Introduction to linear logic

Emmanuel Beffara

To cite this version:

Emmanuel Beffara. Introduction to linear logic. Master. Italy. 2013. �cel-01144229�

HAL Id: cel-01144229

https://hal.archives-ouvertes.fr/cel-01144229
Submitted on 21 Apr 2015

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.

Distributed under a Creative Commons Attribution - NonCommercial - NoDerivatives| 4.0

International License
Introduction to linear logic
Emmanuel Beffara
August 29, 2013

Abstract. is manuscript is the lecture notes for the course of the same title I gave at the Summer school
on linear logic and geometry of interaction that took place in Torino in August 2013. e aim of this course
is to give a broad introduction to linear logic, covering enough ground to present many of the ideas and
techniques of the field, while staying at a (hopefully) accessible level for beginners. For this reason, most
technical development is carried out in the simple multiplicative fragment, with only hints at generaliza-
tions. As prerequisites, some knowledge of classical sequent calculus and some knowledge of the 𝜆-calculus
is useful.

1 e proof-program correspondence . . . . . . . . . . . . . . . . . . . 2
1.1 e Curry-Howard isomorphism . . . . . . . . . . . . . . . . . . . 2
1.2 Denotational semantics . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Linearity in logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Linear sequent calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Multiplicative linear logic . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Cut elimination and consistency . . . . . . . . . . . . . . . . . . . 8
2.3 One-sided presentation . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Full linear logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 e notion of fragment . . . . . . . . . . . . . . . . . . . . . . . . 16
3 A bit of semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1 Provability semantics . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Proof semantics in coherence spaces . . . . . . . . . . . . . . . . . 18
4 A bit of proof theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1 Intuitionistic and classical logics as fragments . . . . . . . . . . . . 19
4.2 Cut elimination and proof equivalence . . . . . . . . . . . . . . . . 21
4.3 Reversibility and focalization . . . . . . . . . . . . . . . . . . . . . 22
5 Proof nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.1 Intuitionistic LL and natural deduction . . . . . . . . . . . . . . . . 23
5.2 Proof structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Correctness criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 30

Reference material on the topics discussed here include

Girard, Lafont, and Taylor, Proofs and types is a book evolved from lecture notes of a graduate course
of the same title. It is a good reference for an introduction to the proof-program correspondence,
although it does not cover all topics of the maer.
Girard, “Linear Logic” is the historical paper introducing linear logic. It is an unavoidable reference,
although not the best way to discover the topic nowadays, since many aspects have been beer
understood since then.
Girard, “Linear Logic: Its Syntax and Semantics” is an updated and more accessible presentation, writ-
ten about ten years later.

1
 1 e proof-program correspondence
Logic is the study of discourse and reasoning. Mathematical logic is the study of mathematical dis-
course and reasoning. e need for mathematical logic historically arose at the end of the nineteenth
century with the search for foundations of mathematics, at a time when abstract mathematics reached
an unprecedented level of complexity and stumbled upon paradoxes. Proof theory is the sub-field of
mathematical logic concerned with the nature and meaning of mathematical statements and proofs.
Gödel’s incompleteness theorem and Gentzen’s cut-elimination method are the first major results that
contributed in its definition in the 1930s. e following years saw the progressive apparition of the
central role of computation in logic, up to the formulation of the Curry-Howard isomorphism in the
1960, as briefly presented below. is correspondence is now a central notion both in proof theory and
in theoretical computer science as it addresses the central question of both fields, which is consistency:
• A logical system is consistent if it is not degenerate, i.e. if it does not prove everything. en one
can search for a meaning of proofs, and the next level of consistency is if there are statements for
which there are different proofs, hence different ways of proving things.
• Consistency in a formal language for computation refers to the ability, by structural considera-
tions like typing, to make sure that a program behaves well (no deadlocks, no infinite loop). is
in turn implies the definition of the meaning of programs.
is search for meaning is known as semantics and is the meeting point between logic and computation.
Linear logic is one of the outcomes of the study of semantics and the interaction between logic and
computer science. See Girard, Lafont, and Taylor, Proofs and types, as a more detailed introduction to
this topic.

1.1 e Curry-Howard isomorphism

e clearest formulation of the proof-program correspondence, also known as the Curry-Howard iso-
morphism, is obtained through the simply typed 𝜆-calculus. We present it here briefly, in its extension
with conjunction types.
In the following, we will assume that one is given, once and for all, a set of propositional variables
(ranged over by Greek leers 𝛼, 𝛽, …) and a set of term variables (ranged over by Latin leers 𝑥, 𝑦, 𝑧…).
1 Definition. Formulas of implicative-conjunctive propositional logic are defined by the following gram-
mar:
𝐴, 𝐵 ∶= 𝛼 propositional variables
𝐴 ⇒ 𝐵 implication
𝐴 ∧ 𝐵 conjunction
Terms of the simply-typed 𝜆-calculus with pairs are defined by the following grammar:

𝑡, 𝑢 ∶= 𝑥 variable
𝜆𝑥𝐴 .𝑡 abstraction, i.e. function
(𝑡)𝑢 application
⟨𝑡, 𝑢⟩ pairing
𝜋𝑖 𝑡 projection, with 𝑖 = 1 or 𝑖 = 2

where 𝑥 in 𝜆𝑥𝐴 .𝑡 is a bound variable. Terms are considered up to injective renaming of bound variables,
assuming all bound variables in a given term are distinct from all free variables.
A typing context is a sequence 𝑥1 ∶ 𝐴1 , …, 𝑥𝑛 ∶ 𝐴𝑛 where the variables 𝑥1 , …, 𝑥𝑛 are all distinct. A
typing judgment consists in a typing context Γ, a 𝜆-term 𝑡 and a formula 𝐴 and is wrien Γ ⊢ 𝑡 ∶ 𝐴. A
term has type 𝐴 in a context Γ if the judgment Γ ⊢ 𝑡 ∶ 𝐴 can be proved using the rules of table 1.

2
 ax Γ, 𝑥 ∶ 𝐴 ⊢ 𝑡 ∶ 𝐵 Γ⊢𝑡∶𝐴⇒𝐵 Γ⊢𝑢∶𝐴
Γ, 𝑥 ∶ 𝐴 ⊢ 𝑥 ∶ 𝐴 ⇒I ⇒E
Γ⊢ 𝜆𝑥𝐴 .𝑡 ∶𝐴⇒𝐵 Γ ⊢ (𝑡)𝑢 ∶ 𝐵
Γ⊢𝑡∶𝐴 Γ⊢𝑢∶𝐵 Γ⊢𝑡∶𝐴∧𝐵 Γ⊢𝑡∶𝐴∧𝐵
∧I ∧E1 ∧E2
Γ ⊢ ⟨𝑡, 𝑢⟩ ∶ 𝐴 ∧ 𝐵 Γ ⊢ 𝜋1 𝑡 ∶ 𝐴 Γ ⊢ 𝜋2 𝑡 ∶ 𝐵

Table 1: Typing rules for the simply typed 𝜆-calculus

In the rules of table 1, if we forget everything that is on the le of colons, i.e. all 𝜆-terms and
variables, we get a deduction system for implicative-conjunctive propositional logic. is system is
known as intuitionistic natural deduction and called NJ (actually NJ refers to the whole natural deduction
system for intuitionistic logic, of which this is a simple subsystem). If we read 𝐴1 , …, 𝐴𝑛 ⊢ 𝐵 as “under
hypotheses 𝐴1 , …, 𝐴𝑛 , I can prove 𝐵”, then this system is correct.
Now if we look only at 𝜆-terms, we get a minimal functional language. An abstraction 𝜆𝑥𝐴 .𝑡 intu-
itively represents the function which, to an argument 𝑥 of type 𝐴, associates the value of 𝑡. Computation
in this language is formalized by a reduction relation ⇝ defined by two simple rules which apply any-
where in a term:
(𝜆𝑥.𝑡)𝑢 ⇝ 𝑡[𝑢/𝑥] 𝜋𝑖 ⟨𝑡1 , 𝑡2 ⟩ ⇝ 𝑡𝑖 for 𝑖 = 1 or 𝑖 = 2

where 𝑡[𝑢/𝑥] is the operation of replacing every occurrence of the variable 𝑥 in 𝑡 by the term 𝑢.
e fact that computation is correct, i.e. respects types and produces no infinite loops, is formulated
by the following theorems, of which we will provide no proof in this introduction.
2 eorem (Subject reduction). If Γ ⊢ 𝑡 ∶ 𝐴 holds and 𝑡 ⇝ 𝑢 then Γ ⊢ 𝑢 ∶ 𝐴 holds.
3 eorem (Confluence). For any pair of reductions 𝑡 ⇝∗ 𝑢 and 𝑡 ⇝∗ 𝑣 of a term 𝑡, there exists a term 𝑤
and a pair of reductions 𝑢 ⇝∗ 𝑤 and 𝑣 ⇝∗ 𝑤.
4 eorem (Strong normalization). A typable term has no infinite sequence of reductions.
eorems 3 and 4 together imply that any 𝜆-term has a unique irreducible reduct and that any
sequence of reduction steps eventually reach it.
In an irreducible typed term, by definition of reduction, there is never an introduction rule (for either
⇒ or ∧) followed by an elimination rule for the same connective. Proofs with this property will be called
normal and they have the following crucial property:
5 eorem (Subformula property). In a normal proof, any formula occurring in a sequent at any point in
the proof is a subformula of one of the formulas in the conclusion.
e reduction operation from the 𝜆-calculus, interpreted on proofs, is a normalization procedure
that computes the normal form of an arbitrary proof. e effect of this, as illustrated by the subformula
property, is that a proof is made explicit in the process, in other words it transforms an arbitrary proof
into a direct proof, without lemmas. Indeed, a lemma for a theorem is a statement that is usually not a
part of the theorem’s statement but is proved as an intermediate step for proving the theorem.

1.2 Denotational semantics

Reduction in the 𝜆-calculus can be seen as a computation method in a world of “pure” functions, indeed
the 𝜆-calculus was initially introduced by Church in order to define a pure theory of functions, in a way
analogous to set theory which describes a world made entirely of sets. It is natural to ask whether this
theory has a model where the terms are actually interpreted as functions and denotational semantics is
precisely the study of such interpretations. By the Curry-Howard isomorphism, this will also interpret

3
 proofs as functions, for instance a proof of 𝐴 ⇒ 𝐵 will actually be a function from proofs of 𝐴 to proofs
of 𝐵. e key ingredients to make this concrete are the following:
• types/formulas are interpreted by spaces (oen as particular sets with additional structure, more
generally by objects in a suitable category),
• terms/proofs are interpreted as morphisms between such spaces,

• reduction preserves the interpretation of terms.

e last point is the most important one: if syntactic reduction actually computes something, then
reducing means geing closer to the final syntactic form of the computed object, but the object itself
does not change. We will describe here a particular model of the simply-typed 𝜆-calculus known as
coherence spaces.¹
6 Example. is simplest and most naive model is obtained by interpreting formulas as plain sets and
terms as (set-theoretic) functions. Propositional variables are arbitrary sets, 𝐴 ⇒ 𝐵 is interpreted by the
set of functions from 𝐴 to 𝐵. is models perfectly well but it does not have much interest because it is
not informative about what proofs and terms mean. at is because there are way too many morphisms,
most of which exhibit behaviours that one could not be defined using terms. A symptom is that this
model fails when considering second order quantification (i.e. quantification over types, also known as
polymorphism), because of cardinality problems.
7 Definition. A coherence space 𝐴 is a set |𝐴| (its web) and a symmetric and reflexive binary relation ¨𝐴
(the coherence). A clique in 𝐴 is a subset of |𝐴| of points pairwise related by ¨𝐴 . e set of cliques of
𝐴 is wrien 𝐶ℓ(𝐴).
e idea is that |𝐴| is a set of possible bits of information about an object of type 𝐴 and the coherence
relation ¨𝐴 indicates which bits can possibly describe aspects of the same object. A clique is a set of
mutually coherent bits, hence a (possibly partial) description for some object.
8 Example. We could represent a type of words over some alphabet using a web |𝐴| made of pairs (𝑖, 𝑥)
that mean “at position 𝑖 there is a leer 𝑥” and pairs (𝑛, $) that mean “at position 𝑛 is the end-of-string
symbol”. Coherence will be defined as (𝑖, 𝑥) ¨ (𝑗, 𝑦) if either 𝑖 = 𝑗 then 𝑥 = 𝑦 or 𝑖 < 𝑗 and 𝑥 ≠ $ or 𝑖 > 𝑗
and 𝑦 ≠ $.
9 Definition. Given coherence spaces 𝐴 and 𝐵, a stable function from 𝐴 to 𝐵 is a function 𝑓 from 𝐶ℓ(𝐴)
to 𝐶ℓ(𝐵) that is
• continuous: if (𝑎𝑖 )𝑖∈𝐼 is a directed family in 𝐶ℓ(𝐴), then 𝑓 (⋃𝑖∈𝐼 𝑎𝑖 ) = ⋃𝑖∈𝐼 𝑓 (𝑎𝑖 );

• stable: for all 𝑎, 𝑎′ ∈ 𝐶ℓ(𝐴) such that 𝑎 ∪ 𝑎′ is a clique, 𝑓 (𝑎 ∩ 𝑎′ ) = 𝑓 (𝑎) ∩ 𝑓 (𝑎′ ).

ese conditions imply monotonicity: the more information we have on the input, the more infor-
mation we have on the output. Continuity means that the value for an infinite input can be deduced
from the values for its finite approximations. Stability is more subtle, it means that if a bit is available
in the output, then there is a minimal set of bits in the input needed to get it.
10 Definition. e trace of a stable function 𝑓 from 𝐴 to 𝐵 is the set

𝑇𝑟􏿴𝑓 􏿷 ∶= 􏿺(𝑎, 𝛽) 􏿖 𝛽 ∈ 𝑓 (𝑎) ∧ ∀𝑎′ ⊊ 𝑎, 𝛽 ∉ 𝑓 (𝑎′ )􏿽.

where 𝑎 and 𝑎′ are cliques in the coherence space 𝐴 and 𝛽 is a point in the coherence space 𝐵.
¹is model was first introduced under the name of “binary qualitative domains” as a semantics for System F, i.e. the simply
typed 𝜆-calculus extend with quantification over types, in Girard, “e system F of variable types, fieen years later”.

