Proof-Theoretical Coherence

Kosta Do
sen and Zoran Petric

Revised Version
September 2007

The version posted here diers from the version printed in 2004 by Kings College Publications (College Publications, London). Besides some relatively slight additions and
corrections, including a small number of additional references, a major correction concerning coherence for dicartesian and sesquicartesian categories, posted already in the
revised versions of May 2006 and March 2007, may be found in 9.6. The present version
diers from the version of March 2007 by having a simpler proof of coherence for lattice
categories in 9.4, and a major correction concerning coherence for lattice categories with
zero-identity arrows in 12.5.

Preface
This is a book in categorial (or categorical) proof theory, a eld of general
proof theory at the border between logic and category theory. In this eld
the language, more than the methods, of category theory is applied to prooftheoretical problems. Propositions are construed as objects in a category,
proofs as arrows between these objects, and equations between arrows, i.e.
commuting diagrams of arrows, are found to have proof-theoretical meaning. They provide a reasonable notion of identity of proofs by equating
derivations that are reduced to each other in a cut-elimination or normalization procedure, or they may be involved in nding a unique normal form
for derivations.
To enter into categorial proof theory one crosses what should be the
watershed between proof theory and the rest of logic. We are not interested
any more in provability onlynamely, in the existence of proofswhich
corresponds to a consequence relation between premises and conclusions.
We have instead a consequence graph, where there may be more than one
dierent proof with the same premise and the same conclusion. We describe
these apparently dierent proofs, code them by terms for arrows, and nd
that some descriptions stand for the same proof, i.e. the same arrow, while
others do not. Our consequence graph is a category, often of a kind that
categorists have found important for their own reasons.
On the other hand, in categorial proof theory proof-theoretical, syntactical, methods are applied to problems of category theory. These are mainly
methods of normalizing in the style of Gentzen or of the lambda calculus.
(In this book, conuence techniques like those in the lambda calculus dominate in the rst part, while cut elimination dominates in the second, bigger,
part.) This syntactical standpoint is something that many categorists do
not favour. Instead of dealing with language, they prefer to work as if
they dealt with the things themselves. We nd that for some problems of
category theory, and in particular for so-called coherence problems, which
make the subject matter of this book, paying attention to language is of
great help.
The term coherence covers in category theory what from a logical
point of view would be called problems of completeness, axiomatizability
and decidability. Dierent authors put stress on dierent things. For our
own purposes we will x a particular notion of coherence, which agrees
completely with Mac Lanes usage of the term in [99], the primordial paper
on coherence.
In the 1960s, at the same time when coherence started being investigated in category theory, the connection between category theory and logic
i

ii

Preface

was established, mainly through Lawveres ideas (see [94]). The roots of
categorial proof theory date from the same yearsthey can be found in a
series of papers by Lambek: [84], [85], [86] and [87]. Lambek introduced
Gentzens proof-theoretical methods in category theory, which Mac Lane
and Kelly exploited in [81] to solve a major coherence problem (see also
[101]).
There are not many books in categorial proof theory. The early attempt
to present matters in [127] has shortcomings. Proofs are not systematically
coded by terms for arrows; only the sources and targets of arrows are mentioned most of the time, and too much work is left to the reader. Some
claims are excessively dicult to verify, and some are not correct (see [69],
[12], Section 3, and [14], Section 1). Lambeks and Scotts book [90] is
only partly about categorial proof theory and coherence, understood as a
decidability problem for equality of arrows in cartesian closed categories.
(Just a short chapter of [128], Chapter 8, touches upon this topic.) The
only remaining book in categorial proof theory we know about, [38], is devoted to showing that cut elimination characterizes fundamental notions of
category theory, in particular the notion of adjunction. Some parts of that
book (Sections 4.10 and 5.9) are about coherence.
Papers in categorial logic often touch upon this or that point of categorial proof theory, but are not very often specically within the eld. And
even when they are within this eld, some authors prefer to advertise their
work as semantical. It should be clear, however, that this is not semantics in the established model-theoretical sensethe sense in which the word
was used in logic in the twentieth century. We nd this semantics of proofs
more proof-theoretical than model-theoretical.
We will try to cover with the references in our book not the whole literature of categorial proof theory, but only papers relevant to the problems
treated. To acknowledge more direct inuences, we would like, however, to
mention at the outset a few authors with whom we have been in contact,
and whose ideas are more or less close to ours.
First, Jim Lambeks pioneering and more recent work has been for us, as
for many others, a source of inspiration. Max Kellys papers on coherence
(see [77], [78], [54] and [79]) are less inuenced by logic, though logical
matters are implicit in them. Sergei Solovievs contributions to categorial

proof theory (see [118], [119] and [120]) and Djordje Cubri
cs (see [28], [29]
and [30]) are close to our general concerns, though they do not deal exactly
with the subject matter of this book; the same applies to some work of
Alex Simpson (in particular, [117]).
We extend Robert Seelys and Robin Cocketts categorial presentation
of a fragment of linear logic, based upon what they call linear, alias weak,
distribution (see [22]; other papers will be cited in the body of the book),

Preface

iii

which we call dissociativity. This is an associativity principle involving two

operations, which in the context of lattices delivers distribution. While
Cockett and Seely are concerned with dissociativity as it occurs in linear
logic, and envisage also applications in the study of intuitionistic logic, we
have been oriented towards the categorication of classical propositional
logic. The subtitle of our book could be General proof theory of classical
propositional logic. We would have put this subtitle were it not that
a great part of the book is about fragments of this proof theory, which
are fragments of the proof theory of other logics too, and are also of an
independent interest for category theory. Besides that, we are not sure our
treatment of negation in the last chapter is as conclusive as what precedes
it. (We also prefer a shorter and handier title.)
Proofs in the conjunctive-disjunctive fragment of logic, which is related
to distributive lattices, may, but need not, be taken to be the same in classical and intuitionistic logic, and they are better not taken to be the same.
Classical proof theory should be based on plural (multiple-conclusion) sequents, while intuitionistic proof theory, though it may be presented with
such sequents, is more often, and more naturally, presented with singular
(single-conclusion) sequents. By extending Cocketts and Seelys categorial
treatment of dissociativity, we present in the central part of the book a
categorication, i.e. a generalization in category theory, of the notion of
distributive lattice, which gives a plausible notion of identity of proofs in
classical conjunctive-disjunctive logic. This notion is related to Gentzens
cut-elimination procedure in a plural-sequent system. By building further
on that, at the end of the book we provide a plausible categorication of
the notion of Boolean algebra, which gives a nontrivial notion of identity
of proofs for classical propositional logic, also related to Gentzen.
It is usually considered that it is hopeless to try to categorify the notion
of Boolean algebra, because all plausible candidates based on the notion
of bicartesian closed category (i.e. cartesian closed category with nite coproducts) led up to now to equating all proofs with the same premises and
conclusions. In our Boolean categories, which are built on another base, this
is not the case. The place where in our presentation of the matter classical
and intuitionistic proof theory part ways is in understanding distribution.
In intuitionistic proof theory distribution of conjunction over disjunction is
an isomorphism, while distribution of disjunction over conjunction is not.
This is how matters stand in bicartesian closed categories. We take that in
classical proof theory neither of these distributions is an isomorphism, and
restore symmetry, typical for Boolean notions.
We reach our notion of Boolean category very gradually. This gradual approach enables us to shorten calculations at latter stages. Moreover, along the way we prove coherence for various more general notions

iv

Preface

of category, entering into the notion of Boolean category or related to it.

Coherence is understood in our book as the existence of a faithful structurepreserving functor from a freely generated category, built out of syntactical
material, into the category whose arrows are relations between nite ordinals. This is a limited notion of coherence, and our goal is to explore the
limits of this particular notion within the realm of classical propositional
logic. We are aware that other notions of coherence exist, and that even
our notion can be generalized by taking another category instead of the
category whose arrows are relations between nite ordinals. These other
notions and these generalizations are, however, outside the connes of our
book, and we will mention them only occasionally (see, in particular, 12.5
and 14.3)
Mac Lanes primordial coherence results for monoidal and symmetric
monoidal categories in [99] are perfectly covered by our notion of coherence. When the image of the faithful functor is a discrete subcategory of
the category whose arrows are relations between nite ordinals, coherence
amounts to showing that the syntactical category is a preordering relation,
i.e. that all diagrams commute. This is the case sometimes, but not always, and not in the most interesting cases. Mac Lanes coherence results
are scrutinized in our book, and new aspects of the matter are made manifest. We also generalize previous results of [72] (Section 1) on strictication,
i.e. on producing equivalent categories where some isomorphisms are turned
into identity arrows. Our strictication is useful, because it facilitates the
recording of lengthy calculations.
For categories with dissociativity, which cover proofs in the multiplicative conjunctive-disjunctive fragment of linear logic, and also proofs in the
conjunctive-disjunctive fragment of classical logic, we provide new coherence results, and we prove coherence for our Boolean categories. These
coherence theorems, which are the main results of the book, yield a simple
decision procedure for the problem whether a diagram of canonical arrows
commutes, i.e. for the problem whether two proofs are identical.
The most original contribution of our book may be that we take into
account union, or addition, of proofs in classical logic. This operation on
proofs with the same premise and same conclusion is related to the mix
principle of linear logic. It plays an important role in our Boolean categories, and brings them close to linear algebra. Taking union of proofs into
account saves Gentzens cut-elimination procedure for classical logic from
falling into triviality, as far as identity of proofs is concerned. This modied cut elimination is the cornerstone of the proof of our main coherence
theorem for classical propositional logic.
We take into account also the notion of zero proof, a notion related to
union of proofsa kind of dual of it. With union of proofs hom-sets become

Preface

semilattices with unit, but we envisage also that they be just commutative
monoids, as in additive and abelian categories. Zero proofs, which are like
a leap from any premise to any conclusion, are mapped into the empty
relation in establishing coherence. Although they enable us to prove anything as far as provability is concerned, they are conservative with respect
to the previously established identity of proofs in logic. We will show that
envisaging zero proofs is useful. It brings logic closer to linear algebra, and
facilitates calculations. We nd also that the notion of zero proof may be
present in logic even when we do not allow passing from any premise to any
conclusion, but restrict ourselves to the types of the acceptable deductions
connecting premises and conclusions, i.e. stick to provability in classical
logic. Negation may be tied to such restricted zero proofs.
Zero proofs resemble what Hilbert called ideal mathematical objects,
like imaginary numbers or points at innity. If our concern is not with
provability, but with proofsnamely, identity of proofszero proofs are
useful and harmless. We dont think we have exhausted the advantages
of taking them into account in general proof theory. We believe, however,
we have fullled to a great extent the promises made in the programmatic
survey [40] (summarized up to a point in the rst chapter of the book),
which provides further details about the context of our research.
We suppose our principal public should be a public of logicians, such as
we are, but we would like no less to have categorists as readers. So we have
strived to make our exposition self-contained, both on the logical and on
the categorial side. This is why we go into details that logicians would take
for granted, and into other details that categorists would take for granted.
Only for the introductory rst chapter, whose purpose is to give motivation,
and for some asides, in particular at the very end, we rely on notions not
dened in the book, but in the standard logical and categorial literature.
We suppose that the results of this book should be interesting not only
for logic and category theory, but also for theoretical computer science. We
do not control very well, however, the quickly growing literature in this
eld, and we will refrain from entering into it. We do not pretend to be
experts in that area. Some of the investigations of proofs of classical logic
that appeared since 1990 in connection with modal translations into linear
logic or with the lambda-mu calculus, in which the motivation, the style
and the jargon of computer science dominate, seem to be concerned with
identity of proofs, but it is not clear to us how exactly these concerns are
related to ours. We leave for others to judge.
This is more a research monograph than a textbook, but the text could
serve nevertheless as the base for a graduate course in categorial proof
theory. We provide after the nal chapter a list of problems left open. To
assist the reader, we also provide at the end of the book a list of axioms

vi

Preface

and denitions, and a list of categories treated in the book (which are
quite numerous), together with charts for these categories indicating the
subcategory relations established by our coherence results.
We would like to thank in particular Alex Simpson and Sergei Soloviev
for encouraging and useful comments on the preprint of this book, which
was distributed since May 2004. We would like to thank also other colleagues who read this preprint and gave compliments on it, or helped us in
another manner.
Dov Gabbay was extremely kind to take care of the publishing of the
book. We are very grateful to him and to Jane Spurr for their eorts and
eciency.
The results of this book were announced previously in a plenary lecture
at the Logic Colloquium in M
unster in August 2002, and in a talk at the
International Congress MASSEE in Borovets in September 2003, with the
support of the Alexander von Humboldt Foundation. We are indebted to
Slobodan Vujosevic and Milojica Jacimovic for the invitation to address the
Eleventh Congress of the Mathematicians of Serbia and Montenegro, held
in Petrovac in September 2004, with a talk introducing matters treated in
the book. We had the occasion to give such introductory talks also at the
Logic Seminar in Belgrade in the last two years and, thanks to Mariangiola
Dezani-Ciancaglini and the Types project of the European Union, at the
Types conference in Jouy-en-Josas in December 2004.
We would like to thank warmly the Mathematical Institute of the Serbian Academy of Sciences and Arts in Belgrade and the Faculty of Philosophy of the University of Belgrade for providing conditions in which we
could write this book. Our work was generously supported by a project of
the Ministry of Science of Serbia (1630: Representation of Proofs).
Belgrade, December 2004

CONTENTS

Preface

Chapter 1. Introduction
1.1. Coherence
1.2. Categorication
1.3. The Normalization Conjecture in general proof theory
1.4. The Generality Conjecture
1.5. Maximality
1.6. Union of proofs and zero proofs
1.7. Strictication

1
1
6
10
15
24
26
29

Chapter 2. Syntactical Categories

2.1. Languages
2.2. Syntactical systems
2.3. Equational systems
2.4. Functors and natural transformations
2.5. Denable connectives
2.6. Logical systems
2.7. Logical categories
2.8. C-functors
2.9. The category Rel and coherence

33
34
36
39
42
44
47
51
53
59

Chapter 3. Strictification

65

3.1. Strictication in general

3.2. Direct strictication
3.3. Strictication and diversication

65
78
84

Chapter 4. Associative Categories

4.1. The logical categories K
4.2. Coherence of semiassociative categories
4.3. Coherence of associative categories
4.4. Associative normal form
4.5. Strictication of associative categories
4.6. Coherence of monoidal categories
4.7. Strictication of monoidal categories

87
88
89
93
96
98
101
103

vii

viii

Contents

Chapter 5. Symmetric Associative Categories

5.1. Coherence of symmetric associative categories
5.2. The faithfulness of GH
5.3. Coherence of symmetric monoidal categories
Chapter 6. Biassociative Categories
6.1.
6.2.
6.3.
6.4.
6.5.

Coherence of biassociative and bimonoidal categories

Form sequences
Coherence of symmetric biassociative categories
Coherence of symmetric bimonoidal categories
The category S

Chapter 7. Dissociative Categories

7.1.
7.2.
7.3.
7.4.
7.5.
7.6.
7.7.
7.8.
7.9.

Coherence of dissociative categories

Net categories
Coherence of net categories
Net normal form
Coherence of semidissociative biassociative categories
Symmetric net categories
Cut elimination in GDS
Invertibility in GDS
Linearly distributive categories

Chapter 8. Mix Categories

8.1.
8.2.
8.3.
8.4.
8.5.

Coherence
Coherence
Coherence
Coherence
Coherence

of
of
of
of
of

mix and mix-dissociative categories

mix-biassociative categories
mix-net categories
mix-symmetric net categories
mix-symmetric biassociative categories

Chapter 9. Lattice Categories

9.1.
9.2.
9.3.
9.4.
9.5.
9.6.
9.7.

Coherence of semilattice categories

Coherence of cartesian categories
Maximality of semilattice and cartesian categories
Coherence of lattice categories
Maximality of lattice categories
Coherence for dicartesian and sesquicartesian categories
Relative maximality of dicartesian categories

107
107
110
112
115
115
117
117
119
121
127
128
132
133
142
143
145
148
156
163
167
167
169
173
176
182
185
185
191
194
199
205
207
213

Contents

ix

Chapter 10. Mix-Lattice Categories

10.1. Mix-lattice categories and an example
10.2. Restricted coherence of mix-lattice categories
10.3. Restricted coherence of mix-dicartesian categories
Chapter 11. Distributive Lattice Categories
11.1.
11.2.
11.3.
11.4.
11.5.

Distributive lattice categories and their Gentzenization

Cut elimination in D
Coherence of distributive lattice categories
Legitimate relations
Coherence of distributive dicartesian categories

Chapter 12. Zero-Lattice Categories

12.1.
12.2.
12.3.
12.4.
12.5.

Zero-lattice and zero-dicartesian categories

Coherence of zero-lattice and zero-dicartesian categories
Maximality of zero-lattice and zero-dicartesian categories
Zero-lattice and symmetric net categories
Zero-identity arrows

Chapter 13. Zero-Mix Lattice Categories

13.1.
13.2.
13.3.
13.4.

Coherence of zero-mix lattice categories

Zero-mix lattice and distributive lattice categories
Coherence of zero-mix dicartesian categories
The category Semilat

Chapter 14. Categories with Negation

14.1.
14.2.
14.3.
14.4.

De Morgan coherence
Boolean coherence
Boolean categories
Concluding remarks

Problems Left Open

List of Equations
List of Categories
Charts
Bibliography
Index

219
219
223
227
231
232
246
263
268
270
275
276
282
285
286
287
295
296
301
304
306
309
310
316
322
328
331
332
345
354
359
371

Chapter 1
Introduction
In this introductory chapter we provide in an informal manner motivation
for the main themes of the book, without giving an exhaustive summary of
its content (such summaries are provided at the beginning of every chapter).
A great deal of the chapter (1.3-6) is based on the survey [40].
While in the body of the book, starting from the next chapter, our
exposition, except for some asides, will be self-contained, both from a logical
and from a categorial point of view, here we rely on some acquaintance with
proof theory (which the reader may have acquired in classic texts like [60],
[111] and [82], Chapter 15, or in the more recent textbook [128]), and on
some notions of category theory (which may be found in [100] and [90]).
Many, but not all, of the notions we need for this introduction will be
dened later in the book.
To have read the present chapter is not essential for reading the rest of
the book. A reader impatient for more precision can move to Chapter 2,
where the book really starts, and return to this introduction later on.

1.1.

Coherence

It seems that what categorists call coherence logicians would, roughly speaking, call completeness. This is the question whether we have assumed
for a particular brand of categories all the equations between arrows we
should have assumed. Completeness need not be understood here as completeness with respect to models. We may have a syntactical notion of
1

CHAPTER 1.

INTRODUCTION

completenesssomething akin to the Post completeness of the classical

propositional calculusbut often some sort of model-theoretical completeness is implicit in coherence questions. Matters are made more complicated
by the fact that categorists do not like to talk about syntax, and do not
perceive the problem as being one of nding a match between syntax and
semantics. They do not talk of formal systems, axioms and models.
Moreover, questions that logicians would consider to be questions of
decidability, which is not the same as completeness, are involved in what
categorists call coherence. A coherence problem often involves the question
of deciding whether two terms designate the same arrow, i.e. whether a
diagram of arrows commuteswe will call this the commuting problem
and sometimes it may involve the question of deciding whether there is in
a category an arrow of a given type, i.e. with a given source and target
we will call this the theoremhood problem (cf. [38], Sections 0.2 and 4.6.1).
Coherence is understood mostly as solving the commuting problem in [90]
(see p. 117, which mentions [84] and [85] as the origin of this understanding).
The commuting problem seems to be involved also in the understanding of
coherence of [79] (Section 10).
Completeness and decidability, though distinct, are, of course, not foreign to each other. A completeness proof with respect to a manageable
model may provide, more or less immediately, tools to solve decision problems. For example, the completeness proof for the classical propositional
calculus with respect to the two-element Boolean algebra provides immediately a decision procedure for theoremhood.
The simplest coherence questions are those where it is intended that all
arrows of the same type should be equal, i.e. where the category envisaged
is a preorder. The oldest coherence problem is of that kind. This problem
has to do with monoidal categories, and was solved by Mac Lane in [99]
(where early related work by Stashe and D.B.A. Epstein is mentioned; see
[122] for historical notes, and also [123], Appendix B, co-authored with S.
Shnider). The monoidal category freely generated by a set of objects is a
preorder. So Mac Lane could claim that showing coherence is showing that
all diagrams commute. We provide in Chapter 4 a detailed analysis of
Mac Lanes coherence result for monoidal categories.
In cases where coherence amounts to showing preorder, i.e. showing that

1.1.

Coherence

from a given set of equations, assumed as axioms, we can derive all equations (provided the equated terms are of the same type), from a logical
point of view we have to do with axiomatizability. We want to show that
a decidable set of axioms (and we wish this set to be as simple as possible,
preferably given by a nite number of axiom schemata) delivers all the intended equations. If preorder is intended, then all equations are intended.
Axiomatizability is in general connected with logical questions of completeness, and a standard logical notion of completeness is completeness of a set
of axioms. Where all diagrams should commute, coherence does not seem
to be a question of model-theoretical completeness, but even in such cases
it may be conceived that the model involved is a discrete category (cf. the
end of 2.9).
Categorists are interested in axiomatizations that permit extensions.
These extensions are in a new language, with new axioms, and such extensions of the axioms of monoidal categories need not yield preorders any
more. Categorists are also interested, when they look for axiomatizations,
in nding the combinatorial building blocks of the matter. The axioms are
such building blocks, as in knot theory the Reidemeister moves are the combinatorial building blocks of knot and link equivalence (see [97], Chapter 1,
or any other textbook in knot theory).
In Mac Lanes second coherence result of [99], which has to do with
symmetric monoidal categories, it is not intended that all equations between arrows of the same type should hold. What Mac Lane does can be
described in logical terms in the following manner. On the one hand, he
has an axiomatization, and, on the other hand, he has a model category
where arrows are permutations; then he shows that his axiomatization is
complete with respect to this model. It is no wonder that his coherence
problem reduces to the completeness problem for the usual axiomatization
of symmetric groups.
Algebraists do not speak of axiomatizations, but of presentations by generators and relations. All the axiomatizations in this book will be purely
equational axiomatizations, as in algebraic varieties. Such were the axiomatizations of [99]. Categories are algebras with partial operations, and we
are here interested in the equational theories of these algebras.
In Mac Lanes coherence results for monoidal and symmetric monoidal

CHAPTER 1.

INTRODUCTION

categories one has to deal only with natural isomorphisms. Coherence questions in the area of n-categories are usually restricted likewise to natural
isomorphisms (see [96]). However, in the coherence result for symmetric
monoidal closed categories of [81] there are already natural and dinatural
transformations that are not isomorphisms.
A natural transformation is tied to a relation between the argumentplaces of the functor in the source and the argument-places of the functor in
the target. This relation corresponds to a relation between nite ordinals,
and in composing natural transformations we compose these relations (see
2.4 and 2.9). With dinatural transformations the matter is more complicated, and composition poses particular problems (see [109]). In this book
we deal with natural transformations, and envisage only in some comments
coherence for situations where we do not have natural transformations. Our
general notion of coherence does not, however, presuppose naturality and
dinaturality.
Our notion of coherence result is one that covers Mac Lanes and Kellys
coherence results mentioned up to now, but it is more general. We call
coherence a result that tells us that there is a faithful functor G from a category S freely generated in a certain class of categories to a manageable
category M. This calls for some explanation.
It is desirable, though perhaps not absolutely necessary, that the functor G be structure-preserving, which means that it preserves structure at
least up to isomorphism (see 1.7 below, and, in particular, 2.8). In all
coherence results we will consider, the functor G will preserve structure
strictly, i.e. on the nose. The categories S and M will be in the same
class of categories, and G will be obtained by extending in a unique way a
map from the generators of S into M.
The category M is manageable when equations of arrows, i.e. commuting
diagrams of arrows, are easier to consider in it than in S. The best is if the
commuting problem is obviously decidable in M, while it was not obvious
that it is such in S.
With our approach to coherence we are oriented towards solving the
commuting problem, and we are less interested in the theoremhood problem. In this book, we deal with the latter problem only occasionally, mostly
when we need to solve it in order to deal with the commuting problem (see

1.1.

Coherence

4.2, 7.1, 7.3-5, 8.2-3 and 11.4). This should be stressed because
other authors may give a more prominent place to the theoremhood problem. We nd that the spirit of the theoremhood problem is not particularly
categorial: this problem can be solved by considering only categories that
are preorders. And ordinary, or perhaps less ordinary, logical methods for
showing decidability of theoremhood are here more useful than categorial
methods. For the categories in this book, the decidability of the theoremhood problem is shown by syntactical or semantical logical tools. Among
the latter we also have sometimes simply truth tables. We have used on
purpose the not very precise term manageable for the category M to leave
room for modications of our notion of coherence, which would be oriented
towards solving another problem than the commuting problem. Besides the
theoremhood problem, one may perhaps also envisage something else, but
our ocial notion of coherence is oriented towards the commuting problem.
In this book, the manageable category M will be the category Rel with
arrows being relations between nite ordinals, whose connection with natural transformations we have mentioned above. The commuting problem in
Rel is obviously decidable. We do, however, consider briey categories that
may replace Relin particular, the category whose arrows are matrices (see
12.5).
The freely generated category S will be the monoidal category freely
generated by a set of objects, or the symmetric monoidal category freely
generated by a set of objects, or many others of that kind. The generating
set of objects may be conceived as a discrete category. In our understanding of coherence, replacing this discrete generating category by an arbitrary
category would prevent us to solve coherencesimply because the commuting problem in the arbitrary generating category may be undecidable. Far
from having more general, stronger, results if the generating category is
arbitrary, we may end up by having no result at all.
The categories S in this book are built ultimately out of syntactic material, as logical systems are built. Categorists are not inclined to formulate
their coherence results in the way we doin particular, they do not deal
often with syntactically built categories (but cf. [131], which comes close to
that). If, however, more involved and more abstract formulations of coherence that may be found in the literature (for early references on this matter

CHAPTER 1.

INTRODUCTION

see [80]) have practical consequences for solving the commuting problem,
our way of formulating coherence has these consequences as well.
That there is a faithful structure-preserving functor G from the syntactical category S to the manageable category M means that for all arrows
f and g of S with the same source and the same target we have
f = g in S i Gf = Gg in M.
The direction from left to right in this equivalence is contained in the functoriality of G, while the direction from right to left is faithfulness proper.
If S is conceived as a syntactical system, while M is a model, the
faithfulness equivalence we have just stated is like a completeness result in
logic. The left-to-right direction, i.e. functoriality, is soundness, while the
right-to-left direction, i.e. faithfulness, is completeness proper.
In this book we will systematically separate coherence results involving
special objects (such as unit objects, terminal objects and initial objects)
from those not involving them. These objects tend to cause diculties,
and the statements and proofs of the coherence results gain by having these
diculties kept apart. When coherence is obtained both in the absence and
in the presence of special objects, our results become sharper.

1.2.

Categorification

By categorication one can understand, very generally, presenting a mathematical notion in a categorial setting, which usually involves generalizing
the notion and making ner distinctions. In this book, however, we have
something more specic in mind. We say that we have a categorication of
the notion of algebraic structure in which there is a preordering, i.e. reexive and transitive, relation R when we replace R with arrows in a category,
and obtain thereby a more general categorial notion instead of the initial
algebraic notion. If the initial algebraic structure is a completely free algebra of terms, like the algebra of formulae of a propositional language, the
elements of the algebra just become objects in a free category in the class
of categories resulting from the categorication. Otherwise, some splitting
of the objects is involved in categorication.

1.2.

Categorification

Categorication is not a technical notion we will rely on later, and so

we will not try to dene it more precisely. What we have in mind should
be clear from the following examples.
By categorifying the algebra of formulae of conjunctive logic with the
constant true proposition, where the preordering relation R is induced by
implication, we may end up with the notion of cartesian category. We
may end up with the same notion by categorifying the notion of semilattice
with unit, where the relation R is the partial ordering of the semilattice.
A semilattice with unit is a cartesian category that is a partial order, i.e.
in which whenever we have arrows from a to b and vice versa, then a and
b are the same object. In the same sense, the notion of monoidal category
is a categorication of the notion of monoid, and the notion of symmetric monoidal category is a categorication of the notion of commutative
monoid, the preordering relation R in these two cases being equality.
There are other conceptions of categorication except that one. One
may categorify an algebra by taking its objects to be arrows of a category.
The notion of category is a categorication in this sense of the notion of
monoid, monoids being categories with a single object. In that direction,
one obtains more involved notions of categorication in the n-categorial
setting (see [2] and [27]).
The motivation for categorication may be internal to category theory,
but it may come from other areas of mathematics, like algebraic topology and mathematical physicsin particular, quantum eld theory (many
references are given in [2]). Our motivation comes from proof theory, as
we will explain in latter sections of this introduction. We are replacing a
consequence relation, which is a preordering relation, by a category, where
arrows stand for proofs. In comparing our approach to others, note that
the slogan Replace equality by isomorphisms!, which is sometimes heard
in connection with categorication, does not describe exactly what we are
doing. Our slogan Replace preorder by arrows! implies, however, the
other one, and so the same categorial notions, like, for example, the notion
of monoidal category, may turn up under both slogans.
In this book one may nd, in particular, categorications, in our restricted sense, of the notions of distributive lattice and Boolean algebra. Alternatively, these may be taken as categorications of conjunctive-disjunctive

CHAPTER 1.

INTRODUCTION

logic, or of the classical propositional calculus. Previously, a categorication

of the notion of distributive lattice was obtained with so-called distributive
categories, i.e. bicartesian categories with distribution arrows from a(bc)
to (a b) (a c) that are isomorphisms (see [95], pp. 222-223 and Session
26, and [20]). Bicartesian closed categories, i.e. cartesian closed categories
with nite coproducts (see [90], Section I.8), are distributive categories in
this sense.
In our categorication of the notion of distributive lattice, distribution
arrows of the type above need not be isomorphisms. This rejection of
isomorphism is imposed by our wish to have coherence with respect to the
category Rel of the preceding section, since the relation underlying the
following diagram:
( a b ) ( a
@ QQ

@ Q
a ( b
HH

 HH
( a b ) ( a

c )




c )
A
A
c )

namely the relation underlying the diagram on the left-hand side below, is
not the identity relation underlying the diagram on the right-hand side:
( a b ) ( a c )
PP

P

PP


P
( a b ) ( a c )

( a b ) ( a c )
( a b ) ( a c )

Our categorication of the notion of distributive lattice is based on

arrows from a (b c) to (a b) c, which Cockett and Seely studied in
their categorial treatment of a fragment of linear logic (see [22]; further
references are given in 7.1 and 7.9). At rst, they called the principle
underlying these arrows weak distribution, and then changed this to linear
distribution in [25]. Since this is a principle that delivers distribution in the
context of lattices, but is in fact an associativity involving two operations,
we have coined the name dissociativity for it, to prevent confusion with
what is usually called distribution. Cockett and Seely were concerned with

1.2.

Categorification

establishing some sort of coherence for dissociativity with respect to proof

nets.
Before appearing in proof nets and in categories, dissociativity was pregured in universal algebra and logic (see 7.1 for references). Dissociativity is related to the modularity law of lattices (see 7.1), and we will see
in 11.3 how in a context that is a categorication of the notion of lattice
this two-sorted associativity delivers distribution arrows of the usual types,
from a (b c) to (a b) (a c) and from (a b) (a c) to a (b c)
(the arrows of the converse types are there anyway), of which neither need
to be an isomorphism. The arrows from (a b) (a c) to a (b c) need
not be isomorphisms in bicartesian closed categories too.
The categorication of the notion of Boolean algebra is usually deemed
to be a hopeless task (see 14.3), because it is assumed this categorication should be based on the notion of bicartesian closed category. In that
notion, as we said above, we have arrows corresponding to distribution of
conjunction over disjunction that are isomorphisms. Natural assumptions
in this context lead to triviality, i.e. to categories that are preorders. Our
categorication of the notion of Boolean algebra is not trivial in this sense.
It incorporates the notion of bicartesian category (i.e. category with nite
products and coproducts), but does not admit cartesian closure. Its essential ingredient is our categorication of distributive lattices, in which the
arrows corresponding to distribution of conjunction over disjunction are not
isomorphisms.
We think it is a prejudice to assume that there must be an isomorphism
behind distribution of conjunction over disjunction. It would likewise be
a prejudice to assume that behind the idempotency law a a = a or the
absorption law a (a b) = a of lattices we must have isomorphisms. The
categorication of the notion of lattice in bicartesian categories is not under
the spell of the latter two assumptions, but the isomorphism corresponding to distribution of conjunction over disjunction is usually presupposed.
This is presumably because in the category Set of sets with functionsthe
central category there isdistribution of cartesian product over disjoint
union is an isomorphism. In the categorication of the notion of distributive lattice with distributive categories, where a (b c) is isomorphic
to (a b) (a c), it is not required that a (b c) be isomorphic to

10

CHAPTER 1.

INTRODUCTION

(a b) (a c), presumably because the latter isomorphism need not exist

in Set. We assume neither of these isomorphisms in our categorication of
the notion of distributive lattice.

1.3.

The Normalization Conjecture in general proof

theory

Categorication is interesting for us because of its connection with general

proof theory. The question What is a proof? was considered by Prawitz
in [112] (Section I) to be the rst question of general proof theory. To
keep up with the tradition, we speak of proof, though we could as well
replace this term by the more precise term deduction, since we have in
mind deductive proofs from assumptions (including the empty collection
of assumptions). Together with the question What is a proof?, Prawitz
envisaged the following as one of the rst questions to be considered in
general proof theory (see [112], p. 237):
In the same way as one asks when two formulas dene the same
set or two sentences express the same proposition, one asks when
two derivations represent the same proof; in other words, one
asks for identity criteria for proofs or for a synonymity (or
equivalence) relation between derivations.
An answer to the question of identity criteria for proofs might lead to
an answer to the basic question What is a proof?. A proof would be the
equivalence class of a derivation. The related question What is an algorithm? could be answered by an analogous factoring through an equivalence relation on representations of algorithms. (Moschovakis stressed in
[107], Section 8, the fundamental interest of identity criteria for algorithms.)
Prawitz did not only formulate the question of identity criteria for proofs
very clearly, but also proposed a precise mathematical answer to it.
Prawitz considered derivations in natural deduction systems and the
equivalence relation between derivations that is the reexive, transitive and
symmetric closure of the immediate-reducibility relation between derivations. Of course, only derivations with the same premises and the same
conclusion may be equivalent. Prawitzs immediate-reducibility relation is

1.3.

The Normalization Conjecture in general proof theory

11

the one involved in reducing a derivation to normal forma matter he

studied previously in [111]. As it is well known, the idea of this reduction stems from Gentzens thesis [60]. A derivation reduces immediately
to another derivation (see [112], Section II.3.3) when the latter is obtained
from the former either by removing a maximum formula (i.e. a formula
with a connective that is the conclusion of an introduction of and the
major premise of an elimination of ), or by performing one of the permutative reductions tied to the eliminations of disjunction and of the existential
quantier, which enables us to remove what Prawitz calls maximum segments. There are some further reductions, which Prawitz called immediate
simplications; they consist in removing eliminations of disjunction where
no hypothesis is discharged, and there are similar immediate simplications involving the existential quantier, and redundant applications of
the classical absurdity rule. Prawitz also envisaged reductions he called
immediate expansions, which lead to the expanded normal form where all
the minimum formulae are atomic (minimum formulae are those that are
conclusions of eliminations and premises of introductions).
Prawitz formulates in [112] (Section II.3.5.6) the following conjecture,
for which he gives credit (in Section II.5.2) to Martin-Lof, and acknowledges
inuence by ideas of Tait:
Conjecture. Two derivations represent the same proof if and
only if they are equivalent.
We call this conjecture the Normalization Conjecture.
This conjecture, together with another conjecture, which will be considered in the next section, was examined in the survey [40]. The present
section and the next three sections give an updated, somewhere shortened
and somewhere expanded, variant of that survey, to which we refer for further, especially historical and philosophical, remarks. (Some other bits of
that survey are in 14.3, where a mistaken statement is also corrected at
the end of the section.)
The normalization underlying the Normalization Conjecture need not
be understood always in the precise sense envisaged by Prawitz. For intuitionistic logic Prawitzs understanding of normalization, which is derived
from Gentzen, is perhaps optimal. There are, however, other logics, and,

12

CHAPTER 1.

INTRODUCTION

in particular, there is classical logic, to which natural deduction is not so

closely tied, and for which we may still have a notion of normalization,
perhaps related to Prawitzs, but dierent. What comes to mind immediately for classical logic is Gentzens plural, i.e. multiple-conclusion, sequent
systems (see below) and cut elimination for them.
Presumably, the notion of normalization we can envisage in the Normalization Conjecture cannot be based on an arbitrary notion of normal
form. It is desirable that this normal form be unique, at least up to some
supercial transformations (like alpha conversion in the lambda calculus).
But uniqueness should not be enough. This normal form and the language for which it is formulated must be signicant, where it is dicult to
say what signicant means exactly. The normal form and the language
for which it is formulated should not be just a technical device, but they
must be deeply tied to the logic, and exhibit its essential features. In the
case of Prawitzs normal form for derivations in intuitionistic natural deduction, besides philosophical reasons having to do with the meaning of
logical connectives, there are important ties with independently introduced
mathematical structures.
The Normalization Conjecture was formulated by Prawitz at the time
when the Curry-Howard correspondence between derivations in natural deduction and typed lambda terms started being recognized more and more
(though the label Curry-Howard was not yet canonized). Prawitzs equivalence relation between derivations corresponds to beta-eta equality between typed lambda terms, if immediate expansions are taken into account,
and to beta equality otherwise.
Besides derivations in natural deduction and typed lambda terms, where
according to the Curry-Howard correspondence the latter can be conceived
just as codes for the former, there are other, more remote, formal representations of proofs. There are rst Gentzens sequent systems, which are
related to natural deduction, but are nevertheless dierent, and there are
also representations of proofs as arrows in categories. The sources and
targets of arrows are taken to be premises and conclusions respectively,
and equality of arrows with the same source and target, i.e. commuting
diagrams of arrows, should now correspond to identity of proofs via a conjecture analogous to the Normalization Conjecture.

1.3.

The Normalization Conjecture in general proof theory

13

The fact proved by Lambek (see [87] and [90], Part I; see also [39],
[37] and [43]) that the category of typed lambda calculuses with functional
types and nite product types, based on beta-eta equality, is equivalent to
the category of cartesian closed categories, and that hence equality of typed
lambda terms amounts to equality between arrows in cartesian closed categories, lends additional support to the Normalization Conjecture. Equality of arrows in bicartesian closed categories corresponds to equivalence of
derivations in Prawitzs sense in full intuitionistic propositional logic (see
[109], Section 3, for a detailed demonstration that the equations of bicartesian closed categories deliver cut elimination for intuitionistic propositional
logic). The notion of bicartesian closed category is a categorication in the
sense of the preceding section of the notion of Heyting algebra. The partial
order of Heyting algebras is replaced by arrows in this categorication.
In category theory, the Normalization Conjecture is tied to Lawveres
characterization of the connectives of intuitionistic logic by adjoint situations. Prawitzs equivalence of derivations, in its beta-eta version, corresponds to equality of arrows in various adjunctions tied to logical connectives (see [94], [38], Section 0.3.3, [41] and [39]). Adjunction is the unifying
concept for the reductions envisaged by Prawitz.
The fact that equality between lambda terms, as well as equality of
arrows in cartesian closed categories, were rst conceived for reasons independent of proofs is remarkable. This tells us that we are in the presence
of a solid mathematical structure, which may be illuminated from many
sides.
Prawitz formulated the Normalization Conjecture having in mind natural deduction, and so mainly intuitionistic logic. For classical logic we
envisage something else. Our categorication of the notion of Boolean algebra, as the categorication of the notion of Heyting algebra with bicartesian closed categories, covers a notion of identity of proofs suggested by
normalization via cut elimination in a plural-sequent system (see Chapters
11 and 14). This is in spite of the fact that for us distribution of conjunction over disjunction does not give rise to isomorphisms, as in bicartesian
closed categories. This disagreement over the isomorphism of distribution
may be explained as follows.

14

CHAPTER 1.

INTRODUCTION

Classical and intuitionistic logic do not dier with respect to the consequence relation between formulae in the conjunction-disjunction fragment
of propositional logic. In other words, they do not dier with respect to
provable sequents of the form A B where A and B are formulae of the
conjunction-disjunction fragment. But, though these two logics do not
dier with respect to provability, they may dier with respect to proofs.
The standard sequent formulation of classical logic, the formulation that
imposes itself by its symmetry and regularity, is based on plural sequents
, where may be a collection with more than one formula, whereas
the standard sequent formulation of intuitionistic logic is based on singular,
i.e. single-conclusion, sequents , where cannot have more than one
formula, while can. There are presentations of intuitionistic logic with
plural sequents (see [103] and [32], Section 5C4, with detailed historical
remarks on pp. 249-250; cf. also [31], where the idea is already present),
but they are not standard, and they do not correspond to natural deduction, as those with singular sequents do. Moreover, in these plural-sequent
formulations of intuitionistic logic, a restriction based on singularity is kept
for introduction of implication on the right-hand side, which corresponds
to the deduction theorem. The deduction theorem enables the deductive
metalogic to be mirrored with the help of implication in the object language, and when it comes to this mirroring, plural-sequent formulations of
intuitionistic logic avow that their deductive metalogic is based on singular
sequents.
The connection of intuitionistic logic with natural deduction, where
there are possibly several premises, but never more than one conclusion,
goes very deep. There are many reasons to hold that the meaning of intuitionistic connectives is explained in the framework of natural deduction,
as suggested by Gentzen (see [60], Section II.5.13). Singular sequents are
asymmetric, i.e. they have a plurality of premises versus a single conclusion.
The asymmetries of intuitionistic logic, and, in particular, the asymmetry
between conjunction and disjunction, can be explained by the asymmetry of
singular sequents that underly this logic. One can suppose that the asymmetry of bicartesian closed categories, which consists in having a (b c)
isomorphic to (a b) (a c) without having a (b c) isomorphic to
(a b) (a c), has the same roots.

1.4.

The Generality Conjecture

15

The dissociativity principle of the arrow that goes from a (b c) to

(a b) c (see 1.2 and 7.1) delivers arrows that go from a (b c) to
(a b) (a c) and from (a b) (a c) to a (b c) (see 11.3; we have
arrows of the converse types without assuming distribution), but neither
of these arrows need to be isomorphisms. So symmetry, which is typical
for Boolean notions, is restored. (Another possibility to restore symmetry
would be to take that a(bc) is isomorphic to (ab)(ac) and a(bc)
is isomorphic to (a b) (a c), which is not the case in Set, but we will
not explore that possibility in this book.)
The dissociativity principle, which is an essential ingredient of our categorication of the notions of distributive lattice and Boolean algebra, is
built into the plural-sequent formulation of classical logic. It is tied to the
cut rule of plural sequents (see 11.1, and also 7.7).
Prawitz envisaged the Normalization Conjecture for classical logic, but
in a natural deduction formulation, i.e. with singular sequents. This is not
the same as considering this conjecture with plural sequents.

1.4.

The Generality Conjecture

At the same time when Prawitz formulated the Normalization Conjecture,

in a series of papers ([84], [85], [86] and [87]) Lambek was engaged in a
project where arrows in various sorts of categories were construed as representing proofs. The source of an arrow corresponds to the premise, and the
target to the conclusion. (Proofs where there is a nite number of premises
dierent from one are represented by proofs with a single premise with
the help of connectives like conjunction and the constant true proposition.)
With this series of papers Lambek inaugurated the eld of categorial proof
theory.
The categories Lambek considered in [84] and [85] are rst those that
correspond to his substructural syntactic calculus of categorial grammar
(these are monoidal categories where the functors a . . . and . . . a have
right adjoints). Next, he considered monads, which besides being fundamental for category theory, cover proofs in modal logics of the S4 kind. In
[86] and [87], Lambek dealt with cartesian closed categories, which cover
proofs in the conjunction-implication fragment of intuitionistic logic. He

16

CHAPTER 1.

INTRODUCTION

also envisaged bicartesian closed categories, which cover the whole of intuitionistic propositional logic.
Lambeks insight is that equations between arrows in categories, i.e.
commuting diagrams of arrows, guarantee cut elimination, i.e. composition
elimination, in an appropriate language for naming arrows. (In [38] it is
established that for some basic notions of category theory, and in particular
for the notion of adjunction, the equations assumed are necessary and sufcient for composition elimination.) Since cut elimination is closely related
to Prawitzs normalization of derivations, the equivalence relation envisaged by Lambek should be related to Prawitzs. (An early presentation of
the connection between Prawitz and Lambek is in [105].)
The normalization of cut elimination does not involve only eliminating
cuts, but also equations between cut-free terms for arrows, which may guarantee their uniqueness. (This is like adding the eta equations to the beta
equations in the typed lambda calculus and natural deduction.)
Lambeks work is interesting not only because he worked with an equivalence relation between derivations amounting to Prawitzs, but also because
he envisaged another kind of equivalence relation. Lambeks idea is best
conveyed by considering the following example. In [86] (p. 65) he says that

the rst projection arrow k 1p,p : p p p and the second projection arrow

k 2p,p : p p p, which correspond to two derivations of conjunction elimina

tion, have dierent generality, because they generalize to k 1p,q : p q p and

k 2p,q : p q q respectively, and the latter two arrows do not have the same

target; on the other hand, k 1p,q : p q p and k 2q,p : q p p do not have the
same source. The idea of generality may be explained roughly as follows.
We consider generalizations of derivations that diversify variables without
changing the rules of inference. Two derivations have the same generality
when every generalization of one of them leads to a generalization of the
other, so that in the two generalizations we have the same premise and
conclusion (see [84], p. 257). In the example above, this is not the case.
Generality induces an equivalence relation between derivations. Two
derivations are equivalent if and only if they have the same generality.
Lambek does not formulate so clearly as Prawitz a conjecture concerning
identity criteria for proofs, but he suggests that two derivations represent

1.4.

The Generality Conjecture

17

the same proof if and only if they are equivalent in the new sense. We will
call this conjecture the Generality Conjecture.
Lambeks own attempts at making the notion of generality precise (see
[84], p. 316, where the term scope is used instead of generality, and
[85], pp. 89, 100) need not detain us here. In [86] (p. 65) he nds that these
attempts were faulty.
The simplest way to understand generality is to use graphs whose vertices are occurrences of propositional letters in the premise and the conclusion of a derivation. We connect by an edge occurrences of letters that
must remain occurrences of the same letter after generalizing, and do not
connect those that may become occurrences of dierent letters. So for the
rst and second projection above we would have the two graphs

k 1p,p

p p
@
@
p

p p

k 2p,p
p

When the propositional letter p is replaced by an arbitrary formula A we

have an edge for each occurrence of propositional letter in A.
The generality of a derivation is such a graph. According to the Generality Conjecture, the rst and second projection derivations from p p to
p represent dierent proofs because their generalities dier.
One denes an associative composition of such graphs, and there is also
an obvious identity graph with straight parallel edges, so that graphs make
a category, which we call the graphical category. If on the other hand it
is taken for granted that proofs also make a category, which we will call
the syntactical category, with composition of arrows being composition of
proofs, and identity arrows being identity proofs (an identity proof composed with any other proof, either on the side of the premise or on the side
of the conclusion, is equal to this other proof), then the Generality Conjecture may be rephrased as the assertion that there is a faithful functor
from the syntactical category to the graphical category. So the Generality
Conjecture is analogous to a coherence theorem of category theory. The
manageable category is a graphical category.
The coherence result of [81] proves the Generality Conjecture for the

18

CHAPTER 1.

INTRODUCTION

multiplicative conjunction-implication fragment of intuitionistic linear logic

(modulo a condition concerning the multiplicative constant true proposition, i.e. the unit with respect to multiplicative conjunction), and, inspired
by Lambek, it does so via a cut-elimination proof. The syntactical category
in this case is a free symmetric monoidal closed category, and the graphical
category is of a kind studied in [54]. The graphs of this graphical category
are closely related to the tangles of knot theory. In tangles, as in braids, we
distinguish between two kinds of crossings, but here we need just one kind,
in which it is not distinguished which of the two crossed edges is above
the other. (For categories of tangles see [134], [129] and [73], Chapter 12.)
Tangles with this single kind of crossing are like graphs one encounters in
Brauer algebras (see [15] and [132]). Here is an example of such a tangle:
(( p q ) p ) p












( p ( q (r r))) (s s)
Tangles without crossings at all serve in [38] (Section 4.10; see also [42])
to obtain a coherence result for the general notion of adjunction, which
according to Lawveres Thesis underlies all the connectives of intuitionistic
logic, as we mentioned in the preceding section. In terms of combinatorial
low-dimensional topology, the mathematical content of the general notion
of adjunction is caught by the Reidemeister moves of planar ambient isotopy. An analogous coherence result for self-adjunctions, where a single
endofunctor is adjoint to itself, is proved in [49]. Through this latter result
we reach the theory of Temperley-Lieb algebras, which play a prominent
role in knot theory and low-dimensional topology, due to Jones representation of Artins braid groups in these algebras (see [74], [97], [110] and
references therein).
In [49] one nds also coherence results for self-adjunctions where the
graphical category is the category of matrices, i.e. the skeleton of the category of nite-dimensional vector spaces over a xed eld with linear trans-

1.4.

The Generality Conjecture

19

formations as arrows. Tangles without crossings may be faithfully represented in matrices by a representation derived from the orthogonal group
case of Brauers representation of Brauer algebras (see also [132], Section 3,
and [70], Section 3). This representation is based on the fact that the
Kronecker product of matrices gives rise to a self-adjoint functor in the category of matrices, and this self-adjointness is related to the fact that in this
category, as well as in the category Rel, whose arrows are binary relations
between nite ordinals, nite products and coproducts are isomorphic.
Graphs like graphs of the tangle type were tied to sequent derivations of
classical logic in [18] and [19], but without referring to categories, coherence
or the question of identity criteria for proofs.
In [108] there are several coherence results, which extend [99], for the
multiplicative-conjunction fragments of substructural logics. But less us
concentrate now on coherence results for classical and intuitionistic logic.
The Normalization Conjecture and the Generality Conjecture agree only
for limited fragments of these two logics. They agree for purely conjunctive
logic, with or without the constant true proposition (see [46] and 9.1-2
below). Proofs in conjunctive logic are the same for classical and intuitionistic logic. Here the Normalization Conjecture is taken in its beta-eta
version. By duality, the two conjectures agree for purely disjunctive logic,
with or without the constant absurd proposition . If we have both conjunction and disjunction, but do not yet have distribution, and have neither
nor , then the two conjectures still agree for both logics, provided the
graphical category is the category Rel whose arrows are relations between
nite ordinals (see [48] and 9.4). And here it seems we have reached the
limits of agreement as far as intuitionistic logic is concerned. With more
sophisticated notions of graphs, matters may stand dierently, and the area
of agreement for the two conjectures may perhaps be wider, but it can be
even narrower, as we will see below.
It may be questioned whether the intuitive idea of generality is caught by
the category Rel in the case of conjunctive-disjunctive logic. The problem is

that if wp : p p p is a component of the diagonal natural transformation,

and k 1q,p : q q p is a rst injection, then in categories with nite products

20

CHAPTER 1.

INTRODUCTION

and coproducts we have

(1q wp ) k 1q,p = k 1q,pp ,

where the left-hand side cannot be further generalized, but the right-hand

side can be generalized to k 1q,pr . The intuitive idea of generality seems to

require that in wp : p p p we should not have only a relation between the

domain and the codomain, as on the left-hand side below, but an equivalence relation on the union of the domain and the codomain, as on the
right-hand side:
p

A


A
p p

p


A


A
p p

(see [50], and also [51]). With such equivalence relations, we can still get coherence for conjunctive logic, and for disjunctive logic, taken separately, but
for conjunctive-disjunctive logic the left-to-right direction, i.e. the soundness part, of coherence would fail (see 14.3). So for conjunctive-disjunctive
logic the idea of generality with which we have coherence is not quite the
intuitive idea suggested by Lambek, but only something close to it, which
involves the categorial notion of natural transformation (cf. the end of
14.3).
Even when we stay within the connes of the category Rel, our understanding of generality does not match exactly the intuitive notion of
generality for conjunctive-disjunctive logic. Intuitively, the relations R of
Rel corresponding to generality should satisfy difunctionality in the sense of
[114]; namely, we should have R R1 R R. But this requirement is not
satised for our images in Rel under G of proofs in conjunctive-disjunctive
logic, even in the absence of distribution (see the end of 14.3). Generality
is caught by Rel only for fragments of logic. Altogether, generality serves
only as a loose motivation for taking Rel as our graphical category. Real
grounds for Rel are in the notion of natural transformation, which has to
do with permuting rules in derivations.
The Normalization Conjecture and the Generality Conjecture agree neither for the conjunction-disjunction fragment of intuitionistic logic with

1.4.

The Generality Conjecture

21

and (see [47] and 9.6), nor for the conjunction-implication fragment of
this logic. We do not have coherence for cartesian closed categories if the
graphs in the graphical category are taken to be of the tangle type Kelly
and Mac Lane had for symmetric monoidal closed categories combined with
the graphs we have in Rel for cartesian categoriesboth the soundness part
and the completeness part of coherence fail (for soundness see a counterexample in 14.3, with p p replaced by p p, and for completeness see
[125]). The soundness part of coherence fails also for distributive bicartesian
categories, and a fortiori for bicartesian closed categories. The problem is
that in these categories distribution of conjunction over disjunction is taken
to be an isomorphism, and Rel does not deliver that, as we have seen in
1.2.
The problem with the soundness part of coherence for cartesian closed
categories may be illustrated with typed lambda terms in the following
manner. By beta conversion and alpha conversion, we have the following
equation:
x x, xy y = y y, z z
for y and z of type p, and x of type pp (which corresponds to p p). The
closed terms on the two sides of this equation are both of type pp pp . The
type of the term on the left-hand side cannot be further generalized, but
the type of the term y y, z z, can be generalized to pp q q . The problem
noted here does not depend essentially on the presence of surjective pairing
, and of product types; it arises also with purely functional types. This
problem depends essentially on the multiple binding of variables, which we
have in x x, x; that is, it depends on the structural rule of contraction.
This throws some doubt on the right-to-left direction of the Normalization
Conjecture, which Prawitz found relatively unproblematic. It might be
considered strange that two derivations represent the same proof if, without
changing inference rules, one can be generalized in a manner in which the
other cannot be generalized.
The area of agreement between the Normalization Conjecture and the
Generality Conjecture may be wider for classical logic, provided normalization is understood in the sense of cut elimination for plural sequent systems
and generality is understood in the sense of the category Rel. It extends

22

CHAPTER 1.

INTRODUCTION

rst to conjunctive-disjunctive logic without distribution (see [48] and 9.4

below). Next, in conjunctive-disjunctive logic with distribution, with or
without and , the agreement also holds (see Chapter 11). And it covers
also the whole classical propositional calculus, with a particular way of understanding normalization involving zero proofs (see 1.6 and 14.2-3). We
do not pretend this particular way of understanding normalization in the
presence of negation is the only possible one, but in the absence of negation
we feel pretty secure, and the match between the two conjectures is indeed
very good. Gentzens cut elimination procedure for plural-sequent systems
needs only to be modied in a natural way by admitting union of proofs, a
rule that in this context amounts to the mix rule of linear logic (see Chapters 8 and 10). Admitting union of proofs saves Gentzens cut-elimination
procedure from falling into preorder and triviality. Our cut-elimination procedure diers also from Gentzens in the way how it treats the structural
rule of contraction, but in this respect it is more in the spirit of Gentzen.
(We will point down at appropriate places in 11.1-2 how our procedure
is related to Gentzens.)
Zero proofs (which were mentioned already in the preface) come up with
negation. Their appearance is imposed by our wish to have coherence with
respect to Rel. With other graphical categories they may disappear, but at
the cost of many problems (which we discuss in 14.3). In particular, the
match between the Normalization and the Generality Conjectures would
be impaired (see 14.3). The price we have to pay with our categorication
of the notion of Boolean algebra is that not all connectives will be tied
to adjoint functors, as required by Lawvere. Conjunction and disjunction
are tied to the usual adjunctions with the diagonal functor (the product
bifunctor is right-adjoint to the diagonal functor, and the coproduct bifunctor is left-adjoint to the diagonal functor), but distribution is an additional
matter, not delivered by these adjunctions, and classical negation and implication do not come with the usual adjunctions. (There are perhaps hidden
adjunctions of some kind here.) Another price we have to pay with zero
proofs is that all theorems, i.e. all propositions proved without hypotheses,
will have zero proofs. So the theorems of classical propositional logic, in
contradistinction to their intuitionistic counterparts, do not serve to encode
the deductive metatheory of classical propositional logic. This metatheory

1.4.

The Generality Conjecture

23

exists, nevertheless, and its categorication is not given by categories that

are preorders.
When we compare the two conjectures we should say something about
their computational aspects. With the Normalization Conjecture we have
to rely in intuitionistic logic on reduction to a unique normal form in
the typed lambda calculus in order to check equivalence of derivations in
the conjunction-implication fragment of intuitionistic propositional logic.
Nothing more practical than that is known, and such syntactical methods
may be tiresome. Outside of the conjunction-implication fragment, in the
presence of disjunction and negation, such methods become uncertain.
Methods for checking equivalence of derivations in accordance with the
Generality Conjecture, i.e. methods suggested by coherence results, often
have a clear advantage. This is like the advantage truth tables have over
syntactical methods of reduction to normal form in order to check tautologicality. However, the semantical methods delivered by coherence results
have this advantage only if the graphical category is simple enough, as our
category Rel is. When we enter into categories suggested by knot theory, this simplicity may be lost. Then, on the contrary, syntax may help
us to decide equality in the graphical category. The Normalization Conjecture has made a foray in theoretical computer science, in the area of
typed programming languages. It is not clear whether one could expect the
Generality Conjecture to play a similar role.
The reexive and transitive closure of the immediate-reducibility relation involved in normalization may be deemed more important than the
equivalence relation engendered by immediate reducibility, which we have
considered up to now. This matter leads outside our topic, which is identity
of proofs, but it is worth mentioning. We may categorify the identity relation between proofs, and consider not only other relations between proofs,
but maps between proofs. The proper framework for doing that seems to
be the framework of weak 2-categories, where we have 2-arrows between
arrows; or we could even go to n-categories, where we have n+1-arrows between n-arrows (one usually speaks of cells in this context). Composition of
1-arrows is associative only up to a 2-arrow isomorphism, and analogously
for other equations between 1-arrows. Identity of 1-arrows is replaced by
2-arrows satisfying certain coherence conditions. In the context of the Gen-

24

CHAPTER 1.

INTRODUCTION

erality Conjecture, we may also nd it natural to consider 2-arrows instead

of identity. The orientation would here be given by passing from a graph
with various detours to a graph that is more straight, which need not
be taken any more as equal to the original graph.
With all this we would enter into a very lively eld of category theory,
interacting with other disciplines, mainly topology (see [96] and papers cited
therein). The eld looks very promising for general proof theory, both from
Prawitzs and from Lambeks point of view, but, as far as we know, it has
not yet yielded to proof theory much more than promises.

1.5.

Maximality

The fragments of logic mentioned in the preceding section where the Normalization Conjecture and the Generality Conjecture agree for intuitionistic
logic all possess a property called maximality. Let us say a few words about
this important property.
For the whole eld of general proof theory to make sense, and in particular for considering the question of identity criteria for proofs, we should
not have that any two derivations with the same premise and conclusion
are equivalent. Otherwise, our eld would be trivial.
Now, categories with nite nonempty products, cartesian categories and
categories with nite nonempty products and coproducts have the following
property. Take, for example, cartesian categories, and take any equation in
the language of free cartesian categories that does not hold in free cartesian
categories. If a cartesian category satises this equation, then this category
is a preorder. We have an exactly analogous property with the other sorts
of categories we mentioned (see 9.3 and 9.5). This property is a kind of
syntactical completeness, analogous to the Post completeness of the usual
axiomatizations of the classical propositional calculus. Any extension of
the equations postulated leads to collapse.
Translated into logical language, this means that Prawitzs equivalence
relation for derivations in conjunctive logic, disjunctive logic and conjunctivedisjunctive logic without distribution and without and , which in all
these cases agrees with our equivalence relation dened via generality in
the sense of Rel, is maximal. Any strengthening, any addition, would yield

1.5.

Maximality

25

that any two derivations with the same premise and the same conclusion
are equivalent.
If the right-to-left direction of the Normalization Conjecture holds, with
maximality we can eciently justify the left-to-right direction, which Prawitz found problematic in [112], and about which Kreisel was thinking
in [83]. In the footnote on p. 165 of that paper Kreisel mentions that
Barendregt suggested this justication via maximality. Suppose the rightto-left direction of the Normalization Conjecture holds, suppose that for
some premise and conclusion there is more than one proof, and suppose
the equivalence relation is maximal. Then if two derivations represent the
same proof, they are equivalent. Because if they were not equivalent, we
would never have more than one proof with a given premise and a given
conclusion. Nothing can be missing from our equivalence relation, because
whatever is missing, by maximality, leads to collapse on the side of the
equivalence relation, and, by the right-to-left direction of the conjecture, it
also leads to collapse on the side of identity of proofs.
Prawitz in [112] found it dicult to justify the left-to-right direction of
the Normalization Conjecture, and Kreisel was looking for mathematical
means that would provide this justication. Maximality is one such means.
Establishing the left-to-right direction of the Normalization Conjecture
via maximality is like proving the completeness of the classical propositional
calculus with respect to any kind of nontrivial model via Post completeness
(which is proved syntactically by reduction to conjunctive normal form).
Actually, the rst proof of this completeness with respect to tautologies was
given by Bernays and Hilbert exactly in this manner (see [135], Sections
2.4 and 2.5; see also [66], Section I.13, and 9.3 below).
Maximality for the sort of categories mentioned above is proved with the
help of coherence in [46] and [48] (which is established proof-theoretically,
by normalization, cut elimination and similar methods; see Chapter 9).
Coherence is helpful in proving maximality, but maximality can also be
proved by other means, as this is done for cartesian closed categories via a
typed version of Bohms theorem in [121], [117] and [45]. This justies the
left-to-right direction of the Normalization Conjecture also for the implicational and the conjunction-implication fragments of intuitionistic logic.
The maximality of bicartesian closed categories, which would justify the

26

CHAPTER 1.

INTRODUCTION

left-to-right direction of the Normalization Conjecture for the whole of intuitionistic propositional logic is, as far as we know, an open problem. (A
use for maximality similar to that propounded here and in [45] and [46] is
envisaged in [133].)
In [38] (Section 4.11) it is proved that the general notion of adjunction is
also maximal in some sense. The maximality we encountered above, which
involves connectives tied to particular adjunctions, cannot be derived from
the maximality of the general notion of adjunction, but these matters should
not be foreign to each other.
Since we nd maximality an interesting property, we pay attention to
it in this book where we could establish it with the help of our coherence
results, and where it is not a trivial property. Besides the maximality results
from Chapter 9, mentioned above, there are analogous results in 12.3, 12.5
and 13.3. We also pay attention to maximality in cases where it cannot
be established (see 10.3 and 11.5). In some cases where it does not hold,
we still have relative maximality results (see 9.7, 11.5 and 12.5).

1.6.

Union of proofs and zero proofs

Gentzens plural-sequent system for classical logic has implicitly a rule of

union, or addition, of derivations, which is derived as follows:
f:AB
R
f
C

contractions

: A B, C

L
R
g) :
cut(C
f, C

g: A B
L
g:
C

C, A B

A, A B, B

f g: A B

R
L
Here C
f and C
g are obtained from f and g respectively by thinning on
R
L
the right and thinning on the left, and cut(C
f, C
g) may be conceived as
obtained by applying to f and g a limit case of Gentzens multiple-cut rule
mix, where the collection of mix formulae is empty. A related principle was
considered under the name mix in linear logic (see 8.1).
In a cut-elimination procedure like Gentzens, f g is reduced either
to f or to g (see [60], Sections III.3.113.1-2). If we have f g = f and
f g = g, then we get immediately f = g, that is collapse and triviality.

1.6.

Union of proofs and zero proofs

27

In [64] (Appendix B.1 by Y. Lafont; see also [67], Section 1) this is taken
as sucient ground to conclude that cut elimination in the plural-sequent
system for classical logic must lead to preorder and collapse. (In [64], the
inevitability of this collapse is compared to the argument presented after
Proposition 1 of 14.3, which shows that a plausible assumption about classical negation added to bicartesian closed categories leads to preorder, but
these are dierent matters.) To evade collapse we may try keeping only one
of the equations f g = f and f g = g, and reject the other; then we must
also reject the commutativity of , but it seems such decisions would be
arbitrary. (For similar reasons, even without assuming the commutativity
of , the assumptions of [127], p. 232, C.12, lead to preorder.) There is,
however, a way to evade collapse here that is not arbitrary. The modication of Gentzens cut-elimination procedure expounded in Chapter 11 (see
also 12.5) and our coherence results (more precisely, the easy, soundness,
i.e. functoriality, parts of these results) testify to that.
The Generality Conjecture tells us that we should have neither f g = f
nor f g = g. The union of two graphs may well produce a graph diering
from each of the graphs entering into the union. It also tells us that union
of proofs should be associative and commutative. The idempotency law
f f = f is imposed by Rel, but it stands apart, and with another graphical
category, we may do without it (see 12.5). Without idempotency, union of
proofs is rather addition of proofs. Our way out of the problematic situation
Gentzen found himself in is to take into account union or addition of proofs.
(Besides [40], section 7, the paper [5], which deals with cut elimination in
ane logic, also makes a similar suggestion.)
If we have union of proofs, it is natural to assume that we also have
for every formula A and every formula B a zero proof 0A,B : A B, with
an empty graph, which with union of proofs makes at least a commutative
monoid; with idempotency, it gives the unit of a semilattice. We may
envisage having zero proofs 0A,B : A B only for those A and B where
there is also a nonzero proof from A to B, as we do in our categorication of
the notion of Boolean algebra, but the more sweeping assumption involving
every A and every B makes sense too.
We should immediately face the complaint that with such zero proofs we
have entered into inconsistency, since everything is provable. That is true,

28

CHAPTER 1.

INTRODUCTION

but not all proofs have been made identical, and we are here not interested
in what is provable, but in what proofs are identical. If it happensand
with the Generality Conjecture it will happen indeedthat introducing
zero proofs is conservative with respect to identity of proofs that do not
involve zero proofs, then it is legitimate to introduce zero proofs, provided it
is useful for some purpose. This is like extending our mathematical theories
with what Hilbert called ideal objects; like extending the positive integers
with zero, or like extending the reals with imaginary numbers.
The use of union of proofs is that it saves the agreement between the
Normalization and Generality Conjectures in the presence of distribution,
as we said in 1.4. The use of zero proofs is that it does the same in
the presence of negation. The idempotency of union is essential in the
absence of zero proofs, but not in their presence. Without idempotency
our graphical category in the case of conjunctive-disjunctive logic turns up
to be a category whose arrows are matrices, rather than the category Rel.
Composition becomes matrix multiplication, and union is matrix addition.
And in the presence of zero matrices, we obtain a unique normal form like
in linear algebra: every matrix is the sum of matrices with a single 1 entry.
A number of logicians have sought a link between logic and linear algebra, and here is such a link. We have it not for an alternative logic, but for
classical logic. We have it, however, not at the level of provability, but at
the level of identity of proofs.
The unique normal form suggested by linear algebra is not unrelated
to cut elimination. In the graphical category of matrices the result of
cut elimination is obtained by multiplying matrices, and the equations of
this category yield a cut-elimination procedure. They yield it even in the
absence of zero proofs, provided we have 1 + 1 = 1. Unlike Gentzens
cut-elimination procedures for classical logic, the new procedure admits a
commutative addition or union of proofs without collapse. So, in classical
logic, the Generality Conjecture is not foreign to cut elimination, and hence
it is not foreign to the Normalization Conjecture, provided we understand
the equivalence relation involved in this conjecture in a manner dierent
from Prawitzs.
This need not exhaust the advantages of having zero proofs. They may
be used also to analyze disjunction elimination. Without pursuing this

1.7.

Strictification

29

topic very far, let us note that passing from A B to A involves a zero
proof from B to A, and passing from A B to B involves a zero proof from
A to B. If next we are able to reach C both from A and from B, we may
add our two proofs from A B to C, and so to speak cancel the two zero
proofs.
Logicians were, and still are, interested mostly in provability, and not
in proofs. This is so even in proof theory. When we address the question
of identity of proofs we have certainly left the realm of provability, and
entered into the realm of proofs. This should become clear in particular
when we introduce zero proofs.

1.7.

Strictification

Strictication is inverse to categorication. While categorication usually

(but not always) involves splitting objects, strictication involves identifying objects. Factoring a set through an equivalence relation, i.e. replacing
the objects of a set by equivalence classes of objects of this set, is a simple
example of strictication. Logicians are very used to a kind of strictication that may be called lindenbaumization, by which the algebra of
formulae of conjunctive logic is replaced by a freely generated semilattice,
or the algebra of formulae of intuitionistic propositional logic is replaced by
a freely generated Heyting algebra, or the algebra of formulae of classical
propositional logic is replaced by a freely generated Boolean algebra. The
equivalence relation involved in these strictications is mutual implication.
In this book we are, however, interested in strictication of categories.
Precise notions of strictication, which we need for our work, will be introduced in Chapter 3. Let us say for the time being that the simpler of these
notions is a kind of partial skeletization of a category. An equivalence relation, induced by a subcategory that is a groupoid and a preorder, is used to
replace the objects of the category by equivalence classes of objects. In the
other, more general and more involved notion, the partial skeletization is
applied to a category generated out of a given category. (We are aware this
preliminary rough description of the matter cannot be very informative.)
After strictication, objects are replaced by equivalence classes, which may
correspond to sequences, or multisets, or sets, or structures of that kind.

30

CHAPTER 1.

INTRODUCTION

The idea is to obtain a strictied category equivalent to the initially

given category in which computations are easier to record, because some
arrows that were not identity arrows, like, typically, associativity isomorphisms, are replaced by identity arrows (see Chapters 5-8 and 11). This
equivalence of categories is not meant to be any equivalence, but an equivalence via functors that preserve a particular categorial structure at least
up to isomorphism. For that we will dene precisely what it means for a
functor to preserve a structure, such as interests us, up to isomorphism (see
2.8).
We were inspired by previous attempts to dene this notion of functor
for monoidal categories, and by the ensuing strictication results of Joyal
and Street in [72] (Section 1) and of Mac Lane in [102] (Sections XI.23). We do not, however, nd these denitions and results sucient for
our purposes, even when only the monoidal structure is strictied. We
need something more general. We envisage strictifying structures other
than just monoidal, and we will have occasion in this book to strictify also
with respect to symmetry (see 6.5, 7.6 and 8.4). Another limitation of
previous strictication results for monoidal categories is that they do not
take into account that the monoidal structure may be just a part of a more
complex ambient structure, and that the functors involved in equivalence
should preserve this ambient, not strictied, structure up to isomorphism.
To have just the monoidal structure preserved is rather useless from our
point of view (see 3.1).
Our results on strictication will be much more general, but they are
not such that they could not be further generalized. In particular, in dening the categorial structure preserved by our functors up to isomorphism
we have presupposed that this structure is dened only with covariant binary endofunctors. A natural generalization is to take here into account
also endofunctors of arbitrary arity, covariant in some argument-places and
contravariant in others. We suppose that our results can be extended to
cover such situations too. For the applications we need it was, however,
enough to cover the simpler situation, excluding contravariance, and we
did just that. We were afraid of complicating further a matter already full
of details, to prove results for which we have no immediate application. (As
Mac Lane says in [100], p. 103: ... good general theory does not search

1.7.

Strictification

31

for the maximum generality, but for the right generality.)

So our notion of logical system in the next chapter involves only conjunction and disjunction as binary connectives, together with the constants
and . Implication is excluded, and negation is left for the end of the
book. To cover these other connectives, we would need to extend our notion of logical structure to permit contravariance. We assume this can be
done in a straightforward manner at the cost of complicating notation. We
refrain, however, from doing so in this book, whose central piece is about
conjunctive-disjunctive logic, and where negation appears only at the end.
Anyway, as far as strictication goes, this limited notion of logical system
is sucient for our purposes.
Classical implication, dened in the usual way in terms of disjunction
and negation, does not come out as a very important connective in our
proof-theoretical perspective. It is not much of an implication, if the role of
implication is to help in mirroring the deductive metatheory in the object
language. Intuitionistic implication plays that role better.
Our results on strictication are still somewhat more general than what
we strictly need. In strictifying a binary connective like conjunction, purely
conjunctive formulae may be replaced by equivalence classes that correspond to sequences, or multisets, or sets, of the atomic formulae joined by
conjunction. For our purposes, we could have stuck to the rst two strictications, but with our general treatment we cover also the third. With
that, we stay within the limits of covariance.
Strictication, though an interesting topic on its own, is not absolutely
essential for our main topiccoherence. It is for us just a tool, we could
have dispensed with in principle. That would, however, be at the cost of
making already pretty long records even longer. So strictication is for us
a rather useful tool.
It is a tool more useful for recording computations than for discovering
how they should be done. Blurring distinctions may sometimes hinder this
discovery.
It is remarkable that the general notion of strictication may be found
implicit in Gentzens sequent systems, as we will try to explain in 11.1, in
the central chapter of the book.

Chapter 2
Syntactical Categories
In this chapter, which is of a preliminary character, we dene the notions of
syntactical categories needed for our work. In particular, we introduce the
notion of logical category (which should not be confused with the homonymous notion of [104], Section 3.4). Logical categories are obtained from
logical systems in a propositional language by replacing derivations with
equivalence classes of derivations. The equivalence producing these classes
is of general mathematical interest, but it has also proof-theoretical meaning, so that the equivalence classes may be identied with proofs. This
presupposes some notions of logic and category theory, which will be duly
dened.
Many of these notions are quite standard, and we go over them just
to x terminology. Something less standard may be found in the section
on denable connectives, where some intricacies inherent in this notion are
made manifest. A new matter is also detailed denitions of notions of
functors preserving the structure of a logical category. We are interested
in particular in those of these functors that preserve the structure up to
isomorphism. These denitions prepare the ground for Chapter 3. We treat
these matters in generality greater than we strictly need after that chapter.
It is not essential to master all the details we go into in order to follow the
exposition later on.
After these syntactical matters, we introduce at the end of the chapter
a category that will serve as the main model of our logical categories. This
model, which is in the realm of a semantics of proofs, and not in the realm
33

34

CHAPTER 2.

SYNTACTICAL CATEGORIES

of the usual semantics of propositions, is the category whose arrows are

relations between nite ordinalsa category tied to the notion of natural
transformation. This category will serve for our coherence results. Our
syntax is linked to this model by functors that preserve the structure on
the nose, i.e. up to an isomorphism that is identity.

2.1.

Languages

A language is a set of words, each word being a nite (possibly empty)

sequence of symbols. A symbol is a mathematical object of any kind. The
length of a word is the number of occurrences of symbols in it, and this
is the most standard measure of the complexity of a word. In particular
cases, however, we may rely on various other measures of complexity, like,
for example, the number of occurrences of some particular kind of symbol.
We introduce rst several languages of the kind logicians call propositional languages. Such languages are generated from a set P of symbols
called letters; logicians would call them propositional letters or propositional
variables. Sometimes we require that P be innite (see the end of 2.8),
but P can also be nite, and even empty. Since nothing in particular is
assumed about P, the symbols of P can be arbitrary mathematical objects,
and the denitions of notions built on P (such as that of logical system and
logical category; see 2.6-7 below) do not depend on the particular P that
was chosen.
Let be a symbol of the kind called in logic n-ary connective, for n 0.
A 0-ary, i.e. nullary, connective is more commonly known as a propositional
constant; 1-ary are unary connectives and 2-ary connectives are binary connectives. We assume, as usual, that P is disjoint from the set of connectives.
Then a language L such as we need is built up with inductive clauses of the
following kind:
(P)

P L,

()

if A1 , . . . , An L, then A1 . . . An L.

It is assumed here that is an n-ary connective. If n = 0, then A1 . . . An is

the empty sequence, and L. We have an analogous convention for all

2.1.

Languages

35

sorts of sequences that will appear in this work: if n = 0, then x1 . . . xn or

x1 , . . . , xn is the empty sequence, and {x1 , . . . , xn } is the empty set .
The elements of L are called formulae; logicians would say propositional
formulae. We use p, q, . . . , sometimes with indices, as variables for letters,
i.e. elements of P, and A, B, . . . , sometimes with indices, as variables for
formulae. The elements of P and nullary connectives are called atomic
formulae. The letter length of a formula is the number of occurrences of
letters in it.
We reserve for nullary connectives and for binary connectives. The
formula pqp, which is in the Polish, prex, notation, is more commonly
written ((p q) p), and we will favour this common, inx, notation for
binary connectives. Polish notation is handy for dealing with n-ary connectives where n 3, but in the greatest part of this work we will have just
nullary and binary connectives. A unary connective appears in Chapter 14.
(Notation for unary connectives that would not be Polish, like Hilberts
is uncommon in propositional logic; for nullary connectives
negation A,
there is no alternative.) We assume that we have as auxiliary symbols the
right parenthesis ) and the left parenthesis (, which are neither letters nor
connectives, with whose help we formulate the clause
if A, B L, then (A B) L.
This clause replaces () for binary connectives. As usual, we take the
outermost parentheses of formulae for granted, and omit them.
Consider a binary relation T on a set of elements called nodes such that
when xT y we say x is the predecessor of y, or y is the successor of x. A path
from a node x to a node y is a sequence x1 . . . xn , with n 1, such that x
is x1 and y is xn , while for every i {1, . . . , n1} we have xi T xi+1 . A root
is a node without predecessors, and a leaf a node without successors. We
say that a node is of n-ary branching, with n 0, when it has n successors.
So leaves are of nullary branching.
A nite tree is such a relation T where the set of nodes is nite, there is
exactly one root and every node except the root has exactly one predecessor.
It is clear that in every nite tree there is exactly one path from the root
to each node.

36

CHAPTER 2.

SYNTACTICAL CATEGORIES

The height of a node in a nite tree is the number of nodes in the path
from the root to this node. A nite tree is planar when all nodes of the same
height n 1 are linearly ordered by a relation <n such that if x1 T x2 and
y1 T y2 and x1 <n y1 , then x2 <n+1 y2 . When for two nodes x and y of the
same height we have x <n y with this linear order, we say that x is on the
left-hand side of y. Every formula of a language corresponds to a planar
nite tree whose leaves are labelled with letters and nullary connectives,
and whose remaining nodes are labelled with n-ary connectives for n 1.
The language L1 ,...,m has exactly 1 , . . . , m as connectives. We are
in particular interested in the languages where {, } and {, }.
These are the languages L , L , L, , L, , L, and L,,, .
We use the word subword as usual: every word is a subword of itself,
and if w1 w2 is a subword of a word w, then w1 and w2 are subwords of
w. A proper subword of a word w is a subword of w dierent from w. A
subformula of a formula is a subword that is a formula. The subformulae A
and B are the main conjuncts of A B, and the main disjuncts of A B.
Let w(A) be the word obtained by deleting all parentheses in a formula
A of a language L. We say that the formulae A and B of L are comparable
when w(A) and w(B) are the same word.
A place in A is a subword w of w(A). There is an obvious deleting map
from subwords of A to places in A. We say that a subword v of A is at a
place w when (v) = w . (Note that dierent subwords of A can be at the
same place.) For A and B comparable, a subword w1 of A and a subword
w2 of B are at the same place when (w1 ) = (w2 ).
We say, as usual, that an occurrence y of a symbol is within the scope of
an occurrence x of an n-ary connective in a formula A when in A we have a
subformula of the form xA1 . . . An with y being in Ai for one i {1, . . . , n}.
We say that y is within the immediate scope of x when y is within the scope
of x and there is no occurrence of a connective z within the scope of x such
that y is within the scope of z.

2.2.

Syntactical systems

A graph is a pair of functions, called the source function and the target
function, from a set of elements called arrows to a set of elements called

2.2.

Syntactical systems

37

objects. We use f , g, . . . , sometimes with indices, as variables for arrows

and a, b, . . . , sometimes with indices, as variables for objects. In many cases
in this work, objects will be formulae of a language such as we introduced
in the preceding section, but we also need the more general notion.
The expression f : a b means that the source function assigns a to f
and the target function assigns b to f ; we call a and b the source and target
of f , respectively. In category theory, is usually written , but we keep
for other purposes (for functions and implication), and we stress with
the logical turnstile symbol the proof-theoretical interpretation of our
work. We call a b, which is just a peculiar notation for the ordered pair
(a, b), the type of f : a b. A hom-set in a graph is the set of all arrows
of the same type for a given type. A graph where for every f, g : a b we
have f = g, i.e. where hom-sets are either empty or singletons, amounts to
a binary relation R on the set of its objects such that (a, b) R i there is
an arrow of type a b in the graph.
For a given graph G, the dual graph G op is dened by interchanging the
source and target functions; namely, the source function of G op is the target
function of G, and the target function of G op is the source function of G,
while the sets of objects and arrows are the same. An object b in a graph
G is terminal when for every object a in G there is a unique arrow of G of
a type a b, and b is initial in G when it is terminal in G op .
A deductive system (in the sense of [90], Section I.1) is a graph that
must have for every object a an identity arrow 1a : a a, and whose arrows
are closed under the partial operation of composition:
f:ab

g: b c

gf : a c
This fractional notation, taken over from the notation for rules in logic,
conveys that if f : a b and g : b c are in the deductive system, then
g f : a c is in the deductive system. We use an analogous notation in
other cases.
A deductive system is discrete when all of its arrows are identity arrows. A deductive system is a preorder when for every f, g : a b in this
deductive system we have f = g. A deductive system that is a preorder
amounts to a preordering, i.e. reexive and transitive, relation on the set

38

CHAPTER 2.

SYNTACTICAL CATEGORIES

of its objects. A preorder is a partial order when the preordering relation is

antisymmetric. Every discrete deductive system is a preorder, but not vice
versa. In principle, one can envisage the empty deductive system, with an
empty set of arrows and an empty set of objects, but we have no interest
in it for our work, and we will exclude it.
The notion of deductive system is a generalization of the notion of category. A category is a deductive system in which the following equations,
called categorial equations, hold between arrows:
(cat 1)

f 1a = 1b f = f : a b,

(cat 2)

h (g f ) = (h g) f.

This notion of category covers only small categories, but in this work, where
we have no foundational ambitions, we have no need for categories whose
collections of objects or arrows are bigger than sets. When we speak occasionally of the category Set of sets with functions, we assume that the
collection of objects of this category is the domain of a model of rst-order
axiomatic set theory, and hence it is a set. The functions between the elements of this domain also make a set. We make an analogous assumption
for other categories mentioned in this book that seem not to be small.
A syntactical system is a particular kind of deductive system where
arrows make an inductively dened language, whose members are called
arrow terms. Arrow terms are words dened inductively out of primitive
arrow terms with the help of symbols tied to partial or total nite operations
on arrow terms and the auxiliary symbols of right and left parentheses. A
subterm of a term is a subword that is a term.
Among the primitive arrow terms we must have the identity arrow
terms, which make the identity arrows of the deductive system (so we must
have them for every object), and among the symbols for operations on
arrow terms we must have one tied to composition:
f:ab

g: b c

(g f ) : a c
(As we said above, this is read: If f of type a b and g of type b c are
arrow terms, then the word (g f ) is an arrow term of type a c.) So,
ocially, parentheses in arrow terms involving are compulsory; but, as

2.3.

