0% found this document useful (0 votes)
12 views

Girard - On the unity of logic

The document discusses a unified sequent calculus that encompasses classical, intuitionistic, and linear logics as fragments of a single logical system. It emphasizes that proofs can utilize lemmas from different logics while maintaining their integrity within the intended logical framework after cut-elimination. This approach contrasts with traditional methods that treat different logics as separate systems, proposing instead that a single logic can adapt to various uses through the manipulation of formulas.

Uploaded by

intimalai
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views

Girard - On the unity of logic

The document discusses a unified sequent calculus that encompasses classical, intuitionistic, and linear logics as fragments of a single logical system. It emphasizes that proofs can utilize lemmas from different logics while maintaining their integrity within the intended logical framework after cut-elimination. This approach contrasts with traditional methods that treat different logics as separate systems, proposing instead that a single logic can adapt to various uses through the manipulation of formulas.

Uploaded by

intimalai
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

Annals of Pure and Applied Logic 59 (1993) 201-217 201

North-Holland

On the unity of logic


Jean-Yves Girard
gquipe de Logique, UA 753 du CNRS Mathtkatiques, Universitt? Paris VII, t. 45-55, 5” &age,
2place Jussieu, 75251 Paris Cedex 05, France

Communicated by D. van Dalen


Received 20 June 1991

Abstract

Girard, J.-Y., On the unity of logic, Annals of Pure and Applied Logic 59 (1993) 201-217

We present a single sequent calculus common to classical, intuitionistic and linear logics. The
main novelty is that classical, intuitionistic and linear logics appear as fragments, i.e. as
particular classes of formulas and sequents. For instance, a proof of an intuitionistic formula A
may use classical or linear lemmas without any restriction: but after cut-elimination the proof
of A is wholly intuitionistic, what is superficially achieved by the subformula property (only
intuitionistic formulas are used) and more deeply by a very careful treatment of structural
rules. This approach is radically different from the one that consists in “changing the rule of the
game” when we want to change logic, e.g. pass from one style of sequent to another: here,
there is only one logic, which-depending on its use-may appear classical, intuitionistic or
linear.

Nous prtsentons un calcul des sequents unifit, commun aux logiques classique, intuitionniste et
lintaire. La principale nouveaute est que les logiques classique, intuitionniste et lineaire
apparaissent comme des fragments, c’est h dire comme des classes particulieres de formules et
de sequents. Par exemple la demonstration d’un &once intuitionniste pourra utiliser des
lemmes classiques ou intuitionnistes sans limitation: simplement apres elimination des
coupures, la demonstration se fera entierement dans le fragment intuitionniste, ce qui est
superficiellement assure par la propriett de la sous-formule (seulement des formules intuition-
nistes sont utilisees) et plus profondement par un traitement tres rigoureux des regles
structurelles. Cette approche est radicalement differente de I’approche habituelle qui consiste
tout bonnement a changer la regle du jeu quand on veut changer de logique, c’est a dire de
style de sequent: ici il n’y a plus qu’une seule logique, qui au grC des utilisations peut
apparaitre classique, intuitionniste ou lineaire.

By the turn of this century the situation concerning logic was quite simple:
there was basically one logic (classical logic) which could be used (by changing
the set of proper axioms) in various situations. Logic was about pure reasoning.
Brouwer’s criticism destroyed this dream of unity: classial logic was not suited for
constructive features and therefore it lost its universality. Now by the end of the
Correspondence to: J.-Y. Girard, Mathematiques Disc&es, UPR A9016, 163 Av. de Luminy, case
930, 13288 Marseille Cedex 09, France.

016%0072/93/$06.00 0 1993 - Elsevier Science Publishers B.V. All rights reserved


202 J.-Y. Girard

century we are faced with an incredible number of logics-some of them only


named ‘logic’ by antiphrasis, some of them introduced on serious grounds. Is
logic still about pure reasoning? In other words, could there by a way to reunify
logical systems-let us say those systems with a good sequent calculus- into a
single sequent calculus ? Is it possible to handle the (legitimate) distinction
classical/intuitionistic not through a change of system, but through a change of
formulas? Is it possible to obtain classical effects by a restriction to classical
formulas? Etc.
Of course, there surely are ways to achieve this by cheating, typically by
considering a disjoint union of systems. However, all these jokes will be made
impossible if we insist on the fact that the various systems represented should
communicate freely (and for instance a classical theorem could have an
intuitionistic corollary and vice versa).
In the unified calculus LU that we present in this paper, classical, linear and
intuitionistic logics appear as fragments. This means that one can define notions
of classical, intuitionistic or linear sequents and prove that a cut-free proof of a
sequent in one of these fragments is wholly inside the fragment; of course a proof
with cuts has the right to use arbitrary sequents, i.e., the fragments can freely
communicate.

