associative
Definition 1.1.2. A category D is called a subcategory of C if Ob D Ob C, Mor D
Mor C, and composition in D agrees with that of C.
Definition 1.1.3. We say that a subcategory D of C is full if for all X, Y Ob D, we have
HomD (X, Y ) = HomC (X, Y ). We say that D is a strictly full subcategory of C if for all
Y Ob C, Y ' X for X Ob D implies Y Ob D.
Yuri Manin divides examples of categories into the following three groups.
Example 1.1.4. The first group of examples consists of categories, where objects are sets
with some additional structure, and morphisms in such categories are just morphisms of sets
which preserve this structure. Here are basic examples (they are very well-known, but we
will mention them to fix the notation):
57

 Set is the category of sets;

Top is the category of topological spaces with morphisms continuous maps between
them;
Gr is the category of groups and group homomorphisms;
Ab is the full subcategory of Gr which consists of abelian groups;
Vect is the category of vector spaces and linear maps;
Ring and ComRing are categories of rings and commutative rings respectively;
Algk and ComAlgk are categories of algebras and commutative algebras respectively
over a fixed field (sometimes a commutative ring) k;
Com(A) for any additive category A denotes the category of complexes over A (see
subsection 2.1 below for the notion of additive category).
Example 1.1.5. The second group consists of categories, where objects are still sets with
some structure, but the morphisms are not maps of the sets.
Ho(Top) is the category, where objects are topological spaces, and morphisms are
homotopy classes of continuous maps.
Rel is the category with objects just sets, but morphisms between two objects X and
Y are defined to be binary relations R X Y . The composition S R of morphisms
R : X Y and S : Y Z is defined by
S R = {(x, z) X Z | y Y, s.t. (x, y) R and (y, z) S} X Z
The identity morphism idX HomRel (X, X) is the equality relation {(x, x) | x X}.
Example 1.1.6. The third group of examples consists of classical structures that sometimes
can be (usefully) considered as categories.
Any (partially) ordered set I can be viewed as a category C(I) with Ob(C(I)) = I
and morphism sets Hom(x, y) = {x y} consisting of one arrow if x 6 y and
Hom(x, y) = otherwise.
For any topological space X we can make topology on X into a category Open(X)
of open sets with morphisms identical inclusions (see section 3.1 in Chapter 1 where
we defined presheaves on topological spaces).
A quiver is a finite directed graph Q = (Q0 , Q1 ). For example,

58

is a quiver. Any quiver Q can be thought of as a category Q, whose objects are vertices
of Q, and morphisms between vertices vi and vj are all paths from vi to vj . Here
path means a sequence of arrows f1 , . . . , fn such that f1 starts at vi and fn ends in
vj . If vi = vj = v then we also include the identity morphism idv into Hom(v, v).
There is another category associated to a quiver. Namely, denote by Qcom the category
with Ob(Qcom ) = Ob(Q). Set of morphisms HomQcom (vi , vj ) contains unique element
if HomQ (vi , vj ) is nonempty, and HomQcom (vi , vj ) = otherwise. Intuitively this
means that all paths between vi and vj define the same morphism in Qcom .

1.2

Functors and natural transformations

Definition 1.2.1. Let C and D be categories. A (covariant) functor F : C D is a map

F : Ob C Ob D and maps F : HomC (X, Y ) HomD (F (X), F (Y )) for all X, Y Ob C,
such that
F (idX ) = idF (X)
F (f g) = F (f ) F (g)
whenever the composition is defined.
There is a dual notion, namely, the notion of a contravariant functor, which is just a
functor on the opposite category C . Here we define C to be the category with Ob(C ) =
Ob(C) and
HomC (X, Y ) = HomC (Y, X),
with composition induced from C.
Example 1.2.2. A an example of a contravariant functor consider the functor () : Vect
Vect which associates to any vector space V its linear dual space V .
Also, we have seen before in sections 2.2 and 3.1 of Chapter 1 that simplicial objects in a
category and (pre)sheaves on topological spaces are also examples of contravariant functors.
Definition 1.2.3. We call a functor F : C D
faithful if the maps F : HomC (X, Y ) HomD (F X, F Y ) are injective for all X, Y
Ob C
full if the maps F : HomC (X, Y ) HomC (F X, F Y ) are surjective
fully faithful if F is both full and faithful
essentially surjective if for all Y Ob D, there exists X Ob C such that Y ' F (X)
in D
Definition 1.2.4. If C and D are categories, then their product C D is defined by Ob(C
D) = Ob C Ob D, and
HomCD ((X1 , Y1 ), (X2 , Y2 )) = HomC (X1 , X2 ) HomD (Y1 , Y2 ).
59

Definition 1.2.5. A bifunctor is just a functor F : C D E, where C, D and E are

categories. For example, the assignment (X, Y ) 7 Hom(X, Y ) is a bifunctor C C Set.
Let F, G : C D be two functors.
Definition 1.2.6. A morphism of functors : F G is given by a collection {X :
F (X) G(X)}XOb C such that whenever f : X Y in C, the following diagram commutes
F (X)


F (f )

F (Y )

/ G(X)


G(F )

/ G(Y )

Example 1.2.7. Consider functor GLn () : ComRings Gr with GLn (R) being the group
of n n invertible matrices with coefficients in the ring R. Also, consider the functor
() : ComRings Gr which associates to a ring R the subgroup R of its units. Then
taking determinant of a matrix defines a natural transformation det : GLn () () .
Example 1.2.8 (Convolution of morphims of functors). Suppose there are categories C, D
and E with functors F : C D and G, G0 : D E, along with a natural transformation
: G G0 . We write
C

/D

6E

G0

We write F : G F G0 F for the morphism of functors given by ( F )X = F (X) :

G(F (X)) G0 (F (X)) for all X Ob C. Similarly, if we have a diagram
F

6D

/E

F0

We write G : G F G F 0 for the morphism of functors determined by (G )X =

G(X ) : G(F (X)) G(F 0 (X)) for all X Ob C.
We call a category C small if Ob C is a set (as opposed to a proper class).
Definition 1.2.9. If C is a small category and D any category, we define Fun(C, D) by
Ob Fun(C, D) = functors D D
Mor Fun(C, D) = morphisms of functors
If C, D carry some extra structure (e.g. D and D are additive categories), then Fun(C, D)
is assumed to consist of functors preserving that structure
Definition 1.2.10. A morphism of functors : F G is an isomorphism (or natural
equivalence, or natural isomorphism) if there is a morphism : G F such that = idF
and = idG .
60

Lemma 1.2.11. A morphism : F G is an isomorphism of functors C D if and only

if X : F (X) G(X) is an isomorphism for all X Ob C.
Proof. If is an isomorphism, then it has an inverse : G F such that = idF and
= idG . But evaluating these equalities pointwise, we see that X X = idX and
X Y = idY for all Y Ob D, i.e. each X is an isomorphism.
1
If each X is an isomorphism, define : G X by X = X
. We need to check that
is a morphism of functors. Given f : X Y , we know that the diagram
F (X)

/ G(X)

F (f )

F (Y )

G(f )

/ G(Y )

/ F (X)

But this is easily seen to imply that

G(X)


G(f )

G(Y )

F (f )

/ F (Y )

commutes, whence is a natural transformation.

Remark 1.2.12. It is convenient to think of the category Cat of all small categories as a
strict 2-category. That is, objects of Cat are small categories, and for each C, D Ob Cat,
the hom-set HomCat (C, D) is actually the category Fun(C, D). The 2-category Cat is strict
because composition of 1-morphisms is strictly associative, i.e. (F G) H = F (G H).
In general, one requires (F G) H = F (G H) being true only up to equivalence of
functors.

1.3

Equivalences of categories

Definition 1.3.1. Let C, D be categories. We say that C is isomorphic to D if there is a

pair of functors F : C D, G : D C such that F G = idD and G F = idC .
This definition is useless in practice because finding functors that are inverses on the
nose is nearly impossible. Instead, we make the following definition.
Definition 1.3.2. A functor F : C D is an equivalence of categories if there exists a
functor G : D C such that F G ' idD and G F ' idC .
We say that C and D are equivalent if there is an equivalence F : C D. We write
C ' D if C and D are equivalent. If F is an equivalence, we call G a quasi-inverse for F .
The quasi-inverse is far from unique, but it is unique up to natural equivalence.
61

Example 1.3.3. Let D = Vectnk be the category of n-dimensional vector spaces over a field
k. Let C be the full subcategory of D consisting of a single object k n . The inclusion functor
i : C D is an equivalence of categories, even though i is not a bijection at the level of
objects. A quasi-inverse of C arises from choosing a basis for each V Ob Vectnk . To each
f : V W , we assign the matrix representation of f in terms of our chosen bases of V and
W.
This example is typical. Equivalent categories may have different objects, but the same
isomorphism classes of objects. Also, the construction of a quasi-inverse typically requires
the axiom of choice.
Recall that F is fully faithful if the maps HomC (X, Y ) HomD (F (X), F (Y )) are
bijections, and F is essentially surjective if for all Y Ob D, there exists X C for which
F (X) ' Y .
Theorem 1.3.4 (Freud). A functor F : C D is an equivalence of categories if and only
if it is fully faithful and essentially surjective.
Proof. We will prove a bit later more general result that will imply this theorem.
If C is a category, a skeleton sk C of C is a full subcategory of C with one object in each
isomorphism class. The theorem shows that sk C , C is an equivalence of categories.
Example 1.3.5 (Groupoids). A groupoid is a (small) category in which all morphisms are
isomorphisms. So a group is just a groupoid with one object. We say that a groupoid is
connected if any two objects can be connected by arrows (possibly in both directions). It
is easy to see that the skeleton of a connected groupoid is a group. The main example is
the fundamental groupoid (X) of a topological space X. Objects of (X) are points in X,
and Hom(X) (x, y) is the set of homotopy classes of paths x y. If X is connected, the
choice of a point x X gives rise to an equivalence of categories 1 (X, x) , (X).
Example 1.3.6. There are several equivalences of categories relating algebra and geometry.
For example, if k is a field, the category of finitely generated commutative reduced k
algebras is anti-equivalent to the category of affine varieties over k via the functor A 7 Spec A.
More generally, the category of all commutative k-algebras is anti-equivalent to the category
of affine schemes over k, once again via Spec. In both cases, Spec has a quasi-inverse, namely
X 7 OX (X).
The category of all associative k-algebras does not have a good geometric analogue. Of
course, one can define the category of non-commutative affine schemes to be the opposite
of the category of associative k-algebras, but this is not reasonable, as is seen by the next
section.

1.4

Representable functors and the Yoneda lemma

For a category C, an object X Ob C yields two functors. The first is contravariant,

denoted hX : C Set, defined by Y 7 HomC (Y, X). For g : Y Z, the induced
62

arrow g = hX (g) : hX (Z) hX (Y ) is s 7 s g. The covariant version of hX is

hX : C Set. On objects, it is defined by Y 7 HomC (X, Y ), and for g : Y X, the map
g = hX (g) : hX (Y ) hY (Z) is s 7 g s.
Definition 1.4.1. A functor F : C Set is representable if there exists X Ob C such
that F ' hX in Fun(C , Set). We say that F : C Set is corepresentable if there exists
X Ob C such that F ' hX .
To simplicy notation, write Cb for the category Fun(C , Set) of presheaves on C. We
b So, h is defined on objects by
want to extend X 7 hX to a (covariant) functor h : C C.
X 7 hX . For a morphism f : X1 X2 , we define hf : hX1 hX2 by
hf (Y ) : hX1 (Y ) hX2 (Y )

, g 7 f g = f g.

We need to check that hf is a morphism of functors. Take a morphism s : Z Y in C. We

need the following diagram to commute.
hX1 (Y )

hf (Y )

/ hX (Y )
2

hX1 (Z)

hf (Z)

/ hX (Z)
2

The upper path sends g : Y X1 to (f g) s, while the lower path sends g to f (g s).
The two are equal by associativity. It is easy to see that hf g = hf hg , because this is
equivalent to (f g) = f g .
b Then the map Hom b(hX , F ) F (X) given by
Theorem 1.4.2 (Yoneda). Let F Ob(C).
C
7 (X)(idX ) is a bijection.
Proof. Let y : HomCb(hX , F ) F (X) be the map 7 (X)(idX ). We show that y is a
bijection by constructing an explicit inverse. Given x F (X), we want a morphism of
functors i(x) : hX F . This would consist of morphisms i(x)(Y ) : hX (Y ) F (Y ) for
each Y . Given f hX (Y ) = Hom(Y, X), we have a map F (f ) : F (X) F (Y ). We define
i(x)(Y )f = F (f )(x). It is not difficult to show that i(x) actually is a morphism of functors.
We will show that i is an inverse to y.
First we show that i is a right inverse to y. For x F (x), i(x) is defined by i(x)(Y )f =
F (f )(x), so y(i(x)) = i(x)(X)idX = F (idX )(x) = x.
Now we show that i is a left inverse for h. Given : hX F , let x = y() = (X)(idX ).
We need i(x) = , i.e. (Z) = i(x)(Z) for all Z Ob C. This is just the claim that
for all f : Y X, we have (Z)(f ) = F (f )(x). Apply the definition of is a natural
transformation to f : Y X. We get a commutative diagram
hX (X)


(X)

hX (Y )

(Y )

63

/ F (X)


F (f )

/ F (Y )

The commutativity of this diagram is the fact that F (f )((X)s) = (X)(s f ) for any
s : X X. If we choose s = idX , then we get F (f )((X)idX ) = (X)(f ), i.e. i(x)f =
F (f )(x) = (X)(f ).
Corollary 1.4.3. Let X, Y Ob C. Then h : HomC (X, Y ) HomCb(hX , hY ) is a bijection.
Proof. Let F = hY . The Yoneda lemma says that the map Hom(hX , hY ) F (X) =
Hom(X, Y ) given by 7 (X)(idX ) is a bijection. It is easy to check that this map is the
inverse to h : Hom(X, Y ) Hom(hX , hY ), so h is a bijection.
In light of the corollary, we can use h to regard C as a full subcategory of Cb consisting
b is representable, then the representing object X is
of representable functors. If F Ob(C)
determined uniquely (up to canonical isomorphism). By this we mean the following. Suppose
b and isomorphisms : F hX , : F hY . Thus 1 : hX hY
we have F Ob(C)
is a natural isomorphism. The Yoneda lemma gives an isomorphism HomC (X, Y )
HomCb(hX , hY ) Applying h1 to 1 gives a canonical isomorphism X Y . We can

make this even more precise. If F : C Set is representable ( : hX

F for some X), one
usually says that F is represented by a pair (X, ), where = X = X (idX ) F (X). The
pair (X, ) is unique up to unique isomorphism. That is, if (X 0 , 0 ) also represents F , there
is a unique isomorphism f : X X 0 such that F (f )() = 0 . Since AutCb(F ) ' AutC (X),
the functor F does not determine X up to unique isomorphism.
There is a dual Yoneda lemma, which states that the assignment X hX is a fully
faithful contravariant functor C Fun(C, Set).
The representability of F can be redefined in terms of a universal property. The pair
(X, ) represents F if and only if for all objects Y Ob C, F (Y ), there exists a unique
f : Y X such that F (f )() = . Dually, (X, ) corepresents F if and only if for all
Y Ob C and F (Y ), there is a unique f : X Y such that F (f )() = .
Many objects in various categories can be constructed by first defining a functor (which
is expected to be representable) and then by proving that the functor is representable. The
representing object is the object one wants. This has an analogy in PDE theory. Given a
system of PDEs, one wants to find a nice (smooth, for example) solutions. A natural way
to do this is to first find a generalized solution (a distribution) and then prove that the
generalized solution is regular enough.
Example 1.4.4 (Matrix representations). Let k be a field, and fix a k-algebra A. A matrix
representation of A over a commutative k-algebra B is a k-algebra homomorphism : A
Mn (B). Given any f : B B 0 , we can define a new representation f : A Mn (B 0 ) by
f = Mn (f ) . We say that f is induced from by f . An obvious question is: is there
a universal representation? That is, we seek a commutative ring An and a representation
n : A Mn (An ) such that for any : A Mn (B), there is a unique homomorphism
f : An B such that = f n .
Define the representation functor Repn (A) : ComAlgk Set to be
Repn (A)(B) = Homk-Alg (A, Mn (B)).
64


We are asking if Repn (A) is corepresentable. The answer is yes! Define n A = (A k Mn (k))Mn (k) ,
where denotes free product. In other words,

n
A = {A A k Mn (k) : [a, m] = 0, m Mn (k)}.
Let eij Mn (k) be the elementary matrix with 1 in the (i, j)-th coordinate. Then eij form
a canonical k-basis for Mn (k). For a A k Mn (k), let
aij =

n
X

eki a ejk .

k=1

n
We claim that [aij , Mn (k)] = 0, and that
A is spanned by the aij . Let An = n A/h[ n A, n A]i

be the abelianization of the algebra n A. It turns out that Repn (A) is represented by An .
This construction is due to Bergman [Ber74].
Example 1.4.5 (Hilbert scheme). Let k be an algebraically closed field of characteristic
zero, and let X be a projective variety over k. Define the functor HilbX : Schk Set by
U 7 {Z U X closed subscheme such that U : Z U is flat}
That is, HilbX (U ) is the set of families of closed subschemes of U X parameterized by U .
One can prove that HilbX is representable. However, the representing object is a scheme,
not a variety. One remedies this by stratifying HilbX via Hilbert polynomials.
For u U , let Zu = 1 (u), and define the Hilbert polynomial of Z at u by PZ,u (m) =
(OZu OX (m)), where denotes Euler characteristic. It is a theorem that PZ,u actually
is a polynomial that is independent of u if U is connected. For some polynomial P , let
HilbPX (U ) = {Z U X : Z is a flat subscheme with PZ = P }.
It is a major theorem of Grothendieck that HilbPX is representable by a projective variety.
Example 1.4.6 (PDEs). Let n > 1, and let U Cn be an open subset. Consider a
differential operator
X
P =
a (z1 , . . . , zn )z
||6m

where = (1 , . . . , n ) ranges over multi-indices and where z = z11 znn . Supposing

the a O an (U ), we are interested in solutions to P u = 0. Denote by D = D(U ) the ring
of all linear differential operators with coefficients in O an (U ). Let D-Mod be the category of
left D-modules, and define a functor SolP : D-Mod Set by
N 7 {solutions of P u = 0 in N }.
It is easy to see that SolP is represented by MP = D/(D P ). Given u N with P u = 0,
define D N by D 7 Du. This is a D-modules homomorphism with kernel D P , and it
is easy to see that this correspondence is a bijection.
65

Given that SolP is determined by MP , we can think of differential equations as D-modules.

If M is some D-module, an N -valued solution to the differential equation determined by M
is a D-linear map M N . This suggests that we think of higher solutions as elements
of ExtiD (M, N ). Even better, we can think of M as an object in the derived category of
D-Mod.

1.5

Adjoint functors

Let F : C D be a functor. For Y Ob D, define the functor FeY : C Set by

F (T ) = HomD (F (T ), Y ) .
In other words, FeY = hY F .
Proposition 1.5.1. Suppose FeY is represented by some XY Ob C for every Y . Then
the assignment Y 7 XY extends to a functor G : D C, for which there is a natural
isomorphism of bifunctors C D Set,
HomD (F (), ) ' HomC (, G()) .
In this situation, we say that F is left adjoint to G, and G is right adjoint to F . This
situation we will denote by F : C  D : G.
Proof. For Y Ob D, choose a natural isomorphism : hXY FeY . This gives bijections
T : HomC (T, XY ) HomD (F (T ), Y ), so we can define XY = XY (idXY ). Define G : D
C on objects by G(Y ) = XY . For f : Y Ye in D, consider
e : HomC (, XYe ) HomD (F (), Ye ).
We have f XY : F (XY ) Ye , so we set
1
e
G(f ) = eG(Y
) (f G(Y ) ) : G(Y ) G(Y ).

It is tedious but straightforward to check that this construction actually makes G a functor.

Example 1.5.2. If F and G are mutual quasi-inverses in an equivalence of categories, then

(F, G) and (G, F ) are adjoint pairs.
Example 1.5.3 (Abelianization). Natural embedding i : Ab , Gr is right adjoint to the
abelianization functor ab : Ab Gr that associates to any group G the group G/[G, G]. Analogous statements are true, for example, for embeddings ComRing , Ring and ComAlgk ,
Algk .
66

Example 1.5.4 (Free objects). Consider forgetful functor for : Gr Set that associates
to each group itself, but viewed as a set (i.e. it forgets the group structure). Then
operatornamef or is right adjoint to the functor free : Set Gr that associates to each set S
free group free(S), generated by S. The same is true with forgetful functors from Ab, Ring,
ComRing, Algk et.c.
Example 1.5.5 (Matrix representations continued). Recall we had a natural isomorphism
HomComAlgk (An , B) ' HomAlgk (A, Mn (B)).

Now we realize that the functor ()n : Algk ComAlgk , A 7 An = n A/[ n A, n A], is left
adjoint to the functor Mn : ComAlgk Algk . It turns out that we have an adjoint pair

n
: Algk  Algk : Mn ().
Example 1.5.6 (Tensor-Hom adjunction). Let R and S be associative unital rings. Let
C = Mod(R), D = Mod(S) be the categories of right modules over R and S respectively. Let
B be an (R, S)-bimodule. Then we have a functor () R B : Mod(R) Mod(S) has right
adjoint HomS (B, ). That is, there is a natural isomorphism

: HomS (M R B, N )
HomR (M, HomS (B, N )).
This is easy to check. Given f : M B N , we can define the map (f )(m) = f (m ).
The map has inverse 1 ()(m b) = (m)b.
Let f : R S be any ring homomorphism. Then we have an adjoint triple of functors

(f , f , f ! ). That is, f is left adjoint to f , which is left adjoint to f ! . We could write

Mod(R)
O

f!

Mod(S)
The functor f is restriction of scalars via f . We have f (N ) = N S S, with right adjoint
f ! = HomS (S, ). The left adjoint of f is f = R S.
As a concrete example, suppose f : R  R/I is the canonical surjection, where I R is
a two-sided ideal. Then f assigns to an R-module the largest quotient killed by I, f ! sends
an R-submodule to the largest submodule which is also a submodule over S, and f sends
an R-module to its quotient by I.
Let (F, G) be an adjoint pair, and let X,Y : HomD (F (X), Y ) HomC (X, G(Y )) be
a natural isomorphism witnessing this adjunction. If we take Y = F (X), we get a map
: X GF (X) corresponding to idF (X) . The maps X = x,F (X) (idF (X) ) define a
morphism of functors : idC GF called the unit of the adjunction. Dually, if we
1
take X = G(Y ), we can define : F G idD by Y = G(Y
),Y (idG(Y ) ). One calls the
counit of the adjunction. Note that we can define the convolutions F : F F GF and
G : F GF F .
67

Lemma 1.5.7. One has (F ) (F ) = idF and (G) (G) = idG .

Lemma 1.5.8. Given functors F : C D, G : D C, if there exist morphisms : idC
GF and : F G idD satisfying the above identities, then (F, G) is an adjoint pair.
Example 1.5.9 (Traces in categories). The following is an abstract version of the Bernstein
trace. Let F : C D be a functor that has both left and right adjoints. So we have an adjoint
triple (E, F, G), where E, G : D C. Suppose we have a natural transformation : G E.
Then for any X, Y Ob C, we have the trace map tr : HomD (F (X), F (Y )) HomC (X, Y )
defined as follows. For a : F (X) F (Y ), we let tr(a) be the composite
X

/ G F (X)

F (X)

/ E F (X) E(a) / EF (Y ) Y / Y

Here : id GF and : EF id come from the adjunction (E, F, G). Setting X = Y , we

get a natural transformation End(F ) End(idC ).

1.6

Limits and colimits

Let J be a fixed category, called the index category. We assume that J is finite, or at least
small. For an arbitrary category C, we write C J for the functor category Fun(J, C). It is
helpful to think of C J as the category of diagrams of shape J in C.
For example, J could be a category with no non-identity morphisms (such categories are
called discrete). Another very useful example is J = { }. Objects of C J are called
pullback data in C. Both of these are special cases of when J is a poset, where we treat J as
a category via
(
{} if i 6 j
HomJ (i, j) =

otherwise
There is an obvious functor : C C J that sends X Ob C to the constant diagram
j 7 (X)(j) = X, with (X)(i j) = idX . Given a morphism : X Y in C, define
() : (X) (Y ) by letting (j : (X)(j) = X (Y )(j) = Y be itself. We call
the diagonal (or constant) functor. It is natural to ask whether has a left or right adjoint.
Fix F : J C (i.e. F Ob(C J )) and define Fe : C Set by Y 7 HomC J ((Y ), F ).
Definition 1.6.1. If Fe is representable, we call the representing object X the limit of F ,
written X = lim F .

Our definition requires that there be a natural isomorphism

HomC J ((Y ), F ) ' HomC Y, lim F .

Any such natural isomorphism comes from a morphism s : (lim F ) F . We can make the

requirement that lim F represent Y 7 Hom((Y ), F ) much more concrete.

For any Y Ob C, a natural transformation t : (Y ) F should be thought of as a

cone over F with vertex Y . Given a morphism Y 0 Y , we can pull back a cone with
68

vertex Y to get a come with vertex Y 0 . It is natural to ask if there is a terminal cone, i.e. a
cone that induces all cones by pullback. This happens precisely when F has a limit. That
is, X = lim F with s : (X) F is the closest (to C) cone over F .

Example 1.6.2 (Pullbacks). Let J be the category { }. Limits of diagrams of

shape J are fibered products. The limit of a diagram X Z Y is called the fiber product
of X and Y over Z, and is denoted X Z Y .
If it happens that for each F Ob(C J ), Fe is representable by lim F , then F 7 lim F

can be extended to a functor lim : C J C. The functor lim is the right adjoint of , i.e.

there is an adjoint pair:

/ C J : lim
:Co

We call a category C complete if all limits of shape J exist in C for all small categories J.
Dually, we can define colimits, which (if they exist) are left adjoint to . That is, there
is an adjoint pair
/C:
lim : C J o

The terminology is slightly confusing. A limit is often called a projective (or inverse)
limit, especially if the index category is a poset and F is contravariant. Dually, colimits are
called inductive (or direct) limits, especially if J is a poset and F is covariant. The notation
colim F is sometimes used instead of lim F .

Example 1.6.3 (Initial and terminal objects). If the index category J = is the empty
category, then there exists unique functor F : J C. The colimit and limit of F are initial
and terminal objects in C respectively.
Example 1.6.4 (Products and coproducts). A (small) category J is called discrete if for
any X, Y Ob(C), HomC (X, Y ) = if X =
6 Y and HomC (X, X) = {idX }. Schematically we
have J = { . . . }. Then a functor
F
:
J C is just a collection Q
of objects {Xj }jJ in C.
`
Then lim F is just the coproduct
Xj and lim F is the product
Xj .

jJ
jJ
Example 1.6.5 (Pushouts). If J = { }, then functors F : J C are just
diagrams X Y Z, also known as pushout data. The colimit of F corresponding to
X Y Z is the pushout of X and Z over Y , denoted X tY Z.
Let I be a small category. If we have a functor F : I C, we think of F as a diagram
of shape I. For i I, write Xi for F (i). We defined the limit of F , denoted lim F , to

be a object X of C together with morphisms i : X Xi , such that for all f : i j, the

following diagram commutes:
X

/ Xi

fj

F (f )

Xj
69

and such that (X, ) is terminal with respect to this property. If G : C D is any functor,
we can apply G to (X, ) to get computable morphisms G(i ) : G(lim F ) G(Xi ), and

thus a canonical morphism F : G(lim F ) lim(G F ).

Definition 1.6.6. The functor G preserves limits if F is an isomorphism whenever lim F

exists.
Dually, given F : I C, the colimit X = lim F has morphisms : Xi X. Applying

G : C D, we get compatible morphisms G() : G(X) G(Xi ), hence a canonical

morphism F : lim(G F ) G(lim F ).

Definition 1.6.7. The functor G preserves colimits if F is an isomorphism whenever

lim F exists.

Theorem 1.6.8. For any object X in a category C, the functor hX = HomC (X, ) preserves
limits and the functor hX = HomC (, X) maps colimits to limits.
Proof. The fact that hX preserves limits is easy to check. Just construct an inverse to F .
Moreover, the fact that hX maps colimits to limits follows trivially from the fact that hX is
limit-preserving. Treat hX : C Set as a covariant functor on the opposite category C . In

fact, hX = hX , so this follows from the fact that hX preserves limits.

Corollary 1.6.9. If a functor F : C Set is corepresentable, then F preserves limits.
Dually, if F : C Set is representable, then F sends colimits to limits.
These are necessary conditions for (co)representability. Unfortunately, they are not
sufficient conditions. In general, it is very difficult to prove that a given functor is
(co)representable.
Corollary 1.6.10. Let F : C  D : G be an adjoint pair of functors. Then F preserves
colimits and G preserves limits.
Proof. This is a standard trick, using Theorem 1.6.8 and the Yoneda Lemma. Consider
H : I C, and arbitrary Y Ob(D). Then
HomD (F (lim H), Y ) ' HomC (lim H, G(Y ))

' lim HomC (H(), G(Y ))

' lim (Hom(, G(Y )) H)

' lim (HomD (F (), Y ) H)

' lim (HomD (, Y ) (F H))

' HomD (lim(F H), Y )

By the Yoneda lemma, we conclude that F (lim H) ' lim(F H).

70

adjointness
by Theorem 1.6.8

by Theorem 1.6.8.

Special topics in category theory

2.1

Brief introduction to additive categories

Example 2.1.1. Let Ab be the category of abelian groups. For two groups A, B, the set
HomAb (A, B) is not only a set it naturally has the structure of an abelian group in which
composition is bilinear.
Definition 2.1.2. A category A is additive if
AB1 For all X, Y Ob A, the set HomA (X, Y ) has the structure of an abelian group, and
the composition maps
HomA (Y, Z) HomA (X, Y ) HomA (X, Z)
are bilinear.
AB2 There is a (unique) object 0 for which Hom(0, X) = Hom(X, 0) = 0 for all X Ob C
AB3 Binary products and coproducts exist (and coincide) in A
If a category C satisfies only the Axiom AB1, we say that C is a preadditive category. We
also call preadditive categories Z-categories, thinking of them as categories enriched over Z.
Axiom 3 means that for any X1 , X2 Ob A, there is an object Y with morphisms
Xo

i1
p1

/ Y o i2 / X
2
p2

such that pa ia = idXa , pa ib = 0 if a 6= b, and i1 p1 + i2 p2 = idY . This is equivalent to the

existence of squares
p1

/ X1

/ X1

p2

X2

/0

X2

i2

i1

/Y

the first of which is cartesian and the second of which is cocartesian.

Example 2.1.3. A prototypical example of a pre-additive category is a ring R. One
considers R as a category with a one object , and sets Hom(, ) = (R, +). The composition
on Hom(, ) is induced by the multiplication on R.
Let A and B be pre-additive categories (also called Z-categories).
Definition 2.1.4. A functor F : A B is additive if for all A, B Ob A, the map
F : HomA (A, B) HomB (F (A), F (B)) is a homomorphism of abelian groups.
From now on, we will tacitly assume that Fun(A, B) is the category of additive functors
from A to B.
71

Example 2.1.5. If R is a Z-category with one object , then R = (Hom(, ), id ) is an

associative ring with unit. An additive functor F : R Ab is just a left module over R.
Indeed, set M = F (), and for R = Hom(, ) and m M , set m = F ()(m). The
definition of a functor fores the action of R on M to be additive and associative. A right
module over R is just a functor R Ab.
The moral is that additive functors on Z-categories with values in Ab should be thought
of as representations.
Example 2.1.6 (Quivers and their representations). If Q is the category corresponding to
a quiver Q (see Example 1.1.6 above), then a representation of Q is just a functor Q Ab.
e have the same objects as Q, but let
We can modify Q to get an additive category. Let Q
HomQe (vi , vj ) be the free abelian group on the set of all paths from vi to vj . The category
e Ab. More explicitly,
of representations of Q is just the category of additive functors Q
a representation F of quiver Q assigns to each vertex of Q an abelian group, and to each
arrow a Z-linear map between abelian groups. So the category of representations of a
quiver Q consists of all diagrams of abelian groups and their morphisms of fixed shape Q.
Similarly we can modify the category Qcom from Example 1.1.6 to get an additive
bcom . Then the category of representations of Q
bcom is just the category of additive
category Q
bcom to Ab. Elements of this category are commutative diagrams of abelian
functors from Q
groups of fixed shape Q.
For a nice introduction to the theory of representations of quivers, see for example
[CB92].
Example 2.1.7. For any additive category A functors F : A will form an additive
category. For example, functors F : Ab are called simplicial modules.
Example 2.1.8. Recall that abelian presheaves on a topological space X are just functors
F : Open(X) Ab. In our context, we should think of a presheaf as being a representation
of the underlying topological space.

2.2

Center of a category and Bernstein trace

Definition 2.2.1. Let A be an additive category. The center of A is defined by

Z(A) = EndFun(A) (idA )
where we define Fun(A) = Fun(A, A).
Proposition 2.2.2. For any ring R, we have Z(R-Mod) ' Z(R), where Z(R) = {x R |
[x, r] = 0, r R} is the usual center of R.
Proof. Suppose Z(R-Mod). It defines a morphism M : M M for every R-module M .
In particular, we have a morphism R : R R. Any morphism of R-modules : R M
is defined by a single element of R, namely, by (1) = m M . Indeed, then we have
72

(r) = (r 1) = r (1) = rm. Moreover, for any element m M there exist unique
morphism R M with 1 7 m. Hence, morphism R is completely defined by R (1) R.
By the definition of natural transformation, for any : R R we have the following
commutative diagram
R
R

/R
R

/R

If (1) = s R, then commutativity of the diagram implies rs = sr. Since this is true for
all s R, r = R (1) is an element of the center Z(R).
We want to show that is completely defined by R . Take any R-module M and any
m M . We want to prove that M (m) M is completely defined by R (1). Again, from
the definition of natural transformation, we have commutative diagram
R
R

: 17m

/M
M

 : 17m 
/

From this diagram we have M (m) = M ((1)) = (R (1)) = rm. This proves the claim.
The opposite direction is easy. Whenever we have an element r Z(R) we can define
r : M M , r (m) = rm. Since
a natural transformation r : idR-Mod idR-Mod by M
M
r Z(R), M will be indeed a morphism of R-modules since (sm) = rsm = srm = s(m).
Any diagram
N
N

/M


/M

will be commutative since for all n N we have N (n) = (rn) = r(n) = M (n).
Example 2.2.3 (Bernstein trace formula). Let V be a finite-dimensional vector space over
a field k, a Endk (V ). We define the trace trV (a) k ' Homk (k, k) to be the composite
k

/ Endk (V )

/ V V a1 / V V h,i / k.

Here End V V V is the inverse of the canonical map vf 7 [x 7 f (x)v]. It is not hard
to check that this agrees with the usual definition of the trace. Let M be any k-vector space
(not necessarily finite-dimensional). Define a linear map trV : Endk (M V ) Endk (M ) by
letting trV (a) be the composite
M

/ M End V

/M V V

73

aidV

/ M V V idM h,i / M .

Lemma 2.2.4. If M is finite-dimensional, the composite trM trV : End(M V ) k is

the usual trace TrM V on the vector space M V .
We can apply this generalized trace to representation theory. Let g be a semisimple
complex Lie algebra, e.g. g = sln (C). Let A = g-Mod be the category of (not necessarily
finite-dimensional) representations of g. The category A is an additive tensor category. That
is, given two g-modules M and V , then M k V is naturally a g-module via x(m v) =
(xm) v + m xv. For fixed V , we define the functor FV : A A by FV (M ) = M V .
For f : M N , the induced map FV (f ) : M V N V is f idV : m v 7 f (m) v.
Lemma 2.2.5. If V is a finite-dimensional g-module, then trV : End(FV (M )) End(M )
is functorial in M , in the sense that it induces a linear map
trV : EndFun(A) (FV ) EndFun(A) (idA ) = Z(A).
Proof. We define trV by the following rule. Given any g-module M , and any a EndFun (FV ),
set
trV (a) = {trV (aM ) : M Ob A} .
We need to check that aM Mor A implies trV (aM ) Mor A. For any f : M N , we need
the following diagram to commute
M

trV (aM )

/M

 trV (aN ) 
/

But this is easy to check.

It is an open problem to compute EndFun(A) (FV ) in general.
Problem Compute EndFun(A) (FV ) in the case where g = sl2 (C) and V = Vn the irreducible (n + 1)-dimensional g-module.
We get around the computation of EndFun(A) (FV ) by constructing a map Z(A)
EndFun(A) (FV ). Send a morphism of functors : idA idA to the convolution FV ,
where ( FV )M = FV (M ) .
Lemma 2.2.6. Let C, D and E be categories with functors

/D

Then ( ) F = ( F ) ( F ).
74

 / 
EE


Proof. We compute directly:

(( ) F )X = ( )F (X) = F (X) F (X) = ( F )X ( F )X = (( F ) ( F ))X .

Since g-Mod ' U(g)-Mod, we have EndFun(A) (idA ) ' Z(U(g)) By the Poincare-BirkhoffWitt theorem, we can write U(g) = U(h) (n+ U(g) + U(g)n ), so we have a canonical
projection : U(g)  U(h) = C[h ]. Here h is a Cartan subalgebra of g. Denote by W the
corresponding Weyl group.PThen W acts on h via the dot action, i.e. (w, ) 7 w =
w( + ) , where = 12 R+ is one-half the sum of the positive roots.
Theorem 2.2.7 (Chevalley). The map |Z(g) : Z(U(g)) C[h ]W is a ring isomorphism.
For the next theorem, we need to set up some notation. We set P (V ) = { h : V =
6 0},

where V = {v V : xv = (x)v, x h}. The convolution of an element f C[h ] with

P (V ) h is
X
(P (V ) f )(x) =
f (x + ).
P (V )

Finally, the discriminant of W is the (skew-symmetric) polynomial

Y
= (x) =
h, xi.
R+

Theorem 2.2.8. The composite map

TrV : C[h ] ' Z(U(g)) ' EndFun(A) (idA )

FV

/ EndFun(A) (FV ) trV / EndFun(A) (idA ) ' Z(U(g)) ' C[h ]

is given by the formula

TrV (f ) =

P (V ) (f )
.

(3.1)

Proof. Take a dominant integral weight h . This is, = c1 1 + +cl l is a non-negative

integral linear combination of fundamental weights i h . Fundamental weights are defined
to be the basis of h dual to the basis of simple roots = {1 , . . . , l } in the sense that
2h , i
hi , j i := hji,jji = ij . Suppose also that is regular, which means that h + , i =
6 0
for all roots R. In other words, is regular if its stabilizer in the Weyl group is trivial.
For such a weight define V to be the irreducible (finite dimensional) g-module with
highest weight . It is known that the the action of an element z Z(U(g)) on V is
given by the multiplication by f (), where f C[h ]W is the function on h corresponding
to z via Harish-Chandra isomorphism. Also, dim(V ) = c (), where c is some constant
(see [FH91][Cor.24.6]). This implies that Tr(z|V ) = c (f )().
So to prove the formula (3.1) it is enough to show that the functions on both sides
coincide for integral dominant regular . Choose such a . Lets compute the trace of
the operator TrV (z) on V . From one hand, it equals to c ( TrV (f ))(). On the other
75

hand,
from Lemma 2.2.4 it equals Tr(z|V V ). It is well-known that V V is isomorphic to
L
V+ , where sum is taken over weights P (V ) with multiplicities. Hence

Tr(z|V V ) =

Tr(z|V+ ) = c

f ( + ) = c (P (V ) (f ))().

This proves the formula (3.1).

2.3

Morita theory

For an associative unital ring A, Mod(A) denotes the category of all right (unital) A-modules,
and A-Mod denotes the category of all left A-modules.
Definition 2.3.1. Let A and B be (possibly noncommutative) rings. We call A and B
M

Morita-equivalent (denoting A B) if the categories Mod(A) and Mod(B) are equivalent.

Remark 2.3.2. It turns out that Mod(A) ' Mod(B) if and only if A-Mod ' B-Mod, so
there are no separate notions of left Morita equivalence and right Morita equivalence.
Recall that a right A-module M is projective if it is the direct summand of a free module,
i.e. there exists an A-module N such that M N ' AI as A-modules, where I is a
(possibly infinite) index set.
Remark 2.3.3. It is obvious that any free module is projective. In general this is not true.
For example, take A = k[x, y, z]/(x2 + y 2 + z 2 = 1) to be the coordinate ring of a 2-sphere.
Consider : A3 A given by (a, b, c) = ax + by + cz. Since (x, y, z) = 1, is onto, and so
we have a splitting A3 = Ker() A. Module P = Ker() is by definition projective, and
one can prove that P is not free, for example see [LS75][p.334].
Though, it turns out that for polynomial rings notions of projective modules and free
modules coincide. The following theorem was originally known as Serres problem.
Theorem 2.3.4 (Quillen-Suslin). If k is a field, then projective modules over k[x1 , . . . , xn ]
are free.
More geometrically, the module P in the example above is not free meaning that that
the tangent bundle on the sphere S 2 is not trivial. Moreover, the Quillen-Suslin theorem
states that every algebraic vector bundle on Ank is trivial. Serres conjecture was proven by
Quillen and Suslin in 1976. Quillens proof was much more intuitive. In the noncommutative
setting, is it an open problem to classify projective modules over the ring of differential
operators on Ank .
Definition 2.3.5. We say that M generates the category Mod(A) if Hom(M, ) is faithful,
i.e. f1 , f2 : K L are equal if and only if f1 g = f2 g for all g : M K. In this case we
also call M a generator for the category Mod(A).
76

An easy example of a generator is A itself, because Hom(A, ) ' idMod(A) .

For a right A-module M , define M = HomA (M, A). The dual M is naturally a left
A-module via (a f )(x) = af (x). In fact, M is an (A, EndA M )-bimodule, just as M is an
(EndA M, A)-bimodule.
Finally, let M M = Im(M A M EndA M ) via the map M M EndA M that
sends m to the map x 7 m(x). Similarly, let M M = Im(M EndA M M A), where
m 7 (m). Note that M M is a two-sided ideal in EndA M , and M M is a two-sided
ideal in A.
Theorem 2.3.6 (Dual basis). Let A be a ring, M an A-module.
1. M is projective if and only if there exists mi M , i M such that for all m M ,
i (m) = 0 for all but finitely many i, and one has
X
m=
mi i (m).
i

2. M is finitely generated and projective if and only if M M = EndA M .

If M is finitely generated
P and projective, one can choose a finite collection of mi M ,
i M such that m =
mi i (m) for all m M . One calls {mi } and {i } dual bases,
even though {mi } may not be a basis of M .
As an exercise, prove the dual basis theorem, and show that M is a generator if and
only if M M = A. There is a kind of duality here. The module M is finitely generated
projective if and only if it is a direct summand of some An , while M is a generator if and
only if A is a direct summand of some M n .
Definition 2.3.7. A right module M is a progenerator if M is a finitely generated projective
generator in Mod(A).
Theorem 2.3.8. Let A and B be rings. The following are equivalent.
1. Mod(A) ' Mod(B)
2. A-Mod ' B-Mod
3. there exists a progenerator M in Mod(A) such that B ' EndA M
Example 2.3.9. Let A be a ring, B = Mn (A) for some n. Since Mn (A) ' EndA (An ), the
theorem shows that A and Mn (A) are Morita equivalent.
Example 2.3.10. Let X be the affine line over a field k of characteristic zero. Let
d
d
A = D(X) = khx, dx
: [ dx
, x] = 1i be the ring of differential operators on X. Let
d
2
M = (x dx 1)A + X A; this is an ideal in A, and is in fact a progenerator for the category
of A-modules. It turns out (Musson, 1991) that EndD(X) (M ) = D(Y ), where Y is the zero
set of y 2 x3 on the affine plane.
As an exercise, show that if A and B are commutative rings, then A and B are Morita
equivalent if and only if they are isomorphic.
77

2.4

Recollement (gluing) of abelian sheaves

Let X be a topological space, Z X a closed subspace, and let U = X \ Z be its open

complement. So we have a closed embedding i : Z , X and an open embedding j : U , X.
We would like to decompose the category of abelian sheaves Sh(X) on X using the
categories Sh(Z) and Sh(U ) of abelian sheaves on Z and U . The adjunctions we will obtain
are a part of Grothendiecks yoga of the six functors, and fit into a diagram:
i

Sh(Z)

j!