Equational systems

39

usual, we will omit outermost parentheses, and other parentheses if this can
be done without engendering ambiguity. Note that here is just a symbol.
The operation of composition tied to this symbol is the operation assigning
to the pair of words (f, g), of the types a b and b c respectively, the
word (g f ), of type a c.
We say that a graph G 1 is a subgraph of a graph G 2 when the objects
and arrows of G 1 are included respectively in the objects and arrows of G 2
and the arrows of G 1 have in G 1 the same source and target as in G 2 .
A deductive system D1 is a subsystem of a deductive system D2 when
D1 is a subgraph of D2 , the identity arrows of D1 are identity arrows in D2
and for every pair of arrows (f : a b, g : b c) of D1 their composition in
D1 is equal to their composition in D2 . A subcategory is a subsystem of a
category. A subcategory must be a category.
An arrow f in a deductive system is mono when for every g and h the
equation f g = f h implies g = h, and f is epi when for every g and h
the equation g f = h f implies g = h.
An arrow f : a b in a deductive system D is an isomorphism when
there is an arrow g : b a in D such that g f = 1a and f g = 1b . The
arrows f and g are here inverses of each other. Isomorphisms in categories
are mono and epi. A category in which every arrow is an isomorphism is a
groupoid. If there is an isomorphism of type a b, then a and b are said to
be isomorphic.
A subsystem D1 of a deductive system D2 is full when for every arrow
f : a b of D2 if the objects a and b are in D1 , then f is in D1 . The partial
skeleton A of a category A is a full subcategory of A such that for every
object a of A there is in A an object a isomorphic to a in A. (So every
category is a partial skeleton of itself.) If in this denition we require that
the object a be unique, then A is called simply a skeleton. A skeleton is
unique up to isomorphism of categories (see 2.4); so it is usual to speak
about the skeleton of a category.

2.3.

Equational systems

An equation in a syntactical system S is a word f = g where f and g are

arrow terms of S of the same type. An equational system E in S is a set of

40

CHAPTER 2.

SYNTACTICAL CATEGORIES

equations in S such that the following conditions are satised:

(re ) f = f is in E for every arrow term f of S;
(sy ) if f = g is in E, then g = f is in E;
(tr ) if f = g and g = h are in E, then f = h is in E;
(co ) if f1 = g1 , . . . , fn = gn for n 1 are in E, then of1 . . . fn = og1 . . . gn
is in E, where of1 . . . fn and og1 . . . gn are arrow terms of S produced
by an n-ary operation o on arrow terms.
For the congruence condition (co ) to make sense, the operation o must be
such that if fi is of the same type as gi , for i {1, . . . , n}, then of1 . . . fn is
of the same type as og1 . . . gn . We envisage only operations of this kind.
As in propositional languages above, when the arity of o is not greater
than 2, we favour the inx notation with parentheses; so we write f1 of2
(with outermost parentheses omitted) instead of of1 f2 .
Consider the smallest relation on the arrow terms of S that satises
f g if f = g is in E, which happens to be the equivalence relation on
the arrow terms of S such that f g i f = g is in E. With the help of
we build a deductive system called S/E. The objects of S/E are the objects
of S, and its arrows are equivalence classes [f ] of the arrow terms f of S
with respect to . The identity arrows of S/E are the equivalence classes
of the identity arrow terms of S, and for every n-ary operation o of S,
including in particular composition, we dene an operation on equivalence
classes by
o[f1 ] . . . [fn ] =df [of1 . . . fn ].
The condition (co ) above guarantees the correctness of this denition.
Most often, we do not write concrete equations, but equations with
variables, like the categorial equations in the preceding section, where f , g
and h are variables for arrows, while a and b are variables for objects. As
usual, we call equations with variables simply equations. We say that such
an equation belongs to an equational system E in S when every instance of
it, with arrow terms of S substituted for variables for arrows and names
of objects of S substituted for variables for objects, is an element of E.
In producing these instances we, of course, pay attention to types. For

2.3.

Equational systems

41

example, in instances of the categorial equation (cat 1) we have that a is

the source of f , while b is its target, and in instances of (cat 2) we have
that f , g and h have types that permit composition.
We say that an equation with variables holds in a graph G when every
substitution instance of it holds in G. (That such an instance holds in G
means, of course, that the names on the two sides of the equation sign
= name the same thing.) It is quite common to understand holding of
equations with variables in this universal manner, and that is how we will
understand it, unless stated otherwise. That is how we understood holding
for the categorial equations in the denition of category. Instead of saying
that an equation holds in G, we may say, synonymously, that it is satised
in G, or, simply, that we have it in G.
To name the arrows of S/E, we use the arrow terms of S, so that an
arrow term names the equivalence class to which it belongs. Synonyms of
name are designate, denote and stand for. Then every equation of E will
hold in S/E. We say that the arrow terms f1 and f2 of S are equal in S/E
when f1 = f2 holds in S/E, which is equivalent with the equation f1 = f2
belonging to E.
If the categorial equations belong to E, then S/E is a category, and, since
such categories arise out of syntactical systems, we call them syntactical
categories. We say that the category S/E is in the system S. If only
instances of f = f are in E, then S/E is S itself.
A set of axioms Ax of an equational system E is a proper subset of the
set of equations E such that E may be generated from Ax by closing under
the rules (sy ), (tr ) and (co ). The set of axioms need not be nite, and it
will usually be innite in this work. Every equation of E is either an axiom
in Ax or derived from previously obtained equations by applying one of the
rules. More formally, a derivation is a nite tree of equations whose leaves
are axioms, and where each node that is not a leaf is obtained from its
successors, i.e. from nodes immediately above, by applying the rules. The
root of the tree is the equation derived.
Instead of saying that an equation holds in S/E because we can derive
it in E, we will sometimes say more simply that we can derive the equation
for S/E. This way of speaking will often prove handier later in the book,
and should not cause confusion.

42

2.4.

CHAPTER 2.

SYNTACTICAL CATEGORIES

Functors and natural transformations

A graph-morphism F from a graph G 1 to a graph G 2 is a pair of maps, both

denoted by F , from the objects of G 1 to the objects of G 2 and from the
arrows of G 1 to the arrows of G 2 , respectively, such that for every arrow
f : a b of G 1 the type of the arrow F f of G 2 is F a F b.
A graph-morphism F from a graph G 1 to a graph G 2 is faithful when,
for every pair (f, g) of arrows of G 1 of the same type, if F f = F g in G 2 ,
then f = g in G 1 .
A functor F from a deductive system D1 to a deductive system D2 is
a graph-morphism from D1 to D2 such that in D2 the following equations
hold:
(fun 1)

F 1a = 1F a ,

(fun 2)

F (g f ) = F g F f.

Note that this denition of functor is more general than the usual one,
which envisages only functors between categories. Otherwise, it is the same
denition. We generalize similarly other notions introduced below.
The product D1 D2 of the deductive systems D1 and D2 is the deductive
system whose objects are pairs (a1 , a2 ) such that a1 is an object of D1 and
a2 an object of D2 , and analogously for arrows. The identity arrows of
D1 D2 are of the form (1a1 , 1a2 ), and composition is dened by
(g1 , g2 ) (f1 , f2 ) =df (g1 f1 , g2 f2 ).
A functor B from D1 D2 to D is called a bifunctor ; for bifunctors,
(fun 1) and (fun 2) amount to the following equations respectively:
(bif 1)

B(1a , 1b ) = 1B(a,b) ,

(bif 2)

B(g1 f1 , g2 f2 ) = B(g1 , g2 ) B(f1 , f2 ),

which we call the bifunctorial equations.

Let D0 be the trivial deductive system with a single object and a
single arrow 1 : . (This deductive system is a category.) Let Dn+1
be Dn D. It is clear that D1 is isomorphic to D. A functor from Dn to
D will be called an n-endofunctor in D. An object of D may be identied

2.4.

Functors and natural transformations

43

with a 0-endofunctor. We call a 1-endofunctor also just endofunctor, and a

2-endofunctor biendofunctor.
The identity functor I of D is the endofunctor in D for which we have
Ia = a and If = f . A functor from D1 to Dop
2 is called a contravariant
functor from D1 to D2 .
A natural transformation from a functor F1 from D1 to D2 to a functor
F2 from D1 to D2 is a family of arrows of D2 indexed by objects of D1
such that a is of the type F1 a F2 a, and the following equations hold in
D2 for f : a b an arrow of D1 :
F2 f a = b F1 f.
Consider now an m-endofunctor M and an n-endofunctor N in D, and
two functions : {1, . . . , m} {1, . . . , k} and : {1, . . . , n} {1, . . . , k}
where m, n 0 and k 0 (if m = 0, then {1, . . . , m} = ; if k = 0, then
we must have m = n = 0). Then M dened by
M (x1 , . . . , xk ) =df M (x(1) , . . . , x(m) )
and N dened analogously are k-endofunctors in D. (If m = 0, then
M (f1 , . . . , fm ) is M (1 ).)
A family of arrows of D such that for every sequence a1 , . . . , ak
of objects of D there is an arrow a1 ,...,ak of the type M (a1 , . . . , ak )
N (a1 , . . . , ak ) is called a transformation of D of arity k. We say that the
arrows f1 , . . . , fk of D, such that for i {1, . . . , k} the arrow fi is of the
type ai bi , ow through in D when the following equation holds in D:
( nat)

N (f1 , . . . , fk ) a1 ,...,ak = b1 ,...,bk M (f1 , . . . , fk ).

By the denition of natural transformation, a transformation is a natural

transformation from the k-endofunctor M of D to the k-endofunctor N
of D when every k-tuple of arrows of D ows through in D. We say
that a1 ,...,ak is natural in a1 , . . . , ak when it is a member of a natural
transformation. The equations ( nat) will be called naturality equations.
A natural transformation in a deductive system D is a natural isomorphism when each member of the family is an isomorphism. Two functors

44

CHAPTER 2.

SYNTACTICAL CATEGORIES

are naturally isomorphic when there is a natural isomorphism from one to

the other.
We say that the deductive systems D1 and D2 are equivalent via a
functor F2 from D1 to D2 and a functor F1 from D2 to D1 when the
composite functor F1 F2 is naturally isomorphic to the identity endofunctor
of D1 and the composite functor F2 F1 is naturally isomorphic to the identity
endofunctor of D2 . It is easy to conclude that the functors via which two
categories are equivalent are faithful functors.
The deductive systems D1 and D2 are isomorphic via a functor F2 from
D1 to D2 and a functor F1 from D2 to D1 when the composite functor
F1 F2 is equal to the identity endofunctor of D1 and the composite functor
F2 F1 is equal to the identity endofunctor of D2 . Two deductive systems
are said to be equivalent when there is a pair of functors via which they are
equivalent, and analogously for isomorphic deductive systems.
Suppose we have two syntactical systems S i , for i {1, 2}, together
with the equational systems E i in S i . A graph-morphism F from S 1 to
S 2 induces an obvious graph-morphism from S 1 /E 1 to S 2 /E 2 , such that
F [f ] = [F f ], provided f = g in E 1 implies F f = F g in E 2 . (We do not
write [f ] usually, but use the arrow term f to designate [f ].) When F from
S 1 to S 2 is a functor, then F is a functor from S 1 /E 1 to S 2 /E 2 .
When we have the graph-morphisms F1 from S 2 to S 1 and F2 from S 1
to S 2 such that S 1 /E 1 and S 2 /E 2 are isomorphic deductive systems via
functors induced by F1 and F2 , we say that S 1 and S 2 are synonymous up
to E 1 and E 2 via F1 and F2 . A stronger notion of synonymity of syntactical
systems, which we will usually encounter, is when F1 and F2 are functors
between S 1 and S 2 , and not any graph-morphisms.

2.5.

Definable connectives

Let L stand for one of the languages L , L , L, , L, , L, and L,,,

of 2.1, generated by an arbitrary set of letters P, and let L2 stand for the
language with the same connectives as L generated by the set of letters
{2}. We use M , N, . . . , sometimes with indices, for elements of L2 . Let
|M | = m 0 be the number of occurrences of 2 in M , and let w1 , . . . , wm
be a sequence of m arbitrary words. Later on in this work, these words will

2.5.

Definable connectives

45

denote either the objects or the arrows of a category. Then M (w1 , . . . , wm )

is the word obtained by putting wi , where i {1, . . . , m}, for the i-th
occurrence of 2 in M , counting from the left.
Let Lcon be a set of pairs (M, ), which we abbreviate by M , where
M L2 and |M | = m 0, while is a function from {1, . . . , m} to
{1, . . . , k} for some k 0. The arity of M is k. We dene M (w1 , . . . , wk )
as M (w(1) , . . . , w(m) ) (cf. the denition of M in the preceding section).
When w1 , . . . , wk are formulae of L, the elements of Lcon stand for
the denable connectives of L. Let us consider some examples of denable connectives. A primitive connective {, } of L is represented
in Lcon by the denable connective (2 2){1,2} where {1,2} is the identity function on {1, 2}, while {, } is represented by where is
the identity function on , which is the empty function (the only possible
function from to ). If {1} is the identity function on {1} (the only
possible function from {1} to {1}), then 2{1} is the identity unary connective, for which we have 2{1} (A) = A. If is the function from {1, 2}
to {1, 2} such that (x) = 3x, then for the denable connective (2 2)
we have (2 2) (A, B) = (2 2)(B, A) = B A. If is the only possible
function from {1, 2} to {1}, then for the denable connective (2 2) we
have (2 2) (A) = (2 2)(A, A) = A A. If is the constant function with
value 1 from {1} to {1, 2}, then for the denable connective 2 we have
2 (A, B) = 2(A) = A.
For M , N11 , . . . , Nkk elements of Lcon , we want to dene the element
M (N11 , . . . , Nkk ) of Lcon resulting from the substitution of N11 , . . . , Nkk
within M . In other words, we want to dene generalized composition
of elements of Lcon . This notion is rather simple when is an identity
function, but in the general case we have the following, more involved,
denition.
Let M be of arity k such that |M | = m, and let Nii , for i {1, . . . , k},
be of arity li with |Ni | = ni . To dene the element M (N11 , . . . , Nkk ) of
Lcon of arity

lj

1jk

we must rst dene what it means to substitute the functions 1 , . . . , k

46

CHAPTER 2.

SYNTACTICAL CATEGORIES

within the function .

Let
(i) =

1ji

n(j) ,

(i) =

lj ,

1ji

and let : {1, . . . , (m)} {1, . . . , m} be dened by

(x) =df min{i | x (i)}.
Next we dene the function (1 , . . . , k ) : {1, . . . , (m)} {1, . . . , (k)}
by
(1 , . . . , k )(x) =df ((x)) (x((x)1)) + (((x))1).
With the help of (1 , . . . , k ) we dene M (N11 , . . . , Nkk ) as
M (N1 , . . . , Nk )(1 ,...,k ) , which is equal to M (N(1) , . . . , N(m) )(1 ,...,k ) .
The denition of (1 , . . . , k ) is pretty opaque, and we must make a
few comments on it. We will consider as an example a simple case of
(1 , . . . , k ), which covers most of our needs in this book.
Let the function 1 + 2 : {1, . . . , n1 + n2 } {1, . . . , l1 + l2 } be dened
by
{
1 (x)
if x n1
(1 + 2 )(x) =df
2 (xn1 ) + l1 if x > n1 .
Then one can check that when m = k and is the identity function on
{1, . . . , m} we have
(1 , . . . , k ) = 1 + . . . + k .
The complications of the general denition of (1 , . . . , k ) above come
from the fact that we want to substitute within the function the functions
1 , . . . , k so that for every ordered pair (x, y) in we have a copy of
y . These complications are not essential for many of the latter parts
of our work. In many cases, we will have for M Lcon that is the
identity function {1,...,m} on {1, . . . , m}. We introduce the convention that
M {1,...,m} is abbreviated by M . With that in mind, the reader can forget
about the indices in M in many places. We have preferred, however, to
state our results later on in greater generality.

2.6.

2.6.

Logical systems

47

Logical systems

We will consider in this work a particular kind of syntactical system called

logical system. A logical system C has as objects the formulae of a language
L, as in the preceding section. We say that such a logical system C is in L.
The primitive arrow terms of C come in families . The members of are
indexed by sequences A1 , . . . , Ak , with k 0, of objects of C. With every
family we associate two elements M and N of L2 , such that |M | = m and
|N | = n, and two functions : {1, . . . , m} {1, . . . , k} and : {1, . . . , n}
{1, . . . , k}. The type of A1 ,...,Ak is M (A1 , . . . , Ak ) N (A1 , . . . , Ak ). So
an is a transformation in C.
In Table 1 we present most of the transformations we need for our
work. In this table, denotes the empty function from the empty set. The
types of the members of are as in Table 2. In the leftmost column of
that table we write down the name of the union of the families on the
right-hand side. So the b family includes the families, i.e. transformations,

b , b , b and b . Within the family b we have the subfamily b, which

includes b and b , and the subfamily b, which includes b and b .

We have, analogously, the subfamilies - and - of the - family, and

analogously in other cases.
Of course, an from L may be found also in a wider language L, .
In practice, one of and will be the identity function, as in the transformations in Table 1, but we allow for greater generality. For the sake
of uniformity, we decided to take always as the identity function in the
transformations with L and L , and as the identity function in the

transformations with L and L , but for c and c we could have done

otherwise. The dierence in indexing c A,B and c B,A sometimes requires

additional care when passing from matters involving to matters involving
, but it helps to enhance the duality underlying and .
The labels b, c, w and k are borrowed from the combinators B, C, W
and K of combinatory logic, d comes from dissociativity (see 1.2) and
m from mix (see 8.1). The Greek labels , and involve and .
As every syntactical system, a logical system C will have the family 1
from Tables 1 and 2, which delivers its identity arrow terms. If we work
with a language in which we have {, }, then for building the arrow

48

CHAPTER 2.

L
L
L
L
L

L,

L,
L,
L,

L,

L,
L,
L,
L
L
L
L

k1

k2

k1

L
L
L
L,
L,
L,
L,
any

k2

dL
dR
m
m1
1

SYNTACTICAL CATEGORIES

(x)

(x)

2 (2 2)
(2 2) 2

(2 2) 2
2 (2 2)

(2 2) 2
2 (2 2)

2 (2 2)
(2 2) 2

1
1
1

2
2
2
2

2
2
2
2

1
1
1

1
1
1

x
x
x

x
x
x

2
2
2
2
22
22
22
2

1
1
1
2
2
1
2

1
1
1
2
2
2
1

x
x
x
x
3x
x
1

x
x
x
3x
x
1
x

1
1
0
3
3
2
2
1

2
0
1
3
3
2
2
1

2
x
=
x
x
x
x
x

x
=
x
x
x
x
x
x

3
3

2
2
2
22
22
2
22
22

22

2
22

2
2

2 (2 2)
(2 2) 2
22
22
2

22

2
(2 2) 2
2 (2 2)
22
22
2

1
1
1
2
2
1
1

2
1
1
3
3
2
2
1

Table 1

2.6.

Logical systems

49

b
A,B,C : A(B C) (AB)C

b
A,B,C : (AB) C A(B C)

w-k

A:AA

A:AA

A:AA

A:AA

A:AA

A:AA

A:AA

c A,B : A B B A

c B,A : A B B A

wA : A A A

wA : A A A

k 1A,B : A B A

k 1A,B : A A B

b
A,B,C : (AB)C A(B C)

A:AA

b
A,B,C : A(B C) (AB)C

k 2A,B : A B B

k 2A,B : B A B

A:A

A:A
L
dA,B,C : A (B C) (A B) C
dR
C,B,A : (C B) A C (B A)
mA,B : A B A B
m1
A,B : A B A B
1A : A A

Table 2

50

CHAPTER 2.

SYNTACTICAL CATEGORIES

terms of C we have the following clause corresponding to a total operation

on arrow terms:
f:AD

g: B E

f g: A B D E
together with the clause corresponding to the partial operation of composition mentioned in 1.2, with a, b and c replaced by A, B and C. This
concludes our denition of logical system.
If is one of the families of primitive arrow terms we have introduced,
except the family 1, then we call -terms the set of arrow terms introduced
inductively as follows: every member of is a -term; if f is a -term, then
for every A in L we have that 1A f and f 1A are -terms.
In every -term there is exactly one subterm that belongs to , which is

called the head of the -term in question. For example, the head of the c

term 1A ( c B,C 1D ) is c B,C . An analogous denition where is 1, yields

arrow terms called complex identities (which are headless). Every complex
identity is equal to an identity arrow term in the presence of bifunctorial
equations.
If we build a language L(B) with the same connectives as L but with
the generating set P replaced by a set B of the same cardinality as P, then
we obtain an isomorphic copy of L. If B is not of the same cardinality
as P, then L(B) and L are not isomorphic, but one can be isomorphically
embedded into the other. So we have a function that assigns to B the
language L(B), and we call L(P) simply L.
Our notion of logical system is such that for a logical system C in L
we have a logical system C(B) in L(B). The logical system C(B) will be
isomorphic to C if P and B are of the same cardinality. The possibility to
build C(B) is ensured by requiring that the transformations be indexed
by all k-tuples of objects of C or C(B).
So what we have really dened with C in L is not a single logical system,
but a function that assigns to an arbitrary generating set P a logical system
C(P) in L(P), which we have chosen to denote by C and L, respectively,
not mentioning P. Applied to a dierent generating set B of letters this
function gives the logical system C(B) in L(B).

2.7.

2.7.

Logical categories

51

Logical categories

For an equational system E in a logical system C in L, we assume whatever

we have assumed for equational systems in a syntactical system, namely
the conditions (re ), (sy ), (tr ) and (co ), plus an additional condition. For
p
an arrow term f : A B of C, let fcp : ApC BC
be the arrow term of C
obtained by uniformly replacing every occurrence of a letter p of P in f
and in its type A B by the formula C of L. Then we assume closure of
E under substitution; namely,
p
(su ) if f = g is in E, then fCp = gC
is in E.

Closure under (su ) means that the letters of L behave like variables for
objects.
The equations of E will be introduced by axiomatic equations with variables in which letters of P do not occur. So we can assume these equations
for an arbitrary set P. This will also guarantee that E is closed under (su ).
When the categorial equations belong to the equational systems E in
a logical system C, so that C/E is a category, and, moreover, for every
{, } in the language L of C we have in E the bifunctorial equations
(bif 1) and (bif 2) of 2.4 with B instantiated by ; namely, the following
equations:
( 1)

1A 1B = 1A B ,

( 2) (g1 f1 ) (g2 f2 ) = (g1 g2 ) (f1 f2 ),

so that C/E has the biendofunctor , we call the syntactical category C/E
a logical category. We say that a logical category C/E is in L when C is
in the language L, and we also say that the category C/E is in the logical
system C.
If for the families of C the naturality equations ( nat) of 2.4 with
a1 , . . . ak and b1 , . . . bk replaced by A1 , . . . , Ak and B1 , . . . , Bk , respectively,
belong to E, then in a logical category C/E we will have the natural transformations from the k-endofunctor M to the k-endofunctor N in C/E.
That M and N are k-endofunctors in C/E is guaranteed by the bifunctorial equations. When the naturality equations belong to E for every of C
and C/E is a logical category we say that C/E is a natural logical category.

52

CHAPTER 2.

SYNTACTICAL CATEGORIES

We separate naturality from bifunctoriality in our denition of logical

category because there are reasons to envisage logical categories that need
not be natural (cf. 14.3), though in this book bifunctoriality and naturality
will go hand in hand. (We do not envisage rejecting bifunctoriality for
logical categories, as some authors do; see 14.3.)
Here are the naturality equations for the transformations in the tables
of the preceding section, with f : A D, g : B E and h : C F :

(b nat) ((f g) h) b
A,B,C = b D,E,F (f (g h)),

(b nat) (f (g h)) b
A,B,C = b D,E,F ((f g) h),

(b nat) ((f g) h) b
A,B,C = b D,E,F (f (g h)),

(b nat) (f (g h)) b
A,B,C = b D,E,F ((f g) h),

nat) f = (f 1 ),
( nat) f

A = D (f 1 ), (
A
D

nat) (f 1 ) = f,
( nat) (f 1 )

A = D f, (
A
D

nat) f = (1 f ),
( nat) f
A = D (1 f ), (
A
D

nat) (1 f ) = f,
( nat) (1 f )
A
D
A = D f, (

( c nat)

( c nat)

(w nat)

(g f ) c A,B = c D,E (f g),

(g f ) c B,A = c E,D (f g),

(f f ) wA = wD f,

(w nat)

f wA = wD (f f ),

(k 1 nat) f k 1A,B = k 1D,E (f g), (k 1 nat) (g f ) k 1B,A = k 1E,D g,

(k 2 nat) g k 2A,B = k 2D,E (f g), (k 2 nat) (g f ) k 2B,A = k 2E,D f,

( nat) 1 A = D f,
(dL nat)

( nat)

f A = D 1 ,

((f g) h) dL
A,B,C = dD,E,F (f (g h)),

(dR nat) (h (g f )) dR
C,B,A = dF,E,D ((h g) f ),

(m nat)
(f g) mA,B = mD,E (f g),
1
1
(m nat) (f g) m1
A,B = mD,E (f g),
(1 nat)

f 1A = 1D f.

One side of the equations ( nat) and ( nat) can, of course, be shortened
by using the categorial equations (cat 1), and (1 nat) is contained in (cat 1).

2.8.

C-functors

53

An arrow term of the form fn . . . f1 , where n 1, with parentheses tied to associated arbitrarily, such that for every i {1, . . . , n} we
have that fi is composition-free is called factorized. In a factorized arrow
term fn . . . f1 the arrow terms fi are called factors. A factorized arrow term fn . . . f1 is developed when f1 is of the form 1A and for every
i {2, . . . , n} we have that fi is a -term for some .
Then by using the categorial and bifunctorial equations we can easily prove by induction on the length of f the following lemma for logical
categories C/E.
Development Lemma. For every arrow term f there is a developed arrow
term f such that f = f .
Note that for a logical category C/E our way of introducing E by axiomatic equations with variables in which letters of the generating set P do
not occur is such that when P is replaced by another generating set B we
have instructions for building another logical category C/E(B) in the logical
system C(B) in the language L(B). This logical category C/E(B) will be
isomorphic to C/E if P and B are of the same cardinality. The axiomatic
equations with variables assumed for E are applied to the arrow terms of
C(B). We have really dened a function that assigns to an arbitrary generating set B a logical category C/E(B), the logical category C/E(P) being
denoted simply by C/E (cf. the end of the preceding section).
When the equational system E of a logical category C/E has as axioms
the elements of a set Ax of equations, and we speak of derivations of equations of E, we need not count (su ) among the rules of derivation if Ax is
closed under (su ). When later we produce sets of axioms, we always assume
that they are closed under (su ), so that the rules of derivation are just (sy ),
(tr ) and (co ). Hence, in general, Ax will be an innite set of equations,
though these equations are instances of a nite number of equations with
variables.

2.8.

C-functors

Let C be a logical system in L. Deductive systems that have

54

CHAPTER 2.

SYNTACTICAL CATEGORIES

an operation on objects and an operation on arrows for every

of L such that for f : a d and g : b e we have f g : a b d e,
an object

for every

of L, and

a transformation for every of C

are called deductive systems of the C kind. A bifunctorial category of the
C kind is a category of the C kind in which the bifunctorial equations hold
for every of C.
Let A1 and A2 be bifunctorial categories of the C kind. The operations
, the objects and the transformations are indexed by 1 and 2 when
they are in A1 and A2 respectively. The type of ai 1 ,...,ak , where i {1, 2},
is Mi (a1 , . . . , ak ) Ni (a1 , . . . , ak ).
A C-functor from A1 to A2 is made of
a functor F from A1 to A2 ,
for every of C a family 22 of arrows of A2 indexed by objects of
A1 whose members are
2 2
a,b
: F a 2 F b F (a 1 b),

for every

of C an arrow : 2 F 1 of A2 .

In practice, when we refer to a C-functor, we mention only F , taking the

families for granted. They will be mentioned explicitly when this is
required.
A dual C-functor from A1 to A2 is obtained from the denition of C2 2
functor by replacing a,b
by
2 2
a,b
: F (a 1 b) F a 2 F b

and by : F 1 2 .
For every C-functor from A1 to A2 and every M Lcon (see 2.5),

we dene in A2 the family of arrows M by induction, with the following

clauses:

2.8.

C-functors

55

a2 = 1F a ,

2 2
( M
aM,bN = M

a
(
a),N (b)

( M 1)
( M 2)

2 N ),
b

where a and b stand for a1 , . . . , am and b1 , . . . , bn respectively,

( M 3)

aM1 ,...,ak = aM(1) ,...,a(m) .

We have a dual denition of M for dual C-functors, where the clauses

while the clause ( M 2)

( M 1) and ( M 3) have just replaced by ,
is replaced by

N 22
) M (a),N (b) .
b

aM,b N = (aM

We say that the members of M are -arrows.

Let Mi be obtained from M by replacing and by i and i respectively. From the bifunctoriality of i in Ai we can deduce that Mi denes
a k-endofunctor in Ai . The word 2 i 2 stands for the 2-endofunctor i ,
the word 2 for the identity 1-endofunctor of Ai , and the word i for the
0-endofunctor i of Ai .

From the inductive denition of M we can deduce the following equation of A2 for every C-functor:

M (L

1
,...,Lk k )

( M ) a1 ,...,a1k

= M1
L1

(a1 ),...,Lk k (ak )

M2 (a11 , . . . , akk )

where ai stands for a sequence b1 , . . . , bl , with l 0, of objects of A1 , and

the Li i from the indices of M on the right-hand side are also from A1 .

It is easier to derive ( M ) when the functions , 1 , . . . , k are all identity

functions, and this easier equation is then used in the derivation of the

general ( M ) equation.
We say that a C-functor from A1 to A2 preserves an M of C (i.e. an

M of Lcon , for C in L) when aM1 ,...,ak is natural in a1 , . . . , ak (see 2.4);

this means that the following equation holds in A2 :
( nat)

F M1 (f1 , . . . , fk ) aM1 ,...,ak = bM1 ,...,bk M2 (F f1 , . . . , F fk ).

It can be checked that it is enough to assume the following instance of

( nat) for every in M :

56

CHAPTER 2.

SYNTACTICAL CATEGORIES

2
( 22 nat) F (f1 1 f2 ) a21,a
= b21,b22 (F f1 2 F f2 )
2

to derive by induction that F preserves M . The following instances of

( nat):
F 1 1 = 1 2 ,
F f a2 = b2 F f
follow from the functoriality of F and categorial equations.
We say that a C-functor from A1 to A2 preserves an of C when the
following equation holds in A2 :

() F a1 aM = aN

2
F
a

where if a is a1 , . . . , ak , then Fa is F a1 , . . . , F ak . We apply an analogous

convention concerning a and Fa also later.
We say that a C-functor is partial when it preserves every M and every
of C.
We say that a C-functor from A1 to A2 is uent in an of C when every
k-tuple of arrows ows through in A2 (see 2.4).
For every C-functor from A1 to A2 that preserves and is uent in
we have the following equation in A2 :
(L)

F 1 1
L1

(a1 ),...,Lk k (ak )

M (L

1
,...,Lk k )

a1 ,...,a1k
N (L

1
,...,Lk k )

a1 ,...,a1k

=
2 1

L1 (Fa1 ),...,Lk k (Fak )

where the Li i in the indices of 1 are from A1 , while those in the indices of

2 are from A2 . To derive (L) we just apply ( M ), () and uency

in . The equation () is the instance of (L) where every Li i is 2,
that is 2{1} .
We are aware that the condition (L), with its multiple indexing, may
look forbidding. Fortunately, in cases we deal with in this book, it will be
equivalent to the simpler uency in condition, as we will see in a moment.
A C-functor is called total when it preserves every M of C and (L)
holds for every of C. Every total C-functor is a partial C-functor. It

2.8.

C-functors

57

can be veried that the composition of two partial C-functors is a partial

C-functor, and the composition of two total C-functors is a total C-functor.
2 2
2 2
2 2
The arrow a,b
tied to the functor F3 F2 is dened as F3 2 a,b 3 F2 a,F2 b ,

and the arrow tied to the functor F3 F2 as F3 2 3 .

We say that a C-functor is groupoidal when it is a C-functor and a dual

C-functor with aM and aM being inverses of each other. For C-functors

where the M arrows are mono, (L) implies uency in , so that for
groupoidal partial C-functors, (L) is equivalent to uency in . The
composition of two groupoidal C-functors is a groupoidal C-functor.
Let us call maps from the set of letters P into the objects of a deductive
system Di of the C kind valuations into Di . A valuation vi into Di is
extended to two maps, both called vi , from the objects and arrow terms of
C to the objects and arrows, respectively, of Di with the obvious clauses
vi (A B) = vi (A) vi (B),
vi ( ) = ,
vi (A1 ,...,Ak ) = vi (A1 ),...,vi (Ak ) ,
vi (f g) = vi (f ) vi (g),
vi (g f ) = vi (g) vi (f ).
We can prove the following.
Proposition. Let A1 and A2 be bifunctorial categories of the C kind.
If F is a total C-functor from A1 to A2 , then for every arrow term f :
M (p1 , . . . , pm ) N (q1 , . . . , qn ) of C, for every valuation v1 into A1 and for
every valuation v2 into A2 such that v2 (p) = F v1 (p), in A2 we have
(t)

F v1 (f ) vM1 (p1 ),...,v1 (pm ) = vN1 (q1 ),...,v1 (qn ) v2 (f ).

Proof. We proceed by induction on the length of f . If f is

L1 (p ),...,Lk (p ) ,
1

then we use (L). If f is f1 f2 , then we use ( nat), the bifunctorial

equations for and the induction hypothesis. If f is f2 f1 , then we use
(fun 2) for F and the induction hypothesis.

Only a C-functor that satises (t) may be said to preserve the Cstructure properly, up to . But not everything in the denition of total

58

CHAPTER 2.

SYNTACTICAL CATEGORIES

C-functor is a consequence of (t). In particular, ( nat) with f1 , . . . , fk

foreign to the C-structure is not such a consequence.
A groupoidal total C-functor is called strong. A C-functor F from A1
2 2
to A2 is called strict when F a 2 F b = F (a 1 b), 2 = F 1 and a,b
and
are identity arrows of A2 . Every strict C-functor is, of course, strong.
For a strict C-functor F the equations ( nat) and (t) become
F M1 (f1 , . . . , fk ) = M2 (F f1 , . . . , F fk ),
F v1 (f ) = v2 (f ), where v2 (p) = F v1 (p),
while (L) is an easy consequence of (). Strict C-functors preserve the
C structure on the nose . (The expression on the nose is used in other
analogous situations, when a structure is homomorphically preserved in an
obvious manner.)
According to what we have said above about the composition of total
C-functors and groupoidal C-functors, the composition of two strong Cfunctors is a strong C-functor. The composition of two strict C-functors is,
of course, a strict C-functor.
These denitions are on the lines of Mac Lanes denition of monoidal

functor of [102] (cf. [7]). For example, with being b , the equation ()
and the inductive clauses for M yield the following equation:

22
2 22
22
22 2

F b
a,b,c a,b1 c (1F a b,c ) = a1 b,c (a,b 1F c ) b F a,F b,F c ,

which is used in [102] (Section XI.2) to dene monoidal functors. Fluency

is, however, only implicit for Mac Lane. Besides that, our denition, which
does not presuppose as Mac Lanes that monoidal functors are between
monoidal categories only, will enable us later to dene monoidal categories
via monoidal functors (see 4.6). We have analogous denitions via strict
C-functors for other sorts of categories too.
When every valuation into a bifunctorial category A of the C kind can
be extended to a strict C-functor from a logical category C/E to A we say
that A is a C/E-category relative to P. The logical system C is here a logical
system in the language L generated by the set of letters P.
If something is a C/E-category relative to an innite set of letters P,
then it must be a C/E(B)-category relative to any other generating set

2.9.

The category Rel and coherence

59

B. (Depending on the number of dierent variables for formulae in the

axiomatic equations with variables assumed for E, we could do here with
a nite set P of at least a certain cardinality instead of an innite set P,
but, assuming uniformly that P is innite, we are on the safe side.) A C/Ecategory relative to an innite set P is then called simply a C/E-category.
If C/E is a natural logical category, then a C/E-category A is a natural
C/E-category when for every in C the naturality equations hold in A.
The bifunctorial equations for every of C are guaranteed in every C/Ecategory, and uency in every of C is guaranteed for every C-functor into
a natural C/E-category.
When for f and g arrow terms of C of the same type we say that the
equation f = g holds in a deductive system A of the C kind, we understand
the letters p, q, . . . of L as variables for objects. If A is a C/E-category, then
that f = g holds in A amounts to saying that F f = F g holds for every
strict C-functor F from C/E to A. Every equation of E holds in this sense
in a C/E-category A, but additional equations f = g, not in E, may hold
in A too.

2.9.

The category Rel and coherence

The objects of the category Rel are nite ordinals (we have 0 = and
n+1 = n {n}), and its arrows are relations between nite ordinals. We
write either (x, y) R or xRy, as usual. In this category, 1n : n n is the
identity relation, i.e. identity function, on n. If n = 0, then 1 : is the
empty relation , with domain and codomain .
For R1 : n m (that is, R1 n m) and R2 : m k, the set of ordered
pairs of the composition R2 R1 : n k is {(x, y) | z(xR1 z and zR2 y)}. For
R1 : n m and R2 : k l, let the set of ordered pairs of R1 +R2 : n+k m+l
be
R1 {(x+n, y+m) | (x, y) R2 }.
With addition on objects, this operation on arrows gives a biendofunctor
in Rel.
The category Rel is a category of the C kind for every logical system
C whose families are from Tables 1 and 2 of 2.6, and, moreover, the

60

CHAPTER 2.

SYNTACTICAL CATEGORIES

appropriate bifunctorial and naturality equations will hold in Rel. The

biendofunctor for every {, } is +, and the object for every
{, } is 0. The natural transformation for every included in the
families b, -, d, m, m1 or 1 is the family 1 of Rel. In other cases, we
have the following:

(x, y) c n,m

(x, y) wn

(x, y) k 1n,m
(x, y)

k 2n,m

i
i

(y, x) c n,m

(y, x) wn

(y, x) k 1n,m

(y, x)

k 2n,m

i
i

(x+m = y or x = y+n);
x y (mod n);

x = y;

x = y+n;

the relations n : n and n : n are the empty relations.

It is not dicult to check that all these families in Rel are natural
transformations. This is clear from the diagrammatical representation of
relations in Rel. Here are a few examples of such diagrammatical representations, with domains written at the top and codomains at the bottom:
0q

1q
2q
3q
4q
@ @

@
@

@
@
q
q
q
@q @q

0q

1q

0q

c 2,3

1q
2q
3q
4q
J J J
J J J
J J J
J J J
J J J
q
q Jq Jq Jq

2q

0q

1q

2q

3q

4q

B
B
B
B
B
B
B
B
B
B
Bq Bq
0

1q

0q

B

B


B
 w
1
B 
B 
Bq

k 12,3

0q

A
A
A

A
A
A

A
A
A w
3



A
A A




A
A A

q
q
q
Aq Aq Aq

c 2,3

1q

2q
















q
q
q
q

k 13,2
q

2.9.

The category Rel and coherence

0q

61

1q

2q
3q
4q






 k 2


2,3






q
q
q

0q

1q

2q

0q

1q
2q
J J J
J J J

J J J
k 23,3
J J J
J J J
q
q Jq Jq Jq

2
q

For the identity relation, i.e. the identity function, 1n we have, of course,
0q

1q

n1
q

n1

Such diagrams are composed in an obvious manner by putting them one

below another; for example,
0q

1q
2q
3q
4q

@ @ @ @
k 14,1
@q @q @q @q

@ @
w2
@q @q

@
c 1,1
q @q
0

0q

J
J

1q

2q

J
q
Jq
0

3q

The equation

(w nat)

(f f ) wn = wm f,

4q

c 1,1 w2 k 14,1

62

CHAPTER 2.

SYNTACTICAL CATEGORIES

which is an instance of ( nat), and which we take as an example, is justied

in the following manner via diagrams:
n

@ @ w
@ @ n
@ @

B

B
 f B
 f B

B

B


B
 f B

B

@ w
@
@ m
@
@
@
m

We can now dene a function G from the objects of C to the objects of

Rel such that for p P, {, } and {, } we have
Gp = 1,
G = 0,
G(A B) = GA + GB.
Hence GA is just the letter length of A.
We can also dene a function, called G too, from the arrow terms of C
to the arrow terms of Rel such that
GA1 ,...,Ak = GA1 ,...,GAk ,
G(f g) = Gf + Gg,
G(g f ) = Gg Gf.
It is easy to check that for f : A B we have that Gf is of type GA GB.
It is also easy to check that if for f : A B we have (x, y) Gf , then the
(x+1)-th occurrence of letter in A, counting from the left, and the (y+1)-th
occurrence of letter in B are occurrences of the same letter.
For many logical categories C/E considered in this work, the two functions G we have just dened induce a strict C-functor G from C/E to Rel
such that G[f ] = Gf (see the penultimate paragraph of 2.4). Whenever
G is such a functor, it is straightforward to show that fact just by checking
that Rel satises the equations of E (with A, B, . . . replaced by n,m, . . .),
and we will not dwell on the proof of that.
The greatest part of our work consists in demonstrating the faithfulness
of such functors G. We call these faithfulness results coherence theorems,
and say that C/E is coherent.

2.9.

The category Rel and coherence

63

If the image of C/E under the functor G is a discrete subcategory of Rel,

which is the case when we exclude c, w, k and , then C/E is coherent i
C/E is a preorder. So, in such cases, our coherence theorems will state that
C/E is a preorder (which is the narrow sense in which Mac Lane understood
coherence originally in [99]).
It is clear that if C/E is coherent in the sense just specied, then it is
decidable whether arrow terms of C are equal in C/E. In logical terms,
one would say that the coherence of C/E implies the decidability of the
equational system E. This is because equality of arrows is clearly decidable
in Rel. So, in the terminology of 1.1, coherence here implies a solution to
the commuting problem.

Chapter 3
Strictification
This chapter is devoted to strictication (a topic announced in 1.7). Our
results are about categories that have as a subcategory a groupoid that
is a preorder. For such a category we nd an equivalent strictied category where the arrows of the groupoid are collapsed into identity arrows.
The functors on which this equivalence of categories is based are functors
that preserve structure up to isomorphism. The interest of strictication
is that it shortens the coding of arrows, and facilitates the recording of
computations.

3.1.

Strictification in general

We will prove a general theorem concerning the possibility of nding for a

C/E-category B a C/E-category B G equivalent to B via strong C-functors
such that some isomorphisms of the C/E structure of B, which make a
subcategory G of C/E, become identity arrows in BG , and may hence be
omitted according to the equation (cat 1) (see Chapter 11). So, instead of
computing in B we can pass to B G , where computations become shorter,
and their recording is simplied. The category BG is here called strict with
respect to the isomorphisms that have become identity arrows, and the
procedure of passing from B to B G is called strictication.
Our theorem will generalize considerably analogous strictication results
of Joyal and Street in [72] (Section 1) and of Mac Lane in [102] (Section
XI.3). First, we strictify with respect to wider classes of isomorphisms
65

66

CHAPTER 3.

STRICTIFICATION

G, such as we will encounter in our work, and not only with respect to
monoidal isomorphisms (for monoidal categories see 4.6). Secondly, even
when we strictify only with respect to monoidal isomorphismsi.e. when G
is a free monoidal categoryour C-functors may preserve a wider structure,
and not just the monoidal structure. They are not just monoidal functors.
As suggested by [72] and [102], strictication opens the way to alternative proofs of coherence results, in the sense of results about certain
categories being preorders. (Some authors go so far as to call strictication
results coherence results, but we believe this usage is confusing.) We will
obtain such alternative proofs of coherence in some cases, but in general
we favour the direct approach to proofs of coherence, in the style of [99]
and [100] (Section VII.2), which is not more dicult. (The alternative
proofs of coherence via strictication may look shorter when their presentation is sketchy.) The prime reason why we deal with strictication is not
the production of alternative proofs of coherencethis is only an occasional
byproduct. Our prime reason is a handy recording of lengthy computations,
as mentioned above. Strictication enables us to have shorter records after
coherence has been proved.
We formulate our strictication results with respect to a language L such
as we specied at the beginning of 2.5, because it is mainly in this context
that we mean to apply them. It will be clear, however, that analogous
results hold also for contexts with richer languages.
Let L be a language such as in 2.5, and let the following condition be
satised:
(I B)

B is a category that for every connective and of L has, respectively, a binary and a nullary operation on objects, denoted
by and .

Let L(Bob ) be the language with the same connectives as L generated by

the set Bob of the objects of B instead of P. The elements of L(Bob ) are
also formulae. To distinguish the and of L(B ob ) from those of Bob we
write G and G for the connectives and of L(Bob ). The connectives G
and G are new connectives, not to be confused with the operations and
involved in the objects of B.
Let G be an equivalence relation on L(B ob ) such that if a and b are

3.1.

Strictification in general

67

objects of B we have a G b only if a is the same object as b. We call such

an equivalence relation generatively discrete. Let [A] be the equivalence
class of the formula A of L(B ob ) with respect to G .
Out of B we build a category B G in the following manner. The objects
of B G are all the classes [A] for A L(B ob ). We use X, Y , Z, . . . , sometimes
with indices, for the objects of BG .
Consider the map E from L(B ob ) to B ob dened inductively by
Ea = a if a B ob ,
EG = ,
E(A G B) = EA EB,
where and on the right-hand side are the and of B. With the help
of E we dene a map F from the objects of B G to B ob in the following
manner. We choose rst a xed representative formula AF [A] so that
if A is an object a of B, then AF is a. We have guaranteed above by
generative discreteness that in [a] we have no object of B dierent from a.
Otherwise, the choice of the representative AF is arbitrary. Then we dene
F [A] as EAF .
The arrows of B G are all the triples (f, X, Y ) such that f : F X F Y
is an arrow of B. The arrow (f, X, Y ) is of type X Y in B G . The arrow
(1F X , X, X) is an identity arrow of B G , and for (f, X, Y ) and (g, Y, Z)
arrows of B G we dene their composition in B G as (g f, X, Z). This denes
the category B G .
Out of the map on objects F we dene a functor F from BG to B by
setting F (f, X, Y ) = f .
We dene a functor F G from B to B G by
F G a = [a],
F G f = (f, F G a, F G b),

for f : a b.

For the denition of F G f to be correct we must have that the type of f is

F F G a F F G b and this is guaranteed by F F G c = F [c] = c. It is trivial to
check that F and F G are indeed functors.
Let (I) be the collective name for (I B) and the condition that G is generatively discrete (i.e. the name for the conjunction of these two conditions).
Then we can prove the following lemma.

68

CHAPTER 3.

STRICTIFICATION

Lemma 1. If (I) holds, then the categories B and B G are equivalent via the
functors F G and F .
Proof. We have F F G a = a, as we noted above, and we have also F F G f =
f . On the other hand, F G F [A] = F G EAF = [EAF ]. Note that EAF is a
generator in L(B ob ), and though [A] = [AF ], the object [A] may well dier
from [EAF ]. However, we have in B G the natural isomorphism whose
members are X = (1F X , F G F X, X).

Note that to prove this lemma, we use just f 1a = 1b f and 1a 1a = 1a ,

which are consequences of (cat 1), so that we could generalize the lemma
to deductive systems B that are not categories.
Let the following conditions, called collectively (II), which strengthen
(I), be satised:
(II C) C is a logical system in L;
(II B) B is a bifunctorial category of the C kind;
(II G) G is a generatively discrete equivalence relation on L(B ob ).
We dene, as before, the category BG and the functors F and F G starting from the equivalence relation G on L(B ob ). We ensure that B G is a
bifunctorial category of the C kind with the following denitions:
X1 X2 =df F G (F X1 F X2 ), where
= [F X1 F X2 ],

is the

of B,

(f1 , X1 , Y1 ) (f2 , X2 , Y2 ) =df F G (F (f1 , X1 , Y1 ) F (f2 , X2 , Y2 ))

=
=

=df F G = [ ], where

F G (f1 f2 ), where is the

(f1 f2 , X1 X2 , Y1 Y2 ).

is the

of B,

of B.

It is easy to check that the bifunctorial equations hold for in B G , because

they are inherited from B. Then we dene the k-endofunctor (M ) out
of the k-endofunctor M , where M Lcon , just by replacing and
everywhere by and respectively, so that we have (M ) (X1 , . . . , Xk ) =
F G M (F X1 , . . . , F Xk ). The following:

X
=df (F X1 ,...,F Xk , (M ) (X1 , . . . , Xk ), (N ) (X1 , . . . , Xk ))
1 ,...,Xk

3.1.

Strictification in general

69

completes the denition of the C structure. It is easy to obtain that

F (M ) (X1 , . . . , Xk ) = M (F X1 , . . . , F Xk ), and analogously for N .
Let B with the C-structure be denoted by B, M, and let BG with the
C-structure we have just dened be denoted by B G , M , . Then we can
check the following lemma in a straightforward manner.
Lemma 2. If (II) holds, then the functors F G and F are strict C-functors
from B, M, to B G , M , and vice versa, respectively.
As a corollary of Lemma 2 we obtain that if B, M, is a C/E-category
for a logical category C/E, then B G , M , is a C/E-category too. A little
bit of work is required only to demonstrate the bifunctorial equations for
, which follow from the denitions of B G and , as we noted above.
It is easy to see that if (II) holds, and for an of C we have that is a
natural transformation in B, then is a natural transformation in BG .
We call a logical system C generatively discrete when for every arrow
term of C of type p q we have that p is the same letter as q.
Let the following conditions, called collectively (III), which strengthen
(II), be satised:
(III C) C is a logical system in L and C is a generatively discrete subsystem of C also in L, so that C and C have both as objects the
formulae of L; we abbreviate this condition involving C, C and
L by C L C;
(III G) G is a logical category of the C kind that is a groupoid;
(III B) B is a G-category and B, M, is a category of the C kind.
Since G is a logical category, it is equal to C /E for an equational system
E in C .
Let G(B) be dened as G save that the letters of P are replaced by the
objects of B (see 2.7). More precisely, we proceed as follows. First, instead
of L generated by P we have L(B ob ) generated by the set Bob of objects
of B, as above, with new G and G . The set L(B ob ) is the set of objects
of G(B). Next, we build the arrow terms of the logical system C (B ob ) by
indexing every of C with k-tuples of formulae of L(Bob ) instead of ktuples of formulae of L, and then closing under G for every of C and

70

CHAPTER 3.

STRICTIFICATION

composition. Finally, the axiomatic equations with variables assumed for

E in C are now interpreted in C (B ob ). The variables for formulae range
over the objects of C (Bob ), and the connectives and in these equations
now apply to G and G of L(B ob ). The logical category C /E (Bob ) is G(B)
(see the end of 2.7).
Let G be the binary relation on the objects of G(B) dened by A G B
i there is an arrow of G(B) of type A B. Since G, and hence also G(B), are
groupoids, G is an equivalence relation. Since C is generatively discrete,
the relation G is generatively discrete; i.e., no two dierent objects of B
are in the relation G . Let [A] be, as before, the equivalence class of an
object of G(B) with respect to G .
If (III) is fullled, then B, which is a G-category, is a G(B)-category too
(see the end of 2.8), and so the valuation that assigns to every generating
object a of G(B) the object a of B itself can be extended to a strict C functor E from G(B) to B. Intuitively, E erases every superscript G in G
and G . We already introduced, when we dened the functor F from B to
B G , the function E on objects of the functor E.
For every object A of G(B) we have an isomorphism A : AF A of
G(B) where AF is the chosen representative from [A]. The inverse of A
is the arrow 1
A : A AF of G(B). A natural choice for AF : AF AF ,
and in particular for a : a a, would be identity arrows, but this choice
is not essential for the time being. It is also not essential to dene A by
induction on the complexity of A. (Such an inductive denition of A is
possible if the representative AF is chosen in a particular canonical way;
cf. 4.5.) If, however, G, and hence also G(B), are preorders, then AF
must be 1AF . For every isomorphism A of G(B) we have the isomorphism
EA : EAF EA in B.
We can then dene the following C structure in B G , dierent from the
B G , M , structure. On objects we have
[A] [B] =df [A G B] = [AF
=df [ G ].

BF ],

That the denition of is correct is guaranteed by the fact that if we have

the isomorphisms f : A A1 and g : B B1 of G(B), then we have also the
isomorphism f G g : A G B A1 G B1 .

3.1.

Strictification in general

71

We dene the following arrows of BG :

2 2
[A],[B]
=df (EAF G BF , [A] [B], [A] [B]),

=df (EG , , ),
2 2
=df (E1
, [A] [B], [A] [B]),
[A],[B]
A G B
F

=df (E1
, , ).
G
It is easy to check that the source of EAF G BF is F ([A] [B]), and its
target F ([A] [B]). The source of EG is F and its target is F (which
is equal to ). Note that in the denition of 22 and 22 we need the
G
G
arrows A and 1
A only for A being of the form AF BF and . We have
1
no use for other and arrows.
We dene the following operation on the arrows of BG :
2
2 2
((f , X , Y ) (f , X , Y ))
(f1 , X1 , Y1 ) (f2 , X2 , Y2 ) =df Y21 ,Y
1
1
1
2
2
2
X1 ,X2 .
2

It is clear that, since EA is an isomorphism, the bifunctorial equations

hold for in B G .
For every M Lcon , let (M ) be obtained by replacing and
2 2
with and . Starting from X
and , with the help of , we de1 ,X2

M
ne X
: (M ) (X1 , . . . , Xk ) (M ) (X1 , . . . , Xk ) by the inductive
1 ,...,Xk

2
clauses X
= 1X = (1F X , X, X), which replaces ( M 1), and the clauses

( M 2), with 2 replaced by , and ( M 3) of 2.8. We dene analoM

M
gously with the help of the inverse X
of X
. Then we have
1 ,...,Xk
1 ,...,Xk
the following denition in B G :

X
=df X
X1 ,...,Xk X1 ,...,Xk .
1 ,...,Xk
1 ,...,Xk

This denes the C structure in BG , and we can conclude that B G , M ,

is a bifunctorial category of the C kind.
It is straightforward to show by induction on M L2 that in B G we
have
(M ) (M ) ((f1 , X1 , Y1 ), . . . , (fk , Xk , Yk )) =

M
.
YM1 ,...,Yk (M ) ((f1 , X1 , Y1 ), . . . , (fk , Xk , Yk )) X
1 ,...,Xk
We introduce the following denitions:

72

CHAPTER 3.

STRICTIFICATION

2 2
2 2
,
X1 ,X2 =df X
1 ,X2

=df ,

2 2
2 2
X1 ,X2 =df X
,
1 ,X2

=df .

22

Then starting from X1 ,X2 and , with the help of

dene

and not

we

X1 ,...,Xk : (M ) (X1 , . . . , Xk ) (M ) (X1 , . . . , Xk )

2

by the inductive clauses X = 1X , which replaces ( M 1), and the clauses

( M 2) and ( M 3). We dene analogously with the help of the inverse

M
M
X1 ,...,Xk of X1 ,...,Xk . Then we have the following in B G :

M
M
X1 ,...,Xk = X
,
1 ,...,Xk

M
M
X1 ,...,Xk = X
.
1 ,...,Xk

To show these equations, we use the equation ( M ) of 2.8, its dual for ,

and the equation (M ) above.

It follows now immediately that in B G we have the following:

X
= X1 ,...,Xk X
X1 ,...,Xk .
1 ,...,Xk
1 ,...,Xk

Since the arrows are dened in terms of , which is easier to handle

than , we will have occasion to apply later this equation, which we call
the alternative denition of . We can prove the following lemma.
Lemma 3. If (III) holds, then the identity functor IBG of B G with 22 and
is a groupoidal partial C-functor from B G , M , to B G , M , , and
IBG with 22 and is a groupoidal partial C-functor from B G , M ,
to B G , M , .
Proof. The equations ( nat) follow from (M ), while the equations ()
follow immediately from the denition of , or the alternative denition
of .

Note that with Lemma 3 we have asserted that and are naturally
isomorphic biendofunctors of BG . Showing that 22 and 22 involved in

3.1.

Strictification in general

73

this isomorphism are natural transformations does not presuppose that the
and 1 arrows in terms of which we dene 22 and 22 are members
of natural transformations. We just assume that and 1 arrows are
isomorphisms of G(B).
Suppose C L C, as in (IIIC). For every logical category C/E we can
determine out of C/E and C a logical subcategory C /E of C/E by restricting
the equations of E to the transformations of C .
We say that C /E ows through C/E when for every of C every ktuple of arrows of C /E ows through in C/E. If C/E is a natural logical
category, then for every subsystem C of C we have that C /E ows through
C/E.
Suppose C L C. For any deductive system B of the C kind there is
a least subsystem of B of the C kind, which we call the C -core of B. We
build the C -core of B by taking for objects all the objects of B; for arrows
we take the members of the transformations of B for every of C with
the same sources and targets as in B, and then close under composition
and the operations on arrows of B for every of C , i.e. of L. Note that
the C -core of B need not be a syntactical system: it inherits the equations
between arrows of B. As a limit case, we can determine also the C-core of
B, which need not coincide with B.
If B is a bifunctorial category of the C-kind, then its C -core is a bifunctorial category of the C -kind; if B is a C/E-category, then its C -core
is a C /E -category; and if B is a natural C/E-category, then its C -core is a
natural C /E -category. The C -core of C/E is C /E .
Consider now the following conditions, called collectively (IV), which
strengthen (III):
(IVC) C L C, as in (III C), and C/E is a logical category in L;
(IVG) the C -core C /E = G of C/E is a groupoid and G ows through
C/E;
(IVB) B, M, is a C/E-category.
Lemma 3 holds if (III) is replaced by (IV) and groupoidal partial Cfunctor is replaced by strong C-functor. Here is this new version of
Lemma 3.

74

CHAPTER 3.

STRICTIFICATION

Lemma 3(IV). If (IV) holds, then the identity functor IBG of BG with 22
and is a strong C-functor from B G , M , to B G , M , , and IBG
with 22 and is a strong C-functor from B G , M , to B G , M , .
Proof. We appeal to Lemma 3, and check, moreover, the uency of the

identity functor, with the arrows M or M , in every , or we check

directly (L) (which is slightly more complicated).

As a corollary, we obtain the following lemma.

Lemma 4. If (IV) holds, then the functor F G , with 22 and , and the
functor F , with F 22 and F , are strong C-functors from B, M, to
B G , M , and vice versa, respectively.
Proof. As we noted in 2.8, the composition of two strong C-functors is a
strong C-functor. Since strict C-functors are strong, Lemmata 2 and 3(IV)
deliver our lemma.

If in Lemma 4 the condition (IV) is replaced by (III), then we can only

arm that the functors in question are groupoidal partial C-functors. If
(IV) holds, then from Lemma 1 we obtain that B, M, and B G , M ,
are equivalent categories via the two strong C-functors of Lemma 4. We
can prove the following lemma.
Lemma 5. If (IV) holds, then B G , M , is a C/E-category.
Proof. Note rst that the bifunctorial equations hold for in B G , as
noted after the denition of on arrows. So B G , M , is a bifunctorial
category of the C kind.
Then take a valuation v that maps the letters of P into the objects of
B G . As mentioned after Lemma 2, the valuation v can be extended to a
strict C-functor v from C/E to B G , M , such that v (p) = v(p).
Suppose that the equation f = g for f, g : M (p1 , . . . , pm ) N (q1 , . . . , qn )
belongs to E. Then we know that v (f ) = v (g) holds in B G , and hence we
have in B G the following equation too:
vN(q1 ),...,v(qn ) v (f ) vM(p1 ),...,v(pm ) = vN(q1 ),...,v(qn ) v (g) vM(p1 ),...,v(pm ) .
By Lemma 3(IV) and the Proposition of 2.8, we conclude that for the

3.1.

Strictification in general

75

maps v from the objects and arrow terms of C to the objects and arrows
of B G , M , dened by v (p) = v(p) and the inductive clauses as for vi ,
which we gave just before that Proposition, we have v (f ) = v (g). So the
valuation v is extended to a strict C-functor v from C/E to B G , M , ,
which proves the lemma.

It is easy to see that if (III) holds, and for an of C we have that is

a natural transformation in BG , then is a natural transformation in B G .
So, together with the comment we made before introducing (III), we can
conclude that if (III) holds, and for an in L we have that is a natural
transformation in B, then is a natural transformation in BG .
A deductive system A of the C kind is called C-strict when for every of C we have that the members of the transformation in A are
identity arrows. This presupposes that the objects M (a1 , . . . , ak ) and
N (a1 , . . . , ak ) are equal in a C-strict deductive system, though in C the
formulae M (A1 , . . . , Ak ) and N (A1 , . . . , Ak ) need not be equal. A bifunctorial category A of the C-kind is C-strict i its C-core is discrete. We
can now prove the following lemma.
Lemma 6. If (IV) holds and G is a preorder, then B G , M , is C -strict.
Proof. Suppose G is a preorder and take an of C . Then in G(B) we have

an arrow A1 ,...,Ak : (M G ) (A1 , . . . , Ak ) (N G ) (A1 , . . . , Ak ) where (M G )

and (N G ) are obtained from M , N Lcon by replacing and by

G and G respectively. Hence [(M G ) (A1 , . . . , Ak )] = [(N G ) (A1 , . . . , Ak )]

and hence (M ) ([A1 ], . . . , [Ak ]) = (N ) ([A1 ], . . . , [Ak ]).

The arrow [A
is so of the type Y Y for some object Y of
1 ],...,[Ak ]
G

B . So the arrow F [A1 ],...,[Ak ] of B is of the type F Y F Y , and since

this arrow belongs to the C -core of B, with G being a preorder we ob

tain F [A
= 1F Y in B. From that we obtain that [A
is
1 ],...,[Ak ]
1 ],...,[Ak ]
G
(1F Y , Y, Y ), which is an identity arrow in B .

Lemmata 1, 4, 5 and 6 yield the following theorem.

Strictification Theorem. If (IV) holds and G is a preorder, then
B, M, is equivalent to the C -strict C/E-category B G , M , via the
strong C-functors F G and F .

76

CHAPTER 3.

STRICTIFICATION

As a corollary we obtain the following.

Strictification Corollary. If C /E is a natural logical category that is
a groupoid and a preorder, then every C /E-category B, M, is equivalent
to the C -strict C /E-category B G , M , via the strong C -functors F G
and F .
In [72] and [102] one nds the instance of this corollary where C /E is the

free monoidal category generated by P, i.e. our category L of 4.6. (From

[72] and [102] one could get the wrong impression that something peculiar to
monoidal categories has been discovered, while a more general result, stated
in our Strictication Theorem and Strictication Corollary, looms behind.)
We have no use, however, for the Strictication Corollary. Instead, we will
rely on the stronger Strictication Theorem to record long computations
concerning B with the help of BG , as mentioned at the beginning of the
section (cf. Chapter 11). The corollary is not sucient for that, because
it does not take into account the unstrictied C structure foreign to the
strictied C structure. This structure is not preserved by a functor that is
just a C -functor and not also a C-functor.
Suppose C L C, as in (III C). Then we say that a C/E-category B can
be (C/E, C )-strictied when there is a C -strict C/E-category B equivalent
to B via two strong C-functors. We can prove the following lemma.
Lemma 7. If C L C holds and every C/E-category can be (C/E, C )strictied, then for the C -core of C/E we have that every arrow in it is an
isomorphism of C/E, this C -core ows through C/E and it is a preorder.
Proof. If every C/E-category can be (C/E, C )-strictied, then the logical category C/E itself can be so strictied. So there is a C -strict C/Ecategory (C/E) equivalent to C/E via the strong C-functors F, 22 ,

from (C/E) to C/E and F , 2 2 , from C/E to (C/E) . Let be the

natural isomorphism of C/E whose members are A : F F A A, and let
1
A
be the inverse of A . Since (C/E) is a C/E-category, every valuation
from P to the objects of (C/E) can be extended to a strict C-functor. Let
v be the strict C-functor from C/E to (C/E) such that v (p) = F p. (Note
that the functors v and F need not coincide.)

3.1.

Strictification in general

77

Let f : M (
p) N (q) be an arrow of the C -core of C/E. Then v (f ) :
v M (
p) v N (q) is in the C -core of (C/E) , and is hence an identity arrow
of (C/E) . By the Proposition of 2.8, we have in (C/E)
M

N M

F f = qN
v (f ) p
= q
p
.

(The functor corresponding to v1 in (t) of the Proposition of 2.8 is here

the identity functor of C/E.) Hence,
1
1

N M
f = N
p) ,
p) = N (
p
) M (
(
q ) F F f M (
q ) F ( q

which proves that f is an isomorphism of C/E.

For f1 , . . . , fk in the C -core of C/E, with fi of type Ki (
pi ) Li (
qi ) for
i {1, . . . , k}, by using the Proposition of 2.8 and ( nat), we have in
(C/E)
F (N (f1 , . . . , fk ) K1 (p1 ),...,Kk (pk ) )
N (L ,...,Lk )

= q1 ,...,q1k

N (L ,...,Lk )

= q1 ,...,q1k

N (K ,...,Kk )
F
p1 ,...,p1k
K1 (p1 ),...,Kk (pk )

M (K ,...,Kk )
v (K1 (p1 ),...,Kk (pk ) ) p1 ,...,p1k

= F (K1 (p1 ),...,Kk (pk ) M (f1 , . . . , fk )).

From that we easily infer, by applying F to both sides and composing with
members of and 1 , that the C -core of C/E ows through C/E.
Take now f, g : M (
p) N (q) in the C -core of C/E. Since v (f ) and
v (g) are the same identity arrow of the C -core of (C/E) , we obtain, as
above,
M

F f = N
q
p
= F g,

from which f = g follows.

It is easy to infer from our Strictication Theorem and Lemma 7 the

following proposition.
Strictification-Coherence Equivalence. If (IV) holds, then every
C/E-category can be (C/E, C )-strictied i the C -core of C/E is a preorder.

78

CHAPTER 3.

STRICTIFICATION

As a corollary of this equivalence we have that if C is generatively discrete and C/E is a natural logical category that is a groupoid, then every
C/E-category can be (C/E, C)-strictied i C/E is a preorder. Such a statement was suggested by [72] and [102] for the particular case when C/E is
the free monoidal category generated by P.
A result analogous to Strictication-Coherence Equivalence is the following proposition:
If (IV) holds, then C/E can be (C/E, C )-strictied i the C -core
of C/E is a preorder.
We can also infer the following:
If (IV) holds, then every C/E-category can be (C/E, C )-strictied
i C/E can be (C/E, C )-strictied.
If our goal is to use strictication to prove preorder, then appealing to
the possibility of strictifying every C/E-category is irrelevant. The following
statement, which is a corollary of Lemma 7, suces.
Strictification-Coherence Implication. If C is generatively discrete,
C/E is a natural logical category that is a groupoid and C/E can be (C/E, C)strictied, then C/E is a preorder.
We will rely on this implication to give alternative proofs of Associative
Coherence and Monoidal Coherence in 4.5 and 4.7.

3.2.

Direct strictification

The procedure of strictication of the preceding section can be simplied

if the category B we want to strictify is a logical category. Then we can
build a C -strict category simpler than B G equivalent to B via two strong
C-functors, one of which is even strict. This category replacing B G , though
not logical, will be a syntactical category, like the logical category B.
Suppose the conditions (IV C) and (IV G) of the preceding section are
fullled. Let G be the binary relation on L dened by A G B i there is
an arrow of type A B in G. Since G is a groupoid, G is an equivalence

3.2.

Direct strictification

79

relation. Since C is generatively discrete, no two dierent letters of P are

in the relation G .
The objects of the syntactical system CG are all the equivalence classes
|[A]| with respect G for A a formula of L. We denote such classes by
X, Y , Z, . . . , sometimes with indices. On the objects of CG we dene the
operations and by
|[A]| |[B]| =df |[A B]|,

=df |[ ]|.

The denition of is correct because G is a logical system in L.

For every arrow term f of C let fG be the arrow term obtained by
replacing every letter p of P in the indices of f by |[p]|. If the arrow term f
of C is of type A B, then the arrow term fG of CG is of type |[A]| |[B]|.
The arrow terms of CG are the arrow terms fG for every arrow term f of
C. As a subsystem of CG we have the syntactical system CG obtained from
C as CG is obtained from C. Since G is a groupoid, every arrow term of CG
is of the type |[A]| |[A]| for some A in L. The equational system EG in CG
is obtained from the axiomatic equations with variables assumed for E (so
that for every equation f = g in E we have fG = gG in EG ) by adding for
every arrow term fG : |[A]| |[A]| of CG the equation fG = 1|[A]| , which we
call a strictifying equation, and then closing under (sy ), (tr ) and (co ) (see
2.3).
It is clear that the syntactical category CG /EG is a C/E-category. If
C/E is a natural logical category, then CG /EG is a natural C/E-category.
Moreover, CG /EG is C -strict. We will show that when G is a preorder the
syntactical category CG /EG is equivalent to C/E via two strong C-functors,
one of which, going from C/E to CG /EG , is even strict.
First we dene as follows a graph-morphism HG from C to CG :
HG A =df |[A]|,
HG f =df fG .
This graph-morphism induces a functor HG from C/E to CG /EG (see the
penultimate paragraph of 2.4).
To dene a functor from CG /EG to C/E we rst choose in every |[A]| a
xed representative AH , so that the representative of |[p]| is p for every p of

80

CHAPTER 3.

STRICTIFICATION

P. Because of the generative discreteness of C , this choice for |[p]| can be

made unambiguously.
Next, we can choose for every A in L an isomorphism A : AH A of G,
whose inverse is 1
A : A AH . A natural choice for AH : AH AH , and
p : p p in particular, is identity arrows, but this choice is not essential
for the time being. If, however, G is a preorder, then AH must be 1AH .
Then we dene as follows a graph-morphism H from CG to C:
H|[A]| =df AH ,

HX1 ,...,Xk =df 1

N (HX1 ,...,HXk ) HX1 ,...,HXk M (HX1 ,...,HXk ) ,

1
2
i

H(fG1 fG2 ) =df 1

HY1 HY2 (HfG HfG ) HX1 HX2 , for fG : Xi Yi ,

H(gG fG ) =df HgG HfG .

It is clear that for fG : |[A]| |[B]| we have that the type of HfG is AH BH ,
that is H|[A]| H|[B]|. We can prove the following lemma.
Lemma 1. If (IVC) and (IVG) hold and G is a preorder, then for every
arrow term f : A B of C, in C/E we have
f .
HHG f = 1
A
B

Proof. We proceed by induction on the length of f . If f is A1 ,...,Ak , then

HHG A1 ,...,Ak = H|[A1 ]|,...,|[Ak ]|

= 1
N (A1 H ,...,Ak H ) A1 H ,...,Ak H M (A1H ,...,Ak H )

1
1

= 1
N (A1 H ,...,Ak H ) N (A1 , . . . , Ak ) A1 ,...,Ak

M (A1 , . . . , Ak ) M (A1H ,...,Ak H ) ,

since is an isomorphism and G ows through C/E,

=

1
N (A1 ,...,Ak )

A1 ,...,Ak M (A1 ,...,Ak ) , by the preordering of G.

In the induction step, for f i : Ai B i where i {1, 2}, we have

1
2

1
2
HHG (f 1 f 2 ) = 1
H|[B 1 ]|H|[B 2 ]| (HHG f HHG f ) H|[A ]|H|[A ]|

= 1
B 1 B 2
H

1
1
2
1
2
1
2
(1
B 1 B 2 ) (f f ) (A A ) AH AH ,

by the induction hypothesis and the bifunctoriality of ,

3.2.

Direct strictification

81

1
2

1
2
= 1
B 1 B 2 (f f ) A A , by the preordering of G;
1
g
HHG (g f ) = 1
B B f A , by the induction hypothesis,
C
g f , since is an isomorphism.
= 1
A
C

We can then prove the following lemma.

Lemma 2. If fG = gG in EG , then HfG = HgG in C/E.
Proof. We proceed by induction on the length of the derivation of fG = gG
in EG . If fG = gG is in EG because f = g is in E, then in C/E we have
1
f
1
A = B g A ,
B

and by Lemma 1 we obtain HfG = HgG . If fG = 1|[A]| is a strictifying

equation of EG , then, by Lemma 1 and the preordering of G, we have HfG =
1AH . The induction step, where fG = gG is obtained by (sy ), (tr ) or (co ),
is straightforward.

Lemma 2 guarantees that the graph-morphism H from CG to C induces a

functor from CG /EG to C/E.
2 2

1
Let X,Y
be the arrow 1
of G.
HX HY of G and let be the arrow
Then we have the following theorem.
Direct-Strictification Theorem. If (IVC) and (IVG) hold and G
is a preorder, then C/E is equivalent to the C -strict C/E-category CG /EG
via the strict C-functor HG from C/E to CG /EG and the strong C-functor
H, 22 , from CG /EG to C/E.
Proof. We have HG H|[A]| = HG AH = |[AH ]| = |[A]|, and we also have
HG HfG = HG HHG f
f ), by Lemma 1,
= HG (1
A
B
H f H ,
= HG 1
G
G A
B

= fG , by the strictifying equations.

On the other hand, HHG A = AH . We have in C/E, of which G is a subcategory, the isomorphism A : AH A. That is a natural transformation

82

CHAPTER 3.

STRICTIFICATION

from HHG to the identity functor follows immediately from Lemma 1. So

C/E and CG /EG are equivalent via HG and H.
It is clear that HG is a strict C-functor. It remains to show only that
H, 22 , is a strong C-functor. That ( 22 nat) holds is built into the
denition of H(fG1 fG2 ).
To prove (L) for H, 22 , , note rst that HM (Z1 , . . . , Zk ) =
M
(M (HZ1 , . . . , HZk ))H . So Z
= 1
M (HZ1 ,...,HZk ) by the preorder1 ,...,Zk

ing of G. Then (L) follows by applying ( M ) and the facts that

is an isomorphism and that G ows through C/E. Since it is clear that
H, 22 , is a groupoidal C-functor, it follows that it is a strong Cfunctor.

As a consequence of this theorem, we have that f = g in C/E i fG = gG

in CG /EG . So instead of computing in C/E, we can pass to CG /EG , in which
equations between arrow terms are easier to record. By omitting according
to (cat 1) arrow terms equated with identity arrow terms, equations become
shorter. We will avail ourselves of this opportunity provided by the DirectStrictication Theorem in Chapters 5-8.
Consider the following subcategory (C/E)at of (C/E)G , where (C/E)G is
dened as in the preceding section by taking that B is C/E. The objects
of (C/E)at are obtained from all the objects [p] and of (C/E)G by closing
under the operations . The category (C/E)at is the full subcategory of
(C/E)G with these objects.
The category (C/E)at is isomorphic to CG /EG . It is easy to show that
there is a bijection between the objects of (C/E)at and CG /EG . For the
arrows of these two categories we have the following bijections. To every
arrow (f, [A], [B]) of (C/E)at , with f : AF BF an arrow of C/E, we assign
the arrow HG f : |[AF ]| |[BF ]| of CG /EG ; and to every arrow fG : |[A]| |[B]| of
at
CG /EG we assign the arrow (BF HfG 1
AF : AF BF , X, Y ) of (C/E) ,
where X and Y are obtained from |[A]| and |[B]|, respectively, by replacing
and in A and B by G and G , and the brackets ]| and |[ by ] and [.
It seems rather natural to assume that a construction like our construction of CG /EG out of C/E will yield a category equivalent to C/E, and indeed
this assumption may have been made tacitly by Mac Lane in [99] (proof
of Theorem 4.2, p. 39) and [102] (Section XI.1, proof of Theorem 1, p.

3.2.

Direct strictification

83

254), where he establishes coherence for symmetric monoidal categories. A

rather obvious interpretation of his text is that he constructs CG /EG out
of the symmetric monoidal category C/E freely generated by P. A more
explicit assumption of a construction similar to our construction of CG /EG
out of C/E is in [116] (proof of Theorem 6.1, p. 98; no details are given, and
no justication that the construction will yield an equivalent category).
Though the assumption that C/E and CG /EG are equivalent is natural,
this assumption is not warranted without assumptions concerning C/E, like
our assumptions (IV C), (IV G) and the condition that G is a preorder. The
main assumption here is that C/E is freely generated, out of a generating
set P. This set can be conceived as a discrete category, and one may
envisage free generation also out of other categories (which does not dier
signicantly from what we have been doing). Free generation is, however,
essential.
Without our assumptions, or assumptions of the same kind, a construction analogous to our construction of CG /EG out of C/E need not yield an
equivalent category, as it is shown by the following counterexample, which
stems from Isbell (see [100], Section VII.1, p. 160).
Consider the logical system C in L whose primitive arrow terms be

sides identity arrow terms are from the b and w-k families, and consider the
equational system E in C for which we assume all the equations that hold

in the category A (see 4.3; these are equations that hold in monoidal cat

egories), the equations (k 1 nat) and (k 2 nat) (see 2.7) and, for i {1, 2},
the following additional equations (which hold in cartesian categories, and
which we will encounter in 9.1):

(wk)

k iA,A wA = 1A .

We build now a syntactical system S of the C-kind that has a single

object N . The operation on the objects of S satises N N = N . The

1
2
primitive arrow terms of S are 1N , b
N,N,N , b N,N,N , k N,N , k N,N and w N ,
all of type N N . With the help of the equational system E above we
obtain the syntactical category S/E, which is a C/E-category.
We want to show rst that S/E is not a preorder. Consider the skeleton
Card of the category Set of sets with functions such that the set of natural
numbers N is in Card. The category Card is equivalent to Set.

84

CHAPTER 3.

STRICTIFICATION

Consider then the functor S that maps S/E into Card such that SN is
N , while for : N N N being a chosen isomorphism between N and

N N we have S k iN,N = k iN ,N , where k 1N ,N and k 2N ,N are respectively

the rst and second projection in Set, and S wN = 1 wN , where wN is

the diagonal map of Set for which wN (n) = (n, n). The image of S/E in
Card under S is not a preorder, and hence S/E is not a preorder.

For C being the logical subsystem of C with the b arrow terms, let
G = C /E be the C -core of C/E. Let us build out of the equations of E a
category SG /EG , analogously to what we did to obtain CG /EG out of C/E,

which boils down to adding the equations b

N,N,N = b N,N,N = 1N to E to
obtain EG . The unique object of SG /EG may be identied with N . We have
in SG /EG

(f g) h = f (g h), by (b nat) and (cat 1),

(f g) k 1N,N = f k 1N,N , by (k 1 nat),

f g = f, by (wk).
We derive analogously f g = g, starting from h (f g) = (h f ) g

and by using (k 2 nat). So f = g, and SG /EG is a preorder. Hence SG /EG is

not equivalent to S/E. This shows that direct strictication is not always
innocuous.
Note that S/E is not freely generated out of a set P, and does not satisfy
our assumption (IV C), since it is not a logical category. The strictication
of S/E as a C/E-category, in the sense of the Strictication Theorem of
the preceding section, is however allowed. The category (S/E)G is not a
preorder, and it is equivalent to S/E. For (S/E)G we cannot, however, nd
its (S/E)at subcategory.

3.3.

Strictification and diversification

For A and B formulae of a language L, let the type A B be called

balanced when there is a bijection between the occurrences of letters in A
and the occurrences of letters in B that maps the occurrence of a letter
to an occurrence of the same letter. Let C be a logical system in L such
that for each transformation of C the type of the arrow term A1 ,...,Ak :
M (A1 , . . . , Ak ) N (A1 , . . . , Ak ) is balanced. This is guaranteed if and

3.3.

Strictification and diversification

85

are bijections. It is easy to show by induction that the type of every

arrow term of C is balanced.
Let a formula A of L be called diversied when every letter occurs in
A at most once. It is clear that for a balanced type A B we have that
A is diversied i B is diversied. A type A B is called diversied when
A and B are diversied, and an arrow term is diversied when its type is
diversied.
Let the conditions (IVC) and (IVG) of 3.1 be satised. Let E pr be an
equational system that is an extension of the equational system E such that
C /E pr is a preorder and for every equation f = g in E pr that is not in E
we have that the type of f and g is not diversied.
Let (C/E)div be the full subcategory of C/E whose objects are the diversied formulae of L, and let (C/E pr )div be the analogous full subcategory
of C/E pr whose objects are all the diversied formulae of L.
Then it is straightforward to show that the categories (C/E)div and
(C/E pr )div are isomorphic. On objects, this isomorphism is just identity,
and the identity map on the arrow terms of C gives rise to a functor from
(C/E)div to (C/E pr )div and to a functor from (C/E pr )div to (C/E)div . To
show the latter, it is enough to appeal to the fact that if f = g in E pr for
f and g diversied, then f = g in E.
Then we can check that (IVC) and (IV G) hold when C/E is replaced
by C/E pr . Now the C -core G of C/E pr is a preorder. So we can apply
the Direct-Strictication Theorem of the preceding section to obtain the
C -strict C/E pr -category CG /EG pr equivalent to C/E pr .
So, for diversied arrow terms f and g of C of the same type, we have
f = g in C/E i f = g in (C/E)div , i f = g in (C/E pr )div , i f = g in
C/E pr , i fG = gG in CG /EG pr . And it is easier to compute in CG /EG pr , as
explained after the Direct-Strictication Theorem. We will take advantage
of that in 7.6-8 and 8.4.

Chapter 4
Associative Categories
In this chapter we scrutinize Mac Lanes proof of coherence, in the sense
of preordering, for monoidal categories (see [99] and [100], Section VII.2),
and present a dierently organized proof, making ner distinctions. We
separate from this proof a proof of coherence for categories like monoidal
categories that lack the unit object, but, in this respect, we do not dier
from Mac Lane, who did the same in [99]. Throughout the book, it will be
our policy to proceed in this manner, by separating coherence results with
and without special objects such as unit objects. Besides obtaining sharper
results in situations where we have coherence both with and without the
special objects, this policy allows us to obtain coherence without the special
objects in cases where adding the special objects causes diculties.
We give a new proof of coherence for subcategories of monoidal categories where associativity arrows are not isomorphisms. Associativity goes
just in one direction, and for coherence we need just naturality and Mac
Lanes pentagonal coherence condition. The proof of Mac Lanes monoidal
coherence may be built on this more basic coherence result, but it has
also a shorter proof, such as Mac Lanes. Associativity that is not an
isomorphism is interesting because of its relationship with dissociativity investigated from Chapter 7 on. Dissociativity is an associativity principle
involving two operations, which is not an isomorphism.
We also explain in this chapter the eect of strictifying the monoidal structure of a category (which may have extra structure besides this
monoidal structure), in accordance with the results of the preceding chap87

88

CHAPTER 4.

ASSOCIATIVE CATEGORIES

ter. The methods of this chapter are based, as Mac Lanes, on conuence
techniques, like those that may be found in the lambda calculus.

4.1.

The logical categories K

For C a logical system in L, let ECnat be the least set of equations we must
have in every equational system in C to make C/ECnat a natural logical
category (see 2.7). So ECnat has as axioms (re ) (see 2.3), the categorial
equations, the bifunctorial equations for every of L and the naturality
equations for every of C.
We will consider in this work a number of natural logical categories
K. Every such K will be C(K)/E(K) for a logical system C(K) and an
equational system E(K). To determine C(K) we will have to specify only
the language L of C(K) and the transformations of C(K). To determine
nat
E(K) it is enough to specify what equations besides those in EC(K)
have to be
assumed as axioms. We call these equations specic equations. We always
take for granted closure of the arrow terms of C(K) under the operations
and composition, the presence in E(K) of (re ) and of bifunctorial and
naturality equations, and the closure of E(K) under (sy ), (tr ), (co ) and
(su ) of 2.3 and 2.7.
We will rst deal with a number of categories K in the language L .
These will make a hierarchy by having transformations included in subfamilies involving of the families specied below. The label K in such
cases is as in the following table, sometimes with additional indices:
K

I
1

A
1, b

S
1, b, c

L
1, b, c, w-k

When we come to natural logical categories K in L, , we take the whole

families mentioned. The label I is derived from identity, A from associativity, S from symmetry, and L from lattice.
If K is one of our logical categories in a language L without and ,
then K , K or K, will be obtained from K by adding to L either or
, or both, and by adding to C(K) transformations included in -, as
appropriate. To obtain E(K ), E(K ) or E(K, ), we enlarge E(K). Our

4.2.

Coherence of semiassociative categories

89

categories K where C(K) has transformations included in the family d will

be named by prexing D to I, A, S and L; in those where C(K) has m we
prex M, which comes from mix, and where C(K) has m1 we prex Z,
which comes from zero. The categories K we will deal with are presented
in the List of Categories at the end of the book.
It is easy to see that if we have proved coherence for the category K
generated by an innite set of letters P, then we have proved coherence also
for K generated by any set of letters P. (If not more than n 0 dierent
letters occur in f = g, then, by substituting, every derivation of f = g can
be transformed into one in which not more than n dierent letters occur.)
So we assume, when this is needed to prove coherence, that K is generated
by an innite set of letters P.