1. Unified sequents

Standard sequent calculi essentially differ by their different maintenances of


sequents:
(i) classical logic accepts weakening and contraction on both sides;
(ii) intuitionistic (minimal) logic restricts the succedent to one formula-
which has the effect of forbidding weakening and contraction to the right;
(iii) linear logic refuses both, but has special connectives ! and ? which - when
prefixed to a formula - allow structural rules on the left (!) and on the right (?).
Our basic unifying idea will be to define two zones in a sequent: a zone with a
‘classical’ maintenance, and a zone with a linear maintenance; there will be no
zone with an intuitionistic maintenance: intuitionistic maintenance, i.e. ‘one
formula on the right’, will result from a careful linear maintenance. Typically, we
could use a notation T;I” F A’;A to indicate that r’ and A’ behave classically,
whereas r and A behave linearly. We could try to identify classical sequents with
those where r and A are empty, and intuitionistic ones as those in which r and A’
are empty, A consisting of one formula. This is roughly what will happen, with
some difficulties and some surprises:
(i) It must be possible to pass between both sides of the semi-colon: surely one
should be able to enter the central zone (we lose information), and also- with
some constraint, otherwise the semi-colon would lose its importance- to move
to the extremes. One of these constraints could be the addition of a symbol, e.g.
move A from I” to r, but now write it as !A.
On the unity of logic 203

(ii) This is not quite satisfactory; typically, a formula already starting with ‘!’
should be able to pass freely. However, it immediately turns out that those guys
that can cross the left semi-colon in both ways are closed under the linear
connectives 63) and $ and under the quantifier Vx. The sensible thing to do is
therefore to distinguish among formulas positive ones, including positive atomic
formulas for problems of substitution. Symmetrically, one distinguishes negative
formulas, while the remaining ones are called neutral: those must pay at both
borders.
(iii) The restatement of the rules of linear logic in this wider context is
unproblematic and rather satisfactory, especially the treatment of ‘!’ and ‘?’
becomes slightly smoother.
(iv) We now have to define three polarities (classes of formulas) and we can
toy with the connectives of linear logic to define synthetic connectives, built like
chimeras, with a head of 8, a tail of &, etc. -only good taste limits the
possibilities. Typically, if we want to define a conjunction we would like it to be
associative (at the level of provability, but moreover at the level of denotational
semantics), hence this imposes some coordination between the various parts of
our chimera. In fact the connectives built have been chosen under two
constraints:
-limitation of the number of connectives: for instance only one conjunction,
only one disjunction, for classical and intuitionistic logics, but unfortunately two
distinct implications for these logics;
- maximimisation of the number of remarkable isomorphisms.
(v) As far as classical logic is concerned, the results presented here are
consistent with the previous work of the author [3]; in fact classical logic is
obtained by limitation to formulas which are (hereditarily) nonneutral. The role
of classical sequents is played by the sequents of the form T;T’ t A’;A when the
nonpermeable part of r, A consists of at most one formula (the stoup of [3]). The
reader is referred to this paper to check the extreme number of isomorphisms
satisfied by the classical fragment (some of them, typically the De Morgan duality
between A and v, do not extend to neutral polarities). There is only one small
defect: a single formula A is interpreted by ; IA; whereas for the other logics, it is
interpreted by ; t ;A. However, if A is negative (right permeable) we can replace
; t-A; by ; k ;A, and if A is positive we can replace ; t A; by ; k ; Vx A (x dummy)
or ;I-;A v (TV).
(vi) As far as disjunction, existence and negation are ignored, intuitionistic
logic is a quite even system in proof-theoretic terms, as shown by various
relations to k-calculus. The neutral intuitionistic fragment is made of (hereditarily)
neutral formulas, and basically accepts intuitionistic 2, A and Ax; besides
sequents ;I’t ;B which were expected, there arise sequents A;Tt ;B
corresponding to the notion of headvariable. Not only the usual intuitionistic
sequent calculus is recovered, but it is improved!
(vii) Surely less perfect is the full intuitionistic system with v, 3x and F (i.e.
204 J.-Y. Girard

