Content uploaded by Robin Cockett
Author content
All content in this area was uploaded by Robin Cockett on Jan 05, 2016
Content may be subject to copyright.
Theory and Applications of Categories, Vol. 26, No. 17, 2012, pp. 412–452.
RANGE CATEGORIES I: GENERAL THEORY
J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
Abstract. In this two-part paper, we undertake a systematic study of abstract partial
map categories in which every map has both a restriction (domain of definition) and a
range (image). In this first part, we explore connections with related structures such
as inverse categories and allegories, and establish two representational results. The
first of these explains how every range category can be fully and faithfully embedded
into a category of partial maps equipped with a suitable factorization system. The
second is a generalization of a result from semigroup theory by Boris Schein, and says
that every small range category satisfying the additional condition that every map is
an epimorphism onto its range can be faithfully embedded into the category of sets and
partial functions with the usual notion of range. Finally, we give an explicit construction
of the free range category on a partial map category in terms of certain types of labeled
trees.
1. Introduction
1.1. Background and Motivation The pervasiveness of partiality both in mathemat-
ics and in theoretical computer science has led several researchers to develop categorical
approaches to partial maps. For the present purposes, the work by Di Paola and Heller on
dominical categories [Di Paola & Heller 1987], which was aimed at a categorical treatment
of classical computability theory and G¨odel’s incompleteness theorems, is particularly rel-
evant, as it was the first approach in which the partiality of a map was captured by an
idempotent rather than a subobject. This prompted Robinson and Rosolini to define a
class of abstract partial map categories called P-categories [Robinson & Rosolini 1988];
furthermore, they proved representational results for those categories in terms of tradi-
tional partial map categories (by which we mean categories where the morphisms are
spans (m, f) where mbelongs to a suitable class of monics; see Section 4 for details.) For
a more detailed account of the history of categories of partial maps, as well as the rela-
tions to semigroup theory, we refer the reader to the introduction of the paper [Cockett
& Manes 2009].
In 2002, Cockett and Lack proposed a more general axiomatization of partial map
categories, called restriction categories. The characteristic feature of restriction categories
is that the partiality of a map f:A→Bis captured by an idempotent f:A→Aon A;
Cockett and Hofstra both acknowledge support through the NSERC Discovery Grant program.
Received by the editors 2011-08-31 and, in revised form, 2012-09-05.
Transmitted by Stephen Lack. Published on 2012-09-09.
2010 Mathematics Subject Classification: 18A15, 18A32, 18D20.
Key words and phrases: Categories of partial maps, restriction category, factorization systems.
c
J.R.B. Cockett, Xiuzhan Guo and Pieter Hofstra, 2012. Permission to copy for private use granted.
412
RANGE CATEGORIES I: GENERAL THEORY 413
in the case of sets and partial functions, this map fwould be defined as
f(x) = xif x∈dom(f)
undefined otherwise.
The behaviour of the operator f7→ fcan then be axiomatized by means of four
equations (see Section 2). What sets apart restriction categories from P-categories, and
hence in particular from the more special dominical categories, is that the definition
doesn’t require products; and unlike traditional partial map categories no pullbacks are
required to define composition. As such, restriction categories are both economical and
general.
Further developments include the study of partial map categories with additional
categorical structure or properties. For example, extensive categories, and in particular
the extensive completion of a category appear in [Cockett 1993], while partial cartesian
closed categories were considered by in [Curien & Obtulowitz 1989]. An improved setting
for the study of abstract computability theory was put forward in [Cockett & Hofstra
2007]; furthermore, a term logic for partial cartesian categories was presented in [Cockett
& Hofstra 2010].
In this paper we continue the development of the theory of partial map categories
with additional categorical structure by studying the class of range categories: these are
categories of partial maps in which every morphism has a range, or image. The category
of sets and partial functions is a typical such category, and in the category of spaces and
partial continuous functions with open domain the maps with range are precisely1the
open maps. Another motivating example is the category of partial recursive functions;
both the domain and range of such a function are recursively enumerable. However, other
similar categories, obtained by replacing the standard recursion theoretic model by a more
general model of partial combinatory logic, need not have enumerable ranges, and hence
this is a non-trivial extra property of a model of computation. Finally, an important class
of examples comes from semigroup theory: every inverse monoid is a range category, as
is the idempotent splitting of an inverse semigroup.
Our interest in range categories and our motivation for developing the basic theory as
reported on in these papers is at least threefold. First, we believe the theory is of intrinsic
interest, much in the same way as the theory of allegories is interesting in its own right.
It has a simple algebraic axiomatization, is more general than other classes of categories
which aim at modeling partial maps with images, and yet captures all of the motivating
examples.
Second, the theory of restriction categories in general and of more specific classes
such as range categories can play a bridging role between (inverse) semigroup theory and
category theory. Various concepts and constructions in semigroup theory, for example
certain kinds of (co)completions, can be carried out on the level of restriction categories
(we shall see several examples in the present papers, but the reader may consider [Cockett
1Assuming a mild separation axiom; see Section 3 for details.
414 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
& Manes 2009] for others); this means that restriction categories play an organizational
role, bringing out the categorical content of semigroup-theoretic results. Since restriction
categories are much closer in spirit to (inverse) semigroups than partial map categories are,
understanding the categorical contents of semigroup theory is much easier when working
with restriction categories.
And third, in the development of computability theory within the setting of [Cockett
& Hofstra 2007] it rapidly became apparent that it would be highly beneficial to have
available a logic which would allow for standard computability-theoretic definitions and
arguments to be interpretable in a suitable class of categories. This class of categories is
that of discrete cartesian range categories, and the term logic, which will be described in
Part II, is precisely the internal language of this class of categories. Consequently, (and
this was in fact the immediate catalyst for the work) a clear understanding of this class
(which is slightly larger than that of regular categories) was required.
1.2. Related work Various aspects of ranges and related concepts in the categorical
context have already been considered in varying levels of detail in the literature. We
mention here some of those, while throughout the text we elaborate on the connections
with the present work more precisely.
The work which is closest (both in spirit and in terms of actual results) to ours is
that of Rosolini, who introduces ranges in the setting of P-categories, notes a connection
with factorization systems, and proves a representational result [Rosolini 1988]. The first
difference with our approach is that we develop ranges without assuming the presence
of finite products (which P-categories have by definition). The second difference is the
approach in loc. cit. imposes a non-algebraic axiom which we do not take as part of
our basic axiomatization. Therefore, the basic theory of ranges which we develop in this
paper is more general; we recover the concepts from P-categories by adding subsequent
layers of structure/properties to our setting. In particular, a restriction category-theoretic
analogue of Rosolini’s representation theorem is presented in part II of this paper.
In [Di Paola & Heller 1987] the authors consider recursion categories with images.
These can be seen as a special class of P-categories with ranges, and thus are again more
special than the range categories we consider here. Their goal, namely to give a categorical
treatment of recursion theory, was also one of our motivations; however, Di Paolo and
Heller did not develop the theory of ranges in much detail, instead they moved straight
to recursion-theoretic applications.
Another related piece of work which should be mentioned here is the paper [Hughes
& Jacobs 2002] on the connection between factorization systems and fibrations; the au-
thors show that every factorization system gives rise to a bicomplete fibration, and that
this fibration satisfies the Beck-Chevalley condition whenever the factorization system is
pullback-stable. In the present paper, we prove a representational result for range cate-
gories in terms of factorization systems; the close relationship between this result and the
one by Hughes and Jacobs is that every range category comes naturally equipped with
a bifibration (in an appropriate restriction-categorical sense). Furthermore, in part 2 of
this paper the Beck-Chevalley condition for range categories with products is examined.
RANGE CATEGORIES I: GENERAL THEORY 415
Since range categories are related to factorization systems, fibrations and regular cat-
egories, there are close connections between results in these areas and some of the results
about range categories developed in this paper. In fact, once one understands the de-
tails concerning the connections, one could in principle translate parts of the theory of,
say, factorization systems into the language of range categories. While we will attempt
throughout the paper to point out when a particular result could have been obtained in
this manner, we believe there are reasons to develop the theory from the perspective of
range categories.
