from display calculi to decision procedures for full ...apt13.unibe.ch/slides/gore.pdf · hyland...
TRANSCRIPT
From Display Calculi to Decision Procedures forFull Intuitionistic Linear Logic
Ranald Clouston, Jeremy Dawson, Rajeev Gore, Alwen Tiu
Logic and Computation GroupResearch School of Computer ScienceThe Australian National University
December 14, 2013
Published in Proc. 22nd EACSL Annual Conference
on Computer Science Logic, Turin September 2013 as
“Annotation-free sequent calculi for Full Intuitionistic Linear Logic”
Overview
What is FILL?
Existing sequent calculi
A Display Calculus for FILL
Nested Sequent Calculus for FILL
Separation
Decidability and Complexity
Further Work
Full Intuitionistic Linear Logic
LL: modal substructural logic without weakening and withoutcontraction ⊗,⊕, 1, 0,(,∧,∨,>,⊥, !, ? Girard 1987
MALL: ⊗,⊕, 1, 0,(,∧,∨,>,⊥ drop exponentials
MILL: ⊗, 1,( intuitionistic
MLL: ⊗, 1,(,⊕, 0 classical ((A ( 0) ( 0) ( A
FILL: ⊗, 1,(,⊕, 0 Hyland and de Paiva 1993
Categorial Semantics for FILL
(⊗, 1,() is a symmetric monoidal closed structure
A⊗ B ( C iff A ( (B ( C ) iff B ( (A ( C )
(A⊗ 1) ( A and A ( (A⊗ 1)
(⊕, 0) is a symmetric monoidal structure
(A⊕ B) ( (B ⊕ A)
(A⊕ 0) ( A and A ( (A⊕ 0)
interaction via either of
weak distributivity ((A⊗ B)⊕ C ) ( (A⊗ (B ⊕ C ))
Grishin(b) ((A ( B)⊕ C ) ( (A ( (B ⊕ C ))
Collapse to (classical) MLL: if we add converse of Grishin(b)Grishin(a) (A ( (B ⊕ C )) ( ((A ( B)⊕ C )
Proof Theory of Full Intuitionistic Linear Logic
LL: substructural logic without weakening and without contraction
MILL ⊗ ( 1 intuitionistic cut-elimination
Γ1 ` A Γ2 ` BΓ1, Γ2 ` A⊗ B
Γ,A ` B
Γ ` A ( B
MLL ⊗ ( 1 ⊕ ⊥ classical cut-elimination
Γ1,A ` B,∆1 Γ2,A ` B,∆2
Γ1, Γ2 ` ∆1,A⊗ B,∆2
Γ ` A,B,∆
Γ ` A⊕ B,∆
Γ,A ` B,∆
Γ ` A ( B,∆
FILL ⊗ ( 1 ⊕ ⊥ intuitionistic cut-elimination fails
Γ1,A ` B,∆1 Γ2,A ` B,∆2
Γ1, Γ2 ` ∆1,A⊗ B,∆2
Γ ` A,B,∆
Γ ` A⊕ B,∆
Γ,A ` B
Γ ` A ( B,∆
Problem and a solutions via annotated derivations
Remember: we need comma on the right to accommodate ⊕
Problem and existing solutions:
multiple conclusions single conclusion existing solutions
Γ,A ` B,∆
Γ ` A ( B,∆
Γ,A ` B
Γ ` A ( B,∆
Γ,A ` B,∆(†)
Γ ` A ( B,∆
unsound no cut-elimination cut-elimination
†: side-conditions which ensure that A is “independent” of ∆
Hyland, de Paiva 1993: type assignments to ensure that thevariable typed by A not appear free in the terms typed by ∆
Further problems and solutions using annotations
Hyland and de Paiva 1993:Γ,A ` B,∆
(†)Γ ` A ( B,∆
†: side-conditions which ensure that A is “independent” of ∆
Bierman 1996: (a⊕ b)⊕ c ` a, ((b ⊕ c) ( d)⊕ (e ( (d ⊕ e))has no cut-free derivation in the Hyland and de Paiva calculus
Bierman, Bellin 1996: refined annotations to regain cut-elimination
Brauner and de Paiva 1997: annotate rules with a binary relationbetween antecedent formulae and succedent formulae, whicheffectively trace variable occurrence
What do derivations look like ?
