Universitext

Matej Brešar

Introduction to
Noncommutative
Algebra
Universitext
Universitext

Series editors
Sheldon Axler
San Francisco State University, San Francisco, CA, USA

Vincenzo Capasso
Università degli Studi di Milano, Milan, Italy
Carles Casacuberta
Universitat de Barcelona, Barcelona, Spain
Angus MacIntyre
Queen Mary University of London, London, UK

Kenneth Ribet
University of California at Berkeley, Berkeley, CA, USA
Claude Sabbah
CNRS, École Polytechnique Centre de mathématiques, Palaiseau, France
Endre Süli
University of Oxford, Oxford, UK

Wojbor A. Woyczynski
Case Western Reserve University, Cleveland, OH, USA

Universitext is a series of textbooks that presents material from a wide variety of

mathematical disciplines at master’s level and beyond. The books, often well
class-tested by their author, may have an informal, personal, even experimental
approach to their subject matter. Some of the most successful and established
books in the series have evolved through several editions, always following the
evolution of teaching curricula, into very polished texts.

Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.

More information about this series at http://www.springer.com/series/223

Matej Brešar

Introduction
to Noncommutative
Algebra

123
Matej Brešar
University of Ljubljana
Ljubljana
Slovenia

and

University of Maribor
Maribor
Slovenia

ISSN 0172-5939 ISSN 2191-6675 (electronic)

ISBN 978-3-319-08692-7 ISBN 978-3-319-08693-4 (eBook)
DOI 10.1007/978-3-319-08693-4

Library of Congress Control Number: 2014943746

Mathematics Subject Classification: 16-01, 16Dxx, 16Kxx, 16Nxx, 16Pxx, 16Rxx, 16Sxx, 16Uxx

Springer Cham Heidelberg New York Dordrecht London

Ó Springer International Publishing Switzerland 2014

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publisher’s location, in its current version, and permission for use must
always be obtained from Springer. Permissions for use may be obtained through RightsLink at the
Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

In memory of my father
Preface

The purpose of this book is to give a gentle introduction to noncommutative rings

and algebras that requires fewer prerequisites than most other books on the subject.
It is based on a series of lectures given to masters students at the University of
Ljubljana and is intended to serve as a first reading on noncommutative algebra for
beginning graduate and advanced undergraduate students. The first two chapters,
in which some results of historic importance are derived by rudimentary tools, can
have a wider audience. For example, they can be used in a course for future high
school teachers, or for an undergraduate seminar. Mathematicians working in areas
that only have some interactions with noncommutative algebra may also benefit
from this book in which important, classical themes are presented in a simple
manner, avoiding excessive generality.
The necessary background needed to follow this text is a standard knowledge of
linear algebra and a basic knowledge about groups, rings, and fields. To make
precise what we mean by this, the book begins with a survey of prerequisites. This
is followed by the first chapter which considers finite dimensional division alge-
bras. Its most prominent results are Frobenius’ theorem on real division algebras
and Wedderburn’s theorem on finite division rings. The second chapter is devoted
to the structure of finite dimensional algebras, featuring the classical Wedderburn’s
theory. After the first two warm-up chapters, which mostly deal with results known
for a hundred years, the next two introduce and study more abstract notions:
modules, vector spaces over division rings, and tensor products. The Double
Centralizer Theorem and the Skolem-Noether Theorem are included therein. The
fifth chapter considers the structure theory of rings. The main themes are primitive
rings, the Jacobson Density Theorem, and the Jacobson radical. If the first five
chapters survey the ‘‘greatest hits’’ of noncommutative algebra, the last two are
slightly more specialized, reflecting my personal taste and interests. They treat
polynomial identities and related notions, such as free algebras and rings of
quotients.
The order of the topics in the book is quite close to the chronological order of
their development. This actually occurred unintentionally, probably as a result of
my attempt to follow the principle that an advanced concept should be introduced

vii
viii Preface

only when truly needed, and not at a higher level of generality than necessary.
Not because I do not appreciate abstract concepts; after all, they are what makes
mathematics beautiful. But it is because they can be better understood and valued
when given some evidence that they are indispensable. Another principle I have
followed, especially in the early chapters, is to choose proofs that appear to be the
simplest. This may not always be the same as the shortest, but rather proofs that
seem very natural and easy to memorize. It has been my desire to write the book in
such a way that readers would not get an impression that one needs supernatural
abilities to create mathematics, but that even they themselves may be able, with
some luck and courage, to discover a little piece of it.
One challenge in the exposition of the book was to find new proofs of classical
theorems that would fulfill the simplicity criteria described above. This has been in
fact my little passion over the last years, which has eventually yielded some
results. The proofs of the following theorems are different from those in the
standard sources: Frobenius’ theorem (Sect. 1.1), the Skolem-Noether Theorem
(Sect. 1.6), Wedderburn’s structure theorems (Sect. 2.9), Martindale’s theorem on
prime GPI-rings (Sect. 7.7), Posner’s theorem on prime PI-rings (Sect. 7.9), and
the Formanek-Razmyslov Theorem on central polynomials (Sect. 7.10). How
much originality is there in these proofs? Frankly, I do not know. I have not found
such proofs in the literature, and have published them in mathematical journals
through a series of papers. On the other hand, so many mathematicians have
known these theorems, especially the first three listed, for so many years that one
hardly imagines that something entirely new about them can still be found. It is not
so rare that some mathematical ideas are discovered, forgotten, and rediscovered.
Anyway, I hope that these proofs provide interesting alternatives to those from
other sources. Two remarks must be added. First, the proof of Frobenius’ theorem
was obtained in a joint work with Peter Šemrl and Špela Špenko. Second, shortly
after the publication of my paper on alternative proofs of Wedderburn’s structure
theorems Edmund Puczylowski informed me that he had used them in his class,
but with a modification—which I immediately liked. I am now happy to use this
modified version in the book.
Each chapter ends with exercises of varying difficulty levels. They are sorted by
topics, not by difficulty. Some of the harder ones are accompanied with hints, but
others are not, in order to give the student the opportunity to fully enjoy the
pleasure of discovering the solution. A fair amount of exercises are original, but
there are also some that appear, either as exercises or as theorems with proofs, in
almost every book on the subject. I tried to avoid long, theoretical exercises, and
gave preference to those that I had found entertaining and appealing.
The list of references at the end of the book is very short. It includes only books
which are explicitly referred to on relatively rare occasions when details are not
provided (say, when closing the discussion on a topic with additional information,
or when omitting tedious details in some example). Most of these books are
textbooks on noncommutative rings and algebras, but written at a more advanced
level than this one. An exception is perhaps J. Beachy’s Introductory Lectures on
Rings and Modules [Bea99] whose emphasis, however, differs from ours. At the
Preface ix

time of writing I was frequently consulting many of the listed books, but mainly
T. Y. Lam’s A First Course in Noncommutative Rings [Lam01] (together with the
accompanying problem book [Lam95]) and L. H. Rowen’s Graduate Algebra:
Noncommutative View [Row08]. I warmly recommend these two texts as further
reading to students who want to dig deeper into the field of noncommutative
algebra.
English is not my mother tongue and I apologize to the reader if this is
sometimes too obvious. ‘‘Translating’’ the English stuck in my mind into the
English as it should be has been a big and often frustrating challenge for me during
the writing process. The bright side of this is that it gives me the opportunity to
thank my sons Jure and Martin for their continuous help in this matter.
The most difficult step in writing a book is forcing yourself to write the first
sentence. I am thankful to my friend and colleague Peter Šemrl for giving me the
necessary encouragement to make it.
Rough drafts of most chapters were first read by Špela Špenko. Her criticism
has been very helpful (as always). Aljaž Zalar has solved most of the exercises and
made me think of modifying some. Igor Klep provided many helpful suggestions
concerning Chaps. 1 and 6. The book has been read cover to cover by Nik Stopar
and Janez Šter. Their comments have contributed to the improvement of the text
and saved me from some errors. My sincere thanks go to all of them, as well as to
the Springer staff for their professional assistance.

Ljubljana and Maribor, March 2014 Matej Brešar

Contents

1 Finite Dimensional Division Algebras . . . . . . . . . . . . . . . . . . . . . . 1

1.1 After the Complex Numbers: What Comes Next? . . . . . . . . . . 1
1.2 Beyond Frobenius’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Simple Rings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Central Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Multiplication Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Automorphisms of Central Simple Algebras . . . . . . . . . . . . . . 13
1.7 Maximal Subfields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.8 Wedderburn’s Theorem on Finite Division Rings. . . . . . . . . . . 16
1.9 Further Examples of Division Algebras . . . . . . . . . . . . . . . . . 18
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 Structure of Finite Dimensional Algebras . . . . . . . . . . . . . . . . . . . 25

2.1 Nilpotent Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Prime and Semiprime Rings . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Unitization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 The Regular Representation. . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 Group Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Matrix Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.7 Idempotents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.8 Minimal Left Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.9 Wedderburn’s Structure Theorems . . . . . . . . . . . . . . . . . . . . . 42
2.10 Algebras Over Special Fields . . . . . . . . . . . . . . . . . . . . . . . . 45
2.11 Scalar Extension (A Naive Approach) . . . . . . . . . . . . . . . . . . 47
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3 Modules and Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.1 Concept of a Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 Basic Module-Theoretic Notions . . . . . . . . . . . . . . . . . . . . . . 56
3.3 Vector Spaces Over Division Rings . . . . . . . . . . . . . . . . . . . . 59
3.4 Endomorphisms and Matrices . . . . . . . . . . . . . . . . . . . . . . . . 61

xi
xii Contents

3.5 Simple Modules. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.6 Maximal Left Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.7 Schur’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.8 Semisimple Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.9 Wedderburn’s Structure Theory Revisited . . . . . . . . . . . . . . . . 69
3.10 Chain Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4 Tensor Products. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1 Concept of a Tensor Product . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2 Basic Properties of Tensor Products . . . . . . . . . . . . . . . . . . . . 81
4.3 Linear (In)dependence in Tensor Products . . . . . . . . . . . . . . . 84
4.4 Tensor Product of Algebras. . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5 Multiplication Algebra and Tensor Products . . . . . . . . . . . . . . 91
4.6 Centralizers in Tensor Products . . . . . . . . . . . . . . . . . . . . . . . 92
4.7 Scalar Extension (The ‘‘Right’’ Approach) . . . . . . . . . . . . . . . 94
4.8 Simplicity of Tensor Products . . . . . . . . . . . . . . . . . . . . . . . . 96
4.9 The Skolem-Noether Theorem . . . . . . . . . . . . . . . . . . . . . . . . 98
4.10 The Double Centralizer Theorem . . . . . . . . . . . . . . . . . . . . . . 99
4.11 The Brauer Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5 Structure of Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.1 Primitive Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 The Jacobson Density Theorem . . . . . . . . . . . . . . . . . . . . . . . 110
5.3 Alternative Versions and Applications . . . . . . . . . . . . . . . . . . 113
5.4 Primitive Rings Having Minimal Left Ideals . . . . . . . . . . . . . . 116
5.5 Primitive Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.6 Introducing the Jacobson Radical . . . . . . . . . . . . . . . . . . . . . . 121
5.7 Quasi-invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.8 Computing the Jacobson Radical . . . . . . . . . . . . . . . . . . . . . . 126
5.9 Semiprimitive Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.10 Structure Theory in Action . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

6 Noncommutative Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.1 Free Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2 Algebras Defined by Generators and Relations . . . . . . . . . . . . 140
6.3 Alternating Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.4 Polynomial Identities: Definition and Examples. . . . . . . . . . . . 145
6.5 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.6 Stable Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.7 T-ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Contents xiii

6.8 The Characteristic Polynomial . . . . . . . . . . . . . . . . . . . . . . . . 153

6.9 The Amitsur-Levitzki Theorem . . . . . . . . . . . . . . . . . . . . . . . 157
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

7 Rings of Quotients and Structure of PI-Rings . . . . . . . . . . . . . . . . 163

7.1 Rings of Central Quotients . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.2 Classical Rings of Quotients . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.3 Ore Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.4 Martindale Rings of Quotients . . . . . . . . . . . . . . . . . . . . . . . . 171
7.5 The Extended Centroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.6 Linear (In)dependence in Prime Rings . . . . . . . . . . . . . . . . . . 177
7.7 Prime GPI-Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.8 Primitive PI-Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7.9 Prime PI-Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.10 Central Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Symbols

N Set of positive integers

Z Ring of integers
Zn Ring of integers modulo n
Q Field of rational numbers
R Field of real numbers
C Field of complex numbers
H Division algebra of quaternions
δij Kronecker delta
jSj Cardinality of the set S
SnT Set difference
idS Identity map on S
detðAÞ Determinant of the matrix A
trðAÞ Trace of the matrix A
M Group of invertible elements in the monoid M
Sn Symmetric group on f1; . . .; ng
sgnðσÞ Sign of the permutation σ
R ffi R0 Isomorphic rings (or algebras, modules, etc.)
ker ϕ Kernel of the homomorphism ϕ
im ϕ Image of the homomorphism ϕ
R=I Factor ring
IJ Product of ideals
½a; b Commutator of a and b
Z(R) Center of the ring R
char(R) Characteristic of R
Ro Opposite ring of R
Mn(R) Ring of all n  n matrices over R
Eij Standard matrix units
eij Matrix units

xv
xvi Symbols

Π i2I Ri Direct product of a family of rings

R1      Rn Direct product of a finite family of rings
R½ω Ring of polynomials over R in one indeterminate
R½Ω Ring of polynomials over R in indeterminates from the set Ω
R½½ω Formal power series ring over R
soc(R) Socle of R
rad(R) Jacobson radical of R
dimF V Dimension of the vector space V over the field F
½A : F Dimension of the algebra A over F
Tn(F) Algebra of upper triangular n  n matrices over F
C½a; b Algebra of continuous functions from [a, b] to R
An nth Weyl algebra
F½G Group algebra of the group G over F
La Left multiplication map x 7! ax
Rb Right multiplication map x 7! xb
M(A) Multiplication algebra of the algebra A
A] Unitization of A
AK Scalar extension of A to the field K
CA(S) Centralizer of the set S in A
annR ðMÞ Annihilator of the R-module M
EndR ðMÞ Ring of endomorphisms of the R-module M
EndD ðVÞ Ring of linear operators of the vector space V over the division
ring D
EndF ðVÞ Algebra of linear operators of the vector space V over the field F
i2I Mi Direct sum of a family of submodules (or additive subgroups, or
subspaces)
M1      Mn Direct sum of a finite family of submodules (or additive
subgroups, or subspaces)
M1 þ    þ Mn Sum of a finite family of submodules (or additive subgroups, or
subspaces)
UV Tensor product of vector spaces
uv Simple tensor
ϕψ Tensor product of linear maps
AB Tensor product of algebras
Fhξ; ηi Free algebra in two indeterminates
FhXi Free algebra in indeterminates from the set X
deg( f ) Degree of the polynomial f
sn Standard polynomial
cn Capelli polynomial
IdðAÞ T-ideal of all polynomial identities of the algebra A
GMn ðFÞ Algebra of generic matrices
b
Z Field of quotients of the commutative domain Z
Symbols xvii

FðΩÞ Field of rational functions in Ω

QZ ðRÞ Ring of central quotients of the ring R
Qrc ðRÞ Right classical ring of quotients of R
Qr ðRÞ Right Martindale ring of quotients of R
RC Central closure of R
Prerequisites

This preliminary chapter aims to give an accurate survey of the background

material needed to follow this text. It is supposed to be used as a handbook, only
when there is a need. We will therefore write it in a concise and condensed
manner, paying less attention to motivation and insight than in the rest of the book.
Some simpler proofs that are as much a part of understanding the subject as
definitions and theorems will be outlined, but deeper results will be just stated as
facts. The reader should be warned that the topics will be presented in a slightly
unbalanced manner. Some will be examined carefully, while other equally or even
more important ones will be only touched upon or even omitted. Of course, the
emphasis will be on topics that are needed later.
Through this chapter we also indicate what knowledge the reader is assumed to
have. We gain different knowledge at different places, so there might be at least
something new for many of the readers. Maybe we will also mention a couple of
things for which it would be pretentious to say that everyone should have learnt
them in one of the undergraduate courses. Some notational and terminological
conventions will also be included. Therefore, even well-prepared readers may
occasionally consult this chapter, say in case they find something ambiguous in the
text. Most of the conventions, however, are standard. Let us just point out two
issues before we start.
Formally, the term ‘‘noncommutative ring’’ means a not necessarily commu-
tative ring. But we shall use this term in a loose manner. By saying that this book is
about noncommutative rings we mean that it considers rings that can be
commutative or not, while by saying that some particular ring is noncommutative
we shall mean that this ring is not commutative. At any rate, even when we use this
term in its precise sense, the stress is on rings that are not commutative. The goals
and the means of the theory of commutative rings are somewhat different.
We will not assume that every ring must have a unity 1, which may be a subject
of dispute. Not that this is uncommon, but the fact is that many mathematicians
believe that, at least in a textbook, the existence of 1 should be added to the axioms
of a ring. This has advantages, but also disadvantages. There are some interesting
and important classes of non-unital rings, e.g., nilpotent rings, nil rings, radical

xix
xx Prerequisites

rings, and simple rings consisting of finite rank operators. In the non-unital setting
we can treat ideals as (sub)rings and we can avoid the unitization process for rings
that seem more natural without unity. On the other hand, sometimes the non-unital
situation is not essentially more interesting, but causes technical difficulties which
are distracting for a newcomer. At such instances we will simply add the
assumption that a ring has a unity. Of course, some price has to be paid for
omitting a possible axiom; admittedly, here and there our exposition would be
slightly simpler if the presence of unity was automatically assumed. The readers
who believe that all rings should have unities can simply skip parts of the text
where the lack of unity creates some problems. There are not that many anyway.

Sets and Relations

First, a word about notation. By the set of natural numbers, N, we mean positive
(not nonnegative) integers:
N ¼ f1; 2; . . .g:

We assume there is no need to explain the meaning of ;, \, [, ,

. By ( and )
we denote the strict inclusion, by S n T the set fs 2 S j s 62 Tg, and by jSj the
cardinality of the set S.
If I is a (possibly infinite) index set and fSi j i 2 Ig is a family of sets, then
Π i2I Si denotesSthe Cartesian product of the sets Si . Thus, Π i2I Si is the set of all
maps f : I ! i2I Si such that f ðiÞ 2 Si for every i. Setting si :¼ f ðiÞ we can write
f as ðsi Þ. If I is finite, say I ¼ f1; . . .; ng, then we write ðsi Þ as ðs1 ; . . .; sn Þ, and
Π i2I Si as S1      Sn . In particular, S1  S2 is the set fðs1 ; s2 Þ j s1 2 S1 ; s2 2 S2 g.
By idS we denote the identity map on the set S. Given a map f : X ! Y and sets
S X, T Y, we write f ðSÞ for ff ðsÞ j s 2 Sg and f 1 ðTÞ for fx 2 X j f ðxÞ 2 Tg.
By f jS we denote the restriction of f to S. If f is bijective, then f 1 denotes its
inverse map. The function composition will be denoted by fg rather than f  g. The
reason behind this is that functions appearing in algebra are often considered as
elements of an algebraic system in which multiplication is defined as composition.
A relation on a set S is said to be an equivalence relation if it is reflexive,
symmetric, and transitive. This means that a a, if a b then b a, and if a b
and b c then a c. The equivalence class of an element a is the set of all b 2 S
such that a b. It is easy to see that S is a disjoint union of its equivalence classes.
An equivalence relation on S thus gives rise to a partition of S.
Our next goal is to state a somewhat less elementary, yet one of the most
frequently used results in mathematics—the so-called Zorn’s lemma. A set S
together with a relation  which is reflexive (a  a), anti-symmetric (a  b and
b  a implies a ¼ b), and transitive (a  b and b  c implies a  c) is called a
partially ordered set. For example, the power set of a given set is partially
ordered by inclusion. A subset C of a partially ordered set S is called a chain if for
Prerequisites xxi

each pair a; b 2 C we have either a  b or b  a. An element u 2 S is called an

upper bound for C if c  u for every c 2 C. Finally, an element m 2 S is called a
maximal element of S if for every a 2 S, m  a implies m ¼ a. We are now in a
position to state
Zorn’s lemma Let S be a nonempty partially ordered set. If every chain in S
has an upper bound in S, then S contains a maximal element.
Zorn’s lemma is equivalent to the axiom of choice which says the following: If
fAi j i 2 Ig is a family of nonempty sets, then there exists a family of elements
fai j i 2 Ig such that ai 2 Ai for every i 2 I.

Groups

A semigroup is a set S together with an associative binary operation. This means

that S is endowed with a map S  S ! S, ðx; yÞ 7! xy, such that ðxyÞz ¼ xðyzÞ for
all x; y; z 2 S. It is easy to see that associativity makes it possible for us to write all
products, including those of more than three elements, without parentheses. In
particular, we can define the nth power xn of every x 2 S and every n 2 N. We say
that elements x; y 2 S commute if xy ¼ yx. If xy ¼ yx holds for all x; y 2 S, then we
say that S is commutative.
An element e 2 S that satisfies ex ¼ x for every x 2 S is called a left identity
element. Similarly, a right identity element is every element e0 2 S satisfying
xe0 ¼ x for every x 2 S. If S has both a left identity element e and a right identity
element e0 , then they are equal; indeed, they both must be equal to ee0 . In this case
we call e ¼ e0 an identity element of S. From now on we will denote an identity
element by 1.
A semigroup with a (necessarily unique) identity element 1 is called a monoid.
An element a in a monoid M is said to be right invertible if there exists b 2 M
such that ab ¼ 1. Every such element b is called a right inverse of a. Analogously
we define left invertibility and left inverses. We say that a is invertible if it is both
right and left invertible. In this case any right inverse of a is equal to any left
inverse of a. Indeed, ab ¼ ca ¼ 1 implies b ¼ ðcaÞb ¼ cðabÞ ¼ c. Thus, there
exists a unique element which is both a left and right inverse of a. We call it the
inverse of a and denote it by a1 . If a is invertible, then so is an for every n 2 N.
Its inverse is ða1 Þn which we also denote by an . For convenience we set a0 ¼ 1
for every a 2 M. Further, if a and b are invertible, then so is ab, and
ðabÞ1 ¼ b1 a1 . The set of all invertible elements of M will be denoted by M  .
A monoid in which every element is invertible is called a group. In the next
paragraphs we give some basic examples.
If M is a monoid, then M  is clearly a group. For example, Z is a monoid under
(the usual) multiplication, and hence Z ¼ f1; 1g is a group. Similarly, if F 2
fQ; R; Cg then F is a monoid under multiplication, and F ¼ F n f0g. The set
Mn ðFÞ of all n  n matrices with entries in F is a monoid under the usual matrix
xxii Prerequisites

multiplication; the corresponding group Mn ðFÞ of all invertible matrices is called

the general linear group (over F).
Let X be a set. A bijective map from X onto X is called a permutation. The set
of all permutations of X is a group under the composition. Indeed, the composition
is associative, idX is the identity element, and inverses clearly exist. We call it the
symmetric group of X. If X ¼ f1; . . .; ng, then we denote the symmetric group of
X by Sn . It is easy to see that jSn j ¼ n!. A permutation in Sn is called a transposition
if it interchanges two of the elements and leaves all the rest fixed. Every
permutation σ 2 Sn can be written as a product of transpositions. If
σ ¼ τ1 . . .τ p ¼ ρ1 . . .ρq , where all τi and ρi are transpositions, then either both p
and q are even or both p and q are odd. In the first case we say that σ is an even
permutation, and in the second case we  say that it is an odd permutation. We define
1 if σ is even
the sign of σ, sgnðσÞ, by sgnðσÞ :¼ .
1 if σ is odd
A commutative group is called abelian. So far we have used only multiplicative
notation for binary operations. For abelian groups we often use additive notation.
This means that the result of the operation is written as x þ y instead of xy. An
abelian group in which this notation is used is called an additive group. (The
convention that an additive group must necessarily be abelian is not entirely
standard, but it is convenient to adopt it in this book.) An identity element in an
additive group is written as 0. Further, we write x instead of x1 and nx instead of
xn . A simple example of an additive group is Z under the usual addition. Another
example is Zn ¼ f0; 1; . . .; n  1g where addition is taken modulo n.
Until further notice we continue using the multiplicative notation. A map ϕ
from a group G into a group G0 is called a group homomorphism, or simply a
homomorphism when it is clear from the context that groups are the subject of
consideration, if ϕðxyÞ ¼ ϕðxÞϕðyÞ for all x; y 2 G. The set ker ϕ :¼ fx 2 G j ϕðxÞ
¼ 1g is called the kernel of ϕ, and the set im ϕ :¼ fϕðxÞ j x 2 Gg ð¼ ϕðGÞÞ is
called the image of ϕ. Note that ϕ is injective if and only if ker ϕ ¼ f1g.
A bijective homomorphism is called an isomorphism. We say that groups G
and G0 are isomorphic, and write G ffi G0 , if there exists an isomorphism ϕ from G
onto G0 . In this case its (set-theoretic) inverse ϕ1 is an isomorphism from G0 onto
G. The relation of being isomorphic is an equivalence relation. Isomorphic groups
have exactly the same properties; informally we consider them as identical,
although they may appear very different
 
at a glance. For example, consider the set
G of all matrices of the form β α with α; β 2 R, α 6¼ 0 or β 6¼ 0. One can
α β

check that G is a group under multiplication, and that the map

 
α β
7! α þ βi
β α
Prerequisites xxiii

is an isomorphism from G onto C ¼ C n f0g. Thus, as long as we are interested

only in multiplication, the set of matrices G is just a disguised form of the more
familiar set of nonzero complex numbers. Another example: The multiplicative
group ð0; 1Þ is isomorphic to the additive group R via the isomorphism x 7! lnðxÞ.
A homomorphism from a group into itself is called an endomorphism. An
automorphism is an endomorphism which is also an isomorphism. An injective
homomorphism ϕ : G ! G0 is sometimes called an embedding. This term is
usually used when we wish to point out that by identifying G with im ϕ we may
view G as a subgroup of G0 . We also say that G embeds in G0 . For example, the
additive group R can be canonically embedded into the additive group C by
α 7! α þ 0i.
Every group G can be embedded into a symmetric group. This result is known
as Cayley’s theorem, and can be proved as follows. For any a 2 G we define
‘a : G ! G by ‘a ðgÞ ¼ ag. One immediately checks that ‘a is a permutation of G,
‘a ¼ idG if and only if a ¼ 1, and ‘ab ¼ ‘a ‘b for all a; b 2 G. Accordingly, a 7! ‘a
is an embedding of the group G into the symmetric group of the set G.
A homomorphism between additive groups is also called an additive map. We
will also come across biadditive maps: We say that a map f : G1  G2 ! G0 ,
where G1 ; G2 ; G0 are additive groups, is biadditive if it is additive in each variable,
i.e.,
f ðx þ y; zÞ ¼ f ðx; zÞ þ f ðy; zÞ and f ðx; z þ wÞ ¼ f ðx; zÞ þ f ðx; wÞ
for all x; y 2 G1 , z; w 2 G2 .
A subset H of a group G is called a subgroup if it contains 1, is closed under
the group operation, i.e., h; h0 2 H implies hh0 2 H, and h1 2 H whenever h 2 H.
Note that H is itself a group. A (left) coset of H in G is a set xH :¼ fxh j h 2 Hg
where x is a fixed element in G. It is easy to see that jxHj ¼ jyHj for all x; y 2 G,
xH ¼ yH if and only if x1 y 2 H, and xH 6¼ yH if and only if xH \ yH ¼ ;. From
these observations one easily infers Lagrange’s theorem which says that the
jGj
number of cosets of a subgroup H in a finite group G is equal to jHj . In particular,
jHj thus divides jGj. Let us point out one corollary to this fundamental result. We
say that an element a in a group G has finite order if ar ¼ 1 for some r 2 N. If r is
the smallest natural number with this property, then r is called the order of a. In
this case hai :¼ f1; a; . . .; ar1 g is a subgroup of G, called the cyclic subgroup
generated by a. If G is finite, then every element in G is of finite order, and hence
Lagrange’s theorem tells us that jhaij divides jGj for every a 2 G. Since jhaij ¼ r
is the order of a it follows that ajGj ¼ 1 for every a 2 G.
Let a be an arbitrary element in a group G. Every element of the form xax1 ,
x 2 G, is called a conjugate of a. A subgroup N of G is called a normal subgroup
if it contains every conjugate of each of its elements, i.e., a 2 N implies xax1 2 N
for every x 2 G. Every subgroup of an abelian group is normal. The center of a
group G, defined as the set of all elements in G that commute with every other
element in G, is a normal subgroup of G. If ϕ : G ! G0 is a homomorphism, then
xxiv Prerequisites

ker ϕ is a normal subgroup of G, while im ϕ is a subgroup of G0 which is not

necessarily normal.
The next construction works for normal subgroups of arbitrary groups, but we
confine ourselves to additive groups as this is all we need. Thus, let H be a
subgroup of an additive group G. A coset of H in G is in the additive notation
written as x þ H. Let G=H denote the set of all cosets of H in G endowed with the
addition defined by
ðx þ HÞ þ ðy þ HÞ :¼ ðx þ yÞ þ H:

It is easy to see that this operation is well-defined, i.e., x þ H ¼ x0 þ H and y þ

H ¼ y0 þ H implies ðx þ yÞ þ H ¼ ðx0 þ y0 Þ þ H, and that G=H is a group under
this operation. We call it the factor group of G by H. The map x 7! x þ H is a
surjective homomorphism from G onto G=H with the kernel H. It is called the
canonical homomorphism.
Let G0 be another additive group. If ϕ : G ! G0 is a homomorphism, then
G= ker ϕ ffi im ϕ:
An isomorphism is given by x þ ker ϕ 7! ϕðxÞ. Indeed, this map is well-defined
since x þ ker ϕ ¼ y þ ker ϕ implies x  y 2 ker ϕ and hence ϕðxÞ ¼ ϕðyÞ. Clearly,
it is additive and a bijection from G= ker ϕ onto im ϕ.
Given groups G1 ; . . .; Gn , we define their direct product G1      Gn as the
Cartesian product of the sets G1 ; . . .; Gn endowed with the componentwise
operation:
ðx1 ; . . .; xn Þðy1 ; . . .; yn Þ ¼ ðx1 y1 ; . . .; xn yn Þ:
It is immediate that G1      Gn is a group, and that each Gi can be naturally
(canonically) embedded in G1      Gn . The direct product Π i2I Gi of a possibly
infinite family of groups fGi j i 2 Ig is defined analogously. The operation is thus
given by ðxi Þðyi Þ ¼ ðxi yi Þ. If Gi are additive groups, then the operation in Π i2I Gi is
also written additively: ðxi Þ þ ðyi Þ ¼ ðxi þ yi Þ.
Let now G1 ; . . .; Gn be subgroups of an additive group G. Then so is their sum
G1 þ    þ Gn :¼ fx1 þ    þ xn j x1 2 G1 ; . . .; xn 2 Gn g:

If
Gi \ ðG1 þ    þ Gi1 þ Giþ1 þ    þ Gn Þ ¼ f0g

for every i, then G1 þ    þ Gn is called the direct sum of the subgroups Gi , and is
denoted by G1      Gn : Every element in G1      Gn can be written as
x1 þ    þ xn where xi 2 Gi in a unique way, so the xi ’s basically play the role of
components. The difference between the groups G1      Gn and G1      Gn
is merely formal. In more accurate terms, they are canonically isomorphic via
x1 þ    þ xn 7! ðx1 ; . . .; xn Þ.
Prerequisites xxv

Rings

A ring R is an additive group endowed with another associative binary operation

ðx; yÞ 7! xy, called multiplication, such that the distributive laws hold in R, i.e.,
ðx þ yÞz ¼ xz þ yz and xðy þ zÞ ¼ xy þ xz

for all x; y; z 2 R. In other words, ðx; yÞ 7! xy is a biadditive map. The following

simple facts follow easily from the definition: 0x ¼ x0 ¼ 0, ðxÞy ¼ xðyÞ ¼
xy, ðxÞðyÞ ¼ xy for all x; y 2 R.
A ring R is commutative if xy ¼ yx for all x; y 2 R. For example, Z, with the
usual addition and multiplication, is a commutative ring. The additive group Zn
becomes a commutative ring if we take multiplication modulo n. The set C½a; b of
continuous functions f : ½a; b ! R, endowed with the usual addition and
multiplication of functions, is a commutative ring.
Commutative rings play a secondary role in this book. But we will be often
concerned with the ‘‘most commutative’’ part of our rings. The set
ZðRÞ :¼ fc 2 R j cx ¼ xc for all x 2 Rg
is called the center of the ring R, and its elements are called central elements.
Obviously, R is commutative if and only if ZðRÞ ¼ R. If R is not commutative then
ZðRÞ is more often than not a relatively small subset of R.
We assume the reader is familiar with at least one example of a noncommu-
tative ring, namely with Mn ðRÞ, the ring of all n  n matrices with entries in R.
Here one takes the usual matrix addition and multiplication. We will introduce
matrix rings in a more accurate and general manner in the next section.
One of the most intriguing properties in rings is that the product of two nonzero
elements can be 0. This cannot happen in, say, Z, but can easily occur in Zn if n is
not prime, or in Mn ðRÞ. A nonzero element x in a ring R is said to be a left zero-
divisor if there exists a nonzero element y 2 R such that xy ¼ 0. Analogously we
define a right zero-divisor. An element that is both a left and right zero-divisor is
called a zero-divisor. For example, a central element that is a left (or right) zero-
divisor is automatically a zero-divisor.
Every ring is a semigroup under multiplication. If it is actually a monoid, then it
will be called a unital ring. Thus, a ring R is unital if it contains an element 1 such
that 1x ¼ x1 ¼ x for all x 2 R. We will call 1 the unity of R. Most rings that first
come into our minds are unital; a simple example of a ring without unity is 2Z, the
ring of even integers. If 1 ¼ 0 in a ring R, then R is the trivial or zero ring whose
only element is 0.
In a unital ring we can speak about left invertible, right invertible, and
invertible elements. A nonzero unital ring D in which every nonzero element is
invertible (i.e., D ¼ D n f0g) is called a division ring. Division rings obviously
have no (left, right) zero-divisors. A commutative division ring is called a field.
xxvi Prerequisites

Obvious examples are Q, R, and C. It is easy to show that Zp with p prime is a

field. Conversely, Zn is not a field if n is not prime.
We say that a ring R has characteristic n if n is the smallest natural number
such that nx ¼ 0 for every x 2 R (note that in a unital ring R the condition nx ¼ 0
for every x 2 R is equivalent to n1 ¼ 0). If no such natural number n exists, then
we say that R has characteristic 0. The characteristic of R will be denoted by
charðRÞ. For example, charðZÞ ¼ 0 and charðZn Þ ¼ n. If R is unital and m 2 N is
such that m1 ¼ 0, then m ¼ rs implies ðr1Þðs1Þ ¼ 0. Hence, either R has zero-
divisors or charðRÞ is prime. In particular, the characteristic of a field is either 0 or
a prime number p. Fields of prime characteristic are called fields of finite
characteristic. The subset F0 of a field F consisting of all elements of the form
ðn1Þðm1Þ1 , n; m 2 Z, m1 6¼ 0, is easily seen to be a field isomorphic to Q if
charðFÞ ¼ 0, and to Zp if charðFÞ ¼ p. We call F0 the prime subfield of F.
A map ϕ from a ring R into a ring R0 is called a ring homomorphism if
ϕðx þ yÞ ¼ ϕðxÞ þ ϕðyÞ and ϕðxyÞ ¼ ϕðxÞϕðyÞ
for all x; y 2 R. Thus, ϕ is an additive group homomorphism and simultaneously a
multiplicative semigroup homomorphism. Its kernel is ker ϕ :¼ fx 2 R j ϕðxÞ ¼ 0g;
and its image is im ϕ :¼ fϕðxÞ j x 2 Rg: The other necessary definitions are the
same as for group homomorphisms (as well as for homomorphisms of other alge-
braic structures). Thus, an isomorphism is a bijective homomorphism, an endo-
morphism is a homomorphism from a ring into itself, an automorphism is an
endomorphism which is also an isomorphism, and an embedding is an injective
homomorphism. The relation of being isomorphic is an equivalence relation. We
write R ffi R0 if R and R0 are isomorphic. Let us modify our earlier

example

on groups
α β
to give one illustration. The set R of all matrices of the form β α
with α; β 2 R, is
easily seen to be a ring isomorphic to C. An isomorphism is defined exactly as
before, just that now there is no need to exclude the case where α ¼ β ¼ 0. Math-
ematicians are inclined to ‘‘forget’’ about the difference between isomorphic objects,
so some might even say that R is the field of complex numbers.
A subset of a ring R which is a subgroup with respect to addition is called an
additive subgroup of R. Let S be an arbitrary nonempty subset of R. The set S of
all elements of the form m1 s1 þ    þ mn sn where mi 2 Z and si 2 S is an additive
subgroup of R which contains S. Moreover, S is the smallest additive subgroup
containing S in the sense that it is contained in every other additive subgroup
which contains S. We call S the additive subgroup generated by S. (For S ¼ ; we
define S ¼ f0g.)
An additive subgroup of R which is closed under multiplication is called a
subring of R. A subring is itself a ring. For example, the center ZðRÞ is a subring
of R which is commutative as a ring. We remark that ZðRÞ can contain only 0 even
Prerequisites xxvii

when R is a nonzero ring. If, however, R is unital, then 1 2 ZðRÞ. A subring D of a

unital ring R that contains 1 and is a division ring is called a division subring of R.
If it is also commutative, then we call it a subfield of R.
The subring generated by the subset S of R is, of course, defined as the smallest
subring of R containing S. Note that it is equal to the additive subgroup generated
by the set of all possible (finite) products s1 . . .sr of elements si 2 S. If the subring
generated by S is R itself, then we call S a generating set for the ring R, and its
elements are called generators of R. We also say that S generates R. A ring is said
to be finitely generated if it contains a finite generating set.
The terms ‘‘generated by’’, ‘‘generating set’’, ‘‘generators’’, ‘‘finitely generated’’
are used, in the same sense as above, for other algebraic structures as well. We
hope that their meaning can be understood without always giving a detailed
explanation.
A subring I of R is called a left ideal of R if xu 2 I for all x 2 R and u 2 I.
Analogously, a subring I of R is called a right ideal of R if ux 2 I for all x 2 R and
u 2 I. If I is both a left and right ideal, then I is called an ideal; we also say a two-
sided ideal when we wish to stress that it is not only left or right. Left and right
ideals are also called one-sided ideals.
For example, the ideals of the ring Z are nZ :¼ fnk j k 2 Zg, n  0. In
commutative rings the adjectives ‘‘left’’ 
and ‘‘right’’ are, of course, superfluous.
The set of all matrices of the form β 0 , α; β 2 R, is an example of a left ideal but
α 0

not an ideal of the ring M2 ðRÞ. By putting zeros in one of the rows instead of in a
column we obtain a right ideal that is not an ideal.
Every ring R has at least two ideals: f0g and R. From now on we will write f0g
as 0. The reader should thus be warned that the symbol 0 can mean two different
things, either the zero element or the zero (or trivial) ideal. From the context it
should always be clear what we have in mind.
A (left, right, two-sided) ideal I different from R is said to be proper. Note that
in this case I does not contain any invertible element if R is unital.
If I and J are ideals of R, then so is their intersection I \ J, their sum
I þ J :¼ fu þ v j u 2 I; v 2 Jg, and their product IJ which is defined as the
additive subgroup generated by the set fuv j u 2 I; v 2 Jg. The same is true if we
replace ‘‘ideal’’ by ‘‘left ideal’’ or ‘‘right ideal’’ (incidentally, the sum and
the product of subrings may not be a subring). We remark that, in accordance with
the definition of the product, I n should be understood as the additive subgroup
generated by the set fu1 . . .un j ui 2 Ig. Let us also point that IJ I \ J if I; J are
two-sided ideals.
Let S be a subset of R and let a; b 2 R. We write
aS :¼ fax j x 2 Sg; Sa :¼ fxa j x 2 Sg; aSb :¼ faxb j x 2 Sg:
Note that Ra is a left ideal and aR is a right ideal. By RaR we denote the additive
subgroup generated by the set fxay j x; y 2 Rg, which is clearly an ideal. If R is
unital, then Ra is the left ideal generated by a, aR is the right ideal generated by a,
xxviii Prerequisites

and RaR is the ideal generated by a. If R is not unital, then the left ideal generated
by a is equal to Ra þ Za where Za :¼ fka j k 2 Zg, the right ideal generated by a
is equal to aR þ Za, and the ideal generated by a, which is sometimes denoted by
ðaÞ, is equal to RaR þ Ra þ aR þ Za.
We have thus described the (left, right, two-sided) ideals generated by one
element. Replacing one element by an arbitrary subset S of R does not make the
task of description much harder. For example, the ideal generated by S is the
additive subgroup of R generated by the set of elements of the form xsy, xs, sy, s
with x; y 2 R and s 2 S.
The sum ofSan arbitrary family of ideals fIi j i 2 Ig is defined as the ideal
generated by i2I Ii . It is easy to see that it is actually equal to the additive
S
subgroup generated by i2I Ii . The same is true for left or right ideals.
Let I be an ideal of a ring R. As I is, in particular, an additive subgroup of R, we
can form the factor additive group R=I. Recall that R=I consists of all cosets x þ I,
x 2 R, which are added according to
ðx þ IÞ þ ðy þ IÞ ¼ ðx þ yÞ þ I:
We can make R=I into a ring by defining the multiplication by
ðx þ IÞðy þ IÞ :¼ xy þ I:
It is quite obvious that the ring axioms are fulfilled. The main issue is to show that
this multiplication is well-defined. However, using RI I and IR I this is also
easy. We call R=I the factor ring of R by I (or the quotient ring, but we will avoid
this term to prevent confusion with the notion of a ring of quotients). Note that the
canonical homomorphism x 7! x þ I, mentioned in the additive group setting, is
now a ring homomorphism.
A simple example: the ring Z=nZ is isomorphic to Zn .
Let ϕ : R ! R0 be a ring homomorphism. Note that ker ϕ is an ideal of R, while
im ϕ is in general only a subring of R0 . Just as in the group-theoretic setting we
have
R= ker ϕ ffi im ϕ

with an isomorphism defined by x þ ker ϕ 7! ϕðxÞ.

Given rings R1 ; . . .; Rn , we define their direct product R1      Rn as the
Cartesian product endowed with the componentwise operations:

ðx1 ; . . .; xn Þ þ ðy1 ; . . .; yn Þ ¼ ðx1 þ y1 ; . . .; xn þ yn Þ;

ðx1 ; . . .; xn Þðy1 ; . . .; yn Þ ¼ ðx1 y1 ; . . .; xn yn Þ:

Clearly, R1      Rn is a ring. We may view Ri as an ideal of R1      Rn via

the canonical embedding xi 7! ð0; . . .; 0; xi ; 0; . . .; 0Þ.
Prerequisites xxix

One analogously defines the direct product Π i2I Ri of a possibly infinite family
of rings fRi j i 2 Ig. The operations are thus given as follows:

ðxi Þ þ ðyi Þ ¼ ðxi þ yi Þ;

ðxi Þðyi Þ ¼ ðxi yi Þ:

Incidentally, the subset of Π i2I Ri consisting of all ðxi Þ such that all but finitely
many xi are 0 is a subring of Π i2I Ri , called the direct sum of the family
fRi j i 2 Ig. The notions ‘‘direct product’’ and ‘‘direct sum’’ thus coincide for finite
families of rings. It is therefore common to use the notation R1      Rn instead
of R1      Rn . However, we will stick with the latter.

Matrices

We have already mentioned that the set of all real n  n matrices, Mn ðRÞ, is a ring
under the usual matrix operations. This ring is not commutative as long as n  2,
and has plenty of zero-divisors. One can replace the role of R in the definition of
addition and multiplication not only by another field, but by any other ring. This
does not affect the proof that the ring axioms are fulfilled.
Let us be precise. Take an arbitrary ring R and n 2 N. Given any aij 2 R,
1  i; j  n, we denote by ðaij Þ the n  n matrix
2 3
a11 a12 . . . a1n
6 a21 a22 . . . a2n 7
6 7
6 .. .. .. .. 7:
4 . . . . 5
an1 an2 ... ann

Define addition and multiplication of such matrices by

ðaij Þ þ ðbij Þ :¼ ðaij þ bij Þ

and
Xn
ðaij Þðbij Þ :¼ ðcij Þ; where cij ¼ k¼1
aik bkj .

One readily checks that in this way the set of all n  n matrices with entries in R
becomes a ring. It will be denoted by Mn ðRÞ: Note that M1 ðRÞ can be identified
with R. If R is unital, then so is Mn ðRÞ. Its unity is the identity matrix In , i.e., the
matrix whose entries on the main diagonal are 1 and all other entries are 0.
Sometimes we call Mn ðRÞ the full matrix ring over R, in order to point out that
it contains all n  n matrices with entries in R. The term matrix ring is often used
as a synonym for the full matrix ring, but it can also mean any ring whose elements
are matrices.
xxx Prerequisites

A special but extremely important case is when R ¼ F is a field. When working

with Mn ðFÞ we often rely on linear algebra methods. We assume the reader is
familiar with the notions such as the eigenvalue, determinant, trace, and
characteristic polynomial of a matrix.

Polynomials

Let R be an arbitrary ring R. We define a polynomial over R as a formal infinite

sum
X
1
ai ωi ¼ a0 þ a1 ω þ a2 ω2 þ . . .;
i¼0

where ai 2 R and all but finitely many ai are 0. The elements ai are called coef-
ficients, and ω is a formal symbol called an indeterminate. The coefficient a0 is
called the constant term. If n  0 is such that ai ¼ 0 for all i [ n and an 6¼ 0, then
we usually write the above polynomial as
a0 þ a 1 ω þ    þ a n ω n
(later we will avoid this lengthy notation and write a polynomial as f ðωÞ or simply f ).
This polynomial is said to have degree n, and an is called its leading coefficient.
If ai ¼ 0 for all i  0, then the corresponding polynomial is denoted by 0. A poly-
nomial that is either 0 or has degree 0 is called a constant polynomial.
The addition and multiplication of two polynomials are defined by
X
1 X
1 X
1
an ωn þ bn ωn :¼ ðan þ bn Þωn ;
n¼0 n¼0 n¼0
! !
X
1 X
1 X
1 X
n
an ωn bn ω n :¼ cn ωn ; where cn ¼ ai bni :
n¼0 n¼0 n¼0 i¼0

This makes the set of all polynomials over R into a ring, which we denote by R½ω
and call a polynomial ring. If R is unital, then so is R½ω. Elements in R commute
with ω, but they do not necessarily commute with each other. In fact, R is com-
mutative if and only if R½ω is commutative. We are primarily interested in the
situation where R is a field. In this case we will write R as F and thus denote the
polynomial ring as F½ω.
So far we have considered polynomials in one indeterminate. The notion of a
polynomial in several indeterminates ω1 ; . . .; ωn is not conceptually much more
demanding, but notationally heavier. It can be defined as an infinite sum
Prerequisites xxxi

X
1 X
1
 ai1 ;...;in ωi11 . . .ωinn ;
i1 ¼0 in ¼0

where ai1 ;...;in 2 R and all but finitely many ai1 ;...;in are 0. Each term ai1 ;...;in ωi11 . . .ωinn
whose coefficient ai1 ;...;in is not 0 is called a monomial. Of course, we may neglect
the terms with zero coefficient, and so every polynomial in several indeterminates
is a finite sum of monomials. The expressions ωi11 . . .ωinn basically play the role of
components. The addition is defined componentwise. The multiplication is
determined by distributivity and the rule

ðai1 ;...;in ωi11 . . .ωinn Þðbj1 ;...; jn ω1j1 . . .ωnjn Þ ¼ ðai1 ;...;in bj1 ;...; jn Þωi11 þj1 . . .ωinn þjn :

These two operations make the set of all polynomials in ω1 ; . . .; ωn into a ring
which we denote by R½ω1 ; . . .; ωn : Incidentally, alternatively one can define
R½ω1 ; . . .; ωn  recursively as R½ω1 ; . . .; ωn1 ½ωn .
One can also start with a possibly infinite set of indeterminates Ω ¼ fωi j i 2 Ig,
and introduce the polynomial ring R½Ω. Each polynomial in R½Ω involves only
finitely many indeterminates, so the above definitions regarding R½ω1 ; . . .; ωn  still
make sense. Only the notation is even more entangled. Let us point out that each
ωi 2 Ω is a central element of R½Ω.

Vector Spaces

A vector space over a field F is an additive group V endowed with an additional

operation F  V ! V, ðλ; vÞ 7! λv, called the scalar multiplication, which
satisfies
(a) ðλ þ μÞv ¼ λv þ μv,
(b) λðv þ wÞ ¼ λv þ λw,
(c) λðμvÞ ¼ ðλμÞv,
(d) 1v ¼ v
for all λ; μ 2 F, v; w 2 V. Note that λv ¼ 0 if and only if λ ¼ 0 or v ¼ 0, and
ðλÞv ¼ λv. Elements in V are called vectors and elements in F are called
scalars. Instead of ‘‘vector space’’ we often say ‘‘linear space’’ or just ‘‘space’’.
A vector space over R is called a real (vector) space, and a vector space over C is
called a complex (vector) space.
The prototype example of a vector space is F n , the Cartesian product of n
copies of F, endowed with componentwise operations

ðλ1 ; . . .; λn Þ þ ðμ1 ; . . .; μn Þ ¼ ðλ1 þ μ1 ; . . .; λn þ μn Þ;

λðλ1 ; . . .; λn Þ ¼ ðλλ1 ; . . .; λλn Þ:
xxxii Prerequisites

In particular, F ¼ F 1 is a vector space. We have introduced F½ω and C½a; b as

rings, but they are also vector spaces if we define scalar multiplication in the usual
way.
A nonempty subset W of a vector space V over F is a (linear) subspace if
λw þ λ0 w0 2 W for all λ; λ0 2 F and w; w0 2 W. A subspace is itself a vector space.
For example, the set of all polynomials in F½ω of degree at most n is a subspace of
F½ω.
A linear combination of vectors v1 ; . . .; vn is every vector of the form λ1 v1 þ
   þ λn vn where λi 2 F. If S is a nonempty subset of a space V, then the subspace
VS generated by S is equal to the set of all linear combinations of vectors in S. It is
called the linear span of S; we also say that VS is spanned by S or that S spans VS .
Note that, by definition, a linear combination involves a finite number of vectors. If
S is an infinite set, then the linear span of S is the union of the linear spans of all
finite subsets of S.
Vector space homomorphisms are linear maps, also called linear operators.
These are maps ϕ : V ! V 0 that satisfy
ϕðλv þ μwÞ ¼ λϕðvÞ þ μϕðwÞ
for all λ; μ 2 F, v; w 2 V. The notions such as ‘‘kernel’’, ‘‘image’’, ‘‘endomor-
phism’’, ‘‘isomorphism’’, etc. should be self-explanatory. Sometimes we will
denote linear operators with capital letters (A, B, etc.). A linear map from a vector
space over F into the space F is called a linear functional. A map
f : V1  V2 ! V 0 , where V1 ; V2 ; V 0 are vector spaces over F, is said to be bilinear
if it is linear in each variable, i.e.,
f ðλu þ μv; zÞ ¼ λf ðu; zÞ þ μf ðv; zÞ and f ðv; λz þ μwÞ ¼ λf ðv; zÞ þ μf ðv; wÞ

for all λ; μ 2 F, u; v 2 V1 , z; w 2 V2 .
A subset S of a vector space V over F is said to be linearly independent if for
all distinct vectors s1 ; . . .; sn 2 S and all scalars λ1 ; . . .; λn 2 F, λ1 s1 þ    þ
λn sn ¼ 0 implies λ1 ¼    ¼ λn ¼ 0. If S is not linearly independent, then we say
that S is linearly dependent. If S ¼ fsi j i 2 Ig, then instead of saying that S is
linearly (in)dependent we can also say that the vectors si , i 2 I, are linearly
(in)dependent.
A linearly independent subset of V that spans V is called a basis of V. For
example, fð1; 0Þ; ð1; 1Þg is a basis of F 2 . The standard basis of F 2 is fð1; 0Þ; ð0; 1Þg,
and the standard basis of F½ω is f1; ω; ω2 ; . . .g. The basis of f0g is ;.
A vector space V over F is finite dimensional if it is finitely generated, i.e., it is
spanned by a finite set. Such a space has a finite basis. Moreover, all its bases have
the same cardinality, called the dimension of V and denoted dimF V. (The concept
of dimension also makes sense for infinite dimensional spaces, but we will not use
it for them.) If V is n-dimensional, i.e., dimF V ¼ n, then every linearly
independent set of n vectors in V is a basis of V, every subset of V having
Prerequisites xxxiii

more than n vectors is linearly dependent, and the dimension of every proper
subspace of V is less than n. An example of such a space is F n , and in fact every
space V with dimF V ¼ n is isomorphic to F n .
We say that a linear operator A from one space to another has finite rank if its
image is a finite dimensional space. In this case we call the dimension of the image
the rank of A.
Every vector space, finite or infinite dimensional, has a basis. We will actually
prove this in Sect. 3.3, but in a more general framework.

Algebras

A vector space A over a field F is an algebra over F, or an F-algebra, if it is

endowed with associative multiplication A  A ! A, ðx; yÞ 7! xy, satisfying
ðλx þ μyÞz ¼ λðxzÞ þ μðyzÞ and xðλy þ μzÞ ¼ λðxyÞ þ μðxzÞ

for all λ; μ 2 F, x; y; z 2 A. We often say only ‘‘algebra’’ without specifying the

associated field. An R-algebra is also called a real algebra, and a C-algebra is
called a complex algebra. There is also the notion of an algebra over a com-
mutative unital ring. The axioms are the same, just the role of a field is replaced by
a commutative unital ring. Rings can be then viewed as algebras over Z, which
conveniently puts rings and algebras under one roof. However, we will consider
only algebras over fields in this book.
Thus, an algebra is a ring which is also a vector space, and multiplication is not
only biadditive but bilinear. Many rings mentioned earlier are actually algebras.
For example, F½ω is an F-algebra and C½a; b is a real algebra. If A is an algebra,
then the ring Mn ðAÞ is an algebra if we define scalar multiplication by
λðaij Þ :¼ ðλaij Þ:

Similarly, the ring A½ω becomes an algebra by defining

X
1  X
1
λ ai ωi :¼ ðλai Þωi :
i¼0 i¼0

We say that A is a finite dimensional algebra if it is finite dimensional as a vector

space. The dimension of A over F will be denoted by
½A : F:
Thus, ½A : F is nothing but dimF A, but it will be used only for algebras.
Algebras are, in particular, rings, so all ring-theoretic notions make sense for
algebras. However, some need small adjustments. A subalgebra is a subring
which is also a subspace. Thus, for example, the real algebra R has no proper
xxxiv Prerequisites

nonzero subalgebras, but the ring R has plenty of subrings. Similarly, algebra
ideals, one-sided or two-sided, are ideals in the ring-theoretic sense which are also
subspaces. If I is an ideal of an F-algebra A, then the factor ring A=I becomes an
F-algebra, called the factor algebra of A by I, if we define scalar multiplication by
λðx þ IÞ :¼ λx þ I:
The direct product of algebras is an algebra with operations defined compo-
nentwise. An algebra homomorphism is a ring homomorphism which is also a
linear map. For example,

the

aforementioned isomorphism between the ring of
matrices of the form β α , α; β 2 R, and the ring C is actually an R-algebra
α β

isomorphism.
The center of an algebra is defined just as the center of a ring. A unital
algebra is an algebra which is unital as a ring, and a division algebra is an algebra
which is a division ring. We define the characteristic of an F-algebra A, charðAÞ,
as the characteristic of A considered as a ring. The axioms readily imply that
charðAÞ coincides with the characteristic of the associated field F as long as A is
nonzero; therefore charðAÞ is either 0 or prime. Hence Zn is not an algebra over
any field if n is not prime (Zp , with p prime, is a field and therefore an algebra over
itself). It is also easy to see that Z is not an algebra over any field.
An F-algebra is, in particular, a vector space. Let fei j i 2 Ig be its basis. For
each pair i; j 2 I there exist αijk 2 F, k 2 I, such that
X
e i ej ¼ αijk ek
k2I

and all but finitely many αijk ’s are 0. We call these formulas the multiplication
table (if A has a small dimension, then one can really write this down in the form
of a table). Two F-algebras A and B are isomorphic, A ffi B, if and only if they
have the same multiplication tables with respect to some bases. By this we mean
that there exist a basis fei jP
i 2 Ig of A, a basis f fiP
j i 2 Ig of B, and scalars αijk 2 F,
i; j; k 2 I, such that ei ej ¼ k2I αijk ek and fi fj ¼ k2I αijk fk for all i; j 2 I. Indeed,
if this holds then the linear map that sends ei into fi is easily seen to be an algebra
isomorphism (note that such a map exists and is unique). Conversely, an algebra
isomorphism from A onto B clearly maps a basis of A onto a basis of B and
preserves the multiplication table. Thus, if ½A : F ¼ n, then the multiplication in A
is determined by n3 scalars αijk . Let us mention that not every choice of the αijk ’s is
appropriate since the associativity condition must be fulfilled.
Let A be a nonzero unital algebra. We may consider F as a 1-dimensional
subalgebra of A via the canonical embedding λ 7! λ1. Accordingly, the element
λ1 2 A will be often written simply as λ. Given a 2 A and a polynomial
f ðωÞ ¼ λ0 þ λ1 ω þ    þ λn ωn 2 F½ω;
Prerequisites xxxv

we define the evaluation of f at a as

f ðaÞ :¼ λ0 þ λ1 a þ    þ λn an 2 A:
We remark that f 7! f ðaÞ is a homomorphism from F½ω into A. If f has constant
term 0, then we can define f ðaÞ also when A is not unital.
We conclude this section with a particularly important example of an algebra.
Let V be a vector space over F. Then EndF ðVÞ; the set of all endomorphisms of V
(i.e., linear operators from V into V) becomes an F-algebra if we define addition,
scalar multiplication, and multiplication by

ðA þ BÞðvÞ :¼ AðvÞ þ BðvÞ;

ðλAÞðvÞ :¼ λAðvÞ;
ðABÞðvÞ :¼ AðBðvÞÞ:

The verification is straightforward. Suppose that dimF V ¼ n\1. Pick a basis

fe1 ; . . .; en g of V. For every A 2 EndF ðVÞ there exist scalars aij 2 F such that
X
n
Aðej Þ ¼ aij ei ; j ¼ 1; . . .; n:
i¼1

A straightforward verification shows that the map A 7! ðaij Þ is an algebra iso-

morphism from EndF ðVÞ onto the matrix algebra Mn ðFÞ (incidentally, in Sect. 3.4
we will prove a generalization of this fact). Thus,
EndF ðVÞ ffi Mn ðFÞ if dimF V ¼ n.

In particular, ½EndF ðVÞ : F ¼ ðdimF VÞ2 .

Fields

We need just a little of field theory in this book. Therefore we will provide the
reader only with basic information, but not even all of that will be used.
We start with a field construction which is usually considered as a part of
commutative ring theory rather than field theory. A commutative ring without
zero-divisors is called a commutative domain (perhaps the reader is familiar with
a more common term ‘‘integral domain’’, but usually one assumes that integral
domains must be unital, which we do not need right now). Typical examples are
the ring of integers Z and the polynomial ring F½ω where F is a field. Of course,
fields are also commutative domains, but they are too ‘‘perfect’’ to be listed among
typical examples.
The field Q is constructed from the commutative domain Z. Essentially the
same construction works for making a field from any nonzero commutative
domain Z. Let us sketch it (for more details the reader can look at Sect. 7.1 where
xxxvi Prerequisites

this construction is presented  in a slightly more general context). Define the

relation on the set Z  Z n f0g by ðr; sÞ ðr0 ; s0 Þ if rs0 ¼ r 0 s. It is easy to
check that this is an equivalence relation. Denote by rs1 the equivalence class of
ðr; sÞ. Thus, rs1 ¼ r 0 s01 if and only if rs0 ¼ r 0 s. Define addition and multipli-
cation on the set Zb of all equivalence classes as follows:

rs1 þ tu1 :¼ ðru þ tsÞðsuÞ1 ;
rs1  tu1 :¼ rtðsuÞ1 :

These operations are well-defined, and Z b is a field. It is called the field of quo-
tients of Z. One can embed Z into Z b and accordingly consider Z as a subring of Z b.
b
We remark that there does not exist a subfield E of Z such that Z E ( Z . If b
Z ¼ Z, then Z b ¼ Q. If Z ¼ F½ω, then Z b is called the field of rational functions in
ω and is denoted by FðωÞ. More generally, if Ω is an arbitrary set of indetermi-
nates, then the field of quotients of F½Ω is called the field of rational functions in
Ω and denoted by FðΩÞ.
We proceed to the genuine field theory. A field K is said to be an extension
field of a field F if F is a subfield of K. For instance, R is an extension field of Q,
and C is an extension field of R. Given a field F, one usually wants to construct
extension fields satisfying certain requirements. The original field F is often
contained in all other fields that appear in the study. In this case we call it the base
field.
If K is an extension field of F, then we may view K as an algebra over F.
Indeed, K is a ring by definition, and the scalar multiplication should be simply
understood as the given multiplication in K restricted to the situation where the
first factor lies in F—if λ 2 F and k 2 K then λk 2 K and all algebra axioms are
satisfied. We say that K is a finite extension of F if ½K : F\1. For example, C is
a finite extension of R with ½C : R ¼ 2, while R is not a finite extension of Q. If K
is a finite extension of F and L is a finite extension of K, then L is a finite extension
of F and ½L : F ¼ ½L : K½K : F.
The main point is that K can be viewed as a vector space over F. Let us give a
simple illustration of the usefulness of this viewpoint. Let K be a finite field. Then
charðKÞ is a prime number p. Hence K can be considered as a finite dimensional
space over the prime subfield F0 ffi Zp , and as such is isomorphic to F0n for some
n 2 N. But then jKj ¼ jF0n j ¼ pn . We have thereby found out that the cardinality of
a finite field can only be pn with p prime and n 2 N. This assertion has a sort of a
converse, which, however, is less obvious: For every prime p and n 2 N there
exists an (up to isomorphism) unique field having pn elements.
An element a in an extension field K of F is said to be a root of the polynomial
pffiffiffi
f ðωÞ 2 F½ω if f ðaÞ ¼ 0. For example, 2 2 R is a root of ω2  2 2 Q½ω. A root of
a polynomial of the form ωn  1, n 2 N, is called a root of unity. We say that a 2 K
is algebraic (over F) if a is a root of some nonzero polynomial in F½ω. More
specifically, a is algebraic of degree m if it is a root of a polynomial in F½ω of
Prerequisites xxxvii

degree m, but is not a root of a nonzero polynomial in F½ω of degree less than m.
In this case the subfield of K generated by a and F is equal to F½a :¼
fλ0 þ λ1 a þ    þ λm1 am1 j λ0 ; . . .; λm1 2 Fg, and ½F½a : F ¼ m. For example,
pffiffiffi pffiffiffi
Q½ 2 ¼ fλ0 þ λ1 2 j λ0 ; λ1 2 Qg.
If a 2 K is a root of a polynomial f ðωÞ 2 F½ω of degree n, then there exists a
polynomial qðωÞ 2 K½ω of degree n  1 such that f ðωÞ ¼ ðω  aÞqðωÞ. From this
one easily infers that f ðωÞ cannot have more than n roots in K. If a1 ; . . .; an are
distinct roots of f ðωÞ in K, then f ðωÞ ¼ cðω  a1 Þ. . .ðω  an Þ where c 2 F is the
leading coefficient of f ðωÞ.
A field K is said to be algebraically closed if every nonconstant polynomial in
K½ω has a root in K. As is evident from the first assertion in the previous
paragraph, this is equivalent to the condition that every polynomial f ðωÞ 2 K½ω of
degree n  1 splits into linear factors: f ðωÞ ¼ cðω  a1 Þ. . .ðω  an Þ where c; ai 2
K (some of the ai ’s may be equal). The Fundamental Theorem of Algebra states
that the field C is algebraically closed. The polynomial ω2 þ 1 does not have a root
in R, so R is not algebraically closed. However, from the fundamental theorem of
algebra it can be easily deduced that every polynomial in R½ω splits into linear
factors a þ bω and quadratic factors c þ dω þ eω2 , where a; b; c; d; e 2 R.
Working with algebraically closed fields is often essentially easier than with
general fields. Let us only point out one matter that is of importance in
noncommutative algebra. Recall that square matrices A and B are similar if there
exists an invertible matrix T such that A ¼ TBT 1 . If K is an algebraically closed
field, then every matrix A 2 Mn ðKÞ is similar to P a block diagonal matrix J 2
Mn ðKÞ whose blocks are matrices in Mki ðKÞ, n ¼ i ki , having a fixed scalar λi on
the main diagonal, the scalar 1 on the superdiagonal, and zeros elsewhere. The
matrix J is called the Jordan normal form of A. The scalars λi appearing in J are
the eigenvalues of A.
A field extension F of a field F is said to be an algebraic closure of F if F is
algebraically closed and every element in F is algebraic over F. For example, C is
an algebraic closure of R. Every field F has an algebraic closure, which is,
moreover, unique up to isomorphism. This is one of the deeper results stated in this
preparatory chapter.
Chapter 1
Finite Dimensional Division Algebras

Finite dimensional division algebras are as simple and “spotless” as a (not necessarily
commutative) ring can be. Their definition gathers together all of the most favorable
properties. They are also the oldest, the most classical topic of noncommutative
algebra. But as it often happens, the simplest objects are not the easiest to study, and
classical themes are not less profound than modern ones. In many algebra textbooks
finite dimensional division algebras are treated in the last chapters by using advanced
tools. Our goals in this first chapter, however, are relatively modest. We will give
an elementary, self-contained introduction to finite dimensional algebras, which in
particular includes two results of extreme importance and beauty: Frobenius’ theorem
on real division algebras and Wedderburn’s theorem on finite division rings. We shall
not strictly confine ourselves to division algebras; presenting some of the results in
the context of (central) simple algebras, even not necessarily finite dimensional ones,
will not change the level of complexity of arguments and shall prove useful later.

1.1 After the Complex Numbers: What Comes Next?

A 2-dimensional real vector space can be transformed into the complex numbers
by appropriately defining multiplication. One just takes two linearly independent
vectors, denotes one of them by 1 to indicate that it plays the role of a unity, requires
that the square of the other one is −1, and then extends this multiplication to the
whole space by bilinearity. If one is interested in the geometric interpretation of this
multiplication, then one will choose these two vectors more carefully. However, from
the algebraic point of view it does not matter which two vectors we take, as long as
they are linearly independent and therefore form a basis of our space. In any case we
get a 2-dimensional R-algebra isomorphic to the complex number field C.
Why confine ourselves to 2-dimensional spaces? Are there other ways to create
some kind of “numbers” from finite dimensional real vector spaces? In any set of
“numbers” one should be able to add, subtract, and multiply elements, and mul-
tiplication should be associative and bilinear. Thus, our space should in particular
© Springer International Publishing Switzerland 2014 1
M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_1
2 1 Finite Dimensional Division Algebras

be an R-algebra (in the early literature elements from algebras were actually called
hypercomplex numbers). In a decent set of “numbers” one should also be able to
divide elements, with the obvious exclusion of dividing by 0. We can now rephrase
our question in a rigorous manner as follows: Is it possible to define multiplication
on an n-dimensional real space so that it becomes a real division algebra?
For n = 1 the question is trivial; every element is a scalar multiple of unity and
therefore up to isomorphism R itself is the only such algebra. For n = 2 we know
one example, C, but are there any other? This question is quite easy and the reader
may try to solve it immediately. What about n = 3? This is already a nontrivial
question, but certainly a natural and challenging one. After finding a multiplication
in 2-dimensional spaces of such enormous importance in mathematics, should not
the next step be finding something similar in 3-dimensional spaces? Furthermore,
what about n = 4, 5, 6, . . . ?
We will give complete answers to the above questions. Throughout the section,
we assume that D is an n-dimensional division R-algebra. We will find all possible n
for which such an algebra exists, and moreover, all possible multiplications making
a finite dimensional real space into a division algebra.
First we recall the notational convention. For λ ∈ R, we write λ1 ∈ D simply as
λ. In fact we identify R with R1, and in this way consider R as a subalgebra of D.

Lemma 1.1 For every x ∈ D there exists λ ∈ R such that x 2 + λx ∈ R.

Proof Since the dimension of D is n, the elements 1, x, . . . , x n are linearly dependent.