/ Sh(X)
[

/ Sh(U )

i!

in which (i , i , i! ) and (j! , j , j ) are adjoint triples.

It is easy to define the functors i , i , j , j because they make sense for any map
between topological spaces. Indeed, if f : X Y is continuous, recall that we defined
f : Sh(X) Sh(Y ) by
(f F )(U ) = F (f 1 (U )).
Formally, one can define f by requiring it to be the left adjoint of f . To show that f
exists, one constructs it in one of several ways. The most common is to let f F be the
sheaf associated to the presheaf
U 7

lim F (V ).

V f (U )
open

Alternatively, we can define f F directly by

(f F )(U ) = {s : U Et(F ) : s(x) Ff (x) for all x U }.
Lets return to the setting where Z X is closed and U = X \Z. The functor j! is extension
by 0, i.e.
(
0
if V 6 U
(j! F )(V ) =
F (V ) if V U
The functor i! : Sh(X) Sh(Z) is restriction with compact support.
If R is a commutative ring I is an ideal in R. Let S = R/I, and write i : Spec(S) ,
Spec(R) for the induced embedding. Then (i , i , i! ) are precisely the functors defined in
the context of a closed embedding i : Z X of topological spaces.
As an exercise, check the following properties of the six functors.
1. (i , i , i! ) and (j! , j , j ) are adjoint triples
2. the unit idSh(U ) j j! and counit j j idSh(U ) are isomorphisms
78

3. the unit idSh(Z) i! i and counit i i idSh(Z) are isomorphisms

4. i! j = 0 and i j! = 0.
5. there are canonical (pointwise) exact sequences

j! j

/ id

/ i i

/ i i!

/ id

/ j j

/0

Later on, we will extend these sequences to exact triangles in the derived category.
Example 2.4.1 (projective plane). Let X = P2C be the projective plane over C. Write
X = Proj C[x, y, z], and let Z = {z = 0} = Proj C[X, Y ] be the line at infinity. The
injection i : Z , X is induced by the graded homomorphism C[x, y, z] C[x, y] that sends
f for some graded C[x, y, z]-module
z to 0. If F is a quasicoherent sheaf on X, then F = M
M = M/z.
] The sequence M i i M 0 is obviously exact,
M . It turns out that i F = ig
so sheafifying we get the exact sequence F i i F 0.
Definition 2.4.2. An additive category A is said to be a recollement of A0 and A00 if there
exist six functors
i

AW 0

j!


/A
W

/ A00

i!

satisfying properties 15 above.

It is an easy consquence of the definitions that i : A0 A is an embedding with image
{A : j A = 0}. A good reference for all of this is [BBD82]. An application to number theory
(the Arvin-Verdier duality theorem for Galois cohomology) can be found in [Maz73].

2.5

Kan extensions

Consider two functors F : C E, K : C D. We are interested in finding F making the

following diagram commute:
C
K

/E
?

D
A typical example is as follows. Let S Mor C and T Mor E be classes of morphisms. We
can construct categories C[S 1 ] and E[T 1 ], called the localizations of C and E at S and T .
79

We have a localization functors `S : C C[S 1 ] and `T : E E[T 1 ]. Given F : C E, can

we extend F to a functor F : C[S 1 ] E[T 1 ]?
Our extension problem is obviously not possible in general. For instance, there might
exist morphisms and in C such that F () 6= F (), but K() = K(). Similarly,
there might exist objects X and Y such that HomC (X, Y ) = and HomE (F X, F Y ) = ,
but HomD (KX, KY ) 6= . We will content ourselves with asking for a best possible
approximation of an extension of F to D. One can do this by looking at universal natural
transformations from F (left Kan extensions), or to F (right Kan extensions).
Definition 2.5.1. A left Kan extension of F along K is a functor LanK (F ) : D E together
with : F LanK (F ) K which is universal among all pairs (G : D E, : F G K),
in the sense that for any such pair there exists a unique : LanK F G such that there is
a commutative diagram

/GK
F
O

LanK F K
A left Kan extension of F : C E along K : C D represents the functor E C (F, K) :
Fun(D, E) Set that sends G : D E to the set HomFun(C,D) (F, G K). In other words,
there is a natural isomorphism
HomFun(D,E) (LanK F, G) = HomFun(C,E) (F, G K).
Let Cat be the category of all categories. A functor K : C D is just a morphism in
Cat. It induces, for any E, a functor K : E D E C given by G 7 G K. The functor
LanK : E C E D is the left adjoint to K .
Dually, we can define right Kan extensions.
Definition 2.5.2. The right Kan extension of F along K, written RanK F , is a functor
RanK F : D E with a natural transformation : RanK (F ) K F which is universal
among all pairs (G : D E, : G K F ), in the sense that for any such pair there exists
a unique : G RanK F such that the following diagram commutes:
GK

/ RanK F K

' 

In other words, there is a natural isomorphism

HomFun(D,E) (G, RanK F ) ' HomFun(C,E) (G K, F )
Assuming left and right Kan extensions exist, we have an adjoint triple (LanK , K , RanK ).
80

Lemma 2.5.3. If K is fully faithful and LanK F exists, then : F LanK (F ) K is a

natural isomorphism.
Example 2.5.4. In this example, denotes the simplicial category, whose objects are the
integers 1, 2, . . . , and whose morphisms f : n m are nondecreasing functions [n] [m].
Define F to be the realization functor Top that assigns to n the n-th simplex
n
o
X
n = (10 , . . . , xn ) Rn+1 :
xi = 1, xi > 0 .
Let K = Y : Set be the Yoneda embedding from into the category of simplicial
sets. Write n for the image of n under Y . We have a diagram:

/ Top
;

||

Set
It is a good exercise to check that LanY F = ||. By Lemma 2.5.3, the natural transformation
F | | Y is an isomorphism, i.e. n = |n |.

81

82

Chapter 4

Classical homological algebra

1
1.1

Abelian categories
Additive categories

Recall a category A is called a pre-additive (or Z-) category if it satisfies the axiom AB1
below. We say that A is additive if A is preadditive, and also satisfies AB2 and AB3.
[AB1] each HomA (X, Y ) is given the structure of an abelian group in such a way that
compositions
HomA (X, Y ) HomA (Y, Z) HomA (X, Z)
are bilinear;
[AB2] A has initial object , terminal object with = ;
[AB3] A has finite products.
We call the zero object, denoted 0. (Note that we can define A = lim( A), where

is the empty category and A is the unique functor.)

Let I be a set, and consider I as a category with no (non-identity) morphisms. Then a
diagram of shape I in A is just a collection {Ai : i I} of objects in A. We set
Y

Xi = lim{I A}

(product)

Ai = lim{I A}

(coproduct)

iI

a
iI

The simplest case is when I = {0, 1}. One obtains products X Y and coproducts X t Y .
Lemma 1.1.1. Let A be an additive category. If X Y exists, then so does X t Y and
X Y ' X t Y (canonically).
83

Proof. Recall that X Y represents the functor A Set defined by

Z 7 Hom(Z, X) Hom(Z, Y ).
That is, there is a natural isomorphism : Hom(Z, X Y ) Hom(Z, X) Hom(Z, Y ),
that maps : Z X Y to the pair (pX , pY ), where pX : X Y X and
pY : X Y Y correspond to idXY .
Define iX : X X Y and iY : Y X Y by iX = 1 (idX , 0) and iY = 1 (0, idY ).
It is easy to check that
pX iX = idX
pY iY = idY
pX iY = 0
pY iX = 0
These identities formally imply iX pX + iY pY = idXY . For, if we call the left-hand map
, we get pX = pX and pY = pY . But idXY also satisfies this, so uniqueness gives
= idXY .
Given iX and iY , we can define an isomorphism
Hom(X, Z) Hom(Y, Z) Hom(X Y, Z)
by (, ) 7 ( pX , pY ), with inverse 7 ( iX , iY ). Since X t Y corepresents
Hom(X, Z) Hom(Y, Z), we have X Y ' X t Y .
Exercise Let A be an additive category. Show that finite products and finite coproducts
exist, and coincide. On the other hand, show that infinite products and coproducts need
not be the same. For any X, Y A, we can define the diagonal X : X X X by
X = idX idX , and the folding map Y : Y t Y Y by Y = idY t idY . (In fact, these
exist in any category with products and coproducts.) Show that the abelian group structure
on HomA (X, Y ) is given by
f + g = Y (f g) X .
This exercise has a very important consequence: being additive is not extra structure
on A, but an intrinsic property of A! An arbitrary category A is additive if and only if
A has finite products and coproducts which coincide. (Note that the condition A ' A
is a consequence of requiring finite products and coproducts to coincide, since is empty
coproduct and is the empty product.)
Exercise Find a categorical definition of f for any f : X Y . Also, show that if A is
additive, then A is additive, and that A B is additive whenever A and B are. Moreover,
if A, B, C are categories, show that there is an equivalence of categories
Fun(A B, C) ' Fun(A, Fun(B, C)).
84

If A, B, C are additive, show this equivalence restricts to an equivalence

Funadd (A B, C) ' Funadd (A, Funadd (B, C)).
Recall that a functor F : A B between additive categories is additive if F :
HomA (X, Y ) HomB (F X, F Y ) is a homomorphism of abelian groups for each X, Y Ob A.
If R, S are rings and B is an (R, S)-bimodule, then the functor M 7 B R M is an additive
functor. On the other hand, if R is a commutative ring, the functor M 7 M 2 on R-modules
is not additive.
Exercise Show that if F : A B is additive, then the canonical map F (X Y )
F (X) F (Y ) is an isomorphism.
Exercise Show that if A is additive, then the functors of points hX : A Set are
actually functors hX : A Ab, and similarly for hX = Hom(X, ). Show that the Yoneda
embedding h : A Fun(A , Ab) is additive.
Lemma 1.1.2. If F : A B is an additive functor and G is an adjoint to F , then G is
additive. If G is right adjoint to F , then the isomorphisms
Hom(F X, Y ) ' Hom(X, GY )
are isomorphisms of abelian groups, and similarly if G is left adjoint to F .
Proof. Let : Hom(F X, Y ) Hom(X, GY ) be the isomorphism 7 Y F (), where
: id GF is the counit of the adjunction. The map 7 F () is additive because F is.
Moreover, 7 Y is additive because B is additive. It follows that is a homomorphism.
From the equivalence
Funadd (A B, C) ' Funadd (A, Funadd (B, C)),
it follows that G must be additive.

1.2

Non-additive bimodules

Let R be an associative unital ring, and let R-Mod be the category of left R-modules. Let
R-Bimod be the category of bimodules over R. Both R-Mod and R-Bimod are abelian
categories. Define F(R) to be the full subcategory of R-Mod consisting of objects isomorphic
to Rn for some n N. Better yet, we could consider P(R), the category of finitely generated
projective R-modules.
Definition 1.2.1. A non-additive bimodule is just a functor T : F(R) R-Mod.
Set F (R) = Fun(F(R), R-Mod). For example, the inclusion F(R) , R-Mod is a nonadditive bimodule. There is a canonical functor : R-Bimod F (R) defined by the rule
M 7 (M : Rn 7 M n ). Note that F (R) is an abelian category.
85

Theorem 1.2.2. The functor is fully faithful, with essential image the full subcategory
of F (R) consisting of additive functors F(R) R-Mod.
Proof. We give a construction showing that the image of consists of additive functors. If
T F (R) is additive, define M = T (R). By definition, M is a left R-module. The right
R-module structure on M is defined by a ring homomorphism : R EndR (M ). Note
that R = EndR (R), where R is treated as a left R-module. If we identify R with EndR (R),
then the fact that T is a functor yields a homomorphism
EndR (R) EndR (T (R)) = EndR (M )
hence M is an R-bimodule.
Remark 1.2.3. It turns out that Hochschild (co)homology can be extended to the category
of non-additive bimodules, yielding topological Hochschild (co)homology. Details can be
found in [BL04].
Theorem 1.2.4. The functor : R-Bimod F (R) has both a left and right adjoint.
Proof. Write for the left adjoint, and ! for the right adjoint of . We will construct
and ! directly. Given X Ob(R-Mod), define six natural morphisms i : X X X,
di : X X X for i {0, 1, 2}. We have
0 (x) = (0, x)
1 (x) = (x, x)
2 (x) = (0, x).
Assume X Ob(F(R)). Then i and di are elements of Mor(F(R)). Define, for any
T Ob(F (R)),
X (T ) = T ( 0 ) T ( 1 ) + T ( 2 ) : T (X) T (X X)
dX (T ) = T (d0 ) T (d1 ) + T (d2 ) : T (X X) X
It is easy to check that for any T ,
X 7 Ker( X (T ))
X 7 Coker(dX (T ))
are additive functors. Tate X = R, viewed as a left R-module. We define
! (T ) = Ker( R (T ) : T (R) T (R) T (R))
(T ) = Coker(dR (T ) : T (R) T (R) T (R)).
We claim that this definition actually gives left and right adjoints to . That is, for any
M R-Mod and T Ob(F (R)), we have isomorphisms
HomR-Bimod ( (T ), M ) ' HomF (R) (T, M R ) = HomF (R) (T, M )
HomR-Bimod (M, ! (T )) ' HomF (R) (M R , T ) = HomF (R) ((M ), T ).

86

Exercise (Bernstein trace) Recall that if we have a functor F : A B that sits in an

adjoint triple (E, F, G), then any natural transformation : G E, we have a canonical
trace tr() : HomB (F X, F Y ) HomA (X, Y ). Take a R and b
a EndR (R) given by
x 7 x a. Let a : ! be a natural transformation, where a (T ) is defined by the
composite
! (T ) 

/ T (R) T (ba) / T (R)

/ / (T )

The exercise is: compute tr(a) := tr(a ).

There is a notion of polynomial approximations of non-additive bimodules, based
on the notion of a polynomial maps between abelian groups. The basic idea goes back to
Eilenberg and MacLane. Given a (set-theoretic) map f : A B between abelian groups,
the defect of f is
(a1 | a2 )f = f (a1 + a2 ) f (a1 ) f (a2 ).
The map f is additive if and only if ( | )f : A A B is the zero mapping. Inductively,
we define the n-th defect of f as
(a1 | a2 | | an )f = (a1 | | an1 + an )f (a1 | | an1 ) (a1 | | b
an1 | an )f .
We say that f is polynomial of degree 6 n if the n-th defect of f is identically zero.
Definition 1.2.5. A (non-additive) functor T : A B between additive categories is
polynomial of degree 6 n if for all X, Y Ob A,
T : HomA (X, Y ) HomB (T X, T Y )
is polynomial of degree 6 n.
Going back to our original example, we can define Fn (R) to be the full subcategory
of F (R) consisting of polynomial functors of degree 6 n. There are canonical injections
Fn (R) , F (R), and F0 (R) ' R-Bimod.
Example 1.2.6 (Theorem of the cube). For a scheme X, the Picard group of X is the set
of isomorphism classes of invertible sheaves, with group operation induced by the tensor

product. One can prove that Pic(X) ' H1 (X, OX

). Let k be a field, and let AbVark be
the category of abelian varieties over k. The theorem of the cube states that the Picard
functor Pic : AbVark Ab is quadratic. There is a way of defining quadratic functors for
any pointed category, and in that generality, the theorem of the cube says that the Picard
functor is quadratic on the whole category of pointed projective varieties.

1.3

Abelian categories

Recall that additive categories are categorized by some basic axioms (see 1.1).
Definition 1.3.1. An additive category A is abelian if it satisfies an extra axiom AB4. We
say A is a Grothendieck category if in addition it satisfies AB5 (see below).
87

Before we can state the extra axioms, we need to define kernels and cokernels in arbitrary
additive categories. Let A be an additive category, : X Y a morphism in A. Consider
the functor Ker() : A Ab defined by
Z 7 KerAb ( : HomA (Z, X) HomA (Z, Y )).
If Ker() is representable, then its representing object is called the kernel of denoted by
Ker(). If Ker() exists, we have a (by definition) an exact sequence
0

/ HomA (Z, Ker )

/ HomA (Z, X)

/ HomA (Z, Y ).

A similar definition for Coker() does not work (the obvious analog of the functor above
is not representable). If we take Z = K, then (idK ) is a morphism k : K X. Hence
Ker() is represented by the pair (K, k : K X). The kernel has a much easier definition.
Let I be the category { }. A diagram of shape I is just a diagram
X0

0
1

// X .
1

A cone over this diagram is essentially a diagram

Y

/ X0

0
1

// X .
1

such that 0 f = 1 f . The equalizer of the diagram is the limit of the corresponding functor
I A. It is a good exercise to check that our definition of the kernel is equivalent to letting
Ker() be the equalizer of the diagram

// Y .

A naive definition of the cokernel would be to look at the functor Z 7 Coker(hX (Z)
hX (Z)). But this does not agree with the classical definition for abelian groups. In fact, this
functor is not usually representable. Indeed, suppose this functor is represented by some C.
Then we would have exact sequences
Hom(Z, X)

/ Hom(Z, Y )

/ Hom(Z, C)

/0

for all Z Ob A. These sequences are not all exact even if A = Ab. For example, let
X = Y = Z, and let be multiplication by n. If C = Z/n, then the sequence is
Hom(Z, X)

/ Hom(Z/n, Z)

/ Hom(Z/n, Z)

/0

which is not exact. (Strictly speaking, this only shows that C = Z/n does not work.)
So let : X Y be a morphism. The correct definition is the following.
88

Definition 1.3.2. Coker() is the representing object (if it exists) of the functor Coker() :
A Ab defined by

Z 7 Ker(Hom(Y, Z) Hom(X, Z)).

If K 0 represents Coker(), we have a canonical morphism c : Y K 0 . The pair (K 0 , c)
has a universal property: if : Y Z is such that = 0, then there is a unique
: K 0 Z such that = c, as in the diagram

/
c

/Z
O


K0

Alternatively, Coker() is the coequalizer of

// Y .

The axiom defining an abelian category is due to MacLane and Grothendieck. We say
an additive category A is abelian if it satisfies
[AB4] Every : X Y can be decomposed in the following way:
K

/X

/I

/Y

k0

/ K0

where
1. = j i
2. (K, k) = Ker() and (K 0 , k 0 ) = Coker()
3. (I, i) = Coker(k) and (I, j) = Ker(k 0 ).
Lets see what this axiom requires in the case A = Ab. Let : A B be a homomorphism
of abelian groups. Then we can decompose as
Ker() 

 k /
X

/ / X/ Ker() = I ' Im()  

/Y

k0

/ / Coker()

So essentially, axiom AB4 requires that the first isomorphism theorem holds in A.
It was Grothendiecks insight that AB4 is equivalent to the combination of the following
two axioms:
[AB4.1] Ker() and Coker() exist for all ;
[AB4.2] If : X Y is such that Ker() = Coker() = 0, then is an isomorphism.
89

Traditionally, if Ker() = 0 we say that is mono in A, and if Coker() = 0 we say that

is epi. Axiom AB4.2 just says that maps which are both mono and epi are isomorphisms.
Suppose we have AB4.1. Then for all : X Y , we can let (K, k) = Ker() ad
(I, i) = Coker(). Similarly, we can let (K 0 , k 0 ) = Coker() and (I 0 , i0 ) = Ker(k 0 ). In
addition, the definition kernels and cokernels imply the existence of a canonical morphism
` : I I 0 such that = j ` i. One can check this by using the universal property of kernels
and cokernels. Moreover, Ker(`) = Coker(`) = 0. If AB4.2 holds, ` is an isomorphism, so
the decomposition required by AB4 exists.
To see that AB4 implies AB4.2, note that Coker(0 X) = idX and Ker(Y 0) = idY .
The axiom furnishes a canonical isomorphism Coker(0 X) Ker(Y 0), hence the fact
that is an isomorphism.
Example 1.3.3 (Abelian categories). Categories of abelian groups, modules over a ring,
quasicoherent sheaves on a scheme. If A is any abelian category and C is a small category,
then Fun(C, A) is an abelian category (we will prove this fact below).
Example 1.3.4. The category Proj(R) of finitely generated projective modules over a ring
R is not generally abelian (it doesnt contain kernels and cokernels). Similarly, if X is
a topological space, we can consider the category Vect(X) of vector bundles on X. This
category is not usually abelian, for exactly the same reason.
The categories Proj(R) and Vect(X) are basic examples of exact categories (a notion due
to Quillen). Every exact category is a full subcategory of an abelian category.
Example 1.3.5 (Filtered abelian groups). Let AbF be the category of filtered abelian groups.
Objects of AbF are abelian groups X equipped with an increasing filtration F X :
F i X F i+1 X . A morphism in AbF from (X, F X) to (Y, F Y ) is a homomorphism
f : X Y such that f (F i X) F i Y for all i. Kernels and cokernels in AbF are defined as
in AbF, with the following filtrations:
F i Ker() = Ker() F i X
F i Coker() = F i Y /F i Y Im()
It is easy to check that part 1 of the axiom AB4 holds, but part 2 does not (i.e. mono + epi
does not imply iso). As an example, choose some abelian group X that admits filtrations
F1 ( F2 , i.e. F1i X F2i X for all i, but F1i X ( F2i X for some i. The map idX : X X is
mono and epi, but is not an isomorphism. In general, in the factorization
K

/X

/I

/ I0

/Y

k0

/ K0

the map ` may not be an isomorphism on each component of the filtration.

Example 1.3.6. The category of topological abelian groups is also exact but not abelian.
The reason is that epimorphisms are maps f : X Y s.t. f (X) Y is dense. But in
general density does not imply surjectivity. As an example, consider dense winding of the
real 2-dimensional torus f : R T2 , f (x) = (eix , eix ) for irrational .
90

Theorem 1.3.7. If C is a small category, A any abelian category, then Fun(C, A) is abelian.
Proof. Let F, G : C A be functors, and let : F G be a natural transformation.
We define Ker() as a pair (K : C A, k : K F ) directly. For X Ob(C), set
K(X) = Ker((X) : F (X) G(X)). On morphisms, given f : X Y , consider the
following commutative diagram:
K(X)


kX

/ F (X) X / G(X)

K(f )

K(Y )

kY

F (f )

G(f )

/ F (Y ) Y / G(Y )

The composite Y F (f )kX : K(X) G(Y ) is equal to Y G(f )X kX = 0. By the universal

property of kernels, there is a unique morphism K(f ) : K(X) K(Y ) making the leftmost
square commute. It follows that k = {kX : K(X) F (X)} is a natural transformation
K F.
It is easy to check that (K, k) is the kernel of . Similarly, we define the cokernel
pointwise. In the canonical decomposition, we have pointwise isomorphisms
I(X) = Coker(Ker X )

`X

/ I 0 (X) = Ker(Coker X )

It follows that ` is a natural isomorphism.

It follows that the category F (R) of nonadditive bimodules is abelian.
Remark 1.3.8. Similarly, one can prove the following useful observation.
Theorem 1.3.9. If A is an abelian category, then the category Com(A) of complexes in A
is also an abelian category.

1.4

Complexes in abelian categories

In what follows, we will deal mostly with abstract abelian categories. How are we to think
of these? The basic idea is that any general statement involving only finitely many objects
and morphisms is true in any abelian category, if and only if it is true in a module category.
This is justified by the following theorem:
Theorem 1.4.1 (Mitchell). If A is an abelian category, there is an associative unital ring
R and a fully faithful exact functor F : A Mod(R).
Remark 1.4.2. In other words, Theorem 1.4.1 says that every abelian category can be
thought of as being a full exact subcategory of some module category. Though, we prefer
not to think about abelian categories this way. For example, the category Qcoh(X) of
quasi-coherent sheaves on a projective scheme is abelian, but there is no obvious way to
embed it into Mod(R) for some R. Moreover, this is not the way one usually thinks about
91

sheaves. Nevertheless, Mitchells theorem can be rather useful in proving facts about general
abelian categories, because viewing objects of abelian categories as modules allows to pick
elements.
Let A be an abelian category. We can define (co)chain complexes and (co)homology
in A, just as in module categories. It is not obvious that the notion of cohomology makes
sense. Suppose we have a chain complex (C , d ), and look at a piece:
Coker(dn )
O

bn+1

Cn

dn

an

/ C n+1
O
$

%
/ C n+2

dn+1

Ker(dn+1 )
The arrows an , bn are uniquely determined by the universal properties of Ker(dn+1 ) and
Coker(dn ). There is a canonical morphism Coker(an ) Ker(bn+1 ), which is an isomorphism
by axiom AB4. Thus we can define Hn (C) = Coker(an ) ' Ker(bn+1 ).
Thus for any abelian category A, we can define the category of complexes in A, written
Com(A), and it is an easy exercise to show that Com(A) is an abelian category.
Definition 1.4.3. For all n Z, the n-th cohomology is the functor Hn : Com(A) A
defined as above.
We can define homotopies between morphisms in Com(A) the same way we did that
earlier for complexes of abelian groups. Homotopic morphisms induce the same morphism
on cohomology.
Remark 1.4.4. There is much more general notion of a homotopy between two morphisms
in a category. Namely, a notion of a homotopy in model categories. We will discuss model
categories later.
Let A , B , C be objects in Com(A). It is easy to check that a sequence
0

/ A

/ B

/C

/0

is exact if and only if it is point-wise exact. Let Exc(A) be the category of short exact
sequences in Com(A). Objects of Exc(A) are short exact sequences as above, and morphisms
are commutative diagrams:
0

/ A
1


/ A
2

f2

/ B
1


/ B
2

92

/ C

g2

/0

/ C
2

Define, for each n Z, two functors from Exc(A) to A by

F n (A B C ) = Hn (C )

F n ( , , ) = Hn ( )

Gn (A B C ) = Hn+1 (C )

Gn ( , , ) = Hn ( )

Definition 1.4.5. The connecting morphism is a natural transformation n : F n Gn

defined as follows. The morphism n : (A B C ) : Hn (C ) Hn+1 (C ) will be
defined using the following diagram:
..
.

..
.


/ An

fn


/ Bn

dn
A

..
.

/ Cn

gn

dn
B

/0

dn
C

n+1
n+1
/ B n+1 g
/ C n+1
/ An+1 f

dn+1
A

/ An+2

f n+2

dn+1
B

/ B n+2

g n+2

dn+1
C

/ C n+2

..
.

/0
/0

..
.

..
.

We use the Mitchell embedding theorem and work as through everything were modules. Choose
c C n such that dcC (n) = 0 (i.e. [c] Hn (C )). Since g n is epi, there exists bn B n
such that g n (b) = c. Note that g n+1 (dnB b) = dnC g n (b) = 0, so dn b Ker g n+1 , which is the
image of f n+1 . It follows that there exists a An+1 such that f n+1 (a) = b. We claim that
n+1 b) = dn+1 (b) = 0, whence dn+1 a = 0 since f n+2
da = 0. Indeed, f n+2 (dn+1 a) = dn+1
B (f
n
is injective. Set (c) = a.
Theorem 1.4.6. For any short exact sequence of complexes
0

/ A

/ B

/ C

/0

()

the sequence

/ Hn (A )

Hn (f )

/ Hn (B )

Hn (g)

/ Hn (C )

n (f ,g )

/ Hn+1 (A )

()

is exact.
The long exact sequence () is functorial in (). So ()7() is a functor Exc(A)
Com(A). Theorem 1.4.6 has a number of useful consequences.
93

Lemma 1.4.7 (Snake lemma). Consider a diagram

0

/ X1

/Y1

f1

/ X2

f2

g1

/ Z1

/Y2

/0

g2

/ Z2

/0

Assume the rows are exact. Then there are two exact sequences with a connecting snake
(natural in the sequences):
0

Ker

Ker

Ker

Coker

Coker

Coker

Proof. Add zeros and think of the vertical sequences as complexes as in:
0


/ X1

f1

/ X2

f2


/Y1


/ Z1

/0

/Y2

g1

g2

/ Z2

/0

We can apply Theorem 1.4.6 to the exact sequence 0 X Y Z 0, obtaining the

result.

Lemma 1.4.8 (5-lemma). Suppose we have a diagram

X1


f1

Y1

/ X2


f2

/ Y2

/ X3


f3

/ Y3

/ X4


f4

/ Y4

/ X5


f5

/ Y5

Assume the rows are exact, f1 is epi, f5 is mono, and that f2 and f4 are isomorphisms.
Then f5 is an isomorphism.
94

Lemma 1.4.9 (3 3 lemma). Consider a commutative diagram

0


/ X1


/ Y1


/ Z1

/0


/ X2


/ Y2


/ Z2

/0


/ X3


/ Y3


/ Z3

/0

If the columns and middle row are exact, then if either the first or last row is exact, so is
the other.
Proof. This is also an easy consequence of Theorem 1.4.6.

1.5

Exact functors

Let A, A0 be abelian categories, F : A A0 an additive functor.

Definition 1.5.1. We say F is left exact if for any short exact sequence in A:
/X

/Y

/Z

/ 0,

the following sequence is exact in A0 .

0

/ FX

/ FY

/ FZ

We say F is right exact if the analogous sequence (with 0 on the right) is exact. We say F
is exact if it is both left and right exact.
As motivation, consider the classical Riemann-Roch problem. Let X be a topological
space, and let F be an interesting sheaf on X. One is usually interested in the global
sections of F . Often the sheaf F can be decomposed via a short exact sequence:
0

/ F1

/F

/ F2

/0

If the functor F 7 (X, F ) = F (X) were exact, we would have a short exact sequence of
abelian groups:
/ F1 (X)
/ F (X)
/ F2 (X)
/0
0
In the classical setting, F is a sheaf of complex vector spaces, and we are only interested
in dimC (X, F ). The short exact sequence of global sections would give dim (X, F ) =
dim (X, F1 ) + dim (X, F2 ). Unfortunately, the functor (X, ) is almost never exact.
95

Example 1.5.2 (hom-functors). If A is an abelian category, the Yoneda functors hX and

hX factor through the category of abelian groups. That is, for X Ob A, we have functors
hX = HomA (X, ) : A Ab
hX = HomA (, X) : A Ab
The fact that these functors are left exact is a direct consequence of our definition of the
kernel.
Example 1.5.3 (Global sections). If X is a topological space, then the functor : Sh(X)
Ab is left exact. Though it is not at all obvious, this is a special case of the previous example.
We will work out the details later.
Example 1.5.4 (Representation theory). If G is a group, we define a functor H0 (G, ) :
G-Mod Ab by assigning to a G-module M the group M G = {m M : gm = m for all g
G}. Similarly, if g is a Lie algebra over a field k, we set H0 (g, M ) = M g = {m M : gx = 0}.
More subtle is the Zuckermann functor. Let h g be a Lie subalgebra (usually the Cartan
subalgebra of a semisimple Lie algebra). We say a g-module M is h-finite if dimk U(h)m < .
We define the functor ()h : g-Mod g-Mod by M 7 Mh = {m M : m is h-finite}.

1.6

Adjointness and exactness

Theorem 1.6.1. If F : A B is an additive functor between abelian categories. Then

If F has a right adjoint, F is right exact.
If F has a left adjoint, F is left exact.
f

Proof. Let F : A  B : G be an adjoint pair. Let 0 X 0

X
X 00 0 be an exact
sequence in A. Take Y Ob(B) and apply Hom(, G(Y )) to this exact sequence. By the
definitions of kernels and cokernels, we get a commutative diagram with exact first row:
0

/ Hom(X 00 , G(Y ))


/ Hom(X, G(Y ))

/ Hom(F X 00 , Y )

/ Hom(X 0 , G(Y ))

/ Hom(F X, Y )

(F f )

/ Hom(F X 0 , Y )

The exactness of the second row, F X 00 represents the functor

(F f )

Y 7 Ker(Hom(F X, Y ) Hom(F X 0 , Y ))
hence (F X 00 , F g) ' Coker(F f ). By definition, this means the sequence
F X0

/ FX

/ F X 00

is exact. The proof when F is a right adjoint is similar.

96

/0

In other words, left adjoints are right exact, and right adjoints are left exact.
Corollary 1.6.2. If F : A B is an additive functor between abelian functors that has
both left and right adjoints, then F is exact.
The converse of this theorem is not true there are functors that are exact but have no
adjoints.
Example 1.6.3. Let X be a topological space, U X an open subset. Let Sh(X) be the
category of abelian sheaves on X, and consider the functor (U, ) : Sh(X) Ab given by
(U, F ) = F (U ). We claim that (U, ) is exact. Indeed, let PSh(X) be the category of
presheaves on X. Recall that the forgetful functor i : Sh(X) , PSh(X) has a left adjoint
denoted ()+ : F 7 F + , were
P + (U ) = {s : U Et(P ) : s = idU }
Here Et(P ) is the total space of P and : Et(P )  X is the canonical projection.
Next, define ZU Ob(PSh(X)) by
(
0 if V U =
ZU (V ) =
Z if V U 6=
The restriction maps are the obvious ones. For any presheaf P on X, we have P (U ) =
HomPSh(X) (ZU , P ). It follows that (U, ) == HomPSh(X) (ZU , ) i, so (U, ) is the
composite of two left-exact functors, hence (U, ) is left-exact.
Example 1.6.4. Let X be a topological space, Z X a closed subspace, and let U = X \ Z
be the complement of Z. Write i : Z , X and j : U , X for the canonical inclusions.
Recall that there is a diagram:
i

Sh(Z)

j!


/ Sh(X)
[

/ Sh(U )

i!

where (i , i = i! , i! ) and (j! , j ! = j , j ) are adjoint triples. It follows that i! and j are left
exact, i and j! are right exact, and i , j are exact.
Example 1.6.5. Let R, S be rings, A = Mod(R), B = Mod(S). Let B be a (R, S)-bimodule.
Then we have an adjoint pair:
R B : Mod(R)  Mod(S) : HomS (B, ).
so R B s right exact. The converse is also true.
97

Theorem 1.6.6 (Watt). Let R, S be rings, and let f : Mod(R) Mod(S) be an additive
functor that is right exact and commutes with direct sums. Then f (R) has the structure of
an (R, S)-bimodule, and there is a natural isomorphism f ' R f (R).
0
Suppose C, C 0 are
}I
` categories with coproducts and F : C C is a functor. For {X`
Ob C, recall that X = limI X. This coproduct`comes with maps i : X X .

Applying F , we get morphisms F (i ) : F (X ) F ( I X ). By the universal properties

of coproducts, we get a canonical morphism
`

F (X )

F (i )

/F

`
We say that F commutes with coproducts if F (i ) is an isomorphism in C 0 for all collections
{X }I . (Clearly this construction works for arbitrary colimits.)
Proof. Write B = f (R). By definition, B is a right S-module. For each x R, define
x : R R, a 7 x a. The map gives us a ring homomorphism
: R HomMod(R) (R, R).
Since f is additive, the following composite is also a ring homomorphism:
R

/ Hom(R, R)

/ HomS (B, B).

Thus B is an (R, S)-bimodule. Explicitly, for x R and b B, we put x b = f (x )(b).

We need to construct a natural transformation t : R B f . For a right R-module
M , we define tM : M R B f (M ) as follows. First, define for each m M the map
m : R M of right R-modules by x 7 m x. This gives f (m ) HomS (B, f M ). We
define
tm (m b) = f (m )(b).
If well-defined, it is easy to see that this is S-linear. We need to check that tM (m xb) =
tM (mx b) for all x R. A simple computation suffices:
f (m )(x b) = f (m )(f (x )(b))
= (f (m ) f (x ))(b)
= f (m x )(b)
= f (mx )(b)
A similar routine computation shows that t : R B f is natural.
Note that everything so far works for every additive functor (no exactness properties
required). To show that t is a natural isomorphism, we need the stated hypotheses. It is
sufficient to show that tM is an isomorphism for each M in three steps:
1. Take M = R. Then tR : R R B f (R) is an isomorphism by the definition of B.
98

L
2. Take any free R-module F = I R = RI , where I is a (possibly infinite) index
set. Since f commutes with direct sums,
f

tF : F R B ' B I

/ f (B)I ' f (B I ).

3. Finally, take any right R-module M . Take a presentation of M :

/M

/ F0

F1

/0

where F0 and F1 are free. Then in the following commutative diagram, the top and bottom
rows are exact (by the right-exactness of f and R B), and the latter two vertical arrows
are isomorphisms.
/ F0 R B

F1 R B


tF1

tF0

/ f (F0 )

f (F1 )

/ M R B


/0

tM

/ f (M )

/0

By the 5-lemma, tM is an isomorphism.

2
2.1

Finiteness conditions
AB5 categories

We would like to define a class of abelian categories which are sufficiently large to have
arbitrary direct sums, but still satisfy some finiteness properties.
Definition 2.1.1 (AB5). We say that an abelian category A satisfies AB5 if
AB5 A has exact (filtered) colimits.
This axiom merits some explanation. Recall that a direct system is just a diagram
Y
Z
indexed by a category that is actually a poset. Let {Xi , X
ij }iI , {Yi , ij }iI and {Zi , ij }iI
be three direct systems, and suppose we have compatible exact sequences
/ Yi

/ Yi

/ Zi

/0

Then AB5 requires the the following sequences to be exact:

0

/ lim Xi

/ lim Yi

/ lim Zi

/0

Theorem 2.1.2. For any ring R, the category Mod(R) satisfies AB5.
99

Example 2.1.3. The axioms for an abelian category are self-dual, so for any ring R, the
category Mod(R) is abelian. However, the category Mod(R) is almost never AB5 because
colimits in Mod(R) are just limits in Mod(R), and these are not necessarily exact.
Indeed, let R be a commutative local domain of dimension one with maximal ideal
m R. Consider the exact sequence of inverse systems:
0

...


/ m3


/ m2


/m


/ R/m3


/ R/m2


/ R/m

i.e. for each n we have 0 mn R R/mn 0. Taking limits, we get 0 0 R

b 0, which is not exact unless R is already complete. Exactness on the left comes from
R
T
lim mn = mn and

Theorem 2.1.4 (Krulls intersection theorem). If R is commutative local Noetherian ring

T
with (unique) maximal ideal m, then
mn = 0.
n=1

Exercise [Mittag-Leffler condition] Let A be an abelian category. Suppose (Xi , ij ),

(Yi , ij ) and (Zi , ij ) are inverse systems indexed by a directed poset I. Suppose there are
compatible exact sequences
/ Xi

fi

/ Yi

gi

/ Zi

/0

Prove that the sequence

0

/ lim Xi

/ lim Yi

/ lim Zi

/0

provided (Xi , ij ) satisfies the Mittag-Leffler condition: for every n I, there exists n0 > n
such that for all i, j > n0 , we have Im(in ) = Im(jn ).
If an abelian category A does not satisfy AB5, then some pathologies may occur. An
object generated by simple subobjects may not be a direct sum of its simple subobjects (i.e.
it may not be semisimple).
100

2.2

Grothendieck categories

Definition 2.2.1. We say an object X in A is Noetherian if every increasing sequence

X1 X2 Xn X
is eventually stationary. The category A is Noetherian if every object in A is Noetherian.
Example 2.2.2. If A = Mod(R) and X Ob(A), then X is Noetherian as an object of A
if and only if X is Noetherian as an R-module. The category Mod(R) is never Noetherian,
except in the trivial case R = 0. This is the case even if R is (right) Noetherian. Similarly, if
X is a Noetherian scheme, the category Qcoh(X) of quasi-coherent sheaves is not Noetherian,
but its full subcategory coh(X) of coherent sheaves is Noetherian.
Definition 2.2.3. A category C has a set of generators if there exists a (small) set {Xi }iI
Ob(C) satisfying the following property. For every f, g : Y Z in C with f =
6 g, there exists
i I and h : Xi Y such that f h 6= g h.
Lemma 2.2.4. Assume A is an abelian category with arbitrary direct sums. Then the
following are equivalent:
1. A has a set {Xi }iI of generators
L
2. X = i Xi is a generator for A
3. every object of A is quotient of X J for some J
Proof. This is a good exercise.
Definition 2.2.5. An abelian category A is called locally Noetherian if it has a set of
Noetherian generators.
Definition 2.2.6. An abelian category A is called Grothendieck category if A satisfies AB5
and A is locally Noetherian.
Remark 2.2.7. Sometimes in the literature the last condition of the definition of a
Grothendieck category is weakened. Namely, sometimes only the existence of a set of
generators is required, without assuming these generators are Noetherian.
Theorem 2.2.8. Let R be a unital ring. Then Mod(R) is locally Noetherian if and only if
R is right Noetherian.
L
Proof. Let S = {Xi }iI be a set of Noetherian generators for Mod(R). Let X = i Xi . By
the lemma, X is a generator for Mod(R). Take any proper right ideal J R, and consider
the projection p : R  R/J. The map p =
6 0 since J 6= R. Thus there exists : X PR such
(X),
that p is not zero as a map X R/J. In particular, (X) 6 J. Thus R =
i.e. R is equal to the ideal generated by the {Im() : HomR (X, R)}. It follows that
R is L
a quotient of some large direct sum of Noetherian objects. In fact, R is a quotient
of X HomR (X,R) . But then 1R must be contained in the image of a finite direct sum of
Noetherian modules, so R is actually a quotient of a finite sum of Noetherian R-modules.
Thus R itself is Noetherian.
The other implication is trivial.
101

2.3

Inductive closure of an abelian category

Let R be a right Noetherian ring. Let mod(R) Mod(R) be the full subcategory of right
Noetherian modules. Similarly, if X is a Noetherian scheme, we have the category coh(X)
as a subcategory of Qcoh(X). It is natural to ask if X and coh(X) determine each other
in a categorical way. Similarly, one could ask if R and mod(R) determine each other. The
answer to this involves the inductive closure of a category.
Definition 2.3.1. A coindex category J is a small nonempty category such that
1. J is connected
2. For all j 0 i j, there is a k with a commutative diagram:
i

/j


/k

j0

3. for pair of arrows u, v : i j there exists w : j k such that wu = wv.

We would like freely add colimits of coindex categories in a category C. Recall that
C embeds in Cb = Fun(C , Set) via the Yoneda embedding. Given a diagram F : J C, we
define Fb : J Cb by Fb(j)(Y ) = Hom(Y, F (j)), i.e. Fb = hF () .
Definition 2.3.2. The inductive closure of C is the full subcategory Ce of Cb consisting of all
possible L = limJ Fb for F : J C with J a coindex category.

Theorem 2.3.3 (Gabriel). Let A be a Noetherian abelian category. Then Ae is a Grothendieck

category. Moreover, A and Ae determine each other up to natural equivalence.
^ and Mod(R) = mod(R).
^ The main reason
Our main examples are Qcoh(X) = coh(X)
e
we are interested in A is because it has enough injectives. We will explain meaning of this
later in 3.1.2.
We would like to have a good characterization of module categories among Grothendieck
categories. First, we need to define projective objects in an arbitrary abelian category. Let
A be an abelian category.
Definition 2.3.4. An object P A is projective if the functor
HomA (P, ) : A Ab
is exact.
102

For any object X, the functor Hom(X, ) is left exact (see example 1.5.2). So P is
projective if and only if Hom(P, ) is right exact. There is an equivalent, but very useful
definition of projective objects. An object P is projective if and only if for all surjections
: X X 0 and morphisms : P X 0 , there exists a lift : P X of as in the
following diagram:
P

/ X0

/0

To see this, suppose we have : X  X 0 . Then if K = Ker , the following sequence is

exact.
/K
/ X / X0
/0
0
Applying Hom(P, ), we get an exact sequence of abelian groups:
0

/ HomA (P, X)

/ HomA (P, K)

/ HomA (P, X 0 )

/ 0.

The surjectivity of is precisely the lifting property we want P to have.

Lemma 2.3.5. Let A = Mod(R) for a ring R. Then P is projective if and only if P is the
direct summand of a free module over R.
Proof. First, we show that free modules are projective. If F is a free R-module, then by
definition F possesses a basis {f }I . Given a surjection : X  X 0 , choose elements
x X such that (x ) = (f ) for each . Since F is free, we can define : F X by
setting (f ) = x . This is well-defined precisely because F is free.
It easily follows that direct summands of free modules are projective. Let P be a direct
summand of a free module F with complement Q. Given a surjection : X  X 0 and
: P X 0 , extend to P Q = F by = 0:
P Q


/ X0
{

/0

Since F is free,
e has an extension to F , and its restriction |P is the desired extension
of to P .
Finally, we show that if P is projective, then there exists Q such that F = P Q is
free. Indeed, choose a free module F with a surjection : F  P . Consider the following
diagram:
P
s



/P
F

id

/0

By the lifting property, has a lift s, which is a splitting of , i.e. F = P Ker().

