Zoran Petrić

270 Minutes on Categorial Proof Theory

Abstract. The aim of these notes is to provide an introduction of

basic categorial notions to a reader interested in logic and proof the-
ory. The first part is devoted to justification of these notions through
a cut elimination procedure. In the second part a classification of
formulae up to isomorphism, and an example of coherence are given.

Mathematics Subject Classification (2010): Primary: 18-01, 03F07,

18A15; Secondary 03F05, 18D10.

Keywords: sequent system, cut elimination, functor, natural transformation,

adjunction, isomorphism, coherence.

Acknowledgements: I am grateful to Mirjana Borisavljević and Kosta Došen for

very useful discussions concerning this text and for correcting some mistakes in
its previous version. I am also grateful to Slavko Moconja, Zoran Petrović and
Marko Radovanović for attending lectures and giving comments during my talks
on which this text is based. I would like to thank very much to Silvia Ghilezan for
her patience during the publishing procedure.
This work was supported by project ON174026 of the Ministry of Education,
Science, and Technological Development of the Republic of Serbia.
 Contents
1. About these notes 2
2. The three propositional languages 2
3. Natural deduction and sequent systems 4
4. From sequent systems to categories 11
5. The classification of formulae 21
6. Categories with products, coherence 25
7. Hints for the exercises 31
References 32

1. About these notes

This text is based on a series of talks delivered by the author at the Algebra
and Logic Seminar of the Mathematical Faculty in Belgrade. Zoran Petrović, the
chairman of the seminar, invited the author to prepare a short course on categorial
proof theory. Once a week, during three weeks of spring 2014, a 90 minutes talk
was organized. Hence the title “270 Minutes...”.
The aim of these notes is to provide an introduction of basic categorial notions
to a reader interested in logic and proof theory. After this introduction, two im-
portant topics of categorial proof theory, namely the classification of formulae up
to isomorphism and coherence, are given by simple examples.
Categorial proof theory as a field of general proof theory was established during
the 1960s. We refer to [2, Preface] for some historical notes on this topic. We can’t
go into details here because our intention is to make a text readable in four and
a half hour. However, two names should be mentioned. The work of Jim Lambek
and Bill Lawvere made a substantial influence to the field.
In order to make it self-contained, the definitions of relevant categorial and other
notions are given in frame boxes at appropriate places throughout this text. Some
proofs are not finished—this is left as an exercise for the reader.

2. The three propositional languages

We deal with fragments of propositional logic throughout this text. Our consid-
erations will vary from one such fragment to another keeping the context as simple
as possible. We introduce three propositional languages whose alphabet consists of
an infinite set of propositional letters p, q, r, . . ., two binary propositional connec-
tives ∧, →, one nullary connective > and two auxiliary symbols ( and ). However,
2
 270 MINUTES ON CATEGORIAL PROOF THEORY 3

it is not the case that all the symbols occur in every language. The words, are the
finite sequences of symbols from the alphabet. A language is a set of words called
formulae.
The formulae are defined inductively as follows.

(1) Propositional letters and > are formulae.

(2) If A and B are formulae, then (A ∧ B) and (A → B) are formulae.
(3) Nothing else is a formula.

We omit the outermost parentheses in formulae taking them for granted. The
implicational language consists of all the formulae in which > and ∧ do not occur.
The conjunction language consists of all the formulae in which > and → do not
occur. The mixed language consists of all the formulae.
The mixed language has its strict variant which we introduce for the sake of sim-
plicity of the forthcoming notions. In this variant, we consider ∧ to be associative
and > to be the unit for ∧. The formulae of this language are defined inductively
as follows.

(1) Propositional letters and > are formulae called atomic.

(2) If A and B are formulae, then hAi → hBi is a formula (called implication),
where hXi is (X), when X is not atomic, or X, otherwise.
(3) If A and B are formulae different from >, then kAk ∧ kBk is a formula,
where kXk is (X), when X is an implication, or X, otherwise.
(4) Nothing else is a formula.

For example, (p ∧ q ∧ r) → p is a formula of the strict mixed language.

In Section 4, where we use this language, we write just A → B instead of the
more complicated, but correct hAi → hBi. Also, we write just A ∧ B, which is A
when B is >, or which is B when A is >, or which is kAk ∧ kBk. This is just in
order not to overburden the notation.
Formal systems, equational systems
A formal system is given by:
1. Alphabet, a set of symbols.
2. Language, a set of words in this alphabet. The elements of the language
are called formulae.
3. Axioms, a set of formulae.
4. Rules of inference, a set of relations on the language of arities greater
or equal to 2, where ρ(A1 , . . . , An , A) means that the formula A is a direct
consequence of the formulae A1 , . . . , An . The rules of inference given by
schematic line rules are called inference figures.
4 ZORAN PETRIĆ

A derivation in a formal system is a finite tree with formulae in its vertices.

The formulae in the leaves, which are not axioms, are called hypotheses,
and the derived formula is in the root. Every branching of this tree cor-
responds to one of the rules of inference. For example, if ρ is a rule of
inference and we have ρ(A1 , A2 , A3 , A), then a derivation may contain the
following branching:

A1 A2 A3
@
@
A
A proof is a derivation without hypothesis. A formula derived without
hypothesis is a theorem of the system.

An equational system is a formal system whose alphabet contains a special

symbol usually denoted by =. The formulae of an equational system are
of the form u = v, where u and v are words in the alphabet, called terms.
The axioms and the rules of inference are such that the relation ≡, defined
on the set of terms by
u≡v when u = v is a theorem of the system,
is an equivalence relation congruent with the operations on terms corre-
sponding to the operation symbols of the alphabet.

3. Natural deduction and sequent systems

In this section, we concentrate on the implicational language only. A Hilbert-
style system H, for implicational fragment of intuitionistic logic, is a formal system
in this language, given by the following axiom schemata
A → (B → A)
(A → (B → C)) → ((A → B) → (A → C))
and modus ponens (MP) as the only inference figure.
A→B A
B
Exercise. Show that (p → q) → ((q → r) → (p → r)) is a theorem of H.

This is not an easy task. It would not be a surprise if one starts with proving a
metatheorem like the deduction theorem. Then, since r is easily deducible from the
hypotheses p → q, q → r and p, the deduction theorem guarantees the existence of
a direct proof, leaving it implicit. Gentzen’s idea was to built in the system all the
deduction power and leave metatheorems for more serious results.

A natural deduction system, for the corresponding fragment of logic, is a formal

system in the implicational language without axioms and with the following two
inference figures. (The notation [A] in the first figure below means that, after
 270 MINUTES ON CATEGORIAL PROOF THEORY 5

applying this rule of inference, an arbitrary number of hypotheses of the form A

above the formula B are no longer active in the derivation.)
[A]
B A→B A
A→B B
Simply-typed λ calculus
This is an equational system whose alphabet consists of symbols λ, ., →,
(, ), :, =, and two disjoint infinite sets, the set of variables x, y, z, . . . and
the set of propositional letters (usually called atomic types).
There is a function, called typing, which maps the set of variables to the
set of implicational formulae. For every implicational formula, there are
infinitely many variables mapped to this formula.
The terms are words in this alphabet, defined inductively as follows:
(1) If x is a variable whose type is A, then x : A is a term.
(2) If M : A → B and N : A are terms, then (M N ) : B is a term.
(3) If x : A and M : B are terms, then (λx .M ) : A → B is a term.
(4) Nothing else is a term.
If x occurs in the scope of λx . in a term, then this occurrence of x is bound ;
otherwise, it is free in this term. The notation x 6∈ F V (M ) means that
every occurrence of x in M is bound. For a term M , a variable x and a
x
term N with the same type as x, we define MN as the result of uniformly
substituting the term N for every free occurrence of x in M . By renaming
of bound variables, one avoids free variables of N to become bound after
substitution.
The language consists of equalities of the form X = Y , where X and Y
are terms with the same type. The axiom schemata are
X = X, λx .M = λy .Myx , y 6∈ F V (M ),
x
(λx .M )N = MN , M = λx .M x, x 6∈ F V (M ),
and we have the following inference figures
X=Y X=Y Y =Z
Y =X X=Z
X=Y Z=T X=Y
XZ = Y T λz .X = λz .Y
The Curry-Howard correspondence assigns to a natural deduction derivation a
simply-typed λ term, following the clauses:
[x : A]
M :B M :A→B N :A
λx .M : A → B MN : B
6 ZORAN PETRIĆ