Hyland , de Paiva and Biermann
(( R) legal as v and (w ⊕ x ( y)⊕ z share no free variables.
v : a ` v : a w : b ` w : bv ⊕ w : a⊕ b ` v : a,w : b x : c ` x : c
(v ⊕ w)⊕ x : (a⊕ b)⊕ c ` v : a,w : b, x : c
(v ⊕ w)⊕ x : (a⊕ b)⊕ c ` v : a,w ⊕ x : b ⊕ c y : d ` y : d
(v ⊕ w)⊕ x : (a⊕ b)⊕ c,w ⊕ x ( y : b ⊕ c ( d ` v : a, y : d z : e ` z : e
(v ⊕ w)⊕ x : (a⊕ b)⊕ c, (w ⊕ x ( y)⊕ z : (b ⊕ c ( d)⊕ e ` v : a, y : d , z : e
(v ⊕ w)⊕ x : (a⊕ b)⊕ c, (w ⊕ x ( y)⊕ z : (b ⊕ c ( d)⊕ e ` v : a, y ⊕ z : d ⊕ e
(v ⊕ w)⊕ x : (a⊕ b)⊕ c ` v : a, λ(w ⊕ x ( y)⊕ z(b⊕c(d)⊕e .(y ⊕ z) : (b ⊕ c ( d)⊕ e ( d ⊕ e
But the type annotations have no computational content
Bellin-style Proof
(( R) is legal because r is not free inlet t be u ⊕ - in let u be v ⊕ - in v
v : a ` v : a w : b ` w : b
u : a ⊕ b ` let u be v ⊕ - in v : a, let u be -⊕ w in w : b x : c ` x : c
t : (a ⊕ b)⊕ c ` let t be u ⊕ - in let u be v ⊕ - in v : a, let t be u ⊕ - in let u be -⊕ w in w : b, let t be -⊕ x in x : c
t : (a ⊕ b)⊕ c ` let t be u ⊕ - in let u be v ⊕ - in v : a, (let t be u ⊕ - in let u be -⊕ w in w)⊕ (let t be -⊕ x in x) : b ⊕ c y : d ` y : d
t : (a ⊕ b)⊕ c, s : b ⊕ c ( d ` let t be u ⊕ - in let u be v ⊕ - in v : a, (s(let t be u ⊕ - in let u be -⊕ w in w)⊕ (let t be -⊕ x in x)) : d z : e ` z : e
t : (a ⊕ b)⊕ c, r : (b ⊕ c ( d)⊕ e ` let t be u ⊕ - in let u be v ⊕ - in v : a, let r be s ⊕ - in (s(let t be u ⊕ - in let u be -⊕ w in w)⊕ (let t be -⊕ x in x)) : d, let s be -⊕ z in z : e
t : (a ⊕ b)⊕ c, r : (b ⊕ c ( d)⊕ e ` let t be u ⊕ - in let u be v ⊕ - in v : a, (let r be s ⊕ - in (s(let t be u ⊕ - in let u be -⊕ w in w)⊕ (let t be -⊕ x in x)))⊕ (let s be -⊕ z in z) : d ⊕ e
t : (a ⊕ b)⊕ c ` let t be u ⊕ - in let u be v ⊕ - in v : a, λr (b⊕c(d)⊕e .(let r be s ⊕ - in (s(let t be u ⊕ - in let u be -⊕ w in w)⊕ (let t be -⊕ x in x)))⊕ (let s be -⊕ z in z) : (b ⊕ c ( d)⊕ e ( d ⊕ e
Again, the type annotations are not given any computationalcontent and this derivation does not even fit on the page!