negation); the translation of this system into linear logic (the starting point of
linear logic, see [2]) made use of the combination !A @ !B, which is awfully
nonassociative (denotationally speaking): compare !(!A 63 !B) G3 !C with !A f?3
!(!B @ !C). However, one could use A instead of !A if A were known to be
positive. Therefore there is room for an associative disjunction provided we
consider not only neutral formulas, but also positive ones. The resulting
disjunction is a very complex chimera which manages to be associative and
commutative, and also works in the classical case. We surely do not get as many
denotational isomorphisms as we would like (typically there is no unit for the
disjunction, or A I> B A C ^- (A I B) A (A I C) only when B and C are neutral),
but the situation is incredibly better than expected. In terms of sequents, we lose
the phenomenon of ‘headvariable’, since a term may be linear in several of its
variables if we perform iterated pattern-matchings.
The system presented here is rather big, for the reason that we used a
two-sided version to accommodate intuitionistic features more directly, and
because there are classical, intuitionistic and linear connectives; last, but not least
rules can split into several cases depending on polarities; the rules for disjunction,
for instance, fill a whole page! But this complication is rather superficial: it is
more convenient to use the same symbol for nine ‘micro-connectives’ correspond-
ing to all possible polarities of the disjuncts. Given A and B, we get at most two
possible right rules and only one left rule, as usual. So LU has a very big number
of connectives but apart from this it is a quite even sequent calculus.

2. Polarities

Each formula is given with a polarity +1 (positive), 0 (neutral), -1 (negative).


We use the following notational trick to indicate polarities: P, Q, R for positive
formulas, S, T, U for neutral formulas, L, M, N for negative formulas. When we
want to ignore the polarity, we shall use the letters A, B, C.
Semantically speaking, a neutral formula refers to a coherent space; a positive
formula refers to a positive correlation space, and a negative formula to a
negative correlation space (see [3] for a definition). Now remember that a
correlation space is a coherent space plus extra structure (in fact PCS generalise
spaces of the form !X, and NCS generalise spaces ?X; both are about structural
rules: a PCS is a space with left structural rules, a NCS obeys right structural
rules); this explains the polaritiy table (see Table 1) for linear logic: we first
combine the underlying coherent spaces S and T to get a coherent space U (e.g.
U = S 8 T), and if possible we try to endow U with a canonical structure of
correlation space (typically, if S and T are underlying coherent spaces for PCS P
and Q, we equip S 8 T with a structure of PCS in the obvious way).
Before we start, we have to make a choice about polarity 0: do we consider the
possibility that something of polarity +1 (or -1) has also polarity O? In that case
On the unity of logic 205

Table 1
Polarities for linear connectives

==I=-
A B

+l +l

0 +1

-1 +l 0 0 0 0 0

+l 0 0 0 0 0 0

0 0 0 0 0 0 0

-1 0 0 0 0 0 0

+1 -1 0 0 -1 0 0

0 -1

-1 -1

it would be normal to indicate that we decide to forget the nonzero polarity,


causing complications and more complications. In fact, if we decide to answer
No, we get a quite reasonable answer: in linear logic we can forget negative
polarity by forming A @ 1, and a positive one by forming A 42 I, hence if I
replace A by (A @ 1);8 I, I can change the polarity to 0. (In a similar way,
V 3 A neutralises any intuitionistic formula.)

3. Sequent calculus: identity and structure

The sequent calculus LU is defined as follows: sequents are of the form


T;T’ t- A’;A where I, r’, A and A’ are sequences of formulas of the language.
The space between the two semi-colons is a space in which the usual structural
rules are available; the intended meaning of such a sequent is that of a proof
which is linear in rand A, i.e. in terms of linear logic of r,!r’ t ?A’,A.
The rules of identity and structure are presented in Fig. 1. They are
independent of any commitment: these rules are meant for all formulas, and do
not refer to any distinction of the form classical/intuitionistic/linear. We have
adopted a two-sided version, which has the effect of doubling the number of
rules; a one-sided version would have been more economical, but we would had
to pay for this facility in the discussion of the intuitionistic fragment which would
look slightly artificial when written on the right. To compensate for this
complication, we decided to use an additive maintenance for the central part of
206 J.-Y. Girard