Hence, a proof, i.e., a derivation without hypotheses, of the formula (p → q) →

((q → r) → (p → r)) reads

[x : p → q] [z : p]
[y : q → r] xz : q
y(xz) : r
λz .y(xz) : p → r
λyz .y(xz) : (q → r) → (p → r)
λxyz .y(xz) : (p → q) → ((q → r) → (p → r))

The Curry-Howard correspondence acts at three different levels. The formulae

correspond to types, the derivations correspond to terms and the normalization of
derivations (see [6, II.3]) corresponds to βη-reduction in simply-typed λ-calculus.
(Roughly speaking, the last two axiom schemata for simply-typed λ calculus, di-
rected from left to right, underly βη-reduction.)
This correspondence is very important from the point of view of general proof
theory. It may serve to formalize the notion of equality of derivations based on
normalization. However, this correspondence is not in the main stream of these
notes. It serves just as an alternative approach to what we intend to do with
sequent systems and categories.
Another way to formalize natural deduction is to introduce sequents, i.e., words
of the form Γ ` A, where Γ is a set of formulae staying for the hypotheses of a
derivation of a formula A. Such a formal system, which is appropriate for our
fragment of logic, is given by the axiom scheme {A} ` A and the inference figures

Γ`B Γ`B Γ`A→B ∆`A

.
Γ`A→B Γ − {A} ` A → B Γ∪∆`B

This motivates the introduction of sequent systems.

The following sequent system is based on Gentzen’s LJ (see [3]). The cut-
elimination theorem that holds for LJ and LK is strong enough to deliver some
very important properties of intuitionistic and classical logic. This gives to sequent
systems an advantage over natural deduction systems.
We envisage a formal system whose alphabet extends the one introduced in
Section 2 by the symbols ` and ,. The language consists of sequents of the form
Γ ` A, where Γ is a finite sequence of formulae, and A is a formula. The sequence Γ,
which may be empty, is the antecedent of the sequent and consists of the antecedent
formulae, while A is the succedent formula of the sequent. The axiom scheme is
A ` A and the inference figures are the following structural inference figures

Γ, A, B, ∆ ` D
interchange
Γ, B, A, ∆ ` D
 270 MINUTES ON CATEGORIAL PROOF THEORY 7

Γ`D A, A, Γ ` D
thinning contraction
A, Γ ` D A, Γ ` D

Γ`A ∆, A, Θ ` D
cut
∆, Γ, Θ ` D
and the following operational inference figures
Γ`A B, ∆ ` D A, Γ ` D
→` `→
Γ, A → B, ∆ ` D Γ`A→D
This formal system is such that we derive sequents in it. However, this system
may serve to define theorems of the corresponding fragment of logic. A formula A
is a theorem of the implicational fragment of intuitionistic logic when the sequent
` A is a theorem of this system.
Theorem 3.1. (Cut-Elimination Theorem) Every derivation can be trans-
formed into a derivation with the same endsequent and in which the cut inference
figure does not occur.
We prove here just a toy example of this theorem. The sequent system IL,
envisaged here, deals with the sequents of the form G ` A, where G and A are
formulae of the strict variant of the mixed language (see Section 2). The reason
for switching to a single premise-single conclusion language is to get closer to the
categorial setting. This system corresponds to the multiplicative conjunction-linear
implication fragment of intuitionistic linear logic. Besides the axiomatic sequents
A ` A, there are only two structural inference figures
G∧A∧B∧E `D C`A G∧A∧E `D
interchange cut
G∧B∧A∧E `D G∧C ∧E `D
and three operational inference figures
C`A B∧G`D A∧G`C A`C B`E
→` `→ ∧`∧
C ∧ (A → B) ∧ G ` D G`A→C A∧B `C ∧E
The formula (p → q) → ((q → r) → (p → r)) is its theorem.
q`q r`r
p`p q ∧ (q → r) ` r
p ∧ (p → q) ∧ (q → r) ` r
(p → q) ∧ (q → r) ` p → r
(q → r) ∧ (p → q) ` (p → r)
p → q ` (q → r) → (p → r)
> ` (p → q) → ((q → r) → (p → r))
8 ZORAN PETRIĆ

However, this system is rich enough to justify some basic categorial notions through
the cut-elimination procedure.
By a formula in a derivation we always mean a particular occurrence of this
formula as an antecedent or a succedent formula in this derivation. The sequent
C ` A is the left-premise and G∧A∧E ` D is the right-premise of the cut inference
figure.
For every inference figure except `→ and ∧ ` ∧, every succedent D of the lower
sequent has the unique successor, an occurrence of the same formula D as the
succedent of an upper sequent. Also, every prime conjunct (one that is not of the
form of conjunction) of the antecedent of the lower sequent of an inference figure,
except A → B in →`, has the unique successor, an occurrence of the same formula
in the antecedent of an upper sequent. Let the rank of a succedent D of a sequent
in a derivation be the number of formulae of that derivation that are related to D
by the reflexive and transitive closure of the successor relation, and let the rank of
a prime conjunct of the antecedent of a sequent in a derivation be defined in the
same manner. If a conjunct of the antecedent of a sequent is not prime, its rank is
1. For example, the red p → q has rank 4 in the above derivation. Let the formula
A in the cut inference figure be called cut formula. Let the degree of a cut in a
derivation be the number of occurrences of implication and conjunction in the cut
formula A. Let the rank of a cut in a derivation be the sum of the rank of the cut
formula in the left premise and the rank of cut formula in the right premise of this
cut inference figure.

Proof of Theorem 3.1. It suffices to show that this theorem holds for a derivation
whose last inference figure is cut and there is no other application of cut in the
derivation. We proceed by induction on lexicographically ordered pairs (d, r), where
d is the degree and r is the rank of the cut in such a derivation.
For the basis, when (d, r) = (0, 2), the derivation is of the form

p`p p`p
(3.1)
p`p

and we transform it into the derivation consisting only of the axiomatic sequent
p ` p.
When d > 0 and r = 2 we have the following cases. The derivations of the
following forms

D D
A`A G∧A∧E `D C`A A`A
(3.2)
G∧A∧E `D C`A

are transformed, respectively, into the following cut-free derivations

D D
G∧A∧E `D C`A
 270 MINUTES ON CATEGORIAL PROOF THEORY 9

If the derivation is of the form

D1 D2 D3
A1 ∧ C ` A2 G ` A1 A2 ∧ E ` D
C ` A1 → A2 G ∧ (A1 → A2 ) ∧ E ` D
(3.3)
G∧C ∧E `D
it is transformed into the following derivation.
D1 D3
D2 A1 ∧ C ` A2 A2 ∧ E ` D
G ` A1 A1 ∧ C ∧ E ` D
G∧C ∧E `D
Finally, if the derivation is of the form
D1 D2
C 1 ` A1 C 2 ` A2 D3
C1 ∧ C2 ` A1 ∧ A2 G ∧ A1 ∧ A2 ∧ E ` D
(3.4)
G ∧ C1 ∧ C2 ∧ E ` D
it is transformed into the following derivation.
D1 D3
D2 C1 ` A1 G ∧ A1 ∧ A2 ∧ E ` D
C2 ` A2 G ∧ C 1 ∧ A2 ∧ E ` D
G ∧ C1 ∧ C2 ∧ E ` D
The two cuts in the transformed derivations for (3.3) and (3.4) are of the lower
degree. Hence, by the induction hypothesis, we can first eliminate the upper cuts
to obtain cut-free derivations. Then, again by the induction hypothesis, the lower
cuts may be eliminated.
When r > 2, either the derivation of the left premise of the cut ends with one
of the following inference figures
G1 ∧ B ∧ C ∧ G2 ` A G1 ` B C ∧ G2 ` A
(3.5)
G1 ∧ C ∧ B ∧ G2 ` A G1 ∧ (B → C) ∧ G2 ` A
or, for A being a prime conjunct, the derivation of the right premise of the cut ends
with one of the following inference figures.
G0 ∧ A ∧ E 0 ` D F1 ` F2 F3 ` D
(3.6) interchange →`
G∧A∧E `D G∧A∧E `D
10 ZORAN PETRIĆ

D1 ∧ E ∧ A ∧ F ` D2 C1 ` D1 C2 ` D2
(3.7) `→ ∧`∧
E ∧ A ∧ F ` D1 → D2 E ∧ A ∧ F ` D1 ∧ D2

It is obvious how to permute the cut with the above inference figures in order to
decrease its rank by 1. For example, a derivation of the form

D2
D1 E ∧ A ∧ F1 ∧ B ∧ C ∧ F2 ` D
G`A E ∧ A ∧ F1 ∧ C ∧ B ∧ F2 ` D
E ∧ G ∧ F1 ∧ C ∧ B ∧ F2 ` D

is transformed into the following derivation.