4
 In each pair (𝑎, 𝛽) in 𝑇𝑟􏿴𝑓 􏿷, 𝑎 is the minimal input (a clique in 𝐴) needed to produce 𝛽 (a point in
𝐵). It can be checked that this trace fully defines the stable function, hence there is a bijection between
stable functions and traces.
Furthermore, it appears that traces are cliques in a suitable coherence space whose web is made of
all pairs of a finite clique in 𝐴 and a point in 𝐵. Coherence between two pairs simply expresses that
they may be part of the same stable function. is coherence space is how we interpret the implication
𝐴 ⇒ 𝐵. An interpretation of simply typed 𝜆-terms as cliques can be deduced from these considerations.
11 Definition. A stable function 𝑓 is linear if for all (𝑎, 𝛽) ∈ 𝑇𝑟􏿴𝑓 􏿷, 𝑎 is a singleton.
In other words, for producing one bit of information in output, the function needs one bit of in-
formation in input. is notion corresponds, in a way that can be made very precise, to the fact the
𝑓 uses its argument exactly once in order to produce each bit of information in output. is is a very
computational property, and through the Curry-Howard correspondence we can turn it into a logical
notion of linearity.

1.3 Linearity in logic

Classical sequent calculus (and intuitionistic sequent calculus as well) derives sequents of the form
Γ ⊢ Δ using le and right introduction rules for each logical connective, plus the special rules for
axiom and cut. It contains particular rules for handling the sets of hypotheses and conclusions, known
as the weakening and contraction rules:
Γ⊢Δ Γ⊢Δ Γ, 𝐴, 𝐴 ⊢ Δ Γ ⊢ 𝐴, 𝐴, Δ
wL wR cL cR
Γ, 𝐴 ⊢ Δ Γ ⊢ 𝐴, Δ Γ, 𝐴 ⊢ Δ Γ ⊢ 𝐴, Δ
ese rules allow each hypothesis to be used any number of times and it is crucial for the expressiveness
of the logic. Note that, in the intuitionistic case (as in NJ) there is always one formula on the right of ⊢
hence weakening and contraction are relevant only in the le-hand side, i.e. the hypotheses.
e rest of classical sequent calculus has two presentations, depending on the treatment of the
context of the active formula in each introduction rule. For instance, the right introduction rule for
conjunction ∧ has two flavours:
Γ ⊢ 𝐴, Δ Γ ⊢ 𝐵, Δ Γ ⊢ 𝐴, Δ Γ ′ ⊢ 𝐵, Δ′
∧Ra ∧Rm
Γ ⊢ 𝐴 ∧ 𝐵, Δ Γ, Γ ′ ⊢ 𝐴 ∧ 𝐵, Δ, Δ′
e first form is called additive and the second one is called multiplicative. For the le introduction
rules, there are similarly two versions:
Γ, 𝐴 ⊢ Δ Γ, 𝐵 ⊢ Δ Γ, 𝐴, 𝐵 ⊢ Δ
∧La1 ∧La2 ∧Lm
Γ, 𝐴 ∧ 𝐵 ⊢ Δ Γ, 𝐴 ∧ 𝐵 ⊢ Δ Γ, 𝐴 ∧ 𝐵 ⊢ Δ
Similar variants exist for the other connectives. e multiplicative and additive variants of each rule are
equivalent in the presence of weakening and contraction, however they have different properties when
studying the structure of proofs.
In the intuitionistic case, weakening and contraction are not necessary if the system is presented in
additive form, as in table 1. However, a multiplicative form could be defined, with explicit contraction
and weakening, and it would bear more precise information about the handling of variables in 𝜆-terms.
Moreover, it is the key to the definition of logical linearity: a proof of Γ, 𝐴 ⊢ 𝐵 is linear in 𝐴 if the 𝐴 in
conclusion is not involved in a contraction or weakening anywhere in the proof.
In some sense, the point of linear logic is to make this notion of linearity explicit in the language of
formulas, and to refine logical deduction accordingly.

5
 Structure
Γ, 𝐴, 𝐵, Γ ′ ⊢ Δ Γ ⊢ Δ, 𝐴, 𝐵, Δ′
exL exR
Γ, 𝐵, 𝐴, Γ ′ ⊢Δ Γ ⊢ Δ, 𝐵, 𝐴, Δ′
Identity

ax Γ ⊢ Θ, 𝐴, Δ Γ ′ , 𝐴, Θ ′ ⊢ Δ′
𝐴⊢𝐴 cut
Γ, Γ ′ , Θ ′ ⊢ Θ, Δ, Δ′
Negation
Γ, 𝐴 ⊢ Δ Γ ⊢ 𝐴, Δ
⊥R ⊥L
Γ⊢ 𝐴⊥ , Δ Γ, 𝐴⊥ ⊢ Δ
Conjunction
Γ ⊢ 𝐴, Δ Γ ′ ⊢ 𝐵, Δ′ Γ, 𝐴, 𝐵 ⊢ Δ
⊗R ⊗L
Γ, Γ ′ ⊢ 𝐴 ⊗ 𝐵, Δ, Δ′ Γ, 𝐴 ⊗ 𝐵 ⊢ Δ
Disjunction
Γ ⊢ 𝐴, 𝐵, Δ & Γ, 𝐴 ⊢ Δ Γ ′ , 𝐵 ⊢ Δ′ &
& R & L
Γ⊢𝐴 𝐵, Δ Γ, Γ ′ , 𝐴 𝐵 ⊢ Δ, Δ′

Table 2: MLL sequent calculus in two-sided presentation

2 Linear sequent calculus

2.1 Multiplicative linear logic
We now define the smallest possible system in linear logic. It has only the multiplicative form of con-
junction and disjunction and propositional variables as atomic formulas. Richer systems, including the
additive form of connectives, units (the analogue of the true and false statements), quantifiers or modal-
ities will be defined later. is minimal system is called MLL, for multiplicative linear logic.
12 Definition. e language of formulas of MLL is defined by the following grammar:

𝐴, 𝐵 ∶= 𝛼 propositional variable
𝐴⊥ linear negation — read “𝐴 orthogonal” or “not 𝐴”
𝐴⊗𝐵 multiplicative conjunction — read “𝐴 tensor 𝐵”
&
𝐴 𝐵 multiplicative disjunction — read “𝐴 par 𝐵”

13 Definition. A two-sided MLL sequent is a pair (Γ, Δ) of possibly empty lists of MLL formulas. Such a
pair is usually wrien Γ ⊢ Δ and read “Γ entails Δ”.
14 Definition. A sequential² two-sided MLL proof is a proof tree built using the rules of table 2. A sequent
Γ ⊢ Δ is provable in MLL if there exists some proof 𝜋 whose conclusion is Γ ⊢ Δ; we write this fact as
𝜋 ∶ Γ ⊢ Δ. By extension, we say that a formula 𝐴 is provable if the sequent ⊢ 𝐴 is provable.
e exchange rules (“exL” on the le and “exR” on the right) state that the order in which the for-
mulas occur in the sequents is irrelevant, only the side on which they occur maers. e axiom rule,
²e word “sequential” conveniently refers to sequent calculus, but it is also used in contrast to proof nets, defined below in
section 5, which are more parallel in nature.

6
 abbreviated as “ax”, is in multiplicative form, with an empty context on both sides. e cut rule is also
multiplicative: contexts from both premises are concatenated. All other rules are introduction rules for
each connective, they have a name that ends with “L” for le rules or “R” for right rules, depending on
the side of the entailment symbol ⊢ on which they introduce a formula in conclusion.
15 Example. e following proof establishes the commutativity of multiplicative conjunction:
ax ax
𝐵⊢𝐵 𝐴⊢𝐴
⊗R
𝐵, 𝐴 ⊢ 𝐵 ⊗ 𝐴
exL
𝐴, 𝐵 ⊢ 𝐵 ⊗ 𝐴
⊗L
𝐴⊗𝐵⊢𝐵⊗𝐴
16 Exercise. Prove that the following sequents are provable in MLL:
&
• multiplicative excluded middle: ⊢ 𝐴 𝐴⊥
& &
• semi-distributivity of tensor over par: 𝐴 ⊗ (𝐵 𝐶) ⊢ (𝐴 ⊗ 𝐵) 𝐶
17 Proposition. Let Γ ⊢ Δ be an MLL sequent. An occurrence of a propositional variable 𝛼 in Γ ⊢ Δ is called
positive if it occurs under an even number of negations in Δ or under an odd number of negations in Γ. If
Γ ⊢ Δ is provable in MLL, then each propositional variable has an equal number of positive and negative
occurrences.
Proof. It is clear that this property holds in the conclusion of an axiom rule and that it is preserved by
all other logical rules.
is linearity property allows to easily recognize as not provable any formula that violates this
&
condition, for instance 𝐴⊥ (𝐴 ⊗ 𝐴) is not provable in MLL.
18 Definition. Two formulas 𝐴 and 𝐵 are linearly equivalent if the sequents 𝐴 ⊢ 𝐵 and 𝐵 ⊢ 𝐴 are provable
in MLL. is fact is wrien 𝐴 ˛ 𝐵.
19 Example. In the proof of example 15, if we replace 𝐴 with 𝐵 and vice-versa, we get a proof of 𝐵 ⊗ 𝐴 ⊢
𝐴 ⊗ 𝐵. As a consequence, 𝐴 ⊗ 𝐵 and 𝐵 ⊗ 𝐴 are linearly equivalent.
20 Exercise. Prove that the linear equivalence of two formulas 𝐴 and 𝐵 is equivalent to the fact that, for
all lists of formulas Γ and Δ, the sequent Γ ⊢ 𝐴, Δ is provable if and only if the sequent Γ ⊢ 𝐵, Δ is
provable.
21 Exercise. Define the system MLL′ as the logical system MLL extended with a new connective 𝐴 ⊸ 𝐵
called linear implication and read “𝐴 implies 𝐵”. e system of inference rules is extended with the
introduction rules for this new connective:
Γ, 𝐴 ⊢ 𝐵, Δ Γ ⊢ 𝐴, Δ Γ ′ , 𝐵 ⊢ Δ′
⊸R ⊸L
Γ ⊢ 𝐴 ⊸ 𝐵, Δ Γ, Γ ′ , 𝐴 ⊸ 𝐵 ⊢ Δ, Δ′
&
Prove that, in this system, for any 𝐴 and 𝐵, the formulas 𝐴 ⊸ 𝐵 and 𝐴⊥ 𝐵 are linearly equivalent.
22 As consequence of this fact, in MLL, it is customary to not consider linear implication 𝐴 ⊸ 𝐵 as a
&
proper connective, but merely as a notation for 𝐴⊥ 𝐵. Nevertheless, this notation is widely used since
it is very meaningful with respect to the proof system, because it carries essentially the same logical
meaning as the entailment symbol ⊢. Indeed, the sequent 𝐴 ⊢ 𝐵 is provable if and only if the formula
𝐴 ⊸ 𝐵 is provable. is also justifies the notation 𝐴 ˛ 𝐵 for linear equivalence.

7
 2.2 Cut elimination and consistency
Sequent calculus was originally introduced by Gentzen in order to prove the consistency or arithmetic,
and this method was later adapted to many logical systems as a standard way to establish consistency
results. e main lemma for consistency, sometimes called Hauptsatz (which means main result in
German), is the cut elimination property, stating that any proof can be transformed into a proof that
does not use the cut rule.³ Linear logic also enjoys this property, and we will prove it now for MLL.
e idea is to start from an arbitrary proof with cut rules and to rewrite each instance of the cut rule
in order to push it upwards to the axiom rules, eventually eliminating all cuts.
We are (for now) working with a very strict system in which reordering the hypotheses and conclu-
sions of a proof takes explicit operation using exchange rules. e le and right exchange rules allow
single transpositions of formulas on either side of a sequent; since transpositions generate all permuta-
tions, it is clear that any permutation of the formulas in a sequent can be implemented by some sequence
of le and right exchange rules. We will write
Γ⊢Δ
ex
Γ ′ ⊢ Δ′
for any such sequence of rules, assuming Γ ′ is a permutation of Γ and Δ′ is a permutation of Δ, when
the precise order of these exchange rules is irrelevant.
23 Definition. e cut elimination relation ⇝ is the congruent binary relation over proofs generated by
the following rules:
• If the active formula is in the context of the last rule of one of the premisses, then this last rule is
exchanged with the cut rule. For instance, if the le premiss ends with a right tensor rule and the
active formula is in the context of the le premiss of this rule, we have

𝜋1𝑏 ∶ Γ 1𝑏 ⊢ 𝐵, Θ 1 , 𝐴, Δ 1𝑏 𝜋1𝑐 ∶ Γ 1𝑐 ⊢ 𝐶, Δ 1𝑐
⊗R
Γ 1𝑏 , Γ 1𝑐 ⊢ 𝐵 ⊗ 𝐶, Θ 1 , 𝐴, Δ 1𝑏 , Δ 1𝑐 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
cut
Γ 1𝑏 , Γ 1𝑐 , Γ 2 , Θ 2 ⊢ 𝐵 ⊗ 𝐶, Θ 1 , Δ 1𝑏 , Δ 1𝑐 , Δ 2
⇝
𝜋1𝑏 ∶ Γ 1𝑏 ⊢ 𝐵, Θ 1 , 𝐴, Δ 1𝑏 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
cut
Γ 1𝑏 , Γ 2 , Θ 2 ⊢ 𝐵, Θ 1 , Δ 1𝑏 , Δ 2 𝜋1𝑐 ∶ Γ 1𝑐 ⊢ 𝐶, Δ 1𝑐
⊗R
Γ 1𝑏 , Γ 2 , Θ 2 , Γ 1𝑐 ⊢ 𝐵 ⊗ 𝐶, Θ 1 , Δ 1𝑏 , Δ 2 , Δ 1𝑐
ex
Γ 1𝑏 , Γ 1𝑐 , Γ 2 , Θ 2 ⊢ 𝐵 ⊗ 𝐶, Θ 1 , Δ 1𝑏 , Δ 1𝑐 , Δ 2
e other cases are similar, we allow such commutations for any rule in the premisses, including
the cut rule itself.
• If the active formula is introduced by an exchange rule in one of the premisses of the cut, then
this exchange rule may be dropped. An instance of this is

𝜋1 ∶ Γ 1 ⊢ Θ 1 , 𝐴, 𝐵, Δ 1
exR
Γ 1 ⊢ Θ 1 , 𝐵, 𝐴, Δ 1 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
cut
Γ 1 , Γ 2 , Θ 2 ⊢ Θ 1 , 𝐵, Δ 1 , Δ 2
𝜋1 ∶ Γ 1 ⊢ Θ 1 , 𝐴, 𝐵, Δ 1 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
⇝ cut
Γ 1 , Γ 2 , Θ 2 ⊢ Θ 1 , 𝐵, Δ 1 , Δ 2
³Of course, by proposition 17, we already know that MLL is not degenerate, since any sequent that does not respect linearity
is not provable. However, cut elimination is by far more general and it carries the actual meaning of the proof system, contrary
to the linearity property, which is nothing more than a secondary observation.