The primary reason is that our aim is to give a coherent and self-contained presen-
tation which does not heavily rely on other results. While it would indeed have been
possible in certain cases to take a shortcut by referring to facts found in the literature, we
usually prefer to give direct proofs which exemplify and illustrate reasoning about range
categories. For example, one can deduce from the fact that one may freely add equalizers
to a category that, as long as one is willing to split idempotents, each restriction category
has a meet completion. However, this does not give a concrete description of what the
resulting category then looks like, and doesn’t explain how the construction relates to the
meet-completion in semigroup theory. Our direct construction (Part II, Section 2.5) is
elementary and immediately makes the connection with the semigroup-theoretic variant
clear.
The second reason is that in some cases definitions or results about ranges do not
correspond nicely to well-known results in neighbouring areas. For example, the term
logic for discrete range categories which we present in Part II is quite natural from a
proof-theoretic point of view, but it turns out to be strictly weaker than regular logic.
As such, it doesn’t really correspond to a well-studied fragment of categorical logic in
the total world. Moreover, the inference rule which would have to be added to the logic
in order to bridge the gap makes the system much less transparent. Closely related is
the fact that while the regular completion of a category with finite limits is a simple and
elegant construction, its partial analogue is a lot more complicated and is best viewed
as a two step process, the first of which freely adds ranges, making the corresponding
factorization system proper and pullback-stable, and the second of which forces the this
factorization system to be the regular epi-mono factorization by means of a category of
fractions construction. Because of such tensions, we prefer to give our primary attention
to the natural unfolding of the theory on the level of restriction categories, only to focus
on connections with related theory after that.
1.3. Contributions The main focus of the first part of this paper is on describing the
fundamental structures and concepts involved, on developing the material in a systematic
and reasonably self-contained fashion, and on generalizing and unifying various strands
of research in this direction. The second part will deal with the interaction of ranges with
other types of structure (in particular, the partial analogue of finite limits) as well as a
term logic for reasoning about such categories.
The main contributions of the present paper are the following:
1. A thorough and systematic development of the basic theory of categories with
416 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
ranges, some examples, and their positioning with respect to other classes of cate-
gories such as allegories, categories with factorization systems, and fibrations with
existential quantification.
2. A representation theorem for range categories, which says that every range category
can be fully embedded into the partial map category associated to a category with
a suitable factorization system, and moreover that every range category in which
idempotents split is of that form.
3. A generalization of a result from semigroup theory by Boris Schein, which says that
every small range category satisfying the additional condition that every morphism
is an epimorphism onto its range can be embedded faithfully (but generally not
fully) in the category of sets and partial functions.
4. A direct description of the free range category on a category or on a restriction
category.
1.4. Outline The paper is organized as follows. First, we consider categories equipped
with a support operator; such an operator captures the idea that to every map we can
associate a domain, in the form of an idempotent. This setting is precisely sufficient
to specify the notion of an open map, and we may then consider categories where all
maps are open. Such categories then turn out to have a cosupport, or range, operator
which is compatible with the support operator. We then develop the elementary theory
of range categories and open maps, and establish various characterizations of ranges,
both equationally and in terms of fibrations. We also prove a technical result needed for
the representation theorem in Section 4, stating that range categories are stable under
idempotent splitting.
Section 3 explores some examples of range categories and connections with more fa-
miliar concepts. Every allegory (abstract category of relations) admits a support and
cosupport operator, and we show how there is a maximal subcategory which is a range
category, namely by selecting the “simple maps” (single-valued relations). Next, we con-
sider categories satisfying the axiom of choice, which in this setting means that every
morphism admits a partial section. Such a category always admits a range operator. As
a special case, we consider inverse categories; these are a generalization of inverse semi-
groups, and have the defining property that every map has a partial inverse. Finally, we
consider ranges and open maps in the context of topological spaces and locales.
Next, we turn to the basic representational results. Sections 4.1– 4.8 show how split-
ting of idempotents of a range category gives a category whose underlying category of
total maps admits a factorization system, and conversely how every suitable factorization
system gives rise to a split range category. Section 4.11 is concerned with a generalization
of Boris Schein’s result about semigroups with domains and ranges (as explained in Jack-
son and Stokes [Jackson & Stokes 2009]); we prove that every range category satisfying
an additional condition can be embedded into sets and partial functions via a faithful
functor which preserves domain and range.
RANGE CATEGORIES I: GENERAL THEORY 417
Finally, Section 5 is devoted to a construction of the free range category on a restriction
category. The key technical idea is that suitable equivalence classes of labeled trees can
be used to represent the domains and ranges in the free category. The main result is then
that the construction is a left adjoint to the forgetful functor from range categories to
restriction categories.
Acknowledgements We would like to express our gratitude towards the anonymous
referee, whose insightful comments, questions and suggestions have improved the quality
of the paper.
2. Basic Theory
This section develops the basic theory of range categories and open maps. While the
presentation is largely self-contained, we only briefly rehearse the main elements of the
theory of restriction categories; the reader is referred to [Cockett & Lack 2002] for a
detailed treatment.
2.1. Support and Restriction The starting point for our investigations is the concept
of a category equipped with a support operator; such an operator captures the idea that
morphisms in such a category have a domain of definition, represented by an idempotent
on the source of the morphism.
Formally, a support (operator) on a category Cis an operation on morphisms
f:X→Y
f:X→X
which satisfies the following four equational axioms:
[R.1] ff =f
[R.2] gf =f g whenever dom(f) = dom(g)
[R.3] gf =gf whenever dom(f) = dom(g)
[wR.4] gf =gf whenever cod(f) = dom(g)
The fourth axiom is actually a weakening of the following axiom (which, for reasons
to become clear later, is called the axiom of determinism):
[R.4] fgf =gf whenever cod(f) = dom(g)
If a support operators satisfies [R.4] then we shall call it a restriction operator, and a
category equipped with a restriction operator is called a restriction category. The general
418 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
theory of such abstract categories of partial maps was developed in detail in [Cockett &
Lack 2002]; here we shall content ourselves with a brief review of those aspects pertinent
to the present objectives and with some observations concerning the differences between
restrictions and the more general supports.
The map frepresents the domain of f; hence maps for which f= 1 are called
total. Total maps are closed under composition, and identities are total; hence we have a
subcategory Tot(C) of Con the total maps.
We call morphisms fwhich satisfy f=frestriction idempotents. (This terminology
will be used both in general support categories and in restriction categories.) For a fixed
object A, the collection of all restriction idempotents on Ais denoted by O(A). This set
(assuming the ambient category is locally small) is in fact a meet-semilattice with top
element. This structure is given by (for restriction idempotents e, e′):
⊤= 1A
e≤e′⇔ee′=e
e∧e′=ee′
Given a morphism f:B→A, there is an induced function
f∗:O(A)→ O(B); e7→ f∗(e) := ef .
which we refer to as pullback along f. This function always preserves the ordering; it
preserves the top element if and only if fis total. If the support operator is a restriction,
then f∗also preserves binary meets.
Categories with support are automatically order-enriched: given parallel maps f, g :
A→Bset
f≤g⇔f=gf .
The ordering on the idempotent lattice O(A) arises by restriction of the ordering on
C(A, A). Importantly, however, the orderings C(A, B) generally do not have a top element,
nor do they have binary meets.
2.2. Examples.
1. The paradigmatic example of a restriction category is Par, the category of sets and
partial functions. For an object A, we have O(A)∼
=P(A), the full powerset of A.
2. Topological spaces and partial continuous functions with open domain form a re-
striction category, with O(A) the lattice of opens of the space A.
3. Partial recursive functions N→Nform a restriction category (monoid). The re-
striction idempotents correspond to recursively enumerable sets.
4. Every category can be regarded as a restriction category in a trivial way via f= 1
for all f. This shows that a category can be a restriction category in more than
one way, and hence that a restriction is additional structure on a category, not a
property.
RANGE CATEGORIES I: GENERAL THEORY 419
5. The category Rel of sets and relations is supported, but not a restriction category.
One verifies easily that a relation fsatisfies fgf =gf for all gif and only if f
is (the graph of) a partial function. The precise connection between categories of
relations and support categories is explored in Section 3.1.
When Cis a category with support, then a map fis called deterministic if f gf =gf
for all g. Deterministic maps form a subcategory and include all restriction idempotents.