D1 D2
G`A E ∧ A ∧ F1 ∧ B ∧ C ∧ F2 ` D
E ∧ G ∧ F1 ∧ B ∧ C ∧ F2 ` D
E ∧ G ∧ F1 ∧ C ∧ B ∧ F2 ` D

By the induction hypothesis the derivation ending with the lifted cut may be
transformed into a cut-free derivation. 

In order to make one step further toward categories, let us consider the following
sequent system ILcat given in the same language as IL. The axiomatic sequents
of the system are A ` A, B ∧ A ` A ∧ B, A ∧ (A → B) ` B and B ` A → (A ∧ B)
and the rules of inference are:

A`B B`C A`C B`E B`C

composition ∧`∧ →`→
A`C A∧B `C ∧E A→B`A→C

Proposition 3.1. The systems IL and ILcat are equivalent in the sense that they
have the same derivable sequents.

Proof. First we show that the inference rules of IL are derivable in ILcat. For the
interchange we have

B∧A`A∧B
G∧B∧A`G∧A∧B
G∧B∧A∧E `G∧A∧B∧E G∧A∧B∧E `D
G∧B∧A∧E `D
 270 MINUTES ON CATEGORIAL PROOF THEORY 11

For the cut rule, we have the following derivation

C`A
G∧C `G∧A
G∧C ∧E `G∧A∧E G∧A∧E `D
G∧C ∧E `D
and for the →`-rule we obtain
A ∧ (A → B) ` B
C`A A ∧ (A → B) ∧ G ` B ∧ G B ∧ G ` D
C ∧ (A → B) ∧ G ` A ∧ (A → B) ∧ G A ∧ (A → B) ∧ G ` D
C ∧ (A → B) ∧ G ` D
Finally, for the `→-rule, we have the following derivation
A∧G`C
G ` A → (A ∧ G) A → (A ∧ G) ` A → C
G`A→C
Now we show that the axioms and inference rules of ILcat are derivable in IL.
Derivations for the axioms are
A∧B `A∧B A`A B`B A∧B `A∧B
, , ,
B∧A`A∧B A ∧ (A → B) ` B B ` A → (A ∧ B)
while the only nontrivial derivation for the inference rule is
A`A B`C
A ∧ (A → B) ` C
,
A→B`A→C
which completes the proof. 

4. From sequent systems to categories

Based on the strict variant of mixed language, we build a new language whose
members are equalities of the form t = s where t and s are terms defined below.
Some of these terms represent IL derivations. The alphabet introduced in Section 2
is extended by the symbols 1, c, η, ε, ◦, =, :, ` and ,.
The terms of this language are defined inductively as follows.
(1) If A and B are formulae, then
1A : A ` A, cB,A : B ∧ A ` A ∧ B,
εA,B : A ∧ (A → B) ` B, ηA,B : B ` A → (A ∧ B)
are terms called primitive.
12 ZORAN PETRIĆ

(2) If f : A ` B and g : B ` C are terms, then (g ◦ f ) : A ` C is a term.

(3) If f1 : A1 ` B1 and f2 : A2 ` B2 are terms, then (f1 ∧ f2 ) : A1 ∧ A2 ` B1 ∧ B2
is a term.
(4) If f : B1 ` B2 is a term and A is a formula, then (A → f ) : A → B1 ` A → B2
is a term.
(5) Nothing else is a term.

A type is a word of the form A ` B where A and B are formulae. We say that
A ` B is the type of a term f : A ` B and we say that this term has A as the
source and B as the target. Sometimes, the subscripts determined by the context
are omitted. Usually, we omit the type in writing a term and by term we mean
just the part before the symbol “:”. We omit the outermost parentheses in terms
taking them for granted.
Every IL derivation of a sequent corresponds to a term whose type is that
sequent. A derivation consisting solely of an axiomatic sequent A ` A corresponds
to 1A : A ` A. A derivation ending with an inference figure corresponds to a term
obtained from the terms corresponding to the derivations of the premises as follows:
f:G∧A∧B∧E `D
interchange
f ◦ ((1G ∧ cB,A ) ∧ 1E ) : G ∧ B ∧ A ∧ E ` D
f:G`A g: E ∧ A ∧ F ` D
cut
g ◦ ((1E ∧ f ) ∧ 1F ) : E ∧ G ∧ F ` D
f:G`A g: B ∧ E ` D
→`
g ◦ ((εA,B ∧ 1E ) ◦ ((f ∧ 1A→B ) ∧ 1E )) : G ∧ (A → B) ∧ E ` D
f:A∧G`D
`→
(A → f ) ◦ ηA,G : G ` A → D
For example, the derivation
p`p q`q
p ∧ (p → q) ` q
(p → q) ∧ p ` q
p ` (p → q) → q
> ` p → ((p → q) → q)
corresponds to the term
(p → (((p → q) → (1q ◦ (εp,q ◦ (1p ∧ 1p→q )) ◦ cp→q,p )) ◦ ηp→q,p )) ◦ ηp,>
whose type is > ` p → ((p → q) → q).
The language consists of words of the form f = g, where f and g are terms with
the same type. Our goal is to define an equational system E in that language, whose
theorems cover the cut-elimination procedure. This means that if a derivation
 270 MINUTES ON CATEGORIAL PROOF THEORY 13

corresponding to a term f is transformed by the cut-elimination procedure into a

derivation corresponding to a term g, then f = g is a theorem of E.
The axiom schemata include the following
f = f, f ∧ 1> = f = 1> ∧ f, (f ∧ g) ∧ h = f ∧ (g ∧ h).
(The other axioms will appear through our analysis.) The inference figures are the
following
f =g f =g g=h
g=f f =h
f1 : A ` B = f2 : A ` B g1 : B ` C = g2 : B ` C
g1 ◦ f1 = g2 ◦ f2
f1 = f2 g1 = g2 f1 = f2
f1 ∧ g1 = f2 ∧ g2 A → f1 = A → f2 .

We start our analysis with two simple instances of Case (3.2) from the proof of
Theorem 3.1, which cover also Case (3.1). A derivation of the form
1A : A ` A g: A ` D
g ◦ 1A : A ` D
is transformed into
g : A ` D.
Also, a derivation of the form
f:G`A 1A : A ` A
1A ◦ f : G ` A
is transformed into
f : G ` A.
Hence, we add to E the following axiom schemata
(4.1) g ◦ 1A = g and 1A ◦ f = f.
We always assume that both sides of our equalities are terms, which, for (4.1),
means that g has A as the source and f has A as the target.
Consider the following instance of Case (3.5) in which the derivation
f:B∧C `A
f ◦ c: C ∧ B ` A g: A ` D
g ◦ (f ◦ c) : C ∧ B ` D
is transformed into
f:B∧C `A g: A ` D
g◦f:B∧C `D
(g ◦ f ) ◦ c : C ∧ B ` D
14 ZORAN PETRIĆ

We see that the axiom scheme

(4.2) h ◦ (g ◦ f ) = (h ◦ g) ◦ f
should be added to E. Let ≡ be the relation on the set of terms defined by
f ≡g when f = g is a theorem of E.
This is an equivalence relation. Let [f ] be the equivalence class of a term f . We
call [f ] an arrow. It is straightforward to check that the following definitions are
correct. The source of [f ] is the source of f , the target of [f ] is the target of f , the
identity arrow on A is [1A ] and, for a term g ◦ f , the composition [g] ◦ [f ] of [g]
with [f ] is [g ◦ f ].

Categories
A category consists of two sets, O of objects and A of arrows, two functions
source, target : A → O,
and two additional functions
1 : O → A, ◦ : A ×O A → A
called identity and composition, where
A ×O A =df {(g, f ) | g, f ∈ A & source(g) = target(f )}
is the set of all composable pairs of arrows. Moreover, for every X, Y ∈ O,
and every f, g, h ∈ A such that (h, g), (g, f ) and (1Y , f ) are composable
pairs of arrows, the following holds
source(1X ) = X = target(1X ),
source(g ◦ f ) = source(f ), target(g ◦ f ) = target(g),
and
g ◦ 1Y = g, 1Y ◦ f = f, h ◦ (g ◦ f ) = (h ◦ g) ◦ f.

