ALGEBRAIC THEORIES

A CATEGORICAL INTRODUCTION

TO GENERAL ALGEBRA

J. Adámek, J. Rosický, E. M. Vitale

With a Foreword by F. W. Lawvere

November 20, 2009 ii
Contents

Foreword v

Introduction ix

0 Preliminaries 1

1 Algebraic theories and algebraic categories 7

2 Sifted and filtered colimits 15

3 Reflexive coequalizers 23

4 Algebraic categories as free completions 31

5 Properties of algebras 39

6 A characterization of algebraic categories 47

7 From filtered to sifted 57

8 Canonical theories 65

9 Algebraic functors 71

10 Birkhoff ’s Variety Theorem 79

11 One-sorted algebraic categories 89

12 Algebras for an endofunctor 103

13 Equational categories of Σ-algebras 111

14 S-sorted algebraic categories 123

15 Morita equivalence 133

16 Free exact categories 141

iii
CONTENTS CONTENTS

17 Exact completion and reflexive-coequalizer completion 157

18 Finitary localizations of algebraic categories 169

A Monads 177

B Abelian categories 195

C More about dualities for one-sorted algebraic categories 199

Bibliography 209

Index 215

November 20, 2009 iv

Foreword

The study that was initiated by Birkhoff in 1935 was named “general algebra”
by Kurosh in his classic text; the subject is also called universal algebra, as in
the text by Cohn. The purpose of general algebra is to make explicit common
features of the practice of commutative algebra, group theory, linear algebra,
Lie algebra, lattice theory, et cetera, in order to illuminate the path for that
practice. Less than twenty years after the 1945 debut of the Eilenberg-Mac Lane
method of categorical transformations, the obvious possibility of its application
to general algebra began to be realized in 1963; that realization continues in the
present book.

Excessive iteration of the passage

T ′ = theory of T
would be sterile if pursued as idle speculation without attention to that funda-
mental motion of theory: concentrate the essence of practice, in order to guide
practice. Such theory is necessary to clear the way for the advance of teaching
and research. General algebra can and should be used in particular algebra (i.e.
in algebraic geometry, functional analysis, homological algebra et cetera) much
more than it has been. There are several important instruments for such appli-
cation, including the partial structure theorem in Birkhoff’s “Nullstellensatz”,
the “commutator” construction, and the general framework itself.
Birkhoff’s theorem was inspired by theorems of Hilbert and Noether in alge-
braic geometry (as indeed was the more general model theory of Robinson and
Tarski). His big improvement was not only in generality: beyond mere existence
of generalized points, he showed they are sufficient to give a monomorphic em-
bedding. Nevertheless, in commutative algebra his result is rarely mentioned
(although it is closely related to Gorenstein algebras). The categorical formula-
tion of Birkhoff’s theorem ( [65] and [91]), like the pre-categorical ones, involves
sub-direct irreducibility and Zorn’s lemma. Finitely-generated algebras, in par-
ticular, are partially dissected by the theorem into (often qualitatively simpler)
finitely-generated pieces. For example, when verifying consequences of a system

v
 . FOREWORD

of polynomial equations over a field, it suffices to consider all possible finite-

dimensional interpretations, where constructions of linear algebra such as trace
are available.
Another accomplishment of general algebra is the so-called commutator the-
ory (named for its realization in the particular category of groups); a categorical
treatment of this theory can be found in [78] and [56]. In other categories this
theory specializes to a construction important in algebraic geometry and num-
ber theory, namely the product of ideals [51]. In the geometrical classifying
topos for the algebraic category of K-rigs, this construction yields an internal
multiplicative semi-lattice of closed sub-varieties.
In the practice of group theory and ring theory the roles of presentations
and of the algebras presented have long been distinguished, giving a syntactic
approach to calculation in particular algebraic theories. Yet many works in gen-
eral algebra (and model theory generally) continue anachronistically to confuse
a presentation in terms of signatures with the theory itself that is presented,
thus ignoring the application of general algebra to specific theories, like that of
C ∞ -rings, for which no presentation is feasible.
Apart from the specific accomplishments mentioned above, the most effec-
tive illumination of algebraic practice by general algebra, both classical and
categorical, has come from the explicit nature of the framework itself. The
closure properties of certain algebraic sub-categories, the functorality of seman-
tics itself, the ubiquitous existence of functors adjoint to algebraic functors,
the canonical method for extracting algebraic information from non-algebraic
categories, have served (together with their many particular ramifications) as
a partial guidance to mathematicians in dealing with the inevitably algebraic
content of their subjects. The careful treatment of these basics, by Adámek,
Rosický, and Vitale, will facilitate future mutual applications of algebra, general
algebra, and category theory. They have achieved in this book a new resolution
of several issues that should lead to further research.

What is General Algebra?

The bedrock ingredient for all its aspects is the use of finite cartesian prod-
ucts. Therefore, as a framework for the subject, it is appropriate to recognize
the 2-category of categories that have finite categorical products and of func-
tors preserving these products. Among such categories there are the linear ones
whose products are simultaneously co-products; that is a crucial property of
linear algebra in that maps between products are then uniquely represented
as matrices of smaller maps between the factors (though of course there is no
unique decomposition of objects into products, so it would be incorrect to say
inversely that maps “are” matrices). General categories with products can be
forced to become linear and this reflection 2-functor is an initial ingredient in
linear representation theory. However, I want to emphasize instead a strong
analogy between general algebra as a whole and any particular linear monoidal
category, because that will reveal some of the features of the finite product
framework that make possible the more profound results.
The 2-category of all categories with finite products has (up to equivalence)

November 20, 2009 vi

. FOREWORD

three characteristic features of a linear category such as the category of modules

over a rig:
1. It is “additive” because if A × B is the product of two categories with
finite products, it is also their coproduct, the evident injections from A, B
having the universal property for maps into any third such category.
2. It is “symmetric closed”; indeed Hom(A, B) is the category of algebras in
the background B according to the theory A. The unit I for this Hom is
the opposite of the category of finite sets. The category J of finite sets
itself satisfies Hom(J, J) = I, and the category Hom(J, B) is the category
of Boolean algebras in B. As dualizer, the case B = small sets is most
often considered in abstract algebra.
3. It is “tensored” because a functor of two variables that is product-preserving
in each variable separately can be represented as a product-preserving
functor on a suitable tensor-product category. Such functors occur in
recent work of Zurab Janelidze [55]; specifically, there is a canonical eval-
uation A ⊗ Hom(A, B) → B, where the domain is “a category whose maps
involve both algebraic operations and their homomorphisms”.
A feature not present in abstract linear algebra (though it has an analogue
in the cohesive linear algebra of functional analysis) is Ross Street’s bo-ff fac-
torization of any map (an abbreviation of “bijective on objects followed by full
and faithful”), see [89] and [90]. A single-sorted algebraic theory is a map I ⊗ A
that is bijective on objects; such a map induces a single “underlying” functor
Hom(A, B) ⊗ B on the category of A-algebras in B. The factorization permits
the definition of the full “algebraic structure” of any given map u : X → B as
follows: the map I → Hom(X, B) that represents u has its bo-ff factorization,
with its bo part the algebraic theory I → A(u), the full X-natural structure
(in its abstract general guise) of all values of u. The original u lifts across the
canonical Hom(A(u), B) → B by a unique u# . This is a natural first step in one
program for “inverting” u, because if we ask whether an object of B is a value
of u, we should perhaps consider the richer (than B) structure that any such
object would naturally have; that is, we change the problem to one of inverting
u# . Jon Beck called this program “descent” with respect to the “doctrine” of
general algebra. (A second step is to consider whether u# has an adjoint.)
Frequently, the dualizing background B is cartesian-closed, i.e., it has not
only products but also finite co-products and exponentiation, where exponenti-
ation is a map
B op ⊗ B → B
in our 2-category. This permits the construction of the important family of
function algebras B op → Hom(A, B) given any A-algebra (of “constants”) in B.
On a higher level, the question whether a given C is a value of the 2-functor
U = Hom(−, B) (for given B), leads to the discovery that such values belong to
a much richer doctrine, involving as operations all limits that B has and all co-
limits that exist in B and preserve finite products. As in linear algebra, where

November 20, 2009 vii

 . FOREWORD

dualization in a module B typically leads to modules with a richer system of

operators, conversely such a richer structure assumed on C is a first step toward
2-descent back along U.
The power of the doctrine of natural 2-operations on Hom(−, B) is enhanced
by fixing B to be the category of small sets, where smallness specifically excludes
measurable cardinals (although they may be present in the categorical universe
at large).

A contribution of Birkhoff’s original work had been the characterization

of varieties, that is, of those full subcategories of a given algebraic category
Hom(A, B) that are equationally defined by a surjective map A → A′ of the-
ories. Later, the algebraic categories themselves were characterized. Strik-
ing refinements of those characterization results, in particular, the clarification
of a question left open in the 1968 treatment of categorical general algebra
[64], are among the new accomplishments explained in the present book. As
Grothendieck had shown in his very successful theory of abelian categories, the
exactness properties found in abstract linear algebra continue to be useful for
the concretely variable linear algebras arising as sheaves in complex analysis;
should something similar be true for non-linear general algebras? More specifi-
cally, what are the natural 2-operations and exactness properties shared by all
the set-valued categories concretely arising in general algebra? In particular, can
that class of categories be characterized by further properties, such as sufficiency
of projectives, in terms of these operations? It was clear that small limits and
filtered colimits were part of the answer, as with the locally-finitely-presentable
categories of Gabriel and Ulmer. But the further insistence of general algebra,
on algebraic operations that are total, leads to a further functorial operation,
needed to isolate equationally the correct projectives and also useful in dealing
with non-Mal’cev categories: that further principle is the ubiquitous preserva-
tion of Linton’s reflexive co-equalizers, which are explained in this book as a
crucial case of Lair’s sifted colimits.

Bill Lawvere
Buffalo, October 31, 2009

November 20, 2009 viii

Introduction

The concept of an algebraic theory, introduced in 1963 by F. W. Lawvere, was

a fundamental step towards a categorical view on general algebra in which vari-
eties of algebras are formalized without details of equational presentations. An
algebraic theory, as originally introduced, is roughly speaking a category whose
objects all are finite powers of a given object. An algebra is then a set-valued
functor preserving finite products and a homomorphism between algebras is a
natural transformation. In the almost half of a century that followed, this idea
has gone through a number of generalizations, ramifications, and applications
in areas such as algebraic geometry, topology, and computer science. The gen-
eralization from one-sorted algebras to many-sorted ones (of particulat interest
in computer science) leads to a simplification: an algebraic theory is now simply
a small category with finite products.

Abstract Algebraic Categories

In the first part of this book, consisting of Chapters 1 – 10, we develop the
approach in which algebraic theories are studied without a reference to sort-
ing. Consequently, algebraic categories are investigated as abstract categories.
We study limits and colimits of algebras, paying special attention to the sifted
colimits since they play a central role in the development. For example, alge-
braic categories are characterized as precisely the free completions under sifted
colimits of small categories with finite coproducts. And algebraic functors are
precisely the functors preserving limits and sifted colimits. This leads to an
algebraic duality: the 2-category of algebraic categories is dually biequivalent
to the 2-category of canonical algebraic theories.
Here we present the concept of equation as a parallel pair of morphisms in the
algebraic theory. An algebra satisfies the equation iff it merges the parallel pair.
We prove Birkhoff’s Variety Theorem: subcategories which can be presented
by equations are precisely those closed under products, subalgebras, regular
quotients, and directed unions. (The last item can be omitted in case of one-
sorted algebras.)

ix
 . INTRODUCTION

Concrete Algebraic Categories

Lawvere’s original concept of one-sorted theory is studied in Chapters 11

– 13. Here the categories of algebras are concrete categories over Set, and we
prove that up to concrete equivalence they are precisely the classical equational
categories of Σ-algebras for one-sorted signatures Σ. More generally, given a
set S of sorts, we introduce in Chapter 14 S-sorted algebraic theories and the
corresponding S-sorted algebraic categories which are concrete over S-sorted
sets. Thus we distinguish between “many-sorted” where sorting is not specified
and “S-sorted” where a set S of sorts is given (and this distinction leads us to
considering the categories of algebras as concrete or abstract ones).
This is supplemented by Appendix A in which a short introduction to mon-
ads and monadic algebras is presented. There we prove a duality between one-
sorted algebraic theories and finitary monadic categories over Set. And again,
more generally, between S-sorted algebraic theories and finitary monadic cate-
gories over Set S .
The non-strict version of some concepts, like morphism of one-sorted theories
and concrete functor, is treated in Appendix C.

Special Topics

Chapters 15 – 18 are devoted to some more specialized topics. Here we

introduce Morita equivalence, characterizing pairs of algebraic theories yielding
equivalent categories of algebras. We also prove that algebraic categories are
free exact categories. Finally, the finitary localizations of algebraic categories
are described: they are precisely the exact locally finitely presentable categories.
Abelian categories are shortly treated in Appendix B.

Other Topics

Of the two categorical approaches to general algebra, monads and algebraic

theories, only the latter is treated in this book, with the exception of the short
appendix on monads. Both of these approaches make it possible to study al-
gebras in a general category; in our book we just restrict ourselves to sets and
many-sorted sets. Thus examples such as topological groups are not treated
here.
Another feature not treated at all in this book are algebraic categories based
on more general tensor products, not only the categorical product. Such tensor
theories were already studied by Mac Lane [71] and they lead to the concept of
operad, which is an important development beyond the scope of this book. The
development of operads was mostly stimulated by homotopy theory with the
aim to define “homotopy invariant algebraic structures” (see [17]). Let us just
remark that sifted colimits play a role also in this more general setting since all
monads associated to an operad preserve these colimits, see [85]. Moreover, the
whole machinery of algebraic theories and sifted colimits can be modified to the
homotopy setting, see [70] and [86].

November 20, 2009 x

Chapter 0

Preliminaries

The aim of this chapter is to fix some notation and recall well-known facts
concerning basic concepts of category theory used throughout the book. The
reader may well skip it and return to it where needed. Only the most usual
definitions and results of the theory of categories are mentioned here, more
about them can be found in any of the book mentioned at the end of this
chapter.
0.1 Foundations. In category theory one needs to distinguish between small
collections (sets) and large ones (classes). An arbitrary set theory making such
a distinction possible is sufficient for our book. The category of (small) sets and
functions is denoted by
Set .
All categories we work with have small hom-sets.
0.2 Properties of functors. A functor F : A → B is
1. faithful if for every parallel pair of morphisms f, g : A ⇉ A′ in A, one has
f = g whenever F f = F g,
2. full if for every morphism b : F A → F A′ in B there exists a morphism
a : A → A′ in A such that F a = b,
3. essentially surjective if for every object B in B there exists an object A in
A with B isomorphic to F A,
4. an equivalence if there exists a functor F ′ : B → A such that both F · F ′
and F ′ ·F are naturally isomorphic to the identity functors. Such a functor
F ′ is called a quasi-inverse of F,
5. an isomorphism if there exists a functor F ′ : B → A such that both F · F ′
and F ′ · F are equal to the identity functors,
6. conservative if it reflects isomorphisms, that is, a : A → A′ is an isomor-
phism whenever F a : F A → F A′ is.

1
 CHAPTER 0. PRELIMINARIES

0.3 Remark. Let F : A → B be a functor.

1. If F is full and faithful, then it is conservative.

2. F is an equivalence iff it is full, faithful, and essentially surjective.

3. If F is an equivalence and F ′ a quasi-inverse of F, it is possible to choose

natural isomorphisms η : IdB → F · F ′ and ε : F ′ · F → IdA such that

F ε · ηF = F and εF ′ · F ′ η = F ′

(compare with 0.8). (Observe that in equations like F ε · ηF = F we

write F for the identity natural transformation on a functor F. We adopt
the same convention in diagrams having functors as vertices and natural
transformations as edges.)

4. F is an isomorphism iff it is full, faithful, and bijective on objects.

0.4 Functor categories and Yoneda embedding.

1. Given a category A and a small category C we denote by A C the category

of functors from C to A and natural transformations, and by C op the dual
category.

2. In case A = Set we have the Yoneda embedding

YC : C op → Set C , YC (X) = C(X, −)

which is full and faithful. This follows from the Yoneda Lemma which
states that for every X ∈ C and for every functor F : C → Set, the map as-
signing to every natural transformation α : YC (X) → F the value αX (idX )
is a bijection natural in X and F.

0.5 Diagrams.

1. A diagram in a category K is a functor from a small category into K.

2. A finite diagram is a diagram D : D → K such that D is a finitely generated

category. This means that D has finitely many objects and a finite set of
morphisms whose closure under composition gives all the morphisms of
D.

3. A category is complete if every diagram in it has a limit; dually: cocom-

plete.

0.6 Colimits in Set. In the category of sets:

1. Coproducts are disjoint unions.

November 20, 2009 2

CHAPTER 0. PRELIMINARIES

2. Coequalizers of i, j : X ⇉ Z can be described as the canonical maps

c : Z → Z/ ∼ where ∼ is the smallest equivalence relation with i(z) = j(z)
for every z ∈ Z. This equivalence relation merges elements z and z ′ of Z
iff there exists a zig-zag of elements

y0′ y1′ ······

yn′
@@  @@ ~ ??
f0 @@ f1 f2  @@ f3 f2n ~~ ??f2n+1
@@  @@ ~~ ??
@@  @@ ~ ?
   ··· ~~~
z y1 y2 yn z′

where each fk is equal to i or to j for k = 0, . . . , 2n + 1.

3. A category D is called filtered if every finitely generated subcategory has

a cocone in D (for more about this concept see Chapter 2). A colimit of
a diagram D : D → Set is described as the quotient
a
Dx/ ∼
x∈obj D

where for elements ui ∈ Dxi we have u1 ∼ u2 iff there exist morphisms

fi : xi → y in D such that Df1 (u1 ) = Df2 (u2 ).

0.7 Construction of colimits. In a category A with coproducts and coequal-

izers all colimits exist. Given a diagram D : D → A form a parallel pair

a i // a
Dfd Dx
j
f ∈mor D x∈obj D

where fd and fc denote the domain and codomain of f. The f -component of

i is the coproduct injection of Dfd , that of j is the composite of Df and the
coproduct injection of Dfc .

1. If a
c: Dx −→ C
x∈obj D

is the coequalizer of i and j, then C = colim D and the components of c

form the colimit cocone.

2. The pair i, j above is reflexive, that is, there exists a morphism

a a
δ: Dx −→ Dfd
x∈obj D f ∈mor D

such that i · δ = id = j · δ. Indeed, the x-component of δ is the coproduct

injection of idX .

November 20, 2009 3

 CHAPTER 0. PRELIMINARIES

0.8 Adjoint functors. Given functors U : A → B and F : B → A, then F

is a left adjoint of U, notation F ⊣ U, if there exist natural transformations
η : IdB → U F and ε : F U → IdA satisfying

εF · F η = F and U ε · ηU = U .

This is equivalent to the existence of a bijection

A(F B, A) ≃ B(B, U A)

natural in A ∈ A and B ∈ B.
1. Every left adjoint preserves colimits.
2. Dually, every right adjoint preserves limits.
3. A solution set for a functor U : A → B and an object X of B is a set
of morphisms fi : X → U Ai (i ∈ I) with Ai ∈ A such that every other
morphism f : X → U A has a factorization f = U h · fi for some i ∈ I and
some morphism h : Ai → A in A.
4. The Adjoint Functor Theorem states that if A has limits, then a functor
U : A → B has a left adjoint iff it
(a) preserves limits, and
(b) has a solution set for every object X of B.
0.9 Reflective subcategories. Given a category B, by a reflective subcategory
of B is meant a full subcategory A such that the inclusion functor A → B has a
left adjoint (called a reflector for B). We denote by R : B → A the reflector and
by rB : B → RB the reflections.
0.10 Representable functors. A functor from a category A to Set is repre-
sentable if it is naturally isomorphic to a hom-functor A(A, −).
1. If A has coproducts, then A(A, −) has a left adjoint assigning to a set X
a coproduct of X copies of A.
2. The colimit of A(A, −) is a singleton set.
3. The Adjoint Functor Theorem can be stated in terms of representable
functors as follow:
A functor F : B → Set, with B complete, has a left adjoint iff it is repre-
sentable. This is the case iff it preserves limits and satisfies the solution
set condition: there exists a set G of objects of B such that for any object
B of B and any element b ∈ F B, there are X ∈ G, x ∈ F X and f : X → B
such that F f (x) = b.
0.11 Example.
1. For every set X the functor X × − : Set → Set is left adjoint to Set(X, −).

November 20, 2009 4

CHAPTER 0. PRELIMINARIES

2. For a category A the diagonal functor

∆ : A → A × A , A 7→ (A, A)

has a left adjoint iff A has finite products. Then

A × A → A , (A, B) 7→ A × B

is a left adjoint to ∆.
0.12 Remark. The contravariant hom-functors B(−, B) : B → Setop , B ∈
obj B, collectively reflect colimits. That is, for every cocone C of a diagram
D : D → B we have: C is a colimit of D iff the image of C under any B(−, B)
is a colimit of the diagram B(−, B) · D in Set.
0.13 Slice categories. Given functors F : A → K and G : B → K, the slice
category (F ↓ G) has as objects all triples (A, f, B) with A ∈ A, B ∈ B and
f : F A → GB, and as morphisms (A, f, B) → (A′ , f ′ , B ′ ) all pairs a : A → A′ ,
b : B → B ′ such that Gb · f = f ′ · Ga.
1. As special cases, we have K ↓ G and F ↓ K where an object K ∈ K is
seen as a functor from the one-arrow category to K.
2. If F is the identity functor on K, we write K ↓ K instead of idK ↓ K.
0.14 Set functors as colimits of representables. Every functor A : T →
Set (T small) is in a canonical way a colimit of representable functors. In
fact, consider the Yoneda embedding YT : T op → Set T and the slice category
El A = YT ↓ A of “elements of A”. Its objects can be represented as pairs
(X, x) with X ∈ obj T and x ∈ A(X), and its morphisms f : (X, x) → (Z, z) are
morphisms f : Z → X of T such that Af (z) = x. We denote by ΦA : El A → T op
the canonical projection which to every element of the set AX assigns the object
X. Then A is a colimit of the diagram of representable functors as follows
ΦA YT
El A / T op / Set T

Indeed, the colimit injection YT (ΦA (X, x)) → A is the natural transformation
corresponding, by Yoneda Lemma, to the element x ∈ AX.
0.15 Kernel pair. A kernel pair of a morphism f : A → B is a parallel pair
f1 , f2 : N (f ) ⇉ A forming a pullback of f and f.
0.16 Classification of quotient objects. A quotient object of an object A
is represented by an epimorphism e : A → B, and an epimorphism e′ : A → B ′
represents the same quotient iff e′ = i · e holds for some isomorphism i : B → B ′ .
We use the same adjective for quotient objects and (any of) the representing
epimorphisms e : A → B :
1. Split means that there exists i : B → A with e · i = idB . Then B is called
a retract of A.

November 20, 2009 5

 CHAPTER 0. PRELIMINARIES

2. Regular means that e is a coequalizer of a parallel pair with codomain A.

3. Strong means that in every commutative square
e /B
A
d
u v
 ~ 
X /Z
m

where m is a monomorphism, there is a “diagonal” morphism d : B → X

such that m · d = v and d · e = u.
4. Extremal means that in every commutative triangle
e /B
A@
@@ ~~>
@@ ~
u @@ ~~m
~~
X
where m is a monomorphism, then m is an isomorphism.
Dually, a subobject of A is represented by a monomorphism m : B → A, and a
monomorphism m′ : B ′ → A represents the same subobject iff m′ = m · i holds
for some isomorphism i : B ′ → B.
0.17 Remark. Let us recall some elementary facts on extremal, strong and
regular epimorphisms.
1. Every regular epimorphism is strong and every strong epimorphism is
extremal. If the category A has finite limits, then extremal = strong.
2. If the category A has binary products, then the condition of being an
epimorphism in the definition of strong epimorphism is redundant. The
same holds for extremal epimorphisms if the category A has equalizers.
3. If a composite f ·g is a strong epimorphism, then f is a strong epimorphism.
The same holds for extremal epimorphisms, but in general this fails for
regular epimorphisms.
4. If f is a monomorphism and an extremal epimorphism, then it is an iso-
morphism.
0.18 Concrete categories. Let K be a category.
1. By a concrete category over K is meant a category A together with a
faithful functor U : A → K.
2. Given concrete categories (A, U ) and (A′ , U ′ ) over K, a concrete functor
is a functor F : A → A′ such that U = U ′ · F.
Further Reading
For standard concepts of category theory the reader may consult e.g. the
following monographs: [4], [27], [71].

November 20, 2009 6

Chapter 1

Algebraic theories and

algebraic categories

In this chapter we introduce algebraic theories and algebraic categories and

study basic concepts, such that limits of algebras, and representable algebras.
We also introduce some of the main examples of algebraic categories.
1.1 Definition. An algebraic theory is a small category T with finite products.
An algebra for the theory T is a functor A : T → Set preserving finite products.
We denote by Alg T the category of algebras of T . Morphisms, called homomor-
phisms, are the natural transformations. That is, Alg T is a full subcategory of
the functor category Set T .
1.2 Definition. A category is algebraic if it is equivalent to Alg T for some
algebraic theory T .
1.3 Remark. An algebraic theory is by definition a small category. However,
throughout the book we do not take care of the difference between small and
essentially small: a category is essentially small if it is equivalent to a small one.
We will see in 10.15, 13.11 and 14.23 that algebraic categories correspond
well with varieties, i.e., equational categories of (many-sorted, finitary) algebras.
Before giving some examples of algebraic categories, we need two simple facts.
1.4 Remark. Every object t of an algebraic theory T yields the algebra YT (t)
representable by t :
YT (t) = T (t, −) : T → Set .
This, together with the Yoneda transformations, defines a full and faithful func-
tor
YT : T op → Alg T .
1.5 Lemma. For every algebraic theory T , the Yoneda embedding
YT : T op → Alg T
preserves finite coproducts.

7
CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

Proof. If 1 is a terminal object of T then T (1, −) is an initial object of Alg T :

for every algebra A we know that A1 is a terminal object, thus there is a unique
morphism T (1, −) → A.
Given two objects t1 , t2 in T then T (t1 ×t2 , −) is a coproduct of T (t1 , −) and
T (t2 , −) since for every algebra A the morphisms T (t1 × t2 , −) → A correspond
to elements of A(t1 × t2 ) = A(t1 ) × A(t2 ).

The simplest examples of algebraic categories are the category of sets and the
category of many-sorted sets. We will treat them in 1.9 and 1.10 as special cases
of the following more general example which plays a prominent role throughout
our book.
1.6 Example. Set-valued functors. If C is a small category, the functor category
Set C is algebraic. An algebraic theory of Set C is a free completion TC of C under
finite products. This means a functor ETh : C → TC such that
1. TC is a category with finite products
and
2. for every functor F : C → B, where B is a category with finite products,
there exists an essentially unique functor (that is, unique up to natural
isomorphism) F ∗ : TC → B preserving finite products with F naturally
isomorphic to F ∗ · ETh .
In other words, composition with ETh gives an equivalence between the category
of finite product preserving functors from TC to B and the category of functors
from C to B. In particular, the categories Set C and Alg TC are equivalent.
1.7 Remark. The free finite-product completion TC can be described as follows:
(i) Objects of TC are all finite families
(Ci )i∈I , I finite
of objects of C, and morphisms from (Ci )i∈I to (Cj′ )j∈J are pairs (a, α)
where a : J → I is a function and α = (αj )j∈J is a family of morphisms
αj : Ca(j) → Cj′ of C. The composition and identity morphisms are de-
fined as expected. A terminal object in TC is the empty family, and
a product of two objects is the disjoint union of the families. Finally,
the functor ETh : C → TC is given by ETh (C) = (C). It is easy to ver-
ify the universal property: since for every object (Ci )i∈I in TC we have
(Ci )i∈I = ΠI ETh (Ci ), then necessarily F ∗ ((Ci )i∈I ) = ΠI F Ci .
(ii) Equivalently, TC can be described as the category of all words over obj C
(the set of objects of C). That is, objects have the form of n-tuples
C0 . . . Cn−1 where each Ci is an object of C (and where n is identified
with the set {0, . . . , n − 1}) including the case n = 0 (empty word). Mor-
′
phisms from C0 . . . Cn−1 to C0′ . . . Ck−1 are pairs (a, α) consisting of a
function a : k → n and a k-tuple of C-morphisms α = (α0 , . . . , αk−1 ) with
αi : Ca(i) → Ci′ .

November 20, 2009 8

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

1.8 Remark. Since the Yoneda embedding YTC : TC op → Alg TC ≃ Set C pre-
serves finite coproducts (1.5), the category TC op is equivalent to the full subcat-
egory of Set C given by finite coproducts of representable functors.
1.9 Example. Sets. The simplest algebraic category is the category of sets
itself. An algebraic theory N for Set can be described as the full subcategory
of Setop whose objects are the natural numbers. In fact, since n = 1 × . . . × 1
in Setop , every algebra A : N → Set is determined, up to isomorphism, by the
set A(1). More precisely, we have an equivalence functor

E : Alg N → Set , A 7→ A(1) .

The category Set has other algebraic theories – we describe them in Chapter 15.
1.10 Example. Many-sorted sets. The algebraic theory N for Set described in
1.9 is nothing else than the theory TC of 1.6 when C is the one-object discrete
category. More generally, if in 1.6 the category C is discrete, i.e. it is a set
S, we get the power category Set S of S-sorted sets and S-sorted functions.
Following 1.7, TSop is equivalent to the full subcategory of Set S
` of finite S-
sorted sets (an S-sorted set hAs is∈S is finite if the coproduct S As is a finite
set). Another equivalent description of TS is obtained by taking finite words
over S as objects (including the empty word). Morphisms from s0 . . . sn−1 to
s′0 . . . s′k−1 are functions a : k → n such that sa(i) = s′i (i = 0, . . . , k − 1).
1.11 Example. Abelian groups. We denote by Tab the category having natural
numbers as objects, and morphisms from n to k are matrices of integers with
n columns and k rows. Composition of P : m → n and Q : n → k is given by
matrix multiplication Q · P = Q × P : m → k, and identity morphisms are the
unit matrices. If n = 0 or k = 0, the only n × k matrix is the empty one [ ].
Tab has finite products. For example, 2 is the product 1 × 1 with projections
[1, 0] : 2 → 1 and [0, 1] : 2 → 1. (In fact, given one-row matrices P, Q : n → 1,
there exists a unique two-row matrix R : n → 2 such that [1, 0] · R = P and
[0, 1] · R = Q : the matrix with rows P and Q.) Here is a direct argument
proving that the category Ab of abelian groups is equivalent to Alg Tab (see also
1.13). Every abelian group G defines an algebra G b : Tab → Set whose object
b n
function is Gn = G . For every morphism P : n → k we define GP b : Gn → Gk
by matrix multiplication
   
g1 g1
b ): 
G(P .   . 
 ..  7→ P ·  .. 
gn gn

The function G 7→ G d : Ab → Alg Tab in a rather obvious

b extends to a functor (−)
way: given a group homomorphism h : G1 → G2 , then ĥ : G b1 → G b 2 is the natural
transformation whose components are hn : Gn1 → Gn2 . It is clear that (−) d is a
well defined, full and faithful functor. To prove that it is an equivalence functor,
we need, for every algebra A : Tab → Set, to present an abelian group G with

November 20, 2009 9

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

A ≃ G. b The underlying set of G is A1. The binary group operation is obtained

from the morphism [1, 1] : 2 → 1 in Tab by A[1, 1] : G2 → G, the neutral element
is A[ ] : 1 → G for the morphism [ ] : 0 → 1 of Tab , and the inverse is given by
A[−1] : G → G. It is not difficult to check that the axioms of abelian group are
fulfilled. For example,
 the axiom x+0 = x follows from the fact that A preserves
1 b (consider
the composition of : 1 → 2 with [1, 1] : 2 → 1. Clearly, A ≃ G
0
the canonical isomorphism An = A(1 × . . . × 1) ≃ A1 × . . . × A1 = Gn = Gn).b

1.12 Example. Modules. Let R be a ring with unit. The category R-Mod
of left modules and module homomorphisms is algebraic. A theory directly
generalizing that of abelian groups above has as objects natural numbers and as
morphisms matrices over R. Algebraic categories of the form R-Mod are treated
in greater detail in Appendix B.
1.13 Remark.
1. In the example of abelian groups we have the forgetful functor U : Ab →
Set assigning to every abelian group its underlying set. Observe that
the above groups Zn are free objects of Ab on n generators, and the full
subcategory of all these objects is the dual of the theory Tab above.
2. Analogously, if the category C of Example 1.6 has object set S, then we
have a forgetful functor which forgets the action of A : C → Set on mor-
phisms:
U : Set C → Set S , U A = hA(s)is∈S .
The functor U has a left adjoint
!
a a
S C
F : Set → Set , F (hAs is∈S ) = C(s, −)
s∈S As

(this easily follows from the Yoneda Lemma because F preserves coprod-
ucts). Following 1.8, the objects of the theory TC are precisely the finitely
generated free objects, that is, the image of finite S-sorted sets under F.
3. In Chapters 11 and 14 we will see that this is not a coincidence: for
every S-sorted algebraic category A the free objects on finitely many S-
sorted generators form a full subcategory whose dual is a theory for A (see
11.22 for one-sorted algebraic categories and 14.13 for S-sorted algebraic
categories).
1.14 Remark.
1. Let Σ be a signature, that is, a set Σ (of operation symbols) together with
an arity function
ar : Σ → N .
A Σ-algebra consists of a set A and, for every n-ary symbol σ ∈ Σ, an n-
ary operation σ A : An → A. An homomorphism of Σ-algebras is a function

November 20, 2009 10

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

preserving the given operations. The category Σ-Alg of Σ-algebras and

homomorphisms as well as its equational subcategories are algebraic, as we
demonstrate in Chapter 13. Thus e.g. Boolean algebras form an algebraic
category.

2. Also the more general many-sorted signatures yield algebraic categories

that we study in Chapter 14. Here is a concrete example:

1.15 Example. Graphs. We denote by Graph the category of directed graphs

G with multiple edges: they are given by a set Gv of vertices, a set Ge of edges,
and two functions from Ge to Gv determining the target (τ ) and the source (σ)
of every edge. The morphisms are called graph homomorphisms: given graphs
G and G′ a graph homomorphism is a pair of functions hv : Gv → G′v and
he : Ge → G′e such that the source and the target of every edge is preserved. A
theory TG of Graph is a special case of 1.6 with the obvious category G :
τ
ide :e // v d idv
σ

The category of algebras of an algebraic theory is quite rich. We already

know that every object t of an algebraic theory T yields the representable al-
gebra YT (t) = T (t, −). Other examples of algebras can be obtained e.g. by the
formation of limits and colimits. We will now show that limits always exist and
are built up at the level of sets. Also colimits always exist, but they are seldom
built up at the level of sets. We will study colimits in the subsequent Chapters.

1.16 Proposition. For every algebraic theory T , the category Alg T is closed
in Set T under limits.

Proof. Limits are formed objectwise in Set T . Since limits and finite products
commute, given a diagram in Set T whose objects are functors preserving finite
products, then a limit of that diagram also preserves finite products.

1.17 Corollary. Every algebraic category is complete.

1.18 Remark.

1. The previous proposition means that limits of algebras are formed object-
wise at the level of sets. For example, a product of two graphs has both
the vertex set given by the cartesian product of the vertex sets, and the
set of edges given by the cartesian product of the edge sets.

2. Monomorphisms in the category Alg T are precisely the homomorphisms

that are componentwise monomorphisms (i.e., injective functions) in Set.
In fact, this is true in Set T , and Alg T is closed under monomorphisms
(being closed under limits) in Set T .

3. In every algebraic category kernel pairs (0.15) exist and are formed ob-
jectwise (in Set).

November 20, 2009 11

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

1.19 Example. One of the most important data types in computer science is
a stack, or finite list, of elements of a set (of “letters”) called an alphabet. Here
we consider stacks of natural numbers: we will have elements of sort n (natural
number) and s (stack) and the following basic operations:

succ, the successor of a natural number,

push, which adds a new letter to the left-most position of a stack,

pop, which deletes the left-most position,

and

top, which reads the top element of the stack.

We will also have two constants: e, for the empty stack, and 0 of sort n. For
simplicity we put top(e) = 0. This leads us to the concept of algebras A of two
sorts
s (stack) and n (natural number)
with operations

succ : An → An ,

push : As × An → As ,

pop : As → As ,

and

top : As → An ,

and with constants 0 ∈ An and e ∈ As .

We can consider stacks as equationally specified algebras of sorts {s, n}, and
the algebraic theory is then obtained from the corresponding finitely generated
free algebras.

1.20 Example. Sequential automata. Recall that a deterministic sequential

automaton A is given by a set As of states, a set Ai of input symbols, a set Ao
of output symbols, and by three functions

δ : As × Ai → As (next-state function)

γ : As → Ao (output)

ϕ : 1 → As (initial state)

Given two sequential automata A and A′ = (A′s , A′i , A′o , δ ′ , γ ′ , ϕ′ ), a morphism

(simulation) is given by a triple of functions

hs : As → A′s , hi : Ai → A′i , ho : Ao → A′o

November 20, 2009 12

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

such that the diagram

δ γ
As × Ai / As / Ao
ϕ ~~
~?
~
~~
~~
hs ×hi 1? hs ho
??
??
??
 ϕ′ ?  
A′s × A′i ′
/ A′
s ′
/ A′
o
δ γ

commutes. This is the category of algebras of three sorts s, i and o given by the
signature
δ : si → s , γ : s → o , ϕ : → s
(see Chapter 14 for the general notion of many-sorted signature). Again, an
algebraic theory of automata is formed by considering finitely generated free
algebras.

Historical Remarks for Chapter 1

Algebraic theories were introduced by F. W. Lawvere in his dissertation [63]. He

considered the one-sorted case, which we study in Chapter 11. This corresponds
to (one-sorted) equational theories of G. Birkhoff [24] which we treat in Chapter
13.
Many-sorted equational theories were first considered by P. J. Higgins [53]
and were later popularized by G. Birkhoff and J. D. Lipson [25]. In the review
of Higgins’ paper A. Heller [52] suggested to look for the connection with Law-
vere’s approach. This was done by J. Bénabou [21] who dealt with many-sorted
algebraic theories. Our definition of an algebraic theory is given without a ref-
erence to sorting. This “sort-free” approach corresponds to the more general
theory of sketches initiated by C. Ehresmann [41] (see [20] for an exposition).
The S-sorted approach is presented in Chapter 14.
The interested reader can find expositions of various aspects of algebraic
theories e.g. in [98], [77], [27], [7] and [80].

November 20, 2009 13

CHAPTER 1. ALGEBRAIC THEORIES AND ALGEBRAIC CATEGORIES

November 20, 2009 14

Chapter 2

Sifted and filtered colimits

Colimits in algebraic categories are, in general, not formed objectwise. In this

chapter we study the important case of sifted colimits, which are always formed
objectwise. Prominent examples of sifted colimits are filtered colimits and re-
flexive coequalizers (see Chapter 3).

2.1 Definition. A small category D is called

1. sifted if finite products in Set commute with colimits over D.

2. filtered if finite limits in Set commute with colimits over D.

Colimits of diagrams over sifted (or filtered) categories are called sifted (or
filtered) colimits.

2.2 Remark. Explicitly, a small category D is sifted iff, given a diagram

D : D × J → Set

where J is a finite discrete category, then the canonical map

!
Y Y
colim D(d, j) → ( colim D(d, j)) [2.1]
D D
J J

is an isomorphism. D is filtered iff it satisfies the same

Q condition, but with
respect to every finitely generated category J (replace J by limJ in [2.1]).

2.3 Example.

1. Colimits of ω-chains are filtered. Here the category D is the poset of

natural numbers, considered as a category.

2. More generally, colimits of chains (where D is an infinite ordinal considered

as a poset) are filtered. These are the “typical” filtered colimits: a category
having colimits of chains has all filtered colimits, see 1.5 and 1.7 in [AR].

15
 CHAPTER 2. SIFTED AND FILTERED COLIMITS

3. Generalizing still further: directed colimits are filtered. Recall that a poset
is called (upwards) directed if it is nonempty and every pair of elements
has an upper bound. Directed colimits are colimits of diagrams whose
schemes are directed posets.
4. An example of filtered colimits that are not directed: the colimits of idem-
potents. Let f be an endomorphism of an object A which is idempotent,
that is, f · f = f. This can be considered as a diagam whose domain D has
one object and, besides the identity, precisely one idempotent morphism.
This category is filtered. In fact, the colimit of the above diagram is the
coequalizer of f and idA . It is not difficult to verify directly (or using 2.18)
that in Set these coequalizers commute with finite limits.
Colimits of idempotents are the only filtered colimits of finite diagrams:
every category with colimits of idempotents has all finite filtered colimits,
and every functor preserving colimits of idempotents preserves all finite
filtered colimits, see e.g. [7].
5. Filtered colimits are of course sifted.
6. Coequalizers are colimits that are not sifted (see 2.16). As we will see
in Chapter 3, reflexive coequalizers are sifted (but not filtered); these are
coequalizers of parallel pairs a1 , a2 : A ⇉ B for which d : B → A exists
with a1 · d = idB = a2 · d.
In fact, in a sense made precise in Chapter 7, we can state that

sifted colimits = filtered colimits + reflexive coequalizers.

2.4 Remark. Sifted categories have an easy characterization: they

(i) are nonempty
and
(ii) have, for every pair of objects, the category of all cospans connected.
This will be proved in 2.14. Before doing that, we need to recall the standard
concepts of connected category and final functor. But we first present a result
showing why sifted colimits are important.
2.5 Proposition. For every algebraic theory T , the category Alg T is closed in
Set T under sifted colimits.
Proof. Since sifted colimits and finite products commute in Set, they do so
in Set T (where they are computed objectwise). It follows that a sifted colimit
in Set T of functors preserving finite products also preserves finite products.
2.6 Example. Coproducts are not sifted colimits. In fact, for almost no alge-
braic theory T is Alg T closed under coproducts in Set T . Concrete example: if
Tab is the theory of abelian groups (1.11), then binary coproducts in Alg Tab are
products and in SetTab they are disjoint union.

November 20, 2009 16

CHAPTER 2. SIFTED AND FILTERED COLIMITS

2.7 Corollary. In every algebraic category sifted colimits commute with finite
products.

In fact, this follows from the fact that the category Alg T is closed under limits
and sifted colimits in Set T and such limits and colimits in Set T are formed
objectwise.

2.8 Example. In the category of abelian groups

1. a directed union of abelian groups carries a canonical structure of abelian

group: the directed colimit of the diagram of inclusion homomorphisms,

2. let a1 , a2 : A ⇉ B be a pair of homomorphisms with a common splitting

d : B → A (i.e., a1 ·d = idB = a2 ·d) and let c : B → C be its coequalizer in
Set. The set C carries a canonical structure of abelian group (the unique
one for which c is a homomorphism). Reflexive coequalizers will be studied
in detail in Chapter 3.

2.9 Remark. Generalizing 2.8.1, a directed union in Alg T is a directed col-

imit of subalgebras. That is, an algebra A is a directed union of subalgebras
mi : Ai → A (i ∈ I) provided that the poset on I given by i ≤ j iff mi ⊆ mj is
directed and A is the colimit (with colimit cocone mi ) of the directed diagram
of all Ai , i ∈ I and all connecting morphisms mij :