103

2.4

Finiteness conditions

Let A be an abelian category, X an object of A.

Definition 2.4.1. The object X is compact (or small) if the functor HomA (X, ) commutes
with direct sums.
L
Let {X }I be a (set-theoretically small) family of objects in A. Recall that X
is by definition the colimit
limI X . By the definition of a colimit, there are canonical
L

morphisms i : X X for each . For any object X in A, we have maps

HomA (X, X )

i,

/ HomA (X,

X ) .

These patch together to yield morphisms

L

HomA (X, X )

/ Hom (X,

X ) .

The object X is compact if X is an isomorphism for every collection {X }.

Definition 2.4.2. The object X is finitely presented if HomA (X, ) commutes with all
small direct limits.
Definition 2.4.3. The object
PX is finitely generated if whenever {X } is a directed system
of subobjects of X such that X = lim X = X, there exists 0 I such that X0 = X.

Definition 2.4.4. The object X is coherent if X is finitely presented and every finitely
generated subobject of X is finitely presented.
Example 2.4.5. Let R = khx1 , . . . , xn i be the free algebra on n generators over a field k.
Clearly R is finitely generated (by the unit) as an R-module. There are ideals I R that are
not finitely generated, and the quotient R/I is finitely generated but not finitely presented.
For many purposes, coherent modules over non-Noetherian algebras are the correct
substitute for finitely generated modules over a Noetherian ring.
Lemma 2.4.6. If A is an abelian category satisfying AB5, then every finitely generated
object is compact.
Proof. This is a good exercise.
The converse is false. There is a counter-example due to Rentschler: there exists a
commutative integral domain R such that the field of fractions K of R is compact but not
finitely generated. (Need reference here!)
Lemma 2.4.7. Let A be an AB5 abelian category, and let P Ob(A) be projective. Then
P is finitely generated if and only if P is compact, if and only if P is finitely presented.
104

Proof. We will prove that if P is compact and projective, then P is finitely presented. First,
observe the following. Let I be a directed set, and {(X , f )}I be a direct system of
objects in A. Then limI X is the colimit of the diagram X : I A. Let |I| be the

L
underlying (discrete) set of I. Then lim|I| X = X by definition, with canonical

L
embeddings i : X , X . We have canonical
X ; these
L morphisms j : X lim

patch together to yield a canonical morphism j : X lim X . The morphism j is

always an epimorphism, and fits into an exact sequence

M
M
/0
/
/ lim X
X(,)
X

(,)S



where S = {(, ) I I : 6 }, f : X X and X(,) = Im i f i .
Assume P is compact projective, and apply HomA (P, ) to the sequence above. We get
 L





L
/ Hom P, lim X
/0
/ Hom P,
Hom P, (,)S X(,)
I X

O
O
O

(,)S

Hom P, X(,)

Hom(P, X )

/ lim Hom(P, X )

/0

The first two vertical arrows are isomorphisms because we assumed P is compact. Since P is
projective, rows are exact. Then 5-lemma implies the third vertical arrow is an isomorphism.
So Hom(P, ) commutes with arbitrary direct limits, hence P is finitely presented.
Projective objects are analogs of vector bundles. This analogy can be made somewhat
precise.
Theorem 2.4.8 (Swan). Let X be a para-compact topological space, and let A = C(X)
be the ring of continuous C-valued functions on X. Then the category Vect(X) of vector
bundles on X is equivalent to the category Proj(A) of projective A-modules via the functor
(X, ).
Theorem 2.4.9. Let X = Spec(A) is a Noetherian affine scheme. Then the category
Vect(X) of vector bundles on X is equivalent to the category Proj(A) of projective A-modules
via the functor (X, ).
Lemma 2.4.10. Let F : A B be an exact functor between abelian categories. Then F is
faithful if and only if F (X) 6= 0 whenever X 6= 0.
Proof. Suppose f is faithful (injective on hom-sets). Then X 6= 0 implies idX =
6 0, which
implies F (idX ) 6= 0, hence F (X) 6= 0. (In other words, arbitrary faithful functors send
nonzero objects to nonzero objects.)
Conversely, suppose that for all X 6= 0 we have F (X) 6= 0. Let f be a nonzero morphism
in A. Then Im(f ) 6= 0. Since F is exact, it commutes with taking kernels, cokernels and
images. So F (Im f ) = Im F (f ) 6= 0, hence F (f ) 6= 0.
105

Corollary 2.4.11. P is a projective generator (i.e. Hom(P, ) is faithful) if and only if

Hom(P, X) 6= 0 whenever X 6= 0.
Theorem 2.4.12. Let A be a Grothendieck category. Assume A has a compact projective
generator P . Then the functor HomA (P, ) : A EndA (P )-Mod is an equivalence of
categories. Conversely, module categories have compact projective generators.
For example, we can use this to show that A = Qcoh(P1 ) is not a module category. In
fact, it has no nonzero projective objects. The same holds for Qcoh(X) whenever X is a
projective variety of positive dimension.

Classical derived functors

3.1

Injectives and injective envelopes

Let A be an abelian category.

Definition 3.1.1. An object E in A is injective if the funtor HomA (, E) is exact.
In other words, E is injective if whenever we have an exact sequence
0

/ X0

/X

/ X 00

/ 0,

the following sequence is exact.

0

/ Hom(X 00 , E)

/ Hom(X, E)

/ Hom(X 0 , E)

/0

We already know it is exact on the left, so the only nontrivial condition is that Hom(, E)
is right-exact.
Definition 3.1.2. An abelian category A has enough injectives if every X Ob(A) is
isomorphic to a subobject of an injective object in A.
Tautologically, E is injective in A if and only if E is projective in the dual category A .
So it would appear that injective and projective objects are dual in some sense. However,
in practice we usually work with Grothendieck categories, and for A Grothendieck, A is
Grothendieck if and only if A is the zero category. Proving this is a good exercise.
We will see that any Grothendieck category has enough injectives. On the other hand, a
Grothendieck category does not necessarily have enough projectives (or any projectives at
all). A typical example is A = Qcoh(X) for X a projective scheme. This category has no
projective objects.
If A is a Noetherian abelian category, then A as a rule does not have enough invectives.
It is possible for such categories to have enough projectives just take the category of finitely
generated modules over a semisimple ring. In any case, even if A does not have enough
injectives, its inductive closure Ae always have enough injectives.
106

Theorem 3.1.3. For any ring R, the categories Mod(R) and R-Mod have enough injectives.
Q
Lemma 3.1.4. Let {E }I be a set of injectives in A. Assume that the product E = E
exists. Then E is injective.
Q
Proof. Note that E = limI E . For all X in A, we have

Y
Hom(X, E) '
Hom(X, E ).
I

The latter is a product of exact functors, hence exact. It follows that E itself is exact.
Definition 3.1.5. Let R be an associative unital ring. An element x R is called right
regular if the right R-module map r 7 xr is injective on R, i.e. if xr = 0 implies r = 0.
Definition 3.1.6. Let M be a right R-module. We say M is divisible if M x = M for any
right regular x R.
m0

In other words, M is divisible if for any m M and right regular x R, there exists
M such that m0 x = m.

Lemma 3.1.7. Every injective module is divisible.

Proof. If E is injective, then consider for any right regular x R the map x : R R
given by r 7 xr. This injective, so by the fact that E is injective, the morphism x : E '
Hom(R, E) E ' Hom(R, E) is surjective. It follows that xE = E, i.e. E is divisible.
Theorem 3.1.8 (Baer). If every ideal in R is principal, then divisible R-modules are
injective.
Proof. Fill this in.
Lemma 3.1.9. If A is an AB5 category, then the canonical morphism
monic.

X is

Example 3.1.10. Let R = Z. Then we claim that Mod(Z) = Ab has enough injectives.
Indeed, Q is obviously divisible, and hence injective. By Lemma 3.1.4, any product of copies
of Q is injective. Moreover, Baers theorem shows that any quotient of injectives is injective.
In particular, Q/Z is injective. But any Z-module can be identified with a subquotient of a
product of copies of Q, via
L
Q

/
MZ
MQ


M.
Thus M is isomorphic to a submodule of an injective.
107

Proposition 3.1.11. Let F : A  B : G be an adjoint pair of functors between abelian

categories. If F is exact, then G preserves injectives. If G is exact, then F preserves
projectives.
Proof. Well prove that if F is exact, then G preserves injectives. Let E be an injective
object in B; we want to show that G(E) is injective in A. But
HomA (, G(E)) ' HomB (, E) F ,
and the composition of exact functors is exact, so G(E) is injective.
We are now ready to prove Theorem 3.1.3. Recall that if f : S R is a ring homomorphism, we have an adjoint triple
f

Mod(R)

/ Mod(S)

f!

where f (M ) = M S R and f ! (M ) = HomS (R, M ). In fact, f ! f = id and f f = id. Since

f has right and left adjoints, it is exact. By the proposition, f maps injectives to injectives.
Now let f : Z R be the canonical unital homomorphism. Then f ! : Mod(Z) Mod(R)
maps injectives to injectives. For any right R-module M , choose an injective abelian group
E with an embedding f M , E. This gives
M = f ! f M , f ! E,
where f ! E is an injective R-modules.

3.2

Canonical constructions on complexes

The following constructions are motivated by topology, but they actually make sense in
much greater generality (i.e. for model categories, as we will see later in the course).
Suspension
Let A be an abelian category and Com(A) be the category of (cohomological) chain complexes
(X , d ).
Definition 3.2.1. Let k Z. The k-th suspension functor [k] : Com(A) Com(A) is
i+k
defined by X 7 X [k], where X [k] = X i+k and diX[k] = (1)k dX
.
Lemma 3.2.2. The functor [k] is an auto-equivalence of categories, with inverse [k].
Moreover, [k] [m] = [k + m].
108

Cone
Well construct the functor cone explicitly. There is a more functorial construction, and we
will describe it later when we will be talking about model categories. Given f : X Y ,
set cone(f ) = Y X [1]. The complex cone(f ) looks like
k
/ Y k X k+1 dcone / Y k+1 X k+2

where

dkcone


=

dkY
0

/ ,


f k+1
.
dk+1
X

k
It is easy to check that dk+1
cone dcone = 0 if and only if
k
dk+1
Y dY = 0
k+1
dk+2
X dK = 0

f k = f k dkX .
dk+1
Y
Proposition 3.2.3. A morphism f : X Y is a quasi-isomorphism if and only if
cone(f ) is acyclic.
Proof. Recall there is a short exact sequence
/Y

/ cone(f )

/ X [1]

/0

Consider the associated long exact sequence:

/ Hi1 (X[1])

Hi (X)

/ Hi (Y )
8

/ Hi (cone f )

fi

/ Hi (X[1])

Hi+1 (X)

/ Hi+1 (Y )
8

f i+1

The arrows f i and f i+1 are isomorphisms, which is easily seen to imply Hi (cone f ) = 0.
Exercise Let Comb (A) be the full subcategory of Com(A) consisting of bounded complexes
(complexes X with X i = 0 for all |i|  0). The category A naturally embeds into Comb (A).
Show that Comb (A) is generated by A in the sense that every X Comb (A) can be
obtained by taking iterated suspensions and cones of objects in A.
Cylinder
Definition 3.2.4. Let f : X Y be a morphism of complexes. The cylinder of f is
p [1]

X
Cyl(f ) = cone(cone(f )[1]
X ).

109

Explicitly,
Cyl(f ) = Y X X [1]
with differentials

dkY
=0
0

dkcyl

0
dkX
0

f k+1
idX k+1
dk+1
X

Note that we have a short exact sequences associated to f : X Y :

iX

/ X

/ Cyl(f ) cone / cone(f )

/0

By Proposition 3.2.3, the morphism iX is a quasi-isomorphism if and only if f is a quasiisomorphism, which happens if and only if cone(f ) is acyclic.
The other short exact sequence associated to f : X Y is
/Y

/ Cyl(f )

/ cone(idX )

/ 0.

Here is the obvious inclusion. It turns out that has a splitting.

Proposition 3.2.5. For f : X Y consider the following diagram
iY

/Y

/ cone(f )

/ X[1]

/ cone(f )

/0

/0

/ X
/ X

/ Cyl(f )
f

/Y

/ Z

/0

where (g, Z ) is defined to be cokernel of f , is defined by (y k , xk , xk+1 ) 7 f k (xk ) + y k and

is given by : (x, y) 7 g(y). Then
1. = idY
2. idCyl(f ) .
3. is quasi-isomorphism.
Mapping cylinders can be characterized for A = Mod(Z). Define the complex I = (0
Z2 Z 0), where (a, b) 7 a b. For any morphism f : X Y , consider the
inclusion X , X I . We claim that there is a cocartesian square:
X
f

i0

/X I

/ Cyl(f )

Proving this is a good exercise.

110

3.3

Classical definition of classical derived functors

We follow Grothendiecks construction. Recall that if A is an abelian category, then A has

enough injectives if every object X in A embeds into an injective object.
Definition 3.3.1. An injective resolution of an object X is an exact complex
0

/X

/ I1

/ I0

/ I2

with each I i injective.

Write I for the complex 0 I 0 I 1 . If we think of X as a complex concentrated
in degree zero, then it makes sense to write : X I for an injective resolution of X. Any
injective resolution : X I is a quasi-isomorphism.
Definition 3.3.2. Dually to the definition 3.1.2, we say A has enough projectives if every
object X in A is the quotient of a projective object.
Similarly, a projective resolution of X is an exact sequence of the form

/ P 2

/ P 1

/ P0

/X

/ 0,

with each P i projective. Again, we denote by P the complex P 1 P 0 0 and

think about projective resolutions as a quasi-isomorphism of complexes : P X.
Lemma 3.3.3. If A has enough injectives, then every object in A has an injective resolution.
Dually, if A has enough projectives, every object in A has a projective resolution.
Proof. Just iterate injective embeddings as follows. Given X Ob(A), we have an embedding
: X , I 0 . We can embed Coker() into another injective I 1 . Then just repeat the
process.
Proposition 3.3.4. Suppose X1 , X2 Ob(A) have injective resolutions 1 : X1 I1 and
2 : X2 I2 . Then any morphism f : X1 X2 admits a lifting fe : I1 I2 making the
following diagram commute:
1
/ I
X1
1


X2

fe

/ I
2

The lift fe is unique up to homotopy.

Proof. Recall that P is projective exactly when hP is exact. So we only need to check the
existence of lifts in the following diagram.
P
X0


/X

111

/0

Reversing arrows, we see that an object I is injective exactly when lifts exist in the following
diagram.
/X
/ X0
0
 ~

An easy iteration of this diagram yields the existence of a lift fe : I1 I2 of f : X1 X2 .

We will show that fe is unique up to homotopy later after gaining some machinery.
A similar theorem holds for projective resolutions. Note that if F is any additive functor
from A to B, then we can extend F in the obvious way to a functor F : Com(A) Com(B). It
is easy to see that F preserves homotopies. In other words, if f fe , then F (f ) F (fe ).
Thus the following definition makes sense.
Definition 3.3.5. Let F : A B be a left exact additive covariant functor. Assume A has
enough injectives. Then we define the (classical) right derived functors of F , as functors
Ri F : A B for i > 0, by
Ri F (X) = Hi (F (I )),
where I is any injective resolution of X. For f : X X 0 , we define
Ri (F )(f ) = Hi (F (fe )),
where fe is a lift of f to injective resolutions.
Theorem 3.3.6. Let A, B and F be as above. Then
1. For all i > 0 the funtors Ri F are additive and independent of the choice of resolution.
2. There is a natural isomorphism R0 F ' F .
3. For any short exact sequence 0 X 0 X X 00 in A, there are morphisms
i : Ri F (X 00 ) Ri+1 F (X 0 ), such that the following sequence is exact.

/ Ri F (X 0 )

/ Ri F (X)

/ Ri F (X 00 )

/ Ri+1 F (X 0 )

4. The i are functorial in exact sequences. In other words, if we define Ri F (0 X 0

X X 00 0) = Ri F (X 00 ) and Ri+1 F (0 X 0 X X 00 0) = Ri+1 F (X 0 ), then
the i are natural transformations Ri F Ri+1 F , where these are viewed as functors
Exc(A) B.
5. If I is injective in A, then Ri F (I) = 0 for all i > 0.
Definition 3.3.7. Let F : A B be a left exact additive functor. An object J Ob(A) is
called F -acyclic if Ri F (J) = 0 for all i > 0.
112

By Theorem 3.3.6, injective objects are universally acyclic, i.e. they are F -acyclic for
any left-exact functor F .
Definition 3.3.8. An F -acyclic resolution of X Ob(A) is an exact complex X , J ,
where J consists of F -acyclic objects.
Proposition 3.3.9. If X J is an F -acyclic resolution, then Ri F (X) ' Hi (F (J )) for
all i.
Example 3.3.10. Flabby sheaves are acyclic relative to the global sections functor. Thus
we can compute sheaf cohomology using flabby resolutions.
We can define (classical) left-derived functors similarly. Given a right-exact additive
functor F : A B, define its left-derived functors Li F : A B as follows. Given X Ob(A),
choose a projective resolution P X, and define
Li F (X) = Hi (F (P )).

3.4

-functors

Classical derived functors satisfy a universal mapping property. Let A, B be abelian categories.
Definition 3.4.1. A (covariant) -functor is a collection T = (T i )i>0 of additive functors
A B given together with natural transformations i : T i T i+1 , where here T i and T i+1
are functors Exc(A) B via T i (0 X 0 X X 00 0) = T i (X 00 ) and T i+1 (0 X 0
X X 00 0) = T i+1 (X 0 ). We require that for any short exact sequence 0 X 0 X
X 00 0, the following sequence be exact:
0

/ T 0 (X 0 )

/ T 0 (X)

/ T 0 (X 00 )

/ T 1 (X 0 )

Definition 3.4.2. A -functor T = (T i )i>0 : A B is universal if for any -functor

T 0 : A B and any natural transformation f 0 : T 0 T 0 0 , there exists a unique extension
f : T T 0 commuting with the i .
Definition 3.4.3. A functor F : A B is called effaceable if for every X in A, there is a
monic f : X , E such that F (f ) = 0.
Dually, a functor F : A B is called coeffaceable if for every X Ob(A), there is an
epic f : P X such that F (f ) = 0.
Theorem 3.4.4 (Grothendieck). Let (E, ) = (Ei , i )i0 : A B be a coeffaceable left
-functor. Then E is universal in the following sense. For any left -functor (T, ) : A B
and any f0 : T0 E0 there exist unique transformations fi : Ti Ei commuting with s.
113

Proof. We will prove this theorem by induction. Given any -functor (T, ) : A B and
f0 : T0 E0 , assume we have constructed fi : Ti Ei for i < k.
We want to construct fk : Tk Ek . For this we need to define fk (X) : Tk (X) Ek (X)
for each X A. For each X, choose p : Y  X, s.t. Ei (Y ) = 0 for all i > 0. Let
X 0 = Ker(p), so that we have a SES in A
0 X0 Y X 0
Since (T, ) and (E, ) are both -functors, by applying T and E we get
/ Tk (X)

/ Tk1 (X 0 )

fk

/ Ek (X)

k1

k1

k1

fk1 (X 0 )

/ Ek1 (X 0 )

k1

/ Tk1 (Y )


/ Tk1 (X)

fk1 (Y )

/ Ek1 (Y )

/ ...

fk1 (X)

/ Ek1 (X)

/ ...

By assumption, Ei (Y ) = 0 for i > 0, so Ek (X) Ek1 (X 0 ) is actually an embedding. Since

E )
k1 fk1 (X 0 ) k1 = 0, exactness of rows implies that Im (fk1 (X 0 ) k1 ) Im(k1


E
Ek1 (X 0 ). This gives fk as k1
fk1 (X 0 ) k1 .
Now we need to check that actually maps {fk } do not depend on the choice Y and
commute with s. Take Y1  X1 and Y2  X2 , where Ei (Y1 ) = 0 and Ei (Y2 ) = 0 for i > 0.
Consider a commutative diagram
/ X0
1

/ X0
2

/ Y1

/ X1

/ Y2

/0

/ X2

/0

This yields to a cubic diagram

/ Tk (X 0 )
1

Tk (X1 )

Tk1 (0 )

Tk ()

z
/ Tk (X 0 )
2

Tk (X2 )
fk (X1 )

fk1 (X20 )

Ek (X1 )

fk (X2 )


/ Ek (X 0 )
1
Ek1 (0 )

Ek ()

Ek (X2 )

fk1 (X10 )

 z
/ Ek (X 0 )
2

114

In this diagram top and bottom faces commute since T and E are -functors. The
front and back faces commute by construction of the maps fk . The right face commutes
by induction. So we only need to show that the left face commutes. Since the front
face commutes, we get fk (X2 ) Tk () = Ek () fk (X1 ). But E(Y2 ) = 0, so
: Ek (X2 ) Ek1 (X20 ) is injective. Hence fk (X2 ) Tk () = Ek () fk (X1 ).

3.5

Main properties of resolutions

The goal of this section is to prove the existence of classical derived functors. Let A be an
abelian category. We assume A has enough projectives, and as before write Com(A) for
the category of chain complexes in A. Recall that if X is a complex in A, its projective
resolution (or approximation) in Com(A) is a quasi-isomorphism P X , where P is
(pointwise) projective.
Lemma 3.5.1. If X Ob(Com(A)) is bounded from above (X i = 0 for i  0), then X
admits a projective resolution.
Proof. The proof is by induction. First, if X i = 0 for i > i0 , put P i = 0 and f i for i > i0 .
Now assume that we have already constructed P i and f i for i > k + 1:
P k+1

/ X k1

/ Xk

dkX

dk+1
P

/ P k+2

f k+1

/ X k+1

dk+1
X

dk+2
P

f k+2

/ X k+2

Consider the cone of the morphism f : P >k+1 X . By assumption, cone(f ) is acyclic in

degree > k + 1.
Recall that cone(f ) = X P >k+1 [1]. The differential dcone : X k P k+1 X k+1 P k+2
is the matrix
 k

dX f k+1
0 dk+1
P
Let Y = Ker(dkcone ). Since A as enough projectives, we can choose P k  Y with P k
projective. Consider the composition
Pk

f k dk+1
P

//Y

/ X k P k+1 .

Since Y = Ker(dkcone ), we have

 k
dX
0

f k+1
dk+1
P

fk
dk+1
P



115


=0

k
which occurs if and only if dkX f k f k+1 dk+1
= 0 and dk+1
P
P dP = 0. Hence, we can extend
f : P X by
dkP

Pk


/ P k+1

fk

f k+1

/ X k+1

Xk

Lemma 3.5.2. Let X : P X and Y : Q Y be projective resolutions for X, Y

Ob(A). Then any f : X Y in A lifts to f : P Q making the following diagram
commute:
f

P


/ Q

/Y

We can choose f so that H0 (f ) = f . Moreover, any two such lifts f , g : P Q are

homotopic.
Proof. Lets write P 1 = X, d0P = X and Q1 = Y , d0Q = Y , so that we have acyclic
complexes P 1 P 0 P 1 0 and Q1 0. Moreover, put f 1 = f . By
induction, assume that f i exists for i > k + 1:
Pk

dkP

f k k
dkQ
Qk

dkP

/ P k+1
"

f k+1

/ Qk+1

dk+1
Q

/ P k+2


/ Qk+2

k+1 k+1 k
k
k
k
Put k = f k+1 dkP . Then dk+1
dP = f k+2 dk+1
Q = dQ f
P dP = 0, so Im( )
k
Ker(dk+1
Q ) = Im(dQ ). Hence we have

Pk
fk

Qk

{ dkQ


/ Im(dk )
Q

/ 0.

By the projectivity of P k , there exists f k such that dkQ f k = k = f k+1 dkP . This finishes the
induction.
116

Now we show that given f , g : P Q lifting f with H0 (f ) = H0 (g ), we want to

construct a homotopy between f and g :
dk1
P

/ P k1

f k1

g k1

 }
/ Qk1

dkP

/ Pk

fk

hk

gk

 }
/ Qk

dk1
Q

/ P k+1


/ Qk+1

hk+1

dkQ

k
k+1 dk . Put h0 = 0. We argue by induction.
such that for all k, f k g k = dk1
P
Q h +h
Assume that we already have constructed hk+1 , hk+2 , . . . and we want to construct hk .
Consider k = f k g k hk+1 dkP , and note that dkQ k = (f k+1 g k+1 dkQ hk+1 )dkP =
k
hk+2 dk+1
P dP = 0. Consider

Pk
hk

Qk1

/ Im(dk1 )
Q

/0

We get the existence of hk by the projectivity of P k .

Note that in the proof of this lemma, we did not use the fact that the Qi are projective
(but we did use the fact that H (Q) = 0). Moreover, we didnt use the fact that P is acyclic.
So we have the following corollary.
Corollary 3.5.3. Let P be (pointwise) projective and bounded above, and let Q be acyclic.
Then any f : P Q is homotopic to zero.

3.6

Definition of classical derived functor via -functors

Recall the definition of a -functor. Let A and B be abelian categories.

Definition 3.6.1. A (left) -functor is a family of additive functors T = (Ti )i>0 , Ti : A B
and morphisms i : Ti+1 (A00 ) Ti (A0 ) of functors on the category of all short exact sequences
0 A0 A A00 0 in A, such that for all such exact sequences, the following sequence
is exact in B.

/ T1 (A)

/ T1 (A00 )

/ T0 (A0 )

/ T0 (A)

/ T0 (A00 )

/0

Definition 3.6.2. Let F : A B be a right-exact additive functor. The classical leftderived functor of F is a pair (LF, ) consisting of a left -functor LF = (Li F, i )i>0 and
117

an isomorphism of functors L0 F
F , which is universal among all left -functors in the
following sense. For all left -functors (T, ) and morphisms f0 : L0 F T0 , there is a
unique sequence of morphisms fi : Li F Ti extending f0 .
Remark 3.6.3. It follows from the definition that (LF, ) is determined up to unique
isomorphism, if it exists. The functor LF is only determined up to automorphisms of F .

F . Then
Indeed, given (LF, ) and (L0 F, 0 ), by definition : L0 F
F and 0 : L00 F
f0 = (0 )1 is an isomorphism L0 F L00 F , which extends uniquely to an isomorphism
f : LF L0 F .
Previously we defined derived functors using injective (resp. projective) resolutions.
Here, we prove that under various assumptions, this agrees with the above definition.
Theorem 3.6.4. Let F : A B be a right-exact additive functor. Assume A has enough
projective objects. For any X Ob(A) choose a projective resolution P X in Com(A).
Define, for i > 0,
Li F (X) = Hi (F (P ))
Then (Li F )i>0 is the classical left-derived functor of F in the sense of definition 3.6.2.
Proof. Our main tools will be lemmas 3.5.1 and 3.5.2. In the proof we need to do the
following steps.
1. prove that the Li F are actually functors;
2. check independence of Li F on the choice of resolutions (up to isomorphism);
3. prove that Li F actually form a -functor. Namely, we need to show that there exist
connecting morphisms i : Li+1 F (A00 ) Li F (A0 ) satisfying the definition 3.6.1;

4. prove the existence of isomorphism L0 F

F;
5. finally, we need to prove the universality of LF .
First we prove the Li F are functors. Let f : X Y be a morphism in A. Let P X
and Q Y be our chosen resolutions. By Lemma 3.5.2, there is f : P Q lifting
f . Define Li F (f ) = Hi (F (f )). By the same lemma, any two lifts f , g : P Q
are homotopic, i.e. we can write f g = dh + hd. Since F is additive, we have
F (f ) F (g ) = F (h)F (d) F (d)F (h), so F (f ) and F (g ) are homotopic as morphisms
F (P ) F (Q ). It follows that Hi (F (f )) = Hi (F (g )), so Li F (f ) is well-defined.
Next we show that Li F is independent of the choice of projective resolutions. Suppose
we have two resolutions P , Q X. By Lemma 3.5.2, the map idX lifts to morphisms
f : P Q and g : Q P such that f g idF (Q ) and g f idF (P ) . This
tells us that F (f ) and F (g ) are inverses (up to homotopy). Thus F (P ) and F (Q ) are
homotopy-equivalent, so Hi (F (P )) ' Hi (F (Q )). Thus Li F is well-defined.
118

Now we construct the connecting homomorphisms i . Consider a short exact sequence

f

0X
Y
Z 0 in A. Replace X and Y by projective resolutions. By Lemma 3.5.2,

f lifts to f : P Q . Let R := cone(f ). We claim that R is a projective resolution of

Z. Recall there is an exact sequence
0

/ Q

/ R

/ P [1]

/0

()

of complexes in A.
Remark 3.6.5. Instead of writing the sequence () we could have written an exact distinguished triangle
P

/ Q

/ P [1] f [1] / Q [1]

/ R

/ ...

This sequence is called triangle since we can write it as a triangle

f

Q
i

P o

[1]

Applying H to () gives a long exact sequence

/ H1 (Q )

/ H1 (R )

/ H1 (P [1])
o

/ H0 (Q )

/ H0 (R )

/ H1 (P [1])

o
/Y

o
/Z

In this diagram we have H1 (Q ) = 0 since Q is acyclic in higher degrees, H1 (P [1]) '

H0 (P ) ' P . Since f is inclusion, the image of H1 (R ) in H1 (P )[1] is zero, and since
H1 (Q ) = 0 we conclude that H1 (R ) is also zero. It is obvious from acyclicity of P
and Q that Hi (R ) = 0 for i 1. Since H0 (P [1]) = H1 (P ) = 0, it follows that
H0 (R ) ' Y /X ' Z. Hence R Z is a quasi-isomorphism, and so R is a resolution of Z.
But Ri = Qi P i+1 is termwise projective, so R is actually a projective resolution of Z.
We can apply F to (). Since P is projective and F is right-exact, we get the short exact
sequence
/ F (Q )
/ F (R )
/ F (P [1])
/0
0
By the additivity of F , F (R ) = cone(F (f )). Moreover F (P [1]) = F (P )[1], so applying
H , we get

/ L0 F (X)

/ L0 F (Y )

/ L0 F (Z)

/ L1 F (X)

We need to prove this exact sequence is functorial in short-exact sequences. Suppose we

replaced the short exact sequence 0 X Y Z 0 by 0 X1 Y1 Z1 0. Pick
119

projective resolutions as above. Then we get the following diagram

f

P1

/ R

h
f1

/ Q

g1

/ R


/X


/Y

/Z

 }
/ X1

/ Q

 }
/ Y1

f1

/0

g1

/ Z1

/0

Then the upper square commutes up to homotopy, i.e. f f1 . Choose

k
k
k
a homotopy hk : P k Qk1
such that k f k f1k k = dk1
1
Q1 h + h dP . Define
k : Rk R1k by
k : Qk P k+1 Qk1 P1k+1
given by the matrix
 k

hk+1
=
0 k+1
k

We claim that this gives us a morphism of exact triangles. Namely, we have the following
diagram:
P

/ Q

P1

f1

/ R

/ Q
1

i1

/ P [1]
[1]

p1

/ R
1


/ P [1]
1

We need to check that the diagram above commutes up to homotopy. We have already seen
that the first square commutes up to homotopy. Lets check the second square.
Qk


Qk1

ik

/ Qk P k+1

q_k 

/ (q k , 0)
_

ik


/ Qk P k+1
1
1

k (q k ) 


/ ( k (q k ), 0)

(4.1)

So the second square commutes. Similarly for the third square.

If we now apply F to the diagram (4.1) and take cohomology, we will obtain the following
commutative diagram between long exact sequences

/ Li+1 F (Z)

/ Li F (X)

/ Li F (Y )

/ Li F (Z)


/ Li+1 F (Z)


/ Li F (X)


/ Li F (Y )


/ Li F (Z)

120

This exactly means functoriality on short exact sequences. Thus, (Li F, ) is a left
-functor.
Next we need to prove existence of natural isomorphism L0 F ' F . Indeed, choose any
X A. Replace X by resolution P  X and look at the first two terms
P 1

/ P0

/X

/0

Applying right exact functor F to this exact sequence, we get an exact sequence
F (P 1 )

F (d)

/ F (P 0 )

/ F (X)

/0

Then L0 F (X) := H0 (F (P )) = Coker F (d) ' F (Coker(d)) ' F (X). Notice that here we
used right exactness of F again, which gave Coker F (d) ' F (Coker(d)).
Finally, we need to check the universal property.
Lemma 3.6.6. If A has enough projectives then the functor Li F, is coeffaceable.
Proof. Given X A we can replace it by a projective cover P  X for some projective
P A. Then we claim that Li F (P ) = 0, i > 0. But this is obvious, since P is projective,
so we can choose its projective resolution to be the complex P = [0 P 0].
Having proved lemma 3.6.6 we can apply theorem 3.4.4 to automatically get the universality of classical derived functors.

3.7

Examples of derived functors

The functors Exti are classical derived functors of Hom : (X, Y ) 7 Hom(X, Y ). We need to
make this precise, because we can think of Hom(, ) as a functor in either variable. We
will assume that A is an abelian category with enough injectives and enough projectives.
Then, for X, Y Ob(A), consider the functors
Hom(, Y ) : A Ab
Hom(X, ) : A Ab
Both these functors are left-exact, so we can define
Exti1 (X, Y ) = Ri Hom(, Y )(X) = Hi (Hom(P , Y ))
Exti2 (X, Y ) = Ri Hom(X, )(Y ) = Hi (Hom(X, I )),
where P X is a projective resolution of X and Y I is an injective resolution of Y .
Theorem 3.7.1. There is a natural isomorphism Exti1 (X, Y ) ' Exti2 (X, Y ).
121

We prove this by giving another (more explicit) construction of Ext, due to Yoneda.
Lets start with i = 1. Fix X and Y in A, and consider the set E 1 (X, Y ) of all short exact
sequences
= (0 Y Z X 0)
Define an equivalence relation on E 1 (X, Y ) by setting 0 if there is a morphism
: Z Z 0 in A such that the following diagram commutes:
/Y

/Z

/X

/0

/X

/0

/Y

/Z

The snake lemma forces to be an isomorphism. Define Ext1 (X, Y ) = E 1 (X, Y )/ . The
additive structure on Ext1 (X, Y ) is defined by + : Ext1 (X, Y ) Ext1 (X, Y ) Ext1 (X, Y ).
Given 1 = (0 Y Z1 X 0) and 2 = (0 Y Z2 X 0), we define 1 + 2
as follows. There is an exact sequence 1 2
0

/Y
O

/Z
O

/X

/0

e
/Z

/X

/0

/Y Y

/Y Y

/ Z1 Z2

/X X

/0

e = X XX (Z1 Z2 ). We define Z
Here Ze is the pullback in the lower-right square, i.e. Z
to be the pushout in the upper-left square, i.e.
Z = Y tY Y Ze = (X XX (Z1 Z2 )) tY Y Y .
The sequence 1 + 2 is 0 Y Z X 0. We need to check that + is compatible
with . The unit is the trivial split-exact sequence 0 Y Y X X 0.
In general, for i = k > 1, we define E k (X, Y ) to be the set of exact sequences of the form

= 0 Y Zk Zk1 Z1 X 0 .
{z
}
|
Z

We say that is elementary equivalent to 0 E k (X, Y ) if there exists a morphism

: Z Z0 such that the following diagram commutes:
0

/Y

/Y

/ Zk


/ Zk1

/ Z0
k

/ Z0

k1

k1

122

/ Z1


/X

/0

/X

/0

/ Z0
1

We put 0 if there is a finite chain of elementary equivalences (of unspecified direction)

(1)

(n)

Z0 .

Z Z Z

More formally,given X, Y Ob(A) and k > 1, define the following category Extk (X, Y ) with
objects complexes of length k having homology

Y if i = k
Hi (Z ) = X if i = 1

0 otherwise
Morphisms in Extk (X, Y ) are morphisms : Z Z0 of chain complexes satisfying H1 ( ) =
idX , H1 ( ) = idY . We could then define Extk (X, Y ) = E k (X, Y ) = 0 (Extk (X, Y )).
(the definition using Ext does not work when k = 1.)
By convention, Ext0 (X, Y ) = HomA (, Y ).
Lemma 3.7.2. The Yoneda construction of Extk defines a bifunctor, covariant in Y and
contravariant in X, i.e. Extk : A A Ab.
p1

Proof. Let = (0 Y Z X 0 and similarly for 0 , and let f : X 0 X be a

morphism. We need to construct f : Extk (X, Y ) Extk (X 0 , Y ). Let Z10 = Ker(Z1 X 0
X), where Z1 X 0 X is the morphism (z, x) 7 p1 (x) f (x). Similarly, define Zi0 = Zi
for i = 2, . . . , k. We have a commutative diagram
0

/Y

/ Z0
k

/ Z0
2

/ Z0
1

/ X0

/0

/Y

/ Zk

/ Z2

/ Z1

p1

/X

/0

Essentially, let V = (0 Y Zk Z2 0. We have a commutative diagram

0

/ V

/ Z0
1

Z0

/0

/ V

/ Z1

/X

/0

We define f to be the class of the exact sequence 0 Y Zk Z2

Z10 X 0 in Extk (X 0 , Y ). As an exercise, verify that when compositions are defined,
(g f ) = f g .
Theorem 3.7.3. If A has both enough injectives and enough projectives, then for all
X, Y Ob(A), there are isomorphisms of abelian groups:
Exti1 (X, Y ) ' Exti (X, Y ) ' Exti2 (X, Y ).
123

Proof. We prove the first isomorphism by showing that{Exti (, Y )} is a universal -functor.

First, we need to show that Ext (, Y ) is effaceable, and second we need to show that it is a functor. For the first, it suffices to show that Exti (P, Y ) = 0 for i > 1 whenever P is projective.
For i = 1, Ext1 (P, Y ) = 0 because any exact sequence 0 Y Z P 0 has a splitting.
For i > 2, it is easy to check that every exact sequence 0 Y Zk Z1 P 0
is equivalent to a (trivial) sequence of the form 0 Y 0 P P 0.
f

Now we show that Ext (, Y ) is a -functor. Given 0 X 0

X
X 00 0, we first
construct ek : E k1 (X 0 , Y ) E k (X 00 , Y ). We want an exact sequence

/ Extk1 (X 00 , Y )

k1
/ Extk1 (X 0 , Y ) / Extk (X 00 , Y )

/ Extk1 (X, Y )

00
Given Z0 E k (X 0 , Y ), define Z100 = X, and Zi+1
= Zi0 for i = 1, . . . , k 1. We claim that
Z00 E k (X 00 , Y ), and that this assignment is compatible with the equivalence relation on
E k (X 0 , Y ).

Exercise Let A = Ab. Prove that ExtiA (X, Y ) = 0 for all i > 2, where X and Y are
arbitrary abelian groups.
Section 3 in the Appendix A demonstrates an application of the classical derived functors
to representation theory of quivers.

124

Chapter 5

Derived categories
1
1.1

Localization of categories
Motivation

We start with an example motivating the use of derived categories. Let k be an algebraically
closed field of characteristic zero. Consider the projective line P1k over k, and let A = coh(P1k )
be the category of coherent sheaves on P1k . Let E be the object OP1 OP1 (1) in A. One
usually calls E the tilting sheaf.
It is easy to compute EndA (E ):
EndA (E ) = HomA (O O(1), O O(1))


HomA (O, O)
HomA (O, O(1))
'
HomA (O(1), O) HomA (O(1), O(1))


k k 2
'
0 k
' k[ ].
Write B = EndA (E ) for this ring.
Note that if F coh(P1 ), then HomA (E , F ) is naturally a right B-bimodule, with the
action of B = HomA (E , E )
HomA (E , F ) HomA (E , E ) HomA (E , F )
coming from composition. Thus Hom(E , ) induces a functor
F : coh(P1 ) Mod(kQ) ' Repk (Q ),
where Q is the quiver and Repk (Q ) is the category of its representations (see Section
3 in the Appendix A on quivers).
125

The functor F has a right-adjoint G, induced by tensoring with E . One might hope
that the functor F is an equivalence of categories, but this is not the case. For example, let
F = OP1 (1). Then F (F ) ' (X, O(1)) (X, O(2)) = 0, but F 6= 0.
Lets compute Ext1 (E , F ). This is, by definition Ext1 (E , O(1)), which is isomorphic
by Grothendieck-Serre duality (Reference here?) to
Ext1 (E (1), O(2)) ' H0 (X, E (1)) = (O(1) O) ' (O) ' k 6= 0.
So we see that even though F (OP1 (1)) = 0, the derived functor RF (OP1 (1)) = Ext1 (E , O(1))
is non-zero, allowing us to retain some more information.
Theorem 1.1.1 (Beilinson, [Bei78]). There is an equivalence of (triangulated) categories
R Hom(E , ) : Db (Coh P1 ) Db (Repk Q ).
Here Db (Repk Q ) denotes the derived category of the category of modules over the
path algebra kQ. The derived category of an abelian category A can be defined as a
category obtain from the category Com(A) of complexes in A by formally inverting all
the quasi-isomorphisms. More precisely, we have the following definition.
Definition 1.1.2. Let A be an abelian category. Let S be the class of all quasi-isomorphisms
in Com(A). Then the derived category D(A) of A is the localization D(A) = Com(A)[S 1 ]
of A at S.
To explain what this definition actually means, we now begin systematic study of
localization of categories.

1.2

Definition of localization

For a category D, let Iso(D) Mor(D) be the class of all isomorphisms in D.

Definition 1.2.1. Let S be a (non-empty) class of morphisms in a category C. The
localization of C at S is a pair (C[S 1 ], Q) consisting of a category C[S 1 ] and a functor
Q : C C[S 1 ] such that
(L1) Q(S) Iso(C[S 1 ])
(L2) Q is universal among all F : C D such that F (S) Iso(D). In other words, for
any such functor F there exists unique (up to isomorphism) functor Fe : C[S 1 ] D
such that F ' Fe Q.
The category C[S 1 ] is called the localization category (or simply localization) of C at S
and Q is called the localization functor. We will usually denote localization simple by C[S 1 ]
omitting the functor Q.
Lemma 1.2.2. If the localization C[S 1 ] exists, then it is unique up to unique equivalence.
126

Proof. It simply follows from the universal property (L2).

Definition 1.2.3. Let S be a class of morphisms in a category C. Denote by S the smallest
class in Mor(C) such that
(S1) S S
(S2) Iso(C) S
(S3) two out of three property: if f, g Mor(C) are composable and two out of three
then the third one is also in S
morphisms f, g, gf are in S,
The class S is called saturation of S. We say that S Mor(C) is saturated if S = S.
Example 1.2.4. The following are examples of saturated classes of morphisms.
1. the class of all isomorphisms in a category C;
2. the class of quasi-isomorphisms in Com(A) for any abelian category A;
3. the class of all weak equivalences in a model category (we will discuss model categories
in Section?).
Remark 1.2.5. We can regard any saturated class S of morphisms in C as a wide subcategory
of C (i.e. Ob(S) = Ob(C)). Indeed, axiom (S3) implies S is closed under composition.
Lemma 1.2.6. For a class S Mor(C), localization C[S 1 ] exists if and only if C[S1 ]
exists, and then they are isomorphic.
Proof. Suppose C[S 1 ] exists. Let Q : C C[S 1 ] be the localization functor, and let
F (S) Iso(D),
Iso(D). Since S S,
F : C D be any functor such that F (S)
1
e
e
so there exists a unique F : C[S ] D such that F ' F Q. Since Q is a functor,
Q(S) Iso(C[S 1 ]) = Iso(C[S 1 ]). Thus if C[S 1 ] exists, so does C[S1 ] and
Q(S)
C[S 1 ] ' C[S1 ].
: C C[S1 ] is the corresponding localization. Let
Conversely, suppose C[S1 ] and Q
F (S) Iso(D) = Iso(D).
F : C D be a functor such that F (S) Iso(D). Then F (S)
1
Note that
This shows that there is a unique Fe : C[S ] D such that F ' Fe Q.
1

Q(S) Q(S) Iso(C[S ]), which completes the proof.