The set OK , which is the set of formulae, and the set AK = {[f ] | f is a term},
with the source and target function, identities and composition defined as above,
make the category K. In the sequel we denote an arrow [f ] just by f . We will
delete occurrences of 1 in the immediate scope of ◦ and omit parentheses tied to ◦
in the immediate scope of ◦ without referring to (4.1) and (4.2).

Consider now a bit more complicated instance of Case (3.2) where

1A : A ` A g: A ∧ B ` D
g ◦ (1A ∧ 1B ) : A ∧ B ` D
is transformed into
g : A ∧ B ` D.
If we add the axiom scheme
(4.3) 1A ∧ 1B = 1A∧B ,
 270 MINUTES ON CATEGORIAL PROOF THEORY 15

to E, then with the help of (4.1), we easily derive g ◦ (1A ∧ 1B ) = g.

According to an instance of Case (3.6), the derivation

g: A ∧ B ∧ C ` D
f:G`A g ◦ (1A ∧ c) : A ∧ C ∧ B ` D
g ◦ (1A ∧ c) ◦ (f ∧ 1C∧B ) : G ∧ C ∧ B ` D
is transformed into
f:G`A g: A ∧ B ∧ C ` D
g ◦ (f ∧ 1B∧C ) : G ∧ B ∧ C ` D
g ◦ (f ∧ 1B∧C ) ◦ (1G ∧ c) : G ∧ C ∧ B ` D
If we add the axiom scheme
(4.4) (g1 ∧ g2 ) ◦ (f1 ∧ f2 ) = (g1 ◦ f1 ) ∧ (g2 ◦ f2 )
to E, then we have
g ◦ (1A ∧ c) ◦ (f ∧ 1C∧B ) = g ◦ ((1A ◦ f ) ∧ (c ◦ 1C∧B )), by (4.4)

= g ◦ ((f ◦ 1G ) ∧ (1B∧C ◦ c)), by (4.1)

= g ◦ (f ∧ 1B∧C ) ◦ (1G ∧ c), by (4.4)

Functors
Given two categories C and D, a functor F : C → D consists of two
functions, both denoted by F , the object function F : OC → OD and the
arrow function F : AC → AD , such that for every C ∈ OC and every
composable pair (g, f ) of arrows of C
F 1C = 1F C , F (g ◦ f ) = F g ◦ F f.
The identity functor on a category consists of the identity function on
objects and the identity function on arrows. The operation of composi-
tion of two functors consists of two compositions—composition of object
functions and composition of arrow functions, hence this operation is as-
sociative.

Product of categories
The product C × D, of categories C and D is the category whose objects
make the cartesian product OC ×OD and whose arrows make the cartesian
product AC × AD . The identity arrow on (C, D) is the pair (1C , 1D ) and
composition is defined componentwise.

If we define ∧(A, B) as A ∧ B and ∧(f, g) as f ∧ g, then (4.3) and (4.4) guarantee

that ∧ is a functor from K × K to K. (Note that the latter definition is correct
since f ≡ f 0 and g ≡ g 0 implies (f ∧ g) ≡ (f 0 ∧ g 0 ).)
16 ZORAN PETRIĆ

According to another instance of Case (3.6), the derivation

g: B ∧ A ` D
f:G`A g ◦ c: A ∧ B ` D
g ◦ c ◦ (f ∧ 1B ) : G ∧ B ` D
is transformed into
f:G`A g: B ∧ A ` D
g ◦ (1B ∧ f ) : B ∧ G ` D
g ◦ (1B ∧ f ) ◦ c : G ∧ B ` D
If we add the axiom scheme
(4.5) cA0 ,B 0 ◦ (f ∧ g) = (g ∧ f ) ◦ cA,B
to E, then with the help of (4.2) we easily derive g ◦ c ◦ (f ∧ 1B ) = g ◦ (1B ∧ f ) ◦ c.

Natural transformations
.
Given two functors F, G : C → D, a natural transformation α : F → G
is a function from OC to AD , i.e., a family of arrows of D indexed by
the objects of C, such that for every C ∈ OC , source(αC ) = F C and
target(αC ) = GC, and for every f : C → C 0 ∈ AC , the following diagram
α
F C −−−C−→ GC
 

F fy
Gf
y
F C 0 −−−−→ GC 0
αC 0

commutes in D. The arrows αC for C ∈ OC , are the components of the

natural transformation α.
Let 2 be the category with two objects, 0 and 1, and one nonidentity
arrow h : 0 → 1. Let I0 , I1 : C → C×2 be functors such that for every object
C and every arrow f of C, we have that I0 (C) = (C, 0), I0 (f ) = (f, 10 ),
and I1 (C) = (C, 1), I1 (f ) = (f, 11 ). Let F, G : C → D be two functors.
.
There is a bijection between the set of natural transformations α : F → G,
and the set of functors A : C × 2 → D such that A ◦ I0 = F and A ◦ I1 = G.
.
This bijection maps α : F → G to A : C × 2 → D defined by
A(C, 0) = F C, A(C, 1) = GC, A(f, 10 ) = F f, A(f, 11 ) = Gf,
0
and for f : C → C ,
A(f, h) = Gf ◦ αC = αC 0 ◦ F f.
.
Its inverse maps A : C × 2 → D to α : F → G defined by αC = A(1C , h).

The scheme (4.5) says that the family indexed by the objects of K × K
c = {cA,B | A, B ∈ Ob(K)}
 270 MINUTES ON CATEGORIAL PROOF THEORY 17

is a natural transformation from the above defined functor ∧ : K × K → K to the

functor G : K × K → K defined so that
G(A, B) = B ∧ A and G(f, g) = g ∧ f.
Isomorphisms, natural isomorphisms
An arrow f : A → B of a category C is an isomorphism when there is an
arrow g : B → A in C, such that g ◦ f = 1A and f ◦ g = 1B in C. A
natural transformation is a natural isomorphism when all its components
are isomorphisms.

The following two schemata say that c is a natural isomorphism, which satisfies
a coherence condition.
(4.6) cB,A ◦ cA,B = 1A∧B

(4.7) cA∧B,C = (cA,C ∧ 1B ) ◦ (1A ∧ cB,C )

According to an instance of Case (3.6), the derivation
g : D1 ∧ A ` D2
f:G`A (D1 → g) ◦ η : A ` D1 → D2
(D1 → g) ◦ η ◦ f : G ` D1 → D2
is transformed into
f:G`A g : D1 ∧ A ` D2
g ◦ (1 ∧ f ) : D1 ∧ G ` D2
(D1 → (g ◦ (1 ∧ f ))) ◦ η : G ` D1 → D2
If we add the axiom schemata
(4.8) A → (g ◦ f ) = (A → g) ◦ (A → f )

(4.9) ηA,B 0 ◦ f = (A → (1A ∧ f )) ◦ ηA,B

to E, then we have
(D1 → g) ◦ η ◦ f = (D1 → g) ◦ (D1 → (1 ∧ f )) ◦ η by (4.9)
= (D1 → (g ◦ (1 ∧ f ))) ◦ η by (4.8)
The scheme (4.8) together with the scheme
(4.10) A → 1B = 1A→B
guarantee that for every A, we have the functor A → : K → K, mapping B to
A → B and f to A → f .
It is not difficult to see how to use (4.1), (4.3), (4.4), (4.10) and (4.8) in order
to show that G : K → K defined by
G(B) = A → (A ∧ B) and G(f ) = A → (1A ∧ f )
18 ZORAN PETRIĆ

is indeed a functor. The scheme (4.9) says that for every A ∈ Ob(K), the family

ηA = {ηA,B | B ∈ Ob(K)}

is a natural transformation from the identity functor on K to G.