Identity

A; c ;A

l-‘;I” ,- A’;A,A A,A;l-” c- A’;II


r,h;I” c A’;A,fI

I’;I” I- A’,A;A A;I” c A’; ;r’ ,- A’;A A;A,r’ c A’;Il


I’;l?’ c A’;A h;r’ c A’;II

Structure

r;r’ c A’;A
a(r) ; Q’ (I-‘) c T’ (A’) ; r(A)

r;r’ B-A’;A r;r’ t- A’;A


rp c A,A’;A r;r’,A + A’;A

r;r’ c A,A,A’;A r;r’,A,A I- A’;A


rp c A,A’;A r;r’,A + A’;A

l?;r’ c A’;A,A r,.4;r’ a- A’;A


J?;r ,- A’,A;A r;A,r’ + A’\;A

r;I” c A’,N;A r;P,r’ + A’;A


I’;r’ ‘- A’;N,A r,p;r’ I- A’;A

Fig._l.

sequents (the same r’ and A’ in binary rules), which is possible since structural
rules are permitted in this area. Another notational trick would be (instead of the
semi-colon) to underline those formulas with a classical maintenance, which
would simplify the schematic writing of our rules, but would not change anything
deep, so this is really a matter of taste.
As expected, weakening and contraction are freely performed in the central
part of the sequent. Besides the exchange rules, which basically allow permuta-
tion of formulas separated by a comma, we get additional permeability rules,
which allow formulas to enter the central zone, and to exit from this zone under
some restriction on polarities. The last group of rules is the only one depending
on polarities.
On the unity of logic 207

The identity axiom is written in a pure linear maintenance. The case of cut is
more complex; in fact it falls into two cases, depending on the style of
maintenance for the two occurrences of A:
(i) if they are both 1inear (i.e. outside the central area), we obtain a rather
expected rule;
(ii) if one of them is linear and the other ‘classical’ we obtain two symmetric
forms of cut; observe that the premise containing the linear occurrence of A is of
the form A;T’ t A; or ;rk A’; A, i.e., the context of A is handled classically.
There is no possibility of defining a cut between two occurrences of A with a
classical maintenance. As a matter of fact, there is no need for that: typically in
classical logic, if we get a cut on A, then A has a polarity +l or -1, and one of
the two occurrences of A can be handled linearily.

4. Logical rules: cases of linear connectives

The calculus presented in Fig. 2 seems rather heavy compared with the usual
formulation of linear logic; but this is just an unpleasant illusion due to the fact
that we have chosen a two-sided version more than twice the size of the one-sided
version.
As expected, the rules for quantifiers (right Ax and left VX) are subject to the
restriction on variables: x not free in T;T’ t A;A’.
This calculus is equivalent to the usual linear logic; more precisely we can
translate the usual linear logic into this new system by declaring all atomic
propositions to be neutral. Then a sequent rl- A in the usual (two-sided) linear
logic becomes r; t ;A. It is easy to translate proof to proof, though the rules for
the exponentials ! and ? are translated by a heavy use of structural manipulations.
For instance, to pass from !r; t ;?A,A to !r; t ;?A,!A, we transit through
;!rk ?A;A, then ;!Tt ?A;!A, and the ultimate moves to :r; t ;?A, !A use the
polarities of ?A and !I’.
Conversely, this new calculus (as long as we restrict ourselves to neutral atomic
propositions) can be translated into the usual linear logic as follows: a sequent
r;r:,r”t A”,AL;A (r: positive, AL negative) translates as r,r!+,!r’1 ?A”,AL, A
in the old syntax for linear logic. Then we have to mimick all rules of the new
calculus in the old one, which offers no difficulty. Of course, we have to prove in
the old calculus a stronger form of the rule for ‘!‘, namely that one can pass from
Tt- A,A to r1 A, !A, as soon as Tis positive and A negative. But since our atoms
are neutral, positive formulas are built from 0, 1, and formulas !A by means of
$, 8 and 3x (and symmetrically for negative formulas), and we can make an
easy inductive argument.
Of all the logical rules of linear logic, only the rules for exponentials do
something to the central part: the right rule for ‘!’ assumes that the context lies
wholly in the central part, whereas the left rule moves a formula from the central
208 J.-Y. Girard