Our rst logical category K will be called I. The logical system C(I)
is in the language L and its only transformation is 1. The equational

system E(I) is just ECnat for C being C(I), with no additional equations. (The
naturality equations follow here from categorial equations.) It is trivial to

show that I is discrete, and hence a preorder. So I is coherent (see the end
of Chapter 2).

A logical category closely related to I is I, where C(I) is in the language

L, , and where the only transformation is again 1. The equational

system E(I) is just E nat

C(I) , and, as trivially as for I, we show that I is discrete.
So I is coherent too.

4.2.

Coherence of semiassociative categories

For the label K of the preceding section being A , let the logical system

C(A ) be in L with the transformations being 1 and b . The specic

equations of E(A ) are the instances of

(b 5)

b
AB,C,D

b
A,B,CD = ( b A,B,C 1D ) b A,BC,D (1A b B,C,D ).

This is Mac Lanes pentagonal equation of [99] (Section 3; see also [100],
Section VII.1).

We call natural A -categories (in the sense of 2.8) semiassociative

categories. These categories where envisaged explicitly in [93] and implicitly
in [8] and [34]. Semiassociative categories dier from Mac Lanes monoidal

90

CHAPTER 4.

ASSOCIATIVE CATEGORIES

categories by lacking b and . We are now going to prove coherence for

the category A . This coherence was proved in dierent manners, which

we nd more complicated, in [93] and [34], and it can be deduced from a
very general theorem of [8] (Theorem 5.2.4), whose proof is only sketched
in that paper, with substantial parts missing.
For every formula A of L and for x and y two dierent occurrences of
in A we write xRA y when BxC is a subformula of A such that y occurs
in C. If A and B are formulae of L that may dier only with respect to
parentheses, then A and B are comparable (see 2.1), and we may take that
RA and RB are relations between the same sets of occurrences of , and
compare these relations. Formally, we could proceed as follows. Let w(A)
be, as in 2.1, the word obtained from A by deleting all parentheses. Then
RA gives rise to a relation RwA between occurrences of in w(A) such that
we have x RwA y i x and y are occurrences of in w(A) corresponding
respectively to the occurrences x and y of in A and xRA y. Then we do
not compare RA and RB , but RwA and RwB . However, to switch all the
time from RA to RwA and back would be tedious, and we will not mention

RwA . It is easy to see that for every arrow term f : A B of C(A ) the
formulae A and B are comparable (namely, w(A) and w(B) are the same
word), and RB RA (which means, ocially, RwB RwA ). Moreover, if

b occurs in f , then RB is a proper subset of RA ; otherwise, RA = RB .

Since RA and RB are conceived as dened on the same sets when A and
B are comparable, we may denote with the same symbol x the occurrences
of or of a letter in A and B that are at the same place (for the notion of
place see 2.1). We proceed analogously in other similar cases in the future
(cf. 7.1, 7.3, 7.5 and 8.3). We can prove the following.
Extraction Lemma. Let A be a formula of L with a subword A1 (m q,
where (m stands for a sequence of m 0 left parentheses. Then there is

an arrow term g : A C of C(A ) such that in C we have as a subword

A1 q at the place where A has A1 (m q. In addition,
()

if x is not the occurrence of in the two subwords above, then

xRA y implies xRC y;

()

the rst index of every b in g is A1 .

4.2.

Coherence of semiassociative categories

91

Proof. We proceed by induction on m. If m = 0, then g is 1A : A A.

Suppose now m > 0. Then in A we have a subword of the form A1 (A2 A3 )

where A2 is either q or beginning with (m1 q. Then there is a b -term

h : A A , whose head is b
A1 ,A2 ,A3 . In A we have (A1 A2 ) A3 at the
place where A has A1 (A2 A3 ), and hence in A we have A1 (m1 q
at the place where A has A1 (m q. We apply the induction hypothesis to

A , and obtain an arrow term g : A C of C(A ) such that in C we have

A1 q at the place where A has A1 (m q, and if x is not our occurrence of
, then xRA y implies xRC y.
Suppose now x is not our occurrence of , and suppose xRA y. Then we
can conclude that xRA y, and we take that g is g h : A C. From this
the lemma follows.

Theoremhood Proposition. There is an arrow term f : A B of

C(A ) i A and B are comparable formulae of L and RB RA .

Proof. The direction from left to right is easy, as we noted above. For the
other direction, we proceed by induction on the letter length n 1 of A,
which is equal to the letter length of B, because A and B are comparable.
If n = 1, then RA = RB = , and f is 1p : p p.
Let n > 1. So in B there must be a subword of the form (pq). Then we
show that RB RA implies that A must have at the same place a subword
p (m q for m 0. Otherwise, A would have a subword p)n ) (m q with
n 0 at the place where B has (l (p q) with l 0. From that it would
follow that RB RA , since for x being the left in (l (p q) and y the
right we have xRB y, but we do not have xRA y.
Then, by the Extraction Lemma, there is an arrow term g : A C of

C(A ) such that in C we have as a subword (p q) at the place where B

has (p q), and, by (), if x is not the occurrence of in (p q), then xRA y
implies xRC y. We replace (p q) in B and C by r, and obtain respectively
B r and C r . We have RB r RC r , and, by the induction hypothesis, there

is an arrow term f : C r B r of C(A ). Let f : C B be obtained

from this arrow term by putting back (p q) at the place of r. Then f is
f g : A B.

It is clear that there is an arrow term of a given type A B in a logical

92

CHAPTER 4.

ASSOCIATIVE CATEGORIES

system C i there is an arrow of type A B in C/E. So, in the terminology

of 1.1, with the Theoremhood Proposition we obtain a solution to the

theoremhood problem for the category A . The questions whether A and

B are comparable and whether RB RA are clearly decidable. This is
not a very dicult theoremhood problem, and we deal with it not so much
because of its intrinsic interest, but because we need it for the proof of
Semiassociative Coherence below. Besides that, it is a good introduction to
our analogous treatment of other theoremhood problems in 7.1, 7.3-5 and
8.3, some of which are less trivial. As here, analogues of the Theoremhood
Proposition will be applied in establishing coherence results.
Note that the existential quantier in the Theoremhood Proposition, as
well as the existential quantier in the Extraction Lemma, is constructive;
namely, when the conditions are satised, we can actually construct the
arrow term of the required type. This applies also to latter versions of the
Extraction Lemma and of the Theoremhood Proposition.
Let d(A) be the cardinality of the set of ordered pairs RA . If f : A B

of A is not equal to 1A : A A, then RB is a proper subset of RA and

d(B) < d(A). We can then prove the following result of [93].

Semiassociative Coherence. The category A

is a preorder.

Proof. Let f, g : A B be arrow terms of C(A ). We proceed by

induction on d(A)d(B) to show that f = g in A . (Until the end of this

proof, we assume that equality between arrow terms is equality in A .) If

d(A) = d(B), then we conclude that A and B are the same formula, and
f = g = 1A .
Suppose d(B) < d(A). By the Development Lemma (see 2.7) we have

that f = f2 f1 and g = g2 g1 for some b -terms f1 : A C and g1 : A D,

and some arrow terms f2 : C B and g2 : D B of C(A ). We have here

d(C), d(D) < d(A). Let the head of f1 be b

E,F,G , and let the head of g1 be

b
H,I,J . The following cases may arise.

(1) The formulae E (F G) and H (I J) have no occurrences of

letters in common within A. Then we use ( 2) of 2.7 to obtain two b terms f2 : C B and g2 : D B such that f2 f1 = g2 g1 . We infer that
RC RD = RB , from which it follows by the Theoremhood Proposition

4.3.

Coherence of associative categories

93

that RB RB . Hence, again by the Theoremhood Proposition, there is

an arrow term h : B B of C(A ). By applying the induction hypothesis,

we obtain that f2 = h f2 and g2 = h g2 , from which f = g follows.
(2) Suppose E(F G) is a subformula of H or I or J in A; or, conversely,
H (I J) is a subformula of E or F or G in A. Then we proceed as in

case (1) by using the equation (b nat).

(3) The subformulae E (F G) and H (I J) coincide in A. Then C
is D and f1 = g1 . We then apply the induction hypothesis to f2 , g2 : C B
and obtain f = g.
(4) The subformula F G is H (I J) or I J is E (F G). Then

we proceed as in case (1) by using the equation (b 5).

The technique used in the proof above is related to the Church-Rosser,

or conuence, property of reductions in the lambda calculus (see [4], Chapter 3). Analogous techniques will be exploited in 7.1, 7.3, 7.5 and 8.3,
where one nds proofs of coherence analogous to our proof of Semiassociative Coherence.
It is not dicult to see that RA = RB implies that A and B are the
same formula of L . Because, if RA = RB , then, by the Theoremhood

Proposition, there is an arrow term f : A B of C(A ), in which b cannot

occur, since RB is not a proper subset of RA . Hence f must stand for an
identity arrow. So there is a bijection between the formulae A of L and
the relations RA , that is RwA . (A relation is not just a set of ordered pairs,
but its domain and codomain must be specied; so w(A) will be mentioned
in specifying RwA .) From Semiassociative Coherence, we can conclude that

A is isomorphic to the category whose objects are the relations RA , and

where an arrow exists between RA and RB when RB RA . Note that this

means that the category A is not just a preorder, but a partial order.

4.3.

Coherence of associative categories

To obtain the natural logical category A, we have that the logical system

C(A) is in L , with the transformations included in 1 and b. The specic

equations of E(A) are those of E(A ) plus

94

CHAPTER 4.

(b b)

b
A,B,C

b
A,B,C = 1A(BC) ,

ASSOCIATIVE CATEGORIES

b
A,B,C

b
A,B,C = 1(AB)C .

Note that it is enough to assume one of the equations (b nat) and

(b nat) to derive the other one with the help of (b b), and (b b) enables us

also to derive an equation analogous to (b 5) involving b . The equations

(b b), together with (b nat) and (b nat), say that b and b are natural
isomorphisms.

We call natural A-categories associative categories. Associative categories are not necessarily monoidal in the sense of [100] (Section VII.1),
because they may lack the unit object (see 4.6). The objects of an associative category that is a partial order make a semigroup.
A formula A of L is said to be in normal form when RA , dened as in
the preceding section, is empty; i.e., when d(A) = 0. Such an A is of the
form (. . . ((p1 p2 ) p3 ) . . . pn ).

We can show that A is a preorder by relying on the proof of Semiassociative Coherence of the preceding section, based on the Theoremhood
Proposition, but we can establish more easily another, weaker, lemma. To

formulate this lemma, we say that the arrow terms of C(A) that are also

arrow terms of C(A ) are -directed. (This terminology will be extended

later to arrow terms other than those of C(A); see 4.6, 6.1 and 14.1.)

Then the following lemma holds in the category A , and hence also in A.
Directedness Lemma. If f, g : A B are -directed arrow terms and B
is in normal form, then f = g.
The proof of this lemma, which is due to Mac Lane (see [99], Section 3),
is a simplication of our proof of Semiassociative Coherence in the preceding
section. The simplication consists in not having to refer to the full force
of the Theoremhood Proposition, but only to a trivial case of it where RB
is empty.
Then we prove the following result of [99] (Section 3).

Associative Coherence. The category A is a preorder.

Proof. For f : A B an arrow term of C(A), there are two -directed

4.3.

Coherence of associative categories

95

arrow terms g : A C and h : B C such that C is in normal form

(these arrow terms are not uniquely determined). By the Development
Lemma (see 2.7), the arrow term f is equal to a developed arrow term
fn . . . f1 . We proceed by induction on n to show that f = h1 g, where
h1 is obtained from the arrow term h by inverting order in composition,

and by replacing b by b and vice versa.

If n = 1, then, since f1 is 1A , by the Directedness Lemma we have
g = h, from which f = h1 g follows.
For n > 1 and fn : B B, we have by the induction hypothesis that
fn1 . . . f1 : A B is equal to (h )1 g for g : A C and h : B C.

If fn is a b -term, then, for h : B C, by the Directedness Lemma we

have h fn = h , and f = h1 g follows. If fn is a b -term, then by the

Directedness Lemma we have h fn1 = h, and f = h1 g follows again.
For f : A B we obtain in the same manner f = h1 g, and so
f = f .

One might suppose that Semiassociative Coherence can be inferred directly from Associative Coherence. This would be so if we could nd an

independent proof that A is isomorphic to a subcategory of A, a proof

that would not rely on Semiassociative Coherence. In fact, we use Semi

associative Coherence to conclude that A is isomorphic to a subcategory

of A. That A is isomorphic to a subcategory of A amounts to showing

that for f and g arrow terms of C(A ) we have f = g in A i f = g in

A. That f = g in A implies f = g in A is clear without appealing to coherence, but for the converse implication we use Semiassociative Coherence
(cf. 14.4).
In the proof of Associative Coherence above, we rely essentially on the

normal form of formulae, and use both b -terms and b -terms. This
is why for the proof of Semiassociative Coherence we could not rely on
the Directedness Lemma, but we needed the Theoremhood Proposition of
the preceding section. The proof of Semiassociative Coherence is not very
dicult, but it is more dicult than the proof of Associative Coherence
based on the Directedness Lemma. The proof of Associative Coherence
can be based on Semiassociative Coherence, but it has also this simpler
proof.

96

4.4.

CHAPTER 4.

ASSOCIATIVE CATEGORIES

Associative normal form

Once we have proved Associative Coherence, we can ascertain that every

arrow term is equal to an arrow term in a normal form, which we are going
to dene. This normal form is unique, in the sense that arrow terms in

normal form are equal in A (i.e., they stand for the same arrow of A) i
they are the same arrow term.
First we prove the following analogue of the Extraction Lemma of 4.2
(see 2.1 for the notion of scope).
Extraction Lemma. If there is an occurrence z of in a formula A of
L , then there is a formula A1 z A2 of L such that there is an arrow term

g : A A1 z A2 of C(A). In addition,
()

for all occurrences x and y of in Ai , where i {1, 2}, we have

that y is in the scope of x in A i y is in the scope of x in Ai ;

()

every subterm of g of the form b

D,E,F is of the type D (E z F )

(D E) z F , and every subterm of g of the form b

F,E,D is of the
type (F z E) D F z (E D).
Proof. We proceed by induction on the number n 0 of occurrences of
connectives in A. If n = 0, then the antecedent of the lemma is false, and
the lemma is trivially true.
If n > 0, then A is A u A with u an occurrence of . If u is z, then
g is 1A . So suppose u is not z, and suppose z is in A . Then, by the

induction hypothesis, we have an arrow term g : A A1 z A2 of C(A)

satisfying the primed version of (). The arrow term g 1A is of the type

A (A1 z A2 ) u A , and we have the arrow term b

A ,A ,A (g 1A ) : A

A1 z (A2 u A ) of C(A).
To verify (), suppose x and y are two occurrences of in A1 . It is clear
that y is in the scope of x in A i it is in the scope of x in A . So, by ()
of the induction hypothesis, we have that y is in the scope of x in A i it
is in the scope of x in A1 . We settle easily in a similar manner cases where
x and y are both in A2 or A . If x is u, then () follows easily again.
The case where z is in A is settled analogously by using the arrow term

4.4.

Associative normal form

97

b
: A u (A1 z A2 ) (A u A1 ) z A2 . We easily check () by going
A ,A
1 ,A2
over the proof above.

The analogue for A of the Theoremhood Proposition of 4.2 would state

simply that there is an arrow term of C(A) of type A B i the formulae

A and B are comparable.
We do not need the assertion () of the Extraction Lemma of this section
for the proof of the Associative Normal-Form Proposition below. We stated
this assertion, nevertheless, because it is analogous to the assertion () of
the Extraction Lemma of 4.2.
We need some preliminary notions to introduce our normal form. For
every formula A of L we assign to every subformula of A a natural number
n 2 in the following manner. We assign to every occurrence of a letter p
in A a prime number iA (p) 2, each occurrence having a dierent number
from all other occurrences. (Note that this assignment is not unique.) Next,
for a subformula B C of A, we have iA (B C) = iA (B) iA (C). For every
subformula D of A, we dene I(D) as follows:
if D is p, then I(D) is p;
if D is B C, then I(D) is I(B) iA (BC) I(C).
So I(D) is like a formula, but with subscripted occurrences of .
Let A be a formula comparable with B, and let x be an occurrence of
in A. The formula B has an occurrence of at the same place, and we
call that occurrence of also x. Let A be obtained from A by adding to
every x of A the subscript that x has in I(B). This subscripting gives rise
to a formula C with subscripted occurrences of for every subformula C

of A. For every arrow term f : A1 A2 of C(A) such that both A1 and A2

are comparable with B, we have an arrow term f : A1 A2 obtained by
replacing every index C of f by C . Then we have the following proposition.
Associative Normal-Form Proposition. If A and B are comparable

formulae, then there is an arrow term f : A B of C(A) such that every

subterm of f : A I(B) of the form b

D,E,F is of the type D l (E k F )
(D l E) k F where for every n in D l E and F we have that n divides

k; analogously, every subterm of f : A I(B) of the form b

F,E,D is of the

98

CHAPTER 4.

ASSOCIATIVE CATEGORIES

type (F k E) l D F k (E l D) where for every n in F and E l D

we have that n divides k.
Proof. We proceed by induction on the number m of occurrences of
connectives in B. If m = 0, then f is 1p : p p. If m > 1, then B
is of the form B1 z B2 for z an occurrence of , and in I(B) we have zk
with k = l n for l, n 2, where l is any other subscript of in I(B).
Then, by the Extraction Lemma of this section, there is an arrow term

g : A A1 z A2 of C(A) such that () is satised. This guarantees that all

the subterms of g from the family b are as required in the statement of the
proposition. By the induction hypothesis, we have arrow terms f1 : A1 B1
and f2 : A2 B2 that satisfy the conditions of the proposition, and f : A B
is (f1 f2 ) g : A B.

The procedure of the proof of this proposition, which presupposes the

Extraction Lemma of this section, gives rise to a unique arrow term, which
we may consider to be in normal form. We may transform this arrow term
into a developed arrow term by replacing (f1 f2 ) g in the proof above by
(f1 1B2 ) (1A1 f2 ) g, or by (1B1 f2 ) (f1 1A2 ) g, when neither of
f1 and f2 is an identity arrow term.

4.5.

Strictification of associative categories

According to our denition of 3.1, an A-category, and in particular an

associative category, is C(A)-strict when for all objects a, b and c we have

a (b c) = (a b) c,

b
a,b,c = b a,b,c = 1a(bc) .

The category Rel of 2.9 with being + is a C(A)-strict associative category.

For G being A, our construction of B G in 3.1 covers a construction

exposed in [102] (pp. 257) and [72] (pp. 29-30), which builds out of an

associative category B a C(A)-strict associative category B G equivalent to

B via two strong C(A)-functors. As a matter of fact, [102] and [72] are

about monoidal categories, i.e. natural A -categories (see the next section),
with whose strictication we deal in 4.7. The result one can extract from

4.5.

Strictification of associative categories

99

[102] and [72] is that every A-category can be (A, C(A))-strictied (A is

C(A)/E(A)). We have shown something more than that in 3.1. We have

shown, namely, that if B is a C/E-category for a logical category C/E where

C(A) L C and the C(A)-core G of C/E, which ows through C/E and is a

preorder, is A, then B can be (C/E, C(A))-strictied. One passes from our

construction to that of [102] (pp. 257) and [72] (pp. 29-30) by realizing

that for G being A, and A and B formulae of G(B) we have A G B i

after deleting every G in A and B we obtain the same nite nonempty
sequence of objects of B. So there is a one-to-one correspondence between
the classes [A] with respect to G and nite nonempty sequences of objects
of B. We are passing from the free groupoid generated by the objects of B
to the free semigroup generated by these objects. The objects of the free
semigroup may be represented by nonempty words.

When G is A, and [A] corresponds to the sequence a1 . . . an of objects

of B in the sense just specied, we can take as the representative AF of [A]
when n 2 the formula (. . . (a1 G a2 ) . . . G an ) (where parentheses are
associated to the left). Then, instead of choosing the arrows A : AF A
and 1
A : A AF of G(B) arbitrarily, as we did in 3.1, we can dene
them inductively in the following manner. First, we dene by induction
the arrows AF G BF : (A G B)F AF G BF and 1
: AF G BF
AF G BF
G
G
G
(A B)F (note that (A B)F = (AF BF )F ):
AF G b = 1
= 1AF G b , for b an object of B,
AF G b

AF G (CF G b) = b
AF ,CF ,b (AF G CF 1b ),

1
= (1
G 1b ) b
AF ,CF ,b .
AF G (CF G b)
AF G CF
We have no need for other and 1 arrows except these to dene 22
2 2
and 1
(see 3.1), but, for the sake of completeness, we can dene
inductively as follows A and 1
A for every object A of G(B):
a = 1
a = 1a , for a an object of B,
AG B = (A G B ) AF G BF ,
1 G 1
(
1
= 1
A B ).
AG B
AF G BF
To check the correctness of these denitions it is enough to verify that
1
AF = 1
AF = 1AF . Note that A and A are dened with arrow terms

100

CHAPTER 4.

ASSOCIATIVE CATEGORIES

that, after deleting identity arrow terms, are in the associative normal form
of the preceding section.
In Lemmata 1-5 of 3.1 we did not appeal to the preordering of G. In
Lemma 6, we had this assumption. We can prove, however, the following

corollary of this lemma without appealing to the preordering of A, i.e. to

its coherence.

Lemma. For G being A, the category A , M , is C(A)-strict.

Proof. For X and Y objects of A , let us write XY for X Y , since

G
in A corresponds to concatenation of sequences of formulae of L . We
have, of course, X (Y Z) = (X Y ) Z = XY Z.

Let [A],[B] stand for EAF G BF : E(A G B)F E(AF G BF ) (see 3.1
for the functor E). Since E(A G B)F is F ([A][B]) and E(AF G BF ) is

1
F [A] F [B], the type of X,Y is F (XY ) F X F Y . If [A],[B] stands for
1

E1
, then X,Y : F X F Y F (XY ) is the inverse of X,Y in A.
AF G BF

To show that for being b we have X,Y,Z

= 1XY Z in A we proceed
as follows. We make an induction on the length of the sequence corresponding to Z, and we use the alternative denition of from 3.1.

If Z is a, then by the denition of U,V and the fact that it is an

isomorphism, we have
1

XY,a (X,Y 1a ) b
F X,F Y,a (1F X Y,a ) X,Y a
1

= (X,Y 1a ) b
F X,F Y,a X,Y a

= 1F (XY a) .
If Z is U a, then we have
1

XY,U a (X,Y 1F (U a) ) b
F X,F Y,F (U a) (1F X Y,U a ) X,Y U a
1

= (XY,U 1a ) ((X,Y 1F U ) 1a ) b
F XF Y,F U,a b F X,F Y,F U a

(1F X b
F Y,F U,a ) b F X,F YF U,a ((1F X Y,U ) 1a ) (X,Y U 1a ),

by denition, (b b) and naturality equations,

1

= (XY,U 1a ) ((X,Y 1F U ) 1a ) (b
F X,F Y,F U 1a )

((1F X Y,U ) 1a ) (X,Y U 1a ), by (b b) and (b 5),

= 1F (XY U a) , by bifunctoriality and the induction hypothesis.

4.6.

Coherence of monoidal categories

101

We proceed analogously when is b .

So A can be (A, C(A))-strictied, and by the Strictication-Coherence

Implication of 3.1, we can conclude that A is a preorder.

4.6.

Coherence of monoidal categories

To obtain the natural logical category A , we have that the logical system

C(A ) is in L, , with the transformations included in 1, b and -.

The specic equations of E(A ) are those of E(A) plus

( )

( )

(b )

A = 1A ,

= 1 ,

A
A
A

= 1

A ,
A
A

A = 1A ,

b
A,,C = ( A 1C ) (1A C ).

From these equations one infers

(b )

(b )

( )

b
A,B, = AB (1A B ),

b
,B,C = ( B 1C ) BC ,

(see [75], Theorems 6 and 7, and 9.1 below).

The specic equations of E(A) are introduced, as all our equations for
logical categories, by axiomatic equations with variables. These equations

with variables are now assumed for E(A ) (see 2.3 and 2.7). The equa
tional system E(A ) will be closed under (su ) for formulae C of L, ,
and not only of L . We assume tacitly from now on that we proceed in
an analogous manner whenever we pass from an equational system formulated originally with respect to a poorer language to an equational system
formulated with respect to a richer language.

The equations ( ) and ( ) above, together with the naturality equa

tions for arrow terms in the family -, say that in subfamilies of this
family we nd natural isomorphisms.

The natural A -categories are commonly called monoidal categories

(see [100], Section VII.1), and sometimes tensor categories (as in [71] and

102

CHAPTER 4.

ASSOCIATIVE CATEGORIES

[72]; in [7] they are called categories with multiplication). The objects of a
monoidal category that is a partial order make a monoid.
For every formula A of L, let k(A) be the number of occurrences of
in A as main conjuncts in subformulae of A (i.e. the visible occurrences
of in subformulae of A of the form B or B). So in L, we do
not count only if A itself is , but our denition of k(A) is adapted to
other languages L too (cf. the end of 6.1). We say that A is in normal
form when d(A) = 0 and k(A) = 0 (where d(A) is dened in 4.2). So, for
example, (p q) r and are in normal form.

An arrow term of C(A ) is called -directed when neither of b ,

and occurs in it. (This denition extends the denition of -directed

arrow terms of C(A) in 4.3.) Then, by extending the proof of the Directedness Lemma of 4.3, which is a simplication of our proof of Semiassociative
Coherence of 4.2, we can prove the following.
Directedness Lemma. If f, g : A B are -directed arrow terms of

C(A ) and B is in normal form, then f = g in A .

Proof. We proceed by induction on d(A)+k(A). In the induction step
we have the following new cases for f = f2 f1 and g = g2 g1 for some
-directed arrow terms f2 : C B and g2 : D B:

(I)

f1 : A C is a b -term and g1 : A D is a -term,

(II)

f1 : A C is a b -term and g1 : A D is a -term,

(III)

f1 : A C and g1 : A D are -terms,

(IV)

f1 : A C and g1 : A D are -terms,

(V)

f1 : A C is a -term and g1 : A D is a -term,

With (I), the only interesting additional case is when the head of f1 is

b
E,F, and the head of g1 is F , where we apply ( b ).
With (II), the only interesting additional cases are when the head of f1

is b
E,,G and the head of g1 is G , where we apply ( b ), and when the

head of f1 is b
,F,G and the head of g1 is F G , where we apply ( b ).
In the other, uninteresting cases, of (I) and (II), which are in principle
covered by what we had in the proof of Semiassociative Coherence in 4.2,
we apply bifunctorial and naturality equations.

4.7.

Strictification of monoidal categories

103

We apply these equations also in cases (III), (IV) and (V); for the last

case we also need the equation ( ). The remaining cases are as in 4.2.
From this Directedness Lemma we infer the following result of [99] and
[100] (Section VII.2), whose proof is analogous to the proof of Associative
Coherence in 4.3.

Monoidal Coherence. The category A is a preorder.

4.7.

Strictification of monoidal categories

In a C(A )-strict A -category for every object a we have

a = a = a,

a = a = a = a = 1a ,

in addition to what was mentioned at the very beginning of 4.5. The cat
egory Rel of 2.9 with being + and being 0 is a C(A )-strict monoidal
category.
What we have said at the beginning of 4.5 concerning the strictication of associative categories and previous results of [102] and [72], applies
mutatis mutandis to the present context. One has to replace associative

by monoidal and A by A .

For G being A , the objects [A] of B G correspond bijectively now to

arbitrary nite sequences of objects of B, including the empty sequence.
The class [G ] corresponds to the empty sequence, and we can take G as
the representative AF of [G ].
The inductive denition of A of 4.5 can now be extended with the
following clauses:
= 1 G ,
G = 1
G

1
=
AF ,
AF G G

1
=
AF .
G G AF

AF G G =
AF ,
G G AF =
AF ,

The correctness of these denitions for the case G G G and 1

is
G G G

guaranteed by ( ).

104

CHAPTER 4.

ASSOCIATIVE CATEGORIES

We can prove as before the following analogue of the Lemma of 4.5,

without presupposing the preordering of A .

Lemma. For G being A , the category A , M , is C(A )-strict.

Proof. We proceed as for the proof of the Lemma of 4.5, with the
following additions.

Let us write instead of [G ]. To show that for being b the equation

X,Y,Z
= 1XY Z holds in A , we have to consider new cases when X, Y or
Z are .
I.

If X is , then
1

Y,Z (,Y 1F Z ) b
,F Y,F Z (1 Y,Z ) ,Y Z
1

= Y,Z (
F Y 1F Z ) b ,F Y,F Z (1 Y,Z ) F (Y Z)

= 1F (Y Z) , by (b ), naturality and isomorphisms.

II.

If Y is , then
1

X,Z (X, 1F Z ) b
F X,,F Z (1F X ,Z ) X,Z
1

= X,Z (
F X 1F Z ) b F X,,F Z (1F X F Z ) X,Z

= 1F (XZ) , by (b ) and isomorphisms.

III.

If Z is , then
1

XY, (X,Y 1 ) b
F X,F Y, (1F X Y, ) X,Y
1

=
F (XY ) (X,Y 1 ) b F X,F Y, (1F X F Y ) X,Y

= 1F (XZ) , by (b ), naturality and isomorphisms.

To show that for being the equation X

= 1X holds in A , we

have
FX

X, = 1F X , since is an isomorphism, and analogously for

So A can be (A , C(A ))-strictied, and we can conclude that A is

a preorder by the Strictication-Coherence Implication of 3.1.
In the presence of the unit object , when we deal with monoidal cate
gories, there is a C(A )-strict monoidal category alternative to B G , inspired

4.7.

Strictification of monoidal categories

105

by Cayleys representation of monoids (see [72], pp. 26-27, and [102], p. 260,
Exercises 1-3). This is a functor category in which on objects is composition of functors. It is not clear how to adapt this functor category to
cases where we have two monoidal structures, while B G covers that and
much more. Moreover, the proof of Proposition 1.3 of [72] (p. 27), which
states the faithfulness of a functor into the functor category, seems to rely
essentially on the presence of the unit object . On the other hand, the approach through the category B G is, of course, possible in situations without
unit objects.

The rst proof proposed for the (A , C(A ))-strictication of monoidal

categories in [72], viz. the proof of Corollary 1.4 on p. 27, does not stand,
since the full image of the functor L in the functor category is not closed
under composition of functors. The other proof of Corollary 1.4 in [72],
on p. 30, is closer to what we have been doing in this section and in 4.5.
(In the presentation of [102], pp. 255, which is more accessible than that
in [72], there is a lapsus on p. 259; one should have there 12G2 and not
G2 21.)

Chapter 5
Symmetric Associative
Categories
We present in this chapter a proof of coherence, with and without unit
objects, for symmetric monoidal categoriesa proof more thorough than
Mac Lanes proof (see [99] and [102], Section XI.1), from which it stems.
We provide with it a proof of the completeness of the usual axiomatization
of symmetric groups via the normal form that stems from Burnside. We
also make explicit the strictication of the monoidal structure involved in
the proof, on which Mac Lane presumably also relies (as we noted in 3.2).
Mac Lane seems to presuppose that this strictication is allowed, while we
justify it by the results of Chapter 3.

5.1.

Coherence of symmetric associative categories

To obtain the natural logical category S, we have that the logical system

C(S) is in L , with the transformations included in 1, b and c. The

specic equations of E(S) are those of E(A) plus

( c c)

(b c)

c B,A

c A,B = 1AB ,

c A,BC = b
B,C,A (1B c A,C ) b B,A,C ( c A,B 1C ) b A,B,C .

The equation ( c c), together with ( c nat), says that c is a natural isomor

phism, while the equation (b c) amounts to Mac Lanes hexagonal equation

of [99] (Section 4; see also [100], Section VII.7, and [102], Section XI.1).
107

108

CHAPTER 5.

SYMMETRIC ASSOCIATIVE CATEGORIES

An alternative way to obtain S is to extend C(A ) with the transfor

mation c, and assume the following denition:

b
A,B,C =df c BC,A b B,C,A

c CA,B

b
C,A,B

c AB,C ,

together with the equations E(S) of S. In that context the equations (b b)

become

c BC,A b
B,C,A

b
A,B,C

c CA,B

c BC,A b
B,C,A

b
C,A,B

c CA,B

c AB,C

b
C,A,B

b
A,B,C = 1A(BC) ,

c AB,C = 1(AB)C ,

while the equation (b c) amounts to

b
C,A,B

c AB,C

b
A,B,C = ( c A,C 1B ) b A,C,B (1A c B,C ).

We call natural S-categories symmetric associative categories. Symmetric associative categories dier from Mac Lanes symmetric monoidal
categories, which we will consider in 5.3, by not necessarily having the
unit object . The objects of a symmetric associative category that is a
partial order make a commutative semigroup.
It is easy to check that the two maps G of 2.9, dened on objects and

on arrows, give rise to a strict C(S)-functor from S to Rel. For that, it is

enough to check that for f = g being one of the equations ( c nat), ( c c)

and (b c) we have Gf = Gg in Rel (for the remaining equations in the

axiomatization of S this is trivial). It is clear that Gf corresponds to a

permutation of a nite nonempty domain. Our goal is to prove coherence

for the category S with respect to Rel; namely, we will prove the following
result of [99] (Section 4).

Symmetric Associative Coherence. The functor G from S to Rel is

faithful.

Coherence here does not mean, as for A , A and A , that S is a preorder.

We do not have G c p,p = G1pp , and hence, by the functoriality of G, we

do not have c p,p = 1pp in S.

For G being A and C/E being S, we have that the conditions (IV C)
and (IV G) of 3.1 are satised, and G is moreover a preorder. To verify

5.1. Coherence of symmetric associative categories

109

that A is generatively discrete, we appeal to the fact that for every arrow

term f : A B of C(A) we have that Gf is a bijection whose ordered

pairs correspond to occurrences of the same letter in A and B. We have
analogous arguments to establish generative discreteness in other cases of
strictication, which we will encounter later, and we will not dwell on this
matter any more.
Then we can apply the Direct-Strictication Theorem of 3.2 to obtain

a category CG /EG , which we will call Sst . We call CG here C(Sst ). The

category Sst is equivalent to S via the strict C(S)-functor HG from S to Sst ,

and the strong C(S)-functor H, 22 from Sst to S.

Consider the composite functor GH from Sst to Rel. It is easy to see

that GH|[A]| = GA, since all the formulae in |[A]| have the same letter
length, and we also have
GH|[A1 ]|,...,|[Ak ]| = GA1 ,...,Ak .
We can conclude that G is equal to the composite functor GHHG . Hence
it is enough to establish that GH is faithful to conclude that G is faithful,
because we know that HG is faithful.

Note that since in Sst the equation (b c) becomes

c X,Y Z = (1Y c X,Z ) ( c X,Y 1Z ),

and since we also have

c XY,Z = ( c X,Z 1Y ) (1X c Y,Z ),

every arrow term of C(Sst ) will be equal to a developed arrow term in which

every c-term is of one of the following forms:

c |[p]|,|[q]| ,

c |[p]|,|[q]| 1X ,

1X c |[p]|,|[q]| ,

(1X c |[p]|,|[q]| ) 1Y .

The c-terms of C(Sst ) and their heads are dened analogously to what we
had in 2.6.
For the rst two arrow terms in this list we use the abbreviation s1 , and
for the third and fourth we use si+1 , where i = GHX. So our developed
arrow term may be written in the form si1 . . . sin 1X , where n 0.

110

CHAPTER 5.

SYMMETRIC ASSOCIATIVE CATEGORIES

It is easy to check that in Sst we have the equations

(s1)
(s2)
(s3)

si si = 1,
si+k si = si si+k , for k 2,
si si+1 si = si+1 si si+1 ,

where 1 stands for 1X for some X. The equation (s3) is derived with the

help of (b c) and ( c nat) (see [102], Section XI.1, p. 254).

It is well known that the equations (s1), (s2) and (s3), together with the
equations corresponding to the categorial equations (cat 1) and (cat 2)
namely, the equations of monoidsaxiomatize symmetric groups (i.e., give
a presentation of these groups by generators and relations; see [26], Section
6.2). A reader with this knowledge may now conclude that the functor GH
is faithful. However, to make the matter self-contained, we will justify this
conclusion in the next section.

5.2.

The faithfulness of GH

Let s[i,j] be an abbreviation for si si1 . . . sj+1 sj if i > j, while s[i,i]

stands for si . For n 0, and 1 standing for 1X for some X, we say that
s[i1 ,j1 ] . . . s[in ,jn ] 1
is in normal form when i1 < i2 < . . . < in (this normal form is implicit in
[17], Note C, pp. 464-465).
Then from (cat 1), (cat 2), (s1), (s2) and (s3) we can prove the following
equations for i k:
s[i,j] s[k,l] = s[k,l] s[i,j] ,
= s[i,l] ,
= s[k1,l] s[i,j+1] ,
= s[i,j+1] ,
= s[k1,l] ,
= 1,
= s[k1,l] s[i,j+1] ,
= s[k1,l1] s[i,j] ,

if
if
if
if
if
if
if
if

k+1 < j,
k+1 = j,
k = j, i > j
k = j, i > j
k = j, i = j
k = j, i = j
k > j l,
k > j < l.

and
and
and
and

k
k
k
k

> l,
= l,
> l,
= l,

5.2. The faithfulness of GH

111

(Note that (s1) is the sixth equation, (s2) is an instance of the rst equation,
and (s3) is an instance of the last equation.) From these equations we easily
infer the following.

Normal-Form Lemma. Every arrow term of C(Sst ) is equal in Sst to an

arrow term in normal form.
We can also prove the following.

Uniqueness Lemma. If the arrow terms f, g : X Y of C(Sst ) are in

normal form and GHf = GHg, then f and g are the same arrow term.
Proof. Let f and g be s[i1 ,j1 ] . . . s[in ,jn ] 1 and s[k1 ,l1 ] . . . s[km ,lm ] 1
respectively. Note that GHs[i,j] corresponds to the following diagram:

j
i1
i

aa

a




aaa


a



a

a





a
j 1
i2
i1
i

j 2 j 1

j 2

i+1

GHX 1

i+1

GHX 1

So, for n > 0, we have in GHf the ordered pair (jn1, in ), with jn1 < in ,
which we call the last falling slope of GHf . Note that for l > in we have
in GHf the ordered pairs (l, l).
Then we proceed by induction on n. If n = 0, then m = 0; otherwise,
in GHf we would have only the ordered pairs (i, i), and in GHg we would
have (lm 1, km ) for lm 1 < km .
If n > 0, then, as we have just shown, m > 0, while in = km and
jn = lm ; otherwise, the last falling slopes of GHf and GHg would dier.
Since for e being sjn sjn +1 . . . sin 1 sin we have
GHf GHe = GHg GHe,
we can conclude that for f and g being s[i1 ,j1 ] . . . s[in1 ,jn1 ] 1 and
s[k1 ,l1 ] . . . s[km1 ,lm1 ] 1 respectively we have GHf = GHg , and we
have by the induction hypothesis that f and g are the same arrow term.
Hence f and g are the same arrow term.

112

CHAPTER 5.

SYMMETRIC ASSOCIATIVE CATEGORIES

As a matter of fact, it would be enough to prove instead of the Uniqueness Lemma that if f is in normal form and GHf = GH1, then f is 1.
(Altogether, this proof would not be shorter than the proof of the Uniqueness Lemma.)
From the Normal-Form Lemma, the functoriality of GH and the Uniqueness Lemma we infer easily that GH is faithful. An alternative proof of
this faithfulness is obtained without the Uniqueness Lemma. Instead we
establish that the number of dierent arrow terms f : X Y in normal
form is n! for n = GHX = GHY , and that for every permutation of
an ordinal n > 0 (this permutation is an arrow of Rel ) there is an arrow

f : X Y of Sst such that GHX = GHY = n and GHf = . Then

we use the fact that every onto function from n! to n! is also one-one. A
proof in this alternative style is suggested by [17] (Note C, pp. 464-465).
Our proof in this section is easily converted into a proof of completeness of
the standard axiomatization of symmetric groups with respect to groups of
permutationsa proof alternative to the proof in [17], mentioned above.

5.3.

Coherence of symmetric monoidal categories

To obtain the natural logical category S , we have that the logical system

C(S ) in L, , with the transformations included in 1, b, c and -.

The specic equations of E(S ) are obtained by taking the union of those

of E(S) and E(A ).

One can derive for S the following equation:

c A, =
A

( c )

(see [75], Theorem 8). This equation says that one of and is superuous:

it can be dened in terms of the other with the help of c. Note that in the

presence of ( c ), the instance of (b c) where A is , namely

c ,BC = b
B,C, (1B c ,C ) b B,,C ( c ,B 1C ) b ,B,C ,

is derivable without using (b c) (we apply (b ), (b ) and (b )), and anal

ogously for the instances of (b c) where B or C is .

5.3. Coherence of symmetric monoidal categories

113

Natural S -categories are usually called symmetric monoidal categories.

The objects of a symmetric monoidal category that is a partial order make
a commutative monoid.
The following result is from [99] (Section 5).

Symmetric Monoidal Coherence. The functor G from S to Rel is

faithful.
The proof of this faithfulness is easily obtained by extending our proof of
Symmetric Associative Coherence in the two preceding sections. To obtain

st
the category Sst
, we take A to be G. Then in S we have c |[A]|,|[]| = 1|[A]| .

Although the categories Sst and Sst

are not preorders, the categories

div
div
S and S are preorders (for the denition of these last two categories
see 3.3). This follows from Symmetric Associative Coherence and Sym

metric Monoidal Coherence. So extending E(Sst ) and E(Sst

) with the equa
tion c A,A = 1AA , which yields preordering (see 6.5), does not add new

div

equations to S

div

and S . We will rely on that in 7.6.

Chapter 6
Biassociative Categories
In this chapter we prove coherence, in the sense of preordering, for categories that have two monoidal structures, with or without unit objects.
We explain what are the eects of strictifying this double monoidal structure. With the help of that, we establish also coherence for categories with
two symmetric monoidal structures. The proofs of the present chapter are
based on the proofs of the preceding two chapters.

6.1.

Coherence of biassociative and bimonoidal categories

Let A be the natural logical category in L isomorphic to A of 4.3. The

only dierence is that is everywhere replaced by . The primitive arrow

terms of C(A) are included in 1 and b, while the equations of E(A) are

obtained by replacing by in those of E(A) (see the List of Equations

and the List of Categories at the end of the book).
To obtain the natural logical category A, we have that the logical system
C(A) is in L, , with the transformations included in the families 1 and
b. The specic equations of E(A) are obtained by taking the union of those

of E(A) and E(A). We call natural A-categories biassociative categories.

An arrow term of C(A) is called -directed when neither of b and b

occurs in it.
We dene inductively as follows formulae of L, in normal form:
every letter is in normal form;
115

116

CHAPTER 6.

BIASSOCIATIVE CATEGORIES

if A and B are in normal form and B is not of the form B1 B2 , then

A B is in normal form;
if A and B are in normal form and B is not of the form B1 B2 , then
A B is in normal form.
So all parentheses within conjunctions and disjunctions are associated to
the left as much as possible.
Then we can easily prove the Directedness Lemma of 4.3 for A by
extending the proof in 4.3. From that we infer as before the following.
Biassociative Coherence. The category A is a preorder.

Let A be the natural logical category in L, isomorphic to A of

4.6. The only dierence is that and are everywhere replaced by and

respectively. The primitive arrow terms of C(A ) are included in 1, b

and -, while the equations of E(A ) are obtained by replacing and

by and respectively in those of E(A ) (see the List of Equations and

the List of Categories).
To obtain the natural logical category A, , we have that the logical
system C(A, ) is in L,,, , with the transformations included in the
families 1, b and -. The specic equations of E(A, ) are obtained by

taking the union of those of E(A ) and E(A ). We call natural A, categories bimonoidal categories.

An arrow term of C(A, ) is called -directed when neither of b , b ,

, , and occurs in it.

We dene inductively as follows formulae of L,,, in normal form:
every letter and the nullary connectives and are in normal form;
if A and B are in normal form, B is not of the form B1 B2 and
neither A nor B is , then A B is in normal form;
if A and B are in normal form, B is not of the form B1 B2 and
neither A nor B is , then A B is in normal form.
So, as for the normal form of formulae of L, , all parentheses within
conjunctions and disjunctions are associated to the left as much as possible,
and, moreover, the conjuncts and disjuncts are deleted.

6.2.

Form sequences

117

Then we can prove the Directedness Lemma of 4.3 for A, by extending the proof in 4.3 and 4.6. From that we infer as before the following.
Bimonoidal Coherence. The category A, is a preorder.

6.2.

Form sequences

The classes [A] involved in (C/E,A)-strictication correspond bijectively to

nite nonempty sequences of objects (see 3.1 and 4.5). With (C/E,A)strictication the classes [A] correspond analogously to a more complicated
notion of sequence where we distinguish concatenation of the kind from
concatenation of the kind. To dene this notion, let X be an arbitrary
set, and let {, }. If is , then c is , and if is , then c is .
We dene inductively as follows the notion of form sequence of X of
colour :
(1) every x X is a form sequence of X of colour ;
(2) if X1 . . . Xn , where n 2, is a sequence of form sequences of colour
c , then the ordered pair (X1 . . . Xn , ) is a form sequence of X of
colour .
(Finite nonempty form sequence would be a more precise, but less concise,
denomination for the notion of form sequence just introduced. We will
introduce below a more general notion of form sequence that covers also
empty form sequences of both colours.)
It is easy to see that every form sequence of X of colour corresponds
to a planar nite tree (see 2.1) with nodes of n-ary branching where n 2,
such that every leaf is labelled by an element of X , every node that is not
a leaf is labelled by or , for every node labelled {, } its successor
is labelled c , provided this successor is not a leaf, and the root is labelled
, provided this root is not a leaf.
We introduce now an alternative notation for form sequences, which is
obtained by writing (X1 X2 . . . Xn ) for (X1 . . . Xn , ) in clause (2) of
the denition above. We call this notation, which we will need in latter
sections, the natural notation for form sequences. The natural notation

118

CHAPTER 6.

BIASSOCIATIVE CATEGORIES

for [A] may be conceived as obtained from any formula in [A] by deleting
parentheses corresponding to in the immediate scope of . For example,
we replace ((p q) r) s by (p q r) s. Note that for form sequences in
natural notation the variables for form sequences X, Y, . . . stand in dierent
contexts for dierent syntactic objects. For example, if X is the form
sequence (p q r), then in X s, the variable X stands for p q r,
while in X s it stands for (p q r). As usually done, with other kinds of
formulae and terms, we omit the outermost parentheses in natural notation.
In the context of (C/E,A)-strictication, we may use as arrow terms for
arrows of the directly strictied category, arrow terms in natural notation,
i.e. arrow terms in which parentheses corresponding to in the immediate
scope of are deleted, as above. For example, we replace (((f1 f2 ) f3 )
f4 ) f5 by ((f1 f2 f3 ) f4 ) f5 . Such arrow terms correspond to planar
nite trees if does not occur in them.
If Bob is the set of objects of a C/E category B, and if G is A, and
is related to C/E as in (IVC) and (IV G) of 3.1, then the objects of B G
correspond to form sequences of B ob of both colours. If X, X1 , . . . , X1 , . . .
stand for form sequences of B ob of colour c (if X B ob , then X is both
of colour c and ), then we dene the operations { , } on the
objects of B G in the following manner:
X1 X2 =df (X1 X2 , ),
X (X1 . . . Xn , ) =df (XX1 . . . Xn , ),
(X1 . . . Xn , ) X =df (X1 . . . Xn X, ),

, ),
, ) =df (X1 . . . Xn X1 . . . Xm
(X1 . . . Xn , ) (X1 . . . Xm
for n, m 2. The operation is, intuitively, concatenation of the kind.
For the classes [A] involved in (C/E,A, )-strictication, we have to
extend the notion of form sequence to take into account the empty sequences
of colours and . For the formal denition that follows, let X be an
arbitrary set, and let c for {, } be as before.
A form sequence of X of colour is dened inductively, as before, with
the clauses (1), (2) and the following additional clause:
(0) if is the empty sequence of elements of X , then (, ) is a form
sequence of X of colour .

6.3.

Coherence of symmetric biassociative categories

119

When we need to distinguish the previous notion of form sequence from the
new notion just introduced, we call the former notion nonextended and the
latter one extended. Planar trees corresponding to form sequences in the
extended sense have leaves labelled by elements of X or by (, ).
For G being A, , we dene the operations and on the objects
of BG conceived as form sequences in the extended sense in the following
manner. For we take (, ), which corresponds to [G ], and for we
take (, ), which corresponds to [G ]. For we enlarge the denition
above with
Y (, ) =df Y,
(, ) Y =df Y,
for Y any form sequence of B ob .

6.3.

Coherence of symmetric biassociative categories

Let S be the natural logical category in L isomorphic to S of 5.1. The

dierence is that is everywhere replaced by . The primitive arrow terms

of C(S) are included in 1, b and c, while the equations of E(S) are obtained
by replacing by , and by permuting the indices of c in the equations of

E(S). So we obtain the equations ( c c) and (b c) (see the List of Equations

and the List of Categories).
To obtain the natural logical category S, we have that the logical system
C(S) is in L, , with the transformations included in the families 1, b and
c. The specic equations of E(S) are obtained by taking the union of those

of E(S) and E(S). We call natural S-categories symmetric biassociative

categories.
For G being A and C/E being S, we have that (IV C) and (IVG) of 3.1
are satised, and G is moreover a preorder. Thus we can apply the DirectStrictication Theorem of 3.2 to obtain a category CG /EG , which we will
call Sst . We call CG here C(Sst ).
As in 3.2, we have the functor GH from Sst to Rel, and it is enough
to show that this functor is faithful to conclude the following.
Symmetric Biassociative Coherence. The functor G from S to Rel is
faithful.

120

CHAPTER 6.

BIASSOCIATIVE CATEGORIES

In the remainder of this section we prove that the functor GH from Sst to
Rel is faithful.

For {, }, a c-term of C(Sst ) is called atomized when, for every

arrow term 1X occurring in it, X is |[p]| for some letter p of P. Atom
ized c-terms may be designated by composition-free arrow terms in natural
notation, as explained in the preceding section, and these arrow terms correspond to planar nite trees analogous to those that correspond to form

sequences. To every atomized c-term we assign a planar nite tree with

nodes of n-ary branching, where n 2, such that exactly one leaf is la

belled with the head cX,Y of our c-term, and all the other leaves are labelled
with arrow terms of the form 1p . Nodes that are not leaves are labelled
with or , and for every node labelled {, } its successor is labelled
c , provided this successor is not a leaf.
Let be either the leaf above when the predecessor of is labelled
with c or is the root, or else let be the predecessor of labelled with

. The level l(f ) of an atomized c-term f is the height of (see 2.1 for

this notion of height). The span s(f ) of an atomized c-term f is the number
of nodes of the same height as on the left-hand side of .
It is easy to see that with the help of the bifunctorial and naturality
equations every arrow term of C(Sst ) is equal in Sst to a developed arrow

term fn . . . f1 1X where every factor fi is an atomized c-term, and if

1 i < j n, then l(fi ) l(fj ) and s(fi ) s(fj ). It is also easy to
see that for every arrow term f : X Y of C(Sst ) there is an arrow term
f 1 : Y X of C(Sst ) such that f 1 f = 1X and f f 1 = 1Y in Sst .
From that we conclude that to show the faithfulness of GH it is enough to
prove for f : X X that if GHf = GH1X , then f = 1X in Sst .
Let f : X X be the developed arrow term fn . . . fk . . . f1 1X of
the kind described above, such that l(fn ) = . . . = l(fk ), s(fn ) = . . . = s(fk ),
and either l(fk ) > l(fk1 ) or s(fk ) > s(fk1 ). Suppose GHf = GH1X .
If GH(fn . . . fk ) = GH1X , then by the faithfulness of GH proved in
5.2 we can conclude that fn . . . fk = 1X in Sst , and we must have
GH(fn . . . fk ) = GH1X , because, otherwise, according to our conditions
on l and s, we could not have GHf = GH1X . We repeat this reasoning
with fk1 . . . f1 1X , until we obtain that f = 1X in Sst . This concludes
the demonstration of the faithfulness of GH, from which we infer Symmetric

6.4.

Coherence of symmetric bimonoidal categories

121

Biassociative Coherence.

6.4.

Coherence of symmetric bimonoidal categories

Let S be the natural logical category in L, isomorphic to S of 5.3.

The dierence is that and are everywhere replaced by and re

spectively. The primitive arrow terms of C(S ) are included in 1, b, c and

-, while the equations of E(S ) are obtained by replacing and by

and respectively, and by permuting the indices of c in the equations of

E(S ).
To obtain the natural logical category S, , we have that the logical
system C(S, ) is in L,,, , with the transformations included in the
families 1, b, c and -. The specic equations of E(S, ) are obtained by

taking the union of those of E(S ) and E(S ) plus

( c )

( c )

c C,C = 1CC ,

c C,C = 1CC ,

provided C is a letterless formula of L,,, .

From ( c ) and ( c ) we can derive the following equations. If h : C D

and h1 : D C are mutually inverse arrows of S, , with C and D
letterless formulae of L,,, , then in E(S, ) we have

( c h)

( c h)

c C,D = (h 1C ) (1C h1 ),

c D,C = (h 1C ) (1C h1 ).

To derive ( c h) we have
(h 1C ) (1C h1 ) = (h 1C ) c C,C (1C h1 ), by ( c )

= c C,D ,

by naturality equations, bifunctorial equations, and by h being an isomor

phism; we proceed analogously for ( c h). Conversely, we obtain ( c ) and

( c ) from ( c h) and ( c h) by putting 1C : C C for h and h1 .

We call natural S, -categories symmetric bimonoidal categories.
The category Set of sets with functions is a bimonoidal category with
being cartesian product, being disjoint union, being a singleton and

being the empty set. But, although the instance c , = 1 of ( c )

122

CHAPTER 6.

BIASSOCIATIVE CATEGORIES

holds in Set, the instance c , = 1 of ( c ) does not hold. So Set is

not a symmetric bimonoidal category in the sense just specied, though it
is a symmetric biassociative category.
st
Let the category Sst
in the preceding section by
, be obtained as S
taking that G is A, instead of A, and that C/E is S, . We call CG here
C(Sst
, ).

A constant object of Sst

, is |[A]| where A is a letterless formula of
L,,, . The remaining objects of Sst
, are called variable objects. The
st
constant objects of S, are denumerable.

The arrow term cX,Y of C(Sst

, ), for {, }, is called basic when
the following two conditions are satised: rst, the form sequences (in the
extended sense) corresponding to X and Y are of colour c , and, secondly,
if the objects X and Y are both constant, then there is no arrow term
of C(Sst
, ) of type X Y (hence there is neither an arrow term of type
Y X).
A developed arrow term fn . . . f1 1Z of C(Sst
, ) such that for every

i {1, . . . , n} the head of fi is a basic arrow term cX,Y is called basically

developed. We can prove the following.
Basic-Development Lemma 1. Every arrow term of C(Sst
, ) is equal in
st
S, to a basically developed arrow term.
Proof. For f : X Y an arrow term of C(Sst
, ), we proceed by induction
on the number n 1 of nodes in the planar nite tree corresponding to
X (which must be equal to the number of nodes in the planar nite tree
corresponding to Y ). If n = 1, then X = Y , and f must be equal to 1|[p]| or
1(,) . For the induction step, we nd rst a developed arrow term equal

to f , and then, by using the equations (b c), (c ) and (c h), together with
the induction hypothesis, we transform this arrow term into a basically
developed one.

Analogously to what we had in the preceding section, we nd for every

c-term f its level l(f ). (Atomization is not here essential, since it leaves
the level invariant.) Then we have the following.

Basic-Development Lemma 2. Every arrow term of C(Sst

, ) is equal in

6.4.

Coherence of symmetric bimonoidal categories

123

Sst
, to a basically developed arrow term hk . . . h1 1X such that k 0
and, if k > 1, then for 1 i < j k we have l(hi ) l(hj ).

To prove this lemma, we use Basic-Development Lemma 1 together with

bifunctorial and naturality equations, which do not spoil basic development.
Basic-Development Lemma 3. Every arrow term f : (X1 . . . Xn , )
st
(Y1 . . . Yn , ) of C(Sst
, ), with n 2, is equal in S, to an arrow term of
the form
(. . . (f1 f2 ) . . . fn ) hk . . . h1 1X
such that fi , for i {1, . . . , n}, is of type X(i) Yi for a permutation
of {1, . . . , n}, while hk . . . h1 1X , with k 0, is basically developed and
l(hj ) = 1 for every j {1, . . . , k}.
To prove this lemma we just apply Basic-Development Lemma 2 and bifunctorial equations. Now we can prove the following.
Symmetric Bimonoidal Coherence. The functor G from S, to Rel
is faithful.
Proof. As before, it is enough to prove that the functor GH from Sst
,
to Rel is faithful. As in the preceding section, it is enough to show for
f : X X that if GHf = GH1X , then f = 1X in Sst
, .
We proceed by induction on the number n 1 of nodes in the planar nite tree corresponding to X. If n = 1, then f is equal to 1|[p]| or
1(,) . For the induction step, suppose X corresponds to the form sequence
(X1 . . . Xn , ) with n 2. Then, by Basic-Development Lemma 3, we

have that f is equal in Sst

, to (. . . (f1 f2 ) . . . fn ) h where h, which

is hk . . . h1 1X , is an instance of an arrow term

h : |[(. . . (p1 p2 ) . . .

pn )]|

|[(. . . (p(1) p(2) ) . . .

p(n) )]|

of C(Sst ), and the type of fi is X(i) Xi , for i {1, . . . , n}.

For every i {1, . . . , n} we must have (i) = i. If Xi is a variable
object of Sst
, , then this follows from the fact that f is of type X X
and GHf = GH1X . If Xi is a constant object of Sst
, , then this follows
from the fact that h is basically developed. Otherwise, there would be in

124

CHAPTER 6.

BIASSOCIATIVE CATEGORIES

h an arrow term cXj ,X(j) or cX(j) ,Xj for some j {1, . . . , n}; but this is
impossible since fj is of type X(j) Xj . So GHh = GH1|[A]| for A being

(. . . (p1 p2 ) . . . pn ). By the faithfulness of GH from Sst to Rel, we have

h = 1|[A]| , from which it follows that h = 1X in Sst
, . We have also that fi
is of type Xi Xi , and since GHfi = GH1Xi , by applying the induction
hypothesis we obtain fi = 1Xi . It follows that f = 1X in Sst

, .
In the induction step of this proof we deal with the least level, while in
the induction step of the proof of the faithfulness of GH from Sst to Rel
in the preceding section we dealt with the greatest level. Because of that,
we had to introduce there the notion of span. The preceding proof could,
however, be reworked in the style of the present sectionwith the least
level.

6.5.

The category S

To obtain the natural logical category S , we take the logical system C(S)
in L, of 6.3. The specic equations of E(S ) are those of E(S) plus the
equations

(c 1)

cA,A = 1AA

for {, }. We call S the natural logical category C(S)/E(S ). The

equations (c h) of the preceding section hold in S with A and B being any

formulae of L, such that h : A B and h1 : B A are mutually inverse

arrows of S . Conversely, we obtain (c 1) from (c h) by putting 1A : A A

for h and h1 .
To show that S is a preorder we proceed analogously to what we had
in the preceding section. Let Sst be obtained as Sst by taking that G is

A and C/E is S . Here CG is C(Sst ). Basic arrow terms cX,Y of C(Sst ) are
those where X and Y are of colour c and there is no arrow term of C(Sst )
of type X Y . Basically developed arrow terms are then dened as in
the preceding section, and we take over also the notion of level. We can
then prove analogues of Basic-Development Lemmata 1-3 of the preceding
section.

6.5.

The category S

125

To conclude the proof that S is a preorder, we prove that every arrow

term f : X X of C(Sst ) is equal in Sst to 1X . We proceed by induction
on the number n 1 of nodes in the planar nite tree corresponding to X.
This proof is analogous to the proof of Symmetric Bimonoidal Coherence
in the preceding section. In the induction step, we reason as in the case
where Xi is a constant object.

Note that, since G cA,A = G1AA , we have no functor G from S to

Rel. The fact that S is a preorder cannot be reformulated as a coherence
theorem stating that G is faithful.

The natural logical category S , whose logical system is C(S) in L , has

the specic equations of E(S) with c A,A = 1AA added. We can show that

S is a preorder by simplifying the argument above.

Chapter 7
Dissociative Categories
In this chapter we prove coherence, in the sense of preordering, for categories with a double monoidal structure without unit objects and with the
linear distribution arrows of [22]. Linear distribution is an associativity
principle involving two binary operations, and we have coined for it the
name dissociativity. This principle will yield arrows based on the usual distribution principle in Chapter 11, where the two monoidal structures are
made of a product and a coproduct.
We prove beforehand coherence for categories such as those mentioned
above that lack the ordinary associativity arrows. We also prove coherence
in cases where dissociativity is allowed only on one side. Our method in
these proofs is based on conuence techniques, like those that may be found
in the lambda calculus.
Next we prove coherence for the case where the two monoidal structures
with dissociativity are symmetric, and we still lack the unit objects. Here
the method of proof is more involved. It is based on a cut-elimination procedure in a sequent system strictied in the symmetric monoidal structure.
We justify this strictication by the results of Chapters 3-6.
We are here at the watershed as far as method is concerned. Up to
this chapter, conuence techniques predominated, while, from now on, cut
elimination, or its simpler version, composition elimination, will take over
the stage. (The two approaches are still mixed in the next chapter.) Cut
elimination could have been employed in the rst part of this chapter too.
For the categories treated there, both approaches are available, and we
127

128

CHAPTER 7.

DISSOCIATIVE CATEGORIES

opted for the rst, the second being well illustrated in the second part of
the chapter.
At the end of the chapter, we consider adding the unit objects, and we
present the linearly distributive categories of [22], for which coherence in
our sense does not obtain. Linearly distributive categories without unit
objects, with which we deal in this chapter, do not seem to have been
considered separately before.

7.1.

Coherence of dissociative categories

To obtain the natural logical category DI, we have that the logical system
C(DI) is in L, , with the transformations included in 1 and d. The
equations E(DI) are just those of E nat
C(DI) (see 4.1). We call natural DIcategories dissociative categories.
We have given in 1.2 our reasons for calling dissociativity the principle
underlying the arrow terms in d. This principle may be found in [1] (Section
15.2), [65], [89] (Section 8), [52] (Section 6.9) and [16]. In category theory
it has been introduced by Cockett and Seely (see [21] and [22]; see also [68],
Section 3.2). The dissociativity principle underlying the arrow terms in d
resembles the modularity law for lattices:
if c a, then a (b c) (a b) c
(see [9], Section I.7). The condition a (b c) (a b) c, without the
assumption c a, has the same force as distribution in lattices (cf. 11.3).
For x and y occurrences of {, } in a formula A of L, we dene

the relation SA
such that xSA
y when x is in the scope of y in A. (For the
notion of scope see 2.1.) Note that for f : A B being a member of the

family d we have that SB

is a proper subset of SA
and SA
a proper subset

of SB . This holds also for f being any arrow term of C(DI) in which a

member of the family d occurs; otherwise SA

= SB
. Here A and B are
comparable formulae of L, (namely, formulae that yield the same word
after deleting parentheses; see 2.1), and the relations SA and SB can be
compared, as the relations RA and RB were compared in 4.2. It is clear
that the following holds.

7.1.

Coherence of dissociative categories

129

Remark. Let the formula A of L, be of the form A1 A2 for {, }.

Then, for x and y being occurrences of or in Ai , where i {1, 2}, we
have that x is in the scope of y in A i x is in the scope of y in Ai .
We have the following analogue of the Extraction Lemma of 4.4, which
is proved by imitating the proof in 4.4.
Extraction Lemma. If there is an occurrence z of in a formula A of

u, then there is a formula A1 z A2

L, such that there is no u with zSA
of L, such that there is an arrow term g : A A1 z A2 of C(DI). In
addition,
()

for all occurrences x and y of or in Ai , where i {1, 2}, we

have that y is in the scope of x in A i y is in the scope of x in
Ai ;

()

every subterm of g of the form dL

D,E,F is of the type D (E z F )
(D E) z F , and every subterm of g of the form dR
F,E,D is of the
type (F z E) D F z (E D).

We do not need () for the proof of the Theoremhood Proposition below, but we stated this condition because it is analogous to () of previous
Extraction Lemmata in 4.2 and 4.4. The following lemma is analogous
to the Theoremhood Proposition of 4.2.
Theoremhood Proposition. There is an arrow term f : A B of C(DI)

i A and B are comparable formulae of L, , and we have SB

SA
and

SA
SB
.
Proof. The direction from left to right is easy. For the other direction,
we proceed by induction on the letter length n 1 of A. If n = 1, then

= SB
= , and f is 1p : p p.
SA
If n > 1 and B is B1 x B2 for x being an occurrence of , then from

SB SA
it follows that A is of the form A1 x A2 . Then, by the Remark,

we have SB
SA
and SA
SB
for i {1, 2}, and so, by the induction
i
i
i
i
hypothesis, we have the arrow terms fi : Ai Bi of C(DI). The arrow term
f is f1 f2 .

130

CHAPTER 7.

DISSOCIATIVE CATEGORIES

If n > 1 and B is B1 z B2 for z being an occurrence of , then from

SB
we conclude that there is no u such zSA
u. So, by the Extraction
Lemma of this section, there is an arrow term g : A A1 z A2 of C(DI) such
that () of the Extraction Lemma holds. Since for x and y in Ai we have

xSA
y i xSA
y, by the induction hypothesis and the Remark we have the
i
arrow terms fi : Ai Bi of C(DI), and f is (f1 f2 ) g.

SA

As explained after the proof of the Theoremhood Proposition of 4.2, with

the Theoremhood Proposition we have just proved we have solved the theoremhood problem for the category DI.
For a formula A let d(A) be the cardinality of the set of ordered pairs

SA
. If f : A B of DI is not equal to 1A : A A, then d(B) < d(A). We
can prove the following.
Dissociative Coherence. The category DI is a preorder.
Proof. Let f, g : A B be arrow terms of C(DI). We proceed by induction
on d(A)d(B) to show that f = g in DI. (Until the end of this proof, we
assume that equality of arrow terms is equality in DI.) If d(A) = d(B),
then we conclude that f = g = 1A .
Suppose d(B) < d(A). By the Development Lemma of 2.7, we have
that f = f2 f1 and g = g2 g1 for some d-terms f1 : A C and g1 : A D,
and some arrow terms f2 : C B and g2 : D B of C(DI). We have
d(C), d(D) < d(A). The following cases may arise.
L
(LL) The head of f1 is dL
E,F,G , and the head of g1 is dH,I,J . Under (LL),
we have the following subcases.

(LL1) The subformulae E (F G) and H (I J) have no occurrences of

letters in common within A. Then we use bifunctorial equations to obtain
two d-terms f2 : C B and g2 : D B such that f2 f1 = g2 g1 . Then

we can infer that SC

SD
= SB
and SC SD = SB , from which it

follows from the Theoremhood Proposition of this section that SB

SB

and SB

S
.
Hence,
again
by
the
Theoremhood
Proposition,
there
is
an

B
arrow term h : B B of C(DI). By applying the induction hypothesis, we
obtain that f2 = h f2 and g2 = h g2 , from which f = g follows.
(LL2)

Suppose E (F G) is a subformula of H or of I or of J; or,

7.1.

Coherence of dissociative categories

131

conversely, suppose that H (I J) is a subformula of E or of F or of G.

Then we proceed as in (LL1) by using (dL nat).
(LL3) Suppose, nally, that E is H, F is I, and G is J. Then C is D,
and f1 = g1 . We then apply the induction hypothesis to f2 , g2 : C B and
obtain f = g.
R
(LR) The head of f1 is dL
E,F,G , and the head of g1 is dJ,I,H . Under (LR),
we have two subcases that are settled analogously to (LL1) and (LL2).
There are no remaining subcases under (LR). It might seem that E could
be J I, while F G is H; in other words, E (F G) and (J I) H
would the same subformula of A of the form (J x I) (F y G) for x and

y occurrences of . Then we would have xSC

y and ySD
x, and, by the

Theoremhood Proposition of this section, we would have both xSB

y and

ySB x, which is a contradiction.

It remains to consider the following cases.

R
(RR) The head of f1 is dR
G,F,E , and the head of g1 is dJ,I,H .
L
(RL) The head of f1 is dR
G,F,E , and the head of g1 is dH,I,J .

The case (RR) is settled analogously to (LL), while the case (RL) is the
same as (LR).

It is not dicult to see that SA

= SB
and SA
= SB
implies that A

and B are the same formula of L, . Because, if SA = SB and SA

= SB
,
then, by the Theoremhood Proposition of this section, there is an arrow

term f : A B of C(DI), in which dL and dR cannot occur, because SB

is not a proper subset of SA

and SA
is not a proper subset of SB
. Hence
f must stand for an identity arrow. So there is a bijection between the

objects A of DI and the pairs of relations (SA

, SA
). From Dissociative
Coherence, we can conclude that DI is isomorphic to the category whose

objects are such pairs, and where an arrow exists between (SA
, SA
) and

(SB
, SB
) when SB
SA
and SA
SB
. Note that, as the category A
4.2, the category DI is not just a preorder, but a partial order.

of

132

CHAPTER 7.

7.2.

DISSOCIATIVE CATEGORIES

Net categories

To obtain the natural logical category DA, we have that the logical system
C(DA) is in L, with the transformations included in 1, b and d. The
specic equations of E(DA) are those of E(A) plus

(dL )

dL
AB,C,D = ( b A,B,C 1D ) dA,BC,D (1A dB,C,D ) b A,B,CD ,

(dL )

dL
D,C,BA = b DC,B,A (dD,C,B 1A ) dD,CB,A (1D b C,B,A ),

(dR )

dR
D,C,BA = (1D b C,B,A ) dD,CB,A (dD,C,B 1A ) b DC,B,A ,

(dR )

dR
AB,C,D = b A,B,CD (1A dB,C,D ) dA,BC,D ( b A,B,C 1D ),

L
L
R

dR
AB,C,D (dA,B,C 1D ) = dA,B,CD (1A dB,C,D ) b A,BC,D ,

(d b)

R
(dR
A,B,C 1D ) dAB,C,D = b A,BC,D (1A dB,C,D ) dA,B,CD .

(d b)

Note that, after replacing by , the arrow term dL

A,B,C is of the same

type as b
A,B,C , and, after replacing by , it is of the same type as b A,B,C .

Dually, dR
A,B,C is of the type of bA,B,C after these replacements. After such
replacements, the equations (dL nat) and (dR nat) become the equations

(b nat) and (b nat) (see 2.7), while all the specic equations of E(DA)
that are added to those of E(A) are related to the pentagonal equations

(b 5) and (b 5) (see 4.2 and the List of Equations at the end of the book).

We may obtain all of these equations by starting from (b 5) and replacing

one or two occurrences of by in each of the types, at the same place.

When only one occurrence is replaced, this forces three or four b-terms to
become dL -terms or dR -terms, and yields the equations (dL ), (dR ) and

(d b). The remaining three equations are obtained analogously from (b 5)

by replacing one occurrence of by . This covers all replacements of by

in (b 5), since the replacements in (b 5) may be conceived as replacements

of two occurrences of by in (b 5). (When all the three occurrences of

are replaced by in (b 5), we obtain (b 5).) There are many symmetries in

these equations.
We call natural DA-categories net categories. Ocially, in our nomenclature they would be called dissociative biassociative categories. A reason
for switching to the handy denomination net is in the connection with
the proof nets of linear logic (see [63] and [33]). The linearly distributive

7.3.

Coherence of net categories

133

categories of [22] (the old denomination of these categories is weakly distributive; cf. [25] for the renaming) are net categories in the sense above,
and all the specic equations of E(DA) may be found in [22] (Section 2.1;
see [21], Section 2.1, for an announcement). However, linearly distributive categories have also two objects and , with which one obtains a
bimonoidal structure (see 7.9).

7.3.

Coherence of net categories

For G being A and C/E being DA, we have that the conditions (IVC)
and (IV G) of 3.1 are satised, and G is moreover a preorder. Thus we
can apply the Direct-Strictication Theorem of 3.2 to obtain the category
CG /EG , which we call DAst . We call CG here C(DAst ).
Coherence for DA, which amounts to DA being a preorder, can perhaps
be deduced from a very general theorem of [8] (Theorem 5.2.4), whose proof
is only sketched in that paper, with substantial parts missing. It is not clear
whether our proof was envisaged in [8].
In order to prove that DA is a preorder, it is enough to prove that DAst
is a preorder. Our proof of the latter will be to a considerable extent analogous to the proofs of Semiassociative Coherence in 4.2 and Dissociative
Coherence in 7.1.
We identify the objects of DAst with form sequences of P (in the nonextended sense; see 6.2), which we call form sequences of letters, or, to simplify the exposition, simply form sequences. In this and in the next chapter,
form sequence will mean form sequence of letters. For these form sequences we use the variables X,Y, . . . , sometimes with indices. For every
form sequence X in natural notation, we dene a relation RX between the
set of occurrences of in X and the set of occurrences of in X. For that
we need some preliminary notions.
For every occurrence x of in a form sequence X in natural notation, if
y is the rightmost occurrence of in X such that X has a subword y (X )
with x in the form sequence X , then l(x) is the leftmost occurrence of
letter in X ; if there are no such occurrences of in X, then l(x) is the
leftmost occurrence of letter in X. Dually, if y is the leftmost occurrence
of in X such that X has a subword (X )y with x in the form sequence

134

CHAPTER 7.

DISSOCIATIVE CATEGORIES

X , then r(x) is the rightmost occurrence of letter in X ; if there are no

such occurrences of in X, then r(x) is the rightmost occurrence of letter
in X. For example, we have
p (( q
y1

l(x1 )
l(x2 )

y2

s ) (( t

y3

x2

u )

x1 r(x1 ) y4

v ))
r(x2 )

Then for an occurrence x of in X and an occurrence y of in X we

stipulate that xRX y when y is on the right-hand side of l(x) and on the
left-hand side of r(x). If X is the form sequence in the example above, then
we have RX = {(x1 , y2 ), (x1 , y3 ), (x2 , y2 ), (x2 , y3 ), (x2 , y4 )}.
We can infer the following from the denitions of l(x) and r(x).
Nonoverlapping Lemma. A form sequence in which in natural notation
an occurrence x1 of is on the left-hand side of an occurrence x2 of
cannot have a subword of the form
l(x1 ) w1 l(x2 ) w r(x1 ) w2 r(x2 ).
Proof. By the denition of r(x1 ) and l(x2 ), we have a subword (X1 )
with X1 containing x1 and ending in r(x1 ), and a subword (X2 ) with X2
containing x2 and beginning with l(x2 ). Then either (X1 ) is a proper
subword of X2 , or (X2 ) is a proper subword of X1 . This is because x1 is
on the left-hand side of x2 , and l(x2 ) is on the left-hand side of r(x1 ).
Suppose (X2 ) is a proper subword of X1 . Then r(x2 ) must be in X1 ,
because (X1 ) has x2 as a subword. But r(x2 ) cannot be in X1 , because
it is on the right-hand side of r(x1 ), which is the last occurrence of letter
in X1 . We conclude analogously that (X1 ) cannot be a proper subword
of X2 .

Note that we cannot prove the Nonoverlapping Lemma without the

assumption that x1 is on the left-hand side of x2 . Here is a counterexample:
( p
l(x1 )

( q
l(x2 )

x2

r )) (( s

t )

x1 r(x1 )

v )
r(x2 )

Note also that it is excluded that a form sequence in natural notation

has a subword of the form

7.3.

Coherence of net categories

135

l(x1 ) w1 r(x1 ) w2 r(x2 )

with l(x2 ) being r(x1 ). Otherwise, for p being r(x1 ) and l(x2 ), we would
have in our form sequence a word (n p)m with n, m 1. (Here (n is a
sequence of n left parentheses, and )m a sequence of m right parentheses,
as in the Extraction Lemma of 4.2.)
When X and Y are the same form sequence, or a pair of form sequences
that in natural notation dier only with respect to parentheses, we say that
X and Y are comparable form sequences. For comparable form sequences
X and Y , we may take that RX and RY are relations between the same
sets, and compare these relations (we did something analogous in 4.2). It
is easy to see that for every arrow term f : X Y of C(DAst ), the form
sequences X and Y are comparable, and RY RX . Moreover, if dL or dR
occurs in f , then RY is a proper subset of RX ; otherwise, RX = RY . For
example, with dL
p,q,r 1s : p (q r) s ((p q) r) s, if x1 and x2 are
respectively the left and right , and y is the , in p q r s, then we
have Rp(qr)s = {(x1 , y), (x2 , y)} and R((pq)r)s = {(x2 , y)}.
Two comparable form sequences X and Y in natural notation correspond to the same words w(X) and w(Y ) written in letters, and ,
which are obtained from X and Y respectively by deleting all parentheses.
A place in X is a subword w of w(X). There is an obvious deleting map
from subwords of X to places in X. We say that a subword v of X is at
a place w when (v) = w . (Note that dierent subwords of X can be at
the same place.) A subword x of X and a subword y of Y are at the same
place when (x) = (y). (These denitions are analogous to those we had
in 2.1.) It is easy to see that the following holds.
Remark. Let X in natural notation be of the form X1 X2 for {, }.
Then, for x and y occurrences of and respectively in Xi , for i {1, 2},
we have xRX y i xRXi y.
The following lemma is analogous to the Extraction Lemmata of 4.2,
4.4 and 7.1.
Extraction Lemma. If there is an occurrence z of in the form sequence

136

CHAPTER 7.

DISSOCIATIVE CATEGORIES

X, then there is a form sequence X1 z X2 in natural notation such that there

is an arrow term g : X X1 z X2 of C(DAst ). In addition,
()

for every occurrence x of in Xi and every occurrence y of in

Xi , where i {1, 2}, if xRX y, then xRXi y;

()

every subterm of g of the form dL

Y,Z,U is of the type Y (Z z U )
(Y Z) z U , and every subterm of g of the form dR
U,Y,Z is of the
type (U z Y ) Z U z (Y Z).

Proof. We proceed by induction on the number n 1 of occurrences of

letters in X. If n = 1, then the antecedent of the lemma is false, and the
lemma is trivially satised.
If n > 1, then X is X X for {, }. If is z, then g is 1X .
Suppose is not z, and suppose rst is and z is in X . Then, by the
induction hypothesis, we have an arrow term g : X X1 z X2 of C(DAst )
satisfying the primed version of (). The arrow term g 1X is of type

X (X1 z X2 ) X , and we have the arrow term dR

X1 ,X2 ,X (g 1X ) :
st

X X1 z (X2 X ) of C(DA ).
Suppose x is an occurrence of and y an occurrence of , and suppose
xRX y.
If x and y are both in X1 , then xRX y by the Remark above, and
hence, by the induction hypothesis, (x, y) RX1 . We settle easily in a
similar manner, with the help of the Remark, cases where x and y are both
in X2 or both in X .
If x is in X2 and y is in X , then r(x) in X1 z (X2 X ) is the rightmost
occurrence of letter of X . Otherwise, r(x) would be a letter p in a subword
p)l y of X2 such that y is an occurrence of . Then, since xRX y and y is
in between x and y, we must have xRX y , and, by the induction hypothesis
and the Remark, we would have (x, y ) RX1 z (X2 X ) , which contradicts
the fact that p in p)l y is r(x).
If x is in X and y is in X2 , then l(x) in X2 X is the leftmost
occurrence of letter of X2 , and so (x, y) RX2 X .
The case where z is in X is settled analogously by using dL
X ,X1 ,X2 .
It remains to consider the case where is but is not z. Suppose
z is in X . Then, by the induction hypothesis, we have an arrow term

7.3.

Coherence of net categories

137

g : X X1 z X2 of C(DAst ) satisfying the primed version of (). So we

have the arrow term g 1X : X X X1 z X2 X of C(DAst ).
Then we verify () by the induction hypothesis and the Remark. The
assertion () is easily checked by going over the proof above.

Note that the implication converse to () in the Extraction Lemma

above holds trivially. We do not need () for the proof of the Theoremhood Proposition below, but we stated this condition because it is analogous to () of previous Extraction Lemmata. Here is the analogue of the
Theoremhood Propositions of 4.2 and 7.1.
Theoremhood Proposition. There is an arrow term f : X Y of
C(DAst ) i X and Y are comparable form sequences and RY RX .
Proof. We have already veried above the easy direction from left to
right. For the other direction, we proceed by induction on the number
n 1 of occurrences of letters in X. If n = 1, then RY = RX = , and f
is 1p : p p.
If n > 1 and Y is Y1 x Y2 for x being an occurrence of , then, since for
every occurrence y of in Y we have xRY y, we have xRX y, which means
that X is of the form X1 x X2 . Then, by the Remark, we have RYi RXi
for i {1, 2}, and, by the induction hypothesis, we have the arrow terms
fi : Xi Yi of C(DAst ). The arrow term f is f1 f2 .
If n > 1 and Y is Y1 z Y2 for z being an occurrence of , then, by the
Extraction Lemma of this section, there is an arrow term g : X X1 z X2
of C(DAst ) such that the assertion () of the Extraction Lemma holds. If
xRYi y, then, since RY RX , by the Remark we have xRX y. By (), we
conclude that xRXi y. So, by the induction hypothesis, we have the arrow
terms fi : Xi Yi of C(DAst ), and f is (f1 f2 ) g.

As explained after the proof of the Theoremhood Proposition of 4.2, with

the Theoremhood Proposition we have just proved we have solved the theoremhood problem for the category DAst . This yields also a solution of
the theoremhood problem for the category DA, but we will examine this
latter problem separately in the next section.
For a form sequence X, let d(X) be the cardinality of the set of ordered

138

CHAPTER 7.

DISSOCIATIVE CATEGORIES

pairs RX . If f : X Y of DAst is not equal to 1X : X X, then RY is a

proper subset of RX and d(Y ) < d(X). We prove the following.
Net Coherence. The category DA is a preorder.
Proof. It is enough to show that DAst is a preorder. Let f, g : X Y be
arrow terms of C(DAst ). We proceed by induction on d(X)d(Y ) to show
that f = g in DAst . (Until the end of this proof, we assume that equality
of arrow terms is equality in DAst .) If d(X) = d(Y ), then we conclude
that X is Y , and f = g = 1X .
Suppose d(Y ) < d(X). By the Development Lemma of 2.7, we have
that f = f2 f1 and g = g2 g1 for some d-terms f1 : X Z and g1 : X U ,
and some arrow terms f2 : Z Y and g2 : U Y of C(DAst ). We have here
d(Z), d(U ) < d(X). The following cases may arise.
L
(LL) The head of f1 is dL
E,F,G , and the head of g1 is dH,I,J . (Here E, F ,
G, H, I and J stand for form sequences.) Due to the presence of (dL )
and (dL ), we can assume that E and H are not of the form (X1 . . . Xn , )
and G and J are not of the form (X1 . . . Xn , ). Under (LL), we have the
following subcases.

(LL1) The form sequences E (F G) and H (I J) have no occurrences

of letters in common within X. Then we use ( 2) and ( 2) to obtain two
d-terms f2 : Z Y and g2 : U Y such that f2 f1 = g2 g1 . Then we
can infer that RZ RU = RY , from which it follows by the Theoremhood
Proposition of this section that RY RY . Hence, again by the Theoremhood Proposition, there is an arrow term h : Y Y of DAst . By applying
the induction hypothesis, we obtain that f2 = h f2 and g2 = h g2 , from
which f = g follows.
(LL2) Suppose E (F G), in natural notation, is a subword of H or of
I or of J; or, conversely, suppose that H (I J) is a subword of E or of
F or of G. Then we proceed as in (LL1) by using (dL nat).
(LL3) Suppose, nally, that E is H and G is J. So F is I. (Due to our
assumptions about E, H, G and J, there are no other remaining subcases
under (LL).) Then Z is U , and f1 = g1 . We then apply the induction
hypothesis to f2 , g2 : Z B, and obtain f = g.