According to an instance of Case (3.3), the derivation

f : A1 ∧ G ` A2 g : B ` A1 h : A2 ` D
(A1 → f ) ◦ η : G ` A1 → A2 h ◦ ε ◦ (g ∧ 1) : B ∧ (A1 → A2 ) ` D
h ◦ ε ◦ (g ∧ 1) ◦ (1 ∧ ((A1 → f ) ◦ η)) : B ∧ G ` D

is transformed into

f : A1 ∧ G ` A2 h : A2 ` D
g : B ` A1 h ◦ f : A1 ∧ G ` D
h ◦ f ◦ (g ∧ 1) : B ∧ G ` D

If we add the axiom schemata

(4.11) εA,B 0 ◦ (1A ∧ (A → f )) = f ◦ εA,B

(4.12) εA,A∧B ◦ (1A ∧ ηA,B ) = 1A∧B

to E, then we have
h ◦ ε ◦ (g ∧ 1) ◦ (1 ∧ ((A1 → f ) ◦ η)) =
= h ◦ ε ◦ (g ∧ 1) ◦ ((1 ◦ 1) ∧ ((A1 → f ) ◦ η)) by (4.1)
= h ◦ ε ◦ (g ∧ 1) ◦ (1 ∧ (A1 → f )) ◦ (1 ∧ η) by (4.4)
= h ◦ ε ◦ (1 ∧ (A1 → f )) ◦ (g ∧ 1) ◦ (1 ∧ η) by (4.1), (4.4)
= h ◦ f ◦ ε ◦ (g ∧ 1) ◦ (1 ∧ η) by (4.11)
= h ◦ f ◦ ε ◦ (1 ∧ η) ◦ (g ∧ 1) by (4.1), (4.4)
= h ◦ f ◦ (g ∧ 1) by (4.12)
The scheme (4.11) says that for every A ∈ Ob(K), the family

εA = {εA,B | B ∈ Ob(K)}

is a natural transformation from the functor F : K → K defined so that

F (B) = A ∧ (A → B) and F (f ) = 1A ∧ (A → f )

to the identity functor on K.

 270 MINUTES ON CATEGORIAL PROOF THEORY 19

Adjunction
Given two categories C and D, an adjunction is given by two functors,
F : C → D and G : D → C, and two natural transformations, the unit
. .
η : 1C → GF and the counit ε : F G → 1D , such that for every C ∈ OC
and every D ∈ OD
GεD ◦ ηGD = 1GD , and εF C ◦ F ηC = 1F C .
These two equalities are called triangular identities. The functor F is a
left adjoint for the functor G, while G is a right adjoint for the functor F .

The schemata (4.9), (4.11), (4.12) together with

(4.13) (A → εA,B ) ◦ ηA,A→B = 1A→B

say that for every A, the functor A ∧ : K → K, mapping B to A ∧ B and f to

1A ∧ f , is a left adjoint for the functor A → : K → K. The unit of this adjunction
is the natural transformation ηA while the counit is the natural transformation εA .
The schemata (4.12) and (4.13) are the triangular identities for this adjunction.

Symmetric monoidal closed categories

A symmetric monoidal closed category is a category C with a functor
⊗ : C × C → C, an object I, and for every object A a functor A : C → C
such that the following holds:
1. There are three natural isomorphisms with components
αA,B,C : A⊗(B ⊗C) → (A⊗B)⊗C, λA : I ⊗A → A, γA,B : A⊗B → B ⊗A.
2. The natural transformation γ is self-inverse, i.e., γB,A ◦ γA,B = 1A⊗B .
A
3. For every object A, the functor A ⊗ is a left adjoint for .
4. The following diagrams commute (coherence conditions).

α α
A ⊗ (B ⊗ (C ⊗ D)) (A ⊗ B) ⊗ (C ⊗ D) ((A ⊗ B) ⊗ C) ⊗ D
1⊗α
α⊗1
A ⊗ ((B ⊗ C) ⊗ D) α (A ⊗ (B ⊗ C)) ⊗ D

α γ
(I ⊗ A) ⊗ B A ⊗ (B ⊗ C) (A ⊗ B) ⊗ C C ⊗ (A ⊗ B)
λ⊗1
α 1⊗γ α

I ⊗ (A ⊗ B) λ
A⊗B A ⊗ (C ⊗ B) α (A ⊗ C) ⊗ B γ⊗1
(C ⊗ A) ⊗ B

A symmetric monoidal closed category is symmetric strict monoidal closed

when all the components of α and λ are identities.
20 ZORAN PETRIĆ

The category K whose objects are formulae and whose arrows are obtained via
the equational system E, axiomatized by

f = f, f ∧ 1> = f = 1> ∧ f, (f ∧ g) ∧ h = f ∧ (g ∧ h)

and (4.1)-(4.13), is a symmetric strict monoidal closed category in which ⊗ is ∧,

I is >, A is A → , and γ is c. Moreover, it is a category of that kind freely
generated by the set of propositional letters that belong to the alphabet. This means
that every function from the set of propositional letters to the set of objects of a
symmetric strict monoidal closed category C extends in a unique way to a functor
from K to C, which preserves the symmetric strict monoidal closed structure.
It is not difficult to show that E covers the cut-elimination procedure in the sense
mentioned just before we started to define this equational system. We leave this as
an exercise for the reader. Not all the schemata (4.1)-(4.13) are necessary for that.
For example, there is no need for (4.7), (4.10) and (4.13). However, we keep these
schemata not only for aesthetic or purely category-theoretic reasons but also for
some other proof-theoretic needs. By using (4.7), the interchange of a formula with
a sequence may be atomized, and if only atomic axiomatic sequents are allowed in
IL, then (4.13) covers the expansion transforming 1A→B into

1A : A ` A 1B : B ` B
εA,B : A ∧ (A → B) ` B
(A → εA,B ) ◦ ηA,A→B : A → B ` A → B

For C and f : A ` B from K, we define f → C : B → C ` A → C as

(A → (εB,C ◦ (f ∧ 1B→C ))) ◦ ηA,B→C .

From (4.3) and (4.13) it follows that

(4.14) 1A → C = 1A→C .

Also, with the help of (4.9), (4.11), (4.4), (4.8) and (4.12) one can derive

(4.15) (f → C) ◦ (g → C) = (g ◦ f ) → C.

Hence, for every object C of K we have the functor → C : Kop → K, mapping

A to A → C and f to f → C.

Opposite category
Given a category C, its opposite category C op consists of the same set of
objects and the same set of arrows. An arrow f of C envisaged as the
arrow of C op is denoted by f op . The functions source and target switch
the roles so that if f : A → B in C, then f op : B → A in C op . The identities
are the same, while the composition ◦ in C op is such that
f op ◦ g op = (g ◦ f )op .
 270 MINUTES ON CATEGORIAL PROOF THEORY 21

5. The classification of formulae

The main problem for a mathematician working in a particular category is to
classify the objects of that category up to isomorphism. This could be a hard
task, and in practise, for a lot of categories from the realm of algebra, geometry
and topology, just some partial results are known. For a proof theorist ready to
accept the program offered in the previous section, the classification of objects, i.e.,
formulae, is of similar importance. Isomorphic formulae may be considered to have
the same proof-theoretical meaning.
Isomorphism ∼ = between formulae is an equivalence relation which does not mean
only that one formula is derivable from the other and vice versa. Since we are in a
category, a formula A is isomorphic to a formula B when there are two derivations
f : A ` B and g : B ` A such that the compositions g ◦ f and f ◦ g are equal
respectively to the identity derivations 1A and 1B .
In this section we explain how to classify the objects of a relaxed version of the
category K. Based on the non-strict variant of mixed language, following the lines
of Section 4, the category Klax is built out of syntax material. The set of objects
of Klax is the set of formulae of this language. The main difference is that ∧ is
not associative on objects and that > is not the unit for ∧. However, we have the
isomorphisms

A ∧ (B ∧ C) ∼
= (A ∧ B) ∧ C, >∧A∼
=A∼
=A∧>

natural in A, B and C. Moreover, the three coherence conditions listed in Section 4

hold.
By an invariant of a category C we mean a function whose domain is the set
of objects of C having the same (isomorphic) value on isomorphic objects. For
example, the object function of a functor whose source is C is an invariant of C. By
a complete invariant of a category C we mean a family {Fi }i∈I of functions whose
domain is the set of objects of C such that for every A and B

A∼
= B in C iff for every i ∈ I, Fi A = Fi B (Fi A ∼
= Fi B).

