On Proof Compression Mechanisms and Deep Inference
Anupam Das
University of Bath
LAC-GeoCal, 24/11/2011
Outline of Talk
1 Motivation
Compression Mechanisms in Proof Systems Proof Complexity
Deep Inference Questions
2 The Setting
Calculus of Structures
Embedding Compression Mechanisms
3 Some Theory Algebra
Measuring Proofs
Motivation
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 3 / 20
Compression Mechanisms
Proof systems often containcompression mechanisms that allow proofs to have smaller representations
, e.g.
I Dag - Reusing an already proved formula/subproof
I Cut - Invoking a lemma
I Substitution - Substituting variables in a proof
I Extension - Naming a theorem or proof Observation
Given a formula A∧B, compression mechanisms candetect similarities between the conjunctsA andB.
E.g. from a proof of A,
I a dag-like system can conclude A∧A
I a system with substitution can concludeA∧Aσ, whereσ is some substitution of variables.
PROBLEM: How can we capture this property of ‘detecting similarities’ formally?
Compression Mechanisms
Proof systems often containcompression mechanisms that allow proofs to have smaller representations, e.g.
I Dag - Reusing an already proved formula/subproof
I Cut - Invoking a lemma
I Substitution - Substituting variables in a proof
I Extension - Naming a theorem or proof
Observation
Given a formula A∧B, compression mechanisms candetect similarities between the conjunctsA andB.
E.g. from a proof of A,
I a dag-like system can conclude A∧A
I a system with substitution can concludeA∧Aσ, whereσ is some substitution of variables.
PROBLEM: How can we capture this property of ‘detecting similarities’ formally?
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 4 / 20
Compression Mechanisms
Proof systems often containcompression mechanisms that allow proofs to have smaller representations, e.g.
I Dag - Reusing an already proved formula/subproof
I Cut - Invoking a lemma
I Substitution - Substituting variables in a proof
I Extension - Naming a theorem or proof Observation
Given a formula A∧B, compression mechanisms candetect similarities between the conjunctsA andB.
E.g. from a proof of A,
I a dag-like system can conclude A∧A
I a system with substitution can concludeA∧Aσ, whereσ is some substitution of variables.
PROBLEM: How can we capture this property of ‘detecting similarities’ formally?
Compression Mechanisms
Proof systems often containcompression mechanisms that allow proofs to have smaller representations, e.g.
I Dag - Reusing an already proved formula/subproof
I Cut - Invoking a lemma
I Substitution - Substituting variables in a proof
I Extension - Naming a theorem or proof Observation
Given a formula A∧B, compression mechanisms candetect similarities between the conjunctsA andB.
E.g. from a proof of A,
I a dag-like system can conclude A∧A
I a system with substitution can concludeA∧Aσ, whereσ is some substitution of variables.
PROBLEM: How can we capture this property of ‘detecting similarities’ formally?
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 4 / 20
Compression Mechanisms
Proof systems often containcompression mechanisms that allow proofs to have smaller representations, e.g.
I Dag - Reusing an already proved formula/subproof
I Cut - Invoking a lemma
I Substitution - Substituting variables in a proof
I Extension - Naming a theorem or proof Observation
Given a formula A∧B, compression mechanisms candetect similarities between the conjunctsA andB.
E.g. from a proof of A,
I a dag-like system can conclude A∧A
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation. Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart? E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like cut-free Gentzen. [Statman]
I More generally the effects of general proof compression mechanisms on proof complexity are studied.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 5 / 20
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation. Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart? E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like cut-free Gentzen. [Statman]
I More generally the effects of general proof compression mechanisms on proof complexity are studied.
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation.
Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart? E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like cut-free Gentzen. [Statman]
I More generally the effects of general proof compression mechanisms on proof complexity are studied.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 5 / 20
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation.
Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart?
E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like cut-free Gentzen. [Statman]
I More generally the effects of general proof compression mechanisms on proof complexity are studied.