This means that there exists a nonzero polynomial f (ω) ∈ R[ω] of degree at most
n such that f (x) = 0. We may assume that the leading coefficient of f (ω) is equal
to 1. As we know, f (ω) splits into linear and quadratic factors in R[ω]:

f (ω) = (ω − α1 ) . . . (ω − αr )(ω2 + λ1 ω + μ1 ) . . . (ω2 + λs ω + μs )

where αi , λi , μi ∈ R. Since f (x) = 0, we have

(x − α1 ) . . . (x − αr )(x 2 + λ1 x + μ1 ) . . . (x 2 + λs x + μs ) = 0.

As D is a division algebra, one of the factors must be 0. That is, x is a root of a linear
or quadratic polynomial in R[ω]. In any case the desired conclusion holds. 

The 1-dimensional subspace Ri of C can be described as the set of all z ∈ C such

that z2 is a nonpositive real number. The analysis of the set of all such elements in
our abstract setting will lead us to the answers to our questions.

Lemma 1.2 The set V = {v ∈ D | v2 ∈ R, v2 ≤ 0} is a linear subspace of D.

Moreover, D = R ⊕ V .

Proof We start with a useful observation. If x ∈ D\V is such that x 2 ∈ R, then x 2 > 0
and so x 2 = α 2 for some α ∈ R. Thus (x − α)(x + α) = 0, giving x = ±α ∈ R.
1.1 After the Complex Numbers: What Comes Next? 3

It is clear that R ∩ V = 0 and that V is closed under scalar multiplication. Let us

show that u + v ∈ V if u, v ∈ V . We may assume that u and v are linearly indepen-
dent. We claim that u, v, 1 are also independent. Indeed, taking α, β, γ ∈ R such that
αu = βv + γ it follows by squaring both sides that βγ v ∈ R, hence β = 0 or γ = 0,
and α = β = γ = 0 now readily follows. By Lemma 1.1 there exist λ, μ ∈ R such
that

(u + v)2 + λ(u + v) ∈ R and (u − v)2 + μ(u − v) ∈ R.

On the other hand,

(u + v)2 + (u − v)2 = 2u2 + 2v2 ∈ R.

Comparing all the relations we obtain λ(u + v) + μ(u − v) ∈ R. Since u, v, 1 are

linearly independent, this yields λ+μ = λ−μ = 0, hence λ = μ = 0, and u+v ∈ V
follows from the first paragraph. Thus, V is a subspace of D. If x ∈ D\R, then, by
Lemma 1.1, x 2 + vx ∈ R for some v ∈ R. Again using the observation from the first
paragraph we see that x + 2v ∈ V . Accordingly, x = − 2v + (x + 2v ) ∈ R ⊕ V . 
We set u ◦ v := uv + vu for u, v ∈ V . Since u ◦ v = (u + v)2 − u2 − v2 it follows
from Lemma 1.2 that u ◦ v ∈ R. If v = 0, then v ◦ v = 2v2 = 0.
Lemma 1.3 If n > 2, then there exist i, j, k ∈ D such that

i2 = j2 = k 2 = −1,
(1.1)
ij = −ji = k, ki = −ik = j, jk = −kj = i,

and 1, i, j, k are linearly independent.

Proof Lemma 1.2 shows that V has dimension n − 1 > 1. Therefore we may choose
linearly independent v, w ∈ V . Let u := w − w◦v
v◦v v. Observe that u  = 0 and u ◦ v = 0.
Set i := √ 1
u, j := √ 1
v, and k := ij. It is straightforward to check that (1.1)
−u2 −v2
holds, from which we obtain that

(α0 + α1 i + α2 j + α3 k)(α0 − α1 i − α2 j − α3 k) = α02 + α12 + α22 + α32 (1.2)

is a positive real number for any α0 , α1 , α2 , α3 ∈ R that are not all 0. This proves,
in particular, that 1, i, j, k are linearly independent. 
The second formula in (1.1) can be easily remembered by the following picture:
4 1 Finite Dimensional Division Algebras

Going clockwise, the product of two consecutive elements is the third one; going
counterclockwise, we obtain the negative of the third one.
Lemma 1.3 rules out the case where n = 3. If n = 4 then D has a basis 1, i, j, k
with multiplication table (1.1). This is a real finding. Let us change the point of view.
Suppose we are given a 4-dimensional real vector space D with basis {1, i, j, k}
in which we define multiplication by (1.1) and the requirement that 1 is a unity.
A direct verification shows that D is indeed an R-algebra (note that it is enough
to check associativity for elements from a basis!). We define the conjugate of an
arbitrary element h = α0 + α1 i + α2 j + α3 k in D by
h := α0 − α1 i − α2 j − α3 k.

If h = 0, then hh (= hh) is a nonzero scalar by (1.2). Therefore h is invertible with

inverse 1 h. Accordingly, D is a division algebra.
hh
The standard notation for this 4-dimensional noncommutative real division alge-
bra is H, in honor of its discoverer, the Irish mathematician W. R. Hamilton. The
elements of H are called (real) quaternions. As the first noncommutative algebraic
system ever introduced, H is of great historic value. Before this discovery Hamilton
had been struggling for a long time to extend the multiplication of complex numbers
to the 3-dimensional space, that is, he had wanted to solve our problem for n = 3.
We have just seen that this had been hopeless. On October 16, 1843, when walking
with his wife along the Royal Canal in Dublin, the idea of 4-dimensional quaternions
suddenly came into his mind, and he carved the formula i2 = j2 = k 2 = ijk = −1
in a stone of Broom Bridge. This story may have a flavor of a romantic legend, but
is documented in Hamilton’s letters.
The following theorem was proved in 1878 by another great mathematician,
F. G. Frobenius.

Theorem 1.4 (Frobenius) A finite dimensional real division algebra D is isomorphic

to R, C, or H.

Proof If n = 1, then D ∼ = R. If n = 2, then V = 0 by Lemma 1.2, so D contains an

element i such that i2 = −1; but then D ∼ = C. By Lemma 1.3 we know that n = 3
and that D ∼
= H if n = 4.
Suppose n > 4. Let i, j, k be the elements from Lemma 1.3. Since the dimension
of V is n − 1, there exists v ∈ V not lying in the linear span of i, j, k. Therefore
j◦v
2 i + 2 j + 2 k is a nonzero element in V and it satisfies i ◦ e = j ◦ e =
e := v + i◦v k◦v

k ◦e = 0. However, from the first two identities we conclude eij = −iej = ije, which
contradicts the third identity since ij = k. Thus, n can only be 1, 2, or 4. 

1.2 Beyond Frobenius’ Theorem

Frobenius’ theorem is a classification theorem par excellence. Three important exam-

ples of finite dimensional real division algebras were discovered by mathematicians,
each of them has turned out to be immensely important, and, up to isomorphism,
1.2 Beyond Frobenius’ Theorem 5

these three are in fact the only ones that exist. Yet this definitive result has various
extensions. However, they are mostly concerned with nonassociative algebras, i.e.,
algebras in which multiplication is not necessarily associative. This goes beyond the
scope of this book. Still, at this point it seems appropriate to at least mention the
algebra of octonions O, the natural successor in the sequence R, C, H. This is an
8-dimensional nonassociative real algebra which contains a copy of H as its subal-
gebra. Its standard basis contains a unity 1 and seven elements whose squares equal
−1. Although not associative, O satisfies weak associativity conditions x 2 y = x(xy)
and yx 2 = (yx)x for all x, y ∈ O; nonassociative algebras satisfying these two condi-
tions are called alternative. Further, O is a nonassociative division algebra, which,
by definition, means that for all a, b ∈ O with a = 0, the equations ax = b and
ya = b have unique solutions in O. Let this be enough in our attempt to make the
reader curious and challenged for seeking further information about octonions else-
where. Two books that include an introduction to nonassociative algebras and can
be recommended are [McC04] and [ZSSS82].
Let us return to our theme. A close inspection of the arguments from the previous
section gives further insight. First, does the last paragraph of the proof of Theorem 1.4
resemble something familiar? Observe that (u, v)
→ − 21 u◦v defines an inner product
on V such that i, j, k are orthonormal vectors, so we were performing nothing but
the Gram-Schmidt process when introducing e. Moreover, this is not just any inner
product. Identifying each u = α1 i + α2 j + α3 k ∈ V with u = (α1 , α2 , α3 ) ∈ R3 , we
have − 21 u ◦ v = u · v where · is the dot product. Another interesting observation is
that 21 (uv−vu) = u×v where × is the cross product. Using uv = 21 u◦v+ 21 (uv−vu)
one now easily sees that H can be equivalently introduced as the vector space R × R3
(with pointwise addition and scalar multiplication) endowed with multiplication

(α0 , u)(β0 , v) = (α0 β0 − u · v, α0 v + β0 u + u × v).

Let us mention in passing that another standard way is to introduce quaternions as

certain 2 × 2 complex matrices. We will arrive at this in Example 2.72.
Our second remark concerning the previous section is that the finite dimensionality
of D was used only at one point, namely at the beginning of the proof of Lemma 1.1
when concluding that every x ∈ D is a root of a nonzero polynomial in R[ω]. This
observation yields a slightly different interpretation of the obtained result. Let us
first introduce some terminology. We say that an element x from an F-algebra A is
algebraic if there exists a nonzero polynomial f (ω) ∈ F[ω] such that f (x) = 0.
This is, of course, just the repetition of the definition given in Prerequisites for field
extensions, and just as before we define the degree of algebraicity. If every element
in A is algebraic, then A is said to be an algebraic algebra. The following simple,
fundamental fact was indirectly observed in the proof of Lemma 1.1.

Lemma 1.5 Every finite dimensional algebra is an algebraic algebra.

Proof The set of all powers of each element is linearly dependent. 

6 1 Finite Dimensional Division Algebras

Our proof of Frobenius’ theorem thus shows that an algebraic real division algebra
is isomorphic to R, C, or H, and hence it is finite dimensional. Let us mention that
infinite dimensional algebraic algebras (over any field) do exist, and will also be met
in this book (for the first time in Example 2.7).
After describing finite dimensional real division algebras, one may wonder what
about the complex ones? The answer is simple and comes easily. Instead of C we
can actually take any algebraically closed field.
Proposition 1.6 If D is a finite dimensional division algebra over an algebraically
closed field F, then D = F.
Proof Let x ∈ D. By Lemma 1.5 there exists a nonzero polynomial f (ω) ∈ F[ω]
such that f (x) = 0. We may assume that its leading coefficient equals 1. Since F
is algebraically closed, we have f (ω) = (ω − α1 ) . . . (ω − αr ) for some αi ∈ F.
Consequently, (x − α1 ) . . . (x − αr ) = 0. As D is a division algebra, this gives
x − αi = 0 for some i, and hence x = αi ∈ F. 
Of course, again we may substitute “algebraic” for “finite dimensional”.
This proposition may seem disappointing after encountering the quaternions. But
later we will see that it can be actually considered as “good news”.
Remark 1.7 The 4-dimensional algebra with basis {1, i, j, k} and multiplication
determined by (1.1) can be formed over any field F, not only over R. However,
if F = C, then this is not a division algebra. In fact, from (1.2) we see that it is
a division algebra if and only if for all αi ∈ F, α02 + α12 + α22 + α32 = 0 implies
α0 = α1 = α2 = α3 = 0. This condition is of course fulfilled in R and some other
fields (say, in Q), but not in algebraically closed fields. Still, one might wonder what
algebra we get if F = C. This is not a hard question, but we will answer it later, in
Example 2.72.

1.3 Simple Rings

Our next main goal is a theorem on finite division rings. In the next sections we shall
progress slowly towards its proof, making several digressions that will turn out to
be important later when dealing with more general rings and algebras. For a while
division algebras will mostly lie dormant in our exposition. This section is devoted
to the following notion.
Definition 1.8 A ring R is said to be simple if R2 = 0 and 0 and R are the only ideals
of R.
The condition that R2 = 0, i.e., xy = 0 for some x, y ∈ R, is needed to exclude patho-
logical cases. In nonzero unital rings it is automatically fulfilled. The condition about
ideals is, of course, more restrictive. Most of rings are not simple, yet some of the
most significant are. At any rate, simple rings form one of the most important classes
of rings.
1.3 Simple Rings 7

Example 1.9 Every division ring D is simple. Moreover, D does not even have one-
sided ideals different from 0 and D. This follows immediately from the elementary
observation that proper one-sided ideals cannot contain invertible elements.

There are many important simple rings that are not division rings. We will now
present the most fundamental example. First we have to introduce some notation and
terminology.
Consider the matrix ring Mn (S) where S is an arbitrary unital ring. Let Eij denote
the matrix whose (i, j) entry is 1 and all other entries are 0. We call Eij , 1 ≤ i, j ≤ n,
the standard matrix units. By aEij we denote the matrix whose (i, j) entry is a ∈ S
and other entries are 0. Every matrix in Mn (S) is a sum of such matrices. Note that
for every A = (aij ) ∈ Mn (S) we have

Eij AEkl = ajk Eil for all 1 ≤ i, j, k, l ≤ n. (1.3)

This formula considerably simplifies computations in Mn (S). Yet the reader might be
more persuaded in the correctness of the arguments based on (1.3) by occasionally
writing down matrices and then making the usual matrix calculations. Just do this
with 2 × 2 matrices, it is enough to feel certain.

Example 1.10 If D is a division ring, then Mn (D) is a simple ring for every n ∈ N.
Indeed, let I be a nonzero ideal of Mn (D). Pick a nonzero (aij ) ∈ I. Choose j and
k so that ajk = 0. From (1.3) we see that ajk Eil ∈ I for all i and l, and hence also
−1
(dajk )Eii · ajk Eil = dEil ∈ I for every d ∈ D. Consequently, I = Mn (D).

Remark 1.11 Rather than in simple rings, we are currently interested in simple
algebras, i.e., algebras satisfying the conditions of Definition 1.8. This may sound
a bit ambiguous. Namely, an ideal of an algebra A must be a vector subspace, while
Definition 1.8 speaks about ideals of a ring which are merely additive subgroups.
Fortunately, both possible interpretations lead to the same conclusion. Indeed, assume
that A is an algebra such that 0 and A are the only algebra ideals of A, and I is a ring
ideal of A such that I = 0 and I = A. Then AI is an algebra ideal of A, so it is either
0 or A. Since AI ⊆ I = A, the only possibility is that AI = 0. This implies that the
algebra ideal J = {x ∈ A | Ax = 0} is nonzero, so that J = A. That is, A2 = 0. This
possibility is excluded in Definition 1.8. Thus, an algebra A is simple as an algebra
if and only if it is simple as a ring.
Without assuming that A2 = 0 it is possible that A has many ring ideals, but no
algebra ideals different from 0 and A. For instance, take the 1-dimensional algebra
A = Ra with trivial multiplication, i.e., a2 = 0. Then clearly A has no algebra ideals
different from 0 and A, but has plenty of ring ideals (say, Za).

Our next goal is to give an example of an infinite dimensional simple algebra

which will accompany us throughout the book. First we introduce a useful notion
that naturally arises in noncommutative algebra.
8 1 Finite Dimensional Division Algebras

Definition 1.12 The commutator of elements a and b in a ring R is the element

[a, b] := ab − ba.

Obviously, a and b commute if and only if [a, b] = 0. We also remark that

[a, b] = −[b, a] and [a, bc] = [a, b]c + b[a, c].

Example 1.13 Let F be a field with char(F) = 0. Recall that End F (F[ω])
denotes the algebra of all linear operators of the vector space F[ω]. Define D, L ∈
End F (F[ω]) by    
D f (ω) = f (ω), L f (ω) = ω f (ω).

Here, f (ω) is the derivative of f (ω). The subalgebra of End F (F[ω]) generated by
D and L is called the Weyl algebra. We denote it by A1 . A more accurate name
is the first Weyl algebra. We will introduce the higher Weyl algebras An later, in
Example 6.4. The origin of these algebras is in quantum mechanics, but we will see
that they are also very interesting algebraic objects. Let us examine some properties
of A1 .
It is straightforward to check that

[D, L] = I, (1.4)

the identity operator. This is the basic relation in A1 . Everything what follows will
be deduced from it. In particular we now see that A1 is a unital algebra. Next, by
induction on n one easily shows that

[D, L n ] = [D, L n−1 ]L + L n−1 [D, L] = nL n−1 (1.5)

for every n ∈ N (here we follow the convention that T 0 = I for every operator T ).
Similarly we get

[Dn , L] = nDn−1 . (1.6)

Let L be the subalgebra of A1 generated by I and L. From (1.5) we see that

[D, L ] ⊆ L ; in particular,

DL ⊆ L + L D. (1.7)

Using [Dm , T ] = D[Dm−1 , T ] + [D, T ]Dm−1 we obtain inductively that

[Dm , L ] ⊆ L + L D + · · · + L Dm−1 for every m ∈ N. (1.8)

This readily implies that the linear span of {TDm | T ∈ L , m ≥ 0} is a subalgebra

of A1 , and hence, since it contains L and D, it is equal to A1 . That is, we have
shown that
1.3 Simple Rings 9

(a) Every element in A1 can be written as T0 + T1 D + · · · + Tn Dn , where n ≥ 0

and T0 , T1 , . . . , Tn ∈ L .
We claim that the elements Ti are uniquely determined. Thus, we have to show that
T0 + T1 D + · · · + Tn Dn = 0 implies Ti = 0 for each i. And this is easy. Considering
the action of the left-hand side operator on 1 we get T0 (1) = 0, and hence T0 = 0.
Similarly, by considering the action on ω we get T1 = 0. We continue with ω2 ,
ω3 , etc. Note that the assumption that char(F) = 0 is used when deriving T2 = 0,
T3 = 0, etc. Since {L m | m ≥ 0} is obviously a basis of L , we have thereby proved
that
(b) The set {L m Dn | m, n ≥ 0} is a basis of A1 .
(Analogously one shows that so is the set {Dm L n | m, n ≥ 0}.)
Let us finally prove that
(c) A1 is a simple algebra.

Let I be a nonzero ideal of A1 . Take S = nk=0 Tk Dk ∈ I , Tk ∈ L , such that
Tn = 0 and n is minimal. Suppose n > 0. Since L commutes with each Ti , it follows,
by making use of (1.6), that


n 
n
[S, L] = Tk [Dk , L] = kTk Dk−1 . (1.9)
k=0 k=1

However, since nTn = 0 and [S, L] ∈ I , this

 contradicts the minimality of n. Thus
n = 0, i.e., I ∩ L = 0. Now take T = m j=0 αj L ∈ I ∩ L , αj ∈ F, such that
j

αm = 0 and m is minimal. Suppose m > 0. Using (1.5) we obtain


m
[D, T ] = jαj L j−1 , (1.10)
j=1

contradicting the minimality of m. Therefore m = 0, i.e., I ∈ I , and so I = A1 .

This proves (c).

1.4 Central Algebras

The center of a unital algebra obviously contains scalar multiples of unity. In many
important examples of algebras these are also the only central elements.

Definition 1.14 A nonzero unital algebra is said to be central if scalar multiples of

unity are the only elements in its center.

Unless specified otherwise, by an algebra we shall mean an algebra over the field
F. If A is a nonzero unital algebra, then we identify F with F · 1, and write λ instead
10 1 Finite Dimensional Division Algebras

of λ1 for every scalar λ ∈ F. Thus, A is central if Z(A) = F. In other words, A is

central if it is an algebra over its center.

Lemma 1.15 Let A be a unital algebra and let n ∈ N. Then A is central if and only
if Mn (A) is central.

Proof If c ∈ Z(A), then the diagonal matrix with all diagonal entries equal to c lies in
Z(Mn (A)). Thus, A is central if Mn (A) is. Now let A be central. Pick C ∈ Z(Mn (A)).
Since C commutes with every Ekl , k = l, one easily shows that C is a diagonal
matrix with all diagonal entries equal. As C also commutes with every aE11 , a ∈ A,
we conclude that this diagonal entry lies in the center of A, and hence it is a scalar.
Thus Mn (A) is central. 

Example 1.16 As the simplest application of Lemma 1.15 we have that Mn (F) is a
central algebra.

Example 1.17 It is easy to verify that H is a central R-algebra.

Example 1.18 Obviously, C is central as a C-algebra, but not as an R-algebra.

Theorem 1.4 therefore implies

Corollary 1.19 A finite dimensional central division R-algebra is isomorphic to R

or H.
n
Example 1.20 The Weyl algebra A1 is central. Indeed, take S = k=0 Tk D ∈
k

Z(A1 ). As [S, L] = 0, it follows from (1.9) that Ti = 0 for i ≥ 1. Therefore

S= m j=0 αj L . Using [D, S] = 0 we see from (1.10) that αi = 0 for i ≥ 1. Thus
j

S = α0 ∈ F.

Example 1.21 If A is a simple unital ring, then its center is a field. Indeed, if c is a
nonzero central element, then cA must be, as a nonzero ideal of A, equal to A. This
implies that c is invertible. If we multiply cx = xc from the left and right by c−1
we see that c−1 is also a central element. Consequently, every simple unital ring
can be viewed as a central algebra over its center. Namely, we can define scalar
multiplication simply as the ordinary multiplication of a central element with an
arbitrary element from A, and all algebra axioms are readily fulfilled.

The main attention in the following sections will be devoted to finite dimensional
central simple algebras, which form an extremely important class of algebras. For
instance, the algebras Mn (F) and H belong to this class, as it is evident from the above
examples. Later we will see that every finite dimensional central simple algebra is
built from matrices and division algebras (Corollary 2.62). However, our immediate
goal is to derive some results on these algebras directly from the definition, without
relying on the description of their structure. For this purpose we will now introduce
a certain algebra of linear operators.
1.5 Multiplication Algebra 11

1.5 Multiplication Algebra

Let A be an algebra. For a, b ∈ A we define maps La , Rb : A → A, called the left

multiplication map and right multiplication map, by

La (x) := ax, Rb (x) := xb.

These maps will frequently play important roles in this book. Of course, they can be
defined if A is merely a ring, but we will usually treat them on algebras. In this case
they can be considered as elements of End F (A), the algebra of all linear operators
from A into itself.
Note that for all a, b ∈ A, λ, μ ∈ F we have

Lab = La Lb , Rab = Rb Ra ,
La Rb = Rb La ,
Lλa+μb = λLa + μLb , Rλa+μb = λRa + μRb .

Hence we conclude that

M(A) := {La1 Rb1 + · · · + Lan Rbn | ai , bi ∈ A, n ∈ N}

is a subalgebra of End F (A).

Definition 1.22 The algebra M(A) is called the multiplication algebra of A.
If A is unital, then La = La R1 and Rb = L1 Rb lie in M(A), and so in this case
we can describe M(A) as the subalgebra of End F (A) generated by all left and right
multiplication maps. We also remark that a = La (1) = Ra (1), and so the conditions
a = 0, La = 0, and Ra = 0 are equivalent if A is unital.

Remark 1.23 Take f ∈ M(A). Then there exist ai , bi ∈ A such that f = ni=1 Lai Rbi ,
i.e.,

n
f (x) = ai xbi , x ∈ A.
i=1

These elements are not unique, f can be presented as a sum of operators of the form
La Rb in many ways. If f = 0, then ai , bi can be chosen so that both the set of all
ai ’s and the set of all bi ’s are linearly independent. For example, this is fulfilled if
we require that n is minimal. Indeed, if, under this assumption, one of the elements,

say bn , was a linear combination of the others, bn = n−1 i=1 λi bi , then f would be
equal to


n−1
Lai +λi an Rbi ,
i=1
12 1 Finite Dimensional Division Algebras

contradicting the minimality of n. Furthermore, if {ui | i ∈ I} is a basis of A, then we

can express each Rb as a linear combination of Rui , from which we see that every
f ∈ M(A) is a (finite) sum of operators of the form Lai Rui . Similarly, f can be written
as a sum of operators of the form Lui Rbi , and, moreover, as a linear combination of
operators Lui Ruj .

The fact that the elements from M(A) can be expressed in different ways through
left and right multiplication maps may create some annoyance. The following lemma
shows that in central simple algebras this problem is controllable.

Lemma
n 1.24 Let A be a central simple algebra, and let ai , bi ∈ A be such that
i=1 Lai Rbi = 0. If the ai ’s are linearly independent, then each bi = 0. Similarly, if
the bi ’s are linearly independent, then each ai = 0.

Proof The two assertions of the lemma are analogous, so we consider only the case
Suppose bn = 0. Since A is simple, the ideal generated
where the ai ’s are independent.
by bn is equal to A. That is, m j=1 wj bn zj = 1 for some wj , zj ∈ A. Hence


m 
n  
n 
m  n
0= Rzj Lai Rbi Rwj = Lai Rwj bi zj = Lai Rci
j=1 i=1 i=1 j=1 i=1

m
where ci = j=1 wj bi zj ; thus, cn = 1. This clearly implies that n > 1. We
may assume that n is the smallest natural number for which the lemma does not
hold. Since


n  
n  n−1
0= Lai Rci Rx − Rx Lai Rci = Lai Rxci −ci x
i=1 i=1 i=1

for every x ∈ A, it thus follows that xci − ci x = 0. Consequently, ci ∈ F. But then


n
0= Lai Rci = Lc1 a1 +···+cn an ,
i=1

which contradicts the linear independence of the ai ’s. 

In the next lemma we consider the finite dimensional situation. Let us recall that
if [A : F] = d, then [End F (A) : F] = d 2 .

Lemma 1.25 If A is a finite dimensional central simple algebra, then M(A) =

End F (A).

Proof Let {u1 , . . . , ud } be a basis of A. Lemma 1.24 implies that the operators
Lui Ruj , 1 ≤ i, j ≤ d, are linearly independent. This becomes clear if we rewrite
d d d
i,j=1 λij Lui Ruj as i=1 Lui Rbi where bi = j=1 λij uj . Therefore
1.5 Multiplication Algebra 13

[M(A) : F] ≥ d 2 = [End F (A) : F],

and so M(A) = End F (A). 

It may be instructive for the reader to find an independent, direct proof of this
lemma for the special case where A = Mn (F) (what is a natural basis of End F (A)?).
Lemmas 1.24 and 1.25 will be used at different places throughout the book. In the
next sections we will see that three classical results can be deduced from them.

1.6 Automorphisms of Central Simple Algebras

The first result that we will derive from Lemmas 1.24 and 1.25 is an important special
case of a celebrated theorem established by T. Skolem in 1927 and somewhat later
rediscovered by E. Noether. We will establish the general Skolem-Noether Theorem
later, in Sect. 4.9, when having appropriate tools at our disposal. Treating a special
case at this early stage of development of the theory gives us an opportunity to
demonstrate a simple, but efficient method of proof.
In unital rings (and algebras) there is a canonical way of constructing automor-
phisms; every invertible element gives rise to one:

Definition 1.26 An automorphism ϕ of a unital ring R is said to be an inner auto-

morphism if there exists an invertible a ∈ R such that ϕ(x) = axa−1 , x ∈ R.
An automorphism that is not inner is called outer. Here are a few examples of
such automorphisms.

Example 1.27 The conjugation z

→ z is an automorphism of the R-algebra C. Of
course it is outer for the identity map is obviously the only inner automorphism of a
commutative ring. We also remark that it is an element of EndR (C), but not of M(C).
Therefore the assumption that A is central cannot be omitted in Lemma 1.25.

Example 1.28 For every α ∈ F\{0}, f (ω)

→ f (ω + α) is an outer automorphism
of the algebra F[ω].

Example 1.29 Let S be a ring and let R = S × S be the direct product of two copies
of S. Then (s, t)
→ (t, s) is an outer automorphism of R.

In light of these examples one can appreciate the next theorem.

Theorem 1.30 (Skolem-Noether) Every automorphism of a finite dimensional

central simple algebra A is inner.

Proof Let ϕ be an automorphism of A. Lemma 1.25 shows that ϕ = ni=1 Lai Rbi for
some ai , bi ∈ A. We may assume that a1 = 0 and the bi ’s are linearly independent
(cf. Remark 1.23). Since ϕ preserves multiplication, we have Lϕ(x) ϕ = ϕLx for every
x ∈ A. Note that this yields
14 1 Finite Dimensional Division Algebras


n
Lϕ(x)ai −ai x Rbi = 0.
i=1

Lemma 1.24 in particular implies that

ϕ(x)a1 − a1 x = 0, x ∈ A. (1.11)

It
remains to show that a1 is invertible. As A is simple, there are wj , zj ∈ A such that
m
j=1 j 1 zj = 1. In view of (1.11) this can be written as
w a


m  
m 
wj ϕ(zj ) a1 = 1, and a1 ϕ −1 (wj )zj = 1.
j=1 j=1

Having both a left and right inverse, a1 is invertible. 

Example 1.27 shows that the assumption that A is central is indispensable in

Theorem 1.30.

Remark 1.31 (a) In light of the end of the proof of Theorem 1.30 it seems appropriate
to mention that in a finite dimensional unital algebra A, a right inverse of an element
is automatically also a left inverse; that is, ab = 1 implies ba = 1. This can be
proved as follows. Take a ∈ A. By Lemma 1.5 there exists a nonzero polynomial
f (ω) ∈ F[ω] such that f (a) = 0. Assume that f has the smallest degree among
all such polynomials. We can write f (a) = 0 as ag(a) = α where α is the negative
of the constant term of f , and g(ω) ∈ F[ω] is a nonzero polynomial with smaller
degree than f (ω), so that g(a) = 0. If α = 0, then a is invertible, a−1 = α −1 g(a).
If α = 0, then ag(a) = g(a)a = 0 and a cannot have neither a left nor right inverse.
(b) The discussion in (a) shows, in particular, that finite dimensional unital algebras
enjoy a rather special property: every nonzero element is either invertible or a zero-
divisor.
(c) A nonzero subalgebra A of a finite dimensional division algebra D is itself a
division algebra. Indeed, take 0 = a ∈ A; then 1 ∈ A for a is algebraic and invertible,
and with reference to (a) we see that a−1 = α −1 g(a) ∈ A.

The field of rational functions F(ω) is a division algebra over F, but its subalgebra
F[ω] is not a division algebra. The finite dimensionality assumption in (c) is thus
necessary. The next example shows that (a) and (b) do not always hold in infinite
dimensional algebras.

Example 1.32 From calculus we know that differentiation and integration are inverse
operations. But one needs to be a little careful to interpret this statement correctly. If
x
by integration we mean the operator J given by J( f )(x) = 0 f (t)dt, then by first
integrating and then differentiating f we get f , but if we first differentiate and then
integrate the constant function 1 we get 0.
1.6 Automorphisms of Central Simple Algebras 15

Let us translate this phenomenon to an appropriate algebraic setting. Let D, J

be elements in the algebra EndR (R[ω]) defined by D(ωn ) = nωn−1 (just as in
Example 1.13) and J(ωn ) = ωn+1 , n ≥ 0 (note that it is enough to specify the
n+1

action of a linear operator on a basis). Then DJ = I, but JD = I. Thus, D is right

invertible but not left invertible, and J is left invertible but not right invertible. From
D(I − JD) = 0 and TDJ = T for any T ∈ EndR (R[ω]) we infer that D is a left
zero-divisor but not a right zero-divisor. Similarly, J is a right zero-divisor but not a
left zero-divisor.
We have chosen this example because of its connection to calculus. Here is a
slightly simpler example of two operators A and B that satisfy AB = I and BA = I:
A(1) = 0, A(ωn ) = ωn−1 , n ≥ 1, and B(ωn ) = ωn+1 , n ≥ 0.

1.7 Maximal Subfields

In this section we get back on track and consider division algebras again. Neverthe-
less, we will introduce the next notions in a more general setting.
If a subalgebra K of a unital algebra A is a field and K ⊇ F (i.e., K contains the
unity of A), then we call it a subfield of A. In this case we may consider A as a vector
space over K. Indeed, one can regard the multiplication of elements from K with
elements from A as a scalar multiplication K × A → A. The vector space axioms are
readily fulfilled. We remark that this does not yet mean that A is also a K-algebra—
this is true only if K is contained in the center Z(A) of A (cf. Example 1.21).

Remark 1.33 If K is a subfield of a finite dimensional unital algebra A, then we have

[A : F] = [A : K][K : F].

This formula is well-known in the special case where A is an extension field of K.

The same proof works in the general case. Indeed, if {ki | i = 1, . . . , m} is a basis of
K over F, and {aj | j = 1, . . . , n} is a basis of A over K, then {ki aj | i = 1, . . . , m, j =
1, . . . , n} is easily seen to be a basis of A over F.

Definition 1.34 A subfield that is not properly contained in a larger subfield of A is

called a maximal subfield of A.

Example 1.35 Each of R ⊕ Ri, R ⊕ Rj, and R ⊕ Rk is a maximal subfield of H,

2-dimensional and isomorphic to C. With some abuse of notation, we can therefore
regard H as a complex vector space (but not as a complex algebra).

If K is a subfield of A, then Ru ∈ EndK (A) for every u ∈ A. Indeed, Ru (kx) =

kxu = kRu (x) for all k ∈ K, x ∈ A. We also remark that given f ∈ EndK (A), its
scalar multiple k f, where k ∈ K, can be interpreted as the operator Lk f. This is a
trivial observation, but should be kept in mind when reading the next proof.
16 1 Finite Dimensional Division Algebras

Theorem 1.36 Let D be a finite dimensional central division F-algebra. If K is a

maximal subfield of D, then [D : F] = [K : F]2 .

Proof Pick a basis {u1 , . . . , ud } of the F-algebra D. We claim that {Ru1 , . . . , Rud }
is a basis of the K-algebra EndK (D). Lemma 1.24 implies that this set is linearly
independent, so it suffices to show that it spans EndK (D). Let f ∈ EndK (D). In
 f ∈ End F (D), so it follows from Lemma 1.25 (and Remark 1.23) that
particular,
f = di=1 Lai Rui for some ai ∈ D. Since f is K-linear, i.e., f Lk = Lk f holds for
every k ∈ K, we have


d
Lai k−kai Rui = 0.
i=1

Lemma 1.24 tells us that ai k − kai = 0 for every i and every k ∈ K. This implies
that ai ∈ K, since otherwise the subalgebra generated by ai and K would be a field
larger than K (cf. Remark 1.31 (c)). Thus, f is a K-linear combination of the Rui ’s,
proving our claim.
Accordingly,

[D : F] = d = [EndK (D) : K] = [D : K]2 .

On the other hand,

[D : F] = [D : K][K : F]

by Remark 1.33, and the result follows. 

Corollary 1.37 The dimension of a finite dimensional central division algebra is a

perfect square.

Proof A finite dimensional central division algebra certainly contains maximal sub-
fields. Indeed, these are exactly the subfields of maximal dimensions. The desired
conclusion therefore follows from Theorem 1.36. 

The main application of Theorem 1.36 will be given in the next section.

1.8 Wedderburn’s Theorem on Finite Division Rings

Finite fields exist and their structure is well-known. Do there exist finite division
rings that are not commutative? The set of nonzero elements of such a ring would
form a nonabelian finite group under multiplication. Understanding this group, which
misses only one element in our ring, should be sufficient for understanding the ring.
As finite groups are one of the most studied algebraic objects, it seems natural to
involve group-theoretic techniques when addressing the above question.
1.8 Wedderburn’s Theorem on Finite Division Rings 17

We shall need the class formula from elementary group theory. Let us recall it and
sketch its proof. Let G be a finite group with center Z(G). For every a ∈ G we let

C(a) := {g ∈ G | ag = ga}.

This is a subgroup of G. The set {xax −1 | x ∈ G} of all conjugates of a has the same
cardinality as the set of all (left) cosets of C(a) in G. Indeed, xax −1
→ xC(a) is
a well-defined bijection. Conjugacy is an equivalence relation on G. Partitioning G
into its disjoint equivalence classes (which are called conjugacy classes) and using
Lagrange’s theorem telling us that the number of cosets is |C|G| (a)| , we obtain the
desired formula
 |G|
|G| = |Z(G)| + , (1.12)
|C(a)|

where the sum is taken over representatives of nontrivial conjugacy classes.

We are now in a position to prove that the answer to our question from the first
paragraph is “no”. This was established by J. H. M. Wedderburn in 1905.
Theorem 1.38 (Wedderburn) A finite division ring is commutative.
Proof Suppose this is not true. Then there exists a noncommutative finite division
ring D of minimal cardinality; all its proper division subrings are therefore commu-
tative. We may regard D as a central division algebra over its center F.
For every a ∈ D we set

C(a) := {x ∈ D | ax = xa}.

Clearly C(a) is a division subring, and hence a subfield of D if a ∈

/ F. Since elements
in any subfield that contains a commute with a, C(a) is actually a maximal subfield.
Let q = |F|. (As we know, q is a prime power, but we will not use this.) Note that an
m-dimensional vector space over F has qm elements. Accordingly, by Theorem 1.36
2
there exists d ≥ 2 such that |D| = qd and |C(a)| = qd for every a ∈ D\F. Now
apply (1.12) to G := D∗ = D\{0}. Since Z(G) = F\{0} and C(a) = C(a)\{0} it
follows that there exists s ∈ N such that
2
2 qd − 1
qd − 1 = q − 1 + s .
qd − 1
2
2 qd −1
Since qd − 1 is a multiple of qd −1
, this leads to a contradiction that

2
qd − 1 (qd )d − 1
= = 1 + qd + · · · + qd(d−2) + qd(d−1)
qd − 1 qd − 1

is a divisor of q − 1. 
18 1 Finite Dimensional Division Algebras

The statement of this theorem is incredibly simple. It provides us an elementary

connection between three fundamental concepts: invertibility, finiteness, and com-
mutativity. Such simple, basic results often have equally simple, few lines proofs.
But no such proof of Wedderburn’s theorem is known (nor expected). There are,
however, numerous different proofs, both elementary and advanced. Probably the
most famous one is a short proof found by E. Witt in 1931. Although Witt’s proof
is a rather standard choice for a textbook (for example, it is included in Lam’s
book [Lam01]), we have demonstrated the proof by T. Nagahara and
H. Tominaga from 1974. It is also quite short and, hopefully the reader will agree,
beautiful. Besides, it is much closer to the spirit of this book than Witt’s and many
other proofs, especially because of the auxiliary results led to it, which will also be
used later.

1.9 Further Examples of Division Algebras

So far we have met only one example of a noncommutative division algebra, namely
H. Moreover, we have proved that there are no noncommutative finite dimensional
division algebras over algebraically closed fields or over finite fields, and that H is
the only such example over the field of real numbers. The lack of examples has now
become apparent. The aim of this section is to bring to light some relatively simple
examples; there are other important examples which, however, are not so easily
approachable by elementary means. Let us warn the reader that some verifications
will be omitted in our discussion. We feel that without indicating additional examples
this chapter would be a torso, but studying them in detail would make exposition a
bit lengthy and tedious.

Example 1.39 In Remark 1.7 we have pointed out that one does not need to limit to
R to get division algebras with the same multiplication table as H. The idea now is
to alter this table a little bit to get new examples of division algebras. Let F be a field
with char(F) =  2, and pick λ, μ ∈ F\{0}. We define the generalized quaternion
λ,μ
algebra A = F as the 4-dimensional unital algebra with basis {1, i, j, k} and
multiplication table

i2 = λ, j2 = μ, k 2 = −λμ,

ij = −ji = k, ki = −ik = −λj, jk = −kj = −μi.

 
Thus, H = −1,−1 R . It is easy to check that A is a central simple algebra; the proof
is left as an exercise for the reader. We define the conjugate a of a ∈ A in the same
way as in H. That is, if a = α0 + v, α0 ∈ F, v = α1 i + α2 j + α3 k, then a = α0 − v.
Noticing that
aa = aa = α02 − λα12 − μα22 + λμα32
1.9 Further Examples of Division Algebras 19
 
λ,μ
we see, by the same argument as in Remark 1.7, that A = F is a division algebra
if and only if for all αi ∈ F,

α02 − λα12 − μα22 + λμα32 = 0 =⇒ α0 = α1 = α2 = α3 = 0.

Now it depends on F whether this can be fulfilled for some choices of λ and μ. Of
course it is never fulfilled if F is algebraically closed. If F = R then it is fulfilled
if both λ and μ are < 0, but for any such choice of λ and μ we get an algebra
isomorphic to H. It turns out that over some fields (including Q) we can get infinitely
many pairwise nonisomorphic division algebras.
The downside of Example 1.39 is the limitation to dimension 4. By Corollary 1.37
this is the smallest possible dimension of a noncommutative central division algebra.
Let us proceed to higher dimensions.
A class of rings is often closed under some constructions. Is it possible to form
new division rings from a given division ring D? The usual suspects, direct products
and matrix ring construction, obviously fail because of abundance of zero-divisors.
What about the polynomial ring D[ω]? It is even farther from a division ring, as
polynomials of degree ≥ 1 will never be invertible in D[ω]. Let us consider a larger
ring D[[ω]] of formal power series over D. Its elements are written as formal sums
 ∞
n=0 an ω , an ∈ D. The main difference from polynomials is that we impose no
n

restriction on the ai ’s. The addition and multiplication are defined in the same way
as for polynomials:
∞
 ∞
 ∞

an ω +
n
bn ω :=
n
(an + bn )ωn ,
n=0 n=0 n=0


∞ 
∞  ∞
 
n
an ωn bn ωn := cn ωn , where cn = ai bn−i .
n=0 n=0 n=0 i=0

This definition obviously makes sense if we replace D by an arbitrary ring R. Thus, for
every ring R we can construct the formal power series ring R[[ω]] which contains
the polynomial ring R[ω] as a subring. However, let us return to the case where
R = D.

Lemma 1.40 Let D be a division ring. Then ∞ n=0 an ω is invertible in D[[ω]] if
n

and only if a0 = 0.

Proof The right invertibility of ∞n=0 an ω is equivalent to the existence of elements
n

bi ∈ D, i ≥ 0, such that

a0 b0 = 1, a0 b1 + a1 b0 = 0, a0 b2 + a1 b1 + a2 b0 = 0, etc.

If a0 = 0, then these equations can be solved. Indeed, the first equation

 gives b0 =
a0−1 , the second one gives b1 = −a0−1 a1 b0 , etc. Thus in this case ∞n=0 an ω is right
n
20 1 Finite Dimensional Division Algebras

invertible.
 Similarly we see that it is left invertible, and hence invertible. Conversely,
if ∞ n=0 n ω is invertible, then a0  = 0 for a0 b0 = 1 holds for some b0 ∈ R.
a n 
Thus, although much closer than D[ω], D[[ω]] is still far from being a division
ring. If only ω was invertible, then the invertibility of elements with zero constant
term would not be a problem. Why not make it invertible?
Example 1.41 For an arbitrary
 ring R we define the Laurent series ring R((ω)) as
the set of all formal series ∞n=−∞ an ω where only finitely many an ’s with n < 0
n

can be nonzero, and with operations defined in the obvious way by formally adding
and multiplying series. Note that multiplication is possible
since, by assumption, for
every f = 0 in R((ω)) there exists m ∈ Z such that f = ∞ n=m an ω and am  = 0.
n

One can check that R((ω)) is a ring containing R[[ω]] as a subring. If R = D is a

division ring and f is a nonzero element written as above, then f ω−m is invertible
by Lemma 1.40, implying that f is invertible. Accordingly, D((ω)) is a division ring.
If A is an algebra, then the ring A((ω)) becomes an algebra by defining scalar
multiplication in the self-explanatory way. Of course, if D is a division algebra, then
D((ω)) is also a division algebra. However, it is infinite dimensional. Therefore this
example is not entirely satisfactory for us. Let us modify it a little bit.
In each of the rings R[ω], R[[ω]], and R((ω)), ω commutes with elements in R
and is therefore a central element. We will now change this as follows. Let σ be a
ring endomorphism of R. Instead of ωa = aω we now require that ωa = σ (a)ω for
all a ∈ R. Accordingly,

(aωi )(bω j ) = aσ i (b)ωi+j for all a, b ∈ R. (1.13)


We can now consider the set of polynomials ni=0 ai ωi endowed with the usual
addition and multiplication determined by (1.13). The resulting ring is called a skew
polynomial ring and is denoted by R[ω; σ ]. Similarly we introduce the skew power
series ring R[[ω; σ ]]. Assuming that σ is an automorphism we see that (1.13) makes
sense also for the negative integers i and j, which enables us to introduce the skew
Laurent series ring R((ω; σ )).
Example 1.42 If D is a division algebra, then so is D((ω; σ )). The proof is practically
the same as for D((ω)). Of course, D((ω; σ )) is infinite dimensional over the base
field F, but we can also view it as an algebra over its center Z. Let us add some
assumptions. First of all, as the noncommutativity is guaranteed by (1.13), we assume
that D = K is a field. Further, assume that the automorphism σ has order n (i.e., n is the
smallest natural number such that σ n = idK ), and F = {x ∈ K | σ (x) = x}. Readers
familiar with Galois Theory surely know concrete examples of such automorphisms.
We leave it as an exercise to prove that the center Z of K((ω; σ )) is F((ωn )), and that
[K((ω; σ )) : Z] = n2 (or see [Lam01, pp. 217–218] for details). That is, K((ω; σ ))
is a central division algebra over Z of dimension n2 .
We recommend [Lam01, Pie82], and [Row08] for more examples of finite dimen-
sional division algebras. One may say that these algebras do exist and are many, but
1.9 Further Examples of Division Algebras 21

it is not so easy to construct them. Also, it is not always easy to study them. Some
rather basic questions about them are still open. Here is an example:

Problem 1.43 Suppose a division algebra D is algebraic and finitely generated (as
an algebra). Is then D finite dimensional?

This question was posed for general (not necessarily division) algebras by A. G.
Kurosh in the early 1940s. However, in 1964 E. S. Golod and I. R. Shafarevitch
showed that in this generality the answer is negative. Thus it is necessary to impose
some restrictions on the algebras in question. Problem 1.43 is often referred to as the
“Kurosh problem for division algebras”.

Exercises

1.1. Prove that the following statements are equivalent for a nonzero ring D:
(i) D is a division ring.
(ii) For all a, b ∈ D with a = 0, the equations ax = b and ya = b have
unique solutions in D.
(iii) For all a, b ∈ D with a = 0, the equation ax = b has a solution in D.
(iv) D2 = 0 and D has no right ideals other than 0 and D.
Remark: As mentioned in Sect. 1.2, nonassociative division algebras are
defined through condition (ii). The equivalence of (i) and (ii) thus means
that in associative algebras this definition coincides with the standard one.
In nonassociative algebras, however, the condition of being a division algebra
is independent of the condition that every nonzero element a has a right and
left inverse. The latter only means that the equations ax = b and ya = b can
be solved for b = 1, while the former does not even imply the existence of
unity. For example, take an associative division algebra (possibly a field) D
having an automorphism ϕ = idD . The vector space D endowed with the new
product x · y := ϕ(xy) is then a nonassociative division algebra without unity.
1.2. Let x, y be elements in a division ring D such that x = 0, y = 0, and x = y−1 .
 −1
Prove Hua’s identity xyx = x − x −1 + (y−1 − x)−1 .
1.3. Show that a division ring D satisfying (xy)2 = (yx)2 for all x, y ∈ D is
commutative.
Hint: When considering an identity in a ring that involves arbitrary elements
it is often useful to replace these elements by the sum (or product) of two
elements.
1.4. Let D be a unital R-algebra. Suppose there exists a linear map x
→ x̃ from D
into D such that x x̃ = x̃x ∈ R for every x ∈ D, and x x̃ = 0 if x = 0. Show
that D is isomorphic to R, C, or H.
22 1 Finite Dimensional Division Algebras

Hint: See the comment following Lemma 1.5, and the hint from the previous
exercise.
1.5. Let α ∈ R. Find all solutions of the equation x 2 = α in H.
1.6. Let R be a unital ring. Show that every ideal of Mn (R) is of the form Mn (I) for
some ideal I of R.
1.7. Show that a (not necessarily unital) ring R is simple if and only if Mn (R) is
simple.
1.8. Show that the center of a simple ring is either 0 or a field. In particular, a
commutative simple ring is thus necessarily a field.
1.9. Give an alternative proof of Proposition 1.6 by considering the eigenvalues of
left multiplication maps.
1.10. Show that a ring R has a left unity (i.e., a left identity element with respect to
multiplication) if and only if every map f : R → R satisfying f (xy) = f (x)y,
x, y ∈ R, is of the form f = La for some a ∈ R. Describe all such maps f in
the case where R = Mn (2Z).
1.11. Let A be a unital F-algebra with [A : F] = n. Show that [M(A) : F] ≥ n, and
find an example where [M(A) : F] = n > 1.
Remark: The other extreme where [M(A) : F] = n2 is treated in Lemma 1.25.
1.12. Find an element in M(H) that cannot be written as La Rb , a, b ∈ H.
1.13. Let A be a finite dimensional central simple algebra. Suppose that a linear map
d : A → A satisfies d(xy) = d(x)y + xd(y), x, y ∈ A. Show that there exists
a ∈ A such that d(x) = [a, x], x ∈ A.
Remark: In terminology introduced in Definition 2.41 and Example 2.43 below,
this can be stated as that every derivation of A is inner.
1.14. Check that if ϕ is an automorphism of the algebra A, then (aij )
→ (ϕ(aij )) is
an automorphism of the algebra Mn (A). Use this to find an outer automorphism
of the R-algebra Mn (C).
1.15. Show that a nonzero algebra endomorphism of a finite dimensional simple
algebra is an automorphism.
1.16. Let A be a finite dimensional central simple algebra. Describe all algebra
endomorphisms of A × A.
1.17. Show that the only inner automorphism of the Weyl algebra A1 is idA1 , and
find an example of an outer automorphism of A1 .
Hint: What restrictions on the action of an automorphism on D and L does
[D, L] = I bring?
1.18. A map x
→ x ∗ from an F-algebra A into itself is called an involution if

(x + y)∗ = x ∗ + y∗ , (xy)∗ = y∗ x ∗ , (x ∗ )∗ = x

for all x, y ∈ A. Immediate examples are the conjugation h

→ h on H, the
transposition M
→ M t on Mn (F), and the conjugate transposition M
→ M ∗
on Mn (C). Given an involution ∗ on A, we set S = {x ∈ A | x ∗ = x} and
Exercises 23

K = {x ∈ A | x ∗ = −x}. Elements in S are called symmetric, and elements in

K are called skew-symmetric. Show that S and K are additive subgroups of
A, xy + yx ∈ S for all x, y ∈ S, xy − yx ∈ K for all x, y ∈ K, and A = S ⊕ K if
char(F) = 2. Find a linear involution on M2 (F), char(F) = 2, such that the
space of symmetric elements S is 1-dimensional.
Hint: Involutions share some properties with the map x
→ x −1 on the group
of invertible elements, which is in the matrix case connected with the notion
of an adjugate matrix.
1.19. A bijective linear map θ from an F-algebra A into itself is called an
antiautomorphism if θ (xy) = θ (y)θ (x) for all x, y ∈ A (linear involutions
are thus antiautomorphisms). Describe all antiautomorphisms of R-algebras
Mn (R) and Mn (H).
Hint: What is the product of two antiautomorphisms?
1.20. Show that every maximal subfield K of H is of the form K = h(R ⊕ Ri)h−1
for some h ∈ H\{0}.
Hint: This can be shown in many ways. To minimize computation, try to apply
Theorem 1.30. More precisely, find an isomorphism from R ⊕ Ri onto K, and
then extend it to an automorphism of H (and thereby prove “by hand” a special
case of Theorem 4.46 below).
1.21. Find a maximal subfield of the F-algebra M2 (F) for F = R, C.
1.22. Prove that a finite dimensional division algebra D is a finite field if and only
if there exists n ∈ N such that x n = x for all x ∈ D.
Remark: In Sect. 5.10 we will establish an incomparably deeper result. This
exercise is meant just as a small illustration of the applicability of rudimentary
tools that are at our disposal now.
 
1.23. Show that 1,1 ∼
= M2 (F) for every field F with char(F) = 2.
F
1.24. Let D be a division ring. Show that every nonzero ideal of D[[ω]] is generated
by ωn for some n ≥ 0.
Chapter 2
Structure of Finite Dimensional Algebras

This chapter is centered around Wedderburn’s structure theory which under rather
mild and inevitable restrictions describes the form of a finite dimensional algebra. In
some sense it reduces the problem of understanding finite dimensional algebras to
the problem of understanding finite dimensional division algebras. The latter can be
difficult, but at least for division algebras over some fields definitive answers were
given in Chap. 1.
Just as in the previous one, in this chapter we also wish to follow the conceptual
simplicity as our basic principle. Rather than building powerful general machineries
and then derive classical theorems as their byproducts, we will give direct proofs
that only illustrate more general phenomena. There will be enough opportunities
for demonstrating more profound theories later. That said, we will usually say a bit
more than needed for achieving our main goals, and make occasional digressions
from the main themes. Partially this will be done in order to make the exposition
less condensed and more illuminating, and partially because of future goals. Thus
we shall give a survey of the very basic theory, but with one eye on more general
topics.

2.1 Nilpotent Ideals

So far we have treated rings without proper nontrivial ideals. Such rings can be
relatively easily handled, they are considered as “nice” from the structural point of
view. We will now briefly consider rings that are in complete contrast; they contain
nontrivial ideals with a particularly “bad” property.

Definition 2.1 An ideal I of a ring R is said to be a nilpotent ideal if there exists

n ∈ N such that I n = 0.

The condition I n = 0 means that u1 u2 . . . un = 0 for all ui ∈ I. To avoid confusion,

we introduce two related notions.

© Springer International Publishing Switzerland 2014 25

M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_2
26 2 Structure of Finite Dimensional Algebras

Definition 2.2 An element a in a ring R is said to be a nilpotent element if there

exists n ∈ N such that an = 0.
Definition 2.3 An ideal I of a ring R is said to be a nil ideal if all elements in I are
nilpotent.
Obviously, a nilpotent ideal is nil. The converse is not true in general (see
Example 2.7).
A trivial example of a nilpotent ideal is I = 0, but of course we are interested in
nontrivial examples. An extreme case is when the ring R itself is nilpotent.
Example 2.4 Take any additive group R, and equip it with trivial product: xy = 0
for all x, y ∈ R. Then R2 = 0.
Example 2.5 A nilpotent element lying in the center Z(R) of the ring R clearly
generates a nilpotent ideal. A simple example of such an element can be obtained by
taking the direct product of the ring from Example 2.4 with any other ring.
The next example is more illuminating.
Example 2.6 Let A = Tn (F) be the algebra of all upper triangular n×n matrices over
a field F, i.e., matrices that have zeros below the main diagonal. Let N be the set of
all strictly upper triangular matrices, i.e., matrices in A that have zeros also on the
diagonal. Then N is an ideal of A such that N n = 0 and N n−1 = 0. One can find a
variety of other ideals of A that are contained in N, and are therefore nilpotent. For
instance, the set I of all matrices whose (1, n) entry is arbitrary and all other entries
are 0 is an ideal of A satisfying I 2 = 0.
Example 2.7 Let A be the set of all infinite, N × N matrices over F that are upper
triangular and have only finitely many nonzero entries. Note that A is an algebra
under the standard matrix operations. Let I be the set of all matrices in A that are
strictly upper triangular. Then I is a nil ideal of A which is not nilpotent.
Remark 2.8 Nilpotent one-sided ideals are defined in the same way as nilpotent
ideals. If L is a nilpotent left ideal of R, then L is contained in a nilpotent (two-sided)
ideal L + LR. Indeed, it is easy to see that L n = 0 implies (L + LR)n = 0. Similarly,
every nilpotent right ideal is contained in a nilpotent ideal.
At this point we can mention a celebrated open problem. From Remark 2.8 we see
that a ring with a nonzero nilpotent one-sided ideal also contains a nonzero nilpotent
ideal. In 1930, G. Köthe conjectured that the same is true for nil ideals. The following
question is thus known as Köthe’s problem.
Problem 2.9 If a ring R has a nonzero nil one-sided ideal, does R also have a nonzero
nil ideal?
In various special situations this is known to be true. But in general the problem is
still open after all these years.
Let us return to nilpotent ideals.
2.1 Nilpotent Ideals 27

Lemma 2.10 The sum of two nilpotent ideals is a nilpotent ideal.

Proof Let I, J be ideals such that I n = 0 and J m = 0. We claim that (I +J)n+m−1 = 0.

That is, the product of n + m − 1 elements of the form u + v, u ∈ I, v ∈ J, is 0. Such
a product can be written as a sum of products w = w1 w2 . . . wn+m−1 where each
wi ∈ I ∪ J. If at least n of these wi s are in I, then w = 0 as I n = 0. If the number of
the wi ’s belonging to I is smaller than n, then at least m of them lie in J, and hence
w = 0 since J m = 0. 

By a maximal nilpotent ideal we mean a nilpotent ideal that is not properly

contained in a larger nilpotent ideal.

Lemma 2.11 If a ring R has a maximal nilpotent ideal N, then N contains all
nilpotent ideals of R.

Proof If I is another nilpotent ideal, then I + N is again a nilpotent ideal by

Lemma 2.10. Because of the maximality of N we must have I + N = N, and thus
I ⊆ N. 

Thus, a maximal nilpotent ideal, if it exists, is unique and is equal to the sum of
all nilpotent ideals. However, not every ring has such an ideal. That is to say, the sum
of all nilpotent ideals of a ring is not always nilpotent (although it is nil for each of
its elements is contained in a nilpotent ideal by Lemma 2.10).

Example 2.12 Let A and I be the algebra and its nil ideal from Example 2.7. For every
k ∈ N, let Ik denote the set of all matrices in I with the property that their nonzero
entries appear only in the first k rows. It is easy to check that Ik is a nilpotent ideal
of A; in fact, Ikk+1 = 0. If A had a maximal nilpotent ideal N, then, by Lemma 2.11,
N would contain each Ik , and hence also ∞ k=1 Ik = I. However, I is not a nilpotent
ideal.

Anyway, a finite dimensional algebra certainly does have a maximal nilpotent

ideal. Indeed, it contains at least one nilpotent ideal (namely 0), and therefore the
(nonempty) set of all nilpotent ideals contains an element of maximal dimension.
This is obviously a maximal nilpotent ideal.

Definition 2.13 The maximal nilpotent ideal of a finite dimensional algebra A is

called the radical of A.

Let us stress that this definition is adjusted to the finite dimensional context. There
are several nonequivalent definitions of radicals of general rings, which agree with
Definition 2.13 in the case of finite dimensional algebras.

Example 2.14 The radical of the algebra A = Tn (F) from Example 2.6 is N, the set
of all strictly upper triangular matrices. Indeed, N is a nilpotent ideal of A, and any
ideal of A that properly contains N cannot be nilpotent since it necessarily contains
a nonzero diagonal matrix.
28 2 Structure of Finite Dimensional Algebras

Example 2.15 The radical of a finite dimensional simple algebra is 0. Namely, a

simple algebra A cannot be nilpotent since An = 0 implies that An−1 is a proper ideal
of A, and hence it is 0.
From now on in this chapter we will be interested in rings that have no nonzero
nilpotent ideals. At the end we will describe the structure of all finite dimensional
algebras with this property.

2.2 Prime and Semiprime Rings

The goal of this section is to introduce two important classes of rings, the prime rings
and the semiprime rings. Let us start, however, with another class with which the
reader is presumably already familiar, at least in the commutative context.
Definition 2.16 A ring R is said to be a domain if for all a, b ∈ R, ab = 0 implies
a = 0 or b = 0.
In other words, R is a domain if it has no left (or right) zero-divisors. Equivalently,
R is a domain if it has the cancellation property: If a = 0, then each of ab = ac and
ba = ca implies b = c.
Commutative domains are of utmost importance in algebra. Noncommutative
domains also form a notable class of rings, but their role in noncommutative algebra
is not entirely parallel to the role of commutative domains in commutative algebra.
The most basic examples of commutative rings are domains, while even the matrix
ring Mn (F), which one might consider as the prototype of a noncommutative ring,
is not a domain. The next lemma introduces a wider class of rings, which can be
regarded as a noncommutative counterpart of the class of commutative domains.
Lemma 2.17 Let R be a ring. The following conditions are equivalent:
(i) For all a, b ∈ R, aRb = 0 implies a = 0 or b = 0.
(ii) For all left ideals I and J of R, IJ = 0 implies I = 0 or J = 0.
(iii) For all right ideals I and J of R, IJ = 0 implies I = 0 or J = 0.
(iv) For all ideals I and J of R, IJ = 0 implies I = 0 or J = 0.
Proof If I and J are left ideals satisfying IJ = 0, then IRJ = 0 since RJ ⊆ J. Hence
we see that (i) implies (ii). Similarly, (i) implies (iii), and each of (ii) and (iii) trivially
implies (iv).
Assume that (iv) holds and that a, b ∈ R satisfy aRb = 0. The product of the
ideals RaR and RbR is then 0. Hence one of them, say RaR, is 0. This implies that Ra
and aR are two-sided ideals such that Ra · R = R · aR = 0. By (iv), Ra = aR = 0.
But then Za is an ideal of R satisfying Za · R = 0, and so (iv) now yields a = 0.
Thus, (iv) implies (i). 
Definition 2.18 A ring R is said to be prime if it satisfies one (and hence all) of the
conditions of Lemma 2.17.
2.2 Prime and Semiprime Rings 29

Lemma 2.19 A commutative ring is prime if and only if it is a domain.

Proof It is enough to observe that ab = 0 implies aRb = 0 if R is commutative. 

Remark 2.20 Let R be a prime ring with char(R) = 0. We first remark that if 0 =
a ∈ R and n ∈ N are such that na = 0, then aR(nb) = 0 for every b ∈ R, and hence
nR = 0. Secondly, if nR = 0 and n = rs for some r, s ∈ N, then rR ·sR = 0. Since rR
and sR are ideals of R, we must have rR = 0 or sR = 0. This implies that char(R) is
a prime number p. Therefore we can consider R as an algebra over Zp in the natural
way. That is, for k ∈ Zp = {0, 1, . . . , p − 1} and x ∈ R we define kx as x + · · · + x
(k times).

We proceed with a yet wider class of rings.

Lemma 2.21 Let R be a ring. The following conditions are equivalent:

(i) For all a ∈ R, aRa = 0 implies a = 0.
(ii) For all left ideals I of R, I 2 = 0 implies I = 0.
(iii) For all right ideals I of R, I 2 = 0 implies I = 0.
(iv) For all ideals I of R, I 2 = 0 implies I = 0.
(v) R has no nonzero nilpotent ideals.

Proof The proof that (i)–(iv) are equivalent is similar to the proof of Lemma 2.17,
so we omit it. It is trivial that (v) implies (iv). Conversely, if (iv) holds and I is an
ideal such that I n = 0, then (I n−1 )2 = 0 and hence I n−1 = 0. Inductively, we get
I = 0. 

Definition 2.22 A ring R is said to be semiprime if it satisfies one (and hence all)
of the conditions of Lemma 2.21.

We remark that as a special case of (i) we have that each of the conditions aR = 0
and Ra = 0 implies a = 0 if R is semiprime.
Let us record the obvious analogue of Lemma 2.19.

Lemma 2.23 A commutative ring is semiprime if and only if it has no nonzero

nilpotent elements.

Remark 2.24 If A is an algebra and I, J are ring ideals of A such that IJ = 0, then the
linear spans of I and J are algebra ideals of A whose product is 0. Hence it follows
that a prime algebra can be equivalently defined as an algebra which is prime as a
ring, or as an algebra in which the product of any two of its nonzero algebra ideals
is nonzero. A similar remark holds for semiprime algebras (as well as for simple
algebras, cf. Remark 1.11).

Remark 2.25 Let I be a nonzero ideal of a prime ring R. If a, b ∈ R are such that
aIb = 0, then aRuRb = 0 for every u ∈ I. Using (i) from Lemma 2.17 twice it
follows that a = 0 or b = 0. This in particular shows that an ideal of a prime ring is
again a prime ring. Similarly, an ideal of a semiprime ring is a semiprime ring.
30 2 Structure of Finite Dimensional Algebras

Let us return to the situation considered in Lemma 2.11. The following can be
added.

Lemma 2.26 If N is a maximal nilpotent ideal of a ring R, then the factor ring R/N
is semiprime.

Proof Let K be an ideal of R/N such that K 2 = 0. Then J = {x ∈ R | x + N ∈ K}

is an ideal of R such that J 2 ⊆ N. As N is nilpotent, it follows that J is nilpotent as
well. Lemma 2.11 therefore implies that J ⊆ N, and hence K = 0. 

Example 2.27 Let A and N be as in Example 2.14. Then A/N is isomorphic to F n ,

the direct product of n copies of F. Indeed, the map (aij ) → (a11 , a22 , . . . , ann )
from A to F n is easily seen to be a surjective algebra homomorphism with kernel N.

The following relations between the classes of rings introduced so far are obvious
from definitions:

division ring =⇒ simple and domain,

simple =⇒ prime,
domain =⇒ prime,
prime =⇒ semiprime.

None of these implications can be reversed.

Example 2.28 We know that the Weyl algebra A1 is simple. Let us show that it is
also a domain, but not a division ring. We use the same notation as in Example 1.13.
Take S, T ∈ L and r, s ≥ 0. Applying (1.8), it follows that


r+s−1
(SDr )(TDs ) = STDr+s + S[Dr , T ]Ds = STDr+s + Si Di
i=s

for some Si ∈ L . Accordingly, for all Sr , Ts ∈ L we have


m 
n  
m+n−1
Sr Dr Ts Ds = Sm Tn Dm+n + Sj D j
r=0 s=0 j=0

where Sj ∈ L . Since L is obviously a domain (isomorphic to F[ω]), it follows from

(b) in Example 1.13 that A1 is a domain, and also that the only invertible elements
in A1 are the nonzero elements from F.

Example 2.29 The ring Z is a domain that is not a simple ring.

Example 2.30 The matrix ring Mn (F), n ≥ 2, is simple, but not a domain.
2.2 Prime and Semiprime Rings 31

Example 2.31 It is easy to check that the ring Mn (Z), n ≥ 2, is prime. But it is
neither a domain nor a simple ring. The obvious, and in fact the only examples of its
ideals are Mn (kZ), k ≥ 0.

Example 2.32 Every commutative ring that has zero-divisors but does not have
nonzero nilpotent elements, is semiprime but not prime. A concrete example is the
ring of continuous functions C[a, b].

The next example points out an important difference between the classes of prime
and semiprime rings, and also indicates at least a slight analogy between prime rings
and prime numbers.

Example 2.33 Let R1 and R2 be nonzero rings. Then their direct product R = R1 ×R2
is not a prime ring, as R1 × 0 and 0 × R2 are nonzero ideals of R whose product is
0. On the other hand, if both R1 and R2 are semiprime, then R is also semiprime.

We shall soon meet an important and non-obvious example of a finite dimensional

semiprime algebra, related to the notion of a group. Before arriving at this specific
topic, we will consider, in the next two sections, some notions that are of general
importance in ring theory.

2.3 Unitization

Let A be an F-algebra. Then the set F × A becomes an F-algebra, which we denote

by A , if we define addition, scalar multiplication and product as follows:

(λ, x) + (μ, y) := (λ + μ, x + y),

μ(λ, x) := (μλ, μx),
(λ, x)(μ, y) := (λμ, μx + λy + xy).

We consider A as a subalgebra of A via the embedding x → (0, x). Note that A is

actually an ideal of A . A crucial observation for us is that A is a unital algebra.
Indeed, (1, 0) is its unity.

Definition 2.34 The algebra A is called the unitization of A.

Remark 2.35 Replacing the role of F by Z one defines the unitization R of a ring
R in exactly the same way (just ignore the scalar multiplication). Alternatively, one
can use Zn instead of Z if char(R) = n > 0.

This construction is intended primarily for algebras (and rings) without unity.
The idea behind it is to be able to reduce some problems for general algebras to
unital algebras; this does not always work, but sometimes it does. In principle one
can construct A even when A is unital. This may seem somewhat artificial at first
32 2 Structure of Finite Dimensional Algebras

glance, especially since the unity of A is then different from the unity of A . However,
constructions of new rings and algebras from the old ones sometimes turn out to be
unexpectedly useful, say when searching for counterexamples.
The unitization of A does not preserve all properties of A. For instance, the sim-
plicity is definitely not preservered since A is an ideal of A . If A is a nonzero prime
unital algebra, then A is not prime since I = {(λ, −λ) | λ ∈ F} is an ideal of A
such that IA = AI = 0. The non-unital case is different.

Lemma 2.36 If A is a prime algebra without unity, then A is also prime.

Proof Let (λ, a), (μ, b) ∈ A satisfy (λ, a)A (μ, b) = 0. Then (λ, a)(1, 0)(μ, b) =
0, and hence λ = 0 or μ = 0. Let us consider the case where λ = 0; the case where
μ = 0 can be treated similarly. We may assume that a = 0. From (0, a)(0, x)(μ, b) =
0 we infer that μax + axb = 0 for every x ∈ A. We may now also assume that μ = 0,
since otherwise aAb = 0 and hence b = 0, as desired. Setting e := −μ−1 b we thus
have ax = axe for every x ∈ A. Accordingly, a(xy)e = axy = (axe)y for all x, y ∈ A.
We can rewrite this as

ax(y − ye) = 0 = ax(y − ey).

Since A is prime it follows that y = ye and y = ey for every y ∈ A. This contradicts

the assumption that A is not unital. 