Therefore:
2.3. Proposition. Given a support category C, the subcategory on the deterministic
maps is a restriction category.
For example, in the category Rel, the deterministic maps are precisely the partial
functions. Note also that the order-enrichment of Rel qua support category is not the
same as that of Rel qua allegory: according to the latter, we have f⊆gwhenever the
graph of fis contained in the graph of g, but according to the former, we have f≤g
whenever fcan be obtained from gby restricting it to a subset. Thus we have f≤g
implies f⊆g, but not vice versa. For deterministic maps f, g, the two orderings coincide.
We conclude this section by organizing restriction categories into a 2-category.
A functor F:C→Dbetween two restriction categories is said to be a restriction
functor if F(f) = F(f) for every f. Restriction categories and restriction functors form
a category, denoted by Rcat0. There is an obvious forgetful functor Ur:Rcat0→Cat0
to the category of categories, which forgets the restriction structure. In [Cockett & Lack
2002] a left adjoint was explicitly given.
Given two restriction functors F, G :C→D, a family of morphisms αA:F A →GA is
called a strict natural transformation if it is a natural transformation in the usual sense
and all of its components are total. It is a lax natural transformation if the naturality
squares commute up to inequality:
F A
≤F f
αA//GA
Gf
F B αB
//GB
Restriction categories, restriction functors, and strict natural transformations thus form
a 2-category, denoted Rcat. Similarly, restriction categories, restriction functors and lax
natural transformations form a 2-category denoted Rcatl.
2.4. Open Maps An important example of a restriction category is the category of
topological spaces and partial continuous functions with open domains. Our notation
O(A) for the poset of restriction idempotents is directly motivated by this example, and
we often wish to think of this poset as the poset of open subobjects of A. It is natural
to investigate to which extent topological notions can be given a sensible interpretation
in the context of restriction categories. In this section we carry this out for the notion of
open map. Before doing so, however, we remark that the idea of axiomatizing open maps
420 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
in a category goes back to work by Joyal and others (see for example [Joyal & Moerdijk
1994]); however, the ambient setting in loc. sit. was one of much more richly structured
categories, such as pretoposes.
In topology, a continuous function f:X→Yis called an open map if the direct image
function which sends a subset U⊆Xto f[U]⊆Yrestricts to a map f!:O(X)→ O(Y),
i.e., when the direct image under fof an open set is again open. In this situation, f!is
left adjoint to f−1.
In order to generalize this, consider a category Cwith support, and a morphism
f:A→B. The associated “inverse image” function f∗:O(B)→ O(A) has the property
that it actually lands in O(A)/f : given any e∈ O(B) we have f∗(e) = ef ≤f. Of course,
the meet-semilattice O(A)/f is the principal downset ↓(f)⊆ O(A); categorically, this is
the same thing as the slice over f.
2.5. Definition. [Open map] A morphism f:A→Bin a category with supports is
called open when the poset map f∗:O(B)→ O(A)/f has a left adjoint f!⊣f∗satisfying
the Frobenius identity
f!(e∧f∗(e′)) = f!(e)∧e′
for any two restriction idempotents e∈ O(A)and e′∈ O(B).
We note that the inequality
f!(e∧f∗(e′)) ≤f!(e)∧f!(f∗(e′)) ≤f!(e)∧e′
holds regardless.
There is a slight reformulation of this definition which is sometimes convenient:
2.6. Lemma. In a support category a map f:A→Bis open if and only if there is a
poset morphism ∃f:O(A)→ O(B)such that
[Open. 1] ∃ff∗(e′)≤e′for all e′∈ O(B)
[Open. 2] e∧f∗(e′)≤f∗(∃f(e)∧e′)for all e∈ O(A)and e′∈ O(B)
[Open. 3] e′∧ ∃f(e)≤ ∃f(f∗(e′)∧e)for all e∈ O(A)and e′∈ O(B)
Note that the last inequality gives Frobenius reciprocity as the inequality running in
the other direction is always present.
Proof.
(⇒)Suppose fis open so that f∗has a Frobenius left adjoint f!:O(A)/f → O(B).
Then define the new ∃f(e) := f!(f∧e), so that
∃f(f∗(e′)) = f!(f∧f∗(e′)) = f!(f)∧e′≤e′
e∧f∗(e′) = e∧f∧f∗(e′)≤f∗(f!(f∧e)) ∧f∗(e′)
=f∗(f!(f∧e)∧e′) = f∗(∃f(e)∧e′)
e′∧ ∃f(e) = e′∧f!(f∧e) = f!(f∗(e′)∧f∧e)
=∃f(f∗(e′)∧e).
RANGE CATEGORIES I: GENERAL THEORY 421
(⇐)Note that ∃f(f∧e) = ∃f(e) as
∃f(ef ∧e)≤ ∃f(e) as f∧e≤e
and
∃f(e) = 1B∧f!(e)≤ ∃f(f∗(1B)∧e) = ∃f(f∧e)
Therefore we may view ∃fas acting on the slice O(A)/f. Moreover, we have
e∧f=e∧f∗(1B)≤f∗(∃f(e)∧1B) = f∗(∃f(e)).
The Frobenius condition is immediate.
Given a support category C, we are now interested in the subcategory on the open
maps. which we will denote by Copen. First, we verify that this is indeed a category with
support.
2.7. Lemma. In any support category all restriction idempotents are open maps and open
maps are closed under composition.
Proof. We must establish the three conditions from the previous lemma for a support
idempotent e. This is easy: set ∃e(e′) = ee′. Then:
∃e(e∗(e′)) = eee′≤e′
e′∧e∗(e′′) = e′ee′′ ≤e∗(∃e(e′)∧e′′)
e′∧ ∃e(e′′) = e′ee′′ =∃e(e∗(e′)∧e′′ )
It remains to show that open maps compose: if fand gare open maps which compose
set ∃gf =∃g∃f. We then have:
∃g(∃f(f∗(g∗(e)))) ≤ ∃g(g∗(e)) ≤e
e∧f∗(g∗(e′)) ≤f∗(∃f(e)∧g∗(e′)) ≤f∗(g∗(∃g(∃f(e)) ∧e′)
e′∧ ∃g(∃f(e)) = ∃g(g∗(e′)∧ ∃f(e)) = ∃g(∃f(f∗(g∗(e′)) ∧e))
(where eis an idempotent on dom(f) and e′an idempotent on cod(g)).
2.8. Remark. As an aside we note that restriction functors do not, in general, preserve
open maps. Therefore taking the subcategory on the open maps is not a functorial
construction.
The open maps in a support category Care those which have a well-defined range.
We put ˆ
f:= ∃f(1),
and refer to ˆ
fas the range, or cosupport of the map f. Thus we have an operator of the
following type: f:X→Y
ˆ
f:Y→Y
This operator satisfies the axioms dual to those of a support operator.
422 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
2.9. Lemma. The operator f7→ ˆ
fis a support operator on the category (Copen)op.
Proof.
ˆ
ff =∃f(1A)f=f∃f(1A)f=f f∗(∃f(1A)∧1B)≥1A∧f∗(1A) = f f =f
ˆ
fˆg=∃f(1A)∧ ∃g(1A′) = ∃g(1A′)∧ ∃f(1A) = ˆgˆ
f
c
ˆ
fg =ˆ
∃f(1A)g=∃∃f(1A)g(1A′) = ∃∃f(1A)(∃g(1A′)) = ∃f(1A)∧ ∃g(1A′) = ˆ
fˆg
c
gˆ
f=∃gˆ
f(1B) = ∃g(∃ˆ
f(1B)) = ∃g(ˆ
f) = ∃g(∃f(1A)) = ∃gf (1A) = c
gf .
What is more, this cosupport operator is compatible with the support, in the sense
that ˆ
f=ˆ
f;ˆ
f=f.
A category Cwith a support and a cosupport which are compatible in this sense will
be called a bisupport category. We now have the main observation of this section:
2.10. Theorem. When Cis a support category, then the subcategory on the open maps
Copen is a bisupport category whose support agrees with that of C, and which contains all
restriction idempotents. Moreover, when the support on Cis actually a restriction, then
Copen is the largest bisupport subcategory of C.
In particular, this says that a bisupport category in which the support is a restriction is
the same thing as a restriction category where all maps are open. See also Proposition 2.13.