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation.
Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart? E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like cut-free Gentzen. [Statman]
I More generally the effects of general proof compression mechanisms on proof complexity are studied.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 5 / 20
Proof Complexity
I Types of question: ‘How short are proofs of formulae in a system, up to a polynomial factor?’
I An important relation: polynomial simulation.
Definition
A systemS polynomially simulates a systemT if for every proof inS there is a proof in T of the same formula whose size is bounded above by a polynomial in the size of the original proof.
I A class of studied questions: When does a tree-like system polynomially simulate its dag-like counterpart? E.g.
I Tree-like Frege polynomially simulates dag-like Frege. [Krajicek]
I Tree-like cut-free Gentzen does not polynomially simulate dag-like
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is called KS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
c↑ −−−−−−A A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag. Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 6 / 20
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is called KS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
c↑ −−−−−−A A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag. Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is called KS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
c↑ −−−−−−A A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag. Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 6 / 20
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is calledKS.
Deep inference systems have their own versions of cut and dag: A∧A¯
i↑ −−−−−−
⊥
c↑ −−−−−−A A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag. Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is calledKS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
A
c↑ −−−−−−
A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag. Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 6 / 20
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is calledKS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
A
c↑ −−−−−−
A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag.
Question
Does KS polynomially simulate KS∪ {c↑}? Conjecture: No. [Straßburger]
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is calledKS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
A
c↑ −−−−−−
A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag.
Question
Does KS polynomially simulate KS∪ {c↑}?
Conjecture: No. [Straßburger]
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 6 / 20
Deep Inference
1 Inference rules operatearbitrarily deep within formulae. A∨(B∧C)
−−−−−−−−−−−−−−
A∨(B∧C0)
2 Derivations are top-down symmetric. A`B ⇒B¯ `A.¯
The basic deep inference system, without compression mechanisms, is calledKS. Deep inference systems have their own versions of cut and dag:
A∧A¯
i↑ −−−−−−
⊥
A
c↑ −−−−−−
A∧A
but these can be applied anywhere inside a formula. They can be considered generalisations of the usual formulations of cut and dag.
Questions
1 How can we capture compression mechanisms in a way that is independent of any particular proof system?
2 Can we give a definition of ‘compression mechanism’, or what it means to be an uncompressed system.
3 Can we give conditions for when a tree-like system is able to polynomially simulate its dag-like counterpart?
4 More generally, can we give conditions for when any system is able to polynomially simulate that system augmented with some compression mechanism?
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 7 / 20
Questions
1 How can we capture compression mechanisms in a way that is independent of any particular proof system?
2 Can we give a definition of ‘compression mechanism’, or what it means to be an uncompressed system.
3 Can we give conditions for when a tree-like system is able to polynomially simulate its dag-like counterpart?
4 More generally, can we give conditions for when any system is able to polynomially simulate that system augmented with some compression mechanism?
Questions
1 How can we capture compression mechanisms in a way that is independent of any particular proof system?
2 Can we give a definition of ‘compression mechanism’, or what it means to be an uncompressed system.
3 Can we give conditions for when a tree-like system is able to polynomially simulate its dag-like counterpart?
4 More generally, can we give conditions for when any system is able to polynomially simulate that system augmented with some compression mechanism?
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 7 / 20
Questions
1 How can we capture compression mechanisms in a way that is independent of any particular proof system?
2 Can we give a definition of ‘compression mechanism’, or what it means to be an uncompressed system.
3 Can we give conditions for when a tree-like system is able to polynomially simulate its dag-like counterpart?
4 More generally, can we give conditions for when any system is able to polynomially simulate that system augmented with some compression mechanism?
The Setting
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 8 / 20
Calculus of Structures
I We will consider those proof systems that can be embedded in the deep inference formalism Calculus of Structures.
I CoS proofs are just sequences of formulae chained together by inference rules from a proof system.
