Lectures on

Topics in Algebraic K-Theory

By
Hyman Bass

Tata Institute of Fundamental Research, Bombay

1967
 Lectures on
Topics in Algebraic K-Theory

By
Hyman Bass

Note by
Amit Roy

No part of this book may be reproduced in any

form by print, microfilm or any other means with-
out written permission from the Tata Institute of
Fundamental Research, Colaba, Bombay 5

Tata Institute of Fundamental Research, Bombay

1967
Preface

These notes are based upon my lectures at the Tata Institute from De-
cember, 1965 through February, 1966. The fact that the volume of ma-
terial treated was excessive for so brief a period manifests itself in the
monotone increasing neglect of technical details in the last chapters.
The notes are often a considerable improvement on my lectures, and I
express my warm thanks to Amit Roy, who is responsible for them.
Portions of the research on which these lectures are based was car-
ried out with NSF contract support at Columbia University.

Hyman Bass

iii
Contents

Introduction 1

1 The exact sequence of algebraic K-theory 7

1 Categories with product, and their functors . . . . . . . . 7
2 Directed categories of abelian groups . . . . . . . . . . . 15
3 K1 C as a direct limit . . . . . . . . . . . . . . . . . . . 18
4 The exact sequence . . . . . . . . . . . . . . . . . . . . 22
5 The category P . . . . . . . . . . . . . . . . . . . . . . 28
6 The category FP . . . . . . . . . . . . . . . . . . . . . 30
7 The category Pic . . . . . . . . . . . . . . . . . . . . . 35

2 Categories of modules and their equivalences 39

1 Categories of modules; faithfully projective modules . . 40
2 k-categories and k-functors . . . . . . . . . . . . . . . . 44
3 Right continuous functors . . . . . . . . . . . . . . . . . 46
4 Equivalences of categories of modules . . . . . . . . . . 47
5 Faithfully projective modules . . . . . . . . . . . . . . . 52
6 Wedderburn structure theory . . . . . . . . . . . . . . . 56
7 Autoequivalence classes; the Picard group . . . . . . . . 60

3 The Brauer group of a commutative ring 65

1 Separable Algebras . . . . . . . . . . . . . . . . . . . . 66
2 Assorted lemmas . . . . . . . . . . . . . . . . . . . . . 69
3 Local criteria for separability . . . . . . . . . . . . . . . 78
4 Azumaya algebras . . . . . . . . . . . . . . . . . . . . . 81

v
vi Contents

5 Splitting rings . . . . . . . . . . . . . . . . . . . . . . . 86
6 The exact sequence . . . . . . . . . . . . . . . . . . . . 87

4 The Brauer-Wall group of graded Azumaya algebras 95

1 Graded rings and modules . . . . . . . . . . . . . . . . 95
2 Separable algebras . . . . . . . . . . . . . . . . . . . . 97
3 The group of quadratic extensions . . . . . . . . . . . . 99
4 Azumaya algebras . . . . . . . . . . . . . . . . . . . . . 101
5 Automorphisms . . . . . . . . . . . . . . . . . . . . . . 102

5 The structure of the Clifford Functor 105

1 Bilinear modules . . . . . . . . . . . . . . . . . . . . . 106
2 The hyperbolic functor . . . . . . . . . . . . . . . . . . 111
3 The Clifford Functor . . . . . . . . . . . . . . . . . . . 116
4 The orthogonal group and spinor norm . . . . . . . . . . 125
Introduction

In order to construct a general theory of (non-singular) quadratic forms 1

and orthogonal groups over a commutative ring k, one should first in-
vestigate the possible generalizations of the basic classical tools (when
k is a field). These are

(I) Diagonalization (if char k , 2), and Witt’s theorem.

(II) Construction of the classical invariants: dimension, discriminant,

Hasse invariant.

This course is mostly concerned with the algebraic apparatus which

is preliminary to a generalization of II, particularly of the Hasse invari-
ant. Consequently, quadratic forms will receive rather little attention,
and then only at the end. It will be useful, therefore, to briefly outline
now the material to be covered and to indicate its ultimate relevance to
quadratic forms.
We define a quadratic module over k to be a pair (P, q) with P ∈
P, the category of finitely generated projective k-modules, and with
=
q : P → k a map satisfying q(ax) = a2 q(x) (aǫk, aǫP) and such that
(x, y) 7−→ q(x + y) − q(x) − q(y) is a bilinear form. This form then in-
duces a homomorphism P → P∗ = Homk (P, k) (by fixing a variable),
and we call (P, q) non-singular if P → P∗ is an isomorphism.
If (P1 , q1 ) and (P2 , q2 ) are quadratic modules, we have the “or-
thogonal sum” (P1 , q1 ) ⊥ (P2 , q2 ) = (P1 ⊕ P2 , q), where q(x1 , x2 ) =
q1 (x1 ) + q2 (x2 ).
Given P ∈ P, in order to find a q so that (P, q) in non-singular we 2
=

1
 2 Contents

must at least have P ≈ P∗ . Hence for arbitrary

 P, we can instead take
P ⊕ P∗, which has an obvious isomorphism, “ 10P 10P∗ ”, with its dual. In-
deed this is induced by the bilinear form associated with the hyperbolic
module
H(P) = (P ⊕ P∗ , qP ),

where q p (x, f ) = f (x)(xǫP, f ǫP∗ ). The following statement is easily

proved:
(P, q) is non-singular ⇔ (P, q) ⊥ (P, −q) ≈ H(P).
Let Q denote the category of non-singular quadratic modules and
=
their isometrics. In P we take only the isomorphisms as morphisms.
=
Then we can view H as the hyperbolic functor

H : P → Q,
= =

−1
where, for f : P → P′ , H( f ) is the isometry f ⊕ f ∗ : H(P) → H(P′ ).
Moreover, there is a natural isomorphism

H(P ⊕ P′ ) ≈ H(P) ⊥ H(P′ ).

With this material at hand I will now begin to describe the course.
In chapter 1 we establish an exact sequence of Grothendieck groups
of certain categories, in an axiomatic setting. Briefly, suppose we are
3 given a category C in which all morphisms are isomorphisms (i.e. a
groupoid) together with a product ⊥ which has the formal properties
of ⊥ and ⊕ above. We then make an abelian group out of obj C in
which ⊥ corresponds to +; it is denoted by K0 C . A related group K1 C ,
is constructed using the automorphisms of objects of C . Its axioms
resemble those for a determinant. If H : C → C ′ is a product preserving
functor (i.e. H(A ⊥ B) = HA ⊥ HB), then it induces homomorphisms
Ki H : Ki C → Ki C ′ , i = 0, 1. We introduce a relative category ΦH, and
then prove the basic theorem:
There is an exact sequence

K1 C → K1 C ′ → K0 ΦH → K0 C → K0 C ′ ,
Contents 3

provided H is “cofinal”. Cofinal means: given A′ ǫC ′ , there exists B′ →

C ′ and CǫC such that A′ ⊥ B′ ≈ HC. This theorem is a special case of
results of Heller [1].
The discussion above shows that the hyperbolic functor satisfies all
the necessary hypotheses, so we obtain an exact sequence

[H] K1 P → K1 Q → K0 ΦH → K0 P → K0 Q → Witt (k) → 0.

= = = =

Here we define Witt (k) = coker (k0 H). It corresponds exactly to the
classical “Witt ring” of quadratic forms (see Bourbaki [2]). The Ki P,
=
i = 0, 1 will be described in chapter 1. K1 Q is related to the stable
=
structure of the orthogonal groups over k.
The classical Hasse invariant attaches to a quadratic form over a 4
field k an element of the Brauer group Br (k). It was given an intrinsic
definition by Witt [1] by means of the Clifford algebra. This necessitates
a slight artifice due to the fact that the Clifford algebra of a form of odd
dimension is not central simple. Moreover, this complication renders the
definition unavailable over a commutative. ring in general. C.T.C. Wall
[1] proposed a natural and elegant alternative. Instead of modifying the
Clifford algebras he enlarged the Brauer group to accommodate them,
and he calculated this “Brauer-Wall” group BW(k) when k is a field.
Wall’s procedure generalizes naturally to any k. In order to carry this
out, we present in chapters 2, 3, and 4, an exposition of the Brauer-Wall
theory.
Chapter 2 contains a general theory of equivalences of categories of
modules, due essentially to Morita [1] (see also Bass [2]) and Gabriel
[1]. It is of general interest to algebraists, and it yields, in particular, the
Wedderburn structure theory in a precise and general form. It is also a
useful preliminary to chapter 2, where we deal with the Brauer group Br
(k) of azumaya algebras, following the work of Auslander-Goldman [1].
In chapter 4 we study the category Az2 of graded azumaya algebras, and
=
extend Wall’s calculation of BW(k), giving only statements of results,
without proofs.
Here we find a remarkable parallelism with the phenomenon wit- 5
nessed above for quadratic forms. Let FP denote the category of “faith-
= 2
 4 Contents

fully projective” k-modules P (see chapter 1 for definition), which have

a grading modulo 2 : P = P0 ⊕ P1 . Then the full endomorphism algebra
END(P) (we reserve End for morphisms of degree zero) has a natural
grading modulo 2, given by maps homogeneous of degree zero and one,
respectively.      
Matricially, ac db = a0 d0 + 0c 0b . These are the “trivial” algebras
in Az ; that is BW(k) is the group of isomorphism classes of algebras in
=2
Az , with respect to ⊗, modulo those of the form END(P). It is a group
=2
because of the isomorphism
A ⊗ A∗ ≈ END(A),
where A∗ is the (suitably defined) opposite algebra of A, for A ∈ Az .
=2
Morever, A is faithfully projective as a k-module. Finally we note that
END : FP → Az
= 2 =2

is a functor, if in both cases we take homogeneous isomorphisms as

morphisms. For, if f : P → P′ and e ∈ END(P), then END( f )(e) =
f e f −1 ∈ END(P′ ). Moreover, there is a natural isomorphism
END(P ⊗ P′ ) ≈ END(P) ⊗ END(P′ ).
6 Consequently, we again obtain an exact sequence:
[END] : K1 FP → K1 Az → K0 ΦEND → K0 FP → K0 Az → BW(k) → 0.
= 2 = 2 = 2 = 2

Chapter 5 finally introduces the category Q of quadratic forms. The

=
Clifford algebra is studied, and the basic structure theorem for the Clif-
ford algebra is proved in the following form: The diagram of (product-
preserving) functors

P H /Q
= =

∧ Clifford

 
FP / Az
= 2 END =2
Contents 5

commutes up to natural isomorphism. Here ∧ denotes the exterior

algebra, graded modulo 2 by even and odd degrees.
This result simultaneously proves that the Clifford algebras lie in
Az , and shows that there is a natural homomorphism of exact sequences
=2

[H] : K1 P / K1 Q / K0 ΦH / K0 P / K0 Q / Witt(k) /0
= = = =

Grassman Clifford Hasse−Wall

      
END : K1 FP
= 2
/ K1 Az / K0 ΦEND / K0 FP
=
/ K0 Az / BW(k) /0
= 2 2 = 2

This commutative diagram is the promised generalization of the

Hasse invariant.
Chapter 1

The exact sequence of

algebraic K-theory

The exact sequence of Grothendieck groups constructed in Bass [K, 7

Chapter 3] is obtained here in an axiomatic setting. The same is done
in a considerably, more general setting by A. Heller in Heller [1]. A
special case of the present version was first worked out by S. Chase
(unpublished).
In the last sections we shall describe the Grothendieck groups of
certain categories of projective modules.

1 Categories with product, and their functors

If C is a category, we shall denote by obj C , the class of all objects of
C , and by C (A, B), the set of all morphisms A → B, A, Bǫ obj C . We
shall assume the isomorphism classes in our categories to form sets.
A groupoid is a category in which all morphisms are isomorphisms.
Definition . A category with product is a groupoid C , together with a
“product” functor
⊥: C × C → C ,
which is assumed to be “coherently” associative and commutative in
the sense of MacLane [1].

7
 8 1. The exact sequence of algebraic K -theory

That is, we are given isomorphisms of functors

⊥ ◦(1C × ⊥) ≈⊥ ◦(⊥ ×1C ) : C × C × C → C

8 and
⊥ ◦T ≈⊥: C × C → C ,
where T is the transposition on C × C . Moreover, these isomorphisms
are compatible in the sense that isomorphisms of products of several
factors obtained from these by a succession of three-fold reassociations,
and two-fold permutations, are all the same. This permits us to write,
n
unambiguously, expressions like A1 ⊥ · · · ⊥ An = ⊥ Ai .
i=1
A functor F : (C , ⊥) → (C ′ , ⊥′ ) of categories with product is a
functor F : C → C ′ which “preserves the product”. More precisely,
there should be an isomorphism of functors F◦ ⊥≈⊥′ ◦(F × F) :
C × C → C ′ , which is compatible, in an obvious sense, with the asso-
ciativity and commutativity isomorphisms in the two categories.
Hereafter all products will be denoted by the same symbol ⊥ (except
for special cases where there is a standard notation) and we will usually
write C instead of C , ⊥).

Examples. 1) Let k be a commutative ring and let P denote the category

=
of finitely generated projective modules over k with isomorphisms as
morphisms. It is a category with product if we set ⊥= ⊕.

2) The full subcategory FP of P with finitely generated faithful projec-

= =
tive modules as objects. Here we set ⊥= ⊗k .

9 3) The full subcategory Pic of FP whose objects are finitely generated

= =
projective modules of rank 1. We set ⊥= ⊗k .

4) The category Q of quadratic modules over k with isometries as mor-

=
phisms. We take ⊥ to be the orthogonal sum of two quadratic mod-
ules.

5) The category Az of Azumaya algebras over k (see Chapter 3). Here

=
take ⊥= ⊗k .
1. Categories with product, and their functors 9

Let C (k) denote one of the categories mentioned above, and let k →
k′ be a homomorphism of rings. Then k′ ⊗k induces a functor C (k) →
(k′ ) preserving product.
If we neglect naturality conditions, then a category with product is
one whose (isomorphism classes of) objects are a commutative semi-
group. The Grothendieck group is got by formally introducing inverses
and making this semi-group into a group.

Definition. Let C be a category with product. The Grothendieck group

of C is defined to be an abelian group K0 C , together with a map

()C : obj C → K0 C ,

which is universal for maps into abelian groups satisfying

K0 . if A ≈ B, then (A)C = (B)C ,

K1 . (A ⊥ B)C = (A)C + (B)C .

In other words if G is an abelian group and ϕ : obj C → G a map sat-

isfying K0 and K1, then there exists a unique homomorphism of groups
ψ : K0 C → G such that ϕ = ψ0( )C .
Clearly K0 C is unique. We can construct K0 C by reducing the free 10
abelian group on the isomorphism classes of obj C by relations forced
by K1.
When C is clear from the context, we shall write ( ). instead of ( )C .

Proposition 1.1. (a) Every element of K0 C has the form (A) − (B) for
some A, Bǫ obj C .

(b) (A) = (B) ⇔ there exists C ∈ obj C such that A ⊥ C ≈ B ⊥ C.

(c) If F : C → C ′ is a functor of categories with product, then the map

K0 F : K0 C → K0 C ′ ,

given by (A)C 7−→ (FA)C ′ is well defined and makes K0 a functor

into abelian groups.
 10 1. The exact sequence of algebraic K -theory

We defer the proof of this proposition, since we are going to prove

it in a more general form (proposition 1.2 below).
Definition. A composition on a category C with product is a sometimes
defined composition ◦ of objects of C , which satisfies the following con-
dition: if A ◦ A′ and B ◦ B′ are defined (A, A′ , B, B′ ∈ obj C ), then so
also is (A ⊥ B) ◦ (A′ ⊥ B′ ), and
(A ⊥ B) ◦ (A′ ⊥ B′ ) = (A ◦ A′ ) ⊥ (B ◦ B′ ).
When this structure is present, we shall require the functors to pre-
serve it: F(A ◦ B) = (FA) ◦ (FB).
11 Definition. Let C be a category with product and composition.
The Grothendieck group of C is defined to be an abelian group K0 C ,
together with a map
( )C : obj C → Ko C ,
which is universal for maps into abelian groups satisfying K0, K1 and
K2. if A ◦ B is defined, then (A ◦ B)C = (A)C + (B)C .
If composition is never defined, we get back the K0 defined earlier.
As before we write ( ) instead of ( )C when C is clear from the
context.
We shall now generalize proposition 1.1.
Proposition 1.2. Let C be a category with product and composition.
(a) Every element of K0 C has the form (A) − (B) for some A, B ∈
obj C .
(b) (A) = (B) ⇔ there exist C, D0 , D1 , E0 , E1 ∈ obj C , such that
D0 ◦ D1 and E0 ◦ E1 are defined, and
A ⊥ C ⊥ (D0 ◦ D1 ) ⊥ E0 ⊥ E1 ≈ B ⊥ C ⊥ D0 ⊥ D1 ⊥ (E0 ◦ E1 ).

(c) If F : C → C ′ is a functor of categories with product and com-

position, then the map
K0 F : K0 C → K0 C ′ ,
given by (A)C 7−→ (FA)C ′ , is well defined and makes K0 a functor
into abelian groups.
1. Categories with product, and their functors 11

Proof. (a) Any element of K0 C can be written as 12

X X
(Ai ) − (B j ) = (⊥ Ai ) − (⊥ B j ).
i j
i j

(b) Let us denote by [A] the isomorphism class containing Aǫ obj C ,

and by M the free abelian group generated by these classes. A re-
P P
lation [Ai ] = [B j ] in M implies an isomorphism ⊥ Ai ≈⊥ B j in
C.
Now, if (A) = (B), then we have a relation of the following type in
M:
X
[A] − [B] = {[Ch0 ⊥ Ch1 ] − [Ch0 ] − [Chh1 ]}
X
′ ′ ′ ′
+ {[Ci0 ] + [Ci1 ] − [Ci0 ⊥ Ci1 ]}
X
+ {[D j0 ] + [D j1 ] − [D j0 ◦ D j1 ]}
X
+ {[El0 ◦ El1 ] − [El0 ] − [El1 ]},

or
X X
′ ′
[A] + {[Ch0 ] + [Ch1 ]} + [Ci0 ⊥ Ci1 ]
X X
+ [D j0 ◦ D j1 ] + {[El0 ] + [El1 ]}
X X
′
= [B] + [Ch0 ⊥ Ch1 ] + {[Ci1 ]
X X
′
+[Ci1 ]} + {[D j0 ] + [D j1 ]} + [El0 ◦ El1 ].

This implies an isomorphism

A ⊥ C ⊥ (D0 ◦ D1 ) ⊥ E0 ⊥ E1 ≈ B ⊥ C ⊥ D0 ⊥ D1 ⊥ (E0 ◦ E1 ).

where 13

′ ′
C = ( ⊥ Ch0 ) ⊥ (⊥ Ch1 ) ⊥ (⊥ Ci0 ) ⊥ (⊥ Ci1 ),
h h i i
 12 1. The exact sequence of algebraic K -theory

D0 =⊥ D j0 , E0 =⊥ E10 ,
j l

D1 =⊥ D j1 , E1 =⊥ E11 .
j l

The other implication is a direct consequence of the definition of

K0 C .

(c) The map obj C → K0 C ′ given by A 7−→ (FA)C ′ satisfies K0, K1

and K2. This gives rise to the required homomorphism K0 C →
K0 C ′ . The rest is straightforward.


Now let C be simply a category with product. For A ∈ obj C , we

write
G(A) = C (A, A),
the group of automorphisms of A. (Recall that C is a groupoid.) If
f : A → B, we have a homomorphism

G( f ) : G(A) → G(B),

given by G( f )(α) = f α f −1 .
We shall now construct, out of C , a new category ΩC . We take
obj ΩC to be the collection of all automorphisms in C . If α ∈ obj ΩC is
an automorphism of A ∈ C , we shall sometimes write (A, α) instead of
α, to make A explicit. A morphism (A, α) → (B, β) in ΩC is a morphism
14 f : A → B in C such that the diagram

f
A /B

α β
 
A /B
f

is commutative, that is, G( f )(α) = β. We define a product in ΩC by

setting (A, α) ⊥ (B, β) = (A ⊥ B, α ⊥ β). There is a natural composition
0 in ΩC : if α, βǫ obj ΩC are automorphisms of the same object in C ,
1. Categories with product, and their functors 13

then we take α ◦ β to be the usual of morphisms. The compatibility of ⊥

and 0 in ΩC is the identity

(α ⊥ β) ◦ (α′ ⊥ β′ ) = (α ◦ α′ ) ⊥ (β ◦ β′ ),

which simply expresses the fact that ⊥ is a functor (of two variables).
Definition. If C is a category with product, we define

K1 C = K0 ΩC .

Let F : C → C ′ be a functor of categories with product. Then F

induces ΩF : ΩC → ΩC ′ , preserving product and composition, so we
obtain homomorphisms

Ki F : Ki C → Ki C ′ i = 0, 1.

We propose now to introduce a relative group to connect the above

into a 5-term exact sequence.
First we construct the relative category ΦF with respect to the func- 15
tor F. Objects of ΦF are triples (A, α, B), A, B ∈ obj C and α : FA →
FB. A morphism (A, α, B) → (A′ , α′ , β′ ) in ΦF is a pair ( f, g) of mor-
phisms f : A → A′ and g: B → B′ in C such that
Ff
FA / FA′

α α′
 
FB / FB′
Fg

is a commutative diagram. We define product and composition in ΦF,

by setting

(A, α, B) ⊥ (A′ , α′ , B′ ) = (A ⊥ A′ , α ⊥ α′ , B ⊥ B′ ),
(B, β, C) ◦ (A, α, B) = (A, βα, C).

We shall see in §4 that under some restriction on F, the Grothendieck

group of this relative category ΦF fits into an exact sequence involving
the Ki′ s of C and C ′ .
We record here a few facts about K0 ΦF which we shall need later:
 14 1. The exact sequence of algebraic K -theory

Remark 1.3. (a) (A, 1FA , A)ΦF = 0 for any A ∈ obj C . This follows
from the fact (A, 1FA , A) ◦ (A, 1FA , A) = (A, 1FA , A) in ΦF.

(b) (A, α, B)ΦF = −(B, α−1 , A)ΦF for any (A, α, B) ∈ obj ΦF. This fol-
lows from (a) and the equation (B, α−1 , A)0(A, α, B) = (A, 1FA , A).

16 (c) Any element of K0 ΦF can be written as (A, α, B)ΦF . For, by propo-

sition 1.2 (a), any element of K0 ΦF can be written as (A, α, B)ΦF −
(A′ , α′ , B′ )ΦF . But this equals (A ⊥ B′ , α ⊥ α′−1 , B ⊥ A′ )ΦF , in
view of (b) above, and the axiom K1.

We close this section with a lemma about permutations that will be

needed in §4. Consider a permutation s of {1, . . . , n}. The axiom of
commutativity for ⊥ gives us, for any A1 , . . . , An , a well defined isomor-
phism
≈
A1 ⊥ · · · ⊥ An −
→ A s(1) ⊥ · · · ⊥ A s(n) ,
which we shall also denote by s. If αi : Ai → Bi , then the diagram
α1 ⊥...⊥αn
A1 ⊥ . . . ⊥ An / B1 ⊥ . . . ⊥ Bn

s s
 
A s(1) ⊥ . . . ⊥ A s(n) αs(1) ⊥...⊥αs(n)
/ B s(1) ⊥ . . . ⊥ B s(n)

is commutative, that is

s(α1 ⊥ · · · ⊥ αn ) = (α s(1) ⊥ . . . ⊥ α s(n) )s. (1.4)

Suppose now that αi : Ai → Ai+1 , 1 ≤ i ≤ n − 1, and αn : An → A1 .

Let s(i) = i − 1( mod n). and set α = α1 ⊥ · · · ⊥ αn . Then (A1 ⊥ · · · ⊥
An , sα )ǫ obj ΩC . If

β = (1A1 ⊥ α−1 −1
1 ⊥ · · · ⊥ (αn−1 · · · α1 ) ),
A1 ⊥ A2 ⊥ · · · ⊥ An → A1 ⊥ · · · ⊥ A1 .

17 then β : (A1 ⊥ · · · ⊥ An , sα) → (A1 ⊥ · · · ⊥ A1 , βsαβ−1 ) in ΩC .

Now αβ−1 = (α1 ⊥ α2 α1 ⊥ · · · ⊥ (αn · · · α1 )), and by (1.4) above,
βs = α−1
1 ⊥ (α2 α1 )
−1 ⊥ · · · ⊥ (α
n−1 · · · α1 )
−1 ⊥ 1 . Consequently:
A1
2. Directed categories of abelian groups 15

Lemma 1.5. Suppose αi : Ai → Ai+1 , 1 ≤ i ≤ n − 1 and αn : An → A1 .

Let s denote the permutation s(i) = i − 1( mod n). Then in ΩC

(A1 ⊥ · · · ⊥ An , s(α1 ⊥ · · · ⊥ αn ))
≈ (A1 ⊥ · · · ⊥ A1 , 1A1 ⊥ · · · ⊥′ A1 ⊥ (αn − α1 ))

In particular, if α : A → B and β : B → C, then

(A ⊥ B, t(α ⊥ α−1 )) ≈ (A ⊥ A, 1)

and
(A ⊥ B ⊥ C, s(α ⊥ β ⊥ (βα)−1 )) ≈ (A ⊥ A ⊥ A, 1)

in ΩC , where t and s are the transposition and the three cycle, respec-
tively.

2 Directed categories of abelian groups

In the next section we shall see that K1 C can be calculated as a kind
of generalized direct limit. We discuss in this section some necessary
technical preliminaries.
In this section G will denote a category of abelian groups. Also, we
shall assume that G is a set.

Definition . A direct limit of G is an abelian group G together with a

→
family of homomorphisms fA : A → G . A ∈ obj G , such that the dia- 18
→
gram
f
A< /B
<< 
<< 
< 
fA << 
  fB
G
→

is commutative for any morphism f : A → B and G is universal for this

→
property.
 16 1. The exact sequence of algebraic K -theory

Clearly G is unique. Also, it follows that G is the sum of its sub-

→ →
groups fA (A). We can describe G as the quotient of ⊕ A by the
→ A ∈ obj G
subgroup generated by the elements of the type f (a) − a, where f is any
morphism A → B and a ∈ A.

Lemma 2.1. Let G be such that given two objects A, B, there exists an
S
object C, with G (A, C) and G (B, C) non-empty. Then G = fA (A) .
→
P
Proof. Any element of G can be written as a finite sum fAi (ai ), ai ∈
→
Ai . To establish our assertion, it is enough to express an element of the
type fA (a) + fB (b) as fC (c) for some C and c ∈ C. We choose C such
that there are morphisms g : A → C, h : B → C. Then c = g(a) + h(b)
serves our purpose. 

It follows in particular that, if G has a “finial” object, that is, an

object C such that G (A, C) , φ for every A ∈ obj G , then fC : C → G
→
is surjective. Let N be the subgroup of C generated by all elements of
19 the type f1 (a) − f2 (a), fi ∈ G (A, C), a ∈ A. Clearly N ⊂ ker fC and this
induces a map C/N → G . On the other hand, all morphisms A → C
→
induce the same map A → C/N, and the latter are clearly compatible
with the morphisms A → B. The universal mapping property gives now
a map G → C/N which is easily checked to be the inverse of C/N → G .
→ →
Thus C/N → G is an isomorphism, that is, N = ker fC .
→

Definition. G is called directed, if

(1) given A, B ∈ obj G , there exists C ∈ obj G , such that G (A, C) and
G (B, C) are both non-empty.

(2) given fi : A → B, i = 1, 2, there exists g : B → C such that

g f1 = g f2 .

We note that lemma 2.1 is valid for directed categories.

Lemma 2.2. Let G be directed and let fA (a) = 0 for some A ∈ obj G
and a ∈ A. There exists then a morphism g : A → B such that g(a) = 0.
2. Directed categories of abelian groups 17

Proof. Since fA (a) = 0, we have, in the direct sum of the C ′ s, a =

P
±( fi (ci ) − ci ). Since only a finite number of terms appear in the re-
i
lation, we can find a C into which all the intervening groups map. In
particular, if G ′ is the full subcategory of G whose objects are those
which have a map into C, then G ′ has C as a final object and we have
fA′ (a) = 0, where fA′ : A → G ′ . Now it follows from the last paragraph
→
that if f : A → C, then there is a family of pairs f1i , f2i : Bi → C,
m
P
and bi ∈ Bi, 1 ≤ i ≤ m, such that f (a) = f1i (bi ) − f2i (bi ). Since
i=1
G is directed, it follows easily, by induction on m, that there exists an 20
h : C → B such that h f1i = h f2i , 1 ≤ i ≤ m. Then h f (a) = 0. 