Proof. We have already established that the open maps form a subcategory with all
the stated properties. It remains to be shown that it is the largest such category when
the support is a restriction. For this note that if a map fis included in a bisupport
subcategory then it must be an open map as one may define ∃f(e) = c
fe. This has the
required properties:
∃f(f∗(e′)) = d
fe′f=c
e′f=e′b
f≤e′
e∧f∗(e′) = e′e′f=e′fe =e′c
fef e ≤e′c
fef =f∗(∃f(e)∧e′)
e′∧ ∃f(e) = e′c
fe =d
e′fe =[
fe′f e =∃f(f∗(e′)∧e)
Hence fis an open map.
It should be noted that the cosupport is hardly ever a corestriction; see Section 3.9 for
the exception. We also point out the the condition that the support is a restriction (which
was used in the first step of the proof) cannot be omitted: in the bisupport category Rel,
a map is open if and only if it is deterministic.
RANGE CATEGORIES I: GENERAL THEORY 423
2.11. Ranges We now turn to our main objects of study, namely range categories. A
range category may be defined as a bisupport category in which the support is a restriction.
Before we turn to an alternative characterization, we observe that given a support operator
on a category there exists at most one compatible cosupport. Indeed, suppose that g
(−)
is a second cosupport combinator; then we compute
ˆ
f=d
e
f f =e
fˆ
f=ˆ
fe
f=g
ˆ
f f =e
f.
Therefore, while a category may have many different support operators, having a compat-
ible cosupport is a property of a support category. (Of course, one may also deduce this
from the unicity of adjoints to pullback functors.) Consequently, being a range category
is a property of a restriction category.
We now give the alternative definition of a range category:
2.12. Definition. [Range Category] A restriction category Cis a range category if it
has an operator f:X→Y
ˆ
f:Y→Y
satisfying the following four axioms:
[RR.1] b
f=b
f
[RR.2] b
ff =f
[RR.3] c
gf =gb
ffor all maps f, g with codom(f) = dom(g)
[RR.4] c
gb
f=c
gf for all maps f, g with codom(f) = dom(g)
So far, we thus have three different ways of describing range categories:
2.13. Proposition. For a restriction category C, the following are equivalent:
(i) Cis a range category
(ii) Chas a compatible cosupport
(iii) all maps in Care open
Proof. We have already shown that the second two conditions are equivalent (The-
orem 2.10). The straightforward calculation that a combinator f7→ ˆ
fa compatible
cosupport if and only if it is a range combinator is omitted.
The following lemma collects some useful basic facts about ranges. The proof is a
straightforward exercise in the use of the axioms and is left to the reader.
424 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
2.14. Lemma. In a range category,
(i) c
ˆ
f g =ˆ
fˆgif codom(f) = codom(g)
(ii) ˆ
f= 1 if fis epi. In particular, ˆ
1 = 1
(iii) b
f=ffor all f
(iv) c
g f ˆg=c
g f if codom(f) = dom(g)
A restriction functor F:C→Dbetween two restriction categories is a range functor
if it preserves not only the restriction but also the range:
F(f) = F(f); F(ˆ
f) = [
F(f).
The latter requirement is not automatic. For example, consider the category of sets,
regarded with trivial restriction (and hence also range) structure. The inclusion into the
category of sets and partial maps preserves the restriction but not the range.
Range restriction categories and range restriction functors form a category, denoted
by RRcat0, which can be enriched with either strict or lax transformations. There are
evident forgetful functors
RRcat0→Rcat0→Cat0,
which forget the restriction and range structures. Since range restriction categories are
defined equationally, we know that U:RRcat0→Cat0is monadic via a finitary monad
so that RRcat0is locally finitely presentable. In Section 5 we give an explicit description
of the free functor.
We note that there is another functor RRcat0→Cat0which sends a range category
to the subcategory on those maps which are total and cototal, i.e. satisfy f= 1,ˆ
f= 1.
The left adjoint to this functor equips a category with the trivial restriction and range.
2.15. Idempotent Splitting We now prove a technical result which will be used in
Section 4, namely that the idempotent splitting of a range category is again a range
category. In [Cockett & Lack 2002] it was already shown that the idempotent splitting of a
restriction category is again a restriction category, and that split restriction categories are
precisely partial map categories arising from systems of monics. Thus here we only focus
on making sure the range structure is also well-behaved with respect to this construction.
2.16. Definition. A restriction idempotent fis split if it can be written as f=mr
with rm = 1. In this case m, the monic part, is called a restriction monic. A restriction
category is said to be split if all of its restriction idempotents split.
If fsplits as f=mr then f=r, because we have f=mr =mr =rsince mis monic,
hence total.
We now consider a range category Cand construct its splitting Split(C) by splitting
the restriction idempotents. Explicitly:
RANGE CATEGORIES I: GENERAL THEORY 425
Objects: restriction idempotents of C,
Maps: a map ffrom (e1:A→A) to (e2:B→B) is given by a map f:A→Bsuch
that e1fe2=f,
Composition: as in C,
Identities: 1e=efor any object eof Split(C).
We observe that Split(C) inherits restriction and ranges from C. To show this, it
suffices to show that the restriction and range of a map are maps of Split(C).
2.17. Lemma. If f:e1→e2is a map of Split(C), then so are f:e1→e1and ˆ
f:e2→e2.
Proof. The equalities
fe1=f e1=fe1=f e1=f,
e1f=fe1=f ,
e2ˆ
f=e2ˆ
f=c
e2f=c
e2f=ˆ
f
and ˆ
fe2=e2ˆ
f=f
show that f:e1→e1and ˆ
f:e2→e2are maps in Split(C), as desired.
This shows that Split(C) is a range category. We record:
2.18. Proposition. If Cis a range category, then so is Split(C)with the split restriction
structure given by the restriction in C.
By the same process in the proof of Proposition 2.27 [Cockett & Lack 2002], we may
regard the 2-category of split range categories RRcatsas a full reflexive sub-2-category
of RRcat: the 2-functor Split :RRcat →RRcats, taking Cto Split(C), is a left adjoint to
the inclusion RRcats֒→RRcat.
We point out that the above proposition also holds when we split all idempotents in
C, not just the restriction idempotents.
3. Examples and Connections
The time is ripe for some examples of range categories. We have already seen that given
a restriction category C, the subcategory of open maps is a range category. In Par, every
map is open, so this is a range category. In the category of spaces and partial continuous
functions with open domain, the open maps which are open in the usual topological sense
are open in our sense. The converse is true provided we impose the T1separation axiom
(points are closed). To see why this is needed, consider the category of Alexandroff spaces;
such spaces are characterized by the fact that for each subset of Xthere exists a smallest
open set containing it. Given a continuous map between such spaces, we may define its
426 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
range to be the smallest open set containing its set-theoretic image. This does define a
range category, but clearly not every map is open in the topological sense.
In the category of partial recursive functions (of one variable, say), every map is open
as well: this may be seen as a consequence of the fact that this category satisfies the
axiom of choice (see Section 3.5). Explicitly, given f, the range of fis the domain of the
function y7→ µx.f(x) = y. (In some texts, the definition of r.e. set is taken to be the
image of a computable function; while this does stress the idea of enumerating a subset,
it does not agree well with the axiomatic categorical approach, according to which one
needs to have domains before one can have ranges.)
In the following sections, we construct more examples by investigating the connections
between range categories and neighbouring areas.
3.1. Allegories Just as restriction categories are abstract categories of partial maps,
allegories ([Freyd & Scedrov 1990]) are abstract categories of relations. For us, two rel-
evant connections between allegories and range categories are that every allegory is a
bisupport category, and that from every allegory one can extract a range category by
taking deterministic maps.
Let us first rehearse the relevant definitions. An involution on a category Cis a
functor (−)◦:C→Cop which is the identity on objects, and which satisfies f◦◦ =ffor
all morphisms f.
3.2. Definition. An allegory is an order-enriched category equipped with an order-
preserving involution. (We denote the ordering on the homsets by ⊆.) Moreover, each
homset is required to have binary meets, and these are to satisfy the modular law
gf ∩h⊆g(f∩g◦h)
Note that this does not imply that the binary meets in the homsets are preserved by
composition. The statement that the involution preserves the ordering means that f⊆g
if (and only if) f◦⊆g◦.