2.4 The Regular Representation

Let us begin with a question. Given a unital algebra A, is it possible that

[a, b] = 1 (2.1)

for some a, b ∈ A? It is not our intention to discuss the reasons for the relevance of
this equation. This problem was primarily chosen in order to illustrate the usefulness
of the concept that we are about to introduce.
The reader’s answer to our question might be “why not?”, and it is difficult to argue
against it. After all, we saw that (2.1) appears in the Weyl algebra A1 ; see (1.4). This
algebra is infinite dimensional. Let us modify our question: Can (2.1) occur in a finite
dimensional algebra? The answer is closer as it may seem. A shortcut to it is based on
left multiplication maps La : x → ax. Recalling the formulas Lλa+μb = λLa + μLb
and Lab = La Lb we see that a → La is an algebra homomorphism.

Definition 2.37 The homomorphism a → La from A into EndF (A) is called the
regular representation of A.

The regular representation is injective, unless aA = 0 for some nonzero a ∈ A.

The latter condition is quite special, in particular it cannot happen if A is unital. Thus,
under a mild assumption a → La is an embedding of A in EndF (A), which makes
2.4 The Regular Representation 33

it possible for one to identify an element a ∈ A with La ∈ EndF (A). When trying
to deal with an element a from an abstract algebra one might feel barehanded at the
start. The advantage of considering La is that it is a linear operator and therefore we
can rely on linear algebra methods.
The answer to our question is now within reach, but let us first record the following
lemma giving a firm basis for using the approach just indicated. The reader will notice
the analogy with Cayley’s theorem from group theory.

Proposition 2.38 Every F-algebra A can be embedded into the algebra EndF (V )
for some vector space V . If A is finite dimensional, then V can be chosen to be finite
dimensional, and so in this case A can be embedded into Mn (F) for some n ∈ N.

Proof If A is unital, then we can take V = A and apply the regular representation.
If it is not unital, then we can embed A into its unitization A , and accordingly take
V = A . In any case V is finite dimensional if A is; if V is n-dimensional, then
EndF (V ) ∼ = Mn (F). 

This proposition enables one to consider elements from a finite dimensional alge-
bra as matrices. The proof was easy, and this observation is often useless. But
occasionally it can be very helpful. Let us return to our question. Thus, assume
that a finite dimensional algebra contains elements a, b satisfying (2.1). Then, by
Proposition 2.38, there exist matrices A, B ∈ Mn (F) such that [A, B] = I (namely, the
regular representation maps 1 into I). Now, the trace of the matrix [A, B] = AB − BA
is 0, while the trace of the matrix I is n. If char(F) = 0, then this is impossible, and
hence (2.1) cannot occur. The same argument yields the following sharper assertion.

Proposition 2.39 Let A be a nonzero finite dimensional unital algebra over a field
F with char(F) = 0. Then 1 cannot be written as a sum of commutators in A.
Example 2.40 Let us show that Proposition 2.39 does not hold if F has prime char-
acteristic p. Define D, L ∈ EndF (F[ω]) just as in Example 1.13. Let I be the ideal
of F[ω] generated by ω p , let V = F[ω]/I, and let φ : F[ω] → V be the canonical
homomorphism. Note that V is a p-dimensional vector space. Since D(ω p ) = 0 it fol-
lows that D(I) ⊆ I. We have L(I) ⊆ I for trivial reasons. These inclusions show that
the maps d,  : V → V given by dφ = φD and φ = φL are well-defined. Clearly
d,  ∈ EndF (V ) ∼= Mp (F), and we have [d, ] = 1. The reader may wish to write
down the matrix representations of d and  to obtain a more explicit interpretation
of this example.

One might wonder whether there is really no other way to handle (2.1) than reduc-
ing the problem to matrices. This does seem to be the most natural and simple way, but
there are others. Let us allow ourselves a digression, and prove a lovely result by N.
Jacobson from 1935, dealing with the more general condition [[a, b], a] = 0. This can
easily occur in finite dimensional algebras even when a and b do not commute, regard-
less of char(F). Say, the standard matrix units E11 , E12 satisfy [E12 , E11 ] = −E12 ,
and hence [[E12 , E11 ], E12 ] = 0. The most convenient way to consider this condition
is through the following concept.
34 2 Structure of Finite Dimensional Algebras

Definition 2.41 Let A be an algebra (resp. ring). A linear (resp. additive) map d :
A → A is called a derivation if d(xy) = d(x)y + xd(y) for all x, y ∈ A.

It is easy to guess where this name comes from.

Example 2.42 The differential operator D from Example 1.13 is a derivation.

The basic example in noncommutative rings is of different nature.

Example 2.43 For every a in an algebra (or ring) A, the map d : A → A given by
d(x) = [a, x] is a derivation. Such a derivation is said to be an inner derivation.

Proposition 2.44 (Jacobson) Let A be a finite dimensional algebra over a field F

with char(F) = 0. If a, b ∈ A satisfy [[a, b], a] = 0, then [a, b] is a nilpotent
element.

Proof Let d be the inner derivation d(x) = [a, x]. Our condition can be written as
d 2 (b) = 0. A simple induction argument shows that

d(d(b)k ) = 0 for every k ∈ N. (2.2)

We claim that

d n (bn ) = n!d(b)n for every n ∈ N. (2.3)

This is trivial for n = 1, so let n > 1 and assume that (2.3) holds for n − 1. From the
definition of a derivation one easily infers that d n satisfies the Leibniz rule that we
know from calculus. Accordingly,
n  
 n i
d n (bn ) = d n (bbn−1 ) = d (b)d n−i (bn−1 )
i
i=0
= bd(d n−1 (bn−1 )) + nd(b)d n−1 (bn−1 ).

Now use the induction assumption together with (2.2) and (2.3) follows.
As an immediate consequence of (2.2) and (2.3) we get d m (bk ) = 0 whenever
k < m. By Lemma 1.5 there exists m ∈ N such that bm is a linear combination
of bk with k < m. Hence d m (bm ) = 0, and so m!d(b)m = 0 by (2.3); that is,
[a, b]m = 0. 

The proof just given was actually discovered by D. Kleinecke, but it is similar to
Jacobson’s original proof.

Remark 2.45 From the proof we see that a slightly more general result than stated
is true: If A is as in Proposition 2.44, d is a derivation of A, and d 2 (b) = 0 for some
b ∈ A, then d(b) is nilpotent. Note that we can reword Proposition 2.44 as follows:
If δ is an inner derivation of A (induced by b), and a ∈ A is such that [δ(a), a] = 0,
2.4 The Regular Representation 35

then δ(a) is nilpotent. Using the regular representation it can be easily shown that
this is also true for every, not necessarily inner, derivation δ. Indeed, [δ(a), a] = 0
implies [Lδ(a) , La ] = 0. The condition that δ is a derivation can be expressed as
Lδ(a) = [δ, La ]. Therefore we have [[δ, La ], La ] = 0. Proposition 2.44 applied to the
(also finite dimensional) algebra EndF (A) tells us that Lδ(a) = [δ, La ] is nilpotent.
Hence δ(a) is nilpotent too.

A more sophisticated application of the regular representation will be given in the

next section.

2.5 Group Algebras

Group theory and ring theory are two branches of algebra. In a first course in abstract
algebra one might get an impression that, although sometimes similar, they are basi-
cally unconnected. But there are many links between groups and rings. The concept
that we are about to introduce gives rise to striking interactions.
Assume temporarily that G is an arbitrary set. If F is, as always, a field, then
 we can
form the vector space over F whose basis is G. Its elements are formal sums g∈G λg g
where λg ∈ F and all but finitely many λg are zero. The definitions of addition and
scalar multiplication are self-explanatory. Assume now that G is a group. Then this
vector space becomes an algebra if we define multiplication by simply extending
the group multiplication on G to the whole space; taking into account the algebra
axioms this can obviously be done in a unique way. Thus,
    
λg g μh h = νk k, where νk = λg μh .
g∈G h∈G k∈G gh=k

We denote this algebra by F[G].

Definition 2.46 The algebra F[G] is called the group algebra of G over F.

It is actually enough to assume that G is merely a semigroup to construct F[G].

In this case we call F[G] a semigroup algebra. Similarly, we can speak about a
monoid algebra if G is a monoid. Furthermore, one can replace the role of the field
F by any ring R, and then define addition and multiplication in formally the same
way. Everything still makes sense, but of course the resulting object is a ring rather
than an algebra. It is called a (semi)group ring and is denoted by R[G]. For example,
considering N0 : = N ∪ {0} as a monoid under addition, we readily see that R[N0 ] is
isomorphic to the polynomial ring R[ω].
It is easy to find group algebras that are domains. Say, F[Z] is such an example.
However, if a group G has an element g = 1 of finite order, then F[G] is not a
domain. Indeed, gn = 1 implies (1 − g)(1 + g + · · · + gn−1 ) = 0. The following
problem has been open for a long time.
36 2 Structure of Finite Dimensional Algebras

Problem 2.47 If a group G has no elements different from 1 of finite order, is then
F[G] a domain?
 
If G is a nontrivial group, then the set of all g∈G λg g such that g∈G λg = 0
is a proper nonzero ideal of F[G]. It is called the augmentation ideal of F[G]. The
algebra F[G] therefore is not simple. The main issue of this section is the question
whether F[G] is semiprime. We begin with the simplest example.

Example 2.48 Let G = Z2 . It is more convenient to use the multiplicative notation,

so we write G = {1, g} where g2 = 1. The group algebra F[G] thus consists of
elements of the form λ1 + μg, λ, μ ∈ F, which are multiplied according to

(λ1 + μg)(λ 1 + μ g) = (λλ + μμ )1 + (λμ + μλ )g.

It is easy to verify that λ1 + μg → (λ + μ, λ − μ) is a homomorphism from F[G]

into F × F. If char(F) = 2, then it is bijective; thus, F[G] ∼
= F × F holds in this case.
In particular, F[G] is semiprime. This is no longer true if char(F) = 2 since F(1 + g)
is then a nilpotent ideal of F[G]. Note that, in this case, F[G]
is isomorphic to the
subalgebra of M2 (F) consisting of matrices of the form α0 βα . An isomorphism is
 
given by λ1 + μg → λ+μ μ
0 λ+μ .

Thus, the group algebra F[Z2 ] is semiprime if and only if char(F) = 2. We will
now generalize this observation to group algebras over arbitrary finite groups. First
we recall some facts from linear algebra. Let V be a finite dimensional vector space,
and let T : V → V be a linear map. Choosing a basis in V , we can represent T as a
matrix relative to this basis. Different bases give rise to similar matrices. Therefore,
no matter which basis we choose, the trace of a matrix representing T is always
the same. We can therefore define tr(T ), the trace of T , as the trace of any matrix
representation of T . If T is nilpotent, then tr(T ) = 0. This can be deduced, for
example, from the Jordan normal form of a matrix representation of T , as 0 is clearly
the only eigenvalue of T . Note that here we have to use the fact that every field F has
an algebraic closure F, and then consider the matrix representation of T as a matrix
in Mn (F).
The following theorem was proved in 1898 by H. Maschke.

Theorem 2.49 (Maschke) Let G be a finite group. Then the group algebra F[G]
is semiprime if and only if either char(F) = 0 or char(F) is a prime number p that
does not divide |G|.

Proof Since F[G] is a finite dimensional vector space over F, we can define a linear
functional ρ : F[G] → F by

ρ(a): = tr(La ).
2.5 Group Algebras 37

Let n = |G| and denote the elements in G by g1 , g2 , . . . , gn where g1 = 1. Obviously,

ρ(g1 ) = n. If i ≥ 2, then
Lgi (gj ) = gi gj ∈ G \ {gj }

for every j. The matrix representation of Lgi with respect to the basis {g1 , g2 , . . . , gn }
has therefore zeros on the diagonal. Consequently, ρ(gi ) = 0. This is a crucial
observation upon which our proof is based.
Suppose now that F[G] has a nonzero nilpotent ideal I. We want to show that
then F has finite characteristic p which divides n = |G|. Pick a nonzero a ∈ I, and
write a = ni=1 λi gi . Without loss of generality we may assume that λ1 = 0. Indeed,
otherwise we choose i such that λi = 0 and replace a by gi−1 a, which is of course
also an element from I. We have

ρ(a) = λ1 ρ(g1 ) + λ2 ρ(g2 ) + · · · + λn ρ(gn ) = nλ1 .

As an element of a nilpotent ideal, a is a nilpotent element. Hence La is a nilpotent

linear map, and so ρ(a) = 0. That is, nλ1 = 0. Since λ1 = 0, this is possible only
when p = char(F) divides n. 
Conversely, assume that p = char(F) divides |G|. Set r = ni=1 gi . Since rgj =
gj r = r for every j, we see that the 1-dimensional space Fr is an ideal of F[G]. As
r 2 = |G|r = 0, Fr is a nilpotent ideal. 

The real meaning of Theorem 2.49 will become clearer later, after describing the
structure of finite dimensional semiprime algebras. This description is, in fact, our
central goal in the rest of the chapter. To this end we need several simple auxiliary
results, which are all of independent interest.

2.6 Matrix Units

We are already familiar with standard matrix units Eij . Let us introduce their abstract
generalization.

Definition 2.50 Let R be a unital ring and let n ∈ N. A set {eij ∈ R | 1 ≤ i, j ≤ n} is

called a set of n × n matrix units if

e11 + e22 + · · · + enn = 1

and
eij ekl = δjk eil

1 if j = k
for all 1 ≤ i, j, k, l ≤ n. Here, δjk stands for the “Kronecker delta”: δjk = .
0 if j = k
38 2 Structure of Finite Dimensional Algebras

Standard matrix units in R = Mn (S), where S is an arbitrary unital ring, of course

provide a basic example. This is not the only set of matrix units in R, not even when
S = F is a field. For example, if eij are matrix units and p is an invertible element,
then fij := p−1 eij p are also matrix units.
The matrix units eij with i = j are substantially different from the matrix units
eii . In particular, the eij ’s satisfy e2ij = 0 and are thus nilpotent elements, while the
eii ’s satisfy e2ii = eii .
Definition 2.51 An element e in a ring R is said to be an idempotent if e2 = e.
Idempotents e and f are called orthogonal if ef = fe = 0.
Thus, the eii ’s are pairwise orthogonal idempotents whose sum is 1. Every eii
gives rise to the subring eii Reii = {eii aeii | a ∈ R} of R, and all of these rings are
isomorphic:

eii Reii ∼
= ejj Rejj .

Indeed,

eii aeii → eji (eii aeii )eij = ejj (eji aeij )ejj

is an isomorphism. Checking this is straightforward and left as an exercise. We just

wanted to indicate why one often considers only e11 Re11 . Another instructive little
exercise is to show that the ideal generated by each matrix unit eij is the whole ring.
We have defined matrix units in an arbitrary unital ring. However, we will now
show that a ring with matrix units is in fact a full matrix ring.
Lemma 2.52 If a unital ring R contains a set of n × n matrix units eij , then R = ∼
Mn (S) where S = e11 Re11 .
Proof Define ϕ : R → Mn (e11 Re11 ) by

ϕ(a) := (aij ), where aij = e1i aej1 .

Note that aij = e11 aij e11 and so aij indeed lies in e11 Re11 . The additivity of ϕ is clear.
The (i, j) entry of ϕ(a)ϕ(b) is equal to


n 
n 
e1i aek1 e1k bej1 = e1i a ekk bej1 = e1i abej1 ,
k=1 k=1

which is the (i, j) entry of ϕ(ab). Thus, ϕ(ab) = ϕ(a)ϕ(b). If aij = 0 for all i, j,
then eii aejj = ei1 aij e1j = 0, and so a = 0 since the sum of the eii ’s is 1. Thus ϕ is
injective. Finally, observe that ϕ(ek1 ae1l ) is the matrix whose (k, l) entry is e11 ae11
and all other entries are 0, from which the surjectivity of ϕ follows. 
Remark 2.53 If R is an algebra, then the isomorphism from Lemma 2.52 is an algebra
isomorphism.
2.7 Idempotents 39

2.7 Idempotents

Let us take a brief digression and say a few general remarks about idempotents. In
what follows e will denote an arbitrary idempotent in a ring R.
As we have seen in the preceding section, the subring eRe of R may be of relevance.
We call eRe the corner ring corresponding to e. This term obviously arises from the
matrix ring example (what is eRe if R = Mn (S) and e = E11 + · · · + Ekk ?). The
corner ring eRe is unital, even if R is not; its unity is e. We shall soon see that eRe is
only one out of four subrings of R that are naturally attached to e.
For simplicity of exposition we assume, until further notice, that R is a unital ring
and that e is a nontrivial idempotent, i.e., an idempotent different from 0 and 1.
Then

f := 1 − e

is also a nontrivial idempotent, e and f are orthogonal, and their sum is 1. A model
for such a pair of idempotents, and a motivating example for what we are about to
say, are matrix units e11 and e22 in a 2 × 2 matrix ring.
Suppose ex1 e + ex2 f + fx3 e + fx4 f = 0 for some xi ∈ R. Multiplying from the
left and right by e we obtain ex1 e = 0. Similarly we see that all other terms are 0.
On the other hand, since every x ∈ R can be written as

x = exe + exf + fxe + fxf

it follows that

R = eRe ⊕ eRf ⊕ fRe ⊕ fRf . (2.4)

Here we have in mind the additive group direct sum. We call (2.4) the Peirce decom-
position of R (with respect to e). Set

R11 := eRe, R12 := eRf , R21 := fRe, R22 := fRf .

Thus R11 and R22 are corner rings corresponding to e and f , respectively, while R12
and R21 are subrings with zero multiplication. Furthermore, we have

Rij Rkl ⊆ δjk Ril

for all 1 ≤ i, j, k, l ≤ 2, which resembles calculations with matrix units. The point
here is that when having a nontrivial idempotent e in our ring, we can, with the help
of its indispensable companion f = 1 − e, mimic 2 × 2 matrices. Sometimes such an
approach is really efficient. We did not say, however, that a ring R with a nontrivial
idempotent e is necessarily isomorphic to a 2 × 2 matrix ring (over some ring). This
does hold true under the additional assumption that the equations
40 2 Structure of Finite Dimensional Algebras

exfye = e and fyexf = f

are solvable in R. Namely, then we can introduce the elements

e11 := e, e12 := exf , e21 := fye, e22 := f ,

which readily form a set of 2 × 2 matrix units of R, yielding R ∼ = M2 (eRe) by

Lemma 2.52.
There are many rings without nontrivial idempotents, for example domains, and
hence in particular division rings. As indicated above, the existence of a sole nontrivial
idempotent can already have an impact on the handling of a ring. But actually the
existence of one idempotent implies the existence of “many”, at least when the
summands eRf and fRe from the Peirce decomposition are nonzero. Namely, if e is
an idempotent, then so are e + exf and e + fxe for every x ∈ R. Also, p−1 ep is an
idempotent for every invertible p ∈ R.
Observations from the previous paragraph are meaningless if e is a central idem-
potent, i.e., an idempotent from the center Z(R) of R. In this case eRf = fRe = 0
and the Peirce decomposition reduces to two summands I := eR = eRe and
J := fR = fRf . Clearly I and J are ideals of R, R = I ⊕ J, and R ∼ = I × J via
the isomorphism x → (ex, fx). A central idempotent thus gives rise to a decomposi-
tion of a ring.
We have assumed at the beginning that R was unital, which made it possible for us
to introduce the idempotent f = 1 − e. The presence of 1 could actually be avoided
without harming the essence of the above discussion. We did not really deal with
f itself, but with products of f with elements from R. Now, if one does not assume
that R is unital and then writes x − ex instead of fx, x − ex − xe + exe instead of
fxf , etc., then most of what was said above makes sense. In particular, the Peirce
decomposition is still available, but the price we have to pay for not assuming the
existence of 1 are lengthy and messy formulas. However, the assertions concerning
central idempotents need only minor changes. The following can be extracted from
the discussion in the previous paragraph: If e is a central idempotent in R, then I = eR
and J = {x − ex | x ∈ R} are ideals of R such that

R=I ⊕J ∼
= I × J.

Considering I as a ring, we see that e is its unity. Ideals are rarely unital rings. In
fact, only those that are generated by central idempotents are.
Lemma 2.54 Let I be an ideal of a ring R. If I is a unital ring, then its unity e is a
central idempotent in R, I = eR, and there exists an ideal J of R such that R = I ⊕ J.
Moreover, R ∼
= I × J.
Proof Since e ∈ I, we have eR ⊆ I, and conversely, I = eI ⊆ eR. Thus I = eR. Next,
since ex, xe ∈ I for every x ∈ R, we have ex = (ex)e and xe = e(xe). Consequently,
ex = xe, and so e is a central idempotent. Now we refer to the discussion before the
lemma. 
2.7 Idempotents 41

Again assuming that R is unital, we can state a kind of converse to Lemma 2.54:
If I and J are ideals of R such that R = I ⊕ J, then there exists a central idempotent
e ∈ R such that I = eR and J = (1 − e)R (so I and J are unital rings). The proof
is easy. Just write 1 = e + f with e ∈ I, f ∈ J, and check that e has the desired
properties.

2.8 Minimal Left Ideals

When considering one-sided ideals we will usually give preference to the left ones.

Definition 2.55 A left ideal L of a ring R is called a minimal left ideal if L = 0

and L does not properly contain a nonzero left ideal of R.

Minimal right ideals and minimal two-sided ideals are defined analogously.

Example 2.56 The only minimal left ideal of a division ring D is D itself.

Example 2.57 Let R = Mn (D), D a division ring, and let L be the set of matrices
in R that have arbitrary entries in the ith column and zeros in all other columns. It
is an easy exercise to show that L is a minimal left ideal of R. We also remark that
L = REii where Eii is the standard matrix unit, and that Eii REii = {dEii | d ∈ D} is a
division subring of R isomorphic to D.

Lemma 2.58 If L is a minimal left ideal of a semiprime ring R, then there exists an
idempotent e ∈ R such that L = Re and eRe is a division ring.

Proof Since R is semiprime, there exist x, y ∈ L such that xy = 0. In particular,

Ly = 0. But Ly is a left ideal of R contained in L, so Ly = L because of the
minimality of L. Accordingly, there exists e ∈ L such that ey = y. Hence it follows
that e2 − e belongs to the set J := {z ∈ L|zy = 0}. Clearly, J is again a left
ideal of R contained in L. Since x ∈ L \ J, this time we conclude that J = 0. In
particular, e2 = e. As e ∈ L, we have Re ⊆ L, and since 0 = e ∈ Re it follows
from the minimality assumption that L = Re. Now consider the corner ring eRe.
Let a ∈ R be such that eae = 0. We must prove that eae is invertible in eRe. We
have 0 = Reae ⊆ Re = L, and so Reae = L. Therefore beae = e holds for some
b ∈ R, and hence also (ebe)(eae) = e. Since ebe is a nonzero element in eRe, by
the same argument there exists c ∈ R such that (ece)(ebe) = e. But a left inverse
coincides with a right inverse, so eae = ece is invertible in eRe, with ebe being
its inverse. 

A similar statement of course holds for minimal right ideals. Also, without any
change in the proof we see that the lemma holds for algebras as well. Needless to
say, minimal left ideals of algebras are defined in the same way as minimal left ideals
of rings.
42 2 Structure of Finite Dimensional Algebras

Semiprime rings with minimal left ideals thus contain particularly nice corner
rings. Unfortunately, the existence of minimal left ideals in rings is an exception rather
than a rule. However, in nonzero finite dimensional algebras they exist for obvious
reasons—just take any nonzero left ideal of minimal dimension. The following result
thus follows immediately from (the algebra version of) Lemma 2.58.

Corollary 2.59 If A is a nonzero finite dimensional semiprime algebra, then there

exists an idempotent e ∈ A such that eAe is a division algebra.

Corollary 2.59 is exactly what we will need in the next section. However, it would
be inappropriate to stop the discussion on minimal left ideals at this point without
mentioning the converse of Lemma 2.58.

Corollary 2.60 The following statements are equivalent for an idempotent e in a

semiprime ring R:
(i) eRe is a division ring.
(ii) Re is a minimal left ideal of R.
(iii) eR is a minimal right ideal of R.

Proof Since (i) is, unlike (ii) and (iii), a left-right symmetric condition, it suffices
to prove that it is equivalent to (ii). As we know that (ii) implies (i) by (the proof
of) Lemma 2.58, we must only show that (i) implies (ii). Thus, assume that eRe is a
division ring, and take a left ideal I of R such that 0 = I ⊆ Re. It is enough to show
that e ∈ I, as this clearly yields I = Re. Pick 0 = u ∈ I. By the semiprimeness of R
we have uru = 0 for some r ∈ R. Note that u = ue for u ∈ I ⊆ Re. Consequently,
eru = erue is a nonzero element of eRe, and hence there exists v ∈ R such that
(eve)(eru) = e. Therefore e ∈ Ru ⊆ I. 

We will continue the study of minimal left ideals later, in Sect. 5.4.

2.9 Wedderburn’s Structure Theorems

In 1907, Wedderburn laid down the foundations of the structure theory of noncom-
mutative algebras. Equipped with simple lemmas from the previous sections we are
now in a position to prove Wedderburn’s theorems. The first one characterizes finite
dimensional simple algebras.

Theorem 2.61 (Wedderburn) Let A be a nonzero finite dimensional algebra. The

following statements are equivalent:
(i) A is prime.
(ii) A is simple.
(iii) There exist n ∈ N and a division algebra D such that A ∼
= Mn (D).
2.9 Wedderburn’s Structure Theorems 43

Proof We already know that (iii) implies (ii) (Example 1.10), and trivially (ii) implies
(i). Thus we only have to prove that (i) implies (iii). Let us first do this under the
additional assumption that A is unital. The proof is by induction on d := [A : F].
If d = 1, then we simply take n = 1 and D = F. Thus, let d > 1. By Corollary 2.59
there exists an idempotent e ∈ A such that eAe is a division algebra. If e = 1, then the
desired result holds (with n = 1). Assume therefore that e is a nontrivial idempotent.
Set f := 1 − e. Note that fAf is a nonzero prime algebra with unity f . Since e does not
belong to fAf , we have [ fAf : F] < d. The induction assumption implies that fAf is
isomorphic to the algebra of m×m matrices over some division algebra. Accordingly,
fAf contains matrix units eij , i, j = 1, . . . , m, such that e11 fAfe11 = e11 Ae11 is a
division algebra. Our goal is to extend the matrix units  of fAf to matrix units of A.
We begin by setting n := m + 1 and enn := e. Then ni=1 eii = f + e = 1, and of
course enn eij = eij enn = 0 for all i, j < n. It remains to find ein and eni , i ≤ n − 1
(informally: the last row and the last column of matrices from our algebra are still
“unknown”, except for their intersection).
Let us first find e1n and en1 . This is the heart of the proof. Using the primeness
twice we see that e11 aenn a e11 = 0 for some a, a ∈ A. As e11 Ae11 is a division
algebra with unity e11 , there is a ∈ A such that

(e11 aenn a e11 )(e11 a e11 ) = e11 .

Setting e1n := e11 aenn and en1 := enn a e11 a e11 we thus have

e1n en1 = e11 .

Since en1 ∈ enn Ae11 , we have en1 = enn en1 and en1 = en1 e11 = en1 e1n en1 . Compar-
ing both relations we get
(en1 e1n − enn )en1 = 0.

The element en1 e1n − enn lies in the division algebra enn Aenn . If it was not zero, then
we could multiply the last identity from the left-hand side by its inverse, leading to
a contradiction 0 = enn en1 = en1 . Therefore

en1 e1n = enn .

Finally, we define enj := en1 e1j and ejn := ej1 e1n for j = 2, . . . , n − 1. Note that
eij = ei1 e1j and e1j ek1 = δjk e11 holds for all i, j, k = 1, . . . , n. Consequently, for
all i, j, k, l = 1, . . . , n we have

eij ekl = ei1 e1j ek1 e1l = δjk ei1 e11 e1l = δjk ei1 e1l = δjk eil .

Thus eij , i, j = 1, . . . , n, are indeed matrix units of A. Lemma 2.52 (together with
Remark 2.53) yields the desired conclusion A ∼ = Mn (D), where D = e11 Ae11 .
It remains to show that (i) implies (iii) without the assumption that A is unital.
Suppose A is prime and not unital. Then A is a prime unital algebra by Lemma 2.36.
44 2 Structure of Finite Dimensional Algebras

Since (i) implies (iii) (and hence (ii)) for unital algebras, it follows that A is simple.
However, this is a contradiction since A is a proper nonzero ideal of A . 

We can add to Theorem 2.61 that D is, of course, also finite dimensional, and that
[A : F] = n2 [D : F]. Let us also mention that the classical version of Theorem 2.61
does not involve prime algebras. Usually it is stated as

A is simple ⇐⇒ A ∼
= Mn (D).

However, for our approach to the structure theory the concept of primeness is more
suitable than the concept of simplicity. At any rate, it is interesting that finite dimen-
sional prime algebras coincide with simple ones. In general, the class of prime alge-
bras is much larger than the class of simple algebras.
Let us also remark that the restriction to finite dimensions in Theorem 2.61 is
absolutely necessary. Just think of the Weyl algebra, which is a simple domain but
not a division algebra (see Example 2.28).

Corollary 2.62 A finite dimensional algebra A is a central simple algebra if and

only if there exist n ∈ N and a central division algebra D such that A ∼
= Mn (D).

Proof Apply Theorem 2.61 and Lemma 1.15. 

Corollary 2.63 The dimension of a finite dimensional central simple algebra is a

perfect square.

Proof We have [Mn (D) : F] = n2 [D : F]. Now use Corollary 1.37. 

The direct product of prime algebras is a semiprime algebra (cf. Example 2.33).
We will now show that in the finite dimensional setting every semiprime algebra is
built in this way.

Theorem 2.64 (Wedderburn) Let A be a nonzero finite dimensional algebra. Then

A is semiprime if and only if there exist n1 , . . . , nr ∈ N and division algebras
D1 , . . . , Dr such that A ∼
= Mn1 (D1 ) × · · · × Mnr (Dr ).

Proof We only have to prove the “only if” part. We proceed by induction on d =
[A : F]. If d = 1, then A = Fa with a2 = 0. Thus a2 = λa, λ = 0, so that
λ−1 a is the unity of A. Hence A ∼ = F. Let d > 1. If A is prime, then the result
follows from Theorem 2.61. We may therefore assume that there exists 0 = a ∈ A
such that I = {x ∈ A | aAx = 0} is a nonzero set. Clearly, I is an ideal of A, and
therefore a semiprime algebra (cf. Remark 2.25). Since a ∈ / I we have [I : F] < d.
The induction assumption implies that I ∼ = Mn1 (D1 ) × · · · × Mnp (Dp ) for some
ni ∈ N and division algebras Di , i = 1, . . . , p. As each factor Mni (Di ) has a unity,
so does I. By Lemma 2.54 there is an ideal J of A such that A ∼= I ×J. We may use the
induction assumption also for J and conclude that J ∼ = Mnp+1 (Dp+1 )× · · · ×Mnr (Dr )
for some ni ∈ N and division algebras Di , i = p + 1, . . . , r. The result now clearly
follows. 
2.9 Wedderburn’s Structure Theorems 45

Theorem 2.64 in particular shows that a nonzero finite dimensional semiprime

algebra is automatically unital.
A more common name for a finite dimensional semiprime algebra is a (finite
dimensional) semisimple algebra. Accordingly, a more standard wording of
Theorem 2.64 is

A is semisimple ⇐⇒ A ∼
= Mn1 (D1 ) × · · · × Mnr (Dr ).

Semisimple rings form another class of rings, which we will meet later. In the finite
dimensional context, semisimple algebras coincide with semiprime ones.
If a finite dimensional algebra A has nonzero nilpotent ideals, then we can factor
out its radical and thereby obtain a semiprime algebra (Lemma 2.26). This yields
the following result which gives a more complete picture of the structure of finite
dimensional algebras.

Theorem 2.65 Let A be a finite dimensional algebra. If N is its radical (i.e., the
maximal nilpotent ideal), then

A/N ∼
= Mn1 (D1 ) × · · · × Mnr (Dr )

for some n1 , . . . , nr ∈ N and division algebras D1 , . . . , Dr , unless A = N is a

nilpotent algebra.

Actually, more can be said. If A is an algebra over a perfect field F (fields of

characteristic 0, algebraically closed fields, and finite fields are all examples of perfect
fields), then there exists a subalgebra A of A such that A ∼ = A/N ∼ = Mn1 (D1 ) ×
· · · × Mnr (Dr ) and A = A ⊕ N, the vector space direct sum. This result is known
as Wedderburn’s principal theorem. We omit the proof. The interested reader can
find it, for instance, in [Row91, pp. 163–164]. For a nice illustration of the theorem
examine the case where A = Tn (F), cf. Examples 2.14 and 2.27.
We will give additional information on Wedderburn’s structure theorems in
Sects. 3.9, 5.3, and 5.4.

2.10 Algebras Over Special Fields

The purpose of this section is to restate Wedderburn’s theorems for algebras over
some special fields. We begin with algebraically closed fields, for which these theo-
rems take on extremely simple forms. The following result was proved (for F = C)
already in 1892 by T. Molien.

Corollary 2.66 (Molien) Let A be a nonzero finite dimensional semiprime algebra

over an algebraically closed field F. Then A ∼ = Mn1 (F) × · · · × Mnr (F) for some
n1 , . . . , nr ∈ N. Moreover, if A is prime then r = 1.

Proof Use Theorems 2.61 and 2.64 along with Proposition 1.6.
46 2 Structure of Finite Dimensional Algebras

In the next result we will indicate the usefulness of Corollary 2.66 by combining
it with Maschke’s theorem. Given a finite group G, we can form the group algebra
F[G] for any field F. Understanding F[G] can give us a clue for understanding G.
We have a freedom in selecting the field F. In view of Corollary 2.66, choosing F to
be algebraically closed seems to be a good idea. Maschke’s theorem (Theorem 2.49)
suggests that by requiring that char(F) = 0 we will be on the safe side. An obvious
choice now is F = C.

Corollary 2.67 Let G be a finite group. Then there exist n1 , . . . , nr ∈ N such that
C[G] ∼
= Mn1 (C) × · · · × Mnr (C).

Thus, starting with an abstract finite group, we were able to represent it through
its group algebra as a “concrete” object, the direct product of full matrix algebras.
Corollary 2.67 is a basic result in the representation theory of finite groups. A rep-
resentation of a group G is a group homomorphism from G into the general linear
group, i.e., the group of all invertible n × n matrices.

Example 2.68 Consider the group algebra C[S3 ], where S3 is the symmetric
 group
of {1, 2, 3}. Representing it as Mn1 (C) × · · · × Mnr (C), we must have ri=1 ni2 = 6,
the dimension of C[S3 ]; moreover, at least one ni must be different from 1 since S3
is noncommutative. The only possibility is that C[S3 ] ∼ = M2 (C) × C × C.

In the next results we confine ourselves to simple algebras.

Corollary 2.69 A finite dimensional simple R-algebra A is isomorphic either to

Mn (R), Mn (C), or Mn (H) for some n ∈ N.

Proof Combine Theorem 2.61 with Theorem 1.4. 

With reference to Corollary 2.62 we get

Corollary 2.70 A finite dimensional central simple R-algebra A is isomorphic either

to Mn (R) or Mn (H) for some n ∈ N.

Corollary 2.71 A finite simple ring R is isomorphic to Mn (F) for some n ∈ N and
some finite field F.

Proof As noticed in Remark 2.20, char(R) is a prime number p and we can consider
R as a Zp -algebra. Therefore Theorem 2.61 tells us that R ∼= Mn (D) where D is a
finite dimensional division algebra over Zp . By Theorem 1.38, D is a field. 

Wedderburn’s theorems reduce the problem of classifying finite dimensional

(semi)prime F-algebras to classifying finite dimensional division F-algebras. The
latter is fully understood if F is one of the fields considered in this section. For
some fields F, however, finding all finite dimensional division F-algebras may be an
extremely difficult problem and their explicit list cannot be given.
2.11 Scalar Extension (A Naive Approach) 47

2.11 Scalar Extension (A Naive Approach)

In most of the above sections the field F has played just a formal role in our study
of F-algebras. In the previous section, however, we saw that F is of essential impor-
tance if we wish to describe algebras precisely. We now aim to present a simple
construction which enables replacing the given field F by a “better” field for which
the structure theory yields definitive results. The standard way for introducing this
construction requires the concept of the tensor product. Following the elementary
style of this chapter we will use an alternative way which has many disadvantages
and one advantage: it is very simple and intuitively clear. It is our hope that a naive
approach used here will help the reader to easily understand (and appreciate) the
“correct” definition given later, in Sect. 4.7.
Let us begin with an illuminating example.

Example 2.72 Consider the algebra HC of “complex quaternions”, i.e., HC is the

4-dimensional C-algebra with basis {1, i, j, k} satisfying the same multiplication
rules (1.1) as the standard real quaternions H (cf. Remark 1.7). That is, we keep the
multiplication table of H, but change the scalars from R to C. Every complex algebra
can be canonically considered as a real algebra (of dimension 2n if it is n-dimensional
as a complex algebra). In this sense we may regard H as an R-subalgebra of HC . As
a sum of squares of nonzero complex numbers can easily be 0, the proof that H is
a division algebra does not work for HC . Moreover, indirectly it shows that HC is
not a domain. Let us determine what HC really is. We shall write i for the complex
imaginary unit (to distinguish it from i ∈ HC ). It is straightforward to check that the
elements
1 − ii j − ik −j − ik 1 + ii
e11 = , e12 = , e21 = , e22 = .
2 2 2 2

form a set of 2 × 2 matrix units of HC , and that e11 HC e11 = Ce11 ∼

= C. Using
Lemma 2.52 we may now conclude that

HC ∼
= M2 (C).

Considering the restriction of this isomorphism to H we see

that H is isomorphic to
z w
the real algebra of all complex matrices of the form −w z .

One might object that by changing the real scalars into the complex ones we did
not make the quaternions nicer. On the contrary, we have spoiled the property of
being a division algebra. However, our aim was to indicate a general principle that
is hidden behind this example. Let us consider another, admittedly less appealing, in
fact quite trivial example.

Example 2.73 The standard matrix units E11 , E12 , E21 , E22 form the standard basis
of M2 (R). Again substituting the scalars we can consider the C-algebra M2 (R)C
48 2 Structure of Finite Dimensional Algebras


consisting of elements 2i,j=1 λij Eij , λij ∈ C, whose multiplication is determined by
Eij Ekl = δjk Eil . It is obvious that

M2 (R)C ∼
= M2 (C).

Let us summarize what has been inferred in these two examples. From
Corollary 2.69 we see that up to isomorphism there are exactly two 4-dimensional
simple real algebras: H and M2 (R) (they both happen to be central). If we take the
standard basis of any of them and introduce the algebra that has the same basis and
the same multiplication table, only the scalars are changed from R to C, then this
new algebra is isomorphic to M2 (C).
Now we proceed to the abstract situation. Take an algebra A over F. For sim-
plicity assume that it is finite dimensional. Choose a basis {e1 , . . . , ed } of A with
multiplication table


d
ei ej = αijk ek . (2.5)
k=1

The constants αijk ∈ F completely determine the multiplication in A. Let K be a

field extension of F. We now introduce the K-algebra AK which has the same basis
{e1 , . . . , ed } and whose multiplication is determined by the same formula (2.5) (this
makes sense since αijk ∈ F ⊆ K). Thus, we only change the scalars but keep the
operations. The associativity of the multiplication in AK is a consequence of the
associativity of the products of the ei ’s. We call AK the scalar extension of A to K.
Since F is a subfield of K, we can canonically consider AK as an F-algebra and
A as its F-subalgebra. Let us also point out that

[AK : K] = [A : F]. (2.6)

Our definition of scalar extension depends on a basis. It would be formally more

correct to denote it by something like AK,{e1 ,...,ed } instead of just AK . However, let us
put aside a natural question of what happens if we take another basis, and stick with
the simplified notation AK . The “correct” definition that will be given in Sect. 4.7 is
independent of basis, and when we reach that point everything should be clear. Let
us now only concentrate on a generalization of our observations on 4-dimensional
(central) simple real algebras.

Theorem 2.74 Let A be a finite dimensional central√simple F-algebra. If F is an

algebraic closure of F, then AF ∼
= Mn (F) where n = [A : F].

Proof Let us show that AF is simple. Take a nonzero ideal I of AF , and pick a nonzero
u = λ1 e1 + · · · + λd ed ∈ I. We may assume that λ1 = 0. Let ϕ ∈ EndF (A) be such
that ϕ(e
 1 ) = 1 and ϕ(ei ) = 0, i ≥ 2. By Lemma 1.25 there exist ai , bi ∈ A such that
ϕ = ri=1 Lai Rbi . Therefore
2.11 Scalar Extension (A Naive Approach) 49


r 
d 
r  d
I ai ubi = λj ai ej bi = λj ϕ(ej ) = λ1 .
i=1 j=1 i=1 j=1

Thus I contains an invertible element, so that I = AF . This means that AF is simple.

From Corollary 2.66 it follows that AF ∼= Mn (F) for some n ∈ N; (2.6) tells us that
[A : F] = n2 . 

Theorem 2.74 can be used to reduce various problems on finite dimensional cen-
tral simple algebras to matrix algebras over fields (even better, over algebraically
closed fields). A quick illustration: Theorem 2.74 immediately yields a new proof of
Corollary 2.63.

Exercises

2.1. Describe the radical of the subalgebra of Mn (F) consisting of all matrices
whose last k rows are zero.
2.2. Find an ideal of the algebra Πn≥2 Tn (F) that is nil but not nilpotent.
2.3. Let R be an arbitrary ring. Show that:
(a) The left ideal generated by a ∈ R is nil if and only if the right ideal
generated by a is nil.
(b) The sum of all nil left ideals of R is equal to the sum of all nil right ideals
of R, and is therefore an ideal of R.
(c) The sum of a nil left ideal of R and a nil ideal of R is nil.
(d) The sum of all nil ideals of R, Nil(R), is a nil ideal of R, and the ring
R/Nil(R) has no nonzero nil ideals.
(e) If R/Nil(R) is commutative (in particular, if R itself is commutative), then
Nil(R) is equal to the set of all nilpotent elements in R.
Remark: We call Nil(R) the upper nilradical of R. There is also the notion of
the lower nilradical (also called the prime radical) of R which is defined as
the intersection of all prime ideals of R, i.e., ideals with the property that the
corresponding factor rings are prime. It can be shown that the lower nilradical
is a nil ideal, and hence contained in Nil(R), and that a ring is semiprime if
and only if its lower nilradical is zero.
2.4. Prove that the following statements are equivalent:
(i) Köthe’s problem has a positive answer.
(ii) The upper nilradical of an arbitrary ring contains all nil one-sided ideals
of this ring.
(iii) The sum of two nil left ideals of an arbitrary ring is nil.
2.5. Show that a prime ring without a nonzero nilpotent element is a domain.
50 2 Structure of Finite Dimensional Algebras

2.6. Show that a semiprime ring R is prime if and only if the intersection of any
two nonzero ideals of R is nonzero. Give an example showing that this is not
always true if R is not semiprime.
2.7. Show that a ring R is prime (resp. semiprime) if and only if Mn (R) is prime
(resp. semiprime).
2.8. Let a, b be elements in a ring R. Note that aRb = 0 implies bRa = 0 in case R
is semiprime, and give an example showing that this is not true in every ring R.
2.9. Let R be a prime ring. Show that if a ∈ R commutes with every element from
a nonzero left ideal L of R, then a ∈ Z(R). Is this still true if R is semiprime?
2.10. Let R be a semiprime ring, and let a ∈ R be such that axa = a2 x for every
x ∈ R. Show that a ∈ Z(R).
2.11. Find an example of a commutative domain without unity such that its unitiza-
tion is not a domain.
Remark: Lemma 2.36 thus does not hold for rings.
2.12. Show that a unital F-algebra B is isomorphic to the unitization A of some
algebra A if and only if there exists a nonzero algebra homomorphism from B
into F. Notice that every group algebra F[G] is an example of such an algebra.
2.13. Let A be a nonzero finite dimensional unital algebra over a field F with
char(F) = 0, and let ϕ be an automorphism of A. Show that ϕ(a) = a + 1 for
every a ∈ A.
Hint: Express Lϕ(a) through ϕ, ϕ −1 , and La .
2.14. Let A be a unital algebra, and let G be a subgroup of the multiplicative group
A∗ . Explain why, in general, the group algebra of G is not isomorphic to the
subalgebra of A generated by G (which is just the linear span of G, so the
elements in both algebras look the same), and find an example (with G = {1})
where these two algebras are isomorphic.
2.15. Show that every group algebra F[G] has a (linear) involution.
2.16. Show that 0 is the only idempotent in a domain without unity.
2.17. Let e1 , e2 be idempotents in an F-algebra A. Note that if char(F) = 2, then e1 +
e2 is an idempotent if and only if e1 and e2 are orthogonal. Show that this can
be extended to any number of idempotents in case A is finite dimensional and
char(F) = 0; that is, the sum of idempotents e1 , . . . , en ∈ A is an idempotent
if and only if e1 , . . . , en are pairwise orthogonal.
Hint: The trace of an idempotent matrix is equal to its rank.
2.18. Show that the subring generated by all idempotents in a ring R contains the
ideal generated by all commutators [e, x] where e is an idempotent and x is an
arbitrary element in R. Hence conclude that a simple ring containing at least
one nontrivial idempotent is generated (as a ring) by its idempotents.
Hint: Express [e, x] as a difference of two idempotents (an inspection of
Sect. 2.7 should give an idea how this can be done), and use the identity
yu = [e, y[e, u]] − [e, y][e, u] where u = [e, x].
Exercises 51

2.19. Show that a unital ring R contains an element which is right invertible but
not left invertible if and only if M2 (R) contains an invertible upper triangular
matrix whose inverse is not upper triangular.
Hint: If ab = 1, then e := 1 − ba is an idempotent and ae = eb = 0.
2.20. Prove that a nonzero idempotent e in a ring R cannot be written as a sum of
two nonzero orthogonal idempotents in R if and only if there does not exist
an idempotent f ∈ R different from 0 and e such that f = ef = fe. Such an
idempotent e is said to be primitive. Show that a sufficient condition for e
being primitive is that Re is a minimal left ideal of R, and that this condition
is also necessary if R = A is a finite dimensional semiprime algebra.
2.21. Let R1 , . . . , Rn be unital rings. Show that every ideal of R1 × · · · × Rn is
of the form I1 × · · · × In where Ii is an ideal of Ri . Hence conclude that the
number of ideals of a finite dimensional semiprime algebra is a power of 2.
2.22. Let A be a nonzero finite dimensional algebra. Show that:
(a) A is nil if and only if A is nilpotent.
(b) A is a domain if and only if A is a division algebra.
(c) A has no nonzero nilpotent elements if and only if A is isomorphic to a
direct product of division algebras.
(d) There exists n ∈ N such that x n = x for every x ∈ A if and only if A is
isomorphic to a direct product of finite fields.

Remark: Later, in Sect. 5.10, we will consider the condition x n = x in arbitrary

rings, even under the assumption that n may depend on x.
2.23. Show that F[Zn ] ∼ = F[ω]/(ωn − 1) for every field F. Use this to find the
Wedderburn decomposition (in the sense of Theorem 2.64) of R[Z3 ].
Remark: This can be also used for describing C[Z3 ]; however, Corollary 2.67
readily implies that C[G] ∼ = C × · · · × C for every finite abelian group G.
2.24. The subset Q8 := {1, −1, i, −i, j, −j, k, −k} of H is clearly a group under
multiplication. It is called the quaternion group. Find the Wedderburn decom-
position of C[Q8 ].
Hint: What is the codimension of the augmentation ideal of a group algebra?
2.25. Let A be a finite dimensional central simple F-algebra such that [x, y][z, w] +
[z, w][x, y] ∈ F for all x, y, z, w ∈ A. Prove that either A = F or [A : F] = 4.
Chapter 3
Modules and Vector Spaces

Modules at last! In most books on a similar subject modules appear on the very
first pages. And there are good reasons for this. Ring theory and module theory are
symbiotic. Many questions about rings can be handled most effectively by using
modules. Moreover, sometimes modules seem to be inevitable. In spite of that, we
have decided to postpone introducing modules up until now. Our viewpoint has
been that an intrinsic approach from the first chapters, where results on rings and
algebras were rather quickly derived from basic definitions, should be easier and
more appealing for a newcomer to the subject. We have now reached the point where
continuing without modules could be inefficient.
This chapter is mostly of a preliminary nature. Several fundamental module-
theoretic notions will be considered at an introductory level. The emphasis will be
on vector spaces and simple modules. Compared with the previous two chapters,
this one lacks striking theorems and surprising proofs. An exception is perhaps a
supplement to Wedderburn’s structure theory, given close to the end of the chapter.

3.1 Concept of a Module

Being familiar with vector spaces, the definition of a module can be easily memorized.
One just replaces the role of a field by an arbitrary ring.
Definition 3.1 Let R be a ring. A left R-module (or left module over R) is an additive
group M together with a map from R×M to M, (r, m) → rm, such that for all r, s ∈ R,
m, n ∈ M:
(a) (r + s)m = rm + sm,
(b) r(m + n) = rm + rn,
(c) r(sm) = (rs)m.
If, additionally, R is a unital ring and
(d) 1m = m
for every m ∈ M, then M is said to be a unital left R-module.
© Springer International Publishing Switzerland 2014 53
M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_3
54 3 Modules and Vector Spaces

A (unital) right R-module is defined similarly, only the order of elements from R
and M has to be interchanged. That is, one requires the obvious analogues of (a)–(d)
for a map (m, r) → mr, where m ∈ M and r ∈ R.
Remark 3.2 If R is commutative, then every left R-module M becomes a right R-
module by defining mr := rm for all m ∈ M, r ∈ R.
Distinguishing between left and right R-modules is therefore more substantial
when R is noncommutative. This does not mean, of course, that the theory of left R-
modules is any different from the theory of right R-modules. Results on left modules
have obvious analogues for right modules, and vice versa. Among two equivalent
choices, let us decide to work with the left version. Thus, unless specified otherwise,
we will henceforth adopt the following convention:

R-module := left R-module.

When it will be clear which ring R we have in mind, or when R will play just a formal
role in our discussion, we will simply use the term “module”.
Modules appear all over the place; finding examples is easy, making a selection
is harder.
Example 3.3 A vector space over a field F is a unital F-module.
Example 3.4 Every additive group M is a unital Z-module, with nm, where n ∈ Z
and m ∈ M, having the usual meaning.
Example 3.5 Let M be an additive group. The set R = End(M) of all endomorphisms
of M (i.e., all additive maps from M into M) becomes a ring if we define the sum
f + g and the product fg of f , g ∈ R as follows:

( f + g)(m) := f (m) + g(m), ( fg)(m) := f (g(m)). (3.1)

We can view M as a unital R-module via fm := f (m), f ∈ R, m ∈ M.

Example 3.6 A left ideal L of a ring R is a left R-module with r, where r ∈ R and
 ∈ L, simply being the ordinary product in R. In particular, R itself is an R-module.
Similarly, a right ideal of R can be considered as a right R-module, while an
ideal (in particular, R itself) is both a left and a right R-module. Moreover, it is an
R-bimodule:
Definition 3.7 Let R and S be rings. If M is both a left R-module and a right S-
module, and also

(rm)s = r(ms) for all r ∈ R, m ∈ M, s ∈ S,

then M is said to be an (R, S)-bimodule. An (R, R)-bimodule is also called an R-

bimodule.
3.1 Concept of a Module 55

Example 3.8 If R is a subring of a ring R , then R is an R-bimodule (via the ordinary

product in R ).

Replacing rings by algebras one has to make some obvious modifications in the
above definitions (as well as in examples). For instance, a (left) A-module, where
A is an algebra over a field F, is defined as a vector space V over F together with a
map A × V → V , (a, v) → av, that satisfies (a), (b), (c), and

(λa)v = a(λv) = λ(av) for all λ ∈ F, a ∈ A, v ∈ V .

In other words, the map must be bilinear, not only biadditive.

Example 3.9 The n-dimensional space F n can be viewed as a unital Mn (F)-module

if we write vectors in F n as columns and define the module operation as the standard
product of a matrix with a column vector.

Example 3.10 A vector space V over F is a unital EndF (V )-module.

Note that Example 3.10 is a refined version of Example 3.5 which also includes
the scalar multiplication, and that Example 3.9 is just a disguised form of its finite
dimensional version. The relevance of these examples will become evident from the
following discussion.
Let A be an algebra over F and let V be a vector space over F. An algebra homo-
morphism from A to EndF (V ) is called a representation of A. If ϕ : A → EndF (V )
is a representation, then V becomes an A-module by defining av := ϕ(a)(v). Con-
versely, every A-module V gives rise to a representation ϕ : A → EndF (V ) given
by ϕ(a)(v) := av. We may therefore regard representations and modules as equiv-
alent concepts. For example, instead of with the regular representation we can deal
with the (left) A-module A; thus, in a hidden manner modules were already treated
before. We will usually present our results in the language of modules rather than
in that of representations. Still, being aware of a possible interpretation in terms of
representations is important for understanding the philosophy of the subject. Con-
sidering a module over an algebra is essentially the same as representing algebra
elements as linear operators of a vector space. This indicates why modules can be
useful for the structure theory. Namely, they can be used to identify some abstract
algebras with algebras of linear operators. In the previous chapter we have shown
that some abstract algebras can be identified with algebras of matrices. But from
linear algebra we know that the two seemingly quite different notions, matrices and
linear operators, are basically equivalent. Thus, as it turns out, constructing matrix
units, the approach taken in the previous chapter, is not the only way to describe
algebra elements as matrices—one can rely on modules instead. Perhaps this sounds
very vague at this point, especially since the matrices from the previous chapter were
not the “ordinary” matrices we know from linear algebra, but matrices with entries
from a division algebra. However, as will become apparent in Sect. 3.4, this is not
really a problem. In Chap. 5 we will show that the module-theoretic approach to the
structure theory yields a profound generalization of Wedderburn’s theory.
56 3 Modules and Vector Spaces

Similar observations as for modules over algebras can be made for modules over
rings: an R-module M can be identified with a ring homomorphism from R into
End(M), the ring of endomorphisms of the additive group of M. Its kernel will later
turn out to be important, so we give it a name.

Definition 3.11 The annihilator of an R-module M is the set

annR (M) := {r ∈ R | rM = 0}.

If annR (M) = 0, then M is said to be a faithful module (here, 0 stands for {0} and
rM for {rm | m ∈ M}).

A glance at the definition shows that annR (M) is a left ideal, but actually it is—as
a kernel of a ring homomorphism—a two-sided ideal.

Example 3.12 Consider the abelian group Zn as a Z-module. Then annZ (Zn ) = nZ.

Example 3.13 The modules from Examples 3.9 and 3.10 are faithful.

3.2 Basic Module-Theoretic Notions

Even the readers that have never met modules before might experience a sense of
déjà vu when reading this section. Basic notions of module theory are very similar
to those one encounters in the first meetings with vector spaces, groups, and rings.
We will therefore introduce them in a rather quick and perhaps slightly dry manner.
The reader is advised to translate these notions into the (presumably more familiar)
language of vector spaces.
For simplicity of exposition we will consider only modules over rings. It should
be obvious how to make the necessary adjustments for modules over algebras. One
just has to require the natural scalar multiplication properties at appropriate places.
Throughout this section, R stands for a ring and M for an R-module. Recall that
a module means a left module by our convention.

Definition 3.14 A subset L of M is called a submodule of M if L is an additive

subgroup of M and r ∈ L for all r ∈ R,  ∈ L.

Of course, a submodule is itself a module.

Example 3.15 Every module M has at least two submodules: 0 and M. Submodules
different from M are said to be proper, and submodules different from 0 are said to
be nonzero or nontrivial.

Example 3.16 If we consider an additive group as a Z-module, then its submodules

are exactly its subgroups. Similarly, viewing a vector space over F as an F-module,
we see that its submodules are its subspaces.
3.2 Basic Module-Theoretic Notions 57

Example 3.17 The submodules of the R-module R are the left ideals of R. This
simple fact will be used frequently.
Example 3.18 Let J be a left ideal of R. The additive subgroup of M generated by
the elements of the form um, where u ∈ J and m ∈ M, is a submodule. We denote it
by JM.
The intersection of submodules is obviously a submodule. Given a subset X of
M, the intersection of all submodules containing X is therefore the submodule gen-
erated by X. Explicitly, it can be described as the additive subgroup generated by the
elements rx + nx, r ∈ R, n ∈ Z, x ∈ X (there is no need to involve nx if M is unital).
If a (sub)module N is generated by some finite set, then we say that N is finitely
generated. Modules generated by only one element are called cyclic. A unital cyclic
module can be written as

Rm := {rm | r ∈ R},

where m is its generator.

We proceed to another fundamental notion.
Definition 3.19 Let N be another R-module. A map ϕ : M → N is an R-module
homomorphism if

ϕ(m + m ) = ϕ(m) + ϕ(m ) and ϕ(rm) = rϕ(m)

for all r ∈ R and m, m ∈ M.

The kernel of ϕ, ker ϕ = {m ∈ M | ϕ(m) = 0}, is clearly a submodule of M,
and the image of ϕ, im ϕ = {ϕ(m) | m ∈ M}, is a submodule of N. An R-module
isomorphism is, of course, a bijective R-module homomorphism. We say that M
and N are isomorphic modules, and write M ∼ = N, if there exists an isomorphism
ϕ : M → N. It is immediate to check that in this case its inverse ϕ −1 : N → M is
also an isomorphism. An R-module endomorphism is an R-module homomorphism
from M into itself, and an R-module automorphism is an endomorphism which is
also an isomorphism.
We define the endomorphism ring

EndR (M)

as the set of all R-module endomorphisms of M endowed with operations defined as

in (3.1), i.e., pointwise addition and composition as multiplication. It is immediate
to check that EndR (M) is indeed a (unital) ring. Note that f ∈ EndR (M) is invertible
if and only if f is a bijective map, i.e., f is an automorphism.
After being familiar with factor groups and factor rings, the following notion
should be easy to grasp. Take a submodule L of M. As L is, in particular, a subgroup
of the additive group of M, one can form the factor group M/L. Recall that its
elements are cosets m + L = {m +  |  ∈ L}, m ∈ M, which are added according to
58 3 Modules and Vector Spaces

(m + L) + (m + L) = (m + m ) + L.

We claim that M/L becomes an R-module if we set

r(m + L) := rm + L, r ∈ R, m ∈ M.

Indeed, this operation is easily seen to be well-defined, and the module axioms are
clearly fulfilled.
Definition 3.20 The R-module M/L is called the factor module.
The map m → m + L is an R-module homomorphism from M onto M/L, called the
canonical homomorphism. Note that its kernel is equal to L. Another fundamental
fact connecting homomorphisms and submodules is the following: If ϕ : M → N is
an R-module homomorphism, then

M/ ker ϕ ∼
= im ϕ.

Indeed, the map m + ker ϕ → ϕ(m) is well-defined and is an isomorphism.

Our last theme is about sums and products of (sub)modules.

Definition 3.21 If {Mi | i ∈ I} is a family of submodules of M, then the submodule

generated
 by i∈I Mi is called the sum of the submodules {Mi | i ∈ I}. We write it
as i∈I Mi . If every Mjhas trivial intersection with the sum of the other submodules,
 
i.e., Mj i∈I\{j} Mi = 0, then we say that this sum is a direct sum, and write it
as ⊕i∈I Mi .
 
Note that i∈I M i consists of all (finite) sums of elements from i∈I Mi . Hence
we see that the sum i∈I Mi is direct if and only if for all distinct i1 , . . . , in ∈ I and
all mi1 ∈ Mi1 , . . . , min ∈ Min , mi1 + · · · + min = 0 implies mi1 = · · · = min = 0. If
the index set I is finite, say I = {1, . . . , n}, then we can write the sum i∈I Mi as
M1 + · · · + Mn , or M1 ⊕ · · · ⊕ Mn if it is a direct sum.
A formally more adequate term for the above defined direct sum is the internal
direct sum. One can, on the other hand, start with an arbitrary family of R-modules
{Mi | i ∈ I}, not necessarily submodules of the given module, and then form their
direct product Πi∈I Mi : this is the Cartesian product of the Mi ’s endowed with the
componentwise operations

(mi ) + (mi ) = (mi + mi ) and r(mi ) = (rmi )

for all r ∈ R, mi , mi ∈ Mi , i ∈ I. Its submodule consisting of all elements for

which all but finitely many of the components are 0 is called the external direct
sum of the Mi ’s. The external direct sum obviously contains isomorphic copies of
the modules Mi as its submodules, and, moreover, it is equal to their internal direct
sum. Therefore we often neglect the difference between internal and external direct
sums, and use the same notation for both. Thus, M1 ⊕ · · · ⊕ Mn may mean an internal
3.2 Basic Module-Theoretic Notions 59

or an external direct sum. Since the index set is finite here, this sum coincides with
n M (written also as M × · · · × M ).
the direct product Πi=1 i 1 n
Given a submodule L of M, does there exist another submodule K such that
M = L ⊕ K? It is easy to see that the answer is negative in general. Say, the proper
nonzero submodules of the Z-module Z are nZ, n ≥ 2, and any two of them intersect
nontrivially. Submodules for which the answer to our question is positive therefore
deserve attention.

Definition 3.22 A submodule L of M is called a direct summand of M if there

exists a submodule K of M such that M = L ⊕ K.

Example 3.23 Let L be a left ideal of a unital ring R. Consider L as a submodule of

the R-module R. We claim that L is a direct summand if and only if there exists an
idempotent e ∈ R such that L = Re. The “if” part is obvious: if L = Re then one
easily checks that K = R(1 − e) satisfies R = L ⊕ K. To prove the converse, suppose
there exists a left ideal K such that R = L ⊕ K. Then there exist e ∈ L and f ∈ K
such that 1 = e + f . For every  ∈ L we thus have  − e = f ∈ L ∩ K = 0, so that
 = e. In particular, e = e2 and L ⊆ Re. Since e ∈ L it follows that L = Re.

3.3 Vector Spaces Over Division Rings

The following familiar notion from linear algebra can be also considered in modules.

Definition 3.24 A subset B of an R-module M is said to be linearly independent

if for all distinct elements b1 , . . . , bn ∈ B and all elements r1 , . . . , rn ∈ R, r1 b1 +
· · · + rn bn = 0 implies ri = 0 for every i. If B is not linearly independent, then we
say that it is linearly dependent.

One should not be misled by the experience with this concept in vector spaces. For
instance, if M is not faithful, then even singletons {x} are never linearly independent.

Definition 3.25 A linearly independent subset B of an R-module M is called a basis

of M if M is generated by B.

Assume that R is unital. A unital R-module is said to be a free R-module if it

has a basis. (A remark for the readers familiar with categories: free R-modules are
free objects in the category of unital R-modules.) Thus, a unital R-module M is free
if it contains a linearly independent subset B with the following property: for every
m ∈ M there exist b1 , . . . , bn ∈ B and r1 , . . . , rn ∈ R such that m = r1 b1 +· · ·+rn bn .
Being free is a rather exceptional property for an R-module, but examples can be
easily found. The simplest one is the R-module R. Indeed, {1} is its basis. It is easy
to see that every free R-module is isomorphic to a direct sum of copies of the R-
module R. The structure of free modules is thus quite simple, apparently just as that
of vector spaces over fields. Yet a closer look shows that not everything we know
about vector spaces holds for free modules. Say, for some rings R it may happen
60 3 Modules and Vector Spaces

that different bases of a free R-module have different cardinalities. This may arouse
the reader’s curiosity, but we abandon the discussion on general free modules at
this point. A more complete information about them can be found in many other
algebra textbooks (for example, [Hun74] gives a gentle introduction). From now on
we will consider only free modules that are relevant for our goals. Before introducing
them in the next definition, let us point out that up until now it went without saying
that “vector space” means a vector space over a field. The following more general
conception turns out to be very useful.

Definition 3.26 Let D be a division ring. A unital (left) D-module is called a (left)
vector space over D.

A right vector space is of course defined as a unital right D-module. Vector spaces
over fields are traditionally left vector spaces, but this is just a matter of notational
convention (cf. Remark 3.2). One must be more careful about “left” and “right” when
considering vector spaces over noncommutative division rings. However, from now
on we assume that our vector spaces are left vector spaces.
We believe that the notions such as a subspace, linear combination, linear span,
etc., need no explanation. They are defined just as for vector spaces over fields.
Fortunately, studying vector spaces over division rings at a basic level is no harder
than studying vector spaces over fields. Every vector space is a free module over the
corresponding division ring, i.e., the following theorem holds.

Theorem 3.27 Every vector space has a basis.

Proof The proof is a typical application of Zorn’s lemma. Let S be the set of all
linearly independent subsets of V . Since ∅ is vacuously a linearly independent set,
S = ∅. Partially order S by inclusion, i.e., define  that A ≤ B if A ⊆ B. If {Ai | i ∈ I}
is a chain in S , then it is easy to see that U := i∈I Ai is a linearly independent set.
Thus U ∈ S and obviously it is an upper bound for the chain {Ai | i ∈ I}. By Zorn’s
lemma S contains a maximal element B.
Let us show that B is a basis of V . Since B ∈ S , it is enough to show that B
generates V . If V = {0}, then B = ∅ and this is trivially true. Thus, let V = {0}.
Pick a nonzero v ∈ V . Our goal is to prove that v is a linear combination of elements
from B. We may assume that v ∈ / B. Then B ∪ {v} contains B as its proper subset, and
therefore it is a linearly dependent set. This means that there exists a finite subset of
B ∪ {v} which is linearly dependent. Since subsets of B are linearly independent, this
subset must contain v. Thus, there exist b1 , . . . , bn ∈ B and d0 , d1 , . . . , dn ∈ D, not
all zero, such that
d0 v + d1 b1 + · · · + dn bn = 0.

The linear independence of b1 , . . . , bn implies that d0 = 0. Therefore

v = (−d0−1 d1 )b1 + · · · + (−d0−1 dn )bn

is indeed a linear combination of elements from B. 

3.3 Vector Spaces Over Division Rings 61

If we introduce S as the set of all linearly independent subsets of V that contain

a given linearly independent set, the same proof gives the following.
Theorem 3.28 Every linearly independent subset of a vector space is contained in
a basis.
A vector space V is said to be finite dimensional if it has a finite basis. Suppose
V contains a basis with n elements. Then:
(a) Every basis of V has n elements.
(b) If U is a subspace of V , then its bases have ≤ n elements.
(c) If a subspace U of V has a basis with n elements, then U = V .
For vector spaces over a field these are the most basic facts that one learns in a standard
linear algebra course. The reader must have seen the proofs. Glancing through them
one notices that they still work if a field is replaced by a noncommutative division
ring. Therefore we feel that we can omit giving these proofs without harming the
reader’s understanding of the subject (the unconvinced reader can take a look at, for
example, [Hun74, pp. 185–187]). In view of (a), we can define the dimension of V ,
dimD V , to be n if V contains a basis with n elements.

3.4 Endomorphisms and Matrices

Let V be a vector space over a division ring D. The endomorphism ring EndD (V ) will
often play an important role in our discussions. Its elements, the endomorphisms, are
also called linear maps or linear operators. Each linear map is uniquely determined
by the action on a basis. More precisely, the following holds.
Remark 3.29 Let B = {bi |i ∈ I} be a basis of V . For each family of vectors {wi | i ∈ I}
in V there exists a unique f ∈ EndD (V ) such that f (bi ) = wi for every i ∈ I. Indeed,
the uniqueness is clear since B is a basis, and the existence is also clear as we may
extend f from B to V by linearity, i.e., we define

f (di1 bi1 + · · · + din bin ) = di1 wi1 + · · · + din win

for every finite subset {i1 , . . . , in } of I and arbitrary di1 , . . . , din ∈ D.

