PROOF-NETS : THE PARALLEL

SYNTAX FOR PROOF-THEORY

Jean-Yves Girard
Institut de Mathématiques de Luminy, UPR 9016 – CNRS
163, Avenue de Luminy, Case 930, F-13288 Marseille Cedex 09

[email protected]

Abstract
The paper is mainly concerned with the extension of proof-nets to additives,
for which the best known solution is presented. It proposes two cut-elimination
procedures, the lazy one being in linear time. The solution is shown to be
compatible with quantifiers, and the structural rules of exponentials are also
accommodated.

Traditional proof-theory deals with cut-elimination ; these results are usually

obtained by means of sequent calculi, with the consequence that 75% of a cut-
elimination proof is devoted to endless commutations of rules. It is hard to be
happy with this, mainly because :
I the structure of the proof is blurred by all these cases ;

I whole forests have been destroyed in order to print the same routine lemmas ;

I this is not extremely elegant.

However old-fashioned proof-theory, which is concerned with the ritual ques-
tion : “is-that-theory-consistent ?” never really cared. The situation changed
when subtle algorithmic aspects of cut-elimination became prominent : typ-
ically the determinism of cut-elimination, its actual complexity, its imple-
mentation cannot be handled in terms of sequent calculus without paying a
heavy price. Natural deduction could easily fix the main drawbacks of cut-
elimination, but this improvement was limited to the negative fragment of
intuitionistic logic.
The situation changed in 1986 with the invention of linear logic : proof-nets
were introduced in [G86] as a new kind of syntax for linear logic, in order to

1
2 Jean-Yves Girard

cope with the problems arising from the intrinsic parallelism of linear sequent
calculus. 9 years later the technology is perfectly efficient and proof-nets are
now available for full linear logic 1 . Using implicit translations, proof-nets are
also available for classical and intuitionistic logics, i.e. for extant logical sys-
tems.
The essential result of [G86] was a sequentialization theorem showing the equiv-
alence of the new syntax with the traditional one. However this original result
was restricted to the multiplicative fragment MLL, non-multiplicative fea-
tures being accommodated by boxes, i.e. sequents under disguise. Later on,
the original sequentialisation proof was extended to quantifiers in [G88, G90] ;
the second version, the only satisfactory one, made a significant use of simpli-
fications of the method of [G86] discovered by Danos & Regnier, [DR88]. The
problem of the extension to full linear logic, and especially to the additives
(let us say to the fragment MALL of multiplicative-additive linear logic) has
remained open for years. In fact two distinct problems must be solved :

I To find the right notion of proof-net : the notion was found quickly -say in the
beginning of 1987- and since that time never really changed. The problem
is to cope with the &-rule of sequent calculus, for which a superimposition
of two proof-nets must be made ; by introducing for each &-link a boolean
variable which distinguishes between the two slices of the superimposition,
one eventually gets a notion of net in which formulas and links are weighted
by boolean polynomials. Although this notion remains the only serious
candidate, it is to be noted that it is far from being absolutely satisfactory :
this is because the question of determining the identity between two formulas
(or two links) in two different proof-nets cannot receive a satisfactory answer,
especially out of MALL.

I To prove sequentialisation : the case of quantifiers was considered as a pre-

liminary case, but the solution found in [G88] was too acrobatic to be ex-
tended. This solution was improved in [G90] by means of a new kind of
switchings, determined by certain formal dependencies. The solution im-
mediately extends to additives, provided certain weight dependencies are
forbidden (this is the dependency condition, which requires all weights to
be monomials). Since 1990 we have been trying to get rid of this technical
limitation. . . and we must admit that this was a stupid attitude : is spite
of their limitations our proof-nets obtained were extremely efficient (elim-
ination of boxes is always a big simplification), whereas getting rid of the
dependency restriction would of course make the proof-nets more intrinsic
in some cases, but would still not make them absolutely satisfactory. This
is the reason why we eventually decided to publish our partial results.
1. Except the additive neutrals
 Proof-Nets 3

We shall first develop the present version of additive proof-nets, which is

enough for most applications, together with a restricted (lazy) cut-elimination.
In an appendix we shall consider possible improvements and the general cut-
elimination procedure.
Of course, proof-nets are not intended for multiplicative and additives only. In
fact we shall devote some sections to the (unproblematic) extension to quan-
tifiers, and to the structural rules in the exponential case (here with limited
problems with weakening). Therefore we get rid of all boxes, but exponential
boxes. This is of course a tremendous improvement over traditional sequent
calculus.

1 PROOF-NETS FOR MALL

1.1 Proof-structures
Definition 1
A link L is an expression
P1 , . . . , P n
L
Q1 , . . . , Q m
involving n formulas (the premises of L) P1 , . . . , Pn and m formulas (the
conclusions of L) Q1 , . . . , Qm

ID −links : 0 premise 2 conclusions : A, A⊥

CUT −links : 2 premises : A, A⊥ 0 conclusion
⊗ −links : 2 premises : A, B 1 conclusion : A⊗B
& &
−links : 2 premises : A, B 1 conclusion : A B
⊕1 −links : 1 premise : A 1 conclusion : A⊕B
⊕2 −links : 1 premise : B 1 conclusion : A⊕B
& −links : 2 premises : A, B 1 conclusion : A&B
&
The premises of ⊗, , &-links are ordered : this means that we can dis-
tinguish a left premise (here A) and a right premise (here B). On the
other hand the premises of a CU T -link and the conclusions of an ID-link
are unordered.

Remark. — It is convenient to consider generalized axioms − A1 , . . . , An (n > 0),

which are interpreted by generalized axiom links (no premise, but ordered conclu-
sions A1 , . . . , An ). Such generalized axioms will occur in the proof of our main
theorem 2 ; they also occur when one wants to accommodate other styles of syntax,
which are foreign to the proof-net technology, in which case they are called boxes.
The idea of a box is that from the outside it looks like a generalized axiom, whereas
it has an inside which can be in turn another proof-net. A box freezes n formulas,
4 Jean-Yves Girard

and can therefore be seen as a sequent, the conclusion of a rule, whose premises are
proven in the box. Traditional sequent calculus is therefore a system of proof-nets
in which the only links are boxes, and all the improvement made in 9 years consist
in progressively restricting the use of boxes : in this paper boxes are limited to
exponential connectives (and to the neutral >).

Remark. — One should never speak of formulas, but of occurrences, which is ex-
tremely awkward. We adopt once for all the convention that all our formulas are
&
distinct (for instance by adding extra indices). In particular ID, ⊗, , ⊕, &-links
are determined by their conclusion(s), and a CU T -link is determined by its premises.