7.3.

Coherence of net categories

139

R
(LR) The head of f1 is dL
E,F,G , and the head of g1 is dJ,I,H . Due to the
presence of (dL ), (dL ), (dR ) and (dR ), we can assume that E and
H are not of the form (X1 . . . Xn , ) and G and J are not of the form
(X1 . . . Xn , ).

Under (LR), we have the following subcases. There are rst two subcases that are settled analogously to (LL1) and (LL2). The remaining
subcases are:
(LR 4)

E is J I and F G is H,

and when F G is J I we have the following two subcases:

(LR 5)
(LR 6)

F is J
F is J F

(so G is I),
(so I is F G).

(There is no subcase named (LR 3), which would be analogous to (LL3).)

(LR 4) Then by (d b) we have

R
L
L

(dR
J,I,F 1G ) dE,F,G = (1J dI,F,G ) dJ,I,H .
R
Let f2 and g2 be obtained from g1 by replacing its head dR
J,I,H by dJ,I,F 1G

and 1J dL
I,F,G respectively. It is clear that f2 f1 = g2 g1 : X Y . Then
we infer that RZ RU = RY , and we continue reasoning as in (LL1), by
applying the Theoremhood Proposition.

(LR 5) Then by (d b) we have

R
L
L

dR
EF,G,H (dE,F,G 1H ) = dE,F,GH (1E dF,G,H ).
R
Let f2 and g2 be obtained from g1 by replacing 1E dR
F,G,H by dEF,G,H

and dL
E,F,GH respectively. It is clear that f2 f1 = g2 g1 : X Y . Then
we infer that RZ RU = RY , and we continue reasoning as in (LL1), by
applying the Theoremhood Proposition.

(LR 6) We prove rst that there is an occurrence z of in J F G

such that for every occurrence x of in E (J F G) H we do not
have xRY z.
Let u be an occurrence of in the word E . For every such u the
occurrence of letter r(u) in Y is either in J or in F . Let p be the rightmost
of these occurrences of letters.

140

CHAPTER 7.

DISSOCIATIVE CATEGORIES

If p is in J, then we take z to be the between J and F . By the

denition of r(u), there is no occurrence u of in E such that uRY z.
Since RY RX , there is no occurrence x of in J such that xRY z, and,
since RY RU , there is no occurrence x of in the word F G) H such
that xRY z.
If p is in F , then we take z to be the occurrence of on the right-hand
side of p nearest to p. This z is either in F , or it is the between F and
G. By the denition of r(u), there is no occurrence u of in E such that
uRY z. Since RY RX , there is no occurrence x of in J such that xRY z.
If there is an occurrence x of in F G) H such that xRY z, then in
Y we have that l(x) is on the left-hand side of p, which is r(u) for some
occurrence u of in E . (As we said after the proof of the Nonoverlapping
Lemma, it is excluded that l(x) coincides with r(u).) Since RY RU , we
must have that l(u) is on the left-hand side of l(x), and, since xRY z, we
must have that r(x) is on the right-hand side of r(u). Since x is on the
right-hand side of u, all this contradicts the Nonoverlapping Lemma. Hence
we do not have xRY z.
There are now two possibilities for the z we have found. Suppose rst
that J F G is of the form KzL. Then we have three subcases:
(LR 6.1)
(LR 6.2)
(LR 6.3)

K is J (so L is F G),
L is G (so K is J F ),
K is J F1 and L is F2 G (so F is F1 F2 ).

(LR 6.1) Then by (dL ) we have

L
L

dL
E,J,I 1H = ((dE,J,F 1G ) 1H ) (dE,F,G 1H ).
L
Let f1 and f1 be obtained from g1 by replacing 1E dR
J,I,H by dE,J,I 1H
L

and (dE,J,F 1G ) 1H respectively. It is clear that f1 = f1 f1 : X Z .
Suppose xRY y. If x is not in E (J y F G) H, then it is easy to infer
that xRZ y. If x is in E (J y F G) H, then it is either in E (J or
in F G) H. In the rst case, y is on the left-hand side of z, and in the
second case, it is on the right-hand side of z. In both cases, we get xRZ y.
So RY RZ , and, by the Theoremhood Proposition of this section, we
obtain an arrow term f2 : Z Y of C(DAst ). By the induction hypothesis,

7.3.

Coherence of net categories

141

we have that f2 = f2 f1 , where d(Z) < d(X). We continue reasoning as

in subcase (LR 5), starting from f1 and g1 . There we apply (d b).

The subcase (LR 6.2) is settled analogously to (LR 6.1) by using (dR ),
and for subcase (LR 6.3) we use both (dL ) and (dR ) to reduce it to

(LR 5), where we apply (d b).

Suppose now that J F G is not of the form K z L. So z is in F ,
but F is not of the form F1 z F2 . Then, by the Extraction Lemma of
this section, there is a form sequence F of the form F1 z F2 in natural
notation such that there is an arrow term h : F F of C(DAst ) with ()
being satised. Let f1 , g1 , h , f2 and g2 be obtained from g1 by replacing
1E dR
J,I,H respectively by
dL
E,JF ,G 1H ,
1E d R
J,F G,H ,
1E (1J h 1G ) 1H ,
((1E (1J h)) 1G ) 1H ,
1E (1J ((h 1G ) 1H )).
Then, by (dL nat) and (dR nat), we have that f1 h = f2 f1 and g1 h =
g2 g1 . For h being of the type X X , we have that RY RX , which
follows easily from our assumption about z and from () of the Extraction
Lemma. For f2 being of the type Z Z and g2 of the type U U , we
infer that RZ = RZ RX and RU = RU RX . So, by the Theoremhood
Proposition of this section, we have the arrow terms f2 : Z Y and g2 :
U Y of C(DAst ). Then we apply the induction hypothesis to f2 , f2 f2 :
Z Y and g2 , g2 g2 : U Y , and also to f2 f1 , g2 , g1 : X Y , where
d(Z), d(U ), d(X ) < d(X).
It remains to consider the following cases:
(RR)

R
the head of f1 is dR
G,F,E , and the head of g1 is dJ,I,H ;

(RL)

L
the head of f1 is dR
G,F,E , and the head of g1 is dH,I,J .

The case (RR) is settled analogously to (LL), while the case (RL) is the
same as (LR).

It is not dicult to see that RX = RY implies that the form sequences

X and Y coincide. Because, if RX = RY , then, by the Theoremhood

142

CHAPTER 7.

DISSOCIATIVE CATEGORIES

Proposition of this section, there is an arrow term f : X Y of C(DAst ), in

which dL and dR cannot occur, because RY is not a proper subset of RX .
Hence f must stand for an identity arrow. So there is a bijection between
the objects X of DAst and the relations RX . From Net Coherence, we
can conclude that DAst is isomorphic to the category whose objects are
the relations RX , and where an arrow exists between RX and RY when
RY RX .

7.4.

Net normal form

In this section we will examine the theoremhood problem (in the sense of
1.1) for the category DA, and we will nd a solution for it dierent from
that suggested by the Theoremhood Proposition of the preceding section.
This solution will also yield a unique normal form for arrow terms of C(DA),
i.e. a normal form such that arrow terms of C(DA) in normal form are equal
in DA i they are the same arrow term.
Consider a formula B of L, . Let B be obtained from B by replacing
every by . Let I(B) be obtained from I(B ) (see 4.4) by putting back
the occurrences of where they were in B, while keeping the subscripts of
I(B ).
Let A be a formula comparable with B (which means that A and B are
the same after deleting parentheses). Next, let A be obtained from A by
adding to every occurrence x of or in A the subscript x has in I(B).
Then we have the following proposition.
Theoremhood Proposition. There is an arrow term f : A B of
C(DA) i A and B are comparable formulae of L, and
()

in A dened with respect to I(B), for every n, m 2 there is

no nm in the scope of n .

Proof. From left to right, suppose we have an arrow term f : A B of

C(DA) such that () fails in A . Then for fG : X Y in DAst we can nd
in X and Y an occurrence x of corresponding to nm and an occurrence
y of corresponding to n . Since () fails in A , we do not have xRX y,
but the subscripts of I(B) tell us that we have xRY y, which contradicts
the easy, left-to-right, direction of the Theoremhood Proposition of 7.3.

7.5.

Coherence of semidissociative biassociative categories

143

For the other direction, we proceed as follows. By the Associative

Normal-Form Proposition of 4.4, there is an arrow term f : A B

of C(A) such that in (f ) : (A ) I(B ) for every subterm of the form

b
D,E,F of type D l (E k F ) (D l E) k F , and every subterm of the

form b
F,E,D of type (F k E) l D F k (E l D), we have that l and
every subscript in D, E and F divides k. We build out of (f ) an arrow
term f : A I(B) by putting back at some places, as required by
A and B. In transforming (f ) into f , some subterms of (f ) in the

family b may remain in that family, and some may be transformed into

arrow terms in the families b, dL or dR . It is excluded that the type of a

subterm of (f ) in the family b becomes D l (E k F ) (D l E) k F or

(F k E) l D F k (E l D), which would prevent its being transformed

in an arrow term in the families b, b, dL or dR . This is guaranteed by

(), and by the fact that for every subterm of (f ) in the family b of a
type (G ) (H ) we have that H satises (), as A does. We obtain
f : A B by deleting the subscripts of f .

The procedure of the proof of the right-to-left direction of this proposition, which presupposes the results of 4.4, gives rise to a unique arrow
term, which we may consider to be in normal form.
We could imagine a proof of Net Coherence where instead of relying on
the Theoremhood Proposition of the preceding section, we would rely on a
strictied version of the Theoremhood Proposition of this section.

7.5.

Coherence of semidissociative biassociative categories

To obtain the natural logical category DLA, we have that the logical system
C(DLA) is in L, , with the transformations included in 1, b and dL . So,
in contradistinction to C(DA), we do not have dR . The specic equations
of E(DLA) are those of E(A) plus (dL ) and (dL ) of 7.2. We call natural
DLA-categories semidissociative biassociative categories.
For G being A and C/E being DLA, we have that the conditions (IV C)
and (IV G) of 3.1 are satised, and G is moreover a preorder. Thus we
can apply the Direct-Strictication Theorem of 3.2 to obtain a category

144

CHAPTER 7.

DISSOCIATIVE CATEGORIES

CG /EG , which we call DLAst . We call CG here C(DLAst ).

Our proof that DLAst is a preorder is to a considerable extent analogous
to the proof that DAst is a preorder, and we assume the notions dened in
7.3. The new proof is somewhat more complicated as far as the denitions
of the relation RX is concerned.
For every object X of DLAst , i.e. for every form sequence X in the
l
r
natural notation of 6.2, we dene two relations RX
and RX
between the
set of occurrences of in X and the set of occurrences of in X. We have
l
xRX
y when the occurrence y of is in between l(x) and the occurrence x
r
of , and we have xRX
y when y is between x and r(x). It is clear that RX
r
l
and RX
.
is the disjoint union of RX
It is easy to verify that, for every arrow term f : X Y of C(DLAst ),
r
l
we have RYr RX
and RX
= RYl . Moreover, if dL occurs in f , then RYr
r
is a proper subset of RX
. It is also easy to verify that the Remark of 7.3
holds when we replace R by Rl and Rr . Then we can prove the following
analogue of the Extraction Lemma of 7.3.
Extraction Lemma. If there is an occurrence z of in the form sequence
X such that there is no occurrence x of in X with xRl z, then there is a
form sequence X1 z X2 in natural notation such that there is an arrow term
g : X X1 z X2 of C(DLAst ). In addition,
()

for every occurrence x of in Xi and every occurrence y of

r
r
l
in Xi , where i {1, 2}, if xRX
y, then xRX
y; moreover, RX
=
i
l
RX1 z X2 ,

()

every subterm of g of the form dL

Y,Z,U is of the type Y (Z z U )
(Y Z) z U .

The proof is obtained by excluding the case where is and z is in X in

the proof of the Extraction Lemma of 7.3.
Next we state the analogue of the Theoremhood Proposition of 7.3.
Theoremhood Proposition. There is an arrow term f : X Y of
C(DLAst ) i X and Y are comparable form sequences, and we have RYr
r
l
RX
and RX
= RYl .

7.6.

Symmetric net categories

145

The proof is again a slight modication of the proof of the Theoremhood

Proposition of 7.3.
The proof that the category DLAst is a preorder is then obtained by
proceeding as in the proof of Net Coherence of 7.3. We keep just the cases
analogous to (LL) cases. So we have the following.
Semidissociative Biassociative Coherence. The category DLA is a
preorder.
Analogously to what we had at the end of 4.2, 7.1 and 7.3, with the
help of Semidissociative Biassociative Coherence, we obtain that DLAst is
r
l
isomorphic to a category whose objects are pairs of relations (RX
, RX
), and
r
l
r
l
r
r
where an arrow exists between (RX , RX ) and (RY , RY ) when RY RX
l
and RX
= RYl .

7.6.

Symmetric net categories

To obtain the natural logical category DS, we have that the logical system
C(DS) is in L, , with the transformations included in 1, b, c and d.
The specic equations of E(DS) are obtained by taking the union of those
of E(DA) and E(S) plus

(dR c) dR
C,B,A = c C,BA ( c A,B 1C ) dA,B,C (1A c B,C ) c CB,A .

We call natural DS-categories symmetric net categories. In 12.4 we

will give a concrete example of a symmetric net category in which and
are not isomorphic. (See 11.3 for the question whether the category Set
of sets with functions is a symmetric net category.)
In the presence of (dR c), the equations (dR nat), (dR ) and (dR ) become derivable from the remaining equations. Note that (dR c) may be

conceived as a denition of dR in terms of dL , c and c. So we may as

well assume that in C(DS) we do not have dR , but only dL , and that dR is
dened by (dR c). We make this assumption in 7.6-8, and we write simply
d for dL , omitting the superscript L . This convention will be in force also
later on whenever we have (dR c) (especially in Chapter 11).
To give some alternative axioms for E(DS) we introduce the following
denitions:

146

CHAPTER 7.

DISSOCIATIVE CATEGORIES

e A,B,C,D =df dA,D,BC (1A c D,BC ) (1A dB,C,D ) b

A,B,CD
is of type (A B) (C D) (A D) (B C);
e A,B,C,D =df e A,B,D,C (1AB c D,C )

is of type (A B) (C D) (A C) (B D).
Dually, we have that

e D,C,B,A =df b
DC,B,A (dD,C,B 1A ) ( c CB,D 1A ) dCB,D,A

is of type (C B) (D A) (D C) (B A);
e D,C,B,A =df ( c C,D 1BA ) e C,D,B,A

is of type (D B) (C A) (D C) (B A).
Then we can state the following equations:

( e)

( e)

c BC,AD

e A,B,C,D = e B,A,C,D ( c A,B 1CD ),

(1DC c B,A ) e D,C,A,B = e D,C,B,A

c DA,CB .

These two equations are mirror images of each other. The equation ( e) can

replace (d b), and the equation ( e) can replace (d b), in our axiomatization
of E(DS).
For every transformation in the logical system C(DS) we have that
in A1 ,...,Ak : M (A1 , . . . , Ak ) N (A1 , . . . , Ak ) the functions and
are bijections, and hence the type M (A1 , . . . , Ak ) N (A1 , . . . , Ak ) is
balanced (see 3.3). Therefore, the type of every arrow term of C(DS) is
balanced.
For C/E being C(DS)/E(DS), that is DS, and C being C(S) of 6.3,
we have that the condition (IVC) of 3.1 is satised. Next, let G be the
C -core C /E of C/E. By Symmetric Biassociative Coherence, and by the
fact that if f = g in E(DS), then Gf = Gg in Rel, we can conclude that
G is the natural logical category S. The category G is a groupoid, and it
ows through DS, so that the condition (IV G) of 3.1 is satised.
Let E pr be the equational system obtained by extending E(DS) with the

equation (c 1) of 6.5, namely cA,A = 1AA for {, }. We know that

7.6.

Symmetric net categories

147

C /E pr , that is C(S)/E pr , which is the category S of 6.5, is a preorder.

Next, for every equation f = g in E pr that is not in E(DS), we can show
that the type of f and g is not diversied. We prove by induction on the
length of derivation that if f = g is in E pr and the arrow terms f and g
are diversied, then every derivation of f = g is made of equations between
diversied arrow terms. (The only problem is when in such a derivation we
pass from f1 = f2 and g1 = g2 to g1 f1 = g2 f2 , in which case we appeal
to the fact that the types of arrow terms of C(DS) are always balanced.)
Then, as in 3.3, we have that (IVC) and (IVG) hold when C(DS)/E(DS)
is replaced by C(DS)/E pr . Now the C -core G of C(DS)/E pr is a preorder.
By the Direct-Strictication Theorem of 3.2, we obtain the C(S)-strict
pr
C(DS)/E pr -category C(DS)G /E pr
G equivalent to C(DS)/E . As in 3.3, for
diversied arrow terms f and g of C(DS) of the same type, we have f = g
in DS i fG = gG in C(DS)G /E pr
G .
Since the type of every arrow term of C(DS) is balanced, for every arrow
term f : A B of C(DS) there is a diversied arrow term f div : Adiv B div
of C(DS) such that f is obtained by substituting uniformly letters for some
letters in f div : Adiv B div . Namely, f is a letter-for-letter substitution
instance of f div . Here we assume that the generating set P is innite (see
4.1).
Our purpose is to show the following.
Symmetric Net Coherence. The functor G from DS to Rel is faithful.
According to what we said above, to prove this coherence we can proceed
as follows. Suppose Gf = Gg in Rel for the arrow terms f and g of C(DS)
of the same type. Then we can nd f div and g div of the same type, and we
will prove
(div )

fGdiv = gGdiv in C(DS)G /E pr

G ,

which implies f div = g div in DS, from which we can conclude, by applying
(su ) (see 2.7), that f = g in DS. So, to prove Symmetric Net Coherence,
we have only to prove (div ) under the assumption Gf = Gg.
We proceed with this proof in the next two sections. In 7.7, we prove a
theorem that says that the equations of C(DS)G /E pr
G cover a normalization

148

CHAPTER 7.

DISSOCIATIVE CATEGORIES

procedure analogous to Gentzens cut-elimination procedure of [60]. In

7.8, we prove additional results, which together with our cut elimination
will yield (div ) under the assumption Gf = Gg. In logic, these results
correspond to inverting rules in derivations, i.e. passing from conclusions
to premises. This invertibility is guaranteed by the possibility to permute
rules, i.e. change their order in derivations, and we show for that permuting
that it is covered by the equations that hold in C(DS)G /E pr
G . This means
that the equations of DS also cover a cut elimination and invertibility, but
this cut elimination and invertibility are more cumbersome to record within
DS than within C(DS)G /E pr
G .

7.7.

Cut elimination in GDS

To formulate the cut-elimination result announced at the end of the preceding section, we need some preliminary notions. The objects of the category
C(DS)G /E pr
G (see the preceding section) correspond bijectively to something
we will call form multisets of letters. We dene this notion as follows.
We say that the form sequences of letters X and Y are c-equivalent
when there is an arrow term of C(S)G of type X Y . It is clear that
c-equivalence is an equivalence relation congruent with the operations
on form sequences of letters for {, }. A form multiset of letters
is the equivalence class of a form sequence of letters with respect to cequivalence. (We exclude here the empty form sequences (, ).) As before
in this chapter, we presuppose that form multisets and form sequences are
of letters, i.e. of P, and omit mentioning that all the time. We can use
form sequences, and, in particular, form sequences in natural notation, to
designate form multisets. For example, p q (p r p) in natural notation
stands for the same form multiset as q (r p p) p.
For A a diversied formula of L, (see 3.3), the form multiset |[A]| is
such that every letter of P occurs in it at most once. Such a form multiset
is called a form set.
Let GDS be the full subcategory of C(DS)G /E pr
G whose objects are
all the objects that correspond to form sets of letters. We write G in
the name of GDS because of the relationship we are going to establish
between this category and Gentzens sequent systems. The category GDS

7.7.

Cut elimination in GDS

149

is a syntactical category in a syntactical system called C(GDS), which is a

subsystem of C(DS)G .
In this section, X, Y , Z, . . ., X1 , . . . will be form sequences of letters that
stand for form sets of letters, and the operations on form sequences, for
{, }, will be written simply , with omitted.
We dene by induction a set of terms for arrows of GDS, which we call
Gentzen terms. First, we stipulate that for every letter p the term 1p : p p,
which denotes the arrow 1|[p]| of GDS, is a Gentzen term. The remaining
Gentzen terms are obtained by closing under the following operations on
Gentzen terms, which we call Gentzen operations. We present these operations by inductive clauses in fractional notation, which are interpreted as
saying that if the terms above the horizontal line are Gentzen terms, then
the term below the horizontal line is a Gentzen term (cf. 2.2). The schema
on the left-hand side of the =dn sign stands for the Gentzen term, while the
schema on the right-hand side stands for the arrow denoted by this term.
Our Gentzen operations correspond to Gentzens rules for cut, introduction
of conjunction on the right and introduction of disjunction on the left:
f:U X Z

g: X Y W

cutX (f, g) =dn (g 1Z ) dY,X,Z (f 1Y ) : U Y Z W

f:U X

g: X Y W

cutX (f, g) =dn g (f 1Y ) : U Y W

f:U X Z

g: X W

cutX (f, g) =dn (g 1Z ) f : U W Z

f:U X

g: X W

cutX (f, g) =dn g f : U W

f1 : U1 X1 Z1

f2 : U2 X2 Z2

X1 ,X2 (f1 , f2 ) : U1 U2 (X1 X2 ) Z1 Z2

where X1 ,X2 (f1 , f2 ) =dn (dX2 ,X1 ,Z1 1Z2 ) dX1 Z1 ,X2 ,Z2 (f1 f2 ),

150

CHAPTER 7.

f1 : U1 X1 Z1

DISSOCIATIVE CATEGORIES

f2 : U2 X2

X1 ,X2 (f1 , f2 ) =dn dX2 ,X1 ,Z1 (f1 f2 ) : U1 U2 (X1 X2 ) Z1

f1 : U1 X1

f2 : U2 X2

X1 ,X2 (f1 , f2 ) =dn f1 f2 : U1 U2 X1 X2

f1 : X1 Z1 U1

f2 : X2 Z2 U2

X1 ,X2 (f1 , f2 ) : (X1 X2 ) Z1 Z2 U1 U2

where X1 ,X2 (f1 , f2 ) =dn (f1 f2 ) dZ2 ,X2 ,X1 Z1 (dZ1 ,X1 ,X2 1Z2 ),
f1 : X1 Z1 U1

f2 : X2 U2

X1 ,X2 (f1 , f2 ) =dn (f1 f2 ) dZ1 ,X1 ,X2 : (X1 X2 ) Z1 U1 U2

f1 : X1 U1

f2 : X2 U2

X1 ,X2 (f1 , f2 ) =dn f1 f2 : X1 X2 U1 U2

Note that X1 ,X2 (f1 , f2 ) = X2 ,X1 (f2 , f1 ) holds in GDS. (In case we have

f1 : U1 X1 Z1 and f2 : U2 X2 Z2 , we apply (d b) of 7.2.) We will

consider the terms on the two sides of this equation as the same Gentzen
term. Analogously, X1 ,X2 (f1 , f2 ) = X2 ,X1 (f2 , f1 ) holds in GDS. (In

case we have f1 : X1 Z1 U1 and f2 : X2 Z2 U2 , we apply (d b) of

7.2.) We will consider also the terms on the two sides of this equation
as the same Gentzen term. We do something analogous for arrow terms
of C(GDS) built with and . Namely, we may omit some parentheses
without ambiguity, and order is irrelevant. For example, f g h stands
for (f g) h, or g (f h), etc., because all these arrow terms are equal
in GDS.
In all the inductive clauses of Gentzen operations above, the Gentzen
terms dened must denote arrows of GDS. So, for example, for f1 : U1 X1
and f2 : U2 X2 in X1 ,X2 (f1 , f2 ) : U1 U2 X1 X2 , we must have that
U1 U2 and X1 X2 correspond to form sets of letters, which means that

7.7.

Cut elimination in GDS

151

U1 and U2 cannot have letters in common, and the same for X1 and X2 .
So all our Gentzen operations are partial operations.
We can then prove the following lemma.
Gentzenization Lemma. Every arrow of GDS is denoted by a Gentzen
term.
Proof. We show by induction on the number of letters in the form set X
that 1X is denoted by a Gentzen term. For that, we rely on the following
equations of GDS:
1X1 X2 = 1X1 1X2 , for

{, },

()

f1 f2 = X1 ,X2 (f1 , f2 ), for f1 : U1 X1 and f2 : U2 X2 ,

()

f1 f2 = X1 ,X2 (f1 , f2 ), for f1 : X1 U1 and f2 : X2 U2 ,

provided f1 and f2 are Gentzen terms (the equations () and () are

trivial).
If for any form set X we have that 1X stands for a Gentzen term, then
we have in GDS
dY,X,Z = cutX (1XZ , 1XY ) = Y,X (1Y , 1XZ ) = Z,X (1Z , 1XY ).
It remains only to note that, besides the equations () and () above,
we have in GDS the equation g f = cutX (f, g) for the Gentzen terms
f : U X and g : X W .

A Gentzen term is cut-free when it has no subterm of the form cutX (f, g).
A Gentzen term of the form cutX (f, g) such that f and g are cut-free is
called a topmost cut.
We dene inductively the depth of a subterm of a Gentzen term:
f is a subterm of f of depth 0;
if is cutX or X1 ,X2 or X1 ,X2 , and (f1 , f2 ) is a subterm of f of
depth n, then f1 and f2 are subterms of f of depth n + 1.
For a topmost cut cutX (f, g) such that X is of colour and is not a
letter, we say that the -rank of cutX (f, g) is n 0 when f has a subterm
X1 ,X2 (f1 , f2 ) of depth n such that X is X1 X2 . Because the objects of

152

CHAPTER 7.

DISSOCIATIVE CATEGORIES

GDS are form sets, i.e., they are diversied, there can be at most one
subterm of f of that form. For a topmost cut cutX (f, g) such that X is of
colour and is not a letter, we say that the -rank of cutX (f, g) is n 0
when g has a subterm X1 ,X2 (g1 , g2 ) of depth n such that X is X1 X2 .
For a topmost cut cutp (f, g), note that 1p must be a subterm of both f and
g, which occurs in each of them exactly once, because of diversication.
We say that the p-rank of cutp (f, g) is n 0 when n is the sum of the depth
of 1p in f and of the depth of 1p in g.
The rank of a topmost cut cutX (f, g) is either its -rank, or -rank, or
p-rank, depending on X.
The complexity of a topmost cut cutX (f, g) is (m, n) where m 1 is
the number of letters in X and n 0 is the rank of this cut. Complexities
are ordered lexicographically; i.e., we have (m1 , n1 ) < (m2 , n2 ) i either
m1 < m2 , or m1 = m2 and n1 < n2 .
We can prove the following theorem for GDS.
Cut-Elimination Theorem. For every Gentzen term t there is a cut-free
Gentzen term t such that t = t in GDS.
Proof. By induction on the complexity of a topmost cut cutX (f, g), we
prove that cutX (f, g) is equal in GDS to a cut-free Gentzen term. From
this the theorem follows. In the remainder of this proof we assume that
equality between arrow terms is equality in GDS.
For the basis we have that if the complexity of cutX (f, g) is (1, 0), then
cutX (f, g) is of the form cutp (1p , 1p ), which is equal to 1p .
Suppose now the complexity is (m, 0) for m > 1, and suppose X is of
colour . Then cutX (f, g) is of the form cutX1 X2 (X1 ,X2 (f1 , f2 ), g), and
we have the following cases.
(1.1)

Consider the Gentzen term

f1 : U1 X1 Z1

f2 : U2 X2 Z2

X1 ,X2 (f1 , f2 ) : U1 U2 (X1 X2 ) Z1 Z2

g : X1 X2 Y W

cutX1 X2 (X1 ,X2 (f1 , f2 ), g) : U1 U2 Y W Z1 Z2

Then consider the Gentzen term

7.7.

Cut elimination in GDS

f1 : U1 X1 Z1
f2 : U2 X2 Z2

153

g : X1 X2 Y W

cutX1 (f1 , g) : U1 X2 Y W Z1

cutX2 (f2 , cutX1 (f1 , g)) : U1 U2 Y W Z1 Z2

We show that
()

cutX1 X2 (X1 ,X2 (f1 , f2 ), g) = cutX2 (f2 , cutX1 (f1 , g)),

and the Gentzen term on the right-hand side has a topmost cut cutX1 (f1 , g)
of lower complexity (m , n ) than the Gentzen term on the left-hand side;
here m < m. Hence, by the induction hypothesis, it is equal to a cut-free
Gentzen term h, and cutX2 (f2 , h) is a topmost cut of lower complexity, to
which we can also apply the induction hypothesis.
To show (), we have to show
(g 1Z1 Z2 ) dY,X1 X2 ,Z1 Z2
(((d
X2 ,X1 ,Z1 1Z2 ) dX1 Z1 ,X2 ,Z2 (f1 f2 )) 1Y ) =
(((g 1Z1 ) dY X2 ,X1 ,Z1 (f1 1Y X2 )) 1Z2 ) dU1 Y,X2 ,Z2 (f2 1U1 Y ),
and to derive this equation for GDS we use essentially (dL ) of 7.2.
(1.2) If we have g : X1 X2 W , while f1 and f2 are as in (1.1), then
to show () we have to show
(g 1Z1 Z2 ) (dX2 ,X1 ,Z1 1Z2 ) dX1 Z1 ,X2 ,Z2 (f1 f2 ) =
((g 1Z1 ) dX2 ,X1 ,Z1 (f1 1X2 )) 1Z2 ) dU1 ,X2 ,Z2 (f2 1U1 ),
which follows readily with the help of (dL nat).
(2.1) If we have f1 : U1 X1 , while f2 and g are as in (1.1), then to
show () we have to show
(g 1Z2 ) dY,X1 X2 ,Z2 ((dX1 ,X2 ,Z2 (f1 f2 )) 1Y ) =
((g (f1 1Y X2 )) 1Z2 ) dU1 Y,X2 ,Z2 (f2 1U1 Y ),
which follows by using essentially (dL ).
(2.2) If we have f1 : U1 X1 , g : X1 X2 W and f2 as in (1.1), then
to show () we have to show

154

CHAPTER 7.

DISSOCIATIVE CATEGORIES

(g 1Z2 ) dX1 ,X2 ,Z2 (f1 f2 ) =

((g (f1 1X2 )) 1Z2 ) dU1 ,X2 ,Z2 (f2 1U1 ),
which follows readily with the help of (dL nat).
(3.1) If we have f1 : U1 X1 , f2 : U2 X2 and g as in (1.1), then to
show () we have to show
g (f1 f2 1Y ) = g (f1 1Y X2 ) (f2 1U1 Y ),
which follows from bifunctorial equations.
(3.2) If we have f1 : U1 X1 , f2 : U2 X2 and g : X1 X2 W , then to
show () we have to show
g (f1 f2 ) = g (f1 1X2 ) (f2 1U1 ),
which follows again from bifunctorial equations.
If the complexity of cutX (f, g) is (m, 0) for m > 1, and X is of colour
, then we proceed analogously.
Suppose now the complexity of cutX (f, g) is (m, n) with m, n 1. Then
we have the following cases.
(4) The form set X is of colour (it may be of the form X1 X2 or p)
and f is V1 ,V2 (f1 , f2 ). So, since n 1, we have
f1 : U1 V1 X Z1

f2 : U2 V2 Z2

V1 ,V2 (f1 , f2 ) : U1 U2 (V1 V2 ) X Z1 Z2

g: X Y W

cutX (V1 ,V2 (f1 , f2 ), g) : U1 U2 Y (V1 V2 ) Z1 Z2 W

Then consider the Gentzen term
f1 : U1 V1 X Z1

g: X Y W

cutX (f1 , g) : U1 Y V1 Z1 W

f2 : U2 V2 Z2

V1 ,V2 (cutX (f1 , g), f2 ) : U1 U2 Y (V1 V2 ) Z1 Z2 W

We show that

7.7.

Cut elimination in GDS

()

155

cutX (V1 ,V2 (f1 , f2 ), g) = V1 ,V2 (cutX (f1 , g), f2 ),

and the complexity (m, n1) of the topmost cut cutX (f1 , g) is lower than
(m, n), so that we may apply the induction hypothesis.
To show the equation (), we have to show
(g 1(V1 V2 )Z1 Z2 ) dY,X,(V1 V2 )Z1 Z2
(((d
V2 ,V1 ,XZ1 1Z2 ) dV1 XZ1 ,V2 ,Z2 (f1 f2 )) 1Y ) =
(dV2 ,V1 ,Z1 W 1Z2 ) dV1 Z1 W,V2 ,Z2
(((g 1
V1 Z1 ) dY,X,V1 Z1 (f1 1Y )) f2 ).

To derive this equation for GDS, we use essentially (d b) and (d b), besides
(dL nat), (dL ) and (dL ) (see 7.2). We have also to consider cases where
we have f1 : U1 V1 X, or f2 : U2 V2 , or g : X W (analogously to
what we had in (1.1)-(3.2)). In all of them, () amounts to equations
simpler than the equation above, which all hold in GDS.
(5) The form set X is of colour , and f is V1 ,V2 (f1 , f2 ), so that we
have f1 : V1 U1 X Z1 , f2 : V2 U2 Z2 and g : X Y W . Then we
have to show the equation
()

cutX (V1 ,V2 (f1 , f2 ), g) = V1 ,V2 (cutX (f1 , g), f2 )

with the complexity (m, n 1) of cutX (f1 , g) lower than (m, n). To show
this equation, we have to show
(g 1Z1 Z2 ) dY,X,Z1 Z2
(((f f ) d
1
2
U2 ,V2 ,V1 U1 (dU1 ,V1 ,V2 1U2 )) 1V ) =
(((g 1Z1 ) dY,X,Z1 (f1 1Y )) f2 ) dU2 ,V2 ,V1 U1 Y (dU1 Y,V1 ,V2 1U2 ),

and to derive that for GDS we use essentially (d b). We have also to
consider cases where we have f1 : V1 X Z1 , or f1 : V1 U1 X, or
f1 : V1 X, or f2 : V2 Z2 , or g : X W . In all of them (), amounts to
simpler equations, which all hold in GDS.
It remains to consider cases with complexity (m, n) where m, n 1, and
X is of colour (it may be of the form X1 X2 or p). These additional

156

CHAPTER 7.

DISSOCIATIVE CATEGORIES

cases are settled dually to cases (4) and (5). Note that in cases with
complexity (m, n) where m, n 1 and X is a letter (hence m = 1) we have
that X is of both colours, and hence in these cases we can proceed either
as in (4) and (5), or as in cases dual to (4) and (5) we have just
mentioned.

7.8.

Invertibility in GDS

In this section, we prove the invertibility results announced at the end of

7.6. First, we cover some preliminary matters. We will need later the
following equations of GDS:
( 1) W1 W2 ,W3 (W1 ,W2 (f1 , f2 ), f3 ) = W1 ,W2 W3 (f1 , W2 ,W3 (f2 , f3 ))
for fi of type Hi Wi Ji or Hi Wi , where i {1, 2, 3},
( 2) R1 ,R3 (W1 ,W2 (f1 , f2 ), f3 ) = W1 ,W2 (R1 ,R3 (f1 , f3 ), f2 )
for f1 of type H1 W1 R1 J1 or H1 W1 R1 ,
f2 of type H2 W2 J2 or H2 W2 , and
f3 of type H3 R3 J3 or H3 R3 ,
( 1) W1 W2 ,W3 (W1 ,W2 (f1 , f2 ), f3 ) = W1 ,W2 W3 (f1 , W2 ,W3 (f2 , f3 ))
for fi of type Wi Ji Hi or Wi Hi , where i {1, 2, 3},
( 2) R1 ,R3 (W1 ,W2 (f1 , f2 ), f3 ) = W1 ,W2 (R1 ,R3 (f1 , f3 ), f2 ),
for f1 of type W1 R1 J1 H1 or W1 R1 H1 ,
f2 of type W2 J2 H2 or W2 H2 , and
f3 of type R3 J3 H3 or R3 H3 ,
()

W2 ,W3 (V1 ,V2 (f1 , f2 ), f3 ) = V1 ,V2 (f1 , W2 ,W3 (f2 , f3 )),

for f1 of type V1 H1 J1 or V1 J1 ,
f2 of type V2 H2 W2 J2 or V2 H2 W2 or
f2 of type V2 W2 J2 or V2 W2 , and
f3 of type H3 W3 J3 or H3 W3 .

7.8.

Invertibility in GDS

157

The equations ( 1) and ( 2), or alternatively ( 1) and ( 2), are

analogous to the two associativity equations for the cut operation one nds
in multicategories (see [85] and [88], Section 3).
To derive these equations for GDS is a rather straightforward, though
pretty lengthy, exercise. We always derive the most complex case, with
all possible parameters present (for ( 1) this means that fi is of type
Hi Wi Ji ), and the remaining cases are obtained by simplifying this most
complex case. For example, to derive the most complex case of ( 1) for
GDS we use essentially (dL ), (dL nat) and Net Coherence.
Let let (X) be the set of letters occurring in the form set X. It is clear
that we have the following.
Balance Remark. For every arrow f : X Y of GDS, we have let (X) =
let (Y ).
A pair of form sets (X1 . . . Xn , Y1 . . . Yn ), where n 2, is
splittable when let (Xi ) = let (Yi ) for every i {1, . . . , n}. A sequence
of form sets X1 , . . . , Xn , Y1 , . . . , Yn is a total split of the pair of form sets
(X1 . . . Xn , Y1 . . . Yn ) when let (Xi ) = let (Yi ) and none of the pairs
(Xi , Yi ) is splittable. For every splittable pair of form sets there is a total
split.
We say that an arrow f : X Y of GDS is splittable when its type
(X, Y ) is splittable, and we say that a total split of (X, Y ) is a total split
of f .
Splitting Remark. Take an arrow f of GDS of type
X X1 . . . Xn Z Y1 . . . Yn
or

X X1 . . . Xn Z R Y1 . . . Yn

and an arrow g of GDS of type

V V1 . . . Vm U W1 . . . Wm
or

V V1 . . . Vm U R W1 . . . Wm

with n + m 1 (if n = 0, then the subword X1 . . . Xn is

just omitted, and analogously in other cases).

158

CHAPTER 7.

DISSOCIATIVE CATEGORIES

Let Z,U (f, g) be splittable with the total split

X V, X1 , . . . , Xn , V1 , . . . , Vm , Y, Y1 , . . . , Yn , W1 , . . . , Wm
where Y is Z U or (Z U ) R, and R is R or R or R R .
If n 1, then f is splittable, and if m 1, then g is splittable.
Here (X, Z) or (X, Z R ) may be splittable, and hence the
forms of the type of f above do not show the total split tied to
this type. If (X, Z R ) is splittable and S1 , . . . , Sk , T1 , . . . , Tk
is its total split, then for every j {1, . . . , k} we have let (Z)
let (Tj ) = . (Otherwise, the total split of Z,U (f, g) mentioned
above would not be a total split.) We have an analogous remark
for (V, U ), (V, U R ) and g.
An analogous remark holds for Z,U (f, g).
It follows from the Splitting Remarks that, for {, }, if Z,U (f, g)
is splittable, then f or g is splittable. Since 1p is not splittable, we can
easily conclude the following with the help of the Cut-Elimination Theorem
of the preceding section.
Splitting Corollary. No arrow of GDS is splittable.
This corollary is related to the connectedness condition of proof nets (see
[33]).
Next we prove the following lemma for GDS.
Invertibility Lemma for . (i) If f : U1 U2 (X1 X2 ) Z1 Z2 is a
cut-free Gentzen term such that let (Ui ) = let (Xi ) let (Zi ) for i {1, 2},
then there are two cut-free Gentzen terms f1 : U1 X1 Z1 and f2 : U2
X2 Z2 such that f = X1 ,X2 (f1 , f2 ).
(ii) If f : U1 U2 (X1 X2 ) Z1 is a cut-free Gentzen term such that
let (U1 ) = let (X1 ) let (Z1 ) and let (U2 ) = let (X2 ), then there are two
cut-free Gentzen terms f1 : U1 X1 Z1 and f2 : U2 X2 such that
f = X1 ,X2 (f1 , f2 ).
(iii) If f : U1 U2 X1 X2 is a cut-free Gentzen term such that let (Ui ) =
let (Xi ) for i {1, 2}, then there are two cut-free Gentzen terms f1 : U1
X1 and f2 : U2 X2 such that f = X1 ,X2 (f1 , f2 ).

7.8.

Invertibility in GDS

159

Proof. We proceed by induction on the length of the cut-free Gentzen

term f . If f is 1p , the lemma holds trivially, since f cannot be of the
required type.
Suppose next that f is Y a ,Y b (f a , f b ) for f a : W a Ra and f b : W b
b
R . Then, under the assumptions of (i), we have two cases:
( i I)
( i II)

X1 X2 is Y a Y b ,
X1 X2 is dierent from Y a Y b .

We deal rst with ( i I).

The cases where let (X1 ) let (X2 ) let (W a ) or let (X1 ) let (X2 )
let (W b ) are impossible.
If let (X1 ) let (W a ) and let (X2 ) let (W b ), then we must have
let (Z1 ) let (W a ) and let (Z2 ) let (W b ). All the other cases are excluded
by the Splitting Corollary. For example, if let (Z1 ) let (Z2 ) let (W a ),
then W a must be U1 U2a , Ra must be X1 Z1 Z2 , W b must be U2b ,
and Rb must be X2 , where U2 is U2a U2b . Then, since let (U2a ) = let (Z2 ),
the arrows f a and f would be splittable, which contradicts the Splitting
Corollary. In the only possible case mentioned above, we take f a for f1 and
f b for f2 .
The case where let (X1 ) W b and let (X2 ) W a is analogous to the
case just settled.
Let (X, Y, Z) abbreviate the conjunction of the following conditions:
let (X) let (Y ) let (Z),
let (X) let (Y ) = ,
let (X) let (Z) = .
If let (X1 ) let (W a ) and (X2 , W a , W b ), then we have as possible cases
let (Z1 ) let (W a ) together with
(1)
(2)
(3)

let (Z2 ) let (W a ), or

let (Z2 ) let (W b ), or
(Z2 , W a , W b ).

The remaining cases are excluded by the Splitting Corollary.

We deal rst with (3). Then f a is of the type U1 U2a (X1 X2a )
Z1 Z2a , while f b is of the type U2b X2b Z2b , where U2a U2b is U2 , X2a X2b

160

CHAPTER 7.

DISSOCIATIVE CATEGORIES

is X2 and Z2a Z2b is Z2 . By the induction hypothesis, f a = X1 ,X2a (f1a , f2a )

for f1a : U1 X1 Z1 and f2a : U2a X2a Z2a . Then by the equation ( 1)
we have
X1 X2a ,X2b (X1 ,X2a (f1a , f2a ), f b ) = X1 ,X2 (f1a , X2a ,X2b (f2a , f b )),
and we take that f1 is f1a , while f2 is X2a ,X2b (f2a , f b ). In cases (1) and
(2) we proceed analogously, using again ( 1) (less complex cases of this
equation, with less parameters).
The three cases where we have let (X1 ) let (W b ) and (X2 , W a , W b ),
or let (X2 ) let (W a ) and (X1 , W a , W b ), or let (X2 ) let (W b ) and
(X1 , W a , W b ), are all settled analogously to the case we have just dealt
with.
The remaining case of ( i I) is when (X1 , W a , W b ) and (X2 , W a , W b ).
Then either let (Zi ) let (W a ), or let (Zi ) let (W b ), or (Zi , W a , W b ),
and we always apply the induction hypothesis and equation ( 1) three
times; namely, we use the equation
X1a X2a ,X1b X2b (X1a ,X2a (f1a , f2a ), X1b ,X2b (f1b , f2b )) =
X1a X1b ,X2a X2b (X1a ,X1b (f1a , f1b ), X2a ,X2b (f2a , f2b )).
Under the assumption ( i II), we have the cases
( i II.1)
( i II.2)

Z1 is Z1 (Y a Y b ),
Z1 is Y a Y b ,

and two more cases obtained by replacing the index 1 in Z1 and Z1 by 2 .

For ( i II.1), we have as possible cases let (X1 X2 ) let (Z2 ) let (W a )
together with
(1) let (Z1 ) let (W a ), or
(2) let (Z1 ) let (W b ), or
(3) (Z1 , W a , W b ),
and three more cases with let (X1 X2 ) let (Z2 ) let (W b ). All the
remaining cases are excluded by the Splitting Corollary.
We deal rst with (3). Then f a is of the type U1a U2 Y a (X1
X2 ) Z1a Z2 , while f b is of the type U1b Y b Z1b , where U1a U1b is U1
and Z1a Z1b is Z1 . By the induction hypothesis, f a = X1 ,X2 (f1a , f2a ) for

7.8.

Invertibility in GDS

161

f1a : U1a Y a X1 Z1a and f2a : U2 X2 Z2 . Then, by the equation

( 2), we have
Y a ,Y b (X1 ,X2 (f1a , f2a ), f b ) = X1 ,X2 (Y a ,Y b (f1a , f b ), f2a )),
and we take that f1 is Y a ,Y b (f1a , f b ), while f2 is f2b . In cases (1) and (2),
and cases obtained by interchanging a and b, we proceed analogously.
For ( i II.2), we have as possible cases let (X1 X2 )let (Z2 ) let (W a )
and let (X1 X2 ) let (Z2 ) let (W b ), for which we apply again the induction hypothesis and the equation ( 2). All the remaining cases are
excluded by the Splitting Corollary.
We proceed analogously when we have ( i II) and Z2 is either Z2
(Y a Y b ) or Y a Y b . With that we have settled ( i II), and also (i).
Under the assumptions of (ii), we have again two cases:
( ii I)
( ii II)

X1 X2 is Y a Y b ,
X1 X2 is dierent from Y a Y b .

We deal with these cases as above, with simplications in cases already

considered.
Under the assumptions of (iii), we must have that X1 X2 is Y a Y b ,
and we have cases simplifying again cases already considered. With that
we have nished dealing with the assumption that f = Y a ,Y b (f a , f b ).
Suppose now f is Y a ,Y b (f a , f b ) for f a : W a Ra and f b : W b Rb .
Then under the assumptions of (i) we have the cases:
( i 1) U1 is U1 (Y a Y b ),
( i 2) U1 is Y a Y b ,
and two more cases with the index 1 of U1 and U1 replaced by 2 .
For ( i 1), we have as possible cases let (X1 X2 ) let (Z2 ) let (W a )
together with
() let (Z1 ) let (W b ), or
() (Z1 , W a , W b ),
and together with
(1) let (U1 ) let (W a ), or
(2) let (U1 ) let (W b ), or
(3) (U1 , W a , W b ),

162

CHAPTER 7.

DISSOCIATIVE CATEGORIES

and six more analogous cases with let (X1 X2 )let (Z2 ) let (W b ) together
with
let (Z1 ) let (W a ), or
(Z1 , W a , W b ).
All the remaining cases are excluded by the Splitting Corollary, or because
let (Y a Y b ) let (X1 X2 ) let (Z1 ).
We deal rst with () together with (3). Then f a is of the type Y a
U1a U2 (X1 X2 ) Z1a Z2 , while f b is of the type Y b U1b Z1b ,
where U1a U1b is U1 and Z1a Z1b is Z1 . By the induction hypothesis,
f a = X1 ,X2 (f1a , f2a ) for f1a : Y a U1a X1 Z1a and f2a : U2 X2 Z2 .
Then, by the equation (), we have
Y a ,Y b (f b , X1 ,X2 (f1a , f2a )) = X1 ,X2 (Y b ,Y a (f b , f1a ), f2a ),
and we take that f1 is Y b ,Y a (f b , f1a ), while f2 is f2a . In all the remaining
cases, we proceed analogously, as well as in ( i 2). This settles (i).
Under the assumptions of (ii), we have cases analogous to those already
treated with Z2 omitted. So we apply again the equation ().
The assumptions of (iii) are excluded if f = Y a ,Y b (f a , f b ).

We prove analogously the following lemma for GDS.

Invertibility Lemma for . (i) If f : (X1 X2 ) Z1 Z2 U1 U2
is a cut-free Gentzen term such that let (Ui ) = let (Xi ) let (Zi ) for i
{1, 2}, then there are two cut-free Gentzen terms f1 : X1 Z1 U1 and
f2 : X2 Z2 U2 such that f = X1 ,X2 (f1 , f2 ).
(ii) If f : (X1 X2 ) Z1 U1 U2 is a cut-free Gentzen term such that
let (U1 ) = let (X1 ) let (Z1 ) and let (U2 ) = let (X2 ), then there are two
cut-free Gentzen terms f1 : X1 Z1 U1 and f2 : X2 U2 such that
f = X1 ,X2 (f1 , f2 ).
(iii) If f : X1 X2 U1 U2 is a cut-free Gentzen term such that let (Ui ) =
let (Xi ) for i {1, 2}, then there are two cut-free Gentzen terms f1 : X1
U1 and f2 : X2 U2 such that f = X1 ,X2 (f1 , f2 ).

7.9.

Linearly distributive categories

163

Let the quantity of letters in an arrow f : X Y of GDS be the

cardinality of let (X) (which is equal to the cardinality of let (Y )). Then we
can prove the following theorem for GDS.
Cut-Free Preordering. For every pair of cut-free Gentzen terms f1 , f2 :
X Y we have f1 = f2 .
Proof. We proceed by induction on the quantity of letters in f1 (which is
equal to the quantity of letters in f2 ). If n = 1, then f1 = f2 = 1p .
Suppose n > 1. If f1 is Z1 ,Z2 (f1 , f1 ), then by the Invertibility Lemma
for we have that f2 = Z1 ,Z2 (f2 , f2 ) for f2 and f2 of the same types as
f1 and f1 respectively. By the induction hypothesis, f1 = f2 and f1 = f2 ,

and hence f1 = f2 . We proceed analogously if f1 is Z1 ,Z2 (f1 , f1 ).

As a corollary of the Cut-Elimination Theorem and of Cut-Free Preordering, we obtain that GDS is a preorder, which, under the assumption
Gf = Gg, implies the assertion (div ) of 7.6. This proves Symmetric Net
Coherence.
Net Coherence of 7.3 could also have been proved via a Cut-Elimination
Theorem and Cut-Free Preordering. Strictication, however, would be in
the associative structure only, and not in the symmetric associative structure.
The category DS corresponds to the multiplicative conjunction-disjunction fragment of linear logic, for which proof nets were developed (see [63]
and [33]). Proof nets, however, serve mainly to solve the theoremhood
problem, while coherence in our sense is maybe implicitly presupposed with
them. The theoremhood problem for DS can also be solved via our results
for GDS in this and in the preceding section, based on cut elimination,
and we do not nd this solution in the style of Gentzen more complicated
than the solution provided by proof nets.

7.9.

Linearly distributive categories

To obtain the natural logical category DA, , we have that the logical
system C(DA, ) is in L,,, , with the transformations included in 1,
b, - and d. The specic equations of E(DA, ) are obtained by taking

164

CHAPTER 7.

DISSOCIATIVE CATEGORIES

the union of those of E(DA) and E(A, ) plus

( dL )

( dL )

( dR )

dL
,B,C = ( B 1C ) BC ,

dL
A,B, = AB (1A B ),

dR
C,B, = (1C B ) CB ,

( dR ) dR
,B,A = BA ( B 1A ).

Natural DA, -categories are called linearly distributive categories in

[25] (the original name from [22] is weakly distributive categories). According to our nomenclature, they could be called dissociative bimonoidal
categories. All of the specic equations above may be found in [22] (Sec
tion 2.1). (These equations should be compared with the equations (d k)

and (d k) of 11.1.)
We have still a functor G from DA, to Rel, but according to [11]
(Section 4.2, pp. 275-278), for

( 1 ) dL
A =df
A
,,A ,
A

A =df dR
,,A ( 1A ) A ,

the equations
A (1 A ) = 1((A)) ,
(1 A ) A = 1((A))
do not hold in DA, , although, when f and g are respectively the lefthand side and right-hand side of one of these equations, we have Gf = Gg
in Rel. So G is not faithful, and coherence fails. The faithfulness of G in
this case would yield preordering, and DA, is not a preorder.
Note that in DA, we have
(1 A ) A = 1A ,
A (1 A ) = 1A ,
which are the triangular equations of an adjunction (for the notion of adjunction see [100], Chapter 4; the functor is left-adjoint to the functor
). What fails is the isomorphism between ( ( A)) and A,
and between ( ( A)) and A ([119] deals with a related problem in symmetric monoidal closed categories). We do not know what other

7.9.

Linearly distributive categories

165

equations, if any, besides those that deliver these isomorphisms, should be

added to the axioms of DA, in order to obtain coherence.
A sort of coherence for linearly distributive categories (symmetric and
not symmetric, without the isomorphisms above) in the context of proof
nets has been investigated in a number of papers (see [22], [11], [23] and
[115]). Coherence in this sense is not quite foreign to what we mean by
coherence, but it is not the same thing. The investigations of [22] appeal
to a connection with the polycategories of [126].

Chapter 8
Mix Categories
In this chapter, we consider categories having what linear logicians call
mixnamely, a natural transformation between the two bifunctors of the
double monoidal structure. The double monoidal structure has or does not
have associativity, symmetry and dissociativity. We prove coherence for
such categories that lack unit objects. The mix principle is an important
addition to Gentzens plural sequent formulation of classical logic, and this
is why we pay particular attention to it.
Our proofs are variations on the cut-elimination theme, and on the techniques of the preceding chapters. There are proofs based on compositionfree languages for our categories, and a proof based on an extension of the
cut-elimination procedure of the preceding chapter.

8.1.

Coherence of mix and mix-dissociative categories

To obtain the natural logical category MI, we have that the logical system
C(MI) is in L, , with the transformations being 1 and m. The equations
nat
E(MI) are just those of EC(MI)
(see 4.1). We call natural MI-categories
mix categories.
A logical principle called mix amounting to mA,B : A B A B
was considered in [63] (Section V.4), [33] (Section 3.3), [56], [10] and [23].
Gentzen called Mischung in Germanwhich is usually translated as mix
a rule that generalizes the cut rule of sequent systems; an instance of Mischung is
167

168

CHAPTER 8.

1 1 ,

MIX CATEGORIES

, 2 2

1 , 2 1 , 2
where is a nonempty sequence of occurrences of the same formula (see
[60], Section III.3.1). The mix principle of mA,B is related to the Mischung
rule above where is the empty sequence. (Gentzen did not envisage
this mix principle because he could prove the conclusion from one of the
premises with the help of the structural rules of thinning on the left and
thinning on the right.)
In C(MI) we dene as follows the binary total operation 3 on arrow
terms:
f:AD

g: B E

f 3 g =df (f g) mA,B : A B D E
for which in MI we have the equations
(3) (g1 3 g2 ) (f1 f2 ) = (g1 g2 ) (f1 3 f2 ) = (g1 f1 ) 3 (g2 f2 ).
From the equation (m nat) of 2.7, in MI we obtain immediately f 3 g =
mD,E (f g), which gives an alternative denition of 3.
A syntactical system C(3MI) synonymous with C(MI) is obtained by
taking as objects the formulae of L, , as primitive arrow terms identity
arrow terms only, and as operations on arrow terms , , and 3. The
equational system E(3MI) is obtained by assuming the categorial and bifunctorial equations for and , and the equations (3). The category
3MI is C(3MI)/E(3MI).
With the denition
mA,B =df 1A 3 1B
in C(MI), we obtain two obvious functors from C(MI) to C(3MI), and vice
versa, which preserve the respective structures on the nose (see 2.8), and
these functors induce functors that give the isomorphism of MI and 3MI.
This means that C(MI) and C(3MI) are synonymous (see the end of 2.4
for the notion of synonymity of syntactical systems). Note that ocially

8.2.

Coherence of mix-biassociative categories

169

C(3MI) is not a logical system, because 3 is not of the kind: it is not a

bifunctor in MI.
We can prove the following proposition for 3MI simply by relying on
the equations (3) and bifunctorial equations.
Composition Elimination. For every arrow term h there is a compositionfree arrow term h such that h = h .
This result is a simple kind of cut-elimination result, such as we had in
7.7.
The composition-free arrow term h can be put into a unique normal
form by applying the bifunctorial equations ( 1) so that every arrow term
1A in h has a letter for A. From that, we obtain immediately that two
dierent arrow terms in normal form must be of dierent types. So 3MI
is a preorder, from which we conclude the following.
Mix Coherence. The category MI is a preorder.
To obtain the natural logical category MDI, we have that the logical
system C(MDI) is in L, , with the transformations included in 1, d
nat
and m. The equations E(MDI) are just those of EC(MDI)
. We call natural
MDI-categories mix-dissociative categories.
To prove that MDI is a preorder, we proceed as in 7.1 for DI by mod

ifying the relations SA

. In xSA
y we have as before that y is an occurrence
of in A, while x can be an occurrence of , or of a letter. With the
help of this relation we proceed analogously to what we had in 7.1. So we
have the following.
Mix-Dissociative Coherence. The category MDI is a preorder.

8.2.

Coherence of mix-biassociative categories

To obtain the natural logical category MA, we have that the logical system
C(MA) is in L, , with the transformations included in 1, b and m. The
specic equations of E(MA) are those of E(A) plus
(bm)

(mA,B 1C ) mAB,C b
A,B,C = b A,B,C mA,BC (1A mB,C ).

170

CHAPTER 8.

MIX CATEGORIES

We call natural MA-categories mix-biassociative categories.

With 3 primitive in a syntactical system synonymous with C(MA), the
equation (bm) is replaced by

((f 3 g) 3 h) b
A,B,C = b D,E,F (f 3 (g 3 h)).

For G being A and C/E being MA, we have that the conditions (IVC)
and (IV G) of 3.1 are satised, and G is moreover a preorder. Thus we
can apply the Direct-Strictication Theorem of 3.2 to obtain a category
CG /EG , which we will call MAst , or 3MAst when 3 is primitive. The
categories MAst and 3MAst are isomorphic, as MI and 3MI are (see
the preceding section). We call CG here C(MAst ), and C(3MAst ) is the
synonymous syntactical system where 3 is primitive.
We can easily prove the Composition Elimination proposition of the
preceding section for MAst . Here is a sketch of how we proceed. By the
Development Lemma of 2.7, there is for every arrow term of C(MA) a

developed arrow term. For a b-term f and a b-term or m-term g we have

in MA that f g = g f for a b-term f and a b-term or m-term g . So

we may say that b-terms can be moved to the right. Analogously, b-terms
can be moved to the left. Eventually, we obtain an arrow term of the form

f1 f2 f3 1A , where in f1 there are no b-terms and m-terms, in f2 there

are no b-terms and b-terms and in f3 there are no b-terms and m-terms.
Then it is enough to apply the Composition Elimination for 3MI of the
preceding section to the arrow term of C(3MAst ) corresponding to f2 to
obtain Composition Elimination for 3MAst .
A composition-free arrow term of C(3MAst ) is atomized when for every
occurrence of 1|[A]| in it we have that A is a letter. We will write 1p instead
of 1|[p]| .
For an atomized composition-free arrow term f of C(3MAst ) let w(f )
be the word obtained from f by deleting parentheses. We already dened
w(X) for a form sequence X in 7.3; it is, analogously, the word obtained
from X by deleting parentheses.
To every pair of parentheses in a form sequence of letters X in natural
notation (see 6.2 and 7.3), we can associate a pair of occurrences of letters
(x, y) in X, where x is the rst occurrence of a letter on the right-hand side
of the left parenthesis and y is the rst occurrence of a letter on the left-hand

8.2.

Coherence of mix-biassociative categories

171

side of the right parenthesis. For example, in p (p (q r)), to the outer

pair of parentheses written down we associate (x, y) where x is the second p
counting from the left and y is r. We suppose that atomized compositionfree arrow terms of C(3MAst ) are written in natural notation (see 6.2),
and we associate analogously pairs (1x , 1y ) to pairs of parentheses in such
arrow terms. (Such arrow terms correspond to a kind of form sequence of
three colours: , and 3.)
For every atomized composition-free arrow term f : X Y of C(3MAst )
there are obvious bijections between occurrences of the same letters in X,
Y and f , or in w(X), w(Y ) and w(f ). We say that such occurrences
correspond obviously to each other. For example, in (1p1 3 1p2 ) 1p3 :
p1 p2 p3 (p1 p2 ) p3 the three occurrences of pi , for i {1, 2, 3},
correspond obviously to each other.
For the proof of the following proposition we rely on the notion of immediate scope of 2.1.
Proposition 1. Let {, , 3}, and let the atomized compositionfree arrow term f : X Y of C(3MAst ), where f , X and Y are written
in natural notation, have a subterm (f1 . . . fn ), for n 2, such that
(1x1 , 1x2 ) is associated to the outermost parentheses of this subterm. Then
there is a pair of parentheses in at least one of X and Y such that (y1 , y2 )
is associated to this pair of parentheses, and, for i {1, 2}, the occurrence
of letter yi corresponds obviously to the occurrence of the same letter xi in
1x i .
Proof. If f is (f1 . . . fn ), then the assertion is trivial. (We usually omit,
however, such outermost parentheses.) Suppose then that (f1 . . . fn ) is
a proper subterm of f .
If is and is within the immediate scope of , then we have both in
X and in Y the required pair of parentheses. If is and is within the
immediate scope of 3, then we have in Y the required pair of parentheses.
If is 3 and is within the immediate scope of , then we have in Y the
required pair of parentheses. If is 3 and is within the immediate scope
of , then we have in X the required pair of parentheses.
It remains to consider two cases where is , which are dual to the two

172

CHAPTER 8.

MIX CATEGORIES

cases above where it is .

Proposition 2. Let f : X Y be an atomized composition-free arrow

term of C(3MAst ), and let f , X and Y be written in natural notation.
Then for every pair of parentheses in the form sequences X or Y to which
(y1 , y2 ) is associated there is a pair of parentheses in f to which (1x1 , 1x2 )
is associated such that, for i {1, 2}, the occurrence of letter yi corresponds
obviously to the occurrence of the same letter xi in 1xi .
Proof. If the pair of parentheses selected in X or Y is outermost (which we
usually do not write), then the assertion is trivial. If the pair of parentheses
selected is in X and belongs to within the immediate scope of , then in
f we must have the required pair of parentheses, which belongs to or 3
within the immediate scope of . If the pair of parentheses selected is in Y
and belongs to within the immediate scope of , then in f we must have
the required pair of parentheses, which belongs to within the immediate
scope of or 3. The cases where is within the immediate scope of are
dual.

Building on Propositions 1 and 2, we can obtain a criterion for the

existence of an arrow of 3MAst of a given type X Y , which solves the
theoremhood problem for 3MAst (see 1.1). Let
(1)

w(X) and w(Y ) coincide save that in w(Y ) we can have an occurrence of at a place where in w(X) we have an occurrence
of .

Let u be obtained from w(X) and w(Y ) by

(2.1) writing 3 at the places where w(X) and w(Y ) dier,
(2.2) adding the parentheses of both X and Y in an obvious manner
(here, two pairs of parentheses, one in X and the other in Y ,
associated to pairs of occurrences of letters that correspond obviously to each other yield a single pair of parentheses added to
u),
(2.3) replacing every occurrence of a letter p by 1p .
Then there is an arrow of 3MAst of type X Y i (1) is fullled and u

8.3.

Coherence of mix-net categories

173

is an arrow term of C(3MAst ) in natural notation. If u is such an arrow

term, then it stands for the required arrow of 3MAst of type X Y . This
will yield a criterion for the existence of arrows of a given type in MA.
We need the following proposition to prove coherence for MA.
Proposition 3. If f1 : X1 Y1 and f2 : X2 Y2 are dierent atomized
composition-free arrow terms of C(3MAst ), then X1 is dierent from X2
or Y1 is dierent from Y2 .
Proof. If w(f1 ) is dierent from w(f2 ), then it is clear that w(X1 ) is
dierent from w(X2 ) or w(Y1 ) is dierent from w(Y2 ). If w(f1 ) coincides
with w(f2 ), but f1 and f2 are dierent arrow terms, then f1 and f2 must
dier with respect to parentheses. In that case, Propositions 1 and 2 yield
the assertion.

From Composition Elimination and Proposition 3 we infer that 3MAst

is a preorder. So we have the following.
Mix-Biassociative Coherence. The category MA is a preorder.

8.3.

Coherence of mix-net categories

To obtain the natural logical category MDA, we have that the logical
system C(MDA) is in L, , with the transformations included in 1, b, d
and m. The specic equations of E(MDA) are those of E(DA) plus

(b mL)

(b mL)

(b mR)

(b mR)

mAB,C b
A,B,C = dA,B,C (1A mB,C ),

L
b
A,B,C mA,BC = (mA,B 1C ) dA,B,C ,

mC,BA b
C,B,A = dC,B,A (mC,B 1A ),

R
b
C,B,A mCB,A = (1C mB,A ) dC,B,A .

We call natural MDA-categories mix-net categories.

The specic equation (bm) of E(MA) is derived as follows for MDA:

174

CHAPTER 8.

MIX CATEGORIES

(mA,B 1C ) mAB,C b
A,B,C

= (mA,B 1C ) dL
A,B,C (1A mB,C ), by ( b mL),

= b
A,B,C mA,BC (1A mB,C ), by ( b mL).

Alternatively, we could have used (b mR) and (b mR).

For G being A and C/E being MDA, we have that the conditions (IV C)
and (IVG) of 3.1 are satised, and G is moreover a preorder. Thus we can
apply the Direct-Strictication Theorem of 3.2 to obtain a category CG /EG ,
which we will call MDAst . We call CG here C(MDAst ).
For every object X of MDAst , i.e. for every form sequence of letters

X in natural notation (see 6.2 and 7.3), we introduce a relation RX

between the set of occurrences of in X and the set of occurrences of ,

and letters in X. We dene xRX

y as xRX y in 7.3 save that y need not
be an occurrence of , but may be an occurrence of , or of a letter (cf.

the version of SA
in 8.1). More precisely, y is an occurrence of , or
of a letter on the right-hand side of l(x) and on the left-hand side of r(x);
here y can also be l(x) or r(x). We dene the occurrences of letters l(x)
and r(x) exactly as before (see 7.3).
Let v(X) be the word obtained from a form sequence X by deleting
every parenthesis and replacing every occurrence of or in X by a single
arbitrary new symbol . When for the form sequences X and Y we have
that v(X) and v(Y ) coincide, we say that X and Y are MDA-comparable.

It is clear that RX
gives rise to a relation Rv X on occurrences of symbols
in v(X) such that we have x Rv X y when x is the occurrence of in v(X)
corresponding to an occurrence x of in X, while y is an occurrence of
or of a letter p corresponding to an occurrence y of , or p in X, and we

have x RX
y.
Then it can be checked that for every arrow term f : X Y of C(MDAst )
the form sequences X and Y are MDA-comparable and Rv Y Rv X . Moreover, if dL , dR or m occurs in f , then Rv Y is a proper subset of Rv X ;
otherwise, Rv X = Rv Y . For example, with mp,q : p q p q we have
v(p q) = v(p q) = pq, while Rv pq = {(, p), (, q)} and Rv pq = .
A place in X is a subword of v(X). We dene when subwords of MDA-

8.3.

Coherence of mix-net categories

175

comparable form sequences are at the same place as in 7.3 and 2.1.
We have a Remark analogous to that of 7.3 with R replaced by R , and
a lemma analogous to the Extraction Lemma of 7.3 with DAst replaced
by MDAst and () that reads:
()

for every occurrence x of in Xi and every occurrence y of ,

or of a letter in Xi , where i {1, 2}, if x is an occurrence
of in X at the same place where X1 z X2 has x, while y is an
occurrence of , or of a letter in X at the same place where

X1 z X2 has y, and x RX
y , then xRX
y.
i

A more precise formulation of () in the Extraction Lemma of 7.3 would

be analogous to this version of (), but there we identied x and y with x
and y . Note that here X and X1 z X2 are comparable and not only MDAcomparable. This means that y and y are occurrences of the same symbol.
The proofs of the Extraction Lemma of 7.3 and of its analogue for MDAst
do not involve the transformation m. They are based on considerations
concerning l(x) and r(x), which are the same both for the relation R and
for the relation R .
With the help of this analogue of the Extraction Lemma, we can prove
the following analogue of the Theoremhood Proposition of 7.3.
Theoremhood Proposition. There is an arrow term f : X Y of
C(MDAst ) i X and Y are MDA-comparable form sequences and Rv Y
Rv X .
Proof. We enlarge the proof of the Theoremhood Proposition of 7.3. If
n > 1 and Y is Y1 z Y2 for z an occurrence of , then there is no guarantee
that X has an occurrence of at the same place. If it has it, then we
proceed as before, applying the analogue of the Extraction Lemma. If, on
the other hand, X has an occurrence u of at that place, then we rst take
an arrow term h : X X of C(MDAst ) made of m, 1 and the operations
and on arrow terms such that X diers from X just by having an
occurrence of instead of u. It is clear that if we exclude from Rv X all
those pairs (u , y ), where u corresponds to u in v(X), then we obtain Rv X .
We continue again as in the proof of the Theoremhood Proposition of 7.3,
and f is (f1 f2 ) g h.

176

CHAPTER 8.

MIX CATEGORIES

For a form sequence X, let d(X) be now the cardinality of RX

. We can
prove the following.

Mix-Net Coherence. The category MDA is a preorder.

Proof. We proceed as in the proof of Net Coherence in 7.3 with the
following additional cases.
(Lm)

The head of f1 is dL
E,F,G and the head of g1 is mH,I . Due to the

presence of (b mL), we may assume that I is not of the form (X1 . . . Xn , ),

and so I cannot be F G. It remains to consider subcases analogous
to (LL1) and (LL2), which are settled with the help of bifunctorial and
naturality equations.
(Rm)

The head of f1 is dR
G,F,E and the head of g1 is mI,H . Here we

invoke (b mR), and deal as in (Lm).

(mm)

The head of f1 is mE,F and the head of g1 is mH,I . Due to the

presence of (b mL) and (b mR), we may assume that E, F , H and I are

not of the form (X1 . . . Xn , ). So, with the previous assumption based

on (b mL) and (b mR), they must be occurrences of letters. We have the

following subcase.
(mm1) The occurrence of letter F coincides with H. Then by (bm),

(m nat), (b mL) and (b mR) we have

L

(1E mF,I ) dR
E,F,I (mE,F 1I ) = (mE,F 1I ) dE,F,I (1E mF,I ).

We continue reasoning by applying the Theoremhood Proposition of this

section. The remaining subcases are settled with the help of bifunctorial
and naturality equations.

8.4.

Coherence of mix-symmetric net categories

To obtain the natural logical category MDS, we have that the logical system C(MDS) is in L, , with the transformations included in 1, b, c,
d and m. The specic equations of E(MDS) are obtained by taking the
union of those of E(MDA) and E(DS) plus
(cm)

mB,A c A,B = c B,A mA,B .

8.4.

Coherence of mix-symmetric net categories

177

We call natural MDS-categories mix-symmetric net categories.

With 3 as in 8.1, the equation (cm) amounts to the equation

(g 3 f ) c A,B = c E,D (f 3 g).

In the arrow terms of C(MDS) we write d instead of dL , as we did for
C(DS), and we take dR as dened by the equation (dR c) of 7.6. Among

the specic equations (b mL), (b mL), (b mR) and (b mR) of E(MDA) (see

8.3), it is enough to keep (b mL) and (b mL); the equations (b mR) and

(b mR) are derivable.

We build the syntactical category GMDS in the syntactical system
C(GMDS) out of MDS as we built GDS out of DS in 7.6-7. The only
dierence is that we replace everywhere DS by MDS. As before, it is
enough to prove that GMDS is a preorder in order to infer the following.
Mix-Symmetric Net Coherence. The functor G from MDS to Rel is
faithful.
We dene the Gentzen terms for arrows of GMDS as for GDS, in 7.7,
with the following additional Gentzen operation:
f:U Z

g: Y W

mix (f, g) =dn (f g) mU,Y : U Y Z W

An alternative notation for mix (f, g) is f 3 g, which we used in 8.1-2,
but mix (f, g) is handier in the present context. Note that due to (cm) we
have mix (f, g) = mix (g, f ) in GMDS. So we will consider the terms on
the two sides of this equation as the same Gentzen term.
Then by enlarging the proof of the Gentzenization Lemma of 7.7 we
can prove that every arrow of GMDS is denoted by a Gentzen term. The
only addition is that mX,Y = mix (1X , 1Y ).
In the denition of depth of 7.7, we now have that can be also mix .
The notions of cut-free Gentzen term, topmost cut, rank and complexity
of a topmost cut are exactly as in 7.7. We can then prove as follows the
Cut-Elimination Theorem where GDS is replaced by GMDS.
Proof of the Cut-Elimination Theorem for GMDS. We proceed as
in the proof of 7.7 until we suppose that the complexity of the topmost cut

178

CHAPTER 8.

MIX CATEGORIES

cutX (f, g) is (m, n) with m, n 1. Then we have the following additional

case.
(6) The form set X is of colour (it may be of the form X1 X2 or p)
and f is mix (f1 , f2 ). So we have
f1 : U1 X V1

f2 : U2 V2

mix (f1 , f2 ) : U1 U2 X V1 V2

g: X Y W

cutX (mix (f1 , f2 ), g) : U1 U2 Y W V1 V2

Then consider the Gentzen term
f1 : U1 X V1

g: X Y W

cutX (f1 , g) : U1 Y W V1

f2 : U2 V2

mix (cutX (f1 , g), f2 ) : U1 U2 Y W V1 V2

We show that
() cutX (mix (f1 , f2 ), g) = mix (cutX (f1 , g), f2 ),
and the complexity (m, n1) of the topmost cut cutX (f1 , g) is lower than
(m, n), so that we may apply the induction hypothesis.
To show the equation (), we have to show
(g 1V1 V2 ) dY,X,V1 V2 ((mXV1 ,V2 (f1 f2 )) 1Y ) =
mW V1 ,V2 (((g 1V1 ) dY,X,V1 (f1 1Y )) f2 ).

To derive this equation for GMDS, we use essentially (b mL) of the preceding section and (cm) (for GMDS, the latter equation reads mX,Y = mY,X ).
We have to consider also subcases where we have f1 : U1 X or g : X
W . In all of them, () amounts to equations simpler than the equation
above, which all hold in GMDS.
If in cutX (f, g) with complexity (m, n) where m, n 1 the form set X
is of colour and g is mix (g1 , g2 ), then we have an additional case treated

dually to (6). In that case, the equation (b mL) of the preceding section,
together with (cm), plays an essential role.

To prove the analogue of the Invertibility Lemmata of 7.8 for GMDS

we need the following equations of GMDS:

8.4.

Coherence of mix-symmetric net categories

(mix)

179

X,Y (mix (f, h), g) = mix (X,Y (f, g), h)

for f of type U X V or U X, and

(mix)

X,Y (mix (f, h), g) = mix (X,Y (f, g), h)

for f of type X V U or X U . To derive these equations for GMDS

we use essentially (b mL) and (b mL) of the preceding section.

Note that (mix) and (mix) hold just with the types indicated for f .
For other types, analogous equations need not hold; take, for example,
pq,r (mix (1p , 1q ), 1r ) : p q r (p q) r.
We will need also the following equation:
(mix mix)

mix (mix (f, g), h) = mix (f, mix (g, h)),

which we derive for GMDS with the help of the equation (bm) of 8.2 (see
the preceding section).
We dene inductively the following abbreviation:
Mix (f ) =df f,
Mix (f1 , . . . , fk1 , fk ) =df mix (Mix (f1 , . . . , fk1 ), fk ), for k 2.
Due to the equation (mix mix), for k 2 we could have also
Mix (f1 , . . . , fk1 , fk ) =df mix (f1 , Mix (f2 , . . . , fk )).
With the help of (mix), (mix) and (mix mix) we can derive for GMDS
the following equations:
(Mix)

X,Y (Mix (f, f1 , . . . , fn ), Mix (g, g1 , . . . , gm )) =

Mix (X,Y (f, g), f1 , . . . , fn , g1 , . . . , gm )

where n, m 0 (if n = 0, then f1 , . . . , fn is just omitted, and analogously if

m = 0), while f is of type U X V or U X, and g is of type W Y S
or W Y . We have also the equation (Mix) where is replaced by ,
while f is of type X V U or X U , and g is of type Y S W or
Y W.

180

CHAPTER 8.

MIX CATEGORIES

The Splitting Remark of 7.8 holds for GMDS as it holds for GDS.
Note that the splittability of f need not entail the splittability of Z,U (f, g).
A counterexample is provided with mix (1p , 1q ) : p q p q, which is splittable, and pq,r (mix (1p , 1q ), 1r ), mentioned above, which is not splittable.
A cut-free Gentzen term f for arrows of GMDS such that every subterm
X,Y (f1 , f2 ) of f for {, } is not splittable is called split-normalized.
Let the quantity of letters for an arrow f : X Y of GMDS be the
cardinality of let (X), as for GDS in 7.8. We can prove the following.
Split-Normalization Lemma. For every Gentzen term h for arrows of
GMDS there is a split-normalized Gentzen term h such that h = h in
GMDS.
Proof. We proceed by induction on the quantity of letters in h. In the
basis of the induction, when this quantity is 1 and h = 1p : p p, the
lemma holds trivially.
Suppose that h is equal in GMDS to a cut-free Gentzen term Z,U (f, g)
with f and g of the same types as in the Splitting Remark of 7.8, and with
the total split mentioned there. Then we apply the induction hypothesis to
f and g to obtain f and g split-normalized. If Z,U (f , g ) is not splittable,
we are done. If Z,U (f , g ) is splittable, then we proceed as follows.
If n 1, then the cut-free Gentzen term f is splittable and can be
written in the form Mix (u, u1 , . . . , un ) with u of type X Z or X Z R
and ui : Xi Yi , for i {1, . . . , n} and all of u, u1 , . . . , un split-normalized.
To put f in this form, we may need to use (mix mix), and we also use the fact
that the subterms of a split-normalized Gentzen term are split-normalized.
Here ui is not splittable and is not of the form mix (u , u ), but u may
be splittable. If u is splittable, and is hence of the form mix (u , u ) with
u : X Y and u : X Y , then we must have let (Z) let (Y ) = and
let (Z) let (Y ) = . This follows from the Splitting Remark. If n = 0,
then f is not splittable, and can be written as Mix (f ).
We put g analogously in the form Mix (v, v1 , . . . , vm ), and we apply
(Mix) to obtain that Z,U (f, g) is equal to
Mix (Z,U (u, v), u1 , . . . , un , v1 , . . . , vm ),
which is split-normalized, because Z,U (u, v) is not splittable. If Z,U (u, v)

8.4.

Coherence of mix-symmetric net categories

181

were splittable, then by the Splitting Remark, either u, which is splitnormalized, would be of the form mix (u , u ) with u : X Y , u : X
Y and let (Z) let (Y ) = or let (Z) let (Y ) = , or v would be of such
a form, which is impossible, as we said above.
We proceed analogously if h is equal in GMDS to a cut-free Gentzen
term Z,U (f, g). If h is equal in GMDS to a cut-free Gentzen term
mix (f, g), then we just apply the induction hypothesis to f and g.

The Invertibility Lemma for is formulated for GMDS as in 7.8,

save that we assume for f not only that it is cut-free, but that it is splitnormalized too. The proof of this lemma proceeds, as before, by induction
on the length of f . In the induction step, when f is Y a ,Y b (f a , f b ) or
Y a ,Y b (f a , f b ), we work as in the proof in 7.8, save that when we eliminated some cases by appealing to the Split Corollary, now these cases are
eliminated by appealing to the fact that f is split-normalized, and hence
not splittable. It remains to consider the case where f is mix (f a , f b ). It
is, however, easy to conclude that we may apply the induction hypothesis
either to f a or to f b in order to obtain a cut-free Gentzen term X1 ,X2 (g, h)
equal to this term in GMDS. Then we apply (mix).
We proceed analogously to prove the Invertibility Lemma for for
GMDS. We also have for GMDS the following new lemma of the same
kind.
Invertibility Lemma for mix . If f : U1 U2 Z1 Z2 is a splitnormalized Gentzen term such that let (Ui ) = let (Zi ) for i {1, 2}, then
there are two split-normalized Gentzen terms f1 : U1 Z1 and f2 : U2 Z2
such that f = mix (f1 , f2 ).
Proof. We proceed by induction on the length of f . The basis of this
induction, when f is 1p , is trivial, as before.
For the induction step, because f is split-normalized, f must be of the
form mix (f a , f b ) for f a : W a Ra and f b : W b Rb . Then we have the
following cases:
(1)

W a is U1 and W b is U2 ,

(2)

W a is U1 U2a and W b is U2b , for U2 being U2a U2b ,

182
(3)

CHAPTER 8.

MIX CATEGORIES

W a is U1a U2a and W b is U1b U2b , for Ui being Uia Uib , where
i {1, 2},

and cases analogous to these. In case (1), we take f a for f1 and f b for
f2 . In case (2), we apply the induction hypothesis to f a to obtain a splitnormalized Gentzen term mix (f a , f a ) equal to f a in GMDS for f a : U1
Z1 and f a : U2a Z2a . Then we take f1 and f2 to be f a and mix (f a , f b ),
respectively, and we apply (mix mix). In case (3), we proceed analogously,
by applying (mix mix) three times (cf. the proof of the Invertibility Lemma
for in 7.8).

We can now prove Cut-Free Preordering of 7.8 for GMDS. The

proof is analogous to the proof of 7.8, with an additional case when f1
is mix (f1 , f1 ), which is settled with the help of the Invertibility Lemma for
mix . As a corollary of the Cut-Elimination Theorem for GMDS and of
Cut-Free Preordering for GMDS, we obtain that the category GMDS is
a preorder, which implies Mix-Symmetric Net Coherence.

8.5.

Coherence of mix-symmetric biassociative categories

To obtain the natural logical category MS, we have that the logical system
C(MS) is in L, , with the transformations included in 1, b, c and m. The
specic equations of E(MS) are obtained by taking the union of those of
E(MA) and E(S) plus the equation (cm). So MS is analogous to MDS, but
with d missing. We call natural MS-categories mix-symmetric biassociative
categories.
We can prove the following.
Mix-Symmetric Biassociative Coherence. The functor G from MS
to Rel is faithful.
We prove this assertion as for MDS via a Cut-Elimination Theorem and
Invertibility Lemmata. We keep in the proofs of the preceding section just
the easy cases.
We do not consider here something that would be called mix-bimonoidal
categories, symmetric or not symmetric, dissociative or not dissociative.

8.5.

Coherence of mix-symmetric biassociative categories

183

(Some kinds of mix-dissociative bimonoidal categories are considered in

[23], Sections 6-7.) We have left open in 7.9 the problem of what axioms
should be added to those of linearly distributive categories in order to obtain coherence with respect to Rel. Mix brings in its own problems in the
presence of and . These problems remain in the next two chapters, and
will disappear in Chapter 11 and later.

Chapter 9
Lattice Categories
This chapter is about coherence for categories with a double cartesian structure, i.e. with nite products and nite coproducts. We take this as a categorication of the notion of lattice. As before, we distinguish cases with and
without special objects, which are here the empty product and the empty
coproduct, i.e. the terminal and initial objects. The results presented are
taken over from [46], [48] and the revised version of [47].
We pay particular attention in this chapter to questions of maximality, i.e. to the impossibility of extending our axioms without collapse into
preorder, and hence triviality. This maximality is a kind of syntactical
completeness. (The sections on maximality, 9.3, 9.5 and 9.7, improve
upon results reported in [46], [48] and [47].)
The techniques of this chapter are partly based on a composition elimination for conjunctive logic, related to normalization in natural deduction,
and on a simple composition elimination for conjunctive-disjunctive logic,
implicit in Gentzens cut elimination.

9.1.