Lemma 1.2.7. Let S1 , S2 Mor(C) be such that C[S11 ] and C[S21 ] exist. If S1 S2 , then
C[S21 ] ' C[S11 ][QS1 (S2 )1 ].
Proof. This is a good exercise.
127

Remark 1.2.8. The two-out-of-three property can be strengthened to give the so-called
two-out-of-six property (see [DHKS04]) which is defined as follows. Given any triple f, g, h
of composable morphisms, then in the diagram

hgf
gf

hg

 

'/ 

if gf, hg S, then f, g, h and hgf are in S. As an exercise, try showing that the collection
of quasi-isomorphisms in Com(A) satisfies this property.

1.3

Calculus of fractions

Let C be a category and S Mor(C). We will always assume S is saturated. Consider

diagrams
Z`
X`
Y
f >
f >
s
s
X

with s S. We call such diagrams left S-fractions (resp. right S-fractions.) We think
of the diagram on the left as an avatar for s1 f , and the diagram on the right as an avatar
for f s1 , and whenever it wont lead to confusion we will denote these diagrams simply
by s1 f and s1 f .
Consider a chain
f1

> Z 1 ` s 1 f2 > Z 2 ` s 2

Y1

Y2

fn

; Z n a sn

Yn1

Yn

1
1
which we formally denote by s1
n fn s2 f2 s1 f1 . Such diagram is called composit
S-fraction.

Definition 1.3.1. Call two composite fractions elementary equivalent if one can be obtained
from the other by one of the following rules
(E1) adding (inserting inside the chain) s1 s or s s1 with s S
(E2) replacing g id1 f by g f
f

> Z _ id g > Y

gf

X
128

>Y

1
1
(E3) replacing s1
1 id s2 by (s2 s1 )

X`

s1 id

? Z ` s2

X`

s2 s1

We say that two composite fractions are equivalent if there is a chain of elementary equivalences relating them.
Definition 1.3.2. Let C be a category, S Mor(C) a saturated class of morphisms. We
define C[S 1 ] to be the category with the same objects as C, and with
HomC[S 1 ] (X, Y ) = {equiv. classes of compos. S-fractions from X to Y }.
Composition is induced by concatenation.
Remark 1.3.3. There is a subtle problem with this definition. Namely, HomC[S 1 ] (X, Y )
might be a proper class and not a set.
Proposition 1.3.4. If C is a small category and S Mor(C), then the category C[S 1 ]
defined in 1.3.2 is indeed the localization of C at S.
Proof. Define a functor Q : C C[S 1 ] that is the identity on objects, and sends a morphism
f : X Y to the formal fraction id1
Y f . It is easy to see that Q is a functor; we have
id1 (f g) (id1 f ) (id1 g).
Suppose we have a functor F : C D with F (S) Iso(D). We define Fe : C[S 1 ] D
by Fe(QX) = F (X) and
1
1
1
Fe(s1
F (f1 ).
n fn sn1 s1 f1 ) = F (sn )

It is easy to check that Fe is unique up to natural isomorphism.

Remark 1.3.5. Another problem with localization is that the set of composite fractions
tends to be very complicated. Moreover, it is difficult to see what structure will be preserved
under localization. For example, if C is additive, it is not clear whether C[S 1 ] is again
additive. For other examples of bad behavior of localization of categories see [Toe11].
1
1
1 g f are equivalent,
Notice that fractions of the form s1
2 f2 s1 f1 and s2 t
provided t S and tf1 = gs1 . In other words, diagram

f1

> Z 1 ` s1 f 2 > Z 2 ` s2

Y1
129

Y2

1 g f provided by the diagram

is equivalent to the fraction s1
2 t
g
f1

=Z a t

> Z 1 ` s 1 f2 > Z 2 ` s 2

Y1

Y2

Indeed, we have the following calculation

1
1
1
s1
f2 s1
2 f2 s1 f1 s2 id id
1 f1
1
s1
t id1 f2 s1
2 id t
1 f1

(ts2 )1 (tf2 ) s1
1 f1
= (ts2 )1 (gs1 ) s1
1 f1
(ts2 )1 (g id1 f1 )
(ts2 )1 (gf1 ).
Remark 1.3.6. If C has a null object (i.e. an object 0 which is both initial and terminal)
then the corresponding object 0 in C[S 1 ] is a null object in the localized category. To see
this, we need to show that HomC[S 1 ] (0, ) and HomC[S 1 ] (, 0) are both constant functors
with value {}. This is easy.
Example 1.3.7 (Universal (Cohn) localization of rings). Let R be a unital ring. Let be
a set of morphisms : P Q, where P, Q range over some finitely-generated projective
(left) R-modules. The universal localization of R at is a ring R[1 ] together with
a homomorphism : R R[1 ] such that that the corresponding induction functor
: R-Mod R[1 ]-Mod sends all to isomorphisms.
Theorem 1.3.8 (Bergman). If R is hereditary (see definition 3.4.16 in the Appendix A),
then all the universal localizations R[1 ] are also hereditary. Moreover, universal localization
1
1
is pseudo-flat in the sense that TorR
1 (R[ ], R[ ]) = 0.

1.4

Ore localization

Definition 1.4.1. We say that S Mor(C) satisfies the left Ore conditions if
(LO1) S is saturated;
(LO2) for any right S-fraction g t1 : X Y , there exist a left S-fraction s1 f such
that f t = sg;
(LO3) if f s = gs with f, g : X Y and s : Z X, s S, then there exists t S such
that tf = tg
Z

/X

f
g

// Y

130

// Y

/W

Similarly, one can define right Ore conditions.

Definition 1.4.2. We say that S Mor(C) satisfies the left Ore conditions if
(RO1) S is saturated;
(RO2) for any left S-fraction s1 f with f Mor(C), s S, there exists a right
S-fraction g t1 with t S, g Mor(C) such that f t = s g;
(RO3) if tf = tg with f, g : X Y and t : Y Z, t S, then there exists s S,
s : W X such that f s = gs.
Exercise(hard) What happens to the localization if in the above definition one replaces
the two-out-of-three property with the 2-of-6 property (see Remark 1.2.8)?
Definition 1.4.3. Fix S Mor(C). For any X, Y Ob(C), we define an equivalence
relation ` on the set of left S-fractions from X to Y as follows. Two fractions

f1

> Z1 `

> Z2 `

f2
s1

s2

are ` -equivalent if there exists t S, t : Z1 Z2 such that f2 = tf1 and s2 = ts1 . In

1 (tf ). This definition can be represented by the
other words, we say that s1
1
1 f ` (ts1 )
following commutative diagram:

f1

> Z1 h

/6 Z 2
`

s2

s1

f2

Remark 1.4.4. A more standard way to define an equivalence relation on left fractions
(under the assumption that S is multiplicatively closed, but not necessarily saturated) is via
the following commutative diagram:
f

f1

>Z `

4 Z2 `

> Z1 j

s2

s1

f2

If S is saturated, then this definition is equivalent to the definition 1.4.3 above. All
interesting examples of localizations one meets in real life are localizations over a saturated
class of morphisms, so our definition 1.4.3 is usually enough. For example, quasi-isomorphisms
in the category of complexes in an abelian category is a saturated class.
131

Lemma 1.4.5. If S is left Ore, then

HomC[S 1 ] = {left S-fractions X Y }/ ` .
Proof. Observe that every left S-fraction is trivially a left composite fraction. We claim that
the following commutative diagram can be filled in:
{left S-fractions X Y } 

/ {left composite fractions X Y }




/ Hom 1 (X, Y )
C[S ]

{left S-fractions}/ `

We begin by showing that the map defined in the obvious way as a map
{left S-fractions X Y } HomC[S 1 ] (X, Y )
actually descends to the quotient {left S-fractions}/ ` .
1
If s1
1 f1 ` s2 f2 , then there exists t S such that s2 = ts1 and f2 = tf1 . We then
compute:
1
1
s1
f1
1 f1 s1 id id
1
s1
t id1 f1
1 id t

(ts1 )1 (tf1 )
= s1
2 f2 .
Now we construct the map inverse to . Take a composite fraction a = s1
n fn
1
f1 . Using (LO2), we replace each right fraction fi s1
with
s

f
i1,i . Similarly,
i1
i1,i
1
1
we can replace fi,i+1 si1,i with si1,i+1 fi1,i+1 . For n = 4, the following commutative
diagram illustrates the described process:
s1
1

f1,4

f1,3

f1,2

f1

> Z1 b

Z
< 1,3 b s1,3 f2,4

Z
< 1,2 c s1,2 f2,3
f2

s1

W1

Z
< 1,4 b s1,4

; Z2 c

Z
< 2,4 b s2,4

Z
; 2,3 c s2,3 f3,4
f3

s2

W2

; Z3 c

Z
; 3,4 b s3,4
f4

s3

W3

< Z4 `

s4

In general, let f be the composite f1,n f1,n f1 and let s = s1,n sn1,n sn .
We would like to define by (a) = s1 f .
132

We need to check that (a) = s1 f is independent (up to ` -equivalence) of the

choices of the si,j and fi,j coming from the Ore axioms. For example, replace s1
1,2 f1,2 by
t1

g
.
We
have
f

s
=
s
f
and
g
s
=
t
f
.
Consider
the
right
fraction
s1,2 t1
1,2
1,2
1
1,2 2
1,2 1
1,2 2
1,2
1,2
and apply (LO2) to find a left fraction u1 v = s1,2 t1
.
Then
us
=
vt
.
Since
S
is
1,2
1,2
1,2
saturated, us1,2 = vt1,2 implies v S. Multiply g1,2 s1 = t1,2 f2 on the left by v. We get
vg1,2 s1 = vt1,2 f2
= us1,2 f2
= uf1,2 s1 .
By (LO3), there exists w S such that wvg1,2 = wuf1,2 . Hence
(s1,2 s2 )1 (f1,2 f1 ) ` (wus1,2 s2 )1 (wvf1,2 f1 )
` (wvt1,2 s2 )1 (wvg1,2 f1 )
` (t1,2 s2 )1 (g1,2 f1 ).
1 (tf ) does not change the value
It remains to check that replacing s1
i
i fi with (f si )
of , and also that and are actually inverses of each other. The proof is a (tedious)
calculation similar to the one weve done above. We leave it as an exercise to the interested
reader.

Lemma 1.4.6. Let C be an additive category, and S Mor(C) be an Ore family of

morphisms. Then C[S 1 ] is additive, and the localization functor Q : C C[S 1 ] is additive.
Proof. (sketch) We define an additive structure on HomC[S 1 ] (X, Y ) as follows. Given two
fractions s1 f, t1 g : X Y , replace s t1 by u1 v for u S, using (LO2). Let
w = vt = us S. Define
s1 f + t1 g = w1 (uf + vg).
It is a straightforward check that this induces a well-defined commutative group structure
on HomC[S 1 ] (X, Y ), that composition in C[S 1 ] is bilinear, and that Q is an additive functor,
i.e. it induces homomorphisms of abelian groups HomC (X, Y ) HomC[S 1 ] (QX, QY ).
We need to check that C[S 1 ] satisfies the axiom (AB3). Since C is additive, for
Y1 , Y2 Ob(C), there exists an object Y1 Y2 , together with canonical projections and
injections pi : Y1 Y2 Yi and ij : Yj Y1 Y2 . These satisfy pk i` = k` and i1 p1 +i2 p2 = id.
Now if we apply the additive functor Q, knowing that Q is identity on objects of C, we
obtain QY1 QY2 = Q(Y1 Y2 ) and all the structure morphisms pi , ii for the category
C[S 1 ]. This shows that C[S 1 ] satisfies (AB3), and so is additive.
One might wonder if localization preserves the property of being abelian. We will see
shortly that the answer is usually no.
133

Localization of abelian categories

2.1

Serre quotients

Let A be an abelian category, and S Mor A a saturated class of morphisms in A. We can

view the localization A[S 1 ] in two different ways:
A[S 1 ] is the localization of A obtained by adding more morphisms (i.e. inverses to
s S)
A[S 1 ] is obtained from A by quotienting out the subcategory of objects in A that
become the zero object in the localization
Morally, there is a formula A[S 1 ] = A/N . We will try to make this quotienting out
precise, and explain why the true. Our main references will be [Gab62], [Ste75] and [AZ94].
For the remainder of this section, assume that the abelian category A has arbitrary
colimits and limits, and moreover has a set of generators (see 2.2.3).
Definition 2.1.1. A (nonempty) full subcategory T of A is called a Serre (or dense)
subcategory if T satisfies: if 0 M 0 M M 00 0 is exact in A, then M Ob T if and
only if both M 0 , M 00 Ob(T ).
Traditionally, objects in T are called torsion objects.
Lemma 2.1.2. Serre subcategories are abelian subcategories.
Proof. Let T be a Serre subcategory of A. Since T is closed under subobjects and quotients.
Moreover, the inclusion T , A preserves kernels and cokernels. Checking the rest of the
definition is easy. For example, since T is closed under extensions, it is closed under finite
direct sums and finite products, and these coincide with the sums and products computed
in A.
Example 2.1.3. Let A = Vectk be the category of all vector spaces over a field k, and
T = Vectfd
k be the subcategory of finite-dimensional vector spaces.
Example 2.1.4. Let A be a commutative ring, S A a multiplicatively closed subset. Let
A = Mod(A). Call an A-module M S-torsion if for all m M , there exists s S such that
s m = 0. We can let T be the full subcategory of all S-torsion modules in A.
Example 2.1.5. Suppose A B Frac(A), and assume B is flat over A. Let A = Mod(A)
and T = {M Ob A : M A B = 0}.
Example 2.1.6. Let F : A B be an additive exact functor between abelian categories.
Put T = {M Ob(A) : F (M ) = 0} = Ker(F ).
134

Example 2.1.7. Let A be aL

graded connected k-algebra, where k is a field. Here, graded
connected means that A = n>0 An , and A0 = k. Let A = GrMod(A) be the category of
right graded A-modules. If M Ob(A), call m M torsion if for all n  0, mAn = 0. The
torsion elements in M form a graded submodule. Let (M ) be this submodule. We say that
M is torsion free if (M ) = 0. Let T = Tors(A) be the full subcategory of A consisting of
torsion A-modules. (Without further hypothesis, this is not a Serre subcategory.)
Lemma 2.1.8. Let A be a Noetherian graded connected k-algebra. Then Tors(A) is a dense
subcategory.
Proof. Use the following easy exercise: any graded connected Noetherian k-algebra is locally
finite, in the sense that each dimk An < . It follows that m (M ) if and only if
dimk (m A) < . Thus (M ) is the sum of all finite-dimensional submodules of M . Now
suppose 0 L M N 0 is an exact sequence of graded A-modules. First, we suppose
M Ob(T ) and show that L, N Ob(T ). But given the above discussion, this is obvious.
Conversely, suppose L, N are in T . Take m M and consider
0 mA L mA mA/(mA L) 0.
Since L is torsion, so is mA L, and since mA/(mA L) ' (mA + L)/L , N , we know
that mA/(mA L) is torsion. Since mA is finitely generated, we get that mA L is
finite-dimensional.
Definition 2.1.9. Let A be an abelian category, T a Serre subcategory. Define the quotient
A/T (Serre quotient) by Ob(A/T ) = Ob(A), and by letting
HomA/T (M, N ) = lim Hom(M 0 , N/N 0 ),

0
0
(M ,N )

the limit being taken over all pairs (M 0 , N 0 ) such that M 0 M and N 0 N satisfying
M/M 0 , N 0 Ob(T ).
Lemma 2.1.10. This definition of the Serre quotient makes sense.
Proof. Let I = {(M 0 , N 0 ) : M 0 M, N 0 N, M/M 0 , N 0 Ob(T )}. We put (M 0 , N 0 ) 6
(M 00 , N 00 ) if M 00 M 0 and N 0 N 00 . The set (I, 6) is directed because whenever
(M10 , N10 ), (M20 , N20 ) I, the pair (M10 M20 , N1 + N2 ) is also in I. The HomA (M 0 , N/N 0 )
form a directed system as follows. If i : M 00 , M 0 , j : N 0 , N 00 are injections, we get
Hom(M 0 , N/N 0 )

/ Hom(M 00 , N/N 0 )

/ Hom(M 00 , N/N 00 )

making {HomA (M 0 , N/N 0 )} a direct system.

Note that (M, 0) I because 0 = M/M Ob(T ). This gives us a map HomA (M, N )
limA (M 0 , N/N 0 ). So every morphism in A/T is represented by a morphism in A.

It is pretty straightforward to define the composition in A/T , we leave it as an exercise.

135

The definition 2.1.9 is motivated by the following theorem.

Theorem 2.1.11 (Serre). Let A be a graded quotient of a polynomial ring k[x1 , . . . , xn ]. Let
X = Proj(A) be the corresponding projective scheme over k. Let Qcoh(X) be the category
of quasi-coherent sheaves on X. Then there is a natural equivalence of categories
GrMod(A)/ Tors(A)

/ Qcoh(X).

Write Tails(A) for the quotient category appearing in this theorem. If M Tails(A) is
finitely-generated, then it turns out that one has HomTails (M, N ) = lim Hom(M/M>n , N )

for any N .
There is a projection functor : A A/T , that assigns to M A the same object M ,
and that sends f to the induced morphism f .
Lemma 2.1.12. An object M is zero in A/T if and only if it M Ob(T ).
Proof. If M = 0 in A/T , then HomA/T (M, M ) = 0, whence idM = 0. Thus there exist
id

M 0 , N 0 M such that the induced map M 0 , M

M M/N 0 is zero. It follows that
M is torsion.
Conversely, suppose M is torsion. It suffices to show that HomA/T (M, M ) = 0. But
every f : M N is represented by some f : M 0 N/N 0 for some pair (M 0 , N 0 ) I. If
M is torsion, then (M 0 , N 0 ) 6 (0, N 0 ), which means that f factors through the zero map.
Corollary 2.1.13. Let f : M N be a morphism in A. Then
1. Ker(f ) = (Ker f )
2. Coker(f ) = (Coker f )
3. f is monic if and only if Ker f Ob(T )
4. is epic if and only if Coker(f ) Ob(T )
5. f is an isomorphism if and only if Ker(f ), Coker(f ) are in Ob(T )
Theorem 2.1.14. Let A be an abelian category. Then : A/ A/T is exact.
Moreover, given any additive exact functor F : A C such that F (T ) = 0, there is a
unique functor Fe : A/T C such that
e = F.
One can check that if S T A are Serre subcategories, then we have (A/S)/(T /S) '
(A/T ). Given a short exact sequence 0 L M N 0 in A, there are isomorphisms
, , in A such that the following diagram commutes:
0

/L


/ L

/M


/ M

136

/N

/0

/ N

/0

Remark 2.1.15. If HomA/T (M, N ) is represented by f : M 0 N/N 0 for some

(M 0 , N 0 ) I, then let i : M 0 , M and p : N  N/N 0 be the canonical maps. By the
Corollary 2.1.13, the morphisms i and p are isomorphisms in A/T . So we can write
= (p)1 f (i)1 . Let S be the class of all morphisms in A with kernel and cokernel
both in T . Then we have the isomorphism A/T ' A[S 1 ].

2.2

Injective envelopes

Let A be an abelian category. Recall (see definition 3.1.1 in Chapter 4) that an object E in
A is injective if the functor Hom(, E) : A Ab is exact. Alternatively, arrows lift as in
the following diagram:
0

/M

/N

 ~

The category A has enough injectives if for any M in A, there is a monic M , E with E
injective.
For convenience we will think of the category A as a subcategory of modules over some
ring. Mitchells theorem 1.4.1 in Chapter 4 guaranties that we are not losing generality.
Definition 2.2.1. Let M N be objects in A. We say that M is essential in N if for all
N 0 M with N 0 6= 0, the intersection N 0 M 6= 0.
We say that a monic f : M N is essential if Im(f ) N is essential.
Theorem 2.2.2 (Eckmann-Schopf,[ES53]). An object E is injective if and only if E has no
proper essential extensions.
Definition 2.2.3. An injective envelope of an object M Ob(A) is any essential injective
extension of M .
Example 2.2.4. Let A be a simple Noetherian hereditary domain (for example, a Dedekind
domain or the Weil algebra khx, yi/([x, y] = 1)). Goldies theorem gives existence of the
field of fractions Q = Frac(A), the localization of A at the (Ore) set A r 0. It turns out that
A , Q is an injective envelope of A.
Theorem 2.2.5 (Baer,[Bae40](not sure if the correct reference)). Let E be an injective
envelope of M with respect to the inclusion i : M , E. Then
1. for any essential monic f : M N , there exists monic g : N , E such that gf = i
2. for any injective extension j : M , E 0 , there exists g : E E 0 such that j = gi.
Corollary 2.2.6. An injective envelope of M is a maximal essential extension of M .
137

Corollary 2.2.7. Given two injective envelopes i : M , E, i0 : M , E 0 , there is an

isomorphism f : E E 0 such that i0 f = i. Hence injective envelopes are unique up to
isomorphism, but not up to canonical isomorphism.
Remark 2.2.8. The category A may have enough injectives but not enough injective
envelopes. For example, in the category Mod(k[t]) , the module k[t] has no injective envelope
(because injective envelope in Mod(k[t]) is the same as projective envelope in Mod(k[t])).
On the other hand, Baer proved that any category of modules Mod(R) has enough injective
envelopes.

2.3

Localizing subcategories

The Serre subcategory T appearing above need not be closed under taking arbitrary (infinite)
direct sums. For example, one can consider Vectfd
k as a subcategory of Vectk .
Definition 2.3.1. Call a Serre subcategory T A localizing if the quotient functor :
A A/T has a right adjoint.
One traditionally denotes the adjoint by , and call the local section functor. We
will see later that this setup captures coherent sheaf cohomology.
Lemma 2.3.2. For a Serre subcategory T A, the following are equivalent:
1. every M Ob(A) has a largest torsion subobject
2. the inclusion i : T , A has a right adjoint (called the torsion functor)
3. the direct sum of torsion modules is torsion
Example 2.3.3. Let A = k[x0 , . . . , xn ] with the standard grading. Let A = GrMod(A) and
T = Tors(A). Recall we defined Tails(A) = A/T . Serres theorem tells L
us that Qcoh(Pnk ) '
0
n ) GrMod(A) is M 7
A/T . Geometrically, our adjoint : Qcoh(P
nZ H (X, M (n)),
L
which has right derived functors Ri ' nZ Hi (X, (n)). Along the same lines, induces
the local cohomology M 7 H0m (M ), where m A is the augmentation ideal.
Theorem 2.3.4. Assume A has injective envelopes, and that T A is closed under essential
extensions (so in particular T is closed under injective envelopes). Then the following are
equivalent:
1. T is a localizing subcategory
2. the inclusion functor T A ha a right adjoint (the torsion functor) : A T
In this case, (A, T ) is called a stable torsion pair.
Corollary 2.3.5. Under the assumptions of Theorem 2.3.4,
1. A/T has enough injectives (so both and have right derived functors)
138

2. ' idA/T (counit adjunction)

3. there is an exact sequence of functors
0

/ idA

/ R1

/ 0,

and Ri = Ri+1 for all i > 1.

Example: projective geometry
L
Let A =
n>0 An be a graded noetherian connected k-algebra. Let A be the category
of graded A-modules (or finitely-generated A-modules). For example, we could have A =
k[x0 , . . . , xn ] with the standard grading. Let T = Tors(A), which under our assumptions
consists of finite-dimensional modules. Recall we defined Tails(A) and tails(A) to be the
quotients of GrMod(A) and grmod(A) by Tors(A). If A is commutative and generated in
degree one, then a theorem of Serre tells us that for X = Proj(A), we have Qcoh(X) '
Tails(X) and coh(X) ' tails(A).
f, and M =
setting, : GrMod(A) Qcoh(X) is the functor M 7 M
L In this
0
i
i
n0 H (X, M (n)). We would like to identify the functors , R , and R .
Start with the exact sequence 0 A>n A A/A>n 0. For any graded A-module
M , applying the functor Hom(, M ) gives the following exact sequence
/ Hom(A/A>n , M )

/ Hom(A, M )

/ Hom(A>n , M )

/ Ext1 (A/A>n , M )

/0

If we now take limn , we get the following isomorphisms:

M = lim Hom(A/A>n , M );

Ri M = lim Exti (A/A>n , M );

L
M = n0 H0 (X, M (n));
L
Ri M = n0 Hi (X, M (n)).
It is a good exercise to derive carefully these isomorphisms.

3
3.1

Derived categories
Definition and basic examples

Let A be an abelian category. Consider the category Com(A) of complexes in A which is

also an abelian category (see Theorem 1.3.9 in Chapter 4). Put S = Qis Mor(Com(A)) to
be the family of all quasi-isomorphisms.
Definition 3.1.1. The derived category D(A) is the localization D(A) = Com(A)[S 1 ].
Denote Q : A D(A) the canonical localization functor.
139

Example 3.1.2. Lets call the complex (C , d ) cyclic if all di = 0. Let Com0 (A) be the
full subcategory of Com(A) consisting of the cyclicQ
complexes. Let i : Com0 (A) , Com(A)
0
be the canonical inclusion. Notice that Com (A) ' nZ A[n], where A just denotes a copy
of A that corresponds to n Z.
Define the functor H : Com(A) Com0 (A) in the opposit direction, which sends a
complex (C , d ) to the cyclic complex (H (C), 0). Notice that H sends quasi-isomorphisms
to isomorphisms.
Hence, by the universal property of localizations, H factors through Q
H

Com(A)
%

/ Com0 (A)
9

D(A)
Recall that an abelian category A is called semisimple if every short exact sequence in
A splits. Equivalently, A is semisimple if Exti (X, Y ) = 0, i > 0 and for any X, Y A. For
example, categories Vect of vector spaces and kGfdMod of finite dimensional representations
of a finite group G are semisimple.
is an equivalence of categories.
Theorem 3.1.3. If A is semisimple, then H
Proof. Define for any complex (C , d) Com(A) two morphisms
fC : (C , d) (H (C), 0)

: (H (C), 0) (C , d)
gC

as follows.
n1

/ Cn
/ . . . ] we put B n = Im(dn1 ), Z n = Ker dn and
/ C n1 d
For C = [ . . .
H n = B n /Z n for each n Z. Then we have the following exam=ct sequences

/ Zn

/ Cn

/ B n+1

/0

(5.1)

/0

(5.2)

Zn

Cn

B n+1

Since A is semisimple, for each n we can choose splittings of the above sequences. This
will give us C n ' Z n B n+1 ' B n H n B n+1 . With these identifications, the map
dn : C n C n+1 will be given by the map B n H n B n+1 B n+1 H n+1 B n+2 which
sends (bn , hn , bn+1 ) to (bn+1 , 0, 0).
as the canonical projection B n H n B n+1 H n
Then we can define the maps fC and gC
n
n
n
n+1
and embedding H , B H B
respectively.
0
We can define the functor L : Com (A) D(A) to be the composition
Com0 (A)

/ Com(A)

140

/ D(A)

 . Namely,
It is a straight forward calculation to show that L is mutually inverse to H
}
the families of morphisms {fC }CCom and {gC
CCom defined above induce isomorphisms of

functors f : id L H and g : H L id.

Remark 3.1.4. In general, D(A) captures much more information than just cohomology
functor. How much more? is a very delicate question. We call A and B derived equivalent if
D(A) ' D(B), i.e. their derived categories are equivalent (as triangulated categories).
For example, take A = Mod(A) and B = Mod(B). Define global dimension (which can
be ) of A as gldim(A) = n s.t. Extn (X, Y ) 6= 0 for some X, Y A, but Extn+1 (X, Y ) =
0, X, Y A. Then gldim is not a derived invariant, but the finiteness of gldim is.
Unfortunate fact: The class Qis is a saturated class, but it does not satisfy any other
Ore conditions.
Hence we cant really conclude anything about the derived category D(A). We dont
even know if it is additive. So we need a bit different construction of D(A). We will give a
construction using the homotopy category K(A).

3.2

The homotopy category

Definition 3.2.1. Let Hot Qis be the subclass of homotopy equivalences (see 1.0.8 on page
2). Then the homotopy category K(A) is defined as the localization K(A) = Com(A)[Hot1 ].
We denote by Qh the corresponding localization functor.
This category can be equivalently defined as follows. To fix the notation, if f, g : X Y
are homotopy equivalent morphisms of complexes, we write f h g. Now define the category
K0 (A) by Ob(K0 (A)) = Ob(Com(A)), and
HomK0 (A) (X, Y ) = HomCom(A) (X, Y )/ h
Key fact: Nilhomotopic morphisms form a two-sided ideal in complexes. It means, if
f h 0 then g f h 0 and f l h 0 for any morphisms g, l that are composable with f .
Lemma 3.2.2. The natural quotient functor h : Com(A) K0 (A) maps homotopy equiva : K(A) K0 (A). The induced functor
lences to isomorphisms, and so it induces a functor h
is an equivalence of categories.
h
The description
So from now on we will identify K(A) with K0 (A) via the functor h.
of the homotopy category as a quotient turns out to be more convenient than the
description K(A) as a localization. We will use the notation K(A) for the homotopy category.
K0 (A)

141

Lemma 3.2.3.
1. The localization functor Q : Com(A) D(A) factors through the
quotient functor h : Com(A) K(A):
Q

Com(A)

/ D(A)
;

Q0

K(A)
2. Functor Q0 factors through the the localization functor Qh
Q0

K(A)

/ D(A)
7

'

Qh

K(A)[h(Qis)1 ]
Moreover, the induced functor G is an equivalence of categories.
Proof. Since Qh h maps Qis to isomorphisms, there exists a functor F : D(A) K(A)[h(Qis)1 ]
s.t. Qh h ' F Q:
h

Com(A)

/ K(A) Qh / K(A)[h(Qis)1 ]
7
$

D(A)
We construct the inverse functor to F as follows. First observe that the functor
Q : Com(A) D(A) maps all nilhomotopic morphisms f to zero.
Indeed, suppose f : X Y , X, Y Com(A) and suppose f h 0. This means that there
is a map h : X Y of degree 1 s.t. f = dh + hd. Consider cone(idX ) = X X[1] and
define a map c : cone(idX ) Y by cn = (f n , hn+1 ). It is a straightforward calculation to
verify that since h is a homotopy, the map c will actually be a morphism of complexes, and
f factors through it.
f

/Y
:
c

cone(idX )
Applying the functor Q to this diagram gives Q(f ) = Q(c)Q(i). Now, since idX is obviously
a quasi-isomorphism, cone(idX ) is acyclic (see 3.2.3 on page 109), and so Q(cone(idX )) = 0.
Therefore, Q(f ) = 0.
But then Q must factor through h : Com(A) K(A), since h is universal among all the
functors L : Com(A) K(A) such that L(f ) = 0 whenever f h 0. Hence Q ' Q0 h.
142

This implies Q0 (h(Qis)) = Q(Qis) Iso. But if Q0 sends h(Qis) to isomorphisms, it must
factor through Qh . This implies that Q0 = G Qh for some functor G : K(A)[h(Qis)1 ]
D(A).
To summarize, we have obtained the following diagram:

/ K(A)

Com(A)


Q0

D(A) o

Qh

/ K(A)[h(Qis)1 ]

Hence we have G F Q ' G Qh h ' Q0 h ' Q. By the uniqueness of Q we get

G F ' id. Similarly, using the diagram above we get F G Qh ' F Q0 ' Qh , and so
F id.

3.3

Verdier theorem

To simplify the notations we will denote the image h(Qis) of Qis under the functor
h : Com(A) K(A) again by Qis.
Question (Yuri): Is h universal among all (not necessarily additive) functors F : Com(A)
C s.t. F (f ) = 0C whenever f h 0.
Theorem 3.3.1 (Verdier, [Ver96]). The class of morphisms Qis in K(A) is both left and
right Ore.
Proof. We want to check the Ore axioms 1.4.1 and 1.4.2. Saturatedness of the localizing
family of morphisms Qis is obvious.
First lets prove the axiom (LO2). Suppose we are given some maps t : W X and
g : W Y , t Qis. We want to find f : X Z and s : Y Z with s Qis.

>Z `
s

X`

>Y
g

W
We are thinking of morphisms in the homotopy category K(A) as honest morphisms of
complexes, since any morphism in K(A) is represented by some morphism in Com(A).
Define = t g : W X Y . Take Z = cone(). Then we have the following exact
sequence
W

/X Y

/ cone()

143

/ W [1]

(5.3)

where i h 0 (see 3.2 on page 109). Write i = (f ) s : X Y Z, where f and s

are defined by
f : X
s: Y

iX

/X Y

/ cone()

iY

/X Y

/ cone()

We claim that Z, f, s are exactly what we are looking for. So first of all we want to prove
that the diagram
>Z `
f

X`

>Y
g

W
commutes up to homotopy. We have 0 h i = (f, s) (t, g) = f t + sg, and so
sg h f t.
Next we want to show that s is actually a quasi-isomorphism. To see that, first note
that the exact sequence 5.3 induces the following long exact sequence
...

/ Hk (W )

Hk ()

/ Hk (X) Hk (Y )

/ Hk (Z)

/ Hk+1 (W )

/ ...

Since = (t, g) and t is a quasi-isomorphism, the map Hk () must be injective. Hence

Hk (Z) Hk+1 (W ) is the zero map. Thus the long exact sequence splits into short exact
sequences
/ Hk (W )
/ Hk (X) Hk (Y )
/ Hk (Z)
/0
0
This follows immediately from the definitions that exactness of such short exact sequence
is equivalent to the square
Hk (W )
Hk (g)

Hk (Y )

Hk (t)

Hk (s)

/ Hk (X)


Hk (f )

/ Hk (Z)

being cartesian and cocartesian. Then since Hk (t) is an isomorphism, Hk (s) also must be
an isomorphism. So the map s is a quasi-isomorphism, and we have proved the axiom
(LO2).

144

Chapter 6

Triangulated categories
1

The basics

1.1

Definitions

Definition 1.1.1. A triangulated category D is an additive category equipped with extra

structure:
an auto-equivalence [1] : D D, which we write X 7 X[1], and call the shift (or
suspension) functor
u

a class E of diagrams of the form X

Y
Z
X[1], called exact (or distinguished)
triangles
Write
/Y

X`
[1]

for such a distinguished triangle.

The class E is required to satisfy the following axioms.
1

TR1 (a) For all X in D, the triangle 0 X

X 0[1] is exact;
(b) The class E is closed under isomorphisms of triangles;
(c) Every morphism u : X Y in D can be completed to an exact triangle in E:
X

/Y

/W

/ X[1] .
u

TR2 The class E is closed under shifts in the sense that if X

Y
Z
X[1] is in E,
v

u(1)

then so is Y
Z
X[1] Y [1].
145

TR3 Given a commutative diagram

/Y

e
/Y

u
e

e
X

/Z

e
/Z

v
e

/ X[1]

w
e

f [1]

e
/ X[1]

e making the rest of the diagram commute. Note that h is not

there exists h : Z Z
functorial.
u

TR4 (Octahedron axiom). Given two exact triangles X

Y
Z
X[1], Y
Ye

W
Y [1], we can construct the following commutative diagram
X

id

e
u

/Y

/Z

e
/Y

e
/Z

e
v

id[1]

/ X[1]

/W

Y [1]

idW

/ X[1]

v[1]

/ Z[1].

such that rows and columns are exact.

1.2

Examples of triangulated categories

Example 1.2.1. Let D = Vectk be the category of vector spaces over a field k. Put the
shift functor [1] to be just [1] = idD . The class E of exact triangles is defined by the following
rule:
u
v
w
E = {U V
V W
W U
U V}
for any triple U, V, W of vector spaces. The maps u, v, w are
u = U V  V , V W
v = V W  W , W U
w = W U  U , U V
It is a good exercise to check that this actually gives Vectk the structure of a triangulated
category.
146

Example 1.2.2. The standard examples are K (A) and D (A), for A an abelian category
and {, b, +, }. In these examples, [1] : C 7 C is the usual shift of complexes. The
exact triangles are diagrams isomorphic to
X

/Y

iu

/ cone(u)

pu

/ X[1]

We will postpone verification that this actually is a triangulated structure to the next section.
Example: vector bundles on projective spaces
This example is due to [BGG78], and is exposited in [OSS11]. Let k be V
a field of characteristic
L
Vi
not 2. Let E be a fixed E-vector space of dimension n + 1. Let = k (E) = n+1
(E)
i=0
be the Z-graded exterior algebra of E. Let M() be the category of graded left -modules,
and let Mb () be the full subcategory of graded modules which are finite-dimensional over k.
Let F Mb () be the full subcategory of free modules over . It turns out that here F
actually coincides with both subcategories of projective and injective modules.
Lets call f : V V 0 in Mb () equivalent to zero if f = V F V 0 for some
F Ob(F). Obviously the class I of morphisms in Mb () which are equivalent to zero is a
b
b
two-sided ideal. Define the stable module category M () to be M () = Mb ()/I. Objects
b
of M () are the same as those in Mb (), and hom-sets are
HomMb () (X, Y ) = Hom (X, Y )/I.
Theorem 1.2.3 (Beilinson, Gelfand, Gelfand). There is a natural structure of a triangulated
b
category on M () in which the shift functor [1] is defined by
V 7 ((n) k V )/i(V )(n)
where W (m) =

iZ W

im ,

(n = dim E 1),

and i(V ) = n+1 (E) V V .

I what follows we will try to explain where this triangulated structure comes from.
Observe that every graded -module V can be viewed as a family of complexes
Le (V ) =

i1
/ V i1 de / V i

die

/ V i+1

parameterized by e E. The fact that e2 = 0 in implies that de (v) = e v is a differential.

Notice that if ee = e for k , then Lee(V ) ' Le (V ). Thus we have an operation
V 7 {Le (V )}eP(E) . Lets re-interpret this using algebraic geometry.
Call a complex of quasi-coherent sheaves on P(E) rigid if it has the form
L = V i OP (i) V i+1 OP (i + 1) .
Let Rig be the full subcategory of Com(Qcoh(P)) consisting of rigid complexes. Note that
the {Le (V )}eP(E) induces a complex L in Rig.
147

In the opposite direction, given a rigid complex L , define V (L ) M() by the rule
V (L ) =

V (L )i =

iZ

P(E), L i (i) .

iZ

We need to give V (L ) the structure of a -module. Recall that

canonically isomorphic to Symk (E ). Consider a map

iZ (P, OP (i))

is

a = (P, diL (i)) : V (L )i V (L )i+1 (P, O(1)) ' V i+1 E .

This allows us to give V (L ) a -module structure by
e v = (1)i (id se )a(v)
where se : E k is ` 7 `(e), e E and v V (L )i .
i
2
It is a straightforward verification that di+1
L ddL = 0 implies that e v = 0. So we get a
graded -module.
Lemma 1.2.4. Two functors defined above by V 7 {Le (V )}eP(E) and L 7 V (L ) are
inverse to each other, and thus give the equivalence of categories Rig ' M().
We call a rigid complex L finite if L i = 0 for |i| >> 0 and dim(L i ) < . Denote by
Rig the full subcategory of Rig of finite complexes.
f

Lemma 1.2.5. The equivalence from Lemma 1.2.4 restricts to the equivalence Rigf ' Mb ().

Consider now the functor : Mb () Rigf Db (coh(P(E))). Be the Theorem 1.2.3 it

b
factors through M ().
b

Theorem 1.2.6. Functor induces an equivalence of triangulated categories M ()

Db (coh(P(E))).
b

So the twisted triangulated structure on M () comes from the usual triangulated

structure on Db (coh(P(E))) when being induced by the functor .

1.3

Basic properties of triangulated categories

Recall that a triangulated category consists of an additive category D together with an

autoequivalence [1] and a class E of exact triangles satisfying some axioms appearing in
Definition 1.1.1. We are interested in developing some basic consequences of these axioms.
For any triangulated category D, there is (by definition) a quasi-inverse [1] : D D
to [1]. For all X D, we have natural isomorphisms X[1][1] ' X ' X[1][1]. Replacing
D by an equivalent category, we may assume [1] and [1] are inverses on the nose, i.e.
X[1][1] = X = X[1][1]. More precisely, we have the following lemma.
148

Lemma 1.3.1. Let D be a category with an auto-equivalence [1] : D D. Then there exists
f:D
e with an automorphism [1]
eD
e and F : D D
e such that the following
a category D
diagram commutes:
D


[1]

/D

e
D

f
[1]

e
/D

e and F is an equivalence of triangulated categories.

Moreover, if D is triangulated, so is D,
e directly. It has as objects sequences (Xn , n )nZ where Xn Ob(D)
Proof. We construct D

and n : Xn [1]
Xn+1 . We define HomDe ((Xn , n )n , (Yn , n )n ) to be the set of tuples
(fn HomD (Xn , Yn ))n such that for all n, the following diagram commutes:
Xn [1]


fn [1]

/ Yn [1]

Xn+1

fn+1

/ Yn+1

f = (Xn+1 , n+1 )nZ .

e is defined by (Xn , n )nZ [1]
The shift automorphism in D
e D by (Xn , n ) 7 X0 . It is a direct calculation to check that this
Define F 1 : D
e satisfies the required properties.
indeed defines an inverse to F , and that the category D
In light of this lemma, we will henceforth assume that the shift functors on our triangulated categories are automorphisms.
Using the automorphisms [1], [1] and the axiom TR2, we can extend every exact
triangle to a helix
X[1] Y [1] Z[1] X X Y Z X[1]
Another way to draw this helix is
T
w[1]

u[2]
v[1]

T
w

u[1]
v

149

A morphism of exact triangles extends to a double helix, and such a double helix is
uniquely determined (by TR3) up to isomorphism by any of the two arrows in the exact
triangle.
For an integer n > 0, let [n] be the n-fold composition of [1], and for n 6 0, let [n] be
the n-fold composition of [1]. Also, to simplify notation we write (X, Y ) = HomD (X, Y ).
u

Proposition 1.3.2. Let X

Y
Z
X[1] be an exact triangle. Then for all U Ob(D),
the following two sequences of abelian groups are exact:

/ (U, Z[1])

/ (Y [1], U )

/ (U, X) u / (U, Y ) v / (U, Z) w / (U, X[1])

/ (X[1], U ) w / (Z, U ) v / (Y, U ) u / (X, U )

/ (U, Y [1])

/ (Z[1], U )

Proof. It suffices to prove that the top sequence is exact at (U, Y ). First we show that
u v = 0. For f : U X, define g = u f = f u, and consider the following diagram.
/0

/Y

/ U [1]
h

/Z

f [1]

/ X[1]

The first row is exact by TR1 and TR2. By TR3, there exists h : 0 Z such that
vg = h 0 = 0, which implies v u f = 0 for all f .
Now assume v g = 0 for some g : U Y and consider the following diagram with exact
rows:
/ U [1] id[1] / U [1]

/0

U
g

/Z

fe

/ X[1]

u[1]

g[1]

/ Y [1]

By TR3, there exists fe : U [1] X[1] such that g[1] = u[1] fe. Applying [1], we see
that g = u fe[1], hence g is in the image of u . This proves exactness of the first sequence.
For the second one the proof is similar.
Corollary 1.3.3 (5-lemma). If the morphisms f, g in axiom TR3 are both isomorphisms,
then so is h.
Proof. Fix some extension h, and choose an arbitrary U Ob(D). Consider the following
sequence (notation as in Proposition 1.3.2):
(U, X)


e
(U, X)

u
e

/ (U, Y )


e)
/ (U, Y

v
e

/ (U, Z) w / (U, X[1]) u[1]/ (U, Y [1])



f [1]

g[1]

u
e[1]
e we / (U, X[1])
e
e [1])
/ (U, Z)
/ (U, Y

150

By Proposition 1.3.2, both rows are exact, so by the usual 5-Lemma for abelian groups,
e is an isomorphism for all U . By the Yoneda Lemma, h : Z Ze is also
h : (U, Z) (U, Z)
an isomorphism.
u

Corollary 1.3.4. In any exact triangle X

Y
Z
X[1], we must have
vu = 0
wv = 0
u[1]w = 0
Proof. The proof is exactly as in that of Corollary 1.3.3. One sees that v u = 0 relative to
all U Ob(D), whence vu = 0 by the Yoneda Lemma. The other parts are similar.
Corollary 1.3.5. If we apply TR1 to u : X Y , then the exact triangle completing u is
determined up to isomorphism.
Proof. Apply Proposition 1.3.2 to (f, g) = (idX , idY ) as in
X

/Y

/Y

u
v
w
u
e
v
e e w
e e
e
Corollary 1.3.6. Let X
Y
Z
X[1] and X
Ye
Z

X[1]
be exact triangles.
e
Assume there exists g : Y Y such that vegu = 0. Then there exists f and h making the
following diagram commute:

X


/Z

e
X

/Y

u
e

e
/Y

v
e

e
/Z

w
e

/ X[1]


f [1]

e
/ X[1]

e
Moreover, if HomD (X, Z[1])
= 0, then f and h are uniquely determined b g.
e yields
Proof. Consider gu : X Ye . Since ve (gu) = 0, Proposition 1.3.2 applied to U = X
e such that gu = u
f :XX
e (f ) = u
e f . Similarly one shows that h exists. The uniqueness
of f and h follows from the long exact sequence.
Remark 1.3.7. For any u : X Y , the axioms TR1 and TR3 yield an exact triangle
u
X
Y Z X[1]. In analogy to the situation with categories of complexes, we write
Z = cone(f ). We could even hope that cone : Mor(D) D is a functor. Unfortunately,
in general one does not always have cone(u v) = cone(u) cone(v). In other words, for
arbitrary triangulated categories, the cone construction is not functorial. Non-functoriality
of cone causes a lot of inconvenience, and later we will see possibilities how to deal with
that.
151

2
2.1

Further properties
Abstract cone and octahedron axiom

Let D be a triangulated category. If u : X Y is a morphism in D, then TR1 shows

u
v
w
that X Y can be extended to an exact triangle X
Y
C(u)
X[1]. Moreover, u
determines C(u) uniquely (up to isomorphism). We call C(u) an abstract cone of u. The
axiom TR3 tells us that C() behaves almost functorially. Define C : Mor(D) D by
u
u
u
e
e
(X
Y ) 7 C(u). For (f, g) : (X
Y ) (X
Ye ), put C(f, g) = h, where h makes the
following diagram commute:
X


/Y

/Z

e
X

e
/Y

u
e

v
e

/ X[1]

e
/Z

w
e

f [1]

e
/ X[1]

The problem is that C(u2 u1 ) = C(u2 ) C(u1 ) does not follow from the axiomatics. What
can we say about C(u2 u1 ) just using the axioms for a triangulated category in particular,
TR4?
Suppose we are given two morphisms u1 : X Y and u2 : Y Z. We can complete
u1
v1
w1
u2
v2
w2
these to exact triangles X
Y
C(u1 )
X[1] and Y
Z
C(u2 )
Y [1].
v1 [1]

2
Define w to be the composite C(u2 )
Y [1] C(u1 )[1].


w
Lemma 2.1.1. C(u2 u1 ) ' C C(u2 )
C(u1 )[1] [1].