Assume now that V is n-dimensional. If D = F is a field, then we know that the ring
EndF (V ) is isomorphic to the matrix ring Mn (F). Representing every endomorphism
by a matrix with respect to a fixed basis gives rise to a standard isomorphism. Let us
see if this still works if D is a noncommutative division ring.
Choose and fix a basis {b1 , b2 , . . . , bn } of V . Given f ∈ EndD (V ), we have


n
f (bj ) = fij bi , j = 1, . . . , n,
i=1
62 3 Modules and Vector Spaces

for some fij ∈ D. We attach the n × n matrix ϕ( f ) to f so that fij is its (i, j) entry, i.e.,
⎡ ⎤
f11 f12 ... f1n
⎢f ... f2n ⎥
⎢ 21 f22 ⎥
ϕ( f ) := ⎢
⎢ .. .. ..
⎥
.. ⎥ .
⎣ . . . . ⎦
fn1 fn2 . . . fnn

Let g be another element from EndD (V ), and let gij ∈ D be such that


n
g(bj ) = gij bi , j = 1, . . . , n.
i=1

Then


n
( f + g)(bj ) = ( fij + gij )bi ,
i=1

and hence ϕ( f + g) = ϕ( f ) + ϕ(g). If ϕ( f ) = 0, then f vanishes on a basis, and

so f = 0. This proves the injectivity of ϕ. The surjectivity is also obvious in view of
Remark 3.29. Indeed, given arbitrary fij ∈ D, we can define f ∈ EndD (V ) according
to f (bj ) = ni=1 fij bi , so that ϕ( f ) = ( fij ). Thus, ϕ is an isomorphism of additive
groups. This is hardly surprising. Let us now examine the action of ϕ on products.
To determine ϕ( fg), we compute


n 
n 
n  n 
 n 
( fg)(bj ) = f (g(bj )) = gkj f (bk ) = gkj fik bi = gkj fik bi .
k=1 k=1 i=1 i=1 k=1

Thus, the (i, j) entry of ϕ( fg) is


n
gkj fik . (3.2)
k=1

We can now compare this with the (i, j) entry of ϕ( f )ϕ(g) which is equal to


n
fik gkj . (3.3)
k=1

As D may be noncommutative, there is, unfortunately, no reason to believe that (3.2)

and (3.3) coincide. But let us look on the bright side: the summations (3.2) and (3.3)
are very similar, the terms differ only in the order of factors. This can be remedied.
We need the following definition.
3.4 Endomorphisms and Matrices 63

Definition 3.30 Let R be a ring. The opposite ring of R, denoted by Ro , is the ring
consisting of the same elements as R, having the same addition as R, and multipli-
cation · given by
x · y := yx,

where yx is the product in R.

Note that Ro is indeed a ring; in particular, the product · is associative. This ring
often naturally appears in the study of modules. For example, an obvious adaptation
of Remark 3.2 shows that there is a canonical correspondence between left R-modules
and right Ro -modules.
The rings R and Ro are not always isomorphic, but they are antiisomorphic. We
say that a bijective additive map θ from a ring R onto a ring S is an antiisomorphism
if θ (xy) = θ (y)θ (x) for all x, y ∈ R. Obviously, the map x → x is an antiisomorphism
from R onto Ro . Antiisomorphic rings are simultaneously division rings. Thus D is
a division ring if and only if Do is.
Let us now consider ϕ( f ) as an element of Mn (Do ) (rather than Mn (D)). This
point of view does not affect the above discussion up until (3.3). In Mn (Do ), the (i, j)
entry of ϕ( f )ϕ(g) is equal to


n 
n
fik · gkj = gkj fik
k=1 j=1

which does coincide with (3.2), the (i, j) entry of ϕ( fg). Thus ϕ is an isomorphism
from EndD (V ) onto Mn (Do ). We have proved

Theorem 3.31 Let V be an n-dimensional vector space over a division ring D. Then
the rings EndD (V ) and Mn (Do ) are isomorphic.

Remark 3.32 It is clear from the proof that Theorem 3.31 also holds for free modules.
That is, EndR (M) ∼
= Mn (Ro ) if M is a free R-module having a basis with n elements.

Remark 3.33 From Theorem 3.31 and Example 1.10 we infer that the ring EndD (V )
is simple if V is finite dimensional. (The converse is actually also true, the interested
reader can look at Example 5.10).

Remark 3.34 Theorem 3.31 yields an alternative aspect of Wedderburn’s structure

theory. We now see that we can add another equivalent condition to (i)–(iii) of
Theorem 2.61: (iv) A ∼= EndΔ (V ) where Δ is a division algebra (in fact, Δ ∼ = Do )
and V is an n-dimensional vector space over Δ. Let us remark that V is automatically
also a vector space over the base field F, since F ( = F · 1) is contained in Δ. The
elements in EndΔ (V ) are, in particular, F-linear maps.
64 3 Modules and Vector Spaces

3.5 Simple Modules

In Sect. 3.3 we have seen that the structure of an R-module is quite simple if R = D is
a division ring. We shall now consider another class of easily approachable modules.
However, they are of a quite different nature.
Definition 3.35 Let R be a ring. An R-module M is said to be simple if RM = 0
and its only submodules are 0 and M.
If M is unital, then the condition RM = 0 reduces to M = 0.
Simple modules are of enormous importance in ring theory. From now on they
will frequently play an important role in our discussions. In Chap. 5 we will see that
the structure theory of rings rests upon them.
Another term for a simple module is an irreducible module. An irreducible
representation of an algebra A is defined as a representation ϕ : A → EndF (V ) for
which there does not exist a proper nonzero subspace W of V that is invariant under
every ϕ(a), a ∈ A. It is easy to check that this is equivalent to the condition that the
associated A-module is irreducible.
The following lemma shows that simple modules are cyclic modules generated
by every nonzero element.
Lemma 3.36 An R-module M = 0 is simple if and only if M = Rm for every
0 = m ∈ M.
Proof Suppose M is simple. Let m ∈ M. As Rm is a submodule of M, we have either
Rm = 0 or Rm = M. Since N := {m ∈ M | Rm = 0} is also a submodule of M and
since RM = 0, we have N = 0. Therefore Rm = M whenever m = 0.
Conversely, if M = Rm for every 0 = m ∈ M and L is a nonzero submodule of
M, then M = R ⊆ L for every 0 =  ∈ L. Thus L = M. 
Viewing modules as generalizations of vector spaces, one might first wonder what
does the simplicity mean for the latter. The answer is easy:
Example 3.37 A vector space V over a division ring D is simple (as a D-module) if
and only if dimD V = 1.
This example, however, is not really illuminating. The next ones better reflect the
meaning of simple modules in the theory that we shall later develop.
Example 3.38 Let V be a nonzero vector space over a division ring D. As in Example
3.10 we may regard V as a module over R = EndD (V ) (via fv := f (v)). We claim
that V is a simple R-module. In view of Lemma 3.36 it suffices to show that for each
pair v, w ∈ V with v = 0 there exists f ∈ R such that f (v) = w. Now, by Theorem
3.28 we may choose a basis of V containing v, and then use Remark 3.29 to conclude
that such an f indeed exists.
Example 3.39 Let L be a left ideal of a ring R such that RL = 0. Then L is simple
as an R-module if and only if L is a minimal left ideal.
3.5 Simple Modules 65

The following lemma indicates the significance of this example.

Lemma 3.40 Let R be a simple ring having a minimal left ideal L. Then every simple
R-module is isomorphic to L.
Proof Let M be a simple R-module. By simplicity of R we have LR = R. Hence
LM ⊇ L(RM) = (LR)M = RM = 0. Choose m ∈ M such that Lm = 0. Since
Lm is a submodule of M it follows that Lm = M. The map ϕ : L → M given by
ϕ() = m is therefore a surjective R-module homomorphism. Its kernel is a left ideal
of R properly contained in L, so it can only be 0. Therefore ϕ is an isomorphism. 
Example 3.41 Let R = Mn (D), D a division ring. Then R is a simple ring (Example
1.10). A typical example of a minimal left ideal is RE11 , the set of all matrices
that have zeros in all columns except in the first one (Example 2.57). Therefore
every simple R-module is isomorphic to RE11 . Neglecting the zero columns, we may
identify this module with Dn , the abelian group of all n-tuple columns on which R
acts via the usual matrix multiplication. (A side remark, related to Theorem 3.31:
We can consider Dn as a left vector space over D, but then the matrices from R do
not act on Dn as D-linear operators; if, however, we consider Dn as a right vector
space over D, then they do.)

3.6 Maximal Left Ideals

Minimal left ideals are prototypes of simple modules. However, they only rarely exist
in rings. Say, even such a simple and basic example of a ring as Z does not have them
(why?). Let us now consider left ideals having the opposite property. As it turns out,
they also have a tight connection with simple modules.
Definition 3.42 A left ideal U of a ring R is called a maximal left ideal if U = R
and U is not properly contained in a proper left ideal.
Maximal right ideals and maximal ideals are defined analogously.
Example 3.43 Let R = Mn (D), D a division ring, and let U = RE11 ⊕ · · · ⊕
REn−1,n−1 . That is, U is the set of all matrices that have zeros in the last column.
It is easy to check that U is a maximal left ideal of R. We remark that R = L ⊕ U
where L is the minimal left ideal REnn .
Example 3.44 If R is a commutative unital ring, then an ideal U of R is maximal if
and only if the factor ring R/U is a field. This is well-known, and the proof is easy
(cf. Lemma 5.38 below, which considers a more general situation). Hence:
(a) pZ is a maximal ideal of Z for every prime number p.
(b) Uc = { f ∈ C[a, b] | f (c) = 0}, where c is a fixed number from [a, b], is a
maximal ideal of the ring of continuous functions C[a, b]. Indeed, f → f (c) is
a homomorphism from C[a, b] onto R with kernel Uc , so that C[a, b]/Uc ∼ =R
is a field.
66 3 Modules and Vector Spaces

It turns out that the Uc ’s are actually the only maximal ideals of C[a, b]. Of course,
the pZ’s are the only maximal ideals of Z.

Most rings have plenty of maximal left ideals.

Lemma 3.45 If R is a unital ring, then every proper left ideal L of R is contained in
a maximal left ideal.

Proof Let S be the set of all proper left ideals of R that contain L. Since L ∈ S ,
S is nonempty. Partially order S by inclusion. If {Ki | i ∈ I} is a chain in S , then
K = i∈I Ki is its upper bound. Indeed, K is easily seen to be a left ideal, and K = R
since 1 ∈
/ K. By Zorn’s lemma S contains a maximal element, which is obviously
a maximal left ideal of R containing L. 

One can imagine that a maximal left ideal U of a ring R is “too big” to be simple
as an R-module. However, we will be interested in the factor module R/U which
should be expected to be “small”. Here we regard R as an R-module and U as its
submodule, so the operations in R/U are given by

(r + U) + (s + U) = (r + s) + U and r(s + U) = rs + U.

Lemma 3.46 Let R be a ring. Every simple R-module is isomorphic to R/U for some
maximal left ideal U of R. Conversely, if U is a maximal left ideal of R such that
R2 ⊆ U (this is automatically fulfilled if R is unital), then R/U is a simple R-module.

Proof Let M be a simple R-module. Choose 0 = m ∈ M. Then Rm = M and hence

ϕ : R → M, ϕ(r) = rm, is a surjective R-module homomorphism. Its kernel U is a
left ideal of R, and R/U ∼= M. We have to show that U is maximal. Let L be a left
ideal of R such that U ⊆ L ⊆ R. Then L/U is a submodule of R/U. Since R/U,
being isomorphic to M, is a simple module, it follows that L/U = 0 or L/U = R/U.
Accordingly, L = U or L = R, showing that U is maximal.
To establish the converse, take a maximal left ideal U of R, and consider a sub-
module N of R/U. Clearly, L := { ∈ R |  + U ∈ N} is a left ideal containing U.
Thus L = U or L = R, yielding N = 0 or N = R/U. Therefore R/U is simple
provided that R · R/U = 0, i.e., R2 ⊆ U. 

Example 3.47 Let R, U, and L be as in Example 3.43. Note that R/U ∼

= L.

Example 3.48 By Example 3.44 it is now clear what are the simple Z-modules and
C[a, b]-modules.

Lemmas 3.45 and 3.46 yield

Corollary 3.49 Every nonzero unital ring has a simple module.

3.7 Schur’s Lemma 67

3.7 Schur’s Lemma

In the previous chapter we saw that minimal left ideals give rise to division rings
(Lemma 2.58), and this was our basis for developing the structure theory of finite
dimensional algebras. Simple modules, which can be regarded as generalizations of
minimal left ideals (Example 3.39), have a similar role in the structure theory of
more general rings and algebras. As a first step towards this more general theory we
will now show that simple modules give rise to division rings. The next lemma may
be just a straightforward observation, but an extremely useful one. It is named after
I. Schur who used it in the representation theory of groups.
Lemma 3.50 (Schur’s Lemma) If M is a simple R-module, then the endomorphism
ring EndR (M) is a division ring.
Proof Let f ∈ EndR (M), f = 0. Then ker f is a submodule of M different from
M, and im f is a submodule of M different from 0. Since M is simple, the only
possibility is that ker f = 0 and im f = M. Thus f is an automorphism, and therefore
an invertible element in EndR (M). 
A slightly more general version of Schur’s Lemma states that a nonzero homo-
morphism from an arbitrary module into a simple module is surjective, and a nonzero
homomorphism from a simple module into an arbitrary module is injective.
A minimal left ideal L of a semiprime ring R thus gives rise to two division rings:
the first one is described in Lemma 2.58 and the other one arises from Schur’s Lemma
if we regard L as a simple R-module. What is the connection between them? The
answer will be given in the corollary to the next lemma.
Lemma 3.51 Let e be an idempotent in a ring R. Then EndR (Re) ∼ = (eRe)o .
Proof Define ϕ : EndR (Re) → eRe by

ϕ( f ) := f (e).

Since f is an endomorphism, we have f (e) = f (ee) = ef (e), and since f maps into
Re we have f (e) = f (e)e. Accordingly, f (e) = ef (e)e indeed belongs to eRe. The
additivity of ϕ is clear. So is the injectivity: f (e) = 0 implies f (re) = rf (e) = 0
for every r ∈ R. Pick a ∈ eRe and define f : Re → Re by f (u) := ua. Then
f ∈ EndR (Re) and ϕ( f ) = f (e) = a, which proves that ϕ is surjective. Next,

ϕ( fg) = f (g(e)) = f (g(e)e) = g(e)f (e) = ϕ(g)ϕ( f ).

Therefore ϕ is an antiisomorphism. Composing ϕ with the canonical antiisomor-

phism x → x from eRe onto (eRe)o we thus obtain an isomorphism from EndR (Re)
onto (eRe)o . 
Corollary 3.52 Let L be a minimal left ideal of a semiprime ring R. Then there
exists an idempotent e ∈ R such that L = Re, D = eRe is a division ring, and
EndR (L) ∼= Do .
Proof Use Lemmas 2.58 and 3.51. 
68 3 Modules and Vector Spaces

3.8 Semisimple Modules

Theorem 3.27 saying that every vector space has a basis can be equivalently stated
as that every vector space is the direct sum of a family of 1-dimensional subspaces.
In view of Example 3.37 this can be further rephrased as follows: If D is a division
ring, then every unital D-module is the direct sum of a (possibly infinite) family of
simple submodules. We will now consider modules with this property over arbitrary
rings.

Definition 3.53 Let R be a ring. An R-module M is said to be semisimple if M is

the direct sum of a family of simple submodules.

The sum of an empty family of submodules should be understood as 0. The zero

module is therefore semisimple. As indicated above, all unital modules over division
rings are semisimple. Trivially, simple modules are semisimple. A more interesting
example, related to Examples 2.57, 3.39 and 3.43, is

Example 3.54 Let R = Mn (D), D a division ring. Pick 1 ≤ k ≤ n, and set M =

RE11 ⊕· · ·⊕REkk . Thus, M is a left ideal of R consisting of all matrices that have zeros
in the last n − k columns. Each REii is a minimal left ideal of R, and hence a simple
module. Therefore M is a semisimple module. In particular, R = RE11 ⊕ · · · ⊕ REnn
is semisimple as an R-module.

Recall a familiar fact from linear algebra: Every generating set of a vector space
V contains a basis of V . In other words, if V is equal to the sum of 1-dimensional
subspaces, then V is equal to the direct sum of some of them. The next lemma extends
this fact to modules over arbitrary rings.

Lemma 3.55 If a module M is the sum of a family of simple submodules {Mi | i ∈ I},
then there exists K ⊆ I such that M is the direct sum of the family {Mk | k ∈ K}.

Proof Let S be the set of all J ⊆ I such that the sum of the family {Mj | j ∈ J}
is direct. Clearly S is a nonempty set since {i} ∈ S for every i ∈ I, and can be
partially ordered by inclusion. The existence of a maximal element in S is a typical
application of Zorn’s lemma. Indeed, just note that the upper bound of every chain
is its union. Denote a maximal element by K, and set M0 := ⊕k∈K Mk . Suppose
there exists i ∈ I such that Mi ⊆ M0 . Then Mi ∩ M0 is a proper submodule of Mi .
As Mi is simple, we must have Mi ∩ M0 = 0. Accordingly, the sum of the family
{Mj | j ∈ K ∪ {i}} is direct, which contradicts the maximality of K. Hence Mi ⊆ M0
for each i ∈ I, so that M0 = M. 

This lemma implies that one can equivalently define semisimple modules by
omitting the word “direct” in Definition 3.53:

Corollary 3.56 If a module is the sum of a family of simple submodules, then it is

semisimple.
3.8 Semisimple Modules 69

Every subspace of a vector space is a direct summand. The reader probably knows
this fact at least for finite dimensional vector spaces over fields. For general vector
spaces it can be derived from Theorem 3.28. Let us show that it actually holds for
semisimple modules.
Lemma 3.57 Every submodule of a semisimple module is a direct summand.
Proof Let M be a semisimple module. Thus M = ⊕i∈I Mi where Mi are simple
submodules. Let L be a submodule of M. By simplicity of Mi we either have L ∩Mi =
Mi or L ∩ Mi = 0, i.e., L either contains Mi or has trivial intersection with it. If the
former holds for every i ∈ I, then L = M and the desired conclusion trivially holds.
We may therefore assume that L ∩ Mi = 0 for some i ∈ I.
As one should expect, we haveto  invokeZorn’s lemma at one point. Consider
the set S of all J ⊆ I such that L ⊕j∈J Mj = 0; S is nonempty by the previous
paragraph. A standard argument shows that S has a maximal  element, let us call it
K, with respect to inclusion. We claim that L⊕ ⊕k∈K M  k = M. If this was not true,
there would exist i ∈ I such that M  ⊆ L ⊕ ⊕ M , and hence, by simplicity of
 
i k∈K k
  
Mi , Mi L ⊕ ⊕k∈K Mk = 0. But then L ⊕j∈K∪{i} Mj = 0, contradicting the
maximality of K. 
It turns out that for unital modules the converse is also true: If every submodule
of a unital module M is a direct summand, then M is semisimple. We will not need
this result, so we omit the proof. One can find it, for example, in [Row08, p. 36].
Do not be misled with subspaces of vector spaces—submodules are only seldom
direct summands. For instance, proper nonzero submodules of the Z-module Z are
nZ, n ≥ 2, and none of them is a direct summand. Thus Z is not a semisimple
Z-module. We will soon learn that the semisimplicity of the R-module R is in fact
an extremely restrictive condition.

3.9 Wedderburn’s Structure Theory Revisited

So far we have introduced several concepts related to modules and considered their
formal properties. It is about time to experience module theory in action. We are going
to do this now by giving a useful supplement to Wedderburn’s structure theory.
We will thus consider finite dimensional algebras and modules over them. In the
proofs we will refer to previous sections which, however, deal only with modules over
rings. This should not confuse the reader. Everything from these sections makes sense
for modules over algebras. Of course, all definitions must be appropriately interpreted
in the context of algebras. For example, a simple module over an algebra should have
no proper nonzero algebra submodules, algebra module homomorphisms must be
linear maps, etc. These changes do not really affect the proofs.
We can add the uniqueness statement to Wedderburn’s Theorem 2.61.
Theorem 3.58 If A is a finite dimensional prime (or equivalently, simple) algebra,
then there exist a unique n ∈ N and an (up to isomorphism) unique division algebra
D such that A ∼= Mn (D).
70 3 Modules and Vector Spaces

Proof We only have to prove the uniqueness part. Assume therefore that A ∼= Mn (D)
and A ∼= Mm (C) for some n, m ∈ N and division algebras D and C. It suffices to
prove that D ∼= C since this in particular implies that D and C are of the same
dimension, from which n = m follows.
Let ϕ : Mn (D) → A be an isomorphism.
 Set e := ϕ(E11 ). Since Mn (D)E11 is a
minimal left ideal of Mn (D), Ae = ϕ Mn (D)E11 is a minimal left ideal of A.By
Lemma 3.51, EndA (Ae) ∼ = (eAe)o . On the other hand, eAe = ϕ E11 Mn (D)E11 ∼ =
∼
E11 Mn (D)E11 = D. Accordingly,

EndA (Ae) ∼
= Do .

Analogously, there exists an idempotent f ∈ A such that Af is a minimal left ideal of

A and

EndA (Af ) ∼
= Co.

As both Ae and Af are simple A-modules, Lemma 3.40 tells us that they are isomor-
phic. This implies that the algebras EndA (Ae) and EndA (Af ) are isomorphic. Indeed,
if θ : Ae → Af is a module isomorphism, then g → θ gθ −1 defines an algebra
isomorphism from EndA (Ae) onto EndA (Af ). Consequently, Do ∼ = C o , and hence
∼
D = C. 

Wedderburn’s Theorem 2.64 can be for unital algebras extended as follows.

Theorem 3.59 Let A be a nonzero finite dimensional unital algebra. The following
statements are equivalent:
(i) A is semiprime.
(ii) A∼= Mn1 (D1 ) × · · · × Mnr (Dr ) for some ni ∈ N and division algebras Di .
(iii) Every unital A-module is semisimple.
(iv) A is a semisimple A-module.
(v) Every left ideal of A is of the form Ae where e ∈ A is an idempotent.

Proof (i) =⇒ (ii). Theorem 2.64.

(ii) =⇒ (iii). We claim that (ii) implies that there exist pairwise orthogonal
idempotents e1 , . . . , es , where s = n1 + · · · + nr , such that e1 + · · · + es = 1,
A = Ae1 ⊕ · · · ⊕ Aes , and each Li := Aei is a minimal left ideal of A. Observe,
first of all, that it suffices to show this for A = Mn (D). In this case we have
A = AE11 ⊕ · · · ⊕ AEnn and each AEii is a minimal left ideal of A (cf. Example
3.54), which proves our claim.
Let M be a unital A-module. For every m ∈ M we have

m = 1m = e1 m + · · · + es m ∈ L1 m + · · · + Ls m.

Thus, M is the sum of the family of its submodules {Li m | m ∈ M, 1 ≤ i ≤ s}. Pick
m and i such that Li m = 0. Let us show that Li m is a simple submodule of M. We
3.9 Wedderburn’s Structure Theory Revisited 71

will actually show that Li m is isomorphic to Li which is, as a minimal left ideal of
A, a simple A-module. Clearly,  → m defines an A-module homomorphism from
Li onto Li m. Its kernel is a proper submodule of a simple module Li , so it must be
0. Therefore Li ∼= Li m indeed holds. We now see that M is the sum of the family of
simple submodules {Li m | m ∈ M, 1 ≤ i ≤ s, Li m = 0}, and thus M is semisimple
by Corollary 3.56.
(iii) =⇒ (iv). Trivial.
(iv) =⇒ (v). Let L be a left ideal of A. Consider L as a submodule of a semisimple
module A. From Lemma 3.57 it follows that L is a direct summand. As shown in
Example 3.23, this is equivalent to the existence of an idempotent e ∈ A such that
L = Ae.
(v) =⇒ (i). A left ideal containing a nonzero idempotent cannot be nilpotent. 

Unlike (iii)–(v), (i) and (ii) are left-right symmetric conditions. Therefore we can,
in particular, add the following equivalent statement to the above list: (v’) Every right
ideal of A is of the form eA where e is an idempotent.

Corollary 3.60 Let A be a finite dimensional simple algebra. If finite dimensional

unital A-modules have the same dimension, then they are isomorphic.

Proof Let M be a unital A-module. By Theorem 3.59 (iii), M is semisimple. If M

is also finite dimensional, then it is obviously a finite direct sum of simple modules:
M = M1 ⊕· · ·⊕Mr , Mi simple. By Lemma 3.40, each Mi is isomorphic to a minimal
left ideal L of A. Therefore M is isomorphic to the direct sum of r copies of L. The
number r thus determines M up to isomorphism. On the other hand, r is determined
by the dimension of M. Thus, if N is a unital A-module of the same dimension as M,
then M ∼ = N. 

A brief inspection of the proof of Theorem 3.59 shows that we have used the
assumption that A is finite dimensional only in the proof of (i) =⇒ (ii). All other
implications hold for every unital algebra, and even for every unital ring (provided
that in (ii) we assume that Di are division rings, not algebras). Certainly (i) =⇒
(ii) does not hold for general unital rings, there are a myriad of counterexamples (Z
to start with). However, it turns out that (v) =⇒ (ii) does hold. Thus, (ii)–(v) are
equivalent if A = R is an arbitrary unital ring. A ring satisfying these equivalent
conditions is called a semisimple ring. Clearly, a ring is isomorphic to Mn (D) if and
only if it is a simple semisimple ring (this may sound strange, but not every simple
ring is semisimple; the terminology becomes a bit awkward here). Moreover, the
uniqueness part of Corollary 3.58 also holds in this context.
The preceding paragraph outlines the so-called Wedderburn-Artin theory,
developed by E. Artin some 20 years after Wedderburn’s seminal work on finite
dimensional algebras. Treating semisimple rings is not substantially harder than
treating semisimple finite dimensional algebras. Nevertheless, we have decided to
stick with the classical finite dimensional context, which is probably intuitively more
clear and therefore easier to understand for a beginner. The interested reader will find
a detailed exposition of the Wedderburn-Artin theory in a number of textbooks, in
72 3 Modules and Vector Spaces

particular in [Bea99, FD93, Her68, Lam01, Pie82, Row91, Row08]. One of the ingre-
dients of this theory are the artinian rings, which will be briefly touched upon in the
next section.

3.10 Chain Conditions

The following three statements are equivalent for a vector space V : (i) V is infinite
dimensional, (ii) there exists an infinite descending chain V1  V2  . . . of sub-
spaces of V , (iii) there exists an infinite ascending chain W1  W2  . . . of subspaces
of V . Indeed, each of (ii) and (iii) obviously implies (i), and the converses follow by
taking a linearly independent subset {v1 , v2 , . . . } of V and defining Vn to be the linear
span of {vn , vn+1 , . . . } and Wn to be the linear span of {v1 , . . . , vn }. The following
two notions can be therefore viewed as generalizations of finite dimensionality.

Definition 3.61 A module M is said to be artinian (or is said to satisfy the descend-
ing chain condition on submodules) if for every chain M1 ⊇ M2 ⊇ . . . of submod-
ules of M, there exists m ∈ N such that Mm = Mm+1 = . . . .

Definition 3.62 A module N is said to be noetherian (or is said to satisfy the

ascending chain condition on submodules) if for every chain N1 ⊆ N2 ⊆ . . . of
submodules of N, there exists n ∈ N such that Nn = Nn+1 = . . . .

Let X be a set and S be a set of subsets of X. We say that A ∈ S is a minimal

element in S (with respect to set inclusion) if there does not exist Y ∈ S such that
A  Y . Similarly, we say that B is a maximal element in S if there does not exist
Z ∈ S such that B  Z.

Lemma 3.63 The following statements are equivalent for a module M:

(i) M is artinian.
(ii) Every nonempty set of submodules of M has a minimal element.

Proof (i) =⇒ (ii). Let S be a nonempty set of submodules of M. Pick M1 ∈ S .

If M1 is not minimal, then there exists M2 ∈ S such that M1  M2 . If M2 is not
minimal, then M2  M3 for some M3 ∈ S . Continuing this process we arrive in
a finite number of steps at a minimal element in S ; for if not, there would exist
an infinite chain M1  M2  . . . . (Note that the axiom of choice was used in this
argument.)
(ii) =⇒ (i). Let M1 ⊇ M2 ⊇ . . . be a descending chain of submodules of M. If
Mm is a minimal element of the set {M1 , M2 , . . . }, then Mm = Mm+1 = . . . . 

Lemma 3.64 The following statements are equivalent for a module N:

(i) N is noetherian.
(ii) Every nonempty set of submodules of N has a maximal element.
(iii) Every submodule of N is finitely generated.
3.10 Chain Conditions 73

Proof (i) ⇐⇒ (ii). This can be proved in the same manner as Lemma 3.63, just
the inclusions must be reversed.
(ii) =⇒ (iii). Let P be a submodule of N. Denote by S the set of all finitely
generated submodules of P. Since it contains 0, S is nonempty. Thus there exists a
maximal element B in S . We have to show that B = P. Let b1 , . . . , br be generators
of B. Take an arbitrary p ∈ P. The submodule generated by p, b1 , . . . , br lies in S
and contains B, and therefore it must be equal to B. In particular, p ∈ B. 
(iii) =⇒ (i). Let N1 ⊆ N2 ⊆ . . . be a chain of submodules of N. Note that ∞ i=1 Ni
is also a submodule of N, and is therefore generated by finitely many elements
c1 , . . . , cq . Each ci lies in one
of the Nj ’s, so there exists n ∈ N such that Nn contains
all c1 , . . . , cq . Hence Nn = ∞ i=1 Ni , i.e., Nn = Nn+1 = . . . . 

Definition 3.65 A ring R is said to be left (resp. right) artinian if R is an artinian

left (resp. right) R-module.

Definition 3.66 A ring R is said to be left (resp. right) noetherian if R is a noetherian

left (resp. right) R-module.

A left artinian (resp. noetherian) ring is not necessarily right artinian (resp.
noetherian), and vice versa. For simplicity we will consider only left artinian and
left noetherian rings in this section. Of course, all statements that we will record
have analogous right versions. If R is both left and right artinian (resp. noetherian),
then we say that R is artinian (resp. noetherian). A left artinian (resp. noetherian)
commutative ring is obviously automatically artinian (resp. noetherian).
Since submodules of the left R-module R are the left ideals of R, R is left artinian if
and only if for every chain L1 ⊇ L2 ⊇ . . . of left ideals of R there exists m ∈ N such
that Lm = Lm+1 = . . . . Analogously, R is left noetherian if it satisfies the ascending
chain condition on left ideals.
There is a non-obvious relation between these two classes of rings: A left artinian
unital ring is left noetherian. But we will not prove this. On the other hand, the ring
Z is noetherian but not artinian; this follows easily from the fact that the nonzero
ideals of Z are nZ, n ∈ N, and that nZ ⊆ mZ if and only if m divides n.
The artinian property turns out to be a perfect ring-theoretic substitute for finite
dimensionality. The basic structure theory of finite dimensional algebras can be,
sometimes with literally the same proofs and sometimes with an extra effort, gener-
alized to artinian rings. But we will not go into details on that. Let us mention only
the fundamental result, often referred in the literature as the Wedderburn-Artin the-
orem, which states that a ring R is simple and left artinian if and only if R ∼ = Mn (D)
where n ∈ N and D is a division ring. In Chap. 5 we will prove a version of the (more
substantial) “only if” part, saying that a simple unital ring with a minimal left ideal is
isomorphic to Mn (D) (Corollary 5.34). Left artinian rings of course do have minimal
left ideals by Lemma 3.63 (the set of all nonzero left ideals has a minimal element).
The noetherian property gives rise to a considerably richer class of rings. Lemma
3.64 implies that a ring R is left noetherian if and only if every left ideal of R is finitely
generated. In particular, if every left ideal of R is principal, i.e., it is generated by a
single element, then R is left noetherian. Besides Z, the basic example of such a ring
74 3 Modules and Vector Spaces

is F[ω] where F is a field. We assume that the reader has learned this fact in a basic
algebra course. A deeper result, known as Hilbert’s basis theorem, says that if T
is a left noetherian unital ring, then so is T [ω]. We will now prove a slightly more
general version which is not limited to polynomials. However, the reader will notice
that the polynomials are imitated in the proof.

Theorem 3.67 Let R be a unital ring, T be a subring of R that contains the unity of
R, and d ∈ R. If T is left noetherian, T + dT = T + Td, and R is generated by the
set T ∪ {d}, then R is left noetherian.

Proof From Td ⊆ T + dT it follows easily that Td s ⊆ T + dT + · · · + d s T for every

s ≥ 1. Similarly, dT ⊆ T + Td yields d s T ⊆ T + Td + · · · + Td s . Consequently,

T + dT + · · · + d s T = T + Td + · · · + Td s for every s ≥ 1. (3.4)

We denote the set in (3.4) by Rs ; we also set R0 := T . Clearly, TRs = Rs T = Rs for

every s ≥ 0.
Take a left ideal L of R. Our goal is to prove that L is finitely generated. Set
W0 := L ∩ T and

Wk := {w ∈ T | d k w ∈ Rk−1 + L}, k ≥ 1.

Using (3.4) one easily verifies that Wk is a left ideal of T . Since T is left noetherian, Wk
rk
is finitely generated: Wk = j=1 Twjk for some wjk ∈ Wk . Next, using dL ⊆ L and
dRk−1 ⊆ Rk we infer that Wk ⊆ Wk+1 for every k ≥ 0. Therefore Wn = Wn+1 = . . .
for some n ≥ 0.
By definition of Wk , k ≥ 1, for each wjk there is jk ∈ L such
that jk − d k wjk ∈
rk
Rk−1 ; for k = 0 we simply take j0 = wj0 ∈ L. Let M := nk=0 j=1 Rjk . We want
to show that L is equal to M and is therefore finitely generated. ObviouslyM ⊆ L.
Since R is generated by T ∪ {d} it can be deduced from (3.4) that R = ∞ s=0 Rs .
Therefore it suffices to show that L ∩ Rs ⊆ M for every s ≥ 0. We proceed by
induction on s. For s = 0 we have L ∩ R0 = W0 ⊆ M, so let s ≥ 1. Take  ∈ L ∩ Rs
and let ws ∈ T be such that  − d s ws ∈ Rs−1 .
rFirst assume s ≤ n. Since, by the very definition, ws ∈ Ws , we have ws =

j=1 j js for some tj ∈ T . By (3.4) there exist elements tj ∈ T such that
s
t w
d s tj − tj d s ∈ Rs−1 . A short calculation, based on js − d s wjs ∈ Rs−1 , shows
s 
that − rj=1 tj js ∈ Rs−1 . Since this is an element from L, we may use the induction
s 
assumption to conclude that it lies in M. As rj=1 tj js ∈ M by definition, it follows
that  ∈ M.
If s > n, then ws ∈ Ws = Wn . As in the previous paragraph we can find tj , tj ∈
rn  n
T such that ws = j=1 tj wjn and d tj − tj d ∈ Rn−1 . One can verify that  −
n
rn s−n 
j=1 d tj jn ∈ Rs−1 . Therefore this element lies in M by induction assumption,
yielding  ∈ M. 
3.10 Chain Conditions 75

Example 3.68 The Weyl algebra A1 satisfies the conditions of Theorem 3.67 with
L playing the role of T and D playing the role of d. Indeed, L is easily seen to be
isomorphic to F[ω] and is therefore (both left and right) noetherian. From (1.7) it
follows that L +DL ⊆ L +L D. Analogously we see that L +L D ⊆ L +DL .
Of course, A1 is generated by L ∪ {D}. Therefore A1 is noetherian.

Exercises

3.1. Let R be a unital ring and M be a unital R-module. Show that M is cyclic if
and only if M ∼
= R/L for some left ideal L of R.
3.2. Let L and N be submodules of a module M. Show that:
(a) N/(N ∩ L) ∼ = (N + L)/L.
(b) If L ⊆ N, then N/L is a submodule of M/L, and (M/L)/(N/L) ∼
= M/N.
Remark: These two results, together with the standard fact that M/ ker ϕ ∼ =N
if ϕ : M → N is a surjective homomorphism, are often referred to as the
isomorphism theorems. They are usually attributed to E. Noether. Analo-
gous theorems, with analogous proofs, hold for other algebraic structures, in
particular for groups and rings. The ring-theoretic versions of (a) and (b) read
as follows: If I and J are ideals of a ring R, then J/J ∩ I ∼= (J + I)/I, and
(R/I)/(J/I) ∼ = R/J if I ⊆ J.
3.3. Let L be a submodule of a module M. Show that every submodule of M/L is
of the form N/L, where N is a submodule of M which contains L. Similarly,
if J is an ideal of a ring R, then every ideal of R/I is of the form J/I, where J
is an ideal of R which contains I.
3.4. Let e and f be idempotents in a ring R. Show that the left ideals Re and Rf are
isomorphic as R-modules if and only if there exist u, v ∈ R such that e = uv
and f = vu. Moreover, if R = Mn (F) with F a field this is further equivalent
to the condition that e and f are similar matrices, i.e., there exists an invertible
t ∈ R such that f = tet −1 .
3.5. Let R be a unital ring, and let M be a free R-module with basis B. Note that
every function from B into a unital R-module N can be (uniquely) extended to
a homomorphism from M into N. Hence deduce that every unital R-module
is the homomorphic image of a free R-module.
3.6. Is a submodule of a free module free?
3.7. A module P is said to be projective if for every surjective homomorphism f :
M → N and every homomorphism g : P → N there exists a homomorphism
h : P → M such that fh = g. Show that every free module is projective.
3.8. Note that a ring R is isomorphic to Ro if and only if R has an antiautomorphism.
Show that if R is such a ring, then so is Mn (R). Find an example of a ring
without antiautomorphisms.
Hint: Where does an antiautomorphism send a left unity?
76 3 Modules and Vector Spaces

3.9. Let V be a vector space over a division ring D. Suppose V has an infinite
countable basis. Show that the ring EndD (V ) is isomorphic to the ring of all
column finite N × N matrices over Do , i.e., matrices with the property that
each of their columns has only finitely many nonzero entries.
Remark: The column finite matrices indeed form a ring under the standard
matrix operations. So do, for instance, the row finite matrices. On the other
hand, the set of all N × N matrices is not a ring since one faces infinite sum-
mations when performing multiplication. (This set is thus merely an additive
group.)
3.10. Describe the center of the ring EndD (V ).
3.11. Let V be a finite dimensional vector space over a division ring D. We define
rank( f ), the rank of f ∈ EndD (V ), as the dimension of im f . Show that
rank( f ) = dimD V − dimD ker f . Further, show that d( f , g) := rank( f − g)
defines a metric on EndD (V ).
3.12. Let D be a division ring and let A = (aij ) ∈ Mn (D). Which of the following
conditions imply that A is invertible?
(a) A is left invertible.
(b) 
A is not a left zero-divisor.
(c) σ ∈Sn sgn(σ )a1σ (1) . . . anσ (n)  = 0.
3.13. Show that every simple module over a unital ring is unital.
3.14. Let V be a nonzero vector space over a division ring D, and let R =
EndD (V ). As in Example 3.38, consider V as a (simple) R-module. Show
that EndR (V ) ∼
= D.
3.15. Let M be a simple module over a commutative unital ring R. Show that the
division ring EndR (M) from Schur’s lemma is a field isomorphic to the factor
ring R/U where U is a maximal ideal of R.
3.16. Show that a homomorphic image of a semisimple module is semisimple.
3.17. Show that finite dimensional simple algebras A and B are isomorphic if (and
only if) there exists n ∈ N such that the algebras Mn (A) and Mn (B) are
isomorphic.
3.18. The left annihilator of a subset I of a ring R is the left ideal lannR (I) := {x ∈
R | xI = 0}, and the right annihilator of I is the right ideal rannR (I) := {x ∈
R | Ix = 0}. Show that if R is semiprime and I is an ideal, then lannR (I) =
rannR (I) is an ideal, and if, additionally, R is a finite dimensional algebra,
then lannR (lannR (I)) = I. Note that this formula is not, in general, valid if I
is merely a left ideal. However, a similar formula that also involves the right
annihilator does hold in this case. Find it!
3.19. Let L be a submodule of a module M. Show that M is artinian (resp. noetherian)
if and only if both L and M/L are artinian (resp. noetherian).
3.20. Show that an injective endomorphism of an artinian module is automatically
surjective. Similarly, a surjective endomorphism of a noetherian module is
injective.
Exercises 77

3.21. Suppose a module M is both artinian and noetherian. Show that for every
endomorphism ϕ of M there exists n ∈ N such that M = im ϕ n ⊕ ker ϕ n .
Remark: This result is known as Fitting’s lemma.
3.22. Show that the polynomial ring F[Ω], where F is a field, is noetherian if
and only if the set Ω is finite. Is the ring of continuous functions C[a, b]
noetherian?
3.23. Let V be a vector space over a division ring D. Show that the following
statements are equivalent:
(i) The space V is finite dimensional.
(ii) The ring EndD (V ) is left artinian.
(iii) The ring EndD (V ) is left noetherian.
Hint: The ring Mn (Do ) has the natural structure of a vector space over Do of
dimension n2 , and its left ideals are automatically subspaces.
Chapter 4
Tensor Products

Sometimes we find some mathematical concept difficult to comprehend at the

beginning. Later, when we get used to it, it appears so natural to us that we do
not understand anymore what appeared so mysterious about it before. Perhaps the
tensor product is a concept with which many have such an experience.
Our intention is to give a gentle introduction to tensor products. After a preliminary,
unavoidable consideration of tensor products of vector spaces, we will proceed with
tensor products of algebras. This will in particular yield a deeper insight into central
simple algebras. In the final part of the chapter we will prove two classical results, the
Double Centralizer Theorem and the Skolem-Noether Theorem, and thereby illus-
trate the usefulness and the strength of the approach based on tensor products and
modules.

4.1 Concept of a Tensor Product

Take two vector spaces U and V over a field F. For simplicity assume, temporarily,
that they are finite dimensional, dim F U = m and dim F V = n. The tensor product
of U and V is a vector space of dimension mn. One is usually interested in algebraic
objects only up to isomorphism, and so might consider this description of the tensor
product as satisfactory. However, it does not really give a clue about the idea upon
which this concept is based. The tensor product serves as a device transforming
bilinear maps into linear ones. This ambiguous description in some sense better
reflects the essence of this concept. Before starting with a rigorous treatment, let us
give an indication in what way the latter could be connected with an mn-dimensional
space. Pick a basis {e1 , . . . , em } of U and a basis { f 1 , . . . , f n } of V . One easily notices
that a bilinear map β from U ×V into another vector space W is uniquely determined
by its mn values β(ei , f j ), 1 ≤ i ≤ m, 1 ≤ j ≤ n. Now, take any vector space of
dimension mn, choose its basis, and define a linear map β from this space into W
so that the elements from the chosen basis are mapped into the elements β(ei , f j ).
It is clear that β together with the chosen bases contains all information about β.
© Springer International Publishing Switzerland 2014 79
M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_4
80 4 Tensor Products

Thus, instead of considering a bilinear map on U × V , we can consider a linear map

defined on a space of dimension dim F U · dim F V .
We now proceed to the formal definition. Its major advantage comparing to the
above discussion is that it is independent of basis. Before stating the definition we
make a few comments in order to avoid a possible misapprehension.
Given any set X we can construct a vector space over F having X as a basis
(cf. Sect. 2.5). Its elements
 are formal linear combinations of elements from X , i.e.,
they can be written as x∈X λx x where all but finitely many λx ∈ F are zero. Thus
X basically plays the role of an index set. Operations in this space are defined in the
obvious way. Ignoring the terms with zero coefficient we get another interpretation:
For every element w in the vector space with basis X there exist a finite subset
{x1 , . . . , xn } of X and unique λ1 , . . . , λn ∈ F such that w = λ1 x1 + · · · + λn xn .
We will deal with the situation where X is the Cartesian product U × V of spaces U
and V . Let us emphasize that here we consider U × V only as a set, not as a vector
space. Therefore, for example, (0, 0) is just an element from the standard basis of the
space with basis U × V , it is not its zero element. Further, (u + u  , v) is an element
from this basis and is therefore different from (u, v) + (u  , v) which is the sum of
two basis elements, while λ(u, v) coincides with (λu, v) only when λ = 1. We may
thus think of the space with basis U × V as being “huge”, incomparably larger than
U and V .
Let us now state the definition. It concerns arbitrary vector spaces, not necessarily
finite dimensional ones.

Definition 4.1 Let U and V be vector spaces over F. Denote by Y the vector space
with basis U × V . Let N be the subspace of Y generated by all elements of the
form

(λu + λ u  , v) − λ(u, v) − λ (u  , v),

(u, λv + λ v ) − λ(u, v) − λ (u, v ),

λ, λ ∈ F, u, u  ∈ U , v, v ∈ V . The tensor product of U and V is the factor vector

space Y /N . It is denoted by U ⊗ V (or by U ⊗ F V when we wish to emphasize
that we consider spaces over F).

This definition is of little practical use. The characteristic property of U ⊗ V ,

described in the following theorem, is the one that matters.

Theorem 4.2 Let U and V be vector spaces over F. Then there exists a bilinear
map U × V → U ⊗ V , (u, v) → u ⊗ v, such that:
(a) Every element in U ⊗ V is a sum of elements of the form u ⊗ v, u ∈ U , v ∈ V .
(b) Given a bilinear map β : U × V → W , where W is a vector space over F,
there exists a linear map β : U ⊗ V → W such that β(u ⊗ v) = β(u, v) for all
u ∈ U, v ∈ V .
Moreover, properties (a) and (b) characterize U ⊗ V up to isomorphism.
4.1 Concept of a Tensor Product 81

Proof We define u ⊗ v as the coset (u, v) + N . Considering the generators of N

we see that the map (u, v) → u ⊗ v is bilinear. By definition, every element in
U ⊗ V can be written as a linear combination of elements of the form u ⊗ v. Since
λ(u ⊗ v) = (λu) ⊗ v, (a) follows.
Let β : U × V → W be a bilinear map. Since U × V is a basis of Y , there exists
a linear map B : Y → W such that B (u, v) = β(u, v) for all u ∈ U , v ∈ V (cf.
Remark 3.29). The bilinearity of β implies that N ⊆ ker B. Accordingly, we can
define a linear map β on Y /N = U ⊗ V by β(x + N ) = B(x). In particular,
β(u ⊗ v) = β(u, v).
To prove the last statement, assume that T is another space for which there exists
a bilinear map U × V → T , (u, v) → u v, with properties (a) and (b). Since
(u, v) → u ⊗ v is a bilinear map, there exists a linear map ϕ : T → U ⊗ V
satisfying ϕ(u v) = u ⊗ v. Analogously, there exists a linear map ψ : U ⊗ V → T
satisfying ψ(u ⊗ v) = u v. Consequently, ϕ and ψ are inverses of each other, and
hence T and U ⊗ V are isomorphic vector spaces. 

Remark 4.3 From (a) it follows that β is a unique linear map from U ⊗ V into W
sending u ⊗ v into β(u, v).

Bilinear maps appear so often in mathematics that we simply cannot avoid them.
However, they may be a bit clumsy to deal with. Theorem 4.2 shows that we can
“glue” spaces U and V into the space U ⊗ V so that a bilinear map defined on U × V
can be credibly represented by a linear map defined on U ⊗ V . Linear maps are not
only simpler than bilinear ones, they are the “proper” maps (i.e., the homomorphisms)
in the context of vector spaces.

Example 4.4 Let A be an F-algebra. The multiplication on A is by definition a

bilinear map (x, y) → x y from A × A into A. Sometimes it is more appropriate to
consider multiplication as a linear map from A⊗ A into A, determined by x ⊗y → x y.

Let us mention that one can also define tensor products of modules. For modules
over commutative rings this can be done in basically the same way as for vec-
tor spaces, the differences are merely terminological. More substantial changes are
needed if rings are noncommutative.

4.2 Basic Properties of Tensor Products

The results of this and of the next couple of sections should convince the reader that
the concept of the tensor product is natural. It is compatible with fundamental linear
algebra concepts.
Let us first summarize the most relevant properties of the tensor product estab-
lished so far. The tensor product of vector spaces U and V is a vector space U ⊗ V
consisting of elements that can be written as

u 1 ⊗ v1 + u 2 ⊗ v2 + · · · + u n ⊗ vn , u i ∈ U, vi ∈ V.
82 4 Tensor Products

Such an expression is far from unique. After all, we have to take into account that
the map (u, v) → u ⊗ v is bilinear, meaning that

(λu + λ u  ) ⊗ v = λ(u ⊗ v) + λ (u  ⊗ v),

u ⊗ (λv + λ v ) = λ(u ⊗ v) + λ (u ⊗ v )

for all λ, λ ∈ F, u, u  ∈ U , v, v ∈ V . These formulas readily imply

u ⊗ 0 = 0, 0 ⊗ v = 0

for all u ∈ U , v ∈ V . Last but not least, we have the so-called universal property
(b) from Theorem 4.2, which might be easier to memorize through a commutative
diagram:

Here, ⊗ denotes the canonical map (u, v) → u ⊗ v. Let us emphasize that ⊗ and β
are bilinear maps, while β is linear.
Elements of the form u⊗v are called simple tensors. One should keep in mind that
these are only generators of the space U ⊗ V , not its typical elements. This warning
may seem redundant at this point, but actually its disregard is a common source of
errors. Standard tensor product considerations deal only with simple tensors, which
may be misleading. Say, Theorem 4.2 describes β by β(u ⊗ v) = β(u, v); this is
all right since β is linear and so its action on U ⊗ V is uniquely determined by this
rule. There are, however, other instances where an information about simple tensors
is not sufficient.
Let us examine further properties of tensor products. First a word about notation:
throughout this and the next section, U, U  , V, V  , W denote arbitrary vector spaces
over the field F. The next lemma introduces the tensor product of linear maps.

Lemma 4.5 If ϕ : U → U  and ψ : V → V  are linear maps, then there exists a

unique linear map U ⊗ V → U  ⊗ V  , which we denote by ϕ ⊗ ψ, such that

(ϕ ⊗ ψ)(u ⊗ v) = ϕ(u) ⊗ ψ(v)

for all u ∈ U , v ∈ V .

Proof The linearity of ϕ and ψ, together with the bilinearity of ⊗, clearly imply that
the map (u, v) → ϕ(u) ⊗ ψ(v) is bilinear. Theorem 4.2 thus shows that the linear
map sending u ⊗ v into ϕ(u) ⊗ ψ(v) indeed exists. Its uniqueness is obvious. 
4.2 Basic Properties of Tensor Products 83

The lemma basically states that the map given by

 
u i ⊗ vi → ϕ(u i ) ⊗ ψ(vi )
i i

is well-defined. The given proof is quite illustrative. In order to establish something

about tensor products one usually begins by introducing a bilinear map, and then, by
the universal property, transforms it into a linear one. As a rule this quickly yields
conclusions which may not be so easily obtained by other means.
It is clear that for linear maps with appropriate domains and codomains we have

(ϕ ⊗ ψ)(ϕ  ⊗ ψ  ) = ϕϕ  ⊗ ψψ  .

Consequently, if ϕ and ψ are isomorphisms, then so is ϕ ⊗ ψ, and

(ϕ ⊗ ψ)−1 = ϕ −1 ⊗ ψ −1 .

This in particular shows the following:

= U  and V ∼
U∼ = V  =⇒ U ⊗ V ∼
= U  ⊗ V .

Let us also mention the formulas

(λϕ + λ ϕ  ) ⊗ ψ = λ(ϕ ⊗ ψ) + λ (ϕ  ⊗ ψ),

ϕ ⊗ (λψ + λ ψ  ) = λ(ϕ ⊗ ψ) + λ (ϕ ⊗ ψ  ),

which follow immediately from the definition.

In the next lemmas we prove that the tensor product is commutative and associa-
tive. One should not understand this literally; such statements can only be true up to
isomorphism. However, such abuse of terminology is quite standard.

Lemma 4.6 U ⊗ V ∼
= V ⊗ U.

Proof The map U × V → V ⊗ U , (u, v) → v ⊗ u, is bilinear, hence there exists

a linear map U ⊗ V → V ⊗ U , u ⊗ v → v ⊗ u. Similarly there is a linear map
V ⊗ U → U ⊗ V , v ⊗ u → u ⊗ v. Obviously, these two maps are inverses of
each other. 

Lemma 4.7 (U ⊗ V ) ⊗ W ∼
= U ⊗ (V ⊗ W ).

Proof The bulk of the proof is to show that the rule (u ⊗ v) ⊗ w → u ⊗ (v ⊗ w)

determines a well-defined map from (U ⊗V )⊗W to U ⊗(V ⊗W ). Since the elements
(u ⊗v)⊗w generate the vector space (U ⊗V )⊗W (why?), the symmetrically defined
map u ⊗ (v ⊗ w) → (u ⊗ v) ⊗ w is then obviously its inverse.
84 4 Tensor Products

Take w ∈ W . The map U × V → U ⊗ (V ⊗ W ), (u, v) → u ⊗ (v ⊗ w), is

bilinear, and hence u ⊗ v → u ⊗ (v ⊗ w) determines a well-defined map from U ⊗ V
to U ⊗ (V ⊗ W ). This is true for every w ∈ W . Accordingly,
  
u i ⊗ vi , w → u i ⊗ (vi ⊗ w)
i i

is a well-defined bilinear map from (U ⊗ V ) × W to U ⊗ (V ⊗ W ). Therefore the

map determined by (u ⊗ v) ⊗ w → u ⊗ (v ⊗ w) is indeed well-defined. 
One usually identifies (U ⊗ V ) ⊗ W and U ⊗ (V ⊗ W ), and denotes each of these
two spaces simply by

U ⊗ V ⊗ W.

Alternatively one can introduce U ⊗ V ⊗ W via trilinear maps; the resulting space is
the same up to isomorphism. Of course, the tensor product of more than three spaces
can also be defined.

4.3 Linear (In)dependence in Tensor Products

An element in U ⊗ V can be written as a sum of simple tensors in many different

ways. Fortunately, this non-uniqueness has its limitations.
Lemma 4.8 Let e1 , . . . , en ∈ U be linearly independent. If v1 , . . . , vn ∈ V are such
that
e1 ⊗ v1 + · · · + en ⊗ vn = 0,

then each vi = 0.
Proof Let f i : U → F be a linear functional such that f i (e j ) = δi j for j = 1, . . . , n
(cf. Theorem 3.28 and Remark 3.29). The map (u, v) → f i (u)v is bilinear, and
so there exists a linear map αi : U ⊗ V → V satisfying αi (u ⊗ v) = f i (u)v.
Consequently,


n 
n 
vi = αi (e j ⊗ v j ) = αi e j ⊗ v j = 0.
j=1 j=1 
All results in this section will follow from this simple lemma. Let us first point
out that in particular it shows that for all u ∈ U and v ∈ V ,

u ⊗ v = 0 =⇒ u = 0 or v = 0.

Incidentally, the tensor product of modules over rings does not always enjoy this
property.
4.3 Linear (In)dependence in Tensor Products 85

We continue with a useful technical improvement of Lemma 4.8.

Lemma 4.9 Let e1 , . . . , en ∈ U be linearly independent. If v1 , . . . , vn ∈ V are

such that

e1 ⊗ v1 + · · · + en ⊗ vn = w1 ⊗ z 1 + · · · + wm ⊗ z m

for some w1 . . . , wm ∈ U and z 1 , . . . , z m ∈ V , then each vi is a linear combination

of z 1 , . . . , z m .

Proof The set {e1 , . . . , en } can be extended to a basis of the linear span of the set
{e1 , . . . , en , w1 , . . . , wm }. Denote the additional vectors (if there are any) in this basis
by en+1 , . . . , e p . Write each w j as a linear combination of basis vectors. Then, using
the bilinearity of ⊗, we see that the given identity can be rewritten as

e1 ⊗ v1 + · · · + en ⊗ vn + en+1 ⊗ z n+1


+ · · · + e p ⊗ z p = 0,

where each vi is the sum of vi and a linear combination of the z j ’s. Lemma 4.8 tells
us that vi = 0. 

Let us also record the obvious analogues of the above lemmas.

Lemma 4.10 Let f 1 , . . . , f n ∈ V be linearly independent. If u 1 , . . . , u n ∈ U are

such that

u 1 ⊗ f 1 + · · · + u n ⊗ f n = 0,

then each u i = 0.

Lemma 4.11 Let f 1 , . . . , f n ∈ V be linearly independent. If u 1 , . . . , u n ∈ U are

such that

u 1 ⊗ f 1 + · · · + u n ⊗ f n = w1 ⊗ z 1 + · · · + wm ⊗ z m

for some w1 , . . . , wm ∈ U and z 1 . . . , z m ∈ V , then each u i is a linear combination

of w1 , . . . , wm .

Bases of U and V give rise to bases of U ⊗ V :

Theorem 4.12 If {ei | i ∈ I } is a basis of U , and { f j | j ∈ J } is a basis of V , then

{ei ⊗ f j | i ∈ I, j ∈ J } is a basis of U ⊗ V .

Proof Writing u ∈ U as a linear combination of the ei ’s, and v ∈ V as a linear

combination of the f j ’s, it follows, by the bilinearity of ⊗, that u ⊗ v is a linear
combination of the ei ⊗ f j ’s. Therefore the set {ei ⊗ f j | i ∈ I, j ∈ J } generates
U ⊗ V . It remains to show that it is linearly independent. By definition, a subset of a
vector space is linearly independent if each of its finite subsets is linearly independent.
86 4 Tensor Products

Every finite subset of {ei ⊗ f j | i ∈ I, j ∈ J } is contained in a set of the form

{eik ⊗ f jl | k = 1, . . . , m, l = 1, . . . , n}. Therefore it suffices to show that such sets
are linearly independent. Assume, therefore, that


m 
n
λkl (eik ⊗ f jl ) = 0
k=1 l=1

for some λkl ∈ F. We can rewrite this as


m 
n 
eik ⊗ λkl f jl = 0.
k=1 l=1
n
Since ei1 , . . . , eim are linearly independent, l=1 λkl f jl = 0 for all k by Lemma
4.8. The linear independence of f j1 , . . . , f jn yields λkl = 0 for all k and l. 

The next two corollaries are immediate.

Corollary 4.13 If {ei | i ∈ I

} is a basis of U , then every element in U ⊗ V can be
written uniquely in the form i∈I ei ⊗ vi .

If I is an infinite set, then this should be understood as that all but finitely many
vi ’s are zero.

Corollary 4.14 If { f j | j ∈ J} is a basis of V , then every element in U ⊗ V can be

written uniquely in the form j∈J u j ⊗ f j .

Remark 4.15 Another useful d observation is that every nonzero element x ∈ U ⊗ V

can be written as x = i=1 u i ⊗ vi where u 1 , . . . , u d are linearly independent
vectors in U and v1 , . . . , vd are linearly independent vectors in V . The proof is easy:
just require the minimality of d with respect to all possible expressions of x, and the
independence of the u i ’s and of the vi ’s clearly follows.

Let now I and J be finite sets. Theorem 4.12 then yields the formula announced
at the beginning of the chapter.

Corollary 4.16 If U and V are finite dimensional vector spaces over F, then

dim F U ⊗ V = dim F U · dim F V.

Observe the analogy with the formula on direct sums:

dim F U ⊕ V = dim F U + dim F V.

Remark 4.17 Corollary 4.16 resembles another formula from basic linear algebra:
The dimension of Hom F (U, V ), the vector space of all linear maps from U to V ,
is equal to dim F U · dim F V . Therefore the spaces Hom F (U, V ) and U ⊗ V are
4.3 Linear (In)dependence in Tensor Products 87

isomorphic. What we wish to point out is a canonical isomorphism—yet not between

exactly these two spaces. Let U ∗ be the dual space of U (i.e., the space of all linear
functionals on U ). Instead of U ⊗ V we consider V ⊗ U ∗ which is a space of
the same dimension. Given v ∈ V and f ∈ U ∗ , define Tv, f ∈ Hom F (U, V ) by
Tv, f (u) = f (u)v. Note that Tv, f has rank one provided that v = 0 and f = 0,
and conversely, every linear map U → V of rank one is of the form Tv, f for some
0 = v ∈ V and 0 = f ∈ U ∗ . The map (v, f ) → Tv, f is bilinear, hence there exists
a linear map

ϕ : V ⊗ U ∗ → Hom F (U, V ), ϕ(v ⊗ f ) = Tv, f .

As every nonzero map in Hom F (U, V ) is a sum of maps of rank one, ϕ is surjective.
But then ϕ is an isomorphism for it is linear and the spaces V ⊗U ∗ and Hom F (U, V )
have the same dimension. One often simply identifies v ⊗ f with Tv, f , and thus
considers simple tensors v ⊗ f as rank one maps given by

(v ⊗ f )(u) = f (u)v.

This is commonly used regardless of whether U and V are finite dimensional or not.
Remark 4.18 The following elementary facts

= U and F ⊗ V ∼
U⊗F∼ =V

also deserve to be mentioned. The proofs are immediate. By bilinearity of ⊗ every

element in U ⊗ F can be written as u ⊗ 1. Therefore u ⊗ 1 → u is an isomorphism
from U ⊗ F onto U . Similarly, 1 ⊗ v → v is an isomorphism from F ⊗ V onto V .

4.4 Tensor Product of Algebras

After a linear algebra inset we are now ready to return to themes that are in the spirit
of this book. Let A and B be algebras over a field F. The vector space A ⊗ B can be
turned into an algebra by defining multiplication in a simple, natural way.

Lemma 4.19 If A and B are F-algebras, then A ⊗ B is an F-algebra relative to

the multiplication determined by

(x ⊗ y)(z ⊗ w) = x z ⊗ yw

for all x, z ∈ A, y, w ∈ B.

Proof Pick z ∈ A and w ∈ B. Let Rz and Rw denote the corresponding right

multiplication maps. By Lemma 4.5 we have the map Rz ⊗ Rw ∈ End F (A ⊗ B) that
sends x ⊗ y into x z ⊗ yw. Since
88 4 Tensor Products

A × B → End F (A ⊗ B), (z, w) → Rz ⊗ Rw ,

is a bilinear map, there exists a linear map

ϕ : A ⊗ B → End F (A ⊗ B), ϕ(z ⊗ w) = Rz ⊗ Rw .

We now define the product of r, s ∈ A ⊗ B as follows:

r s := ϕ(s)(r ).

This multiplication is obviously bilinear, and for all x, z ∈ A, y, w ∈ B we have

(x ⊗ y)(z ⊗ w) = ϕ(z ⊗ w)(x ⊗ y) = (Rz ⊗ Rw )(x ⊗ y) = x z ⊗ yw.

Finally, from the associativity in A and B we immediately infer that

   
(x ⊗ y)(z ⊗ w) (u ⊗ v) = (x ⊗ y) (z ⊗ w)(u ⊗ v) .

Since every element in A ⊗ B is a sum of simple tensors, this implies that our
multiplication is associative. 

Lemma 4.19 basically states that

   
xi ⊗ yi z j ⊗ w j := xi z j ⊗ yi w j
i j i, j

is a well-defined operation on A ⊗ B. From now on we will always consider A ⊗ B,

the tensor product of algebras A and B, as an algebra in which multiplication is
defined by the above formula.
By Lemma 4.6 we know that A ⊗ B and B ⊗ A are isomorphic as vector spaces.
The canonical isomorphism x ⊗ y → y ⊗ x also preserves multiplication. Therefore
we have an algebra isomorphism

A⊗B ∼
= B ⊗ A.

Similarly, just as for tensor products of vector spaces we see that

(A ⊗ B) ⊗ C ∼
= A ⊗ (B ⊗ C),

= A and F ⊗ B ∼
A⊗F ∼ = B,

= A and B ∼
A∼ = B  =⇒ A ⊗ B ∼
= A ⊗ B 

holds for all F-algebras A, A , B, B  , and C.

4.4 Tensor Product of Algebras 89

Let S be a subalgebra of A and T be a subalgebra of B. Regarding S and T

as algebras we can form their tensor product S ⊗ T , which is clearly canonically

isomorphic to the subalgebra of A⊗B consisting of all elements of the form i si ⊗ti ,
si ∈ S, ti ∈ T . In this sense we can consider S ⊗ T as a subalgebra of A ⊗ B.
We continue with illuminating examples.

Example 4.20 Let G and H be groups. What is the tensor product of group algebras
F[G] and F[H ]? From Theorem 4.12 we know that the set {g ⊗ h | g ∈ G, h ∈ H }
is a basis of F[G] ⊗ F[G]. By definition, G × H is a basis of the group algebra
F[G × H ]. Therefore there exists a bijective linear map F[G] ⊗ F[H ] → F[G ×
H ] determined by g ⊗ h → (g, h). This map is clearly also multiplicative, so
we have
F[G] ⊗ F[H ] ∼ = F[G × H ].

Example 4.21 Let A be an arbitrary F-algebra. We claim that

A ⊗ F[ω] ∼
= A[ω].

 Corollary 4.14 we see that every element in A ⊗ F[ω] can be uniquely

Indeed, using
written as i ai ⊗ ωi . It is now straightforward to check that the map
 
A ⊗ F[ω] → A[ω], ai ⊗ ωi → ai ωi ,
i i

is an algebra isomorphism.
In particular, F[ω1 ] ⊗ F[ω2 ] ∼ = (F[ω1 ])[ω2 ] ∼
= F[ω1 , ω2 ]. Note that, via this
isomorphism, the element f ⊗ g ∈ F[ω1 ] ⊗ F[ω2 ] corresponds to the polynomial
f (ω1 )g(ω2 ) ∈ F[ω1 , ω2 ]. It is in fact common not to distinguish between the two,
and thus consider f ⊗ g as the polynomial

( f ⊗ g)(ω1 , ω2 ) = f (ω1 )g(ω2 ).

Example 4.22 For every F-algebra A and every n ≥ 1 we have

Mn (F) ⊗ A ∼
= Mn (A). (4.1)

The proof is similar to that from the preceding example. After noticing that every
element in Mn (F)⊗ A can be uniquely written as the sum of simple tensors E i j ⊗ai j ,
where E i j are the standard matrix units, it readily follows that


n 
n 
n 
n
Mn (F) ⊗ A → Mn (A), E i j ⊗ ai j → ai j E i j ,
i=1 j=1 i=1 j=1
90 4 Tensor Products

is an algebra isomorphism. This map should be considered as canonical—one often

identifies (λi j ) ⊗ a ∈ Mn (F) ⊗ A with the matrix in Mn (A) whose (i, j) entry is
λi j a.
As an important special case of (4.1) we have

Mn (F) ⊗ Mm (F) ∼
= Mnm (F). (4.2)

Namely, Mn (Mm (F)) ∼ = Mnm (F). This is basically just a rewording of the
well-known fact that matrices can be added and multiplied blockwise. A detailed
proof is perhaps a bit tedious, but straightforward.
Alternatively, we can state (4.2) as

End F (U ) ⊗ End F (V ) ∼
= End F (U ⊗ V ),

where U and V are finite dimensional vector spaces over F.

Example 4.23 Let V be a vector space with a countably infinite basis {e1 , e2 , . . . }.
Set A := End F (V ) and define f i j ∈ A, i, j = 1, 2, according to

f 11 (e2k−1 ) = e2k−1 , f 11 (e2k ) = 0,

f 12 (e2k−1 ) = 0, f 12 (e2k ) = e2k−1 ,
f 21 (e2k−1 ) = e2k , f 21 (e2k ) = 0,
f 22 (e2k−1 ) = 0, f 22 (e2k ) = e2k

for all k ∈ N. One can check that { f 11 , f 12 , f 21 , f 22 } forms a set of 2 × 2 matrix

units of A. Therefore A ∼ = M2 ( f 11 A f 11 ) by Lemma 2.52 (and Remark 2.53). It is
an easy exercise to show that f 11 A f 11 ∼= End F (V1 ) where V1 is the subspace of V
generated by all e2k−1 , k ∈ N. As the spaces V1 and V are isomorphic, we actually
= A, and hence A ∼
have End F (V1 ) ∼ = M2 (A). It is not difficult to generalize this to

A∼
= Mn (A) for every n ∈ N.

We can now better understand and appreciate the uniqueness in Wedderburn’s the-
orem (Theorem 3.58). Further, now we see that the tensor product of algebras does
not obey the “cancellation law”, i.e.,

S⊗A∼
= T ⊗ A =⇒
 S∼
= T.

Namely, by (4.1) we have F ⊗ A ∼

= Mn (F) ⊗ A for every n ∈ N.
4.5 Multiplication Algebra and Tensor Products 91

4.5 Multiplication Algebra and Tensor Products

Let A and B be unital algebras. Then 1 ⊗ 1 is clearly the unity of A ⊗ B. One usually
identifies A with A⊗1 := {a⊗1 | a ∈ A} (= A⊗ F) and B with 1⊗ B := {1⊗b | b ∈
B}, and in this way considers A and B as subalgebras of A ⊗ B. Obviously, elements
from A commute with elements from B, and the algebra A⊗ B is generated by A∪ B.
If we add the statement of Lemma 4.8 to these properties, we get a characterization
of the tensor product of unital algebras:

Lemma 4.24 Let T be a unital algebra. Suppose T contains subalgebras A and B

which contain the unity of T and satisfy
(a) ab = ba for all a ∈ A, b ∈ B.
(b) T is generated by A ∪ B.
(c) If a1 , . . . , an ∈ A are linearly independent and b1 , . . . , bn ∈ B are arbitrary,
then a1 b1 + · · · + an bn = 0 implies b1 = · · · = bn = 0.
Then A ⊗ B ∼
= T under the map a ⊗ b → ab.