; I- ;l 1; a- ;

r;r' ,-A';A,A A;r' ,-A';ll,B A,B,r;r' .-A';A


r,A;r' c A';A,II,A@B A@B,r;r' n-A';A

r;r' I-A';A,A,B A,r;T' ,-A';A B,A;r' ,-A';Il


r;r' c A';A,A?8B AqB,r,A;r' + A';A,ll

A,r;r' c A';A,B r;r' s-A';A,A B,h;r' ,-A';ll


r;r' n-A';A,A+B A*B,r,A;r' I-A';A,II

r; e- ;A,T O,l-‘; c ;A

r;r' s-A';A,A r;r'cA';A,B .4,r;r'c A';A B,r;r' c A';A


r;r' c A';A,A&B A&B,r;r' c A';A A&B,r;r' c A';A

r;r' + A';A,A r;r' + A';A,B A,r;r' I-A';A B,r;r' D-A';A


rp + A ';A,A~B r;r’ t A';A,A@B A@B,r;r' c A';A

r;r' N-A';A,A A,rp I-A';A


Al,r;r' b A';A r;r' c A';A,Al

;r' c A';A r;.4,r't A';A


;r' c A';!A !A,r;r' t A';A

r;r' n-A',A;A A;r' c A';


r;r' c A';A,?A ?A;r' c A';

r;r' c A';A,A A[t/x],r;r' + A';A


?;I-'c A';A,hxA hx.4,r;r*c A';A

I-g-9c A ‘;A,A[t/xl A,r;I" c A';A


I-;J?’ c A';A,VxA VxA,l'-;r'
c A';A

Fig. 2.
On the unity of logic 209

area to the extreme left, at the price of a symbol ‘!‘; the new formula !A can now
pass the semi-colon in both ways.

5. Some chime& connectives

It is possible to define new connectives by pattern matching, i.e. by considering


polarities. Below, we shall consider only those connectives and quantifiers which
are of interest to classical and intuitionistic logics: these connectives are A, v , j,
1, V, F, vx, 3x (classical), fl, U, 3, -, V, F, (x) and (Ex) (intuitionistic).
However, it turns out that n, U, V, F, (x) and (Ex) can be chosen to coincide
with A, v, 1, 0, Ax, 3x; moreover, intuitionistic negation is better handled as
-A:AxO.

Our tables (see Tables 2 and 3) have been chosen so as to minimize the total
number of connectives, and to get as many denotational isomorphisms as
possible. It has not been possible to keep the same connective for implication
(owing to conflicts of polarities). Our classical implication has been made up from
1A v B and is quite complicated; another one, built on l(A A lB), would be
simpler, but the discussion is rather sterile since the difference cannot be noticed
on classical formulas.
The rules for our new connectives are presented in Fig. 3.
The author can be accused of bureaucracy: even if one regroups rules, their
number remains. . . frightening. Surely the fact that disjunction is defined by nine

Table 2
Polarities for classical and intuitionistic connectives
210 J.-Y. Girard

Table 3
Classical and intuitionistic connectives defined in terms of linear logic

independent cases counts for something in this inflation. However, observe that
these rules are always variations of the familiar rules for disjunction, and each
line differs from the other by a slightly different structural maintenance. Given a
concrete disjunction A v B, only one of these lines can work, i.e. at most three
rules at usual. Moreover, the usual fragments use at most four lines out of nine.
Also, these rules manage to unify classical and intuitionistic disjunction in the
same ussocia~iveconnective, which is a nontrivial achievement.
All other usual connectives coincide with one of those already introduced, with
the exception of intuitionistic negation: it is impossible to write its rules without
using the constant F (or 0): this minor defect comes from our very cautious
treatment of structural rules; it is therefore better to consider - as defined by
-A:=A zF.

6. Some properties of the calculus

First let us fix once for all a reasonable language:


-atomic predicates are given with their polarity (+ 1, 0, -1);
-two constants 0 and 1, both positive (also denoted F and V);
- unary connectives: !, (e)’ (also denoted 1);
-binary connectives: A, v , ITS, 3, C3,4?, 4, CB, &;
-quantifiers: Vx, 3x, Ax, Vx.
On the unity of logic 211

Rules for conjunction

I’;l?’ I- A';A,P A;r' + A';II,Q P,Q,r;r' c A';A


r,A;r' t-A';A,II,PAQ PhQ,r;r' s-A';A

;r' + A';A Air* c A';~,Q Q,r;r',A .-A';A


A;r' c A';II,AAQ AhQ,r;r' c A';A

r;r' @-A';A,P ;r' c A';B P,rp,B c A';A


r;r' c A';A,PAB PhB,I’;r + A’;A

r;r' c A’;A,A r;r' c A’;A,B A,r;r' + A';A B,r;r' @-A';A


r;r' t A';A,AAB AAB,r;I" c A ';A AAB,r;r' c A';A
Commenfs: P, Q positive; A, B not positive.