I There is no meta/object level distinction.
I We can embed a wide variety of studied proof systems in this formalism, while preserving proof complexity properties. E.g.
[A∨C]¯ ∧[C∨B]
mp−−−−−−−−−−−−−−−−−−−
A∨B
A∧B
mix−−−−−−
A∨B
A
shr−−−−−−
A∧A modus ponens mix dag
>
id−−−−−−
A∨A¯
A
wk−−−−−−
A∨B
A∨A
con−−−−−−
A
[A∨B]∧[C∨D]
∧ −−−−−−−−−−−−−−−−−−−
A∨(B∧C)∨D identity weakening contraction
Calculus of Structures
I We will consider those proof systems that can be embedded in the deep inference formalism Calculus of Structures.
I CoS proofs are just sequences of formulae chained together by inference rules from a proof system.
I There is no meta/object level distinction.
I We can embed a wide variety of studied proof systems in this formalism, while preserving proof complexity properties. E.g.
[A∨C]¯ ∧[C∨B]
mp−−−−−−−−−−−−−−−−−−−
A∨B
A∧B
mix−−−−−−
A∨B
A
shr−−−−−−
A∧A modus ponens mix dag
>
id−−−−−−
A∨A¯
A
wk−−−−−−
A∨B
A∨A
con−−−−−−
A
[A∨B]∧[C∨D]
∧ −−−−−−−−−−−−−−−−−−−
A∨(B∧C)∨D identity weakening contraction
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 9 / 20
Calculus of Structures
I We will consider those proof systems that can be embedded in the deep inference formalism Calculus of Structures.
I CoS proofs are just sequences of formulae chained together by inference rules from a proof system.
I There is no meta/object level distinction.
I We can embed a wide variety of studied proof systems in this formalism, while preserving proof complexity properties.
E.g. [A∨C]¯ ∧[C∨B]
mp−−−−−−−−−−−−−−−−−−−
A∨B
A∧B
mix−−−−−−
A∨B
A
shr−−−−−−
A∧A modus ponens mix dag
>
id−−−−−−
A∨A¯
A
wk−−−−−−
A∨B
A∨A
con−−−−−−
A
[A∨B]∧[C∨D]
∧ −−−−−−−−−−−−−−−−−−−
A∨(B∧C)∨D identity weakening contraction
Calculus of Structures
I We will consider those proof systems that can be embedded in the deep inference formalism Calculus of Structures.
I CoS proofs are just sequences of formulae chained together by inference rules from a proof system.
I There is no meta/object level distinction.
I We can embed a wide variety of studied proof systems in this formalism, while preserving proof complexity properties. E.g.
[A∨C]¯ ∧[C∨B]
mp−−−−−−−−−−−−−−−−−−−
A∨B
A∧B
mix−−−−−−
A∨B
A
shr−−−−−−
A∧A modus ponens mix dag
>
id−−−−−−
A∨A¯
A
wk−−−−−−
A∨B
A∨A
con−−−−−−
A
[A∨B]∧[C∨D]
∧ −−−−−−−−−−−−−−−−−−−
A∨(B∧C)∨D identity weakening contraction
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 9 / 20
Embedding Compression Mechanisms in CoS
We can define compression mechanisms as follows: A∨(B∧B)¯
cut−−−−−−−−−−−−−
A
A
dag−−−−−−
A∧A
Definition
We say that a systemS “distributes over∧” if, for?∈ {∧,∨}, there is a polynomial-size derivation (A ? B)∧(A ? C)`A ?(B∧C).
Theorem
Whenever a system distributes over ∧and has basic equations, cut is equivalent tomp anddag is equivalent toshr. If the system is a deep inference system then cutis equivalent to i↑ anddag is equivalent to c↑.