Coherence of semilattice categories

To obtain the natural logical category L, we have that the logical system

C(L) is in L , with the transformations included in 1, b, c and w-k. The

specic equations of E(L) are those of E(S) plus

185

186

CHAPTER 9.

b
A,A,A (1A w A ) w A = (w A 1A ) w A ,

(b w)

( c w)

c A,A wA = wA ,

LATTICE CATEGORIES

for c m
A,B,C,D =df b A,C,BD (1A ( b C,B,D ( c B,C 1D ) b B,C,D )) b A,B,CD

of type (A B) (C D) (A C) (B D),

(b c w) wAB = c m
A,A,B,B (w A w B ),

k 2A,B = k 1B,A

( c k)

2
(k 1A,B 1C ) b
A,B,C = 1A k B,C ,

(b k)

c A,B ,

k iA,A wA = 1A , for i {1, 2}.

(wk)

We call natural L-categories semilattice categories. Usually, they are

called categories with nite nonempty products. The objects of a semilattice category that is a partial order make a semilattice.

The equation (b c w) is the octagonal equation of [43] (Section 2) and [44]

(Section 1) (cf. [53], Proposition 3.29, p. 235, and, for c m , cf. [55], Section
III.3, p. 517).

The equation (b k) is related to the equation (b ) of 4.6. By using

essentially this equation, we derive for L the equations

(b k 1 )

(b k 2 )

k 1AB,C = (1A k 1B,C ) b

A,B,C ,

k 2C,BA = (k 2C,B 1A ) b
C,B,A ,

which are related respectively to the equations (b ) and (b ) of 4.6. To

derive (b k 1 ), we derive rst

k 1AB,C 1(AB)C = ((1A k 1B,C ) b

A,B,C ) 1(AB)C

with the help of (b 5) of 4.2 and other equations; since for f : E D we

have k 1D,E (f 1E ) wE = f , we have (b k 1 ). We proceed analogously for

(b k 2 ). Conversely, from (b k 1 ) and (b k 2 ) we can derive (b k) with the help

of (b c) from 5.1 and ( c k).

The equation ( c k), which is related to the equation ( c ) of 5.3, says

that k 1 and k 2 are interdenable. In the presence of ( c w) and ( c k), we can

9.1.

Coherence of semilattice categories

187

derive k 2A,A wA = 1A from k 1A,A wA = 1A , and vice versa, so that instead

of (wk) we could have assumed just one of these two equations.

We can also derive for L the equation

(wk k)

(k 1A,B k 2A,B ) wAB = 1AB .

For that we use (b c w), (b k), (b k 1 ), (b k 2 ), ( c k) and (wk).

For fi : C Ai , where i {1, 2}, we have in C(L) the denition

( , )

f1 , f2 =df (f1 f2 ) wC .

Then for f : A D and g : B E the following equations hold in L:

()

(b )

(b )

( c)

(w)

f g = f k 1A,B , g k 2A,B ,

1
2
b
A,B,C = 1A k B,C , k B,C

1
b
C,B,A = k C,B

k 2A,BC ,

k 1CB,A , k 2C,B 1A ,

c A,B = k 2A,B , k 1A,B ,

wA = 1A , 1A .

This shows that with the operation , on arrow terms primitive, together

with k i , where i {1, 2}, we could take the arrow terms on the left-hand
sides of these equations as dened. With these alternative primitives, all

the equations of E(L) can be derived from the categorial equations and the

following equations of E(L):

() k iA1 ,A2 f1 , f2 = fi ,
()

k 1A1 ,A2 h, k 2A1 ,A2 h = h,

for h : C A1 A2 (for these equations see [90], Section I.3). In other

words, we would obtain a syntactical system synonymous with C(L) (see

the end of 2.4 for this notion of synonymity).

Another alternative is to have k i and w primitive, together with the

operation on arrow terms. Then we can assume, besides categorial, bi

functorial and naturality equations, the equations (wk) and (wk k) in order

to obtain a logical system synonymous with C(L).

Synonymity with these alternative syntactical systems can be demonstrated directly, but this is a lengthy exercise. The coherence result for

188

CHAPTER 9.

LATTICE CATEGORIES

semilattice categories we are going to prove will easily yield this synonymity
in an indirect way.
We introduce next still another syntactical system synonymous with

C(L), which will be formulated in the style of Gentzen, and will enable us to
prove a composition-elimination result, i.e. a simple kind of cut-elimination

result, such as we had in 8.1. Let C(GL) be the syntactical system with
formulae of L as objects, with the primitive arrow terms being only identity arrow terms, and with the following operations on arrow terms, besides
the operation :
f1 : C A 1

f2 : C A2

f1 , f2 : C A1 A2
g1 : A1 C

g2 : A2 C

K 1A2 g1 : A1 A2 C

K 2A1 g2 : A1 A2 C

To obtain the equations of E(GL), we assume the categorial equations and

the following equations, for i {1, 2}:

(K 1)

(K 2)

(K 3)

(K 4)

g K iA f = K iA (g f ),

K iA g f1 , f2 = g fi ,
g1 , g2 f = g1 f, g2 f ,

1AB = K 1B 1A , K 2A 1B ,

with appropriate types assigned to f , g, fi and gi . The equation (K 2) is

related to (), while (K 3) and (K 4) are related to (). The syntactical

category GL is C(GL)/E(GL).
It is a straightforward, though somewhat lengthy, exercise to check that

with the denitions corresponding to the equations (), (b ), (b ), ( c),

(w), and the additional denitions

k 1A1 ,A2 =df K 1A2 1A1 ,

k 2A1 ,A2 =df K 2A1 1A2 ,

on the one hand, and the denitions ( , ) and

K 1A2 g1 =df g1 k 1A1 ,A2 ,

K 2A1 g2 =df g2 k 2A1 ,A2 ,

9.1.

Coherence of semilattice categories

189

on the other hand, we can prove that L and GL are isomorphic categories,

and that hence C(L) and C(GL) are synonymous syntactical systems.

We can prove as follows Composition Elimination (see 8.1) for GL.

Proof of Composition Elimination for GL. Take a subterm g f of

an arrow term of C(GL) such that both f and g are composition-free. We

call such a subterm a topmost cut. We show that g f is equal either to a
composition-free arrow term or to an arrow term all of whose compositions
occur in topmost cuts of strictly smaller length than the length of g f .
The possibility of eliminating composition in topmost cuts, and hence every
composition, follows by induction on the length of topmost cuts.
The cases where f or g are 1A are taken care of by (cat 1); the cases

where f is K iA f are taken care of by (K 1); and the case where g is g1 , g2

is taken care of by (K 3). The following cases remain.

If f is f1 , f2 , then g is either of a form covered by cases above, or g is

K iA g , and we apply (K 2).

Note that we do not use the equations (cat 2) and (K 4) in this proof.

An arrow term of C(GL) is said to be in normal form when it is

composition-free and there are no subterms of it of the forms 1AB and

K iA f, g. In GL we have the equations (K 4) and, for i {1, 2},

(K 5)

K iA f, g = K iA f, K iA g,

which is obtained with the help of (cat 1), (K 1) and (K 3). With Composition Elimination and these equations, it can be shown that for every arrow

term of C(GL) there is an arrow term f of C(GL) in normal form such

that f = f in GL. Namely, the following holds for GL.

Normal-Form Lemma. Every arrow term is equal to an arrow term in
normal form.
(The proof of this lemma is incorporated in the proof of the Normal-Form

Lemma for GL in the next section.)

The functor G from L to Rel yields with the isomorphism from GL to

L a functor, which we also call G, from GL to Rel. We can then prove the

following for arrow terms of C(GL).

190

CHAPTER 9.

LATTICE CATEGORIES

Uniqueness Lemma. For every arrow term f there is a unique arrow term
f in normal form such that Gf = Gf .
Proof. Let f be of type A B. It follows from the Normal-Form Lemma
and the functoriality of G that there is at least one arrow term f in normal
form such that Gf = Gf . To show that f is unique we proceed by
induction on the number n(B) of occurrences of in B.
If n(B) = 0, then we make an auxiliary induction on n(A). If n(A) = 0,
then f can be only of the form 1p . If n > 0, then A must be of the form

A1 A2 , and f can be only of the form K 1A2 g or K 2A1 h. Since there are

no g and h such that G K 1A2 g = G K 2A1 h, the arrow term f is uniquely

determined.
Suppose now n(B) > 0. Then B must be of the form B1 B2 , and f
can be only of the form f1 , f2 for f1 : A B1 and f2 : A B2 arrow terms
in normal form. We have that

G K 1B2 f = G K 1B2 f1 , f2 = Gf1 ,

and so, by the induction hypothesis, f1 is unique. Analogously, G K 2B1 f =

Gf2 , and f2 is unique. So f1 , f2 is unique.

(This Uniqueness Lemma is analogous, but not completely analogous, to

the homonymous lemma of 5.2: in the former lemma we do not presuppose
the Normal-Form Lemma, while in the formulation of the present one we
do. Since, however, we have the Normal-Form Lemma in both cases, the
dierence is more in the style of exposition than in mathematical content.)
We can then prove the following (for references concerning this result,
see the references mentioned in the next section before Cartesian Coherence).

Semilattice Coherence. The functor G from L to Rel is faithful.

Proof. We prove that the functor G from GL to Rel is faithful. This

yields Semilattice Coherence.

Suppose that for f1 , f2 : A B arrow terms of C(GL) we have Gf1 =

Gf2 . By the Normal-Form Lemma, for i {1, 2}, there is an arrow term
fi in normal form such that fi = fi . Then from Gf1 = Gf1 = Gf2 = Gf2

9.2.

Coherence of cartesian categories

191

and the Uniqueness Lemma we conclude that f1 and f2 are the same arrow

term, and hence f1 = f2 in GL.

9.2.

Coherence of cartesian categories

To obtain the natural logical category L , we have that the logical system

C(L ) is in L, , with the transformations included in 1, b, c, w-k and

-. The specic equations of E(L ) are obtained by taking the union of

those of E(L) and E(S ) plus

(k )

k 1A, =
A.

There are some redundancies in this axiomatization. The equation

(b ) is derivable with the help of (b k). In E(L ) we can derive the

following:

(w )

w =
.

Natural L -categories are called cartesian categories. These are categories with all nite products, including the empty product. The objects
of a cartesian category that is a partial order make a semilattice with unit.

In C(L ) we have the denition

A =df k 2A,

A,

and for L we have the equations ( nat) and

( 1)

= 1 .

The equations ( nat) and ( 1) amount to

()

A = f,

for f : A ,

which says that is a terminal object in L (see 2.2 for the notion of
terminal object).

A logical system synonymous with C(L ) is obtained by taking as prim

itive instead of -. This is based on the following equations of L :

192

CHAPTER 9.

LATTICE CATEGORIES

A = 1A , A ,

A = A , 1A .

Another alternative logical system synonymous with C(L ) is obtained

by taking as primitives - and instead of k 1 and k 2 . This is based on

the following equations of L :

(1
k 1A,B =
A B ),
A

(
k 2A,B =
A 1B ).
B

If the operation , is primitive together with k 1 and k 2 , then we can

take as primitive, and assume besides the categorial equations, () and

() just ().

Let C(GL ) be the syntactical system dened as C(GL) save that L is

replaced by L, , and besides identity arrow terms we have the arrow terms

A : A as primitive arrow terms. The equations of E(GL ) are those of

E(GL) plus (). The syntactical category GL is C(GL )/E(GL ). It is

easy to ascertain that L and GL are isomorphic, and that hence C(L )

and C(GL ) are synonymous syntactical systems.

We can prove Composition Elimination for GL by enlarging the proof

of the preceding section. The additional cases are where g is A , which is

taken care of by (), and where f is A . In the latter case, g is either 1 ,

or , or g1 , g2 , which are cases already covered.

An arrow term of C(GL ) is said to be in normal form when it is

composition-free and has no subterms of the forms 1AB , K iA f, g, 1

and K iA B . Then we can prove the Normal-Form Lemma of the preceding

section for GL .

Proof of the Normal-Form Lemma for GL . In GL we have the

equations (K 4) and (K 5) of the preceding section, and also

1 = ,

K 1B A = AB ,

K 2A B = AB ,

9.2.

Coherence of cartesian categories

193

which are all instances of () (the rst of these equations is ( 1)).

For f a composition-free arrow term of C(GL ), let n1 be the number

of occurrences of and in the indices of subterms of f that are identity
arrow terms. Next, let n2 be the number of subterms of f of the form

f1 , f2 or C such that there is a subterm K iD f of f with f1 , f2 or C

a subterm of f . Let the grade of f be (n1 , n2 ), and let these grades be
lexicographically ordered (see 7.7, before the Cut-Elimination Theorem).
Then every replacement of subterms of f justied by one of the equations
above reduces the grade of f , and so by induction we obtain that there is

an arrow term f of C(GL ) in normal form such that f = f in GL . It

remains only to appeal to Composition Elimination for GL to obtain the

Normal-Form Lemma.

As in the preceding section, we obtain the functor G from GL to Rel,

with which we can prove the Uniqueness Lemma of the preceding section

for GL .

Proof of the Uniqueness Lemma for GL . We proceed by induction

n(B) with an auxiliary induction on n(A) in the basis, as in the proof of
the preceding section. If n(B) = n(A) = 0, then f can be either of the

form 1p , or p , or , which exclude each other because of their types.

If n(B) = 0 and n(A) > 0, then f can be either of the form K 1A2 g, or

K 2A1 h or A , which exclude each other because of their types or for reasons
mentioned in the proof of the preceding section. For that we use the fact,

easily shown by induction on the length of A, that A is the only arrow

term in normal form of type A , and the fact that Gf is a function
from GB to GA. For the rest of the proof we proceed as in the preceding
section.

We can then infer as in the preceding section the following result, which
stems from [77] (see p. 129, where the result is announced), [106] (Theorem
2.2), [128] (Theorem 8.2.3, p. 207), [108] (Section 7) and [46].

Cartesian Coherence. The functor G from L to Rel is faithful.

It is noteworthy that the functor G maps every arrow of L to a function

194

CHAPTER 9.

LATTICE CATEGORIES

from the target to the source. (We used that fact in the proof of the

Uniqueness Lemma of this section.) The same holds, of course, for L.

9.3. Maximality of semilattice and cartesian categories

A natural logical category C/E is called maximal when every natural C/Ecategory that satises an equation between arrow terms of C that is not in
E is a preorder. In other words, if E is a proper extension of E, then every
natural C/E -category is a preorder. Maximality is an interesting property

when C/E itself is not a preorder, and we will show in this section that L

and L are maximal in this sensein the interesting way. (We take over
these results from [46].)
The maximality property above is analogous to the property of usual
formulations of the classical propositional calculus called Post completeness. That this calculus is Post complete means that if we add to it any
new axiom-schema in the language of the calculus, then we can prove every
formula. An analogue of Bohms Theorem in the typed lambda calculus
implies, similarly, that the typed lambda calculus cannot be extended without falling into triviality, i.e. without every equation (between terms of the
same type) becoming derivable (see [117], [45] and references therein; see
[4], Section 10.4, for Bohms Theorem in the untyped lambda calculus).
Let us now consider several examples of common algebraic structures
with analogous maximality properties. First, we have that semilattices are
maximal in the following sense.
Let a and b be terms made exclusively of variables and of a binary
operation , which we interpret as meet or join. That the equation a = b
holds in a semilattice S means that every instance of a = b obtained by
substituting names of elements of S for variables holds in S (cf. 2.3).
Suppose a = b does not hold in a free semilattice SF (so it is not the case
that a = b holds in every semilattice). Hence there must be an instance
of a = b obtained by substituting names of elements of SF for variables
such that this instance does not hold in SF . It is easy to conclude that
in a = b there must be at least two variables, and that SF must have at
least two free generators. Then every semilattice in which a = b holds is
trivialnamely, it has a single element.

9.3.

Maximality of semilattice and cartesian categories

195

Here is a short proof of that. If a = b does not hold in SF , then there

must be a variable x in one of a and b that is not in the other. Then from
a = b, by substituting y for every variable in a and b dierent from x, and
by applying the semilattice equations, we infer either x = y or x y = y. If
we have x = y, we are done, and, if we have x y = y, then we have also
y x = x, and hence x = y.
Semilattices with unit, distributive lattices, distributive lattices with
top and bottom, and Boolean algebras are maximal in the same sense. The
equations a = b in question are equations between terms made exclusively
of variables and the operations of the kind of algebra we envisage: semilattices with unit, distributive lattices, etc. That such an equation holds
in a particular structure means, as above, that every substitution instance
of it holds. However, the number of variables in a = b and the number of
generators of the free structure mentioned need not always be at least two.
If we deal with semilattices with unit 1, then a = b must have at least
one variable, and the free semilattice with unit must have at least one free
generator. We substitute 1 for every variable in a and b dierent from x in
order to obtain x = 1, and hence triviality. So semilattices with unit are
maximal in the same sense.
The same sort of maximality can be proven for distributive lattices,
whose operations are and , which we call conjunction and disjunction,
respectively. Then every term made of , and variables is equal to a
term in disjunctive normal form (i.e. a multiple disjunction of multiple conjunctions of variables; see 10.2 for a precise denition), and to a term in
conjunctive normal form (i.e. a multiple conjunction of multiple disjunctions of variables; see 10.2). These normal forms are not unique. If a = b,
in which we must have at least two variables, does not hold in a free distributive lattice DF with at least two free generators, then either a b
or b a does not hold in DF . Suppose a b does not hold in DF . Let
a be a disjunctive normal form of a, and let b be a conjunctive normal
form of b. So a b does not hold in DF . From that we infer that for a
disjunct a of a and for a conjunct b of b we do not have a b in DF .
This means that there is no variable in common in a and b ; otherwise,
the conjunction of variables a would be lesser than or equal in DF to the
disjunction of variables b . If in a distributive lattice a = b holds, then

196

CHAPTER 9.

LATTICE CATEGORIES

a b holds too, and hence, by substitution, we obtain x y. So x = y.

For distributive lattices with top and bottom , we proceed analogously via disjunctive and conjunctive normal form. Here a = b may be
even without variables, and the free structure may have even an empty set
of free generators. The additional cases to consider are when in a b we
have that a is and b is . In any case, we obtain , and hence
our structure is trivial.
The same sort of maximality can be proven for Boolean algebras, i.e.
complemented distributive lattices. Boolean algebras must have top and
bottom. In a disjunctive normal form now the disjuncts are conjunctions
of variables x or terms x
, where is complementation, or the disjunctive
normal form is just or ; analogously for conjunctive normal forms. Then
we proceed as for distributive lattices with an equation a = b that may be
even without variables, until we reach that a b , which does not hold
in a free Boolean algebra BF , whose set of free generators may be even
empty, holds in our Boolean algebra. If x is a conjunct of a , then in b we
cannot have a disjunct x; but we may have a disjunct x
. The same holds
for the conjuncts x
of a . It is excluded that both x and x
are conjuncts

of a , or disjuncts of b ; otherwise, a b would hold in BF . Then for

every conjunct x in a and every disjunct y in b we substitute for x
and y, and for every other variable we substitute . In any case, we obtain
, and hence our Boolean algebra is trivial. This is essentially the
proof of Post completeness for the classical propositional calculus, due to
Bernays and Hilbert (see [135], Section 2.4, and [66], Section I.13), from
which we can infer the ordinary completeness of this calculus with respect
to valuations in the two-element Boolean algebranamely, with respect to
truth tablesand also completeness with respect to any nontrivial model.
As examples of common algebraic structures that are not maximal in
the sense above, we have semigroups, commutative semigroups, lattices,
and many others. What is maximal for semilattices and is not maximal
for lattices is the equational theory of the structures in question. The
equational theory of semilattices cannot be extended without falling into
triviality, while the equational theory of lattices can be extended with the
distributive law, for example.

The maximality of L as dened at the beginning of the section diers

9.3.

Maximality of semilattice and cartesian categories

197

from the maximality of semilattices, distributive lattices, etc., we have just

considered, because in L we have types, so that f = g is excluded if f and g

are of dierent types. Hence, the analogue of the trivial semilattice, which

was a one-element structure, is for categories, like L, a preorder.

The maximality of L is, of course, a quite separate result from the

maximality of semilattices we have shown above. None of these results
can be inferred from the other. After some strictication, any semilattice
category yields a semilattice category that is a partial order, and whose
objects will make a semilattice. The maximality of semilattices has to do

with these objects, while the maximality of L has to do with the arrows
between these objects. We will now proceed with the proof of the latter
maximality.

Maximality of L. The category L is maximal.

Proof. Suppose A and B are formulae of L in which only p occurs as a

letter. Suppose f1 , f2 : A B are arrow terms of C(L) such that Gf1 = Gf2 .
As we noted after Cartesian Coherence, at the end of the preceding section,
Gf1 and Gf2 may be conceived as functions from GB to GA. So there must
be an n GB such that Gf1 (n) = Gf2 (n). This means that we must have
GA 2 (i.e., there must be at least two occurrence of p in A), and we have,
of course, GB 1.

Then there is an arrow term hw : p p A of C(L) made of possi

bly multiple occurrences of arrow terms in 1, b, c and w, together with

the operations and on arrow terms, such that Ghw (Gf1 (n)) = 0 and

Ghw (Gf2 (n)) = 1. There is also an arrow term hk : B p of C(L) that is

either 1p or a possibly iterated composition of arrow terms in k 1 and k 2

such that Ghk (0) = n. Then, for i {1, 2}, we have that hk fi hw is

of type p p p and G(hk fi hw ) = G k ip,p . Therefore, by Composition

Elimination for GL (see 9.1) and by the functoriality of G, we obtain that

hk fi hw = k ip,p in L. (This follows from Semilattice Coherence too.) So

in E(L) extended with f1 = f2 we can derive the equation

(k k)

k 1p,p = k 2p,p .

If this equation holds in a semilattice category A, then A is a preorder.

198

CHAPTER 9.

LATTICE CATEGORIES

This is shown as follows. For f, g : a b in A we have

k 1b,b f, g = k 2b,b f, g,
and so f = g in A by the equation () of 9.1.

If for some arrow terms g1 and g2 of C(L) of the same type we have

that g1 = g2 is not in E(L), then by Semilattice Coherence (see 9.1) we

have Gg1 = Gg2 . If we take the substitution instances g1 of g1 and g2 of g2
obtained by replacing every letter by a single letter p, then we obtain again
Gg1 = Gg2 . If g1 = g2 holds in a semilattice category A, then g1 = g2 holds
too, and A is a preorder, as we have shown above.

We have also the following.

Maximality of L . The category L is maximal.

To prove that we proceed as for L. The only modication is that in con

structing hw we envisage also arrow terms in and .

Note that the maximality of L implies that in any semilattice category

A that is not a preorder we can falsify any equation between arrow terms of

C(L) that does not hold in L. This does not mean, however, that there must

be a faithful functor from L to A, which would falsify all such equations

simultaneously. The existence of such a functor is possible for particular
semilattice categories A, but it is another result, which does not follow from

and does not imply maximality. In the case of L and L , the category Set of
sets with functions, with being cartesian product and being a singleton,

is an A such that there is a faithful functor from L and L to A (see [30]

and [117]).
Maximality holds also trivially for all logical categories K that are preorders, because we cannot extend E(K) properly in such cases. The logical
categories that are not preorders that we have considered up to now are

symmetric, i.e. they have c in C(K). Before L and L , however, the symmetric logical categories K from previous chapters that are coherent are not
maximal, in spite of coherence, for the following reason.
All the types of arrow terms of C(K) are balanced (in the sense of 3.3).
Let the balance weight of an equation f = g where f, g : A B are arrow
terms of C(K) be the letter length of A or B. Then it can be shown that

9.4.

Coherence of lattice categories

199

if E is the extension of E(K) with an equation f = g that is not in E(K),

with a single letter occurring in A (and hence also in B), and the balance
weight of f = g is n, then all the equations E that are not in E(K) must
have a balance weight greater than or equal to n.
The notions of maximality envisaged in this section were extreme (or
should we say maximal), in the sense that we envisaged collapsing only
into preorder. (For semilattices, distributive lattices, etc., this is also preorder for a one-object category.) We may, however, envisage relativizing
our notion of maximality by replacing preorder with a weaker property,
such that structures possessing it are trivial, but not so trivial (cf. [38],
Section 4.11). We will encounter maximality in such a relative sense in
9.7.
As an example of relative maximality in a common algebraic structure
we can take symmetric groups. The axioms for the symmetric group S n ,
where n 2, with the generators si , for i {1, . . . , n1}, were given in 5.1.
If to S n for n 5 we add an equation a = 1 where a is built exclusively of
the generators si of S n with composition, and a = 1 does not hold in S n ,
then we can derive si = sj . This does not mean that the resulting structure
will be a one-element structure, i.e. the trivial one-element group. It will
be such if a is an odd permutation, and if a is an even permutation, then
we will obtain a two-element structure, which is S 2 . This can be inferred
from facts about the normal subgroups of S n . Simple groups are maximal
in the nonrelative sense, envisaged above for semilattices.

9.4.

Coherence of lattice categories

Let L be the natural logical category in L isomorphic to the category Lop

(which is L with source and target functions interchanged; see 2.2). We

just replace by , so that the primitive arrow terms of C(L) are included

in 1, b, c and w-k, while the equations of E(L) are duals of those of E(L)
(see the List of Equations and the List of Categories at the end of the

book). Natural L-categories would usually be called categories with nite

nonempty coproducts.

Let C(GL) be the syntactical system with formulae of L as objects,

with the primitive arrow terms being only identity arrow terms, and with

200

CHAPTER 9.

LATTICE CATEGORIES

the following operations on arrow terms, dual to those of C(GL), besides

the operation (cf. 9.1):
g1 : A1 C

g2 : A2 C

[g1 , g2 ] : A1 A2 C
f2 : C A2

f1 : C A1

K 1A2 f1 : C A1 A2

K 2A1 f2 : C A1 A2

To obtain the equations of E(GL), we assume the categorial equations and

the following equations for i {1, 2}, obtained by dualizing the equations

(K 1)-(K 4) of 9.1:

(K 1)

(K 2)

(K 3)

(K 4)

K iA g f = K iA (g f ),

[g1 , g2 ] K iA f = gi f,
g [f1 , f2 ] = [g f1 , g f2 ],

1AB = [K 1B 1A , K 2A 1B ],

with appropriate types assigned to f , g, fi and gi . The syntactical category

GL is C(GL)/E(GL). It is clear that GL is isomorphic to L, and also to

Lop and GLop . For later use, we note that in C(L) we have the denitions

[g1 , g2 ] =df wC (g1 g2 ),

K 1A2 f1 =df k 1A1 ,A2 f1 ,

K 2A1 f2 =df k 2A1 ,A2 f2 .

(We introduce G(L) with so much detail for the sake of notation.)
To obtain the natural logical category L, we have that the logical system
C(L) is in L, , with the transformations included in 1, b, c and w-k.
The specic equations of E(L) are obtained by taking the union of those of

E(L) and E(L).

We call natural L-categories lattice categories. Usually, they would be
called categories with nite nonempty products and nite nonempty coproducts. The objects of a lattice category that is a partial order make a
lattice.

9.4.

Coherence of lattice categories

201

The syntactical system C(GL) has as objects the formulae of L, ,

as primitive arrow terms the identity arrow terms, and as operations on

arrow terms those of C(GL) and C(GL). For the equations of E(GL) we

assume the equations of E(GL) and E(GL). The syntactical category GL

is C(GL)/E(GL), and it is isomorphic to L. This isomorphism is based on

the isomorphism of L with GL, and the isomorphism of L with GL.

We can then prove Composition Elimination for GL by enlarging the
proof in 9.1.
Proof of Composition Elimination for GL. We have rst the cases

where f or g are 1A , where f is K iA f and where g is g1 , g2 . For these

cases we proceed as before. We have next cases dual to the last two, where

g is K iA g , which is taken care of by (K 1), and where f is [f1 , f2 ], which

is taken care of by (K 3). In the remaining cases, if f is f1 , f2 , then g is

either of a form already covered by cases above, or g is K iA g , and we apply

(K 2). Finally, if f is K iA f , then g is either of a form already covered by

cases above, or g is [g1 , g2 ], and we apply (K 2).

Note that we do not use the equations (K 4) and (K 4) in this proof (which
is taken over from [48], Section 3). We can then prove the following.
Invertibility Lemma for . Let f : A1 A2 B be an arrow term of
C(GL). If for every (x, y) Gf we have that x GA1 , then f is equal in

GL to an arrow term of the form K 1A2 f , and if for every (x, y) Gf we

have that x GA1 GA2 , then f is equal in GL to an arrow term of the

form K 2A1 f .
Proof. By Composition Elimination for GL, we can assume that f is
composition-free, and then we proceed by induction on the length of the
target B (or on the length of f ). If B is a letter, then f must be equal in

L to an arrow term of the form K iA3i f . The condition on Gf dictates

whether i here is 1 or 2.

If B is B1 B2 and f is not of the form K iA3i f , then f must be

of the form f1 , f2 (the condition on Gf precludes that f be an identity
arrow term). We apply the induction hypothesis to f1 : A1 A2 B1 and

f2 : A1 A2 B2 , and use the equation (K 5).

202

CHAPTER 9.

LATTICE CATEGORIES

If B is B1 B2 and f is not of the form K iA3i f , then f must be

of the form K jB3j g, for j {1, 2}. We apply the induction hypothesis to
g : A1 A2 Bi , and use the equation

K jB3jK iA3i g = K iA3iK jB3j g ,

which follows from (cat 1), (K 1), (K 1) and (cat 2).

We have a dual Invertibility Lemma for . We can then prove the

following result of [48] (Section 4).
Lattice Coherence. The functor G from L to Rel is faithful.
Proof. Suppose f, g : A B are arrow terms of C(L) and Gf = Gg. We
proceed by induction on the sum of the lengths of A and B to show that
f = g in L. If A and B are both letters, then we conclude by Composition
Elimination for GL that an arrow term of C(L) of the type A B exists i
A and B are the same letter p, and we must have f = g = 1p in L. Note
that we do not need here the assumption Gf = Gg.

If B is B1 B2 , then for i {1, 2} we have that k iB1 ,B2 f and k iB1 ,B2 g
are of type A Bi . We also have

G(k iB1 ,B2 f ) = G k iB1 ,B2 Gf = G k iB1 ,B2 Gg = G(k iB1 ,B2 g),

whence, by the induction hypothesis, we have k iB1 ,B2 f = k iB1 ,B2 g in L.

Then we infer

k 1B1 ,B2 f, k 2B1 ,B2 f = k 1B1 ,B2 g, k 2B1 ,B2 g,

from which f = g follows with the help of the equation () of 9.1. We
proceed analogously if A is A1 A2 .
Suppose now that A is A1 A2 or a letter, and B is B1 B2 or a letter,
but A and B are not both letters. Then by Composition Elimination for
GL we have that f is equal in L to an arrow term of C(L) that is either

of the form f k iA1 ,A2 or of the form k iB1 ,B2 f . Suppose f = f k 1A1 ,A2 .
Then for every (x, y) Gf we have x GA1 .
By the Invertibility Lemma for , which we have proved above, it follows

that g is equal in L to an arrow term of the form g k 1A1 ,A2 . From Gf = Gg

9.4.

Coherence of lattice categories

203

we can infer easily that Gf = Gg , and so by the induction hypothesis

f = g , and hence f = g.

We reason analogously when f = f k 2A1 ,A2 . If f = k iB1 ,B2 f , then

again we reason analogously, applying the Invertibility Lemma for .

This proof of Lattice Coherence is simpler than a proof given in [48].

In the course of that previous proof one has also coherence results for two
auxiliary categories related to L. We need these categories for 9.6, but we
do not need these coherence results. For the sake of completeness, however,
we record them here too.

Let L be the natural logical category in L, obtained as L. The only

dierence is that the arrow terms of C(L ) are closed under the operation
on arrow terms, besides being closed under the operations and ,

and for E(L ) we have in addition to the equations assumed for E(L) the

bifunctorial equations for . Let GL be the syntactical category whose

objects are formulae of L, , which is obtained as GL save that in addition

to the operations on arrow terms of C(GL) we have also the operation

on arrow terms, and for E(GL ) we assume the bifunctorial equations for

in addition to what we had for E(GL). The categories L and GL are

isomorphic, and, hence, C(L ) and C(GL ) are synonymous syntactical

systems.

op
The categories L and GL are isomorphic to Lop
and GL , and to

each other. In them, the and of L and GL are interchanged, and

they are obtained by extending L and GL with the bifunctor .

One can easily prove Composition Elimination for GL (and hence

also for GL ) by abbreviating the proof of Composition Elimination for L

above. For GL we do not have the cases where f is [f1 , f2 ] or K iA f , but

f can be f1 f2 . Then, if g is not of a form already covered by the proof
in 9.1, it must be g1 g2 , and we apply the bifunctorial equation ( 2).

A composition-free arrow term of C(GL ) may be reduced to a unique

normal form, which can then be used to demonstrate coherence for L , i.e.

the fact that the functor G from L to Rel is faithful (see [48], Section 4).
With the help of Lattice Coherence we can easily verify that the following equation holds in L:

204

CHAPTER 9.

(in-out)

LATTICE CATEGORIES

[f, g], [h, j] = [f, h, g, j].

If in C(L) we dene ckA,B,C,D : (A B) (C D) (A C) (B D)

as follows:

ckA,B,C,D =df k 1A,B k 1C,D , k 2A,B k 2C,D ,

then we can easily check that in L we have

ckA,B,C,D = [k 1A,C k 1B,D , k 2A,C k 2B,D ],

which gives an alternative denition of ckA,B,C,D . One passes from one of
these two denitions to the other with the help of the equations () of 9.1
and () of the List of Equations, together with the equation (in-out):

ckA,B,C,D = [k 1A,C

k 1A,B , k 2A,C

k 1A,B , k 1B,D k 2A,B , k 2A,C

= [k 1A,C

k 1C,D ], [k 1B,D k 2A,B , k 2B,D k 2C,D ]

k 1C,D , k 2B,D k 2C,D ].

We can also show by Lattice Coherence that in L we have

wAB = ckA,A,B,B (wA wB ),

wAB = (wA wB ) ckA,B,A,B ,

1
1
2
2
cm
A,B,C,D = k A,B k C,D , k A,B k C,D ,

1
1
2
2
cm
D,C,B,A = [k D,C k B,A , k D,C k B,A ]

(see 9.1 for c m , and the List of Equations for c m ). The last two equations should be compared with the denition of ckA,B,C,D and its alternative
denition. The arrows ckA,B,C,D will be prominent in Chapter 11 (see also
13.2).
Arrows of the type of ckA,B,C,D play in [3] an important role in the
understanding of 2-fold loop spaces. In that paper, one nds a coherence
result in our sense for bimonoidal categories where = to which ck
is added with appropriate specic equations. As a matter of fact, this
coherence result, for which a long proof is presented, covers a hierarchy of ck
principles involving the binary connectives i and j where 1 i < j n,

9.5.

Maximality of lattice categories

205

which are needed for n-fold loop spaces. The role of arrows of the type
of ckA,B,C,D in understanding braiding is considered in [72] (Section 5). In
that context, arrows of the type of ckA,B,C,D may become arrows of the type

of cm
A,B,C,D (cf. [3], Remarks 1.5-6).

9.5.

Maximality of lattice categories

In this section we prove that L is maximal in the sense of 9.3. (This result
is taken over from [48], Section 5.)
Suppose A and B are formulae of L, in which only p occurs as a letter.
If for some arrow terms f1 , f2 : A B of C(L) we have Gf1 = Gf2 , then for
some x GA and some y GB we have (x, y) Gf1 and (x, y) Gf2 , or
vice versa. Suppose (x, y) Gf1 and (x, y) Gf2 .
For every subformula C of A and every formula D let AC
D be the formula
obtained from A by replacing the particular occurrence of the formula C in
A by D. It can be shown that for every subformula A1 A2 of A we have

1 A2
a k j -term h : AA
A of C(L), whose head is k jA1 ,A2 , such that there
Aj
1 A2
is an x GAA
for which (x , x) Gh. Hence, for such an h, we have
Aj
(x , y) G(f1 h) and (x , y) G(f2 h).

We compose fi repeatedly with such k j -terms until we obtain the arrow

terms fi : p . . . p B of C(L) such that parentheses are somehow
associated in p . . . p and for some z G(p . . . p) we have (z, y) Gf1
and (z, y) Gf2 . The formula p . . . p may also be only p. We may

further compose fi with b-terms and c-terms in order to obtain the arrow
terms fi of type p A B or p B such that A is of the form p . . . p
with parentheses somehow associated, and (0, y) Gf1 but (0, y) Gf2 .

By working dually on B with k j -terms, and by composing perhaps fur

ther with b-terms and c-terms, we obtain the arrow terms fi of C(L) of
type pA pB , for A of the form p. . .p and B of the form p. . .p,
or of type p A p, or of type p p B , such that (0, 0) Gf1 and
(0, 0) Gf2 . (We cannot obtain that f1 and f2 are of type p p, since,
otherwise, by Composition Elimination for GL, f2 would not exist.)
There is an arrow term h : p p . . . p of C(L) dened in terms of

w-terms such that for every x G(p . . . p) we have (0, x) Gh . We

dene analogously with the help of w-terms an arrow term h : p. . .p p

206

CHAPTER 9.

LATTICE CATEGORIES

of C(L) such that for every x G(p . . . p) we have (x, 0) Gh . The

arrow terms h and h may be 1p : p p.
If fi is of type p A p B , let fi : p p p p be dened by
fi =df (1p h ) fi (1p h ).
By Composition Elimination for GL, we have that Gfi must be a singleton.

If (1, 0) or (1, 1) belongs to Gf2 , then for fi : p p p dened as wp fi we

have (0, 0) Gf1 and (0, 0) Gf2 . If (0, 1) or (1, 1) belongs to Gf2 , then

for fi : p p p dened as fi wp we have (0, 0) Gf1 and (0, 0) Gf2 .

If fi is of type pA p, then for fi : pp p dened as fi (1p h )
we have (0, 0) Gf1 and (0, 0) Gf2 .
If fi is of type p pB , then for fi : p pp dened as (1p h ) fi
we have (0, 0) Gf1 and (0, 0) Gf2 . In all that we have by Composition
Elimination for GL that Gfi must be a singleton.
In cases where fi is of type p p p, by Composition Elimination for
GL, by the conditions on Gf1 and Gf2 , and by the functoriality of G, we

obtain in L the equation fi = k ip,p . (This follows from Lattice Coherence

too.) So in E(L) extended with f1 = f2 we can derive k 1p,p = k 2p,p ; namely,

the equation (k k), mentioned in the proof of Maximality of L in 9.3.

In cases where fi is of type p p p, we conclude analogously that we

have in L the equation fi = k ip,p , and so in E(L) extended with f1 = f2 we

can derive

(k k)

k 1p,p = k 2p,p .

If either of (k k) and (k k) holds in a lattice category A, then A is a

preorder. We use for that the equation () of 9.1, or its dual with (see

the proof of Maximality of L in 9.3).

It remains to remark that if for some arrow terms g1 and g2 of C(L)
of the same type we have that g1 = g2 is not in E(L), then by Lattice
Coherence we have Gg1 = Gg2 . If we take the substitution instances g1
of g1 and g2 of g2 obtained by replacing every letter by a single letter p,
then we obtain again Gg1 = Gg2 . If g1 = g2 holds in a lattice category
A, then g1 = g2 holds too, and A is a preorder, as we have shown above.
This concludes the proof of maximality for L. (In the original presentation

9.6.

Coherence for dicartesian and sesquicartesian categories

207

of this proof in [48], Section 5, there are some slight inaccuracies in the
denition of fi .)

9.6.

Coherence for dicartesian and sesquicartesian categories

Let L be the natural logical category in L, isomorphic to the category

Lop
. We just replace and by and respectively, so that the primitive

arrow terms of C(L ) are included in 1, b, c, w-k and -, while the

equations of E(L ) are duals of those of E(L ) (see the List of Equations

and the List of Categories; cf. the beginning of 9.4). We have in C(L )
the denition

k2
A =df
A
A, ,

and in L the equations ( nat) and

( 1)

= 1 .

The equations ( nat) and ( 1) amount to

()

A = f,

for f : A,

which says that is an initial object in L (see 2.2 for the notion of initial
object).

Natural L -categories would usually be called categories with nite coproducts, including the empty coproduct. Another possible name would be
cocartesian categories.
To obtain the natural logical category L, , we have that the logical
system C(L, ) is in L,,, , with the transformations included in 1, b,
c, w-k and -. The specic equations of E(L, ) are obtained by taking

the union of those of E(L), E(L ) and E(L ) plus the equations ( c ) and

( c ) of 6.4.
We could replace the last two equations in this denition by their instances

()

()

c , = 1 ,

c , = 1 .

208

CHAPTER 9.

LATTICE CATEGORIES

Another possibility is to have instead the following two equations:

(k )

(k )

k 1, = k 2, ,
k 1, = k 2, .

It is easy to see that from the last two equations we obtain that the
members of the pairs

k 1, = k 2, :

and = w : ,

k 1, = k 2, :

and = w :

are inverses of each other. This shows that every letterless formula of
L,,, is isomorphic in L, either to or to , and this is why above

we could replace ( c ) and ( c ) by their instances () and ().

We call natural L, -categories dicartesian categories. The objects of
a dicartesian category that is a partial order make a lattice with top and

bottom. By omitting the equations ( c ) and ( c ) in the denition of

L, we would obtain the natural logical category L, , and natural L, categories are usually called bicartesian categories (cf. [90], Section I.8).
Dicartesian categories were considered under the name coherent bicartesian
categories in the printed version of [47]. We previously believed wrongly
that we have proved coherence for dicartesian, alias coherent bicartesian
categories. Lemma 5.1 of the printed version of [47] is however not correct.
We prove here only a restricted coherence result for dicartesian categories,
which is sucient for our needs later on. A study of equality of arrows in
bicartesian categories may be found in [24].
Suppose that in the denition of L, we omit one of and from
L,,, , so that we have L,, or L,, . This means that in C(L, )
and E(L, ) we omit all the arrow terms and equations involving the omitted nullary connective. When we omit , we obtain the natural logical
category L , and when we omit , we obtain the natural logical category
L . It is clear that L is isomorphic to Lop
. In the printed version of [47]
natural L -categories were called coherent sesquicartesian categories. We
call them here just sesquicartesian categories.
The category Set, whose objects are sets and whose arrows are functions,
with cartesian product as , disjoint union + as , a singleton set {} as

9.6.

Coherence for dicartesian and sesquicartesian categories

209

and the empty set as , is a bicartesian category, but not a dicartesian

category. It is, however, a sesquicartesian category in the L sense, but
not in the L sense. This is because in Set we have that is equal to
, but {} + {} is not isomorphic to {}.
We have an unrestricted coherence result for sesquicartesian categories,
whose proof is taken over from the revised version of [47]. (This proof diers
from the proof in the printed version of [47], which relied also on Lemma
5.1, and is not correct.)
Sesquicartesian Coherence. The functor G from L to Rel is faithful.
The proof of this result is obtained by enlarging the proof of Lattice Coherence in 9.4, and we will give here just a summary of it. (A detailed proof
may be found in the revised version of [47].)
The syntactical category GL, is obtained as GL save that we have in

addition the primitive arrow terms A : A and A : A, the equations

() and (), and also the equations

(K)

(K)

K 1 1 = K 2 1 ,
K 1 1 = K 2 1 .

We can prove Composition Elimination for GL, by enlarging the proofs

in 9.1-2. Note that we do not need the equations (K ) and (K ) for

this proof, so that we have also Composition Elimination for GL, based
on L, .

Let the category L,, be dened like the category L of 9.4 save

that it involves also and the equations () and (k ), and let the category

L,, be dened like the category L of 9.4 save that it involves also

and the equations () and (k ). Composition Elimination is provable for

syntactical categories isomorphic to L,, and L,, .

An arrow term of C(L, ) is in standard form when it is of the form

g f for f an arrow term C(L,, ) and g an arrow term of C(L,, ). We

can then prove the following.
Standard-Form Lemma. Every arrow term of C(L, ) is equal in L,
to an arrow term in standard form.

210

CHAPTER 9.

LATTICE CATEGORIES

Proof. By categorial and bifunctorial equations, we may assume that we

deal with a factorized arrow term f none of whose factors is a complex
identity (see 2.6-7 for these notions). We may assume moreover that b

and c do not occur in f . For that we use the equations (b ), (b ) and ( c)

of 9.1, and the dual equations with . We may also assume that every

factor of f is either an arrow term of C(L,, ), and then we call it a

-factor, or an arrow term of C(L,, ), when we call it a -factor.

Suppose f : B C is a -factor and g : A B is a -factor. We show
by induction on the length of f g that in L,
()

f g = g f

or

f g = f

or

f g = g

for f a -factor and g a -factor.

We will consider various cases for f . In all such cases, if g is wB , then

we use (w nat). If f is wB , then we use (w nat). If f is k iD,E and g is

g1 g2 , then we use (k i nat). If f is f1 f2 and g is g1 g2 , then we use

bifunctorial and categorial equations and the induction hypothesis.

If f is f1 f2 , then we have the following cases. If g is k iB1 ,B2 , then

we use (k i nat). If g is g1 g2 , then we use bifunctorial and categorial

equations and the induction hypothesis.

Finally, cases where f is B or g is B are taken care of by the equations

() and (). This proves (), and it is clear that () is sucient to prove
the lemma.

We can also prove Composition Elimination and an analogue of the

Standard-Form Lemma for L . Next we have the following lemmata for
L, and L .
Lemma 1. If for f, g : A B either A or B is isomorphic to or , then
f = g.
Proof. If A is isomorphic to or B is isomorphic to , then the matter
is trivial. Suppose i : B is an isomorphism. Then from

k 1, i f, i g = k 2, i f, i g
we obtain i f = i g, which yields f = g. We proceed analogously if A is
isomorphic to .

9.6.

Coherence for dicartesian and sesquicartesian categories

211

Lemma 2. If for f, g : A B we have Gf = Gg = , then f = g.

Proof. This proof depends on the Standard-Form Lemma above. We
write down f in the standard form f2 f1 for f1 : A C and g in the

standard form g2 g1 for g1 : A D. Since k i and do not occur in f1 ,

for every z GC we have an x GA such that (x, z) Gf1 , and since

k i and do not occur in f2 , for every z GC we have a y GB such

that (z, y) Gf2 . So if C were not letterless, Gf would not be empty. We
conclude analogously that D, as well as C, is a letterless formula.
If both C and D are isomorphic to or , then we have an isomorphism
i : C D, and f = f2 i1 i f1 . By Lemma 1, we have i f1 = g1 and
f2 i1 = g2 , from which f = g follows. If i : C and j : D are
isomorphisms, then by Lemma 1 we have

f2 f1 = g2 j i f1 = g2 g1 ,

and so f = g. (Note that = .)

To prove now Sesquicartesian Coherence we have Lemma 2 for the case

when Gf = Gg = , and when Gf = Gg = , we proceed as in the proof
of Lattice Coherence in 9.4, appealing if need there is to Lemma 2, until
we reach the case when A is A1 A2 or a letter, and B is B1 B2 or a
letter, but A and B are not both letters. In that case, by Composition
Elimination, the arrow term f is equal in L either to an arrow term of

the form f k iA1 ,A2 , or to an arrow term of the form k iB1 ,B2 f . Suppose

f = f k 1A1 ,A2 . Then for every (x, y) Gf we have x GA1 . (We reason

analogously when f = f k 2A1 ,A2 .)

By Composition Elimination too, g is equal in L either to an arrow

term of the form g k iA1 ,A2 , or to an arrow term of the form k iB1 ,B2 g . In

the rst case we must have g = g k 1A1 ,A2 , because Gg = G(f k 1A1 ,A2 ) = ,
and then we apply the induction hypothesis to derive f = g from Gf = Gg .
Hence f = g in L .

Suppose g = k 1B1 ,B2 g . (We reason analogously when g = k 2B1 ,B2 g .)

Let f : A1 B1 B2 be the substitution instance of f : A1 B1 B2 obtained by replacing every occurrence of propositional letter in B2 by .
There is an isomorphism i : B2 , and f exists because in Gf , which is

212

CHAPTER 9.

LATTICE CATEGORIES

equal to G(k 1B1 ,B2 g ), there is no pair (x, y) with y GB1 . So we have

an arrow f : A1 B1 , which we dene as

B1 (1B1 i) f . It is easy

to verify that G(k 1B1 ,B2 f ) = Gf , and that G(f k 1A1 ,A2 ) = Gg . By

the induction hypothesis, we obtain k 1B1 ,B2 f = f and f k 1A1 ,A2 = g ,

from which we derive f = g. We reason analogously when f = k iB1 ,B2 f .

From Sesquicartesian Coherence we infer coherence for L , which is
isomorphic to Lop
.
For dicartesian categories we prove here only a simple restricted coherence result sucient for our needs later on in the book. A more general,
but still restricted, coherence result with respect to Rel, falling short of full
coherence, may be found in the revised version of [47] (Section 7, Restricted
Dicartesian Coherence II).
We dene inductively formulae of L,,, in disjunctive normal form
(dnf ): every formula of L,, is in dnf, and if A and B are both in dnf,
then A B is in dnf. We dene dually formulae of L,,, in conjunctive
normal form (cnf ): every formula of L,, is in cnf, and if A and B are
both in cnf, then A B is in cnf.
Restricted Dicartesian Coherence. Let f, g : A B be arrow terms
of C(L, ) such that A is in dnf and B in cnf. If Gf = Gg, then f = g in
L, .
Proof. If Gf = Gg = , then we apply Lemma 2. If Gf = Gg = , then
we proceed as in the proof of Lattice Coherence in 9.4, by induction on the
sum of the lengths of A and B, appealing if need there is to Lemma 2, until
we reach the case when A is A1 A2 or a letter, and B is B1 B2 or a letter,
but A and B are not both letters. In that case there is no occurrence of
in A and no occurrence of in B. We then rely on the composition-free

form of f and g in GL, and on the equation K iC K jD h = K jD K iC h.

Note that if K is one of the categories A, A , S and S , then K is

isomorphic to Kop , while if K is one of the categories L and L , then

K is not isomorphic to Kop . The categories A, A , S and S besides

op
being isomorphic respectively to Aop , Aop
and Sop
, S
, are isomorphic

respectively to A, A , S and S too, while L and L are isomorphic

9.7.

Relative maximality of dicartesian categories

213

respectively only to Lop and Lop

, and not to L and L . So the symmetry
between and is deeper in A, A, , S and S, than in L and L, .

9.7.

Relative maximality of dicartesian categories

The category L, is not maximal in the sense in which L, L and L are

maximal (see 9.3 and 9.5). This is shown with the following counterexample.
Let Set be the category whose objects are sets with a distinguished
element , and whose arrows are -preserving functions f between these
sets; namely, f () = . This category is isomorphic to the category of sets
with partial functions. The following denitions serve to show that Set is
a category of the C(L, ) kind:
I = {},

a = {(x, ) | x a I},

b = {(, y) | y b I},

a b = ((a I) (b I)) I,
a
2 b = (a b) a b ,
+ b = a b I.
a2
Note that a
2 b is isomorphic in Set to the cartesian product a b; the
element of a
2 b corresponds to the element (, ) of a b.

The functions k ia1 ,a2 : a1

2 a2 ai , for i {1, 2}, are dened by

k ia1 ,a2 (x1 , x2 ) = xi ,

k ia1 ,a2 () = ;

for fi : c ai , the function f1 , f2 : c a1

2 a2 is dened by
{
(f1 (z), f2 (z)) if f1 (z) = or f2 (z) =
f1 , f2 (z) =

if f1 (z) = f2 (z) = ;

and the function a : a I is dened by a (x) = . Having in mind

the isomorphism between a
2 b and a b mentioned above, the functions

i
k a1 ,a2 : a1
2 a2 ai correspond to the projection functions, while ,
corresponds to the usual pairing operation on functions.

+ a2 are dened by
The functions k ia1 ,a2 : ai a1 2

k 1a1 ,a2 (x) = (x, ),

k ia1 ,a2 () = ;

k 2a1 ,a2 (x) = (, x),

for x = ,

214

CHAPTER 9.

LATTICE CATEGORIES

+ a2 c is dened by
for gi : ai c, the function [g1 , g2 ] : a1 2

[g1 , g2 ](x1 , x2 ) = gi (xi ), for xi = ,

[g1 , g2 ]() = ;

nally, the function a : I a is dened by a () = .

+ , then it can be checked in a straightIf we take that is
2 and is 2
forward manner that Set and Set without I are lattice categories, and if
in Set we take further that both and are I, then Set is a dicartesian
category.
Consider now the category Set , which is obtained by adding to Set
the empty set as a new object, and the empty functions a : a as new
arrows. The identity arrow 1 is . For Set , we enlarge the denitions
above by

2a =a
2 = ,
+ a =a2
+ = a,
2

k ia1 ,a2 = ai , for a1 = or a2 = ,

a1 , a2 = a12
a2 ,

= I ,

k ia1 ,a2 = a1+

2 a2 , for ai = ,
[f1 , c ] = f1 ,
[c , f2 ] = f2 ,

and dene now the function a : a by a = a . Then it can be checked

+ as before, while is I and is , is
that Set where is
2 and is 2
a dicartesian category too.

In L, the equation k 1p, = p k 2p, does not hold, because G k 1p, =

and G(p k 2p, ) = , but in Set this equation holds, because both sides
are equal to . Since Set is not a preorder, we can conclude that L, is
not maximal.
Although this maximality fails, the category L, may be shown maximal in a relative sense (cf. the end of 9.3). This relative maximality result,
which we are going to demonstrate now, says that every dicartesian category that satises an equation f = g between arrow terms of C(L, ) such
that Gf = Gg (which implies that f = g is not in E(L, )) satises also
some particular equations. These equations do not give preorder in general,

9.7.

Relative maximality of dicartesian categories

215

but a kind of contextual preorder. Moreover, when E(L, ) is extended

with some of these equations we obtain a maximal natural logical category.
If for some arrow terms f1 , f2 : A B of C(L, ) we have Gf1 = Gf2 ,
then for some x GA and some y GB we have (x, y) Gf1 and (x, y)
Gf2 , or vice versa. Suppose (x, y) Gf1 and (x, y) Gf2 . Suppose the
(x+1)-th occurrence of letter in A, counting from the left, is an occurrence
of p. So the (y+1)-th occurrence of letter in B must be an occurrence of p.
Let A be the formula obtained from the formula A by replacing the
(x+1)-th occurrence of letter in A by p , and every other occurrence
of letter or by . Dually, let B be the formula obtained from B by
replacing the (y+1)-th occurrence of letter in B by p , and every other
occurrence of letter or by . Then it can be shown that there is an
arrow term hA : A A of C(L, ) such that GhA = {(0, x)}, and an
arrow term hB : B B of C(L, ) such that GhB = {(y, 0)}. We build

hA with k 1p, : p p and instances of C : C, with the help of

the operations and on arrow terms. Analogously, hB is built with

k 1p, : p p and instances of C : C . It can also be shown that

there are arrow terms j A : p A and j B : B p of C(L, ) such
that Gj A = Gj B = {(0, 0)}. These arrow terms stand for isomorphisms of
L, .
Then it is clear that for fi being
j B hB fi hA j A : p p ,
with i {1, 2}, we have Gf1 = {(0, 0)}, while Gf2 = . Hence, by Composition Elimination for GL, and by the functoriality of G, we obtain in
L, the equations

f1 = k 1p, k 1p, ,

f2 = p k 2p, = k 2p, p .

(This follows from Restricted Dicartesian Coherence too.) If we write 0,

for , which is equal to in L, , then in L, we have

f2 = k 2p, 0, k 2p, .
So in E(L, ) extended with f1 = f2 we can derive

216

CHAPTER 9.

(k k)

LATTICE CATEGORIES

k 1p, k 1p, = k 2p, 0, k 2p, .

The equation

(k )

k 1p, = p k 2p, ,

which holds in Set , and which we have used above for showing the non
maximality of L, , clearly yields (k k), which hence holds in Set , and
which hence we could have also used for showing this nonmaximality.
If we rene the procedure above by building A and B out of A and B

more carefully, then in some cases we could derive (k ) or its dual

(k )

k 1p, = k 2p, p

instead of (k k). We do not replace the x+1-th p by p in building A ,

and we can proceed more selectively with other occurrences of letters and
in A in order to obtain an A isomorphic to p if possible. We can proceed
analogously when we build B out of B to obtain a B isomorphic to p if
possible.
Note that we have the following:

p k 2p, = p , 1 k 2p,

= k 1p, , k 2p, , with (k ),

= 1p .

In the other direction, it is clear that the equation derived yields (k ). So

with (k ) we have that C and are isomorphic, and, analogously, with

(k ) we have that C and are isomorphic. It can be shown that the

natural logical category dened as L, save that we assume in addition

both (k ) and (k ) is maximal. (This is achieved by eliminating letterless

subformulae from C and D in g1 , g2 : C D such that Gg1 = Gg2 , and
falling upon the argument used for the maximality of L in 9.5.)

If f : a b is any arrow of a dicartesian category A and (k k) holds in

A, then we have in A

9.7.

Relative maximality of dicartesian categories

217

k 1b, f k 1a, = k 1b, k 1b, (f 1 )

= k 2b, 0, k 2a, ,
and hence for f, g : a b we have in A

(k k f g)

k 1b, f k 1a, = k 1b, g k 1a, .

So, although L, is not maximal, it is maximal in the relative sense

that every dicartesian category that satises an equation f = g between

arrow terms of C(L, ) such that Gf = Gg satises also (k k) and (k k f g).

Some of these dicartesian categories may satisfy more than just (k k) and

(k k f g). They may satisfy (k ) or (k ), which yields

f k 1a, = g k 1a,
and some may be preorders.

or

k 1b, f = k 1b, g,

Chapter 10
Mix-Lattice Categories
In this chapter we consider categories with nite products and coproducts
in which there is an operation of union on arrows with the same source
and target, so that hom-sets are semilattices with this operation. This is
what the mix principle of Chapter 8 amounts to in the present context. An
example of such a category is the category of semilattices with semilattice
homomorphisms.
We prove restricted coherence results for these categories, the restriction
being on the sources and targets of arrows, which must be in disjunctive or
conjunctive normal form. These coherence results are just an auxiliary for
the proofs of coherence in the next chapter. The technique of proof here is
again based on composition elimination.

10.1.

Mix-lattice categories and an example

To obtain the natural logical category ML, we have that the logical system
C(ML) is in L, , with the transformations included in 1, b, c, w-k and
m. The specic equations of E(ML) are obtained by taking the union of
those of E(MS) and E(L) plus
(wm)

wA mA,A wA = 1A .

We call natural ML-categories mix-lattice categories.

Let C(GML) be the syntactical system with the formulae of L, as
objects, with the primitive arrow terms being only identity arrow terms,
219

220

CHAPTER 10.

MIX-LATTICE CATEGORIES

and with the operations on arrow terms being those of C(GL) plus the
following one:
f:AB

g: A B

f g: A B
To obtain the equations of E(GML), we assume the equations of E(GL)
and the following equations:
( )

(f g) h = (f h) (g h),
( assoc)
( com)
( idemp)

h (f g) = (h f ) (h g),

f (g h) = (f g) h,
f g = g f,
f f = f.

The last equation, ( idemp), can be replaced by 1A 1A = 1A . The

syntactical category GML is C(GML)/E(GML).
It is straightforward to show (by relying on Lattice Coherence of 9.4)
that with the following denition in C(ML):

f g =df wB (f 3 g) wA
(f 3 g is (f g) mA,A , as in 8.1), and the following denition in C(GML):

mA,B =df K 1B K 1B 1A K 2A K 2A 1B ,
together with the denitions involved in showing the synonymity of C(L)
and C(GL), we have that ML and GML are isomorphic categories, and
that, hence, C(ML) and C(GML) are synonymous syntactical systems (see
the end of 2.4 for this notion of synonymity).
It can be checked that for the functor G from ML to Rel we have
G(f g) = Gf Gg,
where on the left-hand side is dened in C(ML) as above, and on the
right-hand side is union of relations with the same domain and codomain
(remember that GmA,B is an identity relation, i.e. identity function; see
2.9).

10.1.

Mix-lattice categories and an example

221

According to the equations ( assoc), ( com) and ( idemp), the homsets in any mix-lattice category are semilattices with the operation . In
ML the following equations hold:
( ) (f1 f2 ) (g1 g2 ) = (f1 g1 ) (f2 g2 ),
for {, }. The derivation of these equations is based on the following
equations of ML:

ckA,C,B,D mAC,BD c m
A,B,C,D = mA,B mC,D ,
m
c A,B,C,D mAC,BD ckA,B,C,D = mA,B mC,D ,
for whose checking we can use Semilattice Coherence of 9.1 (see 9.1 and

the List of Equations for the denitions of c m and c m , and 9.4 for the
denition of ck ).
As an example of a mix-lattice category, we have the category Setsl
,
whose objects are semilattices with unit a, , such that x y = i x =
and y = , and whose arrows are homomorphisms f with trivial kernels;
that is, f (x) = i x = . The unit may be conceived either as top or as
bottom. This category is a subcategory of the category Set of 9.7.
We dene a1 , , a2 , , as the semilattice with unit a1 a2 , , ,
where is as in 9.7, and we have
(x1 , x2 ) (y1 , y2 ) = (x1 y1 , x2 y2 ),
(x1 , x2 ) = (x1 , x2 ) = (x1 , x2 ),
= .
We dene a1 , , a2 , , as the semilattice with unit a1
2 a2 , , ,
where
2 , which corresponds to cartesian product, is dened as in 9.7,
and we have for and the same clauses as above.

The functions k ia1 ,a2 : a1 a2 ai , for i {1, 2}, are dened by

k ia1 ,a2 (x1 , x2 ) = xi ,

k ia1 ,a2 () = ;

for fi : c ai , the function f1 , f2 : c a1 a2 is dened by

{
f1 , f2 (z) =

(f1 (z), f2 (z)) if z =

if z = .

222

CHAPTER 10.

MIX-LATTICE CATEGORIES

The functions k ia1 ,a2 : ai a1

2 a2 are dened by

k 1a1 ,a2 (x) = (x, ),

k 2a1 ,a2 (x) = (, x),

for x = ,

k ia1 ,a2 () = ;
for gi : ai c, the function [g1 , g2 ] : a1
2 a2 c is dened by
[g1 , g2 ](x1 , x2 ) = g1 (x1 ) g2 (x2 ),
[g1 , g2 ]() = .

(The clauses in the denitions of k ia1 ,a2 and k ia1 ,a2 are taken over from 9.7,
and we could have also taken over from there the clause for f1 , f2 , but the
operations in the domains and codomains are changed, and the functions
dened are not the same; the clause for [g1 , g2 ] is new.)
We dene the function ma,b : a b a
2 b by
ma,b (x1 , x2 ) = (x1 , x2 ),

ma,b () = ,

or for the functions f, g : a b we dene the function f g : a b by

(f g)(x) = f (x) g(x).
It can be checked in a straightforward manner that with these denisl
tions, and with being and being
2 , the category Set is a mix-lattice
category (it is easier to rely on than on m in this context). A category
isomorphic to Setsl
is the category Semilat, whose objects are semilattices
and whose arrows are semilattice homomorphisms. We just reject from
the domains of the objects of Setsl
, and the pairs (, ) from the sets of
sl
ordered pairs of the arrows of Set . The mix-lattice structure of Semilat is
then inherited from Setsl
. The domain of a1 , , a2 , , is now a1 a2
instead of a1 a2 , while the domain of a1 , , a2 , , is (a1 a2 )+a1 +a2
instead of a1
2 a2 , which corresponded to a1 a2 (here + is disjoint union).
It is, however, more practical to introduce the mix-lattice structure in Setsl
,
with serving as an auxiliary, than directly in Semilat.
If we replace semilattices above by commutative semigroups, i.e., if we
reject the idempotency law, then we will verify all the specic equations of
ML except (wm) (which amounts to ( idemp)).

10.2.

10.2.

Restricted coherence of mix-lattice categories

223

Restricted coherence of mix-lattice categories

To prove a restricted coherence result for ML, we prove rst Composition

Elimination for GML by extending the proof for GL in 9.4. We use
essentially here the equations ( ).
Next, analogously to what we had in 9.6, we dene inductively formulae
of L, in disjunctive normal form (dnf ): every formula of L is in dnf,
and if A and B are both in dnf, then A B is in dnf. We dene dually
formulae of L, in conjunctive normal form (cnf ): every formula of L is
in cnf, and if A and B are both in cnf, then A B is in cnf.
We dene inductively composition-free arrow terms of C(GML) of type
A B, for A in dnf and B in cnf, that are in normal form. We do that
gradually, relying on two preliminary inductive denitions.
Arrow terms of the form P1 . . . Pn Q1 . . . Qm 1p , where n, m 0, and Pi

for i {1, . . . , n} is of the form K 1C or K 2C , while Qj for j {1, . . . , m} is

of the form K 1C or K 2C , are in atomic bracket-free normal form.

Every arrow term in atomic bracket-free normal form is in bracket-free
normal form. If f : D E and g : D E are in bracket-free normal form,
then f g : D E is in bracket-free normal form.
Every arrow term in bracket-free normal form is in angle normal form.
If f : D E and g : D F are in angle normal form, then f, g : D E F
is in angle normal form.
Every arrow term in angle normal form is in normal form. If f : E D
and g : F D are in normal form, then [f, g] : E F D is in normal form.
We have also the following denitions. Let f be an arrow term of
C(GML) in normal form, and let f be a subterm of f such that f is
in atomic bracket-free normal form, and there is no subterm f of f in
atomic bracket-free normal form with f a proper subterm of f . Then we
say that f is an atomic component of f .
An arrow term f of C(GML) in normal form is said to be in settled
normal form when there are no subterms of f in bracket-free normal form
in which an atomic component occurs more than once.
Let us illustrate all these denitions with an example. The following
arrow terms of C(GML):

224

CHAPTER 10.

MIX-LATTICE CATEGORIES

1 =df K 1q 1p : p q p,

2 =df K 1s K 2p 1q : p q q s,

3 =df K 1pt K 1q 1p : p q p (p t),

4 =df K 2p K 1t K 1q 1p : p q p (p t),
1 =df K 1s K 1p 1p : (p r) s p,
2 =df K 2q K 2pr 1s : (p r) s q s,

3 =df K 1pt K 1s K 1r 1p : (p r) s p (p t),

4 =df K 2p K 1t K 1s K 1r 1p : (p r) s p (p t)
are all in atomic bracket-free normal form. The arrow terms 2 2 , 3 4 ,
(2 2 ) (3 4 ), etc., are in bracket-free normal form. Next,
1 , 2 2 , 3 : p q p ((q s) (p (p t))),
1 , 2 , 3 4 : (p r) s p ((q s) (p (p t)))
are in angle normal form, and
=df [[1 , 2 , 3 , 1 , 2 , 3 4 ], 1 , 2 , 4 ] :
((p q) ((p r) s)) (p q) p ((q s) (p (p t)))
is in settled normal form. This normal from would not be settled if, for
example, 2 in were replaced by 2 2 . The set of occurrences of atomic
components of is made of the two occurrences of 1 , the two occurrences
of 2 , and of the occurrences of 3 , 4 , 1 , 2 , 3 and 4 .
We can then prove the following.
Normal-Form Lemma. Every arrow term f : A B of C(GML) for A
in dnf and B in cnf is equal in GML to an arrow term in settled normal
form.
Proof. We make an induction on the number of occurrences of in A
and in B. If there are no such occurrences of and , then we eliminate
compositions, and by applying the following equations of GML:

K iA (f g) = K iA f K iA g,

K iA K jB f = K jB K iA f,

10.2.

Restricted coherence of mix-lattice categories

225

we obtain an arrow term in bracket-free normal form equal to the original

arrow term.
If there are no occurrences of in A, and B is B1 B2 , then f =

1
K B2 1B1 f, K 2B1 1B2 f in GML, and, by the induction hypothesis, we

have that K 1B2 1B1 f and K 2B1 1B2 f must be equal respectively to f and
f in normal form, which must be in angle normal form, because does
not occur in A. Hence f = f , f , and f , f is in normal form.

If A is A1 A2 , then f = [f K 1A2 1A1 , f K 2A1 1A2 ] in GML, and,

by the induction hypothesis, f K 1A2 1A1 and f K 2A1 1A2 must be equal
respectively to f and f in normal form. Hence f = [f , f ], and [f , f ]
is in normal form.
We easily pass from the normal form to the settled normal form by
applying ( assoc), ( com) and ( idemp).

For an arrow term f of C(GML) in settled normal form, there is a oneto-one correspondence between the set of occurrences of atomic components
of f and the set of ordered pairs of Gf . For example, if f is the arrow term
we had above, then we have the following correspondence:
left 1

left 2

right 1

right 2

(0, 0)

(1, 1)

(0, 3)

(2, 0)

(4, 2)

(2, 3)

(2, 4)

(5, 0)

(6, 1)

(5, 4)

which can be drawn as follows:

((p

q)

((p

r)

s))

0
1
2
3
4
PP
H 4
l2
2 
PP 1 @
3
l1
3 H
HH
PP

@
PP
P
HHH

P@

P
 P@
H

0
1
2
3
4
p ((q s) (p (p

(p

q)
6

r1
r2

5
t)))

This one-to-one correspondence has a ner structure, which we are going

to explain now. For A in dnf let a minimal disjunct of A be a subformula
D of A that belongs to L such that there is no subformula of A in L of
which D would be a proper subformula. We dene analogously the minimal
conjuncts of a formula B in cnf, by replacing by .

226

CHAPTER 10.

MIX-LATTICE CATEGORIES

Let A be the source ((p q) ((p r) s)) (p q), and let B be the
target p ((q s) (p (p t))), of the arrow term we had as an example
above. Then the minimal disjuncts of A are A1 , which is the left occurrence
of p q, next A2 , which is (p r) s, and A3 , which is the right occurrence
of p q. The minimal conjuncts of B are B1 , which is the leftmost p, next
B2 , which is q s, and B3 , which is p (p t).
For an arrow term f : A B of C(GML) in normal form (not necessarily
settled), consider subterms in bracket-free normal form that are not proper
subterms of subterms of f in bracket-free normal form. We call such subterms the molecular components of f . There is a one-to-one correspondence
between the set of occurrences of molecular components of f and the set
of ordered pairs (Ai , Bj ) for Ai a minimal disjunct of A and Bj a minimal
conjunct of B. We call this correspondence the molecular correspondence.
For example, the molecular component 3 4 of corresponds by the
molecular correspondence to the ordered pair (A2 , B3 ).
If f : A B is an arrow term of C(GML) in settled normal form, then
for every molecular component f of f , the set of ordered pairs of Gf is in
one-to-one correspondence with the set of atomic components in f . We call
this correspondence the atomic correspondence. For example, if f is the
molecular component 3 4 of , then Gf = {(0, 0), (0, 1)}, where (0, 0)
corresponds by the atomic correspondence to 3 and (0, 1) corresponds to
4 .
We can then prove the following.
Restricted Mix-Lattice Coherence. Let f, g : A B be arrow terms
of C(ML) such that A is in dnf and B in cnf. If Gf = Gg, then f = g in
ML.
Proof. By the Normal-Form Lemma, we have that f = f and g = g in
GML for f and g in settled normal form. Since Gf = Gf and Gg = Gg ,
because G is a functor, we have that Gf = Gg implies Gf = Gg . If Gf =
Gg , then for the molecular components f of f and g of g such that f
and g correspond by the molecular correspondence to the same ordered
pair (Ai , Bj ), for Ai a minimal disjunct of A and Bj a minimal conjunct of
B, we must have Gf = Gg . Hence, by the atomic correspondence, there

10.3.

Restricted coherence of mix-dicartesian categories

227

is a one-to-one correspondence between the atomic components f in f

and the atomic components g in g such that f and g correspond to
the same ordered pair of Gf , that is Gg . Since Gf = Gg , we may
conclude, by Lattice Coherence, that f = g in GL, and hence also in
GML. (As a matter of fact, f and g must be the same arrow term of
C(GL).) Then, by using the equations ( assoc) and ( com), we must be
able to show that f = g in GML. Since this holds for every pair f and
g of corresponding molecular components, we obtain f = g , and so f = g
in ML.

We will not try to establish here an unrestricted coherence result for

ML, or perhaps a category with E(ML) extended. The result we have
above is sucient for applications in the next chapter, which are our main
concern.

10.3.

Restricted coherence of mix-dicartesian categories

To obtain the natural logical category ML, , we have that the logical
system C(ML, ) is in L,,, , with the transformations included in
1, b, c, w-k, m and -. The specic equations of E(ML, ) are obtained
by taking the union of those of E(ML) and E(L, ) plus

(m)

mA, = k 1A, k 1A, ,

(m)

mA, = k 1A, k 1A, .

It is easy to see that in E(ML, ) we have the equations

mA,C = k 1A,C

k 1A,C ,

mC,A = k 2C,A k 2C,A

for any letterless formula C of L,,, . It is clear that, by relying on the

equation (cm) of 8.4, we could replace (m) above by m,A = k 2,A k 2,A ,

while (m) could be replaced by m,A = k 2,A k 2,A .

We call natural ML, -categories mix-dicartesian categories.
The syntactical category GML, synonymous with ML, is obtained as GL, save that we have in addition the operation on arrow

228

CHAPTER 10.

MIX-LATTICE CATEGORIES

terms of the same type, and the equations ( ), ( assoc), ( com) and
( idemp) of E(GML) (see 9.6 and 10.1), plus the equations

(0) 1A K 2A A = 1A ,

(0) 1A K 2A A = 1A .
That (0) holds in ML, is shown as follows:

1A (k 2A, A )

= wA (1A k 2A, ) mA, (1A A ) wA , by (m nat),

= 1A , with (m).
We proceed analogously for (0) by using (m).
To show that (m) holds in GML, , we have

mA, = K 1 K 1 1A K 2A K 2A 1

= (1A K 2A A ) K 1 K 1 1A , with ( ),

= K 1 K 1 1A , by (0) and (cat 1).

We proceed analogously for (m) by using (0).
The category Setsl
has a terminal object , which is the two-element
semilattice {, x}; this is the free semilattice with unit with a single free
generator x. The initial object of Setsl
is the trivial semilattice with unit
{}; this is the free semilattice with unit with an empty set of generators.

The function a : a is dened by

{
x if y =

a (y) =
if y = ,

while for a : a we have a () = . The category Setsl

, with the structure dened in 10.1 and here, is a sesquicartesian category in the L sense,
but not a dicartesian category, because in Setsl
the object is not isosl
morphic to . Note that in Set the equation (m) holds, but (m) does
not hold. In the category Semilat, which is isomorphic to Setsl
, the terminal
object is the trivial semilattice with a single element, while the initial
object is the empty semilattice, i.e. the empty set.
To prove restricted coherence for ML, , we need rst some preliminary
notions.

10.3.

Restricted coherence of mix-dicartesian categories

229

A null term is an arrow term g : A B of C(L, ) such that Gg is the

empty relation. Let C be a formula of L,, and D a formula of L,, .
Then g : C D is a null term only if for some C of L and some D of L
we have that either C is isomorphic to C or D is isomorphic to D
in L, . This follows easily from Composition Elimination for GL, (see
9.6).
To show that for any such null term g : C D, where i : C C is
an isomorphism of L, , and for any arrow term f : C D of C(L, ), we
have in ML, the equation
(0g)

f g = f,

we rely on the following instance of (0):

1C (C k 2C , ) = 1C .
From this equation we obtain

f i (1C (C k 2C , )) i1 = f i i1 ,

and (0g) follows with the help of ( ) and Lemma 2 of 9.6 (we have

G(f i C k 2C , i1 ) = Gg). We proceed analogously to derive

(0g) for a null term g : C D where D is isomorphic to D in L, ,
and any arrow term f : C D of C(L, ).
Relying on the denition of dnf and cnf of 9.6, we have the following.
Restricted Mix-Dicartesian Coherence. Let f, g : A B be arrow
terms of C(ML, ) such that A is in dnf and B in cnf. If Gf = Gg, then
f = g in ML, .
To prove this result, we proceed as follows. First, by extending the proof
of Composition Elimination for GL, (see 9.6), we obtain Composition
Elimination for GML, .
We dene inductively composition-free arrow terms of C(GML, ) of
type A B, for A in dnf and B in cnf, that are in normal form. The only
dierence with respect to the denition we had in the preceding section

is that arrow terms in atomic bracket-free normal form can now have p ,

230

CHAPTER 10.

MIX-LATTICE CATEGORIES

p , , or 0, instead of 1p ; here 0, stands for either or ,

which are equal in L, . Arrow terms in atomic bracket-free normal form

in which we do not have 1p , but p , p , , or 0, are called zero

atomic bracket-free terms, and those with 1p nonzero atomic bracket-free
terms. We use the same terminology of zero and nonzero for atomic
components. Zero atomic bracket-free terms are null terms in the sense
specied above, and all such arrow terms of the same type are equal in
L, by Lemma 2 of 9.6.
An arrow term f of C(GML, ) in normal form is in settled normal
form when, as before, there are no subterms of f in bracket-free normal form
in which an atomic component occurs more than once, and, moreover, we
do not have subterms of f of the form g h where one of g and h is a
zero atomic component. There is a one-to-one correspondence between the
set of occurrences of nonzero atomic components of an arrow term f of
C(GML, ) in settled normal form and the set of ordered pairs of Gf .
Then, by proceeding as in the preceding section, we can prove the
Normal-Form Lemma where GML is replaced by GML, . We use here

also the equations 1 = and 1 = . To pass from the normal form

to the settled normal form we apply the equations ( assoc), ( com),
( idemp) and (0g). We can then prove Restricted Mix-Dicartesian Coherence as Restricted Mix-Lattice Coherence in the preceding section. In
that proof we use Restricted Dicartesian Coherence where previously we
used Lattice Coherence. As for ML, we will not try to establish here
an unrestricted coherence result for ML, , or perhaps a category with
E(ML, ) extended.
We will not discuss here the maximality of ML and ML, , but we
conjecture that ML is not maximal in the sense in which L was (see 9.5
and 9.3). For example, one could presumably add to E(ML) the equation

mp,p = mp,p c p,p , where Gmp,p = G(mp,p c p,p ), without falling into
preorder. There are other such equations, but we will not go here into the
problem of their classication.

Chapter 11
Distributive Lattice Categories
This is the central chapter of the book. We dene in it the notion that we
take as the categorication of the notion of distributive lattice. Distribution is here based on the dissociativity of Chapter 7, which delivers arrows
corresponding to the common distributions of conjunction over disjunction
and of disjunction over conjunction, but neither of these distributions happens to be an isomorphism (in bicartesian closed categories, the former
distribution is an isomorphism, but the latter is not). For our categorication of distributive lattices, we prove coherence with respect to the category
whose arrows are relations between nite ordinals, as before. We have this
coherence both in the presence and in the absence of terminal and initial
objects.
The essential ingredient of our proof is a cut-elimination theorem for a
category corresponding to a plural sequent system for classical conjunctivedisjunctive logic. This category is obtained by strictifying the double cartesian structure so that arrows of the monoidal structure, i.e. associativity
isomorphisms and isomorphisms tied to the terminal and initial objects,
become identity arrows. This is very much in the spirit of Gentzen, who
based his sequents on sequences of formulae, and Gentzens intuition is here
vindicated by the strictication results of Chapter 3. Our cut-elimination
procedure diers, however, from Gentzens in that it takes into account
union of proofs. Gentzens own procedure would lead to collapse, i.e. preorder. We also dier from Gentzen in how we deal with the structural rule
of contraction. We eliminate cut directly, and do not introduce Gentzens
231

232

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

generalized cut rule, which may be understood as involving several cuts, or

as blending contraction with cut. (Our approach here diers from previously published procedures of eliminating cut directly.)
We believe that one of the achievements of this chapter is notational.
From the very beginning of categorial proof theory, equations imposed by
cut elimination have been a guiding inspiration, but recording these equations precisely proved to be a rather dicult task.

11.1.

Distributive lattice categories and their Gentzenization

The categories we are going to investigate in this chapter, which we call

distributive lattice categories, may be conceived as obtained by the categorication of the notion of distributive lattice. Freely generated categories
of this kind may be conceived as codifying equality of derivations in the
conjunction-disjunction fragment of logic (with or without the empty conjunction and the empty disjunction ). This fragment of logic coincides
in classical and intuitionistic logic, as far as provable sequents of the form
A B are concerned (cf. 1.3). Categories we have considered previously
codify analogously equality of derivations in more restricted fragments of
logic, which were sometimes fragments of nonclassical and nonintuitionistic
logics. In particular, the free symmetric net category of 7.6 corresponds
to a fragment of linear logic (in the jargon of that eld, we have there the
multiplicative conjunction-disjunction fragment of linear logic).
It is remarkable that equations between arrows in the free distributive lattice category cover a procedure of cut elimination in a plural, i.e.
multiple-conclusion, sequent system. A sequent is a singular, or
single-conclusion, sequent when has a single member or is empty; without this restriction, it is a plural, or multiple-conclusion, sequent. The fact
that we are within the realm of plural sequents for conjunctive-disjunctive
logic allows us to assume that we are dealing with classical, rather than
intuitionistic, logic (see the last part of 1.3).
Gentzens cut-elimination theorem of [60] could be phrased as saying
that for every term t coding a derivation of there is a term t coding
a cut-free derivation of . As in the Cut-Elimination Theorem of

11.1.

Distributive lattice categories and their Gentzenization

233

7.7, we prove something more in this chapter. We show also that t = t

in a particular category D. Gentzen did not care about equality between
these termshe did not even introduce terms to code his derivations. His
intuition was, however, good in most cases, and we may copy his procedure
to a great extent. But we cannot copy it completely, because if we did so,
then our category D would be a preorder.
We want equality of arrows in D to correspond to equality of arrows
in the freely generated distributive lattice category, and D should not be
a preorder. Therefore our cut-elimination procedure will not be exactly
Gentzens procedure of [60] restricted to conjunctive-disjunctive logic, but
a modication of it, and we will point out later where precisely we dier
from Gentzen. The main dierence is that we take into account the mix
principle, which in this context yields the operation of union of derivations,
corresponding to the operation on arrow terms of 10.1. The problematic
situation in [60], mentioned at the beginning of 1.6, was noted in [64] (Appendix B1, by Y. Lafont), where it was supposed that there is no alternative
to Gentzens way of dealing with it, and that preorder and triviality are inescapable in the proof theory of classical logic (see also [67], Section 1). We
show that this is not the case, and obtain a coherence result for distributive
lattice categories with respect to the category Rel.
Distributive lattice categories are not the only candidate for codifying
equality of derivations in conjunctive-disjunctive logic. An alternative codication is in a fragment of bicartesian closed categories. The equations of
these categories also cover a cut-elimination procedure in a single-conclusion
sequent system (see [109]). With this alternative codication, we do not
have, however, a coherence result with respect to Rel (see 1.2, [48], Section 1, and [109], Section 1; cf. [76], pp. 95-97). The distribution arrow of
type A (B C) (A B) (A C) is an isomorphism in bicartesian
closed categories. In our distributive lattice categories we have an arrow of
this type, and of the inverse type, but we need not have an isomorphism of
this type. We pass now to the denition of distributive lattice category.
To obtain the natural logical category DL, we have that the logical
system C(DL) is in L, , with the transformations included in 1, b, c,
w-k, m and d. The specic equations of E(DL) are obtained by taking the
union of those of E(DS) and E(ML) plus

234

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

k 2A,BC = (k 2A,B 1C ) dA,B,C ,

(d k)

(d k)

k 1CB,A = dC,B,A (1C k 1B,A ),

(dm)

mA,C = (k 1A,B 1C ) dA,B,C (1A k 2B,C ),

ckA,C,B,D e A,B,C,D = mA,B 1CD ,

e D,C,B,A ckD,C,B,A = 1DC mB,A

(m e)

(m e)

(see 7.6 for e , and 9.4 for ck ). In the arrow terms of C(DL) we write d
instead of dL , as we did for C(DS) and C(MDS), and we take dR as dened
by the equation (dR c) of 7.6.

We call natural DL-categories distributive lattice categories. The objects of a distributive lattice category that is a partial order make a distributive lattice. In DL, the dissociativity arrows dA,B,C enable us to dene
arrows of the type of the common distribution principles of over and of
over (see 11.3). These distribution arrows are, however, not isomorphisms. Note that our distributive lattice categories are not distributive
categories in the sense of [95] (pp. 222-223 and Session 26) or [20], where
distribution of over must be an isomorphism.
The cartesian linearly distributive categories of [22] are symmetric net
categories and are lattice categories, but they are not necessarily distributive lattice categories. The specic equation (wm) is not envisaged in that

paper, nor in [23]. The equations (d k) and (d k) hold in cartesian linearly

distributive categories as a consequence of the presence of the equations

( dL ) and ( dL ) of 7.9 in these categories (see 11.5). The equations

(dm), (bm), (cm), (m e) and (m e) hold in these categories (as can be gathered from the derivations for DL below and from 11.5), though they are
not explicitly mentioned in [22] and [23]. We know the equation (wm) need
not hold in these categories (see 12.5).