? A `@@
mi ~~~ @@mj
~~~ @@
@
~~
Ai / Aj
mij

Thus Alg T is closed under directed unions in Set T .

2.10 Definition. A category A is called connected if it is nonempty and for ev-

ery pair of objects X and X ′ in A there exists a zig-zag of morphisms connecting
X and X ′ :
X → X1 ← X2 → . . . → Xn ← X ′ .
This is equivalent to saying that A cannot be decomposed as a coproduct (that
is, disjoint union) of two nonempty subcategories.

2.11 Remark. A small category A is connected iff the constant functor A → Set
of value 1 has colimit 1.

2.12 Definition. A functor F : D′ → D is called final if, for every diagram

D : D → A such that colim (D · F ) exists in A, then colim D exists and the
canonical morphism colim D · F → colim D is an isomorphism.

2.13 Lemma. The following conditions on a functor F : D′ → D are equivalent:

1. F is final;

November 20, 2009 17

 CHAPTER 2. SIFTED AND FILTERED COLIMITS

2. F satisfies the finality condition with respect to all representable functors

D = D(d, −);
3. For every object d of D, the slice category d ↓ F of all arrows d → F d′ ,
d′ ∈ obj D′ , is connected.
Proof. 1 ⇒ 2 is trivial and 2 ⇒ 3 follows from the usual description of
colimits in Set (see 0.6) and the fact that since the diagram D(d, −) has colimit
1 (see 0.10), so does the diagram D(d, F −) = D(d, −) · F.
To prove 3 ⇒ 1, let D : D → A be a diagram and let cd′ : D(F d′ ) → C (d′ ∈
obj D′ ) be a colimit of D · F. For every object d of D, choose a morphism
ud : d → F d′ for some d′ ∈ obj D′ , and put gd = cd′ · Dud : Dd → C. We
claim that gd : Dd → C (d ∈ obj D) is a colimit of D. In fact, since d ↓ F
is connected, it is easy to verify that gd does not depend on the choice of d′
and ud , and that these morphisms form a cocone of D. The rest of the proof is
straightforward.
2.14 Theorem. A small category D is sifted if and only if it is nonempty and
the diagonal functor ∆ : D → D × D is final.
In order to simplify the proof, let us observe two preliminary facts:
a) Following 2.13 the finality of ∆ means that for every pair of objects A, B
of D the category (A, B) ↓ ∆ of cospans on A, B is connected. That is:
(i) a cospan A → X ← B exists, and
(ii) every pair of cospans on A, B is connected by a zig-zag of cospans.
Therefore, the statement

D is nonempty and the diagonal functor ∆ is final

is equivalent to the statement

D is connected and the diagonal functor ∆ is final.

b) The canonical map

!
Y Y
δ : colim D(d, j) → ( colim D(d, j))
D D
J J

of 2.2 is an isomorphism for every finite discrete category J iff δ is an

isomorphism when J is the empty set and when J is the two-element set.
We are going to prove that D is sifted iff it is connected and the diagonal functor
∆ is final. More precisely, we are going to prove that
1. D is connected iff δ is an isomorphism when J is the empty set,
and

November 20, 2009 18

CHAPTER 2. SIFTED AND FILTERED COLIMITS

2. ∆ is final iff δ is an isomorphism when J is the two-element set (that is,

binary products in Set commute with colimits over D).
Proof. 1. When J is the empty set the codomain of δ is 1 whereas its domain
is 1 iff D is connected (see 2.11).
2. Let J be the two-element set. Given diagrams D, D′ : D → Set, consider the
functor
D × D′ : D × D → Set , (d, d′ ) 7→ Dd × D′ d′ .
Since, for every set X, the functor X × − : Set → Set preserves colimits (see
0.11), the colimit of the diagram D × D′ is

colim′ (Dd×D′ d′ ) ≃ colim ( colim

′
(Dd×D′ d′ )) ≃ colim Dd × colim
′
D′ d′ .
D×D d∈D d ∈D d∈D d ∈D

Consider now the following commutative diagram of canonical maps

colim (Dd × D′ d) δ / ( colim Dd) × ( colim D′ d)

D D D
O
= ≃

colim ((D × D′ ) · ∆) β
/ colim (D × D′ )
D D×D

If ∆ is final, then β is an isomorphism and therefore δ is also an isomorphism.

Conversely, assume that δ is an isomorphism. Given two objects d and d′
in D, the representable functor (D × D)((d, d′ ), −) is nothing but D × D′ : D ×
D → Set, with D = D(d, −) and D′ = D(d′ , −). If δ is an isomorphism, the
previous diagram shows that ∆ satisfies the finality condition with respect to
all representable functors. Following 2.13, ∆ is final.
2.15 Example. Every small category with finite coproducts is sifted. In fact,
it contains an initial object, and the slice category (A, B) ↓ ∆ is connected
because it has an initial object (the coproduct of A and B).
2.16 Example. Consider the category D given by the morphisms
a1
A // B
a2

(identity morphisms are not depicted). D is not sifted. In fact, the slice category
(A, B) ↓ ∆ is the discrete category with objects (f, idB ) and (g, idB ).
2.17 Remark. Filtered colimits are closely related to sifted ones. In fact, our
definition in 2.1 stresses this fact. The more usual definition of filtered category
D is to say that every finitely generated subcategory of D has a cocone in D.
(This includes the condition that D is nonempty.) And a well-known result
states that this implies the property of Definition 2.1.2. The converse is also
true:
2.18 Theorem. For a small category D the following conditions are equivalent:

November 20, 2009 19

 CHAPTER 2. SIFTED AND FILTERED COLIMITS

1. D is filtered,
2. every finitely generated subcategory of D has a cocone,
and
3. D is nonempty and fulfils
(a) for every pair of object A, B there exists a cospan A → X ← B, and
(b) for every parallel pair of morphisms u, v : A ⇉ B there exists a mor-
phism f : B → C merging u and v: f · u = f · v.
Proof. The proof of the implications 3 ⇔ 2 ⇒ 1 is standard, the reader
can find it e.g. in [27], Vol. 1, Theorem 2.13.4. The proof of the implication
1 ⇒ 3 is easy: for (a) argue as in 2.14, for (b) use, analogously, the equalizer of
D(u, −), D(v, −) : D(B, −) ⇉ D(A, −) which is the diagram D of all morphisms
merging u and v : since colim D = 1 the diagram is nonempty.

With the next proposition we establish a further analogy to 2.14:

2.19 Proposition. A small category D is filtered if and only if for any finitely
generated category J the diagonal functor ∆ : D → DJ is final.
Proof. 1. Let all such functors ∆ be final. We are to show that every
finitely generated subcategory J of D has a cocone in D. The inclusion functor
d : J → D is an object of the functor category DJ . Following 2.13.3, the slice
category d ↓ ∆ is connected, thus, nonempty. Since d ↓ ∆ is precisely the
category of cocones of J in D, we obtain the desired cocone.
2. Conversely, let D be filtered. We are to verify that for every object d of DJ ,
where J is finitely generated, the slice category d ↓ ∆ is connected.
2a. If J is a subcategory of D and d : J → D is the inclusion functor, then the
category d ↓ ∆ of all cocones is nonempty. To prove that it is in fact connected,
consider two cocones C1 and C2 . Since J is finitely generated (say, by a finite
set M of morphisms), it has finitely many objects, thus, C1 and C2 are finite
sets of morphisms. Put M = M ∪ C1 ∪ C2 and let J be the subcategory of D
generated by M . Then J has a cocone in D, and this cocone defines an obvious
cocone of J with cocone morphisms to C1 and C2 . Thus, we obtain a zig-zag
of length 2.
2b. Let d : J → D be arbitrary and let M be a finite set of morphisms generating
J . Then the set
d(M ) ∪ {idd(x) ; x ∈ obj J }
is finite and generates a subcategory J0 of D. The slice category d ↓ ∆ is clearly
equivalent to the category of cocones of J0 in D which is connected by 2a.
2.20 Remark. Every colimit can be expressed as a filtered colimit of finite
colimits. That is, given a diagram D : D → A with A cocomplete, then colim D
can be constructed as the filtered colimit of the diagram of all colim D′ where
D′ : D′ → A ranges over all domain restrictions of D to finitely generated sub-
categories D′ of D.

November 20, 2009 20

CHAPTER 2. SIFTED AND FILTERED COLIMITS

2.21 Remark. In Chapter 7 we study functors preserving filtered and sifted

colimits. In case of endofunctors of Set these two properties coincide (see 6.30)
but in general the latter one is stronger, see 2.25. We use the following termi-
nology:

2.22 Definition. A functor is called finitary if it preserves filtered colimits.

2.23 Example. Here we mention some endofunctors of Set that are finitary.

1. The functor
Hn : Set → Set Hn X = X n
is finitary for every natural number n since finite products commute in
Set with filtered colimits.

2. A coproduct of finitary functors is finitary.

3. Let Σ be a signature (1.14). We define the corresponding polynomial

functor
HΣ : Set → Set
as the coproduct of the functors Har(σ) for σ ∈ Σ. Explicitly,
a
HΣ X = Σn × X n
n∈N

where Σn is the set of all symbols of arity n (n = 0, 1, 2, . . .). This is a

finitary functor.

2.24 Example. Let H : Set → Set be a functor. An H-algebra is a pair (A, a)

where A is a set and a : HA → A a function. A homomorphism from (A, a) to
(B, b) is a function f : A → B such that the square

Hf
HA / HB
a b
 
A /B
f

commutes. The resulting category is denoted H-Alg .

1. In 13.23 we will see that if H is finitary then the category H-Alg is alge-
braic.

2. The special case of a polynomial endofunctor HΣ leads to Σ-algebras.

Indeed, for every one-sorted signature Σ the category Σ-Alg is precisely the
category HΣ -Alg : if (A, a) is a HΣ -algebra then the operations σ A : An →
A are the domain restrictions of a to the summand An corresponding to
σ ∈ Σn . This case will be treated in Chapter 13.

November 20, 2009 21

 CHAPTER 2. SIFTED AND FILTERED COLIMITS

2.25 Example. An example of a finitary functor not preserving sifted colimits

is the forgetful functor U : Pos → Set, where Pos is the category of posets
and order-preserving maps. Consider the refexive pair u, v : 1 + 2 ⇉ 2 from
the coproduct of the terminal poset 1 and the two-element chain, where both
morphisms are identity on the second summand, and they map the first one to
the top and bottom of 2, respectively. Whereas the coequalizer in Pos is given
by the terminal poset, the coequalizer of U u and U v in Set has two elements.

Historical Remarks for Chapter 2

Filtered colimits are a natural generalization of directed colimits known from

algebra and topology since the beginning of the twentieth century. The general
concept can already be found in Bourbaki [29] including the fact that directed
colimits commute with finite products in Set. Both [48] and [13] contain the
general definition of filtered colimits and the fact that they are precisely those
colimits which commute with finite limits in Set. P. Gabriel and F. Ulmer even
speculated about general commutation of colimits with limits in Set (see Chapter
15 in [48]) and characterized colimits commuting with finite products in Set; this
is the source of 2.14. This was later re-discovered by C. Lair [61] who called
these colimits “tamisantes”. The term sifted was suggested by P. T. Johnstone.
The concept of a final functor and the characterization 2.13 is a standard
result of category theory which can be found in [71].
The fact that sifted colimits play an analogous role for algebraic categories
that filtered colimits play for the locally finitely presentable ones was presented
in [8].

November 20, 2009 22

Chapter 3

Reflexive coequalizers

An important case of sifted colimits are reflexive coequalizers:

3.1 Definition. Reflexive coequalizers are coequalizers of reflexive pairs, that

is, parallel pairs of split epimorphisms having a common splitting.

3.2 Example. In other words, reflexive coequalizers are colimits of diagrams

over the category M given by the morphisms

a1
/
Ao d /B
a2

(identity morphisms are not depicted) composed freely modulo a1 · d = idB =

a2 · d. This category is sifted: it is an easy exercise to check that the categories
(A, A) ↓ ∆, (A, B) ↓ ∆ and (B, B) ↓ ∆ are connected.
Another method of verifying that M is a sifted category is to prove di-
rectly that reflexive coequalizers commute in Set with binary products. In fact,
suppose that

a1 a′1
c c′
A // B / C and A′ // B ′ / C′
a2 a′2

are reflexive coequalizers in Set. We can assume, without loss of generality, that
c is the canonical function of the quotient C = B/ ∼ modulo the equivalence
relation described as follows: two elements x, y ∈ B are equivalent iff there
exists a zig-zag

A: z1 z2 ······
zk
ai1  === ai ai

=== ai ai2k−1

@@@ ai
 == 2 3

== 4

@@ 2k
 ==

== @@
 =
=


@

···

B: x y

23
 CHAPTER 3. REFLEXIVE COEQUALIZERS

where i1 , i2 , . . . , i2k are 1 or 2. For reflexive pairs a1 , a2 the zig-zags can always
be chosen to have the form

A: z1 z2 ······
z2k [3.1]
a1  ===

===  AAA a
 a
== 2 a ==a1 a2 AA 1
 ==



2

==  AA

 =

=  A
··· 
B: x y

(here for the elements z2i of A we use a1 , a2 and for the elements z2i+1 we use
a2 , a1 ). In fact, let d : B → A be a joint splitting of a1 , a2 .) Given a zig-zag, say,

z
 ??? a
a2
 ?? 1
 ??
 
x y

we can modify it as follows:

dx z>
~ @@@ a >>
a1
~~~ @@ 2 a2 >>a1
>>
~~ @@
~~~ @ >
x x y

Analogously for the general case. Moreover, the length 2k of the zig-zag [3.1]
can be prolonged to 2k +2 or 2k +4 etc. by using d. Analogously, we can assume
C ′ = B ′ / ∼′ where ∼′ is the equivalence relation given by zig-zags of a′1 and a′2
of the above form [3.1]. Now we form the parallel pair

a1 ×a′1
//
A × A′ B × B′
a2 ×a′2

and obtain its coequalizer by the zig-zag equivalence ≈ on B×B ′ . Given (x, x′ ) ≈
(y, y ′ ) in B × B ′ , we obviously have zig-zags both for x ∼ y and for x′ ∼′ y ′
(use projections of the given zig-zag). But also the other way round: whenever
x ∼ y and x′ ∼′ y ′ , then we choose the two zig-zags so that they both have the
above type [3.1] and have the same lengths. They create an obvious zig-zag for
(x, x′ ) ≈ (y, y ′ ). From this it follows that the map

a1 ×a′1 c×c′
// / (B/ ∼) × (B ′ / ∼′ )
A × A′ B × B′
a2 ×a′2

is a coequalizer, as required.
3.3 Corollary. For every algebraic theory T , the category Alg T is closed in
Set T under reflexive coequalizers.
In fact, this follows from 2.5 and 3.2.

November 20, 2009 24

CHAPTER 3. REFLEXIVE COEQUALIZERS

3.4 Example. In a category with kernel pairs every regular epimorphism is a

reflexive coequalizer. In fact, if r1 , r2 is a kernel pair of a regular epimorphism
e: A → B
A
d
id  id
R@
~~ @@
~~~r @@
r2 @@
1
 ~~  
A@ A
@@ ~~
@@ ~
e @@ ~~e
 ~~
B
then e is a coequalizer of r1 , r2 . And since e · id = e · id, there exists a unique d
with r1 · d = id = r2 · d.
3.5 Corollary. For every algebraic theory T , the category Alg T is closed in
Set T under regular epimorphisms. Therefore regular epimorphisms in Alg T
are precisely the homomorphisms which are componentwise epimorphisms (i.e.,
surjective functions) in Set.
In fact, the first part of the statement follows from 2.5 and 3.4. The second one
is clear since the statement is true in Set T .
3.6 Remark. In particular, every algebraic category is co-wellpowered with
respect to regular epimorphisms. This means that, for a fixed object A, the
regular quotients of A constitute a set (not a proper class). In fact, this is true
in Set and therefore, by 3.5, in every algebraic category.
More is true: algebraic categories are co-wellpowered with respect to all
epimorphisms. We do not present a proof of this fact here because we do not
need it. The interested reader can consult [7], 1.52 and 1.58.
3.7 Corollary. Every algebraic category has regular factorizations, i.e., every
morphism is a composite of a regular epimorphism followed by a monomorphism.
Proof. The category Set T has regular factorizations: given a morphism
f : A → B, form a kernel pair r1 , r2 : R ⇉ A and its coequalizer e : A → C.
The factorizing morphism m
r1 f
R // A /B
r2
~~?
e ~
~~m
 ~~
C
is a monomorphism. This follows from the fact that kernel pairs and coequalizers
are formed objectwise (in Set). Since Alg T is closed in Set T under kernel pairs
(1.18) and their coequalizers (3.3), it inherits the regular factorizations from
Set T .

November 20, 2009 25

 CHAPTER 3. REFLEXIVE COEQUALIZERS

3.8 Example. In Ab we know that:

1. Coproducts are not formed at the level of sets. In fact, A + B = A × B

for all abelian groups A, B.

2. Reflexive coequalizers are formed at the level of sets, but general coequal-
izers are not. Consider e.g. the pair x 7→ 2x and x 7→ 0 of endomorphisms
of Z whose coequalizer in Ab is finite and in Set it is infinite.

3.9 Remark. We provided a simple characterization of monomorphisms (1.18)

and regular epimorphisms (3.5) in algebraic categories. There does not seem
to be a simple characterization of the dual concepts (epimorphisms and regular
monomorphisms). In fact, there exist algebraic categories with non-surjective
epimorphisms and with non-regular monomorphisms, as we show in the follow-
ing example.

3.10 Example. Monoids. These are algebras with one associative binary op-
eration and one constant which is a neutral element. The category Mon of
monoids and homomorphisms is algebraic, see 13.14.
An example of an epimorphism which is not regular is the embedding

i: Z → Q

of the multiplicative monoid of integers into that of rational numbers. In fact,

consider monoid homomorphisms h, k : Q → A such that h · i = k · i. That
is, h(n) = k(n) for every integer n. To prove h = k, it is sufficient to verify
h(1/m) = k(1/m) for all integers m 6= 0 : this follows from h(m) · h(1/m) =
k(m) · k(1/m) = 1 (since h(1) = k(1) = 1). Consequently, i is not a regular
epimorphism. Observe that i is also a monomorphism, but not a regular one.

3.11 Remark. Recall that in a finitely complete category A relations on an ob-

ject A are the subobjects of A×A. A relation can be represented by a monomor-
phism r : R → A × A or by a parallel pair r1 , r2 : R ⇉ A of morphisms that are
jointly monic. The following definitions generalize the corresponding concepts
for relations in Set :

3.12 Definition. A relation r1 , r2 : R ⇉ A in a category A is called

1. reflexive if the pair r1 , r2 is reflexive (r1 ·d = id = r2 ·d for some d : A → R),

2. symmetric if there exists s : R → R with r1 = r2 · s and r2 = r1 · s

and

3. transitive provided that for a pullback R of r1 and r2 there exists a mor-

November 20, 2009 26

CHAPTER 3. REFLEXIVE COEQUALIZERS

phism t : R → R such that the diagram

RO
t

r1 R? r2
r1  ?? r
? 2
 ??
?
 
R@ R
~~ @@ ~ ~ @@@
~ @@ ~ @@
~~r r2 @@ ~~r r2 @@
 ~~ 1  ~~ 1  
A A A
commutes.
An equivalence relation is a relation which is reflexive, symmetric and transitive.
3.13 Remark.
1. If r1 , r2 : R ⇉ A is an equivalence relation, then the morphisms d, s and t
of 3.12 are necessarily unique.
2. An equivalence relation in Set is precisely an equivalence relation in the
usual sense.
3. Given a relation r1 , r2 : R ⇉ A and an object X, we can define a relation
∼R on the hom-set A(X, A) as follows: f ∼R g if there exists a morphism
h : X → R such that r1 · h = f and r2 · h = g. It is easy to check that
r1 , r2 : R ⇉ A is an equivalence relation in A iff ∼R is an equivalence
relation in Set for all X in A.
4. Kernel pairs are equivalence relations.
5. In the category of Σ-algebras, see 1.14, an equivalence relation on an al-
gebra A is precisely a subobject of A × A which, as a relation on the
underlying set of A, is an equivalence relation in Set. These relations are
usually called congruences on A. We refer to them as equivalence rela-
tions in Σ-Alg because the concept “congruence” is reserved (with the
only exceptions of 11.31 – 11.33) for congruences of algebraic theories, see
Chapter 10.
3.14 Definition. A category is said to have effective equivalence relations pro-
vided that every equivalence relation is a kernel pair.
3.15 Example. The category of posets does not have this property: take an
arbitrary poset B and an equivalence relation R on the underlying set of B
equipped with the discrete ordering, then the two projections R ⇉ B form an
equivalence relation which is seldom a kernel pair.
3.16 Definition. A category is called exact if it has

November 20, 2009 27

 CHAPTER 3. REFLEXIVE COEQUALIZERS

1. finite limits
2. coequalizers of kernel pairs
and
3. effective equivalence relations,
and if
4. regular epimorphisms are stable under pullback. That is, in every pullback

e′ /B
A
f g
 
C /D
e

if e is a regular epimorphism, then so is e′ .

3.17 Example. Set is an exact category. In fact:
1. If r1 , r2 : R ⇉ A is an equivalence relation, its coequalizer is

q : A → A/ ∼R

where ∼R is as in 3.13 (with X = 1) and q is the canonical morphism. Clearly,

r1 , r2 : R ⇉ A is a kernal pair of q.
2. Given a pullback as in 3.16 and an element x ∈ B, we choose z ∈ C with
g(x) = e(z) using that e is an epimorphism. Then (x, z) is an element of the
pullback A and e′ (x, z) = x. This proves that e′ is surjective.
3.18 Corollary. Every algebraic category is exact.
In fact, since Set is exact, so is Set S . Since equivalence relations are reflexive
pairs, the exactness of Alg T follows using 1.18 and 3.3.
3.19 Definition. We say that colimits in a category A distribute over products
if given diagrams Di : Di → A (i ∈ I) and forming the diagram
Y Y
D: Di → A , Ddi = Di di
i∈I i∈I

then the canonical morphism

Y
colim D → colim Di
i∈I

is an isomorphism.
If all Di are of a certain type we say that colimits of that type distribute
over products.
The concept of distributing over finite products is defined as above but I is
required to be finite.

November 20, 2009 28

CHAPTER 3. REFLEXIVE COEQUALIZERS

3.20 Example. In the category Set it is easy to verify that

1. filtered colimits distribute over products

and

2. all colimits distribute over finite products.

However reflexive coequalizers do not distribute over infinite products. In fact,

consider the coequalizers

fn cn
n+n // n + 1 /1
gn

where the left-hand components of fn and gn are the inclusion maps n →

Q+ 1, and
n Q the right-hand ones are i 7→ i and i 7→ i + 1, respectively. Then
Qn∈N f n , n∈N gn have a coequalizer with infinite codomain, thus, distinct from
n∈N c n .

3.21 Corollary. In every algebraic category:

1. Regular epimorphisms are stable under products: given regular epimor-

phisms ei : Ai → Bi (i ∈ I), then
Y Y Y
ei : Ai → Bi
i∈I i∈I i∈I

is a regular epimorphism.

2. Filtered colimits distribute over products.

3. Sifted colimits distribute over finite products.

In fact, since each of the three statements holds in Set, it holds in Set T , where
limits and colimits are formed objectwise. Following 1.16, 2.5 and 3.5, the
statements hold in Alg T for every algebraic theory T .

3.22 Remark.

1. In the above corollary we implicitly assume the existence of colimits in

an algebraic category. This is not a restriction, because every algebraic
category is cocomplete, as we prove in the next chapter.

2. Although in Set all colimits distribute over finite products, this is not
true in algebraic categories in general: consider the empty diagram in the
category of abelian groups. For I = {1, 2} in 3.19 and D1 = ∅ = D2 we
get colim D = Z and colim D1 × colim D2 = Z × Z.

Historical Remarks for Chapter 3

November 20, 2009 29

 CHAPTER 3. REFLEXIVE COEQUALIZERS

Reflexive coequalizers were probably first applied by F. E. J. Linton [67]. The

fact that they commute with finite products in Set (in fact, in every cartesian
closed category) is contained in the unpublished thesis of P. T. Johnstone writ-
ten in the early 1970s. The importance of reflexive coequalizers for algebraic
categories was first understood by Y. Diers [39] but his paper remained un-
noticed. It was later rediscovered by M. C. Pedicchio and R. Wood [81]. A
decisive step was taken in [5] and [6] where reflexive coequalizers were used for
establishing the algebraic duality (see Chapter 9) and for the study of abstract
“operations” present in algebraic categories except limits and filtered colimits.
Effective equivalence relations were introduced by M. Artin, A. Grothendieck
and J. L. Verdier in [13] and exact categories by M. Barr in [16]. The fact that
filtered colimits distribute with all products in Set goes back to [13].

November 20, 2009 30

Chapter 4

Algebraic categories as free

completions

In this chapter we prove that every algebraic category has colimits. Moreover,
the category Alg T is a free completion of T op under sifted colimits.

4.1 Remark. For the existence of colimits, since we already know that Alg T
has sifted colimits and, in particular, reflexive coequalizers (see 2.5 and 3.3), all
we need to establish is the existence of finite coproducts. Indeed, coproducts
then exist because they are filtered colimits of finite coproducts. And coproducts
and reflexive coequalizers construct all colimits, see 0.7. The first step towards
the existence of finite coproducts has already been done in Lemma 1.5: finite
coproducts of representable algebras, including an initial object, exist in Alg T .

In the next lemma we use the category of elements El A of a functor A : T →

Set introduced in 0.14.

4.2 Lemma. Given an algebraic theory T , for every functor A in Set T the
following conditions are equivalent:

1. A is an algebra,

2. El A is a sifted category

and

3. A is a sifted colimit of representable algebras.

Proof. 2 ⇒ 3 : This follows from 0.14.

3 ⇒ 1 : Representable functors are objects of Alg T (1.4), and Alg T is closed in
Set T under sifted colimits (2.5).
1 ⇒ 2 : Following 2.15, it suffices to prove that (El A)op has finite products. This
is obvious: for example, the product of (X, x) and (Z, z) is (X ×Z, (x, z)) : recall
that (x, z) ∈ AX × AZ = A(X × Z).

31
 CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

4.3 Remark. An analogous result (with a completely analogous proof) holds

for small categories T with finite limits: a functor A : T → Set preserves finite
limits iff El A is a filtered category iff A is a filtered colimit of representable
functors.
4.4 Lemma.
1. If two functors F : D → C and G : B → A are final, then the product
functor F × G : D × B → C × A is final.
2. A product of two sifted categories is sifted.
Proof. 1. This follows from 2.13.3 because, for any object (d, b) in D × B, the
slice category (d, b) ↓ F ×G is nothing but the product category (d ↓ F )×(b ↓ G),
and the product of two connected categories is connected.
2. Obvious from 1 and 2.14.
4.5 Theorem. Every algebraic category is cocomplete.
Proof. As explained at the beginning of this chapter, we only need to estab-
lish finite coproducts A + B in Alg T . Express A as a sifted colimit of repre-
sentable algebras (4.2)
A = colim (YT · ΦA )
and analogously for B. The category

D = El A × El B

is sifted by 4.4 and for the projections P1 , P2 of D we have two colimits in Alg T
over D :
A = colim YT · ΦA · P1 and B = colim YT · ΦB · P2 .
The diagram D : D → Alg T assigning to every pair (X, x) and (Z, z) a coproduct
of the representable algebras (see 1.5)

D((X, x), (Z, z)) = YT · ΦA (x) + YT · ΦB (z) (in Alg T )

is sifted, thus it has a colimit in Alg T . Since colimits over D commute with
finite coproducts, we get

colim D = colim YT · ΦA (x) + colim YT · ΦB (z) = A + B.

(x,z) (x,z)

4.6 Example. Coproducts.

1. In the category Ab of abelian groups finite coproducts are finite products:
the abelian group A × B together with the homomorphisms

(−, 0) : A → A × B and (0, −) : B → A × B

is a coproduct of A and B.

November 20, 2009 32

CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

`
2. `
Infinite coproducts
Q i∈I Ai are directed colimits of finite subcoproducts
j∈J Aj = j∈J A j (for J ⊆ I finite).
3. In the category of sequential automata (1.20) the product A × B of two
automata is the machine working simultaneously in A and B on the given
(joint) input streams. Whereas the coproduct A + B is the machine work-
ing, on a given input streams, completely in A or completely in B.
4. A coproduct of graphs in Graph is given by the disjoint union of vertices
and the disjoint union of edges.
4.7 Example. Coequalizers.
1. In Ab a coequalizer of homomorphisms f, g : A ⇉ B is the quotient c : B →
B/B0 modulo the subgroup B0 ⊆ B of the elements f (a) − g(a) for all
a ∈ A.
2. A coequalizer of a parallel pair f, g : A ⇉ B in Graph is given by forming
the coequalizers in Set of (i) the two vertex functions and (ii) the two edge
functions.
4.8 Remark. Before characterizing algebraic categories as free completions
under sifted colimits, let us recall the general concept of a free completion of a
category C : this is, roughly speaking, a cocomplete category A in which C is a
full subcategory such that every functor from C to a cocomplete category has
an essentially unique extension (that is, unique up to natural isomorphism) to
a colimit-preserving functor with domain A. In the following definition we say
this more precisely. Also, for a given class D of small categories we define a free
completion under D-colimits, meaning that all colimits considered are colimits
of diagrams with domains that are elements of D.
4.9 Definition. Let D be a class of small categories. By a free completion of a
category C under D-colimits is meant a functor ED : C → D(C) such that
1. D(C) is a category with D-colimits
and
2. for every functor F : C → B, where B is a category with D-colimits, there
exists an essentially unique functor F ∗ : D(C) → B preserving D-colimits
with F naturally isomorphic to F ∗ · ED
ED
C= / D(C)
== {{
== {{ ∗
F ==
= {
{ F
 }{{
B

If D consists of all small categories, then ED : C → D(C) is called a free completion

of C under colimits.

November 20, 2009 33

 CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

4.10 Theorem. For every small category C, the Yoneda embedding

op
YC op : C → Set C

is a free completion of C under colimits.

op
Proof. The category Set C is of course cocomplete.
op
Let F : C → B be a
functor, where B has colimits. Since F ∗ : Set C → B should extend F and
preserve colimits, we are forced to define it on objects A : C op → Set (using the
notation of 0.14 applied to T = C op ) by

F ∗ A = colim (F · ΦA ) .
El A

The definition on morphisms (that is, natural transformations) h : A1 → A2

is also obvious: h induces a functor El h : El A1 → El A2 which to every ele-
ment (X, x) of A1 assigns the corresponding element (X, hX (x)) of A2 . By the
universal property of colimits, El h induces a morphism

h′ : colim (F · ΦA1 ) → colim (F · ΦA2 )

and we are forced to define F ∗ h = h′ . op

The above rule A 7→ colim (F · ΦA ) defines a functor F ∗ : Set C → B which
fulfils F ∗ · YC op ≃ F because, for A = YC op (X) = C(−, X), a colimit of F · ΦA =
B(−, F X) is F X. It remains to prove that F ∗ preserves colimits: for this we
prove that F ∗ has the following right adjoint
op
R : B → Set C , RB = B(F −, B) .

We prove the adjunction F ∗ ⊣ R by verifying that there is a bijection

op
B(F ∗ A, B) ≃ Set C (A, RB)

natural in A : C op → Set and B ∈ B. In fact, the definition of F ∗ makes it clear

that the left-hand side consists of precisely all cocones of the diagram F · ΦA
with codomain B in B :

B(F ∗ A, B) = B(colim (F · ΦA ), B) ≃ lim B(F (ΦA (X, x)), B)

The same is true about the right-hand side: recall that A = colim (YC op · ΦA )
(0.14), thus a natural transformation from A to RB is a cocone of the diagram
YC op · ΦA with codomain RB :
op op op
Set C (A, RB) = Set C (colim (YC op ·ΦA ), RB) ≃ lim Set C (YC op (ΦA (X, x)), RB) .

Yoneda Lemma tells us that morphisms from the objects YC op (X) of that dia-
gram to RB = B(F −, B) are precisely the members of the set B(F X, B) :
op
Set C (YC op (X), RB) ≃ B(F X, B) .
op
In this sense, morphisms from A to RB in Set C encode precisely the cocones
of F · ΦA with codomain B.

November 20, 2009 34

CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

4.11 Remark.

1. Although the triangle in Definition 4.9 commutes up to natural isomor-

phism only, in 4.10 it is actually always possible to choose F ∗ so that the
(strict) equality
F = F ∗ · YC

holds. This is easily seen from the above proof since if the algebra A has
the form A = YC op (X) a colimit of F · ΦA can be chosen to be F X.
op
2. Let Colim(Set C , B) be the full subcategory of B C of all functors preserv-
ing colimits. Then composition with YC op defines a functor
op
− · YC op : Colim(Set C , B) → B C .

The above universal property tells us that this functor is an equivalence.

(It is, however, not an isomorphism of categories even assuming the choice
in 1. above.)

4.12 Example.

1. A famous classical example is the free completion under filtered colimits

denoted by
EInd : C → Ind C .

For a small category C, Ind C can be described

op
as the category of all filtered
colimits of representable functors in Set C , and the functor EInd is the
codomain restriction of the Yoneda embedding.

2. One can proceed analogously with sifted colimits: we denote the free
completion under sifted colimits by

ESind : C → Sind C .

For a small category C, Sind C can be described

op
as the category of all sifted
colimits of representable functors in Set C , and ESind is the codomain
restriction of the Yoneda embedding. We are not going to prove these
results in full generality here (the interested reader can find them in [8]).
We only prove 2. under the assumption that C has finite coproducts:

4.13 Theorem. For every algebraic theory T the Yoneda embedding

YT : T op → Alg T

is a free completion of T op under sifted colimits. In other words,

Alg T = Sind (T op ) .

November 20, 2009 35

 CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

Analogously to 4.11 we have, for every functor F : T op → B, a choice of a sifted

colimit preserving functor F ∗ : Alg T → B satisfying F = F ∗ · YT .

Proof. This is analogous to the proof of 4.10 with T = C op . Given a functor

F : T op → B where B has sifted colimits, we prove that there exists an essentially
unique functor
F ∗ : Alg T → B
preserving sifted colimits and such that F ≃ F ∗ · YT . By 0.14 we are forced to
define F ∗ on objects A = colim (YT · ΦA ) by

F ∗ A = colim (F · ΦA ) .
El A

This definition makes sense because, by 4.2, El A is sifted. As in 4.10 all that
needs to be proved is that the resulting functor F ∗ preserves sifted colimits.
In the present situation F ∗ does not have a right adjoint. Nevertheless, since
the inclusion I : Alg T → Set T preserves sifted colimits (2.5), we still have, for
RB = B(F −, B), a bijection

B(F ∗ A, B) ≃ Set T (IA, RB)

natural in A : C op → Set and B ∈ B. The argument is analogous to 4.10: both

sides represent cocones of the (sifted) diagram F · ΦA with codomain B. From
the above natural bijection one deduces that F ∗ preserves sifted colimits: for
every fixed B the functor A 7→ B(F ∗ A, B) preserves sifted colimits, now use
0.12.
4.14 Corollary. A category A is algebraic if and only if it is a free completion
of a small category with finite coproducts under sifted colimits.
4.15 Remark. Let T be an algebraic theory. If B is cocomplete and the functor
F : T op → B preserves finite coproducts, then its extension

F ∗ : Alg T → B

preserving sifted colimits has a right adjoint. In fact, since F preserves finite co-
products, the functor B 7→ B(F −, B) factorizes through Alg T and the resulting
functor
R : B → Alg T , B 7→ B(F −, B)
is a right adjoint to F ∗ .
4.16 Remark. Let T be a finitely complete small category and Lex T denote
the full subcategory of Set T of finite limits preserving functors.
1. YT : T op → Lex T preserves finite colimits.
2. The embedding Lex T → Set T preserves limits and filtered colimits.
3. Lex T is cocomplete.

November 20, 2009 36

CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

The proofs of 1. and 2. are easy modifications of 1.5, 1.16 and 2.5. Using 4.3,
the proof of 3. is analogous to that of 4.5.
4.17 Theorem. For every finitely complete small category T the Yoneda em-
bedding
YT : T op → Lex T
is a free completion of T op under filtered colimits. In other words,

Lex T = Ind (T op ) .

Proof. A functor A : T → Set preserves finite limits iff El A is filtered (4.3).

Moreover, the embedding Lex T → Set T preserves filtered colimits (4.16). The
rest of the proof is a trivial modification of the proof of 4.13, just replace “sifted”
with “filtered” everywhere.
4.18 Remark. Let T be a finitely complete small category. Analogously to
4.15, if B is cocomplete and the functor F : T op → B preserves finite colimits,
then its extension F ∗ : Lex T → B preserving filtered colimits has a right adjoint.

Historical Remarks for Chapter 4

Free colimit completions (see Theorem 4.10) were probably first described by F.
Ulmer [92], see also [48]. The completion Ind was introduced by A.Grothendieck
and J. L.Verdier in [13], but it is also contained in [48]. The completion Sind
was introduced in [8], together with its relation to algebraic categories. But a
general completion under a class of colimits is already treated in [48]. Later,
these completions were studied by a number of authors, see e.g. the results and
the references in [59] and [3].

November 20, 2009 37

 CHAPTER 4. ALGEBRAIC CATEGORIES AS FREE COMPLETIONS

November 20, 2009 38

Chapter 5

Properties of algebras

In this chapter properties of algebras such as finite presentability, regular pro-

jectivity, etc., are studied. These will be later used for a characterization of
algebraic categories.
5.1 Definition. An object A of a category A is called regular projective if its
hom-functor A(A, −) : A → Set preserves regular epimorphisms. That is, for
every regular epimorphism e : X → Z and every morphism f : A → Z there
exists a commutative triangle

A@
~~ @@ f
~~ @@
~ @@
~~
~ 
X /Z
e

5.2 Example.
1. In Set all objects are regular projective.
2. In Ab all free abelian groups are regular projective. Conversely, a regular
projective abelian group is free: express A as a regular quotient e : X → A
of a free group X and apply the previous definition to f = idA . This proves
that X ≃ A × Ker e, and then A is free (because every subgroup of a free
abelian group is free).
3. A graph G is regular projective in Graph (see 1.15) iff its edges are pairwise
disjoint. That is, both functions Ge → Gv are monomorphisms.
5.3 Definition. Let A be a category. An object A of A is:
1. finitely presentable if the hom-functor A(A, −) : A → Set preserves filtered
colimits;
2. perfectly presentable if the hom-functor A(A, −) : A → Set preserves sifted
colimits.

39
 CHAPTER 5. PROPERTIES OF ALGEBRAS

5.4 Remark. Any perfectly presentable object is finitely presentable (because

filtered colimits are sifted) and, assuming the existence of kernel pairs, regular
projective (because regular epimorphisms are coequalizers of reflexive pairs).
We will show in 5.16 that in an algebraic category also the converse implication
holds: perfectly presentable objects are precisely the finitely presentable regular
projectives. In fact, the converse implication holds in any cocomplete exact
category, see 18.3.

5.5 Example. Let T be a small category.

1. In Set T the representable functors are perfectly presentable, in fact, they

have a stronger property: their hom-functors preserve all colimits. This
follows from Yoneda Lemma and the fact that in Set T colimits are formed
objectwise.

2. If T is an algebraic theory, representable functors are perfectly presentable

objects in Alg T . This follows from 1. and the fact that Alg T is closed in
Set T under sifted colimits, see 2.5.

3. Analogously, if T has finite limits, then representable functors are finitely

presentable objects in Lex T . This follows from 1. and the fact that Lex T
is closed in Set T under filtered colimits, see 4.16.

5.6 Example.

1. Every finite set is perfectly presentable in Set.

2. An abelian group A is finitely presentable in the above sense in the cate-

gory Ab iff it is finitely presentable in the usual algebraic sense: that is, A
can be presented by finitely many generators and finitely many equations.
This is easily seen from the fact that every abelian group is a filtered
colimit of abelian groups that are finitely presentable (in the algebraic
sense).
An abelian group is perfectly presentable iff it is free on finitely many
generators.

3. In a poset considered as a category the finitely (or perfectly) presentable

objects are precisely
W the compact elements x, i.e., such that for every
directed join y = i∈I yi , from x ≤ y it follows that x ≤ yi for some i ∈ I.

4. A graph is finitely presentable in Graph iff it has finitely many vertices and
finitely many edges. In fact, it is easy to see that for each such a graph G
the hom-functor Graph(G, −) preserves filtered colimits. Conversely, if G
is finitely presentable, use the fact that G is a filtered colimit of all of its
subgraphs on finitely many vertices and finitely many edges.
A graph is perfectly presentable iff it has finitely many vertices and finitely
many pairwise disjoint edges.

November 20, 2009 40

CHAPTER 5. PROPERTIES OF ALGEBRAS

5.7 Remark. We will see in Chapter 11 that the situation described for Ab in
the above example is a special case of the general fact that:
1. finite presentability has in algebraic categories the usual algebraic mean-
ing (finitely presentable objects are precisely those which can be, in the
classical sense, presented by finitely many generators and finitely many
equations),
2. every free algebra is regular projective,
3. perfectly presentable algebras are just the retracts of the free algebras on
finitely many generators.
5.8 Remark. As pointed out in 5.5, in categories Set C the representable objects
have the property that their hom-functors preserve all colimits. We call such
objects absolutely presentable. In algebraic categories, absolutely presentable
objects are typically rare. For example no abelian group A is absolutely pre-
sentable: for the initial object 1, the object Ab(A, 1) is never intial in Set.
However, the categories Set C are an exception: every object is a colimit of
absolutely presentable objects.
5.9 Lemma. If an object is regular projective (or finitely presentable or perfectly
presentable) then every retract has that property too.
Proof. If f : B → A and g : A → B are such that g ·f = idB , then the natural
transformations

α = A(g, −) : G = A(B, −) → F = A(A, −) and β = A(f, −) : F → G

fulfil β · α = G. Therefore,
α·β β
F // F /G
F

is a coequalizer. By interchange of colimits, G preserves every colimit preserved

by F.
5.10 Remark. Absolutely presentable objects in Set C are precisely the retracts
of the representable functors.
5.11 Lemma.
1. Perfectly presentable objects are closed under finite coproducts.
2. Finitely presentable objects are closed under finite colimits.
Proof. Let us prove the first statement (the proof of the second one is sim-
ilar).
` Consider aQfinite family (Ai )i∈I of perfectly presentable objects. Since
A( I Ai , −) ≃ I A(Ai , −), the claim follows from the obvious fact that a
finite product of functors A → Set preserving sifted colimits also preserves
them.

November 20, 2009 41

 CHAPTER 5. PROPERTIES OF ALGEBRAS

5.12 Corollary. Every object of an algebraic category is

1. a sifted colimit of perfectly presentable algebras

and

2. a filtered colimit of finitely presentable algebras.

In fact, 1. follows from 4.2 and 5.3, and 2. follows from 4.2, 2.20 and 5.11.

5.13 Lemma. Regular projective objects are closed under coproducts.

Proof. Let (Ai )i∈I be a family of regular projective objects and let`e : X → Z
Q a regular epimorphism. The claim follows from the formula A( I Ai , e) ≃
be
I A(Ai , e) and the fact that in Set regular epimorphisms are stable under
products (3.21).

5.14 Corollary. In every category Alg T :

1. The perfectly presentable objects are precisely the retracts of representable

algebras.

2. The regular projective objects are precisely the retracts of coproducts of

representable algebras.

Proof. 1: Following 5.5 and 5.9, a retract of a representable algebra is

perfectly presentable. Conversely, following 4.2, we can express every per-
fectly presentable algebra A as a sifted colimit of representable algebras. Since
Alg T (A, −) preserves this colimit, it follows that idA factorizes through some
of the colimit morphism e : YT (t) → A. Thus e is a split epimorphism and A is
a retract of YT (t).
2: Following 5.5, 5.9 and 5.13, a retract of a coproduct of representable algebras
is regular projective. Conversely, following 4.2, we can express every algebra A
as a colimit of representable algebras. Since Alg T is cocomplete, this implies,
by 0.7, that A is a regular quotient of a coproduct of representable algebras
a
e: T (ti , −) → A . [5.1]
i∈I
`
Therefore, if A is regular projective, it is a retract of i∈I T (ti , −).

5.15 Corollary. Every algebraic category has enough regular projective objects,
i.e., every algebra is a regular quotient of a regular projective algebra.

In fact, use [5.1]

` from the above proof: e is a regular epimorphism and, following
5.5 and 5.13, i∈I T (ti , −) is a regular projective object.

5.16 Corollary. In an algebraic category, an algebra is perfectly presentable if

and only if it is finitely presentable and regular projective.

November 20, 2009 42

CHAPTER 5. PROPERTIES OF ALGEBRAS

Proof. One implication holds in any category, see 5.4. Conversely, if P is a

regular projective object in Alg T , due to 5.14.2 P is a retract of a coproduct
of representable algebras. Since every coproduct is a filtered colimit of its finite
subcoproducts (2.20), if P is also finitely presentable, then it is a retract of a
finite coproduct of representable algebras. Following 5.5, 5.9 and 5.11.1, P is
perfectly presentable.
5.17 Proposition. In every algebraic category the finitely presentable algebras
are precisely the coequalizers of reflexive pairs of homomorphisms between rep-
resentable algebras.
Proof. One implication is obvious: representable algebras are finitely pre-
sentable (5.5), and finitely presentable objects are closed under finite colimits
(5.11).
Conversely, let A be a finitely presentable algebra. Following 4.2, A is a (sifted)
colimit of representable algebras. Thus, A is a filtered colimit of finite colimits
of representable algebras, see 2.20. From 1.5 and 0.7 we deduce that every finite
colimit of representable algebras is a reflexive coequalizer of a parallel pair be-
tween two representable algebras. Since A is finitely presentable, it is a retract
of one of the above coequalizers:
h // Y (t ) e s /
YT (t1 ) T 2
/Qo A
k u

Here e is a coequalizer of h and k, and s · u = idA . Since YT (t2 ) is regular

projective (see 5.4 and 5.5) and e is a regular epimorphism, there exists a ho-
momorphism g : YT (t2 ) → YT (t2 ) such that e · g = u · s · e. Let us observe that
the morphism s · e is a joint coequalizer of h, k and g · h :
h /
/ s·e /A
YT (t1 ) k / YT (t2 )
g·h

In fact, s · e · k = s · e · h = s · u · s · e · h = s · e · g · h. Assume that a morphism

f : YT (t2 ) → X coequalizes h, k, and g · h. Then there exists a unique homomor-
phism v : Q → X with v·e = f. Hence f ·h = f ·g·h = v·e·g·h = v·u·s·e·h. Since
h is an epimorphism (because the pair h, k is reflexive), we get f = v · u · s · e.
Now, from 0.7 we conclude that A is a reflexive coequalizer of a pair of homo-
morphisms between finite coproducts of representable algebras. By 1.5, we get
the claim.
5.18 Remark. Beside finite presentability, an important concept in general
algebra is finite generation: an algebra A is finitely generated if it has a finite
subset not contained in any proper subalgebra. (Or, equivalently, A is a regular
quotient of a free algebra on finitely many generators.) This concept also has a
categorical formulation. For this we need to introduce the following
5.19 Definition. A directed union is a filtered colimit of subobjects. That
is, given a filtered diagram D : D → A where D maps every morphism to a
monomorphism, then colim D is called the directed union.

November 20, 2009 43

 CHAPTER 5. PROPERTIES OF ALGEBRAS

5.20 Remark. In Set directed unions have colimit cocones formed by monomor-
phisms. Thus, the same holds in Set T . Since Alg T is closed under filtered
colimits in Set T , this is also true in Alg T .
5.21 Definition. Let A be a category. An object A of A is finitely generated
if the hom-functor A(A, −) : A → Set preserves directed unions.
5.22 Proposition. In every algebraic category the finitely generated algebras
are precisely the regular quotients of representable algebras.
Proof. 1. Let A be finitely generated. Recall from 4.2 that A is the sifted
colimit of
ΦA
El A / T op YT / Alg T

Let (X, x) be an object of El A and consider the regular factorization of the

colimit morphism x̂ :
x̂ /A
YT (X) ?
FF
FF
FF
qx FF mx
F"
Ix
These objects Ix and the connecting monomorphisms between them one gets
from the morphisms of El A (via diagonal fill-in, see 0.16) form a filtered diagram
of monomorphisms. Indeed, given two elements (X, x) and (Z, z) of A, for
the element (X × Z, (x, z)) we see that mx and mz both factorize through
m(x,z) . It is clear that A = colim Ix , and since A is finitely generated, A(A, −)
preserves this colimit. Thus there exists (X, x) ∈ El A and f : A → Ix such that
idA = mx · f. Therefore mx , being a monomorphism and a split epimorphism,
is an isomorphism. This implies that x̂ : YT (X) → A is a regular epimorphism.
2. For every regular epimorphism e : YT (X) → A in Alg T we prove that A is
finitely generated. Given a filtered diagram D : D → Alg T of subobjects with
colimit cocone cd : Dd → C (d ∈ obj D), since cd are monomorphisms, see 5.20,
it is sufficient to prove that every morphism f : A → C factorizes through some
cd . In fact, since YT (X) is finitely presentable, the morphism f · e : YT (X) → C
factorizes through some cd – and then we just use the diagonal fill-in:
e /A
YT (X)
f
 | 
Cd /C
cd

5.23 Example. In the category N/Set of sets with countably many constants
the finitely generated objects are those that have, beside the constants, only
finitely many elements. Whereas the finitely presentable objects have, moreover,
the property that only finitely many pairs of distinct natural numbers label the

November 20, 2009 44

CHAPTER 5. PROPERTIES OF ALGEBRAS

same constant. (Thus for example the terminal object is finitely generated but
not finitely presentable.) Finally, the absolutely presentable objects are those
finitely generated objects where the constants are pairwise distinct.

Historical Remarks for Chapter 5

The lecture notes [48] by P. Gabriel and F. Ulmer is the source of the concept
of a finitely presentable object. In [8] perfectly presentable objects were intro-
duced under the name of strongly finitely presentable. In algebraic categories,
they coincide with objects “projectif-de-type-fini” of Y. Diers [39] and with the
finitely presentable effective projectives of M. C. Pedicchio and R. J. Wood [81].
The term “perfectly presentable” was suggested by A. Joyal (see [57]), his
motivation comes from perfect complexes as explained in 6.11 below.

November 20, 2009 45

 CHAPTER 5. PROPERTIES OF ALGEBRAS

November 20, 2009 46

Chapter 6

A characterization of
algebraic categories

We have already characterized algebraic categories as free completions (see 4.14).

The aim of this chapter is to characterize them as those cocomplete categories
which have a strong generator formed by perfectly presentable objects.
6.1 Definition.
1. A set of objects G in a category A is called a generator if two morphisms
x, y : A ⇉ B are equal whenever x·g = y ·g for every morphisms g : G → A
with domain G in G.
2. A generator G is called strong if a monomorphism m : A → B is an isomor-
phism whenever every morphism g : G → B with domain G in G factorizes
through m.
6.2 Remark. Here is an equivalent way to express the notions of generator
and strong generator. Consider the functor

A → Set G , A 7→ hA(G, A)iG∈G .

1. G is a generator iff the above functor is faithful.

2. G is a strong generator iff the above functor is faithful and conservative
(see 0.2).
The following proposition suggests that “strong” generator should more
properly be called “extremal”, the present terminology has just historical rea-
sons.
6.3 Proposition. In a category A with coproducts a set of objects G is
1. a generator if and only if every object of A is a quotient of a coproduct of
objects from G,

47
 CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

2. a strong generator if and only if every object of A is an extremal quotient

of a coproduct of objects from G.

Explicitly: G is a (strong) generator iff for every object A in A, an (extremal)

epimorphism a
e: Gi → A
i∈I

exists with all Gi in G. We will see in the proof that this is equivalent to saying
that the canonical morphism
a
eA : G→A
(G,g)∈G↓A

whose (G, g)-component is g, is an (extremal) epimorphism.

Proof. 1. If G is a generator, then ` the canonical morphism eA is obviously

an epimorphism. Conversely, if e : i∈I Gi → A is an epimorphism, then given
distinct morphisms x, y : A ⇉ B, some component ei : Gi → A fulfils x·ei 6= y·ei .
2. Let G be a strong generator. If eA factorizes through a monomorphism m,
then all of its components g : G → A factorize through m. Since G ` is a strong
generator, this implies that m is an isomorphism. Conversely, let e : i∈I Gi →
A be an extremal epimorphism, and consider a monomorphism m : B → A. If
every morphism g : G → A (with G varying in G) factorizes through m, then e
factorizes through m, so that m is an isomorphism.

6.4 Corollary. If A has colimits and every object of A is a colimit of objects

from a set G, then G is a strong generator.

In fact, this follows from 0.7 and 6.3, because any regular epimorphism is ex-
tremal.

6.5 Example.

1. Every nonempty set forms a (singleton) strong generator in Set.

2. The group of integers forms a strong generator in Ab.

3. In the category of posets and order-preserving functions the terminal (one-

element) poset forms a generator – but this generator is not strong. In
contrast, the two-element chain is a strong generator.

6.6 Example.

1. Let C be a small category. Then Set C has a strong generator formed by

all representable functors. This follows from 0.14 and 6.4.

2. Analogously, for an algebraic theory T the category Alg T has a strong

generator formed by all representable algebras: see 6.4 and 4.2.

November 20, 2009 48

CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

6.7 Lemma. Let A be a cocomplete category with a set of perfectly presentable

objects such that every object of A is a sifted colimit of objects of that set. Then
A has, up to isomorphism, only a set of perfectly presentable objects.
Proof. Express an object A of A as a sifted colimit of objects from the given
set G. If A is perfectly presentable, then it is a retract of an object from G.
Since each object from A has only a set of retracts (because each retract of an
object B gives rise to an idempotent morphism e : B → B, e · e = e), our claim
is proved.
6.8 Remark. A result analogous to 6.7 holds for finitely presentable objects
and filtered colimits. The proof is the same.
6.9 Theorem. (Characterization of algebraic categories) The following condi-
tions on a category A are equivalent:
1. A is algebraic;
2. A is cocomplete and has a set G of perfectly presentable objects such that
every object of A is a sifted colimit of objects of G;
3. A is cocomplete and has a strong generator consisting of perfectly pre-
sentable objects.
Moreover if the strong generator G in 3. is closed under finite coproducts, then
the dual of G (seen as a full subcategory) is an algebraic theory of A.
Proof. 1 ⇒ 2 : Let T be an algebraic theory. Then Alg T is cocomplete (4.5),
the representable algebras form a set of perfectly presentable objects (5.5), and
every algebra is a sifted colimit of representable algebras (4.2).
2 ⇒ 3 : Consider the family App of all perfectly presentable objects of A. By
6.7, App is essentially a set. By 6.4, it is a strong generator.
3 ⇒ 1 : Let G be a strong generator consisting of perfectly presentable objects.
Since perfectly presentable objects are closed under finite coproducts (5.11), we
can assume without loss of generality that G is closed under finite coproducts (if
this is not the case, we can replace G by its closure in A under finite coproducts,
which still is a strong generator). We are going to prove that A is equivalent to
Alg (G op ), where G is seen as a full subcategory of A.
(a) We prove first that G is dense, i.e., for every object K of A the canonical
diagram of all arrows from G
DK : G ↓ K → A , (g : G → K) 7→ G
has K as colimit, with (g : G → K) as colimit cocone. To prove this, form a
colimit cocone of DK :
(cg : G → K ∗ ) for all g : G → K in G ↓ K.
We have to prove that the unique factorizing morphism λ : K ∗ → K with λ·cg =
g for all g in G ↓ K is an isomorphism. Consider the coproduct
a
G
(G,g)∈G↓K

November 20, 2009 49

 CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

`
with coproduct injections ρg : G → (G,g)∈G↓K G. We have a commutative tri-
angle
`
(G,g)∈G↓K G
r KK
v rrr KKeK
KK
rr r KK
rx rr KK
%/
K∗ λ
K

where eK is the canonical morphism 6.3 and v is defined by v · ρg = cg for

all (G, g) ∈ G ↓ K. Since eK is an extremal epimorphism (see 6.3), then λ
is an extremal epimorphism. It remains to prove that λ is a monomorphism.
Consider two morphisms x, x′ : X ⇉ K ∗ such that λ · x = λ · x′ , and let us prove
that x = x′ . While G is a (strong) generator, we can assume without loss of
generality that X is in G. Since G ↓ K is sifted (in fact, it has finite coproducts,
because G has, see 2.15) and X is perfectly presentable, both x and x′ factorize
through some colimit morphism: that is, for some (G, g) and (G′ , g ′ ) in G ↓ K
we have a commutative diagram

XB X′
BB ′ {{
BBx x {{
y BB { y′
 B! }{{{ 
G B cg / K ∗ o cg′ G′
BB {{
BB {{
g BBB λ {{{ g′
!  }{
K

Since cλ·x = cg · y = x and analogoulsy cλ·x′ = x′ , we get x = cλ·x = cλ·x′ = x′ .

(b) It follows from (a) that the functor

E : A → Alg (G op ) , K 7→ A(−, K)

is full and faithful. Indeed, given a homomorphism α : A(−, K) → A(−, L),

for every g : G → K in G ↓ K we have a morphism αG (g) : G → L. Those
morphisms form a cocone on G ↓ K, so that there exists a unique morphism
α̂ : K → L such that α̂ · g = αG (g) for all g in G ↓ K. It is easy to check that
E α̂ = α and that Ef c = f for all f : K → L in A. It remains to prove that E is
essentially surjective on objects.
(c) Let us prove first that E preserves sifted colimits. Consider a sifted diagram
D : D → A with colimit (hd : Dd → H). For every object G in G a colimit of
A(G, −) · D in Set is A(G, H) with the colimit cocone A(G, Dd) → A(G, H)
given by composition with hd (because G is formed by perfectly presentable
objects). This implies that (Ehd : E(Dd) → EH) is a colimit of E · D in
Alg (G op ) (sifted colimits are computed objectwise in Alg (G op ), see 2.5).
(d) It follows from (c) that E is essentially surjective on objects. In fact, we

November 20, 2009 50

CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

have the following diagram, commutative up to natural isomorphism,

∗
I E
Alg (G op ) /A
O
/ Alg (G op )
hPPP n n6
PPP nnn
PPP I n
YGop PPP nnn
PP nnnnn YGop
G

where I is the inclusion and I ∗ is its extension preserving sifted colimit (4.13).
Since E · I ∗ · YG op ≃ YG op and E · I ∗ preserves sifted colimits, it follows from 4.13
that E · I ∗ is naturally isomorphic to the identity functor. Thus E is essentially
surjective on objects.