The arrows of Klax do not go beyond derivations in classical logic, which means
that if f : A ` B is an arrow of Klax , then A → B is a tautology. Hence, if
A∼= B, then A ↔ B is a tautology, i.e., A and B correspond to the same Boolean
function. This leads to a simple logical invariant mapping every formula to the
corresponding Boolean function. By relying on this invariant we can show that
p ∧ q is not isomorphic to p in Klax , since p ∧ q ↔ p is not a tautology. On the
other hand, it leaves open the question whether p ∧ p and p are isomorphic in Klax .
22 ZORAN PETRIĆ

The category Set

Suppose there is a universe U satisfying:
(i) x ∈ u ∈ U implies x ∈ U ,
(ii) u ∈ U and v ∈ U imply {u, v}, (u, v), u × v ∈ U ,
(iii) x ∈ U implies P(x) ∈ U and ∪x ∈ U ,
(iv) ω ∈ U , where ω = {0, 1, 2, . . .} is the set of all finite ordinals,
(v) if f : a → b is a surjective function with a ∈ U and b ⊂ U , then b ∈ U .
Every element of U is called a small set.
For a fixed universe U , the category Set has U as the set of objects and
{(f, (u, v)) | f, u, v ∈ U and f ⊂ u × v is a function}
as the set of arrows. The source, target, identity and composition are
defined as expected.

The category Klax is a symmetric monoidal closed category freely generated by

the set of propositional letters that belong to the alphabet. This means that every
function from the set of propositional letters to the set of objects of a symmetric
monoidal closed category C extends in a unique way to a functor, which preserves
the symmetric monoidal closed structure, from Klax to C. The category Set is a
symmetric monoidal closed category with ⊗ being the cartesian product and B A
being the set of all functions from A to B.
Consider the functor from Klax to Set that extends a function mapping the
letter p to a two element set X. This functor maps p ∧ p to the four element set
X × X which is not isomorphic to X in Set. Hence, this functor is an invariant
witnessing that p ∧ p and p are not isomorphic in Klax .
These invariants give us just partial results concerning the classification of ob-
jects of Klax . In order to give the complete classification we introduce the following
invariant. Let S be the equational system in which we write ∼ = instead of =. The
theorems of S are of the form A ∼ = B, where A and B are formulae of non-strict
mixed language. The axiom schemata are,
C= ∼C >∧C = ∼ C,
A ∧ (B ∧ C) ∼
= (A ∧ B) ∧ C, A∧B ∼
= B ∧ A,
(A ∧ B) → C =∼ B → (A → C), >→C= ∼ C,
and we have the following inference figures

A∼
=B A∼
=B B∼ =C
B∼ ∼
A=C
=A

A∼
=B C∼
=D A∼
=B C∼
=D
.
A∧C ∼
=B∧D B→C∼
=A→D
The six axiom schemata are such that for arbitrary A, B and C, these isomor-
phisms hold in Klax . For the first four isomorphisms this is trivial, and the last
 270 MINUTES ON CATEGORIAL PROOF THEORY 23

two are left as an exercise for the reader. The first two inference figures are jus-
tified by the facts that the inverse of an isomorphism is an isomorphism and that
the composition of isomorphisms is an isomorphism. The last two inference figures
are justified by the functoriality of ∧ and → (see (4.3), (4.4), (4.10), (4.8), (4.14),
(4.15)). This suffices to conclude the following.
Lemma 5.1. If A ∼
= B is derivable in S, then A ∼
= B holds in Klax .
The category FinSet*, whose objects are the finite small sets with a selected
base point and whose arrows are base-point-preserving functions, is a symmetric
monoidal closed category. The functor ⊗ is defined on a pair of objects (X, Y ) as
((X − {∗X }) × (Y − {∗Y })) ∪ {∗X⊗Y }.
Any two element set with a base point may serve as the object I, i.e., a neutral (up
to isomorphism) for ⊗. The functor X is defined so that Y X is the set of base-
point-preserving functions from X to Y , whose base point is the constant function,
mapping every element of X to the base point of Y . The definition of ⊗ and X
on arrows is then straightforward.
Let N+ be the set of positive natural numbers and let g be a function, called
valuation, from the set of propositional letters to N+ . Consider the following binary
operations on N+
m  n =df (m − 1)(n − 1) + 1, mn =df mn−1 .
We define the function ḡ, the extension of g, that maps the objects of Klax to
N+ inductively as follows
ḡ(p) = g(p), ḡ(>) = 2,
ḡ(A ∧ B) = ḡ(A)  ḡ(B), ḡ(A → B) = ḡ(B)ḡ(A) .
For a valuation g, let ġ be a function from the set of propositional letters to
the set of objects of FinSet*, satisfying that the cardinality of ġ(p) is g(p). By
the universal property of Klax , there is a unique functor G from Klax to FinSet*
that extends the function ġ and preserves the symmetric monoidal closed structure.
Since every functor preserves isomorphisms, we have that A ∼ = B in Klax implies
GA ∼ = GB in FinSet*.
Two objects of FinSet* are isomorphic when they have the same cardinality. It
is easy to verify that the cardinality of GA is ḡ(A), for ḡ being the extension of g.
Hence, for every valuation g, its extension ḡ is an invariant of Klax in the sense
that if A ∼
= B in Klax , then ḡ(A) = ḡ(B).
We write N+ |= A ∼ = B when ḡ(A) = ḡ(B) holds for every valuation g. By the
preceding paragraph, we have the following result.
Lemma 5.2. If A ∼
= B holds in Klax , then N+ |= A ∼
= B.
From Lemmata 5.1 and 5.2 we have that
if A ∼
= B is derivable in S, then N+ |= A ∼
= B.
Whether the other direction of this implication holds is a question remained open
since it is formulated in [1, Section 9]. The positive answer to this question, together
24 ZORAN PETRIĆ

with Lemma 5.1, guarantees that the family {ḡ | g is a valuation} is a complete
invariant for C. However, for the classification of formulae, we need just a restricted
form of this implication, which is formulated below.
We say that a formula is diversified when every propositional letter occurs in it
no more than once. The following result stems from [1, Arithmetical Completeness
Theorem].
Lemma 5.3. If N+ |= A ∼ = B, for A, B diversified, then A ∼ = B is derivable in S.
From Lemmata 5.1, 5.2 and 5.3, for A and B diversified, we obtain the following
triangle of implications.

A∼
= B is derivable in S N+ |= A ∼
=B

A∼
= B holds in Klax
With the help of these implications we obtain a classification of objects of Klax .
For this we need the following notions. A formula A is an instance of a formula
A0 when for mutually distinct propositional letters p1 , . . . , pn and not necessarily
mutually distinct formulae B1 , . . . , Bn , the formula A is the result of uniformly
substituting the formula Bi for the letter pi .
Similarly, a term f : A ` B is an instance of a term f 0 : A0 ` B 0 when, for
p1 , . . . , pn and B1 , . . . , Bn as above, f is the result of uniformly substituting the
formula Bi for the letter pi in the indices of f 0 . For example, εp,p : p ∧ (p → p) ` p
is an instance of εp,q : p ∧ (p → q) ` q.
If f : A ` B is an instance of f 0 : A0 ` B 0 , then A and B are instances, by the
same substitution, of A0 and B 0 respectively.
Remark 5.1. It is easy to show that if A0 ∼ = B 0 is derivable in S and A and B
are instances, by the same substitution, of A0 and B 0 respectively, then A ∼ = B is
derivable in S.
The following result is taken over from [1, Diversification Lemma].
Lemma 5.4. For every isomorphism f : A ` B of Klax , there is an isomorphism
f 0 : A0 ` B 0 of Klax , such that f is an instance of f 0 , and A0 and B 0 are diversified.
As a corollary of Lemmata 5.1-5.4 we have the following.
Theorem 5.1. A ∼ = B is derivable in S iff A ∼
= B holds in Klax .
Proof. The direction from left to right is Lemma 5.1. For the other direction,
suppose that A ∼= B holds in Klax . By relying on Lemma 5.4, there are diversified
formulae A0 and B 0 such that A0 ∼
= B 0 holds in Klax and A and B are instances, by
the same substitution, of A and B 0 respectively. By Lemma 5.2, N+ |= A0 ∼
0
= B0,
0 ∼ 0
and by Lemma 5.3, A = B is derivable in S. By Remark 5.1, we conclude that
A∼ = B is derivable in S. 
The system S is decidable (see [1, Normal Form Lemma]), hence the relation ∼
=
is decidable. This completes the classification of objects of Klax .
 270 MINUTES ON CATEGORIAL PROOF THEORY 25

6. Categories with products, coherence

Coherence results serve to describe the canonical arrows of categories of a par-
ticular kind. For example, if symmetric monoidal closed categories are concerned,
then the canonical structure consists of all the arrows built out in terms of identities,
α, λ, γ, the units and counits of the adjunctions, and operations of composition,
⊗ and A . Usually, coherence provides a simple decision procedure for equality of
canonical arrows. We refer to [2] as a source of various coherence results.
The best way to formulate and to understand a coherence theorem is to use a
form of a standard logical completeness result. On the side of syntax, one has to
built a category of a desired kind out of syntactical material (like our categories
K and Klax ). All the arrows of such a category are canonical. On the side of
semantics, one has to find a manageable category of the same kind in which the
equality of arrows is easily decidable. Usually, the arrows of such a category are
some special finite relations, or some kind of diagrams representing them.