Proof. This is a consequence of Verdiers octahedron axiom. The statement is equivalent to

the existence of an exact triangle
C(u1 ) C(u2 u1 ) C(u2 ) C(u1 )[1].
One shows that such an exact triangle exists via a direct application of TR4. Simply consider
the diagram
u1
v1
/Y
/ X[1]
/ C(u1 )
X
u2

u2 u1

/Z

/ C(u2 u1 )

C(u2 )

C(u2 )

Y [1]

v1 [1]


/ C(u2 )[1]

152

/ X[1]

1
2
Corollary 2.1.2. In D(A), any X
Y
Z gives rise to

cone(u2 u1 ) = cone (cone(u2 ) cone(u1 )[1]) [1].

In the special case when u1 , u2 are embeddings, the lemma is just the second isomorphism
theorem: Z/Y ' (Z/X)/(Y /X).

2.2

The homotopy category is triangulated (need proofs!!)

The point of this section is to prove the following theorem.

Theorem 2.2.1. Let A be an abelian category. Then K(A) is triangulated.
One puts [1] : K(A) K(A) by (X[1])i = X i+1 and dnX[1] = (1)n dnX . We say that a
u

diagram X
Y
Z
X[1] is exact in K(A) if and only if it is isomorphic to a diagram
of the form
pu
iu /
u /
/ X[1].
X
Y
cone(u)
u

Y
Z 0 is an exact sequence in Com(A) such
Lemma 2.2.2. Suppose 0 X
n
that for every n, the sequence 0 X Y n Z n 0 splits (this is not the same as
0 X Y Z 0 splitting). Let sn : Z n Y n and pn : Y n X n be the right
(resp. left) inverses of v n (resp. un ). Define w : Z X[1] by wn = pn+1 dnY . Then
u
v
w
X
Y
Z
X[1] is an exact triangle in K(A).
Lemma 2.2.3. Every exact triangle in K(A) is isomorphic to one as in Lemma 2.2.2.
Lemma 2.2.4. The axioms of a triangulated category hold for triangles as in Lemma 2.2.2.

2.3

Localization of triangulated categories

Let D be a triangulated category and S Mor(D). Let D[S 1 ] be the (abstract) localization
of D at S. We know that if S is Ore, then D[S 1 ] is additive. We are interested in extra
conditions on S that force D[S 1 ] to be triangulated.
Verdier found such conditions. Assume that S, in addition to being Ore, satisfies the
following two properties:
O4 S is closed under [1], i.e. s S if and only if s[1] S.
O5 If f, g S are as in TR3, then h can be chosen to be in S.
Proposition 2.3.1. If S satisfies O1-O5, then D[S 1 ] has a natural triangulated structure.
Proof. Recall that if S is Ore, then morphisms in D[S 1 ] can be represented by equivalence
classes of left S-fractions. It suffices to define the shift functor on these fractions. Put
(s1 f )[1] = s[1]1 f [1]. Say that a triangle in D[S 1 ] is exact if and only if it is isomorphic
to the image of an exact triangle in D.
All that remains is to check the axioms TR3 and TR4. This is pretty tedious direct
verification.
153

Theorem 2.3.2. Let A be an abelian category. Then the derived category D(A) is triangulated.
Proof. First we use the fact that D(A) ' K(A)[Qis1 ]. Next, check that Qis in K(A) satisfies
O1-O5. The axiom O4 is obvious because Hk (S[1]) = Hk+1 (S). On the other hand, showing
O5 requires some use of TR4. We will omit this part.

2.4

Exact functors

Definition 2.4.1. If D, D0 are triangulated categories and F : D D0 is additive, then F

is called exact (or sometimes triangulated or even triangular) if F respects the triangulated
structures, i.e. if

1. there exists : F [1]

[1] F
2. F sends distinguished triangles ED in D to distinguished triangles ED0 in D0 .
Lemma 2.4.2. Any additive functor F : A B between abelian categories induces an
exact functor F : K(A) K(B).
Proof. This is trivial F preserves the cone construction.
Lemma 2.4.3. If F : D D0 is exact and G : D0 D is an adjoint (either left or right)
to F , then G must also be exact.
Now we want to see why exact sequences give rise to exact triangles. Recall first
that for f : X Y is a morphism in Com(A) we defined cone(f ) = X[1] Y and
cyl(f ) = X X[1] Y , with appropriate twisted differentials, see subsection 3.2.
Lemma 2.4.4. For any f : X Y in Com(A), define : cyl(f ) Y by (xk , xk+1 , y k ) =
f (xk ) + yk. Then there exists a commutative diagram

/ Y
_

/X


/ cyl(f )
f

/ cone(f )

/ X[1]

/ cone(f )

/0

/0

/Y

such that the rows are exact and , are homotopy equivalences. Moreover, = idY and
h idcyl(f ) .
Corollary 2.4.5. Every exact triangle in K(A) (and also D(A)) is isomorphic to one of
the form
X

/ cyl(f )

/ cone(f )

154

pf

/ X[1]

Proof. Just use the commutative diagram

f

/Y
O

/ cone(f )

/ X[1]

/ cone(f )

/ X[1]

/ cyl(Y )

Proposition 2.4.6. Every short exact sequence in Com(A) can be completed to an exact
triangle in D(A). More precisely, every short exact sequence of complexes 0 X Y
Z 0 is quasi-isomorphic to
X

if

/ cyl(f )

pf

/ cone(f )

/ X[1].

Proof. Consider the diagram

0

/X

/Y
O

/X

/0

/Z
O

/ cyl(f )

/ cone(f )

/0

Define : cone(f ) Z by (xk+1 , y k ) := g(y k ). One checks that the diagram above
commutes in Com(A), and that is a quasi-isomorphism. Indeed, since g is surjective, so is
, and Ker() = X[1] Ker(g) = X[1] Im(f ) ' X[1] X. If we show that the kernel of
is acyclic, it will imply that H () is an isomorphism. Now, Ker() is acyclic because one
can exhibit an explicit homotopy idKer() h 0. The homotopy h : Ker() Ker()[1] is
given by (xk+1 , xk ) 7 (xk , 0).
As a conclusion of the results above, we might forget about short exact sequences and
work instead with exact triangles.

2.5

Verdier quotients

Let D be a triangulated category, N D a subcategory.

Definition 2.5.1. The subcategory N is called a triangulated subcategory of D if the
triangulated structure of D induces one on N (in the obvious sense), and if the inclusion
functor N , D is exact.
Lemma 2.5.2. Assume N D is strictly full (i.e. closed under isomorphisms). Then N
is a triangulated subcategory if and only if N is closed under [1]D and taking cones in D.
Sometimes one calls N thick or epaisse.
155

Example 2.5.3. Let N = K (A), where {, +, b}. This is a triangulated subcategory

of the full derived category K(A).
Definition 2.5.4. Let D be a triangulated category, N a triangulated subcategory. The
Verdier quotient of D by N is a pair (D/N , Q : D D/N ), where D/N is triangulated,
Q : D D/N is exact, such that
1. Q(N ) = 0
2. Q is universal among exact functors which kill N .
Definition 2.5.5. Given (D, N ) as above, define S(N ) to be the class of morphisms s :
s
X Y in D such that in any exact triangle X
Y Z X[1], we have Z Ob(N ).
Theorem 2.5.6 (Verdier). The class S(N ) satisfies Ore conditions O1-O3 and is compatible
with the triangulated structure of D (i.e. it satisfies O4-O5). The localization (Q : D
D[S 1 ] is a quotient D/N .
Example 2.5.7. Recall that one way to define D(A) is as the localization of K(A) with respect to quasi-isomorphisms. One could also let D = K(A), and let N be the full subcategory
consisting of acyclic complexes. In that case, S(N ) = Qis, so D(A) = K(A)/{acyclic}.
Example 2.5.8 (Singularity category). Let A = mod(A), or more generally A = coh(X) for
X a projective variety. Consider the subcategory Perf(A) Com(A) consisting of perfect
complexes, i.e. complexes isomorphic to P i P i+1 where P i = 0 for |i|  0,
and the P i are locally free sheaves of finite rank. If A is a regular ring (or X is a smooth
variety), then it is well-known that Perf(A) = Db (A). On the other hand, for A = k[x]/x2
and A = Mod(A), one calls Dsing (A) = D(A)/Perf(A) the singularity category of k[x]/x2 .
For our example, Dsing (k[x]/x2 ) is the category of vector spaces with a previously described
triangulated structure, see example 1.2.1.

2.6

Exact categories

Exact categories were created by Quillen in [Qui73]. Let A be an abelian category, B A

a full additive subcategory. We assume that B is closed under extensions in A. In other
words, if 0 X Y Z 0 is exact in A and X, Z B, then Y B.
Definition 2.6.1. An exact category is a pair (B, E) where E consists of diagrams X
Y Z such that 0 X Y Z 0 is exact in some ambient abelian category A.
Remark 2.6.2. Quillen defined exact categories axiomatically. Given an exact category
(B, E), we can always canonically construct A as a certain subcategory of Funadd , B , Ab).
Example 2.6.3. Let B = A, with E the class of all short exact sequences. Then A is exact.
Example 2.6.4. Let A = Mod(A) for A some associative unital ring. The subcategory B
of projective / flat / free / injective etc. A-modules, is exact.
156

Example 2.6.5. Let B be any additive category. Then B is exact if we let E consist of split
exact sequences X X Y Y .
i

Y
Z
Example 2.6.6. Let A be an abelian category, B = Com(A). Let E consist of X
in

pn

such that for all n, the exact sequence 0 X n Y n Z n 0 splits.

Let (B, E) be an exact category.
Definition 2.6.7. An object I Ob(B) is called E-injective if for all X Y Z in E,
the sequence
HomB (Y, I) HomB (X, I) 0
is exact. Equivalently, I is E-injective if any I X Y in E splits.
Dually, P Ob(B) is E-projective if whenever X Y Z is in E, then
HomB (P, Y ) HomB (P, Z) 0
is exact. Equivalently, all X Y P in E split in B.
Definition 2.6.8. We say that B has enough injectives if for every X Ob(B), there is a
p
i
diagram of the form X
IX
SX in E, with IX injective. Similarly we say that B has
enough projectives if all X Ob(B), there is SX P X X in E with P X projective.
Definition 2.6.9. An exact category (B, E) is called Frobenius if
1. B has enough E-injectives and E-projectives
2. E-injectives and E-projectives coincide
Example 2.6.10. Let A be an associative unital ring, A = Mod(A), and E the class of
exact sequences. The condition 1 always holds, but condition 2 does not hold in general.
But if, for example, A = k[G] for G a finite group and k characteristic zero, then Mod(k[G])
is Frobenius.
V
Example 2.6.11. Let A = (E) for E a finite-dimensional k-vector space. Then A =
Mb (A) is Frobenius.
Example 2.6.12. Let B = Com(A) for an abelian category A. Let E be the class of
term-wise-split exact sequences. For X Ob(B), define IX by (IX)n = X n X n+1 , with
differential dnIX (xn , xn+1 ) = (xn+1 , 0). Define SX by (SX)n = X n+1 with dnSX = dn+1
X .
We now define iX : X IX and pX : IX SX. Put iX (xn ) = (xn , dnX xn ) and
pX (xn , xn+1 ) = dnx xn + xn+1 .
First one checks that IX is injective in (B, E). Next, one shows that iX splits if and only
if X h 0. So an object Z Ob(B) is E-injective if and only if Z h 0. Moreover, IX is
also E-projective, so (Com(A), E) is a Frobenius category.
157

Definition 2.6.13. Let B be a Frobenius category, and let X, Y Ob(B). Define I(X, Y )
g

to be the set of f HomB (X, Y ) such that there exists a factorization f = X

Z
Y,
and Z is projective (hence also injective).
Definition 2.6.14. The stable category of B, denoted B, is the quotient B/I. Objects in B
are objects of B, and
HomB (X, Y ) = HomB (X, Y )/I(X, Y ).
Example 2.6.15. Let B = (Com(A), E) for A an abelian category. Then B = K(A).
Theorem 2.6.16. The stable category of any Frobenius category has a canonical triangulated
structure.
Proof. First we define the suspension functor in B. Given X Ob(B) and injective
resolutions X I Y and X I 0 Y 0 , there is an extension of idX to a commutative
diagram
/I
/Y
X
u

/ I0

/Y0

where v is not unique on the nose, but any two v are the same in HomB (Y, Y 0 ). This allows
u

Su

us to define the suspension functor S : B B by X 7 SX and X

Y 7 SX SY .
One has to check that a) this is a well-defined functor, and that b) S is an equivalence of
categories.
We need to define exact triangles in B. For u : X Y , consider
X


/ IX

/ SX

/C

!w

/ SX

with the leftmost square co-cartesian. There exists a unique w : C SX making the
diagram commute. Exact triangles in B are exactly those of the form X Y C SX
as above.
Let (B, E) be an exact category such that B has enough injectives. Let IB be the class
of injectives in B. Recall that we defined the stable category B, whose objects are objects
in B, and has morphisms Mor(B)/I, where I consists of morphisms which factor through
injectives. The category B is additive, but not exact.
For every Ob(B), choose SX Ob(B) such that there is X IX SX in E with
IX in IB . Given
X

i0

/I


/ I0

158

p0

/Y


/Y0

the injectivity of I 0 yields u making the diagram commute. Exactness gives v making the
diagram commute. Suppose we also have extensions u
e and ve of idX : X X.
Lemma 2.6.17. Given two extensions (u, v), (e
u, ve), we have v = ve in HomB (Y, Y 0 ).
Proof. By commutativity of the diagram, we get ui = i0 = u
ei, whence (u u
e)i = 0. Similarly
(e
v v)p = 0. The first tells us that Ker(u u
e) Im(i) = Ker(p), whence u u
e factors
through I/ Ker(p) as in
p

I/ Ker(p)


/Y

ue
u

p0

I0

ve
v

/Y

We have v ve = p0 (u u
e)p1 , whence the result.
Corollary 2.6.18. The assignment X 7 SX extends to an endofunctor S : B B.
Proof. Suppose we have f : X Y . Then as above, we have an extension
/ IX

X


/ SX

/ IY

/ SY

As above, one sees that h is independent of the choice of g.

u

The functor S is called the suspension functor. Given X

Y
Z in E, we can consider
X

iX

/Y


/ IX

pX

/Z


/ SX

Definition 2.6.19. A triangle X Y Z SX in B is called exact if it is isomorphic

u
v
w
to one of the form X
Y
Z
SX as above.
Note that this defines a functor from the category of diagrams E to the category of
S-triangles. Denote by E the full image of this functor.
Theorem 2.6.20. The triple (B, S, E) is a suspended category, in the sense that all axioms
TR0-TR4 hold, except that S need not be an equivalence. If B is Frobenius, then S is an
equivalence.
Proof sketch. Assume B is Frobenius. Define (I) to be the category of all (unbounded)
acyclic complexes with terms in I. Morphisms in (I) are homotopy-classes of morphisms in
Com(B). We claim that B ' K(I). The corresponding functors are defined are defined as
159

d0

follows. First : K(I) B sends X to Ker(X 0 X 1 ), and : B K(I) sends X to a

complex
P 2 P 1 P 0 I 0 I 1 I 2 ,

P
I
where P
X is a projectives resolution, X
I 0 , and P 0 I 0 is P I . It is easy to
check that [1] ' S , and that sends exact triangles to exact triangles.

Example 2.6.21 (Tate cohomology). Let G be a finite group. Let k be a commutative

ring, A = k[G], and B = mod(A). This is a Frobenius category. Define
b i (G, M ) = HomB (k, S i M )
H
This is the i-th Tate cohomology of G with coefficients in M .

t-structures and the recollement

3.1

Motivation and definition

The same triangulated categories can be identified with the derived categories of completely
different abelian categories. In other words, a triangulated category D can be equivalent to
D(A) and D(B) for unrelated abelian categories A, B. We would like to have some kind of
extra structure that allows us to recover A and B from D. More generally, we would like to
be able to catch abelian subcategories of a triangulated category D. Our main example is
D = Mod(DX ), where X is a variety and DX is the sheaf of differential operators on X.
Example 3.1.1. Let A be an abelian category, D = D (A) for {, +, , b}. Define
D>n (A) = {X D : Hi (X) = 0 for all i < n}
D6n (A) = {X D : Hi (X) = 0 for all i > n}.
The natural inclusion i : A D sending X to 0 X 0 is fully faithful, and
Im(i) = D60 (A) D>0 (A). The general notion of a t-structure is an axiomatization of this
situation.
Definition 3.1.2. Let D be a triangulated category. A t-structure on D is a pair (D60 , D>0 )
of strictly full subcategories satisfying
TS1 D60 D61 and D>1 D>0
TS2 For all X Ob(D60 ) and Y Ob(D>1 ), HomD (X, Y ) = 0.
TS3 For all X Ob(D), there is an exact triangle A X B A[1] in D with
A Ob(D60 ) and B Ob(D>1 ).
Here, we are using the standard notation
D6n = D60 [n]
D>n = D>0 [n].
160

3.2

The core of a t-structure and truncation functors

Definition 3.2.1. Let D be a triangulated category with t-structure (D60 , D>0 ). The core
of this t-structure is the full subcategory D60 D>0 . We will sometimes denote it by D .
Lemma 3.2.2. The core of the standard t-structure on D (A) is A.
Theorem 3.2.3. The core of any t-structure is an abelian category.
During the rest of this subsection we will sketch the main part of the proof. The proof
(which follows the original paper [BBD82]) is based on the following key lemma.
Lemma 3.2.4. Let D be a triangulated category equipped with a t-structure (D60 , D>0 ).
1. For each n Z, the inclusion functors
i6n : D6n D
i>n : D>n D
have right (resp. left) adjoints
6n : D D6n
>n : D D>n
called truncation functors.
2. For all X Ob(D), there is an exact triangle of the form
60 (X) X >1 (X) 60 (X)[1].
Moreover, any two triangles A X B A[1] with A Ob(D60 ) and B Ob(D>1 )
are canonically isomorphic.
Proof. Lets prove the existence of 60 and >1 . The proofs for general n are similar. Using
the axioms for a t-structure, for all X Ob(D), we can choose A X B A[1] with
A D60 and B D>1 . Define
>0 (X) = A
>1 (X) = B
f

Given f : X Y in Mor(D), lets show that the A X

Y in the following diagram
factors uniquely through A0 :
A

B 0 [1]

fe

/ A0

/X
f

/Y

161

/B

/ A[1]


/ B0


/ A0 [1]

Applying functor HomD (A, ) we get a long exact sequence

HomD (A, B 0 [1]) HomD (A, A0 ) HomD (A, Y ) HomD (A, B 0 ) . . .
By the axiom TS2 we have HomD (A, B 0 ) = 0 and HomD (A, B 0 [1]) = 0. The uniqueness of
fe is now clear. Thus we can define >0 (f ) = fe. A similar argument shows that an extension
of f to B B 0 is unique. Our argument gives isomorphisms of functors:
HomD60 (A, 60 (Y )) = HomD (i60 A, Y )
hence the adjunction.
Start with the case D = D (A). We have 6n X = 0 if and only if the adjunction
morphism X >n+1 X is an isomorphism. Similarly, >n X = 0 if and only if 6n1 X X
is an isomorphism.
If m 6 n, there are natural isomorphisms

6m
6n 6m

>n
>n >m

>m 6n
6n >m
we write [m,n] for the last functor.
We can now use these relations to construct kernels and cokernels in the core A = D . Let
f
f : X Y be a morphism in A D. Extend this to an exact triangle X
Y Z X[1]
in D. As before, we denote Z = cone(f ), and call Z the (abstract) cone of f . Recall that
the cone construction is not functorial in this setting, even though the object Z = cone(f )
is unique up to a (non-canonical) isomorphism. Define
K = 61 (Z)[1]
C = >0 (Z)
There are exact triangles
k

K
Y

/ Z[1]
/Z

/X
c

/C

One checks that K is the categorical kernel of f , and that C is the categorical cokernel of f .
Moreover, Coker(k) = Ker(c), so A is an abelian category.

3.3

Cohomological functors

Recall that if D = D (A), then for all i Z, there are cohomology functors Hi : D A,
defined by Hi (X) = H0 (X[i]). The functor H0 () is just X 7 >0 60 X = [0,0] X.
162

Definition 3.3.1. For a general triangulated category D with t-structure and A = D ,

define H0 : D A by X 7 [0,0] X and Hi = H0 [i].
Definition 3.3.2. A t-structure is called non-degenerate if
\
\
D>n =
D6n = 0.
nZ

nZ

Definition 3.3.3. Functor F : D A from a triangulated category to an abelian category

is cohomological if whenever X Y Z X[1] is exact in D, the obvious long sequence
...

/ F (X)

/ F (Y )

/ F (Z)

/ F (X[1])

/ ...

is exact.
Theorem 3.3.4.

1. For any t-structure, H0 is a cohomological functor.

2. If the t-structure is non-degenerate, then f : X Y in D is an isomorphism if and

only if Hi (f ) : Hi (X) Hi (Y ) is an isomorphism for all i.
3. we have
Ob(D6n ) = {X Ob(D) : Hi (X) = 0 for all i > n}
Ob(D>n ) = {X Ob(D) : Hi (X) = 0 for all i < n}

3.4

t-exact functors

e with t-structures, and F : D D

e is an
Suppose we have two triangulated categories D, D
exact (i.e. triangulated) functor.
e >0 . Similarly, F is right
Definition 3.4.1. Functor F is called left t-exact if F (D>0 ) D
60
60
e
t-exact if F (D ) D .
Note that if F is left t-exact, then the functor H0 F : A Ae between two cores of
t-structures is left exact in the sense of abelian categories.
Example 3.4.2. Suppose A and Ae are two abelian categories. Take D = D (A) and
e = D (A).
e Take any left exact functor
D
e If both categories A and Ae have enough injectives, functor has total right
colonA A.
derived functor F = R fitting into a diagram
A

e
/A


 F =R
e
/D

Then we can recover the original functor by only knowing F via = H0 F .

163

3.5

Derived category of t-structure

Question: Take a triangulated category D with a t-structure, take its core A, and take
its derived category D(A). How are D and D(A) related?
Answer: In general there is no relation between them, unless we have some extra conditions.
Indeed, suppose we have some functor F : Com(A) D such that it restricts to the
f

/X
/Y
/ 0] we
embedding A , D. Then for a complex of the form C = [0
should have F (C) = conecD (f ). But the cone construction is not functorial in general. So
there is no such functor F .
We will now look at a situation when such F does exist and we will give a criterion when
F is an equivalence of categories.

Definition 3.5.1. A t-structure is called bounded if

1. its non-degenerate (see definition 3.3.2);
2. for any X Ob(D) we have Hi (X) = 0 for all but finitely many is (see definition
3.3.1).
Example 3.5.2. If A is abelian, then D = Db (A) with the standard t-structure is bounded.
On the other hand D(A) is not bounded.
Definition 3.5.3. For X, Y Ob(A) = Ob(D ) define
ExtnD (X, Y ) := HomD (X, Y [n]).
There a natural map
i+j
(X, Z)
ExtiD (X, Y ) ExtjD (Y, Z) ExtD

(6.1)

called the Yoneda product. To define it, notice that

ExtjD (Y, Z) = HomD (Y, Z[j]) ' HomD (Y [i], Z[i + j])
Then the Yoneda product is just given by the composition map
HomD (X, Y [i]) HomD (Y [i], Z[i + j]) HomD (X, Z[i + j]).
Remark 3.5.4. If X = Y = Z, the construction above turns ExtD (X) into a ring (algebra)
called Youneda algebra.
Theorem 3.5.5 ([BBD82]). Suppose a triangulated category D is equipped with bounded
t-structure, and suppose A = D is its core. Assume that F : Db (A) D is a t-exact
functor.
Then F is an equivalence of triangulated categories if and only if ExtD is generated
by Ext1D in the sense that Exti (X, Y ), X, Y Ob(A), can be written as a linear
combination of monomials 1 bi , with j Ext1D (Xj , Xj+1 ), where X1 = X and
Xi+1 = Y .
164

3.6

Gluing t-structures

Question: How can we get non-trivial t-structures?

One way to do that is to glue new t-structures from old ones. Lets see how we can do
that in more detail.
Let D be a triangulated category, and N D be a triangulated (thick) subcategory. We
call
N

/D

/ D/N

(6.2)

an exact triple of triangulated categories. Call E = D/N .

Definition 3.6.1. Suppose N , D, E have t-structures. They are called compatible if both i
and Q are t-exact functors. Then we also say that such triple is a t-exact triple.
Notice that if the exact triple 6.2 is t-exact, then t-structures on N and E are uniquely
determined by the t-structure on D:
N >0 = N D>0

N 60 = N D60

E >0 = Q(D>0 )

E >0 = Q(D>0 )

Theorem 3.6.2 ([BBD82]). Assume the triple 6.2 is t-exact. Define

N >0 = {X Ob(D) | HomD (X, Y ) = 0, Y N >0 }

N >0 = {X Ob(D) | HomD (Y, X) = 0, Y N <0 }
Then the t-structure on D is given by




D60 = Q1 E 60 N >0




D>0 = Q1 E >0 N <0
Corollary 3.6.3. If N D E is an exact triple, then for any t-structures on N and E
there exists at most one t-structure on D making the sequence N D E t-exact.
Theorem 3.6.4 (Recollement). Let N D E be an exact triple. Assume that i : N D
has both left and right adjoints (equivalently, Q : D E has both left and right adjoints).
Then for any t-structures on N and E there exist unique t-structure on D, compatible with
the given ones.

3.7

Examples of gluing

Classical (topological) recollement

Suppose X is a topological space, Z X is a closed subset, and U = X \ Z is its complement.
Denote by i : Z , X and j : U X the natural inclusions.
Consider the categories AX , AZ , AU of abelian sheaves (i.e. sheaves of abelian groups) on
the spaces X, Z, U respectively. Denote by DX = Db (AX ), DZ = Db (AZ ) and DU = Db (AU )
165

the correspondent derived categories. Then we have the following diagram (see subsection
2.4 in the Chapter 3)
i

DZX

j!


/ DX
X

/ DU

i!

Let us list some properties of these functors.

1. We have adjoint triples (i , I = i! , i! ) and (j! , j ! = j , j ).
2. Also i! j = 0, and so j i = 0 and j j! = 0.
In this case theorem 3.6.2 says that


60
DX
= F DX | j F DU60 and i DZ60
Similarly,
n
o
>0
DX
= F DX | j F DU>0 and i! DZ>0
Fix perversity p(Z) = n and p(U ) = m (this is just a pair of integers associated
to each of the sets Z and U ). Consider shifted t-structures, AZ [n] := DZ6n DZ>n and
AU [m] := DU6m DU>m . Then the gluing theorem allow us to glue these shifted t-structures
into a t-structure on DX called perverse t-structure with perversity p. We denote DX
equipped with such a t-structure by p DX . Its core p A = (p DX ) is called perverse core.
Abstract blow-down
e X is the blow-up of X
Suppose X is a smooth surface, x X is a closed point, and : X
1
at x with the exceptional curve L = (x) X. We would like to know if it is possible to
e of coherent
recover the category coh(X) of coherent sheaves on X from the category coh(X)
e
sheaves on X.
e the structure sheaf of the curve L. Denote
Denote by F = OL coh(X)
e | R Hom
N = {G Db (coh(X))
e (G, F) = 0}.
D(coh(X))
Theorem 3.7.1 (Bondal-Orlov,[BO95]). There is an equivalence of triangulated categories
e
Db (coh(X))/N
' Db (coh(X)).
So we know that we can recover the derived category of coherent sheaves on X, but we
want to recover the category coh(X) itself. The problem is that it sits inside Db (coh(X))
in a twisted way, i.e. it does not coincide with the core of Db (coh(X)) with the standard
t-structure, but rather with a twisted t-structure.
e | Homcoh (T, F) = 0} and
Define subcategories of coherent sheaves T = {T coh(X)
e | Homcoh (F, T ) = 0, T T}.
F = {F coh(X)
166

e by putting p D60 =
Define perverse (or twisted) t-structure on D := Db (coh(X))
{B D60 | H0 (B) T} and p D>0 = {B D>1 | H1 (B) F}. Let p A =p D60 p D>0 be
the core of this perverse t-structure. Then we have the following theorem.
Theorem 3.7.2. If we denote Q : D D/N ' Db (coh(X)) (using notations above) the
canonical quotient functor, then we have equivalence of categories Q(p A) ' coh(X).

3.8

Recollement for triangulated categories

We have already breafly discussed the ideas below in the subsection 2.4 of the Chapter 3,
and also in the example 3.7 above. Here we will look at it more carefully.
Definition 3.8.1. Given three triangulated categories D0 , D, D00 (notice: we do not mention
any t-structures at all) we say that D is recollement (or gluing) with respect to D0 and D00
if we have six functors
i

DZX

j!


/ DX
X

/ DU

(6.3)

i!

satisfying the following properties

R1 (i , I = i! , i! ) and (j! , j ! = j , j ) are adjoint triples;
R2 i! j = 0 (it is equivalent by the adjunction to j i = 0);
R3 The following adjunction morphisms are isomorphisms:

i i idD0 i i!
j j idD00 j j!
R4 There are exact triangles
i i! X X j j ! X i i! X[1]
j! j X X i! i X j! j X[1]
Remark 3.8.2. Notice that the axiom R3 implies that functors i , i! , j , j! are full embeddings.
167

3.9

Example: topological recollement

We have already discussed this example when we were considering gluing theorems for
t-structures, see subsection 3.7.
Recall that we have the following setup. We have a topological space X, Z X is a
closed subset, and U = X \ Z is its complement, and i : Z , X and j : U X are the
natural inclusions.
Consider the categories AX , AZ , AU of abelian sheaves on the spaces X, Z, U respectively.
We denote by D = Db (AX ), D0 = Db (AZ ) and D00 = Db (AU ) the correspondent derived
categories.
Then we will have the recollement diagram (6.3) (where we need to replace functors j!
and j by Rj! and Rj respectively).
We need to say a few words what these functors from the diagram (6.3) actually are.
Functors i , Rj are push-forwards of sheaves, and i, j are pull-backs.
Recall that a continuous map f : X Y is called proper if for any compact subset
K Y the set f 1 (K) X is also compact. For any such map f we can define functor f! as
follows. For any sheaf F AX we define f! F f F to be the sub-sheaf of the push-forward
sheaf, such that
f! F(U ) = {s F(f 1 ) | f |supp(s) : supp(s) , U is proper}
Notice that if f is a closed embedding, then f! = f . If f is open embedding, then f! is
extension by zero functor.
Later we will need the following table
Functor

f
f!
f
Hom
Hom

Exactness
left
left
left

left
left

Acyclic classes
injective, flabby
inj., flabby
inj., flabby, soft

inj.
inj.

Total derived
R
Rf
Rf!

R Hom
RHom

right

flat

Classical derived
H (X, ) (sheaf cohomology)
Ri f
Ri f!

Exti
Exti
i

T ori

Here RHom(F, G) := Hom (F, I ) for an injective resolution G I . Moreover,

T ori (F, G) := Hi (F P ) for a flat resolution P G.

3.10

Example: algebraic recollement

Let be a k-algebra, and assume that it has finite global dimension. Recall that global
dimension is defined to be the supremum over all -modules M of lengths of minimal
projective resolutions of M . Assuma also that we fixed a 2-sided ideal I and denote by
D the quotient D = /I. Denote by i the canonical projection  /I = D.
168

Then we will obtain the following triple of functors

i

Mod(D)

/ Mod()

i!

Lets describe what the functors above are. First of all, i denotes the restriction of
scalars, given by i M = MD D D . Using the usual tensor-hom adjunction, we get
Hom (MD D D , N ) ' HomD (M, Hom(D D , N )).
Thus, using this adjunction we can define the right adjoint functor i! to the functor i
simply by i! = Hom (D D , ).
From the other hand, we might as well define the restriction of scalars i as i =
HomD ( DD , MD ). Again, using the standard tensor-hom adjunction we get the left adjoint
functor i given by i = D. So weve got an adjoint triple (i , i , i! ). We can pass to
the derived categories to obtain the adjoint triple
Li

Db [Mod(D)]

/ Db [Mod()]

Ri!

In general, without any extra-assumptions, we cant extend to full recollement situation.

Assume from now on that the ideal I was also idempotent, i.e. it is generated by an
idempotent e, I = e where e2 = e.
Define an algebra U = ee. Notice that the unit in U is given by e. Moreover, notice
that U ' End (e). Assume also the U has finite global dimension.
Then we indeed have all 6 functors from the recollement. Namely, we have the following
diagram
Li

Db (D)

Lj!


/ Db ()
[

Ri!

/ Db (U )

Rj

In this diagram, functors Ri! , i and Li were defined above. Moreover, functor j is
L

given by j M = e M = eM , functor Lj! is defined as Lj! N = e U N . Finally, we

define Rj N = R HomU (e, N ).
Remark 3.10.1. The functor Mod() Mod(ee) given by M 7 eM is Morita equivalence
if and only if e = .
169

For an application of the above general construction see [BCE08]. The following example
describes the situation the paper is dealing with.
Example 3.10.2. Suppose X is a smooth variety, G is a finite group acting on X. Denote
by D(X) the ring of differential P
operators on X, and consider the algebra = D(X) o G.
1
Define an idempotent e = |G|
g. Note that Mod(D(X) o G) ' Mod(D(X)G ). this
gG

observation is very useful when one is interested in some invariants of D(X)G which are
invariant under Morita equivalence. Indeed, algebra D(X) o G is much easier to understand
than D(X)G .

170

Chapter 7

Applications of homological algebra

1

Perverse sheaves and quivers

In this section we will discuss application of the gluing theorems for t-structures. First we
will discuss abstract recollement, and then we will start talking about perverse sheaves. Our
main references on the subject are [BBD82], [Rei10], [GMV96] and [MV88].

1.1

Stratifications

Let X be a topological space.

Definition 1.1.1. A stratification S = {Si }iI of X is a finite decomposition of X into a
disjoint union of nonempty, locally closed strata Si .
S
In other
S words, X = Si , the Si are pairwise disjoint, and each Si is a union of strata,
i.e. Si = jJ Sj for some J I.
Remark 1.1.2. Any complex-analytic or algebraic variety X over C admits a stratification
with all Si nonsingular, and such that the following equidimensionality condition holds:
for any p, q Si , there is a diffeomorphism of X, preserving all strata, such that (p) = q.
Remark 1.1.3. If S is a stratification of X, then S admits a natural partial order whereby
Si 6 Sj if and only if Si Sj . In this partial order, Si is minimal if and only if Si is closed,
ad Si is maximal if and only if Si is open.
Example 1.1.4. Let Z X be a closed nonempty subspace, U = X r Z. Let S = {S0 =
Z, S1 = U }. We have S0 6 S1 , and clearly X is the disjoint union of S0 and S1 .
Example 1.1.5 (coordinate stratification). Let X = Cn , and define
SI = {(x1 , . . . , xn ) Cn | xi = 0 if and only if i I}.
Then S = {SI : I 2{1,...,n} } is a stratification of Cn . If n = 1, then our stratification is
{0, C }.
171

Example 1.1.6 (Schubert stratification). Let G be a complex semisimple algebraic group,

B G a Borel subgroup, and X = G/B the flag variety. Then X admits a stratification
{Sw : w W }, where the Sw are Schubert cells.
Example 1.1.7 (MacPherson-Vilonen, see [MV88]). Let n 6 m be integers. Put X = C2 ,
S0 = 0, S1 = {(x, y) : y n = xm } r S0 , and S2 = C2 r (S0 S2 ).

1.2

Perversity function and perverse sheaves

Definition 1.2.1. Let (X, S) be a stratified space. A perversity is a function p : S Z.

Given (S, p) and S S, write iS : S , X for the inclusion. Write D(X) = D(ShX ). We
define two full subcategories of D(X):
p

D60 (X) = {F D(X) : Hn (iS F ) = 0 for all S S and n > p(S)}

D>0 (X) = {F D(X) : Hn (i!S F ) = 0 for all S S and n < p(s)}.

Equivalently,
p

D60 (X) = {F D(X) : iS F D(X)6p(S) for all S S}

D>0 (X) = {F D(X) : i!S F D(X)>p(S) for all S S}

p

Theorem 1.2.2. For any (S, p), the pair (p D60 , D>0 ) is a t-structure on D(X), which
induces a t-structure on D (X) for all {b, +, }.
Definition 1.2.3. We put Pervp (X) = p D60 (X) p D>0 (X); this is the (abelian) category
of p-perverse sheaves on X.
If p = 0, then Pervp (X) = ShX . We can replace Z by any constant sheaf O of rings on
X, and replace ShX by the category of sheaves of O-modules. We write Pervp (S, O) for the
resulting category.
The most common setting is where each Si is over C, and p = 21 dimR S = dimC S.
This is called middle perversity.

1.3

Middle extension

We will define a functor j! , sometimes called middle extension, or also Goresky-MacPherson

extension.
i

>0
Theorem 1.3.1 (GM). Let Z
X
U be the open-closed stratification. Let (D60
U , DU )
>0

b
b
and (D60
Z , DZ ) be t-structures on D (U ) and D (Z). Let AU = DU and AZ = DZ be the

corresponding cores. Then for any F AU there exists a unique G AX such that

1. j G = F
2. i G D61
Z
172

3. i! G D>1
Z .
Proof. Recall that j j ' idU and (j! , j , j ) is an adjoint triple. Thus for any F1 , F2 DU ,
HomD (j! F1 , j F2 ) = HomDU (F1 , j j F2 )
' HomDU (F1 , F2 ).
NF

Taking F = F1 = F2 , the identity morphism on F gives a morphism N : j! F j F .

In other words, we have a natural transformation N : j! j . Put j! = Im(N : j! j ).
The functor j! has the desired properties.
Recall the general situation of recollement. We have three triangulated categories with
six functors
i

DW 0

j!


/D
W

/ D 00

i!

Here, (i! , j ! = j , j ) is an adjoint triple, and we require that j j ' idD00 ' j j! .
For any F, G Ob(D0 ), we have
HomD (j! F, j G) ' HomD00 (F, j j G) ' HomD00 (F, G).
Setting F = G, we get for any F Ob(D00 ) a morphism N (F ) : j! F j F . This yields a
natural transformation (called the norm) N : j! j .
Assume D0 , D, D00 are derived categories of abelian categories.
Definition 1.3.2. The (Goresky-MacPherson) middle extension is j! = Im(N : j! j ).
So j! is a functor D00 D. [Note that it is not clear a priori if this makes sense
the image may not exist.]
Definition 1.3.3. Let (X, S) be a stratified space with perversity function p : S Z. Let
R be a noetherian ring, and D = D(X) be the derived category of sheaves of R-modules on
X. Let iS : S , X be the inclusion for S S.
p

D60 (X) = {F D(X) : iS F D6p(S) (X) for all S S}

D>0 (X) = {F D(X) : i!S F D>p(S) (X) for all S S}.

Let Pervp (S) be the core of this t-structure.

Let U be a (union of) open strata, j : U , X. Then we have functors Rj , j! : D(U )
D(X).
Lemma 1.3.4. For any perversity p, the functor Rj (resp. j! ) is left t-exact (resp. right
t-exact) with respect to the p-perverse structure (p D60 (X), p D>0 (X)).
173

Hence, if P Pervp (U ) is a perverse sheaf, then Rj (P) Ob(p D(X)>0 ) and j! (P)
Ob(p D(X)60 ).
Definition 1.3.5. Let the situation be as above. Define functors p j , p j! : Pervp (U )
Pervp (X) by
p

j (P) = p 60 (Rj P) ' p [0,0] (Rj P)

j! (P) = p >0 (j! P) ' p [0,0] (j! P) .

Here p 60 and p >0 are the truncation functors with respect to the p-perverse structure.
The norm transform N : j! Rj induces a perverse norm p j ! p j , which a natural
transformation between functors Perv(p, U ) Perv(p, X).
Definition 1.3.6. The (perverse) Goresky-MacPherson extension is
p

j ! (P) = Im(p j ! (P) p j (P)).

Proposition 1.3.7. For any union of open strata U , if P Pervp (U ), then F = p j ! (P)
Pervp (X), as an extension of P to X (i.e. j F = P), is characterized uniquely by the
following conditions:
Hn (iS F ) = 0

for all n > p(s)

(i!S F )

for all n 6 p(s).

=0

for all S X r U .

1.4

Constructible complexes

Let (X, S) be a stratified space. We assume that all strata are good, i.e. equidimensional,
with t each iS : S , X having finite cohomological dimension. Let R be a noetherian ring.
Definition 1.4.1. A sheaf F Sh(X, R) is constructible with respect to stratification
S if iS F =: F |S is locally constant (i.e. a local system) for each S S. A complex
F Com(ShX (R)) is (cohomologically) constructible if each Hn (F ) is a local system.
Definition 1.4.2. We let DS (X, R) D(X,S
R) be the full subcategory of constructible
complexes with respect to S. Let Dc (X, R) = S DS (X, R), where S ranges over all good
stratifications of X.
One reason to introduce Dc (X, R) is Verdier duality. Recall that for f : X {} the
= f ! R; this is called the dualizing complex.
unique map, we define DX
Theorem 1.4.3 (Poincare-Verdier duality). Let R = k be a field. Then the dualizing complex
DX is in Dc (X, k). If we define D = RH (, DX ), then DX : Dc (X, k) Dc (X, k) is an
equivalence of categories.

om

174

From now on, by p-perverse sheaves we mean the category

p >0
Pervpc (X, R) = Dc (X, R) = p D60
c (X, R) Dc (X, R)

There is a duality theory for perverse sheaves. Define p : S Z by p (S) = p(S)dimR (S).
Notice that
Hn (iS or ! F ) = 0
Hn (iS or ! DX F ) = 0

n > p(S)
n > p (S).