6.10 Example.

1. The category Pos of posets and order-preserving maps is not algebraic:

only the discrete posets are perfectly presentable, and there exists no
strong generator formed by discrete posets.

2. In the category Bool of Boolean algebras consider the free algebras PPn
on n generators (where PX is the algebra of all subsets of a set X). The
dual of the category PPn (n ∈ N) is an algebraic theory for Bool. In fact,
PPn are perfectly presentable and form a strong generator closed under
finite coproducts.

6.11 Example. Let R be a unitary ring. We denote by Ch(R) the category

of chain complexes of left R-modules. Its objects are collections X = (Xn )n∈Z
of left R-modules equipped with a differential, that is, a collection of module
homomorphisms
d = (dn : Xn → Xn−1 )n∈Z
where dn−1 · dn = 0 for each n. Morphisms f : X → Y are chain maps, i.e.,
collections (fn : Xn → Yn )n∈Z of module homomorphisms such that dn · fn =
fn−1 · dn for all n.
A complex X is bounded if there are only finitely many n ∈ N with Xn 6= 0.
Since every complex is a filtered colimit of its truncations, each finitely pre-
sentable complex is bounded. Perfectly presentable objects in Ch(R) are pre-
cisely the bounded complexes of perfectly presentable left R-modules. Since
they form a strong generator of Ch(R), the category Ch(R) is algebraic.

6.12 Notation. We denote by App a full subcategory of A representing all

perfectly presentable objects of A up to isomorphism.

6.13 Corollary. For every algebraic category A the dual of App is an algebraic
theory of A : we have an equivalence functor

E : A → Alg (Aop
pp ) , A 7→ A(−, A) .

In fact, App is a strong generator closed under finite coproducts.

November 20, 2009 51

 CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

6.14 Corollary. Two algebraic categories A and B are equivalent if and only
if the categories App and Bpp are equivalent.

Proof. This follows immediately from 6.13 and the fact that equivalence
functors preserve perfectly presentable objects.

From 1.10 we know that the slice category Set ↓ S, equivalent to the category
Set S of S-sorted sets, is algebraic. This is a particular case of a more general
fact:

6.15 Proposition. Every slice category A ↓ A of an algebraic category A is

algebraic.

Proof. The category A ↓ A is cocomplete: consider a small category D and

a functor
F : D → A ↓ A , F D = (FD , fD : FD → A) .
A colimit of F is given by (C, c), where (C, σD : FD → C) is a colimit of ΦA · F
where ΦA : A ↓ A → A is the forgetful functor, and c · σD = fD for every object
D in D. This immediately implies that an object (G, g) is perfectly presentable
in A ↓ A as soon as G is perfectly presentable in A. Let now G be a strong
generator of A. Then the set of objects G ↓ A = {(G, g) | G ∈ G} is a strong
generator of A ↓ A. This is so because a morphism f : (X, x) → (Z, z) in A ↓ A
is a strong epimorphism iff f : X → Z is a strong epimorphism in A. Following
6.9, A ↓ A is algebraic.

6.16 Lemma. Let the functor I : A → B have a left adjoint R.

1. If I is faithful and conservative and G is a strong generator of B, then

R(G) is a strong generator of A.

2. If I preserves sifted colimits and X is perfectly presentable in B, then RX

is perfectly presentable in A.

Proof. 1: R(G) is a generator because I is a faithful right adjoint. Next,

consider a monomorphism a : A → A′ in A such that every morphism RG →
A′ with G ∈ G factorizes through a. This implies, by adjunction, that every
morphism G → IA′ factorizes through the monomorphism Ia. Since G is a
strong generator, Ia is an isomorphism, and since I is conservative, a is an
isomorphism.
2: Since A(RX, −) ≃ B(X, I−) = B(X, −) · I, we see that A(RX, −) is the
composite of two functors preserving sifted colimits.

6.17 Proposition. Let T be an algebraic theory. Then Alg T is a reflective

subcategory of Set T closed under sifted colimits.

Proof. This is a special case of the adjunction F ∗ ⊣ B(F −, −) obtained in the

proof of 4.10. Indeed, by the Yoneda Lemma, the right adjoint Alg T (YT −, −)
is naturally isomorphic to the full inclusion Alg T → Set T .

November 20, 2009 52

CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

6.18 Theorem. A category is algebraic if and only if it is equivalent to a re-

flective subcategory of Set C closed under sifted colimits, for some small category
C.
Proof. Let T be an algebraic theory. Following 6.17 and 2.5, Alg T is a
reflective subcategory closed under sifted colimits of Set T .
Conversely, let C be a small category. By 1.6 Set C is an algebraic category,
so that it fulfils the conditions of 6.9.3. Following 6.16, those conditions are
inherited by any reflective subcategory closed under sifted colimits of Set C .
6.19 Corollary. The functor category AD of an algebraic category A is alge-
braic.
Proof. Due to 6.18, there exist a small category C and a reflection

o R
A / Set C
I

with the right adjoint I preserving sifted colimits. This induces another reflec-
tion
R·−

AD
o / Set C D
I·−

with I · − preserving sifted colimits because they are formed objectwise. Since

D
Set C ≃ Set C×D , by 6.18 AD is algebraic.
6.20 Remark. Our Characterization Theorem 6.9 shows a strong parallel be-
tween algebraic categories and the following more general concept due to Gabriel
and Ulmer [48].
6.21 Definition. A category is called locally finitely presentable if it is cocom-
plete and has a set G of finitely presentable objects such that every object of A
is a filtered colimit of objects of G.
6.22 Example.
1. Following 6.9, all algebraic categories are locally finitely presentable.
2. If T is a small category with finite limits, then Lex T , see 4.16, is a lo-
cally finitely presentable category. In fact Lex T is cocomplete (4.16), the
representable functors form a set of finitely presentable objects (5.5), and
every object is a filtered colimit of representable functors (4.3).
3. The category Pos of posets (which is not algebraic, see 6.10) is locally
finitely presentable: the two-element chain which forms a strong generator
is finitely presentable. Thus we can apply the following
6.23 Theorem. (Characterization of locally finitely presentable categories) The
following conditions on a category A are equivalent:
1. A is locally finitely presentable;

November 20, 2009 53

 CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

2. A is equivalent to Lex T for a small category T with finite limits;

3. A is cocomplete and has a strong generator formed by finitely presentable
objects.
Moreover if the strong generator G in 3. is closed under finite colimits, then A
is equivalent to Lex (G op ).
Proof. This proof is quite analogous to that of Theorem 6.9: just change
Alg T to Lex T for T = G op , work with finitely presentable objects instead of
perfectly presentable ones, and in the proof of 3 ⇒ 2 use the closure under finite
colimits.
6.24 Corollary. A category A is locally finitely presentable if and only if it is
a free completion of a small, finitely cocomplete category under filtered colimits.
In fact, this follows from 4.17 and 6.23.
6.25 Notation. We denote by Af p a full subcategory of A representing all
finitely presentable objects of A up to isomorphism.
6.26 Corollary. Every locally finitely presentable category A is equivalent to
Lex T for some small, finitely complete category T . In fact, we have the equiv-
alence functor
E : A → Lex (Aopf p ) , A 7→ A(−, A) .
This follows from 6.23 applied to G = Af p .
6.27 Proposition. Let T be a small finitely complete category. Then Lex T is
a reflective subcategory of Set T closed under filtered colimits.
Proof. This is a special case of the adjunction F ∗ ⊣ B(F −, −) obtained in the
proof of 4.10. Indeed, by the Yoneda Lemma, the right adjoint Lex T (YT −, −)
is naturally isomorphic to the full inclusion Lex T → Set T .
6.28 Theorem. A category is locally finitely presentable if and only if it is
equivalent to a reflective subcategory of Set C closed under filtered colimits, for
some small category C.
6.29 Corollary. Let T be a small finitely complete category. Then Lex T is a
reflective subcategory of Alg T closed under filtered colimits.
Proof. Consider the full inclusions
I1
Lex TH / Set T
HH v ;
HH vv
HH vv
I2 HH vv
$ vv I3
Alg T
By 6.27, I1 has a left adjoint, say R. Since I3 is full and faithful, R · I3 is left
adjoint to I2 . Finally, I2 preserves filtered colimits because I1 preserves them
by 6.27 and I3 reflects them.

In 12.12 we will need the following fact.

November 20, 2009 54

CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

6.30 Corollary. Let A be an algebraic category and B a category with sifted

colimits. If in A every finitely presentable object is regular projective, then a
functor F : A → B preserving filtered colimits preserves also sifted colimits.
Proof. Following 6.13 and 6.26
op
Alg (Aop
pp ) ≃ A ≃ Lex (Af p ) .

If moreover finitely presentable objects in A are regular projective, then App =

Af p (5.16). The result now follows from the universal properties stated in 4.13
and 4.17.
6.31 Example. Set (and more generally Set S ) and the category of vector spaces
over a field are examples of algebraic categories where every (finitely presentable)
object is regular projective. More generally, in the category of left modules over
a semi-simple ring every object is regular projective.
Let us finish this chapter by quoting from [30] another characterization theo-
rem similar to 6.9. Recall that Set C is the free completion of C op under colimits
4.10; recall also the concept of absolutely presentable object in 5.8.
6.32 Theorem. A category is equivalent to a functor category Set C for some
small category C if and only if it is cocomplete and has a strong generator con-
sisting of absolutely presentable objects.

Historical Remarks for Chapter 6

The concept of locally finitely presentable category, due to P. Gabriel and F.

Ulmer [48], was an attempt of a categorical approach to categories of finitary
structures generalizing Lawvere’s algebraic theories. This is the source of 6.23.
See [7] for a more recent monograph on locally presentable categories. In [8] the
analogy between locally finitely presentable categories and algebraic categories
was made explicit.
A characterization of categories of algebras for one-sorted algebraic theories
is contained in the thesis of F. W. Lawvere [63]. The characterization 6.9 is
taken from [8]. But the equivalence of 1. and 3. was already proved by Y. Diers
[39]. A first characterization of categories of Set-valued functors is in the thesis
of M. Bunge [30] (the first proof was published in [68]). There is a general result
covering both 6.9, 6.23 and 6.32 (see [34]).

November 20, 2009 55

CHAPTER 6. A CHARACTERIZATION OF ALGEBRAIC CATEGORIES

November 20, 2009 56

Chapter 7

From filtered to sifted

The aim of this chapter is to demonstrate that the “equation”

sifted colimits = filtered colimits + reflexive coequalizers

is almost valid – but not quite. What we have in mind are three facts:

1. A category C has sifted colimits iff it has filtered colimits and reflexive
coequalizers. This holds whenever C has finite coproducts – and in general
it is false.

2. A functor preserves sifted colimits iff it preserves filtered colimits and

reflexive coequalizers. This holds whenever the domain category is finitely
cocomplete – and in general it is false.

3. The free completion Sind C of a small category C under sifted colimits is

obtained from the free completion Rec C under reflexive coequalizers by
completing it under filtered colimits

Sind C = Ind (Rec C) .

This holds whenever C has finite coproducts – and in general it is false.

We begin by describing Rec C in a manner analogous to the description of Sind C

and Ind C, see 4.13 and 4.17. A quite different approach to Rec C is treated in
Chapter 17.
As a special case of Definition 4.9, we get the following

7.1 Definition. By a free completion of a category C under reflexive coequalizers

is meant a functor ERec : C → Rec C such that

1. Rec C is a category with coequalizers of reflexive pairs

and

57
 CHAPTER 7. FROM FILTERED TO SIFTED

2. for every functor F : C → B, where B is a category with reflexive coequal-

izers, there exists an essentially unique functor F ∗ : Rec C → B preserving
reflexive coequalizers with F naturally isomorphic to F ∗ · ERec .
Recall that for a category A we denote by Af p the full subcategory of finitely
presentable objects.
7.2 Lemma. Let T be an algebraic theory. The inclusion I : (Alg T )f p → Set T
preserves reflexive coequalizers.
Proof. This follows from the fact that (Alg T )f p is closed in Alg T under finite
colimits (5.11) and that Alg T is closed in Set T under sifted colimits (2.5), see
also 3.2.
7.3 Theorem. For every algebraic theory T the restricted Yoneda embedding

YT : T op → (Alg T )f p

is a free completion of T op under reflexive coequalizers. In other words,

(Alg T )f p = Rec (T op ) .

Proof. Recall from 3.2 the category M

f1
/
P o d /Q modulo f1 · d = idQ = f2 · d .
f2

We will prove that given a finitely presentable algebra A : T → Set, there exists
a final functor (see Definition 2.12)

M : M → El A .

The rest of the proof is analogous to the proof of Theorem 4.10: given a functor
F : T op → B where B has reflexive coequalizers, we prove that there exists an
essentially unique functor

F ∗ : (Alg T )f p → B

preserving reflexive coequalizers and such that F ≃ F ∗ · YT . In fact, using

the notation of 0.14 for the above final functor M we see that the reflexive
coequalizer of F · ΦA · M (fi ) (i = 1, 2) in B is just the colimit of F · ΦA , thus,
the latter colimit exists, and we are forced to define

F ∗ A = colim (F · ΦA )
El A

on objects. This extends uniquely to morphisms (as in the proof of 4.10) and
yields a functor F ∗ : (Alg T )f p → B. Since the inclusion I : (Alg T )f p → Set T
preserves reflexive coequalizers (7.2), we have for RB = B(F −, B) a bijection

B(F ∗ A, B) ≃ Set T (IA, RB)

November 20, 2009 58

CHAPTER 7. FROM FILTERED TO SIFTED

natural in A and B, from which one deduces that F ∗ preserves reflexive co-
equalizers.
To prove the existence of the final functor M, recall from 5.17 that there
exists a reflexive pair
f1
/
P o d /Q
f2

in T such that A is a coequalizer of YT (fi ) :

YT (f1 )
// c /A
YT (P ) YT (Q)
YT (f2 )

Put c = c · YT (fi ) and define objects of El A as follows:

M P = (P , cP (idP )) and M Q = (Q, cQ (idQ )) .

Since clearly fi : M P → M Q and d : M Q → M P are morphisms of El A, we

obtain a functor M = (−) : M → El A. Let us prove its finality:
1. Every object (X, x) of El A has a morphism into M Q. In fact,

cX : T (Q, X) → A(X)

is an epimorphism, thus, for x ∈ A(X) there exists f : Q → X with

cX (f ) = x, which implies that f : (X, x) → M Q is a morphism of El A.
2. Given two morphisms of El A from (X, x) to M P or M Q, they are con-
nected by a zig-zag in the slice category (X, x) ↓ M. In fact, we can
restrict ourselves to the codomain M Q : the general case is then solved
by composing morphisms with codomain M P by M f1 .
Given morphisms
h, k : (X, x) → M Q
then we have
A(h) · cQ (idQ ) = x = A(k) · cQ (idQ )
which, due to naturality of c, yields

cX (h) = cX (k) .

Now use the description of coequalizers in Set (see 0.6): since cX is the
coequalizer of YT (f1 )and YT (f2 ), there is a zig-zag of this pair connecting
h and k. For example a zig-zag of length 2:

v?
YT (f1 )  ???YT (f2 )
 ??
  ?
h k

November 20, 2009 59

 CHAPTER 7. FROM FILTERED TO SIFTED

for some v : P → X. This means h = v · f1 and k = v · f2 and yields the

following zig-zag in (X, x) ↓ M :

(X, x)
w GG
h www GG k
v GG
www GG
G#
{ww 
MQ o MP / MQ
Mf1 Mf2

Analogously for longer zig-zags.

7.4 Corollary. For a small category C with finite coproducts we have

Sind C = Ind (Rec C) .

More precisely, the composition

ERec EInd
C / Rec C / Ind (Rec C)

is a free completion of C under sifted colimits.

In fact, for T = C op the above theorem yields Rec C = (Alg T )f p from which
6.26 and 4.17 prove Ind (Rec C) = Alg T . Now apply 4.13.

7.5 Remark. In the proof of Theorem 7.3 if B has finite colimits and F pre-
serves finite coproducts, then the extension F ∗ preserves finite colimits. This
follows from the fact that B(F −, B) lies now in Alg T and that (Alg T )f p is
closed in Alg T under finite colimits (5.11), so that we have a bijection

B(F ∗ A, B) ≃ Alg T (A, B(F −, B))

natural in A ∈ (Alg T )f p and B ∈ B.

7.6 Remark. In the introduction to this chapter we claimed that a functor

defined on a finitely cocomplete category preserves sifted colimits iff it preserves
filtered colimits and reflexive coequalizers. The proof of this result can be
found in [11], here we present a (simpler) proof based on 7.4 which requires
cocompleteness of both categories (in fact, sifted colimits are enough as far as
the codomain category is concerned).

7.7 Theorem. A functor between cocomplete categories preserves sifted colimits

if and only if it preserves filtered colimits and reflexive coequalizers.

Proof. Necessity is clear. To prove sufficiency, let F : E → A preserve filtered

colimits and reflexive coequalizers where E and A are cocomplete. For every
sifted diagram D : D → E choose a small full subcategory U : C ֒→ E containing

November 20, 2009 60

CHAPTER 7. FROM FILTERED TO SIFTED

the image of D and closed in E under finite coproducts. Consider the following
diagram
(F ·U)∗
Sind C /A
H O 777 O
 77
77
 77
 77
EInd 77U ∗
ESind  F
 77
7
 Rec; C PPP 77
 xxxx PPPU ′ 77
 xx PPP 77
xxx ERec PPP 7
P'
C /E
U

where U ∗ and (F · U )∗ are the extensions of U and F · U, respectively, preserving

sifted colimits, and U ′ is the extension of U preserving reflexive coequalizers.
Since by 7.4 we have ESind = EInd · ERec , it follows that

(F · U )∗ · EInd · ERec ≃ (F · U )∗ · ESind ≃ F · U ≃ F · U ′ · ERec .

The functor EInd preserves reflexive coequalizers by 4.16 and 4.17, and so do
the functors F, U ′ and (F · U )∗ . Thus, the universal property of ERec yields

(F · U )∗ · EInd ≃ F · U ′ .

Since
U ∗ · EInd · ERec ≃ U ∗ · ESind ≃ U ≃ U ′ · ERec
and, once again, the functors U ∗ , EInd and U ′ preserve reflexive coequalizers,
we have
U ∗ · EInd ≃ U ′ .
Finally, since
F · U ∗ · EInd ≃ F · U ′ ≃ (F · U )∗ · EInd
and the functors F, U ∗ and (F · U )∗ preserve filtered colimits, we have

F · U ∗ ≃ (F · U )∗ . [7.1]

We are ready to prove F (colim D) ≃ colim (F · D). Let

D′ : D → C with D = U · D′

be the codomain restriction of D. Then

F · D = F · U · D′ ≃ (F · U )∗ · ESind · D′

implies, since (F · U )∗ preserves the sifted colimit of ESind · D′ , that

colim (F · D) ≃ (F · U ∗ )(colim (ESind · D′ )) .

November 20, 2009 61

 CHAPTER 7. FROM FILTERED TO SIFTED

From [7.1] and the fact that U ∗ also preserves the sifted colimit of ESind · D′
we derive
colim (F · D) ≃ F (colim (U ∗ · ESind · D′ )) ≃ F (colim D) .

7.8 Remark. We thus established proofs of the affirmative statements 2. and

3. of the introduction of this chapter. The statement 1. is easy: if C has filtered
colimits, reflexive coequalizers, and finite coproducts, then it has all colimits
(4.1).
With the following examples we demonstrate the negative parts of the state-
ments 1.–3.
7.9 Example. Here we give an example of a category not having sifted colimits
although it has (i) filtered colimits and (ii) reflexive coequalizers.
We start with the following category D given by the gluing of two reflexive pairs
at their codomains. That is, D is given by the graph
a1 a′1
/ o
Ao d /Bo d′ / A′
a2 a′2

and the equations making both parallel pairs reflexive:

ai · d = idB = a′i · d′ for i = 1, 2 .
The proof that D is sifted is completely analogous to the proof of Example 3.2:
we verify that colimits over D in Set commute with finite products. Assume
that the above graph depicts sets A, B and A′ and functions between them.
Then a colimit can be described as the canonical function c : B → C = B/ ∼
where two elements x, y ∈ B are equivalent iff they are connected by a zig-zag
formed by a1 , a2 , a′1 and a′2 . Since the two pairs are reflexive, the length of the
zig-zag can be arbitrarily prolonged. And the type can be chosen to be
z1 > z2 > z3 > z4 > z4k B
~ >> a2 >> a1 a′1 >> a′2 a′2 >> a′1 BB a′1
a1
~~~ >> a2 >> >> >> ...... BB
~ >> >> >> >> BB
~~
~    
x y
(here for the elements z4i+1 we use a1 , a2 , for z4i+2 we use a2 , a1 , for z4i+3 we
use a′1 , a′2 and for z4i we use a′2 , a′1 ). From that it is easy to derive that D is
sifted.
We now add to D the coequalizers c1 of a1 , a2 and c′ of a′1 , a′2 : let E be the
category given by the graph
a1 1 a′
/ o
Ao d /~ B ?o ? d′ / A′
a2 ~ ?? a′2
~~ ?
~
~ ′ c ???
~~~ c 
C′ C

November 20, 2009 62

CHAPTER 7. FROM FILTERED TO SIFTED

with the previous equations plus the following ones

c · a1 = c · a2 and c′ · a′1 = c′ · a′2 .

The sifted diagram D → E which is the inclusion does not have a colimit.
However, E has reflexive coequalizers because its only nontrivial reflexive pairs
are a1 , a2 (with coequalizer c) and a′1 , a′2 (with coequalizer c′ ). Moreover, E has
filtered colimits: since the category E is clearly finite, it does not have any
nontrivial filtered diagram except those obtained by iterating an idempotent
endomorphism e, see 2.3. Thus it is sufficient to verify that E has coequalizers
of all pairs e, idX where e is an idempotent endomorhism of X. In fact, the only
idempotents of E are d · ai and d · a′i . The coequalizer of d · a1 and idA is clearly
a1 (because a morphism f with f = f · d · a1 fulfils f 6= a2 and then it uniquely
factorizes through a1 ), analogously for the other three idempotents. Thus, the
above coequalizers demonstrate that E has filtered colimits.
7.10 Remark. For the category D of 7.9 we have

Sind D 6= Ind (Rec D) .

Observe first that

Rec D = E
is obtained from D by freely adding the coequalizers of c and c′ . The category
Ind E does not have a terminal object. In fact, the full subcategory of all objects
X in Ind E having a morphism from at most one of the objects C or C ′ into X
is clearly closed under filtered colimits – thus every object X has that property.
In contrast, Sind D has the terminal object colim D.
7.11 Example. Here we give an example of a functor not preserving sifted
colimits although it preserves (i) filtered colimits and (ii) reflexive coequalizers.
Let E T be the above category E with a terminal object T added. The sifted
diagram D : D → E T which is the inclusion has colimit

T = colim D

in E T . Let A be the category E T with a new terminal objects S added. The

functor

F : E T → A with F T = S and F X = X for all X 6= T

which is the identity map on morphisms of E does not preserve sifted colimits
because
F (colim D) = S and colim F · D = T .
However, F clearly preserves filtered colimits and reflexive coequalizers: the only
nontrivial colimits of these types in E T lie in E and are described in Example
7.9. The same description applies to A.

Historical Remarks for Chapter 7

November 20, 2009 63

 CHAPTER 7. FROM FILTERED TO SIFTED

The open problem of [8] whether preservation of filtered colimits and reflexive
coequalizers implies preservation of sifted colimits was answered by A. Joyal;
his proof even works for quasicategories (see [57]). Another proof was given S.
Lack (see [60]). The present proof is taken from [11] where also the stronger
statements stated at the beginning of this chapter are proved.
The formula 7.4 stems from [8].

November 20, 2009 64

Chapter 8

Canonical theories

Every algebraic category has a number of algebraic theories which are often
non-equivalent, we will study this more detailed in Chapter 15. In the present
chapter we prove that there is always an essentially unique algebraic theory with
split idempotents. We call it a canonical theory. We first discuss splitting of
idempotents and idempotent completions.

8.1 Definition.

1. Given an idempotent morphism

f: X → X , f ·f = f

in a category C, by a splitting of f is meant a factorization f = m · e such

that e · m is the identity morphism:

f idZ
X@ /X Z@ /Z
@@ ~> @@ ~>
@@ ~~~ @@ ~~~
e @@ ~m m @@ ~e
~~ ~~
Z X

2. A category C is called idempotent-complete provided that every idempotent

in C has a splitting.

8.2 Remark.

1. A splitting of an idempotent f is unique up to isomorphism:

(a) for every isomorphism i : Z → Z̄ the morphisms ē = i · e and m̄ =

m · i−1 form a splitting of f, and
(b) for every splitting f = m̄ · ē, ē · m̄ = id, there exists a unique isomor-
phism i such that i · e = ē and m · i−1 = m̄ (just put i = ē · m and
i−1 = e · m̄).

65
 CHAPTER 8. CANONICAL THEORIES

2. To be idempotent-complete is a self-dual notion: C is idempotent complete

iff C op is so.

8.3 Example.

1. Every category which has equalizers is idempotent-complete. In fact, form

an equalizer m of the idempotent f : X → X and id : X → X

m f
Z /X // X
id

Since f · f = id ·f, the morphism f factorizes as f = m · e for some

e : X → Z. Now, m · e · m = f · m = m and m is a monomorphism, so that
e · m = id .
Conversely, if an idempotent f : X → X splits as f = m · e, then m is an
equalizer of f and idX .

2. Every category with coequalizers is also idempotent-complete. Conversely,

if an idempotent f : X → X splits as f = m · e, then e is a coequalizer of
f and idX . This is the dualization of 1.

3. A full subcategory of an idempotent-complete category C is idempotent-

complete iff it is closed in C under retracts.

8.4 Definition. By an idempotent completion of a category C is meant a functor

EIc : C → Ic C

such that

1. Ic C is idempotent-complete

and

2. for every functor F : C → B, where B is an idempotent-complete category,

there exists an essentially unique functor F ∗ : Ic C → B with F naturally
isomorphic to F ∗ · EIc .

8.5 Remark.

1. A category C is idempotent-complete iff the functor EIc : C → Ic C is an

equivalence.

2. Clearly, Ic (C op ) ≃ (Ic C)op .

We give now an elementary description of Ic C.

8.6 Definition. For every category C we denote by Ic C the category of idem-

potents: its objects are the idempotent morphisms of C. Its morphisms from

November 20, 2009 66

CHAPTER 8. CANONICAL THEORIES

f : X → X to g : Z → Z are the morphisms a : X → Z in C such that a = g · a · f

or, equivalently, such that the diagram

a/Z
X@
@@
a
@@
f @@ g
 
X a /Z

commutes. The identity of the object f : X → X is f itself, and composition is

as in C.
We have a full and faithful functor EIc : C → Ic C defined by EIc (X) = idX
and EIc (a) = a.

8.7 Proposition. The functor EIc : C → Ic C defined in 8.6 is an idempotent

completion of C.

Proof. 1. Ic C is idempotent-complete: let a be an idempotent endomorphism

of (f : X → X) in Ic C. A splitting of a is given by

a / (a : X → X) a / (f : X → X)
(f : X → X)

2. Consider a functor F : C → B with B idempotent-complete. Every object

(f : X → X) of Ic C is obtained by splitting the idempotent f : EIc (X) →
EIc (X). Thus, we are forced to define F ∗ on objects f : X → X as the (essen-
tially unique) splitting of F f : F X → F X in B.
To define F ∗ on morphisms a : (f : X → X) → (g : Z → Z) in Ic C. observe
that a is the unique arrow making the following diagram

EIc (f )
EIc (X) / EIc (X)
GG ;
GG f f ww
GG www
GG ww
G# ww
(f )
EIc (a) a EIc (a)

;w (g) GG
g www GG g
GG
www GG
 ww G# 
EIc (Z) / EIc (Z)
EIc (g)

commutative in Ic C. Thus, we are forced to define F ∗ (a) as the unique arrow

November 20, 2009 67

 CHAPTER 8. CANONICAL THEORIES

making the following diagram

Ff
F XG / FX
GG ∗ w;
GGF f F ∗ f www
GG w
GG ww
# ww
F ∗ (f )
Fa F ∗a Fa

F ∗ (g)
F ∗ g ww
w; GG ∗
GGF g
ww GG
w w GG
 ww G# 
FZ / FZ
Fg

commutative in B. Explicitly:
F ∗f Fa F ∗g
F ∗ a : F ∗ (f ) / FX / FZ / F ∗ (g) .

It is easy to verify that this yields a well-defined functor F ∗ with F ∗ · I ≃ F.

8.8 Proposition. For every small category C an idempotent completion of C op
is the codomain restriction of the Yoneda embedding YC : C op → Set to the full
subcategory of all absolutely presentable objects of Set C .
Proof. Recall from 5.10 that a functor C → Set is absolutely presentable iff
it is a retract of a representable one. To prove that the full subcategory of all
absolutely presentable objects of Set C is equivalent to the category Ic (C op ) of
8.6, consider two retracts
e f
o / YC (X) , e · m = idR and o / YC (Z) , f · n = idS
R m
S n

and a morphism g : R → S in Set C . By the Yoneda Lemma, we get two idem-

potents ê = m · e(idX ) and fˆ = n · f (idZ ) in C and a morphism ĝ = n · g · e(idX )
forming a commutative diagram
ĝ
Z NNN /X
NNN ĝ
fˆ
NNN
NNN ê
 N& 
Z /X
ĝ

This is a morphism in the category Ic (C op ). The rest of the proof is straightfor-

ward.
8.9 Corollary. For every algebraic theory T the Yoneda embedding
YT : T op → (Alg T )pp
is an idempotent completion of T op . In other words, (Alg T )pp ≃ Ic (T op ).

November 20, 2009 68

CHAPTER 8. CANONICAL THEORIES

In fact, by 5.10 and 8.8, Ic (T op ) is formed by retracts of representables in Set T ,

that is, by perfectly presentable objects in Alg T (see 5.14).
8.10 Corollary. For two small categories C and D the corresponding functor
categories Set C and Set D are equivalent if and only if C and D have a common
idempotent completion:

Set C ≃ Set D iff Ic C ≃ Ic D .

Proof. The universal property of EIc : C → Ic C clearly implies Set C ≃

Ic C
Set .
Conversely, if Set C ≃ Set D , then the subcategories of absolutely presentable
objects are equivalent. By 5.10 and 8.8, this means that Ic (C op ) ≃ Ic (Dop ).
Now use duality, or 8.5.2.
8.11 Definition. Let A be an algebraic category. An algebraic theory for A is
called canonical if it is idempotent-complete.
8.12 Proposition. Every algebraic category A has a canonical theory unique
up to equivalence. The dual of App (see 6.12) is a canonical theory for A.
Proof. Following 6.13, A ≃ Alg (Aop pp ). By 5.14 and 8.3, App is idempotent-
complete, and then Aop pp is also idempotent-complete, see 8.2. Let us verify the
uniqueness. If A ≃ Alg T for some algebraic theory T , then App ≃ (Alg T )pp ≃
Ic (T op ) by 6.14 and 8.9. Finally, Aop
pp ≃ Ic T by 8.5.

8.13 Example.
1. The canonical theory of the category Set is the theory N of natural num-
bers, see 1.9: it is clear that N is idempotent-complete.
2. The canonical theory of the category Ab is the theory Tab described in
1.13. In fact, we saw in 5.6 that Tab is dual to Abpp .
3. In the category Bool of boolean algebras we have the algebras PX of all
subsets of a set X. The free algebras on n generators PPn form, for n ∈ N,
a strong generator. As noted in 6.10, the dual of this full subcategory of
Bool is a theory for Bool. However, this is not the canonical theory. In
fact, the canonical theory is the dual of the full subcategory of all algebras
Pn for n ∈ N \ {0}, or equivalently, the category of finite nonempty sets
and functions. Since each n > 0 is injective in the category of finite sets,
it is a retract of 2n . Thus Pn is a retract of PPn.

Historical Remarks for Chapter 8

The idempotent completion can already be found in B. Mitchell’s monograph

[75]. M. Bunge [30] presented it as well and used it for proving Corollary 8.10.
She called it idempotent splitting closure; other names were used as well, e.g.
Cauchy completion or Karoubi envelope. Corollary 8.10 was later proved in
[45]. Proposition 8.12 is due to J. J. Dukarm [40].

November 20, 2009 69

 CHAPTER 8. CANONICAL THEORIES

November 20, 2009 70

Chapter 9

Algebraic functors

We have studied algebraic categories as individual categories so far. It turns

out that there is a natural concept of morphism between algebraic categories,
which we call an algebraic functor, so that we obtain a 2-category of all algebraic
categories. We then prove a duality result: this 2-category is biequivalent to
the 2-category of algebraic theories. We first need to introduce a concept of
morphism between algebraic categories – this is quite obvious:
9.1 Definition. Let T1 and T2 be algebraic theories. A functor M : T1 → T2 is
called a morphism of algebraic theories if it preserves finite products.
9.2 Notation. For a morphism of theories M : T1 → T2 we denote by
Alg M : Alg T2 → Alg T1
the functor defined on objects A : T2 → Set by A 7→ A · M.
9.3 Proposition. Let M : T1 → T2 be a morphism of algebraic theories.
1. Alg M : Alg T2 → Alg T1 preserves limits and sifted colimits.
2. Alg M has a left adjoint M ∗ : Alg T1 → Alg T2 which is the essentially
unique functor which preserves sifted colimits and makes the square
YT1
T1op / Alg T1

M op M∗
 
T2op / Alg T2
YT2

commutative up to natural isomorphism.

Proof. The essentially unique functor M ∗ follows from Theorem 4.13. Since
YT2 · M op preserves finite coproducts, see Lemma 1.5, to get the adjunction
M ∗ ⊣ Alg M it suffices to apply 4.15: the right adjoint Alg M is given by
B 7→ Alg T2 (YT2 (M −), B).

71
 CHAPTER 9. ALGEBRAIC FUNCTORS

By the Yoneda Lemma, this is nothing but composition with M. This immedi-
ately implies that Alg M preserves sifted colimits, because they are calculated
objectwise in Alg T1 and Alg T2 (see 2.5).
9.4 Definition. A functor between two algebraic categories is called algebraic
provided that it preserves limits and sifted colimits.
9.5 Example.
1. Every functor Alg M, for a theory morphism M, is algebraic.
2. The forgetful functor Ab → Set is algebraic.
3. Given an algebra A in an algebraic category A, then A(A, −) : A → Set is
algebraic iff A is perfectly presentable.
4. A constant functor with value A between algebraic categories is algebraic
iff A is a terminal object.
5. For every algebraic theory T , the embedding I : Alg T → Set T is an alge-
braic functor, see 1.16 and 2.5.
9.6 Remark. We know from 9.3 that every morphism of theories induces an
algebraic functor between the corresponding algebraic categories. If, moreover,
the algebraic theories are canonical (8.11), then the algebraic functors are es-
sentially just those induced by morphisms of theories, see 9.15 below. This will
motivate us to define “morphisms of algebraic categories” as the algebraic func-
tors. We are now going to prove that every algebraic functor has a left adjoint.
For this we will use Freyd’s Adjoint Functor Theorem (see 0.8).
9.7 Theorem. A functor between algebraic categories is algebraic if and only
if it has a left adjoint and preserves sifted colimits.
Proof. Let G : B → A be an algebraic functor. We are to prove that G has
a left adjoint. That is, for every object A of A we are to prove that the functor
A(A, G−) : B → Set
is representable.
(1) Assume first that A is perfectly presentable. Since G preserves limits, it
remains to prove that A(A, G−) satisfies the Solution Set Condition of 0.10.
For Bpp in 6.12 put
G = {GX | X ∈ Bpp } .
Every object B of B is a sifted colimit of objects from Bpp (see 6.9). Let us
write (σX : X → B) for the colimit cocone. Since G preserves sifted colim-
its, (GσX : GX → GB) is also a colimit cocone. As A(A, −) preserves sifted
colimits, every morphism b : A → GB factorizes as follows
GX
=
ϕ {{{
{ GσX
{{
{{ 
A / GB
b

November 20, 2009 72

CHAPTER 9. ALGEBRAIC FUNCTORS

Thus all G(σX ) form a solution set.

(2) If A is an arbitrary object of A, we express it as a sifted colimit of per-
fectly presentable objects (see 6.9), say A = colim Ai . From (1) we know that
A(Ai , G−) is representable, say A(Ai , G−) ≃ B(Bi , −). Yoneda Lemma now
allows us to define an (obvious) sifted diagram whose objects are the Bi . There-
fore, A(Ai , G−) is representable by colim Bi since
A(A, G−) = A(colim Ai , G−) ≃ lim A(Ai , G−) ≃ lim B(Bi , −) ≃ B(colim Bi , −).

9.8 Remark. The previous theorem can be refined, as we demonstrate in Chap-

ter 18: a functor between algebraic categories is algebraic if and only if it
preserves limits, filtered colimits and regular epimorphisms. One implication
follows from the fact that: a) filtered implies sifted, and b) every regular epi-
morphism is a reflexive coequalizer (of its kernel pair). The converse implication
is a particular case of 18.2.
9.9 Remark. We are going to prove a duality between algebraic categories
and canonical algebraic theories. This does not really mean a contravariant
equivalence of categories. Indeed, a more subtle formulation is needed: just
look at the simplest algebraic category, Set, and the simplest endomorphism,
the identity functor IdSet . It is easy to find a proper class of functors naturally
isomorphic to IdSet – and each of them is algebraic. However, in the category of
all theories no such phenomenon occurs. We thus need to work with morphisms
of algebraic categories “up to natural isomorphism”. For this reason we have to
move from categories to 2-categories. The reader does not need to know much
about 2-categories. Here we summerize the needed facts.
9.10 A primer on 2-categories.
1. Let us recall that a 2-category A has a class obj A of objects and, instead
of hom-sets A(A, B), it has hom-categories A(A, B) for every pairs A, B
of objects. The objects of A(A, B) are called 1-cells and the morphisms
2-cells. A prototype of a 2-category is the 2-category Cat of all small cat-
egories (as objects). Given small categories A, B then 1-cells are functors
from A to B and 2-cells are natural transformations. In our book, we
essentially work just with this 2-category and its full sub-2-categories.
2. Let us recall the concept of a 2-functor F : A → B between 2-categories
A and B : it assigns objects F A of B to objects A of A; for every pair
A, A′ of objects of A, it defines a functor FA,A′ : A(A, A′ ) → B(F A, F A′ )
and fulfils some canonical requirements about compositions and identities.
Example: a full sub-2-category A of a 2-category B is given by a choice of
a class of objects obj A ⊆ obj B. The 1-cells and 2-cells in A are precisely
those of B. The inclusion I : A → B is a 2-functor.
3. Two objects A, B of a 2-category A are called equivalent if there exist
1-cells f : A → B and f : B → A such that f · f is isomorphic to idA in
A(A, A) and f · f is isomorphic to idB in A(B, B).

November 20, 2009 73

 CHAPTER 9. ALGEBRAIC FUNCTORS

4. A 2-functor F : A → B is called a biequivalence if all the functors FA,A′ are

equivalence functors, and every object of B is equivalent to F A for some
object A of A.
5. For every 2-category A we denote by Aop the dual 2-category: it has the
same objects and the direction of 1-cells is reversed (while the direction
of 2-cells remains non-reversed): Aop (A, A′ ) = A(A′ , A).
9.11 Example.
1. We define the 2-category Th of theories to have
objects: algebraic theories,
1-cells: morphisms of algebraic theories (i.e., functors preserving fi-
nite products),
2-cells: natural transformations.
This is a full sub-2-category of Cat, i.e., composition of 1-cells and 2-
cells are defined in Th as the usual composition of functors and natural
transformations, respectively.
2. The 2-category Th c of canonical theories is the full sub-2-category of Th
on all theories that are canonical, that is, idempotent-complete.
3. We define the 2-category ALG of algebraic categories to have
objects: algebraic categories,
1-cells: algebraic functors,
2-cells: natural transformations.
Once again, this is a full sub-2-category of Cat : composition is the usual
composition of functors or natural transformations.
9.12 Remark. We need to be a little careful about foundations here: there
is, as remarked above, a proper class of 1-cells in ALG (Set, Set), for example.
However, if we consider the 1-cells up to natural isomorphism, all problems
disappear: this is one consequence of the duality theorem below. Ignoring the
foundational considerations, we consider ALG as a sub-2-category of the 2-
category of all categories. (The duality we prove below tells us that ALG is
essentially just the dual of Th c .)
9.13 Definition. We denote by

Alg : Th op → ALG

the 2-functor assigning to every algebraic theory T the category Alg T , to every
1-cell M : T1 → T2 the functor Alg M = (−) · M and to every 2-cell α : M → N
the natural transformation Alg α : Alg M → Alg N whose component at a T2 -
algebra A is A · α : A · M → A · N.

November 20, 2009 74

CHAPTER 9. ALGEBRAIC FUNCTORS

9.14 Remark. The 2-functor Alg is well-defined due to 9.3: for every morphism
of theories M, the functor Alg M is algebraic. The fact that for every natural
transformation α we get a natural transformation Alg α is easy to verify.

9.15 Theorem. (Duality of algebraic categories and theories) The 2-category

ALG of algebraic categories is biequivalent to the dual of the 2-category Th c of
canonical algebraic theories.
In fact, the domain restriction of the 2-functor Alg to canonical algebraic theo-
ries
Alg : Th op
c → ALG

is a biequivalence.

Proof. (1) Following 8.12, every algebraic category A is equivalent to Alg T

for the canonical algebraic theory T = Aoppp .
(2) We will prove that for two canonical algebraic theories T1 and T2 the functor

Alg T1 ,T2 : Th c (T1 , T2 ) → ALG (Alg T2 , Alg T1 )

is an equivalence of categories.
(2a) Alg T1 ,T2 is full and faithful: given morphisms M, N : T1 ⇉ T2 and a nat-
ural transformation λ : Alg M → Alg N there exists a unique natural trans-
formation α : M → N such that Alg α = λ. The proof follows the lines of
the proof of Yoneda Lemma. Let us just indicate how to construct α. Con-
sider an object X in T1 . Since T2 (M X, −) ∈ Alg T2 , we have the component
λT2 (MX,−) (X) : T2 (M X, M X) → T2 (M X, N X) and we put

αX = λT2 (MX,−) (X)(idMX ) : M X → N X.

The family (αX )X∈T1 is the required natural transformation α : M → N.

(2b) The functor Alg T1 ,T2 is essentially surjective. In fact, consider an algebraic
functor G : Alg T2 → Alg T1 and let L be its left adjoint, see 9.7. We are going
to prove that L restricts to a functor F which preserves finite coproducts and
makes the square
YT1
T1op / Alg T1

F L
 
T2op / Alg T2
YT2

commutative up to natural isomorphism. For every object X in T1 we have by

adjunction a natural isomorphism

Alg T2 (L(YT1 (X)), −) ≃ Alg T1 (YT1 (X), G−) .

Since YT1 (X) is perfectly presentable (see 5.5) and G preserves sifted colim-
its, the above natural isomorphism implies that L(YT1 (X)) is perfectly pre-
sentable. By 5.14, L(YT1 (X)) is a retract of a representable algebra and,

November 20, 2009 75

 CHAPTER 9. ALGEBRAIC FUNCTORS

since T2 is idempotent-complete, L(YT1 (X)) is itself a representable algebra

(see 8.3.3). This means that there exists an essentially unique object in T2 ,
say F X, such that L(YT1 (X)) ≃ YT2 (F X). In this way we get a map on ob-
jects, F : obj T1 → obj T2 , which, by the Yoneda Lemma, extends to a func-
tor F : T1op → T2op making the above square commutative up to natural iso-
morphism. F preserves finite coproducts because YT1 preserves them by 1.5,
YT2 reflects finite coproducts, and L preserves them. It remains to prove that
G ≃ Alg M, where M = F op : T1 → T2 . Or, equivalently, that L ≃ M ∗ in
the notation of 9.3. But this follows from the essential commutativity of the
above square and the last part of 9.3. This proves that Alg T1 ,T2 is essentially
surjective.
9.16 Corollary. A functor between algebraic categories

G : A2 → A1

is algebraic if and only if it is induced by a morphism of theories. That is: there

exists a morphism of the corresponding canonical algebraic theories M : T1 → T2
and two equivalence functors E1 : Alg T1 → A1 and E2 : Alg T2 → A2 such that
the square
Alg M
Alg T2 / Alg T1

E2 E1
 
A2 / A1
G

commutes up to natural isomorphism.

In fact, given G, the proof of 9.15 yields G ≃ Alg M with the desired property.
The converse implication is clear.
9.17 Remark. The category Set is a kind of dualizing object for the biequiva-
lence Alg : Th op
c → ALG :
(1) Forgetting size considerations (an algebraic theory is by definition a small
category) we have Alg T = Th (T , Set) for every algebraic theory T .
(2) For every algebraic category A there is an equivalence of categories

Aop
pp ≃ ALG (A, Set) .

Indeed, if A ∈ Aoppp , then G = A(A, −) : A → Set preserves limits and sifted

colimits. Conversely, let G : A → Set be an algebraic functor and let L be a left
adjoint of G (see 9.7). Then G ≃ A(L1, −) (with 1 denoting a one-element set)
and L1 is perfectly presentable because G preserves sifted colimits.
9.18 Remark. We conclude this chapter by mentioning the analogous Gabriel-
Ulmer duality for locally finitely presentable categories. The proof is similar to
that of 9.15. Whereas the morphisms of algebraic categories are the functors pre-
serving limits and sifted colimits, the morphisms of locally finitely presentable
categories are the functors preserving limits and filtered colimits. These are the

November 20, 2009 76

CHAPTER 9. ALGEBRAIC FUNCTORS

1-cells of the 2-category LFP , and the 2-cells are natural transformations. We
also denote by LEX the 2-category of small categories with finite limits, finite
limit preserving functors, and natural transformations. The 2-functor

Lex : LEX op → LFP

assigns to every small category T with finite limits the category Lex T (4.16)
and it acts on 1-cells and 2-cells in the analogous way as Alg : Th op → ALG .

9.19 Theorem. The 2-categories LFP and LEX are dually biequivalent.
In fact, Lex : LEX op → LFP is a biequivalence. The converse construction
associates to a locally finitely presentable category A the small, finitely complete
category Aopf p (6.26).

Historical Remarks for Chapter 9

Morphisms of algebraic theories and the resulting algebraic functors were (in
the one-sorted case) introduced by F. W. Lawvere [63] and belong to the main
contribution of his work. The characterization 9.7 of algebraic functors and
the duality theorem 9.15 are contained in [5]. This is analogous to the Gabriel-
Ulmer duality 9.19 for locally finitely presentable categories (see [48]). A general
result can be found in [35].
2-categories have been introduced by C. Ehresmann in [42].

November 20, 2009 77

 CHAPTER 9. ALGEBRAIC FUNCTORS

November 20, 2009 78

Chapter 10

Birkhoff ’s Variety Theorem

So far we have not treated one of the central concepts of algebra: equations. In
the present chapter we prove the famous characterization of varieties of algebras,
due to G. Birkhoff: varieties are precisely the full subcategories of Alg T closed
under
products,
subalgebras,
regular quotients,
and
directed unions.
The last item was not included in Birkhoff’s formulation. The reason is that
Birkhoff only considered one-sorted algebras, and for them directed unions follow
from the other three items (see 11.35). For general algebraic categories directed
unions cannot be omitted, see Example 10.23 below.
Classically, an equation is an expression u = v where u and v are terms (say,
in n variables). We will see in 13.9 that such terms are morphisms from n to 1
in the theory of Σ-algebras. We can also consider k-tuples of classical equations
as pairs of morphisms from n to k. The following concept generalizes this idea.
10.1 Definition. If T is an algebraic theory, an equation in T is a parallel
pair u, v : s ⇉ t of morphisms in T . (Following algebraic tradition, we write
u = v in place of (u, v).) An algebra A : T → Set satisfies the equation u = v if
A(u) = A(v).
10.2 Example.
1. In the theory Tab of abelian groups (1.11) recall that endomorphisms of
1 have the form [n] and correspond to the operations on abelian groups
given by x 7→ n · x. Thus, the equation
[2] = [0]

79
 CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

is satisfied by precisely the groups with

x+x = 0

for all elements x.

2. Graphs whose only edges are loops are given, considering the theory in
1.15, by the equation
τ =σ.

10.3 Remark. Observe that if an algebra A satisfies the equation u = v, then

it also satisfies all the equations of the form u · x = v · x and y · u = y · v for
x : s′ → t and y : t → t′ in T . Moreover, if the equations ui = vi , i = 1, . . . , n,
are satisfied by A for ui , vi : s → ti , then A also satisfies hui i = hvi i, where

hui i, hvi i : s → t1 × . . . × tn

are the correspnding morphisms. For this reason, we will state the definition of
variety using congruences as well as equations.
10.4 Definition. Let T be an algebraic theory. A congruence on T is a col-
lection ∼ of equivalence relations ∼s,t on hom-sets T (s, t), where (s, t) ranges
over pairs of objecs, which is stable under composition and finite products in
the following sense:
1. If u ∼s,t v and x ∼r,s y, then u · x ∼r,t v · y
x // s u //
r t
y v

2. If ui ∼s,ti vi for i = 1, . . . , n, then hu1 , . . . , un i ∼s,t hv1 , . . . , vn i, where

t = t1 × . . . × tn
ui
s /;/ t i
vi w
ww
ww
hv1 ,...,vn i hu1 ,...,un i www
w
ww
www
w
ww
  ww
t1 × . . . × tn

10.5 Example. Consider the theory Tmon of monoids (3.10) whose morphisms
from n to k are all k-tuples of words in n = {0, 1, . . . , n − 1}. The commutative
law corresponds to the congruence ∼ with ∼n,1 defined for words v and w by

w ∼n,1 v iff v can be obtained from w by a permutation of its letters.

The general equivalence ∼n,k is defined by coordinates of the k-tuples of words.