Brauner and de Paiva-style proof
(( R) legal because (b ⊕ c ( d)⊕ e is not related to a
(a, a)a ` a
(b, b)b ` b
(a⊕ b, a), (a⊕ b, b)a⊕ b ` a, b
(c, c)c ` c
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, b), ((a⊕ b)⊕ c, c)(a⊕ b)⊕ c ` a, b, c
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, b ⊕ c)(a⊕ b)⊕ c ` a, b ⊕ c
(d , d)d ` d
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, d), (b ⊕ c ( d , d)(a⊕ b)⊕ c, b ⊕ c ( d ` a, d
(e, e)e ` e
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, d), ((b ⊕ c ( d)⊕ e, d), ((b ⊕ c ( d)⊕ e, e)(a⊕ b)⊕ c, (b ⊕ c ( d)⊕ e ` a, d , e
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, d ⊕ e), ((b ⊕ c ( d)⊕ e, d ⊕ e)(a⊕ b)⊕ c, (b ⊕ c ( d)⊕ e ` a, d ⊕ e
((a⊕ b)⊕ c, a), ((a⊕ b)⊕ c, (b ⊕ c ( d)⊕ e ( d ⊕ e)(a⊕ b)⊕ c ` a, (b ⊕ c ( d)⊕ e ( d ⊕ e
Pure annotational device to track variable occurrences
Cut-free derivation in our display calculus
a ` a b ` ba⊕ b ` a, b c ` c
(a⊕ b)⊕ c ` a, b, c
(a⊕ b)⊕ c < a ` b, c
(a⊕ b)⊕ c < a ` b ⊕ c d ` d
b ⊕ c ( d ` ((a⊕ b)⊕ c < a) > d e ` e
(b ⊕ c ( d)⊕ e ` (((a⊕ b)⊕ c < a) > d), e
(b ⊕ c ( d)⊕ e ` ((a⊕ b)⊕ c < a) > d , e
(b ⊕ c ( d)⊕ e, ((a⊕ b)⊕ c < a) ` d , e
(b ⊕ c ( d)⊕ e, ((a⊕ b)⊕ c < a) ` d ⊕ e
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e > d ⊕ e
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e ( d ⊕ e
(a⊕ b)⊕ c ` a, (b ⊕ c ( d)⊕ e ( d ⊕ e
No annotations, but many extra structural connectives
Cut-free derivation in the deep nested calculus
a⇒ a, (· ⇒ ·)· ⇒ (b ⇒ b)
b ⇒ (· ⇒ b)
a⊕ b ⇒ a, (· ⇒ b)
· ⇒ (c ⇒ c)
c ⇒ (· ⇒ c)
(a⊕ b)⊕ c ⇒ a, (· ⇒ b, c)
(a⊕ b)⊕ c ⇒ a, (· ⇒ b ⊕ c) · ⇒ (d ⇒ d)
(a⊕ b)⊕ c ⇒ a, (b ⊕ c ( d ⇒ d) · ⇒ (e ⇒ e)
(a⊕ b)⊕ c ⇒ a, ((b ⊕ c ( d)⊕ e ⇒ d , e)
(a⊕ b)⊕ c ⇒ a, ((b ⊕ c ( d)⊕ e ⇒ d ⊕ e)
(a⊕ b)⊕ c ⇒ a, (b ⊕ c ( d)⊕ e ( d ⊕ e
No annotations, only commas as structural connective, butsequents are nested (· · · ⇒ · · · ) · · · ⇒ · · · (· · · ⇒ · · · )
Display calculus for (an extension of) FILL
Structural Constant and Binary Connectives: Φ , < >
Antecedent Structure: Xa Ya ::= A | Φ | Xa,Ya | Xa < Ys
Succcedent Structure: Xs Ys ::= A | Φ | Xs ,Ys | Xa > Ys
Sequent: Xa ` Ys (drop subscripts to avoid clutter)
Display Postulates: reversible structural rules
Xa ` Ya > Zs
Xa,Ya ` Zs
Ya ` Xa > Zs
Za < Ys ` Xs
Za ` Xs ,Ys
Za < Xs ` Ys
Display Property: For every antecedent (succedent) part Z of thesequent X ` Y , there is a sequent Z ` Y ′ (resp. X ′ ` Z )obtainable from X ` Y using only the display postulates,thereby displaying the Z as the whole of one side
Logical rules: introduced formula is always displayed
(id) p ` pX ` A A ` Y(cut)
X ` Y
Φ ` X(1 `)1 ` X
(` 1) Φ ` 1
(0 `) 0 ` ΦX ` Φ(` 0)X ` 0
A,B ` X(⊗ `)
A⊗ B ` XX ` A Y ` B(` ⊗)X ,Y ` A⊗ B
A ` X B ` Y(⊕ `)A⊕ B ` X ,Y
X ` A,B(` ⊕)
X ` A⊕ B
X ` A B ` Y((`)A ( B ` X > Y
X ` A > B(`()
X ` A ( B
A < B ` X(−< `)
A−<B ` XX ` A B ` Y(` −<)X < Y ` A−<B