Definition 2
I If L is a &-link, with A & B as its conclusion, we introduce the eigen-

weight pL , which is a boolean variable. The intuitive meaning of p L is

the choice l/r between the two premises A and B of the link, pL for
“left”, i.e. A, ¬pL for “right”, i.e. B ; we use pL to speak of pL or ¬pL .
I If Θ is a structure involving the &-links L 1 , . . . , Lk (with associated

eigenweights p1 , . . . , pk ), then a weight (relative to Θ) is any element of

the boolean algebra generated by p1 , . . . , pk .
Definition 3
A proof-structure Θ consists of :
I A set of formulas (see the previous remark) ;
I A set of links ; each of these links takes its premise(s) and conclusion(s)
among the formulas of Θ ;
I For each formulas A of Θ, a weight w(A), i.e. a non-zero element of the
boolean algebra generated by the eigenweights p1 , . . . , pn of the &-rules
of Θ) ;
I For each link L of Θ, a weight w(L).
satisfying the following conditions :
I Each formula is the the premise of at most one link and the conclusion
of at least one link ; the formulas which are not premises of some link
are called the conclusions of Θ ;
P
I w(A) = w(L), the sum being taken over the set of links with conclu-
sion A ;
I if A is a conclusion of Θ, then w(A) = 1 ;
I if w is any element of the boolean algebra generated by the weights
occurring in Θ, and L is a &-link, then w.¬w(L) does not depend on
pL , i.e. belongs to the boolean algebra generated by the eigenweights
distinct from pL ;
 Proof-Nets 5

I if w is any weight occurring in Θ, then w is a monomial 1 pL1 . . . k pLk

of eigenweights and negations of eigenweights 2 ;
I w(L) 6= 0 ; moreover if L is any non-identity link, with premises A and
(or) B then
&
– if L is any of ⊗, , CU T , then w(L) = w(A) = w(B) ;
– if L is ⊕1 , then w(L) = w(A) ;
– if L is ⊕2 , then w(L) = w(B) ;
– if L is a &-link, then w(A) = w(L).pL and w(B) = w(L).¬pL , (hence
w(L) = w(A) + w(B)).

Remark. —
I Weights are in a boolean algebra, and therefore both algebraic and logical graphism
can be used ; here we decide to use the product notation (instead of the intersec-
tion), but we keep ¬w (instead of 1 − w) ; when we use the sum, we of course
mean the disjoint union, i.e. when I write w(L) = w(A) + w(B), I implicitly mean
that w(A).w(B) = 0.

I The technical condition “w.¬w(L) does not depend on pL ” says that the boolean
variable pL has no real meaning “outside w(L)” ; applying the condition to ¬w(L),
we see that w(L) does not depend on pL , in particular w(L).pL 6= 0.

I There are two ways to think of the dependency condition : either as a technical
restriction needed for the sequentialisation theorem (all our efforts to get rid of
it failed) or as a nice companion to the previous condition, since both are very
natural when a proof-structure is seen as a coherent space, see A.1.1.

1.2 Sequent calculus and proof-nets

Definition 4
Let Θ be a proof-structure and let L be either a CU T -link, or a link with
only one conclusion, which is in turn a conclusion of Θ and such that
w(L) = 1 ; we say that L is a terminal link of Θ. Given such a link, we
define the removal of L in Θ which consists (provided it makes sense) in
one or two proof-structures.

I If L is a ⊗-link (resp. a CU T -link) with premises A, B, and Γ, A ⊗ B

(resp. Γ) is the set of conclusions of Θ : the removal of L consists in
partitioning (if possible) the formulas of Θ distinct from A ⊗ B (resp.
the formulas of Θ) in two subsets X and Y , one containing A, the other
containing B, in such a way that, whenever a link L0 distinct from L
has a premise or a conclusion in X (resp. in Y ), then all other premises
2. This is the dependency condition
6 Jean-Yves Girard

and conclusions of L0 belong to X (resp. to Y ). The restrictions Θ/X

and Θ/Y are defined in an obvious way, and are proof-structures with
respective conclusions Γ0 , A and Γ00 , B. Observe that Γ = Γ0 , Γ00 .
& &
I If L is a -link with premises A, B, and Γ, A B is the set of con-
clusions of Θ : the removal of L consists in removing the conclusion
&
A B and the link L ; this induces a proof-structure with conclusions
Γ, A, B.
I If L is a ⊕1 -link with premise A, and Γ, A ⊕ B is the set of conclusions
of Θ : the removal of L consists in removing the conclusion A ⊕ B and
the link L ; this induces a proof-structure with conclusions Γ, A.
I If L is a ⊕2 -link with premise B, and Γ, A ⊕ B is the set of conclusions
of Θ : the removal of L consists in removing the conclusion A ⊕ B and
the link L ; this induces a proof-structure with conclusions Γ, B.
I If L is a &-link with premises A, B, and Γ, A&B is the set of conclusions
of Θ : the removal of L consists in first removing the conclusion A & B
and the link L (to get Θ0 ) and then forming two proof-structures ΘA
and ΘB :
– In Θ0 make the replacement pL = 1, and keep only those links L0
whose weight is still non-zero, together with the premises and con-
clusions of such links : the result is by definition ΘA , a proof-structure
with conclusions Γ, A.
– In Θ0 make the replacement pL = 0, and keep only those links L0
whose weight is still non-zero, together with the premises and con-
clusions of such links : the result is by definition ΘB , a proof-structure
with conclusions Γ, B.

Definition 5
A proof-structure Θ is sequentialisable when it can be reduced, by iterated
removal of terminal rules, to identity links. In more pedantic terms :

I An identity link is sequentialisable ;

I If the result of removing the terminal link L in Θ yields sequentialisable
proof-structures, then Θ is sequentialisable.
 Proof-Nets 7

Remark. —
I The removal of a given terminal link is not always possible, and its result is
not necessarily unique (however, for proof-nets, it would be easy to show, by a
connectivity argument, that the removal of a ⊗- or CU T -link is unique).

I Each removal step consists in the writing down of a rule of MALL ; therefore a
sequentialisable proof-structure has a sequentialisation, which consists in a proof
in MALL.

I Conversely, given a proof Π of − Γ in sequent calculus, one can build a proof-

structure Π◦ with conclusions Γ, such that Π is a sequentialisation of Π◦ . But
contrarily to the situation of our previous papers [G86, G88, G90] on proof-nets,
Π◦ is no longer unique. The problem is in the interpretation of a &-rule
− Γ, A − Γ, A
&
− Γ, A & B

applied to proofs Π1 and Π2 of − Γ, A and − Γ, B. We must find a proof-structure

Π◦ such that the removal of a terminal &-link yields Π◦1 and Π◦2 . The basic idea is
to merge the two proof-structures by means of the eigenweight pL : anything with
respective weights w1 and w2 in Π◦1 and Π◦2 will now get the weight pL .w1 +¬pL .w2
(we give the weight w1 = 0 to something which is absent in Π◦1 ). This simple idea
yields the following list of cases :

– If X is a formula or a link occurring in both of Π◦1 and Π◦2 , with the weights
w1 (X) and w2 (X), then X will occur in Π◦ with the weight
pL .w1 (X) + ¬pL .w2 (X) ; in particular the formulas of Γ occur in Π◦ with the
weight 1.
– If X is a formula or a link occurring in Π◦1 but not in Π◦2 , with the weight
w1 (X), then X will occur in Π◦ with the weight pL .w1 (X) ; in particular, A
will occur with the weight pL .
– If X is a formula or a link occurring in Π◦2 but not in Π◦1 , with the weight
w2 (X), then X will occur in Π◦ with the weight ¬pL .w2 (X) ; in particular, B
will occur with the weight ¬pL .

and to add to this merge the formula A & B together with the &-link L whose
premises are A, B and whose conclusion is A & B ; both L and A & B receive
the weight 1. But this is not as simple as it might seem : how do we know
that a formula or a link X of Π1 is the same as another formula or link Y of
Π2 ? There is no simple answer (except in some very specific cases, see discussion
A.1.4), and moreover, in cases where we feel entitled to make such identifications,
the resulting weight pL .w1 (X) + ¬pL .w2 (X) is not a monomial, which contradicts
the dependency condition. However there is at least the possibility to decide
that no identification between Π1 and Π2 is made, but for the conclusions, i.e.
the formulas of Γ ; by the way the sequent calculus formulation of the &-rule
8 Jean-Yves Girard

stipulates that the contexts of the two premises must be equal, hence this is a
clear case where there is no doubt as to the identification between a formula in
Π1 and a formula in Π2 .
Anyway, the main problem is to find a sequentialisation theorem ; this means
to give an intrinsic characterization of sequentialisable proof-structures. The
answer (a notion of proof-net) to this problem is important because we want to
carry our program of geometrization of proofs, introduced in [G87A] under the
name of geometry of interaction. Remember that geometry of interaction is
issued from the analysis of multiplicative proof-nets in terms of permutations,
[G86A]. In fact the redaction of this paper was postponed until a geometry of
interaction for additives was found, which is now the case, see [G94].