Proof The standard universal property argument shows that there exists a linear map
ϕ : A ⊗ B → T such that ϕ(a ⊗ b) = ab for all a ∈ A, b ∈ B. According to (a),
(b), and the assumption that A and B contain the unity of T , every element in T is of
the form a1 b1 + · · · + an bn with ai ∈ A and bi ∈ B. Therefore ϕ is surjective. Since
every x ∈ A ⊗ B can be written as x = a1 ⊗ b1 + · · · + an ⊗ bn , where a1 , . . . , an
are linearly independent (cf. Remark 4.15), we see from (c) that ϕ(x) = 0 implies
x = 0. Therefore ϕ is injective. Finally, using (a) we obtain
 
ϕ (a ⊗ b)(a  ⊗ b ) = ϕ(aa  ⊗ bb ) = aa  bb = aba  b = ϕ(a ⊗ b)ϕ(a  ⊗ b ),

which shows that ϕ is an algebra isomorphism. 

Remark 4.25 It is instructive to compare (a)–(c) with analogous properties of the

direct product. Suppose now that an algebra S contains subalgebras A and B such
that
(a’) ab = ba = 0 for all a ∈ A, b ∈ B.
(b’) S is the sum of A and B.
(c’) If a ∈ A and b ∈ B are such that a + b = 0, then a = b = 0.
Then A × B ∼
= S under the map (a, b) → a + b.

As it will be apparent from the next proof, the situation considered in Lemma
4.24 was indirectly met already in Sect. 1.5 when studying the multiplication algebra
M(A) of a central simple algebra A.

Theorem 4.26 If A is a central simple algebra, then A ⊗ Ao ∼

= M(A) under the
map a ⊗ b → L a Rb .
92 4 Tensor Products

Proof The sets L = {L a | a ∈ A} and R = {Rb | b ∈ A} of all left and right

multiplication maps are subalgebras of M(A). By definition, M(A) is generated by
L ∪ R. Obviously, elements from L commute with elements from R. It is clear
that a1 , . . . , an ∈ A are linearly independent if and only if L a1 , . . . , L an ∈ L are
linearly independent. In this case L a1 Rb1 + · · · + L an Rbn = 0 implies bi = 0 (and
hence Rbi = 0) by Lemma 1.24. Lemma 4.24 therefore tells us that

L ⊗R ∼
= M(A) under the map L a ⊗ Rb → L a Rb .

We can polish this a little bit. Note that A ∼

= L under a → L a . Since Rbb = Rb Rb ,
there is no reason why A and R should be isomorphic. However, Ao and R are;
b → Rb is, of course, an isomorphism. Making use of Lemma 4.5 (and the comments
following this lemma) we thus see that

A ⊗ Ao ∼
= L ⊗ R under the map a ⊗ b → L a ⊗ Rb .

Composing the isomorphisms a ⊗ b → L a ⊗ Rb and L a ⊗ Rb → L a Rb we arrive

at the desired conclusion. 
Lemma 1.24 can now be sharpened as follows.
Corollary 4.27 If A is a central
 simple algebra, then for all ai , bi ∈ A we have
n n
i=1 L ai Rbi = 0 if and only if i=1 ai ⊗ bi = 0.
Corollary 4.28 If A is a finite dimensional central simple algebra, then A ⊗ Ao ∼=
Mn (F), where n = [A : F].
Proof Apply Lemma 1.25, which says that M(A) = End F (A), together with the
standard fact End F (A) ∼
= Mn (F). 
Example 4.29 The conjugation h → h is an antiautomorphism of H. Therefore
Ho ∼
= H under the map h → h, and hence H⊗R Ho ∼
= H⊗R H. As H⊗R Ho ∼
= M4 (R)
by Corollary 4.28, we have

H ⊗R H ∼
= M4 (R).

4.6 Centralizers in Tensor Products

The center Z (A) of an algebra A is the set of elements in A that commute with all
elements in A. We now extend this concept by considering sets of elements in A that
commute with some elements in A.
Definition 4.30 Let S be a subset of an algebra A. The centralizer of S in A is the
set

C A (S) := {a ∈ A | as = sa for every s ∈ S}.

4.6 Centralizers in Tensor Products 93

In the same way one defines the centralizer of a subset of a ring, but we shall
deal only with centralizers in algebras. Let us also mention that we have encountered
centralizers of singletons in the proof of Wedderburn’s theorem on finite division
rings.
First we record a few simple remarks to build intuition. The centralizer of any set
S is a subalgebra of A which contains Z (A). On the other hand, the centralizer of S
coincides with the centralizer of the subalgebra generated by S and Z (A). Therefore
one usually considers only centralizers of subalgebras containing the center. One
might expect that, roughly speaking, a large subalgebra has a small centralizer, and
a small subalgebra has a large centralizer. Let us support this inexact statement with
the following obvious facts:

S1 ⊆ S2 =⇒ C A (S2 ) ⊆ C A (S1 ),

C A (A) = Z (A) and C A (Z (A)) = A.

Condition (a) in Lemma 4.24 can be read as B ⊆ C T (A) (or A ⊆ C T (B)). This
suggests that the concept of the centralizer plays a role in tensor products of algebras.
Our starting point is

Proposition 4.31 Let A and B be unital algebras. If S is a subalgebra of A that

contains the unity of A, and T is a subalgebra of B that contains the unity of B, then

C A⊗B (S ⊗ T ) = C A (S) ⊗ C B (T ).

Proof It is enough to prove C A⊗B (S ⊗ T ) ⊆ C A (S) ⊗ C B (T ), since the converse

d
inclusion is trivial. Let c ∈ C A⊗B (S ⊗ T ). We can write c = i=1 yi ⊗ z i where
y1 , . . . , yd ∈ A are linearly independent and z 1 , . . . , z d ∈ B are linearly independent
(cf. Remark 4.15). From c(s ⊗ 1) = (s ⊗ 1)c, s ∈ S, it follows that


d
(yi s − syi ) ⊗ z i = 0.
i=1

Lemma 4.10 implies that yi s − syi = 0 for every i and every s ∈ S. Therefore each
yi lies in C A (S). Analogously, by considering c(1 ⊗ t) = (1 ⊗ t)c, t ∈ T , we infer
that each z i lies in C B (T ). Thus c ∈ C A (S) ⊗ C B (T ). 

Corollary 4.32 If A and B are unital algebras, then Z (A ⊗ B) = Z (A) ⊗ Z (B).

Proof Use Proposition 4.31 for S = A and T = B. 

In the statement of the next corollary we identify A with A ⊗ 1 and B with 1 ⊗ B.

Corollary 4.33 If A and B are central algebras, then A⊗ B is also a central algebra,
C A⊗B (A) = B and C A⊗B (B) = A.
94 4 Tensor Products

Proof The first assertion follows from Corollary 4.32, the second one from
Proposition 4.31 by taking S = A and T = F ( = F · 1), and the third one from
Proposition 4.31 by taking S = F and T = B. 

Corollary 4.34 Let A be a unital algebra and let S be its subalgebra that contains
the unity of A. If B is a central algebra, then C A⊗B (S ⊗ B) ∼
= C A (S). In particular,
C Mn (A) (Mn (S)) ∼= A
C (S) for every n ∈ N.

Proof Proposition 4.31 shows that C A⊗B (S ⊗ B) = C A (S) ⊗ F ∼

= C A (S). The last
assertion follows by taking Mn (F) for B (cf. Example 4.22). 

If S is a subalgebra of A, then we can consider A as a (left) S-module (cf. Example

3.8). Therefore it makes sense to speak about End S (A).

Corollary 4.35 Let A be a finite dimensional central simple algebra. If S is a sub-

algebra of A that contains the unity of A, then C A (S) ⊗ Ao ∼
= End S (A).

Proof Theorem 4.26, along with Lemma 1.25, tells us that a ⊗ b → L a Rb defines
an isomorphism from A ⊗ Ao onto End F (A). Accordingly, the centralizer of S ⊗ F
in A ⊗ Ao is isomorphic to the centralizer of {L s | s ∈ S} in End F (A). The former
equals C A (S) ⊗ Ao by Proposition 4.31, while the latter is nothing but End S (A). 

4.7 Scalar Extension (The “Right” Approach)

We are now in a position to place the notion of the scalar extension, already discussed
in Sect. 2.11, on a solid foundation. Recall the basic idea behind the definition: Given
an F-algebra A and an extension field K of F, we wish to construct a K -algebra
A K that has essentially the same multiplication as A. As we have seen, relying on
a fixed chosen basis this can be achieved by elementary means. We will now show
that by making use of the tensor product this construction can be accomplished
independently of basis.
Recall that we can consider an extension field K of F as an F-algebra. Therefore
K ⊗ F A is an F-algebra. The point is to convert it into a K -algebra.

Lemma 4.36 Let A be an F-algebra and let K be an extension field of F. Then

K ⊗ F A is a K -algebra relative to the scalar multiplication determined by

λ(μ ⊗ x) = (λμ) ⊗ x, λ, μ ∈ K , x ∈ A.

Proof Take λ ∈ K . The map

K × A → K ⊗ A, (μ, x) → (λμ) ⊗ x,

is bilinear, and hence there is an F-linear map

4.7 Scalar Extension (The “Right” Approach) 95

K ⊗ F A → K ⊗ F A, μ ⊗ x → (λμ) ⊗ x

(incidentally, this also follows by using Lemma 4.5 for the (left) multiplication map
L λ on K and the identity map on A). Our scalar multiplication is thus well-defined
and λ(r + s) = λr + λs holds for λ ∈ K , r, s ∈ K ⊗ F A. It is straightforward to
check that we also have

(λ + μ)r = λr + μr, (λμ)r = λ(μr ), λ(r s) = (λr )s = r (λs)

for λ, μ ∈ K , r, s ∈ K ⊗ F A. 
Definition 4.37 Let A be an F-algebra and let K be an extension field of F. The
K -algebra K ⊗ F A (defined in Lemma 4.36) is called the scalar extension of A to
K . It will be denoted by A K .
The following assertions are immediate:
• 1 ⊗ A is an F-subalgebra of A K isomorphic to A.
• A K is linearly spanned by 1 ⊗ A.
• If {ei | i ∈ I } is a basis of A, then {1 ⊗ ei | i ∈ I } is a basis of A K (cf. Corollary
4.14).
Assume now that A is finite dimensional. Then we obviously have

[A K : K ] = [A : F].

If {e1 , . . . , ed } is a basis of A with multiplication table


d
ei e j = αi jk ek , αi jk ∈ F,
k=1

then the elements ei = 1 ⊗ ei form a basis of A K with the same multiplication table:


d
ei e j = αi jk ek .
k=1

Neglecting the insignificant notational differences, we thus see that Definition 4.37
coincides with the definition given in Sect. 2.11. Being independent of basis, the new
definition is not only more clear and elegant, but also more suitable for practical use.
We already saw in Sect. 2.11 that scalar extensions can be used to transform
finite dimensional central simple algebras into the matrix algebras over fields. Let us
consider this phenomenon from the conceptual standpoint.
Definition 4.38 Let A be a finite dimensional central simple F-algebra. An exten-
sion field K of F is said to be a splitting field for A if A K ∼
= Mn (K ) for some
n ∈ N.
96 4 Tensor Products

We can now reword Theorem 2.74 as follows.

Theorem 4.39 An algebraic closure F of F is a splitting field for every finite dimen-
sional central simple F-algebra.
For division algebras we have another, more intrinsic class of splitting fields:
Theorem 4.40 Let D be a finite dimensional central division algebra. Every maxi-
mal subfield K of D is a splitting field for D.
Proof If a ∈ C D (K ), then a and K generate a field. Therefore a ∈ K since K is
maximal. That is, C D (K ) = K . Similarly, C D o (K ) = K . Corollary 4.35 therefore
shows that K ⊗ F D and End K (D o ) are isomorphic as F-algebras. However, from
the proof we see that the isomorphism is given by k ⊗ d → L k Rd , and so they are
actually isomorphic as K -algebras. Since D o is finite dimensional over K , we have
End K (D o ) ∼
= Mn (K ). 
Example 4.41 A basic example of a maximal subfield of H is R ⊕ Ri; but in view
of Theorem 4.39 this is hardly a new example of a splitting field for H. If we take,
for instance, the rational quaternions Q ⊕ Qi ⊕ Q j ⊕ Qk, then Theorems 4.39 and
4.40 do yield different splitting fields.

4.8 Simplicity of Tensor Products

The basic Wedderburn’s structure theorem (Theorem 2.61) can be interpreted as

follows: A finite dimensional algebra is simple if and only if it is isomorphic to the
tensor product of a matrix algebra Mn (F) and a division algebra D (cf. Example
4.22). This shows that, unlike the direct product, a more sophisticated tensor product
may give rise to simple algebras. Let us examine when this can occur.
Theorem 4.42 Let A be an arbitrary unital algebra, and let B be a finite dimensional
central simple algebra. Then every ideal of A ⊗ B is of the form I ⊗ B where I is
an ideal of A. Accordingly, A ⊗ B is simple if and only if A is simple.
Proof Let U be an ideal of A ⊗ B. It is immediate that I := {a ∈ A | a ⊗ 1 ∈ U } is
then an ideal of A, and that I ⊗ B ⊆ U . To prove the converse inclusion,  pick u ∈ U .
Let { f 1 , . . . , f n } be a basis of B. By Corollary 4.14 we have u = nj=1 a j ⊗ f j
for some a j ∈ A. Take ϕ ∈ End F (B) such that ϕ( f 1 ) = 1 and ϕ( f j ) = 0, j ≥ 2.
In view of Lemma 1.25, we can represent ϕ as ri=1 L bi Rci , bi , ci ∈ B. Hence we
have


r 
r 
n 
n
U (1 ⊗ bi )u(1 ⊗ ci ) = a j ⊗ bi f j ci = a j ⊗ ϕ( f j ) = a1 ⊗ 1.
i=1 i=1 j=1 j=1

Thus, a1 ∈ I . Similarly we see that each ai ∈ I , and so u ∈ I ⊗ B. Thus U = I ⊗ B.

4.8 Simplicity of Tensor Products 97

If A is simple, then I = 0 or I = A, and hence U = 0 or U = A ⊗ B. Thus

A ⊗ B is simple. Conversely, assume that A ⊗ B is simple, and choose a nonzero
ideal I of A. Then I ⊗ B is a nonzero ideal of A ⊗ B, so that I ⊗ B
= A ⊗ B. Given
a ∈ A, we can therefore find ai ∈ I and bi ∈ B such that a ⊗ 1 = i ai ⊗ bi . From
Lemma 4.11 we see that a is a linear combination of the ai ’s. Hence a ∈ I . Thus
I = A, proving the simplicity of A. 

The assumption that B is finite dimensional can actually be removed. The proof is
basically the same, one just has to apply Corollary 5.24 from the next chapter instead
of Lemma 1.25. For our current purposes, however, the finite dimensional case is
sufficient.

Corollary 4.43 Let A be an arbitrary unital algebra. Then every ideal of Mn (A) is
of the form Mn (I ), where I is an ideal of A. Accordingly, Mn (A) is simple if and
only if A is simple.

Proof Take Mn (F) for B in Theorem 4.42. 

It should be remarked that Corollary 4.43 can be also proved by entirely elementary
means, avoiding the tensor products.

Corollary 4.44 The tensor product of two finite dimensional central simple algebras
is a finite dimensional central simple algebra.

Proof Use Theorem 4.42 together with Corollary 4.33. 

The tensor product of two finite dimensional simple algebras may not be simple.
If, however, one of them is central, then it is. If both are central, then so is their
tensor product. These phenomena are nicely illustrated by the tensor products of real
division algebras:

Example 4.45 We have

C ⊗R C ∼
= C × C, C ⊗R H ∼
= M2 (C), H ⊗R H ∼
= M4 (R).

The last statement was noticed in Example 4.29. We know already from Sect. 2.11
that the scalar extension of H to C is M2 (C). Therefore C ⊗R H is isomorphic to
M2 (C) even when we consider them as C-algebras (and hence also when we consider
them as R-algebras). We may also regard C ⊗R C as a C-algebra (in light of Lemma
4.36). As such it is isomorphic to C × C since they are both unital 2-dimensional
C-algebras containing an element which is not a scalar multiple of unity and whose
square is −1 (i.e., 1 ⊗ i and (i, −i), respectively). Therefore C ⊗R C and C × C are
isomorphic also as R-algebras.
98 4 Tensor Products

4.9 The Skolem-Noether Theorem

Most parts of this and the preceding chapter were devoted to building machinery.
In these final sections we shall finally be rewarded. We are going to prove two
appealing theorems on finite dimensional central simple algebras by using several
abstract results on tensor products and modules.
In Sect. 1.6 we have proved that all automorphisms of finite dimensional central
simple algebras are inner. As we pointed out then, this result is an important special
case of the so-called Skolem-Noether Theorem. We are now in a position to establish
the general version of this theorem.

Theorem 4.46 (Skolem-Noether) Let A be a finite dimensional central simple alge-

bra. If S is a simple subalgebra of A that contains the unity 1 of A, then every homo-
morphism from S into A that maps 1 into 1 can be extended to an inner automorphism
of A.

Proof As we know, A is isomorphic to a matrix algebra over a central division algebra

(Corollary 2.62). However, for the purposes of this proof it is more convenient to
represent it as an endomorphism algebra. More precisely, according to Theorem
3.31 we have A ∼ = EndΔ (V ), where Δ is a central division algebra and V is a finite
dimensional vector space over Δ. In fact, there is no loss of generality in assuming
that A = EndΔ (V ) (note that the conclusion of the theorem holds simultaneously
for isomorphic algebras).
Let T = S ⊗ F Δ, where F is the base field. As S is simple and Δ is central simple,
T is a simple algebra by Theorem 4.42. Let us point out that V is also a vector space
over F and elements in A are F-linear maps (cf. Remark 3.34). Given v ∈ V , we
see that the map S × Δ → V , (σ, δ) → σ (δv) is bilinear, and hence there is an
F-linear map T → V , σ ⊗ δ → σ (δv). Since v is arbitrary, there exists a bilinear
map T × V → V , (t, v) → tv, such that

(σ ⊗ δ)v = σ (δv) for all σ ∈ S, δ ∈ Δ, v ∈ V . (4.3)

In this way V , considered as a space over F, becomes a unital T -module. Indeed,

since elements in S are Δ-linear maps, we have
      
(σ ⊗ δ)(σ  ⊗ δ  ) v = σ σ  (δδ  )v = σ σ  δ(δ  v)
    
= σ δ σ  (δ  v) = (σ ⊗ δ) (σ  ⊗ δ  )v .

To avoid confusion, we denote V , when viewed as a T -module with respect to (4.3),

by V1 .
Let ϕ : S → A be a homomorphism such that ϕ(1) = 1. By making only obvious
changes in the above discussion we see that V becomes a unital T -module with
respect to the action
4.9 The Skolem-Noether Theorem 99

(σ ⊗ δ) · v = ϕ(σ )(δv) for all σ ∈ S, δ ∈ Δ, v ∈ V . (4.4)

Let us denote V , when viewed as a T -module with respect to (4.4), by V2 .

We have arrived at the crucial point of the proof: since V1 and V2 have the same
dimension over F (indeed, as vector spaces over F they are simply identical) and T
is a simple algebra, Corollary 3.60 implies that they are isomorphic as T -modules.
Let α : V1 → V2 be an isomorphism. Given δ ∈ Δ and v ∈ V , we have

α(δv) = α((1 ⊗ δ)v) = 1 ⊗ δ · α(v) = ϕ(1)(δα(v)) = δα(v).

That is, α is a Δ-linear map. As vector spaces over F, both V1 and V2 are equal to
V . Therefore α ∈ A. Moreover, since α is bijective, it is invertible in A.
Let σ ∈ S. We can multiply α and σ as elements in A. For v ∈ V we have
 
(ασ )(v) = α(σ (v)) = α((σ ⊗ 1)v) = σ ⊗ 1 · α(v) = ϕ(σ )(α(v)) = ϕ(σ )α (v).

Since v is arbitrary, this means that ασ = ϕ(σ )α for every σ ∈ S. That is, ϕ(σ ) =
ασ α −1 , so ϕ can be extended to the inner automorphism a → αaα −1 of A. 

The Skolem-Noether Theorem turns out to be very useful. In particular, as one

can find in several textbooks (e.g., in [FD93], [Her68], and [Hun74]). Frobenius’
theorem on real division algebras and Wedderburn’s theorem on finite division rings
can be derived from it. Of course, these proofs are not as elementary as those given
in Chap. 1. But they are very elegant and therefore worth seeing.

4.10 The Double Centralizer Theorem

Let S be a subalgebra of an algebra A. The centralizer of its centralizer, C A (C A (S)),

is called the double centralizer of S. This is obviously a subalgebra of A which
contains S. Quite possibly S is its proper subset. Say, if S does not contain Z (A),
this is certainly the case. Another example: if A = Mn (F) and S = Tn (F), n ≥ 2,
then C A (S) = F = Z (A), and hence C A (C A (S)) = A  S. On the other hand, if
S = A or S = Z (A), then C A (C A (S)) = S. A more illuminating situation where
the equality holds is described in the following example.

Example 4.47 Considering central algebras S and T as subalgebras of A := S ⊗ T ,

we have C A (S) = T and C A (T ) = S by Corollary 4.33, and hence C A (C A (S)) = S
and C A (C A (T )) = T . If S and T are finite dimensional, then we also have [A : F] =
[S : F][C A (S) : F] = [T : F][C A (T ) : F].

This example illustrates the next theorem in which, however, we will not assume
that S is central.
100 4 Tensor Products

Theorem 4.48 (Double Centralizer Theorem) Let A be a finite dimensional central

simple F-algebra. If S is a simple subalgebra of A that contains the unity of A, then
C A (S) is a simple algebra,

[A : F] = [S : F][C A (S) : F], (4.5)

and C A (C A (S)) = S.

Proof The last claim follows from the first two. Namely, knowing that C A (S) is
simple, we can replace S by C A (S) in (4.5). Thus,

[A : F] = [C A (S) : F][C A (C A (S)) : F].

Comparing this formula with (4.5) we obtain [C A (C A (S)) : F] = [S : F]. Since

S ⊆ C A (C A (S)), the desired conclusion C A (C A (S)) = S follows.
We divide the proof of the first two claims into two steps. In the first one we
will consider the case where S is a division algebra, and in the second one we will
consider the general case.
(I) Assume that S = D is a division algebra. By Corollary 4.35 we have C A (D) ⊗
Ao ∼= End D (A). Theorem 3.31 shows that End D (A) ∼ = Ms (D o ), where s = dim D A.
Therefore End D (A) is a simple algebra. Using the easier “only if” part of Theorem
4.42 it follows that C A (D) is simple. It remains to prove the second claim. On the
one hand, we have

[End D (A) : F] = s 2 [D : F],

and on the other hand,

[End D (A) : F] = [C A (D) : F][A : F].

We claim that
[A : F]
s= .
[D : F]

If D was a field, this would follow from Remark 1.33. Fortunately, the proof remains
unchanged if D is noncommutative. Combining the above three formulas we obtain
(4.5) (for S = D).
(II) Assume now that S is an arbitrary simple subalgebra of A. As we know from
Wedderburn’s theorem (Theorem 2.61), there exist r ∈ N and a division algebra D
such that S ∼= Mr (D). This implies that S contains a set of r × r matrix units ei j .
These are also the matrix units of A for A has the same unity as S. Lemma 2.52
therefore implies that A ∼= Mr (B), where B = e11 Ae11 ⊇ e11 Se11 = D. Since B is
a central algebra (Lemma 1.15), we see from Corollary 4.34 that

C A (S) ∼
= C B (D).
4.10 The Double Centralizer Theorem 101

As B is also a simple algebra (Corollary 4.43), we can apply (I) with B playing the
role of A. Thus, C B (D) is a simple algebra and [B : F] = [D : F][C B (D) : F].
The former establishes the simplicity of C A (S), and the latter implies

[A : F] = r 2 [B : F] = r 2 [D : F][C B (D) : F].

Since r 2 [D : F] = [S : F] and [C B (D) : F] = [C A (S) : F], this proves (4.5). 

Adding the assumption that S is central, we arrive at the situation considered in

Example 4.47.

Corollary 4.49 Let A be a finite dimensional central simple F-algebra. If S is a

central simple subalgebra of A that contains the unity of A, then A ∼
= S ⊗ C A (S).

Proof Consider a linear map ϕ : S⊗C A (S) → A, ϕ(s⊗t) = st. Since elements from
S commute with elements from C A (S), ϕ is multiplicative. The algebra S ⊗ C A (S)
is simple (Theorem 4.42), so that ker ϕ = 0. From [A : F] = [S : F][C A (S) :
F] we see that S ⊗ C A (S) and A are of the same dimension. Hence ϕ is an
isomorphism. 

Corollary 4.49 deals with the case where the center of S is minimal. Let us take
a look at the opposite extreme. Assume that S = K is a maximal subfield of a finite
dimensional central division algebra D. As observed in the proof of Theorem 4.40,
C D (K ) = K in this case. Corollary 4.49 does not hold for S = K ; after all, K ⊗ K
is a commutative algebra. The only nontrivial information that we extract from the
Double Centralizer Theorem is [D : F] = [K : F]2 , which is nothing but the content
of Theorem 1.36.

4.11 The Brauer Group

The basic idea behind the concept of the Brauer group is to use the tensor products
to classify finite dimensional central division algebras over a given field. The tensor
product of two such algebras is not always a division algebra. For instance, H⊗R H ∼ =
M4 (R) (Example 4.29). However, it always is a central simple algebra (Corollary
4.44), and such an algebra conceals within itself a central division algebra. That is,
it is isomorphic, for some n ∈ N, to Mn (D), where D is a central division algebra
(Corollary 2.62). Moreover, D is essentially unique (Corollary 3.58). Thus, ignoring
the matrix part of the resulting central simple algebra, we can say that the tensor
product of two finite dimensional central division algebras gives rise to an algebra
from the same class. Say, the product of H with itself brings out R. (Maybe not a
very exciting result, but what to expect in view of Frobenius’ theorem?)
Let us now make everything precise. Fix a field F. Let A and A be finite
dimensional central simple F-algebras. Then, as we know, there exist n, n  ∈ N
and central division algebras D, D  such that A ∼ = Mn (D) and A ∼ = Mn  (D  ). We
  ∼ 
shall say that A and A are equivalent, and write A ∼ A , if D = D . This definition
102 4 Tensor Products

is unambiguous since D and D  are uniquely determined up to isomorphism. Note

that ∼ is indeed an equivalence relation in the class of all finite dimensional central
simple F-algebras. The equivalence class of A will be denoted by [A]. Thus, if D is a
finite dimensional central division algebra, then [D] = [M2 (D)] = [M3 (D)] = . . . .
As indicated in the previous paragraph, among all representatives of this class we are
really interested only in D (and division algebras isomorphic to D). The problem is
that D may be hidden, so we are forced to treat all representatives equally.

Proposition 4.50 The set of all equivalence classes of finite dimensional central
simple F-algebras, endowed with multiplication

[A] · [B] = [A ⊗ B],

is an abelian group.

Proof We first have to show that this multiplication is well-defined. That is, we have
to prove that for all finite dimensional central simple algebras A, A , B, B  ,

A ∼ A and B ∼ B  =⇒ A ⊗ B ∼ A ⊗ B  .

By definition, A ∼ A means A ∼ = Mn (D) and A ∼= Mn  (D  ) with D ∼= D  . But

then we actually have A ∼
= nM  (D). Similarly, B ∼ B  implies that B ∼ M (C)
= m
and B  ∼
= Mm  (C) for some m, m  ∈ N and a central division algebra C. Since
D ⊗ C is again a finite dimensional central simple algebra by Corollary 4.44, we
have D ⊗ C ∼ = Mk (E) with k ∈ N and E a central division algebra. Now, using
the associativity and commutativity of tensor products, together with basic facts
about matrix algebras represented as tensor products (explained in Example 4.22),
we obtain
   
A⊗B ∼
= Mn (F) ⊗ D ⊗ Mm (F) ⊗ C
   
∼
= Mn (F) ⊗ Mm (F) ⊗ D ⊗ C
 
∼
= Mnm (F) ⊗ Mk (F) ⊗ E
∼
= Mnmk (F) ⊗ E
∼
= Mnmk (E).

Similarly we see that A ⊗ B  ∼

= Mn  m  k (E). Accordingly, A ⊗ B ∼ A ⊗ B  .
The multiplication is associative and commutative since the tensor product is
associative and commutative. Obviously, [F] is the identity element. Finally, from
A ⊗ Ao ∼ = Mn (F) (Corollary 4.28) we infer that [A]−1 = [Ao ]. 

Some of the readers may have noticed that there is a missing point in this proof.
One would first have to verify that the equivalence classes of finite dimensional
central simple F-algebras indeed form a set. This is not a difficult task, but we shall
not go into this matter.
4.11 The Brauer Group 103

Definition 4.51 The group from Proposition 4.50 is called the Brauer group of the
field F. It will be denoted by Br(F).

The name was given after R. Brauer who made fundamental contributions to this
field of research.
Computing Br(F) for a given field F is not an easy task. But at least we know
the following:
• If F is algebraically closed, then Br(F) = {1}.
• If F is finite, then Br(F) = {1}.
• Br(R) ∼ = Z2 .
These are restatements of Proposition 1.6, Theorem 1.38, and Corollary 1.19 with
Example 4.29.
Brauer groups turn out to be important in different mathematical areas. Anyhow,
we close our discussion on this topic at this point, and refer to [FD93] and [Pie82] for
possible further reading. One might protest that we have just warmed up, and besides
the definition and fancy interpretations of some fundamental results from Chap. 1 we
have told nothing. But perhaps the reader will agree that the definition of the Brauer
group is captivating in its own right. This short section is only an invitation into a
rich area.

Exercises

4.1. Let U and V be vector spaces of dimension at least 2. Prove that not every
element in U ⊗ V is a simple tensor. n
4.2. Let U be a real vector space. Show that the identity i=1 u i ⊗ u i = 0 in U ⊗U
implies that each u i = 0.
4.3. Let U be a vector space over a field F with char(F) = 2. Denote by S the
subspace of U ⊗ U generated by all elements of the form u ⊗ u, u ∈ U ,
and by T the subspace generated by all elements of the form u ⊗ v − v ⊗ u,
u, v ∈ U . Show that U ⊗ U = S ⊕ T . What is the dimension of S and T if
dim F U = n < ∞?
Hint: Make use of the map u ⊗ v → v ⊗ u.
4.4. Show that (U1 ⊕ U2 ) ⊗ V ∼ = (U1 ⊗ V ) ⊕ (U2 ⊗ V ) holds for vector spaces.
Extend this to algebras.
4.5. Use the tensor product to construct a 9-dimensional algebra whose radical is
5-dimensional.
4.6. Let S = (si j ) ∈ Mm (F) and T = (ti j ) ∈ Mn (F). The mn × mn matrix
104 4 Tensor Products
⎡ ⎤ ⎡ ⎤
s11 T s12 T . . . s1m T s11 t11 s11 t12 . . . s1m t1n
⎢ s21 T s22 T . . . s2m T ⎥ ⎢ s11 t21 s11 t22 . . . s1m t2n ⎥
⎢ ⎥ ⎢ ⎥
S ⊗ T := ⎢ . .. . . . ⎥=⎢ . .. .. .. ⎥
⎣ .. . . .. ⎦ ⎣ .. . . . ⎦
sm1 T sm2 T . . . smm T sm1 tn1 sm1 tn2 . . . smm tnn

is called the Kronecker or tensor product of matrices S and T . This definition

is clearly in accordance with the convention from Example 4.22. Note that
it is also compatible with the tensor product of linear maps in the following
sense: If S is the matrix corresponding to a linear map f with respect to a basis
{e1 , . . . , em }, and T is the matrix corresponding to a linear map g with respect
to a basis { f 1 , . . . , f n }, then S ⊗ T is the matrix corresponding to f ⊗ g with
respect to the basis {e1 ⊗ f 1 , . . . , e1 ⊗ f n , e2 ⊗ f 1 , . . . , em ⊗ f n }. Show that
tr(S ⊗ T ) = tr(S)tr(T ) and det(S ⊗ T ) = det(S)n det(T )m .
Hint: S ⊗ T = (S ⊗ In )(Im ⊗ T ) where Ik stands for the k × k identity matrix.
4.7. Show that algebras A and B must be unital in case A ⊗ B is unital.
4.8. Show that p ∈ N is prime if and only if there do not exist algebras A = F and
B = F such that A ⊗ B ∼ = M p (F).
4.9. Let A, A , B, B  be finite dimensional simple algebras. Suppose there exist
m, n ∈ N such that Mm (A) ∼ = Mn (B) and Mn (A ) ∼= Mm (B  ). Show that
 ∼
A⊗ A = B⊗B . 

4.10. Let A be a central simple algebra over a field F with char(F) = 2, 3. Show
that a ∈ A satisfies [a, [a, [a, x]]] = 0 for all x ∈ A if and only if there exists
λ ∈ F such that (a − λ)2 = 0. Analyze, in a similar fashion, the condition
[a, [a, [a, x]]] = [a, x].
Hint: Interpret the conditions [a, [a, [a, x]]] = 0 and [a, [a, [a, x]]] = [a, x]
in the multiplication algebra M(A).
4.11. Let A be a 4-dimensional central simple F-algebra. Show that C A ({a}) =
F + Fa for every a ∈ A \ F.
Hint: Apply scalar extension to reduce the problem to the case where A =
M2 (F) and F is algebraically closed (so that the Jordan normal form of a is
applicable).
4.12. Let A be a finite dimensional central simple algebra. Show that the linear span
of all commutators in A has codimension 1 in A.
4.13. Let A be an F-algebra and let K be an extension field of F. True or False:
(a) If A is prime, then A K is prime.
(b) If A K is prime, then A is prime.
4.14. Let A be an R-algebra. Show that the set A × A equipped with operations
 Exercises 105

(x, y) + (z, w) = (x + z, y + w),

(α + iβ)(x, y) = (αx − βy, βx + αy),
(x, y)(z, w) = (x z − yw, xw + yz),

where x, y, z, w ∈ A and α, β ∈ R, is a C-algebra isomorphic to AC .

Remark: This construction provides a shortcut to the notion of the scalar exten-
sion for mathematicians working in areas in which R and C are the only fields
that matter.
4.15. Let D1 and D2 be finite dimensional central division F-algebras. As we know,
D1 ⊗ D2 is isomorphic to Mm (D) for some m ∈ N and a central division
algebra D. Examine Dio ⊗ Mm (D) to show that m divides [Di : F], i = 1, 2.
Hence conclude that D1 ⊗ D2 is a division algebra in case [D1 : F] and
[D2 : F] are relatively prime.
Remark: The latter fact is usually used in the proof of the following important
and beautiful result: If D is a finite dimensional central division F-algebra,
and [D : F] = p1k1 . . . prkr where pi are different primes, then there exist
central division algebras D1 , . . . , Dr such that D ∼= D1 ⊗ · · · ⊗ Dr and
[Di : F] = piki .
4.16. Let char(F) = p > 0. Suppose a unital F-algebra A contains an element
a such that a ∈
/ F and a p ∈ F. Show that then A ⊗ A contains a nonzero
nilpotent element.
Remark: Using this one can show that even the tensor product of fields is not
always a semiprime algebra. On the other hand, if char(F) = 0 and K , L are
field extensions of F such that at least one of them is finite, then K ⊗ F L is a
direct product of fields (and is hence semiprime). The proof uses only standard
field extension tools, and readers can try to produce it themselves.
4.17. Let A and B be simple unital F-algebras. Show that the algebra A ⊗ B is
semiprime if (and only if) its center is semiprime.
n
Hint: Take a nilpotent ideal I of A⊗ B, and consider an element i=1 ai ⊗bi =
0 in I such that n is minimal.
4.18. Show that besides 0, R, and H, the only subalgebras of H are those of the form
h(R ⊕ Ri)h −1 , h ∈ H \ {0}.
4.19. Let A be a finite dimensional central simple algebra over a field F with the
property that λ2 = −1 for every λ ∈ F. Show that if u, v ∈ A satisfy u 2 =
v2 = −1, then there exists an invertible a ∈ A such that u = ava −1 .
Hint: F + Fu and F + Fv are isomorphic subfields of A.
4.20. A linear map δ from an algebra A into itself is called a generalized derivation
if δ(x yz) = δ(x y)z − xδ(y)z + xδ(yz) for all x, y, z ∈ A. Show that every
generalized derivation of a finite dimensional central simple algebra A is the
sum of a left multiplication map and a right multiplication map.
106 4 Tensor Products

Hint: Observe that

   
x 0 x δ(x) − δ(1)x
→
0x 0 x

defines an algebra homomorphism from an isomorphic copy of A in M2 (A)

into M2 (A).
4.21. Let S be a subalgebra of an algebra A. Show that:
(a) If S is commutative, then so is C A (C A (S)), and C A (C A (S)) ⊆ C A (S).
(b) If S is equal to the centralizer of some subset of A, then S = C A (C A (S)).
4.22. Let A and S be as in the Double Centralizer Theorem. Note that the center Z
of S is a field, and that S, C A (S), and C A (Z ) are algebras over Z . Show that
S ⊗ Z C A (S) ∼= C A (Z ).
4.23. Let A be a finite dimensional central simple R-algebra. Show that A contains a
subfield K such that K = C A (K ) if and only if A is isomorphic to R, M2 (R),
or H.
4.24. Let A and A be finite dimensional central simple F-algebras. Show that:
(a) A ∼ A if and only if there exist m, m  ∈ N such that Mm (A) ∼
= Mm  (A ).
(b) A ∼  
= A if and only if A ∼ A and [A : F] = [A : F].
Chapter 5
Structure of Rings

Although most of the concepts introduced in the first four chapters involve
general rings and algebras, the majority of the main results concern algebras of finite
dimension. In this chapter we will make a clean break from the finite dimensional
context.
The goal of the chapter is to extend Wedderburn’s structure theory of finite
dimensional algebras, discussed in Chap. 2, to general rings. As one would expect,
this cannot be done in a straightforward manner and the main results are considerably
more complex.
The topics will be considered in a different order than in Chap. 2. In the first part
we will treat primitive rings, which play a similar role in general rings as simple
algebras do in the context of finite dimensional algebras. The second part will be
primarily devoted to the Jacobson radical, a generalization of the radical of a finite
dimensional algebra.
Most of the theory that we are about to expose was developed around 1945 by
N. Jacobson, a student of Wedderburn.

5.1 Primitive Rings

Wedderburn’s basic structure theorem says that a finite dimensional prime algebra
can be represented as an algebra of all n × n matrices over a division algebra D, or,
equivalently, as an algebra of all linear operators on an n-dimensional vector space
over a division algebra Δ = D o (cf. Remark 3.34). We are going to introduce a class
of prime rings for which a more general structure theorem holds.
The fact that minimal left ideals give rise to division rings (Lemma 2.58) was
one of the main ingredients in our proof of Wedderburn’s theorem. Relying on min-
imal left ideals in more general rings would have a limited reach since their exis-
tence is questionable. But there is a perfect substitute for them, namely, the simple
modules—they exist under rather mild conditions (cf. Lemma 3.46), and they also

© Springer International Publishing Switzerland 2014 107

M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_5
108 5 Structure of Rings

yield division rings by Schur’s Lemma. At this point we advise the reader to glance
through Sects. 3.5–3.7 in order to better understand the discussion that follows.
Let R be a ring and let V be a simple (left) R-module. Schur’s Lemma (Lemma
3.50) tells us that

Δ := End R (V )

is a division ring. Let δ ∈ Δ and v ∈ V . We simplify the notation by setting

δv := δ(v) ∈ V. (5.1)

Obviously, we have

(δ1 + δ2 )v = δ1 v + δ2 v, δ(v1 + v2 ) = δv1 + δv2 ,

(δ1 δ2 )v = δ1 (δ2 v), 1v = v

for all δ1 , δ2 , δ ∈ Δ and v1 , v2 , v ∈ V . This means that V is a vector space over

Δ; more precisely, the additive group of V becomes a vector space over Δ relative
to the operation Δ × V → V defined in (5.1) (that is why we have denoted our
simple module by V instead of, say, M). From now on we shall therefore consider
V simultaneously as a module over R and a vector space over Δ. Since δ ∈ Δ is an
endomorphism of the R-module V , the following fundamental formula

δ(r v) = r (δv) for all δ ∈ Δ, r ∈ R, v ∈ V, (5.2)

connects both structures. Now, δ and r appear symmetrically in (5.2). Accordingly,

why not interpret (5.2) as that r is an endomorphism of the vector space V over Δ?
We will formalize this idea in the next section. However, an extra condition is needed
for its full realization. If r lies in the annihilator of V , then r gives rise to the zero
endomorphism of V over Δ. We will therefore require that V is faithful.

Definition 5.1 A ring R is said to be primitive if it has a faithful simple module V .

Recall from Definition 3.11 and Lemma 3.36 that this means that V = 0 and the
following two conditions are fulfilled:
• r V = 0 for every 0 = r ∈ R.
• Rv = V for every 0 = v ∈ V .
Since our modules are, by convention, left modules, a more adequate name for
such a ring is a left primitive ring. Analogously one defines a right primitive ring,
and it turns out that these two classes of rings do not coincide. It is good to keep this
in mind, but we will omit writing “left” before “primitive ring”.
Remark 5.2 Let R be a primitive ring. If a ring S is isomorphic to R, then S is
primitive too. Indeed, if V is a faithful simple R-module, then the additive group
5.1 Primitive Rings 109

of V becomes a faithful simple S-module by defining sv := ϕ(s)v, where ϕ is an

isomorphism from S onto R.
The discussion before Definition 5.1 indicates a simple example, in fact the
prototype, of a primitive ring:

Example 5.3 Let V be a nonzero vector space over a division ring Δ. Then thering 
R := EndΔ (V ) is primitive. Indeed, V is its faithful simple module (via f v := f v ).
The condition that Rv = V if v = 0 follows from Remark 3.29 and the fact that v
is contained in a basis (cf. Theorem 3.28). Let us stress that here V is of arbitrary
dimension, finite or infinite.

More concrete examples of primitive rings will be met later. Let us now compare
the class of primitive rings to other classes of rings introduced earlier.
Lemma 5.4 A primitive ring R is prime.

Proof Let a, b ∈ R be nonzero elements. Take a faithful simple R-module V . By

faithfulness, there are u, v ∈ V such that au = 0 and bv = 0, and by simplicity there
is r ∈ R such that r (bv) = u. Accordingly, (ar b)v = 0, and so a Rb = 0. 

Lemma 5.5 A prime ring R with a minimal left ideal is primitive.

Proof Let L be a minimal left ideal of R. It is clear that L is simple as an R-module.

The primeness of R implies that it is also faithful. Indeed, r L = 0, where r ∈ R,
implies r R L ⊆ r L = 0, and hence r = 0. 

Lemma 5.6 A simple unital ring R is primitive.

Proof Since R is unital, it has simple modules by Corollary 3.49. The annihilator of
every module is an ideal, and so simple modules over simple rings are automatically
faithful. 

A non-unital simple ring is not always primitive. Such rings may have surpris-
ing properties. For example, in 2002 A. Smoktunowicz constructed a simple nil
ring. Primitive rings, however, cannot be nil. In fact, if R is a nil ring and M is an
R-module, then we clearly have r m = m for every r ∈ R and 0 = m ∈ M. Thus, in
view of Lemma 3.36, nil rings do not even have simple modules.

Lemma 5.7 A commutative primitive ring R is a field.

Proof Let V be a faithful simple R-module, and pick 0 = v ∈ V . We claim that

for every r ∈ R, r v = 0 forces r = 0. Indeed, Rv = V by simplicity of V , hence
r V = r Rv = Rr v = 0, and therefore r = 0 by the faithfulness of V . Now take
0 = a ∈ R. Then av = 0, and therefore, by the simplicity of V , there exists b ∈ R
such that b(av) = v. Consequently, (xba − x)v = 0 for every x ∈ R. Therefore
bax = xba = x, and so ba = 1 (in particular, R is unital). As R is commutative,
this means that b = a −1 . 
110 5 Structure of Rings

Example 5.8 The simplest example of a prime ring that is not primitive is Z.

Example 5.9 The Weyl algebra A1 is a simple unital ring (see Example 1.13), and
hence a primitive ring.

Example 5.10 If V is an infinite dimensional vector space over a division ring Δ,

then the ring EndΔ (V ) is primitive, but not simple. An example of its proper nonzero
ideal is the set of all finite rank operators. These are operators whose rank, i.e.,
the dimension of their image, is finite. It is easy to check that finite rank operators
indeed form an ideal. It is proper since it does not contain the identity operator, and
it certainly is nonzero; in fact, from Remark 3.29 we deduce that for every subspace
W of V there is a linear operator having W as its image.

The ring EndΔ (V ) is also the simplest example of a primitive ring that has minimal
left ideals. An example of a primitive ring without minimal left ideals is the Weyl
algebra A1 . All this is easy to see, but we postpone the explanation until Sect. 5.4.
Let us close this section by mentioning that every F-algebra can be embed-
ded into a primitive ring, namely, into End F (V ) for some vector space V over F
(Proposition 2.38). This is not always true for rings. For example, take a ring that
contains nonzero elements u and v satisfying 2u = 3v = 0 (e.g., Z2 × Z3 ). In light of
Remark 2.20 we see that this ring cannot be embedded into a prime ring, and hence
neither into a primitive ring by Lemma 5.4.

5.2 The Jacobson Density Theorem

The following notion is often one’s first thought when encountering a primitive ring.

Definition 5.11 Let V be a vector space over a division ring Δ, and let R be a
subring of the endomorphism ring EndΔ (V ). We say that R is a dense ring of linear
operators of V if for every n ∈ N, every linearly independent subset {u 1 , u 2 , . . . , u n }
of V , and every (not necessarily linearly independent) subset {v1 , v2 , . . . , vn } of V ,
there exists f ∈ R such that

f (u 1 ) = v1 , f (u 2 ) = v2 , . . . , f (u n ) = vn .

The term “dense” comes from a topological characterization. We will not need it,
so let us not go into detail about this. Still, having in mind the meaning of density
in topology, we may have got the message; a dense ring of linear operators of V
can be considered as a rather large subring of EndΔ (V ), or better, large enough for
certain purposes.

Lemma 5.12 A ring R is a dense ring of linear operators of the vector space V over
Δ if and only if for every g ∈ EndΔ (V ) and every finite dimensional subspace U of
V there exists f ∈ R such that f |U = g|U .
5.2 The Jacobson Density Theorem 111

Proof Take g ∈ EndΔ (V ) and a finite dimensional subspace U of V . Pick a basis

{u 1 , u 2 , . . . , u n } of U . If R is a dense ring, then there exists f ∈ R such that
f (u i ) = g(u i ), i = 1, . . . , n. Consequently, f |U = g|U .
Proving the converse is just as easy, only standard facts about endomorphisms
and bases have to be used (cf. Remark 3.29). 
Corollary 5.13 If V is a finite dimensional vector space over Δ, then EndΔ (V ) is
the only dense ring of linear operators of V .
If V is infinite dimensional, then we have a variety of dense rings of operators on
V . The next lemma gives a partial evidence for this.
Lemma 5.14 If R is a dense ring of linear operators of a vector space V , then so
are its nonzero ideals.
Proof Let I be a nonzero ideal of R. Choose a nonzero h ∈ I , and let u ∈ V be
such that v := h(u) = 0. Now take linearly independent vectors u 1 , . . . , u n ∈ V and
arbitrary vectors v1 , . . . , vn ∈ V . Since R is dense, there exist g1 , . . . , gn ∈ R such
that gi (u i ) = u and gi (u j ) = 0 if j = i. Further, there are f 1 , . . . , f n ∈ R such that
f i (v) = vi . Consequently,

k := f 1 hg1 + · · · + f n hgn ∈ I

satisfies k(u i ) = vi , i = 1, . . . , n. 
Example 5.15 The ring of all finite rank operators on an infinite dimensional vector
space V over Δ is thus a proper dense subring of EndΔ (V ) (cf. Example 5.10).
The next theorem is one of the cornerstones of noncommutative algebra.
Theorem 5.16 (Jacobson Density Theorem) A ring R is primitive if and only if it is
isomorphic to a dense ring of linear operators of a vector space over a division ring.
Proof Let us first resolve the easy part. A dense ring of linear operators of a vector

space V is primitive, since V is its faithful simple module (via f v := f v ). In
fact, one only needs the n = 1 case of Definition 5.11 to establish this. In view of
Remark 5.2, the “if” part is thus proved.
We proceed to the nontrivial part. Let R be a primitive ring and let V be its
faithful simple module. Then Δ = End R (V ) is a division ring by Schur’s Lemma.
The additive group of V becomes  a vector space over Δ with respect
  to the
 operation

defined by (5.1) (i.e., δv := δ v ), and the formula (5.2) (i.e., δ r v = r δv ) holds.
All this is easy to see and was explained in greater detail in the previous section.
Now, for each r ∈ R we define r : V → V according to

r (v) := r v. (5.3)

It is clear that r is an additive map. Moreover, by (5.2) we have r ∈ EndΔ (V ). Note

that
112 5 Structure of Rings

R → EndΔ (V ), r → r ,

is a ring homomorphism; the faithfulness of V implies that it is injective. Thus, R is

isomorphic to

R := {r | r ∈ R} ⊆ EndΔ (V ).

Our goal is to prove that R is dense. The desired condition from Definition 5.11
certainly holds for n = 1, since Rv = V for every v = 0 in V .
Let us pause for a moment. So far we have shown that up to isomorphism primitive
rings can be characterized as rings satisfying the n = 1 case of Definition 5.11. This
was obtained almost for free; more precisely, the arguments were straightforward,
but, of course, an ingenious recasting of the definition of primitivity was behind
them. The second part of the proof is more tricky.
We claim that it is enough to prove the following:
(∗) Let U be a finite dimensional subspace of V . For every w ∈ V \ U there exists
r ∈ R such that r (U ) = 0 and r (w) = 0.
Indeed, assume (∗) and take linearly independent vectors u 1 , . . . , u n ∈ V and arbi-
trary vectors v1 , . . . , vn ∈ V . Given 1 ≤ i ≤ n, use (∗) with the linear span of
u 1 , . . . , u i−1 , u i+1 , . . . , u n playing the role of U and u i playing the role of w. Thus
there exists ri ∈ R such that ri (u j ) = 0 if j = i and ri (u i ) = 0. As we know, we
can find si ∈ R such that si (ri (u i )) = vi . Note that

r := s1r1 + s2 r2 + · · · + sn rn ∈ R

satisfies r (u i ) = vi , i = 1, . . . , n.
Suppose (∗) is not true. Without loss of generality we may assume that U is a
space of the smallest dimension for which (∗) does not hold. Note that dimΔ U ≥ 1,
since (∗) is obviously true for the zero space. Let U0 be an arbitrary subspace of U
such that dimΔ U0 = dimΔ U − 1. We set

L := { ∈ R | (U0 ) = 0}.

Take any z ∈ V \ U0 . Since (∗) does hold for U0 by our assumption, there exists
 ∈ L such that (z) = 0. In view of (5.3) this means that Lz = 0. As L is clearly a
left ideal of R, Lz is a submodule of the R-module V . The simplicity of V therefore
yields

Lz = V for all z ∈ V \U0 . (5.4)

Let w ∈ V \U be a vector for which (∗) does not hold, and choose u ∈ U \U0 .
We claim that the map δ : V → V given by

δ(u) := w for all  ∈ L (5.5)

5.2 The Jacobson Density Theorem 113

is well-defined. First of all, since Lu = V by (5.4), δ is defined on the whole space

V . Suppose that  ∈ L is such that u = 0, i.e., (u) = 0. Then  vanishes on both
U0 and Δu, so that (U ) = 0. Our assumption on w implies that (w) = 0, i.e.,
w = 0. Therefore δ is indeed well-defined. One easily checks that δ is an R-module
endomorphism, i.e., δ ∈ Δ. According to (5.2) we can therefore rewrite (5.5) as

(δu) = w for all  ∈ L,

that is,

L(δu − w) = 0.

From (5.4) it follows that δu − w ∈ U0 . However, this yields

w = δu − (δu − w) ∈ Δu + U0 = U,

contradicting the choice of w. This proves (∗). 

5.3 Alternative Versions and Applications

We have stated the Jacobson Density Theorem in a condensed manner. For appli-
cations it is sometimes useful to know that the vector space from the theorem has
a tight connection to a faithful simple R-module V . Actually, the faithfulness of V
was needed only to assure the injectivity of the canonical homomorphism r → r . If
V is not faithful, then this homomorphism has a nontrivial kernel, namely ann R (V ).
In this case R/ann R (V ) is isomorphic to R. We thus have the following expanded
version of the Jacobson Density Theorem:

Theorem 5.17 Let R be a ring and let V be a simple R-module. Consider V as

a vector space over the division ring Δ = End R (V ). Then the ring R/ann R (V ) is
isomorphic to a dense ring of linear operators of V .

Our next objective is the Jacobson Density Theorem for algebras. Let A be an
algebra over a field F. We say that A is a primitive algebra if it has a faithful simple
A-module V . Here, of course, by a module we mean an algebra module. Therefore
Δ := End A (V ) is a division F-algebra, not merely a ring. Likewise EndΔ (V ) is an
F-algebra, and r → r is an algebra homomorphism. Accordingly, Theorem 5.16
gets the following form.

Theorem 5.18 An F-algebra A is primitive if and only if it is isomorphic to a dense

algebra of linear operators of a vector space over a division F-algebra.

We are now in a position to give a different proof of the fundamental Wedderburn’s

structure theorem (Theorem 2.61).
114 5 Structure of Rings

Corollary 5.19 Let A be a nonzero finite dimensional algebra. The following state-
ments are equivalent:
(i) A is prime.
(ii) A is simple.
(iii) There exist a finite dimensional division algebra Δ and a finite dimensional
vector space V over Δ such that A ∼ = EndΔ (V ) (and hence A ∼ = Mn (D),
where D = Δ and n = dimΔ V ).
o

(iv) A is primitive.

Proof We know that (iii) implies (ii) (Example 1.10), and trivially (ii) implies (i).
Since A is finite dimensional it has a minimal left ideal L. Lemma 5.5, which was
stated for rings but obviously it also holds for algebras, therefore tells us that (i)
implies (iv). It remains to show that (iv) implies (iii).
Let A be primitive. From Theorem 5.18 we know that A is isomorphic to a dense
algebra A of linear operators of a vector space V over a division algebra Δ. Let
u 1 , . . . , u m ∈ V be linearly independent over Δ. Then there exist f 1 , . . . , f m ∈ A
such that f i (u i ) = 0 and f i (u j ) = 0 if j = i. Note that f 1 , . . . , f m are linearly
independent over the base field F. Therefore m ≤ [A : F]. Since A is finite dimen-
sional over F, V must be finite dimensional over Δ. Corollary 5.13 implies that
A = EndΔ (V ) (and hence A ∼ = Mn (D), D = Δo , by Theorem 3.31). Obviously, Δ
is also finite dimensional over F. 

We can now somewhat better understand the meaning and importance of the
Jacobson Density Theorem. It yields a genuine generalization of Wedderburn’s clas-
sical theory. This will become even more apparent later, especially when exploring
the connection between primitive rings and the Jacobson radical.
Looking back at our proof of Theorem 2.61, it may now seem a bit raw and naive.
Why struggle with matrix manipulations when we can obtain the same and more
by refined linear operator methods? Note, however, that the arguments in the last
proofs are much more complex than those in Chap. 2; juggling with several tools and
concepts in one proof may not be so easy for a beginner.
Let us now change the perspective. Suppose we are given a nonzero vector space
V over a division F-algebra Δ and an F-subalgebra A of EndΔ (V ). When is A
dense? We can view V as a faithful A-module (via f v := f (v)), which is simple if
Av = { f (v) | f ∈ A} equals V for every nonzero v ∈ V . Let us therefore assume the
latter. One should not, however, hastily conclude that this already means that A is
dense. We also have to assume that End A (V ) = Δ (in this context we identify δ ∈ Δ
with the map V → V , v → δv). Obviously, Δ ⊆ End A (V ). The converse inclusion
does not always hold.

Example 5.20 Consider C as a 2-dimensional space over R (= Δ = F). Let A be

the subalgebra of EndR (C) consisting of all (left) multiplication maps L z , z ∈ C.
Obviously, Av = C for every 0 = v ∈ C. It is easy to see that End A (C) ∼
= C, so it
properly contains R. Moreover, from Corollary 5.13 we see that A is not dense.
5.3 Alternative Versions and Applications 115

The Jacobson Density Theorem for algebras of linear operators therefore reads
as follows.

Theorem 5.21 Let Δ be a division algebra over a field F, let V be a vector space
over Δ, and let A be an F-subalgebra of EndΔ (V ). If Av = V for every nonzero
v ∈ V and End A (V ) = Δ, then A is a dense algebra of linear operators of V .

Remark 5.22 The condition “Av = V for every nonzero v ∈ V ” (i.e., the condition
that V is simple as an A-module) is equivalent to the condition that there does not
exist a proper nonzero subspace W of V that is invariant under every f ∈ A. Indeed,
just observe that Av is a subspace invariant under every f ∈ A, and the statement
follows.

The next two corollaries are concerned with the simplest situation where Δ = F.
The first one is a classical result which was, in a slightly different form connected to
the representation of groups, obtained by W. Burnside in 1905.

Corollary 5.23 (Burnside) Let V be a finite dimensional vector space over an alge-
braically closed field F. If A is a subalgebra of End F (V ) such that Av = V for
every nonzero v ∈ V , then A = End F (V ).

Proof All we have to show is that End A (V ) = F, and this is easy: End A (V ) is a finite
dimensional division algebra over F, so it must be equal to F by Proposition 1.6. 

If a central simple F-algebra A is finite dimensional, then its multiplication algebra

M(A) coincides with End F (A) (Lemma 1.25). In the next corollary, first proved by
E. Artin and G. Whaples a couple of years before Jacobson’s discovery of the density
theorem, we extend this extremely useful result to infinite dimensional algebras.

Corollary 5.24 (Artin-Whaples) Let A be a central simple algebra. Then its multi-
plication algebra M(A) is a dense algebra of linear operators of A.

Proof Take 0 = v ∈ A. Since A is simple, the ideal of A generated by v is equal

to A. Thus, for every w ∈ A there exist ai , bi ∈ A such that i ai vbi = w. The
condition M(A)v = A is thus fulfilled.
It remains to prove that End M(A) (A) = F. That is, we have to show that each
δ ∈ End M(A) (A) is a scalar multiple of id A . To this end we only have to use

δ(L a (1)) = L a (δ(1)) and δ(Ra (1)) = Ra (δ(1)).

Indeed, these two formulas can be rewritten as δ(a) = aδ(1) = δ(1)a for every
a ∈ A, and therefore δ(1) ∈ F since A is central. 

As already remarked after the proof of Theorem 4.42, Corollary 5.24 makes it
possible for us to get rid of the assumption that the algebra B from that theorem is
finite dimensional.
116 5 Structure of Rings

5.4 Primitive Rings Having Minimal Left Ideals

If the Jacobson Density Theorem says that all primitive rings are, in some sense,
close to rings of matrices over division rings, then those with minimal left ideals are
closer. This is what we intend to show in this section.
Although we are presently interested in primitive rings, it seems more natural to
start things off with general rings.

Lemma 5.25 Let R be a ring having minimal left ideals. Then the sum of all minimal
left ideals of R is a two-sided ideal of R.

Proof We have to prove that the sum of all minimal left ideals of R is a right ideal.
Note that it is enough to show the following: If L is a minimal left ideal of R and
a ∈ R is such that La = 0, then La is also a minimal left ideal. To this end, consider
the map L → La, x → xa. It is obviously a surjective R-module homomorphism.
The minimality of L is equivalent to the condition that L, regarded as an R-module,
has no proper nonzero submodules. The kernel of our map is therefore equal to
0. Thus, L and La are isomorphic as R-modules. Hence La does not have proper
nonzero submodules, meaning that La is a minimal left ideal. 

Minimal one-sided ideals of a semiprime ring R were already considered in

Lemmas 2.58 and 2.60. Recall that a left (resp. right) ideal J of R is minimal if
and only if J is of the form J = Re (resp. J = e R), where e is an idempotent in R
such that e Re is a division ring. Furthermore, given an idempotent e ∈ R, Re is a
minimal left ideal of R if and only if e R is a minimal right ideal of R. In particular,
R has minimal left ideals if and only if R has minimal right ideals.

Lemma 5.26 Let R be a semiprime ring having minimal one-sided ideals. Then the
sum of all minimal left ideals of R coincides with the sum of all minimal right ideals
of R.

Proof Let T be a minimal right ideal of R. As just mentioned, there exists an

idempotent e ∈ R such that T = e R and L := Re is a minimal left ideal of R.
Therefore e = ee ∈ L. That is, e lies in a minimal left ideal, and hence T = e R is
contained in the sum of all minimal left ideals by Lemma 5.25. Similarly, by using
the obvious analogue of Lemma 5.25 for right ideals, we see that every minimal left
ideal is contained in the sum of all minimal right ideals. 

Definition 5.27 Let R be a semiprime ring. If R has minimal one-sided ideals, then
let soc(R) denote the sum of all minimal left ideals of R (equivalently, soc(R) is the
sum of all minimal right ideals of R). Otherwise we let soc(R) = 0. We call soc(R)
the socle of R.

As soc(R) is a two-sided ideal of R, it is equal to R if R is a simple ring having

a minimal one-sided ideal. For example, the ring of matrices over a division ring
R = Mn (D) readily coincides with soc(R) since it is equal to the sum of its minimal
5.4 Primitive Rings Having Minimal Left Ideals 117

left ideals R E ii , i = 1, . . . , n (cf. Example 3.54). However, this example is too

plain to reveal the true nature of the socle. The discussion that follows will give a
better insight.
From now on we consider the case where R is a primitive ring. Let us first recall
what we already know. The Jacobson Density Theorem says that there is an isomor-
phism r → r from R onto a dense ring R of linear operators of a vector space V
over a division ring Δ. We keep this notation until the end of the section. Assume
additionally that R has a minimal left ideal L. Then we can consider L as a faithful
simple R-module (in fact, it is enough to assume the primeness of R for this, see
Lemma 5.5). Corollary 3.52 tells us that V and Δ can be chosen so that V = Re
and Δ ∼ = (e Re)o for some idempotent e ∈ R. However, we shall not elaborate this
view any further. The clue for obtaining additional information about R is to consider
all minimal one-sided ideals of R, not only the chosen minimal left ideal L. More
precisely, the socle of R will play a fundamental role.
The following example is supposed to serve as a motivation for the next lemma.
Example 5.28 An example of a minimal left ideal of R = Mn (D) is L := R E 11 , the
set of all matrices that have zeros in all columns except in the first one (Example 2.57).
By Theorem 3.31 we may consider elements in R as linear operators. Note that all
nonzero elements in L have rank one, and each of them generates the left ideal L.
Lemma 5.29 Let R be a primitive ring, and let a ∈ R. Then the left ideal L generated
by a is minimal if and only if the operator a has rank one.
Proof Suppose a does not have rank one. We may assume that a = 0. Therefore
there exist u, v ∈ V such that a(u) and a(v) are linearly independent over Δ. By
density of R there exists r ∈ R such that r (a(u)) = 0 and r (a(v)) = 0. Accordingly,

I := {x ∈ L | x(u) = 0}

is a left ideal of R which is nonzero (since 0 = ra ∈ I ) and properly contained in L

(since a ∈ / I ). Therefore L is not minimal.
Assume now that a has rank one. Let J be a left ideal of R such that 0 = J ⊆ L.
We have to show that a ∈ J . Choose a nonzero y ∈ J , and let u ∈ V be such
that y(u) = 0. By density we can find r ∈ R satisfying r (y(u)) = a(u). Note
that a(u) = 0 since y ∈ L = Ra + Za and y(u) = 0. Now take an arbitrary
v ∈ V . Since a has rank one, we have a(v) = δa(u) for some δ ∈ Δ. Therefore
v − δu ∈ ker a ⊆ ker y, so that y(v) = δ y(u). Consequently,

r y(v) = r (δ y(u)) = δr (y(u)) = δa(u) = a(v).

Therefore a = r y ∈ J . 
Describing the socle of a primitive ring is now an easy task.
Theorem 5.30 Let R be a primitive ring, and let a ∈ R. Then a ∈ soc(R) if and
only if the operator a has finite rank.
118 5 Structure of Rings

Proof If a ∈ soc(R), then a = a1 + · · · + an where each ai lies in some minimal

left ideal L i . If ai = 0, then ai has rank one by Lemma 5.29. Hence a has rank at
most n.
Conversely, assume that a has finite rank. Let {v1 , . . . , vm } be a basis of the
image of a. Since R is dense, there exist r1 , . . . , rm ∈ R such that ri (vi ) = vi and
ri (v j ) = 0 if j = i. Note that r1 + · · · + rm acts as the identity on the range of a, and
so a = r1 a + · · · + rm a. Since each ri a has rank one, ri a lies in soc(R) by Lemma
5.29. Consequently, a ∈ soc(R). 
Example 5.31 If R = EndΔ (V ), then soc(R) is the set of all finite rank linear
operators of V .
We are, of course, interested in the situation where soc(R) = 0, i.e., R has minimal
left ideals. The next corollaries highlight what Theorem 5.30 says about such rings.
Corollary 5.32 A primitive ring R has minimal left ideals if and only if R contains
a nonzero finite rank operator.
Theorem 5.30 together with Lemma 5.14 gives
Corollary 5.33 A primitive ring R having minimal left ideals contains a nonzero
ideal (in fact, its socle) which is isomorphic to a dense ring of finite rank linear
operators of a vector space over a division ring.
The point here is, of course, that these operators are of finite rank. This brings
them close to operators on finite dimensional spaces, and hence close to matrices.
A primitive ring having minimal left ideals thus contains a particularly nice ideal.
Do we control the whole ring if we control one of its nonzero ideals? We certainly
cannot claim this for general rings. Say, if a ring R is equal to the direct product of
two unrelated rings R1 and R2 , then the properties of the ideal R1 × 0 have nothing
to do with the properties of the other part 0 × R2 . However, prime, and therefore
primitive rings cannot be decomposed in such a way. Problems on prime rings can
be often easily reduced to the same problems on any of their nonzero ideals.
Our final corollary can be considered as an extension of Wedderburn’s
Theorem 2.61.
Corollary 5.34 A simple unital ring R having minimal left ideals is isomorphic to
Mn (D) for some division ring D and n ∈ N.
Proof The simplicity of R implies that soc(R) = R. In particular, 1 ∈ soc(R),
and so 1 has finite rank by Theorem 5.30. Being the unity of a dense ring R, 1
must be the identity map on V . This means that V is finite dimensional over Δ,
and so R = EndΔ (V ) by Corollary 5.13. The desired conclusion now follows from
Theorem 3.31. 
Primitive rings having minimal left ideals form an important, easily approachable,
and also somewhat exclusive subclass of primitive rings. Every such ring must in
particular contain a nontrivial idempotent, unless it is a division ring. Therefore
a primitive domain, which is not a division ring, cannot have minimal left ideals.
A concrete example is the Weyl algebra A1 (cf. Example 2.28). Another important
example is a free algebra, which will be introduced in the next chapter.
5.5 Primitive Ideals 119

5.5 Primitive Ideals

Our aim now is to briefly examine the following notion that links primitive rings
with the central topic of the second part of this chapter, the Jacobson radical.

Definition 5.35 An ideal P of a ring R is said to be a primitive ideal if P is the

annihilator of a simple R-module.

(A formally more appropriate term is a left primitive ideal—not because it was

merely a left ideal, it certainly is two-sided, but because “module” means “left
module” by our convention.)
Clearly, R is a primitive ring if and only if 0 is a primitive ideal of R. More
generally, we have

Lemma 5.36 An ideal P of a ring R is a primitive ideal if and only if R/P is a

primitive ring.

Proof Let P be a primitive ideal of R and let M be a simple R-module such that
P = ann R (M). One immediately checks that the additive group of M becomes a
simple R/P-module if we define

(r + P)m := r m for all r ∈ R, m ∈ M.

This operation is indeed well-defined since P ⊆ ann R (M). On the other hand, from
P ⊇ ann R (M) we infer that M is a faithful R/P-module. Therefore the ring R/P
is primitive.
The proof of the converse is analogous. Starting with a faithful simple R/P-
module N , we note that by defining

r n := (r + P)n for all r ∈ R, n ∈ N ,

the additive group of N becomes a simple R-module (the well-definedness is not

an issue here). It is also clear that ann R (N ) = P. Accordingly, P is a primitive
ideal. 

Example 5.37 Let R = R1 × · · · × Rn , where each Ri is a primitive ring. Then every

Pi = R1 × · · · × Ri−1 × 0 × Ri+1 × · · · × Rn (5.6)

is a primitive ideal. Indeed, R/Pi ∼

= Ri is a primitive ring. In the special case where
Ri ∼= Mn i (Di ) for some n i ∈ N and a division ring Di , the Pi ’s are easily seen to be
the only primitive ideals of R. Moreover, in this case they are the maximal ideals.