read upwards, one rule is a “rewrite” while other “constrains”
Structural rules: no occurrences of formula meta-variables
all sub-structural properties captured in a modular way
X ,Φ ` Y(Φ `)
X ` Y
X ` Φ,Y(` Φ)
X ` Y
W , (X ,Y ) ` Z(Ass `)
(W ,X ),Y ` Z
W ` (X ,Y ),Z(` Ass)
W ` X , (Y ,Z )
X ,Y ` Z(Com `)
Y ,X ` Z
Z ` Y ,X(` Com)
Z ` X ,Y
W , (X < Y ) ` Z(Grnb `)
(W ,X ) < Y ` Z
W ` (X > Y ),Z(` Grnb)
W ` X > (Y ,Z )
((A ( B)⊕ C ) ( (A ( (B ⊕ C ))
Example derivation in our display calculus
a ` a b ` b(⊕ `)
a ⊕ b ` a, b c ` c(⊕ `)
(a⊕ b) ⊕ c ` (a, b), c(ass)
(a⊕ b)⊕ c ` a, (b, c)(drp)
(a⊕ b)⊕ c < a ` b, c(` ⊕)
(a⊕ b)⊕ c < a ` b ⊕ c d ` d((`)
b ⊕ c ( d ` ((a⊕ b)⊕ c < a) > d e ` e(⊕ `)
(b ⊕ c ( d) ⊕ e ` (((a⊕ b)⊕ c < a) > d), e(` Grnb)
(b ⊕ c ( d)⊕ e ` ((a⊕ b)⊕ c < a) > (d , e)(rp)
(b ⊕ c ( d)⊕ e,((a⊕ b)⊕ c < a) ` d , e(` ⊕)
(b ⊕ c ( d)⊕ e, ((a⊕ b)⊕ c < a) ` d ⊕ e(rp)
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e > (d ⊕ e)(`()
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e ( (d ⊕ e)(drp)
(a⊕ b)⊕ c ` a,(b ⊕ c ( d)⊕ e ( (d ⊕ e)
But we implicitly created an occurrence of −< via <
Categorial semantics for bi-intuitionistic linear logic BiILL
(⊗, 1,() is a symmetric monoidal closed structure
A⊗ B ( C iff A ( (B ( C ) iff B ( (A ( C )
(A⊗ 1) ( A and A ( (A⊗ 1)
(−<,⊕, 0) is a symmetric monoidal co-closed structure
A ( (B ⊕ C ) iff (A−<B) ( C iff (A−<C ) ( B
(A⊕ 0) ( A and A ( (A⊕ 0)
interaction via either of
Grishin(b) ((A ( B)⊕ C ) ( (A ( (B ⊕ C ))
dualGrishin(b) ((A⊗ B)−<C ) ( (A⊗ (B−<C ))
Collapse to (classical) MLL: if we add converse of either
Soundness, completeness and cut-elimination
Thm: The sequent X ` Y is derivable iff the formula-translationτa(X ) ( τs(Y ) is BiILL-valid
Proof: the display calculus proof rules and the arrows of the freeBiILL-category are inter-definable.
Thm: If X ` Y is derivable then it is cut-free derivable.
Proof: The rules obey conditions C1-C8 given by Belnap (1982),hence the calculus enjoys cut-admissibility
Method: Essentially Gentzen’s original method
Left-principal cuts: turned into left and right principal cutsRight-principal cuts: turned into left and right principal cutsLeft&right principal cuts: replaced by multiple “smaller” cutsNeed: substitution principles guaranteed by display property
and explicit structural rulesTermination: double induction on grade and size of cut-formula
Proof search via display calculi
Given end-sequent X ` Y how to find a derivation?
Apply (invertible) “rewrite” rules immediately
For “constraint” rules, use structural rules to match structure
Backward Proof-search for display calculus
a ` a b ` ba ⊕ b ` a, b c ` c
(a⊕ b) ⊕ c ` (a, b), c
(a⊕ b)⊕ c ` a, (b, c)(?)
(a⊕ b)⊕ c < a ` b, c(` ⊕)
(a⊕ b)⊕ c < a ` b ⊕ c d ` d((`)
b ⊕ c ( d ` ((a⊕ b)⊕ c < a) > d e ` e(⊕ `)
(b ⊕ c ( d) ⊕ e ` (((a⊕ b)⊕ c < a) > d), e(?)
(b ⊕ c ( d)⊕ e ` ((a⊕ b)⊕ c < a) > (d , e)(?)