The motivating example of an allegory is the category Rel of sets and relations; more
generally, for any regular category Cwe may construct the allegory Rel(C) of relations in
C.
Each map fin an allegory has a support defined by
Af//B
Af=f◦f∩1A
//A
and dually each map has a cosupport:
Af//B
Bˆ
f=ff ◦∩1B
//B
Towards the verification of the axioms of a (co)support, it is useful to observe the
following:
RANGE CATEGORIES I: GENERAL THEORY 427
3.3. Lemma. In an allegory, if h, h′⊆1then hh′=h′h=h∩h′,hh =h, and h=h◦.
Proof. We show that hh′=h∩h′. To this end, note that hh′⊆h1 = hand hh′⊆1h′=
h′, so certainly hh′⊆h∩h′. Conversely, h∩h′⊆h(1 ∩h◦h′)⊆h(1 ∩h′)⊆hh′. All of
the desired equalities follow from this.
We now show:
3.4. Proposition. Every allegory is a bisupport category, with support and cosupport
defined as above.
Proof. We verify the four support axioms.
ff =f(f◦f∩1) ⊆f1 = f
f=f∩f⊆f(f◦f∩1) = ff
so ff =f.
To establish that f g =g f it suffices to show f g =f∩g. However, note that f⊆1
and g⊆1, meaning that f , g are subidentities, so that the result follows from the previous
Lemma.
To establish that g f =gf we show that for any subidentity h⊆1 we have g h =gh.
g h =g∩h= 1 ∩g◦g∩h= 1 ∩h∩g◦g
= 1 ∩h◦h∩g◦g⊆1∩h◦h(1 ∩hh◦g◦g)
⊆1∩h◦g◦gh =gh
gh = 1 ∩h◦g◦gh ⊆1∩g◦g=g
gh = 1 ∩h◦g◦gh ⊆h(g◦gh ∩h◦(1 ∩h◦g◦gh)) ⊆h(h◦g◦gh)⊆h
So gh ⊆g∩h=gh whence the identity holds.
To show that gf =gf we first note that, using the first three restriction identities,
gf ⊆fas f∩gf =f gf =gff =gf . We then have:
gf = 1 ∩fgf = 1 ∩f◦(1 ∩g◦g)f
⊆1∩f◦g◦gf =gf
gf =ggf ⊆gf
The cosupport axioms follow by duality. Both structures agree on their subidentities
and thus the support and cosupport are compatible.
428 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
In general, the support in an allegory is not a restriction: the axiom [R.4] fails, because
not all morphisms need be deterministic. However, we may take the subcategory on the
deterministic maps, and this is a range category.
Another relevant subcategory of an allegory is given by the simple morphisms: a
morphism f:A→Bis simple if ff ◦⊆1B. The simple maps are also closed under
composition and contain all restriction idempotents. Every simple map is deterministic,
so that the simple maps form a range category as well. Moreover, morphisms between
allegories preserve simple maps, so this process is functorial.
3.5. Choice Assuming the axiom of choice, the category Par of sets and partial functions
has as one of its special features that every morphism has a partial section: given a partial
function f:A→Bwe may define m:B→Aby
m(b) = awhere ais such that f(a) = b
↑if no such aexists.
In the category of partial recursive functions, a similar construction works (making use
of the well-ordering structure on the natural numbers).
3.6. Definition. Let Cbe a restriction category.
(i) A map fis called a partial retraction when there exists msuch that fm =mand
fmf =f. In this case mis called a partial section of f.
(ii) A map fis called a partial isomorphism when there exists f−1such that f−1f=f
and ff −1=f−1.
(iii) A restriction category satisfies the axiom of choice if every map has a partial section.
Our aim in this section is to show that such categories always have a range: given f,
choose a partial section mand set ˆ
f=m. In order to prove that this satisfies the axioms
of a range, we need to establish some basic properties of partial retractions and partial
sections.
The first thing to note is that partial retractions do not always compose. However,
we have the following observations:
3.7. Lemma. In any restriction category:
(i) A partial isomorphism is a partial retraction, and a partial retraction which is a
partial section is a partial isomorphism.
(ii) Partial isomorphisms compose and have unique partial inverses.
(iii) If fand gf are partial retractions with partial sections mand krespectively then f k
is a partial section of gm.
(iv) If fis a partial retraction and gis a partial isomorphism then gf is a partial re-
traction.
RANGE CATEGORIES I: GENERAL THEORY 429
(v) If fis a partial retraction with partial section mand eis a restriction idempotent
then ef is a partial retraction with partial section me.
Proof.
(i) The first part of the statement is easy. For the second part, suppose fis both a
partial retraction and a partial section, say with m=fm,f mf =f,f=gf and
gf g =g(so that mis a partial section of f, and fis a partial section of g). We show
that mf =f, which implies that fis a partial isomorphism with partial inverse m.
Note first that
m=mm =mfm =f m.
Then
mf =fmf =gfmf =gf =f
and we’re done.
(ii) If f, g are partial isomorphisms then f−1g−1gf =f−1gf =f−1f gf =f gf =gf, so
that indeed f−1g−1is a partial inverse of gf . The second claim is left to the reader.
(iii) Note that from gfk =kit follows that gfk =f k =k. To show that f k is a partial
section of gm we compute
(gm)(f k) = gf mfk =gf k =k=fk
and
(gm)(f k)(gm) = gf mfkgf m =gf kgf m =gf m =gm
where in the first identity we use m=f m twice, in the second we use f mf =f,
and in the third (gf)k(gf ) = gf .
(iv) Suppose that mis a partial section of f. Then mg−1is a partial section of gf :
(gf )(mg−1) = gmg−1=gg−1mg−1=g−1mg−1=mg−1
and
(gf )(mg−1)(gf ) = gmgf =ggmf =gmf =gf mf =gf.
(v) This follows from the previous item, since any restriction idempotent is a partial
isomorphism.
430 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
Having these elementary facts at our disposal, we now state and prove:
3.8. Proposition. For any restriction category Cwith choice, the assignment ˆ
f=m,
where mis any partial section of f, defines a range on C. Moreover, any restriction
functor out of Cpreserves this range.
The first step in the proof is to observe that this definition of range is independent of
the chosen partial section. So let m, m′be two partial sections of f, and compute
m=fm =f m′fm =m′m
so that m≤m′. By symmetry m=m′.
Next, we verify the range axioms.
For [RR.1], consider a partial section mof f, and calculate ˆ
f=m=m=ˆ
f.
For [RR.2], let mbe a partial section of fagain, so that ˆ
ff =mf =f mf =f.
For [RR.3] we first use part (v) of Lemma 3.7 with e=g: this gives that mg is a
partial section of gf . Then c
gf =mg =g m =gˆ
f.
For [RR.4] suppose two partial retractions f, g are given; let mbe a partial section
of f, and kbe a partial section of gf . By item (ii) of the lemma we know that fk is a
partial section of gm. Then
c
gˆ
f=c
gm =f k =k=c
gf
as needed.
For the last claim, we simply observe that any restriction functor preserves partial
retractions and partial sections.
Looking ahead to the representation of range categories in terms of categories with
factorization systems (Section 4), we point out that range categories with choice can be
embedded into partial map categories of categories with a factorization system in which
the E-maps are split epis.
3.9. Inverse categories Inverse categories generalize inverse monoids, which in turn
are inverse semigroups with unit (see the textbook [Lawson 1998] for an exposition of
these notions). We briefly describe now how inverse categories arise, and discuss some
well-known alternative definitions.
3.10. Definition. A restriction category is said to be an inverse category when each
arrow is a partial isomorphism.
For an easy example, note that any groupoid can be regarded as an inverse category
in which the restriction is trivial (i.e., f= 1 for all f).
Another key example comes from inverse semigroups: when Sis an inverse semigroup,
we may form its idempotent splitting: the objects of the resulting category are the idem-
potents eof S, while a morphism e→fis an element s∈Sfor which fs =s=se. As is
RANGE CATEGORIES I: GENERAL THEORY 431
well-known (see for example [Lawson & Steinberg 2004]) this is an inverse category with
s=s∗s, ˆs=ss∗. The partial inverse of sis of course s∗.