Lemma 5.38 Let W be an ideal of a ring R. If R/W is a simple ring, then W is

a maximal ideal. Conversely, if W is maximal and R 2 ⊆ W (this is automatically
fulfilled if R is unital), then R/W is a simple ring.
120 5 Structure of Rings

Proof Let R/W be simple. If J is an ideal of R such that W ⊆ J ⊆ R, then J/W is

an ideal of R/W , so that J/W = 0 or J/W = R/W . This readily implies J = W
or J = R, proving the maximality of W .
If I is an ideal of R/W , then Z := {r ∈ R | r + W ∈ I } is an ideal of R which
contains W . Assuming that W is maximal it follows that Z = W or Z = R, that is,
I = 0 or I = R/W . Therefore R/W is a simple ring provided that (R/W )2 = 0,
i.e., R 2 ⊆ W . 
Since simple unital rings are primitive (Lemma 5.6), Lemmas 5.36 and 5.38 yield
Corollary 5.39 Every maximal ideal of a unital ring R is a primitive ideal.
Maximal ideals of Z and C[a, b], presented in Example 3.44, are therefore exam-
ples of primitive ideals. For commutative rings the converse to Corollary 5.39 actually
also holds.
Corollary 5.40 Every primitive ideal of a commutative ring R is a maximal ideal.
Proof If P is a primitive ideal of R, then R/P is a primitive commutative ring
by Lemma 5.36, and hence a field by Lemma 5.7. Therefore P is maximal by
Lemma 5.38. 
A primitive ideal of an algebra A is, of course, defined as the annihilator of a
simple (algebra) A-module. As one would guess, all results from this section still
hold if we replace “ring” by “algebra”. Since finite dimensional primitive algebras
are automatically simple by Corollary 5.19, a similar argument as in the proof of
Corollary 5.40 gives
Corollary 5.41 Every primitive ideal of a finite dimensional algebra A is a maximal
ideal.
In general rings it is easy to find examples of primitive ideals that are not maximal.
Say, the zero ideal of a non-simple primitive ring will do (see, e.g., Example 5.10).
Primitive ideals are also related to maximal left ideals.
Lemma 5.42 If P is a primitive ideal of a ring R, then there exists a maximal left
ideal U of R such that P = {x ∈ R | x R ⊆ U }. Conversely, if U is a maximal
left ideal of R and R 2 ⊆ U (this is automatically fulfilled if R is unital), then
P := {x ∈ R | x R ⊆ U } is a primitive ideal of R.
Proof Let M be a simple R-module such that P = ann R (M). By Lemma 3.46
there exists a maximal left ideal U of R such that M ∼ = R/U . Therefore P =
ann R (R/U ) = {x ∈ R | x R ⊆ U }.
Assume now that U is a maximal left ideal of R such that R 2 ⊆ U . Lemma 3.46
tells us that R/U is a simple R-module. Consequently, ann R (R/U ) =
{x ∈ R | x R ⊆ U } = P is a primitive ideal of R. 
Corollary 5.43 Every maximal left ideal U of a unital ring R contains a primitive
ideal P.
5.6 Introducing the Jacobson Radical 121

5.6 Introducing the Jacobson Radical

Recall that the radical of a finite dimensional algebra A is defined as the (necessarily
unique) maximal nilpotent ideal of A. A simple-minded attempt to use the same
definition for general rings immediately fails, simply because rings may not have
maximal nilpotent ideals (Example 2.12). Anyway, generalizations should have some
purpose. Our goal is to define the radical of a ring R as an ideal consisting of all
the most “undesired” elements in R, that is, the elements whose presence creates an
obstacle for understanding the structure of R. Further, we wish the factor ring of R
modulo the radical to be “nice” in the sense that its radical is zero, so that we can say
something about its structure. As witnessed by Theorem 2.65, the radical of a finite
dimensional algebra does satisfy these criteria: the corresponding factor algebra is
the direct product of full matrix algebras over division algebras. Of course, we have
to be realistic in our expectations when trying to extend this result to general rings.
One cannot hope that the structure of “nice” rings is as simple as the structure of
“nice” finite dimensional algebras.
There are different radicals defined for general rings, each of them having its own
relevance. From the point of view of the structure theory, the following radical is
arguably the most important.

Definition 5.44 The Jacobson radical of a ring R, denoted by rad(R), is the inter-
section of all the primitive ideals of R. If R has no primitive ideals (i.e., R has no
simple modules), then we define rad(R) = R.

In other words, rad(R) consists of all elements that annihilate every simple R-module.
It is obviously a (two-sided) ideal of R. Since our modules are actually left modules,
apparently a more appropriate name for rad(R) would be the left Jacobson radical.
Fortunately, it turns out that the Jacobson radical is a left-right symmetric notion, so
this terminological query is unnecessary.
The meaning of the Jacobson radical may not be evident from the definition. In
what follows we will discover several properties and characterizations of rad(R), and
thereby give a better insight into its nature. It truly is an amazing concept; we shall
see that it touches several seemingly unrelated notions, from simple modules and
primitive rings (and hence dense rings of linear operators) to maximal (two-sided
and one-sided) ideals, nilpotency and invertibility.
Let us first clarify the relation to nilpotent ideals. Actually, it is not more demand-
ing to consider more general nil ideals.

Lemma 5.45 Every nil left or right ideal of a ring R is contained in rad(R).

Proof Let I be a nil left (resp. right) ideal. If the lemma was not true, there would
exist a simple R-module M, u ∈ I and m ∈ M such that n := um = 0. Since
Rn = M (Lemma 3.36), we can find x ∈ R satisfying xn = m. That is, v := xu ∈ I
(resp. v := ux ∈ I ) satisfies vm = m (resp. vn = n), and hence vk m = m (resp.
vk n = n) for every k ∈ N. But this is impossible for v is nilpotent. 
122 5 Structure of Rings

Later we will see that even a domain may have a large Jacobson radical. Therefore
rad(R) is not always nil, let alone nilpotent.
It goes without saying that the Jacobson radical of an algebra A, rad(A), is
defined as the intersection of all primitive (algebra) ideals of A. We may, on the
other hand, treat an algebra merely as a ring. Fortunately, the Jacobson radical of the
ring A coincides with the Jacobson radical of the algebra A. We suppress the proof
of this fact (see, e.g., [Hun74, p. 451]). The only information about the Jacobson
radical of an algebra that we need is that it contains all nilpotent ideals. This is clear
from the proof of Lemma 5.45.

Theorem 5.46 If A is a finite dimensional algebra, then rad(A) is the unique max-
imal nilpotent ideal of A.

Proof Let N be the unique maximal nilpotent ideal of A (its existence is explained
in Sect. 2.1). We know that N ⊆ rad(A), so we only need to prove the converse
inclusion. Take q ∈ rad(A). We claim that q + N ∈ rad(A/N ). If this was not true,
there would be a simple A/N -module M such that (q + N )M = 0. Note that the
additive group of M becomes a simple A-module by defining

xm := (x + N )m for all x ∈ A, m ∈ M.

Since q ∈ rad(A), we have q M = 0. This contradicts (q + N )M = 0, and so our

claim is proved.
Let us complete the proof by showing that rad(A/N ) = 0. By Theorem 2.65 we
know that A/N is isomorphic to a finite direct product of full matrix algebras over
division algebras. We have thus arrived at the situation considered in Example
 5.37.
Defining primitive (in fact, maximal) ideals Pi as in (5.6) we clearly have i Pi = 0.
Consequently, rad(A/N ) = 0. 

The Jacobson radical of a finite dimensional algebra A thus coincides with the
radical of A introduced in Sect. 2.1.

5.7 Quasi-invertibility

The Jacobson radical is intimately connected with the concept of invertibility. This
statement does not seem to make much sense for non-unital rings. However, we shall
rely upon the following notions which make it possible to imitate invertibility in
arbitrary rings.

Definition 5.47 An element y in a ring R is said to be left quasi-invertible if there

exists x ∈ R such that

x + y = x y. (5.7)
5.7 Quasi-invertibility 123

We call x a left quasi-inverse of y. Analogously, x ∈ R is said to be right quasi-

invertible if there exists y ∈ R, called a right quasi-inverse of x, which satisfies
(5.7).

If R is unital, then (5.7) can be written as

(1 − x)(1 − y) = 1.

Therefore,

x is a left quasi-inverse of y ⇐⇒ 1 − x is a left inverse of 1 − y.

In particular,

y is left quasi-invertible ⇐⇒ 1 − y is left invertible.

(Thus 1 is not left quasi-invertible.) A similar statement can be made for right quasi-
invertibility.
There is another way of looking at these notions. Define ◦ : R × R → R by

x ◦ y := x + y − x y.

This operation is associative. Indeed, one immediately checks that

(x ◦ y) ◦ z = x + y + z − x y − x z − yz + x yz = x ◦ (y ◦ z).

Clearly x ◦ 0 = 0 ◦ x = x. Accordingly, R endowed with ◦ is a monoid with identity

element 0. Therefore (5.7) can be interpreted as that x is a left inverse of y with
respect to ◦. Since the left and right inverses in a monoid coincide, x ◦ y = y ◦ z = 0
implies x = z. We can reword this as follows.

Lemma 5.48 If an element y has a left quasi-inverse x and a right quasi-inverse z,

then x = z.

Definition 5.49 An element y ∈ R is said to be quasi-invertible if there exists

x ∈ R which is both a left and right quasi-inverse of y. In this case we call x a
quasi-inverse of y.

In other words, x is a quasi-inverse of y if it satisfies (5.7) and x y = yx. Actually,

x is the quasi-inverse of y; its uniqueness follows from Lemma 5.48. We also see
that y ∈ R is quasi-invertible if and only if it is both left and right quasi-invertible.
If R is unital, this is further equivalent to the invertibility of 1 − y.

Example 5.50 Every nilpotent element y ∈ R is quasi-invertible. Indeed, − ∞ n=1 y
n

is its quasi-inverse (the nilpotency of y implies that this sum is actually finite).
 If R nis
unital, we can rewrite this observation in the familiar form: (1 − y)−1 = ∞ n=0 y .
124 5 Structure of Rings

Definition 5.51 We say that S ⊆ R is a quasi-invertible set if all elements in S are

quasi-invertible. Analogously we define a left (resp. right) quasi-invertible set.

Lemma 5.52 If a left ideal L is left quasi-invertible, then L is quasi-invertible.

Proof Take  ∈ L. By assumption  has a left quasi-inverse k ∈ R. From k = k − 

we see that k also lies in L. Thus, k has a left quasi-inverse m. Lemma 5.48 implies
 = m, meaning that k is the quasi-inverse of . 

This simple lemma will eventually lead to the left-right symmetry of the Jacobson
radical.
Assume for a moment that R is unital. If 1 − a ∈ R is not left invertible, then
R(1 − a) is a proper left ideal of R, so it is contained in some maximal left ideal U
(Lemma 3.45), which gives rise to a simple R-module R/U (Lemma 3.46). These
observations can be easily extended to the non-unital situation.

Lemma 5.53 Let R be an arbitrary ring. If a ∈ R is not left quasi-invertible, then

there exists a maximal left ideal U of R such that a ∈
/ U . Moreover, R/U is a simple
R-module and a ∈ / ann R (R/U ).

Proof The assumption that a is not left quasi-invertible can be read as that a does not
lie in the left ideal L := {ra − r | r ∈ R}. Moreover, we claim that if L  is a proper
left ideal containing L, then a ∈ / L  . Indeed, by writing r ∈ R as r = −(ra − r ) + ra
we see that R is the only left ideal of R that contains L ∪ {a}. This makes it possible
for us to apply Zorn’s lemma in a similar way as in the proof of Lemma 3.45. Let S
be the set of all proper left ideals of R that contain
 L. Partially order S by inclusion.
If {K i | i ∈ I } is a chain in S , then K = i∈I K i is its upper bound. Indeed, K is
a proper left ideal since a ∈ / K . Therefore S has a maximal element U . Clearly, U
is a maximal left ideal and a ∈ / U . Furthermore, since a − a 2 ∈ L ⊆ U it follows
that a ∈
2 / U . Lemma 3.46 now tells us that R/U is a simple R-module. Finally,
a(a + U ) = 0 shows that a ∈ / ann R (R/U ). 

This lemma in particular shows that for every element that is not left quasi-
invertible we can find a simple module which is not annihilated by this element.
Elements from rad(R) are therefore necessarily left quasi-invertible, and hence quasi-
invertible by Lemma 5.52. This observation can be sharpened as follows:

Theorem 5.54 Let R be a ring and let q ∈ R. The following statements are equiv-
alent:
(i) The ideal generated by q is quasi-invertible.
(ii) The left ideal Rq is quasi-invertible.
(iii) q ∈ rad(R).

Proof (i) =⇒ (ii). Trivial.

(ii) =⇒ (iii). Suppose q ∈
/ rad(R). Then there exists a simple R-module M such
that qm = 0 for some m ∈ M. Accordingly, Rqm = M by Lemma 3.36. Therefore
5.7 Quasi-invertibility 125

m = m for some  ∈ Rq. Since  is quasi-invertible, there exists r ∈ R such that

r +  = r . Consequently, m = m = r m − r m = 0, which contradicts qm = 0.
(iii) =⇒ (i). As noticed before the statement of the theorem, Lemmas 5.52 and
5.53 show that rad(R) is a quasi-invertible ideal. Its subsets are therefore also quasi-
invertible. 

By making obvious modifications in the proof, one shows that (i) is also equivalent
to (ii’) the right ideal q R is quasi-invertible, and (iii’) q annihilates all simple right
R-modules. We have thereby established that the left Jacobson radical coincides with
the right Jacobson radical (see the comment following Definition 5.44).
The following corollary is immediate.

Corollary 5.55 The Jacobson radical of a ring R is a quasi-invertible ideal which

contains every quasi-invertible one-sided ideal of R.

Let us mention that Corollary 5.55 implies Lemma 5.45 since nilpotent elements
are quasi-invertible (see Example 5.50).

Corollary 5.56 The Jacobson radical of a unital ring R is equal to the intersection
of all the maximal left ideals of R.

Proof Let T denote the intersection of all the maximal left ideals of R. Clearly, T is
a left ideal. Lemma 5.53 along with Lemma 5.52 implies that T is quasi-invertible.
Hence T ⊆ rad(R) by Corollary 5.55. The fact that rad(R) ⊆ T clearly follows
from Corollary 5.43. 

It is interesting in its own right that the intersection of all the maximal left ideals
is a two-sided ideal. Of course we may replace “left” by “right” in the statement of
Corollary 5.56.
For the reader’s convenience we gather together various characterizations of the
Jacobson radical of a unital ring in the next corollary.

Corollary 5.57 Let R be a nonzero unital ring. The following statements are equiv-
alent for q ∈ R:
(i) q ∈ rad(R).
(ii) q M = 0 for every simple R-module M.
(iii) q lies in every ideal P of R such that R/P is a primitive ring.
(iv) 1− xq is invertible for all x ∈ R.
(v) 1 − i xi qyi is invertible for all xi , yi ∈ R.
(vi) q lies in every maximal left ideal of R.

One can add appropriate right versions of (ii)–(iv) and (vi).

126 5 Structure of Rings

5.8 Computing the Jacobson Radical

After discovering a variety of abstract descriptions of the Jacobson radical, it may

still be unclear how to determine rad(R) for a given ring R. There is no magical
device for this, but we can gain intuition by examining some special situations.
Let us begin with two illustrative examples. Since we know that the Jacobson
radical of a finite dimensional algebra is the maximal nilpotent ideal (Theorem 5.46),
our first example is just a repetition of Example 2.14.

Example 5.58 The Jacobson radical of Tn (F), the algebra of all upper triangular
n × n matrices over a field F, is the set of all strictly upper triangular matrices.

The ring in the next example has no nonzero nilpotent elements, in fact it is a
domain. Yet its Jacobson radical is quite big.

Example 5.59 Take a division ring D and consider the ring D[[ω]] of formal power
series. Let T denote its ideal generated by ω, i.e., T is the set of all elements with
zero constant term. Lemma 1.40 says that f ∈ D[[ω]] is not invertible if and only if
f ∈ T . This readily implies that every proper one-sided ideal of D[[ω]] is contained
in T . Therefore T is the unique maximal left ideal of D[[ω]]. Corollary 5.56 shows
that rad(D[[ω]]) = T .

The Jacobson radical is compatible with various ring constructions. Let us con-
sider only the simplest construction, i.e., the direct product.

Proposition 5.60 Let {Ri | i ∈ I } be a family of rings. Then

rad(Πi∈I Ri ) = Πi∈I rad(Ri ).

Proof Note that (ri ) is quasi-invertible in R := Πi∈I Ri if and only if ri is quasi-

invertible in Ri for every i. Accordingly, q = (qi ) ∈ R is such that Rq is quasi-
invertible in R if and only Ri qi is quasi-invertible in Ri for every i ∈ I . Thus, by
Theorem 5.54, q ∈ rad(R) if and only if qi ∈ rad(Ri ) for every i ∈ I . 

Relating the Jacobson radical of a ring to the Jacobson radical of some of its
subrings is another topic of interest. For ideals the relation is simple and nice.

Proposition 5.61 If I is an ideal of a ring R, then rad(I ) = I ∩ rad(R).

Proof Let y ∈ I ∩ rad(R). As an element of rad(R), y is quasi-invertible in R. If x is

its quasi-inverse, then we see from x = x y − y that x also lies in I . This shows that
I ∩ rad(R) is a quasi-invertible ideal of the ring I . Therefore I ∩ rad(R) ⊆ rad(I )
by Corollary 5.55.
Proving the converse inclusion is a bit more tricky. Take q ∈ rad(I ) and r ∈ R.
In view of Theorem 5.54 it is enough to show that u := rq is quasi-invertible in
R. Since u 2 = (rqr )q ∈ I rad(I ) ⊆ rad(I ), it follows that u 2 is quasi-invertible.
We claim that this implies that u is also quasi-invertible. In unital rings this follows
5.8 Computing the Jacobson Radical 127

easily from 1 − u 2 = (1 − u)(1 + u), but of course we do not wish to assume the
existence of unity. At any rate, the proof is easy. A direct computation shows that if
v is the quasi-inverse of u 2 , then vu + v − u is a left quasi-inverse of u and uv + v − u
is a right quasi-inverse of u. Thus u is quasi-invertible (and uv = vu) by Lemma
5.48. 

If the size of rad(R) measures the intricacy of the structure of R, the following
definition considers the worse (yet not the least interesting) situation.

Definition 5.62 A ring R is said to be a radical ring if rad(R) = R.

Let us mention two alternative ways to state this definition.

Proposition 5.63 The following statements are equivalent for a ring R:

(i) R is a radical ring.
(ii) Every element in R is quasi-invertible.
(iii) R has no simple modules.

Proof From Corollary 5.55 we see that (i) and (ii) are equivalent. If R is a radical
ring, then R cannot have simple modules since their annihilators are always proper
subsets of R. Thus, (i) implies (iii), while the converse holds by Definition 5.44.

Proposition 5.64 The Jacobson radical of an arbitrary ring R is a radical ring, i.e.,
rad(rad(R)) = rad(R).

Proof Take rad(R) for I in Proposition 5.61. 

Examples 5.58 and 5.59 therefore also yield examples of radical rings. The one
from the first example, the ring of all strictly upper triangular matrices, is nilpotent.
Such rings are automatically radical; more generally, every nil ring is a radical ring
by Lemma 5.45.

5.9 Semiprimitive Rings

After considering the most unfavorable situation where rad(R) = R, let us now
examine the opposite extreme.

Definition 5.65 A ring R is said to be a semiprimitive ring if rad(R) = 0.

Example 5.66 Every primitive ring is semiprimitive since 0 is its primitive ideal.
More generally, the direct product of primitive rings (in fact, even of semiprimitive
rings) is semiprimitive by Proposition 5.60.

Example 5.67 If R is a nonzero commutative unital ring, then rad(R) is equal to

the intersection of all the maximal ideals of R by Corollary 5.56. With reference to
Example 3.44 we thus have:
128 5 Structure of Rings

(a) Since the intersection of all the maximal ideals pZ of Z is 0 (here, p is a prime
number), Z is a semiprimitive ring.
(b) Similarly, since a≤c≤b Uc = 0, where Uc = { f ∈ C[a, b] | f (c) = 0}, C[a, b]
is a semiprimitive ring.

A semiprimitive ring R cannot have nonzero nil ideals by Lemma 5.45. In

particular, semiprimitive rings are thus semiprime. The converse is not true.

Example 5.68 Let D be a division ring. In Example 5.59 we saw that D[[ω]] has a big
Jacobson radical. Thus, D[[ω]] is an example of a semiprime ring (in fact, a domain)
that is not semiprimitive. The polynomial ring D[ω], however, is semiprimitive.
Indeed, take f ∈ rad(D[ω]). Then f is quasi-invertible by Corollary 5.55; that is,
1 − f is invertible. Since polynomials of degree 0 are clearly the only invertible
elements in D[ω], and since f ∈ rad(D[ω]) cannot be invertible, it follows that
f = 0.

The reader might have noticed that we did not use the full force of the assumption
that D is a division ring. Instead of analyzing what else the above simple argument
brings, we mention a much deeper result by S. A. Amitsur from 1956: Every ring R
has a nil ideal N such that rad(R[ω]) = N [ω]. Consequently, R[ω] is semiprimitive
if R has no nonzero nil ideal. But even Amitsur’s theorem is not the end of the story.
A precise description of N is still a mystery. It turns out that the validity of the
statement “If R is nil, then N = R” is equivalent to a positive answer to Köthe’s
problem.
It is not always easy to find out whether or not a ring is semiprimitive. Here is a
famous old problem:

Problem 5.69 Let G be a group and let F be a field with char(F) = 0. Is the group
algebra F[G] semiprimitive?

The answer is “yes” if G is finite. Indeed, this is just a rewording of Maschke’s

theorem (namely, as is evident from Theorem 5.46, a finite dimensional algebra is
semiprime if and only if it is semiprimitive). Thus, we are asking whether Maschke’s
theorem can be generalized to infinite groups. It turns out that the problem can be
reduced to the case where F = Q.
Let us return to problems we can handle. At the beginning of Sect. 5.6 we said
that a decent radical should have the property that the factor ring of R modulo the
radical has zero radical. We still owe the proof that the Jacobson radical does satisfy
this property.

Lemma 5.70 The ring R/rad(R) is semiprimitive for every ring R.

Proof The case where R is a radical ring is trivial. By Proposition 5.63, we may
therefore assume that R has simple modules. Take such a module M. The additive
group of M becomes a simple module over R/rad(R) by defining

(x + rad(R))m := xm for all x ∈ R, m ∈ M.

5.9 Semiprimitive Rings 129

Checking this is straightforward, including the well-definedness which follows from

rad(R)M = 0. Now, if q + rad(R) ∈ rad(R/rad(R)), then q M = (q + rad(R))M =
0. Since M is an arbitrary simple R-module, this gives q ∈ rad(R). 

Another debt to the reader is an explanation why semiprimitive rings are supposed
to be “nice”. Why would it be easier to deal with a semiprimitive ring than with any
old ring? For primitive rings this is clear as we have the Jacobson Density Theorem.
Direct products of primitive rings can therefore also be viewed as easily approachable.
Unfortunately, not every semiprimitive ring can be represented as such a product
(say, Z cannot). We are forced to deal with the following generalization of the direct
product.

Definition 5.71 Let {Ri | i ∈ I } be a family of rings. A subring R of the direct

product Πi∈I Ri is called a subdirect product of the family {Ri | i ∈ I } if for each
j ∈ I , the map π j : R → R j , π j ((ri )) = r j , is surjective.

If I = {1, 2}, this means that for every r1 ∈ R1 there exists r2 ∈ R2 such
that (r1 , r2 ) ∈ R, and similarly, for every r2 ∈ R2 there exists r1 ∈ R1 such that
(r1 , r2 ) ∈ R. For instance, every ring R is isomorphic to the subdirect product of two
copies of R via r → (r, r ). Subdirect products are thus quite different from direct
products.

Lemma 5.72 A nonzero ring R is semiprimitive if and only if R is isomorphic to a

subdirect product of primitive rings.

Proof Let R be semiprimitive. Denote the family of all the primitive ideals of R by
{P
 i | i ∈ I }. Recall that the rings R/Pi are primitive (Lemma 5.36). By assumption,
i∈I Pi = 0. This shows that the homomorphism

R → Πi∈I R/Pi , r → (r + Pi ) (5.8)

is injective. Its image is clearly a subdirect product of the family {R/Pi | i ∈ I }.

Conversely, take a family of primitive rings {Ri | i ∈ I } and let ϕ : R → Πi∈I Ri
be an injective homomorphism such that π j (ϕ(R)) = R j for every j ∈ I . Set
P j := ker π j ϕ. Since ∼
 R/P j = R j we see, again by Lemma 5.36, that P j is a primitive

ideal of R. If r ∈ i∈I Pi , then ϕ(r ) = 0 and hence r = 0. Thus rad(R) ⊆ i∈I
Pi = 0. 

Example 5.73 Let R = A be a finite dimensional semiprimitive algebra. As men-

tioned above, A is then in particular semiprime, and hence, by Wedderburn’s theorem,
isomorphic to a finite direct product of simple (= primitive in this context) algebras.
Therefore A has only finitely many primitive ideals (cf. Example 5.37). Note that
the homomorphism from (5.8) is surjective.

Example 5.74 From Example 5.67 and the proof of Lemma 5.72 we infer the fol-
lowing:
130 5 Structure of Rings

(a) Z is isomorphic to a subdirect product of the fields Z p , p prime.

(b) C[a, b] is isomorphic to a subdirect product of the fields C[a, b]/Uc ∼
= R,
a ≤ c ≤ b (cf. Example 3.44).

Roughly speaking, Lemma 5.72 shows that a semiprimitive ring can be decom-
posed into primitive factors, however, in a fashion which is not always easily con-
trollable. When making this decomposition we somehow lose track of the way the
components are “glued together”. The example with C[a, b] is illustrative. The direct
product of the fields C[a, b]/Uc ∼ = R, a ≤ c ≤ b, can be identified with the ring
of all functions from [a, b] into R. Thus, when applying the method of Lemma 5.72
to C[a, b] we lose the information that functions from our ring are continuous. The
structure theory is powerful, but has its limitations.
Our final result in this section summarizes some of the main features of the
structure theory. One can compare it with the analogous result concerning finite
dimensional algebras, Theorem 2.65.

Theorem 5.75 If a ring R is not a radical ring, then the ring R/rad(R) is isomorphic
to a subdirect product of dense rings of linear operators of vector spaces over division
rings.

Proof Combine Lemmas 5.70 and 5.72 with the Jacobson Density Theorem. 

5.10 Structure Theory in Action

The structure theory that we have developed is not only beautiful, it is truly applicable.
The purpose of this final section is to indicate how the theory works in “real life”.
Suppose we are given a problem on a ring R. The following procedure, indicated
by Theorem 5.75, sometimes leads to its solution:
(a) By factoring out rad(R) one reduces the problem to the case where R is semi-
primitive.
(b) Using the representation of a semiprimitive ring R as a subdirect product of prim-
itive rings, one further reduces the problem to the case where R is primitive.
(In practice one does not necessarily work with subdirect products, but uses
the fact that the intersection of the primitive ideals of R is 0. We have
introduced subdirect products primarily to indicate the philosophical aspect of
semiprimitivity.)
(c) A problem on a primitive ring R can be usually solved by representing R as a
dense ring of linear operators of a vector space V over a division ring Δ.
(d) If dimΔ V = 1, the density condition does not have the same meaning as other-
wise. In this case R is a division ring (isomorphic to Δo ) and a different argument
may be needed.
Needless to say, we cannot solve just any problem by following this procedure.
There are, however, various important instances where at least some, if not all of
5.10 Structure Theory in Action 131

the steps (a)–(d) can be realized. We will give an evidence of the feasibility of the
procedure by presenting a proof of (yet another) theorem by Jacobson. This theorem
was one of the first applications of the structure theory, and still is one of the most
striking ones.
We first record two elementary lemmas. The reader is probably familiar with the
first one. Anyway, we give the proof for the sake of completeness.

Lemma 5.76 If K is a finite field with char(K ) = p, then there exists n ∈ N such
n
that |K | = p n and k p = k for every k ∈ K . Moreover, if k1 , . . . , k pn −1 are the
nonzero elements in K , then
n
ω p − ω = (ω − k1 ) . . . (ω − k pn −1 )ω. (5.9)

Proof We may consider K as a vector space over the prime subfield F0 ∼ = Zp.
If K is n-dimensional over F0 , then |K | = |F0n | = p n . The multiplicative group
K ∗ = K \{0} thus has p n − 1 elements, so that k p −1 = 1 for every k ∈ K ∗ .
n
n
Consequently, k = k for every k ∈ K . That is to say, every element in K is a root
p
n
of the polynomial ω p − ω. Since the degree of this polynomial is equal to |K |, the
desired conclusion follows. 

The next lemma concerns powers of inner derivations in rings of prime character-
istic. We denote the inner derivation x → ax − xa by ad a (this notation is borrowed
from the theory of Lie algebras). We first remark that in any ring R, regardless of its
characteristic, for all a, x ∈ R we have


m 
m m−k k
(ad a)m (x) = (−1)k a xa . (5.10)
k
k=0

m  m  m+1
Using k−1 + k = k one easily proves this by induction on m.
n
Lemma 5.77 If R is a ring with char(R) = p, where p is prime, then (ad a) p =
n
ad a p for all a ∈ R and n ∈ N.
  p!
Proof If 1 ≤ k ≤ p−1, then p does not divide k!( p−k)!, so it divides kp = k!( p−k)! .
Therefore (5.10) reduces to (ad a) = ad a if m = p (if p = 2 then this is true
p p

as xa 2 = −xa 2 ). This proves the claim of the lemma for n = 1. We may therefore
assume the claim is true for n − 1. Consequently,

n


n−1 p


n−1 p  n−1  p n
(ad a) p = (ad a) p = ad a p = ad a p = ad a p ,

as desired. 

Let us now tackle the Jacobson’s theorem. First, a few words about the background.
The reader is perhaps familiar with the notion of a Boolean ring. This is a ring in
which every element is an idempotent (e.g., a direct product of copies of Z2 ). One
132 5 Structure of Rings

of the first things one learns about such rings is that they are commutative. The
proof is easy: If x and y are elements of a Boolean ring R, then x, y, x + y are
idempotents, which readily implies x y + yx = 0; setting y = x we get char(R) = 2,
so x y + yx = 0 can be written as x y = yx. This was extremely easy, but the fact
itself may appear a bit surprising at first glance since idempotents in rings do not
always commute. Now consider a more general condition where a ring R satisfies
x 3 = x for every x ∈ R. Is R commutative? It turns out that this is true indeed,
but giving a proof may already be a small challenge for the reader. The case where
x 4 = x is yet more entangled, but again the commutativity can be derived by ad hoc
arguments. If we wish to consider the general case where there exists n ≥ 2 such
that

x n = x for all x ∈ R, (5.11)

the need for a conceptual approach becomes apparent. But first of all, why would
anyone be interested in (5.11)? As noticed in Lemma 5.76, every finite field F satisfies
(5.11) with n = |F|. The very same proof shows that the same holds for every finite
division ring. Surely, such rings are automatically fields by Wedderburn’s theorem
(Theorem 1.38), but this is exactly the point we wish to make. We are wondering
whether Wedderburn’s theorem can be generalized to arbitrary rings via (5.11). And
indeed it can. Even better, there is no need to fix n; it may depend on x.

Theorem 5.78 (Jacobson) If a ring R is such that for every x ∈ R there exists an
integer n(x) ≥ 2 such that x n(x) = x, then R is commutative.

Proof We will follow the procedure outlined above.

(a) In the first step we will show that rad(R) = 0 (so that there is no need to factor
out rad(R)). Choose q ∈ rad(R), and set n = n(q). Then q n−1 also lies in
rad(R), so it is quasi-invertible (Corollary 5.55). Thus, there exists r ∈ R such
that q n−1 + r = rq n−1 . Multiplying this identity from the right by q and using
q n = q we obtain q = 0.
(b) We now know that R is semiprimitive. We may assume that  R = 0, so that there
is a family {Pi | i ∈ I } of primitive ideals of R such that i∈I Pi = 0. Therefore
R is commutative if and only if each primitive ring R/Pi is commutative. Since
R/Pi satisfies the same condition, x n(x) = x for every x ∈ R/Pi , we see that
without loss of generality we may assume that R itself is primitive.
(c) In view of the Jacobson Density Theorem, we may assume that R is a dense ring
of linear operators of a vector space V over a division ring Δ. Suppose there exist
linearly independent vectors u 1 , u 2 ∈ V . Choose f ∈ R such that f (u 1 ) = u 2
and f (u 2 ) = 0. Then f n (u 1 ) = 0 for every n ≥ 2, while f (u 1 ) = 0. However,
this contradicts our assumption that f n( f ) = f . The result is thus proved if
dimΔ V ≥ 2.
5.10 Structure Theory in Action 133

It remains to treat the case where dimΔ V = 1, i.e., the case where R =∼ Δo is a
division ring. This is the difficult part of the proof.
(d) Let us write D instead of R. We begin by pointing out two facts that follow
immediately from the condition x n(x) = x, x ∈ D. The first one is that every
nonzero subring of D is automatically a division ring (since the inverse of x is
a power of x), and the second one is that D has finite characteristic p (indeed,
just take 2 = 1 + 1 for x).
Suppose, on the contrary, that D is not commutative. Take a ∈ D\Z (D). Since
D has finite characteristic and a n(a)−1 = 1, we see that the subring generated
by a, let us call it K , is finite. As observed in the previous paragraph, K is
necessarily a field, so we can invoke Lemma 5.76. Thus, there exists n ∈ N
n
such that |K | = p n and, in particular, a p = a. Consider D as a vector space
over K . The inner derivation d := ad a is clearly a K -linear map, so it can be
considered as an element of the K -algebra End K (D). We can therefore evaluate
n
the polynomial from (5.9) at d. Since d p = d by Lemma 5.77, we obtain

(d − k1 ) . . . (d − k pn −1 )d = 0, (5.12)

where the ki ’s are the nonzero elements of K . Note that, in this context, ki
should be understood as the map D → D, x → ki x. Now, since a ∈ / Z (D), d is
a nonzero map. From (5.12) it thus follows that at least one of the maps d−ki has a
nontrivial kernel. This means that there is b ∈ D such that ab−ba = d(b) = ki b
with 0 = ki ∈ K . Accordingly, a and b do not commute and ba ∈ K b.
Let D0 be the subring of D generated by a and b. As observed at the beginning,
D0 is in fact a division ring. Of course, it is noncommutative for ab = ba. From
ba ∈ K b it clearly follows that ba r ⊆ K b for every r ∈ N. Hence bK ⊆ K b,
and consequently, bs K ⊆ K bs for every s ∈ N. This shows that

K + K b + · · · + K bn(b)−2

is a subring of D. Since it is a subset of D0 and since it contains a and b, it

is actually equal to D0 . But then D0 is a finite set. However, this contradicts
Wedderburn’s theorem on finite division rings (Theorem 1.38). 

The first three steps (a)–(c) were quite easy, but nevertheless fascinating. They
illustrated how the structure theory can be used to reduce a problem on a general
ring to a much simpler ring, in this case a division ring. After that we were left to
ourselves, and so the last step (d) had to be more ingenious.
It should be mentioned that Jacobson’s original proof was somewhat different.
We have mainly followed the proof due to I. N. Herstein.
134 5 Structure of Rings

Exercises

5.1. Let e be a nonzero idempotent in a ring R. Show that eV is a simple e Re-

module if V is a simple R-module and eV = 0. Hence infer that e Re is
primitive if R is primitive. Use this to show that the rings Mn (Z) and Mn (F[ω])
are prime but not primitive.
5.2. Show that the ring Mn (R) is primitive if and only if R is primitive.
5.3. Let R be a primitive ring with a faithful simple R-module V . Show that R is
a division ring if and only if for all r ∈ R and v ∈ V , r v = 0 implies r = 0 or
v = 0.
5.4. Let R be a primitive ring. Suppose that a, b ∈ R satisfy (xa)2 = (xb)2 for
every x ∈ R. Show that either a = b or a = −b.
5.5. Give an example showing that the assumption that F is algebraically closed
in Burnside’s theorem (Corollary 5.23) is necessary.
5.6. Give an alternative proof of Lemma 1.24 by using the Artin-Whaples Theorem
(Corollary 5.24).
5.7. Show that the Weyl algebra A1 is a dense algebra of linear operators of F[ω].
Remark: Since A1 is primitive (Example 5.9), the Jacobson Density Theorem
for algebras tells us that A1 is isomorphic to a dense algebra of linear operators
of a vector space over some division algebra. What has to be shown is that
A1 is a dense algebra of linear operators of the original space F[ω], i.e., the
fulfillment of the conditions of Theorem 5.21 has to be verified.
5.8. Let R be a ring, V be an R-module, and ϕ be an automorphism of R. Note
that the additive group of V becomes an R-module if we define the module
operation by x · v = ϕ(x)v for all x ∈ R, v ∈ V . If V is faithful and simple,
then so is the new module. Use this construction to show that the Weyl algebra
A1 has nonisomorphic faithful simple modules.
5.9. Let R be a primitive ring having a minimal left ideal L. Show that every faithful
simple R-module is isomorphic to L.
Remark: This is a generalization of Lemma 3.40.
5.10. Fields are the only commutative primitive rings (Lemma 5.7), and the center of
a simple unital ring is necessarily a field (Example 1.21). However, the center
of a unital primitive ring may not be a field. Find an example of such a ring.
Hint: If R is a primitive ring having minimal left ideals, then every subring of
R that contains soc(R) is also primitive.
5.11. Show that the sum of minimal left ideals of the upper triangular matrix ring
T2 (F) does not coincide with the sum of its minimal right ideals.
5.12. Let R be a primitive ring having minimal left ideals. Show that soc(R) is
simple as a ring, and is contained in every nonzero ideal of R.
5.13. Let R be a unital ring, and let M be a nonzero finitely generated unital R-
module. Show that rad(R)M is a proper submodule of M.
Remark: This result, known as Nakayama’s lemma, is particularly well-
known in commutative algebra and algebraic geometry.
Exercises 135

5.14. A unital ring R is said to be local if the set of all non-invertible elements in
R forms an ideal. Show that this ideal is exactly rad(R), and that R is local
if and only if R/rad(R) is a division ring. Verify that the following rings are
local:
(a) The ring D[[ω]] where D is a division ring (cf. Example 5.59).
(b) The subring of Mn (D), D a division ring, consisting of upper triangular
matrices all of whose diagonal entries are equal.
(c) Z pn where p is prime and n ∈ N.
(d) The unitization of a radical algebra.
5.15. Let a be an element in a unital ring R. Show that the following statements are
equivalent:
(i) a is invertible.
(ii) a + rad(R) is invertible in R/rad(R).
(iii) a + P is invertible in R/P for every primitive ideal P of R.
Hint: If a does not have a left inverse, then there exists a maximal left ideal
containing a. If a has a left inverse b, then (1 − ab)a = 0.
5.16. Let a be an element in a ring R. Show that if there exists n ∈ N such that a n
is quasi-invertible, then so is a.
Hint: After examining the (intuitively more clear) case where R is unital, one
can easily guess how to handle the general case.
5.17. Let a be a nonzero element in a ring R such that a n = a for some n ≥ 2. Show
that a ∈
/ rad(R). (In particular, rad(R) does not contain nonzero idempotents.)
5.18. Let a, b be elements in a ring R. Show that a Rb is a quasi-invertible set if and
only if ba ∈ rad(R).
Hint: This can be proved either directly (it is not difficult) or as an application of
the theory along with the following (quite remarkable) fact: For any u, v ∈ R,
uv is quasi-invertible if and only if vu is quasi-invertible. In a unital ring this
means that 1 − uv is invertible if and only if 1 − vu is invertible.
5.19. Let R be an arbitrary ring. Show that
 
R R rad(R) R
rad = .
0 R 0 rad(R)

Remark: Of course, this can be extended to upper triangular matrices over

R of any size. It is also appropriate to mention the formula rad(Mn (R)) =
Mn (rad(R)), which could as well be included among exercises. Its proof is not
difficult, yet somewhat more involved than the one concerning the triangular
case.
5.20. Show that a nonzero ring R is semiprimitive if and only if R has a faithful
semisimple module.
5.21. Let R be a semiprimitive ring. True or False:
136 5 Structure of Rings

(a) If for every primitive ideal P of R the ring R/P has no nonzero nilpotent
elements, then R has no nonzero nilpotent elements.
(b) If for every primitive ideal P of R the ring R/P has no nontrivial idem-
potents, then R has no nontrivial idempotents.
5.22. Let R be a semiprimitive ring. Show that the following statements are equiv-
alent for a ∈ R:
(i) For every x ∈ R there exists m(x) ∈ N such that [a, x]m(x) = 0.
(ii) For every x ∈ R there exists n(x) ∈ N such that (ax)n(x) = (xa)n(x) .
(iii) a ∈ Z (R).
5.23. Show that an algebraic division algebra over a finite field is commutative.
Hint: One way of proving this result, which is also due to Jacobson, is to use
Jacobson’s Theorem 5.78, together with standard facts about fields and their
extensions.
5.24. Show that a ring R is commutative if every commutator in R is an idempotent.
Remark: This is a very special case of Herstein’s generalization of Theorem
5.78 which says a ring R is commutative if and only if for every pair x, y ∈ R
there exists an integer n(x, y) ≥ 2 such that [x, y]n(x,y) = [x, y].
Chapter 6
Noncommutative Polynomials

Polynomials in several variables are a fundamental topic of interest in commutative

algebra. Their counterparts in noncommutative algebra are the noncommutative poly-
nomials, i.e., polynomials in noncommuting indeterminates. Thus, if ξ ,η are such
indeterminates, then ξ η and ηξ are different noncommutative polynomials. Opera-
tions on such polynomials are defined in the obvious way. For example, if f = ξ η +η
and g = ηξ − η, then

f + g = ξ η + ηξ, fg = ξ η2 ξ + η2 ξ − ξ η2 − η2 , g f = ηξ 2 η − η2 .

The purpose of this chapter is to give a basic introduction to a couple of topics

related to noncommutative polynomials. There will be plenty of examples and simple
observations, more than theorems and proofs. The emphasis will be on polynomial
identities. These are noncommutative polynomials which vanish when replacing
indeterminates with arbitrary elements from a given algebra. However, most of the
deeper results on polynomial identities will be obtained in the next chapter, when we
will have more tools at our disposal. The main goal of this chapter is to introduce an
appropriate framework.

6.1 Free Algebras

After the above informal introduction, let us define noncommutative polynomials

precisely. Take a nonempty set X. It is convenient to consider it as an indexed set, so
let us write X = {ξi | i ∈ I}. A finite sequence of elements from X will be denoted
by ξi1 ξi2 . . . ξim and called a word. The empty sequence is not excluded; we denote
it by 1 and call it the empty word. Defining multiplication by juxtaposition, i.e.,

(ξi1 ξi2 . . . ξim ) · (ξj1 ξj2 . . . ξjn ) := ξi1 ξi2 . . . ξim ξj1 ξj2 . . . ξjn ,

© Springer International Publishing Switzerland 2014 137

M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_6
138 6 Noncommutative Polynomials

the set of all words becomes a monoid (with unity 1), which we denote by X ∗ . It is
called the free monoid on X. Viewing ξi ∈ X as an element of X ∗ , we can write
ξi2 for ξi ξi , ξi3 for ξi ξi ξi , etc. Accordingly, every w = 1 in X ∗ can be written as
w = ξik11 ξik22 . . . ξikrr , where ij ∈ I and kj ∈ N. Of course, ij can be equal to ik if j = k,
but we can achieve that ij = ij+1 .
Recall that given a monoid and a field one can form a monoid algebra.

Definition 6.1 Let F be a field and X a set. The free algebra on X over F is the
monoid algebra of X ∗ over F.

In short, this is the algebra whose basis is the set of all words formed from X
(including 1) and whose multiplication is determined by juxtaposition of words.
The free algebra on X over F will be denoted by FX, or, if X is a finite
(resp. countably infinite) set, also by Fξ1 , . . . , ξn  (resp. Fξ1 , ξ2 , . . . ). The ele-
ments in X are called indeterminates. If X has just two elements, then we will
avoid indices and write Fξ, η. We call elements in free algebras noncommutative
polynomials. However, we will usually omit “noncommutative” and call them sim-
ply polynomials. The classical polynomials in commuting indeterminates will also
occasionally appear in the sequel, but there is little chance for confusion. Already
the notation will be different. We will use ω for denoting commuting indeterminates,
and ξ or η for noncommuting indeterminates.
The notions such as a coefficient, the constant term, and a constant polynomial
are defined for noncommutative polynomials in the same way as for commutative
polynomials.
The role of the field F in the definition of FX is purely formal. One can replace
it with an arbitrary ring R and form the ring RX of (noncommutative) polynomials
with coefficients in R. The operations in RX are defined just as in FX. In particular,
elements from R commute with indeterminates from X, but not necessarily between
themselves. Studying RX at a basic level is not much harder than studying FX with
F a field. Nevertheless, we restrict ourselves to the latter, at least in this chapter. In
the next one we will actually encounter the case where R = Z. One can also describe
ZX as the subring of QX consisting of polynomials with integer coefficients.
Let A be another unital F-algebra. A unital algebra homomorphism (i.e., an algebra
homomorphism that sends 1 into 1) from FX into A is uniquely determined by the
action on X. This is clear since FX is generated by the set X ∪ {1}. What we wish to
point out is a kind of converse: An arbitrary function f from X into A can be extended
to a unital algebra homomorphism f from FX into A. Indeed, there is only one way
to define this extension, and it is well-defined and is a homomorphism. The following
diagram is thus commutative.
x x
X F X
f
f

A
6.1 Free Algebras 139

This may be just a simple observation, but is of fundamental importance. We will refer
to it as to the universal property of FX. (Readers familiar with the categorical
language will recognize that this property shows that FX is a free object in the
category of unital F-algebras.)
Every polynomial f ∈ FX involves only a finite number of indeterminates.
If these are ξi1 , . . . , ξin , then we write f = f (ξi1 , . . . , ξin ). Usually we will simply
write f = f (ξ1 , . . . , ξn ). Given such a polynomial and elements x1 , . . . , xn ∈ A, we
define the evaluation f (x1 , . . . , xn ) in the obvious way. Say, if f = ξ12 ξ2 ξ1 −ξ2 ξ3 +1,
then f (x1 , x2 , x3 ) = x12 x2 x1 − x2 x3 + 1 ∈ A, i.e., one just substitutes xi for ξi . In more
formal terms, the universal property shows that the function ξi → xi gives rise to the
unital homomorphism from Fξ1 , . . . , ξn  into A which maps f into f (x1 , . . . , xn ).
If A is not unital, then f (x1 , . . . , xn ) still makes sense provided that f has zero constant
term.
A polynomial is said to be a monomial if it is a nonzero scalar multiple of a
word. Every nonzero polynomial f ∈ FX is a sum of monomials. More precisely,
f can be uniquely written as f = λ1 w1 + · · · + λm wm where w1 , . . . , wm are pairwise
different words and the coefficients λ1 , . . . , λm are nonzero scalars. Then f is the
sum of the monomials λi wi .
We define the length of a nonempty word w = ξi1 ξi2 . . . ξim as (w) := m, and
the length of the empty word as (1) := 0. The degree of a nonzero polynomial
f = λ1 w1 + · · · + λm wm (λi = 0) is defined as

deg( f ) := max{(w1 ), . . . , (wm )}.

If (w1 ) = · · · = (wm ), i.e., if all monomials of f have the same degree, then f is said
to be a homogeneous polynomial. For instance, ξ 2 η − ξ ηξ + η3 is homogeneous of
degree 3. We will be often concerned with homogeneous polynomials of the following
type: f = f (ξ1 , . . . , ξn ) is called a multilinear polynomial if each indeterminate ξi
appears in every monomial of f exactly once. That is, f is of the form

f = λσ ξσ (1) . . . ξσ (n) ,
σ ∈Sn

where λσ ∈ F and Sn is the symmetric group on {1, . . . , n}. Observe that for every
F-algebra A, the map (x1 , . . . , xn ) → f (x1 , . . . , xn ), xi ∈ A, is multilinear.
Take nonzero polynomials f = λ1 w1 + · · · + λm wm and g = μ1 z1 + · · · + μn zn ,
and assume, for notational simplicity, that deg( f ) = (w1 ) and deg(g) = (z1 ).
Note that w1 z1 = wk zl occurs only when k = l = 1. Hence deg( fg) = (w1 z1 ) =
(w1 ) + (z1 ). Thus,

deg( fg) = deg( f ) + deg(g)

holds for all nonzero f , g ∈ FX (if one defines deg(0) := −∞, then “nonzero”
can be omitted). An immediate consequence is that FX is a domain.
140 6 Noncommutative Polynomials

Remark 6.2 A much deeper result is that FX is a primitive algebra unless X is
a singleton (note that Fξ  is just the usual polynomial algebra F[ξ ] which is not
primitive by Lemma 5.7). We shall not need this result, so we omit the proof. Let us
just give its rough outline for the simplest case where X = {ξ, η}. Take a vector space
V over F with basis {en | n ∈ N}. Define g, h ∈ EndF (V ) by g(e1 ) = 0 and g(en ) =
en−1 for n = 1, and h(en ) = en2 +1 for every n. Let A be the subalgebra of EndF (V )
generated by g and h. It is not hard to prove that A is a primitive algebra, in fact a dense
algebra of linear operators of V . A slightly harder, but still elementary exercise is to
prove that Fξ, η is isomorphic to A. The point is to show that the homomorphism
f (ξ, η) → f (g, h), which exists by the universal property, is injective. Details can
be found, for example, in [Lam01, pp. 184–185].

6.2 Algebras Defined by Generators and Relations

Let A be a unital algebra over a field F. Take a subset X of A that generates A. In

principle, X can be any set of generators, but in practice we are more interested in
“smaller” sets. Forget for the moment that X is a subset of an algebra and consider
it merely as a set. Then we can form the free algebra FX. The function ξ → ξ
from the set X into the algebra A can be, as pointed out in the preceding section,
uniquely extended to a unital algebra homomorphism ϕ from FX into A. As its
range contains X, ϕ is surjective. We have thereby shown that every unital algebra
is a homomorphic image of a free algebra. Setting I = ker ϕ we can rephrase this as
follows: For every unital F-algebra A there exists a free algebra FX and its ideal I
such that

A∼
= FX/I.

These simple observations indicate that free algebras play a special role in non-
commutative algebra. They are hidden behind every algebra. When searching for
counterexamples, one often starts by examining free algebras.
Let us now change the viewpoint. We begin by taking a nonempty set X and a
nonempty subset R of FX. Let us form the factor algebra FX/(R) where (R) is the
ideal of FX generated by R. As noticed above, every unital algebra is isomorphic
to an algebra of this kind. This construction thus gives us endless opportunities
to create interesting examples. Also, it makes it possible for us to get a clearer
picture of some important algebras that we often come across. Before proceeding
to examples, we introduce  some  notation and terminology. We set X = {ξi | i ∈ I}
and R = { fj = fj ξi1 , . . . , ξin( j ) | j ∈ J}. Denote the coset ξi + (R) by xi . Note that
 
fj xi1 , . . . , xin( j ) = 0 for every j ∈ J. Let us switch to a more suggestive notation
and write FX/(R) as
 
Fxi , i ∈ I | fj xi1 , . . . , xin( j ) = 0, j ∈ J.
6.2 Algebras Defined by Generators and Relations 141

We say that this algebra is defined by the generators xi and relations fj . Indeed the
elements xi , i ∈ I, together with 1 generate the algebra, and, unlike the generators ξi of
the free algebra FX, they are not independent but related through the polynomials
fj. These may not be the only polynomials
 that relate the xi ’s. However, if f =
f ξi1 , . . . , ξin ∈ FX is such that f xi1 , . . . , xin = 0, i.e.,
   
f ξi1 , . . . , ξin + (R) = f ξi1 + (R), . . . , ξin + (R) = 0,

then f lies in the ideal of FX generated by the fj ’s.

Example 6.3 Let A := Fx, y | x 2 = 0, y2 = 0, xy + yx = 1. A concrete example

of a pair of elements satisfying these relations are the matrix units E12 and E21 in
M2 (F). Let us show that A is actually isomorphic to M2 (F). We begin by noticing that
xyx = x and yxy = y. Hence it follows that every element in A is a linear combination
of x, y, xy, and yx. Suppose λ1 x + λ2 y + λ3 xy + λ4 yx = 0 for some λi ∈ F. Then the
polynomial f (ξ, η) = λ1 ξ + λ2 η + λ3 ξ η + λ4 ηξ lies in the ideal of Fξ, η generated
by R = {ξ 2 , η2 , ξ η + ηξ − 1}. Accordingly, for every F-algebra B and every pair of
elements a, b ∈ B, the evaluation f (a, b) lies in the ideal of B generated by the set
{a2 , b2 , ab + ba − 1}. Choosing B = M2 (F) and a = E12 , b = E21 , we see that each
λi = 0. Consequently, {x, y, xy, yx} is a basis of A. Note that its multiplication table
is the same as that of {E12 , E21 , E11 , E22 } in M2 (F). Therefore A ∼
= M2 (F).

Example 6.4 Assume that char(F) = 0. Let us show that the first Weyl algebra A1
over F, introduced in Example 1.13, is isomorphic to A := Fx, y | [x, y] = 1. By
the universal property of Fξ, η we can define the unital homomorphism f (ξ, η) →
f (D, L) from Fξ, η onto A1 . Since R := {[ξ, η] − 1} is contained in its kernel,
this homomorphism gives rise to the homomorphism f (x, y) → f (D, L) from A =
Fξ, η/(R) onto A1 . We have to show that this is an isomorphism. To this end, we
refer to the proof of the assertion (a) from Example 1.13. It is evident that it depends
only on the formula [D, L] = I. Therefore the same conclusion holds for every unital
algebra generated by a pair of elements whose commutator equals 1. Accordingly,
every element in A can be written as a linear combination of the elements ym x n ,
m, n ≥ 0. Using (b) from Example 1.13 we now see that our homomorphism is
injective, and hence an isomorphism.
Now take an arbitrary n ∈ N. The nth Weyl algebra, An , is defined as the sub-
algebra of EndF (F[ω1 , . . . , ωn ]), the algebra of all linear operators of F[ω1 , . . . , ωn ],
generated by the operators Di , Li , i = 1, . . . , n, given by

  ∂f  
Di f (ω1 , . . . , ωn ) = (ω1 , . . . , ωn ), Li f (ω1 , . . . , ωn ) = ωi f (ω1 , . . . , ωn ).
∂ωi

One can show that An is isomorphic to

Fxi , yi , i = 1, . . . , n | [xi , yj ] = δij , [xi , xj ] = 0, [yi , yj ] = 0, i, j = 1, . . . , n,

142 6 Noncommutative Polynomials

and that An is, just as A1 (cf. Examples 1.13, 2.28, and 3.68), a simple noetherian
domain. Proving all this for an arbitrary n is not much harder than for n = 1.
To provoke the reader’s curiosity we mention a problem by J. Dixmier from 1968.

Problem 6.5 Is every nonzero algebra endomorphism of An surjective?

Recently it has turned out that this problem, to which one usually refers to as
Dixmier’s conjecture, is equivalent to the famous Jacobian conjecture from alge-
braic geometry. We omit stating this conjecture here; the interested reader will easily
find information about it at many places.

Example 6.6 The algebra A := Fxi , i ∈ I | [xi , xj ] = 0, i, j ∈ I is generated by

the elements xi which commute with each other, and these commuting relations are
the only defining relations. It should come as no surprise that A is the same thing as
F[ωi , i ∈ I], the algebra of polynomials in commuting indeterminates ωi . To give a
formal proof, we use the universal property to define the unital homomorphism from
FX, X = {ξi | i ∈ I}, onto F[ωi , i ∈ I], determined by ξi → ωi . Its kernel contains
the set R := {ξi ξj − ξj ξi | i, j ∈ I}, and every element in A = FX/(R) can be written
as f + (R) where f is such that in each of its monomials the indeterminates ξi appear
in the same fixed order. This readily implies that A ∼ = F[ωi , i ∈ I].

Example 6.7 The Grassmann algebra (or the exterior algebra) over a field F with
char(F) = 2 is defined as

G := Fxi , i ∈ N | xi2 = 0, xi xj + xj xi = 0, i, j ∈ N.

Hence we have, for example, x1 x2 x1 = −x12 x2 = 0; in general, xi Gxi = 0. Note also

that every product of different xi ’s can be written as plus or minus

xi1 xi2 . . . xin , i1 < i2 < · · · < in . (6.1)

Accordingly, these elements, together with 1, linearly span G. Let us show that
each of them is nonzero. It suffices to show that x1 x2 . . . xn = 0. If this was not
true, then ξ1 ξ2 . . . ξn would be contained in the ideal of Fξ1 , ξ2 , . . .  generated by
R = {ξi2 , ξi ξj + ξj ξi | i, j ∈ N}. Then we could write ξ1 ξ2 . . . ξn as a sum of terms of
the form mξi2 m and p(ξi ξj + ξj ξi )p , where m, m , p, p are monomials. The linear
span of all multilinear monomials, i.e., the space of all multilinear polynomials, has
trivial intersection with the linear span of all monomials that are not multilinear.
Hence it follows that ξ1 ξ2 . . . ξn is a linear combination of polynomials
 
ξσ (1) . . . ξσ ( j−1) ξσ ( j) ξσ ( j+1) + ξσ ( j+1) ξσ ( j) ξσ ( j+2) . . . ξσ (n)

where σ ∈ Sn . Such a linear combination is a multilinear polynomial with the pro-

perty that the sum of its coefficients corresponding to even permutations coincides
with the sum of its coefficients corresponding to odd permutations. Since ξ1 ξ2 . . . ξn
6.2 Algebras Defined by Generators and Relations 143

does not share this property, we have arrived at a contradiction. Thus, the elements
from (6.1) areindeed nonzero.
Suppose ni=1 λi bi = 0 where λi ∈ F\{0} and each bi is either 1 or is as in
(6.1). We mayassume that there exists k ∈ N such that xk appears in bn , but not
in b1 . Hence n−1 i=1 λi bi xk = 0 and b1 xk  = 0. Continuing this procedure we arrive
at a contradiction that one of the λi ’s is zero. Thus, the elements of the form (6.1),
together with 1, form a basis of G.
It is obvious that x1 x2 commutes with each xi , and hence it lies in Z(G), the center
of G. The same reasoning shows that every element from (6.1) with n even lies in
Z(G). Moreover, using the assumption that char(F) = 2 it is easy to see that Z(G) is
actually the linear span of all such elements and 1. We will write G0 for Z(G), and
similarly, we will write G1 for the linear span of all elements of the form (6.1) with
n odd. Note that G = G0 ⊕ G1 , the vector space direct sum.
One often considers the Grassmann algebra in n generators. We can describe it as
the subalgebra of G generated by 1, x1 , . . . , xn . By counting the basis elements we
see that its dimension is 2n .
The Grassmann algebra has a big center and plenty of nilpotent ideals, in contrast
to algebras that we prefer in structure theory. However, G has turned out to be
an extremely useful tool in various mathematical areas. A small evidence of its
applicability will be given in the proof of the Amitsur-Levitzki Theorem in Sect. 6.9.

6.3 Alternating Polynomials

In this section we consider special polynomials that play an important role in the
theory of polynomial identities, the theme of the subsequent sections. For simplicity
we will confine ourselves to multilinear polynomials.
Let us begin with an example. Consider the polynomial

h(ξ1 , ξ2 , ξ3 , η) = ξ1 ηξ2 ξ3 − ξ1 ηξ3 ξ2 + ξ2 ηξ3 ξ1 − ξ2 ηξ1 ξ3 + ξ3 ηξ1 ξ2 − ξ3 ηξ2 ξ1 .

Note that by substituting ξ1 for ξ2 we get 0, i.e., h(ξ1 , ξ1 , ξ3 , η) = 0. Similarly,

h(ξ1 , ξ2 , ξ1 , η) = 0 and h(ξ1 , ξ2 , ξ2 , η) = 0.
Definition 6.8 A multilinear polynomial f = f (ξ1 , . . . , ξn , η1 , . . . , ηr ) ∈ FX is
said to be alternating in ξ1 , . . . , ξn if f becomes zero whenever one substitutes ξj
for ξi , 1 ≤ i < j ≤ n.
Replacing ξ1 by ξ1 + ξ2 in f (ξ1 , ξ1 , ξ3 , . . . , ξn , η1 , . . . , ηr ) = 0 we get

f (ξ1 , ξ2 , ξ3 , . . . , ξn , η1 , . . . , ηr ) = −f (ξ2 , ξ1 , ξ3 , . . . , ξn , η1 , . . . , ηr ).

Similarly, f changes sign if we switch any pair of indeterminates ξi and ξj . The

name “alternating” comes from this condition, which is obviously equivalent to the
definition provided that char(F) = 2.
144 6 Noncommutative Polynomials

Our main reason for considering alternating polynomials is their connection to

the notion of linear dependence.

Lemma 6.9 Let A be an F-algebra and let a1 , . . . , an ∈ A be linearly dependent

elements. If a multilinear polynomial f = f (ξ1 , . . . , ξn , η1 , . . . , ηr ) ∈ FX is alter-
nating in ξ1 , . . . , ξn , then f (a1 , . . . , an , x1 , . . . , xr ) = 0 for all x1 , . . . , xr ∈ A.

Proof We may assume that, say, an = n−1 i=1 λi ai for some λi ∈ F. Consequently,


n−1
f (a1 , . . . , an , x1 , . . . , xr ) = λi f (a1 , . . . , an−1 , ai , x1 , . . . , xr ) = 0,
i=1

since f (ξ1 , . . . , ξi , . . . , ξn−1 , ξi , η1 , . . . , ηr ) = 0 for i = 1, . . . , n − 1. 

Two families of alternating polynomials are of special importance. The first one
consists of polynomials that are alternating in all indeterminates (i.e., there are no
ηj ’s).

Definition 6.10 Let n ≥ 2. The polynomial


sn = sn (ξ1 , . . . , ξn ) := sgn(σ )ξσ (1) . . . ξσ (n)
σ ∈Sn

is called the standard polynomial of degree n.

The reader’s first thought may have been that this is similar to the definition of the
determinant of a matrix. There are indeed several analogies.
Choose indices i, j and a permutation σ . Write ξσ (1) . . . ξσ (n) as mξi m ξj m where
m, m , m are monomials in the remaining indeterminates. Then the term mξj m ξi m
corresponds to a permutation whose sign is opposite to that of σ . This readily implies
that sn is alternating.
Note that s2 = ξ1 ξ2 − ξ2 ξ1 , and s3 = h(ξ1 , ξ2 , ξ3 , 1) where h is the above
polynomial (substituting 1 for η simply means that one can “erase” η in the formula
defining h).
The second family is similar, but it involves the ηj ’s.

Definition 6.11 Let n ≥ 2. The polynomial


cn = cn (ξ1 , . . . , ξn , η1 , . . . , ηn−1 ) := sgn(σ )ξσ (1) η1 ξσ (2) η2 . . . ηn−1 ξσ (n)
σ ∈Sn

is called the nth Capelli polynomial.

(Warning: some authors use slightly different definitions.)

Thus, c2 = ξ1 η1 ξ2 − ξ2 η1 ξ1 , and one gets c3 by writing η1 for η and inserting η2

between the last two factors of each term of h. Note that
6.3 Alternating Polynomials 145

sn (ξ1 , . . . , ξn ) = cn (ξ1 , . . . , ξn , 1, . . . , 1).

It is easy to see that cn is alternating in ξ1 , . . . , ξn . Let us also mention the easily

derived formula


n
cn = (−1)i−1 ξi η1 cn−1 (ξ1 , . . . , ξi−1 , ξi+1 , . . . , ξn , η2 , . . . , ηn−1 ), (6.2)
i=1

which holds for every n ≥ 3. Similarly, for every n ≥ 3 we have


n
sn = (−1)i−1 ξi sn−1 (ξ1 , . . . , ξi−1 , ξi+1 , . . . , ξn ). (6.3)
i=1

6.4 Polynomial Identities: Definition and Examples

Much of the rest of the book is devoted to the following notion.

Definition 6.12 A polynomial f = f (ξ1 , . . . , ξn ) ∈ Fξ1 , ξ2 , . . .  is said to be a

polynomial identity (or simply an identity) of an F-algebra A if f (x1 , . . . , xn ) = 0
for all x1 , . . . , xn ∈ A. In this case we also say that A satisfies f .

The reader should not confuse this with algebras defined by generators and relations,
where polynomials vanish only on some of the generators. The condition that f is an
identity of A is quite restrictive, provided, of course, that f is a nonzero polynomial.

Definition 6.13 If a nonzero polynomial is a polynomial identity of A, then A is

called a PI-algebra.

There are two basic examples of PI-algebras.

Example 6.14 Saying that A is commutative is the same as saying that A satisfies the
identity [ξ1 , ξ2 ] (= ξ1 ξ2 − ξ2 ξ1 ). Every commutative algebra is thus a PI-algebra.

This suggests one way of looking at the notion of a polynomial identity, i.e., as a
generalization of commutativity. The second basic example is of a different nature.

Example 6.15 Every finite dimensional algebra is a PI-algebra. If [A : F] = d, then

every multilinear polynomial that is alternating in d + 1 indeterminates is an identity
of A by Lemma 6.9. In particular, A satisfies sd+1 .

Thus, Mn (F) is a PI-algebra. In fact, a considerable part of the theory of poly-

nomial identities is centered around matrices. Roughly speaking, the theory has two
branches, combinatorial and structural. The goal of the first one is understanding
of the set of all identities of a given algebra A, with a special emphasis on the case
146 6 Noncommutative Polynomials

where A = Mn (F). Concerning the second one, let us only mention at this point that
a prime PI-algebra can be embedded, in a particularly nice way, into a matrix algebra
over a field extension of the base field. This indicative result is a byproduct of the
structure theory which will be exposed in the next chapter (see Corollary 7.59 and
Remark 7.60). The combinatorial aspect will be briefly touched upon in the rest of
this chapter.
There are other examples of PI-algebras, quite different from commutative and
finite dimensional ones.

Example 6.16 The commutator [x, y] of any two elements x, y in the Grassmann
algebra G lies in Z(G) = G0 . It is enough to show this for basis elements x =
xi1 . . . xin and y = xj1 . . . xjm . Note that [x, y] can be nonzero only if both n and m are
odd, and in this case it indeed lies in G0 . Thus, G satisfies the identity [[ξ1 , ξ2 ], ξ3 ].
On the other hand, if char(F) = 0, then none of the standard polynomials is an
identity of G. To show this, we first observe that

xσ (1) . . . xσ (r) = sgn(σ )x1 . . . xr (6.4)

holds for every permutation σ on {1, . . . , r}. By writing σ as a product of trans-

positions we see that it is enough to prove (6.4) for the case where σ itself
is a transposition, and in this case the formula is pretty obvious. Consequently,
sr (x1 , . . . , xr ) = r!x1 . . . xr = 0 for every r ≥ 2. From Example 6.15 we therefore
infer that G cannot be embedded into any algebra that is finite dimensional over
some field extension of F; in fact, it cannot even be embedded into Mn (C) for any
commutative algebra C (see Theorem 6.39 below).

Example 6.17 If A is a nilpotent algebra, then A satisfies the identity ξ1 . . . ξn for

some n ∈ N. A concrete example is the algebra of all strictly upper triangular matrices
over a commutative algebra C. Now consider Tn (C), the algebra of all upper triangular
n × n matrices over C. Note that the commutator of any two matrices from Tn (C) is
a strictly upper triangular matrix. Therefore Tn (C) satisfies [ξ1 , η1 ] . . . [ξn , ηn ].