Full and faithful functors

A functor F : C → D is full when for every pair of objects C, C 0 of C and
for every arrow g : F C → F C 0 of D, there is an arrow f : C → C 0 of C with
g = F f . A functor F : C → D is faithful when for every pair f, g : A → B
of arrows of C, F (f ) = F (g) implies f = g.

The interpretation is now given by a functor F from the syntactical category

to the manageable category. The existence of such F corresponds to a soundness
result. The faithfulness of F corresponds to a completeness result.

Products
A product of two objects A and B in a category C consists of an object
1 2
A × B of C and a pair of arrows πA,B : A × B → A, πA,B : A×B → B
of C, and it is characterized by the following universal property: for every
pair of arrows f : C → A and g : C → B of C, there is a unique arrow
h : C → A × B of C such that
1 2
πA,B ◦h=f and πA,B ◦ h = g.

C
f g
h

A A×B B
π1 π2
26 ZORAN PETRIĆ

Categories with products

A category is a category with binary products when every pair of its objects
has a product in this category. In another words, C is a category with
binary products when for every pair (A, B) of its objects, there is an object
1 2
A × B and a pair of arrows πA,B : A × B → A and πA,B : A × B → B.
Moreover, for every pair of arrows f : C → A and g : C → B, there is an
arrow hf, gi : C → A × B such that the following holds:
1 2
(6.1) πA,B ◦ hf, gi = f, πA,B ◦ hf, gi = g,
1 2
(6.2) hπA,B ◦ h, πA,B ◦ hi = h.
It is easy to conclude that a category with binary products has n-ary
products, which are defined analogously, for every n > 1 (cf. [5, III.5,
Proposition 1]). We call these categories shortly categories with products.
Alternatively, this notion may be introduced as follows. The diagonal
functor ∆ : C → C × C for a category C is defined so that for every C ∈ OC
and every f ∈ AC , ∆(C) = (C, C) and ∆(f ) = (f, f ).
If there is a functor × : C×C → C, a right adjoint for ∆, then C is a category
with products. The unit of this adjunction is the natural transformation
from the identity functor on C to the composition × ◦ ∆
w = {wA : A → A × A | A ∈ OC },
while the counit is the natural transformation from the composition ∆ ◦ ×
to the identity functor on C × C
(π 1 , π 2 ) = {(πA,B
1 2
: A × B → A , πA,B : A × B → B) | A, B ∈ OC }.
The equivalence of these two definitions is left as an exercise for the reader.

If categories with products are in question, then the canonical structure consists
of all the arrows built out in terms of identities, projections, composition and the
operation h , i of pairing.
The syntactical category L is obtained in the same manner as we have obtained
the category K in Section 4. The objects of L are the formulae of the conjunction
language introduced in Section 2. Based on that language, we build a new language
whose members are equalities of the form t = s where t and s are terms defined
below. The alphabet introduced in Section 2 is extended by the symbols 1, π 1 , π 2 ,
h, i, ◦, =, :, ` and ,. (Note that π 1 , as well as π 2 , is considered as one symbol.)
The terms of this language are defined inductively as follows.
(1) If A and B are formulae, then
1A : A ` A
1 2
πA,B : A ∧ B ` A, πA,B : A ∧ B ` B,
are terms called primitive.

(2) If f : A ` B and g : B ` C are terms, then (g ◦ f ) : A ` C is a term.

 270 MINUTES ON CATEGORIAL PROOF THEORY 27

(3) If f : C ` A and g : C ` B is a term, then hf, gi : C ` A ∧ B is a term.

(4) Nothing else is a term.

We define the type, the source and the target of a term, and use the same
conventions, as in Section 4. The language consists of the words of the form f = g,
where f and g are terms with the same type. We define an equational system E in
this language in the same manner as in Section 4.
The axiomatic equalities are given by the following schemata
f = f, f ◦ 1A = f = 1B ◦ f, (h ◦ g) ◦ f = h ◦ (g ◦ f )
together with (6.1) and (6.2). The inference figures are
f =g f =g g=h
g=f f =h

f1 : A ` B = f2 : A ` B g1 : B ` C = g2 : B ` C
g1 ◦ f1 = g2 ◦ f2
f1 : C ` A = f2 : C ` A g1 : C ` B = g2 : C ` B
hf1 , g1 i = hf2 , g2 i
Let ≡ be the relation on the set of terms defined by
f ≡g when f = g is derivable in E.
Again, this is an equivalence relation. We define the arrows of L, their sources and
targets, composition, identities and pairing as for the category K in Section 4. This
completes the definition of the syntactical category L. This category corresponds
to the conjunction fragment of both intuitionistic and classical logic. The arrows
wA (the components of the unit of the adjunction between ∆ and ×) are tied to
1 2
the structural rule of contraction and the arrows πA,B and πA,B , called projections,
are tied to the structural rule of thinning.
The category L is a category with products. Moreover, it is a category of that
kind freely generated by the set of propositional letters that belong to the alphabet.
This means that every function from the set of propositional letters to the set of
objects of a category C with products, extends in a unique way to a functor, which
preserves the product structure, from L to C.
Since we have the scheme (4.2), we may omit the parentheses tied to a com-
position being in the immediate scope of another composition. Hence we write
fn ◦ . . . ◦ f1 and reconstruct the parentheses in an arbitrary way. The terms in
normal form are defined inductively as follows. For a propositional letter p, a term
of the form 1p : p ` p and the term of the form fn ◦ . . . ◦ f1 : A ` p, where each fi is
either π 1 or π 2 is in normal form. If f and g are in normal form, then hf, gi is in
normal form.
Theorem 6.1. (Normal-Form Theorem) For every term f there exists a term
f 0 in normal form such that f ≡ f 0 .
28 ZORAN PETRIĆ

Proof. We proceed by induction on the complexity of the target of f .

Let the target of f be a letter p. We may assume that f is of the form fn ◦. . .◦f1 ,
where each fi is not of the form h ◦ g, and if n > 1, then each fi is not of the form
1
1A . If n = 1, then f is either of the form 1p , or of the form πp,A or of the form
2
πA,p and we are done.
If n > 1, then we start a new induction on the number of symbols h in f . If h
does not occur in f , then we are done. Note that if h occurs in fi , then fi is of the
form hg, hi and if h does not occur in fi , then fi is of the form π 1 or π 2 . Since the
target of f is p, we have that h does not occur in fn . For the greatest i such that h
occurs in fi we have that fi+1 is of the form π 1 or π 2 . In either case, by (6.1), f is
equal to a term of the same form with one less occurrence of h and we may apply
the induction hypothesis. This concludes the basis of the first induction.
If the target of f is of the form A ∧ B, then by (6.2) we have that f is equal to
1 2 1
hπA,B ◦ f, πA,B ◦ f i and we may apply the induction hypothesis to πA,B ◦ f and
2
πA,B ◦ f whose targets are of lower complexity. 

Subcategory, full subcategory

Given a category C, its subcategory is a category whose set of objects is a
subset of OC and whose set of arrows is a subset of AC , while its source,
target, identity and composition are just restrictions of the corresponding
functions of C. A subcategory A of C is full, when for every pair of objects
A1 and A2 of A, if f : A1 → A2 is an arrow of C, then it is an arrow of A,
i.e., the embedding of A into C is a full functor.

Our manageable category is very simple. We start with its formal presentation
and then skip to pictures as a more practical way to handle with arrows of this
category. Let O be the category whose objects are the finite ordinals 0 = ∅, 1 = {0}
and in general n = {0, . . . , n − 1}, and whose arrows from n to m are the functions
mapping the set n to the set m. This category is a full subcategory of the category
Set of sets and functions.
Let Oop be the opposite category of the category O. The arrows from n to m of
this category are the functions whose domain is the set m and codomain is the set
n. Since the category O is a category with coproducts (a notion dual to the notion
of product) we have that Oop is a category with products.