Again, assume R = k is a field. Then the duality functor DX : Pervpc (X, k) Pervpc (X, k)
is an equivalence of categories. Assume all S S are even-dimensional, and put p1/2 (S) =
21 dimR (S). This is known as the middle perversity. Poincare-Verdier duality restricts
p
to an autoduality on Pervc 1/2 (X, k). In general, we get a duality between Pervpc (X, k) and

Pervpc (X, k).

Proposition 1.4.4. Let p = p1/2 . Let j : U , X be the inclusion of an open stratum. Let
P Pervpc (X, k) be autodual. Then F = p j ! P is the unique autodual p-perverse extension
of P to F such that Hn (iS F ) = 0 for all S X r U .

1.5

Refining stratifications

Fix a space X. Assume that a stratification T is obtained from S by refining strata in that
each S S is a union of strata in T . There is a natural inclusion DS (X, R) , DT (X, R).
Moreover, we assume that S and T come with compatible perversities p, q, i.e.
p(S) 6 q(T ) 6 p(S) + dim S dim T
for all S S, T T with T S.
Theorem 1.5.1. The t-structure on DT (X, R) of perversity q induces the t-structure of perversity p on DS (X, R), and an exact embedding of abelian categories Pervpc (X) , Pervqc (X).
For any j : U , X, the corresponding functors q j , q j ! , q j ! are compatible with those for p,
in the sense that e.g. the following diagram commutes:
/ Pervq (X)
c

Pervpc (X)


q j

pj


/ Pervq (U )
c

Pervpc (U )

Example 1.5.2. Let X be a complex manifold, R = C, and S a stratification of X by

nonsingular complex submanifolds. Let p : S Z be a perversity such that p(S) only
depends on 2 dimC (S). We also require that p satisfy
0 6 p(n) p(m) 6 m n

n 6 m.

()

Using Theorem 1.5.1, define a t-structure of perversity p on Dbc (X, C). Let Pervpc (X) be the
core of this t-structure.
175

Proposition 1.5.3. For F Dcb (X, C). The following are equivalent:
1. F Pervpc (X)
j

2. All irreducible submanifolds S X contain a Zariski-analytic-open subset U

S,
such that
Hi (Rj F ) = 0
H (j F ) = 0
i

i > p(dimR S)
i < p(S).

Theorem 1.5.4. Under the assumption (), the category PervcS (X) = Pervp (X) DS (X, C)
is Artinian (and Noetherian) and its simple objects are of the form
L (S, E ) = (p iS )! E [p(S)]
where E is a locally constant sheaf on S S. In particular, if all S are simply connected,
then there is a bijection between simple objects and strata in S.
In general, the category Pervp (X) is not Artinian, but it is Noetherian. If p = p1/2 ,
then autoduality + Noetherian Artinian. For this reason, one often restricts to the
(finite-length) category Pervp1/2 (X).

1.6

Gluing

Start with an easy stratification Z , X - U , Z = X r U . Let i : Z X and j : U X

be the inclusions. We will assume that U is a union of open strata, so that p induces
perversities on U and Z. A good example is {0} , C - C , k = Can . In this example, one
typically puts p({0}) = 0 = p(C ). We would like to Describe the category Pervpc (X, k) in
terms of Pervpc (Z, k) and Pervpc (U, k). Also, given G Pervp (U ) and H Pervp (Z), we
would like conditions on G and H that guarantee the existence of F Pervp (X) such that
i F ' H
jF ' G .
For the first question, we have the following theorem.
Theorem 1.6.1. The t-structures of perversity p on Dbc (U, C), Dbc (Z, C) and Db (X, C) are
related by the general recollement construction.
See Section 1.3 for the general notion of recollement.
p

Theorem 1.6.2. Pervc 1/2 (C, Can ) is equivalent to the category of quadruples (V, W, A : V 
W : B) such that AB + 1 and BA + 1 are invertible.
This equivalence comes from the functors vanishing cycles and nearby cycles.
This corresponds to the quiver Q = . From this, we can form the deformed
p
preprojective algebra q,} (Q), and Pervc 1/2 (C, Can ) ' q,} (Q)-Mod.
176

Matrix factorizations

2.1

Introduction

Matrix factorization was first introduced by Eisenbud [Eis80] to describe the stable homological features of modules over a hypersurface singularity. The ideas and phenomena
described in that paper was later formalized in modern language ([Orl04] [Buc86]) as exact
equivalences between the triangulated categories
[MF(R, w)] ' MCM(S) ' Dsg (S),
where
1. R is a regular commutative ring with finite Krull dimension, and w R is a non-unit
non-zerodivisor.
2. The category [MF(R, w)] is the homotopy category associated to the DG category of
matrix factorizations
3. The category MCM(S) is the stable category associated to the Frobenius category of
maximal Cohen-Macaulay modules over the (Gorenstein) ring S
4. The category Dsg (S) is the singularity category of X = Spec(S), defined by the Verdier
quotient Dsg (X) = Db (coh X)/perf(X)
By its construction, The category [MF(R, w)] has a natural enrichment MF(R, w), which
is a Z/2-graded DG category. (i.e., its morphism are 2-periodic complexes). Dycherhoff[Dyc]
then set out to study this DG-category, viewing it as a noncommutative space in the sense of
[KKP08]. Indeed, showing that a suitable enlargement of M F (R, w) has a compact generator,
Dycherhoff has shown that MF(R, w) is derived Morita equivalent to a DG-algebra, in the
sense of Toen [Toe].
Therefore, the noncommutative space in question is DG-affine, and we can use that to
compute the Hochschild cohomology of this DG-category, which can be thought of as Hodge
theory on the noncommutative space.

2.2

The classical exact equivalences

In this section, we describe the three triangulated categories mentioned in the introduction,
and show that they are exact equivalent.
Definition 2.2.1. Let R be a regular domain with finite Krull dimension, and w R is a
nonzero non-unit. The category MF(R, w) of matrix factorization is defined by:
1. An object of this category consists of a Z/2-graded R-module P = P 0 P 1 with
two R-linear maps d0 : P 0 P 1 and d1 : P 1 P 0 satisfying d1 d0 = w idP 0 and
177

d0 d1 = w idP 1 . We will often call such an object a matrix factorization, and denote
it by


d1
1
0
*
P ) P .
d0

2. Between two objects (P, dP ) and (Q, dQ ) there is a Z/2-graded morphism complex
Hom(P, Q) defined as the set of R-linear homomorphism from P to Q with obvious
grading. This complex has the differentials d0 : Hom0 (P, Q) Hom1 (P, Q) and
d1 : Hom1 (P, Q) Hom0 (P, Q) both defined by the formula df = dQ f (1)|f | f dP .
3. Composition of morphism complexes is defined by composition of the R-module maps.
It is then straightforward to check that this is a Z/2-graded DG-category. That is,
its morphism complexes actually satisfy d2 = 0, and its composition maps Hom(Q, R)
Hom(P, Q) Hom(P, R) are chain maps.
Given any Z/2-graded DG-category C, we can define the category [C] = H0 (C) by
Ob([C]) = Ob(C)
Hom[C] (X, Y ) = H0 (HomC (X, Y )) .
In our case C = MF(R, w) and we see that [C] is the category of chain maps between matrix
factorizations modulo homotopy.
Given objects X and Y in MF(R, w), and a chain map f : X Y , (i.e. f
Z 0 (Hom(P, Q))), we can define the cone of f in exactly the same way as the standard
cone construction of chain complexes. The collection of triangles isomorphic to the associated standard triangles will make [C] into a triangulated category. For details, see
[Orl04].
Definition 2.2.2. Let X be a Noetherian scheme of finite Krull dimension which is either
affine or is quasiprojective over a field. Then we define its singularity category as the Verdier
quotient
Dsg (X) = Db (coh X)/perf(X),
where perf(X) denotes the triangulated subcategory of Db (coh X) consisting of complexes
that are isomorphic in Db (coh X) to a bounded complex of locally free sheaves.
The following proposition explains why it is called the singularity category.
Proposition 2.2.3. X is nonsingular if and only if Db (coh X) = perf(X).
Proof. By Serres theorem, coh(X) has enough locally free sheaves. Thus, for any bounded
complex M of coherent sheaves, there is a quasi-isomorphism P M from a bounded
above complex of locally free sheaves. Then, Coker(dn
P ) is always locally free if and only if
the local ring at every stalk has finite global dimension, if and only if X is regular.
178

Proposition 2.2.4. Let S be a Gorenstein ring of Krull dimension n. (i.e., S is a Noetherian

commutative ring and every localization Sp at a prime is a Gorenstein local ring.) Let M be
a finitely generated S-module. Then the following statements are equivalent:
1. M is a high enough syzygy. i.e., there exists an exact sequence
0 M P n P 1 N 0
for some other finitely generated S-module N , where P i are finitely generated projective
modules over S.
2. For any prime p ann(M ), the localization Mp satisfies depth(Mp ) = dim(Sp )
3.
exti (M, S) = 0 for all i > 0
4. There is a quasi-isomorphism M P from M to a complex of finitely generated
projective modules bounded below at 0, i.e.
P = 0 P0 P1
Proof. 1 2. One can show that, in fact, any regular sequence of Sp is actually regular in
Mp [Eis80, Lemma 0.1]. Since Sp is Cohen-Macaulay, we have our claim.
2 3. For any prime p ann(M ), let (a1 , . . . , ad ) be a maximal regular sequence of
Mp , then apply the long exact sequence of
ext(, Sp )
0

/ Mp /(a1 , . . . , ai1 )

ad

/ Mp /(a1 , . . . , ai1 )

/ Mp /(a1 , . . . , ai )

/ 0.

Use the Nakayama lemma to prove inductively that

extd+i (Mp /(a1 , . . . , ad ), Sp ) = 0

extd+i1 (Mp /(a1 , . . . , ad1 ), Sp ) = 0

ext (Mp , Sp ) = 0.
3 4. Use the fact that R Hom(, S) is a duality functor on Db (S), we see that such a
functor restricts to a duality functor on the class of finitely generated modules such that 3
holds, and is simply the functor () there. Thus, we can obtain a free resolution of M ,
and then apply () again.
4 1. This is obvious.
179

Definition 2.2.5. We call a module satisfying any of the equivalent conditions a maximal
Cohen-Macaulay module over S, and they form a full subcategory MCM(S) of Mod(S).
Proposition 2.2.6. The category M CM (S) can be made into an exact category by endowing
it with sequences that are short exact in Mod(S), Then, an object is projective injective
projective as an S-module.
Proof. By Proposition 2.2.4, a maximal Cohen-Macaulay module has enough projective
module in both sides. Hence, if M is either injective or projective in the exact category
MCM(S), then it must be a direct summand of a projective S-module, and hence is itself a
projective S-module. The other direction is easy.
Proposition 2.2.6 tells us that MCM is a Frobenius category in the sense of [Kela]. Thus
we can form its associated stable category.
Definition 2.2.7. Write MCM(S) for the stable category associated to the Frobenius category
MCM(S).
Now, let R be a regular Noetherian domain of finite Krull dimension n, and w R is a
nonzero non-unit. Let S = R/w.
We claim that the triangulated categories [MF(R, w)], MCM(S) and Dsg (Spec S) are
exact-equivalent.

Definition 2.2.8. Given an object X in MF(R, w), we denote by X the complex of Smodules
. . . X0 X1 X0 0
where the rightmost nontrivial module is at degree 0.

Proposition 2.2.9. For any object X in MF(R, w), the complex X is exact except possibly
at degree 0.
Proof. This is a straightforward diagram chasing, merely using the fact that multiplication
by w is injective on finitely generated projective modules over the regular ring R.
Given an object X in MF(R, w), we note that Im(d1 ) Im(d1 d0 ) = Im(w), and hence
Coker(d1 ) is an S-module.
We know that Coker(d1 ) is an arbitrarily high syzygy, and is therefore a maximal CohenMacaulay module over S. One can show that the construction X 7 Coker(d1 ) actually
induces a functor
Coker : [MF(R, w)] MCM(S).
Theorem 2.2.10. The functor Coker : [MF(R, w)] MCM(S) is an equivalence of categories.
180

Proof. The fact that Coker is fully-faithful is a diagram chasing exercise. (It is useful to note
that a map of S-modules factor through a projective iff it factors through a free module)
To show that Coker is essentially surjective, suppose we are given a maximal CohenMacaulay module M , then apply the Auslander-Buchsbaum formula to the localizations of
M at primes p ann(M ), considered as an Rp -module, we have:
pdRp (Mp ) = depth(Rp ) depth(Mp ) = m (m 1) = 1.
Hence, there is an exact sequence
d1

0 X 1 X 0 M 0
where X 1 and X 0 are projective R-module. Then use projectivity of X 0 to show that
w : X 0 X 0 lifts to an R-linear map d0 : X 0 X 1 , as in the following commutative
diagram:
X0
d0

/ X1


/ X0

d1

!
/M

/0

i.e., we have
d1 d0 = w
Hence, we have
d1 d0 d1 = w d1 = d1 w
Then we can use the injectivity of d1 to show that d0 d1 = w as well. This shows essential
surjectivity.
Now, we shall describe the functor
: MCM(S) Dsg (S).
By definition, MCM(S) is a full subcategory of mod(S). We claim that the inclusion
MCM(S) , mod(S) induces as follows:
MCM(S) 


MCM(S)

/ mod(S)


/ Dsg (S)

Indeed, since every projective S-module is a perfect complex in Dsg (S), hence every morphism
that factors through a projective S-module is zero in Dsg (S)
Proposition 2.2.11. The functor : MCM(S) Dsg (S) is an equivalence of categories.
181

Proof. See [Orl04, Prop 1.21] for a proof that is fully faithful.
To show that is essentially surjective, suppose we are given a bounded complex M of
S-modules
M = [0 M e . . . M 0 0]
which we assume, without loss of generality, to be bounded above at degree 0.

Then we can find a projective resolution P

M , where
de

d1

P = [. . . P e . . . P 0 0]
By the characterization of Proposition 2.2.4(1), we see that Im(den1 ) is maximal
Cohen-Macaulay since it is the n-th syzygy of Im(de1 ). Then the good truncation
en1 (P ) = [0 Im(den1 ) P en P 0 0]
is still quasi-isomorphic to M . Moreover, in Db (S), we have exact triangles
en (P ) en1 (P ) Im(den1 )[e + n + 1] en (P )[1]
which translate to an equivalence
en1 (P ) Im(den1 )[e + n + 1]
in Dsg (S). This shows that M ' Im(den1 )[e + n + 1] in Dsg (S).
Use the characterization of Proposition2.2.4(4), we can syzygy down the MCM module
Im(den1 ), and apply the same argument as above, to show that
Im(den1 )[e + n + 1] ' N
in Dsg (S) for some MCM module N . This shows essential surjectivity.
Theorem 2.2.12. The equivalences Coker and are exact equivalences.
Proof. The exactness of is straight-forward checking.
For the exactness of Coker, it will be easier to show that coker is exact. Indeed, by
Proposition2.2.9, this composition is just the functor
X 7 X.
To show that Coker is exact, it suffices to note that X and X[1] only differ by a projective
module; and that cone(f ) and cone(f ) also only differ by a projective module.
More precisely, we have maps in Db (S):
182

X
f

X
f

X1


/ cone(f )


/ cone(f )

/ X 1 [1]

X1


/ X


/ X[1]

/ X 1 [1]

where the first long column is the image under Coker of an exact triangle in [MF(R, w)];
the second long column is an exact triangle in Db (S); the horizontal maps are also exact
triangles in Db (S) This diagram, when read in Dsg (S), shows that Coker is exact.

2.3

The stabilization of the residue field

In the next several sections, we will survey the results from [Dyc], showing that a suitable
enlargement of the triangulated category [MF(R, w)] has a compact generator in the case of
an isolated singularity.
Let R be a regular domain of finite Krull dimension, w R a nonzero nonunit, and
S = R/w. Let (x1 , ..., xm ) is a regular sequence in R. Suppose w is in the ideal (x1 , ..., xn ),
then L := R/(x1 , ..., xm ) is an S-module. We will obtain a description of the object in
[MF(R, w)] that correspond to L in Dsg (S).
Let V be the free R-module V = Re1 . . . Rem , and suppose
V V has homological
degree 1. It is easy to see that on the graded commutative algebra
V , there isVa unique
DG-algebra
structure
with
differential
(e
)
=
x
of
degree
+1.
(Here,
ei V = 1 V and
i
i
V0
xi R =
V ) We
V denote this differential by s0 , which is a degree +1 differential on the
graded complex
V :
0

^m

0
0
V
. . .

^1

0
V

^0

V 0

Moreover, since w is in the

, ..., xn ), we may suppose w =
Viideal (x
V1i+1
define a degree 1 map s1 :
V
V by
s1 (x) = (wi ei ) (x),
So we have a sequence
0

^m

1
1
V
. . .

^1

183

0
V

^0

V .

wi xi . Then we can

Note that we have:

s0 s0 = 0
s1 s1 = 1
s0 s1 (x) = s0 ((wi ei ) (x))
= s0 (wi ei ) (x) (wi ei ) (s0 (x))

by Leibniz rule of s0

= w x s1 s0 (x)
V
So s0 s1 + s1 s0 = w id on V .
Thus, letting s = s0 + s1 , we have s s = w id, and hence


Vodd
s Veven
stab
*
L
=
V
V .
)

is a matrix factorization in MF(R, w).

Proposition 2.3.1. For any complete intersection module L as described above, Lstab
MF(R, w) is the object that corresponds to L Dsg (S) under the exact equivalence MF(R, w) '
Dsg (S).
Proof. We refer the reader to [Dyc] for a proof.
In particular, if R is a regular local ring of dimension n, then its maximal ideal m
is generated by a regular sequence m = (x1 , ..., xn ). Suppose w m is nonzero, and let
S = R/w
Hence, its residue field k = R/m, as an S-module, has stabilization the matrix factorization
hV
i
V
odd
k stab =
V  even V .
This description will be useful in showing that the k stab is a compact generator of [MF (R, w)],
to be defined in the next subsection.

2.4

An enlargement of the matrix factorization category

Definition 2.4.1. Let R be a regular domain with finite Krull dimension, and w R
is a nonzero non-unit. We define MF (R, w) as the category whose objects are matrix
factorization
 1

P  P0
as in Definition 2.2.1, but this time we do not require P 0 and P 1 to be finitely generated.
Morphism complexes are defined exactly as in Definition 2.2.1.
Proposition 2.4.2. Let R, w, S be as above.
1. MF (R, w) is a Z/2-graded DG category, and MF(R, w) MF (R, w) is a full DGsubcategory.
184

2. MF (R, w) has arbitrary coproducts.

3. [MF (R, w)] is a triangulated category with arbitrary coproducts.
Coker

4. [MF (R, w)] Mod(S) is fully faithful.

5. Every object in [MF(R, w)] is compact in [MF (R, w)].
See Definition 2.4.3 for the precise definition of a compact object.
Proof. 1-4. These are either obvious, or have proofs exactly as for [MF(R, w)].
5. Simply note that since the R-modules P 0 and P 1 in an object of [M F (R, w)]
are finitely generated projective modules, the Hom complexes HomMF(R,w) (P, ) already
commute with arbitrary coproducts. The same therefore holds after we take H0 () of the
Hom complexes.
Definition 2.4.3. Let T be a triangulated category. We call an object X T compact if it
commutes with arbitrary coproducts. That is, whenever the coproduct of a family{Y } of
objects T exists, the obvious map
a

HomT (X, Y ) HomT (X,

Y )

is an isomorphism, where the coproduct on the left is taken in the category of abelian groups.
Definition 2.4.4. Let T be a triangulated category. We call an object X T a generator
if the smallest strictly full triangulated subcategory of T containing X is T itself.
Proposition 2.4.5. Let T be a triangulated category with arbitrary coproducts, and X T
be a compact object. Then X is a generator if and only if for all Y T ,
HomT (X[i], Y ) = 0 for all i Z

Y = 0 in T .

Sketch of proof. This is an application of the technique of Bousfield localization. Let S be

the smallest strictly full triangulated subcategory of T containing X, and denote by S
its right orthogonal, defined as the full subcategory of T consisting of objects Y such that
HomT (Z, Y ) = 0 for all Y S. In this context, we are trying to prove that S = 0 T = S.
To prove this, we use the assumption that T has arbitrary coproducts to conclude (using
Brown representability) that T T /S has a right adjoint. This will imply that S T /S
is an exact equivalence. Thus S = 0 T /S = 0 T = S. For details of the argument,
we refer the reader to [Nee].
We state, without proof, the result in Dyckerhoffs paper [Dyc] which is regarded by him
as the main result of that paper.
185

Definition 2.4.6. Let R be essentially of finite type over k, equidimensional of dimension

n, and the sheaf of Kahler differentials R/k is locally free of rank n. We say that (R, w)
has isolated singularity at a rational point if Spec(S/(1 w, ..., n w)) is supported at a closed
point m Spec(R) with residue field R/m = k.
In such cases, R is regular at m, and the same argument as in the previous section shows
that the residue field k = R/m has a free resolution which stabilizes to the reduction mod w
of a matrix factorization, and we denote k stab [M F (R, w)] as that matrix factorization.
Of course, if k is a perfect field, then R is regular, and k stab [M F (R, w)] is the object
that correspond to k = R/m DSg (S) under the equivalence [M F (R, w)]
= DSg (S)
Theorem 2.4.7. Suppose that (R, w) has isolated singularity at a rational point, then k stab
is a compact generator of [M F (R, w)]

2.5

Homotopy theory of DG categories

We present some part of the homotopy theory of DG-categories as developed in [Toe]. We

will later discuss the way Dyckerhoff applied it to the study of the Hochschild cohomology
of the matrix factorization DG-category. We begin by collecting some basic definitons about
DG-categories and their homotopy theory. The exposition is brief, and the reader is advised
to consult [Toe] and [Kelb] as needed.
Definition 2.5.1. Let C, D be (Z/2-graded) DG-categories. A DG-functor F : C D is an
quasi-equivalence if
1. For any two objects X, Y C, the cochain map
HomC (X, Y ) HomD (F X, F Y )
is a quasi-isomorphism.
2. The induced map [F ] : [C] [D] is an equivalence of categories.
We wish to identify DG categories which are quasi-equivalence. In other words, we will
eventually be interested in the category Ho(DGCat) of the category of small DG categories
localized at the class of quasi-equivalence. In order to describe such a category, we use
the theory of model category. The reader is referred to [DS95] for an exposition of model
categories.
Theorem 2.5.2. There is a model category structure on the category of all (small) (Z/2graded) DG-categories such that:
1. The weak equivalences are quasi-equivalences.
2. The fibrations are DG-functors f : T S that satisfy the following two conditions:
186

(a) Surjectivity on Hom complexes: For any objects x, y T , the map of complexes
HomT (x, y) HomS (f (x), f (y))
is surjective.
(b) The induced map [f ] : [T ] [S] lifts isomorphisms. That is, if y 0 f (x) is an
isomorphism in [S], then there exists an isomorphism x0 x that maps to that
given isomorphism.
Proof. See the paper [Tab05].
In order to clarify the definitions that will follow, we first make the following definitions.
Give k[u ] the structure of a DG-algebra with deg(u) = 2. THe category of all modules over
k[u ] is denoted by C(k[u ]). Its objects are simply 2-periodic k-complexes, and morphisms
are 2-periodic cochain maps. Let C(k) be the category of all k-complexes. There are obvious
DG-enrichments of these categories, which we denote by Cdg (k[u ]) and Cdg (k), respectively.
Thus Cdg (k[u ]) is a Z/2-graded DG-category such that Z 0 (Cdg (k[u ])) = C(k[u ]), and
Cdg (k) is a DG-category with Z 0 (Cdg (k)) = C(k).
Write dgcatk for the category of al (small) DG-categories over k. Similarly, write
dgcatk[u ] for the category of all (small) Z/2-graded DG-categories over k. The reason for
the notation dgcatk[u ] is that we can think of a Z/2-graded DG-category as a category
enriched over the monoidal category C(k[u ]). We will write Ho(dgcatk ) and Ho(dgcatk[u ] )
for the respective homotopy categories associated to these two model categories.
Definition 2.5.3. Given a DG categories T , a (left) T -module is a DG functor T
Cdg (k). Similarly, given a Z/2-graded DG-category T , a (left) T -module is a DG functor
T Cdg (k[u ])
For the rest of this subsection, statements and definitions will be given for DG-categories
only. There is obvious counterpart in the Z/2-graded case which we shall not repeat.
Example 2.5.4. If x Ob(T ), then we have a DG-functor
hx : T Cdg (k)
y 7 HomT (y, x).
Note that this is really a DG functor because the map
HomT (y, z) HomCdg (k) (HomT (z, x), HomT (z, x))
is a map of complexes. Thus, hx is a left T -module.
Given two DG functors F, G : T S between two DG-categories. It is straightforward
to define the notion of a natural transformation of given homogenous degree from F to G; it
is likewise straightforward to define the differential of such natural transformation. Then
the class of such natural transformations naturally form a complex. (cf [Kelb] for details.)
187

By the previous paragraph, the class of all DG functors T S can be given the structure
of a DG-category, which we denote by Hom(T , S). In particular, the class of all T -modules
is naturally a DG-category, which we denote by mod(T ), i.e.
mod(T ) = Hom(T , Cdg (k)).
A map between T -modules is defined to be a morphism in Z 0 (mod(T )). By definition,
this means a natural transformation between two DG functors from T to Cdg (k) which is
required to be degree zero and commute with differentials.
Theorem 2.5.5. The DG-category mod(T ) has the structure of a model category, where the
fibrations and weak equivalences are defined levelwise using the model structure on C(k[u ]).
We denote by Int(mod(T )) the full subcategory of fibrant cofibrant objects in this model
category. Write
Tb = Int(mod(T )).
An object in Tb is called perfect or compact if it is perfect or compact as an object in the
triangulated category [Tb ] ' D(T ), and we denote by Tbpf the full subcategory of perfect
objects of Tb .
It is easy to check that the image of the Yoneda functor h : T mod(T ) actually lies in
Tbpf , i.e. h induces a functor T Tbpf . Moreover, there is an enriched Yoneda lemma which
asserts that this functor is always fully faithful.
Definition 2.5.6. A DG category T is called triangulated if the DG functor
h : T Tbpf
is a quasi-equivalence. The DG category Tbpf is called the triangulated hull of T .
For any C(k)-model category M (i.e., a model category with C(k)-enrichment which is
compatible with its model structure), we have
[Int(M)] ' Ho(M).
In particular, for M = mod(T ), we have
[Tb ] ' D(T )
[Tb ]pf ' Dperf (T ).
For more details, refer to [Toe11].
If T is a triangulated DG category, it is clear that the followings hold:
1. For any object x T , the T -module hx [1] is represented by an object, say x[1].
f

2. For any morphism x

y in Z 0 (T ), the T -module cone(hx hy ) is quasi-representable,
(i.e., isomorphism in D(T ) to an object hz ) Denote the representing object by
z = cone(f )
188

3. The shifts and cone above makes [T ] into an idempotent complete triangulated category.
It seems to be a folklore that the converse is also true, but the author knows of no statement
explicitly written down.
Definition 2.5.7. A DG functor F : T S induces (by T S) an extension functor of
modules. These fit into a commutative diagram
T


Tbpf

/S

Fbpf


/ Sbpe

A DG functor F is called a derived Morita equivalence if the induced DG-functor

Fbpf : Tbpf Sbpf
is a quasi-equivalence.
Note that since Tbpe and Sbpe have shift functors, we have:

H0 (HomTbpf (x, y))

H0 (HomSbpf (Fbpf (x), Fbpf (y)))

H (HomTbpf (x, y))

H (HomSbpf (Fbpf (x), Fbpf (y)))

x, y Tbpf
x, y Tbpf

Therefore, F is a derived Morita equivalence if and only if it induces an isomorphism

H0 (Tbpf )

/ H0 (Sbpf )

/ Dperf (mod(S ))

Dperf (mod(T ))
hence the term derived Morita equivalence.

2.6

Derived Morita Equivalence

In the last subsection, we introduced several notions related to DG categories in general.

In this subsection, we will state a general result in the homotopy theory of DG categories,
which is geared towards the application to the matrix factorization category towards the
end of this subsection.
Theorem 2.6.1. Let T be a triangulated 2-periodic DG category which admits coproducts.
Let S be a full DG subcategory of T whose objects are compact in [T ]. Assume that the
smallest triangulated subcategory of [T ] which contains the objects of [S] and is closed under
coproducts is [T ] itself. Then the map
f : T S op mod , x 7 T (, x)|S
189

actually restricts to
f : T Int(S op mod) , x 7 T (, x)|S
and this restricted functor is a quasi-equivalence.
Furthermore, f induces a quasi-equivalence
Tpe
= Int(S op mod)pe = Sbpe
Proof. We proceed in several steps:
1. The first claim that each hx |S is fibrant cofibrant in S op mod is straight-forward.
2. Check that f commute with coproducts.
3. Show that f induces isomorphism on H 0 () of the complexes. i.e., Show by the
following steps that:

Hom[T ] (x, y)
HomD(S op ) (f (x), f (y)

()

(a) True for all x, y S, by the Yoneda lemma.

(b) Fix x S, let A be the full subcategory of those y Ob([T ]) such that (*) holds.
Then by (2), (3a), and 5-lemma, we have A = [T ].
(c) Fix y T , and repeat the argument analogous to (b)
4. Since T is triangulated, it is quasi-equivalence to Tbpe which has a shift functor. Hence,
the fact that f induces isomorphism on H 0 () of Hom complexes implies that f
induces isomorphism on H i () of Hom complexes. (i,e., f is a quasi-fully-faithful)
5. To show that [f ] is essentially surjective, we simply note that its essential image contain
every representable module hx , x Ob(T ). Then, an application
6. The last statement is obvious.

Note that [M F (R, w)] has arbitrary coproducts, and hence is idempotent complete (cf
[Kra07]). Therefore, by the folklore mentioned above, we see that M F inf ty (R, w), defined
in Section 3.2, is a triangulated DG category,
Hence, by Theorem 2.4.7, we may apply Theorem 2.6.1 to the cases:
1. T = M F (R, w) and S = M F (R, w)
2. T = M F (R, w) and S = {k stab }
and we have:
190

Theorem 2.6.2. Let (R, w) has isolated singulated at a rational point. Denote A as the
full subcateogry {k stab } M F (R, w). (Alternatively, A can be viewed as the Z/2-graded
endomorphism DG algebra HomM F (R,w) (k stab , k stab ).) Then there exists quasi-equivalences:

2.

\
M F (R, w)
M F
(R, w)

b
M F (R, w)
A

3.

b
\
MF
(R, w)pe
A
pe

1.

Thus, in particular, (3) says that M F (R, w) is derived Morita equivalent to the Z/2graded DG algebra A.

2.7

Hochschild cohomology

Warning: The author is not completely familiar with the material in this subsection.
Therefore, the material presented might be inaccurate. But the author has tried to make
sure that what is written here is at least morally right.
Recall that, for two Z/2-graded DG categories T ,S, there is a Z/2-graded DG categories
Hom(T , S).
This gives a functor:
Hom : dgcatop
dgcatk[u,u1 ] dgcatk[u,u1 ]
k[u,u1 ]
This functor can be derived to:
RHom : Ho(dgcatk[u,u1 ] )op Ho(dgcatk[u,u1 ] ) Ho(dgcatk[u,u1 ] )
Definition 2.7.1. Given Z/2-graded DG categories T , the image under
Hom(T , T ) RHom(T , T )
of the identity natural transformation idT is an object of RHom(T , T ), which we still denote
by idT . Then the Hochschild complex of T is defined as the complex
Hom(RHom(T ,T )) (idT , idT )
Hochschild cohomology of T is defined to be the cohomology of the Hochschild complex:
HH (T ) := H (Hom(RHom(T ,T )) (idT , idT ))
which is a Z/2-graded algebra.
It turns out that Hochschild cohomology is invariant under derived Morita equivalence
(cf. [Toe11]). Thus, to compute the Hochschild cohomology of M F (R, w), it suffices to
compute that of M F (R, w). To do that, we need to do two things:
191

1. Find a simple description of RHom(M F (R, w), M F (R, w)).

2. Determine which object in that description correspond to the identity functor.
Let (R, w) and (R0 , w0 ) be as in 2.2.1. Given an object of T M F (R R0 , w 1 +
1 w0 ), and an object X M F (R, w), it is straightforward to check that the X R T is
an object in M F (R0 , w0 ). Indeed, keeping the Koszul sign rule in mind, we can calculate:
d2 (x t) = d(dx t + (1)|x| x dt)
= d2 x t + (1)|x|+1 dx dt + (1)|x| dx dt + (1)|x|+|x| x d2 t)
= wx t + x (wt + tw0 )
= w0 x t
Thus, tensoring with a given object T M F (R R0 , w 1 + 1 w0 ) gives a natural
transformation from M F (R, w) to M F (R0 , w0 ). This determines a DG functor
M F (R R0 , w 1 + 1 w0 ) RHom(M F (R, w), M F (R0 , w0 ))

(7.1)

In order for this functor to be a quasi-equivalence, we have to impose conditions on the

pairs (R, w) and (R0 , w0 ).
Definition 2.7.2. Let R be essentially of finite type over k, equidimensional of dimension
n, and the sheaf of Kahler differentials R/k is locally free of rank n. We say that (R, w) has
isolated critical locus at a rational point if Spec(R/(1 w, ..., n w)) is supported at a closed
point m Spec(R) with residue field R/m = k.
Being an isolated critical locus is strictly more stringent than being an isolated singularity (as defined in Defintion 2.4.6) because we require Spec(R/(1 w, ..., n w)) rather than
Spec(S/(1 w, ..., n w)) to be supported at a rational point.
Proposition 2.7.3. If (R, w) and (R0 , w0 ) has isolated critical locus at a rational point,
then the functor
M F (R R0 , w 1 + 1 w0 ) RHom(M F (R, w), M F (R0 , w0 ))
defined in 7.1 is a quasi-equivalence.
Remark 2.7.4. We shall skip the proof of this theorem, but we merely note that the
condition of isolated critical locus at a rational point is used because we have the following
fact:
If (R, w) and (R0 , w0 ) have isolated critical locus,
then so does (R R0 , w 1 + 1 w0 )
Thus, we can use apply the results of Section 3 to study M F (R R0 , w 1 + 1 w0 )
192

Assume from now on that (R, w) has an isolated critical locus at a rational point. Denote
by w
= w 1 + 1 w R R.
Let I R R be the ideal I = (x 1 1 x), hence the module := R R/I is the
diagonal module. (i.e., I is the ideal of the diagonal of Spec(R) Spec(R).)
By an analogy to Fourier-Mukai theory, the following proposition, whose proof we refer
the reader to [Dyc], should not be a surprise.
Proposition 2.7.5. The stabilization stab M F (R R, w)
of the diagonal module corresponds to the identity functor under the quasi-equivalence 2.7.3.
Before we proceed to the explicit calculation of the Hochschild cohomology, we quote
the following result which allows us to pass to the completion of R.
b
Proposition 2.7.6. If (R, w) has an isolated singularity at a rational point, then R R
induces a quasi-equivalence

b w)
M F (R, w)
M F (R,

Applying this proposition to (R R, w),

we have:
Hom(RHom(T ,T )) (idT , idT )

(stab , stab )
=Hom
M F (RR,w)

[
[
stab ,
stab )

(
=HomM F (RR,
\ w)

[
stab = stab R
b is also the stabilization of the R
\
where
R-module Rb
R
\
R must be isoBy the Cohen structure theorem, the regular local complete ring R
morphic to k[[x1 , ..., xn , y1 , ..., yn ]]. Moreover, the diagonal module is given by Rb =
k[[x1 , ..., xn , y1 , ..., yn ]]/I where I is generated by the regular sequence (1 , ..., n ), i =
x i yi .
We denote w
= w 1 + 1 w as w(x,
y) = w(y) w(x). Since w
I, we have
w(x,
y) = wi i for some wi = wi (x, y)
Hence, the matrix factorization stab
has the form
b
R

stab
:=
b
R

 Vodd

V o

/ Veven

where:
\
\
V =R
Re1 ... R
Ren
s0 = DG algebra differential on

V such that s0 (ei ) = i

s1 = (wi ei ) (), where wi are defined by w

= wi i
To compute the cohomology of endomorphism complex of this matrix factorization, we
use the following Lemma:
193

Lemma 2.7.7. Let X and Y be objects in M F (R, w) and assume tat X has finite rank.
Let A be the endomorphism DG algebra of X and that Y is the stabilization of the S-module
L. Then there exists a natural isomorphism
Z/2
HomM F (R,w) (X, Y )
= HomR (X, L)

in the category D(Aop )

Thus, we have
HomM F (RR,
stab , stab )
\ w)

Z/2

\
R/I)
=Hom \ (stab , R
RR

 Vodd

V o

/ Veven

\
where () denotes RR
\ (R R/I)
This last complex is just the Koszul complex with respect to the elements (w1 , ..., wn ) in
\
R
R/I.
\
To compute this complex as an R-complex, it suffices to determine what wn ) R
R/I
corresponds to under the isomorphism
x7x1 \
b
\
R
R R R
R/I

Thus, we calculate:
w(x i ) w(x)
i
w(x1 , ..., xi1 , yi , ..., xn ) w(x)
= lim
i 0
i
i , ..., n )
w(y1 , ..., yn ) w(x)
(mod 1 , ...,
= lim
i 0
i
w(x,
y)
i , ..., n )
= lim
(mod 1 , ...,
i 0
i
= wi (x, y)

i w(x) = lim

i 0

Here, the computation is taken inside the power series ring k[[x1 , ..., xn , y1 , ..., yn ]]. Therefore, the limit operation makes sense.
Theorem 2.7.8. The Hochschild cochain complex of the Z/2-graded DG category M F inf ty (R, w)
is quasi-isomorphic to the Z/2-graded Koszul complex of the regular sequence 1 w, ..., n w
in R. In particular, the Hochschild cohomology is isomorphic, as an algebra, to the Jacobian
algebra
HH (M F (R, w))
= R/(1 w, ..., n w)
concentrated in even degree.
194

Since Hochschild cohomology is invariant under derived Morita equivalence, we have the
following:
Corollary 2.7.9. The Hochschild cohomology of the Z/2-graded DG category M F (R, w) is
isomorphic to the Jacobian algebra
HH (MF(R, w)) ' R/(1 w, . . . , n w)
concentrated in even degree.

Some physics

Well begin with a standpoint that emphasizes physical intuition, at the expense of mathematical rigor. This will be centered on a specific quantum field theory, due to Jones and
Witten.

3.1

Physical motivation

Well approach gauge invariants via an example: the group U (1). The Maxwell equations
are
B =0
E =

B
t

From these equations we know that B = A for some potential A, but A is not unique
it can be replaced by any A + A0 .

Let f = A A , where =
, and dA = 12 f dx dx . Here, the Maxwell
equations are
df = 0
df =0
These equations are invariant under the operation A 7 A .
In quantum mechanics, one starts with a stateR vector, i.e. a function (x, t). A
property E is identified with the functional hEi = d3 x E. If we replace by ei ,
nothing is changed. On the other hand, if we replace by ei(x) , then the operator will not
send to the same value. To account for this, we replace with the covariant derivative
D = + ieA . This plays the role of a connection.
We know have D ei(x) (x) = ei(x) D . This is a type of parallel transport. It can be
defined as
R
R(c; A) = ei c A dX
where c is some curve. This integral is not gauge-invariant, in the sense that if A 7 A0 =
A , then R(c; A0 ) = ei(P2 ) R(c; A)ei(P2 ) , where P1 and P2 are the endpoints of c.
This construction is motivated by the Bohm-Aharonov effect.
195

Lets move to a gauge theory over a general Lie group. In the Lie algebra, we have
a
[Ta , Tb ] = ifabc Tc , and put U = ea T , where the a R. Define A (x) = Aa Ta , where x
lives in spacetime, and the Aa are some functions on spacetime.
Define a parallel transport operator
R(x + dx, x; A) = 1 idx A (x).
Partition the curve, and get
R(c; A) = lim

N
Y

(1 ix` A (x` )) = P e

A(x)dx

`=1

where is the start and is the end of the curve. This is not gauge invariant it transforms
by U ()R(c; A)U 1 (). The Taylor expansion becomes
A0 = U A U 1 + i( U )U 1 .
We want to compute R(; A), where  is a small rectangle at x with side-lengths dx and
x . We have
R(; A) = R(x, x + dx)R(x + dx, x + dx + x)R(x + x)R(x),
which is not gauge invariant. Its trace is
tr R(; A) = Wc (A) = tr(P ei

A(x)dx

).

The operator Wc is Wilson loop. The result of this holonomy calculation is eiF dx
Working out the details, we get a formula for this curvature, which is

F = A A i[A , A ].
In the language of differential forms,
1
dA AA = F dx dx .
2
Lets
switch to some Feynman integrals. Start with the Lagrangian L = T V , and let
R
S = dtL. The Lagrangian L depends only on x and x.
We have the equation
 
L
L

= 0.
t x
x
Over a three-dimensional manifold whose metric has signature , +, +, we put
Z
1
L=
d3 x| |2 .
2 M
196

We also put
Lguage

1
=
2

d3 x|F (A)|2 ,

where F is the curvature

tensor.
R
R (x,T )
If we have Z = eiL , then hW i = 0T(x,t) eiL W , where we average W over all
functions interpolating the beginning and end points.
Let G be the space of maps M G, and let A be the space of connections. Essentially,
Z
H
Y
i K A
iK/4Lcs
WK =
Ae
tr(P e ).
A/G

3.2

Cobordism

Definition 3.2.1. A cobordism is a smooth n-manifold whose boundary is the disjoint

union of two (n 1)-manifolds.
We often write M = 0 t 1 if M is a cobordism from 0 to 1 . We do not require
either M or M to be connected. We do require both M and the i to be oriented. The
choice of an orientation on some i determines one on M in a natural way. So we could have
redefined a cobordism to be two closed oriented (n 1)-manifolds, a smooth n-manifold,

and orientation-preserving diffeomorphisms f0 t f1 : M

0 t 1 . There is an obvious
notion of equivalence of cobordisms, requiring a commutative diagram:
0

/M o

!  }

M0

`
Cobordisms M0 : 0 1 and M1 : 1 2 can be composed: M1 M0 = M1 1 M0 . It
is not immediately obvious that the composite has a smooth structure, but this is in fact
the case. One uses Morse functions fi : Mi [i, i + 1], such that f0 is regular on [1 , 1]
and f1 is regular on [1, 1 + ] for some  > 0.
The composite of cobordisms to be independent of the isomorphism classes of the
individual cobordisms. We can define the n-cobordism category cob(n), whose objects
are closed oriented (n 1)-manifolds, and whose morphisms are differmorphism classes
of cobordisms. It is an easy exercise to check that the cob(n) are in fact categories. The
category cob(n) has a monoidal structure coming from disjoint union. The identity element
for this monoidal structure is the empty (n 1)-manifold. In cob(n), there are natural

isomorphisms : 1C t
, etc. Even better, cob(n) is a symmetric monoidal category.

In other words, there is a natural transformation g1,2 : 1 t 2

2 t 1 satisfying
g1,2 g2,1 = 1.
Definition 3.2.2. Let C, D be monoidal categories. We call a functor F : C D monoidal

if for all X, Y C, there are natural isomorphisms F X F Y

F (X Y ) and F 1C
1D .
Moreover, we require that F commute with the associativity constraints of C and D.
197

3.3

Topological quantum field theories

We think of a TQFT as assigning to each manifold M a vector space F (M ) of fields,

together with an action S : F (M ) R and some functional . One usually writes
Z

S=
L ,
x
M


.

Definition 3.3.1. A topological quantum field theory is a monoidal functor cob(n) Vectk
for some field k.
If we are thinking of physical intuition, one should set n = 4. To each , associate to
the space Fun(O()), i.e. the space of all (not necessarily linear) functions O() R. To
each cobordism M : 0 1 , associate an integral operator ZM : O(0 ) Fun(O(1 ))
defined via the kernel
Z
KM (1 , 2 ) =

O(M )
|0 =1
|1 =2

eiS() .