Rules for intuitionistic implication

P,r;r' c A';A,B r;r' #-A';A,P B,h;r' c A';Il


r;r' c A';A,PlB P3B,r,A;r' k A';A,II

l-‘;l-“,A c A';A,B ;r' I-A';A B,A;r' c A';Il


r;r' c A';A,A)B A)B,A;r + A';II
Comments: P positive; A not positive; B arbitrary.

Rules for “V”

I-;I-’ cA',A;A A[t/x];A’ + II’;


I’;r c A';A,tlxA b&A' c II';

F;I-’ c A’;A,N N[t/x] ,A;A’ c II’;l-I


I’;r’ c A’;A,VxN VXN,A;A’ c II*;II

Comments: A not negative; N negative; x not free in T;T’ k A’;A.

Fig. 3.
212 J.-Y. Girard

Rules for disjunction

I’;r’ c A';A,P I';l?'


.-A';A,Q P,r;r’ I- A';A p,r;r* + A';A
r;r’ + A';A,PVQ r-p c A';A,PVQ pvrj,r;rp + A';A

p I- A';S I';r'c A';A,Q r;s,r’ D-A';A q,r;r’ k A'.;A


;I-'c A';SVQ l-';I-'
c A';A,SVQ svq,rp c A';A

l-';rc A',Q;A,M M,I';r'B-A';A ;I",Q t-A';


I';r'c A';A,biVQ arvp,r;r’ I- A';A

r;r’ @-A';A,P ;I" c A';T p,r;r’ + A';15 rp,T B-A';A


r;I" n-A';A,PVT ;I" c A';PVT wr,r;r’ I- A';A

;I” a- A';S ;I" c A';T rp,s m-A';A I-;I",Tc A';A


;I" c A';SVT ;I" c A';SVT svT,r;r’ c- A';A

r;r’ D-A';A,M ;J?'s-A';M,T M,rp b- A';A ;I",T B-A';


I';l?'
c A';A,MVT ;r' n-A';MVT wT,rp t- A';A

r;r’ c P,A';A,N p;r’ t- A'; N,A;I" t A';ll


I';r c A';A,PVN PVN,A;I" c-A';II

jr c A';S,N I';I"c A';A,N ;r’,s c A'; N,A;I" c A';Il


;r' .-A';SVN r;l-"c A';A,SVN SVN,A;r' c A';l-I

r;r' c A';A,M,N M,r;r’ @-A';A N,A;r'cA';II


I';r'c A';A,blVN bivN,r,h;r’ m-A';A,II
Comments: P, Q positive; M, N negative; S, T neutral.

Rules for “3”

r;r’ k A';A,P[t/xl P,A;A' c II';l-l


l-';I"
c A';A,gxP ElxP,A;A’ + n’;ri

g-3 D-A ';A[t/xl A;A',A c II';II


;r’ + A';hA 3x.4,A;A’ c n’;n

Comments: P positive, A not positive, x not free in ,$A t II’;II.

Fig. 3 (contd.).
On the unity of logic 213

Rules for classical implication

r;l-” c A’;A,P ~,r;r’ + A’;A r;r' c-A’;A,N Q,A;r’ ,- A';l-I


r;r' c A ‘;A,NJP r;r' c A’;A,NJP N+P,r,A;r' t A’;A;Il

p,r;r* I-Q,A’;A r;r' + A’;A,P Q;r’ + A’;


r;r' c A’;A,P+Q pjp,r;r' + A’;A

r;r',M c A';A,N g-9 I-A';M N,A;r' c A';II


r;r' c A’;A,MjN M+N,A;p’ c A’;II

P,r;f” c A’;A,N r;r' t A’;A,P N,A;p’ c A’;l-l


r;l-” c A’;A,P+N P+N,r,A;r’ * A’;A,B

Comments: P, Q positive, M, N negative; this set of rules is incomplete (we have omitted the rules
involving neutral formulas, for which there is no use at present; the reader may reconstitute them
from the rules of disjunction). Observe that the four rules written would have been the same if
implication had been defined from conjunction.

Fig. 3 (cod.).