1.3 A wrong answer : slicing

Definition 6
Let ϕ be a valuation for Θ, i.e. a function from the set of eigenweights of
Θ into the boolean algebra {0, 1}, which induces a function (still denoted
ϕ) from the weights of Θ to {0, 1}. The slice ϕ(Θ) is obtained by restrict-
ing to those formulas A of Θ such that ϕ(w(A)) = 1, with an obvious
modification for the remaining &-links : only one premise is present.
The definition suggests a simple-minded criterion, (which is anyway an ap-
proximation to the real solution) : observe that (if we neglect unary links),
ϕ(Θ) is a multiplicative proof-structure. Therefore we can require ϕ(Θ) to be
multiplicatively correct, i.e. to be a multiplicative proof-net. This condition
is obviously necessary (in fact it is what we get by restricting definition 10 to
normal jumps).
But the condition cannot be sufficient : it corresponds to a separate treatment
of additives and multiplicatives, without real interaction between the condi-
tions. This would mean that all multiplicatives distribute over all additives :
for instance the obvious proof-structure with conclusions
& &
(A⊥ B ⊥ )⊕(A⊥ C ⊥ ) and A⊗(B&C) has two multiplicatively correct slices,
although it states a wrong instance of distributivity, and is not sequentialisable,
& &
whereas the obvious proof-structure with conclusions (A ⊥ B ⊥ ) & (A⊥ C ⊥ )
and A⊗(B ⊕C) (which has essentially the same slices) is sequentialisable. Our
ultimate criterion must therefore force the additive and multiplicative layers
to interact.

1.4 Proof-nets
Our basic idea will be to mimic our criterion of [G90] ; in this paper, certain
switchings for ∀-links were induced by the dependency of some formula upon
 Proof-Nets 9

the eigenvariable of the link.

Definition 7
Let ϕ be a valuation of Θ, let pL be an eigenweight ; we say that the
weight w (in Θ) depends on pL (in ϕ(Θ)) iff ϕ(w) 6= ϕL (w), where the
valuation ϕL is defined by :
I ϕL (pL ) = ¬(ϕ(pL ))
I ϕL (pL0 ) = ϕ(pL0 ) if L0 6= L.
A formula A of Θ is said to depend on pL (in ϕ(Θ)), if A is conclusion
of a link L0 such that ϕ(w(L0 )) = 1 and ϕL (w(L0 )) = 0. This basically
means that A and L0 are present in ϕ(Θ), but that changing the value
of the valuation for pL would make A (or at least L0 ) disappear from the
slice.
Definition 8
A switching S of a proof-structure Θ consists in :
I The choice of a valuation ϕS for Θ ;
&
I The selection of a choice S(L) ∈ l, r for all -links of ϕS (Θ) ;
I The selection for each &-link L of ϕS (Θ) a formula S(L), the jump of
L, depending on pL in ϕS (Θ). There is always a normal choice of jump
for L, namely the premise A of L such that ϕS (w(A)) = 1.
Definition 9
Let S be a switching of a proof-structure Θ ; we define the graph Θ S as
follows :
I The vertices of ΘS are the formulas of ϕS (Θ) ;
I For all ID-links of ϕS (Θ), we draw an edge between the conclusions ;
I For all generalized axiom links with conclusions A1 , . . . , An , we draw
an edge between A1 and A2 , etc., An−1 and An ;
I For all CU T -links of ϕS (Θ), we draw an edge between the premises ;
I For all ⊕-links of ϕS (Θ), we draw an edge between the conclusion and
the premise ;
I For all ⊗-links of ϕS (Θ), we draw an edge between the left premise and
the conclusion, and between the right premise and the conclusion ;
&
I For all -links L of ϕS (Θ), we draw an edge between the premise (left
or right) selected by S(L) and the conclusion ;
I For all &-links L of ϕS (Θ), we draw an edge between the jump S(L) of
L and the conclusion.
10 Jean-Yves Girard

Definition 10
A proof-structure Θ is said to be a proof-net when for all switchings S,
the graph ΘS is connected and acyclic.

We immediately get the

Theorem 1
If Θ is sequentialisable, then Θ is a proof-net.

Proof. — The proof is straightforward and uninteresting. 2

1.5 The sequentialisation theorem

Theorem 2
Proof-nets are sequentialisable.

We shall devote the remainder of the subsection to the proof of the theorem.
We shall make some simplifying hypotheses :

1.5.1 Empires
In this subsection and in the next one, we fix one for all a valuation ϕ and we
concentrate on the slice ϕ(Θ), where Θ is a given proof-net. All switchings S
considered are such that ϕS = ϕ.

Definition 11
If S is a switching and A is a formula of ϕ(Θ), then we modify the graph
ΘS by deleting (in case A is premise of a link L) any edge induced by
L and connecting A to another formula B. (There is at most such an
edge, connecting A with the conclusion of L or with the other premise of
L if L is a CU T -link.) This determines a partition of ΘS into at most
two connected components, and the one containing A is denoted Θ A S . We
T A
define the empire of A as eA := S ΘS , the intersection being taken over
all switchings S.

Lemma 1
Empires are imperialistic ; this means that, as soon B1 and B2 are linked
by L and B1 ∈ eA, then B2 ∈ eA, with only three exceptions :
I B1 is A and is a premise of L : if B2 is the conclusion of L, it is not in
eA ;
&
I L is a -link, B2 is the conclusion of the link, B1 is one of the two
premises of L and the other premise of L is not in eA : in such a case
the conclusion is never in eA ;
 Proof-Nets 11

I L is a &-link, B2 is the conclusion of the link, B1 is the premise of L

(remaining in ϕ(Θ)) and there is a formula C in ϕ(Θ) depending on p L
which is not in eA : in such a case the conclusion is never in eA.

Proof. — straightforward ; see e.g. [G90], lemma 1. We sketch the proof for
the sake of self-containment :

I one first remarks that if B1 is a conclusion of L, then B2 ∈ eA ; the typical

&
example is that of a -link. If S is a switching inducing two components,
and such that B2 6∈ ΘAS , then we see (perhaps by changing our mind about
A
S(L)) that B1 6∈ ΘS . The other cases are handled in the same way. Observe
that the same argument shows that if the conclusion B1 of a &-link L is in
eA, so is any formula whose weight depends on pL .
&
I it only remains to show that if L is a - or &-link and all possible choices
for S(L) are in eA, so is the conclusion B of L, provided B 6= A. But any
switching S must connect B with one of these formulas which are in Θ A S,
A
and this forces B ∈ ΘS . 2

Lemma 2
There is an S such that eA = ΘA
S ; S is called a principal switching for
eA.
&
Proof. — S is obtained as follows : in case A is the premise of a - or
&-link, set S so as to draw an edge between A and the conclusion of the link ;
&
the other and &-links are switched as follows :
&
I if L is a -link with its conclusion C not in eA, then one of its premises,
let us say B is not in eA by lemma 1 and we set S so as to draw an edge
between B and C ;

I if L is a &-link L with its conclusion C not in eA, then a formula of ϕ(Θ),

let us say B, depending on pL , is not in eA by lemma 1 and we set S so as
to draw an edge between B and C.