We do not know how to derive (m e) and (m e) from the remaining

axioms of E(DL). We can however derive from the remaining axioms the

following immediate consequences of (m e) and (m e):

ckA,C,B,D e A,B,C,D (1AB mC,D ) = mA,B mC,D ,

(mD,C 1BA ) e D,C,B,A ckD,C,B,A = mD,C mB,A

11.1.

Distributive lattice categories and their Gentzenization

235

(see the derivation of (m c m ) in this section, and 13.2). Equations that

for the presentation of DL could replace (m e) and (m e) are

(k 1A,C k 1B,D ) e A,B,C,D = mA,B k 1AB,CD ,

e D,C,B,A (k 2D,B k 2C,A ) = k 2DC,BA mB,A ,

and equations that could replace (m e), (m e) and (wm) are the following
two equations:
(wm e) (wA 1CD ) ckA,C,A,D e A,A,C,D (wA 1CD ) = 1A(CD) ,

(wm e) (1DC wA ) e D,C,A,A ckD,C,A,A (1DC wA ) = 1(DC)A .

It is clear that these two equations follow from (m e), (m e) and (wm).

To show the converse, for f being (k 1A,C k 1B,D ) e A,B,C,D and g being

(wAB 1CD ) ckAB,C,AB,D , we have

f g = f w(AB)(CD) ((1AB k 1C,D ) (1AB k 2C,D )),

by Lattice Coherence of 9.4,

= mA,B wAB (k 1AB,C k 1AB,D ), with (dm) and by

Symmetric Net Coherence of 7.6,

= mA,B k 1AB,CD g, by Lattice Coherence,

and then we apply (wm e) to obtain (m e). We proceed analogously for

(m e). From either (wm e) or (wm e) we derive (wm) by applying (m e) or

(m e) (see 11.5 and 13.2 for further comments on (m e) and (m e)).

There are redundancies in our presentation of DL. A synonymous logical system C(DL ) is obtained by omitting m from C(DL). We introduce
m in C(DL ) by (dm) understood as a denition. The specic equations of
E(DL ) are obtained by taking the union of those of E(DS) and E(L) plus

(wm), (d k), (d k), (m e) and (m e), where m is dened. The category DL

is C(DL )/E(DL ). To prove this synonymity, i.e. the isomorphism of DL
with DL (see the end of 2.4 for the notion of synonymity of syntactical
systems), we have the following.
First, we derive for DL the equation

(k 1A,D 1C ) dA,D,C (1A k 2D,C ) = (k 1A,B 1C ) dA,B,C (1A k 2B,C ),

which yields the equation (dm) for DL , because the left-hand side may be
replaced by mA,C . We have, by using naturality equations,

236

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

(k 1A,D 1C ) dA,D,C (1A k 2D,C )

= (k 1A,D 1C ) dA,D,C (1A (k 1D,B 1C )) (1A k 2DB,C )

= (k 1A,DB 1C ) dA,DB,C (1A k 2DB,C )

= (k 1A,B 1C ) ((1A k 2DB ) 1C ) dA,DB,C (1A k 2DB,C )

= (k 1A,B 1C ) dA,B,C (1A k 2B,C ).

Next, we derive the equation (b mL) (see 8.3) for DL :

dA,B,C (1A mB,C )

= dA,B,C (1A (k 1B,D 1C )) (1A dB,D,C )

(1A (1B k 2D,C )), by (dm),

= ((1A k 1B,D ) 1C ) (b
A,B,D 1C ) dAB,D,C b A,B,DC

(1A (1B k 2D,C )), by naturality and (dL ) of 7.2,

= (k 1AB,D 1C ) dAB,D,C (1AB k 2D,C ) b

A,B,C ,

by (b k 1 ) of 9.1 and naturality,

= mAB,C b
A,B,C , by (dm),

and we proceed analogously for (b mL). Hence we have also (b mR) and

(b mR) (see 8.3). We can then derive the equation (bm) (see 8.2) for DL
as we derived it for MDA (see 8.3).
We derive as follows the equation (cm) (see 8.4) for DL . We have the
equation

1
2

(wA 1C ) ((1A k 1A,C ) 1C ) (dR

A,A,C 1C ) dAA,C,C (k A,A k C,C ) =

1
2

(1A wC ) (1A (k 2A,C 1C )) (1A dA,C,C ) dR

A,A,CC (k A,A k C,C ),

by (d b) (see 7.2) and Lattice Coherence of 9.4. We obtain that the two
sides of this equation are equal respectively to the two sides of the following
equation:

(k 1A,C 1C ) dA,C,C (1A k 2C,C ) = (1A k 2A,C ) dR

A,A,C (k A,A 1C ),

by using (d k), Lattice Coherence, naturality and bifunctorial equations,

and from that equation we derive (cm).

11.1.

Distributive lattice categories and their Gentzenization

237

It is easy to derive (m nat) for dened m in E(DL ), so that we have

in E(DL ) all the equations of E(MDS). We have also in E(DL ) all the
equations of E(ML). Since all the equations of E(DL ) are clearly in E(DL),
we obtain that DL and DL are isomorphic.
Note that for DL we can derive

mA,C = (1A k 2B,C ) dR

A,B,C (k A,B 1C ),

which is related to (cm). We can also derive for DL the following equations:

(m c m ) mAC,BD c m
A,B,C,D = e A,B,C,D (1AB mC,D ),

cm
D,C,B,A mDB,CA = (mD,C 1BA ) e D,C,B,A ,

(m c m )

which we will use in 11.2 (see also 13.2). Here is a derivation of (m c m ):

mAC,BD c m
A,B,C,D

= mAC,BD b
A,C,BD (1A b C,B,D ) (1A ( c B,C 1D ))

(1A b
B,C,D ) b A,B,CD

= dA,C,BD (1A mC,BD ) (1A c BD,C ) (1A b

B,D,C )

(1
b
AB c C,D ), by ( b mL) and
A,B,DC
Symmetric Biassociative Coherence of 6.3,

= dA,C,BD (1A c C,BD ) (1A dB,D,C ) (1A (1B mD,C ))

b
A,B,DC

(1AB c C,D ), by (cm) and (b mL),

= e A,B,C,D (1AB mC,D ), by (b nat) and (cm),

and we proceed analogously for (m c m ).

Let C and C be respectively the logical systems C(DL) and C(A), while
E is E(DL). Next, let B be C/E, that is DL. Then it is easy to see that
the conditions (IV C) and (IV B) of 3.1 are satised. Since the C -core of
C/E is the category A, by Biassociative Coherence of 6.1, we have that
the condition (IV G) of 3.1 is also satised. So (IV) holds, and since A is a
preorder by Biassociative Coherence, we can apply the Strictication Theorem of 3.1 to obtain that the category DLA , that is DLG , is equivalent
to DL via two strong C(DL)-functors. Remember that according to 6.2
the objects of DLA may be identied with form sequences of L, in the
nonextended sense. (For understanding the category DLA , see also 4.5.)

238

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

Let D be the category obtained as the disjoint union of DLA and of the
trivial category with a single object (intuitively this is the empty form
sequence of both colour) and a single arrow 1 : . The adding of to
DLA is made for practical reasons, to simplify the exposition of our cutelimination proof by subsuming several cases under a single schema. We
could also do without at the cost of considering more cases in the proof.
The operations { , } on the objects of DLA , i.e. on the form
sequences of L, , are extended in the following manner to operations that
apply also to . For X an object of DLA or , we have
X =

= X.

So all the objects of D are closed under the operations . The operations
{ , } on arrows are extended to operations that apply also to 1
by stipulating that
f

= 1

=f

(the variable f here ranges also over 1 ). So all the arrows of D are closed
under the operations .
The category D will not have the structure of a DL-category. We lack

in D the arrows k 1,X and k 2X, for X dierent from . However, w may be

identied with 1 , and the arrows b

X,Y,Z , bX,Y,Z and cX,Y where one of the
subscripts stands for may also be identied with identity arrows. (So D
would have the structure of something that could be called a relevant net
category; see [43] and [108]; cf. 14.4).
A basic sequence of colour {, } of D is either a form sequence
of L, of the form (A1 . . . An , ), for n 2 and Ai , where i {1, . . . , n},
a formula of L, , or it is a formula of L, , or it is . So the object
, as well as the formulae of L, , is both of colour and of colour .
A basic sequence is a basic sequence of either colour. The members of a
basic sequence (A1 . . . An , ) are the occurrences of formulae A1 , . . . , An ;
the only member of the basic sequence A is A; and the basic sequence
has no members.
We use and , with or without indices, as variables for basic sequences
of colour , and we use and , with or without with indices, for basic

11.1.

Distributive lattice categories and their Gentzenization

239

sequences of colour . For basic sequences in general, we use and ,

with or without indices. We write 1 for (1F , , ) when is not ;
otherwise, 1 is 1 .
A sequent arrow of D is an arrow whose type is , for and
basic sequences dierent from . According to the convention above, is
of colour and is of colour . The type of a sequent arrow is a sequent
(this agrees with Gentzens notion of sequent).
Let 1 2 . . . n be an abbreviation for (. . . (1 2 ) . . . n ), and
f1 f2 . . . fn an abbreviation for (. . . (f1 f2 ) . . . fn ), where n 2.
Next, let f and f be abbreviations for 1 f and f 1 , respectively. Sometimes we will also write, ambiguously, 1 f and f 1 for 1 f
and f 1 , where can be recovered from the context.
We use the following abbreviations:
{

c ,
if = and =
c, =df
1
if = or = ,
{

w
if =
w =df
1
if = ,

y 1 ,2 ,, =df (1 c 2 , ) (1 2 w ) :
1 2 1 2 ,

y ,,2 ,1 =df ( w 2 1 ) ( c ,2 1 ) :
2 1 2 1 .
For n 3, consider the abbreviations dened inductively as follows:

y 1 ,2 ,3 ,...,n ,,

=df y 1 ,2 ,3 ... n , y 1 2 ,3 ,...,n ,, :

1 2 3 . . . n 1 2 3 . . . n ,

y ,,n ,...,3 ,2 ,1 =df y ,,n ,...,3 ,2 1

y , n ... 3 ,2 ,1:

n . . . 3 2 1 n . . . 3 2 1 .
We also have the following abbreviations:
{
d,A,
if = and =
d,A, =df
1 (A ) if = or = ,

240

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

e2 ,1 ,B,A =df d2 ,B,1 A (2 c B,1 A ) (2 d1 ,A,B ) (2 1 c A,B ) :

2 1 (B A) (2 B) (1 A),

eA,B,1 ,2 =df ( c B,A 1 2 ) (dB,A,1 2 ) ( c A1 ,B 2 ) dA1 ,B,2 :

(A 1 ) (B 2 ) (A B) 1 2

(note that e,,B,A = 1B A and eA,B,, = 1AB ),

m, =df (k 1 ,B ) d,B, ( k 2 B, )
(note that for m, we must have = and = ),

ck1 ,2 ,3 ,4 =df w(1 3 ) (2 4 ) ((k 1 1 ,3 k 1 2 ,4 ) (k 2 1 ,3 k 2 2 ,4 ))

(note that none of 1 , 2 , 3 and 4 can be in the subscripts of ck ).
Finally, let 1iA B stand for (1AB , A B, A B), while 1eA B stands for

(1AB , A B, A B). We do not introduce the notation k1, , because we

could not interpret it when is and is not , and analogously with

k2, .
We will now dene by induction a set of terms for sequent arrows of D,
which we call Gentzen terms. First, we stipulate that for every letter p the
term 1p : p p, which denotes the arrow (1p , p, p) of D, is a Gentzen term.
The remaining Gentzen terms are obtained by closing under the following
operations on Gentzen terms, which we call Gentzen operations. As in
7.7, we present these operations by inductive clauses in fractional notation,
which are interpreted as saying that if the terms above the horizontal line
are Gentzen terms, then the term below the horizontal line is a Gentzen
term (cf. 2.2). The schema on the left-hand side of the =dn sign stands for
the Gentzen term, while the schema on the right-hand side stands for the
arrow denoted by this term. First, we have the operations that correspond
to Gentzens structural rules:
f : 1 2 3 4

cL
1 ,2 ,3 ,4 f =dn f (1 c 3 ,2 4 ) : 1 3 2 4

provided 2 and 3 are not (one of the indices 1 and 4 is superuous

as an index of cL ; it is recoverable from the source of f ),

11.1.

Distributive lattice categories and their Gentzenization

241

f : 4 3 2 1

cR
4 ,3 ,2 ,1 f =dn (4 c 2 ,3 1 ) f : 4 2 3 1

provided 2 and 3 are not (one of 1 and 4 is superuous as an index

of cR ),
f:

L
kA
f =dn f k 1 ,A : A

f:

R
kA
f =dn k 2 A, f : A

f : 1 A . . . n A

wL1 ,...,n , f =dn f y 1 ,...,n ,,A : 1 . . . n A

, n 2,

f : A n . . . A 1

R
w,
f =dn y A,,n ,...,1 f : A n . . . 1
n ,...,1

f : 3 2 A 1

, n 2,

g : 1 A 2 3

cut 2 ,2 (f, g) : 1 2 3 3 2 1
where cut 2 ,2 (f, g) denotes

(g 2 1 ) ((1 c 2 ,A ) 2 1 ) d1 2 ,A,2 1

( ( c
1
2
A,2 1 )) (1 2 f ),
and A is called the cut formula of cut 2 ,2 (f, g),
f : 1 1

g : 2 2

mix (f, g) =dn (g 1 ) m2 ,1 (2 f ) : 2 1 2 1

Note that by (m nat) in D we have
mix (f, g) = m2 ,1 (g f ) = (g f ) m2 ,1 .

242

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

If we write d2 ,,1 instead of m2 ,1 , then mix (f, g) can be conceived as

cut , (f, g) where the cut formula A is replaced by .
Had we favoured dR , rather than dL , for f : 1 3 A 2 and g :
2 A 3 1 we could take that cut 1 ,1 (f, g) : 1 2 3 3 2 1
denotes

(3 2 g) (3 2 ( c A,2 3 )) dR
3 2 ,A,2 3

((
3 c 2 ,A ) 2 3 ) (f 2 3 )

where dR stands for dR if 3 2 and 2 3 are not , and otherwise for

1. This would prevent the s and s of f and g to switch from right to
left, as in our ocial denition of the Gentzen operation cut . But since
we favour dL , we have to tolerate this switch, which does not cause serious
trouble, anyway. We have made the same switch in our Gentzen operation
mix , to make it parallel to our cut .
Here are the remaining Gentzen operations, which correspond to rules
for and :
f : A B
L f =dn f ( 1eAB ) : (A B)
f : 1 A 1

g : 2 B 2

R (f, g) =dn (1iAB12 ) eA,B,1 ,2 (f g) : 1 2 (A B)12

g : 2 B 2

f : 1 A 1

L (g, f ) =dn (g f ) e2 ,1 ,B,A (2 1 1eB A ) : 2 1 (B A) 21

f : B A
R f =dn (1iB A ) f : (B A)
This concludes the list of Gentzen operations.
For n 2, we introduce the following abbreviations by induction:
wL1 ,...,n , f =df f ,

for f : 1 . . . n ,

11.1.

Distributive lattice categories and their Gentzenization

243

wLA
f =df wL1 ,...,n , wL
f,

1 ,...,n ,
1 A,...,n A,

for f : 1 A . . . n A ,
for f : n . . . 1 ,

R
w,
f =df f ,
n ,...,1

R
R
R A
f =df w,
w,A
w,
,...,A f ,
n ,...,1
n ,...,1
n
1

for f : A n . . . A 1 .
By Semilattice Coherence of 9.1 (in fact, we use here the relevant coherence
result of [108], Section 5), we have in D the equations

(w y) wL
f = f y 1 ,...,n ,, ,
1 ,...,n ,

R
w,
f = y ,,n ,...,1 f.
n ,...,1

To lighten the burden of notation, in proofs we will sometimes omit

subscripts in Gentzen terms or other terms for arrows of D. A reader
checking the proofs should be able to restore these subscripts. We will also
sometimes take for granted the subscripts of Gentzen operations, and omit
them. We do this in cases where no confusion is likely, and the subscripts
serve no particular purpose. We use , 1 , 2 , . . . as variables for Gentzen
operations (with subscripts omitted or not).
Note that Gentzen terms codify derivations in a plural sequent system
for conjunctive-disjunctive classical propositional logic. (We have mentioned at the beginning of the section that we believe that we are within
classical, rather than intuitionistic, logic; cf. also 1.3.) We have in this
sequent system rules for connectives of the multiplicative kind, to use the
terminology of linear logic. In this terminology, Gentzens rules for conjunction and disjunction of [60] would be called additive. This is not, however,
an essential dierence. We could have worked with additive rules as well.
We took multiplicative rules for practical reasons, which have to do with
our way of dealing with the structural rule of contraction. This dierence
does not bar comparing our cut-elimination procedure with Gentzens, and
it will turn out that, though the two procedures have much in common,
they are not the same.
The main dierence is that we take into account the mix principle, which
yields union of derivations (see 8.1 and 10.1). Gentzen did not take this
principle into account, because, for his more limited purposes, he did not

244

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

need to do so. This mix principle should not be confused with Gentzens
generalized cut, Mischung, also called mix in English (see 8.1), which is
derivable in our system with the help of contractions, that is wL and wR .
We also dier from Gentzen in the way how we deal with contraction,
embodied in the operations wL and wR . We eliminate cut directly, and do
not introduce as Gentzen his generalized cut Mischung, which he used to
deal with problems caused by contraction (see [14], Sections 1 and 2). Eliminating cut directly is handier for notational reasons, because Gentzens
Mischung is more dicult to code in our categorial setting. (Our procedure of direct cut elimination diers from similar procedures in [12], [14],
[130] and [13]; except for [14], where categories are mentioned occasionally, these papers are not concerned with categorial proof theory and the
diculties of notation for arrow terms.)
Another dierence with Gentzen is that we distinguish conjunctive
commas, our (which abbreviates ), from disjunctive commas, our
(which abbreviates ), whereas Gentzen has just one kind of comma. In
other words, we have two-coloured form sequences, whereas Gentzen has
just ordinary sequences. Indeed, if we stay at the level of sequent arrows of
D, then the fact that a sequence is on the left-hand side or on the right-hand
side of dictates whether it is of colour or , and then we could do as
Gentzen. But we do not pay attention only to sequents, as Gentzen does.
For example, in building a sequent arrow denoted by cut 2 ,2 (f, g) we refer

to arrows of D like c 2 ,A , or d1 2 ,A,2 1 , etc., which are not sequent

arrows. With d, we even have that conjunctive commas are nested within
disjunctive commas and vice versa. Gentzen did not have these problems
because he was not considering explicitly arrows and equality between them,
but only types of arrows and, moreover, just sequent types. Gentzen stays
somewhere near the lowest level of D, while we take somewhat more of D
into account when we compute equality of sequent arrows.
In principle, we could have worked with directly strictied DL in the
sense of 3.2, but then we would be less close to Gentzen. In that case, we
would not have operations corresponding to L and R , but at the price of
complications in the computation of rank. We do this computation below
very much in the style of Gentzen. (Cut elimination in something corresponding to our directly strictied DL, but without taking into account

11.1.

Distributive lattice categories and their Gentzenization

245

equality of derivations, may be found in [16].)

Instead of directly strictifying, we have produced D according to the
recipe of 3.1, 4.5 and 6.2. We nd it is interesting to locate Gentzens
sequents within this strictied biassociative structure constructed in the
style of Joyal, Street and Mac Lane (the last author was close to Gentzen
in his youth). This tells us that Gentzen had a sound premonition that
nothing is lost by strictifying with respect to associativity.
Every arrow of D denoted by a Gentzen term is a sequent arrow. We
show in the following lemma that these are all the sequent arrows of D.
Every sequent arrow of D is denoted by a

Gentzenization Lemma.
Gentzen term.

Proof. We prove rst that every sequent arrow (f, A, B) of D is denoted

by a Gentzen term. After that we will pass to the sequent arrows (f, , )
for and with more than one member.
We show by induction on the lenght of A that (1A , A, A) is denoted by
a Gentzen term. If A is p, then (1p , p, p) is denoted by 1p . If A is A1 A2 ,
and (1Ai , Ai , Ai ) is denoted by the Gentzen term 1Ai , for i {1, 2}, then
(1A , A, A) is denoted by L R (1A1 , 1A2 ). If A is A1 A2 , then (1A , A, A)
is denoted by R L (1A1 , 1A2 ). We write, in general, 1A for the Gentzen
term denoting (1A , A, A).

We have that (b
A,B,C , A (B C), (A B) C) is denoted by
L L R (R (1A , 1B ), 1C ),
according to Associative Coherence of 4.3.

The inverse arrow (b

A,B,C , (A B) C, A (B C)) is denoted by

L L
R
R
L cL
,C,AB, c,A B,C, (1A , (1B , 1C )),

according to Symmetric Associative Coherence of 5.1.

R
We have that ( c A,B , A B, B A) is denoted by L cL
,B,A, (1B , 1A ),

and (wA , A, A A) is denoted by wL R (1A , 1A ). Next, (k 1A,B , A B, A) is

L
L
denoted by L kB
1A , while (k 2A,B , A B, B) is denoted by L cL
,B,A, kA 1B .

We proceed analogously for b , b , c, w and k i by using L and R .

246

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

We have that (dA,B,C , A (B C), (A B) C) is denoted by

R L L (R (1A , 1B ), 1C )

or

L R R (1A , L (1B , 1C )).

If the Gentzen terms f and g denote the sequent arrows (f, A, B) and
(g, C, D) respectively, then L R (f , g ) denotes (f g, A C, B D),
while R L (f , g ) denotes (f g, A C, B D). If the Gentzen terms f
and g denote the sequent arrows (f, A, B) and (g, B, C) respectively, then
cut , (f , g ) denotes (g f, A, C).
Take now a sequent arrow (f, A1 . . . An , Bm . . . B1 ) of D where
n, m 2. We have proved above that, for F dened as in 4.5, the sequent
arrow (f, F (A1 . . . An ), F (Bm . . . B1 )) is denoted by a Gentzen term
f . Then for g and h being respectively R (. . . R (1A1 , 1A2 ) . . . , 1An ) and
L (1Bm , . . . L (1B2 , 1B1 ) . . .), the Gentzen term cut , (cut , (g, f ), h) denotes the sequent arrow (f, A1 . . . An , Bm . . . B1 ).

11.2.

Cut elimination in D

In this section we will prove a cut-elimination theorem for the Gentzen

terms of D. Before stating and proving this result, we introduce some
technical notions and prove some auxiliary results.
A cut is a Gentzen term of the form cut , (f, g). A cut-free Gentzen
term is a Gentzen term none of whose subterms is a cut. A cut cut , (f, g)
is called topmost when f and g are cut-free.
We say that a Gentzen term is k-atomized when for every subterm of
L
R
it of the form kA
or kA
we have that A is an atomic formula, which here
means that it is a letter. Then we have the following lemma.
Atomic-k Lemma. For every Gentzen term g there is a k-atomized Gentzen term g such that g = g in D. Moreover, if g is cut-free, then g is
cut-free.
Proof. By Semilattice Coherence of 9.1, in D we have
L
L L
kAB
f = L kB
kA f.

We show next that for f : we have in D

11.2.

Cut elimination in D

247

L
R
L
L
L
kAB
f = w,,
w,,AB
L (kA
f, kB
f ).

The right-hand side (RHS) of this equation is equal to

w ((f k 1 ,A ) (f k 1 ,B )) e,,A,B ( 1eAB ) (w (A B)),

by the equations (w y) of the preceding section. Next we have

RHS = f w

k 1 ,AB ck,A,,B e,,A,B (w (A B)) ( 1eAB ),

by Lattice Coherence of 9.4,

= f k 1 ,AB (w (A B)) (m, (A B)) (w (A B))

( 1eAB ), by (k 1 nat) and (m e),

= f k 1 ,AB , by (mw) and (k 1 nat).

R
R
For kAB
and kAB
we proceed analogously.

We call leaf formulae of a Gentzen term h the following occurrences of

formulae in the type of h:
when h is 1p , the two occurrences of p in the type p p of h,
L
when h is kA
f , or L f , or L (f, g), the rightmost member of the
source of h,
R
when h is kA
f , or R (f, g), or R f , the leftmost member of the target
of h.
L
For example, the rightmost occurrence of A in the source A of kA
f:

L
A is a leaf formula of kA f .
The occurrence of A in the type of wL1 ,...,n , f (i.e. in the source of
L
w1 ,...,n , f ), recognized according to the index , is called the lower contraction formula of wL1 ,...,n , f . For every lower contraction formula A of
wL f there are two or more occurrences of A in the type of f (i.e. in the
source of f ), recognized according to the indices 1 , . . . , n , , which we call
the upper contraction formulae of wL1 ,...,n , f . We determine analogously

248

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

R
the lower and upper contraction formulae of w,
f (the dierence is
n ,...,1
R
that they are now in the targets of w,n ,...,1 f and f ).

For every unary Gentzen operation 1 and every Gentzen term 1 f ,

according to the indices of 1 we can recognize in the type of 1 f what
basic sequences are the basic sequences , 1 , . . . , n , , 1 , . . . , n mentioned in the inductive clause for 1 . We call these basic sequences the
lower parametric basic sequences of 1 f . Our inductive clauses for unary
Gentzen operations are such that for every lower parametric basic sequence
of 1 f there is a unique basic sequence in the type of f , recognized
according to the indices of 1 and the inductive clause for 1 , which we call
an upper parametric basic sequence of 1 f . We determine analogously the
lower and upper parametric basic sequences of 2 (f, g) for a binary Gentzen
operation 2 . Note that our inductive clauses for binary Gentzen operations are such that every lower parametric basic sequence of 2 (f, g) leads
unambiguously to a unique upper parametric basic sequence of 2 (f, g)
in the type of f or in the type of g. (In terms of linear logic, these clauses
correspond to rules for connectives of the multiplicative kind.)
For any Gentzen term h : , and x a member of or , we have
that x is either a leaf formula of h, or a lower contraction formula of h,
or a member of a lower parametric basic sequence of h. We dene the
notion of cluster of x in the following manner (this notion, called Bund in
German, stems from Gentzen; see [61], Section 3.41, [103], Section 2.621,
[36], Section 5, and [44], Section 2).
The cluster of x in h is a nite tree whose nodes are occurrences of the
same formula in the types of subterms of h. We assign to every node a label,
which is a subterm h of h such that the node occurs in the type of h . The
root of the cluster of x in h is x, and the label of the root is h. If a node y is
a leaf formula of its label, then y is a leaf; i.e., it has no successors. If a node
R
y is the lower contraction formula of its label wL1 ,...,n , f or w,
f,
n ,...,1
L
then y has as successors the upper contraction formulae of w1 ,...,n , f or
R
w,
f . These successors, of which there are at least two, all have f
n ,...,1
as labels. If a node y is a member of a lower parametric basic sequence
of its label h , then y has a single successor, which is the occurrence of the
same formula as y, at the same place, as a member of the upper parametric
basic sequence of h . If h is here 1 f , then the label of the successor is

11.2.

Cut elimination in D

249

f , and if h is 2 (f, g), then the label of the successor is f or g, depending

on whether the upper parametric basic sequence occurs in the type of f
or in the type of g. With that, we have dened the cluster of x in h.
For a cut cut 2 ,2 (f, g), the two occurrences of the cut formula A in the
target of f and in the source of g, recognized according to the indices 2 and
2 , are called respectively the left cut formula and the right cut formula of
cut 2 ,2 (f, g). (Note that the left cut formula is on the right-hand side of
, while the right cut formula is on the left-hand side of .)
For any Gentzen term h : , and x a member of or , let h (x)
be the number of nodes in the cluster of x in h. The left rank of a cut
cut 2 ,2 (f, g) is f (x) where x is the left cut formula of cut 2 ,2 (f, g), and
the right rank of cut 2 ,2 (f, g) is g (y) where y is the right cut formula
of cut 2 ,2 (f, g). The rank of a cut is the sum of its left and right ranks.
The least rank of a cut is 2, and in that case the left rank and the right
rank are both 1. This denition of rank is analogous to Gentzens, except
that Gentzen counts the number of nodes in the longest path, while we
count the total number of nodeseither measure is good. (A very formal
denition of rank may be found in [14], Section 3.) As a matter of fact, we
are interested only in ranks of topmost cuts, but our denition applies to
any cut.
We announced in the preceding section (after the equations (w y)) that
we will sometimes omit the subscripts of Gentzen operations. In the denition below, and sometimes later on, we take for granted the subscripts of
wL1 ,...,n , , and write just wL . We do the same with wR , and other Gentzen
operations, when their subscripts are cumbersome, but not important.
We say that a Gentzen term of the form wL f is a wL term. Subterms
that are wL terms are called wL subterms. We have an analogous terminology with wR . The rank of a wL term wL f is wL f (x) for x being the lower
contraction formula of wL f , and analogously for wR terms. We are interested below only in ranks of cut-free wL terms, but our denition applies
to any wL term.
Let x be the left cut formula and y the right cut formula of the cut
cut , (f, g). Then we say that a wL subterm h of g is tied to cut , (f, g)
when h is the label of a node of n-ary branching for n 2 in the cluster of y
in g, and we say analogously that a wR subterm h of f is tied to cut , (f, g)

250

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

when h is the label of a node of n-ary branching for n 2 in the cluster of

x in f .
We say that a wL term is blocked when it is of one of the following forms:
(w 1)

wL , , mix (f, g) : 1 2 1 C 2
1

for f :
(w 2)

1 C 2

and g : 1 C 2 ,

wL , , R (f, g) : 1 2 1 C 2 (A B)
1

for f : 1 C 2 A and g : 1 C 2 B ,
(w 3)

wL , , (AB) L (f, g) : 1 2 1 C 2 (A B)
1

for f : 1 C 2 A and g : 1 C 2 B ,
(w 1) wL1 ,2 , L f : 1 2 (A B)
for f : 1 (A B) 2 A B ,
(w 1) wL , , L (f, g) : 1 2 (A B)
1

for f : 1 (A B) 2 A and g : B ,
(w 2) wL , , L (f, g) : 1 2 (A B)
1

for f : A and g : 1 (A B) 2 B ,
(w 3) wL , , , L (f, g) : 1 2 1 2 (A B)
1

for f : 1 (A B) 2 A and g : 1 (A B) 2 B .
A wL subterm wL h of f2 tied to a topmost cut cut , (f1 , f2 ) is reducible
when it is not blocked and every wL subterm wL t of f2 tied to cut , (f1 , f2 )
such that wL h is a subterm of t is blocked. We can prove the following
lemma.
Reducibility Lemma. For every reducible wL h there is a cut-free Gentzen
term h such that wL h = h in D and after replacing wL h in cut , (f1 , f2 )
by h all the reducible wL subterms of h are of rank lesser than the rank of
wL h.
Proof. We proceed by cases depending on the form of h.

11.2.

Cut elimination in D

251

(W 1) First we have cases where wL h is wL ,..., , h while h is f :

n
1
1 A . . . n A for f : 1 A . . . n A and is either
one of cL , cR , k R , wR and R , or k L , wL and L with the occurrences of
A displayed in 1 A . . . n A members of the lower parametric
basic sequences of f . Then, by Semilattice Coherence of 9.1, we have
either wL f = wL f or wL cL f = wL f .
(W 2) For f : 1 A . . . n A , by Semilattice Coherence, we have
that
L
L
wL1 ,...,n ,, kA
f = cL
1 ... n ,,A, w1 ,...,n , f.

(W 3) Suppose we have f : 1 A . . . n A and

wL1 ,...,n , f : 1 A . . . m A ,
where 1 A . . . m A and 1 . . . n A are designations of the
same basic sequence, and one of the occurrences of A displayed in the rst
designation is the occurrence of A displayed in the second. Then we have
that
wL ,...,m , wL1 ,...,n , f = wL f
1

with appropriate subscripts for wL in wL f . If the rank of wL1 ,...,n , f is

k+1, and the rank of wL ,..., , wL1 ,...,n , f is k+1+l+m, then the rank
m
1
of wL f on the right-hand side is k+l+1. (Here we have m 2.)
(W 4) For f : 1 (A B) . . . n (A B) A B we have that
wL1 ,...,n ,, L f = wL1 ... n ,, L wL1 ,...,n , AB f
by Semilattice Coherence. Here, the right-hand side is blocked according
to (w 1). The wL term wL1 ,...,n , A B f need not be blocked and may
be reducible, but it is of lower rank than the left-hand side (the dierence
is n+1).
(W 5) For f : and g : 1 A . . . n A we have that
wL1 ,...,n , mix (f, g) = mix (f, wL1 ,...,n , g)

252

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

by Semilattice Coherence and (m nat). We proceed analogously when

mix (g, f ) replaces mix (f, g).
For f : 1 A . . . m A and g as above we have that
wL1 ,...,n , 1 ,2 ,...,m , mix (f, g) =
L
f, wL1 ,...,n , g)
wL1 ... n , 1 ... m , mix (w
1 ,...,m ,

by Semilattice Coherence and (m nat). Here, the right-hand side is blocked

L
according to (w 1). The wL terms w
f and wL1 ,...,n , g need not
1 ,...,m ,
be blocked and may be reducible, but they are both of lower rank than the
left-hand side.
We proceed as in case (W 5) when wL h is wL R (f, g) (one of these cases
involves a blocked wL term according to (w 2)). We have cases analogous
to (W 5) also when wL h is wL L (f, g) (here we apply Semilattice Coherence
and (dL nat), and one of these cases involves a blocked wL term according
to (w 3)). We have three additional cases when wL h is wL L (f, g), which
all yield blocked wL terms according to (w 1), (w 2) and (w 3). One
of these cases is the following.
(W 6) For the Gentzen terms f : 1 (A B) . . . n (A B) A
and g : 1 (A B) . . . m (A B) B we have that
wL ,..., , , ,..., , , L (f, g) =
1

wL ... , ... , , L (wL ,..., , A f, wL ,..., , B g)

1

by Semilattice Coherence and (dL nat). Here, the right-hand side is blocked
according to (w 3). The wL terms wL ,..., , A f and wL ,..., , B g
n
m
1
1
need not be blocked and may be reducible, but they are both of lower rank
than the left-hand side.
To conclude the proof of the lemma we have only to check that the
condition on ranks is satised in all cases, even in those where we have not
noted the fact.

We have an analogous denition of blocked wR terms and of reducible wR

subterms. With that, we prove for wR terms a lemma exactly analogous to
the Reducibility Lemma. As a corollary of these two Reducibility Lemmata,
we have the following lemma.

11.2.

Cut elimination in D

253

Blocked-w Lemma. Every topmost cut , (f, g) is equal in D to a topmost cut , (f , g ) in which all wL and wR subterms tied to cut , (f , g )
are blocked.
The proof of this corollary is based on a multiset-ordering induction, which
stems from Gentzen (see [61] and [35]).
The degree of a cut is the number of occurrences of connectives (in this
case, the number of occurrences of and ) in the cut formula.
The complexity of a topmost cut is a pair (d, r) where d is the degree of
this cut and r is its rank. These complexities are ordered lexicographically
(i.e., we have (d1 , r1 ) < (d2 , r2 ) i either d1 < d2 , or d1 = d2 and r1 < r2 ;
cf. 7.7).
According to the Atomic-k Lemma and the Blocked-w Lemma, every
topmost cut cut , (f, g) is equal to a topmost cut cut , (f , g ) with the
same cut formula such that f and g are k-atomized and every wL or wR
subterm of f and g tied to cut , (f , g ) is blocked. We call topmost cuts
such as cut , (f , g ) clean cuts.
We can then prove the following theorem.
Cut-Elimination Theorem. For every Gentzen term t there is a cut-free
Gentzen term t such that t = t in D.
Proof. We show by induction on the complexity of clean cuts that they
are equal in D to cut-free Gentzen terms. This will suce to prove the
theorem.
For the basis of this induction, take a clean cut of complexity (0, 2). This
means that this clean cut is of one of the forms displayed on the left-hand
side of the following equations of D:
cut , (1p , 1p ) = 1p ,
cut , (1p , kpL g) = kpL g,
cut , (kpR f, 1p ) = kpR f,
cut , (kpR f, kpL g) = mix (f, g).
For the rst three equations we use (cat 1), while the fourth holds by denition. With that, we have proved the basis of the induction.

254

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

Note that with the rst three equations we proceed as Gentzen, but not
with the fourth. Instead of reducing the left-hand side of this equation to
the right-hand side, Gentzen would reduce it to a cut-free term obtained
either from f with a number of k L , cL and k R operations, or from g with a
number of k R , cR and k L operations (cf. [60], Section III.3.113.1-2). Such
reductions are, however, not supported by equations of D.
We pass now to the induction step. Suppose rst that the complexity
of our clean cut is (d, 2) for d > 0. When the cut formula is of the form
A B, our clean cut must be of the form
cut , (R (f, g), L h),
for f : 1 A 1 , g : 2 B 2 and h : A B . Then we have in D
the equation
L
cut , (R (f, g), L h) = cR
,2 ,1 , c,2 ,1 , cut 2 , (f, cut , (g, h)).

To show that this equation holds in D, we have that, with subscripts

omitted, the left-hand side is equal to

(h 1) ((1 1e ) 1) d (1 (1i 1)) (1 e ) (1 f g) =

(h 1) d (1 e ) (1 f g),
by (dL nat) and the fact that 1i and 1e are isomorphisms, while, again with
subscripts omitted, for the right-hand side we have

(1 c) (h 1) (d 1) ((1 1 g) 1) ((1 c) 1) d (1 f ) (1 c) =

(h 1) (1 c) (d 1) ((1 c) 1) d (1 c) (1 f g),

by the bifunctorial equation (2) of 2.7, ( c nat) and (dL nat). It suces
to note now that
d (1 e ) = (1 c) (d 1) ((1 c) 1) d (1 c)

holds by Symmetric Net Coherence of 7.6. When is , we have essentially

a case of the equation ( e) of 7.6.

We replaced a clean cut cut , (R (f, g), L h) of complexity (d, 2) by

11.2.

Cut elimination in D

255

L
cR
,2 ,1 , c,2 ,1 , cut 2 , (f, cut , (g, h)).

According to the Atomic-k Lemma and the Blocked-w Lemma, the topmost
cut cut , (g, h) is equal to a clean cut cut , (g , h ) of complexity (d , r) with
d < d, because the cut formula is now B instead of AB. By the induction
hypothesis, cut , (g , h ) = s for a cut-free Gentzen term s. The topmost
cut cut 2 , (f, s) is equal to a clean cut cut 2 , (f , s ) of complexity (d , r)
with d < d, because the cut formula is now A instead of A B. So we can
apply again the induction hypothesis.
We proceed analogously when the cut formula is of the form A B.
With that, we are over with the cases where the complexity of our clean
cut is (d, 2) for d > 0. We dealt with them in the spirit of Gentzen.
Suppose now that the complexity of our clean cut is (d, r) for r > 2,
and suppose the right rank of this clean cut is greater than 1. We proceed
analogously if the left rank is greater than 1, and we need not consider this
case separately.
Suppose rst that in our clean cut cut 2 ,2 (f, g) the cut formula occurs
in a lower parametric basic sequence of g. Depending on various cases for
the unary Gentzen operation , we want to show that one of the following
two equations holds in D:
()

cut 2 ,2 (f, g) = cut 2 ,2 (f, g),

()

cut 2 ,2 (f, g) = cL cut 2 ,2 (f, g)

for a clean cut cut 2 ,2 (f, g) of complexity (d, r ) with r = r 1. The

subscripts omitted in need not be the same on the two sides of () or
(), and often they are not such. We have the following cases.
(1)

If is cL , then () holds by Symmetric Net Coherence.

(2) If is cR , or k R , or wR , or R , then () holds by the bifunctorial

equation (2).
(3) If is k L , then () holds. To show that, we distinguish two cases. In
L
both cases, we have g : 1 A 2 3 and kB
g : 1 A 2 B 3 where
2 B is 2 . In the rst case, the basic sequence 1 2 is not , and in the
second case it is .

256

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

In the rst case, for the left-hand side of () we have

(g 1) (k 1 1 A ,B 1) ((1 c 2 B,A ) 1) d (1 ( c 1)) (1 f ),

while for the right-hand side of () we have

(g 1) ((1 c) 1) d (1 ( c 1)) (1 f ) k 1 1 3 ,B (1 c B,3 )

2

= (g 1) ((1 c) 1) d (1 ( c 1)) k 1 1 (2 A1 ),B

2

(1 c
B,2 A1 ) (1 f )

= (g 1) ((1 c) 1) d (1 ( c 1)) (k 1 1 ,B 12 A1 ) (1 f ),
2
by Semilattice Coherence (this step cannot be made

if 1 2 is , since k 1 1 ,B would not be dened),

2

= (g 1) ((1 c 2 ,A ) 1) ((k 1 1 ,B 1) 1) d (1 ( c 1)) (1 f ),

2

which is equal to the left-hand side by Semilattice Coherence.

In the second case, when 1 2 is , for the left-hand side of () we
have

(g 1) (k 1A,B 1) ( c B,A 1) d (1 ( c 1)) (1 f ),

while for the right-hand side we have

(g 1) ( c 1) f k 1 3 ,B c B,3

= (g 1) k 2 B,A 2 1 (1 ( c 1)) (1 f )

= (g 1) (k 2 B,A 12 1 ) dB,A,2 1 (1 ( c 1)) (1 f ), by (d k),

which is equal to the left-hand side by ( c k) (see 9.1).

(4) If is wL or L , then () holds by various bifunctorial and naturality
equations.
Suppose next that in our clean cut cut 2 ,2 (f, (g, h)), the cut formula
A occurs in a lower parametric basic sequence of (g, h) (here, is a binary
Gentzen operation).
(5) If is mix , then for f : 3 2 A 1 , g : 2 3 and h : 1 A 2
3 we have in D the equation

11.2.

Cut elimination in D

257

cut 2 2 ,2 (f, mix (g, h)) =

L

cR
,2 1 , , c1 ,3 , , mix (g, cut 2 ,2 (f, h)),
3

where the complexity of the clean cut cut

(f, h) is (d, r ) with r = r1.
By bifunctorial and naturality equations, the left-hand side of this equation is equal to (h 1) LHS (1 g f ) where
2 ,2

LHS = (m1 A 2 ,3 1) ((1 c) 1) d (1 ( c 1)),

while the right-hand side is equal to (h 1) RHS (1 g f ) where

RHS = (1 c) ((1 c) 1) (d 1) ((1 ( c 1)) 1)

m1 2 (2 A 1 ),3 (1 c).

We have LHS = RHS by Mix-Symmetric Net Coherence of 8.4. We

proceed analogously if g and h have types interchanged.
(6)

If is R , then we have in D one of the following two equations:

cut (f, R (g, h)) = cR cL R (cut (f, g), h),
cut (f, R (g, h)) = R (g, cut (f, h)),

where the complexity of the clean cuts cut (f, g) and cut (f, h) is (d, r )
with r = r 1, and appropriate subscripts are assigned to cut. Both of
these equations are justied by Symmetric Net Coherence. We proceed
analogously if is L .
With that, we are over with the cases of a clean cut cut 2 ,2 (f, g) or
cut 2 ,2 (f, (g, h)) of complexity (d, r) with r > 2 where the cut formula
occurs in a lower parametric basic sequence of g or (g, h). All these cases
are dealt with in the spirit of Gentzen, except for the case with mix , which
Gentzen did not envisage.
Now we proceed with the cases of complexity (d, r) with r > 2 where
the cut formula does not occur in this manner in a lower parametric basic
sequence. Then our clean cut must be of the form cut (f, wL g) with blocked
wL g tied to our clean cut, and we have to go through the cases (w 1),
(w 2), . . . , (w 3) for blocked wL terms.
(w 1) For f : 1 C 2 , g : 1 C 2 and h : 1 C 2 ,
where is 1 2 , while is 1 2 , and is 1 2 , we have in D the
equation

258

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

cut 2 ,1 (h, wL , , mix (g, f )) =

1

R
L

cR
,, , w , , w , , mix (cut 2 ,1 (h, g), cut 2 ,1 (h, f )),

where the complexity of the clean cut cut 2 ,1 (h, f ) is (d, r ) with r < r,
and analogously for the clean cut cut 2 ,1 (h, g).

If f : C is f (1 c 2 ,C ), while g : C is g (1 c 2 ,C ),

and h is ( c C,1 1) h, then, by Semilattice Coherence and bifunctorial and

naturality equations, the left-hand side of our equation is equal to
(m , 1) ((f g ) 1) LHS (1 h )

where LHS is ((1 c ,C 1C ) 1 ) ((1 wC ) 1) d, while the righthand side is equal to

(1 w ) (1 c , 1 ) m , ((f 1) (g 1)) (d d)

(1 c 1) (1 w ) (1 h ).
C

By the equation (m c m ) of the preceding section, we have

(1 c , 1 ) m , = (m , 1 ) e , ,, .

So the right-hand side is equal to

(m , 1) ((f g ) 1) RHS (1 h )
where

RHS = (1 w ) e C, C,, (d d) (1 c 1) (1 wC )

= (1 w ) e (d d) (1 c 1) (1 ck ) (1 (1 w ))

(1 (w
C 1)), by Lattice Coherence, provided is not ,

= (1 w ) ((1 c 1) 1) d (1 e ) (1 ck ) (1 (1 w ))

(1 (w
C 1)), by Symmetric Net Coherence,

= (1 w ) ((1 c 1) 1) d (1 (1 m, )) (1 (1 w ))

(1 (w
C 1)), by (m e),

= ((1 c 1) 1) d (1 (wC 1)), by bifunctorial and naturality

equations, and (wm),
= LHS , by (dL nat).

11.2.

Cut elimination in D

259

If is , then LHS = RHS by Semilattice Coherence.

(w 2) For f : 1 C 2 A , g : 1 C 2 B and h :
1 C 2 , with , and as in (w 1), we have in D the equation
cut 2 ,1 (h, wL , , R (f, g)) =
1

R
L
R

cR
AB ,, , wAB , , w , , (cut 2 ,1 (h, f ), cut 2 ,1 (h, g)),

where the complexity of the clean cut cut 2 ,1 (h, f ) is (d, r ) with r < r,
and analogously for the clean cut cut 2 ,1 (h, g).
For f : C A , g : C B and h : C dened as in
(w 1), we have by Semilattice Coherence and bifunctorial and naturality
equations that the left-hand side of our equation is equal to
(1iA B 1) (eA,B, , 1) ((f g ) 1) LHS (1 h )

where LHS is ((1 c 1) 1) ((1 wC ) 1) d, while the right-hand side is

equal to

(1 w ) (1 c 1) (1i 1) eA,B, , ((f 1) (g 1)) (d d)

(1 h 1 h ) (1 c 1) (1 w ).

We have by Symmetric Net Coherence that

eA,B, , = (1(A B) c , 1 ) (eA,B, , 1 )

eA ,B ,, .

Then, by bifunctorial and naturality equations, and ( c c) (see the List of

Equations), the right-hand side is equal to
(1iA B 1) (eA,B, , 1) ((f g ) 1) RHS (1 h )

where

RHS = (1 w ) e C, C,, (d d) (1 c 1) (1 wC )
= LHS , as in (w 1).
(w 3) For f : 1 C 2 A , g : 1 C 2 B and h :
1 C 2 , with , and as in (w 1), we have in D the equation

260

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

cut 2 AB,1 (h, wL , , AB L (f, g)) =

1

R
R
L
cL
,,AB, c ,, , w , , w , ,AB
L

L (cL
,A,, cut 2 A,1 (h, f ), c ,B,, cut 2 B,1 (h, g)),

where the complexity of the clean cut cut 2 A,1 (h, f ) is (d, r ) with r < r ,
and analogously for the clean cut cut 2 B,1 (h, g).
For h : C dened as in (w 1), we have by Semilattice Coherence
and bifunctorial and naturality equations that the left-hand side of our
equation is equal to
(f g 1) LHS (1 h ) (1 1eAB 1)

where LHS is (e 1) ((1 c 1) 1) ((1 c) 1) d (1 (wC 1)), while

the right-hand side is equal to
(f g 1) RHS (1 h ) (1 1eAB 1)
where

RHS = (1 w ) (1 c 1) ((1 c) 1 (1 c) 1) (d d)

((1 c) (1 c)) e (1 c) (1 c 1) (1 wC )

= (1 w ) (e 1) ((1 c 1) 1) ((1 c) 1) d (1 e )

(1 w ), by Symmetric Net Coherence,

C
= (1 w ) (e 1) ((1 c 1) 1) ((1 c) 1) d (1 e )

(1 ck ) (1 (1 w )) (1 (w

C 1)),
by Lattice Coherence, provided is not ,
= LHS , by (m e), (wm) and bifunctorial and naturality equations
(cf. the case (w 1)).

If is , then we obtain that LHS = RHS by Semilattice Coherence, in

a simplied version of the derivation above.
(w 1) For f : 1 A B 2 A B and h : 1 A B 2 , with
and as in (w 1), we have in D the equation
cut ,1 (h, wL , , L f ) =
1

R
L
L L
w
,, w ,, cut ,1 (h, c ,AB,, cut AB,1 (h, f )),
2

11.2.

Cut elimination in D

261

where the complexity of the clean cut cut 2 AB,1 (h, f ) is (d, r ) with
r < r.
Then, by the induction hypothesis, cut 2 AB,1 (h, f ) = f , for a cutfree Gentzen term f , which by the Atomic-k Lemma we may assume to be
k-atomized. (As a matter of fact, our procedure of cut elimination is such
that it produces out of a clean cut a k-atomized cut-free Gentzen term.)

The topmost cut cut ,1 (h, L cL

,AB,, f ), whose right rank is 1, is clean,

and its complexity is (d, r ) with r < r.

To justify the equation displayed above, we proceed as follows. For
h : A B dened as in (w 1), we have by Semilattice Coherence
and bifunctorial and naturality equations that the left-hand side of our
equation is equal to
(f 1) ((1 1eAB ) 1) LHS (1 h ),

where LHS is ((1 c 1) 1) d (1 (wAB 1)), while the right-hand side

is equal to
(f 1) ((1 1eAB ) 1) RHS (1 h )
where

RHS = (1 w ) ((1 c) 1) (d 1) ((1 c) 1) d (1 wAB )

= (1 w ) ((1 c 1) 1) d (1 e ) (1 wAB ),
by Symmetric Net Coherence,

= (1 w ) ((1 c 1) 1) d (1 e ) (1 ck ) (1 (1 w ))

(1 (w
AB 1)), by Lattice Coherence, provided is not ,
= LHS , by (m e), (wm) and bifunctorial and naturality equations
(cf. the case (w 1)).

If is , then we obtain that LHS = RHS by Semilattice Coherence, in

a simplied version of the derivation above.
(w 1) For f : 1 A B 2 A , g : B and h :
1 A B 2 , with and as in (w 1), we have in D the equation
cut ,1 (h, wL , , L (f, g)) =
1

R
L
L L

cR
,, , w , , w , , cut ,1 (h, (c ,A,, cut 2 A,1 (h, f ), g)),

262

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

where the complexity of the clean cut cut 2 A,1 (h, f ) is (d, r ) with r < r.
Then, by the induction hypothesis, cut 2 A,1 (h, f ) = f for a cut-free
Gentzen term f , which by the Atomic-k Lemma we may assume to be

k-atomized. The topmost cut cut ,1 (h, L (cL

,A,, f , g)), whose rank is

1, is clean, and its complexity is (d, r ) with r < r.

To justify the equation displayed above, we proceed as follows. For
h : A B dened as in (w 1), we have by Semilattice Coherence
and bifunctorial and naturality equations that the left-hand side of our
equation is equal to (f g 1) LHS (1 h ) where LHS is
(e 1)((1 c 1) 1) d (1 ((1 1eA B ) 1)) (1 (wAB 1)),

while the right-hand side is equal to (f g 1) RHS (1 h ) where

RHS = (1 w ) (1 c 1) ((1 c) 1) (d 1) ((1 c) 1) (e 1) d

e
(1 c 1) (1 (1 1)) (1 w
AB )
A B

= (1 w ) (e 1) ((1 c 1) 1) d (1 e ) (1 (1eAB 1))

(1 w
AB ), by Symmetric Net Coherence,
= (1 w ) (e 1) ((1 c 1) 1) d (1 e ) (1 (1eAB 1))

(1 ck ) (1 (1 w )) (1 (w

AB 1)),
by Lattice Coherence, provided is not ,
= LHS , by (m e), (wm) and bifunctorial and naturality equations
(cf. the case (w 1)).

If is , then we obtain that LHS = RHS by Semilattice Coherence, in

a simplied version of the derivation above.
We proceed analogously in the case (w 2).
(w 3) For f : 1 A B 2 A , g : 1 A B 2 B and
h : 1 A B 2 , with , and as in (w 1), we have in D the
equation
R
L
cut ,1 (h, wL , , , L (f, g)) = cR
,, , w , ,, w , ,,
1

cut ,1 (h, L (cL

,A,, cut 2 A,1 (h, f ), c ,B,, cut 2 B,1 (h, g)),

where the complexity of the clean cut cut 2 A,1 (h, f ) is (d, r ) with r < r,
and analogously for the clean cut cut 2 B,1 (h, g).

11.3.

Coherence of distributive lattice categories

263

Then, by the induction hypothesis, these two cuts are equal to the cutfree Gentzen terms f and g , respectively, which by the Atomic-k Lemma
we may assume to be k-atomized. The topmost cut
L

cut ,1 (h, L (cL

,A,, f , c ,B,, g )),

whose right rank is 1, is clean, and its complexity is (d, r ) with r < r.
To justify the equation displayed above we proceed as follows. For

h : A B dened as in (w 1 ), we have by Semilattice Coherence

and bifunctorial and naturality equations that the left-hand side of our
equation is equal to (f g 1) LHS (1 h ) where LHS is

(e 1) ((1 c 1) 1) ((1 c 1) 1) d (1 ((1 1eAB ) 1))

(1 ((1 w
AB ) 1)) (1 (w AB 1)),
while the right-hand side is equal to (f g 1) RHS (1 h ) where

RHS = (1 w ) (1 w ) (1 c 1) ((1 c) 1 (1 c) 1) (d d 1)

((1 c) (1 c) 1) (e 1) d (1 c 1) (1 (1eAB 1))

(1 w
AB ) (1 w AB )

= (1 w ) (1 w ) (e 1) ((1 c 1) 1) ((1 c 1) 1) d

(1 e ) (1 e ) (1 (1eAB 1)) (1 wAB ) (1 wAB ),

by Symmetric Net Coherence,

= LHS , by Lattice Coherence, (m e), (wm) and bifunctorial and

naturality equations, provided is not (cf. the case (w 1)).

If is , then we obtain that LHS = RHS by Semilattice Coherence, in

a simplied version of the derivation above.
Note that the cases (w 2), (w 3), (w 1), (w 1), (w 2) and (w 3)
are dealt with in the spirit of Gentzen. The case (w 1), which involves
mix, was not envisaged by him. This concludes the proof.

11.3.

Coherence of distributive lattice categories

The essential ingredient in our proof of coherence for the category DL is

the Cut-Elimination Theorem of the preceding section. Another ingredient
is Restricted Mix-Lattice Coherence of 10.2. Before proving coherence for

264

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

DL, we consider some matters that serve to connect DL with the category
ML, but are also of an independent interest.
With the abbreviations
s A,C,D =df e A,A,C,D (wA 1CD ) : A (C D) (A C) (A D),

s D,C,A =df (1DC wA ) e D,C,A,A : (D A) (C A) (D C) A,

t A,C,D =df (wA 1CD ) ckA,C,A,D : (A C) (A D) A (C D),

t D,C,A =df ckD,C,A,A (1DC wA ) : (D C) A (D A) (C A),

we obtain the following equations in DL as an immediate consequence of

(m e), (m e) and (wm):

t A,C,D

s D,C,A

s A,C,D = 1A(CD) ,

t D,C,A = 1(DC)A .

This means that s A,C,D is a right inverse (i.e. section) of t A,C,D , while

s D,C,A is a left inverse (i.e. retraction) of t D,C,A (see [100], Section I.5). It

is easy to see that t A,C,D and s A,C,D are not inverse to each other in DL,

since G( s A,C,D

t A,C,D ) is dierent from G(1(AC)(AD) ); analogously,

t D,C,A and s D,C,A are not inverse to each other. The types of the arrow

terms in the families s and s give what is usually called distribution of

over and distribution of over . However, these arrow terms do not
stand for isomorphisms in DL.
For every formula A of L, , let Adnf be any formula of L, in disjunctive normal form (dnf ; see 10.2) such that there is an arrow term

t A : Adnf A of C(L) and an arrow term s A : A Adnf of C(DL) for which

in DL we have t A s A = 1A . (We do not require the uniqueness of t A and

s A , as we did not require the uniqueness of Adnf .) That for every formula
A of L, there is a formula Adnf is shown by an easy induction on the
number of occurrences of in the scope of an occurrence of .
Dually, for every formula A of L, , let Acnf be any formula in conjunctive normal form (cnf ; see 10.2) such that there is an arrow term

t A : A Acnf of C(L) and an arrow term s A : Acnf A of C(DL) for which

11.3.

Coherence of distributive lattice categories

265

in DL we have s A t A = 1A . That for every formula A of L, there is a

formula Acnf is shown by an easy induction, as above.

For

{, }, the arrow terms t are built out of arrow terms of the

A

form t B,C,D and arrow terms of C(S) with the help of the operations ,

and on arrow terms, while the arrow terms sA are built out of arrow terms

of the form sB,C,D and arrow terms of C(S) with the help of the operations
, , and on arrow terms. For example, if A is p ((q (r s)) q), and
Adnf is ((p q) ((p r) s)) (p q) (this is the source of the arrow term

we had as an example in 10.2), then we can take that t A is

t p,q(rs),q ( t p,q,rs 1pq ) ((1pq b

p,r,s ) 1pq ),

while s A is

((1pq b
p,r,s ) 1pq ) ( s p,q,rs 1pq ) s p,q(rs),q .

It is easy to verify, by referring to denitions, that a cut-free Gentzen

term of D of the type A B for A in L and B in L denotes an arrow
(f, A, B) of D such that f is an arrow term of C(ML). As a consequence
of that and of the Cut-Elimination Theorem of the preceding section, we
obtain the following.
Proposition. For A in L and B in L , every arrow term of C(DL) of
type A B is equal in DL to an arrow term of C(ML).
Then we can prove the following lemma, which appeals to the notion of
settled normal form of 10.2.
Normal-Form Lemma. Every arrow term f : Adnf B cnf of C(DL) is
equal in DL to an arrow term of C(GML) in settled normal form.
Proof. As in the proof of the Normal-Form Lemma of 10.2, we make
an induction on the number of occurrences of in A and in B. If there
are no such occurrences of and , then we apply the Proposition above
and the Normal-Form Lemma of 10.2. For the remainder of the proof we
proceed as in the proof of that lemma in 10.2.

Then we can prove the following.

266

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

Distributive Lattice Coherence. The functor G from DL to Rel is

faithful.
Proof. Suppose f, g : A B are arrow terms of C(DL). If Gf = Gg, then

G( t B f t A ) = G( t B g t A ). By the Normal-Form Lemma above, we

have in DL that t B f t A = f and t B g t A = g for f and g arrow

terms of C(ML). By Restricted Mix-Lattice Coherence of 10.2, we have
that f = g in ML, and hence also in DL. So in DL we have

sB

tB f tA

sA = sB

tB g tA

sA,

and hence f = g.

A logical system synonymous with C(DL) may be obtained by taking

e or e as primitive transformations instead of d, since in DL we have the
equation

dA,B,C = (1AB k 2A,C ) e A,A,B,C (wA 1BC ),

which can easily be checked by Distributive Lattice Coherence. Analo

gously, we could take e or e as primitive, since in DL we have

dC,B,A = (1CB wA ) e C,B,A,A (k 1C,A 1BA ).

Alternative primitive transformations are s and s , whose members occur

in the two equations above. With such alternative primitives, however, we
have not managed to nd an axiomatization simpler than what we have for
E(DL) and E(DL ).

A primitive of the same type as s was considered in [91] and [92] as an

addition to S, extended with the isomorphism of A with . In the
presence of this isomorphism, we cannot expect coherence with respect to
Rel with a functor such as our functors G. The coherence result of [91] is
a restricted coherence result in the sense of preorder, while the coherence
result of [92] is a result about a faithful functor into Rel, which diers from
our functor G with respect to . The equations of those papers without
and hold, however, in DL.

If to E(DL) we add the equation s A,C,D t A,C,D = 1(AC)(AD) or

the equation t D,C,A

s D,C,A = 1(DA)(CA) , then we can derive that all

11.3.

Coherence of distributive lattice categories

267

arrow terms of the same type are equal. Here is a proof of that fact for the
rst equation. (We proceed analogously with the second equation.)
We have that

= s A,B,B

t A,B,B = 1(AB)(AB) ,

= ( c B,A c B,A ) s B,A,A

t B,A,A ( c A,B c A,B ) = 1(AB)(AB) .

So = 1(AB)(AB) , by ( idemp). By Distributive Lattice Coherence, we infer that

[k 1AB,AB k 2AB,AB , k 1AB,AB k 2AB,AB ] = 1(AB)(AB) ;

therefore, with () (see the List of Equations), we have

k 1AB,AB = k 1AB,AB k 2AB,AB = k 2AB,AB .

For f, g : A B, we have

[f 1A , g 1A ] k 1AA,AA = [f 1A , g 1A ] k 2AA,AA
f 1A = g 1A , by (),

f k 1A,A = g k 1A,A , by (k 1 nat),

from which we infer f = g with (wk) (see 9.1). So s and t cannot be

inverses of each other in the context of DL without preorder, i.e. triviality.
(For a result of the same kind, see [22], Proposition 3.1.)
The category Set of sets with functions is a lattice category with being
cartesian product and being disjoint union + (cf. 9.6). Products and
coproducts are unique up to isomorphism (see [100], Sections IV.1-2), and
so there is no alternative lattice-category structure in Set. Since Set is,
of course, not a preorder, we can conclude, according to what we said

above, that it is not a distributive lattice category with s a,b,c : a (b + c)

(a b) + (a c) and t a,b,c : (a b) + (a c) a (b + c) being up to

associativity like identity arrows (for example, with x a and y b, we have

s a,b,c((x, (y, ))) = ((x, y), )). With da,b,c dened as (1ab k 2a,c ) s a,b,c ,

the equation (d b) of 7.2 does not hold in Set, as noted in [22] (Section 3;
the remaining specic equations of E(DA) hold in Set). At the same place,

268

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

an argument is presented that Set with and being and + cannot

satisfy (d b) for any denition of dL and dR . Here is another argument to

the same eect.
The category Set is not a distributive lattice category. If it were that,
then we would have in it for every set a a function ma,a : a a a + a
that satises the following instance of the equation (cm) of 8.4:

ma,a c a,a = c a,a ma,a .

If a = {x}, then, since a a = {(x, x)} is a terminal object, we obtain

ma,a = c a,a ma,a .

There are only two functions from a a to a + a = {(x, ), (, x)}, and

none of them satises the last equation, because c a,a (x, ) = (, x) and

c a,a (, x) = (x, ). This argument shows also that Set is not a mix-lattice
category.

11.4.

Legitimate relations

At the end of 9.2, we made a brief comment on the image under the

functor G of the categories L and L. Once we have proved Distributive

Lattice Coherence, it is of some interest to consider the image under G of
the category DL. We will devote the present section to this matter.
For A and B formulae of L, , we will say that a relation R GAGB
is legitimate when there is an arrow term f : A B of C(DL) such that
Gf = R. We will prove two propositions that will enable us to decide
whether a relation is legitimate.
For k and l nite ordinals, let k +l be the set {n + l | n k}. Other
notions mentioned in the statements and proofs of our two propositions are
dened in the preceding section and in 10.2.
Proposition 1. In GAdnf = k1 + . . . + kn , for n 1, and GB cnf =
l1 + . . . + lm , for m 1, let ki , for i {1, . . . , n}, be GAi for a minimal
disjunct Ai of Adnf , and let lj , for j {1, . . . , m}, be GBj for a minimal
conjunct Bj of B cnf . Then a relation R GAdnf GB cnf is legitimate
i for every i {1, . . . , n} and every j {1, . . . , m} the relation R
+l +...+lj1
+k +...+ki1
(ki 1
lj 1
) is not empty.

11.4.

Legitimate relations

269

Proof. Suppose R is legitimate; i.e., there is an arrow term f : Adnf B cnf

of C(DL) such that Gf = R. By the Normal-Form Lemma of the preceding section, f is equal in DL to an arrow term of C(GML) in normal
form. Then the molecular correspondence (see 10.2) is enough to prove
the proposition from left to right, because for every molecular component
f we have that Gf is not empty. For the other direction, we build out
+l +...+lj1
+k +...+ki1
of the relations R (ki 1
lj 1
) arrow terms of C(GML)
in bracket-free normal form, which we then combine, as molecular components, to build an f in normal form such that Gf = R.

We can check that the relation G, which we have drawn in 10.2,

satises the condition equivalent to legitimacy stated in Proposition 1. For
example, for the couple (A2 , B3 ) we have {(2, 3), (2, 4)} G (3+2
3+1+2 ). The molecular component of corresponding to (A2 , B3 ) is 3 4 .
For the second of our two propositions, remember that, according to
the denition of 2.9, we write composition of relations from right to left,
as composition of functions.
Proposition 2. The relation R GA GB is legitimate i the relation

G t B R G t A GAdnf GB cnf is legitimate.

Proof. The left-to-right direction of the proposition is trivial. For the

other direction, suppose G t B R G t A is legitimate. Then

G sB G t B R G t A G sA,
which is equal to R, is legitimate.

By combining Propositions 1 and 2 we can decide whether any relation

R GA GB is legitimate.
For the maximal relation Rmax GA GB we have (i, j) Rmax when
the i+1-th occurrence of letter in A (counting from the left) and the j+1-th
occurrence of letter in B are occurrences of the same letter. For : A B
being our arrow term of 10.2, we have that G is not Rmax GA GB. If
is obtained from by replacing 3 and 4 by 3 4 , then G coincides
with Rmax GA GB.
We can use maximal relations to solve the theoremhood problem for the

270

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

category DL. If any relation R GA GB is legitimate, then Rmax

GAGB is legitimate. So, to check whether there is an arrow of type A B
in DL, it is enough to check whether Rmax GA GB is legitimate. The
theoremhood problem for DL is, however, solved in a much more familiar
way by noting that there is an arrow of type A B in DL i the implication
A B is a tautology.

11.5.

Coherence of distributive dicartesian categories

To obtain the natural logical category DL, , we have that the logical
system C(DL, ) is in L,,, , with the transformations included in 1,
b, c, w-k, m, d and -. The specic equations of E(DL, ) are obtained
by taking the union of those of E(DL) and E(L, ) minus the equations

(m e) and (m e). We call natural DL, -categories distributive dicartesian

categories. The objects of a distributive dicartesian category that is a
partial order make a distributive lattice with top and bottom.

Note that the equations ( dL ), ( dL ), ( dR ) and ( dR ) of 7.9 hold

in DL, . It suces to derive the rst two of these equations:

( d L )

( dL )

d,B,C = (
B 1C ) BC ,

dA,B, =
AB (1A B );

the remaining two equations then follow easily. For ( dL ), since

BC =

k 2,BC , we have that the right-hand side is equal to

(
B 1C ) (k ,B 1C ) d,B,C

by (d k), and this is equal to d,B,C . Conversely, as we noted in 11.1, one

can derive (d k) from ( dL ) by precomposing with A 1BC . We proceed

analogously for ( dL ).

With the help of (d k), (d k) and Lemma 2 of 9.6 we obtain the following
equations of DL, :

(d)

dA,, = k 1A, (1A ),

(d)

d,,C = ( 1C ) k 2,C .

11.5.

Coherence of distributive dicartesian categories

271

With the help of ( dL ) or (d), together with Restricted Dicartesian

Coherence, we obtain that the equation (m) of ML, (see 10.3) holds

in DL, . For (m) (see 10.3), we rely on ( dL ) or (d), and Restricted Dicartesian Coherence. So in E(DL, ) we have all the equations
of E(ML, ).

The equations (m e) and (m e) of 11.1 are derivable for DL, . To

derive (m e), we rst establish this equation for C and D being . We can
achieve that by relying on

m, [
, ] = 1 ,

for which we use (d). So we have

(k 1A, k 1B, ) e A,B,, = mA,B k 1AB,.

Then we use

k 1A,C = k 1A, (1A C ),

k 1B,D = k 1B, (1B D ),

k 1AB,CD = k 1AB, (1AB (C D )).

We proceed analogously for (m e).

Let C and C be respectively the logical systems C(DL, ) and C(A, ),
while E is E(DL, ). Next, let B be C/E, that is DL, . Then it is easy
to see that the conditions (IVC) and (IVB) of 3.1 are satised. Since
the C -core of C/E is the category A, , by Bimonoidal Coherence of 6.1,
we have that the condition (IVG) of 3.1 is also satised. So (IV) holds,
and since A, is a preorder, by Bimonoidal Coherence, we can apply
A,
the Strictication Theorem of 3.1 to obtain that the category DL,
,
G
that is DL, , is equivalent to DL, via two strong C(DL, )-functors.
A,
According to 6.2, the objects of DL,
may be identied with form
sequences of L,,, in the extended sense.
A,
Let D be now the category DL,
. We use the terminology of 11.1
with the following changes.
A basic sequence of colour {, } of D is a form sequence of L,,,
in the extended sense that is either of the form (A1 . . . An , ), for n 2
and Ai , where i {1, . . . , n}, a formula of L,,, , or it is a formula of
L,,, , or it is (, ). The basic sequence (, ) has no members.

272

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

We dene Gentzen terms by induction as in 11.1 with the following

additions. Besides 1p : p p, which, as before, denotes the arrow (1p , p, p)
of D, we put among atomic Gentzen terms 1i : (, ) , which denotes
(1 , (, ), ), and 1e : (, ), which denotes (1 , , (, )). The
Gentzen operations cL , cR , k L , k R , wL , wR , cut , mix , L , R , L and R
are dened as before, save that c, w and d are now replaced by c , w and
d . In D we have now the equations

c , = 1 , if = (, ) or = (, ), by (c ) (see 5.3),

w (,) = 1(,) , by (w) (see 9.2),

d,A, = 1 (A ) , if = (, ) or = (, ), by ( dL ) or ( dL ).

Note that now k 1(,), and k 2,(,) are dened in D, and they are equal
to . With this in mind, we may continue using the other abbreviations
we had before.
If 1e denotes (1 , , (, )), and 1i denotes (1 , (, ), ), then for
f : we have in D the equations

()

L
k
f = f k 1 , = f ( ) = f ( 1e ),

()

R
k
f = k 2 , f = ( ) f = (1i ) f.

For (), we rely on the fact that k 2A,B =

B

(A 1B ) in L and

X = 1X in D, and analogously for ().

We prove the Gentzenization Lemma as in 11.1, with the following
L i
additions. We have that (1 , , ) is denoted by k
1 , while (1 , , ) is
R e
denoted by k 1 . To show that, we rely on the equations () and (). We

L i
also have that (A , A, ) is denoted by kA
1 , while (A , , A) is denoted
R e
by kA
1 . (We can dene the arrow terms in the family - in terms of
those in the family ; see 9.2.) If we have a sequent arrow (f, , ) and
is (, ) or is (, ), then we proceed as before by using 1i or 1e .
As before, a Gentzen term is k-atomized when for every subterm of it
L
R
of the form kA
or kA
we have that A is a an atomic formula, which now
means that it is a letter or or . We prove the Atomic-k Lemma of 11.2
exactly as before.

11.5.

Coherence of distributive dicartesian categories

273

L
R
Note that we keep the Gentzen operations k
and k
, which need not
be in the spirit of Gentzen. For f : , Gentzen would perhaps equate
L
k
f with an arrow obtained from 1e : (, ) by thinning with on the
R
left and on the right, and he would equate k
f with an arrow obtained
i
from 1 : (, ) by thinning with on the left and on the right. We
do not do that.
To dene clusters and rank, we now count among the leaf formulae of
h also the occurrences of and in the types of 1i : (, ) and
1e : (, ) when h is one of these Gentzen terms.
With the denition of blocked wL and wR subterms copied from what
we had in 11.2, we can prove the Blocked-w Lemma as before. The degree
of a cut is as before the number of occurrences of connectives in the cut
formula; here we count and among these connectives. We can then
prove the Cut-Elimination Theorem for the new category D.

Proof of the Cut-Elimination Theorem. We enlarge the proof in

11.2 with the following cases.
If the complexity of our clean cut is (d, 2) for d > 0 and the cut formula is
L
, then our clean cut can be of the form cut , (1i , k
f ), which for f :
is equal to f in D by relying on () and the equation d,,(,) = 1 of

D, which we obtain by using the equation ( dL ). The remaining possible

R
L
form of our clean cut can be cut , (k
g, k
f ), which is equal to mix (f, g)
in D. (This last step is not in the spirit of Gentzen, who did not envisage
our mix .)
If the complexity of our clean cut is (d, 2) for d > 0 and the cut formula
is , then we proceed analogously. We rely now on the equations () and

( dL ).
If the complexity of our clean cut is (d, r) for r > 2 and the right rank
of this clean cut is greater than 1, then we proceed as in the proof of the

Cut-Elimination Theorem in 11.2. Note that k 1 (,),B is dened in D and

is equal to B . So case (3) can now be handled without distinguishing
cases as in 11.2.
We proceed analogously if the left rank is greater than 1, and the remainder of the proof follows the proof of the Cut-Elimination Theorem of
11.2.

274

CHAPTER 11.

DISTRIBUTIVE LATTICE CATEGORIES

We can then prove the following.

Distributive Dicartesian Coherence. The functor G from DL, to
Rel is faithful.
We proceed as for the proof of Distributive Lattice Coherence in 11.3. In
the Proposition and in the Normal-Form Lemma of 11.3, and at other
appropriate places, we replace L , L , DL and ML by L,, , L,, ,
DL, and ML, , respectively. We now use Restricted Mix-Dicartesian
Coherence of 10.3 instead of Restricted Mix-Lattice Coherence.
When we look for conditions of legitimacy of relations, we take over for
DL, Proposition 2 of the preceding section as it stands. Proposition 1 of
+l +...+lj1
+k +...+ki1
that section can now have R (ki 1
lj 1
) empty, provided
either is in Ai or is in Bj .
As far as the maximality of DL and DL, is concerned, we conjecture
that DL is not maximal. We conjectured at the end of 10.3 that we could

extend E(ML) with mp,p = mp,p c p,p without falling into preorder. We
conjecture the same thing for E(DL). There are other such equations,
which we will not try to classify here. For DL, we can show that it is
relatively maximal in the same sense in which L, is maximal (see 9.7).
Namely, every distributive dicartesian category that satises an equation
between arrow terms of C(DL, ) that is not in E(DL, ) satises also

the equations (k k) and (k k f g). The argument in 9.7 can be transferred

to the present context to demonstrate this fact. Some of these distributive
dicartesian categories may, of course, satisfy more, as indicated at the end
of 9.7.
It can be shown that the arrows of DL that are isomorphisms are denoted by arrow terms of C(S). So S catches the isomorphisms fragment

of DL (cf. [44] for an analogous result showing that S catches the isomor

phisms fragment of L, and S the isomorphisms fragment of L ). That

can be established by an argument based on coherence and on distinguished
disjunctive and conjunctive normal forms.

Chapter 12
Zero-Lattice Categories
A kind of dual of the operation of union of proofs is the notion of zero proof.
With zero proofs, which are mapped into empty relations in establishing
coherence, we disregard provability in logic. With a zero proof we can pass
from any premise to any conclusion.
We rst prove coherence for categories with nite products and coproducts to which we add zero arrows, i.e. arrows that correspond to zero proofs.
We call such categories zero-lattice categories. Zero arrows amount in this
context to the inverse of the mix principle of Chapter 8. Our technique for
the proof of coherence is based on composition elimination. Maximality,
i.e. the impossibility to extend axioms without collapse into preorder, is
easy to establish for zero-lattice categories.
As an example of a zero-lattice category in which the operations corresponding to conjunction and disjunction are not isomorphic, we have
the category Set of sets with a distinguished object and -preserving
functions. By inverting the operations corresponding to conjunction and
disjunction, we have as a subcategory in every zero-lattice category, and in
Set in particular, a symmetric double monoidal category with dissociativity, and without unit objects, such as those for which we proved coherence
in Chapter 7.
We also consider adding only zero arrows that correspond to proofs
in conjunctive-disjunctive logic, in the sense that we have also non-zero
proofs with the same premises and conclusions. We call such zero arrows
zero-identity arrows. We prove coherence when zero-identity arrows are
275

276

CHAPTER 12.

ZERO-LATTICE CATEGORIES

added to the categories of Chapter 11, and restricted coherence when they
are added to the categories of Chapter 10. These categories are interesting
because Gentzens procedure can be modied to incorporate zero-identity
arrows. The modied procedure can yield coherence not only with respect
to the category whose arrows are relations between nite ordinals, but also
with respect to the category whose arrows are matrices, where composition
is matrix multiplication.

12.1.

Zero-lattice and zero-dicartesian categories

To obtain the natural logical category ZL, we have that the logical system
C(ZL) is in L, , with the transformations included in 1, b, c, w-k and
m1 . The specic equations of E(ZL) are those of E(L) plus

1 2
1
1

(m1 0) k 2A,B m1
A,B k A,B = k B,A mB,A k B,A ,
1 2
1
2

(m1 1) k 1A,B m1
A,B k A,B = k B,A mB,A k B,A = 1A .

We call natural ZL-categories zero-lattice categories. The reason for this

name will become clear below.
Note rst that in ZL we have the equation
1

m1
B,A c B,A = c A,B mA,B ,

(cm1 )

which is dual to the equation (cm) of 8.4. We derive (cm1 ) as follows,

with subscripts omitted:

m1 c = [m1 k 2 , m1 k 1 ], by ( c) and (K 3),

= [k 1 m1 k 2 , k 1 m1 k 1 ], [k 2 m1 k 2 , k 2 m1 k 1 ],

by () and (K 3),

= [k 2 m1 k 1 , k 2 m1 k 2 ], [k 1 m1 k 1 , k 1 m1 k 2 ],
by (m1 0) and (m1 1),

= [k 2 m1 k 1 , k 1 m1 k 1 , k 2 m1 k 2 , k 1 m1 k 2 ],
by (in-out),

= c m1 , by (K 3), () and ( c)

(see the List of Equations at the end of the book for all the equations
mentioned in this derivation).

12.1.

Zero-lattice and zero-dicartesian categories

277

We can derive analogously the following dual of the equation (bm) of

8.2:
(bm1 )

1
1
1

(m1
A,B 1C ) mAB,C b A,B,C = b A,B,C mA,BC (1A mB,C ).

The equations (bm) and (cm) are specic equations for the category MS
of 8.5. The remaining specic equations of E(MS) are delivered by the
equations of E(L). Since we have a functor G from ZL to Rel, and we
have coherence for MS, i.e. the faithfulness of G from MS to Rel, we
can conclude that ZL has a subcategory isomorphic to MSop , with the
isomorphism being identity on objects.
If we assume the equation (cm1 ) as primitive for E(ZL), then the
equation (m1 0) becomes superuous, and from the equations (m1 1) it is

1 2
1
2

enough to keep either k 1A,B m1

A,B k A,B = 1A or k B,A mB,A k B,A = 1A .
A logical system C(0ZL) synonymous with C(ZL) (see the end of 2.4
for this notion of synonymity) is obtained by having as primitive instead
of m1 the transformation 0 whose members 0A,B : A B are called zero
arrow terms. Zero arrow terms, which denote zero arrows, are dened in
terms of m1 by

0A,B =df k 2A,B m1

A,B k A,B

1
2

k 1B,A m1
0),
B,A k B,A , by (m

and m1 is dened in terms of 0 by

m1
A,B =df [1A , 0B,A ], [0A,B , 1B ]
= [1A , 0A,B , 0B,A , 1B ], by (in-out).
It is clear that G0A,B is the empty relation GA GB.
The specic equations of E(0ZL) are those of E(L) plus
f 0A,A = 0B,B f = 0A,B ,
for f : A B. These equations deliver immediately the more general
equations
(0)

f 0C,A = 0C,B ,

0B,C f = 0A,C ,

278

CHAPTER 12.

ZERO-LATTICE CATEGORIES

and they also deliver that the arrows 0A,A make a natural transformation from the identity functor to the identity functor; namely, one of these
equations is
(0 nat) f 0A,A = 0B,B f,
which becomes (1 nat) when 0A,A and 0B,B are replaced by 1A and 1B
respectively. It is straightforward to check that ZL and 0ZL are isomorphic
categories. (The notion of zero arrow satisfying (0) is considered in [100],
Section VIII.2, and [95], p. 279.)
Note that in 0ZL, and hence also in ZL, we have the equations
(0 )
for

0A,C 0B,D = 0AB,C D ,

{, }. To derive (0) we have

0A,C 0B,D = 0A,C k 1A,B , 0B,D k 2A,B , by (),

= k 1C,D , k 2C,D 0AB,CD , by (0) and (K 3),

= 0AB,CD , by () and (cat 1), or by (0),
and we proceed analogously for (0). The equations (0 ), which are analogous to the bifunctorial equations ( 1) of 2.7, are null cases of the equations ( ) of 10.1.
Another logical system C(d1 ZL) synonymous with C(ZL), and hence
also with C(0ZL), is obtained by having as primitive instead of m1 , or 0,
the transformation d1 whose members are
d1
A,B,C : A (B C) (A B) C.
R
The type of d1
A,B,C is converse to the type of dA,B,C . The specic equations
of E(d1 ZL) are those of E(L) plus

1
1

(d1 1) k 1AB,C d1
A,B,C k A,BC = k A,B ,
2
2

(d1 2) k 2AB,C d1
A,B,C k A,BC = k B,C ,
2
2

(d1 3) k 1AB,C d1
A,B,C k A,BC = k A,B

k 1B,C .

That 0ZL and d1 ZL are isomorphic is demonstrated with the following

denitions:

12.1.

Zero-lattice and zero-dicartesian categories

279

0A,C =df k 2AB,C d1

A,B,C k A,BC ,

1
2
d1
A,B,C =df 1A k B,C , [0A,C , k B,C ]

= [k 1A,B , 0A,C , k 2A,B 1C ].

This demonstration is quite straightforward; we will here just check that in
d1 ZL we have

1
1
2
1
2

d1
A,B,C = 1A k B,C , [k AB,C dA,B,C k A,BC , k B,C ].

We show rst with the help of (d1 1), (d1 2) and (d1 3) that the righthand side RHS of this equation is equal to the following arrow term with
subscripts omitted:

[k 1 d1 k 1 , k 1 d1 k 2 ], [k 2 d1 k 1 , k 2 d1 k 2 ].
Then it is enough to establish that

k iAB,C RHS = k iAB,C d1

A,B,C ,
for i {1, 2}, and use ().
1
In C(ZL) we dene d1
in terms
A,B,C by combining the denition of d
1
of 0 and the denition of 0 in terms of m , or by the following arrow term:

(1AB (k 2A,C

m1
A,C )) cA,A,B,C (w A 1BC ).

That we can do so will be clear after we have established coherence for ZL

with respect to Rel. Note that Gd1
A,B,C is an identity relation, i.e. identity
function, in Rel.
The dissociativity principle of the type A (B C) (A B) C
of d1
A,B,C is contained in the trivial part of the modularity law, which is
satised in any lattice:
if a c, then a (b c) (a b) c
(see [9], Sections I.5 and I.7). In the presence of an arrow f : A C, we
have in the logical category L an arrow g : A (B C) (A B) C,

280

CHAPTER 12.