November 20, 2009 80

CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

10.6 Example. Let T be an algebraic theory and let M : T → B be a finite

product preserving functor. The kernel congruence of M is denoted by ≈M .
That is, for all u, v ∈ T (s, t) we put

u ≈M v iff M u = M v

This is obviously a congruence on T . In particular, for every T -algebra A we

have a congruence ≈A .
10.7 Remark.
1. Congruences on a given algebraic theory T are ordered in a canonical way:
we write ∼ ⊆ ∼′ in case that for every pair s, t of objects of T we have:

u ∼s,t v implies u ∼′s,t v

2. It is easy to see that every (set theoretical) intersection of congruences is

a congruence. Consequently, for every set E of equations there exists the
smallest congruence ∼E on T containing E. We say that the congruence
∼E is generated by the equations of E.
10.8 Definition. A full subcategory A of Alg T is called a variety if there exists
a set of equations such that a T -algebra lies in A iff it satisfies all equations in
that set.
10.9 Remark.
1. Varieties are also sometimes called equational classes of algebras, or equa-
tional categories. But we reserve this name for the special case of varieties
of Σ-algebras treated in Chapters 13 (for one-sorted signatures) and 14
(for S-sorted ones).
2. We will use the name variety also in the loser sense of a category equivalent
to a full subcategory of Alg T specified by equations. Every time we use
the word variety it will be clear whether the above definition is meant or
the loser version.
10.10 Example.
1. The abelian groups satisfying x + x = 0 form a variety in Ab.
2. All graphs whose edges are just loops form a variety in Graph.
3. Let us consider the category Set × Set of two-sorted sets with sorts called,
say, s and t. This has an algebraic theory TC of all words in {s, t}, see 1.10.
Consider the full subcategory A of Set × Set of all pairs A = (As , At ) with
either As = ∅ or At has at most one element. This can be specified by the
equation given by the parallel pair of projections
//
stt t

November 20, 2009 81

 CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

10.11 Remark. Every variety is also specified by a congruence. More precisely,

given a variety A of Alg T , let ∼ be the congruence generated by the given set
E of equations. A T -algebra A lies in A iff it satisfies all equations in ∼, that
is, iff it fulfils:
u ∼ v implies Au = Av .
In fact, if A satisfies all equations in E, then E ⊆ ≈A and therefore ∼ ⊆ ≈A .
10.12 Notation. For every congruence ∼ on an algebraic theory T we denote
by
T/∼
the algebraic theory on the same objects and with morphisms given by the
congruence classes of morphisms of T :
(T / ∼)(s, t) = T (s, t)/ ∼s,t .
Composition and identity morphisms are inherited from T ; more precisely, they
are determined by the fact that we have a functor
Q: T → T / ∼
which is the identity map on objects and which assigns to every morphism its
congruence class.
10.13 Remark.
1. It is easy to verify that T / ∼ has finite products determined by those of T ,
thus, T / ∼ is an algebraic theory and Q is a theory morphism. Moreover,
the functor Q is full and surjective on objects.
2. A morphism of theories M : T → T ′ factorizes through Q up to natural
isomorphism
Q
T @ /T/∼
@@ yy
@@ yy
@ ≃ y
M @ y ′
@ |yy M
T′
iff the congruence ∼ is contained in the kernel congruence ≈M . When this
is the case, the factorization M ′ is essentially unique and it is a theory
morphism. In fact, it is clear that if M ′ · Q ≃ M, then ∼ ⊆ ≈M . For the
converse define M ′ to be equal to M on objects, and put M ′ [f ] = M f on
morphisms. Clearly, M ′ · Q = M. From the fact that M preserves finite
products and Q reflects them the desired properties of M ′ easily follow.
3. If in 2. we take ∼ equal to the congruence ≈M , then the factorization M ′
Q
T @ / T / ≈M
@@ w
@@ ww
@ www ′
M @@ w
 w{ w M
T′

November 20, 2009 82

CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

is faithful. Therefore, M is full and essentially surjective iff M ′ is an

equivalence of categories.

10.14 Proposition. Let ∼ be a congruence on an algebraic theory T and let

Q : T → T / ∼ be the corresponding quotient. The functor

Alg Q : Alg (T / ∼) → Alg T

is full and faithful.

Proof.
(1) Clearly, Alg Q is faithful because Q is surjective.
(2) Consider objects A, B ∈ Alg T / ∼ and a morphism β : A · Q → B · Q, that
is, a collection βt : At → Bt of homomorphisms natural in t ranging through T .
Then the same collection is natural in t ranging through T / ∼, thus, β : A → B
is a morphism of Alg (T / ∼). Clearly, Alg Q takes this morphism to the original
one.
10.15 Corollary. Every variety is an algebraic category.

More detailed: let ∼ be a congruence on an algebraic theory T , and let A be

the full subcategory of Alg T specified by ∼ . There exists an isomorphism of
categories
E : Alg (T / ∼) → A
such that the triangle

AO  RRRR
v
RRR
RRR
E RRR
RRR
)
Alg (T / ∼) / Alg T
Alg Q

commutes. In fact, for every (T / ∼)-algebra B, given u ∼ v, then B(Qu) =

B[u] = B[v] = B(Qv). This implies that Alg Q factorizes through the inclusion
of A in Alg T . Moreover, if A lies in A, then A = (Alg Q)(B) where B is
the (T / ∼ )-algebra defined by B[u] = Au. This shows that the factorization
E : Alg (T / ∼) → A is bijective on objects. Since Alg Q is full and faithful (see
10.14), E is an isomorphism.

10.16 Proposition. Every variety A of T -algebras is closed in Alg T under

Q
(a) products: given a product B = i∈I Ai in Alg T with all Ai in A, then B
also lies in A,

(b) subalgebras: given a monomorphism m : B → A in Alg T with A in A,

then B also lies in A,

(c) regular quotients: given a regular epimorphism e : A → B in Alg T with A

in A, then B also lies in A,

November 20, 2009 83

 CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

and
(d) sifted colimits: given a sifted colimit B = colim Ai in Alg T with all Ai in
A, then B also lies in A.
Proof. Following 10.15, the inclusion functor A → Alg T is naturally isomor-
phic to Alg Q, which preserves limits and sifted colimits by 9.3. This proves (a)
and (d).
(b) Let m : B → A be a monomorphism with A in A. We prove that B is in A
by verifying that every equation u1 , u2 : s ⇉ t that A satisfies is also satisfied by
B. We know from 1.18.2 that the component mt : Bt → At is a monomorphism.
From A(u1 ) = A(u2 ) and the commutativity of the squares
Bui
Bs / Bt
ms mt
 
As / At
Aui

we conclude Bu1 = Bu2 .

(c) Let e : A → B be a regular epimorphism with A in A. To prove that B lies
in A, observe that a kernel pair k1 , k2 : N (e) ⇉ A of e yields a subobject of
A × A ∈ A, thus N (e) lies in A. by (a) and (b). And since the pair k1 , k2 is
reflexive and e is its coequalizer, B is a sifted colimit (3.2) of a diagram in A,
thus B ∈ A.
10.17 Corollary. Every variety A of T -algebras is closed in Alg T under limits
and sifted colimits.
10.18 Example. Not every full subcategory of an algebraic category Alg T
closed under limits and sifted colimits is a variety. Consider the free completion
T ′ of T under finite products (1.6) and the finite product preserving extension
M : T ′ → T of the identity functor on T . The induced functor Alg M is naturally
isomorphic to the full inclusion Alg T → Alg T ′ ≃ Set T . But Alg T is usually not
closed in Set T under subalgebras (in contrast to 10.16). As a concrete example,
let T be the category of finite sets and functions. The functor

A : T → Set , AX = X × X × X

is clearly an algebra for T , but its subfunctor A′ given by all triples in X ×X ×X

in which at least two different coordinates have the same value is not an algebra
for T .
10.19 Remark. Following 10.13.3 and 10.14, the algebraic functor induced by
a full and essentially surjective morphism of theories is full and faithful. The
morphism M : T ′ → T in 10.18 also demonstrates that this fact cannot be
inverted: Alg M : Alg T → Set T is full and faithful, but M is not full.
10.20 Definition. A reflective subcategory A of a category B (see 0.9) is called
regular epireflective if all reflections rB : B → RB are regular epimorphisms.

November 20, 2009 84

CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

10.21 Corollary. Every variety of T -algebras is a regular epireflective subcat-

egory of Alg T closed under regular quotients and directed unions.

Proof. We already know that the variety A is closed in Alg T under regular
quotients and sifted colimits (10.16) and therefore under directed unions, which
are a special case of sifted colimits (2.9). Moreover, following 10.15, the inclusion
functor A → Alg T is naturally isomorphic to Alg Q, which has a left adjoint
(9.3). It remains to prove that for every T -algebra B the reflection rB : B → RB
is a regular epimorphism. Let rB = m · e

? A aCCC
e ~~~ CCCCCCm
~~ CCCC
~~ ē CCC!
B / RB
rB

be a regular factorization of rB , see 3.7. By 10.16(b) A ∈ A, thus there is a

unique ē : RB → A such that e = ē · rB . Since e is an epimorphism, we see that
ē · m = idA . Also m · ē = idRB due to the universal property of rB . Thus, m is
an isomorphism and rB a regular epimorphism.

10.22 Birkhoff ’s Variety Theorem. Let T be an algebraic theory. A full

subcategory A of Alg T is a variety if and only if it is closed in Alg T under

(a) products,

(b) subalgebras,

(c) regular quotients,

and

(d) directed unions.

Proof. Every variety is closed under (a)-(d): see 10.16. Conversely, let A be
closed under (a)-(d).
(1) We first prove that A is a regular epireflective subcategory. Let B be a T -
algebra. By 3.6 there exists a set of regular epimorphisms e : B → Ae (e ∈ X)
representing all regular quotients of B with codomain in A. Denote by b : B →
Q
e∈X Ae the induced morphism and let

b Q
BA / Ae
AA : e∈X
AA u uu
rB AA
uu
A uuuu mB
RB

be the regular factorization of b (see 3.7). The algebra RB lies in A (because

it is a subalgebra of a product of algebras in A) and rB is the reflection of B

November 20, 2009 85

 CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

in A. Indeed, for every morphism f : B → A with A in A we have a regular

factorization
f = m · e for some e ∈ X

and since e factorizes through b, so does f.

(2) We will prove that A is specified by the congruence ∼ which is the intersec-
tion of the kernel congruences of all algebras in A :

u1 ∼s,t u2 iff Au1 = Au2 for all A ∈ A .

This is indeed a congruence, see 10.6 and 10.7.2. It is our task to prove that
every T -algebra B such that ∼ is contained in ≈B , see 10.7, lies in A.
(2a) Assume first that B is a regular quotient of a representable algebra YT (t).
We thus have a regular epimorphism e : YT (t) → B. We know that the reflec-
tion morphism rt : YT (t) → R(YT (t)) is a regular epimorphism, thus, it is a
coequalizer
u / rt
N v / YT (t) / R(YT (t))

By 4.2 we can express N as a sifted colimit of representable algebras, denote the

colimit cocone by cs : YT (s) → N . Using Yoneda Lemma, we see that for every
s there exist morphisms us , vs : t ⇉ s representing u · cs and v · cs , respectively:

YT (s) PP m6 B
PPPPPPYT (us )
PPPPPP mmmmm O
e m
PPPPPP mmm f
YT (vs ) PPPP' mmmmm
' m rt
/ R(YT (t))
nn6 6 YT (t) RRRR
cs
nn
u nnnnnn RRR
nnnnnnnnvnn h
RRR
RRR
 nnnnnnnn RRR 
(
N A

Moreover, we have us ∼t,s vs because every morphism h : YT (t) → A with

A ∈ A factorizes through rt , so that YT (us ) · h = YT (vs ) · h, and it follows
that Aus = Avs . By assumption on B, this implies Bus = Bvs , and then
e · u · cs = e · v · cs . The cocone cs is jointly epimorphic, thus we have e · u = e · v.
Since rt is the coequalizer of u and v, there exists f : R(YT (t)) → B such that
f · rt = e. Finally, f is a regular epimorphism because e is, so that B, being a
regular quotient of R(YT (t)), lies in A.
(2b) Let B be arbitrary. Express B as a sifted colimit of representable algebras
(4.2) and for each of the colimit morphisms σs : YT (s) → B denote by Bs the
image of σs which, by 3.7, is a subalgebra of B. Since T has finite products, the
collection of these subalgebras of B is directed. Due to assumption (d) we only
need to prove that every Bs lies in A. This follows from (2a): we know that Bs
is a regular quotient of a representable algebra, and Bs has the desired property:
given u1 ∼s,t u2 , we have B(u1 ) = B(u2 ) and this implies Bs u1 = Bs u2 since
Bs is a subalgebra of B.

November 20, 2009 86

CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

10.23 Example. The assumption of closure under directed unions cannot be

omitted: consider the category Set N of N-sorted sets and let A be the full
subcategory of all A = (An )n∈N such that either some An is empty or all An ’s
have precisely one element. This subcategory is clearly closed under products,
subalgebras and regular quotients – we omit the easy verification. However,
it is not a variety, not being closed under directed unions. In fact, every N-
sorted set A = (An )n∈N is a directed union of objects Ak of A : put Ak =
(A0 , . . . , Ak , ∅, ∅, . . .).
10.24 Corollary. Let A be a full subcategory of Alg T . Then A is a variety if
and only if it is a regular epireflective subcategory closed under regular quotients
and directed unions.
In fact, following 10.21 and 10.22, we only need to observe that every regu-
lar epireflective subcategory A is closed in Alg T under products (and this is
obvious) and subalgebras: consider the diagram
m /A
B
rB rA
 
RB / RA
Rm

where m is a monomorphism and A lies in A. Then rA is an isomorphism and

rB is a monomorphism. But rB is also a regular epimorphism, so that it is an
isomorphism, and B lies in A.
10.25 Example. The category Abtf of torsion free abelian groups is a regular
epireflective subcategory of Ab closed under filtered colimits. But this is not a
variety in Ab because it is not closed under quotients. Indeed, Abtf is locally
finitely presentable but not algebraic (it is not exact).

Historical Remarks for Chapter 10

The classical characterization of varieties of one-sorted algebras as HSP-classes

is due to G. Birkhoff [24]; we present this version in 11.35.
For many-sorted algebras the concept of equation in 10.1 corresponds to the
formulas
∀ x1 . . . ∀ xn (t = s)
where t and s are terms of the same sort in (sorted) variables x1 , . . . , xn . Example
10.23 demonstrates that with respect to the above equations we need to add the
closure under directed unions. Another approach, used e.g. in [7], is to admit
infinitely many variables in the equations – that is, to work in the logic Lω∞
(admitting quantification over infinite sets) rather than in the finitary logic Lωω .
As proved in [7], Theorem 3.9, in this logic directed unions can be omitted in
Birkhoff’s Variety Theorem. This is illustrated by Example 14.21 below.

November 20, 2009 87

 CHAPTER 10. BIRKHOFF’S VARIETY THEOREM

November 20, 2009 88

Chapter 11

One-sorted algebraic
categories

Classical algebraic categories, such as groups, modules, boolean algebras, etc.

are not only abstract categories: their objects are sets with a structure, and
their morphisms are functions preserving the structure. Thus, they are concrete
categories over Set (see 0.18), which means that a “forgetful” functor into Set
is given. Also, these classical algebraic categories have theories generated by a
single object X in the sense that all objects of the theory are finite powers X n
of X. (Consider the free group or the free module on one generator.) In order
to formalize this idea, we study in this chapter one-sorted algebraic categories.
In Chapter 14 we will deal with the more general notion of S-sorted algebraic
categories, for which the forgetful functor into a power of Set is considered rather
than into Set.
11.1 Example. The theory
N
of sets (see 1.9), which is the full subcategory of Setop on natural numbers
n = {0, . . . , n − 1}, is a “prototype” one-sorted theory: every object n is the
product of n copies of 1. Moreover, the n injections in Set

πin : 1 → n , 0 7→ i (i = 0, . . . , n − 1)

yield a canonical choice of projections πin : n → 1 in N which present n as 1n .

11.2 Remark.
1. If an algebraic theory T has all objects finite powers of an object X, we
obtain a theory morphism T : N → T as follows: for every n choose an
n-th power of X with projections pni : X n → X for i = 0, . . . , n − 1. Then
T is uniquely determined by

T n = X n and T πin = pni for all i < n in N .

89
 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

2. Conversely, every theory morphism T : N → T represents an object and

a choice of projections of all of its finite powers: put X = T 1 and pni =
T πin : X n → X for all i < n.
This leads us to the following
11.3 Definition.
1. A one-sorted algebraic theory is a pair (T , T ) where T is an algebraic
theory whose objects are the natural numbers, and T : N → T is a theory
morphism which is the identity map on objects.
2. A morphism of one-sorted algebraic theories M : (T1 , T1 ) → (T2 , T2 ) is a
functor M : T1 → T2 such that M · T1 = T2 .
11.4 Remark.
1. We have not requested that morphisms of one sorted theories preserve
finite products. In fact, this simply follows from the equation M · T1 = T2 .
Observe that due to that equation M is the identity map on objects. Since,
moreover, M takes the projections T1 πin to the projections T2 πin , it clearly
preserves finite powers, thus, finite products.
2. The reason why one-sorted theories are requested to be equipped with a
theory morphism from N is that in this way the category of one-sorted
theories and the category of finitary monads on Set are equivalent, as
proved in A.38. (And, by the way, this is the original definition by Lawvere
from 1963.)
3. There is an obvious non-strict version of morphism of one-sorted theories,
where the equality above is weaken to a natural isomorphism between
M · T1 and T2 . See Appendix C for this approach.
11.5 Example. Recall the theory Tab of abelian groups whose objects are
natural numbers and morphisms are matrices (see 1.11). It can be canonically
considered as a one-sorted theory if we define Tab : N → Tab as the identity map
on objects, and assign to πin : n → 1 the one-row matrix with i-th entry 1 and
all other entries 0.
11.6 Remark. Given a one-sorted theory (T , T ), the functor T does not in-
fluence the concept of algebra: the category Alg T thus consists, again, of all
functors A : T → Set preserving finite products. However, the presence of T
makes the category of algebras concrete over Set (see 0.18). Assuming that we
identify Set and Alg N , the forgetful functor is simply

Alg T : Alg T → Set

(see 9.2), which is faithful by 11.8. More precisely, this forgetful functor takes
an algebra A : T → Set to the set A1, and a homomorphism h : A → B to the
component h1 : A1 → B1.

November 20, 2009 90

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

11.7 Example. For the one-sorted theory (Tab , Tab ) of abelian groups the cat-
egory Alg Tab is equivalent to Ab. (But it is not isomorphic to Ab : this si caused
by the fact that algebras for Tab are not required to preserve products strictly.
Consequently, there exist many algebras that are naturally isomorphic to alge-
bras of the form G b (see 1.11) but are not equal to any of those.) The forgetful
functor assigns, for every group G, to Gb the underlying set of G. Observe that,
unlike in Ab, there exist isomorphisms f : A → B in Alg Tab for which A1 = B1
and f1 = id but still A 6= B. In fact, given a group G we usually have many
algebras B 6= Gb naturally isomorphic to G b such that the component of the nat-
ural isomorphism at 1 is the identity. In other words, Alg Tab is not amnestic
(see 13.16).
11.8 Proposition. Let (T , T ) be a one-sorted algebraic theory. The forgetful
functor Alg T : Alg T → Set is algebraic, faithful and conservative.
Proof. Alg T is algebraic by 9.3. Let f : A → B be a homomorphism of
T -algebras. Because of the naturality of f, the following square commutes

fT n
AT n / BT n

≃ ≃
 
(AT )n / (BT )n
fTn

It is now obvious that Alg T is faithful and conservative.

11.9 Corollary. Let (T , T ) be a one-sorted algebraic theory. The forgetful
functor preserves and reflects limits, sifted colimits, monomorphisms and regular
epimorphisms.
11.10 Remark. For every T -algebra A and every subset X of its underlying
set A1 there exists the least subalgebra X of A such that X1 contains X (X
is called the subalgebra generated by X). In fact, consider the intersection of all
subalgebras of A containing X.
11.11 Remark. The concept of algebraic category in Chapter 1 used equiva-
lences of categories. For one-sorted algebraic theories we need more: the equiv-
alence functor must be concrete. This is, in fact, not enough because a quasi-
inverse of a concrete functor is in general not concrete. We are going to require
that the equivalence functor admits a quasi-inverse which is concrete.
11.12 Definition. Given concrete categories U : A → K and V : B → K over
K, by a concrete equivalence between them we mean a pair of concrete functors
′
E / E /A
A@ B
@@ ~~
@@ V ~~
~
U @@
  ~~ U
K

November 20, 2009 91

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

such that both E ·E ′ and E ′ ·E are naturally isomorphic to the identity functors.
We then say that (A, U ) and (B, V ) are concretely equivalent.

11.13 Definition. A one-sorted algebraic category is a concrete category over

Set which is concretely equivalent to Alg T : Alg T → Set for a one-sorted alge-
braic theory (T , T ).

11.14 Remark. A non-strict version of one-sorted algebraic categories is treated

in Appendix C.

11.15 Example. The category Ab with its canonical forgetful functor is a one-
sorted algebraic category. In fact, it is concretely equivalent to the category of
algebras for (Tab , Tab ) : the functor

E / Alg Tab
Ab C
CC vv
CC
C vvv
U CC vv
! {vv Alg Tab
Set

b : Tab → Set of Example 1.11 is

which to every group G assigns the algebra G
concrete. And we have the oncrete functor

E′ / Ab
Alg TabH
HH {{
HH {{
H {{ U
Alg Tab HHH {
# {}
Set

which to every algebra A : Tab → Set assigns the group G with A ≃ G b from
′ ′
1.11. It is easy to verify that both E · E and E · E are naturally isomorphic to
the identity functors.

11.16 Remark. Given a concrete category (A, U ), all subcategories of A are

considered to be concrete by using the domain restriction of U.

11.17 Proposition. Every variety of T -algebras for a one-sorted theory (T , T )

is a one-sorted algebraic category.

Proof. Let ∼ be a congruence on T and let Q : T → T / ∼ be the corre-

sponding quotient functor. Since Q preserves finite products, (T / ∼, Q · T ) is
a one-sorted theory. Consider the full subcategory A of Alg T specified by ∼ .
The equivalence
E : Alg (T / ∼) → A
constructed in 10.15 is a concrete functor (because Alg Q is concrete). Moreover,
E is an isomorphism (not just an equivalence) because it is bijective on objects.
Thus, the inverse functor E −1 : A → Alg (T / ∼) is concrete.

November 20, 2009 92

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

11.18 Example. The category Graph of graphs is algebraic, but not one-sorted
algebraic. In fact, a terminal object in Graph is the graph with one vertex and
one edge, and it has a proper subobject given by the graph G with one vertex
and no edge. Observe that G is neither terminal nor initial in Graph. Now use
the following
11.19 Lemma. In a one-sorted algebraic category a terminal object A has no
nontrivial subobjects: for every subobject m : B → A either B is an initial object
or a terminal one.
Proof. Given a one-sorted algebraic theory (T , T ), denote by A a terminal
object of Alg T and by I the initial one. Since Alg T : Alg T → Set preserves lim-
its (11.9), B(1) is a subobject of A1 = 1. If B1 ≃ 1, then m1 is an isomorphism
and thus m is an isomorphism (since Alg T is conservative). If B1 = ∅, consider
the unique monomorphism a : I → B and the induced map a1 : I1 → B1 = ∅.
Such a map is necessarily an isomorphism, and thus so is a. It is easy to see
that concrete equivalences preserve the above property of terminal objects.
11.20 Example. Even though the category of graphs is not a one-sorted alge-
braic category, the category RGraph of reflexive graphs is. Here the objects are
directed graphs
τ /
G e /G v
σ

together with a map d : Gv → Ge such that τ · d = idGv = σ · d. A morphism

from (Ge , Gv , τ, σ, d) to (G′e , G′v , τ ′ , σ ′ , d′ ) is a graph homomorphism
(he : Ge → G′e , hv : Gv → G′v )
such that he · d = d′ · hv . We consider this category as concrete over Set by
U (Ge , Gv , τ, σ, d) = Ge .
In fact, the category RGraph is concretely equivalent to the following one-
sorted algebraic category A : an object of A is a set Ge equipped with two
maps
t /
G e /G e
s

such that s · t = t and t · s = s. A morphism from (Ge , t, s) to (G′e , t′ , s′ ) is a

map he : Ge → G′e such that he · t = t′ · he and he · s = s′ · he .
1. Define E : RGraph → A by assigning to a reflexive graph (Ge , Gv , τ, σ, d)
the object (Ge , t, s) of A by defining t = d · τ and s = d · σ. This is a concrete
functor.
2. Define E ′ : A → RGraph by assigning to an object (Ge , t, s) in A the reflexive
graph (Ge , Gv , τ, σ, d) by taking as d : Gv → Ge a joint equalizer of
t /
Ge id / Ge
/
s

This yields the canonical factorizations τ : Ge → Gv of t through d, and σ : Ge →

Gv of s through d (such factorizations exist because t = s · t = t · s · t = t · t and

November 20, 2009 93

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

analogously for s). The rest of the definition of E ′ is straightforward. Again, E ′

is a concrete functor.
3. The verification that E ·E ′ and E ′ ·E are naturally isomorphic to the identity
functors is easy.

11.21 Remark. Let (T , T ) be a one-sorted theory.

1. The forgetful functor Alg T : Alg T → Set has a left adjoint. In fact, due
to 4.11 applied to YN : N op → Set we can choose a left adjoint

FT : Set → Alg T

in such a way that the square

T op / T op
N op
YN YT
 
Set / Alg T
FT

commutes. Thus, for every natural number n,

(a) FT (n) = T (n, −)

and
(b) FT (πin ) = − · T πin for all i = 0, . . . , n − 1.

2. The naturality square for η : Id → Alg T · FT applied to πin yields the

commutativity of
η1
1 / T (1, 1)

πin −·T πin

 
n / T (n, 1)
ηn

that is
ηn (i) = T πin for all i = 0, . . . , n − 1
(recall that πin is the inclusion of i).
`
3. Since FT preserves coproducts, FT X = X YT (1) for every set X.

4. T -algebras of the form FT X, for X a set, are called free algebras. If X is

finite, they are called finitely generated free algebras.

11.22 Corollary. Let (T , T ) be a one-sorted theory. T op is equivalent to the

full subcategory of Alg T of finitely generated free algebras.

Indeed by Yoneda Lemma T (n, k) ≃ Alg T (YT (k), YT (n)) = Alg T (FT (k), FT (n)).

November 20, 2009 94

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

11.23 Remark. Every object of a category with finite coproducts defines an

algebraic theory T (A) for which we need to fix coproduct injections

pni : A → nA = A + . . . + A (n summands).

The objects of T (A) are natural numbers, and morphisms from n to k are the
morphisms of A from kA = A+ . . .+ A to nA. Then T : N → T (A) is defined by
T πin = pni . Observe that T (A) is equivalent to the full subcategory of Aop on all
finite copowers of A under the equivalence functor n 7→ nA. The corresponding
category of T (A)-algebras can be equivalent to A, as we have seen in the example
A = Ab and A = Z. (In fact, if A = Alg T for a one-sorted algebraic theory
(T , T ) and A = T 1, then A is always equivalent to Alg T (A).) These theories
T (A) are often seen as the “natural” algebraic theories in classical algebra (e.g.
for A = groups, lattices, monoids, etc.).
11.24 Remark. Extending 11.21, for every one-sorted algebraic category (A, U )
with a left adjoint F ⊣ U a one-sorted theory can be constructed from the full
subcategory of Aop on the objects F {x0 , . . . , xn−1 }. Here a set of “standard
variables” x0 , x1 , x2 , . . . is assumed. In fact, the n injections

{x0 } → {x0 , . . . , xn−1 } , x0 7→ xi

define n morphisms pni : F {x0 , . . . , xn−1 } → F {x0 } in Aop . Let T be the cate-
gory whose objects are the natural numbers and whose morphisms are

T (n, k) = A(F {x0 , . . . , xk−1 }, F {x0 , . . . , xn−1 }) .

The composition in T is inherited from Aop , and so are the identity morphisms.
The functor T : N → T is determined by the above choice of morphisms pni for
all i ≤ n.
11.25 Example. For the one-sorted theory (Tab , Tab ) (11.5) the induced ad-
junction FTab ⊣ Alg Tab is, up to concrete equivalence, the usual adjunction given
by free abelian groups.
Using free algebras we can restate some facts from Chapter 5:
11.26 Proposition. Let (T , T ) be a one-sorted algebraic theory.
1. Free algebras are precisely the coproducts of representable algebras.
2. Every algebra is a regular quotient of a free algebra.
3. Regular projectives are precisely the retracts of free algebras.
Proof. 1: Following 11.21, every free algebra is a coproduct of representable
algebras. Conversely, every representable algebra is free by 11.21. Thus, every
coproduct of representable algebra is free because FT preserves coproducts.
2: Following 4.2, every algebra is a regular quotient of a coproduct of repre-
sentable algebras and then, by 1, it is a regular quotient of a free algebra.
3: This follows from 1. and 5.14.2.

November 20, 2009 95

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

11.27 Remark. We have used “finitely generated” in two different situations

above: for objects of a category, see 5.21, and as a denotation of FT X with X
finite. The following proposition demonstrates that there is no conflict. Note,
however, that a finitely generated free algebra can, in principle, coincide with
a non-finitely generated one. In fact, in the variety of algebras satisfying, for a
pair x, y of distinct variables, the equation x = y, all algebras are isomorphic.

11.28 Proposition. Let (T , T ) be a one-sorted algebraic theory.

1. Finitely generated free algebras are precisely the representable algebras.

These are precisely the free algebras which are finitely generated (in the
sense of Definition 5.21).

2. Perfectly presentable algebras are precisely the retracts of finitely generated

free algebras.

3. Finitely presentable algebras are precisely the coequalizers of (reflexive)

pairs of morphisms between finitely generated free algebras.

Proof. 1. follows from 11.21 and the observation that whenever the object
FT X is finitely generated then it is isomorphic to FT X ′ for some finite set X ′ . In
fact, the set X is the directed union of its nonempty finite subsets X ′ . Since FT
is a left adjoint and directed unions are directed colimits in Set, we see that FT X
is a directed colimit with the colimit cocone formed by all F i, where i : X ′ → X
are the inclusion maps. Moreover, since each i is a split monomorphism in
Set, FT X is the directed union of the finitely generated free algebras FT X ′ .
Since FT X is were finitely generated, there exists a finite nonempty subset
i : X ′ ֒→ X and a homomorphism f : FT X → FT X ′ such that FT i · f = idFT X .
Thus FT i : FT X ′ → FT X, being a monomorphism and a split epimorphism, is
an isomorphism, a contradiction.
2. follows from 1. and 5.14.1, and 3. follows from 1. and 5.17.

11.29 Remark. Let (T , T ) be a one-sorted algebraic theory and X a subset

of the underlying set A1 of a T -algebra A. The subalgebra of A generated by
X (see 11.9) is a regular quotient of the free algebra FT X. Indeed, consider the
homomorphisms iX : FT X → A corresponding to the inclusion iX : X → A1.
Then iX (FT X) is a subalgebra of A because the forgetful functor preserves
regular factorizations. This is obviously the least subalgebra of A containing X
(use diagonal fill-in, see 0.16) and the codomain restriction of iX is a regular
quotient.

11.30 Proposition. Let (T , T ) be a one-sorted algebraic theory and A a T -

algebra. The following conditions are equivalent:

1. A is finitely generated (see 5.21),

2. A is a regular quotient of a finitely generated free algebra,

and

November 20, 2009 96

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

3. there exists a finite subset X of A(1) not contained in any proper subalgebra
of A.
Proof. The equivalence between 1. and 2. follows from 5.21 and 11.18.1.
1 ⇒ 3 : Let A be a finitely generated object of Alg T . Form a diagram in
Alg T indexed by the poset of all finite subsets of A1 by assigning to every such
X ⊆ A1 the subalgebra X of A generated by X (see 11.10). Given finite subsets
X and Y with X ⊆ Y ⊆ A1, the connecting map X → Y is the inclusion
map. Then the inclusion homomorphisms iX : X → A form a colimit cocone of
this directed diagram. Since the functor Alg T (A, −) preserves this colimit, for
idA ∈ Alg T (A, A) there exists a finite set X such that idA lies in the image of
iX – but this proves X = A.
3 ⇒ 2 : If A = X for a finite subset X of A1, then by 11.19 A is a regular
quotient of the finitely generated free algebra FT X.
11.31 Remark.
1. Recall the notion of an equivalence relation on an object A in a category
from 3.12. If the category is Alg T , following the terminology of general
algebra in 11.32 and 11.33 we speak about congruence on the algebra
A (instead of equivalence). This is a slight abuse of terminology since
congruences were previously used for the theory T itself.
2. Similarly to 11.10, for every T -algebra A and every subset X of A1 × A1
there exists the least congruence on A whose underlying set contains X.
Such a congruence is called the congruence generated by X.
3. The finitely generated congruences are those generated by finite subsets of
A1 × A1.
11.32 Lemma. Let (T , T ) be a one-sorted algebraic theory. Given
u //
FT X B
v

in Alg T , the congruence on B generated by the image of hu, vi coincides with

that generated by the image of hu1 · ηX , v1 · ηX i, where ηX : X → (FT X)(1) is
the unit of the adjunction FT ⊣ Alg T.
Proof. Let R be the congruence by the image of hu, vi and S that generated
by the image of hu1 · ηX , v1 · ηX i. To check the inclusion R ⊆ S use the universal
property of ηX and the diagonal fill-in (cf. 0.16). The other inclusion is obvious.

11.33 Corollary. Let (T , T ) be a one-sorted algebraic theory. A T -algebra A

is finitely presentable if and only if there exists a coequalizer of the form
r1 c
// F Y /A
R r2
T

with Y a finite set and R a finitely generated congruence on FT Y.

November 20, 2009 97

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

In fact, this follows from 11.28 and 11.32 by taking B = FT Y with X and Y
finite sets.

As announced at the beginning of Chapter 10, we are going to prove that

condition (d) in Birkhoff’s Variety Theorem 10.22 can be avoided when the
theory is one-sorted. We need a preliminary lemma.

11.34 Lemma. Let (T , T ) be a one-sorted theory with the forgetful functor

U = Alg T. Consider a regular epimorphism e : A → B and a homomorphism
g : A → C in Alg T . If U g factorizes through U e in Set

Ue / UB
UA
Ug
 |
UC

then g factorizes through e in Alg T

e /B
A
g
 
C

Proof. Let f : U B → U C be such that f · U e = U g. Define h : B → C to be

the homomorphism whose component at n ∈ T is

hn = f n : U B n → U C n .

It is strightforward to check the naturality of h using that each component

eT n is surjective (3.5). Since U h = f and Alg T is faithful, we conclude that
h · e = g.

11.35 Proposition. Let (T , T ) be a one-sorted algebraic theory. If a full sub-

category A of Alg T is closed under products, subalgebras, and regular quotients,
then it is a variety.

Proof. By 10.22 all we need is proving that A is closed under directed unions.
Put U = Alg T and F = FT for short. Following the first part of the proof of
10.22 we know that A is regular epireflective in Alg T , with reflectons denoted
by rB : B → RB. Then for every directed union A = ∪i∈I Ai of subobjects
mi : Ai → A with Ai ∈ A for each i we prove that A lies in A. We use the
counit εA : F U A → A. Due to the definition of adjunction, U εA ) is a split
and therefore regular epimorphism. Since U reflects regular epimorphisms (see
11.9), εA is a regular epimorphism. Form the reflection rF UA : F U A → RF U A
and prove that εA factorizes through it: this finishes the proof because given
h : RF U A → A with εA = h · rUF A , then h is a regular epimorphism. Thus, A
is a regular quotient of RF U A, which proves that A is in A.

November 20, 2009 98

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

By Lemma 11.34 it is sufficient to prove that U εA factorizes through U rF UA

in Set. This means that given morphisms x, y : 1 ⇉ U F U A we must prove that
U rF UA · x = U rF UA · y implies U εA · x = U εA · y .
The functor U F U preserves filtered colimits (because U does, see 11.9, and F
is a left adjoint) thus
U F U A ≃ colim U F U Ai .
i∈I

Consequently, there exists i ∈ I such that x, y factorize through U F U mi

UF Umi
U F U AcGi / U; F U A
cGGGG y′
GGGG x xxxxxxx;
GGGG xxxx
x′ GGG
xx y
GG xxxxx
1
Now in Set the monomorphisms with nonempty domains split. Therefore, we
can assume without loss of generality that U mi is a split monomorphism. In
fact, if U A = colim U Ai is nonempty, then some U Ai0 for i0 ∈ I must be
nonempty and it is sufficient to choose i′ ∈ I such that Ai′ contains Ai and
Ai0 ; if U A = ∅ then U Ai = ∅ and U mi = id . Consequently, U RF U mi is a
monomorphism. Consider the following commutative diagram
RεAi
RF U AeKi / RAi
KK rF U A
KK i rAi zz=
KK zz
KK zz ≃
εAi zz
F U Ai / Ai
RF Umi F Umi mi Rmi
 
F U A εA /A
EE
ss EE
s sss E
s r rA EEE
y ss F U A
 s " 
RF U A / RA
Rε A

By applying U it, we see that the fact that U rF UA merges x = U F U mi · x′

and y = U F U mi · y ′ implies that U RF U mi · U rF UAi merges x′ and y ′ . Since
U RF U mi is a monomorphism, also U RεAi ·U rF UAi merges x and y. Consulting
the above diagram again we conclude that U εAi merges x′ and y ′ . This finishes
the proof: from
U εA · x = U εA · U F U mi · x′ = U mi · U εAi · x′
and the analogous equation for y we conclude U εA · x = U εA · y.

Recall the algebraic duality of Chapter 9: if we restrict algebraic theories

to the canonical ones, we obtain a contravariant biequivalence between the 2-
category of algebraic categories and the 2-category of algebraic theories. In the

November 20, 2009 99

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

one-sorted case a better result is obtained, since we do not have to restrict the
theories at all.

11.36 Definition. We define

1. the 2-category Th 1 of one-sorted theories to have

objects: one-sorted algebraic theories,

1-cells: morphisms of one-sorted algebraic theories,
2-cells: natural transformations,

2. the 2-category ALG 1 of one-sorted algebraic categories to have

objects: one-sorted algebraic categories,

1-cells: concrete functors,
2-cells: natural transformations.

11.37 Remark. Every 1-cell in ALG 1 is a faithful and conservative algebraic

functor.

11.38 Definition. We denote by

Alg 1 : (Th 1 )op → ALG 1

the 2-functor assigning to every one-sorted theory (T , T ) the concrete category

Alg 1 (T , T ) = (Alg T : Alg T → Set), to every 1-cell M : (T1 , T1 ) → (T2 , T2 ) the
concrete functor Alg 1 M = (−) · M, and to every 2-cell α : M → N the natural
transformation Alg 1 α : Alg 1 M → Alg 1 N whose component at a T2 -algebra A
is A · α : A · M → A · N.

11.39 Theorem. (One-sorted algebraic duality) The 2-category ALG 1 of one-

sorted algebraic categories is biequivalent to the dual of the 2-category Th 1 of
one-sorted algebraic theories. In fact, the 2-functor

Alg 1 : (Th 1 )op → ALG 1

is a biequivalence.

Proof. (1) Alg 1 is well-defined and essentially surjective (in the sense of
the 2-category ALG 1 , which means surjectivity up to concrete equivalence) by
definition of one-sorted algebraic category.
(2) We will prove that for two one-sorted algebraic theories (T1 , T1 ) and (T2 , T2 )
the functor

Alg 1(T1 ,T1 ),(T2 ,T2 )

Th 1 ((T1 , T1 ), (T2 , T2 )) / ALG 1 ((Alg T2 , Alg T2 ), (Alg T1 , Alg T1 ))

November 20, 2009 100

CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

is an equivalence of categories. The proof that Alg 1(T1 ,T1 ),(T2 ,T2 ) is full and
faithful is the same as in Theorem 9.15. It remains to prove that Alg 1(T1 ,T1 ),(T2 ,T2 )
is essentially surjective: consider a concrete functor
G
Alg T1G o Alg T2
GG ww
GG ww
G
Alg T1 GG ww
# w{ w Alg T2
Set
It is our task to find a theory morphism M : (T1 , T1 ) → (T2 , T2 ) with G ≃
Alg 1 M. We have the left adjoint FT of 11.21 and we denote by F : Alg T1 →
Alg T2 a left adjoint of G. The commutativity of the above triangle yields a
natural isomorphism
ψ : FT2 → F · FT1 .
We are going to prove that F ·YT1 factorizes (up to natural isomorphism) through
YT2 :
YT1
T1op / Alg T1 [11.1]

M op ≃ F
 
T2op / Alg T2
YT2

We define M : T1 → T2 to be the identity on objects. Consider a morphism

f : t → s in T1 . Since YT2 is full and faithful, there exists a unique morphism
M f : M t → M s such that the following diagram commutes
ψs
YT2 M (s) = YT2 T2 (s) = FT2 (s) / F FT1 (s) = F YT1 T1 (s) = F YT1 (s) [11.2]

YT2 M(f ) F YT1 (f )

 ψt

YT2 M (t) = YT2 T2 (t) = FT2 (t) / F FT1 (t) = F YT1 T1 (t) = F YT1 (t)

The equalities FTi · YN = YTi · Tiop (i = 1, 2) for the embedding YN : N op → Set

come from 11.21. The functoriality of
M : T1 → T2
follows from the uniqueness of M f. Moreover, since the isomorphism ψ is nat-
ural, if f = T1 g for some g : t → s in N then M f = T2 g. Diagram [11.2] gives
also the natural isomorphism needed in [11.1]. This finishes the proof: F ≃ M ∗
by 9.3, and then G ≃ Alg M.
11.40 Remark. A related duality involving the categories (not the 2-categories)
of one-sorted algebraic theories and one-sorted algebraic categories can be ob-
tained if we restrict the latter to the uniquely transportable ones. This result,
established in Appendix C, gives an alternative approach to the classical duality
between finitary monads and finitary monadic categories over Set presented in
Appendix A.

November 20, 2009 101

 CHAPTER 11. ONE-SORTED ALGEBRAIC CATEGORIES

Historical Remarks for Chapter 11

In his dissertation [63] F. W. Lawvere presents the name algebraic category for
one equivalent to the category Alg T of algebras of a one-sorted algebraic theory.
Our decision to use concrete equivalence is motivated by the precise analogy one
gets to finitary monadic categories over Set (see Appendix A).
Another variant, based on pseudo-concrete functors in place of the concrete
ones, is to take all categories pseudo-concretely equivalent to the categories
Alg T above. This is shortly mentioned in Appendix C.

November 20, 2009 102

Chapter 12

Algebras for an endofunctor

Throughout this chapter a finitary endofunctor H of Set is assumed to be given.

We discuss the category of algebras for H. We will prove in 13.23 that it is a
one-sorted algebraic category.

12.1 Remark.

1. The concept of H-algebra in 2.24 can be formulated for the endofunctor

H of an arbitrary category K : it is a pair (A, a) consisting of an object
A and a morphism a : HA → A. The category

H-Alg

has as objects H-algebras and as morphisms from (A, a) to (B, b) those

morphisms f : A → B for which f · a = b · Hf.

2. We denote by
UH : H-Alg → K
the canonical forgetful functor (A, a) 7→ A.

3. In the present chapter we restrict ourselves to finitary endofunctors H of

Set. See also 12.17.

12.2 Remark. We want to show how colimits of H-algebras are obtained. We

begin with the simplest case: the initial H-algebra. We will prove that it can
be obtained by iterating the unique morphism u : ∅ → H∅. More precisely, let
us form the ω-chain

u / H∅ Hu / H 2∅ H2 u / H 3∅ / ...
∅

We call it the initial chain of H. Its colimit

I = colim n∈N H n ∅

103
 CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

carries the structure of an H-algebra. Indeed, since H preserve colimits of

ω-chains,
HI ≃ colim n∈N H(H n ∅) ≃ colim n∈N H n ∅ = I .
We denote by i : HI → I the canonical isomorphism. More detailed, denote by

vn : H n ∅ → I (n ∈ N)

a colimit cocone for I. Then i : HI → I is defined by

i · Hvn = vn−1 for all n ≥ 1 .

12.3 Lemma. The H-algebra i : HI → I is initial.

Proof. For every algebra a : HA → A define a cocone fn : H n ∅ → A of the
initial chain as follows: f0 : ∅ → A is unique and

fn+1 = a · Hfn : HH n ∅ → A .

The unique morphism f : I → A with f · vn = fn (n ∈ N) is a homomorphism:

since the cocone (Hvn ) is a colimit cocone, thus collectively epimorphic, this
follows from the commutative diagram
i /
HI cF {= I
FF Hv {
FF n−1 {{
FF
F {{{vn
{
Hf H n∅ f
xx
xx
xx Hfn
 {xx 
HA /A
a

Conversely, if f : I → A is a homomorphism, the above diagram proves that

for every n ≥ 1 the above morphism fn is equal to f · vn . This shows the
uniqueness.
12.4 Example. We describe the initial HΣ -algebra for the polynomial functor
a
HΣ X = Σk × X k
k∈N

by applying labelled trees.

Recall that a tree is a directed graph with a distinguished node (root) such
that for every node there exists a unique path from the root into it. We work
with ordered trees, that is, for every node the set of children nodes is linearly
ordered (“from left to right”). Trees which are isomorphic for an isomorphism
respecting the ordering and the labels are identified.
For every signature Σ, by a Σ-tree is meant a tree labelled in Σ so that every
node has the number of children equal to the arity of its label. More generally,
given a set X of variables, by a Σ-tree on X is meant a tree labelled in Σ + X

November 20, 2009 104

CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

in such a way that every node with n > 0 children has a label in Σn and every
leaf has a label in Σ0 + X.
Initial HΣ -algebra: We can represent HΣ ∅ = Σ0 by the set of all singleton
trees labelled by elements of Σ0 . Given a tree representation of HΣk ∅, we represent
a
HΣk+1 ∅ = Σn × (HΣk ∅)n
n∈N

by the set of all trees

σE
89:;
?>=< [12.1]
yy EE
yy EE
yy EE
y EE
yy
p1 ······ pn

with σ ∈ Σn and p1 , . . . , pn ∈ HΣk ∅. In this way we see that for every k ∈ N

HΣk ∅ = all Σ-trees of depths less than k .

The above ω-chain is the chain of inclusion maps ∅ ⊆ HΣ ∅ ⊆ HΣ2 ∅ ⊆ . . . and its
colimit [
I= HΣk (∅)
k∈N

is the set of all finite Σ-trees. The algebraic structure i : HΣ I → I is given by

tree-tupling: to every n-tuple of trees corresponding to the summand of σ ∈ Σn
it assigns the tree [12.1] above.
12.5 Remark. Free H-algebras. Let H be a finitary endofunctor of Set. We
now describe free H-algebras, that is, a left adjoint of the forgetful functor UH .
For every set X, the endofunctor

H(−) + X

is also finitary. Therefore, following 12.3, it has an initial algebra.

12.6 Proposition. The free H-algebra on X is the initial algebra for the end-
ofunctor H(−) + X.
Explicitly, if H ∗ X is the initial algebra for H(−) + X with structure

iX : HH ∗ X + X → H ∗ X,

then the components

ϕX : HH ∗ X → H ∗ X and ηX : X → H ∗ X

of iX form the algebra structure and the universal arrow, respectively.

Proof. This follows easily from the observation that to specify an algebra
for H(−) + X on an set A means to specify an algebra HA → A for H and a
function X → A.

November 20, 2009 105

 CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

12.7 Corollary. The free H-algebra on X is the colimit of the ω-chain

∅ → H∅ + X → H(H∅ + X) + X → . . .
12.8 Notation. FH : Set → H-Alg denotes the left adjoint of UH . In case
H = HΣ we use FΣ instead of FHΣ .
12.9 Example. We describe the free HΣ -algebra on an set X. Observe that
HΣ (−) + X = HΣ
for the signature Σ obtained from Σ by adding nullary operation symbols from
X. Thus, the description of initial algebra in 12.4 immediately yields a descrip-
tion of the free HΣ -algebra FΣ X on X as the algebra of all finite Σ-trees on
X. That is, finite labelled trees with leaves labelled in X + Σ0 , and nodes with
n > 0 children labelled in Σn . The operations of FΣ X are given by tree-tupling.
The universal arrow assigns to a variable x ∈ X the singleton tree labelled x.
12.10 Example. For the signature Σ of a single binary operation ∗, we have a
description of FΣ X for X = {p1 , p2 , p3 } as all binary trees with leaves labelled
by p1 , p2 , p3 . Examples:

∗
89:;
?>=<
~~ @@@
~ @@
~~~ @@
~ @@
~~
pi
?>=<
89:; pi
89:;
?>=< pj
@ABC
GFED

∗B
89:;
?>=< ∗
89:;
?>=<
~~ BB }} @@@
~ BB } @@
~~~ BB }}} @@
~ BB } @@
~~ }}
∗
89:;
?>=< pk
@ABC
GFED pi
89:;
?>=< ∗
89:;
?>=<
 ???  @@@
 ??  @@
 ??  @@
 ??  @@
 
pi
89:;
?>=< pj
@ABC
GFED pj
@ABC
GFED pk
@ABC
GFED

12.11 Example. A commutative binary operation: this can be expressed by

the functor H assigning to every set X the set HX of all unordered pairs in X,
and to every function f the function Hf acting as f componentwise.
An H-algebra is a set with a commutative binary operation. The free algebra
on X is the colimit of the chain
∅ → H∅ + X = X → HX + X → H(HX + X) + X → . . .
We can represent the elements of HZ as binary non-ordered trees with both
subtrees elements of Z, then we see that the n-th set in the above chain consists
of precisely all binary non-ordered trees of depth less than n with leaves labelled
in X. Consequently the initial algebra is
IX = all unordered binary trees over X .

November 20, 2009 106

CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

12.12 Proposition. For every finitary endofunctor H on Set the category

H-Alg has
(a) limits and
(b) sifted colimits
preserved by the forgetful functor.
We know from 6.30 that H preserves sifted colimits. We can generalize (b) to
say: for every type of colimits preserved by H the category H-Alg has colimits
of that type preserved by UH .

Proof. We prove the more general formulation of (b), the proof of (a) is
analogous. Let D : D → H-Alg be a diagram with objects Dd = (Ad , ad ) and let
A = colim Ad be the colimit of UH ·D in Set with the colimit cocone cd : Ad → A.
If H preserves this colimit, there exists a unique H-algebra structure a : HA →
A turning each cd into a homomorphism. In fact, the commutative squares
Hcd
HAd / HA

ad a
 
Ad /A
cd

define a unique a since (i) cd · ad is a cocone of UH · D and (ii) M cd is the colimit

cocone of H · UH · D. It is easy to see that the algebra (A, a) is a colimit of D
in H-Alg with the cocone cd .
12.13 Theorem. For every finitary endofunctor H on Set the category H-Alg
is cocomplete. It also has regular factorizations of morphisms preserved by the
forgetful functor.
Proof. (1) We start with the latter statement. Observe that regular epimor-
phisms split in Set and therefore H preserves them. Given a homomorphism
h : (A, a) → (B, b) in H-Alg and a factorization h = m · e with

e : A → C a regular epimorphism , m : C → B a monomorphism

in Set, use the diagonal fill-in to obtain an H-algebra structure c : HC → C

turning e and m into homomorphisms:
He / HC
HA
a Hm
 
c
A HB
e b
  
C /B
m

November 20, 2009 107

 CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

Since UH is faithful, m in a monomorphism in H-Alg . And e is a regular epimor-

phism in H-Alg because given a pair u, v : X ⇉ A with coequalizer e in Set, the
corresponding homomorphisms u, v : FH X → A from the free H-algebra have
the coequalizer e in H-Alg .
(2) Arguing as in 4.1, to prove that H-Alg is cocomplete it is sufficient to prove
that it has finite coproducts. Since by 12.3 H-Alg has initial object, it remains
to consider binary coproducts. Thus, we are to prove that the diagonal functor
∆ : H-Alg → H-Alg × H-Alg has a left adjoint, see 0.11. Since by 12.12 the
category H-Alg is complete, it is sufficient (using the Adjoint Functor Theorem
0.8) to find a solution set for every pair (A1 , a1 ), (A2 , a2 ) of H-algebras. That
is, we need a set of cospans
f1 f2
(A1 , a1 ) / (C, c) o (A2 , a2 )

in H-Alg through which all cospans factorize. Consider the coproduct

c1 c2
A1 / A1 + A2 o A2

in Set, denote by f : A1 + A2 → C the morphism induced by the cospan (f1 , f2 ),

and let f : FH (A1 + A2 ) → (C, c) be the homomorphism corresponding to f by
adjunction. We claim that a solution set is provided by those cospans (f1 , f2 )
such that f is a regular quotient of FH (A1 + A2 ). This is indeed a set because
f is a regular epimorphism also in Set. For any cospan
g1 g2
(A1 , a1 ) / (D, d) o (A2 , a2 )

consider the regular factorization in H-Alg

g
FH (A1 + A2 ) / (D, d)
MMM v;
MMM vvv
M v
e MMM vv m
& vv
(C, c)

We get a new cospan in Set by defining

ci ηA1 +A2 e
fi : Ai / A1 + A2 / FH (A1 + A2 ) /C (i = 1, 2)

(where η is the unit of the adjunction FH ⊣ UH ). The cospan (g1 , g2 ) factorizes

through (f1 , f2 ) because gi = m·fi . Moreover, (f1 , f2 ) is a cospan in H-Alg : this
follows easily from the fact that gi and m are homomorphisms of H-algebras and
m is a monomorphism in Set. Finally, (f1 , f2 ) has the desired property because
f = e.
12.14 Remark. So far we have mentioned, besides the polynomial functors
HΣ , only one finitary functor that is not polynomial, see 12.11. And that func-
tor is an obvious quotient of the polynomial functor HΣ X = X × X.