It is immediate that eA = ΘA
S. 2

Lemma 3
Let A and B be distinct formulas in ϕ(Θ) and assume that B 6∈ eA ; then

I if A ∈ eB, then eA ⊂ eB ;
I if A 6∈ eB, then eA ∩ eB = ∅.
12 Jean-Yves Girard

Proof. — see [G90], lemma 3. We also sketch the proof : we define a prin-
cipal switching of eB as in lemma 2 by adding another constraint : if L is a
&
or & whose conclusion is in eB, then in case A is a premise of L set S so
as to connect the conclusion with A, and (otherwise) if some possible choice
for S(L) is not in eA, make this very choice. Now if a formula C belongs to
eA ∩ eB, then C is connected with B (which is not in eA) inside Θ B S = eB,
which means that some edge induced by S inside eB links a formula in eA
with a formula outside eA. This is only possible (because of lemma 1) when
the formula of eA is A. This proves the second part of the claim. If A ∈ eB,
A must be the premise of a link L whose conclusion is not in eA, and S(L)
has been set so as to get two connected components : ΘAS and its complement.
A B A B
It is immediate that ΘS ⊂ ΘS hence eA ⊂ ΘS ⊂ ΘS = eB. 2

Definition 12
A formula B of eA is said to be a door of eA iff

I Either it is the premise of a link L whose conclusion does not belong to

eA ;
I Or it is a conclusion of Θ. Obviously A is a door of eA, the main door ;
the other doors are called auxiliary doors. The set of doors of eA is the
border of eA.

Lemma 4
Let C be an auxiliary door of eA which is not a conclusion of Θ ; then C
&
is the premise of a - or a &-rule.

Proof. — immediate from lemma 1. 2

1.5.2 Maximal empires

The hypotheses are the same as in the previous section ; but we now assume
that A is the conclusion of a &-link L and that eA is maximal w.r.t. inclusion
among similar empires (i.e. that if B is the conclusion of a &-link and eA ⊂ eB,
then A = B).

Lemma 5
Let L0 be any &-link and let B, B 0 ∈ ϕ(Θ) be such that B, B 0 depends on
pL0 , and assume that B ∈ eA ; then B 0 ∈ eA.

Proof. — let w be the weight of the link whose conclusion is B and such that
ϕ(w) = 1 ; then the technical condition “w.¬w(L0 ) does not depend on pL0 ”
ensures that ϕ(¬w(L0 )) = 0, hence if C is the conclusion of L0 , that C ∈ ϕ(Θ).
 Proof-Nets 13

By lemma 1 B ∈ eC, hence eA ∩ eC 6= ∅ ; by maximality eA ⊂ eC is impossi-

ble, hence by lemma 3 we get C ∈ eA. Now by lemma 1 this implies B 0 ∈ eA. 2

Lemma 6
If C is a border formula of eA and L0 is any &-link, then ϕL0 (w(C)) = 1,
i.e. the border formulas are still present in ϕ0L (Θ).

Proof. — C is either a conclusion, in which case w(C) = 1 or C is a premise

of some rule L00 : let B be the conclusion of L00 (in case L00 is a cut, let B be
the other premise of L00 ). Then B 6∈ eA but since B depends on L0 , lemma 5
yields B ∈ eA, a contradiction. 2

1.5.3 Stability of maximal empires

The purpose of this section is to show that if A is the conclusion of a &-link and
is maximal among similar empires w.r.t. a given ϕ, then it remains maximal
w.r.t. any other choice of ϕ. It will be enough to start with a given ϕ and to
show that A is still maximal w.r.t. ϕL . We can only prove this essential fact
under the dependency condition, which has the following consequence :

Lemma 7
Assume that ϕ(w(A)) = 1 and that (w.r.t. ϕ) w(A) depends neither on
pL and pM ; then this remains true w.r.t. ϕL .

Proof. — the monomial w(A) cannot make use of pL etc. 2

This is wrong for non monomials, typically p ∪ q depends neither on p nor on q

if ϕ(p) = ϕ(q) = 1, but depends on q if ϕ(p) = 0. This phenomenon is exactly
the familiar failure of stability in the case of the parallel or.

Lemma 8
Assume that B ∈ eA (w.r.t. ϕ) and that w(B) does not depend on p L ;
then w.r.t. ϕL we still have B ∈ eA.

Proof. — it takes nothing to assume that A is not a conclusion (in which

case eA is ϕ(Θ) and everything is trivial) ; so one can fix a formula C with
the same weight as A such that C cannot belong to eA (take C to be the
conclusion of the link of which A is a premise, or the other premise if this link
is L a CU T ). Let D be the conclusion of the link L, then :
I either D ∈ eA, hence by lemma 5, the formulas outside eA do not depend on
pL , hence by lemma 7, changing ϕ to ϕL does not alter any dependency out-
side. Assume that a switching S has been chosen, (w.r.t. ϕL ) with B ∈ ΘA
S,
14 Jean-Yves Girard

0
C 6∈ ΘAS . Then if we define a switching S (w.r.t. ϕ) essentially by making
the same choices outside eA (which is possible, since the dependencies are
unaffected), then B and C are still connected, hence we still have B ∈ Θ A
S0 ,
contradicting the hypothesis.

I or D 6∈ eA, hence by lemma 5, the formulas inside eA do not depend on p L ,

hence by lemma 7, changing ϕ to ϕL does not alter any dependency inside.
Assume that a switching S has been chosen, (w.r.t. ϕL ) with B 6∈ ΘA S . Then
0
if we define a switching S (w.r.t. ϕL ) essentially by making the same choices
inside eA (which is possible, since the dependencies are unaffected), then B
and A are still not connected, hence we still have B 6∈ ΘA S 0 , contradicting
the hypothesis. 2

Lemma 9
eA is maximal w.r.t. ϕL .

Proof. — assume that (w.r.t. ϕL ) eA ⊂ eB, with eB maximal. Then (still

w.r.t. ϕL ) A ∈ eB but B 6∈ eA. Using the maximality of eA w.r.t. ϕ, of
eB w.r.t. ϕL and lemma 8, we get (w.r.t. ϕ) A ∈ eB and B 6∈ eA, which
contradicts the maximality of eA w.r.t. ϕ. 2

1.5.4 Proof of the theorem

Proof. — by induction on the number n of &-links in Θ :
I if n = 0, then we are basically in the multiplicative case and we are done. To
be precise, we argue by induction on the number m of links of the proof-net :

– if m = 1, then the proof-net consists in a single axiom link and we are

done ;
– if m > 1, then by connectedness Θ must contain some link which is not
an axiom, hence some terminal link (see definition 4) exists, and
&
∗ if there is a terminal or ⊕i -link, then we can remove it, thus getting
another proof-net (immediate) with a smaller value of m, to which the
induction hypothesis applies ; then Θ is sequentialisable ;
∗ if all terminal links are ⊗- or CU T -links, with premises A i , Bi , choose
one of these premises (say A1 ) such that eA1 is maximal inside the eAi
and eBi . If C is a border formula of eA1 and C is not a conclusion of Θ,
then C stands (hereditarily) above some Ai or Bj , let us say above B2 .
Then by lemma 1, C ∈ eB2 , hence by lemma 3 eA1 ⊂ eB2 , contradicting
the maximality of eA1 . This shows that the border of eA1 only consists
of conclusions. But then it is easy to see that ΘA S is always equal to
1
 Proof-Nets 15

eA1 (apply lemma 1), hence ΘB S is always equal to eB1 . This means
1

that the removal of our terminal link induces a splitting of the proof-net
into two connected components, which are obviously proof-nets and to
which the induction hypothesis applies ; Θ is therefore sequentialisable.