8
 • If the active formula is introduced by an axiom rule in one of the premisses of the cut, then
this axiom rule and the cut itself may be dropped, with some exchanges possibly introduced to
preserve the order of conclusions. An instance of this is
ax 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
𝐴⊢𝐴 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2 ex
cut ⇝ 𝐴, Γ 2 , Θ 2 ⊢ Δ 2
𝐴, Γ 2 , Θ 2 ⊢ Δ 2

• If the active formula is introduced on both sides of the cut by the introduction rule for some
connective, then we have a right introduction rule on the le and an le introduction rule on the
right for the same connective. In this case, we can eliminate these introduction rules and the cut
rule and introduce instead a cut rule for each immediate subformula.
For negation we have:
𝜋1 ∶ Γ 1 , 𝐵 ⊢ Δ 1 𝜋2 ∶ Γ 2 ⊢ 𝐵, Δ 2 𝜋2 ∶ Γ 2 ⊢ 𝐵, Δ 2 𝜋1 ∶ Γ 1 , 𝐵 ⊢ Δ 1
⊥R ⊥L cut
Γ1 ⊢ 𝐵⊥ , Δ 1 Γ2 , 𝐵⊥ ⊢ Δ2 ⇝ Γ2, Γ1 ⊢ Δ2, Δ1
cut ex
Γ1, Γ2 ⊢ Δ1, Δ2 Γ1, Γ2 ⊢ Δ1, Δ2
For the tensor we have:

𝜋1𝑏 ∶ Γ 1𝑏 ⊢ 𝐵, Δ 1𝑏 𝜋1𝑐 ∶ Γ 1𝑐 ⊢ 𝐶, Δ 1𝑐 𝜋2 ∶ Γ 2 , 𝐵, 𝐶 ⊢ Δ 2
⊗R ⊗L
Γ 1𝑏 , Γ 1𝑐 ⊢ 𝐵 ⊗ 𝐶, Δ 1𝑏 , Δ 1𝑐 Γ 2 , 𝐵 ⊗ 𝐶 ⊢, Δ 2
cut
Γ 1𝑏 , Γ 1𝑐 , Γ 2 ⊢ Δ 1𝑏 , Δ 1𝑐 , Δ 2
⇝
𝜋1𝑐 ∶ Γ 1𝑐 ⊢ 𝐶, Δ 1𝑐 𝜋2 ∶ Γ 2 , 𝐵, 𝐶 ⊢ Δ 2
cut
𝜋1𝑏 ∶ Γ 1𝑏 ⊢ 𝐵, Δ 1𝑏 Γ 1𝑐 , Γ 2 , 𝐵 ⊢ Δ 1𝑐 , Δ 2
cut
Γ 1𝑏 , Γ 1𝑐 , Γ 2 ⊢ Δ 1𝑏 , Δ 1𝑐 , Δ 2
For the par we have:

𝜋1 ∶ Γ 1 ⊢ 𝐵, 𝐶, Δ 1 & 𝜋2𝑏 ∶ Γ 2𝑏 , 𝐵 ⊢ Δ 2𝑏 𝜋2𝑐 ∶ Γ 2𝑐 , 𝐶 ⊢ Δ 2𝑐 &

& R & L
Γ1 ⊢ 𝐵 𝐶, Δ 1 Γ 2𝑏 , Γ 2𝑐 , 𝐵 𝐶 ⊢ Δ 2𝑏 , Δ 2𝑐
cut
Γ 1 , Γ 2𝑏 , Γ 2𝑐 ⊢ Δ 1 , Δ 2𝑏 , Δ 2𝑐
⇝
𝜋1 ∶ Γ 1 ⊢ 𝐵, 𝐶, Δ 1 𝜋2𝑏 ∶ Γ 2𝑏 , 𝐵 ⊢ Δ 2𝑏
cut
Γ 1 , Γ 2𝑏 ⊢ 𝐶, Δ 1 , Δ 2𝑏 𝜋2𝑐 ∶ Γ 2𝑐 , 𝐶 ⊢ Δ 2𝑐
cut
Γ 1 , Γ 2𝑏 , Γ 2𝑐 ⊢ Δ 1 , Δ 2𝑏 , Δ 2𝑐
ese three cases are called interaction rules.
It is easy to check that for every instance of the cut rule in some proof, at least one of the reduction
rules applies. erefore, a proof that is irreducible for this relation ⇝ contains no instance of the cut
rule. Besides, by construction, if 𝜋 ⇝ 𝜋′ then 𝜋 and 𝜋′ have the same conclusion. e proof of the cut
elimination property thus consists in proving that any proof can be reduced into an irreducible one.
24 Lemma. Every sequence of cut elimination steps that never commutes a cut rule with another cut rule is
finite.
Proof. We use a termination measure to prove this fact. To every instance of the cut rule

9
 𝜋1 ∶ Γ 1 ⊢ Θ 1 , 𝐴, Δ 1 𝜋2 ∶ Γ 2 , 𝐴, Θ 2 ⊢ Δ 2
cut
Γ1, Γ2, Θ2 ⊢ Θ1, Δ1, Δ2
we associate the pair (ℎ(𝐴), ℎ(𝜋1 ) + ℎ(𝜋2 )) where ℎ is the height of a formula or proof as a tree. is pair
is called the measure of this instance of the cut rule. Measures are compared by lexicographic ordering.
It is easy to check that, in each rule in the definition or ⇝, the cuts introduced in the right-hand side
have a strictly smaller measure than the eliminated cut on the le-hand side. Indeed, in the case of
interaction rules, the new cuts apply to subformulas of the original formula, so ℎ(𝐴) decreases strictly.
In all other cases, the new cuts are applied to subproofs of the premisses of the original cut, so ℎ(𝐴) is
unchanged and ℎ(𝜋1 ) + ℎ(𝜋2 ) decreases strictly.
Cut measures take values in ℕ × ℕ with the lexicographic ordering, which is well-founded. If the
measure of a proof 𝜋 is defined as the multiset 𝑚(𝜋) of the measures of the instances of the cut rule in
𝜋, then we deduce that 𝜋 ⇝ 𝜋′ implies 𝑚(𝜋) > 𝑚(𝜋′ ). e multiset ordering over a well-founded set
is itself well-founded, hence there can be no infinite sequence of cut elimination steps starting from an
MLL proof.

Note that it is crucial in this lemma that the rule for exchanging two cuts is never used, otherwise
one could trivially form an infinite sequence by repeatedly commuting the same pair of cut rules.
25 eorem. A sequent Γ ⊢ Δ is provable in MLL if and only if it has a proof that does not use the cut rule.
Proof. Assume Γ ⊢ Δ is provable in MLL. Consider a proof 𝜋0 of Γ ⊢ Δ and a maximal sequence
𝜋0 ⇝ 𝜋1 ⇝ ⋯ ⇝ 𝜋𝑛 ⇝ ⋯ that never commutes two cut rules. By definition of ⇝, each proof 𝜋𝑖
has conclusion Γ ⊢ Δ. By lemma 24, this sequence must be finite. Its last element is thus a proof of
Γ ⊢ Δ that is irreducible by the relation ⇝. is implies that this last proof contains no occurrence of
the cut rule.
Another way of formulating this result is saying that the cut rule is admissible in the proof system
MLL without cut. is means that the cut rule does not affect the expressiveness of the system in terms
of provability. is fact is crucial for establishing consistency because it immediately implies that not
all sequents are provable, so the proof system is not trivial.
26 Corollary. e empty sequent ⊢ is not provable in MLL.

Proof. Every rule of MLL except the cut rule has at least one formula in its conclusion, on one side of
the sequent or the other. erefore no cut-free proof can have the empty sequent as conclusion. By
theorem 25, this implies that ⊢ is not provable.
To sum up, the idea of this proof is to define a computational procedure, the cut-elimination relation
⇝, that transforms an arbitrary proof into a cut-free proof with the same conclusion. e fact that
this relation is well-founded proves that the procedure always terminates, hence all proofs have a cut-
free form. is notion of proof rewriting is the core of the proof-program correspondence and we will
elaborate later on the computational part of this correspondence for MLL in particular.

2.3 One-sided presentation

&
In the rules of table 2, it appears that the connectives ⊗ and are very symmetric: they have the same
rules, only with the le and right sides of sequents permuted. e cut elimination rules of definition 23
are consistent with this and, indeed, the cut elimination procedure behaves in a very similar way for
&
both connectives. Everything that happens on the le for ⊗ happens on the right for , and conversely.
is is a strengthened form of the De Morgan duality, which is well-known in classical logic.

10
 Identity

ax ⊢ Γ, 𝐴 ⊢ 𝐴⊥ , Δ
⊢ 𝐴⊥ , 𝐴 cut
⊢ Γ, Δ
Multiplicatives
⊢ Γ, 𝐴 ⊢ 𝐵, Δ ⊢ Γ, 𝐴, 𝐵 &
⊗ &
⊢ Γ, 𝐴 ⊗ 𝐵, Δ ⊢ Γ, 𝐴 𝐵

Table 3: MLL sequent calculus in one-sided presentation

27 Exercise (De Morgan laws). Prove that, for all formulas 𝐴 and 𝐵, the following linear equivalences
hold:
& &
𝐴 ˛ 𝐴⊥⊥ , (𝐴 ⊗ 𝐵)⊥ ˛ 𝐴⊥ 𝐵⊥ , (𝐴 𝐵)⊥ ˛ 𝐴⊥ ⊗ 𝐵⊥ .
is property is the exact analogue of the De Morgan laws of classical logic, that state the duality
between conjunction and disjunction: ¬(𝐴 ∧ 𝐵) is equivalent to ¬𝐴 ∨ ¬𝐵 and conversely. If we apply
De Morgan laws recursively in subformulas, we can easily establish that any formula of MLL is linearly
equivalent to a formula where negation is only applied to propositional variables. Moreover, the intro-
duction rules for negation imply that a sequent Γ, 𝐴 ⊢ Δ is provable in MLL if and only if the sequent
Γ ⊢ 𝐴⊥ , Δ is provable.
ese considerations motivate the introduction of the one-sided presentation of MLL. e idea is to
combine the above remarks so that
• formulas are considered up to De Morgan equivalences,
• all sequents have the form ⊢ Γ, with an empty le-hand side,
• subsequently, only right introduction rules are used.
Moreover, sequents will be considered up to permutation, since exchange rules make permutation of
formula instances innocuous.
28 Definition. e language of formulas of one-sided MLL is defined by the following grammar:

𝐴, 𝐵 ∶= 𝛼 propositional variable
𝛼⊥ negated propositional variable
𝐴⊗𝐵 multiplicative conjunction
&
𝐴 𝐵 multiplicative disjunction

Linear negation is the operation (⋅)⊥ on formulas inductively defined as

& &
(𝛼)⊥ ∶= 𝛼⊥ , (𝛼⊥ )⊥ ∶= 𝛼, (𝐴 ⊗ 𝐵)⊥ ∶= 𝐴⊥ 𝐵⊥ , (𝐴 𝐵)⊥ ∶= 𝐴⊥ ⊗ 𝐵⊥ .
A one-sided MLL sequent is a possibly empty multiset Γ of one-sided MLL formulas, usually wrien as
⊢ Γ. A sequential one-sided MLL proof is a proof tree built using the rules of table 3.
Note that the axiom and cut rules have to be reformulated to fit in this context. As expected, the sys-
tem is greatly simplified since we now have only four rules, instead of ten in the one-sided presentation.
Expressiveness is preserved, as the following theorem states.
29 eorem. A sequent 𝐴1 , …, 𝐴𝑛 ⊢ 𝐵1 , …, 𝐵𝑝 is provable in two-sided MLL if and only if the sequent ⊢
𝐴⊥
1 , …, 𝐴𝑛 , 𝐵1 , …, 𝐵𝑝 is provable in one-sided MLL.
⊥

11
Proof. Each proof 𝜋 of two-sided MLL with some conclusion Γ ⊢ Δ can be transformed into a proof 𝜋′
of ⊢ Γ ⊥ , Δ, and this transformation is defined inductively, case by case on the last rule:
• if 𝜋 is a two-sided axiom, then the one-sided axiom rule is a suitable 𝜋′ ,

• if 𝜋 is a two-sided cut between 𝜋1 and 𝜋2 , then the one-sided cut between 𝜋′1 and 𝜋′2 is a suit-
able 𝜋′ ,

• right introduction rules are translated into introduction rules for the same connectives,
• le introduction rules are translated into introduction rules for the dual connectives,

• introduction rules for negation and exchange rules have no effect on the conclusion in the one-
sided form so they disappear in 𝜋′ .

For the reverse implication, each rule of one-sided MLL is mapped to the rule of two-sided MLL
with the same shape. Introductions are translated back into le or right introduction rules depending
on whether the introduced formula is an 𝐴𝑖 or a 𝐵𝑗 . Exchange rules and negation introductions are
inserted as needed for the rules to apply.