ZERO-LATTICE CATEGORIES

which is dened like d1

A,B,C in terms of 0 save that 0A,C in this denition
is replaced by f . The relation Gg, however, is not an identity relation in
that case, while Gd1
A,B,C is an identity relation.
With ZL we have abandoned the realm of conjunctive-disjunctive logic
as far as provability is concerned. The type A B A B of m1 does
not correspond in general to a logical consequence; namely, the implication
A B A B is not a tautology. In ZL we have 0A,B : A B for any
formulae A and B. This does not mean, however, that we have abandoned
the realm of logic as far as equality of proofs is concerned. The coherence
results that we have show that adding zero arrows will not enable us to
demonstrate new equations between arrow terms in which 0 does not occur;
namely, the extension with zero arrows is conservative. And this extension
can be useful to facilitate calculations (cf. 13.1).
Bits of zero arrows already existed in all our categories whenever we had
and , and this not only in the - and families, but in other families

as well. For example, 1 : , k 2A, : A and w :

all have an empty image in Rel under G, and behave like zero arrows.
We have seen in 8.1 that m is like Gentzens mix (Mischung) where
is the empty sequence. In a similar vein, m1 is related to the following
version of Gentzens mix:
1 1 ,

, 2 2

1 , 2 1 , 2
where either 1 or 2 is empty, while is any nonempty sequence of formulae, and not necessarily a sequence of occurrences of the same formula,
as Gentzen requires. Such a principle is not logically valid as far as provability is concerned, as we have seen above. It is, however, safe to introduce
it if we are interested not in provability, but in equality of proofs.
Note that the following arrow term of C(L):

wA , wA : A A A A,

which is equal in L to [wA , wA ] and to (wA wA ) ckA,A,A,A (wA wA ),

stands behind Gentzens mix, with a sequence of occurrences of the same

formula. However, GwA , wA is dierent from Gm1

A,A .

12.1.

Zero-lattice and zero-dicartesian categories

281

It is pointless to add m1 to the categories I, A and S, with appropriate

equations that guarantee coherence, since the resulting categories would be
isomorphic to MI, MA and MS. We just interchange and . On the
other hand, the categories ZL and ML are not isomorphic. (Compare this
with the remark on the symmetries of A, S and L, made at the end of
9.6.)
To obtain the natural logical category ZL, , we have that the logical
system C(ZL, ) is in L,,, , with the transformations included in 1,
b, c, w-k, - and m1 . The specic equations of E(ZL, ) are obtained by
taking the union of those of E(ZL) and E(L, ). We call natural ZL, categories zero-dicartesian categories.
Note that in ZL, the following equations hold:

A = f = 0A, ,

A = f = 0,A ,

for f : A ,
for f : A,

= = 0, ,

= 1 = 0, ,

= 1 = 0, ,

k 1, = k 2, = 0, ,

k 1, = k 2, = 0, .

The arrow 0, : is the inverse of 0, : , and so and

are isomorphic in ZL, . Hence and are both terminal and initial
objects in ZL, , which means that they are null objects in the sense of
[100] (Section I.5).
We also have in ZL, the equation
(0)

0A,B = B 0, A ,

according to which 0, could be taken as an alternative primitive. The

equations (0) are derivable from (0). So ZL, can be conceived as
obtained from L, just by adding the arrow 0, , without any new equation, the equation (0) being taken as a denition. In this context, the

equations (k ) and (k ) of 9.6 become derivable from the remaining

equations.
We can also conceive of ZL, as being obtained from ZL whose objects
are formulae of L,,, with the additional equations 1 = 0, and

1 = 0, , and the denitions A =df 0A, and A =df 0,A .

282

12.2.

CHAPTER 12.

ZERO-LATTICE CATEGORIES

Coherence of zero-lattice and zero-dicartesian

categories

Our purpose now is to prove the following.

Zero-Lattice Coherence. The functor G from ZL to Rel is faithful.
We proceed by enlarging the proof of Lattice Coherence in 9.4. Throughout this section, we assume that ZL stands for 0ZL, where 0 is primitive.
The syntactical system C(GZL) is dened as C(GL) of 9.4, save that it
has as primitive arrow terms also the arrow terms 0A,B : A B. The equations of E(GZL) are obtained by adding the equations (0) to the equations
of E(GL). The syntactical category GZL, which is C(GZL)/E(GZL),
is isomorphic to ZL. The syntactical category GZL, , isomorphic to
ZL, , is dened as GZL: we just replace L everywhere by L, .
We can prove Composition Elimination for GZL and GZL, by enlarging the proofs of Composition Elimination in 9.1-2 and 9.4. The
equations (0) take care of all the additional cases.
Let a zero term of C(ZL) be dened inductively by: 1A and 0A,B are
zero terms for every A and B; if f and g are zero terms, then f g for
{, } is a zero term. A proper zero term is a zero term in which 0
occurs at least once.
It is easy to show by induction on the sum of the lengths of h1 : A B
and h2 : B C that if h1 and h2 are zero terms, then h2 h1 is equal in ZL
to a zero term. (If at least one of h1 and h2 is a proper zero term, then
h2 h1 is equal to a proper zero term.)
An arrow term g h f of C(ZL) is in standard form when f is an arrow

term of C(L ) and g is an arrow term of C(L ), while h is a zero term (cf.
9.4). Then we can prove the following.
Standard-Form Lemma. Every arrow term of C(ZL) is equal in ZL to
an arrow term in standard form.
Proof. We proceed as in the proof of the Standard-Form Lemma of 9.6,
save for the following additional cases involving 0. If in the proof of () we
have for f g that f : B C is 0B,C or g : A B is 0A,B , then we apply
the equations (0). Here we treat zero terms rst as -factors, and next as

12.2.

Coherence of zero-lattice and zero-dicartesian categories

283

-factors, or vice versa. We also appeal to the fact noted above that the
composition of two zero terms is equal in ZL to a zero term.

We can also prove the following lemma.

Zero-Term Lemma. For h : A B a zero term of C(ZL), if Gh = , then
h = 0A,B in ZL.
Proof. We proceed by induction on the length of h. If h is 0A,B , then
we are done. Here h cannot be 1A , for if it were 1A , then Gh would
not be empty. If h is h1 h2 for {, }, then from Gh = we infer
Gh1 = Gh2 = , and by the equation (0 ) of the preceding section we
obtain h = 0A,B .

We have also the following strengthening of this lemma.

Empty-Relation Lemma. For h : A B an arrow term of C(ZL), if
Gh = , then h = 0A,B in ZL.
Proof. By the Standard-Form Lemma above, we have h = h3 h2 h1 in

ZL, where h1 is an arrow term of C(L ) and h3 is an arrow term of C(L ),

while h2 : C D is a zero term. From Gh = we conclude that Gh2 must
also be empty. So h2 = 0C,D by the Zero-Term Lemma, and hence, by (0),
we obtain that h = 0A,B .

This last lemma entails Lemma 2 of 9.6 for ZL; namely, the assertion
that if for f, g : A B we have Gf = Gg = , then f = g in ZL. From that
we easily obtain the following.
Invertibility Remark for . Let f : A1 A2 B be an arrow term of
C(GL). If Gf = , then f is equal in ZL to an arrow term of the form

K 1A2 f , and to an arrow term of the form K 2A1 f .

This holds because if Gf = , then f = K 1A2 0A1 ,B = K 2A1 0A2 ,B .

By relying on that, we can prove the analogue of the Invertibility Lemma
for of 9.6 where L is replaced by ZL and it is assumed that Gf = . We
proceed as in the proof in 9.4, save in the case when B is B1 B2 and f is

not of the form K iA3i f . Then f must be of the form f1 , f2 (the condition

284

CHAPTER 12.

ZERO-LATTICE CATEGORIES

on Gf precludes that f be an identity arrow term or a zero arrow term).

We apply either the induction hypothesis or the Invertibility Remark for

to f1 : A1 A2 B1 and f2 : A1 A2 B2 , and use the equation (K 3).

We have, of course, for ZL a dual Invertibility Lemma for , based on
a dual Invertibility Remark for . We can then imitate the proof of Lattice
Coherence in 9.4 to prove Zero-Lattice Coherence.
We can also prove the following in an analogous manner.
Zero-Dicartesian Coherence. The functor G from ZL, to Rel is
faithful.
Zero terms are dened as before, and, as before, the composition of
two zero terms is equal in ZL, to a zero term. We have a denition of
standard form for arrow terms of C(ZL, ) analogous to that for C(ZL),
and the analogue of the Standard-Form Lemma for ZL, .
We can also prove the Zero-Term Lemma with ZL replaced by ZL, .
In the basis of the induction, we have to consider the case where h is 1C
for a letterless formula C. This C is isomorphic in ZL, both to and
to . For i : C and i1 : C being inverse to each other, in ZL,
we have
1C = i1 i
= i1 0, i, since 1 = 0, ,
= 0C,C , by (0).

From the Standard-Form Lemma and the Zero-Term Lemma for ZL,
we infer as above the Empty-Relation Lemma with ZL replaced by ZL, .
For the remainder of the proof of Zero-Dicartesian Coherence, we imitate
the proof of Zero-Lattice Coherence.
An alternative, and presumably simpler, way to prove Zero-Dicartesian
Coherence is to rely on the fact that every object of ZL, is isomorphic
to an object of ZL, or to and . Then we use Composition Elimination
for GZL, and Zero-Lattice Coherence.

12.3.

12.3.

Maximality of zero-lattice and zero-dicartesian categories

285

Maximality of zero-lattice and zero-dicartesian

categories

In this section, we show that ZL and ZL, are maximal in the sense of
9.3. We deal rst with ZL.
Suppose that for some arrow terms f1 , f2 : A B of C(ZL) we have
Gf1 = Gf2 . Suppose that for some x GA and some y GB we have
(x, y) Gf1 and (x, y) Gf2 . Then with the help of the arrow terms

k iA1 ,A2 : Ai A1 A2 , for i {1, 2}, together with

1A1 , 0A1 ,A2 : A1 A1 A2 ,
0A2 ,A1 , 1A2 : A2 A1 A2
and the operation of composition, we can build an arrow term h1 : p A
of C(ZL) such that Gh1 = {(0, x)}. The (x+1)-th occurrence of letter in
A counting from the left is an occurrence of p. Analogously, with the help

of k iB1 ,B2 : B1 B2 Bi , together with

[1B1 , 0B2 ,B1 ] : B1 B2 B1 ,
[0B1 ,B2 , 1B2 ] : B1 B2 B2
and composition, we build an arrow term h2 : B p of C(ZL) such that
Gh2 = {(y, 0)}. The (y+1)-th occurrence of letter in B counting from the
left is an occurrence of the same p we had for h1 : p A. This must be
the case because (x, y) Gf1 . Then for h2 fi h1 : p p, where i {1, 2},
we have that G(h2 f1 h1 ) = {(0, 0)}, while G(h2 f2 h1 ) = . It follows
from Composition Elimination for GZL and from the functoriality of G
that in ZL we have
h2 f1 h1 = 1p ,
h2 f2 h1 = 0p,p .
(This follows from Zero-Lattice Coherence too.) So, if we extend E(ZL)
with f1 = f2 , then we can derive 1p = 0p,p .
If an equation f1 = f2 that is not in E(ZL) holds in a zero-lattice
category A, then by Zero-Lattice Coherence we have Gf1 = Gf2 . So 1p =
0p,p holds in A. Then for f, g : a b in A, with 1a = 0a,a and the equations
(0) we obtain

286

CHAPTER 12.

ZERO-LATTICE CATEGORIES

f = f 0a,a = 0a,b = g 0a,a = g,

and hence A is a preorder. This proves the maximality of ZL.
Exactly the same argument serves to prove the maximality of ZL, .
We just replace ZL by ZL, , and appeal to Zero-Dicartesian Coherence.

12.4.

Zero-lattice and symmetric net categories

The category ZL has a subcategory isomorphic to the category DS of 7.6.

We dene a functor F from DS to ZL in the following manner:
F p = p,
F (A B) = F A F B,

F (A B) = F A F B,

F 1A = 1F A ,

F b
A,B,C = b F A,F B,F C ,

F b
A,B,C = b F A,F B,F C ,

F c A,B

= c F B,F A ,

F b
A,B,C = b F A,F B,F C

F b
A,B,C = b F A,F B,F C ,

F c A,B

= c F B,F A ,

F dA,B,C = d1
F A,F B,F C ,
F (f g) = F f F g,

F (f g) = F f F g,

F (g f ) = F g F f.
To show that F is indeed a functor, we have to check that if f = g holds
in DS, then F f = F g holds in ZL (cf. the penultimate paragraph of 2.4).
So suppose that f = g holds in DS; then Gf = Gg in Rel, and we have
GF f = Gf = Gg = GF g in Rel. By Zero-Lattice Coherence, we obtain
that F f = F g holds in ZL.
It is clear that F establishes a one-to-one correspondence on objects.
To show that F is faithful, which here implies that F is one-one on arrows,
suppose that for f, g : A B arrow terms of C(DS) we have F f = F g in
ZL. Hence in Rel we have Gf = GF f = GF g = Gg, and, by Symmetric

12.5.

Zero-identity arrows

287

Net Coherence, we obtain that f = g in DS. So the subcategory of ZL

that is the image under F of DS is isomorphic to DS.
It was shown in 9.7 that Set is a lattice category with being
2 and
+ . If we dene the function 0a,b : a b by 0a,b (x) = , then it is
being 2
easy to check that Set is a zero-lattice category. It is also a zero-dicartesian
category with both and being {}.
By what we have shown above, a subcategory of Set is a symmetric net
+ and being
category with being 2
2. The claim made in [22] (end of Sec+ is a cartesian linearly (alias
tion 3) that Set with being
2 and being 2
weakly) distributive category, which would imply that it is a symmetric
+ c) (a
+ c
net category, is not correct. The functions da,b,c : a
2 (b 2
2 b) 2
L
dened at that place, which are there called L
, and for which one has
da,b,c (x, (, z)) = and da,b,c (x, ) = ((x, ), ) for every x a and every
z c, do not make a natural transformation d. Take a = {x, }, b = {}
and c = {z, }, and let the function h : c c be dened by h(z) = h() = .
Then we have
+ h))(x, (, z))) = ((x, ), ),
da,b,c ((1a
2 (1b 2
+ h)(da,b,c (x, (, z))) = .
((1a
2 1b ) 2
+ is trivially a symThe category Set with being
2 and being 2
metric net category when we take that da,b,c is dened as 0a 2
(b +
(b +
2 c),a 2
2 c) .
With that denition, however, it is neither a distributive lattice category, in
our sense, nor a linearly distributive category, in the sense of 7.9, because

the equations (d k) and (d k), or ( dL ) and ( dL ), would not hold. For the
same reason, it is also not a cartesian linearly distributive category in the
+ c)) (a
+ c
sense of [22]. That no other denition of da,b,c : a
2 (b 2
2 b) 2
can make of Set a distributive lattice category or a cartesian linearly distributive category is shown in 13.2.

12.5.

Zero-identity arrows

Let the natural logical category ZIL in L, be dened as 0ZL save that
the transformation 0 has as members 0A : A A (we write here 0A instead
of 0A,A ), which are zero-identity arrow terms that stand for zero-identity
arrows, and instead of the equations (0) we have only the following consequence of (0):

288

CHAPTER 12.

(0I)

ZERO-LATTICE CATEGORIES

f 0A = 0B g

for f, g : A B. By putting g for f , this equation delivers that 0 is a natural

transformation from the identity functor to the identity functor; namely,
we obtain the equation (0 nat) of 12.1 with 0A,A and 0B,B replaced by
0A and 0B respectively. As other consequences of (0I), we have
f 0A = g 0A ,
0B f = 0B g,
and the following equation:
(00)

0A 0A = 0A .

In ZIL we dene 0A,B by

0A,B =df f 0A
for some arrow term f : A B of C(ZIL). This denition is correct because
we have f 0A = g 0A , as remarked above. We do not have, however,
0A,B in ZIL for every A and B of L, , as we had it in ZL, but only for
those pairs (A, B) where there is an arrow of L of type A B. It is easy
to show, as in 12.1, that the following instance of (0 ):
(0I )

0A 0B = 0AB

holds in ZIL, and if 0A,C and 0B,D are dened, then (0 ) holds too.
We cannot dene in ZIL every m1
A,B : A B A B of ZL, but only
those where there are arrows of the types A B and B A in L; and we
cannot dene in ZIL every d1
A,B,C : A (B C) (A B) C of ZL, but
only those where there is an arrow of type A C in L.
The natural logical category ZIL, in L,,, is dened as ZIL save
that it is based on L, instead of L. With ZIL and ZIL, , contrary to
what we had with ZL and ZL, , we stay within the realm of conjunctivedisjunctive logic as far as provability is concerned (see 12.1).

12.5.

Zero-identity arrows

289

Problems arise for proving coherence for ZIL and ZIL, with equations like

k 1(Ap)q,p

((k 1A,q 0p ) 0q ) = ((k 1A,p 0q ) 0p ) k 1(Aq)p,q

(cf. the end of the revised version of [47] and 7.9 above). We will not
consider this question here.
Let the natural logical categories ZIML and ZIML, be obtained
from ML and ML, respectively by adding the zero-identity arrow terms
and the equations (0I) and
(0) f 0A,B = f
for every arrow term f : A B of C(ZIML) or C(ZIML, ). The arrow
terms 0A,B are dened in terms of 0A as above. Instead of (0), we could
alternatively assume its instance
1A 0A = 1A ,
which yields (0). In ZIML and ZIML, , a hom-set whose arrows are
of type A B is a semilattice with the unit 0A,B (which can be conceived
as either top or bottom).
Restricted Zero-Identity Mix-Lattice Coherence is formulated as Restricted Mix-Lattice Coherence in 10.2, save that ZIML replaces ML.
To prove this coherence result for ZIML we proceed as for ML, with the
following modications.
The syntactical category GZIML diers from the category GML by
having in C(GZIML) the primitive arrow terms 0A : A A besides 1A :
A A; moreover, we assume for it in addition to the equations of E(GML)
the equations (0I) and (0).
For A in dnf and B in cnf, arrow terms of C(GZIML) of type A B
that are in normal form are dened as in 10.2 save that we allow 0p to
replace 1p in arrow terms in atomic bracket-free normal form. Arrow terms
in atomic bracket-free normal form where instead of 1p we have 0p are
called zero atomic bracket-free terms, and those with 1p nonzero atomic
bracket-free terms. We use the same terminology of zero and nonzero

290

CHAPTER 12.

ZERO-LATTICE CATEGORIES

for atomic components. An analogous terminology was used in 10.3. The

settled normal form is dened as for GML, in 10.3.
We prove Composition Elimination for GZIML by extending the proof
for GML. In that proof we apply the equation (00) and the following
equations of GZIML:

K iAj f 0A1 A2 = K iAj (0B f ),

0A1 A2 K iAj g = K iAj (g 0B ),

for i, j {1, 2} such that i = j, which are easily derived with the help of

(0I) and (K 1), as well as

0A1 A2 g1 , g2 = g1 0C , g2 0C ,
[f1 , f2 ] 0A1 A2 = [0C f1 , 0C f2 ],

which are easily derived with the help of (0I) and (K 3), for {, }.
To prove the Normal-Form Lemma of 10.2 with GML replaced by
GZIML, we proceed as in the proof in 10.2, and we use in addition the
equation f g = f for any zero atomic component g. This equation, which
is analogous to the equation (0g) of 10.3, is derivable from (0I) and
(0). We can then prove Restricted Zero-Identity Mix-Lattice Coherence
as Restricted Mix-Lattice Coherence in 10.2. We use in that proof the fact
that if f and g are zero atomic components of the same type, then they
are equal in GZIML by (0I) and (0).
Restricted Zero-Identity Mix-Dicartesian Coherence is formulated analogously by replacing ZIML with ZIML, (cf. also 10.3), and is proved
in the same manner. The equations

= 1 = 0 ,

= 1 = 0
hold in ZIL, , and hence also in ZIML, .
Let the natural logical categories ZIDL and ZIDL, be obtained from
DL and DL, respectively by adding the zero-identity arrow terms and
the equation (0I). In these categories, we have an arrow of type A B i
the implication A B is a tautology.
We dene 0A,B by f 0A , as before, and we can now derive (0) in the
following manner. We have in ZIDL and ZIDL,

12.5.

Zero-identity arrows

291

1A 0A = wA (k 1A,A 1A ) dA,A,A (1A k 2A,A ) (1A 0A ) wA

= wA (k 1A,A 1A ) dA,A,A (1A k 1A,A ) (1A 0A ) wA , with (0I),

= wA (k 1A,A 1A ) k 1AA,A (1A 0A ) wA , by (d k) of 11.1,

= 1A ,

by Zero-Identity Lattice Coherence, or by applying (k 1 nat) and (wk), for

{, } (see 9.1 and the List of Equations). From 1A 0A = 1A we
easily obtain (0), as we remarked above.
Then we can prove the following.
Zero-Identity Distributive Lattice Coherence. The functor G
from ZIDL to Rel is faithful.
Zero-Identity Distributive Dicartesian Coherence. The functor
G from ZIDL, to Rel is faithful.
Let D be now the category obtained as the disjoint union of the strictied category ZIDLA and the trivial category with the single object and
the single arrow 1 : (cf. 11.1). The Gentzen terms for this category
D are dened as in 11.1 with the addition in the basis of the inductive
denition that 0p : p p, which denotes the arrow (0p , p, p) of D, is a
Gentzen term. We prove the Gentzenization Lemma of 11.1 and the CutElimination Theorem of 11.2 as before. In the proof of the Gentzenization
Lemma, we rely on (0I ). To dene clusters and rank, we count among
leaf formulae the occurrences of p in the type p p of 0p . In the proof of
the Cut-Elimination Theorem, we have as the only additional cases, when
the complexity of our clean cut is (0, 2), the left-hand sides of the following
equations of D:
cut , (0p , 1p ) = 0p ,
cut , (1p , 0p ) = 0p ,
cut , (0p , 0p ) = 0p ,
cut , (0p , kpL g) = kpL g,
cut , (kpR f, 0p ) = kpR f.

292

CHAPTER 12.

ZERO-LATTICE CATEGORIES

For the rst two equations we use (cat 1), for the third (00), for the fourth

(k 1 nat), and for the fth (k 2 nat).

For the remainder of the proof of the Cut-Elimination Theorem, we can
proceed as before, but we can also proceed dierently in the cases where
our clean cut is of the form cut (f, wL g) with blocked wL g tied to our clean
cut. In all these casesnamely, (w 1), (w 2), . . . , (w 3)we applied an
equation of D where on the left-hand side we have a single occurrence of
h : , while on the right-hand side we have more than one occurrence
of h; this is always two occurrences of h, except in case (w 3), where we
have three occurrences of h. Now, we can put h on the right-hand side
for exactly one old occurrence of h, and replace the others by the cut-free
Gentzen term h0 obtained from the cut-free Gentzen term h by replacing
every 1p in h by 0p . It can be shown, by induction on the length of the cutfree Gentzen term h, that in D we have h0 = 0 h. This new procedure
would dispense us from applying the equation (wm) in the remainder of
the proof. Instead, we would rely on the following equations of ZIDL:

wA (0A 1A ) mA,A wA = 1A ,

wA (1A 0A ) mA,A wA = 1A ,
which amount to (0).
For the remainder of the proof of Zero-Identity Distributive Lattice Coherence, we proceed as for the proof of Distributive Lattice Coherence,
relying on Restricted Zero-Identity Mix-Lattice Coherence. We proceed
analogously for the proof of Zero-Identity Distributive Dicartesian Coherence, relying on Restricted Zero-Identity Mix-Dicartesian Coherence.
It can be proved that ZIDL is maximal by imitating the proof of the
maximality of ZL in 12.3. The only dierence in the proof is that for
f1 , f2 : A B we assume that only one letter p occurs in A and B (cf. the

proofs of maximality for L and L in 9.3 and 9.5). The category ZIDL,
is maximal in the relative sense in which L, is maximal (see 9.7).
Consider the natural logical categories ZIDL and ZIDL
, that dier
from ZIDL and ZIDL, respectively by rejecting the equation (wm) for
m, or alternatively the idempotency equation ( idemp) for . In these categories hom-sets are not necessarily semilattices with unitthey must be
only commutative monoids. Union of arrows becomes now disjoint union of

12.5.

Zero-identity arrows

293

arrows, or addition of arrows. The possibility indicated above to prove the

Cut-Elimination Theorem by not relying on (wm), but on (0) instead, indicates that we could prove that there are faithful functors G from ZIDL
and ZIDL
, not into Rel, but into the category Mat, which is isomorphic
to the skeleton of the category whose objects are nite-dimensional vector
spaces over a xed number eld, and whose arrows are linear transformations. (Note that this category of vector spaces is a subcategory of the
category Set of 9.7, where the null vector is .)
More precisely, the objects of the category Mat are nite ordinals, i.e.
natural numbers (the dimensions of our vector spaces), and an arrow of type
n m is an n m matrix. Matrices that are images under the functor G
will have entries that are natural numbers. Composition of arrows is matrix
multiplication, and the identity arrow 1n : n n is the nn identity matrix
with the entries 1n (i, j) = (i, j), where is the Kronecker delta.
Every n m matrix M whose entries are only 0 and 1 may be identied
with a binary relation RM n m such that M (i, j) = 1 i (i, j) RM .
Multiplication of such matrices is the same as composition of relations if
we assume that 1+1 = 1.
For the proof of the faithfulness of G from ZIDL and ZIDL
, into
Mat we would rely on restricted coherence results for the natural logical
categories ZIML and ZIML
, , which are obtained from ZIML and
ZIML, respectively by rejecting the equation (wm). These restricted
coherence results are of the same type as those we had for ML, ML, ,
ZIML and ZIML, . In producing the settled normal form, we just do
not rely on the equation ( idemp).
The fact that (wm) does not hold in Mat shows that this equation cannot
be derived from the remaining equations we have used to axiomatize the
equations of ZIDL and ZIDL, .
We conclude our consideration of zero-identity arrows with some remarks on formulae that are isomorphic in their presence. In ZIML we
have the isomorphism
[1A , 0A ], [0A , 1A ] : A A A A,
{p}

whose inverse is mA,A . Let A and B be formulae of the language L, ,

which is the language L, generated by P = {p}. Then it is clear that

294

CHAPTER 12.

ZERO-LATTICE CATEGORIES

there are in the category L arrows of type A B and B A, and hence in

ZIML we have the isomorphism
[1A , 0B,A ], [0A,B , 1B ] : A B A B,
whose inverse is mA,B . It is the easy to conclude that in ZIML and ZIDL
{p}
for every pair (C, D) of formulae of L, such that in each of C and D
there are n 1 occurrences of p we have that C and D are isomorphic.
(The category ZIML generated by {p} is isomorphic to the category ZML
of 13.1 generated by {p}.) If pn stands for any of these formulae, then the
functoriality of G from ZIML and ZIDL to Rel implies that pn and pm
cannot be isomorphic in ZIML and ZIDL for n = m. This characterizes
{p}
completely the formulae of L, isomorphic in ZIML and ZIDL.
{p}
Let L,,, be the language L,,, generated by P = {p}. Then
{p}

every formula of L,,, is isomorphic in ZIDL, to one of the form pn

for n 1, or pm , or pm for m 0, where if m = 0, then pm is
and pm is . To prove that, we use various isomorphisms of ZIDL, ,
among which isomorphisms of the following types are prominent:
pn (pm ) pn+m ,
pn (pm ) pn+m ,
(A ) (B ) (A B) ,
(A ) (B ) (A B) ,
(A ) A ,
(A ) A .
For example, the following isomorphism is of the last of these types:

(1
((
A ) dA,, ) 1 ) b A,, (1A w ).
A

Since in classical logic formulae in the classes pn , pm and pm

are equivalent respectively to p, and , formulae from distinct classes
among these three cannot be isomorphic. And that, within each class,
formulae with dierent superscripts n or m cannot be isomorphic is shown
by appealing to the functoriality of G from ZIDL, to Rel.

Chapter 13
Zero-Mix Lattice Categories

Zero-mix lattice categories are categories with nite products and coproducts, with or without the terminal and initial objects, to which we add the
union operation on arrows of Chapter 10 and the zero arrows of Chapter 12.
This amounts to making products isomorphic to coproducts. In zero-mix
lattice categories hom-sets are semilattices with unit, and these categories
are related to categories whose hom-sets are commutative monoids, like
linear categories, preadditive categories, additive categories and abelian
categories. In zero-mix lattice categories we have dissociativity, and these
categories are distributive lattice categories in the sense of Chapter 11. We
prove coherence for zero-mix lattice categories with the help of composition
elimination and a unique normal form inspired by linear algebra. Zero-mix
lattice categories are maximal, in the sense that it is impossible to extend
their axioms without collapse into preorder.
The category whose arrows are relations between nite ordinals, on
which we relied throughout the book for our coherence results, is a zero-mix
lattice category. This category is isomorphic to a subcategory of another
zero-mix lattice categorynamely, the category of semilattices with unit,
which is itself a subcategory of the category Set of sets with a distinguished
object , whose arrows are -preserving functions.
295

296

CHAPTER 13.

13.1.

ZERO-MIX LATTICE CATEGORIES

Coherence of zero-mix lattice categories

To obtain the natural logical category ZML, we have that the logical system
C(ZML) is in L, , with the transformations included in 1, b, c, w-k, m
and m1 . The specic equations of E(ZML) are obtained by taking the
union of those of E(ML) and E(ZL) plus
(mm1 )

m1
A,B mA,B = 1AB ,

mA,B m1
A,B = 1AB .

So in ZML we have that A B and A B are isomorphic. The equations

(m nat) and (m1 nat) entail each other in the presence of (mm1 ).
We call natural ZML-categories zero-mix lattice categories. The homsets in a zero-mix lattice category are semilattices with unit (see 13.3 for
references concerning related kinds of categories).
According to what we had in 10.1 and 12.1, we can take for ZML
primitives alternative to m and m1 ; for example, we can take and 0,
which are dened in ZML as before (see 10.1 and 12.1).
In ZML the following equations hold:

(k m)

(k m)

k 1A,B = [1A , 0B,A ] mA,B ,

k 2A,B = [0A,B , 1B ] mA,B ,

k 1A,B = mA,B 1A , 0A,B ,

k 2A,B = mA,B 0B,A , 1B .

Here is a derivation of the rst (k m) equation:

1 1
1 2
1
1

k 1A,B m1
A,B = [k A,B mA,B k A,B , k A,B mA,B k A,B ], by (),

= [1A , 0B,A ], by (m1 1) and (m1 0) of 12.1;

then we apply (mm1 ) (for () see the List of Equations at the end of the
book). Alternatively, we can rely on the following equation of ZL:
m1
A,B = [1A , 0B,A ], [0A,B , 1B ].

We proceed analogously for the remaining three equations of (k m) and

(k m).
For f : A B, let us use the following abbreviations for arrow terms of
C(ZML):

13.1.

Coherence of zero-mix lattice categories

297

Z 1C f =df f k 1A,C m1
A,C ,

Z 2C f =df f k 2C,A m1
C,A .

Z 1C f =df m1
B,C k B,C f,
2

Z 2C f =df m1
C,B k C,B f,

With the equations (k m) and (k m), it is clear that in ZML we have

Z 1C f = f, 0A,C : A B C,

Z 1C f = [f, 0C,B ] : A C B,

Z 2C f = 0A,C , f : A C B,

Z 2C f = [0C,B , f ] : C A B.

Then we can infer that for fi : C Ai and gi : Ai C, where i {1, 2},

the following equations hold in ZML:

(Z )

(Z )

f1 , f2 = Z 1A2 f1 Z 2A1 f2 ,
[g1 , g2 ] = Z 1A2 g1 Z 2A1 g2 .

For (Z ) we have

k 1C,C k 2C,C = wCC (k 1C,C k 2C,C ) mC,C wC

= mC,C wC ,
and from that, with (mm1 ), we obtain

1
2

wC = m1
C,C (k C,C k C,C ),

which yields (Z ). We derive (Z ) analogously via

wC = (k 1C,C k 2C,C ) m1
C,C .

For every arrow term f : A B of C(ZML), in ZML we have also the

equation
(0)

f 0A,B = f,

which we encountered already in the preceding chapter (see 12.5). Here is

a derivation of this equation:

1A 0A,A = wA (1A 0A,A ) mA,A wA

= [1A , 0A,A ] mA,A wA

= 1A , by (k m) and (wk) of 9.1;

298

CHAPTER 13.

ZERO-MIX LATTICE CATEGORIES

from that we easily obtain (0) with the help of ( ).

Conversely, we can derive (mm1 ) from (0). Here is derivation, with
subscripts omitted, of the rst equation of (mm1 ):

m1 m = [1, 0, 0, 1] ((k 1 k 1 ) (k 2 k 2 ))

= ((k 1 0) (0 k 2 )) w, with ( ) of 10.1,

= 1, with ( ) of 10.1 and (0);
we proceed analogously for the second one by using m1 = [1, 0], [0, 1].
So (0), or 1A 0A,A = 1A , could replace (mm1 ) for the axiomatization
of E(ZML). The axiom (0) is more appropriate than (mm1 ) if and 0
are primitive instead of m and m1 .
To prove the coherence of ZML with respect to Rel we introduce, in
the style of the preceding chapters, a syntactical category isomorphic to
ZML for which we can prove Composition Elimination. The syntactical
system C(GZML) is formulated by combining what we had for C(GML)
in 10.1 and for C(GZL) in 12.2 (which is based on 9.4), together with

the primitive operations Z iC on arrow terms, for {, } and i {1, 2}.

The equations of E(GZML) are obtained by assuming in addition to the
equations of E(GML) and E(GZL) the equation (0) and the four equa

tions given above immediately after the denitions of Z iC in C(ZML). Note

that in the presence of (0), which is f 0A,B = f , and of the analogous
equation 0A,B f = f , we can replace ( assoc) and ( com) by the single
equation
(f1 f2 ) (f3 f4 ) = (f1 f3 ) (f2 f4 ).
The syntactical category GZML is C(GZML)/E(GZML), and it is isomorphic to ZML.
We can prove Composition Elimination for GZML by extending the
proofs for GML and GZL, which are based on the proof of Composition
Elimination for GL in 9.4 and 9.1.
Next we introduce some denitions analogous up to a point to those
we had in Chapter 10 and 12.5. Arrow terms of C(GZML) of the form
P1 . . . Pn Q1 . . . Qm , where n, m 0, and is of the form 1p or 0p,q , for

some letters p and q, while Pi for i {1, . . . , n} is of the form K 1C , or K 2C ,

13.1.

Coherence of zero-mix lattice categories

299

or Z 1C , or Z 2C , and Qj for j {1, . . . , m} is of the form K 1C , or K 2C , or Z 1C ,

or Z 2C , are called atomic terms. We have a zero atomic term when is of

the form 0p,q (where p and q can be the same letter), and a nonzero atomic
term when is of the form 1p .
Arrow terms of C(GZML) in normal form are dened inductively by
stipulating that atomic terms are in normal form, and that if f and g are
in normal form, then f g is in normal form.
Let f be an arrow term of C(GZML) in normal form, and let f be a
subterm of f such that f is an atomic term, and there is no atomic subterm
f of f with f a proper subterm of f . Then we say that f is an atomic
component of f .
Let p occur in a formula A of L, as the x+1-th occurrence of letter
counting from the left. Then there is a unique atomic term Q1 . . . Qm 1p :
A p such that G(Q1 . . . Qm 1p ) = {(x, 0)}. We say that the word Q1 . . . Qm
is bound to (x, 0).
Let q occur in a formula B of L, as the y +1-th occurrence of letter
counting from the left. Then there is a unique atomic term P1 . . . Pn 1q :
q B such that G(P1 . . . Pn 1q ) = {(0, y)}. We say that the word P1 . . . Pn
is bound to (0, y).
Hence there is a unique word P1 . . . Pn Q1 . . . Qm such that Q1 . . . Qm is
bound to (x, 0) and P1 . . . Pn is bound to (0, y). We say that the word
P1 . . . Pn Q1 . . . Qm is bound to (x, y) GA GB.
An arrow term f : A B of C(GZML) is in settled normal form when it
is in normal form and there is a one-to-one correspondence between the set
GAGB and the set of atomic components of f , such that for every atomic
component P1 . . . Pn Q1 . . . Qm we have that P1 . . . Pn Q1 . . . Qm is bound to
the ordered pair in GA GB corresponding to it. To every ordered pair
in GA GB corresponds either a zero or a nonzero atomic component
depending on whether is of the form 0p,q or 1p . Then we can prove the
following.
Normal-Form Lemma. Every arrow term of C(GZML) is equal in
GZML to an arrow term in settled normal form.
Proof. Take an arrow term f : A B of C(GZML). By Composition

300

CHAPTER 13.

ZERO-MIX LATTICE CATEGORIES

Elimination for GZML there is a composition-free arrow term f : A B

of C(GZML) equal to f in GZML.

Then we apply the equations (K 4) and (K 4) (see 9.1 and 9.4), and
the following equations of GZML:

0AB,C = Z 1B 0A,C ,

0AB,C = K 1B 0A,C ,

0C,AB = Z 1B 0C,A ,

0C,AB = K 1B 0C,A ,

in order to obtain a composition-free arrow term f of C(GZML), equal

to f in GZML, in which every 1 and and every 0 have subscripts that are
letters. This procedure is arbitrary as far as zero arrow terms are concerned:

we could as well base it on K 2 and Z 2 instead of K 1 and Z 1 . (We could

also use (0) to omit zero arrow terms, which will reappear through (0)
in another garb afterwards; see below.)

Next we apply to f the equations (Z ) and the following equations of

GZML:

X iA (f g) = X iA f X iA g,

K iA K jB f = K jB K iA f,

K iA Z jB f = Z jB K iA f,

Z iA K jB f = K jB Z iA f,

Z iA Z jB f = Z jB Z iA f,

for X {K, Z}, i, j {1, 2} and {, }, to obtain an arrow term f

of C(GZML) in normal form equal to f in GZML.
To transform f into an arrow term in settled normal form, we apply the
equations (0) and (0) to put in the missing atomic components, and delete
the atomic components P1 . . . Pn Q1 . . . Qm 0p,p for which we have already the
atomic components P1 . . . Pn Q1 . . . Qm 1p .

It is easy to establish that if f, g : A B are arrow terms of C(GZML)

in settled normal form and Gf = Gg, then the set of atomic components
of f and the set of atomic components of g must be the same set of atomic
terms. We can then easily prove the following.
Zero-Mix Lattice Coherence. The functor G from ZML to Rel is
faithful.

13.2.

Zero-mix lattice and distributive lattice categories

301

Proof. Suppose f, g : A B are arrow terms of C(GZML) in settled

normal form such that Gf = Gg. By the Normal-Form Lemma, we have
the arrow terms f and g in settled normal form such that f = f and g = g
in GZML. By the functoriality of G, we have Gf = Gf and Gg = Gg ;
hence Gf = Gg . So f and g have equal sets of atomic components,
and they must be equal in GZML by applying ( assoc), ( com) and
( idemp). Therefore, f = g in GZML.

We could have used ( assoc), ( com) and ( idemp) to nd a unique

term in settled normal form equal to an arrow term of GZML. The advantage ZML has over DL is that, due to zero arrows, we can reach this
unique composition-free normal form. For DL, a unique cut-free normal
form was not forthcoming.
We will see in the next section that DL is isomorphic to a subcategory
of ZML. For that we rely on Distributive Lattice Coherence and Zero-Mix
Lattice Coherence. The unique normal form we have for ZML can serve
as a substitute for the missing unique normal form of DL. For every arrow
term of C(DL) we take the arrow term of C(GZML) in normal form whose
image under G is the same.

13.2.

Zero-mix lattice and distributive lattice categories

To obtain the natural logical category ZDL, we have that the logical system
C(ZDL) is in L, , with the transformations included in 1, b, c, w-k, d
and m1 . The specic equations of E(ZDL) are obtained by taking the

union of those of E(DS) and E(ZL) plus (d k) and (d k) of 11.1 and (wm)
of 10.1 for mA,A dened by (dm) of 11.1 understood as a denition. Note

that we do not assume here the equations (m e) and (m e), which would
deliver immediately the equations of E(DL ), and hence of E(DL) (see
11.1). We will see below, however, that all the equations of E(DL) are in
E(ZDL).
We will show that the categories ZML and ZDL are isomorphic with
the denition
(md)

dA,B,C =df mAB,C b

A,B,C (1A mB,C )

302

CHAPTER 13.

ZERO-MIX LATTICE CATEGORIES

in C(ZML) (this denition is derived from the equation (b mL) of 8.3; an

alternative denition can be obtained from (b mL)), and the denition of

mA,B in C(ZDL) corresponding to the equation (dm) of 11.1.
It is easy to conclude from Zero-Mix Lattice Coherence of the preceding
section that with (md) all the equations of E(ZDL) plus (dm) hold in
ZML. To show that with the denition of mA,B corresponding to (dm) all
the equations of E(ZML) hold in ZDL, rst we derive easily (m nat) for
ZDL, by using naturality equations of ZDL. Then we infer with the help

of (wk k) for {, } (see 9.1 and the List of Equations) that in ZDL

we have (k 1 k 1 ) (k 2 k 2 ) = m, with subscripts omitted. Next we derive

easily ( ) for ZDL with the help of naturality and bifunctorial equations.
The equation ( ) of 10.1 is derived for ZDL as indicated in 10.1, and
(0) is derived for ZDL as we derived it for ZIDL in 12.5. With all
that, we obtain (mm1 ) in ZDL as in the preceding section. Since in ZDL
we have all the equations of E(ZL), we have also (bm1 ) and (cm1 ) (see
12.1), which together with (mm1 ) yield (bm) and (cm) in ZDL. (The
equation (m nat) follows from (m1 nat) with the help of (mm1 ), but we
relied on (m nat) in the derivation of (mm1 ).) It remains only to derive
for ZDL the equation obtained from (md) by replacing mAB,C according
to (dm); namely,

dA,B,C = (k 1AB,D 1C ) dAB,D,C (1AB k 2D,C ) b

A,B,C (1A mB,C ).

For that it is enough to derive (b mL), as we did it for DL in 11.1, and

use moreover (mm1 ). We can then conclude that ZML and ZDL are
isomorphic categories.
We can infer from Zero-Mix Lattice Coherence that all the equations of
E(DL) are in E(ZDL). Because of the question concerning the indepen

dence of the equations (m e) and (m e) in our axiomatization of DL (see

11.1), it is, however, of some interest to see how these equations are de

rived in E(ZDL). We derive the equations (b mL), (b mL) and (cm) as we

did it for DL in 11.1. Then we derive the equations (m c m ) and (m c m )

as we did it for DL in 11.1. Note that we do not need the equations (m e)

and (m e) for all these derivations. With (mm1 ), we easily obtain from

(m c m ) and (m c m ) the following equations:

13.2.

Zero-mix lattice and distributive lattice categories

303

e A,B,C,D = mAC,BD c m
A,B,C,D (1AB mC,D ),

e D,C,B,A = (m1
D,C 1BA ) c D,C,B,A mDB,CA .

We also have in ZDL

ckA,C,B,D e A,B,C,D (1AB mC,D )

= ((k 1A,C k 1B,D ) (k 2A,C k 2B,D )) w(AC)(BD) mAC,BD

m
cm
A,B,C,D , by (m c ),
= mA,B mC,D , by naturality equations and Lattice Coherence,
from which (m e) follows easily with the help of (mm1 ). We proceed

analogously for (m e) by using (m c m ).

We have seen in 12.4 that Set is a zero-lattice category with being
+ . (It is also a symmetric net category with being

2 and being 2
+ and being
+ were also a
2
2.) If Set with being
2 and being 2

symmetric net category, and (d k) and (d k) were moreover satised, then

+ b would be isomorphic in Set , which is not the case. Note
a
2 b and a 2
that the equation (wm) played no role in inferring above that A B and
A B are isomorphic in ZDL; namely, in deriving the equation (mm1 )

for ZDL. Since the equations (d k) and (d k) hold in the cartesian linearly
distributive categories of [22] (cf. 11.5), this shows that no denition of
+ c) (a
+ c in Set can support the claim made in
da,b,c : a
2 (b 2
2 b) 2
[22] (end of Section 3), which we have already considered in 12.4. Since
products and coproducts are unique up to isomorphism (see [100], Sections
IV.1-2), there is no alternative lattice-category structure to the lattice+ in Set . (This invalidates also
category structure provided by
2 and 2
Proposition 3.4 of [22].)
In ZML, the arrow ckA,C,B,D : (A C) (B D) (A B) (C D) has
an inverse clA,B,C,D : (A B) (C D) (A C) (B D). The natural
transformation cl could be taken as a primitive instead of m and m1 , or
and 0, because, for f, g : C B, in ZML we have the equations

f g = [k 1B,A , k 2A,B ] clB,A,A,B k 1B,A f, k 2A,B g,

0A,B = [k 1B,A , k 2A,B ] clB,A,A,B k 2B,A , k 1A,B ,

which are easily checked with the help of Zero-Mix Lattice Coherence.

304

CHAPTER 13.

ZERO-MIX LATTICE CATEGORIES

There are many ways to dene clA,B,C,D in ZML. One way is

clA,B,C,D =df e A,B,C,D (m1
A,B 1CD ),

and another

clD,B,C,A =df (1DC m1

B,A ) e D,C,B,A .

These two denitions show that the equations (m e) and (m e) of 11.1 are
in ZML immediate consequences of
ckA,C,B,D clA,B,C,D = 1(AB)(CD) ,
clA,C,B,D ckA,B,C,D = 1(AB)(CD) .

13.3.

Coherence of zero-mix dicartesian categories

To obtain the natural logical category ZML, , we have that the logical
system C(ZML, ) is in L,., , with the transformations included in
1, b, c, w-k, -, m and m1 . The specic equations of E(ZML, ) are
obtained by taking the union of those of E(ZML) and E(L, ). We call
natural ZML, -categories zero-mix dicartesian categories.
Zero-mix dicartesian categories are linear categories in the sense of [95]
(p. 279). The dierence is that linear categories need not satisfy (wm),
which amounts to ( idemp). So the hom-sets of linear categories are commutative monoids, and not necessarily semilattices with unit, as in zeromix dicartesian categories (cf. the categories ZML and ZML
, below).
Closely related notions are the notions of Ab-category (or preadditive category) and additive category, where the hom-sets are abelian groups (see
[100], Sections I.8 and VIII.2, and [57], p. 60). These notions enter into the
notion of abelian category (see [100], Section VIII.3, [57], Chapter 2, and
[59], Section 1.591).
The syntactical category GZML, is dened by combining what we
had for GZML and GL, in 9.6. We can then prove Composition Elimination for GZML, as for GZML.
We dene the atomic terms of C(GZML, ) as we did for C(GZML)
in 13.1, save that the indices p and q of 0p,q (but not those of 1p ) can
be replaced by or . Arrow terms of C(GZML, ) in normal form,

13.3.

Coherence of zero-mix dicartesian categories

305

and their atomic components, are then dened analogously to what we had
in 13.1. Let the settled normal form for an arrow term f : A B of
C(GZML, ) be dened as for C(GZML) when GA = and GB = . If
either GA = or GB = , then f is in settled normal form when it is 0A,B .
We can prove as in 13.1 the Normal-Form Lemma where GZML is
replaced by GZML, , with the following additions. We use the following
equations of GZML, :

A = 0A, ,
1 = 0, ,

A = 0,A ,
1 = 0, ,

together with (0) and (0), to remove superuous atomic components. We

can then prove as before the following.
Zero-Mix Dicartesian Coherence. The functor G from ZML, to
Rel is faithful.
We prove the maximality of ZML and ZML, as we proved the maximality of ZL and ZL, in 12.3.
Let ZML and ZML
, dier from ZML and ZML, by omitting
(wm), or alternatively ( idemp), from the specic equations. In these
categories hom-sets are not necessarily semilattices with unitthey must
be only commutative monoids. We can prove that there are faithful functors
from ZML and ZML
, into the category Mat of 12.5. For these proofs
we proceed as for ZML and ZML, . Note that we did not need (wm) for
Composition Elimination in GZML and GZML, . (We needed (wm) for
the cut elimination of DL in Chapter 11, but not for the cut elimination of
ZIDL in Chapter 12; see 12.5.) The settled normal form is now dened
by making every ordered pair from GA GB correspond to a nonempty
set of atomic components bound to that pair; more precisely, a nonempty
set of occurrences of a single arrow term bound to that pair such that each
occurrence is an atomic component (cf. 13.1). This is a multiset based
on a singleton. Whether zero atomic components are duplicated in this
multiset is without importance, but we count the number of nonzero atomic
components; this number corresponds to an entry n 1 in the matrix. The
categories ZML and ZML
, are clearly not maximal, since we can add
the equation (wm) without falling into preorder.

306

CHAPTER 13.

13.4.

ZERO-MIX LATTICE CATEGORIES

The category Semilat

In this section we will consider as an example of a zero-mix dicartesian

category the category Semilat , whose objects are semilattices with unit,
and whose arrows are unit-preserving semilattice homomorphisms. This
category is a subcategory of the category Set of 9.7. Note that Semilat is
not the category Setsl
of 10.1, which is isomorphic to the category Semilat
of semilattices with semilattice homomorphisms.
We want to summarize matters in this section; so we give again the
following denitions from 9.7:
I = {},

a = {(x, ) | x a I},

b = {(, y) | y b I},

a b = ((a I) (b I)) I,
a
2 b = (a b) a b ,
+ b = a b I.
a2
If a1 , , and a2 , , are semilattices with unit, then we dene the
semilattice with unit a1 , , a2 , , , for {, }, as a1
2 a2 , , ,
where
2 corresponds to cartesian product. For and we have the following
clauses (taken over from 10.1):
(x1 , x2 ) (y1 , y2 ) = (x1 y1 , x2 y2 ),
(x1 , x2 ) = (x1 , x2 ) = (x1 , x2 ),
= .
We have that = = I = {} is the trivial semilattice with unit.

The functions k ia1 ,a2 : a1

2 a2 ai , for i {1, 2}, are dened by

k ia1 ,a2 (x1 , x2 ) = xi ,

k ia1 ,a2 () = ;

for fi : c ai , the function f1 , f2 : c a1

2 a2 is dened by
{
f1 , f2 (z) =

(f1 (z), f2 (z)) if f1 (z) = or f2 (z) =

if f1 (z) = f2 (z) = ;

and the function a : a I is dened by a (x) = .

The functions k ia1 ,a2 : ai a1

2 a2 are dened by

13.4.

The category Semilat

k 1a1 ,a2 (x) = (x, ),

307

k 2a1 ,a2 (x) = (, x),

for x = ,

k ia1 ,a2 () = ;
for gi : ai c, the function [g1 , g2 ] : a1
2 a2 c is dened by
[g1 , g2 ](x1 , x2 ) = g1 (x1 ) g2 (x2 ),
[g1 , g2 ]() = ;

and the function a : I a is dened by a () = . (The clauses in the

denitions of k ia1 ,a2 , f1 , f2 , a and a are taken over from 9.7, where

they were given for Set , while the clauses for k ia1 ,a2 and [g1 , g2 ] are taken
over from 10.1, where they were given for Setsl
.)
For f, g : a b, we dene the function f g : a b by
(f g)(x) = f (x) g(x)
(as for Setsl
in 10.1), and, nally, we have the function 0a,b : a b dened
by
0a,b (x) =
(as for Set in 12.4). It is straightforward to check that with all these
denitions Semilat is a zero-mix dicartesian category.
The category Semilat is a subcategory of the category ComMon of
commutative monoids with monoid homomorphisms. By repeating what we
had above, we can show that ComMon, with both and being cartesian
product, and both and being the trivial single-element monoid, is a
natural ZML
, -category. The category Mat of 12.5 is isomorphic to a
subcategory of ComMon, which is itself a subcategory of Set .
Let us summarize in a table the connections between the three subcategories of Set that we had as examples for various kinds of lattice categories
(see 9.7, 10.1, 10.3 and 12.4):

Setsl

= Semilat
Set
Semilat

category

mix-lattice
zero-dicartesian
zero-mix dicartesian

2
+
2

I {x}
I
I

I
I
I

308

CHAPTER 13.

ZERO-MIX LATTICE CATEGORIES

Note that Setsl

is not a dicartesian category, but only a sesquicartesian
category with and as in the table. We also had in 9.7 the dicartesian
category Set , where, in contradistinction to Set , we had that is ; but
this category is not a zero-lattice category.
The category Rel is a zero-mix dicartesian category with both and
being +, and both and being 0; the operation in Rel is union, and
the zero arrows are empty relations.
The category Rel is isomorphic to a subcategory of Semilat . We dene
a functor F from Rel to Semilat by
F n = Pn, , ,
where Pn is the power set of the ordinal n and is binary union of sets;
for R n m and X Pn we have
(F R)(X) = {y m | for some x X, xRy}.
It is straightforward to check that F is a faithful functor from Rel to
Semilat , which is one-one on objects. So the image under F in Semilat is
isomorphic to Rel. The functor F is a strong, but not strict, C(ZML, )functor (see 2.8). The semilattice with unit Pn, , is, up to isomorphism, the free semilattice with unit with n free generators.
It is known that Rel is the skeleton of a category isomorphic to the
Kleisli category of the power-set monad (or triple) on the category of nite
sets with functions (see [90], Section 0.6, p. 32). In this isomorphism, every
relation R n m is mapped to a function fR : n Pm such that
y fR (x) i xRy, and fR can then be extended in a unique way to an
-preserving semilattice homomorphism F R from Pn, , to Pm, , .

Chapter 14
Categories with Negation
In this, nal, chapter of the book we bring to completion our proposed
codication of the proof theory of classical propositional logic. We rst
prove a general coherence result that enables us to pass from coherence
proved in the absence of negation to coherence with a De Morgan negation
added. De Morgan negation is involutive negation that satises the De
Morgan laws, but does not yet amount to Boolean negation.
To obtain Boolean negation, i.e. an operation corresponding to complementation, we need extra assumptions, which, if we want coherence with
respect to the category whose arrows are relations between nite ordinals,
must be zero arrows. The eect of having these zero arrows, which yield the
zero-identity arrows of Chapter 12, is that all theorems, i.e. all propositions
proved without hypotheses, will have zero proofs. Then we prove coherence
for our Boolean categories, by reducing it to a previously proved coherence
result of Chapter 12.
We end this chapter with comments on alternatives to our approach
in categorifying the proof theory of Boolean propositional logic. Besides
the approach through bicartesian closed categories, i.e. cartesian closed
categories with nite coproducts, which with natural assumptions about
negation collapses into preorder, there are alternatives with relations between nite ordinals being replaced by a more complex kind of relation on
the sum of the ordinals in the domain and codomain. We discuss problems
that arise with these alternatives.
We conclude that if our codication of the general proof theory of clas309

310

CHAPTER 14.

CATEGORIES WITH NEGATION

sical propositional logic is acceptable, then this proof theory is simpler than
the general proof theory of intuitionistic propositional logic, codied in bicartesian closed categories, or, equivalently, in a typed lambda calculus
with product and coproduct types. In particular, equality of derivations is
easily decidable for classical logic. The categorial structure of this classical
proof theory is, however, quite rich. It covers all the categorial structures
considered in this book, except the zero-mix lattice structure of Chapter
13, and extends them conservatively with respect to identity of proofs. It
comes close to the zero-mix lattice structure, through which it is related to
linear algebra.

14.1.

De Morgan Coherence

If L is one of the languages L, and L,,, , then let L be the language

obtained by assuming in the denition of L that we have in addition the
unary (that is, 1-ary) connective of negation. The language Lp is, on
the other hand, dened like L save that the set of letters P is replaced by
the union of P and the set P = {p | p P}.
The syntactical system C(I ) is dened by taking rst for its objects

the formulae of L
, ; next, for every A, B L, we have the primitive
arrow terms
1A : A A,
n
A

n
A : A A,

: A A,

r
A,B : (A B) A B,

r
A,B : A B (A B),

r
A,B : (A B) A B,

r
A,B : A B (A B),

and as the operations on arrow terms we have composition, and . Let

the family n-r be the union of the families n , n , r , r , r and r .

The equations E(I ) are obtained by assuming the categorial equations,
the bifunctorial equations for and , and the isomorphism equations

n
A nA = 1A ,

n
A nA = 1 A ,

r
A,B

r
A,B = 1(AB) ,

r
A,B

r
A,B = 1AB ,

r
A,B

r
A,B = 1(AB) ,

r
A,B

r
A,B = 1AB .

14.1.

De Morgan Coherence

311

The syntactical category I is C(I )/E(I ). Due to the presence of categorial and bifunctorial equations, we can easily prove the Development
Lemma (see 2.7) for I (this presupposes a denition of -term where

can be one of the families n , n , r , etc.).

An arrow term of C(I ) is called -directed when does not occur as
p
a superscript in it. The formulae of L
, that are also formulae of L,
are said to be in normal form. We can then prove the Directedness Lemma
(see 4.3) for I . The proof of this lemma is obtained by relying on the
bifunctorial equations for and . This lemma enables us to prove the
following.
I Coherence. The category I is a preorder.
The proof is analogous to the proof of Associative Coherence in 4.3.
Consider the following denitions in C(I ):
1A =df 1A ,

n
A =df nA ,

n
A =df nA ,

r
A,B =df nAB (nA nB ) r A,B ,

r
A,B =df r A,B (nA nB ) nAB ,

r
A,B =df nAB (nA nB ) r A,B ,

r
A,B =df r A,B (nA nB ) nAB ,

(g f ) =df f g,

(f g) =df r
A,B (f g) r D,E ,

(f g) =df r
A,B (f g) r D,E .

It is easy to see that is a functor from I to I op , i.e. a contravariant

functor from I to I . It follows easily from I Coherence that n , n ,

r and r , for {, }, dene natural transformations between functors

dened in terms of the identity functor, the contravariant functor and
the bifunctors . Our ocial denition of logical category does not cover
I , but it is clear how we can extend this denition to cover also categories
like I .

312

CHAPTER 14.

CATEGORIES WITH NEGATION

The syntactical category I,

is dened as I save that its objects are

from L,,, , and we have in C(I, ) the additional primitive arrow terms

: ,

: ,

: ,

: ,

and in E(I,
) the additional isomorphism equations

= 1 ,

= 1 ,

= 1 ,

= 1 .

We call the union of the families , , and .

By extending the proof of I Coherence, we easily obtain the following.

is a preorder.
Coherence. The category I,
I,

In C(I,
) we can introduce the denition of f as in C(I ) with the
following additions:
,
=df n

=df

,
=df n

=df

n
,

n
,

op

to I,
and obtain a functor from I,
.
Let K be a logical category in L. Then the syntactical category K ,
whose objects are formulae of L , will be obtained from K as I is obtained

from the variant of I in the language L, , or as I,

is obtained from the
variant of I in the language L,,, . For example, in the syntactical

system C(DL
, ), whose objects are formulae of L,,, , we will have
besides the primitive arrow terms in the families 1, b, c, w-k, -, m and
d, those in the families n-r and , and the equations of E(DL
, ) will be

).
obtained by assuming the union of those of E(DL, ) and E(I,

p
We dene a functor F from K to K in the language L ; we call
the latter category Kp . The category Kp is exactly like the old logical
category K save that it is generated not by P but by P P (see the end
of 2.7).
We dene a functor F from K to Kp by the following graph-morphism
from C(K ) to C(Kp ):

14.1.

De Morgan Coherence

313

F p = p,
F = ,

for

{, },

F (A B) = F A F B,

for

{, },

F p = p,
F = ,

F = ,

F A = F A,
F (A B) = F A F B, F (A B) = F A F B;
for f : A B in the families n-r and ,
F f = 1F A

(here F A is equal to F B),

F A1 ,...,An = F A1 ,...,F An ,
F (g f ) = F g F f ,
F (f g) = F f F g,

for

{, }.

It is easy to check that F is indeed a functor; namely, f = g in K implies

F f = F g in Kp (cf. the penultimate paragraph of 2.4).
Next, we dene a functor F from Kp to K by the graph-morphism
from C(Kp ) to C(K ) for which we have F A = A, and F f = f . It is
clear that F and F are strict C(K)-functors.
We dene by induction on the length of A L
, the arrow terms

iA : A F A and i1
:
F
A

A
of
C(I
):
A
iA = i1
A = 1A ,

if A is p or p,

iA1 A2 = iA1 iA2 ,

for

{, },

iB = iB n
B,
i(A1 A2 ) = (iA1 iA2 ) r
A1 ,A2 ,

1
1
iA
= i1
A1 iA2 ,
1 A2

for

i(A1 A2 ) = (iA1 iA2 ) r

A1 ,A2 ,

{, },

1
1

iB
= n
B iB ,
1
1

i1
(A1 A2 ) = r A1 ,A2 (iA1 iA2 ),

1
1

i1
(A1 A2 ) = r A1 ,A2 (iA1 iA2 ).

1
If A L
,,, , then we dene the arrow terms iA : A F A and iA :

F A A of C(I, ) with the additional clauses:

314

CHAPTER 14.
i = 1 , for

CATEGORIES WITH NEGATION

{, },

i = ,

i = ,

i1
= ,

i1
= .

1
i
It is clear that we have i1
A = 1A and iA iA = 1F A in I or I, . We
A
can prove the following.

Auxiliary Lemma. For every arrow term f : A B of C(K ) we have

Ff i
f = i1
A in K .
B
Proof. We proceed by induction on the length of f .
Ff i
If f : A B is in the families n-r and , we have that f = i1
A
B

by I Coherence or I, Coherence.
If f is A1 ,...,Ak : M (A1 , . . . , Ak ) N (A1 , . . . , Ak ), then iM (A1 ,...,Ak )
1
1
is M (iA1 , . . . , iAk ) and i1
N (A1 ,...,Ak ) is N (iA1 , . . . , iAk ); we obtain f =
Ff i
i1
A by using ( nat).
B
If f is f2 f1 , then we have
1
Ff i
f2 f1 = i1
2
C iC F f1 iA , by the induction hypothesis,
B
F (f f ) i .
= i1
2
1
A
B

If f is f1 f2 , for {, }, then iA1 A2 is iA1 iA2 and i1

B1 B2 is
Ff i
we obtain f = i1
by
using
bifunctorial
equations.

A
B

1
i1
B1 iB2 ;

K -Kp -Equivalence. The categories K and Kp are equivalent via the

functors F and F .
Proof. We have F F A = A and F F f = f . That i is a natural isomorphism from the identity functor of K to the composite functor F F
is shown by the Auxiliary Lemma.

Let the functor G from K to Rel be dened by extending the denition

of the functor G from K to Rel with the clauses
GA = GA,
Gf = 1GA ,
for every arrow term f : A B in the n-r and families. Here GA must
be equal to GB, and 1GA is the identity relation, i.e. identity function, on
GA. Then we can prove the following.

14.1.

De Morgan Coherence

315

De Morgan Coherence. If G from K to Rel is faithful, then G from K

to Rel is faithful.
Proof. Suppose that for the arrow terms f, g : A B of C(K ) we have
Gf = Gg. Then we have GF f = GF g, where F is the functor from K
to Kp we have dened above. Since G from K to Rel is faithful, we have
that G from Kp to Rel is faithful and hence F f = F g in Kp . From
K -Kp -Equivalence we conclude that f = g in K .

We can dene the functor from K to Kop by extending the deni

tions we have for I and I,
with the following clauses, provided C(K )
has the required arrow terms on the right-hand side:

b
A,B,C =df r A,BC (1A r B,C ) b A,B,C (r A,B 1C ) r AB,C ,

b
A,B,C =df r AB,C (r A,B 1C ) b A,B,C (1A r B,C ) r A,BC ,

=df r
A,B

=df wA r
A,A ,

c A,B
wA

k iA,B

c A,B

r
B,A ,

=df r
A,B

for i {1, 2},

k iA,B ,

dL
A,B,C =df r A,BC (1A r B,C ) dA,B,C (r A,B 1C ) r AB,C ,

dR
C,B,A =df r CB,A (r C,B 1A ) dC,B,A (1C r B,A ) r C,BA ,

mA,B

=df r
A,B mA,B r A,B ,

m1
A,B

=df r
A,B mA,B r A,B ,

(f g) =df f g,
0A,B

=df 0B,A ;

the clauses for b

A,B,C and b A,B,C are obtained from the clauses for

b
A,B,C and b A,B,C respectively by interchanging and ;

c A,B =df r
B,A

wA

c A,B

r
A,B ,

=df r
A,A w A ,

k iA,B =df k iA,B

r
A,B , for i {1, 2};

316

CHAPTER 14.

CATEGORIES WITH NEGATION

=df r
A, (1A ) A ,

=df
A (1A ) r A, ,

1A ) ,

=df r
,A (
A
A

1A ) r ;
=df

,A
A
A (

the clauses for

A , A , A and A are obtained from the last
four clauses by interchanging with , and with ;

A =df A

A =df A .

The coherence of K is a sucient (though not a necessary) condition for

the correctness of these denitions. By the coherence of K , we also obtain

that n , n , r and r for {, } dene natural transformations

between functors dened in terms of the identity functor, the contravariant
functor and the bifunctor .
The category DL corresponds in the following sense to the system Efde
of tautological entailments of [1] (Section 15, and Section 18 by J.M. Dunn;
see also [6]): there is an arrow of type A B in DL i A B is a theorem
of Efde . The algebraic models with respect to which Efde is complete are
called De Morgan lattices, distributive involution lattices or quasi-Boolean
algebras (see [1], Section 18, and references therein; see also [113], Section
III.3). Complementation in De Morgan lattices is not in general Boolean
complementation (see the next section).

14.2.

Boolean Coherence

The syntactical system C(B) is dened by taking rst for its objects the
formulae of L
,,, ; next, the primitive arrow terms of C(B) are those of
C(ZIDL
),
i.e. those in the families 1, b, c, w-k, -, m, d, 0 (whose
,
members are 0A : A A), n-r and , plus
A : A A,
A : A A ,
for every A in L
,,, , and the operations on arrow terms are composition,
and . The equations of E(B) are obtained by assuming the union of those

14.2.

Boolean Coherence

317

of E(ZIDL, ), i.e. those of E(DL, ) plus (0I), and those of E(I,

). The
syntactical category B is C(B)/E(B). Here, B comes from Boolean.

We dene 0A,B in B by f 0A , as we did in ZIDL, , and we can infer

that the following equation holds in B:
A = A 1 ,

= A 0 , since 1 = = 0, ,
= 0,AA .
We derive analogously A = 0AA, . Since A and A are zero arrows, we
assume that GA and GA are empty relations.
The following equation holds in B:
(0)

(
0A =
A 1A ) dA,A,A (1A A ) A ,
A

since
A = A 0AA
= A (0A 0A ), by (0I ) of 12.5,
which with naturality and bifunctorial equations yields that the right-hand
side RHS of (0) is equal to RHS 0A . The equation (0) shows that we
need not take 0A as a primitive arrow term: we can take it as dened in
terms of A and A . We could then conceive of B as obtained by extending
DL
, with the arrows A and A and the equations (0I) for dened 0A .
The following equations too hold in B:
()

BA = (r
B,A 1BA ) c BA,BA e B,A,B,A

( c B,B c A,A ) (B A )
,

()

BA = (r
B,A 1BA ) e B,A,B,A (B A ) ,

()

AB =
(A B ) e A,B,A,B (1AB r A,B ),

()

AB =
(A B ) ( c A,A c B,B ) e A,B,A,B

c
(1
AB,AB
AB r A,B ),

()

= ( 1 )
,

()

= ( 1 )
,

()

=
(1 ),

()

=
(1 ).

318

CHAPTER 14.

CATEGORIES WITH NEGATION

The syntactical system C(C) is dened by taking rst for its objects

the formulae of Lp
,,, , namely L,,, generated by P P ; next, the
primitive arrow terms of C(C) are those of C(ZIDL, ), i.e. those in the
families 1, b, c, w-k, m, d, - and 0 (whose members are 0A : A A) plus
p : p p,
p : p p ,
for every p P, and the operations on arrow terms are composition,
and . As equations of E(C) we assume those of E(ZIDL, ), i.e. those of
E(DL, ) plus (0I). The syntactical category C is C(C)/E(C).
We dene 0A,B in C by f 0A , as we did in ZIDL, and B, and as in
B we infer p = 0,pp , p = 0pp, and (0) with A replaced by p.

The category C is obtained from C as I,

is obtained from I, ;
namely, we have as objects the formulae of L
,,, , we add to C the

arrows n-r and , and we assume in addition the equations of E(I,

).

The category C is isomorphic to the category B. We dene A and

A in C inductively with the help of ( ) and ( ) for {, }, and
( ) and ( ) for {, }. Our purpose now is to show that the functor
G from C to Rel is faithful, and use this coherence result, together with
results analogous to K -Kp -Equivalence and De Morgan Coherence, to
infer that G from B to Rel is faithful, i.e. that B is coherent.
Let A[] be a formula of Lp
,,, with a particular occurrence of ,
and let A[B] be obtained from A[] by replacing this particular occurrence
of by the formula B of Lp
,,, . Then it is clear that there is an -term
A[p ] : A[] A[p p].

We dene the arrow term g A[B] : A[] B A[B] of C(DL, ) by

induction on the length of A[]:

g B =
B : B B,

g CA[B] = (1C g A[B] ) b

C,A[],B : (C A[]) B C A[B],

g A[B]C = ( g A[B] 1C ) ( c B,A[] 1C ) b

B,A[],C

c A[]C,B :

(A[] C) B A[B] C,

g CA[B] = (1C g A[B] ) dR

C,A[],B : (C A[]) B C A[B],

14.2.

Boolean Coherence

319

g A[B]C = ( g A[B] 1C ) ( c B,A[] 1C ) dL

B,A[],C c A[]C,B :

(A[] C) B A[B] C.
Then we can prove that the following equations hold in C:
(g)

A[p ] = g A[pp] (1A[] 0pp ) (1A[] p )

A[] .

We proceed by induction on the length of A[]. If A is , then

pp (1 0pp ) (1 p )

= 0pp p , by ( ) and naturality equations,

= p , by (0I) and 0 = = 1 .
If A is C D[], then

(1C g D[B] ) b
C,D[],B (1CD[] 0pp ) (1CD[] p ) CD[]

= (1C g D[B] ) (1C (1D[] 0pp )) (1C (1D[] p ))

nat), ( b ) and isomorphisms,

(1C
D[] ), by ( b

= 1C D[p ], by the induction hypothesis.

We proceed analogously when A is D[] C, C D[] and D[] C,
by applying naturality equations and Distributive Dicartesian Coherence.
For A[] a formula of Lp
,,, with a particular occurrence of , and
A[B] obtained from A[] by replacing this particular occurrence of by
B, we have an -term A[q ] : A[q q] A[].

We dene the arrow term g A[B] : A[B] A[] B of C(DL, ) by

g B =
B : B B,

g CA[B] = b
C,A[],B (1C g A[B] ) : C A[B] (C A[]) B,

g A[B]C = c A[]C,B

b
B,A[],C ( c B,A[] 1C ) ( g A[B] 1C ) :

A[B] C (A[] C) B,

g CA[B] = dL
C,A[],B (1C g A[B] ) : C A[B] (C A[]) B,

g A[B]C = c A[]C,B dR
B,A[],C ( c B,A[] 1C ) ( g A[B] 1C ) :

A[B] C (A[] C) B.

320

CHAPTER 14.

CATEGORIES WITH NEGATION

Then, in a dual manner, we can prove by induction on the length of

A[] that the following equations, analogous to (g), hold in C:

A[q ] =
A[] (1A[] q ) (1A[] 0qq ) g A[qq] .

(g)

From (g) and (g), we easily infer with naturality and bifunctorial equations the following equations of C:
for f : B A[],

(gf ) A[p ] f = g A[pp] (f 0pp ) (1B p )

B,
for f : A[] C,

(1
f A[q ] =
C q ) (f 0qq ) g A[qq] .
C

(gf )

Note that in the relations G(A[p ] f ) and G( g A[pp] (f 0pp )) we

have the same sets of ordered pairs, and in the relations G(f A[q ]) and

G((f 0qq ) g A[qq] ) we also have the same sets of ordered pairs.
Suppose now that we have two arrow terms f1 , f2 : B C of C(C) such
that Gf1 = Gf2 . Let p1 , . . . , pn , with n 0, be the set of all occurrences of
letters that are subscripts of subterms of f1 and f2 of the form pi , where
i {1, . . . , n}, and let q1 , . . . , qm , with m 0, be the set of all occurrences
of letters that are subscripts of subterms of f1 and f2 of the form qj ,
where j {1, . . . , m} (the same letter may be repeated in p1 , . . . , pn , or
q1 , . . . , qm ).
We introduce the following abbreviations by induction:
0
= 1B ,
k+1

= k pk+1 ,

00
= 1B ,

0 k+1 = 0 k 0pk+1 pk+1 ,

B 0

B k+1

k+1

B k

= C,

C k+1

= C k ,

Bk

C 0

= 1B ,
=

00
= 1C ,

0 k+1 = 0 k 0qk+1 qk+1 ,

= B,
=

0
= 1C ,
k+1

= k qk+1 ,

k,

k+1

= 1C ,

= k

k
C

14.2.

B p0

B pk+1

h0f

Boolean Coherence

= B,
=

(pk+1 pk+1 ),

B pk

hk+1
=
f

C q0

= C,

C qk+1

= C qk (qk+1 qk+1 ),

h0f

= 1B ,

hk+1
=
f

321

hkf 1pk+1 pk+1

hf

k 1

if pk+1 is not in f ,

B pk ,pk+1 pk+1

hkf 1qk+1 qk+1

if qk+1 is in f

k k

C q ,qk+1 qk+1

= 1C ,
if pk+1 is in f

hkf

if qk+1 is not in f .

Then, by relying on (gf ) and (gf ), for i {1, 2} we obtain in C the

equations

fi = m m 0 m hm
fi fi hfi

0 n n n

where fi is an arrow term of C(ZIDLp

, ), and for fi being

0 m hm
fi fi hfi

0n

we have Gf1 = Gf2 . Since f1 and f2 are also arrow terms of C(ZIDLp
, ),
by Zero-Identity Distributive Dicartesian Coherence of 12.5 we conclude
that f1 = f2 in ZIDLp
, , and hence f1 = f2 in C. This establishes that
the functor G from C to Rel is faithful.
We prove as in the preceding section K -Kp Equivalence for Kp being
C and K being C . (We stipulate that F p = p and F p = p for
{, }.) Then, as in the proof of De Morgan Coherence, we use the
faithfulness of G from C to Rel and K -Kp Equivalence to establish that
G from C to Rel is faithful. The isomorphism of the categories C and
B then yields the following.
Boolean Coherence. The functor G from B to Rel is faithful.
We can dene the functor from B to Bop by extending the denitions
in the preceding section with

322

CHAPTER 14.

CATEGORIES WITH NEGATION

A =df A (n
A 1A ) r A,A ,

A =df r
A,A (1A nA ) A .

The transformations and are dinatural transformations (see [100],

Section IX.4), which means that for f : A B we have
(1A f ) A = (f 1B ) B ,
A (1A f ) = B (f 1B ).
These equations are satised trivially in B, because A , B , A and B are
zero arrows.
We leave open the question of maximality for the category B. In the
light of the results of 9.7, it seems natural to conjecture that this category is
not maximal. Note, however, that the category ZIDL, which is isomorphic
to a subcategory of B (see Chart 3), and covers the conjunction-disjunction
fragment of classical propositional logic, is maximal (see 12.5). For the
conjunction-disjunction fragment of classical propositional logic with the
constants and , we have the category ZIDL, , which is also isomorphic to a subcategory of B, and is maximal in the relative sense in which
L, is maximal (see 9.7 and 12.5). The technique of 9.7 suggests how
to prove some sort of relative maximality also for B.

14.3.

Boolean categories

A distributive dicartesian category A for which we have a functor from

A to Aop , natural isomorphisms like those in the families n-r and , and
dinatural transformations and will be called a Boolean category.
A Boolean category is called a zero-identity Boolean category when for
every 0a dened by (0) of the preceding section, where A is replaced by
the object a, we have for every f, g : a b the equation (0I), namely f 0a =
0b g. The category B is the zero-identity Boolean category generated by
P.
The connection of Boolean categories with Boolean algebras is the following. We have an arrow of type A B in B i A B is a tautology of
propositional logic. This is how B is connected to classical propositional

14.3.

Boolean categories

323

logic. A partially ordered Boolean category, which must be a zero-identity

Boolean category, is a Boolean algebra (in which top and bottom are not
necessarily distinct).
Note that the equations of B cover cut eliminationi.e. they enable us
to prove a cut-elimination theorem, such as we had in 11.2. They cover
rst the cut elimination of ZIDL, . As far as negation is concerned, the
key equation is (0), whose right-hand side corresponds to the cut
f : (, ) A A

g : A A (, )

cut (,),(,) (f, g) : A A

With (0), we have that cut (,),(,) (f, g) is not equal to 1A : A A, but
to 0A : A A.
To prove a cut-elimination theorem, we can rely on Gentzen terms like
those in 11.1, to which we would add dual primitive Gentzen terms and
Gentzen operations where A B, A B, and are replaced respectively by (A B), (A B), and . For example, we would
have the operation
f : A B
L f =dn f ( 1eA B ) ( rA,B ) : (A B)

Our strictication result should be adjusted to support such operations. As

additional primitive Gentzen terms, we would have 1p : p p, 0p : p p,
0p : p p, p : (, ) p p and p : p p (, ), and we would have
the additional Gentzen operations
f : A

L f =dn f (nA
) : A

f : A
R f =dn (nA ) f : A
A similar idea underlies a sequent system in [1] (Section 17) for tautological
entailments, which correspond to De Morgan lattices.
The usual introduction and elimination rules for negation:

324

CHAPTER 14.

CATEGORIES WITH NEGATION

f : A

f : A

L f : A

R f : A

would be admissible in the cut-free system; i.e., we can nd cut-free Gentzen

terms that dene L f and R f . This does not mean that the operations
L and R are dened in terms of the postulated Gentzen operations; in
such a case, we would speak of derivable rules. The arrow terms L f and
R f involve zero arrow terms.
Zero-identity arrows make equal in B many arrow terms of the same
type involving negation. In particular, all arrow terms of the type A
where A is a tautology are equal. However, B is far from being a preorder.
There is an argument from which it is usually concluded that it is hopeless to try to nd a categorication of Boolean algebras. All plausible
candidates seem to be categories that are preorders. To present this argument, we rely on notions dened in [90]. The argument is based on the
fact that in every bicartesian closed category (i.e. cartesian closed category
with nite coproducts), for every object a there is at most one arrow of
type a , for an initial object. In [90] the discovery of that fact is
credited to Joyal (p. 116), and the fact is established (on p. 67, Proposition
8.3) by relying on a proposition of Freyd (see [58], p. 7, Proposition 1.12)
to the eect that if in a cartesian closed category the hom-set Hom(a, )
is not empty, then a
= ; that is, a is isomorphic to . Here is a simpler
proof of the same fact (from [40], Section 5).
Proposition 1. In every cartesian closed category with an initial object
we have that Hom(a, ) is either empty or a singleton.

Proof. In every cartesian closed category with we have k 1, = k 2, :

, because Hom( , )
= Hom(, ). Then for f, g : a

we have k 1, f, g = k 2, f, g, and so f = g.

In [90] (p. 67) it is concluded from Proposition 1 that if in a bicartesian

closed category for every object a we have a
= a, where the negation b
b
is (which corresponds to b ), then this category is a preorder.
If the requirement a
= a is deemed too strong, here are other simi-

14.3.

Boolean categories

325

lar propositions (taken over from [40], Section 5), in which preordering is
inferred from other natural requirements.
Proposition 2. Every cartesian closed category with an initial object in
which we have a natural transformation whose members are n
a : a a
is a preorder.
Proof. Take f, g : a b, and take the canonical arrow n
a : a a,
which we have by the cartesian closed structure of our category. Then we

have (f n
a ) = (g na ) by Proposition 1, and from

(f n
n
(g n
b
a ) = nb
a ),

by the naturality of n , we infer

f n
na .
a
a = g na
n = 1
Since n
a by Proposition 1, we have f = g.
a
a
Then, for terminal, we have

Hom(c, d)

= Hom(, dc )

= Hom(, dc ),

, and Hom(, dc ) is at most a singleton, as we have

since =
shown above.

Proposition 3. Every bicartesian closed category in which we have a

dinatural transformation whose members are a : a a is a preorder.
Proof. Take f, g : a. Then f = g by Proposition 1, and from
(f 1a ) a = (g 1a ) a ,
by the dinaturality of we infer
(1 f ) = (1 g) .
Since : is an isomorphism, we obtain 1 f = 1 g,

from which f = g follows with the help of .

326

CHAPTER 14.

CATEGORIES WITH NEGATION

(For an argument along similar lines, based on De Morgan isomorphisms,

see [124].)
With Boolean categories, and with B in particular, we have, however,
that A is isomorphic to A, and we have also that n is a natural transformation and a dinatural transformation, without falling into preorder.
We believe that our notion of Boolean category, which does not imply preorder, gives a reasonable categorication of the concept of Boolean algebra.
Moreover, this notion delivers cut elimination, as we have indicated above.
That A and A ended up by being zero arrows in B is dictated by
Rel, since we have no other choice in Rel for GA and GA save the empty
relation. With another category, replacing Rel for coherence results, we
need not take A and A as zero arrows.
A category that could replace Rel is the category whose objects are
nite ordinals and whose arrows are split equivalence relations (see [50] and
[51]). These are equivalence relations dened on the sum of the ordinals in
the source and target. For that category, Gp and Gp would correspond
to the diagrams


Gp
p p

p p

Gp

In that context, the left-hand side 0A of (0) would be replaced by 1A ,

and that equation would become similar to a triangular equation of adjunction (see [100], Section IV.1). But in this direction there is a heavy price
to pay. The transformations in the families w-k, - and cannot remain
natural if we want coherence. (Lack of naturality for these transformations
jeopardizes cut elimination.) For example, for the following instance of

(w nat):

wpp p = (p p ) w

we would not have that G(wpp p ) is equal to G((p p ) w ), as

can be seen from the diagrams

14.3.

Boolean categories

327



Gp

p p

QQ QQ
G wpp
Q Q
(p p) (p p)

G w

 
G(p p )
(p p) (p p)

One may perhaps envisage a categorication of Boolean algebras where

these transformations are not always natural, as they are in our Boolean
categories. (In [67], Section 5.4, rejecting bifunctoriality is envisaged for the
same purpose; we have no reason to reject bifunctoriality here.) Problems
would, however, not cease once naturality is rejected for w-k, - and in
the presence of negation.

The question is should G wp be the relation in the left one or in the

right one of the following two diagrams:
p

A


A
p p

p


A


A
p p

The second option, induced by dealing with equivalence relations, or by

connecting all letters that must remain the same in generalizing proofs,

would lead to abolishing the naturality of w even in the absence of negation.

For example, in the following instance of (w nat):

wp p = (p p ) w

we do not have that G(wp p ) is equal to G((p p ) w ):

G p



A
G wp


A
p p

G w

G(p p )
p p

We obtain similarly that cannot be natural.

If, on the other hand, we keep for G wp the relation in the left diagram
the same we had in Relthere would still be problems. We foresee problems
for cut elimination based on equality of arrow terms, such as we understood

328

CHAPTER 14.

CATEGORIES WITH NEGATION

it in this book. This would, however, involve dealing with matters outside
of the scope of the book.
Although in the introduction we have motivated the category Rel by
the Generality Conjecture, it should be stressed that this category does not
always correspond to the intuitive idea of generality. This is so even if we do
not consider the split equivalence relations of [50] and [51], but stay within
Rel. For arrows of Rel capturing generality it is natural to assume that
they are difunctional in the sense of [114] (Section 7), a binary relation R
being difunctional when R R1 R R (in other words, if xRz, yRz and
yRu, then xRu). It is easy to see that the image under G of an arrow of DL

is not necessarily difunctional. For example, G(mp,p (k 1p,p k 2p,p )) is not

a difunctional relation. The claim made in [40] (Section 4) that the image
under G of any arrow of L is difunctional is not correct. A counterexample

is provided by G[k 1p,p , 1p ], [k 2p,p , 1p ], which is not difunctional.