(b ⊕ c ( d)⊕ e,((a⊕ b)⊕ c < a) ` d , e(` ⊕)
(b ⊕ c ( d)⊕ e, ((a⊕ b)⊕ c < a) ` d ⊕ e(?)
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e > (d ⊕ e)(`()
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e ( (d ⊕ e)(?)
(a⊕ b)⊕ c ` a,(b ⊕ c ( d)⊕ e ( (d ⊕ e)
Proof search via display calculi
Given end-sequent X ` Y how to find a derivation?
Apply (invertible) “rewrite” rules immediately
For “constraint” rules, use structural rules to match structure
But there are just too many possibilities for this to be efficient
Need to compile “display”, “undisplay” and “structural rules” intothe logical rules as much as possible (maintaining cut-elimination)
Cut-free derivation in the deep nested calculus
a⇒ a, (· ⇒ ·)· ⇒ (b ⇒ b)
b ⇒ (· ⇒ b)
a⊕ b ⇒ a, (· ⇒ b)
· ⇒ (c ⇒ c)
c ⇒ (· ⇒ c)
(a⊕ b)⊕ c ⇒ a, (· ⇒ b, c)
(a⊕ b)⊕ c ⇒ a, (· ⇒ b ⊕ c) · ⇒ (d ⇒ d)
(a⊕ b)⊕ c ⇒ a, (b ⊕ c ( d ⇒ d) · ⇒ (e ⇒ e)
(a⊕ b)⊕ c ⇒ a, ((b ⊕ c ( d)⊕ e ⇒ d , e)
(a⊕ b)⊕ c ⇒ a, ((b ⊕ c ( d)⊕ e ⇒ d ⊕ e)
(a⊕ b)⊕ c ⇒ a, (b ⊕ c ( d)⊕ e ( d ⊕ e
No annotations, only commas as structural connective, butsequents are nested
Nested sequent calculi: Kashima 1994 (for tense logics)
Nested sequent: where comma is associative and commutative
S T ::= S1, . . . ,Sk ,A1, . . . ,Am ⇒ B1, . . . ,Bn,T1, . . . ,Tl
Context: nested sequent with exactly one hole [ ]
Positive/Negative: if [ ] occurs immediately to right/left of ⇒
Turn Rules: reversible rules using multisets of nested sequentsand formulae
S2 ⇒ (S1 ⇒ T )
S1,S2 ⇒ T(S ⇒ T2)⇒ T1S ⇒ (T1, T2)
Xa ` Ya > Zs
Xa,Ya ` Zs
Ya ` Xa > Zs
Za < Ys ` Xs
Za ` Xs ,Ys
Za < Xs ` Ys
Nested sequent calculi: Kashima 1994 (for tense logics)
Nested sequent: where comma is associative and commutative
S T ::= S1, . . . ,Sk ,A1, . . . ,Am ⇒ B1, . . . ,Bn,T1, . . . ,Tl
Context: nested sequent with exactly one hole [ ]
Positive/Negative: if [ ] occurs immediately to right/left of ⇒
Turn Rules: reversible rules using multisets of nested sequentsand formulae
S2 ⇒ (S1 ⇒ T )
S1,S2 ⇒ T(S ⇒ T2)⇒ T1S ⇒ (T1, T2)
Xa ` Ya > Zs
Xa,Ya ` Zs
Ya ` Xa > Zs
Za < Ys ` Xs
Za ` Xs ,Ys
Za < Xs ` Ys
Nested sequent calculi: Kashima 1994 (for tense logics)
Nested sequent: where comma is associative and commutative
S T ::= S1, . . . ,Sk ,A1, . . . ,Am ⇒ B1, . . . ,Bn,T1, . . . ,Tl
Context: nested sequent with exactly one hole [ ]
Positive/Negative: if [ ] occurs immediately to right/left of ⇒Turn Rules: reversible rules using multisets of nested sequents
and formulae
S2 ⇒ (S1 ⇒ T )
S1,S2 ⇒ T(S ⇒ T2)⇒ T1S ⇒ (T1, T2)
Display Property: For every positive (negative) context X [ ] andevery S, there exists T such that T ⇒ S (respectively S ⇒ T )is derivable from X [S] using only the structural turn rulesabove. Thus S is “displayed” in T ⇒ S (resp. S ⇒ T ).