From the results in the previous section (Proposition 3.8) we immediately get that an
inverse category is a range category. The restriction is then x=x−1xand the range is
ˆx=xx−1. In fact, the range is also a corestriction (meaning that the axiom gˆ
f=c
gf g
holds). However, note that the converse does not hold: there exist codeterministic range
categories which are not inverse. For the simplest counterexample, consider a category
with two objects and only one non-trivial map A→B.
From the fact that partial isomorphisms compose and that every restriction idempotent
is a partial isomorphism we see that given any restriction category C, the subcategory on
the partial isomorphisms is an inverse category containing all restriction idempotents of
C.
Importantly an inverse category may be described in a number of different ways.
A common approach is to define an inverse category as a category equipped with an
involution (−)−1satisfying the following equations:
(x−1)−1=x, (xy)−1=y−1x−1, xx−1x=x, and xx−1yy−1=yy−1xx−1.
This forces the inverse to be unique in the sense that if xyx =xand yxy =ythen
y=x−1. The calculation is well-known but quite tricky. First we note y=yxx−1and by
symmetry y=x−1xy because
y=yxy =yy−1yxx−1xyy−1y=yxx−1y−1yxyy−1y=yxx−1y−1yy−1y=yxx−1.
Then we have y=yxx−1=x−1xyxx−1=x−1xx−1=x−1.
Alternatively, recall that semigroup theorists call a map fregular (category theorists
might prefer – to avoid confusion with regular monics and regular epics – to call fretractive
as in the idempotent splitting it is a retraction) if there is a gsuch that f gf =f. A
regular inverse (or retractive inverse) of a map fis a gsuch that f gf =fand gfg =g.
An inverse category may then be described as a category in which each map fhas a
unique regular (or retractive) inverse f−1.
Finally, just as for categories with choice, it is worth observing that any restriction
functor automatically preserves inverses, so that the image of an inverse category is always
an inverse category.
3.11. Indexed meet-semilattices and frames We have already mentioned the cat-
egory of topological spaces and partial continuous maps with open domain, as well as the
fact that we tend to think of the meet-semilattices of restriction idempotents as lattices
of open subobjects. This section will explain how each restriction category has associated
to it an indexed meet-semilattice, which generalizes the usual subobject fibration of a
category (see e.g. [Jacobs 1999] for an exposition of fibrations and related notions). We
then investigate this fibred structure in the case of range categories, and also consider the
special case where the restriction idempotents actually form frames.
Our starting point is the category ∧SLatpwhose objects are meet-semilattices with
top element and whose morphisms are meet-preserving functions. (It is not required
432 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
that the top element is preserved.) Then ∧SLatop
pbecomes a restriction category when
we define, for a meet-semilattice homomorphism f:X→Y, its restriction to be the
function f:Y→Ygiven by f(y) = f(⊤)∧y. Clearly the total maps in this category
correspond to the homomorphisms which do preserve the top element. We write ∧SLatop
for Total(∧SLatop
p).
For a restriction category C, we may now regard the so-called fundamental functor,
which takes the form
O:C→ ∧SLatop
p;
it sends an object to its meet-semilattice of restriction idempotents, and a morphism
f:A→Bto f∗:O(B)→ O(A). Note that in the special case where C=∧SLatop
pthis
functor is the identity. The fundamental functor is in fact a restriction functor; this leads
to the following lemma (whose proof is immediate):
3.12. Lemma. For any restriction category C, we have a pullback diagram:
Total(C)//
C
O
∧SLatop //∧SLatop
p
The left-hand vertical map may now be regarded as an ordinary subobject fibration,
with the only difference that the subobjects are restriction idempotents. Observe in
particular that this fibration admits comprehension (in the form of a right adjoint to the
functor which sends an object Cto the terminal idempotent on C, namely the identity)
precisely when the restriction idempotents of Csplit.
Naturally, we consider the subcategory of ∧SLatop
pon the open maps which by defini-
tion is now a range category. Explicitly, the open morphisms of ∧SLatop
pare those which,
when regarded in the opposite category, have a Frobenius left adjoint.
Denoting the subcategory of Con the open maps by Copen, we find that the following
diagram is a pullback.
Copen
O
////C
O
∧SLatop
open,p ////∧SLatop
p
Putting this together with the previous lemma, we find that for a range category C,
we have an associated fibration on Total(C) of restriction idempotents which is actually a
bifibration, i.e., which has existential quantification. This opens the way to the connection
between range categories
and factorization systems, see next section for details. Whenever one has a C-indexed
meet-semilattice P, one may form a restriction category whose objects are those of C,
but whose restriction idempotents at an object Aare given by the meet-semilattice P(A)
(see [Cockett & Guo 2006]). This process is part of an adjunction between the category
RANGE CATEGORIES I: GENERAL THEORY 433
of restriction categories and that of indexed meet-semilattices (the right adjoint being the
assignment of the fundamental functor to a category), and is to be thought of as freely
adding partiality (as specified by the indexed semilattice) to a category. The categories
so obtained have interesting structural properties; for example, they are unitary (in the
sense of inverse semigroups).
The above construction can be adapted to produce range categories: suppose that we
are given a category D, and a functor P:D→ ∧SLatop
open, i.e., an indexed meet-semilattice
in which each reindexing functor f∗:P(B)/f → P(A) has a Frobenius left adjoint. Then
from this data we can construct a range category D[P] whose objects are those of Dand
whose idempotent lattice at Ais P(A). (The proof follows the same lines as in [Cockett
& Guo 2006].)
The category of meet-semilattices has an important subcategory, namely the category
of frames Frmpof frames and ∧,W-preserving functions, but where the morphisms are
not required to preserve the top element. Morphisms in this category always have a right
adjoint, and hence are Frobenius left adjoints. This means that the category Locpof
locales, the opposite of Frmp, is a range category. There is a slightly different description
of this category: it is isomorphic to the category Par(Loc,Open) of locales and partial
maps with open domain. Indeed, any map f:A→Bin Locpfactors uniquely through
the open sublocale of Adetermined by f∗(1B).
4. Representation Theorems
The main representational result for restriction categories, proved in [Cockett & Lack
2002], states that every restriction category is embeddable in a category of partial maps.
More precisely, every category of partial maps is a split restriction category, and every
split restriction category is of this form.
In this section we first extend this result to range categories. The partial map cat-
egories corresponding to range categories will be seen to have a factorization system on
their total map category satisfying a certain stability condition.
After that, we prove a different representational result, which states that every small
range category which satisfies an additional axiom can be represented as a (non-full)
subcategory of Par, the category of sets and partial functions. For semigroups with ranges,
this result is due to Boris Schein (see the exposition in [Jackson & Stokes 2009]); the result
presented here extends his surprising proof.
4.1. Factorization systems and partial maps In a category, a collection Mof
monics which includes all isomorphisms and is closed under composition, is called a system
of monics. A system of monics Mis said to be stable if for any m∈ M and any f:A→B
the pullback of malong fexists and belongs to M. In practice, all systems of monics we
discuss here will be stable, so in order to reduce clutter we drop this adjective.
Given a category Cwith a system of monics M, one may form the category of partial
maps Par(C,M) with:
434 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
Objects: A∈C,
Maps: a map from Ato Bis a pair (m, f ), where m:A′→Ais in Mand f:A′→B
is an arbitrary map in C:
A′
m
~~~
~
~
~
~
~
~f
&&
N
N
N
N
N
N
N
N
N
N
N
N
N
A B
factored out by the equivalence relation: (m, f)≈(m′, f ′) whenever there exists an
isomorphism αin Csuch that m′α=mand f′α=f,
Identities: (1A,1A) : A→A,
Composition: (m′, g)(m, f ) = (mm′′, gf ′),where f′and m′′ are given by the pullback
diagram (∗):
A′′
(∗)
m′′
~~|
|
|
|
|
|
|
|f′
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
A′
m
~~~
~
~
~
~
~
~
f
((
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
PB′
m′
~~}
}
}
}
}
}
}g
&&
N
N
N
N
N
N
N
N
N
N
N
N
N
A B C
The original maps in Ccan be embedded into Par(C,M) by f7→ (1, f).
4.2. Theorem. [Cockett & Lack 2002], Proposition 3.1 Let Cbe a category equipped with
a system of monics M. Then Par(C,M)has a split restriction given by (m, f) = (m, m).