November 20, 2009 108

CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

In general a quotient of a functor H is represented by a natural transformation

α : H → H with epimorphic components. We now prove that finitary endofunc-
tors of Set are indeed precisely the quotients of the polynomial ones:
12.15 Theorem. For an endofunctor H on Set the following conditions are
equivalent:
1. H is finitary,
2. H is a quotient of a polynomial functor
and
3. every element of HX lies in the image of Hi for the inclusion i : Y → X
of a finite subset Y.
Proof. 3 ⇒ 1 : Let
D : D → Set
be a filtered diagram with a colimit cocone
cd : Dd → C (d ∈ obj D) .
We prove that the diagram D · H has the colimit
Hcd : HDd → HC
in Set. For that it is by 0.6 sufficient to verify that
(a) every element x of HC lies in the image of Hcd for some d
and
(b) given elements y1 , y2 ∈ HDd merged by Hcd , there exists a connecting
morphism f : d → d′ of D with Hf also merging y1 and y2 .
For (a) choose a finite subset i : Y → X with x lying in the image of Hi. Since
C = colim Dd is a filtered colimit in Set, there exists a factorization
i = cd · j for some d ∈ obj D and j : Y → Dd .
Thus x lies in the image of Hcd .
For (b) choose a finite subset i : Z → Dd such that y1 , y2 lie in the image of
Hi. Since C = colim Dd is a filtered colimit, there exists a connecting morphism
f : d → d′ such that the domain restriction of cd′ to the image i′ : Z ′ → Dd′ of
f · i is a monomorphism. We obtain a commutative diagram
e / Z′
Z
i i′
 Df 
DdB / Dd′
BB {{
BB {{
B
cd BB {
{ c′
}{{ d
C

November 20, 2009 109

 CHAPTER 12. ALGEBRAS FOR AN ENDOFUNCTOR

for some morphism e such that cd′ · i′ is a monomorphism. Without loss of

generality Z ′ 6= ∅ provided C 6= ∅. Then cd′ · i′ is a split monomorphism. Then
H(cd′ · i′ ) also is a split monomorphism, and we conclude that HDf merges
y1 , y2 .
1 ⇒ 3 : Obvious from the description of fitered colimits in Set (see 0.6).
1 ⇒ 2 : This follows from Yoneda Lemma. Define the signature Σ by using,
for every n ∈ N, the elements of Hn as the operation symbols σ of arity n ∈
N. Shortly: Hn = Σn . Then we have a natural transformation α : HΣ → H
which, given an operation symbol σ ∈ Σn (that is, σ ∈ Hn), assigns to the
corresponding n-tuple f : n → Z the value
αZ (σ(f )) = Hf (σ) .
In other words, the component of αZ at the functor Set(n, −) corresponding to
σ ∈ Σn is the Yoneda transformation of σ. Condition 3. tells us precisely that
αZ is a surjective map for all set Z.
2 ⇒ 3 : Every polynomial functor satisfies condition 3. Indeed, to choose an
element x ∈ HΣ X means to fix a symbol σ ∈ Σn , so that x = (x1 , . . . , xn ) ∈
X × . . . × X. Therefore, as Y we can take {x1 , . . . , xn }.
Let now α : HΣ → H be a quotient, and fix an element x ∈ HX. Since
αX : HΣ X → HX is surjective, there exists y ∈ HΣ such that αX (y) = x. Find
a finite subset i : Y → X such that y = (HΣ i)(y) for some y ∈ HΣ . By naturality
of α we have
x = αX (HΣ i)(y) = (Hi)αY (y) .

12.16 Remark. In 13.23 we will see that for every presentation of a functor
H as a quotient functor of HΣ , the category of H-algebras can be viewed as an
equational category of Σ-algebras.
12.17 Remark. Most of the result in this chapter has an obvious generalization
to endofunctors H of cocomplete categories K which preserve sifted colimits.
1. The initial chain of 12.2 is defined by denoting by ∅ an initial object of K
and using the unique morphism u : ∅ → H∅. The corresponding H-algebra
is initial.
2. The free H-algebra on an object X of K is the intial algebra for the
endofunctor H(−) + X.
3. The category H-Alg is complete and cocomplete, and the forgetful functor
into K preserves limits and sifted colimits.
Historical Remarks for Chapter 12
Algebras for an endofunctor were introduced by J. Lambek in [62]. The ini-
tial algebra construction 12.2 and its free-algebra variation 12.7 stem from [1].
Factorizations and colimits in categories H-Alg were studied in [2].
The fact that finitary endofunctors on Set yield one-sorted algebraic cat-
egories follows from the work of M. Barr [15] on free monads, see Appendix
A.

November 20, 2009 110

Chapter 13

Equational categories of
Σ-algebras

In the present chapter we prove that one-sorted algebraic categories are precisely
the equational categories of Σ-algebras for (one-sorted) signatures Σ. The case
of S-sorted signatures is treated in Chapter 14.
13.1 Remark. We described a left adjoint
FΣ : Set → Σ-Alg
of the forgetful functor UΣ : Σ-Alg → Set in 12.9. The more standard description
is that FΣ X is the following Σ-term-algebra: the underlying set is the smallest
set such that
- every element x ∈ X is a Σ-term
and
- for every σ ∈ Σ of arity n and for every n-tuple of Σ-terms p1 , . . . , pn we
have a Σ-term σ(p1 , . . . , pn ).
The Σ-algebra structure on FΣ X is given by the formation of terms σ(p1 , . . . , pn ).
This defines a functor FΣ : Set → Σ-Alg on objects. To define it on morphisms
f : X → Z, let FΣ f be the function which in every term p of FΣ X substitutes
for every variable x ∈ X the variable f (x). More explicitly:
- if x ∈ X, then FΣ f (x) = f (x),
- if p1 , . . . , pn ∈ FΣ X and σ ∈ Σn , then
FΣ f (σ(p1 , . . . , pn )) = σ(FΣ f (p1 ), . . . , FΣ f (pn )).

It is easy to verify that FΣ is a well-defined functor which is naturally isomorphic

to the Σ-tree functor of 12.9. Thus, we have FΣ ⊣ UΣ . The unit of the adjunction
is the inclusion of variables into the set of Σ-terms: ηX : X → FΣ X.

111
 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

13.2 Notation. Suppose that a set of “standard variables” x0 , x1 , x2 , . . . is

given. Then the free Σ-algebras

FΣ {x0 , . . . , xn−1 }

yield, by 11.24, a one-sorted theory for Σ-Alg . We denote this theory by

(TΣ , TΣ ) .

Thus, morphisms from n to 1 in TΣ are the Σ-terms in variables x0 , . . . , xn−1 .

General hom-sets are given by k-tuples of these terms:

TΣ (n, k) = (FΣ n)k .

And TΣ : N → TΣ assigns to every function g : k → n, that is, g ∈ N (n, k), the

k-tuple of terms
xg(0) , . . . , xg(k−1) .
13.3 Lemma. The category Σ-Alg is concretely equivalent to the category of
algebras of (TΣ , TΣ ).
Proof. Define E : Σ-Alg → Alg TΣ on objects as follows. For a Σ-algebra
(A, a), the corresponding functor from TΣ to Set is given on objects by n 7→ An
and on morphisms t : n → 1 by the function An → A of evaluation of the term
t in the given algebra. This function takes a map f : n → A to f (t) ∈ A, where
f : FΣ {x0 , . . . , xn−1 } → (A, a) is the unique homomorphism exteding f.
Conversely, if B is a TΣ -algebra we get a Σ-algebra structure on the set B1
as follows: if σ ∈ Σn , then σ(x0 , . . . , xn−1 ) ∈ FΣ n = TΣ (n, 1) and this yields
an n-ary operation on B1 by applying B to that morphism (recall that Bn is
isomorphic to the n-th power of B1).
This gives a concrete equivalence E : Σ-Alg → Alg TΣ .
13.4 Definition. Given signatures Σ and Σ′ , a morphism of signatures is a
function f : Σ → Σ′ preserving the arities. This leads to the category of signa-
tures Sign – this is just Set/N.
13.5 Definition. For every one-sorted algebraic theory (T , T ) we define the
signature
C(T , T )
as the signature whose n-ary symbols are precisely the morphisms from n to 1
in T :
(C(T , T ))n = T (n, 1) .
This construction can be easily extended to morphisms of one-sorted theories.
13.6 Example. The signature C(TΣ , TΣ ) has all Σ-terms in variables x0 , . . . , xn−1
as n-ary operation symbols. Therefore, there is a canonical morphism of signa-
tures
ηΣ : Σ −→ C(TΣ , TΣ )
given by ηΣ (σ) = σ(x0 , . . . , xn−1 ) ∈ FΣ n for any σ of arity n.

November 20, 2009 112

CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

13.7 Proposition. (A free one-sorted theory on a signature) For every sig-

nature Σ the theory (TΣ , TΣ ) is free on Σ. That is, given a one-sorted theory
(T , T ) for every morphism G : Σ → C(T , T ) of signatures there exists a unique
morphism G : (TΣ , TΣ ) → (T , T ) of one-sorted theories such that C(G) · ηΣ = G
ηΣ
Σ FF / C(TΣ , TΣ )
FF r
FF
rrrrr
F
G FF r
" xrrr C(G)
C(T , T )

Proof. (1) We define a functor G : TΣ → T on objects by n 7→ n and on mor-

phisms p ∈ TΣ (k, 1), that is Σ-terms on {x0 , . . . , xk−1 }, by structural induction:
(i) For variables xi ∈ TΣ (k, 1) put Gxi = T pki , the chosen projection in
T (k, 1).
(ii) Given p = σ(p1 , . . . , pn ) where σ ∈ Σn , pi ∈ T (k, 1) (i = 1, . . . , n) and
Gpi are defined already, put

<Gp1 ,...,Gpn > Gσ

Gp : k /n /1

It is clear that G · TΣ = T and C(G) · ηΣ = G.

(2) Uniqueness: Let M : (TΣ , TΣ ) → (T , T ) be a morphism of one-sorted theories
with C(M ) · ηΣ = G. Since M preserves finite products, all we have to prove is
that it is determined when precomposed by ηΣ . This is the case, indeed:
(i) For variables xi ∈ TΣ (k, 1) use M · TΣ = T, so that M (ηΣ (xi )) is the i-th
projection pki .
(ii) Consider p = σ(p1 , . . . , pn ) with σ ∈ Σn and pi ∈ T (k, 1) (i = 1, . . . , n).
Since σ(p1 , . . . , pn ) = σ(x1 , . . . , xn )· < p1 , . . . , pn > in TΣ , we have that M p =
M (ηΣ (σ)) · hM p1 , . . . , M pm i.
13.8 Remark. Let (T , T ) be a one-sorted theory. If we apply the construction
described in the first part of the proof of 13.7 to the identity morphism on
C(T , T ), we get a morphism

ε(T ,T ) : TC(T ,T ) −→ (T , T )

of one-sorted theories. It is clearly full and then, by 10.13, the unique functor ε′
making commutative the following diagram of morphisms of one-sorted theories
is an isomorphism
Q
TC(T ,T ) / TC(T ,T ) / ∼
FF u
FF uu
F
ε(T ,T ) FFF uuu′
F" uu ε
uz u
T
Therefore:

November 20, 2009 113

 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

1. (T , T ) is a quotient of the free one-sorted theory TC(T ,T ) .

2. Alg T is a variety of C(T , T )-algebras.

In order to improve the previous result we need the notion of equational

category of Σ-algebras. We start comparing the classical notion of equation
with the one introduced in Chapter 10.

13.9 Remark.

1. Classically equations are expressions

t = t′

where t and t′ are terms in variables x0 , . . . , xn−1 for some n. This is

a special case of 10.1: here we have a parallel pair t, t′ : n ⇉ 1 in the
theory TΣ . Also, a Σ-algebra (A, a) satisfies this equation in the classical
sense (that is, for every interpretation f : {x0 , . . . , xn−1 } → A we have
f (t) = f (t′ )) iff the corresponding TΣ -algebra satisfies this equation in the
sense of 10.1.

2. In fact, for the theory TΣ equations in the sense of 10.1 are equivalent to
the classical equations: given a parallel pair

t, t′ : n ⇉ k in TΣ

and given the k projections pki : k → 1 (i = 0, . . . , k − 1) specified by the

functor TΣ : N → TΣ , we get a k-tuple of equations

pki · t = pki · t′

in the classical sense. It is clear that a Σ-algebra satisfies each of these k

equations iff the corresponding TΣ -algebra satisfies t = t′ in the sense of
10.1.

13.10 Definition.

1. By an equational category of Σ-algebras is meant a full subcategory of

Σ-Alg formed by all algebras satisfying a set E of equations. We denote
such a caegory by
(Σ, E)-Alg .
The pair (Σ, E) is called an equational theory.

2. Equational categories are concrete categories over Set which are, for some
signature Σ, equational categories of Σ-algebras.

13.11 Theorem. One-sorted algebraic categories are precisely the equational

categories.

November 20, 2009 114

CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

More detailed: a concrete category over Set is one-sorted algebraic iff it is

concretely equivalent to an equational category of Σ-algebras for some signature
Σ.

Proof. (1) Every equational category of Σ-algebras is a one-sorted algebraic

category. In fact, following 13.9, the concrete equivalence Σ-Alg ≃ Alg TΣ of
13.3 restricts to a concrete equivalence between (Σ, E)-Alg and Alg (TΣ / ∼E ),
where ∼E is the congruence on TΣ generated by E (see 10.7).
(2) Conversely, every one-sorted algebraic category is concretely equivalent to
an equational category of Σ-algebras. In fact, Alg T is equivalent to Alg (TΣ / ∼)
for some congruence ∼ on TΣ (13.8) and, therefore, to (Σ, E)-Alg where E is
the set of all equations u = v where u and v are congruent terms.

13.12 Example. Recall that a semigroup is an algebra on one associative binary

operation. This means that we consider the Σ-algebras with Σ = {∗} which
satisfy the equation
(x ∗ y) ∗ z = x ∗ (y ∗ z) .
Thus, the theory of semigroups is the quotient theory TΣ / ∼ where ∼ is the
congruence generated by the equation above.

13.13 Example. Beside the algebraic theory Tab of abelian groups of 1.11
we now have a different one, based on the usual equational presentation: let
Σ = {+, −, 0} with + binary, − unary and 0 nullary. Then a theory of abelian
groups is the quotient TΣ / ∼ modulo the congruence on TΣ generated by the
four equations
(x + y) + z = x + (y + z)
x+y = y+x
x+0 = 0
x + (−x) = 0

13.14 Example. Recall that a monoid is a semigroup (M, ∗) with a unit. We

can consider the category of all monoids as the category (Σ, E)-Alg where Σ
has a binary symbol ∗ and a nullary symbol e, and E contains the associativity
of ∗ and the equations
x = x∗e
x = e∗x
(equivalently, as the category Alg (TΣ / ∼) where ∼ is the congruence generated
by the associativity of ∗ and the equations above).

13.15 Example. For every monoid M, an M -set is a pair (X, α) consisting of

a set X and a monoid action α : M × X → X (the usual notation is mx in place
of α(m, x)) such that every element x ∈ X satisfies m(m′ x) = (m ∗ m′ )x for all
m, m′ ∈ M, and ex = x. The homomorphisms f : (X, α) → (Z, β) of M -sets are
the functions f : X → Z with f (mx) = mf (x) for all m ∈ M and x ∈ X. We
can describe this category as (Σ, E)-Alg where Σ = M with all arities equal to

November 20, 2009 115

 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

1, and E contains the equations

x = ex
(m ∗ m′ ) x = m (m′ x)

for all m, m′ ∈ M.
13.16 Definition. A concrete category U : A → K is called:
1. Amnestic provided that given an isomorphism i : A → A′ in A with U i =
idUA , then A = A′ . (This implies i = idA because U is faithful.)
2. Transportable provided that for every object A in A and every isomorphism
i : U A → X in K there exists an isomorphism j : A → B in A with
U B = X and U j = i.
3. Uniquely transportable if in 2. the isomorphism j is unique.
13.17 Example.
1. For every one-sorted algebraic theory (T , T ) the concrete category
Alg T : Alg T → Set

is transportable – but almost never uniquely transportable (see 11.7).

In fact, given a T -algebra A and a bijection i : A1 → X, let B : T → Set
be defined on objects by Bn = X n and on morphisms f : n → k in the
unique way making the powers of i natural:
Af
(A1)n ≃ An / Ak = (A1)k

in ik
 
X n = Bn / Bk = X k
Bf

Then these powers form a natural isomorphism j : A → B with j1 = i.

2. For every signature Σ the concrete category
UΣ : Σ-Alg → Set , UΣ (A, σ A ) = A

is uniquely transportable. In fact, given a bijection i : A → X there is a

unique way of defining, for an n-ary symbol σ ∈ Σ, the operation σ X so
that the square
σA /A
An
in i
 
Xn /X
σX

commutes. The same is true for the equational categories of Σ-algebras.

For example, the category of abelian groups is uniquely transportable.

November 20, 2009 116

CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

3. For every endofunctor H of a category K the concrete category

UH : H-Alg → K , (A, a) 7→ A

is uniquely transportable. In fact, given an algebra a : HA → A and an

isomorphism i : A → X in K, the unique algebra x : HX → X for which i
becomes a homomorphism is x = i · a · Hi−1 .

13.18 Remark.

1. For every concrete category we have

transportable + amnestic ⇔ uniquely transportable.

In fact, if (A, U ) is transportable and amnestic, and if in 13.16.2 we have

another isomorphism j ′ : A → B ′ with U j ′ = i, use 13.16.1 on the iso-
morphism j ′ · j −1 : B → B ′ to conclude B = B ′ . Then j = j ′ since U is
faithful.
Conversely, if (A, U ) is uniquely transportable, then by applying 13.16.3
to i = idUA we deduce that it is amnestic.

2. Transportability is not invariant under concrete equivalence, thus, not

all one-sorted algebraic categories are transportable. For example, let
E : Set ′ → Set be the full subcategory of Set consisting of all cardinal
numbers. Then (Set ′ , E) is one-sorted algebraic because it is concretely
isomorphic to (Set, Id). But it obviously fails to be transportable.

3. We will see in 13.21 a converse of 13.17.2: every uniquely transportable

one-sorted algebraic category is (up to concrete isomorphism) an equa-
tional category.

4. In case of groupoids, the concept of (uniquely) transportable is precisely

the concept of (discrete) fibration.

13.19 Definition. Given concrete categories U : A → K and V : B → K by a

concrete isomorphism between them we mean a concrete functor

E /B
A@
@@ 
@@ 
U @@ 
   V
K

and a functor E ′ : B → A such that both E · E ′ and E ′ · E are equal to the

identity functors. (Note that such a functor E ′ is necessarily concrete.) We
then say that (A, U ) and (B, V ) are concretely isomorphic.

13.20 Lemma. A concrete equivalence between uniquely transportable concrete

categories is a concrete isomorphism.

November 20, 2009 117

 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

Proof. Given a concrete equivalence

E /B
A@
@@ 
@@ 
U @@ V
 
K

between uniquely transportable categories, we prove that E is bijective on ob-

jects – thus it is a (concrete) isomorphism.
(a) If A and A′ are objects of A with EA = EA′ , then for the identity morphism
of EA there exists, since E is full, a morphism f : A → A′ with Ef = id . And
f is of course an isomorphism in A. Since U f = V (Ef ) = id, amnesticity of U
implies A = A′ .
(b) For every object B of B there exists an isomorphism i : EA → B in B
yielding an isomorphism V i : U A → V B in K. Let j : A → A′ be the unique
isomorphism in A with U j = V i. The isomorphism Ej · i−1 : B → EA′ fulfils
V (Ej · i−1 ) = U j · V i−1 = id, thus by amnesticity of V we have B = EA′ .

13.21 Corollary. Uniquely transportable one-sorted algebraic categories are,

up to concrete isomorphism, precisely the equational categories.

In fact, this follows from 13.11, 13.17 and 13.20.

Birkhoff’s Variety Theorem (10.22) can be restated in the context of Σ-

algebras:

13.22 Theorem. Let Σ be a signature. A full subcategory A of Σ-Alg is equa-

tional if and only if it is closed in Σ-Alg under

(a) products,

(b) subalgebras,

and

(c) regular quotients.

In fact, this follows from 10.22 and 13.11.

13.23 Proposition. Let Σ be a one-sorted signature and H : Set → Set be a

quotient of the polynomial functor HΣ . The concrete category H-Alg is con-
cretely isomorphic to an equational category of Σ-algebras.

Proof. Let α : HΣ → H be a natural transformation with epimorphic com-

ponents. We get a full and faithful functor

I : H-Alg → HΣ -Alg I(A, a) = (A, a · αA ) .

November 20, 2009 118

CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

Moreover, I is concrete since the diagram

I / HΣ -Alg = Σ-Alg
H-AlgG
GG v
GG vv
G vv
UH GGG vv
# vz v UΣ
Set

clearly commutes.
Since I is injective on objects, H-Alg is concretely isomorphic to the full
subcategory I(H-Alg ) of Σ-Alg . We are to prove that I(H-Alg ) satisfies the
conditions of Proposition 13.22.
(a) Consider the above commutative diagram. Since UH preserves and UΣ
reflects limits and sifted colimits, I preserves them. In particular, I(H-Alg ) is
closed in Σ-Alg under products.
(b) Let f : (A, x) → I(B, b) be a monomorphism in Σ-Alg (and then in Set).
Since αA is a strong epimorphism in Set, we get an H-algebra structure on A
by diagonal fill-in:
HΣ f
HΣ AG / HΣ B
GG α
GG A αB
GG
G# 
Mf
x HA / HB

a b
 { 
A /B
f

This shows that (A, x) = I(A, a).

(c) First observe that if f : I(A, a) → (B, y) is an isomorphism in Σ-Alg , then
(B, y) ∈ I(H-Alg ). Indeed f : (A, a) → (B, b) is an isomorphism in H-Alg ,
where b = f · a · Hf −1 , and I(B, b) = (B, y). (In other words, (H-Alg , I) is
transportable, see 13.16.)
Now consider a regular epimorphism e : I(A, a) → (B, y) in Σ-Alg . Its kernel
pair, being a subobject of I(A, a) × I(A, a), lies in the image of I and we can
take its coequalizer (Q, q) in H-Alg . This coequalizer is preserved by I (because
it is reflexive, see 3.4). Thus I(Q, q) ≃ (B, y).

13.24 Remark. In general algebra the concepts of finitely generated and

finitely presentable Σ-algebra are defined as follows: a Σ-algebra (A, a) is called

(a) finitely generated if it is generated by a finite set (see 11.10), that is, (A, a)
is isomorphic to
FΣ {x1 , . . . , xn }/ ∼

for some congruence ∼ on FΣ {x1 , . . . , xn }

and

November 20, 2009 119

 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

(b) finitely presentable if it is such a quotient modulo a finitely generated

congruence, that is, (A, a) is isomorphic to FΣ {x1 , . . . , xn }/ ∼ for some
congruence ∼ generated by finitely many equations.
It turns out that these concepts coincide with the categorical concepts of 5.21
and 5.3, respectively. Let us first observe that:
(1) Every subalgebra generated by a set X (see 11.10) is a regular quotient
of the free algebra FΣ X. In fact, let B be the subalgebra of A generated
by X and let f : FΣ X → A be the unique homomorphism extending the
inclusion map. Then the image f (FΣ X) is a subalgebra of A because
the forgetful functor preserves regular factorizations, see 11.9. This is,
obviously, the least subalgebra containing X, and the codomain restriction
of f is a regular epimorphism.
(2) Conversely, every regular quotient q : FΣ X → A of a free algebra generated
by X is “generated by X” – more precisely, the image of the map
ηΣ UΣ q
X / UΣ (FΣ X) / UΣ A

generates A.
13.25 Proposition. A Σ-algebra is a finitely generated object of Σ-Alg if and
only if it is a regular quotient of a finitely generated free algebra.
Proof. (1) Let (A, a) be a finitely generated object of Σ-Alg . Form a diagram
in Σ-Alg indexed by the poset of all finite subsets of A by assigning to every
such X ⊆ A the subalgebra X of A generated by X (see 11.10). Given finite
subsets X and Z with X ⊆ Z ⊆ A, the connecting map X → Z is the inclusion
map. Then the inclusion homomorphisms iX : X → A form a colimit cocone of
this directed diagram. Since the functor Σ-Alg (A, −) preserves this colimit, for
idA ∈ Σ-Alg (A, A) there exists a finite set X such that idA lies in the image of
iX – but this proves X = A.
(2) Let us prove that every quotient A = F {x1 , . . . , xn }/ ∼ is finitely generated.
Given a directed diagram of subobjects Bi (i ∈ I) with a colimit B = colim Bi ,
it is our task to prove that Σ-Alg (A, −) preserves this colimit, that is, every
homomorphism h : A → B factorizes through one of the colimit homomorphisms
bi : Bi → B. For the finite set {h(ηX (xk ))}nk=1 there exists i ∈ I such that this
set lies in the image of bi . From that it easily follows that the image of h is
contained into the image of bi . Since bi is a monomorphism, it follows that there
exists a homomorphism g : A → Bi with h = bi · g, as requested.
13.26 Proposition. A Σ-algebra is a finitely presentable object of Σ-Alg if
and only if it is a regular quotient of a finitely generated free algebra modulo a
finitely generated congruence.
Proof. (1) Let (A, a) be a finitely presentable object. By 11.28 there exists
a coequalizer
u / c /
F X Σ /F Z Σ (A, a)
v

November 20, 2009 120

CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

with X and Z finite. Let ∼ be the congruence generated by the finitely many
equations u(ηΣ (x)) = v(ηΣ (x)) where x ∈ X, then (A, a) is clearly isomorphic
to FΣ Z/ ∼ : the canonical morphism q : FΣ Z → FΣ Z/ ∼ is namely also a
coequalizer of u and v.
(2) Conversely, let Z be a finite set and ∼ a congruence on FΣ Z generated
by equations t1 = s1 , . . . , tk = sk . For X = {1, . . . , k} define homomorphisms
u, v : FΣ X ⇉ FΣ Z by

u(i) = ti and v(i) = si for i = 1, . . . , k .

Then the canonical map q : FΣ Z → FΣ Z/ ∼ is a coequalizer of u and v. There-

fore, FΣ Z/ ∼ is finitely presentable by 11.28.

Historical Remarks for Chapter 13

This chapter contains classical results one can find in every introduction to
general algebra, e.g. [37].
The concepts of amnestic and transportable concrete categories are taken
from [4].

November 20, 2009 121

 CHAPTER 13. EQUATIONAL CATEGORIES OF Σ-ALGEBRAS

November 20, 2009 122

Chapter 14

S-sorted algebraic
categories

In the previous chapters we have considered one-sorted algebraic categories.

They are categories equipped with a forgetful functor into Set, like groups,
abelian groups, lattices, etc. In computer science one often considers S-sorted
algebras, where S is a given nonempty set (of sorts), and algebras are not sets
with operations, but rather S-indexed families of sets with operations of given
sort. This means that the forgetful functor is into Set S rather than into Set.
In this chapter we revisit one-sorted algebraic categories generalizing definitions
and several results to the S-sorted case, see 14.3.
Analogously to the one-sorted case, where the theory has objects X n (which
we represented by n alone) and projections πin : X n → X are specified, in case
of S-sorted theories we have objects Xs for s ∈ S that generated the whole
theory in the sense that every object of T is a product

Xs0 × . . . × Xsn−1

for some word w = s0 . . . sn−1 over S. We, again, suppose that projections

πiw : Xs0 × Xsn−1 → Xsi (i = 0, . . . , n − 1)

are chosen. And, again, instead with the above product we work with the word
s0 . . . sn−1 alone. In other words, the theory N which plays a central role for
one-sorted theories is generalized to the following:

14.1 Notation. For every nonempty set S (of sorts) we denote by

S∗

the category whose objects are the finite words on S, and whose morphisms
from s0 . . . sn−1 to t0 . . . , tk−1 are all functions f : k → n with sf (i) = ti for all

123
 CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

i = 0, . . . , k − 1.
In particular, for every word w = s0 . . . sn−1 we have the projections

πiw : s0 . . . sn−1 → si (i = 0, . . . , n − 1)

given by the i-th injection 1 7→ n in Set.

14.2 Example. N = {s}∗ provided that we identify every natural number n

with the word ss . . . s of length n.

14.3 Remark. We know from 1.10 that S ∗ is an algebraic theory for Set S , and
every word w is a product of one-letter words with the projections π0w , . . . , πn−1
w
S ∗
above. We are going to identify Set with Alg S . The full embedding

YS ∗ : (S ∗ )op → Set S

assigns to a word w = s0 . . . sn−1 the S-sorted set

YS ∗ (w)s = {i = 0, . . . , n − 1 ; si = s} .

14.4 Definition. Let S be a nonempty set.

1. A S-sorted algebraic theory is a pair (T , T ) where T is an algebraic theory

whose objects are the words over S, and T : S ∗ → T is a theory morphism
which is the identity map on objects.

2. A morphism of S-sorted algebraic theories M : (T1 , T1 ) → (T2 , T2 ) is a

functor M : T1 → T2 such that M · T1 = T2 .

14.5 Remark. Analogously to 11.3 we have not requested that morphisms of

one sorted theories preserve finite products since this simply follows from the
equation M · T1 = T2 . Observe that due to that equation M is the identity map
on objects.

14.6 Example.

1. For the category Graph of graphs we have an S-sorted theory with S =

{v, e} and T : S ∗ → Tgraph is determined by T v = v and T e = e (recall
that Tgraph is the free completion under finite products of the category
Cgraph described in 1.15).

2. Let C be a small category, put S = obj C, and let ETh : C → TC denote the
free completion of C under finite products (1.6); recall from 1.7 that the
objects of TC can be viewed as words over S. Moreover, if C is discrete,
then TC = S ∗ . Therefore, we have a unique theory morphism TC : S ∗ → TC
which is the identity map on objects. We obtain an S-sorted theory

(TC , TC ) .

November 20, 2009 124

CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

14.7 Remark. Precisely as in the one-sorted case, the functor T does not
influence the concept of algebra: the category Alg T thus consists, again, of
all functors A : T → Set preserving finite products. However, the presence of
T makes the category of algebras concrete over Set S : the forgetful functor is
simply
Alg T : Alg T → Set S
(see 14.4). More precisely, this forgetful functor takes an algebra A : T → Set
to the S-sorted set hAsis∈S , and a homomorphism h : A → B to the S-sorted
function with components hs : As → Bs.
14.8 Proposition. Let (T , T ) be an S-sorted algebraic theory. The forgetful
functor
Alg T : Alg T → Set S
is faithful, algebraic and conservative. It thus preserves and reflects limits, sifted
colimits, monomorphisms and regular epimorphisms.
The proof is analogous to that of 11.8.
14.9 Remark. The concept of one-sorted algebraic category in Chapter 11 used
concrete equivalences of categories over Set. For S-sorted algebraic theories we
need, analogously, concrete equivalences over Set S (see 11.12).
14.10 Definition. A S-sorted algebraic category is a concrete category over
Set S which is concretely equivalent to Alg T : Alg T → Set S for an S-sorted
algebraic theory (T , T ).
14.11 Proposition. Every variety of T -algebras for an S-sorted theory (T , T )
is an S-sorted algebraic category.
The proof is analogous to that of 11.17.
14.12 Remark. Let (T , T ) be an S-sorted theory.
1. The forgetful functor Alg T : Alg T → Set S has a left adjoint. In fact, due
to 4.11 applied to YS ∗ : (S ∗ )op → Set S we can choose a left adjoint

FT : Set S → Alg T

in such a way that the square

T op
(S ∗ )op / T op

YS ∗ YT
 
Set S / Alg T
FT

commutes.
2. T -algebras of the form FT (X), for X an S-sorted set, are called free alge-
bras. If X is finite (1.10), they are called finitely generated free algebras.

November 20, 2009 125

 CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

14.13 Corollary. Let (T , T ) be an S-sorted theory. T op is equivalent to the

full subcategory of Alg T of finitely generated free algebras.
14.14 Remark. Results about finitely presentable and perfectly presentable
algebras and regular projectives generalize easily from the one-sorted case, see
11.26 – 11.33, to the S-sorted one, we leave this for the reader. Let us just stress
that when working with variables, in the S-sorted case to every variable a sort
is assigned. That is, the corresponding object of variables also lives in Set S .
14.15 Theorem. (One-sorted algebraic duality) The 2-category ALG S of S-
sorted algebraic categories, concrete functors and natural transformations is
biequivalent to the dual of the 2-category Th S of S-sorted algebraic theories,
morphisms of S-sorted theories and natural transformations.
The proof is completely analogous to that of 11.39.

The concept of signature also generalizes from one-sorted to S-sorted easily.

Note however that we cannot speak about n-ary operations. Instead we need to
specify n + 1 sorts: one sort for each of the variables, and the sort of the result.
14.16 Definition.
1. An S-sorted signature is a set Σ (of operation symbols) together with an
arity function assigning to every symbol σ in Σ an element of S ∗ × S; we
write
σ : s0 . . . sn−1 → s
if the pair is (s0 . . . sn−1 , s), including the case

σ: s

for n = 0.
2. A Σ-algebra is a pair (A, a) consisting of an S-sorted set A = hAs is∈S and
a function assigning to every element σ : s0 . . . sn−1 → s of Σ a mapping

σ A : As0 × . . . × Asn−1 → As .

(In case n = 0 we have a constant σ A ∈ As .)

3. Σ-homomorphisms from (A, a) to (B, b) are S-sorted functions

f = hfs i with fs : As → Bs (s ∈ S)

such that for every element σ : s0 . . . sn−1 → s of Σ the square

σA
As0 × . . . × Asn−1 / As

fs0 ×...×fsn−1 fs
 
Bs0 × . . . × Bsn−1 / Bs
σB

November 20, 2009 126

CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

commutes. This yields a concrete category

Σ-Alg

of Σ-algebras with the forgetful functor

UΣ : Σ-Alg → Set S , UΣ (A, a) = A .

14.17 Example.
1. The category of graphs has the form Σ-Alg for S = {v, e} and Σ consisting
of two operations of arity e → v (called τ and σ in 1.15).
2. For Σ = ∅ we have Σ-Alg = Set S .
3. For sequential automata (see 1.20) put S = {s, i, o} and Σ = {δ, γ, ϕ} with
arities δ : si → s, γ : s → o and ϕ : s.
4. For the example of stacks 1.19 put S = {s, n} and Σ = {succ, push, pop,
top, 0, e} with the arities given in 1.19.
14.18 Remark.
1. The description of a left adjoint

FΣ : Set S → Σ-Alg

of UΣ is completely analogous to 13.1. Given an S-sorted set X of vari-

ables, we form the smallest S-sorted set FΣ X (of terms) such that every el-
ement x ∈ Xs is a term of sort s, and for every σ ∈ Σ of arity s0 . . . sn−1 →
s and for every n-tuple of terms p0 , . . . , pn−1 of sorts s0 , . . . , sn−1 respec-
tively, we have a term σ(p0 , . . . , pn−1 ) of sort s. The Σ-algebra structure
on FΣ X is given by the formation of terms σ(p0 , . . . , pn−1 ).
This defines a functor FΣ : Set S → Σ-Alg on object. To define it on
morphisms, proceed as in 13.1.
2. We obtain, assuming a countable set of “standard variables xsi of sort s”
for every s ∈ S, an S-sorted theory

(TΣ , TΣ )

analogous to the one described in 13.2: the words s0 . . . sn−1 represent the
sn−1
free Σ-algebra FΣ {xs00 , . . . , xn−1 }. The categories Σ-Alg and Alg TΣ are
concretely equivalent over Set S , this is analogous to 13.3.
3. Equations in the sense of 10.1 can be substituted by expressions

t = t′

where t and t′ are two elements of FΣ X of the same sort (for some finite
S-sorted set X of standard variables). This is analogous to 13.9, except

November 20, 2009 127

 CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

that in the S-sorted case the quantification of variables must be made

explicit. If Σ is a one-sorted signature and t, t′ are terms in FΣ X, then in
place of X we can take the set Z ⊆ X of all variables that appear in t or
t′ . An algebra satisfies t = t′ independently of whether we work with FΣ Z
or FΣ X. This is not so in S-sorted signatures, as we demonstrate below.
We therefore need the following
14.19 Definition. Given an S-sorted signature Σ, by an equation is meant an
expression
∀x0 ∀x1 . . . ∀xn−1 (t = t′ )
where xi is a variable of sort si (i = 0, . . . , n − 1) and t, t′ are elements of
FΣ {x0 , . . . , xn−1 } of the same sort s. A Σ-algebra (A, a) satisfies the equation
provided that for every S-sorted function f : {x0 , . . . , xn−1 } → A the unique
homomorphism f : FΣ {x0 , . . . , xn−1 } → (A, a) extending f fulfils f s (t) = f s (t′ ).
In case n = 0 we write ∀ ∅ (t = t′ ).
14.20 Example.
1. In the signature of graphs, see 14.17, consider variables x, x′ of sort v and
a variable y of sort e. the equation
∀x ∀x′ (x = x′ )
describes graphs on at most one vertex. Whereas
∀x ∀x′ ∀y (x = x′ )
describes all graphs that either have no edge or have just one vertex.
2. In the theory of stacks there are several equations one expects to be re-
quired. For example, if a natural number x is inserted into a stack y and
then deleted, the stack does not change:
∀x ∀y (pop(push(x, y)) = y .
Other such equations are
∀x ∀y (top(push(x, y)) = x
and (due to our definition of top)
top(e) = 0 .

14.21 Example. Let us return to Example 10.23 explaining that Birkhoff’s

Variety Theorem requires, in general, the use of directed unions. The example
worked with Set N which is Σ-Alg for the empty N-sorted signature. Let xn and
yn be variables of sort n ∈ N, and consider the equation quantifying yn and all
x0 , x1 , x2 , . . . :
∀yn ∀x0 ∀x1 ∀x2 . . . (xn = yn ) .
Then algebras, i.e., N-sorted sets, satisfy these equations iff they lie in the
category A. However, infinite quantification brings us out of the finitary logic
(and out of the realm of Definition 10.1).

November 20, 2009 128

CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

14.22 Definition.

1. Let Σ be an S-sorted signature. By an S-sorted equational category of

Σ-algebras is meant a full subcategory of Σ-Alg formed by all algebras
satisfying a set E of equations (in the sense of 10.1 or, equivalently, 14.18).

2. S-sorted equational categories are concrete categories over Set S which are,
for some signature Σ, S-sorted equational categories of Σ-algebras.

14.23 Proposition.

1. S-sorted algebraic categories are precisely the S-sorted equational cate-

gories. More detailed: a concrete category over Set S is S-sorted alge-
braic iff it is concretely equivalent to an S-sorted equational category of
Σ-algebras for some signature Σ.

2. Uniquely transportable S-sorted algebraic categories are, up to concrete

isomorphism, precisely the S-sorted equational categories.

The proof is completely analogous to that of 13.11 and 13.21.

14.24 Remark. Every S-sorted equational category is, of course, complete

and cocomplete. In particular initial algebras exist in all S-sorted equational
categories. In theoretical computer science these algebras are used as a formal-
ization of “abstract data types”: these are given by operations and equations
and consist of elements generated by the given operations (no extra variables
are used) and satisfy only the equations that are consequences of the given ones.
An abstract data type is thus, precisely as initial objects should be, determined
only up to isomorphism. We illustrate this on a couple of examples:

14.25 Example.

1. Natural numbers is a one-sorted abstract data type given by a constant 0

and a unary operation s (successor). This correspond to the initial algebra
of the one-sorted signature Σ = {s, 0} with arity 1 and 0, respectively. In
fact, every initial Σ-algebra is a representation of natural numbers.

2. Stacks of natural numbers. Here we need the two-sorted signature Σ of

14.17.4. Its initial algebra does not resemble stacks because we will have
formal terms such as top(e), top(top(e)), etc. However, the equational
category given by the three equations of 14.20.2 has an initial algebra
I = hIn , Is i where In is the abstract data type of natural numbers (no
equation involves the operation succ) and Is consists of stacks

e = [ ], [x], [x, y], [x, y, z], . . .

of elements x, y, z, . . . of In .

November 20, 2009 129

 CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

14.26 Remark. For one-sorted

` signatures we have Σ-Alg concretely equivalent
to HΣ -Alg , where HΣ A = σ∈Σ An (for n = arity of σ), see 2.24.
Analogously for S-sorted signatures Σ : define
HΣ : Set S → Set S
on objects A = hAs is∈S by setting the sort s of HΣ A as follows; we denote by
Σs ⊆ Σ the set of all symbols of output sort s and put
a
(HΣ A)s = As0 × . . . × Asn−1
σ∈Σs

for the arity s0 . . . sn−1 → s of σ. Then there is a concrete equivalence

E / HΣ -Alg
Σ-AlgH
HH u
HH uu
HH uuu
UΣ HH u
# uz u UHΣ
Set S
assigning to every Σ-algebra (A, a) the HΣ -algebra (A, a) where the coproduct
components of a
as : As0 × . . . × Asn−1 → As
σ∈Σs

are the given operations σ A .

14.27 Proposition. Every finitary endofunctor H of Set S is a quotient of a
polynomial functor HΣ for some S-sorted signature Σ. Moreover, the concrete
category H-Alg is concretely isomorphic to an equational category of Σ-algebras.
Proof. For the first statement, the argument in 12.15 using Yoneda Lemma
generalizes without a problem: define an S-sorted signature Σ whose operations
σ of (an arbitrary) arity s0 . . . sn−1 → s are precisely the elements of sort s in
HX, where the S-sorted set X is given by
Xt = {i = 0, . . . , n − 1; si = t} for all t ∈ S .
Then define α : HΣ → H by taking such an operation symbol σ and putting
αZ (σ(f )) = (Hf )s (σ)
for all S-sorted function f from the above set X into Z.
For the second statement, the only difference to the proof of 13.23 is that in
the S-sorted case we must also check that I(H-Alg ) is closed in Σ-Alg under
directed unions. This follows from the commutativity of the diagram
I / HΣ -Alg = Σ-Alg
H-AlgH
HH u
HH uu
HH uuu
UH HH u
$ zuu UΣ
S
Set
since UH preserves sifted colimits and UΣ reflects them (see 12.17.3).

November 20, 2009 130

CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

14.28 Remark.
1. The converse implication of Theorem 12.15 does not generalize to the
S-sorted case: a quotient of a polynomial functor on Set S need not be
finitary.
A simple example can be presented in Set N : start with the constant
functor of value 2 = 1 + 1 (the S-sorted set having two elements in every
sort). This functor is clearly polynomial. Let H be the quotient with
HX = 1 whenever all sorts of X are nonempty, else, HX = 2. This
functor does not preserve the filtered colimit of all finitely presentable
subobjects of 1.
2. For finite sets S of sorts Theorem 12.15 fully generalizes: finitary endo-
fuctors of Set S are precisely the quotients of polynomial functors. In fact,
the proof of 12.15 easily modifies: in part (b) of the implication 3 ⇒ 1
choose the S-sorted set Z ′ in such a way that for every sort s we have
Zs′ 6= ∅ iff Cs 6= ∅; since D is filtered and S is finite, this choice is clearly
possible. Then, again, c′d · i′ is a split monomorphism.

Historical Remarks for Chapter 14

Historical comments on S-sorted algebras have been mentioned already at the

end of Chapter 1. For a short introduction to applications of S-sorted algebras
see [97].

November 20, 2009 131

 CHAPTER 14. S-SORTED ALGEBRAIC CATEGORIES

November 20, 2009 132

Chapter 15

Morita equivalence

In this chapter we study the problem of the presentation of an algebraic category

by different algebraic theories. This is inspired by the classical work of Kiiti
Morita who in 1950’s studied this problem for the categories R-Mod of left
modules over a ring R. He completely characterized pairs of rings R and S such
that R-Mod and S-Mod are equivalent categories; such rings are nowadays called
Morita equivalent. We will recall the results of Morita below, and we will show
in which way they generalize from R-Mod to Alg T where T is an algebraic
theory.
We begin with a particularly simple example.
15.1 Example. In 1.9, we described a one-sorted algebraic theory N of Set :
N is the full subcategory of Setop whose objects are the natural numbers. Here
is another one-sorted theory of Set : T2 is the full subcategory of Setop whose
objects are the even natural numbers 0, 2, 4, 6, . . . . T2 obviously has finite prod-
ucts. Observe that T2 is not idempotent-complete (consider the constant func-
tions 2 → 2) and that N is an idempotent completion of T2 : for every natu-
ral number n we can find an idempotent function f : 2n → 2n with precisely
n fixed points. Then n is obtained by splitting f. Following 6.14 and 8.12,
Alg T2 ≃ Alg N ≃ Set.
In fact, we can repeat the previous argument for every natural number k > 0.
In this way we get a family Tk , k = 1, 2, . . . of one-sorted algebraic theory of Set
(with T1 = N ). We will prove later that, up to equivalence, there is no other
one-sorted algebraic theory of Set.
Clearly, if T and T ′ are algebraic theories and if there is an equivalence
T ≃ T ′ , then Alg T and Alg T ′ are equivalent categories. The previous example
shows that the converse is not true.
15.2 Definition. Two algebraic theories T and T ′ are called Morita equivalent
if the corresponding categories Alg T and Alg T ′ are equivalent.
From 6.14 and 8.12. we already know a simple characterization of Morita
equivalent algebraic theories: two theories are Morita equivalent iff they have

133
 CHAPTER 15. MORITA EQUIVALENCE

equivalent idempotent completions. In case of S-sorted algebraic categories a

much sharper result can be proved. Before doing so, let us recall the classical
result of Morita.

15.3 Example. Let R be a unitary ring (not necessarily commutative) and

denote by R-Mod for the category of left R-modules. There are two basic con-
structions:

1. Matrix ring R[k] . This is the ring of all k × k matrices over R with the
ususal addition, multiplication, and unit matrix. This ring R[k] is Morita
equivalent to R for every k > 0, i.e., the category R[k] -Mod is equivalent
to R-Mod.

2. Idempotent modification uRu. Let u be an idempotent element of R, uu =

u, and let uRu be the ring of all elements x ∈ R with ux = x = xu with
the binary operation inherited from R and the neutral element u. This ring
is Morita equivalent to R whenever u is pseudoinvertible, i.e., eum = 1
for some elements e and m of R.

Morita’s original result is that the two operations above are sufficient: if a ring
S is Morita equivalent to R, i.e., R-Mod and S-Mod are equivalent categories,
then S is isomorphic to the ring uR[k] u for some pseudoinvertible idempotent
k × k matrix u.

We generalize now Morita constructions to one-sorted algebraic theories. For

these theories we take N as the set of objects, see 1.9, and write T instead of
(T , 1).

15.4 Definition. Let T be a one-sorted algebraic theory.

1. The matrix theory T [k] , for k = 1, 2, 3, . . . is the one-sorted algebraic theory

whose morphisms f : p → q are precisely the morphisms f : kp → kq of T ;
composition and identity morphisms are defined as in T .

2. Let u : 1 → 1 be an idempotent morphism (u · u = u). We call u pseu-

doinvertible provided that there exist morphisms m : 1 → n and e : n → 1
such that e · un · m = id1 . The idempotent modification of T is the theory
uT u whose morphisms f : p → q are precisely the morphisms of T satis-
fying f · up = f = uq · f. The composition is defined as in T , the identity
morphism on p is up .

15.5 Remark.

1. Both T [k] and uT u are well defined. In fact, T [k] has finite products with
p = 1 × . . . × 1 : the i-th projection is obtained from the i-th projection in
T of kp = k × . . . × k. Also uT u has finite products with p = 1 × . . . × 1 :
the i-projection πi : p → 1 of T yields a morphism u · πi : p → 1 of uT u
(i = 1, . . . , k) and these morphisms form a product p = 1 × . . . × 1 in uT u.

November 20, 2009 134

CHAPTER 15. MORITA EQUIVALENCE

2. Idempotent modifications are much more “concrete” than (the equality

of) idempotent completions. This will be seen on examples illustrating
Morita equivalence below.

15.6 Theorem. Let T be a one-sorted algebraic theory.

1. the matrix theories T [k] are Morita equivalent to T for all k > 0,

and

2. the idempotent modifications uT u are Morita equivalent to T for all pseu-

doinvertible idempotents u.

Proof. 1: Matrix theory T [k] . We have a full and faithful functor T [k] → T
defined on objects by n 7→ nk and on morphisms as the identity mapping.
Every objects of T is a retract of an object coming from T [k] : in fact, for every
n consider the diagonal morphism ∆ : n → nk = n × . . . × n. Consequently, T
and T [k] have the same idempotent completion. Thus, by 8.12, they are Morita
equivalent.
2: Idempotent modification uT u. Here we consider T as a full subcategory of
(Alg T )op via the Yoneda embedding (1.4)

YT : T → (Alg T )op .