Coproducts
A coproduct of two objects A and B in a category C consists of an object
A + B of C and a pair of arrows ι1A,B : A → A + B, ι2A,B : B → A + B
of C, and it is characterized by the following universal property: for every
pair of arrows f : A → C and g : B → C of C, there is a unique arrow
h : A + B → C of C such that
h ◦ ι1A,B = f and h ◦ ι2A,B = g.
 270 MINUTES ON CATEGORIAL PROOF THEORY 29

ι1 ι2
A A+B B
h g
f
C
The coproducts in O are given by addition on objects and by putting “side by
side” on arrows, i.e., the coproduct of f : n → m and f 0 : n0 → m0 is given by the
function g : n + n0 → m + m0 such that
(
f (i), when 0 6 i 6 n − 1,
g(i) =
m + f (i − n), when n 6 i 6 n + n0 − 1.
0

By inverting the arrows, this gives the products in Oop . We proclaim Oop to be
the manageable category.
The objects of Oop are presented so that n is presented by n vertices. For
example, 5 is presented by the following picture.
0 1 2 3 4
q q q q q
The object n × m is presented by the picture for n + m.
The arrows of Oop are presented by the standard pictures for finite functions.
Note that the picture is such that the set of vertices at the bottom line is mapped
to the set of vertices at the top line. For example, an arrow from 5 to 3 of Oop is
presented by the following picture.
0 2 1 3 4
q q q q q

@

q
 q @q
0 1 2

1 2
The arrows π3,2 : 3 + 2 ` 3 and π3,2 : 3 + 2 ` 2 are presented respectively by
0 1 2 3 4 0 1 2 3 4
q q q q q q q q q q
 
q q q q 
q


0 1 2 0 1

while w2 : 2 ` 2 + 2 is presented by the following picture.

0 1
q q
@ @
q q @q @q
0 1 2 3

For f : 3 ` 2 and g : 2 ` 3 presented respectively by

0 1
q q 2q 0
q 1
q

q q
 q q q
0 1 0 1 2
30 ZORAN PETRIĆ

the arrow f × g is presented by the following picture.

0 1 2 3 4
q q q q q
 
q q q q q


0 1 2 3 4

It is not difficult to show directly, without using the above categorial arguments
involving dualities, that π 1 , π 2 and the pairing defined as hf, gi =df (f × g) ◦ w,
satisfy (6.1) and (6.2). Using the above pictures as arguments is very convincing.
This is left as an exercise for the reader.
Let a constant function from the set of propositional letters to the set of objects
of Oop be such that every letter is mapped to 1. Since L is a category freely
generated by the set of propositional letters, this function extends in a unique way
to a functor from L to Oop . For example, the arrow f
1 1 2 1 1
hhπp,q ◦ πp∧q,p , πp∧q,p i, πp,q ◦ πp∧q,p i : (p ∧ q) ∧ p ` (p ∧ p) ∧ p
of L is mapped by this functor to the following arrow of Oop
0 1 2
q q q
HH
q q Hq
H
0 1 2

which is usually drown as edges connecting letters in the source and the target of f .

( p ∧ q )∧ p
q q q
HH
q q Hq
H
( p ∧ p )∧ p
Hence, we have a soundness result—to check that f and g are different in L, it
suffices to draw the corresponding pictures and find a difference.
Theorem 6.2. (Soundness) If f and g represent the same arrow of L, then they
have the same picture.
For a completeness result, it is necessary to find a touching point of the syntax
and semantics. This is why we represent the arrows of L by terms in normal form.
Let f : A ` B be a term in normal form. The nesting of h and i in f corresponds
to the nesting of ( and ) in B and the subterms of f of the form 1p or fn ◦ . . . ◦ f1 ,
with A as the source and a letter as the target, are in one to one correspondence
with the occurrences of letters in B.
It is not difficult to see that every such subterm corresponds to a unique edge in
the picture, and vice versa, one can read such a subterm from an edge. This is the
touching point for the syntax and semantics. For the term and the picture from
the preceding example, we have the following correspondence.
1 1 2 1 1
hhπp,q ◦ πp∧q,p , πp∧q,p i, πp,q ◦ πp∧q,p i : (p ∧ q) ∧ p ` (p ∧ p) ∧ p
 270 MINUTES ON CATEGORIAL PROOF THEORY 31

( p ∧ q )∧ p
q
H q q
HH
q q Hq
( p ∧ p )∧ p
A formalization of these observations is left as an exercise for the reader. Hence,
if two terms, in normal form and with the same type, are not identical, then the
corresponding pictures are different.
Theorem 6.3. (Completeness) If f and g are two terms with the same picture,
then f and g represent the same arrow of L.
Proof. Let f and g be two terms with the same type and having the same picture,
i.e., mapped to the same arrow of Oop . By the Normal-Form Theorem, there are
f 0 and g 0 in normal form such that f ≡ f 0 and g ≡ g 0 . By Theorem 6.2, f 0 and g 0
have the same picture and by the preceding paragraph, these terms are identical.
Since ≡ is an equivalence relation, we conclude that f ≡ g, and f and g represent
the same arrow of L. 
Hence, we have a completeness result—to check that f and g are equal in L, it
suffices to draw the corresponding pictures and find no difference.

7. Hints for the exercises

1. In order to prove that the isomorphisms (A ∧ B) → C ∼ = B → (A → C) and
>→C∼ = C hold in Klax , let the components in Klax of the natural isomorphism
that replaces associativity be denoted by
aA,B,C : A ∧ (B ∧ C) ` (A ∧ B) ∧ C, a−1
A,B,C : (A ∧ B) ∧ C ` A ∧ (B ∧ C),
and let the components in Klax of the natural isomorphism that replaces neutral
conditions for > be denoted by
lC : > ∧ C ` C, l−1
C : C ` > ∧ C.

Write down the naturality conditions for a, a−1 , l and l−1 .

For f : A ∧ B ` C, denote by f ∗ the arrow
(A → f ) ◦ ηA,B : B ` A → C.
Show that for g : D ` B, the following equations hold:
(1) f ∗ ◦ g = (f ◦ (1A ∧ g))∗ ,
(2) εA,C ◦ (1A ∧ f ∗ ) = f ,
(3) ε∗A,B = 1A→B .
In terms of ε, a, a−1 and ∗, define arrows of the type
(A ∧ B) → C ` B → (A → C) and B → (A → C) ` (A ∧ B) → C
and show that they are inverse to each other.
In terms of ε, l, l−1 and ∗, define arrows of the type
>→C`C and C ` > → C
32 ZORAN PETRIĆ

and show that they are inverse to each other.

2. In order to show that the two definitions of a category with binary products
are equivalent, let us first assume that C is a category of binary products according
to the first definition. In terms of π 1 , π 2 and pairing h , i define the product f × g
of two arrows and show that it has functorial properties. In the same terms define
the components of natural transformations that may serve as the unit and counit
of the adjunction between the diagonal functor and the product functor. Show that
the unit and counit are natural and that the triangular equations hold.
On the other hand, if the second definition is assumed, then define the pairing
h , i in terms of the right adjoint × of the diagonal functor, and the unit of this
adjunction. It remains to show that the equations (6.1) and (6.2) hold.

References
[1] K. Došen and Z. Petrić, Isomorphic objects in symmetric monoidal closed categories, Math.
Struct. in Comp. Science 7 (1997), 639–662.
[2] ——–, Proof-Theoretical Coherence, KCL Publications, London, 2004 (revised version avail-
able at: http://www.mi.sanu.ac.yu/~kosta/publications.htm)
[3] G. Gentzen, Untersuchungen über das logische Schließen, Math. Z. 39 (1935), 176–210, 405–
431 (English translation: Investigations into logical deduction, in [4], 68–131)
[4] ——–, The Collected Papers of Gerhard Gentzen M.E. Szabo (ed.), North-Holland, Amster-
dam, 1969
[5] S. Mac Lane, Categories for the Working Mathematician, Springer, Berlin, 1971 (expanded
second edition, 1998).
[6] D. Prawitz, Ideas and results in proof theory, in: J.E. Fenstad (ed.), Proceedings of the Second
Scandinavian Logic Symposium, North-Holland, Amsterdam, 1971, 235–307.