Consistency of one-sided MLL is a consequence of this theorem. However, it could be established

directly, using the same technique based on proof rewriting, and the cut-elimination procedure is ac-
tually simpler because there are much less commutation rules and only one interaction rule, for the
&
pair ⊗/ . Besides, the cut elimination relation ⇝ of definition 23 is mapped by the proof translation of
theorem 29 to the appropriate cut elimination relation for one-sided MLL.

2.4 Full linear logic

So far, we have focused on the very simple MLL system, which contains only multiplicative forms
of conjunction and disjunction. e one-sided presentation, which includes “hard-wired” De Morgan
duality, provides a simple sequent-style proof system with good properties, so we will keep using one-
sided presentations whenever possible.
e full system of linear logic is obtained by introducing in the system all the other connectives
needed to get back the expressiveness of full classical logic. We will now briefly describe all these,
together with their cut elimination procedure.

Additives
e additive form of conjunction is wrien & and pronounced “with”, the additive form of disjunction
is wrien ⊕ and pronounced “plus”. ey are dual, in the sens that the definition of negation is extended
with the rules
(𝐴 & 𝐵)⊥ ∶= 𝐴⊥ ⊕ 𝐵⊥ and (𝐴 ⊕ 𝐵)⊥ ∶= 𝐴⊥ & 𝐵⊥ .

eir proof rules are

⊢ Γ, 𝐴 ⊢ Γ, 𝐵 ⊢ Γ, 𝐴 ⊢ Γ, 𝐵
& ⊕1 ⊕2
⊢ Γ, 𝐴 & 𝐵 ⊢ Γ, 𝐴 ⊕ 𝐵 ⊢ Γ, 𝐴 ⊕ 𝐵
e names ⊕1 and ⊕2 refer to the fact that the rules are very similar, to the point that they may be
formulated as a single rule:
⊢ Γ, 𝐴𝑖
⊕𝑖
⊢ Γ, 𝐴1 ⊕ 𝐴2

12
 where 𝑖 is either 1 or 2. eir treatment in cut elimination is similar to the one for multiplicative
connectives. ere are now two interaction rules, for & against ⊕1 and for & against ⊕2 , and they have
a very similar shape:

𝜋1,1 ⊢ Γ, 𝐴1 𝜋1,2 ⊢ Γ, 𝐴2 𝜋2 ⊢ Δ, 𝐴⊥
𝑖
& ⊕𝑖 𝜋1,𝑖 ⊢ Γ, 𝐴𝑖 𝜋2 ⊢ Δ, 𝐴⊥
𝑖
⊢ Γ, 𝐴1 & 𝐴2 ⊢ Δ, 𝐴⊥ ⊥
1 ⊕ 𝐴2
⇝ cut
cut ⊢ Γ, Δ
⊢ Γ, Δ
Remark that one of the premisses of the & rule is erased. As for the commutation rules, the case of
a cut against a formula in the context of one of the additive rules is treated in the only possible way.
Commutation with ⊕𝑖 is straightforward and commutation with & implies duplicating the other proof:

𝜋1,1 ⊢ Γ, 𝐴, 𝐵1 𝜋1,2 ⊢ Γ, 𝐴, 𝐵2
&
⊢ Γ, 𝐴, 𝐵1 & 𝐵2 𝜋2 ⊢ Δ, 𝐴⊥
cut
⊢ Γ, Δ, 𝐵1 & 𝐵2
⇝
𝜋1,1 ⊢ Γ, 𝐴, 𝐵1 𝜋2 ⊢ Δ, 𝐴⊥ 𝜋1,2 ⊢ Γ, 𝐴, 𝐵2 𝜋2 ⊢ Δ, 𝐴⊥
cut cut
⊢ Γ, Δ, 𝐵1 ⊢ Γ, Δ, 𝐵2
&
⊢ Γ, Δ, 𝐵1 & 𝐵2
e termination measure used in lemma 24 does not work directly, because this rule might duplicate cuts
in 𝜋2 that are bigger than the cut it reduces. However, if we restrict this rule so that it only applies when
cuts in 𝜋2 are small enough (at worst, apply it only if 𝜋2 is already cut-free!), then we get a reduction
strategy that reaches a normal form, so consistency is preserved. e logical system we obtain is called
multiplicative-additive linear logic and abbreviated as MALL.
30 We have the following notations for the connectives:
conjunction disjunction
&
multiplicative ⊗
additive & ⊕
is may look like a strange decision to use these symbols since connectives of the same kind do not
look alike. One of the justifications of this choice is that connectives that are similar in appearance
interact nicely together.
31 Exercise. Prove the following linear equivalences in MALL:
& & &
𝐴 ⊗ (𝐵 ⊕ 𝐶) ˛ (𝐴 ⊗ 𝐵) ⊕ (𝐴 ⊗ 𝐶) and 𝐴 (𝐵 & 𝐶) ˛ (𝐴 𝐵) & (𝐴 𝐶).
ese distributivity laws are similar to the standard calculus laws of distribution of multiplication
over addition, the similar shape of the connectives help in remembering which distributivities hold.
& & &
Indeed, 𝐴 (𝐵 ⊕ 𝐶) is not linearly equivalent to (𝐴 𝐵) ⊕ (𝐴 𝐶).
32 Exercise. Prove that.
In the computational interpretation of proofs deduced from the cut elimination procedure, we inter-
pret a proof of 𝐴 ⊢ 𝐵 (or equivalently a proof of ⊢ 𝐴⊥ , 𝐵) as a way to map a proof of 𝐴 to a proof of 𝐵.
e cut elimination procedure defines how a given proof of 𝐴 is used to produce the resulting proof of
𝐵. Comparing the cut elimination behaviour of multiplicative and additive connectives illustrates the
difference in meaning between them. Assuming 𝛿 (the data) is a proof of ⊢ 𝐴 ∗ 𝐵 for some connective ∗
and 𝜑 (the function) is a proof of 𝐴 ∗ 𝐵 ⊢ 𝐶, what will happen when performing cut elimination in the
cut of 𝛿 against 𝜑?

13
 • for 𝐴 ⊕ 𝐵, 𝛿 provides either a proof of 𝐴 or a proof of 𝐵;

• for 𝐴 & 𝐵, 𝛿 provides a proof of 𝐴 and a proof of 𝐵 but 𝜑 must use one of them;
• for 𝐴 ⊗ 𝐵, 𝛿 provides a proof of 𝐴 and a proof of 𝐵 and 𝜑 must use both of them;
&
• for 𝐴 𝐵, 𝛿 provides one proof with two conclusions 𝐴 and 𝐵, and 𝜑 must use them independently,
with some subproof for 𝐴 and some other subproof for 𝐵.

Units
In classical logic, there are two logical units, standing for true and false. A way of characterizing them is
as neutral elements for conjunction and disjunction, respectively. For this reason, in linear logic, there
are multiplicative and additive variants of these units, hence there are four units. eir notations follow
the paern of the conjunctions and disjunctions:
“true” “false”
multiplicative 1 ⊥
additive ⊤ 0
e definition of negation extends in the natural way:

1⊥ ∶= ⊥, ⊥⊥ ∶= 1, ⊤⊥ ∶= 0, 0⊥ ∶= ⊤.
e proof rules are deduced from the neutrality property:

1 ⊢Γ ⊤
⊢1 ⊥ ⊢ Γ, ⊤
⊢ Γ, ⊥
33 ere is no rule for 0. is can be justified by the computational interpretation and the neutrality
for ⊕: a proof of 𝐴 ⊕ 𝐵 is either a proof of 𝐴 or a proof of 𝐵, so the set of proofs of 𝐴 ⊕ 𝐵 is mostly the
disjoint union of the set of proofs of 𝐴 and the set of proofs of 𝐵. But if 𝐴 ⊕ 0 and 𝐴 must be equivalent,
this means that the set of proofs of 𝐴 ⊕ 0 and the set of proofs of 𝐴 are essentially the same, therefore
there must be no proof of 0.
e cut elimination rules are extended in the straightforward way: interaction of 1 against ⊥ simply
cancels both rules, interaction between ⊤ and 0 never happens since there is no rule for 0. e context
rule for ⊥ is a simple commutation and the context rule for ⊤ erases the cut and its other premiss:
⊤
⊢ Γ, 𝐴, ⊤ 𝜋 ⊢ Δ, 𝐴⊥ ⇝ ⊤
cut ⊢ Γ, Δ, ⊤
⊢ Γ, Δ, ⊤
is is consistent with the context rule for &.
34 Exercise. Prove the following linear equivalences:
&
• neutralities: 𝐴 ˛ 𝐴 ⊗ 1 ˛ 𝐴 ⊥ ˛ 𝐴 & ⊤ ˛ 𝐴 ⊕ 0,
&
• nullity: 𝐴 ⊗ 0 ˛ 0, 𝐴 ⊤ ˛ ⊤.

14
Exponentials
Limiting the logical system to multiplicative and additive variants of classical conjunction and disjunc-
tion, without any contraction and weakening, does provide an interesting system, however its logical
expressiveness is rather limited. A witness of this fact is that the height of cut-free proofs of a given
sequent is obviously bounded by the number of connectives in the sequent, since each rule introduces
one connective. As a consequence, proof search in MALL can be performed by exhaustive enumera-
tion of all possible proofs and the decision problem of finding whether a given sequent is provable is
decidable.⁴
In order to recover the expressiveness of classical logic, it is thus necessary to allow some contraction
and weakening. Introducing them directly would collapse the system into classical logic itself, so we
introduce them in a controlled way, by means of modalities. e language of formulas is extended with
a pair of new constructs:
𝐴, 𝐵 ∶= ⋯
!𝐴 replicable and erasable — read “of course 𝐴”
?𝐴 possibly contracted or weakened — read “why not 𝐴”

De Morgan duality is extended so that these are dual: (!𝐴)⊥ ∶= ?(𝐴⊥ ) and (?𝐴)⊥ ∶= !(𝐴⊥ ). e modal-
ity ? is a marker for formulas that can be contracted and weakened:
⊢ Γ, 𝐴 ⊢Γ ⊢ Γ, ?𝐴, ?𝐴 ⊢ ?𝐴1 , …?𝐴𝑛 , 𝐵
? w c !
⊢ Γ, ?𝐴 ⊢ Γ, ?𝐴 ⊢ Γ, ?𝐴 ⊢ ?𝐴1 , …?𝐴𝑛 , !𝐵
e introduction rule for ? is usually called “dereliction”, as it is a kind of regression: the linearity
information is lost. e introduction rule for !, usually called “promotion”, is very particular since it
imposes a constraint on the context, indeed if we use the proof with conclusion 𝐵 several times later in
our proof, we will have to use the other conclusions the same number of times, hence these cannot be
linear.
e cut elimination rules naturally imply duplicating or erasing the proof above a promotion rule.
For instance, when a cut occurs between a contraction and a promotion, we have
𝜋2 ⊢ ?Δ, 𝐴⊥
!
𝜋1 ⊢ Γ, ?𝐴, ?𝐴 𝜋2 ⊢ ?Δ, 𝐴⊥ 𝜋1 ⊢ Γ, ?𝐴, ?𝐴 ⊢ ?Δ, !𝐴⊥ 𝜋2 ⊢ ?Δ, 𝐴⊥
c ! cut !
⊢ Γ, ?𝐴 ⊢ ?Δ, !𝐴⊥ ⇝ ⊢ Γ, ?Δ, ?𝐴 ⊢ ?Δ, !𝐴⊥
cut cut
⊢ Γ, ?Δ ⊢ Γ, ?Δ, ?Δ
c
⊢ Γ, ?Δ
where ?Δ stands for a sequence of formulas all starting with the ? modality, and where the right-hand
side ends with as many instances of the contraction rule as necessary to get the appropriate conclusion.
Interaction with weakening is similar except that the promotion is erased and not duplicated, interaction
with dereliction is a simple cancellation. Commutation rules can be formulated too, paying aention
to the contextual constraint in the promotion rule. e termination argument of lemma 24 does not
work anymore in this context, because of duplication: the rule above does not decrease this measure.
is can be fixed by taking into account the number of contraction rules applied to the active formula
in each cut, which does decrease.
ese modalities are called exponential because they have the fundamental property of turning ad-
ditives into multiplicatives.
⁴e problem of deciding the provability of an arbitrary MALL sequent was actually proved to be complete for polynomial
space complexity, whereas the same problem for full propositional linear logic is undecidable. See Lincoln et al., “Decision prob-
lems for propositional linear logic”.

15
35 Exercise. Prove the linear equivalences
&
!(𝐴 & 𝐵) ˛ !𝐴 ⊗ !𝐵, ?(𝐴 ⊕ 𝐵) ˛ ?𝐴 ?𝐵, !⊤ ˛ 1, ?0 ˛ ⊥.

antifiers
Our exposition so far has been restricted to propositional logic, since it is already very expressive in
linear logic. However, it is possible to include quantifiers and get the full mathematical expressiveness
of traditional predicate calculus, in a refined way.
First-order quantifiers are introduced in the standard way: choose a language for first-order objects
(variables, function symbols, predicate symbols), replace propositional variables by atomic formulas
made of a predicate symbol applied to the right number of first-order terms, and extend the language of
formulas with the usual quantifiers ∀𝑥 and ∃𝑥, dual of each other: (∀𝑥 𝐴)⊥ = ∃𝑥(𝐴⊥ ). Proof rules and
cut elimination steps are the same as in classical logic, cut elimination keeps the same complexity, and
on the whole, the proof theory is not very much affected by this change.
Second-order quantifiers are more interesting from the point of view of the structure of proofs. Now
we extend the language of formulas with quantification over propositional variables ∀𝛼 and ∃𝛼. ese
quantifiers are also dual of each other, the proof rules and cut elimination rules are as simple as for first
order, even more so since the language is simpler. e real difference lies in the argument for termina-
tion of cut elimination: now we cannot proceed by induction on formulas anymore, since substitution
of propositional variables make formulas grow. It is now necessary to use arguments comparable to
reducibility candidates (as used for second-order intuitionistic natural deduction, a.k.a. system F, a.k.a.
polymorphic 𝜆-calculus).
An interesting side-effect of second-order quantification is that they make additives and units un-
necessary, with respect to linear equivalence:

0 ˛ ∀𝛼 𝛼 1 ˛ ∀𝛼(𝛼 ⊸ 𝛼) 𝐴 ⊕ 𝐵 ˛ ∀𝛼(!(𝐴 ⊸ 𝛼) ⊸ !(𝐵 ⊸ 𝛼) ⊸ 𝛼)

& &
⊤ ˛ ∃𝛼 𝛼 ⊥ ˛ ∃𝛼(𝛼 ⊗ 𝛼⊥ ) 𝐴 & 𝐵 ˛ ∃𝛼(!(𝐴 𝛼) ⊗ !(𝐵 𝛼) ⊗ 𝛼⊥ )

e semantic study of quantifiers in linear logic is rich topic with many open questions, and we
will not elaborate on this in the present notes. Besides, the propositional part already illustrates most
aspects of the proof theory of linear logic.