Following 8.3, the idempotent YT (u) : YT (1) → YT (1) has a splitting in (Alg T )op ,
say
YT (u) idA
YT (1) / YT (1) AE /A
EE y< EE yy<
EE yy EE yy
E
ǫ EE
y E yyǫ
E" yy η η EE
" yy
yy
A YT (1)

Consider also the subcategory TA of (Alg T )op of all powers An , n ∈ N. TA is a

one-sorted algebraic theory, and it is Morita equivalent to T . In fact, every object
of T is a retract of one in TA and vice-versa – this clearly implies that T and
TA have a joint idempotent completion (obtained by splitting their idempotents
in (Alg T )op ). Indeed, since A is a retract of YT (1), Ap is a retract of YT (p).
Conversely, consider m : 1 → n and e : n → 1 in T such that e · un · m = id1
as in 15.4.2. Then YT (1) is a retract of An via ǫn · YT (m) : YT (1) → An and
YT (e) · η n : An → YT (1), and then YT (p) is a retract of Anp .
To complete the proof, we construct an equivalence functor Ȳ : uT u → T1 . On
objects it is defined by Ȳ (p) = Ap , and on morphisms f : p → q by

Ȳ f
Ap / Aq
O
ηp ǫq

YT (p) / YT (q)
YT (f )

November 20, 2009 135

 CHAPTER 15. MORITA EQUIVALENCE

in (Alg T )op . Observe that Ȳ (idp ) = idAp because ǫ · η = idA . Now, we check
the equation
YT (f ) = η q · Ȳ (f ) · ǫp [15.1]
Indeed:

YT (f ) = YT (u)q · YT (f ) · YT (u)p = η q · ǫq · YT (f ) · η p · ǫp = η q · Ȳ (f ) · ǫp .

From equation [15.1] since ǫp is a (split) epimorphism and η q is is a (split)

monomorphism, we deduce that Ȳ preserves composition (because YT does),
and that Ȳ is faithful (because YT is). Since Ȳ is surjective on objects, it
remains to show that it is full: consider h : Ap → Aq in (Alg T )op , we define
k = η q · h · ǫp : YT (p) → YT (q). Since YT is full, there is a f : p → q in T such
that YT (f ) = k. Now:

Ȳ (f ) = ǫq · YT (f ) · η p = ǫq · η q · h · ǫp · η p = h .

It remains to check that f is in uT u :

YT (f ) · YT (up ) = η q · h · ǫp · η p · ǫp = η q · h · ǫp = k = YT (f )

and then f · up = f because YT is faithful; analogously, uq · f = f.

15.7 Theorem. For two one-sorted algebraic theories T and S the following
conditions are equivalent:
1. S is Morita equivalent to T ;
2. S is, as a category, equivalent to an idempotent modification uT [k] u of a
matrix theory of T for some pseudoinvertible idempotent u of T [k] .
Proof. Consider an equivalence functor

E : Alg S → Alg T

and the Yoneda embeddings YS : S op → Alg S, YT : T op → Alg T (recall, from

1.5, that YS and YT preserve finite coproducts). Since YS (1) is perfectly pre-
sentable in Alg S (5.5), then A = E(YS (1)) is perfectly presentable in Alg T ,
therefore, due to (5.14) it is a retract of YT (n) for some n in N :

idA
AE /A
EE yy<
EE yy
E
η EE yyǫ
" yy
YT (n)

There exists a unique u : n → n in T such that YT (u) = η · ǫ, and such a u is

an idempotent. We consider u as an idempotent on 1 in T [n] and prove that u
is pseudoinvertible there. For this, choose an S-algebra Ā and an isomorphism

November 20, 2009 136

CHAPTER 15. MORITA EQUIVALENCE

i : YT (n) → E Ā. Since E is an equivalence functor, Ā is perfectly presentable,

thus it is a retract of YS (k) for some k ∈ N :
idĀ
Ā DD /
DD zz= Ā
DD zz
η̄ DD zzzǭ
! z
YS (k)

Consider now the composites

i E η̄ kη
YT (n) / E Ā / EYS (k) ≃ kA / kYT (n) ≃ YT (nk)

and
kǫ / kA ≃ EYS (k) Eǭ / E Ā i−1 / YT (n)
YT (nk) ≃ kYT (n)

then there exist unique morphisms e : nk → n and m : n → nk in T which YT

maps on the above composites. One immediately checks that YT (e · uk · m) = id,
that is e · uk · m = id . Thus u is pseudoinvertible in T [n] .
To complete the proof, we construct an equivalence functor Ē : S → uT [n] u. It
is the identity map on objects. If f : p → q is a morphism in S, Ēf is the unique
morphism np → nq in T such that
YT (Ēf )
qYT (n) ≃ YT (nq) / YT (np) ≃ pYT (n)
O
qǫ pη

qA ≃ E(YS (q)) / E(YS (p)) ≃ pA
E(YS (f ))

commutes. Using once again YT (u) = η · ǫ and the faithfulness of YT , one easily
checks that up · Ēf · uq = Ēf, so that Ēf is a morphism p → q in uT [n] u. The
proof that Ē is a well defined, full and faithful functor is analogous to that in
Theorem 15.6 and is left to the reader.
15.8 Example. All one-sorted theories of Set. These are, up to equivalence
of categories, precisely the theories Tk of 15.1. In fact, it is easy to see that
[k]
Tk ≃ T1 is the matrix theory for every k ≥ 1. Moreover, given an idempotent
u : 1 → 1 of Tk , then the function u : k → k in Set is pseudoinvertible iff it is
invertible, thus u = id . Consequently, there are no other one-sorted theories of
Set.
15.9 Example. Let R be a ring with unit. Following 11.22, we can describe a
one-sorted theory TR of R-Mod : TR is the full subcategory of R-Modop of the
finitely generated free R-modules Rn (n ∈ N). Every one-sorted algebraic theory
of R-Mod is equivalent to TS for some ring S which is Morita equivalent to R.
Indeed, the two constructions of 15.3 fully correspond to the two constructions
of 15.4:

November 20, 2009 137

 CHAPTER 15. MORITA EQUIVALENCE

1. T(R[k] ) is equivalent to (TR )[k] ;

2. given an idempotent element u ∈ R, the corresponding module homomor-
phism ū : R → R with ū(x) = ux fulfils: uRu is equivalent to ū(TR )ū.
15.10 Example. For every monoid M, consider the category M -Set (cf. 13.14).
Two monoids M and N are called Morita equivalent if M -Set and N -Set are
equivalent categories. Here we need just one operation on monoids: if N is
Morita equivalent to M, then N is isomorphic to an idempotent modification
uM u for some pseudoinvertible idempotent u of M.
In contrast with the situation of 15.9, M -Set has, in general, many one-sorted
theories not connected to any Morita equivalent monoid. (This is true even
for M = {∗}, since M -Set = Set has infinitely many theories which are not
equivalent as categories, see 15.1.) However, all unary theories of M -Set have
the form which correspond to Morita equivalent monoids. By a unary theory
we mean a one-sorted algebraic theory T on objects T n (n ∈ N) which is a free
completion of its full subcategory on {T } (i.e., of the endomorphism monoid of
T ) under finite products. The category M -Set has an obvious one-sorted theory
T[M] : the theory of free M -sets M + . . . + M on n generators (n ∈ N) as a full
subcategory of (M -Set)op . Consequently, for every Morita equivalent monoid
N we have a unary theory T[N ] for the category M -Set. And these are, up to
categorical equivalence, all unary theories. In fact, let T be a unary theory with
Alg T equivalent to M -Set. For the monoid N = T (T, T ), there is an obvious
categorical equivalence between Alg T and N -Set : every N -set A : T (T, T ) →
Set has an essentially unique extension to a T -algebra A′ : T → Set, and (−)′
is the desired equivalence functor. Therefore, N is Morita equivalent to M, and
T is equivalent to T[N ] .
15.11 Remark. Another approach to classical Morita theory for rings is based
on the following result, due to Eilenberg and Watts (see [19], Theorem 2.3):
Let R, S be unitary rings; the assignement
M 7→ M ⊗S (−) : S-Mod → R-Mod
induces a bijection between isomorphism classes of R-S-bimodules and isomor-
phism classes of colimit preserving functors.
Using Eilenberg-Watts Theorem one can prove that R and S are Morita equiv-
alent iff there are bimodules M, N and bimodule isomorphisms
M ⊗S N ≃ R and N ⊗R M ≃ S
These facts are easy to generalize to algebraic theories. The generalization of
Eilenberg-Watts Theorem is given by the following lemma, which is a conse-
quence of 4.15.
15.12 Lemma. Let T be an algebraic theory. The functor
YT : T op → Alg T
is a free completion of T op conservative with respect to finite coproducts. This
means that

November 20, 2009 138

CHAPTER 15. MORITA EQUIVALENCE

1. Alg T is cocomplete and YT preserves finite coproducts

and
2. for every functor F : T op → B preserving finite coproducts, where B is a
cocomplete category, there exists an essentially unique functor F ∗ : Alg T →
B preserving colimits with F naturally isomorphic to F ∗ · YT .
15.13 Definition. Let T , S be algebraic theories. A bimodule

M: T ⇒S

is a finite coproduct preserving functor M : T op → Alg S.

15.14 Remark.
1. The functor YT : T op → Alg T is a bimodule T ⇒ T . More generally, every
morphism of theories T → S induces a bimodule T ⇒ S by composition
with YS .
2. Bimodules compose: given bimodules M : T ⇒ S and N : S ⇒ R we
define N ◦ M : T ⇒ R by N ∗ · M, where N ∗ : Alg S → Alg R is the colimit
preserving extension of N (15.12). This composition is associative up to
isomorphism and YS ◦ M ≃ M ≃ M ◦ YT . (In other words, the 2-category
T hBim defined below is in fact a bicategory in the sense of [22].)
15.15 Definition.
1. We define the 2-category T hBim to have
objects: algebraic theories,
1-cells: bimodules,
2-cells: natural transformations.
2. We define the 2-category ALG colim to have
objects: algebraic categories,
1-cells: colimit preserving functors,
2-cells: natural transformations.
15.16 Corollary. Let T and S be algebraic theories.
1. The assignement

M : T ⇒ S 7→ M ∗ : Alg T → Alg S

extends to a biequivalence T hBim ≃ ALG colim .

2. T and S are Morita equivalent if and only if there exist two bimodules
M : T ⇒ S and N : S ⇒ T such that N ◦ M ≃ YT and M ◦ N ≃ YS .

November 20, 2009 139

 CHAPTER 15. MORITA EQUIVALENCE

In fact, T hBim → ALG colim is an equivalence on hom-categories by 15.12. The

rest of the proof is obvious.

Historical Remarks for Chapter 15

The classical results concerning equivalences for categories of modules were

proved by K. Morita in [76]. Thirty years later J.J. Dukarm proved a gen-
eralization to one-sorted algebraic theories in [40].
For one-sorted theories an approach to Morita equivalence via bimodules is
due to F. Borceux and E. M. Vitale [28]. Theorem 15.7, a many-sorted version of
Morita equivalence, is due to J. Adámek, M. Sobral and L. Sousa [12]. Example
15.10 is due to B. Banaschewski, see [14].
The Eilenberg-Watts theorem quoted in 15.11 was independently proved by
S. Eilenberg in [43] and C. E. Watts in [96]. An exhaustive treatment of Morita
theory for rings in terms of bimodules appears in the monograph of H. Bass
[19].

November 20, 2009 140

Chapter 16

Free exact categories

We know that every algebraic category is an exact category having enough

regular projective objects (see 3.18 and 5.15). In the present chapter, we study
free exact completions and prove that every algebraic category is a free exact
completion of its full subcategory of all regular projectives. This will be used
in the next chapter to characterize algebraic categories among exact categories,
and to describe all finitary localizations of algebraic categories. The trouble with
regular projective objects in an algebraic category is namely that they are not
closed under finite limits. Luckily they have weak finite limits. Recall that weak
limits are defined as limits except that the uniqueness of the factorization is not
requested (see 16.7). The main point is that the universal property of a free
exact completion is based on left covering functors. These are functors which
play, for categories with weak finite limits, the role that functors preserving
finite limits play for finitely complete categories.
We will be concerned with regular epimorphisms (3.4) in an exact category
(3.16). For the comfort of the reader, we start by listing some of their (easy
but) important properties. In diagrams, regular epimorphisms are denoted by
// .

16.1 Lemma. Let A be an exact category.

1. Any morphism factorizes as a regular epimorphism followed by a monomor-

phism.

2. Consider a morphism f : X → Z. The following conditions are equivalent:

(a) f is a regular epimorphism;

(b) f is a strong epimorphism;
(c) f is an extremal epimorphism.

Proof. 1: Consider a morphism f : X → Z and its factorization through the

141
 CHAPTER 16. FREE EXACT CATEGORIES

coequalizer of its kernel pair

f1 f
N (f ) // X /Z
f2 ?

e 
  m
I
We have to prove that m is a monomorphism. For this, consider the following
diagram, where each square is a pullback

N (f ) // /X
q
e
 #  m1

/ / N (m) /I

m2 m
  
X //I /Z
e m

Since in A regular epimorphisms are pullback stable, the diagonal q is an epi-

morphism. Now, m1 · q = e · f1 = e · f2 = m2 · q, so that m1 = m2 . This means
that m is a monomorphism.
2: a ⇒ b: Let u, v and m be morphisms such that v · f = m · u. If f is the
coequalizer of a pair (x, y), and m a monomorphism, then u also coequalizes x
and y.
b ⇒ c: If f = m · u with m a monomorphism, we can write v · f = m · u with
v = id . By condition (b) m is a split epimorphism, but a monomorphism which
is also a split epimorphism is an isomorphism.
c ⇒ a: Just take a regular epi-mono factorization f = m · e (which exists by 1.);
if condition (c) holds, then m is an isomorphism and therefore f is a regular
epimorphism.
16.2 Corollary. Let A be an exact category.
1. The factorization stated in Lemma 16.1.1 is essentially unique;
2. The composite of two regular epimorphisms is a regular epimorphism;
3. If the triangle
f
X@ //Z
@@ ~~?
@@ ~
g @@ ~~
~~ h
A
commutes and f is a regular epimorphism, then h is a regular epimor-
phism;
4. If a morphism is a regular epimorphism and a monomorphism, then it is
an isomorphism.

November 20, 2009 142

CHAPTER 16. FREE EXACT CATEGORIES

In fact, everything follows easily from condition 2.b of Lemma 16.1.

16.3 Lemma. Every exact category has the following properties:

1. The product of two regular epimorphisms is a regular epimorphism;

2. Consider the following diagram

a1
A0 // A1
a2
f0 f1
 b1 
B0 // B1
b2

with f1 · ai = bi · f0 for i = 1, 2. If f0 is a regular epimorphism and f1 is

a monomorphism, then the unique extension to the equalizers is a regular
epimorphism;

3. Consider the following commutative diagram

a1 a2
A0 /Ao A1
f0 f f1
  
B0 /Bo B1
b1 b2

If f0 and f1 are regular epimorphisms and f is a monomorphism, then

the unique extension to the pullbacks is a regular epimorphism.

Proof. 1: Observe that f × id is the pullback of f along the suitable projec-

tion, and the same holds for id ×g. Now f × g = (f × id) · (id ×g).
2: Since f1 is a monomorphism, the pullback of the equalizer of (b1 , b2 ) along
f0 is the equalizer of (a1 , a2 ).
3: This follows from 1. and 2., using the usual construction of pullbacks via
products and equalizers.

For the sake of generality, let us point out that in 16.1, 16.2 and 16.3 we do
not need that in A equivalence relations are effective.

16.4 Remark. In 16.3.2 if f1 is any morphism (not necessarily a monomor-

phism) it is no longer true that the pullback of an equalizer e : E → B0 of (b1 , b2 )
along f0 is an equalizer of (a1 , a2 ). What remains true (in any category with
finite limits) is the following fact: let e′ : E ′ → A0 be a pullback of e along f0 ,
let k1 , k2 : N (f1 ) ⇉ A1 be a kernel pair of f1 , and n : E ′ → N (f1 ) the unique
morphism such that ki · n = ai · e′ (i = 1, 2). Then the following is a limit

November 20, 2009 143

 CHAPTER 16. FREE EXACT CATEGORIES

diagram
E′ E
} EE
e′ }}} EEn
}} EE
}~ } E"
A0 QQ a2 N (f1 )
QQQ
QQQ mmkm1mmm
a1 mQ k2
mmm QQQQQ
 vmmmmm QQ( 
A1 A1
We will use this fact in the proof of Theorem 16.24.

From Propositions 3.18 and 5.15 we know that an algebraic category is an

exact category having enough regular projective objects. In fact, each algebra
is a regular quotient of a regular projective algebra. In the following we study
categories having enough regular projectives; we introduce the concept of a
regular projective cover for a subcategory of regular projectives in case there is
“enough of them”.

16.5 Definition. Let A be a category. A regular projective cover of A is a full

and faithful functor I : P → A such that

1. for every object P of P, IP is regular projective in A;

2. for every object A of A, there exists an object P in P and a regular

epimorphism P → A (we write P instead of IP and we call P → A a
P-cover of A).

16.6 Definition. A functor is exact if it preserves finite limits and regular

epimorphisms.

The present chapter is devoted to the study of exact functors defined on

an exact category A having a regular projective cover P → A. First of all,
observe that regular projective objects are not closed under finite limits, so that
we cannot hope that P inherits finite limits from A. Nevertheless, a “trace” of
finite limits remains in P. In fact, P has weak finite limits.

16.7 Definition. A weak limit of a diagram D : D → A is a cone pX : W →

DX (X ∈ obj D) such that for every other cone aX : A → DX there exists a
morphism a : A → W such that pX · a = aX for all X.

Observe that, unlike limits, weak limits are very much “non-unique”. For
example, any non-empty set is a weak terminal object in the category Set.

16.8 Lemma. If P → A is a regular projective cover of a finitely complete

category A, then P has weak finite limits.

Proof. Consider a finite diagram D : D → P. If

h πX : L → DX iX∈D

November 20, 2009 144

CHAPTER 16. FREE EXACT CATEGORIES

is a limit of D in A, then we choose a P-cover l : P → L. The resulting cone

h πX · l : P → DX iX∈D

is a weak limit of D in P.

In the situation of the previous lemma, apply an exact functor G : A → B.

Since G preserves finite limits, the factorization of the cone

h G(πX · l) : GP → G(DX) iX∈D

through the limit in B is Gl : GP → GL, which is a regular epimorphism because

G is exact.
We can formalize this property in the following definition.

16.9 Definition. Let B be an exact category and let P be a category with

weak finite limits. A functor F : P → B is left covering if, for any finite diagram
D : D → P with weak limit W, the canonical comparison morphism F W →
lim F · D is a regular epimorphism.

16.10 Remark. To avoid any ambiguity in the previous definition, let us point
out that if the comparison w : F W → lim F · D is a regular epimorphism for
a certain weak limit W of D, then the comparison w′ : F W ′ → lim F · D is
a regular epimorphism for any other weak limit W ′ of D. This follows from
Corollary 16.2 because w factorizes through w′ .

16.11 Example.

1. If a finite diagram D : D → A has a limit L, then the weak limits of D

are precisely the objects W such that L is a retract of W. Therefore any
functor preserving finite limits is left covering.

2. If P → A is a regular projective cover of an exact category A, then it is a

left covering functor.

3. The composition of a left covering functor with an exact functor is a left

covering functor.

16.12 Example. Let P be a category with weak finite limits, and consider
the (possibly illegitimate) functor category [P op , Set]. The canonical Yoneda
embedding YP op : P → [P op , Set] is a left covering functor.

Proof. Consider a finite diagram D : D → P in P, a weak limit W of D and

a limit L of YP op · D. The canonical comparison τ : YP op (W ) → L is a regular
epimorphism whenever, for all Z ∈ P, τZ : YP op (W )(Z) → LZ is surjective.
Since the limit L is computed pointwise in Set, an element of LZ is a cone from
Z to L, so that the surjectivity of τZ is just the weak universal property of
W.

November 20, 2009 145

 CHAPTER 16. FREE EXACT CATEGORIES

16.13 Remark. In the main result of this chapter (16.24) we show that an
exact category with enough regular projective objects is a free exact completion
of any of its regular projective covers. This is one of the results that request
working with the left covering property (instead of the seemingly more natural
condition of preservation of weak finite limits). In fact, the basic example 16.11.2
would not be true otherwise. This can be illustrated by the category of rings:
the inclusion of the full subcategory P of all regular projective rings does not
preserve weak finite limits. For example, the ring Z of integers is a weak terminal
object in P, but it is not a weak terminal object in A because the unique
morphism from Z to the one-element ring does not have a section.
A remarkable fact about left covering functors is that they classify exact
functors. Before stating this in a precise way, see 16.24, we need some facts about
left covering functors and pseudoequivalences. A pseudoequivalence is defined
“almost” as an equivalence relation, but (a) using a weak pullback instead of
a pullback to express the transitivity, and (b) without the assumption that the
graph be jointly monic.
16.14 Definition. Let P be a category with weak pullbacks. A pseudoequiva-
lence is a parallel pair
x1
/
X′ x / X
2

which is
1. reflexive, i.e., there exists r : X → X ′ such that x1 · r = idX = x2 · r,
2. symmetric, i.e., there exists s : X ′ → X ′ such that x1 · s = x2 and x2 · s =
x1 ,
and
3. transitive, i.e., in an arbitrary weak pullback

x′1
P / X′

x′2 x2
 
X′ /X
x1

there exists t : P → X ′ such that x1 · t = x1 · x′1 and x2 · t = x2 · x′2 . The

morphism t is called a transitivity morphism of x1 and x2 .
16.15 Remark.
1. Observe that the existence of a transitivity morphism of x1 and x2 does
not depend on the choice of a weak pullback of x1 and x2 .
2. Recall that a regular factorization of a morphism is a factorization as a
regular epimorphism followed by a monomorphism. In a category with

November 20, 2009 146

CHAPTER 16. FREE EXACT CATEGORIES

binary products, we speak about regular factorization of a parallel pair

p, q : A ⇉ B. What we mean is a factorization of (p, q) as in the follow-
ing diagram, where e is a regular epimorphism and (p′ , q ′ ) is a jointly
monomorphic parallel pair
p
A NNN q // B
O O
NNN
NNN p′ q′
e NNN
N& &
I
obtained by a regular factorization of hp, qi : A → B × B. Since jointly
monomorphic parallel pairs are also called relations, we call (p′ , q ′ ) the
relation induced by (p, q).
3. If P has finite limits, then equivalence relations are precisely those parallel
pairs which are, at the same time, relations and pseudoequivalences. The
next result, which is the main link between pseudoequivalences and left
covering functors, shows that any pseudoequivalence in an exact category
is a composition of an equivalence relation with a regular epimorphism.
(The converse is not true: if we compose an equivalence relation with
a regular epimorphism in general we do not obtain a pseudoequivalence.
Consider the category of rings, the unique equivalence relation on the one-
element ring 0, and the unique morphism Z → 0. The parallel pair Z ⇉ 0
is not reflexive, because there are no morphisms from 0 to Z.)
16.16 Lemma. Let F : P → B be a left covering functor. For every pseudoe-
quivalence x1 , x2 : X ′ ⇉ X in P, the relation in B induced by (F x1 , F x2 ) is an
equivalence relation.
Proof. Consider a regular factorization in B
F x1
F X ′ PP // FO XO
PPPF x2
PPP i i2
p PPP 1
PPP'
'
I
Since the reflexivity and transitivity are obvious, we only check the transitivity
of (i1 , i2 ). The pullback of i1 · p and i2 · p factorizes through the pullback of i1
and i2 , and the factorization, v, is a regular epimorphism (because p is a regular
epimorphism and B is an exact category):
j1
WC / F X′
CC
CCv p
CC
C! ! i′1

j2 Q /I

i′2 i2
  
F X′ //I / FX
p i1

November 20, 2009 147

 CHAPTER 16. FREE EXACT CATEGORIES

Consider also a transitivity morphism t : P → X ′ of (x1 , x2 ) as in Definition

16.14. Since F : P → B is left covering, the factorization q : F P → W such that
j1 · q = F x′1 and j2 · q = F x′2 is a regular epimorphism. Finally, we have the
following commutative diagram

v·q
FP //Q
τ
p·F t hi1 ·i′1 ,i2 ·i′2 i
 v 
I / FX × FX
hi1 ,i2 i

Since v · q is a regular epimorphism and hi1 , i2 i is a monomorphism, there exists

τ : Q → I such that hi1 , i2 i · τ = hi1 · i′1 , i2 · i′2 i. This implies that (i1 , i2 ) is
transitive.

16.17 Remark. Generalizing the fact that functors preserve finite limits iff
they preserve finite products and equalizers, we are going to prove the same
for left coverings. We use the phrase “left covering with respect to weak finite
products” for the restriction of 16.9 to discrete categories D. Observe that this is
equivalent to being left covering with respect to weak binary products and weak
terminal objects. Analogously, we use “left covering with respect to equalizers”.

16.18 Lemma. A functor F : P → B, where P has weak finite limits and B is

exact, is left covering if and only if it is left covering with respect to weak finite
products and weak equalizers.

Proof. 1. Using Lemma 16.3 and working by induction, one extends the
left covering character of F to joint equalizers of parallel n-tuples, and then to
multiple pullbacks.
2. Consider a finite diagram D : D → P. We can construct a weak limit of D
using a weak product ΠX∈D DX, weak equalizers Ed , one for each morphism
d : X → X ′ in D, and a weak multiple pullback E as in the following diagram

Ed J DX ′
e′d ~~> JJJ Dd·π ss ssss9 9
e
JJJd X
~~ sssss
~~~ JJJ
$ s sssssssπX ′
~ s
E@ .. ΠX∈D DX ..
@@ . t: KKKKKKKDc·π .
@@ tt KKKKKK Y
@ tt KKKKKK
e′c @@ ttt ec π Y′
tt K% %
Ec DY ′

November 20, 2009 148

CHAPTER 16. FREE EXACT CATEGORIES

Perform the same constructions in B to get limits as in the following diagrams

Ld F DX ′
l′d ?
KK
KK l F Dd·π̃ q qqqqq8 8
 KKd X qq q

KK qqqqqq
 KK
% qqqqqqq π̃X ′
L? .. ΠX∈D F DX ..
?? . s9 MMMMMMMF Dc·π̃ .
?? ss MMMMMM Y
? ss MMMMMM
l′c ?? sss lc π̃ MMM& &
Y′
ss
Lc F DY ′

? Sd LLLL p 7 7 DX ′
ppF
s′d  LLLsd F Dd·F π pp p
 pXpppp
 LLL
L p pppppppFpπ ′
 % pp X

S? .. F (Π X∈D DX) ..
?? . r 9 NNNNNNF Dc·F πY .
N
?? rrr NNNNNN
?
s′c ?? rrrrrsc F π
NNNNNN
NNN' '
 Y′
rr
Sc F DY ′
By assumption, the canonical factorization qd : F Ed → Sd is a regular epimor-
phism. By Lemma 16.3, this gives rise to a regular epimorphism q : Q → S,
where Q is the multiple pullback of the F ed . By part 1., the canonical factor-
ization t : F E → Q is a regular epimorphism. Finally, a diagram chase shows
that the pullback of ld · ld′ along the canonical factorization p : F (ΠX∈D DX) →
ΠX∈D F DX is sd · s′d . By part 1., p is a regular epimorphism, so that we get a
regular epimorphism p′ : S → L. The regular epimorphism p′ · q · t : F E → Q →
S → L shows that F is left covering.
16.19 Lemma. A left covering functor F : P → B preserves finite jointly mono-
morphic sources.
Proof. A family of morphisms (fi : A → Ai )i∈I is jointly monomorphic if and
only if the span formed by idA , idA is a limit of the corresponding diagram:

id/A x / FA id /
A 00 L 44 FA F A4
00 4 44
id fi 00 y F fi 44 id F fi 44
  00   44   44
fi F fi 44F fj F fi 44F fj
A OOO / Ai 00fj F A RRR / F Ai 44 F A RRR / F A i 4
OOO . 00 RRR R RRR . . 44
OOO . . 0 RRR . . . 44 R
R . 44
O 0 RRR RR
fj OOOO 0
 F fj RRR 44 F fj RRRR 4
O' R ) R) 
Aj F Aj F Aj

Now apply F and consider the canonical factorization q : F A → L, where L is a

limit in B of the corresponding diagram. By assumption, q is a regular epimor-
phism. It is also a monomorphism, because x·q = id, and so it is an isomorphism.
This implies that idF A , idF A is a limit, thus the family (F fi : F A → F Ai )i∈I is
jointly monomorphic.

November 20, 2009 149

 CHAPTER 16. FREE EXACT CATEGORIES

16.20 Lemma. Consider a functor F : P → B. Assume that P has finite limits

and B is exact. Then F is left covering if and only if it preserves finite limits.
Proof. One implication is clear, see 16.11.1. Thus, let us assume that F
is left covering and consider a finite non-empty diagram D : D → P. Let
(πX : L → DX)X∈D be a limit of D and (π̃X : L̃ → F DX)X∈D a limit of
F · D. Since the family (πX )X∈D is jointly monomorphic, by Lemma 16.19 also
the family (F πX )X∈D is monomorphic. This implies that the canonical factor-
ization p : F L → L̃ is a monomorphism. But it is a regular epimorphism by
assumption, so that it is an isomorphism.
The argument for the terminal object T is different. In P, the product of T
with itself is T with the identity morphisms as projections. Then the canonical
factorization F T → F T × F T is a (regular) epimorphism. This implies that the
two projections π1 , π2 : F T × F T ⇉ F T are equal. But the pair (π1 , π2 ) is the
kernel pair of the unique morphism q to the terminal object of B, so that q is
a monomorphism. Since F is left covering, q is a regular epimorphism, thus an
isomorphism.

Let us point out that in 16.16 and 16.18 we do not need to assume that
equivalence relations are effective in B. Moreover, if in Definition 16.9 we replace
regular epimorphism by strong epimorphism, then 16.19 and 16.20 hold for all
categories B with finite limits.
16.21 Definition. Let P be a category with weak finite limits. A free exact
completion of P is an exact category Pex with a left covering functor

Γ : P → Pex

such that for every exact category B and for every left covering functor F : P →
B, there exists an essentially unique exact functor F̂ : Pex → B with F̂ · Γ
naturally isomorphic to F.
Note that, since a free exact completion is defined via a universal property,
it is determined uniquely up to equivalence.
16.22 Remark. Since the composition of the left covering functor Γ : P → Pex
with an exact functor Pex → B clearly gives a left covering functor P → B, the
previous universal property can be restated in the following way: Composition
with Γ induces an equivalence functor

− · Γ : Ex[Pex , B] → Lco[P, B]

where Ex[Pex , B] is the category of exact functors from Pex to B and natural
transformations, and Lco[P, B] is the category of left covering functors from P
to B and natural transformations.
16.23 Remark. In order to prepare the proof of Theorem 16.24, let us explain
how an exact category with enough regular projective objects can be recon-
structed using any of its regular projective covers. Let P → A be a regular

November 20, 2009 150

CHAPTER 16. FREE EXACT CATEGORIES

projective cover of an exact category A. Fix an object A in A and consider

a P-cover a : X → A, its kernel pair a1 , a2 : N (a) ⇉ X, and again a P-cover
x : X ′ → N (a). In the resulting diagram
a1 ·x a
X′ // X //A
a2 ·x

the left-hand part is a pseudoequivalence in P (not in A !) and A is its coequal-

izer. Consider a morphism ϕ : A → B in A and the following diagram

f′
X′ // > Z ′
g′

Σ
a1 ·x a2 ·x b1 ·z b2 ·z

  f
//  
X g
Z

a b

 
ϕ
A /B

Using the regular projectivity of X and X ′ and the universal property of the
kernel pair of b, we get a pair (f ′ , f ) such that ϕ · a = b · f and f · ai · x = bi · z · f ′
for i = 1, 2. Conversely, a pair (f ′ , f ) such that f · ai · x = bi · z · f ′ for i = 1, 2
induces a unique extension to the quotient. Moreover, two such pairs (f ′ , f ) and
(g ′ , g) have the same extension if and only if there is a morphism Σ : X → Z ′
such that b2 · z · Σ = f and b2 · z · Σ = g.
16.24 Theorem. Let I : P → A be a regular projective cover of an exact cate-
gory A. Then A is a free exact completion of P.
Proof. 1. Extension of a left covering functor F : P → B to a functor F̂ : A →
B. To define F̂ on objects A ∈ A, construct the coequalizer
x1 =a1 ·x a
// //A
X′ x2 =a2 ·x
X

as in 16.23. By 16.16, the relation (i1 , i2 ) induced by F x1 , F x2 : F X ′ ⇉ F X

in B is an equivalence relation. Since B is exact, we can define F̂ A to be a
coequalizer of (i1 , i2 )
F x1 α
F X ′ OO // FO X
O
/ / F̂ A
OOOF x2
OOO i i2
p OOOO 1
OO' '
R

November 20, 2009 151

 CHAPTER 16. FREE EXACT CATEGORIES

Let now ϕ : A → B be a morphism in A. We can costruct a pair f : X → Z,

f ′ : X ′ → Z ′ as in 16.23 and define F̂ ϕ to be the unique extension to the
quotients as in the following diagram
F f′
F X′ / F Z′

F x1 F x2 F z1 F z2
   
FX / FZ
Ff
α β
 
F̂ A / F̂ B
F̂ ϕ

The discussion in 16.23 shows that this definition does not depend on the choice
of the pair f, f ′ . The preservation of composition and identity morphisms by F̂
comes from the uniqueness of the extension to the quotients. It is clear that
F̂ · I is naturally isomorphic to F and that a different choice of P-covers X and
X ′ for a given object A ∈ A produces a functor naturally isomorphic to F̂ .
2. F̂ is the essentially unique exact functor such that F̂ ·I is naturally isomorphic
to F. Indeed, using once again the notations of 16.23, (F̂ a1 , F̂ a2 ) is the kernel
pair of F̂ a, and F̂ x and F̂ a are regular epimorphisms:

F̂ x F̂ a1 F̂ a
F X ′ ≃ F̂ X ′ / / F̂ N (a) // F̂ X ≃ F X / / F̂ A
F̂ a2

This implies that F̂ A is necessarily a coequalizer of (F x1 , F x2 ). In a similar way

one shows that F̂ is uniquely determined on morphisms.
3. The extension F̂ : A → B preserves finite limits. In fact, it is sufficient to
show that F̂ is left covering with respect to the terminal object, binary products
and equalizers of pairs, see Lemma 16.18 and Lemma 16.20.
3a. Products: let A and B be objects in A. Working as in 16.23 we get coequal-
izers
x1 a / / z1 b / /
X′ // X A and Z ′
//
Z B
x2 z2

Consider the following diagram, where both lines are products in A and c : R →
X × Z is a P-cover
R
c

πX πZ
Xo X ×Z /Z

a a×b b
  
Ao πA A×B πB
/B

By 16.3.1 a × b is a regular epimorphism, so that (a × b) · c : R → A × B is a

P-cover. Moreover, by 16.8 (R, πX · c, πZ · c) is a weak product of X and Z in

November 20, 2009 152

CHAPTER 16. FREE EXACT CATEGORIES

P. Applying F̂ , we have the following diagram in B

F (πX ·c) F (πZ ·c)

FX o FR / FZ

α γ β
  
F̂ A o F̂ πA
F̂ (A × B)
F̂ πB
/ F̂ B

from which we get the following commutative square

hF (πX ·c),F (πZ ·c)i

FR / / FX × FZ

γ α×β
 
F̂ (A × B) / F̂ A × F̂ B
hF̂ πA ,F̂ πB i

The top morphism is a regular epimorphism because F is left covering, and the
right-hand morphism is a regular epimorphism by 16.3.3, so that the bottom
morphism also is a regular epimorphism, as requested.
3b. Equalizers: let

e ϕ
E /A // B
ψ

be an equalizer in A. The idea is once again to construct a P-cover of E using

P-covers of A and B. For that, consider the following diagram:

R ′
m mmm6 Z
n′ mmm
c
mm mm z
 mm m 
m
E ′′ m6
N (b)
n mmmmm
z′ mmm b1 b2
 mmmmm f  
m
E′ ′
/X // Z
e g
a′ a b
  ϕ 
E e
/A // B
ψ

where a : X → A and b : Z → B are P-covers, E ′ is a pullback of a and e, the

equations b · f = ϕ · a and b · g = ψ · a hold, N (b) is a kernel pair of b, n is the
unique morphism such that b1 · n = f · e′ and b2 · n = g · e′ , E ′′ is a pullback of
z and n, and z : Z ′ → N (b) and c : R → E ′′ are P-covers. From 16.4 we know

November 20, 2009 153

 CHAPTER 16. FREE EXACT CATEGORIES

that the following is a limit diagram in A

E′ D
~ DD
e′ ~~~ DDn
~~ DD
~ ~ D"
X QQQg N (b)
QQQ b1 mm
QQmQmmmm
f b2
mmm QQQQQ
 vmmmmm QQ( 
Z Z

Clearly, it remains a limit diagram if we paste it with the pullback E ′′ and, by

16.8, we get a weak limit in P by covering it with c : R → E ′′

R
~ AAA ′
e′ ·z ′ ·c
~~ AAn ·c
~~~ AA
~
X PPPg Z′
PPP bn1 n·znnn
PnPn
f
nnn PPPPP b2 ·z
 wnnnn P' 
Z Z

Consider the following diagram in B :

6 F Z′
lllll
′
mlll
q
l llll 
l ll
′′ N (β)
=V lll6
k m lll
q′ lll β1 β2
 llllll Ff  
l / FX //
FR V′ v′ FZ
Fg
γ α ′ α β
   F̂ ϕ

h v
F̂ E /V /
6 F̂ A // F̂ B
F̂ ψ
F̂ e

where V is an equalizer of F̂ ϕ and F̂ ψ, α : F X → F̂ A, β : F Z → F̂ B and

γ : F R → F̂ E are the coequalizers defining F̂ A, F̂ B and F̂ E (as explained in
the first part of the proof), V ′ is a pullback of α and v, N (β) is a kernel pair of
β, m is the unique morphism such that β1 · m = F f · v ′ and β2 · m = F g · v ′ , q
is the unique morphism such that βi · q = F (bi · z) (i = 1, 2), V ′′ is a pullback
of q and m, and h : F̂ E → V is the unique morphism such that v · h = F̂ e. We
have to prove that h is a regular epimorphism. By 16.4 the following is a limit

November 20, 2009 154

CHAPTER 16. FREE EXACT CATEGORIES

diagram:
V ′′ E
′ ′
v ·q zz
z EE ′
z EEm
zz EE
|zz E"
F X RR F g F (b1 ·z) l F Z ′
RRR l
RRRlllll
Ff l l R RRRR F (b2 ·z)
ll
 vllll RR( 
FZ FZ
Since F is left covering, the unique morphism k : F R → V ′′ such that m′ · k =
F (n′ ·c) and v ′ ·q ′ ·k = F (e′ ·z ′ ·c) is a regular epimorphism. As B is exact and F
is left covering, by 16.16 (β1 , β2 ) is the equivalence relation induced by (z1 , z2 ),
so that q is a regular epimorphism, and then q ′ also is a regular epimorphism.
Since α′ also is a regular epimorphism, it remains to check that α′ · q ′ · k = h · γ.
By composing with the monomorphism v, this is an easy diagram chasing.
3c. Terminal object: let 1A and 1B be terminal objects of A and B, and T → 1A
a P-cover. Applying F̂ we get a commutative diagram

FT / F̂ 1
A
zz
zz
zz
 z| z
1B

Since T is weak terminal in P and F is left covering, F T → 1B is a regular

epimorphism. Therefore, F̂ 1A → 1B also is a regular epimorphism.
4. The extension F̂ : A → B preserves regular epimorphisms. This is obvious:
if ϕ : A → B is a regular epimorphism in A and a : X → A is a P-cover, we can
choose as a P-cover of B the morphism ϕ · a : X → B. Applying F̂ we get a
commutative diagram
id /
FX FX
α β
 
F̂ A / F̂ B
F̂ ϕ

which shows that F̂ ϕ is a regular epimorphism.

16.25 Corollary.
1. Let A and B be exact categories and I : P → A a regular projective cover.
Given exact functors G, G′ : A ⇉ B with G · I ≃ G′ · I, then G ≃ G′ .
2. Let A and A′ be exact categories, P → A a regular projective cover of A
and P ′ → A′ one of A′ . Any equivalence P ≃ P ′ extends to an equivalence
A ≃ A′ .
16.26 Remark. For later use, let us point out a simple consequence of the
previous theorem. Consider the free exact completion I : P → A as in 16.24 and

November 20, 2009 155

 CHAPTER 16. FREE EXACT CATEGORIES

a functor K : A → B, with B exact. If K preserves coequalizers of equivalence

relations and K · Γ is left covering, then K is exact.
16.27 Corollary. Let A be an algebraic category and P its full subcategory of
regular projective objects. The inclusion I : P → A is a free exact completion of
P.
In fact, this follows from Theorem 16.24 because A is exact (see 3.18) and P is
a projective cover of A (see 5.15).

Historical Remarks for Chapter 16

Following a suggestion of A. Joyal, the exact completion of a category with

finite limits was presented by A. Carboni and R. Celia Magno in [32]. The more
general approach working with categories with weak finite limits is due to A.
Carboni and E. M. Vitale [33].
The connection between the exact completion and the homotopy category
of topological spaces (see 17.4) was established by M. Gran and E. M. Vitale in
[49].

November 20, 2009 156

Chapter 17

Exact completion and

reflexive-coequalizer
completion

The present chapter is devoted to elementary constructions of two free com-

pletions described in Chapters 16 and 7, respectively, in a different manner: a
free exact completion (of categories with weak finite limits) and a free reflexive-
coequalizer completion (of categories with finite coproducts). The reader may
decide to skip this chapter without losing the connection to other chapters.

In Chapter 17 we have seen that if I : P → A is a regular projective cover

of an exact category A, then P has weak finite limits and I is a free exact
completion of P. We complete here the study of the exact completion showing
that for any category P with weak finite limits it is possible to construct a free
exact completion Γ : P → Pex . Moreover, Γ is a regular projective cover of Pex .
The following construction of Pex is suggested by 16.23.
17.1 Definition. Given a category P with weak finite limits, we define the
category Pex as follows:
1. Objects of Pex are pseudoequivalences x1 , x2 : X ′ ⇉ X in P (we sometimes
denote such an object by X/X ′).
2. A premorphism in Pex is a pair of morphisms (f ′ , f ) as in the diagram

f′
X′ / Z′
x1 x2 z1 z2
   
X /Z
f

such that f · x1 = z1 · f ′ and f · x2 = z2 · f ′ .

157
 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

3. A morphism in Pex is an equivalence class [f ′ , f ] : X/X ′ → Z/Z ′ of premor-

phisms, where two parallel premorphisms (f ′ , f ) and (g ′ , g) are equivalent
if there exists a morphism Σ : X → Z ′ such that z1 · Σ = f and z2 · Σ = g.
4. Composition and identities are obvious.
17.2 Notation. We denote by Γ : P →Pex the embedding of P into Pex assign-
ing to a morphism f : X → Z the following morphism
f
X /Z

id id id id
   
X /Z
f

17.3 Remark.
1. The fact that the above relation among premorphisms is an equivalence
relation can be proved (step by step) using the assumption that the
codomain z1 , z2 : Z ′ ⇉ Z is a pseudoequivalence. Observe also that the
class of (f ′ , f ) depends on f only (compose f with a reflexivity morphism
of (z1 , z2 ) to show that (f ′ , f ) and (f ′′ , f ) are equivalent); for this reason,
we often write [f ] instead of [f ′ , f ].
2. The fact that composition is well-defined is obvious.
3. Γ is a full and faithful functor. This is easy to verify.
4. Observe that if P is small (or locally small), then Pex also is small (or
locally small, respectively).
17.4 Remark. The above equivalence relation among premorphisms in Pex can
be thought of as a kind of “homotopy” relation. And in fact, this is the case in a
particular example: let X be a topological space and X [0,1] the space of continu-
ous maps from the interval [0, 1] to X; the evaluation maps ev0 , ev1 : X [0,1] ⇉ X
constitute a pseudoequivalence. This gives rise to a functor E : Top → Topex .
Now two continuous maps f, g : X → Z are homotopic in the usual sense pre-
cisely when E(f ) and E(g) are equivalent in the sense of Definition 17.1. More
precisely, E factorizes through the homotopy category, and the factorization
E ′ : HTop → Topex is full and faithful (and left covering).
We are going to prove that the above category Pex is exact and the func-
tor Γ : P → Pex is a regular projective cover. For this, it is useful to have
an equivalent description of Pex as a full subcategory of the functor category
[P op , Set].
17.5 Lemma. Let P be a category with weak finite limits, and let
YP op : P → [P op , Set]
be the Yoneda embedding. The following properties of a functor A : P op → Set
are equivalent:

November 20, 2009 158

CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

1. A is a regular quotient of a representable object modulo a pseudoequiva-

lence in P, i.e., there exists a pseudoequivalence x1 , x2 : X ′ ⇉ X in P and
a coequalizer
x1
YP op (X ′ ) // YP op (X) //A
x 2

in [P op , Set];

2. A is a regular quotient of a representable object modulo a regular epimor-

phism a : YP op (X) → A such that N (a), the domain of a kernel pair of a,
is also a regular quotient of a representable object:

x a1 a
YP op (X ′ ) / / N (a) // Y op (X) //A (for some X ′ in P).
P
a2

Proof. Consider the diagram displayed in 2. Since a is the coequalizer of

(a1 · x, a2 · x), we have to prove that (a1 · x, a2 · x) is a pseudoequivalence in P.
Let us check the transitivity: consider the following diagram

YP op (W ) XXX
11 LLL XXXXXX
11 LLLv XXXXX x′1
LLL XXXXX
11 XXXXX
11 LL& XXXXX
u1 X,
11 P JJ′ / YP op (X ′ )
11 JJ
1 JJu
JJ x
x′2 11 JJ 
11 J$ ′
a1
11 u2 P / N (a)
11
11 a′2 a2
1   
YP op (X ′ ) x / / N (a) a1
/ YP op (X)

where P and P ′ are pullbacks, and W is a weak pullback. Since YP op (W ) is

regular projective and x is a regular epimorphism, the transitivity morphism
t : P → N (a) of (a1 , a2 ) extends to a morphism t′ : YP op (W ) → YP op (X ′ ) such
that t · u · v = x · t′ . This morphism t′ is a transitivity morphism for (a1 · x, a2 · x).
The converse implication follows from Lemma 16.16, since YP op : P → [P op , Set]
is left covering (see 16.12).

Note that the fact that (x1 , x2 ) is a pseudoequivalence in P does not mean
that (YP op (x1 ), YP op (x2 )) is a pseudoequivalence in [P op , Set] because YP op does
not preserve weak pullbacks.