Definition . A subcategory G ′ of a directed category G is called domi-

nating, if
(1) given A ∈ obj G , there exists A′ ∈ obj G ′ and a map A → A′ in G ,
(2) given fi : A′i → B in G , i = 1, 2, with A′i ∈ G ′ , there exists
g : B → C ′ with C ′ ∈ obj G ′ such that g fi ∈ G ′ , i = 1, 2.
We note first that G ′ is also directed. For given A′1 , A′2 ∈ obj G ′ ,
we can find fi : A′i → B in G . There exists then ag : B → C ′ with
C ′ ∈ obj G ′ such that g fi is a morphism in G ′ , i = 1, 2. Next suppose
f1 , f2 : A′ → B′ are maps in G ′ . There exists g : B′ → C in G such
that g f1 = g f2 . We can find h : C → D′ with hg in G ′ . Thus we have a
morphism hg in G ′ with (hg) f1 = (hg) f2 .
Proposition 2.3. If G ′ is a dominating subcategory of the directed cat-
egory G , then the natural map ϕ : G ′ → G is an isomorphism.
→ →
Proof. Write fA′ ′
: → A′ G′
and fA : A → G for the canonical maps. If
→ →
A ∈ obj G , A ∈ obj G and A′ g A is a map in either direction in G , then
′ ′

g
A′ > /A
>>
>>
′
>>
fA′ > fA
fA′ >>>
′

 >> 
G′ /G
→ ϕ →
 18 1. The exact sequence of algebraic K -theory

is commutative. 
21 Given A, we can find g : A → A′ by (1), so that im fA ⊂ Im fA′ =
im ϕ fA′ ⊂ im ϕ. It follows from lemma 2.1, that ϕ is surjective. Suppose
now x ∈ ker ϕ. Since G ′ is directed, lemma 2.1 is applicable to G ′ and
we can write x = fA′ (a′ ). By lemma 2.2 there exists g : A′ → A such
that g(a′ ) = 0. Choose h : A → B′ in G such that hg is in G ′ . Thus
hg(a′ ) = 0, so that x = fA′ ′ a′ = fB′ ′ hg(a′ ) = 0, which shows that ϕ is
injective.

3 K1C as a direct limit

Let C be a category with product. If A is an object of C , we write
G(A) for its automorphism group, [A] for the isomorphism class of A,
and G[A] for the abelianization of G(A), that is, the quotient of G(A)
by its commutator subgroup. This notation is legitimate because any
two isomorphisms A → B induce the same isomorphism G[A] → G[B],
since G(A) → G(B) is unique up to inner automorphisms.
We now propose to construct a directed category G of abelian
groups, in the sense of §2. The objects of G are the G[A], A ∈ obj C .
As for morphisms in G , we set G (G[A], G[B]) = φ if there exists no
A′ with A ⊥ A′ ≈ B. Otherwise let h : A ⊥ A′ → B be an isomor-
phism for some A′ . We have a homomorphism G(A) → G(B) given by
α 7→ G(h)(α ⊥ 1A′ ). This induces a homomorphism f : G[A] → G[B]
which is independent of the isomorphism h chosen and depends only
22 on the isomorphism class [A′ ] of A′ . The homomorphism f will be de-
noted by G[A] ⊥ [A′ ]. Now we define G (G[A], G[B]) to be the set of all
homomorphisms G[A] → G[B] which are of the form G[A] ⊥ [A′ ] for
some A′ with A ⊥ A′ ≈ B.
We define composition of morphisms in G by
(G[B] ⊥ [B′ ])(G[A] ⊥ [A′ ]) = G[A] ⊥ [A′ ⊥ B′ ],
where A ⊥ A′ ≈ B.
Since G (G[Ai ], G[A1 ⊥ A2 ]) is not empty for i = 1, 2, G satisfies
the condition (1) in the definition of a directed category. To verify (2),
suppose given f1 , f2 : G[A] → G[B],
3. K1 C as a direct limit 19

fi = G[A] ⊥ [A′i ]. Then [A ⊥ A′i ] = [B] so if we set

g = G[B] ⊥ [A] : G[B] → G[B ⊥ A], we have

g fi = (G[B] ⊥ [A])(G[A] ⊥ [A′i ]) = G[A] ⊥ [A′i ⊥ A] = G[A] ⊥ [B],

which is independent of i.

Definition . A functor F : C ′ → C of categories with product is called

cofinal if every A ∈ obj C “divides” FB’ for some B′ ∈ obj C ′ , that is,
if A ⊥ A1 ≈ FB′ for some object A1 of C . A subcategory C ′ ⊂ C is
called cofinal if the inclusion functor is.

Theorem 3.1. Let C be a category with product.

(1) Let G be the directed category of abelian groups constructed

above. There is a canonical isomorphism
≈
G −
→ K1 C .
→

(2) Let C ′ be a full cofinal subcategory of C . The inclusion of C ′ in

C induces an isomorphism
≈
K1 C ′ −
→ K1 C .

Proof. (a) If α ∈ G(A) let (α) denote its image in K1C . 23

Since (αβ) = (α) + (β), the map α 7→ (α) is a homomorphism

G(A) → K1 C which, since K1 C is abelian, induces g[A] : G[A] →
K1 C . In particular, since (1A′ ) = 0, we have (α ⊥ 1A′ ) = (α) +
(1A′ ) = (α) and this implies that the g[A] actually define a map
of the directed category G into K1 C . Hence we have a homo-
morphism G → K1 C . To construct its inverse we need only ob-
→
serve the obvious fact that the map assinginig to each α its image
(viaG(A) → G[A] → G ) in G satisfies the axioms defining K1 , so
→ →
that by universality, we get the desired homomorphism K1 C → G .
→
 20 1. The exact sequence of algebraic K -theory

(b) If A′ and B′ are two objects of C ′ ⊂ C , the symbols G(A′ ), G[A′ ]

and G[A′ ] ⊥ [B′ ] are unambiguous since C ′ is full in C . Let G ′ be
the directed category associated with C ′ . Evidently G ′ ⊂ G , and
we need only show that G ′ is a dominating subcategory of G (in the
sense of §2), for then we can invoke proposition 2.3.

Given G[A] ∈ obj G , choose A ⊥ B ≈ C ′ , B ∈ obj C , C ′ ∈ obj C ′ .
This is possible because C ′ is cofinal in C . Then G[A] ⊥ [B] : G[A] →
G[C ′ ], and G[C ′ ] ∈ obj G ′ . This verifies condition 1) for G ′ to be dom-
inating in G . Condition 2) requires that if f1 , f2 : G[A′ ] → G[B],
A′ ∈ obj C ′ , then there exists g : G[B] → G[C ′ ] such that g fi is a
morphism in G ′ , i = 1, 2. Let fi = G[A′ ] ⊥ [Ai ].
24 Choose D ∈ obj C so that B ⊥ D ≈ D′ ∈ obj C ′ . Set C ′ = A′ ⊥ D′
and let g = G[B] ⊥ [A′ ⊥ D]. Then g fi = (G[B] ⊥ [A′ ⊥ D])(G[A′ ] +
[Ai ]) = G[A′ ] ⊥ [Ai ⊥ A′ ⊥ D] = G[A′ ] ⊥ [B ⊥ D] = G[A′ ] ⊥ [D′ ]
which is a morphism in G .
Definition . An object A of C is called basic if the sequence An = A ⊥
· · · ⊥ A (n factors) is cofinal; that is, every B ∈ obj C divides An for
some n.
If A is basic the full subcategory C ′ whose objects are the An , n ≥ 1,
is a full cofinal subcategory (with product) to which we may apply the
last theorem. If we assume that An ≈ Am =⇒ n = m, then G ′ is an
ordinary direct sequence of abelian groups. The groups are G[An ], n ≥
1, and there is a unique map, G[An ] → G[An+m ], namely G[An ] ⊥ [Am ],
which is induced by α 7→ α1A m. These are only non-identity morphisms
in G ′ .
In this case we can even make a direct system from the G(An ), by
G(An ) → G(An+m ); α 7→ α ⊥ 1A m.
If we write
G(A∞ ) = lim G(An )
→
then it is clear that
lim G[An ] = G(A∞ )/[G(A∞ ), G(A∞ ].
→
3. K1 C as a direct limit 21

Theorem 3.2. Suppose that A is a basic object of C , and that An ≈ Am 25

implies n = m.
(a) K1 C is the direct limit

K1 C ≈ lim(G[An ]; G[An ] ⊥ [Am ] : G[An ] → G[An+m ])

→
≈ G(A∞ )/[G(A∞ , G(A∞ )].

(b) If α, β ∈ G(An ), then (α) = (β) in K1 C ⇔ there exist γ ∈ G(Am )

and δ1 , δ2 , ε1 , ε2 ∈ G(A p ), for some m and p, such that

α ⊥ γ(δ1 δ2 ) ⊥ 1A p ⊥ ε1 ⊥ ε2

and
β ⊥ γ ⊥ δ1 ⊥ δ2 ⊥ (ε1 ε2 ) ⊥ 1A p
are conjugate in G(An+m+4p ).
(c) (α) = 0 in K1 C ⇔ there exists γ ∈ G(Am ) for some m, such that

α ⊥ γ ⊥ γ−1

is a commutator. Moreover, α2 ⊥ 1A2m is a product of two com-

mutators.
Proof. (a) Follows directly from Theorem 3.1 and the preceding re-
marks.
(b) The implication ⇐ is clear.
For ⇒, we apply Proposition 1.2 (b) to the category C ′ consisting 26
of An , and use part (a) above to obtain γ, δ1 , δ2 , ε1 , ε2 such that

ᾱ = α ⊥ γ ⊥ (δ1 δ2 ) ⊥ ε1 ⊥ ε2

and
β̄ = β ⊥ γ ⊥ δ1 ⊥ δ2 ⊥ (ε1 ε2 )
are isomorphic Write n = n(α) if α ∈ G(An ), and similarly for
β, γ, . . .. Our hypothesis shows that n(α) is well defined and that

n(α) + n(γ)+n(δ1 δ2 ) + n(ε1 ) + n(ε2 )

 22 1. The exact sequence of algebraic K -theory

= n(β) + n(γ) + n(δ1 ) + n(δ2 ) + n(ε1 ε2 ).

Since n(α) = n(β), n(ε1 ) = n(ε2 ) = n(ε1 ε2 ) and n(δ1 ) = n(δ2 ) =

n(δ1 δ2 ), we conclude that n(δi ) = n(εi ); call this integer p, and write
m = n(γ).


Since ᾱ ≈ β̄, we have ᾱ ⊥ 1A p ≈ β̄ ⊥ 1Ap . Both of these are

in G(An+m+4p ), and we can conjugate by suitable permutations of the
factors to obtain

α′ = α ⊥ γ ⊥ (δ1 δ2 ) ⊥ 1Ap ⊥ ε1 ⊥ ε2

and
β′ = β ⊥ γ ⊥ δ1 ⊥ δ2 ⊥ (ε1 ε2 ) ⊥ 1Ap .
27 Now, two elements of G(An+m+4p ) are isomorphic if and only if they are
conjugate (recall the definition, in §1, of isomorphism in ΩC ). There-
fore α′ and β′ are conjugates. This completes the proof of (b).
Moreover, β′−1 α′ = (β−1 α) ⊥ 1Am ⊥ δ2 ⊥ δ−1 −1
2 ⊥ ε2 ⊥ ε2 is
a commutator. Conjugating by a permutation of factors, we find that
(β−1 α) ⊥ 1Am ⊥ (δ2 ⊥ ε2 ) ⊥ (δ2 ⊥ ε2 )−1 is a commutator. Since we
could have chosen m = 2m′ , we could take γ1 = 1Am ⊥ δ2 ⊥ ε2 , and a
further conjugation shows that

(β−1 α) ⊥ γ1 ⊥ γ1−1

is a commutator. Assuming β = 1An we have proved the first part of

(c) : α ⊥ γ1 ⊥ γ1−1 is a commutator. Since α ⊥ γ1 ⊥ γ1−1 is conjugate
to α ⊥ γ1−1 ⊥ γ1 , their product α2 ⊥ 1A 2m1 , m1 = n(γ1 ), is a product of
two commutators. This proves the last assertion in (c).

4 The exact sequence

Throughout this section F : C → C ′ denotes a cofinal functor of cate-
gories with product.
We define
d : K0 ΦF → K0 C
4. The exact sequence 23

to be the homomorphism induced by the map (A, α, B) 7→ (A)C − (B)C

from obj ΦF to K0 C . This is clearly additive with respect to ⊥ to 0 28
in ΦF so it does define a homomorphism d. The composite of d and
K0 F : K0 C → K0 C ′ sends (A, α, B)ΦF to (FA)C ′ − (FB)C ′ , which is
zero, since FA and FB are isomorphic.
Suppose (A)C − (B)C ∈ ker K0 F. Using Proposition 1.1, we can find
a C ′ ∈ C ′ and an α : Fa ⊥ C ′ → FB ⊥ C ′ . Cofinality of F permits us
to choose C ′ = FC for some C ∈ obj C . Then d maps (A ⊥ C, α, B ⊥ C)
into (A)C − (B)C . Thus we have proved that the sequence
d K0 F
− K0 C −−−→ K0 C ′
K0 ΦF →
is exact.
Let C1 denote the full subcategory of C ′ whose objects are all FA,
A ∈ obj C . By Theorem 3.1 (b), we have an isomorphism
θ : K1 C1 → K1 C ′ .
Let
d1 : K1 C1 → K0 ΦF
be the homomorphism induced by the map (FA, α) 7→ (A, α, A)ΦF from
obj ΩC1 to K0 ΦF. This map is additive with respect to ⊥ and 0 in ΩC1 ,
so that d1 is well defined. The composite d ◦ d1 sends (FA, α)ΩC1 to
(A)C − (A)C = 0. Thus d ◦ d1 = 0.
We define now a homomorphism 29

d′ : K1 C ′ → K0 ΦF
by setting d′ = d1 ◦ θ−1 .
Clearly d ◦ d′ = 0. Suppose (A, α, B)ΦF ∈ ker d. Then, by Propo-
sition 1.1(b), there is an isomorphism β : A ⊥ C → B ⊥ C for some
C ∈ obj C . We have then a commutative diagram
α⊥1FC
FA ⊥ FC / FB ⊥ FC

F1A⊥C Fβ−1

 
FA ⊥ FC / FA ⊥ FC
α′
 24 1. The exact sequence of algebraic K -theory

for a suitable α′ , showing that the triples (A ⊥ C, α ⊥ 1FC , B ⊥ C) and

(A ⊥ C, α′ , A ⊥ C) are isomorphic. Thus

(A, α, B)ΦF = (A ⊥ C, α ⊥ 1FC , B ⊥ C)ΦF = (A ⊥ C, α′ , A ⊥ C)ΦF .

The third member is the image of (F(A ⊥ C), α′ )ΩC1 by d1 .

Hence the sequence
d′ d
K1 C ′ −→ K0 ΦF →
− K0 C

is exact.
30 Next we note that d′ ◦ K1 F = 0. This follows from the fact that
′
d ◦ K1 F sends (A, α)ΩC to (A, F, α, A)ΦF and the triples (A, Fα, A) and
(A, 1FA , A) are isomorphic in view of the commutative diagram

Fα / FA
FA
Fα F1A
 
FA / FA
1FA

Theorem 4.6. If F : C → C ′ is a cofinal functor of categories with

product, then the sequence
K1 F d′ d K0 F
K1 C −−−→ K1 C ′ −→ K0 ΦF →
− K0 C −−−→ K0 C ′

is exact.

We have only to show that ker d′ ⊂ im K1 F. For this we need an

effective criterion for recognizing the triples (A, α, B) with the property
(A, α, B)ΦF = 0. This is given in Lemma 4.7 below, for which we now
prepare.
In ΩC ′ let E denote the smallest class of objects such that

(i) α ≈ β, α ∈ E ⇒ β ∈ E

(ii) α, β ∈ E ⇒ α ⊥∈ E

(iii) α, β ∈ E , α ◦ β defined ⇒ α ◦ β ∈ E
 4. The exact sequence 25

(iv) (FA, 1FA ), (FA ⊥ FA, t) ∈ E for all A, t being the transposition.
31 These properties imply the following:

(v) If α ∈ E , then (α)ΩC ′ ∈ im K1 F ⊂ ker d′ .

We need only note in (iv), that ′′ t = Ft′′ , with the obvious abuse
of notation.

(vi) (FA ⊥ · · · ⊥ FA, s) ∈ E for any permutation s.

Using (i), (ii), (iii) and (iv), this reduces easily to the fact that
transpositions generate the symmetric group.

(vii) If α : FA → FB and β : FB → FC, then

(FA ⊥ FB, t)(α ⊥ α−1 )) ∈ E

and
(FA ⊥ FB ⊥ FC, s(α ⊥ β(βα)−1 )) ∈ E ,
where t and S are the appropriate transposition and 3-cycle re-
spectively.

This statement follows from (i), (iv) and Lemma 1.5.

Now in ΦF we call an object of the form (A, α, A) an automorphism.
We call it elementary if (FA, α) ∈ E . For any α = (A, α, B), we write

α∼1

if α ⊥ 1FC ≈ ε for some c ∈ C and some elementary automorphism ε.

We also write
α∼β
if and only if α ⊥ β−1 ∼ 1.

Lemma 4.7. For α, β ∈ ΦF, (α)ΦF = (β)ΦF ⇔ α ∼ β. In particular, 32

(α)ΦF = 0 ⇔∝∼ 1.

Before proving this lemma, let us use it to finish the

 26 1. The exact sequence of algebraic K -theory

Proof of Theorem 4.6. Given (FA, ∝) such that (α)ΩC ′ ∈ ker d′ we have
to show that (α)ΩC ′ ∈ im K1 F. The hypothesis means that (α)ΦF = 0, so
that by Lemma 4.7, there is a c ∈ obj C , an elementary automorphism
ε = (E, ε, E), and an isomorphism ( f, g) : α ⊥ 1FC → ε. This means
that the diagram

α⊥1FC
FA ⊥ FC / FA ⊥ FC

F f¯ Fg

 
FE / FE
ε

is commutative. Hence α ⊥ 1FC = Fg−1 F f (F f )−1 εF f = F(g−1 f )ε′ ,

where ε′ = (F f )−1 εF f ≈ ε in ΩC ′ . By properties (i) and (v) above,
(ε′ )ΩC ′ ∈ im K1 F, so we have (α)ΩC ′ = (α ⊥ 1FC )ΩC ′ = (F(g−1 f ))ΩC ′ +
(ε)ΩC ′ ε im K1 F, as required.

Proof of Lemma 4.7. If α ∼ β, then (α)ΦF = (β)ΦF by virtue of (v)

above. For the converse, we will prove:

(a) ∼ is an equivalence relation

(b) ⊥ induces a structure of abelian group on M = obj ΦF/ ∼.

(c) α ◦ β ∼ α ⊥ β whenever α ◦ β is defined.

33 Once shown, these facts imply that the map obj ΦF → M satisfies
the axioms for K0 ΦF, so it induces a homomorphism K0 ΦF → M,
which is evidently surjective. Injectivity follows from the first part of
the proof above

(1) If α and β are elementary automorphisms, then so are α−1 , α ⊥ β,

and α ◦ β (if defined).
This is obvious.

(2) If β ∼ 1 and α ⊥ β ∼ 1, then α ∼ 1.

4. The exact sequence 27

For, by adding an identity to β, we can find elementary automor-

phisms ε1 = (E1 , ε1 , E1 ) and ε = (E, ε, E) and an isomorphism
( f, g) : α ⊥ ε1 → ε. Thus
α⊥ε1
FA ⊥ FE1 / FB ⊥ FE1

Ff Fg
 
FE / FE
ε

commutes. Set ε2 = (A ⊥ E1 , 1FA ⊥ ε−1 1 , A ⊥ E 1 ); ε2 = 1FA ⊥ ε1

−1

is clearly elementary. Set ε′2 = (E, F f ε2 F f −1 , E). Since F f :

(FA ⊥ FE1 , ε2 ) → (FE, ε′2 ) in ΩC ′ , ε′2 is also an elementary auto-
morphism. Moreover, we have

( f, g) : (α ⊥ ε1 ) ◦ ε2 → ε ◦ ε′2 ,

clearly, and (α ⊥ ε1 ) ◦ ε2 = α ⊥ 1FE1 . Since ε ◦ ε′2 is elementary,

we have shown α ∼ 1, as claimed.
If α = (A, α, B) and β = (B, β, C), then 34

(4) α ⊥ α−1 ∼ 1,
and

(5) α ⊥ β ⊥ (βα)−1 ∼ 1.
For,

(1A⊥B , t) : α ⊥ α−1 → (A ⊥ B, t(α ⊥ α−1 ), A ⊥ B)

and

(1A⊥B⊥C , s) : α ⊥ β ⊥ (βα)−1
→ (A ⊥ B ⊥ C, s(α ⊥ β ⊥ (βα)−1 , A ⊥ B ⊥ C),

and the latter are elementary by property (vii) of E .

Now, for the proof of (a), we note that (4) ⇒ reflexivity, (1) ⇒
symmetry, and (3) plus (4) ⇒ transitivity. The statements (b) and (c)
follow respectively from (1) and (5).
 28 1. The exact sequence of algebraic K -theory

5 The category P
Let k be a commutative ring. We define P(k) (or P) to be the category
= =
of finitely generated projective k-modules and their isomorphisms, with
product ⊕.
The groups Ki P are denoted in [K, §12] by Ki (k). Strictly speaking,
=
the definitions do not coincide since the Ki (k) are defined in terms of
exact sequences, and not just ⊕. Of course this makes no difference for
K0 since all sequences split. For K1 , however, the exact sequences of
automorphisms 0 → α′ → α → α′′ → 0 need not split. In terms of
matrices this means that α has the form
α′ β
!
α=
0 α′′
35 It is clear that α can be written in the form α = (α′ ⊕ α′′ )ε′ , where ε is of
the form id0 idγ , and the equivalence of the two definitions results from
the fact that (ε)ΩP = 0 in K1 P. The last fact is seen by adding a suitable
= =
identity automorphism to ε to put it in GL(n, k), for some n, and then
writing the result as a product of elementary matrices (see (5.3) below).
We summarize now some results from [K].
The tensor product ⊕k is additive with respect to ⊕ so that it induces
on K0 P a structure of commutative ring.
=
If Y ∈ spec(k) and P ∈ P, then PY is a free kY - module and its
=
rank is denoted by rk p (Y ). The map
rk p : spec (k) → Z,
given by Y → rk p (Y ), is continuous, and is called the rank of P. Since
rkP⊕Q = rkP + rkQ and rkP⊕Q = rkP rkQ , we have a rank homomorphism
rk : K0 P → C,
=
where C is the ring of continuous functions spec (k) → Z.
(5. 1) The rank homomorphism rk is split by a ring homomorphism
C → K0 P, so that we can write
=

K0 P ≈ ⊕K̃0 P,
= =
5. The category P 29

where K̃0 P = ker(rk). K̃0 P is a nil ideal.

= =
This result is contained in [K, Proposition 15.4]. 36

(5.2) Suppose max(k) the space of maximal ideals of k, is a noetherian

space of dimension d. Then

(a) If x ∈ K0 P and rk(x) ≥ d, then x = (P)P for some P ∈ P.

= = =
(b) If rk(P)P > d and if (P)P = (Q)P , then P ≈ Q.
= = =

(c) (K̃0 P)d+1 = 0.

=

Since k is a basic object for P in the sense of §3, we deduce imme-

=
diately from Theorem 3.2 and [K, Theorem 3.1 and Proposition
12.1], that

(5.3) There is a natural isomorphism

K1 P ≈ GL(k)/[GL(K), GL(k)],
=

where GL(k) = lim GL(n, k)(= AutK n ) with respect to the maps
! →
α 0
α 7→ from GL(n, k) to GL(n + m, k) · [GL(k), GL(k)] =
0 Im
E(k), the group generated by all elementary matrices in GL(k),
we have also E(k) = [E(k), E(k)]. The determinant map det :
GL(K) → U(k) is split by U(k) → GL(k) (defined via GL(1, K)).
Thus we have a natural decomposition

K1 P ≈ U(k) ⊕ S K1 P,
= =

where S K1 P ≈ S L(k)/E(k) = S L(k)/[S L(k), S L(k)].

=
We have also the following interesting consequence of Theorem 37
3.2.

(5.4) If α ∈ [GL(n, k),

! GL(n, k)], then for some m and some γ ∈ GL
α 0 0 2 0

(m, k), 0 γ 0−1 is a commutator in GL(n + 2m, k) and α0 I2m is
0 0γ
a product of two commutators.
 30 1. The exact sequence of algebraic K -theory

6 The category FP
Let k be a commutative ring.

Proposition 6.1. The following conditions on a k-module P are equiv-

alent:

(a) P is finitely generated, projective, and has zero annihilator.

(b) P is finitely generated, projective, and has every where positive rank
(that is PY , 0 for all Y ∈ spec(k)).

(c) There exists a module Q and an n > 0 such that P ⊗k Q ≈ kn .

Proof. The equivalence (a) ⇔ (b) is well known. 

(b) ⇒ (c). The module P is ”defined over” a finitely generated

subring k0 of k. By this we mean that there exists a finitely generated
projective k0 -module P0 such that P ≈ k⊗k0 P0 . To see this, we express P
as the cokernel of an idempotent endo-morphism of a free k-module kn .
Let α be the matrix of this endomorphism with respect to the canonical
38 basis of kn . We take for k0 , the subring of k generated by the entries
of α. It is easily seen that P0 can be takes to be the cokernel of the
endomorphisms of k0n determined by α.
So we can assume that k is noetherian with dim max(k) = d < ∞.
Let x = (P)P ∈ K0 P. Then rk(x) is a positive continuous functions spec
=
(k) → Z, and it takes only finitely many values, since spec (k) is quasi
- compact. Hence we can find y ∈ C (in the notation of (5.1)) such
that rk(x)y = m > 0 (the constant function m). Now x = rk(x) − z′
with z′ nilpotent, so that xy = m − z with z = yz′ nilpotent. It follows
that n = mh = wxy for some h > 0 and ω ∈ K0 P ; for instance we
=
can take h ≥ d + 1, in view of (5.2) (c). By enlarging h we can make
rk(wy) > d. Then we have wy = (Q)P for some Q by (5.2) (a), it follows
=
that P ⊗k Q ≈ kn .
(c) ⇒ (a). Assume P ⊗k Q ≈ kn . There is a finite set of ele-
p
P
ments x1 , . . . , x p ∈ P such tht P ⊗k Q = xi ⊗ Q. We have then a
i=1
6. The category FP 31

homomorphisms f : F → P, F a free k-module of finite rank, such that

f ⊗ 1Q : F ⊗K Q → P ⊗k Q is surjective and therefore splits. Hence
f ⊗ 1Q⊗P : F ⊗ Q ⊗ P → P ⊗ Q ⊗ P is surjective and splits. Thus
P ⊗ kn (≈ P ⊗ Q ⊗ P), being a direct summand of F ⊗ kn (≈ F ⊗ Q ⊗ P), is
finitely generated and projective. It follows that P is finitely generated
and projective.
That P has zero annihilator is clear.
Remark . The argument in (b) ⇒ (c) can be used to show, more pre- 39
cisely, that if P is a finitely generated projective k−module of constant
d+1
rank r > 0, then P ⊗k Q ≈ kr for some projective k-module Q and
some d ≥ 0. If max(k) is a noeterian space of finite dimension, then this
number can be chosen for d.
Modules satisfying (a), (b) and (c) above will be called faithfully
projective. They are stable under ⊗(= ⊗k ). The faithfully projective
modules together with their isomorphisms form a category
FP(k) (or FP)

with product ⊗, in the sense of §1. Condition (c) in the proposition

above shows that the free modules are cofinal in FP. We propose now
to calculate the groups Ki FP.
We write
Q ⊗z K0 P = (Q ⊗Z C) ⊕ (Q ⊗Z K̃0 P)

in the notation of (5.1). Thus Q ⊗Z C is the ring continuous functions

from spec(k) (discrete) Q. Let U + (Q ⊗Z K0 P) denote the unit whose
“rank” (= projection on Q ⊗ C) is a positive function.
Theorem 6.2. K0 FP ≈ U + (Q ⊗Z K0 P) 40
= =

≈ U + (Q ⊗Z C) ⊕ (Q ⊗Z K̃0 P).
=

Example. Suppose spec(k) is connected, so that C = Z. The

K0 FP ≈ (positive rationals) ⊕ (Q ⊗Z K̃0 P),
=

the direct sum of free abelian group and a vector space over Q.
 32 1. The exact sequence of algebraic K -theory

Proof. If P is faithfully projective, then P ⊗ Q ≈ kn for some n > 0, so

that (P)P (Q)P = n in K0 P. It follows that 1 ⊗ (P)P ∈ U + (Q ⊗Z K0 P),
= = = =
and this homomorphism, being multiplicative with respect to ⊗, defines
a homomorphism
K0 FP → U + (Q ⊗Z K0 P). (6.3)
=


We first show that this map is surjective. Any element of the right
1
hand side can be written as ⊗ x, x ∈ K0 P, and rk(x) is a positive
n =
function of spec (k) into Z. Since x is defined over a finitely generated
subring of k, we can assume without loss of generality, that k is finitely
generated with max(k) of dimension d, say. By increasing n by a mul-
tiple we can make rk(x) exceed d, so that x = (P)P so some P ∈ P by
= =
1
(5.2)(a). Clearly P ∈ FP. Thus ⊗ x = (1 ⊗ (kn )P )−1 (1 ⊗ (P)P ) is in the
n = =
image of (6.3).
41 Next we prove the injectivity of (6.3). Suppose 1 ⊗ (P)P = 1 ⊗ (Q)P .
= =
Then, for some integer n > 0, n((P)P − (Q)P ) = 0, so that (kn ⊗k P)P =
= = =
(kn ⊗k Q)P . By choosing n large we can make rank (kn ⊗k P) large and
=
then invoke (5.2)(b) to obtain kn ⊗ P ≈ kn ⊗ Q. Hence (P)FP = (Q)FP .
This establishes the first isomorphism in the theorem.
To prove the second isomorphism, we note that
U + (Q ⊗Z K0 P) = U + (Q ⊗Z C) × (1 + (Q ⊗Z K̃0 P)),
= =

and, since Q ⊗Z K̃0 P is a nil algebra over Q, we have an isomorphism

exp : Q ⊗Z K̃0 P → 1 + (Q ⊗Z K̃0 P).
In order to compute K1 FP, we prove a general lemma about direct lim-
its. Let
L = (Wn , fn,nm : Wn → Wnm )n,m,∈N
be a direct system of abelian groups, indexed by the positive integers,
ordered by divisibility. We introduce an associated direct system
L′ = (Wn , fn,′ nm : Wn → Wnm ),
6. The category FP 33

′
where fn,nm = m fn, nm , and a homomorphism

(n.1Wn ) : L → L′

of direct systems. For the latter we note that 42

n·1Wn
Wn / Wn

fn,nm 1
m fn,nm = fn,nm

 
Wnm / Wnm
nm1Wn,m

is commutative. L′ is a functor of L. We have an exact sequence of

direct systems
L → L′ → L′′ → 0, (6.4)
where L′′ = (Wn /Wnm , fn,nm
′′ ) is the cokernel of L → L′ .

