Girard - On the unity of logic
Girard - On the unity of logic
North-Holland
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.
1. Unified sequents
(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
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
Identity
A; c ;A
Structure
r;r’ c A’;A
a(r) ; Q’ (I-‘) c T’ (A’) ; r(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.
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; e- ;A,T O,l-‘; c ;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.
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.
Fig. 3.
212 J.-Y. Girard
Fig. 3 (contd.).
On the unity of logic 213
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.).
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.
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
Conclusion
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.