The category Rel captures, however, the intuitive idea of generality for
all categories in Chart 1 except for L and the three categories above L. It
captures this idea for the category MDS, too, and for all categories below
MDS in Chart 2.

14.4.

Concluding remarks

Our coherence results show that a number of logical categories that we have
investigated here are isomorphic to subcategories of the Boolean category
B, and B is isomorphic to a subcategory of ZML
, . We record all these
results about subcategories in Charts 1-3. Such results about subcategories
are sometimes taken for granted, and, indeed, they are not surprising, but
it is not trivial to establish them. One means of proving them is via coherence, which, as we have seen, is often established with considerable eort.
(Another means can be via maximality.)
In general, we have the following situation. Suppose a syntactical system
S is a subsystem of a syntactical system S. Suppose also that we have the
syntactical categories S /E and S/E such that the set of equations E is
a subset of E, the functor G from S /E to Rel is faithful, and there is a
functor from S/E to Rel that extends G. Then S /E is isomorphic to a
subcategory of S/E, with the isomorphism being identity on objects.

14.4.

Concluding remarks

329

To show that, it is enough to show that the identity maps on the objects
and arrow terms of S induce a functor from S /E to S/E that is inclusion
on objects and one-one on arrows (see the penultimate paragraph of 2.4).
This amounts to showing that for f and g arrow terms of S of the same
type we have f = g in S /E i f = g in S/E. It is clear that f = g in
S /E implies f = g in S/E, since E is included in E. For the converse,
from f = g in S/E we infer Gf = Gg, and then, by the faithfulness of G
assumed above, we obtain that f = g in S /E (cf. the end of 4.3). Note
that S /E is only isomorphic to a subcategory of S/E, and is not actually
a subcategory of S/E, because an arrow term f of S stands in S /E for an
equivalence class of arrow terms (this is an arrow of S /E ) that is a subset,
maybe proper, of the equivalence class for which f stands in S/E (see 2.3).
Note that if for S a subsystem of S and E a subset of E we have that
S /E is a preorder, then we can ascertain that S /E is isomorphic to a
subcategory of S/E without appealing to the functor G and Rel. We have
such a situation with many of our categories where coherence amounts to
preorder, but we also have it where preorder does not amount to coherence,

as with the categories S and S of 6.5.

There are coherence results with respect to Rel, related to the coherence
results of this book, about categories that have arrows of the w kind, but
not those of the k kind, and vice versa, about categories that have arrows
of the k kind, but not those of the w kind. These categories are tied to
substructural logics: the former to relevant logic, and the latter to ane
logic. These coherences are proved in [108] for logical categories in the

language L, in between S and L .

Speaking of categories tied to relevant logic, there is in the neighbourhood a sort of category interesting for strictication. If to the relevant

natural logical category in L , which is like L save that it lacks the natural

transformations k i , we add the natural isomorphism w1 , whose members

1

are wA : A A A, and whose inverse is w, then we obtain a groupoid

that is a preorder. (The logical principle standing behind w1 is called

mingle in relevant logic; see [1], Section 8.15.) If this groupoid, which is
a preorder, happens to be the category G involved in the strictication of
some category, the equivalence classes introduced by strictication will cor

respond to nite nonempty sets. When G is S plus c A,A = 1AA , then we

330

CHAPTER 14.

CATEGORIES WITH NEGATION

have multisets instead of sets (cf. 7.7), and with A we have sequences (see
4.5).

We need not, however, assume that w1 is an isomorphism. We can keep

1
w
just w1
A = 1A , reject w A w A = 1AA , and have only equations that
A

will yield coherence with respect to Rel with the assumption that G w1
A is

1
2
equal to G(k 1A,A k 2A,A ). Dening w1
A by k A,A k A,A we have all these

equations in DL, and hence also in B. (Another possibility is to dene w1

A

as just k 1A,A , or just k 2A,A .)

If we are right that B provides a reasonable notion of identity of proofs
in classical propositional logic, and if bicartesian closed categories provide
the notion of identity of proofs in intuitionistic propositional logic, we can
conclude that the general proof theory of the former logic is simpler than
that of the latter. Equality of derivations in classical propositional logic,
i.e. equality of arrow terms in B, would be decided via Rel, in an elementary
way. It was assumed before that classical general proof theory should be
simpler, because it was assumed that all derivations with the same premise
and conclusion are equal in classical logic. In other words, it was assumed
that for a given premise and conclusion we cannot have more than one proof.
We do not agree with that, and though we provide with B a relatively simple
codication of that proof theory, it is not that simple.
It is true that all theorems, i.e. all propositions proved without hypotheses, will have zero proofs, which is not the case in the standard formulations
of intuitionistic general proof theory. If we are right, with the theorems of
classical logic we do not nd a record of the deductive metatheory. But
this metatheory of proofs from hypotheses exists, and it is not trivial. Our
charts (see Charts 1-3) give an idea of the number of important mathematical structures that enter into the notion of Boolean category. We can
also note in Chart 3 how with B we have come close to ZML
, , which is
related to linear algebra.

Problems Left Open

1) How to axiomatize the equations E such that C(DA, )/E is a preorder (see 7.9)?
2) Let C(DS, ) be like C(DA, ) with the transformation c added.
How to axiomatize the equations E such that C(DS, )/E is coherent
with respect to Rel (see 7.9)?
3) How to axiomatize equations for mix-bimonoidal categories, symmetric or not symmetric, dissociative or not dissociative, for which one
could prove coherence with respect to Rel (see Chapter 8)?
4) How to axiomatize the equations E such that C(L, )/E is coherent with respect to Rel (see 9.6, and the revised version of [47]in
particular the end)?
5) For K being ZIL or ZIL, , how to axiomatize the equations E such
that C(K)/E is coherent with respect to Rel (see 12.5)?
6) Can one prove coherence with respect to Rel for ML, ML, , ZIML
and ZIML, (see 10.2-3 and 12.5)? If not, what extended axiomatization delivers coherence?
7) For K being ML, ML, , DL or DL, , prove that one could properly extend E(K) without falling into preorder (see 10.3 and 11.5).
Classify the equations that give such extensions.

8) Can one derive the equations (m e) and (m e) from the remaining

axioms of E(DL) (see 11.1)?
9) Find concrete examples, distinct from DL, DL, and B, of distributive lattice, distributive dicartesian and Boolean categories in which
and are not isomorphic (see Chapters 11 and 14).
10) Consider the maximality question for the category B (see the end of
14.2).

331

List of Equations
We list here the equations assumed as axioms for the logical categories
in our book. Besides that, we list prominent equations and denitions
that were used in derivations or alternative axiomatizations. We mention
in parentheses the sections where the equations were rst introduced. A
number of equations for were not stated explicitly in the main text, but
appear here for the rst time. We put the equations for immediately
below the dual equations for . Otherwise, the list follows the order in
which the equations appear in the book.

Categorial equations:
(cat 1)

f 1a = 1b f = f : a b

(2.2)

(cat 2)

h (g f ) = (h g) f

(2.2)

Bifunctorial equations:
(1)

1A 1B = 1AB

(2.7)

(1)

1A 1B = 1AB

(2.7)

(2)

(g1 f1 ) (g2 f2 ) = (g1 g2 ) (f1 f2 )

(2.7)

(2)

(g1 f1 ) (g2 f2 ) = (g1 g2 ) (f1 f2 )

(2.7)

Naturality equations (for f : A D, g : B E and h : C F ):

(b nat)

(b nat)

(b nat)

(b nat)

( nat)

( nat)

( nat)

( nat)

((f g) h) b
A,B,C = b D,E,F (f (g h))

((f g) h) b
A,B,C = b D,E,F (f (g h))

(f (g h)) b
A,B,C = b D,E,F ((f g) h)

(f (g h)) b
A,B,C = b D,E,F ((f g) h)

f
A = D (f 1 )

f
A = D (f 1 )

(f 1 )
A = D f

(f 1 )
A = D f

332

(2.7)
(2.7)
(2.7)
(2.7)
(2.7)
(2.7)
(2.7)
(2.7)

List of Equations

333

f
A = D (1 f )

( nat)

( nat)

f A = D

(2.7)

(1 f )

(1 f ) A = D

(2.7)

(1 f )
A = D f

(2.7)

( nat)
( nat)

(2.7)

(g f ) c A,B = c D,E (f g)

( c nat)

(2.7)

(g f ) c B,A = c E,D (f g)

( c nat)

(f f ) wA = wD f

f wA = wD (f f )

(w nat)

(w nat)

(2.7)

(2.7)

(2.7)

f k 1A,B = k 1D,E (f g)

(k 1 nat)

(2.7)

(g f ) k 1B,A = k 1E,D g

(k 1 nat)

(2.7)

g k 2A,B = k 2D,E (f g)

(k 2 nat)

(2.7)

(g f ) k 2B,A = k 2E,D f

(k 2 nat)

( nat)

(2.7)

1 A = D f

(2.7)

( nat)

f A = D 1

(2.7)

(dL nat)

((f g) h) dL
A,B,C = dD,E,F (f (g h))

(2.7)

(dR nat)

(h (g f )) dR
C,B,A = dF,E,D ((h g) f )

(2.7)

(m nat)

(f g) mA,B = mD,E (f g)

(2.7)

(m

nat)

(f

g) m1
A,B

m1
D,E

(f g)

Specic and other equations:

(b 5)

(b 5)

(b b)

(b b)

( )

( )

b
AB,C,D

b
AB,C,D

b
A,B,C

b
A,B,CD = ( b A,B,C 1D ) b A,BC,D (1A b B,C,D )

b
A,B,C = 1A(BC) ,

b
A,B,C = 1A(BC) ,

b
A,B,CD = ( b A,B,C 1D ) b A,BC,D (1A b B,C,D )
(4.2)

b
A,B,C

A = 1A ,

A = 1A ,

b
A,B,C

b
A,B,C

A = 1A

A = 1A

b
A,B,C = 1(AB)C (4.3)

b
A,B,C = 1(AB)C
(4.6)

334

List of Equations

= 1

A ,
A
A

= 1

A
A
A

( )
( )

(b )

(b )

A = 1A ,

b
A,,C = ( A 1C ) (1A C )

(b )

(b )

( )

( )

( c c)

( c c)

b
A,B, = AB (1A B )

b
,B,C = ( B 1C ) BC

(4.6)

b
,B,C = ( B 1C ) BC

(4.6)

c B,A

c A,B

c A,B = 1AB

(5.1)

c B,A = 1AB

c BC,A = b
B,C,A (1B c C,A ) b B,A,C ( c B,A 1C ) b A,B,C

( c )

( c )

( c )

( c 1)

( c 1)

( c )

(4.6)

c A,BC = b
B,C,A (1B c A,C ) b B,A,C ( c A,B 1C ) b A,B,C
(5.1)

(b c)

b
A,B, = AB (1A B )

(b c)

(4.6)

(b )

A = 1A

b
A,,C = ( A 1C ) (1A C )

(b )

(4.6)

c A, =
A

c ,A =
A

(5.3)

c C,C = 1CC ,

c C,C = 1CC ,

for letterless C

(6.4)

for letterless C

(6.4)

c A,A = 1AA

(6.5)

c A,A = 1AA

(6.5)

(dL )

dL
AB,C,D = ( b A,B,C 1D ) dA,BC,D (1A dB,C,D ) b A,B,CD
(7.2)

(dL )

dL
D,C,BA = b DC,B,A (dD,C,B 1A ) dD,CB,A (1D b C,B,A )
(7.2)

List of Equations

335

(dR )

dR
D,C,BA = (1D b C,B,A ) dD,CB,A (dD,C,B 1A ) b DC,B,A
(7.2)

(dR )

dR
AB,C,D = b A,B,CD (1A dB,C,D ) dA,BC,D ( b A,B,C 1D )
(7.2)

(d b)

(d b)
(dR c)

L
L
R

dR
AB,C,D (dA,B,C 1D ) = dA,B,CD (1A dB,C,D ) b A,BC,D
(7.2)

R
(dR
A,B,C 1D ) dAB,C,D = b A,BC,D (1A dB,C,D ) dA,B,CD
(7.2)

R
L
dC,B,A = c C,BA ( c A,B 1C ) dA,B,C (1A c B,C ) c CB,A
(7.6)

e A,B,C,D =df dA,D,BC (1A c D,BC ) (1A dB,C,D ) b

A,B,CD :
(A B) (C D) (A D) (B C) (7.6)
e A,B,C,D =df e A,B,D,C (1AB c D,C ) :
(A B) (C D) (A C) (B D)

( e)

c BC,AD

e A,B,C,D = e B,A,C,D ( c A,B 1CD )

(7.6)
(7.6)

e D,C,B,A =df b
DC,B,A (dD,C,B 1A ) ( c CB,D 1A ) dCB,D,A:
(C B) (D A) (D C) (B A) (7.6)

e D,C,B,A =df ( c C,D 1BA ) e C,D,B,A :

(D B) (C A) (D C) (B A)

(1DC c B,A ) e D,C,A,B = e D,C,B,A

( e)

( dL )

( dL )

( dR )

( d )

c DA,CB

dL
,B,C = ( B 1C ) BC

dL
A,B, = AB (1A B )

dR
C,B, = (1C B ) CB

dR
,B,A

= BA ( B 1A )

f 3 g =df (f g) mA,B ,
(3)

for f : A D and g : B E

(7.6)
(7.6)
(7.9)
(7.9)
(7.9)
(7.9)
(8.1)

(g1 3 g2 ) (f1 f2 ) = (g1 g2 ) (f1 3 f2 ) = (g1 f1 ) 3 (g2 f2 )

(8.1)
mA,B =df 1A 3 1B

(8.1)

336
(bm)

List of Equations

(mA,B 1C ) mAB,C b
A,B,C = b A,B,C mA,BC (1A mB,C )
(8.2)

((f 3 g) 3 h) b
A,B,C = b D,E,F (f 3 (g 3 h))

(8.2)

(b mL) mAB,C b
A,B,C = dA,B,C (1A mB,C )

(8.3)

L
(b mL) b
A,B,C mA,BC = (mA,B 1C ) dA,B,C

(8.3)

(b mR) mC,BA b
C,B,A = dC,B,A (mC,B 1A )

(8.3)

R
(b mR) b
C,B,A mCB,A = (1C mB,A ) dC,B,A

(cm)

(8.3)

mB,A c A,B = c B,A mA,B

(8.4)

(g 3 f ) c A,B = c E,D (f 3 g)

(b w)

(b w)

( c w)

( c w)

(8.4)

b
A,A,A (1A w A ) w A = (w A 1A ) w A

(9.1)

wA (1A wA ) b
A,A,A = w A (w A 1A )

c A,A wA = wA

wA

(9.1)

c A,A = wA

cm
A,B,C,D =df b A,C,BD (1A (b C,B,D ( c B,C 1D ) b B,C,D )) b A,B,CD :
(A B) (C D) (A C) (B D) (9.1)

(b c w)

wAB = c m
A,A,B,B (w A w B )

(9.1)

cm
A,B,C,D =df b A,B,CD (1A ( b B,C,D ( c B,C 1D ) b C,B,D )) b A,C,BD :
(A C) (B D) (A B) (C D)

(b c w)

(b k)

(b k)

( c k)

( c k)

(wk)

(wk)

wAB = (wA wB ) c m
A,A,B,B

2
(k 1A,B 1C ) b
A,B,C = 1A k B,C

(9.1)

1
2

b
A,B,C (k A,B 1C ) = 1A k B,C

k 2A,B = k 1B,A

k 2A,B = c A,B

c A,B

k 1B,A

k iA,A wA = 1A ,

(9.1)

wA k iA,A = 1A ,

for i {1, 2}
for i {1, 2}

(9.1)

List of Equations

(b k 1 )

(b k 1 )

(b k 2 )

(b k 2 )

(wk k)

(wk k)

337

k 1AB,C = (1A k 1B,C ) b

A,B,C

(9.1)

k 1AB,C = b
A,B,C (1A k B,C )

k 2C,BA = (k 2C,B 1A ) b
C,B,A

(9.1)

k 2C,BA = b
C,B,A (k C,B 1A )

(k 1A,B k 2A,B ) wAB = 1AB

(9.1)

wAB (k 1A,B k 2A,B ) = 1AB

( , ) f1 , f2 =df (f1 f2 ) wC , for f1 : C A1 and f2 : C A2 (9.1)

[g1 , g2 ] =df wC (g1 g2 ), for g1 : A1 C and g2 : A2 C (9.4)

()
()

(b )

(b )

(b )

f g = f k 1A,B , g k 2A,B , for f : A D and g : B E

f g = [k 1D,E f, k 2D,E g], for f : A D and g : B E

1
2
b
A,B,C = 1A k B,C , k B,C

1
b
C,B,A = k C,B

c A,B = [k 2A,B , k 1A,B ]

wA = 1A , 1A

(9.1)

(9.1)

()

k iA1 ,A2 f1 , f2 = fi ,

()

(9.1)

wA = [1A , 1A ]

()

k 2B,C ]

1
1
2

b
C,B,A = [k CB,A k C,B , k C,B 1A ]

(w)

()

c A,B = k 2A,B , k 1A,B

(w)

(9.1)

k 1CB,A , k 2C,B 1A

( c)

k 2A,BC

1
2
b
A,B,C = [1A k B,C , k A,BC

( c)

(b )

(9.1)

[g1 , g2 ] k iA1 ,A2 = gi ,

for fi : C Ai and i {1, 2}

for gi : Ai C and i {1, 2}

k 1A1 ,A2 h, k 2A1 ,A2 h = h,

[h k 1A1 ,A2 , h k 2A1 ,A2 ] = h,

K 1A2 g1 =df g1 k 1A1 ,A2 ,

K 1A2 f1 =df k 1A1 ,A2 f1 ,

(9.1)

for h : C A1 A2

(9.1)

for h : A1 A2 C

for g1 : A1 C

(9.1)

for f1 : C A1

(9.4)

338

List of Equations

K 2A1 g2 =df g2 k 2A1 ,A2 ,

K 2A1 f2 =df k 2A1 ,A2 f2 ,

for g2 : A2 C

(9.1)

for f2 : C A2

(9.4)

for i {1, 2}, and f , g, fi and gi of appropriate types,

(K 1)

(K 1)

(K 2)

(K 2)

(K 3)

(K 3)

(K 4)

(K 4)

(K 5)

(K 5)

g K iA f = K iA (g f )

K iA g f = K iA (g f )

K iA g f1 , f2 = g fi

(9.1)
(9.4)
(9.1)

[g1 , g2 ] K iA f = gi f

(9.4)

g1 , g2 f = g1 f, g2 f

(9.1)

g [f1 , f2 ] = [g f1 , g f2 ]

(9.4)

1AB = K 1B 1A , K 2A 1B
1AB = [K 1B 1A , K 2A 1B ]

K iA f, g = K iA f, K iA g

(9.1)
(9.4)
(9.1)

K iA [f, g] = [K iA f, K iA g]

k 1A1 ,A2 =df K 1A2 1A1

(9.1)

k 1A1 ,A2 =df K 1A2 1A1

k 2A1 ,A2 =df K 2A1 1A2

(9.1)

k 2A1 ,A2 =df K 2A1 1A2

(k )

(k )

(w )

(w )

k 1A, =
A

(9.2)

k 1A, =
A

w =

(9.2)

w =

A =df k 2A,

k2
A =df
A
A,

( 1)

( 1)

= 1

= 1

(9.2)
(9.6)
(9.2)
(9.6)

List of Equations

()

()

339

for f : A

(9.2)

for f : A

(9.6)

A = f ,
A = f ,

A = 1A , A

(9.2)

A = A , 1A

(9.2)

(1
k 1A,B =
A B )
A

(9.2)

(
k 2A,B =
A 1B )
B

(k k)

(k k)

(9.2)

k 1p,p = k 2p,p

(9.3)

k 1p,p = k 2p,p

(9.5)

(in-out) [f, g], [h, j] = [f, h, g, j]

(9.4)

ckA,B,C,D =df k 1A,B k 1C,D , k 2A,B k 2C,D :

(A B) (C D) (A C) (B D)

ckA,B,C,D = [k 1A,C k 1B,D , k 2A,C k 2B,D ]

(9.4)

wAB = ckA,A,B,B (wA wB )

(9.4)

wAB = (wA wB ) ckA,B,A,B

(9.4)

1
1
2
2
cm
A,B,C,D = k A,B k C,D , k A,B k C,D

(9.4)

1
1
2
2
cm
D,C,B,A = [k D,C k B,A , k D,C k B,A ]

()

()

(k )

(k )

(K)

(K)

(k k)

(k )

(k )

(9.4)

c , = 1

(9.6)

c , = 1

(9.6)

k 1, = k 2,

(9.6)

k 1, = k 2,

(9.6)

K 1 1 = K 2 1

(9.6)

K 1 1 = K 2 1

(9.6)

k 1p, k 1p, = k 2p, 0, k 2p, ,

k 1p, = p k 2p,

(9.7)
(9.7)

k 1p, = k 2p, p

for 0, = =

(9.4)

(9.7)

(k k f g) k 1b, f k 1a, = k 1b, g k 1a, ,

for f, g : a b

(9.7)

340

List of Equations

(wm)

wA mA,A wA = 1A

(10.1)

f g =df wB (f 3 g) wA ,

for f, g : A B

mA,B =df K 1B K 1B 1A K 2A K 2A 1B
( )

(f g) h = (f h) (g h),

(10.1)
(10.1)

h (f g) = (h f ) (h g) (10.1)

( assoc)

f (g h) = (f g) h

(10.1)

( com)

f g =gf

(10.1)

( idemp)

f f =f

(10.1)

()

(f1 f2 ) (g1 g2 ) = (f1 g1 ) (f2 g2 )

(10.1)

()

(f1 f2 ) (g1 g2 ) = (f1 g1 ) (f2 g2 )

(10.1)

mAC,BD c A,B,C,D = mA,B mC,D

(10.1)

k
cm
A,B,C,D mAC,BD cA,B,C,D = mA,B mC,D

(10.1)

ckA,C,B,D

(m)
(m)

mA, = k 1A, k 1A,

mA, = k 1A, k 1A,
mA,C = k 1A,C

k 1A,C ,

mC,A = k 2C,A k 2C,A ,

(0)

(10.3)

(10.3)
for letterless C

(10.3)

for letterless C

(10.3)

1A K 2A A = 1A

(10.3)

(0)

1A K 2A A = 1A

(10.3)

(0g)

f g = f,

(10.3)

(d k)

(d k)
(dm)

for a null term g

k 2A,BC = (k 2A,B 1C ) dA,B,C

(11.1)

k 1CB,A = dC,B,A (1C k 1B,A )

(11.1)

mA,C = (k 1A,B 1C ) dA,B,C (1A k 2B,C )

(11.1)

mA,C = (1A k 2B,C ) dR

A,B,C (k A,B 1C )

(11.1)

ckA,C,B,D e A,B,C,D = mA,B 1CD

(11.1)

e D,C,B,A ckD,C,B,A = 1DC mB,A

(11.1)

(m e)
(m e)

List of Equations

341

(k 1A,C k 1B,D ) e A,B,C,D = mA,B k 1AB,CD

e D,C,B,A (k 2D,B k 2C,A ) = k 2DC,BA

mB,A

(11.1)
(11.1)

(wm e) (wA 1CD ) ckA,C,A,D e A,A,C,D (wA 1CD ) = 1A(CD)

(11.1)
(wm e) (1DC wA ) e D,C,A,A ckD,C,A,A (1DC wA ) = 1(DC)A
(11.1)

(m c m ) mAC,BD c m
A,B,C,D = e A,B,C,D (1AB mC,D )

(11.1)

(m c m ) c m
D,C,B,A mDB,CA = (mD,C 1BA ) e D,C,B,A

(11.1)

s A,C,D =df e A,A,C,D (wA 1CD ) : A (C D) (A C) (A D)

(11.3)

s D,C,A =df (1DC wA ) e D,C,A,A : (D A) (C A) (D C) A

(11.3)

t A,C,D =df (wA 1AD ) ckA,C,A,D : (A C) (A D) A (C D)

(11.3)

t D,C,A =df ckD,C,A,A (1DC wA ) : (D C) A (D A) (C A)

(11.3)

t A,C,D

s D,C,A

s A,C,D = 1A(CD)

(11.3)

t D,C,A = 1(DC)A

(11.3)

dA,B,C = (1AB k 2A,C ) e A,A,B,C (wA 1BC )

dC,B,A = (1CB wA ) e C,B,A,A (k 1C,A 1BA )

(d)
(d)

dA,, = k 1A, (1A )

d,,C = ( 1C ) k 2,C

(11.3)

1 2
1
1

(m1 0) k 2A,B m1
A,B k A,B = k B,A mB,A k B,A
1 2
1
2

(m1 1) k 1A,B m1
A,B k A,B = k B,A mB,A k B,A = 1A

(11.3)
(11.3)

(11.3)

(12.1)
(12.1)

1
1
1

(bm1 ) (m1
A,B 1C ) mAB,C b A,B,C = b A,B,C mA,BC (1A mB,C )
(12.1)
1

(cm1 ) m1
B,A c B,A = c A,B mA,B

(12.1)

342

List of Equations

1 2
1
1

0A,B =df k 2A,B m1

A,B k A,B = k B,A mB,A k B,A

(12.1)

m1
A,B =df [1A , 0B,A ], [0A,B , 1B ] = [1A , 0A,B , 0B,A , 1B ] (12.1)
for f : A B

f 0A,A = 0B,B f = 0A,B ,

(0)

f 0C,A = 0C,B ,

for f : A B

0B,C f = 0A,C ,

(0)

0A,C 0B,D = 0AB,CD

(0)

0A,C 0B,D = 0AB,CD

(12.1)
(12.1)
(12.1)

(12.1)

1
1

(d1 1) k 1AB,C d1
A,B,C k A,BC = k A,B

(12.1)

2
2

(d1 2) k 2AB,C d1
A,B,C k A,BC = k B,C
2
2

(d1 3) k 1AB,C d1
A,B,C k A,BC = k A,B

A = f = 0A, ,

A = f = 0,A ,

(12.1)

for f : A

(12.1)
(12.1)

= 1 = 0,

(12.1)

= 1 = 0,

(12.1)

k 1, = k 2, = 0,
k 1, = k 2, = 0,

(0) 0A,B = B 0, A

(00)

(12.1)
(12.1)

k 1B,C

for f : A

= = 0,

(0I)

=df 1A k 1B,C , [0A,C , k 2B,C ] = [k 1A,B , 0A,C , k 2A,B 1C ] :

A (B C) (A B) C (12.1)

0A,C =df k 2AB,C d1

A,B,C k A,BC

(12.1)

for f : A B

(0 nat) f 0A,A = 0B,B f ,

d1
A,B,C

(12.1)

(12.1)
(12.1)
(12.1)

f 0A = 0B g,

for f, g : A B

(12.5)

f 0A = g 0A ,

for f, g : A B

(12.5)

0B f = 0B g,

for f, g : A B

(12.5)

0A 0A = 0A
0A,B =df f 0A ,

(12.5)
for f : A B

(12.5)

List of Equations

343

(0I)

0A 0B = 0AB

(12.5)

(0I)

0A 0B = 0AB

(12.5)

(0)

f 0A,B = f ,

for f : A B

(12.5)

1A 0A = 1A

(12.5)

(mm1 ) m1
A,B mA,B = 1AB ,

(k m)

(k m)

mA,B m1
A,B = 1AB

k 1A,B = [1A , 0B,A ] mA,B , k 2A,B = [0A,B , 1B ] mA,B

k 1A,B = mA,B 1A , 0A,B , k 2A,B = mA,B 0B,A , 1B

(13.1)
(13.1)
(13.1)

for f : A B,

Z 1C f =df m1
B,C k B,C f = f, 0A,C : A B C

Z 1C f =df f k 1A,C m1
A,C = [f, 0C,B ] : A C B

Z 2C f =df m1
C,B k C,B f = 0A,C , f : A C B

Z 2C f =df f k 2C,A m1
C,A = [0C,B , f ] : C A B

(Z )

(Z )

(13.1)
(13.1)
(13.1)

f1 , f2 = Z 1A2 f1 Z 2A1 f2 , for f1 : C A1 and f2 : C A2 (13.1)

[g1 , g2 ] = Z 1A2 g1 Z 2A1 g2 , for g1 : A1 C and g2 : A2 C (13.1)

1
2

wC = m1
C,C (k C,C k C,C )

(13.1)

wC = (k 1C,C k 2C,C ) m1
C,C

(13.1)

(f1 f2 ) (f3 f4 ) = (f1 f3 ) (f2 f4 )

(13.1)

(md)

(13.1)

dA,B,C =df mAB,C b

A,B,C (1A mB,C )

(13.2)

e A,B,C,D = mAC,BD c m
A,B,C,D (1AB mC,D )

(13.2)

e D,C,B,A = (m1
D,C 1BA ) c D,C,B,A mDB,CA

(13.2)

clA,B,C,D =df e A,B,C,D (m1

A,B 1CD ) :
(A B) (C D) (A C) (B D)

(13.2)

clD,B,C,A =df (1DC m1

B,A ) e D,C,B,A

(13.2)

f g = [k 1B,A , k 2A,B ] clB,A,A,B k 1B,A f, k 2A,B g

0A,B = [k 1B,A , k 2A,B ] clB,A,A,B k 2B,A , k 1A,B

(13.2)
(13.2)

344

List of Equations
ckA,C,B,D clA,B,C,D = 1(AB)(CD)

(13.2)

clA,C,B,D ckA,B,C,D = 1(AB)(CD)

(13.2)

n
A

(14.1)

n
A

= 1A

n
A nA = 1 A

rA,B

(14.1)

rA,B = 1(AB)

(14.1)

rA,B

rA,B = 1AB

(14.1)

r
A,B

r
A,B = 1(AB)

(14.1)

rA,B

rA,B = 1AB

= 1

= 1

= 1

(14.1)

(14.1)

(14.1)

(14.1)

= 1

(14.1)

(A 1A ) dA,A,A (1A A )
A

(0)

0A = A

()

BA = (r
B,A 1BA ) c BA,BA e B,A,B,A

()

(14.2)

( c B,B c A,A ) (B A )

BA = (r
B,A 1BA ) e B,A,B,A (B A )

()

AB =
(A B ) e A,B,A,B (1AB r A,B )

()

AB =
(A B ) ( c A,A c B,B ) e A,B,A,B

c
AB,AB (1AB r A,B )

()
()
()
()

= ( 1 )

(14.2)

(14.2)
(14.2)

= ( 1 )

(14.2)

(14.2)

(14.2)

(1 )

(14.2)

=
(1 )

(14.2)

(1A f ) A = (f 1B ) B ,

for f : A B

(14.2)

A (1A f ) = B (f 1B ),

for f : A B

(14.2)

List of Categories
We list in the table below the logical categories and some other related
categories we deal with in the book. We present the categories involving
immediately below the dual categories involving . Otherwise, we follow
the ascending order in which the categories appear in the Charts that follow,
which is close to the order in which they appear in the book.

category

I
I

language

families

L
L,

1
1

1, b

1, b

1, b

L,

1, b,

A
A

L,
L,

4.2

(b 5)

4.3
6.1

(b 5), (b b)

4.6

(b 5), (b b),

( ), ( ), (b )

1, b,

4.1

(b 5), (b b)

section
4.1

specic equations

6.1

(b 5), (b b),

( ), ( ), (b )

1, b

(b 5), (b b),

6.1

(b 5), (b b)
A,

L,,,

1, b,

6.1

(b 5), (b b),

(b 5), (b b),

( ), ( ), (b ),

( ), ( ), (b )

1, b, c

1, b, c

1, b, c

L,

1, b, c,

L,

5.1

6.5

6.3

5.3

6.4

(b 5), (b b), ( c c), (b c)

(b 5), (b b), ( c c), (b c),

( c 1)

(b 5), (b b), ( c c), (b c)

(b 5), (b b), ( c c), (b c),

( ), ( ), (b )

1, b, c,

(b 5), (b b), ( c c), (b c),

( ), ( ), (b )
345

346

List of Categories
category

language
L,

families
1, b, c

specic equations

(b 5), (b b), ( c c), (b c),

section
6.3

(b 5), (b b), ( c c), (b c)

S

L,

1, b, c

(b 5), (b b), ( c c), (b c),

6.5

(b 5), (b b), ( c c), (b c),

( c 1), ( c 1)
S,

L,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

6.4

(b 5), (b b), ( c c), (b c),

( ), ( ), (b ),

( ), ( ), (b ),

( c ), ( c )

1, b, c,

9.1

9.4

9.2

9.6

9.4

9.4

(b 5), (b b), ( c c), (b c),

w-k

(b w), ( c w), (b c w),

(b k), ( c k), (wk)

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k

(b w), ( c w), (b c w),

(b k), ( c k), (wk)

L,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, -

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

( ), ( ), (b ),

(k )

L,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, -

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

( ), ( ), (b ),

(k )

L,

1, b, c,

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

w-k

(b k), ( c k), (wk)

L,

1, b, c,

w-k

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b k), ( c k), (wk)

List of Categories
category

language
L,

347
families
1, b, c,

specic equations

(b 5), (b b), ( c c), (b c),

w-k

section
9.4

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk)

L

L,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, -

9.6

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

( ), ( ), (b ),

(k ), ( c )
L

L,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, -

9.6

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

( ), ( ), (b ),

(k ), ( c )
L,

L,,,

1, b, c,
w-k, -

(b 5), (b b), ( c c), (b c),

9.6

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

( ), ( ), (b ),

( ), ( ), (b ),

DI

L,

L,

D A

(k ), (k ), ( c ), ( c )
1, d
1, b, d

(b 5), (b b), (d ),
L

(b 5), (b b), (d )
L

7.1
7.5

348
category
DA

List of Categories
language
L,

families
1, b, d

specic equations

section

(b 5), (b b), (d ), (d ),
L

7.2

(b 5), (b b), (d ), (d ),

(d b), (d b)
DA,

L,,,

1, b, -,
d

(b 5), (b b), (dL ), (dR ),

7.9

(b 5), (b b), (d ), (d ),

(d b), (d b),

( ), ( ), (b ),

( ), ( ), (b ),

( dL ), ( dL ), ( dR ), ( dR )
DS

L,

1, b, c,
d

(b 5), (b b), ( c c), (b c),

7.6

(b 5), (b b), ( c c), (b c),

(dL ), (dL ), (dR ), (dR ),

(d b), (d b), (dR c)

MI
MA
MS

ML

L,

1, m

L,

1, b, m

L,

L,

(b 5), (b b), (b 5), (b b), (bm)

1, b, c,

(b 5), (b b), ( c c), (b c),

(b 5), (b b), ( c c), (b c),

(bm), (cm)

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m

8.1
8.2
8.5

10.1

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm)
ML,

L,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
(m), (m)

10.3

List of Categories

349

category
MDI

language
L,

families
1, d, m

MDA

L,

1, b, d,
m

specic equations

(b 5), (b b), (dL ), (dR ),

section
8.1
8.3

(b 5), (b b), (d ), (d ),
L

(d b), (b mL), (b mR),

(d b), (b mL), (b mR)
MDS

L,

1, b, c,
d, m

(b 5), (b b), ( c c), (b c),

8.4

(b 5), (b b), ( c c), (b c),

(dL ), (dL ), (dR ), (dR ),

(d b), (b mL), (b mR),

(d b), (b mR), (b mL),
(dR c), (cm)
DL

L,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

11.1

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm), (dm),
(dL ), (dL ), (dR ), (dR ),

(d b), (d k), (m e),

(d b), (d k), (m e), (dR c)

DL,

L,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

d, -

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm), (dm),
(dL ), (dL ), (dR ), (dR ),

(d b), (d k),
(d b), (d k), (dR c),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c )

11.5

350
category
ZIL

List of Categories
language
L,

families
1, b, c,

specic equations

(b 5), (b b), ( c c), (b c),

w-k, 0A

section
12.5

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(0I)
ZL

L,

1, b, c,
w-k, m

(b 5), (b b), ( c c), (b c),

1

12.1

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(m1 0), (m1 1)
ZIL,

L,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, -,

(b 5), (b b), ( c c), (b c),

0A

(b w), ( c w), (b c w),

12.5

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(0I)

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
ZL,

L,,,

1, b, c,
w-k, -,
m

(b 5), (b b), ( c c), (b c),

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(m1 0), (m1 1)

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),

12.1

List of Categories
category
ZIML

language
L,

351
families
1, b, c,

specic equations

(b 5), (b b), ( c c), (b c),

(b 5), (b b), ( c c), (b c),

0A

(b w), ( c w), (b c w),

12.5

w-k, m,

section

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),
(0I), (0)
ZIML,

L,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

-, 0A

(b w), ( c w), (b c w),

12.5

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
(m), (m), (0I), (0)
ZIDL

L,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

d, 0A

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm), (dm),
(dL ), (dL ), (dR ), (dR ),

(d b), (d k), (m e),

(d b), (d k), (m e), (dR c),

(0I)

12.5

352
category
ZIDL,

List of Categories
language
L,,,

families
1, b, c,

specic equations

(b 5), (b b), ( c c), (b c),

(b 5), (b b), ( c c), (b c),

d, -,

(b w), ( c w), (b c w),

0A

(b w), ( c w), (b c w),

12.5

w-k, m,

section

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm), (dm),
(dL ), (dL ), (dR ), (dR ),

(d b), (d k), (m e),

(d b), (d k), (m e), (dR c),

(0I),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c )
B

L
,,,

1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,

(b 5), (b b), ( c c), (b c),

d, -,

(b w), ( c w), (b c w),

0A , n-r,

(b w), ( c w), (b c w),

, ,

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm), (dm),
(dL ), (dL ), (dR ), (dR ),

(d b), (d k), (m e),

(d b), (d k), (m e), (dR c),

(0I),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
n-r and isomorphisms

14.2

List of Categories
category
ZML

language
L,

353
families
1, b, c,

(b 5), (b b), ( c c), (b c),

w-k, m,
m

specic equations

section
13.1

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),
(m1 0), (m1 1), (mm1 )
ZML,

L,,,

1, b, c,
w-k, -,
m, m

(b 5), (b b), ( c c), (b c),

13.3

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),
(m1 0), (m1 1), (mm1 ),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
ZML
,

L
,,,

1, b, c,
m, m

n-r,

(b 5), (b b), ( c c), (b c),

w-k, -,
1

(b 5), (b b), ( c c), (b c),

(b w), ( c w), (b c w),

(b w), ( c w), (b c w),

(b k), ( c k), (wk),

(b k), ( c k), (wk),

(bm), (cm), (wm),
(m1 0), (m1 1), (mm1 ),

( ), ( ), (b ),

( ), ( ), (b ),

(k ), (k ), ( c ), ( c ),
n-r and isomorphisms

14.1

Charts
The charts we present on the next three pages are to be read as follows.
When the category A is joined by an upward-going line to the category B,
this means that A is isomorphic to a subcategory of B, with the isomorphism being identity on objects. The assertion that A is isomorphic to a
subcategory of B is established by appealing to the fact that A is a preorder or that A is coherent with respect to Rel, as explained in 14.4. The
three charts could be combined into a single chart by pasting them together
in growing order over the parts in which they overlap. For practical, and
aesthetical, reasons we have preferred not to make this pasting, and have
three separate charts.
We have established coherence with respect to Rel for all categories in

the charts except S , S , DA, , L, , ML, ML, , ZIL, ZIL, , ZIML

and ZIML, . For S and S we have that they are preorders, though they
are not coherent with respect to Rel (see 6.5). The category DA, was
considered in 7.9. For L, , ML, ML, , ZIML and ZIML, we have
proved only a restricted form of coherence (see 9.6, 10.2-3 and 12.5).
This explains the absence of some lines in Charts 2 and 3.
Of the categories with negation, we have mentioned only two, B and
ZML
, , at the top of Chart 3, which is also the top of all the charts pasted
together. We have, however, coherence for categories with negation where
we have coherence without negation (see 14.1), and there are replicas of
Charts 2 and 3 involving such categories.

354

Charts

355

L,
r
HH L


Lr

Hr
(
 (((((
Q

Q

Lr(
(
(

Q

H
H

Q

HH

Q


H

Q

H

L
L

Q

r 
HrH
Q

S
S
,
H

Q
rH
r

r L

r 


H

H


H





HHL  

S  
HHr
H


r
r
L
HH
 
L
 H
HH
HH




H

HH


HH


 
HH
H

S 


A, r H
r
r
H

r S



H

H
HH

H






S 
A 
HH HHr
r 
r

H

S


H
H

H


S
HH
HH
HH
 
HH


H

HH

HH


r 
H


r A



H




H A


r 
r

Hr
A
H
A
H

 I H
HH


HH

r

A
HH

H

HHr
r 


Chart 1

356

Charts

rDL,
ML,
r
@
@

@
@
@
@

rDL











@


MDS r

ML rPP
PPP
@

PP
P

@

PP
PP
 
PP
PP @

PP
P

P@
r
Pr
P


 DS


L








MDA r
r
MS P
PP 
PP
DA,
 PPP
PP

r
PP
PP



PP
PP

PPr

PP
r

 DA
S







MDI r
r 
r

MA P
PPP
L

PP

 PD A

PP

PP

PP

PP
PP 
PP
PPr
PrDI




A








r


MI PP
PP


PP

PP

PP
Pr
I
L,

Chart 2

Charts

357

r
 ZML,






r 
B

rZML,
 








r 



ZIDL,











r 
 rZML
DL,
 












r 


ZIDL













ZL, r


r 
 DL



rZIML,


,


rML,
,




,
rZIML
,





, 

rML


,
ZL r


L, r
, 




, 



, rZIL, 

,




,  



 


,,

rZIL



,




,


,



r
L

Chart 3

Bibliography
[... ] The sections where a reference is mentioned are listed at the end in
square brackets.
[1] A.R. Anderson and N.D. Belnap, Entailment: The Logic of
Relevance and Necessity, Vol. I , Princeton University Press,
Princeton, 1975 [7.1, 14.1, 14.3-4]
[2] J.C. Baez and J. Dolan, Categorication, Higher Category Theory (E. Getzler et al., editors), Contemporary Mathematics, vol. 230,
American Mathematical Society, Providence, 1998, pp. 1-36 [1.2]
nzl and R. Vogt, It[3] C. Balteanu, Z. Fiedorowicz, R. Schwa
erated monoidal categories, Advances in Mathematics, vol. 176
(2003), pp. 277-349 [9.4]
[4] H.P. Barendregt, The Lambda Calculus: Its Syntax and Semantics, North-Holland, Amsterdam, 1981 [4.2, 9.3]
[5] G. Bellin, Two paradigms of logical computation in ane logic?,
Logic for Concurrency and Synchronization (R.J.G.B. de
Queiroz, editor), Kluwer, Dordrecht, 2003, pp. 111-144 [1.6]
[6] N.D. Belnap, A useful four-valued logic, Modern Uses of Multiple-Valued Logic (J.M. Dunn and G. Epstein, editors), Reidel,
Dordrecht, 1977, pp. 8-37 [14.1]
nabou, Categories avec multiplication, Comptes Rendus de
[7] J. Be
lAcad
emie des Sciences, Paris, Serie I, Mathematique, vol. 256
(1963), pp. 1887-1890 [2.8, 4.6]
[8] , Structures algebriques dans les categories, Cahiers de
Topologie et G
eom
etrie Di
erentielle, vol. 10 (1968), pp. 1-126
[4.2, 7.3]
359

360

Bibliography

[9] G. Birkhoff, Lattice Theory , third edition, American Mathematical Society, Providence, 1967 (rst edition, 1940) [7.1, 12.1]
[10] R. Blute, Linear logic, coherence and dinaturality, Theoretical
Computer Science, vol. 115 (1993), pp. 3-41 [8.1]
[11] R. Blute, J.R.B. Cockett, R.A.G. Seely and T.H. Trimble,
Natural deduction and coherence for weakly distributive categories,
Journal of Pure and Applied Algebra, vol. 113 (1996), pp. 229296 [7.9]
, A cut-elimination proof in intuitionistic predi[12] M. Borisavljevic
cate logic, Annals of Pure and Applied Logic, vol. 99 (1999), pp.
105-136 [Preface, 11.1]
[13] , Two measures for proving Gentzens Hauptsatz without mix,
Archive for Mathematical Logic, vol. 42 (2003), pp. 371-387
[11.1]
, K. Do
, On permuting cut
sen and Z. Petric
[14] M. Borisavljevic
with contraction, Mathematical Structures in Computer Science, vol. 10 (2000), pp. 99-136 [Preface, 11.1-2]
[15] R. Brauer, On algebras which are connected with the semisimple
continuous groups, Annals of Mathematics, vol. 38 (1937), pp.
857-872 [1.4]
nnler and A.F. Tiu, A local system for classical logic, Logic
[16] K. Bru
for Programming, Articial Intelligence and Reasoning , Lecture Notes in Computer Science, vol. 2250, Springer, Berlin, 2001, pp.
347-361 [7.1, 11.1]
[17] W. Burnside, Theory of Groups of Finite Order , second edition, Cambridge University Press, Cambridge, 1911 (reprint, Dover,
New York, 1955) [5.2]
[18] S.R. Buss, The undecidability of k-provability, Annals of Pure and
Applied Logic, vol. 53 (1991), pp. 75-102 [1.4]
[19] A. Carbone, Interpolants, cut elimination and ow graphs for the
propositional calculus, Annals of Pure and Applied Logic, vol.
83 (1997), pp. 249-299 [1.4]
[20] J.R.B. Cockett, Introduction to distributive categories, Mathematical Structures in Computer Science, vol. 3 (1993), pp. 277307 [1.2, 11.1]

Bibliography

361

[21] J.R.B. Cockett and R.A.G. Seely, Weakly distributive categories,

Applications of Categories in Computer Science (M.P. Fourman et al., editors), Cambridge University Press, Cambridge, 1992,
pp. 45-65 [7.1-2]
[22] , Weakly distributive categories, Journal of Pure and Applied Algebra, vol. 114 (1997), pp. 133-173 (updated version available in 2004 at: http://www.math.mcgill.ca/rags) [Preface, 1.2,
Chapter 7, 7.1-2, 7.9, 11.1, 11.3, 12.4, 13.2]
[23] , Proof theory for full intuitionistic linear logic, bilinear logic
and mix categories, Theory and Applications of Categories, vol.
3 (1997), pp. 85-131 [7.9, 8.1, 8.5, 11.1]
[24] , Finite sum-product logic, Theory and Applications of Categories, vol. 8 (2001), pp. 63-99 [9.6]
[25] J.R.B. Cockett, J. Koslowski and R.A.G. Seely, Introduction
to linear bicategories, Mathematical Structures in Computer
Science, vol. 10 (2000), pp. 165-203 [1.2, 7.2, 7.9]
[26] H.S.M. Coxeter and W.O.J. Moser, Generators and Relations for Discrete Groups, Springer, Berlin, 1957 [5.1]
[27] L. Crane and D.N. Yetter, Examples of categorication, Cahiers
de Topologie et G
eom
etrie Di
erentielle Cat
egoriques, vol. 39
(1998), pp. 3-25 [1.2]

, Interpolation property for bicartesian closed categories,

[28] Dj. Cubri
c
Archive for Mathematical Logic, vol. 33 (1994), pp. 291-319
[Preface]
[29] , On the semantics of the universal quantier, Annals of Pure
and Applied Logic, vol. 87 (1997), pp. 209-239 [Preface]
[30] , Embedding of a free cartesian closed category into the category
of sets, Journal of Pure and Applied Algebra, vol. 126 (1998),
pp. 121-147 [Preface, 9.3]
[31] H.B. Curry, A note on the reduction of Gentzens calculus LJ*,
Bulletin of the American Mathematical Society , vol. 45 (1939),
pp. 288-293 [1.3]
[32] , Foundations of Mathematical Logic, McGraw-Hill, New
York, 1963 (second edition, Dover, New York, 1977) [1.3]

362

Bibliography

[33] V. Danos and L. Regnier, The structure of multiplicatives,

Archive for Mathematical Logic, vol. 28 (1989), pp. 181-203 [7.2,
7.8, 8.1]
[34] P. Dehornoy, The structure group for the associativity identity,
Journal of Pure and Applied Algebra, vol. 111 (1996), pp. 59-82
[4.2]
[35] N. Dershowitz and Z. Manna, Proving termination with multiset
orderings, Communications of the Association for Computing
Machinery , vol. 22 (1979), pp. 465-476 [11.2]
[36] K. Do
sen, Modal translations in substructural logics, Journal of
Philosophical Logic, vol. 21 (1992), pp. 283-336 [11.2]
[37] , Deductive completeness, The Bulletin of Symbolic Logic,
vol. 2 (1996), pp. 243-283, 523 (for corrections see [38], Section 5.1.7,
and [39]) [1.3]
[38] , Cut Elimination in Categories, Kluwer, Dordrecht, 1999
[Preface, 1.3-5, 9.3]
[39] , Abstraction and application in adjunction, Proceedings of
the Tenth Congress of Yugoslav Mathematicians (Z. Kadelburg, editor), Faculty of Mathematics, University of Belgrade, Belgrade, 2001, pp. 33-46 (available at: http://arXiv.org/math.CT/
0111061) [1.3]
[40] , Identity of proofs based on normalization and generality, The
Bulletin of Symbolic Logic, vol. 9 (2003), pp. 477-503 (version with corrected remark on difunctionality available at: http://
arXiv.org/math.LO/0208094) [Preface, Chapter 1, 1.3, 14.3]
[41] , Models of deduction, Proof-Theoretic Semantics (R. Kahle
and P. Schroeder-Heister, editors), Proceedings of the Workshop
Proof-Theoretic Semantics, T
ubingen 1999, Synthese, vol. 148
(2006), pp. 639-657 [1.3]
[42] , Simplicial endomorphisms, Communications in Algebra
vol. 36 (2008), pp. 2681-2709 (available at: http://arXiv.org/math.
GT/0301302) [1.4]
, Modal functional completeness, Proof
[43] K. Do
sen and Z. Petric
Theory of Modal Logic (H. Wansing, editor), Kluwer, Dordrecht,
1996, pp. 167-211 [1.3, 9.1, 11.1]

Bibliography

363

[44] , Cartesian isomorphisms are symmetric monoidal: A justication of linear logic, The Journal of Symbolic Logic, vol. 64
(1999), pp. 227-242 [9.1, 11.2, 11.5]
[45] , The maximality of the typed lambda calculus and of cartesian closed categories, Publications de lInstitut Math
ematique
(N.S.), vol. 68(82) (2000), pp. 1-19 (available at: http://arXiv.org/
math.CT/9911073) [1.5, 9.3]
[46] , The maximality of cartesian categories, Mathematical Logic
Quarterly , vol. 47 (2001), pp. 137-144 (available at: http://arXiv.
org/math.CT/911059) [1.4-5, Chapter 9, 9.2-3]
[47] , Coherent bicartesian and sesquicartesian categories, Proof
Theory in Computer Science (R. Kahle et al., editors), Lecture
Notes in Computer Science, vol. 2183, Springer, Berlin, 2001, pp.
78-92 (revised version, with major corrections, available at: http://
arXiv.org/math.CT/0006091) [1.4, Chapter 9, 9.6]
[48] , Bicartesian coherence, Studia Logica, vol. 71 (2002), pp.
331-353 (version with some corrections in the proof of maximality
available at: http://arXiv.org/math.CT/0006052) [1.4-5, Chapter
9, 9.4-6, 11.1]
[49] , Self-adjunctions and matrices, Journal of Pure and Applied Algebra, vol. 184 (2003), pp. 7-39 (unabridged version available at: http://arXiv.org/math.GT/0111058) [1.4]
[50] , Generality of proofs and its Brauerian representation, The
Journal of Symbolic Logic, vol. 68 (2003), pp. 740-750 (available
at: http://arXiv.org/math.LO/0211090) [1.4, 14.3]
[51] , A Brauerian representation of split preorders, Mathematical
Logic Quarterly , vol. 49 (2003), pp. 579-586 (available at: http://
arXiv.org/math.LO/0211277) [1.4, 14.3]
[52] J.M. Dunn and G.M. Hardegree, Algebraic Methods in Philosophical Logic, Oxford University Press, Oxford, 2001 [7.1]
[53] B. Eckmann and P.J. Hilton, Group-like structures in general categories I: Multiplications and comultiplications, Mathematische Annalen, vol. 145 (1962), pp. 227-255 [9.1]

364

Bibliography

[54] S. Eilenberg and G.M. Kelly, A generalization of the functorial

calculus, Journal of Algebra, vol. 3 (1966), pp. 366-375 [Preface,
1.4]
[55] , Closed categories, Proceedings of the Conference on Categorical Algebra, La Jolla 1965 (S. Eilenberg et al., editors),
Springer, Berlin, 1966, pp. 421-562 [9.1]
, The mix rule, Mathematical Struc[56] A. Fleury and C. Retore
tures in Computer Science, vol. 4 (1994), pp. 273-285 [8.1]
[57] P.J. Freyd, Abelian Categories: An Introduction to the Theory of Functors, Harper & Row, New York, 1964 [13.3]
[58] , Aspects of topoi, Bulletin of the Australian Mathematical Society , vol. 7 (1972), pp. 1-76 [14.3]
[59] P.J. Freyd and A. Scedrov, Categories, Allegories, NorthHolland, Amsterdam, 1990 [13.3]
ber das logische Schlieen, Mathe[60] G. Gentzen, Untersuchungen u
matische Zeitschrift, vol. 39 (1935), pp. 176-210, 405-431 (English
translation: Investigations into logical deduction, in [62], pp. 68-131)
[Chapter 1, 1.3, 1.6, 7.6, 8.1, 11.1-2]
[61] , Neue Fassung des Widerspruchsfreiheitsbeweises f
ur die reine
Zahlentheorie, Forschungen zur Logik und zur Grundlegung
der exakten Wissenschaften (N.S.), vol. 4 (1938), pp. 19-44 (English translation: New version of the consistency proof for elementary
number theory, in [62], pp. 252-286) [11.2]
[62] , The Collected Papers of Gerhard Gentzen (M.E. Szabo,
editor), North-Holland, Amsterdam, 1969
[63] J.-Y. Girard, Linear logic, Theoretical Computer Science, vol.
50 (1987), pp. 1-101 [7.2, 7.8, 8.1]
[64] J.-Y. Girard, P. Taylor and Y. Lafont, Proofs and Types,
Cambridge University Press, Cambridge, 1989 [1.6, 11.1]
[65] V.N. Grishin, A generalization of the Ajdukiewicz-Lambek system
(in Russian), Investigations in Nonclassical Logics and Formal Systems (in Russian), Nauka, Moscow, 1983, pp. 315-334
(Mathematical Reviews 85f:03068) [7.1]

Bibliography

365

[66] D. Hilbert and W. Ackermann, Grundz

uge der theoretischen
Logik , Springer, Berlin, 1928 (English translation of the second edition from 1938, Principles of Mathematical Logic, Chelsea, New
York, 1950) [1.5, 9.3]
[67] J.M.E. Hyland, Proof theory in the abstract, Annals of Pure and
Applied Logic, vol. 114 (2002), pp. 43-78 [1.6, 11.1, 14.3]
[68] J.M.E. Hyland and V. de Paiva, Full intuitionistic linear logic
(extended abstract), Annals of Pure and Applied Logic, vol. 64
(1993), pp. 273-291 [7.1]
[69] C.B. Jay, Coherence in category theory and the Church-Rosser property, Notre Dame Journal of Formal Logic, vol. 33 (1992), pp.
140-143 [Preface]
[70] V.F.R. Jones, A quotient of the ane Hecke algebra in the Brauer
algebra, LEnseignement Math
ematique (2), vol. 40 (1994), pp.
313-344 [1.4]
[71] A. Joyal and R. Street, The geometry of tensor calculus, I, Advances in Mathematics, vol. 88 (1991), pp. 55-112 [4.6]
[72] , Braided tensor categories, Advances in Mathematics, vol.
102 (1993), pp. 20-78 [Preface, 1.7, 3.1, 4.5-7, 9.4]
[73] C. Kassel, Quantum Groups, Springer, Berlin, 1995 [1.4]
[74] L.H. Kauffman and S.L. Lins, Temperley-Lieb Recoupling
Theory and Invariants of 3-Manifolds, Annals of Mathematical Studies, vol. 134, Princeton University Press, Princeton, 1994
[1.4]
[75] G.M. Kelly, On Mac Lanes conditions for coherence of natural
associativities, commutativities, etc., Journal of Algebra, vol. 1
(1964), pp. 397-402 [4.6, 5.3]
[76] , Many-variable functorial calculus, I, in [80], pp. 66-105 [11.1]
[77] , An abstract approach to coherence, in [80], pp. 106-147 [Preface, 9.2]
[78] , A cut-elimination theorem, in [80], pp. 196-213 [Preface]

366

Bibliography

[79] G.M. Kelly and M.L. Laplaza, Coherence for compact closed categories, Journal of Pure and Applied Algebra, vol. 19 (1980),
pp. 193-213 [Preface, 1.1]
[80] G.M. Kelly, M.L. Laplaza, G. Lewis and S. Mac Lane, Coherence in Categories, S. Mac Lane, editor, Lecture Notes in Mathematics, vol. 281, Springer, Berlin, 1972 [1.1]
[81] G.M. Kelly and S. Mac Lane, Coherence in closed categories,
Journal of Pure and Applied Algebra, vol. 1 (1971), pp. 97-140,
219 [Preface, 1.1, 1.4]
[82] S.C. Kleene, Introduction to Metamathematics, North-Holland, Amsterdam, 1952 [Chapter 1]
[83] G. Kreisel, A survey of proof theory II, Proceedings of the
Second Scandinavian Logic Symposium (J.E. Fenstad, editor),
North-Holland, Amsterdam, 1971, pp. 109-170 [1.5]
[84] J. Lambek, Deductive systems and categories I: Syntactic calculus
and residuated categories, Mathematical Systems Theory , vol. 2
(1968), pp. 287-318 [Preface, 1.1, 1.4]
[85] , Deductive systems and categories II: Standard constructions
and closed categories, Category Theory, Homology Theory and
their Applications I , Lecture Notes in Mathematics, vol. 86,
Springer, Berlin, 1969, pp. 76-122 [Preface, 1.1, 1.4, 7.8]
[86] , Deductive systems and categories III: Cartesian closed categories, intuitionist propositional calculus, and combinatory logic,
Toposes, Algebraic Geometry and Logic (F.W. Lawvere, editor), Lecture Notes in Mathematics, vol. 274, Springer, Berlin, 1972,
pp. 57-82 [Preface, 1.4]
[87] , Functional completeness of cartesian categories, Annals of
Mathematical Logic, vol. 6 (1974), pp. 259-292 [Preface, 1.3-4]
[88] , Multicategories revisited, Categories in Computer Science
and Logic (J.W. Gray and A. Scedrov, editors), American Mathematical Society, Providence, 1989, pp. 217-239 [7.8]
[89] , From categorial grammar to bilinear logic, Substructural
Logics (K. Dosen and P. Schroeder-Heister, editors), Oxford University Press, Oxford, 1993, pp. 207-237 [7.1]

Bibliography

367

[90] J. Lambek and P.J. Scott, Introduction to Higher Order Categorical Logic, Cambridge University Press, Cambridge, 1986 [Preface, Chapter 1, 1.1-3, 2.2, 9.1, 9.6, 13.4, 14.3]
[91] M.L. Laplaza, Coherence for distributivity, in [80], pp. 29-65 [11.3]
[92] , A new result of coherence for distributivity, in [80], pp. 214-235
[11.3]
[93] , Coherence for associativity not an isomorphism, Journal of
Pure and Applied Algebra, vol. 2 (1972), pp. 107-120 [4.2]
[94] F.W. Lawvere, Adjointness in foundations, Dialectica, vol. 23
(1969), pp. 281-296 [Preface, 1.3]
[95] F.W. Lawvere and S.H. Schanuel, Conceptual Mathematics:
A First Introduction to Categories, Cambridge University Press,
Cambridge, 1997 [1.2, 11.1, 12.1, 13.3]
[96] T. Leinster, A survey of denitions of n-category, Theory and
Applications of Categories, vol. 10 (2002), pp. 1-70 [1.1, 1.4]
[97] W.B.R. Lickorish, An Introduction to Knot Theory , Springer,
Berlin, 1997 [1.1, 1.4]
[98] J.-L. Loday et al., editors, Operads: Proceedings of Renaissance Conferences, Contemporary Mathematics, vol. 202, American Mathematical Society, Providence, 1997
[99] S. Mac Lane, Natural associativity and commutativity, Rice University Studies, Papers in Mathematics, vol. 49 (1963), pp. 2846 [Preface, 1.1, 1.4, 2.9, 3.1-2, Chapter 4, 4.2-3, 4.6, Chapter
5, 5.1, 5.3]
[100] , Categories for the Working Mathematician, Springer,
Berlin, 1971 (incorporated in the expanded edition [102]) [Chapter 1,
1.7, 3.1-2, Chapter 4, 4.2-3, 4.6, 5.1, 7.9, 11.3, 12.1, 13.23, 14.2-3]
[101] , Why commutative diagrams coincide with equivalent proofs,
Algebraists Homage: Papers in Ring Theory and Related
Topics (S.A. Amitsur et al., editors), Contemporary Mathematics,
vol. 13, American Mathematical Society, Providence, 1982, pp. 387401 [Preface]

368

Bibliography

[102] , Categories for the Working Mathematician, second edition, Springer, Berlin, 1998 [1.7, 2.8, 3.1-2, 4.5, 4.7, Chapter 5,
5.1]
[103] S. Maehara, Eine Darstellung der intuitionistischen Logik in der
klassischen, Nagoya Mathematical Journal , vol. 7 (1954), pp.
45-64 [1.3, 11.2]
[104] M. Makkai and G.E. Reyes, First Order Categorical Logic,
Lecture Notes in Mathematics, vol. 611, Springer, Berlin, 1977 [Chapter 2]
[105] C.R. Mann, The connection between equivalence of proofs and cartesian closed categories, Proceedings of the London Mathematical
Society (3), vol. 31 (1975), pp. 289-310 [1.4]
[106] G.E. Mints, Category theory and proof theory (in Russian), Aktualnye voprosy logiki i metodologii nauki , Naukova Dumka,
Kiev, 1980, pp. 252-278 (English translation, with permuted title,
in: G.E. Mints, Selected Papers in Proof Theory, Bibliopolis, Naples,
1992) [9.2]
[107] Y.N. Moschovakis, What is an algorithm?, Mathematics Unlimited2001 and Beyond (B. Engquist and W. Schmid, editors),
Springer, Berlin, 2001, pp. 919-936 [1.3]
, Coherence in substructural categories, Studia Logica,
[108] Z. Petric
vol. 70 (2002), pp. 271-296 (available at: http://arXiv.org/math.CT/
0006061) [1.4, 9.2, 11.1, 14.4]
[109] , G-Dinaturality, Annals of Pure and Applied Logic, vol.
122 (2003), pp. 131-173 (available at: http://arXiv.org/math.CT/
0012019) [1.1, 1.3, 11.1]
[110] V.V. Prasolov and A.B. Sosinski, Knots, Links, Braids and
3-Manifolds (in Russian), MCNMO, Moscow, 1997 [1.4]
[111] D. Prawitz, Natural Deduction: A Proof-Theoretical Study ,
Almqvist & Wiksell, Stockholm, 1965 [Chapter 1, 1.3]
[112] , Ideas and results in proof theory, Proceedings of the Second
Scandinavian Logic Symposium (J.E. Fenstad, editor), NorthHolland, Amsterdam, 1971, pp. 235-307 [1.3, 1.5]

Bibliography

369

[113] H. Rasiowa An Algebraic Approach to Non-Classical Logics,

North-Holland, Amsterdam, 1974 [14.1]
[114] J. Riguet, Relations binaires, fermetures, correspondances de Galois, Bulletin de la Soci
et
e math
ematique de France, vol. 76
(1948), pp. 114-155 [1.4, 14.3]
[115] R.R. Schneck, Natural deduction and coherence for non-symmetric
linearly distributive categories, Theory and Applications of Categories, vol. 6 (1999), pp. 105-146 [7.9]
[116] M.C. Shum, Tortile tensor categories, Journal of Pure and Applied Algebra, vol. 93 (1994), pp. 57-110 [3.2]
[117] A.K. Simpson, Categorical completeness results for the simply-typed
lambda-calculus, Typed Lambda Calculi and Applications (M.
Dezani-Ciancaglini and G. Plotkin, editors), Lecture Notes in Computer Science, vol. 902, Springer, Berlin, 1995, pp. 414-427 [Preface,
1.5, 9.3]
[118] S.V. Soloviev, The category of nite sets and cartesian closed categories (in Russian), Zapiski Nauchnykh Seminarov (LOMI),
vol. 105 (1981), pp. 174-194 (English translation in Journal of Soviet Mathematics, vol. 22 (1983), pp. 1387-1400) [Preface]
[119] , On the conditions of full coherence in closed categories, Journal of Pure and Applied Algebra, vol. 69 (1990), pp. 301329 (Russian version in Matematicheskie metody postroeniya
i analiza algoritmov , A.O. Slisenko and S.V. Soloviev, editors,
Nauka, Leningrad, 1990, pp. 163-189) [Preface, 7.9]
[120] , Proof of a conjecture of S. Mac Lane, Annals of Pure and
Applied Logic, vol. 90 (1997), pp. 101-162 [Preface]
[121] R. Statman, -denable functionals and -conversion, Archiv
f
ur mathematische Logik und Grundlagenforschung , vol. 23
(1983), pp. 21-26 [1.5]
[122] J.D. Stasheff, The pre-history of operads, in [98], pp. 9-14 [1.1]
[123] , From operads to physically inspired theories, in [98], pp. 53-81
[1.1]
[124] M.E. Szabo, A categorical characterization of Boolean algebras, Algebra Universalis, vol. 4 (1974), pp. 192-194 [14.3]

370

Bibliography

[125] , A counter-example to coherence in cartesian closed categories,

Canadian Mathematical Bulletin, vol. 18 (1975), pp. 111-114
[1.4]
[126] , Polycategories, Communications in Algebra, vol. 3(8)
(1975), pp. 663-689 [7.9]
[127] , Algebra of Proofs, North-Holland, Amsterdam, 1978 [Preface, 1.6]
[128] A.S. Troelstra and H. Schwichtenberg, Basic Proof Theory ,
Cambridge University Press, Cambridge, 1996 (second edition, 2000)
[Preface, Chapter 1, 9.2]
[129] V.G. Turaev, Operator invariants of tangles and R-matrices (in
Russian), Izvestiya Akademii Nauk SSSR, Seriya Matematicheskaya, vol. 53 (1989), pp. 1073-1107 (English translation in
Mathematics of the USSR-Izvestiya, vol. 35, 1990, pp. 411-444)
[1.4]
[130] J. von Plato, A proof of Gentzens Hauptsatz without multicut,
Archive for Mathematical Logic, vol. 40 (2001), pp. 9-18 [11.1]
[131] R. Voreadou, Coherence and Non-Commutative Diagrams
in Closed Categories, Memoirs of the American Mathematical Society, no 182, American Mathematical Society, Providence, 1977 [1.1]
[132] H. Wenzl, On the structure of Brauers centralizer algebras, Annals
of Mathematics, vol. 128 (1988), pp. 173-193 [1.4]
ck, Identity of Proofs, doctoral dissertation, Univer[133] F. Wideba
sity of Stockholm, Almqvist & Wiksell, Stockholm, 2001 [1.5]
[134] D.N. Yetter, Markov algebras, Braids (J.S. Birman and A. Libgober, editors), Contemporary Mathematics, vol. 78, American Mathematical Society, Providence, 1988, pp. 705-730 [1.4]
[135] R. Zach, Completeness before Post: Bernays, Hilbert, and the development of propositional logic, The Bulletin of Symbolic Logic,
vol. 5 (1999), pp. 331-366 [1.5, 9.3]

Index

A, 115, 345
A, , 116, 345

A, 93, 345

A , 89, 345

A , 101, 345

A, 115, 345

A , 116, 345
Ab-categories, 304
abelian categories, 304
Acnf , 264
addition of proofs, 26
additive categories, 304
adjunction, 164
admissible rule, 324
Adnf , 264
alternative denition of , 72
angle normal form, 223
arrow, 36
arrow term, 38
Artin, E., 18
associative category, 94
Associative Coherence, 94
associative normal form, 98
Associative Normal-Form Proposition,
98
atomic bracket-free normal form, 223
atomic component, 223, 299
atomic correspondence, 226
atomic formula, 35
atomic term, 299
Atomic-k Lemma, 246, 272
371

atomized c-term, 120

atomized arrow term, 170
axiom, 41
axiomatizability, 3
B, 317, 352
b, 47

b-term, c-term, etc., 50

Bohms Theorem, 25, 194
Balance Remark, 157
balance weight, 198
balanced type, 84
Barendregt, H.P., 25
basic arrow term, 122
basic sequence, 238, 271
Basic-Development Lemma, 122, 123
basically developed arrow term, 122
Bernays, P., 25, 196
biassociative category, 115
Biassociative Coherence, 116
bicartesian category, 208
biendofunctor, 43
bifunctor, 42
bifunctorial category, 54
bifunctorial equations, 42, 51
bimonoidal category, 116
Bimonoidal Coherence, 117
binary connective, 34
blocked wL term, 250
blocked wR term, 252
Blocked-w Lemma, 253, 273
Boolean category, 322

372
Boolean Coherence, 321
Boolean negation, 309
bound, 299
bracket-free normal form, 223
branching, 35
Brauer algebras, 18
Brauer, R., 19
Bund, 248
C, 318
c, 47
c-equivalent form sequences, 148
C-functor, 54
C-strict category, 75
C-strict deductive system, 75
Card, 83
cartesian category, 191
Cartesian Coherence, 193
cartesian linearly distributive categories, 234, 303
categorial equations, 38
categorication, 6
category, 38
category of the C kind, 54
Cayley, A., 105
C/E-category, 59
(C/E, C )-strictied category, 76
Church-Rosser property, 93
clean cut, 253
cluster, 248
C , 318
(co ), 40
cocartesian category, 207
Cockett, J.R.B., ii, iii, 8, 128
coherence, 1, 17, 62
coherent bicartesian category, 208
coherent sesquicartesian categories,
208
colour of form sequence, 117
ComMon, 307
commuting problem, 2

Index
comparable form sequences, 135
comparable formulae, 36
completeness, 1
complex identity, 50
complexity of topmost cut, 152, 253
Composition Elimination, 169, 189,
201, 203, 282, 290, 298, 304
composition of arrows, 37
conuence property, 93
conjunctive normal form, cnf, 212,
223
connectedness in proof nets, 158
constant object, 122
contravariant functor, 43
correspond obviously, 171
C -core, 73

Cubri
c, Dj., ii
Curry-Howard correspondence, 12
cut, 246
cut formula, 241
Cut-Elimination Theorem, 152, 177,
253, 273, 291, 323
cut-free Gentzen term, 151, 246
Cut-Free Preordering, 163, 182
D, 238, 271, 291
d, 47, 145
DA, 132, 348
DAst , 133
DA, , 163, 348
De Morgan Coherence, 315
De Morgan lattices, 316
De Morgan negation, 309
decidability, 2
deductive system, 37
deductive system of the C kind, 54
denable connective, 45
degree of cut, 253
, 47
-, 47

Index
depth of subterm of Gentzen term,
151
derivable rule, 324
derivation, in equational system, 41
developed arrow term, 53
Development Lemma, 53, 311
DI, 128, 347
dicartesian category, 208
difunctional relation, 20, 328
direct strictication, 78
Direct-Strictication Theorem, 81
-directed arrow term, 94, 102, 115,
116, 311
Directedness Lemma, 94, 102, 311
discrete deductive system, 37
disjunctive normal form, dnf, 212,
223
dissociative biassociative category, 132
dissociative bimonoidal category, 164
dissociative category, 128
Dissociative Coherence, 130
dissociativity, 8, 127, 128
distributive dicartesian category, 270
Distributive Dicartesian Coherence,
274
distributive involution lattices, 316
distributive lattice category, 234
Distributive Lattice Coherence, 266
diversied arrow term, 85
diversied formula, 85
diversied type, 85
DL, 233, 349
DLA , 237
DLA, 143, 347
DLAst , 144
A,
, 271
DL,
DL , 235
DL, , 270, 349
DS, 145, 348
dual C-functor, 54

373
dual graph, 37
Dunn, J.M., 316
d1 ZL, 278
Efde , 316
Empty-Relation Lemma, 283
endofunctor, 43
epi arrow, 39
Epstein, D.B.A., 2
equation, in syntactical system, 39
equational system, 39
equivalent categories, 44
equivalent deductive systems, 44
Extraction Lemma, 91, 96, 129, 136,
144, 175
factor, 53
-factor, 210
-factor, 210
factorized arrow term, 53
faithful functor, 42
faithful graph-morphism, 42
nite tree, 35
owing through, 43, 73
uent C-functor, 56
form multiset of letters, 148
form sequence, 117, 118
form sequence of letters, 133
form sequence, extended sense, 119
form sequence, nonextended sense,
119
form set of letters, 148
formula, 35
fractional notation, 37
Freyd, P.J., 324
full subsystem, 39
functor, 42
GDS, 148
generality, 16
Generality Conjecture, 17

374
generatively discrete equivalence relation, 67
generatively discrete logical system,
69
Gentzen operations, 149, 177, 240,
323
Gentzen term, 291
Gentzen terms, 149, 177, 240, 323
Gentzen, G., iiv, 11, 12, 14, 22, 26
28, 31, 148, 163, 167, 168,
185, 188, 231233, 239, 240,
243245, 248, 249, 253255,
257, 263, 273, 276, 280
Gentzenization Lemma, 151, 177, 245,
272, 291
GL, 201
GL, , 209
GL, , 209

GL, 188

GL , 203

GL , 192

GL, 200

GL , 203
GMDS, 177
GML, 220
GML, , 227
grade, 193
graph, 36
graph-morphism, 42
graphical category, 17
groupoid, 39
groupoidal C-functor, 57
GZIML, 289
GZL, 282
GZL, , 282
GZML, 298
GZML, , 304
head, 50
height of a node, 36

Index
hexagonal equation, 107
Hilbert, D., v, 25, 28, 196
holding, of equation, 41
hom-set, 37
I, 89, 345

I, 89, 345

I, 345
I , 311
I Coherence, 311

I,
, 312

I,
Coherence, 312
identity arrow, 37
identity arrow term, 38
identity functor, 43
immediate scope, 36
inducing a graph-morphism, 44
inx notation, 35
initial object, 37
inverse, 39
invertibility, 148
Invertibility Lemma for , 159, 178,
181, 201, 283
Invertibility Lemma for , 162, 178,
181, 202, 284
Invertibility Lemma for mix , 181
Invertibility Remark for , 283
Invertibility Remark for , 284
Isbell, J., 83
isomorphic categories, 44
isomorphic deductive systems, 44
isomorphic objects, 39
isomorphism, 39
Jones, V.F.R., 18
Joyal, A., 30, 65, 245, 324
K, 88
k, 47
k-atomized Gentzen term, 246
, 47

Index
Kelly, G.M., ii, 4, 21
Kleisli category, 308
K , 312
K -Kp -Equivalence, 314
Kp , 312
Kreisel, G., 25
L, 200, 347
L , 208, 347
L, , 207, 347
L, , 208
L , 208, 347

L, 185, 346

L , 203, 346

L , 191, 346

L, 199, 346

L , 203, 346

L , 207, 346
L, 34
L , L , etc., 36
L , 310
Lp , 310
label, 248
Lafont, Y., 27, 233
Lambek, J., ii, 13, 1518, 20, 24
language, 34
last falling slope, 111
lattice category, 200
Lattice Coherence, 202
Lawvere, F.W., ii, 13, 18, 22
leaf, 35
leaf formulae, 247
left cut formula, 249
left rank, 249
left-hand side, in a planar tree, 36
legitimate relation, 268
length of a word, 34
letter length, 35
letters, 34

level of atomized c-term, 120

375
lexicographical order, 152, 193, 253
linear category, 304
linearly distributive category, 164
logical category, 51
logical system, 47
lower contraction formula, 247
lower parametric basic sequences, 248
m, 47
m1 , 47
MA, 169, 348
MAst , 170
Mac Lane, S., i, ii, iv, 24, 21, 30, 58,
63, 65, 82, 8789, 94, 107,
108, 245
main conjunct, 36
main disjunct, 36
manageable category, 4
Martin-Lof, P., 11
Mat, 293
maximal relation, 269
maximality, 24, 194, 197, 198, 213,
274, 285, 292, 305, 322
MDA, 173, 349
MDA-comparable, 174
MDAst , 174
MDI, 169, 349
MDS, 176, 349
member of basic sequence, 238
MI, 167, 348
mingle, 329
minimal conjunct, 225
minimal disjunct, 225
Mischung, 167, 244, 280
mix, 26, 167
mix category, 167
Mix Coherence, 169
mix-biassociative category, 170
Mix-Biassociative Coherence, 173
mix-dicartesian category, 227
mix-dissociative category, 169

376

Index

Mix-Dissociative Coherence, 169

mix-lattice category, 219
mix-net category, 173
Mix-Net Coherence, 176
mix-symmetric biassociative category,
182
Mix-Symmetric Biassociative Coherence, 182
mix-symmetric net category, 177
Mix-Symmetric Net Coherence, 177
ML, 219, 348
ML, , 227, 348
modularity law, 128
molecular component, 226
molecular correspondence, 226
mono arrow, 39
monoidal category, 101
Monoidal Coherence, 103
MS, 182, 348
multiple-conclusion sequent, 232

nonzero atomic bracket-free term, 230,

289
nonzero atomic term, 299
normal form, 94, 110, 115, 116, 189,
192, 223, 229, 299
Normal-Form Lemma, 111, 189, 192,
224, 265, 290, 299, 305
Normalization Conjecture, 11
nose, 58
null object, 281
null term, 229
nullary connective, 34

n, 310
n-ary branching, 35
n-ary connective, 34
n-categories, 4, 7, 23
n-endofunctor, 42
natural C/E-category, 59
natural in, 43
natural isomorphism, 43
natural logical category, 51
natural notation, 117
natural transformation, 43
naturality equations, 43, 52
naturally isomorphic functors, 44
negation, 310
net category, 132
Net Coherence, 138
net normal form, 143
node, 35
Nonoverlapping Lemma, 134

P, 34
partial C-functor, 56
partial order, 38
partial skeleton, 39
path, 35
pentagonal equation, 89, 132
place, in form sequence, 135, 174
place, in formula, 36
planar nite tree, 36
plural sequent, 14, 232
P , 310
Polish notation, 35
polycategories, 165
Post completeness, 24, 194
power-set monad, 308
Prawitz, D., 1013, 15, 16, 21, 24,
25, 28
preadditive categories, 304
predecessor, 35

0, 277
0A , 287
object, 37
obvious correspondence, 171
octagonal equation, 186
on the nose, 58
operation on arrow terms, 38
0ZL, 277

Index
prex notation, 35
preorder, 37
presentation by generators and relations, 3
primitive arrow term, 38
product of deductive systems, 42
proof nets, 132, 158, 163, 165
proper subword, 36
proper zero term, 282
propositional constant, 34
propositional formula, 35
propositional language, 34
propositional letter, 34
propositional variable, 34

377
root, 35
S, 119, 346
S, , 121, 346

S, 107, 345

S , 112, 345

S, 119, 345

div
S , 113

div
S , 113

S , 121, 345
S , 124, 346

S , 125, 345
Sst , 124
Sst , 119
quantity of letters in arrow, 163

Sst , 109
quasi-Boolean algebras, 316

Sst
, 113
Sst
r, 310
, , 122
same place, in form sequence, 135
rank, 249
same place, in formula, 36
rank of topmost cut, 152
scope, 36
-rank, 151
Scott, P.J., ii
-rank, 152
Seely, R.A.G., ii, iii, 8, 128
p-rank, 152
semiassociative category, 89
(re ), 40
Semiassociative Coherence, 92
Reducibility Lemma, 250
semidissociative biassociatice category,
reducible subterm, 250, 252
143
Reidemeister moves, 3, 18
Semilat, 222, 228, 306, 307
Rel, 59
Semilat , 306, 307
relevant net categories, 238
Restricted Dicartesian Coherence, 212 semilattice category, 186
Restricted Mix-Dicartesian Coherence, Semilattice Coherence, 190
sequent, 239
229
Restricted Mix-Lattice Coherence, 226 sequent arrow, 239
Restricted Zero-Identity Mix-Dicarte- sesquicartesian category, 208
Sesquicartesian Coherence, 209
sian Coherence, 290
Set, 9, 38, 83, 121, 208, 267
Restricted Zero-Identity Mix-Lattice
Set , 213, 214, 275, 287, 303, 307
Coherence, 289
, 312
Set , 214
right cut formula, 249
Setsl
, 221, 228, 307
right rank, 249
settled normal form, 223, 230, 299

378
Shnider, S., 2
, 47
Simpson, A.K., ii
single-conclusion sequent, 232
singular sequent, 14, 232
skeleton, 39
small categories, 38
Soloviev, S.V., ii
source, 36, 37

span of atomized c-term, 120

specic equation, 88
split equivalence relations, 326
Split-Normalization Lemma, 180
split-normalized Gentzen term, 180
splittable arrow, 157
splittable pair of form sets, 157
Splitting Corollary, 158
Splitting Remark, 158
standard form, 209, 282
Standard-Form Lemma, 209, 282
Stashe, J.D., 2
Street, R., 30, 65, 245
strict C-functor, 58
strict category, 65
strictication, 29, 65
Strictication Corollary, 76
Strictication Theorem, 76
Strictication-Coherence Equivalence,
77
Strictication-Coherence Implication,
78
strictied category, 65, 76
strictifying equation, 79
strong C-functor, 58
structural rules, 240
structure-preserving functor, 55, 56,
58
(su ), 51
subcategory, 39
subformula, 36

Index
subgraph, 39
subsystem, 39
subterm, 38
subword, 36
successor, 35
(sy ), 40
symbol, 34
symmetric associative category, 108
Symmetric Associative Coherence, 108
symmetric biassociative category, 119
Symmetric Biassociative Coherence,
120
symmetric bimonoidal category, 121
Symmetric Bimonoidal Coherence, 123
symmetric groups, 110
symmetric monoidal category, 113
Symmetric Monoidal Coherence, 113
symmetric net category, 145
Symmetric Net Coherence, 147
synonymous syntactical systems, 44
syntactical category, 6, 17, 41
syntactical system, 38
Tait, W.W., 11
tangles, 18
target, 36, 37
Temperley-Lieb algebras, 18
tensor category, 101
terminal object, 37
theoremhood problem, 2
Theoremhood Proposition, 91, 129,
137, 142, 144, 175
tied, subterm tied to cut, 249
topmost cut, 151, 189, 246
total C-functor, 56
total split, 157
(tr ), 40
transformation, 43
tree, 35
type, 37

Index
unary connective, 34
union of proofs, 26
Uniqueness Lemma, 111, 190, 193
upper contraction formulae, 247
upper parametric basic sequence, 248
valuation, 57
variable object, 122
w, 47
weakly distributive category, 164
wL subterm, 249
wL term, 249
word, 34
zero arrow, 277
zero arrow term, 277
zero atomic bracket-free term, 230,
289
zero atomic term, 299
zero proof, 27
zero term, 282
zero-dicartesian category, 281
Zero-Dicartesian Coherence, 284
zero-identity arrow, 287
zero-identity arrow term, 287
zero-identity Boolean category, 322
Zero-Identity Distributive Dicartesian
Coherence, 291
Zero-Identity Distributive Lattice Coherence, 291
zero-lattice category, 276
Zero-Lattice Coherence, 282
zero-mix dicartesian category, 304
Zero-Mix Dicartesian Coherence, 305
zero-mix lattice category, 296
Zero-Mix Lattice Coherence, 300
Zero-Term Lemma, 283, 284
ZIDL, 290, 351
ZIDLA , 291
ZIDL , 292

379
ZIDL, , 290, 352
ZIDL
, , 292
ZIL, 287, 350
ZIL, , 288, 350
ZIML, 289, 351
ZIML , 293
ZIML, , 289, 351
ZIML
, , 293
ZL, 276, 350
ZL, , 281, 350
ZML, 296, 353
ZML , 305
ZML, , 304, 353
ZML
, , 305
ZML
, , 353