(better: introduce space of distributions)

For example, consider
Z
1
2 1
2 2
K = xe 2 mx + 2 mu x .
In physics, all the fundamental parameters ~, mass, coupling, etc. are contained in the
Lagrangian, and in S.
If M is a closed n-manifold, we can think of M as a cobordism from the empty space
to itself. By convention, O() = 0, so Fun(O()) = R. So ZM : R R is just a number,
which turns out to be
Z
eiS() .

K=
O(M )

Example 3.3.2 (Witten, Dijkgraaf). Let G be a finite group. Let M be an n-cobordism,

and let F(M ) be isomorphism classes of principal G-bundles on M . In this case, for
M = 0 t 1 , the kernel is
X

KM (0 , 1 ) =

|0 =0
|1 =1
F (M )

1
.
# Aut()

Moreover,
ZM (f ) =

KM (0 , 1 )f (0 ).

0 F (0 )

198

3.4

Chern-Simons

Let M be a three-dimensional spacetime manifold, G a Lie group, and let

2
CS = A dA + A A A
3
where A is some connection on the trivial G-bundle on M . Our functional is
Z
1
2
S(A) =
tr(A dA + A A A).
i M
3
We have

3.5

S
A

= FA = 0, so the space of solutions consists of flat connections.

Atiyah-Segal axioms

Recall that a TQFT is a functor Z : cob(n) Vectk for some field k.

Perfectoid rings, almost mathematics, and the cotangent

complex

Scholze introduced perfectoid fields (and more generally, perfectoid spaces) in his paper
[Sch12], in which he proved a wide range of special cases of Delignes Weight Mondromy
Conjecture for p-adic fields.

4.1

Perfectoid fields

Recall that a valued field is a field k together with a homomorphism | | : k for

some totally ordered abelian group (whose operation we will write multiplicatively). One
requires | | to satisfy the ultrametric inequality:
|x + y| 6 max{|x|, |y|}.
As is standard, write 1 for the unit in , and introduce an extra element 0, stipulating that
0 < for all . We define |0| = 0, thus extending | | to a function k {0}. One
calls the rank of | | the dimension dimQ (|k | Q). We will say that the valuation | | is
non-discrete if |k | 6' Z.
Put
k = {x k : |x| 6 1}.
This is called the ring of integers of k. It is a valuation ring with (unique) maximal ideal
k + = {x k : |x| < 1}.
We call k /k + the residue field of k.
199

Definition 4.1.1. A perfectoid field is a complete valued field k with respect to a nondiscrete rank-one valuation, with residue characteristic p > 0, such that the Frobenius
Fr : k /p k /p is surjective.
A typical example of a perfectoid field is Qp (p ) , the completion of Qp (pn : n > 1)

with respect to the p-adic topology. Similarly, Qp (p1/p ) and Cp = Qp are perfectoid.

An example in characteristic p is the t-adic completion of Fp (t1/p ). For a perfectoid field k

of residue characteristic p, choose k with |p| 6 || < 1. Note that the Frobenius map
a 7 ap is defined on A/ for any k -algebra A.
Definition 4.1.2. Let k be a perfectoid field. A perfectoid k-algebra is a Banach k-algebra
A such that A = {x A : |x| 6 1} is open and for which Fr : A / A / is surjective.
If k is a perfectoid field, let Perf(k) denote the category of perfectoid k-algebras, with
continous k-maps as morphisms. We will construct, for any perfectoid field k with residue
characteristic p, a perfectoid field k [ of characteristic p. Start by defining
(
)
Y
p
k [ = lim(k /) = (xi )
k / : xi+1 = xi .

Fr

i>0

It is not too difficult to check directly that k [ is a valuation ring, and we put k [ = Frac(k [ ).
There is a canonical map ()] : k [ k defined by
n

(x0 , x1 , . . . )] = lim x
fn p ,
n

where x
fn is an arbitrary lift of xn k / to k . The map ()] is not additive unless k already
has characteristic p, in which case k = k [ . In general, ()] extends to an isomorphism
of multiplicative groups k [ k , and we can use this to define a valuation on k [ by
|x|k[ = |x] |k . See Lemma 3.4 of [Sch12] for a proof that ()] has the claimed properties,
and that k [ is a perfectoid field with the same value group as k.
Example 4.1.3. Let N be a lattice (i.e. a finite free Z-module) and let NR be a strongly
convex polyhedral cone. Let NR be its dual. If we put M = N , then the spectra of
algebras of the form k[] = k[ M ] form affine charts for toric varieties over k. There is
a perfectoid version of this. Write khi = kh M [ p1 ]i for the ring


1

k [ M Z[ ]]
k.
p
Then khi is a perfectoid algebra over k. (Note: khi is not, in this context, a noncommutative polynomial algebra over k.
If A is a perfectoid k-algebra, define A[ in much the same way, via
!
A[ =

lim A /

Fr

200

k[ k [ .

It turns out that if A is perfectoid, then A[ is also perfectoid, and we have the following
deep theorem:
Theorem 4.1.4 (Scholze). The functor ()[ : Perf(k) Perf(k [ ) is an equivalence of
categories.
In fact, much more can be shown, e.g. ()[ induces equivalences of categories between
FEt(A) and FEt(A[ ) for all A, where FEt(A) is the category of finite etale algebras over A
(it turns out that such algebras are perfectoid).
Example 4.1.5. If A = khi = kh M [ p1 ]i as in Example 4.1.3, then A[ = k [ hi =
k [ h M [ p1 ]i.
Theorem 4.1.4 is proved without introducing much heavy machinery in Section 3.6 of
[KL13]. The basic idea is that the map ()] : k [ k induces a ring homomorphism
: W (k [ ) k , where W () is the ring of p-typical Witt vectors. The inverse to the
functor ()[ is A] = W (A ) W (k[ ) k. Scholzes proof is more conceptual, and passes
through a diagram
Perf(k)


/ Perf(k a )

/ Perf(k a /)

/ Perf(k [a )

/ Perf(k [a / [ )

()[

Perf(k [ )

in which most of the categories have yet to be defined. The superscript ()a should be read
almost, following the almost mathematics initially created by Faltings, and developed
systematically in [GR03].

4.2

Almost mathematics

Almost mathematics was first introduced by Faltings in [Fal88], where he proved a deep
conjecture of Fontaine on the etale cohomology of varieties over p-adic fields. We follow the
treatment in Section 2.2 of [GR03].
Let V be a valuation ring with maximal ideal m. Throughout this section, we assume
that the value group of V is non-discrete. This implies m2 = m. In fact, much of the theory
works for any unital ring V with idempotent two-sided ideal m, but we have no need to
work at this level of generality. The reader should keep in mind the example V = k for a
perfectoid field k.
Let Mod(V ) be the category of all V -modules, and let Ann(m) be the full subcategory
consisting of those modules killed by m.
Lemma 4.2.1. Ann(m) is a Serre subcategory of Mod(V ).
Proof. We need to show that if 0 M 0 M M 00 0 is an exact sequence of V -modules
for which M 0 and M 00 are killed by m, then M is also killed by m. We trivially have m2 M = 0,
but m2 = m, whence the result.
201

Definition 4.2.2. The category of V a -modules is the Serre quotient Mod(V a ) = Mod(V )/Ann(m).
Write ()a : Mod(V ) Mod(V a ) for the quotient functor. Even though V a does
not exist as a ring we will write HomV a (, ) for hom-sets in Mod(V a ). While in general,
hom-sets in quotient categories can be difficult to describe, the category Mod(V a ) is relatively
easy to understand via the following theorem.
Theorem 4.2.3. There is a natural isomorphism HomV a (M a , N a ) = HomV (m M, N ).
It follows that Mod(V a ) is naturally a Mod(V )-enriched category. We define two functors
()! , () : Mod(V a ) Mod(V ):
M = HomV a (V a , M )
M! = m M .
Theorem 4.2.4. The triple (()! , ()a , () ) is adjoint. Moreover, these adjunctions induce
natural isomorphisms (M )a = M = (M! )a .
This is suggestive of the situation in which j : U , X is an open embedding of topological
spaces. The restriction functor j : Sh(X) Sh(U ) fits into an adjoint triple (j! , j , j ) in
which j j = 1 = j j! . So we should think of Mod(V a ) as the category of quasi-coherent
sheaves on some subscheme of Spec V , even though there is no such subscheme. In fact,
since Mod(V a ) does not contain enough projectives, it should be thought of as some kind of
non-affine object.
The category Mod(V a ) inherits the structure of a tensor category from Mod(V ). In fact,
we have internal tensor and hom defined by
M a N a = (M N )a
Homa (M a , N a ) = Hom(M, N )a
There is a tensor-hom adjunction Hom(L N, M ) = Hom(L, Homa (M, N )). This allows us
to speak of algebra objects in Mod(V a ) as commutative unital monoid objects in the tensor
category (Mod(V a ), ). We call such objects V a -algebras. It is easy to check that they are
all of the form Aa , for A some V -algebra.
If A is a V a -algebra, we can form the category Mod(A) of A-modules in the obvious way.
This is also an abelian tensor category, so it makes sense to speak of flat objects in the
usual way, i.e. an A-module M is flat if the functor M A is exact.
Definition 4.2.5. Let k be a perfectoid field. A perfectoid k a -algebra is a flat, -adically
complete k a -algebra A for which Fr : A/ 1/p A/ is an isomorphism.
Note that even though 1/p may not exist as an actual element of k , the ideal ( 1/p )
is well-defined, so it makes sense to write M/ 1/p if M is any k - (or k a )-module. Write
Perf(k a ) for the category of perfectoid k a -algebras.

Theorem 4.2.6. The functor A 7 Aa induces an equivalence of categories Perf(k)

Perf(k a ).
Idea of proof. See the first part of Section 5 in [Sch12]. Besides some technicalities, one has
the existence of an inverse functor A 7 A k k.
202

4.3

The cotangent complex

We have seen in the last section that localization induces an equivalence between the category
of perfectoid k-algebras and the category of perfectoid k a -algebras. Since k and k [ are
canonically isomorphic via ()] , also write for the element of k [ corresponding to k.
It is easy to check that k / = k [ /. It easily follows that Mod(k a /) = Mod(k [a /).
The idea of this section is to pass from Perf(k a ) to a suitable category of perfectoid
k a /-algebras.
Definition 4.3.1. Let k be a perfectoid field. A k a /-algebra A is perfectoid if it is flat
and Fr : A/ 1/p A is an isomorphism.
Let Perf(k a /) denote the category of perfectoid k a /-algebras. We will show that the
functor A 7 A/ from Perf(k a ) to Perf(k a /) is an equivalence of categories by introducing
an almost version of the cotangent complex. Lets start by recalling the classical theory:
Theorem 4.3.2. There is a functorial way of assigning to a flat map A B of (commutative, unital) rings an object (the cotangent complex) LB/A of D60 (B). This complex satisfies
the following properties:
e  A is a
1. There is a functorial way of assigning to a square-zero extension 0 I A
square-zero extension an obstruction class
o(I) ext2 (LB/A , IB )
e
e such that B
e e A = B.
such that o(I) = 0 if and only if there is a flat A-algebra
B
A
e a deformation of B to A.)
e
(One calls such a B
e exists, then the set Def Ae(B) of deformations of B to A
e is
2. If a deformation of B to A
A
a
ext1 (LB/A , IB )-torsor.
e and B
e 0 are deformations of A-algebras B, B 0 to A,
e and if f : B B 0 is an
3. If B
A-algebra map, then there is a functorial way of assigning to f an obstruction class
o(f ) ext1 (LB/A , IB 0 )
e
e e
e0
such that the set Def A
A (f ) of isomorphism classes of lifts of f to f : B B is nonempty
if and only if o(f ) 6= 0.

e exists, then Def Ae(f ) is a Hom(LB/A , IB 0 )-torsor.

4. If a lift of f to A
A
Proof. See Proposition 2.1.2.3 in Chapter III of [Ill71] for parts 1 and 2.
Gabber and Ramero were able to generalize the classical theory of the cotangent
complex to an almost setting. To be precise, Theorem 4.3.2 remains true if we work V a algebras, for any non-discrete valuation ring V . In other words, there is a canonical object
LaB/A D60 (B) such that the theore still works.
203

Theorem 4.3.3 (Scholze). Let k be a perfectoid field. If A is a perfectoid k a /-algebra,

then LaA/(ka /) = 0 as an object of D(A).
Idea of proof. This is a deep result, but at the heart of its proof is the fact that if R is a
smooth perfect Fp -algebra, then LR/Fp = 0. Smoothness yields LR/Fp = 1R/Fp [0], and from
d(rp ) = pdrp1 = 0 we see that 1R/Fp = 0.
Corollary 4.3.4 (Scholze). The functor A 7 A/ induces an equivalence of categories

Perf(k a )
Perf(k a /).
Proof sketch. Let An = k a / n+1 . We content ourselves with showing that objects and
morphisms in Alg(A0 ) lift uniquely to each Alg(An ). Let B0 be a perfectoid A0 -algebra.
Theorem 4.3.3 tells us that LaB0 /A0 = 0, so B0 and any morphisms from B0 to other perfectoid
A0 -algebras lift uniquely to A1 by the almost version of Theorem 4.3.2. All that remains
for the induction to work is to show that LaBn /An = 0 implies LaBn+1 /An+1 = 0. There is an
exact sequence
n
/ B0 / Bn+1

/ Bn

/ 0,

and general theory gives us an exact triangle in D(An+1 ):

/ La

LaB0 /A0

Bn+1 /An+1

/ La
Bn /An

/.

Since LaB0 /A0 = 0 by Theorem 4.3.3 and LaBn /An = 0 by assumption, we get that LaBn+1 /An+1 =
0.

Nearby and Vanishing Cycle Functors

We work in the following setting: let X be a complex-analytic manifold, f : X C a

non-constant analytic map, for which we suppose that Z = f 1 (0) is non-empty. Put
U = X r Z. Throughout, we work with the middle perversity p1/2 (S) = 12 dimR S.
Denote by D(X) the bounded derived category Dbc (X, C) of constructible sheaves of
complex vector spaces. Let Perv(X) = Pervp1/2 (X, C) be the corresponding category of
perverse sheaves. Recall that Perv(X) is an abelian category in which every object has finite
length.

5.1

Construction

The following is originally from Delignes section La formalisme des cycles evanescents in
[DK73].
Consider the diagram
Z


{0} 

/Xo

/Co

204

? _U o

? _ C o p

f
X


f
C

f C is the universal covering; and X

f . We denote the projection
f = X C C
where p : C

f
X  X by .
f , C) : The functor is
Recall that we have the adjunction : Sh(X, C)  Sh(X
exact, and is left-exact, so we have an adjunction at the level of derived categories:
f ) : R .
: D(X)  D(X
Definition 5.1.1. The nearby cycles functor is the functor f : D(X) D(Z) defined by
f (F ) = i R (F ).
f ) = U X, the functor f factors through D(U ). In other words,
Note that since (X
there exists f : D(U ) D(Z) such that f = f j . In other words, f (F ) only
depends on F |U . Moreover, f comes with a natural transformation
: i f
defined by applying i to the adjunction transformation 1 R = f
Definition 5.1.2. The vanishing cycles functor is the functor f : D(X) D(Z) given by
f (F ) = cone(F : i (F ) f (F )).
Proposition 5.1.3. Both f and f restricts to functors on the corresponding subcategories
of perverse sheaves:
f : Perv(U ) Perv(Z)
f : Perv(X) Perv(Z).

5.2

Monodromy

f C be the universal cover. The fundamental

As in the previous section, let p : C
f by deck transformations. Denote by t : C
f C
f the deck
group 1 (C ) = Z acts on C

transformation corresponding to the generator 1 of 1 (C ). The transformation t satisfies

pt = p. If we base-change to X, we have a commutative diagram
/ X

It follows that = , whence we have a natural isomorphism at the level of derived

functors:
R R = R .
205

The standard adjunction gives us a morphism of functors : id R . Thus we have

a commutative diagram
R

/ R R

where T : f f is the endomorphism coming from . One calls T the monodromy

operator.
Moreover, T restricts to identity under : i f , i.e. it extends to a morphism of
triangles:
i

/ f

/ f

/ f

/ i [1]


/ i [1].

For any F D(X), this gives us a morphism of triangles in DZ :

F

i (F )

/ f (F ) canF / f (F )


T id

/ f (F )

/ i (F )[1]

varF

f (F )


/0

It follows that
var can = T id.

5.3

Gluing problems

Given i : Z , X - U : j as above, we want to glue Perv(U ) and Perv(Z).

Definition 5.3.1. Define a category Glue(Z, U ) whose objects are quadruples (G , H , a, b),
where:
1. G Perv(U )
2. H Perv(Z)
3. a : f (G ) H is a morphism in Perv(Z)
4. b : H f (G ) is a morphism in Perv(Z)
5. b a = TG id
Theorem 5.3.2. The functor Perv(X) Glue(Z, U ) given by
F 7 (j F , f (F ), canF , varF )
is an equivalence of categories.
206

Example 5.3.3. Consider Z = {0} , X = C - U = C , via f (z) = z. One has that

Perv(Z) = Vectfin
C D(VectC ). The category Perv(U ) consists of locally constant sheaves on
C , i.e. Perv(U ) ' RepC (1 (C )), which is just Modfin (C[t ]). To be explicit,
n
o

Perv(U ) = (V, A) : V Vectfin

and
V
:
A

V
.
C
The functor f acts via (V, A) 7 V and T(V,A) : f (V, A) f (V, A) is the operator A.
Theorem 5.3.2 now says that
Perv(X) = {(W, V, E : V W, F : W V ) : F E + id is invertible} .
There is a general way of assigning an algebra q,h (Q) (called the multiplicative preprojective
algebra) to a quiver Q such that
Perv(X) = q,h ( ).
Example 5.3.4. Let n > 1, let X = Cn , and let S = {XI }I[n] , where
XI = {(x1 , . . . , xn ) Cn : xi = 0 if and only if i I}.
Let Perv(X) be the category of perverse sheaves on X with respect to S. Combinatorialists
are interested in perverse sheaves because of the following theorem. We first define a category
A(n) whose objects are tuples


{VI }I[n] , EI,i : VI VI{i} for i
/ I, FI,i : VI{i} VI for i
/I
such that
1. EI{j},i EI,j = EI{i},j EI,i for i
/I
2. FI,j FI{i},i = FI,i FI{i},j for i 6= j both not in I
3. FIr{j},i IFIr{j},j = FI{i}r{j} FI,i for i
/ I and j I
4. FI,i EI,i + id is invertible for all i
Morphisms are {fI : VI VI0 } commuting with the Es and F s.
Theorem 5.3.5. With X as above, there is an equivalence of categories Perv(X) ' A(n) .
Proof. See [GMV96].
Example 5.3.6. Let X = C2 and {0} {xn = ym} C2 be our stratification (we assume
n 6 m). Then there is an equivalence of categories Perv(X) ' A(n,m) , where A(n,m) consists
of tuples (A, B1 , . . . , Bn , C) with arrows
Ao

qk
pk

/ Bk o

Bk+m

sk
rk

k
(mod n)

207

/C

satisfying certain requirements, including that 1 + qk pk and k are all invertible.

As a special case, if we put (n, m) = (2, 3), define
H = ChA, Bi/(ABA + A2 + A, BAB + B 2 + B).
Define H by the algebra homomorphism H H , which is universal among algebra
homomorphisms : H R such that (A + 1) and (1 + B) are units. Alternatively,
H = HhX, Y i/((1 + A)X = X(1 + A) = 1, (1 + B)Y = Y (1 + B) = 1).
We have A(2,3) ' Modfg (H ).

5.4

Monadology

Recall that an exact category (in the sense of Quillen) is an additive category A together
p
i
with a special class E of exact pairs, consisting of diagrams of the form A
B
C such
that (A, i) = Ker(p) and (C, p) = Coker(i). These exact pairs are required to satisfy certain
axioms. [reference earlier section for these]
Example 5.4.1. If A is an abelian category, then we can let E be the class of all short exact
sequences. More interestingly, we could restrict to the class of all split exact sequences.
Example 5.4.2. The category of finitely-generated projective modules over a unital ring is
exact (when given the obvious class of exact pairs).
Example 5.4.3. Let A = Com(B) for some abelian category B. We can set E to be the
class of termwise-split short exact sequences in A.
In an exact category, there is are notions of admissible monic (a morphism which fits
into the first half of an exact pair) and admissible epic (a morphism which fits into the
second half of an exact pair). We write , for admissible monics and  for admissible epics.
Fix an exact category A.

Definition 5.4.4. A monad in A is a complex of the form P = [P P P+ ] with

admissible monic and + admissible epic.
By complex we require + = 0. Denote by H(P) the quotient Ker(+ )/ Im( )
[prove that this quotient exists]. We call this the cohomology of the monad P. Let Ae
be the category of monads in A. This is naturally an exact category with H : Ae A an
exact functor.
1
0
Let Ae1 be the category whose objects are P1 = [P1 , P0 ,
P1 ]. We call Ae1 the
( , )
(+ ,+ )t
category of 3-step filtations. Finally, let Ae2 be the category of [L AB L+ ]
with monic and + epic.
Proposition 5.4.5. There are canonical exact equivalences of categories Ae ' Ae1 ' Ae2 .
208


+
Proof. We consider the equivalence Ae ' Ae1 . To a sequence P P P+ associate the

0
sequence P Ker(+ ) , P , and in reverse send P1 , P0 ,
P1 to the sequence
0 1
P1 P1  P1 /P0 .
For Ae1 ' Ae2 , consider

(0 ,can)

[P1 P0 P1 ] 7 [P0 P1 P0 /P
1 /P1 ]
...

The motivation for all of this is quite classical. Recall the following theorem in [Bar77].
Let M(n, r) be the moduli space of framed torsion-free coherent sheaves on P2C of rank r
and Chern class c2 = n. Let i : ` , P2 be the inclusion of the line at infinity. Elements
of M(n, r) are pairs (F , ), where F is a rank-r torsion free sheaf on P2C with c2 (F ) = n,

and : i F
OPr
1 is a framing at infinity. There is a natural way to give M(n, r) the
structure of a scheme.
Theorem 5.4.6 (Barth). There is an isomorphism between M(n, r) and the moduli space
of isomorphism classes of triples (B1 , B2 , i, j), where B1 , B2 Mn (C), i : Cr Cn and
j : Cn Cr , satisfying [B1 , B2 ] + ij = 0. Also, the triples are stable in the sense that there
are no S Cn such that B (S) S for {1, 2} and Im(i) S.
The linear-algebraic data in this theorem can be encoded in terms of the quiver [couldnt
live-TEXthe quiver: it has i : W V , j : V W and Bi : V V ]
Sketch of proof. We will construct a functor in one direction. To a tuple (B1 , B2 , i, j), assign


a
b
F = H V OP2 (1)
(V OP2 )2 W OP2
V OP2 (1) .
We can identify P2C with the proj of C[z0 , z1 , z2 ], where we can regard the zi as global
sections of OP2 (1). Even better, we think of the zi as maps OP2 (i) OP2 (i + 1) for all i.
The maps a and b are given by

z0 B z1
a = z0 B 2 z 2
z0 j


b = (z0 B2 z2 ) z0 B1 z1 z0 i
One can check that ba = 0 if and only if [b1 , b2 ] + ij = 0.
Now we define the operation F 7 (B1 , B2 , i, j). Define a rank-two vector bundle E on
P2 by the Euler exact sequence
0 OP2 (1) OP3
2 E 0.
209

We put


F 7 H1 (P2 , F (2)) OP2 (1) H1 (P2 , F (1) E ) OP2 H1 (P2 , F ) OP2 (1) .

There is an example from [BW02]. Let r = 1 and n be any integer. Let Cn be the space
of isomorphism classes of tuples (X, Y, i, j) such that X, Y Mn (C), i Cn and j (Cn )
such that [X, Y ] + 1 = ij. This is generally known as the Calosero-Moser space.
`
There is a bijection between n>0 Cn and the set of isomorphism classes of right ideals
in A = A1 (C) = ChX, Y i/([X, Y ] 1).

5.5

Reflection functors

Let A be an exact category.

Definition 5.5.1. The category A] of dyads consists of commutative diagrams
C 



/A



/ C+

where we do not require that + = 0.

Let A]1 be the category of special monads, i.e. those of the form
C+ 

 /
+
/ / C+
AB

where + = 0. Finally, let A]2 be the category of special exact sequences, i.e. those of the
form

+
2

+
2
1
2
+
 ( , , )
/ A B1 B2
/ / D+
D
i ) is mmonic for each i, and such that ( , i ) is epic for each i.
such that ( ,
+ +

Proposition 5.5.2. There are canonical exact equivalences A] ' A]1 ' A]2 .
Proof. We first construct a functor A] A]1 . Send a dyad to the monad

C 

 ( , )

/AB

210

+
+

/ C+ .

For the functor A]1 A]2 , send a monad Q1 to

Ker(+ ) , A B H(Q1 )  Coker( ).

A key observation is that on A]2 , here is a natural auto-equivalence r : A]2 A]2 , sending
Q2 to r(Q2 ), which is
D A B 2 B 1 D+ .
Our equivalences transport this to an auto-equivalence of A] , which we call reflection.A
dyad
 /
A

C 


B
is sent by r to

Ker(+ ) 




/ C+

/A

H(Q1 )


/ Coker( )

0
+

where
0

= Ker(+ ) , Ker(+ , + )  H(Q1 )

0
+
= Ker( , )  B Coker( ).

Exercise 1. Check that any Barth or C-M monad is naturally special. Compute the
reflection functor on the corresponding dyads.
Exercise 2. Recall the moduli space M(n, 1) of torsion-free coherent sheaves on P2C of

rank one, framed over ` . In other words, we specify : i F

OP1 . We claim that
M(n, 1) ' Hilbn (C2 ). To see this, first identify Hilbn (C2 ) with the space of isomorphism
classes of triples (X, Y, i) where X, Y Mn (C), i Cn , [X, Y ] = 0, and i generates C2 as a
C[X, Y ]-module. This is clearly just the space of isomorphism classes of (V, i), where V is
an n-dimensional C[X, Y ]-module and i V is a cyclic vector. This in turn is isomorphism
classes of surjections C[X, Y ]  V . Recall Barths description of F M(n, 1) in terms of
monads. Every such sheaf is the cohomology of a monad of the form
OP2 (1) V

/ (O 2 V )2 O 2
P
P

/ O 2 (1) V.
P

Via the restriction functor j : coh(P2C ) coh(C2 ) ' mod(C[X, Y ]), this corresponds to
0

/ A0 V


/0

/ (A0 V )2 A0


/ A0

211

/ A0 V


/V

/0

/0

where we write A0 = C[X, Y ]. Everything comes down to showing that (e

, ) is a quasiisomorphism of A0 -modules.
Exercise 3. Replace A0 = C[x, y] = Chx, yi/[x, y] by the Weyl algebra A~ = Chx, yi/([x, y] =
~). In fact A~ ' A1 for ~ 6= 0. We replace Hilbn (C2 ) with the C-(?)-Moeser space Cn of
quadruples (X, Y, i, j) satisfying [X, Y ] + ~ = ij. We have an exact sequence
0

/AV

/ (A V )2 A


/AV



/A

/0

/0

/V

/ 0.

The claim is that there is such a quasi-isomorphism in the category mod (A) of A -modules
over A.
This is worked out in the paper [BC07]. It reflects the fact (from commutative algbra)
that any `
ideal in C[x, y] is isomorphic to a unique ideal of finite codimension. So we can
think of n>0 Hilbn (C2 ) as the moduli space of all ideals in C[x, y]. Similarly, we could
`
L A
interpret n Cn as the space of isomorphism classes of right ideals in
2 ~ .

Abstract Nearby Cycles functor

6.1

The setup

Consider a general recollement of triangulated categories:

DZ e

%
/ DU
9

/ DX

We recall the following

0
0
0
Theorem 6.1.1 ([BBD82] ). For any two t-structures (D0
Z , DZ ) and (DU , DU ) There
0
exists a unique t-structure (D0
X , DX ) compatible with both.
0
0
0
0
0
We denote by PervZ = D0
Z DZ , PervX = DX DX , PervU = DU DU and call
them the abelian categories of perverse sheaves.
Note that for any M PervX , we have:

i! M D0
Z
i M D0
Z
Definition 6.1.2. M PervX is called a tilting perverse sheaf if
i M PervZ

and

i! M PervZ

Lemma 6.1.3. Given any M PervX , then M is tilting perverse if and only if both of the
followings hold:
212

1. j j M and j ! j! M are both in PervX

2. The adjunction maps
M j j M

is surjective

j! j M M

is injective

Proposition 6.1.4. Let MU PervU be such that j (MU ) PervX and j! (MU ) PervX .
Then there exists a tilting perverse sheaf M PervX such that j M
= MU .
Sketch of proof. Define A, B PervX by the following exact sequence:
M

U
0 A j! MU
j MU B 0

Case 1. Ext2 (B, A) = 0

Then we have a three-step filtration, and this allows us to construct a minimal tilting
extension
M = (MU )
Case 2. Ext2 (B, A) 6= 0
This case requires more work.
Denote by Pervmte (U ) the full subcategory of Perv(U ) consisting of objects MU which
have a minimal tilting extension (MU ).
Then (MU ) is functorial in MU Pervmte (U ).
Denote by Pervmte (X) the full subcategory of Perv(X) consisting of those objects M
such that j M Pervmte (U ).
Definition 6.1.5. Define a category Glue(Z, U ) whose objects are quadruples (MU , , , ),
where:
1. MU Pervmte (U )
2. Perv(Z)
3. : i! (MU ) is a morphism in Perv(Z)
4. : i (MU ) is a morphism in Perv(Z)
5. = : i! (MU ) i (MU ) is the canonical map
Theorem 6.1.6. There is a canonical commutative diagram of equivalences of categories:
Pervmte X

/ Glue(Z, U )

diads

where:
213


/ diads

1. r : diads diads is the reflection functor.

2. Q is defined by the following:
Given M Pervmte (X), define the diad

Q(M ) :=
j! j M

(j M )
can

can

can

can

%%

j9 j M

3. R is defined by the following:

Given (MU , , , ) Glue(Z, U ), define the diad

(MU )
8

can

+
!
R(MU , , , ) := i i ((MU ))

&
i

6.2

&&

i i (MU )
8

can

Multiplicative Preprojective Algebra Associated to a Quiver

The main references are [CBS06] and [BRT13].

Let Q = (I, Q) be a finite quiver, where I denotes the set of vertices and Q denotes the
set of arrows.
Denote Q as the double quiver of Q. i.e.,
1. Vertices are the same : I(Q) = I(Q)
2. Arrows are doubled in the opposite direction.
Any construction on a quiver that does not depend on orientation should be a construction
on the double quiver
For any arrow a Q, we its double (in the opposite direction) by a Q. By defining

a = a, we have an involution
() : k(Q) k(Q)
a 7 a
214

Also, define  : Q {1} by


(a) =

1
:aQ
1 : a Q

Given two sets of parameters:

{qv }vI ,

{~a }aQ

Extend this set of parameters to Q by setting ~a = ~a

Definition 6.2.1. The multiplicative preprojective algebra q,~ (Q) associated to the quiver Q
and the sets of parameters q = {qv }vI , ~ = {~a }aQ is defined as the algebra homomorphism
that is initial among all algebra homomorphism : k(Q) R satisfying:
1. (aa + ~a ) is a unit in R

Y
X
2.
(aa + ~a )(a)
qv ev = 0
vI

aQ

Lemma 6.2.2. Such q,~ (Q) always exists for any quiver Q and any sets of parameters
q = {qv }vI , ~ = {~a }aQ .
In [CBS06], they defined the preprojective algebra only for the case ~a = 1 for all a Q.
Example 6.2.3. Consider the one-loop quiver. (i.e., the quiver Q = Qloop with one vertex
v and one arrow a).
Write qv = q and ~a = ~.
Then the multiplicative preprojective algebra is:
k(Q) q,~ (Q) =

khx, y, (~ + xy)1 i
(xy qyx ~)

Note that when ~ = 0, we have:

q,0 (Q) =

khx1 , y 1 i
(xy qyx)

while for ~ = 1, we have:

q,1 (Q) =

khx, y, (1 + xy)1 i
(xy qyx 1)

Example 6.2.4. Consider the one-arrow quiver.

h
i
a
Q =
0
215

Then the multiplicative preprojective algebra is:

q,~ (Q) =

khe0 , e , a, a , (aa + ~)1 , (a a + ~)1 i

(aa + ~ = q0 e0 , (a a + ~)1 = q e , e0 e = e e0 = 0,

e2i = ei )

while for ~ = 1, we have:

q,1 (Q) =

(e +

a a

k(Q)
= q e , e0 + aa = q0 e0 )

Theorem 6.2.5. We have a canonical bijection:


h
i
Y
a
Repf.d. q,1 (
0 )
(q0 ,q CC )

= {F PervC : j (F ) has scalar monodromy}

Example 6.2.6. Let Q be the quiver with two vertices 0 and , one loop x from 0 to 0,
and one arrow i from to 0.
Denote their opposite arrows by y = x and j = i .
Set ~x = ~y = 0, ~i = ~j = 1, (q0 , q ) = (q, q n ) for some q n 6= 1.
Now we have:
k(Q )


q,~ (Q )

e 70/ /

k((Qloop ))


/ / q,~ (Qloop )

khx1 ,y 1 i
(xyqyx)

Theorem 6.2.7. There is a natural recollement set-up:

r

Db (Modq,~ (Qloop ))
l

/ Db (Modq,~ (Q ))

,
/ Db (ModUq,~ (Qloop ))
2

Moreover, the map

i :



Rep q,~ (Q ), n = (1, n) //GL(n)

n=0

{Isomorphism class of ideals of Aq }

is the Calogero-Moser map.

216

Appendix A

Miscellaneous topics
1

Characteristic classes of representations (after Quillen)

Let A be an associative unital ring. Consider the (discrete) group GLn (A) = Mn (A) of
invertible n n matrices with coefficients in A. We call GLn (A) the general linear group
over A. We have embeddings GLn (A) GLn+1 (A) given by


0
7
.
0 1
We define
GL(A) = GL (A) = lim GLn (A).

Put
H (GL(A), k) =

Hi (GL(A), k) = H (BGL(A), k),

i>0

where BGL(A) is the classifying space of GL(A) (see 2.5.3). The diagonal map GLn (A)
GLn (A) GLn (A) gives H (GL(A), k) the structure of a coassociative cocommutative
coalgebra over k.
Note that H (GL(A), k) has also the structure of a graded-commutative k-algebra. There
is a natural map taking direct sum : GLn (A) GLm (A) GLn+m (A), defined by


A 0
AB =
, A GLn (A), B GLm (A)
0 B
It induces a map : GL(A) GL(A) GL(A), which in turn induces a map BGL(A)
BGL(A) BGL(A). The latter map turns out to be associative and commutative up to
homotopy. It follows that H (GL(A), k) naturally has the structure of a graded k-algebra.
We want to define a functor of points GrComAlgk Set for which the representing
object is H (GL(A), k). First we need to extend the notion of representation. Let G be a
(discrete) group. We can think of a group G as a category (actually, groupoid) with one
object.
217

Let A be any (small) additive category. Let Isom(A) be the groupoid of isomorphisms
in A. That is, the only morphisms in A which we keep in Isom(A) are isomorphisms. Let
S = 0 (Isom(A)) be the set of isomorphism classes of objects in A.
Definition 1.0.8. A representation of G in A is a functor : G Isom(A).
We denote by Isom(G, A) the groupoid Isom(Fun(G, A)) of isomorphism classes of representations of G in A. That is,
a
Isom(G, A) =
Hom(G, Out(Ps )),
sS

where Ps is a representative of the class of s S.

Exercise Check that if G and G0 are groups then Isom(Fun(G, G0 )) is the quotient of
Hom(G, G0 ) by inner automorphisms in G0 , i.e.
Isom(Fun(G, G0 )) = Hom(G, G0 )/ Inn(G0 ).
Definition 1.0.9. We say that representation : G Isom(A) is trivial if f G,
(f ) = id HomIsom(A) ((), ()), where Ob(G) is the unique object of G.
Definition 1.0.10. Two representations E, E 0 : G Isom(A) are stably isomorphic if
E ' E 0 0 , where and 0 are trivial representations.
Define St(G, A) = Isom(G, A)/ , where E E 0 if E and E 0 are stably isomorphic.
Recall that S = 0 (Isom(A)) is the set of isomorphism classes of objects in A. The set
S is naturally a commutative monoid, via [P ] + [Q] = [P Q]. We denote by S n S the
category with Ob(S n S) = S, and
HomSnS (s, s0 ) = {t S : s0 = s + t}.
There is a natural functor [G, Aut()] : S n S Set, which sends an object s to [G, Aut(Ps )].
Given a morphism t from s to s0 (i.e. s0 = s + t), there is a natural map u 7 i id from
Aut(Ps ) to Aut(Ps Ps0 ) ' Aut(Ps0 ). Denote by t the induced morphism [G, Aut(Ps )]
[G, Aut(Ps0 )].
Lemma 1.0.11. There is a natural isomorphism St(G, A) ' limSnS [G, Aut()].

We should think of Hom(G, Out(P )) as a functor from (S, S) to Set, where s 7

[G, Aut(Ps )] = Hom(G, Out(Ps )). An arrow (s, t) : s s0 is send to u 7 u id as a
function [G, Aut(Ps )] [G, Aut(Ps Pt )].
Example 1.0.12. Let A be an associative unital ring. Let A be the category of all finitelygenerated projective (right) A-modules. A general representation of G in A will be a
homomorphism G AutA (P ) for some P . Certainly the category A contains the objects
An . An automorphism of An is just an element of GLn (A), so a representation of G in
An is just a homomorphism G GLn (A).
218

There is a monoid homomorphism (N, +) (S, ) given by n 7 [An ]. This induces a

functor between groupoids F : N n N S n S. It turns out that this functor is cofinal. In
other words, for every s S n S, there exists n > 1 and a morphism f : s F (n) in S n S.
Indeed, if s = [P ] for some projective P , then since P is projective there exists another
projective Q such that P Q ' An for some n. Setting f = [Q], we have f : s F (n).
It follows that colimits over the category S n S are the same as colimits over the
subcategory N n N. This gives us the much more manageable description of St(G, A):
St(G, A) = lim[G, GLn (A)].

There is a canonical map : limn [G, GLn (A)] [G, GL(A)]. Concretely, let E : G

Aut(P ) represent an element of St(G, A). Choose Q with P Q ' An . We let E be the
composite
E

G
GLA (P ) GLA (P Q) = Aut(An ) = GLn (A) , GL(A).
Remark 1.0.13. Our construction is parallel to the topological situation. Namely, let X be
a paracompact topological space. We replace representations of G in A by complex vector
bundles on X. Let VB(X) be the set of`isomorphism classes of vector bundles on X. There
is an obvious decomposition VB(X) = n VBn (X), where VBn (X) consists of isomorphism
classes of n-dimensional vector bundles on X. In fact, VBn (X) = [X, BUn ], where BUn is
e 0 (X) := [X, BU ],
the classifying space of the group Un . Instead of St(G, A), we think of K
the (reduced) topological
K-theory. The analogue of our map : St(G, A) [G, GL(A)]
`
e 0 (X), which is an isomorphism if X is compact or a
is the map [X, n VBn (X)] K
finite-dimensional CW complex.
L
Let M = i>0 Mi be a graded abelian group. For an arbitrary group G, define
Y
H0 (G, M ) =
Hi (G, Mi ),
i>0

where we view each Mi as a trivial G-module.

Definition 1.0.14. A characteristic class (of representations) in A with coefficients in M
is a natural transformation of contravariant functors
: Isom(, A) H0 (, M )
between functors Grp Set. We call stable if (E) = (E) for any trivial representation
.
Now suppose M is a graded-commutative ring. Recall that graded commutativity means
ab = (1)|a||b| ba for any homogeneous a, b M . Then H0 (G, M ) has a natural ring
structure given by the cup product `: Hi (G, Mi ) Hj (G, Mj ) Hi+j (G, Mi+j ). Given
cocycles f1 : Gi Mi and f2 : Gj Mj , the cup-product f1 ` f2 is given by
[f1 ` f2 ](g1 , . . . , gi+j ) = f1 (g1 , . . . , gi )f2 (gi+1 , . . . , gi+j ).
219

Definition 1.0.15. A characteristic class is called exponential if (0) = 1 and (E1

E2 ) = (E1 ) ` (E2 ).
Let E : G Isom(A) be a representation. We apply to E the homology functor
H (E) : H (G) H (Isom(A)) = H (Nerve(Isom(A)), Z).
Observe that H0 (G, M ) = Hom0 (H (G), M ).
Let k be a field. Then we have constructed a functor GrComAlgk Set which assigns to
M the set of stable exponential characteristic classes with coefficients in G.
Theorem 1.0.16. Assume that X is nice. Then H (Isom(A)) represents this functor.
It turns out that H (Isom(A)) = H (GL (A), k). So, at least morally, H (GL(A), k)
is the universal algebra of stable exponential classes. More concretely, given such a
characteristic class : Isom(, M ) H0 (, M ), then given any f : M M 0 , we have
f

f : Isom(, M 0 ) H0 (, M )
H0 (, M 0 ). One says that f is induced from by f .
The theorem asserts that there is a natural bijection between stable exponential characteristic
classes in A with coefficients in M and algebra maps H (Isom(A)) M .

Generalized manifolds

The ideas in this example come from Freed and Hopkins paper [FH13].
Let Man be the category of smooth (finite-dimensional) manifolds and smooth maps.
A generalized manifold is a sheaf on Man. That is, a generalized manifold is a functor
F : Man Set such that whenever M Ob Man has an open cover {U }I , the following
diagram is an equalizer:
F (M )

F (U )

//

F (U U )

,I

The main idea is to try to extend differential geometry to the category of generalized smooth
manifolds. One example is differential forms. We have theVfunctor : Man Set, which
assigns to a manifold M the de Rham complex (M ) = 1 (M ).
g allows us to treat any manifold as a generalized
The Yoneda embedding Man , Man
manifold. Recall that the Yoneda lemma shows that there is a natural bijection
HomMan
g (hX , F ) ' F (X).
Definition 2.0.17. For any generalized manifold F , we can define the de Rham complex
of F by

(F ) = HomMan
g (F , ).
220

The Yoneda lemma tells us that this definition agrees with the usual one if F is
representable. Can we compute the de Rham cohomology
L of F ? We have a direct-sum
decomposition of functors on generalized manifolds, = q q . Let F = 1 : M 7 1 (M ).
It is a highly nontrivial theorem that (1 ) is isomorphic to
R

/R

/R

Thus HdR (1 ) = R, concentrated in degree zero. (This is Theorem 7.19 in [FH13].)

1
q
q
1
Recall that (by definition) q (1 ) = HomMan
g ( , ). So ( ) should be thought
of as a natural way to associate q-forms to 1-forms. For M fixed, 1 (M ), we have
M () q (M ), such that for any f : M 0 M , (f ) = f (M ).
In his paper [Che77] Chen generalized notion of differential forms to loop spaces. Namely,
take X to be a smooth manifold. Consider the path space P (X), which is a set of all smooth
paths : I X endowed with compact-open topology. Then P (X) is not strictly speaking
a smooth manifold, but we can still define what is a smooth map N P (X) and what is
differential form on P (X).
First, any map f : N P (X) defines unique map f: N I X given by f(n, t) =
f (n)(t). We call a map f smooth if he correspondent f is smooth.
We can define differential q-form on P (X) by assigning to each smooth map fN : N
, in such a way that for any morphism of manifolds
P (X) a q-form on N , denoted by fN
) = f is satisfied. This notion allows us
g : M N the compatibility condition g (fN
M
to define in the similar fashion de Rham algebra of differential forms on P (X).
It turns out that the Chens construction is coherent with the notion of generalized
manifold given in [FH13]. Indeed, the path space P (X) is just an exponent object X I in
g i.e. it is defined by
Man,
I
HomMan
g (N I, X) ' HomMan
g (N, X ),

g via Yoneda embedding. By definition

where N, X and I are viewed as elements of Man

I
I

I
2.0.17, (X ) = HomMan
g (X , ). Having a morphism of functors (X ) is the same
as having a compatible family of maps N : X I (N ) (N ). But this is exactly how Chen
defined differential forms on X I . So the two constructions are coherent.
Now take forms 1 , . . . , k on X of degrees s1 , . . . , sk respectively. For any smooth map
f : N P (X) these forms define pull-back forms f (1 ), . . . , f (k ) on N I. Each of
these pull-backs can be written as f (i ) = i0 (n) + i00 (n)dti . Then we can integrate the
form 100P
(n1 ) . . . k00 (nk )dt1 . . . dtk over the simplex k1 = {(t1 , . . . , tk ) I k | t1
tk and
ti = 1} to obtain a (s1 1) + + (sk 1)-form on N N = N k . Then
i

using the diagonal embedding i : N , N k , we get a (s1 1) + + (sk 1)-form i () on

N . The RR
map that associates to the set of forms 1 , . . . , k the form is called iterated path
integral
.
221

RR
One of the main results in Chens paper is that
induces a morphism of Hopf algebras
c

B ( (X)) (P (X)) which turns out to be a quasi-isomorphism. Here B c ( (X))

denotes the cyclic bar construction of the algebra (X).