2.5 e notion of fragment

As already mentioned before, it is oen useful to study subsystems of linear logic. ese are oen
simpler because they have less connectives, less proof rules and less cut-elimination rules. ey may
still be expressive enough to represent some other systems or to illustrate particular features. For these
reasons, many works focus on various subsystems, some of which have become standard systems in
their own right.
A fragment of linear logic is a system obtained by imposing restrictions on the language of formu-
las, sequents or proofs. e most standard ones, based on the restriction of the set of formulas, are
propositional fragments, with no quantifiers of any order and no units:
• LL: all connectives allowed
• MLL: only multiplicatives
• MALL: only multiplicatives and additives
• MELL: only multiplicatives and exponential modalities — this is enough to represent and study
the simply typed 𝜆-calculus and has been used extensively for this purpose

16
 In these systems, the rules that involve forbidden connectives are simply never used. e same sys-
tems extended with the units are named similarly with a 0 as subscript. e systems with quantifiers
are named with a 1 or 2 (or both) as subscripts, for first and second order quantification. With these
& &
conventions, MLL02 allows ⊗, , 1, ⊥ and second-order quantification; MALL1 allows ⊗, , ⊕, & and
first-order quantification but no units, and LL012 is the full system.
Fragments can also be defined by restricting the shape of sequents. e most standard example of
that is intuitionistic linear logic, which designates the same family of systems but in two-sided presenta-
&
tion, with always exactly one formula on the right. is effectively forbids the rules for and negation,
so these systems include linear implication ⊸ as a primary connective, as in exercise 21. Such systems
are named like the classical systems, except that “LL” is replaced by “ILL”. Section 5.1 elaborates on
MILL, the minimal intuitionistic system.

3 A bit of semantics
3.1 Provability semantics
e cut elimination property provides an argument for consistency of linear logic since it provide a com-
putational interpretation of proofs and implies that the system is not degenerate. However, it provides
no interpretation of formulas that would allow for the formulation of a completeness theorem meaning
that there are enough rules for the connectives of the logic. In classical logic, this role is played by
Boolean algebras, with the fact that a formula is provable in LK if and only if its interpretation is true
in any such algebra. e proper notion in linear logic is that of phase spaces.
36 Definition. A phase space is a pair (𝑀, ⊥) where 𝑀 is a commutative monoid and ⊥ is a subset of 𝑀.
Two points 𝑥, 𝑦 ∈ 𝑀 are called orthogonal if 𝑥𝑦 ∈ ⊥. For a subset 𝐴 ⊆ 𝑀, let define the orthogonal of 𝐴
as 𝐴⊥ ∶= 􏿺𝑦 ∈ 𝑀 􏿖 ∀𝑥 ∈ 𝐴, 𝑥𝑦 ∈ ⊥􏿽. A fact is any subset of 𝑀 of the form 𝐴⊥ .
Points of 𝑀 are interpreted as tests, or possible interactions between a potential proof of a formula
and a potential refutation of it. Elements of ⊥ are successful tests or interactions in this set. Formulas will
be interpreted by facts, which play the role of truth values and describe the set of possible behaviours
for potential proofs.
is is parametricity in ⊥ the key point in phase spaces; it is where the notion of interaction appears
explicitly in the logic and how it differs from classical systems with traditional Boole-like semantics.
37 Exercise. Prove that, for any subsets 𝐴 and 𝐵 of 𝑀, 𝐴 ⊆ 𝐵 implies 𝐵⊥ ⊆ 𝐴⊥ . Prove that for any 𝐴,
𝐴 ⊆ 𝐴⊥⊥ and 𝐴⊥⊥⊥ = 𝐴⊥ .
38 Definition. Let (𝑀, ⊥) be a phase space and let 𝑒 be the neutral element of 𝑀. For any subsets 𝐴, 𝐵 ⊆ 𝑀,
define
⊥⊥ &
𝐴 ⊗ 𝐵 ∶= 􏿺𝑝𝑞 􏿖 𝑝 ∈ 𝐴, 𝑞 ∈ 𝐵􏿽 𝐴 𝐵 ∶= (𝐴⊥ ⊗ 𝐵⊥ )⊥
𝐴 ⊕ 𝐵 ∶= (𝐴 ∪ 𝐵) ⊥⊥ 𝐴 & 𝐵 ∶= 𝐴 ∩ 𝐵 0 ∶= ∅⊥⊥ ⊤ ∶= 𝑀
!𝐴 ∶= (𝐴 ∩ 𝐼) ⊥⊥ ⊥
?𝐴 ∶= (𝐴 ∩ 𝐼) ⊥ 1 ∶= {𝑒}⊥⊥
where 𝐼 is the set of idempotent elements of 𝑀 that are elements of the set 1.
An interpretation J⋅K𝑀 is defined by the choice of a fact J𝛼K𝑀 for each propositional variable 𝛼. If
𝐴 is a formula of LL, its interpretation J𝐴K𝑀 is deduced using the definitions above.
It is easy to check that if propositional variables are interpreted as facts, then for any formula 𝐴 the
interpretation J𝐴K𝑀 is a fact. Moreover, linear negation is indeed interpreted by the orthogonal. e
interpretation of linear implication is instructive:
&
𝐴 ⊸ 𝐵 = 𝐴⊥ 𝐵 = 􏿺𝑥 ∈ 𝑀 􏿖 ∀𝑦 ∈ 𝐴, 𝑥𝑦 ∈ 𝐵􏿽

17
 A point 𝑥 is in 𝐴 ⊸ 𝐵 if, whenever it is composed with a point of 𝐴, the result is in 𝐵.
39 eorem. A formula 𝐴 of LL is provable if and only if 𝑒 ∈ J𝐴K𝑀 in all interpretations in phase spaces.
Proof. Soundness is easily checked by induction on proofs. For completeness, the idea is to take for
𝑀 the set of all sequents, considered up to permutation and up to duplication of ? formulas, with con-
catenation of sequents as the monoid operation. For ⊥ we take the set the provable sequents. en one
checks by induction that for any formula 𝐴 the set J𝐴K𝑀 is the set of all Γ such that ⊢ Γ, 𝐴 is provable, if
we take this as a definition for propositional variables. e neutral element 𝑒 of 𝑀 is the empty sequent,
so ⊢ 𝐴 is provable when 𝑒 ∈ J𝐴K𝑀 .
Remark that, as a degenerate case, we can always choose ⊥ = ∅, then the set of facts is the ele-
mentary Boolean algebra {∅, ⊤}. In this case, formulas are interpreted by their naive truth value, both
conjunctions are interpreted as the classical one, similarly for disjunctions, all linearity is lost and we
get back classical logic.

3.2 Proof semantics in coherence spaces

e idea of linearity historically appeared in the study of coherence spaces as a denotational model of
the 𝜆-calculus. Hence it is natural that proofs in linear logic have an interpretation in these spaces. We
first define the interpretation of formulas as coherence spaces.
40 Definition. Let 𝐴 and 𝐵 be coherence spaces. e coherence spaces 𝐴⊥ , 𝐴⊗𝐵, 𝐴⊕𝐵 and !𝐴 are defined
as follows:
• 􏿖𝐴⊥ 􏿖 = |𝐴| and 𝑥 ¨𝐴⊥ 𝑥′ if either 𝑥 = 𝑥′ or 𝑥 and 𝑥′ are not coherent in 𝐴.
• |𝐴 ⊗ 𝐵| = |𝐴| × |𝐵| and (𝑥, 𝑦) ¨𝐴⊗𝐵 (𝑥′ , 𝑦′ ) if 𝑥 ¨𝐴 𝑥′ and 𝑦 ¨𝐵 𝑦′ .
• |𝐴 ⊕ 𝐵| = ({1} × |𝐴|) ∪ ({2} × |𝐵|) and (𝑖, 𝑥) ¨𝐴⊕𝐵 (𝑗, 𝑥′ ) if 𝑖 = 𝑗 and 𝑥 ¨ 𝑥′ .
• |!𝐴| is the set of finite cliques of 𝐴 and 𝑥 ¨!𝐴 𝑥′ if 𝑥 ∪ 𝑥′ is a clique in 𝐴.
&
e spaces 𝐴 𝐵, 𝐴 & 𝐵 and ?𝐴 are defined by duality.
&
e definition by duality can be easily reformulated directly for 𝐴 𝐵 and 𝐴 & 𝐵. In particular, if
we expand the definition for 𝐴 ⊸ 𝐵, we get that (𝑥, 𝑦) ¨𝐴⊸𝐵 (𝑥′ , 𝑦) if, whenever 𝑥 ¨𝐴 𝑥′ , we have
𝑦 ¨𝐵 𝑦′ : cliques in 𝐴 ⊸ 𝐵 are relations that map coherent points to coherent points. If the coherence
spaces 𝐴 and 𝐵 are flat (i.e. coherence is equality) then cliques in 𝐴 ⊸ 𝐵 are graphs of partial functions
from 𝐴 to 𝐵.
& &
A proof of a sequent ⊢ 𝐴1 , …, 𝐴𝑛 will be interpreted by a clique in the coherence space 𝐴1 ⋯ 𝐴𝑛 .
We will derive judgments of the form ⊢ 𝛼1 ∶ 𝐴1 , …, 𝛼𝑛 ∶ 𝐴𝑛 where each 𝛼𝑖 is a point in 𝐴𝑖 , meaning
that the tuple (𝛼1 , …, 𝛼𝑛 ) is an element of the interpretation of the considered proof. e coherent
interpretation, in table 4 then appears as a decoration of the proof system.
Remark the particular shape of the rules for &: it appears as identical to the rules for ⊕, and indeed
the difference comes form the fact that a proof 𝐴 & 𝐵 does contain a proof of 𝐴 and a proof of 𝐵, and the
interpretation makes a disjoint union of the interpretations of these proofs. Another point to remark
is that, in the interpretation of the cut rule, it is required that both premisses use the same point 𝛼; the
other rules have no such constraint.
41 eorem. e set of tuples in the interpretation of a proof is always a clique.
Proof. By a simple induction of proofs.
42 eorem. e interpretation of proofs in coherence spaces is invariant by cut elimination.
Proof. By case analysis on the various cases of cut elimination.

18
 Identity

ax ⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴 ⊢ 𝛼 ∶ 𝐴⊥ , 𝛿 ∶ Δ
⊢𝛼∶ 𝐴⊥ , 𝛼 ∶𝐴 cut
⊢ 𝛾 ∶ Γ, 𝛿 ∶ Δ
Multiplicatives
⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴 ⊢ 𝛽 ∶ 𝐵, 𝛿 ∶ Δ ⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴, 𝛽 ∶ 𝐵 &
⊗ &
⊢ 𝛾 ∶ Γ, (𝛼, 𝛽) ∶ 𝐴 ⊗ 𝐵, 𝛿 ∶ Δ ⊢ 𝛾 ∶ Γ, (𝛼, 𝛽) ∶ 𝐴 𝐵
Additives
⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴𝑖 ⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴𝑖
&𝑖 ⊕𝑖
⊢ 𝛾 ∶ Γ, (𝑖, 𝛼) ∶ 𝐴1 & 𝐴2 ⊢ 𝛾 ∶ Γ, (𝑖, 𝛼) ∶ 𝐴1 ⊕ 𝐴2
Exponentials
⊢ 𝛾 ∶ Γ, 𝛼 ∶ 𝐴 ⊢𝛾∶Γ ⊢ 𝛾 ∶ Γ, 𝑎 ∶ ?𝐴, 𝑎′ ?𝐴
? w c
⊢ 𝛾 ∶ Γ, {𝛼} ∶ ?𝐴 ⊢ 𝛾 ∶ Γ, ∅ ∶ ?𝐴 ⊢ 𝛾 ∶ Γ, 𝑎 ∪ 𝑎′ ∶ ?𝐴

􏿺 ⊢ 𝑎1,𝑖 ∶ ?𝐴1 , …𝑎𝑛,𝑖 ∶ ?𝐴𝑛 , 𝑏𝑖 ∶ 𝐵 􏿽

𝑖∈𝐼
!
⊢ ⋃𝑖∈𝐼 𝑎1,𝑖 ∶ ?𝐴1 , … ⋃𝑖∈𝐼 𝑎𝑛∶𝑖 ∶ ?𝐴𝑛 , {𝑏𝑖 | 𝑖 ∈ 𝐼} ∶ !𝐵

Table 4: Interpretation of proofs in coherence spaces

4 A bit of proof theory

4.1 Intuitionistic and classical logics as fragments
In section 1.2, we introduced coherence spaces as a denotational semantics for the 𝜆-calculus. e study
of this model revealed the notion of linearity, from which linear logic was built, and proofs in linear logic
also got an interpretation in coherence spaces in section 3.2. It is natural to ask how these models relate
from a logical point of view. e answer to this question lies in the definition of the trace of a stable
function, as of definition 10. It can be formalized as the definition of a translation from intuitionistic
natural deduction into linear logic.
43 Definition. e translation (⋅)∗ of formulas from implicative-conjunctive propositional logic into linear
logic is defined inductively as
(𝛼)∗ ∶= 𝛼 (𝐴 ⇒ 𝐵)∗ ∶= !𝐴∗ ⊸ 𝐵∗ (𝐴 ∧ 𝐵)∗ ∶= 𝐴∗ & 𝐵∗

Sequents are translated as (𝐴1 , …, 𝐴𝑛 ⊢ 𝐵)∗ ∶= ⊢ ?(𝐴∗1 )⊥ , …, ?(𝐴∗𝑛 )⊥ , 𝐵∗ .