I if n > 0, then choose ϕ (e.g. ϕ(pL ) = 1 for all L), and let A be a conclusion of
some &-link such that eA is maximal w.r.t. ϕ. Now by lemma 9 eA remains
maximal w.r.t. to any ϕ0 , and its border remains the same by lemma 6.
Moreover, if L is any & link whose conclusion B belongs to eA w.r.t. ϕ, B
still belongs to eA w.r.t. any ϕ0 such that ϕ0 (w(L)) = 1 (this is another use
of the dependency condition : since w(L) is a monomial, is is possible to
write ϕ0 as ϕL1 ...Lk for a suitable sequence L1 . . . Lk with the property that
all intermediate valuations ϕL1 ...Li yield the same value ϕL1 ...Li (w(L)) = 1.
Then lemma 8 is enough to conclude.) This means that a given eigenweight
pL cannot occur both in eA for some valuation and in the complement of
eA for some other valuation. In other terms, we can split the eigenweights
into two disjoint groups, those who may occur in eA (group I) and those
who may occur in its complement (group II). Now let us introduce two new
proof-nets :

– a proof-net Θ0 whose conclusions are those of the (constant) border of eA ;

a formula B is in Θ0 when B ∈ eA for some valuation. This immediately
induces a proof-structure (take those links whose conclusion is in Θ 0 , with
the weight they had in Θ). Θ0 contains all occurrences of eigenvariables
from group I, and no occurrence from group II. It is immediate that Θ 0 is
a proof-net. Moreover Θ0 has a terminal &-link and removing this link as
in definition 4 induces two proof-nets (immediate) to which the induction
hypothesis applies, and which are therefore sequentialisable. Then Θ 0 is
sequentialisable. In case A is a conclusion of Θ, then Θ = Θ0 and we are
done ; otherwise we introduce
– a proof-net Θ” consisting of those formulas B such that B 6∈ eA for some
valuation ϕ0 such that ϕ0 (w(B)) = 1 and of the border formulas of eA
which are not conclusions (there are some). The links of Θ” are all the
links of Θ whose conclusion is outside eA and a new axiom link whose
conclusions consist in the border of eA. The links and formulas coming
from Θ receive the same weight, whereas the new axiom link is weighted
1. Θ” contains all occurrences of eigenvariables from group II, and no
occurrence from group I. Now it is an easy exercise in graph theory to
show that Θ” is a proof-net. The induction hypothesis applies and Θ” is
therefore sequentialisable. By the way observe that if we replace our new
axiom link in Θ” by the proof-net Θ0 , the result is the original Θ.
16 Jean-Yves Girard

The sequentialisation of Θ is obtained by taking the sequentialisation of Θ”

and replacing the sequent calculus axiom corresponding to the new axiom
link with the sequentialisation of Θ0 . 2

1.6 Cut-elimination in MALL (lazy procedure)

There is a Church-Rosser cut-elimination procedure for additive proof-nets,
which enjoys the subformula property. For practical uses this procedure is
challenged by a coarser one, lazy cut-elimination, which can be proved to
achieve the same result in some important cases. The more general procedure
is described in appendix A.1.3.

1.6.1 Lazy cut-elimination

Lazy cut-elimination is concerned with only a restricted kind of cut :

Definition 13
A cut-link L is said to be ready iff

I the weight of the cut-link is 1

I both premises of the cut are the conclusion of exactly one link (these
links are therefore also of weight 1)

Theorem 3
Let Θ be a proof-net whose conclusions do not contain the connective &
and without ready cut ; then Θ is cut-free.

Proof. — if Θ contains no &-link, we are done, since all weights are 1. Oth-
erwise, choose a &-link L in such a way that the empire eA of its conclusion
is maximal among similar empires (w.r.t. any valuation). Then A is weighted
1 as well his (hereditary) conclusions. Below A sits a terminal link L 0 with
no conclusion (otherwise A would be a subformula of this conclusion, and &
would occur in a conclusion), hence L0 is a cut of weight 1. In fact A sits
hereditarily above the premise B of L which is in turn the conclusion of a link
of weight 1 (either L or a link ”below” L). Now the premise B ⊥ of L0 might
be the conclusion of several links, in which case there is a valuation ϕ and
a &-link L00 (with conclusion C) such that B ⊥ depends on pL00 . In the slice
ϕ(Θ) observe that B ⊥ 6∈ eA, but A ∈ eC (this is because B ⊥ ∈ eC and so
B ∈ eC and then A ∈ eC by lemma 1). But then eA ⊂ eC by lemma 3, a
contradiction. 2
 Proof-Nets 17

Remark. —

I The theorem therefore establishes that the lazy procedure which only removes
ready cuts is enough in the absence of additive connectives. In fact since the
general procedure extends the lazy one and is Church-Rosser, the lazy procedure
yields the same result in this important case.

I The theorem still holds for full linear logic : if a proof-net is without ready cuts
and its conclusions mention neither & nor existential higher order quantifiers,
then it is cut-free. The restriction on existentials is just to forbid the hiding of a
& by a ∃-rule, which can be the case with higher order.

Definition 14
Let L be a ready cut in a proof-net Θ, whose premises A and A⊥ are
the respective conclusions of links L, L0 . Then we define the result Θ0 of
reducing our cut in Θ :

I If L is an ID-link then Θ0 is obtained by removing in Θ the formulas

A and A⊥ (as well as the cut-link and L) and giving a new conclusion
to L0 : the other conclusion of L, which is another occurrence of A⊥
(ID-reduction).
&
I If L is a ⊗-link (with premises B and C) and L0 is a -link (with
⊥ ⊥ 0
premises B and C ), then Θ is obtained by removing in Θ the for-
mulas A and A⊥ as well as our cut links and L,L0 and adding two new
cut links with respective premises B,B ⊥ and C,C ⊥ (⊗-reduction).
I If L is a &-link (with premises B and C) and L0 is a ⊕1 -link (with
premise B ⊥ ), then Θ0 is obtained in three steps : first we remove in
Θ the formulas A and A⊥ as well as our cut link and L,L0 ; then we
replace the eigenweight pL by 1 and keep only those formulas and links
that still have a nonzero weight : therefore B and B ⊥ remain with
weight 1 whereas C disappears ; finally we add a cut between B and
B ⊥ (⊕1 -reduction).
I If L is a &-link (with premises B and C) and L0 is a ⊕2 -link (with
premise C ⊥ ), then Θ0 is obtained in three steps : first we remove in
Θ the formulas A and A⊥ as well as our cut link and L,L0 ; then we
replace the eigenweight pL by 0 and keep only those formulas and links
that still have a nonzero weight : therefore C and C ⊥ remain with
weight 1 whereas B disappears ; finally we add a cut between B and
B ⊥ (⊕2 -reduction).

Proposition 1
If Θ0 is obtained from a proof-net Θ by lazy cut-elimination, then Θ 0 is
still a proof-net.
18 Jean-Yves Girard

Proof. — we consider all possible cut-elimination steps :

I ID-reduction : the only important thing to remark is that the other con-
clusion of L (which is an occurrence of A⊥ ) cannot be the same occurrence
(otherwise the structure would bear a cycle).