Quivers and path algebras

Path algebras of quivers provide a nice class of examples of non-commutative algebras. We

will use the machinery of classical derived functors to see why quivers can be viewed as
noncommutative spaces.

3.1

Basic definitions

First we recall (see 1.1.6) basic definitions about quivers. Also, it will fix the notations.
Definition 3.1.1. A quiver is just a directed graph with finitely many vertices. More
formally, a quiver is a quadruple Q = {Q0 , Q1 , s, t}, where Q0 is the (finite) set of vertices,
Q1 is the set of arrows, and s, t : Q1 Q0 are the incidence maps assigning to an arrow
its source (resp. target).
For example, a quiver may look like

Definition 3.1.2. A (nontrivial) path in Q, = 1 2 m , is a sequence {1 , . . . , m } of

edges such that s(i ) = t(i+1 ). A trivial path is just a vertex.
Remark 3.1.3. We can naturally regard any quiver Q as a category Q, where Ob(Q) = Q0 ,
and HomQ (i, j) is the set of paths from i to j. Composition in Q is concatenation of paths.
There is another category that can be naturally associated to quiver, see example 1.1.6 in
the Chapter 3.
Remark 3.1.4. One can also think of a quiver as a finite noncommutative space. In other
words, vertices are the points, and arrows represent some kind of higher homological link
between points. If A is a finitely generated reduced commutative algebra over an algebraically
closed field k of characteristic zero, then we can consider its maximal spectrum X =
mSpec(A), whose points are maximal ideals m A, or equivalently k-algebra homomorphisms
u : A k. Given points p, q X, we can consider the corresponding skyscraper sheaves
Op , Oq . It turns out that Exti (Op , Oq ) = 0 if p 6= q. In contrast, for a quiver Q, the modules
corresponding to different vertices can have nontrivial Ext-groups, whose dimensions encode
the number of arrows between points. We will make this precise below, see 3.4.3.
Remark 3.1.5. We will usually denote vertices of a quiver simply by natural numbers:
1, 2, . . . , n, sometimes . But when we are thinking of vertices as trivial paths, or as elements
of Path(Q) then we will denote them by e1 , e2 , . . . , en , or e .
222

Let Path(Q) be the set of paths in Q. We can naturally extend s and t to maps
s, t : Path(Q) Q0 by
s() = s(m )
t() = t(1 ).
Definition 3.1.6. The length function ` : Path(Q) Z+ is a map where `() is the number
of arrows in . We put `(e) = 0 if e is a vertex.

3.2

Path algebra of a quiver

Definition 3.2.1. Let Q be a quiver. The path algebra of Q over k, denoted kQ, is the
k-vector space with basis Path(Q). The product on kQ is given by
(
if t() = s()
=
0
otherwise
Example 3.2.2. Let the vertices of Q be {1, 2, 3}, with arrows : 1 2 and : 2 3.
Then Path(Q) = {e1 , e2 , e3 , , , }, so kQ is the 5-dimensional k-algebra spanned by the
above elements. Some basic computation yields the following presentation of kQ:
kQ = khe1 , e2 , e3 , , , : eij = ij ei , e1 = e2 = , e1 = e2 = e3 = e3 = = 0i
It is easy to see that e1 + e2 + e3 = 1 in kQ.
Example 3.2.3. Let Q be the quiver with a unique edge and vertex. Then kQ ' k[x],
where the unique vertex corresponds to q, and the edge corresponds to x. Similarly, let Qn
be the unique quiver with a single vertex and n edges. One has kQn = khx1 , . . . , xn i.
Example3.2.4. In general, if A and B are k-algebras and M is an (A, B)-bimodule, then
A M
we write
for the algebra, which as a vector space is isomorphic to A M B,
0 B
and with multiplication given by

 
 

a1 m1
a2 m2
a1 a2 a1 m2 + m1 b2

=
.
0 b1
0 b2
0
b1 b2



A 0
.
N B


k 0
If now Q is the Kronecker quiver 1 2, then kQ '
. The isomorphism is
k 2 k
given by








1 0
0 0
0
0
0
0
e1 7
, e2 7
, 1 7
, 2 7
,
0 0
0 1
(1, 0) 0
(0, 1) 0
Similarly we define for (B, A)-bimodule N an algebra

223

Example 3.2.5. Let Q be a quiver with at most one path between any two vertices. Write
n = #Q0 . Then kQ ' {A Mn (k) : Aij = 0 if there is no path j i}. For example, if
Q = 1 2 n, then kQ is the algebra of n n lower-triangular matrices.
Example 3.2.6. Finally, let Q be the quiver



k[x] k[x]
Then kQ '
. This is called a framed 1-loop quiver. Path algebra Path(Q) gives
0
k
an example of the ring which is right Noetherian, but not left Noetherian.
L
Let Q be
a
quiver.
We
write
kQ
=
0
eQ0 ke kQ. This is a semisimple k-algebra.
L
Let kQ1 = :ij k be the span of all arrows in Q.
Lemma 3.2.7. Let Q be a quiver. Then kQ1 is naturally a kQ0 -bimodule, and kQ '
TkQ0 (kQ1 ).
Proof. Recall that if S is a k-algebra and M is an S-bimodule, then the tensor algebra
TS M satisfies the following universal property: given any k-algebra map f0 : S A and
an S-bimodule map f1 : M A, there is a unique f : TS M A such that f |S = f0 and
f |M = f1 . We apply this to the case S = kQ0 , M = kQ1 . These both embed into kQ, giving
a map TS M kQ. Surjectivity follows from the definition of kQ, and injectivity follows
from induction on the grading.

3.3

The structure of the path algebra

First of all, note that {ei : i Q0 } is a complete set of orthogonal idempotents in kQ, i.e.
(
X
ei if i = j
ei = 1,
ei ej =
0 otherwise
i
Moreover, if A = kQ, then Aei is the span of all paths starting at i, while ej A is the span of
all paths ending at j. L
Thus ej Aei is the span of all paths starting
L at i and ending at j. As
a left A-module, A = i Aei , and as right A-modules, A = i ei A. This implies that the
Aei are projective left ideals and the ei A are projective right ideals.
Lemma 3.3.1. Let Q be a quiver and put A = kQ.
(a) For any left A-module M and right A-module N ,
HomA (Aei , M ) ' ei M and HomA (ej A, N ) ' N ej .
(b) If 0 6= a Aei and 0 6= b ei A, then ab 6= 0.
(c) If ej Aei A, then i = j.
224

Proof. (a) Any f : Aei M is determined by f (ei ) ei M , etc.

(b) Write a = a0 (longest path in a) + and b = b0 (longest path in b) + , where
a0 b0 6= 0. Then ab = a0 b0 (longest path) + 6= 0.
(c) The module Aei A has a basis given by all paths going through i, so ej Aei A implies
i = j.
Lemma 3.3.2. Each ei kQ is a primitive idempotent, i.e. Aei is an indecomposable
A-module. Similarly, each ej A is an indecomposable right projective.
Proof. Recall, if A is a k-algebra and M is a left A-module, we say that M is decomposable
if M ' M0 M1 for nonzero M0 , M1 . Note that M is decomposable if and only if EndA (M )
has a nontrivial idempotent (projector onto one of the factors). Consider EndA (Aei ) =
HomA (Aei , Aei ) ' ei Aei . The last isomorphism is an isomorphism of k-algebras if we
consider ei as the unit in ei Aei . If Aei were decomposable, then ei Aei would have a
nontrivial idempotent f . Then f 2 = f = f ei , so f (f ei ) = 0. Part (b) of Lemma 3.3.1
tells us that this cannot be the case unless f = 0.
Lemma 3.3.3. Let A = kQ. Then Aei 6' Aej (as A-modules) unless i = j. Therefore {ei }
is a complete set of primitive idempotents in A.
Proof. Let f HomA (Aei , Aej ) = ej Aei and g HomA (Aej , Aei ) = ei Aej . Then f g
ei Ae2j Aei Aej A. If f were an isomorphism with inverse g, we would have ei Aej A, but
part (c) of Lemma 3.3.1 implies i = j.
Remark 3.3.4. In general, it is not easy to distinguish projectives up to isomorphism.
In general, a k-algebra A can have many non-equivalent projectives, even if A has no
idempotents.

3.4

Representations of quivers

Definition 3.4.1. For any quiver Q and a field k, define the category Repk (Q) to be just
the category Fun(Q, Vect) of functors from the quiver Q viewed as a category to the category
of vector spaces.
Explicitly, objects of Repk (Q) are assignments i 7 Xi Vectk , together with k-linear
maps X : Xs() Xt() for any edge Q1 . A morphism : X X 0 is given by a
collection of k-linear maps {i : Xi Xi0 }iQ0 such that for all edges Q1 , the following
diagram commutes.
Xs()


/ Xt()

s()

0
Xs()

t()

/ X0
t()

Proposition 3.4.2. There is a natural equivalence of categories kQ-Mod ' Repk (Q). Moreover, Mod(kQ) ' Repk (Q ), where Q is the opposite quiver of Q, with the same vertices
and edges as Q, but s = t, t = s.
225

Proof. There is an obvious functor F : kQ-Mod Repk (Q), which assigns to a kQ-module
X, the Q-representation F (X) with F (X)i = ei X. Given a morphism f : X X 0 of
kQ-modules, we let F (f ) : ei X ei X 0 be the restrictionL
to ei X of f : X X 0 .
Conversely, given {Xi } Ob Repk (Q), define X =
iQ0 Xi . There are canonical
projection and injection maps i : Xi , X and i : X  Xi . Give X a kQ-module structure
by
(1 m ) x = t(1 ) X1 Xm s(m ) x
The similar proof works for the category of right kQ-modules Mod(kQ).
If A is a k-algebra, write Irr(A) for the set of isomorphism classes of irreducible Amodules. Let Ind(A) be the set of isomorphism classes of indecomposable projectives over
A.
Theorem 3.4.3. Assume Q has no oriented cycles. (Equivalently, A = kQ has finite
dimension over k.)
1. The following assignments are bijections:
Q0 Irr(A)

i 7 S(i)

Q0 Ind(A)

i 7 Aei ,

where S(i)j = k ij .
2. There is a natural linear isomorphism for every i, j Q0 :
Ext1A (S(i), S(j)) ' Span Q(i, j)
where Q(i, j) is the set of all arrows from i to j.
3. ExtkA (M, N ) = 0 for all k > 2 and all A-modules M, N .
Remark 3.4.4. Part 1 of the theorem obviously fails if Q has oriented cycles. For example,
let Q consist of a single loop. Then kQ = k[x], and Irr(A) = Spec(A) k, which is much
larger than Q0 .
Before proving Theorem 3.4.3 we will need some other lemmas and facts.
Proposition 3.4.5. For any Q and any left A-module X, the following sequence is exact.
M
M
f
g
/
/
/X
/0
0
Aet() es() X
Aei ei X
()
Q1

iQ0

Here, g(a x) = a x for all a Aei , x ei X. The map f is defined by

f (aet() es() x) = a x a x
= aes() es() x aet() et() x
(Aes() es() X) (Aet() et() X),
where we used that = es() = et() .
226

Remark 3.4.6. Note that () is a projective resolution because each Aei k V is a direct
summand of the free module A k V , where V = kQ1 . Moreover, if we put X = A, then
we get a L
projective bimodule resolution of A. Indeed, Aei ej A is a direct summand of
A A = i,jQ0 Aei ej A, which is a rank one A-bimodule.
Proof. (of Prop.3.4.5) First we show that g is surjective. This can be easily seen from the
fact that any element of X can be written as

X
X
X
x=1x=
ei x =
ei x = g
ei ei x .
iQ0

iQ0

iQ0

The fact that Ker(g) Im(f ) is just a direct computation. Indeed,

g f (a et() es() x) = g(a x a x) = ax ax = 0.
To show that Ker(g) Im(f ), we first note that any
uniquely as
n
X
X
=
a xa

Ln

i=1 Aei

ei X can be written

i=1 paths a
s(a)=i

where all but finitely many of xa es(a) X are zero. Let deg() be the length of the longest
path a such that xa 6= 0. If a is a nontrivial path, we can factor it as a = a0 , with
s(a0 ) = t() and a0 a single edge. We have a0 xa = a0 es(a) es(a) x = a0 et() es(a) . By
definition,
f (a0 xa ) = a0 xa a0 xa = a xa a0 xa
We claim that for any , the set +Im(f ) contains elements of degree zero. For, if deg() = d,
then

n
X
X

f
a0 xa

i=1 s(a)=i
`(a)=d

has degree strictly less than d. The claim follows by induction on `(a) = d.
Let
now Ker(g), and take 0 + Im(f ) an element of degree zero. In other
P
P words,
n
0) =
0 = i=1 eL

x
.
If
g()
=
0,
then
because
g

f
=
0,
we
get
g()
=
g(
xei , an
i
ei
n
element of i=1 ei X, which can be zero if and only each xei = 0. But each xei = 0 implies
0 = 0, i.e. Im(f ). This finishes the proof that Ker(g) = Im(f ).
Finally, we prove that Ker(f ) = 0. Suppose f () = 0. Then we can write
=

Q1

path a
s(a)=t()

a x,a = a x,x + ,

227

where a is a path of maximal length. We get

f () =

XX

a x,a

XX

a x,a = a x,a + lower terms.

This contradicts our choice of a.

Definition 3.4.7. If X is a finite-dimensional representation of Q, define dim(X) =

(dimk X1 , . . . , dimk Xn ) Zn , the dimension vector of X.
Definition 3.4.8. The Euler form of Q is a bilinear form h, i : Zn Zn Z, defined by
h, i =

i i

iQ0

t() s() .

Q1

Associated quadratic form is called the Tits form of Q, q : Zn Z, and is defined by

q() = h, i.
Corollary 3.4.9. For any two finite-dimensional representations X and Y of Q, we have
hdimX, dimY i = dimk HomA (X, Y ) dimk Ext1A (X, Y ).
Proof. This comes from applying HomA (, Y ) to the standard resolution ():
0

/ HomA (X, Y )

HomA (Aei ei X, Y )

iQ0

HomA (Aet() es() X, Y )

/ Ext1 (X, Y )
A

Q0


M


M

Homk (ei X, ei Y )

iQ0

Homk (es() X, et() Y ).

Q1

The standard Euler characteristic yields the formula.

Remark 3.4.10. Taking X = Y , the Corollary 3.4.9 tells us that
dim EndA (X) = dim Ext1 (X, X) q(dimX).
Lemma 3.4.11. Let X, Y be (nonzero) simple A-modules. Then an extension of X by Y
0

/Y

/Z

/X

/0

is nonsplit if and only if (Y ) is the only proper submodule of Z.

228

()

/0

Proof. Suppose () splits. Then there exists : X Z such that = idX . Then (X) Z
is a submodule 6= (Y ) because (X) = X 6= (Y ) = 0.
Conversely, suppose K is a proper submodule of Z and K 6= (Y ). Define = |K : K
X. Since X is simple, is either 0 or surjective. If = 0, then K Ker = Im = (Y ),
a simple module. This implies K = (Y ), a contradiction. Thus is surjective. If k K
is killed by , then K (Y ) 6= 0, so K = (Y ), a contradiction. We have shown that
: K X is an isomorphism. We define the splitting by = 1 .
Lemma 3.4.12. Let X, Y be two non-isomorphic, simple A-modules. Then
a) If A is commutative, then every extension of X by Y splits.
b) If z Z(A) such that z AnnA (Y ) \ AnnA (X), then every extension of X by Y
splits.
Proof. a) Note that if A is commutative and 0 6= x X, then the map A X given by
a 7 ax is surjective, with kernel a maximal ideal mx . After choosing nonzero y Y , then
X ' Y iff mx = my . Let Z be an extension of X by Y . Choose a my \ mx , and define
b
a : Z Z by z 7 az. Then b
a(Y ) = 0 and X ' Z/ Ker(b
a) ' X. This lets us create a
splitting of this exact sequence.
b) Define b
a : Z Z as in part (a). Everything else works.
We return to the proof of Theorem 3.4.3.
Proof. The only non-trivial part of the Theorem 3.4.3 is about Exts.
Lemma 3.4.13. Let Q be a quiver, and let i, j be vertices in Q. Then there exists an arrow
: i j if and only if there is a nonsplit extension
0

/ S(j)

/X

/ S(i)

/ 0.

Proof. Assume i 6= j and there is : i j. We can construct an extension X by

(
k if n = i = j
Xn =
0 otherwise
and X0 = id if 0 = , with X0 = 0 otherwise. There is an obvious non-split exact sequence
0

/ S(j)

/X

/ S(i)

/ 0.

If i = j and there is : i i, let Xn = k 2 if n = i, Xn = 0 otherwise, with

!

0 1
if 0 =
X0 =
0 0

0
otherwise
The opposite direction is obvious.

229

(A.1)

Lemma 3.4.14. If i and j are two vertices in a quiver Q, then

dimk Ext1kQ (S(i), S(j)) = #{arrows i j}.
Proof. It follows from the Euler formula
hdimX, dimY i = dimk HomA (X, Y ) dimk Ext1A (X, Y ) Ext .
Apply this when X = S(i) and Y = S(j). We know that HomA (S(i), S(j)) = 0 if i 6= j, and
has dimension 1 otherwise. Moreover, dimS(i) = (ik )k and dimS(j) = (jk )k . It follows
that
( P
i,s() j,t()
if i 6= j
hdimS(i), dimS(j)i =
P
1 i,s() j,t() if i = j
which yields the desired fact.
These two lemmas prove the part (b) of the theorem. Now we only need to show that
for n 2 all functors Extn vanish.
Lemma 3.4.15. For any k-algebra A, the following three conditions are equivalent:
1. every submodule of a (left) projective A-module is projective
2. every (left) A-module has a projective resolution of length 6 1
3. ExtnA (, ) = 0 for all n > 2
Proof. Directions 1 2 3 are obvious.
To see 3 1 consider the sequence 0 X P P/X 0, where P is any
projective and X P . Apply HomA (, Y ) for any Y . As above, the long exact sequence
yields Extn (X, Y ) ' Extn+1 (P/X, Y ) = 0. Hence Ext1 (X, Y ) = 0 for all Y , and so X is
projective.
Now the statement (c) about vanishing of Extn for n 2 follows from the lemma above,
knowing that for a path algebra A = kQ every (left) A-module has a projective resolution
of length 6 1, see Proposition 3.4.5.
Definition 3.4.16. Algebras, satisfying one of the three equivalent conditions of Lemma
3.4.15 are called (left) hereditary.
Theorem 3.4.17. Let k be a perfect field. Then any finite-dimensional hereditary k-algebra
is Morita equivalent to the path algebra of a quiver.
Definition 3.4.18. A k-algebra A is called formally smooth if it satisfies any of the following
equivalent conditions:
230

1. A behaves like a free algebra with respect to nilpotent extensions: whenever I R

has I n = 0, maps lift in the following diagram

=R


/ R/I

2. A has cohomological dimension 6 1 with respect to Hochschild cohomology, i.e.

HHn (A, M ) = 0,

A-bimodule M,

n > 2

3. the kernel of the multiplication map A A A is a projective A-bimodule

Lemma 3.4.19. If A is formally smooth, then A is (left and right) hereditary.
m

Proof. Let 1 (A) = Ker(A A A). Then there is a short exact sequence of bimodules:
0

/ 1 (A)

/AA

/A

/ 0.

The fact that A is formally smooth tells us that 1 (A) is a projective bimodule. Tensor by
M for some left A-module M , to get
0

/ 1 (A) A M

/AM

/M

/ 0.

1
(we have exactness on the left because TorA
1 (A, M ) = 0 since A is projective.) Since (A)
is projective, so is 1 (A) M , whence the result.

Theorem 3.4.20 (CuntzQuillen). Every hereditary finite-dimensional k-algebra is formally

smooth.
Proof. This is Proposition 6.2 in [CQ95].
Example 3.4.21. A commutative k-algebra A is formally smooth if and only if Spec(A)
is smooth with Krull dimension 6 1. For example, A = k[x, y] is not formally smooth (in
the category of all associative algebras) though it is smooth in the category of commutative
k-algebras.
Example 3.4.22. Let A = khx, yi/([x, y] = 1). It is hereditary but not formally smooth.
Example 3.4.23. Let G be an abstract group. Then A = k[G] is formally smooth if and
only if G is virtually free, i.e. it contains a free group of a finite index.

231

232

Appendix B

Exercises
1

Standard complexes in Algebra and Geometry

Exercise 1. For a morphism of sheaves : F G show that the presheaf Im() defined
in section 3.3 is actually a sheaf.
Exercise 2. Show that if X is an irreducible topological space, then any constant sheaf on
X is flabby.
Exercise 3. Suppose G is a group and A is G-module. Check that the set Z 1 (G, A) =
{f C 1 (G, A) : d1 f = 0} of 1-cocycles is exactly the set of derivations d : G A, i.e.
Z 1 (G, A) = Der(G, A). Moreover, the set B 1 (G, A) = Im(d0 ) of 1-coboundaries is exactly
the set of inner derivations.
Exercise 4. Let A be a k-algebra for a commutative ring k, and (M , dM ), (N , dN ) be
two complexes of left A-modules. In 2.5 we defined graded Hom-space HomA (M, N ). Prove
that if M is a finitely generated (as A-module) then HomA (M, N ) = HomA (M, N ).
Exercise 5. Show that any associative star product (see 3.2) on At = A k kJtK is unital,
and that for any , there exists 0 such that 10 = 1A .

Classical homological algebra

Exercise 1. Let A be a ring, M an A-module. Prove the dual basis theorem 2.3.6. Moreover,
show that M is a generator for the category Mod(A) in the sense of definition 2.3.5 if and
only if M M = A.
Exercise 2. Show that if A and B are commutative rings, then A and B are Morita
equivalent if and only if they are isomorphic.
Exercise 3. Prove the properties 2.4 of functors (i , i , i! ) and (j! , j , j ) from the yoga
of six functors.
233

Exercise 4. Suppose F : Top is the functor defined by F ([n]) = n , where by n we

denote the standard topological n-simplex. Denote by Y : , sSet the Yoneda embedding.
Prove that the left Kan extension LanY F of F along Y exists and is isomorphic to the
geometric realization functor, i.e. LanY F = | |.
Exercise 5. Let A be an additive category. Show that finite products and finite coproducts
exist, and coincide. On the other hand, show that infinite products and coproducts need
not be the same.
Exercise 6. For an additive category A and any X, Y A, we can define the diagonal
X : X X X by X = idX idX , and the folding map Y : Y t Y Y by
Y = idY t idY . Show that the abelian group structure on HomA (X, Y ) is given by
f + g = Y (f g) X .
Exercise 7. For additive A find a categorical definition of f for any f : X Y , X, Y A.
Exercise 8. Show that if A is additive, then A is additive, and that A B is additive
whenever A and B are.
Exercise 9. If A, B, C are categories, show that there is an equivalence of categories
Fun(A B, C) ' Fun(A, Fun(B, C)).
If A, B, C are additive, show this equivalence restricts to an equivalence
Funadd (A B, C) ' Funadd (A, Funadd (B, C)).
Exercise 10. Show that if F : A B is additive, then the canonical map F (X Y )
F (X) F (Y ) is an isomorphism.
Exercise 11. Show that if A is additive, then the functors of points hX : A Set are
actually functors hX : A Ab, and similarly for hX = Hom(X, ). Show that the Yoneda
embedding h : A Fun(A , Ab) is additive.
Exercise 12. Let A be an additive category, : X Y a morphism in A. We defined
kernel of (if it exists) as the object Ker() representing the functor Ker() : A Ab
defined by
Z 7 KerAb ( : HomA (Z, X) HomA (Z, Y )).
Prove that this definition is equivalent to letting Ker() be the equalizer of the diagram

// Y .

Exercise 13. If A is an abelian category show that Com(A) is also an abelian category.
More generally, show that for any small category I the category Fun(I, A) is naturally an
abelian category.
234

Exercise 14. Prove the Mittag-Leffler condition 2.1.

Exercise 15. Prove lemma 2.2.4.
Exercise 16. If A is an abelian category satisfying AB5, then every finitely generated
object is compact. This was the lemma 2.4.6.
Exercise 17. For a Grothendieck category A show that A is also Grothendieck if and only
if A is the zero category.
Exercise 18. Show that the category Comb (A) of bounded complexes (see subsection 3.2) is
generated by A, i.e. every X Comb (A) can be obtained by taking iterated suspensions
and cones of objects in A.
Exercise 19. Take the complex I to be I = (0 Z2 Z 0). For any morphism of
complexes f : X Y there is a natural inclusion X , X I . Show that the mapping
cylinder Cyl(f ) (see 3.2) can be equivalently defined as a push-forward
X
f

i0

/X I

/ Cyl(f )

Residues and Lie cohomology

Let V be a vector space over a field k. We do not assume that V is finite-dimensional.

3.1

Commensurable subspaces

Exercise 1. Let A and B be subspaces of V . Prove that the following conditions are
equivalent:
(i) dim (A + B)/(A B) <
(ii) dim A/(A B) < and dim B/(A B) <
(iii) dim (A + B)/A < and dim (A + B)/B < .
Call the subspaces A and B commensurable (and write A B) if they satisfy the above
conditions.
Exercise 2. Prove that commensurability is an equivalence relation on the set of subspaces
of V . (Hint: prove first that if A B and B C then dim (A + B + C)/(A B C) < .)
Exercise 3. Let A, B, A0 and B 0 be subspaces in V such that A A0 and B B 0 . Prove
then that
A + B A0 + B 0 and A B A0 B 0 .
235

Exercise 4. Given two commensurable subspaces A and B in V , define the relative dimension
of A and B by
[A|B] := dim A/(A B) dim B/(A B).
Prove that if A, B and C are pairwise commensurable, then
[A|B] + [B|C] = [A|C].
Exercise 5. Let A, B, A0 and B 0 be subspaces in V such that A A0 and B B 0 . Prove
that
[A|A0 ] + [B|B 0 ] = [A B|A0 B 0 ] + [A + B|A0 + B 0 ].

3.2

Traces

Exercise 6. For a subspace A in V , define

End(V, A) := {g End V : A + gA A}.
and
Endfin (V, A) := {g End V : dim gA < and A + gV A}.
Prove that Endfin (V, A) and End(V, A) are subalgebras of End V depending only on the
commensurability class of A. Moreover,
f End(V, A) , g Endfin (V, A)

f g , gf Endfin (V, A).

Exercise 7. Let Endfin (V ) = {g End V : dim gV < }. Prove that

g1 , g2 Endfin (V, A)

g1 , g2 Endfin (V ).

Conclude that, for any g Endfin (V, A), dim g 2 V < .

Exercise 8. Using the previous exercise 7, we can define a trace map trV : Endfin (V, A) k
as follows. Given g Endfin (V, A), choose any finite-dimensional g-invariant subspace U
containing g 2 V , restrict g to U and define
trV (g) := trU (g|U ).
where trU is the usual trace on U . Prove that this definition is independent of the choice of
U.
Exercise 9. Prove that trV is a linear map: that is, for any g1 , g2 Endfin (V, A),
trV (1 g1 + 2 g2 ) = 1 trV (g1 ) + 2 trV (g2 ) ,

1 , 2 k.

(Hint: Use the (finite-dimensional) subspace U = g12 V + g22 V + g1 g2 V + g2 g1 V .)

Exercise 10. For any f, g End V , define [f, g] = f g gf . Notice that, by part (a),
if f End(V, A) and g Endfin (V, A), then [f, g] Endfin (V, A), and thus trV [f, g] is
well defined. Prove that trV [f, g] = 0. (Hint: Use the (finite-dimensional) subspaces
U1 = gf gV (gf )2 V and U2 = (f g)2 V .)
Exercise 11. For any f, g, h End(V, A), with one of them lying in Endfin (V, A), prove
trV ([h, f ]g) = trV (h[f, g]) .
236

3.3

Residues

Let A be a subspace of V . Fix a projection : V A.

Exercise 12. Prove that if f End(V, A), then [, f ] Endfin (V, A).
V : End(V, A) End(V, A) k by
Exercise 13. Using 12, define the function ()
V
()
(f, g) = trV ([, f ]g).
V is a skew-symmetric bilinear form on End(V, A).
Prove that ()

Exercise 14. Let f, g End(V, A). Let U be a subspace of V containing A and invariant
V (f, g) = U (f, g). (Because of this property we can drop
under f and g. Prove that ()
()
V .) Check that, if A is invariant under f and g, then
the superscript V in ()
() (f, g) = 0.
Exercise 15. Let A0 be another subspace of V , and let 0 : V A0 be a projection onto
A0 . Assume that A A0 . Then 0 Endfin (V, A), and


(0 ) (f, g) () (f, g) = trV ( 0 )[f, g] .
(Hint: Use property 11 of 3.2.)
Exercise 16. Using 15, prove that, if f, g End(V, A) are such that [f, g] = 0, then
() (f, g) is independent of the choice of the projection : V A and depends only on the
commensurability class of A. (We will write A (f, g) instead of () (f, g) when [f, g] = 0.)
Exercise 17. Prove that if f, g End(V, A) commute and if dim A < or dim V /A < ,
then A (f, g) = 0.
Exercise 18. Let A and B be subspaces of V . Choose projections A : V A, B : V B
and AB : V A B such that A + B AB is a projection A+B of V onto A + B
(this is clearly always possible). Prove that, for any f, g End (V, A) End (V, B), the
following formula holds
A (f, g) + B (f, g) = A+B (f, g) + AB (f, g).
Conclude that, if f, g End (V, A) End (V, B) are such that [f, g] = 0, then
A (f, g) + B (f, g) = A+B (f, g) + AB (f, g).
This last formula is called the (abstract) Tate Residue Formula.
237

3.4

Interpretation in terms of Lie algebra cohomology

We denote the Lie algebras End(V, A), Endfin (V, A) etc. by gl(V, A), glfin (V, A) etc. respectively, with the usual (commutator) bracket.
Exercise 19. Check that () is a 2-cocycle on the Lie algebra gl(V, A). Hence it defines a
canonical central extension
/k

e
/ gl(V,
A)

/ gl(V, A)

/0

Exercise 20. Show that the cohomology class cA := [() ] H2 (gl(V, A), k) is independent
of the choice of and depends only on the commensurability class of A. (Hint: Use 15 from
3.3.)
Exercise 21. Show that cA induces a cohomology class cA H2 (gl(V, A)/glfin , k).

3.5

Adeles and residues on algebraic curves

The above formalism comes from algebraic geometry. In what follows, we briefly outline
a classic construction of residues of differential forms on curves due to Tate [Tat68]. This
material requires familiarity with basic algebraic geometry.
Let X be a smooth connected algebraic curve over k, which we now assume to be
algebraically closed. Let K = k(X) be the field of rational functions on X. For a (closed)
point x X, let Ox denote its local ring. Write Ax = Ox for the completion of Ox and Kx
for the field of fractions of Ax . Note that Ax is canonically a subspace of Kx , so we can
consider End(Kx , Ax ) defined as in 3.2.
Now, choose a local parameter (coordinate) t at x and identify Ax ' kJtK and Kx ' k((t))
in the usual way, where kJtK and k((t)) are the rings of formal power and Laurent series in t,
respectively. In addition, identify the elements of Kx with the corresponding multiplication
operators on Kx : this gives us the embedding
: Kx , End(Kx )

f 7 [f : g 7 f g].

Exercise 22. Show that Im() End(Kx , Ax ).

Exercise 23. Prove that, for all f, g Kx ,
Ax (f, g) = coefficient of t1 in f (t)g 0 (t) ,
which is the usual residue of the differential form = f dg at x.
Exercise 24. For a set S of closed points of X, denote O(S) :=
X

Kx
K
A
(f, g) = O(S)
(f, g) f, g Kx .
x

xS

238

xS

Ox K. Prove that

(Hint: To prove the above formula, consider the spaces

Y
AS =
Ax
xS

(
VS =

(fx )

Kx : fx Ax for all but a finite number of x ,

xS

(The elements of VX are called adeles on the curve X.) Note that K embeds diagonally in
VS . Using the short exact sequence of sheaves
0

/ OX

/ F0

/ F1

/ 0.

where F 0 (U ) = K AU and F 1 (U ) = VU for an open U X and is the diagonal map,

show that
VX /(K + AX ) ' H1 (X, OX ) .
Conclude that dimk [VX /(K + AX )] < . Then, use the abstract Residue Formula from
exercise 18 in 3.3 for V = VS , A = AS and B = K.)
Exercise 25. Conclude from 24 that the sum of residues of = f dg over all points of X is
zero.

239

240

Bibliography
[AZ94]

M. Artin and J. J. Zhang. Noncommutative projective schemes. Adv. Math.,

109:228287, 1994. 134

[Bae40]

Reinhold Baer. Abelian groups that are direct summands of every containing
abelian group. Bull. Amer. Math. Soc., 46:800806, 1940. 137

[Bar77]

W. Barth. Moduli of vector bundles on the projective plane. Invent. Math.,

42:6391, 1977. 209

[BBD82]

A.A. Beilinson, J. Bernstein, and P. Deligne. Faisceaux pervers.

Mathematique de France, 100, 1982. 79, 161, 164, 165, 171, 212

[BC07]

Yuri Berest and Oleg Chalykh. A -modules and Calogero-Moser spaces. J.

Reine Angew. Math., 607:69112, 2007. 212

[BCE08]

Yuri Berest, Oleg Chalykh, and Farkhod Eshmatov. Recollement of deformed

preprojective algebras and the Calogero-Moser correspondence. Mosc. Math. J.,
8(1):2137, 183, 2008. 170

[Bei78]

A. A. Beilinson. Coherent sheaves on Pn and problems in linear algebra. Funktsional. Anal. i Prilozhen., 12(3):6869, 1978. 126

[Ber74]

George M. Bergman. Coproducts and some universal ring constructions. Trans.

Amer. Math. Soc., 200:3388, 1974. 65

[BGG78]

I. N. Bernstein, I. M. Gelfand, and S. I. Gelfand. Algebraic vector bundles on

Pn and problems of linear algebra. Funktsional. Anal. i Prilozhen., 12(3):6667,
1978. 147

[BKS03]

D. Benson, H. Krause, and S. Schwede. Realizability of modules over Tate

cohomology. Trans. AMS, 356(9):36213668, 2003. 41

[BL04]

Andrew Baker and Andrey Lazarev. Topological Hochschild cohomology and

generalized Morita equivalence. Algebr. Geom. Topol., 4:623645 (electronic),
2004. 86
241

Societe

[BO95]

Oleg Bondal and Dmitry Orlov. Semiorthogonal decompositions for algebraic

varieties. preprint, 1995. 166

[BRT13]

Yuri Berest, Ajay Ramadoss, and Xiang Tang. The Picard group of a noncommutative algebraic torus. J. Noncommut. Geom., 7(2):335356, 2013. 214

[Buc86]

Ragnar-Olaf Buchweitz. Maximal cohen-macaulay modules and tate-cohomology

over gorenstein rings. 1986. 177

[BW02]

Yuri Berest and George Wilson. Ideal classes of the Weyl algebra and noncommutative projective geometry. Int. Math. Res. Not., (26):13471396, 2002. With
an appendix by Michel Van den Bergh. 210

[CB92]

William Crawley-Boevey. Lectures on representations of quivers. 1992. available

at http://www1.maths.leeds.ac.uk/~pmtwc/quivlecs.pdf. 72

[CBS06]

William Crawley-Boevey and Peter Shaw. Multiplicative preprojective algebras,

middle convolution and the Deligne-Simpson problem. Adv. Math., 201(1):180
208, 2006. 214, 215

[Che77]

Kuo-Tsai Chen. Iterated path integrals. Bulletin of Am. Math. Soc., 83, Number
2:831879, 1977. 221

[CQ95]

Joachim Cuntz and Daniel Quillen. Algebra extensions and nonsingularity. J.

Amer. Math. Soc., 8(2):251289, 1995. 32, 231

[DHKS04] William G. Dwyer, Phillip S. Hirschhorn, Daniel M. Kan, and Jeffry H. Smith.
Homotopy limit functors on model categories and homotopical categories, volume
113 of Mathematical Surveys and Monographs. American Mathematical Society,
Providence, RI, 2004. 128
[DK73]

Pierre Deligne and Nicholas Katz. Groupes de monodromie et geometrie algebrique.

II, volume 340 of Lecture Notes in Mathematics. Springer-Verlag, 1973. Seminaire
de Geometrie Algebrique du Bois-Marie 19671969 (SGA 7 II). 204

[DS95]

W. G. Dwyer and J. Spali

nski. Homotopy theories and model categories. In
Handbook of algebraic topology, pages 73126. North-Holland, Amsterdam, 1995.
186

[Dyc]

T. Dyckerhoff. Compact generators in categories of matrix factorizations.

arXiv:0904.4713. 177, 183, 184, 185, 193

[EH00]

David Eisenbud and Joe Harris. The geometry of schemes, volume 197 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 2000. 17

[Eis80]

David Eisenbud. Homological algebra on a complete intersection, with an

application to group representations. Trans. Amer. Math. Soc., 260(1):3564,
1980. 177, 179
242

[ES53]

B. Eckmann and A. Schopf. Uber

injektive Moduln. Arch. Math. (Basel), 4:7578,
1953. 137

[Fal88]

Gerd Faltings. p-adic Hodge theory. J. Amer. Math. Soc., 1(1):255299, 1988.
201

[FH91]

William Fulton and Joe Harris. Representation theory, volume 129 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1991. A first course, Readings
in Mathematics. 75

[FH13]

Daniel S. Freed and Michael J. Hopkins. Chern-Weil forms and abstract homotopy
theory. Bull. Amer. Math. Soc. (N.S.), 50(3):431468, 2013. 220, 221

[FL98]

Daniel R. Farkas and Gail Letzter. Ring theory from symplectic geometry. J.
Pure Appl. Algebra, 125(1-3):155190, 1998. 43

[Gab62]

Pierre Gabriel. Des categories abeliennes. Bull. Soc. Math. France, 90:323448,
1962. 134

[Gin05]

Victor Ginzburg. Lectures on noncommutative geometry. 2005. Available at

http://arxiv.org/abs/math/0506603. 32

[GM03]

Sergei I. Gelfand and Yuri I. Manin. Methods of homological algebra. Springer

Monographs in Mathematics. Springer-Verlag, second edition, 2003. 41

[GMV96] Sergei Gelfand, Robert MacPherson, and Kari Vilonen. Perverse sheaves and
quivers. Duke Math. J., 83(3):621643, 1996. 171, 207
[GR03]

Ofer Gabber and Lorenzo Ramero. Almost ring theory, volume 1800 of Lecture
Notes in Mathematics. Springer-Verlag, 2003. 201

[Hak72]

Monique Hakim. Topos anneles et schemas relatifs. Springer-Verlag, 1972.

Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 64. 15

[Ill71]

Luc Illusie. Complexe cotangent et deformations. I, volume 239 of Lecture Notes

in Mathematics. Springer-Verlag, 1971. 203

[Kela]

Bernhard Keller. Derived categories and their uses. Available at http://

school2011.incarne.net/pdf/Literature/Keller-derived.pdf. 180

[Kelb]

Bernhard Keller. On differential graded categories. Available at http://atlas.

mat.ub.edu/grgta/articles/Keller.pdf. 186, 187

[Kel03]

Bernhard Keller. Notes for an introduction to Kontsevichs quantization theorem.

2003. availeble at http://www.math.jussieu.fr/~keller/publ/emalca.pdf.
41
243

[KKP08]

L. Katzarkov, M. Kontsevich, and T. Pantev. Hodge theoretic aspects of mirror

symmetry. In From Hodge theory to integrability and TQFT tt*-geometry, volume 78 of Proc. Sympos. Pure Math., pages 87174. Amer. Math. Soc., Providence,
RI, 2008. 177

[KL13]

Kiran Kedlaya and Ruochuan Liu. Relative p-adic hodge theory, i: Foundations.
2013. arXiv:1301.0792. 201

[Kra07]

Henning Krause. Derived categories, resolutions, and Brown representability.

In Interactions between homotopy theory and algebra, volume 436 of Contemp.
Math., pages 101139. Amer. Math. Soc., Providence, RI, 2007. 190

[LS75]

T. Y. Lam and M. K. Siu. K0 and K1 an introduction to algebraic K-theory.

Amer. Math. Monthly, 82:329364, 1975. 76

[Mat89]

Hideyuki Matsumura. Commutative ring theory, volume 8 of Cambridge Studies

in Advanced Mathematics. Cambridge University Press, Cambridge, second
edition, 1989. Translated from the Japanese by M. Reid. 31

[Maz73]

Barry Mazur. Notes on etale cohomology of number fields. Ann. Sci. Ecole
Norm.
Sup. (4), 6:521552, 1973. 79

[MLM92] Saunders Mac Lane and Ieke Moerdijk. Sheaves in geometry and logic. Universitext. Springer-Verlag, New York, 1992. A first introduction to topos theory.
15
[MV88]

R. MacPherson and K. Vilonen. Perverse sheaves with singularities along the

curve y n = xm . Comment. Math. Helv., 63(1):89102, 1988. 171, 172

[Nee]

A. Neeman. Triangulated categories. Available at http://hopf.math.purdue.

edu/Neeman/triangulatedcats.pdf. 185

[Oka50]

Kiyoshi Oka. Sur les fonctions analytiques de plusieurs variables. VII. Sur
quelques notions arithmetiques. Bull. Soc. Math. France, 78:127, 1950. 17

[Orl04]

D. O. Orlov. Triangulated categories of singularities and D-branes in LandauGinzburg models. Tr. Mat. Inst. Steklova, 246(Algebr. Geom. Metody, Svyazi i
Prilozh.):240262, 2004. 177, 178, 182

[OSS11]

Christian Okonek, Michael Schneider, and Heinz Spindler. Vector bundles on

complex projective spaces. Modern Birkhauser Classics. Birkh
auser/Springer,
2011. Corrected reprint of the 1988 edition, with an appendix by S. I. Gelfand.
147

[Qui73]

Daniel Quillen. Higher algebraic K-theory. I. In Algebraic K-theory, I: Higher

K-theories, number 541 in Lecture Notes in Mathematics, pages 85147. Springer,
Berlin, 1973. 156
244

[Rei10]

Ryan Reich. Notes on Beilinsons How to glue perverse sheaves. J. Singul.,

1:94115, 2010. 171

[Sch12]

Peter Scholze. Perfectoid spaces. Publ. Math. Inst. Hautes Etudes

Sci., 116:245
313, 2012. 199, 200, 202

[Ste75]

Bo Stenstrom. Rings of quotients, volume 217 of Die Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, New York, 1975. 134

[Tab05]

Goncalo Tabuada. Une structure de categorie de mod`eles de Quillen sur la

categorie des dg-categories. C. R. Math. Acad. Sci. Paris, 340(1):1519, 2005.
187

[Tat68]

John Tate. Residues of differentials on curves. Ann. Sci. Ecole

Norm. Sup. (4),
1:149159, 1968. 238

[Toe]

Bertrand Toen. The homotopy theory of dg-categories and derived morita theory.
arXiv:math/0408337. 177, 186

[Toe11]

Bertrand Toen. Lectures on dg-categories. In Topics in algebraic and topological

K-theory, volume 2008 of Lecture Notes in Math., pages 243302. Springer, Berlin,
2011. 129, 188, 191

[Ver96]

Jean-Louis Verdier. Des categories derivees des categories abeliennes. Asterisque,

(239):xii+253 pp. (1997), 1996. With a preface by Luc Illusie, Edited and with a
note by Georges Maltsiniotis. 143

245