Polarity: note that ⇒ always points to the positive contexts
Shallow nested sequent calculus for BiILL
Logical rules:
p ⇒ p idS ⇒ S ′,A A, T ⇒ T ′
S, T ⇒ S ′, T ′ cut
0⇒ · 0lS ⇒ TS ⇒ T , 0 0r
S ⇒ TS, 1⇒ T 1l · ⇒ 1
1r
S,A,B ⇒ TS,A⊗ B ⇒ T ⊗l
S ⇒ A, T S ′ ⇒ B, T ′
S,S ′ ⇒ A⊗ B, T , T ′ ⊗r
S,A⇒ T S ′,B ⇒ T ′
S,S ′,A⊕ B ⇒ T , T ′ ⊕lS ⇒ A,B, TS ⇒ A⊕ B, T ⊕r
S ⇒ A, T S ′,B ⇒ T ′
S,S ′,A ( B ⇒ T , T ′ (lS ⇒ T , (A⇒ B)
S ⇒ T ,A ( B(r
S, (A⇒ B)⇒ TS,A−<B ⇒ T −<l
S ⇒ A, T S ′,B ⇒ T ′
S,S ′ ⇒ A−<B, T , T ′ −<r
Shallow nested sequent calculus for BiILL
Structural Rules: Grishin (b) analogues
T , (S ⇒ S ′)⇒ T ′
(S, T ⇒ S ′)⇒ T ′gl
S ⇒ (S ′ ⇒ T ′), TS ⇒ (S ′ ⇒ T ′, T )
gr
W , (X < Y ) ` Z(Grnb `)
(W ,X ) < Y ` Z
W ` (X > Y ),Z(` Grnb)
W ` X > (Y ,Z )
Thm: Every formula has a cut-free nested shallow sequentderivation iff it has cut-free display calculus derivation
Thm: if a nested sequent is derivable it is cut-free derivable
Proof: essentially the same as for display calculus
Prof search: how to absorb the turn and gl and gr rules ?
Deep nested sequents: just apply the rules inside contextsX [ ] and U and V are hollow.
X [U , p ⇒ p,V]idd similarly for units (no cut rule)
X [S,A,B ⇒ T ]
X [S,A⊗ B ⇒ T ]⊗d
l
X1[S1 ⇒ A, T1] X2[S2 ⇒ B, T2]
X [S ⇒ A⊗ B, T ]⊗d
r
X1[S1 ⇒ A, T1] X2[S2,B ⇒ T2]
X [S,A ( B ⇒ T ](d
l
X [S ⇒ T , (A⇒ B)]
X [S ⇒ T ,A ( B](d
r
X1[S1,A⇒ T1] X2[S2,B ⇒ T2]
X [S,A⊕ B ⇒ T ]⊕d
l
X [S ⇒ A,B, T ]
X [S ⇒ A⊕ B, T ]⊕d
r
X [S, (A⇒ B)⇒ T ]
X [S,A−<B ⇒ T ]−<d
l
X1[S1 ⇒ A, T1] X2[S2,B ⇒ T2]
X [S ⇒ A−<B, T ]−<d
r
Hollow: X [] contains no formulae (⇒-tree of empty nodes)
Merge: X [ ] ∈ X1[ ] • X2[ ] and S ∈ S1 • S2 and T ∈ T1 • T2
Deep nested sequents: just apply the rules inside contexts
Propagation rules: allow formulae to be moved in a context
X [S ⇒ (S ′,A⇒ T ′), T ]
X [S,A⇒ (S ′ ⇒ T ′), T ]pl1
X [S ′, (S ⇒ A, T )⇒ T ′]
X [S ′, (S ⇒ T )⇒ A, T ′]pr1
X [S, (S ′ ⇒ T ′),A⇒ T ]
X [S, (S ′,A⇒ T ′)⇒ T ]pl2
X [S ⇒ A, (S ′ ⇒ T ′), T ]
X [S ⇒ (S ′ ⇒ A, T ′), T ]pr2
Note: these are not just reversible rules expanded into two!