Furthermore, the total maps in Par(C,M)are precisely those in the image of C.
We now consider factorization systems (E,M) in a category C. The factorization
systems we consider are strong, in the sense that diagonal fillers are unique. (See [Adamek
et. al 1990] for an exposition of the general theory of factorization systems.) We say that
(E,M) is M-stable when for every map f=mfef, with mf∈ M and ef∈ E, and a∈ M
we have that
•
a′′
m′
f
@
@
@
@
@
@
@
•
e′
f
88
p
p
p
p
p
p
pf′//
a′
•
a
•
mf
@
@
@
@
@
@
•f//
ef
p
p
p
p
p
p
88
p
p
p
p
p
p
•
RANGE CATEGORIES I: GENERAL THEORY 435
where f′and m′
fare pullbacks of fand mfalong a, and e′
fis the pullback of efalong
a′′, respectively, then f′=m′
fe′
fis the (E,M)-factorization of f′.
It is clear that a factorization system in a category where pullbacks along M-maps
exist is M-stable when the E-maps are stable under pullback along M.
4.3. Example. It is well-known that the category Top of spaces and continuous maps is
not regular because the regular epi-monic factorization is not pullback-stable. However,
when we let Mbe the regular monics and Ebe the class of epimorphisms, then Mis a
system of monics, and we get a factorization system which is pullback stable (along all
maps, not just M-maps).
In [Hughes & Jacobs 2002] it is explained how an M-stable factorization system (E,M)
on a category Cinduces a fibration whose total category is M, regarded as a subcategory of
C→. This fibration then has some distinguishing features: it has existential quantification,
and it has full subset types. Conversely, it is proved in loc. cit. that such fibrations always
arise from factorization systems. In the next two subsections we shall prove directly (i.e.,
without invoking fibrations) that split range categories correspond to stable factorization
systems. In principle, we could have deduced this result from the correspondence exhibited
in loc. cit., by using the fact that, as explained in Section 3.11, the total map category of
a range category comes equipped with a fibration in which reindexing functors have left
adjoints. However, we opt to complete the triangle by directly showing how the range
structure gives rise to a factorization system and vice versa.
4.4. Range categories from factorization systems Next, we extend Theorem 4.2
by showing that if a category with a system of monics Mhas the additional feature of
having an M-stable (E,M) factorization system, then the resulting partial map category
has ranges. Formally:
4.5. Theorem. Let Cbe a category equipped with an M-stable factorization system
(E,M). Then the partial map category Par(C,M)is a split range category.
Proof. From Theorem 4.2 we already know that Par(C,M) is a split restriction category
with restriction (m, f) = (m, m). Therefore it suffices to define the range operator and to
show that the axioms [RR.1]-[RR.4] are satisfied.
To define the range of a morphism (m, f ), consider the (E,M)-factorization of f=
mfef. Then define
\
(m, f) := (mf, mf).
We now have:
[RR.1] \
(m, f) = (mf, mf) = (mf, mf) = \
(m, f).
[RR.2] \
(m, f)(m, f ) = (mf, mf)(m, f) = (m, f ) since the following (*) is a pullback
436 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
diagram:
A′
(∗)
1A′
~~}
}
}
}
}
}
}
}ef
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
A′
m
~~~
~
~
~
~
~
~
f
((
P
P
P
P
P
P
P
P
P
P
P
P
P
P
PD
mf
~~~
~
~
~
~
~
~mf
&&
M
M
M
M
M
M
M
M
M
M
M
M
M
A B B
[RR.3] Let (n′, f ′) be the pullback of (n, f) and (n′′ , m′
f) the pullback of (n, mf). Then
there is a unique map e′
f:D→D′in Csuch that m′
fe′
f=f′and n′′e′
f=efn′and so
(n′, e′
f) is a pullback of (n′′, ef):
E
n′
~~}
}
}
}
}
}
}f′
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
e′
f//_______ F
n′′
}
}
}
~~}
}
}m′
f
A′
m
~~~
~
~
~
~
~
~
ef//
f
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
PD
mf
B′
n
~~}
}
}
}
}
}
}
}n
&&
N
N
N
N
N
N
N
N
N
N
N
N
N
A B B
Hence, by hypothesis, f′=m′
fe′
fis the (E,M)-factorization of f′and therefore nf ′=
(nm′
f)e′
fis the factorization of nf ′since nm′
f∈ M and e′
f∈ E. Thus,
\
(n, g)(m, f ) = \
(n, n)(m, f)
=\
(mn′, nf′)
= (mnf′, mnf ′)
= (nm′
f, nm′
f) (since nf′= (nm′
f)e′
f)
= (n, n)(mf, mf) (since (n′′, m′
f) is the pullback of (n, mf))
= (n, g)\
(m, f).
[RR.4] Suppose that (f′, n′) and (m′
f, n′′) are pullbacks of (f, n) and (mf, n), respectively.
Again, there is a unique map e′
f:E→Fsuch that m′
fe′
f=f′and n′′e′
f=efn′and so
(n′, e′
f) is a pullback of (n′′, ef):
E
n′
~~}
}
}
}
}
}
}f′
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
e′
f//_______ F
n′′
}
}
}
~~}
}
}m′
f
e1
&&
N
N
N
N
N
N
N
N
N
N
N
N
N
A′
m
~~~
~
~
~
~
~
~
ef//
f
''
P
P
P
P
P
P
P
P
P
P
P
P
P
P
PD
mf
B′
n
~~}
}
}
}
}
}
}
}
g
&&
N
N
N
N
N
N
N
N
N
N
N
N
NG
m1
A B C
RANGE CATEGORIES I: GENERAL THEORY 437
If gm′
f=m1e1is the factorization of gm′
f, then gf ′=g(m′
fe′
f) = m1(e1e′
f) is the factor-
ization of gf ′. Thus,
\
(n, g)\
(m, f) = \
(n, g)(mf, mf)
=\
(mfn′′, gm′
f)
= (mgm′
f, m1)
= (mgf ′, mgf ′)
=\
(mgf ′, mgf ′)
=\
(n, g)(m, f ),
as desired.
4.6. Factorization systems from range categories We now embark on the proof
of the converse of Theorem 4.5, namely that any split range category gives rise to a stable
factorization system on its category of total maps.
4.7. Theorem. Let Cbe a split range category. Then Total(C), the subcategory on the
total maps, admits an M-stable (E,M)-factorization system with
E={f|f= 1,ˆ
f= 1};M={m|m= 1, m is a partial isomorphism}
Proof. We first show that E-maps are stable under pullback along M-maps. Let e∈ E
and m∈ M. Since mis a partial section there exists rsuch that m=rm, rmr =r. By
the fact that Mis a stable system of monics, we know that the pullback of ealong m
exists:
Ae′=rem′
//
m′
B
m
Ce//D
r
aa
Moreover, we may take it to be e′=rem′, where we denote the pullback of malong eby
m′; the latter is again a partial section, say with partial retraction r′. Now
b
e′=drem =\
rem′br=\
rem′r′=d
rere =bre =br= 1
This shows that b
e′= 1; since we already know that e′= 1, it follows that e′∈ E.
Next, we show that (E,M) is a factorization system. It is clear that Eis closed under
composition, since if e, e′∈ E then c
e′e=c
e′be=b
e′= 1. Suppose that f:X→Yis a total
map; since ˆ
fis a split restriction idempotent, we may factor it as ˆ
f=mfrffor some maps
rf:Y→Zand mf:Z→Y. with rfmf= 1Z. We claim that the (E,M)-factorization
of fis f=mf(rff). First, note that f=ˆ
ff =mfrff, so that findeed factors as such.
Next, note that mf∈ M since it is a restriction monic. Next, we prove that rff∈ E: we
have rff=mfrff=mfrff=f= 1, and d
rff=d
rfb
f=\rfmfrf=brf= 1.
438 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
To prove that E-maps are orthogonal to M-maps, consider a commutative square
Ae//
p
B
q
Cm//D
r
where e∈ E, m ∈ M and where ris a partial retraction of m. We claim that rq :B→C
is a diagonal filler. That this map is total follows right away from the fact that q=mrq
is total. It is also easy to see that rqe =rmp =p. To show that mrq =q, we derive first
that
rcmp =\
mrmp =cmp
whence cmp ≤r. Using this, we find
mr =r≥cmp =bqe =bq
and therefore mrbq=bq. Now the desired equality
mrq =mrbqq =bqq =q
follows.