If we interpret this at the level of coherence spaces, we see that a point in (𝐴 ⇒ 𝐵)∗ is a point in
&
?(𝐴∗ )⊥ 𝐵∗ , that is a pair of a clique in (𝐴∗ )⊥ and a point in 𝐵∗ . Developing further, we see that a
clique in (𝐴 ⇒ 𝐵)∗ is a coherent set of such pairs, which is the condition for being the trace of a stable
function.
is translation of formulas is naturally extended as a translation of proofs, which defines a transla-
tion of 𝜆-terms into sequential proofs in linear logic. e introduction rule for ⇒ exactly corresponds
&
to the introduction rule for , but the elimination rule requires more work:

19
 ⊢ ?(Δ∗ )⊥ , 𝐴
! ax
Γ⊢𝐴⇒𝐵 Δ⊢𝐴 ⊢ ?(Δ∗ )⊥ , !𝐴 ⊢ 𝐵⊥ , 𝐵
→ & ⊗
Γ, Δ ⊢ 𝐵 ⊢ ?(Γ ∗ )⊥ , ?𝐴⊥ 𝐵 ⊢ ?(Δ∗ )⊥ , !𝐴 ⊗ 𝐵⊥ , 𝐵
cut
⊢ ?(Γ ∗ )⊥ , ?(Δ∗ )⊥ , 𝐵
44 Exercise. Write the translation of all typing rules of the 𝜆-calculus into linear logic for this translation
of formulas.
is translation of formulas has the crucial property that it preserves provability: a sequent Γ ⊢ 𝐴
is provable in intuitionistic natural deduction if and only if its translation is provable in linear logic. For
this reason, when dealing with sequents that are translations of intuitionistic ones, it is not a restriction,
with respect to provability, to only accept proofs that are translations of intuitionistic proofs. Hence
intuitionistic can be considered as a fragment of linear logic e simply typed 𝜆-calculus can be seen
as a fragment of MELL, since the purely implicative part does not use additives.
If we study cut elimination in this fragment, we subsequently get a decomposition of the 𝛽-reduction
of 𝜆-calculus into simpler steps. We will not develop this point in the present notes, however it does
provide insights on the operational semantics of the 𝜆-calculus and in particular makes the status of
linearity explicit.
is translation of intuitionistic logic into linear logic suggests the possibility of also translating
classical logic. It is well known that double-negation translations and their refinements map classical
logic into intuitionistic logic; translation into linear logic simplifies the translation a bit further.
e crucial simplification is that, as for the translation (𝐴 ⇒ 𝐵)∗ = !𝐴∗ ⊸ 𝐵∗ , everything happens
through the use of modalities, carefully introduced in the translation in order to allow contraction and
weakening when required by the original logic. In a classical sequent Γ ⊢ Δ, these rules are allowed
both on the le (hence the ! on the le of linear implication) and on the right, which requires a ? on
the right of linear implication. Translating 𝐴 ⇒ 𝐵 into !𝐴 ⊸ ?𝐵 does not work, and systematic study
reveals two possible translations:

(𝐴 ⇒ 𝐵)𝑡 ∶= !?𝐴𝑡 ⊸ ?𝐵𝑡 or (𝐴 ⇒ 𝐵)𝑞 ∶= !𝐴𝑞 ⊸ ?!𝐵𝑞 ,

and variations around these (for instance, replacing ! with !! would not change the result very much).
e translation extends to conjunction and disjunction by choosing the appropriate variant (additive or
multiplicative) and requires some care; we will dot develop it here.
45 eorem. A sequent is provable in classical sequent calculus if and only if its translation in linear logic,
by any of the above translations, is provable.
Proof. Preservation of provability through the translation is proved by providing a translation of classi-
cal proof rules, in a similar way as for the intuitionistic case above. For the reverse implication, it suffices
to remark that there is a simple reverse translation that erases modalities, translates all conjunctions to
∧ and all disjunctions to ∨ and that this translation also preserves provability.
is theorem implies that classical logic too can be considered as a fragment of linear logic. Actually,
since there are two different translations here, two fragments of LL correspond to classical logic (for
these translations, these systems are known as LKT and LKQ respectively). e study of these fragment
applied to classical extensions of the 𝜆-calculus reveal that the dichotomy between the two forms of
translations is a proof-theoretical form of the computational dichotomy between call-by-name (for the
𝑡 variant) and call-by-value (for the 𝑞 variant). See Danos, Joinet, and Schellinx, “A new deconstructive
logic: linear logic” for a thourough study of such translations.

20
 4.2 Cut elimination and proof equivalence
e proof rewriting system for cut elimination is an operation of normalization for proofs. However,
confluence of this system fails, but for bad reasons: depending on the order in which we apply cut
elimination steps, we might end up with proofs that differ in the order of rules (moreover, the various
uses of the exchange rules is unspecified, but this is a secondary problem). So we only get confluence
up to commutation of independent rules.
On the other hand, any denotational semantics that is an invariant of cut elimination must interpret
these equivalent proofs as equal. e point of proof nets, as introduced in section 5, is to avoid this prob-
lem by providing a formalism without rule commutation. We can also study what these commutations
mean for sequential proofs.
We could indeed prove that the various rewriting rules for cut elimination form a confluent system
modulo this quotient, so a given proof always has one cut-free form up to commutation, and it is obtained
by the cut-elimination procedure (the details for proving this are not of particular interest in the context
of these notes and we will stay at a more informal level). Lemma 24 then implies strong normalization,
which means that any sequence of reductions eventually yields the same result. Hence the following
definition is consistent.
46 Definition. For two proofs 𝜋 ∶ Γ ⊢ 𝐴, Δ and 𝜋′ ∶ Γ ′ , 𝐴 ⊢ Δ, let [𝜋, 𝜋′ ] be the cut-free form up to rule
commutation of the proof obtained by applying the cut rule on 𝜋 and 𝜋′ .
is defines a notion of composition for cut-free proofs. We can elaborate on this notion by defining
a proper category of cut-free proofs: the objects are the formulas of MLL and the morphisms from 𝐴
to 𝐵 are the cut-free proofs of 𝐴 ⊢ 𝐵. e composition operation of the above definition provides
composition in the category, the axiom rule provides the identity morphisms. It is an easy but tedious
lemma to prove the associativity of this composition.
Many models identify proofs a bit more than what cut elimination imposes. Consider these three
proofs of the identity for 𝐴 ⊕ (𝐵 ⊗ 𝐶):
1. A simple axiom rule:

& ax
⊢ 𝐴⊥ & (𝐵⊥ 𝐶⊥ ), 𝐴 ⊕ (𝐵 ⊗ 𝐶)

2. An axiom rule for 𝐵 ⊗ 𝐶:

ax & ax
⊢ 𝐴⊥ , 𝐴 ⊢ 𝐵⊥ 𝐶⊥ , 𝐵 ⊗ 𝐶
⊕1 & ⊕2
⊢ 𝐴⊥ , 𝐴 ⊕ (𝐵 ⊗ 𝐶) ⊢ 𝐵⊥ 𝐶⊥ , 𝐴 ⊕ (𝐵 ⊗ 𝐶)
& &
⊢ 𝐴⊥ & (𝐵⊥ 𝐶⊥ ), 𝐴 ⊕ (𝐵 ⊗ 𝐶)

3. Axiom rules for propositional variables only:

ax ax
⊢ 𝐵⊥ , 𝐵 ⊢ 𝐶⊥ , 𝐶
⊗
⊢ 𝐵⊥ , 𝐶 ⊥ , 𝐵 ⊗ 𝐶 &
ax &
⊢ 𝐴⊥ , 𝐴 ⊢ 𝐵⊥ 𝐶⊥ , 𝐵 ⊗ 𝐶
⊕1 & ⊕2
⊢ 𝐴⊥ , 𝐴 ⊕ (𝐵 ⊗ 𝐶) ⊢ 𝐵⊥ 𝐶⊥ , 𝐴 ⊕ (𝐵 ⊗ 𝐶)
& &
⊢ 𝐴⊥ & (𝐵⊥ 𝐶⊥ ), 𝐴 ⊕ (𝐵 ⊗ 𝐶)

Along these lines, we could systematically define expanded proofs of the identity for any formula, the
set of logical rules is such that this expansion is always possible.

21
 • Associativity and commutativity

(𝐴 ⊕ 𝐵) ⊕ 𝐶 ≃ 𝐴 ⊕ (𝐵 ⊕ 𝐶) (𝐴 ⊗ 𝐵) ⊗ 𝐶 ≃ 𝐴 ⊗ (𝐵 ⊗ 𝐶)
𝐴⊕𝐵≃𝐵⊕𝐴 𝐴⊕𝐵≃𝐵⊕𝐴
𝐴⊕0≃𝐴 𝐴⊗1≃𝐴

• Distributivity

𝐴 ⊗ (𝐵 ⊕ 𝐶) ≃ (𝐴 ⊗ 𝐵) ⊕ (𝐴 ⊗ 𝐶) 𝐴⊗0≃0

• Exponentiation

!(𝐴 & 𝐵) ≃ !𝐴 ⊗ !𝐵 !⊤ ≃ 1

• For any 𝐴 and 𝐵, 𝐴 ≃ 𝐵 iff 𝐴⊥ ≃ 𝐵⊥ .

Table 5: Standard isomorphisms in LL

47 is property is valid in any reasonable proof system, and if we look at it from the point of view of
terms in the 𝜆-calculus, we get the standard notion of 𝜂-expansion. For this reason, we keep the same
terminology in the present context. We use the symbol =𝜂 to represent proof equivalence modulo rule
commutation and 𝜂-expansion.
48 Definition. Two formulas 𝐴 and 𝐵 are isomorphic (wrien 𝐴 ≃ 𝐵) if there are proofs 𝜋 of 𝐴 ⊢ 𝐵 and
𝜌 of 𝐵 ⊢ 𝐴 such that [𝜋, 𝜌] is 𝜂-equivalent to the axiom for 𝐴 ⊢ 𝐴 and [𝜌, 𝜋] is 𝜂-equivalent to the
axiom for 𝐵 ⊢ 𝐵.
Among the linear equivalences that we have seen so far, many are actually isomorphisms. Table 5
enumerates the standard ones for LL. e additives and the exponentials provide examples of linear
equivalences that are not isomorphisms:
𝐴⊕𝐴˛𝐴 !𝐴 ⊗ !𝐴 ˛ !𝐴 !!𝐴 ˛ !𝐴 !?!?𝐴 ˛ !?𝐴

49 Exercise. Explain why these are not isomorphisms.

4.3 Reversibility and focalization

&
e proof rule for suggests that in a sequent ⊢ 𝐴1 , …, 𝐴𝑛 , the comma means mostly the same thing as
&
. is is consistent with the definitions of the semantics of proofs, and it is also reminiscent of the fact
that, in intuitionistic and classical logic, 𝐴 ⊢ 𝐵 means mostly the same thing as 𝐴 ⇒ 𝐵 (and in linear
&
logic it means 𝐴 ⊸ 𝐵, which is the same as 𝐴⊥ 𝐵). is property has an interesting consequence
about the structure of proofs.
&
50 Proposition. Every proof of a sequent ⊢ Γ, 𝐴 𝐵 is equivalent modulo commutations of rules to a proof
&
of ⊢ Γ, 𝐴, 𝐵 followed by a rule.
& &
Proof. is can be established by induction on the initial proof of ⊢ Γ, 𝐴 𝐵 by showing that the
& &
rule that introduces 𝐴 𝐵 can be commuted down with any rule coming aer it. If 𝐴 𝐵 is introduced
& &
by an axiom rule, this axiom can be 𝜂-expanded so that the is effectively introduced by a rule.

22
 &
is property is known as reversibility. It is remarkable that is not the only reversible connective,
indeed half of the connectives enjoy this property. e following theorem is a simple generalization of
the above proposition and is no harder to prove.
&
51 eorem. e connectives , &, ⊥, and ⊤ are reversible.
e dual connectives have an associated property called focalization, as defined below. For the pur-
pose of explanation, it is useful to introduce the notion of polarity, which is based on these observations.⁵
52 Definition. A formula is positive if its main connective is ⊗, ⊕, 1, 0 or !. It is negative if its main
&
connective is , &, ⊥, ⊤ or ?.
e reversibility property for negative formulas states in any proof, modulo rule commutations, one
may assume that all negative connectives are introduced in the last rules.
53 eorem. Let Γ = 𝑃1 , …, 𝑃𝑛 be a provable sequent consisting of positive formulas only. en there is a
formula 𝑃𝑖 and proof of ⊢ Γ of the form
𝜋1 ⊢ Γ 1 , 𝑁1 ⋯ 𝜋 𝑘 ⊢ Γ 𝑘 , 𝑁𝑘
𝑅
⊢ Γ 1 , …, Γ 𝑘 , 𝑃𝑖
where the 𝑁𝑗 are the maximal negative subformulas of 𝑃𝑖 and the last set of rules 𝑅 builds 𝑃𝑖 from the 𝑁𝑗 .
In other words, in a positive sequent, there is a “principal” formula 𝑃𝑖 such that we may assume that
the last rules of the proof build 𝑃𝑖 in one big step, without affecting formulas in the context. Hence, the
proof has a “focus” on 𝑃𝑖 .

5 Proof nets
e study of sequent calculus proofs and their dynamics in cut elimination is not a straightforward
task, mostly because the system has heavy notations and a lot of technical details, like commutation
rules, that do not feel crucial but make things obfuscated. In proof theory, natural deduction is oen
a preferred formalism because it is indeed more natural and easier to manipulate, as illustrated by the
symptomatic example of the 𝜆-calculus, which is a compact and efficient syntax for natural deduction
in intuitionistic logic. is calculus contains the essence of intuitionistic logic and its cut elimination
dynamics, as witnessed by Böhm’s separation theorem and similar results. One would like a similar tool
for linear logic, and proof nets are the answer to this question.