We now define remarkable fragments; they are all defined by a restriction of


the possible atomic formulas and of the possible connectives and quantifiers:
(1) The classical fragment:
-positive and negative atoms (including V and F); closed under 1, A, v, +, ‘dx
and 3x.
(2) The intuitionistic fragment:
-positive and neutral atoms (including V and F); closed under A, v , 3, Ax, 3x.
(3) The neutral intuitionistic fragment:
- neutral atoms; closed under A, 3, Ax.
(4) The linear fragment:
-all atoms; closed under (.)I, 8, 9, +I, @, &, !, ?, Ax, Vx.
The interest of these various fragments is that they enable us to formalise
arguments belonging to various logical systems inside LU, with the advantage of a
unique proof-maintenance. Each fragment uses a very small part of our kolossal
sequent calculus. But LU is not the union of its fragments: there must be
interesting formulas outside of these fragments (and also other interesting
fragments; for instance a positive intuitionistic fragment based on the implication
!(A XI B) should be investigated).
The classical fragment is based on the idea of staying within the positive or
negative formulas; the intuitionistic fragment stays within the positive and neutral
formulas; the neutral intuitionistic fragment is wholly neutral; the linear fragment
admits all three polarities.
214 .I. -Y. Girard

An important property of these fragments is the substitution property: let a be a


proper predicate symbol of arity n, and let A be a formula of the same polarity as
a, in which distinct free variables x1, . . . , x,, have been distinguished. Then one
can define for any formula B the substitution B[IZx,, . . . , x,.A/a] as the result of
replacing any atom at1 * - - t, of B by A[t,, . . . , t,] (with usual precautions
concerning free and bound variables, comrade Tchenienko). All the fragments
considered are closed under mutual substitution.
Each fragment gets its own notion of sequent: first all formulas must belong to
the fragment; but some additional properties may be required:
(i) A classical sequent T;T’ 1 II’;II is such that if we make the sum of the
number of negative formulas in r and of positive formulas in II, we get the total
number 0 or 1.
(ii) An intuitionistic sequent is of the form T;T’ k ;A.
(iii) A neutral intuitionistic sequent is a sequent T;T’ t ;A, with at most one
formula in lY

Theorem. If a sequent of one of the fragments considered provable, it is provable


within the fragment.

Proof. We limit our search to cut-free proofs; by the subformula property, all the
formulas occurring in the proofs belong to the fragment; in particular this is
enough for linear logic, since no additional restriction has been imposed on linear
sequents. Let us consider the remaining cases: in all cases we have to check that
the restriction on the shape of the sequent can be forwarded from the conclusion
to the premise(s).
Neutral intuitionistic fragment. First observe that the restriction “A’ empty”
will be easily forwarded (this holds for both intuitionistic fragments). Then
observe that for any cut-free rule of LU ending with a neutral intuitionistic
sequent r;r’ I-;S,
- all premises are of the form A;A’ k ;T,
- one of these premises, say A;A’ k ;T, is such that the number of formulas in A
is greater or equal to the number of formulas in r, with only one exception,
namely the identity axiom. In particular there is no way to prove a sequent
I’$” I- ;S of formulas in this fragment when r has two formulas or more; the
formula of r (if there is one) is the analogue of the familiar headvariable of typed
&calculi, which are based on neutral intuitionistic fragments.
This proves that all premises of the rule must be also neutral intuitionistic
sequents.
Classical fragment. If S is the sequent r;I” t A’;A let us define p(S) to be the
sum of the number of negative formulas in r and of positive formulas in A. Now
for any rule with a conclusion S made of classical formulas and such that
p(S) > 1, there is premise S’ such that p(S’) 2 p(S), with only one exception:
On the unity of logic 215

the axiom F, r; I- ;A.Furthermore, there are only two rules with a premise S’ and
a conclusion S such that p(S’) > p(S): the two permeability rules enabling a
formula to enter the central zone. Now it is an easy exercise, given any cut-free
proof of a sequent S made of classical formulas with ,u(S) > 1, to produce another
proof of any sequent S’ obtained by removing as many formulas among those
which contribute to p(S). In particular, a ‘bad’ permeability rule can be replaced
with a weakening and so we stay among classical sequents.

lntuitionisticfragment. If Y(S) counts the number of formulas in the part A of a


sequent S = T;T’ !- ;A, then the restriction Y(S) 6 1 is forwarded from the
conclusion to the premises of all rules involving intuitionistic formulas but for the
case of a rule

l-;T’k ;C,P B,A;l-’ t ;