I ⊗-reduction : let us fix a valuation ϕ, and let X be the set ϕ(Θ), that we
can write as the union Y ∪ Z of ϕ(Θ0 ) and {B, C, A, A⊥ , B ⊥ , C ⊥ }, so that
Y ∩ Z = {B, C, B ⊥ , C ⊥ }. Given a switching S of Θ0 , it induces a graph
Θ0S on Y and we can consider its subgraph G (on the same Y ) in which the
edges between B, B ⊥ and C, C ⊥ have been removed, as well as the subgraph
H (on Y ∩ Z) consisting of these two edges : we have Θ0S = G ∪ H. We
want to show that Θ0S is connected and acyclic. In order to do this, we
”extend” our switching S to a switching S 0 of Θ ; it is immediate that ΘS 0
can be written as G ∪ H0 , where H0 is a graph on Z with one edge between
B, A, one edge between C, A, one edge between A, A⊥ , one edge between
A⊥ , B ⊥ (or between A⊥ , C ⊥ , depending on the switching of L0 ). Since G ∪H0
is connected and acyclic, then G has 3 connected components, and each of
them meets Y ∩ Z. Now B, C are not in the same component (otherwise
there would be a cycle in G ∪ H0 ; B, B ⊥ are neither in the same component,
since, if we switch L0 to “left”, we get a cycle in G ∪ H0 ; the same is true,
for symmetrical reasons, of B, C ⊥ and C, C ⊥ . Then B ⊥ , C ⊥ must lie in the
same component. From this it is immediate that G ∪H, i.e. Θ0S , is connected
and acyclic.

I (⊕ − 1-reduction) : a valuation ϕ for Θ0 can be extended to a valuation ϕ0

for Θ by setting ϕ0 (pL ) = 1 (and giving any value to the other eigenweights
which might have disappeared when making the cut-elimination step). One
can also extend a switching S of Θ0 into a switching S 0 of Θ by setting
S(L) = B. Then it is almost immediate that ΘS 0 and Θ0S are of the same
type.

I (⊕ − 2-reduction) : symmetrical to the previous case. 2

Definition 15
The size of a proof-net is defined to be the number of its links.

Theorem 4
Lazy cut-elimination converges to a unique lazy normal form (i.e. a proof-
net without ready cuts) in a time which is linear in the size of the proof-
net.
 Proof-Nets 19

Proof. — unicity of the normal form is an easy consequence of the fact that
our procedure is Church-Rosser. To prove termination, observe that some of
the eigenweights receive definite values pL = 0 or pL = 1 during the cut-
elimination of Θ. Let ϕ be a valuation of Θ which extends these choices. Now,
if Θ 7→ Θ1 7→ . . . 7→ Θn is a reduction sequence, we can observe the slices ϕ(Θ),
ϕ(Θ1 ), . . . , ϕ(Θn ), and observe that the number of links in each of these slices
is strictly decreasing. Then n cannot be bigger than the the number of links in
ϕ(Θ), hence, n is smaller than the number of links in Θ. Lazy normalization
therefore takes a linear number of steps, which is not quite the same as linear
time. However, it is quite easy to see how to transform this into a linear time
algorithm : we keep track of the values taken by the eigenweights and we delay
the substitutions occurring in the additive reductions, i.e. we perform them
only in case they yield values 0 (in which case we may erase) or 1 . . . When
the process is completed, then we perform the remaining substitutions. 2

Remark. — Only a limited part of the correctness criterion is actually used to prove
proposition 1 ; the most important part is devoted to the absence of deadlock, i.e.
theorem 3.

2 THE CASE OF QUANTIFIERS

Quantifiers have been treated in previous papers [G88, G90], hence we shall
be brief. Roughly speaking the role of eigenweights is taken by eigenvariables,
and everything is adapted mutatis mutandis.

2.1 Proof-structures
∀ −links : 1 premise : A[e/x] 1 conclusion : ∀xA
∃t −links : 1 premise : A[t/x] 1 conclusion : ∃xA

∀-links make use of eigenvariables ; for each ∀-link L there is a specific variable
eL which is associated with this link. Each existential link comes with the name
of a term, namely the term t of the premise of the ∃-link. This remark would
be pure pedantism if we had not to take care of the case of fake dependencies
(i.e. when x does not occur in A) where we cannot recover t from the premise.
We shall say that a formula A depends on an eigenvariable eL when
I either A is the premise of L

I or eL occurs in A

I or A is the premise of a link ∃t and eL occurs in t

20 Jean-Yves Girard

Proof-structures are defined as expected i.e. as in definition 3 :

I if L is a quantifier link with premise A, then w(L) = w(A) ;

I we require that w(A).¬w(L) = 0 for any formula A depending on e L ;

I we require that no eigenvariable occurs in a conclusion.

One easily defines what it means for a sequent calculus proof to be a sequen-
tialisation of a given proof-structure, by extending definition 4 :

I If L is a ∀-link with premise A[eL /x], and Γ, ∀xA is the set of conclusions of
Θ : the removal of L consists in removing the conclusion ∀ and the link L and
replace everywhere eL with a fresh variable e ; this induces a proof-structure
with conclusions Γ, A[e/x].

I If L is a ∃t -link with premise A[t/x], and Γ, ∃xA is the set of conclusions of

Θ : the removal of L consists in removing the conclusion ∃xA and the link
L ; this induces a proof-structure with conclusions Γ, A[t/x]. Observe that
this removal might be impossible, typically if some eigenvariable occurs in
t.

2.2 Proof-nets
Definitions 8 and 9 are adapted as follows :

I S selects for each ∀-link L of ϕS (Θ) a formula S(L), the jump of L, depending
on pL in ϕS (Θ). There is always a normal choice of jump for L, namely the
premise A of L.

I In ΘS we draw an edge between the conclusion and the premise of any ∃-

link, and for all ∀-links L of ϕS (Θ), we draw an edge between the jump S(L)
of L and the conclusion of L.

The condition for being a proof-net is exactly defined as in definition 10. We

have only to check the extension of theorems 1 and 2. We only give some
indications as to the proof of the latter :

I We shall mainly be concerned with a maximal empire eA where A is the

conclusion of either a & or a ∀-link.

I If eA is maximal, then its border formulas depend on no eigenvariable

I If eA is maximal and if B, B 0 ∈ ϕ(Θ) are such that e0L occurs in B, B 0 ; then