5.1 Intuitionistic LL and natural deduction

What makes the Curry-Howard correspondence work in 𝜆-calculus is the intrinsic asymmetry of in-
tuitionistic logic: from some set of hypothesis (the free variables) deduce one statement (the term, or
rather its result). is asymmetry is a restriction with respect to classical logic but it allows nice sim-
plifications and beer property that stem from the functional nature of the resulting system.
e same restriction can also be applied in linear logic: consider the original two-sided presentation
and impose that exactly one formula appears on the right of each sequent. e rules for negation and
multiplicative disjunction are now forbidden, hence the associated connectives are dropped. However,
linear implication ⊸, which is usually defined using them, is acceptable if we use the rules of exercise 21.
e system we get is multiplicative intuitionistic linear logic, in short MILL.
⁵e focalization property was introduced in Andreoli, “Proposition pour une synthèse des paradigmes de la programmation
logique et de la programmation par objets” with motivations related to computational interpretation of proof search in linear
logic. It led to the definition of polarities in Girard, “A new constructive logic: classical logic” as properties of formulas in classical
logic.

23
 Identity
ax
𝑥∶𝐴⊢𝑥∶𝐴
Implication
Γ, 𝑥 ∶ 𝐴 ⊢ 𝑡 ∶ 𝐵 Γ⊢𝑡∶𝐴⊸𝐵 Δ⊢𝑢∶𝐴
⊸R ⊸E
Γ ⊢ 𝜆𝑥.𝑡 ∶ 𝐴 ⊸ 𝐵 Γ, Δ ⊢ (𝑡)𝑢 ∶ 𝐵
Tensor
Γ⊢𝑡∶𝐴 Δ⊢𝑢∶𝐵 Γ, 𝑥 ∶ 𝐴, 𝑦 ∶ 𝐵 ⊢ 𝑡 ∶ 𝐶 Δ⊢𝑢∶𝐴⊗𝐵
⊗R ⊗E
Γ, Δ ⊢ (𝑡, 𝑢) ∶ 𝐴 ⊗ 𝐵 Γ, Δ ⊢ 𝑡(𝑥,𝑦∶=𝑢) ∶ 𝐶

Table 6: Intuitionistic linear logic as natural deduction

54 Definition. e language of formulas of MILL is defined by the following grammar:

𝐴, 𝐵 ∶= 𝛼 propositional variable
𝐴⊸𝐵 linear implication — read “𝐴 implies 𝐵”
𝐴⊗𝐵 multiplicative conjunction — read “𝐴 tensor 𝐵”

Proof terms are defined by the following grammar:

𝑡, 𝑢 ∶= 𝑥 variable — axiom
𝜆𝑥.𝑡 linear abstraction — introduction of ⊸
(𝑡)𝑢 linear application — elimination of ⊸
(𝑡, 𝑢) pair — introduction of ⊗
𝑡(𝑥,𝑦∶=𝑢) matching — elimination of ⊗

where 𝑥 and 𝑦 are taken from a set of variables. e variable 𝑥 is bound in 𝜆𝑥.𝑡, the variables 𝑥 and 𝑦 are
distinct and bound in 𝑡(𝑥,𝑦∶=𝑢). Standard conventions apply: bound names are assumed to be distinct
from free names and terms are considered up to renaming of bound names.
A context (wrien Γ or Δ) is a partial mapping from variables to formulas with a finite domain.
Proof terms are typed according to the rules of table 6, where Γ, Δ stands for the union of contexts, with
the assumption that they have disjoint domains.
Remark that the fragment without the tensor and its associated constructs is exactly the 𝜆-calculus,
with contexts handled in multiplicative style, restricted so that contraction and weakening are forbidden.
As a direct consequence, in typed terms, each variable that is either bound or declared in the context
has exactly one occurrence in the term. e tensor rules provide a simple extension with pairs.
55 Lemma. Let 𝑡[𝑢/𝑥] denote the term 𝑡 where the variable 𝑥 is replaced by the term 𝑢. If the typings Γ, 𝑥 ∶
𝐴 ⊢ 𝑡 ∶ 𝐵 and Δ ⊢ 𝑢 ∶ 𝐴 hold and Γ and Δ have disjoint domains, then Γ, Δ ⊢ 𝑡[𝑢/𝑥] ∶ 𝐵 holds.
Proof. Straightforward induction on the typing of 𝑡.
In other words, the following cut rule is admissible:
Γ, 𝑥 ∶ 𝐴 ⊢ 𝑡 ∶ 𝐵 Δ⊢𝑢∶𝐴
cut
Γ, Δ ⊢ 𝑡[𝑢/𝑥] ∶ 𝐵
is allows the definition of the cut elimination reduction using substitution, in a way similar to the
𝛽-reduction of usual 𝜆-calculus.

24
 Γ Γ Δ Γ Δ Γ Δ
𝑥∶𝐴

𝑥∶𝐴 𝑡 𝑡 𝑢 𝑡 𝑢 𝑢
𝐴⊸𝐵 𝐴 𝐴⊗𝐵
ax. .
⊸ ⊸. ⊗. 𝑡. ⊗

𝐴 𝐴⊸𝐵 𝐵 𝐴⊗𝐵 𝐶

𝑥 𝜆𝑥.𝑡 (𝑡)𝑢 (𝑡, 𝑢) 𝑡(𝑥,𝑦∶=𝑢)

Figure 1: Graphical notation for MILL proof terms

56 Definition. Cut elimination for MILL is the congruent binary relation ⇝ over proof terms generated
by the following rules:
(𝜆𝑥.𝑡)𝑢 ⇝ 𝑡[𝑢/𝑥] 𝑡(𝑥,𝑦∶=(𝑢, 𝑣)) ⇝ 𝑡[𝑢/𝑥][𝑣/𝑦]
57 eorem. Cut elimination in MILL is strongly normalizing.
Proof. Confluence of ⇝ is easy to prove: contrary to the case of usual 𝜆-calculus, this relation in MILL
is strongly confluent (i.e. it has the diamond property) thanks to linearity.
Termination is also easy: each step strictly decreases the number of rules in the typing of the terms.
Indeed, (𝜆𝑥.𝑡)𝑢 ⇝ 𝑡[𝑢/𝑥] removes three rules (⊸R, ⊸E, and the axiom for 𝑥) and 𝑡(𝑥,𝑦∶=(𝑢, 𝑣)) ⇝
𝑡[𝑢/𝑥][𝑣/𝑦] removes four rules (⊗R, ⊗E, and the axioms for 𝑥 and 𝑦).
e structure of terms and cut-elimination may be illustrated using graphical notations. In figure 1,
we propose such a notation: a term is represented by a graph, with nodes corresponding to logical rules
and edges corresponding to occurrences of formulas in a proof. Orientation is relevant in our notations:
dangling edges above a graph represent hypotheses, there is one such edge for each element of the
typing context; the unique dangling edge below is the conclusion of the proof. Nodes pointing down
represent introduction rules and nodes pointing up represent elimination rules, edges between nodes
represent how each rule relates formulas, with the intuition that inputs are above and outputs are below.
Cut elimination rules in this context are formulated as graph rewritings, as shown in figure 2. ey
occur when the output of an introduction rule is connected to the input of an elimination rule. Indeed,
because of the intuitionistic structure of MILL, there is no need for commutation rules. Interaction steps
appear as purely local graph rewriting operations: the two interacting rules plus the associated axiom
rules are replaced by edges connecting the relevant parts of proofs. Note in particular that, again thanks
to linearity, substitution is a very simple and local operation.

5.2 Proof structures

We now extend the graphical language of proofs to the full system of MLL. e transition from the
intuitionistic fragment to the full one-sided system follows the same path as the transition between
these systems in sequent calculus:
• Firstly, we allow several formulas on the right hand side of sequents. Subsequently, in the graph-
ical language, there will be any number of outputs (or conclusions, i.e. dangling edges below).
&
is allows for the reintroduction of , which graphically simply joins together two conclusions
of a proof, and linear negation, which transforms a hypothesis into a conclusion or vice-versa.

25
Linear implication:

Γ Δ Γ Δ

ax
𝑢
𝑡 𝑢
⇝ 𝑡
⊸

. 𝐵 . 𝐵
Tensor:
Γ Δ Θ Γ Δ Θ

𝑢 𝑣 𝑢 𝑣

ax ⊗ ⇝
ax
𝑡 ⊗ 𝑡

𝐶. 𝐶.

Figure 2: Cut elimination steps in MILL

26
 𝜋 𝜌 𝜋 𝜌 𝜋
ax 𝐴⊥ 𝐴 𝐴 𝐵 𝐴 𝐵
cut
&
⊗
𝐴.⊥ 𝐴
Γ. Γ. Γ.
&
Δ 𝐴⊗𝐵 Δ 𝐴 𝐵
𝜋⊢ Γ, 𝐴⊥ 𝜌 ⊢ 𝐴, Δ 𝜋 ⊢ Γ, 𝐴 𝜌 ⊢ 𝐵, Δ 𝜋 ⊢ Γ, 𝐴, 𝐵 &
ax cut ⊗ &
⊢ 𝐴⊥ , 𝐴 ⊢ Γ, Δ ⊢ Γ, 𝐴 ⊗ 𝐵, Δ ⊢ Γ, 𝐴 𝐵

Figure 3: Translation of MLL sequent calculus rules into proof structures.

• Secondly, we hard-wire De Morgan duality, so that negation becomes an operation on formulas

and sequents become one-sided. Graphically, this means that there are no inputs (i.e. dangling
edges above) anymore and that axiom and cut now involve two edges bearing opposite formulas.
e four rules of MLL are translated graphically as shown in figure 3 (where 𝜋 ⊢ Γ states the fact that
𝜋 is a graph representing a proof of conclusion Γ). Formally, the kind of graphs built with these rules
are called proof structures.
58 Definition. An MLL proof structure is a directed multigraph with edges labelled by MLL formulas and
nodes labelled by either rule names or the special symbol “c”. Moreover, for each node, incoming edges
(called premisses) and outgoing edges (called conclusions) are equipped with a total order.
e node of each label imposes constraints on the number of incoming and outgoing edges as well as
their labels, according to the corresponding proof rules. e labels must match the following constraints:
&
node label c ax cut ⊗
labels of incoming edges 𝐴 — 𝐴, 𝐴⊥ 𝐴, 𝐵 𝐴, 𝐵
&
labels of outgoing edges — 𝐴⊥ , 𝐴 — 𝐴⊗𝐵 𝐴 𝐵
e nodes labeled “c” are called the conclusions of the structure.
In the translation of proof trees into proof structures given by figure 3, there are some implicit ele-
ments with respect to the above definition. Firstly, all edges are oriented from top to boom. Secondly,
conclusion nodes (labelled “c”) are not drawn, instead we write the label of their incoming edge. We
use double lines leading to a sequence Γ or Δ to represent an unspecified number of conclusions.
Finally, the natural handling of conclusion nodes when building structures is also implicit. For
instance, in the rule for ⊗, the final structure is built using two structures 𝜋 and 𝜌, introducing a new
node labelled ⊗, changing the target of the edges for the premisses 𝐴 and 𝐵 so that they lead to this new
node instead of their conclusion in 𝜋 and 𝜌, introducing a new conclusion node for 𝐴 ⊗ 𝐵 and removing
the conclusions nodes for 𝐴 and 𝐵.
59 Example. A variant of the semi-distributivity rule from exercise 16, equivalent by commutativity of ⊗
& & & &
and , is wrien (𝐶 𝐵)⊗𝐴 ⊢ (𝐴⊗𝐵) 𝐶. In one-sided form, it is thus wrien ⊢ (𝐶⊥ ⊗𝐵⊥ ) 𝐴⊥ , (𝐴⊗
&
𝐵) 𝐶. A possible sequential proof of this sequent, with its translation as a proof structure, is given in
figure 4. Note that we have placed the various nodes correctly so that the edges do not intersect, but
this is just a maer of graph representation.
60 In this example, if we only look at the proof structure, we cannot deduce the sequential proof it was
&
deduced from. Indeed, we get the same structure if we permute the two rules (this is consistent with
the reversibility property described in section 4.3) and we also get the same structure if we exchange
the order of the ⊗ rules. is defines four sequential proofs that have the same proof structure, and by

27
 ax
𝐶⊥
ax ax 𝐶
𝐵⊥ ax 𝐵
⊢ 𝐶⊥ , 𝐶 ⊢ 𝐵⊥ , 𝐵
ax ⊗ 𝐴
⊢ 𝐴⊥ , 𝐴 ⊢ 𝐶⊥ ⊗ 𝐵⊥ , 𝐵, 𝐶 𝐴⊥ ax
⊗ ⟹
⊢ 𝐶⊥ ⊗ 𝐵⊥ , 𝐴⊥ , 𝐴 ⊗ 𝐵, 𝐶 & ⊗ ⊗
&
⊢ (𝐶⊥ ⊗ 𝐵⊥ ) 𝐴⊥ , 𝐴 ⊗ 𝐵, 𝐶 & & &
& &
⊢ (𝐶⊥ ⊗ 𝐵⊥ ) 𝐴⊥ , (𝐴 ⊗ 𝐵) 𝐶
(𝐶. ⊥ ⊗ 𝐵⊥ )
& &
𝐴⊥ (𝐴 ⊗ 𝐵) 𝐶

Figure 4: Example of a sequential proof and its translation as a proof structure.

ax
𝐶⊥
𝐶
𝐵⊥ ax 𝐵
ax 𝐴
𝐴⊥
& &

⊗ ⊗

(𝐶. ⊥
& &
𝐵⊥ ) ⊗ 𝐴⊥ (𝐴 𝐵) ⊗ 𝐶

Figure 5: An incorrect proof structure.