Easy commutation arguments reduce the use of this rule to the case where C is
positive or atomic. From this it can be ensured that all sequents with Y(S) > 1
occurring in the proof of an intuitionistic sequent have a succedent made of
positive or atomic formulas. Now one can easily produce given a proof of a
sequent T$” t- ;A with Y(A)f 1 (this includes v(A)= 0) and all formulas
intuitionistic, another proof of I’$’ t $7 where n has been obtained from A by
adding formulas, or removing atomic or positive ones. In particular, we can
replace the ‘bad’ rule above by the ‘good’ one:

r;rlk ;P B,A;r't;C
~~B,r,n;r+;c

and this shows that we can stay within the intuitionistic sequents. 0

Remarks. (i) We implicitly used a cut-elimination theorem for LU that is more


or less obvious (but perhaps a bit too long to write down explicitly).’
(ii) The results of the theorem concern not only provability, but also proofs, in
the sense of denotational semantics; there would be nothing to prove in the
classical and the intuitionistic case if the constant 0 were not allowed (only the
axioms involving 0 and its negation prevent us to draw conclusions like in the
neutral fragment). Now the proofs we look at with the wrong p(S) or Y(S) are in
fact interpreted in a coherent space with an empty web: all proofs of such
sequents are denotationally equal, and we therefore replace a proof by another
one with the same semantics!
(iii) The paper of Schellinx [4] investigates the faithfulness of the translation
intuitionistic I--+ linear and our proof is roughly inspired from this paper.

’ See forthcoming paper by J. Vauzeilles to appear in this journal


216 .I. -Y. Girard

Now, it remains to compare the systems LU restricted to various fragments


with the sequent calculi for the corresponding logics:
(i) The two intuitionistic fragments are OK: just translate T;T’ E ;A as r,r’ FA
and observe that all rules are correct. The other way around might be slightly
more delicate at least if we investigate cut-free provability.
(ii) The classical fragment translates not to LK, but to LC (see [3]); more
precisely besides the superficial difference one-sided/two sided, LC uses the
semi-colon in a different way: one tries to put as many formulas as possible in the
central zone: in particular, starting with T;T’ t A’;A, the idea is to move all
positive formulas from r to the right, and all negative formulas of A to the left:
with the result that T,A can consist of at most one formula.

Conclusion

As a matter of conclusion let us observe that this attempt at unification is


orthogonal to synchretic attempts of the style ‘logical framework’: too often
unification is at the price of a loss of structure (we lose properties: cut-
elimination, nay consistency). Here it goes the other way around. All fragments
considered are better as subsystems of LU than they were as isolated systems:
(i) Classical logic is handled by LC which is much better than LK.
(ii) The neutral intuitionistic fragment gets a legalislation of the notion of
headvariable and its normalisation procedure should be of the style ‘linear
head-reduction’.
(iii) The intuitionitic fragment gets a subtler approach to pattern-matching,
typically a denotationally associative disjunction.
(iv) Linear logic gets a smoother sequent calculus, in particular for exponential
connectives; this formulation has some similarties with the linear sequent calculi
proposed by Andreoli and Pareschi [l].
. . . Not to speak of the fact that all these systems are part of the same calculus,
i.e., are free to interact . . . .
There is of course the obvious question: is this LOGIC, i.e., did we catch here
all possible logical systems? Surely not, and there are additional parameters on
which one can play to broaden the scope of a unified approach to logic:
(i) The consideration of additional polarities: a polarity can be abstractly seen
as the permission to perform certain structural rules on the left or the right of a
sequent. Many other cocktails (from the absolute noncommutative polarity to
classical polarities) are possible, and all the combinations between weakening,
exchange and contraction yield up to 15 possible polarities. Most of these
combinations presumably make no sense, and one should not hurry to invent
polarities with no concrete application. However, if one insists on experimenting
in that way, it seems that a good criterion for the consideration of additional
On the unity of logic 217

polarities could be the possibility of extending the definition of disjunction so as


to preserve its denotational associativity.
(ii) The extension of these results to systems which have always been on the
border line of logic: systems of arithmetic, and more generally inductive
definitions.
(iii) The extension to second-order; in particular, it will be possible to cope
with the loss of subformula property by considering quantifications ranging on
various fragments (example: for all positive and classical (u).

NON SI NON LA

References

[l] J.-M. Andreoli and R. Pareschi, Linear objects: logical processes with built-in inheritance, to
appear in New Generation Computing (1991).
[2] J.-Y. Girard, Linear logic, Theoret. Comput. Sci. 50 (1987) l-102.
[3] J.-Y. Girard, A new constructive logic: classical logic, Preprint, universite Paris VII, March 1991,
to appear in Math. Structures Comput. Sci. l(3).
[4] H. Schellinx, Some syntactical observations on linear logic, Preprint 1990, to appear in J. Logic
and Computation, 1991.

You might also like