B ∈ eA implies B 0 ∈ eA.
 Proof-Nets 21

2.3 Cut-elimination in MALLq

Lazy cut-elimination is defined by adding a case to definition 14 :
I If L is a ∀-link (with premises B[eL /x]) and L0 is a ∃t -link (with premise
B ⊥ [t/x]), then Θ0 is obtained in three steps : first we remove in Θ the
formulas A and A⊥ as well as our cut link and L,L0 ; then we replace the
eigenvariable pL by t ; finally we add a cut between B[t/x] and B ⊥ [t/x]
(∃t -reduction).
The procedure might not work for general proof-structures : if e L actually
occurs in t, then substituting t for eL changes B ⊥ [t/x] into B ⊥ [t0 /x] which
does not match B[t/x], and also eL is still present. But the proof-net condition
forbids this (just set S so that S(L) = B ⊥ [t/x]).
It is of course essential to prove that the proof-net condition is preserved
through cut-elimination, which essentially amounts looking at our new step :
given a switching S of Θ0 , we want to back it up into a switching S 0 of Θ,
which offers no difficulty as to L (set S 0 (L) = B[eL /x]) ; but this might be
problematic for some ∀-link L0 , since a formula C may not depend on e0L ,
whereas C[t/eL ] does, and if S(L0 ) = C[t/eL ], we cannot set S 0 (L0 ) = C.
However observe that :
I in such a case, e0L must occur in t, and this forces (w.r.t. Θ) B[t/x] ∈ eA0 ,
hence ∀xA ∈ eA0 , where A0 is the conclusion of L0 . But since B[t/x] 6∈ eA,
we also get A0 6∈ eA, hence eA ⊂ eA0
I C must also depend on eL , hence C ∈ eA, which implies C ∈ eA0
I but it is immediate that the illegal jump S 0 (L0 ) = C to an element of eA0
will not alter the correctness criterion.
We can therefore prove that the proof-net condition is preserved by induction
on the number of illegal jumps (with an induction loading : we consider proof-
nets in which fake dependencies have been declared). The basis step being
trivial, the induction step (n + 1 illegal jumps) consists in choosing L 0 , A0 , C
as above, and to modify our proof-net by introducing a fake dependency i.e.
to pretend that e0L occurs in C. This induces another proof-net (as observed
above) in which our jump is now legal. There are only n illegal jumps in
this new proof-net, and the induction hypothesis applies, yielding a connected
acyclic graph.

2.4 Complexity of cut-elimination

As before, cut-elimination involves a linear number of steps. This is clearly
not enough for linear time normalization, since one of these steps is a substi-
tution, whose iteration is exponential. However, the algorithm remains linear
22 Jean-Yves Girard

if we do not perform the substitutions, i.e. we keep the eigenvariables in the

proof-net and we use an auxiliary stack to remember which term should be sub-
stituted for them. This process is particularly efficient in the following case :
we normalize a proof of A without & or ∃ ; then the lazy cut-elimination will
eventually yield a cut-free proof (in which some substitutions have not been
performed) in linear time. Now, imagine that we perform the delayed substi-
tutions in a given formula B of our net : we get a subformula B 0 of A which is
completely determined by the choice of the links below B, hence we can forget
the substitutions and directly write this subformula.

3 EXPONENTIALS
We shall deal with only two rules, which are central in the study of expo-
nentials, namely weakening and contraction. The point of the introduction of
exponentials is precisely to allow weakening and contraction for certain for-
mulas, those which are prefixed by ?. This is not enough, and in the standard
version of linear logic, two additional rules, promotion and dereliction, are
added. We now know that the choice of additional rules is much more open,
and we shall therefore ignore them. As a consequence, it will be impossible to
speak of cut-elimination, and we shall concentrate on correctness.
Our basic ingredient will be the notion of a discharged formula (terminology
strictly inspired from natural deduction). Besides usual formulas we shall al-
low in a proof-net discharged ones, denoted [A]. These formulas are handled
in a very specific way :
I If [A] is the premise of L, then L is a ?-link (see below).
I If [A] is a conclusion of the proof-net, we no longer require its weight to be 1.
We therefore modify definition 10 as follows : a non-discharged conclusion
has weight 1.
I We shall assume that [A] is the conclusion of (unspecified) generalized ax-
ioms (i.e. boxes). This will handle all the additional rules we know, but
usual dereliction 3 .
I Since discharged formulas are bound to be merged by ?-links (see below),
there is no need to superimpose such formulas, hence we require that dis-
charged formulas are the conclusions of a unique link.
The only link we consider is the n-ary link ?n , with n unordered premises, which
are all occurrences of the same discharged formula [A], and one conclusion,
3. Usual dereliction easily fits into our pattern : just allow any formula which is the
conclusion of a link to be discharged.
 Proof-Nets 23

namely ?A. The case n = 0 is allowed, and accounts for weakening. The
condition for being a proof-structure is the following : the weight increases,
i.e. if [Ai ] is any premise of L, then w([Ai ]) 6 w(L). The condition for being
a proof-net depends on an additional datum : for any ?-link L, a default
jump L is chosen : L can be any formula B (discharged or not) such that
w(L) 6 w(B). Given a valuation ϕ, a switching S will select a jump for each
?-link L : this jump may be the default one L or any premise [A i ] of L such
that ϕ(w([Ai ])) = 1. We then state the usual connected-acyclic condition.
One must introduce a sequent calculus with two kinds of formulas, usual ones
and discharged ones. This calculus is identical to the usual but for :
I the &-rule : the contexts may differ on some discharged formulas. From
− Γ, [∆], A and − Γ, [∆0 ], B, deduce − Γ, [∆], [∆0 ], A & B.
I the new rule of weakening/contraction : from − Γ, [A], . . . , [A], deduce
− Γ, ?A.
Sequentialisation is an easy exercise.

A APPENDIX
A.1 More about additives
A.1.1 An alternative to weights
Let Θ be a proof-net ; then we can define a structure of coherent space, on the
set consisting of all formulas and all links of Θ : X _ ^ Y iff w(X).w(Y ) 6= 0.
Now, due to the fact that weights are monomials, as soon as X1 , . . . , Xn are
pairwise coherent, the intersection w(X1 ) ∩ . . . ∩ w(Xn ) is nonzero. This means
that maximal cliques in the coherent space correspond to slices. This also
means that we can replace the weighting of Θ by the coherent space structure 4 .
This alternative presentation has many advantages, in particular that we have
not to name the eigenweights. This becomes crucial with the technique of
spreading that we now introduce, since a spreading may introduce new eigen-
weights.

A.1.2 Spreading of a proof-net

Let Θ be a proof-net, let A be a formula in Θ and let p be an eigenweight.
The spreading of Θ above A w.r.t. p consists in replacing every formula or
link hereditarily above A (including A) and whose weight does not depend on
p by two copies, X1 and X2 . This induces a new coherent space structure :
I Xi _
^ Yj iff i = j and X _
^ Y in Θ
4. In this way the dependency condition looks very natural
24 Jean-Yves Girard

I X1 _
^ Y iff w(X).w(Y ).p 6= 0 in Θ
I X2 _
^ Y iff w(X).w(Y ).¬p 6= 0 in Θ
I X_
^ Y iff X _
^ Y in Θ
The resulting coherent space needs not be a proof-structure. However, in the
case we shall use this construction, this will be the case, and the resulting
coherent space will indeed be a proof-net.

A.1.3 Cut-elimination
Let us say that a cut is almost ready, when both premises are the conclusions
of a unique link. For almost ready cuts there is an obvious cut-elimination,
the same as in the ready case. We shall now explain how to reduce a general
cut to almost ready ones.
Assume that the cut is between A and A⊥ , with the same monomial weight w ;
let w1 , . . . , wm be the weights of the links with conclusion A, and w10 , . . . , wn0
the weights of the links with conclusion A⊥ , so that
w = w1 + . . . + wm = w10 + . . . + wn0 ; we assume that the cut is not at most
ready, so m + n > 2. Let us say that an eigenweight p splits A when we can
partition w1 , . . . , wm into two non-empty subsets such that the partial sums
are respectively equal to w.p and w.¬p ; as soon as m > 1 there is at least one
splitting for A.
I if the eigenweight p is a common splitting for A and A⊥ , then we can du-
plicate A into two copies of respective weights w.p and w.¬p, the same for
A⊥ , and replace our cut with two cuts
I otherwise, there is a splitting, of say, A⊥ w.r.t. an eigenweight p ; we spread
Θ above A w.r.t. p and we are back to the previous case. (The spreading
makes sense mainly because A⊥ ∈ eA w.r.t. any valuation, hence that all
formulas above A⊥ are in eA).
This procedure is Church-Rosser and terminating. It should be considered
when theoretical questions (subformula property etc.) are at stake. Its com-
putational value is limited, since iterated spreadings may induce exponentially
many duplications.