17.6 Remark. The full subcategory of [P op , Set] of all objects satisfying 1. or

2. of the above lemma is denoted by P ′ ex . In the next lemma, the codomain re-
′
striction of the Yoneda embedding YP op : P → [P op , Set] to Pex is again denoted
by YP , and Γ is the functor from 17.2.
op

November 20, 2009 159

 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

′
17.7 Lemma. There exists an equivalence of categories E : Pex → Pex such that
E · Γ = YP .
op

′
Proof. Consider the functor E : Pex → Pex sending a morphism [f ] : X/X ′ →
′
Z/Z to the corresponding morphism ϕ between the coequalizers as in the fol-
lowing diagram
f′
YP op (X ′ ) / YP op (Z ′ )

x1 x2 z1 z2
   
YP op (X) / YP op (Z)
f
a b
 
A /B
ϕ

The functor E is well-defined because a is an epimorphism and b coequalizes y0

′
and y1 . Moreover, E is essentially surjective by definition of Pex . Let us prove
that E is faithful: if E[f ] = E[g], then the pair (f, g) factorizes through the kernel
pair N (b) of b, which is a regular factorization of (y0 , y1 ). Since YP op (X) is regular
projective, this factorization extends to a morphism YP op (X) → YP op (Z ′ ), which
shows that [f ] = [g].
E is full: given ϕ : A → B, we get f : YP op (X) → YP op (Z) by regular projectivity
of YP op (X). Since b · f · x1 = b · f · x2 , we get f : YP op (X ′ ) → N (b). Since N (b) is
the regular factorization of (z1 , z2 ) and YP op (X ′ ) is regular projective, f extends
to f ′ : YP op (X ′ ) → YP op (Z ′ ). Clearly, E[f ′ , f ] = ϕ.
17.8 Proposition. For every category P with weak finite limits, the functor

Γ : P → Pex

of 17.2 is a left covering functor into an exact category. Moreover, this is a

regular projective cover of Pex .
Proof. 1. Pex has finite limits. Since the construction of the other basic
types of finite limits is completely analogous, we explain in details the case of
equalizers, mentioning the construction for binary products and terminal object
just briefly.
1a. Equalizers: Consider a parallel pair in Pex together with what we want to
be their equalizer

e′ f′
E′ / X′ // Z ′
g′
e1 e2 x1 x2 z1 z2
    f
//  
E /X Z
e g

This means that we need the following equations: x1 ·e′ = e·e1 and x2 ·e′ = e·e2 .
Moreover, we request f ·e and g ·e being equivalent, that is, we need a morphism

November 20, 2009 160

CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

ϕ : E → Z ′ such that z1 · ϕ = f · e and z2 · ϕ = g · e. Let us take E and E ′ to