Lemma 6.5. With the notation introduced above, the exact sequences

lim L → lim L′ → lim L′′ → 0

−−→ −−→ −−→
and
lim L ⊗ (Z → Q → Q/Z → 0)
−−→
are isomorphic. (here ⊗ = ⊗Z .).

Proof. Let E = (Zn , en,nm ) with Zn = Z and en, nm = 1Z for all n, m ∈ N.

Evidently the exact sequence of direct systems

L → L′ → L′′ → 0

and
L ⊗ (E → E ′ → E ′′ → 0)
are isomorphic. The lemma now follows from the fact that lim E = Z, 43
→
lim E ′ = Q, and standard properties of direct limit. 
→
 34 1. The exact sequence of algebraic K -theory

Theorem 3.1 allows us to compute K1 FP using only the free mod-

ules. Let
Wn = GL(n, k)/[GL(n, k), GL(n, k)]
and let fn,nm and gn,nm be the homomorphisms Wn → Wnm , induced
α 0
I
 α 0!
 n 
respectively by α 7→  . . .  and α →
  . .. from GL(n, k) to
0 In 0 α
GL(nm, k). Then it follows from theorem 3.1 that

K1 P = lim(Wn , fn,nm )
−−→
and
K1 FP = lim(Wn , gn,nm ).
−−→
Lemma 6.6. If α ∈ GL(n, k) and if nm ≥ 3, then
 m
α 0  α 0 
  
  

 In   α 
.  ≡  .  mod [GL(n, k), GL(n, k)].

 ..  
 
  .. 
   
0 In 0 α
(See [K, Lemma 1.7]).
′
It follows from lemma 6.6, that gn,nm = fn,nm = m fn,nm , and hence,
using lemma 6.5, we have the following
44 Theorem 6.7. K1 FP ≈ Q ⊗Z K1 P
=

≈ (Q ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P).
=

If we pass to the limit before abelianizing, we obtain the groups

GL⊗ (k) = lim(GL(n, k), α 7→ α ⊗ Im )n,m∈N ,

−−→
which consists of matrices of the type
 
α . 0 
 . . 

 α 

. . 
.

0
7. The category Pic 35

where α is in GL(n, k) for some n. The centre of this group consists of

scalar matrices (the case n = 1) and is isomorphic to U(k). We write

PGL(k) = GL⊗ (k)/ centre = GL⊗ (k)/U(k).

Now

K1 FP = GL⊗ (k)/[GL⊗ (k), GL⊗ (k)] = (Q ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P)

=

and we have projective on the first summand

det′ : K1 FP → Q ⊗Z U(k),

which is induced by the determinant. Explicitly, if α ∈ GL(n, k), then

 
α . 0 
 . .  1
det′  α  = ⊗ det α.
. .  n
 
0 .

This evaluates det′ , in particular, on elements of the centre (the case. 45

n = 1); so we see easily that:

coker (U(k) → K1 FP) (6.8)

= (Q/Z ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P)
=
= PGL(k)/[PGL(k), PGL(k)]
= lim PGL(n, k)/[PGL(n, k), PGL(n, k)],
−−→
where the maps are induced by the homomorphisms α 7→ α ⊗ Im from
GL(n, k) to GL(nm, k).

7 The category Pic

Pic(k) (or Pic) is the full subcategory of FP consisting of projective k-
modules of rank one, with ⊗k as product. We shall denote K0 Pic by Pic
(k).
 36 1. The exact sequence of algebraic K -theory

A module P in Pic satisfies

P ⊗k P∗ ≈ k,

where P∗ = Homk (P, k). So any object of Pic, in particular k, is cofinal.

Theorem 3.1 then shows that

K1 Pic ≈ Aut (k) ≈ U(k).

The inclusion Pic ⊂ FP induces homomorphisms

Pic(k) → K0 FP (7.1)0

46 and
U(k) → K1 FP. (7.1)1
The latter is induced by U(k) = GL(1, k) ⊂ GL⊗ (k), which identifies
U(k) with the centre of GL⊗ (k). So the co-kernel is PGL(k). Thus, we
have from (6.8),

coker (7.1)1 ≈ (Q/Z ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P) (7.2)

=
≈ PGL(k)/[PGL(k), PGL(k)],

and

ker(7.1)1 = the torsion subgroup of U(k)

(that is, the roots of unity in k).

The last assertion follows from the fact that (7.1)1 is the natural map
U(k) → Q ⊗Z U(k) followed by the inclusion of the latter into K1 FP =
(Q ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P).
(7.3) The kernel of the natural map (7.1)0 ; Pic(k) → K0 FP is the
torsion subgroup of Pic(k).

Proof. If (L)Pic ∈ ker (7.1)0 , then L ⊗k P ≈ k ⊗k P ≈ P for some

P ∈ FP. By (6.1)(c), we can choose P to be kn , in which case we have
L ⊗ · · · ⊗ L ≈ kn . Taking nth exterior powers, we get L ⊗ · · · ⊗ L ≈ k, so
that (L)Pic is a torsion element in Pic(k). 
 7. The category Pic 37

Conversely, suppose (L)Pic has order n, that is that L ⊗ · · · ⊗ L ≈ k.

47 We have to show that (L)FP = 0. This amount to showing that L ⊗ P ≈ P
for some P in FP. It is immediate that we can take for P, the module
k ⊕ L ⊕ L⊗2 ⊕ · · · ⊕ L⊗(n−1) , where L⊗i denotes the i-fold tensor product
of L with itself.
Chapter 2

Categories of modules and

their equivalences

In this chapter we first characterize (up to equivalence) categories of 48

modules as abelian categories with arbitrary direct sums and having a
faithfully projective object. Then we show that any equivalence from
the category A-mod of left modules over a ring A into the category B-
mod for another ring B, is of the form P⊗A , where P is a B-A-bimodule,
unique up to isomorphism. We deduce a number of consequence of the
existence of such an equivalence, and we characterize the modules P that
can arise in this manner. A detailed account of the Wedderburn structure
theory for semi-simple algebras is obtained in this context, Finally, for
an algebra A over a commutative ring k, the study of autoequivalences
of A-mod leads to the introduction of a group Pick (A) for a k-algebra A,
which generalizes the usual Picard group Pic(k) = Pick (k).

Most of this material is folklore. The main sources are Gabriel [1]
and Morita [1]. I have borrowed a great deal from an unpublished expo-
sition of S.Chase and S.Schanuel.

39
 40 2. Categories of modules and their equivalences

1 Categories of modules; faithfully projective mod-

ules
Let A and B be two categories. We recall that A and B are said
to be equivalent if there exist functors T : A → B and S : B → A
such that S T and T S are isomorphic to the identity functors of A and B
respectively. By abuse of language we shall say that T is an equivalence.
49 We call a functor T : A → B faithful (resp. full, fully faithful) if
the map
T : A (X, Y) → B(T X, T Y) (1.1)
is injective (resp surjective, bijective) for all X, Y ∈ obj A , where
A (X, Y) denotes the set of morphisms from X into Y. If T is an equiv-
alence, then obviously it is fully faithful; also, given Y ∈ obj B, there
exists X ∈ obj A , such that T X ≈ Y. Conversely, these two conditions
together imply that T is an equivalence. This gives us a
(1.2) Criterion for equivalence: Let T : A → B be a functor
satisfying the following conditions:

(i) T is fully faithful

(ii) Given Y ∈ obj B, there exists X ∈ ob jA with T X ≈ Y.

Then T is an equivalence.

Proof. Using (ii) we can choose, for each Y ∈ obj B, an S Y ∈ obj A

and an isomorphism
f (Y) : Y → T S Y.

These induce bijections B(Y, Y ′ ) → B(T S Y, T S Y ′ ), and by (i), we

have bijections A (S Y, S Y ′ ) → B(T S Y, T S Y ′ ). The first map, followed
by the inverse of the second, defines a bijection

S : B(Y, Y ′ ) → A (S Y, S Y ′ ).

It is easy to see that S , so defined, is a functor satisfying our require-

ments. 
 1. Categories of modules; faithfully projective modules 41

50 We shall now consider abelian categories. We shall discuss them

only provisionally, mainly for the purpose of characterizing categories
of modules. Definitions can be found in Gabriel [1], Freyd [1], and
Mitchell [1].
A functor T : A → B between abelian categories is called additive
if the maps (1.1) are homomorphisms. T is left exact if it preserves
kernels, right exact if it preserves cokernels, and exact if it does both.
We call T faithfully exact if it is faithful, exact, and preserves arbitrary
direct sums. We shall often call direct sums coproducts, and use the
`
symbol in place of the more familiar ⊕.
Let P be an object of the abelian category A . Then

h p = A (P, )

defines a functor from A to the category of abelian groups. We call P a

generator of A if hP is faithful, projective if hP is exact, and faithfully
projective if hP is faithfully exact.
Lemma 1.3. Let A be an abelian category with arbitrary direct sums.
(a) An object P of A is a generator of A ⇔ every object of A is a
quotient of a direct sum of copies of P.

(b) A class of objects of A which contains a generator is suitable un-

der arbitrary direct sums, and which contains the co-kernel of any
morphism between its members, is the whole of obj A .
`
Proof. (a). ⇒. Let X be any object of A and let S = f ∈A (P,X) P f , 51
where P f = p, with inclusion i f : P → S . There is a morphism F :
S → X such that Fi f = f for all f . Let g : X → coker F. We want
to show that g = 0, and, by hypothesis, if suffices to show that h p (g) =
A (P, g) = 0. But h p (g)( f ) = g f = gFi f = 0.
(a) ⇐ Suppose g : X → Y be a non-zero morphism. We want h p (g) , 0,
i.e. g f , 0 for some f : P → X. Choose a surjection F : S → X
`
with S = Pi , each Pi = P. The morphism F is defined by a
i
family of morphisms fi : P → X, and since gF , 0, we must have
g fi , 0 for some i.
 42 2. Categories of modules and their equivalences

(b) is a trivial consequence of (a).



The theorem below gives a characterization of categories of mod-

ules. We shall denote by
A− mod (resp. mod − A)
the category of left (resp. right) modules over a ring A.
Theorem 1.4 (See Gabriel [1] of Mitchell [1]). Let A be an abelian cat-
egory with arbitrary direct sums. Suppose A has a faithfully projective
object P. Let A = A (P, P). Then
h p = A (P. ) : A → mod − A
is an equivalence of categories, and h p (P) = A.
Proof. Clearly h p (P) = A, and since h p is faithful,
h p : A (X, Y) → HomA (h p X, h p Y) (1.5)
52 is a monomorphism. Using the criterion for equivalence (1.2), it remains
to show that
(i) h p is full (that is, that (1.5) is surjective), and
(ii) each A-module is isomorphic to some h p X.

For X = P we see easily that (1.5) is the standard isomorphism
h p (Y) → HomA (A, hP Y). As contravariant functors in X, the two side
of (1.5) are both left exact and convert direct sums into direct products.
This follows for the functor on the right, because h p is faithfully exact.
It follows from these remarks and the 5-lemma that the collection of X
for which (1.5) is an isomorphism satisfies the hypothesis of (1.3)(b),
and hence is the whole of obj A . This proves (i).
d
If M is an A-module, there is an exact sequence F1 →
− F0 → M → 0
with Fi free. Up to isomorphism we can write Fi = hPGi , with Gi
a direct sum of copies of P. By (i), we can write d = hP g for some
s : G1 → G0 . Then, from exactness, M ≈ coker hP g ≈ hP coker g.
1. Categories of modules; faithfully projective modules 43

Proposition 1.5. Let P be a right module over a ring A. The following

statements are equivalent:

(i) P is faithfully projective.

(ii) P is finitely generated, projective, and is a generator of mod A.

generator of mod − A.

Proof. In view of the definition of faithful projectivity, we have only 53

to show if P is projective, then P is finitely generated if and only the
functor HomA (P, ) preserves coproducts. 

Suppose P is finitely generated. Any homomorphism of P into a

coproduct has its image in a finite coproduct (a finite number of factors
is enough for catching the non-zero coordinates of the images of a finite
system of generators of P). Thus such a homomorphism is a (finite) sum
of a homomorphisms of P into the factors.
Conversely, suppose HomA (P, ) preserves coproducts. Consider a
`
homomorphism e : P → Ai (each Ai = A) with a left inverse (such
i
a map exists since P is projective). By hypothesis, e is a finite sum of
homomorphisms ei : P → Ai , i ∈ S , S a finite set. Thus P is a direct
`
summand of Ai and hence finitely generated.
i∈S

Remark. If P is not projective, then finite generation is no longer equiv-

alent with HomA (P, ) preserving coproducts. For, we have obviously,

(1.6) P is finitely generated ⇔ the proper submodules of P are induc-

tively ordered by inclusion.
On the other hand

(1.7) HomA (P, ) preserves coproducts ⇔ the union of any ascending

sequence of proper submodules of P is a proper submodule.

If P is the maximal ideal of a valuation ring, where the value group 54

has a suitably pathological order type, then P will satisfy (1.7) but not
(1.6).
 44 2. Categories of modules and their equivalences

`
Proof of (1.7) ⇐. If f : P → Mi is a homomorphism such that f (P)
i∈I
is not contained in a finite direct sum of the Mi′ s, then we can choose a
` `
countable subset J of I such that if g : Mi → M j is the projection,
i∈I j∈J
then g f (P) is like wise not in a finite sum. Letting S expand through a
sequence of finite subsets of J, with J as their union, we find that the
submodules (g f )−1 ( M j ) violate the assumed chain condition on P.
P
j∈S
⇒. Suppose P1 ⊂ P2 ⊂ · · · ⊂ Pn ⊂ · · · are proper sub - modules
S
of P with Pn = P. The projections fn : P → P/Pn define a map of
Qn≥1 `
f : P → P/Pn , whose image is clearly in P/Pn , but not in a finite
n
sum of the P/Pn .

2 k-categories and k-functors

Let A be a ring and let M be a right A-module. For an element a ∈
centre A, the homothetie h(a) M : M → M (defined by h(a)M (x) = xa) is
A-liner. These homomorphisms define an endomorphism of the identity
functor Idmod−A .
Proposition 2.1. The homothetie map
h : centre A → End (Idmod−A )
is an isomorphism of rings.
55 Proof. If h(c) = 0, then c = h(c)A (1) = 0. Let f be an endomorphism
of the functor Idmod−A . fA is the left multiplication in A by c = fA (1).
The element c belongs to the centre of A. This follows from the fact
that fA commutes with all left multiplications in A, since f is a natural
transformation. Set f ′ = f − h(c). We shall show that f ′ = 0. Let M be
a right A-module. For an x ∈ M, consider the A-linear map t : A → M
given by t(a) = xa. We have fM′ ◦ t = t ◦ fA′ . It follows that f M′ (x) = 0.
Thus f ′ = 0. 

The proposition suggests the definition

centre A = End (IdA )
2. k-categories and k-functors 45

for any abelian category A . Let k be a commutative ring and k → cen-

tre A a homomorphism. This converts the A (X, Y) into k-modules so
that the composition is k-bilinear. Conversely, given the latter structure,
we can clearly reconstruct the unique homomorphism k → centre A
which induces it. An abelian category A with a homomorphism k →
centre A will be called a k- category. A functor T : A → B be-
tween two such categories will be called a k-functor if the maps (1.1)
are k-linear. The k-functors forms a category, which we shall denote by
k-Funct (A , B).
If A is a k-algebra, then by virtue of (2.1), mod-A is a k-category.
Let A and B be k-algebras and suppose M is a left A−, right B-module.
If B-module. If t ∈ k and x ∈ M, then tx and xt are both defined. The
following statement is easily checked:
(2.2) tx = xt for all t ∈ k, x ∈<⇔ ⊗A M : mod − A → mod − B 56
is a k-functor.
This condition simply means that M can be viewed as left module
over A ⊗ B0 . We will often follow the Cartan-Eilenberg convention of
k
writing A MB to denote the fact that M is left A−, right B-bimodule, and
when a ground ring k is fixed by the context, it will be understood that
M satisfies (2.2).

Proposition 2.3. h( A MB ) ⊗A M : mod − A → mod − B defines a

fully faithful functor

h : (A ⊗k B0 ) − mod → k − Funct( mod − A, mod − B).

In particular, A MB ≈A NB as bimodules ⇔ ⊗A M ≈ ⊗A N as functors

from mod − A to mod − B.

Proof. If f : A MB → A NB is a bimodule homomorphism, then h( f ) =

⊗A f is a morphism of functors. Thus h is a functor. If h( f ) = 0, then
1A ⊗A f = h( f )(1A ) = 0, i.e, f = 0. So h is faithful. 

Suppose t : hM → hN is a natural transformation. We will conclude

 46 2. Categories of modules and their equivalences

by showing that t = h( f ), where f is the unique B-morphisms rendering

f
M /N

≈ ≈
 
A ⊗A M / A ⊗A N
tA

57 commutative. The vertical maps are bimodule isomorphisms. Since left

multiplications in A are right A-linear, tA must respect it, by naturality.
Thus tA , and hence also f , is a bimodule homomorphism, so h( f ) is
defined. Let s = t−h( f ) : hM → hN. The class C of X in obj mod − A
for which sX = 0 contains A. Since hM and hN are right exact and
preserve coproducts, it follows from (1.3) (b), that C = obj mod − A.

3 Right continuous functors

We will here describe the image of the functor of proposition 2.3. Func-
tors of the type ⊗A M : mod − A → mod − B are (i) right exact,
and (ii) preserve arbitrary coproducts. It follows that they also preserve
direct limits. A functor satisfying (i) and (ii) will be called right contin-
uous. The next theorem says that they are all tensor products.
Theorem 3.1 (Eilenberge-Watts). The correspondence A MB 7→ ⊗A M
induces a bijection from the isomorphism classes of left A⊗k B0 - modules
to the isomorphism classes of right continuous k-functors from mod −
A to mod − B. In the situation ( A MB , B NC ), A (M ⊗B N)C corresponds
to the composite of the respective functors.

Proof. The last statement follows from

((⊗B N) ◦ (⊗A M))(X) = (⊗B N)(X ⊗A M)

= (X ⊗A M) ⊗B N
= X ⊗A (M ⊗B N)
= ⊗A (M ⊗B N)(X).

58 Injectivity is just the last part of proposition 2.3. 

4. Equivalences of categories of modules 47

Let T : mod − A → mod − B be a right continuous k-functor.

The composite
A → HomA (A, A) → HomB (T A, T A),
where the first map is given by left multiplications, is a homomorphism
of k-algebras. This makes M = T A into a left A ⊗k B0 - module. We will
conclude by showing that the functors T and ⊗A M are isomorphic. If X
is a right A-module, we have maps
≈ T
X−
→ HomA (A, X) −
→ HomB (T A, T X) = HomB (M, T X),
and the composite fX is A-linear (for the action of A on M just con-
structed). Now, there is a canonical isomorphism
HomA (X, HomB (M, T X)) ≈ HomB (X ⊗A M, T X),
and fX is an element of the first member. Let gX be the corresponding
element in the second member. The homomorphisms gX define a natural
transformations of functors g : ⊗A M → T . For X = A, we have gA as the
obvious isomorphism A ⊗A M → T A = M. Using the right continuity
of T and ⊗A M, we now see that the class of objects x for which gx
is an isomorphism, satisfies the conditions of (1.3) (b). Thus g is an
isomorphism of functors.
Definition 3.2. We shall call a bimodule A MB invertible, if the functor
⊗A M : mod − A → mod − B is an equivalence.
This equivalence is evidently right continuous (indeed, any equiv- 59
alence is). It therefore follows from theorem 3.1. that the invertibil-
ity of M is equivalent to the existence of a bimodule B NA such that
M ⊗B N ≈ A and N ⊗A M ≈ B as bimodules (over appropriate rings).
This shows that the definition of in vertibility is left-right symmetric. In
particular, M⊗B : B − mod → A − mod is also equivalence.

4 Equivalences of categories of modules

We have just seen that an equivalence is, up to isomorphism, tensoring
with an invertible bimodule. We now summarize.
 48 2. Categories of modules and their equivalences

Proposition 4.1. Let A and B be a k-algebras and suppose

T
mod − A−
→ mod − B
←−
S

are k-functors such that S T and T S are isomorphic to the identity func-
tors of mod − A and mod − B respectively. Set P = T A and Q = S B.
Then we are in the situation ( A PB , B QA ), and :

(1) T ≈ ⊗A P, and S ≈ ⊗B Q.

(2) There are bimodule isomorphisms

f : P ⊗B Q → A and g : Q ⊗A P → B.

(3) f and g may be chosen to render the diagrams

f ⊗1P g⊗1Q
P ⊗B Q ⊗A P / A ⊗A P Q ⊗A P ⊗B Q / B ⊗B Q

1P ⊗g and 1Q ⊗ f
   
P ⊗B B /P Q ⊗A A /Q

commutative.

60 Proof. Statements (1) and (2) follow immediately from theorem 3.1,
since an equivalence is automatically right continuous. To prove the
statement (3), we first note that all the intervening maps are isomor-
phisms of bimodules. If a : A ⊗A P → P and b : P ⊗B B → P are the
natural maps, then we have b(1 ⊗ g) = ua( f ⊗ 1) for some A − B - au-
tomorphism u of P. In particular, u ∈ HomB (P, P) = HomB (T A, T A) ≈
HomA (A, A) = A. So u is a left multiplication by a unit in A, which we
shall denote by the same letter u. Since u is also an A-homomorphism,
we must have u ∈ centre A. Now, evidently ua = a(u ⊗ 1P ). So if we
replace f by u f we have made the first square commutative. Assume
that this has been done. 

Write f (p ⊗ q) = pq and g(q ⊗ p) = qp for p ∈ P, q ∈ Q. We

have arranged that (pq)p′ = p(qp′ ), and we will prove that the desired
4. Equivalences of categories of modules 49

equality (qp)q′ = q(pq′ ) follows automatically. For, if p, p′ ∈ P, q,

q′ ∈ Q we have

((qp)q′ )p′ = (qp)(q′ p′ ) (g is left B-linear)

′ ′
= q(p(q p )) (g is right B-linear)
′ ′
= q((pq )p ) ( by assumption)
′ ′
= (q(pq )p ) (q ⊗ ap = qa ⊗ p, a ∈ A).

Hence, if d = (qp)q′ −q(pq′ ), then dp′ = 0 for all p′ ∈ P. Let h : A → Q

be defined by h(a) = da. Then h ⊗ 1 p : A ⊗A P → Q ⊗A P followed by
the isomorphism g is zero. So h ⊗ 1P = 0. But ⊗A P is a fully faithful
functor. Therefore h = 0, that is d = 0.

Definition 4.2. A set of pre-equivalence data (A, B, C, P, f, g) consists 61

of k-algebras A and B, bimodules A PB and B QA , bimodule homomor-
phisms
f : P ⊗B Q → A and g : Q ⊗A P → B,
which are “associative” in the following sense: Writing f (p ⊗ q) = pq
and g(q ⊗ p) = qp, we require that

(pq)p′ = p(qp′ ) and (qp)q′ = q(pq′ ) p, p′ ∈ P, q, q′ ∈ Q.

We call it a set of equivalence data if f and g are isomorphisms.

Theorem 4.3. Let (A, B, P, Q, f, g) be a set of pre-equivalence data. If

f is surjective, then

(1) f is an isomorphism

(2) P and Q are generators as A-modules

(3) P and Q are finitely generated and projective as B-modules.

(4) g induces bimodule isomorphisms

P ≈ HomB (Q, B) and Q ≈ HomB (P, B)

 50 2. Categories of modules and their equivalences

(5) The k-algebra homomorphisms

HomB (P, P) ← A → HomB (Q, Q)0

induced by the bimodule structures, are isomorphisms.

Proof. The hypothesis on f means that we can write

X
1= pi qi in A.

p′j ⊗ q′j ∈ ker f . Then

P
62 (1) Suppose
X X X
p′j ⊗ q′j = (p′j ⊗ q′j )pi qi = p′j ⊗ ((q′j pi )qi ) =
j,i j,i
X X
= (p′j (q′j pi )) ⊗ qi = ( (p′j q′j )(pi ⊗ qi )
j,i j,i
X X X
=( p′j q′j )( pi qi ) = 0, since p′j q′j = 0.
j i

(2) We have A-linear maps hi : P → Ai given by hi (p) = pqi . These

`
define an A-linear map h : Pi → A (each Pi = P), which is
i
surjective. It follows by (1.3) (a), that P is a generator of A− mod ,
since A is so. The argument for Q is similar.
e
(3) Define P→ − ` B (each B = B), by e(p) = (q p) and h((b )) =
h i i i i
P ←− i P P
pi bi . Then he(p) = i pi (qi p) = ( pi qi )p = p. Thus P is finitely
i
generated and projective. Similarly Q also is finitely generated and
projective.

(4) g induces an A-B-bimodule homomorphism h : P → HomB (Q, B),

P P
given by h(p)(q) = qp. If h(p) = 0, then p = (pi qi )p = pi
i P i
(qi p) = 0. If f : Q → B is B-linear, then f (q) = f ( q(pi qi )) =
P P P iP
f ( (qpi )qi ) = i (qpi ) f (qi ) = q(pi f (qi )), so f = h( pi f (qi )).
i i i
Similarly Q ≈ HomB (P, B).
 4. Equivalences of categories of modules 51

63 (5) Define h : A → HomB (P, P) by h(a)p = ap. If h(a) = 0, then

P P
a = a(pi qi ) = (api )qi = 0. If f : P → P is a B-linear, then
i P i P P
f (p) = f ( (pi qi )p) = f ( pi (qi p)) = ( f (pi )qi )p, so that f =
i i i
h( f (pi )qi ). Similarly A ≈ HomB (Q, Q)0 via right multiplication.
P
i

Theorem 4.4. Let (A, B, P, Q, f, g) be a set of equivalence data (see

definition 4.2). Then

(1) The functors P×B , ⊗A P, Q⊗A , and ⊗B Q are equivalences between

the appropriate categories of A-modules and B-modules.

(2) P and Q are faithfully projective both as A-modules and B-modules.

(3) f and g induce bimodule isomorphisms of P and Q with each others

duals with respect to A and to B.

(4) The k-algebra homomorphisms

HomB (P, P) ← A → HomB (Q, Q)0

and
HomA (P, P)0 ← B → HomA (Q, Q),

induced by the bimodule structures on P and Q, are isomorphisms.

(5) The bimodule endomorphism rings of A, B, P and Q are all isomor-

phic to the centres of A, B mod − A and mod − B.

(6) The lattice of right A-ideals is isomorphic, via U 7→ U P, with the 64

lattice of B-submodules of P, the two sided ideals corresponding to
A − B-submodels, or equivalently, to fully invariant B-submodules.
Similar conclusions apply with appropriate permutations of (A, B),
(P, Q), (left, right). In particular, by symmetry, A and B have iso-
morphic lattices of two-sided ideals.
 52 2. Categories of modules and their equivalences

Proof. (1) is immediate.

(2), (3) and (4) follow immediately from (2), (3), (4) and (5) of
theorem 4.3.
We have isomorphisms

⊗A P
centre A ≈ HomA−A (A, A) −−−→ HomA−B (P, P),
≈
P⊗B
centre B ≈ HomB−B (B, B) −−−→ HomA−B (P, P),
≈

and similarly for Q also. The statement (5) follows from these isomor-
phisms plus proposition 2.1. 