Note: the rules do not need to track polarity
Thm: the turn rules and rules gl and gr are (cut-free) admissible
Thm: if a nested sequent is (cut-free) derivable in the deepcalculus then it is cut-free derivable in the shallow calculus
Thm: if a nested sequent is cut-free derivable in the shallowcalculus then it is (cut-free) derivable in the deep calculus
Cor: the deep and shallow nested calculi derive the same sequents
Deep nested sequents: just apply the rules inside contexts
Propagation rules: allow formulae to be moved in a context
X [S ⇒ (S ′,A⇒ T ′), T ]
X [S,A⇒ (S ′ ⇒ T ′), T ]pl1
X [S ′, (S ⇒ A, T )⇒ T ′]
X [S ′, (S ⇒ T )⇒ A, T ′]pr1
X [S, (S ′ ⇒ T ′),A⇒ T ]
X [S, (S ′,A⇒ T ′)⇒ T ]pl2
X [S ⇒ A, (S ′ ⇒ T ′), T ]
X [S ⇒ (S ′ ⇒ A, T ′), T ]pr2
Note: these are not just reversible rules expanded into two!
Note: the rules do not need to track polarity
Thm: the turn rules and rules gl and gr are (cut-free) admissible
Thm: if a nested sequent is (cut-free) derivable in the deepcalculus then it is cut-free derivable in the shallow calculus
Thm: if a nested sequent is cut-free derivable in the shallowcalculus then it is (cut-free) derivable in the deep calculus
Cor: the deep and shallow nested calculi derive the same sequents
From BiILL back to FILL
a ` a b ` ba ⊕ b ` a, b c ` c
(a⊕ b) ⊕ c ` (a, b), c
(a⊕ b)⊕ c ` a, (b, c)(?)
(a⊕ b)⊕ c < a ` b, c(` ⊕)
(a⊕ b)⊕ c < a ` b ⊕ c d ` d((`)
b ⊕ c ( d ` ((a⊕ b)⊕ c < a) > d e ` e(⊕ `)
(b ⊕ c ( d) ⊕ e ` (((a⊕ b)⊕ c < a) > d), e(?)
(b ⊕ c ( d)⊕ e ` ((a⊕ b)⊕ c < a) > (d , e)(?)
(b ⊕ c ( d)⊕ e,((a⊕ b)⊕ c < a) ` d , e(` ⊕)
(b ⊕ c ( d)⊕ e, ((a⊕ b)⊕ c < a) ` d ⊕ e(?)
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e > (d ⊕ e)(`()
(a⊕ b)⊕ c < a ` (b ⊕ c ( d)⊕ e ( (d ⊕ e)(?)
(a⊕ b)⊕ c ` a,(b ⊕ c ( d)⊕ e ( (d ⊕ e)
From BiILL back to FILL
Problem: all our calculi are for BiILL ... are they sound for FILL?
Display calculus: must create antecedent < structures in itsderivation of FILL-formulae in order to display and undisplay
Question: is BiILL a conservative extension of FILL?
we were not able to find a categorial proof for BiILL
Compare: to tense logic Kt say where there is a simple semanticproof that Kt is a conservative extension of K (same frames)
From BiILL back to FILL
Problem: all our calculi are for BiILL ... are they sound for FILL?
Display calculus: must create antecedent < structures in itsderivation of FILL-formulae in order to display and undisplay
Question: is BiILL a conservative extension of FILL?
we were not able to find a categorial proof for BiILL
Compare: to tense logic Kt say where there is a simple semanticproof that Kt is a conservative extension of K (same frames)
From BiILL back to FILL
Nested FILL-sequent: nested sequent that has no nesting ofsequents on the left of ⇒ and no occurrences of −<
Why? because then it is free of any traces of −< since only the−<d
l rule creates such nestings
FILLdn: remove −<dl , −<d
r , pl2 and pr1
Separation Thm: nested FILL-sequents are (cut-free) derivable inFILLdn iff they are derivable in the original deep calculus
Proof: if the end-sequent contains only FILL-formulae thenbackward proof search will never create the offending nesting
Thm: every rule of FILLdn preserves FILL-validity downwards
Cor: FILLdn is sound and cut-free complete for FILL-validity
Cor: BiILL is a conservative extension of FILL
From BiILL back to FILL
Nested FILL-sequent: nested sequent that has no nesting ofsequents on the left of ⇒ and no occurrences of −<
Why? because then it is free of any traces of −< since only the−<d
l rule creates such nestings
FILLdn: remove −<dl , −<d
r , pl2 and pr1
Separation Thm: nested FILL-sequents are (cut-free) derivable inFILLdn iff they are derivable in the original deep calculus
Proof: if the end-sequent contains only FILL-formulae thenbackward proof search will never create the offending nesting
Thm: every rule of FILLdn preserves FILL-validity downwards
Cor: FILLdn is sound and cut-free complete for FILL-validity
Cor: BiILL is a conservative extension of FILL
Complexity of deciding FILL
Backward proof search: need to try all ways to partition a contextand apply all propagation rules in all ways, but this terminates
Subformula property: every formula in the premises is asubformula of a formula in the conclusion
Complexity: FILL is in NP
Proof: guess a cut-free derivation and check it in time polynomialin size of the given formula. Possible since each formula isintroduced exactly once in any derivation
Complexity: our calculus handles the problem of decidingconstants-only multiplicative linear logic (COMLL) since thennf-formulae of this logic contain no propositional atoms andno implication connectives and it is known to be NP-hard
Thm: deciding FILL-validity is NP-complete
Our Methodology
Start with a categorially formulated logic L
Give a display calculus for L
X Easily shown to be complete.X Easily shown to support a cut-elimination procedure.X Embeds L into an extension so conservativity, and hence
soundness, not clear.X Not well suited for proof search.