Example 6.18 Suppose ξ 2 is an identity of A. This does not necessarily mean that
A2 = 0 (for example, consider the 3-dimensional subalgebra Fx1 + Fx2 + Fx1 x2
of the Grassmann algebra G). However, A3 = 0, provided that char(F) = 2. The
proof is easy. We have x 2 = 0 for every x ∈ A. Replacing x by x + y, and using
x 2 = y2 = 0, we get xy + yx = 0 for all x, y ∈ A (i.e., ξ η + ηξ is an identity of A).
Consequently, for all x, y, z ∈ A we have

(xy)z = −z(xy) = −(zx)y = y(zx) = (yz)x = −x(yz),

yielding xyz = 0.
Much more can be said. If char(F) = 0 and A satisfies the identity ξ n , then A is a
nilpotent algebra, i.e., A satisfies ξ1 ξ2 . . . ξd for some d, depending on n. This result
is called the Nagata-Higman Theorem. The proof uses elementary tools, but we
6.4 Polynomial Identities: Definition and Examples 147

omit it (it can be found, for example, in [ZSSS82]). Finding the exact lower bound
for d is an open problem.
Example 6.19 Let n ≥ 2. Theorem 5.78 implies that an algebra satisfying the identity
ξ n − ξ is commutative. Concrete examples are Boolean rings and finite fields. Of
course, Theorem 5.78 says more than that. However, the condition x n(x) = x cannot
be interpreted in terms of polynomial identities.
How to obtain new PI-algebras from the old ones? In the next two examples
we give some answers to this question. The first example, as trivial as it is, is of
fundamental importance.
Example 6.20 If f is an identity of an algebra A, then f is also an identity of every
subalgebra of A as well as of every homomorphic image of A (equivalently, of A/I
for every ideal I of A). In particular, a subalgebra or a homomorphic image of a
PI-algebra is again a PI-algebra.
Example 6.21 Let {Ai | i ∈ I} be a family of algebras. If each Ai satisfies the same
identity f , then so does the direct product Πi∈I Ai . On the other hand, if each Ai is
a PI-algebra, then there is no reason to believe that Πi∈I Ai is a PI-algebra too. If,
however, the set I is finite, say I = {1, . . . , n}, then this is true. Namely, if fi is an
identity of Ai , then f1 . . . fn is an identity of A1 ×· · ·×An . Not only the direct product,
even the tensor product of two PI-algebras is again a PI-algebra. But this is anything
but obvious. It was proved by A. Regev in the early 1970s.
All these examples may give the wrong impression that most algebras are PI-
algebras. Infinite dimensional algebras with small centers, e.g., infinite dimensional
central algebras, are actually only rarely PI-algebras. This will become clear as the
theory unfolds. Let us mention, at this point, just an obvious example of an algebra
without nonzero identities.
Example 6.22 The free algebra Fξ1 , ξ2 , . . .  is not a PI-algebra.
One can define polynomial identities for rings, not only for algebras. We will in
fact deal with PI-rings in the next chapter. In this one, however, we consider only
algebras in order to make the exposition simpler.

6.5 Linearization

The purpose of this section is to get familiar with the process of linearization.
The concept of this process is applicable to various situations throughout mathemat-
ics. Our goal, however, is to show that it can be used to recast an arbitrary polynomial
identity into a multilinear one.
In Example 6.18 we have already met a simple example of a linearization: If A
satisfies ξ 2 , then it also satisfies the multilinear identity ξ η + ηξ . The proof is very
easy, one just replaces an arbitrary element by the sum of two arbitrary elements.
This simple idea, which is the essence of “linearization”, can be pushed further.
148 6 Noncommutative Polynomials

Example 6.23 Consider a slightly more complicated situation where A satisfies

f (ξ ) = ξ 3 . All relevant calculations can be made in a free algebra, so we can ignore
A for a while. Let us define two new polynomials g = g(ξ1 , ξ2 ) and h = h(ξ1 , ξ2 , ξ3 )
as follows:

g := f (ξ1 + ξ2 ) − f (ξ1 ) − f (ξ2 ),

h := g(ξ1 , ξ2 + ξ3 ) − g(ξ1 , ξ2 ) − g(ξ1 , ξ3 ).

Of course, h can be explicitly expressed by f , i.e.,

h = f (ξ1 + ξ2 + ξ3 ) − f (ξ1 + ξ2 ) − f (ξ1 + ξ3 ) − f (ξ2 + ξ3 ) + f (ξ1 ) + f (ξ2 ) + f (ξ3 ).

However, to compute h it is easier to first compute g,

g = ξ1 ξ22 + ξ2 ξ1 ξ2 + ξ22 ξ1 + ξ12 ξ2 + ξ1 ξ2 ξ1 + ξ2 ξ12 ,

and hence derive


h= ξσ (1) ξσ (2) ξσ (3) .
σ ∈S3

Since A satisfies f , it also satisfies h. We have thus found a multilinear polynomial

identity of A, but at the cost of increasing the number of indeterminates.

To consider arbitrary polynomials, we first introduce an auxiliary notion. The

degree of a polynomial f in ξi is the maximal number of occurrences of ξi in the
monomials of f . For example, f = ξ12 ξ2 ξ3 ξ1 + ξ2 ξ3 ξ2 has degree 3 in ξ1 , degree 2 in
ξ2 , and degree 1 in ξ3 .

Theorem 6.24 If an algebra A satisfies a nonzero polynomial identity, then A also

satisfies a nonzero multilinear polynomial identity of the same or lower degree.

Proof Let f = f (ξ1 , . . . , ξn ) be a nonzero identity of A. Denote by di the degree of

f in ξi . The proof is by induction on d := max{d1 , . . . , dn } > 0.
If d = 1, then each ξi appears in every monomial of f at most once—“at most”
and not “exactly”, so f is not necessarily multilinear. But this can be easily remedied.
If, say, λξ1 . . . ξm , where λ is a nonzero scalar, is a monomial of f of minimal degree,
then we just substitute 0 for the remaining indeterminates, and hence obtain a nonzero
multilinear identity f (ξ1 , . . . , ξm , 0, . . . , 0) of degree less or equal to deg( f ).
Let d > 1. Without loss of generality we may assume that there exists k ≤ n
such that dk = · · · = dn = d and di < d for i < k. Define a new polynomial which
involves an additional indeterminate, g = g(ξ1 , . . . , ξn , ξn+1 ), by

g := f (ξ1 , . . . , ξn−1 , ξn + ξn+1 ) − f (ξ1 , . . . , ξn−1 , ξn ) − f (ξ1 , . . . , ξn−1 , ξn+1 ).

6.5 Linearization 149


Note that g is also an identity of A. Let us write f = i λi wi , where the wi ’s are
 different words and the λi ’s are nonzero scalars. Then g can be written as
pairwise
g = i λi gi , where gi is obtained from wi in the same way as g is obtained from f .
If ξn does not occur in wi , then gi = −wi . If ξn occurs only once in wi , then gi = 0.
If ξn occurs at least twice in wi , then gi is the sum of all possible words obtained by
replacing at least one, but not all, of the ξn ’s in wi by ξn+1 . Thus, if in any of these
words one substitutes ξn for ξn+1 , then one gets back the word wi . Therefore the
words appearing in gi are different from the words appearing in gi whenever i = i.
As d > 1, the indices i such that ξn occurs at least twice in wi do exist. This shows
that g = 0.
The following conclusions can be drawn from the preceding paragraph:
• g is a nonzero identity of A.
• deg(g) ≤ deg( f ).
• For j = 1, . . . , n − 1, the degree of g in ξj is less or equal to dj .
• The degree of g in ξn and ξn+1 is d − 1.
Now repeat this procedure, first with g in place of f and ξn−1 in place of ξn , and then
with other indeterminates down to ξk . At the end we arrive at a situation where a
nonzero identity has degree at most d − 1 in every indeterminate. 

This theorem is of crucial importance if one is interested in the structure of an

algebra satisfying some nonzero polynomial identity. With regard to particular iden-
tities, it should be remarked that in the finite characteristic case some information
can be lost when performing the linearization process. Take our first example where
an F-algebra A satisfies ξ 2 . Linearizing we get that it also satisfies ξ η + ηξ . If
char(F) = 2, then the two identities are equivalent, each one implies the other one.
If char(F) = 2, then the multilinear one only tells us that the algebra is commutative,
while the original one shows much more, every element has square zero.

6.6 Stable Identities

Let A be an F-algebra and let f ∈ FX be an identity of A. Take an extension field

K of F. Is f an identity of AK = K ⊗ A, the scalar extension of A to K? Since AK has
the same multiplication table as A, an affirmative answer seems plausible. However,
the next simple example shows that this may not be true even for A = F.

Example 6.25 Let F be a finite field with n elements. Then f = ξ n − ξ is an identity

of F (see Lemma 5.76). Since f cannot have more than n zeros in any field, f is not
an identity of any extension field K of F that properly contains F.

The main goal of this section is to show that the above question does have an
affirmative answer if F is infinite. Moreover, we may replace the role of the field K
by an arbitrary commutative algebra C.
150 6 Noncommutative Polynomials

Definition 6.26 An identity f of an F-algebra A is said to be stable (for A) if f is an

identity of C ⊗ A for every commutative F-algebra C.

Example 6.25 thus shows that ξ n − ξ is not a stable identity for A = F if |F| = n.
Let us turn to positive results.

Lemma 6.27 Every multilinear identity is stable.

Proof Let f = f (ξ1 , . . . , ξn ) be a multilinear polynomial, and consider, for arbitrary

A and C, f (x 1 , . . . , x n ) with x i ∈ C ⊗ A. Each x i is a sum of simple tensors, and so,
by multilinearity of f , f (x 1 , . . . , x n ) can be written as a sum of elements of the form
f (c1 ⊗ x1 , . . . , cn ⊗ xn ). Since C is commutative we have

f (c1 ⊗ x1 , . . . , cn ⊗ xn ) = c1 . . . cn ⊗ f (x1 , . . . , xn ),

which is 0 if f is an identity of A. 

The classical polynomials in commuting indeterminates play a similar role in

commutative algebra as the noncommutative polynomials do in noncommutative
algebra. The algebra of commutative polynomials F[Ω] satisfies the universal prop-
erty in the context of commutative unital algebras, the evaluation of a commutative
polynomial in elements from a commutative algebra can be defined in the obvious
way, and every commutative unital algebra is a homomorphic image of F[Ω] for
some set Ω. A detailed explanation of all this is essentially the same as for FX
in Sects. 6.1 and 6.2. At any rate, these are just elementary observations—but very
useful ones, as we shall see in the rest of the chapter.

Lemma 6.28 Let f = f (ξ1 , . . . , ξn ) be an identity of an F-algebra A. If f is an

identity of F[ω1 , . . . , ωs ] ⊗ A for every s ∈ N, then f is stable for A.

Proof Let C be a commutative unital F-algebra. Since every commutative algebra

is a subalgebra of a commutative unital algebra (say, of its unitization),
 i it suffices
to show that f is an identity of C ⊗ A. Take arbitrary elements x i = m j=1 cij ⊗ xij ,
i = 1, . . . , n, in C ⊗ A. Let us set


mi
x̃i := ωij ⊗ xij ∈ F[Ω] ⊗ A,
j=1

where Ω = {ωij | i = 1, . . . , n, j = 1, . . . , mi }. By assumption, f (x̃1 , . . . , x̃n ) = 0.

Therefore, if ϕ is the (unital) homomorphism
 from
 F[Ω] into C sending ωij into cij ,
then f (x 1 , . . . , x n ) = (ϕ ⊗ idA ) f (x̃1 , . . . , x̃n ) = 0. 

Let us warn the reader that henceforth our argumentation will be somewhat less
formal than in the proof of Lemma 6.28. We will avoid speaking about the univer-
sal property and homomorphisms from F[Ω], but simply substitute elements from
commutative algebras for indeterminates in formulas involving polynomials.
6.6 Stable Identities 151

Theorem 6.29 Let F be an infinite field. Then every polynomial identity of an arbi-
trary F-algebra A is stable.

Proof Let f =f (ξ1 , . . . , ξn ) be an identity of A. Pick s ∈ N and take arbitrary

elements x̂i = j pij ⊗ xij in F[ω1 , . . . , ωs ] ⊗ A, i = 1, . . . , n. By Lemma 6.28 it is
enough to show that f (x̂1 , . . . , x̂n ) = 0. Suppose this is not true. Then we can write

f (x̂1 , . . . , x̂n ) = qk ⊗ ak , (6.5)
k

where q1 = 0 and the ak ’s are linearly independent (cf. Remark 4.15). Since F is
infinite, there exist λ1 , . . . , λs ∈ F such that q1 (λ1 , . . . , λs ) = 0 (the proof is a
straightforward induction on s; for s = 1 just use the fact that the number of zeros
of a polynomial cannot exceed its degree). Let us set αij := pij (λ1 , . . . , λs ) and
βk := qk (λ1 , . . . , λs ); thus β1 = 0. Substituting λi for ωi in (6.5) we get
   
f α1j ⊗ x1j , . . . , αnj ⊗ xnj = βk ⊗ ak .
j j k

   
Since αij , βk ∈ F, we have j αij ⊗ xij = 1 ⊗ α
j ij ij and
x k βk ⊗ ak =
 
1⊗ k βk ak . The above identity can be therefore rewritten as

  
1⊗f α1j x1j , . . . , αnj xnj =1⊗ βk ak ,
j j k

which is a contradiction. Namely, the left-hand side is 0 since f is an iden-

tity of A, while the right-hand side is not 0 since β1 = 0 and the ak ’s are
linearly independent. 

Theorem 6.29 can be proved in a more direct manner, avoiding the polynomial
algebra. But we wanted to point out the general principle. If one wishes to prove a
result for general commutative algebras, then it is sometimes enough to prove it for
polynomial algebras. This approach will be used a few more times in the rest of the
chapter.
Returning to scalar extensions, we now see that, as long as F is an infinite field, an
F-algebra A and its scalar extension AK satisfy the same identities (with coefficients
in F). Indeed, an identity of AK is also an identity of A since we can embed A into
AK via x → 1 ⊗ x, and an identity of A is also an identity of AK by Theorem 6.29.
For instance, from Example 2.72 we now see that H and M2 (C) satisfy the same
identities (with real coefficients). More generally, by making use of splitting fields,
the problem of determining identities of a finite dimensional central simple F-algebra
A can be reduced to the case where A = Mn (K). This is true even when F is finite
since then A is already isomorphic to Mn (K) by Corollary 2.71.
152 6 Noncommutative Polynomials

6.7 T-ideals

Throughout this section, X denotes the countably infinite set {ξ1 , ξ2 , . . . } of non-
commuting indeterminates. By Id(A) we denote the set of all polynomial identities
of an F-algebra A. Obviously, Id(A) is an ideal of the free algebra FX. One can
say more about it, as we shall soon see. We restrict our discussion to nonzero unital
algebras, so that evaluations make sense also for polynomials with nonzero constant
term. This will simplify the exposition.

Definition 6.30 An ideal I of FX is called a T-ideal if ϕ(I) ⊆ I for every algebra
endomorphism ϕ of FX.

Remark 6.31 If ϕ is an endomorphism of FX and gi := ϕ(ξi ), i ∈ N, then

 
ϕ f (ξ1 , . . . , ξn ) = f (g1 , . . . , gn )

for every f ∈ FX. On the other hand, for any gi ∈ FX, i ∈ N, there exists
an endomorphism ϕ of FX satisfying ϕ(ξi ) = gi . Thus, an ideal I of FX is a
T-ideal if and only if for all f = f (ξ1 , . . . , ξn ) ∈ I and g1 , . . . , gn ∈ FX we have
f (g1 , . . . , gn ) ∈ I.

If A is a nonzero unital algebra, then Id(A) is obviously a proper T-ideal of FX.

The next lemma in particular shows that the converse is also true. That is, for every
proper T -ideal I of FX there exists a nonzero unital F-algebra A such that I = Id(A).

Lemma 6.32 If I is a proper T-ideal of FX, then I = Id(FX/I).

Proof Take an arbitrary polynomial f = f (ξ1 , . . . , ξn ) ∈ FX, and consider its

evaluations on FX/I. Note that

f (g1 + I, . . . , gn + I) = f (g1 , . . . , gn ) + I

for all g1 , . . . , gn ∈ FX. Using Remark 6.31 we thus see that f ∈ I if and only if f
is an identity of FX/I. 

Assume temporarily that I is just a nonempty subset of FX. The variety deter-
mined by I, denoted by V (I), is the class of all unital algebras A suchthat I ⊆ Id(A);
that is, every polynomial from I is an identity of A. For example, V {ξ1 ξ2 − ξ2 ξ1 }
is the class of all commutative unital algebras. Note that the variety determined by
the set I coincides with the variety determined by the T-ideal generated by I. There-
fore, assume from now on that I itself is a T-ideal; to avoid the trivial situation,
assume also that I = FX. Lemma 6.32 shows that the algebra FX/I belongs to
V (I). Moreover, FX/I is a distinguished representative of the class V (I) due to
the following property. If A ∈ V (I) and {a1 , a2 , . . . } is a countably infinite subset
of A, then there exists a unique unital homomorphism ϕ : FX/I → A such that
ϕ(ξi + I) = ai , i = 1, 2, . . . . Indeed, just take the unital homomorphism from FX
6.7 T-ideals 153

to A determined by ξi → ai , note that I lies in its kernel since A ∈ V (I), and the
existence of ϕ follows. The uniqueness is obvious. Because of this property, we call
FX/I the relatively free algebra (with respect to I).
To each unital algebra A we can assign the T-ideal Id(A) and the corresponding
relatively free algebra FX/Id(A).

Example 6.33 If A is not a PI-algebra, then Id(A) = {0} and so the corresponding
relatively free algebra is just the free algebra FX.

Example 6.34 Let

A = Mn (F). We begin by introducing  a seemingly unrelated
(i)
notion. Let Ω = ωjk | j, k = 1, . . . , n, i = 1, 2, . . . be a set of commuting inde-
terminates. For each i we set
⎡ (i) (i) (i) ⎤
ω11 ω12 . . . ω1n
⎢ (i) (i) (i) ⎥
⎢ω21 ω22 . . . ω2n ⎥  
g (i)
=⎢
⎢ .. .. . .
⎥
.. ⎥ ∈ Mn F[Ω] .
⎣ . . . . ⎦
(i) (i) (i)
ωn1 ωn2 . . . ωnn

We call g(i) an n × n generic matrix. Thus, the entries of a generic matrix are distinct
(independent) commuting indeterminates. The subalgebra of Mn (F[Ω]) generated
by all n × n generic matrices g(i) , i ∈ N, and the identity matrix is called the algebra
of n × n generic matrices. It will be denoted by GMn (F). Assume from now on that
F is an infinite field. We claim that

FX/Id(Mn (F)) ∼
= GMn (F),

i.e., the relatively free algebra corresponding to Mn (F) is isomorphic to the algebra
of n×n generic matrices. To prove this, take an identity f of Mn (F). Since F  is infinite,

Theorem 6.29 tells us that f is also an identity of F[Ω] ⊗ Mn (F) ∼ = Mn F[Ω]  (cf.

Example 4.22). But then f is also an identity of the subalgebra GMn (F) of Mn F[Ω] .
Accordingly, GMn (F) ∈ V (I) where I = Id(Mn (F)). As observed above, this
implies that there exists a unital homomorphism ϕ : FX/I → GMn (F) such that
ϕ(ξi +I) = g(i) , i = 1, 2, . . . . It is clear that ϕ is surjective. Let us
 prove that it is also
injective. Take f = f (ξ1 , . . . , ξn ) such that ϕ( f + I) = 0, i.e., f g(1) , . . . , g(n) = 0.
Since the entries of the matrices g(i) are independent commuting indeterminates, we
can replace them by arbitrary elements in F (cf. the previous section). This shows
that f ∈ I. Therefore f is injective, and hence an isomorphism.

6.8 The Characteristic Polynomial

This section may appear as an intruder in the chapter. The main reason for its inclu-
sion is that it is needed in the next section for the proof of the Amitsur-Levitzki
Theorem. On the other hand, although the results of this section do not directly
154 6 Noncommutative Polynomials

involve polynomial identities, they consider some other kind of identities that have
many interactions with them. At any rate, the topics of this section are of general
interest and importance.
Some of the readers may have already met what we are about to say in a linear
algebra course, yet probably at a lower level of generality. A typical course considers
only matrices with entries from a field. Many concepts, however, still make sense if a
field is replaced by a more general ring. Say, the determinant of a matrix A ∈ Mn (C),
which we denote by det(A), can be defined in the usual way as long as C is a
commutative ring. Accordingly, we can define the characteristic polynomial of
A as

pA (ω) := det(A − ωI) ∈ C[ω].

For any polynomial q ∈ C[ω] we define q(A) in the same way as when C is a field.
The Cayley-Hamilton Theorem says that

pA (A) = 0.

Quite possibly the reader is familiar with this theorem only in the case where C is a
field. However, from many of its proofs it is evident that C can be any commutative
ring. On the other hand, the Cayley-Hamilton theorem for fields easily implies the
Cayley-Hamilton Theorem for commutative rings. Indeed, assuming the truth of the
theorem for fields, it follows that it is also true for commutative domains since they
can be embedded into their fields of quotients. In particular, the theorem holds for
the ring Z[ω11 , ω12 , . . . , ωnn ]. Thus, if g = (ωij ) is a generic matrix, then pg (g) = 0.
Now take any commutative ring C, choose cij ∈ C, 1 ≤ i, j ≤ n, substitute cij for
ωij in this identity, and the validity of the Cayley-Hamilton Theorem for C follows.
For any matrix A ∈ Mn (C) we define tr(A), the trace of A, in the usual way. Our
goal is to express the coefficients of pA through the traces of the powers of A. To
this end, we need formulas on symmetric polynomials in commuting indeterminates
with integer coefficients, which were discovered already by Isaac Newton. Recall
that a polynomial is said to be symmetric if it remains unchanged after permuting
the indeterminates. Important examples are the so-called elementary symmetric
polynomials, defined by e0 := 1 and

ek = ek (ω1 , . . . , ωn ) := ωi1 ωi2 . . . ωik ,
1≤i1 <i2 <···<ik ≤n

k = 1, . . . , n, i.e., ek is the sum of all distinct products of k distinct indeterminates;

in particular, e1 is the sum of all ωi ’s, and en is their product. Let ω be another
indeterminate. Note that


n
(−1)n (ω − ω1 )(ω − ω2 ) . . . (ω − ωn ) = (−1) j en−j ω j . (6.6)
j=0
6.8 The Characteristic Polynomial 155

Elementary symmetric polynomials are thus intimately connected with the roots of
polynomials. Another reason for their importance is that they generate the algebra
of all symmetric polynomials, but we will not need this fact. Our goal is to connect
the elementary symmetric polynomials with the symmetric polynomials

n
j
pj = pj (ω1 , . . . , ωn ) := ωi , j ≥ 1.
i=1

One can verify that

e1 = p1 , 2e2 = e1 p1 − p2 , 3e3 = e2 p1 − e1 p2 + p3 , (6.7)

which indicates the general rule.

Lemma 6.35 (Newton’s formulas) For k = 1, . . . , n, we have


k
kek = (−1) j−1 ek−j pj . (6.8)
j=1

Proof From (6.6) it follows that


n
j
(−1) j en−j ωi = 0, i = 1, . . . , n.
j=0

Summing up over all i we obtain (6.8) for k = n. The case where k < n follows
easily from this one, basically we are facing only a problem in notation. The one
that we are using suggests that n is fixed and k varies. While this setting is natural in
view of applications, in the proof that follows it is inconvenient to regard n as fixed.
Let us therefore write, just for the purpose of this proof, ek,n for ek and pj,n for pj .
Our goal is to show that the polynomial

k
P := kek,n − (−1) j−1 ek−j,n pj,n
j=1

is 0. Note that for any i, j ≤ k we have

ei,n (ω1 , . . . , ωk , 0, . . . , 0) = ei,k (ω1 , . . . , ωk ),

pj,n (ω1 , . . . , ωk , 0, . . . , 0) = pj,k (ω1 , . . . , ωk ).

Since (6.8) holds for n = k, it follows that P(ω1 , . . . , ωk , 0, . . . , 0) = 0. This means

that P does not contain nonzero monomials in ω1 , . . . , ωk . Similarly, by setting
zeros at other places, we see that P does not contain nonzero monomials in any set
of k indeterminates. However, from the definition of P it is obvious that its nonzero
156 6 Noncommutative Polynomials

monomials could only be of degree k, and therefore cannot involve more than k
indeterminates. Hence P = 0. 

In the characteristic zero case we can therefore express the ek ’s through the pj ’s.
Say, from (6.7) we obtain
1 2  1 
e1 = p1 , e2 = p1 − p2 , e3 = p31 − 3p1 p2 + 2p3 .
2 6
In general, for each k = 1, . . . , n there exists a polynomial qk = qk (ω1 , . . . , ωk ) ∈
Q[ω1 , ω2 , . . . ] such that

ek = qk (p1 , . . . , pk ). (6.9)

For our goals it is important to note that qk has zero constant term. We also remark that
m
if λω1m1 . . . ωk k , λ ∈ Q, mj ≥ 0, is a monomial of qk , then m1 +2m2 +· · ·+kmk = k.
For convenience we also set q0 := 1, so that (6.9) holds for k = 0, 1, . . . , n.

Theorem 6.36 If C is a commutative Q-algebra, then for every A ∈ Mn (C) we have


n
 
pA (ω) = (−1) j qn−j tr(A), tr(A2 ), . . . , tr(An−j ) ω j .
j=0

Proof A standard reasoning shows that it suffices to consider the case where A
is a generic matrix. Therefore we may assume that C = Q[ω11 , ω12 . . . , ωnn ]. In
particular, C is thus a commutative domain and as such it can be embedded into a
field (e.g., its field of quotients), which can be further embedded into an algebraically
closed fields (e.g., its algebraic closure). But then there is no loss of generality in
assuming that C itself is an algebraically closed field. Hence we have pA (ω) =
(−1)n (ω − λ1 ) . . . (ω − λn ), where the λi ’s are the eigenvalues of A. From the Jordan
normal form of A it readily follows that tr(Ai ) = pi (λ1 , . . . , λn ), i = 1, . . . , n. Using
(6.6) we now obtain

n
pA (ω) = (−1) j en−j (λ1 , . . . , λn )ω j
j=0
n
 
= (−1) j qn−j p1 (λ1 , . . . , λn ), . . . , pn−j (λ1 , . . . , λn ) ω j
j=0
n
 
= (−1) j qn−j tr(A), . . . , tr(An−j ) ω j ,
j=0

as desired. 
6.8 The Characteristic Polynomial 157

For example, for n = 2 we have

1 
pA (ω) = ω2 − tr(A)ω + tr(A)2 − tr(A2 ) (6.10)
2
 
(and hence det(A) = 21 tr(A)2 − tr(A2 ) in this case). In general, the coefficient at
ω j is a Q-linear combination of expressions of the form
 m  m
tr(A)m1 tr A2 2 . . . tr An−j n−j

where m1 + 2m2 + · · · + (n − j)mn−j = n − j.

Corollary 6.37 Let C be a commutative Q-algebra. If A ∈ Mn (C) is such that

tr(A) = tr(A2 ) = · · · = tr(An ) = 0, then An = 0.

Proof Since the polynomials qk , k = 1, . . . , n, have zero constant terms, it follows

that pA (ω) = (−1)n ωn . 

6.9 The Amitsur-Levitzki Theorem

The matrix algebra Mn (F) has dimension n2 , and thus satisfies sn2 +1 (see Example
6.15). From (6.3) we therefore see that it also satisfies sn2 +2 , sn2 +3 , etc. But does
it satisfy sk with k < n2 + 1? Moreover, what is the minimal degree of a nonzero
identity of Mn (F)? We first approach these questions from the opposite direction.

Lemma 6.38 A nonzero polynomial of degree less than 2n is not an identity of

Mn (F).

Proof By Theorem 6.24 it suffices to treat multilinear polynomials. Take a nonzero

multilinear polynomial f of degree 2n − 1. Thus, f = σ ∈S2n−1 λσ ξσ (1) . . . ξσ (2n−1)
with, say, λ1 = 0. The sequence of matrix units E11 , E12 , E22 , . . . , En−1,n , Enn has
the property that their product in the given order is E1n , while the product in any other
order is zero. Therefore f (E11 , E12 , . . . , Enn ) = λ1 E1n = 0. Similarly, by taking an
appropriate subsequence we see that a nonzero multilinear polynomial of degree less
than 2n − 1 is not an identity of Mn (F). 

Thus, the best we can hope for is that Mn (F) satisfies s2n —and this is indeed the
case! This was proved by S. A. Amitsur and J. Levitzki in 1950. The charming proof
that we are about to give is due to S. Rosset.

Theorem 6.39 (Amitsur-Levitzki) The standard polynomial s2n is an identity of

Mn (C) for every commutative algebra C.

Proof By assumption, C is an algebra over some field F. Let F0 be the prime subfield
of F, and consider C as an algebra over F0 . Since Mn (C) ∼
= C ⊗F0 Mn (F0 ) (Example
4.22), and since multilinear identities are stable (Lemma 6.27), it is enough to prove
158 6 Noncommutative Polynomials

that s2n is an identity of Mn (F0 ). Suppose this is true for F0 = Q. Then s2n is
also an identity of Mn (Z). But Mn (Zp ) is a homomorphic image of Mn (Z).  Indeed,

if ϕ is the canonical homomorphism from Z onto Zp , then (mij ) → ϕ(mij ) is
a homomorphism from Mn (Z) onto Mn (Zp ). Therefore s2n is also an identity of
Mn (Zp ). We have thus reduced the problem to proving that s2n is an identity of
Mn (Q).
Take A1 , . . . , A2n ∈ Mn (Q). We must show that s2n (A1 , A2 , . . . , A2n ) = 0. We
will accomplish this with the help of the Grassmann algebra G = G0 ⊕ G1 over Q
with generators x1 , x2 , . . . (see Example 6.7). For B = (bij ) ∈ Mn (Q) and x ∈ G we
define Bx to be the matrix (bij x) ∈ Mn (G). Note that the product (Bx)(B x ) of two
such matrices is equal to (BB )(xx ). Let us set

A := A1 x1 + A2 x2 + · · · + A2n x2n ∈ Mn (G).

From xi Gxi = 0 we easily infer that


A2n = Aσ (1) Aσ (2) . . . Aσ (2n) xσ (1) xσ (2) . . . xσ (2n) .
σ ∈S2n

Recall that xσ (1) . . . xσ (r) = sgn(σ )x1 . . . xr (see (6.4)). Consequently,

A2n = s2n (A1 , A2 , . . . , A2n )x1 x2 . . . x2n .

 n
Hence we are reduced to proving that A2n = 0, i.e., A2 = 0. Note that A2 lies
in Mn (G0 ). As G0 is a commutative algebra, Corollary 6.37 tells us that it suffices
 k
to show that A2 = A2k has zero trace for k = 1, . . . , n. To this end, notice that

A2k−1 = j Bj yj for some Bj ∈ Mn (Q) and yj ∈ G1 . Consequently,

        
tr A2k = tr AA2k−1 = tr Ai Bj xi yj = tr Ai Bj xi yj .
i,j i,j

On the other hand,

      
tr A2k = tr A2k−1 A = tr Bj Ai yj xi .
i,j

   
Comparing both expressions and using tr Ai Bj = tr Bj Ai (a general prop-
 of the trace) and xi yj = −yj xi (a consequence of yj ∈ G1 ), we obtain
erty
tr A2k = 0. 
Exercises 159

Exercises

6.1. Suppose that f ∈ Fξ, η commutes with ξ η. Show that there exists p ∈ F[ω]
such that f = p(ξ η).
Remark: This is a simple special case of a theorem by G. Bergman: If f , g ∈ FX
commute, then there exist h ∈ FX and p, q ∈ F[ω] such that f = p(h) and
g = q(h).

6.2. Find an automorphism of Fξ, η different from the identity map.

6.3. Show that Fξ1 , ξ2 , . . .  can be embedded into Fξ, η via ξi → ξ i η.
6.4. A linear subspace J of an algebra A is called a Jordan subalgebra of A if
u ◦ v := uv + vu ∈ J for all u, v ∈ J. For example, the set of symmetric
matrices in Mn (F), n ≥ 2, is a Jordan subalgebra of Mn (F), but not a subalgebra.
Elements of the Jordan subalgebra of FX generated by X ∪ {1} are called
Jordan polynomials. For example, if char(F) = 2 then ξ12 = 21 ξ1 ◦ ξ1 and
ξ1 ξ2 ξ1 = 21 ξ1 ◦ (ξ1 ◦ ξ2 ) − 41 (ξ1 ◦ ξ1 ) ◦ ξ2 are Jordan polynomials. Show that
ξ1n , ξ1 ξ2 ξ3 + ξ3 ξ2 ξ1 , (ξ1 ξ2 − ξ2 ξ1 )2 are also Jordan polynomials. Find a Jordan
subalgebra J of the Grassmann algebra G such that u1 u2 u3 u4 + u4 u3 u2 u1 ∈ /J
for some u1 , u2 , u3 , u4 ∈ J. Hence, conclude that ξ1 ξ2 ξ3 ξ4 + ξ4 ξ3 ξ2 ξ1 is not a
Jordan polynomial, and neither is ξ1 ξ2 . . . ξn + ξn ξn−1 . . . ξ1 for any n ≥ 4.
Remark: A nonassociative (= not necessarily
 associative)
 algebra J is called a
Jordan algebra if xy = yx and x 2 (yx) = x 2 y x for all x, y ∈ J. Every Jordan
subalgebra of an associative algebra is a Jordan algebra under the product ◦.
Such Jordan algebras are called special. There exist Jordan algebras that are not
special; they are called exceptional.

6.5. Analogously, a Lie subalgebra of an algebra A is defined as a linear subspace of

A, say L, such that [u, v] ∈ L for all u, v ∈ L. For example, the set of all trace zero
matrices in Mn (F), n ≥ 2, is a Lie subalgebra of Mn (F), but not a subalgebra.
A Lie polynomial is an element of the Lie subalgebra of FX generated by X.
Find a multilinear Lie polynomial of degree 3 which is simultaneously a Jordan
polynomial. Show that ξ1 ξ2 . . . ξn − ξn ξn−1 . . . ξ1 is not a Lie polynomial for
n ≥ 3.
Remark: A nonassociative algebra L is called a Lie algebra if x 2 = 0 and
(xy)z + (zx)y + (yz)x = 0 for all x, y, z ∈ L. A Lie subalgebra of an associative
algebra is a Lie algebra under the product [ . , . ]. Conversely, the Poincaré-
Birkhoff-Witt Theorem shows that every Lie algebra is isomorphic to a Lie
subalgebra of an associative algebra. (The product in Lie algebras is actually
usually denoted by [ . , . ], so the defining identities are written as [x, x] = 0 and
[[x, y], z] + [[z, x], y] + [[y, z], x] = 0; the latter is called the Jacobi identity.)
160 6 Noncommutative Polynomials

6.6. Describe the R-algebras C, R × R, and H in terms of generators and relations.

6.7. Find all invertible elements in Fx, x | xx = x x = 1.
Remark: A more standard notation for this algebra is F[x, x −1 ]. Its elements
are called Laurent polynomials (cf. Example 1.41). Note also that F[x, x −1 ]
is isomorphic to the group algebra F[Z].
6.8. Let q ∈ F ∗ . Show that Fx, x , y, y | xx = x x = 1, yy = y y = 1, xy = qyx
is a simple noetherian domain provided that q is not a root of unity.
6.9. Show that Fx, y | xy = 1 is a primitive algebra with nonzero socle, and that
it contains a left zero-divisor which is not a right zero-divisor.
Hint: Examine the element e := 1 − yx.
6.10. Prove that Fξ1 , ξ2  ⊗ Fη1 , η2  = ∼ Fx1 , x2 , y1 , y2 | [xi , yj ] = 0, i, j = 1, 2.
6.11. Prove that the Weyl algebras satisfy An ∼ = Ak ⊗ An−k , 1 ≤ k < n.
6.12. Find the Jacobson radical of the Grassmann algebra G.
6.13. Let fi (ξ1 , . . . , ξn ) ∈ FX, i = 1, . . . , m, be linearly independent polynomials.
Show that an F-algebra A must be a PI-algebra if for arbitrary x1 , . . . , xn ∈ A
the evaluations fi (x1 , . . . , xn ), i = 1, . . . , m, are linearly dependent in A.
6.14. We say that an F-algebra A is an algebraic algebra of bounded degree if
there exists n ∈ N such that for every x ∈ A there is f (ω) ∈ F[ω] of degree
n satisfying f (x) = 0. Show that the class of algebraic algebras of bounded
degree properly contains the class of finite dimensional algebras, and is properly
contained in the class of PI-algebras.
6.15. Show that the algebra of all finite rank operators on an infinite dimensional
vector space over F is an algebraic algebra, but not a PI-algebra.
6.16. Let A be an algebra and let M be an A-bimodule. Equip A × M with pointwise
addition and scalar multiplication, and the product (x, m)(y, n) = (xy, xn+my).
Check that in this way A × M becomes an algebra; let us call it AM . Show that
A is a PI-algebra if and only if AM is a PI-algebra.
6.17. Find an identity of minimal degree of the subalgebra of Mn (F) consisting of
all matrices that have zeros in all columns except in the first one.
6.18. Show that, in general, Id(A) = Id(A[ω]), although A and A[ω] satisfy the same
multilinear identities.
6.19. Let f = f (ξ1, . . . , ξn ) ∈ FX  be a multilinear polynomial, and let k ∈ N.
Prove that tr f (A1 , . . . , An ) = 0 for all A1 , . . . , An ∈ Mk (F) if and only if
there exist f0 ∈ Id(Mk (F)) and gi , hi ∈ FX such that f = f0 + i [gi , hi ].
Hint: Use mξn m = m mξn + [mξn , m ] for monomials m, m .
   
6.20. Show that every generic matrix in Mn F[Ω] is invertible in Mn F(Ω) . Hence
conclude that if F is an infinite field, m ∈ FX is a nonzero monomial, and
f ∈ FX is such that mf ∈ Id(Mn (F)), then f ∈ Id(Mn (F)).
Remark: A theorem by Amitsur says that this is true even when m is an arbitrary
polynomial which is not an identity of Mn (F). See also the comment following
Theorem 7.65.
Exercises 161

6.21. Let A be a nonzero finite dimensional semiprime algebra. Show that there exists
n ∈ N such that A satisfies s2n , but does not satisfy nonzero identities of degree
less than 2n.
6.22. Given A, B ∈ M2 (F), let

PA,B (ξ, η) := ξ η + ηξ − tr(B)ξ − tr(A)η + tr(A)tr(B) − tr(AB) ∈ Fξ, η.

Recalling the form (6.10) of the characteristic polynomial, we see that by

applying the linearization method to the Cayley-Hamilton Theorem we obtain
PA,B (A, B) = 0. Show that the identity
  
sgn(σ )PAσ (1) Aσ (2) ,Aσ (3) Aσ (4) Aσ (1) Aσ (2) , Aσ (3) Aσ (4) = 0
σ ∈S4

yields the Amitsur-Levitzki Theorem for n = 2 (provided that char(F) = 2,

but from the first paragraph of our proof of the Amitsur-Levitzki Theorem it is
evident that this can be assumed without loss of generality).
Remark: We have confined ourselves to n = 2 for clarity. The Amitsur-Levitzki
Theorem can be, in a similar fashion, derived from the Cayley-Hamilton theo-
rem for any n. This was discovered by Yu. P. Razmyslov, and is actually one of
the standard proofs of the Amitsur-Levitzki Theorem. A deeper result, obtained
independently by Razmyslov and C. Procesi, shows that in fact every polyno-
mial identity of Mn (F), char(F) = 0, is a consequence of the Cayley-Hamilton
identity pA (A) = 0. Moreover, the same is true for the more general trace iden-
tities. We omit giving their definition, but just give a hint by saying that the
characteristic polynomial pA gives rise to a trace identity that is not a polynomial
identity.
Chapter 7
Rings of Quotients and Structure of PI-Rings

A commutative domain can be embedded into a field, its field of quotients. One
usually learns this simple and highly useful fact in one’s first encounters with abstract
algebra. Therefore it is natural to seek its analogies in noncommutative rings. The
lack of commutativity causes considerable difficulties. Nevertheless, various gener-
alizations of the construction of the field of quotients exist in noncommutative ring
theory. We will discuss a few of them: rings of central quotients, classical rings of
quotients, and Martindale rings of quotients. Although now the issue is the noncom-
mutativity, we cannot entirely escape from the commutative context. The centers of
the constructed rings will be of crucial importance for us. The so-called extended
centroid of a prime ring, i.e., a field defined as the center of the Martindale ring of
quotients, will enable us to extend a part of the theory of central simple algebras
to general prime rings. We will also use it as a principal tool in the second part
of the chapter, devoted to the structure theory of prime rings satisfying polynomial
identities (or their generalizations, the generalized polynomial identities). As a very
rough summary, it can be stated that prime PI-rings are close to finite dimensional
simple algebras.

7.1 Rings of Central Quotients

A glance at the construction of the field of quotients of a commutative domain shows

that it highly depends on the commutativity. There is no obvious way to avoid it.
However, in this first section we will take an easy way; we are going to use essentially
the same construction for a noncommutative ring R, but allow only the elements from
the center Z = Z(R) as “denominators”. More precisely, we will assume that the
nonzero elements from Z satisfy the (obviously necessary) condition from the next
definition, and enlarge R in such a way that these elements become invertible in the
new, larger ring.

© Springer International Publishing Switzerland 2014 163

M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4_7
164 7 Rings of Quotients and Structure of PI-Rings

Definition 7.1 A nonzero element a in a ring R is said to be regular if it is neither

a left nor right zero-divisor.

Constructing the ring of central quotients QZ (R). We assume that R is any ring
such that its center Z is nonzero and all elements in Z\{0} are regular; in particular,
 
Z is a commutative domain. We claim that the relation ∼ on the set R × Z\{0}
defined by

(r, z) ∼ (r  , z ) ⇐⇒ rz = r  z

is an equivalence relation. Only the transitivity is not entirely obvious. Thus, assume
that (r, z) ∼ (r  , z ) and (r  , z ) ∼ (r  , z ), i.e., rz = r  z and r  z = r  z . Then
(rz −r  z)z = r  zz −r  z z = 0, and so (r, z) ∼ (r  , z ) follows from the regularity
of z . Let rz−1 denote the equivalence class of (r, z). Thus, rz−1 = r  z−1 if and only if
rz = r  z; in particular, rz(zw)−1 = rw−1 . Let QZ (R) denote the set of all equivalence
classes, equipped with addition and multiplication defined by

rz−1 + sw−1 := (rw + sz)(zw)−1 ,

rz−1 · sw−1 := rs(zw)−1 .

To prove that these operations are well-defined, we must show that (r, z) ∼ (r  , z )
and (s, w) ∼ (s , w ) imply (rw + sz, zw) ∼ (r  w + s z , z w ) and (rs, zw) ∼ (r  s ,
z w ). But this is straightforward. As one would expect, QZ (R) equipped with these
operations is a ring. Verifying this requires a little patience, but no more than verifying
that Q is a ring.

Definition 7.2 The ring QZ (R) is called the ring of central quotients of R.

No matter whether R is unital or not, QZ (R) is. Its unity is z0 z0−1 , where z0 is an
arbitrary element in Z\{0}. The center of QZ (R), which we denote by  Z, is easily
seen to consist of elements of the form zw−1 , z, w ∈ Z, w = 0. If z = 0, then wz−1
is the inverse of zw−1 . Therefore Z is a field; of course, 
Z is nothing but the field
of quotients of Z. Every ring whose center is a field can be considered as a central
algebra over its center. Thus, QZ (R) is a central algebra over Z.
One immediately checks that

ι : R → QZ (R), ι(r) = (rz0 )z0−1 ,

is an embedding of R into QZ (R) (we remark that the definition of ι is independent

of the choice of z0 ∈ Z\{0}). It is convenient to identify R with its isomorphic copy
ι(R) inside QZ (R), and thus consider R as a subring of QZ (R). The expression rz−1 ,
so far used just as a notation for the equivalence class, can then be interpreted as the
product of r ∈ R and the inverse of z ∈ Z\{0}. Every element in Z\{0} is invertible
in QZ (R). Intuitively, QZ (R) is the smallest ring containing R that has this property.
The next proposition says this in more formal terms, through a universal property.
7.2 Rings of Central Quotients 165

Proposition 7.3 If T is a unital ring and ϕ : R → T is a homomorphism such that

ϕ(z) is invertible in T for every z ∈ Z\{0}, then ϕ can be uniquely extended to a
homomorphism ϕ : QZ (R) → T .
Proof It is straightforward to check that ϕ(rz−1 ) := ϕ(r)ϕ(z)−1 is a well-defined
homomorphism which extends ϕ. The uniqueness is clear. 
Let us add that if ϕ is injective and every element in T can be written as ϕ(r)ϕ(z)−1 ,
then ϕ is an isomorphism.
The next theorem summarizes our discussion up to now.
Theorem 7.4 Let R be a ring such that its center Z is nonzero and all elements in
Z\{0} are regular. The ring QZ (R) has the following properties:
(a) QZ (R) is a unital ring containing R as a subring.
(b) Every element in Z\{0} is invertible in QZ (R).
(c) Every element in QZ (R) is of the form rz−1 , where r ∈ R and z ∈ Z\{0}.
Moreover, these properties characterize QZ (R) up to isomorphism.
It is easy to see that if R is a prime ring with nonzero center Z, then QZ (R) exists
(i.e., nonzero elements in Z are regular) and is a prime ring.
We continue with examples. The first two are obvious, but worth mentioning.
Example 7.5 If the center of a unital ring R is a field, then QZ (R) = R.
Example 7.6 If R is a commutative domain, then QZ (R) is the field of quotients of R.
Example 7.7 We claim that QZ (Mn (Z)) ∼ = Mn (Q). Obviously, Mn (Q) has the prop-
erty (a). Commuting with matrix units we easily notice that the center of Mn (Z)
consists of matrices nI with n ∈ Z (cf. the proof of Lemma 1.15). Therefore Mn (Q)
has the property (b), as well as (c); to notice the latter, make use of the common
denominator of the entries of a matrix in Mn (Q).
 Z 2Z 
Example 7.8 A small variation of the preceding example: If R = Z Z , i.e., the
subring of M2 (Z) consisting of matrices whose (1, 2) entry is an even integer, then
QZ (R) = M2 (Q). The point we wish to make here is that even if a ring is a bit “odd”,
its ring of central quotients may be “spotless”.
Example 7.9 If R is the ring of quaternions with integer coefficients, R = Z ⊕
Zi ⊕ Zj ⊕ Zk, then QZ (R) is the ring of quaternions with rational coefficients,
QZ (R) = Q ⊕ Qi ⊕ Qj ⊕ Qk. Indeed, Z(R) = Z and checking (a)–(c) is immediate.
Certain simple modifications of our construction of QZ (R) are also important and
can be found in standard textbooks, usually under the name “(central) localization”.
We will not need them, but let us just give a clue what changes can be made. The role
of Z \{0} can be replaced by any subset S of Z that is closed under multiplication, and
instead of requiring that the elements from S are regular, one defines the equivalence
relation ∼ on R × S as follows: (r, z) ∼ (r  , z ) ⇐⇒ (rz − r  z)z = 0 for some
z ∈ S.
166 7 Rings of Quotients and Structure of PI-Rings

7.2 Classical Rings of Quotients

Given a ring R, is it possible to construct a larger ring Q such that all regular elements
in R, not only those from the center, are invertible in Q? More specifically, can we
construct Q in such a way that Theorem 7.4 with the set of all regular elements
playing the role of Z\{0} holds for Q? The answer is not as straightforward as for
the ring of central quotients. It certainly is positive for some rings (for commutative
domains, for example). But it can also be negative, as we will see.
Let us make it precise what ring Q we wish to have. To avoid trivialities, we tacitly
assume that our rings contain at least one regular element.

Definition 7.10 Let R be a ring and S be the set of all regular elements in R. A ring
Q is called a right classical ring of quotients of R if it has the following properties:
(a) Q is a unital ring containing R as a subring.
(b) Every element in S is invertible in Q.
(c) Every element in Q is of the form rs−1 , where r ∈ R and s ∈ S.

It is clear that elements of R that are not regular cannot be invertible in any larger
ring. The existence of a right classical ring of quotients of R is thus the best we can
hope for if we wish to make as many elements in R invertible as possible.
A left classical ring of quotients of R is defined analogously, one just substitutes
s−1 r for rs−1 in (c). However, up until the end of the section we will consider only
right classical rings of quotients.

Example 7.11 Let  ZR2Z

be any of the rings from Examples 7.6–7.9, i.e., a commutative
domain, Mn (Z), Z Z , or Z ⊕ Zi ⊕ Zj ⊕ Zk. It is easy to verify that the ring of
central quotients QZ (R) is also a right classical ring of quotients of R.

Example 7.12 If every regular element in R is already invertible, then R is its own
right classical ring of quotients. Every finite dimensional unital algebra thus has this
property (see Remark 1.31 (b)).

Suppose R has a right classical ring of quotients Q. Given r ∈ R and s ∈ S, we

have s−1 r ∈ Q by (a) and (b), and so by (c) there exist r1 ∈ R and s1 ∈ S such that
s−1 r = r1 s1−1 . Multiplying by s from the left and by s1 from the right we obtain
rs1 = sr1 . Thus, the following holds:

rS ∩ sR = ∅ for every r ∈ R and s ∈ S. (7.1)

It is easy to find rings in which (7.1) is not fulfilled, and hence rings not having
right classical ring of quotients (see Example 7.18 below). The main goal of this
section is to prove that (7.1) is also sufficient, not only necessary, condition for a ring
R to have a right classical ring of quotients. This result is basically due to O. Ore;
more precisely, he proved it, in 1931, for the special case where R is a domain. The
condition (7.1) is called the right Ore condition.
7.2 Classical Rings of Quotients 167

Unlike in the previous section, the noncommutativity is now a serious obstacle.

There are two standard ways to overcome it. The first one, used already by Ore, is
based on a careful modification of the construction from the previous section. While
this approach is intuitively clear, working out the details is extremely tedious. We
will therefore choose the other way. To gain intuition, we sketch it for the seminal
case R = Z. Via the regular representation we may identify Z with the maps x → nx,
n ∈ Z, from Z into itself. Our goal is to describe Q without explicit use of fractions,
i.e., by staying inside the framework of the ring Z. Consider, for example, the map
f : 2Z → Z, f (2x) = x. Note that f halves each number from its domain, so it carries
information about the rational number 21 . Similarly, g : mZ → Z, g(mx) = nx, where
m ∈ N and n ∈ Z, corresponds to mn ∈ Q (in the sense that g(u) = mn u for all u ∈ mZ).
Given another map of this kind, g : m Z → Z, g (m x) = n x, we can consider g+g
and gg as maps defined on mm Z (maybe one can choose larger domains, but this
 
is irrelevant for us). Note that g + g corresponds to mn + mn  , and gg to mn · mn  . By
identifying maps that correspond to the same rational number we can obtain, in this
way, an alternative definition of the field Q. Perhaps an overcomplicated one, but
having an advantage of being suitable for generalizations to noncommutative rings.
The crucial point is that the above maps are Z-module homomorphisms from ideals
of the ring Z into Z.
Constructing the right classical ring of quotients Qrc (R). Let R be a ring sat-
isfying the right Ore condition (7.1). Denote by I the set of all right ideals of R
that contain at least one regular element. A right ideal can be considered as a right
R-module. Consider the set of all pairs ( f , I) where I ∈ I and f : I → R is a right
R-module homomorphism. Define an equivalence relation on this set as follows:
( f , I) ∼ (g, J) if f and g coincide on some K ∈ I such that K ⊆ I ∩ J. In order
to verify the transitivity, it is enough to notice that I1 ∩ I2 ∈ I if both I1 , I2 ∈ I .
Let us therefore prove this. Take si ∈ Ii ∩ S, i = 1, 2, and use (7.1) to find s ∈ S and
r ∈ R such that s1 s = s2 r. Hence I1 ∩ I2 contains a regular element, namely s1 s, so
it belongs to I .
We write [ f , I] for the equivalence class determined by ( f , I). Let Qrc (R) denote
the set of all equivalence classes, endowed with addition and multiplication

[ f1 , I1 ] + [ f2 , I2 ] := [ f1 + f2 , I1 ∩ I2 ],
[ f1 , I1 ][ f2 , I2 ] := [ f1 f2 , f2−1 (I1 )].

We must prove that these operations are well-defined. First of all, f2−1 (I1 ) = {x ∈
I2 | f2 (x) ∈ I1 } indeed lies in I . Namely, from (7.1) we infer that given si ∈ Ii ∩ S,
i = 1, 2, there exist s ∈ S and r ∈ R such that f2 (s2 )s = s1 r, showing that the
right ideal f2−1 (I1 ) contains the regular element s2 s. Assume now that ( f1 , I1 ) ∼
(g1 , J1 ) and ( f2 , I2 ) ∼ (g2 , J2 ). Thus, there exist Ki ∈ I such that Ki ⊆ Ii ∩ Ji
and fi = gi on Ki , i = 1, 2. Consequently, f1 + f2 = g1 + g2 on K1 ∩ K2 ∈ I
and hence ( f1 + f2 , I1 ∩ I2 ) ∼ (g1 + g2 , J1 ∩ J2 ), which proves that addition is
well-defined. Concerning multiplication, note that f1 f2 = g1 g2 on f2−1 (K1 ) ∩ K2 ∈
I and hence ( f1 f2 , f2−1 (I1 )) ∼ (g1 g2 , g2−1 (J1 )). Thus, both operations are indeed
168 7 Rings of Quotients and Structure of PI-Rings

well-defined. Verifying that Qrc (R) is a ring with zero element [0, R] and unity
[idR , R] is straightforward.
This construction is the basic example of the so-called “noncommutative local-
ization”. Its general version considers an arbitrary set closed under multiplication
instead of the set of all regular elements. This does not make the construction sub-
stantially more complicated, but we will not go into this matter.
We are now in a position to prove the result announced above.

Theorem 7.13 (Ore) A ring R has a right classical ring of quotients if and only if R
satisfies the right Ore condition.

Proof We only have to prove the “if” part. Assume that R satisfies the right Ore
condition and let us show that Qrc (R) is a right classical ring of quotients of R.
We embed R into Qrc (R) via

ι : R → Qrc (R), ι(r) = [Lr , R];

here, as always, Lr stands for a left multiplication map. Let us only check the injec-
tivity. If [Lr , R] = 0, then Lr = 0 on some I ∈ I . Since I contains a regular element,
it follows that r = 0.
Let s ∈ S. Then the map fs : sR → R, fs (sx) = x, is well-defined. Moreover, fs is
a right R-module homomorphism. Note that sR ∈ I since s2 ∈ sR ∩ S. The reader
can readily verify that [ fs , sR] is the inverse of ι(s). In particular, ι(s) is invertible.
Now take an arbitrary [ f , I] ∈ Qrc (R). Pick s ∈ I ∩ S and set r := f (s) ∈ R. We
have f (sx) = rx = Lr fs (sx) for all x ∈ R. Hence

[ f , I] = [ f , sR] = [Lr fs , sR] = [Lr , R][ fs , sR] = ι(r)ι(s)−1 .

Identifying R with ι(R) we thus see that Qrc (R) has all the desired properties. 

The ring Qrc (R) has a universal property, analogous to that satisfied by QZ (R)
(see Proposition 7.3).

Proposition 7.14 Let R be a ring satisfying the right Ore condition. If T is a unital
ring and ϕ : R → T is a homomorphism such that ϕ(s) is invertible in T for
every regular element s ∈ R, then ϕ can be uniquely extended to a homomorphism
ϕ : Qrc (R) → T .

Proof Take [ f , I] ∈ Qrc (R). Pick s ∈ I ∩ S and define ϕ : Qrc (R) → T by

ϕ([ f , I]) := ϕ( f (s))ϕ(s)−1 .

One can immediately check that ϕ(ι(r)) = ϕ(r), so ϕ extends ϕ. However, we

must first prove that ϕ is well-defined. Assume, therefore, that ( f , I) ∼ (g, J).
Our goal is to show that ϕ( f (s))ϕ(s)−1 = ϕ(g(t))ϕ(t)−1 for every t ∈ J ∩ S.
Let K ∈ I be such that K ⊆ I ∩ J and f = g on K. It suffices to show that
7.2 Classical Rings of Quotients 169

ϕ( f (s))ϕ(s)−1 = ϕ( f (u))ϕ(u)−1 holds for u ∈ K ∩ S. Namely, by analogy we then

also have ϕ(g(t))ϕ(t)−1 = ϕ(g(u))ϕ(u)−1 , and so the desired conclusion follows
from f (u) = g(u). By regularity of s and u we have sR, uR ∈ I , and hence sR∩uR ∈
I . Therefore there exist x, y ∈ R such that v := sx = uy ∈ S. Since ϕ(s) and ϕ(v)
are invertible in T by assumption, v = sx gives ϕ(s)−1 = ϕ(x)ϕ(v)−1 . Consequently,

ϕ( f (s))ϕ(s)−1 = ϕ( f (s))ϕ(x)ϕ(v)−1 = ϕ( f (v))ϕ(v)−1 .

Similarly, from v = uy we infer ϕ( f (u))ϕ(u)−1 = ϕ( f (v))ϕ(v)−1 . Comparing we

get ϕ( f (s))ϕ(s)−1 = ϕ( f (u))ϕ(u)−1 .
The additivity of ϕ is clear. To prove the multiplicativity, take [ fi , Ii ] ∈ Qrc (R),
i = 1, 2. Pick s1 ∈ I1 ∩ S. By definition we have ϕ([ f1 , I1 ][ f2 , I2 ]) = ϕ( f1 ( f2 (s2 )))
ϕ(s2 )−1 where s2 is an arbitrary element in f2−1 (I1 ) ∩ S; since [ f1 , I1 ] = [ f1 , s1 R]
we may choose s2 ∈ f2−1 (s1 R). Thus f2 (s2 ) = s1 x and hence ϕ(x) = ϕ(s1 )−1
ϕ( f2 (s2 )), yielding

ϕ( f1 ( f2 (s2 ))) = ϕ( f1 (s1 ))ϕ(x) = ϕ( f1 (s1 ))ϕ(s1 )−1 ϕ( f2 (s2 )).

This readily implies that ϕ([ f1 , I1 ][ f2 , I2 ]) = ϕ([ f1 , I1 ])ϕ([ f2 , I2 ]). Thus, ϕ is a

homomorphism extending ϕ. The uniqueness is obvious because of condition (c)
from Definition 7.10. 

Corollary 7.15 If a ring satisfies the right Ore condition, then its right classical
ring of quotients is unique up to isomorphism.

Remark 7.16 Similar results of course hold for the left classical ring of quotients.
One defines the left Ore condition as Sr ∩ Rs = ∅ for every r ∈ R and s ∈ S, and
constructs the left classical ring of quotients Qlc (R) of R by using left ideals and right
multiplication maps for defining the canonical embedding. Proposition 7.14 together
with its left analogue easily implies that if R has both a left and right classical ring
of quotients, then they are isomorphic. In other words, if both the left and right Ore
condition are fulfilled, then Qrc (R) ∼= Qlc (R).

7.3 Ore Domains

In this section we take a closer look at the classical situation where R is a domain.
The right Ore condition (7.1) in this case gets a slightly simpler form:

rR ∩ sR = 0 for every r, s ∈ R \ {0}. (7.2)

Definition 7.17 A nonzero domain satisfying the right Ore condition (7.2) is called
a right Ore domain.
170 7 Rings of Quotients and Structure of PI-Rings

Theorem 7.13 implies that a ring R is a right Ore domain if and only if R is a subring
of a division ring D (=Qrc (R)) such that every element in D is of the form rs−1 with
r, s ∈ R, s = 0.
Not every domain is right Ore.

Example 7.18 Let R = Fξ, η. We obviously have ξ R ∩ ηR = 0. Therefore R is not

a right Ore domain (and neither a left Ore domain, i.e., the condition Rr ∩ Rs = 0
for every r, s ∈ R \ {0} is not fulfilled). 

A domain that is neither right nor left Ore may still be embedded into a division
ring, but not in such a way that condition (c) of Definition 7.10 would be fulfilled.
The free algebra Fξ, η turns out to be an example of such a domain. On the other
hand, there do exist domains that cannot be embedded into division rings. For details
we refer the reader to [Lam99, Chap. 9].
The connection between Fξ, η and the nonfulfillment of the right Ore condition
is even tighter.

Proposition 7.19 Let A be a nonzero unital algebra over a field F. If A is a domain

that is not right Ore, then A contains a subalgebra isomorphic to Fξ, η.

Proof By assumption there exist r, s ∈ A \ {0} such that rA ∩ sA = 0. Let A0 be the

unital subalgebra of A generated by r and s. We claim that the unital homomorphism
f (ξ, η) → f (r, s) from Fξ, η onto A0 is injective, and hence an isomorphism.
Suppose this is not true. Then there exists a nonzero polynomial f (ξ, η) such that
f (r, s) = 0. We may assume that f is of minimal degree. Let us write f = λ +
ξ g + ηh, where λ is the constant term of f and g, h ∈ Fξ, η. There is no loss of
generality in assuming that h = 0. Setting u := g(r, s) and v := h(r, s) we thus have
λ + ru + sv = 0. Multiplying on the right by r we obtain r(λ + ur) + svr = 0, and
so r(λ + ur) = −svr ∈ rA ∩ sA = 0. Since deg(h) < deg( f ), v = 0. As s and r are
also nonzero and A is a domain, we have arrived at a contradiction. 

Trivial examples of right Ore domains are division rings and commutative domains.
As another simple example we mention the ring of quaternions with integer coeffi-
cients (cf. Examples 7.9 and 7.11). The next proposition gives rise to less obvious
examples.

Proposition 7.20 (Goldie) A right noetherian domain R is a right Ore domain.

Proof Take r, s ∈ R \ {0}, and set Jn := rsR + r 2 sR + · · · + r n sR. Since the Jn ’s

obviously form an ascending chain of right ideals, there exists n ∈ N such that
Jn = Jn+1 = . . .. Pick any x0 ∈ R \ {0}. Then r n+1 sx0 is a nonzero element from
Jn+1 , and hence also from Jn . Thus there exists k ≤ n such that

r k sxk + r k+1 sxk+1 + · · · + r n sxn = r n+1 sx0

for some xk , . . . , xn ∈ R with xk = 0. The cancellation property of R yields

7.3 Ore Domains 171

sxk + rsxk+1 + · · · + r n−k sxn = r n+1−k sx0 .

Therefore sxk = 0 lies in rR. This shows that (7.2) holds. 

Example 7.21 The Weyl algebra An is a noetherian domain (Example 6.4), therefore
Qrc (An ) exists and is a division ring. Moreover, Qrc (An ) ∼
= Qlc (An ) by Remark 7.16
and the left version of Proposition 7.20.

Proposition 7.20 is just a small extract from the theory developed by A. W. Goldie
in the late 1950s. His main result characterizes rings whose right classical ring of
quotients is isomorphic to Mn1 (D1 ) × · · · × Mnr (Dr ) where Di are division rings.
A corollary of special importance is that right noetherian semiprime rings do have
this property. Goldie’s theory is treated in many books, including [Her68, Lam99,
MR01, Row08].

7.4 Martindale Rings of Quotients

In this section we take another step forward in the study of rings of quotients—
or perhaps a step backwards. Namely, the next construction is similar to, and in
fact slightly simpler than the construction of Qrc (R). Roughly speaking, we will
just replace the role of the right ideals containing regular elements by the (easier to
handle) nonzero two-sided ideals. In spite of the technical similarities, the philosophy
behind these two constructions is different. The goal of the next one is not making
regular elements invertible, but merely providing a larger ring that can be used as
a tool for solving problems concerning the original ring. Still, the field of quotients
of a commutative domain serves as a prototype here as well. We will assume that
our rings are prime. This is a natural setting; a commutative ring is a domain if and
only if it is prime, and the role of prime rings among general rings is often parallel
to the role of domains among commutative rings. With a little extra effort we could
handle more general semiprime rings, but we will stick with the classical prime ring
situation, as considered in the seminal work of W. S. Martindale from 1969.
Constructing the right Martindale ring of quotients Qr (R). Let R = 0 be a
prime ring. The set J of all nonzero ideals of R is closed under products, and hence
also under finite intersections. We endow the set of all pairs ( f , I), where I ∈ J
and f : I → R is a right R-module homomorphism, by the following relation, which
is readily seen to be equivalence: ( f , I) ∼ (g, J) if f and g coincide on some K ∈ J
such that K ⊆ I ∩ J. Write [ f , I] for the equivalence class determined by ( f , I),
and denote by Qr (R) the set of all equivalence classes, equipped with addition and
multiplication

[ f1 , I1 ] + [ f2 , I2 ] := [ f1 + f2 , I1 ∩ I2 ],
[ f1 , I1 ][ f2 , I2 ] := [ f1 f2 , I2 I1 ]
172 7 Rings of Quotients and Structure of PI-Rings

(note that f1 f2 is indeed defined on I2 I1 since f2 (I2 I1 ) = f2 (I2 )I1 ⊆ I1 ). It is easy to

verify that these operations are well-defined. Indeed, assume that ( f1 , I1 ) ∼ (g1 , J1 )
and ( f2 , I2 ) ∼ (g2 , J2 ), i.e., there exist Ki ∈ J such that Ki ⊆ Ii ∩ Ji and fi = gi on
Ki , i = 1, 2. Then f1 +f2 = g1 +g2 on K1 ∩K2 ∈ J and f1 f2 = g1 g2 on K2 K1 ∈ J .
Thus, ( f1 + f2 , I1 ∩ I2 ) ∼ (g1 + g2 , J1 ∩ J2 ) and ( f1 f2 , I2 I1 ) ∼ (g1 g2 , J2 J1 ). One
immediately checks that Qr (R) is a ring with zero element [0, R] and unity [idR , R].

Definition 7.22 The ring Qr (R) is called the right Martindale ring of quotients
of R.

The left Martindale ring of quotients Ql (R) is constructed analogously through

the left R-module homomorphisms. In general, Qr (R) ∼ = Ql (R).

Theorem 7.23 Let R = 0 be a prime ring, and let J be the set of all nonzero ideals
of R. The ring Qr (R) has the following properties:
(a) Qr (R) is a unital ring containing R as a subring.
(b) For every q ∈ Qr (R) there exists I ∈ J such that qI ⊆ R.
(c) For every q ∈ Qr (R) and I ∈ J , qI = 0 implies q = 0.
(d) If I ∈ J and f : I → R is a right R-module homomorphism, then there exists
q ∈ Qr (R) such that f (x) = qx for all x ∈ I.
Moreover, these properties characterize Qr (R) up to isomorphism.

Proof We already know that Qr (R) is a unital ring. Note that

ι : R → Qr (R), ι(r) = [Lr , R],

where Lr is a left multiplication map, is an embedding of R into Qr (R) (injectivity

follows from the primeness of R). Identifying R with its isomorphic copy ι(R) we
may consider R as a subring of Qr (R), so that (a) holds.
Take q = [ f , I] ∈ Qr (R). For every x ∈ I we have

qx = [ f , I][Lx , R] = [ fLx , RI] = [Lf (x) , R] = f (x) ∈ R.

This proves (b). Moreover, note that qI = 0 implies f (I) = 0 and hence q = [ f , I] = 0.
If I  is another member of J and qI  = 0, then also (qI)I  ⊆ qI  = 0. Since R is
prime and qI ⊆ R, this yields qI = 0, and hence q = 0. Thus (c) is proved. Verifying
(d) is also easy: If I ∈ J and f : I → R is a right R-module homomorphism, then,
as noticed at the beginning of this paragraph, q := [ f , I] ∈ Qr (R) satisfies f (x) = qx
for all x ∈ I.
Let Q be an arbitrary ring with properties (a)–(d). Take q ∈ Q. By (b) there exists
I ∈ J such that qI ⊆ R. Therefore we can define the right R-module homomorphism
f : I → R by f (x) = qx for all x ∈ I. If I  ∈ J also satisfies qI  ⊆ R, then f coincides
with f  : I  → R, f (x  ) = qx  , on I ∩ I  ∈ J . Hence we see that the map

ϕ : Q → Qr (R), ϕ(q) = [ f , I],

7.4 Martindale Rings of Quotients 173

is well-defined. It is easy to check that ϕ is a ring homomorphism. The injectivity of

ϕ follows from (c), and the surjectivity from (d). Therefore Q ∼= Qr (R). 

Further properties of Qr (R) can be extracted from the basic ones (a)–(d). For
example, (c) can be extended as follows.

Lemma 7.24 For every q1 , q2 ∈ Qr (R) and I ∈ J , q1 Iq2 = 0 implies q1 = 0 or

q2 = 0. Accordingly, Qr (R) is a prime ring.

Proof By (b) we may choose Ii ∈ J such that qi Ii ⊆ R, i = 1, 2. Hence we may

conclude from (q1 I1 )I(q2 I2 ) ⊆ (q1 Iq2 )I2 = 0 that q1 I1 = 0 or q2 I2 = 0 (see Remark
2.25). From (c) we thus get q1 = 0 or q2 = 0. 

For further reference we record a straightforward extension of (b).

Lemma 7.25 Given q1 , . . . , qn ∈ Qr (R) there exists I ∈ J such that qi I ⊆ R,

i = 1 . . . , n.

Proof By (b), for each i there exists Ii ∈ J such that qi Ii ⊆ R. Hence I := I1 ∩· · ·∩In
has the desired property. 