We now prove (6). Since P is A-projective, the canonical map U ⊗A

P → U P is an isomorphism. That U 7→ U P is an isomorphism of
the lattice of right ideals of A onto the lattice of B-submodules of P,
now follows from the fact that ⊗A P : mod − A → mod − B is an
equivalence. The fully invariant right A-submodules of A, i.e., the two-
sided ideals of A, correspond to the fully invariant B−submodules of P,
which, by virtue of (4), are just the A − B- submodules of P.
65 The remaining assertions in (6) are clear. The isomorphism between
the lattices of two-sided ideals of A and B can be made explicit: U ↔ b
if U P = Pb, where U and b are two-sided ideals in A and B respec-
tively. The conclusion above show that given U , the ideal b exists and
is unique.

5 Faithfully projective modules

Let B be a k−alegbra and let P be right B-module. From B and P we
will construct a set of pre-equivalence data and then determine in terms
of B and P alone, what it means for them to be equivalence data.
We set
A = HomB (P, P),
and
Q = HomB (P, B).
5. Faithfully projective modules 53

Then A is a k−algebra and P is an A − B-bimodule, that is, a left A ⊗k B0 -

module. Moreover, Q is a B − A - bimodule with the following prescrip-
tion:
(bq)p = b(qp) (5.1)
and
(qa)p = q(ap), (5.2)
a ∈ A, b ∈ B, p ∈ P, q ∈ Q. Next we define pq ∈ A for p ∈ P and q ∈ Q,
by requiring that

(pq)p′ = p(qp′ ), p′ ∈ P (5.3)

This permits us to define a homomorphism of A − A-bimodules

f p : P ⊗B Q → A, by fP (p ⊗ q) = pq,

and a homomorphism of B − B-bimodules 66

g p : Q ⊗A P → B, by g p (q ⊗ p) = qp.

Finally, we claim that

(qp)q′ = q(pq′ ), (5.4)
for p ∈ P, q, q′ ∈ Q. Since these are linear maps P → B, we need only
show that they have the same value at any p′ ∈ P. But

((qp)q′ )p′ = (qp)(q′ p′ ) by (5.1)

′ ′
= q(p(q p )) by B - linearity of q
= q((pq′ )p′ ) by (5.3)
′ ′
= (q(pq )p ) by (5.2).

We have now proved

Proposition 5.5. Let B be a k-algebra, P a right B-module, and f p and
g p be as constructed above. Then

(HomB (P, P), B, P, HomB (P, B), f p .g p )

is a set of pre-equivalence data.

 54 2. Categories of modules and their equivalences

Example. Let P = eB, where e is an idempotent. Then B = P ⊕ (1− e)B.

Any B-linear map f : P → B can be extended to a B−linear map f¯ :
B → B by setting f¯(1 − e) = 0. Thus we have inclusions HomB (P, P) ⊂
HomB (P, B) ⊂ HomB (B, B). With this identification, HomB (P, P) = eBe
and HomB (P, B) = Be.

67 Proposition 5.6. In the notation of proposition 5.5:

(a) im fP = HomB (P, P) ⇔ P is a finitely generated projective B-modu-

le, in which case f p is an isomorphism.

(b) im g p = B ⇔ P is a generator of mod − B, in which case g p is an

isomorphism.

(c) (HomB (P, P), B, P, HomB (P, B), f p , g p ) is a set of equivalence data
⇔ P is faithfully projective.

Proof. (c) follows from (a), (b) and proposition 1.5.

In view of theorem 4.3, it remains only to show the implications ⇐
in (a) and (b).
Suppose P is a finitely generated projective B−module. We can
`
find a free B-module ei B with a basis e1 , . . . , en , and B−linear maps
h1 ` h2
P −−→ ei B −−→ P such that h2 h1 = 1P . If qi : P → B denotes the com-
posite of h1 and the ith coordinate linear form on ei (B), we can write
`
P P
h (p) = ei (qi p). Let pi = h2 (ei ). Then p = h2 h1 p = h2 ( ei (qi p)) =
P1 P P
pi (qi p) = ( pi qi )p. So 1 p = pi qi ∈ im f p , and the latter is a
two-sided ideal in HomB (P, P). Hence im f p = hom B (P.P).
Next, suppose P is a generator of mod − B. Then B is a quotient of
a sum (which we mau take finite) of copies of P. This means that we can
P
find qi ∈ HomB (P, B) such that qi P = B. Hence g p is surjective. 

68 Lemma 5.7. A right B−module P is projective ⇔ there exist pi ∈ P,

qi ∈ HomB (P, B), i ∈ I, such that

(i) given p ∈ P, qi p = 0 for almost all i, and

P
(ii) i pi (qi p) = p, p ∈ P.
5. Faithfully projective modules 55

The family (pi ) which arise in this manner are precisely the gener-
ating systems of P. If U = im g p , then U is generated, as a two-sided
ideal, by the qi p j . Moreover, PU = P and U 2 = U .

Proof. Projectivity of P is equivalent to the existence of a free B-module

` h1 ` h2
i∈I ei B and B-linear maps P −−→ ei B −−→ P such that h2 h1 = 1P . The
i∈I
latter condition, in turn, is equivalent to the existence of the pi and qi .
For, given pi and qi , one can construct h1 and h2 in an obvious fashion.
On the other hand, given h1 and h2 , we can take pi to be h2 (ei ), and qi
to be the composite of h1 with the ith coordinate linear form on ei B.
`
i∈I
If P is projective, it is clear that families (pi ) are precisely the systems
of generators for P. 

Setting Q = HomB (P, B) we can write U = QP (the set of sums

P
of elements of the form qp, q ∈ Q, p ∈ P). But qp = q pi (qi p) =
P P i
q(p j (q j pi ))(qi p) = (qp j )(q j pi )(qi p), which shows that U is gen-
i, j i, j
erated, as a two-sided ideal, by the q j pi . Moreover (ii) shows that
P = PU = PQP, and therefore U = QP = QPQP = U 2 .

Lemma 5.8. Let B be a commutative ring, M a finitely generated B- 69

module, and U an ideal of B such that MU = M. Then M(1 − a) = 0
for some a ∈ U .

Proof. If x1 , . . . , xn generate M, we can find ai j ∈ U such that xi =

P P
xi ai j , that is, xi (δi j − ai j ) = 0, i = 1, . . . , n. It follows by a well-
j j
known argument, that xi det(δi j − ai j ) = 0, that is, M det(δi j − ai j ) = 0.
But det(δi j − ai j ) is of the form 1 − a for some a ∈ U . 

Proposition 5.9. Let B be a commutative ring and P a projective B-

module. If either B is noetherian or P is finitely generated, the ideal im
g p of B is generated by an idempotent e, and ann P = (1 − e)B. Hence P
is a generator of mod − B if and only if P is faithful ( i.e., annP = 0).

Proof. The hypotheses guarantee that U = img p is a finitely generated

ideal of B, using (5.7) in the second alternative. From (5.7) we also
 56 2. Categories of modules and their equivalences

have PU = P and U 2 = U . Taking M = U in (5.8) we find an

e ∈ U such that U (1 − e) = 0. So U = U e and e2 = e. Moreover,
P
P = Pe so P(1 − e) = 0. If Pa = 0, then, since e = q j p j , we have
P
ea = q j p j a = 0 and thus a = (1 − e)a. Hence ann P = (1 − e)B.
Finally, P is a generator ⇔ im g p = B ⇔ e = 1 ⇔ annP = 0. 

The following corollary shows that for a commutative ring, B, the

concept of a faithfully projective object of mod − B is the same as that
of faithfully projective B- modules (as defined in §6 of Chapter 1).
70 Corollary 5.10. Let P be a module over a commutative ring, B. Then P
is a faithfully projective object of mod − B ⇔ P is finitely generated,
projective, and faithful.

Example
  1. Let k be a field and
 let B be the ring of matrices of the form
a b , a, b, c ∈ k. Let e = 1 0 . The right ideal P = eB is a finitely
0 c 00
generated, projective, faithful B−module. However, im g p = P , B, so
P not a generator of mod − B. Of course, B is not commutative.

Example 2 (Kaplansky). Let B be the (commutative) ring of continuous

real valued functions on the interval [0, 1], and let P be the ideal of
all functions vanishing in a neighbourhood of 0. It is known that P is
projective, and clearly it is faithful. However, it is easy to show that
im g p , B, so P is not a generator of mod − B. Of course, P is not
finitely generated.

6 Wedderburn structure theory

Given a ring, B, we shall denote by Mn (B), the ring of n × n matrices
with entries in B. If P is a B-module, we shall write p(n) for the direct
sum of n copies of P. There is a natural isomorphism

HomB (P(n) , P(n) ) ≈ Mn (HomB (P, P)).

71 We recall
Schur’s lemma. A homomorphism from a simple module over a ring
into another simple module is either an isomorphism or the zero map.
6. Wedderburn structure theory 57

Theorem 6.1. Let P be a faithfully projective right module over a ring

B. Suppose further that P is simple (This is rare !). Then

(1) A = HomB (P, P) is a division ring

(2) P is a finite dimensional left vector space over A, say P ≈ A(n) .

(3) B ≈ HomA (P, P)0 ≈ Mn (A) (via right multiplication ).

(4) B is a simple ring whose lattice of left ideals is isomorphic, via

b 7→ Pb, to the lattice of A-subspaces of P.

(5) Centre B ≈ HomA−B (P, P) ≈ centre A, and these are fields.

(6) P⊗B : B − mod → A − mod is an equivalence of categories.

Conversely, if P , 0 is a finite dimensional left vector space over a

division ring A, and if B = HomA (P, P0 ), then P is a faithfully projective
simple right B-module, and A ≈ HomB (P, P) (via left multiplication).

Proof. (1) follows from Schur’s lemma (2), (3), (4), (5) and (6) follows
from theorem 4.4 and proposition 5.6 (c).
If P , 0 is a finite dimensional left vector space over a division ring 72
A, then evidently P is a finitely generated projective generator of A-mod,
that is, a faithfully projective A-module. Moreover, B = HomA (P, P)0
operates transitively on the non-zero elements of P, so that P is a simple
B-module. It follows, as before, form theorem 4.4(4) and proposition
5.6 (c), that A ≈ HomB (P, P).
We now describe the classical method for finding a P as above. 

Lemma 6.2. If P is a minimal right ideal in a ring B, and if P2 , 0,

then P = eB for some idempotent e.

Proof. Since P2 , 0, there exists x ∈ P such that xP , 0. Schur’s

x
lemma then implies that P → − P (left multiplication by x) is an isomor-
phism, so that x = xe for a unique e ∈ P. But this implies x = xe2 , so
e2 = e. In particular, 0 , eB ⊂ P, and thus P = eB. 
 58 2. Categories of modules and their equivalences

Proposition 6.3. Let B be a ring having no idempotent two-sided ideals

other than 0 and B, and let P be a minimal right ideal such that P2 , 0.
Then P is a faithfully projective and simple B−module, so we have the
consequences of theorem 6.1.
Proof. P is finitely generated projective thanks to lemma 6.2. Moreover
0 , P ⊂ im g p is, according to lemma 5.7, an idempotent two sided
ideal. The hypothesis therefore implies that im g p = B, that is, P is a
generator of mod-B. Thus P is faithfully projective. Also P is simple by
hypothesis. 

73 Example. A right artinian ring B having no two-sided ideals other than

0 and B satisfies the hypothesis of the above proposition. For, it has
a minimal right ideal P , 0 and P2 cannot be zero (otherwise the two-
sided ideal BP , 0 would be distinct from B since it would be nilpotent).
We now generalize these results to the semi-simple case. Recall that
a module is called semi-simple if it is a direct sum of simple modules.
Lemma 6.4. Suppose a module M is the sum of a submodule N and a
family (S i )i∈I of simple submodules. Then there is a subset J of I such
that the map aa
fJ : N ( S j ) → M,
j∈J

induced by inclusions, is an isomorphism.

Proof. Among the subsets J for which f j is a monomorphism, we can
choose a maximal one, say J0 , by Zorn’s lemma. If fJ0 is not surjective,
there exists j ∈ I − J0 such that S j 1 im f J0 . Since S j is simple, im
f J0 ∩ S j = 0. Thus J0 ∪ { j} contradicts the maximality of J0 . 

Corollary. A submodule of a semi-simple module is a direct summand.

Proposition 6.5. Suppose B has a faithfully projective right B− module
P which is semi-simple. Then

P ≈ S 1(n1 ) ⊕ · · · ⊕ S r(nr ) ,

74 where S 1 , . . . , S r are a complete set of non-isomorphic simple B− mod-

6. Wedderburn structure theory 59

ules, and each ni > 0. If Di = HomB (S i , S i ), then Di is a division ring,

and Y
HomB (P, P) ≈ Mni (Di ).
1≤i≤r

Moreover, B is itself a semi-simple B− module.

Proof. Since P is finitely generated and semi-simple, it is a finite direct

sum of simple modules, and we can write P ≈ S 1(n1 ) ⊕ · · · ⊕ S r(nr ) , where
each S i is simple, S i not isomorphic to S j for i , j, and each ni > 0.
If S is any simple module, then S is a quotient of a coproduct of copies
of P and this clearly implies that S is isomorphic to some S i . Since
HomB (S i , S j ) = 0 for i , j (Schur’s lemma), we have HomB (P, P) ≈
HomB (S i(ni ) , S i(ni ) ) ≈
Q Q
Mni (Di ). Since B is a quotient, and hence
1≤i≤n 1≤i≤r
a direct summand of a coproduct of copies of P, B is also semi-simple.


Proposition 6.6. Let B be right artinian and let B have no nilpotent

two-sided ideals , 0. Then B is a semi-simple right B− module. As a
ring, B is a finite direct product of full matrix rings over division rings.
In particular, the center of B is a finite product of fields.

Proof. Once we know that B is a semi-simple right B− module, the

remaining conclusions follow from (6.5), since B is obviously faithfully
B−projective and the ring of endomorphisms of the right B-module B is
isomorphic to B. 

If b is a minimal (i.e. simple) right ideal of B, then b = eB with e2 = 75

e. This follows from lemma 6.2, provided b2 , 0. But b2 = 0 implies
that Bb , 0 is a nilpotent two-sided ideal contradicting our hypothesis.
We note that b, being a direct summand of B, is a direct summand of
any right ideal which contains b.
Now, if B is not semi-simple we can find a right ideal O minimal
with the property that O is not semi-simple. Choose a simple right ideal
b in O. Then O = b + O ′ (direct sum) for some right ideal O ′ ⊂
,
O. Then O ′ is semi-simple and thus O also is semi-simple, which is a
contradiction.
 60 2. Categories of modules and their equivalences

Proposition 6.7. B is a semi-simple B−module ⇔ every B− module is

projective.
Proof. ⇒ Let P be a right B− module. Then P is a quotient of a free
right B−module F which is semi-simple by assumption. It follows from
the corollary to (6.4), that P a direct summand of F. 

⇐ Let O be the sum of all simple right ideals of B. Then O is semi-

simple, by (6.4). By hypothesis, B/O is projective, so that B = O ⊕ b
for some right ideal b. If b , 0, then, being finitely generated, it has
a simple quotient module and hence a simple submodule (because the
simple quotient is projective). This contradicts the defining property of
O, and hence O = B. Thus B is semi-simple.
76 Definition 6.8. We call a ring B semi-simple if it is semi simple as a
right module over itself.
The results above show that is equivalent to B being a finite product
of matrix rings over division rings. In particular, the definition of semi-
simplicity of a ring is left-right symmetric.

7 Autoequivalence classes; the Picard group

If A is a k−category, k a comutative ring, we define
Pick (A )
to be the group of isomorphism classes (T ) of k−equivalences T : A →
A . The group law comes from composition of functors.
If A is a k−algebra, we define
Pick (A)
to be the group of isomorphism classes (P) of invertible A − A− bi-
modules (see definition 3.2) with law of composition induced by tensor
product: (P)(Q) = (P ⊗A Q). It follows from proposition 4.1 and the-
orem 4.4(3), that this is indeed a group with (P)−1 = (HomA (P, A)). In
the latter we can use either the left or the right A−module structure of P.
77 According to theorem 3.1:
7. Autoequivalence classes; the Picard group 61

Proposition 7.1. (P) 7→ (P⊗A ) and (T ) 7→ (T A) define inverse isomor-

phisms
Pick (A)−→
←− Pick (A − mod ).

Let P be an invertible A − A bimodule. If α, β ∈ Autk (A) are

k−algebra automorphisms, write

αP β

for the bimodule with additive group P and with operations

a · p = α(a)p, p · a = pβ(a) (p ∈ P, a ∈ A).

Thus P =1 P1 .
Suppose f : P → Q is a left A−isomorphism of invertible A −
A−bimodules. Since, via right multiplication, A = HomA (A P,A P)0 , we
can define α ∈ Autk (A) by

pα(a) = f −1 ( f (p)a)

or
f (pα(a)) = f (p)a, pεP, a ∈ A.
Then f :1 Pα → Q is a bimodule isomorphism. This proves, in particu-
lar, the statement (4) in the following

Lemma 7.2. For α, β, γ ∈ Autk (A) we have 78

(1) αA β ≈ γαA γβ

(2) 1A α ⊗A 1A β ≈ 1A αβ

(3) 1 Aα ≈1 A1 ⇔ α ∈ In Aut (A), the group of inner automorphisms of

A.
(In all cases above the symbol ≈ denotes bimodule isomorphism.)

(4) If P is an invertible A−A−bimodule and if P ≈ A as left A− modules,

then P ≈1 Aα as bimodules for some α ∈ Autk (A).
 62 2. Categories of modules and their equivalences

Proof. (1) The map α Aβ →γα Aγβ given by x 7→ γ(x) is the required
isomorphism.

(2) Using (1) we have 1 Aα ⊗A 1 Aβ ≈ α−1 A1 ⊗A1 Aβ ≈ α−1β ≈ 1 Aαβ .

(3) If f : 1 Aα →1 A1 is a bimodule isomorphism, then as a left A-auto-

morphism f (x) = xu, where u = f (1) is a unit in A. Moreover,
f (α(a)) = f (1.a) = f (1)a, which gives α(a)u = ua, that is, α(a) =
uau−1 for all a ∈ A.


Conversely, if α(a) = uau−1 for some unit u ∈ A, then f (x) = xu

defines a bimodule isomorphism 1 Aα →1 A1 .
The group Pick (A − mod ) ≈ Pick (A) operates on the isomorphism
classes of faithfully projective left A−modules. We now describe the
stability group of a faithfully projective module under this action.

79 Proposition 7.3. Let Q be a faithfully projective left A−module, and let

B denote the k−algebra Homa (Q, Q)0 . Then there is an exact sequence
ϕQ
1 → In Aut (B) → Autk (B) −−→ Pick (A) (*)

with

im ϕQ = {(P) ∈ Pick (A) P ⊗A Q ≈ Q as left A- modules}.

Proof. Suppose first that Q = A, so that B = A. Define ϕA (α) = (1 Aα ).

Lemma 7.2 tells us that this is a homomorphism with kernel InAut (A),
and with the indicated image. 

In the general case, we set Q∗ = HomA (Q, A). Then the func-
tor T = HomA (Q, ) ≈ Q∗ ⊗A : A − mod → B − mod is an
equivalence, with T Q = B. This induces an isomorphism Pick (A −
mod ) → Pick (B − mod ). By proposition 7.1, we obtain an isomor-
phism Pick (A) → Pick (B), and this maps (P) ∈ Pick (A) into (Q∗ ⊗A P ⊗A
Q) ∈ Pick (B).
We define now ϕQ : Autk (B) → Pick (A) as the composite Autk (B) →
≈
Pick (B) −
→ Pick (A), where the first map is defined as in the special case
7. Autoequivalence classes; the Picard group 63

treated in the beginning, and the second is the inverse of the isomor-
phism just mentioned. The exactness of (∗) follows from the special
case. Also, if (P) ∈ Pick (A), then P ⊗A Q ≈ Q as left A− modules
⇔ Q∗ ⊗A P ⊗A Q ≈ Q∗ ⊗A Q as left B−modules. Since Q∗ ⊗A Q ≈ B as
B − B−bimodules, the last statement in the proposition follows from the
special case.
Let C = center A. If P is an invertible A − A− bimodule, we can 80
define a map
αP : C → C
by requiring that
pt = αP (t)p, p ∈ P, t ∈ C.
This is possible because, the map p 7→ pt, being a bimodule endo-
morphism of P, is the left multiplication by a unique element in the
centre. Now αP is a k−algebra homomorphism (tp = pt for t ∈ k). If
p ⊗ q ∈ P ⊗A Q and t ∈ C, then (p ⊗ q)t = p ⊗ αQ (t)q = pαQ (t) ⊗ q =
α p αQ (t)(p ⊗ q). Thus
αP⊗Q = αP αQ .
Since, evidently αA = IdC , it follows from the invertibility of P, that
αP is an automorphism of C, and that (P) 7→ αP is a homomorphism
Pick (A) → Autk (C). The kernel is clearly PicC (A). Summarization gives

Proposition 7.4. If A is a k-algebra with center C, then there is an exact

sequence
0 → PicC (A) → Pick (A) → Autk (C).
If A is commutative, then

0 → PicA (A) → Pick (A) → Autk (A) → 1

is exact and splits.

Proof. The map α 7→ (1 Aα ) (see lemma 7.2) gives the required splitting 81
Autk (A) → Pick (A). 

Example . Let A be the ring of integers in an algebraic number field k,

and let G(k/Q) be the group of automorphisms of k. (k need not be
64 2. Categories of modules and their equivalences

Galois.) Evidently AutZ (A) ≈ G(k/Q), and PicA (A) is just the ideal class
group of A. Thus PicZ (A) is the semi-direct product of the ideal class
group of A with G(k/Q), which operates on the ideal group, and hence
on PicA (A). This is also the group of autoequivalences of the category
A−mod. In particular, PicZ (A) is finite (finiteness of class number) and
PicZ (Z) = {1}. Thus any autoequivalence Z − mod → Z − mod is
isomorphic to the identity functor.
Chapter 3

The Brauer group of a

commutative ring

In this chapter we prove the fundamental theorem on Azumaya alge- 82

bras, following largely the paper of Auslander Goldman [1]. In §4 we
obtain Rosenberg and Zelinsky’s generalization of the Skolem-Noether
theorem (see [1]). Finally we introduce the Brauer group Br(k) of a
commutative ring k. The functor End : FP → Az is cofinal, in the sense
of chapter 1, and we obtain an exact sequence

K1 FP → K1 Az → Ko Φ End → K0 FP → K0 Az → Br(k) → 0.

We have computed the groups Ki FP in chapter 1, and we further show

here that K0 Φ End ≈ Pic(k). The final result is that the functors Pic →
FP → Az yield an exact sequence

U(k) → K1 FP → K1 Az → Pic(k) → K0 FP → K0 Az → Br(k) → 0,

from which we can extract a short exact sequence

0 → (Q/Z ⊗Z U(k)) ⊕ (Q ⊗Z sK1 P) → K1 Az →t Pic(k) → 0,

the last group being the torsion subgroup of Pic(k). This gives a fairly
effective calculation of K1 Az.

65
 66 3. The Brauer group of a commutative ring

1 Separable Algebras
Let k be a commutative ring. If A is a k−algebra, we write A0 for the
83 opposite algebra of A, and Ae = A ⊗k A0 . A two-sided A− module M can
be viewed as a left A ⊗k A0 module: We define the scalar multiplication
by
(a ⊗ b)x = axb, x ∈ M, a, b ∈ A.
In particular, A is a left Ae − module, in a natural manner, and we have
an exact sequence
0 → J → Ae → A → 0 (1.1)
of Ae − linear maps (where a ⊗ b ∈ Ae goes to ab ∈ A). If needed, we
shall make the notation more explicit by writing

Ae = (A/k)e ,

and
J = J(A) = J(A/k).
We define k−linear map
δ : A → J,
by setting δ(a) = a ⊗ 1 − 1 ⊗ a.
Lemma 1.2. Im δ generates J as a left ideal, and δ satisfies δ(ab) =
a(δb) + (δa)b.
P P
Proof. Clearly im δ ⊂ J. If x = ai ⊗ bi ∈ J, that is, if ai bi = 0, then
P P P P
84 x = ai ⊗bi − ai bi ⊗1 = (ai ⊗1)((1⊗bi )−(bi ⊗1)) = − ai δbi . Finally,
δ(ab) = ab⊗1−1⊗ab = (a⊗1)(b⊗1−1⊗b)+(a⊗1)(1⊗b)−(1⊗b)(1⊗a) =
a(δb) + (1 ⊗ b)δa = a(δb) + (δa)b. 

Corollary 1.3. If M is a left Ae − module and N is a right Ae −module,

there are natural isomorphisms

HomA e(A, M) ≈ {x ∈ M ax = xa, for all a ∈ A},

and

N ⊗Ae A ≈ N/ (Submodule generated by ax − xa, a ∈ A, x ∈ N).

1. Separable Algebras 67

Proof. Since A ≈ Ae /J, HomAe (A, M) ≈ {x ∈ M Jx = 0} = {x ∈

M (δa)x = 0∀a ∈ A} = {x ∈ M ax = xa∀a ∈ A}. The other part is
trivial. 

For a two-sided A−module M we shall denote the subgroup {x ∈

M ax = xa∀a ∈ A} by M A. Note that if A is a subalgebra of a k−algebra
B, then BA is just the centralizer of A in B. In particular, AA = centre A.
We denote by Derk (A, M) the k−module of all k−derivations of A
into M, that is, k−linear maps d : A → M satisfying d(ab) = ad(b) +
(da)b, a, b ∈ A. If f : M → N is Ae −linear, then d 7→ f d defines a
k−linear map Derk (A, M) → Derk (A, N).. For example, if x ∈ M and if
f : Ae → M is defined by f (1) = x, then the composite

δ f
− J ֒→ Ae −
A→ →M

is a derivation, called the inner derivation dx defined by x. Thus, if

a ∈ A, dx (a) = (δa)x = ax − xa.

Proposition 1.4. For an Ae -module M, the map f 7→ f δ defines an 85

isomorphism
HomAe (J, M) → Derk (A, M),

with inner derivations corresponding to those f which can be extended

to Ae .

Proof. Since im δ generated J, we have f δ = 0 ⇒ f = 0. 

Suppose d ∈ Derk (A, M). We can define a k−linear map f : Ae → M

P P
by setting f ( ai ⊗ bi ) = − ai d(bi ). This satisfies f δa = f (a ⊗ 1 −
1 ⊗ a) = −ad(1) + 1d(a) for all a ∈ A. But d(1) = d(12 ) = 1d(1) +
d(1)1 = 2d(1) so that d(1) = 0. Thus f δ = d. It remains to show
that f /J is Ae −linear. If x = ai ⊗ bi ∈ J, we must show that f ((a ⊗
P
P P
b)x) = (a ⊗ b) f (x). But f ((a ⊗ b)x) = f ( aai ⊗ bi b) = − aai d(bi b) =
P
− aai (bi db + d(bi )b) = (a ⊗ b) f (x).
The derived functors of M 7→ M A are called the Hochschild co-
homology groups of A with coefficients in M. We denote them by
 68 3. The Brauer group of a commutative ring

H i (A, M). By virtue of (1.3), H i (A, M) ≈ ExtiAe (A, M). The exact se-
quence (1.1) gives us an exact sequence

0 → HomAe (A, M) → HomAe (Ae , M)

→ HomAe (J, M) → Ext1Ae (A, M) → 0,

which we can rewrite, using (1.3) and (1.4), to obtain:

Proposition 1.5. There is an exact sequence

0 → M A → M → Derk (A, M) → H 1 (A, M) → 0.

86 so that H 1 (A, M) = the k−module of k−derivations of A into M, modulo

the k−submodule of inner derivations.
If C = AA = centre A, then C ⊗ 1 ⊂centre Ae , so we can view the
above exact sequence as a sequence of C−modules and C−linear maps.
Proposition and Definition 1.6. A k−algebra A is called separable, if
it satisfies the following conditions, which are equivalent:
(1) A is a projective Ae −module.

(1)bis M 7→ M A is an exact functor on Ae − modules.

(1)ter (Ae )A → AA → 0 is exact.

(2) If M is an Ae −module, then every k−derivation A → M is inner.

(2)bis the derivation δ : A → J is inner.

Proof. Since M A ≈ HomAe (A, M), the implications (1) ⇔ (1)bis ⇒
(1)ter are clear. If HomAe (A, Ae ) → HomAe (A, A) → 0 is exact, then
1A factors through Ae , so that A is Ae projective, thus proving (1)ter ⇒
(1). 