A.1.4 Discussion
We would like to discuss several issues connected with additive proof-nets.
Contrarily to the multiplicative case, the extant solution is not perfect (al-
though it has the virtue to exist). Let us discuss the weaknesses of our solution
and potential improvements :
the main question which occurs when translating sequent calculus to proof-nets
is the problem of superimposition. There are two difficulties :
 Proof-Nets 25

A.1.5 Weights
Weights must be monomials. However, weights of the form p ∪ q will naturally
occur if we want to allow more superimpositions. The present state of affairs
is as follows :

I in spite of years of efforts, I never succeeded in finding the right correctness

criterion for these more liberal proof-nets

I general boolean coefficients might be delicate to represent (on the other

hand, the case we consider has a natural presentation in terms of coherent
spaces)

I normalization in the full case might be messy

A.1.6 Identity of formulas

Imagine that we have no problem with weights, and that we try to maximize
the identifications. One idea is to adopt a binary ⊕-rule, which is quite natural.
Besides that, we immediately stumble on a lot of problems :

I if I have a cut on A in both proofs, should I superimpose them. . . worse, if

both proofs have two cuts on A, which ones should be superimposed ? The
same question occurs with the contraction rule : how do we superimpose
two contraction rules (remember that the premises are indiscernible).

I what about existential witnesses ? Should one superimpose two portions of

proofs with the same structure, but different witnesses ?

I what about normalization ? During this process, distinct portions of proofs

might become equal, hence it would be necessary to superimpose them. . .

These limitations do not apply if we restrict to cut-free proofs in MALL, and

if we make the extra assumption that all identity links are atomic : there is a
well defined notion of net (provided one can fix the problem of weights) which
enjoys the maximum number of identifications.
If identification is difficult, its converse is easy, i.e. there is not the slightest
problem to forget that two formulas are equal. This could indicate a possible
theoretical way out, namely considering a proof-net as the set of its slices. The
computational value of this idea is limited (exponential growth of the net), but
this might be valuable for theoretical considerations. However, we have no idea
how to define a correctness condition for such sets of slices.
There are other problems connected to normalization :
26 Jean-Yves Girard

A.1.7 Normalization
Multiplicative proof-nets have a local cut-elimination, and this is still true for
quantifiers, provided we do not compute the existential witnesses. The addi-
tive case involves a real global move, which consists in setting an eigenweight
to 0 or 1, and erase everything with weight 0. This is rather brutal, and
completely foreign to the parallel asynchronous spirit of proof-nets. In [G94]
(section 3.4.) a variant of usual sequent calculus is introduced, with a local
cut-elimination, i.e. the erasing is performed in a lazy way, which means that
some useless ”beards”, which are bound to be erased, are still hanging. The
calculus with these beards is connected with the additive neutral > (which has
still no satisfactory treatment), and we can expect that there is a notion of
additive bearded proof-net with a local elimination. This is perhaps the most
promising open question in this paper.

A.2 Multiplicative neutrals

There are two multiplicative neutrals, 1 and ⊥, and two rules, the axiom − 1
and the weakening rule : from − Γ, deduce − Γ, ⊥. Both rules are handled by
means of links with one conclusion and no premise ; however ⊥-links are treated
like 0-ary ?-links, i.e. they must be given a default jump. Sequentialisation is
immediate.
At first sight, cut-elimination is unproblematic : replace a cut between the
conclusions 1 and ⊥ of zero-ary links with. . . nothing. But we notice a new
problem, namely that a cut formula A can be the default jump of a ⊥-link
L, and we must therefore propose another jump for L. Usually one of the
premises of the link with conclusion A works (or the jump of L 0 if A is the
conclusion of a ⊥-link) works. Worse, this new jump is by no ways natural (if
A is B ⊗ C, the new jump can either be B or C), which is quite unpleasant. As
far as we know, the only solution consists in declaring that the jumps are not
part of the proof-net, but rather of some control structure. It is then enough
to show that at least one choice of default jump is possible. This is not a very
elegant solution : we are indeed working with equivalence classes of proof-nets
and if we want to be rigorous we shall have to endlessly check that such and
such operation does not depend on the choice of default jumps. In practice
one can be rather sloppy. . .
Of course everything would be nicer without any default jump. But then a
proof-net for a multiplicative combination A of occurrences of 1 and ⊥ would
basically be nothing more than A itself : the correctness criterion for proof-
nets without jumps encompasses the decision problem for such combinations,
and this problem is known to be NP-complete by [LW92]. . . The existence of
a correctness criterion of the same style as the familiar ones is therefore very
 Proof-Nets 27

unlikely 5 .
This discussion is fully relevant to the exponential case : as soon as we start to
normalize exponential cuts, then the same problems (with the same solution)
arise.

A.3 Additive neutrals

There is still no satisfactory approach to additive neutrals, which are fortu-
nately extremely uninteresting in practice. The only way of handling > is by
means of a box or, if one prefers, by means of a second order translation : on
this Kamtchatka of linear logic, the old problems of sequent calculus are not
fixed. The absence of a satisfactory treatment of > calls for another notion of
proof-net. . . presumably a solution to the wider question of bearded proof-nets.

BIBLIOGRAPHY
[DR88] Danos, V. & Regnier, L. : The structure of multiplicatives,
Archive for Mathematical Logic 28.3, pp. 181-203, 1989

[G86] Girard, J.-Y. : Linear Logic, Theoretical Computer Science 50.1,

pp. 1-102, 1987

[G86A] Girard, J.-Y. : Multiplicatives, Rendiconti del Seminario Matem-

atico dell’Università e Policlinico di Torino, special issue on Logic
and Computer Science, pp. 11-33, 1987

[G87A] Girard, J.-Y. : Towards a Geometry of Interaction, Categories in

Computer Science and Logic, Contemporary Mathematics 92, AMS
1989, pp. 69-108

[G88] Girard, J.-Y. : Quantifiers in linear logic, Temi e prospettive

della logica e della filosofia della scienza contemporanee, CLUEB,
Bologna, 1989.

[G90] Girard, J.-Y. : Quantifiers in linear logic II, Nuovi problemi della
logica e della filosofia della scienza, CLUEB Bologna 1991.

5. There is still a possibility, namely to work with the necessary condition acyclic graph
with n + 1 connected components, where n is the number of ⊥-links ; but we would have to
check that this condition is preserved through cut-elimination (which seems likely), and we
would have to restrict to proof-nets in which the formula ⊥ cannot occur in the absence of
cuts, so that our necessary condition eventually becomes sufficient. . .
28 Jean-Yves Girard

[G94] Girard, J.-Y. : Geometry of interaction III : the general case,

Proceedings of the Workshop on Linear Logic, MIT Press, submit-
ted.

[LMSS90] Lincoln, P. & Mitchell, J. & Scedrov, A. & Shankar N. :

Decision Problems in Propositional Linear Logic, Annals of Pure
and Applied Logic 56,1992, pp. 239-311.

[LW92] Lincoln, P. & Winkler, T. : Constant-only Multiplica-

tive Linear Logic is NP-complete, Theoretical Computer Science
135,1994,pp.155-159.