be the following weak limits
e1 e2
EA Eo E′ /E
~ A
e ~~~ AAϕ
~~ AA e e′ e
~~~ A   
′ /X
X PPPg z1 nn Z Xo x1 X′ x2
PPP nnn
f
PnPn z2
nnn PPPPP
 wnnnn P' 
Z Z

(i) It is straightforward to check that (e1 , e2 ) is a pseudoequivalence in P

(just use the fact that (x1 , x2 ) is a pseudoequivalence).
(ii) To show that [e] equalizes [f ] and [g], use the morphism ϕ : E → Z ′ .
(iii) The morphism [e] is a monomorphism: in fact, consider two morphisms
in Pex
h′
A′ // E ′
′
k
a1 a2 e1 e2
  h //  
A E
k

such that [e] · [h] = [e] · [k]. This means that there is a morphism Σ : A → X ′
such that x1 · Σ = e · h and x2 · Σ = e · k. By the weak universal property of
E ′ , we have a morphism Σ′ : A → E ′ such that e1 · Σ′ = h and e2 · Σ′ = k. This
means that [h] = [k].
(iv) We prove that every morphism
h′ / X′
A′
a1 a2 x1 x2
   
A /X
h

in Pex such that [f ] · [h] = [g] · [h] factorizes through [e]. We know that there is
Σ : A → Z ′ such that z1 ·Σ = f ·h and z2 ·Σ = g ·h. The weak universal property
of E yields then a morphism k : A → E such that e · k = h and Σ · k = ϕ. Now,
x1 · h′ = e · k · a1 and x2 · h′ = e · k · a2 . The weak universal property of E ′ yields
a morphism k ′ : A′ → E ′ such that e1 · k ′ = k · a1 and e2 · k ′ = k · a2 . Finally,
the needed factorization is [k ′ , k] : A/A′ → E/E ′ .
1b. Products: Consider two objects x1 , x2 : X ′ ⇉ X and z1 , z2 : Z ′ ⇉ Z in Pex .
Their product is given by
x′ z′
X′ o P′ / Z′
x1 x2 p1 p2 z1 z2
     
Xo x P z
/Z

November 20, 2009 161

 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

where
x z
Xo P /Z

is a weak product of X and Z in P, and P ′ is the following weak limit

′
P P
nnn} @@P@PPPP ′
x nnnn }}}
′
@@ PzP
n nnnn }}} p1 p2 @@@
PPP
PPP
v
n n ~} ' ′
X′ B P PPP z xnnnn
P AA } Z
BB ||x PPP
B|B| PPnPnnn z AA}}
x1 A z
||| BBBx2 nnnnn PPPP z1 }}}AAA 2
 ~| wnn PP' ~
} 
X X Z Z

1c. Terminal object: For any object T of P, the projections from a weak product
π1 , π2 : T × T ⇉ T form a pseudoequivalence. If T is a weak terminal object in
P, then (π0 , π1 ) is a terminal object in Pex .
2. Pex is closed under finite limits in [P op , Set]. In fact, by Lemma 17.7, we
′
can identify Pex with Pex . We prove that the full inclusion of Pex into [P op , Set]
preserves finite limits. Because of Lemma 16.20, it is enough to prove that
the inclusion is left covering. We give the argument for equalizers, since that
for products and terminal object is similar (and easier). With the notations of
part 1., consider the following diagram, where ǫ, α and β are extensions to the
coequalizers, the triangle on the right is a regular factorization, and the triangle
at the bottom is the factorization through the equalizer

e′ f′
YP op (E ′ ) / YP op (X ′ ) // YP op (Z ′ )
g′ II
II z
II
II
    f   b1
I$ $
e / YP op (X) // Y op (Z) oo
YP op (E) P N (b)
g b2
c a b
  
ǫ /k5 A
α // 
C B
kkk β
′ kkkkk
ǫ kkkkl
 kkkkkk
L

We have to prove that ǫ′ is a regular epimorphism. Using ϕ : E → Z ′ , we check

that α · a · e = β · a · e, so that there is p : YP op (E) → L such that l · p = a · e, and
then p = ǫ′ · c. So, it is enough to prove that p is a regular epimorphism, that
is, the components p(P ) : YP op (E)(P ) → L(P ) are surjective. This means that,
given a morphism u : YP op (P ) → A such that α · u = β · u, we need a morphism
û : P → E with l · p · û = u. First of all, observe that, since a is a regular
epimorphism and YP op (P ) is regular projective, there is u′ : P → X such that
a · u′ = u. Now, b · f · u′ = b · g · u′, so that there is u′′ : YP op (P ) → N (b) such that
b1 · u′′ = f · u′ and b2 · u′′ = g · u′ . Moreover, since z is a regular epimorphism

November 20, 2009 162

CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

and YP op (P ) is regular projective, there is ũ : P → Z ′ with z · ũ = u′′ . Finally,

z1 · ũ = f · u′ and z2 · ũ = g · u′ , so that there is û : P → E such that ϕ · û = ũ
and e · û = u′ . This last equation implies that l · p · û = u.
3. Pex is closed in [P op , Set] under coequalizers of equivalence relations. In fact,
consider an equivalence relation in Pex , with its coequalizer in [P op , Set]
α // c
B A //C
β

We have to prove that C lies in Pex . For this, consider the following diagram:

K / B ′′ / YP op (X)

a
  
/B α /A
B′
β c
  
YP op (X) //A //C
a c

with each square except, possibly, the right-hand bottom one, a pullback. The
remaining square is, then, also a pullback because [P op , Set] is exact, X ∈ P
and a is a regular epimorphism. Since A, B and YP op (X) are in Pex , which is
closed in [P op , Set] under finite limits (see part 2.), also K lies in Pex . So, K is
a regular quotient of a representable object. But K is also the kernel pair of
the regular epimorphism c · a : YP op (X) → C. By Lemma 17.5, this means that
C is in Pex .
17.9 Corollary. For every category P with weak finite limits, the functor
Γ : P → Pex
of 17.2 is a free exact completion of P.
In fact, this follows from 16.24 and 17.8.
17.10 Remark. Let A be an algebraic category. From Chapter 7 we know that
there are equivalences
A ≃ Ind (Af p ) and Af p ≃ Rec (App )
where App and Af p are the full subcategories of A of perfectly presentable
objects and of finitely presentable objects, respectively, and Rec is the free
completion under finite colimits conservative with respect to finite coproducts
(see 17.11). An analogous situation holds with the exact completion. In fact,
there are equivalences
A ≃ (Arp )ex ≃ (FCSum (App ))ex and Arp ≃ Ic (FCSum (App ))
where Arp is the full subcategory of regular projective objects and FCSum is the
free completion under coproducts conservative with respect to finite coproducts.

November 20, 2009 163

 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

The second part of this chapter is devoted to an elementary description of

the free completion under reflexive coequalizers

ERec : C → Rec C

of a category C with finite coproducts, already studied in Chapter 7. Let us

start by observing that 7.5 can be restated as follows.

17.11 Proposition. Let T be an algebraic theory. The functor

YT : T op → (Alg T )f p

is a free completion of T under finite colimits, conservative with respect to finite

coproducts. This means that

1. (Alg T )f p is finitely cocomplete and YT preserves finite coproducts

and

2. for every functor F : T op → B preserving finite coproducts, where B is

a finitely cocomplete category, there exists an essentially unique functor
F ∗ : Alg T → B preserving finite colimits with F naturally isomorphic to
F ∗ · YT .

We pass now to the elementary description of Rec C.

17.12 Definition. Given a category C with finite coproducts, we define the

category Rec C as follows:

1. Objects of Rec C are reflexive pairs x1 , x2 : X1 ⇉ X0 in C (that is, parallel

pairs for which there exists d : X0 → X1 such that x1 · d = idX0 = x2 · d,
see 3.12).

2. Consider the following diagram in C

Z1
z1 z2
f  
// Z
V g
0

with z1 , z2 a reflexive pair. We write

h : f 7→ g

if there exists a morphism h : V → Z1 such that z1 · h = f and z2 · h = g.

This is a reflexive relation in the hom-set C(V, Z0 ). We write f ∼ g if f
and g are in the equivalence relation generated by this reflexive relation.

November 20, 2009 164

CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

3. A premorphism in Rec C from (x1 , x2 ) to (z1 , z2 ) is a morphism f in C as

in the diagram
X1 Z1
x1 x2 z1 z2
   
X0 / Z0
f

such that f · x1 ∼ f · x2 .

4. A morphism in Rec C from (x1 , x2 ) to (z1 , z2 ) is an equivalence class [f ] of

premorphisms with respect to the equivalence ∼ of 2.

5. Composition and identities in Rec C are the obvious ones.

6. The functor ERec : C → Rec C is defined by

X Z
id id id id
f    
P (X / Z) = X /Z
[f ]

17.13 Remark.

1. Consider
Z1
z1 z2
f  
V // Z 0
g

as in 17.12. Explicitly, f ∼ g means that there exists a zig-zag

@f1 ? ; fn >
h1
?? h
?? 2 {{{ ~>>> hn+1
hn
{ >>
?? {{ >>
 }{{ 
f f2 ...... fn−1 g

2. Using the explicit description of f ∼ g, it is straightforward to prove that

Rec C is a category and P : C → Rec C is a full and faitful functor.

3. The above description of P : C → Rec C does not depend on the existence

of finite coproducts in C

17.14 Lemma. Let C be a category with finite coproducts. The category Rec C
of 17.12 has finite colimits and ERec : C → Rec C preserves finite coproducts.

November 20, 2009 165

 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

Proof. (1) Finite coproducts in Rec C are computed componentwise, i.e., if

x1 , x2 : X1 ⇉ X0 and z1 , z2 : Z1 ⇉ Z0 are objects of Rec C, their coproduct is
`
X1 X1 Z1 Z1
x1 x2 z1 z2
‘ ‘
x1 z1 x2 z2
   
`  
X0 / X0 Z0 o Z0
[iX0 ] [iZ0 ]

(2) Reflexive coequalizers in Rec C are depicted in the following diagram

`
X1 Z1 X0 Z1
x1 x2 z1 z2 hf,z1 i hg,z2 i
  [f ]
// Z   
X0 0
/ Z0
[g] [id]

17.15 Lemma. Consider the diagram

Z1
z1 z2
f  
// Z
V g
0

as in 17.12. If a morphism w : Z0 → W is such that w · z1 = w · z2 and f ∼ g,

then w · f = w · g.
Proof. Clearly if h : f 7→ g, then w · f = w · g. The claim now follows from
the fact that to be coequalized by w is an equivalence relation in C(V, Z0 ).
17.16 Remark. For every reflexive pair x1 , x2 : X1 ⇉ X0 in C, the diagram

ERec x1 [idX0 ] x1
ERec X1 // ERec X0 / (X1 // X 0 )
ERec x2 x2

is a reflexive coequalizer in Rec C. Therefore, if two functors F, G : Rec C → B

preserve reflexive coequalizers and F · ERec ≃ G · ERec , then F ≃ G.
17.17 Proposition. Let C be a category with finite coproducts. The functor

ERec : C → Rec C

of 17.12 is a free completion of C under finite colimits, conservative with respect

to finite coproducts. This means that
1. Rec C has finite colimits and ERec preserves finite coproducts

November 20, 2009 166

CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

and
2. for every functor F : C → B preserving finite coproducts, where B is
a finitely cocomplete category, there exists an essentially unique functor
F ∗ : Rec C → B preserving finite colimits with F naturally isomorphic to
F ∗ · ERec .
ERec
C? / Rec C
?? y y
?? ≃ yy
F ?? yy ∗
 |yy F
B

Proof. Given F : C → B as above, we define F ∗ : Rec C → B on objects by

the following coequalizer in B :

F x1
F X1 // F X / F ∗ (x1 , x2 )
0
F x2

Lemma 17.15 makes it clear how to define F ∗ on morphisms. The argument

for the essential uniqueness of F ∗ is stated in 17.16. The rest of the proof is
straightforward.
The previous universal property allows us to give a different prove of the
equation Sind C ≃ Ind (Rec C) already established in 7.4.
17.18 Corollary. Let C be a small category with finite coproducts. There exists
an equivalence of categories

Ind (Rec C) ≃ Sind C .

Proof. Let B be a cocomplete category. By 4.18, the functors Ind (Rec C) →

B preserving colimits correspond to the functors Rec C → B preserving finite
colimits and then, by 17.17, to the functors C → B preserving finite coproducts.
On the other hand, the functors C → B preserving finite coproducts correspond,
by 15.12, to the functors Sind C → B preserving colimits. Since both Ind (Rec C)
and Sind C are cocomplete (4.16 and 4.5), we can conclude that Ind (Rec C) and
Sind C are equivalent categories.

Historical Remarks for Chapter 17

The reflexive coequalizer completion of a category with finite coproducts is due

to A. M. Pitts (unpublished notes [82]). It appeared in press in [31]. The con-
nection between the exact completion and the reflexive coequalizer completion
was established by M. C. Pedicchio and J. Rosický in [79], see also [87].

November 20, 2009 167

 CHAPTER 17. EXACT COMPLETION AND
REFLEXIVE-COEQUALIZER COMPLETION

November 20, 2009 168

Chapter 18

Finitary localizations of
algebraic categories

In 6.9, we characterized algebraic categories among cocomplete categories by the

existence of a suitable generator. In this chapter, we will analogously character-
ize algebraic categories among exact categories. We also study the relationship
between algebraic categories and exact, locally finitely presentable categories.

18.1 Theorem. Let E be an exact category with sifted colimits, A a category

with finite limits and sifted colimits, and F : E → A a functor preserving finite
limits and filtered colimits. Then F preserves reflexive coequalizers if and only
if it preserves regular epimorphisms.

Proof. Necessity is evident, because every regular epimorphism is a reflexive

coequalizer of its kernel pair. For sufficiency, let F preserve finite limits, filtered
colimits and regular epimorphisms.
1. Since every equivalence relation in E is a kernel pair of its coequalizer and
since every regular epimorphism is a coequalizer of its kernel pair, F preserves
coequalizers of equivalence relations. Since every pseudoequivalence in E can be
decomposed as a regular epimorphism followed by an equivalence relation (cf.
16.16), F preserves coequalizers of pseudoequivalences.
2. Consider a reflexive and symmetric pair r = (r1 , r2 : X ′ ⇉ X) of morphisms
in E. We construct a pseudoequivalence r containing r (the transitive hull of r)
as a (filtered) colimit of the chain of compositions

r ◦ r ◦ ...◦ r n-times

(the composition r ◦ r is depicted in the following diagram, where the square is

169
 CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

a pullback)
r◦r
r1′
x FFF r′
x xx FF 2
xx FF
|xx F"
′
X E X′ B
r1 | || EE y y BB r
| EEr2 r1 y
y BB 2
|| EEE yy BB
~|| " |yy B
X X X
Since F preserves filtered colimits and finite limits, we have F r = F r. The pseu-
doequivalence r has a coequalizer, which is preserved by F. But a coequalizer
of r is also a coequalizer of r, and so F preserves coequalizers of reflexive and
symmetric pairs of morphisms.
3. If r = (r1 , r2 ) is just a reflexive pair, then a reflexive and symmetric pair
containing r is given by r ◦ r−1 , that is

r ◦ r−1G
ww GG
s1 ww GGs2
w GG
www GG
{w #
′
X H X′ B
| HH vv BB r
r1 || HHr2 r2 vv BB 1
|| HH vv BB
|| HH# v B
~| {vv
X X X

Once again, a coequalizer of r ◦ r−1 is also a coequalizer of r, so that F preserves

reflexive coequalizers.
18.2 Corollary. Let E be a cocomplete exact category, A a category with finite
limits and sifted colimits and F : E → A a functor preserving finite limits. Then
F preserves sifted colimits if and only if it preserves filtered colimits and regular
epimorphisms.
In fact, this follows from 7.7 and 18.1.
We can now generalize 5.16.
18.3 Corollary. In a cocomplete exact category, perfectly presentable objects
are precisely finitely presentable regular projectives.
Proof. One implication is established in 5.4. For the converse implication,
apply 18.2 to the hom-functor hom(G, −) of a finitely presentable regular pro-
jective object G.
18.4 Corollary. A category is algebraic if and only if it is cocomplete, exact
and has a strong generator consisting of finitely presentable regular projectives.
Proof. Necessity follows from 3.18 and 6.9. Sufficiency follows from 18.3 and
6.9.
In the previous corollary, the assumption of cocompleteness can be reduced
to asking for the existence of coequalizers of kernel pairs, which is part of the

November 20, 2009 170

CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

exactness of the category, and the existence of coproducts of objects from the
generator. In fact, we have the following result (recall that a category is well-
powered if, for a fixed object A, the subobjects of A constitute a set (not a
proper class):

18.5 Lemma. Let A be a well-powered exact category with a regular projective

cover P → A. If P has coproducts, then A is cocomplete.

Proof. 1. The functor P → A preserves coproducts. Indeed, consider a

coproduct a
si : Pi → Pi
I

in P, and a family of morphisms hxi : Pi → XiI in A. Let q : Q → X be a regular

epimorphism with Q ∈ P. For each i ∈ I, consider ` a morphism yi : Pi → Q such
that q · yi = xi . Since Q is in P, there is y : I Pi → Q such that y · si = yi ,
and then q · y · si = xi , for all i ∈ I.
As far as the uniqueness
` of the factorization is concerned, consider a pair of
morphisms ` f, g : P
I i ⇉ X such that si · f = si · g for all i. Consider also
f ′ , g ′ : I Pi ⇉ Q such that q · f ′ = f and q · g ′ = g. Since q · f ′ · si = q · g ′ · si ,
there is ti : Pi → N (q) such that q1 · ti = f ′ · si and q2 · ti = g ′ · si , where
q1 , q2 · N (q) ⇉ Q is a`kernel pair of q. From the first part of the proof, we
obtain a morphism t : I Pi → N (q) such that t · si = ti for all i. Moreover,
q1 · t · si = f ′ · si for all i, so that q1 · t = f ′ because Q is in P. Analogously,
q2 · t = g ′ . Finally, f = q · f ′ = q · q1 · t = q · q2 · t = q · g ′ = g.
2. Recall that SubA (A) is the ordered class of subobjects of A. For every category
A, we denote by θ(A) its “ordered reflection”, i.e., the ordered class obtained
from the preorder on the objects of A given by A ≤ B iff A(A, B) is nonempty.
We are going to prove that for any object A of A, SubA (A) and θ(P/A) are
isomorphic ordered classes. In fact, given a monomorphism m : X → A, we
consider a P-cover q : Q → X and we get an element in θ(P/A) from the com-
position m · q. Conversely, given an object f : Q → A in P/A, the monomorphic
part of its regular factorisation gives an element in SubA (A).
3. A has coequalizers. Consider a parallel pair (a, b) in A and its regular
factorization
a //
B NNN A
O O
NNNb
NNN i1 i2
r NNN
N& &
R
Consider now the equivalence relation a1 , a2 : A′ ⇉ A generated by (i1 , i2 ), that
is the intersection of all the equivalence relations on A containing (i1 , i2 ). Such
an intersection exists: by part 2., SubA (A) is isomorphic to θ(P/A), which is
cocomplete because P has coproducts. Since, by assumption, A is well-powered,
SubA (A) is a set, and a cocomplete ordered set is also complete. Since A is exact,
(a1 , a2 ) has a coequalizer, which is also a coequalizer of (i1 , i2 ) and then of (a, b).
4. A has coproducts. Consider a family of objects (Ai )I in A. Each of them can

November 20, 2009 171

 CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

be seen as a coequalizer of a pseudoequivalence in P as in the following diagram,

where the first and the second columns are coproducts in P (and then in A,
see 1.), x0 and x1 are the extensions to the coproducts, the bottom row is a
coequalizer (which exists by Part 3.), and σi is the extension to the coequalizer.
xi1 ai
Pi′ // Pi / / Ai
xi2
s′i si σi
` x1 ` 
I Pi′ // I Pi q
//Q
x2

Since coproducts commute with coequalizers, the third column is a coproduct

of the family (Ai )I .
18.6 Corollary. A category is algebraic if and only if it is exact and has a
strong generator G consisting of finitely presentable regular projectives such that
coproducts of objects of G exist.
Proof. Let A be an exact category and G a strong generator consisting of
regular projectives. Since a coproduct of regular projectives is regular projective,
the full subcategory P consisting of coproducts of objects from G is a regular
projective cover of A. Following 18.5, it remains to prove that A is well-powered.
Consider an object A and the map
F : Ω(G ↓ A) → SubA (A)
assigning to a subset M of G ↓ A the subobject of A represented by the
monomorphism sM : SM → A, where
a
eM : G→A
(G,g)∈M

is the canonical morphism whose (G, g)-component is g, and

` qM
(G,g)∈M G
/ SM sM / A

is the regular factorization of eM . We are to prove that F is surjective, so that

SubA (A) is a set. For this, consider a monomorphism m : S → A and let M(s)
be the set of those (G, g) ∈ G ↓ A such that g factorizes through s :
g
G? /A
?? ?
?? 
h ?? s
 
S
We get the commutative diagram
` ŝ / ` qM(s)
(G,h)∈G↓S G (G,g)∈M(s) G
/ SM(s)
QQQ
QQQ sM(s)
Q σ
eS QQQQ d
QQQ  
( w /A
S s

November 20, 2009 172

CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

where the (G, h)-components of eS and ŝ are h and g, respectively, and the
(G, g)-component of σ is h. By diagonal fill-in there exists d : SM(s) → S such
that d · qM(s) = σ and s · d = sM(s) . Such a d is an isomorphism: it is a
monomorphism because sM(s) is, and an extremal epimorphism because eS is.
Thus, F (M(s)) = s and the proof is complete.

From Propositions 3.18 and 6.22, we know that an algebraic category is exact
and locally finitely presentable. The converse is not true because of the lack of
projectivity of the generator. In the remaining part of this chapter we want to
state in a precise way the relationship between algebraic categories and exact,
locally finitely presentable categories.
18.7 Definition. Given a category B, by a localization of B is meant a reflective
subcategory A whose reflector preserves finite limits. And A is called a finitary
localization if, moreover, it is closed in B under filtered colimits.
18.8 Remark. More loosely, we speak about localizations of B as categories
equivalent to full subcategories having the above property. We use the notation

o R
A /B
I

that is, R is left adjoint to I and I is full and faithful.

Let us start with a general lemma.
18.9 Lemma. Consider a reflection

o R
A /B
I

1. If I preserves filtered colimits and an object P ∈ B is finitely presentable,

then R(P ) is finitely presentable;
2. If the reflection is a localization and B is exact, then A is exact.
Proof. 1. Same argument as in the proof of 6.16.1.
2. Let r1 , r2 : A′ ⇉ A be an equivalence relation in A. Its image in B is an
equivalence relation, so that it has a coequalizer Q and it is the kernel pair of
its coequalizer (because B is exact)

Ir1 q
IA′ // IA //Q
Ir2

If we apply the functor R to this diagram, we obtain a coequalizer (because R

is a left adjoint) and a kernel pair (because R preserves finite limits)
r1 Rq
// / / RQ
RIA′ ≃ A′ r2
A ≃ RIA

November 20, 2009 173

 CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

and this means that (r1 , r2 ) is effective. It remains to prove that regular epi-
morphisms are stable under pullbacks. For this, consider a pullback

f′
P /C

g′ g
 
A /B
f

in A, with f a regular epimorphism. Form its image in B, computed as a

two-step pullback of Ig along the regular factorization m · e of If

e′ /Q m′ / IC
IP
Ig′ h Ig
  
IA /E / IB
e m

so that e′ is a regular epimorphism. If we apply the functor R to the second

diagram, we come back to the original pullback, computed now as a two-step
pullback (because R preserves finite limits)

Re′ / RQ Rm′ / RIC ≃ C

P ≃ RIP
g′ Rh g
  
A ≃ RIA / RE / RIB ≃ B
Re Rm

Now observe that Rm is a monomorphism (because R preserves finite lim-

its) and also a regular epimorphism (because f is a regular epimorphism, and
Rm · Re = f ), so that it is an isomorphism. It follows that Rm′ is an isomor-
phism. Moreover, Re′ is a regular epimorphism (because R, being a left adjoint,
preserves regular epimorphisms). Finally, f ′ is a regular epimorphism because
f ′ = Rm′ · Re′ .
18.10 Theorem. Finitary localizations of algebraic categories are precisely the
exact, locally finitely presentable categories.
Proof. Since an algebraic category is exact and locally finitely presentable,
necessity follows from 6.16.1 and 18.9. For the sufficiency, let A be an exact and
locally finitely presentable category. Following 6.26, A is equivalent to Lex T ,
where T ≃ Aop f p , and Lex T is a reflective subcategory of Alg T closed under
filtered colimits (see 6.29). Consider the full subcategory P of Alg T consisting
of regular projective objects. Such an object P is a retract of a coproduct
of representable algebras (5.14.2). Since every coproduct is a filtered colimit
of its finite subcoproducts, and a finite coproduct of representable algebras is
representable (1.5), P is a retract of a filtered colimit of representable algebras.
Following 4.3, we have that P is contained in Lex T . Moreover, P is a regular

November 20, 2009 174

CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

projective cover of Alg T (5.15). Since, by 16.27, the full inclusion of P into
Alg T is a free exact completion of P and, by assumption, Lex T is exact, it
remains just to prove that the inclusion P → Lex T is left covering. Once this
is done, we can apply 16.26 to the following situation

P EE / Alg T
EE u u
EE
EE uuu
uu
E" zuu R
Lex T
where R is the reflector, and we conclude that R is an exact functor. But the
inclusion P → Alg T ≃ Pex is left covering, and Lex T is closed in Alg T under
limits, so that also the inclusion P → Lex T is left covering.
18.11 Remark. To finish this chapter, we observe that Corollary 18.2 can
be used to prove the characterization of varieties established in 10.24 without
using Birkhoff’s Variety Theorem 10.22. We sketch the argument. Let T be an
algebraic theory and
I : A → Alg T , R ⊣ I
a regular epireflective subcategory closed under regular quotients and directed
unions. From the closedness under regular quotients it immediately follows
that I preserves regular epimorphisms and that in A equivalence relations are
effective, so that A is exact. To prove that I preserves filtered colimit, consider
a functor F : D → A with D filtered and let hσd : F d → Bid∈D be its colimit
cocone in Alg T . Let
ed
Fd / Gd md / B
be the regular factorization of σd . For any morphism f : d → d′ in D there
exists a unique Gf : Gd → Gd′ such that md′ · Gf = md (use diagonal fill-in,
cf. 0.16). This defines a new functor G : D → A (indeed Gd ∈ A because it
is a regular quotient of F d ∈ A) and B is the directed union of the Gd’s, so
that B ∈ A. Following 18.2 I preserves sifted colimits and then by 6.18 A is
algebraic. Moreover, an algebraic theory TA of A can be described as follows
(cf. 6.16):
TAop = {R(T (X, −)) | X ∈ T }
The functor
T → TA , X 7→ R(T (X, −))
preserves finite products (by 1.5) and is surjective on objects. It remains to
prove that it is also full: let η be the unit of the adjunction R ⊣ I and con-
sider f : R(T (X, −)) → R(T (Z, −)). Since ηT (Z,−) is a regular epimorphism
and T (X, −) is regular projective, there exists g : T (X, −) → T (Z, −) such that
f · ηT (X,−) = ηT (Z,) · g. Since η is natural and ηT (X,−) is an epimorphism, it
follows that Rg = f. By 10.13 TA is a quotient of T . Finally
A ≃ Alg TA ≃ Alg (T / ∼)
so that A is a variety of T -algebras.

November 20, 2009 175

 CHAPTER 18. FINITARY LOCALIZATIONS OF ALGEBRAIC
CATEGORIES

Historical Remarks for Chapter 18

The first systematic study of localizations is due to P. Gabriel [47] who also
together with N. Popesco characterized Grothendieck categories, see [83]. This
is an ancestor of Theorem 18.10.
Arbitrary (i.e., non necessarily finitary) localizations of one-sorted algebraic
categories and, more in general, of monadic categories over Set are characterized
in [93] and in [94]. Essential localizations are studied in [10], which generalizes
the original result for module categories due to J. E. Roos [88].
One of the results of Lawvere’s thesis [63] is a characterization of one-sorted
algebraic categories, compare with Corollary 18.6. The only difference is that
in Lawvere’s original result the generator is required to be abstractly finite,
a notion that without the other conditions of the characterization theorem is
weaker than finitely presentability.

November 20, 2009 176

Appendix A

Monads

An important aspect of algebraic categories which has not yet been treated in
this book are monads. The aim of this appendix is to give a short introduction to
monads on a category K, and then to explain how finitary monads for K = Set
precisely yield one-sorted algebraic theories, and for K = Set S the S-sorted
ones.
The word “monad” stems from monoid: recall that a monoid in a category
K is an object M together with a morphism m : M × M → M which

1. is associative, that is, the square

m×idM
M ×M ×M / M ×M

idM ×m m
 
M ×M /M
m

commutes, and

2. has a unit, that is a morphism e : 1 → M such that the triangles

5 M iS
kkkk O SSSSSSS
idM kkkkk Sid
SSMSS
kkkk m
SSSS
kkkk
M = 1 × M e×id / M ×M o M ×1=M
M id ×e
M

commute.

A.1 Definition.

1. A monad M on a category K consists of an endofunctor M on K and

natural transformations

(a) µ : M M → M (monad multiplication) and

177
 APPENDIX A. MONADS

(b) η : IdK → M (monad unit)

such that the diagrams

Mµ
MMM / MM [A.1]
µM µ
 
MM /M
µ

and
6 M hP [A.2]
nnn O PPPPP
Mnnnn PM
PPP
n µ
nnn PPP
nnn / M ·M o
P
M ηM Mη
M

commute.
2. The monad is called finitary if M is a finitary functor.
A.2 Example.
1. The functor M : Set → Set given by

MX = X + 1

carries an obvious structure of a monad: the unit is the coproduct injection

ηX : X → X + 1 and the multiplication µX : X + 1 + 1 → X + 1 merges
the two copies of 1 to one.
2. The word monad on Set assigns to every set X the set

M X = X∗

of all words on it, that is, the (underlying set of the) free monoid on X.
This yields an endofunctor on Set together with natural transformations
ηX : X → X ∗ , the formation of one-letter words, and µX : (X ∗ )∗ → X ∗
given by concatenation of words.
3. Recall that for every finitary endofunctor H of Set free H-algebras exist,
giving a left adjoint FH : Set → H-Alg (see 12.7). The corresponding
monad H ∗ on Set is called the free monad on H. For example, if H is
the polynomial functor of a signature Σ, the free monad is the monad of
Σ-terms.
In the notation of 12.6 we have two natural transformations

η : Id → UH FH = H ∗

and
ϕ : HH ∗ → H ∗

November 20, 2009 178

APPENDIX A. MONADS

yielding a natural transformation

ψ = ϕ · Hη : H → H ∗ .

We will see in A.28 that it has the universal property explaining the name
free monad.

A.3 Example. The basic example of a monad is that induced by any adjoint
situation
o F /
A K where F ⊣ U .
U

Let η : IdK → U F and ε : F U → IdA denote the unit and counit of the adjunc-
tion. Recall the equalities

U = U ε · ηU and F = εF · F η [A.3]

characterizing adjoint situations. Then for the endofunctor

M = UF : K → K

we have the natural transformation

µ = U εF : M M = U F U F → U F = M [A.4]

which, together with the unit η : IdK → M, forms a monad on K. In fact, the
commutativity of the two triangles [A.2] follows from [A.3] and the square [A.1]
follows from the naturality of ε :

ε · F U ε = ε · εF U [A.5]

yielding
µ · M µ = U (ε · F U ε)F = U (ε · εF U )F = µ · µM .
Observe that whenever U is a finitary functor, this monad is finitary because
F, being a left adjoint, always preserves filtered colimits.

A.4 Example.

1. Every one-sorted algebraic category U : A → Set defines a monad on Set

assigning to every set X the free algebra generated by it. In other words,
this is the monad induced by the adjunction F ⊣ U as in A.3, where F is
the free-algebra functor of 11.21.
Since U preserves filtered colimits by 11.8, all these monads are finitary.

2. Analogously, every S-sorted algebraic category defines a finitary monad

on Set S .

A.5 Remark. Recall from 12.1 the category M -Alg of M -algebras for the
endofunctor M of K. If M is the monad induced by an adjoint situation F ⊣ U

November 20, 2009 179

 APPENDIX A. MONADS

as in A.3, then every object A of A yields a canonical M -algebra on X = U A :

put
x = U εA : M X = U F U A → U A = X .
This algebra has the property that the triangle

XE [A.6]
EE
ηX
EEidX
EE
 E"
MX x / X

commutes: see [A.3]. Also the square

µX
MMX / MX [A.7]
Mx x
 
MX /X
x

commutes, see [A.5]. This leads to the following

A.6 Definition. An Eilenberg-Moore algebra for a monad M = (M, µ, η) on

K is an algebra (X, x) for M such that [A.6] and [A.7] commute. The full
subcategory of M -Alg formed by all Eilenberg-Moore algebras is denoted by

KM .

A.7 Remark. The Eilenberg-Moore category K M is considered a concrete cat-

egory on K via the faithful functor

UM : K M → K , (X, x) 7→ X

It is easy to verify that this concrete category is uniquely transportable (same

argument as in 13.17.3).

A.8 Example.

1. For every category K we have the trivial monad Id = (IdK , id, id). The
only Eilenberg-Moore algebras are idX : X → X. Thus, K I ≃ K.

2. For the monad M X = X + 1 of A.2.1 an Eilenberg-Moore algebra is a

pointed set: given x : X + 1 → X with [A.6], the left-hand component of
x is idX , thus, x just chooses an element 1 → X. Here [A.7] always com-
mutes. Homomorphisms are functions preserving the choice of element.
Shortly: Set M is the category of pointed sets.

3. For the word monad A.2.2 the category Set M is essentially the category
of monoids. In fact, given an Eilenberg-Moore algebra x : X ∗ → X, then
[A.6] states that the response to one-letter words is trivial: x(a) = a. And

November 20, 2009 180

APPENDIX A. MONADS

[A.7] states that for words of length larger than 2 the response is given by
the binary operation
a1 ∗ a2 = x(a1 a2 ) .
In fact, for example in length 3 we get

x(a1 a2 a3 ) = x(a1 (a2 a3 )) = a1 ∗ (a2 ∗ a3 )

as well as
x(a1 a2 a3 ) = x((a1 a2 )a3 ) = (a1 ∗ a2 ) ∗ a3 .
Thus, ∗ is an associative operation. And [A.7] also states that the response
of x to the empty word is a unit for x.
Conversely: every monoid defines an Eilenberg-Moore algebra: see Re-
mark A.5. The monoid homomorphisms are easily seen to be precisely
the homomorphisms in Set M . Thus, Set M is isomorphic to the category of
monoids.
A.9 Example. Free Eilenberg-Moore algebras. For every monad M, the M -
algebra
(M X, µX : M M X → M X)
is an Eilenberg-Moore algebra: [A.6] and [A.7] follow from the definition of
monad. This algebra is free with respect to ηX : X → M X. In fact, given an
Eilenberg-Moore algebra (Z, z) and a morphism f : X → Z in K, the unique
homomorphism extending f is f = z · M f :
(a) f = z · M f is a homomorphism:
µX
MMX / MX

MMf Mf
 µY 
MMZ / MZ

Mz z
 
MZ /Z
z

due to [A.5] and the naturality of µ.

(b) Conversely, if f : (M X, µX ) → (Z, z) is a homomorphism, then f = f · ηX
implies f = z · M f :
µX
MMX / MX
dJJ w;
JJMηX id www
JJ w
JJ ww
J ww
Mf MX f
ttt
tt
tt
 ztt Mf 
MZ /Z
z

November 20, 2009 181

 APPENDIX A. MONADS

A.10 Corollary. Every monad is induced by some adjoint situation.

In fact, given a monad M on a category K, we have the above adjoint situation

FM
o /K
KM
UM

where FM is the free-algebra functor

FM X = (M X, µX ) .

It is defined on morphisms by FM f = M f. Thus, the monad induced by the

adjunction FM ⊣ UM has the underlying endofunctor UM · FM = M, and the unit
η (recall the universal arrows ηX above). We need to verify that for the counit
ε of the adjoint situation we have

µ = UM ε FM .

In fact, the component of ε at an Eilenberg-Moore algebra (X, x) is the unique

morphism
ε(X,x) : (M X, µX ) → (X, x)
with UM ε(X,x) · ηX = idX . But due to [A.6] and [A.7] the morphism x carries
such a homomorphism. Therefore, UM ε(X,x) = x for all algebras (X, x). In
particular,
(UM ε FM )X = UM ε(MX,µX ) = µX .

A.11 Definition. For every adjoint situation

o F
A /K with F ⊣ U
U

let M be the monad of A.3. The comparison functor is the functor

K: A → KM

which assigns to every object A the Eilenberg-Moore algebra

KA = (U A, U εA )

of A.5. The definition of K on morphisms f : A → B uses the naturality of ε

which shows that Kf = U f is a homomorphism:

UεA
MUA / UA

MUf Uf
 
MUB / UB
UεB

November 20, 2009 182

APPENDIX A. MONADS

A.12 Remark. The comparison functor K : A → K M of A.11 is the unique

functor such that UM · K = U and K · F = FM .
A.13 Example.
1. For the concrete category of monoids U : Mon → Set the comparison func-
tor K : Mon → Set M is an isomorphism. The inverse K −1 was described
in A.8.2 on objects, and acts trivially on morphisms: K −1 f = f.
2. The concrete category U : Pos → Set of partially ordered sets yields the
trivial monad Id = (IdSet , id, id) of A.8: recall that the left adjoint of
U assigns to every set X the discrete order on the same set. For this
monad we have an isomorphism between Set and Set M , and the comparison
functor is then simply the forgetful functor U.
3. For the free monad on H, see A.2.3, the Eilenberg-Moore ∗category is con-
cretely isomorphic to H-Alg . Indeed, the functor J : Set H → H-Alg tak-
ing an Eilenberg-Moore algebra x : H ∗ X → X to the H-algebra obtained
by composing with ψX (see A.2.3) is easily seen to be a concrete isomor-
phism.
A.14 Definition. A concrete category (A, U ) on K is monadic if U has a left
adjoint F and the comparison functor K : A → K M is an isomorphism.
A.15 Remark.
1. The fact that (A, U ) is monadic does not depend on the choice of the
left adjoint of U. Indeed, if F ′ is another left adjoint, then the canonical
natural isomorphism F ≃ F ′ induces an isomorphism of monads M ≃ M′
(see A.24 for the notion of monad morphism), where M′ is the monad
induced by the adjunction U ⊣ F ′ . As we will see in A.26, this implies
′
that K M and K M are concretely isomorphic.
2. In other words, monadic concrete categories are precisely those which, up
to concrete isomorphism, have the form K M . It is not surprising, then,
that monoids are an example and posets are not.
A.16 Definition. A coequalizer in a category K is called absolute if every
functor with domain K preserves it.
A.17 Example. For every Eilenberg-Moore algebra (X, x) we have an absolute
coequalizer
Mx / x
MMX µ / MX /X
X

in K. In fact, x merges the parallel pair by [A.7] and, moreover, the morphisms
ηX and ηMX are easily seen to fulfil the following equations:
(i) µX · ηMX = idMX ,
(ii) ηX · x = idX ,

November 20, 2009 183

 APPENDIX A. MONADS

and

(iii) M x · ηMX = ηX · x.

It is easy to derive from (i)–(iii) that x is a coequalizer of M x and µX . Since every

functor G preserves the equations (i)–(iii), it follows that Gx is a coequalizer of
GM x and GµX .

A.18 Beck’s Theorem. (Characterization of monadic categories) A concrete

category (A, U ), with U a right adjoint, is monadic if and only if

(a) it is uniquely transportable,

and

(b) A has coequalizers of all reflexive pairs f, g such that U f, U g have an

absolute coequalizer, and U preserves these coequalizers.

A proof of A.18 can be found in [71], Chapter 6, Section 7. The reader has just
to observe that the parallel pairs of morphisms used in that proof are reflexive,
and U creates the coequalizers involved in condition (b) since it is amnestic and
conservative.

A.19 Proposition. Every equational category is monadic.

Proof. Condition (a) of A.18 follows from 13.17 and (b) from 13.11 and 11.8:
the forgetful functor is algebraic, thus it preserves reflexive coequalizers.

A.20 Example.

1. Pointed sets, monoids, groups, abelian groups, . . . , with their forgetful

functors are monadic.

2. For a one-sorted algebraic theory (T , T ), the concrete category (Alg T , Alg T )

in general is not monadic, as Example 11.7 shows: Alg Tab is not amnes-
tic whereas UM always is. What remains true is that (Alg T , Alg T ) is
pseudo-monadic, as we will see in C.4,

A.21 Theorem. Equational categories are up to concrete isomorphism precisely

the categories Set M of Eilenberg-Moore algebras for finitary monads M on Set.

Proof. In fact, every equational category is, by A.19, concretely isomorphic

to Set M where M is the monad of its free algebras.
Conversely, given a finitary monad M = (M, µ, η) on Set, we know from 13.23
that M -Alg is concretely isomorphic to an equational category. Therefore, it
is sufficient to prove that Set M is closed in M -Alg under products, subobjects,
and regular quotients. Then the result follows from Birkhoff’s Variety Theorem
in the form 13.22.

November 20, 2009 184

APPENDIX A. MONADS

Q
(a) Products: let (X, x) = i∈I (Xi , xi ) where each (Xi , xi ) is an Eilenberg-
Moore algebra. (X, x) satisfies [A.6] because the projections (πi )i∈I are a limit
cone, thus collectively monomorphic, and the diagram

X
w DDD
πiwww DDidX
www DD
{w
w D!
x /X
Xi MX AA
ww AA πi
ηXi www AA
ww AA
 {ww Mπi
M Xi / Xi
xi

commutes for every i. Thus xi · ηXi = id implies x · ηX = id . (X, x) satisfies

[A.7] for similar reasons:

µXi
M M XeLi / M Xi
LLL w w;
LMMπ
LLL i Mπi ww
w
LL ww
µX ww
MMX / MX

Mxi Mx x xi
 
MX /X
x GG
rrrr GG
GG
rrr πi GGG
 yrrr Mπi # 
M Xi / Xi
xi

(b) Subalgebras m : (X, x) → (Z, z) of Eilenberg-Moore algebras (Z, z). In the

following diagram

ZO
m

ηZ
XA
zz AA id idZ
z ηX AA X
zz
z AA
|z
z
MX /X
w x @@
ww @@
w
wMm @@
w m @
{w
 w
MZ /Z
z

the outward triangle commutes and all parts except the middle triangle also com-
mute. Thus so does the middle triangle since m is a monomorphism. Therefore,
(X, x) satisfies [A.6]. The proof of [A.7] is analogous.
(c) Regular quotients e : (Z, z) → (X, x) of Eilenberg-Moore algebras (Z, z). In

November 20, 2009 185

 APPENDIX A. MONADS

the following diagram

Z
e

X
ηZ
ηX zz
z AAA id idZ
zz AA X
z AA
|zz
MX /X
w; x `@@
w @@
www @@
w e
 ww Me @
MZ /Z
z

again all parts except the middle triangle commute. Since e is an epimorphism,
so does the middle triangle. Therefore, (X, x) satisfies [A.6]. For [A.7] use the
analogous argument plus the fact that M preserves epimorphisms (because they
split in Set).

A.22 Corollary. One-sorted algebraic categories are up to concrete equivalence

precisely the categories Set M of Eilenberg-Moore algebras for finitary monads M
on Set.

In fact, this follows from A.21 and 13.11.

A.23 Corollary. For every finitary monad M on Set the category Set M is
cocomplete and the forgetful functor UM : Set M → Set preserves sifted colimits.

In fact, the category Set M is equational by A.21 and then one-sorted algebraic
by 13.11. Use now 4.5 and 11.9.

A.24 Definition. Let M = (M, µ, η) and M′ = (M ′ , µ′ , η ′ ) be monads on

a category K. A monad morphism from M to M′ is a natural transformation
ρ : M → M ′ such that the diagrams

IdK D
η {{ DD η′
{{ DD
{{ DD
}{{ D!
M / M′
ρ

and
Mρ ρM ′
MM / MM′ / M ′M ′
µ µ′
 
M / M′
ρ

commute.

November 20, 2009 186

APPENDIX A. MONADS

A.25 Example. Let M be any endofunctor of Set. Following 4.11 and 4.13 there
exists a finitary endofunctor M f and a natural transformation ǫM : M f → M
universal among the natural transformations N → M with N finitary.
If M is part of a monad M = (M, µ, η), then M f becomes part of a finitary
monad and ǫM a monad morphism: the monad multiplication for M f is

i µf
Mf · Mf / (M · M )f / Mf

where µf is the unique natural transformation with ǫM · µf = µ · ǫMM and i is

the unique natural isomorphism with in = id for all n ∈ N , and analogously for
the monad unit.

A.26 Proposition. Every monad morphism ρ : M → M′ induces a concrete

functor
′
Hρ : K M → K M
assigning to every Eilenberg-Moore algebra (X, x) for M′ the algebra (X, x · ρX )
for M. Conversely, given a concrete functor

′ H / KM
KM C
CC ||
CC
C ||
UM′ CC ||| UM
! ~|
K

there exists a unique monad morphism ρ : M → M′ with H = Hρ .

In fact, for a free Eilenberg-Moore algebra (M ′ A, µ′A ), the algebra H(M ′ A, µ′A )
has the form (M ′ A, σA : M M ′ A → M ′ A). We get a monad morphism
′
MηA σA
ρ : M → M′ , ρA : M A / M M ′A / M ′A .

A full proof can be found in [27], Vol. 2, Proposition 4.5.9.

A.27 Corollary. The category of finitary monads on Set and monad morphisms
is dually equivalent to the category of finitary monadic categories on Set and
concrete functors.

In fact, this follows from A.26.

A.28 Corollary. The free monad on a finitary endofunctor (see A.2.3) is indeed
free on H : for every finitary monad M = (M, µ, η) and every natural transfor-
mation α : H → M there exists a unique monad morphism α∗ : H ∗ → M with
α = α∗ · ψ.
∗
In fact, recall from A.13.3 that H-Alg is concretely isomorphic to Set H and
use, in place of α∗ , the concrete functor from Set M to H-Alg assigning to every
Eilenberg-Moore algebra x : M X → X the algebra x · αX : HX → X.

November 20, 2009 187

 APPENDIX A. MONADS

A.29 Remark. We observed above that every monad M is induced by an

adjoint situation using the Eilenberg-Moore algebras. There is another way to
induce M : the construction of the Kleisli category of M. This is, as we note
below, just the full subcategory of K M on all free algebras:
A.30 Definition. The Kleisli category of a monad M is the category
KM
with the same objects as K and with morphisms from X to Z given by mor-
phisms f : X → M Z in K :
KM (X, Z) = K(X, M Z) .
The identity morphisms are ηX and the composition of two morphisms f ∈
KM (X, Z) and g ∈ KM (Z, W ) is given by the following composition in K :
f Mg µW
X / MZ / MMW / MW

A.31 Example. For the monad M X = X + 1 of A.2.1 the Kleisli category is

the category of sets and partial functions. A partial function from X to Z is
represented as a (total) function from X to Z + 1.
A.32 Notation. For every monad M we denote
1. by KM : KM → KM the functor which assigns to X ∈ KM the free Eilenberg-
Moore algebra (M X, µX ) and to f ∈ KM (X, Z) the morphism
Mf µZ
MX / MMZ / MZ

2. by JM : K → KM the functor which is the identity map on objects and

which to every morphism u : X → Z of K assigns
u ηZ
X /Z / MZ

A.33 Lemma. The functor JM is a left adjoint of

KM UM
KM / KM /K

and M is the monad induced by the adjunction JM ⊣ UM · KM . The functor KM

is the corresponding comparison functor; it is full and faithful.
Proof. JM is a left adjoint of UM · KM with unit
ηX : X → M X = UM KM JM X
and counit given by the morphism in KM (JM UM KM X, X) which is ε = idMX in
K. The two axioms
εJM · JM η = JM and UM KM ε · ηUM KM = UM KM
are easy to check.

November 20, 2009 188

APPENDIX A. MONADS

A.34 Theorem. Given monads M and M′ on a category K, there is a bijective

correspondence between monad morphisms ρ : M → M′ and functors G : KM →
KM′ for which the triangle
KC
|| CC J
JM CC M′
||| CC
| C!
~|
|
KM / KM′
G

commutes.
We will see in the proof that the bijective correspondence assigns to every monad
morphism ρ : M → M′ the functor ρ̂ : KM → KM′ which is the identity map on
objects, and assigns to p : X → M Z in KM (X, Z) the value
p ρZ
X / MZ / M ′Z

Proof. (a) ρ̂ is well-defined: Preservation of identity morphisms follows from

′
ρX ·ηX = ηX . Preservation of composition follows from the commutativity of the
diagram below, where p : X → M W and q : W → M Z are arbitrary morphisms:
µZ
MMZ / MZ
Mq jjjj
jjj4 JJ
JJ ρZ
j JJ
jjjj JJ
p j jjjj J%
X / MW MC ′ Z
HH 
HH
HH ρM Z 
ρW HHH
$ 
M ′ WJ  ′
JJ  µZ
JJ
JJ 
M ′ q JJ% 


M ′M Z / M ′M ′Z
M ′ ρZ

(b) The equality ρ̂ · JM = JM′ follows from ρ · η = η ′ .

(c) If ρ, σ are different monad morphisms, then ρ̂ 6= σ̂ due to the fact that the
component ρX is obtained from ρ̂ by ρ̂(idMX ) = ρX .
(d) Let G : KM → KM′ be a functor such that G · JM = JM′ . Observe that this
condition tells us that G is the identity map on objects. The identity morphism
idMX : M X → M X can be seen as a morphism εX : M X → X in KM . Applying
G we get a morphism GεX : M X → X in KM′ , that is, a morphism M X → M ′ X
in K that we denote by ρX . We claim that ρX is the component at X of a monad
morphism ρ : M → M′ with ρ̂ = G.
The fact that G is a functor means that
′
(i) GηX = ηX (preservation of identity morphisms)
(ii) given p : X → M W and q : W → M Z in K we have
p Mq µZ
G( X / MW / MMZ / MZ ) =

November 20, 2009 189

 APPENDIX A. MONADS

Gp M ′ Gq µ′Z
=(X / M ′W / M ′M ′Z / M ′Z )

(preservation of composition)

and

(iii) for every u : X → W in K we have

′
u ηW u ηW
G( X /W / MW ) = ( X /W / M ′W )

(due to G · JM = JM′ ).

This implies for all u : X → W in K the equation

(iv) GM u = Gu · ρX

since we apply (ii) to p = idMX and q = ηW · u (thus the previous objects X, W

′
and Z are now M X, X and W, respectively, and Gq = ηW · u by (iii)) and use
′
[A.2] for M and M . Also for every q : W → M Z we have

(v) G(q · u) = Gq · u

by applying (ii) and (iii) to p = ηW · u. From this we derive the naturality of ρ :

(vi) ρW · M v = M ′ v · ρX for all v : X → W

since (v) yields for q = idMW and u = M v

ρW · M v = GM v

and then we apply (iv). The equality ρ · η = η ′ follows from (v) by u = ηW and
q = idMW , and the equality ρ · µ = µ′ · ρM ′ · M ρ follows from (ii) by p = idMMX
and q = idMX ; this proves that the right-hand side is equal to GµX , whereas
the left-hand side is GµX by (v) applied to q = µX and u = idMMX . Thus,

ρ : M → M′

is a monad morphism. Finally we need to prove ρ̂ = G, that is, for every

p0 : X → M W
Gp0 = GεX · p0
and for this apply (v) to u = p0 and q = idMW .

A.35 Remark. Let M be a finitary monad on Set. Then the functor M is

essentially determined by its domain restriction to N op (the full subcategory
of natural numbers). Also the natural transformations η and µ are uniquely
determined by their components ηn and µn for natural numbers n. In fact, this
follows from 4.11 applied to YN : N op → Set.

This leads us to the following restriction of the Kleisli category:

November 20, 2009 190

APPENDIX A. MONADS

A.36 Notation. For every finitary monad M on Set we denote by

SetfM

the full subcategory of the Kleisli category KM on all natural numbers, and by
f
JM : N op → SetfM

the domain-codomain restriction of JM .

A.37 Corollary. Given finitary monads M and M′ on Set, there is a bijec-

tive correspondence between monad morphisms ρ : M → M′ and those functors
G : SetfM → SetfM′ for which the triangle

N opF
f xx FF J f
JM FF M′
xxx FF
x
|xx F#
SetfM G
/ Setf ′
M

commutes. It assigns to ρ the functor

p p ρk
ρ̂f : ( n / M k ) 7→ ( n / Mk / M ′k )

The proof is completely analogous to that of A.34. All we need to notice is

that a monad morphism ρ : M → M′ is uniquely determined by its components
ρn , n ∈ N op . This follows once again from 4.11 applied to YN : N op → Set.

In 11.36 we defined the 2-category Th 1 of one-sorted algebraic theories. In

the following theorem we consider it as a category, that is, we forget the 2-cells.

A.38 Theorem. The category Th 1 of one-sorted algebraic theories is equivalent

to the category of finitary monads on Set.

Proof. Denote by FMon the category of finitary monads and monad mor-
f
phisms. Every object M defines a one-sorted theory by dualizing JM of A.36:
f op
(JM ) : N → (SetfM )op .

In fact, since JM : Set → Set M is a left adjoint, it preserves coproducts. In other

words, coproducts are the same in Set and in Set M .The same, then, holds for fi-
nite coproducts in the full subcategories N op and SetfM , respectively. Therefore,
f op
the identity-on-objects functor (JM ) preserves finite products and we obtain
a one-sorted theory  
E(M) = (SetfM )op , (JM f op
) .

This yields a functor

E : FMon → Th 1

November 20, 2009 191

 APPENDIX A. MONADS

which to every monad morphism ρ : M → M′ between finitary monads assigns

E(ρ) = (ρ̂f )op .

The commutative triangle of A.37 tells us that this is a morphism of one-sorted

theories. It is easy to verify that E is a well-defined functor.
Next E is full and faithful due to the bijection in A.37. Finally, E is essen-
tially surjective. Recall from 11.21 that for every one-sorted theory (T , T ) we
have a left adjoint FT to the forgetful functor Alg T (of evaluation at 1) such
that
FT (n) = T (n, −) and FT (πin ) = − · T πin .
Moreover, the naturality square for η : Id → Alg T · FT applied to πin yields

ηn (i) = T πin for all i = 0, . . . , n − 1 .

Let M be the monad corresponding to this adjoint situation. We know that M

is finitary and the values at n ∈ N are

M n = T (n, 1) .

The theory (SetfM )op thus has as morphisms from n to k precisely all functions

p : k → T (n, 1) .

This k-tuple of morphisms defines a unique morphism

Ip ∈ T (n, k)

characterized by
T πik · Ip = p(i) (i = 0, . . . , k − 1) .
We obtain an isomorphism of categories

I : (SetfM )op → T

and it remains to prove that the triangle

N>
f op
(JM ) vv >>
vvv >>T
vv >>
v{ v >
(SetfM )op I
/T

commutes. Given a morphism u ∈ T (n, k), then T u is the unique morphism

with
T πik · T u = T πu(i)
n

for all i. And we have

f op
T πin · (I · (JM ) (u)) = T (πik )I(ηn · u) = ηn · u(i) = πu(i)
n
.

November 20, 2009 192

APPENDIX A. MONADS

A.39 Remark. The situation with S-sorted theories and S-sorted algebraic
categories is entirely analogous: all the above results translate without problems
from Set to Set S , only Theorem A.21 needs some work:
A.40 Theorem. S-sorted equational categories are up to concrete isomorphism
precisely the categories Set M of Eilenberg-Moore algebras for finitary monads M
on Set S .
Proof. The only difference with respect to the proof of A.21 is that in the
S-sorted case we have to check also that (Set S )M is closed in M -Alg under
directed unions. Let ki : (Xi , xi ) → (X, x) (i ∈ I) be a colimit cocone of a
filtered colimit in M -Alg , where each (Xi , xi ) is an Eilenberg-Moore algebra.
Then (X, x) satisfies [A.6] because the cocone (ki ) is collectively epimorphic:

Xi
ki

ηXi XB idXi
ηX yy
y BB
y BBidX
yy BBB
|yy
M /X
; X x `AA
www AA
w AA
wwwMk k AA
 ww i i

M Xi / Xi
xi

And it satisfies [A.7] because the cocone (M M ki ) is collectively epimorphic

being a colimit cocone (since M M preserves filtered colimits):
µXi
M M XLi / M Xi
LLL w
LMMk Mki www
LLL i w
LL ww
% µX {ww
MMX / MX

Mxi Mx x xi
 
MX /X
cGG
rr9
x
Mki r GG ki
r GG
rrr GG
 rrr G 
M Xi / Xi
xi

A.41 Theorem. The category Th S of S-sorted algebraic theories is equivalent

to the category of finitary monads on Set S .
The proof is completely analogous to the proof in the one-sorted case. We just
observe that the functor J : Set S → (Set S )M can, in case of M finitary, be
restricted to J f : (S ∗ )op → (Set S )fM analogously to A.36.

November 20, 2009 193

 APPENDIX A. MONADS

November 20, 2009 194

Appendix B

Abelian categories

The categories R-Mod of left modules over a unitary ring R are algebraic and
abelian. The aim of the present appendix is to prove that these are the only
one-sorted abelian algebraic categories. We also prove the many-sorted gener-
alization of this result.

B.1 Remark. In the following we use the standard terminology of the theory
of abelian categories:

1. A zero object is an object 0 which is initial as well terminal. For two

objects A, B the composite A → 0 → B is denoted by 0 : A → B.

2. A biproduct of objects A and B is a product A × B with the property that

the morphisms

hidA , 0i : A → A × B and h0, idB i : B → A × B

form a coproduct of A and B.

3. A category is called preadditive if it is enriched over the category Ab of

abelian groups, i.e., if every hom-set carries the structure of an abelian
group such that composition is a group homomorphism.

4. In a preadditive category, an object is a zero object iff it is terminal, and a

product of two objects is a biproduct. A preadditive category with finite
products is called additive.

5. A functor F : A → A′ between preadditive categories is called additive if

it is enriched over Ab, i.e., the derived functions A(A, B) → A′ (F A, F B)
are group homomorphisms. In case of additive categories this is equivalent
to F preserving finite products.

6. Finally, a category is called abelian if it is exact and additive.

195
 APPENDIX B. ABELIAN CATEGORIES

B.2 Example. Just as one-object categories are precisely the monoids, one-
object preadditive categories are precisely the unitary rings R. Every left R-
module M defines an additive functor M : R → Ab with M (∗) = M and M r =
r · − : M → M for r ∈ R. Conversely, every additive functor F : R → Ab is
naturally isomorphic to M for M = F (∗).

For a small, preadditive category C, we denote by Add [C, Ab] the abelian
category of all additive functors into Ab (and all natural transformations). The
previous example implies that R-Mod is equivalent to Add [R, Ab].

B.3 Theorem. Abelian algebraic categories are precisely those equivalent to

Add [C, Ab]

for a small additive category C.

Proof. Sufficiency. For every small additive category C, we prove that the
category Add [C, Ab] is equivalent to Alg C. We denote by hom(C, −) : C → Set
and Hom(c, −) : C → Ab the hom-functors. Consider the forgetful functor
U : Ab → Set. Since U preserves finite products, it induces a functor

b = U · − : Add [C, Ab] → Alg C

U

Let us prove that U b is an equivalence functor.

(a) Ub is faithful. This is obvious, because U is faithful.
(b) Ub is full. In fact, we first observe that U b preserves sifted colimits. This
follows from the fact that sifted colimits commute in Ab (as in any algebraic
category, see 2.7) with finite products, and the functor U = hom(Z, −) preserves
sifted colimits.We first verify that Ub is full for morphisms (natural transforma-
tions)
b
α : U(Hom(C, −)) → U bG

where C ∈ obj C and G : C → Ab is additive. By the Yoneda Lemma, for all

X ∈ C and for all x : C → X, we have αX (x) = Gx(a), where a = αC (idC ),
thus, αX is a group homomorphism. Consequently, α lies in the image of U b.
b b
The general case of a morphism β : U F → U G reduces to the previous one by
using the fact that F is a sifted colimit of representables and that U b preserves
sifted colimits. To see that F is a filtered colimit of representables, observe that,
following 4.2, Ub F is a sifted colimit of representables. Now

b F = colim hom(Ci , −) = colim (U · hom(Ci , −)) =

U

b (Hom(Ci , −)) = U (colim Hom(Ci , −)) .

= colim U

This implies F = colim Hom(Ci , −) because U b reflects sifted colimits (since it

preserves sifted colimits and reflects isomorphisms).
b is essentially surjective. We first take the representable functors hom(C, −)
(c) U

November 20, 2009 196

APPENDIX B. ABELIAN CATEGORIES

for C ∈ obj C, which are objects of Alg C. Since C is preadditive, for every object
C ′ the hom-set hom(C, C ′ ) is an abelian group, and hom(C, −) factorizes as
Hom(C,−) U
C / Ab / Set

b
with Hom(C, −) : C → Ab additive. Therefore, U(Hom(C, −)) ≃ hom(C, −).
Once again, the general case follows from the previous one by using the fact
that any C-algebra is a sifted colimit of representable C-algebras and that U b
preserves sifted colimits.
Necessity. Let T be an algebraic theory, and assume that Alg T is abelian.
Since T op embeds into Alg T , T is preadditive (with finite products), and then
it is a small additive category. Following the first part of the proof, Alg T is
equivalent to Add [T , Ab].
B.4 Corollary. Abelian algebraic categories are precisely the additive cocom-
plete categories with a strong generator consisting of perfectly presentable ob-
jects.
In fact, this follows from 6.9 and B.3.
B.5 Remark. In B.3 the condition that C is additive can be weakened: preaddi-
tivity is enough. In fact, let C be a small preadditive category. We can construct
the small and preadditive category Mat (C) of matrices over C as follows:
- Objects are finite (possibly empty) families (Xi )i∈I of objects of C;
- Morphisms from (Xi )i∈I to (Zj )j∈J are matrices M = (mi,j )(i,j)∈I×J of
morphisms mi,j : Xi → Zj in C;
- The matrix multiplication, the identity matrices, and matrix addition,
as well known from Linear Algebra, define the composition, the identity
morphisms and the preadditive structure, respectively.
This new category Mat (C) is additive. Indeed, it has a zero object given by the
empty family, and biproducts ⊕ given by disjoint unions. Let us check that the
obvious embedding C → Mat (C) induces an equivalence between Add [Mat (C), Ab]
and Add [C, Ab]. Indeed, given F ∈ Add [C, Ab], we get an extension F ′ ∈
Add [Mat (C), Ab] in the following way: F ′ (M ) is the unique morphism such
that the following square
L F ′ (M) L
F (Xi ) / F (Zj )
I O J


F (Xi ) / F (Zj )
F (mi,j )

commutes for all (i, j) ∈ I × J, where the vertical morphisms are injections in
the coproduct and projections from the product, respectively. It is easy to verify
that the functor F 7→ F ′ is an equivalence from Add [C, Ab] to Add [Mat (C), Ab].

November 20, 2009 197

 APPENDIX B. ABELIAN CATEGORIES

B.6 Remark. Observe that an object G of an additive, cocomplete category

A is perfectly presentable iff its enriched hom-functor Hom(G, −) : A → Ab
preserves colimits. (Compare with the absolutely presentable objects of 5.8.) In
fact, if G is perfectly presentable, then Hom(G, −) preserves finite coproducts
(because they are finite products) and reflexive coequalizers (because U : Ab →
Set reflects them). This implies that Hom(G, −) preserves finite colimits. In-
deed, given a parallel pair a, b : X ⇉ Z in A, its coequalizer is precisely the co-
equalizer of the reflexive pair (a, idZ ), (b, idZ ) : X + Z ⇉ Z. Finally, Hom(G, −)
preserves arbitrary colimits because they are filtered colimits of finite colimits.
B.7 Example. The group Z is perfectly presentable in Ab. Indeed,

Hom(Z, −) : Ab → Ab

is naturally isomorphic to the identity functor. Observe that Z is of course not

absolutely presentable.
B.8 Corollary. One-sorted abelian algebraic categories are precisely the cate-
gories equivalent to R-Mod for a unitary ring R.
Proof. Following B.3, a one-sorted abelian algebraic category A is of the form
Add [T , Ab] for T a one-sorted additive algebraic theory with objects T n (n ∈
N). Any F ∈ Add [T , Ab] restricts to an additive functor T (T, T ) → Ab, where
the ring T (T, T ) is seen as a preadditive category with a single object. Moreover,
F is uniquely determined by such a restriction, because each object of T is a
finite product of T. Finally, Add [T (T, T ), Ab] is equivalent to T (T, T )-Mod.
B.9 Corollary. Finitary localizations of abelian algebraic categories are pre-
cisely the abelian locally finitely presentable categories.
Proof. Let A be an abelian, locally finitely presentable category. Following
the proof of 18.10, we have that A = Lex T is a finitary localization of Alg T , with
T an additive algebraic theory. Due to B.3, Alg T is equivalent to Add [T , Ab].

November 20, 2009 198

Appendix C

More about dualities for

one-sorted algebraic
categories

Throughout our book we took the “strict view” of what a theory morphism or
a concrete functor or a monadic functor etc. should be. That is, the condition
put on the functor in question was formulated as equality between two func-
tors. There is a completely natural “non-strict view” where the conditions are
formulated as natural equivalences between functors. This has a number of ad-
vantages. For example, we can present a characterization of one-sorted algebraic
categories (see Theorem C.6 below) for which we know no analogous result in
the strict variant. Also the duality between one-sorted algebraic theories and
uniquely transportable one-sorted algebraic categories can be directly derived
from the non-strict version of the biduality 11.39 without using monads. In the
present appendix we shortly mention the non-strict variants of some conepts in
our book.

C.1 Definition.

1. Given concrete categories U : A → K and V : B → K by a pseudo-concrete

functor between them we mean a functor F : A → B such that V · F is
naturally isomorphic to U.

2. Given concrete categories U : A → K and V : B → K by a pseudo-concrete

equivalence between them we mean a functor F : A → B which is at the
same time an equivalence and pseudo-concrete. (Note that any quasi-
inverse of F is necessarily pseudo-concrete.) We then say that (A, U ) and
(B, V ) are pseudo-concretely equivalent.

C.2 Definition. A concrete category (A, U ) on K is pseudo-monadic if there

exists a monad M on K and a pseudo-concrete equivalence A → K M .

199
 APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

C.3 Beck’s Theorem. (Characterization of pseudo-monadic categories) A

concrete category (A, U ), is pseudo-monadic if and only if
(a) U has a left adjoint,
(b) U is conservative,
and
(c) A has coequalizers of all reflexive pairs f, g such that U f, U g have an
absolute coequalizer, and U preserves these coequalizers.
A proof of C.3 can be found in [27], Chapter 4, Section 4. The reader needs just
to observe that the parallel pairs of morphisms used in that proof are reflexive.
C.4 Proposition. For every one-sorted algebraic theory (T , T ), the concrete
category (Alg T , Alg T ) is pseudo-monadic.
Proof. The functor Alg T is algebraic and conservative (11.8), and algebraic
functors are right adjoint and preserve reflexive coequalizers. Following C.3,
(Alg T , Alg T ) is pseudo-monadic.
C.5 Definition. A pseudo-one-sorted algebraic category is a concrete category
over Set which is pseudo-concretely equivalent to Alg T : Alg T → Set for a one-
sorted algebraic theory (T , T ).
C.6 Theorem. (Characterization of pseudo-one-sorted algebraic categories)
The following conditions on a concrete category (A, U ) over Set are equivalent:
1. (A, U ) is pseudo-one-sorted algebraic;
2. A is cocomplete and U is a conservative right adjoint preserving sifted
colimits.
More detailed: let A be a cocomplete category. Given a faithful functor

U : A → Set

with A cocomplete, there exists a one-sorted algebraic theory (T , T ) and an

equivalence functor
E / Alg T
AA
AA x x
AA ≃ xx
U AA
A xx
{xx Alg T
Set
making the above diagram commutative up to natural isomorphism iff U is a
conservative right adjoint preserving sifted colimits.

Proof. The conditions are necessary by 11.8. Let us prove that they are
sufficient: let U : A → Set be as above, with F a left adjoint of U. The set
F = {F n : n ∈ N} is closed in A under finite coproducts and, by 6.16, it is a

November 20, 2009 200

APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

strong generator formed by perfectly presentable objects. Following the proof

of 6.9 (implication 3 ⇒ 1), the functor

E : A → Alg (F op ) , K 7→ A(−, K) : F op → Set

is an equivalence. Moreover, the codomain restriction T : N → F op of the func-

tor F op · YN : N → Aop is a morphism of theories. Let us check that there exists
a natural isomorphism Alg T · E ≃ U : for every A ∈ A and n ∈ N the functor
Alg T · E(A) assigns to n ∈ N the set A(T n, A), and since T n = F (YN (n)), the
adjunction F ⊣ U clearly yields a natural isomorphism A(T n, A) ≃ U A(n).
Factorize T as a functor T ′ : N → T which is the identity on objects and equal
to T on arrows, followed by a functor T ′′ : T → F op which is the identity on
arrows and equal to T on objects. Therefore, T ′′ is an equivalence (because T is
surjective on objects) and (T , T ′ ) is a one-sorted theory. The following diagram
concludes the proof:

E Alg T ′′
A HH / Alg (F op ) / Alg T
HH ≃ s ss
HH Alg T ss
H s
U HHH s s Alg T′
$  yss
Set

C.7 Corollary. Pseudo-one-sorted algebraic categories are up to pseudo-concrete

equivalence precisely the categories Set M of Eilenberg-Moore algebras for finitary
monads M on Set.

In fact, this follows from C.4, C.6 and A.23.

It is easy to extend the biequivalence of Theorem 11.39 to pseudo-one-sorted

algebraic categories.

C.8 Definition. Given one-sorted algebraic theories (T1 , T1 ) and (T2 , T2 ), a

pseudo-morphism
M : (T1 , T1 ) → (T2 , T2 )
is a functor M : T1 → T2 with M · T1 naturally isomorphic to T2 .

C.9 Remark. Pseudo-morphisms preserve finite products.

C.10 Theorem. (Non-strict one-sorted algebraic duality) The 2-category of

objects: one-sorted algebraic theories,

1-cells: pseudo-morphisms,

2-cells: natural transformations,

is dually biequivalent to the 2-category of

November 20, 2009 201

 APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

objects: pseudo-one-sorted algebraic categories,

1-cells: pseudo-concrete functors,

2-cells: natural transformations.

Proof. The proof is analogous to that of 11.39, just observe that from a
pseudo-concrete functor

(G, ϕ) : (Alg T2 , Alg T2 ) → (Alg T1 , Alg T1 )

we get a natural isomorphism ψ : FT1 → F · FT2 between the left adjoints. The
rest of the proof of 11.39 can now be repeated with no other changes.

C.11 Remark. In Appendix A we have obtained the duality between the cate-
gory of one-sorted algebraic theories and the category of uniquely transportable
one-sorted algebraic categories using finitary monads. We are going to derive
such a duality from the biequivalence of Theorem C.10. In order to perform this
we need several preliminary steps. We start by an observation explainin why
when we restrict our attention to uniquely transportable one-sorted algebraic
categories we do not need a non-strict version.

C.12 Lemma. Let (A1 , U1 ), (A2 , U2 ) be concrete categories over K.

1. If (A2 , U2 ) is transportable, then for every pseudo-concrete functor

G : (A1 , U1 ) → (A2 , U2 )

there exists a concrete functor H : (A1 , U1 ) → (A2 , U2 ) naturally isomor-

phic to G.

2. If (A1 , U1 ) and (A2 , U2 ) are transportable, then they are pseudo-concretely

equivalent if and only if they are concretely equivalent.

Proof. 1. Let ϕ : U1 → U2 · G be a natural isomorphism. For every object

A ∈ A1 , consider the isomorphism

ϕA : U1 A → U2 GA .

Since (A2 , U2 ) is transportable, there exists an object HA in A2 and an isomor-

phism
ψA : HA → GA
such that U2 ψA = ϕA . This gives a map on objects H : A1 → A2 . For f : A → B
−1
in A1 , put Hf = ψB · Gf · ψA . In this way H is a functor such that U2 · H = U1 ,
and ψ : H → G is a natural isomorphism.
2. follows immediately from 1.

C.13 Corollary. Every transportable pseudo-one-sorted algebraic category is

one-sorted algebraic.

November 20, 2009 202

APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

Recall from 11.36 the 2-categories Th 1 of one-sorted algebraic theories, and

ALG 1 of one-sorted algebraic categories. In the remaining part of this ap-
pendix we consider Th 1 and ALG 1 as categories, that is, we forget the 2-cells.
Explicitely:
C.14 Definition. We define
1. the category Th 1 of one-sorted theories to have
objects: one-sorted algebraic theories,
morphisms: morphisms of one-sorted algebraic theories,
2. the category ALG 1 of one-sorted algebraic categories to have
objects: one-sorted algebraic categories,
morphisms: concrete functors.
3. the category ALG 1u to be the full subcategory of ALG 1 of all uniquely
transportable one-sorted algebraic categories.
C.15 Lemma. The category Th 1 (seen as a 2-category with only identity 2-
cells) is biequivalent to the 2-category PsTh 1 having
objects: one-sorted algebraic theories (T , T ),
1-cells from (T1 , T1 ) to (T2 , T2 ) : pairs (M, µ) with M : T1 → T2 a pseudo-
morphism of one-sorted theories and µ : M · T1 → T2 a natural isomor-
phism,
2-cells from (M, µ) to (N, ν) : natural transformations α : M → N which
are coherent, i.e., such that ν · αT1 = µ.
Proof. (1) Let us start by observing that the coherence condition ν · αT1 = µ
on a 2-cell α of PsTh 1 immediately implies that α is invertible and that between
two parallel 1-cells of PsTh 1 there is at most one 2-cell.
(2) The inclusion Th 1 → PsTh 1 is a biequivalence: since Th 1 and PsTh 1 have
the same objects, we have to prove that the induced functor
Th 1 ((T1 , T1 ), (T2 , T2 )) → PsTh 1 ((T1 , T1 ), (T2 , T2 ))
is full and essentially surjective (it is certainly faithful because Th 1 has only
identity 2-cells).
(2a) Full: let M, N : (T1 , T1 ) → (T2 , T2 ) be 1-cells in Th 1 and α : (M, =) →
(N, =) a 2-cell in PsTh 1 . The coherence condition gives αn = id for every
n ∈ N , and then M = N by naturality of α.
(2b) Essentially surjective: consider a 1-cell (M, µ) : (T1 , T1 ) → (T2 , T2 ) in
PsTh 1 . We define a functor
N : T1 → T2 , N (f : x → y) = µy · M f · µ−1
x : x → y.

It is easy to check that N · T1 = T2 and that µ : (M, µ) → (N, =) is a 2-cell in

PsTh 1 .

November 20, 2009 203

 APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

C.16 Lemma. The category ALG 1u (seen as a 2-category with only identity
2-cells) is biequivalent to the 2-category PsALG 1 having
objects: pseudo-one-sorted algebraic categories (A, U ),
1-cells from (A1 , U1 ) to (A2 , U2 ) : pseudo-concrete with G : A1 → A2 a
functor and ϕ : U1 → U2 · G a natural isomorphism,
2-cells from (G, ϕ) to (H, ψ) : natural transformations α : G → H which
are coherent, i.e., such that U2 α · ϕ = ψ.
Proof. (1) Let us start by observing that, since U2 is conservative, the coher-
ence condition U2 α·ϕ = ψ implies that α is invertible and, since U2 is faithful, it
implies that between two parallel 1-cells of PsALG 1 there is at most one 2-cell.
(2) The inclusion ALG 1u → PsALG 1 is essentially surjective (in the sense of
the 2-category PsALG 1 ) : let (A, U ) be an object in PsALG 1 . We are going to
construct the diagram
E1
AB / A1 E2 / A2
BB {{
BB U1 {{
U BB
B {{
 }{{ U2
Set
where E1 and E2 are pseudo-concrete equivalences, (A1 , U1 ) is transportable,
and (A2 , U2 ) is uniquely transportable. Therefore, (A2 , U2 ) is a uniquely trans-
portable pseudo-one-sorted algebraic category. By Corollary C.13 we conclude
that (A2 , U2 ) is an object of ALG 1u .
(2a) Objects in A1 are triples (A, πA,X : U A → X, X) with A ∈ A, X a set
and πA,X an isomorphism. A morphism from (A, πA,X , X) to (A′ , πA′ ,X ′ , X ′ )
is a pair of morphism a : A → A′ , x : X → X ′ such that x · πA,X = πA′ ,X ′ · U a.
Clearly the forgetful functor

U1 : A1 → Set , U1 (A, πA,X , X) = X

is transportable. Moreover, we have an equivalence

E1 : A → A1 , E1 A = (A, idUA , U A)

with quasi-inverse the forgetful functor

E1′ : A1 → A , E1′ (A, πA,X , X) = A .

Note that E1 is a concrete functor, whereas E1′ is pseudo-concrete: a natural

isomorphism π : U · E1′ → U1 is given by

π(A,πA,X ,X) = πA,X .

(2b) Consider any concrete category U1 : A1 → Set. Objects of A2 are equiva-

lence classes of objects of A1 , with X equivalent to X ′ if there exists a U1 -identity
i : X → X ′ , that is, an isomorphism i such that U1 i = idU1 X . We denote by

November 20, 2009 204

APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

[X] the equivalence classe of X. The hom-set A2 ([X][Z]) is the quotient of the
disjoint union of the A1 (X ′ , Z ′ ) for X ′ ∈ [X], Z ′ ∈ [Z], with f : X → Z equiv-
alent to f ′ : X ′ → Z ′ if there exist U1 -identites i : X → X ′ and j : Z → Z ′
and U1 f = U1 f ′ . The composition of [f ] : [X] → [Z] and [g] : [Z] → [W ] is
[gjf ] : [X] → [W ], where j : Z → Z ′ is any U1 -identity. The functors E2 and U2
are defined by

E2 : A1 → A2 , E2 (f : X → Z) = [f ] : [X] → [Z] ,

U2 : A2 → Set , U2 ([f ] : [X] → [Z]) = F f : F X → F Z .

Clearly, U2 · E2 = U1 and E2 is full and surjective on objects, so that it is
an equivalence (because U1 is faithful). Finally, it is easy to check that U2 is
amnestic if U1 is conservative, and that U2 is transportable if U1 is transportable.
(3) The induced functor

ALG 1u ((A1 , U1 ), (A2 , U2 )) → PsALG 1 ((A1 , U1 ), (A2 , U2 ))

is an equivalence:
(3a) Full: let G, H : (A1 , U1 ) → (A2 , U2 ) be 1-cells in ALG 1u and α : (G, =) →
(H, =) a 2-cell in PsALG 1 . The coherence condition gives U2 (αA ) = id for every
A ∈ A. Since U2 is amnestic, αA is the identity.
(3b) Faithful: obvious because ALG 1 has only identity 2-cells.
(3c) Essentially surjective: let (A1 , U1 ), (A2 , U2 ) be objects in ALG 1u and
(G, ϕ) : (A1 , U1 ) → (A2 , U2 ) a 1-cell in PsALG 1 . As in the proof of C.12 we get a
concrete functor H : (A1 , U1 ) → (A2 , U2 ) and a natural isomorphism ψ : H → G.
To end the proof observe that ψ : (H, =) → (G, ϕ) is a 2-cell in PsALG 1 . Indeed,
the condition U2 (ψA ) = ϕA is precisely the coherence condition on ψ.

Recall the 2-fucntor Alg 1 : (Th 1 )op → ALG 1 from 11.38. There is an obvious
version of this 2-functor in the present context: all we need to observe is that
given a coherent natural transformation α in PsTh 1 , then Alg α is also coherent.
By a slight abuse of notation we denote this 2-functor by Alg 1 again:
C.17 Notation. We denote by

Alg 1 : (PsTh 1 )op → PsALG 1

the 2-functor assigning to every one-sorted theory (T , T ) the one-sorted alge-

braic category (Alg T , Alg T ), to every 1-cell (M, µ) : (T1 , T1 ) → (T2 , T2 ) the
1-cell (Alg M, Alg µ−1 ), and to every 2-cell α : (M, µ) → (N, ν) the 2-cell Alg α
whose component at a T2 -algebra A is A · α : A · M → A · N.
C.18 Theorem. (One-sorted algebraic duality) The category ALG 1u of uniquely
transportable one-sorted algebraic categories is equivalent to the dual of the cat-
egory Th 1 of one-sorted algebraic theories.
Proof. All that remains to be proved after C.15 and C.16 is that the 2-functor

Alg 1 : (PsTh 1 )op → PsALG 1

November 20, 2009 205

 APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

of C.17 is a biequivalence.
(1) The 2-functor Alg 1 is well-defined by 11.8.
(2) The 2-functor Alg 1 is essentially surjective (in the sense of the 2-category
PsALG 1 ) : following C.6, for every object (A, U ) of PsALG 1 there exists a
pseudo-concrete equivalence E : A → Alg T with natural isomorphism

ϕ : Alg T · E → U .

Recall from 0.3 that it is possible to choose a quasi-inverse E ′ : Alg T → A and

natural isomorphisms

η : IdAlg T → E · E ′ and ε : E ′ · E → IdA

such that
Eε · ηE = E and εE ′ · E ′ η = E ′ .
We get a natural isomorphism

ψ = ϕ−1 E ′ · Alg T η : Alg T → U · E ′ .

It follows that

η : (IdAlg T , =) → (E · E ′ , ϕE ′ · ψ) and ε : (E ′ · E, ψE · ϕ) → (IdA , =)

are 2-cells in PsALG 1 . Indeed, the coherence condition on η is just the definition
of ψ, and the coherence condition on ε follows from the equation Eε · ηE = E :

U ε·ψE ·ϕ = U ε·ϕ−1 E ′ E ·Alg T ηE ·ϕ = ϕ−1 ·Alg T Eε·Alg T ηE ·ϕ = ϕ−1 ·ϕ = U .

We conclude that (A, U ) and (Alg T , Alg T ) are equivalent objects in PsALG 1 .
(3) We will prove that for two one-sorted algebraic theories (T1 , T1 ) and (T2 , T2 )
the functor
Alg 1(T1 ,T1 ),(T2 ,T2 )
PsTh 1 ((T1 , T1 ), (T2 , T2 )) / PsALG 1 ((Alg T2 , Alg T2 ), (Alg T1 , Alg T1 ))

is an equivalence of categories.
(3a) Full and faithful: the proof as in Theorem 9.15, indeed α : M → N is
coherent iff Alg α : Alg M → Alg N is coherent.
(3b) Essentially surjective: we follow the proof of 11.39. Let

(G, ϕ) : (Alg T2 , Alg T2 ) → (Alg T1 , Alg T1 )

be a 1-cell in PsALG 1 and ψ : FT2 → F · FT1 the induced natural isomorphism

on the adjoint functors. As in 11.39 we get a 1-cell

(M, =) : (T1 , T1 ) → (T2 , T2 )

in PsTh 1 and we have to construct a 2-cell

α : (Alg M, =) → (G, ϕ)

November 20, 2009 206

APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

in PsALG 1 . Since Alg T1 · Alg M = Alg T2 , there exists a natural isomorphism

i : FT2 → M ∗ · FT1 . Let

ψ| : YT2 · M op → F · YT1 , i| : YT2 · M op → M ∗ · YT1

be the restrictions of ψ and i to T1 . By 9.3, there exists a natural isomorphism

α∗ : M ∗ → F. Moreover, α∗ is unique with the condition α∗ YT1 · i| = ψ| (apply
4.11 to YT1 ). Since FT2 and F ·FT1 preserve colimits, the previous equation gives
α∗ FT1 · i = ψ (apply 4.11 to YN ). Passing to the right adjoints, we get a natural
isomorphism α : Alg M → G such that (Alg T1 )α = ϕ, as required.

November 20, 2009 207

 APPENDIX C. MORE ABOUT DUALITIES FOR ONE-SORTED
ALGEBRAIC CATEGORIES

November 20, 2009 208

Bibliography

[1] J. Adámek, Free algebras and automata realizations in the language of cat-
egories, Comment. Math. Univ. Carolin. 14 (1974), 589-602.
[2] J. Adámek, Colimits of algebras revisited, Bull. Austral. Math. Soc. 17
(1977), 433-450.
[3] J. Adámek, F. Borceux, S. Lack and J. Rosický, A classification of accessible
categories, J. Pure Appl. Algebra 175 (2002), 7-30.
[4] J. Adámek, H. Herrlich and G. Strecker, Abstract and Concrete Categories,
Dover Publications 2009.
[5] J. Adámek, F. W. Lawvere and J. Rosický, On the duality between varieties
and algebraic theories, Alg. Univ. 49 (2003), 35-49.
[6] J. Adámek, F. W. Lawvere and J. Rosický, How algebraic is algebra?, Theory
Appl. Categ. 8 (2001), 253-283.
[7] J. Adámek and J. Rosický, Locally Presentable and Accessible Categories,
Cambridge University Press 1994.
[8] J. Adámek and J. Rosický, On sifted colimits and generalized varieties, The-
ory Appl. Categ. 8 (2001), 33-53.
[9] J. Adámek and J. Rosický, Toward a characterization of algebraic exactness,
J. Algebra 272 (2004), 730-738.
[10] J. Adámek, J. Rosický and E. M. Vitale, On algebraically exact categories
and essential localizations of varieties, J. Algebra 244 (2001), 450-477.
[11] J. Adámek, J. Rosický and E. M. Vitale, What are sifted colimits?, preprint
2009.
[12] J. Adámek, M. Sobral and L. Sousa, Morita equivalence for many-sorted
algebraic theories, J. Algebra 297 (2006), 361-371.
[13] M. Artin, A. Grothendieck and J. L. Verdier, Théorie des topos at coho-
mologie étale des schémas, Lect. Notes in Math. 269, Springer-Verlag 1972.
[14] B. Banaschewski, Functors into categories of M -sets, Abh. Mat. Sem. Univ.
Hamburg 38 (1972), 49-64.

209
BIBLIOGRAPHY BIBLIOGRAPHY

[15] M. Barr, Coequalizers and free triples, Math. Z. 116 (1970), 307-322.
[16] M. Barr, A. P. Grillet and D. H. van Osdol, Exact Categories and Categories
of Sheaves, Lect. Notes in Math. 236, Springer-Verlag 1971.
[17] M. Barr and C. Wells, Toposes, Triples and Theories, Springer-Verlag 1985.
[18] M. Barr and C. Wells, Category Theory for Computing Science, Prentice
Hall 1990.
[19] H. Bass, Algebraic K-theory, Benjamin 1968.
[20] A. Bastiani and C. Ehresmann, Categories of sketched structures, Cah.
Top. Géom. Différ. Catég. 13 (1972), 104-214.
[21] J. Bénabou, Structures algébriques dans les catégories, Cah. Top. Géom.
Différ. Catég. 10 (1968), 1-126.
[22] J. Bénabou, Introduction to bicategories, In Reports of the Midwest Cate-
gory Seminar Lect. Notes in Math. 40, Springer-Verlag 1967, 1-77.
[23] D. B. Benson, M. Nečesal and J. Rosický, Algebras over variable theories,
Algebra Universalis 47 (2002), 55-64.
[24] G. Birkhoff, On the structure of abstract algebras, Proc. Cambr. Phil. Soc.
31 (1935), 433-454.
[25] G. Birkhoff and J. D. Lipson, Heterogenous algebras, J. Combin. Theory 8
(1970), 115-133.
[26] J. M. Boardman and R. M. Vogt, Homotopy invariant algebraic structures
on topological spaces, Lect. Notes in Math. 347, Springer-Verlag 1973.
[27] F. Borceux, Handbook of Categorical Algebra, Cambridge Univ. Press 1994.
[28] F. Borceux and E. M. Vitale, On the notion of bimodel for functorial se-
mantics, Appl. Categ. Structures 2 (1994), 283-295.
[29] N. Bourbaki, Théorie des Ensembles, Herrman 1956.
[30] M. Bunge, Categories of Set-valued functors, Dissertation, University of
Pennsylvania 1966.
[31] M. Bunge and A. Carboni, The symmetric topos, J. Pure Appl. Algebra
105 (1995), 233-249.
[32] A. Carboni and R. Celia Magno, The free exact category on a left exact
one, J. Austral. Math. Soc. Ser. A 33 (1982), 295-301.
[33] A. Carboni and E. M. Vitale, Regular and exact completions, J. Pure Appl.
Algebra 125 (1998), 295-116.
[34] C. Centazzo, J. Rosický and E. M. Vitale, A characterization of locally
D-presentable categories, Cah. Top. Géom. Différ. Catég. 45 (2004), 141-146.

November 20, 2009 210

BIBLIOGRAPHY BIBLIOGRAPHY

[35] C. Centazzo and E. M. Vitale, A duality relative to a limit doctrine, Theory

Appl. Categ. 10 (2002), 486-497.
[36] C. C. Chang and J. Keisler, Model Theory, North-Holland 1973.
[37] P. M. Cohn, Universal Algebra, Harper and Row 1965.
[38] M. Coste, Localisations, spectra and sheaf representation, In Appl. of
Sheaves, Lect. Notes in Math. 753, Springer-Verlag 1979, 212-238.
[39] Y. Diers, Type de densité d’un sous-catégorie plein, Ann. Soc. Sc. Bruxelles
90 (1976), 25-47.
[40] J. J. Dukarm, Morita equivalence of algebraic theories, Colloq. Math. 55
(1988), 11-17.
[41] C. Ehresmann, Sur les structures algébriques, C. R. Acad. Sci. Paris, 264
(1967), A840-A843.
[42] C. Ehresmann, Catégories structurées. Ann. Sci. École Norm. Sup. (3) 80
(1963), 349-426.
[43] S. Eilenberg, Abstract description of some basic functors, J. Indian Math.
Soc. (N. S.) 24 (1961), 231-234.
[44] S. Eilenberg and J. C. Moore, Adjoint functors and triples, Illinois J. Math.
9 (1965), 381-398.
[45] B. Elkins and J. A. Zilber, Categories of actions and Morita equivalence,
Rocky Mountain J. Math. 6 (1976), 199-225.
[46] P. J. Freyd, Aspects of topoi, Bull. Austral. Math. Soc. 7 (1972), 1-76.
[47] P. Gabriel, Des catégories abéliennes, Bull. Sci. Math. France 90 (1962),
323-448.
[48] P. Gabriel and F. Ulmer, Lokal Präsentierbare Kategorien, Lect. Notes in
Math. 221, Springer-Verlag 1971.
[49] M. Gran and E. M. Vitale, On the exact completion of the homotopy
category, Cahiers Top. Géom. Diff. Cat. 39 (1998), 287-297.
[50] G. Grätzer, Universal Algebra, Van Nostrand 1968.
[51] J. Hagemann and C. Hermann, A concrete ideal multiplication for algebraic
systems and its relation to congruence modularity, Arch. Math. (Basel) 32
(1979), 234-245.
[52] A. Heller, MR0163940, Math. Reviews (1965).
[53] P. J. Higgins, Algebras with a scheme of operators, Math. Nachr.
27(1963/64), 115-132.
[54] M. Hyland and J. Power, The category theoretic understanding of universal
algebra: Lawvere theories and monads, Electron. Notes in Theor. Comput.
Sci. 172 (2007), 437-458.

November 20, 2009 211

BIBLIOGRAPHY BIBLIOGRAPHY

[55] Z. Janelidze, Closedness properties of internal relations I: A unified ap-

proach to Mal’tsev, unital and subtractive categories, Theory Appl. Categ.
16 (2006), 236-261.
[56] G. Janelidze and M. C. Pedicchio, Pseudogroupoids and commutators, The-
ory Appl. Categ. 8 (2001), 408-456.
[57] A. Joyal, Notes on Logoi, preprint 2008.
[58] A. Kurz and J. Rosický, Strongly complete logic for coalgebras, preprint
2007.
[59] G. M. Kelly, Basic Concepts of Enriched Categories, Cambridge Univ. Press
1982.
[60] S. Lack and J. Rosický, Notions of Lawvere theory, Appl. Cat. Structures
(to appear).
[61] C. Lair, Sur le genre d’esquissabilité des catégories modelables (accessibles)
possédant les produits de deux, Diagrammes 35 (1996), 25-52.
[62] J. Lambek, A fixpoint theorem for complete categories, Math. Z. 103
(1968), 151-161.
[63] F. W. Lawvere, Functorial semantics of algebraic theories, Dissertation,
Columbia University 1963.
[64] F. W. Lawvere, Ordinal sums and equational doctrines, In Seminar on
Triples and Categorical Homology Theory, Lect. Notes in Math. 80, Springer-
Verlag 1969, 141-155.
[65] F. W. Lawvere, Core varieties, extensivity, and rig geometry, Th. Appl.
Categ. 20 (2008), 497-503.
[66] F. E. J. Linton, Some aspects of equational categories, In Proc. Conf. Cat-
egorical Algebra La Jolla 1965, Springer-Verlag 1966, 84-94.
[67] F. E. J. Linton, Coequalizers in categories of algebras, In Seminar on
Triples and Categorical Homology Theory, Lect. Notes in Math. 80, Springer-
Verlag 1969, 75-90.
[68] F. E. J. Linton, Applied functorial semantics, II., In Seminar on Triples
and Categorical Homology Theory, Lect. Notes in Math. 80, Springer-Verlag
1969, 53-74.
[69] J.-L. Loday, Generalized Bialgebras and Triples of Operads, Astérisque 320
(2008).
[70] J. Lurie, Higher Topos Theory, Princeton Univ. Press 2009.
[71] S. Mac Lane, Categories for the Working Mathematician, Springer-Verlag
1998.
[72] S. Mac Lane, Natural associativity and commutativity, Rice Univ. Stud.
49 (1963), 38-46.

November 20, 2009 212

BIBLIOGRAPHY BIBLIOGRAPHY

[73] M. Makkai and R. Paré, Accessible categories: The Foundation of Categor-

ical Model Theory, Contemp. Math. 104, AMS 1989.
[74] E. G. Manes, Algebraic Theories, Springer-Verlag 1976.
[75] B. Mitchell, Theory of categories, Academic Press 1965.
[76] K. Morita, Duality for modules and its applications to the theory of rings
with minimum conditions, Sci. Rep. Tokyo Kyoiku Daigaku, Sect. A6 (1958),
83-142.
[77] B. Pareigis, Categories and Functors, Academic Press 1970.
[78] M. C. Pedicchio, A categorical approach to commutator theory, J. Algebra
177 (1995), 647-657.
[79] M. C. Pedicchio and J. Rosický, Comparing coequalizer and exact comple-
tions, Theory Appl. Categ. 6 (1999), 77-82.
[80] M. C. Pedicchio and F. Rovatti, Algebraic categories, In: Categorical Foun-
dations, Cambridge Univ. Press 2004, 269-310.
[81] M. C. Pedicchio and R. J. Wood, A simple characterization of theories of
varieties, J. Algebra 233 (2000), 483-501.
[82] A. Pitts, The lex reflection of a category with finite products, unpublished
notes, 1996.
[83] N. Popesco and P. Gabriel, Caractérisation des catégories abéliennes avec
générateurs et limites inductives exactes, C. R. Acad. Sci. Paris 258 (1964),
4188-4190.
[84] J. Power, Enriched Lawvere theories, Theory Appl. Categ. 6 (1999), 83-93.
[85] C. W. Rezk, Spaces of Algebra Structures and Cohomology of Operads,
Thesis, MIT 1996, http://www.math.uiuc.edu/ rezk
[86] J. Rosický, On homotopy varieties, Adv. Math. 214 (2007), 525-550.
[87] J. Rosický and E. M. Vitale, Exact completion and representations in
abelian categories, Homology Homotopy Appl. 3 (2001), 453-466.
[88] J. E. Roos, Caractérisation des catégories qui sont quotients des modules
par des sous-catégories bilocalisantes, C. R. Acad. Sci. Paris 261 (1965), 4954-
4957.
[89] R. Street, Elementary cosmoi I, In Category Seminar Sydney, Lect. Notes
Math 420, Sprinegr-Verlag 1974, 75-103.
[90] R. Street and R. F. C. Walters, Yoneda structures on 2-categories, J. Al-
gebra 50 (1978), 350-379.
[91] W. Tholen, Variations on the Nullstellensatz, unpublished manuscript.
[92] F. Ulmer, Properties of dense and relative adjoints, J. Algebra 8 (1968),
77-95.

November 20, 2009 213

BIBLIOGRAPHY BIBLIOGRAPHY

[93] E. M. Vitale, Localizations of algebraic categories, J. Pure Appl. Algebra

108 (1996), 315-320.
[94] E. M. Vitale, Localizations of algebraic categories II, J. Pure Appl. Algebra
133 (1998), 317-326.
[95] A. A. Voronov, Notes on universal algebra, In Proc. Symp. Pure Math 73,
AMS 2005, 81-103.
[96] C. E. Watts, Intrinsic characterizations of some additive functors, Proc.
Amer. Math. Soc. 11 (1960), 5-8.
[97] W. Wechler, Universal Algebra for Computer Science, Springer-Verlag
1992.
[98] G. C. Wraith, Algebraic Theories, Aarhus Lecture Note Series 22 (1970).

November 20, 2009 214

Index

ALG, 74 Lex, 77
ALG1u , 203 Lex T , 36
ALG1 , 100, 203 M -set, 115
ALG S , 126 M on, 26
ALGcolim , 139 P os, 22
Ab, 9 P sALG1 , 204
Add, 196 P sT h1 , 203
Alg, 74 R-Mod, 10
Alg M , 71 RGraph, 93
Alg T , 90 Rec C, 57, 164
Alg T , 7 S ∗ , 123
Alg 1 , 100, 205 Set, 1
Bool, 51 Set S , 9
EIc , 66 Set M , 184
EInd , 35 SetfM , 191
ERec , 57, 165 Sind C, 35
ESind , 35 TΣ , 112
ETh , 8 T h, 74
ED , 33 T h1 , 100, 203
El A, 5 T h S , 126
Ex, 150 T hc , 74
F M on, 191 T hBim , 139
FH , 106 UH , 103
FT , 94 UΣ , 111
FΣ , 106, 111 UM , 180
FM , 182 YC , 2
Graph, 11 ∆, 5
H-Alg , 21, 103 Γ, 150, 158
HΣ , 21 ΦA , 5
Ic C, 66 Σ, 10
Ind C, 35 Σ-Alg , 11
f
JM , 191 M, 177
JM , 188 A C, 2
KM , 188 Af p , 54
LEX, 77 App , 51
LF P , 77 C(T , T ), 112
Lco, 150 C op , 2

215
INDEX INDEX

G, 47 category of algebras for a theory, 7

K ↓ K, 5 category of algebras for an endofunc-
K M , 180 tor, 21
KM , 188 chain, 15
N, 9 co-wellpowered category, 25
T, 7 cocomplete category, 2
T / ∼, 82 comparison functor, 182
T [k] , 134 complete category, 2
TΣ , 112 concrete category, 6
Tab , 9 concrete equivalence of categories, 91
TC , 8 concrete functor, 6
uT u, 134 concrete isomorphism of categories, 117
P ′ ex , 159 concretely equivalent categories, 92
Pex , 150, 157 concretely isomorphic categories, 117
⊣, 4 congruence on a theory, 80
πin , 89 congruence on an algebra, 97
(F ↓ G), 5 connected category, 17
(A, U ), 6 conservative functor, 1
(T , T ), 90 cospan, 18
2-category, 73
2-functor, 73 dense, 49
diagram, 2
abelian category, 195 directed colimit, 16
absolute coequalizer, 183 directed graph, 11
absolutely presentable object, 41 directed union, 17, 43
additive category, 195
additive functor, 195 effective equivalence relation, 27
Adjoint Functor Theorem, 4 Eilenberg-Moore algebra, 180
adjoint functors, 4 Eilenberg-Moore category, 180
algebra for a signature, 10, 126 enough regular projective objects, 42
algebra for a theory, 7 equation, 79, 114, 128
algebra for an endofunctor, 21, 103 equational category, 114
algebraic category, 7 equational theory, 114
algebraic functor, 72 equivalence of categories, 1
algebraic theory, 7 equivalence relation, 27
amnestic concrete category, 116 equivalent objects in a 2-category, 73
arity, 10 essentially surjective functor, 1
essentially unique extension, 33
bicategory, 139 exact category, 27
biequivalence of 2-categories, 74 exact functor, 144
bimodule, 139 extremal epimorphism, 6
biproduct, 195
Birkhoff’s Variety Theorem, 85 faithful functor, 1
fibration, 117
canonical algebraic theory, 69 filtered category, 3, 15
category of algebras for a signature, 11 filtered colimit, 15

November 20, 2009 216

INDEX INDEX

filtered colimit in Set, 3 kernel pair, 5

final functor, 17 Kleisli category, 188
finitary functor, 21
finitary localization, 173 left covering functor, 145
finitary monad, 178 localization, 173
finite S-sorted set, 9 locally finitely presentable category, 53
finite diagram, 2
finitely generated Σ-algebra, 119 matrix theory, 134
finitely generated category, 2 monad, 177
finitely generated free algebra, 94, 125 monad induced by an adjunction, 179
finitely generated object, 44 monad morphism, 186
finitely presentable Σ-algebra, 119 monadic category, 183
finitely presentable object, 39 Morita equivalent algebraic theories, 133
free Σ-algebra, 111 Morita equivalent monoids, 138
Morita equivalent rings, 133
free H-algebra, 105
morphism of S-sorted algebraic theo-
free algebra, 94, 125
ries, 124
free completion under D-colimits, 33
morphism of algebraic theories, 71
free completion under colimits, 33
morphism of one-sorted algebraic the-
free completion under filtered colimits,
ories, 90
37
morphism of signatures, 112
free completion under finite products,
8 one-sorted algebraic category, 92
free completion under reflexive coequal- one-sorted algebraic theory, 90
izers, 57 one-sorted signature, 10
free completion under sifted colimits,
35 perfectly presentable object, 39
free Eilenberg-Moore algebra, 181 polynomial functor, 21
free exact completion, 150 preadditive category, 195
free monad, 178 pseudo-concrete equivalence, 199
free theory on a signature, 113 pseudo-concrete functor, 199
full functor, 1 pseudo-concretely equivalent categories,
199
generator, 47 pseudo-monadic category, 199
pseudo-morphism of one-sorted theo-
homomorphism of Σ-algebras, 10 ries, 201
homomorphism of H-algebras, 21 pseudo-one-sorted algebraic category, 200
homomorphism of algebras, 7 pseudoequivalence, 146
pseudoinvertible morphism, 134
idempotent complete category, 65
idempotent completion, 66 quasi-inverse functor, 1
idempotent modification, 134 quotient functor, 109
idempotent morphism, 16 quotient object, 5
initial H-algebra, 103
isomorphism of categories, 1 reflective subcategory, 4
reflexive coequalizer, 23
kernel congruence, 81 reflexive pair of morphism, 23

November 20, 2009 217

INDEX INDEX

reflexive relation, 26 zero object, 195

regular epimorphism, 6
regular epireflective subcategory, 84
regular factorization, 25
regular factorization of a parallel pair,
147
regular projective cover, 144
regular projective object, 39
relation, 26
representable functor, 4
retract, 5

S-sorted algebraic category, 125

S-sorted algebraic theory, 124
S-sorted equational category, 129
S-sorted set, 9
S-sorted signature, 126
sifted category, 15
sifted colimit, 15
signature, 10
slice category, 5
solution set, 4
split epimorphism, 5
splitting of an idempotent, 65
strong epimorphism, 6
strong generator, 47
subalgebra generated by a set, 91
subobject, 6
symmetric relation, 26

term for a signature, 111

transitive relation, 26
transportable concrete category, 116
tree, 104

unary theory, 138

uniquely transportable concrete cate-
gory, 116

variety, 81

weak limit, 144

well-powered category, 171
word, 8

Yoneda embedding, 2
Yoneda Lemma, 2

November 20, 2009 218