Finally, from the fact that mis monic it follows that the filler rq is unique.
4.8. Main Result We now make precise in which sense categories with factorization
systems are the same thing as range categories.
Let MFac be the 2-category with
0-cells: categories Cequipped with a factorization systems (E,M), which is stable under
pullbacks of M-maps, and where all M-maps are monic.
1-cells: a 1-cell from (C,E,M) to (D,F,N) is a functor F:C→Dwhich sends E-maps
to F-maps and M-maps to N-maps, and which preserves pullback squares along
M-maps.
2-cells: M-cartesian natural transformations, i.e., natural transformations whose natu-
rality squares involving M-maps are pullback squares.
We are now in a position to state the main results. Recall that RRcatsis the 2-category
whose 0-cells are split range categories, whose 1-cells are range-preserving restriction
functors and whose 2-cells are total natural transformations.
4.9. Theorem. The 2-categories MFac and RRcatsare 2-equivalent.
RANGE CATEGORIES I: GENERAL THEORY 439
Proof. We have shown that the assignments
Total :RRcats→ MFac,Par :MFac →RRcats
are well-defined on objects. We know that these constructions form a 2-equivalence when
regarded as 2-functors between the 2-category of split restriction categories and the 2-
category of categories equipped with systems of monics. Hence we only need to show that
they are well-defined on 1-cells.
Concretely, this means that we have to show that a range functor Fbetween split
range categories gives a functor Total(F) which preserves E- and M-maps. Conversely,
we need that the restriction functor Par(G) associated to a 1-cell Gin MFac preserves
ranges.
For the first of these claims, recall that the E-maps are defined to be those ffor which
f= 1 and b
f= 1; it is clear that any range functor preserves these. Similarly, the M-maps
are preserved because any restriction functor preserves restriction monics.
For the second part, consider a morphism (m, f ) in Par(C,M,E), and its range
(mf, mf) (where mfis the M-part of f). Since the morphisms in MFac preserve factor-
izations, any such morphism Fwill send f=mfefto F f =mF f eF f . This shows that
(F mf, F mf) = (mF f , mF f ) is the range of F(m, f ), as needed.
4.10. Theorem. Any range category embeds via a full and faithful range preserving func-
tor into a range category of the form Par(C,E,M).
Proof. Given a range category C, consider Split(C), which is a split range category.
There is a full and faithful range preserving functor C→Split(C). Now apply the pre-
vious theorem, which implies that Split(C)∼
=Par(Total(Split(C)),E,M), for the induced
factorization system (E,M).
4.11. A Generalization of Schein’s Theorem While the representational results
in the previous section hold for all range categories, this section is concerned with those
which satisfy the additional axiom
[RR.5] gf =hf ⇒gb
f=hb
f.
In particular, this axiom says that maps for which b
f= 1 are epimorphisms, and hence
it excludes pathological examples of range categories such as the category of Alexandroff
spaces, or the trivial restriction and range on a category. It is interesting to note that
earlier authors [Di Paola & Heller 1987, Rosolini 1988], took this axiom together with
[RR.2] as definition of range. The connection then is:
4.12. Lemma. A combinator satisfying [RR.1],[RR.2] and [RR.5] also satisfies the
other range axioms.
440 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
Proof. We first prove [RR.4]: we have
c
gf gf =gf by [RR.2]
⇔c
gf g ˆ
f=gˆ
fby [RR.5]
⇔c
gf c
gˆ
f=c
gˆ
fby [RR.5]
so that c
gˆ
f≤c
gf , and similarly
c
gˆ
fg ˆ
f=gˆ
fby [RR.2]
⇔c
gˆ
fgf =gf by [RR.5]
⇔c
gˆ
fc
gf =c
gf by [RR.5]
so that c
gf ≤c
gˆ
f.
Next, observe that we have k=kkand hence by [RR.5] ˆ
k=kˆ
k=k. Now [RR.3]
follows, since
gˆ
f=c
gˆ
fas this is a restriction idempotent
=c
gf by [RR.4]
In terms of the representation of range categories using factorization systems, the range
categories satisfying this extra axiom are precisely those corresponding to factorization
systems of which the E-maps are epimorphisms.
The representation to which we now turn is much more specific: it says that we can
faithfully represent a small range category satisfying [RR.5] into the category of sets
and partial functions. The price we pay for this gain in concreteness is the fact that this
representation is generally not full.
Our reworking of Schein’s original result (which was concerned with the special case of
semigroups with ranges) follows the exposition in Jackson and Stokes [Jackson & Stokes
2009].
The aim is, for a range category Csatisfying [RR.5], to define a faithful range functor
S:C→Par.
Towards the definition of this functor, consider an object Xof C. A stickleback γon
Xis a zigzag of maps of the following form:
Y0
γ0
{{v
vvvv
vvv
vγ′
0
A
A
A
A
A
A
AY1
γ1
~~}
}
}
}
}
}
}γ′
1
A
A
A
A
A
A
A
... Yn
γn
~~}
}
}
}
}
}
}
}
X=Z0Z1Z2... Zn
RANGE CATEGORIES I: GENERAL THEORY 441
where γ0=γ′
0,b
γ′
0=bγ1,γ1=γ′
1, ..., d
γ′
n−1=bγn. This stickleback has length n: the
shortest stickleback has one arrow.
Suppose that, in the above diagram, there exists a map xmaking
Yi
γi
γ′
i
!!
C
C
C
C
C
C
C
C
ZiZi+1
x
hh
commute. Then it follows that x≥b
γ′
i, since
xb
γ′
i=c
xγ′
i=d
γ′
iγi=d
γ′
iγ′
i=b
γ′
i.
This allows us to define a new stickleback, which is the same as γexcept for that we
replace the fragment
Yi
γi
γ′
i
!!
C
C
C
C
C
C
C
CYi+1
γi+1
||x
x
x
x
x
x
x
xby
Yi+1
xγi+1
}}{
{
{
{
{
{
{
{
ZiZi+1 Zi
The stickleback so obtained is called a shortening of γvia x, and the new map xγi+1
is called the contracted map. We say that a stickleback is short when it does not admit
any shortenings. The key fact about shortenings is the following lemma:
4.13. Lemma. Any stickleback on Xcan be shortened to a short stickleback, which does
not depend on the choice or order of shortenings.
In effect, this says that shortening, considered as a reduction relation on sticklebacks,
is strongly normalizing. It is clear of course that it is terminating, since shortening strictly
decreases the length of a stickleback.
Proof. Consider a shortening as above, induced by a morphism x:Zi+1 →Zi. Suppose
that x′:Zi+1 →Zialso induces a shortening. Then we have:
xγ′
i=γi=x′γ′
i⇒xbγi=x′b
γ′
i
⇒xdγi+1 =x′dγi+1
⇒xdγi+1 γi+1 =x′dγi+1γi+1
⇒xγi+1 =x′γi+1
so that xand x′give rise to the same shortening. The first implication is where the axiom
[RR.5] is used.
Finally, we have to show that shortenings can be applied in any order. This is clear
when the shortenings are not adjacent. When they are (say via x:Zi+1 →Zi, y :
Zi+2 →Zi+1) then first shortening via xleaves a new stickleback which we shorten via
yx; alternatively, first shortening via yleaves a stickleback which we shorten via x. Either
way, the resulting stickleback has xyγi+2 as contracted map.
442 J.R.B. COCKETT, XIUZHAN GUO AND PIETER HOFSTRA
The short stickleback associated to γwill be denoted by sh(γ). We now define S(X)
to be the set of all short sticklebacks on X. For a morphism f:X→Y, let S(f) be the
partial function which sends the stickleback γto the shortening of
Y0
fγ0
{{v
v
v
v
vv
v
v
vγ′
0
A
A
A
A
A
A
AY1
γ1
~~}
}
}
}
}
}
}γ′
1
A
A
A
A
A
A
A
... Yn
γn
~~}
}
}
}
}
}
}
}
Y=Z0Z1Z2... Zn
provided fγ0=γ0, and which is undefined otherwise.
The