case analysis it is easy to check that there can be no other sequential proof for this sequent and that all
such proofs lead to this structure.
61 Approaching the question differently, one could try to enumerate all cut-free proof structures that
& &
have the conclusions (𝐶⊥ ⊗ 𝐵⊥ ) 𝐴⊥ and (𝐴 ⊗ 𝐵) 𝐶 (assuming that 𝐴, 𝐵 and 𝐶 are propositional
variables). Clearly the proof structure of figure 4 is the only one we can write.
e situation would be different if, say, 𝐶 was replaced by 𝐵. In this case, there would be two
different structures, depending on which premiss 𝐵⊥ of the le tensor is connected to which premiss 𝐵
&
on the right. ese would correspond to two different classes of sequential proofs of ⊢ (𝐵⊥ ⊗ 𝐵⊥ )
&
𝐴⊥ , (𝐴 ⊗ 𝐵) 𝐵 that are not identified by the translation into structures.
62 Definition. A proof net is a proof structure that is the translation of some sequential proof.
is notion refers to the fact that not any proof structure is actually the translation of a sequential
proof, because the definition of proof structures only mentions the local structure of each logical rule.
&
Indeed, the constraints for ⊗ and are exactly the same, while their sequent calculus counterparts are
&
fundamentally different since has one premiss but ⊗ has two.
&
63 Example. e proof structure of figure 5, a variant of the one of figure 4 with ⊗ and exchanged, is
&
not a proof net. To establish this fact, we can simply prove that the conclusion sequent ⊢ (𝐶⊥ 𝐵⊥ ) ⊗
&
𝐴⊥ , (𝐴 𝐵) ⊗ 𝐶 is not provable without cut, since the structure of figure 4 has no cut. is is done by
case analysis on the last rule of a hypothetical cut-free proof of this sequent:
& &
⊢ 𝐶⊥ 𝐵⊥ , Γ ⊢ 𝐴⊥ , Δ ⊢ Γ, 𝐴 𝐵 ⊢ Δ, 𝐶
& & ⊗ or & & ⊗
⊢ (𝐶⊥ 𝐵⊥ ) ⊗ 𝐴⊥ , (𝐴 𝐵) ⊗ 𝐶 ⊢ (𝐶⊥ 𝐵⊥ ) ⊗ 𝐴⊥ , (𝐴 𝐵) ⊗ 𝐶
&
for some Γ and Δ. In the first rule, among Γ and Δ, one must be empty and the other must be (𝐴
𝐵) ⊗ 𝐶. In either case, the premisses do not respect the linearity property for propositional variables

28
 Axiom (assuming the right premiss of the cut node is not the le conclusion of the axiom node):

ax 𝐴 𝐴
⇝
𝐴. cut 𝐴.

Tensor versus par:

𝐴 𝐵 𝐴⊥ 𝐵⊥ 𝐴 𝐵 𝐴⊥ 𝐵⊥
&
⊗ cut
⇝
. cut . cut
Plus the same rules with the le and right premisses of the cut exchanged.

Figure 6: Cut elimination steps in MLL.

(proposition 17) so they cannot be provable. A similar argument applies to the other possibility for the
last rule.
In this example, we already see that the crucial point for the tensor is that it splits the context in
two, which implies spliing the proof in two independent subproofs.
& &
64 Exercise. Enumerate all the cut-free proof structures with conclusions (𝐴⊥ ⊗𝐴⊥ ) 𝐴⊥ and (𝐴⊗𝐴) 𝐴
and identify which ones are proof nets.
We now have a graphical language in which we can represent proofs, in a way that identifies se-
quential proofs that differ only by commutation of independent rules. In this seing, cut elimination
steps are much simpler than in sequential proofs because only interaction rules are relevant: each cut
has premisses that are introduction nodes or axiom nodes.
65 Definition. Cut elimination for MLL proof structures is the binary relation ⇝ over proof structures
such that 𝜋 ⇝ 𝜌 if and only if 𝜌 is obtained from 𝜋 by applying one of the local graph rewriting rules
of figure 6.

66 Proposition (Strong normalization). In any MLL proof structure, all maximal sequences of cut elimina-
tion steps are finite, all have the same length and they all reach the same irreducible proof structure (up to
graph isomorphism).
Proof. e finiteness of cut elimination sequences is obvious since each step strictly decreases the num-
ber of nodes. at all sequences have the same length and reach the same irreducible structure is easy
since the relation is strongly confluent. e only critical pairs are when a cut is between the conclusions
of two distinct axiom nodes and when an axiom has both its conclusions connected to different cuts,
and in these cases the reducts are isomorphic.

Observe that this property does not imply cut elimination, because there is one case of cut that is
not reducible:

29
 ax

.
cut

is kind of “vicious circle” is not a proof net, and thankfully so since it would be a proof of the empty
sequent, since it has no conclusion node! Subsequently, it is clear that an irreducible proof net is always
cut-free.
is leaves one question open in order to prove cut elimination for proof nets: how do we know
that the reduct of a proof net is always a proof net, and not an incorrect proof structure? Correctness
criteria provide the answer to this question.

5.3 Correctness criteria

A correctness criterion is a characterization of correct proofs among proof structures. In other words, it
must characterize whether a given proof structure is sequentializable, i.e. if there is a sequential proof of
which it is the graphical translation. Several such criteria exist in the literature, for various fragments
of linear logic, motivated by various uses:
• It should be reasonably easy to prove that correctness is preserved by cut elimination.

• e efficiency of actually computing whether a structure satisfies the criterion should be known
(and minimized), because it is directly related to the complexity of the decision problem for the
considered logic.
Although the original criterion is the so-called long trip criterion, we will present here the Danos-
Regnier criterion,⁶ which is in the standard one in practice and is used as reference in the development
of refined criteria.
As a gentle introduction, we will approach the problem of sequentializability from what we know
of sequential proofs. Let us consider an arbitrary MLL proof structure 𝜋. e first interesting thing is
&
reversibility of , as shown in theorem 51. It implies that 𝜋 is sequentializable if and only if it is the
&
image of a sequential proof that ends with the introduction of any that is in conclusion. In other
&
words, in conclusion are irrelevant for correctness.
&
We can actually generalize this property to any rule, not only in the conclusion of a proof. Con-
&
sider a cut-free proof that contains a rule acting on two formulas 𝐴 and 𝐵; since the proof has no cut,
& &
𝐴 𝐵 is a subformula of a conclusion, which we can write Φ(𝐴 𝐵):
⋮
𝐴 𝐵
⊢ Γ, 𝐴, 𝐵 &
& &
⊢ Γ, 𝐴 𝐵
⋮ 𝐴
&
𝐵
&
⊢ Δ, Φ(𝐴 𝐵) . &
Δ Φ(𝐴 𝐵)
&
It is easy to check that the rule can be removed, which changes the conclusions but not the validity
&
of the proof. Instead of 𝐴 𝐵, we have two formulas 𝐴 and 𝐵, and we may choose to apply the same
&
rules, not to 𝐴 𝐵 but to 𝐴, for instance, leaving 𝐵 as an extra conclusion:
⁶is criterion was introduced in Danos and Regnier, “e structure of multiplicatives” together with other criteria, as part of
a thorough study of proofs in MLL.

30
 ⋮ 𝐴 𝐵
⊢ Γ, 𝐴, 𝐵
⋮
⊢ Δ, Φ(𝐴), 𝐵 𝐴
.
Δ Φ(𝐴) 𝐵

Of course, we could have made the opposite choice, kept 𝐵 inside the proof and turned 𝐴 into an extra
&
conclusion. Such a choice for a is called a switching.
&
67 Definition. A switching 𝑠 of a proof structure 𝜋 is a function mapping each -labelled node of 𝜋 to
&
either 1 or 2. e switched structure 𝑠(𝜋) is is the structure obtained by replacing each node 𝑛 in 𝜋
by an edge from its 𝑠(𝑛)-th premiss to its conclusion and its other premiss to a new conclusion node.
&
As a consequence of the above remark, the operation of switching a in a proof structure preserves
&
correctness. If we switch all rules (assume there are 𝑛 of them), then we get new proof structures
&
with no cut and no (there are 2𝑛 of them, if we consider all possible switchings).
68 Lemma. If 𝜋 is a correct cut-free proof structure, then for all switchings 𝑠 of 𝜋, 𝑠(𝜋) is correct.
Now, what can we say about a sequentializable proof structure that only contains axiom and ⊗
nodes? e shape of the translation of a ⊗ rule (as in figure 3) shows that applying such a rule means
taking two proof nets and connecting one node of each to a new node (which is connected to a fresh
conclusion). Because there are two independent proof nets as premisses in the rule, this cannot create
a cycle (even in the underlying undirected graph). Moreover, if the premisses were connected graphs,
then the resulting structure is connected too. is is the property we are aer.
69 Lemma. A proof structure with only axiom and ⊗ nodes is correct if and only if, as an undirected graph, it
is acyclic and connected.

Proof. e direct implication is justified by the above remarks and can be proved by induction on a
sequential proof. e reverse implication is proved by recurrence on the number of ⊗ nodes in the
structure, it suffices to remark that removing any ⊗ node connected to a conclusion effectively discon-
nects the structure into two connected components, which are correct by recurrence hypothesis; the
base case of no ⊗ node corresponds to a single axiom node, which is obviously a correct proof.
is condition of acyclicity and connectedness of switchings is the Danos-Regnier criterion.
70 eorem (Danos-Regnier). An MLL proof structure is sequentializable if and only if all its switchings are
acyclic and connected.

Sketch of proof. e “only i” part is essentially contained in the arguments above. For the “i” part, the
key point is to prove that the condition does imply the existence of a spliing ⊗ node in a structure
with only ⊗ as conclusions that respects the criterion. ere are several methods to establish this in the
literature but we will not elaborate on this point in the present notes.

Note that the above theorem does not exclude the presence of cut nodes. Indeed, cuts do not change
the problem significantly: by similar arguments as above, one can justify that a structure with cuts is
correct if and only if the same structure where cuts are replaced by tensor rules is correct. is is a
consequence of the fact that the tensor and cut rule have strictly the same geometrical structure (two
subproofs as premisses, with one formula taken from each).
71 eorem. Cut elimination preserves correctness of proof structures.

31
Proof. e case of the elimination of the axiom is straightforward, so we focus on the case of tensor
versus par. We have a reduction with the shape

𝐴 𝐵 𝐴⊥ 𝐵⊥
& 𝐴 𝐵 𝐴⊥ 𝐵⊥
⊗
⇝
cut cut
. cut
.
such that all switchings of the le-hand side are acyclic and connected, and we must show that the
switchings of the right-hand side are acyclic and connected too. Consider a switching 𝑠 of the right-
hand side. It can be extended as 𝑠1 , the switching of the le-hand side that keeps the first premiss, and
𝑠2 , the other way around. e switched structure for 𝑠1 must have the shape

𝜋1 𝜋2 𝜋3 𝜋4

. cut
where the 𝜋𝑖 are connected and acyclic subgraphs, with 𝜋4 connected to one of the other components
because of connectedness of 𝑠1 (𝜋). If 𝜋4 was connected to 𝜋1 or 𝜋2 , then in the switching 𝑠2 we would
have a cycle, since the tensor is connected to 𝜋4 , but that cannot happen by hypothesis. Hence 𝜋4 is
necessarily connected to 𝜋3 and in the reduct we have

𝜋1 𝜋2 𝜋3 𝜋4

cut cut
which is acyclic and connected. Hence the
. reduct satisfies the Danos-Regnier criterion, so it is a proof
net.

is fills the missing part in the proof of cut elimination for proof nets, since we had strong normal-
ization in theorem 66. As a consequence, proof nets are a proper syntax for proofs in MLL for which
we can prove the key properties directly, without reference to proof trees. e graphical formulation
gets rid of all unimportant commutations between rules, so we can even consider that proof nets are
canonical proof objects.

e extension of this technique to wider fragments of linear logic is a rich topic, with various open
questions. e formulation of proof nets as canonical as in MLL is oen unknown but approximations
exist.
• e introduction of multiplicative units requires some technique to formulate the proper notion
of proof net, because naive extensions with a ⊥ node break the connectivity condition.

• For MELL, the situation is almost as good, except that the promotion rule forces sequentiality in
the form of boxes: each instance of this rule, together with its subproof, becomes a single node in
the proof net, with as many outputs as there are formulas in the conclusion of the rule; the node
is associated with an independent proof net for the subproof (the “contents” of the box).

32
 • Additive connectives are known to be difficult, and the proper solution has been found nearly
twenty years aer the introduction of proof nets, in Hughes and Glabbeek, “Proof nets for unit-
free multiplicative-additive linear logic”.

References
Andreoli, Jean-Marc. “Proposition pour une synthèse des paradigmes de la programmation logique et
de la programmation par objets”. èse de doctorat. Université de Paris VI, 1990.
Danos, Vincent, Jean-Baptiste Joinet, and Harold Schellinx. “A new deconstructive logic: linear logic”.
In: Journal of Symbolic Logic 62 (1996), pp. 755–807.
Danos, Vincent and Laurent Regnier. “e structure of multiplicatives”. In: Archive for Mathematical
Logic 28 (1989), pp. 181–203.
Girard, Jean-Yves. “A new constructive logic: classical logic”. In: Mathematical Structures in Computer
Science 1.3 (1991), pp. 255–296.
— “Linear Logic”. In: eoretical Computer Science 50 (1987), pp. 1–102.
— “Linear Logic: Its Syntax and Semantics”. In: Advances in Linear Logic. Ed. by Jean-Yves Girard, Yves
Lafont, and Laurent Regnier. Cambridge University Press, 1995, pp. 1–42.
— “e system F of variable types, fieen years later”. In: eoretical Computer Science 45.2 (1986),
pp. 159–192. : http://dx.doi.org/10.1016/0304-3975(86)90044-7.
Girard, Jean-Yves, Yves Lafont, and Paul Taylor. Proofs and types. Cambridge University Press, 1989.
Hughes, Dominic J. D. and Robert J. van Glabbeek. “Proof nets for unit-free multiplicative-additive linear
logic”. In: 18th IEEE Symposium on Logic in Computer Science (LICS). June 2003, pp. 1–10.
Lincoln, Patrick et al. “Decision problems for propositional linear logic”. In: Annals of Pure and Applied
Logic 56.1-3 (1992), pp. 239–311. : http://dx.doi.org/10.1016/0168-0072(92)90075-B.

33