Example 7.26 If R is a commutative domain, then, as already indicated, Qr (R) is

isomorphic to the field of quotients QZ (R) of R. We must show that QZ (R) has
properties (a)–(d). Only (d) is not entirely obvious. Assume, therefore, that f : I → R
is a right R-module homomorphism. Take a nonzero a ∈ I. We have f (x)a = f (xa) =
f (ax) = f (a)x for every x ∈ I, and hence f (x) = qx where q := f (a)a−1 ∈ QZ (R).
Thus, for a commutative domain R we have QZ (R) ∼ = Qcr (R) ∼ = Qr (R). The
same is true if R is any of the rings from Examples 7.7–7.9, i.e., Mn (Z), Z 2Z
Z Z , or
Z ⊕ Zi ⊕ Zj ⊕ Zk. One can prove this from scratch. On the other hand, one can apply
Corollary 7.61 below.

Example 7.27 If R is a simple unital ring, then Qr (R) = R. Indeed, R obviously has
properties (a)–(c), and checking that it also satisfies (d) is immediate: If f : R → R
is a right R-module homomorphism, then f (x) = f (1x) = f (1)x for all x ∈ R.

Example 7.28 Let V be an infinite dimensional vector space over a field F. Denote
by K the subring of EndF (V ) consisting of all finite rank operators. Clearly, every
operator in K is a sum of rank one operators v ⊗ h, defined by (v ⊗ h)(u) = h(u)v;
here, v ∈ V and h ∈ V ∗ , the dual of V (cf. Remark 4.17). It is an easy exercise to show
that K is a simple ring without unity. Let R be any ring such that K ⊆ R ⊆ EndF (V ).
Note that R is prime and that K is its minimal ideal. Let us show that

Qr (R) ∼
= EndF (V ).

We must verify that EndF (V ) has the four properties (a)–(d). The first three are
easy: (a) is trivial, (b) follows from the fact that K is a (left) ideal of EndF (V ), and
174 7 Rings of Quotients and Structure of PI-Rings

(c) from the fact that every nonzero ideal of R contains K. To verify (d), take a
nonzero ideal I of R and a right R-module homomorphism f : I → R. Pick v0 ∈ V
and h0 ∈ V ∗ so that h0 (v0 ) = 1. Define q : V → V by q(v) := f (v ⊗ h0 )(v0 ).
Obviously, q(v + v ) = q(v) + q(v ). Since

f (λk)k  = f (λkk  ) = f (k(λk  )) = λf (k)k 

 
holds for all λ ∈ F and k, k  ∈ K, i.e., f (λk) − λf (k) K = 0, it follows that
f (λk) = λf (k), k ∈ K. This yields q(λv) = λq(v). That is, q ∈ EndF (V ). Next, for
all v ∈ V and h ∈ V ∗ we have

f (v ⊗ h) = f (v ⊗ h0 · v0 ⊗ h) = f (v ⊗ h0 ) · v0 ⊗ h
= f (v ⊗ h0 )(v0 ) ⊗ h = q(v) ⊗ h
= q · v ⊗ h.

Thus, f (k) = qk for every k ∈ K, and hence f (x)k = f (xk) = qxk for all x ∈ I and
k ∈ K. That is, ( f (x) − qx)K = 0, which yields f (x) = qx. This completes the proof.

7.5 The Extended Centroid

Throughout this section we assume that R is a nonzero prime ring.

Definition 7.29 The center of Qr (R) is called the extended centroid of R.

The extended centroid of R will be denoted by C. As usual, Z will stand for the
center of R, and, as in the preceding section, J for the set of all nonzero ideals of
R. We will continuously refer to properties (a)–(d) of Theorem 7.23.

Lemma 7.30 If q ∈ Qr (R) is such that qr = rq for all r ∈ R, then q ∈ C. Accord-

ingly, Z is a subring of C.

Proof Pick q ∈ Qr (R). We must show that [q, q ] = 0. By (b) there exists I ∈ J
such that q I ⊆ R. Take x ∈ I. Since q commutes with x and q x, we have qq x =
q xq = q qx. Thus, [q, q ]I = 0, and hence [q, q ] = 0 by (c). 

If elements in Qr (R) correspond to right R-module homomorphisms, then

elements in C correspond to R-bimodule homomorphisms (these are, of course,
maps that are both left and right R-module homomorphisms):

Lemma 7.31 If f : I → R, where I ∈ J , is an R-bimodule homomorphism, then

there exists λ ∈ C such that f (x) = λx for all x ∈ I.

Proof Since f is, in particular, a right R-module homomorphism, by (d) there exists
q ∈ Qr (R) such that f (x) = qx, x ∈ I. On the other hand, since f is also a
7.5 The Extended Centroid 175

left R-module homomorphism it follows that q(rx) = r(qx) for all r ∈ R and
x ∈ I. That is, [q, r]I = 0, and so [q, r] = 0 by (c). Therefore λ := q ∈ C by
Lemma 7.30. 

Conversely, if λ ∈ C and I ∈ J are such that λI ⊆ R, then x → λx is an

R-bimodule homomorphism from I into R. The extended centroid, although defined
through the right Martindale ring of quotients, is thus a left-right symmetric notion.

Remark 7.32 Suppose R is an algebra over a field F. Every α ∈ F gives rise to the R-
bimodule homomorphism x → αx from R into R. Using Lemma 7.31 one therefore
easily infers that F embeds canonically in C. Accordingly, F can be considered as a
subfield of C.

Theorem 7.33 The extended centroid C of a nonzero prime ring R is a field.

Proof Take λ = 0 in C and choose I ∈ I satisfying (b). Clearly, λI ∈ J by (c).

We claim that f : λI → R, f (λx) = x, is well-defined. Indeed, λx = 0 obviously
implies λIx = 0, where Ix is the ideal of R generated by x, and hence x = 0 by
(c). Since f is an R-bimodule homomorphism, Lemma 7.31 yields the existence of
μ ∈ C such that f (y) = μy, y ∈ λI. Writing y = λx, x ∈ I, we thus get μλx = x.
That is, (μλ − 1)I = 0, and so μλ = 1 by (c). Thus, λ is invertible. 

We may therefore regard Qr (R) as an algebra over C.

Definition 7.34 The C-subalgebra of Qr (R) generated by R is called the central

closure of R.

of R, which will be denoted by RC , obviously consists of

The central closure
elements of the form i λi ri , where λi ∈ C, ri ∈ R. Thus, RC is, like C, a left-right
symmetric object. The following symmetric version of Lemma 7.25 therefore holds
for elements in RC .

Lemma 7.35 Given q1 , . . . , qn ∈ RC , there exists I ∈ J such that qi I ⊆ R and

Iqi ⊆ R, i = 1 . . . , n.
Proof It is enough to prove the lemma for n = 1. Namely, just as in the proof
of Lemma 7.25, we can take  the intersection of ideals corresponding to particular
elements. Thus, pick q = i λi ri ∈ RC . By Lemma 7.25 there exists I ∈ I such
that λi I ⊆ R for each i. This clearly implies that qI ⊆ R and, since λi commutes
with elements from I, also Iq ⊆ R. 

Assume that R has a nonzero center Z. Then the field of quotients 

Z of Z may be
viewed as a subfield of C. For clarity we record the following simple lemma.

Lemma 7.36 If Z = 0, then QZ (R) can be embedded in RC . Moreover, if 

Z = C,
then QZ (R) ∼
= RC .
Proof The first statement follows from Proposition 7.3, and the second one is
obvious. 
176 7 Rings of Quotients and Structure of PI-Rings

Rings from Example 7.26 satisfy the condition  Z = C. In general, 

Z is not equal
to C, not even when Z is already a field and thus equal to 
Z.

Example 7.37 Let F0 be a subfield of F. Consider the special case of Example 7.28
where R = K + F0 · 1 (here, 1 is the identity operator). Note that Z = F0 · 1, and
Z=Z∼
hence  = F0 . In Example 7.28 we have shown that Qr (R) ∼= EndF (V ), which
implies that C ∼
= F. Accordingly, RC = K + F · 1.

This example indicates the intrinsic nature of the notion of the extended centroid.
Namely, in the framework of the F-linear operators it is more natural to deal with F
than with its subfields.

Definition 7.38 If R = RC , then R is said to be centrally closed.

We remark that Z = 0 implies C ⊆ RC . Indeed, λ = (λz−1 )z ∈ RC for every

λ ∈ C, where z ∈ Z \ {0}. Therefore, in this case R is centrally closed if and only if
Z = C. But R may be centrally closed also when Z = 0.

Example 7.39 Every simple ring R, regardless of whether its center is 0 or not, is
centrally closed since λR ⊆ R for all λ ∈ C by (b). A concrete example of a simple
ring with zero center is the ring K of finite rank operators from Example 7.28.

By the term “centrally closed prime algebra” we will mean a centrally closed
prime ring viewed as an algebra over its extended centroid. As pointed out in the
preceding paragraph, the center of such an algebra A is either 0 or coincides with the
extended centroid. In the latter case, A is a central algebra. The notion of a centrally
closed prime algebra can be thus considered as a generalization of the notion of
a central simple algebra. An example of a centrally closed prime algebra that is
not simple is EndF (V ) where V is an infinite dimensional vector space over F (cf.
Example 7.28).

Lemma 7.40 The central closure RC of any prime ring R is a centrally closed prime
algebra over C.

Proof Since RC ⊇ R ∈ J , the primeness of RC follows from Lemma 7.24.

Let us denote the extended centroid of RC by E. By Remark 7.32 we may
consider C as a subfield of E. The lemma will be established by showing that E = C.
Take ε ∈ E. By (b), there exists a nonzero ideal U of RC such that εU ⊆ RC . Note
that I := {x ∈ U ∩ R | εx ∈ R} is an ideal of R. We claim that I = 0. Indeed, choose
u = 0 in U and use Lemma 7.25 to get J ∈ J such that uJ ⊆ R and (εu)J ⊆ R.
Then uJ is contained in I and is not 0 by (c). We can now define the R-bimodule
homomorphism f : I → R, f (x) = εx. By Lemma 7.31 there exists λ ∈ C such that
εx = λx, x ∈ I. This readily implies that εy = λy for every y from the subspace  I of
RC generated by I. Observe that  I is an ideal of RC , so that (c) yields ε = λ ∈ C. 
7.6 Linear (In)dependence in Prime Rings 177

7.6 Linear (In)dependence in Prime Rings

Numerous applications of the Martindale rings of quotients stem from the results
we are about to establish. To get a feel for the topic, we first examine the condition
studied in the next lemma in the simple setting of division rings. Thus, let D be a
division ring, and let a, b ∈ D be such that axb = bxa for every x ∈ D. A natural
possibility where this occurs is when a and b are linearly dependent over the center
of D. Actually, this is the only possibility. Indeed, assuming that a = 0 and then
multiplying axb = bxa by a−1 on both sides we get xba−1 = a−1 bx, which readily
implies that λ := ba−1 (= a−1 b) lies in the center of D and satisfies b = λa. Thus,
we have a simple connection between the multiplicative structure of D (axb = bxa)
and the linear structure of D (b = λa). This little observation can be generalized in
various directions.
We use the same notation as in the preceding sections. In particular, R always
denotes a nonzero prime ring and J the set of all nonzero ideals of R. By (b) and
(c) we refer to conditions of Theorem 7.23.
Lemma 7.41 Let a, b ∈ Qr (R) and let I ∈ J . If axb = bxa for all x ∈ I, then a
and b are linearly dependent over C.
Proof By Lemma 7.25 there exists J ∈ J such that aJ ⊆ R and bJ ⊆ R. By
replacing the role of I by I ∩ J we see that there is no loss of generality in assuming
that aI ⊆ R and bI ⊆ R.
We may assume that a = 0. Then aI = 0 by (c), and hence K := IaI ∈ J . We
claim that the map f : K → R,



f xi ayi = xi byi , xi , yi ∈ I,
i i

is well-defined. Assume that i xi ayi = 0.
 Multiplying
 on the right by zb, z ∈ R, and

using a(yi z)b = b(yi z)a, it follows that x by
i i i za = 0. But then i xi byi = 0
since R is prime. Note also that f indeed maps into R for bI ⊆ R. Obviously, f is an
R-bimodule homomorphism, so it follows from Lemma 7.31 that there exists λ ∈ C
such that f (w) = λw for every w ∈ K. In particular, xby = λxay for all x, y ∈ I. That
is, I(b − λa)I = 0, and hence b = λa by Lemma 7.24. 
Even in the basic case where a, b ∈ R and I = R, the involvement of the extended
centroid C in Lemma 7.41 is indispensable, i.e., it cannot be replaced by the center
Z even when Z is a field. Indeed, a glance at Example 7.37 convinces us of that.
The next lemma treats a considerably more general condition, but, as we shall
see, the proof can be easily reduced to the situation treated in Lemma 7.41.
Lemma 7.42 Let ai , bi ∈ Qr (R) and I ∈ J be such that

n
ai xbi = 0 for all x ∈ I.
i=1
178 7 Rings of Quotients and Structure of PI-Rings

If a1 , . . . , an are linearly independent over C, then each bi = 0. Similarly, if

b1 , . . . , bn are linearly independent over C, then each ai = 0.

Proof Let us only prove the first statement. The second statement can be handled
similarly. Thus, assume that a1 , . . . , an are linearly independent. The n = 1 case
follows from Lemma 7.24. We may therefore assume that the desired conclusion
holds if the number of summands is smaller than n. Suppose that, say, bn = 0. By
(b) we can choose J ∈ J such that bn J ⊆ R. Then xbn y ∈ I whenever x ∈ I
and y ∈ J, so that ni=1 ai (xbn y)bi = 0. Since the last term, an xbn ybn , is equal to
 n−1 
− i=1 ai xbi ybn , we can rewrite this identity as

n−1
ai x(bn ybi − bi ybn ) = 0 for all x ∈ I, y ∈ J.
i=1

The induction assumption yields bn ybi − bi ybn = 0 for all y ∈ J and i = 1, . . . ,

n − 1. As bn = 0 by assumption, Lemma 7.41 tells us that there nexist λi ∈ C
such that
n bi = λ i bn , i = 1, . . . , n − 1. But then we
n can rewrite i=1 ai xbi = 0
as i=1 λi ai xbn = 0, where λn = 1. Hence i=1 λi ai = 0 by Lemma 7.24,
contradicting the linear independence of the ai ’s. 

Lemma 7.42 will play a crucial role in the rest of the chapter. Let us point out that
this lemma generalizes Lemma 1.24, which was used as one of the main tools in the
first chapters.
The following theorem is just a refined version of Lemma 7.42.

Theorem 7.43 Let ai , bi , cj , dj ∈ Qr (R) and I ∈ J be such that

m
ai xbi = cj xdj for all x ∈ I.
i=1 j=1

If a1 , . . . , an are linearly independent over C, then each bi is a C-linear combination

of d1 , . . . , dm . Similarly, if b1 , . . . , bn are linearly independent over C, then each ai
is a C-linear combination of c1 , . . . , cm .

Proof We only prove the first statement. The proof is practically the same as that of
Lemma 4.9, but we give it anyway. Extend {a1 , . . . , an } to a basis of the C-linear
span of {a1 , . . . , an , c1 , . . . , cm }. Denote the additional basis elements (if there are
any) by an+1 , . . . , ap , and write each p cj as a C-linear combination of a1 , . . . , ap . Our
identity can then be rewritten as i=1 ai xbi = 0, where bi , i ≤ n, is the sum of bi
and a linear combination of the dj ’s. Lemma 7.42 shows that each bi = 0. 

Mentioning a result on tensor products (Lemma 4.9) in the last proof was no
coincidence. The above results can be interpreted in terms of tensor products and
multiplication algebras. The next theorem is a generalization of Theorem 4.26 which
considers central simple algebras.
7.6 Linear (In)dependence in Prime Rings 179

Theorem 7.44 If A is a centrally closed prime algebra, then A ⊗ Ao ∼

= M(A) under
the map a ⊗ b → La Rb .

Proof The proof is literally the same as that of Theorem 4.26. The only difference
is that one has to refer to Lemma 7.42 instead of to Lemma 1.24. 

The next theorem and its corollary make a passage to the topics of the next
sections. The theorem shows that the linear dependence of elements in Qr (R) can be
characterized through an identity involving the Capelli polynomials cm (introduced
in Sect. 6.3).

Theorem 7.45 Elements a1 , . . . , am ∈ Qr (R), m ≥ 2, are linearly dependent over

C if and only if cm (a1 , . . . , am , x1 , . . . , xm−1 ) = 0 for all x1 , . . . , xm−1 ∈ R.

Proof Since cm is alternating in the first m indeterminates, the “only if” part follows
from Lemma 6.9. The “if” part will be proved by induction on m. The m = 2 case
is exactly the content of Lemma 7.41 (with I = R). We may thus assume that there
exist x2 , . . . , xm−1 ∈ R such that cm−1 (a1 , . . . , am−1 , x2 , . . . , xm−1 ) = 0. Applying
the formula (6.2) from Sect. 6.3 we see that cm (a1 , . . . , am , x1 , . . . , xm−1 ) = 0 can
be written as

m
(−1)i−1 ai x1 cm−1 (a1 , . . . , ai−1 , ai+1 , . . . , am , x2 , . . . , xm−1 ) = 0,
i=1

and so the linear dependence of a1 , . . . , am follows from Lemma 7.42. 

Note that the theorem remains valid, with the same proof, if the xi ’s are taken
from a nonzero ideal of R.

Corollary 7.46 For every n ≥ 2, the Capelli polynomial cn2 is an identity of every
proper subalgebra of Mn (F), but not of Mn (F) itself.

Proof A proper subalgebra B of Mn (F) has dimension less than n2 , so cn2 is an

identity of B by the “only if” part of Theorem 7.45 (the extended centroid of
Mn (F) is of course F, cf. Example 7.27). Conversely, choosing linearly indepen-
dent a1 , . . . , an2 ∈ Mn (F) it follows by the “if” part that there exist xi ∈ Mn (F) such
that cn2 (a1 , . . . , an2 , x1 , . . . , xn2 −1 ) = 0. 

7.7 Prime GPI-Rings

We now change the subject of discussion. The theory developed in the previous
sections will still be present, but just as a tool used in the proofs. The central topic from
now on will be rings satisfying polynomial identities and generalized polynomial
identities. Although it may appear strange at a glance, we will first treat the latter
and after that the former. This order is more suitable for our approach.
180 7 Rings of Quotients and Structure of PI-Rings

Informally, one can think of a generalized polynomial identity as of a polynomial

identity in which some of the indeterminates are replaced by fixed elements. Thus,
there exist elements aik such that

ai0 xj1 ai1 . . . ain−1 xjn ain = 0 (7.3)

holds for arbitrary xjl .

Example 7.47 Let R be the ring from Example 7.28, i.e., R is any ring such that
K ⊆ R ⊆ EndF (V ) where V is an infinite dimensional vector space over F and K is
the subring of EndF (V ) consisting of all finite rank operators. In fact, what we are
about to say also makes sense if V is finite dimensional (in this case we clearly have
R = EndF (V ) ∼= Mn (F)), we are now just more interested in rings that do not satisfy
nontrivial polynomial identities. Take any rank one operator a ∈ K. One can readily
check that for every x ∈ R there is λx ∈ F such that axa = λx a. Given x, y ∈ R, we
thus have (axa)ya = λx λy a = ay(axa). Thus,

axaya = ayaxa for all x, y ∈ R. (7.4)

This is a model of a generalized polynomial identity.

We will confine ourselves to prime rings. Otherwise axb = 0, x ∈ R, was a
nontrivial generalized polynomial identity, which would make the subject rather
muddled. The standard definition of a prime GPI-ring requires the introduction of
an appropriate generalization of a free algebra, which provides a suitable setting for
“generalized polynomials”. To make our exposition as simple as possible, we will
avoid this and give a rather straightforward definition which involves only multilinear
identities, i.e., such that each xj appears in each term of (7.3) exactly once (as for
example in (7.4)). More precisely, by a multilinear generalized polynomial identity
of a prime ring R we mean an identity of the form

nσ
σ σ σ σ σ
a0i xσ (1) a1i xσ (2) a2i . . . an−1,i xσ (n) ani =0 (7.5)
σ ∈Sn i=1

for all x1 , . . . , xn ∈ R, where atiσ are fixed elements in Qr (R). Assuming that the atiσ ’s
belong to R does not make the treatment much easier, and because of applications it is
really useful to involve Qr (R). Applying the linearization process, which works just
as well as for ordinary polynomial identities, it is possible to show that restricting to
multilinear identities does not cause loss of generality; our definition of a prime GPI-
ring will in fact be equivalent to the standard one (as given in the treatise [BMM96]
of the theory of generalized identities). Before giving it, let us ask ourselves what
should be required in order to consider the identity (7.5) as nontrivial, in the sense
that it can yield some information about the structure of R. Assuming that every
atiσ = 0 is certainly not enough. Say, ax − xa = 0 only means that a is in the center of
R, and says nothing about R. More generally, from Lemma 7.42 (and Theorem 7.43)
7.7 Prime GPI-Rings 181


we know that the identities of the form i ai xbi = 0 are merely consequences of
the linear dependence relations among the ai ’s and the bi ’s. A similar statement can
be made for any multilinear generalized polynomial
 identity in which the xi ’s appear
in the same order in each of the terms, i.e., i a0i x1 a1i x2 a2i . . . an−1,i xn ani = 0 (cf.
Theorem 7.45 and its proof). Therefore, the least one has to require in order to regard
(7.5) as nontrivial is that one of the summations

nσ
σ σ σ σ σ
a0i xσ (1) a1i xσ (2) a2i . . . an−1,i xσ (n) ani (7.6)
i=1

is not always zero (i.e., is not an identity itself). As we shall see, this is already suffi-
cient. Let us call (7.6) the σ-term of (7.5). Every multilinear generalized polynomial
identity is thus the sum of its σ-terms.

Definition 7.48 A nonzero prime ring R is said to be a GPI-ring if there exists a

multilinear generalized polynomial identity of R such that one of its σ-terms is not
an identity of R. For convenience, we also consider R = 0 as a GPI-ring.

The ring R from Example 7.47 is thus a GPI-ring since axaya = 0 for some
x, y ∈ R. We remark that the center of R may be 0 (say, if R = K), and that the
extended centroid of R may properly contain the field of quotients of the center of R
(Example 7.37). This shows that prime GPI-rings do not share all the nice properties
of prime PI-rings that will be studied in the next sections; cf. Theorem 7.56 and
Corollary 7.57.
The study of generalized polynomial identities was initiated by S. A. Amitsur. In
the first paper on the subject, in 1965, he described primitive GPI-rings. Our goal
in this section is to prove a more general result on prime GPI-rings, established by
W. S. Martindale in 1969.
As always, C stands for the extended centroid of a prime ring R, RC for the central
closure of R, and Qr (R) for the right Martindale ring of quotients of R.

Lemma 7.49 If R is a nonzero prime GPI-ring, then there exist a, b ∈ R \ {0} such
that dimC aRC b < ∞.

Proof Among multilinear generalized polynomial identities of R whose σ-terms

are not all identities, choose one with the minimal number of summands. There
must be at least two permutations, σ1 and σ2 , such that the corresponding σ1 -term
and σ2 -term are not identities. As σ1−1 = σ2−1 , we may assume without loss of
generality that σ1−1 (1) < σ1−1 (2) and σ2−1 (2) < σ2−1 (1). As usual, let x1 , . . . , xn be
variables in our identity. Denote by f1 (x1 , . . . , xn ) the sum of all σ-terms such that
σ −1 (1) < σ −1 (2), and by f2 (x1 , . . . , xn ) the sum of all σ-terms such that σ −1 (2) <
σ −1 (1). Thus, f1 (x1 , . . . , xn ) consists of summands in which x1 appears before x2 ,
and f2 (x1 , . . . , xn ) consists of summands in which x2 appears before x1 . Of course,
f1 (x1 , . . . , xn ) + f2 (x1 , . . . , xn ) = 0 for all x1 , . . . , xn ∈ R. In view of the minimality
assumption there exist w1 , . . . , wn ∈ R such that f1 (w1 , . . . , wn ) = 0. Note that the
182 7 Rings of Quotients and Structure of PI-Rings

identity f1 (x, y, w3 , . . . , wn ) = −f2 (x, y, w3 , . . . , wn ), where x, y are arbitrary and

the wi ’s are fixed, can be written as

m
pi xqi yri = sj ytj xuj for all x, y ∈ R (7.7)
i=1 j=1

and some pi , qi , ri , sj , tj , uj ∈ Qr (R). We may assume that {p1 , . . . , pk } is a maximal

linearly independent subset of {p1 , . . . , pl }. Expressing each pi , i > k, as a linear
combination of p1 , . . . , pk , we see that (7.7) gets the form

m
pi xhi (y) = sj ytj xuj for all x, y ∈ R, (7.8)
i=1 j=1

where hi belong to M(Qr (R)), the multiplication algebra of Qr (R). We are now in a
position to apply Theorem 7.43. Accordingly, each hi (y) is a C-linear combination
of u1 , . . . , um . This holds for every y ∈ R, but since the elements in A := RC are
just C-linear combinations of elements in R, it actually holds for every y ∈ A. As the
summations in (7.8) are not 0 if we take x = w1 and y = w2 , hi (y) = 0 for some i
and y. To summarize, the algebra M(Qr (R)) contains nonzero elements that map A
into some finite dimensional spaces. Among them choose h so that it can be written
as h = di=1 Lai Rbi with d minimal. We have to prove that d = 1. Suppose this is not
true, i.e., d > 1. We will derive a contradiction by a similar trick as in the proof of
Lemma 7.42. First observe that the minimality of d implies that the sets {a1 , . . . , ad }
and {b1 , . . . , bd } are linearly independent. By Lemma 7.41 there exists r ∈ R such
that bd rb1 − b1 rbd = 0. Note that h := hRbd r − Rrbd h ∈ M(Qr (R)) can be written
d−1
as h = i=1 Lai Rbi where bi = bd rbi − bi rbd . But this contradicts the assumption
that d is minimal. Namely, dimC h(A) < ∞ implies dimC h (A) < ∞, and the linear
independence of a1 , . . . , ad−1 together with b1 = 0 implies h = 0 by Lemma 7.42.
Thus, there exist nonzero a1 , b1 ∈ Qr (R) such that dimC a1 Ab1 < ∞. This read-
ily implies that dimC a1 rAb1 s < ∞ for any r, s ∈ R. By Theorem 7.23 (b)–(c), we
may choose r and s so that a := a1 r and b := b1 s are nonzero elements in R. 

We are now just a step away from Martindale’s theorem. However, let us pause
for a moment and examine the special case where R is a simple unital ring. This is
needed for applications to PI-rings in the next section.

Lemma 7.50 A simple unital GPI-ring R is a finite dimensional algebra over its
center Z.

Proof Recall that Qr (R) = R and hence C = Z, the center of R (cf. Example 7.27).
By Lemma 7.49 there exist a, b ∈ R \ {0} such that La Rb is a finite rank operator.
Since R is simple
 and unital, u av
j j j = k wk bzk = 1 for some uj , vj , wk , zk ∈ R.
Consequently, j,k Luj Rzk (La Rb )Lvj Rwk is the identity operator, and has finite rank.
This means that [R : Z] < ∞. 
7.7 Prime GPI-Rings 183

Theorem 7.51 (Martindale) A nonzero prime ring R is a GPI-ring if and only if

A := RC is a primitive ring containing an idempotent e such that Ae is a minimal
left ideal of A and eAe is a finite dimensional division algebra over C.

Proof We start with the easier “if” part. Let d := [eAe : C]. Then

sd+1 (ex1 e, . . . , exd+1 e) = 0

for all x1 , . . . , xd+1 ∈ R (see Example 6.15). This is a generalized polynomial identity
whose 1-term ex1 ex2 e . . . exn e is not identically zero since R is prime.
We proceed to the “only if” part. By Lemma 7.49, there exist a, b ∈ R \ {0} such
that dimC aAb < ∞. If L = 0 is a left ideal of A such that L ⊆ Ab and M = 0 is
a right ideal of A such that M ⊆ aA, then ML ⊆ aAb and hence dimC ML < ∞;
moreover, ML = 0 for A is prime (Lemma 7.40). Choose L and M so that ML has
minimal dimension. We claim that AML is a minimal left ideal of A. Let L  be a left
ideal such that 0 = L  ⊆ AML. Then L  ⊆ L, hence ML  ⊆ ML, and so ML  = ML by
the dimension assumption. Consequently, L  ⊇ AML  = AML, proving that AML is
indeed minimal. We can now invoke Lemma 2.58 to obtain an idempotent e ∈ A such
that Ae = AML and eAe is a division algebra. Moreover, using e ∈ AML ⊆ AaAb it
can be easily shown that eAe is finite dimensional over C. Finally, Lemma 5.5 tells
us that A is a primitive ring. 

The results from Sect. 5.4 therefore hold for the central closure RC of a prime
GPI-ring R. Moreover, the associated division ring of RC is a finite dimensional
algebra in this context. A general prime GPI-ring is therefore quite close to the one
considered in Example 7.47.

7.8 Primitive PI-Rings

We proceed with polynomial identities. So far we have considered them only in

algebras, not yet in rings. Thus, the first question that presents itself is how to define
a polynomial identity of a ring. An obvious choice is to give the same definition
as for algebras, but requiring that the noncommutative polynomial in question has
coefficients in Z (instead of in F). This is all right, but the problem, then, is identifying
identities that should be treated as nontrivial. For instance, if a ring R has characteristic
p, then the polynomial pξ vanishes on R. As there is no reason why an algebra of
characteristic p should be a PI-algebra, we cannot regard pξ , in spite of being a
nonzero polynomial in Zξ , as a nontrivial identity of R. Any definition of a nontrivial
polynomial identity of a ring is therefore somewhat technical. One possibility is to
admit those polynomials whose coefficient at one of their monomials of highest
degree is 1. Assuming that such a polynomial is an identity and then employing
the linearization process, one derives a multilinear identity in Zξ1 , ξ2 , . . . such
that one of its coefficients is 1. This can be easily shown by inspecting the proof of
184 7 Rings of Quotients and Structure of PI-Rings

Theorem 6.24, but there is no absolute need for us to bother with details. Let us simply
restrict our attention to multilinear identities. Unlike in Chap. 6, we are now interested
in the structure of rings satisfying some nontrivial polynomial identity, rather than
in particular polynomial identities. Therefore we can afford this restriction without
any loss of generality.

Definition 7.52 A ring R is said to be a PI-ring if there exists

f = kσ ξσ (1) . . . ξσ (n) ∈ Zξ1 , ξ2 , . . .

σ ∈Sn

such that k1 = 1 and f is a (polynomial) identity of R, i.e., f (x1 , . . . , xn ) = 0 for all

x1 , . . . , xn ∈ R.

Of course, it is only important that one of the kσ ’s equals 1. But it is convenient

to require that k1 = 1. This does not harm the generality for we can always rename
the indeterminates.
If R is an algebra over a field, then R is a PI-algebra in the sense of Definition 6.13
if and only if R is a PI-ring in the sense of Definition 7.52. This is not obvious, but
we omit the proof.
We will consider only prime PI-rings. This will make it possible for us to develop
a decent structure theory. Without any restriction on rings we would face examples
such as nilpotent rings, Boolean rings, the Grassmann algebra, etc. This is just too
much to handle.
We will study general prime PI-rings in the next section, while this one is devoted
to a smaller class of primitive PI-rings. What are the examples of such rings? Recall
that there are two basic examples of PI-algebras: commutative algebras and finite
dimensional algebras. Since a commutative primitive ring is a field (Lemma 5.7), and
hence a 1-dimensional algebra over itself, the only immediate examples of primitive
PI-rings are finite dimensional primitive (and hence simple, see Corollary 5.19)
algebras. We will show that these obvious examples are, in fact, the only examples.
For simple unital rings, which form a subclass of primitive rings (see Lemma 5.6),
this follows immediately from  Lemma 7.50. Namely, every prime PI-ring R is also a
GPI-ring. Indeed, the identity σ ∈Sn kσ xσ (1) . . . xσ (n) = 0, k1 = 1, can be regarded
as a multilinear generalized polynomial identity (7.5) with atiσ being integer multiples
of 1; if R = 0, then its 1-term x1 . . . xn is not an identity since R, as a prime ring, is
not nilpotent. Thus:

Lemma 7.53 A simple unital PI-ring R is a finite dimensional algebra over its
center Z.

The assumption that R is unital is actually redundant. Theorem 7.56 below shows
that a simple PI-ring has a nonzero center, from which one easily deduces that it
is automatically unital. Anyway, Lemma 7.53 in its present form is sufficient for
7.8 Primitive PI-rings 185

our purposes. We will use it not only in the proof of the next theorem, but also
in the next section. Since this lemma was (indirectly) derived from Lemma 7.49
which concerns prime GPI-rings, one might wonder whether the proof of the latter
becomes simpler if we adapt it to the case where R is a simple unital PI-ring. However,
besides notational simplifications and the omission of the last paragraph, there are
no essential differences.
The following theorem was proved in 1948 by I. Kaplansky in his pioneering
paper on polynomial identities.

Theorem 7.54 (Kaplansky) A primitive PI-ring R is a finite dimensional simple

algebra over its center Z.

Proof In light of the Jacobson Density Theorem (Theorem 5.16), we may assume
that R is a dense ring of linear operators of a vector space V over a division ring
Δ. We will now mimic the proof of Lemma 6.38. Let f be a polynomial as in
Definition 7.52. Suppose there exist linearly independent vectors u0 , u1 , . . . , un ∈ V .
By density, we can choose h1 , . . . , hn ∈ R so that hi (uj ) = δij uj−1 , 1 ≤ i, j ≤ n.
Then h1 . . . hn (un ) = u0 , while for any product of the hi ’s in a different order
we have hi1 hi2 . . . hin (un ) = 0. Hence f (h1 , . . . , hn )(un ) = u0 = 0, a contra-
diction. Consequently, dimΔ V ≤ n, and so R = EndΔ (V ) (Corollary 5.13).
Thus, R is a simple unital ring (cf. Theorem 3.31 and Example 1.10). Now apply
Lemma 7.53. 

7.9 Prime PI-Rings

The class of prime PI-rings is substantially larger than the class of primitive PI-rings.
First of all, it
 Zcontains
 all commutative domains. Rings from Examples 7.7–7.9,
i.e., Mn (Z), Z 2Z , and Z ⊕ Zi ⊕ Zj ⊕ Zk are all examples of prime PI-rings.
Z
Indeed, proving that they are prime is easy, and, as subrings of finite dimensional
algebras (Mn (Q), M2 (Q), and H, respectively) which satisfy standard polynomials,
they are PI-rings. A description of prime PI-rings is therefore necessarily somewhat
entangled, especially when compared with that of primitive PI-rings. Examples just
given indicate that rings of quotients naturally come into play in this context.
Our first lemma already says a lot, yet it does not give the full picture. We use the
standard notation: R is a nonzero prime ring with center Z, extended centroid C, and
central closure RC .

Lemma 7.55 If R is a prime PI-ring, then A := RC is a finite dimensional central

simple algebra.

Proof Let U be a nonzero ideal of A. Every multilinear  identity of R is obviously

also an identity of A, and hence also of U. Let f = σ ∈Sn kσ ξσ (1) . . . ξσ (n) ∈
Zξ1 , ξ2 , . . ., k1 = 1, be a multilinear identity of U of minimal degree n.
186 7 Rings of Quotients and Structure of PI-Rings

Write

f = gξn + gi ξn mi , (7.9)
i

where g, gi are multilinear polynomials and each mi is a monomial with leading

coefficient 1 and of degree at least 1. Since g has degree n − 1 and ξ1 . . . ξn−1 is
one of its monomials, g is not an identity of U. Let u1 , . . . , un−1 ∈ U be such
that u := g(u1 , . .
. , un−1 ) = 0. According to (7.9), f (u1 , . . . , un−1 , x) = 0 can be
written as ux = vi xwi for all x ∈ U and some vi ∈ A ∪ Z, wi ∈ U. This in
particular holds for all x in R ∩ U, which is an ideal of R different from 0 (the latter
follows from Theorem 7.23). Writing ux as ux1, we thus infer from Theorem 7.43
that 1 is a C-linear
 combination of the wi ’s. Therefore 1 ∈ A and so C ⊆ A,
yielding 1 ∈ i Cwi ⊆ U. Thus U = A, showing that A is a simple unital algebra.
By Lemma 7.53, A is finite dimensional over its center (which is equal to C by
Lemma 7.30). 
Given an arbitrary ring R and its proper ideal I, it is quite likely that I does not
contain nonzero elements from the center Z of R. For example, this is certainly the
case if Z is a field. On the other hand, take Mn (Z), which is a typical example of a
prime PI-ring. One can easily check that its nonzero ideals have the form Mn (kZ) for
some k ∈ N (cf. Corollary 4.43). Therefore each of them contains nonzero integer
multiples of the identity matrix. Thus, nonzero ideals of Mn (Z) intersect the center
of Mn (Z) nontrivially. This is actually true for every prime PI-ring:
Theorem 7.56 If I = 0 is an ideal of a prime PI-ring R, then I ∩ Z = 0.
Proof Take an arbitrary nonzero C-linear functional γ on A := RC . Since we know
by Lemma 7.55 that A is a finite dimensional central simple algebra, Lemma 1.25
nexists g ∈ M(A) such that g(x) = γ (x)1, x ∈ A. Choose ai , bi ∈ A
shows that there
so that g = i=1 Lai Rbi and the sets {a1 , . . . , an } and {b1 , . . . , bn } are linearly
independent (cf. Remark 1.23). Lemma 7.35 implies that there exists a nonzero ideal
J of R such that, in particular, ai J ⊆ R and Jbi ⊆ R for every i. Then K := JIJ is also
a nonzero ideal of R and satisfies g(K) ⊆ I ∩ C. Since g(K) = 0 by Lemma 7.42, it
follows that I ∩ C = 0. Of course, I ∩ C = I ∩ Z for I ⊆ R. 
Theorem 7.56 actually holds even for semiprime rings. This was proved by
L. H. Rowen in 1973 (for the proof see, e.g., [Row08, p. 419]).
Corollary 7.57 If R is a prime PI-ring, then Z = 0 and C =  Z, the field of quotients
of Z. Moreover, if Z is a field, then R is a finite dimensional central simple algebra.
Proof Taking R for I in Theorem 7.56 we obtain Z = 0. Let λ ∈ C. By (b) (of
Theorem 7.23) there exists an ideal I = 0 of R such that λI ⊆ R. Using Theorem 7.56
again we obtain a nonzero z ∈ I ∩ Z. Hence λz ∈ R ∩ C = Z, and so λ = z z−1 with
z, z ∈ Z. This proves that C = 
Z. Finally, if Z is a field, then Z = 
Z = C, and hence
R = RC is a finite dimensional central simple algebra by Lemma 7.55. 
7.9 Prime PI-rings 187

The next theorem sharpens Lemma 7.55. It is usually attributed to E. Posner who
proved a slightly weaker version in 1960.

Theorem 7.58 (Posner) If R is a prime PI-ring, then QZ (R) is a finite dimensional

central simple algebra.

Proof Corollary 7.57, together with Lemma 7.36, shows that QZ (R) ∼
= RC . Now use
Lemma 7.55. 

By Wedderburn’s theorem, QZ (R) is isomorphic Mn (D) for some finite dimen-

sional central division algebra D and n ∈ N. The point of Theorem 7.58 is that every
prime PI-ring is very close to such an algebra. Namely, every element in QZ (R) can
be written as rz−1 where r ∈ R and z ∈ Z. For example, every matrix in Mn (Q) is of
the form 1z T , where z ∈ Z \ {0} and T ∈ Mn (Z).
The converse to Theorem 7.58 is trivially true: If QZ (R) is finite dimensional,
then it satisfies a standard identity so that it is a PI-ring, and so is its subring R. A
somewhat rougher characterization of prime PI-rings reads as follows.

Corollary 7.59 A prime ring R is a PI-ring if and only if R can be embedded into
Mm (K) for some m ∈ N and some field K.

Proof The “if” part is clear. Let R be a PI-ring. Then QZ (R) is a finite dimensional
simple algebra over its center 
Z. Let K be a splitting field for QZ (R) (for example,
we can take an algebraic closure of Z). The scalar extension of QZ (R) to K is then
isomorphic to Mm (K). 

Remark 7.60 If R is a prime PI-algebra over a field F, then F is a subfield of 

Z (see
Remark 7.32), and so K is an extension field of F.

Corollary 7.59 implies that every prime PI-ring R satisfies a standard polynomial.
In fact, if R embeds in Mm (K), then, by the Amitsur-Levitzki Theorem, R satisfies
s2m .
As observed in the proof of Theorem 7.58, QZ (R) ∼= RC holds for a prime PI-ring.
For clarification, we add the following corollary.

Corollary 7.61 If R is a prime PI-ring, then QZ (R) ∼

= Qr (R) ∼
= Qcr (R).

Proof It is clear that QZ (R) satisfies conditions (a)–(c) of Theorem 7.23. It remains to
prove (d). Take a nonzero ideal I of R and a right R-module homomorphism f : I → R.
By Theorem 7.56 there exists 0 = c ∈ I ∩Z. We have f (c)x = f (cx) = f (xc) = f (x)c
for every x ∈ I, and so f (x) = c−1 f (c)x. Since c−1 f (c) ∈ QZ (R), this proves that
QZ (R) ∼= Qr (R).
To prove that Qcr (R) ∼= QZ (R), we have to verify that QZ (R) satisfies conditions
of Definition 7.10. Only the fulfillment of (b) is not entirely obvious. Take a regular
element s in R. Then s is clearly also regular in QZ (R). But then, since QZ (R) is a
finite dimensional unital algebra, s is invertible in QZ (R) (cf. Remark 1.31 (b)). This
proves (b). 
188 7 Rings of Quotients and Structure of PI-Rings

7.10 Central Polynomials

In this final section we apply the above theory to a different theme.

Definition 7.62 Let A be an algebra over a field F. A polynomial f (ξ1 , . . . , ξn ) ∈

Fξ1 , ξ2 , . . . with zero constant term is said to be a central polynomial for A if
f (x1 , . . . , xn ) lies in the center of A for all x1 , . . . , xn ∈ A, but f is not an identity
of A.

In other words, f (ξ1 , . . . , ξn ) is a central polynomial if f is not an identity, but

[ f , ξn+1 ] is. This shows that central polynomials can exist only on PI-algebras. The
condition that f must have zero constant term excludes the trivial possibility where
f is a sum of a constant polynomial and an identity.

Example 7.63 The polynomial [ξ1 , ξ2 ] is a central polynomial for the Grassmann
algebra G; see Example 6.16.

Example 7.64 The polynomial [ξ1 , ξ2 ]2 is a central polynomial for the algebra
M2 (F). One can check this by a short calculation. On the other hand, this follows
from the Cayley-Hamilton Theorem applied to the commutator of two 2 × 2 matri-
ces. Indeed, recall that the characteristic polynomial is given by the formula (6.10)
from Sect. 6.8, and that the trace of a commutator is 0. A similar, but multilinear
polynomial [ξ1 , ξ2 ][ξ3 , ξ4 ] + [ξ3 , ξ4 ][ξ1 , ξ2 ] is also central for M2 (F).

There is no obvious way to generalize Example 7.64 to matrices of larger size. In

fact, finding examples of central polynomials for Mn (F) is not an easy task. However,
they do exist, as we will soon find out.
Let us recall a few facts about the T-ideal In := Id(Mn (F)) from Sect. 6.7. We
assume, until further notice, that F is an infinite field. The relatively free algebra
FX/In is then isomorphic to the algebra of n × n generic matrices GMn (F); this is
(i)
the subalgebra of Mn (F[Ω]), Ω = {Ωjk | j, k = 1, . . . , n, i = 1, 2, . . .}, generated
by all n × n generic matrices g(i) = (Ωjk(i) ), i ∈ N, and the identity matrix (Example
6.34). The algebras Mn (F) and GMn (F) satisfy the same polynomial identities
(Lemma 6.32).
The next theorem was established independently by E. Formanek and Yu. P.
Razmyslov in the early 1970s.

Theorem 7.65 For every n ∈ N, there exists a central polynomial for Mn (F).

Proof Let F(Ω) be the field of rational functions in Ω (i.e., the field of quo-
tients of F[Ω]). Consider GMn (F) as a subring of the F(Ω)-algebra Mn (F(Ω)).
Denote by GMn (F) the F(Ω)-subalgebra of Mn (F(Ω)) generated by GMn (F).
This is, of course, just the linear span of GMn (F). Corollary 7.46 states that the
Capelli polynomial cn2 is not an identity of Mn (F), therefore it is not an iden-
tity of GMn (F), and neither of GMn (F). Using Corollary 7.46 again it follows
that GMn (F) = Mn (F(Ω)). This implies that GMn (F) is a prime ring. Indeed,
7.10 Central Polynomials 189

if a, b ∈ GMn (F) satisfy axb = 0 for every x ∈ GMn (F), then this obviously
also holds for every x ∈ GMn (F) = Mn (F(Ω)), and hence a = 0 or b = 0 since
Mn (F(Ω)) is prime.
Thus, GMn (F) is a prime PI-ring. The F-subalgebra of GMn (F) generated by all
generic matrices g(i) (but without the identity matrix) is clearly an ideal of GMn (F).
Theorem 7.56 implies that it contains a nonzero central element. Hence, the center
of GMn (F) contains elements different from scalar multiples of unity. The same
therefore holds for the isomorphic algebra FX/In . Let f + In be an element of the
center of FX/In which is not of the form α + In , α ∈ F. Note that f − α0 , where
α0 is the constant term of f , is a central polynomial for Mn (F). 

With some extra effort one can prove Amitsur’s theorem saying that GMn (F) is not
only prime, but a domain (see, e.g., [Row08, p. 467]). Consequently, QZ (GMn (F))
is also a domain; moreover, Theorem 7.58 thus implies that it is a finite dimensional
central division algebra. It is called the universal division algebra of degree n.
Requiring the infinity of F in our proof of Theorem 7.65 is the price we had to
pay for relying merely on the abstract theory. Constructive proofs, such as given by
Formanek and Razmyslov, work regardless of the cardinality of F. (On the other
hand, the case where F is finite can be reduced, by elementary means, to the case
where F is infinite.)
As indicated in the last two paragraphs, there is much more to say. But every book
must have an end.

Exercises

7.1. Let R be a ring such that its center Z is nonzero and all elements in Z \ {0} are
regular. Show that QZ (S) ∼ = QZ (R) for every ring S such that R ⊆ S ⊆ QZ (R).
7.2. Find an example of a ring R such that R  QZ (R)  Qrc (R).
7.3. Let F ⊆ K be fields, and let T be a ring such
 that F ⊆ T ⊆ K and K is the
field of quotients of T . Note that R = F0 TT is a ring under the standard matrix
operations. Prove that R satisfies the right Ore condition and describe its right
classical ring of quotients. Show by an example that R does not necessarily
satisfy the left Ore condition.
7.4. Show that every PI-algebra A satisfies a nonzero identity in two indeterminates.
Hence derive that if A is a domain then it is a right (and left) Ore domain.
Remark: The latter can be also deduced from Proposition 7.19, as well as from
the structure theory (cf. Corollary 7.61).
7.5. Let I be a nonzero ideal of a prime ring R. Show that Qr (I) ∼
= Qr (R).
7.6. Show that an automorphism of a nonzero prime ring R can be uniquely extended
to an automorphism of Qr (R).
190 7 Rings of Quotients and Structure of PI-Rings

7.7. Let ϕ be an automorphism of a prime ring R, and let 0 = q ∈ Qr (R) be such

that qϕ(x) = xq for all x ∈ R. Show that q is invertible in Qr (R) (and so
ϕ(x) = q−1 xq for all x ∈ R).
7.8. Let R = 0 be a prime ring. Denote by Qs (R) the set of all q ∈ Qr (R) for which
there exists a nonzero ideal J of R such that Jq ⊆ R. Show that Qs (R) is a
subring of Qr (R) which contains R and whose center is the extended centroid C
(thus, R ⊆ RC ⊆ Qs (R) ⊆ Qr (R)). We call Qs (R) the symmetric Martindale
ring of quotients of R. Note that q from the preceding exercise actually lies in
Qs (R). Show that if R is a domain,then so is Qs (R).
7.9. Show that ξ is a zero-divisor in Qr (Fξ, η). Thus, Qr (Fξ, η) is not a domain
and it properly contains Fξ, η. Show, however, that Fξ, η is centrally closed.

7.10. Let R be a ring. Note that the set of all R-bimodule homomorphisms from R
into R forms a ring under the standard operations. We denote it by W and call
it the centroid of R. Show that:
(a) If R is unital, then W is isomorphic to the center Z of R.
(b) If R2 = R, then W is commutative.
(c) If R simple, then W is isomorphic to the extended centroid C of R.
(d) If R is prime, then W can be embedded into C; show by an example that
W is not necessarily isomorphic to C.
7.11. Let d and d  be derivations of a ring R. Check that dd  −d  d is again a derivation,
while there is no reason to believe that dd  is a derivation. In fact, show that
if R is prime and char(R) = 2, then dd  is a derivation only when d = 0 or
d  = 0. The assumption on char(R) is necessary; show that if R is semiprime,
then d 2 is a derivation if and only if 2d = 0.
Remark: The first of these two results was obtained by E. Posner in 1957,
before the discovery of the extended centroid. However, the proof becomes
shorter by making use of it.
7.12. Let R be a prime ring with extended centroid C. Suppose that a nonzero
derivation d of R and q ∈ Qr (R) \ C satisfy d(x)q = −qd(x) for all x ∈ R.
Show that there exists λ ∈ C such that d(x) = λ[q, x] for all x ∈ R, and that
q2 ∈ C.
7.13. Let R be a ring. A biadditive map D : R ×R → R is called a biderivation if the
maps x → D(x, y) and x → D(y, x) are derivations for every y ∈ R. Suppose
that R is prime and noncommutative. Show that then there exists λ from the
extended centroid C such that D(x, y) = λ[x, y] for all x, y ∈ R. Hence derive
that every additive map f : R → R satisfying f (x)x = xf (x) for every x ∈ R is
of the form f (x) = λx + μ(x) where λ ∈ C and μ : R → C.
Hint: Begin by computing D(xy, zw) in two different ways.
7.14. Let R be a domain. Show that if R is a GPI-ring, then it is a PI-ring.
7.15. Let R be a simple unital ring. Prove that R is isomorphic to the ring of n × n
matrices over a field if and only if there exists 0 = a ∈ R such that axaya =
ayaxa for all x, y ∈ R.
Exercises 191

7.16. Let R be a prime ring, and let ai ∈ R. Prove that a1 xa2 xa3 = 0 for all x ∈ R
implies that at least one ai is 0. Show by an example that a1 xa2 xa3 xa4 = 0 for
all x ∈ R does not, in general, imply that one of the ai ’s is 0.
Hint: Theorem 7.51 suggests where to look for examples. However, they exist
only under a strong restriction on char(R).
7.17. True or False:
(a) Every nonzero ideal of a prime PI-ring contains a regular element.
(b) Every nonzero ideal of a prime GPI-ring contains a regular element.
(c) Every left zero-divisor in a prime PI-ring is also a right zero-divisor.
(d) Every left zero-divisor in a prime GPI-ring is also a right zero-divisor.
7.18. True or False:
(a) If a nonzero ideal of a prime ring R is a PI-ring, then R is a PI-ring.
(b) If a nonzero left ideal of a prime ring R is a PI-ring, then R is a PI-ring.
7.19. Show that a prime PI-ring with finite center is isomorphic to the ring of n × n
matrices over a finite field.
7.20. Let R be prime PI-ring, let  Z be the field of quotients of the center Z of R, and
let n ∈ N. Prove that the following statements are equivalent:
(i) QZ (R) is a central simple algebra over  Z with dim Z QZ (R) = n .
2

(ii) R satisfies s2n , but does not satisfy nonzero identities of degree less than
2n.
(iii) There exists a field K such that R can be embedded in Mn (K), and Mn (K)
satisfies the same multilinear identities as R (accordingly, R cannot be
embedded in Mn−1 (H) for any commutative ring H).
(iv) The degree of algebraicity over  Z of every element in R is at most n, and
there exist elements in R whose degree of algebraicity is n.
7.21. A polynomial f (ξ1 , . . . , ξm ) ∈ Fξ1 , ξ2 , . . . is called a weak identity of
Mn (F) if f (a1 , . . . , am ) = 0 for all trace zero matrices ai ∈ Mn (F). Find a
weak identity of M2 (F) of degree 3.
Hint: Example 7.64.
7.22. Show that [ξ1 , ξ2 ] cannot be a central polynomial for a prime algebra.
Remark: This can be proved by applying the theory, but also directly, by
elementary means.
7.23. Show that a central polynomial for Mn (F) is an identity of Mn−1 (F).
References

[Bea99] Beachy, J.A.: Introductory Lectures on Rings and Modules. London Mathematical
Society Student Texts, vol. 47. Cambridge University Press, Cambridge (1999)
[BMM96] Beidar, K.I., Martindale 3rd, W.S., Mikhalev, A.V.: Rings with Generalized Identi-
ties. Monographs and Textbooks in Pure and Applied Mathematics, vol. 196. Marcel
Dekker Inc, New York (1996)
[FD93] Farb, B., Dennis, A.V.: Noncommutative Algebra. Graduate Texts in Mathematics,
vol. 144. Springer, Berlin (1993)
[Her68] Herstein, I.N.: Noncommutative Rings. Carus Monographs in Mathematics, vol. 15.
The Mathematical Association of America, Eau Claire (1968)
[Hun74] Hungerford, T.W.: Algebra. Graduate Texts in Mathematics, vol. 73. Springer, Berlin
(1974)
[Lam01] Lam, T.Y.: A First Course in Noncommutative Rings. Graduate Texts in Mathematics,
vol. 131, 2nd edn. Springer, New York (2001)
[Lam95] Lam, T.Y.: Exercises in Classical Ring Theory. Problem Books in Mathematics.
Springer, New York (1995)
[Lam99] Lam, T.Y.: Lectures on Modules and Rings. Graduate Texts in Mathematics, vol. 189.
Springer, New York (1999)
[MR01] McConnell, J.C., Robson, J.C.: Noncommutative Noetherian Rings. Revised Edition.
Graduate Studies in Mathematics, vol. 30. American Mathematical Society, Provi-
dence (2001)
[McC04] McCrimmon, K.: A Taste of Jordan Algebras. Universitext. Springer, New York
(2004)
[Pie82] Pierce, R.S.: Associative Algebras. Graduate Texts in Mathematics, vol. 189. Springer,
New York (1982)
[Row91] Rowen, L.H.: Ring Theory, Student edn. Academic Press, San Diego (1991)
[Row08] Rowen, L.H.: Graduate Algebra: Noncommutative View. Graduate Studies in Math-
ematics, vol. 91. American Mathematical Society, Providence (2008)
[ZSSS82] Zhevlakov, K.A., Slinko, A.M., Shestakov, I.P., Shirshov, A.I.: Rings that are Nearly
Associative. Pure and Applied Mathematics, vol. 104. Academic Press, New York
(1982)

© Springer International Publishing Switzerland 2014 193

M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4
Index

A Biderivation, 190
Abelian group, xxii Bilinear map, xxxii
Additive group, xxii Bimodule, 54
Additive map, xxiii Boolean ring, 131, 147, 184
Additive subgroup, xxvi Brauer group, 103
Algebra, xxxiii Brauer, R., 103
Algebra homomorphism, xxxiv Burnside’s theorem, 115, 134
Algebraic algebra, 5 Burnside, W., 115
Algebraic algebra of bounded degree, 160
Algebraic closure, xxxvii
Algebraic element, xxxvi, 5 C
Algebraically closed field, xxxvii Capelli polynomial, 144, 179, 188
Alternating polynomial, 143 Cartesian product, xx
Alternative algebra, 5 Cayley’s theorem, xxiii, 33
Amitsur, S. A., 128, 157, 160, 181, 189 Cayley-Hamilton Theorem, 154, 161, 188
Amitsur-Levitzki Theorem, 143, 153, 157, Center of a group, xxiii
161, 187 Center of a ring, xxv
Annihilator (left, right) of a subset of a ring, Center of an algebra, xxxiv
76 Central algebra, 9
Annihilator of a module, 56 Central closure, 175, 176, 181, 183, 185
Antiautomorphism, 23 Central element, xxv
Antiisomorphism, 63 Central idempotent, 40
Artin, E., 71, 115 Central polynomial, 188, 191
Artin-Whaples Theorem, 115, 134 Central simple algebra, 10, 12, 13, 18,
Artinian module, 72, 76, 77 22, 44, 46, 48, 49, 51, 91, 92,
Artinian ring, 72, 73, 77 94–98, 100–102, 104–106, 151,
Ascending chain condition, 72 178, 185–187, 191
Augmentation ideal, 36, 51 Centralizer, 92, 94, 99, 100, 106
Automorphism, xxiii, xxvi, 57 Centrally closed, 176, 179, 190
Axiom of choice, xxi, 72 Centroid, 190
Chain, xx
Characteristic of a field, xxvi
B Characteristic of a prime ring, 29
Base field, xxxvi Characteristic of a ring, xxvi
Basis, xxxii, 59 Characteristic of an algebra, xxxiv
Bergman, G., 159 Characteristic polynomial, 154
Biadditive map, xxiii Class formula, 17
© Springer International Publishing Switzerland 2014 195
M. Brešar, Introduction to Noncommutative Algebra, Universitext,
DOI 10.1007/978-3-319-08693-4
196 Index

Classical ring of quotients, 166, 168, 169, Evaluation, xxxiv, 139, 150
171, 187, 189 Extended centroid, 163, 174–179, 181, 185,
Coefficient of a polynomial, xxx, 138 190
Commutative ring, xxv Extension field, xxxvi
Commutator, 8
Complex algebra, xxxiii
Complex space, xxxi F
Conjugate element (in a group), xxiii Factor algebra, xxxiv
Conjugate quaternion, 4 Factor group, xxiv
Constant polynomial, xxx, 138 Factor module, 58
Constant term, xxx, 138 Factor ring, xxviii
Corner ring, 39 Faithful module, 56
Coset, xxiii Field, xxv
Cyclic module, 57 Field of quotients, xxxvi, 156, 163–165, 171,
173, 175, 181, 186, 188, 189, 191
Field of rational functions, xxxvi
D Finite characteristic, xxvi
Degree of a polynomial, xxx, 139 Finite dimensional, xxxii, xxxiii, 61
Degree of a polynomial in an indeterminate, Finite extension, xxxvi
148 Finite field, xxxvi
Degree of algebraicity, xxxvi, 5 Finite order, xxiii
Dense ring of linear operators, 110 Finite rank, xxxiii, 110
Derivation, 22, 34–35, 190 Finitely generated, xxvii
Derivation (generalized), 105 Fitting’s lemma, 77
Descending chain condition, 72 Formal power series ring, 19, 126
Dimension, xxxii, xxxiii, 61 Formanek, E., 188
Direct product of algebras, xxxiv Formanek-Razmyslov Theorem, viii, 188
Direct product of groups, xxiv Free algebra, vii, 118, 138–161, 170, 180,
Direct product of modules, 58 183–191
Direct product of rings, xxviii Free module, 59, 60, 63, 75
Direct sum of rings, xxix Free monoid, 138
Direct sum of subgroups, xxiv Frobenius’ theorem, vii, 1, 4–6, 99, 101
Direct sum of submodules, 58 Frobenius, F. G., 4
Direct summand, 59 Full matrix ring, xxix
Division algebra, xxxiv Fundamental Theorem of Algebra, xxxvii
Division ring, xxv
Division subring, xxvii
Dixmier’s conjecture, 142 G
Dixmier, J., 142 General linear group, xxii, 46
Domain, xxxv, 28–31, 35, 36, 40, 44, Generalized polynomial identity, 163,
47, 49–51, 118, 122, 126, 128, 179–184
139, 142, 154, 160, 163–166, Generators, xxvii
169–171, 173, 189, 190 Generators and relations, 141
Double centralizer theorem, vii, 79, 99, 101, Generic matrix, 153, 154, 156, 160, 188, 189
106 Goldie, A. W., 171
Golod, E. S., 21
GPI-ring, 181–185, 190, 191
E Grassmann algebra, 142–143, 146, 158–160,
Elementary symmetric polynomials, 154 184, 188
Embedding, xxiii, xxvi Group, xxi
Empty word, 137 Group algebra, 35–37, 46, 50, 51, 89, 128,
Endomorphism, xxiii, xxvi, 57 160
Endomorphism ring, 57 Group homomorphism, xxii
Equivalence relation, xx Group ring, 35
Index 197

H Left unity, 22
Hamilton, W. R., 4 Length of a word, 139
Herstein, I. N., 133, 136 Levitzki, J., 157
Hilbert’s basis theorem, 74 Lie algebra, 159
Homogeneous polynomial, 139 Lie polynomial, 159
Hua’s identity, 21 Linear combination, xxxii, 60
Linear dependence, xxxii, 59
Linear functional, xxxii
I Linear independence, xxxii, 59
Ideal, xxvii, xxxiii Linear map (operator), xxxii, 61
Idempotent, 38 Linear span, xxxii, 60
Identity element (left, right), xxi Linearization, 147
Indeterminate, xxx, 138 Local ring, 135
Inner automorphism, 13, 22, 98 Lower nilradical, 49
Inner derivation, 22, 34, 131, 133
Inverse (left, right), xxi
Invertible (left, right), xxi M
Involution, 22, 50 Martindale, W. S., 171, 181
Irreducible representation, 64 Martindale ring of quotients, 171–179, 181,
Isomorphism, xxii, xxvi, 57 187, 190
Isomorphism theorems, 75 Martindale’s theorem, viii, 182, 183
Maschke’s theorem, 36, 46, 128
Maschke, H., 36
J Matrix units (standard), 7, 37
Jacobi identity, 159 Maximal element, xxi
Jacobian conjecture, 142 Maximal ideal, 65, 76, 119–122, 127, 128
Jacobson Density Theorem, vii, 110–118, Maximal left ideal, 65, 66, 120, 121,
129, 130, 132, 134, 185 124–126, 135
Jacobson radical, vii, 107, 114, 119, Maximal subfield, 15–17, 23, 96, 101
121–136, 160 Minimal ideal, 41, 173
Jacobson, N., 33, 107, 115, 131, 133, 136 Minimal left ideal, 41–42, 51, 64, 65, 67,
Jordan algebra, 159 68, 70, 71, 73, 107, 109, 110, 114,
Jordan normal form, xxxvii 116–118, 134, 183
Jordan polynomial, 159 Module homomorphism, 57
Module over a ring, 53
Module over an algebra, 55
K Molien, T., 45
Köthe’s problem, 26, 49, 128 Monoid, xxi
Köthe, G., 26 Monoid algebra, 35, 138
Kaplansky, I., 185 Monomial, xxxi, 139
Kleinecke, D., 34 Multilinear polynomial, 139
Kronecker (tensor) product of matrices, 104 Multiplication algebra, 11–13, 22, 91, 92,
Kronecker delta, 37 104, 115, 179, 182, 186
Kurosh, A. G., 21 Multiplication table, xxxiv

L N
Lagrange’s theorem, xxiii, 17 Nagahara, T., 18
Laurent polynomials, 160 Nagata-Higman Theorem, 146
Laurent series ring, 20 Nakayama’s lemma, 134
Leading coefficient, xxx Newton’s formulas, 155
Left (right) vector space, 60 Newton, I., 154
Left ideal, xxvii Nil ideal, 26, 27, 49, 51, 121, 128
Left multiplication map, 11 Nilpotent element, 26
198 Index

Nilpotent ideal, 25–30, 36, 37, 45, 105, 121, Q

122, 126, 143 Quasi-inverse (left, right), 123
Noether, E., 13, 75 Quasi-invertible (left, right), 122
Noetherian module, 72, 76, 77 Quaternion algebra (generalized), 18
Noetherian ring, 73–75, 77, 142, 160, 170, Quaternion group, 51
171 Quaternions, 4
Nonassociative algebra, 5, 159
Nonassociative division algebra, 5
Noncommutative polynomial, 138 R
Noncommutative localization, 168 Radical of a finite dimensional algebra, 27,
Noncommutative ring, xix 28, 45, 49, 103, 121, 122
Nontrivial idempotent, 39 Radical ring, 127, 128, 130, 135
Normal subgroup, xxiii Rank, xxxiii, 76, 110
Razmyslov, Yu. P., 161, 188
Real algebra, xxxiii
O Real space, xxxi
Octonions, 5 Regev, A., 147
Opposite ring, 63 Regular element, 164
Order, xxiii Regular representation, 32
Ore condition, 166–170, 189 Relatively free algebra, 153, 188
Ore domain, 169–171, 189 Representation of a group, 46
Ore, O., 166 Representation of an algebra, 55
Orthogonal idempotents, 38 Right multiplication map, 11
Outer automorphism, 13, 22 Right ideal, xxvii
Ring, xxv
Ring homomorphism, xxvi
P Ring ideal of an algebra, 7
Partially ordered set, xx Ring of central quotients, 163–166, 187
Peirce decomposition, 39 Root of a polynomial, xxxvi
Permutation, xxii Root of unity, xxxvi
PI-algebra, 145–147, 153, 160, 183, 184, Rosset, S., 157
187–189 Rowen, L. H., 186
PI-ring, 147, 184–187, 189–191
Poincaré-Birkhoff-Witt Theorem, 159
Polynomial, xxx, 138 S
Polynomial identity, 145, 184 Scalar, xxxi
Polynomial in several indeterminates, xxx Scalar extension, 48, 94, 95, 97, 104, 105,
Polynomial ring, xxx 149, 151, 187
Posner’s theorem, viii, 187 Scalar multiplication, xxxi
Posner, E., 187, 190 Schur’s Lemma, 67, 76, 108, 111
Prime ideal, 49 Schur, I., 67
Prime radical, 49 Semigroup, xxi
Prime ring, 28–32, 42–46, 49, 50, 69, 104, Semigroup algebra, 35
107, 109, 110, 114, 117, 118, 134, Semiprime ring, 28–31, 36, 37, 41, 42, 44,
146, 165, 171–191 45, 49–51, 67, 70, 76, 105, 116,
Prime subfield, xxvi 128, 161, 171, 186, 190
Primitive ideal, 119–121, 127, 129, 130, 132, Semiprimitive ring, 127–130, 132, 135, 136
135, 136 Semisimple module, 68, 69, 71
Primitive idempotent, 51 Semisimple ring, 45, 71
Primitive ring, vii, 136, 140, 160, 181, Shafarevitch, I. R., 21
183–185 Sign of permutation, xxii
Procesi, C., 161 Similar matrices, xxxvii
Product of ideals, xxvii Simple module, 53, 64–69, 71, 76, 107–109,
Projective module, 75 111, 121, 124, 127, 128, 134
Index 199

Simple ring (see also central simple algebra), Unital homomorphism, 138
6–10, 22, 28, 30, 31, 36, 42, 44, Unital module, 53
46, 65, 69, 71, 73, 76, 96, 97, 104, Unital ring, xxv
105, 109, 110, 114, 142, 160, 173, Unitization, 31
176, 182, 184, 185, 190 Unity, xxv
Simple tensor, 82 Universal division algebra, 189
Skew Laurent series ring, 20 Universal property, 82, 83, 91, 139–142, 150,
Skew polynomial ring, 20 164, 168
Skew power series ring, 20 Upper bound, xxi
Skew-symmetric element, 23 Upper nilradical, 49
Skolem, T., 13 Upper triangular matrix algebra, 26, 27, 126,
Skolem-Noether Theorem, vii, 13, 79, 98, 99 134, 135, 146
Smoktunowicz, A., 109
Socle, 116–118, 160
Splitting field, 95
V
Stable identity, 150
Variety, 152
Standard polynomial, 144, 146, 157, 161,
Vector, xxxi
185, 187
Vector space, xxxi, 60
Subalgebra, xxxiii
Subdirect product, 129
Subfield, xxvii
Subgroup, xxiii W
Submodule, 56 Weak identity, 191
Subring, xxvi Wedderburn’s principal theorem, 45
Subspace, xxxii, 60 Wedderburn’s structure theorems, vii, 25,
Sum of ideals, xxvii, xxviii 42–46, 53, 55, 63, 69–72, 90, 96,
Sum of submodules, 58 107, 113, 114, 118, 129, 187
Symmetric element, 23 Wedderburn’s theorem on finite division
Symmetric group, xxii rings, vii, 1, 16–18, 93, 99, 132,
Symmetric polynomial, 154 133
Wedderburn, J. H. M., 17, 42, 71, 107
Wedderburn-Artin theory, 71, 73
T Weyl algebra, 8–10, 22, 30, 32, 44, 74, 110,
T-ideal, 152–188 118, 134, 141, 142, 160, 171
Tensor product of algebras, 88 Whaples, G., 115
Tensor product of linear maps, 82 Witt, E., 18
Tensor product of vector spaces, 80 Word, 137
Tominaga, H., 18
Trace identity, 161
Z
U Zero-divisor (left, right), xxv
Unital algebra, xxxiv Zorn’s lemma, xx, xxi, 60, 66, 68, 69, 124