Embedding Compression Mechanisms in CoS
We can define compression mechanisms as follows:
A∨(B∧B)¯
cut−−−−−−−−−−−−−
A
A
dag−−−−−−
A∧A
Definition
We say that a systemS “distributes over∧” if, for?∈ {∧,∨}, there is a polynomial-size derivation (A ? B)∧(A ? C)`A ?(B∧C).
Theorem
Whenever a system distributes over ∧and has basic equations, cut is equivalent tomp anddag is equivalent toshr. If the system is a deep inference system then cutis equivalent to i↑ anddag is equivalent to c↑.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 10 / 20
Embedding Compression Mechanisms in CoS
We can define compression mechanisms as follows:
A∨(B∧B)¯
cut−−−−−−−−−−−−−
A
A
dag−−−−−−
A∧A
Definition
We say that a systemS “distributes over∧” if, for?∈ {∧,∨}, there is a polynomial-size derivation (A ? B)∧(A ? C)`A ?(B∧C).
Theorem
Whenever a system distributes over ∧and has basic equations, cut is
Some Theory
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 11 / 20
Rings and Modules
I Asemiring Ris a tuple(R,+,×,0,1)such that:
1 (R,+,0)is a commutative monoid.
2 (R,×,1)is a commmutative monoid.
3 ×distributes over +.
I Amodule V over a semiringR= (R,+,×,0,1)is a tuple (V,+V,0V,•) such that:
1 (V,+V,0V)is a commutative monoid.
2 •:R×V →V. •left-distributes over+V and right-distributes over+.
3 ×is associative over•. 0•v= 0V and1•v=v.
Observation
(TAUT,∧,>,·) forms a module over N, wheren·A≡
n
z }| {
A∧· · ·∧A.
Rings and Modules
I Asemiring Ris a tuple(R,+,×,0,1)such that:
1 (R,+,0)is a commutative monoid.
2 (R,×,1)is a commmutative monoid.
3 ×distributes over +.
I Amodule V over a semiring R= (R,+,×,0,1)is a tuple (V,+V,0V,•) such that:
1 (V,+V,0V)is a commutative monoid.
2 •:R×V →V. •left-distributes over+V and right-distributes over+.
3 ×is associative over•. 0•v= 0V and1•v=v.
Observation
(TAUT,∧,>,·) forms a module over N, wheren·A≡
n
z }| {
A∧· · ·∧A.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 12 / 20
Rings and Modules
I Asemiring Ris a tuple(R,+,×,0,1)such that:
1 (R,+,0)is a commutative monoid.
2 (R,×,1)is a commmutative monoid.
3 ×distributes over +.
I Amodule V over a semiring R= (R,+,×,0,1)is a tuple (V,+V,0V,•) such that:
1 (V,+V,0V)is a commutative monoid.
2 •:R×V →V. •left-distributes over+V and right-distributes over+.
3 ×is associative over•. 0•v= 0V and1•v=v.
Observation
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
If, in addition we have µ(v+w) =µ(v) +µ(w) then we say thatµ is an L1 norm.
We will be interested in measures that form norms over (TAUT,∧,>,·).
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 13 / 20
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
If, in addition we have µ(v+w) =µ(v) +µ(w) then we say thatµ is an L1 norm.
We will be interested in measures that form norms over (TAUT,∧,>,·).
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
If, in addition we have µ(v+w) =µ(v) +µ(w) then we say thatµ is an L1 norm.
We will be interested in measures that form norms over (TAUT,∧,>,·).
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 13 / 20
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
If, in addition we have µ(v+w) =µ(v) +µ(w) then we say thatµ is an L1 norm.
We will be interested in measures that form norms over (TAUT,∧,>,·).
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
If, in addition we have µ(v+w) =µ(v) +µ(w) then we say thatµ is an L1 norm.
We will be interested in measures that form norms over (TAUT,∧,>,·).
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 13 / 20
Measures and Norms
For a proof system S:
I Ameasuring is a functionM :PRFS→R such thatM ∼ | · | andM is monotone with respect to subproofs.