Refine to a nested sequent calculus with deep inference.
X Conservativity, and hence soundness, now clear.X Supports proof search and complexity analysis.
Question: how far can we push this methodology?
We know: it works for the exponentials ! and ?
But: BiILL plus additives not conservative over FILL plusadditivies e.g > ` p,> is BiILL-valid but not FILL-valid
Our Methodology
Start with a categorially formulated logic L
Give a display calculus for L
X Easily shown to be complete.X Easily shown to support a cut-elimination procedure.X Embeds L into an extension so conservativity, and hence
soundness, not clear.X Not well suited for proof search.
Refine to a nested sequent calculus with deep inference.
X Conservativity, and hence soundness, now clear.X Supports proof search and complexity analysis.
Question: how far can we push this methodology?
We know: it works for the exponentials ! and ?
But: BiILL plus additives not conservative over FILL plusadditivies e.g > ` p,> is BiILL-valid but not FILL-valid
Happy (belated) birthday Gerhard!
Belnap’s Eight Conditions a la Kracht
(C1) Each formula variable occurring in some premise of a rule ρ is asubformula of some formula in the conclusion of ρ.
(C2) Congruent parameters is a relation between parameters of theidentical structure variable occurring in the premise and conclusion
(C3) Each parameter is congruent to at most one structure variable inthe conclusion. Equivalently, no two structure variables in theconclusion are congruent to each other.
(C4) Congruent parameters are either all antecedent or all succedentparts of their respective sequent.
(C5) A formula in the conclusion of a rule ρ is either the entireantecedent or the entire succedent. Such a formula is called aprincipal formula of ρ.
(C6/7) Each rule is closed under simultaneous substitution of arbitrarystructures for congruent parameters.
Belnap’s Eight Conditions a la Kracht
(C8) If there are rules ρ and σ with respective conclusions X ` A andA ` Y with formula A principal in both inferences (in the sense of C5)and if cut is applied to yield X ` Y , then either X ` Y is identical toeither X ` A or A ` Y ; or it is possible to pass from the premises of ρand σ to X ` Y by means of inferences falling under cut where thecut-formula always is a proper subformula of A.
X ` C > DX ` C ( D
U ` C D ` ZC ( D ` U > Z
cutX ` U > Z
U ` C
X ` C > DX ,C ` D D ` Z
cutX ,C ` Z
C ` X > Zcut
U ` X > ZX ,U ` Z
X ` U > Z
Devil’s Advocate Questions
Can’t we do all of this with labelled sequent calculi?
Γ,R(x , y , z), x : A ` z : B,∆x and y do not occur in the conclusion
Γ ` y : (A ( B),∆
Almost certainly we can, but the decidability cannot be read off aswe are able to do, and I can’t see the “Grishin” condition
Where are the Grishin rules in the deep calculus?In the logical rules and propagation rules
X1[S1,A⇒ T1] X2[S2,B ⇒ T2]
X [S,A⊕ B ⇒ T ]⊕d
l
X [S ⇒ A, (S ′ ⇒ T ′), T ]
X [S ⇒ (S ′ ⇒ A, T ′), T ]pr2
Isn’t “merge” just another way of hiding annotations? No, it is avery natural definition ... superpose two identical ⇒-trees