(1) ⇔ (2) by virtue of the identifications Ext1Ae (A, M) = H 1 (A, M) =

derivations modulo inner derivations. Also the implication (2) ⇒ (2)bis
is obvious. Finally, proposition 1.4 shows that δ is inner ⇔ 1J extends
to a homomorphism Ae → J that is, ⇔ the exact sequence (1.1) splits.
This proves (2)bis ⇒ (1).
 2. Assorted lemmas 69

87 Corollary 1.7. If A/k is separable with centre C, then for an Ae −module

M, there is a split exact sequence of C−models,
0 → M A → M → Derk (A, M) → 0.
In particular, C is a C−direct summand of A.
This follows directly from (1.5) and the definition above.
Corollary 1.8. If A → B is an epimorphism of k−algebras, with A/k
separable, then B/k is separable, and centre B= image of centre A.
Proof. If M is a two-sided B−module, then evidently M B = M A , so that
M 7→ M B is an exact functor, that is, B is separable. Also, AA → BA =
BB → 0 is exact. 

2 Assorted lemmas
The reader is advised to skip this section and use it only for references.
fi
Lemma 2.1 (Schanuel’s lemma). If 0 → Ni → Pi − → M → 0 are exact
with Pi projective, i = 1, 2, then P1 ⊕ N2 ≈ P2 ⊕ N1 .
Q
Proof. If P1 P2 = {(x1 , x2 ) ∈ P1 ⊕ P2 f1 (x1 ) = f2 (x2 )}, then the
M Q
coordinate projections give us maps P1 M P2 → Pi , i = 1, 2, and a 88
commutative diagram
0 0

 
N2 N2

 
0 / N1 / P1 Π M P2 / P2 /0


0 / N1 / P1 /M /0

 
0 0
 70 3. The Brauer group of a commutative ring

with exact rows and columns, as is easily checked. Since the Pi are
Q
projective, we conclude that P1 ⊕ N2 ≈ P1 P2 ≈ P2 ⊕ N1 . 
M

Let A be a ring. An A− module M is called finitely presented, if

there exists an exact sequence

F1 → F0 → M → 0

of A−linear maps with Fi a finitely generated free A− module, i = 0, 1.

Corollary 2.2. (a) If 0 → M ′ → M → M ′′ → 0 is an exact sequence

of A−modules, with M and M ′′ finitely presented, then M ′ is finitely
generated.

89 (b) If A is commutative, and M and N are finitely presented A−modules,

then so is M ⊗A N.

(c) If A is an algebra over a commutative ring k, and if A is finitely pre-

sented as a k−module, then A is finitely presented as an Ae −module.

Proof. (a) Case I. Suppose M is projective. Then the result follows

easily from the definition and Schanuel’s lemma.
General Case. Let f : P → M be surjective with P finitely generated
and projective, and let f ′′ : P → M ′′ be the epimorphism obtained by
composing f with M → M ′′ . We have a commutative diagram

0 /P P /0

f′ f f ′′
  
0 / M′ /M / M ′′ /0

with exact rows. The exact sequence

0 = ker f ′ → ker f → ker f ′′ → coker f ′ = M ′ → coker f = 0

shows that M ′ is finitely generated, since by case I, ker f ′′ is.

(b) follows easily from right exactness.

2. Assorted lemmas 71

(c) If A is finitely presented as a k− module, then Ae is finitely pre-

sented as a k−module, This implies that J is finitely generated as
a k−module and a fortiori as an Ae − module.


Lemma 2.3. Let Ki be a commutative k− algebra and let Mi, Ni be 90

Ki −modules, i = 1, 2. There is a natural isomorphism
≈
(M1 ⊗k M2 ) ⊗K1 ⊗k K2 (N1 ⊗k N2 ) −
→ (M1 ⊗K1 N1 ) ⊗k (M2 ⊗K2 N2 )

given by (m1 ⊗ m2 ) ⊗ (n1 ⊗ n2 ) 7→ (m1 ⊗ n1 ) ⊗ (m2 ⊗ n2 ). If the Mi and

Ni are Ki −algebras, then the above map is an isomorphism of K1 ⊗k
K2 −algebras.
Proof. Straightforward. 

Corollary 2.4. (a) If Ki is a commutative k-algebra, and Ai a Ki -alge-

bra, i = 1, 2, then (2.3) defines a natural isomorphism

(A1 /K1 )e ⊗k (A2 /K2 )e ≈ (A1 ⊗k A2 /K1 ⊗k K2 )e

(b) If K and A are k−algebras, K commutative, then (K ⊗k A/K)e ≈

K ⊗k (A/k)e .
Proof. (a) In (2.3) we take Mi = Ai and Ni = A0i . Evidently (A1 ⊗k
A2 )0 = A01 ⊗k A02 .

(b) Set A1 = K1 = K, A2 = A and K2 = k in (a).



Lemma 2.5. Let Ki be a commutative k-algebra, let Ai be a Ki -algebra,

and let Mi and Ni be Ai -modules, i = 1, 2. The k-bilinear map ( f1 , f2 )
7→ f1 ⊗ f2 defines a K1 ⊗k K2 -homomorphism

HomA1 (N1 , M1 ) ⊗k HomA2 (N2 , M2 ) → HomA1 ⊗k A2 (N1 ⊗k N2 , M1 ⊗k M2 ).

It is an isomorphism in either of the following situations: 91

(i) Ni is a finitely generated projective Ai −module, i = 1, 2.

 72 3. The Brauer group of a commutative ring

(ii) N1 and M1 are finitely generated projective A1 −modules, A1 is

k−flat, and N2 is a finitely presented A2 −module.

Proof. The first assertion is clear.

(i) By additivity we are reduced to the case Ni = Ai , i = 1, 2, and then

the assertion is clear.

(ii) By additivity again we can assume that N1 = M1 = A1 .

Write S N2 = A1 ⊗k HomA2 (N2 , M2 ), T N2 = HomA1 ⊗k A2 (A1 ⊗k N2 ,

A1 ⊗k M2 ). We have a map S N2 → T N2 . This is a isomorphism for
N2 = A2 , and therefore, for N2 = A(n) (n) (m)
2 . Let now A2 → A2 → N2 → 0
be an exact sequence. S and T being left exact contravariant functors in
N2 , we obtain a commutative diagram

0 / S N2 / S A(m) / S A(n)
2 2

  
0 / T N2 / T A(m) / T A(n)
2 2

with exact rows. The second and third vertical maps being isomor-
phisms, if follows that the first one is also an isomorphism. 

Corollary 2.6. If Ki is a commutative k−algebra, and if Ai /Ki is a sep-

92 arable algebra, i = 1, 2, then A1 ⊗k A2 /K1 ⊗k K2 is a separable algebra
with centre = (centre A1 ) ⊗k (centre A2 ). More generally, if Mi is an
(Ai /Ki )e module, then the natural map

M1A1 ⊗k M2A2 → (M1 ⊗k M2 )A1 ⊗k A2

is an isomorphism.

Proof. We have, by hypothesis, (2.4), and (2.5), an isomorphism

M1A1 ⊗k M2A2 = Hom(A1 /K1 ) e(A1 , M1 ) ⊗k Hom(A2 /K2 ) e(A2 , M2 ) →

Hom(A1 /K1 )e ⊗k (A2 /K2 )e (A1 ⊗k A2 , M1 ⊗k M2 )
= (M1 ⊗k M2 )A1 ⊗k A2
2. Assorted lemmas 73

Applying this to (A1 /K1 )e ⊗k (A2 /K2 )e = (A1 ⊗k A2 /K1 ⊗k K2 )e → A1 ⊗k

A2 → 0 we get a commutative diagram

((A1 ⊗k A2 /K1 O ⊗k K2 )e )A1 ⊗k A2 / (A ⊗ A )A1 ⊗k A2

1 k O2

((A1 /k1 )e )A1 ⊗k ((A2 /K2 )e )A2 / (AA1 ⊗k AA2 )

1 2

in which the vertical maps are isomorphisms and the lower horizontal
map is surjective, by hypothesis and right exactness of ⊗k . It follows
that the upper map is also surjective, and this finishes the proof, using
criterion (1.6)(1)ter for separability. 

Corollary 2.7. If Ai /k is a separable algebra, i = 1, 2, then (A1 ⊗k A2 )/k 93

is separable with centre A1 ⊗k centre A2 as its centre.
Corollary 2.8. Suppose K and A are k-algebras. Suppose further that
K is k-flat.
(a) If M and N are A-modules with N finitely presented, then

K ⊗k HomA (N, M) → Homk⊗k A (K ⊗k N, K ⊗k M)

is an isomorphism.

(b) If K is commutative, and if A is finitely presented as anAe -module

then for an Ae -module M, the map

K ⊗k (M A ) → (K ⊗k M)K⊗k A

is an isomorphism.
Proof. The statement (a) follows from (2.5) (ii) with N1 = M1 = A1 =
K1 = K, and K2 = k. The statement (b) follows from (a) by substituting
Ae for A, and A for N. 

Corollary 2.9. Let K and A be k-algebras, with K commutative.

(a) A/k separable ⇒ K ⊗k A/K is separable with centre (K ⊗k A) = K⊗k
(centre A).
 74 3. The Brauer group of a commutative ring

(b) If K is faithfully k-flat and if A is a finitely presented Ae -module,

then (K ⊗k A)/K separable ⇒ A/k separable.

94 Proof. (a) is a special case of (2.6).

(b) Suppose K ⊗k A/K is separable. Corollary 2.4 implies that (K ⊗k

A/K)e → K ⊗k A is isomorphic to K ⊗k ((A/k)e → A). Then
(2.8)(b) further implies that ((K ⊗k A/k)e )K⊗k A → (K ⊗k A)K⊗k A
is isomorphic to K ⊗k (((A/k)e )A → AA ). Therefore, by hypothesis,
K ⊗k (((A/ke )A → AA ) is surjective. Since K is faithfully k-flat, this
implies that ((A/k)e )A → AA is surjective, so that A/k is separable
(see (1.6)(1) ter).


Example. If K is a noetherian local ring, in (2.9)(b) we can take K to be

completion of k.

Corollary 2.10. If A/k is a finitely Ae -presented k-algebra, then A/k is

separable ⇔ AM /kM is separable for all maximal ideals M of k.
Q
Proof. Take K = kM in (2.9)(b). Alternatively, repeat the proof of
M
(2.9)(b) and at the end use the fact that a k-homomorphism f is surjec-
tive ⇔ fM is surjective for all M . 

Corollary 2.11. Suppose Ai is a k-algebra and that Pi is a finitely gen-

erated projective Ai -module, i = 1, 2. Then

EndA1 (P1 ) ⊗k EndA2 (P2 ) → EndA1 ⊗k A2 (P1 ⊗k P2 )

is an algebra isomorphism.

Proof. Set Ni = Mi = Pi and Ki = k in (2.5) (i). 

95 Corollary 2.12. Suppose P1 and P2 are finitely generated projective

k-modules. Then

Endk (P1 ) ⊗k Endk (P2 ) → Endk (P1 ⊗k P2 )

is an algebra isomorphism.
2. Assorted lemmas 75

Proof. Set Ai = k in (2. 11). 

Proposition 2.13. Let P be a finitely generated projective k-module.

Then A = Endk (P) is a separable algebra with centre k/annP.

Proof. Both centre A and B ann P commute with localization, and hence
we can use (2.10) to reduce to the case when P is free, say P ≈ k(n) , so
that A ≈ Mn (k). Denoting the standard matrix algebra basis by (ei j ), we
n
ei1 ⊗ e1i ǫAe . Then (ers ⊗ 1)e = ers ei1 ⊗ e1i = er1 ⊗ e1s and
P P
set e =
i=1 P
(1⊗ers )e = ei1 ⊗e1i ers = er1 ⊗e1s . Hence eǫ(Ae )A . Under (Ae )A → AA ,
e maps into ei1 e1i = eii = 1. Hence (Ae )A → AA is surjective. Thus
P P
A is separable, by (1.6)(1)ter . Theorem (6.1)(5) of chapter 2 implies that
centre Mn (k) = k. 

Lemma 2.14. Let f : M → M be a k-endomorphism, k a commuta-

tive ring. Suppose that M is either noetherian or finitely generated and
projective. Then, if f is surjective, it is an automorphism.

Proof. If ker f , 0, then f surjective implies that ker f n is a strictly

ascending chain of submodules, an impossibility if M is noetherian, If M
is projective, then M ≈ M⊕ ker f and localization shows that ker f = 0 96
if M is finitely generated. 

Proposition 2.15. Let k be a local ring with maximal ideal M , and let
A be a k-algebra, finitely generated as a k-module. Suppose that either
k is noetherian or that A is k-projective. Then if A/M A is a separable
(k/M )-algebra, A is a separable k-algebra.

Proof. Consider δ : A → J = J(A/k). Let k′ denote reduction modulo

M ; e. g. k′ = k/M . Then δ induces a k′ -derivation δ′ : A′ → J ′ , where
J ′ is a two-sided A′ -module. By hypothesis and criterion (1.6)(2), δ′
must be inner, δ′ (a′ ) = a′ e′ − e′ a′ = ae′ − e′ a = δ(a)e′ , for some e′ ǫ J ′ ,
coming from say eǫ J. It follows that δ(a)e ≡ δ(a) mod M J, so (1.2)
implies J = Je + M J. The exact sequence

0 → J → Ae → A → 0
 76 3. The Brauer group of a commutative ring

shows that J is noetherian if k is, and that J is k-projective and finitely

generated if A is. Hence we can apply lemma 2.14 to the k-homomor-
e
phism J → − J, provided the latter is surjective. But this follows from
Nakayama’s lemma, since J = Je + M J.
e
Now the composite J ֒→ A → − J is an automorphism of J, so that J
is an Ae -direct summand of Ae . This proves that A is Ae -projective, as
required. 

Lemma 2.16. Let f : P → M be a k-homomorphism with P finitely

97 generated and projective. Denote the functor Homk (, k) by ∗ . Then f has
a left inverse ⇔ f ∗ : M ∗ → P∗ is surjective. If M is finitely presented,
then (coker f ∗ )M = coker ( fM )∗ , so that f has a left inverse ⇔ fM does
for all maximal ideals M .

Proof. f left invertible ⇒ f ∗ right invertible ⇒ f ∗ surjective ⇒ f ∗ right

invertible (because P∗ is projective) ⇒ f ∗∗ : P∗∗ → M ∗∗ left invertible.
The commutative square

f
P /M

≈
 
P∗∗ f ∗∗
/ M ∗∗

shows that f ∗∗ left invertible ⇒ f left invertible. 

The natural homomorphism

(M ∗ )M = (Homk (M, K))M → (MM )∗ = HomkM (MM , kM )

is an isomorphism for M finitely presented, by (2.8)(a). Hence since P

is also finitely presented, we have ( f ∗ )M ≈ ( fM )∗ in this case, so that
by exactness of localization, (coker f ∗ ) M ≈ coker ( fM )∗ .

Corollary 2.17. If A is a faithfully k-projective k-algebra, then k is a

direct summand of A.
 2. Assorted lemmas 77

Proof. We want k → A to have a left inverse, and (2. 16) plus our
98 hypothesis makes it sufficient to prove this for k local, say with maxi-
mal ideal M . Then 1ǫA/M A is a part of a k/M -basis for A/M A, so
Nakayama’s lemma implies that 1ǫA is a part of a k-basis of A. 

Proposition 2.18. Let A and B be k-algebras with A a faithfully projec-

tive k-module. Then A ⊗k B/k separable ⇒ B/k separable.

Proof. A is k-projective implies that Ae is k-projective. Hence (A ⊗k

B)e ≈ Ae ⊗k Be is Be-projective. Thus, if A ⊗k B is (A ⊗k B)e -projective,
then it is Be -projective. Corollary 2.17 and our hypothesis implies A ≈
k ⊕ A′ as a k-module, so that A ⊗k B ≈ B ⊕ (A′ ⊗k B) as a Be-module.
Thus Be -projectivity of A ⊗k B ⇐ Be-projectivity of B. 

Proposition 2.19. Suppose A is a K-algebra and that K is a k-algebra.

Then

(1) A/K and K/k separable ⇒ A/k separable.

(2) A/k separable ⇒ A/k separable.

If A faithfully K-projective, then A/k separable ⇒ K/k separable.

Proof. (1) K/k separable means that

0 → J(K/k) → K e → K → 0

splits as an exact sequence of K e -modules. Hence

0 → (A/k)e ⊗K e J(K/k) → (A/k)e → (A/k)e ⊗ke K → 0

splits as an exact sequence of (A/k)e -modules, so that (A/k)e ⊗K e K 99

is a projective (A/k)e -module. it follows easily from corollary 1.3
that (A ⊗k Ao ) ⊗K e K ≈ A ⊗K Ao = (A/K)e . Hence if we further
assume that A is (A/K)e -projective, it follows from the projectivity
of (A/K)e over (A/k)e , remarked above, that A is (A/k)e -projective.
 78 3. The Brauer group of a commutative ring

(2) In the commutative diagram with exact rows and columns

(A/k)e /A /0


(A/K)e /A /0


0
if the top splits, then so much the bottom. Suppose A is faithfully
K-projective. Then (A/k)e is (K/k)e -projective, so that A is (K/k)e -
projective, assuming that (A/k) is separable. By corollary 2.17 K is
a K-direct summand, of A, hence a (K/k)e -direct summand, so we
conclude that K is (K/k)e -projective, as claimed.

Proposition 2.20. (a) If A1 and A2 are k-algebras, then A1 × A2 /k is
separable ⇔ A1 /k and A2 /k are.
(b) If Ai is a ki -algebra, i = 1, 2, then A1 × A2 /k1 × k2 is separable
⇔ A1 /k1 and A2 and A2 /k2 are.
100 Proof. (a) (A1 ×A2 )e = (A1 ⊗k A01 )×(A2 ⊗k A02 )×(A1 ⊗k A02 )×(A2 ⊗k A01 ) =
Ae1 × Ae2 × B, and A1 × A2 is an (A1 × A2 )e -module annihilated by B.
As such it is the direct sum of the Aei -modues Ai . Thus A1 × A2 is
(A1 × A2 )e -projective ⇔ Ai is Aei -projective
(b) Any k1 × k2 -module or algebra splits canonically into a product of
one over k1 and one over k2 . In particular (A1 × A2 /k1 × k2 )e =
(A1 /k1 )e × (A2 /k2 )e , so A1 × A2 is (A1 × A2 /k1 × k2 )e -projective ⇔ Ai
is (Ai /ki )e -projective.


3 Local criteria for separability

Theorem 3.1. Let A be a k-algebra, finitely generated as a k-module.
Suppose either that k is noetherian or that A is a projective k-module.
Then the following statements are equivalent:
3. Local criteria for separability 79

(1) A/k is separable.

(2) For each maximal ideal M of k, A/M A is a semi-simple k/M -

algebra whose centre is a product of separable field extensions of
k/M .

(3) For any homomorphism k → L, L a field, L ⊗k A is a semi-simple

algebra.

We will deduce this form the following special case:

Theorem 3.2. Let A be a finite dimensional algebra over a field k. The

following statements are equivalent:

(1) A/k separable

(2) L ⊗k A is semi-simple for all field extensions L/k.

(3) For some algebraically closed field L/k, L ⊗k A is ε product of full 101
matrix algebras over L.

(4) A is semi-simple and centre A is a product of separable field exten-

sions of k.

We first prove that (3.2) ⇒ (3.1):

(1) ⇒ (3). If k → L, then L ⊗k A/L is separable, by (2.9), and we
now apply (1) ⇒ (4) of (3.2).
(3) ⇒ (2). Apply (2) ⇒ (4) of (3.2), where L ranges over field
extensions of k/M .
(2) ⇒ (1). From (4) ⇒ (1) of (3.2) we know that A/M A is a sepa-
rable (k/M )-algebra so the hypothesis on A and proposition 2.15 imply
that AM /kM is separable, for all M . (1) now follows from corollary
2.10.

Proof of Theorem 3.2. (1) ⇒ (2)′ . Since A/k separable implies that
(L ⊗k A)/L is separable, it suffices to show that A/k separable ⇒ A
is semi-simple. Let M, N be left A-modules. Then Homm (M, N) is a
two-sided A-module, i.e., an Ae -module, and HomAe (A, Homk (M, N)) =
HomA (M, N) clearly (see (1.3)). Since k is a field, Homk (M, ) is an exact
 80 3. The Brauer group of a commutative ring

functor. Since A is Ae -projective (by assumption), HomAe (A, ) is exact.

Hence HomA (M, ) is an exact functor, so every A-module is projective.
Proposition 6.7 of Chapter 2 now implies that A semi-simple.
102 (2) ⇒ (3). This follows from the structure of semi-simple rings plus
the fact that there are no non-trivial finite dimensional division algebras
over an algebraically closed field.
(3) ⇒ (1). By assumption, L⊗k A is a product of full matrix algebras
over L. Proposition 2.13 and 2.20 imply that L⊗k A/L is separable. Since
L is faithfully k-flat (k is a field!), (2.9)(b) implies that A/k is separable.
The proof of (1) ⇔ (4) will be based upon the next two lemmas,
which are special cases of the theorem.

Lemma 3.3. A finite field extension C/k is separable as a k-algebra ⇔

it is separable as a field extension of k.

Proof. If k ⊂ K ⊂ C, then C/k is separable, in either sense ⇔ C/K

and K/k are, in the same sense. This follows from proposition 2.18 in
one case, and from field theory in the other. An induction on degree
therefore reduces the lemma to the case C = k[X]/( f (X)). Let L be an
algebraic closure of k, and write f (X) = (X−ai )ei in L[X], with a′i s dis-
Q
i
tinct. Then C is a separable field extension ⇔ L ⊗k C = L[X]/ f (X)L[X]
has no nilpoint elements ⇔ L⊗k C is a product of copies of L ⇔ L⊗k C/L
is separable ⇔ C is a separable algebra over k, by (1) ⇔ (3) of (3.2),
which we have already proved. 

We now prove the implication (1) ⇔ (4) of (3.2). We have already

proved that A/k is separable implies that A is semi-simple. Hence the
103 centre C of A must be a finite product of field extension of k. in par-
ticular A is a faithfully projective C-module, so by proposition 2.19,
A/k separable ⇒ C/k separable. The last part of (4) now follows from
proposition 2.20 and lemma 3.3 above.

Lemma 3.4. Let k be a field, and suppose that A is a finite dimensional

k-algebra, simple and central (i.e.centre A = k). If B is any k-algebra,
every two-sided ideal of A ⊗k B is of the form A ⊗k J for some two sided
ideal J of B.
4. Azumaya algebras 81

Proof. According to theorem 6.1 chapter 2 there is a division algebra

D and an n > 0 such that A ≈ Mn (D) = D ⊗k Mn (k). Theorem 4.4 of
chapter 2 contains the lemma when A = Mn (k), in which case A ⊗k B =
Mn (B) = EndB (B(n) ). It therefore suffices to prove the lemma for A = D,
a division algebra. If (ei ) is a k-basis for B, then (1 ⊗ ei ) is a left D-basis
for D ⊗k (B). Let I be a two-sided ideal of D ⊗k B. Then I is a D-
subspace of D ⊗k B, and it is (clearly) generated by the “primordial”
elements of I with respect to the basis (1 ⊗ ei ), i.e. by those elements
P
x = (di ⊗1)(1⊗ei ) , 0 of I such that S (x) = {i|di , 0} does not properly
contain S (y) for any y , 0 in I, and such that the least one di = 1. If
x such an element and if d , 0 is in D, then x(d ⊗ 1) ∈ I, because I is
P P
a two-sided ideal. Now x(d ⊗ 1) = (di ⊗ ei )(d ⊗ 1) = di d ⊗ ei so
S (x(d ⊗ 1)) = S (x). Subtracting (d′ ⊗ 1)x from x(1 ⊗ d) will therefore
render S ((d′ ⊗ 1)x − x(1 ⊗ d)) a proper subset of S (x), for a suitable
d′ ǫD. 

Since x is primordial, this implies (d′ ⊗ 1)x = x(d ⊗ 1), i.e. that 104
d′ di
⊗ ei = di d ⊗ ei . Some di = 1 so we have d′ = d. Moreover,
P P
di d = ddi for all i. By assumption centre D = k, so xǫk ⊗ B = 1 ⊗ B.
Setting 1 ⊗ J = I ∩ (1 ⊗ B), we therefore have I = D ⊗ J.
We shall now prove the implication (4) ⇒ (1) of theorem 3.2. Let
C = centre A. To show that A/k is separable Proposition 2.19 makes it
sufficient to show that A/C and C/k are separable. In each case, more-
over, proposition 2.20 reduces the problem to the case when C is a field,
Separability of C/k then results from the hypothesis and lemma 3.3.
Let L be an algebraic closure of C. Then it follows from lemma 3.4
that L ⊗C A is simple, hence a full matrix algebra over L. Separability of
A/C now follows from the implication (3) ⇒ (1) which we have proved.
Thus the proof of the theorem 3.2 is complete.

4 Azumaya algebras
Theorem and Definition 4.1. An azumaya algebra is a k-algebra A
satisfying the following conditions, which are equivalent:

(1) A is a finitely generated k-module and A/k is central and separable.

 82 3. The Brauer group of a commutative ring

(2) A/k is central and A is a generator as an Ae -module.

(3) A is a faithfully projective k-module, and the natural representation

Ae → Endk (A) is an isomorphism.

105 (4) The bimodule Ae Ak is invertible (in the sense of definition 3.2 of
chapter 2), i.e. the functors

(N  / A ⊗k N)
/
k − mod o Ae − mod

(M A o M)
are inverse equivalences of categories.

(5) A is a finitely generated projective k-module, and for all maximal

ideals M of k, A/M A is a central simple k/M -algebra.

(6) There exists a k-algebra B and a faithfully projective k-module P

such that A ⊗k B ≈ Endk (P).

Proof. (1) ⇒ (2). Let M be a maximal two-sided ideal of A, and set

M = M ∩ k. According to (1.8) and our hypothesis A/m is a separable
k-algebra with centre k/M . Since A/M does not have two-sided ideals,
its centre is a field. Thus M is a maximal ideal of k, so A/M A is a
central separable algebra over k/M , and it follows from theorem 3.2
that A/M A is simple. Consequently m = M A. Applying this to Ae ,
which, by (2.7), is also a separable k-algebra, we conclude that every
maximal two-sided ideal of Ae is of the form M Ae for some maximal
ideal M of k. 

106 Viewing A as a left Ae -module we have the pairing gA : A ⊗k HomAe

(A, Ae ) → Ae , and its image is a two -sided ideal which equals Ae ⇔ A is
a generator as an Ae -module (see(5.6) of chapter 2). If imgA , Ae , then
im gA is contained in some maximal two-sided ideal of Ae , so, according
to the paragraph above, im gA ⊂ M Ae , for some maximal ideal M of k.
Now lemma 5.7 of chapter 2, plus our hypothesis that A is Ae -projective,
4. Azumaya algebras 83

imply that A = (imgA )A ⊂ M A. But from (1.7), k = centre A is a direct

summand of A. So A = M A ⇒ k = M , which is a contradiction.
(2) ⇒ (3). A is a generator of Ae − mod , so the pairing

gA : HomAe (A, Ae ) ⊗k A → Ae

is surjective. It follows now from theorem 4.3 and proposition 5.6 of

chapter 2 that A is a finitely generated projective k-module, and that
Ae → Endk (A) is an isomorphism. Since A is a faithful k-module (k
being centre of A), corollary 5.10 of chapter 2 implies that it is faithfully
projective.
(3) ⇒ (4). This follows directly from proposition 5.6 and definition
3.2 of chapter 2.
(4) ⇒ (1). is trivial once we note that centre A = HomAe (A, A).
(1) ⇒ (5) follows from (1) ⇒ (3) of theorem 3.1.
(5) ⇒ (1). Theorem 3.1 shows that A/k is separable.
Let C =centre A. Then C is a C-direct summand of A, and hence
a finitely generated projective k-module, since A is so. We have a ho- 107
momorphism k → C, and (1.8) implies that k/M → C/M C is an
isomorphism for all maximal ideals M of k. This implies that k → C
is surjective, since the cokernel is zero modulo all maximal ideals of k,
and hence it splits, because C is projective. The kernel of k → C is also
zero, since it is zero modulo all maximal ideals of k. Thus k → C is an
isomorphism.
(3) ⇒ (6). Take B = A0 and P = A.
(6) ⇒ (1). Endk (P) is faithfully projective, since P is. Since A ⊗k
B ≈ Endk (P), it follows from proposition 6.1 of chapter 1 that B is
faithfully projective. Proposition 2.13 says that Endk (P)/k is central and
separable, so that, by proposition 2.18, A/k is separable. Similarly B/k
is separable. It follows from (2.7), that (centre A)⊗k (centre B) = centre
Endk (P) = k. Hence centre A has rank 1, as a projective k-module, and
so centre A = k, since k is a direct summand of A ⊗k B and therefore of
centre A.

Corollary 4.2. If A/k is an azumaya algebra, then U 7→ U A is a

bijection from the ideals of k to the two-sided ideals of A.
 84 3. The Brauer group of a commutative ring

Proof. This follows from theorem 4.4 of chapter 2, since two-sided ide-
als of A are simply Ae -submodules of A. 

Corollary 4.3. Let A ⊂ B be k-algebras with A azumaya. Then the

natural map A ⊗k BA → B is an isomorphism.

Proof. This is a special case of the statement (1) of theorem 4.1. 

108 Corollary 4.4. Every endomorphism of an azumaya algebra is an au-

tomorphism.