I Ameasure is a compositionM ◦P whereM is a measuring and P is an assignment of tautologies to proofs that are minimal, up to a polynomial, and closed under subproofs.
I Anorm on a moduleV over a semiring R⊆Ris function µ:V →R such that:
1 µ(a•v) =|a|µ(v). (Homogeneity)
2 µ(v+w)≤µ(v) +µ(w). (Triangle inequality, orM)
3 µ(v) = 0 impliesv= 0V.
Main Result
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 14 / 20
The Characterisation
Theorem
Let S be a system with an L1 norm. The following are equivalent:
1 dag-S has a norm.
2 dag-S has an L1 norm.
3 dag-S is polynomially equivalent toS.
Proof Idea.
I (2) implies (1): Trivial.
I (1) implies (3): Ifµ=M◦P is a norm for S then it is also a measure for dag-S, and so a norm for dag-S.
I (3) implies (2): Transform proofs by just expanding out dag-steps. Evaluating the size of the resulting proof in a clever way gives only a polynomial blow-up; this relies crucially on homogeneity of the norm.
The Characterisation
Theorem
Let S be a system with an L1 norm. The following are equivalent:
1 dag-S has a norm.
2 dag-S has an L1 norm.
3 dag-S is polynomially equivalent toS.
Proof Idea.
I (2) implies (1): Trivial.
I (1) implies (3): Ifµ=M◦P is a norm for S then it is also a measure for dag-S, and so a norm for dag-S.
I (3) implies (2): Transform proofs by just expanding out dag-steps.
Evaluating the size of the resulting proof in a clever way gives only a polynomial blow-up; this relies crucially on homogeneity of the norm.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 15 / 20
Further Results
Proposition
Cut-free tree-like Gentzen and deep inference systems have L1 norms.
Corollary
Dag-like cut-free Gentzen has no L1 norm. A cut-free deep inference systems with c↑ has an L1 norm just if it is polynomially simulated by the system without c↑.
Proposition
Tree-like Frege systems have an L1 norm just if dag-like Frege systems do. Conjecture
Neither tree-like nor dag-like Frege systems have an L1 norm.
Further Results
Proposition
Cut-free tree-like Gentzen and deep inference systems have L1 norms.
Corollary
Dag-like cut-free Gentzen has no L1 norm. A cut-free deep inference systems with c↑ has an L1 norm just if it is polynomially simulated by the system without c↑.
Proposition
Tree-like Frege systems have an L1 norm just if dag-like Frege systems do.
Conjecture
Neither tree-like nor dag-like Frege systems have an L1 norm.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 16 / 20
Further Results
Proposition
Cut-free tree-like Gentzen and deep inference systems have L1 norms.
Corollary
Dag-like cut-free Gentzen has no L1 norm. A cut-free deep inference systems with c↑ has an L1 norm just if it is polynomially simulated by the system without c↑.
Proposition
Tree-like Frege systems have an L1 norm just if dag-like Frege systems do.
Conclusions
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 17 / 20
Summary
I We can propose a definition of compression mechanism, or rather of being an uncompressed system:
Definition
A system is uncompressed just if it has an L1 norm.
Under this definition, deep inference would not be regarded as a compression mechanism, as without dag, cut and substitution the basic system still has an L1 norm.
I We formulated compression mechanisms dag and cut in a way that is independent of any particular system, but nonetheless does the same job from the point of view of proof complexity.
I We characterised the situations when an uncompressed system p-simulates its dag-like counterpart for a wide variety of systems.
Summary
I We can propose a definition of compression mechanism, or rather of being an uncompressed system:
Definition
A system is uncompressed just if it has an L1 norm.
Under this definition, deep inference would not be regarded as a compression mechanism, as without dag, cut and substitution the basic system still has an L1 norm.