Proof. Suppose f : A → A is an endomorphism of an azumaya algebra

A/k. By (4.2), ker f = U A for some ideal U of k and hence ker f = 0.
Therefore (4.3) implies A ≈ f (A) ⊗k A f (A) . Counting ranks we see that
A f (A) = k. 

Corollary 4.5. The homomorphism

Pick (k) −→ Pick (A),

induced by L 7→ A ⊗k L, is an isomorphism.

Proof. This follows (see (4) of (4.1)) from the fact that A⊗k : k −
mod → Ae − mod is an equivalence which converts ⊗k into ⊗A ; the
latter is just the identity

(A ⊗k M) ⊗A (A ⊗k N) ≈ (A ⊗A A) ⊗k (M ⊗k N) ≈ A ⊗k (M ⊗k N).

Corollary 4.6 (Rosenberg-Zelinsky). If A/k is an azumaya algebra,

then there is an exact sequence
ϕA
0 → InAut(A) → Autk−alg (A) −−→ Pic(k),
 
where imϕA = (L)|A ⊗k L ≈ A as a left A-module

Proof. This follows immediately from (4.5) and proposition 7.3 of chap-
ter 2 
 4. Azumaya algebras 85

Corollary 4.7. If A/k is an azumaya algebra of rank r as a projective

109 k-module, then Autk−alg (A)/InAut(A) is an abelian group of exponent rd
for some d > 0.

Proof. Let A ⊗k L ≈ A as a left A-module, hence as a k-module. The re-

mark following proposition 6.1 of chapter 1 provides us with a k-module
d d d
Q such that Q ⊗k A ≈ k(r ) for some d > 0. So L(r ) ≈ k(r ) . Taking rd th
d
exterior powers we have L⊗r ≈ k. By virtue of (4.6), the corollary is
now proved. 

Corollary 4.8 (Skolem-Noether). If Pic (k) = 0, then all automorphisms

of an azumaya k-algebra are inner.

Corollary 4.9. If A is a k-algebra, finitely generated as a module over

its centre C, then A/k is separable ⇔ A/C and C/k are separable.

Proof. In view of (2.19) it is enough to remark that, if A/k is separa-

ble, then A is faithfully C-projective. This follows from (1) ⇔ (3) of
theorem 4.1. 

Proposition and Definition 4.10. Call two azumaya k-algebras A1 and

A2 similar, if they satisfy the following conditions, which are equivalent:

(1) A1 ⊗k Ao2 ≈ Endk (P) for some faithfully projective k-module P.

(2) A1 ⊗k Endk (P1 ) ≈ A2 ⊗k Endk (P2 ) for some faithfully projective k

modules P1 and P2 .

(3) A1 − mod and A2 − mod are equivalent k-categories.

(4) A1 ≈ EndA2 (P) for some faithfully projective right A2 -module P.

Proof. (1) ⇒ (2). A2 ⊗k Endk (P) ≈ A1 ⊗k A2 ⊗k Ao2 ≈ A1 ⊗k Endk (A2 ). 110

(2) ⇒ (3). Since Ai ⊗k Endk (Pi ) ≈ EndAi (Ai ⊗k (Pi ) (see (2.8) (a)),
and since Ai ⊗k Pi is a faithfully projective Ai -module, it follows from
theorem 4.4 and proposition 5.6 of chapter 2, that Ai − mod is k-
equivalent to (Ai ⊗k Endk (Pi ) − mod , i = 1, 2.
(3) ⇒ (4) follows proposition 4.1 and theorem 4.4 of chapter 2.
 86 3. The Brauer group of a commutative ring

(4) ⇒ (1). A1 ⊗k A02 ≈ EndA2 (P) ⊗k Ao2 ≈ EndA1 ⊗k A0 (P ⊗k A02 ).

2
Now P ⊗k A02 is faithfully projective A2 ⊗ A02 -module, so P ⊗k A02 ≈
A2 ⊗k Q, where Q = (P ⊗k Ao2 )A2 is faithfully projective k-module. Hence
A1 ⊗k A02 ≈ EndAe2 (A2 ⊗k (Q), since A2 ⊗k : k − mod → Ae2 − mod is
an equivalence. 

It follows from this proposition that similarly is an equivalence re-

lation between azumaya algebras, and that ⊗k induces a structure of
abelian group on the set of similarity classes of azumaya k-algebras.
We shall call this group the Brauer group of k, and hence denote it by
Br(k). The identity element in Br(k) is the class of k, and the inverse of
the class of an azumaya k-algebra A is the class of A0 .
If K is a commutative k-algebra, then A 7→ K ⊗k A induces a ho-
momorphism Br(k) → Br(K), by virtue of (2.9)(a), and this makes Br a
functor from commutative rings to abelian groups.

5 Splitting rings
111 If P is a projective k-module, denote its rank by [P : k]. This is a
function spec (k) → Z. If L is a commutative k-algebra, denote by ϕL
the natural map spec (L) →spec (k). Then, for a projective k-module
P, we have ϕL o[P : k] = [P ⊗k L : L]. If A/k is an Azumaya algebra,
denote its class in Br(k) by (A).

Theorem 5.1. Let A/k be an azumaya algebra.

(a) If L ⊂ A is a maximal commutative subalgebra, then A ⊗k L ≈

EndL (A) as L-algebras, viewing A as a right L-module. Hence if A
is L-projective, then (A) ∈ ker(BR(k) → Br(L)), and ϕL o[A : k] =
[A : L]2 . If also L is k-projective, then ϕL o[L : k] = [A : L]. If L/k
is separable, A is automatically L-projective.

(b) Suppose L is a commutative faithfully k-projective k-algebra, and

suppose (A) ∈ ker(Br(k) → Br(L)). Then there is an algebra B,
similar to A, which contains L as a maximal commutative subalge-
bra. If Endk (L) is projective as a right L-module, then so is B.
6. The exact sequence 87

Proof. (a) EndL (A) is the centralizer B in Endk (A) = A ⊗k A0 of l ⊗ L ⊂

A ⊗ A0 . Since A = A ⊗ l ⊂ A ⊗ L ⊂ B, it follows from (4.3), that
B = A ⊗k BA. Now BA ⊂ (A ⊗k A0 )A = l ⊗ A0 , and BA commutes
with l ⊗ L, a maximal commutative subalgebra of l ⊗ A0 . Hence
BA = l ⊗ L, so B = A ⊗k L, as claimed.
If A is L-projective, then ϕL ◦ [A : k] = [A ⊗k L : L] = [EndL (A) : 112
L] = [A : L]2 . If, further, L is k-projective, ϕL ◦ [A : k] = [A ⊗k L :
L] = [A⊗L (L⊗k L) : L] = [A : L].[L⊗k L : L] = [A : L]. (ϕL ◦[L : k]),
so that [A : L] = ϕL ◦ [L : k].
Suppose E = (0 → J → Le → L → 0) splits. Then A ⊗L E also
splits. So A = A ⊗L L is projective over A ⊗L (L ⊗k L0 ) = A ⊗k L.
But A ⊗k L/L is an azumaya algebra, so A ⊗k L is projective over L.
Hence A is L-projective.

(b) If (A)ǫ ker(Br(k) → Br(L)), then A0 ⊗k L ≈ EndL (P) for some faith-
fully projective L-module P. Using the isomorphism to identity, we
o
have Ao ⊗k L = EndL (P) ⊂ Endk (P) = D. Let B = DA . Then
(4.3) implies that D = A0 ⊗k B. Since L is faithfully k-projective,
so also is P. So D/k is a trivial azumaya algebra and (B) = (A) in
o
Br(k). Clearly L = l ⊗ L ⊂ B. Since B = DA , BL = DAo ∩ DL .
Further, DL = Endk (P)L = EndL (P), so BL = EndL (P)Ao . Since
EndL (P) = Ao ⊗k L, we have EndL (P)A0 = centre EndL (P) = L.
Now D = Ao ⊗k B with L ⊂ B, so D is locally (with respect to k) a
direct sum of copies of B as an L-module. Hence B is L-projective
as soon as we show that D = Endk (P) is. Again we localize (with
respect to k), whereupon P becomes L-free and I becomes k-free.
Then We can write P = P0 ⊗k L, P0 a free k-module, and we have
D = Endk (P) = Endk (P0 ) ⊗k Endk (L). By hypothesis, Endk (L) is
right L-projective, so D is L-projective.


6 The exact sequence

We now make out of the azumaya k-algebras, a category with product, 113
 88 3. The Brauer group of a commutative ring

in the sense of chapter 1. We write

Az = Az(k)
= =

for the category whose objects are azumaya k-algebras, whose mor-
phisms are algebra isomorphisms, and with product ⊗k .
Recall that the category FP = FP(k) (see §6 of chapter 1) of faith-
fully projective k-modules also has ⊗k as product. More over, (2.12)
says that functor
End = Endk : FP → Az
= =
preserves products. (If f : P → Q in FP, then f is an isomorphism,
=
and End ( f ) : End(P) → End(Q) is defined by End ( f )(e) = f e f −1 .)
Theorem 4.1 (6) asserts that End is a cofinal functor, so we have the five
term exact sequence from theorem 4.6 of chapter 1:
K1 End Ko End
K1 FP −−−−−→ K1 Az → Ko Φ End → Kc FP −−−−−→ Ko Az. (6.1)
= = = =

It follows immediately from (4.10)(2), that

coker (K0 End) = Br(k). (6.2)

Consider the composite functor

I End
Pic ֒→ FP −−−→ Az, (6.3)

which sends every object of Pic to the algebra k ∈ Az. Hence the com-
114 posites (Ki End) ◦(Ki I) = 0 for i = 0, 1. We will now construct a con-
necting homomorphism K1 Az → K0 Pic = Pic(k) and use it to identity
(6.1) with the sequence we will thus obtain from (6.3).
Recall that K1 Az is derived from the category ΩAz, whose objects
are pairs (A, α), A ∈ Az, α ∈ Autk−alg (A). Let 1 Aα denote the invertible
two-sided A-module constructed in lemma 7.2 of chapter 2. We have
A
1 Aα ≈ A ⊗k Lα , where Lα = ( 1 Aα ) , according to theorem (4.1)(4). In
this way we have a map

obj ΩAz → obj Pic,

6. The exact sequence 89

given by (A, α) 7→ Lα = (1 Aα )A . If f : (A, α) → (B, β) is an isomor-

phism in ΩAz, then f induces (by restriction) an isomorphism Lα → Lβ ,
thus extending the map above to a functor. If α, β ∈ Autk−alg (A), then
we have from (II, (7.2)(2)) a natural isomorphism

1 Aαβ ≈ 1 Aα ⊗ A1 Aβ .

Since A⊗k : k − mod → A − mod converts ⊗k into ⊗A , it follows

that Ldβ ≈ Lα ⊗k Lβ . Finally, given (A, α) and (B, β), we have

Lα⊗β = (1 (A ⊗ B)α⊗β )A⊗k B

= ( 1 Aα ⊗k 1 Bβ)A⊗k B
= (1 Aα )A ⊗k (1 Bβ )B
= L α ⊗k L β .

We have thus proved: 115

Proposition 6.4. (A, α) 7→ Lα = (1 Aα )A defines a functor

J : ΩAz → Pic

of categories with product, and it satisfies, for α, β ∈ Autk−log (A), A ∈

Az,
Lαβ ≈ Lα ⊗k Lβ .
Now we define a functor.

T : Pic → Φ End

by setting T (L) = (L, αL , k), where αL is the unique k-algebra isomor-

phism Endk (L) ≈ k → Endk (k) ≈ k. Clearly T preserves products.
Suppose (P, α, Q) ∈ Φ End. Thus α : A = End(P) → B = End(Q)
is an algebra isomorphism. α permits us to view left B-modules as left
A-modules. Since P⊗k : k − mod A − mod is an equivalence, the
inverse being HomA (P, ) : A − mod → k − mod , we can apply this
functor to the B−, hence A-module, Q and obtain a k-module

L = HomA (P, Q)
 90 3. The Brauer group of a commutative ring

such that Q ≈ P ⊗k L as a left A-module. It follows that Lα ∈ Pic.

If ( f, g): (P, α, Q) → (P′ , α′ , Q′ ) is in Φ End, then the map HomEnd(P)
(P, Q) → HomEnd(P′ ) (P′ , Q′ ), given by e 7→ ge f −1 , is in Pic, thus giving 116
us a functor
S : Φ End → Pic.

Moreover, S preserves products because

HomEnd(P⊗k P′ ) (P ⊗k P′ , Q ⊗k Q′ )
≈ HomEnd(P)⊗k End(P′ ) (P ⊗k P′ , Q ⊗k Q′ ) by (2.12)
′
≈ HomEnd(P) (P, Q) ⊗k HomEnd(P′ ) (P , Q). by (2.5)(i).

If L ∈ Pic, then S T L = S (L, αL , L) = Homk (k, L) ≈ L.

We have now proved all but the last statement of

Proposition 6.5. There are product-preserving functors

T /
Pic o Φ End
S

defined by T L = (k, αL , L) and S (P, α, Q) = HomEnd(P) (P, Q), such that

S T ≈ IdPic . If (P, α, R) ∈ Φ End, then S (P, βα, R) ≈ S (Q, β, R) ⊗k
S (P, α, Q).

Proof. The prove the last statement we note that composition defines a
homomorphism

HomEnd(Q) (Q, R) ⊗k HomEnd(P) (P, Q) → HomEnd(P) (P, R).

117 The module above are projective and finitely generated over k. There-
fore it is enough to check that the map is an isomorphism over residue
class fields k/M . 

Proposition 6.6. S and T define inverse isomorphisms

Pic(k) = K0 Pic ⇆ K0 ΦEnd.

6. The exact sequence 91

Proof. S T ≈ IdPic so it suffices to show that K0 S is injective. If K0 S

(P, α, Q) = k, then Q ≈ P ⊗k k = P as a left End(P)-module. Let
f : Q → P be such an isomorphism. This means that for all e ∈ End(P)
and q ∈ Q, f (α(e)q) = e f (q), that is, that
α / End(Q)
End(P)

End(1P ) End( f )

 
End(P) / End(P)
1End(P)

commutes. Thus (1P , f ) : (P, α, Q) → (P, 1End(p) , P) in Φ End, so

(P, α, Q)Φ End = 0 in K0 Φ End. 

Theorem 6.7. The sequence of functors

ΩI Ω End J I End
ΩPic −−→ ΩFP −−−−→ ΩAz −
→ Pic →
− FP −−−→ Az

of categories with product defines an exact sequence which is the top

row of the following commutative diagram:

U(k) / K1 FP / K1 Az / Pic(k) / K0 FP / K0 Az / Br(k) /0

O
K0 S K0 T

K1 FP / K1 Az / K0 Φ End / K0 FP / K0 Az

The bottom row is the exact sequence of theorem 4.6 of chapter 1

for the functor End : FP → Az · K0 S and K0 T are the isomorphisms of
proposition 6.6.

Proof. We first check commutativity: If L ∈ Pic, then K0 I(L)Pic = 118

(L)FP , while K0 T (L)Pic = (L, αL , L)Φ End is sent to (L)FP (k)−1
FP = (L)FP .


Now that the diagram commutes, exactness of the top row follows
from that of the bottom row, wherever the isomorphisms imply this. At
 92 3. The Brauer group of a commutative ring

K0 Az and Br(k) exactness has already been remarked in (6.2) above. At

K1 FP the composite is clearly trivial. So it remains only to show that
ker(K1 End) ⊂ Im K1 I.
Now we know that the free modules kn are cofinal in FP, and hence
(by cofinality of End) that the matrix algebras Mn (k) = End(kn ) are
cofinal in Az. Hence we may use them to compute K1 as a direct limit
(theorem 3.1 of chapter 1).
Write GLn (FP) = GL(n, k) = Autk (kn ), and GLn (Az) = Autk−alg
(Mn (k)). We have the “inner automorphism homomorphism”

fn : GLn (FP → GLn (Az)

with ker fn = centre GLn (FP = GL1 (FP) = U(k)), and im fn ≈ PGL
(n, k).
Tensoring with an identity automorphism defines maps

GLn (FP) → GLnm (FP)

119 and
GLn (Az) → GLnm (Az)
making ( fn )n∈N a map of directed systems. Writing

GL(FP) = lim GLn (FP),

→
GL(Az) = lim GLn (Az),
→

and
f = lim fn : GL(FP) → GL(Az),
→
we see that K1 End is just the ablianization of f . In §6 of chapter 1 we
computed

K1 FP = GL(FP)/[GL(FP), GL(FP)]
= (Q ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P).

Moreover K1 I is induced by the inclusion U(k) = GL1 (FP) → GL(FP).

K1 I is the map K1 Pic = U(k)−Z⊗Z U(k) → Q⊗Z U(k) ⊂ K1 FP, so coker
6. The exact sequence 93

(K1 I) = (Q/Z⊗Z U(k))⊗(Q⊗Z S K1 P = PGL(k)/[PGL(k), PGL(k)]. Thus

exactness at K1 FP means that: the inclusion PGL(k) ⊂ GL(Az) induces
a monomorphism

PGL(k)/[PGL(k), PGL(k)] → GL(Az)/[GL(Az), GL(Az)].

Suppose α, β ∈ GLn (Az) = Autk−alg (Mn (k)). Let τ : kn ⊗ kn →

kn → kn be the transposition. Write E(τ) for the corresponding inner 120
automorphism of MN 2 (K). Then E(τ)(α ⊗ 1Mn (k) )E(τ)−1 = 1Mn (k) ⊗ α
commutes with β ⊗ 1Mn (k) . Now τ is just a permutation of the basis
of kn ⊗k kn , so it is a product of elementary matrices, provided it has
1
determinant +1 (which happens when n(n − 1) is even). For exam-
2
ple, if we restrict our attention to values of n divisible by 4, then τ
is a product of elementary matrices, hence lies in [GLn2 , GLn2 (k)], so
E(τ) ∈ [PGLn2 (k), PGLn2 (k)]. It follows that, for n divisible by 4, the
image of GLn (Az) in GLn2 (Az)/[PGLn2 (k), PGLn2 (k)] is abelian. Note
that PGLn2 (k) = In Aut (Mn (k)) is normal in GLn (Az), hence so also is
[PGL2n (k), PGL2n (k)], so the factor group above is defined. Finally, since
the n divisible by 4 are cofinal N we can pass to the limit to obtain

[GL(Az), GL(Az)] ⊂ [PGL(k), PGL(k)],

are required. Q.E.D.

Proposition 6.8. In the exact sequence of theorem 6.7,

ker(U(k) → K1 FP) = the torsion subgroup of U(k),

ker(Pic(k) → K0 FP) = the torsion subgroup of Pic(k),

Hence there is an exact sequence 121

 the torsion 
0 → (Q/Z ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P) → K1 Az → subgroup of → 0.
Pic(k)

This sequence splits (not naturally) as sequence of abelian groups.

94 3. The Brauer group of a commutative ring

Proof. For K1 I : U(k) = K1 Pic → K1 FP we have from chapter 1, §7,

ker K1 I = the torsion subgroup of U(k),

and
coker K1 I = (Q/Z ⊗Z U(k)) ⊕ (Q ⊗Z S K1 P).
From the same source we have
ker K0 I = the torsion subgroup of Pic(k).
The last assertions now follow from the exact sequence plus the fact
that the left hand term is divisible, hence an injective Z-module. 
Chapter 4

The Brauer-Wall group of

graded Azumaya algebras

This chapter contains only a summary of results, without proofs. They 122
are included because of their relevance to the following chapter, on Clif-
ford algebras.

1 Graded rings and modules

All graded objects here are graded by Z/2Z. A ring A = A0 ⊕A1 is graded
if Ai A j ⊂ Ai+ j (i, j ∈ Z /2Z ), and an A-module M = M0 ⊕ M1 is graded
if Ai M j ⊂ Mi+ j . (We always assume modules to be left modules unless
otherwise specified.) If S is a subset of a graded object, hS will denote
the homogeneous elements of S , and ∂x = degree of x, for x ∈ hS .
If A is a graded ring, then |A| will denote the underlying ungraded
ring. If A is ungraded, then (A) denotes the graded ring with A concen-
trated in degree zero. An A-module is graded or not according as A is or
is not. If M is an A-module (A graded) we write |M| for the underlying
|A|-module. If A is not graded, we write (M) for the (A)-module with M
concentrated in degree zero.
Let A be a graded ring. For A-modules M and N,

HOMA (M, N)

95
 96 4. The Brauer-Wall group of graded Azumaya algebras

is the graded group of additive maps from M to N defined by: f ∈

hHOMA (M, N) ⇔ (i) f is homogeneous of degree ∂ f (i.e. f (Mi ) ⊂
Ni+∂ f ); and (ii) fax = (−1)∂ f ∂a a f x(a ∈ hA, x ∈ M).
123 The degree zero term of HOMA (M, N) is denoted by HomA (M, N).
Let A′ denote the graded group A with new multiplication

a · b = (−1)∂a∂b ab (a, b, ∈ hA).

If M is an A-module let M ′ denote the A′ -module with M as the under-

lying graded group and operators defined as

a · x = (−1)∂a∂x ax (a ∈ hA, x ∈ hM).

Then it is straightforward to verify that

HOMA (M, N) = Hom|A′ | (|M ′ |, |N ′ |),

an equality of graded groups.

A-mod refers to the category with A-modules as objects and homo-
morphisms of degree zero (i.e. HomA (, )) as morphisms.

Lemma 1.1. The following conditions on an A−module P are equiva-

lent:

(1) HomA (P, ) is exact on A − mod .

(2) Hom|A| (|P|, ) is exact on |A| − mod .

(3) HomA (P, ) is exact on A − mod .

(4) P is a direct summand of A(I) ⊕ (τA)(J) for some I and J, where τA

is the A-module A with grading shifted by one.

This lemma tells us that the statement “P is A-projective” is unam-

biguous.
124 If S ⊂ A (graded), we define the centralizer of S in A to be graded
subgroup C such that c ∈ hC ⇔ cs = (−1)∂c∂s sc for all s ∈ hS . It is
easy to see that C is actually a subring of A. We say that two subrings
of A commute, if each lies in the centralizer of the other. If B1 and B2
2. Separable algebras 97

are subrings generated by sets S 1 and S 2 , respectively, of homogeneous

elements, then B1 and B2 commute ⇔ S 1 and S 2 commute. We write
AA = CENTRE (A) = centralizer of A in A.
The degree zero term will be denote centre A. One must not confuse
centre A, centre |A|, and CENTRE (A). They are all distinct in general.
Let k be a commutative ring which is graded, but concentrated in de-
gree zero. Even though k and |k| are not essentially different, k − mod
and |k|-mod are. A k-algebra is a graded ring A and a homomorphism
k → A of graded rings such that the image k1 lies in AA , and hence in
centre A.
If A1 and A2 are k-algebras and if M i is an Ai -module i = 1, 2 we
define the k-module M 1 ⊗ M 2 by

(M 1 ⊗ M 2 )n = (M01 ⊗ Mn2 ) ⊕ (M11 ⊗ Mn+1

2
).

We define an action of A1 ⊗ A2 on M 1 ⊗ M 2 by

(a1 ⊗ a2 )(x1 ⊗ x2 ) = (−1)∂a2 ∂x1 a1 x1 ⊗ a2 x2

for ai ∈ hAi , xi ∈ hM i , i = 1, 2. This makes A1 ⊗ A2 a k-algebra (for

M i = Ai ) and M 1 ⊗ M 2 an A1 ⊗ A2 -module. The subalgebras A1 ⊗ 1
and 1 ⊗ A2 commute, and the pair of homomorphisms Ai → A1 ⊗ A2 is 125
universal for pairs of homomorphisms f i : Ai → B of k-algebras such
that im f 1 and im f 2 commute. In practice it is useful to observe that: if
Ai is generated by homogeneous elements S i , and if f 1 (S 1 ) and f 2 (S 2 )
commute, then f 1 (A1 ) and f 2 (A2 ) commute.

2 Separable algebras
In this section also k denotes a commutative ring concentrated in degree
zero. If A is a k-algebra, then A0 denotes the opposite algebra, and we
write
A∗ = (A′ )0 = (A0 )′
for the algebra with graded group A and multiplication

a × b = (−1)∂a∂bba (a, b ∈ hA).

 98 4. The Brauer-Wall group of graded Azumaya algebras

A right A-module M will be considered a left A∗ -module by setting

ax = (−1)∂a∂x xa (a ∈ hA∗ , x ∈ hM).
If M is a left A-, right B-module such that (ax)b = a(xb) and tx = xt for
all a ∈ A, x ∈ M, b ∈ B, t ∈ k, then we view M as an A ⊗k B∗-module by
(a ⊗ b)x = (−1)∂b∂x axb (a ∈ hA, b ∈ hB, x ∈ hM).
In particular, two-sided A-modules will be identified with modules over
A e = A ⊗k A ∗
126 We have an exact sequence
0 → J → Ae → A → 0
(a ⊗ b) 7→ ab
of Ae -modules. We call A a separable k-algebra if A is Ae -projective.
This means that the functor
Ae − mod → k − mod
M 7→ M A = HOMAe (A, M)
is exact.
The stability of separability and CENTRES under base change and
tensor products all hold essentially as in the ungraded case. In particular
ENDk (P) = HOMk (P, P) is separable with CENTRE k/annP, for P a
finitely generated projective k-module. Moreover :
Proposition 2.1. Let A be finitely generated as a k−module and suppose
either that k is noetherian or that A is k−projective. Then A is separable
⇔ (A/M A)/(k/M ) is separable for all maximal ideals M of k.
Suppose now that k is a field. If a ∈ k. write k < a >= k[X]/(X 2 −a),
with grading k.1 ⊕ k · x, x2 = a. It can be shown that if char k , 2 and if
a , 0, then k < a > is separable with CENTRE k. Moreover
!
a, b
k < a > ⊗k k < b >≈ ,
k
the k-algebra with generators α, β of degree one defined by relations:
α2 = a, β2 = b, αβ = −βα.
 3. The group of quadratic extensions 99

127 Theorem 2.2. Let A be a finite dimensional k-algebra, k a field. Then

the following conditions are equivalent:

(1) A/k is separable.

i
(2) A = ΠAi , where Ai is a simple (graded) k-algebra and AiA is a
separable field extension of k, concentrated in degree zero.

(3) For some algebraically closed field L ⊃ k, L ⊗k A is a product of

algebras of the types

(i) ENDL (P), P a finite dimensional L-module, and

(ii) L < 1 > ⊗L ENDL (P), P a finite dimensional L-module with P1 = 0.

If char k = 2, then type (ii) does not occur.

Corollary 2.3. Let k be any commutative ring and A a k-algebra finitely

generated as a k-module. Suppose either k is noetherian or that A is
k-projective. Then if A/k is separable, |A|, |A0 |, |AA |, and |AA0 | are sepa-
rable |k|-algebras.

3 The group of quadratic extensions

A quadratic extension of k is a separable k-algebra L which is a finitely
generated projective k-module of rank 2. By localizing and extending 1
to a k-basis of L we see that |L| is commutative.

Proposition 3.1. If L/k is a quadratic extension, then there is a unique

k-algebra automorphism σ = σ(L) of L such that Lσ = k.

Proposition 3.2. If L1 and L2 are quadratic extensions of k, then so also 128

is
L1 ∗ L2 = (L1 ⊗k L2 )σ1 ⊗σ2 ,
where σi = σ(Li ). Further, ∗ induces on the isomorphism classes of
quadratic extensions the structure of an abelian group,

Q2 (k).
 100 4. The Brauer-Wall group of graded Azumaya algebras

If we deal with |k|-algebras, then we obtain a similar group,

Q(k)

of ungraded quadratic extensions. Each of these can be viewed as a

graded quadratic extension of k, concentrated in degree zero, and this
defines an exact sequence

0 → Q(k) → Q2 (k) → T,

where T = continuous functions spec (k) → Z/2Z, and right hand map
is induced by L 7→ [L1 : k] = the rank of the degree one term, L1 , of L.
In particular, if Spec (k) is connected, we have

0 → Q(k) → Q2 (k) → Z/2Z,

and the right hand map is surjective ⇔ 2 ∈ U(k). In this case L =

k < u > is a quadratic extension for u ∈ U(k). L = k · 1 ⊕ k · x
with x2 = u, and σ(x) = −x for σ = σ(L). If u1 , u2 ∈ U(k), then
u1 , u2
k < u1 > ⊗k k < u2 >= has k-basis 1, x1 , x2 , x3 = x1 x2 = −x2 x1
k
with x21 = u1 , x22 = u2 , x23 = −u1 u2 . If σi = σ(k < ui >), then σ1 ⊗ σ2
sends x1 7→ −x1 , x2 7→ −x2 , x3 7→ x3 .
129 It follows that

k < u1 > ∗k < u2 >= k[−u1 u2 ],

where k[u] = k[X]/(X 2 − u), concentrated in degree zero.

Proposition 3.3. Suppose 2 ∈ U(k). Then

(a) the sequence 0 → Q(k) → Q2 (k) → Z/2Z → 0 is exact; and

(b) there is an exact sequence

2 2
U(k) →
− U(k) → Q(k) → Pic(k) →
− Pic(k),

where the map in the middle are defined by u 7→ k[u] and L 7→

(L/k), respectively.
4. Azumaya algebras 101

Next suppose that char k = 2.