I We formulated compression mechanisms dag and cut in a way that is independent of any particular system, but nonetheless does the same job from the point of view of proof complexity.
I We characterised the situations when an uncompressed system p-simulates its dag-like counterpart for a wide variety of systems.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 18 / 20
Summary
I We can propose a definition of compression mechanism, or rather of being an uncompressed system:
Definition
A system is uncompressed just if it has an L1 norm.
Under this definition, deep inference would not be regarded as a compression mechanism, as without dag, cut and substitution the basic system still has an L1 norm.
I We formulated compression mechanisms dag and cut in a way that is independent of any particular system, but nonetheless does the same job from the point of view of proof complexity.
I We characterised the situations when an uncompressed system p-simulates its dag-like counterpart for a wide variety of systems.
Summary
I We can propose a definition of compression mechanism, or rather of being an uncompressed system:
Definition
A system is uncompressed just if it has an L1 norm.
Under this definition, deep inference would not be regarded as a compression mechanism, as without dag, cut and substitution the basic system still has an L1 norm.
I We formulated compression mechanisms dag and cut in a way that is independent of any particular system, but nonetheless does the same job from the point of view of proof complexity.
I We characterised the situations when an uncompressed system p-simulates its dag-like counterpart for a wide variety of systems.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 18 / 20
Summary
I We can propose a definition of compression mechanism, or rather of being an uncompressed system:
Definition
A system is uncompressed just if it has an L1 norm.
Under this definition, deep inference would not be regarded as a compression mechanism, as without dag, cut and substitution the basic system still has an L1 norm.
I We formulated compression mechanisms dag and cut in a way that is independent of any particular system, but nonetheless does the same job from the point of view of proof complexity.
Further Directions
I Can we extend the main result to account for other compression mechanisms?
Conjecture
A system with an L1 norm cannot polynomially simulate any extension of it that does not have an L1 norm.
I A specific application of this result might be a robustnesstheorem for cut-free deep inference systems.
Conjecture (Idea)
All regular analytic deep inference systems with L1 norms are polynomially equivalent.
for some definitions of ‘regular’ and ‘analytic’.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 19 / 20
Further Directions
I Can we extend the main result to account for other compression mechanisms?
Conjecture
A system with an L1 norm cannot polynomially simulate any extension of it that does not have an L1 norm.
I A specific application of this result might be a robustnesstheorem for cut-free deep inference systems.
Conjecture (Idea)
All regular analytic deep inference systems with L1 norms are polynomially equivalent.
for some definitions of ‘regular’ and ‘analytic’.
Further Directions
I Can we extend the main result to account for other compression mechanisms?
Conjecture
A system with an L1 norm cannot polynomially simulate any extension of it that does not have an L1 norm.
I A specific application of this result might be a robustnesstheorem for cut-free deep inference systems.
Conjecture (Idea)
All regular analytic deep inference systems with L1 norms are polynomially equivalent.
for some definitions of ‘regular’ and ‘analytic’.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 19 / 20
Further Directions
I Can we extend the main result to account for other compression mechanisms?
Conjecture
A system with an L1 norm cannot polynomially simulate any extension of it that does not have an L1 norm.
I A specific application of this result might be a robustnesstheorem for cut-free deep inference systems.
Conjecture (Idea)
All regular analytic deep inference systems with L1 norms are polynomially
for some definitions of ‘regular’ and ‘analytic’.
Further Directions
I Can we extend the main result to account for other compression mechanisms?
Conjecture
A system with an L1 norm cannot polynomially simulate any extension of it that does not have an L1 norm.
I A specific application of this result might be a robustnesstheorem for cut-free deep inference systems.
Conjecture (Idea)
All regular analytic deep inference systems with L1 norms are polynomially equivalent.
for some definitions of ‘regular’ and ‘analytic’.
Anupam Das (University of Bath) Proof Compression Mechanisms LAC-GeoCal, 24/11/2011 19 / 20