Proposition 3.4. Suppose k is a commutative ring of characteristic 2.

Then if a ∈ k, k[a] = k[X]/(X 2 + X + a) is a quadratic extension, con-
centrated in degree zero. k[a] = k · 1 + k · x with x2 + x + a = 0, and
σ(x) = x + 1. Q2 (k) ≈ Q[k] and there is an exact sequence
℘
k−
→ k → Q(k) → 0,

where ℘(a) = a2 + a, and a 7→ k[a] induces k → Q[k].

4 Azumaya algebras
k is a commutative ring concentrated in degree zero.

Theorem and Definition 4.1. A is an azumaya k- algebra if it satisfies

the following conditions, which are equivalent:

(1) A is a finitely generated k- module, AA = k, and A/k is separable.

(2) AA = k and |A| is a generator as an |Ae |-module.

(3) A is a faithfully projective k-module and Ae → ENDk (A) is an iso- 130

morphism.

(4) The functors

(M  / M A)
/
Ae − mod o k − mod

(A ⊗k N) o N)
are inverse equivalences of categories.

(5) For all maximal ideals M ⊂ k, A/M A is simple, and CENTRE

(A/M A) = k/M .

(6) There exists a k-algebra B and a faithfully projective k-module P

such that A ⊗k B ≈ ENDk (P).
 102 4. The Brauer-Wall group of graded Azumaya algebras

Corollary 4.2. Let A and B be k-algebras with A azumaya. Then b 7→

A ⊗ b is a bijection from two-sided ideals of B to those of A ⊗k B.
Corollary 4.3. If A ⊂ B are k-algebras with A azumaya, then B ≈
A ⊗k B A .
Call two azumaya algebras A and B similar if A ⊗k B∗ ≈ ENDK (P)
for some faithfully projective module P. With multiplication induced by
⊗k , the similarity classes form a group, denoted
Br2 (k),
and called the Brauer-Wall group of k.
Theorem 4.4. If A is an azumaya algebra, define L(A) = AA0 . Then
L(A) is a quadratic extension of k, and L(A ⊗k B) ≈ L(A) ∗ L(B). A 7→
L(A) induces an exact sequence
0 → Br(k) → Br2 (k) → Q2 (k) → 0.

5 Automorphisms
131 If A is a k-algebra a = a0 + a1 ∈ A write σ(a) = a′ = a0 − a1 . Let U(A)
denote the group of units in A, and hU(A) the subgroup of homogeneous
units. If u ∈ hU(A), we define the inner automorphism, αu , by
αu (a) = uσ∂u (a)u−1 .
This is clearly an algebra automorphism of A, and a simple calculation
shows that αuv = αu αv . Thus we have a homomorphism
hU(A) → Autk−alg (A); u 7→ αu .
The kernel consists of those u such that uσ∂u (a) = au for all a ∈ A.
Taking a homogeneous, σ(a) = (−1)∂a a, so the condition becomes
(−1)∂u∂a ua = au, for all a ∈ hA, i.e. u ∈ CENTRE(A) = AA .
Thus we have an exact sequence
1 → hU(AA ) → hU(A) → Autk−alg (A). (5.1)
Now just as in the ungraded case one can prove:
5. Automorphisms 103

Theorem 5.2. Let A be an azumaya k-algebra. If α ∈ Autk−alg (A),

let 1A α denote the Ae -module A with action a · x · b = axα(b) for
a, x, b ∈ A. Then Lα = (1 Aα )A is an invertible k-module, and α 7→ Lα in-
duces a homomorphism g making the sequence 1 → U(k) → U(A0 ) →
g
Autk−alg (A) →
− Pic(k) exact. im g = {(L)|A ⊗k L ≈ A as left A-modules}

Here U(k) = U(AA ) ⊂ U(A0 ) ⊂ hU(A), and the left hand portion of 132
the sequence is induced by (5.1) above. Pic (k) is the group of “graded
invertible k-modules”. If u is a unit of degree one in A then Lαu is just k,
but concentrated in degree one. This explains why we have U(A0 ), and
not hU(A), in the exact sequence.
This theorem will be applied, in Chapter 5, §4, to the study of or-
thogonal groups. We conclude with the following corollary:

Corollary 5.3. Autk−alg (A)/ (inner automorphisms) is a group of expo-

nent rd for some d > 0, where r = [A : k].
Chapter 5

The structure of the Clifford

Functor

In this chapter we introduce the category, Quad(k) of quadratic forms on 133

projective k-modules, and the hyperbolic functor, H : P → Quad. This
satisfies the conditions of chapter 1 to yield an exact sequence,

K1 P → K1 Quad → K0 ΦH → K0 P → K0 Quad → Witt(k) → 0,

where Witt (k) = coker (K0 H) is the classical “Witt ring” over k.
The Clifford algebra is constructed as a functor from Quad to k-
algebras, graded mod 2, and the main structure theorem (§3) asserts
that the Clifford algebra are (graded) azumaya algebras, in the sense of
Chapter 4 and that the diagram

P H / Quad

∧ C1
 
FP / Az
2 END 2

commutes up to natural isomorphism. Here ∧ denotes exterior algebra,

graded mod 2 by even and odd degrees. The proof is achieved by a
simple adaptation of arguments in Bourbaki [2].

105
 106 5. The structure of the Clifford Functor

This commutative diagram leads to a map of exact sequences

K1 P / K1 Quad / K0 ΦH / K0 P / K0 Quad / Witt(k) /0

= =

     
K1 FP K1 Az / K0 ΦEND / K0 FP / K0 Az / Br2(k) /0
2 2 2 2

134 This map of exact sequences is the “generalized Hasse-Wall invari-

ant.”
In §4 we indicate briefly what the construction of the spinor norm
looks like in this generality.

1 Bilinear modules
We shall consider modules over a fixed commutative ring k, and we shall
abbreviate,

⊗ = ⊗k , Hom = Homk , M ∗ = Hom(M, k).

Bil (P × Q) denotes the module of k-bilinear maps, P × Q → k.

Let P be a k-module. If x ∈ P and y ∈ P∗ write

hy, xiP = y(x).

If f : P → Q then f ∗ : Q∗ → P∗ is defined by

h f ∗ y, xiQ = hy, f XiP (x ∈ P, y ∈ Q∗ ).

There are natural isomorphisms

s d
Hom(P, Q∗ ) ← − Hom(Q, P∗ )
− Bil(P × Q) →
sB ←− B −→ dB

defined by

hS B x, yiQ = B(x, y) = hdB y, xiP (x ∈ P, y ∈ Q).

1. Bilinear modules 107

Applying this to the natural pairing

h , iP : P∗ × P → k,

we obtain the natural homomorphism

dP : P → P∗∗ ; hdP x, yiP∗ = hy, xiP .

We call P reflexive if dP is an isomorphism, and we will then often 135

identify P and P∗∗ via dP .
Suppose Bǫ Bil(P × Q), x ∈ P, and yǫQ. Then

hdB y, xiP = hsB x, yiQ = hdQ y, sB xiQ∗ = hs∗B dQ y, xiP .

From this and the dual calculation we conclude:

dB = s∗B dQ and sB = d∗B dP . (1.1)

We call B non-singular if dB and sB are isomorphisms. In view of (1.1)

this implies P and Q are reflexive. Conversely, if P and Q are reflexive,
and if dB is an isomorphism, then (1.1) shows that sB is also.
A pair (P, B), B ∈ Bil(P × P), is called a bilinear module. f : P1 →
P2 is a morphism (P1 , B1 ) → (P2 , B2 ) if B2 ( f x, f y) = B1 (x, y) for x,
yǫP1 . We define

(P1 , B1 ) ⊥ (P2 , B2 ) = (P1 ⊕ P2 , B1 ⊥ B2 )

and
(P1 , B1 ) ⊗ (P2 , B2 ) = (P1 ⊗ P2 , B1 ⊗ B2 )
by (B1 ⊥ B2 )((x1 , x2 ), (y1 , y2 )) = B1 (x1 , y1 ) + B2 (x2 , y2 ), and (B1 ⊗
B2 )(x1 ⊗ x2 , y1 ⊗ y2 ) = B1 (x1 , y1 )B2 (x2 , y2 ). Identifying (P1 ⊕ P2 )∗ =
P∗1 ⊕ P∗2 we have dB1 ⊥B2 = dB1 ⊕ dB2 . Moreover, dB1 ⊗B2 is dB1 ⊗ dB2
followed by the natural map P∗1 ⊗ P∗2 → (P1 ⊗ P2 )∗ . The latter is an
isomorphism if one of the Pi is finitely generated and projective.
If (P, B) is a bilinear module we shall write B∗ (x, y) = B(y, x). If P 136
is reflexive and we identify P = P∗∗ then (1.1) shows that d(B∗ ) = sB =
(dB )∗ . We call (P, B) or B symmetric if B = B∗ . For any B, B + B∗ is
clearly symmetric.
 108 5. The structure of the Clifford Functor

If (P, B) is a symmetric bilinear module we have a notion of orthog-

onality. Specifically, if U is a subset of P, write

PU = {x ∈ P|B(x, y) = 0 ∀y ∈ U}.

When P is fixed by the context we will sometimes write

U ⊥ = PU .

The following properties are trivial to verify:

U ⊥ is a submodule of P.
U ⊂ V ⇒ V⊥ ⊂ U⊥
U ⊂ U ⊥⊥
U ⊥ = U.⊥⊥⊥

We say U and V are orthogonal if U ⊂ V ⊥ , and we call a submodule

U totally isotropic if U ⊂ U ⊥ , i.e. if B(x, y) = 0 for all x, yǫU. The
expression P = U ⊥ V denotes the fact that P is the direct sum of the
orthogonal submodules U and V.

Lemma 1.2. Let (P, B) be a non-singular symmetric bilinear module. If

U is a direct summand of P then U ⊥ is also a direct summand, and B
induces a non-singular pairing on U × (P/U ⊥ ).

137 Proof. Since 0 → U → P → P/U → 0 splits so does 0 → (P/U)∗ →

P∗ → U ∗ → 0. By hypothesis dB : P → P∗ is an isomorphism, so
U ⊥ = d−1 ∗
B (P/U) is a direct summand of P. Moreover the composite
dB
P −−→ P∗ → U ∗ is surjective, with kernel U ⊥ , so B induces an isomor-
phism (P/U ⊥ ) → U ∗ . Since U and (P/U ⊥ ) are reflexive this implies the
pairing on U × (P/U ⊥ ) is non-singular (see(1.1)). 

Lemma 1.3. Let f : (P1 , B1 ) → (P2 , B2 ) be a morphism of symmetric

bilinear modules, and suppose that (P1 , B1 ) is non-singular. Then f is a
monomorphism, and
P2 = f P1 ⊥ P(2f P1 )
1. Bilinear modules 109

Proof. If xǫ ker f then 0 = B2 ( f x, f y) = B1 (x, y) for all y ∈ P1 so

x = 0 because B1 is non-singular. Now use f to identify P1 ⊂ P2 and
B1 = B2 |P1 × P1 . Then P1 ∩ P2P1 = 0 because B1 is non-singular. If
x ∈ P2 define h : P1 → k by h(y) = B2 (x · y). Since B1 is non-singular
h(y) = B1 (x1 , y) for some x1 ǫP1 , and then we have x = x1 + (x − x1 ) with
x − x1 ∈ P2P1 . 

Lemma 1.4. Let (P, B) be a non-singular symmetric bilinear module

and suppose that U is a totally isotropic direct summand of P.
(a) We can write P = U ⊥ ⊕ V, and, for any such V, W = U ⊕ V is a
non-singular bilinear submodule of P. Hence P = W ⊥ PW .
(b) V ≈ U ∗ , so if U is finitely generated and projective then so is W,
and [W : k] = 2[U : k].
(c) If B = B0 + B∗0 and if B0 (x, x) = 0 for all xǫU then we can choose V
above so that B0 (x, x) = 0 for all xǫV also.
Proof. (a) According to Lemma 1.2, P = U ⊥ ⊕ V, and for any such 138
V, B induces a non-singular form on U × V. Thus B induces iso-
morphisms f : U → V ∗ and g : V → U ∗ . If B1 = B|W × W then
0 g
dB1 : U ⊕ V → U ∗ ⊕ V ∗ is represented by a matrix f dB , where
2
B2 = B|V × V. Evidently dB1 is an isomorphism. Lemma 1.3 now
implies P = W ⊥ PW .
(b) is clear
(c) Identifying U = U ∗∗ and V = V ∗∗ , the symmetry of B implies
f ∗ = g. Let B3 = B0 |V × V, where B = B0 + B∗0 (by hypothesis), and
set k = f −1 dB3 : V → U. Then for v ∈ V we have
B(v, hv) = h f hv, viV = h f f −1 dB3 v, viV = B3 (v, v) = B0 (v, v).

Let t : V → U ⊕ V by t(v) = v − h(v). Then if V1 = tV it is still
clearly true that P = U ⊥ ⊕ V1 (in fact, W = U ⊕ V1 ). We conclude the
proof by showing that B0 (v, v) = 0 for vǫV1 . Suppose vǫV. Then
B0 (tv, tv) = B0 (v − hv, v − hv) = B0 (v, v) + B0 (hv, hv)−
 110 5. The structure of the Clifford Functor

− B0 (v, hv) − B0 (hv, v).

Since hv ∈ U and B0 (x, x) = 0 for x ∈ U, by hypothesis, and since

B = B0 + B∗0 , we have B0 (tv, tv) = B0 (v, v) − B0 (v, hv) − B∗0 (v, hv) =
B0 (v, v) − B(v, hv). This vanishes according to the calculation above, so
lemma 1.4 is proved.
Let P be a module and B ∈ Bil(P × P). We define the function

q = qB : P → k; q(x) = B(x, x).

139 q has the following properties:

q(ax) = a2 x (a ∈ k, x ∈ P), (1.5)

If Bq (x, y) = q(x + y) − q(x) − q(y), then Bq ∈ Bil(P × P). Indeed,

direct calculation shows that Bq = B + B∗.
Lemma 1.6. Suppose P is finitely generated and projective, and that q:
P → k satisfies (1.5). Then there is a B ∈ Bil(P × P) such that q = qB .
In particular, Bq = B + B∗.
P P 2
Proof. If P is free with basis (ei )1≤i≤n then q( ai ei ) = ai q(ei ) +
P i i
ai a j Bq (ei , e j ). Set bii = q(ei ), bi j = Bq (ei , e j ) for i < j, and bi j =
i< j P P P P
0 for i > j. Then q( ai ei ) = ai a j bi j = B( ai ei , ai ei ), where
P P P i i, j i i
B( ai ei , ci ei ) = ai c j bi j .
i i i, j
In the general case choose P′ so that F = P ⊕ P′ is free and extend
q to q1 on F by q1 (x, x′ ) = q(x) for (x, x′ ) ∈ P ⊕ P′ . If q1 = qB1 then
q = qB where B = B1 P × P.
We define a quadratic form on a module P to be a function of the
form qB for some B ∈ Bil(P × P). B is then uniquely determined modulo
“alternating forms,” i.e. those B such that B(x, x) = 0 for all x ∈ P.
We shall call the pair (P, q) a quadratic module, and we call it non-
singular if Bq is non-singular. f : P1 → P2 is a morphism (P1 , q1 ) →
(P2 , q2 ) of quadratic modules if q2 ( f x) = q1 (x) for all xǫP1 . Evidently f
then induce a morphism (P1 , Bq1 ) → (P2 , Bq2 ) of the associated bilinear
modules.
 2. The hyperbolic functor 111

140 If f is an isomorphism we call f an isometry. If qi = qBi then we

define q1 ⊥ q2 = qB1 ⊥B2 on P1 ⊕ P2 , and q1 ⊗ q2 = qB1 ⊗B2 on P1 ⊗ P2 .
It is easily checked that these definition are unambiguous. 

2 The hyperbolic functor

Let P be a k-module and define

B0P ∈ Bil((P ⊕ P∗ ) × (P ⊕ P∗ )) by B0P((x1 , y1 ), (x2 , y2 )) = hy1 , x2 iP ,

and let qP = qBP be the induced quadratic form:

0

qP (x, y) = hy, xiP (x ∈ P, y ∈ P∗ ).

Let BP = B0P + (B0P)∗ be the associated bilinear form, BP = BqP . Then

BP((x1 , y1 ), (x2 , y2 )) = hy1 , x2 iP + hy2 , x1 iP .

If dP : P → P∗∗ is the natural map then it is easily checked that

dBP : P ⊕ P∗ → (P ⊕ P∗ )∗ = P∗ ⊕ P∗∗

is represented by the matrix

!
0 1 P∗
.
dP 0

Consequently, BP is non-singular if and only if P is reflexive.

 If, in this
case, we identify P = P∗∗ then the matrix above becomes 10P 10P∗ .
We will write 141
H(P) = (P ⊕ P∗ , qP )
and call this quadratic module the hyperbolic form on P.
Suppose f : P → Q is an isomorphism of k-modules. Define

H( f ) = f ⊕ ( f ∗ )−1 : H(P) → H(Q).

qQ (H( f )(x, y)) = qQ ( f x, ( f ∗ )−1 y) = h( f −1 )∗ y, f xiQ
 112 5. The structure of the Clifford Functor

= hy, f −1 f xiP = qP (x, y), so H( f ) is an isometry.

If we identify (P1 ⊕ P2 )∗ = P∗1 ⊕ P∗2 so that

h(y1 , y2 ), (x1 , x2 )iP1 ⊕P2 = hy1 , x1 iP1 + hy2 , x2 iP2

then the natural homomorphism

f : H(P1 ) ⊥ H(P2 ) → H(P1 ⊕ P2 ),

f ((x1 , y1 ), (x2 , y2 )) = ((x1 , x2 ), (y1 , y2 )). is an isometry.

Summarizing the above remarks, H is a product preserving functor
(in the sense of chapter 1) from (modules, isomorphisms, ⊕) to (quadra-
tic modules, isometries, ⊥). We now characterize non-singular hyper-
bolic forms.

Lemma 2.1. A non-singular quadratic module (P, q) is hyperbolic if

and only if P has a direct summand U such that q|U = 0 and U = U ⊥ .
In this case (P, q) ≈ H(U) (isometry).
142 Suppose P is finitely generated and projective. If U is a direct sum-
mand such that q|U = 0 and [P : k] ≤ 2[U : k] then (P, q) ≈ H(U).

Proof. If (P, q) ≈ H(U) = (U ⊕U ∗ , qU ) then the non-singularity of (P, q)

implies U is reflexive, and it is easy to check that U ⊂ U ⊕ U ∗ satisfies
qU |U = 0 and U = U ⊥ . 

Conversely, suppose given a direct summand U of P such that q|U =

0 and U = U ⊥ . Write q = qB0 , so that Bq = B0 + B∗0 . According to
Lemma 1.4 we can write P = U ⊥ ⊕ V = U ⊕ V and Bq induces a non-
singular pairing on U × V. Moreover we can arrange that B0 (v, v) = 0
for all v ∈ V, i.e. that q|V = 0. Let d : V → U ∗ be the isomorphism
induced by Bq ; hdv, uiU = Bq (v, u) for u ∈ U, v ∈ V.
Let
f = 1U ⊕ d : P = U ⊕ V → U ⊕ U ∗ .
This is an isomorphism, and we want to check that

qU ((u, dv)) = q(u, v) for uǫU, vǫV.qU ((u, dv)) = hdv, uiU = Bq (v, u),
2. The hyperbolic functor 113

while q(u, v) = q(u) + q(v) + Bq (u, v) = Bq (u, v), since q/U = 0 and
q/V = 0.
The last assertion reduces to the preceding ones we show that U =
U ⊥ . Lemma 1.2 shows that U ⊥ is a direct summand of rank [U ⊥ : k] =
[P : k] − [U : k] ≤ [U : k], because, by assumption, [P : k] ≤ 2[U : k].
But we also have q/U = 0 so U ⊂ U ⊥ , and therefore U = U ⊥ , as
claimed.

Lemma 2.2. A quadratic module (P, q) is non-singular if and only if

(P, q) ⊥ (P, −q) ≈ H(P),

provided P is reflexive.

Proof. P reflexive implies H(P) is non-singular, and hence likewise for 143
any orthogonal summand.
Suppose now that (P, q) is non-singular. Then so is (P, q) ⊥ (P, −q) =
(P ⊕ P, q1 = q ⊥ (−q)).
Let U = {(x, x)ǫP ⊕ P|x ∈ P}. Then q1 /U = 0, and U is a direct
summand of P ⊕ P, isomorphic to P. If U & U ⊥ we can find a (0, y)ǫU ⊥ ,
y , 0. Then, for all x ∈ P,

0 = Bq1 ((x, x), (0, y)) = q1 (x, x + y) − q1 (x, x) − q1 (0, y)

= q(x) − q(x + y) + q(y)
= −Bq (x, y).

Since Bq is non-singular this contradicts y , 0. Now the Lemma follows

from Lemma 2.1. 

Lemma 2.3. Let P be a reflexive module and let (Q, q) be a non-singular

quadratic module with Q finitely generated and projective. Then

H(P) ⊗ (Q, q) ≈ H(P ⊗ Q).

Proof. The hypothesis on Q permits us to identify (P ⊗ Q)∗ = P∗ ⊗ Q∗ ,

so it follows that (W, q1 ) = H(P) ⊗ (Q, q) is non-singular. We shall apply
Lemma 2.1 by taking
 114 5. The structure of the Clifford Functor

U = P⊗Q ⊂ W = (P⊗Q)⊕(P∗ ⊗Q). If xi ⊗yi ǫU, then q1 (Σxi ⊗yi ) =

P
ΣqP (xi )q(yi ) + i< j Bq1 (xi ⊗ yi , x j ⊗ y j ) = i< j BP(xi , x j )Bq (yi , y j ) = 0,
P P
because qP /P = 0 in H(P). Thus U ⊂ U ⊥ , and to show equality it
suffices clearly to show that (P∗ ⊗ Q) ∩ U ⊥ = 0. If Σxi ⊗ yi ∈ U
and Σw ⊗ z j ∈ (P∗ ⊗ Q) ∩ U ⊥ then 0 = Bq1 (Σxi ⊗ yi , Σw j ⊗ z j ) =
P P j
B (xi , w j )Bq (yi , z j ). 
i, j

144 Since (P∗ ⊗ Q)∗ = P ⊗ Q∗ (P is reflexive) the non-singularity of q

guarantees that all linear functionals on P∗ ⊗ Q have the form i BP (xi , )
P
Bq (yi , ), so Σw j ⊗ z j is killed by all linear functionals, hence is zero. We
have now shown U = U ⊥ so the lemma follows from Lemma 2.1.
A quadratic space is a non-singular quadratic module (P, q) with
P finitely generated and projective, i.e. Pǫ obj P, the category of such
=
modules. We define the category
Quad = Quad(k)

with
objects : quadratic spaces
morphisms : isometries
product :⊥
The discussion at the beginning of this section shows that
H : P → Quad
=

is a product preserving functor of categories with product (in the sense

of chapter 1), and Lemma 2.1 shows that H is cofinal. We thus obtain
an exact sequence from Theorem 4.6 of chapter 1. We summarize this:
Proposition 2.4. The hyperbolic functor
H : P → Quad
=

is a cofinal functor of categories with product. It therefore induces (The-

orem 4.6 of chapter 1) an exact sequence
K1 P → K1 Quad → K0 ΦH → K0 P → K0 Quad → Witt(k) → 0,
= =
2. The hyperbolic functor 115

where we define Witt (k) = coker (K0 H).

We close this section with some remarks about the multiplicative 145
structures. Tensor products endow K0 Quad with a commutative multi-
plication, and Lemma 2.3 shows that the image of K0 H is an ideal, so
Witt (k) also inherits a multiplication. The difficulty is that, if 2 is not
invertible in k, then these are rings without identity elements. For the
identity should be represented by the form q(x) = x2 on k. But then
Bq (x, y) = 2xy is not non-singular unless 2 is invertible.
Here is one natural remedy. Let Symbil denote the category of
non-singular symmetric bilinear forms, (P, B) with Pǫ obj P. If (P, B) ∈
=
Symbil and (Q, q)ǫQuad define

(P, B) ⊗ (Q, q) = (P ⊗ Q, B ⊗ q), (2.5)

where B ⊗ q is the quadratic form qB⊗B0 , for some B0 ǫ Bil(Q × Q) such

that q = qB0 . It is easy to see that B ⊗ q does not depend on the choice
of B0 . Moreover, the bilinear form associated to B ⊗ q is (B ⊗ B0) + (B ⊗
B0 )∗ = (B ⊗ B0 ) + (B∗ ⊗ B∗0 ) = B ⊗ (B0 ⊗ B∗0 ) = B ⊗ Bq , because B = B∗.
Since B and Bq are non-singular so is B ⊗ Bq so (P ⊗ Q, B ⊗ q) ∈ Quad.
If aǫk write hai for the bilinear module (k, B) with B(x, y) = axy for
x, yǫk. If a is a unit then haiǫSymbil.
Tensor products in Symbil make K0 Symbil a commutative ring, with 146
identity h1i, and (2.5) makes K0 Quad a K0 Symbil−module. The “forget-
ful” functor Quad → Symbil, (P, q) 7−→ (P, Bq ), induces a K0 Symbil-
homomorphism K0 Quad → K0 Symbil, so its image is an ideal. The
hyperbolic forms generate a K0 Symbil submodule, image K0 H), of
K0 Quad, so Witt (k) is a K0 Symbil-module. This follows from an ana-
logue of Lemma 2.3 for the operation (2.5)
Similarly, the hyperbolic forms, (P ⊕ P∗ , BP), generate an ideal in
K0 Symbil which annihilates Witt(k). Lemma 2.2 says that h1i ⊥ h−1i
also annihilates Witt (k).
 116 5. The structure of the Clifford Functor

3 The Clifford Functor

If P is a k-module we write
T (P) = (k) ⊕ (P) ⊕ (P ⊗ P) ⊕ . . . ⊕ (P⊗n ) ⊕ . . .
for its tensor algebra. If (P, q) is a quadratic module then its Cliford
algebra is
Cl(P, q) = T (P)/I(q),
where I(q) is the two sided ideal generated by all x ⊗ x − q(x)(x ∈ P). If
we grade T (P) by even and odd degree (a (Z/2Z)-grading) then x ⊗ x −
q(x) is homogeneous of even degree, so
Cl(P, q) = Cl0 (P, q) ⊕ Cl1 (P, q)
is a graded algebra in the sense of chapter 4, We will consider Cl(P, q) to
be a graded algebra, and this must be borne in mind when we discuss
tensor products.
147 The inclusion P ⊂ T (P) induces a k-linear map
C P : P → Cl(P, q)
such that C P (x)2 = q(x) for all x ∈ P, and C P is clearly universal among
such maps of P into a k-algebra.
Cl evidently defines a functor from quadratic modules, and their
morphisms, to graded algebras, and their homomorphisms (of degree
zero). Moreover it is easy to check that both T and Cl commute with
base change, k → K. The next lemma says Cl is “product preserving,”
in an appropriate sense.
Lemma 3.1. There is a natural isomorphism of (graded) algebras,
Cl((P1 , q1 ) ⊥ (P2 , q2 )) ≈ Cl(P1 , q1 ) ⊗ Cl(P2 , q2 ).
C P1 ⊕P2
Proof. Pi → P1 ⊗ P2 −−−−−→ Cl((P1 , q1 ) ⊥ (P2 , q2 )) induces an algebra
homomorphism,
fi : Cl(Pi , qi ) → Cl((P1 , q1 ) ⊥ (P2 , q2 )), i = 1, 2.

3. The Clifford Functor 117

If xi ∈ Pi then ( f1 x1 + f2 x2 )2 = (q1 ⊥ q2 )(x1 , x2 ) = q1 x1 + q2 x2 =

( f1 x1 )2 + ( f2 x2 )2 , so f1 x1 and f2 x2 , being of odd degree, commute (in
the graded sense). Therefore so also do the algebras they generate, im f1
and im f2 . Hence f1 and f2 induce an algebra homomorphism

F : Cl(P1 , q1 ) ⊗ Cl(P2 , q2 ) → Cl((P1 , q1 ) ⊥ (P2 , q2 )),

and this is clearly natural. To construct its inverse let

g : P1 ⊕ P2 → Cl((P1 , q1 ) ⊥ (P2 , q2 )) by g(x1 , x2 ) = C1 x1 ⊗ 1+ 1⊗C2 x2 ,

where Ci = C Pi . If g extends to an algebra homomorphism from Cl ⊥ 148

((P1 , q1 ) ⊥ (P2 , q2 )) it will evidently be inverse to F, since this is so
on the generators, C P1 ⊕P2 (P1 ⊕ P2 ), and C1 P1 ⊗ 1 + 1 ⊗ C2 P2 , respec-
tively. To show that g extends we have to verify that g(x1 , x2 )2 = (q1 ⊥
q2 )(x1 , x2 ). g(x1 , x2 )2 = (C1 x1 )2 ⊗ 1 + 1 ⊗ (C2 x2 )2 + C1 x1 ⊗ C2 x2 +
(−1)(deg C1 x1 )(deg C2 x2 ) (C1 x1 ⊗ C2 x2 ) = q1 x1 + q2 x2 = (q1 ⊥ q2 )(x1 , x2 ).

Examples 3.2. If q is the quadratic form q(x) = ax2 on k, denote this

quadratic module by (k, a). Then C1(k, a) = khai = k1 ⊕ kx, x2 = a.
Thus C1(R, −1) ≈ C = R1 ⊕ Ri, for example.
a, b
C1((k, a) ⊥ (k, b)) ≈ C1(k, a) ⊗ C1(k, b) ≈ ( ). (3.3)
k
The latter denotes the k-alegbra with free k-basis 1, xa , xb , y, where
deg xa = deg xb = 1, x2a = a, x2b = b, y = xa xb = −xb xa . The degree zero
component is !
a, b
= k[y], y2 = −ab.
k 0
−1 − 1
For example, as a graded R-algebra, C ⊗R C ≈ ( ), the standard
k
quaternion algebra (plus grading).

H(k) = (k ⊕ k∗ , qk ). (3.4)

Let e1 be a basis for k (e.g. e1 = 1) and e2 the dual basis for k∗ .

Writing q = qk we have q(a1 e1 + a2 e2 ) = a1 a2 . Hence C1(H(k)) is
 118 5. The structure of the Clifford Functor

generated by elements x1 and x2 (the images of e1 and e2 ) of degree

1 with the relations 2 2
  x1 = 0 = x2 and x1 x2 + x2 x1 = 1. In M2 (k) the
matrices y1 = 00 10 and y2 = 01 00 satisfy these relations, so there is a 149
homomorphism

Cl(H(k)) → M2 (k)
xi 7→ yi

It is easy to check that this is an isomorphism. This isomorphism

is the simplest case of Theorem below, which we prepare for in the
following lemmas.

Lemma 3.5. Let P be a k-module. There is a k-linear map P∗ → Homk

(T (P), T (P)), f 7→ d f , where d f is the unique map of degree −1 on T (P)
such that d f (x ⊗ y) = f (x)y − x ⊗ d f (y) for xǫP, y ∈ T (P). Moreover
d2f = 0 and d f dg + dg d f = 0 for f , gǫP∗ . If q is a quadratic form on
P then d f I(q) ⊂ I(q), so d f induces a k-linear map, also denoted d f , of
degree one on Cl(P, q).

Proof. d f is defined on P⊗(n+1) = P ⊗ P(⊗n) by induction on n, from the

formula given. This shows uniqueness, and that
X
d f (x0 ⊗ . . . ⊗ xn ) = (−1)i f xi (x0 ⊗ . . . î . . . xn ).
0≤i≤n

Hence
X
d2f (x0 ⊗ . . . ⊗ xn ) = (−1)i+ j ( f xi )( f x j )(x0 ⊗ . . . ĵ . . . î . . . ⊗ xn )
0≤ j<i≤n
X
+ (−1)i+ j−1 ( f xi )( f x j )(x0 ⊗ . . . î . . . ĵ . . . ⊗ xn )
0≤ j<i≤n

= 0.

150 It is easy to check that f 7→ d f is k-linear, so we have 0 = (d f +g )2 =

(d f + dg )2 , and hence d f dg + dg d f = 0 for f , gǫP∗ .
3. The Clifford Functor 119

The formula above shows that if x, x, y ∈ hT (P) (the set of homo-

geneous elements) then

d f (x ⊗ y) = d f x ⊗ y + (−1)∂x x ⊗ d f y.

If q is a quadratic form on P write u(x) = x⊗ x− q(x) for xǫP. Since u(x)

has even degree the formula above shows that d f (u(x)⊗y) = d f u(x)⊗y+
u(x) ⊗ d f y, and d f (u(x)) = f (x)x − f (x)x = 0, so d f (u(x) ⊗ y) ∈ I(q). If
v ∈ hT (P) then d f (v ⊗ u(x) ⊗ y) = d f v ⊗ u(x) ⊗ y + (−1)∂v v ⊗ d f (u(x) ⊗ y) ∈
I(q). Since I(q) is additively generated by all such v ⊗ u(x) ⊗ y it follows
that d f I(q) ⊂ I(q).

Lemma 3.6. If B ∈ Bil(P × P) there is a unique k-linear map λB :

T (P) → T (P) satisfying

(i) λB (1) = 1

(ii) λB L x = (L x + dB(x, ) )λB for x ∈ P.

(Here L x denotes left multiplication by x in T (P).)λB also has the fol-

lowing properties:

(a) λB preserves the ascending filtration on T (P) and induces the iden-
tity map on the associated graded module.

(b) For f ǫP∗ , λB d f = d f λB .

(c) λ0 = 1T (P) and λB+B′ = λB ◦ λB′ for B, B′ ǫ Bill (P × P).

(d) If q is a quadratic form on P, then λB I(q) = I(q−qB ), and λB induces 151

an isomorphism Cl(P, q) → Cl(P, q − qB ) of filtered modules.

Proof. Writing xy in place of x ⊗ y in T (P), (ii) reads:

λB (xy) = xλB (y) + dB(x, ) (λB (y))(xǫP, yǫT (P)).

Starting with λB (1) = 1 this gives an inductive definition of λB on P(⊗n) ,

since the right side is k-bilinear in x and y. Moreover (a) follows also
from this by induction on n.
 120 5. The structure of the Clifford Functor

(b) We prove that λB d f = d f λB by induction on n, the case n = 0

being clear (from (a)). For x ∈ P, y ∈ hT (P),

λB d f (xy) = λB (( f x)y − x(d f y))

= ( f x)(λB y) − (x(λB (d f y) + dB(x, ) (λB (d f y)))
d f λB (xy) = d f (x(λB y) + dB(x, ) (λB y))
= ( f x)(λB y) − x(d f (λB y)) + d f dB(x, ) (λB y)

Their equality follows from d f λB y = λB d f y (induction) and the fact

(Lemma 3.5) that d f dB(x, ) = −dB(x, ) d f .
(c) If B = 0 then dB(x, ) = 0 for all x so (ii) reads λ0 L x = L x λ0 , and
1T (P) solves this equation for λ0 . We prove λB 0λ′B = λB+B′ by checking
(i) (which is clear) and (ii):

λB ◦ λB′ (xy) = x(λB ◦ λB′ y) + dB+B′ (x, ) (λB ◦ λB′ y).

λB λB′ (xy) = λB (x(λB′ y) + dB′ (x, ) (λB′ y))
= xλB λB′ y + dB(x, ) (λB λB′ y) + dB′ (x, ) (λB λB′ y)
= x(λB λB′ y) + (dB(x, ) + dB′ (x, ) )(λB λB′ y).
and dB(x, ) + dB′ (x, ) = d(B+B′ )(x, ) .

152 (d) Let I = {u ∈ T (P)|λB (u)ǫI(q − qB )}. λB (xu) = x(λB u) +

dB(x, ) (λB u), so, thanks to Lemma 3.5, I is a left ideal. λB ((X 2 − (qx))y) =
xλB (xy) + dB(x, ) λB (xy)

− (qx)(λB y) = x(xλB y + dB(x, ) (λB x)) + dB(x, ) (xλB y + dB(x, ) λB y)

− (qx)(λB y) = x2 (λB y) + xdB(x, ) (λB y) + B(x, x)(λB y) − xdB(x, ) (λB y)
− (qx)(λB y) (we have used d2B(x, ) = 0; Lemma 3.5) =
= (x2 − (qx − qB x))λB y ∈ I(q − qB ).

Thus I is a left ideal containing all (x2 − qx)y, so it contains I(q). We

have proved

λB I(q) ⊂ I(q − qB ) = λB λ−B I(q − qB ) ⊂ λB (I(q − qB − q−B ) = λB I(q),

using (c). Now (a) implies λB induces an isomorphism Cl(P, q) →

Cl(P, q − qB ) of filtered modules. 
3. The Clifford Functor 121

Corollary 3.7. Giving Cl(P, q) the filtration induced by the ascending

filtration on T (P), the structure of Cl(P, q) as a filtered module is inde-
pendent of q. In particular, taking q = 0, we have an isomorphism

Cl(P, q) ≈ Λ(P)

of filtered modules.
Proof. Writing q = qB for some B ∈ Bil (P × P) we obtain an isomor-
phism Cl(P, q) → Cl(P, 0) = Λ(P), induced by λB . 

Corollary 3.8. C P : P → Cl(P, q) is a monomorphism. If U is a direct

summand of P then the map

Cl(U, q/U) → Cl(P, q),

induced by the inclusion U ⊂ P, is a monomorphism.

Proof. The first assertion follows from the commutativity of 153

λB
Cl(P, q) / ∧(P)
cG
GG {=
GG
G {{{
C P ) GGG {{
{{
P
and the fact that P → ∧(P) is a monomorphism. Let B′ = B/U × U, so
q/U = q′B . Then it is easily checked that

λB
Cl(P, q) / ∧(P)
O O

Cl(U, q′ ) / ∧(U)
λB′

is commutative, so the second assertion follows since ∧(U) → ∧(P) is

injective. 

∧(P) = T (P)/I(), I(0) being the (homogeneous) ideal generated by

all x ⊗ x, xǫP.
∧(P) = k ⊕ ∧1 P ⊕ ∧2 P ⊕ . . .
 122 5. The structure of the Clifford Functor

and ∧1 P ≈ P. Lemma 3.1 givens a natural isomorphism

∧(P ⊕ Q) ≈ ∧(P) ⊕ ∧(Q)

of (Z/2Z)- graded algebras. ∧ therefore defines a product preserving

functor
∧ : P → FP ,
= 2
if, for P finitely generated and projective, we view ∧(P) as a faithfully
projective module, graded modulo 2. Similarly, by virtue of Lemma 3.1,
the Clifford algebra defines a product preserving functor,

Cl : Quad → (graded algebras, ⊗)

154 Theorem 3.9. If (P, q) ∈ obj Quad, then Cl(P, q) ∈ obj Az , i.e. it is
2
a graded azumaya algebra. The resulting functor Cl : Quad → Az
2
renders the diagram
P H / Quad
=

∧ Cl
 
FP / Az
2 END 2

commutative up to natural isomorphism, i.e. for P finitely generated and

projective,
Cl(H(P)) ≈ END (∧(P))
as graded algebras.
Corollary 3.10. There is a natural map of exact sequences.

K1 P / K1 Quad / K0 ΦH / K0 P / K0 Quad / Witt (k) /0

=

     
K1 FP / K1 Az / K0 ΦEND / K0 FP / K0 Az / Br2 (k) /0
2 2 2 2

In Theorem 4.6 of Chapter 4 we exhibited an exact sequence

0 → Br(k) → Br2 (k) → Q2 (k) → 0,

3. The Clifford Functor 123

where Q2 (k) was “the group of graded quadratic extensions of k.” The
map above assigns to the class of (P, q) in Witt (k) and element β of
Br2 (k). The projection of β in Q2 (k) corresponds, in the classical case
when k is a field, to the discriminant, if char k , 2, and the Arf invariant
if char k = 2. The remaining contribution from Br(k) is essentially the
Hasse invariant.

Proof of 3.9. We want to construct a natural isomorphism 155

ϕP : Cl(H(P)) → END (∧(P)).

for P ∈ obj P. Suppose this is done. Then if (P, q) ∈ obj Quad, we

have (P, q) ⊥ (P, −q) ≈ H(P) (Lemma 2.2), so Cl(P, q) ⊗ Cl(P, −q) ≈
Cl(P, q) ⊥ (P, −q)) (Lemma 3.1) ≈ Cl(H(P)) = END(∧(P)), by assump-
tion. Therefore, by criterion (6) of Theorem 4.1, Chapter 4, Cl(P, q) is a
graded azumaya algebra. Thus we only have to construct ϕP .
H(P) = (P ⊕ P∗ , qP ) with qP (x, y) = hy, xiP
= y(x) for (x, y) ∈ P ⊗ P∗ . Define

P ⊗ P∗ → END(∧(P))

by (x, y) 7→ L x + dy . Then, using Lemma 3.5, (L x + dy )2 = L x2 + L x dy +

dy L x + dy2 = L x dy + dy L x , because x2 = 0 in ∧(P) and dy2 = 0. If u ∈ ∧(P)
then (L x dy + dy L x )u = xdy (u) + dy (xu) = xdy (u) + y(x)u − xdy (u) =
y(x)u. Thus (L x + dy )2 is multiplication by y(x) = qP (x, y) on ∧(P), i.e.
(L x + dy )2 = qP (x, y) in END (∧(P)). Thus we have defined an algebra
homomorphism

ϕP : Cl(H(P)) → END(∧(P)),

and since L x +dy has degree 1, it is a homomorphism of graded algebras.

Suppose f : P1 → P2 is an isomorphism. Then on ∧(P2 ), L f (x) =
∧( f )L x ∧ ( f )−1 and d f ∗ − 1y (x2 ) = ( f ∗−1 y)(x2 ) = y( f −1 x2 ), so ∧( f )dy ∧
( f )−1 x2 ) = ∧( f )dy ( f −1 x2 ) = ∧( f )y( f −1 x2 ) = d f ∗−1y (x2 ) for x2 ∈ P2 ,
because y( f −1 x2 ) has degree zero in ∧(P2 ). Therefore ∧( f )(L x + dy ) ∧ 156
( f )−1 = L f (x) + d f ∗−1(y) so it follows that ϕP is natural, recalling that
H( f ) = f ⊕ f ∗−1 .
 124 5. The structure of the Clifford Functor

Next we will show that, for P = P1 ⊗ P2 the following diagram is

commutative:
ϕP1 ⊕P2
Cl(H(P1 ⊕ P2 )) / END(∧(P1 ⊕ P2 ))

≈ ≈
 
Cl(H(P1 ) ⊥ H(P2 )) END(∧PO 1 ⊗ ∧P2 )
≈ ≈

Cl(H(P1 )) ⊗ Cl(H(P2 )) ϕ / END(∧P1 ) ⊗ END(∧P2 )
P1 ⊗ϕP2

To see this, we trace the images of ((x1 , x2 ), (y1 , y2 ))

∈ (P1 ⊗ P2 ) ⊕ (P∗1 ⊗ P∗2 ) ⊂ Cl(H(P1 ⊕ P2 )) :

((x1 , x2 ),_(y1 , y2 ))  / L(x1 ,x2 ) + d(y1 ,y2 )

_


 L x1 ⊗1 + L1⊗x2
((x1 , y1 ),_(x2 , y2 ))
+dy1 ⊗ 1∧P2 + 1∧P1 ⊗ dy2
O

_

((x1 , y1 ) ⊗ 1) + (1 ⊗ (x2 , y2 ))  / ((LX1 + dy1 ) ⊗ 1∧P2 )
(1∧P1 ⊗ (L x2 + dy2 ))

157 Since all of these algebras are faithfully projective k-modules we con-
clude that ϕP1 ⊕P2 is an isomorphism ⇔ ϕP1 ⊕ ϕP2 is an isomorphism
⇔ ϕP1 and ϕP2 are isomorphisms. (In Chapter 2 we showed that the
functor Q⊗ is faithfully exact for Q faithfully projective.)
Now given P1 we choose P2 so that P1 ⊕P2 ≈ k⊕· · ·⊕k, and then the
problem of showing that ϕP1 is an isomorphism reduces to the special
case P1 = k. We do this case now by a direct calculation.
H(k) ≈ (ke1 ⊕ ke2 , q) with q(a1 e1 + a2 e2 ) = a1 a2 . Here ke2 = (ke1 )∗
and e2 is the dual basis to e1 , i.e. e2 (e1 ) = 1. Therefore ∧(ke1 ) = k[e1 ] =
k1 ⊕ ke1 with e21 = 0, and de2 (1) = 0, de2 (e1 ) = 1. Moreover, Le1 (1) = e1
and Le1 (e1 ) = 0.
4. The orthogonal group and spinor norm 125

Cl(H(ke1 )) = k[e1 , e2 ] with e21 = 0 = e22 , and e1 e2 + e2 e1 = 1,

because 1 = q(e1 + e2 ) = (e1 + e2 )2 .

ϕk : Cl(H(ke1 )) → END(∧(ke1 ))

is defined by ϕk (e1 ) = Le1 and ϕk (e2 ) = de2 . With respect to the basis
 e1 for ∧(ke
1,  1) these endomorphisms are represented by the matrices
0 0 and 0 1 , respectively, and these clearly generate M (k). Thus ϕ
10 00 2 k
is surjective. On the other hand Corollary 3.7 says that, as a module,
Cl(H(k)) ≈ ∧(k ⊕ k∗ ), a free module of rank four (because ∧(k ⊕ k∗ ) ≈
∧(k) ⊗ (k∗ )). A surjective homomorphism of free modules of the same
finite rank must be an isomorphism, so ϕk is an isomorphism as claimed.

4 The orthogonal group and spinor norm

We assume here that spec (k) is connected. Suppose (P, q) is a quadratic 158
space (i.e. ∈ obj Quad (k)) and that [P : k] = n. If n is odd then
2 ∈ U(k); otherwise reduce (P, q) modulo a maximal ideal containing
2k, and we contradict the fact that non-singular forms over fields of char
2 have even dimension. We propose to use the Clifford algebra,

A = Cl(P, q) = A◦ ⊕ A1

to study the orthogonal group

Ω = Ω(P, q),

i.e. the group of isometries of (P, q).

We take the position from Chapter 4 that everything is graded. Thus

Pic(k) = Pic |k| ⊕ Z/2Z (4.1)

is the group of invertible k-modules. The first summand describes the

underlying ungraded module (|k|-module) and the Z/2Z summand des-
ignates the degree (0 or 1) in which it is concentrated.
Write
G(A) = Autk−alg (A),
 126 5. The structure of the Clifford Functor

and hU(A) for the homogeneous units of  A. Recall that if u ∈ hU(A)

uau−1 if ∂u = 0


then αu ∈ G(A) is defined by αu (a) =  . Here, for
ua′ ν−1 if ∂u = 1

a = a0 + a1 , a′ = a0 − a1 . Thus, for homogeneous a, we can write this
as
αu (a) = (−1)∂u∂a uau−1 (u ∈ hU(A), a ∈ hA).
159 According to Theorem 3.9 A is azumaya k-algebra. Therefore The-
orem 5.2 of chapter 4 gives us an exact sequence

1 → U(k) → U(A◦ ) → G(A) → Pic(k) (4.2)

u → αu

To apply this we first embed Ω in G(A). Indeed, since the Clifford

algebra is a functor of (P, q) there is a canonical homomorphism, α 7→
C(α), of Ω into G(A). If we identify P ⊂ A (in fact P ⊂ A1 ) then C(α)
is the unique algebra automorphism of A such that C(α)(x) = α(x) for
x ∈ P. For example, the automorphism a 7→ a′ described above is just
C(−1P ). We will use this monomorphism to identify Ω with a subgroup
of G(A). We can characterise it:

Ω = {α ∈ G(A) αP ⊂ P}.

For if αP ⊂ P then for x ∈ P we have q(αx) = (αx)2 = α(x2 ) = α(qx) =

qx, so α induces an isometry, α′ : P → P. Evidently then α = C(α′ ).
Next we introduce the Clifford group

Γ = {u ∈ hU(A)|αu ∈ Ω}.

and the special Clifford group

Γ◦ = Γ ∩ A◦ = {u ∈ U(A◦ ) αu ∈ Ω}.

If u ∈ U(k) ⊂ U(A◦ ) then αu = 1, so U(k) ⊂ Γ◦ . Therefore the exact

sequence (4.2) induces a sub-exact sequence,

1 / U(k) / U(A◦ ) / G(A) / Pic(k) (4.3)

S S

1 / U(k) / Γ◦ /Ω / Pic(k)
 4. The orthogonal group and spinor norm 127

160 Now Pic(k) = Pic |k| ⊗ Z/2Z (see (4.1)), so we obtain homomorphisms

Ω → Pic(k)

and
Ω → Z/2Z,
the second being the first followed by projection on the second factor.
We shall write

S Ω = ker(Ω → Z/2Z)
[

VS Ω = ker(Ω → Pic(k)),

the special, and very special orthogonal groups (of (P, q)), respectively.
With this notation we can extract from (4.3) an exact sequence

1 → U(k) → Γ◦ → vS Ω → 1. (4.4)

If x ∈ P then x2 = qx in A, therefore also in A◦ , so the identity map on

P extends to an isomorphism A → A◦ , or, in other words, an antiauto-
morphism of A. We shall denote it by a 7→ ã. All α ∈ Ω commute with
this antiautomorphism; just check it on P. For a∈ A we will define its
conjugate, ā, by
ā = ã′ = ã′
and write
Na = aā.
The last remark shows that α(ā) = α(a) for αεΩ.
Let 161
n = {a ∈ A Na ∈ k}.
If x ∈ P then x̄ = x̃′ = x′ = −x so N x = −x2 = −q(x) ∈ k, and P ⊂ n.
Suppose a, b ∈ n. Then N(ab) = (ab)ab ¯ = abb̄ā = aN(b)ā =
aāN(b) = N(a)N(b), because N(b) ∈ U(k).

a, b ∈ n ⇒ ab ∈ n and N(ab) = N(a)N(b).

 128 5. The structure of the Clifford Functor

Next suppose u ∈ Γ. Then for x ∈ P we have αu (x) = (−1)∂u uxu−1 ,

or u′ x = αu (x)u. Therefore ūαu (x) = x̄ū′ , and αu (x) = αu ( x̄) with
x̄ = −x. Setting y = αu ( x̄) we have ūy = α−1 ′
u (y)ū , so, by definition,
ū ∈ Γ and αū = α−1 u . In particular αuū = αu αū = 1 so N(u) = uū ∈
ker(Γ → G(A)) = U(k). Summarizing, we have proved:
If u ∈ Γ then ū ∈ Γ and αū = α−1
u .
Moreover N(u) ∈ U(k) (i.e. Γ ⊂ n) so

u−1 = N(u)−1 ū.

Thus N defines a homomorphism

N : Γ → U(k).

We now introduce the groups

N
Pin = ker(Γ −→ U(k))

162 and
N
Spin = ker(Γ◦ −→ U(k)).
If u ∈ U(k) then ū = u so N(u) = u2 . Therefore if we apply N to the
exact sequence (4.4) we obtain a commutative diagram with exact rows
and columns:

1 1 1 (4.5)

  
1 / 2 U(k) / Spin / VS Ω′ /1

  
1 / U(k) / Γ0 / VS Ω /1

N σ
  
1 / U(k)2 / U(k) / U(k)/U(k)2 /1


1
4. The orthogonal group and spinor norm 129

Here 2 U(k) denotes units of order 2 (square roots of 1). σ : VS Ω →

U(k)/U(k)2 is called the spinor norm, and its kernel, VS Ω′ , the spino-
rial kernel.
So far we have the following subgroups, with indicated successive
quotients, of Ω:
Ω)

⊂ Pic(k) = Pic |k| ⊕ Z/2Z

VS Ω
)
σ
→ U(k)/U(k)2
− (4.6)
VS Ω′
)
≈ Γ◦ /U(k) ⊂ U(A◦ )/U(k).
1
163
Of course the bulk of the group is the bottom layer. We shall now
investigate this for small values of n = [p : k].
n = 1 (so 2 ∈ U(k)). A◦ = k, A1 = P, and Γ◦ = U(k). Ω = {±1} in
this case.
n = 2 A1 = P so Γ◦ = U(A◦ ). A◦ is a quadratic extension of k (in
the sense of chapter 4, §3,) so Γ◦ is abelian group. If (P, q) = H(k) then
A◦ = k × k so Γ◦ = U(k) × U(k) and VS Ω ≈ U(k).
n = 3 (so 2 ∈ U(k)). Then A1 = P ⊕ L1 , where L1 is the degree
one term of L = |A||A| the centre of the ungraded algebra A. A◦ is a
“quaternion |k|-algebra,” i.e. azumaya |k|-algebra of rank 4, and N(a) ∈ k
for all a ∈ A◦ . If u ∈ U(A◦ ) then. Since N(u) = uū ∈ U(k), we have 164
u−1 = N(u)−1 ū. Therefore, for a ∈ A,

αu (a) = uau−1 = ū−1 ā ū = (N(u)u−1 )−1 ā(N(u)u−1 )

= uāu−1 = αu (ā).

Consequently αu leaves invariant the eigenspaces of − ; these behave

=
nicely because a = a and 2 ∈ U(k).
Now x̄ = −x for x ∈ P. If we localize k then P has an orthogo-
nal basis, e1 , e2 , e3 , and it is easy to see that L1 = ke1 e2 e3 , e1 e2 e3 =
 130 5. The structure of the Clifford Functor

(−1)3 e3 e2 e1 = (−1)3+2+1 e1 e2 e3 . Therefore, under the action of − , A1 =

P ⊕ L1 is the eigenspace decomposition. In summary we have observed
that u ∈ U(A◦ ) ⇒ Nu ∈ U(k) ⇒ αu leaves the eigenspaces of − invariant
⇒ αu P ⊂ p, i.e. u ∈ Γ◦ . Therefore

Γ◦ = U(A◦ ) and VS Ω = U(A◦ )/U(k).

In case A◦ = M2 (k) we have Γ◦ = GL2 (k), the norm N is just the deter-
minant, and
VS Ω = PGL2 (k) = GL2 (k)/U(k).
n = 4. In this case L = A◦A◦ is a quadratic extension of k (in the sense
of Chapter 4, §3), and A◦ is a quaternion L-algebra. The norm N takes
values in L. In case 2 ∈ U(k) a calculation like that for the case n = 3
(localize k and diagonalize (P, q) first) shows that

Γ◦ = {u ∈ U(A◦ ) Nu ∈ U(k)}.

and this is probably in general. In case A◦ M2 (L) then N : U(A◦ ) =

GL2 (L) → U(L) is just the determinant. Hence S L2 (L) ⊂ Γ0 and
Γ◦ /S L2 (L) ≈ U(k), in this case. VS Ω = Γ◦ /U(k) ⊃ S L2 (L)/2 U(k),
165 i.e. modulo element of order 2 in the centre, and modulo this sub-
group VS Ω lands in U(k)/U(k)2 . Note that if L = k × k then S L2 (L) =
S L2 (k) × S L2 (k).
Suppose k happens to be a Dedekind ring of arithmetic type in a
global field. Then one knows that Pic(k) is finite (finiteness of class
number) and U(k) is finitely generated (Dirichlet unit theorem). Hence
VS Ω = Γ0 /U(k) is of finite index in Ω. The discussion above shows,
therefore, that the finite generation of Ω is equivalent to the finite gen-
eration of Γ0 , and that for n ≤ 4 this is “usually” equivalent to the finite
generation of U(A0 ). The point is that U(A0 ) is often an easily recog-
nized linear group.
One can similarly use this procedure to reduce the study of normal
subgroups of Ω to those of U(A0 ), at least in many cases, for n ≤ 4.
Bibliography

[1] Auslander M. and Goldman O. -[1] The Brauer group of a com- 166
mutative ring, Trans. A.M.S., 97 (1960), pp. 367-409.

[2] Azumaya G. -[1] On maximally central algebras, Nagoys Math. J.,

Vol. 2 (1951), pp. 119-150.

[3] Bass H. -[K] K-theory and stable algebra I.H.E.S, n0 22 (1964),

pp. 489-544.
-[2] The Morita theorems, mimeographed notes.

[4] Bourbaki N. -[1] Algébre, Chapitre VIII, Hermann, Paris, 1958.

-[2] Algébre, Chapitre IX, Hermann, Paris, 1959.

[5] Freya P. -[1] Abelian categories, Harper and Row, New York,
1964.

[6] Gabriel P. -[1] Des catégories abélinnes, Bull. Soc. Math. France,
90(1962), pp. 323-448.

[7] Heller A. -[1] Some exact sequences in algebraic K-theory, Topol-

ogy, Vol.3 (1965), pp.389-408.

[8] Maclane S. -[1] Natural associativity and commutativity, Rice Uni-

versity Studies, 49(1963), pp.28-46.

[9] Mitchell B. -[1] The full imbedding theorem, Amer. J. Math., Vol.
86 (1964), pp. 619-637.

131
132 BIBLIOGRAPHY

[10] Morita K. -[1] Duality for modules and its applications to the the-
ory of rings with minimum condition. Science reports of the Tokyo
Kyoiku Daigaku, Vol. 6 Sec A (1958), pp. 83-142.

[11] O’Meara O.T. -[1] Introduction to quadratic forms, Springer-

Verlag,1963.

[12] Rosenberg A. and Zelinsky D. -[1] Automorphisms of separable

algebras, Pacific J. Math., Vol. 11 (1961), pp. 1109-1117.

[13] Wall C.T.C -[1] Graded Brauer groups, Crelle’s J., Vol. 213 (1964),
pp. 187-199.

[14] Witt E. -[1] Theorie der quadratischen Formen in beliebigen

Körpern, Crelle’s J., Vol.176 (1937), pp. 31-44.